N. K. Bose, Multidimensional Systems Theory and Applications-Springer Netherlands (1995) PDF

MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS
Mantle and Lower Crust

Exposed in Oceanic Ridges
and in Ophiolites
Contributions to a Specialized Symposium of the
VII EUG Meeting, Strasbourg, Spring 1993
Edited by
R.L.M. VISSERS
Faculty of Earth Sciences,
Geodynamics Research Institute,
Utrecht, The Netherlands
and
A. NICOLAS
Laboratoire de Tectonophysique,
Universite des Sciences et Techniques du Languedoc,
Montpellier, France
Springer-Science+Business Media, B.V.

A c.I.P. Catalogue record for this book is available from the Library of Congress
ISBN 978-90-481-6424-0 ISBN 978-94-017-0275-1 (eBook)

DOI 10.1007/978-94-017-0275-1
Printed on acid-free paper
All Rights Reserved

© 1995 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 1995
Softcover reprint of the hardcove 1st edition 1995
No part of the material protected by this copyright notice may be reproduced or
utilized in any form or by any means, electronic or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owner.
This book is dedicated to the
belief that the only certainty
in the fostering of
understanding between
people scattered around the
globe lies in the harnessing
and cultivation of the
intellectual and spiritual
resources of individuals,
irrespective of origin, for the
ultimate benefit and
sustenance of the global
community. On a personal
note, N. K. Bose wishes to
thank those he had failed to
acknowledge in his earlier
books for their faith and
friendliness. Specific
mentions are made of
chotomashi, balumamu,
shejomashi, and chotopishi.
Table of Contents
List of Acronyms xi
List of Notations xii

Preface xiii
Acknowledgments xvii
Introduction xix
1. TRENDS IN MULTIDIMENSIONAL SYSTEMS THEORY

1 Introduction 1
2 Multidimensional Systems Stability 2
2.1 Multidimensional Digital Filters 3
2.2 Multivariate Networks 9
2.3 Delay-Differential Systems 12
2.4 Stiff Differential Systems 13
2.5 Multipass Processes 15
3 Multivariate Realization Theory 15
3.1 State-space Realization Theory 16
3.2 Transform Domain Realization Theory 18
3.2.1 Integral Representation of Positive Real Functions 20
3.2.2 Sum of Squares Representation of a Form in Network Synthesis 22
4 n-D Problem of Moments and It., Applications in Multidimensional Systems
Theory 23
5 Role of Irreducible Polynomials in Multidimensional Systems Theory 26
6 Hilbert Transform and Spectral Factorization 27
7 Conclusions 28
8 Updates 30
vii
viii MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS
2. CAUSAL AND WEAKLY CAUSAL 2-D FILTERS WITH APPLICATIONS

IN STABILIZATION 35
1 Scalar 2-D Input I Output Systems 35
2 Stability 38
3 Structural Stability 43
4 Multi-Input Multi-Output Systems 43
5 Stabilization of Scalar Systems 44
6 Characterization of Stabilizers for Scalar Systems 47
7 Stabilization of Strictly Causal Transfer Matrices 53
7.1 MIMO Feedback Systems and Their Stabilization 55
8 Characterization of Stabilizers for MIMO Systems 59
9 Stabilization of Weakly Causal Systems 63
10 Stabilization of MIMO Weakly Causal Systems 73
11 Conclusions 77
12 Updates 77
3. THE EQUATION Ax = b OVER THE RING qz, w] 79

1 Introduction 79
2 Sufficient Condition For Solution 79
3 Appendix A. Zero-Dimensional Polynomial Ideals 86
4. GROBNERBASES: AN ALGORITHMIC METHOD IN POLYNOMIAL

IDEAL THEORY 89
1 Introduction 89
2 Grobner Bases 91
3 Algorithmic Construction of Grobner Ba<;es 95
4 An Improved Version of the Algorithm 98
5 Application: Canonical Simplification, Decision Of Ideal Congruence
And Membership, Computation In Residue Class Rings 103
6 Application: Solvability and Exact Solution of Systems of Algebraic
Equations 108
7 Application: Solution of Linear Homogeneous Equations With Polynomial
Coefficients 116
8 Grobner Bases For Polynomial Ideals Over The Integers 120
9 Other Applications 124
10 Specializations, Generalizations, Implementations, Complexity 125
11 Updates 126
Table of Contents ix
5. MULTIVARliITEPOLYNOMIALS, MATRICES, AND MATRIX-FRACTION

DESCRIPTIONS 129
1 Introduction 129
2 Relative Primeness and GCD Extraction from Multivariate Polynomials 131
2.1 Test') for Relative Primeness 131
2.2 Primitive factorization algorithms for GCD extraction 133
2.2.1 Multivariate GCD extraction from the Sylvester matrix 133
2.3 Bezout Matrix Ba<;ed Method 135
2.3.1 Multivariate Polynomial GCD Extraction from Bezout Matrix 136
3 Polynomial Matrix Primitive Factorization in the Bivariate Ca')e 138
3.1 Primitive Factorization of Bivariate Polynomial Matrices 140
3.2 General Factorization of Bivariate Polynomial Matrices 147
4 Multivariate Polynomial Matrix Factorization 151
4.1 Various Types of Multivariate Polynomial Matrix Factorizations 151
4.2 Modules 151
4.3 Various Degrees of Primeness 155
4.4 R-Modules and Vector Spaces 159
5 Computations for Coprimeness Using Grabner Ba<;es 161
5.1 Algorithm for Computing the Right Inverse of ZLC Matrices 162
5.2 Minor coprimeness and Grobner Bases 163
5.3 Problems in Factor Coprimeness and GeLD (GCRD) Extraction 168
6 Generalization of the Serre Conjecture and its Consequences 170
6.1 Normal Full Rank Matrix Case 170
6.1.1 Factorization Problem in Normal Full Rank Case 174
6.2 Degenerate Rank Ca<;e 174
7 Factorization as a Product of Elementary Matrix Factors 176
8 Applications in Multidimensional Systems Stabilization 178
9 Behavioral Approach 184
10 Conclusions 187
6. RECENT IMPACTS OF MULTIDIMENSIONAL SYSTEMS RESEARCH 191
1 Introduction 191
2 Inference of Stabilty of Sets of Multidimensional Systems From Subset<;
of Low Cardinality 192
2.1 Univariate Complex Hurwitz Polynomial Ca<;e 192
2.2 Multivariate Scattering Hurwitz Polynomial Case 193
2.3 Robust Wavenumber Response 198
2.4 Stability Of Interval Matrix 199
3 Multiple Deconvolution Operators for Robust Superresolution 200
x MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS
3.1 Recursive Least Squares (RLS) Algorithm for Superresolution 202

3.2 Backprojection Algorithm for Superresolution 206
4 Multisensor Array-Based Superresolution 207
4.1 MuItisensor High Resolution Image Reconstruction Model 209
4.2 Regularization Methods 211
5 Wavelets for Superresolution 213
6 Other Recent Applications 216
7 Conclusions 218
7. MULTIVARIA1ERATIONALAPPROXIMANTS OFTHEPADE-TYPE 221
1 Introduction and Motivation 221
2 Multivariate Pade-lYpe Approximanlo; (Scalar Cao;e) 222
3 Pade-lYpe Matrix Approximants 226
4 Conclusions 227
5 Updates 228
8. OLD PROBLEM SOLUTIONS AND SOME NEW PROBLEMS 231
1 Progress Towards Solution of Open Problems since 1985 231
1.1 Nonessential Singularities of the Second Kind (NSSK) 231
1.2 Rational Spectral Factorization 232
1.3 Feedback Stabilization 233
1.4 Robust Continuous System Stability 235
1.5 Stability of Discrete Systems Under Coefficient Perturbation 236
1.6 Robust Stability of Multivariable Systems 237
2 New Open Problems 238
2.1 Greatest Common Left (Right) Divisor Computation 238
2.2 DeterminantaI (Left Right Joint) Factorization 239
2.3 Unimodular Completion Over Ring of Proper Stable Multivariate
Rational Functions 240
2.4 Minimax Controller and Grobner Bao;es 242
References 243
Index 265
xi
List of Acronyms
BCCB block circulant with circulant blocks
BIBO bounded inputlbounded output
BTHTHB block Toeplitz-plus-Hankel with Toeplitz-plus-Hankel blocks
BTIB block Toeplitz with Toeplitz blocks
CFf continuous Fourier transform
OCF double coprime factorization
OCT discrete cosine transform
DFD determinantal factorization domain
DFf discrete Fourier transform
FIR finite impulse response
FLC factor left coprime
FLP factor left prime
FQQP first quadrant quarter plane
GCD greatest common divisor
GCDD greatest common divisor domain
GCLD greatest common left divisor
GCRD greatest common right divisor
GFLP generalized factor left prime
IIR infinite impulse response
LCM least common multiple
LCR least common reducible
LMFD left matrix fraction description
LMSD linear multistep multiderivative
LSI linear shift-invariant
MDKN multidimensional Kirchhoff network
MDWD multidimensional wave digital
MFD matrix fraction description
MIMO multi-input multi-output
MLA minimal left annihilator
MLC minor left coprime
MLP rrrinor left prime
MRA minimal right annihilator
MRP minor right prime
NSSK nonessential singularity of the second kind
PCGA preconditioned conjugate gradient algorithm
RMFD right matrix fraction description
SHS scattering Hurwitz stable
SOS sum of squares
SSS strictest sense Hurwitz stable
UFD unique factorization domain
WLP weakly zero left prime
WSS widest sense Hurwitz stable
ZLC zero left coprime
ZLP zero left prime
ZRC zero right coprime
ZRP zero right prime
xii MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS
List of Notations
Notations are described when they occur in text proper. The list below is included
for quick reference to further facilitate reading.
Symbol Meaning
"Equals by definition"
(/1, ... , fr) Ideal generated by /1, ... , Ir
[hij]mxle m x k matrix with entries hi;
(m:l) (mH)!
mIl!
Omxle
A + or AdjA or Aadi
AE9B
A(Un )
a(z,w)
C
DM
1M
Di,!i
9* (PI , ... ,Pn)
lie
K[z] = K[Zl, ... , zn]
K[z,w]
]R
lR[z\zd
Ra[z,wj
]RmXl(p)
R mxk
Torus: TxT x ... x T~ T: {z E c: Izl = I}

n-copies
Open unit polydisc: U xU; ... xU', U: {z E C: Izl < I}
..
n-copies
Closed unitpolydisc: iT x U x ... x U, U: {z E C: Izl ~ I}

Ring of integers
Set of positive integers
Finite field with prime q
z Indeterminates (Zl, Z2, ... , Zn)
(Zll, ... ,Z;l)
Preface
Preface to Second Edition
From the time that the original edition was published in 1985, multidimensional
systems theory has matured into a discipline of research and teaching with an
expanding array of applications. The international journal on Multidimensional
Systems and Signal Processing, founded in 1990, is now in its fourteenth year. A
biannual international workshop on n-D systems was launched in 1998 and the
impressive number of special sessions, mini-symposia, monographs and special
issues that have emerged bear testimony to the growing popularity and importance
of the subject-matter among scientists in various disciplines including engineering,
computer science, geophysics and mathematics.
This second edition builds on the fundamentals expounded in the original book
with the addition of important developments in theory as well as practice since
1985. Particular attention has been given to the consolidation of basic results, uni-
fication of theory and the diversification of applications. Chapters that remain have
been reordered and updated in content and references. Some chapters, considered
to be somewhat outdated, have been replaced with newer proven as well as poten-
tially significant results, inspired by some groundbreaking research and directions
which are likely to stimulate further research. In addition to the description of
some challenging open problems, posed in 1985, which have since been solved,
new problems yet to be tackled are also included.
The chapters kept intact from the original edition but with updates are Chapter
1, Chapter 2 (present Chapter 7), Chapter 3 (present Chapter 2), Chapter 6 (present
Chapter 4) and Chapter 7 (present Chapter 3). Chapters 4 and 5 in the original
edition are replaced by more relevant and current results in Chapters 5 and 6 in
the current edition. The present Chapter 5 has been necessitated by the theoret-
ical developments following the work of Guiver and Bose and, subsequently of
Oberst in 1990, which led to a flurry of papers during the last decade both in theory
and in applications of the theory. The contents of this chapter interface well with
xiii
xiv MULTIDIMENSIONAL SYSTEMS THEORY AND APPUCATIONS
earlier chapters by supplementing as well complementing the material there. The

present Chapter 6 was motivated, first, by Kharitonov's groundbreaking results on
root clustering of a set of polynomials followed by Bose's nontrivial generaliza-
tion to the multivariate case (1988) which generated considerable research activity.
Chapter 6 also describes the progress towards the attainment of robust blind su-
perresolution from a sequence of low resolution degraded images, motivated by
fundamental results expounded in earlier chapters on multivariate polynomial co-
primeness. This chapter also alerts the reader to other arenas of application of the
developed theory. The ordering of the chapters allow the potential instructor to
proceed from an overview (Chapter I), to systems over rings with a selected domain
of application for illustration and emphasis (Chapter 2), solution of a linear sys-
tem of equations over a bivariate polynomial ring, for the mathematically minded,
(Chapter 3), algorithmic algebra and computational complexity of computer al-
gebra algorithms (Chapter 4), the most fundamental issues in polynomial matrix
algebra and their analysis from the polynomial ring, ideal and module-theoretic
standpoints with applications (Chapter 5), selected cutting-edge theory and appli-
cations (Chapter 6), and a brief exposure to rational approximants (Chapter 7). A
status of attempts made to solve the open problems in the first edition and, because
of page constraints, only a selected handful of new open problems are compiled
in Chapter 8. The list of acronyms and the subject index should further facilitate
reading of the text.
This second edition should serve as a guide to research as well as education till
the time a textbook (acknowledged to be a mammoth and time-consuming task)
emerges. The attractiveness of the new edition stems from the cohesiveness of
the fundamental results presented concisely and compactly, but without loss of
essential details necessary for complete comprehension, and without losing sight
of the applicability of these results.
The research that led to this update was supported at various stages by grants
from the National Science Foundation, Army Research Office, Air Force Research
Laboratory and the Humboldt Foundation. I also wish to acknowledge the support
and cooperation of the staff of Kluwer Academic publishers and many colleagues
and students from allover the world who made this challenging task worthwhile
and enjoyable.
N. K. BOSE
Preface to First Edition (1985)
Towards the end of 1981, Professor M. Hazewinkel invited me to write a book

on multidimensional systems theory in his series devoted to mathematics and its
applications. At that time, however, I had already completed the manuscript for
the text entitled, 'Applied Multidimensional Systems Theory', which was pub-
lished by Van Nostrand Reinhold Co. during the first quarter of 1982. Though, at
that time, the need for a textbook on the subject was realized, it was felt that the
area of multidimensional systems theory was expanding at a prolific rate from
the standpoints of both theoretical and applied research. This was brought about,
sometimes, by independent groups of researchers with different academic back-
grounds who worked without the benefit of interaction between groups. Since the
subject has a rich mathematical flavor, any attempt to promote dialogue between
engineers emphasizing theory and mathematicians willing to gear their research
towards applications, is, unquestionably, very beneficial. This book, then, gradu-
ally evolved out of the realization of the necessity of documenting the significant
progress and novel directions of research in the arena of multidimensional systems
theory, since the publication of my textbook, with the ultimate objective of alert-
ing the reader about the existence of an apparently inexhaustible source of open
problems, only some of which could be explicitly stated in this book.
The first chapter contains a description of very recent results in topics of fun-
damental importance in multidimensional systems theory. Perusal of this chapter
is likely to convince the reader of the broadening scopes for applications of either
some key theoretical results or limitations to such applications caused by problems,
which still defy complete, satisfactory, solutions. Since, approximation theory is
a very vast topic, only multivariate rational approximants of the PadC-type has
been briefly surveyed in Chapter 2, because this type of rational approximation
technique and its variants are popular in various problems of multidimensional
systems theory. A rigorous analysis of a type of 2-D feedback system is provided
in Chapter 3, where the advantages of incorporating the weakly causal feature in
the design of stabilizing compensators is highlighted. Multidimensional feedback
xv
xvi MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS
systems have been proposed in several applications, and the future role of feed-
back in the design of multidimensional structures should not be underestimated. In
Chapter 4, the problems of existence and construction of feedback compensators
for spatio-temporal systems, whose inputs and outputs are functions of a tempo-
ral and a spatial variable, are studied. The approach in Chapter 3 is a transform
domain approach based on tools from algebra while that of Chapter 4 is based on
techniques from the theory of linear systems with coefficients in a commutative
ring and the criteria for stabilizability are specified in terms of a state-space rep-
resentation (and also a transfer matrix representation). Chapter 5 introduces the
reader to the modeling, analysis, and applications of linear, shift-variant multidi-
mensional systems while Chapter 6 describes the theory of Grobner bases which
are known to have proven as well as potential applications in multidimensional
systems theory problems. Chapter 7 considers conditions for the solution of a
system of equations over the ring C[z, w], motivated by the availability of some
recent results in the mathematical literature and the applicability of the solution to
the problem considered to one formulated in Chapter 3.
This book is aimed towards fulfilling the needs of those mathematicians who
want to learn about current applications in an area requiring a wide variety of
mathematical resources and also towards scientists, engineers, and researchers
in industries and universities who want to keep abreast of latest developments
in multidimensional systems theory, the directions along which this subject is
expanding and some of the problems that are, currently, open to investigation. The
book could also be, effectively, used in advanced seminars, selected continuing
education courses and, possibly, as either a text or supplement to a text adopted
for courses in multidimensional systems theory or spatial and temporal signal
processing.
The book could not be written without the cooperation of the contributors.
Special thanks go to Dr. J. P. Guiver, who stimulated many interesting discussions
and provided useful insights to some research problems during his stay, here, at the
University of Pittsburgh. I have greatly benefited from interaction with colleagues
at various universities, who made available to me many of their research results
before those formally appeared in journals. The brief survey in Chapter 2 was
partially influenced by the doctoral dissertation of A. Cuyt. The motivation and
help provided by other colleagues is partly reflected in the contents of this book
and I shall not attempt to acknowledge everyone here because, inadvertently, I
could omit a name that should appear. I wish to express my sincere gratitude to
the Air Force Office of Scientific Research, where I have been fortunate to work
in Dr. Joseph Bram's program in the Directorate of Mathematical and Information
Sciences, and to the National Science Foundation for continuing support of the
research I have been conducting with my group of researchers. Some results of
that research are included in this book. I thank Professor M. Hazewinkel for inviting
me to contribute in his useful series and to the D. Reidel Publishing Company for
their cooperation in the successful completion of this project.
Acknowledgments
While two-dimensional systems theory was understood well prior to 1985, the
possibilities and intrinsic difficulties in n-D, n ~ 3 have been understood more
fully during the last fifteen years or so. This is largely due to the progress in n-D,
n ~ 3 polynomial matrix theory and matrix fraction descriptions on one hand and
the behavioral approach on the other. The impact of the study of a collection of
trajectories or behavior on multidimensional systems theory has been briefly noted
paving the way for a fuller documentation, possibly in a textbook setting, later.
The author benefited from direct or indirection interaction with several former stu-
dents, who wrote their doctoral dissertations in the area, and colleagues with whom
the author had either prolonged research interactions or communication through
e-mail exchanges and discussions during meetings and workshops. Thanks are
conveyed to Sankar Basu, Ki-Jin Boo, Chalie Charoenlarpnopparut, Alfred Fet-
tweis, Surapong Lertrattanapanich, Zhiping Lin, Michael Ng, Ulrich Oberst, Isao
Yamada, and Li Xu. The previous collaboration with John P. Guiver and the past
as well as current interaction with Bruno Buchberger are also explicit in this book.
Assistance from Mahesh Babu Chappali and Surapong Lertrattanapanich in the for-
matting, computer processing and proofreading of this book is acknowledged. The
cooperation extended by Lynn Brandon and Marlies Vlot from Kluwer Academic
Publishers is gratefully noted. The most recent grants from the Army Research
Office and the National Science Foundation in the programs managed by William
Sander and Radhakishen Baheti provided the impetus for some of the new results
reported, respectively, in Chapter 6 and Chapter 5. Some of the research and doc-
umentation was facilitated by the support for research from the Alexander von
Humboldt Foundation through the Humboldt Research Award. This book would
not have been completed as scheduled without the sabatticalleave granted to N.
K. Bose by the Pennsylvania State University.
xvii
Introduction
In the Editor's Preface for the first edition, Professor Michiel Hazewinkel noted
in March 1985, "Modem technological demands such as for image processing and
the remarkable successes of 1-D theory (i. e. the usual kind of mathematical
system theory) caused the emergence of 2-D (and n-D) theory. This is a field in
which things are in rapid flux calling for an occasional taking stock of the situation.
This is exactly what this book aims to do. "
In July 2003, more than eighteen years after the above words of wisdom were
communicated, we find the need again to take stock of things for future teaching
and research in mathematical system theory and its expanding applications. A
need for studying systems over rings and modules has been felt. Technical devices
from algebraic analysis, algebraic geometry, arithmetic geometry, theory of partial
differential equations, algorithmic algebra, computational efficiency in the imple-
mentation of algorithms in computer algebra are being used routinely by different
groups of scientists conducting research in multidimensional systems and multi-
dimensional signal processing. It is fair to say that the 2-D theory is resting on
a solid foundation and courses are being taught based on textbooks available for
that purpose. The jump in difficulty in the comprehension of higher dimensional
systems theory, permitting as well as defying generalizations, from 2-D has been
appreciated and much better understood during the last decade or so. This second
edition brings the reader on par with these latest developments and prepares him
to apply the theoretical knowledge gained to cutting-edge technological problems.
At the same time, the reader is left with the satisfying feeling that a dead-end is
nowhere in sight. An abundance of problems, in theory as well as applications,
will, undoubtedly, produce many fine research papers and dissertations in the future
based on a thorough perusal of this comprehensive edition.
xix
Chapter 1
TRENDS IN MULTIDIMENSIONAL SYSTEMS THEORY

BYN.K.BOSE
1. Introduction
The theories of functions and polynomials of several complex and/or real vari-
ables along with their numerous applications in several areas of systems theory
primarily concerned with the topics of multidimensional digital filter stability, sta-
bilization and design, multivariate network realizability theory, digital array pro-
cessing in the general framework of multidimensional signal processing techniques
needed to process signals carried by propagating wave phenomena, in addition to
problems occurring in control theory concerned with 2-D state-space models, no-
tions of controllability, observability, minimality, feedback and pole-placement as
well as stabilization via output feedback provided the subject matter of a Special
Issue devoted to Multidimensional Systems [1]. Though the vast majority of the
papers in [1] were concerned with multidimensional deterministic systems, a paper
concerned with methods for inference about models of random processes on mul-
tidimensional Euclidean space from observed data was also included especially
because the mathematical tools employed partially fitted those widely used in the
Issue and also because it was felt at that time that the future scopes for develop-
ment of multidimensional systems theory should not be restricted to deterministic
systems only. In fact, broadly speaking, multidimensional systems theory spans
deterministic and statistical approaches to the modeling, analysis and design of
spatio-temporal continuous and discrete systems. A collection of reprints geared
particularly to the developments in the area of multidimensional signal processing
(deterministic as well as statistical) till about 1977 has been compiled in [2]. As a
result of increasing research activity in the area and the realization of the necessity
to encourage interaction between computer scientists, engineers, and mathemati-
cians, a collection of reprints, which characterized the significant developments
that took place in the domain of multidimensional systems theory till about 1979,
1
2 MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS
was published in [3], where an introductory survey article and a list of open prob-
lems were also included. The importance of documenting the scattered research
results in an unified form so that the theoretical fundamentals, even though based
on relatively advanced and broad range of mathematical topics, could be presented
to advanced graduate students and research scientists in a classroom or seminar
types of settings, gave rise to the need to select and expound fundamental results
of proven and potential significance in the area of mathematical multidimensional
systems theory. This was especially so because though a sizeable number of
books had appeared of an applied nature, especially in the area of image pro-
cessing, books devoted exclusively to the more mathematical aspects that support
such applications were non-existent at that time. One complete chapter in [4] was
devoted to multiparameter systems including two-dimensional filters, distributed
processes and statistical and probabilistic models for random fields. More recently,
a comprehensive account of the developments in the theory of deterministic mul-
tidimensional systems along with an exposition of the supporting mathematical
tools required in selected branches of study where such tools are used, was given
in [5].
The objective here will be to document the progress in multidimensional systems
theory since [5] was published so that the reader is alerted to the flurry of activities
generated by certain fundamental results, some of which are finding applications
in more than one area of research. In the process it will become apparent that the
topic of multidimensional systems continues to provide challenging theoretical
problems which arise in the continuously expanding domain of applications. Ef-
forts will be made to relate, briefly but succinctly, past efforts, present status and
future trends so that the reader following his perusal of the contents of this chapter
will not only be able to appreciate better the contents of the succeeding chapters but
also will become cognizant of the resources available to tackle the open problems,
implicitly as well as explicitly mentioned in this book. Though it is recognized
that a fundamental concept, result, or theoretical limitation, usually affects more
than one area of application and attempts to segment the range of applications to
distinct domains is sometimes futile, it is felt that when, for the sake of clarity in
exposition, it becomes necessary to consider one particular topic in one particular
section then the relevant links with other topics or applications need to be cited
and attended to. It is, therefore, hoped that the topics selected for discussion in
the following sections will represent and reveal the existence of strong coupling
between theoretical fundamentals and applications that fall under the umbrella of
multidimensional systems theory.
2. Multidimensional Systems Stability
The concept of stability plays an important role in various areas that fall un-
der the jurisdiction of multidimensional systems theory. Akin to the occurrence
Trends in Multidimensional Systems Theory 3
of several fundamental stability criteria in other fields of science (for example,

stability of solutions of differential equations in nonlinear systems theory could
be investigated under the definitions of Lyapunov stability, asymptotic stability,
conditional stability, orbital stability, stability under perturbation, etc.), the topic
of multidimensional systems also uses the concept of stability in various context-
dependent forms. An exposition of the state of this topic until about the end of 1981
is given in the book by Bose [5], whose coverage of stability is geared towards lin-
ear shift-invariant multidimensional discrete systems, delay-differential systems,
lumped-distributed and variable-parameter networks and to a lesser extent, stiff
differential systems occurring in discussions of stability questions for numerical
methods. Here, the important characteristics and developments of stability theory
in these and related areas will be highlighted and attention will be directed to some
questions that need to be satisfactorily resolved.
2.1 MULTIDIMENSIONAL DIGITAL FILTERS
The most commonly used criterion for assessing the stability of linear shift-
invariant (LSD multidimensional digital filters is the bounded inputlbounded output
(BmO) criterion, which is well known to be equivalent to the requirement of
absolute summability of the impulse response sequence characterizing the filter.
A LSI n-dimensional recursive filter is also characterizable by a transfer function,
H( Zl, Z2,···, Zn ) _- B(
A(Zl' Z2,···, zn)
) (1.1)
Zl,Z2,···, Zn
whereH(Zl, Z2,.·., zn) is viewedasarationalfunctionofzl, Z2,···, Zn, zit, zit,
... , z;;l (recursible filters include causal and weakly causal filters). Fornotational
convenience, the z-transform H(Zl' Z2, .. · ,zn) of an n-D sequence,
h[k 1, k2, ... , kn], will be defined as a power series involving the superposition of
the products of monomials of the type Z~l , Z~2 , ... ,z~n and the generic element,
h[k 1, k2, ... , kn], of the sequence. Physically, the indeterminates ZI, Z2, ... , Zn
are the respective delay variables along the spatial or temporal directions of sam-
pling during the analog to digital conversion of a multidimensional spatio-temporal
signal. In the case of first quadrant quarter-plane filters, A(Zl' Z2, ... , zn) and
B(Zl' Z2,'" ,zn) are polynomials. Since a ring isomorphism (see Chapter 2)
maps, bijectively, a weakly causal filter onto a first quadrant quarter-plane fil-
ter, A(Zl' Z2,." ,zn), B(Zl' Z2,·.·, zn), will be understood to be relatively prime
polynomials in the delay variables Zl, Z2,· .. , Zn unless mentioned otherwise.
For a first quadrant quarter-plane (FQQP) digital filter, H(Zl' Z2, ... , zn), is
assumed to be holomorphic around the origin ( this is assured by assuming
B(O, 0, ... ,0) -1= 0 ), thereby permitting a Taylor series expansion,
L ... L h[kl, k2, ... , kn]Z~l ... z~n

00 00
H(Zl' Z2,· .. , zn) = (1.2)

kl=O kn=O
The investigation into BIBO stability reduces to the determining of conditions

under which the sum
00 00
L ... L Ih[k1, k2"'" kn]1 (1.3)

kl=O kn=O
converges. 2 Denoting by fr, un, and Tn the closed unit polydisc (IZil :::; 1, i =
1, ... n), the open unit polydisc (IZil < l,i = 1, ... n), and the distinguished
boundary (IZil = 1, i = 1, ... n), itis well-known that convergence of(1.3) implies
uniform convergence in fr of (1.2), which in turn implies that H(ZI"'" zn) is
holomorphic in un and continuous on fr. Also, if H(zI, ... , zn) is holomorphic
in a neighborhood offr, then (1.3) converges. In the n 1 case, for a rational =
function, H(zt} = [A(zt}]j[B(zd], it is simple to establish that (1.3) with n 1 =
is absolutely summable if and only if the polynomial B(zt} f= 0, IZll :::; 1. This
fact does not generalize to the n > 1 case and B(ZI"'" zn) f= 0 in fr is only
a sufficient condition for BIBO stability of H(ZI, ... , zn) in (1.1). Though, in
this context a set of necessary and sufficient conditions has not yet been obtained,
some progress made towards the attainment of that goal is worth recording. The
notations in [7] will be adopted. Letmn be the Lebesgue measure divided by (21r)n
that is carried with the compact Abelian group (with componentwise multiplication
as group operation) Tn, so that mn(rn) = 1. Also, for 0 < p < 00, let HP(U n )
be the class of all holomorphic functions, F(zI, ... , zn) in un, for which,
IlFllp ~ sup {{ !F(rz)I P dmn }I/P < 00 (1.4)

O~T<1 JTn
The following results, due to Dautov, will be stated and proved because of its
unavailability in the western literature.
THEOREM 1.1 (Dautov [8]) For the convergence of the summation in (1.3), it is
sufficient that the function
G( ZI, ... , Zn ) -~ an[H(ZI"'"

aa Zn)ZI '" zn]
a
ZI Z2··· Zn
belongs to Hl(U n ).
Proof. By differentiation,
2Absolute convergence in (1.3) implies that h[kl, k2, ... , knl's are uniformly bounded [6, p.1021, i.e. there
exists a constant K such that, Ih[kl ,k2, ... , knll ~ K, kl, k2, ... , k n = 0, 1,2, ...
Denote g[k 1, ... ,kn] £ [k1 + 1] ... [kn + 1]h[k1' ... ' kn]. The Hardy-Littlewood
inequality generalized to the n-D case yields
(1.6)
From (1.5) and (1.6), the proof of the theorem is easily completed. 0
The following definition must be given before stating the next result of Dautov,
which is useful to check whether a function belongs to H1{U n ).
DEFINITION 1.1 If F{Zl, ... , zn) is any junction in un, we define F{Zl, ... , zn)
!Jy (note that z=(Zl,.· . •zn)),
F{z) = r-t1
lim F{rz)
at every z E Tn at which this radial limit exists'.

Let £1 (Tn) denote the space of all functions integrable with respect to the
measure m n . Then, the following result was proved by Dautov.
THEOREM 1.2 [8J If F{z) is a rational junction (in reducedform) whose denomi-
natoris zero onlyatafinite number ofpoints belonging to ~ and F{z) E £1{~),
then F{z) E H1{U n ).
Dautov, subsequently, arrived at results on stability for a special class of rational
functions, whose denominator polynomials have a finite number of zeros on Tn.
These results are based on the definition for the order of a zero introduced next.
DEFINITION 1.2 For the sake of brevity, denote Ikl = k1 + ... + kn, where ki >
0, Vi and k £ (k1. k2, ...• kn ), so that the power series in (1.2), rewritten as (to
avoid cluttering of notation let zk denote Z~l , •..• z~n ),
H{z) = L· .. L h[k]zk
is holomorphic in some polydisc nn centered at the origin in en (en £ c x
c. .. x C, is the Cartesian product ofn copies ofthe complex .field C). For s = 0,
1,2, ... let Fs{z) be the sum of those terms, h[k]zk,jor which Ikl = S, so that
00
H{z) = LF8{Z).
8=0
If H{z) is not identically zero in nn, then there is a smallest S = 00 such that Foo
is not the zero polynomial. This 00 is the order of the zero which H{z) has at the
origin. Furthermore, if H{z) is holomorphic in a neighborhood of z(O), then the
order of the zero of H{z) at z(O) is the order ofthe zero of H{z + z(O» at z = O.
Next, consider an irreducible rational function,
H(z) = A(z) (1.7)

1 - blZI - '" - bnzn
where the bk'S belong to the complex field Co It is elemen~ to verify that
the denominator polynomial of H(z) in (1.7) is nonzero in U if and only if
L~=I Ibkl < 1. In case L~=I Ibkl = 1, this denominator polynomial is zero in
U'" if and only if
b*
Z - k
k - jb;f'
k = 1,2, ... , n (1.8)
The zero, then, is at

z (0) = (b*~, ... , b* )
~ (1.9)
Note that z(O) is on rn.
THEOREM 1.3 [8J H(z) in (1.7) characterizes a BIBO stable system even when it
hasanonessentialsingularityatz(O) onrn (due to the constraint, L~=llbkl = 1)
iffor polynomial A(z), the zero order at point z(O) satisfies
oz(O) (A) ~ 2(n + 1)
Proof. B(z) =1- bizi - ... - bnzn

For any complex variable w,
Iwl ~ 1 ++ 2 Re(l- w) ~ 11 - wl 2 (1.10)
In (1.9), ziO) = bklbkl- I . Define,

wk~(ziO)-lzk=lbkl[bkrIzk' k=I, ... ,n (1.11)
Using the constraint L~=llbkl = 1 and (1.11), .

n
IB(z)1 ~ ReB(z) ~ Re L Ib l(l- Wk)
k (1.12)
k=1
Since IZkl ~ 1 implies that IWkl ~ 1, therefore, z E U'" implies that W E U"'.
Substituting (1.10) in (1.12),
1 n 1 n
IB(z)1 ~ 2" LIbklll - wkl 2 ~ 2"min{l bl l,···, Ibnl} L
11- wkl 2 (1.13)
k=1 k=1
Substituting (1.11) in (1.13),
IB(z)1 ~ Klz - z(O) 12 , zEU'" (1.14)
where,
n
Iz - Z(O) 12 = L IZk - ZiO) 12
k=l
is the Euclidean norm of (z - z(O)).

By successive differentiation, it is possible to verify that
anH{Z)Zl ... Zn A1{z)

(1.15)
aZ1 •.. aZn [B{z)]n+l
where A1{z) is a polynomial whose zero order is not less than the zero order of
A{z). From the condition in the theorem, it then follows that the zero order of
A1{z) at z = z(O) is given by,
Therefore,
(1.16)
where Kl is a positive constant. Therefore, from (1.14), (1.15), and (1.16),
an H{Z)Zl ... Zn Kl
--~~~--~ < ---- z E=
U , (1.17)
az 1 .•• aZn - Kn+l'
o
Subsequently, using Theorems 1.1 and 1.2, Dautov mentioned the feasibility
for demonstrating validity of this Theorem by generalizing techniques employed
to tackle special cases of the denominator polynomial in (1.7). For an arbitrary
rational function H{z), the problem of determining conditions for BmO stability
in the presence of a finite number of nonessential singularities of the second kind on
Tn is complex and, at present, no general solution is available. Dautov [9] clarified
the problem in the 2-D case by showing that for a certain class of denominator
polynomials B{z), H{z) = A{z)/B{z) is BmO stable if and only if it can be
continuously extended to U2 from U2 ; furthermore, he conjectured that this is
-2 2
true for any B{z) whose only zeros in U are on T and these are also zeros of
A{z) (note that in the 2-D case the finiteness condition for the number of common
zeros on rn is automatically satisfied). Of course, H{z) can never be extended
analytically to if when there is a nonessential singularity of the second kind at
Z = z(O) E Tn since in any open neighborhood of z = z(O), H{z) will always
be unbounded and therefore Riemann's analytic continuation theorem will not
apply. However, if H{z) extends continuously to if one can take limits from any
direction within un to achieve the same result. The example given next shows a
8 MULTIDIMENSIONAL SYSTEMS THEORY AND APPUCATIONS
rational function H(z) which is analytic in U 2 , has a nonessential singularity of

the second kind at z = (1, 1) and does not extend on U 2 to a continuous function.
Example 1.1 Let,
H( )_ (Zl - 1)n + (Zl -1)(Z2 - 1) + (Z2 - 1)n n » 1

Zl, Z2 - (zl+ Z2- 2)2 '
For (Zl,Z2) = (e1'O'</J

,e1 ) E T 2 ,
'</J _ (e1 0 - 1)n + (e1 0 - 1)(e1</J -1) + (e1</J _l)n
'0
H(e1 ,e1 ) - (ejO -1 + ej</J _1)2
the line (J = ¢ is !.
° °
Then, the limit as (J ~ along the line ¢ = is 0, but the limit as (J ~
(This example is due to Dr. J. Murray).
° along
The famous counterexample of Goodman [10] also exhibits this property. Through
a series of neatly constructed examples, Goodman, in a prize-winning paper [to],
was the first to point out difficulties in the prevailing concept of BillO stability for
multidimensional filters. In addition to other results, he showed that
m~O, n~O (1.18)
is BillO unstable when m = n = 1, and BillO stable when m = n = 8. For the

case when m = n = 1 it is clear that H(zI, Z2) in (1.18) does not extend on U 2 to
a continuous function, since
zl~l zl~l
(1 - zt}2
· H( Zl, Z2 ) = lim 2(
11m 1 - Zl ) = °
while
lim H(1 - x 2 + jx, 1 _ x 2 _ jx) = lim x4 ~ x2 = 1
x~o x~o x
Therefore, the validity of Dautov's conjecture, will enable one to arrive at Good-
man's conclusions concerning specializations of(1.18) in a relatively simple man-
ner. Furthermore, it follows from Dautov's results [9], that H(Zl, Z2) is BillO
stable when m = n for all positive powers of the numerator in (1.18) except when
m = n = 1. In fact, it is easy to infer from [9] that H(zI, Z2) in (1.18) is BillO
stable except when the 2-tuple (m, n), takes values (0,0), (0,1), (1,0), (1,1), (2,0),
and (0,2).
Though Dautov's conjecture on 2-D filter BillO stability in the presence of
nonessential singularities of the second kind on T2 is still unresolved at least in the
open literature, a sufficient condition and a necessary condition expressed in terms
of tangents to the algebraic curve at a zero of B (Zl' Z2) on T2 have been advanced
by Alexander and Woods [11]. This necessary condition uses the fact that when
n = 2 in (1.1) the number of nonessential singularities of the second kind is finite
and isolated, while in the n > 2 case the complexity of the problem increases
considerably since the locus of those singularities is of real dimension 2n - 4
in a space of real dimension 2n. The sufficient condition for Bmo stability, as
given by Alexander and Woods, uses a result by Zak [12] involving an extension to
two variables of a well-known theorem on the absolute convergence of the Fourier
series expansion of a periodic function of bounded variation and belonging to class
Lipschitz a. Though the problem of the resolution of nonessential singularities of
the second kind in the context of BmO stability of multidimensional filters remains
to be satisfactorily tackled, the tests for the absence of zeros of a multivariate
polynomial in a unit polydisc,
(1.19)
can be implemented in various forms, proceeding from criteria resulting from

Rudin's theorem, which itself is provable by using simple one-variable arguments
exclusively, as done by Delsarte-Genin-Kamp [13]. Note that B(O, 0, ... ,0) =I- 0,
the condition necessary to permit the power series expansion in (1.2) of the rational
function H(z), allows the test in (1.19) to be replaced by the test in (1.20), where
ml is the degree in ZI of polynomial B(ZI,"" zn)
-;-;11-1
(Z2,"" zn-d E U (1.20)
The various algebraic tests to test for the conditions in (1.19) and (1.20) are given
in [5], [14], and [15].
2.2 MULTIVARIATE NETWORKS
Various types of electrical networks like lumped-distributed and variable pa-

rameter networks can be studied within the setting of multidimensional systems
theory. The various definitions for multivariate Hurwitz polynomials and its ram-
ifications have been discussed by Bose [5] in his review of the realizability theory
for multivariate rational functions in the complex variables p ~ (pl,P2,'" ,Pn)'
Fettweis [16] introduced a particular class of Hurwitz polynomials, referred to as
principal HUIWitz polynomials or scattering HUIWitz polynomials in view of their
central role in the scattering transfer matrix description of multivariate lossless two
ports encountered in the synthesis of reference filters as a prelude to the derivation
of wave digital filter structures. For the sake of generality, complex coefficients
will be permitted in a polynomial.
DEFINITION 1.3 A polynomial, g(p) will be called a scattering HUIWitz polyno-

mial or, equivalently, a principal HUIWitz polynomial, if it satisfies the following
properties:
(a) g(p) =1= 0, Re p > 0, i.e. the open right-halfpolydomain and

(b) g(p) is relatively prime to the paraconjugate polynomial, described next.
The paraconjugate of g(p) is the polynomial g*(p) ~ g*( -p*)

= g*( -pi, -P2"'" -p~), where the star superscript represents the operation
of complex conjugation. (Note that when the polynomial coefficients are real,
g*(p) ~ g( -p) = g( -PI, -P2,"" -Pn).
Subsequently, Fettweis showed the equivalence of Definition 1.3 to Definition
1.4 given next in a slightly specialized but somewhat simpler form than originally
proposed.
DEFINITION 1.4 A polynomial, g(p), is a scattering HUIWitz polynomial if and

only if
(a) g*(p) and g(p) are relatively prime and
(b) I~(~)II ~ 1, Re p > 0
The key to the proof of the equivalence of the preceding two definitions is the
development (and use) of the maximum-modulus principle to situations where
nonessential singularities of the second kind are allowed to occur on the boundary
of the open polydomain, Re p > 0, for the rational function [Ig* (p) 1l/[lg(p) I]. An
interesting outcome of the introduction of the previous two definitions is the mul-
tivariate counterpart of a well-known result in passive network synthesis. Define
the para-even and para-odd parts, respectively, of g(p),
ge(P) = f.q(p) ~ g*(p)], go(p) = f.q(p) - g*(p)] (1.21)

2
Then, the following result holds.
FACT 1.1 If g(p) is a scattering HUIWitz polynomial of degree then the

~ 1,
quotient of ge(P) and go(p) is a nontrivial (that is not zero or an imaginary
constant) multivariate reactance junction in irreducible form. Vice versa, if the
quotient of two relatively prime polynomialv, ge(P) and go(p), is a nontrivial
reactance junction then their sum is a scattering HUIWitz polynomial of degree
~1.
It is mentioned that other definitions used with regard to multivariate Hurwitz
° °°
polynomials include the imposing of restrictions g(p) =1= 0, Re p > (g(p) is
then, referred to as widest sense Hurwitz) and g(p) =1= 0, Re p ~ (g(p) is, then,
j =1= ifori = 1,2, ... ,n,j = 1,2, ... ,n,g(p) =1= O,RePi = O,RePj > °
referred to as strictest sense Hurwitz), g(p) =1= 0, Re Pi > 0, Re Pj = when
when j =1= i for i = 1,2, ... , n,j = 1,2, ... , n (g(p), in the previous two cases,
when n = 2 has been called narrow sense Hurwitz) and various other obvious
ramifications. Widest sense Hurwitz polynomials have been recently considered
by Gregor [17]. It should be noted that the test for the scattering Hurwitz property
requires the test for a polynomial to be widest sense Hurwitz. While algebraic
tests exist to determine whether or not a polynomial is strictest sense Hurwitz,
efficient tests for establishing whether or not a polynomial is devoid of zeros in the
open right-half polydomain are now being developed. A strictest sense Hurwitz
polynomial is obtainable from a polynomial B(ZI, Z2, ... , zn) devoid of zeros in
rr after bilinearly transforming each variable Zi,
1 - Pi
Zi --+ -1--' i = 1,2, ... , n (1.22)
+Pi
and then multiplying the resulting rational function by the denominator polyno-
mial in Pl,P2, ... ,Pn. However, for a strictest sense Hurwitz polynomial to be
'bilinearly transformable' to a polynomial devoid of zeros in un
requires some ad-
ditional restrictions. These restrictions have been treated by Bose [5] with pertinent
references. The counterparts of tests for absence of multivariate polynomial zeros
in compact polydomains like the unit polydisc may not exist when the polydomain
is unbounded as illustrated next.
Example 1.2 Consider the following polynomial [5]
Ql (PI, P2) = 2p?p~ + p? + p~ + 2PIP2 + 2Pl + 2P2 + 2

=1 6 -1 -1
Clearly, Ql(Pl, 0), Ql(0,P2) are each nonzero in D = D U {oo}, where D is
the closed right-half plane. Note that at P2 = 0, or PI = 0, Ql(Pl,P2) does not
experience degree reduction. Also,
Ql(jWl,jW2) = 2w~w~ - w~ - 2WIW2 - w~ + 2 + j2(WI + W2)

ForQl (jWl,jw2) to be zero, itis necessary that WI = -W2, under which constraint
the real part OfQl (jWl,jw2) in the last equation is 2(wt + 1), which is nonzero for
° -2 -2 -1
any real WI. Therefore Ql(jWl,jw2) I- in R ,where R = R x R ,and R is
-1 -1
obtained by explicitly adjoining {oo} to the real line lit However, Ql (PI, P2) has
=2
zeros in D ,e.g., when PI = j,P2 = (1 + j + (1 + 4j)1/2).
In this section, it is pointed out that microwave circuits which can be modeled
by lumped network elements and commensurate or non-commensurate (at least,
for theoretical reasons) transmission lines can be analyzed and synthesized via
transform techniques utilizing rational functions in several complex variables. In
effect, an univariate transcendental function characterization is transformed into
a multivariate rational function description of the system. This idea permeates
an approach to analysis for delay-differential systems, to be described in the next
section. A valid question in the stability analysis of lumped-distributed networks
has been posed as a problem [5].
PROBLEM. Prove or disprove the following statement: The (n + 1)-variate polyno-

mial G (pI, P2, ... ,Pn+1) has no zeros in the polydomain Re Pi > 0, i = 1, ... , n+
°
1 (Le. it is widest sense Hurwitz) if and only if the function G(pI, tanh aIPI, ... ,
tanh anPt} has no zeros in Re PI > for all ai > 0, i = 1,2, ... , n. It may
be assumed that G (PI, P2, ... ,Pn+ I) is devoid of polynomial factors of the form
I1~~i(pk - 1)mk or more specifically, albeit restrictively, G(PI, 1, 1, ... , 1) ¢ 0.
A proof of the statement in the preceding problem has been given by Delsarte,
Genin, Kamp [18], [19]. It is noted that the stability criterion under discussion
relates the zero exclusion of a multivariate polynomial from a specified polydo-
main to the stability problem associated with passive lumped-distributed networks,
independent of delay. For any specified fixed delay, the assessment of stability be-
comes a more difficult problem, and it is not clear whether multivariate techniques
will offer any advantage over those techniques that are currently available to handle
zero exclusion problems for classes of univariate transcendental functions.
2.3 DELAY-DIFFERENTIAL SYSTEMS
Given a delay differential system of the retarded type with delays equal to integer
multiples of a fixed delay h ~ 0, it is well-known that the system is asymptotically
stable independent of delay if and only if the characteristic function, Q(P, e- hp ),
which may be viewed as a polynomial in P and e- hp , satisfies the condition,
(1.23)
Kamen [20] showed that after introducing an additional complex variable z to

replace e- hp , the bivariate polynomial,
QI (p, z) ~ Q(p, e-hp)le-hp-tz
displays the following property.
FACT 1.2 QI (p, z) =1= 0, Re P ~ 0, Izl = 1 if and only if

(a) Equation(l.23) holds for all real h ~ °and
(b) Q(Po, e- hPo ) =1= O,Po = jw/h, wE [0,211"], lim h -+ 00
For delay-differential equations of a more general type (including neutral type)

with commensurate and incommensurate delays, some complications may arise.
Consider a delay-differential equation of the generic type,
(1.24)
where akr and hr (k = 0, 1, ... , m 1; r = 0, 1, ... , m2) are constants; furthennore,
o = ho < hI < ... < hm2 (1.25a)
and
akm2 i- 0 for some k, amlO i- 0,

a m1r i- 0 for some ri-O (1.25b)
Brumley [21] showed that even if every zero, PI = p~O) of the characteristic
equation,
ml m2
Q(p,e-P1 ) = LLakrp~e-hrPI (1.26)

k=Or=O
satisfies Re p~O) < 0 and the hr's are commensurable, it is possible for (1.24) to
have unbounded solutions unless an additional polynomial condition is satisfied.
The quasipolynomial in (1.26) is an entire analytic function in the complex variable
PI, possessing an infinite number of zeros, the only limit point of which is infinity.
Even with the spectra of (1.26) in the left half plane, it is possible that in some
isolated cases due to the presence of roots which approach the imaginary axis at
infinity, one may be able to construct a solution of (1.24), subject to appropriate
initial conditions, which increases without limit for some sequence of values of the
argument. Gromova [22] called such a distribution of roots of (1.26), for which
there exists sequence of roots which approach the imaginary axis at infinity, the
asymptotically critical case. When the hr's in (1.25a) are incommensurable, the
technique of Brumley in [21] does not apply.
Setting aside the pathological or isolated cases referred to above, the situations
when the asymptotic stability of neutral differential equations is equivalent to their
spectra lying in the left-half plane are of very great interest. Guiver and Bose [23]
have considered several equivalences in the bivariate polynomial fonnulation of the
test for asymptotic stability independent of delay for delay-differential equations
of the neutral type.
2.4 STIFF DIFFERENTIAL SYSTEMS
Genin [24] associated a bivariate canonical polynomial to a multistep - mul-

tiderivative fonnula used for integrating stiff differential equations. Specifically,
consider a first order differential equation,
dy
Dy = dt = h(y,t) (1.27)
An approximate solution x of (1.27) can be obtained through the use of the fol-
lowing linear multistep-multiderivative (LMSD) formula
n k
L~:)-1)iaijhiDiXt_j = 0, t = k,k + 1, ... (1.28)
i=O j=O
initialized by a set of k starting values (Xo, Xl, ... , xk-d for fixed integers n
and k, real constants aij and DOxt_j = Xt-j is the computed value of y at time
tt-j, DiXt_j 6 fi{Xt-j, tt-j) with Diy = fi{y, t) == (8!i-t/8y)ft +(8!i-t/8t)
and h is the step size. The associated bivariate canonical polynomial is:
n
H(Pl,P2) = Lbk{pdp~
k=O
where
bi(Pd = (Pl - 1) ?: aij (PlPl +_ 11)k-

k k
3=0
j
The (k, n) method based on (1.28) is weakly stable provided the polynomial
bo(Pd = H(Pl,O) has exact degree k - 1, does not vanish in Re Pl > 0 and
has at most zeros of multiplicity 1 on Re Pl = O. When bo (Pl) =1= 0 on Re Pl = 0,
the (k, n) method becomes strongly stable. A weakly stable (k, n) integration for-
mula is said to have the error order v and the corresponding error constant Kv+l
if its associated canonical polynomial satisfies
1 (Pi,
Ie H - +
log~1 1) f"V Kv+1 (-2)£1+ 1
for Pi -+ 00
Pi Pi - Pl
The above remains satisfied if single variable factors, when present, of H(Pl,P2)
are factored out. Under these conditions, Genin [24] showed that the integration
formula (1.28) is A -stable if and only if H (Pl , P2) is a bivariate Hurwitz polynomial
in the narrow sense. According to the Daniel-Moore conjecture, the maximum
error order achievable by an n-derivative formula equals 2n and the maximum
value of the corresponding error constant is
(2n)! {2n + 1)1

Recently, Delsarte-Genin-Kamp [25] proved the validity of the Daniel-Moore con-
jecture when n =2 (others proved the validity in general via different approaches)
using certain interesting properties of the bivariate canonical polynomial associated
with a (k, 2) integration formula. These properties are expressible in terms of the
positive realness of certain univariate rational functions. The property of positive
realness is known to form the nucleus of the topic of passive network synthesis.
The possibility for relating bivariate positive functions to stability questions for
numerical methods has been raised and remains largely unexplored except for a
brief mention of Dahlquist [26] in reference to a study of A -stability 'in an implicit
alternating direction scheme for hyperbolic and parabolic equations' .
2.5 MULTIPASS PROCESSES
A large variety of operations in the coal cutting, metal shaping, and automatic
agricultural ploughing can be modeled as multipass processes, which are charac-
terized by repetitive cycles of operations, each of finite duration, and interaction
between the state and/or output variables generated during successive cycles of
operation. Each individual cycle of operation, is referred to as a pass. Successful
control systems performing such operations demand modeling in a 2-D framework
where the two independent directions of information propagation are along the trial
(pass) and from trial-to-trial. Edward and Owens [27] have noted that the notion of
stability along each pass coincides with the notion of BmO stability in the theory
of 2-D systems. Consequently, the well developed stability tests in the area of 2-D
digital filters can be directly applied to the stability analysis of multipass processes
that may be characterized by linear shift-invariant discrete state-space models.
The status of stability theory for multidimensional systems has been reported.
The broad scope of applications of fundamental tests is highlighted. The occur-
rence and utilization of definitions on various forms of stability are underscored.
During the early years, when LSI multidimensional systems were the prime tar-
gets of research, transform techniques were primarily used. With shift of attention
more and more to spatially-varying and nonlinear multidimensional systems, the
need for new tools and approaches to stability analysis was felt.
3. Multivariate Realization Theory
In response to the scopes for applications of multidimensional systems theory,

there has been a substantial advance in the status of realization theory. Generally
speaking, the approaches adopted in the development of such a theory may be
broadly classified under either state-space of transform domain (or polynomial
matrix) categories. The state-space approach towards realization has been more
popular in estimation, identification and control-theoretic applications, while the
techniques based on transform methods occur more within the settings of network
theory, digital filtering, and related fields of study. However, any attempt to restrict
a particular approach to selected areas of application may be unwise and futile.
3.1 STATE-SPACE REALIZATION THEORY
Linear dynamical systems, discrete or continuous, characterized respectively by
x(t) = Fx(t) + Gw(t), y(t) = Hx(t) + Jw(t) (1.29a)

or
x[k + 1] = Fx[k] + Gw[k], y[k] = Hx[k] + Jw[k] (1. 29b)
where F, G, H, J are constant matrices and x, w, yare the state, input and output
vectors have been widely studied for over two decades. Since the presence or
absence of J in (1.29a) or (1.29b) does not complicate in any way, the treatment of
the realization problem from the mathematical standpoint, for the sake of brevity
in exposition, it will, henceforth be considered absent. Then, in the discrete case,
the realization problem involves the determination of {F, G, H}, if they exist, such
that for a specified impulse response sequence, {TI' T2, ... }, the following holds.
k = 1,2, ... (1.30)
When the elements of Tk and likewise of F, G, H are restricted to belong to any

specified field, finite or infinite, procedures to obtain a realization {F, G, H}, the
notions of minimality, reachability, observability and equivalence or isomorphism
between two realizations have been thoroughly documented. The need for analysis
and synthesis of a wider class of systems than those characterized by (1.29a) or
(1.29b) necessitated the initiation of research into the realization theory of systems
over rings. The first question that occurs when lifting the restrictions imposed
by a field, is under what circumstances the finiteness criterion imposed on the
sequence {TI, T2, ... } (for realization to be possible, the infinite block-Hankel
matrix generated with TI , T 2 , ••• in its first block row must have a finite rank)
continues to hold. Via imposition of a very mild restriction on a commutative
ring R, Rouchaleau, Wyman, and Kalman [28], were able to prove the following
result based on the definition that a ring R is Noetherian if every ideal is finitely
generated. R is assumed to be an integral domain permitting the consideration of
its quotient field K.
FACT 1.3 Let R be a Noetherian integral domain, Kits quotientjield, T I , T 2 , •••

an input/output sequence over R. This has a realization over R if and only if it is
realizable over K.
The next step in classical linear system theory relates the notions of canonical
(i.e. both reachable and observable) and minimal realizations under the imposition
of equivalence for the dimensions of realizations; this dimension equals the rank of
the infinite block Hankel matrix formed from TI , T2 , ••• as referred to previously.
Rouchaleau and Sontag [29] defined a realization of an input/output sequence over
R to be absolutely minimal if and only if its dimension is the same as that of a
minimal (canonical in this case) realization of the specified input/output sequence

over the quotient field K and proved the following general result.
FACT 1.4 The canonical realization ofevery input/output sequence over a Noethe-
rian domain R is absolutely minimal if and only if R is a principal ideal domain,
i.e. R is Noetherian and every pair of elements rl, r2 in R hru' a greatest common
divisor which can be expressed as a linear combination, r3rl + r4r2 where r3, r4
also belong to R.
The existence of absolutely minimal realizations which, however, may not be

canonical can be guaranteed over rings more general than principal ideal domains.
In multidimensional systems theory, these general rings include polynomials in
not more than two indeterminates, having coefficients in a field. Sontag [30] also
gave a lattice characterization for the class of minimal-rank realizations over a
commutative Noetherian integral domain.
It has been noticed that unlike in the case of systems over fields, where a system
is canonical (reachable and observable) if and only if it is minimal, for systems over
rings a reachable and observable system is minimal but not, necessarily, vice-versa.
Sontag [31] studied the observability properties of realizations of linear response
maps over commutative rings and gave a characterization for those maps which
admit realizations which are simultaneously reachable and observable in a strong
sense. To do this (and also to tackle the problem of regulation of linear systems
over commutative rings), he introduced the concept of split map described below
for the case when R is a Noetherian integral domain, as this restriction is usually
met by rings encountered in systems theory and the description also provided an
algebraic criterion to test for the split property.
FACT 1.5 Let R be a Noetherian integral domain and K be its quotient field. Let
{Tl' T 2 , ••• } be an input/output matrix sequence over R and let n be the dimension
of a canonical realization of this sequence over K. Let Hn be the block Hankel
matrix whose first block row is T1 , T2 , ••• , Tn and the last block column is formed
from Tn, Tn+l,"" T 2n - 1 arranged sequentially from top to bottom. Then the
input/output map is split if and only if the ideal generated by all (n x n) minors of
Hn is R (or in other words the greatest common divisors of these minors is a unit,
when R is a Bezout domain).
A survey of early developments in the state-space realization theory over com-

mutative rings outlined above is given in [32]. In multidimensional systems theory
Eising [33] viewed two-dimensional causal and weakly causal digital filter transfer
matrices as linear systems over the rings of proper rational functions and stable
proper rational functions, as a prelude to the development of a state-space realiza-
tion model, for which the state can be recursively computed so long as the support
of the impulse response sequence belongs to a causality cone [Chapter 2]. Other
state-space models encountered in the realization of 2-D systems have been dis-
cussed by Bose [5]. State-space modeling of 3-D systems has been considered by
Tzafestas and Pimenides [34], [35].
It is relevant to point out that some research into the realization theory over
a noncommutative ring has also been conducted. The distinguishing fact here is
that the Cayley-Hamilton theorem which provides the finiteness condition in the
realization theory over a field or a commutative ring, fails to hold. Fliess [36]
began the study on realization of rational power series in.several non-commuting
variables, after recognizing their relevance in the analysis of bilinear systems and
possibly, in a larger class of nonlinear systems. Sontag [37] obtained some purely
algebraic extensions of results in the theory of linear dynamical systems for the
case when the coefficient ring is arbitrary. Fomasini [38] studied the possibility of
using rational noncommutative power series to realize spatial filters. He presented
an extension of Ho's algorithm in classical linear system theory, tackled the partial
representation problem in which one is attempting to obtain a recursive model for
the coefficients of a series on the basis of incomplete data, and finally considered
the problem of generating all minimal realizations of a specified filter.
3.2 TRANSFORM DOMAIN REALIZATION THEORY
The polynomial matrix approach to linear systems, initiated by Rosenbrock, is

very useful in the study of realization theory and related problems of dynamic
compensation, stabilization, output regulation in the presence of disturbances,
input tracking, etc. It is well known that for a realizable input-output map over
a field, the transfer matrix T(P), whose elements are rational functions in the
complex variable p (l-D case) over an arbitrary but fixed field of coefficients, is
factorable as (matrix fraction description)
T = AB- I = D-IC (1.31)
where A, B, C, D are polynomial matrices over the ground field and A, B are right
coprime while C, D are left coprime. The coprimeness conditions are equivalent
to the requirement of existence of polynomial matrices X, Y, T, S with coefficients
over the base field such that the respective conditions given below are satisfied. I
denotes an identity matrix of appropriate order.
XA+YB=I (1.32a)
CT+DS=I (1.32b)
It was shown by Guiver and Bose [39] (see Chapter 5, Section 3) that matrices
whose entries are rational functions in PI, P2 (2-D case) over a field of coefficients
are factorable in a form similar to (1.31) with polynomial matrices A, B devoid
of any nontrivial common right factor and polynomial matrices C, D devoid of
any nontrivial common left factor. More importantly, the computations to extract
a greatest common right divisor or a greatest common left divisor from two poly-
nomial matrices having entries that are polynomials in Pl,P2, need be performed
in the specified ground field (and not in any extension field) containing the co-
efficients of the polynomial matrices (or the rational matrix from which the two
initial non coprime polynomial matrices are derived). The primitive factorization
algorithm which is central to the procedure under discussion cannot be extended
to cases involving more than two indeterminates. Youla and Gnavi [40] delin-
eated the various types of coprimeness (factor, minor, and zero) that are natural
in multidimensional systems theory. These three types of coprimeness are equiv-
alent in the 1-D case and are mutually distinct in the n-D case (n > 2). When
n = 2, factor and minor coprimeness, are, interestingly, equivalent concepts and
this fact justifies the feasibility of primitive factorization in the 2-D matrix case.
Of course, the primitive factorization algorithm holds in the n-D case for any n,
when attention is restricted to n-variate polynomials instead of polynomial ma-
trices. Zero primeness imposes the most stringent restrictions on the (m x m)
minors of a m X k, m ~ k, polynomial matrix A(p} ~ A(Pl, P2, ... , Pn}.A(p} is
called zero-prime provided all its (m x m) minors are devoid of common zeros,
while it is minor-prime provided all its (m x m) minors are devoid of common
factors. Zero-primeness of A(p} implies its unimodularity (and vice-versa); that
is, zero primeness of A(p} is equivalent to the existence of a polynomial matrix,
B(p} such that A(p}B(p} is the identity matrix of order m. Youla and Pickel
[41], after defining that A(p} is projectively free if it can be included at the first m
rows of some (r X r) elementary polynomial matrix showed how the Serre con-
jecture (proved independently by Quillen and Suslin as documented by Lam [42])
translates into the following fact. See also [43].
FACT 1.6 Any (m x r) zero-prime polynomial matrix A(p}, m ~ r can be row-

bordered up into a square (r x r) elementary polynomial matrix V(p}. Moreover,
if A(p} is real, i.e. if all its entries are polynomials with real coefficients, V(p}
can also be constructed real.
Youla and Pickel proved Fact 1.6 using only a minimum of modern abstract
algebra and justified the validity of the following equivalence.
Zero-prime ++ Unimodular ++ Projectively Free

For polynomial matrices over an arbitrary commutative ring, validity of (1.32a)
and (1.32b), imply, respectively, right and left zero (not, in general, factor or
minor) coprimeness of polynomial matrices under consideration. In (1.32a), the
polynomial matrices A, B are called right Bezout and in (1.32b) the polynomial
matrices C and D are called left Bezout. Recently, Khargonekar [44], studied
the existence and realization theory of matrix fraction representations in the forms
given in (1.31) for linear systems over commutative rings. He was able to bring
out the expected result that a transfer matrix T associated with an input/output map
admits a left or right Bezout matrix fraction representation if and only if the map
is split (see Fact 1.5). Split maps and zero-coprime fractional representations are,
therefore, related.
Fuhrmann [45] developed a polynomial model approach to linear dynamical
systems over a field with the objective of providing the exposition of state-space
and matrix fraction representation schemes for realization, in a unified setting.
He based his approach on representation theorems for submodules and quotient
modules of spaces of pol ynomial matrices and vectors. A correspondence between
fractional representations of the transfer matrix of a given input/output map and
its reachable or observable realizations was established. Generalization of this
approach to the case of systems over a principal ideal domain has been done by
Conte and Perdon [46], whose results are documented in a more general setting in
the work of Khargonekar, already referred to here.
Significant developments have taken place during the last decade in the con-
solidation, unification, and correlation of research on the realization theory over
rings. These results directl y apply to problems in the areas of multidimensional sig-
nal processing, delay-differential systems, realizations with parameters (Byrnes,
[47]), realizations incorporating system robustness and parameter uncertainty, and,
in general, in the domain of analysis and design of families of systems, instead
of only particular systems. Often, realizations have to be obtained under severe
constraints. This is, particularly, brought out in network theory, where the con-
straints of positive realness or restrictions on topology (like the doubly terminated
lossless multidimensional two-port synthesis problem) can lead to serious, if not
insurmountable problems. In fact, whether or not the multivariate positive realness
property is sufficient for the synthesis of arbitrary multidimensional passive mul-
tiports remains an unresolved problem (see Bose, [5]). Multidimensional wave
digital filters have a lot of attractive properties and it would be useful to provide
general synthesis schemes for reference filters (see Fettweis, [16]). The stumbling
blocks encountered in the realization of systems geared towards the types of ap-
plications referred provide impetus for further research and renewed challenge in
an arena where new developments never seem to obliterate the scope and need for
additional procedures, techniques, and refinements. Recent results which benefit
the area of multivariate network realizability theory are summarized next.
3.2.1 Integral Representation of Positive Real Functions

Multivariate rational positive real functions and matrices form the nucleus in
the study of multivariate network realizability theory.
DEFINITION 1.5 An m x m matrix Z{p) whose elements are (rational)functions

in the complex variables p d (Pi, ... ,Pn) is called (rational) positive real if
(a) Z{p) is holomorphic in Re p >0
(b) [Z*{p)]t = Z{p*) in Re p >0
(c) Z{p) + [Z*{p)]t is Hermitian nonnegative definite in Re p > 0

Condition (a) implies that Z{p) need not be BillO stable, condition (b) im-
plies the realness of coefficients in the entries of the (rational) matrix Z{p), while
condition (c) is a consequence of passivity of the network characterized by Z (p).
Positive real functions as well as their Poisson integral representations playa fun-
damental role in the theory of networks. In the n = 1 case, the following results
are due to Cauer [48] who made use of the Riesz-Herglotz representation of a
function, regular in Izl < 1, with its real part nonnegative in Izl < 1, and the
bilinear transformation, z = (PI - 1) / (PI + 1).
FACT 1.7 Afunction Z(PIJ is positive real ifand only if,
Z{pt} = 100
-00
jXPI - 1
.
JX - PI
dtt{x) + CPI
where tt{ x) is a real nondecreasing function of bounded variation on (-oo, 00)

and C is a nonnegative real constant. It is possible to select tt{x) = -tt{ -x), and
the integral is taken to be in the Stieltjes sense.
FACT 1.8 [48]: Any positive real function Z(Pt} can be represented as,
Z(Pt} = PI [c + tx) d:{X)]

10 PI +x
where C is a nonnegative constant, tt is a nondecreasing function and the inte-
gral is to be taken in the Stieltjes sense. Conversely, if the preceding integral
representation exists, it represents a positive real function.
For a representation theorem for positive real matrices in a single complex
variable, which is the matrix counterpart of the result in Fact 1.7, see [49, Appendix
1]. More importantly, the multivariate counterpart of the Riesz-Herglotz result was
given in 1963 [50].
FACT 1.9 [50]: Thefunction f : un -+ Cis holomorphic and has nonnegative

real part in un if and only if it admits a representation,
f (z) = jIm f (O) + ( ...

lTn ! IT [2
k=I
1 1
- Zkwk
* - 1] dtt (w)
with a positive measure tt on Tn such that
hn ... ! W~l •.. w~ dtt{w) =0

unless
for all k = 1, ... , m or

for all k = 1, ... , m
Note the orthogonality constraint that the positive measure J.t must satisfy on Tn
when n > 1. See also [51, pp. 80-91}.
Aizenberg and Dautov [52] extended the above result for functions holomorphic
in other poly domains besides un,but were not able to cover unbounded polydo-
mains, which occur in the context of positive real functions of matrices. For these
types of results relating to polydomains other than un,
see [53], [54], [55], and
[51] for more references.
3.2.2 Sum of Squares Representation of a Form in Network Synthesis

It has been seen (see Chapter 5 of [5]) that the infeasibility of sum of squares
representation for a six variable positive (in the projective plane) form of degree 4
provides a theoretical limitation to a synthesis procedure for multivariate positive
real matrices, which can characterize lumped-distributed and variable parameter
multiports. Interestingly, this theoretical limitation also influences the extendabil-
ity problem in multidimensional spectral estimation [56], and is linked to the fact
that strongly positive [57] (see Fact 1.10 also) or completely positive [58] linear
maps rather than positive linear maps are the natural generalization of positive lin-
ear functionals. Many examples exist to substantiate this fact and, quite recently,
Schmiidgeon [59] gave the polynomial in (1.33) below as an example of a positive
polynomial which is not a sum of squares of polynomials, and therefore it is a
positive but not a strongly positive functional.
F(Xl, X2) = 200(x~ - 4xd 2 + 200(x~ - 4X2)2+
(X2 - Xd(X2 + XdXt{Xl + 2)(Xt{Xl - 2) + 2(x~ - 4)) (1.33)
In the context of network realizability theory, the class of functions synthesizable
by Koga's procedure (see [5]) can, however, be identified from the theorem to be
stated next.
Let Pn,m be the set of all positive semidefintite forms in n variables of degree
m. Pn,m belongs to a closed convex cone in a finite dimensional Euclidean space.
Let pJ;,m denote the family of members of Pn,m in which all the intervening
exponents are bounded by a given integer r, i.e. if F(Xl, X2,···, xn) E pJ;,m
then F(Xl, X2,"" xn) involves only monomials X~lX~2 '" x~n with kl + k2 +
... + kn = m and 0 ~ kl' k2, ... , k n ~ r. Let Sn,m be a subcone of Pn,m
formed from all finite sums of squares of polynomials. Clearly, Sn,m ~ Pn,m' Let
C[Xl' X2, ... ,xn] be the algebra of all polynomials with complex coefficients in
n commuting indeterminates Xl, X2, ... , Xn . By a positive functional is meant a
linear functional on C[Xl' X2, ... ,xn] which is nonnegative on Sn,m, \In, m. The
functionals with nonnegative values on the cone Pn,m, \In, m, are strongly positive.
THEOREM 1.4 [60}: Suppose r, m are even and m ~ 4 (m = 2 case is trivial).
(1) Let n ~ 4. Then, P~,m n Pn,m ~ Sn,m, iffm ~ rn - 2

(2) Let n = 3. Then, P~,m n Pn,m ~ Sn,m, iff m = 4, or m ~ 3r - 4
Since Koga's synthesis procedure depends upon the validity of P;,m nPn,m ~
Sn,m it only holds if and only if either m ~ 2{n - 1) or m = 2.
4. n- D Problem of Moments and Its Applications in

Multidimensional Systems Theory
This brief section is included to provide further scope for appreciation of the
effects of distinction between the notions of positive and completely positive func-
tionals in multidimensional problems. The problem of moments has an extensive
literature. Shohat and Tamarkin [57] have given a concise but through documenta-
tion of the results in that area. Though the classical moment problem including its
various ramifications like the trigonometric moment problem, and the Hausdorff
moment problem is completely and satisfactorily solved in the I-D case, the n-D
(n > 1) counterpart provides some very interesting insights into the complexities
encountered when attempting extensions of the 1-D results to several dimensions
[61], [62]. The n-D moment problem is formulated as follows.
PROBLEM FORMULATION: Let there be given an infinite multiple sequence of
real constants
ii, i 2 ,···, in = 0, 1,2, ...
in an n-D Euclidean space. The objective is to find necessary and sufficient con-
ditions for a n-D distribution function F{x) to exist (a distribution function is
nonnegative, defined and finite over the family of all Borel sets in the n-D Eu-
clidean space r and is completely additive) whose spectrum (defined as the set
°
of all points x in]Rn such that F(1) > for every open set 1 containing x) is to
be contained in a closed set 10, given in advance, and which is a solution to,
(1.34)
for ii, i 2 , .•• , in = 0, 1,2, ...

For the n = 2 case let P{Xl' X2) be any polynomial in Xl, X2,
P{Xl' X2) = L Qil,Bi2Xilx~2 (1.35)
il,i2
where ail' ,Bi2 are real or complex valued constants. Introduce the functional,
(..t{P) = L (..til.i2 Q il,Bi2 (1.36a)

it,i2
where
(1.36b)
FACT 1.10 A necessary and sufficient condition that the 1o-moment problem de-
fined by the sequence ofmoments {J.Lil ia} shall have a solution is that the junctional
J.L{P) be (1o)-nonnegative i.e. J.L{P) ~ 0 whenever P{XI,X2) ~ 0 on 10• When
the junctional J.L{ P) satisfies the preceding condition it is called strongly positive.
The preceding result suggests that for any positive polynomial P{XI, X2) in
(1.35), the linear functional T defined by
is positive. T is multiplicatively positive if < T, ¢¢* > ~ 0 for every test function
¢(XI' X2). In the linear space of all real polynomials in variables Xl, X2, a mul-
tiplicatively positive functional may not be positive and therefore the condition
< T, ¢2{XI' X2) >~ 0 for all real polynomials ¢(XI' X2) is not sufficient for the
solvability of the 2-D moment problem.
In the trigonometric moment problem, 10 is the distinguished boundary Tn of
a polydisc and (1.34) is replaced by,
J.LiIia ...in = J.L(i) = ! T~ ! ei<i,z> dF(z) (1.37)
where z = (Zl,"" zn), i = (il,'" in) and < i, z >= itzl + ... + inzn.
A representation similar to the one given above occurs in probability theory [63]
and the multidimensional spectrum estimation problem [56]. It is also pointed out
that the notion of strong positivity has important consequences in the theory of
unbounded operators [64]. In the problem of multidimensional spectrum estima-
tion occurring in diverse fields of applications of the theory of multidimensional
signal processing Equation (1.38) below relates a specified finite set of correlation
samples {J.L{i) : i E Ll = (O, ±Ol,"" ±om), Oi E Z, i = 1,2, ... , m)} to a real
positive spectrum F{x), x E nn c r, to be estimated, provided it exists, such
! D~ !
that for i E Ll,
J.L{i) = F{x)ei<i,x> dx (1.38)
where nn is the domain over which F(x) is specified to be nonzero. If a solution

to (1.38) for a positive F{x) exists, then the specified correlation samples are said
to be extendible. Without specializing i to take values only on finite Ll, J.L{i) in
(1.38) may be viewed as a continuous function of i which is positive definite in
the following sense. For any points iI, i2, ... im in r and any complex numbers
6, ... , ~m one has
m m
LL lL(ik - ij)~k~; ~ 0 (1.39a)

k=lj=l
for every positive integer m, where

• • A (. • • .)
Ik - Ij = Zkl - Zj1, ... , Zkn - Zjn (1.39b)
Indeed (1.39a) follows from the use of (1.38) in the following manner
ff
k=1 r=l
lL(ik - ir) ~k C = f f ~k ex!...
k=1 r=1 Dn
... t XJ
ei<i",-ir,x> F(x) dx
i-oo
L ei<ir,x> ~r 12 F(x) dx ~ 0, since F(x) is positive.
m
1
r=1
In the I-D case, for a specified finite set of samples 1L(i), i E ~ = {O, ±81 , ... , ±
8M}, the positive definiteness of the Toeplitz matrix with [1L(0)1L(8d·· .1L(8MW
W
in the first row and [IL(O)IL( -8t} .. ·IL( -8M in the first column is necessary and
sufficient for the extendability of ({IL(i)} : i E ~). This can be substantiated
via results of Rudin. Rudin [65] considered a finite set S in a discrete group G
and related the extension problem to the sum of squares representation problem of
certain positive trigonometric polynomials. He then showed that the problem of
extending the class of all continuous complex-valued functions IL on S - S (the set
of all points ik - ir E G, with ik E Sand ir E S) which satisifes the I-D (finite S)
counterpart of (1.39a) has a solution if G = Z, the additive group of integers and
if S is a finite arithmetic progression in G. Rudin also proved that the analogous
result fails to hold in higher dimensions i.e when n > 1. In fact, he showed that not
all positive trigonometric polynomials on T2 are representable as sums of squares
of trigonometric polynomials and used this fact to demonstrate that the extension
problemmayfailtohaveasolutionifG = ZxZ ~ Z2, the group ofalllatticepoints
in the plane and if S is a square of lattice points. Subsequently, Rudin transferred
the result of infeasibility, in general, of extension from Z2 to ~2 and hence of Znto
r, n ~ 2. A Toeplitz form, like in (1.39a), on S x S, where S is a subset of, say
G = zn = Z X Z x ... x Z is a function, IL on the Cartesian product S x S with
the property that lL(i k , ij ) is a function only of i k - ij, i k E S, ij E S. Then, the
preceding discussion centering around the validity of (1.39a) as a consequence of
the positive definiteness property of forms induced by positive measures via the
representation given in (1.37) or (1.38), together with Rudin's results, lead to the
following important fact.
FACT 1.11 [66J: Every Toeplitzform induced by a positive measure is positive

definite but the converse is false if n > 1.
Fact 1.11 imposes serious limitations on the multidimensional spectral estima-
tion problem, where a procedure for verifying whether or not a finite multidimen-
sional Toeplitz form is the restriction of a form induced by a positive measure
is required. One procedure to attain this objective has been reported in [66, pp.
362-363].
5. Role of Irreducible Polynomials in Multidimensional

Systems Theory
The role of irreducible polynomials has been noted, from the very beginning, in
[5] and these types of polynomials playa crucial role in the problem of unambigu-
ous determination of a function (object) from the modulus of its Fourier transform
(energy spectra) or equivalently from its autocorrelation function. This problem is
important at high frequency applications like optics where the only experimentally
measurable quantity is the scattered-intensity distribution. In the discrete case,
i.e. when the brightness distribution is specified in a rectangular array, the image
reconstruction problem reduces to the determination of a polynomial (uniquely, if
possible) p(z) £ P(Zl' Z2, ... , zn) from a specified knowledge of the product (in
(1 .40) Z -1 =
l> (-1 -1 -1) .
Zl ,Z2 , ... , Zn
A(Z, Z -1) = P (Zl, Z2,···, Zn )P (-1

Zl ,Z2-1 , ... , Zn-1) (l.40)
subject to the physical constraint that the coefficients (representing the brightness
distribution) of p( Zl, Z2, ... , zn) are nonnegative. The question of uniqueness of
p(Zl' Z2, ... , zn) is related to its irreducibility.
The coefficients of A(z) form the sequence of discrete autocorrelation function
of the unknown image required to be restored. In the 1-D case, the brightness
distribution cannot, in general, be reconstructed unambiguously. Uniqueness of
reconstruction up to trivial factors (like multiplication by a positive real constant,
and turning of the brightness distribution by 1800 ) is possible when the polynomial
p(z) turns out to be irreducible [67]. Carlitz [68] gave an asymptotic formula for
the number of irreducible polynomials in several variables of total degree m and
arrived at the conclusion that almost all multivariate polynomials are irreducible
(for this, see also [69]). For enumeration of irreducible multivariate polynomials
whose coefficients are in the finite field Zq, q prime, see [5, Chapter1]. These
results justify the reduced ambiguity of the image reconstruction problem in the
discrete multidimensional case. Manolitsakis [70] considered the reconstruction
problem of two-dimensional continuous objects from their energy spectra. He
showed that the properties of two-dimensional fields differ inherently from those
of one-dimensional fields (especially because the zero-sets are not isolated) and
that they lead to a reduced ambiguity for object reconstruction from intensity data.
Irreducible polynomials with coefficients over finite field Zq, where q is a prime
integer, are known to play useful roles in the construction of arrays with special
properties, which are generalizations of some of the properties of pseudorandom
sequences (see [5, Chapter 6]). In the I-D case particular types of irreducible
polynomials called primitve polynomials are used to design linear feedback shift
registers for generating pseudorandom sequences, which have the two-level dis-
crete autocorrelation property when q = 2. Binary arrays with a maximum of three
distinct levels of autocorrelation have been constructed in [5] and explicit expres-
sions for the values of these levels have been given. In Zq, q prime, arrays with
not more than q distinct levels of autocorrelation are discussed in [71], where how-
ever, a distinction between the notions of period and the maximum area property
of linear arrays from their counterparts in the case of pseudorandom sequences are
drawn. A natural generalization of pseudorandom (maximum period) sequences
to maximum period arrays is given in [72], where, however, the concept of period
is different from that in [5] an [71]. The delineation of properties of arrays studied
in [72] have been motivated by applications in 2-D cyclic code construction and
such constructive schemes have been subsequently considered in [73].
6. Hilbert Transform and Spectral Factorization

The problem of determining the phase of certain function analytic in one half
of the complex plane from the knowledge of its magnitude on the axis bounding
the half planes has been of interest in circuit theory, communication theory (in
the context of modulation and demodulation of signals) and in various branches
of physics including the theory of decay of elementary particles, X-ray crystal-
lography, coherence theory, and light scattering. The theory of analytic functions
via the Hilbert transform pair (dispersion relations) technique provides a relation-
ship between the real and imaginary parts or the magnitude and phase of certain
analytic functions (for the magnitude-phase relationships to hold the analytic func-
tion should also have no zeros in the polydomain of interest, i.e. the function is
minimum phase). It has been pointed out in [5] that problems occur in trying to
extend the I-D Hilbert transform relationships to the multidimensional setting. In
[74], this fact is substantiated by demonstrating that the 2-D dispersion relations
constitute a pair of inhomogeneous singular integral equations with no solution.
Bose [5] discussed the pitfalls in applying the Hilbert transform to stabilize
unstable continuous filters. Given an amplitude-squared function,
it may not be possible to guarantee (spectral) factorization of the type
F 2(jWl,j W2) I w? = _P~ = H(Pl,P2)H(-Pl, -P2) (1.41)

w~ = -p~
with the restriction that the real function H(pI, 112), if it exists, has no zeroes of its
numerator or denominator polynomials in Re PI > 0, Re P2 > 0 simultaneously.
Example 1.3 [5] Consider the function,
F 2(jWI,jW2) = (WI - W2)2 +1 (1.42)
In fact, H(pI,P2) = PI - P2 + 1 or P2 - PI + 1 associated with (1.42) for the

factorization in (1.41) has zeros in Re PI > 0, Re 112 > 0 as well as in Re PI < 0
and Re P2 < O. For an example where H(pI,P2) in (1.41) does not even exist,
consider
Another limitation in the applications of multidimensional methods occur in the

inability to factor a positive (on the multi-axis or poly-boundary) function in the
form shown in (l.40) subject to the constraint that the zero-set of the finite order
polynomial factor be constrained in a half-plane. In the discrete case, the zero-set
of a spectral factor should be excluded from a unit polydisc. Note that zero sets of
multivariate polynomials form continuous algebraic curves which extend through-
out the space, i.e. cannot be constrained to belong to bounded polydomains. Helson
and Lowdenslager [75], however, defined half-plane spectral factors, which satisfy
stability constraints, but, in general have infinite support. Due to greater versatility
of a half-plane filter frequency response over that of a quarter-plane filter (fortu-
nately the theoretical foundations of half-plane filter design are well established
in contrast to the quarter-plane case), half-plane filters are being increasingly used
in multidimensional filtering, prediction [76] and spectrum estimation problems
[77], [56].
7. Conclusions
The progress and directions of research in several areas of multidimensional
systems theory since about 1981 has been reported. In the process, some open
problems have been brought to the attention of the reader. The breadth of applica-
tions of the theoretical results discussed is, indeed, impressive, and we list below
some additional specific areas of applications that have not, yet, been cited in this
chapter. First, in addition to the flurry of activities associated with spatio-temporal
signal processing [78], optical and electronic feedback systems (these systems in-
corporating spatial and temporal variables are called hybrid systems) have been
proposed for various purposes like iterative image processing and image restora-
tion. Chapter 2 by Guiver and Bose considers the problem of stabilization of 2-D
systems by causal and weakly causal compensators and contains references to the
recent literature on spatio-temporal feedback systems. Stability conditions for 1-D
causal linear shift-variant systems have been shown to be similar in content to the
well established theorems on the BmO stability of 2-D LSI systems [79]. Third,
stability results for n-dimensional linear shift invariant digital filters also become
applicable in stability investigations of externally bilinear systems (which form a
subclass of nonlinear systems) whose output response is a bilinear function of two
inputs, each applied to a different input terminal. Kamen has shown that a n-D
shift invariant externally bilinear system is BmO stable if a particular member of
the associated class of linear shift invariant (LSI) 2n-D system is BmO stable, but
not vice-versa [SO].
In Section 3, the thrust towards research into the realization theory of linear
systems over commutative rings was evident. This was to a large extent motivated
by application of the ring structure in systems characterizable by delay-differential
equations when those are interpretable as differential equations whose coefficients
belong to a polynomial ring in one (in case of commensurate delays) or several
variables (in case of incommensurate delays) [81], in image processing or 2-D
digital filtering systems (see Chapters 2 and 6), and in the global study of families
of linear systems initiated in [82]. Besides questions related to realization, other
questions can also be meaningfully posed about linear systems over rings. The
answers to these questions, however, are usually difficult to obtain and often the
use of new tools and methods of approach are required in comparison with those
used to tackle analogous problems for classical linear systems over fields. For
results, conjectures, and open problems on 'pole-shifting' over rings, see [83]. An
excellent survey of various recent results on linear systems over commutative rings
was prepared by Sontag [84].
Multidimensional linear passive systems are frequently encountered in math-
ematical physics and electrical engineering. The nucleus to the study of such
systems is provided by the concept of positive realness. Multivariate positive real
matrices occurring in network theory characterize only a subclass of passive sys-
tems. In a more general context passivity of an operator can be defined relative to an
acute, closed convex solid cone in ]RR with vertex at zero. Then, a large number of
physical processes describable by partial differential equations are passive systems
relative to certain cones. A general theory of shift-invariant multidimensional sys-
tems based on the theory of multivariate positive real matrices has been presented
by Vladimorov [85]. Aside form electrical network theory, I-D linear passive
system theory is known to be useful in the descriptions of thermodynamic sys-
tems and the scattering theory of electromagnetic waves and elementary particles.
Vladimorov gave an exposition of the considerably increasing scopes for applica-
tions of multidimensional linear passive systems theory and noted the relevance of
the fundamental concept of positive realness (with respect to an appropriate cone)
in equations of the theory of elasticity, magnetohydrodynamics, the equations of
rotating fluid and acoustics, Dirac's equation in quantum physics and the transfer
equation. Subsequently, Drozzinov [86] investigated a class of first order partial
differential equations with infinitely differentiable coefficients and identified the
subclass of systems that become passive relative to a cone. In [87], Drozzinov also
proved some results on the quasi-asymptotic behavior of fundamental solutions

of passive shift-invariant systems after developing a multidimensional Tauberian
theorem for holomorphic functions of bounded arguments. Note that the argu-
ment of any positive real function is continuous and bounded in the polydomain
of holomorphy.
In Section 4, the multidimensional moment problem has been considered. Be-
sides the applications already cited, the n-D moment problem, when n = 2, is
useful in image reconstruction [88], where given some finite set of moments,
M(j, k) = ff f(x, y)a;iyk dx dy
of an image plane irradiance distribution, f(x, y), the object is to determine how
well the image can be reconstructed. In Sections 5 and 6, respectively, the roles of
irreducible polynomials in multidimensional systems theory and the consequence
of problems in extending I-D Hilbert transform and spectral factorization results
to the multidimensional setting are briefly summarized.
There are other mathematical limitations in the extension of I-D results to the
n-D (n > 1) case which influence applications of the theory of multidimensional
systems. For example, it is known that in the unit disc in the complex plane, every
function f in the Nevanlinna class N has a factorization, f = I x E, where I is
inner and E is outer. However, there exist bounded analytic functions on a polydisc
(or for that matter on a ball), that have no inner-outer factorization [7]. Rubel
[89] introduced the notions of internal and external analytic functions of several
complex variables and proved that on any simply-connected complex analytic
manifold of any finite dimension every f in the RP-Nevanlinna class, RP : N,
(the class RP : N is defined as the usual Nevanlinna class N, except that one uses
pluriharmonic majorants rather than n-harmonic majorants), has a factorization,
f = I x E, where I is internal and E is external. The notion of weakly inner-
strongly outer factorization has been used in [90] to study the asymptotic behavior
of planar least-squares inverse polynomials occurring in studies of stabilization of
2-D digital filters.
In order to satisfy constraints of size this chapter, and for that matter, this book
is concerned with tools relevant to the study of deterministic multidimensional
systems. Random fields or distributed disordered systems, characterized by ran-
dom variation over space and time are not considered and the reader is referred
to [91] for information on two-dimensional and multidimensional local average
processes.
8. Updates
Two papers, one published in 1990 [92] and the other in 2001 [93], in the area
of multidimensional signal processing and systems, which review the problems
and progress that followed the publication of the first edition in 1985 may be
helpful to the readers. A new approach to the stability of multivariate polynomi-
als, with respect to a unit polydisc, was advanced in [94] as a special case of an
algebraic characterization of the exact multidimensional unwrapped phase. The
proposed phase unwrapping algorithm was also applied to the classical problem
concerned with the zero distribution, with respect to the unit circle, for an arbi-
trary complex coefficient univariate polynomial without encountering the plethora
of singular cases. By applying the theory of Cauchy indices, a symbolic algebra
based analytic expression was also provided for the unwrapped phase (and, conse-
quently, zero distribution with respect to the imaginary axis) associated with any
complex coefficient characteristic polynomial of a continuous-time system [95].
The proposed algorithms in [94] and [95] do not require any zero-finding and,
very importantly, force the singular case problem in all division algorithm based
procedures, to be absent.
The distinguished boundary (referred to also as the torus) Tn of the open unit
polydisc un is a compact Abelian group under the operation of the component-
wise multiplication and as such carries a Lebesgue measure mn(Tn) = (21r)n.
The polydisc algebra A(Un ) is the class of all continuous complex functions on
the closure un of un whose restriction to un is holomorphic. Several BillO
stability conditions in multidimensional signal processing emerge as special cases
of a theorem due to Rudin for functions belonging to the class A(Un ) [7, Theorem
4.7.2, p.87]. This theorem essentially states that every value assumed on un by
a function belonging to A(Un ) is already assumed on a small subset K U Tn
of un, for K = <p(U) where <P = (<PI,"" <Pn) is a continuous map of U into
un, which carries T into Tn such that for a loop E in T the winding number
Ind( <pjoE) > 0 for j = 1,2, ... , n. Further simplifications on the subset K,
which have appeared in the signal processing literature, also follow directly by
choosing </>j{>') = >.,1 :-::: j :-::: n, or indirectly from the result that if a function
f E A(Un ) has no zeros on Tn, /j(>') = f (1, ... , 1, >., 1, ... , 1) (>. in jth place)
and kj is the number of zeros of /j in U, then f(>', ... , >.) has kl + ... + kn zeros
inU.
The zero set of a multivariate polynomial is unbounded and, often it is nec-
essary to determine whether or not a specified polydomain (or its complement)
is zero-free for a polynomial. In [96], the zero set of a multivariate polynomial
is enclosed by unions and intersections of unbounded sets. Two additions to the
extensive literature on implementation of 2-D stability tests are available in [97]
and [98]. A member of the class A(Un ) qualifies as the transfer function of a
normal (first quadrant BillO-stable linear shift-invariant) filter if it has finite norm
(defined as equivalent to the filter being BillO-stable). The work of Dautov ([9])
suggested the conjecture that all rational functions in A(Un ), including those that
have nonessential singularities of the second kind (NSSK) on Tn, have finite norm.
Youla [99] advanced a system-theory motivated proof of Rudin's result that every
element belonging to the special class of all-pass functions in A{Un ) is rational,
devoid of NSSK on Tn and of finite norm. In a recent paper [l00],it was proved
that a rational multivariate first quadrant quarter-plane digital filter transfer func-
tion, analytic on the open polydisc, is BmO-stable if and only if it has a uniform
extension to the distinguished boundary of the polydisc. See solution regarding
Conjecture 1 of Open Problem 1 in Chapter 8.
Based on an abstract model in a Banach space setting, a stability theory for
linear repetitive processes has been presented in [101]. The synthesis, however, is
restricted to the use of simple types of pre-selected controllers so that parameters
involved in the meeting of the closed-loop system requirements can be determined
easily. More general design techniques without a priori restrictions on the con-
troller structure and capable of incorporating general boundary conditions, uncer-
tainty in the defining models and a type of singularity along the pass dynamics are
being considered [102]. Iterative learning control (ILC) increases the performance
of feedback control systems, purely in software, simply by iteratively adjusting the
command. In ILC the main idea is to use information from previous repetitions to
generate new control actions that improve the tracking accuracy as the number of
repetitions increases. Robots with high tracking accuracy, when performing a high
speed maneuver, are now being delivered with built-in learning control. Repetitive
control does similar kind of learning for systems with periodic commands, for
example, to cancel vibrations.
ILC systems can be described by 2-D models and the asymptotic tracking prob-
lem for such systems was considered in [103]. A 2-D repetitive process structure
also occurs in the analysis of the local convergence and stability properties of itera-
tive algorithms for solving nonlinear optimal control problems. In an algorithm of
this type, a trial solution is updated from one iteration to the next. The time hori-
zon of the dynamic system under investigation comprises one dimension while the
other dimension is the progress of iterations [104]. Publications in the subject of
multipass and repetitive processes has been quite extensive during the last couple
of decades and the reader is referred to the survey chapter [105].
The monograph [106] has been devoted to state-space realization theory of linear
multidimensional systems while the other recent monograph [107] emphasizes the
polynomial approach via the transform domain theory. An attempt at developing a
general framework with the objective of covering variable-parameter and nonlinear
multidimensional systems was started [108], which has been somewhat eclipsed
by the flurry of research in the behavioral approach [109] to multidimensional
systems that followed the seminal paper by Ulrich Oberst in 1990 [110] for the
multidimensional continuous as well as the discrete cases. For the latter case, see
also [111]. A brief description of behavioral systems has been included in Chapter
5, Section 9.
The infeasibility, in general, of the representation of a positive multivariate
polynomial in real variables as a sum of squares of polynomials was seen to be
linked to limitations in general multivariate passive synthesis and the extendabil-
ity problem in multidimensional spectrum estimation. The feasibility of fac-
toring a symmetric nonnegative definite multivariate polynomial matrix A(x),

where x £ [Xl X2 ••• Xn]t, Xl E lR for i = 1,2, ... , n, in Cholesky form as
A(x) = H(x)tH(x), where the elements of H(x) are rational functions follows
from the results in [112]. For results on nonnegativity constrained spectral factor-
ization, described in Section 5, see [113], [114].
One of the recent contributions in system theory, particularly control, has been
the development of a computational framework based on the sum of squares (SOS)
representation of a positive multivariate polynomial. According to a result of
Landau [115], a positive bivariate polynomial in real variables can be expressed as a
sum of squares (SOS) of at most four polynomials in one variable, the coefficients of
which are rational functions in the other variable. A constructive proof of Landau's
result is given in [116]. A rational function f E R(XI, ... , xn) is called positive
semidefinite if f is nonnegative wherever it is defined. A n-variate semidefinite
rational function f E R(XI,··· , xn) can be expressed as a sum of 2n squares of
rational functions [117, p. 302] (Hilbert had done this for n = 2 in 1893). In
this multivariate case, Cassels showed in 1964 that a SOS of rational functions
can be reduced to a SOS of rational functions in which anyone variable does not
occur in the denominator; furthermore, this is possible to do without increasing the
number of summands. For a comprehensive discussion of representation results
of positive polynomials as a SOS of rational functions with bounds on the number
of terms in the summand and a constructive proof of the affirmative answer to
the question, "Is there a continuous solution to Hilbert's 1rth Problem" (whether
every rational positive definite form may be represented as a quotient of SOS of
forms or, alternatively, as a SOS of quotients because it is possible to convert
between a quotient of SOS and a SOS of quotients proved nonconstructively by
Emil Artin, [117, Chapter 11, p. 302], [112]), see [118]. The role of positive
polynomials in system theory [5, Chapter 2] has received a recent boost from its
use along with Linear Matrix Inequalities in formulations of problems in robust and
nonlinear control. See [119] for computational considerations and a easy-to-read
exposition of the semidefinite programming formulation of the SOS decomposition
of multivariate positive polynomials.
The following theorem (and its discrete counterpart), is presented in [116].
THEOREM 1.5 Let A be a (m x m) polynomial matrix in two variables p

(PI,P2) with normal rank r such that (i) A is para-Hermitian (i.e. A. = A or,
more explicitly A(PI, P2) = [A· (-pi, -P2)]t), (ii) A is nonnegative definite for
allp = jw.
Then there exists a (2r x m) rational matrix H which is holomorphic in Re p > 0
and is a spectral factor of A in the sense
A=H.H. (1.43)
Moreover, it is possible to have an H which is polynomial in one of the variables,

say P2 and is rational in the other variable Pl.
The above theorem and its discrete counterpart are useful in multiport synthesis
of lossless networks with prescribed scattering matrix, though Kummert based his
procedure [120] mainly on the factorization of an univariate rational matrix and
known properties of block companion matrices. It should be noted that the spectral
factor in the preceding theorem is matrix-valued even when m = 1, unlike the
spectral factors in the sense of Helson-Lowdenslager cited in Section 6 or Marzetta
[5, pp. 227-228]. Finite order bivariate polynomial approximant to the spectral
factor of the form in [5, p. 228], which is a polynomial in one variable but a
two-sided power series in the other variable, has been given in [121].
Chapter 2
CAUSAL AND WEAKLY CAUSAL 2-D FILTERS

WITH APPliCATIONS IN STABILIZATION
BY J. P. GUIVER AND N. K. BOSE
1. Scalar 2-D Input I Output Systems

A discrete 2-D scalar input/output system is described by means of an equation
{y[m, n]} = F( {x[m, n]}) (2.1)
where F is an operator which maps doubly indexed input sequences {x[m, n]}
into doubly indexed output sequences {y[m, n]} where we assume that x[m, n]
and y[m, n] take values in ~ the field of real numbers.
An important class of systems are the linear shift-invariant (LSI) systems. The
input/output relation of an LSI system can be represented by means of a convolution
equation
y[m, n] = L h[m - k, n - l]x[k, If (2.2)
[k,ljEZ 2
where
h[m, n] = F(&[m, n]) (2.3)
is the system response to a unit impulse &[m, n] applied at [0,0], and 'Ii} = Z x Z,
where Z is the set of integers.
We will denote the support of the sequence {h[m, n]} by Supp(h), i.e.
Supp(h) = {[r, s] E Z2 : h[r, s] =1= O} (2.4)
Usually some sort of causality is assumed which is specified by conditions on

Supp(h). Q1, defined below, is called first quadrant quarter plane (FQQP).
DEFINITION 2.1 We will say an LSI ~ystem F is causal if
Supp(h) C Q1 where Q1 = {(x, y) E ]R2 : x ~ 0, y ~ O} (2.5)
35
We will say F is weakly causal if (defined by R. Eising in [33])
Supp(h) cC (2.6)
where C is a closed convex cone in ]R2 satisfying
(i) C n (-C) = {(O, On

(ii) Q1 n C (2.7)
The first condition merely says that the cone makes an angle of less than 1r at
(0,0). Under condition (2.6) any input x[m, n] which satisfies
Supp( u mn ) n C is finite for all [m, n] E Z2 (2.8)
where
Umn[r, s] A x[m - r, n - s]
will give rise to a unique output by means of the convolution equation (2.2).
If a[m, n] is any doubly indexed sequence, we define its 2-D z- transform by
Z(a[m, n)) = A(z, w) = L a[m, n]zmw n (2.9)

[m,njEZ 2
If we then take the z-transform of each side of (2.2) we get
Y(z,w) = H(z, w)X(z, w) (2.10)
DEFINITION 2.2 H(z, w) = ~[m,njEZ2 h[m, n]zmw n is the system function.
H(z, w) can be regarded either as a formal power series or as a function of z

and w when the power series converges. From a practical standpoint it is desirable
to approximate H(z, w) by means of a rational function in z, w, z-1 and w- 1 for
purposes of recursive implementation of the input/output relation.
I-et
a(z,w) = L a[m,n]zmw n
[m,njEA
where A c Z2 is finite and
(0,0) E A and a(O, 0) =1 (2.11)
Then for recursive implementation of H(z, w) = [b(z, w))/[a(z, w)] by means

of a 2-D difference equation it is required (see O'Connor and Huang [122]) that A
be contained in a lattice sector with vertex (0,0) of angle less than 1r.
Here b(z, w) = ~[m,njEB .B[m, n]zmwn where it is assumed that B C Z2 is
finite, but, for recursive computability, no other conditions are needed on B.
Causal and Weakly Causal 2-D Filters with Applications in Stabilization 37
However, we will assume that
A,BeC (2.12)
where C is a closed convex cone satisfying (2.7). We then have the following
observation.
PROPOSITION 2.1 Let a(z, w), b(z, w) E lR[z, z-I, W, w- 1] with supports, A, B
respectively where A satisfies (2.11) and A and B satisfy (2.12). Then
[b(z, w)]j[a(z, w)] has a unique expansion L h[m, n]zmw n such that Supp(h)
eC.
Proof. The existence of the expansion follows since

b(z,w) b(z, w)
a(z,w) l-a*(z,w)
00
= b(z, w) I)a*(z, w)]i

i=O
where
a* = - L a[m,n)zmw n
A-HO,O)}
It is then clear that all exponents must be in C. D

The uniqueness follows from the work of Eising ([33, Lemma 2-10]), which
we will discuss later in this chapter. Eising [123] shows that a causal impulse
response has a finite dimensional local state space realization if (and clearly only if)
H(z, w) has a representation ofthe form fp(z, w)]j[q(z, w)] where p, q E R[z, w)
and q(O, 0) f= O.
DEFINITION 2.3 A rational junction [P(z, w))j[q(z, w)) with p, q E R[z, w] is

called causal if q(O, 0) f= O. It is called strictly causal ifin addition p(O, 0) = O.
It follows from Proposition 2.1 that this definition is consistent with the defini-
tion in (2.5). Strict causality will imply that h[O, 0] = 0, i.e. the output will not
depend on the current input. Eising [33] also shows that if H(z, w) has a repre-
sentation of the form [b(z, w)]j[a(z, w)) where a, b E R[z, z-l, w, w- 1] and a, b
have supports A, B respectively, satisfying (2.11) and (2.17), then H(z, w) has a
finite dimensional local state space realization.
DEFINITION 2.4 Ajunction of the form [b(z, w)]j[a(z, w)] where a and b are as
described in the above paragraph is called weakly causal. It is called strictly
weakly causal ifin addition b(z, w) has zero constant term.
This definition is, from Proposition 2.1 seen to be consistent with (2.6). Strictly
weakly causal will imply that the output will not depend on the current input - i.e.
there is no direct coupling.
The systems we consider in this chapter will be of the type described in Defini-
tions 2.3 and 2.4 i.e. (strictly causal and (strictly) weakly causal rational transfer
functions.
2. Stability
DEFINITION 2.5 A YO system (2.1) is said to be BIBO (bounded input/bounded
output) stable iffor all M > 0 there exists N > 0 such that for any input sequence
{x} satisfying Ix[m, n] I < M for all [m, n], the corresponding output sequence
satisfies Iy[m, n]1 < N for all [m, n].
For a system described by a convolution equation (2.2), BmO stability is well
known to be equivalent to
LL Ih[m,n]1 < 00 (2.13)

m n
Given a cone C satisfying (2.7), C c Z2 is contained in a lattice sector

(with vertex angle less than 11") of the form S[(Ml, Nd, (M2, N2)] defined by
the relatively prime integer pairs (Ml, Nd and (M2' N 2) (see [122]). We assume
that (MI' Nd and (M2' N2) are independent vectors; otherwise, we can map
S[(MI' Nd, (M2' N2)] onto Z+ and we would thus be dealing with an essentially
I-D situation (see [122]).
In general S[(MI' Nd, (M2' N2)] can be mapped into (but not necessarily onto)
QI n Z2 by means of the map
m' = kIm + k2n

n' = k3m + k4n (2.14)
where
ki = sgn(D)N2
k2 = -sgn(D)M2
k3 = -sgn(D)NI
k4 = sgn(D)MI (2.15)
and
D = MIN2 - M2Nl (2.16)

(Note: D =1= 0 by the assumption that (MI' Nt) and (M2' N2) be independent).
Then we have the following theorem of O'Connor and Huang [122].
THEOREM 2.1 LetSupp(h) c S[(MI' Nt), (M2' N 2)] (with vertex angle less than
11"). Then L h[m, n]zmw n is BlBOstableijandonlyijL g[m', n']zm' w n ' is BlBO
stable where
Theorem 2.1 says that we can determine the stability of a weakly causal transfer
function by checking that of an appropriately causal one.
As noted earlier, the map given in (2.14) does not in general map onto the whole
of QI n Z2. Eising [33, Lemma 2-5] remedies this by showing that any cone C
satisfying (2.7) is contained in a causality cone Ce. Ce n Z2 can then be mapped
bijectively (i.e in a 1-1, onto manner) to QI n Z2. For convenience we reproduce
(but don't prove) some details of Eising's work. (See [33] for full details).
DEFINITION 2.6 A causality cone C e is an intersection of two half-planes Hp,r

and Hq,t where
Hp,r = {(x, y) : (x, y) E ]R2 ,px + ry ~ O}

Hq,t = {(x,y): (x,y) E ]R2,qX+ty ~ O} (2.18)
where p, q, r, t are non-negative integers satisfying
pt-qr=1 (2.19)
Note that (Hp,r n Hq,t) n Z2 = S[(t, -q), (-r,p)].

LEMMA 2.1 Let C be a closed convex cone satisfying (2.7). Then there exists a
causality cone Ce such that C C Ce.
LEMMA 2.2 Let C e = Hp,r n Hq,t be a causality cone. Then the map 4> :
Ce n Z2 -+ QI n Z2 given by
4>[m, n] = [pm + rn, qm + tn] (2.20)
is bijective.
(4) is the same map as described in (2.14) and (2.15) with MI = t, NI =

-q, M2 = -r, N2 = p and D = pt - qr = 1).
It is important to note that for Ce = Hp,r n Hq,t, Ce n Z2 is a semigroup under
addition, with (0,0) as 'identity', which is generated by (t, -q) and (-r, p). (In
general a lattice sector S[(MI' Nt), (M2' N 2)] is a semi group but is not generated
by (MI' Nt) and (M2' N 2), or by any two vectors). Thefactthat [t, -q] and [-r,p]
generate Ce n Z2 follows from the fact that the 4> given above is a semi group
isomorphism which maps [t, -q] to [1,0] and [-r, p] to [0,1]. Given [m, n] E
Ce n Z2, <form, n] = [h, k] = h[l, 0] + k[O, 1].
Therefore,
[m, n] = h[t, -q] + k[-r,p]

= [ht - kr, -hq + kp]
where
h=pm+rn k = qm+tn
NOTATION 2.1 ForanyCe = Hp,r n Hq,t. weletSp,r,q,t = {H(z,w) : 3h[m,n]
such that H(z, w) = L: h[m, n]zmw n and Supp(h) C Ce }
With this notation the ring of causal power series is represented by S1,0,0,1 since
FQQP Q1 = H 1,0 n HO,l.
THEOREM. Sp,r,q,t is a ring which is isomorphic to S1,0,0,1. The ring
isomorphism
,,,
iP:sp,r,q, t-+S1001 (2.21)
is given by
L L h(<fo- [m, n])am{3n

00 00
iP(H)(a, {3) = 1 (2.22)

m=On=O
where <fo is as in (2.20).
LEMMA 2.3 The isomorphism iP can be described by the substitution
(2.23)
with inverse
(2.24)
The map ~ is of course different for different values of p, r, q, t but use of the
same notation should not cause any confusion.
NOTATION 2.2 For any Ce = Hp,r n Hq,t we let Pp,r,q,t represent the ring of
sums of the form L:[m,n]EA arm, n]zmw n where A is finite and A C Ce• Thus
P1,0,0,1 is just R[z, w]. the ring of two variable polynomials over R
NOTATION 2.3 For Ce = Hp,r n Hq,t we let Rp,r,q,t denote the subring of Sp,r,q,t
consisting oftransfer junctions representable in theform [b(z, w)]/[a(z, w)] where
a, bE Pp,r,q,t and a(z, w) has non-zero constant term.
NOTATION 2.4 We let Rp,r,q,t denote the subring of Rp,r,q,t consisting of those
junctions for which the numerator b(z, w)( E Pp,r,q,t) has zero constant term.
Thus, in the above notations, RI,o,O,1 and RI,O,O,1 stand for the rings of causal
and strictly causal rational functions respectively.
It follows from (2.23) and (2.24) that the cI> given in (2.22) maps
Pp,r,q,t bijectively onto PI,O,O,1

Rp,r,q,t bijectively onto RI,o,O,1 (2.25)
and
Rp,r,q,t bijectively onto RI,o,O,1

LEMMA 2.4 Given a(z, w) E Pp,r,q,t it follows from the observations following
(2.20) that a(z, w) can be written uniquely as a polynomial in ztw-q and z-rwp
and then cI> is just that map (see (2.23) and (2.24» which replaces ztw-q by a and
z-rwp by {:3.
Similar comments apply for Rp,r,q,t and Rp,r,q,t which are just quotients of
elements in Pp,r,q,t.
DEFINITION 2.7 A representation (bla)(a, b E Pp,r,q,t with a having non-zero

constant term) of an element in Rp,r,q,t will be called irreducible if there does not
exist aI, bl , c E Pp,r,q,t, with c(z, w) not a constant, such that a = alc, b = blc;
in other words if a(z, w) and b(z, w) have no non-trivial common factors in the
ring Pp,r,q,t.
Now, in general, given two arbitrary elements, a(z, w) and b(z, w) in Pp,r,q,t
n Pp' ,r',q',t', with no restriction on a(z, w)we can have a and b being relatively
prime in one ring but not in the other. For example, b(z, w) = z, a(z, w) = z + w
are relatively prime in PI,O,O,1 but have w as a common factor when considered as
elements of PI,O,I,I. However, the following proposition is true.
PROPOSITION 2.2 Let a, bE Pp r q t n Pp' r' q' t'. Then if a and b have no (non-
trivial) commonfactor in Pp,r,q,t: thena(z,~) 'a:W b(z, w) can only have common
factors oftheform (const.)zkw'([k, l] E Hp',r' n Hq"t') in Pp',r',q',t'
Proof. Supposea(z, w) andb(z, w) have a (non-constant) common factorc(z, w) E
Pp',r',q',t' which is not of the form (const.)zkw'((k, l) E Hp',r' n Hq',t'). Then
there exist a1 ( z, w), b1 (z, w) E Pp' ,r' ,q',t' such that
a(z, w) = adz, w)c(z, w)

b(z,w) = bl(z,w)c(z,w) (2.26)
After multiplying each side of (2) by appropriate powers of z and w

we see that there exist m, n ~ 0 such that
zmwna(z,w) and zmwnb(z,w) (2.27)

are in .IR[z, w] and have a common factor in .IR[z, w] not of the form
(const. )z"'w" , /-L, 1/ ~ O.
Let
~ : Pp,r,q,t --+ .IR[z, w] (see(2.25))
take
a(z, w) to ao(a, (J)
and
b(z,w) to bo(a,{J).
Then ~ takes (see (2.23»
and
zmwnb(z, w) to aI'ffl+rn{Jqm+tnbo(a, (J).

Now (2.27) indicates that a pm+rn {Jqm+tn ao (a, (J) and apm+rn{Jqm+tnbo(a, (J)
have a common factor not of the form (const. )ai {Jj , i, j ~ O. Therefore, ao (a, (J)
and bo(a, (J) must have such a common factor, say d(a, (J).
Then d(ztw-q, z-r wP) is a non-trivial factor of
a(z,w) = ao(ztw-q,z-rwP)
and
b(z,w) = bo(ztw-q,z-rwP)
contradicting the hypothesis that a( z, w) and b( z, w) are relatively prime in Pp,r,q,t.
D
COROLLARY 2.1 Letp(z,w) = [b(z,w)]j[a(z,w)]E Rp,r,q,t n Rp',r',q',t' (So

a, b E Pp,r,q,t n Pp',r',q',t' and a(z, w) has a non-zero constant term). Then
[b(z, w)]j[a(z, w)] is irreducible in Rp,r,q,t ifandonly ifitisirreducible inRp' ,r' ,q' ,t'.
Proof. Since a(z, w) = const.+ other terms, a factor of the form zkw l would
mean a(z, w) = zkw'((const.)z-kw-l+ other terms). But [k, I] and [-k, -I]
cannot lie in the same causality cone unless k = I = O. So a(z, w) cannot have
any non-trivial factors of the form zkw' either in Pp,r,q,t or in Pp' ,r' ,q' ,t'. The result
now follows from the previous proposition. D
3. Structural Stability
Given H E R1,o,O,1 it is convenient, for many purposes, to consider a slightly
more restricted definition of stability.
DEFINITION 2.8 H(z, w) = (P(z, w)]/[q(z, w)] E Rl,O,O,b where p and q are
relatively prime polynomials, is said to be structurally stable if q(z, w) f= 0 for
-2 -
(z, w) E U where U = {u E C : lui < 1} and U is its closure.
It follows from Theorem 2.1 that H E Sp,r,q,t is BmO stable if and only if
~(H) is BmO stable.
DEFINITION 2.9 H E Rp,r,q,t is said to be structurally stable if~(H) is struc-

turally stable.
This definition says that in any irreducible representation [b(z, w)]/[a(z, w)]
of H(a, b E Pp,r,q,t), (b/a) is structurally stable if and only if it is BmO stable
and b and a have no common zeroes on T2. (Note that by (2.23) and (2.24),
(z, w) E T2 if and only if (a, (1) E T2). By Corollary 2.1, therefore, Definition
2.9 is independent of which Cc (and therefore which~) we choose (for in general
there will be infinitely many causality cones containing the support of a(z, w) and
b(z, w».
NOTATION 2.5 We let e;,r,q,t stand for the ring of structurally stable elements of
Rp,r,q,t. Similarly, R;,r,q,t will stand for the structurally stable elements of Rp,r,q,t
We will denote by lRs [z, w] those polynomiaL~' which have no zeroes in U2 •
4. Multi-Input Multi-Output Systems

A discrete multi-inputlmulit-output (MIMO) 2-D system is described by means
of an equation
{y[m, n]} = F( {x[m, n]}) (2.28)
where now the input {x[m, n]} and output {y[m, n]} are doubly indexed vector
sequences. As in the scalar case one can put conditions on the system such as
linearity, shift-invariance, causality and rationality. The systems we consider will
be characterized by their transfer matrix H which will be assumed to be in Rmxk
where R is one of the rings Rp,r,q,t, Rp,r,q,t, e;,r,q,t, or R;,r,q,t. Conditions for a
matrix to be over any of these rings for p = t = 1, q = r = 0 «strictly) causal and
structurally stable (strictly) causal transfer matrices) will be given subsequently in
terms of a matrix fraction description.
U2
e1 ,
U1
: .. C
Y1 ..
+ e~
~
P
Y2 ..
j - +
Figure 2.1. Scalar feedback system
5. Stabilization of Scalar Systems

Let
n(z,w) -
p(Z, w) = d(z, w) E Rl,O,O,l (2.29)
where
n(z, W), d(z, w) E lR(z, w] (2.30)
The assumption that p is strictly causal as opposed to just causal is made to
avoid the possibility of delay free loops from within the feedback system (see, for
example, Cadzow [124]). Since any h E Rl,O,O,l can be split into its strictly causal
part and its direct coupling, this is not restrictive. We see from Definition 2.3 that,
in terms of n(z, w) and d(z, w), the restriction that p E Rl,O,O,l becomes
d(O, 0) #0 (2.31)
n(O, 0) = ° (2.32)
We also assume that n(z, w) and d(z, w) have no common factors. Note, how-
ever, that n and d will in general have zeroes in common but by the coprimeness
of n and d we have that
n(z, w), d(z, w) have only a finite number of common zeroes (2.33)
The feedback system we will consider is shown in Figure 2.1.

When we speak of a feedback system we will always mean one with the above
structure. The approach and notation used will essentially be that of Desoer et ai.
[125]. We will refer to p = p(z, w) as (the transfer function of) the plant and to
c = c(z, w) as (the transfer function of) the compensator. For the time being we
assume c E R1,o,O,1 although this restriction will be lifted in the next chapter.
Let
Then (see [125])

e = Heuu
where
(2.34)
and
y = Hyu u
where
Hyu = K(Heu - 1) (2.35)
and
K is the matrix [~1 ~] (2.36)
Consequently (Desoer and Chan [126]) we can restrict our attention just to Heu
in discussing the stability of the feedback system.
Now let us represent c(z, w) as a quotient of relatively prime polynomials.
( ) _ x(z,w) (2.37)
c Z,w -
y ( Z,w )
Then it easily follows from (2.35) that
yd
h e1u1 = y d + xn (2.38)
xd
he2u1 = y d + xn (2.39)
-ny
he1u2 = y d + xn (2.40)
yd
he2u2 = yd + xn (2.41)
DEFINITION 2.10 We will say pis stabilizable ifwe can find c(z, w) such that the
he;uj' (i,j = 1,2) are each in RtO,O,l'
We wish to characterize those compensators c(z, w) E R1,o,O,1 which stabilize

p = (n/d).Now clearly if we can find x(z, w) and y(z, w) such that
yd + xn is in llts [z, w]
then [x(z, w)]/[y(z, w)] is one such compensator. Suppose x(z, w) and y(z, w)
are such that yd + xn is not a stable polynomial. Can we still achieve stability due
to some cancellation in Equations (2.38) to (2.41) ?
Suppose yd + xn = TIT2 where Tl E lRs[z, w), T2 E lR(z, w)\:IR s [z, w). Now in
Equation (2.39), the only way cancellation can occur is if x and d have a common
factor since n and d are relatively prime, and x and y are relatively prime. Therefore
for he2t£1 to be (structurally) stable every irreducible factor of T2 must divide both d
and x; (in practice, though, such cancellation cannot be perfectly achieved and any
small fluctuation in the system will introduce instabilities). But a similar argument
for the stability of hel U2 would imply that every factor of T2 must also divide both
nand y contradicting the relative primeness of nand d, and x and y. So we have
the following proposition:
PROPOSITION 2.3 c(z,w) = [x(z,w»)/[y(z,w») will stabilize thefeedback ~ys

tem given in Figure 2.1 if and only if
yd + xn is in :IRs [z, w)
So to determine whether a given plant p(z, w) is stabilizable we must look at
the set of all yd + xn as y and x range over lR(z, w). Note that for x / y to be causal
we must also specify that y(O, 0) =I- 0; but in fact this follows automatically from
our assumption that n(O, 0) = 0. For if 4l>(z, w) E :IRs [z, w) is such that
yd+ xn = 4J
Then
y(O, O)d(O, 0) = 4J(0,0)

Therefore
y(O, 0)
4J(0, 0)
= d(O,O) =I- °
So p(z, w) is stabilizable by means of a causal compensator [x(z, w)]/[y(z, w)]
if and only if the ideal, < d(z, w), n(z, w) > (c lR(z, w)), intersects:IRs [z, w] non
trivially. Thus we get the following theorem.
THEOREM 2.2 p(z, w) = [n(z, w))/[d(z, w») is stabilizable by means ofa causal
compensator [x(z, w»)/[y(z, w») if and only if n(z, w) and d(z, w) have no com-
-2
mon zeroes in U .
°
Proof. NECESSITY: Ifn(z, w) andd(z, w) have a common zero (zo, wo) E U 2 i.e.
n(zo, wo) = d(zo, wo) = and Izol ~ 1, Iwol ~ 1, then for any 4J(z, w) E :IR[z, w]
with 4J = yd + xn we have
4J(zo, wo) = y(zo, wo)d(zo, wo) + x(zo, wo)n(zo, wo) = °

and therefore ifJ cannot be in lRs [z, w]

-2
SUFFICIENCY: Suppose n(z, w) and d(z, w) have no common zero in U . Let
us list the common zeros of n and d as
(Zl' wt), (Z2' W2), ... , (zs, ws), (ZsH' WsH),···, (Zt, Wt)
where we can choose the indexing such that
i = 1, ... ,8
i = s + 1, ... , t
Now consider the real polynomial
s t
1j;{z,w) = IT{Z-Zi)(Z-zi) IT (W-Wi)(W-Wi)
i=l i=s+l
where * denotes complex conjugation. Clearly
i = 1, ... , t
So1j;(Z, w) vanishes wheneverbothd{z, w) andn{z, w) do. Therefore, by Hilbert's
Nullstellensatz, there exists
N E Z+ \{O} such that 1j;N E< d, n >

where < d, n > denotes the polynomial ideal generated by d and n. In other
words, there exist x{z, w), y(z, w) E lR[z, w] such that
yd + xn = 1j;N
Now 1j; and hence 1j;N clearly has no zeroes in U 2. Hence 1j;N is in lRs [z, w] and p
is stabilizable. 0
6. Characterization of Stabilizers for Scalar Systems

In the previous section we found a necessary and sufficient condition for the
stabilizability of a given transfer function p(z, w) and also found a compensator
which would achieve this stabilization. Now we wish to characterize in some way
the set of all such compensators.
We use a technique which is becoming increasingly popular in the I-D literature
(see in particular Desoer et al. [125] and Vidyasagar [127]); that is, instead of
regarding the compensator c(z, w) as a quotient of polynomials, we consider c as
a quotient of elements taken from a different ring, in our case the ring Rto,o,l' of
(structurally) stable causal 2-D rational transfer functions. It is clear that the ring
of polynomials R[z, w] is contained in Rf,O,O,1' So, the representation nj d of p

as a quotient of polynomials is also a representation as a quotient of elements in
RS1,0,0,1'
Now in the axiomatic approach of Desoer et aI. [125], the fundamental
hypothesis required for their characterization theorem for compensators
is the existence of a coprimeness equation of the type
vd + 'Un = 1
where p = nd- 1 is a representation of the plant p as a quotient of elements from
the ring that one is working in. If we work in the ring of polynomials R[z, w]
then such an equation will not in general hold, since even if n and d are relatively
prime they will still have common zeroes which must then also appear on the right
hand side of the equation. So what we do is loosen the restriction that 'U and v be
polynomials. This leads to the following proposition.
PROPOSITION 2.4 Let n, d E R[z, w] be relatively prime polynomials. Then there

exist 'U, VERt ,0,0 ,1 such that
vd + 'Un =1
if and only if n and d have no common zeroes in U2 •
Proof. Suppose vd + 'Un = 1 for some 'U, VERt ,0 ,0 ,l' Let
where
and
Then
Now let (zo, wo) E c'2 be any common zero of nand d. Then either
'U2(ZO, wo) = 0 or V2(ZO, wo) = 0 which implies that (zo, wo) = 0 ¢ U 2
since 'U2, V2 E lRs [z, w].
-2
Conversely, suppose n and d have no common zero in U . then we know from
the previous section that there exist x, y E lR[z, w] and <P E :IRs [z, w] such that
yd+xn = <P
Therefore, taking u = (x/</J) , v = (y/</J) we get vd+ un = 1 with u, vERt 001'

, "0
Now combining Proposition 2.4 with Theorem 2.2 gives us
PROPOSITION 2.5 Ifn, dE lR[z, w] are relatively prime, then n/d is stabilizable
if and only if there exist u, VERI ,0 ,0,1 such that
vd+un = 1
This immediately leads via the work of Desoer et al. [125] to a characterization
of the stabilizing compensators. For completeness, we present all of the relevant
details here.
THEOREM 2.3 Let n, d E lR[z, w] be relatively prime polynomials with no common

-2
zeroes in U. Let u, VERI 0 0 1 be such that vd + un = 1. then for any
s(z, w) E RIo , the compens~t'o~
, ,01
-sd+u
C=--- (2.42)
sn+v
is causal and gives rise to a (structurally) stable feedback ~ystem. Conversely, if
c( z, w) is a causal (rational) compensator which gives rise to a stable feedback
~ystem, C can be written in the form of (2.42).
Proof. Let
s E RI ,0 ,0,1 be arbitrary.
Let
where
and
Then
-sd+u
C=---
sn+v
x(z, w)
y(z, w)
where
x = -81V2U2d + 82V2U1
Y= 81 V2u2n + 82V1U2
Now
and since n(O, 0)

V1(0,0) # 0.
= °and V2(0, 0), U2(0, 0) are both different from 0, we have that
So
y(O,O) = 82(0, 0)V1 (0, 0)U2(0, 0) # °
and c is therefore a causal (though not necessarily stable) transfer function.
Furthermore,
yd + xn = (81 V2U2n + 82V1 u2)d + (-81 V2U2d + 82v2uI)n

= 82(V1U2d + V2U1n) = 82U2V2
So, from (2.39) to Definition 2.10 we see that the feedback system is stable.
Conversely, suppose that c(z, w) is a causal (rational) transfer function which
stabilizes the plant. Represent c( z, w) as a quotient of relatively prime polynomials
( ) _ x(z,w)
c Z,w - (
y Z,w
) y(O, 0) # °
Then, from Proposition 2.3 we have that
yd + xn = 4> E IRs [z, w]

i.e.
(2.43)
is a solution of
where 1], eE R!, , ,

001 (2.44)
The homogenous equation
(2.45)
has general solution 1]h = 8n, eh = -8d where 8 E R~ ,00

, ,1 is arbitrary.
For certainly 'fJh = sn, eh = -sd is a solution of the homogenous equation and
if
then
°°
So we just need to show s = (-e h j d) is in Rf , , ,I; this follows since
vd+un = 1
and therefore
h e h = -v."ch + u'fJ h E RSI 00 I

- -e = -v."ch - un-
d d ",
So from (2,45) and the hypothesis of the theorem, the general solution of
'fJd + en = 1 for 'fJ, eE RtO,O,1

is
'fJ = v + sn, e= u - sd sERfoOI

,,,
°°
Therefore, there exists sERf , , ,I such that
y x
- = v + sn - = u - sd
<P <P
c=:' = :'j'Jt = -sd+u
y <p <p sn + v
D
COROLLARY 2.2 Let nand d be relatively prime polynomials with no common

-2
zeroes in U . Let x, y E lR[z, w] be such that
yd + xn E lRs [z, w]
Then the stabilizing compensators are characterized by the expression
-sId + S2X
c= (2.46)
where
SI E lR[z, w)
and
S2 E IRs [Z, w) are arbitrary

Proot If yd + xn = if>, let v = y j if> , u = x j if> and use the previous theorem. D
To summarize: given a plant p{z, w) = [n{z, w»)j[d{z, w») where d and n are
relatively prime, proceed as follows:
(1) If n and d have common zeroes in U 2 then p cannot be stabilized by a causal

(rational) compensator.
-2
(2) If n and d have no common zeroes in U ,then find x{ z, w) and y{ z, w) such
that yd + xn is stable. A method for doing this is discussed in the proof of
Theorem 2.2 using the constructive approach given in Chapter 4 to determine
whether a given polynomial is in a given polynomial ideal specified by a finite
number of generators.
(3) The compensators are then given by
-SId + S2X
c=
SIn + S2Y
where SI is an arbitrary element of IRs [z, w)
(4) From (2.34) we then obtain with the above c
1
Example 2.1
(a)
n{z,w) zw - z - 2w
d{z,w) Z W - 2zw2 - 2zw + 4z + 4
2 2
We construct a Grobner basis (see Chapter 4 of this text) for the polynomials
z2w 2 - 2zw 2 - 2zw + 4z + 4 and zw - z - 2w and in the process generate
the polynomials z2 + 4z + 4 E IRs [z, w). In fact we have:
z2 + 4z + 4 = d{z, w) - z{w + 1)n{z, w)
The compensators are then characterized by the expression
-SI{Z, w)(z 2w 2 - 2zw 2 - 2zw + 4z + 4) - S2{Z, w)z{w + 1)
SI{Z,W)(zw - z - 2w) + S2{Z,W)
where
81 E lR[z, w] and 82 E lRs [z, w]
(b)
n(z, w) z+w
d(z,w) 1-z+w
n(z, w) and d(z, w) have a common zero at the point (~, -~) E U2 • Conse-
quently we cannot stabilize this transfer function by means of a causal compen-
sator; however we will see in Section 9 (Example (2.4» that it can be stabilized
by means of a weakly causal compensator.
7. Stabilization of Strictly Causal Transfer Matrices

Let P be an m x k matrix with entries
D .. ( ) _ nij(z, w) (2.47)
.r~3 Z,W - d .. ( )
'3 Z,W
where
(2.48)
and are relatively prime
i = 1, ... ,m, j = 1, ... ,k
We will be interested in those P for which

dij (0,0) =1= 0 (2.49)
and
(2.50)
i= 1, ... ,m, j = 1, ... , k
Recall (Section 4) that such a P is called a strictly causal rational transfer matrix.
The following two propositions illustrate the relationship of Conditions (2.49) and
(2.50) with matrix fraction descriptions of P.
PROPOSITION 2.6 (i) If P satisfies (2.49) then any irreducible left matrixfraction
description (IMFD) A-I B of P has (det A) (0, 0) =1= o. (see references [39],
[128] for definitions of IMFD, irreducible IMFD or coprime LMFD)
(ii) Conversely, if one can find a (not necessarily irreducible) IMFD A-I B of P
such that (det A)(O, 0) f= 0, then P satisfies (2.49).
54 MULTIDIMENSIONAL SYSTEMS THEORY AND APPliCATIONS
Proof. Let
di = LC M {dij : j = 1, ... , k} i = 1, ... , m
DI =diag{dl, ... ,dm }
and
Clearly, P = DII NI
Fonn [DI NI] and extract the greatest common left divisor L of [DI Nd using
the procedure in [39] to obtain
[DI NI] = L[DL NL]

(i) It is clear that (det Dd(O, 0) f= 0 (for P satisfying (2.49» and therefore (det
DL)(O,O) f= O.
Given any other irreducible LMFD A-IB of P we know (follows transitively
from results in [128, Theorem 5.2])
det A = k det DL
where k is a non-zero constant.

(ii) A-I B = (A+ B)/(det A) where A+ is the adjoint matrix of A. The result
now follows from (2.48).
o
PROPOSITION 2.7 Let A-IB be any (not necessarily irreducible) LMFD of P
such that (det A)(O, 0) f= O. Then the following are equivalent.
(i) Equation (2.50) is satisfied.
(ii) bij(O,O) = Ofor any entry bij of the matrix B.

(iii) All m th order minors of [A B], except for det A, are zero when evaluated at
(0,0)
Proof. Follows immediately from A[Im P] = [A B]. 0
Similar statements for the above two propositions follow of course for any right
matrix fraction description (RMFD) of P.
PROPOSITION 2.8 Let P be an m x k strictly causal rational transfer matrix. We
can represent P by means of an irreducible LMFD
(2.51)
~2
~1 Ir+
~1 + ~1 ~~ ~
C P
- +
Figure 2.2. Matrix feedback system
where from Proposition 2.6 and Proposition 2.7
(det DL)(O, 0) -=1= 0 (2.52)
and
(2.53)
Since Di/ NL is irreducible we have that the mth order minors of [DL NLJ
have no common factor and hence have only a finite number of common zeroes.
7.1 MIMO FEEDBACK SYSTEMS AND THEIR STABILIZATION
As in the scalar case we will be considering a feedback system of the type shown
in Figure 2.2
Here,
Yl, U2 and e2 are k x 1 vectors

Y2, Ul and el are m x 1 vectors
and C is assumed to be a k x m causal rational transfer matrix.

Let
e = [ el ]
e2 '
then
e = Heuu
where
-(Im+PC)-lp ]
h - C(Im + PC)-l p
(2.54)
Note that Heu is well defined since at (0,0), P and C are both well defined
(being causal) and P(O, 0) = 0mxk. So (1m + PC)(O, 0) = 1m.
We also have
where
(2.55)
and K is the matrix
[ Okxm h ] (2.56)
-1m 0mxk
So as in the scalar case, we can restrict our attention just to Heu in discussing
the stability of the feedback system.
Now represent C by an irreducible RMFD
C = XRYi1 (2.57)
Since C is assumed to be causal, we have from Proposition 2.6 that

(2.58)
It now follows from (2.54) that
he1u1 = YR(DLYR + NLXR)-l DL (2.59)

h e2u1 = XR(DLYR + NLXR)-l DL (2.60)
h e1u2 = -YR(DLYR + NLXR)-l NL (2.61)
he1u1 = h - XR(DLYR + NLXR)-l NL (2.62)
We could also work in terms of an irreducible RMFD for P, and an irreducible
LMFDforC;
P=NRD"R 1 (2.63)
It is not difficult to show that
H - [ 1m - P(h + CP)-lC -P(lk + CP)-l ] (2.64)
eu - (lk + CP)-lC (lk + CP)-l
It then follows from (2.63) that
h e1u1 = 1m - NR(YLDR + XLNR)-lXL (2.65)

h e2u1 = DR(YLDR + XLNR)-l XL (2.66)
h e1u2 = -NR(YLDR + XLNR)-lYL (2.67)
he1u1 = DR(YLDR + XLNR)-lYL (2.68)
PROPOSITION 2.9 det(YLDR + XLNR) = a det(DLYR + NLXR) where a is a

non-zero constant.
Proof.
[ YL
NL
XL] [DR
-DL NR 1m
0] = [ YLDR + XLNR
0
XL]
-DL
Therefore, since from [128, Theorem 5.2] detDR = const. det DL, we have
that
det [~ :~L] = const. det(YLDR + XLNR) (2.69)
Similarly,
[~~ -~L] [~ ~~] - [~~ DLYR1NLXR]

implies (since det YL = const. det. YR) that
det [~~ ~L] = const. det(DLYR + NLXR) (2.70)
The result now follows from (2.69) and (2.70). D

COROLLARY If P = Di l NL and C = XRYill are any other irreducible
2.3
MFD's of P and C then det(DLYR + NLXR) = det(DLYR + NLXR)
Proof. Follows transitively from Proposition 2.9. D
We now show that det(DLYR + NLXR) (which by the above is independent
of the representations of P and C provided they are irreducible) contains all the
information about the (structural) stability of the matrix He'/},.
PROPOSITION 2.10 The feedback ,system in Figure 2.2 is (structurally) stable if
and only if det( D LYR + N LXR) is a stable polynomial.
Proof. The feedback system is stable if and only if the matrix He'/}, is stable. This
is true if and only if
But
hel '/},l - hel '/},2 ]
[
he2 '/},l h - he2 '/},2
= [ YR(DLYR + NLXR)-l DL YR(DLYR + NLXR)-l NL] (2.71)
XR(DLYR + NLXR)-l DL XR(DLYR + NLXR)-l NL
= [ .i~ ](DLYR + NLXR)-l [DL NL ]

Now
since YR and X R are right coprime
and D L and N L are left coprime,
we have that
(2.72)
and
so if
[i~ ] (DLYR + NLXR)-l [DL NL] = A- l B
where A-l B is irreducible, then from (2.72) and (2.73) it follows that det(DLYR +
N LXR) = const. det A. The result now follows since A -1 B is (structurally) stable
if and only if det A is a stable polynomial on using Theorem 2.4 below. 0
THEOREM 2.4 Let P = ALl BL = VT-1U, where ALl BL, VT- l and T-1U
are all irreducible. Then det T = k· det AL where k is a non-zero constant. See
[128J for a proof,
NOTE 2.1 Proposition 2.9 and Corollary 2.3 also follow as corollaries to (the
proof of) Proposition 2.10 by uniqueness of'denominators'.
LEMMA 2.5 Let NL,DL satisfy (2.52) and (2.53). Then ifdet(DLYR + NLXR)
is stable, (det YR)(O,O) of O.
Proof. Sincedet(DLYR+NLXR) is stable, therefore, (det(DL YR+NLXR»(O, 0) of
O. The result now follows from the Cauchy-Binet theorem and (2.52) and (2.53).
o
We now state a slightly extended version of a result due to Youla and Gnavi
[40]. The following Proposition though also stated and proved in Chapter 3 (see
Lemma 3.10) in a manner different from that in [40], is quoted here to facilitate
comprehension of the proof of Theorem 2.5. For a simple proof of Proposition
2.11, see the proof of Lemma 3.4 in Chapter 3.
PROPOSITION 2.11 Let A be an m x l normal full rank matrix with entries in
D, m ~ l (here D is an Euclidean domain)
Let ail ...i m = A ([ .1. .. ':" ]) 2

Sl·· ,Sm
2This denotes the determinant of a submatrix of a matrix A obtained by taking rows 1,2, ... ,m and columns
it,i2,'" ,im
and let I be the ideal of D generated by the ail ...im 'so

Then, given any a E I, there exists an i x m matrix B with entries in D such
that
AB=a1m
THEOREM 2.5 P(z,w) = Di1(z,w)NL(z,w) is stabilizable by means of a
causal compensator if and only if the mth order minors of [DL NL] have no
-2
common zero in U .
Proof. NECESSITY: If the mth order minors of [DL NL] have a common zero
(zo, wo), then for any YR, XR we have, by the Cauchy-Binet theorem, (zo, wo) is
]]
a zero of
det(DLYR + NLXR) = det [[DL NL] [ r~
So if Izol ~ 1, Iwol ~ 1, then det(DLYR + NLXR) cannot be stable.
SUFFICIENCY: Suppose the mth order minors of [DL NL] have no common
zero in U 2 . Since Di 1NL is irreducible, the mth order minors of [DL NL] have
only a finite number of common zeroes. We can therefore, precisely as in the proof
of Theorem (2.2) find a stable polynomial 'IjJ(z, w) which vanishes at each common
zero of the mth order minors of [DL NL]. Therefore by Hilbert's Nullstellensatz
and by Proposition 2.11 there exist N E Z+ \ {O}, and matrices Y and X, of order
mxmandkxm, respectively, with entries in R.[z, w], such that [DL NL] [ i- ] =
'ljJN I.
Consequently Xy- 1 is a compensator which stabilizes P, and is causal by
Lemma 2.5. D
8. Characterization of Stabilizers for MIMO Systems

Like in the scalar case we can characterize in a systematic way those (causal)
compensators which stabilize P = Di 1 NL.
ASSUMPTION 2.1 We again assume for the remainder of this chapter that Di 1 N L
satisfies (2.51), (2.52), and (2.53).
PROPOSITION 2.12 There exist V E (R 1,0,0,1

S )mxm and U E (R 1,0,0,1
S )kXm such
that
DLV+NLU=lm
if and only if the mth order minors of[DL NL] have no common zeroes in U 2
Proof. SupposeDL V +NLU = 1m with V E (Rto,O,l)mxm andU E (Rto,O,l)kxm.
Then
where VI E RmXm[z, W), U1 E RkXm[Z, w), and V2, U2 E Rs[z, w).

Thus U2DL VI + V2NLU1 = U2v2I, so by the Cauchy-Binet theorem, any
common zero (zo, wo) E ((J of the mth order minors of [DL NL) must also be a
zero of U2 or V2. Since U2 and V2 are in Rs[z, w) such a zero cannot be in U2.
Conversely, if the mth order minors of [DL NL) have no common zeroes in
-2
U , we know from the sufficiency part of the proof of Theorem 2.5 that there
exist Y E JlFxm[z,w) and X E RkXm[z,w) and </J E ~xm[z,w) such that
DLY + NLX = </JIm.
Therefore taking
gives
with
V E (R~oodmxm,
,,, U E (R 1,0,0,1
S )kXm
o
Combining Proposition 2.12 with Theorem (2.5) gives
PROPOSITION 2.13 D'L1 NL is stabilizable by means of a causal compensator

if and only if there exist VR E (Rf,o,O,l)mxm, UR E (Rf,o,O,l)kXm such that
DLVR +NLUR = 1m.
As in the scalar case this leads, using the work of Desoer et al. [125] to a
characterization of the stabilizing compensators. Again, we give all details. The
next Theorem explicitly parametrizes the stabilizing compensators based on the
MFD description of a bivariate MIMO plant.
THEOREM 2.6 Let D"i 1NL and NRDi/ be irreducible representations of the
strictly causal plant P(z, w) such that the mth order minors of [DL NL] have
no common zeroes in U2. Let UR E (Rto,O,l)kXm and VR E (Rto,o,d mxm be
such that DLVR + NLUR = 1m (such UR and VR exist by previous proposition).
, ,dkxm the compensator
Then for any S E (Rr ,00
(2.74)
is causal and stabilizes the feedback system. Conversely, if C (z, w) is a rational

causal compensator which stabilizes the feedback !Jystem, C can be written in the
form of (2.74).
Proof. Let S E (Rf ,0,0,d kxffi be arbitrary. Let
S -- SI ,
82
where
SI, Ul E rXffi[z, w], VI E IRffiXffi[Z, w]

82, V2, U2 E IRs [z, w]
Then
C = (-DRS + UR)(NRS + VR)-l

=Xy- l
where
x -V2U2DRSI + 82V2Ul
=
Y = V2U2NRSI + 82U2Vl
Now
(2.75)
Since all the mth order minors of [DL NL] except for det DL are zero when
evaluated at (0,0) (by (2.52) and (2.53», and since U2(0, 0) =1= 0, V2(0, 0) =1= 0 (U2
and V2 being in IRs [z, w]), we have from (2.75) and the Cauchy-Binet theorem that
(det Vl)(O,O) =1= O.
Therefore,
(det Y)(O, 0)
= det (Y(O, 0))
= det((V2UINRSt} (0, 0) + (82U2Vt) (0, 0))
= det(82u2Vt} (0, 0) (since NR(O, 0) = 0 by Proposition 2.7(ii)
= 82(0, 0)U2(0, O)(det Vt}(O, 0) =1= 0
and hence C is causal.
Furthermore,
DLY+NLX
= V2U2DLNRSl + 82U2DLVl - V2U2NLDRSl + 82V2N LUI
= 82U2DLVl + 82V2NLUl (since DLNR = NLDR)
= 82u2v2Im (see (2.75»
So, from (2.59), (2.60), (2.61), and (2.62) we see that the feedback system is
stable.
Conversely, suppose C{z, w) is a rational causal transfer matrix giving rise to a
stable feedback system.
Let XY-1 be an irreducible RMFD of C.
Since we are assuming C causal we have from Proposition 2.6 that
(det Y)(O, 0) =I 0 (2.76)
From Proposition 2.10 we have that (DLY +NLX) = ep where det ep E Rs[z, w].
Therefore,
H = yep-I, S = Xep-1 (2.77)

is a solution of
DLH + NLS = 1m where H E {Rto,o,d mxm and S E (Rto,o,d kxm (2.78)
The homogenous equation
DLHh +NLSh = Omxm
has general solution
Sh = -DRS (2.79)
where S E {Rf 00 l)kXm is arbitrary.
For certainly (2.79) is a solution (since NLDR = DLNR) for arbitrary S E
(RI,o,O,l)kXm; and if DLHh + NLS h = Omxm where
Hh E {R 1,0,0,1
S )mxm ,.....
,=h E {R 1,0,0,1
S )kXm
then
H h -- -D- ,=h -- N R (_D-R 1 .....

L N L .....
1 ,=h)
D -
- -R
.....h -
= (RD-=
1 .....h )
So we just need to show

1 ,=h E (R S
- D-
R ..... 1,0,0,1 ·)kxm., this follows since
VLDR+ULNR=h forsome
UL E (Rto,o,d kxm , VL E (Rto,o,d kXk
and therefore
So from, (2.79), the general solution of DLH + NL'B = 1m for

10 1 ° 1 °°
H E (R S , , , )mxm ';:;' E (R S , , , )kXm .
, ~ 1 IS
(2.80)
where
S E (R 1,0,0,1
S )kXm
Therefore, from (2.77) and (2.78) 3S E (Rto,o,.)kxm such that
YtP- 1 = VR + NRS
XtP- 1 = UR - DRS
Therefore
c = Xy- 1 = XtP- 1 tPy- 1

= (XtP- 1 )(YtP- 1 )-1
o
COROLLARY 2.4 Let NL, DL, NR, DR be as in the previous theorem. Let XR E
rxm[z, w] and YR E lIFxm[z, w] be such that
det(DLYR + NLXR) E lRs[z, w]

Then the stabilizing compensators are characterized by
where
R1 E lRk xm [z, w]
and
r2 E lRs [z, w] are arbitrary.
9. Stabilization of Weakly Causal Systems

NOTATION 2.6 Let
p,q,r, t E Z+
such that
pt-qr=l
V(p, T, q, t).
PROPOSITION 2.14 Let
b(z, w)
a(z, w)
be an irreducible representation (see Definition 2.7) ofp(z, w) E Rp,r,q,t where
a, bE Pp,r,q,t and a(z, w) has a non-zero constant term. Then p E R;,r,q,t if and
only if a(z, w) has no zeroes in V(p, T, q, t).
DEFINITION 2.11 By a zero of a(z, w) we mean a pair (zo, wo) E ~ for which
each term in a(z, w) is well defined and such that a(zo, wo) = O. Note that by this
definition (0,0) cannot be a zero of a(z, w).
Proof. (Of Proposition 2.14)
( ) _ b(z,w)
p z,w - a (z,w )
We can write (see Lemma 2.4) b(z, w) and a(z, w) as polynomials in ztw-q
and z-rwp:
a(z, w) = ao(ztw- q, z-r wP) (2.81)

b(z, w) = bo(zt w - q, z-r wP) (2.82)
and
~ (b(Z, w)) = ~(b(z, w)) = ~(bo(ztw-q, z-r wP )) = bo(a, (3) (2.83)

a(z, w) ~(a(z, w)) ~(ao(ztw-q, z-rw p)) ao(a, (3)
Now if there exists (zo, wo) E V(p; T, q, t) such that a(zo, wo) = 0, then setting
t -q f3 - -r p
ao -- zowo , 0 - Zo Wo
we get that
(ao, f3o) E U 2 and ao(ao, f3o) =0
Therefore, from the definition of structural stability in Rp,r,q,t we have
b(z, w) s
a(z, w) ¢ R;,r,q,t
Conversely, suppose
b(z,w) ¢ s
a(z, w) R;,r,q,t
°
Then
-2
ao(ao, .80) = for some (ao,.80) E U
By the continuity of the zeroes of a polynomial we can assume that
Then, defining
Zo = %.8g, Wo = ao.85
we obtain a zero (zo, wo) of a(z, w) lying in V(p, r, q, t). 0
We now consider the problem of stabilizing a plant p(z, w) in Rp,r,q,t using a
compensator c(z, w) in Rp,r,q,t. We use the same feedback configuration as in
Figure 2.1. Because of the ring isomorphism cp which maps Rp,r,q,t bijectively
onto Rl,O,O,l and Rp,r,q,t bijectively onto Rl,O,O,l we can convert the problem to the
causal case, solve it using the methods of Section 5 and then convert back again.
So from Theorem 2.2 we have the following proposition.
PROPOSITION 2.15 The plant p E Rp,r,q,t is stabilizable by means of a compen-

sator c E Rp,r,q,t if and only if
CP(a)(a,.8) and cp(b)(a,.B)
have no common zeroes in U2 (where [b(z, w)]j[a(z, w)] is irreducible}.
One might expect that in terms of a(z, w) and b(z, w) the condition for sta-
bilizability is that a(z,w) and b(z,w) have no common zeroes in V(p,r,q,t).
Unfortunately this is not the case since CP(a)(a,.8) and CP(b)(a,.8) may have ze-
roes in U 2 of the form (ao, 0) or (0, .80) and these are not describable in terms of
z andw.
Example 2.2 p =t=q=1 r=O

b(z, w)
a(z,w)
It is easily shown that b(z,w) and a(z,w) only have a common zero at (-2,-2)
which does not lie in V(1, 0,1,1).
However
and
cp(a)(a,.8) = 1+a +.8
have common zeroes at (1,-2) and (-1,0). Therefore [b(z, w)]j[a(z, w)] is not
stabilizable in Rl,O,l,l.
p, r, q, t with pt - qr = 1, we let ap,r(z, w) denote that portion of

If, given
a(z, w) involvingjustpowersofwP z-r (including the constantterm) andaq,t (z, w)
denote that portion of a(z, w) involving just powers of ztw-q (including the con-
stant term), and similarly define bp,r(z, w) and bq,t(z, w), then the criterion for
stabilizability clearly is as follows.
PROPOSITION 2.16 With notations described, the stabilizability criterion becomes
(i) b(z, w) and a(z, w) have no common zero in V(p, r, q, t)

(ii) bq,t(z, w) and aq,t(z, w) have no common zero in
(iii) bp,r(z, w) and ap,r(z, w) have no common zero in
Given a weakly causal rational transfer function, the region of support of the
numerator and denominator is not contained in an uniquely defined causality cone.
In this section we consider the following problem: given a plant p(z, w) E Rp,r,q,t
which is not stabilizable in ....1?~'"
r q t, find
p* , r* , q* , t*
such that
[Hp,r n Hq,t] c [Hp*,r* n Hq*,t*]
(and therefore p(z, w) E Rp* ,r* ,q* ,t*) and p(z, w) is stabilizable in Rp* ,r* ,q* ,t*.
Example 2.3 Consider the p(z, w) given in Example (2.2).
. b(z, w) Z2W-2 + zw- I + w

I.e. a(z,w) = 1 + zw- I + w
We chose as our region of support HI,O n HI,I' However, we found that while
conditions (i) and (iii) of Proposition 2.16 were satisfied, condition (ii) was not
since z2w-2 + zw- I and 1 + zw- I have a common zero at (1,-1).
However, if instead we chose p = t = 1, q = 2, r = 0 we now have that
aq,t(z, w) == 1 and hence condition (ii) is satisfied as well. In terms of ~(a) and
~(b) we have
~b(z,w) =~(ZW-2)2W2+(zw-2)w+W = u 2r:J2+uf3+f3

a(z, w) 1 + (zw- 2)w + w 1 + uf3 + f3
Now it is easily checked that a 2f32 + af3 + f3 and 1 + af3 + f3 only have a
common zero at (ao, f3o) = (-1/2, -2).
We conclude that although [b(z, w )]/[a(z, w)] is not stabilizable in Rl,O,l,l it is
stabilizable in R 1,o,2,1.
In order to draw conclusions for the general case we first prove the following
lemma.
LEMMA 2.6 Let
p,q,r, t E Z+
be such that
pt-qr=l
Let
be a finite set of points in
Ql \ {(O, 0), (1, 1)} (Ql is the first quadrant ojJR.2 ) such that
for all i = 1, ... , n
X~~ <
- Y~~ .J! <
and Yi - x":~
Then
3p', q', r', t' E Z+ with p't' - q'r' = 1
and
[Hp,r n Hq,t] C [Hpl,r l n Hql,t']
such that for each i = 1, ... ,n either
Proof. We first note that

t q
xi = Yi
implies either (Xi, Yi) = (0,0) or (1,1).
But we have excluded (0,0) and (1,1) from our values for (Xi, Yi). Therefore we
can split the (Xi, Yi)'S into two sets:
Z1 = {(Xi,Yi) : xf < yi,yf ~ xi}
and
Let
(Then
{1, ... , n} = 11 U 12
and 11 n 12 is empty).
We can assume that q and r are not zero since for example if q = 0, then
p = t = 1 and we define a new (larger) causality cone HPl,Tl n Hql,h by
P1 =P
q1 =P r1 =r
Since p, t can never be zero we therefore assume that p, t, q, r are all greater than
zero.
For i E I1, xf < Yl and therefore
(~1)k -+00 as k-+oo
(Note that Xi cannot be zero since this would imply Yi = 0 contradicting x: < yi).
It follows that for each i E 11 there exists k i E Z+ such that
Now let
k = maxki
iElt
p' =p+kq
r'=r+kt
then
p't - qr' = pt - qr =1
and
for i E 11 yfz. ' >x·zr' (2.84)
Certainly, [Hp,r n Hq,tl c [Hp' ,r' n Hq,t] for if x, y E ]R2 is such that
px+ry ~ 0, qx + ty ~ 0
then
p'x + r'y = px + ry + k(qx + ty) ~ 0
Now consider (Xi, Yi) E Z2
Since
x~z = y9z
we have
yfz.' <X·zr'
Therefore for i E 12
x.+r t ' > x·-.1!

t ' = x.xr: t' , +'
z z z ,Yi = y9-.1!
z Yi = y9z p
So if we let
t'=t+r', q'=q+p'
then
p't' - q'r' = p't - qr' = 1

and
(2.85)
Also by an argument similar to that given before
[Hp,r n Hq,t] C [Hp',r' n Hq,t] C [Hp',r' n Hq"t']

and so by (2.84) and (2.85), the proof of the lemma is complete. D We can now
state and prove the main theorem of this chapter.
THEOREM 2.7 Let
b(z, w)
a(z,w)
be an irreducible representation ojp(z, w) E Rp,r,q,t. Then there exists

a causality cone C; = Hp * ,r* n H q* ,t* such that
Cc ~ [Hp,r n Hq,t] C C;
and such that p( z, w) is stabilizable by means oj a compensator
c(z, w) E Rp* ,r* ,q* ,t*
if and only if
b(z,w) and a(z,w)
have no common zeroes on T2.
Proof. If b(z,w),a(z,w) have a common zero (zo,wo) E T2, then in whatever
causalityconeCc = Hp,r n Hq,t thesupportsofa(z, w) andb(z, w) are considered
to be, ~(b)(a, (:J) and ~(a)(a, (:J) have a common zero at (ao, (:Jo) E T2 where,
Conversely, suppose a(z, w) and b(z, w) have no common zero on T2.

Since [b(z,w)]j[a(z,w)] is irreducible b(z,w) and a(z,w) have only finitely
many zeroes in common. (If a, b involve powers of z-l and w- 1 these zeroes
are found by multiplying b(z, w) and a(z, w) to make them polynomials and then
finding those common zeroes (zo, wo) of these polynomials for which Zo =1= 0
and Wo =1= O. If a, b involve powers of z-l but not w- 1 just multiply each by an
appropriate power of z and find zeroes (zo, wo) for which Zo =1= 0; similarly if a, b
involve powers of w- 1 but not Z-l ).
Let the common zeroes of a(z, w) and b(z, w) be
(Zl, wIl, (Z2' W2), ... , (zm, wm)

and let the labelling be such that
(Zi, Wi) E V(p, r, q, t) i = 1, ... , n

(Zi' Wi) ~ V(p, r, q, t) i =n+ 1, ... ,m
Note that (0,0) cannot be a common zero of a(z, w), b(z, w). Indeed, the only
time this could happen is when a(z, w) and b(z, w) are polynomials but we have
assumed a(z, w) has non-zero constant term. Now by Lemma 2.6, there exists a
causality cone C~ = Hp',r' n Hfj,r' such that
[Hp,r n Hq,t] C C~
and such that
fori = 1, ... ,n (2.86)

Now it is easy to show that
vpl ,r' ,q',t' C Vp,r,q,t

so from (2.86) we have that
(Zi' wd ¢ V(p', r', q', t') fori = 1, ... ,m
In other words, b(z, w) and a(z, w) satisfy condition (i) of Proposition 2.16 (with
p', r', q', t' in place of p, r, q, t). To satisfy Proposition 2. 16(ii) and (iii) let
q* = q' + p', t* = t' + r'

p* = p' + q*, r* = r' + t*
Then,
p*t* - q*r* = 1,
[Hp,r n Hq,t] C [Hpl,r l n Hq',t'] c [Hp*,r* n Hq*,t*],
and
aq*,t*(z,w) = ap*,r*(z,w) = const. i= 0.

So (i), (ii), and (iii) are satisfied (with p*, r*, q*, t* replacing p, r, q, t) and therefore
[b(z, w )]j[a(z, w)] is stabilizable in Rp*,r* ,q* ,t*. D
NOTE 2.2 The construction in the proof of Theorem 2.7 may not in practice be
the best way offinding p* , r* , q* , t*.
Example 2.4
b(z,w) z +w
-
a(z,w) 1-z+w
The only common zero of a and b is at (~, - ~). Take
p* = t* = 1, q* = 0, r* = 2.
Then (see Figure 2.3), a(z, w) andb(z, w) have no common zeroes in V(1, 2, 0,1)
and so Proposition 2.16 is satisfied (conditions (ii) and (iii) are easily checked).
Hence
z+w
is stabilizable in Rl,2,O,1.
1-z+w
If we wish to stabilize it we first map Rl ,2,01
, to Rl ,001
, , under q,. So
q,( z+w )_q,( z+(wz- 2)z2 )_ a+{3o? _bo(a,{3)
1-z+w - 1-z+(wz-2)z2 -1-a+{3a2 - ao(a,{3)
(1;2, Yz) •
Figure 2.3. The shaded region shows the region {(Izl, Iwl) : (z, w) E V(I, 2, 0, I)}
We then stabilize [bo(a, {1)]/[ao(a, {1)] using the methods of Section 5 to obtain a
causal compensator cl(a,{1). Then 11>-1 : R1,o,O,1 -+ Rl,2,O,1 will map cl(a,{1)
to a c(z, w) E R 1,2,O,1 which will stabilize [b(z, w)]j[a(z, w)].
Example 2.5
b(z,w) Z2 W -2 + zw- 1 + w + 4z _
a(z,w) 1 + zw- 1 + 2w E R1,o,1,1'
a( z, w) and b( z, w) have just one common zero, occurring at (1, - i).

aq,t(z, w) and bq,t(z, w) also have a common zero, occurring at zw- 1 = -1.
Let
PI = P + q = 2, ql = q = 1
rl = t = 1, tl = t = 1.
Then
Now let
p' = PI = 2, q'= ql + 2Pl = 5,

r' = rl = 1, t' = tl + 2rl = 3.
Then
p't' - q'r' = 1,
and
So
Z2 W -2 + zw- 1 + W + 4z
1 + zw-1 + 2w is stabilizable in R 2 ,1,5,3.
10. Stabilization of MIMO Weakly Causal Systems

The map,
tp : Sp,r,q,t -+ SI,O,O,1
induces a map
. p,r,q,t -+
tp . smxk sr;n.Xk
1,0,0,1
where for
H= [hij]mxk E S;::r~:'t
<P(H) = [<P(hij)]mxk E srO~t.l (2.87)
tp also restricts obviously to mappings
from Rmxk to R mxk

~"p,r,q,t 1,0,0,1'
from -=1nR xk
p,r,q,t to ""l'r1,0,0,1'
Xk
and
firom Pp,r,q,t
mxk t pmxk
0 1,0,0,1·
Let
P(z, W) E R;;r~:'t
P = fpij]mxk
74 MULTIDIMENSIONAL SYSTEMS THEORYAND APPLICATIONS
where
.. ( ) _ bij(Z, w) (2.88)
PI, Z, W - ( ) E Rp,r,q,t
aij z,w
is irreducible i = 1, ... , m, j = 1, ... , k and

bij(Z, w), aij (z, w) E Pp,r,q,t (2.89)
with aij(z, w) having non-zero constant tenn i = 1, ... , m,j = 1, ... , k. Let
¢(aij(Z, w)) = ~j(a, f3) (2.90)

¢(bij(Z, w)) = nij(a, f3) i = 1, ... , m, j = 1, ... , k
Then
[ nij( a, f3)]
q,(P(z, w)) = H ( a, f3) = ~j(a, f3) (2.91)
where
nij(Z, w)
is irreducible i= 1, ... ,m, j = 1, ... ,k (2.92)
dij(Z, w)
DEFINITION 2.12 A representation of P of the form P = A-I B where

and
is called a left matrix fraction description (IMFD) of P over Pp,r,q,t.
Similarly one can define an RMFD for P over Pp,r,q,t.
DEFINITION 2.13 An IMFD P = A-I B of P over Pp,r,q,t is called irreducible if

any left common factor of A and B over Pp,r,q,t has as determinant a constant. A
left common factor of A and B over Pp,r,q,t is a matrix L E P;:"~qr:: such that
B = LBI for some

AI E P mxm
p,r,q,t
Because of the ring isomorphism properties of q, Rp,r,q,t -+ RI,O,O,I the
correspondence
defines a bijective correspondence between (irreducible) LMFD's of P over Pp,r,q,t

and (irreducible) LMFD's of H= q,(P) over PI,O,O,I'
Since from (2.89) we are assuming that the aij(z, w)'s have non-zero constant
term, then any irreducible representation A-I B will have the property that
det(A)(z, w) (2.94)
has a non-zero constant term. It then follows (see proof of Proposition 2.2 and
Corollary 2.1) that if
A E R mxm
p,r,q,t
n Rp',r',q',t'
mxm
and
B E R p,r,q,t
kxm n R p'kxm
,r' ,q' ,t' th
en
FACT 2.1 A-IBis irreducible in Rp,r,q,t ifandonlyifitisirreducible in By ,r' ,q',#!'
We have, as in the causal case, that if A-I B is an irreducible LMFD of P over

Pp,r,q,t then
FACT 2.2 P E K;,r~:,t if and only if all entries of B have zero constant term.
Due to the isomorphism «P, various other properties ofLMFD's over Pt,O,O,1 go
through for LMFD's over Pp,r,q,t but these will not be of interest to us here.
Matrix fraction descriptions of elements of R;:-r~:'t and K;,~:,t allow us very
easily to derive the matrix counterparts of the scalar results of Section 9.
Let
-=mxk
P E Rp,r,q,t
and let
p=AL1 BL
be an irreducible LMFD of P over Pp,r,q,t. Then (2.94) and Fact 2.1 imply that
det(A L ) has a non-zero constant term and

the entries of B L have zero constant term
If d(z, w) is an mth orderminorof[AL B L] we denote by dp,r(z, w) that portion

of d(z, w) involving only powers of wpz- r (induding the constant term) and by
dq,t(z, w) that portion of d(z, w) involving only powers of ztw-q (induding the
constant term).
Then the matrix counterpart of Proposition 2.15 is
PROPOSITION 2.17 P = ALl BL satisfying Fact 2.1 and Fact 2.2 is stabilizable
by a compensator in R;,~,r;,t if and only if
(i) the mth order minors offAL BL] have no common zeroes in V(P, r, q, t)
(ii) the polynomials in ztw-q,d~:l(z,w), have no common zero in Iztw-ql ::;

1 (where d(i) (z, w) i = 1, ... , ( m.:, k ) are the mth order minors of
[AL BL])'
(iii) the polynomials in wpz- r , d~,~(z, w), have no common zero in IwPz-rl ::; 1
Proof. Follows from the ring isomorphism q, : Rp,r,q,t -+ Rl,O,O,l and the results
of Section 5. D
The matrix version of Theorem 2.7 is
THEOREM 2.8 Let

-1 -
P = AL BL E Rp,r,q,t
satisfy Fact 2.1 and Fact 2.2. Then there exists a causality cone
c; = Hp*,r* n Hq*,t*
such that
Hp,r n Hq,t C C;
and such that P is stabilizable by means of a compensator
C(z, w) E R!~;!! ,q* ,t*
if and only if the mth order minors offAL BL] have no common zeroes on T2.
Proof. A common zero (zo, wo) E T2 of the mth order minors would imply
that whatever causality cone we choose, (q,(A»-lq,(B) will be an irreducible
representation of q,(P) such that mth order minors of
have a common zero on T2 and q,(P) is therefore not stabilizable by a causal

system (Lemma 2.5).
Conversely, because ALI BL is irreducible, the mth order minors of [AL BL]
have only a finite number of common zeroes which, by hypothesis, do not lie on
T2. Also as in the scalar case, (0,0) cannot be one of these common zeroes. We
can then (following the same proof as in the scalar case) find p*, r*, q*, t* such
that Hp,r n Hq,t C Hp. ,r* n Hq*,t* and such that (i), (ii), and (iii) of (2.17) are
satisfied. As in the scalar case once we have p* , r* , q* , t* as in the above theorem,
the compensators can be obtained by mapping R p*,r* ,q* ,t* to Rl,o,O,l using the
causal techniques of Section 5, and then mapping back again. D
11. Conclusions
The general problem investigated in this chapter is the structural stability of 2-D
feedback systems where the plant and compensator each correspond to discrete 2-D
causal or weakly causal multi-inputlmulti-output systems. In particular, necessary
and sufficient conditions are obtained for an unstable plant to be stabilizable and
a classification of the stabilizing compensators is given. A compensator required
to stabilize an unstable plant (provided of course this plant is stabilizable), may
be constructed using the techniques described in Chapter 4. Also, the contents of
Chapter 3 dealing with a derivation of sufficient conditions for the solution of a
system of linear equations over a bivariate polynomial ring are useful in the design
of compensators to provide 2-D feedback systems with certain desirable properties.
This chapter should motivate further research, especially because multidimensional
feedback systems have been proposed for various purposes like iterative image
processing and image restoration [129], [130]. Such image processing systems
that contain feedback loops are sometimes known to oscillate in space and time
[130] and these undesirable oscillations can only be avoided if proper stability
conditions are imposed on the feedback systems. In the design of 2-D feedback
systems, the possibility of incorporating the weakly causal property in the design
of recursive compensators broadens considerably the class of plants that can be
stabilized (see Examples (2.3) and (2.4) in this chapter).
12. Updates
The extension to the general case of an integral domain of the earlier results over
Bezout domains [131] and the work presented in this chapter was reported in 1992
[132] for the case of a scalar feedback system. Consider the closed loop scalar
feedback system in Figure 2.1 of this chapter but consider now the n-D instead of
the two-dimensional case. The closed loop feedback system is structually stable if
and only if each entry of the transfer matrix in Equation 2.34is holomorphic in fr .
For a multivariate rational plant p(z) = ~f:~, where J(z) and g(z) are relatively
prime n-variate polynomials in z = (Zl' Z2, ... ,zn), the rational compensator
c(z) = Zf:~ stabilizes the system if and only if Jh + gk is zero-free in fr. In
[132], as a generalization of the work presented in this chapter, it was shown
that a necessary and sufficient condition for the existence of a rational stabilizing
compensator c( z) is thatllz) and g( z) have no common zeros in Un. The condition
V(f(z)) n V(g(z)) n U = 0 implies that around any point in fr, a unit is
contained in the ideal generated by J(z) and g(z). Then technical devices in the
theory of analytic functions in several complex variables ensures that this property
extends globally over fr. The plantp(z) is then stabilizable and, furthermore, it
was shown that if the plant has real coefficients, then the controller also has real
coefficients [132, Theorem 3.1.21]. The compensator can be constructed by using

Grobner bases that can solve for k(z) and h(z) in, by Hilberts' Nullstellensatz,
k{z)g(z) + h{z)J(z) = (1fJ{z))N,

where 1fJ{z) may be calculated for zero-dimensional algebraic varieties by using the
procedure delineated in this chapter for the 2-D case. Constructive and effective
algorithms for constructing a stabilizing compensator for a MIMO 2-D plant, using
Grobner bases has also been given in [133], following the work described here and
reported, first, in 1985. Research counterparts in the n-D case concerned with
stabilizability and stabilizability margin are worth pursuing.
If the compensator itself is constrained to be stable, then the plant p{z) is said
to be strongly stabilizable . A topological criterion to determine the existence of
a stable complex compensator was advanced in [134], where it was shown that
feedback stabilizability entails overcoming an obstruction that does not always
vanish. The obstruction to the stabilizability of an n-D transfer function p(z) =
un
~f:~, J, g, relatively prime in the subset of given by V(f) n V{g) n un, where
V(f) denotes the algebraic variety of J, i.e. V(f) = {z E ClIJ{z) = O}. The
further condition for h(z) and k(z) to have real coefficients when J(z) and g(z)
have real coefficients was investigated in [135], where a computational method
(based on quantifier elimination by cylindrical algebraic decomposition [5, Chapter
2]) for testing the strong stabilizability conditions was given. From the construction
procedure for the compensator for the 2-D case, described in this chapter, it can be
inferred that a real coefficient compensator can be constructed for a real coefficient
plant provided the relevant condition is satisfied. Necessary conditions for strong
stabilizability for MIMO LSI multidimensional plants were presented in [136].
For the MIMO case, it has been shown in this chapter that for the 2-D normal
full rank case, if Di1NL is a minor coprime left MFD (maximal order minors of
[DL NL] have no nontrivial common factor) of an unstable plantP, andXRYi1 is
a minor coprime right MFD (maximal order minors of [Xh YA]t have no common
factor) of a compensator C, then C stabilizes P if and only if det(DL YR + N LXR)
is zero-free in U 2 • See Chapter 5 for n-D generalizations of these results as well
as difficulties encountered.
Chapter 3
THE EQUATION Ax = b OVER THE RING C[z, w]

BY J. P. GUIVER
1. Introduction
In Chapter 2 we saw that a strictly causal MIMO system Di 1 N L was stabilizable
by a causal compensator XRYi1 if and only if
det(DLYR + NLXR) E lRs [z, w].

DLYR + NLXR represented in some sense the 'denominator' of the feedback
system (see (2.65) to (2.68 ».
It is therefore of interest (see also Emre [137]) to
study the equation DLYR + NLXR = q, where q, is a polynomial matrix and we
look for a polynomial solution YR, XR. This equation can be written as
where
and
B= [l~].
In turn, AB = q, can be studied as m equations of the form
Ae = <Pi
where <Pi is the ith column of q, and q, has m columns.
2. Sufficient Condition For Solution

Let
A E cmXl[z, w] (3.1)
79
and
bE cnXl[z,w]. (3.2)
We want to know if there exists x E ct Xl [z, w] such that
Ax=b. (3.3)
We note here that if A E ]Rmxi[z, w] and b E ~xl [z, w] and if Ax = b has
a solution with x E ct Xl [z, w], then by taking the real parts of each side of the
equation it follows that (3.3) has a solution in ~Xl[z,w].
Let D be an integral domain and let K be its ring of quotients. Then for any
maximal ideal M of D, its localization with respect to M is defined to be the ring
DM = {~ E K : fED, 9 E D\M} . (3.4)
For the time being we consider a more general situation to (3.3), namely:
Ax = b A E D mxi , b E D mx 1 , m < t. (3.5)
The following result is not difficult to prove (see Gustafson [138, Lemma 1]).
LEMMA 3.1 Equation (3.5) has a solution x E D ixl ifand only ifithas a solution
over D M for every maximal ideal M of D.
LEMMA 3.2 Assume A in (3.5) has full rank when considered as a matrix over
K. In other words A has a non-zero mth order minor. Let leD be the ideal
generated by the mth order minors of A. Let
1M ~ {~E K : f EI, 9 ¢ M}. (3.6)
Then 1M is an ideal in the ring D M.

LEMMA 3.3 Equation (3.5) has a solution over D M if and only if it has a solution
overDM/IM.
Before proving the preceding lemma, we prove two preliminary results, the first
of which appears in (Youla-Gnavi [40]), but is proved differently here. The proof
below essentially validates Propostion 2.11, stated in Chapter 2.
LEMMA 3.4 Let A be an m x i matrix with entries in D, m ~ i. Let
a' . =A(l ... m)l
ZI···Im •.
0Z •••
•
om
lA [i.1 . .. im]
. denotes the submatrix of a matrix A obtained by taking rows iI, ... , im and columns
31 ... 3m
jl, .. ·,jm·
A (i,t31 ...... ~m)
3m
denotes the determinant of A [i.31
1 ...
'"
i.
3m
m].
The Equation Ax = b over the Ring C[z, w] 81
and let I be the ideal of D generated by the ail ... im 'so

Then, given any a E I, there exists an l x m matrix B with entries in D such that
AB = aIm·
Proof. For each sequence 1 ~ i l < i2 < ... < im ~ l we construct an l x m

matrix Bil ... i m as follows:
In rows il, ... , im of Bil ...i m we place the m rows of the adjoint of the matrix
A [~l
Jl
...
...
i.m ].
Jm
In the remaining rows we place zeros. It is clear then that
. =a·Zi ...~m
AB·Zi ···~m . 1m·
Now if a E I then 3Cil ...im ED for each sequence 1 ~ il < i2 < ... < im ~ i
such that
a = L Ci1 ...imail ...i m ·
l-::;il <... <im-::;l
Let
B = L Ci1 ...imBit ...i m ·
l-::;il <... <im-::;l
Then
AB L...J r~
A ('" .)
. B-H ···~m
"'l···~m
'L...J . AB·Zi . ··~m

" c·Zi·· ·~m .
' " c·Zi···~m
L...J . 1m
. a·~l···~m
COROLLARY 3.1 Let A, D and I be as in Lemma (3.4). Then if <P is any m x k

matrix with entries in I there exists an i x k matrix B with entries in D such that
AB = CPo
Proof. Let Eij denotes the m x k matrix with zeroes everywhere except in the
(i, j) position where there is a one.
Let Bij be an i x m matrix with entries in D such that
ABij = <Pij 1m
where l/>ij is the (i, j) th entry of q,.

Then
Let
B = LBijEij
i,j
Then
AB = LABijEij
i,j
o
Proof.(OfLemma 3.3): Suppose Equation (3.5) has a solution over nM /IM then
there exists x E n~l and y E .Pz:}Xl such that
Ax=b+y.
Multiplying both sides by a common denominator of x and y gives us
Ax· = gb + y.
where
x· E nix!, y. E 1 mxl and 9 E n\M.
Now from Corollary 3.1, there exists x E n ixl such that

Ax = y •.
Therefore
A(x· - x) = gb
where
x E nix 1 ft M
.
and 9
.
x· -
So
nM
~ ~
AX -x =b and x - X iXl
---E '
9 9
The proof the other way is trivial. o
Now let us return to Equation (3.3), i.e. when n = qz, w]. We assume, in
addition, the following.
ASSUMPTION 3.1 The order minors are relatively prime. 2 This in particular
mth
implies that they have only finitely many zeroes in common.
2This assumption holds when the composite matrix A = [DL N L 1with Dr; 1 N L irreducible
The Equation Ax = b over the Ring C{z, w] 83
Denote the (non-zero) mth order minors by h, ... , Ir and let the common zeroes
of h, ... , Ir be PI,···, Pk where
i = 1, ... ,k.
Now it is well known (see for example Fulton [139]) that the maximal ideals of
C{z, w] are precisely those ofthe form M = (z - a, w - b) for some (a, b) E c2.
Suppose P = (a, b) is not one of the Pi'S. Form the polynomial
s t
h(z, w) = II (z - al-'J II (w - b/l j ) (3.7)
i=l j=l
where
{Ill, ... ,Ils, VI, ... , vt} = {1, ... , k}
and
i = 1, ... , s
j = 1, ... ,to
Then h(a, b) # 0 but h(ai' bi) = 0 i = 1, ... ,k.

Therefore
hd E I for some positive integer d
by Hilbert's Nullstellensatz. So we have that
hdb E I mxl
and hence by Corollary 3.1 there exists
x E c!x 1 [z, w] such that Ax = hdb.
Therefore Axl hd = b is a solution of Equation (3.3) over C{z, W]M (since

h(a, b) # 0). Therefore, Equation (3.3) has a solution over C{z, W]M for every
M = (z - a, W - b) such that (a, b) is not one ofthe Pi's.
NOTATION 3.1 We let Mi denote the maximal ideal (z - ai, W - bi ) ofC{z, w].
Let Di denote the localization ofC{z, w] with respect to Mi, and let Ii = IMi (see
Equation (3.6)).
Then the above discussion together with Lemmas 3.1 and 3.3 imply the following
lemma.
LEMMA 3.5 Equation (3.3), where A satisfies Lemma 3.2 and Assumption 3.1,
has a solution over C{z, w] if and only if it has a solution over Dil Ii for each
i = 1, ... ,k.
C is embedded naturally in Dil Ii and Dil Ii is a vector space over C (as well as
having a ring structure) which is finite dimensional (see Appendix A). The easiest
case to deal with is when Dil Ii ~ C, i = 1, ... ,k; here the isomorphism is given
by
(3.8)
and the condition Dil Ii c, i = 1, ... , k is equivalent to I = Rad(I) (see

f'V
Appendix A). Lemma 3.5 and Equation (3.8) therefore lead to the following con-
dition.
FACT 3.1 If I = Rad(I), then Equation (3.3) (with A satisfying Lemma 3.2 and
Assumption 3.1) has a solution over C[z, w] if and only if
Rank A = Rank [A b] when evaluated at (ai, bi) i = 1, ... , k.

In general Dilli is not a field but is a Noetherian full quotient ring.
DEFINITION 3.1 (see Ching [140]) A commutative ring with identity is said to
be a Noetherian full quotient ring if it is Noetherian with the property that every
element is .either a unit or a zero divisor.
LEMMA 3.6 Dil Ii is a Noetherian full quotient ring.
Proof. Iff I 9 + Ii E Dil h g( ai, bi) # 0, is such that f (ai, bi) # 0, then 9I f + Ii
is the inverse and fig + Ii is a unit.
Suppose fig + Ii E Dilh g(ai' bi) # 0, is such that f(ai, bi) = O.
Let s be the smallest integer such that there exists q E C[z, w] with q(ai, bi) # 0
and qr E I. Certainly a finite minimum 8 does exist since if we choose h such
that h(aj, bj) = 0, j # i, h(ai' bi) # 0 (in a similar manner to the construction
in Equation (3.7)) then by Hilbert's Nullstellensatz hd fd E I for some d.
Let q be such that qr E I.
Then
(qr- 1 +Ii) (~+ Ii) = q:S Eli (8) 0 since 1 ¢ Ii).
But qr- 1 is not in Ii, for if it were, qjB-l = qI/q2 where q2(ai, bi) # 0 and
ql E I; therefore, q2qjB-l E I contradicting the minimality of s.
So qr- 1 + Ii is non-zero element of Dil Ii annihilating fig + Ii, and fig + Ii
is a zero divisor.
Finally we observe that Dil Ii is Noetherian since the Noetherian property is
closed under the operations of localization and formation or factor rings. D
An important property of Noetherian full quotient rings is the following:
The Equation Ax = b over the Ring q z, w] 85
THEOREM 3.1 (see Ching [140J, Kaplansky [141, Theorem 82]). Afmite set of
elements in a Noetherian full quotient ring either contains only zero divisors in
which case these have a common annihilator, or contains a unit and therefore
generates the whole ring.
Ching [140, Theorem 6] presents a sufficient condition for solution of a linear
equation over a Noetherian full quotient ring in terms of McCoy rank.
DEFINITION 3.2 Let R be a commutative ring with identity. Let A be a matrix

with entries in R. The McCoy rank of A denoted by rankM A is the greatest integer
r such that the rth order minors of A do not have a common annihilator. RankM A
is defined to be zero if all the entries have a common annihilator.
If R is a Noetherian full quotient ring, Theorem 3.1 allows us to describe McCoy
rank in a simpler manner; namely rankM A is the maximum integer r such that there
is an rth order minor of A which is a unit. We therefore have:
FACT 3.2 For A E cmxl[z, w], rankMAoverthe ring DdIi isjustthe usual rank
of A over C when A is evaluated at (ai, bi).
We let Pi denote the usual rank of A at (ai, bi), i = 1, ... , k. Ching's condition,
mentioned above, is as follows:
FACT 3.3 Let A E Rmxl, bE R mx1 where R is a Noetherian full quotient ring
and A is not offull McCoy rank over R. Then Ax = b has a solution over R if
those (rankMA + 1)th order minors of the matrix [A b] which are not minors of
A (i.e. which involve the column vector b) are all zero in R.
Note that if A has full McCoy rank over R then it has an m th order minor which
is a unit in R (when R is Noetherian full quotient) and therefore Ax = b will have
a solution in R. As an immediate consequence we get the following theorem.
THEOREM 3.2 Let A E cmxl[z, w] satisfy Lemma 3.2 and Assumption 3.1. Let
Pi, i = 1, ... , k be as in Fact 3.2. Then a sufficient condition for Equation (3.3) to
have a solution in ct x 1 [z, w] is that for i = 1, ... , k, each (Pi + 1) th order minor
of [A b) involving b lies in h See Appendix A for the problem of determining
when a given polynomial is in h
COROLLARY 3.2 Let A, Pi be as in the above theorem. Let P = min(Pl, ... ,Pk}
Then Equation (3.3) has a solution ifeach (p+ 1)th order minor of [A b] involving
b lies in I.
Proof. The (p + 1)th order minors involving b will in particularly lie in Ii (in
fact they lie in Ii for i = 1, ... , k, if and only if they lie in I - see Corollary A2).
Therefore so will the (Pi + 1)th order minors involving b. Now use the previous
theorem. D
3. Appendix A. Zero-Dimensional Polynomial Ideals
We recall in this appendix a few facts from the theory of polynomial ideals (see
for example Van der Waerden [142] and Fulton [139]). For simplicity we work
over C although any algebraically closed field would do.
LetIbe an ideal of the polynomial ring C(ZI,'" ,zn). We let V(I), called the it
variety denote the set {(al,"" an) E en : I(al,"" an) = 0, I E I}. V(I) is
called the zero-set of I. Hilbert's basis theorem tells us that any polynomial ideal
is finitely generated, so V(I) can be characterized as the set of common zeroes of
a finite set of generating polynomials of I. Given It, ... ,IT E C[ZI' ... ,Zn), we
denote by (It, ... , In) the ideal in C[ZI , ... , zn) generated by the Ii'S.
Given an ideal I in any ring R, the radical of I is defined as Rad( I) = {I E
R : In E I for some positive integer n}. An ideal is called a radical ideal if I =
Rad(I).
Given a set X c en we denote by I( X) the ideal of polynomials in C[ZI , ... , zn)
which are zero at each point of X. Clearly I(X) is a radical ideal for any
X c en. If I is any ideal of C[Zl , ... , zn) it is also clear that Rad( I) C I (V (I) ).
Hilbert's Nullstellensatz gives us the inclusion the other way, which then implies
that Rad(I) = I(V(I)); the fact that C is algebraically closed is essential here.
. In fact, given a polynomial ideal I = (It, ... , In) C C[Zl , ... , zn), there exists
a number d such that if I E I(V(I)), then Id E I; in other words the exponent d
depends only on I and not on the particular polynomial I E I(V(I)).
Closely related to Hilbert's Nullstellensatz is the fact that the maximal ide-
als of crZI,'" ,zn] are precisely those ideals of the form M = (Zl - aI, Z2 -
a2,.·., Zn - an) where ai E C, i = 1, ... , n. Given such an ideal we can define
the localization of C[ZI' ... ,zn) with respect to M to be the ring DM = {I / 9 E
C(ZI,"" zn) : g(al, ... , an) = O}. Let 1M denote the ideal of DM defined by
1M = {J /g E C(ZI,"" zn) : g(al, ... , an) # 0, I E I}.
From now on until the end of the Appendix we will assume I is an ideal of
C[ZI,"" zn) such that V(I) contains only a finite non-zero number of points;
such an ideal is called zero-dimensional. Label these points PI, ... ,Pk where
Pi = (a~i), ... , a~)), i = 1, ... , k. Let Mi denote the maximal ideal (Zl -
a~i), ... , Zn -a~)) andletDi and Ii denoteDMi andIMi respectively, i = 1, ... ,k.
Then Fulton [139, Chapter 2, Proposition 6] presents the following result.
(AI) PROPOSITION. There is a natural isomorphism from C(ZI' ... ,zn)/I onto
the direct product Xf=l Dd h
(A2) COROLLARY. f E I if and only if f E 1;, i = 1, ... , k.

The field C is naturally embedded in Di/ Ii and Dd Ii is in fact finite dimen-
sional as a vector space over C because (see Fulton [139, Chapter 1, corollary 4])
C[ZI' ... ,zn)/ I is finite dimensional when V(I) is finite. For each i, i = 1, ... ,k,
The Equation Ax = b over the Ring C[ z, w] 87
there exists a minimum positive integer di such that h E Ii for any monomial h in
ZI - a~i), ... , Zn - a~) of degree di or greater. This follows from Hilbert's Null-
stellensatz; for let 1 E C[ZI"'" zn] be such that 1{pj) = 0, j i= i, 1{Pi) i= 0
(see for example the construction in (3.13», then {Zj - a)i))d 1d E I, where d is the
exponent in Hilbert's Nullstellensatz, and therefore {Zj _a)i))d E h j = 1, ... , n.
Consequently any monomial in the (Zj - a)i)'s of degree greater than n(d -1)
is in Ii which gives us an upper bound for di. Dil Ii as a vector space is then a
subspace of the space of polynomial in ZI - a~i), ... , Zn - a~) (strictly speaking
residue classes of polynomials mod Ii) of degree less than di. The simplest case
is when di = 1 in which case Dil Ii is isomorphic to C because every element
1I 9 + Ii E Dil Ii is equal to a constant mod h the isomorphism is given by
1 19 + Ii -+ [1 (Pi)]/[g(Pi)] because if c E C is such that 1 19 - c E Ii then 1 19 - c
evaluated at Pi must be zero. So if ~ = 1, (J 19) E Ii if and only if 1(Pi) = O.
Consequently, if di = 1, i = 1, ... , k, then, for 1 E C[ZI, ... , zn], 1 E I if and
only if f E Ii for each i which is true if and only if f (Pi) = 0 for each i. Therefore
if di = 1, i = 1, ... , k, then RadI = I; it is also clear that the converse is true
since if RadI = I it is easy to see that for each i, j, i = 1, ... ,k, j = 1, ... , n,
(Zj - ay)) E Ii and therefore di = 1.
A slightly different approach makes use of the primary decomposition theorem
for ideals in Noetherian rings. In particular, if I C C[ZI"'" zn] is such that
V(I) is finite then I = Ql n Q2 n ... n Qk where Qi is the ideal M[i + I, Mi
is the maximal ideal (ZI - ali), ... , Zn - a~)) and (Ji is the smallest number (J
such that Mi C Mt+ 1 + Ii (see van der Waerden [142, 7.6, § 96] for this and
for what follows). The condition that a given polynomial 1 is in I if (and only
if) f E Mti + Ii for i = 1, ... , k is known as the fundamental theorem of Max
Noether.
It is not difficult to show that (Ji is the same as the di discussed earlier in
the appendix. It is also possible to construct polynomials hI"'" hk in ZI -
a~i) , ... ,Zn - a~) each of whose degrees is less than (Ji such that f E Qi if and
only if f, when expanded in a Taylor series about the point (ali) , ... ,a~)) and after
omitting all terms of degree ~ (Ji, is equal to a linear combination of hI"'" hk
with constant coefficients. This is also equivalent to having a basis for those in
ZI - ali), ... ,Zn - a~) of degree less than di {= (J) which belong to Ii; in fact
Qi = Ii n C[Zl' ... ,zn] - this follows, for example, form [143, § 2.6 Cor. 2] and
the easily seen fact that DiQi = h although it is not difficult to prove from first
principles. So we can actually establish whether or not a given polynomial is in
Qi (or equivalently Ii).
The simplest case that arises is when (Ji = 1{= di), i = 1, ... , k. As we have
seen earlier this occurs if and only if Dd Ii ~ C i = 1, ... , k which in turn is
equivalent to RadI = I. This case is also considered by Van der Waerden as an
application of Noether's theorem. He shows that O"i = 1, i = 1, ... , k occurs

if the linear terms of It, ... , !r when expanded in a Taylor Series about Pi span
the whole n dimensional space of linear terms in Zl -
a~i), ... ,Zn - a~); the
converse of this statement is also seen to be true because O"i = 1 ::::} Zj - a;i) E
h j = 1, ... ,n ::::} (Zj - ay»)g = L:=l
gsl8 for some g, g8 E qZl, ... ,zn]
with g(Pi) 1= O. Consequently we have the following result:
(A3) PROPOSITION. Let I = (It, ... , !r), Pi = (a~i), ... , a~») be as above.
Then the following are equivalent:
(i) Dd1i ~ C i = 1, ... ,k

(ii) Rad I = I
(1'1'1') J( a (i) (i»).IS 0 f rank n,

l , ... , an i = 1, ... , k where J(Zl,"" zn) is the
matrix
8z1
811 '" !!h...]
8z
...
n
· ..
[
·· .
8fr ... 8fr
8z1 8z;;
We also have in the above Proposition that (i) is equivalent to (iii) for each i
separately.
Chapter 4
GROBNER BASES: AN ALGORITHMIC METHOD

IN POLYNOMIAL IDEAL THEORY
BY B. BUCHBERGER
1. Introduction
Problems connected with ideals generated by finite sets F of multivariate poly-

nomials occur, as mathematical subproblems, in various branches of systems the-
ory, see, for example, [5]. The method of Grobner bases is a technique that provides
algorithmic solutions to a variety of such problems, for instance, exact solutions of
F viewed as a system of algebraic equations, computations in the residue class ring
modulo the ideal generated by F, decision about various properties of the ideal
generated by F, polynomial solution of the linear homogeneous equation with
coefficients in F, word problems modulo ideals and in commutative semi groups
(reversible Petri nets), the bijective enumeration of all polynomial ideal over a
given coefficient domain etc. .
For many years, the work of G. Hermann [144] was the only algorithmic method
for tackling problems in polynomial ideal theory. Still, her paper is a rich source.
However, as pointed out in [145] and [146], the solution of her main problem
"is a multivariate polynomial f in the ideal generated by F?" does not yet give a
feasible solution to the "simplification problem modulo an ideal" (i.e. the problem
of finding unique representatives in the residue classes modulo the ideal) and to
the problem of effectively computing in the residue class ring modulo an ideal.
The method of Grobner bases, as its central objective, solves the simplification
problem for polynomial ideals and, on this basis, gives easy solutions to a large
number of other algorithmic problems including Hermann's original membership
problem. Also, when compared with Hermann's algorithms, our algorithm that
constructs Grobner bases is of striking simplicity and, depending on the example
considered, may get through with intermediate computations using polynomials
of relatively low degree. On the other hand, as shown in [147] and [148], the
decision of polynomial ideal congruence intrinsically is a complex problem. In
89
the worst case, therefore, also the method of Grobner bases may lead to exploding
computations. Much work is going on to analyze and predict these phenomena
and to extend the applicability of the method.
The method of Grobner bases was introduced in 1965 by this author in [149],
[150] and, starting from 1976, was further refined, generalized, applied and an-
alyzed in a number of papers [151]-[152]. The basic idea of the method is the
transformation of the given set of polynomials F into a certain standard form G,
for which in [151] the author introduced the name 'Grobner bases', because Prof.
W. Grobner, the thesis advisor of [149] stimulated the research on the subject by
asking how a multiplication table for the associative algebra, which is formed by
the residue ring modulo a polynomial ideal, can be constructed algorithmically
and by presenting a first sketch of an algorithm: He proposed to 'complete' the
basis F by adjoining the differences of different representations of power products
(modulo the ideal). This, however, is no finite procedure. It was the author's main
contribution to see and prove in [149], [150] that it suffices to adjoin the differences
of (the reduced forms of) certain 'critical pairs' (or, equivalently, the reduced form
of'S-polynomials' [149]), which are finite in number.
A concept analogous to the concept of 'Grobner bases', with the name 'standard
bases', for the domain of formal power series was also introduced in Hironaka's
famous paper [153]. However, Hironaka only gave a nonconstructive existence
proof for these bases, whereas in [149], together with the concept of Grobner bases,
we also presented an algorithm for constructing the bases and only this algorithm
allows an algorithmic solution to the various problems shortly mentioned above.
A nonconstructive existence proof for Grobner bases may also be found in [154].
Hilbert's basis theorem, then, follows as a corollary.
Later (1967) the two basic ideas of our method, critical pairs and completion,
were also proposed by Knuth and Bendix [155] in the more general context of
equations between first order terms. The Knuth-Bendix algorithm now plays an
important role in various branches of computer science (abstract data type transfor-
mations, equational theorem proving and applications in automated program ver-
ification). Recently, the Knuth-Bendix algorithm and the author's own algorithm
for constructing Grobner bases were brought together under a common algorithm
structure by R. Llopis de Trias [156] and, independently, by P. Le Chenadec [157];
see also [145] for a general introduction to the "critical-pair completion" algorithm
type. On the other hand, the improvements of the author's algorithm were carried
over to the Knuth-Bendix algorithm, see [155]. A lot of challenging questions
remain to be treated, which, in the future, might also affect systems theory (for ex-
ample, decision methods for boolean algebra based on the critical-pair/completion
approach, see [158]).
In the present chapter, a survey on the method of Grobner bases is given. In
Section 2, the concept of Grobner bases is defined and, in Section 3, the basic
form of the algorithm for constructing Grobner bases is described. In Section 4 an
improved version of the algorithm is presented. The improvements are important
Grobner Bases: An Algorithmic method in Polynomial Ideal Theory 91
for the practical feasibility of the computations. In Section 5, the algorithm is

applied to the simplification problem, the congruence problem and related prob-
lems in polynomial ideal theory. In Section 6, the algorithm is applied to the exact
solution of systems of algebraic equations and related problems. In Section 7, it is
demonstrated that the S-polynomials have also a significance as the generators of
the module of solutions for linear homogeneous equations with polynomial coef-
ficients and an algorithm for a systematic solution of such equations is presented.
Grobner bases for polynomial ideals with integer coefficients are treated in Sec-
tion 8. Some other applications are summarized in Section 9. Finally, in Section
to, some remarks about specializations, generalizations, implementations and the
computational complexity of the algorithm are made.
The emphasis of this paper is on explicit formulation of algorithms (in an easy
notation) and on examples. With the exception of some sketches, no proofs of the
underlying theorems can be given. However, complete references to the original
publication are provided.
2. Grobner Bases
Notation
K afield.
K[Xl' ... , xn] ring of n-variate polynomials over K.
The following typed variables will be used:
f, g, h, k,p, q polynomials in K[Xl, ... , xn].
F, G finite subsets of K[Xl, ... , xn].
s, t, u power products of the form xi! ... x~ .
a, b, c, d elements in K.
i, j, I, m natural numbers.
Let F = {ft, ... , fm}. By 'Ideal(F)' we will denote "the ideal generated by F"
(i.e. the set
{L hi· filhi E K[Xl, ... , xn](i = 1, ... , m)}).

l:'S;i:'S;m
Furthermore, we will write 'f =F g' for "f is congruent to 9 modulo Ideal(F) " (i.e.
f - 9 E Ideal(F». Before one can define the notion of Grobner bases the notion of
'reduction' must be introduced. For this it is necessary to fix a total ordering <T
of the power products xi! ... x~n, for example, the 'total degree ordering' (which
is 1 <T X <T Y <T x2 <T xy <T y2 <T x3 <T x2y <T xy2 <T y3 <T ...
in the case of two variables) or the 'purely lexicographical ordering' (which is
1 <T X <T X 2 <T X 3 <T··· Y <T xy <T X 2 Y <T ... <T Y2 <T xy 2 <T···
in the case of two variables). In fact, any total ordering is suitable, which at least
has the following two properties:
(Tl) 1 <T t for all t =1= 1,
(T2) if S <T t then s . u <T t· u.
A total ordering satisfying (TI) and (T2) will be called 'admissible'. For the sequel,
assume that an arbitrary <T has been fixed. With respect to the chosen <T, we
use the following notation.
Notation
Coefficient(g, t) the coefficient of tin g.
LeadingPowerProduct(f) the maximal power product (w.r.t. <T)
occurring with non-zero coefficient in f.
LeadingCoefficient(f) the coefficient of the
LeadingPowerProduct(f)·
DEFINITION 4.1 [149],[150] 9 -+F h (read: 'g reduces to h modulo F') ijfthere
exists f E F, b and u such that
9 -+ /,b,u and h = 9 - b. u . f
g-+/,b,u (read: 'g is reducible using f,b,u') ijfCoefficient(g,u· LeadingPower-
Product( f» =1= 0, b = Coefficient(g, U· LeadingPowerProduct( f» /
LeadingCoefficient( f)
Hence, roughly, 9 reduces to h modulo F iff a monomial in 9 can be deleted by
the subtraction of an appropriate multiple b· u . f of a polynomial f in F yielding
h. Thus, the reduction may be viewed as one step in a generalized division.
Example 4.1 Consider F:= {II, 12, fa}, where

II := 3x2y + 2xy + y + 9x2 + 5x - 3,
12 := 2x3 y - xy - y + 6x 3 - 2x2 - 3x + 3,
fa := x 3 y + x 2y + 3x3 + 2x2.
The polynomials II, 12, fa are ordered according to the purely lexicographical
ordering. The leading power products are x 2y, x 3 y, x 3 y, respectively, and the
leading coefficients are 3, 2, and 1. Consider
9 := 5y2 + 2x2y + {5/2)xy + {3/2)y + 8x2 + {3/2)x - 9/2.

Modulo F, 9 reduces, for example, to
h := 5y2 + {7/6)xy + {5/6)y + 2x2 - {1l/6)x - 5/2,

Grabner Bases: An Algorithmic method in Polynomialldeal Theory 93
namely,
9 -+/,b,'U for f:= f1,b:= 2/3,u:= 1
because Coefficient(g, 1 . x 2 y) = 2 =1= 0, b = Coefficient(g, 1 . x 2 y) I
LeadingCoefficient(j1), and
h = 9 - (2/3) . 1 . h.
DEFINITION 4.2 h is in normal form (or reduced form) modulo F iff there is no
hi such that h -+ F hi.
h is a normal form of 9 modulo F iff there is a sequence of reductions
9 = ko -+ F k1 -+ F k2 -+ F ... -+ F km = h
and h is in normal form modulo F.
An algorithm S is called a normal form algorithm (or simplifier) iff for all F
andg:
S (F, g) is a normal form of 9 modulo F.
LEMMA 4.1 [149J, [151] Thefollowing algorithm is a normalform algorithm:
ALGORITHM 4.1 (h: = NormaIForm(F, g)).
h :=g
while exist f E F, b, u such that h -+ /,b,'U do choose fER, b, u such
that h -+ /,b,'U and u· LeadingPowerProduct( f) is maximal
(w.r.t. <T)
h:= h - b· u· f
The correctness of this algorithm should be clear. For the correctness, the selection
of the maximal product u· LeadingPowerProduct(j) is not mandatory. However,
this choice is of crucial importance for efficiency. The termination of the algorithm
is guaranteed by the following lemma.
LEMMA 4.2 [149J, [151] For all F :-+F is a Noetherian relation (i.e. there is no
infinite sequence ko -+ F k1 -+F k2 -+ F ... ).
Example 4.2 h in the Example 4.1 is in normal form modulo F: no power product
occurring in h is a multiple of the leading power product of one of the polynomials
in F. Thus, no reduction is possible. Another example:
gl can be further reduced:
gl -tit (1/9)xy + (2/9)y - 3x3 + (1/3)x 2 + {19/9)x - 2/3 =: g~.

gi is in normal form modulo F. gi, hence, is a normal form of x 3y modulo F.

Actually, gi may be the result of applying the algorithm 'NormalForm' to x 3 y
(depending on how the instruction 'choose 1 E F, such that .. .' in the algorithm
is implemented). In this example, a second reduction is possible:
x 3 y -+/1 {1/2)xy + {1/2)y - 3x3 + x 2 + {3/2)x - 3/2 =: g2·

g2 is already in normal form modulo F.
From the example one sees that, in general, it is possible that, modulo F, gl and
g2 are normal forms of a polynomial g, but gl = g2. Those sets F, for which such
a situation does not occur, play the crucial role for our approach to an algorithmic
solution of problems in polynomial ideal theory:
DEFINITION 4.3 [149], [151] F is called a Grobner basis (or Grobner set) ijffor
all g, hI, h 2:
if hI and h2 are normal forms of 9 modulo F then hI = h 2.
It is the central theme of this chapter to show that

(a) for those sets F that are Grobner bases, a number of important algorithmic
problems (that are formulated in terms ofIdeal(F» can be solved elegantly and
(b) those sets F, which are not Grobner bases, can be transformed into sets G,
that are Grobner bases and generate the same ideal.
Most of the algorithmic applications of Grobner bases are based on the following
fundamental property of Grobner bases.
THEOREM 4.1 [149], [151], [159](CharacterizationTheoremforGrobnerbases).

Let S be an arbitrary normal form algorithm. The following properties are equiv-
alent:
(GB1) F is a Grobner basis.
(GB2) For all I,g: 1 =F 9 ijfS{F,f) = S{F,g).

(GB1) is also equivalent to:
(GB3) -+F has the 'Church-Rosser' property.
(GB3) links Grobner bases with analogous concepts for equations of first order
terms and the Knuth-Bendix algorithm. For details see [145]. (GB3) is not needed
in this chapter. The following lemma is helpful in establishing this link.
LEMMA 4.3 [159], [160] (Connection between reduction and congruence): For
all F, I,g:
Grobner Bases: An Algorithmic method in Polynomialldeal Theory 95
(Here, Bp is the reflexive, ~ymmetric, transitive closure of -+F i.e.

/ B Piff there exists a sequence
/ = ko B F kl B F k2 B F ... B F k m = g,
where
/ BF g iff(f -+F g or g -+F I).
(GB2) immediately shows that, for Grobner bases F, the decision problem
is algorithmically decidable (uniformly in F). For Grobner bases, other
' / =F g'
computability problems will have similarly easy solutions: see Sections 5-9.
3. Algorithmic Construction of Grobner Bases

Before we give the algorithmic applications of Grobner bases we show how it
may be decided whether a given set F is a Grobner basis and how Grobner bases
may be constructed. For this the notion of an 'S-polynomial' is fundamental:
DEFINITION 4.4 [149J, [150J, [151J The 'S-polynomial corresponding to It, 12,
is
SPolynomial(fl, h): = Ul • It - (Cl/C2) . U2 . 12,
=
where Ci LeadingCoejJicient( h),
Ui is such that 8i . Ui = the least common multiple of 81,82 and
8i = LeadingPowerProduct(fi) (i = 1,2).
Example 4.3 For It, 12 as in Example 4.1, the SPolynomial(lt, h) is
2x 2 y + {5/2)xy + {3/2)y + 8x2 + {3/2)x - 9/2.

Note that the least common multiple of 81 and 82 is the minimal power product
that is reducible both modulo It and modulo h. The algorithmic criterion for
Grobner bases is formulated in the following theorem, which forms the core of the
method:
THEOREM 4.2 (Buchberger [149J, [150J, [151], [159J; Algorithmic Character-

ization of Grobner bases). Let S be an arbitrary normal form algorithm. The
following properties are equivalent:
(GB1) F is a Grobner basis.
(GB4) For all It, h E F: S{F, SPolynomial(lt, 12» =o.

(GB4), indeed, is a decision algorithm for the property 'F is a Grobner ba-
sis': one only has to consider the finitely many pairs It, h of polynomials in F,
compute the corresponding S-polynomials and see whether they reduce to zero by
application of the nonnal fonn algorithm S. In addition, Theorem 4.2 is the basis
for the central Algorithm 4.2 of this paper for solving the following problem.
PROBLEM 4.1 Given F.
Find G, such that Ideal(F) = Ideal(G) and G is a Grobner basis.
ALGORITHM 4.2 (Buchberger [149), [150})jor Problem 4.1
G:=F
B:= {{/I,h}lb,h E G,b =1= h}
while B =1= 0 do
{/I,h}: =apairinB
B : = B - {{b, hH
h : = SPolynomial(b, h)
h' : =NormalForm(G, h)
if h' =1= 0 then
B: = B U /{ {g, h'}lg E G}
G:=GU/{h'}
The partial correctness of this algorithm, essentially, relies on Theorem 4.2.

The termination can be shown in two ways, see [150], [161]. (Sketch of the first
method [161]: One considers the sequence of ideals Ideal(Pd C Ideal(P2 ) c ... ,
where ~ is the set of leading power products of polynomials in Gi and Gi is the
value of G after G has been extended for the i-th time. It is easy to see, that the
inclusions in this sequence are proper. Hence, by Hilbert's theorem on ascending
chains of ideals in K[XI' ... ,xn], see [l6~], the sequence must be finite. Sketch
of the second method [150]: One uses Dickson's lemma [163], which, applied
to the present situation, shows that a sequence tl, t2, ... of power products with
the property that, for all j, tj is not a multiple of any of its predecessors, must be
finite. Actually, if ti is the leading power product of the i-th polynomial adjoined
to G in the course of the algorithm (i = 1,2, ... ), then the sequence tl, t2,'"
has this property and, hence, must be finite. This is the way, the termination of
the algorithm was first proven in [150], where Dickson's lemma was reinvented.
Hilbert's basis theorem can be obtained as a corollary in this approach, see [154].)
Example 4.4 Starting from the set F of Example 4.1, we first choose, for instance,
the pair b, 12 and calculate
SPolynomial{b, h) = 2x2y + {5/2)xy + {3/2)y + 8x2 + {3/2)x - 9/2
Reduction of this polynomial to a reduced fonn yields
(7/6)xy + (5/6)y + 2X2 - (11/6)x - 5/2.
We adjoin this polynomial to G in the fonn
/4 := xy + {5/7)y + {12/7)x2 - {11/7)x - 15/7,
Grabner Bases: An Algorithmic method in Polynomial Ideal Theory 97
where we normalized the leading coefficient to 1. (This normalization is not

mandatory. However, as a matter of computational experience, it may result in
drastic savings in computations over the rationals. Theoretically, this phenomenon
is not yet well understood. Investigations of the kind done for Euclid's algorithm
should be worthwhile, see [164] for a survey on these questions.)
Now we choose, for example, the pair It and 14:
SPolynomial{It,/4) = 1· It - (3/1) . x· /4
= -{1/7)xy + y - {36/7)x 3+ {96/7)x 2 + {80/7)x - 3.
Reduction of this polynomial, by subtraction of -(1/7) ·14 (and normalization),
yields the new polynomial.
15 := y - {14/3)x 3 + {38/3)x 2 + (61/6)x - 3.
Furthermore, SPolynomial(f4, 15) = 1· 14 - {l/l)·x· 15. By subtracting (5/7)· 15

and normalization we obtain
16 := x4 - 2x3 - {15/4)x 2 - (5/4)x.

Finally, the reduction of SPolynomial(lt, /3) = X· II - (3/1) . 1 . /3 leads to
17 := x 3 - {5/2)x 2 - (5/2)x.
The reduction of the S-polynomials of all the remaining pairs yields zero and,
hence, no further polynomials need to be adjoined to the basis. For example,
SPolynomial(f6, h) = {1/2)x 3 - {5/4)x 2 - {5/4)x

reduces to zero by subtraction of 1/217. Hence, a Grobner basis corresponding to
F is
G:= {It, .. · ,h}·
DEFINITION 4.5 [165J F is a reduced Grobner basis iff F is a Grobner basis and
for all I E F : I is in normal form modulo F - {f} and l£adingCoejJicient(j) =
1.
Example 4.5 G in Example 4.4 is not a reduced Grobner basis: For example, II
reduces to zero modulo {h, ... ,h}. By successively reducing all polynomials of
a Grobner basis modulo all the other polynomials in the basis and normalizing the
leading coefficients to 1, one always can transform a Grobner basis into a reduced
Grobner basis for the same ideal. We do not give a formal description of this
procedure, because it will be automatically included in the improved version of
the algorithm below. In the example, also 12, /3, 14, and 16 reduced to zero and
15 reduces to
I~ := y + x 2 - {3/2)x - 3.
Hence, the reduced Grobner basis corresponding to F is
G' := {f~, h} = {y + X2 - {3/2)x - 3, x 3 - {5/2)x 2 - {5/2)x}.

THEOREM 4.3 (Buchberger [165J: Uniqueness of reduced Grobner bases). If
Ideal( F) = Ideal( F' ) and F and F' are both reduced Grobner bases then F = F'.
DEFINITION 4.6 Let GB be the junction that associates with every FaG such
that ldeal(F) = Ideal(G) and G is a reduced Grobner basis.
By what was formulated in Theorems 4.2, 4.3, Algorithm 4.2 and the remarks
in Example 4.5 we, finally, obtain the following main theorem, which summarizes
the basic algorithmic knowledge about Grobner bases.
THEOREM 4.4 (Buchberger 1965, 1970, 1976) G B is an algorithmic junction that

satisfiesfor all F, G:
(SGB1) Ideal(F) = ldeal(GB(F»,
(SGB2) ijldeal(F) = Ideal(G), then GB{F) = GB{G),

(SGB3) GB{F) is a reduced Grobner basis.
4. An Improved Version of the Algorithm

For the tractability of practical examples it is crucial to improve the algorithm.
There are three main possibilities for achieving a computational speed-up:
(1) The order of selection of pairs {iI, h} for which the.S-polynomials are
formed, though logically insignificant, has a crucial influence on the complexity of
the algorithm. As a general rule, pairs whose least common multiple of the leading
power products is minimal with respect to the ordering <T should be treated first.
This, in connection with (2), may drastically reduce the computation time.
(2) Each time a new polynomial is adjoined to the basis, all the other polynomials
may be reduced using also the new polynomial. Thereby, many polynomials in G
may be deleted again. Such reductions may initiate a whole cascade of reductions
and cancellations. Also, if this procedure is carried out systematically in the course
of the algorithm, the final result of the algorithm automatically is a reduced Grobner
basis. The reduction of the polynomials modulo the other polynomials in the basis
should also be performed at the beginning of the algorithm.
(3) Whereas (1) and (2) are strategies that need no new theoretical foundation,
the following approach is based on a refined theoretical result [166], which has
proven useful also in the general context of 'critical-pairlcompletion' algorithms,
in particular for the Knuth-Bendix algorithm: The most expensive operations in the
algorithm are the reductions of the h' modulo G in the while-loop. We developed
Grobner Bases: An Algorithmic method in Polynomialldeal Theory 99
a 'criterion' that, roughly, allows to detect that certain S-polynomials h can be re-
duced to zero, without actually carrying out the reduction. This can result in drastic
savings. Using this criterion, in favorable situations, only 0(1) S -polynomials must
be considered instead of 0(1 2 ), where I is the number of polynomials in the basis.
(Of course, in general, I is dynamically changing and, therefore, the effect of the
criterion is very hard to assess, theoretically).
Strategy 1. was already used in [149], [150]. Also, the correctness of the
reduction and cancellation technique sketched in (2) was already shown in [149],
[150]. The criterion described in (3) was introduced and proven correct in [166],
details of the correctness proof may be found in [167].
Before we give the details of the improved version of the algorithm based on
(1)-(3) we present a rough sketch:
In addition to G and B, we use two sets R and P. R contains polynomials
of G, which can be reduced modulo the other polynomials of G. As long as R
is non-empty, we reduce the polynomials in R and store the resulting reduced
polynomials in P. Only when R is empty, we adjoin the reduced polynomials in
P to G and determine the new pairs in B for which the S-polynomials have to be
considered. If an S-polynomial for a pair in B is reduced with a non-zero result
h', h' is put into P and, again, polynomials in G are sought that are reducible with
respect to h'. Such polynomials are put into R and we continue with the systematic
reduction of R. We now give the details.
PROBLEM 4.2 Given: F.

Find: G, such that Ideal(F) = Ideal(G) and G is a reduced Grobner basis.
ALGORITHM 4.3 (Buchberger [166]) for Problem 4.2

R := Fj P:= 0j G:= 0j B:= 0
Reduce AIl(R, P, G, B); New Basis (P, G, B)
while B f:= 0 do
{h,h}: = a pair in B whose LCM(LP(fd, LP(h))
is minimal w. r. t <T
B:= B - {{1I,hH
if (not Criterionl(II, 12, G, B) and
not Criterion2(II, h» then
h:= NormaIForm(G, SPolynomial(lI, h»
if h f:= 0 then
Go := {g E GILP(h) ~M LP(g)}
R:=Goj P:={h}j G:=G-G o
B:= B - {{II,h}lft E Go or 12 EGo}
ReduceAll(R, P, G, B); NewBasis( P, G, B).
Subalgorithm ReduceAll(transient : R, P, G, B):

while R i= 0 do
h: =an element in R; R:= R - {h};
h : = NormalForm(G UP, h)
ifh i= 0 then
Go := {g E GILP(h) ~M LP(g)}
Po:= {p E PILP(h) ~M LP(P)}
G:=G-G o
P:=P-Po
R:=RUGoUPo
B := B - {{/I, h} E Bib E Go or 12 EGo}
P:= PU {h}.
Subalgorithm NewBasis(transient : P, G, B):

G:=GUP
B:= B U {{g,p}lg E G,p E P,g i= p}
H:=Gj K:=0
while H i= 0 do
h: = an element in Hj H:= H - {h}
k: = NormalForm(G - {h}, h)j K:= K U {k}
G:=K.
Subalgorithm Criterionl( /I, 12, G, B) :{:} there exists apE G such that
b i=P,p i= 12,
LP(P) ~M LCM(LP(ft}, LP(h)),
{/I,p} not in B and {p, h} not in B.
Subalgorithm Criterion2( /I, h) :{:}

LCM(LP(/I) , LP(h)) = LP(/I)· LP(h).
Abbreviations
LP(g) the leading power product of g,
LC M (s, t) the least common multiple of sand t,
s ~M t t is a multiple of s.
The correctness of this improved version of the algorithm is based on the fol-
lowing lemma and theorem.
LEMMA 4.4 [149J, [150J For arbitrary F, /I, 12:

If LP(fd . LP(h) = LCM(LP(/I), LP(h)), then SPolynomial(/I, h) can
always be reduced to zero modulo F.
Grobner Bases: An Algorithmic method in Polynomial/deal Theory 101
THEOREM 4.5 (Buchberger 1979 [166J; detection of unnecessary reductions of

S-polynomials). Let S be an arbitrary normal form algorithm. The following
properties are equivalent:
(GBl) F is a Grobner bavis.
(GB5) For all I, 9 E F there exist hI, h2' ... ,hk E F such that
l=hI,g=hk
LCM(LP(ht} , ... , LP(h k )) '5,M LCM(LP(f) , LP(g)),
S(F, SPolynomial(hi, hHI)) = 0 (for 1 '5, i '5, k).
Lemma 4.4 guarantees that we need not consider the S-polynomial of two
polynomials /I and 12, whose leading power products satisfy the condition stated
in the lemma (Criterion 2). The iteration of Criterion 1 in Algorithm 4.3 guarantees
that, upon termination of the algorithm, condition (GB5) is satisfied for G and,
hence, G is a Grobner basis.
Example 4.6 Let F:= {/I, 12, h}, where
The total degree ordering of power products is used in this example: first order
by total degree and, within a given degree, order lexicographically. We took an
example with a particularly simple structure of the polynomials in order to make the
reduction process simple and to emphasize the crucial point: the difference of the
crude version of the algorithm and the improved version, which is reflected in the
pairs of polynomials {/I, h}, for which the S-polynomials have to be considered.
A trace of the crude form of the algorithm could be as follows (if the selection
strategy 1. for pairs of polynomials is used: in the trace, we write Ii, Ii -+ ik
for indicating that the reduction of the S-polynomial of Ii and Ii leads to the
polynomial Ik):
12,13 -+ 14 := x 2yz - z3,
/I,/4 -+ 15 := xz3 - xz2,
12, /4 -+ 16 := yz3 - z3,
13, /4 -+ 0,
15, 16 -+ 17 := xyz2 - xz2,
14,17 -+ 18 := z4 - X 2Z 2,
12,17 -+ 0,
15,17 -+ 0,
16,17 -+ 0,
15,18 -+ /9 := x 3 z 2 - xz2,
16, Is -+ O.
The S -polynomials of all the other pairs are reduced to zero. All together one
has to reduce (9 . 8)/2 = 36 S-polynomials.
In the improved algorithm, first, by ReduceAll, It, 12, 13 are reduced with re-
spect to each other. In this example, this reduction process leaves the original basis
unchanged. Then, by NewBasis, It, 12, 13 are put into G. Simultaneously the set
of pairs B for which the S-polynomial have to be considered is generated. The
first pair, again, is
h,h~/4
In this phase, again a call to ReduceAll is made. It is detected that, modulo
{h, 13, /d, It can be reduced to /5, hence, It can be deleted from G and, corre-
spondingly, the pairs {It, h} and {It, h} can be deleted from B. By NewBasis,
/4 and /5 are adjoined to G and B is updated. The consideration of the next pair
in B yields
12,/4 ~ 16
ReduceAll has no effect in this case. Thus, /6 is adjoined to the basis immediately
and B is updated. The consideration of the next pair {f3, Id in B can be skipped
by application of Criterion 1: LP(h) = xy2z divides LCM(LP(f3), LP(/4)) =
x 2y2z and {f3, h} and {h, Id are not in B any more, because they already were
considered. The consideration of the next pairs in B yields
15,/6 ~ 17
14,17 ~ is
with the corresponding updating of G and B (no reductions and cancellations of
polynomials in G are possible!). The S-polynomials of the next pairs reduce to
zero
h,h~O
/5,/7 ~ 0
The criterion does not detect this fact a priori! However, the consideration of the
next pair {f6, h} can, again, be skipped by application of Criterion I: 15 is a
suitable p in the criterion. Then, the following pairs are considered:
/5,/8 ~ fg
/6,/8 ~ 0
/4,19 ~ 0
The next pair {f7, /9} may, gain, be skipped by application of Criterion 1. Finally,
15,19 ~ 0
From now on, the application of Criterion I detects a priori, without actually carry-
ing out the reductions, that all the remaining pairs may be skipped. Hence, instead
of 36 reductions, only 11 have to be carried out with the improved algorithm. The
pair {fa, Is} is an example of a pair, for which Criterion 2 is successful. The
gain by using the criteria, in particular Criterion 1, becomes more drastic as the
complexity of the examples, in terms of the number of variables, the degrees of
polynomials and the number of polynomials, increases.
5. Application: Canonical Simplification, Decision Of Ideal

Congruence And Membership, Computation In Residue
Class Rings
In this section, it is shown how our algorithm for constructing Grobner bases
may be applied for algorithmic solutions to the canonical simplification problem
modulo polynomial ideals, the decision problems 'I =F g' and 'I E Ideal(F)',
and the problem of effectively computing in the associative algebra K[Xl, ... , xn]
IIdeal(F). Actually, the three problems are intimately connected with each other.
This connection is summarized in the following definitions and lemmas whose
proof may be found in [145]. The concepts involved in these lemmas have been
developed and refined in various papers by B. Caviness, J. Moses, D. Musser, H.
Lausch and W. Nobauer, R. Loos, M. Lauer, and the author: see [145] for a detailed
reference to the literature.
Let T be an arbitrary (decidable) set (for example, T := K[Xl,"" xn)) and '"
an equivalence relation on T (for example, '" = =F).
DEFINITION 4.7 An algorithm C with inputs and outputs in T is called a 'canon-

ical simplifier' (or 'ample function') for", (on T) ifffor all objects I, 9 in T
(SE) C(J) '" I and
(SC) if f '" 9 then C(f) = C(g)
(i.e. C singles out a unique representative in each equivalence class. C(f) is

called a canonical form of I ) .
LEMMA 4.5 '" is decidable if there exists a canonical simplifier C for "'.
Proof. By (SE) and (SC): I'" 9 iffC(J) = C(g). The converse of the lemma
is true, also. However, the simplification algorithm constructed in the proof of the
converse is of no practical value, see [145], [146]. D
LEMMA 4.6 Let R be a computable (binary) operation on T, such that", is a

congruence relation with respect to R. Assume we have a canonical simplifier C
for "'. Define:
Rep(T): = {I E TIC(J) = I}(setof'canonical
representatives', ample set),
R'(f,g) := C(R(f,g»(forali I,g E Rep(T)).
Then, (Rep(T), R') is isomorphic to (T/ !'V, R/ rv), Rep(T) is decidable, and R'
is computable. (Here, R/ IV([fJ, [g]): = [R(j,g)], where [fJ isthe congruence
class of / with respect to !'V).
Lemma 4.6 shows that, having a canonical simplifier for an equivalence relation
that is a congruence with respect to a computable operation, one can algorithmically
master the factor structure. The theorem is proven by realizing that i(f) := [J] (f E
Rep(T» defines an isomorphism between the two structures and by checking the
computability properties. Applying these general concepts and facts to the case of
polynomial ideals we first note:
COROLLARY 4.1 (to Theorem 4.1) Let S be an arbitrary normal form algorithm
in the sense of Definition 4.2 and Fa Grobner basis. Then C := >./ . S(F, /),
i.e. the algorithm, that takes the fixed F and a variable / as input and computes
S(F, f), is a canonical simplifier for =F.
Proof. (SE) is fulfilled because, clearly, /=F 9 if / ~F 9 (see Definition 4.1).

By iteration, / =F S(F, f). (SC), in case of =F, is just the content of Theorem
4.1. [J
In addition, one can prove the following lemma.
LEMMA 4.7 [149], [150] Let F be a Grobner basis. Then B := {[u]lu is a

power product that is not a multiple of the leading power product of any of the
polynomials in F} is a linearly independent vector space basis for the vector space
K[XI, ... , xn]/ldeal(F) (the residue class ring modulo Ideal(F)).
Proof. Assume that there is a linear dependence
Cl • [utJ + C2 . [U2] + ... + Cz • [uzJ = 0

for some [Ui] in B. Then
/ := Cl • Ul + C2 • U2 + ... + C, • Ul E Ideal(F).
Hence, by Theorem 4.1,/ must be reducible to 0 modulo F. However, / is already

in normal form because, by definition of B, none of the Ui can be reduced modulo
F. Thus, / = 0, i.e. Cl = ... = Cl = O. [J
Based on the above lemmata, the following problems can be solved by the
following methods (for S use the normal form algorithm NormalForm described
in Algorithm 4.1):

Find a canonical simplifier C for the congruence =F modulo Ideal(F).
METHOD 4.1 [168], [151] Compute G := GB(F). Then the normalform algo-
rithm S(G, f) is a canonical simplifier for =F.
PROBLEM 4.4 Given F, f, g.

Decide, whether f == F g.
METHOD 4.2 [151J Compute G := GB(F).
Then: f ==F 9 iffS(G,f) = S(G,g).
PROBLEM 4.5 Given E, a finite set of equations between generators of a commu-
tative semigroup and two words f, g.
Decide whether the equality f = 9 is derivable from E.
METHOD 4.3 [166J, [169J Let Xl, ... , xn be the finitely many generators of the
commutative semigroup. Conceive every equation p = q in E as a polynomial
p - q in Q[XI,"" x n ].
Compute G:= GB(E).
Then: f = 9 is derivable from E iff S(G, f) = S(G, g).
PROBLEM 4.6 Given F, f.
Decide whether f E Ideal( F).
METHOD 4.4 [151] Compute G:= GB(F).
Then: f E Ideal(F) iff S(G, f) =0.
PROBLEM 4.7 Given F I , F2:
Decide whether Ideal( F I ) ~ Ideal( F2).
METHOD 4.5 [151], [165J Compute G 2 := GB(F2)'
Then: Ideal(Fd ~ Ideal(F2) ifffor all f E FI : S(G 2, f) = O.
Find a linearly independent ba'iis B for the vector space K[XI' ... , xn] / Ideal( F)
(the residue cla'is ring modulo Ideal(F» and,
for any two basis elements [uJ and [vJ in B find a linear representation of[u] .
[v] in terms of the basis elements in B (i.e. find the 'multiplication table' for
K[XI, ... , xn] /Ideal(F».
METHOD 4.6 [149J, [150J Compute G := GB(F).

Take B := {[u] lu is a power product that is not a multiple of the leading power
product of any of the polynomials in G}.
S(G, U· v) yields a linear representation of[u] . [v].
PROBLEM 4.9 Given F, f, h (where K[XI, ... , xnVIdeal(F) is assumed to be

finite-dimensional as a vector space).
Find g, such that f . 9 ==F h (ifsuch a 9 exists).
METHOD 4.7 Compute G:= GB{F).

Represent I and h as a linear combination of the elements in B (see Method 4.6)
and represent 9 as a linear combination with unknown coefficients. Thus, one gets
a linear ,\ystem of equations for the unknown coefficients, which is solvable iff a
solution 9 exists.
Note, that all the above methods are 'unifonn' in the sense that F is a free
parameter in the respective algorithms. Thus, for example, Method 4.3 is a solu-
tion to the unifonn word problem for finitely generated commutative semigroups
(which is equivalent, for example, to the reachability problem for reversible Petri
nets). It has been proven [147], [148] that the unifonn word problem for finitely
generated commutative semigroups and, also, the unifonn congruence problem
for polynomial ideals in Q[Xl' ... ,xn ] is exponentially space complete, i.e. is an
intrinsically hard problem. Method 4.2 shows that this problem can be 'easily'
reduced to the problem of constructing Grobner bases. Hence, the problem of
constructing Grobner bases must be an intrinsically hard problem. For practice,
this means that the worst case behavior of the Algorithm 4.2 and 4.3 may be ex-
tremely bad. However, this does not mean that it is useless to construct Grobner
bases, because in the particular cases at hand, the algorithm may perfonn well (for
example, if the input F is 'nearly' a Grobner basis). Also, if for a given F the
Grobner basis G has been constructed, an infinite number of particular algorithmic
problems of the kind 'I E Ideal(F)?', 'compute a representation of [u] . [v]' etc.
can be solved extremely easily.
Example 4.7 For F as in Example 4.1, I := xy is not in Ideal(F), because
S{GB{F), xy) = -x2 + 1/2x 1= O.
I =F 9 := x2y + {3/2)xy + {1/2)y + 3x2 + (3/2)x - 3/2,
because S{GB{F) , g) is also -x2 + {1/2)x.
Example 4.8 The following reversible Petri net is a Petri net with places a, b, e, I, s
and three transitions that may be described by the rules
as -+ ~s,
bs -+ es,
s -+ I,
where it is implicitly assumed that the 'reverse' rules
~s -+ as
etc. are also available. Let
F = {as - ~ s, bs - cs, s - f}.
Then: a configuration v is reachable from configuration w iff v = F w. For
example, a5bc3 /2s 3 is reachable from a5b2~s5 iff a5be3 p s 3 =F a5b2~s5. In
order to answer such questions, we first compute (w.r.t the total degree ordering)
G:= GB{F) = {s - I,el - bl,b21 - a/}.
a5 bc3 p s 3 is reachable from a5 b2C2 s 5, because the normal forms of both markings
are a7 /5 (with respect to G), whereas cs 2 is not reachable from c2 s, because their
respective normal forms are distinct, namely bP and af.
Example 4.9 For F of Example 4.1,
is a linearly independent vector basis for K[x, y]lIdeal(F), see the corresponding
reduced Grobner basis G in Example 4.5.
because
S{GB{F), x 3 ) = {5/2)x 2 + (5/2)x.
Example 4.10 As an application of the construction of inverses in polynomial
residue class rings, we take the simplification of radical expressions. For the
formulation of the problem see [170]. Consider, for example, the problem of
rationalizing the denominator of
1
x + 21/ 2 + 32/ 3
This problem may be solved by considering the given expression as an element
in Q(x)[2 1/ 2,32/ 3 ], which is isomorphic to Q(X)[Yl'Y2]lIdeal(y~ - 2,y~ - 3),
i.e. the polynomial ring in the two indeterminates Yl, Y2 over the rational function
field Q(x) modulo the ideal generated by the polynomials y? - 2 and y~ - 3. The
application of the algorithm yields the equivalent Grobner-basis
G:= {y? - 2,y~ - 3},
i.e. it is shown by the application of the algorithm that the given basis is already
a Grobner-basis. (In fact, in this simple case, this can be shown by Criterion 2 in
Algorithm 4.3). The residue classes of
1'Y1'Y2'Y1Y2'Y~'Y1Y~
form a vector space basis for Q(x) [Y1, Y2]lIdeal(y? - 2, y~ - 3). In order to obtain
the inverse of x + 21/2 + 32/ 3 we merely have to solve the equation
(x + Y1 + y~) . (a1 + a2Y1 + a3Y2 + a4Y1Y2 + a5Y~ + a6Y1Y~) = l.

By using the reductions Y? -+G 2, y~ -+G 3 this yields a linear system of equations
in the unknowns a 1, ... ,a6 (by comparison of coefficients at the power products
1, Y1, ... , Y1Y~)' whose solution is
a1 = (x 5 - 4x3 + 9x 2 + 4x + 18)/d,
a2 = (_x 4 + 4x 2 + 18x - 4)/d,
a3 = (3x 3 + 18x + 27)/d,
a4 = (-9x 2 - 6)/d,
a5 = (_x 4 - 9x + 4)/d,
a6 = (2x 3 - 4x - 9)/d,
where d = x 6 - 6x 4 + 18x3 + 12x2 + 108x + 73
6. Application: Solvability and Exact Solution of Systems of

Algebraic Equations
In this section, it is shown how the algorithm for constructing Grobner bases may
be used for the exact solution of systems of algebraic equations and questions about
the solvability of such systems. The significance of Grobner bases for problems in
this category stems from the fact that, for Grobner bases, the explicit construction of
all the elimination ideals is extremely simple. This is particularly true for Grobner
bases with respect to the purely lexicographical ordering of power products. It is
not so easy for Grobner bases with respect to other orderings, for example, the total
degree ordering. Still, it is also reasonable to construct Grobner bases with respect
to the total degree ordering for solving algebraic systems because, in extensive
computational experiments, it turned out recently [171] that the complexity of the
algorithm for constructing Grobner bases is extremely sensitive to a permutation
of variables when the purely lexicographical ordering is used, whereas it is nearly
stable, when the total degree ordering is used. Furthermore, the complexity with
respect to the total degree ordering is approximately in the same range as the
complexity with respect to the purely lexicographical ordering, when the most
favorable permutation of variables is used. Since, for a given example, there is
no a priori method to predict which permutation of the variables will give the best
computation times, it, therefore, is also a good method to compute the Grobner
basis with respect to the total degree ordering and then accept the disadvantage
that the computation of the elimination ideals is not so easy as in the case of the
purely lexicographical ordering. In the sequel, we present the method with respect
to both orderings of power products.
LEMMA 4.8 [172] Let F be a Grobner basis with respect to the purely lexico-
graphical ordering oj power products. Without loss oj generality let us assume
Xl <T X2 <T ... <T Xn . Then
/deal(F) n K[Xl, ... , Xi] = /deal(F n K[Xl, ... , Xi]

(fori = 1, ... , n), where the ideal on the right-hand side isJormedinK[Xl,"" Xi].
If,
This lemma shows that the 'i-th elimination ideal' of F is generated by just those
polynomials in F that depend only on the variables Xl, ... , Xi.
Proof. E Ideal(F) nK[Xl' ... ,Xi], then' can be reduced to 0 modulo F (use
Theorem 4.1). With respect to the purely lexicographical ordering determined by
Xl <T X2 <T '" <T Xn , this means that' can be reduced to zero by subtraction
of appropriate multiples bj . Uj ·ii (ii E F) such that LP(fj) contains only inde-
terminates from the set {Xl, ... ,Xi} and, hence, all power products occurring in
ii contain only indeterminates in this set. Also Uj can contain only indeterminates
in this set. Adding all these bj • Uj . ii, one gets a representation of, of the form
I = L aj . Uj . Ij
which shows that' is in Ideal(F n K[Xl, ... , Xi]) o

Decide, whether F has a solution (i.e. whether there exist al,.'" an in an alge-
braic extension oj K such thatJor all, in F : ,(al,"" an) = 0.)
METHOD 4.8 [149] ,[150] Compute G := GB(F).

Then: F is unsolvable iff 1 E G.
Proof. It is well known that F has a solution iff 1 ¢ Ideal(F), see, for example,
[173]. Now, Ideal(F) = Ideal(G) and 1 E Ideal(G) iff 1 is reducible w.r.t. G (by
Theorem 4.1). The latter is true iff 1 E G. 0

Decide, whether F has finitely or infinitely many solutions.
METHOD 4.9 [149], [150] Compute G := GB(F).

Then: F has finitely many solutions ifffor all i( 1 ::; i ::; n): a power product of
the form xli occurs among the leading power products of the polynomials in G.
Proof. It is well known that F has finitely many solutions iff the vector space
K[Xl' ... ,xn ]lIdeal(F) has finite vector space dimension, see, for example, [173].
Because of Lemma 4.7 this is true iff the set B considered in Lemma 4.7 is finite.
It is easy to see from the definition of B that B is finite iff the condition stated in
Method 4.9 is satisfied. D
About the exact dimension of polynomial ideals, one can say more than is
expressed above by using Grobner bases for computing the Hilbert function of
polynomial ideals. Many details are given in [174], [175].
PROBLEM 4.12 Given F (solvable, with finitely many solutions).

Find all the solutions of the system F.
METHOD 4.10 [172] Compute G := GB(F) with re!Jpect to the purely lexico-
graphical ordering of power products.
The polynomials in G, then, have their variables "separated" in the precise sense
ofLemma 4.8 (G is 'triangularized' ). G contains exactly one polynomial ofK[Xl]
(actually, it is the polynomial in /deal(G) n K[xd with smallest degree).
The successive elimination can, then, be carried out by the following process:
p: = the polynomial in G n K[xd
Xl := {(a)lp(a) = O}
for i : = 1 to n - 1 do
Xi+l:= 0
for all (a" ... , ai) E Xi do
H:= {g(a" ... ,ai,Xi+1)1
9 E G n K[x" ... ,Xi+1] - K[Xl' ... ,Xi]}
p: = greatest common divisor of the polynomials in H
(Actually, {p} = G B (H); in the ca'ie of univariate
polynomials the algorithm G B specializes to
Euclid's algorithm!)
Xi+l := Xi+1 U {(al, ... , ai, a)lp(a) = O}.
Upon termination, Xn will contain all the solutions. (Note that some of the p may
be 1, i.e. the corresponding partial solution (al, ... ,ai) cannot be continued.)
Of course, for the univariate polynomials p occurring in the algorithm, the 'ex-
act' determination of all their zeros may not be possible effectively. However, of
course, this is not a deficiency of the particular method but an intrinsic limitation of
algorithmic solvability of polynomial equations. Still, Method 4.10 is an algorith-
mic method (using only arithmetic in K) for completely reducing the multivariate
problem to the univariate one.
Before we can give a method for Problem 4.11 that is based on Grobner bases
with respect to arbitrary orderings of power products we must solve the following
problem.
PROBLEM 4.13 Given a Grobner basis G, such that G, as a .system of equations,

has only finitely many solutions.
Find the p E/deal(G) n K[XI] with minimal degree.
METHOD 4.11 [150] (In case the purely lexicographical ordering with Xl <T
X2 <T ... <T Xn is used, the solution of the problem is easy, see Method 4.10. In
the other cases proceed by the following method.)
Determine do, ... , dL by the following process, which involves the solution of
:,ystems of linear equations in every step:
i:= 0
repeatpi := S(G, xi)
i:= i + 1
until there exists (do, ... , l4-d 1= (0, ... ,0) such that do· Po + ... +
+di-l . Pi-l = 0
1:= i - 1
Then, P = do . 1 + d l . Xl + ... + d, . xi
METHOD 4.12 [150] for solving Problem 4.12.

Compute G:= GB(F).
The successive elimination can, then, be carried out by the following process:
p: =the polynomial in Ideal(G) n K[XI] of minimal degree
(see Method 4.11)
Xl := {(a)lp(a) = O}
for i := 1 to n - 1 do
Xi+l:= 0
for all (al, ... , ail E Xi do
H := {g(al, ... , ai, Xi+! , ... , xn)lg E G}
H:=GB(H)
p: =the polynomial in Ideal( H) U K[Xi+!]
of minimal degree
Xi+! := Xi+l U {(al, ... , ai, a)lp(a) = O}
Upon termination, Xn will contain all the solutions. (Note, again, that some
of the p may be one, i.e. the corresponding partial solution (aI, ... , ai) cannot be
continued. Also, of course, one will store the Grobner basis H corresponding to
a particular partial solution (al, ... , ai) and use it instead ofG for construction
of H corresponding to (al, ... , ai, a).)
Example 4.11 The system F of Example 4.1 is solvable, because G = GB(F)

does not contain the polynomial 1 (see Example 4.5). The system
F := {X2y - X2, x3 - x 2 + y, xy2 - xy + 2}
is unsolvable. Let us use the total degree ordering in this example.
SPolynomial(x 2y - x 2, x 3 _ x 2 + y)
= x 2y _ x 3 _ y2 -+F -x3 _ y2 + x 2 -+F _y2 + y.
Thus, we have to adjoin y2 - Y to the basis.
SPolynomial(xy2 - xy + 2, y2 - y) = 2,
which cannot be reduced further. Hence, we have to adjoin 1 to the basis. This is
the signal that F is unsolvable.
Example 4.12 F of Example 4.1 has only finitely many solutions, because x 3 and
y appear as leading power products in GB(F).
F := {x 2y _ y2 _ x 2 + y, x 2 _ y}
has infinitely many solutions. Actually, F is already a Grobner basis (with respect
to the total degree ordering of power products): check by applying Algorithm 4.3
which, in this case, does not adjoin any new polynomial to F. No power products
of the form yi occurs among the leading power products. Hence, F has infinitely
many solutions.
Example 4.13 For F of Example 4.1,
GB{F) = {x 3 - {5/2)x 2 - (5/2)x, Y + x 2 - (3/2)x - 3}

The solutions a of the first (univariate!) polynomial are 0, (5 + v'65)/4, (5 -
v'65)/4. Each of these solutions can be continued to a solution (a, b) of F by
solving the second polynomial in the form y + a 2 - 3/2a - 3 for y. This yields
(0,3), «5 + v'65)/4, -(3 + v'65)/4), «5 - v'65)/4, (-3 + v'65)/4) as the three
solutions of the system.
Example 4.14 The same example can also be treated by Method 4.12. With
respect to the total degree ordering, G := GB(F) = {91,92,93} where
91 := x 2 + y - (3/2)x - 3,
92 := xy - y + x + 3,
93 := y2 - (5/2)y - 4x - 3/2,
We now compute the normal forms of 1, X, x 2 , ••• :
S(G, 1) = 1,
do . 1 = 0 has no non-trivial solution

S(G, x) = x
do' 1 + dl . X = 0 has no non-trivial solution
S(G, x 2) = -y + (3/2)x + 3
do . 1 + d l . X + d2 • x 2 = 0 has no non-trivial solution.
S(G, x 3) = -(5/2)y + (25/4)x + 15/2
do . 1 + d l . X + d2 • x 2 + d3 • x 3 = 0 leads to
the following linear system of equations:
-(5/2)d3 - d2 = 0
(25/4)d3 + (3/2)d2 + d l = 0
(15/2)d3 + 3d2 + do = 0
which has (after normalization d3 = 1) the unique solution d3 = 1, d2 = -5/2,
dl = -5/2, do = O. This means that
p := x 3 - (5/2)x 2 - (5/2)x
is the polynomial in Ideal(G) n K[x] with minimal degree (in accordance to what
we already have seen in Example 4.13). p has the three solutions al = 0, a2 =
(5 + V65)/4, a3 = (5 - V65)/4. Substitution of al yields
gl(at) := Y - 3,
g2(at} := -y + 3,
g3(at} := y2 - (5/2)y - 3/2,
The Grobner basis corresponding to these three polynomials is
G' := {y - 3}
By computing the normal forms 1, y, y2, ... and looking at the corresponding
systems of linear equations as above one detects that
p':= y - 3
is the polynomial in Ideal(G')nK[y] of minimal degree. Of course, in this particu-

larly simple example, this can be seen immediately from the Grobner basis. Hence,
(aI, btl with bl := 3 is the first solution of the system. Similarly, substitution of
a2 yields
gda2) := y + (3 + V65)/4,
g2(a2) := (1 + V65)/4y + (17 + V65)/4,
g3(a2) := y2 - 5/2y - (13 + V65),
The Grobner basis corresponding to these three polynomials is
G" := {y + (3 + V65)/4} andp":= y + (3 + V65)/4
is the polynomial in Ideal(G") n K[y] of minimal degree. Hence, (a2'~) with
~ := - (3 + v'65) /4 is the second solution of the system.
Finally, substitution of a3 yields, again, three polynomials in K[y] whose

Grobner basis consists of the polynomial y + (3 - v'65)/4. Hence, the third
solution is (a3, b3) with b3 := (-3 + v'65)/4.
Example 4.15 Given F consisting of

4x2 + xy2 - Z + 1/4,
2x + y2z + 1/2,
x 2z - (1/2)x _ y2,
the corresponding Grobner basis G (with respect to the purely lexicographical
ordering, where z <T y <T x) consists of
Z7 - (1/2)z6 + (1/16)z5 + (13/4)z4 + (75/16)z3 + (171/8)z 2-t
+(133/8)z - 15/4,
y2 _ (19188/497)z6 + (318/497)z5 - (4197/1988)z4-
- (251555/1988)z3 - (481837/1988)z2 +
+(1407741/1988)z - 297833/994,
x + (4638/497)z6 - (75/497)z5 + (2111/3976)z4+
+(61031/1988)z3 + (232833/3976)z2 - (85042/497)z+
+144407/1988.
Applying Method 4.10 for solving G, one first had to find all the solutions of the
first polynomial, which is univariate. Each of these solution aI, can be continued
to two solutions (aI, a2) of the second polynomial and each of these (aI, a2) can
be continued to a solution (aI, a2, a3) of the third polynomial. The solutions of the
first polynomial can be determined systematically with any guaranteed precision,
see [176]. It has not yet been studied systematically how, numerically, the precision
of the solutions of the first equation must be fixed in order to guarantee a given
precision for all the solutions of the last equation. This is a near-at-hand important
problem for future study.
Example 4.16 Sometimes, it is necessary to solve systems of algebraic equations

with 'symbolic' coefficients. For example consider F consisting of
II := X4 + (b - d),
h := X4 + X3 + X2 + Xl + (-a - c - d),
h := X3 X4 + XIX4 + X2 X3 + (-ad - ac - cd),
14 := XIX3X4 + (-acd),
where Xl <T X2 <T X3 <T X4 are the polynomial indeterminates and a, b, c, d
are 'symbolic' coefficients. One might like to solve this system for Xl, X2, X3, X4.
This is nothing else then saying that one conceives the polynomials as elements in
Q(a, b, c, d)[XI' ... ,X4], where Q(a, b, c, d) is the field of rational functions over
Q. Our algorithm works over arbitrary fields and, hence, in particular also over
Q(a, b, c, d). Some steps of Algorithm 4.3 are:
Reduction of II modulo h (by subtraction of h from II and nonnalizing the

coefficient of the leading power product to 1) yields
Ii := X3 + X2 + Xl + (-a - b - c)(1I may be cancelled).

Reduction of h modulo Ii yields
I~ := X4 + (b - d)(h may be cancelled).

Reduction of h modulo the other polynomials (starting with the subtraction of
X3 . I~ and, then executing several other reduction steps) yields
I~ := x~+ 2XIX2 - (a + 2b + c - d)X2 + x~-

-(a + b + C)XI + (ab + ac + b2 + be - bd)
(f3 may be cancelled).
Reduction of 14 yields
I~ := XIX2 + x~ - (a + b + C)XI - aedj{b - d)(canceI/4).

(Note here that division in Q(a, b, c, d) has to be perfonned. I~ can now be further
reduced (using 14) yielding If
I~' := x~ - (a+ 2b + C - d)X2 - x~ + (a + b + C)XI +

+(ab2 + abc - abd + aed + b3 + b2c-
-2b2d - bed + b~)j(b - d).
Cancel I~. No further reduction is possible. Therefore, we consider
SPolynomial(f~', I~) = Xl· If - X2 . I~

Reduction of this polynomial yields
15 := X2 + (b 2 - 2bd + d2)j(aed)x~ + (abe + abd - ad2 + bed - ed2)

j(aed)XI + (-b + d)
Now, again, a number of reductions are possible yielding, finally,
91·.= X 3 + -b2 ±2bd-d

acd
2 x2+
1
+ -abc-abd±ac:jad2 -bcd±cd2 Xl + (-a - C- d),
92 := + (b - d),
X4
._ + (abc±abd-2acd±bcd-ad2 -cd2 ) 2+
3
93 .- Xl (b2-2bd±d2) Xl
+(a 2ed + a~d + acJl)j(b2 - 2bd + ~)XI +
+(a2eJl)j(b3 - 3b2d + 3bJl + ~,
94 := x2 + (b 2 - 2bd + ~)j(acd)x~+
+ (abC+abd-a~:cP+bcd-cd2) Xl + (-b + d)
By Criterion 1, the reduction of the S-polynomials of these polynomial may be
skipped. Hence, G := {gl, ... ,gd is the reduced Grobner basis. By Methods 4.8
and 4.9 it can be seen that the system has finitely many solutions. The system must
contain a univariate polynomial in Q(a, b, c, d)[xd : g3. A particular solution of
g3 is
al := (-ad)/(b - d),
which can be extended to a solution (aI, a2, a3, a4) of the entire system, where
a2 := (ab + b2 - bd)j(b - d),
a3 := c,
a4:= -b+d.
Dividing g3 by (Xl - ad one gets a quadratic polynomial whose solutions can be
extended to solutions of the entire system in the same way as before.
7. Application: Solution of Linear Homogeneous Equations

With Polynomial Coefficients
In this section, it is shown how the algorithm for constructing Grobner bases may
be used for determining a finite set of generators for all the polynomial solutions of
a linear homogeneous equation with polynomial coefficients. Before the method
can be described, it must be shown how one can find a linear representation of the
polynomials in a basis F in terms of the polynomials in its corresponding Grobner
basis G and vice versa.
PROBLEM 4.14 Given a Grobner basis G = {gl,'" ,gm} and some f.

Find hI"'" hm such that f = hI 'g, + ... +hm ·gm (and LP(hi ·gd <5.T LP(f)
fori = 1, ... ,mY.
METHOD 4.13 Roughly, reduce f to zero modulo G and collect the multiples of
the gi necessary in the reduction. In more detail: take Algorithm 4.1 (the normal
form algorithm) and insert instructions that collect the multiples of the gi used in
the reduction.
hI := ... := h m := 0
while f -I- 0 do
choose i, b, u such that f ~gi,b,u and u . LP(gi)
is maximal W.r.t. <T
f := f - b . u . gi
hi := hi + b· u
PROBLEM 4.15 Given F = {II, ... , fz} and G = {gl,'" ,gm} such that G =
GB(F).
Find Y such that Y is a matrix of polynomials with m rows and I columns and
fj = L gi· Yi,j (forj = 1, ... ,1)

l:'S:i:'S:m
METHOD 4.14 The j-th column ofY consists of hI, ... , hm that are obtained by
the Method 4.13 for the representation of h(j = 1, ... ,1).
PROBLEM 4.16 Given: F = {II,···, Il}.

Find G = {gl, ... , gm} and X such that G = GB (F), X is a matrix of polyno-
mials with I rows and m columns and
gZ- -- " f-3 ·X--

'~ 3,Z (fori = 1, ... , m).
l:'S:j:'S:l
METHOD 4.15 Augment Algorithm 4.2 or Algorithm 4.3 by instructions that keep
track of the multiples of h that are used in the reduction of those polynomials
whose normalform is adjoined to the ba<;is G (compare Method 4.13)
PROBLEM 4.17 Given a reduced Grobner ba<;is G = {gl, ... ,gm}.

Find a matrix R with m columns such that the finitely many rows of R constitute
a set of generators for the linear homogeneous equation
hI· gl + ... + hm . gm = O(h l , ... , hm E K[XI, ... , xn])
i.e. R should consist ofm-tuples (kl,l, ... ,kl,m), ... , (kr,l, ... , kr,m) ofpolyno-
mials such that
kj,l . gl + ... + kj,m . gm = 0 (jor j = 1, ... ,r)
andfor all (hi' ... ' hm)for which
hI . gl + ... + hm . gm = 0
there exist polynomials PI, ... ,Pr such that
(hl,··.,hm)=
= PI . (kl,l, ... , kl,m) + ... + Pr . (kr,l, ... ,kr,m)
METHOD 4.16 [177], [178], [179], [180], [174], [175] R: = empty matrix
for all pairs (i,j)(l:CS: i < j:CS: I):
Consider h := SPolynomial(gi,gj) = Ui . gi - (Ci/Cj) . Uj . gj, where Ci is
the leading coefficients of gi, Ui is such that 8i . Ui is the LCM(81, 82), 8i is the
leading power product of gi (i = 1,2).
Reduce h to zero modulo G and store the multiples of the gl, ... ,gl necessary
for this reduction. This gives a representation of h of the form
h = kl . gl + ... + kl . gl(compare Method 4.131).

Add (... , Ui,"" -(Ci/Cj) . Uj,"') - (kI , ... , kz) as last row in R
t t
position i position j
PROBLEM 4.18 Given F = {h, ... , III arbitrary.
Find a matrix Q with I columns such that the finitely many rows of Q constitute a
set of generators for the linear homogeneous equation
hI' h + ... + hi' Il = O{hI , ... , h, E K[XI, ... , xnD·
METHOD 4.17 [178] By Method 4.15, compute G = GB{F) = {gl,'" ,gm}

and a matrix X with i rows and m columns such that
gi = L Ii· Xj,i(jor i = 1, ... ,m).

I~j~l
By Method 4.14, compute a matrix Y with m rows and i columns such that
Ii = L gi' Yi,j(jor j = 1, ... ,i).

l~i~m
By Method 4.16 compute a matrix R with m columns such that the r rows of R
constitute a set of generators for the linear homogeneous equations
hI . gl + ... + hm . gm = O.
Then,
Q= [ .~ ~ .~~.' .~t. ] (a block matrix)

R·xt
(I is the unit matrix with i rows and columns, X t is the trampose of X).
Example 4.17 Let F:= {h, 12, h}, where

h := x2y - xy,h:= xy2 - x 2,
h := x 3 y - x 2y + x3 - x 2.
We use the total degree ordering. First, G := GB{F) has to be computed with
simultaneous determination of the matrix X. We start with a reduction of h:
h- x . h = x3 - x 2 =: I~·
The representation
I~ = (-x) . h + 0·12 + 1· h
must be stored. Then we reduce the S-polynomial of hand 12:
h: = SPolynomial{h, h) = y . h - x . 12,
h+ h - I~ = 0 =: 14.
If 14 was not zero, the following representation of 14 in tenns of /1, h and /3
could be obtained from this reduction:
14 = y . /1 - x . h + h - /3 + x . h =
= (y + x)· /1 + (-x + 1)· h + (-1)· /3.
This example of a reduction should suffice to demonstrate how the linear rep-
resentations of the new polynomials in G in tenns of the polynomials in F can
be obtained in general. Since, however, /4 is zero, nothing has to be adjoined
to G in this stage of the algorithm, The S-polynomial of /1 and I~ and also the
S-polynomial of h and I~ reduce to zero. Hence,
G:= {91,92,93}
where
· - 11, 92·-
91.- .- f 2, 93·-
.- x 3 - x 2 ,
is the reduced Grobner basis corresponding to F and
1 0 -x
X:= [ 0 1 0
1
001
is the transfonnation matrix.
The matrix Y for the reverse transfonnation (i.e. the linear representation of
the elements of Fin tenns of the elements in G) is obtained by Method 4.14:
/1 reduces to zero modulo G by subtraction of 91,

h reduces to zero modulo G by subtraction of g2,
h reduces to zero modulo G by subtraction of X· 91 and 93.
Hence,
Y:= [~0 0~ 1~ 1
For getting R, we have to reduce the S-polynomials of the pairs (gi,gj):
h 1,2: = SPolynomial(gI,92) = y. gl - X· 92·

hl,2 + 92 - 93 = O.
, = SPolynomial(gl, 93) = X· gl - y. g3·
h13:
hl,3 = O.
h 2,3: = SPolynomial(g2, g3) = x 2 . 92 - y2 . 93.
h2,3 - Y . gl - X . g3 - g2 + g3 = O.
From the first reduction:
Y . 91 - X • 92 + 92 - 93 =
(y) ·91 + (-x + 1) ·92 + (-1) ·93 = o.

Hence, the first row in R is the solution (the 'syzygy')
(y, -x + 1, -1).
The other rows of R are obtained analogously:
(-x + 1)
R= [ ~~~
(-y)
(0)
(-1)
(-y)
(x 2 - 1) (_y2 +x+ 1)
]
Finally, the computation of Q requires only some matrix multiplications: First, we

note that yt . X t = I in this particular example. Hence,
(0) (0) (0)

(0) (0) (0)
(0) (0) (0)
(y+ x) (-x + 1) (-1)
(x + xy) (0) (-y)
(xy2 - x - Y - x) (x - 1) (_y2 + X + 1)
2 2
Of course, the first three rows can be cancelled in this particular example, the last
three rows constitute a complete set of generators for the solutions (hI, h2' h3) to
the equation hI . h + h2 . h + h3 . fa = O.
For K[Xl, ... ,xn]-modules, as for example the module of all the solutions to
the above linear equation, a notion of 'Grobner bases' and 'reduced Grobner bases'
can be introduced, see [180], [174], [175]. Then the matrices Q can be reduced to
a minimal set of generators and the construction can be carried over to obtain the
whole 'chain of syzygies' or the 'free resolution' of a polynomial ideal.
8. Grobner Bases For Polynomial Ideals Over The Integers

The concept of Grobner bases, the essential properties of Grobner bases and the
algorithm for constructing Grobner bases as reflected by Definitions 4.2, 4.3, 4.5,
4.6, Lemmata 4. 1, 4.2, 4.3, Theorems 4.1,4.2,4.3,4.4, Algorithms 4.1,4.2,4.3 and
most of the applications in Sections 5 and 7 can be carried over to polynomial ideals
in Z[Xl, ... ,xn ] and, in fact, to ideals in certain other rings, see [160]. However,
a subtle analysis of the notion of reduction and, more essentially, of the notion of
'S-polynomial' must be carried out for this purpose. We cannot go into the details
of the theoretical foundations of the algorithm for integer polynomials. Rather, we
explain the steps of the generalized algorithm in the style of the preceding sections.
The problem of deciding ideal membership for ideals in Z[Xb ... ,xn ], the sim-
plification problem for these ideals and related problems have a long and interesting
history. For some of the details of the history, see [181]. The first general solu-
tion of both the simplification and (hence,) the membership problem, was given
by Lauer [182] based on the Grobner bases approach but needing two different
types of 'S-polynomials'. Other solutions based on the Grobner bases approach,
but destroying the simple structure of the algorithm, were given in [172], [178],
[179]. The first general solution based on a different approach was given only in
[181]. Our own solution [160], which will be presented here, seems to be much
more concise than the solutions given so far and leaves the simple structure of the
algorithm untouched.
In addition to some ordering ofthe power products, in the case of Z[Xb ... ,X n ],
one also must fix some ordering of the integers, for example, 0< -1 < 1 < -2< 2 < -3 <
3 < ... (An axiomatic characterization of the admissible orderings is possible but will
not be used in this paper). The crucial difference, then, to the case of polynomials
with field coefficients is that, in the definition of 'reduction' (Definition 4.1) it is not
possible to totally cancel Coefficient(g, t), where t = u· LeadingpowerProduct(f),
because the element Coefficient(g, t)! LeadingCoefficient(f), in general, will not
be in Z. In the following, the typed variables a, b, c, d will be used for integers in-
stead offield elements, f, g, h, k,p, q will beusedforpolynomialsinZ[xl, ... , x n ],
and F, G for finite sets in Z[Xb ... ,Xn ].
DEFINITION 4.8 [160] 9 -+F h (read: 'g reduces to h modulo F') iff there exists
f E F, band u such that
9 -+ /,b,!'/' and h = 9 - b . u . f·
9 -+ /,b,!./, (read: 'g is reducible using f, b, u') iff
ai=0anda-b·c<a,
where,
a = CoejJicient(g, U· LeadingPowerProduct( f)), and
c = LeadingCoejJicient(f )
Example 4.18 The b in Definition 4.8 can be determined by the following algo-
rithm M(a, c), for example:
M(a, c): = if a and c have the same sign
then if a - c < a then M (a - c, c) + 1
else 0
else if a + c < a then M (a + c, c) - 1
else 0
In practice, M may be realized by a modified integer division.
The definitions, theorems, algorithms and lemmata of Section 2 can now be carried
over without any change: In particular, we have again the algorithm NormalForm
that produces a nonnal fonn for every polynomial, we have the notion of a Grobner
basis, the characterizations (GB2) and (GB3) of Grobner bases and the connection
between reduction and ideal congruence stated in Lemma 4.3. For the fonnulation
of the algorithm that constructs Grobner bases, however, we need some additional
preparation.
DEFINITION 4.9 [160] The least common reducible ofcl, C2 is defined asfollows:
LCR(cl, C2) := max{L{ct) , L{c2))(max taken w.r.t. <),

where
L{c) : =abs{c)/2, ifc is even
-(abs{c) + 1)/2, ifcis odd.
DEFINITION 4.10 [160] Pl and P2 constitute the critical pair corresponding to It

and h iff
Pi = a· U - M{a,q)· 'Ui· fi,where
U = LCM(Sl, S2),
a = LCR{cl, C2).
Si = LeadingPowerProduct( Ii),
q =LeadingCoejJicient(Ii),
'Ui is such that 'Ui . Si = U{i = 1, 2).
The difference of the two components of a critical pair is the analogue to the S-
polynomial in the case of field coefficients. We formulate the algorithm for critical
pairs instead of S -polynomials, because, at present, we do not have a formal
proof that, in fact, the algorithm below is correct with S -polynomials instead of
critical pairs, although it is very likely. Also, we would like to introduce the
concept of a critical pair to the reader, because this concept may be applied to
domains without any operation of subtraction also. See [145] for an introduction
to 'critical-pair/completion' algorithms.
Example 4.19 0, -1,1, -2,2, -3,3, -4,4 are the values of L for the arguments 0,
1,2, 3,4,5,6,7, 8, respectively, and LCR(3, 1) = -2, LCR(7, 8) = 4. Note that
L{c) = L( -c).
The main theorem of Section 3, which gives an algorithmic characterization of

Grobner bases, and the main algorithm for the main problem can now be carried
over in the following form:
THEOREM 4.6 (Buchberger [160]) Let S be an arbitrary normal form algorithm.

The following properties are equivalent:
(GBl) F is a Grobner bavis.
(GB3) For all /I,h E F,PbP2:

ifPI and P2 constitute the critical pair corresponding to ft, 12, then S(F, It} =
S(F,h)·
Find G, such that Ideal(F) = Ideal(G) and G is a Grobner basis.
ALGORITHM 4.4 (Buchberger [160])jor solving Problem 4.19.
G:=F
B := {/I, h}lft, 12 E G}
while B i= 0 do
ft, 12 : = a pair in B
(Pb P2): = the critical pair corresponding to /I, 12
,
(pi,p~): , = (S(G,PI,
, S(G,P2)
h := PI - P2
if h' i= 0 then
B := B U {{g, h'}lg E G}
G:=GU{h'}.
Also the various improvements of the algorithm, the notion of reduced Grobner
bases and the theorem on the uniqueness of the reduced Grobner bases (Section 3)
can be carried over. We do not explicitly state the details.
Example 4.20 Take F as in Example 4.1. Note that the leading coefficients of
the polynomials in F can not be simply set to 1 by dividing the whole polynomial:
the ideal would change! We fix the 'purely lexicographical' ordering for the
bivariate power products with the ordering x <T y of the two indeterminates. In
order to 'complete' F by Algorithm 4.4, one has to consider the 'critical pairs' of
polynomials in F. We start with 12, Is: LC(h) = 2, LC(Is) = 1, LCR(2, 1) =
1, LCM(LP(h), LP(f3)) = x 3y. Thus, x 3y is the monomial that has to be
reduced in one step modulo 12 and fa in order to get the critical pair corresponding
to 12, fa. The polynomial x 3 y maybe reduced by 12 in the following way:
x 3y -+/2 _x3y + xy + y - 6x 3 + 2X2 + 3x - 3 =: p.
P may be further reduced modulo fa:

p -+ faX 2y + xy + y - 3x 3 + 4x2 + 3x - 3 =: p'.
p' is irreducible with respect to F. The polynomial x 3 y may also be reduced by
fa:
x 3 y -+ fa _x 2y - 3x 3 - 2x2 =: q.
Also q is irreducible with respect to F.p' i= q and, hence,
14 := p' - q = 2x 2y + xy + y + 6x2 + 3x - 3
must be adjoined to the basis.

Similarly, one now has to consider the next critical pair, for example, the one
corresponding to h, 14 : _2x2y is the 'least common reducible' of h and 14,
which has to be reduced in one step modulo hand 14, yielding
p := x2y + 2xy + y + 9X2 + 5x - 3 and
q := xy + y + 6x 2 + 3x - 3,
respectively. Reduction to normal forms yields
p' := _x 2y + xy + 3x2 + 2x (using /4) and
q' := xy + y + 6x 2 + 3x - 3.
Thus, the difference of these two polynomials must be adjoined to the basis:
15 := _x 2y - y - 3x 2 - X + 3.
Similarly, the consideration of the critical pair of 14 and 15 leads to
16 := -xy + y - x - 3.
The consideration of the critical pair of 15 and 16 leads to
17 := 2y + 2x2 - 3x - 6.
Finally, the consideration of the critical pair of 16 and 17 leads to
Is := 2x3 - 5x2 - 5x.
The consideration of all the other critical pairs leads to identical normal forms.
Hence, G := {h, ... , Is} is a Grobner basis corresponding to F. Actually, the
consideration of most of these critical pairs can be avoided a priori by the improved
version of the algorithm. Furthermore, some of the polynomials in the basis can
also be cancelled in the course of the algorithm. Reduction of all the Ii modulo
G - {Ii} leaves us with the reduced Grobner basis G' := {/~, 17, Is} .where
I~ := -xy - y - 2x2 + 2x + 3.
Note that the reduced Grobner bases corresponding to F are different depending
on whether we work in Q[Xl, ... , x n ] or in Z[Xl, ... ,xn ].
9. Other Applications
A number of other applications of Grobner bases have been reported in the
literature: decision, whether a given polynomial ideal is principal [150], Hilbert
functions of polynomial ideals [149], [180], [174], [175], Lasker-Noether decom-
position of polynomial ideals [183], free resolutions of polynomial ideals and
syzygies (a generalization of the above linear equation problem with polynomial
coefficients) [180], [175], multidimensional integration [184] and bijective enu-
meration of polynomial ideals. The latter problem asks for an algorithm that
enumerates bases for ideals in R[Xl' ... ,xn ] (R a ring) such that every ideal is
represented exactly once in the enumeration. By Theorem 4.4, it is clear that
a bijective enumeration of all ideals in K[xl, ... ,xn ] and Z[Xl, ... ,xn ] can be
achieved by bijectively enumerating all Grobner bases in these polynomial rings,
which is easily possible (see [154]). The applicability of Grobner bases to other
problems is extended, for example, to the construction of Hensel codes for rational
functions.
10. Specializations, Generalizations, Implementations,

Complexity
The algorithm for constructing Grobner bases specializes to Gau{3' algorithm
in case F consists only of linear polynomials, it specializes to Euclid's algorithm
in case F consists only of univariate polynomials, it specializes to an algorithm
for the word problem for finitely generated commutative semigroups in case F
consists only of polynomials of the form u - v (differences of power products)
[166], [169]. The algorithm for Z[Xl' ... ,xn ] specializes to Euclid's algorithm in
Z in case n = 0 [160].
The algorithm has been generalized for polynomials over various rings, in par-
ticular, over Z [182], [172], [178], [179], [160], and for associative algebras [161].
The Knuth-Bendix generalization [155] was already discussed in the introduction.
An interesting generalization was also undertaken by G. Bauer [185], who gives
an axiomatic definition of the concept of 'substitution' and was able to define the
notion of 'critical pair' in this general context.
The algorithm has been implemented various times, [149], [183], [186], [179].
[186] is an implementation in SAC-I. R. Gebauer and H. Kredel [171], Univ. of
Heidelberg, ER.G., work on the implementation of the algorithm in SAC-2, which
will be included in the next release of SAC-2 (announced for December 1983).
SAC-2 is a large software system for symbolic computation in algebraic domains,
in particular in polynomial domains. It is written in the ALOES language, whose
compiler is written in FORTRAN. Thus, SAC-2 is installed easily whenever FOR-
TRAN is available. G. E. Collins (University of Wisconsin-Madison, Departments
of Computer Science) and R. Loos (Universitiit Karlsruhe, Institut fur Informatik
n are the authors of the SAC-2 system. The implementation of our algorithm
in SAC-2 by R. Gebauer and H. Kredel gives the user the choice to use various
orderings of power products, to work over various coefficient domains (including
the field of rational functions over Q) and to communicate in convenient input and
output format with the computer .
Various analyses of the complexity of the algorithm have been carried out: [149],
[166], [187], [148], [188]. Summarizing, these analyses show that the degrees of
the polynomials in the reduced Grobner bases, with probability 1, stay below
d, + ... + dt - n + 1, where the dis are the degrees of the input polynomials. In
exceptional cases, this bound does not hold. Many theoretical questions remain
open. Typical running times in SAC-2 on an IBM 3701168: several seconds
for F with 3 polynomials of degree 3 in 3 variables, 20 sec for the example in
[172] with 6 polynomials of degree 3 in 6 variables. However, this computing
time may drastically change if a different permutation of the variables and purely
lexicographical ordering is used. For the worst permutation, the computation was
as high as 10000 sec, whereas in the total degree ordering the computation time for
the same example was always in the range 20-30 sec independent of the permutation
of variables. See Section 6 for the consequences of these observations. A number
of papers on the complexity and applications of Grobner bases have appeared in
the literature. Some of them are collected in the two proceedings [189] [190]. The
Journal of Symbolic Computation (Academic Press) carries a significant number
of articles in this area.
11. Updates
Since the appearance of this chapter in the first edition of this book, my theory
and method of Grobner bases have received quite some attention by the research
community resulting in
• a number of textbooks devoted to Grobner bases
• various generalizations and improvements of the Grobner bases method, and
• numerous new applications of the method in many fields of mathematics.

Instead of giving a complete list of textbooks, I just mention the book by M.
Kreuzer and L. Robbiano [191] because this book, in addition to being a nice
blend of theory and algorithms, contains a complete list of all other textbooks on
Grobner bases that appeared so far. The vast original literature on Grobner bases
that meanwhile has been published can then be easily obtained from the references
in these textbooks.
In 1998, we organized a conference on Grobner bases with the intention of
getting. an overview on the various application areas of Grobner bases (in alge-
braic geometry, invariant theory, automated geometric theorem proving, integer
programming, partial differential equations, hypergeometric functions, symbolic
summation, coding theory, cryptography, statistics, numerics, etc.). The proceed-
ings [192] of this conference contain survey articles on all these application areas.
In addition, the proceedings contain an introductory article on Grobner bases and
applications based on a concise proof of the main theorem, [192, pp. 3-31], and
they also contain an English translation, [192, English translation of [150],pp.535-
545], of my original article [150]. Unfortunately, the proceedings [192] do not
contain a survey on the numerous applications of Grobner bases in systems theory
that have been stimulated and discovered since the appearance of my paper in 1985.
However, Z. Lin and L. Xu now published a special issue of the Journal of Mul-
tidimensional Systems and Signal Processing, [193], which is entirely devoted to
applications of Grobner bases in systems theory. This issue also contains another
introduction to Grobner bases that is meant for systems theorists who do not want
to go into formal details but just would like to see and understand how the method
works in examples, see [193, pp. 223-253]. In this introduction, we also give the
following list of known applications of Grobner bases in systems theory with the
respective references to the literature:
• solvability and solution of unilateral and bilateral polynomial matrix equations,

Bezout identity,
• design of FIRIIIR multidimensional filter banks,
• stabilizability/detectability of feedback stabilizing compensator/asymptotic ob-

server,
• synthesis of deadbeat or asymptotic tracking controller/regulator,
• minimax controller design,
• constructive solution of the n-D polynomial matrix completion problem,
• computation of minimal left/right annihilators,
• elimination of variables for latent variable representation of a behavior,
• computation of controllable part; controllability test,
• observability test,
• computation of transfer matrix and minimal realization,
• solution of the Cauchy problem for discrete systems,
• testing for inclusion; addition of behaviors,
• test zero/weak zero/minor primeness,
• finite dimensionality test,
• computation of sets of poles and zeroes; polar decomposition,
• achievability by regular interconnection,
• computation of structure indices.
Acknowledgement
The work described in this paper is supported by the Austrian Research Fund,
Project No. 4567. I am indebted to R. Gebauer, H. Kredel and F. Winkler for
valuable support in the preparation of the examples.
Chapter 5
MULTIVARIATE POLYNOMIALS, MATRICES,

AND MATRIX-FRACTION DESCRIPTIONS
BYN.K.BOSE
1. Introduction
In this chapter the polynomial/polynomial matrix approach to multidimensional
systems theory is presented, based on ideals, varieties, and modules. The factor-
ization algorithms for bivariate polynomial matrices are now well developed but
not presented elsewhere in sufficient detail. The generalizations or lack of ade-
quate and full generalizations of the bivariate results to the multivariate case are
explained and, to the extent possible, justified, based on current knowledge. The
previous chapter should also prepare the reader to implement the algorithmic theory
presented here. The n-variate polynomial ring K[z], where K is an arbitrary but
fixed field and z = (ZI' Z2, ... , zn), is the base ring of interest. A result of crucial
importance in polynomial algebra is Hilbert's basis theorem, according to which
any ideal I C K[z] is finitely generated in the sense that there exist polynomials,
9i(Z) E K[z], i = 1,2, ... , k such that I = (91,92, ... , 9k). Furthermore, the
ascending chain condition is satisfied on K[z] i. e. if
It ~ 12 ~ 13 ... ~ h ~ ...
is an ascending chain of ideals of K[z], then there exists a positive integer N such
that IN = IN+! = IN+l .... A commutative ring that satisfies the ascending
chain condition is called a (Noetherian ring). The ring K[z] is a Noetherian ring.
A ring that satisfies a condition dual to the Noetherian condition: the descending
chain condition of ideals, is called an Artinian ring i. e. a ring is Artinian if
every descending chain of ideals is finite. An Artinian ring is also Noetherian. A
Noetherian ring is zero-dimensional (see Appendix A in Chapter 3) if and only
if the ring is also Artinian[194, Corollary 9.1, p. 227]. The ring K[ZI' Z2] is
Artinian, a fact which has implications in 2-D systems theory (see, for example,
Chapter 2 regarding feasibility of parametrization of stabilizing controllers for a
multivariable (MIMO) 2-D plant and, of course, Section 3 of this chapter).
129
K[z] is the simplest example of a graded ring, being graded by degree. Its
grading is the direct sum decomposition,
K[z] = 8 0 ffi 8 1 ffi "', such that 8i 8j C 8i+j, Vi, j ~ 0,

where 8k is the vector space of homogeneous polynomials (forms) of degree k.
This induces a representation, as in Definition 1.2 of Chapter 1. A polynomial
g(z) E K[z] with coefficients in field K defines a function g : K n -+ K; the
value of g at point a E K n is g(a). Thus K[z] is a ring of polynomial functions
on K n, the affine n-space over K, written An(K) or, simply An, when K is
algebraically closed. For an ideal I c K[z], the correspondence, I -+ V(I) is
provided by the affine algebraic set (variety),
V(I) = {a E Kn I g(a) = 0, Vg E I}.

For I C K[z], the set
rad(I) = {g E K[z] I grn E I, for some integer m,}

is an ideal, called the radical of I. An ideal I is called a radical ideal if I = rad( I).
According to Hilbert's Nullstellensatz, when K is algebraically closed and
I C K[z], then I(V(I)) = rad(I), which suggests that there is a bijective
correspondence between radical ideals and affine varieties.
The multivariate polynomial ring R ~ K[Zl, ... , zn] in indeterminates Zl,
Z2, ... , Zn over a field K of coefficients is commutative and it is an integral
domain. Every finite set of elements of R has a greatest common divisor (GCD)
and, therefore, it is a GCD domain (GCDD). A GCD is unique up to multiplication
by an invertible element of R. Two elements, c, d E R are relatively prime
(factor coprime or weakly coprime) if GCD(c, d) = 1 and have no common zeros
(zero coprime, strongly coprime) if there exist x, y E R such that xc + yd = 1.
A GCDD R is a Bezout domain if any pair of relatively prime elements in R
is also zero coprime, and in that case, the GCDD is principal. A GCD of any
finite set of elements belonging to K[Zl' ... ,zn] can be extracted. Polynomial
matrices with elements in K[z], however, behave quite differently when n, the
number of indeterminates varies. In the case of K[Zl], a principal ideal domain,
multivariable (MIMO) systems theory via the polynomial matrix approach has
been in existence for over thirty years. This chapter documents the bivariate
polynomial matrix counterpart, a consequence of more recent research and not
completely documented in texts or monographs, of the earlier univariate results.
The possibilities and difficulties encountered in generalization to the trivariate case
is explained and illustrated through selected examples.
The issues of factoring a multivariate polynomial a(z) E K[z], where K is
an arbitrary but fixed field, as a product of irreducible factors, tests for relative
primeness and greatest common divisor (GCD) extraction from a specified finite
set of polynomials belonging to K[z] have been presented in [5, Chapter 1]. The
Multivariate Polynomials, Matrices, and Matrix-Fraction Descriptions 131
GCD extraction by the primitive factorization scheme for multivariate polynomials

is presented here as a prelude to the important result on primitive factorization of
a bivariate polynomial matrix [39]. This result can be used to extract the greatest
common right (left) divisor (GCRD (GCLD» of two bivariate polynomial matrices.
The questions of existence, uniqueness and construction of the GCRD (GCLD)
become much more complicated and difficult to answer for polynomial matrices
in three or more variables. The constructive approach towards bivariate matrix
fraction description (MFD) in Chapter 2 involves factors with elements belonging
to the field of fractions associated with a general ring instead of a Bezout domain.
It is also possible to get the bivariate MFD with polynomial matrix factors by using
the primitive factorization scheme for bivariate polynomial matrices presented in
this chapter.
LetKqXP[z] = KqxP[Zl, Z2, ... , zn] be the ring of (q xp) n-variate polynomial
matrices whose elements belong to the n - variate polynomial ring R = K[z]
where K is an arbitrary but fixed field. Let K (z) be the field of rational functions
(which equals the quotient field of K[z] ) in Zl, Z2, ... ,Zn with coefficients in K.
Let K s (z) be the set of rational functions in K (z) whose denominator is zero-free
in the closed unit polydisc un and let K:"xt (z) be the set of (m x l) matrices with
entries in Ks(z). Let It and Ot,m denote, respectively, the l x l identity matrix and
the l xm null matrix. The subscripts in these matrices will be used only when
necessary for notational clarity.
2. Relative Primeness and GCD Extraction from Multivariate

Polynomials
DEFINITION 5.1 An element b(Zl' Z2, ... ,zn)/a(zl, Z2, ... ,zn) E K(Zl' Z2, ... , zn)
is said to be of reduced form ifpolynomials b(Zl' Z2, ... , zn), a(zl' Z2,··., zn) are
relatively prime in K[Zl' Z2, ... , zn] i.e., are devoid ofcommon polynomial factors
other than units.
2.1 TESTS FOR RELATIVE PRIMENESS
To test for n-variate polynomial relative primeness, using the classical results
based on the theory of resultants, the numerator and denominator polynomials are
each written in recursive canonical form in the main variable, say Zl:
Tl
b(Zl' Z2,··· ,zn) = L bk(Z2, Z3,"" zn)zf

k==O
(5.1)
ml
a(zl' Z2, ... ,zn) = L ak (Z2' Z3, ... ,zn)zf·
k==O
Without any loss of generality, it may be assumed that the polynomial sequence
{bo, ... , bTl' aO, ... , amI} is devoid of a common factor. It is also assumed that
bk's and ak's belong to a unique factorization domain (UFD) D[Z2' Z3, ... ,zn],
and that bTl (Z2' Z3, ... ,Zn) ¢ 0 and amI (Z2' Z3, ... ,Zn) ¢ 0, rl > 0, ml > O.
Form the matrix R of order rl + ml involving bk'S and ak's, the indeterminates
not being shown for brevity, as follows.
bTl bTI - 1 bl bo 0 0 0
0 bTl bTI - 1 b1 bo 0 0
0 0 bTl bn - 1 bl bo 0
........................................................
R= 0 0 0 bTl bTI - 1 ... b1 bo (5.2)
0 0 amI aml-l al ao
o amI aml-l ao o
amI aml-l o o
R is referred to as a Sylvester or inner matrix.
THEOREM 5.1 With amI ¢ 0, bTl ¢ 0, rl > 0, ml > 0, b(Zl, ... , zn) and
a(zl, ... ,zn) in (5.1) are relatively prime if and only if the resultant det R ~
r(z2' Z3, ... , zn) is not identically equal to O. Without loss of generality, it is
assumed that the sequence {bo , ... , bTl' ao, ... , amJ is devoid ofa commonfactor.
Proof Observe that because the sequence {bo, ... , bTl' ao, ... , amI} is devoid of
a common factor, any factor common to the two given polynomials must involve
the variable Zl. Consequently, the given polynomials in (5.1) have a noncon-
stant common factor if and only if there exist two polynomials f (Zl' ... , zn) and
g(Zl' ... ,zn) such that
with
6Z1 [! (Zl' ... ,zn)] < 6Z1 [b(Zl' ... , zn)] (5.4)
6Z1 [g(zt, ... ,zn)] < 6Z1 [a(zl, ... ,zn)].
After writing all polynomials in recursive canonical form in the main variable Zl,
(5.3) leads to a system of equations which has a nontrivial solution for f (Zl , ... , zn),
g(Zl, ... , zn) if and only if det R == O. Therefore a(zl, ... , zn), b(Zl, ... , zn)
are relatively prime if and only if det R ¢ O.
Note that det R == 0 also when amI == 0 and bTl == O. The test for absence of
a common factor in {bo,···, bTl' ao, ... ,amI} is a special case of the result just
given in the sense that a lesser number of indetermina~s are involved.
2.2 PRIMITIVE FACTORIZATION ALGORITHMS FOR GCD EXTRACTION
A primitive factorization algorithm was developed by this author in 1976 [5],

based on the use of the Sylvester matrix or inner matrix. Without any loss of
generality, it can be assumed that the set of bk'S and the set of ak 's in (5.1) each has
no nontrivial common factor. In case this assumption of primitivity with respect
to main variable Zl in a(zl' Z2,· .. , zn) and b(ZI' Z2, ... , zn) is not satisfied, the
content of each can be found by applying repeatedly the test to be discussed here
on the (n -l)-variate polynomials (ao, al, ... , amI)' when finding the content of
a(zl' ... ,zn). It is well known that
GCD(a, b) = {GCD [cont (a) , cont(b)]} {GCD[pp(a), pp(b)]}. (5.5)
Therefore, unless stated otherwise the given polynomials in (5.1) will be assumed
to be primitive.
2.2.1 Multivariate GCD extractionfrom the Sylvester matrix

Running parallel to the established results in the single-variable case, the inner-
wise [14] matrix [~(a, b)i,j] associated with (5.1) is written as follows:
amI aml-l aml -2 amI +l-i-j

0 amI aml-l am1 +2-i-j
0 0 amI aml +3-i-j
[~(a,bkj] ~ ....................................................................... (5.6)
0 0 brl b3+rl-i-j
0 brl brl - 1 b2+rl-i-j
brl brl-l bn - 2 b1+n-i-j
In (5.6), there are i rows containing the ak's, j rows containing the bk'S, and
ai == 0, i < 0, bj == 0, j < 0. Next, consider the following n-variate subresultant
matrix:
~(a, b, Zr)i,j == det (5.7)

b3+rl -i-j
b2+rl-i-j
obtained by replacing only the last column in (5.6) by the last column defined in
(5.7). The new column can be expanded as follows:
~(a, b, Zdi,j = f(zl, Z2,···, Zn)a(zl, Z2,··· ,zn) +

+ g(Zl' Z2,···, Zn)b(Zl, Z2,···, zn) (5.8)
where the degrees in Zl of f (Zl' ... , zn) and g( Zl , ... , zn) are, respectively i - 1
and j - 1. V sing the fact that the GCD of two primitive polynomials is primitive,
the result stated next follows. Note that det [ll (a, b)rl ,ml) is the resultant of the
two polynomials in (5.1).
THEOREM 5.2 SupposeJortwoprimitiven-variate polynomials, a(zl' Z2, ... , zn)

and b(zl, Z2, ... , zn) in (5.1), the associated [ll(a, b)i,j) and ll(a, b, Zdi,j are as
defined in (5.6), (5.7). Also, define ll(a, b)i,j == det[ll(a, b)i,j)' Then if,
ll(a,b)rl,ml == ll(a,b)rl-l,ml-l == ... == ll(a,b)Tl-j+1,ml-j+l == 0 (5.9)
and
ll(a, b)rl-j,ml-j ~ 0, (5.10)
then GCD( a, b) is oj degree j in Zl and,
GCD(a,b) == pp{ll(a,b,zdrl-j,ml-j}. (5.11)
Conversely, if GCD( a, b) is oj degree j in Zl then it is given by (5.11), and (5.9)

and (5.10) are valid.
The proof of the above theorem can be constructed from the arguments given a
priori.
Example 5.1 It is required to extract the GCD from
a(zl' Z2, Z3) = z~zt + (z~ + 2Z~Z3)Zr + (1 + Z~Z3 + z~z~)zr

+ (Z2 + 2z3)Zl + (Z2Z3 + z~)
b(zl, Z2, Z3) = (Z2 + z:)z? + (2zg + Z2Z: + 2Z2Z3 + Z~)Zl
+ (zg + z2zg + Z2Z~ + Z~Z3),
if they are not relatively prime.
Considering Zl as the main variable,
cont{a(Zl' Z2, Z3)} = GCD{a4' a3, a2, al, ao}

cont{b(Zl' Z2, Z3)} = GCD{~, bl, bo}
The fact that conti a(zl, Z2, Z3)} = 1 is verifiable by inspection, since a4
Z2 . Z2, ao = Z3(Z2 + Z3) are relatively prime. Also, cont{b(zl, Z2, Z3)}
GCD{b2, bl , bo} = (Z2 + z1). Therefore,
GCD[cont(a) , cont(b») = 1. (5.12)
With respect to the main variable Zl
pp{a(Zl,Z2,Z3)} = a(zl,z2,Z3)
PP{b(Zl' Z2, Z3)} = zr + (Z2 + 2Z3)Zl + (Z2Z3 + z~)
For pp{a(zl,z2,Z3)} and PP{b(Zl,Z2,Z3)}, the matrix corresponding to (5.6) is

formed. It is found that .6.[pp(a), pp(b)h,4 == 0, implying that pp(a) and pp(b) are
not relatively prime. Also,
.6.[pp(a), pp(b)h,3 == °
.6.[pp(a), pp(b)]O,2 = 1 ~ 0,
implying that the degree in Zl of the g.c.f. of pp(a) and pp(b) is 2. The GCD of
pp(a) and pp(b) is
GCD[pp(a), pp(b)] = ppdet [~ Z~~~~~}]

(5.13)
= z? + (Z2 + 2z3)Zl + (Z2Z3 + z~)
= (Zl + Z2 + Z3)(Zl + Z3).
Using (5.5), it follows from (5.12) and (5.13) that
GCD[a(zl, Z2, Z3), b(Zl, Z2, Z3)] = (Zl + Z2 + Z3)(Zl + Z3)

2.3 BEZOUT MATRIX BASED METHOD
The greatest common divisor (GCD) of two multivariate polynomials could also
be extracted from the corresponding Bezout matrix. This result is of importance
because it has been shown that Bezoutiants provide computational algorithms
which are often, if not always, superior to those provided by either the Euclidean
algorithm or bigradient determinants.
An element !(Zl, Z2, ... , znr of K[Zl, Z2, ... , zn] can be written in recursive
canonical form with ZI as the main variable. In that case, !(ZI, z\zd, where
Z\Zl = (Z2, Z3, ... , zn) is said to belong to K[Z\Zl][Zt], where K[Z\Zl] is the
ring of multivariate polynomials in (Z2, Z3, ... , zn) and !(Zl, z\zt) is expressible
as
nl
!{Zl,Z\Zt} = L ak(z\zt}zf, (5.14)
k=O
when nl is the partial degree of a(zl, z\zt} in the variable Zl. Similarly, consider
g(Zl, z\zt} to be another element of K[Zl, Z2, ... , zn] and express it as a member
of K[Z\Zl][zd. If ml ~ nl is the partial degree in Zl of g(Zl, z\zt}, then
= L bk(Z\zt}zf,
ml
g(Zl,Z\Zt} (5.15)
k=O
Obviously, in (5.14) and (5. 15), ak (z), k = 0, 1, ... , nl andb,{z\zt}, l = 0, 1, ... ,

ml, are elements of K[z\zd. The GCDofthe (nl +1) polynomials ak(z\zt}, k =
0,1, ... ,nl, is the content of f{zI, z\zt} and the quotient polynomial obtained
after dividing f{Zl, z\zt} by its content is the primitive part of f{Zl, z\zt}. Since
GCD(J,g) = GCD{contf,contg) x GCD{ppf, ppg) (5.16)
and since the first term on the right-hand side of (5.16) involves the computa-
tion of the GCD of (n - 1) variate polynomials, for our purpose, it is sufficient
to consider the problem of extracting the GCD from two primitive polynomials.
Therefore, f{Zl, z\zt} andg{zl' z\zt} will each be assumed to be primitive when
considered to be an element of K[Z\Zl][Zl]. Note that a polynomial primitive in
K[Z2, ... ,Zn][Zl] need not be primitive in K[Zl' Z3, ... , Zn][Z2]. Furthermore, we
eliminate a trivial case by requiring that ao{z\zt} and bo{z\zt} are not identically
zero. A procedure [5] for extracting the GCD of f{Zl, z\zt} and g{Zl' Z\Zl) was
given by using the Sylvester or inner matrix and then computing the bigradient
determinants or inner determinants. Here it is confirmed that the same is possible
by using the Bezout matrix.
2.3.1 Multivariate Polynomial GCD Extractionfrom Bezout Matrix

Remember that, without loss of generality, f{Zl, z\zt} and g{Zl' z\zt} are
primitive in K[z\zIJ[zIJ, and a not too interesting trivial case is avoided by re-
quiring ao{z\zt} and bo{z\zt} to be not zero polynomials. For the sake of brevity,
ak{z\zt}, b,{Z\Zt} will be denoted simply by ak and b, and let
(5.17)
Associate the Bezout matrix B(J, g)in (5.18) below with the polynomials in (5.14)
and (5.15). B(J,g) is symmetric and square of order nl.
U (5.18)
The first result is written below as a theorem, which is well-known for univariate
polynomials.
THEOREM 5.3 The nullity oj B(J,g), is equal to the partial degree in the main
variable Zl oJ the greatest common divisor oj f{Zl, z\zt} and g{Zl' z\zt}.
Suppose that the rank of B(J,g) is r, so that its nullity is (nl - r). Next, define
the matrix C(J,g) of order r in (5.19) below. For the sake of brevity, the matrix
in (5.19) has been written to correspond to the case when r = 3.
laobll lao~1 laob3 1+ laob41 zl

laob2 1 laob3 1+ lal~1 laob41+ lal b3 1+
C(j,g) = (l aob51 + la l b41) zl
laob3 1 laob41 + lal b3 1 laob51+ lal b41+ la 2 b3 +
1
(Iaobtil + lal b51+ la2 b41)zl

(5.19)
THEOREM 5.4 The greatest common divisor of the primitive multivariate polyno-
mials f(Zl, z\zr) and g(Zl' z\zr) is given by the primitive part in K[Z\Zl][Zl] of
the determinant of matrix C (j, g), patterned after the matrix for the r = 3 case
in (5.19).
The proof of Theorem 5.4 (and also Theorem 5.3) may be constructed by ex-
ploiting the relationship between a Sylvester matrix and a Bezout matrix. The two
are linked through a non-singular linear transformation matrix whose elements
belong to K[Z\Zl]. This same matrix links the matrix used for OCD extraction in
[5] to the matrix C (j, g) above.
Example 5.2 Consider bivariate polynomials f(Zl, z\zr) and g(Zl' z\zr) shown
below to be elements of R[Z2][Zr].
f(Zl, Z2) = Z2 + (Z2 + 1)Zl + z? (5.20)
g(ZI, Z2) = 1 + (Z2 + 1)Zl + Z2Z? (5.21)

Comparing (5.20) and (5.21) with (5.14) and (5.15), respectively,
aO(Z2) = Z2, al(z2) = Z2 + 1, a2(z2) = 1

bO(Z2) = 1, bl (Z2) = Z2 + 1, b2(Z2) = Z2
Clearly, both sets of polynomials are primitive in lR[Z2][Zl]. The Bezout matrix
B(j, g) in (5.18) for this case is,
z2 -1 z2 -1 )
B(j,g)= ( z~-1 z~-1
the nullity of B(j, g) is 1 and its rank is 1. The partial degree in Zl of the OCD,
therefore, is 1 by Theorem 5.3. The OCD can be computed by Theorem 5.4 to be
PP (Iaobll + laob2 lzr)

PP ((z~ - 1) + (z~ - 1)zr)
Zl +1
3. Polynomial Matrix Primitive Factorization in the

Bivariate Case
For notational simplicity, it is convenient to replace, in this Section 3, K[Zl' Z2]
by K[z, w] (with independent complex variables z, w),and this is done.
DEFINITION 5.2 Let A(z, w) E Kmxl[z, w]. It is assumed to have (normal) full
rank and to be such that m ::; i. Then A(z, w) is said to be primitive with respect
to K[w] if and only if the (m x m) minors of A(z, w) have no common divisor
that is a polynomial of w only.
An equivalent but more useful characterization is given next.
DEFINITION 5.3 A(z, w) is primitive with respect to K[w] if and only if
Ap(z, w) = A(z, w) modulo p(w)

has full rank over Kp[w] = K[z, w]jp(w)K[z, w] for all irreducible polynomials
p(w) E K[w].
Let K be a field. Consider a(z, w) E K[z, w],
n
a (z, w) = L aiwi, where ai = ai(z) E K[z].
i=O
FACT 5.1 : Let p(z) E K[z]. Then one of the two possibilities below hold.
I a(z, w) == 0 mod p(z) (i.e. p(Z)lai(Z), i = 0, ... , n)
nl
2 a(z, w) = E (3iwi mod p(z)
i=O
where (3i = (3i(Z) E K[z], deg (3i(Z) < deg p(z), nl ::; nand (3nl [z] # O.
The (3i(Z)'S can be obtained by means of the division algorithm:
Note that nl = n if and only if p(z) does not divide an(z).

The following notation will be of importance in the algorithm described later:
given any polynomial a(z, w) in K[z, w] we denote its representation in Fact 5.1
by a(z, w), i.e.,
nl
a(z,w) == L(3iwi
i=O
and a(z, w) == 0, if and only if a(z, w) == 0 mod p{z).

Example 5.3 If p(z) = Z2 + 1 E R[z) and if

a(z,w) (Z3 + z)w2 + (z3 + 3z + l)w + z - 1
b(z,w) (Z2 + 1)w2 + z3 + z2 + z + 1,
then a(z, w) = (2z + l)w + z - 1, b(z, w) = O.
Assume now that p(z) is irreducible in K[z). Then, since (3nt (z) is a nonzero
polynomial of degree less than deg p(z) (assuming a(z, w) # 0), therefore (3nt (z)
andp(z) are relatively prime. Therefore by the Euclidean algorithm one can find
x(z) and y(z) in K[z) such that
x(z){3nt (z) = 1 - y(z)p(z)
Therefore, if a(z, w) # 0, then

x(z)a(z, w) = a*(z, w) mod p(z)
where a*(z, w) is monic (i.e. the coefficient of the highest power of w is the
constant 1). Hence if aj(z, w) is any polynomial in K[z, w) with
degw aj(z, w) ~ degwa*(z,w) = degwa(z,w)
then there exist qj(z, w) and Tj(Z, w) in K[z, w) such that
where either
Therefore,
aj(z, w) = qj(z, w)x(z)a(z, w) + Tj(Z, w) modp(z)

where either
Tj(Z, w) == 0 or degw Tj(Z, w) ::; degw Tj(Z, w) < degw a(z, w).
Example 5.4 Letp(z) = z2 + 1 E lR[z),
a(z, w) = zw 2 + (z3 + l)w + 1, aj(z, w) = w3 + z2 W + z.
Then
a(z, w) = zw 2 + (1 - z)w + 1, aj(z, w) = w 3 - w + z.
Take x(z) = -z; then

x(z)a(z, w) _Z 2W 2 + (Z2 - z)w - Z
w2 - (1 + z)w - z mod (z2 + 1).
So take a*(z, w) = w2 - (1 + z)w - z. Then
aj(z, w) = w3 - W + z = (w + z + 1) (w 2 - (1 + z)w - z) +
"'----...----'
qj(z,w)
+ ,(z2 + 3z)w + z2 + 2z.,

"
Tj(Z,W)
Therefore,
Tj(Z, w) = (3z - l)w + 2z - 1
and
aj(z, w) = -z(w + z + l)a(z, w) + (3z - l)w + 2z - 1 mod (z2 + 1).

Indeed it can be checked that
aj(z, w) = -z(w + z + l)a(z, w) + (3z - l)w + 2z - 1

+(z2 + 1) (w 3 + (z2 +z)w2 + (z3 +z2 - z)w + 1).
3.1 PRIMITIVE FACTORIZATION OF BNARIATE POLYNOMIAL MATRICES
If M = [mij] is a matrix with entries mij in K[z, w], where K is an arbitrary

but fixed field, we denote by M the matrix [mij].
THEOREM 5.5 (Primitive factorization theorem [39, Guiver-Bose j): Let A be an

m x £ normal full rank matrix with entries in K[z, w], m ~ £. Let g{z) E K[z]
be the content of the greatest common divisor of the set of minors oforder m (i.e.,
major determinants) of A. Then there exists L, R, m x m and m x £ matrices,
re!Jpectively, with entries in K[z, w], such that A = LR and det L = g(z).
Proof Letp(z) be a factor of g(z) with p(z) irreducible in K[z]; the irreducible
factors will be extracted one at a time.
Step 11: Let Al = A. If a row of Al is zero modp(z), say row SI, then
1
o
1
sthrow ........... p (5.22)
1
o 1
is a left divisor of A and we terminate the algorithm here.

If no row of Al is zero modp(z) (i.e. if no row of Al is identically zero) then
proceed to Step Ill.
Step Ill: Pre and postmultiply A 1 by premutation matrices Un and VI respectively
-
so that
UnAIVI = [:2~;'~) . X 1 (5.23)
am(z,w)
where a(z, w) =1= 0 and either aj(z, w) == 0, j = 2, ... , m or deg w a(z, w) ~

degwaj(z,w). Postmultiplication by VI simply ensures the first column is not
identically zero. Now let x(z), qj(z, w), Tj(Z, w) be as in the scalar polynomial
case for j = 2, ... , m (if aj(z, w) == 0 take qj(z, w) = Tj(Z, w) == 0). Then
premultiplication by the unimodular matrix
1
-xq2 1 0
ViI = 1 (5.24)
0
-xqm 1
gives
a(z,w)
T2(Z, w)
VnUnAI VI = X mod p(z) (5.25)
Tm(Z,W)
where
degw Tj(Z, w) < degw a(z, w) or Tj(Z,W) == O. (5.26)
Now premultiply by U12 to move the nonzero polynomial of least degree in w

to the (1,1) position. In this way the degree in w of the polynomial in the (1,1)
position is reduced and continuing the process in the manner above leads eventually
to
d11 IX
o mod p(z) (5.27)
o
where Ui is of the form
(5.28)
and d11 = d ll (z, w) =1= 0 since the first column of Ai V is not zero modp(z).
End of Step III; proceed to Step 12.
Step Ik.: From Step lI(k-l) one has
d11
x
o mod p(z) (5.29)
o
where Uk-i, Vk-i are unimodular and
djj = djj(z, w) =1= 0 for j = 1, ... , k - 1 (5.30)
If a row of Ak, say row Sk, is zero, then
1
o
U- i
k-i p .......... . row k -1 + Sk (5.31)
1
o 1
is a left divisor of A the algorithm terminates here. If no row of Ak is identically

zero proceed to Step Ilk.
Step Ilk: By proceeding in the manner of Step III find unimodular matrices
W k , Zk such that
W.A.Z. [ ~..:...~... ~~;..::.'.. ] mod p(z) (5.32)
and dkk = dkk(Z, w) i= O. A k+1 is (m - k) x (l - k).

If Uk = (Ik-l E9 Wk)Uk- 1 Vk = Vk-l(Ik- 1 E9 Zk), then
* x
o mod p(z) (5.33)
where djj = djj(z, w) i= 0 j = 1, ... , k, and Uk, Vk are each unimodular. End
of Step IIk:; proceed to Step I(k+ 1).
The algorithm either terminates at one of the steps 11, ... , I(m-1) or one proceeds
to step 1m. In this case from Step lI(m-1), one gets
x x mod p{z), (5.34)

o dm-1,m-l
o 0 ...
where djj = djj(z,w) i= 0 j = 1, ... ,m-1 and where Am is a 1 x (l-m+1)

row vector.
It is claimed that Am must be identically zero. This is true since the minors of
order m of A and therefore of Um-l AVm-l are all divisible by p( z). This follows
in general from the Cauchy-Binet theorem but in our case it is easy to see since
Vm-l is just a permutation matrix with permutes the columns of A. But a non zero
entry d(z, w) of Am would give
[TI' d;,(Z,W)] d(z,w) (5.35)

as a non zero minor (mod p(z ), contradicting the fact that p(z) divides each mth
order minor of A. Therefore, Am == 0, which implies that
1
1 0
U- 1
m-l (5.36)
0 1
p
is a left factor of A. The proof of the theorem is now complete.
REMARK 5.1 In practice it is not necessary to permute the columns at all - just
work with the chosen column in its original position.
REMARK 5.2 Suppose p(z) is linear, p(z) = z - a. Then given any polynomial
q(z) in K[z] we can expand it as a polynomial in (z - a) with constant term q(a).
Thus
q(z) = q(a) mod p(z). (5.37)
So given a polynomial a(z, w) in K[z, w]
n
a(z, w) = L aiwiai = ai(z) E K[z] (5.38)
i=O
one has
n
a(z,w) = Lai(a)wi
i=O (5.39)
= a(a, w)
So iJp(z) is linear, the algorithm is essentially that presented by MorJ, Levy, and
Kung [128J. An important feature of the algorithm presented by Guiver and Bose
is that the primitive factorization is obtained only through computations in the
ground field.
REMARK 5.3 There is, of course, a similar primitive factorization theorem for
m ~ l where p( z) is a factor of the order l minors and an l x l matrix with
det p( z) is extracted on the right.
REMARK 5.4 The primitive factorization theorem holds when the coefficient ring
K[z] is replaced by any Euclidean domain D.
Example 5.5 Let
Z2 +z
(z - l)w z ]
-z (5.40)
w2 +z 2w -w+z
Then, detA = {z2 + 1)w{w 2 + (z2 + z + l)w - Z2 - z). We wish to extract a

left factor with detenninant z2 + 1. So we work mod z2 + 1.
W+Z+1 z-l
[ zw z ] (5041)
(z - l)w -z
{z + l)w w -w -w+z
2
Step 11: None of the rows of At are identically zero; so proceed to Step Ill.
Step Ill: We will reduce column 3.
Un = 13x3 (5042)
vn=[z{-wi+ z) ~ ~1] 0
(5043)
Vi1UuAl
w+z+1 z-l Z]
_ [ {z + l)w + (z + 1) {z - l)w + (z - 1) 0
- -zw 2 + w + z3 + z2 w2 - {Z2 - Z + l)w+ (z2 + 1)( -w + z)
+z3 - Z2
w+z+1 z-l z ]
[
= L--{_Z_+_1_)W_+_{_Z_+_1_)_{_Z_-_1)_W_+_{Z_-_1)---lA2
_ -zw 2 + w - (z + 1) w2 + zw - (z - 1)
00 mod z2 +1
(5.44)
Step I2: A2 has no rows that are identically zero. So proceed to step II2.
Step 112:
U2k = 12x 2 (5.45)
= [ -2{Z
V21 1
1 01] (5.46)
- l)(zw - z - 1)
TT U. A _ [{Z + l)w + z +
Y21 21 2 - -2{z + 1)
1
(z - l)w + z -
-2{z - 1)
1] mod z2 +1
(5.47)
(5.48)
-2{z -1) ] mod z2 +1

(z-l)w+{z-l)
(5.49)
(5.50)
mod Z2 +1 (5.51)
Therefore,
[ _1 (z - 1)(zw - z - 1) 1]
W 2 = V22 U22 V21 U21 = 1 - !(~ + l)(z - l)(zw - z -1) ~(w + 1)
(5.52)
U2 = [~ 0 0] [ ~ ~~]
o W2 -z{w - z) 0 1
1 o o
-~(z-l)x -~(z - l)x 1
x{zw - z - 1) - z{w . . :. z) x{zw - z -1)
1-!{w+1)(z-1)x 1- !(w + l)x ~(w + 1)

x{zw - z -1)- x{z - l)(zw - z - 1)
-~z{w - z)(w + 1)
(5.53)
+ l)(z
(z2 z - w) ]
~(z2 + l)(w + l){z - w)
(5.54)
where
P1(z,w) = t(3z 3 + 3z2 - Z - 1 + Z2W + w - Z3W2 - ZW 2)

P2(Z, W) = 2(3z4 + 3z2 - Z2 - Z - Z4W - Z3W - ZW+
w - Z3W2 + 2Z2W2 - ZW 2 + 2W2)
P3(Z, W) = !'(W + l)(3z + 3 + w - ZW 2)
P4(Z, W) = t(3z 2 + 3z + 2z2w + 2zw - 3w - Z2W2 - 2zw2
+3w 2 - ZW 3 + 2w 3)
u.2- 1 -- [
1
-1 _!'(w + 1) 00] 1
(5.55)
z(w - z) 1-1(w+1)x ~(z-l)x
X (z - 1){zw - z - 1) X (zw - z - 1)
Therefore
detenninant= - (z2 +1)
"
A- '[ -1
1
- z(w - z)
- ~ (~ +
1-~(w+1)x
1) Z2
~(z-l)x
~ 1 ]' x
(5.56)
x(z -l){zw - z -1) x(zw - z -1){z2 + 1)
Z2 +z
P2(z,w) (z - w)(z2 +
P4(Z,W) ~(w + l)(z - w)
1)]
where P1(z, w), P2(z, w), P3(z, w) and P4(Z, w) are as before.
3.2 GENERAL FACTORIZATION OF BNARIATE POLYNOMIAL MATRICES
Let E be an Euclidean domain. Let A be a normal full rank matrix of size

(m X f), m ~ £ (such matrices, when m ~ £ may be called wide matrices) with
entries in' E[Z2], where Z2 is a complex variable. Let gEE be any common factor
to the set of mth order minors of A. It has been shown by Guiver and Bose [39]
that A can be factored as A = LB, where L, B are matrices of size (m X m) and
(m x f), respectively, and det L = g. This type of factorization will be called a
determinantalJactorization of A with respect to the common factor 9 and E[Z2] is
called a determinantalJactorization domain (DFD). Examples of E include the ring
of integers and K[zI] , where K is a field and Zl is a complex variable, independent
of Z2. Since E is an UFD, for the sake of brevity in ensuing discussions, 9 may be
replace by d(Z2), dEE, where d(Z2) denotes the greatest common divisor ofthe
mth order minors of A.
THEOREM 5.6 Suppose that A is a m X £ matrix (m ~ £) with entries in E[Z2],

where E is an Euclidean domain. Let d(Z2), dEE, denote the greatest common
divisor oj the mth order minors oj A. Then A can be Jactored as the product
A = LB of compatible matrices with entries in E[Z2] with det L = d(Z2)'
Proof The GCD d(Z2) =c.p(Z2) where C E E and p(Z2) are, respectively, the
content and primitive part of d(Z2) E E[Z2]. From Theorem 5.5, it follows that
A = AlA2' where det Al = c, and AI, A - 2 have entries in E[Z2]. There exists
a matrix V with entries in E[Z2] such that the corresponding Hermite form is
where det VEE, F = FlF2 is a right primitive factorization of F i.e. det Fl =

P(Z2) is a primitive polynomial in E[Z2]' Then,
A _ Fl [F2 1O]Adj V
2 - detV
has entries in E[Z2] as shown next. Let 9 be a prime divisor of det VEE.
Then since E is an Euclidean domain (and, therefore, a principal ideal domain),
(g) is a maximal ideal and E / (g), the set of elements modulo g, is a field. Now,
Fl [F2 IO]Adj V == 0 in E / {g}[Z2]' But det Fl is primitive in E[Z2] and, therefore,
det Fl ¢ 0 in E / (g) [Z2]. Therefore,
[F21 O]Adj V == 0 in E/{g)[Z2] ,

implying that 9 divides every entry of [F2 IO]Adj V. Cancel this common factor and
repeat the argument for another prime divisor of det V. Eventually, it is possible
to conclude that
det VI[F2 1O]Adj v.
L::,.
So, to conclude the proof define L = AlFl and
B ~ [F2 I O]Adj V
detV '
because clearly, det L = det Al det Fl = c.p(Z2) = d(Z2)'
Next, let A, d(Z2) be as in the preceding theorem, with d(Z2) = c.p(Z2) rep-
resenting a primitive factorization of d(Z2) in E[Z2]. Suppose that dl (Z2) divides
d(Z2) so that
where CiPi(Z2) is a primitive factorization of di(Z2) in E[Z2] , i = 1,2. Thus,
C = ClC2, P(Z2) = PI (Z2)P2(Z2).

THEOREM 5.7 There exist matrices L1, Bl of size (m x m) and (m x i), respec-
tively, with entries in E[Z2] such that A = LIBI and det Ll = d l (Z2)'
Proof By Theorem 5.5,

Furthermore, there exist matrices U and V, det U E E, det VEE such that the
corresponding "Smith form" is
where D is diagonal and, therefore, decomposable as
where,
Therefore,
U A2 = D l D2[D o I O]V- l ,
and as in the previous theorem, [Do I O]V- l has entries in E[Z2] and, therefore, so
does D2[Do I o]V- l . A left primitive factorization of D2[Do I O]V- l yields
D2[Do I O]V- l = L2B 2,

where det L2 E E, and B2 is primitive. Now perform a left factorization of DlL2
so that
where B3B2 is primitive and,
It is then possible to show by argument similar to that in the preceding theorem

that U- l L3 has entries in E[Z2]. If gEE is irreducible (is prime) in E and divides
detU, then
(Adj U)L3B3B2 == 0 in E / (g)[Z2].
But since B3B2 is primitive in E[Z2] it is possible to find an mth order submatrix
M of B3B2 such that det M :t 0 in E / (g)[Z2]. Consequently, (Adj U)L3 == 0 in
E/(g)[Z2] and so 9 divides (Adj U)L3. Therefore,
A = AlU- l L3B3B2,
where det(AlU- l L 3) = c, det B3 = Pl(Z2). By, primitive factorization algo-

rithm,
Al U- l L3 = C l C2,
where det C l = Cl, det C2 = C2. Also, by primitive factorization theorem,
C2B 3 = B4C3,
wheredet B4 = Pl(Z2), det C 3 = c2.Finally,A = LlB l , whereLl = ClB4, Bl =

C3B2, det Ll = ClPl (Z2).
A generalization theorem in E[Z2] is given next, without loss of generality, in

the case of a square matrix A of order m with entries in E[Z2].
THEOREM 5.8 The square matrix A of order m with entries in E[Z2] can be fac-
toredas
i=l
where detAi = di (Z2) = CiPi(Z2) , di E E,Pi(Z2) is primitive in E[Z2] and
detA = I1~=1 ~(Z2).
Proof.Let A = BlB2 be a primitive left factorization of A i.e. det Bl E E and
det B2 is a primitive polynomial in E[Z2]. Then, there exist matrices U, V with
det U E E, det VEE such that
UB2V= D = DoDl···Dk,
where D is diagonal, Do E E, det Di = Pi(Z2). Therefore,
B 2V = U-lDoD l · ··Dk ·
By arguments similar to that in the preceding theorem, U- l Do has entries in
E[Z2] and by primitive right factorization with det Ul E E and det Fl = PI (Z2),
it follows that
U-lDODl = FlUl ·
By repeating sequentially the preceding arguments it follows that,
Ul D2 = F2U2, U2 Da = FaUa, ... , Uk-lDO = FkUk ,
where detFi = Pi(Z2), and det Uk E E. So,
B2V Fl F2 ··· FkUk,
B2 = (fI Fi)
z=l
UkV- l ,
where Uk V-I has entries in E[Z2] by a similar argument, and is also unimodular.
Therefore
A~ B, (g F;) U.V- ~ (g Ci) (g Fi) (U.V-

1 1 ),
where det Ci =Ci. Now by interchanging left and right primitive factorizations,
the required factorization,
i=l
det Ai = CiPi(Z2) = di(Z2) is obtained where the entries of each Ai are in E[Z2].
o
4. Multivariate Polynomial Matrix Factorization

The absence of primitive factorization for n ~ 3 is illustrated by a simple
counterexample due to Levy [195]. This example is cited next, without proof. An
earlier counterexample is due to Youla and Gnavi [40], [5].
Example 5.6 Consider the trivariate matrix
c = [ZiZ2 0]
Z3
whose determinant is ZlZ3. It is possible to show that a matrix factorization ofthe
form C = AB, where the determinants of A and B are, respectively Z3 and Zi,is
not possible.
4.1 VARIOUS TYPES OF MULTIVARIATE POLYNOMIAL MATRIX

FACTORIZATIONS
Considerable research has been conducted in the case when R = K[Zl' ... ,zn]
[40], [110], [111] is the n-variate polynomial ring where K is an arbitrary but fixed
field of coefficients. R = K[Zl], is a principal ideal ring and K[Zl] has been the
setting for the well-understood theory of the polynomial approach to multivariable
(multi-input multi-output (MIMO» systems theory initiated by Rosenbrock [196].
The case when R = K[Zl' Z2] has been completely tackled, as detailed in Section
3, where it was shown that polynomial matrix factorization could be constructed
via computations in the ground field K without the need of any algebraically
closed extension field. Factor and minor primeness are equivalent in this case but
zero primeness is not implied by either, even though the cardinality of the set of
common zeros of the maximal minors is always finite for this bivariate case. In
fact, the constructive results in [39] hold even when the matrix elements belong to
R = E[Zl], where E is any Euclidean domain.
4.2 MODULES
DEFINITION 5.4 Let R be a commutative ring with identity. A module over the ring
R, also called R - module, is an Abelian group M, usually written additively,
together with a map (r, m) -+ rm for R x M to M sati.sfying the following
conditions:
(a) r{mi + m2) = rmi + rm2, (b) {ri + r2)m = rim + r2m, (c) (rir2)m =
ri{r2m), (d) 1m = m.
Sometimes, the R - module just defined, is called a left R - module. There
is a similar definition for a right R - module in which the elements of R occur
on the right. Let A be an integral domain. The integral domain K[z], where K is
a field is of prime interest here.
DEFINITION5.5 A submodule of an A - module M is a subset of M which is an

A - module in its own right.
DEFINITION 5.6 The torsion submodule T(M) of an A - module M is the set

of all m E M for which there is a (nonzero divisor) rEA, r =I=- 0 such that
rm = O. M is called a torsion -free module ifT(M} = (0) and a torsion module
ifT(M} = M.
Since A = K[z] is Noetherian, therfore every submodule over A = K[z] has

a finite generating set.
DEFINITION 5.7 An A - module that has a basis is called a free-module over

A. A projective-module is a direct summand offree modules.
A free module is projective and a projective module is torsion-free. An important
problem in commutative algebra is to obtain the set of all polynomial solutions,
called the syzygy-module, for a system of homogeneous linear equations with
multivariate polynomial coefficients.
DEFINITION 5.8 A .syzygy on a set of polynomials {gI,'" ,gs} is an s-tuple

(al,'" ,as) suchthatalgl +.. .+asgs = O. The set ofall.syzygies on {gl,'" ,gs}
forms a module called the (first) syzygy module on {gl, ... ,gs} and is denoted by
SYZ(gl, ... , g s}. The kth .syzygy module is defined inductively to be the module of
the !Jyzygies of the (k - l}th syzygy module (see also 5.9 below).
The module M is torsion-free if and only if M is a first syzygy and reflexive if and
only if M is a second syzygy. [197, Theorem 2.17]
A module homomorphism f : M -+ N between A-modules M and N satisfies
f(ax} = af(x), a E A, x E M and is called an A-homomorphism. The sub-
A-module Imf = f(M) of L is called the image of f and the sub-A-module
Kerf = {x I x E M, f(x) = O} of M is called the kernel of f. Cokerf = L/lmf
is called the cokernel of f.
DEFINITION 5.9 A sequence F of A-homomorphisms ofA-modules Mn (n E Z)
F : .. ,
In+l M n
M n+l --=----:.....t ~
In M n-l ~ ...
is called an exact sequence

if Imfn+1 = Ker f n,for all n.
The image of the map fn is the kth syzygy of module M = coker h.
The set of all A-homomorphisms from an A-module M to an A-module N,

denoted by HomA(M, N} is called the module of A-homomorphisms. There is a
canonical isomorphism,
homA(A, N) ~ N : f -+ f(1).
To, x E N, the linear map A --+ N : a --+ ax, a E A, x E M, can be associated

that leads to a --+ af(l) = f(a).
DEFINITION 5.10 If M is a left A-module, then the module M* ~ homA(M, A)
is called the dual of M. A module M is called reflexive if M ~ M**.
The classification for modules over Noetherian rings is
FREE ~ PROJECTIVE ~ REFLEXIVE ~ TORSION-FREE

Classification of modules that are not finitely generated, not of direct interest
here, is given in [198, Volume 1, Chapter 2].
DEFINITION 5.11 [194, pp.617-618] Consider a sequence F offree A-modules Fi
and maps h given by
1m+! F.m
F : . . . F.m+ 1 --=--.:.......t 1m
~
F.m-l ----+ ...
with the properties that fm+! maps Fm+! onto the kernel of fmfor each m ~ 1,
and that M is the cokernel of h. A sequence offree modules Fm and maps f m with
these properties is called a free resolution of M. If the F:n are merely projective,
then the sequence is called a projective resolution. If for some finite m, we have
Fm+l = 0 but Fk t=- 0, 0 ~ k ~ m, then F is a finite resolution of length m.
Every module has a free resolution, and, particularly, every module over a
polynomial ring has a finite free resolution. Free or projective resolutions serve to
compare a module with free modules. The projective dimension of module M is
the minimum of the lengths of projective resolutions of M (it is 00 if M has no
projective resolution.) A very discursive account of free or projective resolution
of R-modules over a polynomial ring R = K[z] and projective dimension of a
finitely generated module is given in [194, Chapter 19].
DEFINITION 5.12 For a module M, the rank of M is defined as the vector space
dimension of the sub.space generated by the generators of M over the quotient field
of the base ring.
Consider the n-variate polynomial matrix F E nmxt, n = K[z], where K
is the algebraic closure of field K. Consider the associated finitely generated
module M = coker(F) = n ixt / n Ixm F. Systems that admit a normal full rank
representation are those whose projective dimension is at most one [107], [110].
Therefore, it becomes necessary to consider the non full rank case for polynomial
rings in more than two indeterminates. Therefore, let rankF = r, r ~ £ ~ m.
The module ker (F) is generated by the rows of a matrix G E n mxt i.e.
ker(F) = nIxmG,
where rank(F)+ rank{G) = r. Therefore, in the normal full rank case (r = i),
rank(F) = l -{=} rank(M) = l - rank(F) = o.
That the above equivalence in tum is necessary and sufficient for M to be a torsion
module follows from the Fact below.
FACT 5.2 If R is an integral domain and M is an R-module, then M is a torsion

module if and only if the rank of M is O.
Therefore, the normal full rank case does not permit consideration of torsion-
free modules like free and reflexive modules.
Example 5.7 Consider the non-full rank matrix F
Ft=[~ ~ ~], sothatm=3,£=2jorF.
Then r =rank(F) = 1 and since GCD(Zl, Z2) = 1, therefore F is minor co-

prime. Its associated module is M = D 1 x2 / D 1 x3 F and the torsion submodule
of Mis t(M) = K[Zl' Z2]/(Zl, Z2) f=. O. Therefore M is not torsion-free. The
equivalence,
M is torsion-free -<===> F is minor coprime
does not hold in general. It only holds if the projective dimension of M is at most
one [110], [107].
Example 5.8 Consider F as in the preceding example. The variety of rank sin-
gularities of F is V(Zl' Z2) = {(O, On, which is finite. But, it has been shown
that the associated cokernel module M is not torsion -free, and thus not reflexive.
Oberst has given an example to show that if M is reflexive, then the variety of rank
singularities of F or, for that matter, the rank singularities RS(M) of M need not
be finite.
It has been pointed out by Oberst [110] that the reason why system theory in
two dimensions is more complete and manageable than in higher dimensions is
because every second syzygy module is free over the polynomial ring K[Zl' Z2]
since every module has projective dimension at most two. This implies that over
K[Zl' Z2], every finitely generated reflexive module is free [199]. When n ~ 3,
existence of an unique (up to unimodular matrices or units in the polynomial
matrix ring) greatest common left (or right) divisor cannot be guaranteed and the
problem of construction under the tacit assumption of existence has only been
solved in certain important special cases. The framework of D-modules adopted
by several researchers like Oberst [110], uses a framework of algebraic analysis
based on homological algebra and differential operators. The ring D of differential
operators bears resemblance to a polynomial ring. For an integration of the theory
of partial differential operators within the framework of D-modules, see [198]
4.3 VARIOUS DEGREES OF PRIMENESS
Let N be a set of integers and let a vector v = (VI"'" Vp)T E KPXl[z] for
somep E N.
DEFINITION 5.13 The vector v is called a unimodular column vector ifits com-
ponents generate K[z] i.e. if there exist 91, ... ,9p E K[z] such that V191 +
... + vp9p = 1; a matrix A E KqxP[z] is called a unimodular matrix ifits major
determinants generate the unit ideal in K[z].
Let F E Dlxm, i ~ m be a normal full rank matrix (normal full rank implies
that at least one major determinant or maximal minor is not identically zero))
whose entries are in the unique factorization domain (UFD) D. Let M be the set
of major determinants of all (i x i) submatrices of H and since M is a subset of
D, denote 1= (M) to be the ideal in D generated by the elements of M.
DEFINITION 5.14 The matrix F is (a) zero prime if 1= D, (b) minor prime if the
elements of M have no nontrivial common divisor in D (i.e. excluding units in
D) and (c) (ieft) factor prime ifwhenever F isfactorable as F = U F 1 , U E
D lxl , Fl E D lxm then U is unimodular i.e. its determinant is an unit in D. The
matrix is called weakly zero prime if the maximal minors have only a finite number
of common zeros.
The notion of primeness degree, which acts as a numerical measure of the amount
of primeness of a matrix was introduced recently [200].
DEFINITION 5.15 The height or codimension of a proper ideal I in the n-variate

polynomial ring K(z), where the field K is algebraically closed, is n minus the
dimension of the variety on which I vanishes.
FACT 5.3 Let F be a matrix over the ring K[z] where K is an algebraically closed
field of characteristic zero and let F have at least as many columns as rows. Let
V denote the variety on which F loses rank (i.e. the set of points in Kn at which
full-size minors of F have common zeros). Then:
(a) F has full row rank if and only if V f= K n

(b) F is minor left prime (MLP) or, for brevity, minor prime (MP) if and
only if V has dimension at most n - 2 (i.e. the variety V does not contain a
hypersurface)
(c) F is weakly zero left prime (WLP) or, for brevity, weakly zero prime
(WP) if and only if V is finite (i.e. the variety V comprises of isolated points
or, equivalently, is zero-dimensional)
(d) F is zero left prime (ZLP) or, for brevity, zero prime (ZP) if and only if
V = <P, the empty variety.
Therefore, it follows that
ZLP => WLP => MLP
The characterizations of the preceding notions of primeness can also be given

in terms of ideals rather than varieties and this allows generalization of primeness
notions to arbitrary coefficient fields without the imposition of algebraic closure.
The kernel and image of a n-variate wide polynomial matrix F E Kmxl[z] are
defined next
KerR F = {w E EIFw = O},ImR F = {w E Rmlw = Fv, vEE}

The cokernel of F is the factor module,
CokerR F = RmjlmR F.
The ideals of the minors of a matrix are determined solely by its cokernel.
DEFINITION 5.16 A minimal left annihilator (MIA) of a n-variate polynomial

matrix F is a n-variate polynomial matrix Q which is a left annihilator of F i.e.
QR = 0 and any left annihilator of F is a multiple ofQ i.e. QIR = 0 implies
Ql = YQ for a n-variate polynomial matrix Y. A minimal right annihilator
(MRA) can be similarly defined.
FACT 5.4 [200] Afull row (column) rank matrix over K[z] is an MIA (MRA) if
and only if it is minor left (right) prime.
A consequence of Fact 5.4 is that the row-space of a normal full rank n-variate
polynomial matrix F has a kernel representation if and only if F is minor left
prime or, equivalently, it is an MLA. In contrast to zero and minor primeness,
factor primeness cannot be decided by the variety Valone. An example to justify
this is given next [107], [201].
Example 5.9 The non-zero maximal minors of each of the two full rank matrices
F and G are z~, ZlZ2, and ZlZ3.
F = [Zl 0
o Zl
o -Z2]
Zl Z3
Though the varieties of the ideals of the minors are identical, F is not factor (left)
prime (FLP) but G can be shown to be FLP. Note that G is not minor (left) prime
and, therefore, does not have a kernel representation, in accordance with Fact 5.4.
DEFINITION 5.17 Let F E Kmxl[z], m ~ t. The left primeness degree pdF) of

F over the ring K[z] in n indeterminates is the height of the ideal generated in
K[z] by the order m minors of F or is defined to be 00 if this ideal is K[z]. The

matrix F is left d-prime (d E {O, 1, ... , n} or d = 00) if its left primeness degree
is at least d. The right primeness degree of Ft is defined similarly, on the order m
minors of Ft.
Therefore, PL (F) = 0 corresponds to F not having full rank. F has full row
rank if and only if pdF) ~ 1, F is minor left prime if and only if pdF) ~ 2,
F is weakly zero prime if and only if pdF) ~ n, and F is zero left prime if
and only if pdF) = 00. It follows that in the bivariate n = 2 case, minor left
primeness and weak zero left primeness are equivalent, each being associated with
a zero-dimensional (isolated points) variety of the ideal of maximal minors. ill the
trivariate case, the minor, weakly zero, and zero primeness concepts are all distinct.
For n > 3, other distinct primeness concepts appear depending on the dimension
of the variety of maximal minors. Note that the concept of factor primeness does
not fit naturally into the current discussion. ill fact, simple algebraic (in terms
of ideals) or geometric (in terms of variety) characterizations of factor primeness
does not seem to exist.
The left coprimeness degree of a pair of multivariate polynomial matrices N, D
with the same number of rows (encountered in matrix fraction description of a
rational multivariate matrix) is defined to be the left primeness degree of the com-
pound matrix [N D]. The right coprimeness degree of a pair of matrices with the
same number of columns can be defined analogously [200].
The notion of factor left (right) primeness (FL(R)P) has been generalized to
matrices that are not of normal full rank However, as observed in [200], the GFLP
is more a generalization of MLP rather than FLP, because a full rank matrix is
GFLP if and only if it is minor left prime.
DEFINITION 5.18 A matrix F is called generalized factor left prime (GFLP) if

the existence ofafactorization F = UFl with rank(F) = rank(Ft) implies the
existence of a polynomial matrix E such that Fl = EF. The notions of MLP and
GFLP coincide in the normal full rank case, while ZLP implies GFLP.
ill summary, for the normal full rank case (see also [107, pp. 13-14]),
ZLP => WLP ... => MLP
and
MLP ¢:} GFLP => FLP
(analogously for right primeness) whereas in the general case
zero prime => weakly zeroprime => ... => minor prime
and
zeroprime => generalized factor prime => minor prime
The dots refer to the gap in the corresponding ideal dimensions that are associated
with the missing links between the various primeness degrees. The primeness
notions also have module-theoretic interpretations. Suppose that F E KqXP(z),
so that F can be viewed as a mapping from Dlxq to Dlxp where D is a polynomial
ring. The module-theoretic interpretation is in terms of the cokernel module
M = DIXp jDlxqF
Zero primeness of F then corresponds to module M being projective or, equiva-

lently free (from Serre's conjecture (see below), proved, independently, by Quillen
and Suslin). It has been shown by Oberst that F is GFLP if and only if M is
torsion-free. As a free module is torsion-free, therefore, zero primeness implies
generalized factor (left/right) primeness, as previously asserted. Furthermore, a
full row rank GFLP matrix is factor left prime (FLP), but a full row rank FLP need
not be GFLP.
The next example constructed using the matrix in a previous Example 5.6 il-
lustrates that factor right primeness is not equivalent to minor right primeness for
n ~ 3.
Example S.10 The maximal minors of the matrix
M= [Zl
o
Z2
Z3
0]
Zl
t = [ Ct o
Zl
]
t
are al = ZIZ3, a2 = Z2ZI, and a3 = zf. Since C cannot be factored as proved in

Example 5.6, therefore, M is factor prime even though the maximal minors share
a common nontrivial factor.
Let C E KqxP[z] be the composite matrix C ~ [A I B], of normal full rank

where without loss of generality, q ~ p. Let G E Kqxq[z] be a greatest common
left divisor (GCLD) of A and B (assuming G exists). Note that G is restricted to
be a square matrix as required in matrix-fraction descriptions of rational matrices
in multidimensional system theory.
DEFINITION 5.19 [202J Let FE Kmxl[z] with m ~ f, and let al, ... , af3 denote
the f x .e minors of the matrix F, where {3 = (7) = (m~i)!l!' Extracting the
greatest common divisor (GCD) d of al, ... , af3 gives:
i = 1, ... ,{3. (5.57)
Then, bl , ... , bf3 are called the generating set [203J or reduced minors [202J of
F.
For somen-variate polynomial matrices A I and B I , consider the representation,
assumed to exist,
C = [A I B] = G[ Al I Bd.
/;:,. A
(5.58)
It is clear from Eq.(5.58) that all columns of C belong to the module (over the
polynomial ring K[z]) generated by the columns of G. Furthermore, it is assumed
that the determinant of G equals the greatest common divisor (GCD) of the major
determinants of C (this is needed because neither determinantal factorization nor
primitive factorization is possible, in general, for the n-D, n > 2 case [5, pp.64-
65]). This fact is an implicit requirement for Eq.(5.59) below and is explicitly
noted here for the sake of clarity. When the reduced minors (minors after GCD
extraction) of C have no common zero, there exists a polynomial matrix H [5]
such that,
[All BdH =1. (5.59)
Consequently, using Eqs.(5.58) and (5.59) one can write
G= G[AI I Bl]H = CH. (5.60)
From Eq.(5.60), the columns of G also belong to the module generated by the
columns of C. Therefore, the columns of G and the columns of C generate the
same module over the polynomial ring K[z], subject, of course, to the assumptions
made.
4.4 R-MODULES AND VECTOR SPACES
There are some similarities as well as fundamental difference between mod-

ules (over rings) and vector spaces (over fields), submodules and subs paces etc.
[204]. One such difference which influences the factorization problem is stated
and illustrated next.
FACT 5.5 The maximal set oflinearly independent elements ofthe reduced Grobner
basis for module M may only generate a proper submodule of M.
The reader may wish to consult the very clear presentation in [204].
Example 5.11 Consider the module M generated by the three polynomial vectors
over K[Zl, Z2].
1/1 [0 z~ - 3 - zr + 2 ] t
1/2 [z~-3 0 -ZlZ2+1]t
1/3 [zr-2 -ZlZ2+ 1 O]t.
Let 1 = (nl = ZlZ2 - 1, n2 = -zr + 2, n3 = zi - 3) be an ideal of the ring
K[Zl' Z2]. Furthermore, note that
The syzygies of
are the three polynomial vectors generating module M. Therefore, it is easily

seen that the reduced Grobner basis of N, whose elements are zero-coprime, and
Syz(N), are, respectively,
GB(N) = (1), Syz(N) = M.

The matrix,
G= [lIl 112 lI3]
has linearly dependentcolwnn vectors over the ring K[ZI, Z2]' Also lII, 112, lI3 form
the reduced Grobner basis. If we choose any two vectors in the set {lib lI2, lI3},
then we get a maximal set of linearly independent elements of G. But such a set
does not generate the whole module M.
The next example also illustrates that the maximal set of linearly independent
elements of the Grobner basis for a module generated by the columns of a multi-
variate polynomial matrix C using any lexicographical ordering may not form a
minimal generating set of the column space of C.
Example 5.12 The reduced minors of the factorable matrix C E K2X3[z], shown
below, are zero-coprime.
c=[ ZIZ~Z3
2 2
ZIZ3 + Z3
0
-Z3
2 2 - 1
-z1Z2
-Z~Z3 - ZI ]
t;,. h
= G[All Bd (5.61)
where,
1 2 -
[ -z2z2 1 ZIZ~Z3 ]
G -ZI Z3 '
[ z3I Z22Z32 ]
Al
ztZ~Z3 + ZfZ3 +1 '
and
[ -ZIZ~Z3 4 2 z3 + 1]
-Zlz2
BI - 1 -z~(zfz~ + 1) .
-z l2 z22 -
Using degree reverse lexicographical ordering with ZI ~ Z2 ~ Z3, the matrix

G whose columns are the reduced Grobner basis vectors of the module gener-
ated by all columns of C is calculated (by the program SINGULAR [205]) to
be G = [gi g2 g3], where gi = [0 Z3]t, g2 = [Z3 O]t, and g3 =
[ (zf z~ + 1) ZI]t. However, here the GeLD Gcannot be computed using Algo-
rithm A in [206], because no proper linearly independent subset of G can generate
the column space of C. Note that though the columns of Ggenerate the same mod-
ule as the columns of C, it cannot be derived from G by applying the algorithmic
theory of Grobner basis.
In general, the construction of G will depend on the validity of a conjecture

advanced in [207], [208].
CONJECTURE 5.1 Let d be the greatest common divisor ofall major determinants
ofC E KqxP[z), where q < p.lfthe reduced minors ofC have no common zeros
in K n , then C can be factored as C = G Co with Co E KqXP[z] being ZLp,
G E Kqxq[z] and det G = d.
When q > p, the counterpart of the conjecture is easy to state by cosidering ct.
The preceding conjecture is proved for the case p = q + 1 [207], the condition
satisfied in (5.12). For further results on the preceding Conjecture, see Section 6
below.
5. Computations for Coprimeness Using Grobner Bases

Let two normal full rank matrices A{z) and B{z) belong to Kmxq[z] and
Kmxi[z] respectively, where integers q, l, m satisfy q + l > m > 1. Form
the composite matrix C.
C ~ [A I B].
DEFINITION 5.20 The matrices A{z), B{z) are
1 zero left coprime (ZLC) if there exists no n-tuple z = (Zl' Z2, •.• , zn) which is
a zero of all the m x m minors ofC{z),
2 minor left coprime (MLC) if these m x m minors (or major determinants of

C{z) are relatively prime,
3 factor left coprime (FLC) ifin any polynomial matrix decomposition C(z) =
C1 {Z)C2 {z) in which C1 {z) is square, the determinant ofC1 {z) is an element
of K, i.e. C1 {z) is unimodular.
NOTE 5.1 In dualfa\'hion, A{z) and B{z) are zero right coprime (ZRC) etc. if
At{z) and Bt{z) are zero left coprime etc, where {.)t denotes transposition.
FACT 5.6 [40J For n = 1, ZLC == MLC == FLC;for n = 2, ZLC ¢ MLC == FLC
and for n = 3, ZLC ¢ MLC ¢ FLC
5.1 ALGORITHM FOR COMPUTING THE RIGHT INVERSE OF ZLC

MATRICES
The right inverse, [xt[z] yt[zW of [A[z] B[z]] exists i.e.
A(z)X(z) + B(z)Y(z) = 1m , (5.62)
where 1m is a (m x m) identity matrix, if and only if A(z), B(z) are ZLC.

FACT 5.7 (Hilbert) Let an ideal in the polynomial algebra belonging to K[z]
= 1, 2, ... , m }.
be generated by the set of zero-coprime polynomials, {!:lk (z), k
Then there exist polynomials Ak(z) E K[z], k = 1,2, ... ,m such that
m
1 = LAk(z)!:lk(z). (5.63)
k=l
The construction ofthe class of all possible solutions for {Ak (z)} in Eq.(5.63)
is presented next.
ALGORITHM 5.1 1 Compute one particular solution by using Buchberger's al-

gorithm for computing Grobner bases [209] and keeping track of how each
element of the Grobner basis is generated.
2 Compute the jYZYgy module (which is analogous to the null ,space in linear
algebra) generated by the set ofpolynomials, {!:lk (z)}.
3 Combine the previous two results to obtain a set of all possible solutions as in
[209].
ALGORITHM 5.2 1 Defineamx(q+i)polynomialmatrix. C(z) = [A(z)IB(z)].

Find all q+lCm m x m minors, !:lj(z),j = 1,2, ... ,q+l Cm, ofC(z).
2 Use Algorithm 5.1 to compute all possible sets of polynomials, {Aj (z), j =
1,2, ... ,Hi Cm} such that Eq. (5.63) holds.
3 Pick W to be any (q + i) x m real constant matrix whose m x m minors,
Wj,j = 1,2, ... ,q+i Cm are all nonzero. Note that such a W always exists
forq x i ~ m.
4 Form the m x m polynomial matrix D(z, A), defined as
D(z, A) ~ C(Z)A(A)W (5.64)
where A(A) = diag[Al, A2,"" Aq+l]' Then use this matrix to form the prod-
uct,
F(z, A) ~ A(A)WDa(z, A) (5.65)
where Da(z, A) denotes the adjoint matrix of D(z, A).
5 Compute the polynomial matrices, Zj{z) where,
Zj{z) = ~ [JmF{z, A) (5.66)

Kj [JAil [JAh ... [JAjrn
where {it, h, ... , jm} is a set of corresponding column indices, used to form
Wj.
6 Obtain the set of all right inverses Z{z) ofG{z) from the sum ofproduct of the
following form
Z(z) ~ ~Aj(Z)Zj(z) ~ [ :i:1 1 (5.67)
5.2 MINOR COPRIMENESS AND GROBNER BASES
By applying the theory of Grobner bases, a procedure for construction of poly-

nomial matrices Xi{Z) and Yi{z) in Eq.(5.68) below [40] is introduced.
THEOREM 5.9 [40J The m x q and m x f polynomial matrices A{z) and B{z),
q + f ~ m ~ 1, are MLC if and only if f oreveryi = 1, 2, ... , n, there exist
polynomial matrices Xi{Z) and Yi{z) such that
(5.68)
where 'l/1i{Z\Zi) is a nontrivial (n -i)-variate polynomial which is independent

of the variable Zi. Moreover, if A{z) and B{z) have real coefficients so also can
Xi{Z) and Yi{z).
Given two polynomial matrices, A{z) and B{z), which are minor left-coprime,
the next two algorithms compute polynomial matrices Xi(Z) and Yi(z) such that
Eq.(5.68) holds for i = 1,2, ... ,n.
ALGORITHM 5.3 1 Dejineamx{q+f)polynomialmatrixG{z) = [A{z)IB{z)].

UsingAlgorithm5.1,jindthe set of polynomials {Ak,i{Z), k = 1,2, ... ,q+l Gm}
such that
q+lCm
L Ak,i{Z)~k{Z) = "pi (Z\Zi), i = 1,2, ... ,n (5.69)

k=l
where the ~k{Z)'S denote the m x m minors ofG{z) and the "pi (Z\Zi) 's are
obtainable from the Grobner basis computed by using lexicographical ordering,
Zi >- Zj, 'Vj -I i.
2 Subsequently, calculate the .\yzygy module of {~k' k = 1,2, ... ,q+l Gm}. The
set of all solutions to Eq.(5.69) is thenformed in the same manner as in step 3
of Algorithm 5.1.
3 Using Algorithm 5.2, calculate the set of polynomial matrices {Zj(z),j =

1,2, ... ,q+l Cm} as in Eq.(5.66).
4 All possible polynomials matrices, {Xi(Z), Yi(z), i = 1,2, ... , n} that satisfy
Eq. (5.68) are
Example 5.13 Let A(z) = [ZlZ2 + 1 -1 1] and B(z) = [ 1 ] so

Z2 ZlZ2 - -Zl
that [A(z) I B(z)] is minor left-coprime. The objective is to find all polynomial
matrices X1(z), X 2 (z), Yi (z) and Y2(Z) such that,
A(Z)Xl(Z) + B(z)Y1(z) = 'Ijldz\zt}I2

A(z)X2(z) + B(Z)Y2(Z) = 'Ijl2(z\z2)I2 (5.71)
where 'Ijli(Z\Zi) is nontrivial and independent of Zi.

Let C(z) = [ ZlZ2Z2+ 1 ZlZ2-1- 1 1.
-Zl
] The set M of all 2 x 2 mmors
. are
~l(Z) = z?z~ + Z2 - 1, ~2(Z) = -Z?Z2 - Zl - Z2, ~3(Z) = -ZlZ2 + Zl + 1.

Using lexicographical order Zl ~ Z2, the Grobner basis G 1 of an ideal generated
by M is calculated to be,
G1 = {z~ + 2z~ + Zl + 1, z~ + 2Z1 + Z2}

Alternatively, using lexicographical order Z2 ~ Zl the Grobner basis G2 of an ideal
generated by M is calculated to be,
G 2 = {z~ - 2z~ + 3Z2 + 1, z~ - Z2 + 2 + Zl}

Thus, the polynomial 'Ijll (z \ Zl) is (z~ - 2z~ + 3Z2 + 1) and similarly, the polynomial
'Ijl2(Z\Z2) is (z~ +2z~ + Zl + 1). In addition, by keeping track of how each element
of Gi, i = 1,2 is generated, the following relations are obtained,
Zr + 2z? + Zl + 1 = (O)~l(Z) - Zl~2(Z) + (z? + 1)~3(Z) (5.72)

z~ - 2z~ + 3z2 + 1 = (O)~dz) - (Z2 - 1)2~2(z)
+ (ZlZ~ - ZlZ2 + 2Z2 - 1)~3(z) (5.73)
Next the syzygy modules of the ideal generated by M are, by using Zl ~ Z2

ordering,
Syz(M) = ((O,ZlZ2 - Zl -1, -Z~Z2 - Z2 - zt}, (1, l,ZlZ2 + 1)} (5.74)

and by using Z2 >- Zl ordering,

Syz(M) = ((0, ZlZ2 - Zl - 1, -Z?Z2 - Zl - Z2), (1, 1, ZlZ2 + 1)) (5.75)
The generic solutions ofEq.(5.71) are obtained by taking the sum of the particular
solutions obtained in Eqs. (5. 72) and (5.73) and the Syz(M) solutions in Eqs.(5.74)
and (5.75), respectively.
A1(zj 1) = v
A 2 (zj 1) = + U(ZlZ2 - Z2 - 1) + v
-Zl
A 3 (zj 1) = (z? + 1) + u( -Z?Z2 - Z2 - zt} + V(ZlZ2 + 1)

A1(zj 2) =t
A 2(zj 2) = -(Z2 - 1)2 + S(ZlZ2 - Zl - 1) + t
A 3 (zj 2) = (ZlZ~ - ZlZ2 + 2Z2 - 1) - S(Z?Z2 + Z2 + zt} + t(ZlZ2 + 1)
where u, v, s, and t are arbitrary polynomial in K[Zl' Z2].
By using the method of Algorithm 5.2, the polynomial matrices Zj,j = 1,2,3 in
Eq.(5.66) are
Finally, the class of all solutions is parameterized by
The generalization of Grobner bases from polynomial ideals to modules has

been described in detail in [209, Chapter 3]. By computing a Grobner basis ofthe
module generated by the columns of C(z), the polynomial 'l/Ji(Z\Zi) in Eq.(5.68)
can be obtained directly. The next algorithm will solve for the polynomial matrices
Xi(Z) and Yi(z) for i = 1,2, ... ,n such that Eq.(5.68) holds.
ALGORITHM 5.4 1 Define am x (q+l) polynomial matrixC(z = [A(z)IB(z)].
Compute n Grobner bases for the module generated by the q +l column vectors
Ck(Z), k = 1,2, ... ,q+l ofC(z), using the n lexicographical orderings where
the ith ordering gives Zi the highest grade, for i = 1,2, ... ,n.
2 For each ordering, the Grobner basis reveals the polynomial 'if1i{Z\Zi), asso-
ciated with the polynomial matrices Xi{Z) and Yi{z) for i = 1,2, ... ,n that
can be computed so that Eq.(5.68) holds'.
3 By keeping track ofhow each element of the Grobner basis is generated calcu-
late the solution set {Pk,j{Z), k = 1,2, ... ,q + f, j = 1,2, ... ,m} satisfying
q+i
'if1i{z\zdej = LPk,j{Z)Ck{Z), fori = 1,2, ... , n, j = 1,2, ... , m
k=l
where ej is the lh column of 1m.

4 For each i in the above step, define the (q + f) x m polynomial matrices
M{z : i) which has the element Pk,j{Z) in its kth row and lh column Vk,j.
Note that M{z : i) has the property thatC{z)M{z : i) = 'if1i{z\zi)lm . Similar
to Algorithm 5.3, the matrices Xi{Z) and Yi{z) can be defined naturally by
[ Xi{Z) ] ~ M{ .')
Yi{z) - Z • Z
5 Find the set of all possible solutions after solving for the ,syzygy module (the
set of all syzygies, where each syzygy is a vector ofpolynomials [209, Chapter
3, p. 161]) generated by the set of column vectors {Ck' k = 1,2, ... , q + f}.
The above (second) method is applied to the previous example, which was
already solved by the first method. First, compute the Grobner basis of modules
generated by the column vectors in
By using the program SINGULAR, the Grobner basis becomes, for Zl ~ Z2

ordering,
For Z2 ~ Zl ordering, the Grobner basis is
G2 = ({O, z~ - 2z~ + 3z2 - l)t, {O, Zl + z~ - Z2 + 2)t, (1, z~ - Z2 + 2)t)

This implies that polynomials 'if11(Z\Zl) and 'if12(Z\Z2) are (z~ - 2z~ + 3z2 - 1)
and (z~ + 2z? + Zl + 1) respectively.
Furthermore, by keeping track of how each element of the Grobner basis is
generated, the following relationships are obtained,
[ zf + 2z?0+ Zl + 1 ] = (zr)cr(z) - (zr + 1)c2{z) - (ZrZ2 + zr + Zl + 1)c3{z)

and
0
[ Z23 - 2Z22 + 3Z2 - 1 ] = (Z~ - 2Z2 + l)Cl{Z) + (ZlZ2 - ZlZ~ - 2Z2 + 1)c2{z)
- (ZlZ~ - ZlZ~ + Z~)C3(Z)
The next step is to write the polynomial modules (zt +2Z~+Zl + l)t and(z~-2z~+
3Z2 - l)t as a combination of the column vectors of C{z), Ck{Z), k = 1,2,3 with
the polynomial coefficients. Using division algorithm in K{Z2)[Zl] and K{zd[Z2]'
respectively, one arrives at
[
zt + 2z~o+ Zl + 1 ] 2 3 3
= (ZdCl{Z) + (zl + ZdC2{Z) - (Z l Z2 - Zl -1)c3{z)
2
and
[ z~ - 2z~ o+ 3z2 - 1] = ( ) () (Z2 - 1 Cl Z - ZlZ2 - z22 ) C2 ()

+ Z2 Z
- (ZlZ~ - z~ + z~ + Z2)C3(Z)
Combining equations above, the polynomial matrices M{z; i), i = 1,2 as defined
in the previous algorithm are,
M(z; 2) = [ -(zt~ zt)

-(ztz2 - z~ - 1)
. -(z?\
1)
-(Z~Z2 + z~ + Zl + 1)
1
M(z; 1) = [
Z2 - 1
-(ZlZ2 - z~ Z2) +
z~ - 2Z2 1
ZlZ2 - ZlZ~ - 2Z2
+
1 +
1
-(ZlZ~ - z~ + z~ + Z2) -(ZlZ~ - ZlZ~ + z~)
Lastly, the syzygy module generated by the set of column vectors of C(z) is
computed as
SYZ(CI C2 C3) = [(ZlZ2 - Zl - 1) (-Z~Z2 - Z2 - zd (-z~z~ - Z2 + 1W
Thus, the class of all solutions is parameterized by
Z(z; 1) = M(z; 1) + [u v]g(z)

Z{z; 2) = M(z; 2) + [s t]g(z)
where u, v, s, and t are arbitrary polynomial in K[Zl' Z2].
For a full row rank matrix F with entries in the n-variate polynomial ring
D = qZl, ... ,zn], consider the possibility of existence of a matrix G, that satisfies
the extended Bezout identity [210],
FG = wI
where G has entries in C[z], W E C[z]. The existence of G has been discussed
when F is ZLP, in which case W = 1. When F is MLP(WZLP), G exists with
W an (n - l)-variate polynomial (a univariate polynomial). Woods et al. [200]
proved the existence of a one-to-one correspondence between the number of Zi'S
in wand the n different primalities (including WZLP, ZLP, MLP) defined by the
dimension of the algebraic variety formed from the zeros of all the maximal order
minors of F. In [210], among other results, it is shown how to pass from one
type of primeness to another by inversion of certain w E C[z] and how to obtain
extended Bezout identities for each type of primeness.
5.3 PROBLEMS IN FACTOR COPRIMENESS AND GCLD (GCRD)

EXTRACTION
Unlike in the n = 1 and also n = 2 cases, it is not possible to speak of

greatest common polynomial matrix divisors as these may not be unique up to
unimodularity when n ~ q3. However, under the hypothesis of existence and
uniqueness upto unimodularity, constructive algorithmic procedures like Grobner
bases have been used to obtain the factor provided specific constraints are met as
in [206].
Let D E Klxl[z], N E KffiXl[z], where K is the ground field of interest. Then
G E Klxl[z] is called a GCRD (greatest common right divisor) of D and N if and
only if
1) G is a right factor of both D and N
2) G is a left multiple of any common right factor of D and N
The next example shows that the left (or right) GCD of a polynomial matrix in
three indeterminates may not exist. This counterexample was discussed in Lin's
Ph.D dissertation at Cambridge University in England.
Example 5.14 Consider the matrices
D(zI, Z2, Z3) = ZIZ2I, N(ZI' Z2, Z3) = [ZIZ~

Z2 Z3
Zl]
Z2
Consider the composite matrix F = [Dt Ntt Clearly, two factorizations of

Fare
F=[~
2
-Zlz3
ZlZ2
0
Zl
Z3(1 - Z3)
Z2
r[ z~ o1 ] =~ FIGI
z3
F=[~
-Z2Z3
ZlZ2
Z3(Z3 -
Zl
1) o
Z2 f[ Zl
Z3
o1 ] =~ FG
2
It is easy to show that F I , though not MRP (the GCD of the six maximal minors is
2
Zl), is FRP(factor right prime). This follows from the fact that in FI (0, Z2, Z3), the
only non-zero elements, z3(1 - z3)and Z2 (in the last row), are not zero coprime
and therefore cannot annihilate aZRPvector. Suppose that G 3 E K2X2[Zl, Z2, Z3],
a GCRD of D and N exists implying that F = F3 G 3 , G 3 = K1G 1 = K 2 G 2 , for
F3 E K4X2[ZI,Z2,Z3],KI,K2 E K 2X2 [ZI,Z2,Z3]. Then, it follows that F =
F3K I G I , and since GI is non-singular, therefore FI = F3K I . But FI is FRP,
implying that KI is an unimodular square matrix with a constant determinant
ki. Therefore, det(G 3) = k Iz 2 and, similarly, det(G 3) = k 2z 1 (from F =
F3K2G2, F2 = F3K 2), where det{K2) = k 2. This leads to a contradiction, and
the hypothesis that GCRD G 3 exists is incorrect.
The possibility of nonunique irreducible (in the factor coprime sense) realiza-
tions in the n ~ 3 case is illustrated in the next example, due to Oberst [110]
that uses an earlier counterexample in [40], [5] that was presented to prove the
infeasibility of polynomial matrix factorization when n ~ 3. This example was
first used by Northcott [143] to prove that the power of a prime ideal need not be
primary, as was mentioned in [5, Chapter 1, p. 64]. The infeasibility, in general,
of GCRD (or GCLD) extraction from two polynomial matrices, seen through the
previous example is linked to the nonuniquess exhibited.
Example 5.15 Consider the trivariate polynomial matrix F and its adjugate matrix
Fadj
whose elements, defined next, are polynomials in ZI, Z2, Z3.

FI =zi- ZIZ3, h=Z2Z3-Z~, h=Z~-Z?Z2
The determinant of F is IFI = d I d2 where the irreducible polynomials dl and d2
are defined to be
Oberst showed that the rational matrix H, defined below, has the following two
different irreducible left MFDs (12 is the identity matrix of order 2).
H = F- I {z I I 2) = (d2h)-1 Fadj
where each of the matrix pairs (F, z 1 I2) and (d212, Fadj ) is factor left coprime i.e.
the composite matrices [F z l I2] and [d2h Fadj] are both factor left prime (and,
here, also minor left prime). This gives rise to different minimal realizations of
the same transfer class.
See Section2 in Chapter 8 for the open problem on the finding of an algorithm for
computing the GCLD (in the polynomial matrix factor sense), assuming it exists,
of a composite multivariate polynomial matrix in three or more indeterminates,
when its reduced minors are not zero coprime.
6. Generalization of the Serre Conjecture and its

Consequences
LetK[z] = K[ZI' ... ,zn] denote the set of polynomials in n variables Zl, ..• ,Zn
with coefficients in a specified field K; KlXm[z] the set of l x m matrices with
entries in K[z]; K* the set of all nonzero elements in K, i.e., K* = K\ {O}; Ol,m
the l x 'Tn zero matrix and 1m the m x m identity matrix.
Henceforth, the argument (z) is omitted whenever that does not cause confusion.
Here we consider the case where F is a normal full rank matrix and postpone
discussion of the complementary (degenerate) case.
6.1 NORMAL FULL RANK MATRIX CASE
To facilitate reading, some definitions and notations are briefly reviewed.
DEFINITION 5.21 [208] Let FE Klxm[z] with l ~ m be a multivariate polyno-

mial matrix. Then F is said to be:
(i) zero right prime (ZRP) if the maximal minors aI, ... ,a~ are zero coprime,
i.e., there exist hI' ... ' h~ E K[z] such that 'Ef=l hi ai = ko E K*;
(ii) minor right prime (MRP) ifthe maximal minors a I , ... , a~ are factor coprime,
i.e., d = kl E K*.
G E Klxm[z] with m ~ l is said to be zero left prime (ZLP) or minor left

prime (MLP) if Gt is ZRP or MRP, where Gt is the transpose of G.
In 1955, J.P. Serre raised the question as to whether finitely generated projective
modules over polynomial rings (algebraic vector bundles over an affine space of
n-variate polynomials whose coefficients are in field K) are free (trivial bundles).
This posed problem came to be referred to as Serre's conjecture [42] which includes
also the more general case when K is a principal ideal domain. Both nonconstruc-
tive and constructive methods have been advocated to solve Serre's conjecture .
The property, projective modules over polynomial rings are free (projective-free),
has been shown to be equivalent to the unimodular matrix completion question:
can a ZRP matrix (the maximal minors of the matrix generate the unit ideal in
the polynomial ring) F E KPXq[z] (p > q) be completed to a square matrix
U = [F E] E KPXP[z] such that E E KPx(P-q)[z] and det U = ko E K*is a
nonzero element of K.? The evolution of Serre's problem from 1955 to 1976 led to
two independent nonconstructive proofs by D. Quillen and A. A. Suslin of Serre's
Conjecture [42]. A purely algebraic formulation that is a consequence of their
work is stated next when q ~ p (the case complementary to the one considered
above).
THEOREM 5.10 [Quillen-Suslin][211JLetAbeaunimodularqxpmatrix(q :S p)

over K[z] , where K is afield. Then there exists a unimodular p x p matrix U over
K[z] such that
AU = [Iq I 0qx(p-q)]. (5.76)
The result remains true if the field K is replaced by a principal ideal domain.
This important result was proved constructively by Logar-Sturmfels [211] in
1992, where, for simplicity, the exposition was restricted to the field C of complex
numbers. They produced an algorithm for computing the unimodular matrix U.
Since A equals the first q rows of U- i , therefore, the finding of U is equivalent
to completing A to a square invertible matrix. Their constructive procedure is
implementable using Grobner bases. The Quillen-Suslin theorem proves that the
ring K[z] of multivariate polynomials is, actually, a Hermite ring.
DEFINITION 5.22 A ring is a Hermite ring if every unimodular row can be com-
plemented i. e. the row can be embedded in a square unimodular matrix with
elements in the base ring.
For a discursive discussion on the relationship between Hermite rings and coprime
factorizations over a commutative ring see [131, Chapter 8].
A syzygy-based heuristic algorithm for unimodular completion was advanced
by Park (see [206, Algorithm B, p.1478]). The main problem of the syzygy-based
heuristic algorithm is the lack of an effective procedure for finding a minimal
syzygy basis. In order to compute a globally minimal syzygy basis, the following
result is required[212].
FACT 5.8 Let A E KqxP[z] be of rank q, with q < p and let r = p - q. The
syzygy basis, Syz(A) of A har; a generating matrix of minimal dimension p x r
i. e. a generating matrix is globally minimal if and only if there exists a minor right
prime (MRP) matrix H E KPXT[Z] (i.e. the major determinants of H form a set
of relatively prime polynomials) such that AH = OqXT. Let Hi E KPX8[Z] be a
generating matrix of Syz(A), with s > r. Then Syz(A) har; a generating matrix
of dimension p x r if and only if Hi can be factored ar; Hi = H E for some
HE KPXT[Z], E E K TX8 [Z] with H being MRP.
For the algorithm for computing this globally minimal syzygy basis based on Fact
5.8, see [206, Algorithm C, p.1479]. The example below illustrates the application
of the algorithm and the reader should be able to apply the result to other examples.
Example 5.16 Given a matrix A = [ zfz~ + 1 ZfZ3 + 1 ZlZ~Z3], the goal is

to compute a 3 x 3 unimodular matrix whose first row is identical to A. Note that
the polynomials in A are zero-coprime as required.
First, by using the program SINGULAR [205], a Grobner basis of the module
generated by columns of A, with respect to the degree reverse lexicographical
ordering ZI ~ Z2 ~ Z3, is {1} (verifying zero-coprimeness) and a 3 x 1 polynomial

matrix B (that must exist because of zero-coprimeness) such that AB = 1 is
B = [1 -z~z~ z~ ]t. The generating matrix of Syz A is
HI = [hI h2 h 3 ], where, hI = [ z~z~ + Z3 -Z3 -Z~Z3 - ZI ]t, h2 =
[ZIZ~Z3 0 -z~z~ - 1 ]t, and h3 = [ Z~Z3 + 1 -z~z~ - l O t
Since hI, h2' h3 are linearly independent, therefore, H2 = HI. It can be
verified that none of the three sets of two columns of HI can be used to fonn
a MRP matrix. By using the n-D primitive factorization algorithm (of course,
applicable to a class of n-D matrices to which the matrix in this example belongs)
[207] on submatrix H3 fonned by the first two columns of H2, the following factor
H can be extracted, H3 ~ [hI h 2 ] = H4E, where,
E = [ Z3
-ZI
The matrix H is indeed MRP and
AH = OIx2.
It is possible to conclude that H is a globally minimum generating matrix [212]

of Syz(A).The unimodular matrix 0 and its associated inverse A are 0
(B I H), and
Z~Z~ +1 Z~Z3 +1 ZIZ~Z3

6 4 2 2 542
ZI Z2 Z3 - ZI Z2+ zrz~z~+ ZI Z 2 Z3-
A 0- 1 = + zl4 Z22 Z3 +ztZ~Z3 -1 -ZIZ~Z3
-zlz~- 7 2 5 2
-Z l Z2 Z3 - Z l Z2- 1- 6 4 Z3-
ZlZ2
-2z15 z 22 - z31 4 2
-ZrZ3 - 2z~ -Z l Z2 Z3
There is considerable interest in the localizations of multivariate polynomial

rings K[Zl' ... ,Zn]. If one such ring, say R, is projective-free, then the polynomial
rings over R are likewise projective-free. No general results, however, are available
that will show how many of these rings are projective-free.
Consider a generalization of Serre's conjecture: let d be the GCD of all the max-
imal order minors of F, and bI , ... ,bfJ be the reduced minors of F. If bI , ... , bfJ
are zero coprime, can F be completed to a square matrix U = [F E] E Kmxm[z]
such that E E Kmx (m-l) [z] and det U = d? When d = ko E K*, the generalized
Serre's conjecture reduces to the original Serre's conjecture. Henceforth, it is as-

sumed that d is not a unit in K[z], and the field of coefficients K is an algebraically
closed field, such as the field of complex number C.
The generalized Serre's conjecture is equivalent to several other, possibly more
tractable, conjectures. This set of conjectures is complete in the sense that solution
of any would automatically solve the remaining ones. The detailed proofs are
available in [208].
Conjectures 1- 4 Let F be given in Definition l. !fbI,"" bf3 are zero coprime,
then we have the following conjectures:
1. There exists E E Kmx(m-l)[z] such that U = [F E] E Kmxm[z] with
detU = d.
2. F can be factored as' F = Fo Go, for some Fo E KmXI[z], Go E Klxl[z]
with det Go = d.
3. There exists H E KIXm[z] such that H F = Go,for some Go E KIXI[z] with
detGo = d.
4. There exists B E K(m-l)xm[z] with B being ZLp, such that BF = Om-l,l.
It is important to emphar;ize that when the zero coprimeness condition is not
satisfied by the reduced minors, Conjectures 1 to 4 cannot hold in general. For
example, consider the 1 x 2 matrix F = [ZIZ2 zlz3]T E CIX2[ZI, Z2, Z3]. Clearly,
it is not possible for F to be completed into a square matrix U E C 2X2 [ZI, Z2, Z3],
with F being the first column ofU, such that det U = ZI. This is because although
the reduced minors Z2 and Z3 are factor coprime, they are not zero coprime.
PROPOSITION 5.1 Conjectures 1-4 are equivalent.
So far we have restricted our discussion to the car;e where the reduced mi-
nors bt , ... , bf3 are zero coprime. However, for n-variate (n > 1) polynomials,
bl , ... , bf3 may not be zero coprime even if they are factor coprime. In [208J
another conjecture was raised for this car;e.
Conjecture 5 [208J Let F be given in 5.19. !fd, bl ,.·., bf3 are zero coprime, then
F can be factored as
F = FoGo, (5.77)
for some Fo E Kmxl[z], Go E KIXI[z] with det Go = d.
The difference between Conjecture 2 and Conjecture 5 is that bl , ... , bf3 in Con-
jecture 5 may not be zero coprime, but d, bl , ... , bf3 are zero coprime. Surprisingly,
these two apparently different conjectures turn out to be equivalent.
PROPOSITION 5.2 Conjecture 2 and Conjecture 5 are equivalent.

We now consider another related matrix factorization problem: given F E
K'XI[z] with det F = nf=1
Ij (fj E K[z], j = 1, ... , J), to factorize F as F =
nf=l E
Fj with Fj K'Xl[z] and det Fj = /j (j = 1, ... ,J). This is the so-called
detenninantal factorization problem. It is well-known [39] that univariate and
bivariate polynomial square matrices always admit determinantal factorizations.
However, it has been pointed out [40] that some n-variate (n > 2) polynomial
matrix does not have a detenninantal factorization (see Example 5.15). Thus, it is
interesting to know whether or not a given n-variate polynomial matrix admits a
detenninantal factorization. Another conjecture was raised for this problem [208].
Conjecture 6 Let F E K'Xl[z] with det F = nf=l/j (/j E K[z], j = 1, ... , J).
If iI, ... ,!J are pairwise zero coprime, i.e., Ii and fk are zero coprime for 1 ~
i, k ~ J, i f= k, then F can befactored as F = nf=l E
Fj with Fj K'XI[z] and
det Fj = fj (j = 1, ... , J).
6.1.1 Factorization Problem in Normal Full Rank Case

C5
.J..t
C1 ~C2-+C4
'\tt.J../
C3
The strategy behind the proof, given in [208] of the equivalence of the first
five conjectures C1, C2, C3, C4, and C5 in the subsection above is given by the
flowchart above. The truth of Conjecture 2 was recently proved[213]. Therefore,
Conjectures 1,2, 3, 4, and 5 all hold in the full normal rank case. It was also proved
that C2 implies the sixth conjecture C6. Therefore, the only unsolved problem in
the full rank case requires the settling of whether or not the implication C6 -+ C2
holds.
6.2 DEGENERATE RANK CASE
Another reason, besides the ones already encountered, why polynomials matri-
ces in K[Zl' Z2] can be handled better than their multivariate counterpart is because
of a consequence of the Hilbert basis theorem (see the introductory section), stated
next.
FACT 5.9 Refer to Definition 5.11. If K is a field then any finitely generated
module over the polynomial ring K[Zl' Z2, ... ,zn] has a free resolution length
~ n.
In the n = 2 case, Fact 5.9 can be used to prove the following result of Oberst
[110], as done in [107, p. 52, Theorem 12].
FACT 5.10 A minimal left (right) annihilator oj a bivariate polynomial matrix

whose elements are in K[z] = K[Zl, Z2], is either zero or can be chosen to be
matrix with full row (column) rank.
From Facts 5.4 and 5.10, it can be inferred that the full rank condition becomes
restrictive in the case of three or more indeterminates. The rank deficient case
and other notions of primeness that do not require the imposition of the restriction
of full rank was studied, recently, in [200]. For a discursive documentation, see
the monograph by Zerz [107]. Systems that admit a full column or row rank
factorization are those whose projective dimension (the minimum of the lengths
of projective resolutions) is at most one [194].
ASSUMPTION 5.1 Let FE KmXT[Zl, ... , zn) be of rank I with I < m and I < r.
Assume that F l , ... , Fk are all the m x I full rank submatrices ofF, and di is the
OCD of all the I x I minors of Fi, i = 1, ... ,k. It can be shown that F l , ... , Fk
have the same reduced minors, denoted by bl , ... ,bfJ. For simplicity, we call
bl , ... ,bfJ the column reduced minors ofF.
Note that for a matrix not of full rank, its column reduced minors may not be
the same as its row reduced minors. With the above notation and assumption, we
can raise the following conjectures concerning multivariate polynomial matrices
not of full rank.
Conjectures 7 - 9 Let F be given in Assumption 1. If bl , ... , bfJ are zero coprime,
then we have the following conjectures:
7. F can be factored as F = Al A 2,jor some Al E KmXl[z), A2 E K'XT[Z]
with Al being ZRP, det Gi = di, where Gi is the ith I x I submatrix of A2
corre.~ponding to Fi, i = 1, ... , k.
8. There exists H E K'Xm[z] such that H F = A 2,jor some A2 E K'XT[Z) such

that det Gi = di, where Gi is the ith I x I submatrix of A2 corre.~ponding to
Fi, i = 1, ... , k.
9. There exists B E K(m-l)xm[z] with B being ZLp, such that BF = Om-l,T'
Conjectures 7 - 9 may be considered as generalizations of Conjectures 2 - 4. It
should be pointed out that when F is not a full rank matrix, it cannot be completed
into a square matrix whose determinant is nonzero. In the following proposition,
we show that Conjectures 7 -9 are equivalent to Conjectures 2-4. Forconvenience
of exposition, we re-state Conjecture 2 in the following.
Conjecture 2' Let F be given in Assumption 1. If bl , ... ,bfJ are zero· coprime,
then Fl can befactored as Fl = Al Gl,jor some Al E KmXl[z) with Al being
ZRP, G l E K'X'[z] with det G l = dl .
PROPOSITION 5.3 Conjectures 2', 7-9 are equivalent.
Proof: We only show the equivalence of Conjecture 2' and Conjecture 7. The
other equivalences can be shown similarly as in the proof of Proposition 1. See
[208].
Without loss of generality, assume that FI is fonned from the first I columns of
F, i.e., F = [PI 0], where C E Kmx(r-l) [z].
Conjecture 7 -+ Conjecture 2': Assume that F can be factored as F = Al A 2,
for some Al E KmXl[z], A2 E Klxr[z] with Al being ZRP, det Gi = di, where
Gi is the ith I x I submatrix of A2 corresponding to Fi, i = 1, ... , k.
Since G I is the I x I submatrix of A2 corresponding to FI, it is fonned from
the first I columns of A 2. Hence, we can re-write A2 as A2 = [G I D], where
D E Klx(r-l)[z]. We then have
(5.78)
It is then obvious that

(5.79)
with G I E KlXl[z] and det G I = d i .
Conjecture 2' -+ Conjecture 7: Assume that FI can be factored as FI = Al GI,
for some Al E KmXl[z] with Al being ZRP, GI E K'xl[z] with det G I = d i .
Since FI is a full rank m x I submatrix of F and Al is ZRP, by a known result in
[40] (see, also [212]), it can be asserted that F admits a polynomial factorization:
(5.80)
for some A2 E K'Xr[z]. Let Gi be the ith I x I submatrix of A2 corresponding to

Fi" i = 1, ... ,k. It is then straightforward to show that det Gi = di, i = 1, ... , k.
o
Finally, we generalize Conjecture 5 to the case where F is not of full rank.
Similarly to the proof of Proposition 4, it can also be shown that Conjecture 10 is
equivalent to Conjecture 5. It is omitted here to save space.
Conjecture 10 Let F be given in Assumption 1. If di, bI , ... , bf3 are zero coprime
for some i E {i, ... , k}, then F can be factored as F = Al A 2,jor some Al E
KmXl[z], A2 E K'Xr[z] with Al being MRP, det Gi = ~, where Gi is the ith I x I
submatrix of A2 corresponding to Fi , i = 1, ... , k.
7. Factorization as a Product of Elementary Matrix Factors

An elementary matrix, denoted by eij (P), differs from the identity matrix by a
single entry p, which occurs in the (i, j) - th position with i i= j. Multiplying a
matrix by eij (p) on the left results in the row operation of replacing the i - th row
of the matrix by the i - th row plus p times the j - th row. Multiplying a matrix
by eij (P) on the right results in the corresponding column operation. The central
question in this section is the following.
"When is an element of the special linear group S Lm (R) expressible as a product
of elementary matrices over the ring R or, in other words, when does it also belong
to em(R), the subgroup of SLm(R) generated by m x m elementary matrices ?"
In 1977 A. A. Suslin answered the question by proving the (Suslin stability

theorem)
This result is striking because the Euclidean division algorithm that produces the
elementary reduction in the univariate case does not hold in the multivariate case.
Parks and Woodburn [214] produced an algorithm that expresses any square ma-
trix of order m, where m 2:: 3 and whose entries belong to the polynomial ring
K[Zl' Z2,··· ,zn] over an arbitrary but fixed field K, as a product of elementary
matrices with entries in the same ring. Thus, they provided an algorithmic proof of
the Suslin stability theorem, which can be implemented to realize biorthogonal
multiband filter banks. When the answer to the above question for a specified
matrix is in the affirmative (by the Suslin stability theorem, this is so whenever
R is a multivariate polynomial ring over an arbitrary but fixed field and the matrix
order m is at least 3) the matrix is said to be realizable and it can be synthesized
in the ladder topology.
Any matrix in S£m(R) when R is an Euclidean domain is realizable. Thus,
though any element of S£m(K[Zl]) is realizable for any positive integer m , the
fact that not all elements of S£2(K[Zl,Z2]) are realizable follows directly from
the work of Cohn [215] who showed that the matrix,
V(z z)
1, 2
= [ 1 + ZlZ2
-Z~
zr ]
l-Z 1Z2 (5.81)
cannot be expressed as a product of elementary matrices. It can, however, be

proved [214] that though the above matrix, is not realizable as a product of 2 x 2
elementary matrices over the base ring K[Zl' Z2], it can be embedded in the matrix
which is realizable in the group S£3(K[Zl' Z2]) of all (3 x 3) unimodular matrices

over the base ring. Besides, an algorithm has been developed that determines pre-
cisely when a given matrix in S£2(K[Zl, Z2,··· ,znD allows a factorization into
products of elementary matrices and, in that case, also produces the factorization
[216]. Cynthia Woodburn has shown that the unimodular matrix,
M = [
1 + V2Z1Z2 + zrz~
zi 1-
zt
V2Z1Z2 + z~z~
0
0
1
o 0 1
can be factored as a product of 109 elementary matrix factors. Let E3 (K[Zl, Z2]) be
the group under matrix multiplication generated by the elementary matrices with
polynomial entries from K[ZI' Z2]. She found matrices L, R E E3(K[zI, Z2])
such that LM R = I so that M = L -1 R- 1 will yield a factorization of Minto
109 elementary matrices. In this case, M = L -1 R- 1 =
e2dzj(l - v'2z1z2 + z~z~»e23( -z~zj)e23( -1)e32(l)e23( -1)
.e21 (zlzj(l - v'2z1z2 + z~ z~))e23( -zfzj)e23 (-1)e32(1)e23( -1)
.e21 (z~zi{1 - v'2Z1z2 + z~z~))e23( -zfzi)e23 (-1)e32(1)e23( -1)
.e21 (zfzi(l - v'2z1z2 + z~ z~))e23( -zfzi)e23 (-1)e32 (1)e23( -1)
.e21(1- v'2z1z2 + z~z~)e12(-1- v'2z1z2 - z~z~)
.e32( -1)e23 (1)e32( -1)e21 (-v'2z2 + zlz~)e12(zr)e21 (-1)e12(1)e2r( -1)
.e12(l)e12(l)e21 (-1)e12(l)e12(1)e2d -1)e12( -v'2z2 - zlz~)e23( -1)
.e32(l)e23( -1)e23(1)e32 (zi zi)e31 (-Zl (1 - v'2Z1z2 + z~z~»
.e32( -1)e23(l)e32( -1 )e21 (-v'2z2 + zlz~)e12(zr)e21 (-1)e12(l)e21 (-1)
.e12(1)e12(1)e21 (-1)e12(l)e12(1)e21 (-1)e12( -v'2z2 - zlz~)e23( -1)
.e32(1)e23( -1)e23(zr)e32(zf zi)e31 (-Zl (1 - v'2z1Z2 + z~z~))
.e32( -1)e23(1)e32( -l)e21( -v'2z2 + zlz~)e12(zr)e21( -1)e12(1)e21 (-1)e12(1)
.e12(l)e21 (-1)e12(1)e12(1)e21 (-1)e12( -v'2z2 - zlz~)e23( -1)
.e32(1)e23( -1)e23(Zne32(z~zi)e31 (-Zl (1 - v'2z1z2 + z~ z~))
.e32( -1)e23 (1)e32( -1)e2r( -v'2z2 + zlz~)e12(zl)e21 (-1)e12(1)e21 (-1)e12(1)
.e12(1)e21 (-1)e12(1)e12(1)e21 (-1)e12( -v'2z2 - zlz~)e23( -1)e32 (1)
.e23( -1)e23(zf)e32(zlzi)e31 (-Zl (1 - v'2z1Z2 + z~ z~))
.e21 (-1)e12(1)e21 (-1)
8. Applications in Multidimensional Systems Stabilization

Parametrization of stabilizing controllers for two-dimensional systems was first
reported in 1985 (see Chapter 2) and since that time considerable progress has been
made even for models in which some stabilizable transfer matrices do not have
rightlleft coprime factorizations or matrix-fraction description (MFD). The MFD
description approach, however, provides, conveniently, a parametrization of all
stabilizing compensators. The appropriateness of Grobner bases over polynomial
rings and modules continue to be exploited in the development of computational
methods for output feedback stabilizability and stabilization of 2-D systems [133],
[217], matrix-fraction descriptions of multidimensional systems (when such de-
scriptions exist) with tests for various primeness or coprimeness issues (Chapter
2) [203], [202],[107], and construction of strong stabilizers of multidimensional
systems and related problems [136], [135].
Though the polynomial ring based approach has been emphasized in this chapter,
the definitions of coprimeness and coprime factorizations can be over a more
general commutative ring as in [131, Chapter 8] and in Chapter 2. To wit, if
R denotes a commutative domain with identity, F denotes the field of fractions
associated with R, and M(R), (M(F)) denote, respectively, the set of matrices
with elements in R, (F), then A, B E M(R} are (zero) right coprime if there exist
X, Y E M(R} such that the right Bezout identity
XA+YB=[
holds. Suppose G E M(F}. Then, a pair (N, D) where N, D E M(R) is a (zero)
right coprime factorization of G if
(a) D is square and detD ~ 0,
(b) G = ND- l ,
and N, D are zero right-coprime. Zero left-coprimeness and zero left-coprime
factorization are defined as standard variants of above. The coprime factorizations
are referred to as rightlleft matrix fraction descriptions (MFDs). The MFDs may
not exist unless R is a Bezout domain i. e. every finitely generated ideal of R is
principal. The element generating this principal ideal is the GCD. Since the GCD
exists for any two or more elements in a Bezout domain, it is also a GCD domain
(a domain in which any finite set of elements has a GCD). Matrices of right size
whose elements belong to K[Zl' Z2] belong to a GCD domain but not a Bezout
domain (from results in Section 3 of this chapter). However, right-sized matrices
with elements in K[Zl, Z2, Z3] may not belong to a GCD domain (see an example in
[40]). The MFD approach, when the MFDs exist, is very powerful in problems of
stabilizability, stabilization, and parametrization of all stabilizing compensators of
an unstable multi variable plant, irrespective of whether it is continuous or discrete,
lumped or distributed, l-D or n-D.
DEFINITION 5.23 Let P E Kmxl(z}. Then P is said to have a double coprime

factorization (DCF) if there exist
(a) N l , N2 E KmXl(z), a nonsingular Dl E K!Xl(z), and a nonsingular D2 E
K~xm(z);
(b) Y l , Y 2 E K!xm(z), a nonsingular Xl E K~xm(z), and a nonsingular

X 2 E K!xl(z); such that
P = NlDll = D2l N2
and the matrix equation below hold\'
Ol,m]
1m .
In the l-D (due to Youla-Jabr-Bongiorno (1976» and 2-D cases (due to Guiver-
Bose (1985), Chapter 2), the parametrization problem of all stabilizing controllers
of a multivariable plant is completely solved. For possibility of parametrizing
all stabilizing compensators of a plant whose transfer matrix has elements in the
quotient field of another integral domain, namely Hinfty using coprime MFDs,
see [218].
Based on the bivariate polynomial matrix approach in Section 3 of this chapter
consider the MFDs of a bivariate rational matrix. Let H(ZI, Z2) E Kmxl(ZI, Z2).
Then, H(ZI, Z2) can always be factored as
H(ZI, Z2) = NIDII = Di l N2 (5.82)
where NI, D I , D 2, N2 are bivariate polynomial matrices of suitable size. Equation

(5.82) may be rewritten as
(-N2 D2][~~]=0
For specified N, D with elements in K(ZI, Z2]' The syzygy module V is the set of
polynomial row-vectors P(ZI, Z2) such that
Di l N2 is a left coprime MFD of H(ZI, Z2) if and only if the rows of( - ih D2 ]
form a basis for the module V. The generalized resultant matrix has been used
in [219] to get the coprime MFDs in the bivariate case. Of course, the procedure
described in Section 3 of this chapter could also be used for the purpose.
In the n-D (n ~ 3) case, it can be shown that stabilizability is a property that
lies between freeness and projectivity of certain modules associated to the plant
[107, Chapter 3] and the proof of admitting a DCF has only been solved in special
cases. Suppose that the plant P E Kmxl(z) admits a DCE Then, the set of all
stabilizing controllers, C, are given (like in the 2-D case, presented in Chapter 2)
by
C = (DIS + Yd( -NIS + Xd- 1
= (X2 - SN2)(Y2 + SD2)-1
where det(-N1S + Xd t=. 0, det(Y2 + SD 2 ) t=. 0 and the parametrizer matrix
S E K~xm(z). However, in the n-D case, the rational matrix P(z) does not
always admit a minor right coprime decomposition. With S as the design param-
eter, instead of the controller C, a matching problem results. If the performance
function is affine in S, one has an interpolation problem. When performance in
addition to internal (structural) stability of the feedback control system is desired,
an n-D matrix Nevanlinna-Pick interpolation (for results in the bidisc case, see
[220]) [221] problem results. Helton [222] pointed out the connection between
the feedback stabilization problem for n-D plants considered in this chapter and
the n-D matrix Nevanlinna-Pick problem.
NOTATION 5.1 Let BHOO(m) denote the functions analytic on un and bounded
in the supremum norm by 1.
The Nevanlinna-Pick interpolation problem on U 2 is defined below.
DEFINITION 5.24 Given afinite setojpoints {{zi k) , z~k))} in U2 andnumberwk in

the complex plane, and a positive number" find afunction f (Zl, Z2) E ,BHOO(2)
which takes the value Wk at (zik), z~k)).
The Nevanlinna-Pick interpolation problem on U2 has a bivariate rational func-

tion solution if and only if a linear matrix inequality (LMD has a solution. There-
fore, currently documented LMI tools in the control literature are suitable for use,
especially because in H oo -control, the problem is to design a controller which
minimizes the largest energy error signal over all disturbances of £2- norm at most
1 subject to the constraint that the controller must make the closed-loop system
structurally stable with norm equal to at most a specified tolerance level, > O.
DEFINITION 5.25 Two polynomial matrices D, N with the same number oj rows
are said to be left coprime if the composite matrix, R = [N - D] is GFLP.
Right comprimeness is defined by transposition. The matrix R is left prime if and
only if it is a minimal left annihil~tor of agolynomial ma~x M = [Dr !iff
i.e.
(a) NDI - DNI = 0 and (b) if NDI - DNI = 0, then D = XD and N = XN
for some polynomial matrix X.
The submatrices {Dl,Nt} of M is an arbitrary right factorization of a multi-
variate rational matrix H = D-l N = NIDil. The result, det D = det{Dt} (up
to nonzero constant factors), which is known to hold in the univariate and bivariate
cases, fails to generalize to the trivariate case [203]. Therefore, classical MFDs,
which work in the univariate (I-D) and bivariate (2-D) cases, are inadequate in
the trivariate case. The two examples, namely Example 5.6 and Example 5.10,
introduced earlier, can be used to illustrate this other difference in the n 2 3 case
from what is known to hold when n = 1 and n = 2.
Example 5.17 In the right MFD of the rational matrix

-1
[ Zl 0
H = BA- 1 = [0 Zl]
Z2 Z3
]
the matrices A = [Zl 0], B = [0 Zl] are factor right coprime. In a left
Z2 Z3
MFD of the same matrix,
(5.83)
the matrices D = Z3 and E = [ - Z2 Zl] are obviously, factor left coprime.

However, in the irreducible MFDs in (5.83),
jDj = Z3 i= jAj = ZlZ3

The previous example also illustrates that left and right MFDs, where one is
minor coprime and the other not, do exist. This differs from the known result for
zero coprimeness : a rational matrix with a zero left coprime MFD also has a zero
right coprime MFD [41]. It has been further shown, recently, that the left and right
coprimeness degrees of a rational matrix are equal on the condition that each is at
least 2 [200].
Unlike in the I-D and 2-D cases, minor coprime factorizations are not necessary
and sufficient for stabilizability. By defining coprimeness over rings of fractions,
it is possible to give necessary and sufficient conditions for stabilizability in the
general case. Thus, stabilizability is characterizable as a property that lies between
freeness and projectivity of certain modules. By Quillen-Suslin theorem, projective
modules over polynomial rings are free. However, the question, over which rings
projective modules are free, is a fundamental problem of algebra. The Quillen-
Suslin theorem holds for the ring of stable rational funcions of one variable. It
can also be seen to hold for bivariate case (via coprime factorization of rational
matrices etc.[39]; see also Chapter 2).
The existence of double coprime factorization, parametrization of all stabilizing
compensators and construction procedures for DCFs have been reported for special
classes of 3 - D and n - D systems in [223] and [224], respectively. A compre-
hensive review of stability and stabilization of linear n - D systems together with
some open problems is given in [225].
A MIMO n - D, n > 2 system may not admit a minor coprime MFD and,
therefore, the 2 - D case results of Chapter 2 are not applicable. The results
available in the n- D case include those in [217] (using polynomial matrix algebra)
and in [202] (using commutative algebra and topology but over a general integral
domain). The stabilizability condition in the n - D MIMO case is equivalent to
the absence of zeros in ff'" of the maximal order reduced minors of [DL I NL],
where Dr;l NL is a left MFD (not necessarily minor left coprime) of P. Lin [217]
presented a method for obtaining a strictly causal stabilizing compensator using
the family of reduced minors of the causal plant. However, this method, though
simple conceptually, is not fully constructive because of the need for solving a
multivariate polynomial equation,
f3
L q{Z)bi{Z) s{z) (5.84)
i=l
where polynomials bl (z), ... , bf3{z) are the reduced minors specified and poly-
nomials CI (z), ... ,cf3{z), s{z) have to be found with s{z) devoid of zeros in the
closed polydisc. Sule [202] based his procedure for constructing a causal stabi-
lizing compensator for an unstable multi variable multidimensional strictly causal
plant on a family of elementary factors, defined by him, for the plant.
Consider the right and left MFD's (not restricted to be right or left coprime and,
therefore, always exist)
where
Define composite matrices
T = [Nh Dk]t E K(Hm)xm[z],

W = [-NL DL] E Kix(Hm)[z].
DEFINITION 5.26 Let 1't, ... ,Tp denote the family of all nonsingular (m x m)
submatrices of the matrix T. Define Bj £. T(Tj)-l, where the entries of Bj are
already in reduced form. Let Ii be the radical of the least common multiple of
all the denominators of Bj . Then, the set {It,··., Jp } comprises the family of
elementary factors ofT.
In fact, it can be shown that after proper ordering of the (£ x £) submatrices ofW,
the elementary factors of T are identical to those of W. This is similar to what is
known about the reduced minors ofT and W i.e. after proper ordering the families
of reduced minors of T and W are identical up to signs. Inspite of this similarity,
the family of elementary factors is quite different from the family of reduced
minors and the former is more difficult to compute. The stabilizability conditions
for a specified n-D system can be given in terms of either the set of elementary
factors or the set of reduced minors, i.e. zero coprimeness in un of the elements of
either set is both necessary and sufficient for structural stability. Construction of a
stabilizing compensator can also be feasible by either approach while the problem
of parameterization of all stabilizing compensators for an arbitrary stabilizable
plant remains to be fully solved in the n-D (n > 2) case.
The elements of this set of elementary factors, besides being redundant, are,
more difficult to compute than the family of reduced minors. Furthermore, Sule's
approach cannot parametrize all the controllers by a Q-matrix as in classical Q-
parametrization (Youia parametrization) of 1 - D plants and its nontrivial 2 - D
counterpart reported in Chapter 2. Mori and Abe [226] introduced the family of
generalized elementary factors to give a necessary and sufficient condition for
feedback stabilizability. The coordinate free approach in [202] and [226] does
not require the right-left-coprime factorization approach adopted in this chapter
but but is constrained by the need for an already obtained stabilizing controller
for parameterization of all compensators that stabilize a plant. Furthermore, a
Q-matrix (S-matrix in Chapter 2) of bigger size, in general, is required. The mul-
tivariate double-coprime factorization would produce a Q (S) which is of global
minimal size while the approach in [226], [227] only produce a local minimal
one, in the sense of global minimal and local minimal generators defined in [212].
Whether or not a plant that is output feedback stabilizable admits a minor coprime
MFD was an interesting question posed in [202] and [217]. Necessary condi-
tions for strong stabilizability for MIMO linear shift-invariant multidimensional
plants were presented in [136]. Sufficient conditions, for strong stabilizability of
classes of linear n - D systems which, if satisfied, lead to the construction of a
stable stabilizing compensator was given recently [228]. Necessary and sufficient
conditions for strong stabilizability of single-input-multiple-output (SIMO) and
multiple-input-single-output (MISO) systems were given in [136].
9. Behavioral Approach
In behavioral theory, a dynamical system is viewed primarily as a collection of
trajectories called a behavior, without reference to some specific representation
like inputs to and outputs from a system as in transfer matrix and state-space
descriptions. Therefore, the behavioral approach to the modeling of dynamical
systems is not tied in any input-output or input-state-output structure [229]. The
system, or its behavior, is a subset of time-trajectories, in the case of temporal
systems, or a subset of space trajectories in the case of spatial systems. A behavior
is also associated with a spatio-temporal system. The behavior, thus, consists
of a set of trajectories that map the independent spatio-temporal variables into
the dependent variables. The behavioral theory extends to infinite-dimensional
distributed systems. The main objects of study in [229] is the behavior of linear
time-invariant lumped systems. The trajectories w with the so-called external or
manifest variables as components in the behavior B is the kernel of a smooth map
Q (!) : (coo)l -+ (coo)m
for some (l x m)polynomial matrix Q(e) E JRixm [e] and Coo is the space of
smooth functions on Ilt The behavior B is the kernel representation
(5.85)
A different representation of the behavior B involves auxiliary or latent variables

in the d-dimensional vector z, namely,
B = { w E (COO (JR))l I M (!) w = N (:t) Z} (5.86)
where N(e) E ]R9 Xd [e], MCe) E ]R9 xl [e]. For COO-behavior, a process of elimi-
nation can be used to filter the latent vector z from the representation in 5.86 to
arrive at the kernel representation in 5.85. The behavioral theory in [229] for the
case of constant coefficient ordinary differential equations can be generalized to
the case of constant coefficient partial differential operators, where the differential
ring lR(a1 , ..• ,an] is the n-variate polynomial ring over lR of partial differential
operators. For brevity, it is denoted by lR(8]. The solution space could be the
space Coo on JR1I' and is an infinite dimensional topological vector space. Here, the
problem becomes much more complicated because the ring lR[a1 , •.. , an] is not a
principal ideal domain. The existence of kernel representation in this case follows
from the Fundamental Principle, which is a deep refinement of Hilbert's Nullstel-
lensatz, and is credited to Ehrenpreis and Palmadov for systems of linear constant
coefficient partial differential operators. Given a p(8) E lR[a1 , ... , an] £. lR(8],
its zeros are not points but functions f in some space F such that p(8)f = O. More
generally, a zero ofp(8) ~ (Pl(8), ... ,Pl(8)) in Fi is ai-tuple (ft, ... , h) such
that
In the case of constantcoefficientPDE's, the underlying ring A = qa1 , ... ,an]

is Noetherian. The solution space coo(lRn) is an infinite dimensional topological
vector space of smooth functions on lRn. The linear shift-invariant distributed
behavior B is
where matrices M(8) E APxi, N(8) E APxm. Via a process of elimination, the
behavior B has a kernel representation so that
The kernel representation is a consequence of the Fundamental Principle, which

states that the image of P(8) equals the kernel of another matrix Q(8) E AqxP.
Thus, solvability of the system of equations
P(8)x =y
for x E (coo)i given y E (Coo)P depends on the existence of Q(8) such that
Q(8)P{8) = 0, so that Q(8)y = O. Thus, one can solve for all y in the image
of the map in 5.87 and this image is contained in the kernel of another map
The Fundamental Principle states that the image of P( 8) actually equals the kernel
of Q( 8). Note that Q( 8) is finitely generated because q a1 , ... , an] is Noetherian,
and, therefore Q(8) is generated by all the relations, finite in number, between
the rows of P(8). Thee reader is referred to [230] for an exposition of the be-
havioral approach to multidimensional systems and for an example illustating that
a controllable n-D distributed system over the differential ring JR[al , ... ,an] is
not necessarily given by free submodules, when n ~ 3. Their counterexample
associates, with a matrix of partial differential operators, a (2 x 3) trivariate poly-
nomial matrix, which can easily be shown to be factor left prime. However, the
autoregressive behavior determined by the matrix of partial differential operators
is not controllable, because it is a submatrix of the matrix representation of the
curl operator whose associated polynomial matrix is not factor left prime, though
the system defined by the curl operator is controllable. It was seen that every
torsion-free module can be embedded into a free module. Similarly, a torsion-free
differential module (D-module) can be embedded into a free differential module.
Accordingly, among differential operators there is a subclass composed of those
that admit a parametrization. The differential module determined by an operator is
torsion-free if and only if this operator is parametrizable [198, Volume II, p. 614]
The curl operator in vector calculus is parametizable by the grad operator.
Let D = K[al , ... , an] be the ring of partial differential operators 8 =
[al , ... ,an]' LetG E cm xi [8] be a polynomial matrix. G = [gij] = [glg2 ... gi]
gives rise to a differential map
DG : eoo(JRn , ei ) -+ e OO (JRn , em)
given by the set of linear partial differential equations (PDEs)
DG[WI W2 ... wilt = Li g[

(a a ' ... , -a
-a ' -a a ) Wj
j=l Xl X2 Xi
after replacing ai by Ix; ,i = 1,2, ... , k, where Xi is a space or time variable.

View each row of G as an element of the free module Di. Let (G) denote the
D-submodule of D i , generated by the rows of G.
It is assumed that the behavior B is comprised of only smooth trajectories i.e.
The behavior B corresponds to some kernel representation given by the polynomial

matrix G. Here a multivariate polynomial is viewed as a constant coefficient
partial differential (polynomial) operator gij (8) so that instead of zeros of n-
variate polynomials in en, one is interested in the solutions (counterpart of zeros)
of constant coefficient linear homogeneous partial differential equations in function
spaces of interest. The matrix G (8) defines a D-module morphism. The questions
of interest are:
1 when is the kernel of this morphism equal to some image i.e. when is it exact?
and
2 when is the image of the morphism equal to the kernel of some differential
system?
The questions, in general, relate to the embedding of a partial differential system

in a two-sided complex of modules and maps, and, in particular to the finding of
the counterpart of Hilbert's Nullstellensatz in the function space of interest.
There exists a bijective correspondence between modules and behaviors if one
considers the space of smooth functions [110]. This is similar to the bijective cor-
respondence between radical ideals and affine varieties in polynomial ring theory,
exhibited through Hilbert's Nullstellensatz, seen in this and earlier chapters. How-
ever, in multidimensional behavioral system theory, the bijective correspondence
fails for some solution spaces, such as the space of tempered distributions [231].
Identification of spaces or necessary and sufficient conditions with respect to a
space for the bijective correspondence to hold has been a topic of some interest
[232].
The mathematical basis for multidimensional behavioral theory follows from
the work of Oberst [110] and the Fundamental Principle.
FACT 5.11 Thefollowing results hold for multidimensional linear shift-invariant

systems. All matrices have elements in A = R[z]. M is an A-module.
• For multivariate polynomial matrices G, H (factorization)
kerMH ~ kerMG {:}:3X such that G = XH
• Let G, H E Alixm(i = 1,2) be the matrices of normal full rank. Then

(uniqueness offull row rank kernel representations),
kerMH = kerMG {:} il = i2 and:3 X such that G = XH

• Given G and H with elements in A, then (elimination of latent variables)
G(kerMH) = kerMX for some suitable X
The preceding results stated in Fact 5.11 follows from the fact that M is a "large in-
jective cogenerator" over R[z], as proved by Oberst using the EhrenpreislPalmodov
theorem for injectivity.
10. Conclusions
A flurry of research activity has been witnessed in the last few years when
D = K[ZI, ... , zn], n ~ 3 and the problem is still open, though much better
understood than was the case a decade back. The three notions of primeness
described above are all different in this case. The notions of zero primeness,
unimodular matrix completion and polynomial matrix inverse are linked by the
Quillen-Suslin proof of the freeness of projective modules over polynomial rings
(referred to in the literature as Serre's conjecture) [42]. However, unlike zero and
minor primeness, the notion of factor primeness cannot be characterized exclu-
sively by the variety of the ideal generated by the maximal order minors in the ring
D = K[Zl' ... , Zn], n ~ 3 and also the exclusive consideration ofthe matrix full
rank condition is not sufficient. This necessitated the generalization of the factor
primeness concept to that of factor primeness in the generalized sense with both
being equivalent in the full-rank case [110]. Multivariate polynomial matrix prime-
ness, coprimeness, and greatest common divisor (if it exists) extraction results have
been advanced by several researchers in special cases [206],[207]. Oberst [110]
established the duality between multidimensional linear shift-invariant linear sys-
tems and finitely generated modules over D = K[Zl, ... ,zn]' He showed that for
n ~ 3, a transfer function might have two minimal matrix fraction descriptions or
realizations that are not comparable.
The recent theoretical developments in the multivariate polynomial matrix fac-
tor primeness problem has simultaneously stimulated applications, most notably
in multidimensional filter bank design [206]. A Groebner basis-based proof for
perfect reconstruction with linear phase was advanced for the two-band multi-
dimensional FIR (finite impulse response) filter banks. Although the problem
of two-band multidimensional linear phase perfect reconstruction filter bank has
been constructively tackled, the multiband counterpart demands further attention.
Further progress is desirable towards multidimensional wavelet construction from
filter banks and their use in the problem of image and video signal compression.
In 1977, A. A. Suslin went on to prove an analogue of his proof of Serre's con-
jecture over the polynomial ring E[z] = E[ Zl, ... , zn], where E is a commutative
Noetherian ring. This result is referred to in the literature as Suslin's stability the-
orem for which an algorithmic proof is available in the special case when E is a
field K [214]. Suslin's stability theorem provides the machinery for biorthog-
onal multiband filter bank realization for perfect reconstruction using the ladder
topology, via the use of Grobner bases. Precise conditions under which a given
multivariate polynomial matrix of order 2, for which Suslin's stability theorem
does not apply, is factorable as a finite product of elementary matrices and the
problem of constructively obtaining this type of factorization, in that case, were
advanced in [216].
Scopes for further applications of the developed theory include multidimen-
sional convolutional code construction with intriguing possibilities for influencing
compression and transmission of image, video, and multimedia signals over ban-
dlimited channels, wired as well as wireless. Though the algorithms in [196] and
[39] always apply in the I-D and 2-D cases, respectively, such a conclusion is
not valid when D = K[Zl, ... , zn], n ~ 3. The field K is application-dependent
(for example in multidimensional convolutional coding, K must be a finite field).
Polynomial matrix representation of convolutional codes is well-known [233]. The
parity-check polynomial matrix with elements in (GF2) associated with convolu-
tional code generator polynomial matrix is directly linked to the syndrome, which
in turn belongs to the syzygy of solutions or syzygy module of the code generator
matrix. Also, when the right inverse of this code generator matrix exists, the code
is called a non-catastrophic convolutional code. All catastrophic convolutional
codes suffer error propagation effects. Inspite of the documented diverse appli-
cations of one-dimensional convolutional codes [234] in space communications,
satellite communications, mobile communications, and voice-band data commu-
nications, the scopes and promises of multidimensional convolutional codes have
been realized providing, therefore, another fertile arena for the in-depth pursuit of
recent results in multivariate polynomial factorization.
Convolutional codes are often used to improve the performance of radio and
satellite links. See [235] and [201] for definitions of multidimensional convolution
al codes.
DEFINITION 5.27 A convolutional code is a type of error correction code in which

(a) each k-bit information ,\ymbol (each k-bit string) to be encoded is transformed
into an m-bit ,\ymbol, where m 2: k and
(b) the transformation is afunction of the last k information ,\ymbols, where k is
the constraint length of the code.
Convolutional codes may be viewed as discrete-time behaviors, the main dif-
ference being that in the context of coding one works over a finite field instead of a
real field. The duality between convolutional codes and a class of behaviors called
complete behaviors follows as a special case of a seminal result of Oberst [110,
p.62] which states that a behavior, comprised of the trajectories of the system, has
a kernel representation if and only if it is complete.
Details on 2-D and n - D codes can be found in [233] and [236]. They provide
error detection and correction capabilities in the transmission of multidimensional
data like 2-D pictorial data, 3-D motion picture, and 4-D animated holograms.
Given a finite field, K q, consider the polynomial ring R = Kq[Zl, Z2, ... , zn] in
n determinates over K q . A generic message word is a vector fj E K: and can be
associated with a vector Fj(z) E K:(z) = Rk after applying the multidimen-
sional z-transform operator. Note that R is Kq-isomorphic to the n-dimensional
finite sequence space
S = {J : N'" -+ Kq I f has finite support},

where ~ is the integer lattice. To realize an encoding scheme that is shift invariant
with respect to the coordinate axes in multidimensions, one needs to construct
an injective module homomorphism 4> : Rk -+ Rm. Clearly, 1m (4)(R k )) is a
submodule of Rm. Since R is a Noetherian ring, each code C c Rm is finitely
generated. This implies that for the n-D C of length £, there is a positive integer
m and a polynomial matrix F E Rmxi, called the generator matrix of C so that C
= ImRF = {Fv I v E Rl}. C has a well-defined rank, say k = rank F. The rank
190 MULTIDiMENSIONAL SYSTEMS THEORY AND APPLICATIONS
k (the largest integer k for which there is a nonidentically zero k x k minor of F)

of the rowspace (F) is the dimension of the vector space generated by the rows of
Fin Q{R)n over the field of fractions Q{R) of R and the rate of C is kim. An
n-D convolutional code, C, of length I is an R-submodule of the free module Rl,
the space of i-component row vectors with entries in R. For the code C to be a
free R-module, it must have a R-basis i.e. C is required to have a generator matrix
G E Rmxk, where rank G = k = rank C. With G = [gI,g2, ... ,gk] E Rmxk
satisfying this property, it is called an encoder and each codeword can be written
as a R-linear combination of vectors gI, g2, ... , gk. This is always possible for
the K[zl]-module but not for the K[Zb ... ,zn]-module. The n-D convolutional
code C isfree if it is free as an R-module and, in that case, G is an encoder of C.
Any 1-D convolutional code is free because in that case the polynomial ring R is
a principal ideal domain. The developed algebraic structure of 1-D convolutional
code have been investigated with regard to possibilities and limitations in the
nontrivial generalization to multidimensions. The many differences between the
algebraic structure of 1-D and 2-D codes and between 2-D and n-D (n > 2) codes
are akin to the difference in notions of primeness and factorization, discussed in
this chapter, as one proceeds from the 1-D to 2-D and subsequently, to the n-D
(n> 2) case.
Finally, the relationship between multidimensional convolutional codes and
multidimensional behavioral systems, a subject not formally introduced very briefly
in this Chapter to keep its size within page constraints set, has been appreciated
starting with the dissertation of Paul Wiener [201].
Chapter 6
RECENT IMPACTS OF MULTIDIMENSIONAL

SYSTEMS RESEARCH
BYN. K.BOSE
1. Introduction
A polynomial or a rational function (matrix), characterizing a single-input

single-output (multi-input multi-output) system, has the coefficients for param-
eters. The number of such free parameters defines the dimension of the space and
a system with fixed parameters may be represented by a point in parameter space.
If at least one coefficient varies about its nominal value, a region in parameter space
is generated. This region characterizes a family of systems instead of one fixed
system. When the coefficients vary independently of each other within specified
compact intervals, an interval !Jystem is generated. A well explored case when the
coefficients do not vary independently occurs when the region in parameter space
is a bounded polyhedral set. A polyhedral set is formed from the intersection of a
finite number of closed half-spaces and could be unbounded. A bounded polyhe-
dral set is a convex polytope and vice versa. For an interval system, the polytope
degenerates into a boxed domain or a hyper-rectangle. Extensive documentation
of research results concerned with the extraction of information about the complete
polytope from a very small subset of the polytope with respect to the property of
stability for both continuous-time and discrete-time systems is available in several
recent texts ([237], [238], [239], for example), since Kharitonov's trend-setting
publications [240] followed by the edge theorem[241]. The goal is to obtain tests
for invariance of useful properties of sets of distinguished classes of functions
from tests on a small subset of such functions. Properties of concern in control
and signal processing include stability and frequency response.
191
2. Inference of Stabilty of Sets of Multidimensional Systems

From Subsets of Low Cardinality
Over the last fifteen years many results on stability involving the study of zero-
sets of polynomials have been generalized to the multivariate situation, where
different types of stability conditions (that become equivalent in the univariate
case) appear. Here, starting with the enunciation of Kharitonov's pathbreaking
result on complex coefficient univariate interval polynomials [240], the nontrivial
generalizations to the multivariate case are described.
Consider the complex interval polynomial of nth degree, in the complex variable
s,
n
P(s) = ~)[ak,ak] + i[Qk, bkDs k. (6.1)
k=O
The set of polynomials derived from (6.1) by specializing, respectively, the real and
imaginary parts of each coefficient to either its lower or upper bound is defined to
be the set of extreme (vertex) polynomials. The cardinality of this set is obviously,
at most, 4n+ 1 . Each polynomial belonging to the set P( s) can be represented as
a point in R2n+2. In parameter space, P(s) in (6.1) is associated with a boxed-
domain, whose vertices correspond to the extreme polynomials. We define the
zero-set Z(P) of P(s) in (6.1) as:
Z(P) A {s I p(s) = 0, p(s) E P(s)}, (6.2)
where p{s) E P{s) is a polynomial with fixed coefficients. A polynomial p(s) E

P{ s) is called strict Hurwitz (SH) if all its zeros are in the open left-half plane and
P{ s) will be defined to have the SH property if each of its elements is SH.
Kharitonov showed how the SH property of a set of complex interval polynomi-
als may be established by testing a specific set of eight extreme complex coefficient
[240] polynomials for the SH property. Kharitonov's results were proved via a net-
work realizability theory approach in [242] based on the isomorphism between the
the space of SH polynomials and the space of reactance functions. A Kharitonov-
type of result wa obtained in [243] for an interval set of multivariate scattering
Hurwitz polynomials, encountered in the topic of multivariate realizability theory.
2.1 UNIVARIATE COMPLEX HURWITZ POLYNOMIAL CASE
In this case it is required to determine the conditions which are necessary and
sufficient for the zero-set of P( s) to be in the open left-half plane. In the case when
~ = bk = 0, k = 0,1, ... , n it was shown by V. L. Kharitonov that four extreme
polynomials Aj{s), j = 1,2,3,4 whose coefficients in ascending powers of s are
Recent Impacts of Multidimensional Systems Research 193
given below, suffice.
Qo al a2 ~ ~ a5 a6
Qo Q.l a2 a3 ~ Q.5 a6
(6.3)
ao al ~ ~ a4 a5 Q.s
ao Q.l ~ a3 a4 Q.5 Q.s
For n = 2, the coefficients are required only to be of like sign. For n = 3,4, 5, the
numbers of extreme polynomials required to be tested are 1,2, and 3, respectively.
For n ~ 6, all the polynomials are required in general. For the general case in
(6.1), define a set {Cj,k(8)} of sixteen extreme polynomial,
j, k = 1,2,3,4 (6.4)
where the coefficient of Bk(8), k = 1,2,3,4 are similarly patterned as the coef-
ficients of Ak(8). It was proved in [240], [242] that the interval polynomial has
zero-set in the open left-half plane if and only if each of the eight extreme polyno-
mials C1,2(8), Cl ,3(8), C2,1 (8), C2,4(8), C3,1 (8), C3,4(8), C4,2(8) and C4,3(8) is
Hurwitz, i.e. has zeroes in the open left-half plane. Since only a subset of the ver-
tices of the boxed domain in parameter space is required in the test of the Hurwitz
property for P(8), we call Kharitonov's results the "reduced vertex results".
2.2 MULTIVARIATE SCATTERING HURWI1Z POLYNOMIAL CASE
To facilitate reading and for the sake of brevity, the notation
p = (pl,P2, ... ,Pn)

will be used to denote a set of n complex variables Pl,P2, ... ,Pn. This type of
polynomial was also cited in Chapter 1.
DEFINITION 6.1 A polynomial g(p) will be called a scattering Hurwitz polyno-

mial or, equivalently, a principal Hurwitz polynomial, if it satisfies the following
properties:
(a) g(p) i= 0, for Re p > 0, i.e. in the open right half polydomain and
(b) g(p) is relatively prime to the paraconjugate polynomial,
gr(P) £ g*(-p*) = g*( -pr, -P2,···, -p~)
where the asterisk * represents the operation of complex conjugation.
ml m2
B(Pl,P2) = L L bklk2Ptlp~2 (6.5)

kl=O k2=O
in the two complex variables PI and P2, wlwse coefficients belong to the field of real
numbers. This restriction is imposed only for the sake of brevity in exposition; the
main ideas enable one to also handle the complex coefficient case. The polynomial
in (6.5) can be rewritten as a sum of homogeneous polynomials,
m
B(PI,P2) =L ak(PI,P2), (6.6)
k=O
where m ~ mi + m2 and ak(PI,P2) is a homogeneous bivariate polynomial of
degree k for k = 0, 1, ... , m. The objective in [243] was to establish a criterion to
test for the scattering Hurwitz property of a set of interval polynomials generated
from (6.5) by allowing each coefficient to vary within the closed interval defined
by
bklk2 E [fuclk2' bklk2] . (6.7)
It is easy to see that in (6.6),
k
ak (P I, P2 ) ="
L...Jbi(k-i)PIP2
k k-i . (6.8)
i=O
The immediate objective is to assign to bi(k-i) either the value bi(k-i) or Qi(k-i)
so that, in the case of keven, ak(jwI,jw2) attains respectively its maximum
and minimum values for an arbitrary but fixed pair (WI, W2) when the interval of
variation of each coefficient in ak(PI,P2) is as given in (6.7). For example, when
k = 2,
a2(PI,P2) b02P~ + bl1PIP2 + b20P~ ,
a2(jwI,jw2) -b02W~ - bl1WIW2 - b20W~,
For an arbitrary but fixed pair (WI, W2), the following assignments of coefficients
lead, respectively, to the maximum and minimum values of a2(jwI,jw2) in the
closed first and third quadrants of the (WI, W2) plane:
l!.o2 Ql1 Q20

b02 bn ~O
The corresponding assignments for an arbitrary but fixed value of (WI, W2) in the
closed second and fourth quadrants are
l!.o2 bn Q20
b02 Qn ~O
For the sake of brevity in notation, with keven, ak(PI,P2) and ~(PI,p2)' re-
spectively, will denote the assignments of coefficients in ak(PI,P2) which yield
maximum and minimum values of ak(jwI,jw2) for an arbitrary but fixed pair
(WI, W2) in the sub-domain of interest (when k is odd, the words "maximum" and
"minimum" are replaced by "maximum of the imaginary part" and "minimum of
the imaginary part"). Again, for conciseness, the case when m is even in (6.6)
is considered; the situation when m is odd can then be similarly tackled. The
following result was proved in [243].
THEOREM 6.1 The set ofpolynomial defined in (6.5), (6.6), and (6.7) are scatter-
ing Hurwitz if and only if the following polynomials are scattering Hurwitz:
Blj(P) = am(p) + ~m-l (p) + ~m-2(P) + am-3(p) + am-4(p) + ... ,

B2j (p) = am(p) + ~m-l (p) + ~m-2(P) + ~m-3(P) + am-4(p) + ... ,
B1j(p) = ~m(P) + ~m-l (p) + am-2(p) + am-3(p) + ~-4(P) + ... ,
Blj(p) = ~m(P) + am-l(p) + am -2(p) + ~m-3(P) + ~m-4(P) + ... ,
where again, the subscript j is associated with he quadrants in the (WI, W2) plane
andp £ (PI,P2).
The proof was motivated by the isomorphism between scattering (or princi-
pal) Hurwitz polynomials and multivariate reactance functions, studied in depth
by Fettweis [244], because of their central role in the scattering transfer matrix
description of multivariate lossless two-ports encountered in the synthesis of refer-
ence filters prior to the derivation of wave digital filter structures. In [243], it was
noted that there are dependencies in the preceding set of sixteen polynomials (from
the four quadrants). These dependencies are due to the redundancies inherent in
the enforcement of upper and lower bounds in the first and third quadrant (and also
in the second and fourth quadrants). So, the number that is necessary and sufficient
for the purpose is reduced from 16 to 8. In [243], it was stated that the number
of extreme n-variate polynomials whose scattering Hurwitz property suffices to
test for the same property in a set of real (complex) coefficient n-variate interval
polynomials is 4(2n-l) (8(2n-1 )). These results are further discussed in [245]. In
[246], some basic results in [243], [245] were extended to the case of strict sense
stable n-variate polynomials, defined next.
DEFINITION 6.2 A polynomial g(p) is strictest sense Hurwitz stable (SSS )(widest
sense Hurwitz stable (WSS» if g(p) i= 0 for Re P ~ 0 (Re p > 0), i.e. in the
closed right-half polydomain (open right-half polydomain).
In the univariate case the classes of strict sense (Hurwitz) stable (SSS) and scat-
tering Hurwitz stable (SHS) polynomials coincide while in the multivariate case
the class SSS is a proper subclass of SHS. Other three classes of multivariate poly-
nomials, not of direct interest here, have, for their respective generic element, a
se!fparaconjugate polynomial (a widest sense Hurwitz polynomial that equals its
paraconjugate upto a complex constant of unit magnitude), a reactance Hurwitz
polynomial (a self-paraconjugate Hurwitz polynomial without multiple irreducble

factors), a immittance Hurwitz polynomial (product of a scattering Hurwitz poly-
nomial and a reactance Hurwitz plynomial). In [247], it was shown that arbitrary
small coefficient variations may destroy SSS stability and to overcome this unde-
sirable situation a new class of n-variate stable polynomials was defined. Given
a set of nonnegative integers, (mil m2, ... , m n ), define the set S of multivariate
polynomials, whose generic element is
mt m2 mn
B(PI,P2, ... ,Pn) = B{p) = L L '" L bktk2 ... knPttp~2 ... p~n. (6.9)
kt=Ok2=O kn=O
The polynomial B (p) in (6.9) can be written in recursive canonical form in the main
variablepk, k = 1,2, ... , nasB{p) = E:kO b~k){PI' .. ' ,Pk-I,Pk+1,'" ,Pn)pt,
whose mam . coeffi' . b(k)
clent IS ( )
mk PI,··· ,Pk-I,Pk+I,··· ,Pn .
DEFINITION 6.3 The n-variate polynomial B(p) in (6.9) will be called robust sta-
ble provided it satisfies the following conditions: (a) B{p) is SSS, (b) the main co-
efficients, b~l (PI, ... ,Pk-I,Pk+I,··· ,Pn)for k = 1,2, ... , n are (n-1)-variate
robust stable polynomials and (c) deg{b~l) = (ml"'" mk-l, mk+l,"" m n )
for k = 1, 2, ... , n.
The robust stable class of polynomials is the largest class of polynomials pre-
serving stability under small coefficient variations [247]. It has been shown in
[248] that the vertex implication, in coefficient space, for SHS and SSS properties
of multivariate interval polynomials [243], [245] hold also for the robust stable
property. The dual version of this result (i.e. for the diamond family in coefficient
space) also holds [247], [249]. The robust stability property is invariant under par-
tial differentiation but the SSS property is not. For example, the SSS polynomial
PIP2 + P2 + 5 loses the SSS property after partial differentiation with respect to
Pl. The reader may wish to answer whether or not the SHS and WSS properties
are invariant under the operation of partial differentiation.
The phase function ljJ{WI, W2, ... , wn ) ~ ljJ{w) associated with a scattering
Hurwitz polynomial B{p) is defined as
ljJ{w) = tan- l !~:~ ~ argB(jw) (6.10)
where B(jw) ~ a{w) + jb{w). The following result, proved in [250], has been
used to reduce stability testing sets for polyhedral families of multivariate polyno-
mials [251]. Previously, in [248] and [247], the result below was also used together
with the notion of convex directions (for multivariate polynomials) in robust sta-
bility studies of interval and diamond families of multivariate polynomials in terms
of a finite number of vertex polynomials of the respective families.
THEOREM 6.2 The phase function </J(w) associated with a n-variate scattering
Hurwitz polynomial B(p) satisfies, for all real valued vectors w where </J(w) and
its partial derivatives are well-defined, the inequality,
~I l°</J(w) ISin[2</J(W)]I (6.11)

L- Wk ow ~ 2 .
k==l k
The n-variate polynomial B(p) is an n-variate convex direction if the inequality
~I lo[argB(jw)] ISin[2[ar g B(jW)]1 (6.12)

L- Wk ow ~ 2 .
k==l k
holdsfor all values ofw E ]Rn, where B(jw) i= 0.

Next, consider the convex combination B 12 (P) of two n-variate real-coefficient
SHS polynomials Bl(P) and B2(p), each of degree m = (ml,m2, ... ,mn ),
where mi is the partial degree in Pi, i = 1,2, ... , n.
B 12 (P) = >'Bt{p) + (1 - >')B2(P), >. E [0,1].

Rewrite B12(P), after defining Bt{p) - B2(P) ~ A 12 (P), as
B1 2(P) = B2(P) + >.A 12 (P), >. E [0,1]. (6.13)
The following theorem, proved in [250], establishes the SH stability of an edge

in coefficient space from the SH stability of its end points.
THEOREM 6.3 Suppose that B t{p ) and B2 (p ), cited above, are SHS. Then B12 (p)
in (6.13) is SHSfor all >. E [0,1], if the in equality,
E Iw.laarg~:(iW)J ~ ISin{2arg~AJ2(;W)]}1
hold for all real-valued vectors w, where the partial derivatives are well-defined.
The sufficient condition in the preceding theorem becomes both necessary and
sufficient in the univariate case. In [251], the preceding result which gives a
sufficient condition for convex directions, has been encountered in studies of robust
stability of conic sets of multivariate polynomials. The stability (in the sense of
Definition 6.3) theory of conic sets of polynomials is more general than that of
compact sets. The frequency domain approach, including value sets and the zero-
exclusion principle, could not be included and the reader is referred to [252] for
a reformulation of zero-exclusion principle for application in the robust stability
study of multivariate polynomials and to [246] for earlier results. The main tool
in the frequency domain approach to robustness analysis is the zero-exclusion
principle. Not only both continuous and discrete space-time systems with real or
complex parameter perturbation (interval, disc, poly topic etc.) are tackled by this
appraoch but more general D-stability problem with reference to a poly domain D
can be considered. See [253] for the frequency domain criteria in the case of 2-D
systems.
2.3 ROBUST WAVENUMBER RESPONSE
Robustness property is crucial in the performance evaluation of spatio-temporal

systems used to process analog video signals, multidimensional transform-based
coding and in multidimensional filtering for picture quality improvement. Bounds
on the wavenumber response of multidimensional digital filters, where the uncer-
tainty structure for the coefficients could be of the interval, poly topic, and nonlinear
types are important to obtain. Such filters are important in video data processing.
The purpose of the filtering, which can be one, two, and three-dimensional is to
improve NTSC by limiting the crosstalk that would otherwise occur between lu-
minance and chrominance components. Finite impulse response (FIR) filters that
are widely used because of their linear phase properties are inherently stable. For
greater speed of implementation, when infinite impulse response (IIR) filters are
employed, the analysis of robust wavenumber response should be coupled with
robust stability. The wavenumber response of a two-dimensional (2-D) finite im-
pulse response (FIR) filter, characterized by the unit impulse response bisequence
{h[k 1, k 2]} of finite support is
H(Wl,W2) = LLh[kl,k2]exp(-j(w1k1 + W2k2)).

kl k2
The Fourier transform of the bisequence {h[k 1, k 2]} given above contains trigono-
metric terms of the following four types:
cos (wlkI)cos (w2k2)' sin (w1kI)sin (W2k2),
cos (wlkI)sin (w2k2), sin (w1kI)cos (w2k2).

In the special case, when for all indices kl and k2 that define the region of support
of the filter unit impulse respose,
holds, the wavenumber response simplifies to a product of I-D filter frequency

responses and the 2-D filter is then called product separable. Otherwise, the filter
is nonseparable. The 3-D and n-D counterparts of the preceding formulations are
straightforward. In general, fewer operations are required to realize a separable
2-D (or n-D) filter than an arbitrary 2-D (or n-D) filter ofthe same order. However,
the price to be paid is a potential loss in performance because of constraints on the
type of response that can be achieved. In particular, if filters with rapid cut-offs
are desired, the passband of separable filters will have a rectangular shape, which
may be inadequate if circular, aster, or diamond-shaped passbands are desired as
in multidimensional sampling for image and video applications.
Some degree-dependent vertex implication results are available when the prop-
erty of interest is the frequency-wavenumber response. Interval trigonometric
polynomials and the ratios of such polynomials are useful for assessing the robust
frequency responses of, respectively, FIR and IIR digital filters. For such studies
it has been concluded that degree-dependent vertex-implication results hold in the
univariate case [254]. Such degree-dependent vertex implication results hold also
for some special classes of two-dimensional interval digital filters, called product
separable filters. Latest results are documented in [255], where consequences of
imposing various symmetry conditions in the unit impulse response of multidi-
mensional FIR filters on the bounds of wavenumber responses of such filters are
investigated. Fourfold and eightfold symmetry constraints on the unit impulse
response are satisfactorily tackled for lower order filters. Higher order cases and
the twofold symmetry case remain as difficult open problems from the compu-
tational complexity standpoint. Amplitude and phase envelopes of an interval
rational transfer function of continuous-time and discrete-time systems have also
been considered in [256], [257], respectively.
2.4 STABILITY OF INTERVAL MATRIX
See [258] for an early discussion of this subject. Let [A] be an interval matrix,
[A] = [A A], A ~ A, with A, A E Rnxn and A = (~ij), A = (aij). The ex-
treme matrices (or vertex matrices) of [A] are matrices of the form A = (aij) with
aij E {~j, aij} for all i, j. The set of the extreme matrices of [A] is denoted by
V[A]. Vertex implication enables us to infer the satisfaction of a certain property
of an interval matrix from the satisfaction of this property for certain vertices of
the matrix interval. A vertex implication is known to be true for the classes of the
inverse-nonnegative (or monotone) matrices and M-matrices for some subclasses
of the totally nonnegative matrices and for the classes of the P-matrices and diag-
onally stable matrices. The number of extreme matrices involved in these vertex
implications are 2,2,2, 2n - 1 , and 2n (n-l)/2, respectively.
Ifr c R nxn we denote by a(r) the spectrum ofr.
a(r) := {A E Cl3x E C n {O}, MEr: Mx = AX}

An interval matrix [A] is called HUlWitz (Schur) stable if a([A]) is contained in
the open left-half of the complex plane (inside the unit disk). It is well known
that the Hurwitz stability of the extreme matrices of an interval matrix does not
imply the Hurwitz stability of the matrices contained in the interval matrix. The
fundamental differences between the polynomial stability problem and the matrix
stability problem was illustrated through counterexamples in [259]. Both the poly-
tope of matrices and the interval matrix cases were considered. Counterexamples
were given to justify that the edge theorem [237], [238] for stability of a poly-
topic set of polynomials does not generalize to either a polytope of matrices or a
hyper-rectangle of matrices (interval matrix). Characterizations of Hurwitz and
Schur stability, among other properties, of interval matrices are given in [260]. It is
proved that a symmetric interval matrix is positive definite (Hurwitz stable, Schur
stable) if and only if it contains at least one symmetric matrix with the respective
property and is nonsingular (for Schur stability, two interval matrices are to be
nonsingular). As a consequence, verifiable sufficient conditions are obtained for
positive definiteness and Hurwitz and Schur stability of symmetric interval matri-
ces. These results are useful in robust stability analysis in a state-space setting.
For further results on stability of interval matrices and polytopes of matrices and
some problems, still open, in this area, see [237]. There has been some attempts
at developing a necessary and sufficient condition leading to a computable test for
checking the robust stability of interval matrices [261] but a complete proof for
such a result has not appeared as of now in the open literature.
3. Multiple Deconvolution Operators for Robust

Superresolution
A very fertile arena for applications of some of the developed theory of mul-
tidimensional systems has been spatio-temporal processing following image ac-
quisition by, say a single camera, mutiple cameras or an array of sensors. Due to
hardware cost, size, and fabrication complexity limitations, imaging systems like
CCD detector arrays or digital cameraas often provide only multiple low-resolution
(LR) degraded images. However, a high-resolution (HR) image is indispensable in
applications including health diagnosis and monitoring, military surveillance, and
terrain mapping by remote sensing. Other intriguing possibilities include substi-
tuting expensive high resolution instruments like scanning electron microscopes
by their cruder, cheaper counterparts and then applying technical methods for
increasing the resolution to that derivable with much more costly equipment. Res-
olution improvement by applying tools from digital signal processing technique
has, therefore, been a topic of very great interest.
The low resolution images are either captured as a temporal sequence or simul-
taneously with different sensors. The high resolution filtered image is constructed
from the aliased (undersampled) noisy and blurred frames with subpixel shifts.
The process for obtaining a HR image from a sequence ofLR frames is referred to
as superresolution imaging. The term, superresolution, used here is different from
the usage in optics of the same term to imply recovery of information beyond the
diffraction limit. The HR image realizes an increase in spatial resolution (distinct
from temporal resolution that sets the frame rate or number of frames captured per
second), measured in pixels per unit distance. There is a fabrication limit on how
Recent Impacts o/Multidimensional Systems Research 201
small the size of the pixels can be because, besides cost constraints, the ubiquitous
shot noise lowers the SNR.
The exact deconvolution approach overcomes the ill-posedness associated with
the single image restoration problem by using multiple sensors whose PSFs (point
spread functions) are constrained to satisfy a specific condition. For the gen-
eral case of m multisensors in the real n-dimensional Euclidean space r, let
hI (x), h 2(x), ... ,hm(x) be the compactly supported blur PSFs, where x ~
[Xl X2 ... Xn]t, Xl E lR for i = 1,2, ... , n, and the superscript t denotes the
transpose operator. The following result provides a necessary and sufficient con-
dition for a compactly supported solution {d1 (x), d2(x), ... ,dm(x)} to exist so
that 'E.f=l hj(x) * dj(x) = o(x), where o(x) denote the Dirac function in lRn.
THEOREM 6.4 [262J For the compactly supported distribution hI (x), h2(X), ... ,
n
hm(x) inlR , there exist compactly supported distributions d1 (x), d2(x), ... , dm{x)
such that 'Ei hi{x) * di{X) = o(x) if and only ifpositive constants Gl, G2 and a
positive integer N can be found such that
pEen, (6.14)
where Hi(p) is the n-dimensional Laplace tramform ofhi(x), 1·1 is the absolute
value operator, p ~ [PI P2 .. , Pn]t is the complex-valued vector in the complex
n-dimensional Euclidean space en, 1m p denotes the imaginary part of p, and
IIpll ~ ('E~=l IpjI2)1/2 is the Euclidean norm.
A set of sensor blur PSFs hI (x), h2(X), ... ,hm(x) that satisfy the inequality in
the above theorem is referred to as strongly coprime (zero coprime). This strongly
coprime condition guarantees that the transforms of blur PSFs, hl{x), h 2 (x), ... ,
hm{x), have no common zeros and satisfy certain growth conditions for large IIpll
which allow recovery of the original signal in the noiseless case. For the uniform
distribution with circular support on lR2, it is known [262, page 726] that the PSFs
of two sensors become strongly coprime, when the ratio of the radii of the sensors
assumes any integer value from 2 to 200. Henceforth, the blur PSFs, he (Xl, X2) and
h,(Xl, X2), responsible for the two luminance images, will be assumed to satisfy
the strongly coprime condition.
In some special cases, the compactly supported solution set for exact deconvolu-
tion has been found explicitly. However, as described in [262], the non-compactly
supported solution (inverse filter) set together with a wideband lowpass filter, called
e
moUifier , is more desirable especially in noisy environments. In lR2 , let (WI, W2)
denote that wavenumber response of a mollifier O(Xl' X2) and let Ww;(Wl, W2) be
the Fourier transform, assuming it exists, of the noise in the ith observed image.
Then, fori = 1,2, ... ,m, the Fourier transform, Di(Wl, W2), of the non-compactly
supported solution cascaded with a mollifier 8(WI, W2) is given by [262]
(6.15)
whereHi(WI, W2) denotes the complex conjugate of Hi (WI, W2). Actually, Di(Wl, W2)
describes the inverse filter cascaded with the mollifier 8(WI, W2), which is used to
cut-off the high-wavenumber noise.
3.1 RECURSNE LEAST SQUARES (RLS) ALGORITHM FOR

SUPERRESOLUTION
Using sequential estimation theory in the wavenumber domain, an efficient

method was developed [263] for interpolating and recursively updating to pro-
vide filtering provided the displacements of the frames with respect to a reference
frame were either known or estimated. It was observed that the performance dete-
riorated when the blur produced zeros in the wavenumber domain and theoretical
justification for this can be provided.
Let f(x, y) be a continuous spatial image. Consider a set of p spatially shifted
versions
fx(x, y) = f(x + Oxk, y + Oyk), k = 1,2, ... ,p,
where Oxk and Oyk are arbitrary but either known or estimated shifts of f(x, y)
along the x and y axes, respectively. Consider the digitized image frame of size
(M x N), k = 1,2, ... ,p,
with sampling periods Tx, Ty along the x and y axes, and let
L L fk[i,l]exp (-21rj (r; + ~»

M-IN-l . 1
Fk[m,n] =
i=O [=0
denote the discrete Fourier transform (DFT), where j = A. The original

continuous object is assumed to be bandlimited so that its 2-D continuous Fourier
transform (CFT), FC( u, v) satisfies
The positive non-unity integers Lx and Ly are

so that the sampling frequencies Wx = ~: and Wy = ~: are less than the Nyquist
sampling frequencies,
The CFT of fk(X, y) is

Fk(u, v) = exp(j(6xkU + 6yk v))FC(u, v)
The DFT Fk[m, n] of the kth undersampled frame can be expressed as
1 L",-1 Ly-1 2 2
Fk[m, n] = T T. L L Fk ( ; : ; + iwx, ; ; + lwy)
x Y i=-L", l=-Ly x Y
for k = 1,2, ... ,p. Write Fk[m, n] in matrix form with lexicographical ordering
(-Lx, -L y),(-Lx +1, -Ly),(-Lx +2, -L y), ... ,(L x -2, Ly -l),(L x -1, L y-1).
Define the map
21rm . 21rn
(MTx + 'tWx, NTy + lw y) -+ r
wherei = (r-1)mod(2Lx) -Lx and I = L(~~~) J -Ly with r = 1,2, ... , 4L x L y.
Then, the preceding equation can be expressed in matrix form for all the frames.
. min I
4>[k,r] = ex p (21rJ [6xk (MT + T-) + 6yk(NT. + T.y )])
x x y
At frequency point [m, n] in the presence of noise, the DFT of the jth observation
frame is of size M x N,
Zj[m,n] = Fj[m,n] + Nj[m,n], m = 0, 1, ... , M -1, n = 0, 1, ... ,N-1
j = 1, ... , k, where p > k > 4LxLy are the low resolution frames. Set Zk =
[Z1 Z2 ... Zk]t. Here, q, is the k x 4LxLy matrix [Y 1 , ... , Yk]t, where
1
Y i = TxTy Xi
and XL is the kth row of the matrix [</>[k, r]] which can be expressed as the product
of a diagonal matrix D =diag(d1 , ... , dp ) and a matrix H = [h[k, r]], where
. m Lx) + 6 (n Ly)))
dk = ex p(J21r(6xk (MT - T Yk NT. - To ' k = 1,2, ... ,p
x x y y
and
j27r6Yk
Wyk = exp{ T. )
y
a = (r - 1)mod{2Lx),
The RLS algorithm is described as follows. Set
R[k] = <jJH<jJ
(where superscript H denotes the complex conjugate transpose) and
r[k] = <jJHZk
Let P[k] = R[k]-I and Q[k + 1] = P[k]Yk+1 (where Y denotes the complex
conjugation), then the basic RLS algorithm is given by :
F(k+I) = P[k + l]r[k + 1]

P[k + 1] = P[k] - Q[k + l]Q*[k + l]fa[k + 1]
where
r[k + 1] = r[k] + Y k+IZk+1
H - - -
r[k + 1] = [YIY2"'Yk+1][ZIZ2",Zk+1] = r[k] + Yk+IZk+1
t - t
= <jJk+1Zk+1
P(k) = R-I[k]r[k], R-1[k] = P[k]
By matrix inversion formula,
(A + BCD)-l =A-I_A-IB(DA-IB+C-I)-IDA-I
Setting A = R[k],B = Yk+1, C = 1,D = Yk+1

P(k+l) = R-1[k + l]r[k + 1] = [R[k] + Yk+1 Yk+1tl[r[k] + Y k+ I Z k+1]
= [R-I[k] _ R- 1[k]Yk+1 Yk+1 R - 1 [k]][r[k] +Y Z ]

1 + Yk+1 P[k]Y k+l k+1 k+1
P[k]Y yt P(k)
= P(k) + P[k + l]Y Z _ k+1 k+1 (6.16)
k+l k+l 1 + Yk+1 P[k]Y k+1
Define the gain vector
(6.17)
Therefore,
- t -
K[k + 1] = P[k]Yk+l - K[k + I]Yk+1P[k]Yk+l
= [P[k] - K[k + I]Yk+1P[k]]Yk+1
= P[k + I]Yk+l'
(Note from matrix inversion formula: P[k + 1] = P[k] - K[k + I]Yk+lP[k].)
Therefore, from (6.16) and (6.17),
~(k+l) ~(k) - t ~(k)
F =F +P[k+l]Yk+l(Zk+1-Yk+1F )
(Zk+l - Yt+1 P(k)) may be viewed as the a-priori estimation error vector and
K[k] is like the Kalman gain vector.
Alternately, with the answer known, from the derivation above, the expression for
the gain vector can also be derived as follows (this possibility was suggested by
Professor Hans Dieter Fischer following a lecture delivered at Ruhr University at
Bochum, Germany during the summer of 200 1, as part of a course on multidimen-
sional signal processing offered by the author) :
Start with the matrix identity
R[ktl = R[k + ItlR[k + I]R[ktl

= R[k + Itl[R[k] + Yk+1Yt+1]R[ktl
= R[k + Itl + R[k + ItlYk+1Yk+lR[ktl
= R[k + Itl + K[k + I]Yk+l R[ktl
where,
K[k + 1] = R[k + I t 1Yk+1'
-
Therefore,
R[k + Itl = R[krl - K[k + I]Yk+1R[krl,
K[k + 1] = R[k + lr l Y k +1 = [R[ktl - K[k + I]Yk+lR[ktl]Yk+l'
1-
K[k + 1] = ~[kt Yk~l .
1 + Yk+1 R[k]- Y k+1
The problem of reconstruction in the wavenumber domain with errors present both
in the observation and data was tackled by the method of Total Least Squares (TLS)
[264].
3.2 BACKPROJECTION ALGORITHM FOR SUPERRESOLUTION
According to [265] and [266], the image acquisition process is modelled by
gk(m, n) = ak (h(Tk (f(x, y))) + 'f/k(X, y)), k = 1,2, ... , K, (6.18)
where
• gk is the kth observed degraded (low resolution, blurred, geometric trans-

formed, and noisy) image.
• K is the number of observed images.

• / is a original high resolution image.
• Tk is the 2-D geometric transformation in the kth frame.
• h is a blurring operator.
• 'f/k is an additive noise in the kth frame.
• ak is the downsampling operator.
An attempt is made to construct a higher resolution image j, which approximates

/ as accurately as possible. The proposed high resolution reconstruction technique
is based on iterative backprojection (mp) algorithm adopted from computer aided
tomography (CAT). Starting with an initial guess 1(0) for the high resolution image,
the imaging process is designed to obtain a set of simulated low resolution images
{g1°) }f"=1 corresponding to the set of actual observed images {9k}f"=1' Ideally, if
1(0) is the correct high resolution image, then the set of simulated images {g10) }~1
should be identical to the actual observed images {9k}f"=1' The difference images
{gk - g10) }f"=1 are used to improve the initial guess by "backprojecting" each value
in the difference images onto its receptive field in 1(0), yielding an improved high
resolution image /(1). This process is repeated to minimize the error function
The imaging process of gk at the nth iteration is simulated by
gin) = (Tk(f(n)) * h) .j.. s, (6.19)
where.j.. s denotes a downsampling operator by a factor of s, and * is the 2-D

convolution operator. The iterative update scheme for the high resolution image
is expressed by
K
l(n+1) = I(n) + ~2:T;1 (((9k - gin)) t s) * p) , (6.20)
k=l
where t s is an upsampling operator by a factor of s and p is a "backprojection"

kernel according to Theorem 6.5 ..
THEOREM 6.5 The iterations of (6.20) converge to the desired deblurred image
I, if the following condition holds':
1
(6.21)
118 - h * pll2 < 1- "K
K
lifT! II '
LJk=1 .Lk 2
where 8 denotes the unity pulse junction centered at (0,0).
REMARK 6.1 lfthe 2-D geometric tran:,formation Tk, k = 1,2, ... , K consist
only of 2-D translations and rotations, then (6.2l)reduces to 118 - h * pII2 < 1.
THEOREM 6.6 Given condition (6.21), the algorithm converges at an exponential
rate, i.e. the norm ofthe error converges to zerofasterthan qn for some 0 < q < 1),
regardless of the choice of initial guess 1(0).
The detailed proofs of above theorems and remarks are given in [265] and [266].
4. Multisensor Array-Based Superresolution

DefineT(v) to be the (mxm) symmetric Toeplitzmatrix with v = [vo, ... , Vm_l]t
as the first column. LetH(v, w) be the (m x m) Hankel matrix (always symmet-
ric) with v as the first column and w as the last column. Let the shift of vector v
be denoted by u(v) = [VI, V2, ... , Vm-l, of
and let C be the space of matrices
that can be diagonalized by the discrete cosine transform (DCT) matrix C whose
entry in the jth row and ith column is
[C] J~.. -- J2 - 8jl cos ((j - l)(2i

2m
- 1)11")
' 1 ~ i,j ~ m
m
Then the set of symmetric Toeplitz-plus-Hankel matrices, defined next, comprise
the class C, where J is the m x m anti-identity (exchange) matrix.
C = {T(v) + H(u(v), Ju(v)) Iv = [vo, ... , vm_d t E ]Rm}

FACT 6.1 A matrix H of order m
H = T(v) + H((1(v), J(1(V)) , v E r

is diagonalizable by Cn #:. C
i.e. H = CtA(Ct)-l,Het = etA
where A is a diagonal matrix.
Example 6.1 Let v = ~ [1 ~ 0 '" 0] t E r. Then the tridiagonal symmetric
Toeplitz-plus-Hankel matrix
H = T(v) + H((1(v) , J(1(v))
is diagonalizable by Cm #:. C and the eigenvalues of H can be computed as follows.
From the eigenequation,
vtct = etA
where et is the first row of C t , it follows from the jth column of C t and A that
Aj = cos2 [U 2m
- 1)11"] , J. = 1,2, ... ,m
Example 6.2 Consider the symmetric Toeplitz-plus-Hankel matrix of order m,
Sk = T(ek+1) + H(ek' J(ek)), 1:::;; k :::;; m - 1

where ek is the kth unit vector with 1 in its kth position and zeros elsewhere.
Again,
SkCt = CtA
where the element in the (i, j)-th position of C t is
J2 - 6 j1
cos(
U- 1)(2i - 1)11"
2 ), 1:::;; i,j:::;; m
m m
Multiplying the first row of Sk by the lh column of ct,
cos [U- 1)(2k

2m
- 1)11"]
+ cos
[U - 1)(2k
2m
+ 1)11"] [U - 1)11"] \
= cos 2m /\j
Simplifying,
Aj(Sk) = 2 cos [u -2m1)k1l"] '

Let Hi be the Toeplitz-plus-Hankel symmetric blur matrix when the width of
the band of the Toeplitz component is (L + 1). Then,
HX _
L -
{f L;~i S2j-l H~,
£ "C L - 2 )/4 S . H X
L = 4k
otherwise
L~j=O ~ 2'
Then, if L = 4k, then for i = 1,2, ... ,m,
"(HX) _
I\z L -
±
L cos
2 [(i -2 1)11"] ~ [(i - 1)(2j - 1)11"]
~ cos
m . 1
J=
m
and in the complementary case,

. (L-2)/4 . .
"(HX) _
I\z L
i 2 [(Z -2 1)11"]"
- L cOS ~ cos
[2(Z - 1b1l"]
m .
J=
0 m
4.1 MULTISENSOR mGH RESOLUTION IMAGE RECONSTRUCTION MODEL
A specific high-resolution image reconstruction problem, based on a model for

a prefabricated multi sensor image acquisition system, proposed and studied by
Bose and Boo [267] is considered.[267]. Consider a sensor array with Ll x L2
sensors in which each sensor has Nl x N2 sensing elements (pixels) and the size
of each sensing element is T1 x T2. The goal is to reconstruct an M1 x M2 image
of improved resolution from several captured images of size Nl x N 2, where
Ml = LINI and M2 = L2N2. The pixel size of high resolution image, therefore,
isR- f;:.
x To maintain the aspect ratio of the reconstructed image, the case where
L1 = L2 = L is considered. For simplicity, L is assumed to be an even positive
integer in the following discussion. To generate enough information to resolve the
high-resolution image, subpixel displacements between sensors are necessary. In
the ideal case, the sensors are shifted from each other by a horizontal and a vertical
displacement proportional to T1 / Lx T2/ L. However, in practice there can be small
perturbations around these ideal sub pixel locations due to imperfections of the
mechanical imaging system during fabrication. Thus, for it, 12 = 0, 1, . .. ,L - 1
with (1t,12) f (0,0), the horizontal and vertical displacements d hl2 and dhl2'
respectively, of the [1t, 12]-th sensor with respect to the [0, O]-th reference sensor
are given by
dftl2 =
Tl( X)
L it + lLJl2 and
d!!hl2 -_T2( -y)
L 12 + €h12 '
where €hl2 and lY1h denote, respectively, the actual normalized horizontal and
vertical displacement errors. The estimates, €hl2 and €K 2' of these parameters,
'
Ehl2 and lK 2' can be obtained by manufacturers during camera calibration.
'
It is reasonable to assume that
I€hh I < 21
_X
an
d I-€hh
Y I 1
< 2'
because if that is not the case, then the low-resolution images acquired from two
different sensors may have more than the desirable overlapping information for
reconstructing satisfactorily the high-resolution image [267].
Let f(XI, X2) denote the original bandlimited high-resolution scene, as a func-
tion of the continuous spatial variables, Xl, X2. Then the observed low-resolution
digital image Yhl2 acquired from the (It, l2)-th sensor, characterized by a point-
spread function, is modeled by
T2 (n2 +! )+dj'112 iTI (nl +! )+di112

YhI2[nl, n2] = i f(Xl, x2)dxldx2+nlth[nl, n2],
T2(n2-! )+dJ'1 12 TI (nl -! )+di112
(6.22)
for nl = 1, ... , Nl and n2 = 1, ... , N 2. Here nltl2[nl, n2] is the noise corre-
sponding to the (hl2)-th sensor. These low-resolution images are combined to
yield the observed Ml x M2 high-resolution image Y by assigning its pixel values
according to
(6.23)
for 1t,12 = 0,1,··· , (L - 1), nl = 1, ... , Nl and n2 = 1, ... , N2, and

Ylth[nl, n2] is the shifted downsampled version of the observed high resolution
image.
The continuous image model f(xl, X2) in (6.22) can be discretized by the rect-
angular rule and approximated by a discrete image model. Let g and f be, respec-
tively, the vectors formed from discretization of g(XI' X2) and f(xl, X2) using a
column ordering. The Neumann boundary condition [268] is applied to the im-
ages. This assumes that the scene immediately outside is a reflection of the original
scene at the boundary, i.e.,
k = 1- i, i < 1,
{ k = 2Ml + 1- i, i > MI,
f [i, j] = f [k, l] where
1 = 1- j, j < 1,
1 = 2M2 + 1- j, j > M 2•
Under the Neumann boundary condition, the blur matrices are banded matrices
with bandwidth L + 1, but the entries at the upper left part and the lower right part
of the matrices are changed. The resulting matrices, denoted by Hh12 (f ,12) and h
(G
Hb.12 1 ,12)' have a Toeplitz-plus-Hankel structure. Using the MATLAB notation
Toeplitz(c,r) to denote the Toeplitz matrix having c as its first column and r as its
firstrow,H~12(fil,12)' Z E {x,y} can be written as
where Hhh (€h,12) is of size Ml x M l , Hhh (lhh) is of size M2 x M2 and
T(llll2) = ~Toep([ 1, ... ,1,! + llll2'0, ... ,O]t, [1, ... ,1,! - li1 l 2'0, ... ,On,
, L --.....--.. 2 ' --.....--.. 2 '
L /2 ones L /2 ones
-z) =
11 (flth L1 Toepl''ttz ([0, ... ,0]t , [0, ... ,0, 21 + fl-z h'
1 ~ ]t ,
L/2-l ones
T r (li1 h) = ±Toeplitz([O, ... ,0, ~- li1 h' ~ ]t, [0, ... ,O]t),
L/2-l ones
and J denotes the anti-identity (exchange) matrix. The blur matrix correspond-
ing to the (lr, h)-th sensor under the Neumann boundary condition is the square
Kronecker product matrix of order MlM2 given by
(6.25)
where the 2 x 1 vector lh , l2 is denoted by (f-El1,2 l )t. The blur matrix for the
l lYl1,2
sensor array is made up of blur matrices from each sensor:
L-l L-l
HL(l) = LL DI 112H1tl2(lI1h), (6.26)

11=0 12=0
where the 2L2 x 1 vector l is defined as
- _ [-X -Y -x -Y -x -Y -x -Y ]t
10 - 10 1000 00 10 01 1001 •• , fL-lL-2 fL-lL-2 fL-lL-l fL-IL-l .
Here Dh12 are diagonal matrices with diagonal elements equal to 1 if the corre-
sponding component of g comes from the (lr, l2)-th sensor and zero otherwise,
(see [267] for more details). Using H = HL(l) we have the linear system
Hf=g (6,27)
The resulting matrix H is either block Toeplitz with Toeplitz blocks (BTTB) (for
zero boundary conditions) or block Toeplitz plus Hankel with Toeplitz plus Hankel
blocks (BTHTHB) (for Neumann boundary conditions). For these matrices the
operation H v or Htu may be done in 0 (L 2 log L) flops. The linear system (6.27)
is very ill-conditioned since the underlying problem is ill-posed. Thus its exact
solution will be dominated by high-wavenumber noise amplification rendering it
meaningless. Some form of regularized least squares solution will be necessary.
4.2 REGULARIZATION METHODS
The most well-known regularization methods to solve the ill-posed problem is

the Tikhonov regularization technique. The regularized solution for f is the one
that minimizes the following expression
min IIHf - gll~ + allLfll~ (6.28)

f
where L is the regularization operator and a is the regularization parameter. The

common regularization operators are the identity matrix and the discrete Laplacian
operator. It is not difficult to show that the solution ofEq. (6.28) becomes
fa = [HTH + aLTLrl HT g
This is obtained by equating to zero the gradient vector (with respect to f) '\7<L> =
-2Htg + 2HtHf + 2aLtLf of the cost function from Eq. (6.28) expanded as
<L> = (g - Hf)t(g - Hf) + a(Lf)t(Lf).
The regularized solution by Tikhonov regularization may suffer from undesir-
able artifacts such as ringing near edges and insufficient noise smoothing in flat
regions. Many types of space-variant regularization methods have been proposed
to reduce these artifacts [269], [270]. The image restoration problem is modeled
as
g=Hf+n (6.29)
where, after scanning, f, g and n are the original image, the observed degraded
image and additive white Gaussian noise (with variance 0";) vectors, respectively.
The space-variant regularized solution for f in Eq. (6.29) is given by [269]
(6.30)
where a is the regularization parameter, L is the regularization operator and S
is the diagonal matrix which weights the relative degree to which the restored
image will be regularized at each pixel. Assume that the original image vector f is
generated by a noise source whose variance is not constant over the entire image.
Then, f can be described by
f=KWv (6.31)
where K is the generating matrix, W is a diagonal weight matrix which scales the
variance over the image and v is a zero-mean white Gaussian noise process vector
with variance O"~. The Wiener filter solution ofEq. (6.29) can be written as [269]
(6.32)
2
If we select a = ~,
(Tv
L = K- 1 and S = (WWt)-l, the space-variant regular-
ized solution in Eq. (6.30) is identical to the Wiener filter solution in Eq. (6.32).
These choices require a great deal of a priori information, most of which is not
available. Therefore, L is often approximated by the discrete Laplacian operator
and a is estimated by the GCV[271] (generalized cross validation) or the L-curve
method[272]. From Eq. (6.31), S = (WWt)-l is proportional to the reciprocal
Recent Impacts a/Multidimensional Systems Research 213
of the local variance ofLf. Assuming that the local variance ofLf, u2 [i,j], cen-
tered at pixel location [i, j], is approximately constant over a small region of size
(2P + 1) x (2Q + 1), it can be estimated as
i+P i+Q
1 "" "" 2 [k I] (6.33)
(2P + 1)(2Q + 1) ~
2 [ . .]
U Z,J = ~ x ,
k=~-PI=J-Q
where x[k, I] is a 2-D representation of the decorrelated image Lf. The entries of
weight matrix S are then given by s[i,j] = 1ju2 [i,j].
The regularization parameter selection methods are divided into two categories:
(l)Method 1: which needs a priori information about noise variance. (2)Method
2: which does not require knowledge of the noise variance. The GCV (Generalized
Cross Validation) and L-curve belong to Method 2. The GCV has proven to be
effective for estimating the regularization parameter. However, the minimum of
the GCV function is often very flat and therefore, difficult to locate numerically
[273]. Also, it fails to compute the correct regularization parameter when the errors
are highly correlated [274]. The L-curve is the parametric plot of the norm of the
regularized solutions versus the norm of the corresponding residual for all valid
regularization parameters. The L-curve method is more robust than GCV method
because it combines information about the residual norm and information about
regularized solutions while GCV method uses only information about the residual
norm. The effectiveness of L-curve method for regularizing the blurred and noisy
image was shown in [275]. It is reported that the L-curve method can fail to
retrieve proper regularization parameter when the noise to signal ratio goes to zero
[276] or the solution is "rough" [277]. Despite its usefulness, the L-curve method
requires massive computations for the calculation of the points on the plot. When
the blur matrix has a structure like BCCB and symmetric BTHTHB (which cannot
be satisfied in many applications) one can calculate the points in transform domain.
In the reconstruction of high resolution image with multi sensors, the blur matrix
with the Neumann boundary condition becomes non-symmetric BTHTHB, which
cannot be diagonalized by a DCT matrix. In [278], L-curve points are calculated
by use of PCGA. For different regularization parameter a, the iterative procedure
for calculation of L-curve points is repeated. To reduce computational complexity,
the so-called "Hybrid Methods" have been introduced [273]. Here, the original
least squares problem is projected onto a smaller dimensional subspaces and then
the regularization technique is applied to the smaller dimensional problem.
s. Wavelets for Superresolution

Wavelets 'l/Jj,k are traditionally defined as. the dyadic translates and dilates of the
mother wavelet"p. That is "pj,k(X) = "p(2J x - k). Such wavelets are refered as
HP
Input
LP
(a) Filter bank
)------r-~ HP
Input
LP
(b) Lifting scheme
Figure 6.1. Comparison between implementation of the fast wavelet transfonns. In (a), filterbank:
is used for the first generation wavelet while the lifting scheme is for the second generation wavelet
as shown in (b).
first generation wavelets. However, the application of these kinds of wavelets is

limited by many situations [279] [280] such as those listed next.
1 Bounded domain: When working with finite data, it is natural to deal with
data on the boundary. Therefore, it is desirable to have wavelets adapted to the
data on the boundary directly without assuming boundary condition (such as
periodic, zero, or Neumann boundary condition) for the data.
2 Irregular sampling: In many practical applications including superresolution,

the samples are not necessarily on a regular grid. However, the first generation
wavelets cannot be adapted to irregular sampling geometry.
In order to tackle aforementioned problems, the second generation wavelets

were introduced. The basic idea of second generation wavelets is to give up
translation and dilation in order to construct wavelets adapted to general settings
(bounded domains and irregular samples) and simultaneously to maintain many
powerful properties of first generation wavelets such as time-frequency localiza-
tion, orthogonality or biorthogonality, and fast transform.
In the implementation aspect, the first generation wavelet transform is performed
by using filter bank while the second generation wavelet transform is done by using
the lifting scheme as shown in Fig. 6.1(a) and (b), respectively. The input to the
filter bank is passed through the highpass and lowpass filters, G(z) and H(z) in
the upper bank and lower bank, respectively, prior to downsampling. The resulting
outputs are the highpass (HP) and the lowpass (LP) components as shown in Fig.
6.1(a). The filters G{z) and H{z} are designed in a way that permits perfect
reconstruction of the input.
On the other hand, in the lifting scheme, the input is downsampled and split
into odd-indexed and even-indexed samples in, respectively, the upper bank and
lower bank. Then the remaining process is divided into two main steps: prediction
(P) and update (U). The prediction step estimates the odd-indexed samples from
the even-indexed samples. Then the difference between the actual odd-indexed
samples and predicted odd-indexed samples is used to update the original even-
indexed samples. The resulting outputs in the upper and lower banks are the HP
and LP components, respectively, as depicted in Fig. 6.1(b).
There are two approaches that could be used in the prediction step: interpo-
lating subdivision [281] and averaging subdivision [282]. The order NEZ of
subdivision can be chosen by selecting the polynomial of degree N - 1 in the
prediction step. According to [279], it can be shown that the order N of su~divi
sion corresponds to the number of vanishing momen~ of the dual wavelet 'ljJ. On
the other hand, the number of vanishing moments N E 7l, of primal wavelet 'ljJ
is chosen in the update step. The same result could also be obtained in the filter
bank by designing the filter G{z} and H{z) to have the same vanishing moments
in primal and dual wavelets. It is seen that the lifting scheme leads also to new
insights into the construction of first generation wavelets.
According to [283], the advantages of the lifting scheme over the filter bank
implementation are listed next.
1 Faster implementation: Lifting scheme leads to a faster implementation when

compared to the filter bank implementation.
2 In-place implementation: Lifting scheme allows for an in-place implemen-

tation of the fast wavelet transform. That is, the wavelet transform can be
calculated without allocating auxiliary memory.
3 Parallel implementation: The parallel implementation is possible in the lifting

scheme.
The superresolution system produces high resolution (HR) image from a set
of low resolution (LR) frames. The relative motions in successive frames are
estimated and used for aligning the sample points in each frame into a HR grid.
There are various types of models [284][285] used to represent camera motion:
translation, rigid, affine, bilinear, and projective. The most general model is the
projective model which has eight motion parameters. Following the registering of
the LR frames in a HR grid, the available samples distribute irregularly as shown
in Fig. 6.2. It is noticed that the number of registered LR samples is finite and
confined to some bounded domain. Also, they distribute irregularly on the HR
:- ~.~~
n....,.. ""
.......... _- ............ ifiijiu..r".mpl.··DlLR .. ··.. ···· · -.......... -.-.-~.-~ .......... '.-. i:j~. ~;~~
~~~~~a ---.--.. --.--.~-. .::. :::.::. ::··.::·.:::·.:::::"li;<~jC:~:·.::··~:::::·.:·:':::·.::.,:::. ::.t?:::":::.;.~

I
i~ ~\W
........n f~L......y'
a,..-...r;;;l~;
L.......v'~..."mm ~
.. .. .. .. ........... -....... ...... .. ....-.. ....-..... .. ..-.. ...-..-.. ..-..-..--...
'.~~ ~ ~ ~-~ ---~ ~ "- ~ ,, ~ ~ ~ ~-
Figure 6.2. The superresoiution system.
grid. Due to these features (bounded domain and irregular sampling), the second
generation wavelet seems to be the right tool for superresolution.
6. Other Recent Applications

Here attention is directed to applications, not described or cited before in the
text, of multidimensional systems. Fomasini [286] considered the problem of river
pollution modeling and the associated selfpurification phenomenon by using 2-D
state-space models [5, Chapter 4, Section 4.5]. He discussed several models to
accomodate physical phenomena of increasing complexity.
Over the years, various concepts from multidimensional systems theory have
been of use in the design of multidimensional wave digital filters [287], [288]. The
possibility of integrating linear as well as nonlinear partial differential equations
(PDEs) (which are known to characterize a wide range of problems in diverse
branches of science and engineering including electromagnetics, elasticity and
strength of materials, fluid mechanics, thermodynamics, plasma physics, etc), by
multidimensional wave digital (MDWD) principles, has been demonstrated dur-
ing the past decade [289] as well as in more recent publications [290], [291],
[292]. Fries [290] showed how to transform a multidimensional Kirchhoff net-
work (MDKN) to a new coordinate system without altering the structure of the
system, so that MDKNs in curvilinear coordinates can be described. Consider
two coordinate systems in the vector variables x = [Xl X2 • • • Xn] t and
y = [YI Y2 . . . Yn] t , each having real-valued components. The two coor-
dinate systems are related by a diffeomorphic orientation preserving transforma-
tion,
y = f(x),
where the Jacobian J of the transformation vector f has a globally positive deter-
minant. The vector of differential operators,
Dx - -[...L OXl
...L
OX2
... -.L]
OXn '
~n ]
are transformed according to
Fries' research was based on a result from tensor calculus (v being an arbitrary
vector),
(6.34)
whose validity was substantiated in [292], by proving, inductively, the identity,

referred to as Piola's identity (Harley Flanders offered a deductive proof based on
Cartan's exterior calculus), stated below.
FACT 6.2 Let Dn = [Dl D2 . . . Dn] t be the vector differential operator

and let I n , In(adj) denote, re!Jpectively, the Jacobian matrix of order n and its
adjoint matrix. Then,
(6.35)
implying that the multidimensional differential operator D~ annihilates the adjoint
(adjugate) matrix of the Jacobian.
The design of multidimensional wave digital algorithm rests on the derivation of
a multidimensional Kirchhoff network (MDKN), in the continuous space and time
domains, endowed with the property of passivity and., in particular, losslessness
and dissipativity. Subsequently, MDWD principles lead to discrete dynamic sys-
tems that implement discrete conservation laws akin to the conservation of energy
properties in continuous systems. While linear shift-invariant passive or lossless
multidimensional circuits and systems have been heavily researched and docu-
mented [5, Chapter 5], the incorporation of nonlinear and time-varying elements
in the Kirchhoff network creates challenging problems that invite development of
new methods and techniques.
The simplest building blocks for the passive multidimensional Kirchhoff net-
works are resistors, inductors, ideal transformers and ideal gyrators. While the
first type of element is passive but dissipative, the last two are passive as well as
lossless. The inductances, possibly nonlinear and time-varying, have to be defined
in a particular way in order to ensure that the crucial property of multidimensional
passivity is intact. This general notion of multidimensional passivity is important
for achieving stability and computational robustness of the wave digital algorithm
designed to integrate the PDEs. For the case of a coupled multiport of inductors
L, the multiport characterizations,
u AtD(Ai), L = AtA
u ~(LDi + D(Li))
both account for passivity. However, the first representation of the multiport is more
general than· the second representation. Both representations are equivalent if and
only if a factorization L = A&Ao exists, where Ao = U A, UtU = UU t = I,
and A is the Cholesky triangle such that (A&DAo) is symmetric. The solution Ao,
which has been shown to exist, makes the defect matrix
vanish as A --t Ao. It was shown that skew-symmetry is the underlying property
which characterizes the equivalent representation problem and is crucial to the
existence and uniqueness of a time-varying orthogonal matrix U as a solution to
the matrix differential equation (involving a a time-variant skew-symmetric matrix
M in linear form) satisfied by U [293]. An explicit solution for U was given
subject to the validity of commutativity in multiplication of M and its derivative
matrix. Two proofs were advanced to show that the commutativity condition of
two skew-symmetric matrix holds if and only if their product is symmetric. It
was subsequently shown that a matrix differential equation may also be setup that
involves a skew-symmetric matrix Z in quadratic form. An isomorphism was set up
between skew-symmetric and orthogonal matrices through a generalized Cayley
transformation so that U can then be constructed, if one wants, after solving
for Z[294]. Further research is required for handling the nonlinear time-varying
case. In the case of nonlinear current controlled coupled inductor matrix of order
2, the expression for the defect required to construct the Cholesky factor was
deduced [294] to expose the additional difficulty introduced by the incorporation
of nonlinearity.
7. Conclusions
Following the first paper,published in 1988, concerned with the invariance of a
type of property called the scattering Hurwitz property, unique to multidimensional
systems, a handful of articles dealing with zero distributions of sets of bivariate and
multivariate polynomials were published in several journals before a Special Issue
on Robustness of Multidimensional Systems appeared in 1994 [295]. Inference
of properties about sets from subsets with regard to other properties of concern in
multidimensional control system design and signal processing include multivariate
positivity, positive realness, and absolute stability of delay-differential systems.
Some research has been reported on these topics during the last decade following
what was summarized by this author in 1993 (see Chapter 4 in the book entitled
"Multivariate Analysis: Future Directions," edited by C. R. Rao and published by
North-Holland Elsevier Science Publishers, Amsterdam, The Netherlands).
Multiple undersampled images of a scene are often obtained by using a charge-
coupled device (CCD) detector array of sensors which are shifted relative to each
other by subpixel displacements. This geometry of sensors, where each sensor
has a sub array of sensing elements of suitable size has recently been popular
in the task of attaining spatial resolution enhancement from the acquired low-
resolution degraded images that comprise the set of observations. The multisensor
array technology is particularly suited for use in a broad spectrum of applications
ranging from satellite imaging to microelectromechanical systems (MEMS) where
accuracy, reliability, and low transducer failure rates are essential. This also extends
to chronic implantable sensors, monitoring of semiconductor processes, mass-
flow sensors, optical cross-connect switches, pressure and temperature sensors.
The benefits include application of the methods for development to any sensor
array or cluster, reduced calibration and periodic maintenance costs (desirable
especially for space-based prefabricated multi sensor arrays), higher confidence in
sensor measurements based on statistical average on multiple sensors, extended
life of the array compared to a single-sensor system, improved fault tolerance,
lower failure rates, and low measurement drift. Further research is required on the
incorporation of robustness and efficiency, features that are crucial in multispectral
sensing and subsequent multiframe processing. The sparseness and structure of the
matrix in the image deblurring and superresolution problem needs to be exploited.
Rapid progress in computer and semiconductor technology is making it possible
to implement image sequence processing tasks reasonably quickly, but the need
for processing in real time requires attention to design of efficient and robust
algorithms for implementation on current and future generations of computational
archi tectures.
Analysis from the wavelet point of view of the construction of a high resolution
(HR) image from low resolution (LR) images acquired through a multisensor array
in the approach of Bose and Boo [267] was recently conducted in [296]. Absence
of displacement errors in the LR samples was assumed and this resulted in a
spatially invariant blurring operator. The algorithms developed decomposed the
function from the previous iteration into different wavenumber components in the
wavelet transform domain and, subsequently, added them into the new iterate to
improve the approximation. Extension of the approach when some of the LR
images are missing, possibly due to sensor failure was also implemented in [296].
The wavelet approach towards HR image formation was generalized to the case of
spatially varying blur associated with the presence of subpixel displacement errors
due to improper alignment of the sensors [297].
Chapter 7
MULTIVARIATE RATIONAL APPROXIMANTS

,
OF THE PADE-TYPE
BYN. K.BOSE
1. Introduction and Motivation
While the study ofPade approximants is extensively documented in the literature

[298], [299], its generalization to the multivariate case (henceforth to be referred
to as Pade-type approximants) is of relatively recent origin. Nevertheless, though
an early paper dealing with Pade-type approximants to a double power series
was published as late as 1973, the ensuing decade witnessed a flurry of activities
resulting in the documentation of those research not only in journals, but books
[299, Section 1.4], [300, Section 4.2], special issue [301], and proceedings of
conference exclusively devoted to the topic ofPade approximation. Due to the very
recent publication of the special issue on Rational Approximation for Systems,
identified in reference [301], this chapter will be more concise than what was
originally planned. The primary goal will be to alert the reader to the vast, albeit,
scattered literature on multivariate rational approximants of the Pade type, and to
acquaint him (her) with the possibility of applying these available resources to
problems originating in systems theory. It is hoped that the reader will be able,
without too much effort, to locate numerous other references, as needed, from
the sources listed at the end of this chapter because these sources contain the
necessary keys to those references, which, for practical limitations on space could
not be included here.
Pade approximation theory has found numerous applications and interpretations
in systems theory including the minimal partial realization problem [302], multi-
port network synthesis [303], time domain design of digital filters from a prescribed
impulse response sequence [304], [305], and model reduction of control systems
[306], [307]. The Pade-type of approximation theory also plays a significant role in
multidimensional systems theory. In [308] algorithms for representation of double
power series by branched continued fraction [309] has been considered with the
221
objective of approximating 2-D input-output characteristics of nonlinear systems,

described by a double power series. In [310], the problem of order reduction of
2-D systems using Pade-type approximation has bee considered. The approach is
based on the matching of a finite number of coefficient in the double power series
expansion of the transfer function of the system with the corresponding coefficients
associated with the lower-order model, so that a number of linear equations in an
equal number of unknowns are formed. No constraints, however, were placed on
the geometry of the lattice points associated with the matched coefficients which
increased the problem of computational complexity and, furthermore, an unstable
lower-order model could result even though the original system transfer function
was stable.
In the area of digital processing of 2-D signals, the problem of factoring spectral
density function occurs. It is well known that unlike in the I-D case the 2-D
spectral factorization of rational density functions into rational spectral factors
is not possible, in general [5, Chapter 4]. However, it is possible to extend the
concept of spectral factorization to the 2-D case provided non-rational infinite-
order factors are allowed. Since, a recursive filter must be of finite order for
implementation, it becomes necessary to obtain a rational approximation to the
infinite order factor. Pade-type approximation is possible in this case and, in order
to attend to the problem of possible instability, it may be advisable to truncate
as well as smooth the double power series characterizing the impulse response
sequence (already obtained as an infinite-order spectral factor) by 2-D windows,
as suggested in [311], prior to obtaining a desirable rational approximant. In the
rational modeling of 2-D systems, the system to be modeled may not only be
characterized by its impulse response sequence (first-order data) but also by its
autocorrelation sequence (second-order data). In [312], a generalization of the
Pade-type of approximation procedure has been considered to model a 2-D system
characterized by both first-and-second-order datas.
2. Multivariate Pade-Type Approximants (Scalar Case)

For the sake of brevity in notation and exposition, we restrict ourselves to bi-
variate rational approximants of the Pade-type. It is noted that the generalization
to the n-variate (n > 2) case is usually conceptually straightforward, even though
it may be computationally quite cumbersome.
DEFINITION 7.1 A rational junction, [P(Zll Z2)]/[Q(Zll Z2)] is a bivariate ap-

proximant ofthe Pade-type to a double power series T (Zl 1 Z2) = :E~o :E~o tijzi z4
provided
00 00
TQ -P = LLTijztz4 (7.1)
i=O ;=0
Multivariate Rational Approximants of the Pade-Type 223
with Tij = ofor (i,j) E E, where E denotes a suitable interpolation set. The
choice of E and the set of point where the coefficients of P(ZI' Z2) and Q(ZI, Z2)
are allowed to be nonzero determine the various types of approximants.
The Canterbury group's main effort may be summarized by the following

choices of polynomial coefficients support and interpolation set
nl n2
P(ZI' Z2) = 2: 2:Pij z fz4 (7.2)
i=O j=O
ml m2
2: 2: qijzf z4 (7.3)
i=O j=O
E ((i,j) I 0 ~ i ~ max(nl' md, 0 ~ j ~ min(n2' m2)}
U{(i,j) I 0 ~ i ~ min(nl,md, 0 ~ j ~ max(n2,m2)}
U{(i,j) I max(n2,m2) < j ~ n2 + m2, max(n2,m2) <
< i + j ~ n2 + m2, 0 ~ i ~ min(nl,md}
U{(i,j) I max(nl, md < i ~ nl + ml, max(nl, md <
< i + j ~ nl + ml, 0 ~ j ~ min(n2, m2)} (7.4)
with
Tij = 0 for (i,j) E E (7.5)
and (the symmetry conditions),
Tml +nl +1-£,i + Ti,m2+n2+1-i = 0,

i = 1,2, ... , min(ml, nI, m2, n2)' (7.6)
Figure 7.1 a shows the geometry of the lattice points forming the interpolation set
E, described above. Chisholm [313] initiated the research of the Canterbury group
by investigating approximants with nl = n2 = ml = m2 = min (7.2)-(7.6),
which were shown to reduce to diagonal Pacte approximants when one variable
is equated to zero (projection), besides possessing properties of symmetry, exis-
tence and uniqueness, invariance under a group of certain bilinear transformations
(homographic invariance), duality or reciprocal invariance, factorization and addi-
tivity. Chisholm and McEwan [314] provide the generalization of these results to
the n > 2 case. Certain generalization of the bivariate and n-variate approximants
in [313] and [314], respectively, were proposed [315], [316], as generalized ana-
logues of off-diagonal Pade approximants. For determinantal expressions for the
numerator and denominator polynomials of these approximants, see Levin [317].
Lutterodt [318], [319], considered approximants with P(ZI, Z2) and Q(Zl, Z2)
defined as in (7.2) and (7.3), respectively, but his interpolating sets E were chosen
E E
t----------, + (m)+I)(m2+1)
- 1 points
(a) Canterbury grid (b) Lutterodt grid 1
E
m
E
n
i I + 112(m+lXm+l)-lpoints
I
I I
I I
n
(c) Lutterodt grid 2 (d) Karlsson-Wallin grid
Figure 7.1. Geometry of Lattice Points of Interpolation Sets for Various Grids
as in Figures 7.1b and 7.1c. For Figure 7.1c [319],
E = {(i,j) I ° j ~ n2}
~ i ~ nl, 0 ~
U{(i,j) I ni +1 ~ i ~ ni + ml, n2 + 1 ~ j ~ n2 + m2}
U{(i,O)lnl+1~i~nl+ml} (7.7)
U{(O,j) I n2 + 1 ~ j ~ n2 + m2}
In both Figures 7.1b and 7.1 (c) the set E has (ni + 1)(n2 + 1) + (mi + 1)(m2 +
1) -1 elements which are required in the finding of (the same number of) unknown
coefficients {Pij}, {qij} with qoo ~ 1 by solving a set of linear equations derived
from (7.1) by equating coefficients of monomials zfz4, (i,j) E E, on both sides
of the equation. These approximants do not satisfy many of the properties of
Chisholm's diagonal approximants. A detailed comparison of the properties of
various types of approximants has been given by Cuyt [320]. Lutterodt considered
the problem of generalization his results to the n-variate case in [318].
Karlsson and Wallin considered Pade-type rational approximants [321] where
n
P{Zl, Z2) - L
i+j=O
Pijzi~ (7.8)
m
Q{Zl,Z2) L
i+j=O
qijzf~ (7.9)
E ::J {(i, j) Ii + j ~ n}. (7.10)
The geometry oflattice points comprising set E, in this case, is shown in Figure
7.1d. Note that the number of unknowns in (7.8) and (7.9) with qoo £ 1 is
~(n + l)(n + 2) + ~(m + l)(m + 2) - 1 which equals the number of elements
in set E.
The computational complexity associated with the computation of the Pade-
type of multivariate rational approximants is an important problem that deserves
attention. In [322] and in [315], a computational scheme, called the prong method,
has been used to obtain a solution to the Canterbury approximants and numerical
difficulties encountered in the construction of such approximants were discussed.
In many situations, it is advisable from the practical standpoint, that the approx-
imation problem be so formulated such that recursive schemes for computation
may be developed by exploiting certain structure properties of the characteriz-
ing matrix in the system of linear equations, which is set up for solution. The
recursive scheme may also be developed as in the univariate case by applying
the so called E-algorithm due to Wynn [300, pp. 159-177]. It is well-known
that if this E-algorithm is applied to the partial sums of the univariate power series,
T{z) = :E~o tizi, then E~~m) (this is standard notation for the E-algOrithm) is the
[11m] Pade-approximant to T{z) i.e. [11m] has polynomials of degrees 1 and m,
respectively, for its numerator and denominator. Hillion [323] used the E-algorithm
for obtaining rational approximants for multiple power series. Cuyt [324] applied
the E-algorithm for recursive computation of multivariate Padetype approximants
introduced by her [325] and proved that these approximants satisfy the property
analogous to the one, referred to above, which is valid for Pade approximants
(univariate case). In [326], it was shown how the I-D Pade approximants may
be applied several times to obtain a bivariate rational approximant with the power
series coefficient matching property (i.e. matched with a finite set of coefficients
of a specified power series) over a rectangular grid. The technique generalizes
naturally to the n-D, n > 2, case, sine the generalization calls for the repeated
application of the I-D Pade approximation technique to a univariate power series
expansion whose coefficients belong either to the real number field or the field
of rational functions in one or more independent complex variables. Though the
technique is conceptually simple and appears to be computationally expedient, the

properties of the rational approximants generated remain to be studied in depth.
3. Pade-Type Matrix Approximants

Parle approximants to an operator T(z), analytic in z and possessing a formal
Taylor series expansion around z = 0 with coefficients in a noncommutative alge-
bra, were considered by Bessis in 1973 [327], and the various invariance properties
of such approximants were also studied. The feasibility of recursive computation of
1-D matrix Parle approximants was demonstrated in [328]. Matrix Stieltjes series
was defined in [303], and the RC-ideal transformer network realizability property
for matrix Parle approximants of certain orders to such series was proved via the
artifices of continued-fraction expansion and matrix Cauchy index. A multivariate
scalar Stieltjes series can be defined. For the sake of brevity a 2-D Stieltjes series
is defined next.
DEFINITION 7.2
tij = 10 10 xi ~da (Xl, X2)
00 00
be finite, 'Vi, j, where a(xI, X2) on 0 ~ Xl < 00, 0 ~ X2 < 00 is bounded,

monotone nondecreasing in Xl for fixed X2, and monotone nondecreasing in X2
for fixed Xl. Then, the series
00 00
T(ZI,Z2) = LLtijzlz4
i=O j=O
is a double Stieltjes series with an integral representation,
Actually, the integral representation in Definition 7.2 may be considered as

one possible generalization of a Stieltjes function from the 1-D to the 2-D case.
Another possible representation for a bivariate Stieltjes function is,
whose power series expansion is [329],

where tij's are the moments defined in Definition 7.2. It will be useful to define
a bivariate matrix Stieltjes series with the objective of relating Pade-type approxi-
mants of specific orders to such series with the impedance matrices of a subclass
of passive multivariate multiports [5, Chapter 5]. In fact, this study may be ini-
tiated with the scalar bivariate Stieltjes series and one-ports with driving-point
impedances,
Tl T2
Z(Zl,Z2) = F+ - +-
Zl Z2
where F, T 1, T2 are real nonnegative functions in the complex variables Zl, Z2 and
Z(Zb Z2) is a real rational function in ZI, Z2. In the 1-D case, Pacte approximants
of certain orders to a Stieltjes series are impedance functions of the form
Tl
Z(zd =F+-
Zl
whose poles and zeros are simple and alternate on the negative real axis with
the pole begin closest to the origin. Studies relating matrix Stieltjes series to
multivariate multiport realizability theory will be useful in the derivation of new
properties ofPacte-type approixmants to matrix Stieltjes series in addition to those
which are known [329].
The bivariate Pade-type approximation to a general bivariate matrix power se-
ries has been considered in [330], where recursive schemes for computing the
approximants have been developed. Since these results have potential applications
in multivariable (mult-inputlmulti-output) 2-D system design, a sufficient condi-
tion for testing the stability of the approximants in terms of the specified impulse
response sequence has been obtained [331].
4. Conclusions
This chapter provides a concise exposition of the status of multivariate rational
approximation theory of the Pacte-type and its links to multidimensional systems
theory. It is, however, emphasized, that problems of interest in multidimensional
systems theory like the identification of linear systems, stabilization of spatio-
temporal and distributed recursive filters and the modeling of random fields by
autoregressive moving average processes might not only be based on either the
specified impulse-response or autocorrelation data but on both. When only the
first order (impulse response) data is provided (in, for instance, modeling from
a deterministic point of view) one might proceed with Pacte-type of apprximants
while when only the second order datas are provided (modeling from a stochastic
point of view by fitting the system's input-output covariance properties), the least
squares approximation theory is widely used. In the later case, for 2-D problems the
normal equations involve a positive definite block Toeplitz matrix, 2-D Levinson
type algorithm is available [332] for fast computation of the solution vector. There
is a significant body of current literature available on approximants based on the

least-squares criterion [5, Chapter 4], [333], but the objective here is not to describe
that aspect of research. Also, the use of a modified least-squares procedure to obtain
approximants using first and second order datas was cited [312].
Several problems remain to be investigated. First, the availability of a three re-
currence formula relating the denominator polynomials of bivariate approximants
of successive orders [330], suggests that these polynomials may belong to a se-
quence of bivariate polynomial which are orthogonal over real intervals [334]. This
possibility has to be checked for validity. Second, it is of mathematical interest
to develop a theory of bivariate matrix orthogonal polynomials over real intervals
analogous to what is recently available in the single variable case [303], [335].
The 'denominator' bivariate matrix polynomials in the matrix rational approxi-
mants might form such an orthogonal sequence. Third, as described in Section
2, a study of multivariate matrix Stieltjes series would be useful, especially in the
context of its possible relation to multivariate network realizability theory. Fourth,
in the context of 2-D digital filtering, it is useful to determine whether or not an
approximation scheme similar to that in [330] can be developed for the design of
half-plane filters [5, Chapter 3]. It is expected that a judicious blend of techniques
used in approximating a formal Laurent series by means of rational functions [336]
with the techniques used in [330] might help in reaching this goal.
5. Updates
Among a number of generalizations of Pade approximants, one of the best
known in system theory is 2-D Pad6 approximants.. A Pade approximation of
a function h(z) = h(Zl' Z2, ... ,zn) is a rational function a(z)lb(z) in reduced
form, b(O) =1= 0, so that specified initial segments of alb and h agree. This can be
considered as a special case of solving the congruence a == bh mod I, where I is an
ideal in R = K[z]. for the univariate and bivariate cases it has been shown in [337]
that a unique reduced solution (a, b) exists when certain conditions are satisfied
by the leading terms, It( a) and It(b) with respect to a particular term order.
For an early comparison of several multivariate Pade approximants, see [338].
Applications of 2-D Pade theory can occur in the synthesis of nonlinear trans-
fer characteristics [308], besides the design of analog and digital filters. Given
a bivariate power series, the rational approximant is computed in [339] by an
accuracy-through-order principle. The index set E for matching is required to
satify a set of three conditions, namely (a) the inclusion property (this requires that
for a point in E, the rectangular subset of points from the origin with the given
point as its furthermost comer also lies in E), (b )the index set N of the numerator
coefficients is a subset of E, and (c) and the cardinality 1 E \ N 1 of the difference
set E \ N is equal to 1D 1- 1, where D is the index set of the denominator coeffi-
cients of the rational approximant. Compuatability via the accuracy-through-order
principle requires the index set E to satisfy the inclusion property. The imposed
condition NeE permits the decoupling of the set of equations, as in I-D Pade
theory, for a homogeneous part defining the denominator coefficients from an in-
homogeneous part which yields, subsequently, the numerator coefficients by direct
substitution. The restiction imposed on I E \ N I guarantees the existence of a
nontrivial denominator. Determinantal expressions for the numerator and denom-
inator polynomials in the rational approximant are available like in the I-D case.
A nonbranched continued fraction representation can also be constructed for the
Pade approximant. For these and related results, the reader is referred to [339].
When discussing convergence results of univariate Pade approximants, one com-
pares a selected (based on information about the function being approximated)
sequence of approximants in the Pade table (each element of this table has a dou-
ble entry with the numerator and denominator degrees indexing, respectively,the
rows and columns), is compared with the function being approximated. Uniform
convergence, which requires the Chebyshev-norm of the error tending to zero, and
convergence in measure or capacity (meaning that convergence excludes an un-
known area of disruption that can be made arbitrarily small) have been extensively
studied. Many of these univariate results have a multivariate counterpart [340].
A list of more than 2500 references has been gathered over the past 20 years, to
help study and reliably compute generalizations of Pade approximants. His list is
available at http://win-www.uia.ac.be/u/cantlbib/refs9912.bib
Chapter 8
OLD PROBLEM SOLUTIONS AND

SOME NEW PROBLEMS
1. Progress Towards Solution of Open Problems since 1985

A list of open problems in multidimensional systems theory were compiled in
the original edition published in 1985. The progress made towards their solution
are summarized next.
1.1 NONESSENTIAL SINGULARITIES OF THE SECOND KIND (NSSK)
PROBLEM #1
Problem on BIBO Stability With Presence ofNonessential Singularities of the Sec-
ond Kind on Tn
Let H{z) = P{z)]j[Q{z)] be a n-variate rational function where P{z) and

Q{z) are relatively prime polynomials having real coefficients and Q{O) i= O.
Then, H{z) is expandable as a power series around z = O.
L ... L
00 00
H{z) = h[k]z~l ... z!n

kl=O kn=O
Clearly, k ~ (k1 , ... , k n ) and z ~ (Zl,'" ,zn).

Investigate the truth or falsity of the following conjectures.
Conjecture 1. Let there be a finite number of points 2 on Tn where P(z) and Q(z)
have common zeros. then, the series {h[k]} is absolutely summable if and only
if H(z) is continuously extendable from un to fr.
2 Also the case when this restriction is not satisfied remains to be investigated
231
Conjecture 2. A BillO stable rational H(z) with nonessential singularities of the

second kind on Tn cannot have a BillO stable inverse.
In a recent paper [lOO],it was proved that a rational multivariate first quadrant
quarter-plane digital filter transfer function, analytic on the open unit polydisc,
is BillO-stable if and only if it has an uniform extension to the distinguished
boundary of the polydisc. Though continuity on the distinguished boundary is not
equivalent to a function (not restricted to be rational) having an uniform extension
to the boundary (there are functions, defined and continuous on the distinguished
boundary, whose Fourier series expansions do not converge to the functions), in
tests for BillO stability, one needs only check for continuity on the torus (this also
settles Conjecture 1 in Open Problem 1 in Chapter 8).
In Conjecture 2 of Open Problem 1, the precise effects on BillO stability of
nonessential singularities of the second kind (NSSK), when present on the unit
polydisc distinguished boundary (Tn), and the existence of a BillO stable rational
function with a BillO stable inverse in the presence of such type of singularity were
conjectured. In [341], it was shown that a BillO stable bivariate rational function
in reduced form (and, therfore, having a finite number of common zeros) could
have a BillO stable inverse. It was shown that stable systems, in the 3-D case, can
admit uncountably infinite NSSKs in T3 and TxT x U, where T denotes the unit
circle and U denotes the interior of the unit disk [342]. A sufficient codition for
the inverse of a stable transfer function with uncountably infinite NSSKs in those
domains to be unstable was also given in the same paper.
1.2 RATIONAL SPECTRAL FACTORIZATION
PROBLEM #2
Rational Spectral Factorization
Consider the problem of spectral factorization of a multivariate rational spectral

density function, S (Z1' ... , Zn, zI 1 , ... , Z; 1 ), having real coefficients, and which
is strictly positive on ~.
S(Z1' ... ,Zn, zI 1 , ... ,z;1) = H(zl, ... ,Zn)H(ZI 1, ... ,z;1)
where it is desired (but, this may not be possible, in general) that the spectral factor,
H(Z1' ... ,zn) be a real, rational, minimum phase function. We define a minimum
phase function characterizing a 'causal' or positive cone filter to be one which is
BillO stable with a BillO stable inverse; also, the support of the coefficients of the
power series expansions about (0, 0, ... ,0) of the function and its inverse are con-
=
strained to belong to the positive n-D cone (i.e. the first quadrant, when n 2; the
support can be analogously defined for weakly causal and asymmetric half-plane
Old Problem Solutions and Some New Problems 233
filters). Since, it is well-known that a n-D (n > 1) spectral density function almost
never has a rational factorization, investigate the possibility of obtaining a set of
necessary and sufficient conditions which a rational S(Zl,"" Zn, zI l , ... , z;l)
must satisfy so that a desired rational minimum phase spectral factor is obtained.
Note that in [77], it has been shown that if the numerator and denominator of
S(Zl,Z2, zIl,z2l) are strictly positive on T2, then it is possible to express it as a
quotient of two minimum phase functions, each with support which is of infinite
extent along one coordinate axis and of finite extent along the other coordinate axis.
1.3 FEEDBACK STABILIZATION
PROBLEM #3
Problem in Stabilization of Feedback Systems
(Contributed by J. P. Guiver)
It has been seen that if n, d are polynomials in lR[z, w), which are relatively
prime and devoid of common zeros in U 2 , then, by using Hilbert's Nullstellensatz,
it was shown that there e polynomials y, x, and'I/J in lR[z, w) and an integer N > 0
can be found such that
yd+xn = 'l/JN
where'I/J has no zeros in U 2 . The objective is to replace lR[z, w) by Q[z, w), where
Q is the field of rational numbers and investigate into the following problem.
Let d(z, w), n(z, w) be relatively prime polynomials in Q[-?', w), devoid of com-
mon zeros in U 2. Find q(z,w),y(z,w),x(z,w) in Q[z,w], after proving their
existence such that
yd+xn = q
and
-2
q has no zeros in U
For motivation and illustration consider the following simple example. Let
n(z, w) = z2 + 2z - 1
d(z, w) = z - w + 2
which have common zeros for values (z,w) equalling (-1 -...;2,1 -...;2) and
(-1 + ...;2, 1 + ...;2). Clearly, via the procedure advanced in Chapter 2, together
with the use of Grobner basis,detailed in Chapter 4
'I/J(z,w) = zw + (1 + ...;2)(w - z) - (1 + ...;2)2

x(z,w) = 1
y(Z, w) = -(Z + 1 + 0)
satisfy yd + xn = 'I/J with 'I/J(z, w) having no zeros in U2 . However, 'I/J has non-
rational coefficients.
By the continuous dependence of the zeros of a polynomial on its coefficients
and because of the compactness of the unit bidisc, there exist intervals within
which the coefficients of x(z, w) and y(z, w) can vary whilst not destroying the
property that xn + yd has no zeros in U2 • In particular, we perturb the coefficients
of x(z, w) and y(z, w) so that those are rational.
This argument will hold in general and proves the existence of q(z, w) E
Q[z, w]. Let the perturbed polynomials be,
x(z,w) = 1
y(z, w) = -(z + f3)
where the parameter f3 will be determined so that
x(z, w)n(z, w) + y(z, w)d(z, w)

= zw + f3( w - z) - (1 + 2f3)
~ q(z,w)
has no zeros in U2 . Invoking the tests for absence of zeros of q( z, w) in U2 , i.e.

(i) q(z, w) -=1= 0 in T2, (ii) q(z, 1) -=1= 0 in U I and (iii) q(1, w) -=1= 0 in U I , it is easy
to infer that a suitable choice for f3 is f3 = 1.
The problem posed here, with the polydomain of interest U 2 , could be extended
easily to the case of an arbitrary compact polydomain.
Given the transfer function of the plant, p = ~, in reduced form, where f and
g belong to the multivariate polynomial ring lR[Zl' Z2, ... ,zn], when is there a
stabilizing compensator c = ~~, with, first, hI, h2 E lR[Zl' Z2, ... ,zn] If this is
found to be true, the second part seeks the answer to the same question when hI, h2
are further restricted to belong to Q[ZI' Z2, •.. ,zn].
The first part of the above question was answered as a generalization of the
result for the 2-D case (Chapter 2). Making use of the fact that the number of
common zeros in any two relatively prime bivariate polynomials is finite, Guiver
and Bose were able to construct a desired compensator after equating the sta-
bilizability issue of the plant to the criterion for absence of zeros in the closed
unit bidisc of the numerator and denominator polynomials of the plant transfer
function. They also posed the second question with some supporting arguments
in the 2-D case. To generalize the first part of the 2-D result to n-D, where the
common zeros of two relatively prime n-variate polynomials need not be isolated
and, indeed, can belong to a (2n - 4)-dimensional real Euclidean space, Shankar
and Sule [132], employing mathematical resources available in commutative alge-
bra and topology, obtained necessary and sufficient conditions for stabilizability
in an arbitrary compact polydomain (obviously, including the polydisc) of plants

whose transfer functions are fractions over a general integral domain. Their re-
sult, expressible in terms of coprimeness of ideals instead of zero-coprimeness of
polynomials is a generalization of the 2-D result over unit bidiscs. Thus, their
approach is coordinate-free in the sense that a particular coprime fraction descrip-
tion is not required because the ideals are invariant with respect to the fractional
representation in the ring of, say stable causal transfer functions (instead of just the
polynomial ring). They also studied robustness issues after defining a topology on
the set of transfer functions and concluded that the stabilizability property is not
generic for the transfer function set. The problem of construction, however, was
not addressed in [132]. An approach towards construction, using Rabinowitsch's
trick (this, due to Rabinowitsch in 1929, introduces an additional indeterminate
in a way that makes it possible to go from the weaker form of Hilbert's Nullstel-
lensatz in Fact 5.7 to Hilbert's (strong) Nullstellensatz on p.130 that supplies an
isomorphism between radical ideals and affine varieties) and Grabner bases was
presented in [133] for the 2-D case. In the n-D case, the problem should be possi-
ble to tackle in the case of zero-dimensional variety (isolated common zeros as in
the bivariate case). The challenge to effective construction, in general, lies when
the variety is of dimension greater than zero.
1.4 ROBUST CONTINUOUS SYSTEM STABILITY
PROBLEM #8
Invariance of Stability Property under Coefficient Perturbation
In system design, it is often necessary to preserve one or more characteristics of

the system when system element values fluctuate about their respective nominal
values. Investigations into the conditions for invariance of the multivariate polyno-
mial positivity property under coefficient perturbation were undertaken in [343].
Investigate into the possibility of obtaining the multivariate mathematical counter-
part of Kharitonov's theorem for polynomial f(Pl,P2, ... ,Pn) which is required
to have no zeros in the closed right-half polydomain, Re Pi ~ 0, i = 1, ... , n. In-
vestigate also other feasible n-dimensional generalizations of Kharitonov's result,
centering around those which find relevance in the context of studies of different
types of stability for multidimensional systems.
Open Problem 8 was concerned with the obtaining of multivariate counterparts
of Kharitonov's celebrated theorems on vertex implications (in the case of boxed
domains in parameter space) for inferring invariance of the strict Hurwitz property
of interval polynomials in both the real and complex coefficient cases. A degree-
independent vertex-implication result in parameter space (like Kharitonov's, where
a small number, independent of polynomial degree, of points in parameter-space
allows the test for invariance of a property of interest to be implemented for the
complete boxed domain associated with the interval system) on a bivariate interval
polynomial for invariance of the scattering Hurwitz property was proved in [243],
where also the n-variate result was conjectured. This n-variate conjecture was
proved in [246]. In [246] , a subclass of scattering Hurwitz (stable) polynomials
called strict sense Hurwitz (stable) polynomials was also studied for robust interval
strict sense Hurwitz stability. In the univariate case the scattering Hurwitz and
strict sense Hurwitz definitions coincide. More recently [247], the maximal class
of real and complex coefficient multivariate polynomials were defined such that
polynomials from this class do not lose the particular stability property under small
coefficient variations~ All the coefficients of any polynomial in this maximal class
must be of like sign, a property which implies that such a polynomial cannot be
lacunary. It is also noted that an univariate strict Hurwitz polynomial (devoid of any
zero in the closed right-half plane) cannot have missing coefficients and, therefore,
cannot be lacunary. In [248],[247] it was further shown that the maximal class of
polynomials referred to satisfy edge implications in parameter space for poly topic
coefficent (parameter) variations and vertex implications for interval and diamond
families of polynomials. The zero sets of families of multivariate polynomials
as well as quasipolynomials are also considered, albeit briefly, in [239] and a
Special Issue of the International Journal of Multidimensional Systems and Signal
Processing (vol. 5, no. 4, 1994) was entirely devoted to the topic of robustness of
multidimensional systems.
1.5 STABILITY OF DISCRETE SYSTEMS UNDER COEFFICIENT

PERTURBATION
PROBLEM #9
Stability of discrete systems under coefficient perturbation
An univariate polynomial which has all its zeros within the unit circle is a strict
Schur (SS) polynomial. Such a polynomial occurs in the stability studies of linear
time-variant digital filters. Investigate the scope for obtaining the counterpart of
Kharitonov's result stated in Problem # 8, for the case of a set of strictly Schur
polynomials.
In response to Open Problem 9, it was shown via a counterexample by N. K. Bose
and E. Zeheb [344] that unlike for the continuous-time case, the vertex implication
for BmO stability does not hold for an interval polynomial in the discrete-time
case. This is because, in general, an interval polynomial cannot be inferred to
belong to the SS class by testing only a finite number of extreme polynomials. The
example given next substantiates this.
Example 8.1 [344] Consider the polynomial,

4 3 3 2 1
g(b1,z) = z +b1z + 2"z - 3'
where the coefficients b1 varies over the interval, [_1;, ~]. It can be verified that
g(O, z) is not strict Schur, even though the distinct extreme polynomials, g( _1; ,z)
and g( 1; , z), are each strict Schur.
Bartlett, Hollot, and Huang [241] advanced the edge theorem which is widely
quoted in the literature [237] [238]. This edge implication in parameter space
holds not only for univariate interval polynomials but is applicable to the problem
of ascertaining robust D-stability of a poly topic set of polynomials, each of whose
members are required to have all its zeros in a region D of the complex plane. The
multivariate counterpart of the edge theorem has also appeared in the literature
[246]and is available for use in robust Dn -stability investigations of various classes
of Schur polynomials [345], where Dn is the corresponding polydomain of interest.
Constructive methods for the robust stability analysis of nonaffine families of
multivariate polynomials need to be developed. Extreme point algorithms for
construction of the minimal testing sets need also to be synthesized in that case at
least for particular classes.
1.6 ROBUST STABILITY OF MULTIVARIABLE SYSTEMS
PROBLEM #10
Stability of Matrix Polynomials Under Coefficient Perturbation
Let H(s) be a matrix whose entries are rational functions in s with real coef-
ficients. If H(s) is represented by a typical coprime matrix fraction description,.
H(s) = A-l(s)B(s) (where A(s) and B(s) are relatively left prime polynomial
matrices with det A( s) to), then H (s) characterizes a stable system if and only
if det A(s) has all its zeros in the open left half-plane, Re s < O. Of course, H(s)
could also be represented by a typical right coprime matrix fraction description,
D(s)C- 1 (s) (where C(s) andD(s) are relatively right prime polynomial matrices
with det C(s) to) and then for stability it would be necessary and sufficient that
det C (s) =1= 0 in Re s ~ O. For the sake of brevity, a C (s) satisfYing det C (s) =1= 0
in Re s ~ 0 will be called a strictly Hurwitz polynomial matrix. In attempting
to provide a generalization of the result linking strictly Hurwitz polynomials (Le.
polynomials devoid of zeros in Re s ~ 0) and reactance functions, Anderson and
Bitmead obtained certain interesting results [346] for determining whether or not
all the zeros of the determinant of the polynomial matrix belong to the open left-
half plane in terms of whether or not the reactance matrix property is satisfied by
a derived rational matrix. It should be noted, however, that the derivation of this
rational matrix from the specified polynomial matrix is not straightforward like in
the scalar case. Possibly exploiting the results in [346], obtain conditions under
which the linear convex combination of two strictly Hurwitz polynomial matrices
will be strictly Hurwitz. Investigate the scope for obtaining the polynomial matrix
counterpart of Kharitonov's result, stated in Problem # 8, when the coefficients of

the elements of the polynomial are bounded from above and below by specified
real numbers.
In the coprime MFD of a multivariate I-D plant, the system stability depends
on the restriction of the zeros of the determinant of a polynomial matrix in the
open left-half plane. In robust stability the problem of determining whether or
not the determinant of each element of an interval polynomial matrix (Le. each
coefficient of the matrix varies in a closed interval, independent of the remaining
coefficients) has all its zeros in the open left-half plane occurs. This problem was
investigated in [347], based on the use of a standard multidimensional digital filter
stability criterion and a generalization of the Kharitonov theorem considered in
[238, Chapter 7]
2. New Open Problems

2.1 GREATEST COMMON LEFT (RIGHT) DIVISOR COMPUTATION
Let F E K'Xm[z] with I ~ m be a multivariate polynomial matrix. Let d be

the GCD of all the maximal order minors of F. A conjecture made in Chapter 5
was that if the reduced minors of F, obtained after extraction of this GCD from the
maximal order minors, are zero-coprime then F can be factored as F = Go Fo,
for some Fo E KmXI[z], Go E K'XI[z] with det Go = d. A proof of the validity
of this conjecture has been given in [213]. It was also seen that the GCLD (or
GCRD) of a composite multivariate polynomial matrix (i. e. the matrix is viewed
as a concatenation of two or more matrices so that the usage of the adjective
"common" is appropriate; otherwise GCLD (GCRD) should be replaced by GLD
(GRD» in three or more indeterminates may not exist. Under the assumptions
that the reduced minors of F are not zero coprime and the GCLD exists, an open
problem is to find an algorithm for computing the GCLD. Only partial answers are
available for this problem [206]. The following example illustrates the validity of
the open problem posed above.
Example 8.2 The matrix C and its three 2 x 2 minors are
Zt{Zl -1) Zl - Z2Z3 - Z3

C= ZlZ3 - 2Z1 - Z3]
[ Zl -1 z~ + 2Z2 + 2 Z2+ Z3 -1 '
M12 (Zl - + I)(ZlZ2 + Zl + Z3),

l)(z2
M23 (-Z2Z3 - Z3 + 2Z2 + 3)(ZIZ2 + ZI + Z3),
MI3 (ZI - I)(ZIZ2 + ZI + Z3).
It is obvious that the GCD of M 12 , M23, and M 13 is (ZIZ2 + ZI + Z3); thus the
set of reduced minors is {(ZI - 1)(z2 + 1), (-Z2Z3 - Z3 + 2Z2 + 3), (ZI - 1)}.
The common zeros of the reduced minors are determined from the solution set,
{ ZI = 1, Z2 = ~;-=~, Z3 is arbitrary} . To show that a GLCD exists, first note
the factorization
ZI -Z3 ] [ ZI - 1 1 Z3 - 2 ]
C=GCl = [ 1
Z2 + 1 0 Z2 +1 1 '
where G is not unimodular. To prove that G is a GCLD, unique upto unimodularity,

it suffices to show that there exists no other nonunimodular matrix Gl E K2X2[z]
such that C 1 = G 1C2 for some matrix C2 E K 2X 3[z]. By contradiction, if there
exists such a matrix Gl then the major determinants of C1 must be divisible by the
determinant of Gl, which is impossible since the the minors of C1 are the same
as the reduced minors of C and have no nontrivial common factor. Thus, G is a
GCLD which is unique upto unimodularity.
The question is to advance a general method for factorization when the factor exists
because algorithms applicable in special cases (like when the set of reduced minors
are zero coprime) in [206] do not work.
2.2 DETERMINANTAL (LEFT RIGHT JOINT) FACTORIZATION
Let for brevity, D (like K[z]) be a Noetherian unique factorization domain

(UFO). Consider ap x p matrix FE Dpxq, where p ~ q(p ~ q). Let 9 E D be a
common divisor of the p x p minors (q x q minors) of F. A left (right) determinantal
factorization of F with respect to 9 is F = GFI (F = FIG), where G E DPxp,
Fl E Dpx q, and det(G) = 9 (G E Dqx q, Fl E Dpx q, and det(G) = 9). When
D = K[ZI' Z2], determinantal factorization is always possible. It is well known
that D = K[z] is not a DFD, when n ~ 3, because here factor primeness of
a matrix does not, necessarily, imply minor primeness. The joint left and right
factorization, considered next, may be possible when left or right determinantal
factorization is not possible.
Consider F E DPxp, where p ~ q. Suppose that the greatest common divisor of
the p x p minors of F is not irreducible. Let 91 E D and 92 E D be two common
divisors of the p x p minors of F. A determinantal left and right factorization
of F with respect to 91 and 92 is F = G1FIG2 where G1 E DPxp, F E DPxq,
G2 E Dqxq, det(Gt} = 91, and det(G 2) = 92. The case when p ~ q can be
similarly defined.
It is possible that a multivariate polynomial matrix with elements in K[z] might
not yield a left or right determinantal factorization but a left and right joint factor-
ization may be possible as illustrated next. Consider the 4 x 5 matrix A shown
below whose elements belong to the ring K[Zl' Z2, Z3]. This particular matrix over
a binary field was, possibly, first used by Paul Wiener in his study of multidimen-
sional convolution codes [201].
A=
o
Z2 ]
Z3
Zl +Z3
It is easily shown by routine manipulation that the 5 reduced minors of the 5

major determinants of A have a common zero at (0,0,0). The above matrix A does
not have a determinantal (left) factorization. A joint left and right factorization
is exhibited below both when K is the finite binary field Z2 and also when it is
the field of real numbers JIt The computations are easily implemented by using
the computer algebra system SINGULAR which is suitable for calculating Grabner
bases of ideals in K[z] as well as modules whose elements belong to K[z]. The
factorizations of A in fields Z2 and ~ respectively, are given below.
1 0 0 0 0
0 1
[Z'
z 201 1
0 11][ 1 0
0 0 1 Z2 1
Z2 +Z3 ]
0
0
1
0
0 0
0
0
0
0 Zl + Z2 + Z2 + Z3 Z3
Zl 1 0 1 1 0 Zl
0 0 0 1 0
Zl 0 1 1 0 Z2 1 Zl Zl +Z2
0 0 0 0 1
1 0 0 0 0
[Zl-Z2 -2 0 -1
o
o
1 -1
0
1 0
1
~ ][ ;2
o Zl
Zl
Z2
1
0
Zl
Z2
-1
0
Z3
] 0
0 0
1 0
Z3
0
0 0
0
0 0 0 1 0
o 0 0 1 Zl Zl 0 0 Zl +Z3
0 0 0 0 1
The problem is to investigate necessary and sufficient conditions for a left and right
joint multivariate polynomial matrix factorization to hold and explore algorithmic
approaches to construct such a factorization. The case when the specified matrix
is not of full nonnal rank needs to be addressed subsequently. The limitations of
Grabner basis theory for the purpose needs to be fully delineated.
2.3 UNIMODULAR COMPLETION OVER RING OF PROPER STABLE

MULTIVARIATE RATIONAL FUNCTIONS
Unlike FIR filters, llR filters are not guaranteed to be stable The set of all proper
structurally stable rational functions in n variables is known to fonn a commutative
ring [348]. It is natural to limit the scope of focus to the noncommutative ring of
square matrices whose entries are proper (structurally) stable multivariate rational
functions. Perfect reconstruction (PR) is possible by constraining the analysis and
synthesis filter banks to be unimodular or in algebraic tenns, restricting those to
be units in the ring. In the IIR case, the units of the ring are matrices whose
determinants are minimum phase transfer functions. The problem of designing
aPR IIR filter bank is then reduced to the finding of a polyphase matrix H(z),
associated with the analysis filter bank, that is a stable proper rational matrix
with minimum phase determinant. The polyphase matrix G(z), associated with
the synthesis filter bank, can be obtained by directly computing the inverse of
the polyphase matrix of the analysis filter bank. Since the determinant of the
polyphase matrix H(z) is minimum phase, the synthesis filter bank associated
with the polyphase matrix G(z) is guaranteed to be stable.
If one of the filters in the analysis IIR filter bank is specified and its polyphase
components are devoid of common zeros, similar to the FIR case, is it always
possible to obtain the complete analysis-synthesis IIR filter bank by unimodular
completion [348]? The solution to this type of unimodular completion problem is
likely because a set of all proper stable rational multivariate functions is known to
form a commutative ring. One then needs to determine if the Quillen-Suslin theo-
rem always holds for this ring. Then find a constructive procedure for unimodular
completion over the ring of stable rational multivariate functions.
Summarizing, given is a set of m rational functions hi = hi(z), of several
variables z = (Zi' Z2, ... , zn) such that:
1 the hi'S are devoid of nonessential singularities of the first kind (that is, their
denominators are devoid of zeros on fl'" (NSFK) on the closed unit polydisc
fl'" .,
2 the set of hi'S is a zero-coprime set on fl"'.
One question is to determine if it possible to complete the (m x m) rational matrix

H{z) for which
1 [hi, h2 , ... , hml is the first row of H{z) .
2 all elements of H (z) have no NSFK on If''.
3 det (H) is a minimum phase function (in the sense that both its denominator
and its numerator are zero-free in If'' i. e. structurally stable with a structurally
stable inverse), which is a unit in the ring.
First verify if it is true when n=l, 2 and the hi'S are rational and then for arbitrary
n with hi's polynomials. Then the open problem is for the case when n ~ 3 and
the hi'S are rational; the second question, a constructive procedure for unimodular
construction for the specific base ring, also remains to be settled. For background
material, see [133], [349], [349, Reference 21] after studying Chapter 2 and, pos-
sibl y, referring to [131]. Some of the results and references for Section 2.1 are of
relevance here and vice-versa.
2.4 MINIMAX CONTROLLER AND GROBNER BASES
GivenaplantF(s) = l/m(s), where

n
m(s) = II (S2 + (3i), 0 < {3i < {3i+1
i=l
is an even polynomial with simple roots on the imaginary axis, in the complex s-
plane, the problem is to find from the uncountably infinite set of odd polymonials
{ni (s )}, whose generic element has the form
n-1
ni(s) = ki S II (s2 + 'Yl(i))' ki > 0, (3, < 'Yl(i) < (3,+!

1=1
the one, denoted for brevity, by
n-1
n(s) = ks II (s2 + 'Yl), k > 0, {3, < 'Yl < (3,+!
1=1
so that the characteristic polynomial, m(s) + n(s), of the resulting optimal rate
feedback system has the fastest slowest mode among the set of strict Hurwitz
polynomials, {m( s) + ni (s)}. In other words, the right most roots of m( s) + n( s)
are required to be the farthest to the left of the imaginary axis in comparison to
similar roots for any other polynomial in the set {m(s) + ni(s)}.
The solution for the n = 1 and n = 2 cases are straightforward to obtain
[350]. The complete analytic characterization and solution construction (either
explicitly or by recursion) for the minimax control problem- using optimal rate
feedback is given for the case when the plant consists of a known fixed set of
coupled oscillators of cardinality not exceeding tbree[351]. The tools used are those
originating in realizability theory (especially with respect to the determination of
qualitative properties of solutions of polynomial equations) and Grobner bases
in polynomial ideal theory. The computations for reaching the optimal design
are recursive in nature and easily implementable with arbitrary precision. The
extent to which such a characterization and recursive mode solution construction
is possible, is shown for the n = 4 case[352]. When the plant is of higher order, the
construction of an optimal solution appears to be analytically and computationally
intractable necessitating the need for numerical methods for generating satisfactory
suboptimal solutions. Itremains to investigate if this problem, in particular, belongs
to the class of N P - complete problems and, in general, what the alternatives
are to the percieved limitations of resources in algorithmic algebra (like Grobner
bases for tackling hard problems of the type encountered here and elsewhere (like,
for example, in Section 2.1) that can be recast within the framework of polynomial
ideal and module theory.
References
[1] N; K. Bose (ed.), "Special issue on multidimensional systems," Proc. IEEE, vol. 65, June
1977.
[2] M. G. Ekstrom and S. K. Mitra (eds.), Two-Dimensional Signal Processing, Dowden,

Hutchinson and Ross, New York:, 1978.
[3] N. K. Bose (ed.), Multidimensional Systems: Theory and Applications, IEEE Press, New
York, 1979.
[4] A. S. Will sky, Digital Signal Processing and Control and Estimation Theory: Points
of Tangency, Areas of Intersection, and Parallel Directions, The MIT Press, Cambridge,
Massachusetts, 1979.
[5] N. K. Bose, Applied Multidimensional Systems Theory, Van Nostrand Reinhold, New York,
1982.
[6] S. M. Nikolskii, Approximation of Functions of Several Variables and Imbedding Theorems,

Springer-Verlag, Berlin, 1975.
[7] W. Rudin, Function Theory in Polydiscs, W. A. Benjamin Inc., New York, 1969.
[8] Sh. A. Dautov, "Some questions of multidimensional complex analysis," Akad. Nauk SSSR
Sibirisk, Otdel. Inst. Fiz. Krasnoyarsk, p. 19, 1980, (in Russian; translation help provided
to the author by Professor Y. Tsypkin through the courtesy of Professor E. I. Jury).
[9] Sh. A. Dautov, "On absolute convergence of the series of taylor coefficients of a rational
function of two variables: Stability of two-dimensional recursive digital filters," Soviet Math.
Dokl., vol. 23, no. 2, pp. 448-451, 1981, (American Mathematical Society Translations).
[10] D. Goodman, "Some Stability Properties of1Wo-dimensional Linear Shift-invariant Digital

Filters," IEEE Trans. Circuits and Systems: I, vol. 24, pp. 201-208, 1977.
[11] R. K. Alexander and J. W. Woods, "2-d digital filter stability in the presence of second-kind
nonessential singularities," IEEE Trans. Circuits and Systems: /, vol. 29, pp. 604--612,
September 1982.
[12] I. E. Zak, "On absolute convergence of double fourier series," Soobsceniya Akad. Nauk
Gruzin. SSR, vol. 12, pp. 129-133, 1951.
243
[13] P. Delsarte, Y. V. Genin, and Y. V. Kamp, "A simple proof of Rudin's multivariable stability
theorem," IEEE Trans. Acoust., Speech, and Signal Process., vol. 28, pp. 701-705, December
1980.
[14] E. LJury, "Stability of multidimensional systems and related problems," in Multidimensional

Systems: Techniques and Applications, Spyros G. Tzafestas, Ed., New York:, NY, 1986, pp.
89-159, Marcel Dekker, Boston, MA.
[15] T. S. Huang (ed.), Two-Dimensional Digital Signal Processing 1: Linear Filters, vol. 42,
Springer-Verlag, Topics of Applied Physics Series, New York, 1981.
[16] A. Fettweis, ''On the scattering matrix and the scattering transfer matrix of multidimensional
10ssless two-parts," Archiv for Elektronik und Ubertragungstechnik, vol. 36, pp. 374--381,
1982.
[17] J. Gregor, ''Biquadratic n-dimensional impedances," Int. J. Circuit Theory and Applications,
vol. 9, pp. 369-377, 1981.
[18] P. Delsarte, Y. V. Genin, and Y. V. Kamp, "An equivalence between bounded multivariable
functions and a class of bounded single variable functions," Int. J. Control, vol. 34, pp.
383-389, 1981.
[19] P. Delsarte, Y. V. Genin, and Y. V. Kamp, "Koga's multivariable stability criterion," Proc.
IEEE, vol. 70, pp. 298-299, March 1982.
[20] E. W. Kamen, "Linear systems with commensurate time delays: Stability and stabilization
independent of delay," IEEE Trans. Automatic Control, vol. 27, pp. 367-375, April 1982,
also corrections in IEEE Transactions on Automatic Control, Vol. 28, Feb. 1983, pp. 248-249.
[21] W. E. Brumley, ''On the asymptotic behavior of solutions of differential-difference equations

of neutral type," J. Differential Equations, vol. 7, pp. 175-188, 1970.
[22] P. S. Gromova, "Stability of solutions of nonlinear equations of the neutral type in the
asymptotically critical case," Mathematical Notes of the Academy of Sciences of the USSR,
vol. 1, no. 5, pp. 472-479, May-June 1967, (Translated from Mathematicheskie Zametki,
Vol 1., No.6, June 1967, pp. 715-726).
[23] J. P. Guiver and N. K. Bose, ''On test for zero-sets of multivariate polynomials in noncompact
polydomains," Proc. IEEE, vol. 69, pp. 467-469, April 1981.
[24] Y. V. Genin, "An algebraic approach to A-stable linear multistep multiderivative integration
formulas," BIT, vol. 14, pp. 382-406, 1974.
[25] P. Delsarte, Y. V. Genin, and Y. V. Kamp, "A proof of the Daniel-Moore conjectures for
A-stable multistep two-derivative formulae," Philips J. Research, vol. 36, pp. 79--88, 1981.
[26] G. Dahlquist, Positive Functions and Some Applications to Stability Questionsfor Numerical
Methods, Academic Press, New York, 1978, in de Boor, C. and Golub, G. (eds.), Recent
Advances in Numerical Analysis.
[27] J. B. EdwarQs and D. H. Owens, Analysis and Control of Multipass Processes, Research
Studies Press, a division of John Wiley and Sons Ltd., Chichester, England, 1982.
[28] Y. Rouchaleau, B. F. Wyman, and R. E. Kalman, "Algebraic structure of linear dynamical

systems iii: Realization theory over a commutative ring," Proc. Nat. Acad. Science, USA,
vol. 69,pp.3404-3406, 1972.
REFERENCES 245
[29] Y. Rouchaleau and E. D. Sontag, "On the existence of minimal realizations of linear systems
over noetherian integral domains," 1. Computer and Systems Science, vol. 18, pp. 65-75,
February 1979.
[30] E. D. Sontag, "The lattice of minimal realizations of response maps over rings," Math.
Systems Theory, vol. 11, pp. 169-175, 1977.
[31] E. D. Sontag, "On split realizations of response maps over rings," Info. and Control, vol.
37, pp. 23-33, April 1978.
[32] E. D. Sontag, "Linear systems over commutative rings: A survey," Ricerche di Automatica,
vol. 7, pp. 1-34, 1976.
[33] R. Eising, "State-space realization and inversion of 2-D systems," IEEE Trans. Circuits and
Systems: I, vol. 27, pp. 612-619, July 1980.
[34] S. G. Tzafestas and T. G. Pimenides, "Feedback characteristic polynomial controller design

of3-d systems in state-space," 1. Franklin Institute, vol. 314, pp. 169-189, September 1982.
[35] S. G. Tzafestas and T. G. Pimenides, "Feedbackdecoupling-controller design of3-d systems

in state-space," Mathematics and Computers in Simulation, vol. 24, pp. 341-352, 1982.
[36] M. Fliess, "Sur la realisation des systemes dynamiques bilineaires," C. R. Acad. Sci., Paris,
vol. 227A, pp. 923-926, 1973.
[37] E. D. Sontag, "On linear systems and noncommutative rings," Math. Systems Theory, vol.
9, pp. 327-344, 1976.
[38] E. Fomasini, "On the relevance of noncommutative power series in spatial filter realization,"
IEEE Trans. Circuits and Systems: I, vol. 25, pp. 290-299, May 1978.
[39] J. P. Guiver and N. K. Bose, "Polynomial matrix primitive factorization over arbitrary
coefficient field and related results," IEEE Trans. Circuits and Systems: I, vol. 29, pp.
649-657, 1982.
[40] D. C. Youla and G. Gnavi, "Notes on n-dimensional system theory," IEEE Trans. Circuits
and Systems: I, vol. 26, pp. 105-111, 1979.
[41] D. C. Youla and P. F. Pickel, ''The Quillen-Suslin theorem and the structure of n-dimensional
elementary polynomial matrices," IEEE Trans. Circuits and Systems: I, vol. 31, pp. 513-518,
June 1984.
[42] T.Y. Lam, Serre's Conjecture, Springer, Berlin, lecture notes in mathematics, 635 edition,
1978.
[43] J. J. Rotmann, An Introduction to Homological Algebra, Academic Press, New York, 1979.
[44] P. P. Khargonekar, "On matrix fraction representation for linear systems over commutative
rings," SIAM 1. Control and Optimization, vol. 20, pp. 172-197, March 1982.
[45] P. A. Fuhrmann, "Algebraic systems theory: An analyst's point of view," 1. Franklin Institute,
voI.305,pp.521-540,1976.
[46] G. Conte and P. M. Perdon, "Systems over a principle ideal domain: A polynomial model
approach," SIAM 1. Control and Optimization, vol. 20, pp. 112-124, January 1982.
[47] c. I. Byrnes, "On the control of certain deterministic infinite-dimensional systems by

algebro-geometric techniques," Amer. J. Math., vol. 100, pp. 1333-1381, 1978.
[48] W. Cauer, '"The Poisson integral for functions with positive real part," Bull. Amer. Matk
Soc., vol. 38, pp. 713-717,1932.
[49] E. J. Beltrami and M. R. Wohlers, Distributions and the Boundary Values of Analytic
Functions, Academic Press, New York, 1966.
[50] A. Koranyi and L. Pulalnszky, "Holomorphic functions with positive real part on polycylin-
ders," Trans. Amer. Matk Soc., vol. 108, pp. 449-456, 1963.
[51] I. A. Aizenberg and A. P. Yuzhakov, "Integral respresentation and residues in multidimen-

sional complex analysis," Trans. of Math. Monagraphs, Amer. Math. Society, Providence,
R. I., vol. 58, 1983.
[52] L. A. Aizenberg and S. A. Dautov, "Holomorphic functions of several variables with nonneg-
ative real parts; traces of holomorphic and pluriharmonic functions on the silov boundary,"
Mat. Sb. USSR, vol. 28, no. 3, pp. 301-313.
[53] A. Koninyi, '"The Poisson integral for generalized half-planes and bounded symmetric
domains," Annals Matk, vol. 82 (2nd series), pp. 332-350, 1965.
[54] v. S. Vladimorov, "Holomorphic functions with nonnegative imaginary part in a tubular

domain over a cone," Math. USSR Sbomik, vol. 8, no. 1, pp. 126-146, 1969.
[55] V. S. Vladimorov, "Holomorphic functions with positive imaginary part in the future tube
I and IT," Math. USSR Sbomik, vol. 22, pp. 1-16, 1974, (Math. USSR Sbornik, 23, No.4,
1976, pp. 467-482).
[56] J. H. McClellan, "Multidimensional spectral estimation," Proc. IEEE, vol. 70, pp. 1029-
1039, September 1982.
[57] J. A. Shohat and J. D. Tamarldn, '"The problem of moments," Amer. Matk Society, Mathe-
matical Surveys, New York, , no. 1, 1943.
[58] W. F. Stinespring, "Positive functions on C* -algebras," Proc. Amer. Matk Society, vol. 6,
pp.211-216,1955.
[59] . K. Schmiidgeon, "An example of a positive polynomial which is not a sum of squares of
polynomials - a positive but not strongly positive functional," Math. Nachr., vol. 88, pp.
385-390, 1979.
[60] M. D. Choi and T. Y. Lam, "External positive semidefinite forms," Annals Matk, vol. 231,
pp. 1-18, 1977.
[61] E. K. Haviland, "On the momentum problem for distributions in more than one dimension,"
Amer. J. Matk, vol. 57, pp. 562-568, 1935.
[62] E. K. Haviland, "On the momentum problem for distribution functions in more than one
dimension 2," Amer. J. Math., vol. 58, pp. 164-168, 1936.
[63] H. Cramer, Random Variables and Probability Distributions, Cambridge University Press,
1970.
REFERENCES 247
[64] R. T. Powers, "Self-adjoint algebras of unbounded operators," Comm. Math Phy., vol. 21,
pp. 85-124, 1971. .
[65] W. Rudin, ''The extension problem of positive definite functions," Illinois J. Maths, vol. 7,
pp.532-539,1963.
[66] G. Cybenko, ''Moment problems and low rank Toeplitz approximations," Circuits, Systems
and Signal Process., vol. 1, no. 3-4, pp. 345-366, 1982.
[67] Y. M. Bruck and L. G. Sodin, "On the ambiguity of the image reconstruction problem,"
Optical Communications, vol. 30, pp. 304--308, September 1979.
[68] L. Carlitz, ''The distribution of irreducible polynomials in several indeterminates," Illinois

J. Math., pp. 371-375, 1963.
[69] M. M. Hayes and 1. H. McClellan, "Reducible polynomials in more than one variable,"
Proc. IEEE, vol. 70, pp. 197-198, February 1982.
[70] I. Manolitsakis, ''Two-dimensional scattered fields: A description in terms of the zeros of

entire functions," J. Math Phys., vol. 23, pp. 2291-2298, December 1982.
[71] T. Nomura, H. Miyakawa, H. Imai, and A. Fukuda, "A theory of two-dimensional linear
recurring arrays," IEEE Trans. Inform. Theory, vol. 18, pp. 775-785, November 1972.
[72] S. Sakata, "General theory of doubly periodic arrays over an arbitrary finite field and its
applications," IEEE Trans. Inform. Theory, vol. 24, pp. 719-730, November 1978.
[73] S. Sakata, "On determining the independent point set for doubly periodic arrays and encoding
two-dimensional cyclic codes and their duals," IEEE Trans. Inform. Theory, vol. 27, pp.
556--565, September 1981.
[74] M. Nieto Vesperinas, "Dispersion relations in two dimensions: Application to the phase
problem," Optik, vol. 56, pp. 377-384, 1980.
[75] H. Helson and D. Lowdenslager, "Prediction theory and Fourier series in several variables,"
Aita Math, vol. 99, pp. 165-202, 1958.
[76] T. L. Marzetta, ''Two-dimensional linear prediction: Autocorrelation arrays, minimum-phase

prediction error filters and reflection coefficient arrays," IEEE Trans. Acoust., Speech, and
Signal Process., vol. 28, pp. 725-733, December 1980.
[77] T. L. Marzetta, "Additive and multiplicative minimum-phase decompositions of 2-D rational

power density spectra," IEEE Trans. Circuits and Systems: I, vol. 29, pp. 207-214, April
1982.
[78] N. K. Bose Guest Editor, ," in Special Issue: Aspects of Spatial and Temporal Signal
Processing, Birkhauser Boston, Cambridge, MA, 1984, vol. 3, Birkhiiuser, Boston.
[79] D. I. Schmidlin, ''The characterization of causal shift-variant systems excited by causal

inputs," IEEE Trans. Circuits and Systems: I, vol. 28, pp. 981-994, October 1981.
[80] E. W. Kamen, "On the relationship between bilinear maps and linear two-dimensional
maps," Nonlinear Analysis: Theory, Methods and Applications, vol. 3, pp. 467-481, 1979.
[81] N. S. Williams and V. Zakian, "A ring of delay operators with applications to delay differ-
ential systems," SIAM J. Control and Optimization, vol. 15, pp. 247-255, 1977.
[82] M. Hazewinkel and R. E. Kalman, "Invariants, canonical fonns and modulii for linear
constant finite dimensional dynamical systems," Lecture Notes in Economics and Math.
Systems, vol. 131, pp. 48--60, 1976.
[83] R. Bumby, E. D. Sontag, H. I. Sussman, and W. Vasconcelos, "Remarks on the pole-shifting

problem over rings," J. Pure and Applied Algebra, vol. 20, pp. 113-127, 1981.
[84] E. D. Sontag, ''Linear systems over commutative rings: A (partial) updated survey," Proc.
lFAC, Kyoto, Japan, August 1981.
[85] V. S. Vladimorov, Generalized Functions in Mathematical Physics, 'Nauka', Moscow, 2nd

edition, 1979, English transl., 'Mir'Moscow, 1979.
[86] I. U. N. Drozzinov, "Passive linear systems of partial differential equations," Math. USSR
Sbomik, vol. 44, no. 3, pp. 269-278, 1983.
[87] I. U. N. DrozZinov, "A multidimensional Tauberian theorem for holomorphic functions of

bounded argument and the quasi-asymptotics of passive systems," Math. USSR Sbomik,
vol. 45, no. I, pp. 45-61, 1983.
[88] M. R. Teague, "Image analysis via the general theory of moments," J. Opt. Soc. America,
vol. 70, pp. 920-930, August 1980.
[89] L. A. Rubel, "Internal and external analytic functions of several complex variables," preprint
provided to the author.
[90] P. Delsarte, Y. V. Genin, and Y. V. Kamp, "Planar least-squares inverse polynomials part 2:
Asymptotic behavior," SIAM J. Alg. Disc. Meth., vol. 1, pp. 336-344, September 1980.
[91] E. Vanmarcke, Random Fields: Analysis and Synthesis, The MIT Press, Cambridge, MA,
1983.
[92] N. K. Bose, "Multidimensional digital signal processing: problems, progress, and future
scopes," Proc. IEEE, vol. 78, pp. 590-597, 1990.
[93] N. K. Bose, ''Two decades of multidimensional systems research and future trends," in
Multidimensional Signals, Circuits and Systems, K. Galkowski & I. Wood, Ed., London, U.
K., 2001, pp. 5-27, Taylor & Francis.
[94] I. Yamada, K. Kurosawa, H. Hagesawa, and K. Sakaniwa, "Algebraic phase unwrapping

and zero distribution of complex polynomials - characterization of multivariate stable poly-
nomials," IEEE Trans. Signal Process., vol. 98, no. 6, pp. 1639-1664, 1998.
[95] I. Yamada and N.K. Bose, "Algebraic phase unwrapping and zero distribution of polynomials
for continuous-time systems," IEEE Trans. Circuits and Systems: I, vol. 49, no. 3, pp. 298-
304,2002.
[96] F. G. Boese and W. I. Luther, "Enclosure of the zero-set of polynomials in several complex
variables," Multidimensional Systems and Signal Process., vol. 12, pp. 165-197,2001.
[97] Yuval Bistritz, "Stability testing of two-dimensional discrete-time systems by a scattering-

type stability table and its telepolation," Multidimensional Systems and Signal Process., vol.
13,pp.55-77,2002.
REFERENCES 249
[98] Minoru Yamada, Li Xu, and Osami Saito, "Further results on Bose's 2-D stability test," in
Proc. Fourteenth Int. Symp. Mathematical Theory of Networks and Systems (MTNS-2000),
Perpignan, France, June 2000.
[99] D. C. Youla, "Two observations regarding first-quadrant causal BIBO-stable digital filters,"
Proc. IEEE, vol. 78, pp. 598-603, 1990.
[100] Andrew T. Tomerlin and William W. Edmonson, "BIBO stability on D-dimensional filters,"
Multidimensional Systems and Signal Process., vol. 13, no. 3, pp. 333-340, 2002.
[101] E. Rogers and D. H. Owens, Stability Analysisfor Linear Repetitive Processes, Lecture Notes
in Control and Information Sciences 175. Springer-Verlag, Berlin Heidelberg, Germany,
1992.
[102] K. Galkowski, E. Rogers, S. Xu, J. Lam, and D. H. Owens, "LMIs- a fundamental tool
in analysis and controller design of linear repetitive processes," IEEE Trans. Circuits and
Systems: I, vol. 49, no. 6, pp. 768-778, 2002.
[103] L. Xu, Synthesis of Linear Discrete Multidimensional Control Systems, Ph.D. thesis, Toy-
ohashi University of Technology, Japan, Dept. Electrical Engineering, 1993.
[104] P. D. Roberts, "Two-dimensional analysis of an iterative nonlinear optimal control algo-

rithm," IEEE Trans. Circuits and Systems: I, vol. 49, no. 6, pp. 872-878, 2002.
[105] E. Rogers and D. H. Owens, 'Two decades of research in linear repetitive processes," in
Multidimensional Signals, Circuits and Systems, K. Galkowski & J. Wood, Ed., London, U.
K., 2001, pp. 107-147, Taylor & Francis.
[106] K. Galkowski, State-space Realizations of Linear 2-D Systems with Extensions to the
General n-D (n > 2) Case, Lecture Notes in Control and Information Sciences 263. Springer-
Verlag London Limited, London, Great Britain, 2001.
[107] E. Zerz, Topics in Multidimensional Linear Systems Theory, Lecture Notes in Control and
Information Sciences 256. Springer-Verlag London Limited, London, Great Britain, 2000.
[108] J. Gregor, "A formal approach to discrete systems theory," Appl. Math. and Compo Sci, vol.
7,no.4,pp. 775-794,1997.
[109] H. Pillai, J. Woods, and E. Rogers, "On homomorphisms of n-D behaviors," IEEE Trans.
Circuits and Systems: I, vol. 49, no. 6, pp. 732-742,2002.
[110] U. Oberst, "Multidimensional constant linear systems," Acta Applicandae Mathematicae,

pp. 1-175, 1990.
[111] S. Kleon and U. Oberst, "Transfer operators and state spaces for discrete multidimensional
linear systems," Acta Applicandae Mathematicae, pp. 1-82, 1999.
[112] N. K. Bose, "Matrix factorization in a real field," Linear Algebra and Its Applications, vol.
11,pp.21-25,1975.
[113] Y. Q. Shi and N. K. Bose, "Some results in nonnegativity constrained spectral factorization,"
Optics Comm., vol. 68, no. 4, pp. 251-256, October 1988.
[114] D. W. Richardson and N. K. Bose, "Necessary conditions for nonnegativity constrained

spectral factorization," Optics Comm., vol. 76, no. 1, pp. 8-12, April 1989.
[115] E. Landau, "Uber die darstellung definiter Funktionen durch Quadrate," Mathematische
Annalen, vol. 62, pp. 272,1906.
[116] S. Basu, "A constructive algorithm for 2-D spectral factorization with rational spectral
factors," IEEE Trans. Circuits and Systems: I, vol. 47, pp. 1309-1318, September 2000.
[117] T. Y. Lam, The Algebraic Theory of Quadratic Forms, Mathematics Lecture Note Series.
W. A. Benjamin, Inc., Reading, Massachusetts, 1973.
[118] C. N. Delzell, A Constructive Continuous Solution to Hilbert's 17th Problem and Other
Results in Semi-Algebraic Geometry, Ph.D. thesis, Stanford University, Math. Dept., June
1980.
[119] P. A. Parrilo, "Semidefinite programming relaxations for semialgebraic problems," Math.

Prog. Sere B, vol. 96, no. 2, pp. 293-320, 2003.
[120] A. Kummert, "Synthesis of two-dimensional lossless m-ports with prescribed scattering

matrix," Circuits, Systems and Signal Process., vol. 8, no. 1, pp. 97-119, 1989.
[121] N. K. Bose and Y. Q. Shi, "2-D Wilson spectral factorization," IEEE Trans. Acoust., Speech,
and Signal Process.; vol. 36, pp. 125-128, January 1988.
[122] B. O'Connor and T. Huang, "Stability of general two dimensional recursive digital filters,"
IEEE Trans. Acoust., Speech, and Signal Process., vol. 26, pp. 550-560, December 1978.
[123] R. Eising, "Realization and stabilization of 2-D systems," IEEE Trans. Automatic Control,
vol. AC-23, no. 5, pp. 793-799, October 1978.
[124] J. Cadzow, Discrete-Time Systems: An Introduction with Interdisciplinary Applications,

Prentice Hall, NJ, 1973.
[125] C. Desoer, R-W. Liu, J. Murray, and R. Saeks, "Feedback system design: The fractional
representation approach to analysis and synthesis," IEEE Trans. Automatic Control, vol.
AC-25, no. 3, pp. 399-412, June 1980.
[126] C. Desoer and W. Chan, ''The feedback inteconnection of lumped linear time-invariant
systems," J. Franklin Institute, vol. 300, no. 5/6, pp. 335-351, NovemberlDecember 1975.
[127] M. Vidyasagar, "On the use of right-coprime factorizations in distributed feedback systems
containing unstable subsystems," IEEE Trans. Circuits and Systems: I, vol. CAS-25, no.
11, pp. 916-921, November 1978.
[128] M. Morl, B. Levy, and S-Y. Kung, "New results on 2-D systems theory, Part I: 2-D poly-
nomial matrices, factorization and coprimeness," Proc. IEEE, vol. 65, pp. 861--S72, June
1977.
[129] G. Ferrani and G. Hausler, ''TV optical feedback systems," Optical Engineering, vol. 19,
pp. 442-451, July/August 1980.
[130] G. Hausler and N. Streibl, "Stability of spatio-temporal feedback systems," Optica Acta,
vol. 1930,pp. 171-187,1983.
[131] M. Vidyasagar, Control System Synthesis: A Factorization Approach, The MIT Press, The
MIT Press, Cambridge, MA, 1985.
REFERENCES 251
[132] S. Shankar and V. R. Sule, "Algebraic geometric aspects of feedback stabilization," SIAM
J. Control and OptimiZlJtion, vol. 30, no. 1, pp. 11-30, 1992.
[133] L. Xu, O. Saito, and K. Abe, "Output feedback stabilizability and stabilization algorithms
for 2-D systems," Multidimensional Systems and Signal Process., vol. 5, pp. 41-60, 1994.
[134] S. Shankar, "An obstruction to the simultaneous stabilization of two n-D plants," Acta
Applicandae Mathematicae, vol. 36, pp. 289-301, 1994.
[135] Jiang Qian Ying, "A computational method for determining strong stabilizability of n-D
systems," J. Symbolic Computation, vol. 27, pp. 479-499, 1999.
[136] Jiang Qian Ying, "On the strong stabilizability of MIMO n-dimensional linear systems,"
SIAM J. Control and Optimization, vol. 38, no. 1, pp. 313-335, 2000.
[137] E. Emre, "The polynomial equation QQ c +RPc = 4> with application to dynamic feedback,"
SIAM J. Control and Optimization, vol. 18, no. 6, pp. 611-620, November 1980.
[138] W. Gustafson, "Roth's theorem over commutative rings," Linear Algebra and Its Applica-
tions, vol. 23, pp. 245-251, 1979.
[139] W. Fulton, Algebraic CunJes: An Introduction to Algebraic Geometry, Benjamin/Cummings,

Massachusetts, 1969.
[140] W. S. Ching, "Linear equations over commutative rings," Linear Algebra and Its Applica-
tions, vol. 18, pp. 257-266, 1977.
[141] I. Kaplansky, Commutative Rings, Allyn and Bacon, Boston, 1970.
[142] B. L. Van der Waerden, Modem Algebra, vol. II, Ungar, NY, 1950.
[143] D. Northcott, Ideal Theory, Cambridge University Press, 1953.
[144] G. Hermann, "The question of finitely many steps in polynomial ideal theory (German),"
Mathematische Annalen, vol. 95, pp. 736--788, 1926.
[145] B. Buchberger and R. Loos, "Algebraic simplification," Computer Algebra - Symbolic and
Algebraic Computation (B. Buchberger; G. Collins, R.Loos (eds.)), pp. 11-43, 1982.
[146] A. Blass and Yu Gurevich, "Equivalence relations, invariants and normal forms," SIAM J.
Computing, vol. l3, no. 4, pp. 682-689, 1984.
[147] E. Cardoza, R. Lipton, and A. R. Meyer, "Exponential space complete problems for petri
nets and commutative semi groups," Con! Record of the 8th Annual ACM Symp. Theory of
Computing, pp. 50--54, 1976.
[148] E. W. Mayr and A. R. Meyer, "The complexity of the word problems for commutative
semigroups and polynomial ideals," Advances in Math., vol. 46, no. 3, pp. 305-329, 1982.
[149] B. Buchberger, An Algorithm for Finding a Basis for the Residue Class Ring of a Zero-
dimensional Polynomial Ideal (Gennan), Ph.D. thesis, Univ. OfInnsbruck (Austria), Math.
Inst., 1965.
[150] B. Buchberger, "An algorithmical criterion for the solvability of algebraic systems of equa-
tions (German)," Aequationes Mathematicae, vol. 4, no. 3, pp. 374-383, 1970.
[151] B. Buchberger, "A theoretical basis for the reduction of polynomials to canonical form,"
ACM SIGSAM Bull., vol. 10, no. 3, pp. 19-29, 1976.
[152] A. Galligo, ''The divison theorem and stability in local analytic geometry (French)," E:xtrait
des Annales de l'lnstitut Fourier, Univ. of Grenoble, vol. 29, no. 2,1979.
[153] H. Hironaka, "Resolution of singularities of an algebraic variety over a field of characteristic

zero: I, II," Annals Matk, vol. 79, pp. 109-326, 1964.
[154] B. Buchberger, "Miscellaneous results on Griibner-bases for polynomial ideals n," Tech.
Rep., University of Delaware, Dept. Compo and Inform. Scie., 1983, Tech Rep. 83-1.
[155] D. E. Knuth and P. B. Bendix, "Simple word problems in universal algebras," Proc. Conf.
Computational Problems in Abstract Algebra, 1967, I. Leech (ed.).
[156] R. Llopis de Trias, "Canonical forms for residue classes of polynomial ideals and term
rewriting systems," Tech. Rep., Univ. Autonoma de Madrid, Division de Matematicas,
1983, Rep. 84-03, Univ. Simon Bolivar, Caracas, Venezuela.
[157] P. Le Chenadec, Canonical Forms in Finitely Presented Algebras (French), Ph.D. thesis,
Univ. of Paris-Sud., Centre d'Orsay, 1983.
[158] I. Hsiang, Topics in Automated Theorem Proving and Program Generation, Ph.D. thesis,
Univ. of lllinois at Urbana-Champaign, Dept. of Compo Scie., 1982.
[159] L. Bachmair and B. Buchberger, "A simplified proof of the characterization theorem for
Grobner bases," ACM SIGSAM Bull., vol. 14, no. 4, pp. 29-34,1980.
[160] B. Buchberger, "A critical-pair/completion algorithm for finitely generated ideals in rings,"
in Proc. Conf. Rekursive Kombinatorik, E. Borger, G. Hasenjager, and D. ROdding, Eds.,
Miinster, May 1983, Lecture Notes in Computer Science 171, pp. 137-161, Springer-Verlag.
[161] G. M. Bergman, ''The diamond lemma for ring theory," Advances in Math., vol. 29, pp.
178-218, 1978.
[162] B. L. Van der Waerden, Modem Algebra: 1,1/, Frederick Ungar, New York, 1953.
[163] L. E. Dickson, "Finiteness of the odd perfect and primitve abundant numbers with n distinct
prime factors," Am. J. Math., vol. 35, pp. 413-426, 1913.
[164] R. Loos, "Generalized polynomial remainder sequences," in Computer Algebra - Symbolic

and Algebraic Computation, B. Buchberger, R. Loos, and G. E. Collins, Eds. 1983, pp.
115-138, Springer, Wien - New York.
[165] B. Buchberger, "Some properties of Griibner bases for polynomial ideals," ACM SIGSAM
Bull., vol. 10, no. 4, pp. 19-24, 1976.
[166] B. Buchberger, "A criterion for detecting unnecessary reductions in the construction of
Grobner bases," Proc. EUROSAM 79, Marseille, Lecture Notes in Computer Science, vol.
72, pp. 3-21, 1979, W. Ng (ed.).
[167] B. Buchberger and F. Winkler, "Miscellaneous results on the construction of Griibner bases
for polynomial ideals I," Tech. Rep., Tech. Rep. No. 137, University of Linz, Math. Inst.,
1979.
REFERENCES 253
[168] M. Lauer, "Canonical representatives for the residue classes of a polynomial ideal," Proc.
1976 ACM Symp. Symbolic and Algebraic Computations, Yorktown Heights, N. Y., pp. 339-
345, August 1976, R. D. Jenks (ed.).
[169] A. M. Ballantyne and D. S. Lankford, "New decision algorithms for finitely presented
commutative semigroups," Computers and Matk with Applications, vol. 7, pp. 159-165,
1981.
[170] B. E Caviness and R. Fateman, "Simplification of radical expressions," Proc. 1976 ACM
Symp. Symbolic and Algebraic Computations, Yorktown Heights, N. Y., pp. 329-338, August
1976, R. D. Jenks (ed.).
[171] R. Gebauer and H. Kredel, "Buchberger's algorithm for constructing canonical bases
(Grobner bases) for polynomial ideals," Program documentation, Univ. of Heidelberg, Dept.
of Applied Math., 1983.
[172] W. Trinks, "On B. Buchberger's method for solving systems of algebraic equations," 1.
Number Theory, vol. 10, no. 4, pp. 475-488, 1978.
[173] W. Grobner, Modem Algebraic Geometry (German), Springer, Wien-Innsbruck, 1949.
[174] E Mora and H. M. Moller, "The computation of the Hilbert function," Proc. EUROCAL
83, London, Lecture Notes in Computer Science 162, Springer, pp. 157-167, March 1983,
H. van Hulzen (ed.).
[175] H. M. E Mora and Moller, "New constructive methods in classical ideal theory," 1. Algebra,
vol. 100, pp. 138-178, 1986.
[176] G. E. Collins and L. E. Heindel, "The SAC-1 polynomial real zero system," Tech. Rep.,
Univ. of Wisconsin-Madison, Compo Scie. Dept., 1970, Techn. Rep. No. 93 .
. [177] D. Spear, "A constructive approach to commutative ring theory," Proc. MACSYMA Users
Con!, Berkeley, pp. 369-376, July 1977, R. J. Fateman (ed.), published by MIT.
[178] G. Zacharias, "Generalized Griibner bases in commutative polynomial rings," Bachelor's

Thesis, MIT, Dept. ofComp. Scie., 1978.
[179] S. Schaller, Algorithmic Aspects of Polynomial Residue Class Rings, Ph.D. thesis, Univ. of
Wisconsin-Madison, Compo Scie. Dept., 1979, Tech. Rep. No. 370.
[180] D. Bayer, The Division Algorithm and the Hilbert Scheme, Ph.D. thesis, Harvard University,
Cambridge, Mass., Math. Dept., 1982.
[181] c. W.Ayoub, "On constructing bases for ideals in polynomial rings over integers," 1.
Number Theory, vol. 17, pp. 204-225, 1983.
[182] M. Lauer, "Canonical representatives for the residue classes of a polynomial ideal (German),"
Diploma Thesis, University of Kaiserslautern (ER.G.), Dept. of Mathematics, 1976.
[183] R. Schrader, "Contributions to constructive ideal theory (German)," Diploma Thesis, Uni-
versity of Karlsruhe (ER.G.), Math. Inst., 1976.
[184] H. M. Moller, "Multi-dimensional Hermite interpolation and numerical integration (Ger-

man)," Matk ZeitschriJt, vol. 148, pp. 107-118, 1976.
[185] G. Bauer, The Representations of Monoids by Confluent Rule Systems, Ph.D. thesis, Uni-
versity of Kaiserslautern (FRG), Dept. of Compo Scie., 1981.
[186] E Winkler, "Implementation of an algorithm for constructing Grobner bases (German),"

Diploma Thesis, University of Linz (Austria), Dept. of Math., 1978.
[187] B. Buchberger, "A note on the complexity of constructing Grobner bases," Proc. of the
EUROCAL 83, London, Lecture Notes in Computer Science 162, Springer, pp. 137-145,
March 1983, H. van Hulzen (ed.).
[188] D. Lazard, "Grobner bases, Gaussian elimination, and resolution of systems of algebraic
equations," Proc. EUROCAL83, London, Lecture Notes in Computer Science 162, Springer,
pp. 146-156, March 1983, H. van Hulzen (ed.).
[189] J. Fitch, Ed., Proc. of the EUROSAM 84 Symp., Cambridge. Springer Lecture Notes in
Computer Science 174, 1984.
[190] Proc. of the EUROCAL85 Symposium, Linz, April 1985.
[191] M. Kreuzer and L. Robbiano, Computational Commutative Algebra I, Springer-Verlag,

Heidelberg - New York, 2000.
[192] B. Buchberger and E Winkler, Eds., Grobner Bases and Applications, vol. 251 of London
Mathematical Society Lecture Notes Series, Cambridge, 1998. Cambridge University Press,
Proc. IntI. Conf. "33 Years of Grobner Bases".
[193] Z. Lin and L. Xu, Eds., Applications of Grobner Bases in Multidimensional Systems and
Signal Processing, vol. 12 of Special Issue of the J. Multidimensional Systems and Signal
Process., October 2001.
[194] E. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, Springer-
Verlag, New York, 1995.
[195] R. Levy, 2-D Polynomial and Rational Matrices, and Their Applicationsfor the Modeling of
2-D Dynamical Systems, Ph.D. dissertation, Dept. Electrical Engineering, Stanford Univ.,
Stanford, California, 1981.
[196] H.H. Rosenbrock, State-Space and Multivariable Theory, John Wiley & Sons Inc., New
York, NY, 1970.
[197] M. Auslander and M. Bridger, Stable Module Theory, Number 94 in Memoirs Amer. Math.
Soc. Amer. Math. Society, Providence, R.I., 1969.
[198] J.E Pommaret, Partial Differential Control Theory: Mathematical Tools (Vol. I) and Control
Systems (Vol. II), Kluwer Academic Publishers, Dordrecht, The Netherlands, 2001.
[199] E. G. Evans and P. Griffith, Syzygies, Cambridge University Press, Cambridge, 1985.
[200] Jeffrey Wood, Eric Rogers, and David H. Owens, "A formal theory of matrix primeness,"
Math. Control Signals and Systems, vol. 11, pp. 40-78, 1998.
[201] P. A. Weiner, Multidimensional Convolutional Codes, Ph.D. thesis, University of Notre

Dame, Dept. of Mathematics, 1998.
[202] V. R. Sule, "Feedback stabilization over COrrlmutative rings: the matrix case," SIAM J.
Control and Optimization, vol. 32, pp. 1675-1695, 1994.
REFERENCES 255
[203] Z. Lin, "On matrix fraction descriptions of multivariable linear n-D systems," IEEE Trans.
Circuits and Systems: I, vol. 35, pp. 1317-1322, 1988.
[204] D. A. Cox, J. Little, and D. O'Shea, Using Algebraic Geometry, Springer, New York, 1998.
[205] G.M. Greuel, G. Pfister, and H. Schoenemann, Singular version 1.2 User Manual, Centre
for Computer Algebra, University of Kaiserslautern, Germany, June 1998.
[206] C. Charoenlarpnopparut and N. K. Bose, "Multidimensional FIR filter bank design using
Grobner bases," IEEE Trans. Circuits and Systems: I, pp. 1475-1486, 1999.
[207] Z. Lin, "Notes on n-D polynomial matrix factorization," Multidimensional Systems and
Signal Process., vol. 10, pp. 379-393, 1999.
[208] Z. Lin and N. K. Bose, "A generalization of Serre's conjecture and some related issues,"
Linear Algebra and Its Applications, vol. 338, pp. 128--138,2001.
[209] W. W. Adams and P. Loustanou, An Introduction to Grabner Bases, vol. 3, Amer. Math.
Society, Providence, R.I., 1994.
[210] A Quadrat, "Extended Bezout identities," Pmc. European Control Conf, pp. 1673-1678,
September 2001, Portugal.
[211] A. Logar and B. Sturmfels, "Algorithms for the Quillen-Suslin theorem," J. Algebra, vol.
145, pp. 231-239, 1992.
[212] Z. Lin, "On syzygy modules for polynomial matrices," Linear Algebra and Its Applications,
vol. 298, pp. 73-86,1999.
[213] J. F. Pommaret, "Solving Bose conjecture on linear multidimensional systems," Pmc.

European Control Conf, pp. 1853-1855, September 2001.
[214] H. Park and C. Woodburn, "An algorithmic proof of Suslin's stability theorem forpolynomial
rings," J. Algebra, vol. 178, pp. 217-298, 1995.
[215] P. M. Cohn, "On the structure of GL 2 of a ring," Inst. Hautes Etudes Sci. Publ. Math., vol.
30,pp.365-413,1966.
[216] H. Park, "A realization algorithm for SL2(R[Xl,·, Xn]) over the Euclidean domain," SIAM
J. Matrix Anal. and Applications, vol. 21, pp. 178-184, 1999.
[217] Z. Lin, "Feedback stabilizability of MIMO n-D linear systems," Multidimensional Systems
and Signal Process., vol. 9, pp. 149-172, 1998.
[218] M. C. Smith, "On stabilization and the existence of coprime factorizations," IEEE Trans.
Automatic Control, vol. 34, no. 9, pp. 1005-1007, September 1989.
[219] Y. Lai and C. T. Chen, "Coprime fraction computation of 2-D rational matrices," IEEE
Trans. Automatic Control, vol. 32, no. 4, pp. 333-336, April 1987.
[220] J. Agler and J. E. MacCarthy, "Nevanlinna-Pick interpolation for the bidisk," J. Reine
Angewendte Mathematik, vol. 506, pp. 385-407, 1999.
[221] J. A. Ball and M. Rakowski, "Unitary colligations, reproducing kernel Hilbert spaces, and
Nevanlinna-Pick interpolation in several variables," J. Functional Analysis, vol. 157, pp.
1-61, 1998.
[222] J. W. Helton, "Some adaptive control problems which convert to a "classical problem in
several complex variables" ," IEEE Trans. Automatic Control, vol. 26, no. 12, pp. 2038~2043,
December 2001.
[223] Z. Lin, ''Feedback stabilization of MIMO 3-D linear systems," IEEE Trans. Automatic
Control, vol. 44, no. 10, pp. 1950-1955, 1999.
[224] Z. Lin, "Feedback stabilization of n-D MIMO linear systems," IEEE Trans. Automatic
Control, vol. 45, no. 12, pp. 2419~2424, 2000.
[225] Z. Lin, ''Output feedback stabilizability and stabilization of linear n-D systems," in Multi-
dimensional Signals, Circuits and Systems, K. Galkowski & J. Wood, Ed., London, U. K.,
2001, pp. 59~76, Taylor & Francis.
[226] K. Mori and K. Abe, ''Feedback stabilization over commutative rings: further study of
coordinate-free approach," SIAM 1. Control and Optimization, vol. 39, no. 6, pp. 1952~
1973,2001.
[227] K. Mori, "Parametrization of stabilizing controllers over commutative rings with application
to multidimensional systems," IEEE Trans. Circuits and Systems : I, vol. 49, no. 6, pp. 743~
752,2002.
[228] Z. Lin, J. Q. Ying, and L. Xu, "An algebraic approach to strong stabilizability of linear n-D
MIMO systems," IEEE Trans. Automatic Control, vol. 47, no. 9, pp. 1510-1514,2002.
[229] J.W. Polderman and J.C. Willems, Introduction to Mathematical Systems Theory, Springer-
Verlag, BerlinlHeidelberg, 1998.
[230] H. Pillai and S. Shankar, "A behavioral approach to control of distributed systems," SIAM
1. Control and Optimization, vol. 37, no. 2, pp. 388-408, 1998.
[231] S. Shankar, 'The Nullstellensatz for systems of PDE," Advances in Applied Math., vol. 23,
pp.360-374,1999.
[232] A. J. Sasane, E. G. F. Thomas, and J. C. Willems, ''Time autonomy versus time controlla-
bility," Systems and Control Letters, vol. 45, pp. 145~153, 2002.
[233] M. E. Valcher and E. Fornasini, "On 2-D finite support convolutional codes:an algebraic
approach," Multidimensional Systems and Signal Process., vol. 5, pp. 231~243, 1994.
[234] Ph. Piret, Convolutional Codes: An Algebraic Approach, The MIT Press, Cambridge, MA,
1988.
[235] Bruce Kitchens, ''Multidimensional convolutional codes," SIAM 1. Discrete Math., vol. 15,
no.3,pp. 367~381,2002.
[236] H. Gluesing-Luerssen, J. Rosenthal, and P. A. Weiner, ''Duality between multidimensional

convolutional codes and systems," in Advances in Math. Systems Theory, F. Colonius,
U. Helmke, D. Proetzler-Wolters, and F. Wirth, Eds., Boston, 2001, pp. 135-150, Birkhauser.
[237] B. Ross Barmish, New Tools for Robustness of Linear System, MacMillan Publishing
Company, New York, 1994.
[238] S. P. Bhattacharyya, H. Chappellat, and L. H. Keel, Robust Control: The Parametric Ap-
proach, Prentice Hall PTR, Upper Saddle River, NJ, 1995.
REFERENCES 257
[239] J. Kogan, Robust Stability and Convexity, Springer-Verlag, London, Great Britain, 1995.
[240] V. L. Kharitonov, "On a generalization of a stability criterion," (in Russian), Izv. Akad.
Nauk. Kazakk SSR Ser. Fiz Mat., , no. 1, pp. 53-57, 1978.
[241] A. C. Bartlett, C. V. Hol1ot, and L. Huang, "Root locations of an entire polytope of poly-
nomials: it suffices to check the edges," Math. Control, Signals, and Systems, vol. 1, pp.
61-71, 1988.
[242] N. K. Bose and Y. Q. Shi, "A simple general proof of Kharitonov's generalized stability
criterion," IEEE Trans. Circuits and Systems: I, vol. 34, pp. 1233-1237, October 1987.
[243] N. K. Bose, "Robust scattering multivariate Hurwitz polynomials," Linear Algebra and Its
Applications, vol. 98, pp. 123-136, 1988.
[244] A. Fettweis, "A new approach to Hurwitz polynomials in several variables," Circuits,
Systems and Signal Process., vol. 5, no. 4, pp. 405-417, 1986.
[245] S. Basu, "On the multidimensional generalization of robustness of scattering Hurwitz prop-
erty of complex polynomials," IEEE Trans. Circuits and Systems: I, pp. 1159-1169, Septem-
ber 1989.
[246] S. Basu, "On boundary implication of stability and positivity properties of multidimensional
systems," Proc. IEEE, vol. 78, no. 4, pp. 614-626, April 1990.
[247] V. L. Kharitonov and J. A. Torres Munoz, "Robust stability of multivariate polynomials.

Part 1: Small coefficient variations," Multidimensional Systems and Signal Process., vol.
10, no. l,pp. 7-20,1999.
[248] V. L. Kharitonov, J. A. Torres Munoz, and M. I. Ramfrez-Sosa, "Robust stability of mul-

tivariate polynomials. Part 2: Polytopic coefficient variations," Multidimensional Systems
and Signal Process., vol. 10, no. 1, pp. 21-32, 1999.
[249] Y. Q. Shi and S. F. Zhou, "Stability of a set of multivariate complex polynomials with
coefficients varying in diamond domain," IEEE Trans. Circuits and Systems: I, vol. 39, no.
8,pp.683-688,1992.
[250] N. K. Bose, "Edge property from end-points for scattering Hurwitz polynomials," Auto-
matica, vol. 32, no. 4, pp. 655-657, 1996.
[251] V. L. Kharitonov and 1. A. Torres Munoz, "Stability of multivariate polynomials. Part 4:

Conic sets," Multidimensional Systems and Signal Process., vol. 14, no. 4, pp. 343-363,
2003.
[252] V. L. Kharitonov, M. I. Ramfrez-Sosa, and 1. A. Torres Munoz, "Stability of multivariate

polynomials. Part 3: Frequency domain approach," Multidimensional Systems and Signal
Process., vol. 11, no. 3, pp. 213-231, 2000.
[253] B. T. PolyakandS. B. Shmulyian, "Frequency domain criteria for robust stability of bivariate
polynomials," IEEE Trans. Circuits and Systems: I, vol. 41, no. 2, pp. 161-167, February
1994.
[254] N. K. Bose and K. D. Kim, "Boundary implications for frequency response FIR and IIR
filters," IEEE Trans. Acoust., Speech, and Signal Process., vol. 39, pp. 2167-2173, 1991.
[255] N. K. Bose and C. Yang, "Robust wavenumber response of multidimensional FIR filters,"
in Advances in Control: Highlights of ECC'99, P. M. Frank:, Ed., London, Great Britain,
1999, pp. 285--292, Springer-Verlag London Limited.
[256] A. Levkovich, E. Zeheb, and N. Cohen, "Frequency response envelopes of a family of

uncertain continuous-time systems," IEEE Trans. Circuits and Systems: I, vol. 42, pp.
156-165, 1995.
[257] N. Cohen, A. Levkovich, P. de Oliveira, and E. Zeheb, ''Frequency response envelopes of a

family of uncertain discrete-time systems," Circuits, Systems and Signal Process., vol. 22,
pp. 19--41,2003.
[258] J. Garloff and N. K. Bose, "Boundary implications for stability properties: present status,"
in Reliability in Computing: The Role of Interval Methods in Scientific Computing, R. E.
Moore, Ed., San Diego, CA, 1988, pp. 391--402, Academic Press, Inc.
[259] B. Ross Barmish, M. Fu, and S. Saleh, "Stability of a polytope of matrices: Counterexam-
ples," IEEE Trans. Automatic Control, vol. 33, no. 6, pp. 569-572, 1988.
[260] J. Rohn, ''Positive definiteness and stability of interval matrices," SIAM 1. Matrix Anal. and
Applications, vol. 15, no. 1, pp. 175-184, 1994.
[261] R. E. Yedavalli, "It suffices to check only two special vertex matrices in Kronecker space to
analze the robust stability of an interval matrix," in Proc. Amer. Control Con!, Anchorage,
Alaska, May 2002, pp. 1266-1271.
[262] C. A. Berenstein and E. V. Patrick, "Exact deconvolution for multiple convolution operator
- an overview, plus performance characterization for imaging sensors," Proc. IEEE, vol. 78,
no.4,pp.723-734,199O.
[263] S. P. Kim, N. K. Bose, and H. M. Valenzuela, ''Recursive reconstruction of high resolution

image from noisy undersampled multiframes," IEEE Trans. Acoust.• Speech. and Signal
Process., vol. 38, pp. 1013-1027, 1990.
[264] N. K. Bose, H. C. Kim, and H. M. Valenzuela, ''Recursive total least squares algorithm
for image reconstruction from noisy, undersampled frames," Multidimensional Systems and
Signal Process., vol. 4, no. 3, pp. 253-268, July 1993.
[265] Michal Irani and Shmuel Peleg, "Improving resolution by image registration," CVGIP:
Graphical Models and Image Process., vol. 53, no. 3, pp. 231-239, May 1991.
[266] Michal Irani and Shmuel Peleg, "Motion analysis for image enhancement: Resolution,
occlusion, and transparency," 1. Visual Comm. and Image Representation, vol. 4, pp. 324-
335, December 1993.
[267] N. K. Bose and K. J. Boo, "High-resolution image reconstruction with multi sensors," Int.
1. Imaging Systems and Tech, vol. 9, pp. 294--304, 1998.
[268] Michael K. Ng, Wilson C. Kwan, and Raymond H. Chan, "A fast algorithm for high-
resolution color image reconstruction with multi sensors," in Lecture Notes in Computer
Science, P. Yalamov L. Vulkov, J. Wasniewski, Ed., New York, April 2001, vol. 1988, pp.
615--627, Springer-Verlag.
[269] S. J. Reeves, ''Optimal space-varying regularization in iterative image restoration," IEEE

Trans. Image Process., vol. 3, no. 3, pp. 319-324, May. 1994.
REFERENCES 259
[270] M. R. Banham and A. K. Katsaggelos, "Digital image restoration," IEEE Signal Process.
Magazine, vol. 14, no. 2, pp. 24-41, Mar. 1997.
[271] S. Chardon, B. Vozel, and K. Chehdi, "Parametric blur estimation using the generalized
cross-validation criterion and smoothness constraint on the image," Multidimensional Sys-
tems and Signal Process., vol. 10, no. 4, pp. 395-414, October 1999.
[272] Per Christian Hansen, Rank-Deficient and Discrete Ill-Posed Problems: Nwnerical Aspects
of Linear Inversion, SIAM, Philadelphia, PA, 1997.
[273] M. E. Kilmer and D. P. O'Leary, "Choosing regularization parameters in interative methods

for ill-posed problems," SIAM J. Matrix Anal. and Applications, vol. 22, pp. 1204-1221,
2001.
[274] Per Christian Hansen, "Analysis of discrete ill-posed problems by means of the L-curve,"
SIAM Review, vol. 34, no. 4, pp. 561-580, December 1992.
[275] N. K. Bose, S. Lertrattanapanich, and laehoon Koo, "Advances in superresolution using

L-curve," Proc. IEEE Int. Symp. Circuits and Systems, vol. II, pp. 433-436, May 6-9, 2001.
[276] M. Hanke, ''Limitations of the L-curve method in ill-posed problems," BIT, vol. 36:2, pp.
287-301, 1996.
[277] C. R. Vogel, "Non-convergence of the L-curve regularization parameter selection method,"

Inverse Problems, vol. 12, pp. 535-547, 1996.
[278] Michael K. Ng,laehoon Koo, andN. K. Bose, "Constrained total least-squares computations
for high-resolution image reconstruction with multisensors," Int. J. Imaging Systems and
Tech, vol. 12, no. 1, pp. 35-42, 2002.
[279] Wim Sweldens and Peter Schroder, "Building your own wavelets at home," Wavelets in
Computer Graphics ACM SIGGRAPH Course notes, pp. 15-87, 1996.
[280] Wim Sweldens, ''The lifting scheme: A construction of second generation wavelets," SIAM
J. Mathematical Anal., vol. 29, no. 2, pp. 511-546, March 1998.
[281] G. Deslauriers and S. Dubuc, "Interpolation dyadique," in Fractal, dimensions non entieres
et applications, pp. 44-55, 1987.
[282] D. L. Donoho, "Smooth wavelet decomposition with blocky coefficient kernels," in Recent
Advances in Wavelet Analysis, L. L. Schumaker and G. Webb, eds., Academic Press, New
York, pp. 259-308, 1993.
[283] Ingrid Daubechies and Wim Sweldens, "Factoring wavelet transforms into lifting steps," J.
Fourier Anal. Appl., vol. 4, no. 3, pp. 245-267, 1998.
[284] Steve Mann and Rosalind W. Picard, "Video orbits of the projective group: A simple
approach to featureless estimation of parameters," IEEE Trans. Image Process., vol. 6, no.
9, pp. 1281-1295, September 1997.
[285] S. Lertrattanapanich and N. K. Bose, ''Latest results on high-resolution reconstruction from

video sequences," Technical Rep. IEICE, DSP99-140, The Inst. Electronic, Info. and Comm.
Engineers, Japan, pp. 59-65, December 1999.
[286] E. Fomasini, "A 2-D systems approach to river pollution modelling," Multidimensional
Systems and Signal Process., vol. 2, no. 3, pp. 233-265, 1991.
[287] A. Fettweis, "Wave digital filters: theory and practice," Proc. IEEE, vol. 74, pp. 270--327,
1986.
[288] A. Fettweis, "Discrete passive modelling of physical systems described by partial differential
equations," in Multivariate Analysis: Future Directions, C. R. Rao, Ed. 1993, pp. 115-131,
North-Holland Elsevier Science Publishers.
[289] A. Fettweis, "Transformation approach to numerically integrating PDEs by means of WDF

principles," Multidimensional Systems and Signal Process., vol. 2, pp. 127-159, 1991.
[290] M. Fries, "Multidimensional reactive elements on curvilinear coordinate systems and their
MDWDF generalizations," Proc. IEEE Int. Con! Acoust., Speech, and Signal Process., vol.
6,pp. V1-9-V1-12, 1994.
[291] S. Bilbao, Wave and Scattering Methods for the Numerical Integration ofPartial Differential
Equations, Ph.D. thesis, Stanford University, Department of Electrical Engineering, 2001.
[292] A. Fettweis and N. K. Bose, "A property of Jacobian matrices and some of its consequences,"
IEEE Trans. Circuits and Systems: I, vol. 50, no. 1, pp. 153-155, 2003.
[293] N. K. Bose and A. Fettweis, "Skew-symmetry in the equivalent representation problem of a

time-varying multiport inductor," Proc. IEEE Int. Symp. Circuits and Systems, vol. III, pp.
662---665, 2003.
[294] N. K. Bose and A. Fettweis, "Isomorphism between skew-symmetric and orthogonal ma-
trices via generalized Cayley transform," Proc. European Con! Circuit Theory and Design
(ECCTD 2(03),2003.
[295] J. Kogan Editor, ," in Special Issue: Robustness of Multidimensional Systems, Dordrecht,
The Netherlands, October 1994, vol. 5, Kluwer Academic Publishers.
[296] R. H. Chan, T.E Chan, L.X. Shen, and Z.W. Shen, "Wavelet algorithms for high resolution
image reconstruction," SIAM 1. Scientific Computing, vol. 24, pp. 1408-1432,2003.
-
[297] R. H. Chan, T.E Chan, L.X. Shen, and Z.W. Shen, "Wavelet deblurring algorithms for
spatially varying blur from high resolution image reconstruction," Linear Algebra and Its
Applications, vol. 366, pp. 139-155,2003.
[298] G. E. Baker Jr. and Peter Graves-Morris, Pade Approximants Part I: Basic Theory Encyclo-
pedia of Mathematics and Its Applications, vol. 13, Addison-Wesley Publ. Co., Reading,
MA,1981.
[299] G. E. Baker Jr. and Peter Graves-Morris, Pade Approximants Part 2: Extensions and Appli-
cations, Encyclopedia of Mathematics and Its Applications, vol. 14, Addison-Wesley Publ.
Co., Reading, MA, 1981.
[300] C. Brezinski, Pade-Type Approximation and General Orthogonal Polynomials, ISNM50,

Birkhiiuser Verlag, Basel, 1980, (Section 4.2).
[301] J. S. R. Chisholm, "Generalizations of Pade approximants," in Special Issue: Rational

Approximations for Systems, A. Bultheel and P. Dewilde, Eds., Birkhiiuser Boston Inc.,
1982, vol. 1, pp. 279-287, Circuits, Systems and Signal Process.
[302] B. W. Dickinson, M. Morf, and 1'. Kailath, "A minimal realization algorithm for matrix
sequences," IEEE Trans. Automatic Control, vol. 10, pp. 31-38, February 1974.
REFERENCES 261
[303] S. Basu and N. K. Bose, "Matrix Stieltjes series and network models," SIAM J. Math. Anal.,
vol. 14, pp. 209-222, March 1983.
[304] R. Hastings James and S. K. Mehra, "Extensions of the Pade-approximation technique for
the design of recursive digital filters," IEEE Trans. Acoust., Speech, and Signal Process.,
vol. 25, pp. 501-509, December 1977.
[305] C. S. Burrus and W. Parks, "Time domain design of recursive digital filters," IEEE Trans.
Audio and Electroacoustics, vol. 18, pp. 137-141, June 1970.
[306] Y. Shamash, "Stable reduced order model using Pade-type approximation," IEEE Trans.
Automatic Control, vol. 19, pp. 615--616, October 1974.
[307] Y. Shamash, "Model reduction using the routh stability criterion and the Pade approximation
technique," Int. J. Control, vol. 21, pp. 475--484, 1975.
[308] A. Cichocki, "Nested-feedback-loops realization of 2-D systems," Circuits, Systems and

Signal Process., vol. 1, no. 3-4, pp. 321-343,1982.
[309] W. Siemaszko, "Branched continued fraction for double power series," J. Compo Appl.
Math., , no. 2, pp. 121-125, 1980.
[310] P. N. Paraskevopoulos, "Pade-type order reduction of two-dimensional systems," IEEE

Trans. Circuits and Systems: I, vol. 27, pp. 413--416, 1980.
[311] M. P. Ekstrom and J. W. Woods, "Two-dimensional spectral factorization with applications

in recursive digital filtering," IEEE Trans. Acoust., Speech, and Signal Process., vol. 24, pp.
115-128, April 1976.
[312] L. F. Chaparro and E. I. Jury, "Rational approximation of 2-D linear discrete systems," IEEE
Trans. Acoust., Speech, and Signal Process., vol. 24, pp. 115-128, April 1976.
[313] I. S. R. Chisholm, "Rational approximation defined from double power series," Mathematics
of Computation, vol. 27, no. 124, pp. 841-848, October 1973.
[314] J. S. R. Chisholm and J. McEwan, "Rational approximations defined from power series in
n-variables," Proc. Royal Soc. London A, vol. 336, pp. 421--452, 1974.
[315] R. Hughes Jones and G. J. Makinson, 'The generation of chisholm rational polynomial
approximants in power series in two variables," J. Inst. Math. Applications, vol. 13, pp.
299-310, 1974.
[316] R. Hughes Jones, "General rational approximants in n-variables," J. Approx. Theory, vol.
16,pp.201-233,1976.
[317] D. Levin, "General order Pade-type rational approximants defined from double power
series," J. Inst. Maths. Applications, vol. 18, pp. 1-8, 1976.
[318] C. H. Lutterodt, "Rational approximants to holomorphic functions in n-dimensions," J.

Math. Anal. and Applications, vol. 53, pp. 89-98, 1976.
[319] C. H. Lutterodt, "A two-dimensional analogue ofPade approximant theory," J. Phys. Math,
vol. 7,pp. 1027-1037, 1974.
[320] Annie A. M. Cuyt, Pade Approximants for Operators: Theory and Applications, Springer-
Verlag, Heidelberg, Germany, 1984.
[321] J. Karlsson and H. Wallin, "Rational approximation by an interpolation procedure in several

variables," in Pade and Rational Approximation: Theory and Applications, E. B. Saff and
R. S. Varga, Eds., London, 1977, Academic Press.
[322] P. R. Graves-Morris, R. Hughes, and G. J. Makinson, 'The calculation of rational approxi-

mants in two variables," 1. Inst. Maths. Applics., vol. 13, pp. 299-310, 1974.
[323] P. Hillion, "Remarks on rational approximations of multiple power series," 1. Inst. Math
Applications, vol. 19, pp. 281~293, 1977.
[324] Annie A. M. Cuyt, 'The f-algorithm and multivariate Pade approximants," Numerische
Mathematik, vol. 40, pp. 39-46, 1982.
[325] Annie A. M. Cuyt, "Abstract Pade approximants in operator theory," in Pade Approximation
and its Applications, L. Wuytack, Ed., Berlin, 1979, LNM 765, pp. 61~87, Springer-Verlag.
[326] N. K. Bose, "Two-dimensional rational approximants via one-dimensional Pade technique,"

in Proc. of European Signal Proc. Con!, Kunt and deCoulon, Eds., Lausanne, Switzerland,
1980,pp.409-411.
[327] D. Bessie, 'Topics in the theory of Pade approximants," in Padi Approximants, P. R.

Graves-Morris, Ed., London and Bristol, 1973, pp. 20-44, The Institute of Physics.
[328] N. K. Bose and S. Basu, 'Theory and recursive computation of 1-D matrix Pade approxi-
mants," IEEE Trans. Circuits and Systems: I, vol. 27, pp. 323~325, April 1980.
[329] M. F. Bamsley and P. D. Robinson, "Rational approximant bounds for a class of two-variable
Stieltjes functions," SIAM 1. Math. Anal., vol. 9, pp. 272~290, April 1978.
[330] N. K. Bose and S. Basu, "'I\vo-dimensional matrix Pade approximants: Existence,

nonuniqueness, and recursive computation," IEEE Trans. Automatic Control, vol. 25, pp.
509~514, June 1980.
[331] S. Basu and N. K. Bose, "Stability of 2-D matrix rational approximants from input data,"
IEEE Trans. Automatic Control, vol. 26, pp. 540-541, April 1981.
[332] J. H. Justice, "A Levinson-type algorithm for two-dimensional Wiener filtering using bi-
variate Szego polynomials," Proc. IEEE, vol. 65, pp. 582~586, June 1977.
[333] T. L. Marzetta80, "Two-dimensional linear prediction: Autocorrelation arrays, minimum-

phase prediction error filters and reflection coefficient arrays," IEEE Trans. Acoust., Speech,
and Signal Process., vol. 28, pp. 725~733, December 1980.
[334] D. Jackson, "Formal properties of orthogonal polynomials in two variables," Duke Math
1., vol. 2, pp. 423-434, 1936.
[335] J. S. Geronimo, "Scattering theory and matrix orthogonal polynomial on the real line,"
Circuits, Systems and Signal Process., vol. 1, no. 3-4, pp. 471-495, 1982.
[336] W. B. Gragg and G. D. Johnson, 'The Laurent-Pade table," Info. Process., vol. 74, pp.
632-637, 1974, Proc. IFIP Congress.
[337] P. Fitzpatrick and J. Flynn, "A Grobner basis technique for Pade approximation," 1. Symbolic
Computation, vol. 13, pp. 133~138, 1992.
REFERENCES 263
[338] Annie A. M. Cuyt, "A comparison of some multivariate Pade approximants," SIAM 1. Math.
Anal., vol. 41, pp. 195-202, 1983.
[339] Annie A. M. Cuyt, Shiro Ogawa, and Verdonk Brigitte, "A comparison of some multidi-
mensionallinear shift-invariant recursive systems using Pade techniques," Multidimensional
Systems and Signal Process., vol. 3, pp. 309-321, 1992.
[340] Annie A. M. Cuyt, Abouir Jilali, Drive Ramon, and Gonz31ez Pablo, "On the convergence
of general order multivariate Pade-type approximants," 1. Approx. Theory, vol. 86, pp.
216--228, 1996.
[341] Z. Lin and L. T. Bruton, "BmO stability of inverse 2-D digital filters in the presence of
nonessential singularities of the second kind," IEEE Trans. Circuits and Systems: I, vol.
36,pp.244-254,1989.
[342] L. Wang, D. Xiyu, and F. Xia, "Nonessential singularities of the second kind and stability
of multidimensional digital filters," Multidimensional Systems and Signal Process., vol. 3,
pp. 363-380, 1992.
[343] N. K. Bose and J. P. Guiver, "Multivariate polynomial positivity invariance under coefficient
perturbation," IEEE Trans. Acoust., Speech, and Signal Process., vol. 28, pp. 660-665,
December 1980.
[344] N. K. Bose and E. Zeheb, "Kharitonov's theorem and stability test of multidimensional
digital filters," Proc. lEE Circuits and Systems, Part G, vol. 133, pp. 187-190, 1986.
[345] S. Basu and A. Fettweis, "New results on stable multivariate polynomials - part II: discrete
case," IEEE Trans. Circuits and Systems: I, vol. 21, no. 11, pp. 1264-1274, 1987.
[346] B. D. O. Anderson and R. R. Bitmead, "Stability of matrix polynomials," Int. 1. Control,

vol. 26, pp. 235-247, 1977.
[347] H. Kokame and T. Mori, "A Kharitonov-like theorem for interval polynomial matrices,"
System and Control Letters, vol. 16, no. 2, pp. 107-116, February 1991.
[348] S. Basu, "Multidimensional filter banks and wavelets- a system-theoretic perspective," 1.

Franklin Institute, vol. 335B, pp. 1367-1409, 1998.
[349] C. Charoenlarpnopparut and N. K. Bose, "Grabner bases for problem solving in multidi-
mensional systems," Multidimensional Systems and Signal Process., vol. 12, no. 3/4, pp.
365-376,2001.
[350] R. S. Bucy, M. K. Narniri, and J. R. Velman, "Minimax control," Computers Math. Applic.,
vol. 19, no. 4, pp. 51-63, 1990.
[351] N. K. Bose and C. Charoenlarpnopparut, "Minimax controller design using rate feedback,"
Circuits, Systems and Signal Process., vol. 18, no. 1, pp. 17-25, 1999.
[352] N. K. Bose and C. Charoenlarpnopparut, "Minimax controller using rate feedback: latest
results," Proc. Int. Federation Automatic Control, vol. G-2e-20-4, pp. 525-530, 1999.
Index
Abelian group, 151 weakly, 37

absolute convergence, 4 causality cone, 39
additive group of integers, 25 characterization
admissible ordering, 92 theorem for GrObner bases, 94
algebraic Ching's condition, 85
equations, 91, 108, 114 Cholesky triangle, 218
extension, 109 Church-Rosser property, 94
triangularized equations, 11 0 closed convex cone, 39
Artinian ring, 129 closure
ascending chain condition, 129 reflexive, symmetric, transitive, 94
associative algebra, 125 codimension (height), 155
of multiplication table, 105 cokerne1, 152, 156
asymptotically critical, 13 complete behavior, 189
asymptotically stable i. o. d., 12 completely positive, 22
asymptotically stable i. o. d., 13
completion of bases, 90
automated program verification, 90
complexity of Grobner bases, 106, 125
congruent modulo an ideal, 91, 104, 121
backprojection kernel, 207
content, 133, 136
behavior, 184
convex directions, 196
behavioral approach, 184
convex polytope, 191
Bezout
convolutional code, 189
domain, 130, 179
free, 190
identity, 127, 179
coordinate-free approach, 235
identity extended, 167
coprime
left, 19
matrix, 135 factor, 130
right, 19 strongly, 130
bijective enumeration of polynomial ideals, 89 weakly, 130
bilinearly transformable, 11 zero, 130
bounded inputlbounded output (BmO) stability coprimeness conditions
criterion, 3, 4, 38, 231 I-D, 18
boxed domain, 191 coprimeness equation, 48
counterexample
canonical Bose, 28
form, 103 Bose-Zeheb, 236
representatives, 104 Charoenlarpnopparut's, 160
simplifier, 103, 104 Cohn's, 177
Canterbury grid, 223 Goodman's, 8
causal,35 Guiver-Bose's, 11
strictly, 37 Levy's, 151
strictly weakly, 37 Lin's, 168
265
Murray's, 8 GCD extraction

Northcott-Youla-Gnavi-Oberst, 1('f) from Sylvester Matrix, 133
criteria for reduction, 98, 101 from Bezout matrix, 135, 136
criticrupair, 122, 123 generru factorization, 147
criticru pairs, 90 generruized cross vruidation, 212,213
criticru-paircompletion rugorithms, 90, 99, 122 generruized factor left prime, 157, 181
generating set, 158
D-module, 186 generators
Daniel-Moore conjecture, 14 for linear homogeneous equation, 117, 118
Dautov's conjecture, 7, 8 for solutions of linear equations, 116
degenerate rank case, 174 of a commutative sernigroup, 105
delay-differentiru systems, 12 Grabner Bases
descending chain condition, 129 (basis or set), 89, 94
determinantal factorization, 239 rugorithmic construction, 95
determinantal factorization, 147 characterization theorem, 94
left and right, 239 for polynomiru iderus over integers, 120
problem, 174 improved algorithm, 98
deterrninantal factorization domain, 147 reduced, 97,98
diagonalizability by DCf, 208 running time, 125
Dickson's lemma, 96 Grobner basis, 52, 164
direct product, 86
discrete cosine transform matrix, 200 Hankel matrix, 200
division algorithm, 138 Hardy-Littlewood inequality, 5
double coprime factorization, 179 Hermite form, 148
Hermite ring, 171
Hilbert function, 110, 124
elementary factors, 182
Hilbert transform, 27
generruized, 183
Hilbert's basis theorem, 86
elementary matrix, 176
Hilbert's basis theorem, 90, 96, 129
elimination, 184
Hilbert's Nullstellensatz, 47, 59, 83, 84, 86, 87,
ideru, 108, 109
130,161,187,233
of variables, 127
holomorphic functions, 4
successive, 110
Hurwitz
elimination ideal, 109
strict, 192
equationru theorem proving, 90
Hurwitz polynomial
equations
narrow sense, 11
rugebraic, 89,91, 108, 114
principru, 9, 193
linear homogeneous, 89, 91,118
scattering, 9, 192, 193
triangularized, 11 0
strictest sense, 10, 195
Euclid's rugorithm, 97, 110, 125
widest sense, 10, 12, 195
Euclidean norm, 7
hyper-rectangle, 191
exact sequence, 152
exchange matrix, 211
ideru,86
extendible, 25
congruence, 89,94, 122
extension problem, 25 generated by polynomirus, 89
Lasker-Noether decomposition, 124
factor membership, 120
coprime, 167 principru, 124
factorleft coprime, 161 image, 152, 156
factor structure implementation of Grobner bases, 125
rugorithms for, 104 injective module homomorphism, 189
factorization innerwise matrix, 133
deterrninantal, 174 interval
primitive, 131, 133, 140, 144 matrix, 199
feedback stabilization, 233 matrix stability, 199
Fundamental Principle, 185 polynomial, 192
system, 191
Gauss' algorithm, 125 inverse filter, 202
GCD domain, 130 irreducible polynomirus, 26
INDEX 267
irreduicible representation, 41 modules
iterative backprojection, 206 R-module and vector space, 159
iterative learning control (ILC), 32 mollifier, 201, 202
moment problem, 23
Jacobian matrix, 217 trigonometric, 24
monic, 139
Kalman gain vector, 205 multidimensional
Karlsson-Wallin grid, 224 (vector)differential operator, 217
kernel, 152, 156 Kirchhoff network, 216
kernel representation, 156, 184, 185, 189 passivity, 217
Knuth-Bendix algorithm, 90, 94,98 wave digital principles, 216
multidimensional integration, 124
multipass processes, 15, 32
L-curve, 213
multiplicatively positive, 24
Lasker-Noether decomposition, 124
multistep-multiderivative
least common reducible, 122, 124
formula, 14
left coprimeness degree, 157
multivariate networks, 9
lifting scheme, 215
multivariate Pade-type approximants, 225
linear shift-invariant (LSI), 3, 35
localization, 80, 86
Neumann boundary condition, 210, 211
Lutterodt grid 1, 223
Nevanlinna-Pick interpolation, 181
Lutterodt grid 2, 223
Noetherian
full quotient ring, 84, 85
matrix fraction description, 179 relation, 93
I-D, 18 ring, 87
2-D MFD, 60, 180 Noetherian full quotient ring, 84
left (LMFD), 53 Noetherian ring, 16, 129
n-D MFD, 180 nonessential singularities of the first kind, 241
right (RMFD), 54 nonessential singularities of the second kind, 7, 231
matrix inversion formula, 204 normal form (reduced form), 93
Max Noether's fundamental theorem, 87 normal form algorithm, 93
maximal ideal, 83, 86 normal full rank, 147, 155, 157, 161
McCoy rank, 85
MlMO systems, 43 order of a zero, 5
minimal ordering
left annihilator, 156 admissible, 92
right annihilator, 156 purely lexicographical, 110
minimax controller, 242 total degree, 91, 101, 106, 108, 112, 118,
minimum phase function, 233, 241 126
minorleft coprime, 161
minor right prime, 170 para-even polynomial, 10
modeling para-odd polynomial, 10
projective, 215 paraconjugate polynomial, 10
river pollution, 216 parametrizable, 186
selfpurification, 216 parametrization of stabilizing compensators, 61, 63
module, 151 Petri nets
A-homomorphisms, 152 reversible, 89, 106
A-module, 152 phase function, 196
D-module, 154 Piola's identity, 217
dual, 153 polydisc
free, 152 algebra, 31
homomorphism, 152 closed unit, 4
projective, 152 distinguished boundary(torus), 4
R-module, 151 open unit, 4
rank, 153 polyhedral set, 191
reflexive, 153 polynomial
syzyg~ 152, 162, 164, 166 homogeneous (form), 194
torsion, 152, 154 paraconjugate, 193
torsion -free, 152 reactance Hurwitz, 196
self-paraconjugate, 195 reduced GrObner basis, 97

strictly Hurwitz matrix, 237 reduced GrObner basis algorithm, 99
strictly Schur, 236 reduced minors, 158, 159, 173
polynomial ideal theory, 94 reduced vertex result, 193
positive functional, 22 regularization
positive real operator, 212
function, 21 parameter, 212
function integral representation, 21 parameter selection, 213
rnatrix,20 space-variant, 212
positive real function Tikhonov, 212
integral representation, 20 regularization methods, 211
positive semidefinite, 33 relative primeness, 131
prime tests, 131
minor, 155 residue class ring, 89, 104, 107
minor left, 155 residue classes of polynomials, 87
weakly zero, 155 resolution
weakly zero left, 155 free, 120, 153
zero, 155 projective, 153
zero left, 155 spatial, 200
primeness, 155 temporal, 200
degree, 155 resultant, 134
degree left, 156 right coprimeness degree, 157
degree right, 157 right inverse, 161
primitive, 136 ring
primitive factorization differential, 154, 185
bivariate polynomial matrix, 140 ring isomorphism, 3, 40, 65, 74, 76
for multivariate polynomials, 133 ring of stable multivariate rational functions, 241
theorem, 140 robust wavenumber response, 198
primitive matrix, 138 Rudin's theorem, 9
primitive part, 136 running time for GrObner bases, 125
primitive polynomials, 27, 133
principal ideal domain, 130 S-polynomials, 90,91, 95,97,98, 121
projective dimension, 153 sernigroup, 39
isomorphism, 40
projective model, 215
sernigroups
projectively free, 19
pseudorandom sequences, 27 finitely generated commutative, 89, 106, 125
separated variables, 11 0
purely lexicographical ordering, 92,108,109, Ill,
Serre's conjecture, 170
114, 123, 126
generalized, 172, 173
simplification
quasipolynomial, 13 algorithm, 103
Quillen-Suslin, 19 canonical, 103
Quillen-Suslin theorem, 170, 241 of radical expressions, 107
simplifier, 93
radical ideal, 86, 130 solvability and exact solution of systems of alge-
reachability problem braic equations, 108
for reversible Petri nets, 106 spectral factorization, 27, 232
realizable, 177 split map, 17, 20
realization stability
absolutely minimal, 17 Billa, 38
canonical, 17 matrix polynomials, 237
minimal, 17 structural, 43
multivariate, 15 stabilizability, 182
state-space, 16, 32 stabilizable, 45, 46, 49, 59, 60,65, 66,71, 73, 76,
transfonn domain, 18 77
recursive canonical fonn, 131 stabilization of
recursive canonical fonn, 132, 135 scalar feedback system, 44
recursive implementation, 36 weakly causal systems, 63
reduced fonn, 131 weakly causal transfer matrices, 53
INDEX 269
stabilizer characterization, 47 Toeplitz-plus-Hankel matrix, 207, 208

MIMO systems, 59 total degree ordering, 126
stabilizing compensator characterization, 52 transfer mtrix, 43
standard bases, 90 triangularized algebraic equations, 110
Stieltjes series (2-D), 226
stiff differential systems, 13 uniform word problem, 106
strictly Schur polynomial, 236 unimodular
strongly coprime (zero coprime), 201 column vector, 155
strongly coprime condition, 201 matrix, 155
strongly positive, 22, 24 unimodular completion, 240, 241
strongly stabilizable, 78 unimodular matrix completion, 170
strongly stable, 14 unique representation for residue classes, 89
structurally stable, 43 unsolvable, 109, 112
subdivision
averaging, 215 variety, 78, 86, 130, 155
interpolating, 215 empty, 155
order, 215 zero-dimensional, 157
submodule, 152 vertex
subpixel displacements, 209 (extreme) polynomials, 192
sum of squares matrix, 199
trigonometric polynomials, 25
sum of squares (SOS), 22, 33 wavelets
superresolution, 200 filter bank, 214
backprojection algorithm, 206 first generation, 214
multisensor array-based, 207 lifting scheme, 214
multisensor array-based model, 209 mother, 213
RLS update algorithm, 202 second generation, 214
robust, 200 weakly causal, 36, 39, 66
system, 215 weakly stable, 14
wavelets, 213 weakly zero prime, 155
support of sequence, 35 wide matrices, 147
Suslin's stability theorem, 177, 188 Wiener filter solution, 212
Sylvester matrix, 132 word problems
symbolic coefficients, 114 for commutative sernigroups, 106, 125
syndrome, 188 for ideals, 89
system function, 36
syzygies z-transform, 3, 36
chain of, 120 zero boundary conditions, 211
syzygy, 120, 152 zero left coprime, 161
first, 152 zero right
second, 152 coprime (ZRC), 179
coprime factorization, 179
Tikhonov regularization, 212 zero right coprime, 161
Toeplitz form, 25 zero right prime, 170
positive definite, 26 zero-dimensional, 157
Toeplitz matrix zero-dimensional polynomial ideal, 86
symmetric, 207 zero-set, 86

N. K. Bose, Multidimensional Systems Theory and Applications-Springer Netherlands (1995) PDF

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

N. K. Bose, Multidimensional Systems Theory and Applications-Springer Netherlands (1995) PDF

Uploaded by

Copyright:

Available Formats

MULTIDIMENSIONAL SYSTEMS THEORY AND APPLICATIONS

Mantle and Lower Crust

Springer-Science+Business Media, B.V.

ISBN 978-90-481-6424-0 ISBN 978-94-017-0275-1 (eBook)

Printed on acid-free paper

All Rights Reserved

List of Notations xii

1. TRENDS IN MULTIDIMENSIONAL SYSTEMS THEORY

2. CAUSAL AND WEAKLY CAUSAL 2-D FILTERS WITH APPLICATIONS

3. THE EQUATION Ax = b OVER THE RING qz, w] 79

4. GROBNERBASES: AN ALGORITHMIC METHOD IN POLYNOMIAL

5. MULTIVARliITEPOLYNOMIALS, MATRICES, AND MATRIX-FRACTION

3.1 Recursive Least Squares (RLS) Algorithm for Superresolution 202

Torus: TxT x ... x T~ T: {z E c: Izl = I}

Closed unitpolydisc: iT x U x ... x U, U: {z E C: Izl ~ I}

Preface to Second Edition

earlier chapters by supplementing as well complementing the material there. The

Towards the end of 1981, Professor M. Hazewinkel invited me to write a book

TRENDS IN MULTIDIMENSIONAL SYSTEMS THEORY

2. Multidimensional Systems Stability

of several fundamental stability criteria in other fields of science (for example,

2.1 MULTIDIMENSIONAL DIGITAL FILTERS

L ... L h[kl, k2, ... , kn]Z~l ... z~n

H(Zl' Z2,· .. , zn) = (1.2)

The investigation into BIBO stability reduces to the determining of conditions

L ... L Ih[k1, k2"'" kn]1 (1.3)

IlFllp ~ sup {{ !F(rz)I P dmn }I/P < 00 (1.4)

G( ZI, ... , Zn ) -~ an[H(ZI"'"

at every z E Tn at which this radial limit exists'.

Next, consider an irreducible rational function,

H(z) = A(z) (1.7)

The zero, then, is at

Note that z(O) is on rn.

Proof. B(z) =1- bizi - ... - bnzn

Iwl ~ 1 ++ 2 Re(l- w) ~ 11 - wl 2 (1.10)

In (1.9), ziO) = bklbkl- I . Define,

Using the constraint L~=llbkl = 1 and (1.11), .

is the Euclidean norm of (z - z(O)).

anH{Z)Zl ... Zn A1{z)

where Kl is a positive constant. Therefore, from (1.14), (1.15), and (1.16),

rational function H(z) which is analytic in U 2 , has a nonessential singularity of

Example 1.1 Let,

H( )_ (Zl - 1)n + (Zl -1)(Z2 - 1) + (Z2 - 1)n n » 1

For (Zl,Z2) = (e1'O'</J

m~O, n~O (1.18)

is BillO unstable when m = n = 1, and BillO stable when m = n = 8. For the

can be implemented in various forms, proceeding from criteria resulting from

2.2 MULTIVARIATE NETWORKS

Various types of electrical networks like lumped-distributed and variable pa-

DEFINITION 1.3 A polynomial, g(p) will be called a scattering HUIWitz polyno-

(a) g(p) =1= 0, Re p > 0, i.e. the open right-halfpolydomain and

The paraconjugate of g(p) is the polynomial g*(p) ~ g*( -p*)

DEFINITION 1.4 A polynomial, g(p), is a scattering HUIWitz polynomial if and

(b) I~(~)II ~ 1, Re p > 0

ge(P) = f.q(p) ~ g*(p)], go(p) = f.q(p) - g*(p)] (1.21)

FACT 1.1 If g(p) is a scattering HUIWitz polynomial of degree then the

It is mentioned that other definitions used with regard to multivariate Hurwitz

Example 1.2 Consider the following polynomial [5]

Ql (PI, P2) = 2p?p~ + p? + p~ + 2PIP2 + 2Pl + 2P2 + 2

Ql(jWl,jW2) = 2w~w~ - w~ - 2WIW2 - w~ + 2 + j2(WI + W2)

has been posed as a problem [5].

PROBLEM. Prove or disprove the following statement: The (n + 1)-variate polyno-

2.3 DELAY-DIFFERENTIAL SYSTEMS

The paraconjugate of g(p) is the polynomial g(p) ~ g( -p*)

ge(P) = f.q(p) ~ g(p)], go(p) = f.q(p) - g(p)] (1.21)