Matrices and graphs theory and applications to economics

Matrices and Graphs
Theory and Applications
to Economics



Proceedings of the Conferences on

Matrices and Graphs
Theory and Applications
to Economics

University of Brescia, Italy

8 June 1993
22 June 1995

Sergio Camiz
Dipartimento di Matematica "Guido Castelnuovo"
Universita di Roma "La Sapienza", Italy

Silvana Stefani
Dipartimento Metodi Quantitativi
Universita di Brescia, Italy

b World Scientific

II

'

Singapore· New Jersey· London· Hong Kong


Published by

World Scientific Publishing Co. Pte. Ltd.
POBox 128, Farrer Road, Singapore 912805
USA office: Suite IB, 1060 Main Street, River Edge, NJ 07661
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library.

MATRICES AND GRAPHS
Theory and Applications to Economics
Copyright © 1996 by World Scientific Publishing Co. Pte. Ltd.
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.

ISBN 981-02-3038-9

This book is printed on acid-free paper.

Printed in Singapore by Uto-Print

v

FOREWORD
the editors

The idea to publish this book was born during the conference « Matrices
and Graphs: Computational Problems and Economic Applications» held in
the far June 1993 at Brescia University. The conference was such a success that
the organizers, actually the present editors themselves, after a short talk with
the lecturers, decided on the spot to apply to the Italian Consiglio Nazionale
delle Ricerche (CNR) for a contribution to publish the conference proceedings.
The second editor did and the contribution came after a while.
In the meantime the editors organized another conference on «Matrices
and Graphs: Theory and Economic Applications», held as the previous in
Brescia during June 1995, partly with different invited lecturers. The conference was a success again and therefore the first editor applied to the Italian
National Research Council and got a second contribution, that came only recently.
While the lecturers of the first conference, who were not at the second one,
were a bit upset, having submitted their paper without seeing any proceedings
published at that time, the lecturers common to the first and the second conference suggested to join the contributions and publish a unique book for both
conferences. This is what we did.
During all these years, both authors were very busy lecturing, researching,
publishing, raising more funds to make their research possible. Most papers
arrived late and were carefully read by the editors, then the search of suitable
referees was not easy, so that the reviewing process took also a while, some
papers being sent back to the authors for corrections and then submitted again
to referees. A complete re-editing was necessary in order to get the uniform
editor's style ... , well, these are the reasons of such a delay, but eventually here
we are.

The book reflects our scientific research background: for academic and scientific reasons both of us were drawn to different research subjects, both shifting from pure to applied mathematics and statistics, with particular attention
to data analysis in many different fields the first editor, and to operational
research and mathematical finance the second. So, in each of the steps of this
long way, we collected a bit of knowledge.
The fact that in most of investigations we dealt with matrices and graphs
suggested us to investigate in how many different situations they may be used.


vi

This was the reason that led to the conferences; as a result, this book looks
like a patchwork, as it is composed of different aspects: we submit it to the
readers, hoping that it will be appreciated, as we did.
In fact, the numerous contributions come from pure and applied mathematics, operations research, statistics, econometrics. Roughly speaking, we
can divide the contributions by areas: Graphs and Matrices, from theoretical
results to numerical analysis, Graphs and Econometrics, Graphs and Theoretical and Applied Statistics.

Graphs and Matrices contributions begin with John Maybee: in his paper New Insights on the Sign Stability Theorem, he finds a new characterization
of a sign stable matrix, based on some properties of the eigenvectors associated to a sign semi-stable matrix. Szolt Thza in Lower Bounds for a Class
of Depth- Two Switching Circuits obtains a lower bound for a certain class of
(0,1) matrices. It is interesting to note that the problem can be formulated
in terms of a semicomplete digraph D, if one wants to determine the smallest
sum of the number of vertices in complete bipartite digraphs, whose union is
the digraph D itself. Tiziana Calamoneri and Rossella Petreschi's Cubic
Graphs as Model of Real Systems is a survey on cubic graphs, i.e. regular
graphs of degree three, and at most cubic graphs, i.e. graphs with maximum
degree three and show a few applications in probability, military problems, and
financial networks. Silvana Stefani and Anna Torriero in Spectral Properties of Matrices and Graphs describe from one hand how to deduce properties
of graphs through the spectral structure of the associated matrices and on the
other how to get information on the spectral structure of a matrix through

associated graphs. New results are obtained towards the characterization of
real spectrum matrices, based on the properties of the associated digraphs.
Guido Ceccarossi in Irreducible Matrices and Primitivity Index obtains a
new upper bound for the primitivity index of a matrix through graph theory
and extends this concept to the class of periodic matrices. Sergio Camiz and
Yanda Thlli in Computing Eigenvalues and Eigenvectors of a Symmetric Matrix: a Comparison of Algorithms compare Divide et Impera, a new numerical
method for computing eigenvalues and eigenvectors of a symmetric matrix, to
more classical procedures. Divide et Impera is used to integrate those procedures based on similarity transformations at the step in which the eigensystem
of a tridiagonal matrix has to be computed.
Among contributions on Graphs and Econometrics we find Sergio Camiz
paper I/O Analysis: Old and New Analysis Techniques. In this paper, Camiz
compares various techniques used in I/O analysis to reveal the complex struc-


vii

ture of linkages among economic sectors: triangularization, linkages comparison, exploratory correspondence analysis, etc. Graph analysis, with such concepts as centrality, connectivity, vulnerability, turns out to be a useful tool for
identifying the main economic flows, since it is able to reveal the most important information contained in the I/O table. Manfred GilU in Graphs and
Macroeconometric Modelling deals with the search of a local unique solution of
a system of equations and with necessary and sufficient conditions for this s0lution to hold. He shows how through a graph theoretic approach the problem
can be efficiently investigated, in particular when the Jacobian matrix is large
and sparse, a typical case of most econometric models. Manfred Gilli and
Giorgio Pauletto in Qualitative Sensitivity Analysis in Multiequation Models
perform a sensitivity analysis of a given model when a linear approximation is
used, the sign is given and there are restrictions on the parameters. They show
that a qualitative approach, based on graph theory, can be fruitful and lead
to conclusions which are more general than the quantitative ones, as they are
not limited to a neighborhood of the particular simulation path used. Mario
FaUva in Hadamard Matrix Product, Graph and System Theories: Motivations
and Role in Econometrics shows how the analysis of a model's causal structure can be handled by using Hadamard product algebra, together with graph

theory and system theoretical arguments. As a result, efficient mathematical
tools are developed, to reveal the causal and interdependent mechanisms associated with large econometric models. At last, International Comparisons and
Construction of Optimal Graphs, by Bianca Maria Zavanella, contains an
application of graph theory to the analysis of the European Union countries
based on prices, quantities and volumes. Graph theory turns out to be a most
powerful tool to show which nations are more similar.

Graphs and Statistics papers are represented by three contributions. Giovanna lona Lasinio and Paola Vicard in Graphical Gaussian Models and
Regression review the use of graphs in statistical modelling. The relative merits of regression and graphical modelling approach are described and compared
both form the theoretical point of view and with application to real data.
Francesco Lagona in Linear Structural Dependence of Degree One among
Data: a Statistical Model models the presence of some latent observations using
a linear structural dependence among data, thus deriving a particular Markovian Gaussian field. Bellacicco and Tulli in Cluster identification in a signed
graph by eigenvalue analysis establish a connection between clustering analysis
and graphs, by including clustering into the wide class of a graph transformation in terms of cuts and insertion of arcs to obtain a given topology.


viii

After this review, it should be clear how important is the role of matrices
and graphs and their mutual relations, in theoretical and applied disciplines.
We hope that this book will give a contribution to this understanding.
We thank all the authors for their patience in revising their work. A
special thanks goes to Anna Torriero and Guido Ceccarossi for their constant
help, but especially we would like to thank Yanda Tulli, who did the complete
editing trying to (and succeeding in) making order among the many versions
of the papers we got during the revision process. Last, but not least, thanks
to Mrs. Chionh of World Scientific Publishers in Singapore, whom we do not
know personally, but whose efficieny we had the opportunity to know through
E-mail.

October, 1996.
Sergio Camiz and Silvana Stefani

The manuscripts by Sergio Camiz, Guido Ceccarossi, Manfred Gilli, Giovanna lona Lasinio and Paola Vicard, Francesco Lagona, and Bianca Maria
Zavanella, referring to the first Conference, have been received at the end of
1993. The manuscripts of Antonio Bellacicco and Yanda Tulli, Tiziana Calamoneri and Rossella Petreschi, Sergio Camiz and Yanda Tulli, Mario Faliva,
Manfred Gilli and Giorgio Pauletto, John Maybee, Silvana Stefani and Anna
Torriero, and Szolt Tuza, referring to the second Conference, were received at
the end of 1995.
This work was granted by the contributions from Consiglio Nazionale delle
Ricerche n. A.I. 94.00967 (Silvana Stefani) and Consiglio Nazionale delle
Ricerche n. A.I. 96.00685 (Sergio Camiz).


ix

Sergio Camiz is professor of Mathematics at the Faculty of Architecture of
Rome University ~La Sapienza~. In the past, he was professor of Mathematics at Universities of Calabria, Sassari, and Molise, of Statistics at Benevento
Faculty of Economics of Salerno University, and of Computer Science at the
American University of Rome. He spent periods as visiting professor at the
Universities of Budapest (Hungary), Western Ontario (Canada), Lille (France)
and at Tampere Peace Research Institute of Tampere University (Finland),
contributed to short courses on numerical ecology in the Universities of Rome,
Rosario (Argentina), and Leon (Spain), held conferences on data analysis applications at various italian universities, as well as the Universities of New
Mexico (Las Cruces), Brussels (Belgium), Turku and Tampere (Finland), and
at IADIZA in Mendoza (Argentina), contributed with communications to various academic congresses in Italy, Europe, and America. After a long activity in
the frame of computational statistics and data analysis for numerical ecology,
and in programming numerical computations in econometrics and in applied
mechanics, his present research topics concern the analysis, development, and
use of numerical mathematical methods for data analysis and applications in

different frames, such as economical geography, archaeology, sociology, and
political sciences. He was co-editor of two books, one concerning the analysis
of urban supplies and the other on pollution problems, and author of several
papers published on scientific journals.
Silvana Stefani is a Full Professor of Mathematics for Economics at the
University of Brescia. She got her Laurea in Operations Research at the University of Milano. She has been visiting scholar in various Universities in Warsaw (Poland), Philadelphia (USA), Jerusalem (Israel), Rotterdam (the Netherlands), New York and Chicago (USA). She has been Head of the Department of
Quantitative Methods, University of Brescia, from November 1990 to October
1994 and is currently Coordinator of the Ph.D. Programme ~Mathematics
for the Analysis of Financial Markets~. She was co-editor of two books, one
concerning the analysis of urban supplies and the other on mathematical methods for economics and finance, and author of numerous articles, published in
international Journals, in Operations Research, applied Mathematics, Mathematical Finance.

Typeset by Jj.TEX
Edited by Vanda Tulli


x

AUTHORS ADDRESSES

Antonio Bellacicco Dipartimento di Teoria dei Sistemi e delle Organizzazioni
Universita di Teramo
Via Cruccioli, 125, 64100 Teramo, Italia.
Tiziana Calamoneri Dipartimento di Scienza dell'Informazione
Universita "La Sapienza"di Roma
Via Salaria, 113, 00198 Rama, Italia.
E-mail:
Sergio Camiz Dipartimento di Matematica "Guido Castelnuovo"
Universita "La Sapienza"di Roma
P.le A. Moro, 2, 00185 Roma, Italia.

E-mail
Luigi Guido Ceccarossi Dipartimento Metodi Quantitativi
Universita di Brescia
Contrada S. Chiara 48/b, 25122 Brescia, Italia.
E-mail
Mario Faliva Istituto di Econometria e Matematica per Ie Decisioni Economiche
Universita Cattolica di Milano
Via Necchi 9, 20100 Milano, Italia.
Manfred Gilli Departement d':Econometrie
U niversite de Geneve
Boulevard Carl-Vogt 102, 1211 Geneve 4, Switzerland.
E-mail
Giovanna lona Lasinio Dipartimento di Statistica, Probabilita e Statistiche
Applicate
Universita "La Sapienza"di Roma
P.le A. Moro, 2, 00185 Rama, Italia.
E-mail
Francesco Lagona Dipartimento di Statistica, Probabilita e Statistiche Applicate
Universita "La Sapienza"di Roma
P.le A. Moro, 2, 00185 Rama, Italia.


xi

John Maybee University of Colorado
265 Hopi PI. Boulder, Co 80303 USA.
E-mail
Giorgio Pauletto Departement d'Econometrie
Universite de Geneve
Boulevard Carl-Vogt 102, 1211 Geneve 4, Switzerland.

E-mail
Rossella Petreschi Dipartimento di Scienza dell'Informazione
Universita "La Sapienza di Roma "
via Salaria, 113, 00198 Roma, Italia.
E-mail:
Silvana Stefani Dipartimento Metodi Quantitativi
Universita di Brescia
Contrada S. Chiara 48/b, 25122 Brescia, Italia.
E-mail
Anna Torriero Istituto di Econometria e Matematica per Ie Decisioni Economiche
U niversita Cattolica di Milano
Via Necchi 9, 20100 Milano, Italia.
E-mail
Vanda Tulli Dipartimento Metodi Quantitativi
Universita di Brescia
Contrada S. Chiara 48/b, 25122 Brescia, Italia.
Zsolt Tuza Computer and Automation Research Institute
Hungarian Academy of Sciences
1111 Budapest, Kende u. 13-17, Hungary.
e-mail
Paola Vicard Dipartimento di Statistica, Probabilita e Statistiche Applicate
Universita "La Sapienza"di Roma
P.le A. Moro, 2, 00185 Roma, Italia.
Bianca Maria Zavanella Istituto di Statistica, Facolta di Scienze Politiche
Universita Statale di Milano
Via Visconti di Modrone, 20100 Milano, Italia.
E-mail

xiii

Contents
New Insights on the Sign Stability Problem
John Maybee

1

Lower Bounds for a Class of Depth-two Switching Circuits
ZsoltTuza

7

Cubic Graphs as Model of Real Systems
Tiziana Calamoneri and Rossella Petreschi

19

Spectral Properties of Matrices and Graphs
Silvana Stefani and Anna Torriero

31

Irreducible Matrices and Primitivity Index
Luigi Guido Ceccarossi

50

A Comparison of Algorithms for Computing the Eigenvalues
and the Eigenvectors of Symmetrical Matrices

Sergio Camiz and Yanda Tulli
I/O Analysis: Old and New Analysis Techniques
Sergio Camiz

72

92

Graphs and Macroeconometric Modelling
Manfred Gilli

120

Qualitative Sensitivity Analysis in Multiequation Models
Manfred Gilli and Giorgio Pauletto

137

Hadamard Matrix Product, Graph and System Theories: Motivations
and Role in Econometrics
Mario Faliva
International Comparisons and Construction of Optimal Graphs
Bianca Maria Zavanella

152

176


xiv

Graphical Gaussian Models and Regression
Giovanna Jona Lasinio and Paola Vicard
Linear Structural Dependence of Degree One among Data: A
Statistical Model
Francesco Lagona
Cluster Identification in a Signed Graph by Eigenvalue Analysis
Antonio Bellacicco and Yanda Tulli

200

223

233


NEW INSIGHTS ON THE SIGN STABILITY PROBLEM
J. MAYBEE
Progmm in Applied Mathematics
University of Colomdo
We obtain a new characterization of when a matrix is sign stable. Our results
makes use of properties of eigenvectors of sign semi-stable matrices. No classical
stability theorems are required in proving our results.

1

Introduction

We deal with n x n real matrices. Such a Matrix A is called semistable (stable)
if every A in the spectrum of A, u(A) lies in the closed (open) left-half of the

complex plane. The real matrix sgn(A) = [sgn aij] is called the sign pattern of
A and two real matrices A and B are said to have the same sign pattern if either
aij bij > 0 or both aij and bij are zero for all i and j. When A is a real matrix
we let Q(A).be the set of all matrices having the same sign pattern as A. We
also will write A in the form A = Ad + A where Ad = diag[an, a22, ... ,ann]
and A= A-Ad.
Let u be a complex vector u = (Ul' U2, ... , un). We say that U is qorthogonal to Ad if au # 0 implies Ui = O. Notice that if U is q-orthogonal to
Ad, then U is q-orthogonal to Bd for every matrix BE Q(A).
Let A be a real matrix satisfying aij # 0 if and only if aji # 0 for all i # j.
Then A is called combinatorially symmetric and we may associate with A the
graph G(A) having n vertices and an edge joining vertices i and j if and only if
i # j and aij # o. The graph G(A) is a tree if it is connected and acyclic. We
also use, for any matrix, the directed graph D(A) defined in the usual way.
The real matrix A is called sign semi-stable (sign-stable) if every matrix
in Q(A) is semi-stable (stable). We will deal only with the case where A is
irreducible in order to keep the arguments simple (Gantmacher, 1964). All of
our results can be readily extended to the reducible case.
We will prove the following results about sign semi-stable matrices.
Theorem 1 The following are equivalent statements:
1. The matrix A is sign semi-stable.

2. Matrix A satisfies

(i) ajj S; 0, for all j,


2

(ii) aijaji ::; 0 for all i and j, and
(iii) every product of the form ai(1)i(2)ai(2)i(3) ... ai(k)i(l) = 0 for k 2:

3 where {i(I), i(2), ... , i(k)} is a set of distinct integers in N =
{l,2, ... ,n}.
3. There exists a positive diagonal matrix D = diag[dl, d2 , ... , dnl, di >
0, i = 1,2, ... , n such that the matrix DAD- 1 = Ad + S where S is skew
symmetric (Gantmacher, 1964) and satisfies (ii), and (iii).
Theorem 2 The following are equivalent statements about a sign semi-stable
matrix:

I'. The matrix A has .A = 0 as an eigenvalue.
2'. Every matrix in Q(A) has .A

=0

as an eigenvalue.

3'. There is an eigenvector u satisfying Au = 0 which is q-orthogonal to Ad.
Theorem 3 The following are equivalent statements about a sign semi-stable
matrix:

I". The matrix A does not have a purely imaginary eigenvalue.
2". No matrix in Q(A) has a nonzero purely imaginary eigenvalue.

3". There is no eigenvector u satisfying Au = ij.Lu, which is q-orthogonal to
Ad.
The equivalence of conditions (1) and (2) of Theorem 1 is a well known
result due to Maybee, Quirk, and Ruppert (see Jefferies et al, 1977 for one
proof of this result). All the known proofs of this equivalence make use of one
of the classical stability theorems. By proving that (1) ::::} (2) ::::} (3) ::::} (1) we
can avoid the use of any stability theorem, a fact of some independent interest.
A consequence of Theorem 1 is that the family of sign semi-stable matrices

can be identified with the family of matrices of the form A = Ad + S, where
S is skew-symmetric and A satisfies (i),(ii), and (iii). This fact is used in an
essential way to prove Theorem 3.
Our proofs of Theorems 2 and 3 lead directly to simple algorithms for
testing a given matrix satisfying the conclusions (i),(ii), and (iii) to determine
whether or not it is sign-stable.
Finally, given Theorems 1,2, and 3 we can state the following sign stability
result.


3

Theorem 4 The real matrix A is sign stable if and only if the following four
conditions are satisfied.

(i) ajj ::; 0 for all j;
(ii) aijaji ::; 0 for all i and j;
(iii) every product of the form ai(1)i(2)' ai(2)i(3) ... ai(k)i(l) = 0 for k 2: 3
where {i(l), i(2), ... ,i(k)} is a set of distinct integers in N = {I, 2, ... ,n}
(iv) the matrix A does not have an eigenvector q-orthogonal to Ad

2

The proof of Theorem 1

Suppose first that the matrix A is sign semi-stable. The fact that (i),(ii),
(iii), are then true follows by a familiar continuity argument which we omit.
Hence (1) ::::} (2). Given that (2) is true and A is irreducible it follows that,
if aij :f 0 then aji :f 0 also. For suppose aij :f 0 and aji = O. Since there
is a path from j to i in A, (iii) is violated. Thus A is combinatorially symmetric. But then G(A), the graph of A exists, is connected and has no cycles. Hence G(A) is a tree. Then by a theorem of Parter and Youngs (1962),

there exists a positive diagonal matrix D such that DAD- 1 = Ad + S where
Ad = diag[all' a22, ... , ann], S = [Sij], with Sii = 0 for i = 1,2, ... , n, and
Sij = -Sji for all i :f j. Thus (2) ::::} (3). Now set A = DAD-l and suppose Au = AU. Taking scalar products on the right and left with U yields
u·Au = U· Adu+u· Su = U· AU = ~lul2 and Au·u = Adu·u+Su·u = Alul 2.
We have U· AdU = Adu· u and U· Su = -Su· u. Hence 2Adu· u = (A + ~)luI2
so we obtain
Re(A) =

AI~~ u

L aiilu l2 where Io =

(1)

{j I ajj :f O}. Hence condition (i)
iE10
implies that for any A in a(A), Re(A) ::; O. Thus (3) implies (1) and Theorem
1 is proved.
Now a sign semi-stable matrix is sign stable if and only if it has no eigenvalues on the imaginary axis in the complex plane. On the other hand, if u is
an eigenvector of A belonging to an eigenvalue on the imaginary axis, then we
must have Ui = 0 if i E Io by (1), i.e. u is q-orthogonal to Ad.
Note also that it follows from the proof of Theorem 1 that, if aii = 0, i =
1 ... n, A is skew-symmetric and all the eigenvalues of A are purely imaginary,
hence A is not sign stable. If aii < 0, i = 1 ... n, then every nonzero vector
But AdU . u =


4

satisfies Re()..) < 0 so A is sign stable. Hence the interesting case for sign

stability is 1 :5 1101 < n, which we assume to hold henceforth.

U

3

The proof of Theorem 2

Let A be a sign semi-stable matrix. By conditions (i) and (ii) every term in
the expansion of det A has the same sign. Therefore if det A = 0, it must
be combinatorially equal to zero and hence every matrix in Q(A) also has
determinant equal to zero. It follows that (1') implies (2'). Our task is to
discover when there exists a non-zero vector U such that Au = O. Now U must
vanish on the (nonempty) set 10 so we partition the components of a candidate
vector U initially into the sets Z(Io), N(Io) where Z(Io) = {i liE I o}, Ui = 0
if i E Z(Io), and Ui =J 0 if i E N(Io). Now given a set Ip ;2 10 and a partition
of the components of U such that Ui = 0 if i E Z(Ip) and Ui =J 0 if i E N(Ip).
We look at the equations

L

SijUj

(2)

= 0

jEN{I,,)

If such an equation has exactly one nonzero term, it has the form SikUk = 0

for some fixed value of k. Since Sik =J 0 and k E N(Ip), this is a contradiction.
Hence we must place k E I p +1 ' We do this for each such occurrence. Thus
IpH ;2 Ip and Z(Ip) ~ Z(Ip+l) , N(Ip) ;2 N(IpH)' If the system (2) contains
no equation having only a single non-zero term, then IpH = Ip and Z(IpH) =
Z(Ip) , N(Ip+l) = N(Ip). We will examine this case below. Suppose that
Ip+l = N. Then Z(IpH) = Nand u=O, i.e. no matrix in Q(A) has zero as
an eigenvalue. It remains to consider the case where we have some Ip = IpH
with IIpl < n so Ui = 0 for i E Z(Ip) and Ui =J 0 for i E N(Ip). Clearly
every equation in system (2) at this point contains either no non-zero terms
or at least two non-zero terms. We have N(Ip) ~ 2 and the induced graph
(N(Ip)) is a forest. We claim that this forest consists of isolated single trees, i.e.
S(N(Ip)) = O. For suppose (N(Ip)) has a nontrivial tree To. This tree has a
vertex of degree one and there would then exist an equation in the subsystem
S(N(Ip))u = 0 having exactly one nonzero term, a contradiction. Next let
IN(Ip) I = q and suppose there exists r rows in the subsystem
SijUi =

L

uEN{Ip)

0, i E Z(Ip), having two or more non-zero entries. We have r ~ 1 so the set
of such rows is nonempty. Let this set be Zo(Ip) and consider the submatrix
S(N(Ip))UZo(Ip). The graph of this submatrix is a forest on the q+r vertices.
If IZo (Ip) I ~ q then the numbers of edges in this forest is at least 2r ~ r + q, a
contradiction. Similarly, there cannot be two directed paths from vertex k to


5

vertex l in the directed graph D(So) where So is the matrix of the subsystem
(2) for i E Zo(Ip). It follows that the subsystem Sou = 0 uniquely determines
one or more eigenvectors U belonging to .A = o. Hence each matrix in Q(A)
has at least one eigenvector U belonging to .A = 0 and vanishing upon the set
Z(Ip) ~ 10. Thus Theorem 2 is proved.
4

The proof of Theorem 3

We look for an eigenvector U such that U is q-orthogonal to Ad and Au = iJL,
for some JL -=I o. As in the proof of Theorem 2, we partition a candidate vector
u initially into the sets Z(Io), N(Io) with Ui = 0 if i E Z(Io) and Ui -=I if
i E N(Io). Now given Ip ;2 10 and a partition of U into Z(Ip) and N(Ip) we
look at the equations

°

L

SjkUk = O,j E Z(Ip),

(3)

kEN(Ip)

and

L

SjkUk = iJLuj,j

E

N(Ip),

(4)

kEN(Ip)

If any equation in the subsystem (3) contains exactly one nonzero term,
we have SjkUk = 0 and, as in the proof of Theorem 2, we adjoin each such k to
Ip. Similarly, if any sum on the left hand side of an equation in the subsystem
(4) contains no nonzero term, we have iJLuj = 0 and this contradiction compels
us to add the index j to Ip. Doing this for every such occurrence produces
the new set IpH ~ Ip and thereby the new partition, Z(IpH), N(IpH). If the
subsystem (3) contains no equation having a single term and the subsystem
(4) contains no empty sums, then Ip+l = Ip. We examine this case below.
Suppose that Ip+l = N. Then Z(IpH ) = Nand U = O. Since every matrix
in Q(A) has the same zero-nonzero pattern, it follows that no matrix in Q(A)
has a purely imaginary nonzero eigenvalue.
It remains to consider the case where we have some Ip = IpH with IIpl < n,
so Ui = 0 for i E Z(Ip) and Ui -=I 0 for i E N(Ip). At this point every equation
in the subsystem (3) has either no nonzero tenns or at least two nonzero terms.
Also the induced subgraph (N(Ip)) is a forest and contains no nontrivial trees,
since every sum in the subsystem (4) contains at least one term. This forest
must contain at least two trees, because if it were a single tree the subsystem
(3) would have an equation containing exactly one nonzero term. Moreover, if
a tree in the forest is adjacent to a vertex j E Z (I p) there must also be another
tree in the forest adjacent to vertex j for the same reason. We must therefore

6

have IN(Ip)1 2:: 4. Let jo be the index of a row in subsystem (3) containing
q 2:: 2 nonzero terms. Then vertex jo in G(A) must be adjacent to q distinct
trees in the forest (N(Ip)).
Now choose a pair of trees in (N(Ip)) adjacent to vertex jo. Set Uk = 0
if k is not a vertex of one of these trees. Let the trees be TI and T2, respectively. Then the submatrices S(TI ) and S(T2) are disjoint nonzero skew
symmetric submatrices of S. Hence they have nonzero purely imaginary eigenvalues iJ.LI and iJ.L2. Let VI and V2 be nonzero vectors satisfying S(TI)VI =
iJ.LI VI, S(T2)V2 = iJ.L2v2. Then any vectors aVI and f3v2 also satisfy these equations where a and f3 are nonzero constants. If J.LI = J.L2 then we choose a and
f3 to satisfy

where kl is a vertex in one tree and k2 is a vertex in the other. Thus aVI and
f3v2 determine a vector u such that Au = iJ.LI u with U q-orthogonal to Ad. If
J.LI #- J.L2 then choose ao such that aoJ.L2 = J.L2 and modify S by multiplying
S(T2) by ao to obtain S'(T2). The resulting matrix is in Q(A) and has iJ.LI as
an eigenvalue.
This proves that some matrix in Q(A) has an eigenvector q-orthogonal
to Ad and it belongs to a purely imaginary eigenvalue. Hence Theorem 3 is
proved.
References
Jefferies C., V. Klee, and P. van den Driessche , 1977.
matrix sign stable?~ Can. J. Math., 29: p. 315-326.
Gantmacher, F. R. , 1964.
New-York.

«When is a

The Theory of Matrices, Vols 1,2. Chelsea,

Parter S. and J. W. T. Youngs, 1962. «The symmetrization of matrices
by diagonal matrices~ J. Math. Anal. Appl., 4: p. 102-110.


7

LOWER BOUNDS FOR A CLASS OF DEPTH-TWO
SWITCHING CIRCUITS
Z. TUZA

Computer and Automation Institute
Hungarian Academy of Sciences
Let M = (aij) be an (m x m) matrix with zero diagonal and ar + a 2i > 0 for
all i =I j, 1 ~ i,j ~ m. For a set R of rows and a set 0 of colutnns, ~enote by
R x 0 the set of the IRI . 101 entries lying in the intersection of those rows and
columns. We prove that if Rl, . .. , Rl and 01, . .. ,Ol are l sets of rows and l sets
l

U(Rk x Ok) is identical to the

of columns of M, respectively, such that the set
set of nonzero entries of M (i.e., aij

=I 0

k=l

if and only if the i-th row is in Rk and
l

the j-th column is in Ok for some k ~ l), then

L

(IRkl

+ lOkI) ~ mlog2 m.

k=l

1

The problem

In this note we investigate an extremal problem on a class of m by m 0-1
matrices, motivated by switching theory. The particular case in question can
be formulated in several equivalent ways, as follows.
• Suppose that the square matrix M = (aij) E {o,l}mxm has zero diagonal, and at least one of aij and aji is 1 for all i f. j, 1 :::; i, j :::; m.
Minimize the total number of rows and columns in a collection of 1-cells
(submatrices with no 0 entry) such that each aij = 1 occurs in at least
one of those I-cells.
• Let B = (X U Y, E) be a bipartite graph with vertex classes X =
{Xl, ... , xm} and Y = {Yl. ... , Ym}, and edge set E, such that XiYi ~ E
for aliI:::; i :::; m, and at least one of XiYj and XjYi belongs to E for all
if. j, 1 :::; i,j :::; m. Find the smallest total number IV(Bl)I+·· .+IV(Be)1
of vertices in a collection of complete bipartite subgraphs Bi C B, Bi =
(Xi U Yi,Ei ), Ei = {xy I X E Xi, Y E Yi} (1 :::; i :::; e), such that
El U···uEe =E.
• Given a semi-complete directed graph D = (V, E) on m vertices, without
loops and parallel edges (i.e., each pair x, Y E V is adjacent either by

just one oriented edge, or by precisely two oppositely oriented edges


8

Xy, YX E E), determine the smallest sum of the numbers of vertices in
complete bipartite digraphs Di C D (with all edges oriented in the same
direction between the two vertex classes in each D i ) whose union is D .

• Suppose that a circuit has to be designed with inputs Xl, . .. ,X m and
outputs YI, . .. ,Ym, where a set of conditions Cij prescribes whether there
exists a directed path of length 2 from Xi to Yj (written as Cij = 1;
otherwise we put Cij = 0). Assuming Cii = 0 for alII::; i ::; m, and
Cij = 1 or Cji = 1 (or both) for all i '" j, 1 ::; i,j ::; m, minimize the
number of links (adjacencies) in such a circuit.
The equivalence of the matrix problem and the two types of graph theoretical formulations is established by the corresponding adjacency matrices:
In the bipartite case we define aij := 1 if and only if Xi is adjacent to Yj; or,
conversely, we join Xi to Yj if and only if aij = 1. For digraphs, the entry
aij = 1 of the matrix corresponds to the edge oriented from vertex i to vertex
j.
To see that the switching circuits also give an equivalent formulation, notice first that each link involved in a path of length 2 verifying Cij = 1 for some
pair i,j either starts from an input node or ends in an output node. Now, each
internal node Zk of a length-2 path connects a set Xk of inputs with a set Yk
of outputs, and the number of links incident to Zk is IXkl + IYkl. Therefore,
X k x Yk must be a I-cell in the 0-1 matrix (Cij). Conversely, each I-cell R xC
of r rows and c columns in a 0-1 matrix M can be represented by an internal
node Z connected to r input nodes and c output nodes in the circuit to be
constructed.
Notation

We denote R xC := {aij I ri E R, Cj E C}, where R ~ {r}, ... ,rm } is a
set of rows and C ~ {CI,"" em} is a set of columns. (We may also view
the 0-1 matrices as subsets of {rl,' .. , rm} X {c}, ... , em}.) The shorthand
l

U (Rk x Ck) =

M means that the entry aij has value 1 in M if and only if

k=l

ri E Rk and

Cj E C k for some k, 1 ::; k ::; i (and aij = 0 otherwise). The
complexity, IT(M), of M is defined as

where the value of i is unrestricted.


9

Obviously, the definition of u(M) can be extended for arbitrary (not necessarily square) 0-1 matrices, but in this paper we do not consider the more
general case; i.e., M E {O, 1}mxm will be assumed throughout.

2

The results

It can be shown (Tuza, 1984) that
cm2

u(M) < logm
holds for every (m x m) matrix M, for some constant c, and also that this
upper bound is best possible apart from the actual value of c. For some restricted classes of matrices, however, the complexity can be much smaller. This
is the case, for example, in the following two particular sequences, as proved
by Tarjan (1975).

Theorem 1 If m = 2n and M = (aij) is the upper triangle matrix (aij = 1
for 1 ::; i < j ::; m and aij = 0 otherwise), then
u(M)

= n . 2n = m log2 m .

/2

Theorem 2 If m = (L n J)' where LxJ is the lower integer part of x, i. e. the
largest integer not exceeding x, and M = (aij) is the matrix J - I with aii = 0
and aij = 1 for all i =I- j, 1::; i,j ::; m, then

Our main goal is to show that the lower bound of m log2 m in Theorem 1 is
valid for a much larger class of (m x m) matrices. Namely, we will prove the
following result:

Theorem 3 If an (m xm) matrix M
and aij + aji > 0 for all i =f:. j, then

= (aij) (aij E {O, I}) has zero diagonal

Theorems 1 and 3 are best possible in general, as it is discussed in the concluding section. On the other hand, we are going to observe that the complexity

10

a(M) of a typical member of the class of matrices involved in Theorem 3 is
much larger than O(mlogm). To formulate this assertion more precisely, denote
E

{O, 1}7nX7n I aii = 0, aij

+ aji > 0

M~ := {M = (aij) E

{O, l}7nX7n I aii = 0, aij

+ aji =

Mm := {M = (aij)

for j

=f i} ,

1 for j

=f i} .

There is a constant c > 0 such that

Theorem 4

cm 2

a(M) 2: - logm

holds for (1 - 0(1)) IMml matrices M E Mm and (1 - 0(1)) IM~I matrices
M E M~ as m ~ 00.
The proofs of Theorems 3 and 4 are given in Sections 3 and 4, respectively.
Some open problems are mentioned in Section 5.

3

The general lower bound

The subject of this section is to prove Theorem 3, i.e., that a(M) 2: m log2 m
holds for all matrices M = (aij) of order m with zero diagonal, containing at
least one nonzero entry in each {aij, aji}, i =f j.
Suppose that an optimal collection of all-l submatrices Rk x Ck C M

e

has been chosen,

U(Rk x Ck) =

e

M,

k=l

L (IRk I + ICkl) =

a(M). Let X

=

i=l

{XI, ... , xl'l, and let us define the following two sets for i = 1,2, ... , m:

Ai

.-

{xklriERk},

Bi

.-

{Xk

I Ci E Ck}.

Notice that the ordered pairs (i,j) with Ai n B j =f 0 correspond to precisely
those entries aij of M which occur in at least one Rk x Ck, therefore the

e

assumption

.

U(Rk x Ck) = M and the initial conditions on M imply that
k=l

(i)
(ii)

n Bi = 0
Ai n B j =f 0

Ai

for alII:::; i :::; m, and
or Aj

n Bi =f 0

for all i

=f j,

1:::; i,j :::; m.