CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler

Lecture Notes in Economics and Mathematical Systems
For information about Vols. 1-156, please contact your bookseller or Springer-Verlag
Vol. 157: Optimization and Operations Research. Proceedings 1977 Vol. 184: R. E. Burkard and U. Derigs, Assignment and Matching
Edited by R. Henn. B. Korte. and W. Oettll. VI, 270 pages. 1978. Problems: Solution Methods with FORTRAN-Programs. VIII, 148
pages. 1980.
Vol. 158: l. J. Cherene, Set Valued Dynamical Systems and Econom,c
Flow. VIII. 83 pages. 1978. Vol. 185: C. C. von Weizsacker, Barriers to Entry. VI, 220 pages. 1980.
Vol. 159: Some Aspects of the Foundations of General Equilibrium
Vol. 186: Ch.-l. Hwang and K. Yoon, Multiple Attribute Decision
Theory: The Posthumous Papers of Peter J. Kalman. Edited by J. Green.
Making - Methods and Applications. A State-of-the-Art-Survey. XI,
VI, 167 pages. 1978.
259 pages. 1981.
Vol. 160: Integer Programming and Related Areas. A Classified Vol. 187: W. Hock, K. Schittkowski, Test Examples for Nonlinear
Bibliography. Edited by D. Hausmann. XIV, 314 pages. 1978. Programming Codes. V. 178 pages. 1981.
Vol. 161: M. J. Beckmann, Rank in Organizations. VIII, 164 pages.1978.
Vol. 188: D. Bos, Economic Theory of Public Enterprise. VII, 142 pages.
Vol. 162: Recent Developments in Variable Structure Systems, Eco· 1981.
nomics and Biology. Proceedings 1977. Edited by R. R. Mohler and
Vol. 189: A. P. LUthi, Messung wirtschaftlicher Ungleich~eit. IX, 287
A. Ruberti. VI, 326 pages. 1978.
pages. 1981.
Vol. 163: G. Fandel, Optimale Entscheidungen In Organisationen. Vol. 190: J. N. Morse, Organizations: M~ltiple Agents with Multiple
VI, 143 Seiten. 1979. Criteria. Proceedings, 1980. VI, 509 pages. 1981.
Vol. 164: C. l. Hwang and A. S. M. Masud, Multiple Objective De- Vol. 191: H. R. Sneessens, Theory and Estimation of Macroeconomic
cision Making - Methods and Applications. A State-of-the-Art Survey. Rationing Models. VII, 138 pages. 1981.
XII, 351 pages. 1979.
Vol. 192: H. J. Bierens: Robust Methods and Asymptotic Theory in
Vol. 165: A. Maravall, Identification in Dynamic Shock-Error Models. Nonlinear Econometrics. IX, 198 pages. 1981.
VIII, 158 pages. 1979.
Vol. 193: J. K. Sengupta, Optimal Decisions under Uncertainty. VII,
Vol. 166: R. Cuninghame-Green, Minimax Algebra. XI, 258 pages.
156 pages. 1981.
1979.
Vol. 194: R. W. Shephard, Cost and Production Functions. XI, 104
Vol. 167: M. Faber, Introduction to Modern Austrian Capital Theory.
pages. 1981.
X, 196 pages. 1979.
Vol. 195: H. W. Ursprung, Die elementare Katastrophentheorie. Eine
Vol. 168: Convex Analysis and Mathematical Economics. Proceedings
Darstellung aus der Sicht der Okonomie. VII, 332 pages. 1982.
1978. Edited by J. Kriens. V, 136 pages. 1979.
Vol. 196: M. Nermuth, Information Structures in Economics. VIII, 236
Vol. 169: A. Rapoport et aI., Coalition Formation by Sophisticated pages. 1982.
Players. VII, 170 pages. 1979.
Vol. 197: I nteger Programming and Related Areas. A Classified
Vol. 170: A. E. Roth, Axiomatic Models of Bargaining. V, 121 pages. Bibliography. 1978 - 1981. Edited by R. von Randow. XIV, 338
1979. pages. 1982.
Vol. 171: G. F. Newell, Approximate Behavior of Tandem Queues. XI, Vol. 198: P. Zweifel, Ein okonomisches Modell des Arztverhaltens.
410 pages. 1979. XIX, 392 Seiten. 1982. .
Vol. 172: K. Neumann and U. Steinhardt, GERT Networks and the Vol. 199: Evaluating Mathematical Programming Techniques. Pro-
Time-Oriented Evaluation of Projects. 268 pages. 1979. ceedings, 1981. Edited by J.M. Mulvey. XI, 379 pages. 1982.
Vol. 173: S. Erlander, Optimal Spatial Interaction and the Gravity Vol. 200: The Resource Sector In an Open Economy. Edited by
Model. VII, 107 pages. 1980. H. Siebert. IX, 161 pages. 1984.
Vol. 201: P. M. C. de Boer, Price Effects in Input-output-Relations:
Vol. 174: Extremal Methods and Systems Analysis. Edited by A. V.
Fiacco and K. O. Kortanek. XI, 545 pages. 1980. A Theoretical and Empirical Study for the Netherlands 1949-1967.
X, 140 pages. 1982.
Vol. 175: S. K. Srinivasan and R. Subramanian, Probabilistic Analysis
Vol. 202: U. Witt, J. Perske, SMS - A Program Package for
of Redundant Systems. VII, 356 pages. 1980.
Simulation and Gaming of Stochastic Market Processes and Learning
Vol. 176: R. Fare, Laws of Diminishing Returns. VIII, 97 pages. 1980. Behavior. VII. 266 pages. 1982.
Vol. 203: Compilation of I nput-Output Tables. Proceedi ngs, 1981.
Vol. 177: Multiple Criteria Decision Making-Theory and Application.
Edited by J. V Skolka. VII, 307 pages. 1982.
Proceedings, 1979. Edited by G. Fandel and T. Gal. XVI, 570 pages.
1980. Vol. 204: K. C. Mosler, Entscheidungsregeln bei Rislko: Multivariate
stochastische Dominanz. VII, 172 Seiten. 1982.
Vol. 178: M. N. Bhattacharyya, Comparison of Box-Jenkins and Bonn
Monetary Model Prediction Performance. VII, 146 pages. 1980. Vol. 205: R. Ramanathan, Introduction to the Theory of Economic
Growth. IX, 347 pages. 1982.
Vol. 179: Recent Results in Stochastic Programming. Proceedings,
1979. Edited by P. Kall and A. Prekopa. IX, 237 pages. 1980. Vol. 206: M. H. Karwan, V. Lotfi, J. Teigen, and S. Zionts, Redundancy
in Mathematical Programming. VII, 286 pages. 1983.
Vol. 180: J. F. Brotchie, J. 'Iv. Dickey and R. Sharpe, TOPAZ - General
Planning Technique and its Applications at the Regional, Urban, Vol. 207: Y. Fujimori, Modern Analysis of Value Theory. X, 165 pages.
and Facility Planning Levels. VII, 356 pages. 1980. 1982.
Vol. 181: H. D. Sherali and C. M. Shetty, Optimization with Dislunctive Vol. 208: Econometric Decision Models. Proceedings, 1981. Edited
Constraints. VIII, 156 pages. 1980. by J. Gruber. VI, 364 pages. 1983.
Vol. 182: J. Wolters, Stochastic Dynamic Properties of Linear Eco- Vol. 209: Essays and Surveys on Multiple Criteria Decision Making.
nometric Models. VIII, 154 pages. 1980. Proceedings, 1982. Edited by P. Hansen. VII, 441 pages. 1983.
Vol. 183: K. Schittkowski, Nonlinear Programming Codes. VIII, 242 Vol. 210: Technology, Organization and Economic Structure. Edited
pages. 1980. by R. Sato and M.J. Beckmann. VIII, 195 pages. 1983.
continuation on page 407

Lectu re Notes
in Economics and
Mathematical Systems
Managing Editors: M. Beckmann and W. Krelle
294
J. Jahn W. Krabs (Eds.)
Recent Advances
and Historical Development
of Vector Optimization
Proceedings of an International Conference
on Vector Optimization
Held at the Technical University of Darmstadt, FRG
August 4-7, 1986
Spri nger-Verlag
Berlin Heidelberg New York London Paris Tokyo
Editorial Board
H.Albach M. Beckmann (Managing Editor) P. Dhrymes

G. Fandel G. Feichtinger J. Green W. Hildenbrand W. Krelle (Managing Editor)
H.P.Kunzi K.Ritter R.Sato U.Schittko P.Sch6nfeld R.Selten
Managing Editors
Prof. Dr. M. Beckmann

Brown University
Providence, RI 02912, USA
Prof. Dr. W. Krelle

Institut fUr Gesellschafts- und Wirtschaftswissenschaften
der Universitat Bonn
Adenauerallee 24-42, D-5300 Bonn, FRG
Editors
Prof. Dr. Johannes Jahn

Institut fur Angewandte Mathematik
Universitiit Erlangen-Nurnberg
Martensstr. 3, D-8520 Erlangen, FRG
Prof. Dr. Werner Krabs

Fachbereich Mathematik
Technische Hochschule Darmstadt
SchloBgartenstr. 7, D-6100 Darmstadt, FRG
ISBN-13: 978-3-540-18215-3 e-ISBN-13: 978-3-642-46618-2

DOl: 10.1007/978-3-642-46618-2
This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation,
broadcasting, reproduction on microfilms or in other ways, and storage in data banks. Duplication
of this publication or parts thereof is only permitted under the provisions of the German Copyright
Law of September 9, 1965, in its version of June 24, 1985, and a copyright fee must always be
paid. Violations fall under the prosecution act of the German Copyright Law.
© Springer-Verlag Berlin Heidelberg 1987
2142/3140-543210
PREFACE
This volume contains refereed contributions to the "International

Conference on Vector Optimization" held at the Technical University
of Darmstadt from August 4-7, 1986. The aim of this conference was to
bring together economists, engineers and mathematicians being inter-
ested in recent advances in vector optimization. This meeting was an
interdisciplinary forum devoted to new results in the theory, to ap-
plications as well as to the solution of vector optimization problems
which are relevant in practice. The exchange of ideas and experiences
of its participants improved the insight into various disciplines in
which vector optimization plays an important role.
Because of the great variety of topics covered by the contribu-
tions the articles of this volume are organized in different sections:
Historical retrospect, mathematical theory, goal setting and decision
making, engineering applications, and related topics. The papers of the
invited State-of-the-Art Tutorials given by Professors J.M. Borwein,
H. Eschenauer, W. Stadler and P.L. Yu are also included.
This international conference was made possible by the financial
support by the German Science Foundation (DFG), the Minister for Sci-
ence and Arts of the State of Hessen, the Department of Mathematics of
the Technical University of Darmstadt, SIEMENS AG, IBM Deutschland GmbH
and Deutsche BP AG. We are very grateful to all of these sponsors.
Furthermore, we would like to thank all members of the international
program committee as well as our colleagues Professors K. Keimel,
J. Kindler and P. Spellucci for their support in the local program
committee of this conference. Moreover we are very indebted to Drs.
U. Lamp and G. Leugering for their help and also to Mrs. A. Garhammer
for her assistance.
Johannes Jahn Werner Krabs

Erlangen Darmstadt
TABLE OF CONTENTS
I. HISTORICAL RETROSPECT
W. STADLER:
"Initiators of Multicriteria Optimization" 3
II. MATHEMATICAL THEORY
E. BEDNARCZUK:
"Well Posedness of Vector Optimization Problems" 51
J.M. BORWEIN:
"Convex Cones, Minimality Notions, and Consequences" 62
A.R. DA SILVA:
"Evaluation Functionals are the Extreme Points of a Basis
for the Dual of c~[a,bl" 86
S. DOLECKI and C. MALIVERT:

"Polarities and Stability in Vector Optimization" 96
R. DURIER:
"Sets of Efficiency in a Normed Space and Inner Product" 114
K.-H. ELSTER and A. GOPFERT:

"Recent Results on Duality in Vector Optimization" 129
J. GWINNER:
"A Result of Farkas Type and Duality in Vector
Optimization" 137
S. HELBIG:
"Parametric Optimization with a Bottleneck Objective and
Vector Optimization" 146
VI
J. JAHN:
"Duality in Partially Ordered Sets" 160
I. KALISZEWSKI:
"Generating Nested Subsets of Efficient Solutions" 173
U. KRAUSE:
"Hierarchical Structures in Multicriteria Decision Making" 183
R. LUCCHETTI:
"Well Posedness, Towards Vector Optimization" 194
L.I. POLISHCHUCK:
"On Some Applications of Stochastic Dominance in
Multiobjective Decision-Making" 208
P. SERAFINI:
"Some Considerations About CompUtational Complexity for
Multi Objective Combinatorial Problems" 222
A. STERNA-KARWAT:
"On the Existence of Cone-Efficient Points" 233
A. WIECZOREK:
"Pseudo-Utilities" 241
III. GOAL SETTING AND DECISION MAKING
J. ESTER:
"A Fuzzy Concept of Efficiency" 257
G. ISLEI:
"An Approach to Measuring Consistency of Preference
Vector Derivations Using Least Square Distance" 265
Lj. VLACIC, A. WIERZBICKI and B. MATIC:

"Aggregation Procedures for Hierarchically Grouped
Decision Attributes with Application to Control System
Performance Evaluation" 2.85
VII
H.R. WEISTROFFER:
"A Flexible Model for Multi-Objective Optimization" 311
P.L. YU and I.S. CHIEN:

"Foundation of Effective Goal Setting" 317
IV. ENGINEERING APPLICATIONS
H.A. ESCHENAUER:
"Multicriteria Optimization Procedures in Application on
Structural Mechanics Systems" 345
V. RELATED TOPICS
H. KIRSTEN and R. TICHATSCHKE:

"The Efficiency of a Method of Feasible Directions for
Solving Variational Inequalities" 379
P.D. ROBINSON and P.K. YUEN:

"Bivariational Bounding Methods" 400
I. HISTORICAL RETROSPECT
INITIATORS OF MULTICRITERIA OPTIMIZATION
W. Stadler
Division of Engineering
San Francisco State University
1600 Holloway Avenue,
San Francisco, CA 94132, U.S.A.
Suwmary
riulticriteria optimization is what one attempts to carry out in

every applied decision process. Perhaps this was most evident to the
early political economists such as Edgeworth and Pareto who laid the
foundations of the subject in their competitive equilibrium and
welfare theories. For ~athematicians, such as Cantor and Hausdorff,
it was reflected in the consideration of ordered sets. For anyone who
studied economics, a theory of psychological games would seem to be
useful in outtrading the other participants to achieve one's economic
ends. Not surprisingly, Borel first constructed such a theory, and
economics and game theory were properly united by von Neumann and
11orgenstern. The subject was further developed by Koopmans within the
niceties of a linear theory of the firm. It eventually gained
mathematical recognition when Kuhn and Tucker provided a proper
definition of a vector maximum, and mathematical acceptance with
Hurwicz's treatment of the subject in topological vector spaces.
~Vhat follows is a biographical treat~ent of these authors and their

fundamental contributions to multicriteria optimization.
1. Introduction
As with most mathematical disciplines, multicriteria optimization

has its roots in a variety of other mathematical disciplines and it
has come into its own only recently. Nearly all of the early work was
done in connection with welfare and utility theory. This was followed
by some work in game theory and in production theory; eventually, the
subject evolved to what now is often termed the vector maximum problem.
4
The subject of this paper is the evolution of the following

formal multicriteria optimization problem:
Let Q (open) c nn and introduce the inequality constraints
the equality constraints
with corresponding functional constraint set
x = {XEJRn : XEQ, f(x):<> 0, h(x) = OJ.
The criterion functions are
gi(') : X - R, i = 1,2, ... ,N,
with corresponding criteria map
The attainable set is
Y = g (X) = {y EJRN : y = g (x), x EX} .
The basic problem statement is: Obtain "optimal" decision (s) x EX for
g (x) subject to x EX.
In the preceding the following notation has been used for the
natural order on JRn: For x, y ERn,
x :<> y iff x. :<> Yi' ViE I = {1,2, ... ,n};

~
x < Y iff x :<> Y, and x t y;
x Y iff xi =Y i , ViE I;
x« Y iff x 1 < y i ' ViE I.
In a broad sense, multicriteria optimization is nothing more than

an organized approach for selecting good decisions from a decision
set D in the presence of conflicting desiderata. As in the case of a
single desideratum, there are two fundamentally different approaches
to accomplish such a task. In the one-dimensional case, one may
introduce a preference i on D directly, in some fashion, or one may
5
induce such a preference by means of a mapping ~(.) D .... JR., with
(see Figure 1). That is, the linear order $ on the real numbers
'/'(2) OR,:s)
•
Figure 1. Decisions with one Desideratum
is used to induce a preference ~ on D via the mapping ~(.). In the

physical sciences, one generally specifies a physical criterion
function, such as weight, whereas a good deal of mathematical
economics deals with the construction of preferences ~ on D. However,
even there one quickly investigates conditions subject to which there
exists an orderpreserving map from D to JR., a utility function.
An analogous approach is used in multicriteria optimization - one

may either have competing preferences on the decision set directly, or
one may introduce a vector-valued map g(.) : D .... JR.N and use a suitable
preference on JRN to induce a preference on D (see Figure 2). The
former approach is more prevalent in economics, where one may, in fact,
have different decision sets along with the different preferences; the
192
g(·Hg,(·),9.(·»
~---1f--t-+----...lL~
9,
Figure 2. Decisions with many Desiderata
latter multicriterion approach han become the accepted one in the

sciences and in engineerin~As before, one ma~ of course, ir-terpret g1 (.),
g2(·), ••• ,gN(·)' as the utilities of the corresponding preferences.
6
Generally, the fundamental purpose of a preference on a decision

set is to enable a decision maker to indicate his preference for one
decision over another. For a sufficiently structured preference, one
may also define optimal elements, a best choice. In this context, the
maximal elements with respect toAa preference ~ have found the
greatest acceptance: A decision d ED is maximal with respect to ~ on
D iff
d ~ d ==* d'" d.
If the preference is induced by means of a mapping g(.) = (g1 (.),

••• ,gN(·» with ~onlRN, then one writes: A decision dED is ~
maximal iff
g(d) ~ g(d) ==l> g(d) - g(d).
That is, one simply uses the maximal values over the range of g(.) to
induce maximal decisions.
In order to make use of traditional single criterion optimization

methods and because of an ingrained desire for unique solutions, one
often scalarizes the multicriteria problem. In a sense, this is
nothing more than a continuation of the previous process. Namely, one
introduces a mapping
W( .) : (lRN ; ~) -+ (lR s)
with
x;iy <F;> W(x) sW(y),
x, Y ElRN Thus, W(.) may again be viewed as a utility function for ~

on lRN or it may be taken as given a priori.
Ideally, maximizing W(g (d» with respect to dE D then generates the

set of maximal decisions; at the very least, the maximizing decision
belongs to this set. In economics, W(.) has long been termed the
welfare function, since its maximum presumeably provides the
collective maximum welfare which competing consumers may attain bas.ed
on their individual preferences. In vector optimization, N(·) is
termed a compromise function, regret function, or the like.
7
These are the essentials of multicriteria optimization. ~1any have

contributed to its evolution and the following discussion presents
only what the author deems to be initiating contributions to the
subject from its inception until 1958. Much of the discussion is
based on an earlier survey article (Ref. 1), which als-o contains
additional related works.
2. A Historical Sketch
Vector optimization is an inherent part of economic equilibrium

theory and as such it can be said to have been founded by Adam Smith
in his treatise The Wealth of Nations in 1776. According to Arrow
(Ref. 2) "the full recognition of the general equilibrium concept can
unmistakeably be attributed to Leon Walras". Within equilibrium theory,
the work most relevant to the topic at hand was Walras', Jevons' and
Menger's work on utility theory, and the work on welfare theory by
Edgeworth and Pareto spanning the period from 1874 to 1906.
A second contributing branch was the inception of the theory of

psychological games and the notion of strategy by Borel in 1921, and
subsequently by von Neumann in 1927.
In 1951, Koopmans introduced the efficient point set in the context

of production theory.
Hathematically, the subject can be said to have had its origin in

the early work of Cantor and Hausdorff on ordered sets during the
period 1895 to 1906. However, the subject became a mathematical
discipline only as recently as 1951 when the concept of a vector
maximum problem was introduced by Kuhn and Tucker at one of the many
Berkeley symposia on mathematical statistics and probability. It became
socially accepted among the pure mathematicians when it was treated in
abstract spaces by Hurwicz in 1958.
As indicated, multicriteria optimization arose naturally in

mathematical economics and in game theory. Its independent application
to other problems began with the work of Koopmans in production theory,
and with its use to solve water resources planning problems by Marglin
in 1967. Although the subject was introduced in the engineering
literature with a short note by Zadeh in 1963, its use in engineering
8
a.nd in the science.s did not begin in earnes.t until the 1970s. This
latter evolutjon of the subject is traced in Refs. 3 and 4.
The following papers present a selective history of multicriteria

optimization. Clearly, their selection reflects the author's
preferences. From an objective point of view, however, they were
selected because they represent large steps in the evolution of the
subject, they are still well-springs of ideas, and because the authors
carried their ideas to formal mathematical conclusions. Last, not least,
they were a pleasure to read and to study.
3. Welfare Theory and Scalarization
Welfare theory traditionally deals with competing economic agents,

all of whom presumeably reach their points of maximum utility at the
Pareto point obtained from the maximization of a welfare function W(.).
Such consumers with competing preferences were first treated with

reasonable rigor from a utility point of view by Francis Ysidro
Edgeworth, Drummond Professor of Political Economy at Oxford, and a
fellow of All Souls College. Apparently, he did not have much influence
on the current thinking of his time. However, modern readers "find him
a stimulating and reassuring intellectual forbear, surprisingly up-to-
date in many respects, and still instructive" (Ref. 5). It will soon
be clear just how up-to-date he was with respect to vector optimization.
Along with Pareto, he bases his approach to economics on qualitative

mathematical reasoning similar to that used in the physical sciences,
such as hydrodynamics, for example. He defines economics with (Ref. 6,
p. 6): "... economics investigates the arrangements betw·een agents each
tending to his own maximum utility; ..• ", and he had high hopes for
economics as a science (Ref. 6, p. 12):" 'Mecanique Sociale' may one day
take her place along with 'Mecanique Celeste' throned each upon the
double-sided height of one maximum principle,
Many of Edgeworth's conclusions pertaining to multicriteria

optimization were stated in the context of two individuals acting
within a simple exchange economy. In economics texts the argument is
often presented graphically by means of the Edgeworth Box with
Mathematical Psychics as the relevant reference. Although th8
9
discussion may, of cou~se, be found there, the Edgeworth Box was

first used by the statistician Arthur Lyon Bowley (1869 - 1957),
holder of the first chair of statistics at the University of London
(Ref. 7, p. 4). We shall use it here also, since it does serve nicely
to visualize Edgeworth's discussion. Indeed, we shall include some
game theoretic aspects with an eye on the next section.
Figure 3. The Edgeworth Box and the Core
The relevant discussion begins on p. 20 of Ref. 6. In today's

language, we consider two economic agents X and Y, with convex
utilities P(x,y) and n(x,y), and trading in two commodities (x,y). The
consumption set of both agents is the positive quadrant. The quadrants
are overlapped in such a way that agent X has null of each commodity
at the lower left and the total endowment (a,b) at the top right
whereas agent Y has (0,0) at top right and (a,b) at lower left. Thus,
every point (x,y) in the box may be interpreted as a distribution of
(a,b) with the first agent receiving (x,y) and the second (a -x, b -y).
The level curves P(x,y) = p of agent X (dashed lines) and n(x,y) = n of
agent Y (solid lines) are the lines of indifference; that is, any point
on such a line is a point of equal utility for the agent. He then
derives the following conditions characterizing the equilibrium points
(x,y) for the economy:
(1) Ref. 6, p. 21. "It is required to find a point (x,y) such that
in whatever direction we take an infinitely small step, P and n do not
increase together but that, while one increases, the other decreases."
He terms the collection of these points the contract curve,
10
corresponding to what now is termed the Pareto set, efficient set,

and so on.
He continues (p. 22): "If then we enquire in what direction X and

Y will consent to move together, the answer is, in any direction
between their respective lines of indifference, in a di,rection
positive as it may be called for both. At what point will they refuse
to move at all? When the lines of indifference are coincident (and
lines of preference not only coincident, but in opposite directions)."
Thus, in essence, the gradients must be codirectional with
I7P (x,y) = k 17 n (x,y) ,
~ ap _ k an Cln
ax - ax and k Cly'
(jp ClIT
Cly ax'
which he notes is a necessary but not sufficient condition for a point

to lie on the contract curve.
(2) Ref. 6, p. 23. "Or again, we may consider that motion is

possible so long as, one party not losing, the other gains. The point
of equilibrium, therefore, may be described as a relative maximum, the
point at which, e.g. IT being constant, P is a -maximum."
He forms the Lagrangian
L(X,y) = P(x,y) - c(IT(x,y) -IT)
and notes that a necessary condition for a maximum then is given by
(~ - c ~) dx + (ClP - c ClIT) dy 0
Clx Clx Cly Cly ,
again yielding points of the contract curve. The result is the same if
agent Y is used.
On p. 27 he then generalizes the result to m agents and n

commodities. With the exception of the inclusion of inequality
constraints he has thus noted that conditions for (Pareto) optimality
may most easily be derived when based on the following lemma.
11
Lemma 1. Let dE D be Pareto opt~mal. Then, for an arbitrary choice

of j E:r = {1 ,2, ••• ,N} one has that d maximizes gj (d) subject to dE D
and
(3) Ref. 6, p. 53 (footnote). In a case where both agents

cooperate with one another, he strives for a unique solution by
defining the utilitarian point (i,y) with
P <,c,y) + n (x,y) = max {P (x,y) + n (x,y)}.
Furthermore, he notes that (x,y) belongs to the contract curve and

that the result is quite general; namely, max {P(x,y) +An(x,y)}, A
being the coefficient of effective sympathy, also yields a point on
the contract curve.· He thus has essentially asserted:
Lemma 2. Assume c i > 0, i 1,2, •.• ,N, with
and assume d ED maximizes W(d) subject to d ED. Then d" ED is a Pareto

optimal decision.
(4) Ref. 6, pp. 22 - 23. He indicated that ~~ > 0 would imply that
both agents could increase their utilities; that if ~~ < 0, motion is
impossible in any direction, thus implying that this is a sufficient
condition for (Pareto) optimality (the condition is only necessary,
in general). Finally, on p. 25, he indicates that the maxima and
minima of P(x,y) and n(x,y) represent demarkation points for the 'pure'
and 'impure' parts of the contract curve.
(5) Ref. 6, p. 37 - 42. Although not directly relevant here, it

was first pOinted out by Shubik (Ref. 8) that Edgeworth also presented
the fundamental concepts concerning the core of a cooperative game.
It is evident from this discussion that Edgeworth provided much of

the foundation for multicriteria optimization; and, he clearly laid
the groundwork for Pareto's generalizations.
Edgeworth's results and arguments were based on the assumed

existence of each agent's utility. Pareto was the first to take the
12
indifference curves of each agent as given, a fundamentally different

approach, whos·e merits and logic are still being discussed.
Pareto was a civil engineer with a social and political conscience

whose earlier socialist leanings eventually gave way to a viewpoint
which made him one of the founders of fascism. His interest in
economics evolved through his close friendship with the economist
Maffeo Pantaleoni who introduced him to Walras and Walras' writings
on economic equilibrium. When Walras retired, he supported Pareto as
his successor to the chair of political economy at the University of
Lausanne.
Pareto's early training in mathematics during his engineering

studies at Turin, and his mastery thereof quickly placed him ahead of
his contemporaries in economics and allowed him, along with Edgeworth,
to lay the foundations for the mathematical treatment of economic
equilibrium. Many contributed to the further evolution and refinement
of the subject (see Ref. 1, within the context of vector optimization)
culminating in Debreu's Theory of Value (1959) which is the basic text
for a modern treatment of economic equilibrium. Its extensions and
general use for the prediction of economic phenomena led to the 1983
Nobel Prize in economics for Gerard Debreu.
Generally, Pareto derives his results for n consumers and m goods.

A detailed and annotated discussion of these results was given in Ref.
1. Here, we shall restrict the presentation to a simple exchange
economy with two consumers and two goods, where each consumer has a
fixed initial allocation and where an economic equilibrium is an
allocation which maximizes each consumer's ophelimity as well as the
ophelimity of the collectivity. He provides the following definition
for the maximum ophelimity of the collectivity (Ref. 9, Chapter 6,
Section 33):
"We will say that the members of a collectivity enjoy maximum

ophelimity in a certain position when it is impossible to find a way
of moving from that position very slightly in such a manner that the
ophelimity enjoyed by each of the individuals of the collectivity
increases or decreases. That is to say, any small displacement in
departing from that position necessarily has the effect of increasing
the ophelimity which certain individuals enjoy, and decreasing that which
others enjoy, of being agreeable to some and disagreeable to others."
13
Let the two goods be given as X,Y and let a specific allocation
for a consumer be denoted by wi = (xi'Yi)' i 1,2. Assume that the
initial allocations of the consumers are woi (xoi'Y oi ) and that their
trade is constrained by obstacle surfaces fi(w i ) = O. The ophelimity
(preference) of each consumer is characterized by a family of
indifference surfaces whose integration along some specific path
(when possible) yields the index of ophelimity I = F(~) (a nonunique
utility function ~(.), since any monotonically increasing transformation
F(·) thereof is another utility function with the same indifference
curves). Pareto coined the term ophelimity to express the economic
aspects of consumer satisfaction, since he believed that the term
"utility" had long served to describe the broader sociological aspects
of satisfaction.
He then takes an exchange equilibrium to be an allocation w = (w 1 ,w 2 )

at which the following conditions are satisfied: The total allocation
must match the initial allocation with
Satisfaction of the constraints together with maximum individual

ophelimity implies
fi (w.) dx. + fi (w.) dYi 0,

x 1. 1. Y 1.
~~(wi) dx.1. + ~i (w.) dYi 0,

Y 1.
and hence,
~'li th the good X taken as money, we have Px 1 for the price of X and
for the price of the good Y at equilibrium. If the prices are constant,
then it furthermore follows that
~i (w. )
Y 1.
~i (w.)
X 1.
14
It remains to characterize the maximum ophelimity of the collectivity

at equilibrium. Towards this purpose, he first notes that
Ocp1(W 1 ) cp~ (w 1 ) oX 1 + cp~ (w 1 ) oy 1

O<jl2(W 2 ) CP~(W2) oX 2 + CP~(W2) oY2
and then imposes the condition
Since the cpi(.) are essentially positive, this condition can be

x .
satisfied only if one of the Ocpl(.) is negative, thus characterizing
a Pareto optimum for (cp1 ,cp2). Note that the expression is not the
variation of a welfare function in its usual sense. With the previously
defined variations and prices, the condition may also be written in
the form
the balance sheet of the collectivity, implying that the variation of

the receipts and expenditures is zero.
Pareto always emphasized his debt to those who worked before him,
even when he considered their theories wrong or inadequate. To his
detractors he wrote: "Perfecting a theory is completely different from
seeking to destroy it by foolish and pedantic sUbtleties."
Most of the early scalarization occured in the context of welfare

theory based on a desire for a unique equilibrium. Usually, a general
welfare function was introduced which was then endowed with certain
properties.
Marshall and his pupil Pigou represented the total welfare of the
society by the sum
where
cp i ( . ) the utility of the i-th member,

15
Xi the consumption of the i-th member.
Bergson (Ref. 10) is the first to generalize this approach by

introducing an economic welfare function W(~1'~2' ••• '~N)' assumed to
be an increasing function of the individual utilities. He writes:
"Evidently, also to maximize W(·) would satisfy the criteria of
Pareto and Barone." He thus asserts:
Lemma 3. Suppose W(·) : RN ~ E is a compromise function which is

monotonically increasing in each of its arguments. Then d ED such that
W(g(d» max {w 0 g (d) : dE D}
is a Pareto optimal decision.
4. Game Theory
Arrow (Ref. 2) writes: "The historical relation between game theory

and economic equilibrium theory is paradoxical. In principle, game
theory is a very general notion of equilibrium which should either
replace the principle of competitive equilibrium or include it as a
special case. In fact, while game theory has turned out to be
extraordinarily stimulating to equilibrium theory, it has been through
the use of mathematical tools developed in the former and used in the
latter with entirely different interpretation."
Game theory is included in the present discussion because of the

first of these statements, since it does include the notions of
economic equilibrium and the concepts of optimality which are used in
vector optimization problems. In fact, while such notions as min max
and Nash equilibrium have been introduced in a competitive player
context, there is no reason whatsoever why they should not acquire
useful meaning in vector optimization, as desirable optima for a single
decision maker.
As mentioned earlier, games of chance have an ancient history,

eminently traced, for example, in Gods, Games and Gambling by Florence
David. Borel is the initiator of psychological game theory. There, in
addition to chance, the psychology of the player and his choice of a
suitable strategy, may affect the outcome of a game. He writes (Ref.
11): "We will consider a game in which the winning depends
16
simultaneously on chance and on the skill of the player." He also

points out that the only one who had considered a particular case of
the problem prior to him, was Joseph Bertrand in his Calcul de
Probabilities.
The min max theorem is fundamental to the theory of games. Borel

demonstrated it for n = 3, and n = 5, and he indicated that it would
be of interest to know whether it held for general n or not. Von
Neumann's own claim to being the originator of the theory is based
largely on his proof of the min max theorem for general n (Ref. 12).
His results were based on showing that every two-person, zero-sum
game, with mixed strategy and expectation as the payoff, has a min
max solution. He introduced the following assumptions:
AssumEtion, 1. (Ref. 12, p. 307). Let
EI; {I; E:nf1 I; 2: 0, 1;1 + + I;M ,; 1} ,
E
n
{n ERN n 2: 0, n
1
+ ... + n N ,; 1},
and let f(·) satisfy:
(i) n EEn' 1;' ,1;" EEl;' f(I;' ,n) 2:A, f(I;" ,n) 2:A ~ f(l;,n) 2:A,
for every I; = e 1;' + (1 - e) 1;", 0,; e ,; 1;
(ii) I; EEl;' n' ,n" EEn' f(l;,n') ';A, f(l;,n") ,;A -" f(l;,n) ';A,
for every n = en' + (1 - e) n", 0,; e ,; 1.
With this assumption, von Neumann then proves the following theorem.
Theorem 1. (Ref. 12, p. 307). Assume that f(.) is continuous and

that it satisfies Assumption 1. Then
min f(l;,n) min max f(l;,n).

nEE n EE n I;EE I;
n
Frechet (Ref. 13) indicates, however, that proofs of such theorems

and more general ones had appeared before the notes of Borel and von
Neumann. Frechet's eloquent exposition and von Neumann's reply to his
allegations make i~teresting reading. Frechet asks the reader to judge.
17
I have no doubt that von Neumann developed the theory independent

of Borel. After discovering, however, that Borel had worked along a
similar vein earlier, it would have detracted nothing from his own
accomplishments and genius if he had acknowledged Borel as the
initiator of the theory of psychological games, and had cited his
contributions.
A similar situation was dealt with nicely when Kuhn and Tucker
discovered that the optimality conditions attributed to them had been
stated and proven earlier by W. Karush (1939) in an unpublished
~aster's thesis, and by F. John (1948) in a paper appropriately
titled "Extremum Problems with Inequalities as Subsidiary Conditions".
Kuhn (Ref. 14) wrote a historical paper in which these related
efforts are put into proper perspective and he rectifies his earlier
omission by publishing Karush's results as an appendix to his paper.
5. Efficiency in Production Theory
Tjalling Koopmans gave an extensive treatment of the linear vector

optimization problem in the context of production theory. In the
writing, he acknowledges that he benefitted from discussions with
Morton Slater, George Dantzig, Oskar Lange, and others, an enviable
discussion group. Also, he was subsequently made aware of an article
by Robert Remak, who queries: "Can Economics Become an Exact Science?"
This is an excellent article which can be recommended to any serious
(germanspeaking) student of mathematical economics.
The paper by Koopmans (Ref. 15) is long and the present cow~ents
can only hint at the extent of his results. Suppose that the total
output of a commodity n is given by
k
1:: a .x.,
j=1 nJ J
where a nj is the contribution of the j-th activity to the output of

the n-th commodity given in rate of flow per unit time of the n-th
commodity for a unit amount of the j-th activity. For N commodities
and k activities one may thus write
v = A x
"'" --'
18
where ~ is termed the technology matrix. He makes postulates which are

sufficient to insure that the range of attainable commodities is a
cone which does not include the positive quadrant in its interior. The
attainable cone is denoted by (~). He then defines a (maximizing)
efficient point as a possible point yE (~) for which there is no
~ E (~) such that
y - y ~ Q.
Furthermore, he defines a local possible cone (R) at the point yE (~)
as the set of all vectors 9 of the form
9 = idy-y), A. > 0, yE (~),
to avoid having to deal with cones whose vertex is not at the origin.
~~~
~·~UM~------------------------------------Y,
(0)
-Y2
Figure 4. The Attainable Cone (A) and the Local Possible Cone (D~ at y
Numerous necessary and sufficient conditions for efficiency are

given. Among them the following theorem.
Theorem 2. (Ref. 15, p. 61). A point y E (~) is efficient iff

there exists a vector p such that
0, p»Q.
The vector R is termed a "normal" to (~) in y. He goes on to

distinguish efficient facets, splits y into primary, intermediate and
final commodities, ~ = (~pri' ~int' ~fin) and develops further
theorems with respect to ~fin only, including the introduction of
19
commodity availibility constraints.
6. The Vector Maximum Problem
From a mathematical point of view, ordered spaces and orderpreserving

maps provide the basis for mathematical economics as well as for
multicriteria optimization. The foundation for infinite dimensional
ordered spaces was laid by Cantor (Ref. 16) and by Hausdorff (Ref. 17).
Cantor also introduces equivalence classes (so important to preference
theory), denoting equivalence by -, and he introduces the notation
in connection with the generalization of inequality. He deals with
"simply" ordered sets and sets on which several different orders have
been introduced. Cantor states the first sufficient conditions for the
existence of a utility function and Hausdorff gives the first example
of a complete ordering (the lexicographic ordering) for which there
does not exist a utility function. It is of interest to note the
evolution of mathematical terminology by comparing Cantor's statement
of the sufficient conditions to that which may be found in modern
textbooks.
Theorem 3(a). (Ref. 16, p. 511). Suppose an ordered set M has a

structure which (1) makes it perfect; (2) contains a subset S with
cardinality S =~O' which is related to M in .such a way that for any
two elements mO and m1 of M there are elements of S which lie "between"
mo and m1 with respect to the order on M, then M= 8.
Theorem 3 (b) • (A modern statement). A chain D is isomorphic with

a subchain of the chain of real numbers iff D contains a countable,
order-dense subset.
Any mathematician would benefit from a study of Georg Cantor's work

and life. His writing is lucid and dense, and even when one thinks to
have understood, yet another layer appears underneath.
Fifty years later, Kuhn and Tucker (Ref. 18) extended the methods
of linear programming to nonlinear programming including inequality
constraints. Prompted, in part, by Koopmans' paper on efficient
solutions, they also posed a vector maximum problem. YH th only
inequality constraints considered, they define the vector maximum
problem as: "Find x satisfying f (x) ~ 0 such that g (x) > g (x) for no x
20
satisfying f (x) ~ O. "
They impose what are now termed the Kuhn-Tucker constraint

qualifications (resulting in proper solutions). They introduce
o
x
for the partial derivatives of
 (x,u) v~ g(x) + u T f(x)
at (x,a) and prove the following necessary conditions (among others).
Theorem 4. (Ref. 18, p. 483). In order that x be a proper solution

of the vector maximum problem, it is necessary that there be some
vo » 0 such that x and some a satisfy the conditions:
(a)
o
x " 0,
OT
x
x 0, x~ O.
(b) o
u ~ 0, OT
x
a 0, a ~ O.
Vector maximization became of age mathematically when Hurwicz (Ref.

19) generalized the Kuhn-Tucker results to topological vector spaces.
This is another excellent paper which is recommended to anyone
wishing to do vector optimization in Banach spaces.
In view of the length of the paper, let it suffice to state the

generalization of the previously cited Kuhn-Tucker theorem. Let X and
Y be arbitrary spaces and let Z be a vector space with origin Oz and
with convex wedge P z in Z. Then Z becomes an ordered vector space
(Z, ~) with
Pz
z1:S z2 iff z2 - z1 EP z ' z1,z2 EZ •
Let Q eX, and define the constraint map
the feasible set
x = Q n f- 1 (7) ) {x EX
z
21
the criterion map
g(o) : X ~ Y
and denote
Y = g (X) = {y E Y : y g (X), x EX}
as the attainable set.
The set of support points of Y is denoted by
f1(Y*(o),y) = {yEY : y*(y) = max{y*(y) : yEY}}.
A functional y*(.) EY* is strictly positive on Py iff
and implies y*(y) > o.
Now, for some Yo E Y, let Ko be the closure of the intersection of

all of the convex wedges containing the set (Y -YO) U (-py)' Then, if
A
K contains no y' ~ 0y and y is ~ - maximal, y is called properly

y Py
maximal.
A subset X* of X* is regularly convex if, given an element

x* ( .) $ x*, there exists an element x of the underlying space X such
that
where hx(o) : X* ~ m is defined by
h-(x*(o))
x
= x*(x).
A mapping f(o) is regular if, for some X ;0:0, f(x»> o.
Theorem 5. (Ref. 19, p. 99). Assume the following:
(i) X, Y, Z are Banach spaces with closed convex wedges Px

and Pz ;
(ii) X; is regular convex and f(·) is regular;
(iii) Y is a topological vector space with the property that, for

22
every closed convex wedge K s: Y, there is a linear

continuous functional y*(.) EY*, strictly positive on K:
(iv) g(.) and f(·) are Frechet differentiable on X:
(v) x is properly maximal.
Then, for some y~(.) >~ 0;, the function

Py
(x,z*(.» = y*[g(x)] + z*[f(x)]
has a nonnegative quasi saddle-point at (x,z~(.): y~(.»: that is,
0x«x,z;(.»;~) :$;0 x+ ~,
0x«X,z;(.»;x) =0:
(ii)
for all z * ( .) :> 0*
z'
ff*
z
0z*«X,z;(.»: z;(.» = Z;[f(X)] = o.
A detailed survey pertaining to the vector maximum problem in

topological vector spaces is given by Dauer and Stadler (Ref. 20).
7. Conclusions
Mathematical economics clearly provided the impetus and early

context for vector optimization. Nearly all of those cited were
practitioners of the art, or had acce$to them. There is a nice
thread of continuity leading from Edgeworth to Hurwicz. The only
exception thereto are Cantor and Hausdorff who seem to have developed
a theory of ordered sets purely out of mathematical interest and,
perhaps, necessity. Borel was extremely active politically and thus
most certainly consulted economists and their work. So much for the
theory.
The story is somewhat less encouraging on the applied side. To my

knowledge, there are only three areas in which one can claim to have
used multicriteria optimization, the wide use of equilibrium models
23
in economic forecasting, the multicriteria ecological requirement in

U.S. Water Resources Management, and its use in the construction of
highly focused systems. If the subject is to gain the wide acceptance
which it deserves, we must emphasize the ~ with which one can
consistently and rationally include several criteria in the design
process. Although the mathematical intricacies of the subject may be
enjoyable and challenging, they are generally of little interest to
the economically and practically oriented designer or manager. The
step from the treatment of one criterion to several is easy and it
should be presented as such.
The subject is by no means finished. The following areas of research

are clearly indentifyable:
(1) A wider concept of vector optimization erobodied in the

definition of a vector optimum. Much of the mathematical theory is
based on maximality with respect to partial orders or partial preorders,
at best. More general concepts of optimality based on preferences
satisfying conditions other than merely those of reflexivity and
transitivity are needed.
(2) Computational algorithms and display of the optima. An

essential aspect of multicriteria optimization is that there generally
are an infinity of optima. A designer can usually be convinced that
the final choice should lie within this infinite set; however, he must
come up with a single final design. To enable him to do so in an
informed manner, time efficient numerical algorithms for the generation
of the Pareto set of nonlinear programming and control problems must be
developed, as well as methods of displaying the optimal solution set
when a large number of criteria is involved.
(3) Multicriteria aspects of natural phenomena. To date all of the

so-called natural laws have been based on optimality concepts which are
maximum or minimum principles with respect to a single criterion.
Although there still is very little theoretical or practical evidence
of more general principles, it seems obvious that natural law quite
generally achieves an optimal balance among several criteria.
(4) Applications. There are numerous papers whose titles hint at

application; however, the emphasis then turns out to be yet another
discussion of multicriteria fundamentals with an academic example
24
addended. Indeed, nearly every field of endeavor permitting an

adequate mathematical model lends itself to multicriteria optimization.
References
1. Stadler, W., A Survey of Multicriteria Optimization or the Vector

Maximum Problem, Part I: 1776 - 1960, Journal of Optimization
Theory and Applications, Vol. 29, No.1, pp. 1 - 52, September 1979.
2. Arrow, K.J., Economic Equilibrium, International Encyclopedia of

the Social Sciences, edited by David L. Sills, Vol. 4, The
MacMillan Company & The Free Press, New York, New York, pp. 376 -
389, 1968.
3. Stadler, W., Multicriteria Optimization in Mechanics (A Survey),

Applied Hechanics Reviews, Vol. 37, No.3, pp. 277 - 236, Harch 1984.
4. Stadler, W., Applications of Multicriteria Optimization in

Engineering and in the Sciences (A Survey), in MCDM - Past Decade
and Future Trends (Source Book for Multiple Criteria Decision
Making), edited by Hilan Zeleny, JAI - Press, Greenwich,
Connecticut, 1984.
5. Hildreth, C., Edgeworth, Francis Ysidro, International Encyclopedia

of the Social Sciences, edited by David L. Sills, Vol. 4, The
MacMillan Company & The Free Press, New York, New York, pp. 506 -
509, 1968.
6. Edgeworth, F.Y., Uathematical Psychics, C. Kegan Paul & Co.,

1 Paternoster Square, London, England, 1331.
7. Bowley, A.L., Mathematical Groundwork of Economics: An Introductory

Treatise, Clarendon Press, Oxford, England, 1924. ~eprints of
Economic Classics, Augustus M. Kelley, New York, New York, 1960.
8. Shubik, M., Edgeworth Market Games, Contributions to the Theory of

Games, Vol. 4, edited by R.D. Luce, and A.\v. Tucker, Annals of
Mathematical Studies, Princeton University Press, Princeton, New
Jersey, pp. 267 -278, 1959.
9. Pareto, V., Manuale di Economica Politica, Societa Editrice

Libraria, Milano, Italy, 1906. Translated into French, with
revised Mathematical Appendix, by Girard and Briere, as Manuel
D'Economie Politique, Giard, Paris, France, First Edition, 1909
and Second Edition, 1927. Translated into English by A.S. Schwier,
as Manual of Political Economy, The MacMillan Company, New York,
New York, 1971.
10. Bergson, A., Socialist Economics, A Survey of Contemporary

Economics, Vol. 1, edited by H.S. Ellis, Irwin, Homewood,
Illinois, pp. 412 - 447, 1948.
11. Borel, E., The Theory of Play and of Integral Equations with Skew
Symmetric Kernels (in French), Comptes Rendu5 Academie des Sciences,
Vol. 173, pp. 1304 -1308. English translation in Econometrica, Vol.
21, No.1, pp. 97 - 100, January 1953.
25
12. von Neumann, J., On the Theory of Parlor Garnes (in German),
Hathematische Annalen, Vol. 100, pp. 295 - 320, 1923.
13. Frechet, M., Borel, Emile, Initiator of the Theory of

Psychological Garnes and its Application, Econometrica, Vol. 21,
pp. 95-127,1953.
14. Kuhn, H.W., Nonlinear Programming: A Historical View, Proceedings

of SIAM-AMS, Volume 9, pp. 1 - 26, 1976.
15. Koopmans, T.C., Analysis of Production as an Efficient Combination

of Activities, Activity Analysis of Production and Allocation,
Cowles Commission Monograph No. 13, edited by T.C. Koopmans, John
i.viley and Sons, New York, New York, pp. 33 - 97, 1951.
16. Cantor, G., Contributions to the Foundation of Transfinite Set

Theory (in German), Mathematische Annalen, Vol. 46, pp. 431 -512,
1895, and Vol. 49, pp. 207 - 246, 1897.
17. Hausdorff, F., Investigations Concerning Order Types (in German),

Berichte tiber die Verhandlungen der Koniglich Sachs is chen Gesell-
schaft der Wissenschaften zu Leipzig, 11athematisch-Physische
Klasse, Vol. 58, pp. 106 -169, 1906.
18. Kuhn, H.W., and Tucker, A.W., Nonlinear Programming, Proceedings

of the Second Berkeley Symposium on Mathematical Statistics and
Probability, edited by J. Neyman, Berkeley, California, pp. 481-
492, 1951.
19. Hurwicz, L., Programming in Linear Spaces, Studies in Linear and

Nonlinear Programming, edited by K.J. Arrow, L. Hurwicz, and
H. Uzawa, Standford University Press, Stanford, California, 1958;
Second Printing, Oxford University Press, London, England, pp.
38-102,1964.
20. Dauer, J.P., and Stadler, W., A Survey of Vector Optimization in

Infinite-Dimensional Spaces, Part II, Journal of Optimization
Theory and Applications, Vol. 51, No.2, pp. 205 - 241, November
1986.
26
Adam Smith (1723 - 1790)
Adam Smith's extraordinary influence on economics is based on two

books, The Theory of Moral Sentiments (1759), and The Wealth of Nations
(1776). The latter was the first book which attempted a systematic
treatment of the economic process in a competitive environment. He
wrote and published the former during his tenure as a professor of
moral philosophy, a position which he held from 1751 to 1763 at the
University of Glasgow, and the latter was written at home in Kirkcaldy
and was eventually published in London.
Adam was born in 1723, in Kirkcaldy, Fifeshire, near Edinbourgh,

the son of Adam Smith, a comptroller of customs, and Margaret Douglas,
daughter of a wealthy landowner in Fifeshire. He completed elementary
school in Kirkcaldy, entered the University of Glasgow in 1737, and
graduated with an M.A. in 1740. Subsequently, he went to Oxford
University as a Snell fellow at Balliol College where he presumeably
formed some of the views which he expounded in public between 1748 and
1751. In 1764 he went to France as tutor of the young duke of
Buccleuch, a service for which the elder duke provided him with a
generous pension for life.
He eventually returned to England, and in early 1778 he was

appointed a commissioner of customs for Scotland and as a commissioner
of salt duties. These posts required his presence in Edinbourgh, where
he lived until his death on July 17, 1790.
27
Jam Smith
1723 = 1790
28
Francis Ysidro Edgeworth (1845-1926)
Edgeworth was raised on the family estate of Edgeworthstown,

County Longford, Ireland.
At 17 he entered Trinity College, Dublin, and then proceeded to

Balliol College, Oxford, receiving first class honors in litterae
humaniores. After several years of practicing law in London, he
accepted a lectureship at King's College and was later appointed
Tooke professor of political economy.
In 1891 he became Drummond professor of political economy at

Oxford and a fellow of All Souls College. Never marrying, he resided
principally at All Souls for the remainder of his life.
He became the first editor of the Economic Journal in 1891. At the

time of his death in 1926 he was joint editor with Keynes.
In spite of the respect he enjoyed as a teacher and original

thinker, he seems not to have influenced the then current thinking
in economics and statistics. Many of todays methods in economic
equilibrium theory, however, have their roots in Edgeworth's work.
29
Francis Edgeworth
18~5 = 1926
30
Vilfredo Federico Damaso Pareto (1848 -1923)
Vilfredo Pareto was born in Paris, France, on July 15, 1848. His
father, the Marchese Raffaele Pareto had fled there from Italy for
political reasons. In 1858 they returned to Italy and Vilfredo
completed primary school and eventually attended the Polytechnic
Institute of Turin where he attained a degree in civil engineering.
He graduated in 1870, worked first as a director of the Rome

Railway Company and subsequently became managing director of an iron
processing plant with headquarters in Florence. During these years,
his disillusionment with the Italian government grew, he dispised the
Italian aristocracy for sucking the blood of the poor and he
ultimately resigned his directorship. His essentially liberal and
socialistic views, however, changed and many of Pareto's later ideas
served as the foundations of Mussolini's fascist economics.
Pareto was a scholar of Greek and Roman antiquity and he developed

an abiding interest in economics through his close friendship with
the economist Maffeo Pantaleoni. The mathematical skills he had
acquired during his studies at the Politechnic Institute soon made
him a leader in economic analysis. His books on economic theory, in
particular, his Cours d'economie politique became classics in
fundamental economics.
In 1893 Pareto succeeded Leon Walras as Professor of Political

Economy at Lausanne. He died at Celegny, Switzerland, on August 19,
1923.
31
JL848 = JL923
32
Emile Felix Edouard Justin Borel (1871 - 1956)
Emile was born on 7 January, 1871, in Saint-Affrique, Aveyron,

France. His father, Honore Borel was the village's protestant pastor
and his mother, Emilie Teissie-Solier came from a local merchant
family. By age 11 he was known as a prodigy and left the local school
to attend the lycee at Montauban. In 1889 he took first place in the
entrance exams for the Ecole Polytechnique, the Ecole Normale
Superieure, and in the general competitions. He elected to attend
the Ecole Normale.
During his first year there he published two papers, eventually

writing more than 300 papers including 35 books. He graduated first
in his class at age 22 and was immediately offered a teaching
position at the University of Lille. During the next three years he
wrote his thesis "Sur quelques points de la theorie des fonctions",
which included what is now termed the Heine-Borel theorem; in
addition, he wrote some 22 papers. By age 28 he was well-known to
analysts around the world.
His work was strongly influenced by Cantor and he had such

enviable discussion partners as Darboux, Hermite and Goursat and
he married Appell's eldest daugther. In 1909 a chair for the theory
of functions was created for him at the Sarbonne and in 1910 he
became vice director in charge of science students at the Ecole
Normale. He was elected to the Academy in 1921.
In the following years, Borel became extremely active in politics,

first as mayor of Saint-Affrique, then as Radical-Socialist member of
the chamber of deputies (1924 - 1926) and Minister of the Navy (1925).
After his retirement from politics in 1936 and from the Sarbonne in
1940, Borel still managed to write another 50 articles and books
before his death in Paris, France, on February 3, 1956.
33
Emille BOIrcn
1871 = 1956
34
Johann (John) von Neumann (1903 - 1957)
John von Neumann was born in Budapest, Hungary, on December 28,

1903, the eldest of three sons of Max von Neumann, a banker. He
attended private schools for his elementary education and he entered
the Gymnasium in 1914. His mathematical abilities were quickly
recognized and furthered by private tutoring. He published his first
paper when he was nineteen.
Von Neumann was a Privatdozent at Berlin from 1927 to 1929 and at

Hamburg from 1929 to 1930. He then visited Princeton for three years,
at the end of which he was invited to join the Institute for Advanced
Studies, becoming its youngest member. L.C. Young writes: "In
Princeton it used to be said that if something was obvious for
Hermann Weyl, then von Neumann could prove it; if it was obvious for
Marsten Morse, it was most certainly false". - A tribute to von
Neumann's quick grasp of the fundamentals of even the most complex
issues. He was equally at horne in Set Theory, Measure Theory or Lie
Groups and at his best in the Spectral Theory of Operators in Hilbert
Space. Along with Borel, he is the co-founder of the Theory of Games,
culminating in his book with Oscar von Morgenstern Theory of Garnes and
Economic Behavior, in 1944.
He was one of the many scientists who fled Nazi Germany along with
such greats as Kac, John and Bochner. At the outbreak of World War II,
he was asked to participate in war related projects, worked as a
consultant for Los Alamos in the construction of the atomic bomb from
1943 on, continued to work on many advisory committees after the war,
and became a member of the Atomic Energy Commission in 1954. He died
three years later, in Washington, D.C., on February 8, 1957.
35
John Von
1903 = 1957
36
Tjalling C. Koopmans (1910 - 1985)
He was born on August 28, 1910, in 'sGraveland, the Netherlands.

He received his M.A. from the University of Utrecht in 1933 and his
Ph.D. from the University of Leiden in 1936. Thereafter, he became a
lecturer at the Netherlands School of Economics, an economist for the
League of Nations, and a Research Associate at Princeton University.
From 1942 to 1944, he served as a statistician with the British
Merchant Shipping Mission in Washington, D.C.
From 1944 to 1954 he was with the Cowles Commission for Research in
Economics at the University of Chicago; in 1948, he became its director
and he joined the faculty as a Professor of Economics. In 1955 the
Cowles Commission moved to Yale University where it became the Cowles
Foundation for Research in Economics; Koopmans joined the Yale faculty
as Professor of Economics at the same time. He served as Director of
the Cowles Foundation from 1961 to 1967.
He received honorary doctorates from several universities; among

them, the Catholic University of Louvain (1967) and the University of
Pennsylvania (1976). He was a member of the National Academy of
Sciences, President of the American Economic Society and of the
American Econometric Society. In 1975, he shared the Nobel Prize for
Economics with the Russian economist Leonid Kantorovich for his
contributions to the theory of optimal allocation of recources.
He was author and editor of numerous books and papers ranging from
Tanker Freight Rates and Tankship Building (1939) to Activity Analysis
of Production and Allocation (1951). He continued to serve on the Yale
faculty until his retirement in 1981 and he remained actively involved
in the Cowles Foundation until his death in New Haven, Connecticut, on
February 26, 1985.
37
1910 = 1985
38
Georg Cantor (1645 - 1918)
Georg Cantor was born 3. March, 1845 in St. Petersburg. The family
moved to Germany in 1856. He wanted to study mathematics, his father
insisted on engineering. In 1862 his father relented. Georg started
studying in 1862 in Zurich and in 1863 continued in Berlin where Kummer,
WeierstraB and Kronecker were teaching.
Cantor is the initiator of set theory and transfinite analysis.

WeierstraB accepted Cantor's ideas whereas Kronecker explicitly warned
his students against Cantor's ideas.
Cantor got his doctorate at 22. In 1869 he settled in Halle (now

in East Germany). The continuum hypothesis was a discouraging work for
him which almost caused him to quit mathematics altogether. After 10
years of presenting his new ideas, at the age of 40, he had had no
recognition of his work among his peers. In particular, Kronecker and
partially also Hermite's open opposition to Cantor gave him
difficulties which led to a nervous collapse in 1884. His work finally
began to receive recognition from age 52 on. He died 6. January, 1918.
39
1845 = 1918
40
Felix Hausdorff (1868 - 1942)
Hausdorff was born to a wealthy merchant family in Breslau, Germany

(now Wroclaw, Poland) on 8 November, 1868. He completed his secondary
education in Leipzig, and subsequently studied mathematics and astronomy
at Leipzig, Freiburg and Berlin. He graduated from Leipzig in 1891.
He became a Dozent at Leipzig in 1896 and an Associate Professor in

1902. He took a position in Bonn in 1910, and became a Professor at
Greifswald in 1913. In 1921 he returned to Bonn, where he retired in
1935.
Hausdorff's main interests lay in philisophy and literature to the

extent that he published poems and aphorisms under the pen-name of Dr.
Paul Hongre. In mathematics, his work on topological and metric spaces
made him the founder of general topology and the general theory of
metric spaces. His major work was the Grundzlige der Mengenlehre which
he wrote in Leipzig during the period from 1910 to 1914.
He was not active politically, but as a Jew, the threat of internment

in Nazi Germany propted him to commit suicide with his wife and her
sister in Bonn, Germany, on 26 January, 1942.
41
868 = 9-42
42
Harold William Kuhn (1925)
Harold Kuhn was born in Santa Monica, California, on July 29, 1925.
After a brief interruption of his studies by service in the U.S. Army
(1944 -1946), he graduated from the California Institute of Technology
in 1947 with a B.S. in Mathematics. He then continued his studies at
Princeton, where he received his M.A. and Ph.D. degrees in 1948 and
1950, respectively. He was a Fulbright Research Scholar in Paris from
1950 to 1951. In 1952 he joined the mathematics department of Bryn Mawr
College and he left there in 1959 for a year's study at the London
School of Economics as a National Science Foundation Senior Post
Doctoral Fellow. He returned to Princeton in 1960 as an Associate
Professor of Mathematical Economics; he was appointed to the rank of
Professor in 1963.
He has served as consultant to various government organizations and

he is currently a senior consultant to Mathematics', Inc., a research
firm in Princeton, New Jersey. He is a past president of the Society
for Industrial and Applied Mathematics (1953 -1954) and he was the
executive secretary for the Division of l~athematics of the National
Academy of Sciences from 1953 until 1960.
He has authored numerous articles in operations research and

mathematical economics and he is co-editor of Mathema,tical Systems
Theory and Economics, and Associate Editor for the Journal of Economic
Theory, the ONR Logistics Quarterly and Systems Theory.
He presently has a joint appointment in Mathematics and in

Economics at Princeton University as a Professor of Mathematical
Economics.
43
44
Albert William Tucker (1905)
Albert Tucker was born on November 28, in Oshawa, Ontario, Canada,

he had his basic education in Toronto and he eventually attained his
B.A. and M.A. degrees from the University of Toronto. In 1929 he came
to Princeton to do graduate work in mathematics and he completed his
Ph.D. in 1932 with a dissertation on combinatorial topology. He then
became a National Research Fellow at Cambridge University, England,
Harvard University, and Chicago University. He joined the Princeton
mathematics department in 1933, became an Assistant Professor in 1934,
Associate Professor in 1938 and Professor in 1946. He succeeded Emil
Artin as Albert Baldwin Dod Professor in 1953 and simultaneously
succeeded Solomon Lefschetz as department chairman.
He has been a visiting lecturer at numerous universities such as

Stanford University and the Massachusetts Institute of Technology; he
was a Fulbright Lecturer at several Australian universities. During
World War II he taught for the Army Specialized Training Program and
he worked on gunnery direction problems for the Office of Scientific
Research and Development and for Frankfort Arsenal. Since 1948 he has
directed a research project sponsored by the Office of Naval Research,
centering on the theory of games, mathematical programming and
combinatorics and he has been editor of the Princeton University Press
series on advanced mathematics since its inception in 1938. He has
authored papers in topology and differential geometry but he is best
known for his contributions in nonlinear programming.
He has been extremely active in professional societies serving as

president of the Mathematical Association of America and vice president
of the American Association for the Advancement of Science, among others.
He has been a consultant to both the RAND Corporation in Santa Monica,
California, and to the IBM Research Center in York Town Heights, New
York. Most recently, he has been involved in improving the mathematical
curriculum in U.S. schools as chairman of the Commission on Mathematics
of the College Entrance Examination Board. He is now an Emeritus
Professor at Princeton University.
45
1905 =
46
Leonid Hurwicz (1917)
Hurwicz was born on August 21, 1917, in Moscow, Russia, where his
parents had fled from Poland during World War I. The family returned
to Poland after the war. After his basic education, he eventually
studied at the University of Warsaw. He left Poland in 1938 to s·tudy
economics at the London School of Economics (1938-1939), at the
Institut des Hautes Etudes Internationales in Geneva (1939 -1940),
Harvard University (1941) and he eventually completed his studies at
the University of Chicago in 1942.
He became an Associate Professor at Iowa State College, Ames, Iowa

(1946 - 1949), a Professor of Economics and Mathematical Statistics at
the University of Illinois (1949 -1951) and he has been a Professor of
Economics and Mathematics at the University of Minnesota since 1951.
He has been a visiting professor at numerous universities, among them
the University of California in Berkeley (1976) and the Wuhan Institute
of Technology in the Peoples Republic of China.
His earlier work in economics involved the use of statistical

techniques of analyzing economic policies. He has worked in game theory
and the application of linear and nonlinear programming in economics,
resulting in over 90 papers. More recently, he has focused on
decentralization in economic systems.
He has served as associate editor and as a member of the advisory

board of several economic journals. He has been, and still is, a
member of numerous advisory boards and he is a member of the National
Academy of Sciences.
47
LeoniJ H1ULlrwicz
1917 =
II. MATHEMATICAL THEORY
WELL POSEDNESS OF VECTOR OPTIMIZATION PROBLEMS
Ewa Bednarczuk
Systems Research Institute
Polish Academy of Sciences
01-447 Warsaw, Newelska 6
Summary
We investigate upper semicontinuity of solutions in parametric

vector optimization. In particular we distinguish a class of well posed
problems for which sufficient conditions for upper semicontinuity of
solutions are formulated without any compactness assumptions of
feasible solution sets.
1. Introduction
Let X and Y be vector topological spaces and let C be

a closed convex cone in Y, int c#0. A relation < defined as y<x
iff x E y+C introduce a partial order in Y.
Let S be a subset of Y. An element YOE S is said to be mini-
mal (efficient) if there is no y, Y#Yo such that Yo E y+C. An
element Yo is said to be weakly efficient if there is no y, Y#Yo'
such that Yo Ey+intC. The set of all efficient elements of a set S
with respect to a cone C will be denoted by E(S,C), or shortly
E(S). The set of all weakly efficient elements of a set S will be
denoted by wE(S,C) or wE(S).
Let f be a function defined on X and takinq its values in Y
and let r be a subset of X.
The vector optimization problem (VOP) for the pair (f, r) is
the problem of finding all elements x such that
f (x) E wE (f ( r »
where weak efficiency is understood in the sense of a given cone C.
52
The following examples are often used as illustrations in the

sequel.
1. f=id. The solutions of the corresponding (VOP) for (id,r) are

weakly efficient elements of r.
2. Let Y=R and C=R+={xlx>O}. is a real-valued function
If f:X+R
and rc X, then for (f,r) the corresponding (VOP) does not
coincide, in general, with the scalar optimization problem
inf{f (x) IXE:r} •
In particular, if f=id and r is not closed the solutions of the

~calar optimization problem need not coincide with weakly efficient
elements of r.
On the analogy of the scalar case the set N, N=wE(f(r» is

called the optimal value of the vector optimization problem. The set M
defined as
M = {x E: r: fIx) E:wE(f(r»}
is called the solution set.

The sets E(S) and wE(S) may be empty. Recently, conditions
for these sets to be nonempty are given by Borwein 119831, Sterna-
-Karwat 119801, Papageorgiu 119851. In the sequel we assume the set
wE(f(r» to be nonempty.
Moreover, the sets E(S) and wE(S) need not be closed. However,
in the finite-dimensional case, when Y=Rn , and C is a polyhedral,
cone wE(S) is closed (see Papageorgiu 119851).
2. Parametric optimization problems
Let U be a topological space. Let r:u+x be a multifunction and

f:XxU+Y
A family of problems (VOP) u of finding all elements x E: r (u)
such that
f (x , u) E: wE (f (r (u) , u) )
is a parametric vector optimization problem. The parametric vector

optimization problem gives rise to the following multifunctions.
The optimal value multifunction N:U+Y is defined as
53
N (u) = wE (fi r (u) ) ,C)
and the optimal solution multifunction M:U+Y is defined as
M(u) = {xE: r(u) If(x,u) E: N(u)}
Moreover, for a E: N (u), we define the set
{XE: r(u) If(x,u) a}.
Different continuity of the above multifunctions were studied.

Closedness of M and N was investigated by Sawaragi, Tanino,
Nakayama 119861, Penot and Sterna-Karwat 119861. Upper semicontinuity
of M under compactness assumptions was studied by Penot and Sterna-
-Karwat 119861 and under compactness combined with convexity assumptions
by Sterna-Karwat 119861. In a more general setting stability of
solutions to vector optimization problems was investigated by Dolecki
and Malivert 119861.
Here we investigate upper semicontinuity of the solution multi-
function M without any compactness assumptions on feasible solution
set. We formulate sufficient conditions for M to be upper semiconti-
nuous in the class of well posed optimization problems defined in the
next section.
3. Well posed optimization problems
From now on we assume the function f:X+Y to be continuous.

We start with the analysis of parametric problems (VOP)u of the
form
f(x) E:wE(f(r(u))
Let u E: U. We define a multifunction II :Y+X as
lI(a) = {xlf(x) = a} 1'\ r(u) .
Definition. The problem (VOP)u is well posed if
(1) N(U) " 0,

(2) for each a o E: N (u) the multifunction II is upper
semicontinuous at a o •
Here we use the classical definition of upper semicontinuity of a
multifunction. A multifunction n is upper semicontinuous (u.s.c.) at
54
Yo if, for each open set Q containing n (Yo)' there exists a neigh-
bourhood W of Yo such that n (y) E: Q for y E: W.
The following proposition gives an equivalent formulation of well
posedness of a (VOP)u.
Proposition 1. Let X and Y be Hausdorff topological spaces satisfy-

ing the first countabi1ity axiom. The following conditions are equiva-
lent:
(1) the problem (VOP)u is well posed
(2) N (u);i~, for each a o E: N (u) any sequence {x n }, xn E: r (u)\Madu),
lim f(x )=a , contains a converqent subsequence with the limit
n+oo n 0 A -
point belonging to r(u).
Proof (1)-+(2). Suppose on the contra~y that (2) does not hold. If
N(U)=O, then (VOP)u is not well posed and the implication is proved.
Hence, assume that N(.u)/O. Now, there are a o E:N(u), and a sequence
{xn }, xn E: r (u) \ Mao (u), n=1, 2, ••• , ~!! f (x n ) =a o such that {xn } does
not contain any convergent subsequence or contains a subsequence with
the limit point not belonging to r (u). In both cases, there exists
a subsequence {xnk} of {xn } such that
But x nk E: 11 (a. k ), where Ilk=f (X nk ) , and {ak} tends to a o • And conse-

quently 11 is not u.s.c. at a o
(2)-+(1). Follows directly form the definition.
Q.E.D.
On the analogy of the scalar case a sequence (x n }, xn E: r (u) \Mllo(U),
lim f (x ) =a for Il E: N (u), is called a minimizing sequence of the
n+oo n 0 0
problem (VOP)u. It is a simple observation that when r(u) is compact
then the problem (VOP)u is well posed since N(u);l~ and each sequence
belonging to r(u) contains a convergent subsequence.
Below we give an example of a vector optimization problem which is
not well posed.
Example. Let X=R, Y=R2, r=r (u) =R, f:R + R2, f (x) = (sinx,cosx). The set
f(r) is simply the unit circle of the euclidean plane and wE(f(r» =
=E(f(r» is the part of the circle contained between points (-1,0) and
(0,-1). The resulting solution set is of the form
M=M(u) = {x I (2n-1)'IT ~x~ (2n-1)'IT +2'

A 'IT n=O, ± 1 , .•• }
The sets are depicted in Fig. 1, 2.

55
FIG.1
• • • •
---- ----
' . . . . . t---- ------
/
M-M(u)
FIG.2
By choosing x n ' xn=(2n-1)TI-~n' n=1,2, ... , where lim Bn=O we obtain

a minimizing sequence which does not contain any coR~erging subsequence.
If the problem (VOP)u is well posed the set M(u) need not be compact as
shows the following example.
Example. Let X=R2, Y=R2, r (13.) = {(x,y) iY~-x}, f=id. Hence,

M(u)=N(u) = {(x,y) iy=-x}. The multifunction , is a functions,
'(r)=r, and , is upper semicontinuous at each r o =(r 1 ,r 2 ), r 2 =-r 1 ,
for , is a continuous function on Y.
If the problem (VOP)u is well posed then the sets
<lo s N(u),
56
are compact. This follows from the fact that for different a the sets
~(a) are disjoint.
4. Upper semicontinuity of solutions
Before we prove the main result we recall the definition of a lower

semicontinuous multifunction which will be used in the sequel. A mUlti-
function Q is lower semicontinous (l.s.c.) at Yo if, for each open
open set Q which meets Q(yo)' Q n Q(Yo)~0, there exists a
neighbourhood W of Yo Q n Q(y)~0 for y E: W. A multi-
such that
function Q which is u.s.c. and l.s.c. at Yo is said to be continuous
at Yo'
Theorem 1. Let X, Y, and U be Hausdorff topological spaces

satisfying the first countability axiom.
If
(i) r is continuous at il,

(ii) f is continuous on X,
(iii) the problem (VOP)u is well posed,
(iv) N is u.s.c. at il,
then the solution multifunction M is u.s.c. at u
Proof. Suppose that M is not u.s.c. at u. There exists

a closed set F, F",M(u) = 0, and a sequence {uk}' lim uk = U,
such that Fn M(U k ) ~ 0, i.e. there exists a sequence kt~k}'
x k E: F 1'\ M( uk)' k =1 , 2 , •..
First, it may happen that there exists a subsequence {xk } of the
m
sequence {x k } such that x k £ r(uk ) \ r(u), m=1,2, . • . • Denote
m m
x k =z, m=1,2, • • . . If {zm} does not contain any convergent sub-
semquemnce or contains a convergent subsequence with a limit point
Xo i r(u), then r is not u.s.c. at u contradictory to the assump-
tion. Consider now the situation when {zm} contains a convergent
subsequence {zm} with the limit point ZoE: r(u). Without loss of
n
generality we take zm =zm' Hence, lim zm=zo' Aand by the continuity
of f, lim f(zm)=f(zoY' Clearly, f~~oo) E: f(r(u». We show that we
m+ oo
o
also have f (zo) E: N (u). _.Put f (zo) =Yo Since f is continuous and
r is l.s.c. at il, the multifunction fr, fr(u)=f(r(u» is lower
57
semicontinuous at u. Yo i N(U).
Now, suppose on the contrary that
By the definition, there exists yos f(r(u» such that Yo s Yo + intC.
By the lower semicontinuity of fr, there exists a sequence
wm E r(uk ), m=1,2, ..• , such that f(wm) E fr(u k ) and Yo = lim f(wm).
m m m+oo
Since Yo=Yo+c, c E int C, hence, for all m sufficiently large
f(zm)=f(wm)+c m, for some c m belonging to int C. Thus,
f(zm) E f(wm)+intC contradictory to the assumption that zmEM(Uk).
m
Now, the only possibility is that x k E r(uk ) " r (u) for all k suffi-
ciently large. Since x k i M(u), it must be f(xk)E N(Uk ) \ N(U).
Since N is assumed to be u.s.c. at u
the sequence f(x k ) contains
a convergent subsequence with the limit point as N(u). Without loss
of generality we can assume lim f(xk)=a. Hence, according to the
definition {xk } is a minimi~tng sequence, and by well posedness of
the problem {xk } contains a convergent subsequence with the limit
point Xo belonging to r(u). By the continuity of f, f(xo)=a,
and consequently xos M(u)" F, contradictory to the assumption
M(U)" F=f/l.
Q.E.D.
Remark 1. In Theorem 1 we consider the problems (VOP)u for which
N(u)=wE(f(r(u))) and this fact is essential for the proof. Now we can
ask a question whether Theorem 1 remains true if instead of (VOP)u
we consider vector optimization problems (VOP)~ of finding all
elements x of r(u) such that
fIx) S E(f(r(u))) ,
i.e. the optimal value multifunction is given by all efficient elements

of the set f(r(u)). Let us denote i t by N1 , N1 (u)=E(f(r(u))). The
answer to this question is negative. This can be visualized by the
following example
Example. Let X=Y=R 2 , U=R+, u=O, f=id and
r (u) = {(x,y) I -2~x,y~2} {(X,y) Iy~g(x,u)},
f
where g:RxR+R is defined as
max{ I x+1 I,u} for x~o
g(x,u)
1 -x+1 for x~o
For u=O g(x,O)=g(x), where

58
Since for each

g(x)
u, g(' ,u)
are closed. Now consider the solutions
r -x+1
'1
for
for
x~o
x~O
is a continuous function the sets

M1 (u) of the problem
r (u)
(VOP)~,
i.e. efficient elements of the sets r (u) . We have
M1 (0) {(x,y) iy=-x+1, -2~x~-1 , 1<x~2},
M1 (u) {(x,y) iy=-x+1, -2~x~-1-u 1 ~x~2} .
The multifunction M1 =N 1 is not u.s.c at u=O since zo=(1,0) belongs

to each M(u) and does not belong to M(O).
Remark 2. By examining the proof of Theorem 1 one can observe that

the upper semicontinuity of N at u is exploited only once. It is
used to deduce convergence of a sequence {f(x k )} to a limit point
a S N(u). What is in fact sufficient at this point is convergence of
the sequence {f(x k )} to a limit point belonging to fr(u). Hence,
instead of upper semicontinuity of N at u we can make use of other
assumptions assuring the above convergence, e.g. compactness of the
set fr (u) •
Basing ourselves on the above remarks we can formulate the following
results
Theorem 2. Let X, Y, and U be as in Theorem 1 .

If
(i) r is continuous at u,
(iii) the problem (VOP) G. is well posed,
(iv) fr (u) is compact,
then M is u.s.c. at u.
Proof. The first part of the present proof is the same as the
first part of the proof of Theorem 1. Only the case x k E r (uk) /"I r (u)
for all k sufficiently large should be analysed here again. Since
f(x k ) E fr(u), the sequence {f(x k )} contains a convergent subsequence
with the limit point YOE fr(u). But Yo actually belongs to N(u).
The proof of this fact is exactly the same as the proof of the same
fact in the first part of the proof of Theorem 1. Hence, {x k } is
a minimizing sequence and contains a convergent subsequence with the
limit point xo' Finally XoEM(U)/"IF, which leads to a contradiction.
Q.E.D.
For fully parametric problems (VOP)u of finding all elements
x E r (u) such that
59
f(x,u) e: wE(f(r(u»,u),
where f:XxU+Y and f(x,u)=f(x), Theorem 2 can be formulated in the

following way.
Theorem 3. Let X, Y, and U be as in Theorem 1.

If
(i) r is continuous at il,
(ii) f is uniformly continuous on r(u)xu,
(iii) (VOP)il is well posed,
(iv) fr (il) is compact
then the solution multifunction M is u.s.c. at u.
Proof. As in the proof of Theorem 2 we should consider only the

case where x k e: r(uk ) ~ r(u), for all k sufficiently large. By (iv) ,
lim f(xk,u k ) = y and y e: fr(il). By the same arguments as in the
k+oo 0 0 A
proof of Theorem 1, it must be Yo e: N(u). Now, we have
If(xk)-yol = If(xk,il)-yol If(xk,~)-f(xk,uk) 1+lf(xk,uk)-yol

and, by the joint uniform continuity of f, lim f(xk)=yo. Hence,
{xk } is a minimizing sequence and, by the welf+posedness of the problem
(VOP)u' {xk } contains a convergent subsequence with a limit point
xo. By the continuity of f, f(xo)=yo' and consequently xoe: M(u)nF
which contradicts the hypothesis M(~)nF=~.
Q.E.D.
5. Comparison with the scalar case
Consider now a scalar optimization problem (SOP)
inf{f (x) Ix e: r} ,
where f:X+R is a continuous function and r is a subset of X.

Together with a given problem we consider also a parametric problem
(SOP)u of the form
inf{f(x) Ix e: ru},
where r:u+x is a multifunction. The optimal value of the problem

(SOP)u is denoted by fr(u). The optimal solution set is denoted by
M(u) .
Definition. Let U e: U. The problem (SOP)u is well posed if

fr(u) is finite and the multifunction ~:R+X defined in Sec. 3 is
60
upper semicontinuous at u.
The above definition was used in Bednarczuk 119841 to formulate
sufficient conditions for upper semicontinuity of solutions in scalar
case. The next theorem is a scalar counterpart of Theorem 1.
Theorem 4. Let X and Y be Hausdorff topological spaces satisfying

the first countability axiom.
If
(i) r is continuous at u,
(iii) the problem (SOP)u is well posed,
then the solution multifunction M is u.s.c. at u.
Proof. The proof of this result can be found in Bednarczuk 119841~

There is a full coincidence between the assumptions of Theorem 1
and that of Theorem 4 except the assumption (iv) of Theorem 1. The
reason is that in the scalar case continuity of the multifunction r
entails upper (and lower) semicontinuity of the optimal value multi-
function N. In the general case it is not true as shown by Nakayama,
Tanino, Sawaragi 119851, and Penot, Sterna-Karwat 119861.
6. Conclusions
Most of existing results on upper semicontinuity of solutions in

vector optimization are based on the assumption that the set r(u) is
compact. If r(u) is compact, then the optimal value multifunction N
is u.s.c. at a (proposition 2.8, Penot, Sterna-Karwat 1986), and as
was mentioned in Sec. 3 of this paper the problem (VOP)a is well posed.
Hence, in the compact case Theorem 1 reduces to the results of Penot,
Sterna-Karwat 119861 (Y-Hausdorff) and the results of Nakayama, Tanino,
Sawaragi 119851 (Y-finite-dimensional). An approach to stability
questions, based on the idea of well posed problems was also presented
by Lucchetti 119861.
References
Bednarczuk E. (1984) - Continuity properties of solutions in perturbed

constrained minimization problems, Ph.D. thesis.
Borwein J. (1983) - On the existence of Pareto efficient points,
Mathematics of Op.Res. vol.9, n.1, pp 64-73.
Dolecki S., Malivert Ch. (1986) - Polarities and stability in vector
optimization, this volume.
Lucchetti R. (1986) - Well posedness, towards vector optimization, this
volume.
Nakayama H., Tanino T., Sawaragi Y. (1985) - Theory of multiobjective
optimization, Academic Press.
Papageorgiu N. (1985) - Pareto efficiency in locally convex spaces I,
Numer. Func. Anal. and Opt. vol. 8, n. 1, pp. 83-116.
Penot J-P., Sterna-Karwat A. (1986) - Parametrized multicriteria
optimization, continuity and closedness of optimal multifunctions,
J.M.A.A. vol. 117.
Sterna-Karwat A. (1980) - On existence of cone-maximal points in real
topological linear spaces, to appear in Israel J. of Math.
Sterna-Karwat A. (1986) - A note on the solution set in optimization
with respect to cones, Optimization vol. 17, n. 3, pp. 297-303.
CONVEX CONES. nlNlnAUTY NOTIONS. AND CONSEQUENCES
J.n. Borweln
Dalhousie University
HaUfax. Nova Scotia
1. Introduction
2. Cones with structure
3. nlnimaUty notions
4. Lattice complementarity
5. Conclusion
? A VECTOR OPTlnlZATION PROBLEn IS ANY OPTlnlZATION

PROBLEn USING PARTIAL ORDERS IN A NONTRIVIAL WAY?
This paper Is a sUghly reworked and extended version of my

"state of the art" tutorial. A blbUography has been added. but
prlmarUy of my own papers where more comprehensive
references are to be found.
1. INTRODUCTION. nost problems to which functional analysis

appUes can be modelled by
(VP) h(b) := mtns{f(x): g(x) Sp bl
where X. Y. Z are vector spaces. Sand P are appropriate

closed convex cones. and f :X-tY.S and g:~Z.P are convex.
Lipschitz. differentiable or otherwise. [y ~P b means b - YEP]
THREE THEnES
a) Exoloit soecial structure: If it always works it has been

discovered (e.g. Slater"s condlt1on).
63
bO Exploit the trade off betveeu topololY of X.g and order

structure of l.P [and betveen f.X and U.S for obJecl1ve
structure and for attainment]: We write the constraint

mult1funcl1ona11y as b E H(x):.=I(X) + P to l11ustrate. If P ... X
then g Is Irrelevant; if P - 0 then g Is critical (See Borveln
(Ola.b).)
bit) Especia11y exploit latl1ce structures: e.g. duaUty in

HUbert space betveen A and ( • ) orthogonaUty: XAY'" 0 111
x1.0. y1.0. and <x. y>= 0 (See Borvein(05). Borveln and Uost
(04a). Borveln and Dempster(07b). and below.)
c) Exploit scalar results: Firstly. the constraint structure of

(VP) Isnat changed by objective changes. Secondly. many
(most?) vectorial results rely on the underlying totany
ordered scalarlzed problem (e.g. dlfferenttabUitg results for
convex operators relhlclI to scalar results. See Borveln
(02a.06b) and below).
Nov ve proceed vlth:

2) An outltne of useful special cone structure
3) A sketch of mlnlmaUty notions
4) A more careful case study of lattice complementarity.
2. CONE STRUCTURE IN BANACH SPACE. We restrIct ourselves

when convenient to closed. painted cones in Banach space
where the situation ts cleanest. The general case can be found
tn the references.
2a. Order Intervals
2.1. Order Interyals are defined by [a.b]s :- Ic: a ~s c ~s b}.

Normalltg Is the requirement that if 0 ~s an S5 bn and
bn-) 0 then an-) O.
64
lormanty of the cone ties the topology firmly to the order

and Is often a necessa.,. restrlcUon.
2.2. PROPOSITIOI
1) Intervals are bounded Iff S Is normal In In this holds
ll1 S Is pointed (sn-S - Ol-
11) lorm and order bounded sets coincide Iff S Is normal
with nonempty Interior.
111) S Is normal 111 S Is weakly normal 1U. S+ Is
generating (s+- s+ • y.) U1 s+ Is a (strict) Hone: for
some 8)0
t EY· =--> t =It-1Z' 1I1tll!1. 1I1zll! 1

IItll S Ii lt ES+, lzES+
(and dual1y replacing S by s+). Strict Hone/normal cone

duaUty (Peresslnt(61)) is nicely explained via adjoint convex
process duality (Borvein(B6c)).
2.3. ABSTRACT 0111 THEOREM (Peressini(61)) If a decreasing

sequence (net) In a normal cone converges weakly then It
converges In norm.
[The classical result follows stnce pointwise convergence to a
continuous limit ImpUes weak convergence.]
2.4. lonnallty In acUon: S Is normal 1U. when g Is continuous

at x.. f Is &-convex and ftnlte It x.. and
f(x) Ss g(x)
then f is conUnuous at x.. [The ·only If· direction is

given In Borveln(B2a). The -If- dlrectton makes a good
exercise.]
65
2.5. Aall1cllttlnl (Borveln(8211.86b)): (I) Aigebrllic lubgradlentl

of a conUnuous convex operator Ire conttnuous. (II) Centtnutty
at one point of a convex operator translatel throughout the
core of the domlin. (itO Order-bounded convex operlltors are
conttnuous.
2b. nonotone sequences and nets
2.6. 0 S hal the monotone net (sequence) aropertg or nNP

(nSP) If decreallng netl (sequences) In S have Infima.
it) S hal the (lequenttllll Danlen pro"rtg If the Inftmum
II also the topological limit of the net (sequence).
110 5 II order complete If every non-empty set In 5 hal
an Inftmum.
2.1. PROPOSITION (Borvein(82a.84b))

o S is Danien if S has veakly compact intervall or if S
Is veak* closed vlth veak* compact Intervals.
iO If S admits a strict monotone funcUonal then
sequenUIII order completeness (sequenUal Danien
property) impl1es order completeness (Oanien property).
2.8. Example: Intervals in c. +(10 or 1,+(1'0 (Up<m) are norm

compact. In Lp +[0.11 (Up<m) they are veak compact. In
L. +[0.11 intervals are v* compact. In C+[O.11 they are bounded
but not v* compact.
2c. Bases
2.9. S hilS a (cloled. respecUvely bounded. veak compact) uu.

B tf B is (cloBed. respectively bounded, vellk compllct) and
convex and S = Ut)O tB vith 0 tclB (uniquely if desired).
66
2.10. PROPOSITION (Borvetn(B2a))

1) S has a base U1 S admits a strIctly posItive linear
funcUonal t (t(x) >0 if XES\O).
tt) Every separable closed cone is based.
un S has norm intartor iff S+ has a vti-compact basa.
(Duany: S+ has Mackey InterIor 111 S has a veakly
compact basa"
iv) S has a non-ampty norm intarior iff S+ 1s pOinted and

locany vti-compact.
(Any locally compact S can be represented as S =KeL
vhere K has a compact basa and L is a finne dimensional
vector space. Borvein(BS)"
v) Intervals ara bounded or compact vhenever the base is.
vO S is Daniell vhen S has a closed bounded base.
Adjoint procass duality nicaly axplains iii).
2.11. Example. i) The base in It(IO is bounded; as the dual of

all the convergent sequences c+(lO It Is v· compact; vlth
B :-tXElt: Ex n- 1} in each case.
li)The BishoD-Phelps cone 5:= cone(B(O.1).1} in XxII has a
bounded base ilnd so Is oanlen. In thIs order.bounded and
order-bounded sets coincide. This yields Ekeland's theorem.
(See Borveln(81)"
67
2d. Lattices
2.12. 0 A vector lattice is an ordered vector space in which

aVb:= max(a,b}:= sup(a,b} always exists. Then
Ixl: = max{x.-x}. x+:=max{x.O}. x-:=(-xt.
iO A B.acb (or comolltl Rilsz) lattice is a Banach space

vector lattice in which the norm is
absolute Ilxll == Illxlll
and
mpnotone 0 Ss x Ss y -=> IIxll ! lIyll.
UO A Hilbert latUce is a Hilbert space vector laUice in

which the inner product induces a Riesz norm.
2.1l. PROPOSITION
o A Danien laUice is order complete.
it) An order complete space is a lattice iff the cone is
IlneraUng.
Ui) A closed cone with base Is latticlal iff B is a
(ChoqueU simplex. In Un eVlry closed pOinted cone Is
Danlen but only the usual simplicial cones are lattices.
(Sel Plressini(fi7).)
tv) A (countably)-arder completa Banach lattice is Daniell
H1 it contains no lattice copy of 1m (10. (See Schaefer
(74), Undanstrauss and Tzafrtrt(7B).)
v) An absolute lattice Is Rlesz 111 Intervals are bounded.
(See Borwein and Yost(B4a).)
vt) A Banach lattice is Daniell (order-conUnuous) iff It
has weak compact intervals iff intervals are strongly
exposed. (See Borwein(B6b). also, Undenstrauss and
Tzafrlrl(78)J
68
2.14. ADDltclt'on: (Borveln (B6b))

Let X have a Frlchet (6ateaux) smooth norm. Let Y be an
order complete lattice. Every ordar-bounded S-Convex operator
f:x-..y.S Is Frechet (6ateaux) differentiable densely iff S has
veakly compact Intervals.
Reason. When S is DanleU f inherits the differentiabUity

properUes of •• f. The scalar result is In Borveln and
Preiss(B71). When S is not DanleU ve embed 1m.
A STROIl6LY EXPOSED INTERVAL
rp - 1(8)
2.15. A Hilbert space vector lattice is absolute iff xAy =0

forces oc. II> - o. This Impl1es that s+ c S. A HUbert space
vector lattice is Rlesz iff s+- S iff
XAy • 0 <==> OC. y>. 0 and XES. yES.
These and many other relationships are studied In Borveln and

Yost(B4a).
69
2.16. Example: 1) lz[O,ll is a Hnbert lattice In the usual

pointvise order 5. Then 5 = 5+ and 5 has veakly compact
intervals, an unbounded base, and empty interior.
11) ACz[O,11 (the space of absolutely continuous
functions vith derivatives in lz) is a Hilbert space vith the
tnner product
<x .Y> := x(O)y(O) + I~·(t)y·(t)dt
o
vhich uses the derivatives. The point vise order 5 Is an
absolute lattice order (since Ix·(t)1 -lxl'U) a.e.) vhich has
1E int 5 and unbounded order-incomplete tntervals.
Then 5+ has a veak compact base and so is Daniell. Nov.

5+ consists of positive Borel measures on [0.11 and Y:== 5+- s+
is only dense in X. Thus x.s+ is not a lattice, but I.S+ Is a
Daniell lattice.
These distinctions are critical tn study of vartaUenal and

control problems. (See Borweln (B6a).)
2.17. Tha Oaniall proparty ·charactarlzas· tha vaUdlty of tha

max-formula for convex operators:
vllare iJsf(x):= (TEL[X.Y]: T(h)!S f(x+1I) - f(x)."h). (5aa

Namath(85). and Borvein (84b).) Pracisely. tlla diractional
der1vaUve d+f(X; II) exists for every convex operator vlth
values In Y.S ll1 S is countably Oanten. and tha Danian
property ensures (MAX). Thus if 5 admits a strictly monotone
functional (as vhen Y is separable and normad). (MAX) and
Oanien coincide.
Order completeness characterizes the full Hahlt:8anach
extension proparty and most othar similar properties. (Sea
Borvein(82b).)
70
3. n,",nALlIU NOTIONS. Ihere are really only tva concepts:

dlmlnaURI and non- domlnat,d pllnts (If a set Ir Its
closure).
3.1. n An s-eareto (mInima" efnclent. non-domlnated) Pllnt

of a closed set C is a point x in C such that some section
(C - x)n S saUsnes
(C-x) n -5 c S
-s
11) An 5-least element (strIng minimum. dlmlnaUng) poInt
If iI cllsed set C is a pltnt x In C vlth
(C - X) c 5
EXISTENCE OF PARETO POINTS is EQUIVALENT to the AXlOn OF

CHOICE eyen if C is a dual 8anach lattice unit ball and S the
dual Irdering. (See 8Irveln(83)J Geometricallu eyeruthlng Is
plausible and v· closed.
3.2. PROPOSITION (Barveln(03))

Conditions for existence of minimal pOints in any locally
conyex setting include:
1. C contains a closed section vlth iI lover bound and S is
Daniell; or
2. C Is cllsed and bounded vhU, 5 Is blundedlu Irder
complete (toplloglcal1y bounded monotone nets haye
infima) and Daniel1; or
3. C contaIns a compact secUon.
71
l.l. blJ.: Case 2 applies in Lp (1 ~ P <ID). while Case 1 applies

to the Bjshop-phelps cone if cc Xxi has jnf{r: r E C) )- ID.
and yIelds Ekeland"s theorem In Banach space. Also. the
positive semi-defin1te operators on HUbert space form a
Daniell cone in the Weak Operator Topology. (As a result.
square roots exist. See Borwein(8l).)
l.4 PROPOSITION (Borweln and Dempster(B7b»

Suppose that C and S are such that minimal paints exist
for all sections of C. Then a strong minimum exists for C
1U. C has a unIque mInImal poInt.
Thus conditions for strong minima reduce to verifying

uniqueness In the Pareto setting.
l.S. An S-yeak mInImum of C Is a poInt x of C such that
(C - X) n inH-S) C ~~~~,~,S"";~,,,;",,;,,~~S~
"""
~"""
"" "' \ C '""""
" '"/;
" ',,
" " ' " """" "
",,,,, " " "
',',',"',,',',', ,',','... ',','".',','
"""""""'"
X
These paints are exactly K-efficient pOints of C. where
K :- {OJ U int S. At the expense of using non-closed cones theIr
theory can be subsumed. (See Borweln(80b).) Also. theIr only
real virtue is the ease of analysis. In the picture. zero is a
strong minimum while x is a domin4/ed weak minimum'
Compare ProposlUon l.4.
l.6. Proper Efficiency: An S-efftcient point of C Is properly

efficient if
72
cone(C-x) n -9 c S
proper
One may similarly define 10cally prooer efficiency to

study first order condlttons. Any S-veak minimum is {O)UintS
-proper. (See Borvein(77.BOb).) AlI these definitions simplify
vhen 5 Is pointed
3.7. SCALARIZATIUM LEnnA. Suppose that S has a veak

compact base B. that C is closed and convex. and that
XEC.
Then x is properly efficient iff there is a strictly

S-posttive functional t such that
t(x) • mlnCEC t(c).
If {O)U int 5 is used the base condition is superfluous. In

general it is necessary.
Proof. strict
73
I claim that almost all knovn vector duality thearies can

essentially be derived by this or other types of scalarizatton.
l.B. Examole: Consider again
(YP) h(b):- mtnstf(x): U(x) !p b} .
Assume any scalar reuularlty condition vlth b := 0 and ISS. . .

the Scalarization Lemma appUes. Suppose that x, is feasible.
Then f(x,) is a properly efficient solution i11 there is a
conttnuous Unelr operator T :Z..... Y such that T(P) C S.
T(g(x.» .. 0 and
(PE) f(x.) Is properly efficient for f(x) + T-g(x).
J!nI(: 1) There Is 1 vlth l(P) 1 O. 1-U(X,) =0 and for al1 x
11) Select s In 5 vith t(s). 1 and set T(z):- l(z)s. Then

.-T -1 and the result foUovs.
Hahn4manach theory gives the corresponding strong dualtty

result and constraint quaUf1cattons are unchanged from the
scalar case. As alvays this looks more like the scalar result:
If 5 Is order complete. subgradients of the value functton h
correspond to Lauranue mulUpUer operators of (YP). See
Borveln(82i1.82b.84b).
l.9. ggq: Is there an Improper analogue of (PE)? This vould

Intrinsically not follov by scalarizatton. The only type of
characterization of Improper paints vhich I knov Involves
lexicographic multipUers. Thts Is treated in Borveln(BOa). The
difficulties are illustrated by the next example.
74
3.10. Example: 0 An onlered normed space can be given a

monotone renorm 111 the cone Is normal. To see this suppose
that S Is normal, define a nev narm by
IIxliS =- supOt(x)l: IItll ~ 1, t E S+},
and use Proposition 2.2.i10 to shov that this Is an equivalent

norm.
iO Nonetheless many closed (laUice) cones .mlt no
striCtly lRanetone functtonals and In parttcular no striCtly
monotone renorms. Consider the space I.(n of bounded
functions on r. vhere r Is uncountable. Let Y be the subspace
spanned by the constant function 1 and the functions of
countable support. One verlftes that Y Is countably order
complete but not order complete. [Also Y may be vleved as a
continuous function space C[O] vhere 0 Is countably Stonlan
but not Stonian. See Schaefer(7S).] Directly. or from
Proposition 2.7.i0 ve see that the cone of nonnegative
-I.
bounded functtons In Y admit no strlcUy monotone functtonals.
In particular S (n+ admits no strlcUy monotone
funcUonals. Thus no scalarlzaUon result applies. Despite this.
S Is an order complete Banach laUlce cone vlth tnterlor. Thus
It Is a significant problem to characterize some or all of the
5-properly enlcient paints of subsets of I. (n.
4. LATTICE ConPLEnENTARITY PROBLEnS. nany opttmlzatlon

problems can be reduced to solving an appropriate
complementarity problem: for F:II->II solve
x 1. o. F(x) 1. 0, <X,F(xb =0 . (1)
(The ordering Is coordinate-vise.) This problem has milny

generaUzaUons. some fitting some not. In vhlch II Is
75
replaced by an ordered vector sp.ce .nd the inner product is

repl.ced by the .....ropri.t. btllne.r p.irinl. On. c.n.
however. rewrite (1) as a lattice or order cOlnplelnnt.ritg
problem: for F:.I _).1 solve
Inin {x. F(x)} - 0 . (2)
.Iso .dmlts • •y extensions In wlltch .1

(The lattice Ininilnuln is t.lcen coordln.tevlse.) Thts problem
Is replaced by a
v.cter lattice and the -Inln- Is r.plac.d by the .pproprl.t.
lattlc. Ineet.
rarely b••n expl1cltly studied. .v.n In .1.

lonetheless. the I.ttice calnplamentarity problam 1Ia.
I will now outlln.
the theory of the lattice complalnantarlty problaln and vln
tra to Indlc.te vhy (2) oftn le.ds to • Inore .ppropri.te
Inodal th. doas the traditional complement.rlty problem.
"oreover, m.ny Important but rather obscurely defined m.trlx
cl.sses (•.g. Leontl.f-matrlces. Z-matrlc.s. ft-matrlces.
P-m.trlces) are .fforded simple l.ttice-theoretlc descriptions.
The matrix theory can be found in Berman and Plemmons (79).
4.1. Example: (Linear Programming Duality) Consider a standard

linear programming pair in finite dimensions written in matrix
form:
(LP) p: - In.x{cx: Ax ~ b. x~O}

(LO) d: = Intnlby: A·Y ~c, ylO}
.nd p = d vlth .tt.inment is equlv.lent to solving eltller

(LCP) T(z) + q ! O. z! O. <T(z) + q.z> - 0
or
(DCP) mln{ T(z)+q, z} = o.
vh.re
76
ner• •naralil. consider. 8.n.c" spac. X•• Unear IR.p

T:X-> X·•• c.v.x c•• K In X. _d • du.l co. K+ In X· . We
trg solving
(LCP) T(z) + q ~ K+O. z !KO' <T(z) + q.z> 0.=0
Equ.ng. consider • 8.n.c" I.Ulce X vU" positive c.e K

.nd T:X -> X. W. trg solving
(OCP) IRlnK, T(z)+q. zJ =0 0.
Ag.ln. a 8.n.c" l.ttlc. Is a compl.t. norm.d sp.c. wit" a

c.....Ubl. l.ttlc. structur. such t ...t Ixl S Igl ItnpU.s t ...t
IIxll ~ 11111.
4.2. EX.IR,Ia: cun. L,(f). 1,(1). or Co(I). aach vlth K as tha

polntvlsa anI.rlnl. Is • 8an.ch l.ttlce.
4.3. nET A-THEOREn (Lindenstrauas _d Tzafrlri(18))

If .n ·idanttty· "aids In • it "aids In .ny 8.n.c" lattlc•.
Wh.n do the tva p.-.l.lRs calncld.? Precls.ly vh.n X • X·

.nd K - K+. Thla ".ppens .x.ctly v ...n X la • Htlb.rt I.Uic.
er aqulv.l.ntly vhen
for x. y! °( See 8erwaln .nd DaIRpst.r(81b)). Ag.ln. the

only Hllb.rt laUlces are Lz(Jl) (Jl I R. . m.asur.) In the
polntvis. order (Sc....f.r(14)). Unfortun.taly. no Sobelov
spIC.. of v ..lch ACz Is a prototgp.. Is I Hllb.rt laUice.
77
4a. Least Elements a.... Pre-Leantief "atrices
A lelst el...nt of tile feasible set
F(q): = Ix 1 0: T(x)+q 1 OJ
Is a feasible element which is smaUar than an others. Thus a

least alament is precisaly a K-strong minimum of tha faasibla
set. Dne of the rationales for their study Is that minimtzation
of Iny strictly posltlva linear functional ovar F(q) will find
tha lalst alamant if it axists Ind so ana can hopa to reduca
solution of appropriately structured complementartty problems
to lInelr programming. (See Berman and Plemmona(1g).)
4.4. THEDIEn (Borvein and Dempster(87b))

Lat T:X -) Y be continuous and linear batvaan Banach
Ilttices. Suppose X has veakly compact intervals. Then
every non-emptg F(q) has a least element HI. T Is 1UD (5)
[also called (H+)]:
(5) T(XAy) ! T(X)A T(y) for x.y ! O.
leason:<- (1) Closed convex sets In X have minimaUPareto)

points. (2) (5) forcas unlq_nass. (3) .0'41 usa Proposition 3.4.
=) Consider x _d y with xAy = 0 a....
F(q):={z 1 0: T(z) - T(X)A T(y) 1 Ot.
(1) The least element is O. (2) Thus T(X)A T(y) s o.

(3) By lInelrity (5) holds.
In 1ft type (5) or (H+) matrices are those vith It mast one
positive entry per rov and Ire called pre-Leonttef.
78
4b. P natrices and Unigueness
T:X -) X is !UIUt (P) if
T(x)vx ~ 0 ~ T(X)AX ImDUes x = o.

This says T can not -raverse the Sign- of x. In In this holds
iff T has positive principal minors.
4.5. THEOREn (Borveln and oempster(B1b))

o Every feasible (OCP)(q) has a unique solution exacUy
vhen T Is type (Pl.
10 noreover. In In. If T Is type (P) then (OCP)(q) Is
alvags feasible.
The II result is due to Gole. Eoves ond others.
R.osons: (0) direct 10tUc. menlpuleUon.

(b) is more deUcate as w. now indicate belovo Th.
arguments cln be feneved in detlll In Barveln Ind Dempster
(B1b) and Berveln(BS).
4bO Salylna (nenUneild comDlementilr1tu problems

bU vilriilUonal t.chniques
Finding a solution to x E C ilnd
(VI) <f(x). x - u> ~ 0 for 111 y E C
is equivalent to solving
(EQ) 0 E F(x) + ilc(x)
vhere IC is the indicltor of C. In 1ft it Is equivillent to
(FP) x - p[(x + F(x))

79
vtth Pc tha EucUdaan naaralt potnt prOjection. If F tl

continuoul and C is compact and convex this is possible from
Brouwer"s theorem. (ActuaUy. Brouwer Is ROV equivalent to
(VI): ule F:- f- I and set y:- f(x) E C. Thts gives f(x) = x.)
4blt) Application to (lLCP) in ,8

"'a solve tha (VI) for Kn(x :lIxll S m}. This yields x E K vith
(*) <T(x)+q. y - X> l 0
for aU y in K vith lIyll ! m. If ve can shov that IIxll < m tt

will fonov that <T(x)+q. x ) = 0 and that (*) holds for an y In
K. This is the desired conclusion. "'e assuma tha folloving
lattlca coarctvtty condition: there is a constant c ) 0 such
that for all x In K
Nov va reean that in a dual normed lattice the foUoving

holds: for f In X· and x l 0
Than (.) gives
Also
HeRce (LC) shows that

80
and IIxll ~ IIqll/c as needed. Finally. in II, type (P) implies

(LC) ilnd so ensures solvilbnUy.
Thus (LC) milY hold for non-coercive matrices.
4c. Least Element Solutions to (OCP)
T:X -) X is 1!lU (Z) If
XAy = 0 ImoUes T(X)AY S O.
T:X -) X is .!!11m (11) if
Tx ~ lx for all x! 0
U.e. T = 11 - p. P positive and Unear).

In II both mean that only positive entries ilre diagonal In
generill they d1ffer but:
4.6. THEOREH (Borweln and Dempster(B1b))

Let T:X -) X be Uneilr conttnuous yUh X iln order complete
(e.g. reflexive) Bilnach lattice. TFAE
U) T is type (ll. un T is type (ll).

(iii) Each feasible (OCP)(q) possesses a least element
solution.
In I" thIs Is due to Cottle and Ve1nott.
Reason: (a) In iI countably order complete normed lattice (n =>

un yUh 1:- IITII. (This fans In a Frechet lattice: use (TX)I =
nx.. J
(b) In any order complete vector lattice: if T Is type
(11) consider
81
Then f is isotone. flO) l O. and f(x) S x iff x Is feasible.

Use Tarski's fixed point theorem (1955). [In essence this Is a
value improvement method: 1"(0) converges Increasingly to the
solution.]
(c) If UIO holds we start wltll xAy - 0 and iteratively
solve (OCP)(qn) with qn: - xn - T(Xn) and x.:- x. This yields a
decreasing sequence (xn) with Umlt 0 and T(Xn-Xn+1) ! x n.
Then T(x) ! nx + T(xn) so that T(X)Ay ! o. This is (t).
4.7. Remarks: (a) In In (or any atomic lattice) this Iteration

can be done in one step.
(b) Operators which are both type (P) and (1) are said
to be .lUlU! (K) (the only (OCP) solution is the least element).
In I· these are the non-singular M matrices:
(*) T:- ll-P with 1 l r .,(P). (r., the spectral radius)
They may be characterized in II by
T(X)AX ~ 0 ImpUes x S o.
Moreover. with f as above. in any Banach space (*) ensures

that the Contraction principle appUes. This holds without any
order completeness and estlbUshes the solvability of (OCP)(q)
for an q.
(c) In C[O.l1 feasible type (ll) problems can fan to have
solution.
4d. An Order Duality principle
4.8. THEOREM (Borwein and Dempster(87b))

Let X be In order complete Banach laUice. Let T:X -> X be
continuous type (K) (te. (P) and (l)). For f.g. and h in X
consider solving
82
Xo:- LE(x: Tx 1. qA Th. x 1. f}

yO:= 6E(y: Ty i qVTf. x ih).
with f S h. Then the soluUons extst and
Reason: The corresponding (OCP) are solvable. Now lattice

manipulation takes over.
4.9. coronaru (Dual esUmates) Consider
Xo:= LE(x: Tx ! q. x ! f} .
Then
TKo S qVTf .
Reason: Apply the Theorem with h:- x •.
This is relevant in various obstacle problems and POE

formulations. It also Inustrates an entirely new duality
pa1r1ng wh1ch 1nvolves no dual spaces. A s1ml1i1r mixed
topologlcal-Iatticlill formulilUon is due to nosco.
5. CONCLUSIONS. We suggest that lattlca-theoretic _lysis is

worthy of attenUon tn most optimization problems In which
positivity plays a slgnUcant role:
(a) It ratses the posslbUlty of avoiding messy dual
spaces such as L.· .
(b) It aUows us to model conUnuous or tnflntte horizon
input-output models tn more reasonable setttngs than Hilbert
83
space (e.g. Loo' [[0.11. c.). The open Leonlief model leilds to iI
type (l)operator: T: • I-A. A a technology miltrix.

(c) It yields new insights 1n old settings; much Df the
theory of the classlcill (L[P) Is reillly about the (O[P).
(d) In general when theorists hilve abstrilcted optimization
models. they have used totally ilrbitrary convex cones as
orderings. As the coerctvlty ilnd other results illustrate much
more can be said about reasonilble orderings (e.g. latticial.
compactly based. with interior. generating); but this Is rarely
tilken into account.
84
6.REFEREIICES
(79) A. Bennan and R.J. Plemmons. ~tlN IllllriCIIII is

1IIItlllllhllllllltlcilI Sci6ll&ft'(Academlc Press. lew Yorlc.1979)
(87a) J.H. Borwein and D. Preiss. -A smooth vartational
principle and applications; to appear.
(87b) J.H. 80rwein and H.A.H. Dempster, -The linear order
complementarity problem, - to appear.
(86a) J.H. Borwein, -Epi-Lipschitz-like sets: Theorems and
examples, - Nonli/lllar Analgsis.- Theorg .Ife/hods and
Applications. (1986).
(86b) J.H. 80rwein, -Generic differentiability of order-
bounded convex operators, - Journal Aust /fath. Soc. B.
23(1986), 22-29.
(86c) J.H.80rwein, -Norm duality for convex processes and
applications, - .JOTA. 48(1986), 53-64.
(85) J.H. 80rwein, -Alternative theorems for complementarity
problems,- in Inrinitll Programming (Springer, New York,
1985).
(84a) J.H. 80rwein and D.A. Yost, -Absolute norms on vector
lattices,- Proc. Edinburgh /lath. Soc., 27(1984), 215-222.
(84b) J.H.80rwein, -On subgradients of convex operators,-
/lath. OpIlrationsrorschllng. 15(1984), 179-191-
(83) J.H. Borwein, -On the existence of Pareto efficient
points, - /lath. or Opllr.. Hils.. 9(1983), 64-73.
(82a) J.H. Borwein. -Continuity and differentiability
properties of convex operators: Proc. london /lath. Soc..
44(1982), 420-444.
(82b) J.H. 80rwein, -On the Hahn-Banach extension property,-
Proc. AI1IIlr.. /lath. Soc.. 86(1982), 42-46.
(8Ia) J.H. 80rwein, -Convex relations in optimization and
analysis,- pp. 335-377, in: 61l/lllralizIld ConvllKitg in
Optimization and Economics (Academic Press, New York.
1981)
(8Ib) J.H. 80rwein, -A Lagrange multiplier theorem and a
sandwich theorem for convex relations. - /lath. Sclllld.,
48(1981). 198-204.
(80a) J.H. 80rwein. -Lexicographic multipliers,- J. /lath. Anal
Appl.78(1980), 309-327.
(80b) J.H. 80rwein. -The Geometry of Pareto optimality,-
/lath. Oplffa/ionsforchung. 11(1980), 235-248.
(77) J.H. Borwein. -Proper efficient points for maximizations
with respect to cones. - SIA/I J. Control & Opt. 15(1977),
57-63.
85
(18) J. Undenstrauss and L. Tzafrtrt .. CllI66iul""" II/IIICII6

FlIIIClill/l ~(Sprtnger. New yort. 1918)
I~
(85) A.B. Nemeth. -About subdifferentiabllUy of convex
operators-. to appear.
(61) A.L. Peressini .. I1rtItInHIlfI/ItIIBgic.' rtlCltIr $JIII&t16 (Harper
and Row. New Yort. 1961)
(15) H.H. Schaefer. BlIIIIICh lllllic.1I ad PosiliVil OpllrlllllrS
(Springer. Nev York. 1914)
EVALUATION FUNCTIONALS ARE THE EXTREME POINTS OF A BASIS
FOR THE DUAL DE C~ [a,b]
A.R. da Silva
Instituto de Matematica
Universidade Federal do Rio de Janeiro
Brazil
Abstract
We show via a Sobolev's imbedding theorem that the evaluation functio-
nals are the extreme points of a basis for the dual cone of the cone
formed by real non-negative continuously differentiable functions on a
bounded closed interval.
1. Introduction
Let C [a,b] be the real vector space formed by all real continuous
functions defined on [a,bJ.
As it is well-known if we equip C[a,b] with the norm
1/2
(f € C[a,b])
the natural ordering cone
K : = {f € C [a,bJ: f::: O}
has no inner point (for a proof see e.g. BOHL [lJ page 38) . In this
paper we show that if we consider Cl [a,b] be the real vector space
formed by all real continuously differentiable functions defined on
[a,b] equipped with the to 11.11 2 likened norm

87
(g € CI [a,bJ)
then it follows from a Sobolev' s imbedding theorem that the natural
ordering cone
does have a non-empty interior. As a consequence we can find a
w*-compact basis for the dual cone Kl* with the evaluation functio-
nals as extreme points. All concepts not explicitly defined can be
found in HOLMES [3J or SMIRNOW [6J.
2. The natural cone of cl[a,bJ has a non-empty interior
Let w; [a,bJ denote the real vector space of all generalised derivable
functions on the interval [a,bJ possessing generalised derivative in
L 2 ( [a,bJ) equipped with the norm
II f II ~ ,= \ r: (f2) dA + J: (Df) 2 dA 1112
where Df denotes the generalised derivative of f and A the Lebesgue
measure on [a,b] .
As it is well-known in the case f is a continuously differentiable
function on [a,b] the generalised derivative of f coincides with
the usual derivative of f on [a,b].
Lemma 2.1
Let ( Cl [a, bJ, II. Ill) (resp. Kl ) be the real n<lrmed linear space (resp.
the natural ordering cone on cl[a,bJl defined in the section 1.
Then the interior of Kl is non-empty.

88
Proof:
According to a Sobolev's imbedding theorem (see e.g. SMIRNOW [6J page
306 ff.) there exists a compact linear operator
1jJ: (W;[a,bJ, 11.11;) -(C[a,bJ, 11.11 (0 ) ,
where Ilh II 00: = max{ Ih (t) I : t e [a,b]}
which imbeds (W~ [a,b], II.II~) into (C [a,b] 11.11 (0 ).
In particular 1jJ restricted to (C l [a,b] , 11.11 1 ) imbeds
(C l [a,b] , 11.11 1 ) into (C [a,b], 11.11 (0 ) .
Therefore there exists a constant M> 0 such that
Let fecI [a,b] be such that
f(t»O, for all te[a,bJ.
Choose t e [a,bJ with f (t) = min{f (t): t e [a,b]} > 0
and <5 = f (t) .

2M
We claim that the ball
ist contained in Kl . In fact, let g e B <5 (f) .
According to (*) ~olds:
whence
f ( t » ! (f(t) -g(t», for all te [a,b].

2M - M
89
that implies
g(t) >f(t)- f(t) >0, for all te [a,bJ

- 2
that is Bo (f) c Kl •
Remark. Closed ordering cones with non-empty interior seem to be useful
for scalarization of certain vector optimization problems (compare
with JAHN [4J).
We recall that whenever the ordering cone of a partially ordered normed
space has a non-empty interior its dual cone has a w*-compact basis.
Indeed, if e is a interior point of the ordering cone K of a normed

space X then the set
B*: ={1/J eX*: 1/J(e) =l}
is a w*-compact basis for K* (For a proof see e.g. JAMESON [5J page
123) •
Theorem 2.2
Let (C l [a,bJ, II. Ill) (resp. Kl ) be the real normed linear space (resp.
the natural ordering cone on Cl[a,bJ) defined in the section 1.
Then the dual cone Ki has a w*-compact basis B* with the evaluation
functionals as extreme points, that is,
ext(B*) = {e t e (Cl[a,b])*: t e [a,bJ)}.
where e t : Cl [a,bJ ---+- IR is defined by
et(g):=g(t).
Proof:
Since e =1 is in the interior of Kl (see the proof of Lemma 2.1),

90
the convex set
is a w*-compact basis for
Now let
where et denotes the evaluation functional
e t : Cl [a,b] - IR defined by e t (g) : = g (t) .
Each such functional is continuous. In fact, choose to E [a,b].
According to a Sobolev's imbedding theorem (see the proof of Lemma 2.1)
we have
where M is a positive constant independent of f.
Therefore
let (f) I = If (t ) I
o 0
~Mllflll,
what shows the continuity of et.

o
So we have F* c:B*. We claim that F* is w*-closed and consequently
w*-compact. In fact, let (e t ) be a net in F* converging w* to

d
1/1. That is,
et (f) -1/I(f) , for all fE (Cl[a,b], 11.11 1 ).

d
whence
91
Now the net (t d ) c::: [a,b] has a convergent subnet still denoted by (t d ).
Set s: =lim td. Since each f £ C1 [a,b] is continuous we have

d
f (s) = lim f (t d )
d
= 1/1 (f) •
Therefore 1/1 = e s € F*, what shows that F* is w*-c1osed.
w*
Now let A*: = COiiV (F*).
We claim that A* = B*.
The inclusion A* c::: B* is trivial, so we have to show the proper inc1!:;!

sian does not hold. Suppose by contradiction we could find ~ £ B*"A*.
In that case, since K~ n (-K~) = {a} B* is contained in the comple-

ment of (-A)* so we have
Hence ~ ¢ (A* U (-A*)), what implies that the following holds:
w*
(*) ~ £ COiiV (A* U (-A*)) .
In fact, otherwise there exist nets (t A) c::: [O,lJ and (n A) c::: A* (resp.
(-~A) c::: (-A*)) such that
Since ~ £ B*. In particular
l=~(e) = lim [tAnA(e) + ( l - t A) (-~A(e)J

A
= lim etA + (-1) (1 - t A)]
= lim [2t A - 1J .
A
Therefore
92
Now since (n A) c.A* and A* is w*-compact th,ere exists a convergent
subnet (nA,) of (n A).
So we have
 = w* - lim n I
A' A
Therefore € A*, in contradiction to the choice of . Consequently
(*) holds, thus there exists a w*-closed hyperplane, that strictly

w*
separates from conv (A * U (-A *», that is, there exists
f € Cl[a,b]"'-{O} and Ci. € IR such that
(f) > Ci. ~ let(f) I, for all t € [a,b].
In particular, we have
Since (e) = 1, on one hand
On the other hand, since f is bounded
so since €B*c.K*
1
a contradiction. Therefore A* = B*.
Consequently (see e.g. DUNFORD - SCHWARTZ [2J page 440 - Lemma 5)
(I) ext(B*) c.F*.
Now let
B*: = {1jJ € Kl*: 1jJ (e) < l } .

o -
93
We claim that for each f € (Cl[a,b], /1.11 1 ) holds:
(lj! (f) ) 2 :: lj! (f2), for all lj! € B*

o
(compare with TATE [7]).
In fact,
lj!( (af + e) 2) ;:: 0, for all lj! € B* and a € IR.

o
what implies
whence
Thus, since lj! € B*

o
We show next that each multiplicative functional is a extreme point of
B*.
o Let ~ be a multiplicative functional and be v and ~ linear
functionals in B* such that
o
~ = (1/2) v + (1/2) ~.
On one hand,
;:: (1/2) (v (f»2 + (1/2) (~(f» 2.
On the other hand,
(~(f»2 =[(1/2)v(f) + (1/2)]1(f)J 2
= (1/4) v (f) 2 + (1/2) v (f)]1 (f) + (1/4)]1 (f) 2.

94
Since s is multiplicative
o> (1/4) (v (f) ) 2 - (1/2) v (f) \l (f) + (1/4) (11 (f» 2
Hence v = 11 = S, what shows that s € ext (B~) •

Now since B* is an extremal subset of B* and each eva1uationfunctio
o
na1 is multiplicative, we have
(II) F*c:::ext(B*).
Therefore it follows from (I) and (II) that
F* = ext (B*) ,
what concludes the proof.

CJ
References
[lJ Boh1, E. (1974) Monotonie: Loesbarkeit und Numerik bei Operator-
gleichungen (Springer Tracts in Natural Philosophy - Vo1. 25) •
[2J Dunford, N. & Schwartz, J.T. (1957) Linear O~erators - Part I
(John Wiley & Sons) •
[3J Holmes R.B. (1975) Geometric Functional Analysis and its A~p1ic~
tions. (Springer Verlag, New York - Heidelberg) .
~J Jahn, J. (1986) Parametric Approximation Problems Arising in
Vector Optimization (to appear).
[5J Jameson, G. (1970) Ordered Linear Spaces (Lectures Notes in Mathe
matics - Vol. 141).
~J Smirnow, W.I. (1971) Lehrgang der Hoeheren Mathematik - Tei1 V
(VEB Deutscher Ver1arg der Wissenschaften - Berlin) •

95
[7J Tate, J. (1951) On the relation between extremal points of convex
sets and homomorphisms of algebras, Corom. Pure Appl. Math.
4, 31-32.
A.R. da Silva, Instituto de Matematica, Universidade Federal do Rio de
Janeiro, Caixa Postal 68530, CEP 21.944 - Rio de Janei-
ro, Brazil.
Polarities and Stability in Vector Optimization
S. Dolecki and C. Malivert

Departement de Mathematiques , Faculte des Sciences
87100 Umoges Cedex , France.
Abstract :This paper studies the closedness, upper and lower semicontinuity of the solution and
value multifunctions of a parametrized multiobjective problem . The analysis is based upon the use
of polarities in a characterization of the set of maximal points .
Introduction:
In this paper we study the behavior of the set of solutions and the set of
values for a parametrized vector maximization problem. This topic has been
investigated before by several authors as Sawaragi / Nakayama / Tanino [12]
Lucchetti [9] ,PenotlSterna-Karwat [11] . A special attention has been given
recently to the upper semicontinuity of solutions by Bednarczuk [1] .To our
knowledge, Penot and Sterna-Karwat provide the most general results on the
lower semicontinuity and graph-closed ness . Here we slightly weaken their
hypothesis on the behavior of the constraints and objective functions .For
instance in Theorem 2.2 , Conditions 2.9 and 2.10 are substituted for the
closed ness and the lower semicontinuity of G , and in Theorem 2.3 our second
assumption is weaker than the continuity of f .The relation that we consider
on the value space is only supposed to be transitive and we use a notion of
maximality which is sometimes referred to, as efficiency up to
indifference .This choice seems to be natural in the case of a non symmetric
97
relation and amounts to saying that a point is maximal if there exists no

other non equivalent point better than it .Note that this notion allows the set
of optimal values to contain different points, which are then equivalent one
to another with respect to the preference relation on the value space .
Because of its practical interest we pay a special attention to the case
where the value space is ordered by a constant convex cone . However our
framework is much more general since no linear structure is involved and we
consider the case of moving preferences. More precisely, to each parameter
we associate a transitive relation on the value space and a maximization
problem with respect to this relation . Thus , the domination structure on Y ,
the feasible set on X and the objective function will be allowed to vary
simultaneously , while in [12] moving preferences are only considered with a
fixed feasible set and a constant objective function . Doing so allows us to
encompass in the same work the behavior of the values of the optimization
problem and also of the set of its solutions. The graph-closedness in the case
of mobile transitive relations has been already investigated in [11] .However
here , all the results are provided in the generality of mobile relations .
In Section 1 we give several characterizations of the set of maximal points
using a polarity associated to the preference relation . We use these
characterizations in Section 2 in order to get the graph closed ness and the
upper semicontinuity of the value and the solution multifunctions. These
results are all derived from our theorem on general moving polarities . In the
last section we give sufficient conditions ensuring the lower semicontinuity
of the value multifunction. The case of non topological spaces is studied in
[6] where stability results are established in spaces endowed merely with a
convergence structure .
- Vector optimality and polarities.
Let be Y a set equipped with a transitive relation denoted by !;; ( read "smaller
98
than" ) . We denote by ;d the associated inverse relation and by !;;?; I ~ their
complementary relations. The symbol ::::: stands for the "equivalence" relation
(j;; and ;d ) and c for the "strict" relation (j;; and ~) relative to j;;. We will use
also the inverse relation of c I denoted by ::J I and their complementary
relations It: and :tI. For any relation 0 on Y and for each x,y e Y , we will
denote indifferently (x,Y) e 0 or ye Ox or y 0 x, in contrast with certain
other notations . A point Yo e Y is said non dominated by A c Y if for every
yeA, Yo j;; y entails y j;; Yo . If besides yoeA then it is called a maximal point
of A. This notion of maximality was introduced in [7] in the case of a

topological vector space ordered by a convex cone . With this type of
maximality , a point of A equivalent to a maximal point of A is also maximal .
Given an arbitrary relation 0 on Y ,or equivalently a subset of Y x Y , we
associate to 0 the polarity [0] by setting [3], for each Ac Y,
[O]A = n {Ox : Xe A }.
The subset [O]A of Y is called the J:l.Q..l.ar. of A . In what follows we will take
for 0 , the relation It: ("non strictly smaller than") which may be defined by
x It: Y iff x!;;?; y or y j;; x .
The main interest of this relation is that the corresponding polar [1t:]A of A
is exactly the set of points which are non dominated by A .Thus we get the
following formula for the set of maximal points of A
j;; -max A = [1t:]A n A (1.1 )
Thus we are naturally induced to use general results on the continuity of

polarities in order to inquire the stability of the value set of a
parametrized vector maximization problem .
We shall derive other formulae from this first one , using the closure
operator [3] associated to our polarity [It:]. Recall that a closure operator is
an operator which is isotone, expansive and idempotent . The closure

associated with the polarity [It:] of a subset A of Y is defined by
(1.2)
Another formulation ,more handy to check whether a point belongs or not

99
to the closure of A, is the following

y e cl¢A iff \t z c y (1.3)
Note that we have always

[It:] A = [It:] cl¢A (1.4)
Indeed the inclusion ::::> is obvious and conversely if z ~ [It: ]cl¢A there exists
yecl¢A such that zcy , implying by (1.3) the existence of x eA such that zcx.
In the sequel we shall need the following characterization of transitivity .
Proposition 1.1 The following are equivalent

(i) c is transitive ;
(ii) cA c cl¢A for each A c Y ;
(iii) !;-max A = !;-max (~A) for each A c Y where ~ means (c or = ).
.E.rQ.Qf:{i) ~ (ii) Let be ye cA and z c y, by transitivity we have z e cA, and
ye cl¢A from (1.3) . In fact cI¢ being expansive we have ~A c cl¢A .

(ii) ~ (iii) A maximal point of ~A being non dominated by ~A ,we have
!;-max{!;A) c A and using (1.1) we get !;-max (~A) = [1t:]~A n A. From (ii)
A c ~A c cl¢A. Thus !;-max (~A) = [1t:]A n A = !;-max A .

(iii) ~ (i) Suppose c non transitive. There exist x,y and z with xcy, y::J z and
Xlt: z . Then we have x e !;-max{x,z} and x ~ !;-max ~{ x,z} since xcy with
ye !;{ x,z} . @
Since the preference relation !; is always supposed to be transitive , the
corresponding strict relation c is also transitive and from the previous
proposition we get for each A c Y:
!;-max A = [It: ]~A n ~A (1.5)
!;-max A = [It: ]~A n A (1.6)
!;-max A = [1t:]A n ~A (1.7)
In the sequel, we shall also need the formulae

100
if x c y and y ~ Z, then x c Z (1.8)
if x ~ y and y c Z, then xc Z (1.9)
2- Graph closed ness of the value and solution multjfunctions.
Consider topological spaces W,X and Y, an objective function f :WxX- Y and a
multifunction D:W ==~ X . Furthermore for each w € W let be given a
transitive relation ~w defined on Y . In all the sequel if (OW)WEW is a family
of relations on Y, 0 will denote the multifunction from WxY to Y such that for
each (w,Y) € WxY ,
o (w ,Y) = { z€ Y:Z € Ow Y } .
We consider for each w€ W the problem:
lJ>(w)
where Gw= { f(w,x) € Y : x €Dw} .

The choice of a varying domination structure may be justified in an
economical setting by the fact that preferences of the decision maker may
change for example with the time . But on the other hand this situation arises
naturally in classical examples .
Consider for instance the solution set of the scalar problem
max {f(w,x) : x € X }
where f is a function from WxX to IR .The solution set of such a problem may
be viewed as the set of maximal points of X under the relation ~w defined by
V x,t € X x ~wt iff f(w IX) ~ f(w,t) , (2.1)
and we observe that the relation above , changes accordingly to w .
Now ,let us introduce two notions of convergence which are related to

graph-closed ness and lower semicontinuity of multifunctions . If F is a
multifunction from X to Y we recall that y belongs to the classical upper
101
limit of F at x :
Y EO ( LimX-I y- F)x
iff, for every neighborhood V of y and U of x , V intersects F(U) , or

equivalently if there exists a net (Xj,Yj)jEI in the graph of F,tending to (x,y).
A point Y belongs to the classical lower limit of F at x :
Y EO ( LimX+ I y- F)x
iff, for every neighborhood V of y there exists a neighborhood U of x such that

for each u EO U , V intersects F(u) . In other words it means that every net (Xj)j€I
tending to x admits a subnet (Xj)j€J such that there exists a net (Yj)j€J tending
to Y with (Xj,Y)j€J in the graph of F .The link with graph closed ness and lower
semicontinuity is the following : The inclusions
(LimX-y- F)x c Fx ( LimX+ y- F)x ::::> Fx

I I
mean respectively that F is graph-closed at x and lower semicontinuous at x .

The notations we introduce here are motivated by the theory of [' -limits
[2], and make sense even when the considered spaces are no more topological
but only endowed with convergence structures .
The following proposition generalizes to the case of non constant relations
a result of [3].
2.1 Proposition Let a : wxY ===* Y and A: W ===* Y be multifunctions.

Suppose that
( LimWxY- y- a)(w ,y)

I
c a(w,y) for each yEO Y ; (2.2)
( LimW+ y- A)w ::::> Aw (2.3)

I
Then
(LimW-y- [a]A)w c [aw]Aw .

I
Proof: Let be z~ [aw]Aw . There exists y EO Aw such that z~ a(w ,Y) .
By (2.2) and (2.3) ,we have

102
z~ ( LimWxY- y- a}(w ,y) (2.4)

I
and
ye ( Limw+ y- A)w (2.5)

I
From (2.4) there exist neighborhoods T of z , U of wand V of y such that
T n a(UxV) = +. (2.6)
By virtue of (2.5) there exists a neighborhood U' of w, such that U' c U and for
each u e U', Au n V ¢ + . Consequently , for each u e U', [a u]Au n T= +

proving that z ~ ( LimW- y- [a]A)w .@
I
Substituting for a the relation It: we get sufficient conditions for the
closed ness of the value multifunction ~ -max G at w. In fact, it follows from
our next result that if It: is a closed set in WxYxY , G is graph-closed and
lower semicontinuous at w ,then ~ -max G is graph-closed at w .
2.2 Theorem Suppose that
( LimWxy-y-lt:)(w,y) c It:wY for each yeY; (2.8)

I
( LimW+ y- ~G)w ~ Gw ; (2.9)

I
( LimW-y- G)w c ~wGw (2.10)

I
Then
( LimW- y- ~-max G )w c ~w-max Gw .

I
.E.!:.QQ.! : Assumptions (2.8) and (2.9) allow to apply Proposition 2.1 . We have
(2.11 )
Since graph-closedness is preserved under finite intersection we may use

(2.10) and (2.11) together with caracterizations (1.6) and (1.7) to get
( LimW- y- ~-max G )w c ~w-max Gw . @

I
103
Note that the first condition of Theorem 2.2 means that the graph of the
multifunction r;t: is closed in the product space WxYxY . When we consider a
fixed quasi ordering given by a convex cone C in a topological vector space
this amounts to say that C \ (C n -C) is open .
As seen at the beginning of this section ,the solution set of a scalar
maximization problem is equal to the set of maximal points of the feasible
set Dw with respect to the relation defined on X by (2.1) . Thus in the scalar
case the solution multifunction is graph closed at w as soon as (2.8) is
satisfied for the relation (2.1) and (2.9), (2.10) are satisfied with D instead
of G .These conditions are fulfilled for instance if we suppose f continuous on
{w}xDw and D continuous at w as in Theorem 8 of [8] and also under weaker
assumptions, as in Corollary 5.4 of [2] .
In the non scalar case the graph-closedness of the solution multifunction
may be derived directly from Theorem 2.2 . Let us denote the solution set of
the problem (l1> w) by :
Sw = {x eDw : f(w,x) e ~w-max Gw} (2.12)
Analogously to the scalar case we have

Sw = ~fw-max Dw (2.13)
where ~fw is the relation on X defined by
\:I x,t e X x ~fw t iff f(w ,x) ~w f(w,t) (2.14)
2,3 Theorem Suppose that
(i) ( LimWxY-y- r;t:)(w ,y) c r;t:wY for each ye Y ;

J
(ii) f(w,x) c (LimWxX+y- ~f)(wJx)

J
n
for each x eX;
(iii) (LimW- X- D)w c Dw c ( LimW+ X- D)w ;

J J
Then
( LimW- X- S)w c Sw
J
104
f.rQ.Q1 : According to (2.13), we may use Theorem 2.2 (with 0 substituted for
G), provided that we prove the graph-closed ness of the multifunction
i;Cfw' WxX ==~ X defined with the aid of (2.14). Suppose that p cfw q , that is,
f(w,p) C w f(w ,q). By (i), there exist neighborhoods T, U, V of w, f(w,p) and
f(w,q), respectively, such that,
for each t €T, U €U , v €V ,
(2.15)
In view of (ii), there exist neighborhoods To of w, P of p and 0 of q such that
for each t € To ' r € P and s € 0 :
;;;;!t f(t,r) n U :¢: + and !;t f(t,s) n V:¢: +

The above, entails that for every t in To, r in P and s in 0 there exist u €U and
v€V such that f(t,r)!;t u and V!;t f(t,s) . In view of (2.15) and the
transitivity of!; t ' (1.8), (1.9), this amounts to f(t,r) C t f(t,s) for each
t€T n To, rEP and s €O , so proving the graph-closedness of (2.14). @

The second assumption of Theorem 2.3 is slightly weaker than the continuity
of f . In fact , considering a topological vector space quasi ordered by a
convex cone C , this condition means that for every neighborhood V of f(w,x)
there exists a neighborhood TxU of (w,x) such that :
f(TxU) c (V-C) n (V+C) .

If C is a normal cone this condition is equivalent to the continuity of f .
To conclude this section we provide sufficient conditions on f and 0 ensuring
(2.9) and (2.10) in Theorem 2.2 . Since Gw = f(w, Ow) these results are
consequences of theorems on the composition of two relations stated in [4]
and we recall their proof only for completness sake.
2.4 Proposition Suppose that
( LimW+ X- O)w ::::> Ow (2.16)

I
f(w,x) € (LimWxX+ y- r;;;;f)(w IX) for each x € Ow (2.17)

I
Then (LimW+ X- r;;;;G)w ::::> Gw

I
105
E.!:Q.Q! : Let yd( w ,Ow} and let V be a neighborhood of y. There exists Xe 0 W
such that y=f(w,x} , and by (2.17) there exist neighborhoods T,U of wand x
respectively such that for each (t,u) e TxU
V n (!;;f(t,u» *- + (2.18)
Moreover it follows from (2.16) that there exists a neighborhood T 1 of w such

that
(2.19)
Thus ,for each t eTnT l , V n !;;G(t} *-+ . @
Contrary to the case of the lower semicontinuity ,the composition of two

graph-closed relations is not necessarily graph closed. Following [5] we say
that a multifunction O:W ==:; X is compactoid at We W (or subcontinuous) if ,
for every open covering {Uj}jEl of X there exists a neighborhood T of wand
{i l , ... ,i n } c I such that O(T} c Ui U ... U U·\ . It is well known that if 0 is

1 n
compactoid at w ,then there exists an Xe X such that for each neighborhood Q
of x and T of w, Q n OT *- + .In term of nets, 0 is compactoid if and only if for
each net (w j,Xj}jEI in the graph of 0 with (w j}jEI tending to w , there exists a
subnet (X)jEJ convergent in X .
2.5 Proposition Suppose that 0 is compactoid at wand that
, O}w c Ow
(LimW-X- (2.20)
( LimWxX-,y- f)(w,x} c !;;w f(w,x} for each Xe Ow ; (2.21)
Then
( LimW-,y- G}w C !;;w Gw .
E.!:QQL: Let be ye ( LimW-y-

, G}w ,for every neighborhood T of wand V of y
we have
106
GT n v :# t (2.22)
or in other words the subset of X defined by
AT = U {xEDw: f(w,x) E V} (2.23)

wET
is non empty .We shall prove that there exists XE X such that for every
neighborhoods T of wand U of x
(2.24)
Indeed suppose on contrary ,that for each x EX there exist neighborhoods Tx of
wand Ux of x such that
ATx n Ux = t (2.25)
Then (Ux)xEX containing an open covering of X , and D being compactoid there
exists a neighborhood T' of wand {xl' ... 'xn} c X such that
DT' c UXl U ... U UX n (2.26)
Choosing To c T' n TXl n ... n TXn we have by (2.23) AT c ATx

o 1
n ... n ATx n
and by (2.25) , AT
o
n ( UXl U ... U UXn) = t . Since AT c DTo, we get from
0
(2.26) , AT = t , a contradiction with (2.23) . Now (2.24) and (2.20) entail

o
that x EDw . Thus for every neighborhoods T,U,V of respectively W,X and y we
have from (2.23) and (2.24)
f(TxU) n V:# t .
Hence , from (2.21) we have y E !; wf(w ,x) with xED w or equivalently
yE !;wGw. @
3 Upper semicontjnujty of the value and solution multifunctions
Let us recall that a multifunction F :W ==~ V is said upper semicontinuous
(u.s.c) at WE W if for each neighborhood V of Fx there exists a neighborhood

107
T of w such that FT c V . It is well known that if F 1, F2 are two
multifunctions, F 1 graph closed at w, F2 u.s.c at w with F2 w compact (in our
convention "compact" does not imply "Hausdorff"), then the intersection

multifunction F1 n F2 is u.s.c. at w
( LimWxY-y- It)(w ,Y) c ItwY for each y€ Y (3.1)

J
( LimW+ y- !;;;G)w:::) Gw ; (3.2)

J
G is u.s.c. at w with Gw compact; (3.3)
Then
~-max G is u.s.c. at w .
Proof: Denoting by A : W ==~ Y the multifunction defined by At = !;;;tGt
for t;tW and Aw =Gw we have by Proposition 2.1 (with Q= [It])
(LimW-y- [It]A)w c [ltw]Aw.

J
This means that the multifunction [It]A: W ==~ Y is graph closed at w
Now , in view of (3.3) we get that [It]A n G is u.s.c. at w . But this
multifunction is exactly ~ -max G Indeed, for t;tw we have by (1.6)
[It t]!;;;tGt n Gt = ~t-max Gt, and at w I [It w] Gw n Gw = ~w-max Gw by

(1.1). @
Theorem 3.1 recovers in the scalar case Theorem 7 from Hogan [8] saying that
max f(w,Dw) is continuous as soon as D is continuous at w,uniformly compact
near wand f is continuous on {w}xDw . Indeed, in the scalar case, (3.1) is
always satified and D lower semicontinuous with f continuous on {w}xDw
entails (3.2) (Proposition 2.4) . Moreover, if f is continuous on {w}xDw and D
uniformly compact near w with a closed graph at w , (3.3) is also satisfied .
As in the previous section, the upper semicontinuity of the solution

multifunction is a particular case of the result above about the value
108
multifunction .
(i) (LimWxY-,y- It:)(w ,Y) c It: w y for each yE Y ;
(ii) f(w,x) c (LimWxX+ ,y- !; f)(w ,x) n (LimWxX+ ,y- ;df)(w,x)
for each x EX;
(iii) (LimW+ ,X- D )w ~ Dw;
(iv) D u.s.c. at w , Dw compact;
Then
the solution multifunction S is u.s.c. at w .
.E.!:.Q.Qi : As in Theorem 2.3 ,assumptions (i) and (ii) imply that the relation
defined on X using (2.14) satisfies (3.1) . Moreover (iii),(iv) are equivalent to
conditions (3.2) and (3.3) with D instead of G .Thus using the formulation
(2.13) of Sw, as the set of maximal points of Dw for the relation (2.14) we
get from Theorem 3.1 that S is u.s.c. at w . @
4 Lower semicontinuity of the marginal multifunction .
The following simple example shows that apparently "good assumptions" are
not sufficient to ensure the lower semicontinuity of the value or the solution
multifunction . Suppose that W = IR and X = Y = 1R2 , f(w,x) = x for each WEIR
and that 1R2 is quasi ordered by the half space {(r1,r2) : r1 ~ O} .we take for Dw
the segment [(w ,0),(0,1)] . The multifunction D is obviously lower
semicontinuous graph-closed and compactoid (thus upper semicontinuous) at

O,also f is continuous and open . However the solution and the value
multifunctions are not lower semicontinuous at ° . Indeed for w;cO these

109
multifunctions reduce to the singleton (w,O) while at ° they are equal to the
segment [(0,0),(0,1)]
In the sequel we shall give two kinds of results ensuring the lower
semicontinuity of the value multifunction. The first one is a variant for the
case of maximality up to indifference of Proposition 3.1 of [11] .
Given a transitive relation !:;; on Y ,we say that Y is !:;; -normal if for every
y€ Y there exists a neighborhood basis ~ (y) such that each V€ ~ (y) satisfies
V :::> (!:;; V) n ( ;d V) . (4.1)
If !:;; is reflexive the above inclusion becomes an equality .
4.1 Theorem Suppose that Y is !:;;w-normal and
(LimWxY-Y-
, ;d)(w,y) C ;dwY for each y€Y; (4.2)
( LimW-,Y- G)w C Gw c ( LimW,Y- G)w; (4.3)
There exists a neighborhood To of w such that
\7't€To \ly€Gt (;dtY) n !:;;t-max Gt::l= ~ ; (4.4)
The multifunction !:;;-max G is compactoid at w ; (4.5)
Then
( LimW+,Y- !:;;-max G )w :::> !:;;w-max Gw
ErQ.QL: If we suppose that the result is false, then there exist

Zo € !:;; W -max Gw and an open neighborhood V0 of Zo such that for each
neighborhood T of w the subset
(4.6)
is non empty. Let V 1 be another neighborhood of zo' included in Vo ' satisfying
(4.1) . By the lower semicontinuity of G (4.3), and (4.4) ,for each neighborhood
V of Zo ' there exists a neighborhood To of w such that
(4.7)
Thus, for each neighborhoods T of wand V of zo' the subset

110
is non empty, because FTn T C FT n To' Since ~ -max Gt is supposed to be

o
compactoid at w I (4.5) the filter ff = (ff TV h,v admits a cluster point zl such
that for each neighborhoods T of w, V of Zo and Q of zl
3 teFT (4.8)
Now,remark that (4.8) entails zl e (LimW- y- G)w n ( LimWxY- y- ;d )(w,zo)

I I
Therefore, using (4.2) and (4.3), zl ;dwZo and zleGw which entails, by the
maximality of Zo ' zl ~w Zo . Thus, by (4.1), zl e V 1C V 0 and V0 is also an open
neighborhood of zl' Writing (4.8) with Vo instead of Q we get, for each
neighborhood T of w,
3 teFT ~t-max Gt n Vo ¢~
contradicting (4.6). @
Remark: Instead of Y ~w-normal it would be sufficient to suppose that there
exists a neighborhood basis ~ (y) such that
\;j Ve ~(y) \;j Ze V (~z) n (;dZ) C V.
Our next result specializes Y to be a topological vector space on which a
family {Athe:w of convex cones (with vertex at the origin) is defined. We shall
consider the family of transitive relations {~the:w defined by Yo ~tYl
whenever y 1 -Yo eAt· The corresponding strict relations C t, t e W , may
be characterized by : Yo Ct Yl' whenever Yl -yoeAot, where
Aat = At \ {-At}·
We say that a multifunction G :W ==4 Y is attracted by the family {~the:w at
(w,z) if, for every neighborhood V of Z there exist neighborhoods U of Z and To
of w such that, for every t e To, Gt nU ¢ ~ and [It: t]Gt nU ¢ ~ imply
~rmax Gt n V ¢ ~ .
111
4.2 Iheorem Suppose that
( LimW+ y- Ao C)w:::> Acw C (4.9)

J
(Limw-x- G)w c Gw c (LimW+ J x- G)w; (4.10)

J
~-max G is compactoid at w ; (4.11 )
Ihere is a neighborhood I of w such that int(n Aot : t EI) *~ (4.12)
G is attracted by (Ath€w at (w,z) for each Z E ~w-max Gw (4.13)
Ihen
( Limw+ y- ~-max G }w :::> ~w -max Gw

J
Remark: In the case where At=A for each t EW (4.9) is automatically satisfied
and (4.12) is satisfied if int A *~ .
f.[QQf : Suppose that ZOE (~w-max Gw) \ ( LimW+ y- ~-max G)w . Ihen there
J
exists a neighborhood Vo of 0 such that w belongs to the closure of
H = {tEW: ~t-max Gt n (zo+Vo) = ~}.
By (4.13), there exist neighborhoods U of 0 and 10 of w such that
From the lower semicontinuity of G (4.10) ,there exists a neighborhood 11 of
w such that
Ihus, for each t EIon 11 n H, we have [¢ t]Gt n (U+zo) = ~ or equivalently
(4.14)
Using (4.12) there exists a neighborhood 12 of w, with I 2 cI o nI 1 ,such that
Ao = int( n Aot : t E12 ) is non empty. Let roE Ao and to>O be such that to roE U .
In view of (4.14), for each t EI2 nH, there exists YtEGt such that
Zo +~roE Yt - Aot
Since Act is a convex cone we have, for each t E12 n H , Aot + Ac c Aot and
112
(4.15)
where Uo = to ro - Ao is a neighborhood of 0 . As G is compactoid there exists Yo
such that, for every neighborhood V of Yo and each neighborhood T of w there
is t e: T with Yt e: V . By (4.10) , Yo e: G wand, by the maximality of zo'
Yo-zo e: Aow c . Let S be a neighborhood of 0 such that S+Sc -U o . It follows
from (4.9) that there is a neighborhood T3 c T 2 of w such that
Vte:T 3 (Yo-zo+S)nAotc*<} (4.16)
Choosing t e:T3 with Yt e: Yo- S we get Yo - Zo + S c Yt - Zo - Uo and by (4.16),
(Yt - Zo - Uo) n Ao{ *<}, contradicting (4.15) . @
References :
[1 ]Bednarczuk E., Well posedness of vector optimization problems,

International Conference on Vector Optimization , Darmstadt , (1986) .
[2]Del Prete I.,Dolecki S.,Lignola M.B., Continuous convergence and

preservation of convergence of sets ,J .. Math. Anal. Appl. 120 (1986),
454-471.
[3] Dolecki S., Convergence of minima in convergence spaces ,Optimization 17

(1986) , 553-572 .
[4] Dolecki S., Continuity of bilinear and non bilinear polarities, Int. School of
Math. G.Stampacchia ,Erice 1984 (R.Conti, E. De Giorgi, F. Giannessi, eds.)
Springer Verlag ,1986.
[5] Dolecki S.,Greco G.H.,Lechicki,A., Compactoid and compact filters , Pacific

J. Math., 117,No 1 ,(1985).
[6] Dolecki S.,Malivert C. , Stability of efficient sets: continuity of mobile

113
polarities, to appear.
[7] Hartley R., On cone-efficiency , cone-convexity and cone-compactness

SIAM J. Appl. Math. 34, (1978) , 211-222 .
[8] Hogan W., Point-to-set maps in mathematical programming ,SIAM Review

15,(1973), 591-603.
[9] Lucchetti R., Stability in Pareto problems , Quaderno 14 ,Universita di

Milano ,Dipartimento di Matematica ,(1985).
[10] Naccache P.H., Stability in multicriteria Optimization ,J. Math. Anal.

Appl. ,68, (1979), 441-453 .
[11] Penot J.P., Sterna-Karwat A., Parametized multicriteria optimization;

Continuity and closedness of optimal multifunctions , J. Math. Anal. Appl. ,
120 ,(1986), 150-168 .
[12] Sawaragi Y.,Nakayama H.,Tanino T., Theory of Multiobjective Optimization

,Academic Press Inc ,1985.
[13] Tanino T., Sawaragi Y., Stability of non dominated solutions in

multicriteria decision-making , J. Opt. Th. Appl. , 30,(1980), 229-253.
SETS Of EffICIENCY IN A NORMED SPACE
AND INNER PRODUCT
Roland DURIER
Laboratoire d'Analyse Numerique
Universite de Bourgogne
B.P. 138 - 21004 - DIJON CEDEX
Abstract : In a normed space X the distances to the points of a given set A being
considered as the objective functions of a multicriteria optimization problem, we de-
fine four sets of efficiency (effiCient, strictly efficient, weakly efficient and pro~
perly efficient points). Instead of studying properties of the sets of efficiency ac-
cording to properties of the norm, we investigate an inverse problem : deduce proper-
ties of the norm of X from properties of the sets of efficiency, valid for every finite
subset A of X.
We thus obtain some new characterizations of inner product spaces. The tools are
classical James' characterizations of inner product spaces and a description of each
set of efficiency which looks like the description of the convex hull of A.
The same machinery allows to study polyhedral spaces where only partial results
are obtained. Some characterizations of strictly convex spaces are also given.
INTRODUCTION
In the context of multiobjective programming; a natural problem is the

study, according to properties of the objective functions, of properties
of sets of efficiency (e.g. existence, topological properties, stability).
We investig~te here, in a particular framework, an inverse problem: de-
duce properties of the objective functions from properties of the sets
of efficiency.
Our framework will be the following : a non-empty finite set A of a
normed linear space X over ffi being given, we consider the multiobjective
programming problem where the objective functions are the distances, re-
lative to the norm of X, to each point of A. With such a multiobjective
problem four sets of efficient points can be associated -they are formal-
ly defined in Section 1. - : the sets of strictly efficient, efficient,
weakly efficient and properly efficient points, respectively denoted by
c(A), E(A), C(A), M(A). They satisfy the inclusion relations (where
y ~ Z means Y c Z) :
A~C(A)~
E(A)~ C(A).
~ M(A).......==r
115
When X is an inner product space, i.e. when the norm of X is derived

from an inner product, then c(A), E(A) and C(A) are all equal to the con-
vex hull of A and M(A) is the union of A and the relative interior of the
convex hull of A. When X is strictly convex, i.e. when the unit ball of
X is strictly convex, then c(A), E(A) and C(A) are equal. When X is ~
hedral, i.e. when X is finite dimensional and its unit ball is a polytope,
then E(A) and M(A) are equal.
The aim of the paper is to investigate sufficient conditions of the
above type dealing with the sets of efficiency, valid for every finite
subset A of X, such that X is an inner product space, or X is strictly
convex, or X is polyhedral.
Thus we shall obtain some criteria characterizing inner product spaces
whose dimension exceeds two. The major tools will be
a description of each set of efficiency, which looks like a descrip-
tion of the convex hull and which uses ad-hoc subsets instead of half-
spaces,
• classical results of James [10] on inner products in normed linear
spaces.
Criteria characterizing strictly convex spaces will be directly dedu-
ced from definitions. Only partial results will be obtained about poly-
hedral spaces.
The paper is divided as follows : definitions, notations and a list

of known results are given as preliminaries in Section 1 .. Section 2. is
devoted to a geometric description of each set of efficiency associated
with a given set A. Section 3. provides the central theorems of the pa-
per : they give new characterizations of inner product spaces. In Section
4. we obtain some characterizations of strictly convex spaces and in
Section 5. we investigate polyhedral spaces. Concluding remarks are in
Section 6 ..
- 1 - PRELIMINARIES.
The norm in the normed space X is denoted by I. I.

Definition 1.1. Let A be a (not necessarily finite) subset of X :
c(A) is the set of x in X for which no y i x exists in X such that,
for every a in A, Iy - al < Ix - al ;
E(A) is the set of x in X for which no y exists in X such that, for
every a in A, Iy - al < Ix a I, with at least a strict inequality
C(A) is the set of x in X for which no y exists in X such that, for
every a in A, Iy - al < Ix a I.
116
The members of c(A) (resp. E(A), C(A)), are called strictly efficient
(resp. efficient, weakly efficient) points.
Let A be a finite subset of X
M(A) is the set of x in X for which there exists a family of positive
numbers (wa)a E A such that, for every y in X,
E9I Jx-aJ.s. Ew Jy-aJ

aEA a aEA a
The members of M(A) are called properly efficient points.
The sets c(A), E(A), C(A) and M(A) are empty if A is empty.
In vector optimisation some other terminology is used: Pareto optima,

minimal solutions, admissible solutions. In geometry of normed spaces,
strictly efficient points are also called minimal points (see [2], where
the notations c(A) is used) and weakly efficient points are also called
closest points to A (see [14] where the notation C(A) is used). Properly
efficient points are initially defined by Geoffrion [8] ; the present
equivalent formulation emphasizes the fact that they are solutions of a
family of weighting optimization problems.
We list now some known facts about the sets of efficiency in a normed
space. The first proposition deals with the direct problem, while the
others deal with the inverse problem.
We denote by coCA), coCA) and ri coCA) the convex hull, the closed
convex hull and the relative interior of the convex hull of A.
Proposi tion 1.2.

(1) If X is a Hilbert space, then for every subset A,
c(A) = E(A) = C(A) = coCA).
(2) If X is an inner product space, then for every finite subset A,
c(A) = E(A) = C(A) = co(A),
and M(A) = A u ri co(A).
(3) If X is strictly convex, then for every subset A,
c(A) E(A) = C(A),
and for every two-points subset A,
c(A) = E(A) = C(A) = coCA).
(4) If X is two-dimensional, then for every compact subset A,
c(A) c coCA) c C(A),
and if moreover X is strictly convex,
c(A) = E(A) = C(A) = coCA).
(5) If X is polyhedral, then for every finite subset A,
E(A) = M(A).
117
Proof: Some of thes~ assertions are classic (see [7], [14], [12], [2]).
Proofs of relations about c(A), E(A) can be found in [6] (Proposition
1.3. and Corollary 4.1.). Proofs of relations about M(A) can be found
in [3] (Remark 3.6. and Corollary 4.3.). Another proof of (5) will be
given in Section 5. 0
Remark 1.3. Proposition 1.2. suggests inverse results which will be ob-
tained further. The necessary condition of (2) will be shown to imply
that X is an inner product space, provided the dimension of X exceeds
two, an unavoidable restriction from (4). The necessary condition of (3)
focusing on two-points sets will be shown to imply that X is strictly
convex. On the other hand we shall show that the necessary condition of
(5) does not imply that X is polyhedral.
The next Proposition 1.4. summarizes known characterizations of inner
product spaces, using assumptions on every subset A of X, or at least in-
finite subsets.
Proposition 1.4. Suppose the dimension of X is at least three.

(1) If for every subset A we have C(A)C co(A), then X is a Hilbert
space.
(2) If X is finite dimensional and if for every subset A we have
E(A) = co(A), then X is an inner product space.
(3) If for every closed hyperplane H we have c(H)c H, then X is an
inner product space.
Proof. For (1), see [14], Theorem 5. ; the assumption may be replaced by
C(H)c H for every closed hyperplane H of X.
For (2), see [15], Theorem 1.19. and for (3), see [1] (17.10). 0
Proposition 1.5. If for every two-points set A, we have

(i) co(A) cc(A) or
(ii) C(A) cc(A),
then X is strictly convex.
Proof. For (i) see [2], Proposition 3. and for (ii) see [13], Proposition
4.1. 0
Before proceeding further we give below basic notations for conveni-

ence of reference. Throughout the paper, B is the unit ball and S the
unit sphere in X ; DC, int D and IT are the complementary set, the inte-
rior and the closure of a set D. For x and y in X, [x,y] is
{ax + (1 - a) y / 0 <a < 1}, while [x, y[ and ]x, y[ are respectively
118
{ax + (1 - a) y / 0 <a~ 1} and{ax + (1 - a) y /0 <a< 1} ; these three

sets are called segments. Inner product and duality product between X and
its dual are both denoted by (.,.). For x in X and 0 in S, y' (x ; 0) is
the directional derivative of the norm at x with respect to 0 and oy(x)
is the subdifferential of the norm at x.
Generally we use terminology of Rockafellar's book [17].
- 2 - A DESCRIPTION OF THE SETS OF EFFICIENCY.
The next definition introduces cones of X, associated with a direc-

tion represented by a unit vector 0 they are essential tools in our
description of the sets of efficiency.
Definition 2.1. Let 0 be in S :
is the set of x in X satisfying Ix - Aol > Ixl for every A > 0,

is the complementary set of
is the set of x in X satisfying y'(x; - 0)
°0 ' > O.
Remark 2.2. For every 0 in S, 00 (resp. Uo ) is a cone with vertex 0 ; a

point u in S belongs to 00 (resp. Uo ) if and only if the half-line emana-
ting from u and directed by - 0 goes out of the ball B (and this half-line
is not tangent to S).
If X is an inner product space, 00 is the closed half-space
{x / (x,o) ~ O} and Uo is {O} U {x / (x,o) < O}.
The closure of Po takes place in many times later an explicit des-

cription is now given.
Proposition 2.3. Let 0 be in S. Then
18 = {xEX / 'rjA > 0, Ix + HI ~ Ixl}

{xEX / ,y'(x;o) ~ O}
Proof: This result can be deduced from Proposition 1.5. of [6] ; an in-
dependent proof is obtained by considering only convex function on ffi
t -+ Ix + tol, x and 0 being given. The same idea will be applied in
Proposition 2.5.
We verify successively
if x is in 18 ,then, for every A > 0, we have I x + AO I > I x I,
if x satisfies Ix + Aol ~ Ixl for every A > 0, then
xn = x + (1/n) 0 defines a sequence (xn)n>o such that, for every n > 0,
IX n - (1/n) 01 = Ixl < Ix + (1/n) Q I = Ixnl
i.e. xn EP o·
Thus Po is dense in the obviously closed set
{x EX / 'rjA > 0, I x + AO I ~ I x I }. o
119
Theorem 2.4.
(1) Let A be a compact subset of X. Then x E X is in C(A) if and only
if, for every oES,
A n (x + P -0) ~ 0
(2) Let A be a finite subset of X. Then x E X is
(i) in c(A) if and only if, for every oES,
An(x + P- o ) ~ 0,
(ii) in E(A) if and only if, for every oES,
An(x+Qo)~0 or Acx + fl_ o .
(3) Let A be a finite subset of X and let X be finite dimensional.
Then x E X is in M(A) if and only if, for every o E S,
A n (x + Uo ) ~ 0 or A c x + P -0 .
Proof. (1) See Theorem 2.1 . of [6 ].

( 2) See Remark 2.1 . of [6] .
(3) Denote, for the time being, by N(A) the set
N(A) {xEX / 'o'oES, An(x + Uo ) ~ 0 or Acx + P- o }
We prove that 01N(Alis equivalent to °1M(A).
We h a v e : 0 1 N( A) ¢'~ 3 0 E S, ('0' a E A, a 1 U0 and 3 a E A, a1 P_ 0 ) •
Using the equivalences (Proposition 3.1. of [6]) :
a1Uo¢'~'o'p E oy(a),
a 1 P _ 0 ¢'~ '0' p E 0 y ( a) ,
(p,o)
(p, 0)
>
°
> 0,
we get : o 1N(A) if and only if there exists, in the dual space of X, a
closed half-space containing every oy(a) ( a E A) , one at least of these
sets being contained in the open half-space.
From Theorem 3.5. of [3], this last assertion is equivalent to
01M(A). 0
When X is an inner product space, the former description of C(A), c(A)
and E(A) (resp. M(A)) can be viewed as a description of coCA) (resp.
Au ri co(A)). The geometrical character of conditions of Theorem 2.4. is
attractive and turns out to be useful.
Some partial results about E(A), essentially in ffi2 with the 11- norm
or with a polyhedral norm are given in [9] and in [18]. These papers deal
with the so-called location theory; the problem consists, for example,
in locating a facility shared by many users in such a manner that the fa-
cility is as close as possible to each of the users.
It is also possible to yield an analogous description of the interior

of c(A), even if it is empty; it implies a useful inclusion relation bet-
ween int c(A) and int M(A).
120
Proposition 2.5. Let A be a finite subset of X and let X be finite dimen-

sional. Then x E X is in int c(A) if and only if, for every 0 E 5,
An(x + int Qo) i 0.
Proof. Denote, for the time being, by d(A) the set
d(A) = {xEX / 'tIoE5, An("x +int Qo)i 0}
It is clear, from Theorem 2.4., that d(A) is included in c(A).
We show first that (d(A»c is closed. Let (x n ) be a sequence conver-
ging to x in X, with xn jt d (A) for every n. Thus we have, for every n,
3 on E 5, A e xn + Po'
n
As X is finite dimensional,S is compact; whence we extract a sub-
sequence of (On)' denoted by (on) as well, converging to 0 in 5.
We have
'tI aEA, a - xnEP o '
n
which means (Proposition 2.3.)
'tIaEA,'tIA>O, la-xn+AcSnl~lxnl.
Passing to the limit, we get
'tI a E A, 'tIA > 0, I a - x + Aol > I x I ,
which means
A ex + Po ,Le. xjtd(A).
Hence (d(A»c is closed.
We show now that (c(A»c is dense in (d(A»c.
Let xjtd(A). We have (Proposition 2.3.)
30E5, 'tIaEA, 'tIA>O, la-x+AcSI~la-xl.
Let (xn)n>o be the sequence xn = x - (1/n)O.
For a E A and for every n,
la - xn ~ la - xl = la - xn -(1/n)ol·
That means :
'tI n, 'tI a EA, a - xn E Po'
which is nothing else than
'tI n, A n (x n + Qo) = 0
i. e. xnjtc(A).
As (x n ) converges to x, (c(A»c is dense in (d(A»c. o
Corollary 2.6. Let A be a finite subset of X and let X be finite dimen-
sional ; then
int c(A) e int M(A).
Proof: We remark first that int Qa is included in Uo ' for 0 E 5.
Indeed the inclusion
(Uo)c ePo
121
is exactly the implication

y'(x;-o) ~ 0 => y'(x;o) ~ 0,
which is obvious by considering the convex function on m
t -+ Ix + tol (x and 0 fixed).
Thus int c(A) is an open set contained in M(A) from Theorem 2.4. and
then
int c(A) C int M(A). o
- 3 - CHARACTERIZATIONS OF INNER PRODUCT SPACES.
We recall first important results of James [10], using Birkhoff's

orthogonality in a normed space.
Let x and 0 be in X x is said to be orthogonal to 0 - a relation
denoted by x 1. 0 - if I x + to I ~ I x I for every t Em.
A subset D of X is said to be orthogonal to 0 - a relation denoted
by D..Lo - if x..L 0 for every x in D.
We summarize three theorems of [9] (Theorems 2., 5. and 4.)
In a normed space X of three or more dimensions, if

(A) for every x and y in X, and 0 in 5, x..L 0 and y..L 0 imply x + y..L 0,
or (B) for every 0 in 5, there exists a closed hyperplane H satisfying
H ..L 0,
or (C) for every closed hyperplane H in X, there exists 0 in 5 satisfying
H ..L 0,
then X is an inner product space.
Phelps [14] proves moreover that, under condition (C), X is a Hilbert
space.
Using Proposition 2.3., it is easy to translate orthogonality rela-
tion and conditions (A), (B) and (e)
x..Lo means: xE"PonP_ o .
(A) means 'd 0 E 5 , Po n P _0 is a convex cone, and in fact a subspace.
(B) means 'd 0 E 5, Po n P _0 contains a closed hyperplane.
(C) means for every closed hyperplane H, there exists C in 5 such that
HcPonP_ o •
Some easy characterizations of inner product spaces can now be
produced
Proposition 3.1. Suppose dimension of X greater than two.

I f (i) 'do E 5, Po is convex,
or ( ii) 'd 0 E 5, int Qc is convex,
or (i i i) 'd 0 E 5, 0 % co(int Qo) ,
122
Proof: (i) implies obviously (A).

For (ii), as 0 does not belong to int Qo ' if int Qo is convex, (iii)
just works.
(iii) implies, for every 0 in 5 existence of a closed hyperplane H
such that one of the open half-spaces associated with H contains int Qo '
Denote by H+ and H the two open half-spaces associated with H.
Suppose int Qo c H+ ; then int Q- o c H-. Then we have
H = (H+)c n (H-)c c (int Qo)c n (int Q_o)c = Po n P- o '

what is condition (8). o
The next two theorems are central in the paper.
Theorem 3.2. Suppose dimension of X greater than two.

If (i) for every finite subset A,
ri co(A) c C(A),
or (ii) dimension of X is finite, equals n and for every subset A with
n + 1 affinely independent points,
int co(A) c C(A),
Proof: The negation of the proposition

for every finite subset A, ri co(A) c C(A)
is
there exist a finite subset A and x in X such that
x Eri co(A) and x %C(A).
With x = 0, a non restrictive assumption with a translation of A, and
using Theorem 2.4., we obtain equivalently:
there exist a finite subset A and 0 in 5 such that
o Eri co(A) and A cint(Q_o)'
Thus assumption (i) is equivalent to
for every finite subset A and for every 0 in 5,
Acint(Q_o) ~ O%ri co(A).
That means : for every 0 in 5, 0 % co (int Q-o) ;
we conclude with Proposition 3.1.
If X in n-dimensional, Caratheodory Theorem leads to the conclusion
with assumption (ii) and the same reasoning. o
Corollary 3.3. Suppose dimension of X greater than two.
(1) X is an inner product space if and only if, for every finite subset
A, one at least of the sets C(A), E(A) and c(A) is convex.
(2) X is an inner product space if and only if, for every finite subset
A, one at least of the conditions is satisfied :
123
C(A) = coCA),
or E(A) = coCA),
or c(A) = coCA),
or, i f X is finite dimensional,
M(A) = A uri coCA).
Proof: All these conditions imply coCA) c C(A), except the last one which
implies ri coCA) c C(A). o
In Theorem 3.2. we assume that a set included in coCA) is contained

in C(A), the greatest of the sets of efficiency. It is natural to inves-
tigate the case where c(A), the most little of the sets of efficiency
-or better a set included in c(A)- is contained in coCA) -or better in a
set including coCA). It is the aim of the theorem below, where aff(A)
denotes the affine hull of A. Technical reasons force us to assume dimen-
sion of A is finite.
Theorem 3.4. Suppose dimension of X is finite and greater than two.

If, for every finite subset A,
int c(A) c aff(A),
Proof: Our aim is to prove condition (C) of James. Let H be a hyperplane
of X and let u%H.
st
- 1 step: For every finite subset A of u + H, 0 is not a member of aff(A)
and then 0 does not belong to int c(A). From Proposition 2.5., that is
equivalent to : A cP& for some & in 5 ; & depends on u and A and is deno-
ted by &(A,u).
Thus we have for every finite subset A included in u + H, there
exists &(A,u) in 5 such that ACP&(A,u)'
- 2 nd step : Let (An)n>o be an increasing sequence of finite subsets of
u + H such that B = h~o An is dense in u + H. Then there exists, for each

n, &n &(An,u) in 5 such that AncP& .
n
Due to compactness of 5, a subsequence denoted by &n as well , con-
verges to &u ES. Let x EB, i.e. x EAn for some n ; x E P & implies from
n
Proposition 2.3. x EP& . Whence u + H is included in p& .
u u
3 rd step : We apply result of 2 nd step to EU, where E is intended to

converge to O. We have EU + H C P & for some &EUE S. Due to compactness
EU
of 5, there exists a sequence (En) converging to 0 and & E 5 such that
&EnU converges to &. Passing to the limit in EnU + HCP& ,and using
EnU
124
Proposition 2.3., we get
The proof of Theorem 3.4. is now complete. o

Remark.
If X is not finite dimensional, our proof is not valid since we use
essentially the compactness of S ; on the other hand existence of the
countably infinite set B dense in u + H can be avoided by using nets in-
stead of sequences. But a proof valid if X is infinite dimensional cannot
lead to condition(C) of James, which implies that X is a Hilbert space.
- 4 - CHARACTERIZATION OF STRICTLY CONVEX SPACES.
Theorem 4.1. If
(i) \j o E S, P-0 c Q 0 or P -0 c int Q o '
or (i i) for every two-points set A, coCA) cc(A),
or(iii) for every two-points set A, E(A) c co(A),
or (i v) for every two-points set A, E(A) c c(A),
or (v) for every three-points set A, C(A) c E(A),
then X is strictly convex.
Proof :
1. If a segment [u, u-ao] (with a > 0 and 0 E S) is included in S,
then [u, u-ao[ is included in P- o ' but does not meet Q o .
If a segment [u, u+80] (with B > 0 and 0 E S) is included in S,
then [u, u+80[ is included in P- o but does not meet int Qo .
Sufficiency of (i) is proved by absurd.
2. Sufficiency of (ii) is proved in [2], Proposition 3. This condi-

tion is included in our theorem for the sake of completeness.
3. Let u and v be non colinear ; we claim that, under (iii),

lu + vi < lui + Ivl· We may suppose lul.s. lu + vi.
Denote a = 0, b = u + v and A = {a, b}. From assumption (iii), we
have u 1/ E (A) and therefore there exists y -I u with I y I .s. I u I and
Iy - bl .s. lu - bl, with at least a strict inequality. Introducing
"yE[a, b] with I"YI = IYI, we get
lu + vi < lui + Ivl·
4. We have, whatever X is if A {a, b}, then coCA) c E(A). Then
(iv) implies (ii).
5. Suppose X is not strictly convex. Then some segment

[u, u ao] (a > 0 and o E S) is included in S. Let a = u, b u - ao
and A {O, a, b} . Consider x ( a + b) / 2 : it is not zero, it is in
C (A) but y = x / 2 satisfies
125
I y - a I = I x - a I = I y - b I = I x - b I = 1 and I y I < I x I •
Thus x is not in E(A) and consequently we have a three-points subset A
such that C(A).¢E(A). 0
- 5 - POINTS OF EFFICIENCY IN A POLYHEDRAL SPACE.
In this section X is a finite dimensional space.
Theorem 5.1. X is polyhedral if and only.if, for every 6 in 5, U6 = Q6.

Proof : We have always U6 c Q6.
Suppose first U6 i Q6 for some 6 in S. Then there exists x E Q6 and
x%U 6 i.e.
'd A > 0, I x - HI < I x I and y' (x; - 6) < O.
That implies the convex function t ~ Ix - t61 on [0 + ~[ is strictly in-
creasing with a right derivative null at a ; thus it cannot be piecewise
linear, from what follows that X is not polyhedral.
Conversely suppose that, for every 6 in 5, Q- 6 = U_ 6 or equivalently
P- o = (U_o)c i.e.
'd x EX, 'd 0 E 5 , Y , (x ; 6) < a ¢:::>:3 A > 0, Ix+HI~lxl·
Klee [11] denotes, for xES, by cone(x, B) the smallest cone which
contains B and has vertex x. It is clear that cone(x, B) is the cone
which has vertex x and is generated by the vectors 0 in 5 such that there
exists A > 0 with Ix + Aol ~ Ixl.
With our assumption, cone(x, B) is generated by the vectors 0 in 5
such that y'(x; 6) ~ O. As the function 0 ~ y'(x; 0) is continuous, each
set cone(x, B) (x E 5) is closed. Proposition 5.B. of [11] implies that
B is a polytope. 0
Corollary 5.2. If X is polyhedral then, for every finite subset A of X,

E(A) = M(A).
Proof: A consequence of Theorems 2.4. and 5.1. 0
Corollary 5.2. can also be deduced from [3] and [4] ; in these papers,
using completely different methods, more general situations are studied,
in which each efficient point is properly efficient. 0
Remark 5.3. It is now natural to investigate if a sufficient condition

such that X is polyhedral is the fact that M(A) and E(A) are equal for
every finite subset A. A counterexample shows the answer is negative.
In m2 consider the norm associated with a unit ball B so defined
let eO = (1, 0) and, for n ~ 1, en is the point of the unit euclidean
circle whose polar angle is TI / 2 n ; B is the convex hull of the set of
points {en' - en} n>O·
126
With this norm m2 is not a polyhedral space we show that however

we have E(A) = M(A) for every finite set A.
To be able to apply Theorem 2.4., we look into the sets Qo' Uo and
P- o ; Remark 2.2. is at that time a good tool.
We note first that for 0 i (0, 1) and 0 i (0, -1), we have Q0 = U o.
On the other hand,
if8 (0, 1), we have
Q8 = P-8 = {(u, v) / v ~ OJ,
U8 {(u, v) / v < O} u {(u, v) / u > 0 and v = O}
Q- 8 , U_ 8 and P8 are obtained by symmetry.
More precisely if 0 = (en - e n + 1 ) / le n + 1 - enl (n..c 1), Q o is the
convex cone with vertex 0 generated by the segment [e n + 1 , - en] and P- 8
is the complementary set of the cone with vertex 0 generated by the open
segment ]e n , - e n + 1 [.
Suppose now that a non empty finite set A exists with M(A) i E(A) ;
then there exists x satisfy-ing x E E(A) and xjtM(A). We can suppose x = 0,
with a translation of A. According to Theorem 2.4. and the remarks above,
we have :
(*) An U8 0, A;iP_ 8 and An Q8 i 0,
or the same relations with -8.
It is sufficient to investigate consequences of (*). They are
A1 = An {(u, v) / u < 0 and v = O} i 0,

A2 = An {(u, v) / v > O} i 0,
and A = A1 u A2 .
As A2 is finite there exists n such that A2 is "above the line

(0, en)"' more precisely is in the open half-space limited by the line
(0, en) and containing (0, 1).
With 0 = (en - e n + 1 ) / len - e n + 1 1, we have A2 nP_8 = 0, and also

A2 n Q o = 0. Of course Al n Q o = i'J. That is in contradiction with the as-
sumption 0 E E(A).
- 6 - CONCLUDING REMARKS.
This paper considers four sets of efficient points in a normed space

X associated with the distances to the points of a finite subset A. A
description of these sets and of the interior of one of them is given
using cones of X which are actually half-spaces when X is an inner pro-
duct space.
From direct results about the sets of efficiency when X is an inner
product space or strictly convex or polyhedral, it is natural to inves-
127
tigate inverse problems.

First we obtain new characterizations of inner product spaces of dim-
mens ion greater than two. In a forthcoming book, D. Amir [1] goes through
more than three hundred such characterizations. Those yielded here are
answers to natural questions and deserv~s perhaps to be known. Moreover
from the point of view of multiobjective programming, precisely in gene-
ral location theory, they close the discussion about, roughly speaking,
convexity of the sets of efficiency. Of course there are still open ques-
tions : for example, in a normed space which is not an inner product
space, what are the sets A such that c(A) (or E(A) or C(A» is convex?
[5] and [16] deal with such a problem. On the other hand, considering
infinite subsets A, is c(A) (or E(A) or C(A» convex when A is convex
compact? This question is asked in [2].
A second inverse problem leads to a characterization of strictly
convex spaces. The third is more particular but seems promising : it
deals with polyhedral spaces. We get a characterization using the so-
called Qo and Uo but only a sufficient condition with equality of E(A)
and M(A) ; one may ask what is a necessary and sufficient condition on
the geometry of X such that we have E(A) = M(A) for every finite subset
A.
Moreover some results are obtained with the restriction of finite
dimensionality of X ; is it possible to avoid it ?
Finally the remarks above and most theorems and propositions of this
article show that the paper deal with multicriteria optimization and
with geometry of normed spaces.
REFERENCES
[1] D. Amir, Linear-isometric characterizations of inner-product spaces

(to appear).
[2] B. Beauzamy and B. Maurey, Points minimaux et ensembles optimaux dans
les espaces de Banach, J. Funct. Anal. 24 (1977) 107-139.
[3] R. Durier, On efficient points and Fermat-Weber problem,(submitted).
[4] R. Durier, Weighting factor results in vector optimization, J. Optim.
Theory Appl., forthcoming.
[5] R. Durier, Meilleure approximation en norme vectorielle et theorie
de la localisation, RAIRO Model. Math. Anal. Numer., forthcoming.
[6] R. Durier and C. Michelot, Sets of efficient points in a normed space,
J. Math. Anal. Appl. 117 (1986) 506-528.
[7] L. Fejer, Uber die Lage der Nullstellen von Polynomen, die aus Mini-
mumforderung gewisser Art entspringen, Math. Ann. 85 (1922) 41-88.
[8] A.M. Geoffrion, Proper efficiency and the theory of vector minimiza-
tion, J. Math. Anal. Appl. 22 (1968) 618-630.
128
[9] P. Hansen, J. Perreur and J.F. Thisse, Location theory, dominance

and convexity: some further results, Oper. Res. 28 (1980) 1241-1250.
[10] R.C. James, Inner products in normed linear spaces, Bull. Amer.
Math. Soc. 53 (1947) 559-566.
[11] V. Klee, Some characterizations of convex polyhedra, Acta Math. 102
(1959) 79-107.
[12] H.W. Kuhn, A pair of dual non linear programs in : J. Abadie, ed,
Methods of non linear programming, Amsterdam (North-Holland, 1967) pp.
37-54.
[13] P.L. Papini, Minimal and closest points,nonexpansive and quasi-non-
expansive retractions in real Banach spaces in : P.M. Gruber and J.M.
Wills, eds, Convexity and its applications (Birkhauser, Basel, 1983) pp.
248-263.
[14] R.R. Phelps, Convex sets and nearest points. II, Proc. Amer. Math.
Soc. 9 (1958) 867-873.
[15] F. Plastria, Continuous location problems and cutting plane algo-
rithms, Thesis (Vrije Universiteit Brussel, 1983).
[16] F. Robert, Meilleure approximation en norme vectorielle et minima
de Pareto, RAIRO Model. Math. Anal. Numer. 19 (1985) 89-110.
[17] R.T. Rockafellar, Convex analysis (Princeton University Press,
Princeton, 1970).
[18] J.E. Ward and R.E. Wendell, Characterizing efficient points in lo-
cation problems under the one-infinity norm in : J.F. Thisse and H.G.
Zoller, eds, Locational analysis of public facilities (North-Holland,
Studies in mathematical and managed economics, 31,1983) pp. 413-429.
RECENT P~SULTS ON DUALITY
IN VECTOR OPTIMIZATION
K.-H. Elster
Technische Hochschule Ilmenau
Sektion 1·1athematik, Rechentechnik
und okonomische Kybernetik
Ilmenau, DDR - 6300
and
A. Gopfert
Technische Hochschule "Carl Schorlemmer"
Leuna-Uerseburg
Sektion Mathematik
Merseburg, DDR - 4200
1. Introduction
In connection with the remarkable and r'apid development of linear

and nonlinear vector optimization, an increasing number of papers
were devoted to the construction and investigation of dual problems.
Several concepts of duality were treated, especially Lagrange duality
and Fenchel duality (cf. Elster /1/, Elster/Iwanow /3/, Jahn /8/).
Apart of linear problems we can find extensive investigations of
convex problems and more and more of nonconvex problems. In the last
case some difficulties can arise concerning p. e. s'eparation theorems
which are sufficiently general. In the present paper a direct duality
theorem is derived for rather general (nonconvex) vector optimization
problems. For the proof a separation theorem is needed where the
separating functional is nonlinear (more precisely: convex). The
results are formulated in finite dimensional spaces, but an extension
to infinite dimensions is possible.
The results may be of interest not only for theoretical reasons

but for the foundation of numerical methods, too. Moreover, duality
assertions are important for classes of such vector optimization
problems where a special structure of the primal (and/or the dual)
problem gives a possibility for additional results. Especially let
us emphasize to geometrical vector optimization problems which often
arise in connection with engineering problems and where the dual
constraints are linear (Elster, K.-H./Elster, R. /2/).
130
2. Separation Assertions
We consider the space Z = RP equipped with the usual (closed convex)

ordering cone R~. To obtain duality theorems (with respect to proper
efficient points but for rather general nonconvex problems) we use
separation theorems which separate a non convex set and a convex one
by a convex (nonlinear) functional.
The idea is due to C. Gerstewitz-Gerth, some versions of the

separation theorem are given by Gerstewitz/Iwanow /5/ and Gerth/
Weidner /6/. Proper efficient points of a vector minimum problem are
obtained via scalarization not in the usual way by a linear continuous
functional z* EZ* (z* is the dual space of Z) but by a convex functional
on Z. Such a functional can be gained (in the simplest case) when - K
is embedded in an open convex set C such that the boundary of C
includes the origin. More precisely we have the assertion of
Theorem 2.1 (Separation Theorem). Let A c Z, A t (2) and let
CcZ, C =I z, C =I' ¢ be an open convex set with

cl C - int K c C. Then holds
A n C (2)
if and only if there is a functional s Z ... R with

the following properties:
(i) s is convex on Z (and thus continuous 1 »,
(ii) s is strictly monotonic with respect to
int K, i.e. z1 Ez 2 +int K ~ s(z1) > s(z2),
(iii) s (Z) R,
(iv) s (x) < °
< s (y) V x E C, VY E int A,
s(x') ~O ~ s(y') Vx' Ecl C, vy' EA.
Proof. In a space with absorbing neighborhoods of ,as in the°

present case), by choosing kO E int K and translating - K along k °,
each point z E RP "( - K) has a representation z = AkO + k with at least
°
one A > and a k E - K. This is true all the more if - K is replaced
by C. Now let us translate using negative A. too. For
1)
In more general spaces the continuity can be proved.
131
SA. : = cl C + A.k,
o A. E R, (2.1)
and each fixed z E Z there exist A., A.' E R such that z ~ SA. and z E SA. ' •
For the functional s we are interested in one can take the following
modified Minkowski-functional
s : z .. R, s(z) := inf{A.\ ZESA.}. (2.2)
If z is fixed then s(z) is defined uniquely since along the line

generated by k O a Dedekind's cut is introduced which determines the
number s (z) 2) •
Both classes of that cut are nonempty which was shown above. The
"upper" class is connected according to
because of
(2.3)
Each A. E R is in exactly one of those classes. Hence s· is always

finite (s is proper).
For any fixed r E R the following relations (2.4) - (2.0) hold.
s(z) :> r <$9> ZEcl C+rkO. (2.4)
In (2.4) the implication (~) is clear by definition.
We assume s (z) :> rand z ~ cl C + rkO. Then z is in the exterior of

cl C + rkO and hence there is a ~ > r such that z ~ cl C + ~kO and thus
s (z) > r which is a contradiction.
By (2.4) we prove the property (i), i.e. the convexity of s. If

zi ECl C+r. kO, r. = s(zi), i = 1,2 and A.E(O,1), thenA.z 1 +
2 ~ ~ °
(1-A.)z Ecl C+ (A.r 1 + (1-?,.)r 2 )k and by (2.4)
s(A.Z 1 + (1-A)z2) :>Ar 1 + (1-A)r 2 = As(z1) + (1-A.)s(z2).
Moreover, we have the equivalence
2) By that Dedekind's cut we obtain in Z the separation.

132
s (z) < r ~ z E C + rkO. (2.5)
If s (z) < r, then z E SA for at least one A < r; hence we obtain z E C +

rkO similar to the proof of (2.3).
Conversely, if c E C then we have c + tkO E C for a sufficient small

t > ° and thus
c + rkO = c + tkO + (r-t)kO E C + (r-t)kO.
Hence follows z E SA. if A r - t < r and finally s (z) < r.
By (2.5) we prove the property (ii), i.e. the monotonicity of s. If

z1 Ez 2 +int K, r<:s(z1), then
z2 E z 1 - int K c cl C + rkO - int K
and, by the assumption concerning C,
By (2.5) we have s(z2) < r
Obviously, together with (2.4) and (2.5) we have the two

equivalences
s (z) > r ~ z ~ cl C + rkO, (2.6)
s(z) <: r ~ z ~ C
°
+ rk . (2.7)
Using (2.4) and (2.7), we obtain
s(z) = r ~ ZEdC + rkO. (2.8)
In the case r =0 we conclude from (2.4) - (2.7) all the assertions (iv)
concerning separation. Since r E R is arbitrary we obtain from (2.8)
the assertion (iii), i.e. s(Z) = R. This completes the proof.
Remarks.
1. The Separation Theorem given above holds also in the cases of

a) linear topological spaces, b) non convex sets (but cl C has to
be convex if s convex is required), c) more general monotonicity
assertions.
133
2.. Essentially for applications of the Separation Theorem is the

avoidance of the interior-point-condition concerning K. Let Z be
a linear topological Hausdorff space with a closed convex ordering
cone K, K ~{O} + ¢. A functional s : Z ~ R is called strictly
monotonic with respect to K ~{O} if
z1 Ez2 + (K~{O}) ~ s(z1) > s(z2).
Then the following assertion holds.
Separation Theorem. Let Z,K as above. A, C c Z, C open and convex,

C + z, AnC = ¢, cl C - (K~{O}) cC and let
Z U (c 1 C - Ak 0 ) ,
AER
where kO E K ~ {a} is fixed. Then there exists a functional s Z~R
such that
(i) s convex and continuous on Z
(ii) s(Z) =R
(iii) s strictly monotonic with respect to K ~ {O} ,
(iv) s(x) < 0:0: sty) VXEC, vyEA,

s(x'):o:O<s(y')VX'EclC, Vy'E intA.
3. If we assume int K +¢ and kO E int K, then we can conclude
Z U (int K - AkO)
7I.ER
and hence, together with cl C - int K c C, the assumption used by

Gerstewitz/Iwanow /5/:
Z = U (cl C - AkO) .
7I.ER
4. To prove the continuity of s in more general spaces than RP let

z EC. Then there is a neighborhood U(x) cC with s(z) <0, z EU(x)
because of proposition (iv) of the Separation Theorem (Remark 3).
Therefore the convex function s is bounded from above and with
that continuous on Z.
3. Duality Assertions
Let X = Rn, Y = Rm, V c Y, Z = RP ; RP partially ordered by R~. We

consider the vector optimization problem
(P) f(x) ~ v - min,

134
xEB:= {UEX! g(u) EV}. (3. 1 )
A pareto-optimal point f(xO) is called properly efficient if there

exists a functional s (which is constructed according to 2.) such
that xO is a solution of the problem
s(f(x» ~ min, xEB. (3.2)
The advantage of the Separation Theorem proved in 2. consists in the

fact, that proper efficiency of f(xO) can be introduced by a
partially order relation too:
f(xO) is pareto-optimal not only with respect to - K but also with

respect to the set C introduced in 2. To prove this we assume that xO
is a solution of the problem (3.2) and define the set C according to
c:= {yEA = f(B)! s(y) < s(O)}, (3. 3)
where ° is presumed as pareto-optimal (translation of f(xO) into the

origin). Conversely, if ° is pareto-optimal with respect to a set C
(cf. Section 2.), then the Separation Theorem yields the scalarizing
functional s.
Due to the formulation of duality assertions we introduce the dual

problem
(D) h ~ v - max, (3.4)
hED:= {hEZ!3S,3LEW s(h)" inf{s(f(x» +L(g(x»}}

XEX
where W = {L : Y ~ 'R! L (v) ,,0 V v E V} and s is a convex strictly

monotonic with respect to R~ (continuous) functional as introduced
in Section 2.
The problem (3.4) can be considered as a generalization of the

dual problem in Jahn /8/.
Always weak duality holds since the dual problem (D) is of the
"Lagrange type". Moreover there exists a strong duality theorem and,
under strong assumptions, an inverse duality theorem, too (/5/, /6/,
/8/) .
135
Theorem 3.1 (Duality Theorem). Each properly efficient point f(xO)

of the problem (P) is pareto-optimal with respect to the problem (D).
Proof. Let f(xO) be properly efficient with respect to (P). Then

there is a functional s according to (3.2) such that
s(f(xO)) = min s(f(x)).

xEB
We set
° if u E V,
L(u) 1 + = otherwise
and obtain
inf[s(f(x)) +L(g(x))] min{inf[ ..• ], inf [ .•. ]

xEX xEB xEX"--,B
inf s (f (x) ) s(f(xO)),

xEB
hence f (xO) ED. Since by the weak duality for any h ED the relation
h-f(xO) ER~"{O} must fail, f(xO) is dual efficient.
In a paper of Klotzler /9/ concerning scalar optimization problems

we can find the idea, the Lagrange functional of the dual problem (D)
not to consider as a linear mapping but as a general one. If we combine
that approach with the construction of the scalarizing functional s we
get the possibility to generate general vector optimization problems
without duality gaps.
Using a result of Ester/Schwartz /4/, Gerstewitz/Iwanow /5/ showed

the explicit construction of a functional s definitively nonlinear.
Further investigations will be devoted to use dual vector optimization
in connection with interactive methods.
4. References
/1/ ELSTER, K.-H.: Neuere Entwicklungen der Vektoroptimierung.

X. Internat. KongreB tiber Anwendunqen der Math. i. den Ingenieur-
wissenschaften, vleimar 1984. Berichte Heft 4, S. 63 - 66.
/2/ ELSTER, K.-H./ELSTER, R.: Relationships between posynomial

programming and vector optimization (To appear).
136
/3/ ELSTER, K.-H./IWANOW, E.: tiber Fortschritte auf dem Gebiet der
mehrkriteriellen Entscheidungsprobleme. Wis-s. Zei tschr. TH
Ilmenau 11. (1985) 2, 15 - 36.
/4/ ESTER, J./SCHWARTZ, B.: Ein erweitertes Effizienztheorem. Math.
Operationsforschung u. Statist., sera optimization 1! (1983),
331 - 342.
/5/ GERSTEWITZ, C./IWANOW, E.: Dualitat fur nichtkonvexe Vektor-
Optimierungsprobleme. Wiss. Zeitschr. TH Ilmenau 31 (1985) 2,
61 - 81. -
/6/ GERTH, C./WEIDNER, P.: Nonconvex separation theorems and some
applications to the vector optimizatiori (To appear in JOTA).
/7/ G~PFERT, A.: Multicriteria duality, examples and advances:
Lecture Notes in Econ. and Math. Systems Nr. 273, Springer
Verlag 1986, 52 - 58.
/8/ JAHN, J.: Mathematical Vector Optimization in Partially Ordered
Linear Spaces. Meth. u. Verf. der math. Physik Bd. 31. Vlg.
Peter Lang, Frankfurt/M. (1986).
/9/ KL~TZLER, R.: Dualitat bei diskreten Steuerproblemen. Math.
Operationsforsch. u. Statist., sera optimization 12 (1981) 3,
411-420. -
/10/ WEIDNER, P.: Charakterisierung von Mengen effizienter Elemente
in linearen Raumen auf der Grundlage allgemeiner Bedingungen.
Diss. A, MLU Halle-Wittenberg (1995).
A RESULT OF FARKAS TYPE AND DUALITY
IN VECTOR OPTIMIZATION
Joachim Gwinner
Fachbereich Mathematik, Technische Hochschule Darmstadt
SchloBqartenstr. 7. D-6100 Darmstadt
ABSTRACT. By a minimax approach we prove a result of Farkas type and

use it to derive duality for vector optimization problems with and with-
out constraint qualification.
1. INTRODUCTION
Using a minimax approach we established duality for a general class of

homogeneous programs in [Gw-1]. Here we modify this approach and first
prove a result of Farkas type. Similar to [Gw-1] we replace the usual
bilinear coupling <y,g(x» between the dual variable y and the con-
straint g by a convex/concave-positively homogeneous function G
Instead of a scalar objective function we introduce a convex/concave-
positively homogeneous function F which admits not necessarily linear
scalarizations. More precisely, Theorem 2.2 in section 2 asserts that
under certain topological and regularity assumptions the subsequent two
statements are equivalent:
xeX , (vyeY) G(x,y) ~ 0 =9 (3zeZ\{O}) F(x,z) ~ 0 (1. 1 )
(3zeZ\{O}, yeY) (VxeX) G(x,y) + F(x,z) ~ 0 . (1 .2 )
Then we essentially follow the approach of Oettli in [Oe-1] and investi-

gate weakly minimal, respectively weakly maximal elements of a pair of
general convex vector programs. After proving a weak duality result
(Proposition 3.1), we apply Theorem 2.2 to establish a strong duality
result imposing a constraint qualification (Theorem 4.1).
Moreover, stimulated by the work of Craven and Zlobec [Cr-Zl] and of

Borwein and Wolkowicz [Bo-Wo], similar to [Gw-1], we split the con-
straints, modify the dual program and can thus drop the constraint quali-
fication (Theorem 5.2).
Finally in our concluding remarks we compare our results to other known

duality results in the literature and indicate possible extensions.
For background, definitions and notations not explained in this paper

we refer to the book of Jahn on vector optimization [Ja-1].
138
2. A RESULT OF FARKAS TYPE
Throughout the paper, a real-valued function ~ defined on some product

set XxY of convex sets X,Y in linear spaces is termed a saddle
function I if for each yeY, ~(.,y) is convex on X and for each xeX,
~(x,') is concave and moreover upper semicontinuous on Y, provided
Y is contained in a linear topological space. If in addition Y is a
cone, a saddle function ~: XxY - R is called a partly homogeneous
saddle function, if for each xeX, ~(x,.) is positively homogeneous.
Instances of such functions ~ are positive sums of functions of the
following type
~ fx, y) = f (x) ·0 (y) + g (x) • ~ (y) ,
where f is convex and nonpositive on X ,0 is sublinear and non-

negative on Y, A is a linear form, nonnegative on Y and g is ar-
bitrarily convex on X. Nonpositive convex functions abound. Indeed, on
the real line we have the following nonlinear examples (x e R ):
fIx) -(ax + b)a (0 < a < 1 ; a,b > 0) ,

-bx
f (x) ale - 1) (a,b > 0)
f (x) a arc tan (-bx) (a,b > 0)
fIx) -a In (1 +x) (a > 0; x 2:. 0) ,

f (x) a In ( 1 + bx - b /x 2 +2x) (a,b > 0; x 2:. 0) ,
3
fIx) a sin ["2 11 - (bx + ~) -1 ]
11
(a,b > 0, x 2:. 0)
Instances of nonnegative sublinear functions in Hilbert spaces are given

by
1/2
o(y) = (y,By) ,
where B is a symmetric, positive semidefinite operator; or more gen-

erally in linear topological spaces by
o(y) = max {<y,n>1 n e ~} ,
where ~ is a weak * compact subset in the dual space containing zero.

After these preliminaries we can state the basic result in our minmax
approach.
Theorem 2.1. Let C be a convex set in a linear space and let K be

a convex, compact set in a linear topological space. Let ~: C x K R
be a saddle function. Then the following statements are equivalent.
(Vx e C) (3y e K) ~(X:,y) 2:. 0; (2.1 )
(3y e K) (Vx e C) q,(X,y) 2:. O. (2.2)

139
The p~oo6 of this minmax type result can be given using the separation
theorem [see Gw-1, §2]; Theorem 2.1 can also be directly derived from a
more general minmax theorem due to Fuchssteiner and Konig [Fu-KO], based
itself on a general Hahn-Banach principle.
Here and throughout the paper we work in the following setting. X is a

convex set in a linear space; Y ,respectively Z is a closed convex
cone in a linear topological space that admits a convex compact base
Y
0
, respectively Z
0
not containing the respective zero element; i.e.
we have Y = R Y and Z = R +Zo with R the set of nonnegative
-
+ 0 +
reals. Moreover let G : XxY R and F : XxZ R be partly homo-
geneous saddle functions. Then we can assert the subsequent Farkas type
result.
Theorem 2.2. Suppose,
y e Y , (Vx e X) G(x,y) ~ 0 ~ y = 0 . (2.3)
Then the following statements are equivalent.
x e X (vy e Y) G ( x , y) ~ 0 ~ ( 3Z e Z\ { O}) F ( x , z) ~ 0 (2.4 )
(3ye Y , Z e Z\{O}) (Vx e X) G(x,y) + F(x,z) ~ 0 • (2.5 )
P~oo6. Defining
tp(x) max{F(x,Z>1 Z e Zo}
y(x) max{G(x,y)1 y e Yo}
(2.4) is equivalent to
x eX, y(x) ~ 0 ~ tp(x) ~ 0 .
Using the simplex
this implication results in
Theorem 2.1 applies to the function
und thus we obtain
(3s = (s1,s2) e S)(Vx e X) s1Y (x) + s2"(x) ~ 0 .

We keep s # 0 fixed and have more detailed
Z e Z )
o
140
Once again we apply Theorem 2.1 now to the function
cjJ (x, (y, Z) )
on X x (Y xz ) • Hence we arrive at the statement

o 0
(2.6)
Finally we set y = s1 Yo e Y, Z = s2zo e z. Since (s1 ,s2) ~ 0, Z = °

cannot occur in virtue of (2.3). Therefore (2.6) yields (2.5). - The
reverse implica~ion (2.5) ~ (2.4) is trivial. q.e.d.
2
We remark that the conclusion (2.6) with (s1,s2) e R+\{O,O} is re-
lated to F. John's optimality conditions [Ma-1, 5.4.1]. To guarantee
s2 ~ ° or Z ~ ° we need a constraint qualification (CQ), here (2.3),
which is a natural extension of Karlin's CQ [Ma-1, 5.4.4; Ka-1, Th. 7.1.1]
to our constraints (Vy e Y) G(x,y) ~ 0 . The more familiar Slater's CQ ,
formulated here as
(3 ~ eX) (Vy e Y \ { 0 } ) G( x , y) < 0 (2.7)
clearly implies (2.3). Actually (2.7) is equivalent to (2.3); this can be

seen by a minmax argument, since Y is assumed to be convex and compact.
o
3. THE DUAL PAIR OF VECTOR PROGRAMS
In addit~on to the setting of §2 we introduce the linear topological space

E containing the*cone Z, and let Ez b~ paired with another linear to-
z
pological space E z . The duality on *Ez~Ez will be denoted by <.,.> ,
and etc. are linear forms in E. Then we can introduce the feasible
Z
set of our primal vector program, respectively of our dual vector program
as follows:
V 0.1 (3x e Xl: (Vy e y) G(x,y) ~ 0 t.,(Vz e Z) F(x,z)~ <A,Z>} (3 •1 )
W {AI (3y e Y, z e Z\{O}) (Vx e X) G(x,y) + F(x,z).2: <A,Z>} (3.2)
This gives rise to determine V-minimal vectors A (primal program P

defined by
X e V A (VA e V) (3z e Z\{O}) <A-X,Z> .2: 0, (3.3)
respectively to determine W-maximal vectors X (dual program D) defined

by
X e W A (VA e W) (3z e Z\{O}) <A-X,Z> ~ °. (3.4)
Comparing to the optimality notions in [Ja-1], (3.3), respectively (3.4)

gives a dual characterization of weakly minimal, respectively weakly maxi-
mal elements using the dual cone of an ordering cone with non empty interior.
A direct consequence of the definitions is the subsequent weak duality
141
result.
Proposition 3.1. Let A1 e V, A2 e W. Then there exists some z e Z\{O}

such that <A 1 -A 2 ,Z> ~ O.
Pltoo6. Let x 1 e X be associated to A1 e V by (3.1), and let Y2 e Y,

z e Z\{O} be associated to A2 e W by (3.2). This gives in particular
(3.5 )
G{x 1 'Y2) :>. 0 ,
F(X 1 ,z) :>. <~1 ,z> , (3.6)
(3.7)
G(x, 'Y2)+ F(X, ,z) ~ <~2'z>
Multiplying (3.5) and (3.6) by -1 and adding up all three inequalities,
we come up with the claimed inequality <A _A z>:>. o.
2 l' q.e.d.
At once we recover from Proposition 3.1 the well-known weak duality

relation inf(P) ~ sup(D) in scalar optimization by the choice
Z = R+,ZO = {+1}.
4. STRONG DUALITY WITH CONSTRAINT QUALIFICATION
This section presents our strong duality theorem for the dual pair of
veotor programs introduced in §3. Here we reqnire Karlin's CQ (2.3)
or equivalently Slater's CQ (2.7) as hypothesis.
Theorem 4.1. Let ~ e V. Suppose,
y e Y, (vx eX) G( x , y) ~ O"q Y = O. (4.1 )
Then ~ is V-minimal, if and only if ~ is W-maximal.
The pltoo6 essentially follows the lines of Oettli [Oe-1] and consists
of three parts.
1 .) We show: ~ is V-minimal ~~ e W.
By (3.3) we have
x eX, Ae E *z ' (vy e Y) G(x,y):>. 0 A(VZ e Z) F(x,z) -<A,z>:>. 0
=0> (3.z e Z\{O}) <A-~,Z> ~ O.
This implication writes as follows
x e X, (Vy e y) <Hx,y) :>. 0 :::=> (3 z e Z\{O}) <A -~, Z> ~ 0,
where x = (x, A) varies in X= XxE * y varies in Y = YxZ, and

z'
G(x,y) = G(x,y) + F(x,z) - <A,Z> .
To apply Theorem 2.2, we have to check (2.3). Therefore we consider the

142
system
(Y.z) e Y"Z • (Vx e X. A e E*z ) G(x.y) + F(x.z) - <A.Z> ~ 0 .
Now fix x and vary A in the cone E*z • this implies
hence z = 0 . It remains to consider
y e Y • (Vx e X) G(x.y) ~ 0 .
In virtue of (4.1) • y = 0 follows and (2.3) is satisfied.
Theorem 2.2 yields the existence of y e Y. z1 e Z. z2 e Z\{O} such that
Now vary A e E*
z only; this implies z1 and we obtain
(3y e Y. z e Z\{O}) (Vx e X) G(x.y) + F(x.z) ~ <I.z> •
hence I e W
2.) We show: I is V-minimal -? I is W-maximal.

By 1.) I e Wand by Proposition 3.1, for any A e W there exists
z e Z\{O} such that <I-A.Z> ~ 0
3.) I e V is W-maximal -? r is V-minimal.

This implication follows immediately from Proposition 3.1. q.e.d.
We remark that this strong duality result extends earlier work of Aubin
[Au-1]. Corley [Co-1]. and using the convex core topology also the neces-
sary optimality conditions of Kirsch-Warth-Werner [K-W-W, II §3]. where,
however. affine equality constraints are also considered.
5. STRONG DUALITY WITHOUT CONSTRAINT QUALIFICATION
First we reformulate the primal feasible set V in (3.1) as follows:
V {AI (3X e X) (Vz e Z) F(x.z) ~ <A.Z>} •
where
(x e X I (V Y e y) G ( x • y) ~ O}. (5.1 )
Now let S be a generating set of X, i.e. S is a subset of Y such
that x e X belongs to il: if and only if
(vy e S) G(x,y) ~ 0
holds. S may coincide with Y. Then motivated by recent work of Craven

and Zlobec [Cr-Zl] and of Borwein and Wolkowicz [Bo-Wo] in scalar opti-
143
mization we introduce the sets

{y e sl (Vx e 50 G(x,y) O} ,
We note that Slater's CQ (2.7) implies that S= contains the zero ele-
ment only and thus excludes equality constraints.
With the splitting
S< SUR,
x {x e X I (Vy e S= U R) G( X t y) ~ O} ,
Y R +5
the primal feasible set V is given by
V = tAl (3x e X) : (Vy e Y) G(x,y) ~ 0 A(VZ e Z) F(x,z) ~ <A,z>t(5.2)
Replacing X by X and Y by ~ we arrive at the modified dual pro-

gram D to determine W-maximal vectors analogously defined as in (3.4),
where
-
W = {AI (3y e Y, z e Z \{O}) (Vx e X): G(x,y) + F(x,z)~ <A,Z>} • (5.3 )
In the sequel we need the following hypothesis:

(H) There exists a splitting S< = SuR with § convex compact.
By the choice S = {Yo} for some Yo e s<, (H) can be trivially satis-
fied. However, in this case Y is only one-dimensional and the dual D
is not very meaningful. Therefore S in (H) should be taken as large
as possible such that most constraints G(o,y) remain explicit con-
straints not hidden in the set X.
To establish strong duality without constraint qualification, we rely
on the subsequent lemma, which is a variant of Lemma 2.2 in [Bo-Wo].
Lemma 5.1. Let S< be defined as above and let (H) be satsfied. Then
there exists some x e X such that G(x,y) < 0 holds for all yeS.
For the p~oo6 we refer to [Bo-Wo] or to the proof of the more closely
related Lemma 6.1 in [Gw-1].
Now we can present the promised strong duality result.
Theorem 5.2. Let S< be defined as above and let (H) be satisfied.
Let Xe V. Then X is V-minimal, if and only if X is W-maximal.
P~oo6. In virtue of Lemma 5.1, the constraint qualification (4.1) is

satisfied (with
-
X and y) for the primal program P with its feasi-
ble set V given by (5.2). Therefore Theorem 4.1 applies to conclude
the claimed duality result. q.e.d.
144
A related splitting approach to dispense with constraint qualifications

in vector optimization can already be found with Ben-Israel, Ben-Tal
and Charnes [B-B-C]. In contrast to the present paper, they characterize
an efficient point (Pareto optimum) in finite dimensional convex pro-
grams with polyhedral cones Y and Z.
6. CONCLUDING REMARKS
There are several other concepts of optimality in vector optimization

[see Ja-1, 11.4]. The most useful notion from the practical point of
view is that of a Pareto optimum (minimal, respectively maximal element).
However using this notion, the symmetry between the dual problems is
lost. Indeed, Jahn [Ja-2, Th.2.3; Ja-1,II. 8.2] established strong dual-
ity between p~ope~iy minimal elements of the primal problem and maximal
elements of the dual problem. On the other. hand, for any properly mini-
mal element there is an associated scalar convex optimization problem
by definition. Thus the Farkas type result, Theorem 2.2, can be applied
(with Z = R+, Zo = {1}) to conclude the strong duality theorem of Jahn,
also for the constraints (Vy e Y) G(x,y) ~ 0; moreover the splitting
technique can be carried through to dispense with any constraint quali-
fication or normality hypothesis.
The convexity assumptions can be relaxed in §2 and §4. Note that

Proposition 3.1 holds without convexity, whereas Lemma 5.1 makes deci-
sive use of convexity. Already Oettli [Oe-1] presupposes for his strong
duality theorem that a certain composed set, only, is convex and closed.
Here instead of Theorem 2.1, we can directly apply the minmax theorem of
Fuchssteiner and Konig [Fu-Ko, Th. 5.2] which works for a class of weak-
ened convex-like / concave-like functions. This concept is not stable
under addition, however. Therefore all relevant composed functions must
then be assumed to belong to this relaxed class. For investigations in
this direction - in the scalar optimization case - we refer to the forth-
coming paper [C-G-J].
Added Note: Another weakening of convexity assumptions can be found in

the recent paper of Craven, Glover, and Zlobec [C-G-Z]. They extended
the original Borwein - Wolkowicz lemma to pseudoconvex constraints and
obtained modified necessary Lagrangean conditions for a differ-
entiable objective function and pseudoconvex constraints, also without
requiring any constraint qualification.
145
REFERENCES
[Au-1] J.P. Aubin: A Pareto minimum principle. In: Differential Games

and Related Topics (H.W. Kuhn, G.P. Szego, eds.). North-Holland,
Amsterdam, 1971; pp. 147-175.
[B-B-C] A. Ben-Israel, A. Ben-Tal, and A. Charnes: Necessary and suffi-

cient conditions for a Pareto optimum in convex programming.
Econometrica 45 (1977), 811-820.
[Bo-Wo] J.M. Borwein and H. Wolkowicz: Characterizations of otimality

without constraint qualification for the abstract convex pro-
gram. Math. Programming Study 19 (1982), 77-100.
[Co-l] H.W. Corley: Duality theory for maximations with respect to

cones. J. Math. Anal. Appl. 84 (1981), 560-568.
[C-G-J] B.D. Craven, J. Gwinner, and V. Jeyakumar: Nonconvex theorems

of the alternative and minimization (to appear in Optimization).
[C-G-Z] B.D. Craven, B.M. Glever, and S. Zlobec: On minimization subject

to cone constraints. Numer. Funct. Anal. Optimization 6 (1983),
363-378.
[Cr-Zl] B.D. Craven and S. Zlobec: Complete characterization of opti-

mality for convex programming in Banach spaces. Applicable Anal.
11 (1980), 61-78.
[Fu-Ko] B. Fuchssteiner and H. Konig: New versions of the Hahn-Banach

theorem. In: General Inequalities 2 (E.F. Beckenbach, ed.).
Birkhauser, Basel, 1980; pp. 255-266.
[Gw-l] J. Gwinner: An extension lemma and homogeneous programming.

J. Optimization Theory Appl. 47 (1985), 321-336.
[Ja-l] J. Jahn: Mathematical Vector Optimization in Partially Ordered

Linear Spaces. Methoden und Verfahren der mathematischen Physik,
Vol. 31. Peter Lang, Frankfurt, 1986.
[Ja-2] J. Jahn: Duality in vector optimization. Math. Programming 25

(1983), 343-353.
[Ka-1 ] S. Karlin: Mathematical Methods and Theory in Games, Pro-

gramming, and Economics, Vol. 1. Addison-Wesley, Reading, Mass.,
1959.
[K-W-W] A. Kirsch, W. Warth, and J. Werner: Notwendige Optimalitats-

bedingungen und ihre Anwendung. Lecture Notes in Economics and
Mathematical Systems, 152. Springer, Berlin, 1978.
[Ma-1] O.L. Mangasarian: Nonlinear Programming. McGraw-Hill, New-York,

1969.
[Oe-1] W. Oettli: A duality theorem for the nonlinear vector-maximum

problem. In: Progress in Operations Research (A. Prekopa, ed.).
North-Holland, Amsterdam, 1976; pp. 697-703.
PARAMETRIC OPTIMIZATION WITH A BOTTLENECK OBJECTIVE
AND VECTOR OPTIMIZATION
Siegfried Helbig
Johann Wolfgang Goethe-Universitat
Fachbereich Mathematik
Robert Mayer-StraBe 6-10, D-6000 Frankfurt/Main
Abstract
In this paper, efficient and weakly efficient points of a set are

characterized by an optimization problem with a parameter in the bottle-
neck objective function. The investigation of the structure of the pa-
rameter set leads to a necessary and sufficient criterion for the non-
emptiness of the set of efficient points. The continuous dependence of
the optimization problem on the parameter is investigated. Finally, the
results are applied to a scalarization of multi-objective optimization
problems.
I. Introduction
Let M be a non-empty subset of RI, I € M. Assume that RI is par-

tially ordered by the cone
C := {y € RI I Yi ~ 0 for i=1, ••• ,I}.
In this paper, we consider two types of optimal points of M.
Definition 1.1.: a) A point v € M is called efficient if and only if
Y € M, Y $ v - Y = v.
b) A point v € M is called weakly efficient if and only if
Y € M, Y ~ v - 3 Yi Vi.
i€{1, .•• ,I}
Let S be the set of all weakly efficient and Seff the set of all effi-
cient points of M. Then consider the vector optimization problem
147
VOP: Determine efficient (resp. weakly efficient) points Qf M.
In section II, we present an optimization problem with a parameter in

the bottleneck objective function (called OP(q», which has a close re-
lationship to VOP. This enables us to characterize the efficient and
weakly efficient points of M. The investigations of the structure of
the parameter set in section III leads to a necessary and sufficient
criterion for Seff ~ ~ and a decomposition of the parameter set. Sec-
tion IV deals with the question of the continuous dependence of OP(q)
on the parameter. These results can be applied to a scalarization of
VOP, which is due to BROSOWSKI and CONCI [4] (see section V).
II. The bottleneck optimization problem
Let M be a non-empty subset of RI and let q € RI. Then consider the

following bottleneck optimization problem:
OP(q): min
subject to y € M.
For abbreviation, define (q,y) ;= I

max (qi+Yi) for q, y € I
i.=1 , ••• ,I
Denote the optimal value and the set of optimal pOints of OP(q) by
:= inf (q,y)
y€M
and
;= {y € M I (q,y) = E }.
q
Let LM
be the set of all parameters q € RI, for which the problem OP(q)
has a solution, i.e.
The next theorem shows the relationship between VOP and OP(q) and bet-
ween the sets Seff (resp. S) and Pq :
148
Theorem 11.1.: a) For q € LM, let v be an optimal point of OP(q). Then

v is a weakly efficient pOint of M.
b~ For q € LM, let v be the unique solution of OP(q), i.e. P q {v}.
Then v is an efficient point of M.
c) Let v € Seff (resp. v € S). Then P q = {v} (resp. v € P q ) for all
q € Rl with q = -v + ae, where a € Rand e = (1, ••• ,1) € Rl.
Remark: BROSOWSKI and CONCI [4] proved part c) for a vI in the spe-
cial case of a set M described in section V.
Proof: a) If v is not weakly efficient, then there exists u € M with

u i < vi for i=1, .•. ,1, which implies (q,u) < (q,v). This is a contra-
diction to the optimality of v.
b) If v is not efficient, then there exists u € M with u ~ v and u ~ v.
The inequalities
for i=1, ... ,1
leads to (q,u) ~ (q,v). Since u € M, u is an optimal point for OP(q).

This is a contradiction to the uniqueness of v.
c) Obviously, we have
(q,v) max (qi+vi) max a=a.

i=1, .•• ,1 i=1, ..• ,1
If there exists u € M with (q,u) < a, then
for i=1, ..• ,1,
which implies u i < vi flir i=1, ... ,1. This contradicts v € Seff (resp.
v € S). In the case of efficiency of v, we have
({v} - C) n M = {v},
which proofs the assertion. o
A direct consequence of this theorem is a characterization of efficient

and weakly efficient points of VOP:
149
Corollary II.2.: a) A pOint v € M is efficient if and only if there ex-

ist real numbers q1, .•• ,ql such that
(q,v) ..; (q,y) for all y E M, Y ~ v.
b) A point v E M is weakly efficient if and only if there exist real

numbers q1, ... ,ql such that
(q,v) s (q,y) for all y E M. o
This result shows that even for non-convex and non-compact sets M the
efficient and weakly efficient points can always be characterized by an
optimization problem, which is nearly a discrete Tschebyscheff-approxi-
mation problem. Other authors corne to similiar conclusions:
JAHN [6] uses a "parametric norm" (1. e. a koind of weighted discrete

Tschebyscheff-norrn) to characterize efficient and weakly efficient
points under the assumption that M is bounded from below.
In the case of compactness of M, BOWMAN [1] gives a necessary condition

for efficient points as a best discrete Tschebyscheff-approximation. We
get his result (see [1], Theorem 1) as corollary of theorem II.1.c):
For this, let M be a compact set and put Y1* := min Y1 and K1 := M. Then
define recursively for i=2, ••• ,1 YEM
Ki - 1
Yj* for j
Ki := {y < i}
E
Yj
Yi* := min Yi·
y€K i
It follows:
Corollary II.3.: If v is an efficient point of M, then there exists

an element PERI with P1 = 0 such that v is a solution of
min
YEM
Proof: With Pi := Vi - y~ - (v 1 - y~) for i=1, ••• ,1, the assertion

follows with theorem II.1.c). 0
150
Part c) of theorem 11.1. shows how we can restrict ourselves to a sub-

set of the parameter set LM, because of
Lemma 11.4.: Let q ELM' Then P P q + ae and Eq+ae Eq + a for all a in

1 q
R, where e = (1, .•• , 1) E R •
Proof: Let v E P q . If there exists y E M with
(q+ae,y) a + (q,y) < a + (q,v)
then this will be a contradiction to v P q . Hence p

E
q c P q + a e and
Eq+ae Eq + a. The converse inclusion follows immedia~ely. 0
This lemma implies that it suffices to take the parameter set
III. The parameter sets LM and LM
In this section, we assume that M is a convex, closed and non-empty

subset of Rl. For the following, we need the useful
Lemma 111.1.: Let (vk ) be a sequence in M with ~vkll2 + 00, which is

bounded from above. Then P q is not compact for all q ELM'
Remark: The main idea of the proof below is due to COLGEN and SCHNATZ.
In [5], they showed a similiar result for linear optimization with vari-
able objective function.
Proof: Define an index set I by
k
I := {i E {1, ..• ,1} I (Vi) is unbounded},
which is non-empty by assumption. The sequence (Vk/nvkU2) has a cluster

point y E Jl.l with IIyll2 = 1. W.l.o.g. let
y.
151
Since Oy02 = 1, the set I := {i {1, ... ,1} I Yi f O} is non-empty.

€
Hence, Yi < 0 for i € I c I and Yi = 0 for i € {1, ... ,1}\I. Let v be an

L
optimal point of OP(q), q € M. Since M is convex, for arbitrary A 2: 0,
A € R, there exists ko € N such that for all k 2: ko
Since M is closed, we obtain
v + AY E M for all A 2: o.
From this, we conclude far A 2: 0
for i E {1, .•. , 1 hI

and
for i E I,
which implies (q,V+AY) Eq and V+AY E P q for all A 2: O. Hence, P q is

not compact. o
With this lemma, we get
Theorem 111.2.: Let S ff be non-empty. Then

1 e 1
a) = RLM(resp. LM = {q E R I ql = a}).
b) P q is compact for all q ELM.
Proof: a) Let p be in Rl with P =~. This implies the existence of a
k p k
sequence (vk ) in M with Ov "2 + 00 and (p,v ) + Ep (Ep may be -00). Hence,
the index set
I := {i E {1, .•. ,I} I (v~) is unbounded}
is non-empty. To apply lemma 111.1., we have to show that the sequence

(vk ) is bounded from above. This is obvious for i E {1, .•• ,1}\I. If
k
vi + 00 for some i E I, then
which is a contradiction to the definition of Ep. For that reason, the

152
lemma above implies that P q is not compact for all q € LM•

Since Seff ~ ¢, theorem II.1.c) implies that
{v} for v € Seff and q -v.
Obviously, P q is a compact set. This is a contradiction to the above

conclusion.
b) If Pp is not compact for some p € LM, then there exists a sequence
(vk ) in Pp with ~vk~2 + 00. Since v k is optimal for k € N, we obtain
for i=1, ... ,l.
Hence, the sequence (vk ) is bounded from above. Applying lemma 111.1.,
we get a contradiction to the compactness of P q for q = -v, V € Seff. 0
This theorem leads to a necessary and sufficient criterion for the non-
emptiness of Seff.
Corollary 111.3.: The following statements are equivalent:

a) Seff ~ ¢;
b) For each q € L
M, the set P q is compact.
c) There exists some q € LM
such that P q is compact.
Proof: Because of theorem 111.2. it suffices to show c) ~ a).

Since P q is compact, there exists an efficient point v in P q . To show
that v is an efficient point of M, i.e. v € Seff' let u be in M
such that u ~ v. Then the inequality
(q,u) ~ (q,v)
implies u € P q • Hence, u v and the assertion holds. o
The next example shows that the assumption Seff ~ ¢ is necessary in

part a) of the above theorem even in the case that M is described by
linear inequalities.
Example 111.4.: Let T = [O,1J u {2} be given. Furthermore, let 1 3.

Define the continuous mappings
153
B(t) := (O,-1,t 2 ) for t E [0,1 ]

B(2) := (-1,0,0)
bet) := -2t for t E [0,1 ]
b(2) := O.
Let M := {x E R3 I <B(t),x> ~ bet) for all t E T}, where <.,.> denotes

the usual inner product. Then consider for q E R3
OP (q) : min (q,y)

subject to y E M.
Let q = (0,-2,0). We claim that P q is non-empty, but not compact. For

A ~ 0, A E R, let VA := (O,2,-A). Obviuously, VA E M for A ~ O. To show
that v A E P q , let u E M with
(q,v A ) = 0 > (q,u) ~ qi + u i for i=1,2,3.
Especially, this implies u 1 < 0, which is a contradiction to u E M.

Hence, VA are in the optimal set for A ~ 0 and P q is not compact. With
corollary III.3., it follows Seff = ¢, and with theorem II.1.b), we ob-
tain s t- ¢.
Next, we show that Pp is empty for p = (0,0,0). For £ > 0, £ E R, the
pOints v£ := (0,£,-1/£) are in M, since
o = b(2)
and
bet) for t E [0, 1 ] •
An upper bound for the optimal value Ep is given by the estimation
E inf max (0,£,-1/£) O.

P £>0
But v (O,O,V 3 ) is not in M for any choice of v 3 • o
These investigations leads to a decomposition of the parameter spaces

LM and LM· For this purpose, define for y E M
Note, that in general the sets r(y), y E M, are not convex.

154
Theorem 111.5.: a) Let Seff ~ ~. Then
LM = u r (v) •
VES eff
b) Let Seff ~ and S ~ ~. Then
u r (v) •
YES
Remark: By definition of LM, its decomposition is obvious.
Proof: a) With theorem 111.2., we obtain that P q is compact for q ELM.

Like in the proof of corollary 111.3., we conclude that there exists
an efficient point v in P q , which is efficient in M. Hence, q E r(v).
b) Follows immediately with theorem II.1.a). 0
We give some examples:
Example III. 6. :
a)
All points on the boundary of M are efficient. Theorem 111.2. implies

2
that LM = R • Some of the sets r(v), v E Seff' are shown in the right
figure.
b) In this example the set of efficient points is drawn more intensive.
Theorem 111.2. implies that LM = R2.
155
y."--~
c) In this example even the set of weakly efficient points is empty,

which implies that LM = ~.
IV. Continuity properties of the optimal value and the optimal set
Let M be a non-empty subset of Rl. The optimization problem OP(q)

depends on the parameter q € Rl. Therefore, we are interested in the in-
vestigations of the mappings E:LM -->R, which is called the optimal-
value-mapping and assigns to each q E LM the optimal value E q , and
P:LM -->POW(R l ), which is called the optimal-set-mapping and assigns
to each q € LM the optimal set P . Here, POW(R1) denotes the power set
1 q
of I .
Theorem IV.1.: The mapping E is continuous in q E int LM, where int de-
notes the topological interior.
Proof: Let E > 0 be given. Furthermore, let a > 0, such that a $ E/2
and Ua(q) eLM' where Uo(q) := {p E R
1
max Ip.-q. I < a}.
i=1, .•. ,1 ~ ~
156
Define a set K by
K := {y E M I (q,y) ,;:; Eq + Ill.
Let p be in UIl{q) and y E K. It follows
Pi < qi + I)
for i=1, •.• ,1,
qi + Yi ,;:; Eq + I)
which implies
for i=1, ..• ,1.
Hence,
On the other side, we obtain for u E Pp
for i=1, ..• ,1.
For this,
which completes the proof. 0
lUth the continuity of the obj~Gtive function of OP(q), it follows

immediately
Theorem IV.2.: The optimal-set-mapping P is a closed mapping. o
It should be mentioned that theorems IV.1. and IV.2. hold without as-
sumption to M. Another continuity property of P is shown in
Theorem IV.3.: Let M be a convex, closed and non-empty set. If Seff is

non-empty, then the optimal-value-mapping P is an upper-semiconti-
nuous mapping.
157
Proof: If P
is not upper-semicontinuous in q € M, then there exist an L
open neighbourhood W of Pq and a sequence (qk) in M, which converges L
to q, such that P k\W ~ ~. Let v k € P k\W. If (vk ) is a bounded sequen-
q q
ce, tben we get a contradiction with theorem IV.2 .. Therefore, let (vk )
be an unbounded sequence with ~vkn2 + 00. Since the optimal-value-mapping
E is continuous (note that LM
= BI is an open set if Seff ~ ~), for
sufficiently large k
For this, (qk,vk ) < Eq + E/2 for sufficiently large k. Let ko be in M

such that qk is in {p € RI I max Iq.-p. I < E/2} for k ~ k o . This
i=1, ..• ,1 1 1
implies
for i=1, .•• ,I.
Hence, the sequence (vk ) is bounded from above. Applying lemma 111.1.,
the optimal sets P p ' p € L
M, are not compact. With corollary 111.3.,
this is a contradiction to Seff ~ ~. D
v. A scalarization of VOP
Let T be a compact Hausdorff-space, n € m, and let
A: TxRn ->R
a: T ->R
I
Pi: R ->R for i=1, •.. ,1
be continuous mappings. Furthermore, let
Z := {x € Rn I A(t,x) s a(t) for all t € T}
be the set of feasible points. We call x € Z p-efficient (resp. weakly

p-efficient) if and only if p(x) := (P1 (x), .•. ,PI(x» is efficient
(resp. weakly efficient) in the image set p(Z). In order to solve
VOP: Determine p-efficient (resp. weakly p-efficient) points in Z
one replaces it by a family of scalar (i.e. ordinary) optimization pro-

158
blems. One of these scalarizations is due to BROSOWSKI and CONCI [4]:
1
For each A E A := {p E R I Pl O} consider the minimization problem
MP(A): min z
subject to
A(t,x) s a(t) for all t E T
Pi(x) - Ai s Z for i=1, ••• ,l.
Denote by EA and P A the optimal value and the optimal set of MP(A).
With M p(Z), it is obvious that
(i) (x,z) EPA' A E A p(x) E P q for q -A.

(ii) p(x) E P q , q E LM (X,E q ) E P A for A = -q.
For that reason, we can apply the results of the former sections:
A) Characterization theorem
BROSOWSKI characterizes in [3J weakly p-efficient points of Z by a
Kolmogoroff-type-theorem. In the linear case, i.e. that Z is described
by linear inequalities and Pi are linear for i=1, ••. ,l, a Kuhn-Tucker-
type-theorem holds for weakly p-efficiency (see [3]). Our characteri-
zation theorem reads as follows (compare corollary 11.2.):
B) Properties of the parameter set

Suppose that p(Z) is closed and convex. Then example 111.4. shows
that the solvability of MP(A), i.e. P A ~ ¢, for some A E A does not im-
ply the solvability of the scalarized problem for all parameters. But
corollary 111.3. says:
MP(A) is solvable for some A E A with compact set PA'

~ MP(A) is solvable for all A E A with compact set PA'
~ The set of p-efficient points is non-empty.
These statements are of interest in connection with an algorithm of

BROSOWSKI [3] for VOP in the linear case. This algorithm starts with
A E A and reduces in each step the number 1 of objective functions to
a number k s 1-1 by solving the scalarized problem. Now, the above con-
clusions of corollary 111.3. say
159
either: for given A E A the algorithm determines after at most 1

steps an efficient point, and hence will determine for each
pEA an efficient point;
~ for given A E A the algorithm does not determine an effi-

cient point. This means that the set of p-efficient pOints
is empty.
In the case of a finite set T, this algorithm and a computer programm

for it is in K5LLNER [7].
C) Continuity properties of EA and P A

To obtain continuity properties of the mappings E:A -->R, which
assigns to each A E A the optimal value E A, and P:A __>POW(Rn +1 ), which
assigns to each A E A the optimal set P A, we refer either to the inves-
tigations of BROSOWSKI [2] or to the results of section IV of this pa-
per.
References
[1] V.J. BOWMAN: On the relationship of the Tchebycheff norm and the ef-
ficient frontier of multi-criteria objectives
in: THIRIEZ and ZIONTS, Multiple Criteria Decision Ma-
king, Springer-Verlag, Berlin, 1976
[2] B. BROSOWSKI:Parametric semi-infinite optimization
Peter Lang-Verlag, Frankfurt/Main, 1982
[3] B. BROSOWSKI:A criterion for efficiency
Preprint, University of Frankfurt/Main, 1986
[4] B. BROSOWSKI and A. CONCI: On vector optimization and parametric pro-
gramming, Segundas Jornados Latino Americanas de Mate-
matica Aplicada 2, 1983, 483-495
[5} R. COLGEN and K. SCHNATZ: Continuity properties in semi-infinite
parametric linear optimization, Numerical Functional
Analysis and Optimization 3, 1981, 451-460
[6} J. JAHN: Parametric approximation problems arising in vector
optimization, to appear in IOTA
[7] C. K6LLNER: Ein interaktives Verfahren zur Losung finiter, li-
nearer Vektoroptimierungsprobleme, Diplomthesis, Uni-
versity of Frankfurt/Main, 1987
DUALITY IN PARTIALLY ORDERED SETS
Johannes Jahn
Institute of Applied Mathematics
University of Erlangen-Nurnberg
Martensstrasse 3, 8520 Erlangen, F. R. Germany
Abstract. In this paper a duality principle derived from abstract opti-

mization is developed in a general setting. Strong duality theorems
are presented and a specialization to topologic·al linear spaces is
given.
1. Introduction and Definitions
The literature on duality in vector optimization and multi-objective

programming, respectively, is very much grown during the last 5 years. A
survey on this literature can be found in the overview paper of Naka-
yama [14] but also in the book of Jahn [9, p. 205-206]. For more recent
duality investigations we refer to [1], [12], [13], [4], [6], [7],
[ 1 5], [1 9], [11], [1 6], [2] and [5]. Most of the published duality
results are obtained by extending the Lagrange theory, the Fenchel-
Rockafellar theory or similar theories known from scalar optimization.
If one checks these duality results, one can see that there exists also
a duality principle which is implicitly used and which is independent
from the mentioned theories (we quote [18] for an early paper with
this underlying idea). This duality principle is described in [9, sec-
tion 8.1] for partially ordered linear spaces. It is the aim of this
paper to show that this duality principle can be formulated in a very
simple way for partially ordered sets. The first duality investigations
in such an algebraic setting were carried out by Rosinger [17]. More-
over, we show that in partially ordered topological linear spaces the
so-called dual set can be transformed with the aid of certain conti-
nuous linear functionals and the connection with known duality results
in multi-objective linear programming can be demonstrated.
161
In the following let (X, S) be a partially ordered set where X is

a given nonempty set and S is a partial order on X assumed to be
reflexive and transitive. We are concerned with the problem of charac-
terizing minimal elements of an arbitrary non empty subset P of X.
Recall ~hat an element pEP is called a minimal element of P, if
pEP, p S P ==> p S p.
Next, we associate another set D c X to the set P by defining
D : = {d E X I pEP => p!l; d or d S pl. (1 )
The set P is also called the primal set and the set D is said to be
the dual set of P. In the next section we characterize minimal ele-
ments of the primal set P by maximal elements of the dual set D.
Recall that an element d E D is called a maximal element of D, if
-
d ED, d::; d => d S d.
Notice that in the case of an antisymmetric partial order S the dual

set of P can also be written as
D : = {d E X I pEP => P :f d or d p} .
If the dual set D is nonempty, then we have for each pEP and d E D
the relation p :f d or d ::; p. In optimization theory this result is
often called a weak duality theorem. In the context of this paper it is
an immediate consequence of the definition of the dual set D.
2. Duality Relations
In this section we consider an arbitrary nonempty primal set and

investigate how to characterize its minimal elements by the maximal
elements of its dual set. The next lemma is useful for the proof of
strong duality theorems.
162
Lemma 1:
Let (X,S) be a partially ordered set, let P be a nonempty subset of X,
and let D be defined as in (1). If x E P n D, then x is a minimal ele-
ment of the set P and also a maximal element of the set D.
Proof:
Since xED, it follows that for each pEP with P ~ x we obtain x ~ p.
This implies that x E P is a minimal element of P. Next, take an arbi-
trary element d E D with x S d. x belongs to P and therefore we get
d S x from the def ini tion of D. Consequently, xED is a maximal ele-
ment of D. c
The next result is a strong duaZity theorem for minimal elements

of the primal set.
Theorem 2:
Let (X,~) be a partially ordered set, let P be a nonempty subset of X,
and let D be defined as in (1). Then each minimal element of the set
P is also a maximal element of the set D.
Proof:
Let pEP be any minimal element of the set P, i.e. for each pEP
with P ~ P we have p ~ p. So, p belongs to the set D, and by Lemma
p is also a maximal element of the set D. c
In order to get a converse strong duaZity theorem we need an addi-

tional assumption which ensures that each maximal element of the dual
set is also a minimal element of the primal set.
Theorem 3:
Let (X,~) be a partially ordered set, let P be a nonempty subset of X,
-
and let D be defined as in (1 ) • For each maximal element d E D of the
set D let the following implication hold:
d f P => 3 d E D with d s d and d :$ d. (2)
Then each maximal element of the set D is also a minimal element of

the set P.
163
Proof:
Let d E D be an arbitrary maximal element of the set D. Then we have
for each d E D with d S d also d S d. Consequently, by the implication
(2) d belongs to the set P, and Lemma 1 leads to the assertion. D
The implication (2) is a trivial rewriting of the statement that

every maximal element of D belongs to P. In general, a converse strong
duality theorem cannot be proved without an additional assumption. A
simple sufficient condition for the implication (2) can be derived in
the case of a partially ordered real linear space (X,S) (where S is a
linear partial order). Recall that a nonempty subset S of a real
linear space X is called aZgebraiaaZZy open. if S equals its core, i.e.
S {xESI hEX => 3X>Owithx+)'hES V).E[O,X]};
and S is said to be aZgebraiaaZZy aZosed. if S equals its algebraic

closure, i.e.
S S U {x E xl 3 s E S\{x} with AS + (1-).)x E S V A E(O,1]}.
Lemma 4:
Let (X,S) be a partially ordered real linear space, let P be a non-
empty subset of X. Moreover, let there exist some hEX with Ox S h
and h ~ OX' and let an arbitrary d E D be given.
(a) If D\P is algebraically open, then the implication (2) is satis-

fied.
(b) I f the set
P+:=Pu{xExl 3pEP with psxand x~p} (3)
is convex and algebraically closed, then the implication (2) is

satisfied.
(c) If the space is equipped with a topology and the set P+ given by
(3) is closed, then the implication (2) is satisfied.
Proof:
Let d E D be arbitrarily given and assume that d $ P.
(a) Since D\P is algebraically open, for hEX with Ox Shand h ~ Ox

there exists some positive real number A with d := d + )'h E D.
164
Obviously we get d $ d and d ~ d. Therefore the implication (2) is

satisfied.
(b) Since the set P+ is convex and algebraically closed, the set X\P+
is algebraically open (e.g., compare [9, p. 13] for a similar
result). Notice that
X\P+ = X\P n {x E xl pEP ==> P ~ x or x $ p} = D\P,
and the assertion follows from part (a) of this lemma.
(c) Because the set P+ is assumed to be closed, the set X\P+ = D\P is
open, and the assertion is a consequence of part (a) of this lemma.
o
Remark 5:
If, in addition to the assumptions of Lemma 4, the partial order is
also antisymmetric, then the ordering cone ex := {x EX I Ox $ x} in-
duced by the partial order $ is pointed, and therefore in this case
the set P+ given by (3) equals the algebraic sum of the sets P and ex'
i.e. P+ = P + ex' In order to see this equality notice that
P+ P U {x EX! 3 pEP with P $ x and x +p}

{XEX! 3pEP with P$x}
P + ex'
The assumption in Lemma 4, (c) that P+ = P + ex is closed is a stand-
ard assumption used for the proof of a converse strong duality theo-
rem in vector optimization (e.g., see [8, Thm. 2.5 and Lemma 2.4]).
3. Specialization to Topological Linear Spaces
In this section we investigate the dual set (1) in a partially

ordered topological real linear space. Under appropriate assumptions
we rewrite this dual set with the aid of certain continuous linear
functionals. Moreover, we apply this result to a multi-objective
linear programming problem.
165
3.1 Formulation of the Dual Set
In the following (X,~) is assumed to be a partially ordered topo-

logical real linear space with an ordering cone Cx := {x Exl Ox $ x}
where $ denotes the partial order on x. Obviously, Cx is a convex cone.
Recall that the dual opdeping eone Cx * of the topological dual space X*
is given by
Cx* := {l E x*1 l(x) ~ ° V x E Cx},

and the set
c~* := {l E x*1 l(x) > °

denotes the quasi-interiop of the dual ordering cone Cx *' For our in-
vestigations we assume that C;* t
~. By a theorem of Krein-Rutman [10]
(compare also [9, p. 86]) the quasi-interior C~* is nonempty, if X is
a separable normed space and the ordering cone Cx is closed and
pOinted. Because of the assumption that C~* is non empty it follows
immediately that the ordering cone Cx is pointed or, in other words,
the partial order is antisymmetric.
For completeness we summarize our assumptions concerning the space:
Let (x,$) be a partially ordered topological real

linear space with an ordering cone Cx for which
C;* 9= ~.
In this setting we show under appropriate assumptions that the dual
set D of a suitable primal set P equals
15 := {d E xl 3 * * with ltd) $l(p)

l E CX V pEP
V P E P\{d}}. (5)
Lemma 6:
Let the assumption (4) be satisfied, and let P be any nonempty subset
of x. Then 15 c D where 15 is given by (5) and D denotes the dual set of
P.
166
Proof:
In the case of 0 = 0 the assertion is obvious. If 0 is nonempty, take
any d E D. Then we consider two cases:
(a) Assume that there exists some l E C*'X* with
led) s l(p) for all pEP.
Then we get for all p E P that p ~ d or d p. This means d E D.
(b) If there exists some l E Cx*\ {Ox*} with
l (d) < l(p) for all p E P\{d} ,
it follows for all pEP that p ~ d or d p. So we obtain d E D
as well. c
For the proof of the converse inclusion D c 0 we need additional

assumptions on the space, the primal set and the ordering cone.
Lemma 7:
Let the assumption (4) be satisfied, and, in addition, let X be
locally convex. Let P be a nonempty subset of X for which the set
P+C x is closed and convex. Assume that there exists a continuous
linear functional l *' E Cx*' * with - co < inf l *' (p).
pEP
Moreover, let the following condition be satisfied:
For each minimal element p of the set P there exists
either a continuous linear functional l E C~* with
l(p) ~ l(p) for all pEP (6)
or a continuous linear functional l E Cx*\{Ox*} with
l(p) < l(p) for all p E P\{p}.
Then D c 0 where 0 is given by (5) and D denotes the dual set of P.
Proof:
Since the assertion is evident for D = 0, in the following we assume
+
D 0 and we pick an arbitrary d E D. Then we have for all pEP
P ~ d or d = p which implies that d *
P + (Cx\{Ox}). Next we consider
two cases:
(a) If we assume that d E P + Cx' we get d E P. Since d E P n D,
by Lemma 1 d is a minimal element of the set P. Then by the
condition (6) d belongs to D.
167
(b) Next, we assume that d $ P + Cx. By a known separation theorem

(e.g., see [9, Thm. 3.18]) there exist a continuous linear func-
tional l E Cx*\{Ox*} and a real number a with
led) < a S l(p) for all pEP.
For an arbitrary positive real number A we define
II II
lA := l + Ai ECX*' and we set e := a - led) > 0. Then we get
for a sufficiently small A > °
e
lA (d) = led) II
+ Ai (d) <
II
a + Ai (d) S a - 2
and
II II
l{p) + Ai (p) ~ a + Ai (p) ~ a - ~
for all pEP.
II
So we obtain for such a functional lA E Cx *
lA (d) S lA{p) for all pEP

which implies d ED. c
Remark 8:
The assumption in Lemma 7 that the infimum of a continuous linear
functional ill E C~* on P is finite is fulfilled for instance, if the
primal set has a lower bound, say E x. In this case P c x
+ Cx {x}
and
inf l
II
(p) ~ l
II
(x) > -
A
=.
pEP
Remark 9:
The condition (6) holds for instance, if the set P is strictly convex.
A set P which is assumed to have a nonempty interior is called
striatly aonvex, if for every P1' P2 E P with P1 P2 +
AP1 + (1-A)P2 E int{P) for all A E(O,1).
In fact, if p is any minimal element of the set P, i.e.
{{p} - Cx) n P = {p}, by Eidelheit's separation theorem (e.g., see
[9, Thm. 3.16]) there exists some l E Cx*\{Ox*} with
l{p) < l{p) for all p E int{P).
168
Then we conclude for an arbitrary pEP with p +p

l(p) < leAp + (1-A)p) for all A E (0,1)
and l(p) < l(p), respectively. So the condition (6) is satisfied.
Remark 10:
If the set P is a convex polytope in ~m equipped with the compo-
nentWise partial order, then by a known result of Gale [3, p. 308]
for each minimal element p of P there exists a vector l E R n with
positive components such that
lTp ~ lTp for all pEP.
Consequently, the condition (6) is fulfilled r In this case the set

D can be written as
D = {d E ~I 3 l1, ... ,lm > 0 with lTd ~ lTp V pEP}.
Finally we summarize our results for a locally convex space.
Corollary 11:
Let the assumptions of Lemma 7 be satisfied. Then an element of the
primal set P is minimal if and only if it is a maximal element of the
dual set
o = {d E xl 3 l E C~* with led) ~ l(p) V pEP
or 3 l E Cx*\{Ox*} with led) < l(p) V p E P\{d}}.
Proof:
With Lemma 6 and Lemma 7 we obtain that the dual set 0 can be written
as above. Then we get the assertion from Theorem 2 and Theorem 3 to-
gether with Lemma 4, (c) and Remark 5. 0
3.2 An Application to Multi-Objective Linear Programming
In this subsection we investigate a special type of the primal

set P which arises in multi-objective linear programming. We have the
following assumption:
169
Let e be a real m x n matrix;

let A be a real p x n matrix;
let b E R P be a given vector;
let the space ~m be partially ordered
(7)
in the componentwise sense;
and let the set {x ~ o~ni Ax ~ b}
be nonempty (~ has to be understood
in a componentwise sense).
Under this assumption we consider the multi-objective linear program-

ming problem
min ex
subject to
Ax ~ b
x ~ 0 ~n
which has to be understood in the folloWing way: Determine minimal

elements of the image set {exi Ax ~ b, x ~ 0 ~n}. Therefore we choose
the primal set
P := {exi Ax ~ b, x ~ 0 ~n}. (8 )
The next lemma gives a formulation of the dual set of P.
Lemma 12:
Let the assumption (7) be satisfied, and assume that the set P given
by (8) is bounded. Then the dual set of P reads
D = {d E ~ml 3 l1, ... ,lm > 0 and u 1 , ..• ,u p ~ 0

with ATU ~ eTl and lTd ~ uTb}. (9 )
Proof:
Since the primal set P is a convex polytope, by the Lemmas 6 and 7
(in connection with Remark 10) the dual set of P can be written as
D = {d E~ml 3 l1, ... ,lm > 0 with lTd ~ lTp Vp E pl.

Using standard optimality conditions from linear programming we obtain
immediately
D = {d E ~ml 3 l1, ... ,lm > 0 and u 1 , .•• ,u p ~ 0 with
lTd ~ lTex + u T (b-Ax) V x ~ 0 ~n}.

170
With simple transformations this set can be rewritten as

o = {d Em m I 3 .t 1 , ••• ,.tm > ° and u 1 , ••• , up ~ ° with
ATu ~ cT.t and .tTd ~ uTb}. D
By Corollary 11 an element of the primal set P given by (8) is

minimal if and only if it is a maximal element of the dual set 0
given by (9). In the case of b +
0IRP it is shown in [8] that a vector
in IR m is a maximal element of 0 if and only if it is a maximal element
of the set
{d EIRml 3 .t 1 , ••• ,.tm > ° and a real m x p matrix
U with d = U(b), ATUT.t ~ cTt
and uT.t ~ amP} c: D.
Consequently, for b +
o~ the dual multi-objective linear programming
problem is equivalent to the following problem formalized by
max Ub
subject to
ATUT.t ~ cTt
uT.t ~ °mP
t 1 ,···,.tm > °
U EIRmxp.
It is obvious that for m = 1 this problem reduces to the well-known dual

linear programming problem.
4. Conclusion
This paper shows that a duality principle which is implicitly

used by many authors publishing duality results in vector optimization
can be formulated in a very general setting. Under the implication (2)
satisfied by a standard assumption in a topological linear space we
present already strong duality theorems. Therefore this duality prin-
ciple seems to be a key for many duality investigations. But for the
derivation of duality results under weak assumptions in vector opti-
mization, it is better to develop a direct approach than to use ex-
plicitly this duality principle.
171
References
[1] H.W. Corley, "Duality Theory for the Matrix Linear Programming
Problem", J. Math. Ana~. App'L 104 (1984) 47-52.
[2] K.-H. Elster and A. Gopfert, "Recent Results on Duality in Vector
Optimization", in this volume.
[3] D. Gale, The Theory of Linear Economic Mode~s (McGraw-Hill,
New York, 1960).
[4] Ch. Gerstewitz and E. Iwanow, "Dualitat fur nichtkonvexe Vektor-
optimierungsprobleme", Wiss. Z. TH Itmenau 31 (1985) 61-81.
[5] J. Gwinner, "A Result of Farkas Type and Duality in Vector Opti-
mization", in this volume.
[6] E.H. Ivanov and R. Nehse, "Some Results on Dual Vector Optimiza-
tion Problems", Optimization 16 (1985) 505-517.
[7] E. Iwanow, "Verallgemeinerte Konvexitat und die Dualitat in der
Vektoroptimierung", Wiss. Z. TH I~menau 31 (1985) 39-44.
[8] J. Jahn, "Duality in Vector Optimization", Math. Programming 25
(1983) 343-353.
[9] J. Jahn, Mathematicat Vector Optimization in Partiatty Ordered
Linear Spaces (Peter Lang, Frankfurt, 1986).
[10] M.G. Krein and M.A. Rutman, "Linear Operators Leaving Invariant
a Cone in a Banach Space" (Uspehi Mathematiceskih Nauk (N.S.) 3,
no. 1 (23), 3-95 (1948)), AMS T:t'anstation Series 1. Volume 10,
Providence (1962), p. 199-325.
[11] H.C. Lai and C.P. Ho, "Duality Theorem of Nondifferentiable
Convex Multiobjective Programming", J. Optim. Theory App~. 50
(1986) 407-420.
[12] C. Malivert, "Dualite en Programmation lineaire Multicritere",
Math. Operationsforsch. Statist. Sere Optim. 15 (1984) 555-572.
[13] H. Nakayama, "Geometric Consideration of Duality in Vector Opti-
mization", J. Optim. Theory App~. 44 (1984) 625-655.
[14] H. Nakayama, "Duality Theory in Vector Optimization: An Over-
view", in: Y.Y. Haimes and V. Chankong (eds~, Decision Making
with Muttipte Objectives (Springer, Berlin, 1985).
[15] H. Nakayama, "Lagrange Duality and its Geometric Interpretation",
in: P. Serafini (ed.), Mathematics of Mu~ti Objective Optimi-
zation (Springer, Wien, 1985), p. 105-127.
[16] V. Postolica, "Vectorial Optimization Programs with Multifuncti-
ons and Duality", Ann. Sci. Math. Quebec 10 (1986) 85-102.
172
[17] E.E. Rosinger, "Duality and Alternative in Multiobjective Opti-

mization", Proc. Amer. Math. Soc. 64 (1977) 307-312.
[18] P. Schonfeld, "Some Duality Theorems for the Non-Linear Vector

Maximum Problem", Unternehmensforschung 14 (1970) 51-63.
[19] T. Tanino, "Conjugate Maps and Conjugate Duality", in: P. Sera.fini

(ed.), Mathematics of Multi Objective Optimization (Springer,
Wien, 1985), p. 129-155.
GENERATING NESTED SUBSETS OF EFFICIENT SOLUTIONS
Ignacy Kaliszewski
Systems Research Institute
01-447 warszawa, nl. Newelska 6
Abstract
Given a vector optimization problem, the solution set is composed

of nondominated elements of a set. All nondominated elements are
equivalent in the sense that each of them is a solution as good as any
other one. However, if the vector optimization problem is a formaliza-
tion of a practical multiple objective decision problem, one may expect
that some solutions are more preferred than others. To assist a deci-
sion maker, most of methods for multiple objective decision making rely
upon additional information, to be obtained from the decision maker, to
that contained in the definition of the vector optimization problem. In
contrast to that, in this paper we show how it is possible to impose
a certain order over nondominated elements of a set ,based only on the
information contained in the formulation of the corresponding vector
optimization problem. We also present a numerically attractive method
to deternine nested subsets of properly efficient elements.
1. Introduction
We address the vector optimization problem
(VP) Vmax (f(x))

X E. S ,
k
where f is a vector function, f:S~R , and Vmax denotes the operation of
deriving all efficient elements of S. The solution of VP is then any
element x, XE.S, such that f(x) is efficient in the set Z={yE.Rkly=f(x)
x e. s} (i. e. f (x) is nondominated in Z with respect to the preference
174
cone ~ • Throughout this paper, for the sake of notational

simplicity, we shall address only the set Z and therefore it will be
not confusing when we call efficient elements of Z the solutions of VP.
The only assumptions made about the problem VP are the existence
of a point y x e. Rk such that Z ~ {yx }- ~nt(
. k and closedness of Z.
R+)
The reason of interest in determining efficient elements of VP is
that almost all concepts of solutions in multiple objective optimiza-
tion problems (of which VP is a mathematical formulation ) assume
solutions to be efficient. Therefore, it is important to have compu-
tationally simple methods to determine all efficient elements of a
set, or at least to determine a certain representative subset of the
efficient set. Such methods are provided by modifications of the
weighted Tchebycheff norms, but the price of their simplicity is that
they provide slightly less than the set of all efficient solutions,
namely the set of properly efficient solutions.
2. Generating properly efficient solutions
The following definition is due to Geoffrion [4].

Definition 2.1. YEo Z is properly efficient if y is efficient and
there exists a finite number M> 0 such that, for each i, we have
Yi-Yi
" M
Yj-Yj
for some j such that Yj <Y j , whenever YEo Z and y i >y i .

This definition and in particular the number M are referred to
in Theorem 2.1 and Theorem 2.2. Those theorems are implicit in the
general theory of vector optimization developed by ~ierzbicki [15,16]
and recently by Jahn [6,7J and also in the results of Rolewicz II3J,
Borwein [1], Gearhart [4], Nieuwenhuis [12], and Henig [5], and may be
derived from these theories as shown e.g. in [9]. However, both
theorems were originally formulated and proved for the particular form
of the vector optimization problem addressed in this paper. The formu-
lation of Theorem 2.1 is different but equaivalent to that given by
Choo and Atkins [2]. The equivalence was shown in [IOj.
Let e denote the vector in Rk whose entries are each unity.
Let 1\ = fA 1Ai>O, i=I, •.. ,k, t
~=I
Ai=I}.
Theorem 2.1. ye: Z is properly efficient if and only if there exist

he: Aand p>O such that y solves the following program
175
For each 'I E. Z, Y

properly efficient, there exist A €. A such
that 'I uniquely solves pA for A=~ and p satisfying
M < (l+kp)(kp)-I.
For each y€. Z solving pA, M < (l+(k-l)P)p-l.
Theorem 2.2 [9J. YE.Z is properly efficient if and only if there

exist A£ A and p>O such that y solves the following program
For each YE. Z, y properly efficient, there exists IE." such

that y uniquely solves PA for A=I and p satisfying
M < ~in I.(kp)-I.
i l -1
For each 'IE. Z solving P A, M < (p+max Ai )(p(k-l)) .
i
The program pA was studied by Choo and .Atkins in [2], the program
by Steuer and Choo in [14J and by Kaliszewski in [8]. The main
conclusions are as follows.
- In the case of Z either finite or polyhedral, where all
efficient ele~ents are also properly efficient, there exist
formulae for calculating values of the parameters p such that
the program pA as well as P A when solved for all values of
A~ A yield all efficient solutions. Only the formula to derive
p for pA and Z finite is operational.
- In the general case, no matter how small p is, by solving the
program pA or PA for all A~ A and fixed p only a subset of
properly efficient elements can be determined.
- pA has certain computational advantages over PA'
The main feature which distinguishes the programs pA and P A from
other methods of scalarizing vector optimization problems is that these
programs do not introduce any additional nonlinearities to that of the
original problem. This is due to the well known transformation:
min max Yi is equivalent to min t

y~Z i tIi.R
t~Yi' i=I, ... ,k,
Y E. Z .
This feature can hardly be overestimated when dealing with linear
or integer vector optimization problems. This and the fact that only
176
~
the assumption about the existence of y is needed,makes both
programs very attractive. Moreover, these programs permit one to get some
extra information about properly efficient solutions determined by them,
which might be useful to a decision maker. Such information is provided
by the "if" parts of Theorem 2.1 and Theorem 2.2: for each y E. Z which
solves either pA or PA not only do we know that there exists a finite
number M (because is properly efficient) but also a certain upper
bound on this number is established.
Below, whenever y~ Z is properly efficient, we shall assume that
M is the smallest possible number for which the inequality of Defini-
tion 2.1 is fulfilled. In other words we shall refer to a slightly
modified definition of M.
Definition 2.2. Y E. Z is properly efficient if y is efficient and

there exists a finite number M>O defined as
Yi-Yi
M max max min
y i;Yi-Yi>O j;Yj-Yj>O Yj-Yj
where Y€. Z.
Obviously, with respect to proper efficiency both definitions are
equivalent. The difference is Definition 2.2 addresses one particular
number whereas Definition 2.1 addresses rather an interval of numbers.
One confines oneself to properly efficient solutions not only
because this set of solutions is technically more easy to deal with
but also because improperly efficient solutions in the context of
multiple objective decision making are regarded as anomalous. Indeed,
interpreting coordinates Yi of elements Y£ Z as values of objective
functions fi(x}, i=1, ••• ,k over certain set S, for an improperly
efficient solution Y €. Z
there exists an objective function fi for which the marginal
gain can be made arbitrarily large to ~ach of the marginal losses
in other criteria (compare [4]).
It is quite natural then to confine ones attention to properly
efficient solutions for which
for each objective function fi the gain relative to at least one
of the losses in other criteria is' not larger than a certain
prescribed value.
In other words one may be interested in properly efficient solutions
for which M is not greater than a given number. The motivation of
this is that one may want to avoid solutions for which the gain for
each objective function fi relative to at least one of the losses in
177
other criteria is excessively high. Too high a gain to loss ratio of a

solution in the presence of other solutions with moderate values of
this ratio may suggest that the considered vector optimization problem
does not model correctly the underlying decision makinq problem.
It is easily seen from the definition of M that its values
provide a global information about (properly) efficient solutions. This
contrasts with the weighting method (which is of limited use for
non-convex problems) in which weighting vectors A provide local
information about trade-offs of solutions.
Theorem 2.1 and Theorem 2.2 suggest two algorithms for generating
nested subsets of properly efficient solutions with values of n1
playing the role of a controlling parameter. Each subset contains only
those solutions for which the value of M is not greater than
a certain value. The algorithms are as follows.
The algorithm pA(P A)

Step 0 Set i=l
Step 1 Select a value for M, denoted by Mi , and determine Pi
from the relation H.«1+(k-1)p.)p7 1 (set a=max A1' ,
1 1 1
O<a<l, and determine Pi from the relation
-1
Hi < ( P.t+a) ( Pi( k-1 ) ) ).
Step 2 Determine a subset Z.1 0=
properly efficient solutions
by solving pA with Pi for all A E.I\ (by solving PA
with Pi for all AE. A, max A·=a, O<a<l).
i 1
Step 3 Set i=i+1. Go to Step 1 to continue, otherwise STOP.
It is obvious that if the sequence of values of Mi is such that
Mi+l>M i , i=l,2, .•. then Pi+1<Pi and we get nested subsets of properly
efficient solutions i.e. Zi~Zi+1·
If the set Z of all efficient solutions is not finite, then in
Step 2 we can determine a certain representation of this set e.g. using
the filtering method of Steuer and Choo [14].
The above algorithms can be used in the decision making situations
in the following way. A decision maker is provided with a certain
subset of (properly) efficient solutions and makes a temporal decision.
Next, he is provided with a larger subset. If his previous decision
remains unchanged, than it is perhaps reasonable to stop generating
other efficient solutions. If not, the decision maker is provided with
a subsequent subset and so on. This process is relatively fastly
convergent if solutions with small values of M are likely to be
accepted as final solutions.
Let us observe that for the convex case the solutions with low
values of M are likely to be located in the "central" part of the
178
efficient set, but the notion of the central part is difficult to make
precise (compare Fig. 1 a,b,c). If Z is non-convex the situation is
even less clear as is shown by the example.
a) b) c)
Fig. 1. Values of M: a - high, b - moderate, c - small.
Example 2.1. Let Z be the figure composed of elements of the line

segments {(O,O),(O,6)}, {(O,6),(3,S)}, {(3,S),(4,4)}, {(4,4),(S,2)},
{(S,2),(6,1)}, {(6,1),(8,O)}, {(S,O),(O,O)} and all elements enclosed
by these segments (Fig. 2). Values of M for properly efficient
solutions of Z are as follows
3 for O~Yl<3,
1 for 3~Yl~4,
2 for 4<Yl~S,
2 for
1 for
4 8
Fig. 2.
179
On the ot~er hand, however, the solution resulting form the maxi-
mization of 2::y. over Z, which is often assumed to be a good
i=1 ~
starting point for interactive decision making processes, will be
always included in the first subset generated by one of the algorithms
proposed above. This fact is implicit in Lemma 2.2. To prove this lemma
we will need the following fact.
Lemma 2.1. Given a nonempty set Z, there is at most one properly
efficient element with M<1.
Proof. Suppose that there exist two properly efficient elements of Z,

y and y, such that My<1 and My<1. _BY the definition of M
y i -y i Yi -Y i
1>My max max min ~ for each
Y£Z i;Yi-Yi>O j;Yj-Yj>O Yj-Yj Yj-Yj
i=1, .•. ,k and some j such that Yi-Yi>O, Yj-Yj>O
But since
...
-Yj-Yj>
for each i=1, ... ,k and some j such that Yi-Yi> 0 ,
Yi-Yi
Yj-Yj>O
then IV
Yi-Yi
Lemma 2.2. Among all properly efficient solutions of a set Z the

smallest value of M is attained by the element which solves the program
k
max L Yi
i=1
Y£ Z.
Proof. The element of Z,
k
say Y,
which solves the above_ program
lies on the hyperplane t;1Y i =a and all elements Y € z, _Y'fY lie on
or below this hyperplane. Therefore, for each y€ Z, yfy
Yi-Yi
, 1
whenever
Yj-Yj
and therefore My~1.
If My< 1 then by Lemma 2.1 there is no other Y (. Z, Y efficient,
such that My <1. Thus, Y has the smallest value of M.
If M-=1 then observe that for each Y £ Z, yfy
Y
180
Yi-Yi
and this entails that for each Y € Z, yiy
Yi-Yi
min ~ 1
j;y·-Y·>O Yj-Yj
J J
which completes the proof.
For the case k=2 one can also make use of values of M to get
the approximate shape of the efficient frontier of the set Z. Suppose
that for a certain value of M an efficient element yEo Z has been
determined by the algorithm pA or PA' Then, by Definition 2.1,
Y2-Y2
or - _ - - ~ ~1 for all Y E'. z, ytY. This can be
Y1 -Y 1
represented graphically as in Fig. 3 - all elements Y E. Z must lie on
or below the line (YI-Yj);:; H(Y2- Y t
or (Y2- Y2) ~ H(Y 1 -Yl) which
gives a rough approximation of Z (the shaded region in Fig. 3).
Repeating this procedure for another efficient element, sa:' y, and
taking the common part of approximations resulting form y and y we
get a nore accurate approximation.
Fig. 3.
Example 2.2. Consider the set Z given in Example 2.1. Suppose that
the following efficient elements of Z have been determined
(3,5) for M=I,
( 5~, Ii) for M= .~ ,

3
(2,5!.) for
3
H=3 .
Drawing the corresponding lines we get a sequence of approximations of
Z as shown in Fig. 4.
Fig. 4.
3. Conclusions
In this paper we have shown how to generate nested subsets of

efficient elements, each subset containing only elements characterized
by a given number. The method proposed does not require any additional
information to that contained in the standard formulation of vector
optimization problems. The method seems to be of theoretical interest
and also of value for many multiple objective decision making algorithms.
It would be even more attractive if it generated only those elements
for which an increase in one coordinate was always compensated by
limited decreases in all other coordinates or, in other words, only
those solutions for which for each objective function fi the gain
relative to all of the losses in other criteria was not larger than
a certain prescribed value. In principle this would mean that only a
subset of properly efficient elements had to be generated. This problem
will be addressed in Author's subsequent paper [IIJ.
182
References
[1] J.Borwein: Proper efficient points for maximizations with respect

to cones. SIAM J.Contr. Opt. 15 (1977) 57-63.
[2] E.U.Choo, D.R.Atkins: Proper efficiency in nonconvex programming.
Math. Oper. Res. 8 (1983) 467-470.
[3] W.B.Gearhart: Characterization of properly efficient solutions by
generalized scalarization methods. JOTA 41 (1983) 491-502.
[41 A.M.Geoffrion: Proper efficiency and the theory of vector maximiza-
tion. J.Math.Anal.Appl. 22 (1968) 618-630.
[5] M.I.Henig: Proper efficiency with respect to cones. JOTA 36 (1982)
387-407.
[61 J.Jahn: Scalarization in vector optimization.Math.Progr.29(1984)
203-218.
17} J.Jahn: A characterization of properly minimal elements of a set.
SIAM J.Contr. Opt. 23 (1985) 649-656.
[81 I.Kaliszewski: Characterization of properly efficient solutions by
an augmented Tchebycheff norm. Bull. Polish Ac. Sci., Ser. Technical
Sciences 33 (1985) 415-420.
[9} I.Kaliszewski: Norm scalarization and proper efficiency in vector
optimization. Foundations of Contr. Engin. 11 (1986) 117-131.
rOJ I.Kaliszewski: A modified weighted Tchebycheff metric for multiple
objective programming. To appear in Compt. Oper. Res.
[11] I.Kaliszewski: Substantial efficiency in vector optimization
problems. In preparation.
[12} J. Ttl. Nieuwenhuis : Properly efficient and efficient solutions for
vector maximization problems in Euclidean space. J.Hath.Anal.Appl.
84 (1981) 311-317.
[13] S.Rolewicz: On a norm scalarization in infinite dimensional Banach
spaces. Contr. a. Cyb. 4 (1975) 85-89.
[141 R.E.Steuer, E.U.Choo: An interactive weighted Tchebycheff procedure
for multiple objective programming. Hath. Progr. 26 (1983) 326-344.
[15J A.P.Wierzbicki: Basic properties of scalarizing functionals for
mUltiobjective optimization. Math. Operationsforsch. Statist. Ser.
Optimization 8 (1977) 55-60.
[16] A.P.~'lierzbicki: The use 0:: reference objectives in multiobjective
optimization. In: G.Fandel and T.Gal, (eds.): Hultiple Criteria
Decision Making Theory and Application, Springer Verlag (1980)
468-486.
HIERARCHICAL STRUCTURES
IN MULTICRITERIA DECISION MAKING
Ulrich Krause
Fachbereich Mathematik und Informatik
Universitat Bremen, 2800 Bremen, West Germany
1. Introduction
Confronted with decisions involving many criteria in real lif. people
sometimes apply a rankin~ of importance to the criteria, instead of
trying to balance the criteria on some common scale. The principle
"first things first" is at the heart of any hierarchical procedure
and one may ask for the reasons people are applying it. Beside em-
pirical investigation, an answer may also be given by looking analy-
tically for possible appealing features which hierarchical structures
could possess. In this respect, not only the individual decision
taker is of interest, but also his cultural environment, including
e.g. the dictionary or the decimal system both being based on lexi-
cographical ordering. It is the main result of the present paper
that hierarchical decision procedures may be characterized by some
specific stability property. Roughly speaking, the latter means that
a decision taken in favor of one alternative against the other, is
not upset by a small perturbation in the assessment of any of the
criteria. For the result, we have to assume the decision rule under
consideration yields decisions in sufficiently many cases of conflict
between the different criteria. Without this assumption the stability
property is not very demanding and it is shared by various different
rules.
A precise definition of the stability property together with some

discussion is given in section 2. In section 3 the notion of hier-
archical structures as used in this paper is introduced and a theo-
rem characterizing those structures by the stability property is
stated together with some consequences. The proof of the theorem is
given in section 4. Although the context is rather different, one
may be tempted to consider the famous Possibility Theorem proved
by the economist K.J.Arrow [1] in 1951 as the first characterization
of a hierarchical structure, namely one single dictator in this
case of social choice, in terms of quite appealing properties.
184
2. A stability property for quasi-orderings
Consider an arbitrary set of alternatives, say A and A*0, among

which decisions have to be taken on the basis of multiple criteria,
the set of which we denote by N = {1, ... ,n}. Suppose for simplicity
the evaluation of A according to criterion i is in real numbers,
which means there is a function fi:A .ffi, ~ the set of real numbers
and fi(a) being the value of alternative a with respect to criterion i.
For any given alternative a the decision taker thus obtains a vector
of data (f 1 (a), f 2 (a), ... ,f n (a». To come out finally with decisions
the decision taker, very probably, will try to aggregate the dif-
ferent data fi(a) for i=I, ... ,n in some way or the other into a single
real number. Thus he is looking for some aggregation rule or scala-
r i z a t ion F: Rn • IR, wher eRn ={x =( xI' ... , x n) I xi ER }. Equi Ppe d wit h s u c h
a rule F decisions within A can be taken by preferring a to b pre-
cisely when F(f 1 (a), ... ,fn(a» > F(f 1 (b), ... ,fn(b». But another
decision problem arises, namely how to choose from the big set of
those aggregation rules. Tradition, custom or something else may help
n
and suggest some rule like F(x)= .2:,r.x. with certain weights r 1.>0,
n 1 1=1 1 1
f(X)=(.2: X~)2, F(x)=min{x.ll$i$n} etc. (For a discussion of such
1 =1 1 1
format10n rules cf. [6], whffe these rules playa major role in what
is called the character of a decision taker. The above problem em~~es
in measurement theory, cf. [5], as well as in multicriteria decision
making, cf. [7]. Both have a lot in common.)
It should be clear however, and may easily be checked that the choice
of the rule F may matter by affecting the decisions taken within A.
Hence it is perhaps more appropriate instead of picking up some
particular rule to look for classes of them having convenient pro-
perties. A property which certainly would be desirable is the sta-
bility property in the sense that decisions induced within A by F
are not upset by small perturbations in any of the f i . More precisely,
call an aggregation rule F:lR n • ~ stable, if there exists some
number o<d$l such that F(x)$F(y) is equivalent to F(TX)$F(TY) for
all x,yEIR n and all small perturbations T defined by u=((1+£1)z l +01'
... , (l+£n)zn \On),I£il,loil<d for i=l, ... ,n. (Here
z=(zl, ... ,Zn)EIR, 1·1 denotes the absolute value for real numbers.
Usually d is a small number.) It is easy to see that no one of the
rules mentioned above is stable in this sense. But perhaps there are
others? this problem can be tackled on directly (cf.[4] for some
special cases.) Instead we prefer to generalize the problem in a
first step and then to solve it in the next section. This, on the
185
one hand simplifies the mathematics and gives on the other hand more
room to applications. The reason for the latter being that not in
every application real numbers are available. Obviously, for taking
decisions within A not the function F:R n + R is really needed, but
only the quasi-ordering defined by x~y iff F(x)~F(y). A quasi-
ordering~ on IRn is a binary relation~ on IRn which is reflexive
(x.{x for all x), transitive (x4y and y.{z imply x~z) and total
(x~y or Y~ x). By -< we denote strict ordering, i.e. x<y means x<4y
and not y~ x. By'" we denote indifference, i.e. x""'y means x~y and
y~x. It should be noted that~ is not assumed to be anti-symmetric
(x4 y and y4. x imply x=y). Thus more generally what we are looking
for is a stable quasi-ordering ~ on IR n , meaning there exists some
o<d~l such that x4y is equivalent to ,x~ ,y, for all x,y ElR n and all
small perturbations l. Before turning to the final version of the
stability property, it is useful to look at the following alternative
descriptions of its present version.
Lemma. For: any quas i-order; ng 4 on IRn the foll owi ng properti es are
equivalent.
(i) 4 is stable
(i i ) x~y <= > l X ~ l Y
for all X,YER n , all scale transformations l defined by lz=(a 1 z 1+b 1 ,
.•. , anzn+b n ) with ai,b i E IR arbitrary, but ai>o for i=l, ... ,no
( iii) If for x, y , u , vEIR n s i gn (y i-x i ) =s i gn ( v i - u i) hold s for all i,
the n x -< y i seq u i val en t to u -< V. ( The reb y for r EIR, s i gn r =-fr-r if r", 0
and sign 0=0.)
Proof.
(i) => (ii): By accumulation of small perturbations. Let ai)o, biEIR
i =1 , ... ,n. De fin e for kE {I ,2 ,3 , ...• }
bi t bi
= ~(ai - 1) if a i ",l and 0i=r- if a i =l.
1
By assumption 4 is stable for some o<d~l. Choose k such that for
Ei' 0i defi ned above one has IEi I, I 0i I<d. Then
TZ = (E1z1+01, ... ,Enzn+on) defines a small perturbation and by iter-
ating it k-times the scale transformation l is obtained,
Cb ...• 't)z=(a1z1+b1' ... ,anzn+bn)=lz. Therefore (i) implies
x 4 y <= > Tx 4 Ty <=> ('(. '( ) x ~ ( T• '( ) y <= > ... <=> LX -4 l Y .
(ii) => (iii): Let x,y,u,vER n and suppose sign(y.-x.)=sign(v.-u.)
1 1 1 1
for all i. Define
186
vi-u i
a.
1
=- --
Yi-Xi
if and a i =1 otherwise. Then a i >0 for all i.
Define bi=ui-aix i . It follows ui=aixi+bi and vi=aiYi+bi for

all i and U=TX, V=TY, where Tz=(alz1+b1, ..• ,anzn+bn)' Property
(ii) yields: x"y <=> u~v. Analogously: y..t,.x <=> v~u. Therefore:
x<..y <=> u-<,v.
(iii) => (i): Let X,YElR n and choose d=1. Let T be a small perturba-
tion, i.e. TZ=((1+£1)z1+01 .... '(1+£n)zn+On) with I£i 1.lo i 1<1. Then
for U=TX, V=TY one has vi-ui=(1+£i)(Yi-xi) and hence
sign(vi-ui)=sign(Yi-xi) for all i. By (iii) therefore x<"Y is
equivalent to u.<,.v. Analogously: Y'<"x <=> v.<..u. Taking together:
X",y <=> u4v <=> TX4TY.
•
Property (ii) in the lemma we call scale-independence of~, meaning
the ordering given by~ does not depend on how origin and unit of
measurement of any of the coordinates i=1 •...• n are choosen.
Property (iii) is nice in that it does not really depend on real
numbers, but only on component ordering,
1. Xi <y i
sign (Yi-xi) = { 0, xi=Y i This allows us to discard real numbers
-1, xi >Y i
also in the construction of the cartesian product ~n. and we arrive
at the following version of stability free of any real numbers.
Let for i=1, ... ,n (Xi'~i) be a non-empty set Xi equipped with some
quasi-ordering ~i' Furthermore, let X = .~ Xi be the Cartesian
1=1
product of the Xis equipped with some quasi-ordering ~ . Define
for elements x=(x 1 ..... xn ) etc. in X.
1. x.-«. y.
si(x,y) = { o. x~(V~ Y~ • i=1 .... ,n.
-1, Yi<i Xi
Now, we call ~ a stable quasi-ordering on X, if for any x,y,u,vEX
the equality si(x,y)=si(u,v) for all i implies that x-'..Y is equiva-
lent to u-'..v.
According to the lemma, for (Xi'~i)=(IR.~) the quasi-ordering ~ on

X is stable in the above sense precisely when it is stable as we
defined it earlier.
In the next section we will determine the structure of stable

quasi-orderings.
187
3. Hierarchical structures
n
Let (Xi'~i) for i=l, ... ,n be quasi-ordered sets and let X = IT X.
i =1 1
be the Cartesian product. A typical element of X we denote by
x=(x 1 , ... ,x n ). A quasi-ordering ~on X we call hierarchical, or of
hierarchical structure, if there exist
some number m € {l, ... ,n},
an injection a : {l, ... ,m} + {l, ... ,n}
and a mapping t, :{l, ... ,m} + {-<., }-}
such that for any X,YEX the relation x<..y holds i f and only i f
there exists some iE{1, .•. ,m} with xo(i)t,(i)o(i) Yo(i) and
x o(j )("\Jo( j) Yo( j ) fo r all jE {l, ... , i -1 }.
The hierarchical structure is described by the parameters m,o and t,

and we call (m,o,t,) the type of a hierarchical quasi-ordering~.
To illustrate the concept consider a hierarchical quasi-ordering 4

of type (n,o,'..('). (Thereby '..(' is a short name for the mapping
f:.: {1, .•• , n} + {-<, >},
t, ( i ) = .L... a 11 i. a i sap e r muta t ion 0 f {l , ..• , n} . )
This obviously gives the usual lexicographic (quasi-)ordering on the
product space X with respect to its component (quasi-) orderings.
One may imagine an arbitrary hierarchical quasi-ordering as obtained
from such a lexicographic (quasi-) ordering by cutting the set of
components whi ch matter from {1, .•• , n} down to some subset
{o(l), ... ,o(m)} and by admitting beside ..(i also order reversal >i.
A degenerated case of a hierarchical quasi-order ~ is that one for
m=1. The definition given then says that x..(y is equivalent to
xo(l)t,(l)o'(l)Y (1): meani~g that~ on X is equivalent tO~o(l) or to
the reversed quasl-orderlng ~(1) a
on the single factor X0(1)·
Now we have the following characterization of stable quasi-orderings

n
on X = IT X. as defined in the preceding section. A quasi-ordering
i =1 1
4 on a set we call trivial, i f x~y for all x and y in the set.
Theorem. Suppose each of the quasi-ordered sets (Xi'~i),i=l, ... ,n,

[ possesses at least 3 indifference classes. Then a quasi-ordering on
n
X = IT X. is stable i f and only i f i t is hierarchical or i t is trivial.
i =1 1
The proof of the theorem will be given in the next section. It is

quite obvious that any lexicographic quasi-ordering is stable and
is strictly increasing in the following sense. A quasi-ordering
188
n
on X = IT Xi is called strictly increasing, if xi~iYi for all
i=l
i implies x~y and X-4.y holds provided there is in addition some
j with xj"'(jYj' Converse1Y~ consider a quasi-ordering40n X which
is stable and strictly increasing. By the theorem~ must be hier-
archical and because it is strictly increasing we must have m=n
and 6(i)=.( for all i. Hence{ muH be lexicographical. Thus the
theorem yields immediately the following corollary.
Corollary 1. Suppose each of the quasi-ordered sets (Xi'~i)'

i=l, ...
. ,n, possessesn at least 3 indifference classes. Then a quasi-
ordering on X = IT X. is stable and strongly increasing if and
i =1 1
only if it is lexicographic.
By putting for all i (Xi'~i) = (R,S), the theorem provides an

answer also to the question of stable quasi-orderings on Rn dealt
with in the preceding section. For, the assumption made in the
theorem holds trivially in this particular case, because there are
as many indifference classes in (~,S) as there are real numbers.
Moreover, the theorem yields also an answer to the question of
stable aggregation rules F : Rn - R. By the theorem the quasi-ordering
on Rn defined by x, y iff F(x)SF(y) must be trivial or hierarchical.
The former means that F is constant on ~n. The latter forces 4 to
be of degenerated type, i.e. m=l, quite by the same familiar argument
used to demonstrate that the usual lexicographic ordering on IR n can-
not be represented by a real valued function. Thus we obtain also
the following corollary from the theorem.
Corollary 2.
(a) Any quasi-ordering on Rn, n~2, which is stable (with respect to
S on R) must be hierarchical. It is stable and strictly increasing
if and only if it is lexicographic.
(b) Any stable aggregation rule F:R n - R, n~2, must be trivial in
the sense that F is constant or is a function of exactly one coordi-
nate only.
Without the assumption made on the number of indifference c1asses,the

theorem and its corollary 1 may become false as is shown by the follow-
ing example.
Example: For i=1,2 take (Xi,:4 i )=({o,ll,S). On X=X1xX2 define a
quasi-ordering ~ by (0,0)-< (0,1)", (1,0)< (1,1). The only possibilities
of four points x,y,u,V€X satisfying the precondition si(x,y)=Si(U,v)
189
for all i in the definition of a stable quasi-ordering are given

by x=(o,o), y=(l,o), u=(o,l), v=(l,l) and the case with y and u
interchanged. Then x-<.y and u-<. v, i .e. ~ is stable. Obviously, -4 is
strictly increasing. But4 cannot be lexicographic because of
(o,l)tY(l,o). This is in agreement with corollary 1 because
( {O, I} ,::;) possesses only 2 i ndi fference cl asses.
Before turning to the proof of the theorem we like to add some re-
marks concerning the relationship of the corollaries to two dif-
ferent branche's in the literature.
Remark 1. For the particular situation depicted below Corollary 1

reduces to the main theorem proved by P.C. Fishburn in [2]. On the
n
cartesian product X IT X. given is some quasi-ordering ~ from
i =1 1
which orderings 4i on the factors Xi are derived as follows. For
x = (x 1 , ... ,X n )EX, iE{l, ... ,n} and aEX i let x(i,a)EX be obtained
by replacing xi in x by a. Assume that x(i ,a)<x(i ,b) i f and only
if y(i,a)«y(i,b), for arbitrary x,yEX, a,bEX p iE{1, ... ,n}
(Axiom 2 in [2].) Defining a~ib for a,bEX i i f there exists some
xEX such that x(i ,a)&,x(i ,b), a quasi-ordering 4i on Xi is derived.
It turns out that with respect to these derived quasi-orderings the
given quasi-ordering is strictly increasing. With~i as deri~ed
above from ~ , corollary 1 becomes the theorem of Fishburn. His idea
of proof will be very useful for our proof in the next section.
Remark 2. Suppose ~ is not only a quasi-ordering but an ordering on

Rn which is stable. Then (IR n , 4) is a fully ordered vector space and
by the Hausner-Wendel or Hahn-Erdos theorem there exists a fully
ordered basis of ~n with respect to which ~ can be represented lexi-
cographically (cf.[3].) For this it is actually not necessary to
assume full stability, scale-independence for changes in the origin
and uniform changes in the unit of measurement would be sufficient.
The above theorems however, do not yield lexicography with respect
to the usual orderings on the given factors or lexicography with
respect to the canonical basis in Rn. In case of a stable quasi-
ordering 4 on IR n one could proceed by considering the quotient
space IR nIN
190
4. Proof of the theorem
It is easy to see, that a hierarchical or trivial quasi-ordering on

X is stable. For the reverse implication let ~ be a stable quasi-
ordering on X which is not trivial. By the following four steps we
shall show (cf. also [2]) that" must be hierarchically in the sense
n
defined at the beginning of section 3. Remember that X= nx.,(x.,~.)
i=l' , ,
a quasi-ordered set and { 1, xi -}, such that for
all x,yEX we have:
xi - xL'.y.
Suppose the implication does not hold. Then there exist X,YEX
satisfying xi <iYi for all i but xrJy. Let x,yO be arbitrary.
Since by assumption (X i ,4 i ) has at least 3 indifference classes,
we can find u,v,zEX satisfying the following relations:
Zi ~i u i .( i v. , for xi .(i Yi
zi .L..i vi.L.. i u i for y i ..(i xi

, , ,
zi ~ i u. r.J. v. for xi tVi Yi
It follows that zi <i u i and zi -<i vi for all 1. Hence
si(z,U) = si(z,v) = si(x,y) for all 1. By assumption ~ is stable
and therefore XNy implies that ZNU and z-vv. Transitivity of IV
yields u..,v. By construction of u and v si(u,v) = si(x,y) for all i
and using stability again it follows that x""'y. Thus XNy for any
two x,y EX and ~ must be trivial. This contradicts the assumptions.
~~ Ch 0 0 s e L'.E { .t..., :>} a c cor din g to s t e pl. Fix t his L'. and calla
subset f/I"'I~ U, •.. ,n} decisive (cf.[l],[2]) if there exist
x,yEX such that the following is true:
si(x,y)=l for all iEI and Sj(x,y)=-l for all HI =>XL'.Y.
Due to stability of ~ , for I decisive the above implication holds
for any x,yEX. By step 1, {l, ... ,n} is decisive. Hence there exiSts
a minimal decisive subset I of {l, ... ,n}. We shall show that I is
a singleton. Suppose the contrary, that is there exists an hEI
with I\{h} '" f/I. Since (Xi'~i) has at least 3 indifference classes,
we can find x,y,zEX satisfying the following relations:
191
Zh -<h xh <.
h Yh for the h choosen
Xj ..( j Yj -< j Zj for all jEI\{h}
Yk -< k zk -< k xk fa)\' all kE{l, ... ,n}\I (which may be 0).
Concerning x,y we have Sj(x,y)=l for all JEI and sk(x,y)=-l for all
keI. Hence XAY because I is decisive. Concerning X,Z we have
Sj(x,z)=l for all jEI\{h} and Sj(x,Z)=-l for all kU\{h}. Since I
is a minimal decisive subset, I\{h} can not be decisive and we thus
can not have XAZ. With respect to y,z finally we have sh(z,y)=l and
Sj(z,y)=-l for all J*i. Since I is minimal, {i} can not be decisive
and we can not have ZAy. Thus we arrive at XAY, not XAZ, not ZAY.
For A=<" this amounts to x..t..y, z{x, y~z which contradicts transi-
tivity. For A=>- we have y.l.,x, x{ z, z4y which too contradicts trans-
itivity. Therefore we must have I = {h} and I is a singleton.
Step 3. By step 2 there exists a decisive subset {h}. We shall

show (with A as in step 2):
xh <
h Yh => XAY
Suppose x,yEX with xh..(h Yh. Using the assumption of at least 3 indif-
ference classes we can find U,v,ZEX satisfying the following relations:
uh <.. h zh -<. h v h for I = {h}

Zj..(j u j <j Vj for hh and Xj <j Yj
Zj <j Vj .( j u j for hh and Yj <j Xj
Zj <"j u j "'j Vj for j*h and Xj "'j Yj
Since I = {h} is decisive and sh(u,z)=l,Sj(u,z)=-l for j*h it follows
that UAZ. Since zh <h v h and Zj.(j Vj for hh it follows by step
1 that ZAV. Transitivity of AE{< ,,>} implies UAV. By construction
of u and v si(u,v)=si(x,y) for all i and stability of ~ yields XAy.
Step 4. By step 3 there exists an element of {l, ... ,n}, which we

denote by 0(1), and an element t.E{.(, '>} such that for all x,yEX the
following implication is true:
x 0 ( 1) .( 0 ( 1) Yo ( 1) => xt.y
From this we obtain
xt.y <=> (x o (l)-<o(l)Yo(l)) or (xo(l)~o(l)Yo(l) and xt.y)
and, possibly by interchanging x and y,
x..(y <=> (XO(l)t.(~)(l/G(ll or (xo(l)~o(l)Yo(l) and x<y),
with t.(l)E{<",»-}. Now, if for any x,yEX x o (l)"'o(l)Yo(l) already

implies that y~x, then the second term in the above formula disappears.
192
In that case the theorem is proven with m=l. Otherwise there exist
x,yEX such that x a(l),va(l)Ya(l)butX{Y. Lets for simplicity assume
that a(l)=l. Choose aEX l and define for U,VEX'=X 2X... xX n
u4'v if and only if (a,u)~(a,v) .
Obviously 4' is a quasi-ordering on X'. Defining si on X' for
i=2, ... ,n analogously to si on X it follows that si(u,v)=si«a,u),(a,v»
for U,VEX'. Since ~ is stable, 4' must be stable too. Furthermore,
~' is not trivial for the following reason. For x,yEX choosen above
we may write x=(x 1 ,u),y=(Y1'v) with u,vEX'. Because of Xl ~lY1 the
stability of=, implies that (xl'u)..«Y1'v) holds iff (a,u)«a,v)
holds. Now x<y, and hence u <'v. Thus ~' is not trivial. There-
fore the beginning of step 4 we may apply also to '" on X'. Thus
we may conclude as follows. Let x,yEX and write
x=(x 1 ,u), y=(Yl'v) with u,vEX'. If Xl -l Y1 then stability of ~
implies that x-<y is equivalent to u -<'v. Therefore from what we
had at the beginning it follows that (with a(l)=l)
x<y <=> (x 1 l:1(1)l Y1) or (xl "'lY1 and u <'v). If applied to ~'
on X' we obtain a(2)E{1, ... ,n}\{a(l)} and ~(2)E{.(,,>}such that
u <'v <=>(Xa(2)~~H/a(2» or (x a(2) ""a(2)Ya(2) and u -('v).
Putting together (and inserting a(l) for 1) we obtain that x<y holds
if and only if
(Xa(l)~(~~l)Ya(l» or (xa(l)~a(l)Ya(l) and xa(2)~(~~2)Ya(2»
and X
a(2)
~ y
a(2) a(2) and u .( 'v).
This process we may continue until we arrive after m iterations
with m~n at a trivial quasi-ordering. This completes the proof of
the theorem.
References
[1] Arrow, K.J., "Social Choice and Individual Values", 2nd ed.,
Yale University Press, New Haven and London, 1963 (first
edition 1951).
[2] Fishburn, P.C., Axioms for lexicographic preferences, Rev.
Econ.Studies 42 (1975), 415-419.
[3] Fuchs, L., "Partially Ordered Algebraic Structures", Pergamon
Press, Oxford etc., 1963.
[4] Krause, U., Recht und Gerechtigkeit, Okonomie und Gesellschaft
Jahrbuch 2: Wohlfahrt und Gerechtigkeit (1984), 158-172.
[5] Roberts, F.S., "Measurement Theory", Addison-Wesley, Reading,
1979.
193
[6] Steedman, I. and Krause, U., Goethe's Faust, Arrow's

possibi1ity theorem and the individua1 decision taker, in:
E1ster, J. (ed.), "The Mu1tip1e Se1f", Cambridge University
Press, Cambridge, 1986.
[7] Yu, P.L., "Mu1tip1e-Criteria Decision Making", P1enum Press,
New York and London, 1985.
WELL POSEDNESS,TOWARDS VECTOR OPTIMIZATION
Roberto Lucchetti
Department of ~1athematics, University of Milano
Via Saldini 50, 20133 Milano, Italy
Introduction. The paper has two goals: to revisit the concept of well
posedness in some fields where it has already been defined and to give
a tentative definition of well posedhess for vector optimization pro-
blems. As far as the first part is concerned, I shall refer to other,
more specific papers, for a deeper analysis: here I want only to put a
few of properties of well posedness in some evidence: in particular I
shall spend some words about the ordinal character of the property, star-
ting from Patrone 1987, because of the originality of this approach,
that seems to me of some interest. This analysis leads to a definition
of well posed problems that does not involve the given objective fun-
ction, but only the preference system with respect to which the opti-
mization process must be done. To be more detailed about the organiza-
tion of the paper, section two will deal with minimum problems, section
three with variational inequalities and section four with saddle point
problems. The first part is then terminated with section five, where
some words are spent to find a common origin between previous defini-
tions and some comments are presented about the Nash equilibrium pro-
blem: this is the starting point for the second part, which is develo-
ped in section six. Here a tentative definition is given for well posed-
ness in vector optimization problems, furthermore I shall present some
simple examples and initial remarks. A deeper study of the definition,
a further analysis showing how these (and other) properties here pre-
sented for minimum problems can be translated in the vector optimiza-
tion case will be made somewhere else.
Section 2. Minimum problems

Let X be a topological space, K a subset of X, f: X + lli a given function.
We shall consider the problem (P) : minimize f on the subset K. Someti-
195
mes (P) will also denoted by (f,K) to emphasize the data of the problem.
Definition 2.1 {Tyhonov 1966) The problem (f,K) is (Tyhonov) well posed
if:
there exists Xo E K such that f{xo)~f(x) ~ x E K and
v {x } E K such that{f{x »+ f{x o ), then{x }+ xo.
n n n
For motivations,comments and further references see,for instance, Nas-
hed 1982,Lucchetti 1983,Dontchev-Zolezzi:here I make only the following
remarks: the definition actually requires uniqueness of the minimum po-
intra definition of generalized wellposedness (g.w.p.) can be given as-
king that every minimizing sequence has a subsequence converging to a mi-
nimum point;this implies compactness of the solution set:see for instan-
ce Lucchetti 1982.Finally an alternative definition is given in Bednar-
czuk(1984) requiring only upper semicontinuity of the level set multi-
function of f , restricted to the set K.This definition is very close
to g.w.p. as it easy to see. Observe also that in definition 2.1 one
can pay attention only to sequences{x} such that{f{x )}is decreasing
n n
to f (x o ) •
Example 2.2. (Rockafellar 1970). Let f: mn + (_oo, +~ be a convex,lo-
wer semicontinuous function,with one and only one minimum point.Then
(f,mn ) is T.w.p .. Observe that this result can be false for an analo-
gous function defined on an infinite dimensional Hilbert space.
Example 2.3. Let X be a Banach space,x E X,K a closed convex subset of
X. Let flu) = iu- xl. This is the so called best approximation problem.
It is easy to show the following statements:
if X is reflexive,then f has a minimum point on K, for every x and K,
if X is strictly convex,then f has at most one minimum point on K;
if moreover the norm on X is Kadec,that is{x}+ weakly x,{lx I}+ Ix
n n
imply {xn}+x, then (f,K) is T.w.p.
Less easy,but very relevant,is showing the o~posite result: namely if the
best approximation problem, for a fixed x E X,is T.w.p. on every closed
convex set K, then X necessarily fulfills the three mentioned proper-
ties{Holmes '1972, where T.w.p. is called strong solvability). There are
many characterizations of T.w.p.,mainly in the convex case: many of
them are collected, for instance, in Lucchetti 1983. Here I recall only
one of them,both for its elegance and simplicity and for later uses.
196
Theorem 2.4 (Furi-Vignoli 1970) Let X be a complete metric space and f:X
7m lower semicontinuous and bounded from below. Then diam {x : fIx) ~
inf f + a } 7 0 as a 7 0 if and only if (f,x) is T.w.p ..

A similar statement involving the measure of non compactness of the le-
vel sets of f holds also for g.w.p.,see Furi-Vignoli(1970]. Now I am
going. to present a choice of properties of T.w.p. mainly in the convex
setting.The first one, illustrated by the following three theorems,rela-
tes well posedness with continuous dependence of the solution sets on
some data defining the problem (p). We start by intoducing some termi-
nology.Given a metric space X and a sequence of (closed)convex sets Kn
let us call
liK x,x n E K for all large n }
XEX : J {x } 7
n n n
lsK XEX : 3 {x k}
7
x'X k E Kn ,n k a subsequence of the integers}
n k
Definition 2.5 We say that {K } 7 K in the Kuratowski sense if lsK c K
n n
C liK
n
When X is a Banach space a useful tool,for instance for approximation
problems,is to combine strong and weak convergence in the previous de-
finition,this leads to the following convergence of sets:
{K } 7 K in the Kuratowski-Mosco sense if wlsK c K cliK Here w
n n n
stands for weak convergence:namely the approximating sequence in the
definition of ls must here converge in weak sense. Moreover let epif
= { r } the epigraph of f:we say that {f } conver-
(x,r) E XXlR : f (x) ~
n
ges to f in some sense if { epif n } 7 epif in the same sense.Next,for
a set C and a positive real number a let Ca { x : dist (x,C) <a}.
Definition 2.6 {K} 7 K in the Hausdorff sense if
n
Va>O 3ft K C Ka
Vn > ft and
n
\!a>O 3 ft Vn > ft K C Ka
n
Definition 2.7 A multifunction K :P 7 X is said to be upper semicon-
tinuous at Po if V 0 open in X ,0::> Kp",
-
V {p n } 7 Po then KPnC 0 for
all large n, K is said to be lower semicontinuous at Po ifV{p } + Pothen
n
liKPn~ Kpo. Here P is a metric space with the meaning of a parameter
space .
Theorem 2.8 (Lucchetti-Patrone 1982) Let f:X + ~ be a convex function
which is uniformly continuous on bounded sets. Moreover suppose that
(f,K) is T.w.p. for every closed convex set K. Then ,if {K } 7 Ko in
n
Kuratowski-Mosco sense, we have that mn converges to mo,where mn are
197
the minimum points of f on the (closed convex) sets K •

n
Theorem 2.9 (Lucchetti-Patrone 1982) Let f : X +m be a convex cQnti~
nuous function with one and only one minimum point on every closed con-
vex set K. Suppose that for every {K } converging to Ko in the Hausdorff
n
sense we have that {m } converges to mo. Then (f,K) is T.w.p. for all K.
n
Theorem 2. 10 (Lucchetti 1984) Let f :XxP +m be uniformly continuous
on bounded sets,f(.,p)a convex function for every p,let K :P +X_be a lo-
wer and upper semicontinuous multifunction at the point po.Consider the
problem (Pp) : minimize f(.,p) on Kp.
Let Mp the solution set of (Pp). If (f(.,po) ,Kpo) is T.w.p.,then M is
upper semicontinuous at Po.
Some comments about previous theorems are needed.The first one shows
how T.w.p. can improve convergence of solutions: it is a standard argu-
ment that in the hypotheses and notations of theorem 2.8,without T.w.p.
it is only possible to get weak convergence of m to mo 1hence T.w.p.
n
improves it to strong convergence.Theorem 2.9 is a kind of converse of
theorem 2.8 since it shows that if we want strong convergence of the
solutions,then we must deal with T.w.p. problems.Both theorems 2.8 and
2.9 have,in my opinion, an important theoretical meaning;namely they
show a connection between T.w.p. and continuous dependence on the da-
ta under perturbations(sometimes this property is called Hadamard w.p.).
Theorem 2.10 can be considered as a similar result,but with a different
emphasisjhere T.w.p. is assumed only for the limit problem PPo:hence
the result can be considered as easier to handle. Moreover it clarifies,
in the (restricted) setting previously presented,famous Berge's theorem
(Berge 1963) and its successors,as it shows how T.w.p.,rather than com-
pactness of the constraint set, plays a central role in the upper semi-
continuity property of the solution set. By the way,obs~ethat T.w.p.
is a property of both the function and the constraint set,considered
together. Previous theorems are only some examples of results in this
direction:some more are in the papers already cited and, for instance,
in Lucchetti1982,Bednarczuk1982,1984.Some consequences of theorem 2.8
applied to the best approximation problem are presented in Lucchetti1985.
The next aspect related to T.w.p. I shall consider here is the follo-
wing:how many T.w.p. problems are there in some specified classes of
198
minimum problems? Let us call GURn ) (or simply G) the set of functions
f: mn + [_00 , +00) which are convex and lower sernicontinuous. On G we
consider Kuratowski convergence.It is well known that G then becomes a
metrizable compact space. Inside G we consider the set N = {f £ G such
that f is Frechet differentiable and (f(.) _(y,.)~n) is T.w.p. for e-
very y£ mn }.
Theorem 2.11 (Patrone 1986) N is dense in G and G-N is a first catego-
ry set.
Last theorem then shows that a very particular subset of T.w.p. func~
tions is a "large" set inside G. The paper proving the theorem was ori-
ginated by a previous one (Lucchetti-Patrone 1978) where it is possi-
ble to find other results in this direction, for different classes of
problems,and also the infinite dimensional case is investigated:it
should be noticed that here Kuratowski-Mosco convergence does not pro-
vide the same strong result. Last property I present in this section
is the so called ordinal property. I mention that in game theory or in
mathematical economics we often have some preference systems, among
the data of the model, (defined on certain sets),rather than a repre-
sentative function of these preferences. Then it is not a nonsense to
ask whether T.w.p. of an optimization problem depends or not on the par-
ticular function we choose to represent the preference system. We have
the following:
Theorem 2.12 (Patrone 1987) Let X a connected topological space , ~ a
total preorder on X, f,g : X +m continuous functions both represen-
ting :$ • Then (f,X) is T.w.p. if and only if (g,X) is T.w.p.
The connectedness hypothesis can be accepted without any trouble, whi-
le continuity seems to be rather unnatural: in minimum problems it is
lower semicontinuity that people accept without comments. But it is an
obvious remark that if f has on X one and only one minimum point xo'
then we can substitute the problem (f,X) by the equivalent problem
(g,X) where g(x) = fIx) for every x r Xo and g(x o ) = f(xo) - 1. We do
not change the underlying preference system ,we do not destroy lower
semicontinuity and we get a (trivial!) T.w.p. problem. However this is
a rather stupid way to go on. We can observe the g has the (strange)
property that every minimizing sequence eventually coincides with the
minimum point. Then, from now on, we consider problems without this
199
kind of triviality. Then we can state the following:

Theorem 2.13 Let ~ be a total preorder on a topological space X. Let
f: X ... JR be a function representing ~ (namely x ~ y iff f (x) :;; f (y) ) •
Then (f,X) is T.w.p. if and only if:
there exists Xo E: X such that Xo < x Vx t- Xo and
v {x } E: X such that '" y t- Xo xn -< y for all large n, then {x }.... Xo
n n
Proof. Let x } such that {f(x )} ... f(x o ). Given y E: X such that f(y»
n n
> f (x o ) then f(x ) < f(y) which implies x < y; hence { x } ... xo. This
n n n
shows one implication. Conversely suppose we have a sequence'{ x n } such
that, for every y < y for all large n. If limsup f(x n ) = a >f(x o)
E: X, X
n
then there exists some z E: X such that z t- Xo and f(z) < a ( here I am
using the existence of non trivial minimizing sequences). But this im-
plies z< x n ' at least for a subsequence, against the hypothesis. Then
{ f (x )} ... f (x o )' hence {x } ... Xo and this concludes the proof.
n n
As a final remark of this section I remark that the condition expressed
on the preorders in the previous theorem can actually be taken as the
definition of T.w.p. for a minimum problem. It is a little more invol-
ved than the original one, but it has the property of emphasizing the
role of the preference system underlying the optimization process and
it avoids trivialities.
Section 3. Variational Inequalities
This section and the next one are really a quick overview of w.p. in
variational inequalities and in saddle point problems; this is due to
two main reasons: these topics are not so strictly related to vector
optimization as the minimum problem ;moreover they are not so widely
studied and of cornmon knowledge. It should be at first observed that
the minimum case is in both extensions at the basis of the given defi-
nitions. To state the problem let X be a reflexive Banach space and K
a closed convex set of X; let A: X ... X* be a monotone operator and let
f:X .... JR be a convex function. Here the problem (V.I) is to find
x E: K such that (Ax, x-y) + f (x) - f (y) :;; 0 for every y E: K.
Let H (a) ={ x E: K : (Ax,x-y) + f (x) - f (y) :;; a Ix - y I for every y }
Then we have the following:
Definition 3.1 (Lucchetti-Patrone 1981) We say that the problem (V.I)
200
is w.p. if H(a) # ~ for every a > 0 and diarnH (a) ... 0 as a .... o.
It is apparent that theorem 2.4 inspired the definition, if the sets
H(a) are considered as a kind of level sets for the variational inequa-
lity. Furthermore it is well known that a convex minimum problem can
be stated in terms of a variational inequality by means of the deriva-
tive of the functional to be minimized: hence next theorem can repre-
sent a further motivation for the given definition:
Theorem3.2 (Lucchetti-Patrone 1981) If A = g' for some function g which
is convex, lower semicontinuous and Gateaux differentiable and if f + g
is bounded from below, then (V.lj is w.p. if and only if (f+g,K)is T.w.p.
Next, I present one example and two theorems on w.p. variational ine-
qualities. Both theorems deal with a linear operator : the first one
exibits some conditions which are equivalent to w.p.,while the second
one focuses attention on an equation, rather than an inequality. Namely
(V.I) reduces to find x such that Ax = f when the set K is the whole X
and f is linear.
Example 3.3 (Lucchetti-Patrone 1981) If A is hemicontinuous and stron-
gly monotone, if f is convex lower semicontinuous and not identically
+ 00 , then (V.I) is w.p ••
Theorem 3.4( Lucchetti-Patrone 1982-83) Let X be a Hilbert space and A
a linear bounded operator. Let f E: X. Then the three following conditions
are equivalent:
(V. I) is w. p. on every closed convex set K
- (V. I) has one and only one solution on every closed convex set K
- A is coercive( namely there is a > 0 such that (Ax,x)~ alxl 2 for all x)
Theorem 3.5 (Lucchetti-Patrone 1982-83) Under the same hypotheses of the-
orem 3.4, suppose moreover that K = X • Consider the problem: find x E: X
such that Ax = f (as a V.I) . Then the following are equivalent:
- A is a homeomorphism
(V. I) is w.p.
for every g E: X,Ax = g is w.p. (as a V.I ).
Observe that this last theorem deals, as theorems 2.8,2.9,2.10, with
the problem of relating well posedness with continuous dependence on
the data of the solution sets:here this fact is stated by means of the
continuity of the inverse operator A- 1 .
201
Section 4. Saddle point problems
In this section I shall have in mind the saddle point problem as a spe-
cialization to the zero sum case of t~e Nash equilibrium prob~em of a
two players non cooperative game. This can give the idea of extending
concepts to this last case: in fact something is done, but in my opi-
nion the general case must be studied more deeply; so here I confine
myself to the zero sum case. Next section will contain some remarks on
the Nash problem. As usual, some terminology is needed. Let X,Y be me-
tric spaces, f: XxY +~ a given function. The saddle point problem (S)
is to find (x,y) EXXY such that f(x,y);'; f(x,y);'; f(x,y) 'i XEX,'i YEY.
Let h(x) = inf f(x,y), g(y) = s~pf(x,y) and w = g - h. Observe that
y
Wf;O; a sequence {(x ,y )} is said maximinimizing if {w(x ,y )}+ O.
n n n n
Finally let E (a) = { (x,y) E XxY: f (x,y) f; sup f (x,y) - a,f (x,y);'; inff (x,y)
x
+a } •
Definition 4.1 (Cavazzuti-Horgan 1983) (S) is w.p. if:

there exists a saddle point (xo,Yo) and
every maximinimizing sequence converges to (xo,Yo).
Next theorem is not hard to prove: hence it is stated without proof.It
should be compared with theorem 2.4 and definition 3.1.
Theorem 4.2 Let X,Y be complete, f(.,y) upper semicontinuous for every
y, f(x,.) lower semicontinuous for every x. Then the following state-
ments are equivalent:
(S) is w.p.
- diam E(a) + 0 as a + O.
Following the usual procedure I shall present one example and some the-
orems on w.p. of saddle point problems.
Example 4.3 T"et f : ~mx ~n + ~ be a function with one and only one sad-
dle point (x,y) and such that f (. ,y) is concave for every y and f (x,.)
is convex for every x. Then (S) is w.p ..
To see this simply observe that for such a function f the associated g
and h are convex and concave respectively : moreover the minimum pro-
blem for g and the maximum problem for h have as unique solution y and
X. Hence (h,lRn) and (g ,]Rm) are T.w.p. ( see example 2.2) . Now we can con-
clude by means of theorem 3.2 in Cavazzuti-Morgan 1983.
Theorem 4.4 (Patrone-Torre 1986) X,Y are Hilbert spaces, A: X + X,
B: Y + Y, M: Y + X are linear bounded operators , A and Bare self-
202
adjoint ,-A and B are positive, a and b are elements of X and Y respec-
tively. Let f(x,y) = (Ax,x) + (By,y) - 2(a,x) - 2(b,y) + (x,My).
Then the following statements are equivalent:
(S) is w.p. on every HxK,Hc::: X, K c: Y closed convex sets
- (S) has one and only one solution on every HxK
- -A and B are coercive operators.
This is the analogous of theorem 3.4 in previous section. Next theorem
deals with the ordinal property already investigated in the minimum case.
Theorem 4. 5 (Patrone 1987) Let X, Y connected,~ a total preorder on XxY,
f1 and f2 continuous functions both representing ~ • Then (S) with f1
is w.p. if and only if (S) with f2 is w.p .•
I want to conclude this part by saying that perhaps it should be of so-
me interest to investigate density and genericity properties for w.p.
(V.I) and (S) problems and connections between w.p. and continuous de-
pendence on the data: I do not know results in this direction.
SectionS Conclus-ions of the first part
I draw here some conclusions arising from previous sections and I in-
troduce w.p. in the context of vector optimization problems (V.O ): de-
finition of w.p. for (V.O) and some consequences are object of the next
section. To find out a possible schema for w.p. we can see that there
is the introduction of a notion of (value of the problem and of) appro-
ximate solution sets, with increasing precision in the approximation.
Hence w. p. will mean:
there is a solution of the problem and
every sequence of approximate solutions converges to the solution, when
the error in the approximation goes to zero.
An alternative to the second condition is:
the diameter of the approximate solution sets goes to zero.

Some words should be spent about the value of the problem. For saddle
point problems ,when it happens that infsup f(x,y) supinf f(x,y)
y x - x y
then the problem has its value, which is univoquely determined,even if
the problem has no solution or more than one solution. Furthermore,e-
xistence of the value is equivalent to the existence of maximinimizing
sequences(theorem 2.1 in Cavazzuti-Morgan 1983). We can conclude that,
203
as in minimum problems, approximate pOints are characterized by having

values near to the value of the problem. Let us remark that for Nash
problems the situation radically changes. Given two functions f and g
it should be natural to define the approximate solution set E(a) by
E(a) = {(i,9)
e:XxY : f(i,9)1i: sup f(x,9) - a, CJ($C,~) Ii: sup g(i,y) - a }.
x y
This was effectively done( see Tijs 1981, Lucchetti-Patrone 1986). Fur-
thermore the fact that diam E(a) goes to zero when a goes to zero is
related to a tentative definition of w.p. Nash problems(see Patrone 1987).
But it can happen, for instance, that there is {(x ,f )}e:E(a ) for so-
n n n
me {a } -+- O,such that f(x ,'1 ) and g(x ,'1 ) both tend to - 0 0 . This situa-
n n n n n
tion seems to be heavily unpleasant for both players. Here the lack of
a natural idea of value can be at the basis of this feature. Moreover
the given definition of w.p. for Nash problems does not have in general
the ordinal property. These considerations justify in my opinion need
of deeper investigation. We saw before a possible schema for w.p •• Of
course it is also possible to give a schema for g.w.p.: it is enough
to ask for compactness of the solution set( rather than uniqueness of
the solution) and convergence of a subsequence of approximating sequen-
ces. Finally it is the time to deal with vector optimization. We start
with the following remarks: the idea of approximate solution set is far
from being univoquely determined: it is enough to recall the paper of
White 1986, that lists six e:-efficient sets. One of them was already
object of a previous paper:seeLoridan1984. See also Staib 1986.Moreo-
ver it is a clear non sense to ask for uniqueness, or even compactness
of the solution set. What is said in this section can suggest the idea
of selecting a set of values and, for every value, to look for appro-
ximating sequences. The convergence of these sequences to the solution
characterized by the value, will be called w.p •. We shall see this in
the next section.
Section 6 Vector optimization and well posedness
I do not want to introduce here too many notations and too heavy termi-
nology: as we only give an attempt to start with the argument, we can
confine ourselves to a simple setting. Thus (V.O problem can be sta-
ted in such a way: suppose we have a function f: X -+- Y, where X, Yare
204
Banach spaces and there exists a closed convex cone C in Y which gene-
rates a preorder in the natural way: x ~ y iff y-x E C. Given a set K c
C X the vector optimization problem is: to find
x ~ K such that (f (x) - C ) " f ( K ) {f (xl).
The first step for giving a definition of w.p. is to define inf f(K).
Given a set AcY the most natural idea for defining the set inf A is
to keep all the elements which are minimal for the closure of A; namely:
Definition 6.1 x E inf A if x E A and there exist~ no y f x such that
x- YE C.
This definition can be found in Cesari-Suryanarayana 1976 and it is u-

sed for instance in Staib 1986. Slight different definitions can be
found in Caligaris-Oliva 1981 and in Azimov 1985. However, from now on,
the set inf A is intended as in definition 6.1. Now we can give the de-
finition of well posedness:
Definition 6.2 The(V.O ) problem (f,K) is w.p. if
the set inf f(K) is nonempty
for every 1 E inf f (K) there exists xl E K such that f (xl) = 1 and
for every {xnh K such that f(x n ) - f(x n + 1 ) E C and{f(x n )} ->- 1, then
{x ->- l.
n
As initial comments, I observe that some conditions guaranteeing that
inf f(K) is nonempty can be found for example in Staib 1986. Moreover
the monotonicity condition focuses attention only to really relevant
approximating sequences; finally it is apparent how to define g.w.p ..
Now, two simple examples of w.p. problems:
2 2 2
Example 6. 3 . f: JR ->- JR f(x,y) (x,y). K = {(x,y): x+y?; 1},C =JR
+
2 2
Example 6.4 f: JR ->- JR f(x,y) (-Ixl ,-Iyl), K = {(x,y): Ixl+lyl;;;1}
2
C = JR + . Then (f ' K) lS
. g.w.p ..
For what previous examples are concerned, I want to pay attention to
the scalarization problem: this means, roughly speaking, finding the
efficient points solving scalar optimization problems. In the first e-
2
xample, if we look for a E JR+ such that a·f gives efficient points, the
unique possible choice ( up to normalization) is a (1,1). The related
scalar problem is not T.w.p .. In the second example only the efficient
values (-1,0) and (0,-1) are the unique solution( value) of some scala-
rization process. So we can conclude that w.p. (V.O) problems can ge-
205
nerate, via scalarization, minimum problems which are not T.w.p •• It

is not surprising, however, that scalarization can destroy w.p.,as it
is essentially a linearization, and linear problems often do not beha-
ve well with respect to w.p •• The next example presents a simplified
version of the vector best approximation problem. A vectorial norm, as
introduced by Kantorovitch 1939, is a function II II : X ->- Y suct that:
II x II E C 'V x and II x II 0 iff x =0
II Ix II = I 1 I II x II \f IE JR, 'V X EX
Ilxll+llzll-llx+zIIE C, 'V x,z EX.
Fixed an element x EX, a vector space U, a closed convex set K c:. U and
a function f: U ->- X the vector best approximation problem, in its full
generality, is defihed as:
"minimize" II f (u) - x II , u E K (see for instance Jahn 1984).
Example 6.5 Suppose we have a space X which is reflexive Banach space
with respect to m equivalent norms I 11 , •.• ,1 1m. The space is also stric-
tly convex for at least one norm, sayl 11 , which is also Kadec( see e-
xample 2.3). Given m points x 1 , •.• ,xm and a closed convex set K, consi-
der the problem of "minimizing" the function f:X ->- JRm defined as f (x) =
= ( Ix - x111, ••• , Ix - x I ) where onJRm we consider the usual cone JRm•
rn m +
To verify w.p. of the problem, observe at first that inf f(K) ~ ¢, for
m m m
instance because f(K) - JR+ haSJR+-compact sections; moreover f(K) -JR+
is a closed set, as it is not difficult to see; this imples that to e-
very IE inf f(K) it corresponds at least one element xl such that f(x l )=
1. Suppose now that there exists some y E K, Y ~ xl' such that f (y) =
f(x l ). Hence IXI -xiii = Iy - xiii for i = 1, ••• ,m. By convexity of K
and strict convexity of I 11 , it follows that f(x l ) - f(!(xl-y)) EC-{O},
against optimality of xl. Hence the solution is unique. Moreover let
{xnh K and {f(xn )} ->- f(x l ), then {I xn - x111}->- IXI - x111· As {x n }
is a bounded sequence, there exists a subsequence weakly converging to
m
an element y E X. This in turn implies f(x l ) - fly) E JR+. Hence xl y,
all the sequence xn - x 1 converges weakly to xl - x 1 and we conclude by
using Kadec property of I 11 • To conclude the paper I recall that a de-
finition of g.w.p. for vector optimization problems is given also in
Bednarczuk, together with connections between this property and conti-
nuous dependence of solutions on the data, with results similar to those
206
here recalled for minimum problems ( see in particular theorem 2.10).
References
A.Y.Azimov 1985 Duality theorems for multicriteria problems, Soviet

Math. Dokl. Vol.31
E.Bednarczuk 1982 On upper semicontinuity of global minima in constrai-
ned optimization problems, J. Math. Anal. Appl. Vol.86
E.Bednarczuk 1984 Continuity properties of solutions in perturbed con-
strained minimization problems, Ph. D. Thesis
E.Bednarczuk to appear. Paper presented at the International Conference
on Vector Optimization,Darmstadt 1986.
C.Berge 1963 Topological spaces, Macmillan Co. N.Y.
O.Caligaris-P.Oliva 1981 Necessary and sufficient conditions for Pareto
problems, Boll. Un. Mat. Ital. Vol.18
E.Cavazzuti-J.Morgan 1983 Well posed saddle point problems, Lectures
Notes Pure Appl. Math. Vol.86
L.Cesari-M.B.Suryanarayana 1976 Existence theorems for Pareto optimiza-
tion in Banach spaces, Bull. Amer. Math. Soc. Vol.82
A.Dontchev-T.Zolezzi Book in preparation
M.Furi-A.Vignoli 1970 About well posed optimization problems for func-
tional in metric spaces, J. Optim. Theory Appl. Vol.5
R.B.Holmes 1972 A course on optimization and best approximation, Sprin-
ger Verlag, Berlin
J.Jahn 1984 Scalarization in vector optimization, Math. Progr. Vol.29
L. Kantorovitch 1939 The method of successive approximation for functio-
nal equations, Acta Math. Vol.71
P.Loridan 1984 e-solutions in vector optimization problems, J.Optim.
Theory Appl. Vol.43
R.Lucchetti 1982 Some aspects of the connections between Hadamard and
Tyhonov well posedness of convex programs, Boll. Un. Mat. Ital.C Vol.41
R.Lucchetti 1983 Hadamard and Tyhonov well posedness in optimal control,
Methods Oper. Res. Vol.45
R. Lucchetti 1984-85 On the continuity of the minima for a family of
constrained optimization problems, Numer. Funct. Anal. Optim. Vol.7
R.Lucchetti 1985 Convergence of sets and of projections, Boll. Un. Mat.
Ital. C Vol. 4
R.Lucchetti-F.Patrone Sulla densita e generic ita di alcuni problemi di
minimo ben posti, Boll. Un. Mat., Ital. B Vol. 15
R.Lucchetti-F.Patrone 1981 A characterization of Tyhonov well posedness
for minimum problems, with applications to variational inequalities,
Numer. Funct. Anal. Optim. Vol.3
R.Lucchetti-F.Patrone 1982 Hadamard and Tyhonov well posedness of a cer-
tain class of convex functions, J. Math. Anal. Appl. Vol.88
R.Lucchetti-F.Patrone-S.Tijs .1986 Determinateness of two person games,
Boll. Un. Mat. Ital. B Vol.5
M.Z.Nashed 1982 On well posed and ill posed extremal roblems, Nonli-
near phenomena in mathematical sciences, Proc. Int. Coni. Arlington,Te-
xas, (1980)
F.Patrone 1986 Most convex functions are nice, to appear on Numer. Funct.
Anal. Optim.
207
F.Patrone 1987 Well posedness as an ordinal property, to appear

F.Patrone-A.Torre 1987 Characterization of existence and uniqueness for
saddle point problems and related topics, Boll. Un. Mat. Ital.to appear
R.T.Rockafellar 1970 Convex analysis, Princeton University,Princeton
T.Staib 1986 On two generalizations of Pareto minimality,Preprint 979
Technische Hochschule Darmstadt
S.Tijs 1981 Nash equilibria for non cooperative n-person games, Siam
Rev. Vol.23
A.N.Tyhonov 1966 On the stability of the functional optimization pro-
blem, USSR Compo Math. Math. Phys. Vol.6
D.J.White 1986 Epsilon efficiency, J. Optim. Theory Appl. Vol.49
ON SOME APPLICATIONS OF STOCHASTIC
DOMINANCE IN MULTIOBJECTIVE
DECISION-MAKING
Leonid I. Polishchuck
Institute of Economics and Industrial Engineering
Siberian Branch of the USSR Academy of Sciences
Prospect Lavrentjeva 17, 630090 Novosibirsk 90, USSR
Abstract: Some multiobjective problems of economic origin are

considered in which both quantitative and qualitative goals are
incorporated, the latter represented by binary relation of stochastic
dominance. As an example can be mentioned the problem of attaining of
a compromise between two goals, namely the increase of average well-
being of a population and the reduction of living standards
differentiation between individuals. The quantitative indicator is put
in correspondence with the first goal, whereas the second one is
represented by partial ordering known as Lorenz domination. The second
example is related to decision-making under uncertainty, when on a level
with the conventionally interpreted efficiency the high reliability of
plans is to be ensured. The necessary and sufficient conditions for the
effectiveness of solutions in the similar problems are obtained. Some
properties and methods of construction of these solutions are considered.
1. Introduction
The situations are widespread in decision-making when the objective

function ~O : X ~ ~ ordering the elements of the admissible set X is
unknown explicitly, but the family ~ of such functions which is known
to contain ~O can be characterized. In this case the quasi-ordering
R~ c X x X, defined by the following conditions
xR~y ~ ~ (x) :? ~(y), V ~ E ~(x,y EX) (1)
and known as stochastic dominance generated by family ~ (see e.g. [1J)

can be used for the preliminary reduction of the domain where the
optimum should be searched. The totality of R~-nondominated elements
of X
{x E X I (2)
209
provided being exteriory stable (Le. for any x E X there exists

x * EE(R~,X) such that x *
R~x) a .
fort~ory contains the points of
maximum of ~O (if the latter exist).
The well known special case of stochastic dominance is the classical

Pareto ordering on set X clRn when the family ~ = ~o is formed by all
monotonically non-decreasing functions on X. Here for x = (x 1 , ••. ,x n ) ,
y = (Y1' ••• 'Yn) we have
Vi 1, •.. ,n,
and E(R~ ,X) is the Pareto frontier E(X) of the set X.

o
The Pareto optimality principle playing a key role in vector
optimization theory does not reflect, however, interdependences of
variables x 1 , ••• ,xn (interpreted as partial criteria) and therefore
usually is to be strengthened. Hereafter we'll confine ourselves to
the case of the equal importance (equity) of x 1 , •.• ,x n which appeares
in a number of practically important decision-making problems.
We indicate here two most typical examples. The first one relates
to the social welfare problem (hereafter SW-interpretation), when
variables xi are comparable quantitative welfare characteristics (say,
incomes) of n individuals. As a second example can serve the problem
of decision making under uncertainty (UDM-interpretation): in this
case the components of x represent the performance characteristics of
stochastic system in n potentially possible and equiprobable states
of nature 1 ).
The strengthening of Pareto-optimality with regard to equal

importance of x 1 , .•• ,x n can be provided by narrowing of ~O to the set
n
~ 1 {~(x1, ••• ,xn) = L: f(x i ) I f : R .... lR is a
i=1 (3)
monotonically increasing function}.
In this case [2] xR~ y (x,yEX) iff

1
card {i I x. ::; t} ::; card {i I y. ::; t} , V tEE, (4)

~ ~
which corresponds to the classical notion of stochastic dominance

210
(see e.g. [1,2]) for discrete random variables with equiprobable

values 2 ). In SW-interpretation the domination XR~1Y means that for
any income level t the size of population with incomes exceeding t is
in variant x not less than in y.
This requirement being based on egalitarian principles does not

exclude, however, anyhow large differentiation of x 1 , ••. ,xn and
therefore does not ensure the reduction of individual welfare
difference. The latter problem of social policy has an equivalent in
UDM-interpretation, namely the increase of the reliability of plans
in presence of uncertainty - if we assume that reliability can be
represented by the scatter of performance characteristics.
The common base for the qualitative characterization of quantities

x 1 , .•• ,xn differentiation is the Lorenz domination [2,5). We recall
that for non-zero x, y ER~ vector x Lorenz-dominates y (written as
xLy) iff
k k
1: Xi 1: Yi (5)
i=1 i=1
x·lI ~
y.1I Vk 1, .•. ,n-1,
where 11 = (1, •.• ,1) and x,y are the permutations of x,y components
respectively such that 1 ~ x
relation also is a special case of stochastic dominance: L = R~ with
2
n x.
{(j) (x) =1: f (x.~1I) I f R + R is a convex
i=1 (6)
continuous function}.
Relations R~ and R~ are in general position with each other so

1 2
that the radical levelling of differentiation (increase of
reliability) may have adverse impact on the total value of x 1 ' •.. ,xn .
Therefore when evaluating alternatives x E X it makes sense to be
efficiency m(x) = *
guided in line with the Lorenz domination by the average (expected)
(x·1I).
The decision making problem with two goals represented by Lorenz

domination L and the quantitative criterion m, and the corresponding
conflict between equity (reliability) and average (expected)
efficiency is the main subject of our analysis 3 ) .
211
Comparisons of alternatives from these requirements point of view

lead to the stochastic dominance relation a~3' where
(7)
We note that relation R~3 induces on E~~{O} convex but non-conical

domination structure (in sense of [6]): sets D (x*) == {x EJR~ ~ {O} I
XR~ x*} can be described as
3
D(X*) (8)
(here En is the totality of all possible permutations of indices

* - (x* (1), •.• ,x* (n»)' and, being convex, aren't
{1, .•. ,n}, an~)xcr cr cr
however cones
The R~3-nondominated elements of the set X c JR~ which can be

interpreted as effective compromises between equi,ty (reliability)
and average optimality (hereafter - effective EAO (from Equity-
Average Optimality) compromises) can naturally be treated as potential
solutions of the problem of two mentioned goals reconciliation on X.
We concretize the corresponding definition as follows: the non-zero
vector x* E X is said to be effective EAO-compromise iff no x EX
exists such that xLx*, m(x) ~m(x*) and at least one of these
dominations is strict.
The properties of effective EAO-compromises have much in common

with the properties of classical Pareto optimums but at the same time
possess the considerable specificity. It is confirmed already by
conditions of EAO-effectiveness.
2. Conditions for EAO-compromises effectiveness
The following proposition provides conditions for effectiveness of

EAO-compromises in terms of dual variables associated with x 1 ' ••. ,xn •
By dual variables meant are the components of vector p = (P1, ••• ,Pn)
being an external normal to the set X in the point x* under
consideration. From SW-interpretation point of view the quotients of
dual variables Pi/Pj are the upper bounds for trade-offs characterizing
possible income redistributions: with income shrinking of the i-th
212
individual by /:, > 0, the income of the j-th individual may increase
not more than by (Pi/Pj)/:'. With regard to what has been said the
variable Pi can be considered as a comparative characteristic of
mutual efforts necessary to ensure the i-th individual's welfare.
Proposition 1. The sufficient condition for the EAO-compromise

x* EX effectiveness is the existence of a vector p EJRn such that
p'x* ~p.y, Vy EX, (9)
1, .•• ,n, (10)
p.x*>O. ( 11 )
If the set X is convex, the necessary condition for the compromis~

x* E X to be effective is the existence of a vector p EJRn , p t 0
satisfying condition (9) and such that
Vi,j = 1, ..• ,n, (10' )
p'x* ~O
5) • o (11 ' )
Thus effective EAO-compromises can be characterized by the opposite

ordering of individuals according to~ first, their incomes, and,
second, the outlays ensuring these incomes. In UDM-interpretation this
condition can be interpreted as risk-aversion [9]. The appearance of
this requirement when the decisions are being chosen with regard to
their reliability seems to be quite natura1 6 ).
As it is shown by simple examples, conditions- (9), (10'), (11') in

general case are insufficient for the EAO-effectiveness, and conditions
(9) - (11), generally speaking, are not necessary even for the convex
X. The gap revealed between necessary and sufficient conditions is
characteristic also for classical Pareto-optimality (see e.g. [10]).
Another reason for this analogy is that the definition of EAO-
effectiveness can easily be strengthened or weakened s-o that conditions
(9) - (11) (respectively (9), (10'), (11')) turn out to be necessary
and sufficient.
In first case the concept in question is a proper effectiveness

[11, 12] which in the problem considered is to exclude the possibility
of a finite increment to average (expected) efficiency with
213
"infinitely small" increase of inequality, and vice versa. In order

to arrive at the precise definition, let us call the transfer
(redistribution) of a part of the i-th individual's income to j-th
individual as admissible, if Xj > x j , and forbidden if otherwise. It
should be noted, that according to Dalton condition [2,5] every
admissible transfer of a small enough value leads to an income vector
that L-dominates the initial one. We call the standardized vector
x/(x·~) the distribution of an income vector x; the distribution of x
is said to be obtained from the distribution of x* by a system of

transfers Sij ~O, i,j = 1, ..• ,n, i '" j, if
*
~ ~+ j 1, ... ,n. (12 )
x • 11 x*· ~
Let us agree to call the effective EAO-compromise x* properly

effecti ve, if for certain a > 0 and for all sufficiently small £ > 0
the following conditions are met:
(i) If in the system of transfers transferring the distribution

of x* into the distribution of x E X there is an admissible transfer of
value £, then in the same system a forbidden transfer of value not less
than a £ can be found, or x • 11 :s; x* • 11 - a £ ;
(ii) If x • 11 > x* • 11 + £, x E X, then in any system of transfers

transferring the distribution of x* to the distribution of x, a
forbidden transfer of a value not less than a £ can be found.
Proposition 2. For EAO-compromise x* E X proper effectiveness is

sufficient, and in case of set X convexity-necessary as well the
existence of the vector p = (P1, ... ,Pn) meeting conditions (9) - (11).0
Conversely, we call admissible plan x* '" 0 weakly effective EAO-

compromise if x* is the boundary point of x7 ) and there is not x EX
such that m(x) > m(x*), xLx* and corresponding inequalities (5) for all
k = 1, ... ,n-1 are fulfilled strictly.
Proposi tion 3. For the compromise x* E X weak EAO-effecti veness is

sufficient, and in case of set X convexity-necessary as well the
existence of the vector p = (P1, ..• ,Pn) '" 0 meeting conditions (9),
(10'), (11').0
214
The symmetry of conditions (10'), (11') with respect to p,x

indicates that the problem of construction of (weakly) EAO-effective
compromises admits the dual statement identical to the initial one.
Since similar self-duality can be useful in economic applications
(see e.g. [13) we adduce the exact formulation of the corresponding
statement. t~ith this aim in mind let us suppose that admissible set X
satisfies the following assumptions:
X is convex, compact, saturated 8 ) and 0 E int(X -JR~). (13)
Let
xO = {p EJR~! p. x ~ 1, v x EX} ( 1 4)
by a polar set of X. It is known [14) that the transformation X 1-+ xO

is a one-to-one mapping of the totality of sets distinguishing by
00
conditions (13) onto itself, and moreover, X = X. We say that
elements x EX, P E XO are in the polar correspondence, if p • x = 1.
One can observe that the point x~O E argrnax{m (x)! x E X} where the
average efficiency m(xl attains maximum on X is in polar correspondence
with the point of absolute equity (reliability) p; = max{A! A~ EXO}~.
In general case the following statement holds.
Proposition 4. If set X satisfices conditions (13), then vector

x* E X is weakly EAO-effective compromise iff there exists weakly EAO-
effecti ve compromise p E x O of polar set x O beinq in polar correspondence
with x ~ 0
3. Some properties of effective EAO-compromises
We begin with connection of EAO-effectiveness with the usual

stochastic nondominatedness in the sense (4). Since families of
functions ~1 and ~3 are as before in general position, the EAO-
effectiveness generally does not imply neither R~ -nondominatedness
1
nor even Pareto-optimality. However under certain assumptions the
latter condition together with EAO-effectiveness is sufficient for
the R~1-nondominatedness.
Proposition 5. If admissible set X is saturated, then effective

EAO-compromise provided its Pareto-optimality is stochastically R~1-
nondominated:
215
E(R~ , X) n E(R~ , X) c E(R~ , X). 0

3 0 1
We note that the existence of effective EAO-compromises and what is

more, the exterior R~3-stability of the set E(R~ , X) is guaranteed,
for example, by compactness of X, because [7, 1sT
according to (8) the
quasi-ordering R~ induces closed domination structure.
3
The existence of properly effective EAO-compromises is also typical.

Besides, the following statements hold which also have well-known
analogies in Pareto-optimization theory [16].
Proposition 6. If X is convex and closed, then the totality of

properly effective EAO-compromises is dense in the set E(R~ , X). 0
3
Proposition 7. If X c]R~ is a convex polyhedron, each effective

EAO-compromise is properly effective, and the totality of effective
EAO-compromises E(R~ , X) is a union of a finite number of convex
polyhedrons of dimen~ion not above n - 1. 0
We indicate one simple method which leads to an effective EAO-

compromise or permits to improve (in terms of both equity and average
efficiency) an admissible plan not possessing the property of EAO-
effectiveness.
Assuming as before the convexity and closedness of X, let us fixe

the distribution a EJI?n_1 of income vector (JI?n-1 = {a E:R~I a' 11 = 1}
is (n-1) -dimensional simplex) and consider the s-calar convex
programming problem
t - max,
rda) x = at, (15 )
X E X9).
The solution of this problem has a form x = x(u) = at{a) where

t(a) is the maximum value of the functional coinciding with x(x(a».
In addition vector p = (P1"'.,Pn) of constraints Xj = ajt (j =
1, .•• ,n) dual prices in the problem n(a) is the external normal to
the set X in the point x(u).
As can easily be verified, any effective EAO-compromise x* E X is

automatically the solution of the problem n(a*) with a* = x /(x*. 11).
216
In order to determine whether the effective EAO-compromise keeps in

line with the preassigned distribution a* EF n _ 1 , one is to solve the
problem n (a *), and, obtaining vector of dual variables p, to test
whether it satisfies proposition 1 if coupled with vector a*.
Naturally, the fulfillment of sufficient conditions (10), (11)
guarantees the effectiveness of EAO-compromise x* = a*t(a*). At the
same time the violation of necessary conditions (10'), (11') implies
non-effectiveness, generally, only in case of the dual variables
vector being unique.
Thus, one can construct effective (or close to such) EAO-compromises

by fitting suitable distribution a, solving the problem n(a), checking
effectiveness conditions, and, possibly, adjusting the distribution
if its choice proved to be unsuccessful. We describe now such
adjustment in a comparatively simple situation.
Let p be a unique vector of dual prices in the problem n(a*), and

moreover p.a*>O, but the set J = J(a*) = {(i,j)1 a~>a~, p. >p.}
1 J 1 J
is non-empty. In this case vector x* = a* t(a*) is not an effective
EAO-compromise, but, as can be easily verified, with a sufficiently
small r-. > 0 and any vector h from a non-empty cone
K(a*) = { l:: ~ .. (e. -e.) I ~1'J' ;:'0, l:: ~ij>O}

(i,j)EJ 1J 1 J (i, j) EJ
an admissible vector x (a * + r-.h) = (a * + r-.h) t (a * + r-.h) dominating x* in

terms of both the average value m(x) and equity, is in line with the
distribution a * + r-.h.
4. Effective EAO-compromises and inequality measures
Another approach to the construction of effective EAO-compromises

consists in the employment for this purpose of the quantitative
indicators (measures) f : X + ~ satisfying the condition
yLx ==;> f (x) ;0, f (y) . ( 1 6)
Diverse variants of these indicators are proposed in the literature

[2,5]; as typical examples can be quoted the standardized moments of
order r ;0, 1
217
n
f(x) E (17)
m(x) i=1
the range function
f(x) = ( 1 8)
m(x)
and the Gini coefficient
n
f(x) E ( 19)
m(x) i,j=1
It should be stressed that with n > 2 the ordering of vectors x EJRn

generated by any inequality measure is not equivalent to relation L
because the former represents the complete ordering while the latter
only the partial one.
If in the analysis of conflict between average optimality and

equity we replace the Lorenz domination by its arbitrary indicator
satisfying conditions (16), then Pareto-optimal solutions x* E X of
thus formed bi-criteria problem «m,f) ,X) with vector criterion
(m,f) : X ~ JR2 more often than not appear to be effective EAO-
compromisei?)conversely, for any effective EAO-compromise x* one can
find the inequality indicator f such that x* is a Pareto-optimal
solution of the corresponding problem «m,f) ,X) (if not all the
components of x* are identical, the required indicator can be chosen
in the form
f(x) min{tl x E 1 11 + t(B(x*) n 11)},

x:1! Ii
where B(x*) = {y EJR~I y' 11 1, yLX*}).
Problems «m,f) ,X) have certain advantages over the initial

statement. The main one is that when the multiobjective choice is
implementing in the interactive regime, the decision-maker has the
quantitative characteristics of both goals attained levels.
This approach, however, has at the same time serious deficiences.

First, essential arbitrariness is introduced into the initial
statement of the problem, because the set of Pareto optimums in the
problem «m,f) ,X) depends essentially on the choice of f, and
218
generally by far not exhaust the totality of effective EAO-compromises.

Indeed, the former set is generally one-dimensional, whereas the latter
one according to proposition 7 can have dimension n-1. In general the
following statement holds.
Proposition 8. With n ~ 3 for any continuous inequality indicator

f there exist set X CJR~ and effective EAO-compromise x* EE(Rcp3'X)
such that x* is not a Pareto-optimal solution in the problem
«m,f),X). 0
The second drawback is that Pareto-optimal solutions of «m,f),X),

as well as effective EAO-compromises in general, don't have to be
RCP1-nondorninated11). Meanwhile the stochastic nondominatedness in
usual sense is a very desirable requirement incarnated the natural
ideas about the rationality of the choice. In this connection we shall
describe a bi-criteria model destined for the reconciliation of goals
considered in the paper, for which the Rcp1-nondominatedness of
corresponding Pareto-optimal solutions is guaranteed.
Namely, let us extend the description of admiss'ible set X by means

of adjoining with variables x 1 " " ' x additional (n+1)-th variable z;
n
let x :: ~
(x,y) EX:: X x:R+. In UDl.f-interpretation the variable z is to
be considered as a planned (declared) level of efficiency (which does
not have to coincide with the expected efficiency m(x». In SW-
interpretation this value is interpreted as an established welfare
standard.
The criterion m X --:R of the form
m(x,z) = z (20)
which requires the increase of planned efficiency (welfare standard),

can be considered in line with the minimizing value
_ 1 n
f(x,z) ::: Ii: E [z-xil+ (21)
i=1
([al+ :: max(a,O». The latter represents the expected excess of

planned efficiency over actual one (in SW-interpretation f(x,z) is
the average income deficient for the individuals to achieve the
standard z) and can therefore be considered as a (one-sided)
reliability (equity) measure.
219
Proposition 9. If the set X is co~pa:t, then for any Pareto-

optimal solution i* = (x*,z*) of «m,f),X) problem there exists R -
- * * - * ~1
nondominated plan x eE (R~1'X) such that f(x ,z ) = f(x,z ). 0
Thus, provided the proposition 9 condition satisfied, the uniqueness

of equation f (x, z*) = f (x*, z*) solution with respect to x E X guarantees
that x* is stochastically nondominated in the usual sense.
Notes
1. The results stated below can be reformulated with minor

amendments for more general case when x 1 , ••• ,x n characterizes the
welfare of n social groups with arbitrary size of population, or, in
UDM-interpretation, when the probabilities of n states of nature are
not necessarily equals.
2. The similar approach to formalization of the notions of partial

criteria comparative importance is developed in [2].
3. A somewhat different statement of this problem will be described

in no. 4.
4. The general optimality conditions in similar problems are

studied in [7].
5. This proposition and the next ones are given here without proofs
which can be found in [8].
6. We remind in this connection that "risk measurement in von

Neumann - Morgenstern utility function is also applicable to inequality
measurement" [4, p. 252].
7. This condition which seems to be somewhat artificial is

dictated by subsequent proposition and can be deleted provided not all
components of x are identical.
8. This means that (x EX, x ~x' ~O) ===9 x' EX.
9. The problem n(a) is said to be generated by assortment

scalarization widely used in analysis of multicriteria economic
220
models [17].
10. It is fulfilled automatically if the system of equations

m(x) = m(x*), f(x) = f(x*) has a unique solution in X.
11. For the inequality measure (17 ) with r =2 this fact is

mentioned in [4].
References
1. Hazen G.B. Preference convex unanimity in multiple criteria

decision making. Math. of Operations Research, v. 8, n.4,1983, pp. 505-
516.
2. Marshall A.W. and Olkin I. Inequalities: Theory of Majorization

and its Applications. Academic Press, N.Y. a.o., 1979.
3. Podinovsky V.V. Axiomatic solution of the problem of the

criteria importance assesment in multicriteria problems (in Russian).
In: Sovremennoe sostojanie teorii issledovania operatsij. Nauka,
lo1os cow, 1 979, pp. 117 - 1 49 •
4. Sawaragi Y., Nakayama H. and Tanino T. Theory of Multiobjective

Optimization. Academic Press, Orlando a.o., 1985.
5. Sen A. On Economic Inequality. Clarendon Press, Oxford, 1973.
6. Yu P.L. Cone convexity, cone extreme points, and uondominated

solutions in decision problems with multiple objectives. J. of
Optimization Theory and Applications, v. 14, n. 3, 1974, pp. 319 -377.
7. Hazen G.B. and Morin L.T. Optimality conditions in nonconical

multiple-objective programming. J. of Optimization Theory and
Applications, v. 40, n. 1, 1983, pp. 25-60.
8. Polishchuck L.I. Effective equity-efficiency compromises (in

Russian). In: Modelirovanie i analis ekonomicheskih processov. Nauka,
Novosibirsk, 1985, pp. 40-55.
9. Keeney R.L. and Raiffa H. Decisions with Multiple Objectives:

Preferences and Value Tradeoffs. John Wiley & Sons, N.Y. a.o., 1976.
10. Podinovsky V.V., Noghin V.D. Pareto-optimal Solutions of

Multicriteria Problems (in Russian). Nauka, MOscow, 1982.
11. Geoffrion A.M. Proper efficiency and the theory of vector

maximization. J. of Mathematical Analysis and Applications, v. 22,
n. 3, 1968, pp. 618 - 630.
12. Borwein J. Proper efficient points for maximization with

respect to cones. SI~~ J. of Control and Optimization, v. 15, n. 1,
1 977, pp • 57 - 6 3 •
221
13. Sato R. Theory of Technical Change and Economic Invariance.

Application of Lie Groups. Academic Press, N.Y. a.o., 1981.
14. Ruys P.H.M. and Weddepohl H.N. Economic theory and duality. In:
Convex Analysis and Mathematical Economics. Springer-Verlag, Berlin
a.o., 1979, pp. 1-72.
15. White D.J. Optimality and Efficiency. John Wiley & Sons,
Chichester a.o., 1982.
16. Arrow K.J., Barankin E.N'. and Blackwell D. Admissible points
of convex sets. In: Contributions to the Theory of Games, v. 2.
Princeton Univ. Press, Princeton, 1953.
17. Polishchuck L. I. Models of ~!ul ticri teria Optimization of
Economic Systems (in Russian). Novosibirsk State Univ. Press,
Novosibirsk, 1984.
18. Bakhtin A.E., Polishchuck L.L, Sokolov V.G. Multiobjective
optimization of planned decisions with regard for reliability
properties (in Russian). In: Osnovnie napravlenia matematicheskogo
modelirovania economicheskih processov. Central Economic-Mathematical
Inst., Moscow, 1985.
SOME CONSIDERATIONS ABOUT COMPUTATIONAL COMPLEXITY
FOR MULTI OBJECTIVE COMBINATORIAL PROBLEMS
Paolo Serafini
University of Udine
Department of Mathematics and Computer Science
Via Zanon 6, 1-33100 Udine
1 INTRODUCTION
In the field of vector (or multi objectivel optimization there has been a relatively little
interest in solving combinatorial or discrete problems. During the 70's just a few papers
have been published on multi objective (m.o.) integer linear programming. However no
special emphasis was put on the important aspect of computational complexity. This can
be certainly ascribed to the fact that the theory of NP-completeness was developing at a
fast pace in those same years.
A pioneering paper by P. Hansen 161 appeared in 1980; he attacked the shortest path
problem with two objectives (actually several versions of it) with a truly computational
attitude. Following that paper only a few papers appeared adopting the same perspective,
almost always on shortest path problems (17,17,13,14,81). Other papers on m.o. integer
programming can also be quoted (11,21,4,9,16,31).
From all these results one receives the impression that the computational complexity
of m.o. problems has not been formulated yet in precise terms in spite of the fact that real
problems naturally expressed as combinatorial m.o. problems are rather common. Just to
quote an example consider problems defined on graphs with costs and failure probabilities
associated to the graph components, and requirements of finding a certain sub graph with
low cost and high reliability.
It is also important to define precisely the basic concepts in order to increase the
'awareness' of anyone facing a m.o. combinatorial problem, as it has already happened in
scalar optimization after the development of the theory of NP-completeness.
Hence the aim of this paper is at making some preliminary considerations which can
be helpful in facing m.o. combinatorial problems with a computational perspective.
223
:4 GENERAL CONSIDERATIONS
We shall be concerned with the following problem:
given I:X _ Zm
"find" minimal elements in the subset F f X
An element z e F is minimal (or nondominated) if, for no 11 e F, 1(11) $ I{z).

As usually the relations a < b and a ~ b between vectors are intended componentwise,
whereas a $ b is intended ai $ bi for each component i with a:F b. If z is minimal, I(z) is
called a minimal (or nondominated) value. In general the aim of the problem may consist
in finding just one, or some, or even all minimal elements. Hence the exact meaning of the
term "find" is not specified yet and this accounts for the inverted commas in the problem
statement. The objective function I has as codomain integers instead of reals, as it is
usually assumed for computational complexity purposes (see [5,15]). The domain of I is
a finite set X which contains the feasible set F.
What characterizes a particular instance is the pair (I, F). As usually it is assumed
that each instance (I, F) is implicit ely defined by a string L. "Implicitely" means that
there exist two algorithms A! and AF which, after receiving as input the encoding L of
the instance and an element z e X, output in polynomial time the value I(z) and the
truth value of z e F respectively.
We are interested in problems for which the cardinality of F is very large, in the sense
that:
max{lFI:F defined by L with ILl = n} = n(cn ) for some c > 1
(refer to [5,15] for the notations 0(·),0(·) and E>(.)).
This assumption rules out the possibility of explicit ely enumerating the feasible solu-
tions to find minimal elements (when this is indeed possible see [101).
Since F is finite an upper bound K such that
II/(z) -/(11)1100 < K 'tr/z,lI e F
always exists and sometimes can be easily computed. In case the number K enters a
complexity measure, it is useful to remember that ILl = 0(log2 K) and hence K = 0(2ILI).
One consideration seems to be particularly important: the number m of objectives
enters the encoding L and it is likely to be present in complexity measures. It is crucial
to note that, in contrast to the asymptotic meaning of a complexity measure, m is never
too large, least of all tends to infinity. Of course we are speaking of those m.o. problems
with truly incomparable objectives and not, for instance, of those problems facing the
minimization of a vector in R m as a m.o. minimization on all components, in which case
m can be rather large. Therefore we will consider m as a fixed parameter.
Now the original problem will be particularized into a list of several versions, each one
ofthem giving a precise meaning to the term "find". Our aim is at examining these versions
and at possibly selecting one of them as a standard version to measure complexity. This
list comprises nine versions and by no means pretends to be exhaustive. Other significant
versions could be envisaged and attached to the list.
Given any instance (I, F)
Vi : find all minimal elements;
224
V2 : find all minimal values and at least one minimal element for each minimal value;
Va : find all minimal values;
V. : find at least p > 1 minimal values or, in case p minimal values do not exist, report
this fact and find all minimal values. Furthermore find at least one minimal element
for each found minimal value;
Vs : find at least p > 1 minimal values or, in case p minimal values do not exist, report
this fact and find all minimal values;
V6 : find one minimal element;
V7 : find one minimal value;
Va : given z e zm, does there exist z e F: /(z) ~ z?;
Vg : given z, we zm, W > 0, does there exist z e F: w/(z) ::; wz?
Vl is sometimes regarded as the genuine m.o. problem and consequently some state-
ments about computational complexity have been derived with respect to this version for
particUlar problems (for the Shortest Path problem see [6,8]). As a matter of fact this is
always an intractable problem: just consider a constant objective function and Vl calls
for a complete enumeration of the feasible set F. However, and this is a point apparently
missed by the above quoted authors, Vl is not the m.o. counterpart of a scalar opti-
mization problem, which requires the computation of only one minimal element among
possibly many minimal elements, all of them with the same objective function value. In
this sense it is V2 the natural counterpart of a scalar optimization problem.
Quite often it is realized that the computational effort demanded by V2 is too high
and so one is just trying to solve V4 or even V6 • In particular, the parameter p in V4
should be regarded as part of the instance description. Theoretically no a priori upper
bound exists for p and this makes V4 not so different from V2 (as we shall see soon).
Va, Vs and V7 are usual evaluation versions.
The remaining Va and V9 are recognition versions. In particular Va is the exact
counterpart of the recognition version of a scalar problem (see [151), whereas V9 is the
recognition version of a scalarized m.o. problem.
It should be noted that V2 and Va can be intractable as well for the simple reason they
could easily require an exponential amount of time. In fact the number of nondominated
values is 0(Km-l) in the worst case, where K is the bound previously introduced, and
computing all these values requires obviously a proportional amount of time. Therefore,
unless K is a polynomial function of the input size, no algorithm designed to solve V2 or
Va can be efficient. This dependence on the range of the objective function is an intrinsic
feature of multi objective problems. In view of these observations we shall assume that
the bound K is a polynomial function of the input string length ILl.
Then it is possible to prove some polynomial reductions between some versions:
Va oc V2 , Vs oc V., V7 oc V6, V4 oc V2 , Vs oc Va, V6 oc V2 , V7 oc Va, V6 oc V41 V7 oc Vs ,
Vg oc Va: easy to show;
V6 oc V7 : this is indeed possible in most combinatorial problems;
V2 oc Va, V4 oc Vs : as above;
V7 oc V9 : by binary search;
V2 oc Va : consider a box with side lenghts K containing all feasible values. Then by binary
search along one coordinate, keeping fixed the other coordinates (as in the constraint
method) it is possible to determine a nondominated value. Hence O(K m - 1 log2 K) calls
of Va are required in the worst case. Recall that this reduction is in any case of complexity
O(Km-l);
225
Vs oc V2 : it is necessary to examine all minimal values; hence complexity O(Km-l) in

the worst case.
V. oc Vs : as previously it is possible to determine one minimal elements with O(log2 K)
calls of Vs via the constraint method. Once this solution has been obtained the box with
side lenghts K has been devided into 2m smaller boxes. Other nondominated solutions
have to be found in anyone of 2m - 2 of them. Hence the constraint method has to
be invoked a certain number of times until p solutions are found. In the worst case this
number can be as high as Km-l; just think of a problem with only one nondominated
solution (like an utopia solution); once this solution is found, then the stated amount of
time is simply spent to realize that the solution is unique.
V2 oc V. : just take p = Km-l;
Note that no polynomial reduction Vs oc V9 is expected unless P=NP (see later). Note
also that V2 , Vs , V., Vs and Vs are equivalent problems, within the assumption on the
bound K. It is perhaps surprising that V. is equivalent to V2 • As a matter offact we are
considering worst case complexity and an unbounded parameter p. In a normal situation
one is to expect that V. can be solved with just plog2 K calls of Vs. One could also be
tempted of solving V. with a certain number of calls (hopefully exactly p) ofV9 by varying
the parameter w. This certainly makes things easy from the computational point of view.
However the 'quality' of the solutions can be rather poor, because solutions 'hidden' in
the convex hull of J(F) can never be discovered this way. In fact, if by repetitive calls of
V9 it is not possible to find more than q different solutions, this does not mean at all that
the number of nondominated solutions is q.
As a conclusion it is suggested to take Vs as the standard reference version to measure
computational complexity for m.o. combinatorial problems. This is indeed justified by
the following reasons
1) Vs is the m.o. counterpart of a scalar recognition version to which complexity
considerations usually refer.
2) With the exception of VI, which is always intractable and also not really inter-
esting, all versions can be reduced to Vs. This makes Vs highly representative from a
computational point of view. Of course these reductions suffer from the dependence on
the bound K, which is also affected by the presence of fast growing coefficients with re-
spect to the number of objectives. However, as explained previously this is an inherent
difficulty of m.o. problems.
3) This reason is more subtle. It has been perhaps noted that the concept of a decision
maker's utility function has not been yet introduced. Actually, either the utility function
is known and so the problem is immediately reduced to a scalar one, or it has to be inferred
through some interaction with the decision maker. This interaction is then an intrinsic
part of the computational process and should be precisely defined in order to measure the
worst case complexity of the overall process.
It should be emphasized that a priori hypotheses on the decision maker's behaviour,
like consistency or convexity of his underlying utility function etc., can be assumed only
if it is possible to verify them through a computational effort whose complexity must be
eventually taken into account. Otherwise the decision maker should be regarded as a
random function taking values in some preassigned range.
In our opinion the new problem originated in this way, although interesting in its
own, is a step farther from the original m.o. problem and this is why the concept of utility
function has not been used here. Hoewever, and so we come finally to the third reason,
226
if the approach to a m.o. problem does call for some interaction with the decision maker,
it is sensible to expect V8 playing a fundamental role in soliciting the decision maker's
preferences.
Hence Irom now on, when speaking 01 a m.o. optimization problem, we wall alwa1ls
reler to Va.
Of course a problem known to be NP-complete when m = 1 will remain NP-complete
for m > 1, if, as it usually happens, it will still be in the class NP. The unfortunate fact is
that most polynomial single objective problems become NP-complete for m > 1. This will
be matter of investigation in the next sections. In particular we shall investigate the role
of total unimodularity and matroidal structure which are the most common structures
leading to polynomial algorithms for single objective problems.
3 TOTAL UNIMODULARITY
In some scalar problems it may happen that
V(P) ~ F ~ P (1)
where V (P) is the set ofvertices of a polyhedron P defined by a set of linear inequalities.
Then there is no difference in solving either the combinatorial Problem A:
min ex
s.t. x E F
or the linear programming Problem B:
min ex
s.t. x E P.
For instance, if F is a subset of zn, (1) holds iff the linear inequalities defining P
form a totally unimodular matrix with arbitrary integers as r.h.s. coefficients [20].
At first glance it would seem that the inclusions (1) benefit m.o. problems as well,
since they refer to constraints only. However this is not the case.
Let I(x) = Ox and let Q be the conical hull of the rows of OJ let Q* be the negative
dual of Q (i.e. Q* := {x : xy ~ 0 Tty E Q}). Then x is minimal for Problem A if and only
if (x + Q*) n F = x. Similarly x is minimal for Prolem B if and only if (x + Q*) n P = x.
The latter condition is stronger in general if Q* is not a halfspace, that is if there are
more objectives. Consider the following simple example: F = {0,1?, Cl = (-l,S) and
C2 = (1, -4). Then Q* is the conical hull of (S,l) and (4,1). It is easily seen that the
origin is nondominated for Problem A and dominated for Problem B.
Another property of scalar linear programming inherited by some combinatorial prob-
lems, i.e. the fact that neighbourhoods of adjacent vertices are exact (local minima are
necessarily global minima) no longer holds. Consider a simple example with F = {O, I?,
Cl = (-2,1) and C2 = (1, -2). Then Q* is the conical hull of (1,2) and (2,1) and so the
origin is locally but not globally optimal.
These considerations have a serious impact on the type of properties and results
needed when dealing with more objectives. In fact a current line of research consists in
227
characterizing the polyhedron of feasible solutions (as done by Edmonds for the match-
ing problem or by other authors for the Travelling Salesman Problem [12]) so that the
inclusions (1) hold, at least near the optimum. But, for m.o. objective problems, this
is a useless effort, as explained by the previous considerations and also by the following
general statement.
Consider the m.o. problem
"min" Cz
s.t. z e F
where V(P) f;; F f;; P and P is a polyhedron assigned through linear inequalities.
Then this problem is NP-complete, since it has binary knapsack as a particular case:
c1z ~ Kl ( objective)
-r:'z ~ -K2 ( constraint )
ze{O,I}'" ( F = V(P) with P the unit cube)
Note that
"min" Cz
s.t. z e V(P)
is NP-complete as well and this result casts a shadow on the idea of looking for minimal
elements of m.o. linear programming among the vertices of the feasible polyhedron, an
idea which has received a considerable attention in the past.
Since the above statement is a general one, the question arises whether particular
problems exploiting total unimodularity for their solution, like Shortest Path and Assign-
ment, are still difficult. The negative answer is that both are NP-complete for m > 1,
even with nonnegative costs.
We shall limit ourselves to consider first the Biobjective Shortest Path and then the
Biobjective Assignment, both with nonnegative costs. As mentioned in the introduction
the Biobjective Shortest Path has been investigated rather extensively ([6,8,14]).
Biobjective Shortest Path is equivalent to Constrained Shortest Path; as the latter
is claimed to be NP-complete ([51 p. 214) Biobjective Shortest Path is NP-complete as
well. However the reference provided in [51 is a private communication and so it may be
interesting to give here a simple proof of this fact.
The proof goes through a transformation from the following Binary Knapsack:
"given c1 ~ 0, c2 ~ 0, K 11 K 2, does there exists z e {O, l}n such that c1 z ~ K 1 and
c2z ~ K2 7"
Consider now the network with nodes Vo, ... , V", and a pair of arcs from Vk - 1 to Vk
for k = 1, ... , n with costs (c~, 0) and (O,~) for each pair. Then the question "is there a
path from Vo to Vn such that c1z ~ Kl and c2z :5 Ek c% - K2 7" is equivalent to Binary
Knapsack.
It is immediate to see that the problem is in NP and so it is NP-complete. However
things are not so bad because a pseudopolynomial algorithm is immediately suggested for
it from the previous transformation from a pseudopolynomial problem. The algorithm we
shall describe can be applied to those instances with nonnegative costs for at least one of
the objectives.
228
Algorithm for Biobjective Shortest Path

Let G be the given graph, V(I), ... ,V(n) its nodes, V(I) the source, V(n) the desti-
nation, K- and K+ a lower and an upper bound respectively with respect to the second
objective for all paths V(I) -+ VU) and K = K+ - K-.
Form a new graph G' by copying K times the nodes of G. Let V(i,j) be the nodes
of G' with 1 :5 i :5 nand K- :5 j :5 K+. For any arc (V(h),V(k)) in G there arcs
(V(h,j), V(k,j +c2 (h, k))) for all j : K- - min{0,c2 (h, k)} :5 j :5 K+ - max{0,c2 (h, k)}.
The complexity of constructing G' is O(n 2 K).
Then the problem is to find the shortest paths in G' from V(I,O) to V(n,j) for all
K- :5 j :5 K+. This can be achieved by applying Dijkstra's algorithm with complexity
O(n 2 K 2 ). Let aU) be the shortest path value from V(I, 0) to V (n, j). All minimal values
(vi, v~) can be found in the following way with complexity 0 (K):
j:=K-j
while aU) = +00 do j := j + 1;
(vI. v~) := (aU),j);
i:= 1;
j:= j + 1;
while j :5 K+ do
begin
if aU) < vi
then begin
i:= i + 1;
(vi, v~) := (aU),j);
end
j:= j + 1;
end
Minimal shortest paths can be easily derived from the minimal values and the output
of Dijkstra's algorithm. The overall complexity is O(n 2 K 2 ).
The proof that Biobjective Assignment is NP-complete goes through a transformation
from the following particular case of Partition, which is NP-complete too (15] p. 223):
" given a set A, IAI = 2n, weights w : A -+ Z+, LaEA w( a) = 2W, is there a subset
Be A, IBI = n, such that LaEB w(a) = W ? ".
Let w > maxaEA w(a). Then consider an assignment from A to {I, 2, ... , 2n}; if
a E A is assigned to an odd number this has a cost (w + w(a), w - w(a)), otherwise the
cost is (w, w). Then the question" is there an assignment with cost less or equal to
(2nw + W,2nw - W) ? " is equivalent to the above stated Partition.
Then Biobjective Assignment is NP-complete since it is obviously in NP.
4 MATROIDS
Let (E, M) be a matroid and M' be the family of maximal subsets (see [11]). Let
c : E -+ Z be given costs. It is well known that the problem
min" c(e)
AEM'L.t
eEA
can be easily solved via the greedy algorithm.

229
Let us now consider the biobjective case: «given cost c : E -+ Z2, bound K E Z2 is
there a maximal A E ,M' s.t. EeeA c(e) ~ K 1"
This is clearly an NP problem. Its NP-completeness it easily proved by recalling that
linearly constrained matroidal optimization problems are NP-complete (see !2j) and by
observing that these problems are equivalent to biobjective matroidal optimization.
Besides this result there is another negative fact. Note that E is partially ordered by
the cost and one is tempted to apply the greedy algorithm to any topological sorting of
E in order to find a nondominated solution (recall that a topological sorting of E is any
total order < on E such that c(a) ~ c(b) ~ a < b). Unfortunately the following holds:
Fact 1 : the greedy algorithm applied to a topological sorting of E does not neces-
sarily output a nondominated solution.
Proof: Just consider as a counterexample the graphic matroid for the complete graph
K. with costs C(V1,V2) = (3,0); C(V1,VS) = (0,3); C(V1,V.) = (1,1); C(V2,VS) = (4,4);
C(V2,V.) = (2,2); c(vs,v.) = (2,2). The greedy algorithm applied to the topologi-
cal sorting (V1IV4), (vs,v.), (V2,V.), (V1,VS), (V1IV2), (V2,VS), gives the spanning tree
{(V1, v.)( Vs, V.)(V2, v.)} with total cost (5,5). However this solution is dominated by the
tree {(V1, V.)(V1, V2)(V1, vs)} with cost (4,4). Moreover, for arbitrary costs, the solution
can be arbitrarily bad. •
A converse of the above statement does hold however, and this result can provide
some insight into the problem.
Fact 2 : for any nondominated solution A there exists a topological sorting for which
the greedy algorithm outputs A.
Proof: we may associate to the set E a graph G in the following way: the nodes
correspond to the elements E and an arc exists from ej to ej iff c( ed ~ c( ej ). Obviously G
is a directed acyclic graph. Given a maximal subset A ~ E and its complement B := E\A,
we may form a new graph GA by adding to G the following set of arcs
FA := ((a,b) : a E C(b) n A Vb E B}
where C(b) is the circuit generated by b. Then proving the thesis is equivalent to proving
that G A is acyclic: in fact GA represents a partial order in this case and any topological
sorting of it is the claimed sorting.
Hence let us suppose that GA contains a directed cycle. Since G is acyclic, the cycle
must contain some arc of FA' More precisely, due to the fact that G is transitively closed,
the cycle may be assumed to consist alternatively of one arc of G and one of FA and so it
is of the form
with aj E A and bj E B. We may also suppose without loss of generality that this is the
cycle through a 1 with the fewest arcs.
This last property enables us to obtain a new maximal subset A' by swapping the ai's
with the bi's (that this is indeed possible can be proven by using the same arguments as for
the two matroid intersection algorithm !11], to which this construction is quite similar).
Moreover, since c(bd ~ c(aiH) for i = 1, ... ,k -1 and c(bk ) ~ c(at), A' dominates A
contrarily to our assumption. I
Thus we have identified two families of maximal subsets, namely M*, the family of
maximal nondominated subsets of E, and M+, the family of maximal subsets A for which
230
G A is acyclic (or equivalently there exists a topological sorting of E such that the greedy
algorithm outputs A). Moreover Fact 1 and Fact 2 have proved that the strict inclusion
oM* c oM+ holds.
Let us now consider also the family oM' of maximal subsets A which are locally non-
dominated, in the sense that they are not dominated by any other maximal subset obtained
by swapping an element b f. A with any a e O(b). Obviously if a subset A is locally dom-
inated GA does have a cycle. So the inclusion oM+ ~ oM' is proved. Furthermore the
inclusion is strict as proven by the following example: Let oM be the graphic matroid for
=
the graph (V, E) with V {Vl,V2,VS,V4,V5} and
Let the costs be Cn = (2,2), ClS = (1,5), C2S = (1,1), Cu = (3,3), C15 = (0,4), C45 =
(1,1). Let A = {(V2,VS), (Vl,VS), (VltV4), (V4,V5)}' Then A is locally nondominated but·
GA is not acyclic.
The interesting conclusion is that, even if the concept of maximal subset obtained via
the greedy algorithm applied to some topological sorting of E is weaker than the concept
of nondominated solution, it is anyway stronger than the one of locally nondominated
solution. The reason is that, as apparent from the proof of Fact 2, a larger neighbourhood
is considered by requiring acyclicity of GA. The important fact is that exploration of this
larger neighbourhood is not computationally expensive; in fact it can be carried out with
complexity O(lEI2).
If we consider other ways of ordering the elements of E, no better conclusions can be
drawn. Obviously the greedy algorithm applied to any lexicographic ordering of E does
produce a nondominated solution. However, roughly speaking, these solutions are not the
ones we are looking for, since they are by definition too much biased in favour of one
particular objective.
Let us also consider the ordering of E induced by IIclloo (break any tie via topological
sorting) or by IIc-clioo for some constant c. We recall that minimization of II/(:I:)-cil oo is a
well known scalarization procedure yielding all nondominated solutions by suitably varying
c. Nonetheless the greedy algorithm applied to this ordering of E does not necessarily
output a nondominated solution, as seen by reconsidering the counterexample provided
in Fact 1.
5 OONOLUSIONS
The previous sections showed that in m.o. objective combinatorial problems one is
typically faced with NP-hard problems. Hence it is almost compulsory to attack these
problems through heuristics combined with clever implicit enumeration techniques.
But this is not the only reason. In fact the modelling itself of a real problem through
a m.o. mathematical model is due to a lack of information about the actual preferences of
the decision maker. Therefore the interest in having guaranteed optimal solutions should
shift in favour of having a rather large number of acceptably 'good' solutions.
Moreover the common approach of getting solutions through some kind of interaction
with the decision maker is also to some extent a heuristic procedure.
It is perhaps convenient to bring together these two kinds of solution approximation
into a unique procedure. In other words one should direct the ideas for assessing the
231
decision maker's preferences into the procedure generating solutions. For example, since
generating solutions of a combinatorial problem usually corresponds to visiting a very
large search tree, possible cutoffs of the search tree could be also controlled by heuristic
demands for speeding up the computation and by the interaction with the decision maker.
For linear discrete problems which are polynomial in case of a single objective, it may
be convenient to solve the scalarized version V9 (a weighted linear combination) and then,
once discovered some nondominated solutions on the boundary of the convex hull of I(F),
to build some implicit enumeration scheme to 'peel off' the external strata of I(F) in
order to discover other hidden nondominated solutions. Again this can be done through
some interaction with the decision maker.
This is a research field almost unexplored at the time. A preliminary work in this sense
exploiting duality concepts developed by the author can be found in [18,19]. Certainly this
kind of research would turn more into a patient handcraft work on particular problems
rather than produce general theoretical breakthroughs. However the practical utility of
such an approach should be quite promising and worth the effort.
6 REFERENCES
(1) Bitran,G.R., Theory and algorithm. for linear multiple objective program, with zero-one vari·
able" Math. Progr. 17 (1979) 362-390.
(2) Camerini,P.M., and Vercellis,C., The mafroidal knap.ack: a claBB of (often) well·.olvable
problemr, Operations Research Letters 3 (1984) 157·162.
(3) Chalmet,L.G., Lemonidis,L., and Elzinga,D.J., An algorithm for the bicriterion integer pro·
gramming problem, European J. of Operations Resear (1986) 292·300.
(4) Dyer,M.E., and WaIker,J., A note on bicriterion programming, Math. Progr. 27 (1983)
355-361.
(5) Garey,M.R., and Johnson,D.S., Oomputer. and intractability, Freeman, San Francisco, USA,
1979.
(6) Hansen,P., Bicriterion path problemr, in Multiple criteria deciBion making, Fandel,G., and
Gal,T., editors, 109-127, Springer, Berlin, 1980.
(7) Hartley,R., Survey of algorithm. for vector optimi.ation problemr, in Multi.objective deci.ion
making, French,S., et aI., editors, Academic Pw York, 1983.
(8] Henig,M.I., The ,horte,t path problem with two objective function" European J. of Operations
Research 25 (1985) 281·291.
(9) Ignizio,J.P., Multiobjective mathematical programming via the Multiples model and algorithm,
European J. of Operations Research 22 (1985) 338·346.
(10) Kung,H.T., Luccio,F., and Preparata,F., On finding the ma:tima of a .et of vector" J. of
ACM 22 (1975) 469-476.
(11) Lawler,E., Combinatorial optimization: network, and matroid., Holt, Rinehart and Winston,
New York, 1976.
(12) Lawler,E., Lenstra,J.K., Rinnooy Kan,A.H.G. and Shmoys,D.B., The traveling .aleBman
problem, Wiley, Chichester, UK, 1985.
(13) Martins,E.Q.V., On a multicriteria .horte.t path problem, European J. of Operations Research
16 (1984) 236-245.
232
(14] Martins,E.Q.V., On a .pecial clu. 01 bicriterion path problem., European J. of Operations

Research 17 (1984) 85-94.
(15] Papadimitriou,C.H., and Steiglitz,K., Oombinatorial optimization: algorithm. and compleat"
Prentice Hall, 1982.
(16] Ramesh,R., Karwan,M.H., and Zionts,S., A clu/J 01 practical interactive branch and bound
algorithm. lor multicriteria integer programming, Working paper 688 (1985), School of Man-
agement, State University of New York at Buffalo.
(17] Ruiz-Diaz,F., and French,S., Multi objective combinational /Jcheduling, in Multi-objective de-
ci.ion malcing, French,S., et a1., editors, Academic Press, New York, 1988.
(18] Serafini,P., Dual relasation and branch-and-bound technique/J lor multi objective optimiza-
tion, in Interactive decision anali., Grauer,M., and Wierzbicki,A.P., editors, 84-90, Springer,
Berlin, 1984.
(19] Serafini,P., A unified approach lor /Jcalar and vector optimization, in Mathematic/J lor multi
objective optimization, Seranni,P., editor, Springer, Vienna, 1985.
(20] Veinott,A.F., and Dantzig,G.B., Integral e:rireme point., SIAM Rev. 10 (1968) 871-872.
(21] Villareal,B., and Ka.rwa.n,M.H., Multicriteria integer programming: a (hybrid) dynamic pro-
gramming recur/Jive approach, Math. Progr. 21 (1981) 204-223.
ON THE EXISTENCE OF <X>NE-EFFICIENT POINTS
Alicia Sterna-Karwat
Department of Mathematics, Monash University
Clayton 3168, Australia
1. Introduction
In multicriteria optimization the decision-maker's way of making a choice is

represented by some binary relation, called preference, defined over a set of
one's alternatives. In this paper we shall deal with the case in which the set A
of alternatives is a subset of a real Hausdorff topological vector space (t.v.s.)
X and the preference relation is defined through a convex cone C £ X (i.e. C
is nonempty, convex and tC £ C for t ~ 0). We say that an alternative e € A
is efficient if there is no other alternative x € A which is preferred to e
with respect to the given preference [14] (see Definition 1.1.below). Our
attention will be focussed on the existence of efficient points in the above
setting. References [7,14] provide an extensive bibliography on this topic. We
quote only the papers which are explicitly used in this note. For a study of the
existence of efficient alternatives with respect to other preference relations the
reader is referred, for example, to [4, 13, 15].
Throughout the paper, mn, A\B, A, 2(C) = C n(-C) denote n-dimensional

Euclidean space, the set-theoretic difference of subsets, the closure of a subset
A and the largest vector space contained in a convex cone C, respectively. The
cone C is called pointed if 2(C) = {O}.
1.1 Definition. Let Xl' ~ € X. We say that Xl is preferred to ~
if Xl - ~ € c\2(C).
Thus, according to the definition of efficiency, if A £ X, e € A is
efficient in A if X € A and X - e € C together imply e - X € C [2,3]. The
set of the efficient points in A will be denoted by E(A;C). Such points are
described in [1,10-12] as maximal up to indifference.
In 1974 Yu [16] showed that in mn , if C is pointed and A is nonempty

and compact, then E(A;C) is nonempty. Next, in 1978, Hartley [3] proved that in
mn , E(A;C) is nonempty for every convex cone C and for every C-compact subset
A (C-compact means that there exists x € X such that A ncx + C) is nonempty
and compact). In 1980 Corley [2] extended the result of Yu to infinite
234
dimensions for subsets satisfying a compactness-like condition similar to

C-compactness. In 1983 Borwein [1] proved that E(A;C) ¢ $ in the infinite
dimensional case if C is closed and A is C-compact. The interest of the
present paper lies in the fact that the convex cone C is not supposed to be
closed or that C\{O} is open. The only result previously known. apart from
Corley·s. which covered the existence of efficient points for nonclosed cones was
that which assumed existence of a C-strictly positive. continuous functional f
on X. i.e. f(x) ~ 0 for x € C and f(x) >0 if x € C\t(C) [1.5.7.14]. The
above case covers each convex cone C with C\{O} open. The class ~ defined
by the author in [10] is the largest possible to ensure the existence of the
efficient points (see Definition 1.2 below).
The general decomposition construction given in Theorem 2.3 provides new
insight into the problem on the existence of efficient points. Moreover it
provides better understanding of the structure of convex cones in topological
vector spaces. Indeed. having a preference relation associated with a convex cone
C either (i) we are sure that the efficient points with respect to this
relation. exist in every compact set. or. (ii) we can remove some part C1 of C
such that the remaining part C2 ~ C is still a convex· cone and satisfies (i).
Moreover. the removed part C1 is not a vector subspace but its closure is a
vector subspace. This somewhat pathological part of C prevents the existence of

efficient points in some compact sets in the metrizable setting (see Theorem 3.2
below).
1.2 Definition. We say that a convex cone C belongs to the class ~ if the
following condition holds: for every vector subspace L contained in X. C nL
is a vector subspace whenever C n L is a vector subspace.
The following results are proved in [10].
1.3 Theorem (sufficient condition). Let C be a convex cone in a t.v.s. X.

If C € ~ then E(A;C) ¢ $ for every C-compact subset A of X.
1.4 Theorem (necessary and sufficient condition). Let C be a convex cone in

a t.v.s. X such that t(C) is a metrizable vector subspace. Then E(A;C} ¢ $
for every nonempty compact subset A of X if and only if C € ~.
The assumption that t(C) is metrizable is essential in the above theorem

(see Example 1.1 in [12]). The above results cannot be extended either to closed
bounded sets or nonconvex cones [1.2].
More details on the existence of efficient points are given in Section 3.

235
Now we turn our attention toa study of the class ~ .
2. On the structure of convex cones
The proof of the following proposition can be partially found in [10].

and is not difficult to verify using separation arguments.
2.1 Proposition. Let C be a convex cone in a t.V.S. X. Then C € ~
whenever one of the following conditions holds:

(i) X is finite dimensional;
or more generally.
(ii) the vector subspace 2(C) is finite dimensional; in particular if C
is pointed;
(iii) C is closed;
(iv) C ~ {O} is open;
or more generally.
(v) C admits a continuous C-strictly positive functional on X.
The ubiquitous cone of Klee [6.9] is an example of a proper convex cone C

such that C = X; thus C (~. Moreover. in every infinite dimensional normed
space X. there exist cones with nonempty interiors and yet which are not members
of ~ [10].
The class ~ can be characterized equivalently either by relaxing the

closedness assumption imposed on L in Definition 1.2. or by assuming that
L ~ 2(C) [11].
Let us observe the follOWing useful criterion ensuring that C € ~.
2.2 Proposition. Let C be a convex cone in a t.V.s. X such that
2(C) = C n 2(C). Then C € ~.
Proof. Suppose that 2(C) = C n 2(C) and C (~. Then there exist a vector
subspace L of X and x € C nL such that C nL is a vector subpsace and
-x (C. Since -x € C nL ~ C. we must have that x € C n 2(C) = 2(C). which

contradicts -x ( C. D
The converse of Proposition 2.2 is not true (see the last paragraph in this
section).
The following theorem proved in [11] shows that an arbitrary convex cone can
be uniquely decomposed as a union (hence as a sum) of a convex cone from the class
~ and a convex cone whose closure is a vector subspace. which meet at the origin.
236
2.3 Theorem. Let C be a convex cone in a t.v.s. X. Then there exist

convex cones K(C) and p(C) in X such that
(i) C = K(C) U p(C) (in particular C = K(C) + p(C»;
(ii) p(C) belongs to ~ and is pointed;
(iii) K(C) is a vector subspace;
(iv) K(C) n p(C) = {O}.

Moreover. conditions (i) - (iv) determine K(C) and p(C) uniquely.
We shall recall briefly the construction which leads to the above

decomposition.
Let ~ be the set of all convex cones in X. Define the operation
T : ~ ~~ by T(C) = C n tee)
and its transfinite iteration. for a an ordinal
number. by ~(C) =
T(~-l(C» if a-1 exists. and ~(C) n ~(C) if a is =
~<a
a limit ordinal. Since {~(C)} is a chain with respect to inclusion of subsets
of X we must have an ordinal a(C) such that ~(C)(C) = ~(C) for every
a ~ a(C). Put K(C) =~(C)(C). Lemma 2.1 in [11] ensures that K(C) is a
vector subspace and that K(C) = C n K(C). Moreover. C, K(C) is a convex

subset (Lemma 2.2 [11]). obviously closed under positive scalar multiplication.
and we put p(C) = (C , K(C» U {O}.
We shall show the second part of Theorem 2.3 after the following observation.
Observe that t(C) £ K(C) and if K(C) is a vector space then it must be
that K(C) = t(C). We can give the following characterization of the class ~.
2.4 Theorem [11]. Let C be a convex cone in a t.v.s. X. Then C € ~ if

and only if K(C) = t(C).
The proof of the uniqueness of the decomposition in Theorem 2.3. Let C1
and C2 be convex cones such that (i) C = C1 U C2 (ii) C2 € ~ and is pointed
(iii) C1 is a vector space (iv) C1 n C2 = {O}. Since C1 is a vector
subspace we obtain that t(C) = C1 U t(C2 ). which is only possible if either
T(C) = C n t(C) = C n C1 = C1 • because C1 n C2 = {O}. Since C1 is a vector
subspace we have that T(C) = K(C) = C1 .
transfinite induction. and that C1 is a vector subspace and C1 n C2 = {O}. one

237
can show that fl(c) = C1 U fl(C2 ) for every ordinal a. Thus

K(C) = C1 U K(C2 ). Since C2 € ~ and is pointed we have, by Theorem 2.4, that
K(C) = C1 . Consequently, using (i) and (iv) for both decompositions, one obtains
that p(C) = C2 ' which establishes the uniqueness. C
Observe that operations K, p : ~ 4~, assigning K(C) and p(C)

respectively to every C € ~, are orthogonal idempotents, that is, K2 = K, p2 =P
and K(p(C» = p(K(C» = {O} for every C €~. Moreover, Theorem 2.2 says that
K + P = id:1/.
2.5 Theorem. The class ~ is closed under arbitrary intersections.
Proof. Let {Ci}i € I be a family of convex cones in X such that Ci € ~
for every i € I. Put C - n Ci . Then

i € I
T(C) = C n iCC) £ n Ci n i(C.) = n T(C i ).

i € I 1 i € I
One can show that that fl(C) £ n fl(C i ) for every ordinal a. Hence, by the
i € I
definition of the operation K and by Theorem 2.4 we obtain that
K(C) £ n K(C.) = n i(C.). On the other hand
i € I 1 i € I 1
iCC) £ K(C) £ n i(C l.) £ iCC). Hence K(C) = i(C), which again by Theorem 2.4,
i € I
means that C € '€. []
Using Theorem 2.5 and Proposition 2.1 (iv) we obtain
2.6 Corollary. If C is a convex cone such that C\{O} is evenly convex (i.e.
C\{O} is an intersection of open half spaces) then C € '€.
Let us observe here that, although the existence of continuous C-strictly

positive functionals on X implies C € ,€, no such conclusion can be drawn from
the existence of nontrivial continuous C-positive functionals on X (i.e.,
functionals f on X such that f ¢ 0 and f{x) ~ 0 for every x € C). If X
is a locally convex t.v.s. and C is a convex cone in X then for every
Xo € C 'i{C) there exists a continuous C-positive functional such that
f{xO) > O. However, the following condition implies C € '€. Let us denote by
C+ the set of continuous C-positive functionals on X.
2.7 Remark. Let C be a convex cone in a t.v.s. X satisfying the condition

(+) : For all x € C 'i(C) there exists f € C+ such that f{x) > O. Then
C € '8.
238
The proof follows from Proposition 2.2 since the condition (+) implies
iCC} = C n iCC}. D
Moreover, let us observe here, that in a locally convex space X the

condition (+) is equivalent to iCC} = C n iCC}.
We say that a convex cone C has a base B if B is convex, 0 f. B and
C = U tB. If X is locally convex and C has a base then X admits f € C+
QO
which is C-strictly positive; thus C € ~. The next proposition shows that in a
t.v.s. X, C € ~ whenever C has a base.
2.8 Proposition. Let C be a convex cone in a t.v.s. X with a base B.

Then C€ ~.
The proof follows from the fact that if K = UtB then K n i(K} {O} and
t~O
iCC} ~ C n iCC} ~ K n i(K} = {O}. Hence iCC) = C n iCC} and applying

Proposition 2.2 we obtain that C € ~. D
Moreover, let us observe that for (i) closed cones, (ii) cones satisfying
condition (+) of Proposition 2.7 (in particular for cones for which continuous
C-strictly positive functionals exist), (iii) cones with bases, (iv) cones whose
closures are pointed, the transfinite sequence {~(C}} stops at the first place,
i.e. K(C} = T(C} = C n iCC} = iCC). However, in Rn , the convex cone
C = Cl U ... ~,where k ~ n and
Ci = {(xl,···,xn ) : Xl - ... = x i - l = 0, Xi > O} U {O} for i = l, ... ,k,

k
is such that a(C} =k and K(C} =T (C) = {O}. A similar construction in
infinite-dimensional vector spaces endowed with the strongest locally convex
topology [8] gives cones with arbitrary large a(C).
3. More on the existence of the efficient points
Observe first that if X is a locally convex t.v.s. and C € ~ then C

also belongs to ~ in the weak topology of X. Moreover, we say that a subset
A of X is C-weakly compact if A n (x+C) is nonempty and weakly compact for
some x € X.
3.1. Theorem [12]. Let X be a locally convex t.v.s. and let C € ~. Then
239
(i) E{A;C) # ~ for every C-weakly compact subset A of X;

(ii) E{A;C) n ext{A) # ~ for every nonempty convex, weakly compact subset
A of X, where ext{A) is the set of extreme points in A.
The existence of efficient points which are extreme in weakly compact convex
subsets was observed by Borwein [1] for closed convex cones. However, if X is
the space of sequences x = (xn ) with at most a finite number of non-zero terms,
wi th the norm IIxll = }; IXn I. then for every convex corie C and for every convex
compact subset A of X, E{A;C) # ~; yet if C (~ there exists a compact set A

such that E{A;C) =~ [12].
Using Theorems 2.3 and 2.4 we can show
3.2 Theorem [11]. Let X be a t.v.s., let C £ X be a convex cone and let
A £ X. Then
(i) E{A;C) = E{A;K{C» n E{A;p{C».
(ii) If C € ~, then E{A;C) = E{A; p{C».
(iii) If K{C) is metrizable (in particular if X is metrizable), then

C € ~ if and only if E{A;C) = E{A;p{C» (compare Theorem 1.4).
(iv) Let M be a vector subspace of X such that K(C) n M is finite
dimensional. Then E{A;C) # ~ whenever A £ M and A is C-compact (compare
Proposition 2.1(ii».
References
[1] J.M. Borwein, On the existence of Pareto efficient points, Maths. Oper.
Res. 8 (1983), 64-73.
[2] H.W. Corley, An existence result for maximization with respect to cones,
J. Optim. Theory Appl. 31 (1980), 277-281.
[3] R. Hartley, On cone-efficiency, cone-convexity and cone-compactness,
SIAM J. Appl. Math. 34 (1978), 211-222.
[4] G.B. Hazen, T.L. Morin, Optimality conditions in nonconical multiple
objective programming, J. Optim. Theory Appl. 40 (1983), 25-60.
[5] M.I. Henig, Existence and characterization of efficient decisions with
respect to cones, Mathematical Programming 23 (1982), 111-116.
[6] R.B. Holmes, Geometric Functional Analysis and Its Applications,
Springer-Verlag, New York, 1975.
[7] J. Jahn, Mathematical Vector Optimization in Partially Ordered Linear
spaces, Peter Lang, Frankfurt am Main, 1986.
[8] J.L. Kelley, I. Namioka et al., Linear Topological Spaces, Princeton,
D. Van Nostrand Co., Inc., 1963.
[9] V.L. Klee, Convex sets in linear spaces, Duke Math. J. 18 (1951),
443-466.
[10] A. Sterna-Karwat, On existence of cone-maximal points in real topological
linear spaces, Israel J. Math. 54 (1986) 33-41.
[11] A note on convex cones in topological vector spaces,
Bull. Austral. Math. Soc. 35 (1987) 97-109.
240
[12] , A note on cone-maximal and extreme points in

topological vector spaces, Numerical Functional Analysis and Optimization
(to appear).
[13] , On a structure of preorder relations on a topological
space, Analysis Paper 53, Monash University (1986).
[14] Y., Sawaragi, H. Nakayama, T. Tanino, Theory of Multiobjective
Optimization, Academic Press, Inc., Orlando, 1985.
[15] N.C. Yannelis, N.D. Prabhakar, Existence of maximal elements and
equilibria in linear topological spaces, J. Math. Econom. 12 (1983),
233-245.
[16] P.L. Yu, Cone convexity, cone extreme points and nondominated solutions
in decision problems with multi-objective, J. Optim. Theory Appl. 14
(1974), 319-377.
Acknowledgement.
The author would like to thank Jean-Paul Penot for his comments which improved
the presentation of the paper.
Current address.
Department of Mathematics, The University of Newcastle, New South Wales 2308,
Australia.
PSEUDO-UTILITIES
Andrzej Wieczorek
Institute of Computer Science
P.O. Box 22, 00-901 Warsaw
Abstract. A pseudo-utility representation of a binary relation> in a

set X is a function p:x2 ~ R with p(x,y) > 0 equivalent to x > y. The
concept seems to have appeared in many different contexts in mathema-
tical literature but it still requires systematic studies within the
framework of game or utility theory.
The paper contains basic definitions and a survey of already exist-

ing results including author-s own. We try to present a uniform theory
also exhibiting the relevance of the concept in game theory where it
occurs to be an indispensable tool for proving the existence of equi-
libria in certain types of games.
The paper presents the concept of "parametric pseudo-utility".Given

a set of "parameters" T and a family of binary relations (>t1tET),
respectively in sets X(t), we search for a real function 11 ("paramet-
ric pseudo-utility representation") defined over {(t,x,y) ItET,x,yEX(t)}
with some "good" properties (continuity, quasi-concavity etc.) and
such that for all t,x,y, 11 t (X,y»O is equivalent to x>t Y ' In Sec-
tion III we formulate some results extending earlier results of the
author (however, no proofs are given in this paper). In Section IV we
prove the relevance of the problem in game theory: the existence of a
"good" parametric pseudo-utility representation is very useful when
proving the existence of equilibria in noncooperative games.
In turn, it is also interesting to look back at the simpler case

of just one relation> in a set X and its "pseudo-utility represen-
242
tation". In the literature this case is only considered in connection

with ideas of difference measurement (see Fishburn [1970], Chapter 6
or Krantz et al. [1971], Chapter 4 for a survey). In Section I we give
basic definitions and in Section II we make a survey of the very few
known results, the most interesting of which are due to Kiruta et al.
[1980], Chapter 1.
The relation between the case of a single relation and the "parame-
tric case" resembles, to some extent, the relation between the usual
problem of the existence of a utility representation of a preference
relation and its "measurable" version, the latest considered by Aumann
[1969], Wieczorek [1980] and many others.
I. BASIC DEFINITIONS
Throughout this section> is a fixed binary relation in a set X.

First recall the classic definition of a utility representation of >
it is a function u:X ~ R such that for every x,yEX, x>y is equi-
valent to u(x) - u(y»O.
We define a pseudo-utility representation of > as a function

p:XxX ~ R such that for every x,yEX, x>y is equivalent to p(x,y»O;
thus a pseudo-utility representation of > is a real function defined
over XxX whose support is just equal to the graph of >.
Every utility representati.on u of > generates its pseudo-utility

representation
U(x,y) := u(x) - u(y), for x,yEX;
if the formula above holds for a function U:X.xX ~ R then we say that
U is ~enerated by a utility u.
A pseudo-utility representation of > can be regarded as something

"less" than a pseudo-utility generated by a utility function; this
justifies the use of the term "pseudo-utility" introduced in the pre-
sent paper.
One can immediately see that the usual problem of finding a utility
representation of > can be split, by means of the previously defined
243
notions, into two problems:

1. find a pseudo-utility representation of >;
2, find a utility function generating a function p:X.xX'" R.
Other problems analogous to (2) are also considered in the litera-

ture, for instance, (2.b) and (2.c) are in Chapter 4 of Krantz et ala
[1971]; (2.a) is in Kiruta et al. [1980], p. 53:
2.a. for a given function p:XxX ... R find functions u,v:X ... R such
thatp(x,y) = u(x) - v(y);
2.b. for a given p:XxX R find ... u:X ... R with p(x,y) I u (x) -u (y) I ,
2.c. for a given p:XxX R find ... u:X ... R with p(x,y) u(x)/u(y).
Real functions p defined over XxX and appearing in utility

contexts assume a very natural interpretation: p(x,y) is a measure of
difference in utility between x and y. Typically authors begin with
a binary relation >* in x2 and find a function p:XxX'" R such that
(x,y»*(z,t) is equivalent to p(x,y) - p(z,t) > 0; thus p is a
utility representation of >* in the usual sense. Examples of such
procedures can be found in Chapter 4 of Krantz et ala [1971] (Theorem
1, p. 147). However, more often authors directly seek for a utility u
generating this p, thus having the property that
(x,y»*(z,t) is equivalent to u(x)-u(y»u(z)-u(t)
(see e.g. Chapter 6 of Fishburn [1970] for a survey of results of this

type) •
Another approach takes a function p:XxX'" R, interpreted as measur-

ing differences in utility, as a primitive concept; this is the way
things are presented by Kiruta et ala [1980] in Chapter 1.
The problem (1) is trivial as it stands: the function p:XXX'" R

defined by
3. p(x,y) := if x > y
:= 0 otherwise;
clearly represents >. However, (1-) which is a slight modifi-

cation of (1) may lead to more interesting problems;
1-. find a pseudo-utility representation of > with some "good" pro-
perties.
Here we only notice an immediate consequence of Urysohn-s Lemma:

if the graph of > is an open FE> while X is normal then > has
244
a continuous pseudo-utility representation.
In turn, (2) is generally nontriv.ial, while its modification (2#)

becomes even more interesting:
2~. find it "good" utility function generating a function p:XxX .... R.
Obviously, having passed through (1) and (2) we get a utility re-
presentation of > but notice that, generally, finding a utility re-
presentation of > does not have to go through (1) and (2). The fol-
lowing observation shows that this procedure may be sometimes even
misleading: even if > admits a utility representation u and p is
a pseudo-utility representation of >, p need not be generated by
u nor by any other utility representation of >. A trivial example
of this kind is the usual order > in the real line and its pseudo-
representation p defined by (3).
Scew-symmetric pseudo-utility representations play a special role

and may be interpreted specially. Obviously, every relation with such
a representation must be antisymmetric. In turn, if p is a pseudo-
utility representation of an anti symmetric relation then so is the
function
q(x,y) := p(x,y) - p(y,x)
which is clearly antisymmetric.
II. A FEW KNOWN RESULTS
Also in this section > denotes a fixed binary relation in a set X.

We begin with some immediate observations. Notice that every pseudo-
utility representation p of > immediately generates some others:
the functions
k
sgn p, max(p,O), P , for any positive integer k,
are also pseudo-utility representations of >i if q is another

pseudo-utility representation of > then so are the functions
min(p,q), max(p,q), p+q.

245
If r:XxX ~ R is any function such that for all x,y, x > y

implies r(x,y) > 0 then the function
r~(x,y) := r(x,y).~(x>y)
is a pseudo-utility representation of >.
The book [1980] of Kiruta, Rubinov and Yanovskaya contains a lot of

elementary but very useful results concerning the existence or even in-
dicating a particular form of a pseudo-utility or utility representa-
tion. We just quote some of them «6) and (7) seem to have been in-
spired by Theorem 1, p. 147 in the book [1971] by Krantz et al.).
4. Suppose that X is equipped with a metric. If for every xxX, the

set L(x) := {ylx>y} is open then p:XxX ~ R defined by
p(x,y) := dist (y,X'L(x))
is a pseudo-utility representation of >, continuous in y.
If, moreover, X is separable and, for every x, L(x) is an in-

terior of its closure while U(x) := {yly>x} is Borel then p is
measurable.
5. Suppose that X is a convex set in a vector space. Any two biaffine

pseudo-utility representations of > differ at most by a multiplica-
i;ive constant.
Now suppose that p is a scew-symmetric pseudo-utility representa-

tion of >.
6. If P is negatively additive, i.e. non x>y and non y>z implies

p(x,z) p(x,y) + p(y,z) then it is generated by a utility, unique up
to an additive constant.
7. Suppose that X is a convex set in a vector space while p is bi-

affine. Then p is monotone, i.e. x>y>z and x>z implies
p(x,z) > max (p(x,y),p(y,z)), if and only if [> is negatively transi-
tive and p is generated by an affine utility function].
246
III. PARAMETRIC PSEUDO-UTILITY
Suppose that T is a set (of parameters) 1 for each tET we are

given a set X(t) equipped with a binary relation >t' There arises a
question of the existence of a real function n defined over the set
+ := {(t,x,y) ItET, x,yEX(t)}
which has some "good" required properties. We are specially interested

in continuity of n and quasi-concavity of n(t,x,y) in x, in the
case where T is a topological space while all the X(t) are includ-
ed in a topological vector space X.
The importance of this problem will be exhibited in the next section.
We shall be considering the following properties of a family e of

subsets of a topological space X:
8.a. for every C E t and its neighborhood G there exists DE e ~
~ C c Int D cDc G,
B.b. for every C,C~E e
and every compact set K c Int(CnC~) there
exists D E ~ such that K c Int D cDc C n C~.
We shall also consider a jOint property of families t and (:j of

subsets of a topological space X:
8.c. for every C,C~E C there exists D E t such that CUC~c D
=O{O E ~IC u C~= O}.
Obviously, every family of compact sets e which is closed under

finite intersections, satisfies (8.b).
~hefollowing theorem extends the results included in Theorem A.10,

p. 32, of Wieczorek [1986]:
8. THEOREM. Let T and X be locally compact spaces and let C and

~ be families of subsets of X, we assume that all elements of t ~
compact and that conditions (11.a-c) are satisfied.
For every tET, let X(t) = X and let >t be a binary relation in
X(t) such that the joint graph
~ := {(t,x,y) I x >t y }
is relatively open in
247
t := {(t,x,y) ItET, x,yEX(t)}.
Assume that * there exist sequences: (K i ) of compact subsets of TxX

and (C i ) of elements of e
such that
If for every tET and yEX(t)
Ut(y) := {xEX(t) I x >t y} E ~
then there exists a continuous function ~: i ~ [Oj1] such that for

every tET, ~t("') is a pseudo-utility representation of >t and,
for every yEX(t) and rational qE[Oi1], {xl~t(x,y) ~ q} is an in-
tersection of an element of with X(t). e
Condition (*) requires a very special form of the ,graph of ~.
When is it satisfied automatically? Here is a sufficient condition:
9. PROPOSITION. If T is ~-compact metrizable while X is compact

metrizable and its (unique) uniformity has a closed base whose every
element Vc:XxX has all sections V(x):={yl(x,y)EV} in t then
every open set in TxXxX has the form U{KixCili=1,2, ... } where
all Ki ~ TxX are compact while all Ci are in e (thus, in the
framework of Theorem (8), (*) is satisfied).
We shall also consider a corollary to Theorem (8) which is con-

cerned with special case of C being a class of convex sets in a
topological vector space:
10. COROLLARY. Let T be a locally compact space and let X be a

locally convex Hausdorff convex set in a topological vector space,
satisfying
** for every compact convex set C c: X and its every neighborhood
G there exists a convex compact neighborhood of C included in G.
For every tET, let X(t) c: X be convex and let >t be a binary
relation in X(t) such that the jOint graph
+) Formally, due to the change of the order of factors, KixC i should

be written here as {(t,x,y) I (t,y) EK i , XEC i }.
248
0/ := {(t,x,y) Ix >t y}
is relatively open in
.p := {(t,x,y) ItET, x,yEX(t)}.
Assume that * there exist sequences: (K i ) of compact subsets of TxX
-
and (C.) of elements of
~
e such that
n U{K.xC.li
~ ~
1,2, ... }.
If for every tET and yEX(t) the set Ut(y) :={xEX(t)1 x >t y} is
convex then there exists a continuous function lP : ~ .... [0,1] such that
for every tET, lP(x,y) is a pseudo-utility representation of >t'
quasi-concave in x.
Since (*) has been already discussed, we only notice that (**) holds,
for instance, whenever X is also compact (see A.6,p. 31 in Wieczorek
[1986]).
IV. PSEUDO-UTILITIES IN GAME THEORY
In [1952] Debreu wrote a paper being a mathematical preparation for

his famous joint work with Arrow [1954] on the existence of economic
equilibria. The paper dealt with a conflict situation generalizing the
n-person non-cooperative game scheme. Slightly reformulated, this
situation looks as follows:
given a positive integer n (number of players), non-empty sets

S1""'S
n
(respective players- sets of strategies), relations r.J c
-
n.
~
s.~ x S., j=1, ••• ,n (players- constraining relations) and functions
J
u. : n. s ..... R (players- utility functions), an equilibrium (I) is a
J ~ ~ *
system of strategies sEn. S. such that for every player j,
* * * ~ ~ *
(s ,s.)E
J
r.J (i.e. s. is available for j at
J
s ) and
u. (s*) = max{u. (s) Is. = s~ for all ih and (s*,SJ.)E f J.}

J J ~ ~
(i.e. j cannot increase his utility by changing the jth coordinate in

s * with accordance to rj ) •
249
Debreu [1952] proved the following theorem generalizing Nash-s

[1950] equilibrium existence theorem:
11. THEOREM (Debreu). An equilibrium I exists whenever for every j:
a. Sj is a contractible polyhedron in a Euclidean space;

b. fj is closed;
c. the function
is continuous;
d. u. is continuous, and
]
e. for every s E nisi the set
Mj (s) := {~I (s,~) E fj and uj(s with Sj replaced by~)
is contractible.
Conditions (c) and -(e) do not look very nicely but they were adopted
by Debreu just for greater generality. Notice that for every j, (b), (d)
and
c-. rj is lower semi-continuous;
imply (c). Since obviously
a-. S. is a compact convex set in a Euclidean space;
]
is a special case of (a), (e) would be, for every j, an immediate con-
sequence of (a-) and
e-. u j is quasi-concave in Sj'
In seventies the research in some areas of mathematical economics

concentrated at models in which utility functions were replaced by more
general preference relations. Here is a counter-part of the previous
equilibrium concept in this setting (the interpretation is as before) :
given a positive integer nr nonempty sets S1""'Sn' relations
f. c n.s.xs., j=1, ••• ,n and relations W. c n.sixs., j=1, ••• ,n (pre-
]-~~] ]-~.]
ference relations; (s, ~) E Vj is interpreted as l'S with Sj replaced
by ~ is better than s, for j"), an equilibrium (II) is a system of
strategies s * E nisi such that for all j, * *
(s ,s.) E r. and for every
~ ,
*
(s , ~) E rj excludes
*
(s , ~) E¥j •
]]
In [1975] Shafer and Sonnenschein proved the following:
12. THEOREM (Shafer, Sonnenschein). An equilibrium II exists if for

every j:
a. Sj is a compact convex set in a Euclidean space;
250
b. fj is closed;
c. rj is lower semi-continuous and has nonempty convex sections
d. Y J. is open and irreflexive, i.e. there is no s with (s,s.) E'I'.;

J J
e. ~. has convex sections
J
Obviously, a special case of the theorem of Debreu with assumptions

(a-), (b), (c-), (d) and (e-) can be derived from the theorem of Shafer
and Sonnenschein just by taking, for every j,
'f' j .. =
The crucial step in the proof of Theorem (11) (this would be also
the case in the proof of Theorem (12) if the authors were not restric-
ting themselves just to Euclidean spaces) relies on the use, for every
j, of a function f.: n.s.xs. ~ R with the following properties:
J 1. 1. J
13.a. fj(S'~) > 0 if and only if the player j can improve upon at
s by changing the jth coordinate of s to ~;
13.b. Zj:= {(s,~) EniSixSjlfj(s,~) max{f j (s, t )I(s, f;.) E fj }} is

a closed set;
13.c. for every s E nisi' {~E Sjl (s, ~) E Zj} is nonempty and con-
~ (version: ~).
Let me set the matters more precisely (I shall begin with a defini-
tion and continue with a theorem) ;
given a positive integer n, nonempty sets S1""'Sn' relations

r.Jcn.s.xs., j=1, ••• ,n and functions f.:n.S.xs. ~ R, j=1, ••• ,n, an
1. 1. J J 1. 1. J
equilibrium (III) is a system of strategies s*En.S. such that for
-"'------*"" * 1. 1.
all~, (s ,Sj) E fj and for every ~, (s*,~) E rj implies
f.(s,~»O.
J
14. THEOREM. An equilibrium III exists if for every j:

a. Sj is a compact convex set in a locally convex Hausdorff topological
vector space (version: a compact acyclic metric ANR)
b. rj is closed;
c. the set ZJ' := {Is,!::) E n.s.xs.lf.(s, 1:) = max{f.(s,/:') I (s,t.) Ef'J'}}
'i 1. 1. J J '> J
is closed and has nonempty sections Zj(s) := {~I (s, ~) E Zj};
d. for every s E nisi and (){ > 0, {~E Sj Ifj (s, ~) ~ (){} is convex
(version: ~).
251
Notice that condition (c) is a consequence of (b),

c-. fj is lower semi-continuous~ and
II
c • fj is continuous.
Condition (2) simply means that Pj is quasi-concave in the second

coordinate~ the version is also known and considered in the literature
(see e.g. McClendon [1984], p. 151).
Theorem (14), with conditions (c-), (e") replacing (c) (but not the
version) is in Wieczorek [1986], Theorem 1.4, p. 11 or Ichiishi,
Theorem 4.7.2, p. 70.
Theorem (14) is an immediate consequence of the Fixed Point Theorem

of Fan Ky [1952] and Glicksberg [1952]~ the version follows from
Eilenberg-Montgomery Theorem [1946]. In the proof it suffices to con-
sider the set
1, ••• ,n}
and to notice that, by a fixed point theorem Z contains some (s*,s*)~
obviously every such s* is an equilibrium III.
One can also think of other versions of Theorem (14)~ the appearing
conditions would then guarantee the existence of an (s*,s*) E Z by
means of a Lefschetz type theorem.
The reader will realize that the definitions of equilibrium I, II

and III can be automatically transferred to the case with an arbitrary
number of factors j. Obviously, Theorem III still holds in this case~
its version is true for a countable number of factors.
Thus we see that investigating special cases of Theorem (14) (in

particular stating that Theorems (11) and (12) are its special cases'
heavily relies on the following:
15. PROBLEM. Given sets S1' ••• 'Sn in a topological vector space
(version: ~~) and a relation =
Wj nisixs j when does there
exist a continuous function f J. : n.s.xs. ~ R such that
~ ~ J
i'j = {( s, ~ ) If j (s, ~) > O} which is quasi-concave in ~ (version:
such that for all s E n.s.
~ ~~
~ a,{~lpj(s,~) ~a} is
acyclic or empty).
~
252
This problem can be reformulated in an obvious way:
15~. PROBLEM. Given a topological space T, a set X in a topological

vector space (version: a metric space X) and a family (>t1tET) of
~
binary relations in X, when does there exist a continuous function

IT: TxXxX ~ R such that for every tET, ITt(x,y) is a pseudo-utility
representation of >t' quasi-concave in x (version: such that for
~
~ yEX ~ a, {xlrrt(x,y) > a} is acyclic or empty)?

~
This problem has been already answered in Section III by Corollary

(10) which actually arose as a result of investigating the relations
between Theorems (12) and (14).
Let me also formulate another (trivial but important) sufficient

condition illustrating the fact that Theorem (11) is indeed a special
case of Theorem (14): a function whose existence is required in Problem
(15~) exists whenever there exists a continuous "parametric utility"
function u:TxX ~ R such that for every tET, u t is a quasi-concave
utility representation of >t'
Darmstadt, August 1986
REFERENCES
K. Arrow and G. Debreu, 1954, Existence of an equilibrium for a compe-

titive economy, Econometrica 22, 265-290
R.J. Aumann, 1969, Measurable utility and the measurable choice theorem,
in: La Decision (Editions du Centre National de la Recherche
Scientifique), 15-26
G. Debreu, 1952, A social equilibrium existence theorem, Proc. Nat.

Acad. Sci. USA 38, 886-893
S. Eilenberg and D. Montgomery, 1946, Fixed point theorems for multi-

valued transformations, Amer. J. Math. 58, 214-222
253
Fan Ky, 1952, Fixed point and minimax theorems in locally convex topo-
logical linear spaces, Proc. Nat. Acad. Sci. USA 38, 271-275
P. Fishburn, 1970, Utility theory for decision making, Wiley, New York
I.L. Glicksberg, 1952, A further generalization of the Kakutani fixed

point theorem with application to Nash equilibrium points, Proc. Amer.
Math. Soc. l, 170-174
T. Ichiischi, 1983, Game theory for economic analysis, Academic Press,

New York
A.Ya. Kiruta, A.M. Rubinov and E.B. Yanovskaya, 1980, Optimal~nyr vybor
raspredeleni~ v slozhnykh social~no-ekonomicheskikh zadachakh, Nauka,
Leningrad
D. Krantz, R. Luce, P. Suppes and A. Tversky, 1971, Foundations of

measurement, Vol. I, Academic Press, New York
J. Nash, 1950, Equilibrium points in n-person games, Proc. Nat, Acad.

Sci. USA ~, 48-49
J.F. McClendon, 1984, Minimax theorems for ANR~s, Proc. Amer. Math.
Soc. 90, 149-154
W. Shafer and H. Sonnenschein, 1975, Equilibrium in abstract economies

without ordered preferences, J. Math. Econ. ~, 345-348
A. Wieczorek, 1980, On the measurable utility theorem, J. Math. Econ.

I, 165-173
A. Wieczorek, 1986, Constrained and indefinite games and their applica-

tions, Dissertationes Math. ~, 1-43
III. GOAL SETTING AND DECISION MAKING
A FUZZY CONCEPT OF EFFICIENCY
;Jochen Ester
Technische Universitat Karl-Marx-Stadt
PSF 964, Karl-Marx-Stadt, 9010, DDR
O. MOTIVATION
We have been working in the field of MCDM since the early seventies.
With the aim of practical applications and technical and economic
effects we have achieved some results in theory, methodology and
applications. Recently, we could help engineers, economists and
others in solving the follOWing practical problems:
o optimal design of freezing systems,
0 optimal design of mechanical elements,
0 optimal design of textile articles,
0 optimal selection in training of high-performance swimmers,
0 optimal distortion of the command variable in position
feedback control,
o optimal control of robots.
Our last result in development of computer-tools for Computer Aided
Decision Making is an interactive program system POLYP.BC for micro-
computers. At present, we have reached the following stage:
MADM: Incomplete weighting method, Prevalence method, Method with
fuzzy dominating sets, Generalized Hodges-Lehmann criterion,
MODM: Global stochastic search, Antiparallel gradients method,
Generalized antiparallel gradients method (interactive),
Interactive method for linear t~ODM-problems, some single
criterion optimization methods.
These methodS reqUire a lot of important subroutines, e.g. normali-
zation, ordering (partial ordering, fuzzy partial ordering) and
methods for scalariz~ng functions and model-building, which are
implemented as single modules of the system. Due to the simplicity
of the structure, of the connections and the interfaces between all
modules in POLYP.BC, new modules can be written and connected with
the system.
Our experience is that there is no method, which can be used for
each problem or for each deciSion maker.
In many cases we could find out that the experts in modelling of the
258
decision object are not very sure in the

o determination of the feasible region,
o determination of the relations between input variables (deci-
sion variables) and output variables (criteria) of the system.
Many discussions with the decision makers have shown that they could
give us the following expression:
"The result of the decision must have the property 1 and the proper-
ty 2 ~ ••• .2!. the property 1 ~ the property 4 ••• I"
On the other hand, these "and" and "or" of the decision makers are
not so strict as, for instance, in Boole's Algebra.
We believe that with Fuzzy Theory a lot of these problems can be des-
cribed and solved. Therefore, we have developed the following method.
1. QUALITY OF A DECISION - FUZZY NOTION
For the evaluation of different decisions we have a set of single

features, attributes or objectives, which characterize the "quality"
G of a decision.
G = aggr(A 1 • •• Am) (1)
G - set of fuzzy decisions with respect to all criteria,
Ai- set of fuzzy decisions with respect to the single
criterion qi' i=1(1)m ,
aggr - aggregation rule, depends on the decision maker(s); indirect-
lyon influences from outside.
The set of attributes, objectives, criteria can be divided in such
of
o admissibility, practicability,
a usefulness,
o economic utility,
o expense,
o certainty of the "expected" outcomes.
Admissibility and/or practicability lead to the fuzzy set of feasib-
le decisions, but certainty or uncertainty in this sense lead to the
fuzzy mapping from the decision space into the criteria space.
20 GENERALIZED PROBLEM SETTING
In general, we can write

"max" {q(X)1 X€x} • ( 2)
where "max" is not specified and means that all criteria
259
( q1 ...
q )T = q shall be maximized.
m
X~Rn _ decision space,
Q: X~ZCRm - criteria space,

X - set of feasible decisions in the
decision space,
Z - set of feasible deciSions in the
criteria space,
Q - mapping , Qc Rnx Rm •
X and Q may be fuzzy sets. Then, consequently, Z is also a fuzzy set.
X = {xl jix(X)} ,
Q = {( x,q) I fiQ( x ,q)}
Z ;: {q lfi( q) }
where;MX(x), JUQ(x,q),;«(q) are the membership functions of the
fuzzy sets. The following relation holds
Z = aggro(x,Q) ( 3)
fo r instance,
~(q) = m~x minyux(x)~(X,q» •
Now, we define a supporting set of Z.

Zo={ ql;'(q» o}, ZED cl(Zo)
The membership functionjU(q) is, in fact, an important criterion.
Therefore, we get a new problem setting, which is a substitute
problem of (2).
(4)
Now, we norm'alize this new MCDM-Problem
qm+1 =;U(q) , m:= m+1 ,
"max" [q I q€Ze}= PMCR m , (5)
where PM denotes the Pareto set.
I
With the individual maximums qimax= max { qi q E: Ze} and the lower
bounds for each criterion in the Pareto set qimin= min [ qi I q E PM}
we transform qi - q
_________ + _d
__
i_m_i_n~
qi = , 0< d«1, i=1(1)m , (6)

qimax- qimin + d
q' = (qi1 ••• q~)T£ Z' - feasible set of q' •
Due to the transformation (6) the next relation holds
PM'~Z', PM'CE={q/o.(qi"1,i=1(1)m},
which is important for a simple definition of membership-funct.of Ai-
260
3. CHOICE OF PARETO-OPTIMAL ELEt'lENTS (FINAL DECISION)

For the fuzzy set G (c.f.(1» it holds
G ={ q I,lAG( q)} - fuzzy set of fuzzy decisions,
max~(q) I qE Z'f\ R:} - optimal fuzzy decision.
For the single attributes we have the fuzzy sets Ai'
Ai = f q Iti ( q )}
and we choose
0 for qi ~ 0
[
jUi(q) = q1 i for 0 ~ qi ' 1
for 1 ~ qi
From Fuzzy Theory we know a lot of aggregation rules, for instance,
(i) • • • intersection
( ii) • • • • .0. • • 0 • • • • intersection
( iii) .. . . . . . • intermediate
(iv) m~xfi) • • • • • • •••• • • • • • union (7)
(v) p( minyvri) )+( 1-p)( m~xYUi»' 0 ~ p ,£ 1 •• • union-inter-

0 ••
section
(vi) p(minY'i»+(1-P)(~L (Ui»,O ~ p ,£ 1 • 0 • intermediate-
• 0 •
i i '!: intersection
(vii) p(maX(Mi»+(1-p)(~[
i I . i
(M
1 i
» ,0 ' p 61 • • • • • • intermediate-
union
(i)-(iv) can be covered by
m
I'G( p ,g, r y«> = ( L
i=1
gi(rfi)P
m
)1/p (8 )
with g = (g1 gm)T, gi a 0 , [gi = 1
...
i=1
-oo<p<+oo,
~ = ~ ".~m)T 0
It must be noted that~G in (8) is not normalized o This normalization

can be carried out (c.f.(22», but it is not so important for the
following conclusions.
Especially, we get from (7)
261
( 11) )iG(P=O) = n(r0

i i
)9i
(iii) I'G(p=1) • ~(gir¢'i) (9)
(iv) I'G(P=+OO) • m~x( r¥'i) _. 0
If gi=(1/m),i=1(1)m and r i =1,i=1(1)m we obtain (i),(ii),(iii) and

(iv) from (7)0
(I) Global requirements

In which direction lies a reference point qO=;UO ? This direction is
determined by the vector ro The following relations hold
= min(r i )
i
= (rlr i ) i=1(1)m (10)
~ can be found with (10) by given r. Or, if;U° is given with
mixyt~) =fj = 1
r can be found by
r i = 1fU~ , i=1(1)m 0 (11)
What is the character of the reference point?
G =Ui Ai ( st rict or soft?) , +00> p ~ 1 ,

G =nAi (st rict or soft?) , -00< p < 1 0
i
In Fig.1 the dominating sets are shown, if the character o~o is
strict "or" and strict "and".
o
_ _ ifnon~Rnve.x dominating set
of p- in the case of
/ . strict "or"
convgx dominating set
'11 ~. strict "and"
'\.
of ft: in the case of
fJ' ~
~~------~~~--~-.~-------
Figure 1: Dominating sets
fiG(p=+oo) and;Us(p=-oo) are the corresponding achievement functions.

In Fig.2 the shape of the niveau surfaces ofjUG(p) in the criteria
space R2 is shown.
In Iii are explained and proofed the following interesting theorems
262
()
x-w
Figure 2: Niveau surfaces of~G(p), r,g are fixed
for r fixed, which describe relations between important solution

sets if p is decreasing or increasing.
We define
Argm~X~(p,g,q) ( qeZ'nR:} = e~ , (12)
eP = U eP
g g
(13)
It holds
I. ePf PM' for -00< p < +00 (14)
263
II. EPf ES for -00< s tI p< +00 (15)

III. Suppose (p~EP) is continuous for p<p'o Then, for a given E ~ 0
we find p< p' with:
V q' e PM' : 3 q" € EP: f q' - q "/< [ 0 (16)
(14)-(16) lead to the following conclusions:
The "optimal fuzzy decision" is always a paretooptimal solutiono If
g is not fixed, the "optimal fuzzy solution set" is increasing with
decreasing power p. On the other hand, if p is small enough, the
corresponding optimal fuzzy decision set is an approximation of the
Pareto seto These properties are independend of the vector of para-
meters r (or of the choosen reference point).
(II) Local requirements - sensitivity of the solution

g determines the marginal rates of substitution in the optimal fuzzy
solution qOpt. It is easy to show that the next relations hold
qopt_- (opt
q1 000 qopt)T
m
. dZG
oq g (r )p(qopt)p-1
=+...:.:L=+ j j j o (17)
dZ G gi(ri)P(q~Pt)p-l
oqi
Therefore, it is very simple to compute the substitution-rates in
the found optimal fuzzy decision. This allows to quantify the effec-
tive rates of compensation between the single attributes. On the
other hand, we can control the values Kij by varying the parameters
gi and gj without any variation in the global requirements.
If the marginal rates of substitution are given by the decision
maker, for a fixed reference point we can estimate the parameters
gi,i=1(1)m and the degree of the logical properties.
40 HIERARCHICAL STRUCTURE
It is very easy to build up a hierarchical model of efficiency or
quality of a decision o If, for instance, G is given in the following
form
G = un
j iEl j
Ai =
j
Gj U (18 )
we can bUild up a hierarchical model in the following wayo

Each Gj has its own power Pj (degree of logical properties)o
264
G j = {q lfj<t(q»} (19)
t jJM( q» (L J g~(r~M (q »Pj//Pj

=
ieI. 1. "( i i
(20)
JAG ('1- gj(rt» Po l/po , Po ~ 1

=
~ j
(211j )
For these hierarchical rules it is reasonable to normalize the

membership functions. This can be done on each level, if the next
relation is fulfilled.
L
kfI
gS(r s / s = 1
k k
( 22)
s
Using (22) we do not change the direction in which lies the refe-
rence point, but its fictive locus in this direction.
5. SUMMARY
a) Admissibility and/or certainty can be handled as a single
criterion.
b) The logical property of the aggregation rule can be controlled
by a parameter p.
c) Given reference points and given weighting coefficients can be
taken into consideration.
d) With P and the reference point we can control the global pro-
perties (locus) of the solution.
e) With the weighting coefficients we can control the substitution-
rates in the locus of the solution.
f) The solution is an optimal fuzzy solution and it is paretooptimal.
g) By a hierarchical structure we can bUild up a more complicated,
multilevel decision making model.
REFERENCES
/1/ Ester,J.:Troltzsch,F. (1986). On generalized notions of efficiency
in MCDM, Syst.Anal.Model.Simul., 3(2)
/2/ Werners,B. (1984). Interaktive EntscheidungsunterstDtzung durch
ein flexibles mathematisches Programmierungssystem,Minerva,MOnchen
/3/ Peschel,M. (1980).Ingenieurtechn.Entscheidungen, Technik, Berlin
/4/ Ester,J.;HolzmOller,R. (1986).About some methods and applications
in MCDM, Syst.Ana~.Model.Simul., 3(5)
/5/ Dubois,D.:Prade,H. (1980). Fuzzy sets and systems - theory and
applications, Academic Press, New York
AN APPROACH TO r.1EASURDIG CONSISTENCY OF
PREFERENCE VECTOR DERIVATIONS
USING LEAST SQUARE DISTANCE
Gerd Islei
Manchester Business School
Booth Street West, Manchester M15 6PB
There is a wide interest in the use of judgement models for real life
decision making. However recent research in the psychology of decision
making has pointed to serious shortcomings which are often not recog-
nized by the practitioner and are also neglected by the analyst. We
shall examine some. conceptual inferences by looking at different ver-
sions of calculating preferences and discuss their usefulness for real
life applications.
A geometric representation is used to facilitate the interpretation of

the results and to permit a systematic analysis. As a consequence an
approach is suggested which is based on least square derivations and
which overcomes major shortcomings.
At the same time a measure of consistency can be defined which permits

investigation into the behaviour of the decision maker towards his/her
own inherent uncertainty.
1. Survey of the Problem
The use of judgemental models for real life applications is receiving

increased attention. There are a variety of approaches based on
slightly different assumptions which have been applied with varying
degrees of success. This paper examines some conceptual inferences
by looking at different versions of calculating preferences and dis-
cussing their usefulness. Particular attention is paid to the way
judgements are aggregated (weighted and normalised) and its compati-
bility with empirical observations.
We shall consider a decision problem with n attributes A1 , A2 , ..• ,An

and call a matrix of the form
266
a judgement matrix if the entry a ij represents the degree of preference

of attribute Ai over Aj (i.e. Ai = a ij A j ). Therefore a judgement mat-
rix represents all pairwise comparisons between the n attributes A1 ,A 2 ,
••• ,An'
In direct assessing methods (Keeney and Raiffa (1976)ljudgements are
usually made against some fixed scale, therefore assuming transitivity
of preference judgements. That is, if we transform these assessments
into ratio comparisons the corresponding judgement matrix is consistent,
i.e. has the property :
aij=a ik a kj for all i,j,k=1, • • • , n
(with other words, all entries of the matrix can be deduced from n-1
pairwise comparisons) •
The eigenvector method (Saaty (1980)) does not require consistency but
assumes that judgement matrices are reciprocal (i.e. a ij = 1/a ji for all
i,j=1, • • • , n) and requires that all entries of the upper triangular
half of the matrix be given (i.e. n(n-1)/2 comparisons have to be made).
The philosophy behind direct assessment is that you elicit consistent

(i.e. minimal) information from the decision maker (O.M.) and ask for
adjustments of the comparisons if the results are considered unsatis-
factory (all judgements being made with one frame of reference). Our
research, however, indicates (see also Kahneman and Tversky (1982))
that these adjustments usually are quite insufficient, new estimates
are biased towards the initial values (i.e. anchoring occurs).
The eigenvector method (E.M.) asks for redundant information by using

judgements with several frames of reference +) in order to combat noise
and to check the O.M.'s consistency. However we found (see also Coombs
(1964) that if you ask for too much (redundant) information then the
effectiveness of people's judgements deteriorates. (A typical real
+) row/column i corresponds to comparisons of all attributes with resp-

ect to the reference attribute Ai
267
life case study evaluates 8-10 choices with respect to 8-10 attributes,
i.e. requires a total number of 200-300 comparisons for the E.M.).
Also in our applications (Lockett and Hetherington (1983» a precise

final result was often rather less important than a comprehensive anal-
ysis of the D.M.'s uncertainty. The consistency index derived from the
E.M. unfortunately provides only a very crude measure with limited sta~
istical properties (for example it does not enable us to examine how

much individual assessments contribute to this overall index). There-
fore its practical value has been quite insignificant.
There are two more observations from our applications which should be
mentioned
(i) The D.M. often considered the spread of the E.M. results to be too
large (Lockett, Stratford, Cox, Hetherington and Yallup (1986»
(ii) Sometimes D.M.s appear to have difficulties in interpreting ratio

scales correctly (that is in handling the multiplicative nature of
the judgements).
The aim of this paper is twofold :
First to provide a comprehensible geometric interpretation of the E.M.

approach and to illustrate some shortcomings.
Second to introduce a method using a Geometric Least Squares technique

and to demonstrate its usefulness in analysing inconsistencies.
2. A Geometric Interpretation of the Eigenvector Method
To illustrate several aspects of our discussion we shall consider a

problem with three attributes where the following three pairwise compar-
isons have been supplied :
A1 =a 12 A2 A1 = a 13 A3 A2 =a 23 A3
As a concrete example we shall use (Saaty and Vargas (1984»
A1=2 A2 A = 5 A3 A = 7 A3
1 2
This can be recorded in the following judgement matrices
A1 a 11 a 12 a 13 2 5
A2 a 21 a 22 a 23 A2 1/2 7
A3 a 31 a 32 a 33 A31/5 1/7
268
+)
These jUdgements can also be represented as geometric graphs the
following way :
A1
A3 ~---. .- - -... A2
7
A first approach to solving this problem might be to demand that the

result should be as 'close' as possible to the input comparisons.
Least Square Methods (L.S.M.) (Cogger and Yu (1983)) do precisely that
with 'close' being defined with respect to a suitable distance function
(usually the Euclidean metric). However it can be seen (Saaty and
Vargas (1984)) that in the case of the above concrete example the re-
sult from L.S.M .. appears intuitively wrong++)
The reason for this is that L.S.M. only uses direct comparisons but ig-
nores (what we call) indirect comparisons (i.e. it ignores that for ex-
ample from A1 = 2 A2 and A2 = 7 A3 we can deduce the indirect comparison
A1 = 14 A3 )·
Since the E.M. produces an intuitively more acceptable result
let us now take a closer look at the principle on which the calculation
of the E.M. is based.
Given a reciprocal judgement matrix, E.M. takes the normalized princi-

pal eigenvector as preference vector and the principal eigenvalue is
+) One should always keep in mind that a matrix is a convenient device

for recording the comparisons of a judgement process, it is not an
integral part of it. Therefore particular properties of the matrix
used for the evaluation need careful empirical justification.
++) Input (direct) comparisons reveal preferences A1 > A2 ; A2 > A3

and A1 > A3 . Therefore we might not accept a solution whlch
produces a preference A2 > A1 as is the case with L.S.M.
269
used to measure consistency. The snag in these calculations, however,

is that the principal eigenvector can be derived from the limiting
value (suitably normalised) of the powers of the judgement matrix. (To
be more precise from the limiting values of a column).
Hence it is more appropriate to examine what assumptions are implied

by the use of the powers of a judgement matrix rather than to discuss
properties of principal eigenvectors.
If we take a general 3 x 3 matrix to the power of 2 we obviously ob-

tain :
(2) (2) (2)

all a 12 a 13
(2 ) (2) (2)
a 2l a 22 a 23
(2) (2)
a 32 (2)
a 3l a 33
(2) _
with all - a ll 'a ll + a 12 'a 2l + a 13 'a 3l
(2)
a 12 = a l ,. a 12 a 12 'a 22 + a 13 'a 32
+
etc
Therefore if we use the representation of our geometric graph the ele-

ment al~) is nothing but the sum of all paths of length 2 from Ai to
A.. (Such a path of length 2 shall be called indirect comparison of
J
order 2).
Similarly the element a!~)of the matrix to the power of 3 represents

lJ
precisely the sum of all paths of length 3 from Ai to Aj . (These paths
are denoted indirect comparisons of order 3). Etc.
Therefore the result of the E.M. is essentially based on the arithmet-

ic average of all paths from Ai to Aj with length going to infinity.
But if we take a look for example at the element a~~)of the general
3 x 3 matrix
all,all'all
+ all'a12'a2l
+ all'a13,a3l
then the following observations can be made:
(i) Higher order comparisons include lower order comparisons with

270
increasing weight
(ii) Inconsistent loops are handled unsatisfactorily
e.g. in our concrete example the entry a~~)contains the loops
an' a 32 o a 21 5-1/7'1/2 0.35
a12oa23·a31 2-7 ·1/5 2.80
that is, it contains the same loop twice only traversed in opposite di-
rection (as indicated by the arrows) .
Since obviously a13·a32·a21 = (a12oa23.a31)-1 the averaging procedure

used in E.M. generates mistakes as soon as inconsistent loops (i.e. in-
direct comparisons of order 3 or more) are included.+) Also these mis-
takes become more significant with increasing inconsistency.
These observations are largely responsible for the following consequen-

ces
(1) Results of the E.M. for judgement matrices with inconsistent in-
puts exhibit a comparatively wide spread
(2) The E.M. can produce odd results
Two illustrative examples will be discussed in the next section.
3. Illustration of some Shortcomings
Consider the following judgement matrix (Saaty and Vargas (1984))
Example 1
A1 A2 A3 A4 AS Solution E.M.
A1 1 1/7 1/2 1/8 2 A1 = 6.12%
--+ A2 7 3 8 A = 37.43%
2
A3 2 1/3 1/4 5 A = 13.42%
3
----to A4 8 4 5 A4 = 38.69%
AS 1/2 1/8 1/5 1/5 A =

5
4.34%
It has been claimed (ibid.) that the E.M. produces the "intuitively"
correct ranking of the attributes A2 and A4 (namely A4 >A 2 ). We shall
use the following (partial) geometric graph of this problem to examine
this claim.
-1
+) Note that for all xE;R we have: x;x l!! 1 with equality iff x=1
271
The direct comparison a 24 =1 shows indifference between A2 and A4 and

the graph contains all comparisons of order 2 between A2 and A4 . The
· t h e E.M. assumes t h at t h e rat~o
averaging proce d ure use d ~n (-)
. A2 / A4 =24
is approximated by the average of all comparisons of order 2 between
A2 and A4 ·
Altogether we have 'the following comparisons of order 2 between the
attributes A2 and A4 +) : A = 7·1/8 A4 0.88 A4
2
A = 3 ·1 /4 A4 0.75 A4
2
A = 8'1/5 A4 1. 60 A4
2
A2 = 1 -1 A4 1.00 A4
A = 1 -1 A4 1.00 A4
2
But if we do the same calculation now going from A4 to A2 we obtain
A4= 8 -1 /7 A2 1. 14 A2
A = 4 -1 /3 A2 1. 33 A2
4
A = 5·1/8 A2 0.63 A2
4
A = 1 •1 A2 1.00 A2
4
A = 1•1 A2 1.00 A2
4
with average A4 =1.02 11. 2 (i. e. A4 >A 2 ) ++)
+) Note that in the case of this second order approximation the dir-
e~t compariso~ a?4=1 appears twice in the calculation, i.e. it is
g~ven more we~ghE.
++) Note that this peculiar discrepancy to the previous result occurs
because the arithmetic mean is taken over (basically multiplica-
tive) comparisons which should be combined multiplicatively.
272
If anything then the dominance in the first case (resulting in A2 >A 4 )

is stronger than the dominance obtained in the second calculation (re-
sulting in A4 >A 2 ). Furthermore the geometric mean for both calcula-
tions is of course identical and gives the result: A2 =1.01 A4 (i.e.
A2 >A4 ). Therefore it is not "intuitively obvious" that the E.M. pro-
vides the correct ranking.
It is even more instructive to consider the following judgement matrix

(Saaty (1980» :
ExamEle 2
A1 A2 A3 A4 AS A6 Solution E.M.
-
A1 1 4 3 1 3 4 A1 =32.08%
A2 1/4 7 3 1/5 A2 =13.9S%
A3 1/3 1/7 1/5 1/5 1/6 A3= 3.48%
A4 1/3 5 1/3 A4 =12.8S%
AS 1/3 5 5 3 AS =23.74%
-A6 1/4 6 3 1/3 A6 =13.91%
and to examine the ranking of A2 in comparison to A6 •

The direct comparison a26~1 shows indifference between A2 and A6 ,
and using the comparisons of order 2 first going from A2 to A6 yields
the average A2 =0.96 A6 (i.e. A6 >A 2 ).
However going from A6 to A2 produces in this case the average A6 =1.09A 2

(Le. A6 >A 2 ). Similarly the geometric mean gives the result A2 =0.94A 6
(Le. A6 >A 2 ).
Therefore we do not expect higher order comparisons (which use lower

order comparisons with increasing weight) to produce a rank reversal as
the E.M. does.
We conclude that the E.M. can produce 'odd' results simply because it
uses an arithmetic average for ratio comparisons, hence including in-
consistent loops in a very unsatisfactory way.
There is also an empirical reason why it may not be advisable to use

comparisons of order 3 or higher in our calculations. It has been ob-
served that in most judgemental processes people generate patterns with
respect to frames of reference, and we found in our applications that
D.M.s treat the (direct) com~arisons of an individual row/column as if
273
they correspond to the same frame of reference+). They also take ac-
count of tradeoffs (i.e. comparisons of order 2) between these frames
of reference (i.e. rows/columns). However in our experience they do
not consider tradeoffs of tradeoffs (i.e. higher order comparisons).
Nevertheless our applications also suggest that slight differences in
the results are rather less important in discriminating between differ-
ent approaches than a potential to reduce data-requirements and to per-
mit a more comprehensive feedback for the D.M.
4. Introduction to Geometric Least Square
In this section we shall briefly introduce a method of calculating pre-

ference vectors which takes account of these problems. (Full details
are in Islei and Lockett (1986». Again we are lead by empirical ob-
servations we encountered in most of our case studies. We found that
if people do pairwise (i.e. relative) comparisons of attributes, for
example, they do not merely deliver value estimates (i.e. sample points)
but rather they determine a functional relationship between these at-
tributes (i.e. they implicitly declare a marginal rate of substitution).
Let us consider the following examples
A1 2 5 A1 2 14
(I) A2 1/2 7 (II) A2 1/2 7
A3 1/5 1/7 A3 1/141/7
In the case of matrix(II) the entries of the upper triangular half re-
present the following equations :
A1 - 2 A2 0 (1)
A1 14 A3 0 (2)
A2 - 7 A3 0 (3)
and normalising requires
A1 + A2 + A3 = 1 (c)
Now taking up our empirical argument it is not too difficult to real-
ise that these equations represent planes in a three-dimensional space
with coordinates A 1 , A 2 , A3 . For example, the equation A1 - 2 A2 =O
+) row i representing the comparisons of all attributes with respect

to the reference attribute A ..
1
274
represents the plane (see Figure 1) which is spanned by the straight

line A1 - 2 A2 =0 of the A1A2 -plane and the A3 -axis. The normalising
requirement (c) is represented by the plane through the points (1,0,0),
(0.1.0), (0,0,1). Therefore, if the comparisons of the judgement mat-
rix are consistent (as is the case for matrix (II) ), the solution is
the unique point of intersection of all those planes (1) - (3) and (c).
(The circled point in Figure 1). Hence our overdetermined system of
linear equations has a unique solution. However, if the comparisons
are not consistent, then this is not the case.
For example, matrix (I) with equations:
A1 - 2 A2 0 (4)
A1 - 5 A3 0 (5)
A2 - 7 A3 0 (6)
A1 + A2 + A3 (c)
is not consistent.
Therefore the solution in this case should lie inside the shaded tri-
angle of Figure 2. The optimal solution will be the point of the tri-
angle which has least square distance from the planes (4) - (6). (For
details and further discussion see Islei and Lockett (1986).
If we now take into account the previous arguments then of course we

suggest as optimal solution the point which has least square distance
from the planes representing all comparisons of order 2. This will be
referred to as Geometric Least Square method (G.L.S.). Applying G.L.S.
to the example of matrix (I) gives the solution :
A 1 = 53.64%
Obviously our method does not require all direct comparisons for a cal-
culation but can be used as soon as the available direct comparisons
permit us to calculate at least one indirect comparison between any two
attributes. We are thus able to generate a succession of results by
including more and more comparisons in our calculation.
The judgement matrix of Example 1 will be used to illustrate how the

optimal result changes if additional comparisons are successively in-
cluded.
If we start with the comparison of the top row
comp.1:a 12 =1/7 comp.2:a 13 =1/2 comp.3:a 14 =1/8 comp.4:a 15 =2

275
then G.L.S. produces the result A = 5.41%

1
A2 = 37.84%
A~=
.:)
10.81%
A4 = 43.24%
A5= 2.70%
Now we add successively all the other comparisons represented in the

top triangular half of the matrix :
compo 5 compo 6 compo 7
compo 8 compo 9 comp.l0
and obtain :
compo 1 - 5 compo 1 - 6 compo 1 - 7
Al = 5.51% A = 5.58% A = 5.90%
1 1
A2 = 37.16% A2 = 38.81% A2 = 38.06%
A = 11. 53% A = 12.10% A = 12.04%
3 3 3
A = 43.05% A4 = 40.70% A4= 40.16%
4
A = 2.75% A = 2.82% A = 3.85%
5 5 5
comp 1 - 8 compo 1 - 9 compo 1 - 10
A1 = 5.80% A1 = 5.88% A1 = 6.89%

A2 = 37.33% A2 = 36.50% A2 = 37.94%
A3= 11.40% A3= 12.40% A3= 13.97%
A4= 41.62% A4= 41.79% A4= 36.23%
A5= 3.84% A5= 3.43% A5 = 4.97%
A cursory look at these results indicates that individual attributes

show little variation, but that by adding the last comparison 10 (i.e.
going from the results of compo 1-9 to those of compo 1-10) some dist-
inct noise is generated. This can be made more precise by carrying
out an error analysis (theoretical details are in Islei and Lockett
(1986) ) •
The results of this analysis for compo 1-9 and compo 1-10 are shown in
TABLE 1 and 2 respectively. (See Appendix for some comments on this
error analysis) .
Clearly the comparison a 45 =5 is not very consistent with the previous

inputs, for example the solution generated by compo 1-9 has an average
distance of 1.11 to the planes representing these comparisons (with
maximum distance 3.744) whereas a 45 has a distance of 4.830 to it.
Also if we include comparison a 45 =5 into our calculations we then ob-
tain a result which shows a marked increase for the standard error and
276
and for the average distance of the solution to the direct comparisons.
Nevertheless a comparison with random inputs (forthcoming paper) shows

that the judgements of this example are still very compara.ble. +)
5. A Practical Application
Now we shall look at some figures from a typical real life case study
and show how the error analysis was used to analyse inconsistencies.
This application required 9 attributes and the following judgement mat-
rix was supplied for them
A1 A2 A3 A4 A5 A6
A1 1 7 7 1 1 3 7
A2 1/7 1 5 1/7 1/5 1/3
A3 1/7 1/5 1/9 1/7 1/5
A4 7 9 1 3 5 7
A5 1 ·5 7 1/3 1 3 5
A6 1/3 3 5 1/5 1/3 1 3
A7 1/7 1/7 1/5 1/3
A8 1/7 7 3 1/5 1/5 1/5 5
A9 9 9 3 1/5 7
The E.M. produced the following results
A = 18.29% A2 = 2.71% A3 = 1.67% A4 = 23.02%

1
A = 13.99% A6= 15.31% A7 = 2.18% A8= 5.07%
5
A = 17.74%
9
and gave a consistency index of 83.12%.
Although a consistency idex of 83.12% is below what Saaty considers

satisfactory (Saaty (1980» in most of our applications a D.M. is hap-
py (as he was in this case) with a figure above 80% (Lockett and
Hetherington (1983» mainly because it seems to indicate that the
inputs are quite consistent (and also because he usually does not know
which comparisons should be adjusted) :+)
+) Note that G.L.S. gives the {ollowing solution for Example 2

A 1 = 24.60% A2 = 15.51% A3= 7.12%
A4= 15.04% A5= 21.43% A6= 16.31 %
++) Furthermore in this particular case using the E.M. he still had to
supply some additional 200 judgements for the choice comparisons.
277
However when we presented to the D.M. the results of G.L.S. and the
error analysis (TABLE 3 shows the printout for the complete matrix) ,
indicating that comparison a 69 =S clearly stood out (i.e. that the in-
clusion of this comparison generated a considerable noise), he real-
ised that he should have input a 69 =1/S.
After adjustment of this comparison we obtained the following results

for the E.M.
A = 18.97% A2 = 2.86% A = 1.78% A = 22.50%

1 3 4
AS= 13.74% A = 8.01% A7 = 2.26% A8 = 5.68%
6
A = 24.19 %
9
with consistency index now being 90.54%.
The new results for G.L.S. with error analysis (for the complete mat-
rix) can be found in TABLE 4.
A comparison of these results with the previous ones shows that rank
reversals occurred for the three leading attributes with either method.
+)
However only E.M. produced a rank reversal for the attributes AS' A6 .
The D.M. was quite satisfied with the new results and accepted the
noise as indicated in our error analysis.
Now let us briefly examine if in this example the total number of 36

(=n(n-1)/2) comparisons made by the D.M. could have been effectively
reduced.
We found in most of our applications that the successive results tend

to stabilise after 2(n-·1) comparisons had been supplied ++) (except
when some distinct inconsistency occurred which with our method could
be located).
In fact the solution for G.L.S. after input of 2(n-1) comparisons was
A1 = 18.58% A2 = 3.44% A3= 2.18% A4= 19.18%

AS= 16.34% Afi= 7.77% A7= 3.38% A8= 7.33% A9= 21.80%
indicating the correct ranking of all attributes and little discrepancy

to the final solution. (Lack of space prevents a complete presentation
of all 29 successive solutions).
We have done many more applications using our analysis with very simi-
+ It can in fact be shown that in general G.L.S. is more stable to

input variations than the E.M.
++ This is roughly equivalent to the total number of 'judgements for two
frames of reference. Comparisons with respect to one frame of
reference usually are not sufficient to generate satisfactory
results.
278
lar results. They confirm that requesting more comparisons in general

does not 'improve' the results but often causes a deterioration of
people's concentration and hence tends to impair the quality of their
judgements.
6. Conclusion
We have discussed some shortcomings of the Eigenvector Method and con-

trasted this with a new Geometric Least Square approach. The message
apparently is that our technique (which is based on various empirical
observations) has several advantages over the E.M. :
(1) Far less data may be required

(2) A more comprehensive analysis of the inconsistencies can be
derived
(3) A more frequent and detailed feedback of the results to the de-
cision maker is possible
(4) The spread of the results seems to be more acceptable
More importantly, it also offers the potential of a research tool in de-

cision making permitting investigation into the behaviour of the deci-
sion maker towards his/her own inherent uncertainty.
279
APPENDIX (Comments on the error analysis in TABLES 1-4)
Using our method there are basically two procedures to assess inconsis-
tencies.
First, since we apply a least squares approach to calculate our solu-

tions, we can use the distance (of the comparisons to this solution)
as an indicator of how good a comparison is with respect to the solu-
tion.
Second, we can produce a succession of results for each attribute and

therefore are able to calculate a standard error of the previous re-
sults with respect to the last available one (that is, we apply pre-
cisely the formula for standard deviation except that we do not use
square deviation from the mean but from the last solution, which in
our procedure is the best available estimate (Islei and Lockett (1986».
In TABLE 1 the latest comparison supplied is : a(3,5)=5 (which is

a 35 =5 in our notation).The distance of this comparison a 35 to the
previous result (not on this printout) generated by the comparisons
a 12 , a 13 , a 14 , a 15 , a 23 , a 24 , a 25 , a 34 is given as :D[a(3,5»)= 1.53.+)
The next figures give the latest calculated results for the attributes
A 1 , . . . , A5 (in percent).
This is followed by the standard error E(i) for each attribute Ai (the
figures in the brackets are the standard errors expressed as percent
of the actual results).
Below this are listed the distances of each individual input to the
final solution,
e.g. INPUT a(2,4)= 1 (which is a 24 =1 in our notation) is with

D[a(2,4»)= -3.74 the least compatible one.
We are also able to use the average distance of all direct comparisons
as an indicator of the inconsistencies.
(A comparison of results using data from previous case studies (Lockett

and Hetherington (1983) with random inputs for different sized matrices,
average distances derived from them and the dispersion around the aver-
age will be discussed in a forthcoming paper) .
+) Note that this distance is not precisely the Euclidean distance but
is enlarged by a factor of 100 and retains a sign (indicating if the
input comparison has to be increased (+) or decreased (-) in order to
be more compatible with the calculated solution).
280
REFERENCES
Cogger, K.O. and Yu, P.L. (1983), "Eigen Weight Vectors and Least
Distance Approximation for Revealed Preference in Pairwise Weight
Ratios", ORSA/TIMS Conference.
Coombs, C.H. (1964), "A Theory of Data", .Tohn ltliley, New York.
Islei, G. and Lockett, A.G. (1986), "An Approach to Preference Vector
Derivation Using Geometric Least Square", Proceedings of VIIth
International Conference of M.C.D.M.
Kahneman, D. and Tversky, A. (1982), "Judgement under Uncertainty
Heuristics and Biases", Cambridge University Press, Cambridge.
Keeney, R.L. and Raiffa, H. (1976), "Decisions with Multiple Object-
ives : Preferences and Value Tradeoffs", John Wiley, New York.
Lockett, A.G. and Hetherington, B. (1983), "Subjective Data and MCDM",
in "Essays and Surveys on Multiple Criteria Decision Making", Ed.P.
Hansen, Springer Verlag, pp.247-259.
Lockett, A.G., Stratford, M., Cox, B., Hetherington, B., and Yallup,P.
(1986) "Modelling a Research Portfolio Using AHP: A Group Decision
Process", R&D Management, May.
Saaty, T.L. (1980), "The Analytic Hierarchy Process", McGraw-Hill,
New York.
Saaty, T.L. and Vargas, L.G. (1984), "Inconsistency and Rank Preserva-
tion", Journal of Math. Psych., 28, pp 205-214.
281
(0,0,1)
/
/
(0,1,0)
-- - /
/
/
/ ~ Incon~1.tent. Ca ••
/
/
/
282
TABLE 1
DISTANCE OF THIS COMPARISON TO LAST SOLUTION D[a( 3 , 5 )]- -1.53
1 ) A1 5.88%
2 ) A2 36.50%
3 ) A3 12.40%
4 ) A4 41.79%
5 ) A5 3.43%
STANDARD ERROR
E( 1 )- 0.31 5.2%)
E( 2 ) .. 1.46 4.0%)
E( 3 ) .. 0.95 7.7%)
E( 4 )= 1.23
E( 5 ) 0.58 .. 2.9%)
17.0%)
DISTANCE OF COMPARISONS TO SOLUTION

D[a( 1 2 ) ]- 0.66 YOUR INPUT a( 1 , 2 )= .1428571
D[a( 1 3 )]- -0.29 YOUR INPUT a( 1 , 3 )- .5
D[a( 1 4 ) ] - 0.65 YOUR INPUT a( 1 4 )- .125
D[a( 1 5 )]= -0.44 YOUR INPUT a( 1 , 5 )- 2
D[a( 2 3 )]- -0.22 YOUR INPUT a( 2 3 )= 3
D[a( 2 4 ) ] = -3.74 YOUR INPUT a( 2 4 )- 1
D[a( 2
D[a( 3 ..
5 ) ]- 1.12
4 ) ] , 1.90
YOUR
YOUR
INPUT
INPUT
a(
a(
2
3
, 5 )-
, 4 )-
8
.25
D[a( 3 5 )]= -0.93 YOUR INPUT a( 3 5 )- 5
AVERAGE DISTANCE = 1.11
TABLE 2
DISTANCE OF THIS COMPARISON TO LAST SOLUTION D[a(4.5)]= 4.83
1 A1 6.89%
2 A2 37.94%
3 A3 13.97%
4 A4 36.23%
5 A5 4.97%
STANDARD ERROR
E( 1 ) .. 1. 23 17.8%)
E( 2 )= 0.80 2.1%)
E( 3 )- 2.31 16.6%)
E( 4 )- 5.64 15.6%)
E( 5 )= 1.81 36.4%)

D[a( 1
D[a( 1
2 )]- 1.46
3 )]= -0.08
YOUR
YOUR
INPUT
INPUT
a(
a(
1
·
2 )= .1428571
1 , 3 )- .5
D[a( 1 , 4 )]= 2.34 YOUR INPUT a( 1 • 4 ) ... 125
D[a( 1 , 5 )]- -1.37 YOUR INPUT a( 1 • 5 ) .. 2
D[a( 2 3 )]= -1.25 YOUR INPUT a( 2 • 3 )- 3
D[a( 2 4 )]- 1. 20 YOUR INPUT a( 2 4 )- 1
D[a( 2 5 )]= -0.23 YOUR INPUT a( 2 5 )= 8
D[a( 3 4 )]= 4.76 YOUR INPUT a( 3 , 4 ) - ,25
D[a( 3 , 5 )]= -2.14 YOUR INPUT a( 3 , 5 )= 5
D[a( 4 5 )]- 2.23 YOUR INPUT a( 4 5 )= 5
283
TABLE 3
DISTANCE OF THIS COMPARISON TO LAST SOLUTION D[a( 8 • 9 )]- 4.66

( 1 A1 19.20%
( 2 A2 4.55%
( 3 A3 3.41%
( 4 A4 21.41%
( 5 A5 14.12%
( 6 A6 9.73%
( 7 A7 4.00%
( 8 A8 6.53%
( 9 A9 17.05%
STANDARD ERROR
E( 1 )- 0.81 4.2%)
E( 2 )= 0.83 18.2%)
E( 3 )- 1.01 29.5%)
E( 4 ) = 1.18 5.5%)
E( 5 )- 2.75 19.5%)
E( 6 )- 1.92 19.8%)
E( 7 )= 0.72 18.1%)
E( 8 ) .. 1. 70 26.1%)
E( 9 )- 4.53 26.6%)
·· ·• 32 )) =.. 77
D[a( 1 2 )]- -1-.79 YOUR INPUT a( 1
D[a( 1 3 )]=-0.66 YOUR INPUT a( 1
D[a( 1 4 )]--1.56
D[a( 1
· 5 )].. 3.59
YOUR
YOUR
INPUT
INPUT
a(
a(
1
1
•
•
4
5
)- 1
)- 1
D[a( 1 6 )] .. -3.15 YOUR INPUT a( 1 ) .. 3
··
D[a( 1 7 )]'" -1.25 YOUR • 6 ) .. 7
INPUT a( 1 • 7
D[a( 1 8 )]- -3.75 YOUR INPUT a( 1 ) .. 7
D[a( 1 9 ) ] - 1.53 • 8
YOUR INPUT a( 1 • 9 )- 1
D[a( 2
D[a( 2
· ....
3 )]- -2.45
4 )] 1.48
YOUR
YOUR
INPUT
INPUT
a(
a(
2
2
•
•
3
4
) .. 5
) ... 1428571
Dla( 2 5 ) ] 1.69 YOUR INPUT a( 2 )- .2
··
• 5
Dla( 2
D[a( 2 ..
6 ) ]= 1.24
7 ) ]
.. 0.39
YOUR
YOUR
INPUT
INPUT
a(
a(
2
2
•
•
6
7
)= .3333333
)- 1
Dla( 2 8 ) ] 3.58 YOUR INPUT a( 2 • 8 ) ... 1428571
D[a( 2
Dla( 3 ..
9 ) ] = 2.64
4 ) ] 1.02
YOUR
YOUR
INPUT
INPUT
a(
a(
2
3
• 9
4
)= .1111111
) ... 1111111
Dla( 3 5 ) ] = 1.37 YOUR INPUT a( 3 ) - .1428571
··
Dla( 3 6 )] = 1.43 • 5
YOUR INPUT a( 3 • 6 ) - .2
D[a( 3 7 )] = -0.42 YOUR INPUT a( 3 • 7 )- 1
Dla( 3
D[a( 3 ·· 9 )]..
8 ) ]- 1.16
1.50
YOUR
YOUR
INPUT
INPUT
a(
a(
3
3
•
•
8
9
)- .3333333
)- .1111111
Dla( 4 5 )]- -6.63 YOUR INPUT a( 4 • 5 ) .. 3
D[a( 4
Dla( 4 , · 6 )]= -5.34
7 )] = -0.93
YOUR
YOUR
INPUT
INPUT
a(
a(
4
4
•
,
6
7
)- 5
) .. 7
Dla( 4
Dla( 4 ,
D[a( 5
· 9 )]..
8 )]- -2.21
3.09
YOUR
YOUR
INPUT
INPUT
a(
a(
4
4
• 8
9
)- 5
)- 1
6 )]- -4.76 YOUR INPUT a( 5 6 ) .. 3
D[a( 5 7 )] .. -1.15 YOUR INPUT a( 5 7 ) .. 5
D[a( 5
D[a( 5
Dla( 6 · ..
8 )]= -3.64
9 ) ] 8.01
7 )]= -0.72
YOUR
YOUR
INPUT
INPUT
a(
a(
5
5
,
•
8
9
)- 5
)= .3333333
YOUR INPUT a( 6 7 )- 3
Dla( 6 8 )]=-4.50 YOUR INPUT a( 6 , 8 ) .. 5
Dla( 6 , 9 )]=-14.81 YOUR INPUT a( 6 ) .. 5
D[a( 7 8 ) ]= 2.64 • 9 ) ... 2
YOUR INPUT a( 7 8
Dla( 7 9 )] = 1.55 YOUR INPUT a( 7 , 9 ) ... 1428571
Dla( 8 , 9 ) ]- 4.06 YOUR INPUT a( 8 , 9 ) ... 1428571
284
'I'ABLE 4
DISTANCE OF THIS COMPARISON TO LAST SOLUTION D[a( 8 , 9 )] = 2.54
1 A1 18.75%
2 A2 3.73%
3 A3 2.66%
4 A4 21.41%
5 A5 13.57%
6 A6 7.87%
7 A7 3.21%
8 A8 5.64%
9 A9 23.15%
STANDARD ERROR
E( 1 )= 0.69 3.7%)
E( 2 ) = 0.25 6.7%)
E( 3 ) = 0.34 12.8%)
E( 4)= 1.18 5.5%)
E( 5 )= 3.10 22.8%)
E( 6 )= 0.52 6.7%)
E( 7 )= 0.17 5.3%)
E( 8 )= 1.63 28.9%)
E( 9 )= 1.59 6.9%)
D[a( 1 2 )]= -1.04 YOUR INPUT a( 1 , 2 )= 7
D[a( 1 , 3 )] =' 0.02 YOUR INPUT a( 1 3 )= 7
D[a( 1 4 )]= -1.89 YOUR INPUT a( 1 4 )= 1
D[a( 1 5 )] = 3.66 YOUR INPUT a( 1 5 )= 1
D[a( 1 6 )]= -1.54 YOUR INPUT a( 1 6 )= 3
D[a( 1 7 )]= -0.53 YOUR INPUT ae 1 , 7 )= 7
D[a( 1 8 »)= -2.94 YOUR INPUT a( 1 8 )= 7
D[a( 1 9 »)= -3.12 YOUR INPUT a( 1 9 )= 1
D[a( 2 , 3 )] = -1.87 YOUR INPUT a( 2 , 3 )= 5
D[a( 2 4 )] = 0.66 YOUR INPUT a( 2 , 4 )= .1428571
D[a( 2 5 )] = 1.00 YOUR INPUT a( 2 5 )= .2
D[a( 2 6 )] = 1.05 YOUR INPUT a( 2 6 )= .3333333
D[a( 2 7 )] = 0.37 YOUR INPUT a( 2 , 7 )- 1
D[a( 2 8 )]= 2.90 YOUR INPUT a( 2 8 )= .1428571
D[a( 2 9 )] = 1.15 YOUR INPUT a( 2 9 ) = .1111111
D[a( 3 4 )] = 0.28 YOUR INPUT a( 3 4 ) = .1111111
D[a( 3 5 )] = 0.71 YOUR INPUT a( 3 5 )= .1428571
D[a( 3 6 )] = 1.06 YOUR INPUT a( 3 6 )= .2
D[a( 3 7 )]= -0.39 YOUR INPUT a( 3 7 )= 1
D[a( 3 8 )]= 0.74 YOUR INPUT a( 3 8 )= .3333333
D[a( 3 9 ) ] = 0.08 YOUR INPUT a( 3 , 9 )= .1111111
D[a( 4 5 )]= -6.10 YOUR INPUT a( 4 , 5 ) = 3
D[a( 4 6 )]= -3.52 YOUR INPUT a( 4 6 )= 5
D[a( 4 7 )]= -0.15 YOUR INPUT a( 4 7 )= 7
D[a( 4 8 ») = -1. 33 YOUR INPUT a( 4 8 )= 5
D[a( 4 9 )]= -1.23 YOUR INPUT a( 4 9 )= 1
D[a( 5 6 ) ] = -3.18 YOUR INPUT a( 5 6 )= 3
D[a( 5 7 )]= -0.49 YOUR INPUT a( 5 7 )= 5
D[a( 5 8 »)= -2.87 YOUR INPUT a( 5 8 )= 5
D[a( 5 9 )] = 5.55 YOUR INPUT a( 5 9 ) = .3333333
D[a( 6 7 )]= -0.56 YOUR INPUT a( 6 7 )= 3
D[a( 6 8 )]= -3.99 YOUR INPUT a( 6 8 )= 5
D[a( 6 9 ») = 3 .18 YOUR INPUT a( 6 9 )= .2
D[a( 7 8 )]= 2.04 YOUR INPUT a( 7 8 ) = .2
D[a( 7 9 ) ] = -0,09 YOUR INPUT a( 7 9 )= .1428571
D[a( 8 9 )] = 2.31 YOUR INPUT ae 8 9 )= .1428571
AVERAGE DISTANCE = 1. 77
AGGREGATION PROCEDURES FOR HIERARCHICALLY GROUPED
DECISION ATTRIBUTES WITH APPLICATION TO
CONTROL SYSTEM PERFORMANCE EVALUATION
Lj. Vlaci6 1 , A. Wierzbicki 2 and B. Mati6 1

lInstitute for Control and Computer Sciences - IRCA, University of
Sarajevo
2Institute of Automatic Control, Technical University of Warszaw
1. INTRODUCTION: HIERARCHICALLY GROUPED

ATTRIBUTES OF A DECISION PROBLEM
It is a known fact in decision theory that attributes of decisi-

ons often have hierarchical structure and that the attributes of a lo-
wer level must be grouped and aggregated into an attribute of the upper
level. Until now, however, the theoretical foundations of such aggrega-
tion require a rather complicated testing of various independence ass-
umptions - see Keeney and Raiffa (1976) - hence most of applied examp-
les of such aggregation are based on heuristics. We shall consider here,
first, an outline of an example of such hierarchical structure of att-
ributes, and then proceed to alternative theoretical foundations of the-
ir aggregation and to more detailed description of this aggregation in
the example considered.
The example concerns the selection of alternative configurations
of a integrated control system - see Vlaci6 and Mati6 (1984, 1985). For
comparing alternative microcomputer-based process control systems, the
following main attributes were proposed:
- a composite attribute evaluating the dynamic properties of the system;
- a composite attribute evaluating the flexibility and uniformity of sy-
stems algorithms;
- a subjectively assessed attribute corresponding to system safety and
redundancy;
- a subjectively assessed attribute expressing the quality of systems
design and fUnctional configuration;
- a subjectively assessed attribute expressing the quality of systems
documentation;
- subjectively assessed attribute corresponding to the standardization
of in-plant equipment.
286
Integrated
Microcomputer-based
Process Control System
I-- Dynamic ___________________-,__ System Utilization

Properties Factor
Completeness of system System Response
I-- associated functions Time
Flexibility & Uniformity Extent of System
-- of Algorithms Cen tra li za tion
_ Re liabi li ty Safe ty &
Redundancy
__ Design & Functional
Configuration
-- Discontinuity of Process Flow
-- Documentation
-- Standardization of In-plant
Equipment
the main attributes subattributes
FIGURE 1. Hierarchically Grouped Decision Attributes
other attributes can be also considered - see Fig. 1; the impor-

tant point, however, is that most of the attributes can be only subjec-
tively assessed, while some of the attributes express composite evalua-
tions of several more detailed properties of the alternative systems
that can be analytically modelled and thus expressed quantitatively.
For example, the composite attribute "dynamic properties" expresses jo-
intly the evaluation of the following second-level attributes that can
all be modelled analytically - see Vlacic (1986):
- a system utilization factor;
- system response time;
- extent of system centralization.
We shall consider here a decision process in which several alter-

native microcomputer-based process control systems are evaluated by a
committee of experts; full descriptions of the alternatives together
with the corresponding values of lower-level attributes, obtained with
help of an analytical model, are presented to all committee members. The
questions adressed in this paper are how to assist each committee mem-
ber in aggregating the lower-level attributes into an assessment of the
composite upper-level attribute, then how to assist him in aggregating
287
all subjective assessments of upper-level attributes into an overall

assessment of an alternative, finally, how to organize the decision-ma-
king process in the committee in a way that supports information exchan-
ge and discussions between committee members and leads to an agreement
on the selection of the best alternative.
2. AGGREGATION BY CARDINAL UTILITY ~UNCTI0NS
Aggregation procedures based on cardinal utility theory have wery

well developed theoretical foundations ~ see Keeney and Raiffa (1976)
for details. We recall here only that a constructive assessment of a
multiattribute value function or, in the case of decisions under uncer-
tainty with probabilistic representation of outcomes, of a multiattri-
bute cardinal utility function requires an experimental testing of va-
rious independence assumptions. The simplest and most general of these
independence assumptions is preferential independence for each attribu-
te: in most applied cases, we can choose the attributes in such a way
that the preference ordering for each attribute while other attribute
values are fixed is independent on the level of other attributes. Under
this assumption, an overall value function aggregating all attribute
assessments can be written as a composition of individual value functi-
ons for each attribute:
(1 )
where p is the number of attributes, the a(:lsessment value of an attri-

bute denoted by Yi' Y = (Yl"'Yi"'Y p ) is the vector of attribute valu-
es and v.1 ('1.)
1
are individual value functions:
. •
the function h: RP ~ Rl
is an arbitrary (strictly) monotone function.
A much stronger assumption is that of mutual preferential indepen-
dence or, eqUivalently but much better expressing the essence of this
assumption, of pairwise preferential independence. ~or two attributes
Yl' Y2' this amounts to the assumption that the tradeoff coefficients
tr(Yl,Y2) - that is, the values of deterioration of one attribute, say
Yl' one is willing to .accept in order to improve another attribute, '12'
by a unit-satisfy the following proportionality condition:
(2)
for each Y1'Y2,Y1'Y2 in the range of values of attributes. For more a-

ttributes, the assumption of pairwise preferential independence amounts
to the requirement that the relation (2) holds for each pair of attri-
butes independently of the values of other attributes. Pairwise prefe-
288
rential independence holds if and only if the overall value function

can be represented (possibly, after a monotone transformation) in the
additive form:
p
v (y) = E v. (y. ) (3)
i=l ~ ~
However, testing pairwise preferential independence is a task of

fast growing complexity with the number of attributes; typically, such
independence is just heuristically assumed. This might be related to
the fact that theorems of Debreu (1959) were sometimes misinterpreted
to indicate that preferential independence for each attribute implies
pairwise preferential independence. The following example shows that
such an interpretation is not valid.
Suppose all attributes are evaluated on a uniform dimensionless
scale 0 .•• 10 and consider the following overall value function:
p
v(y) = min Yi + 0.5 E Yi (4)
l~i;;:p i=l
Such a form of a value function is rather plausible: the decision
maker attaches importance to the minimal individual value of attribute
assessment while compensating with some coefficient for the average of
individual attribute assessments; in fact, we shall use a similar con-
struction to obtain a nonstationary approximation of cardinal utility
functions of decision makers. Consider p = 2 and take Yi = 2~ Y2 = 3,
y! = 4~ Yi = 5; it is easy to check that trey!, Y2) = trey!, Yi)
=tr(y!,yi) = 1/3 whereas trey!, Y2) = 3 z hence the condition (2) does
not hold and the overall value function (4) cannot be represented in an
additive form; neither can it be represented in an additive form if p=3
or more, although this function is preferentially independent for each
attribute. Thus, the assumption of pairwise preferential independence
is quite strong and we cannot expect it to be satisfied in some typical
decision situations.
Cardinal utility functions are a more sound basis for aggregating
attributes than general value functions. Cardinal utility functions are
such value functions which preserve, in the probabilistic expectation
sense, given preferences between probability distributions of consequ-
ences of uncertain decision alternatives under any positively monotono-
us and linear or affine transformations of attribute assessment scales.
Thus, such functions are themselves independent (up to a positively mo-
notone affine scaling transformation) of any positively monotone affine
transformation of their arguments, and any function that possesses such
independence properties can be interpreted as a cardinal utility function.
289
In order to select appropriate classes of utility functions for

assessment in applied cases, vartous additional independence conditions
have been postulated. The most stringent is that of additive independen-
ae~ equivalent to postulating an additive form of multiattribute cardi-
nal utility function. Less exacting condition is that of mutuaZ utiZity

independenae~ where the preferences on lotteries on any subset of att-
ributes do not depend on the values of assessments of other attributes;
this is equivalent to a multiplicative form of multiattribute cardinal
utility function:
(5)
However, testing such independence assumptions and assessing pa-

rameters in corresponding utility functions in applications to group
decision making requires a large number of pairwise comparisons of hy-
pothetical alternatives or of lotteries of such alternatives performed
by each decision maker in a committee. Experience shows that most deci-
sion makers do not,want to spend their time in answering many hypothe-
tical questions; in order to distort and break a decision process in a
committee of experts, it is sufficient that one committee member takes
such attitude. Thus, simpler and more practical means of communicating
with decision makers in a committee are needed.
Another, perhaps more important drawback of aggregation methods
based on unmodified value theory and cardinal utility theory is that
these methods assume stationarity of value functions or utility functi-
ons: it is assumed that these functions do not change in time, are al-
ready well formed. However, many recent studies of principles of deci-
sion support - see, e.g., Dreyfus (1985) - show that one of the most
important characteristics of a decision process is Zearning by deaision
makers. If a committee member is gathering new information and learning
during discussions in the committee, he most probably changes his uti-
lity function. It is quite impractical to assume that, firstly, the
utility functions of all committee members will be laboriously assessed,
then the results will be discussed in the committee with the most pro-
bable outcome that the utility functions must be assessed anew.
Therefore, for the purpose of simple but plausible approximations
of nonstationary utility functions under changing preferences, a speci-
al class of aahievement funations has been proposed - see Wierzbicki
(1982, 1985). Lewandowski, Johnson and Wierzbicki (1986). These functi-
ons can have cardinal form and are preferentially independent in each
attribute, but not pairwise preferentially independent nor mutually u-
tility independent. The main ways of communication with decision makers
290
are then not pairwise comparisons of alternatives or lotteries, but the

elicitation of decision maker~s aspiration levels (possibly, together
with reservation levels) for each attribute. Much research in the beha-
vioral aspects of decision processes - see Simon (1958), Tietz (1985) -
indicates that such parameters characterizing preferences are easily
understood and accepted by decision makers.
3. AGGREGATION BY ORDER-CONSISTENT ACHIEVEMENT FUNCTIONS
Consider a decision maker who, supported by a team of analysts or

a decision support system, faces a problem of choosing a decision out
of an admissible set Xo and has specified p attributes; without loss of
generality, suppose that all attributes of decision improve if their va-
lues increase. Assume that the decision maker, either relying on his
own experience or supported by the team of analysts, has established
(judgementally or analytically) a mapping f: Xo + RP that specifies the
values of attributes for each alternative (the outcomes of decisions)
and has learned about the ranges of attainability (Yi,min; Yi,max) of
each attribute. These ranges need not be very tight bounds of the set
of attainable attribute values Yo = f (X o ); they might, in judgemental
evaluations, correspond just to the scale of attribute assessments which
can be simply specified as between 0 and 10.
Suppose that, in this situation, the decision maker specifies his
reservation level i~~ and his aspiration level ~~,
~
where these levels
satisfy Yi,min<~i<Y~<Yi,max' for each attribute, i = 1, ... p. What can a
team of analysts or a decision support system conclude on the basis of
such information about the preferences of the decision maker?
One way would be to select an established theoretical tool; for
this type of information, an appropriate tool would be the theory of fu-
zzy sets - see, e.g., Sakawa (1983). Membership functions for the asse-
ssment of satisfying decision makers requirements on each outcome could
be postulated in the form:
0, if Yi
:::;. i'i '.
-,
(y;
k
- i~)/(~'.'
~ ~
- y~),
~
if Yi < Yi < ~"i '. (6)
-<
-II
1, if Yi Yi;
The convolution of these membership functions could be interpreted in

the sense of minimum operation:
291
(7)
However, the logic of fuzzy sets is stil too sharp to describe

fully the preferences of the decision maker: the membership function
does not describe his dis utility of not reaching his reservation levels,
nor his possible utility of reaching more than his aspiration levels -
- this would require an extention of the membership function (6,7).
Such specific utility function, explicitely dependent on the aspiration
and reservation levels of the decision maker, can be constructed when
accepting following assumptions:
(i) The decision maker prefers outcomes that satisfy all his reserva-
tion levels to any outcome that does not satisfy at least one of
his reservation levels; similarly for aspiration levels;
(ii) The satisfaction of the decision maker at reaching (all, or the
last of) his reservation levels can be measured by 0, while his
satisfaction at reaching (all, or the last of) his aspiration le-
vels can be measured by 1;
(iii) The satisfaction of the decision maker at reaching the maximum of
the range of all outcomes can be measured by 1 + a, where a ~ 0
is a parameter (if a = 0, then the decision maker behaves in a
strict satisficing way); the dissatisfaction of the decision ma-
ker at reaching the minimum of the range of at least one of the
outcomes can be measured by - b, where b > 0 is another parame-
ter,
(iv) Since all available informations for the construction of this spe-
cial utility function has been already used, the simplest form of
this function that would satisfy (i), (ii), (iii), obtained thro-
ugh linear interpolation, is postulated.
Such a function has the following form:
b«Yi-Yi,min)/(Yi-Yi,min)-I) , if Yi,min ~
Yi ~ y!1
Ui(Yi) (Yi-Yi)/(Yi-Yi) , if y!1 <
Yi < -n
Yi (8)
a(Yi-Yi)/(Yi,max-Yi)+I, if Y~1 ~ Yi ~
Yi,max
while the aggregation is defined as in (7):
(9)
The above function belongs to the class of order representing aa-

hievement funations, see Wierzbicki (1977, 1982, 1986) for axiomatic
292
definitions of this class; its level sets coincide with the shifted po-
sitive cone RP+ that defines the partial preordering of the attribute
space. This function has several interpretations. One of them is a L-
-shaped utility function, consistently summarizing the information con-
tained in the points Yi,min' yi, yi,
Yi,max' and thus serving as an
approximation to the preferences of the decision maker. Observe that
this particular function is a cardinal utility function, since it is
defined by ratios of intervals and thus independent of any affine trans-
formations of outcomes; therefore, it can be even used in interpersonal
comparisons of utility in a committee.
Another interpretation of this function is a transformed and we-
ighted (with changing, but piece-wise constant weighting coefficients)
Chebyshev or 100 norm of the difference between the point Ymax and the
actual attribute assessment y = f(x). To ilustrate this interpretation
more clearly, consider a case when the points Ymin' y' are not specifi-
ed and denote y" = y; the function (8,9), after an affine transformati-
on, simplifies then to the form:
s(y,y) = min (Yi - Y·)/(y.

l~i:::p ~ ~ ,max
- y.)
~
- 1 (10)
which can also be written in the form:
(11)
-
Since we assume that Yi < Yi,max for all i = l, ••. p, the achieve-
ment function corresponds in this case to the weighted Chebyshev norm
with changed sign: however, the weights Ai are not specified explicite-
ly by the decision maker, but defined implicitely through his statement
of aspiration levels y as compared to the upper bound point Ymax. In-
terpreted as an achievement function, this function was used, for exam-
ple, by Wierzbicki (1984) and Lewandowski, Johnson, Wierzbicki (1986);
as a Chebyshev norm, by Nakayama (1984) and Steuer and Choo (1983). The
case when a decision maker specifies a reservation and an aspiration
point was also investigated, though not in the form of the achievement
function (8,9), by Gorecki et al. (1984), and by Weistroffer (1984).
The simplification (10,11) of the achievement function (8,9) is
not the only one possible. If the deCision maker knows both upper bo-
und point Ymax and the lower bound point Ymin but sp~cifies only one
reference point y, this point should be interpreted as an aspiration
level point rather than reservation level point; but in mathematical
modifications of (8,9) we must treat it as reservation levels, y y',
and let the aspiration levels coincide with the upper bounds, y" = Ymax
293
(since the function (8) would become di~cont~nuous, if we would let

y' = y"). If both aspirat~on and reservation levels are specified but
the lower bounds are not available, the definition of the function (8)
can be also suitable modified.
The achievement function (8,9) or (10,11) has, however, one dis-
advantage. It is not strictly monotone with respect to the attribute
evaluations Yi = fi(X) t in fact, it is constant when one of the attri-
bute evaluations increases much above other attribute evaluations that
are kept constant. This implies, theoretically, that the maxima of this
function might be not efficient but only weakly efficient and, behavio-
rally, that the decision maker does not pay any attention to overachi-
evements, however large, in some attributes, as long as other attribu-
tes show underachievements. Since most decision makers would allow at
least some degree of compensation of underachievements by large over-
achievements in other attributes, it is necessary to modify the functi-
on (8,9). We can do it when adding the following postulate to the list
(i), (ii), (iii), (iv):
(v) If an attribute shows underachievement when compared to its re-

servation (or aspiration) level, and other attributes show over-
achievements, the decision maker is willing to accept a compen-
sation of the underachievement by the average overachievement in
other attributes (all measured relative to the scales implied by
points Yi,min' yi, yi, Yi,max) with a weighting coefficient c,
where 0 < c < p.
This postulate leads to the following form of the achievement

function which belongs to the class of order approximating aahievement
funations (whereas order representing and order approximating achieve-
ment functions are jOintly called order-consistent achievement functi-
ons, see Wierzbick~ 1986):
p
s(y,y',~") =u(y) = (minu.(y.) + (c/p) E u.(y.»/(l +c) (12)
l~iS.p ~ ~ i=l ~ ~
with ui(Yi) defined as in (8). This achievement funct~on is also a car-

dinal utility function. Its maxima are not only efficient, but also
properly efficient (with a priori bounded trade-off coefficients, see
Wierzbicki (1986». Its level sets approximate from outside the shifted
positive cone RP+" Beside these properties, it has all the interpreta-
tions of the function (8,9).
Functions of similar types have been extensively used in so called
DIDAS decision support systems, see Grauer, Lewandowski and Wierzbicki
(1982). Here, however, we are interested in the application of the
294
specific form (12) ~n a group decision process ot $electing between

discrete alternatives.
4. GROUP DECISION PROCESS AND SCDAS METHOD
We consider here a decision process similar as described by Le-

wandowski, Johnson and Wierzbicki (1985) in the SCDAS (Selection Co-
mmittee Decision Analysis and Support) system; however, we use here
consequently both aspiration and reservation levels for all attributes
and consider here a hierarchy of attributes as described in the intro-
duction.
Let us concentrate first on aggregating lower-level attributes.
Since these attributes are modelled analytically for each alternative,
the ranges ymin~ Ymax for these attributes can be computed and presen-
ted to all experts in the committee. On the basis of this information
and their own experience, the committee members can form own aspiration
and reservation.levels that should be communicated to a decision supp-
ort system. For each committee member and each alternative, the decisi-
on support system can compute then the values of the achievement func-
tion (12), treated in this case as proposed evaluations of the corres-
ponding composite upper-level attribute. For this purpose, it is useful
to assume a = b = 1 and multiply the values of (12) by 10, thus obtai-
ning a scale of evaluations between -10 (if some attributes are at the
lower bound of their assessment scale) through 0 (if some attributes
are at reservation levels) and through 10 (if all attributes satisfy
aspirations) up to 20 (in the unlikely case when all attributes are
evaluated at the upper bounds of their assessment scale). At the same
time l the decision support system presents to the committee member a
ranking of alternatives according to this composite attribute. If the
committee member disagrees with the ranking or the proposed evaluations
of the composite attribute for any alternative, he can either change
his aspirations or reservations, or move to change the parameter c in
(12), or to use his own evaluations of this attribute in the next stage
of the decision process - where anyway all upper-level attributes are
assessed judgementally.
Further stages of the deCision process can be organized similarly
as in SCDAS system. These stages include:
Discussing procedural details and ways of aggregating opinions in the
committee; discussing and setting scales of assessment for each upp-
er-level attribute.
295
- Specifying aspirations and reservations for all upper~level attribu-

tes by all committee members; computing special disagreement coeffi-
cients in the decision support system and discussing in the committee
the differences of opinions in all cases when these coefficients are
high. agreeing on the use of aggregated aspirations and reservations
in the committee.
Scoring, that is, assessing each attribute of each alternative by

each committee member. individual work of each committee member with
the decision support system that computes and displays various ran-
kings of alternatives - according to scores on various attributes or
values of function (12), this time used to aggregate scores for all
upper-level attributes of a given alternative - together'with appro-
priate graphical interpretations.
- Aggregating opinions across the committee - firstly, by computing

disagreement coefficients on scores and discussing differences of o-
pinions about alternatives in cases when these coefficients are high
and~ secondly, by taking averages across the committee of function
(12) interpreted as a cardinal utility function, for each alternati-
ve; these averages serve as the basis for preparing an overall ran-
king of alternatives for a final discussion in the committee.
The full adapted procedure of SCDAS method is shown in Fig. 2.
In following, we go through this procedure in two stages: first for the
lower-level attributes, then for the upper-level attributes.
5. LOWER-LEVEL AGGREGATION: THE EVALUATION OF THE

MULTICRITERIA MODEL FOR THE JUDGEMENT OF A
DISTRIBUTED PROCESS CONTROL SYSTEM STRUCTURE
Step 1: Defining the proaedure. In procedural planning, we stress only

that a committee of K=5 committee members with equal voting factor v k
was considered, and formulars such as shown in Tables T.1. and T.2. we-
re used to assist the decision process.
Step 2: Betting and saaling attributes. The following criteria were se-
lected for the evaluation of the hierarchical structural form of a con-
trol system (on the lower level of attributes):
System utilisation factor (p) - in the sense of definition and analy-

tical relation in Vlacie and Matie (1983).
System response time (T r ) - in the sense of definition and analytical

relation in Vlacie and Matie (1983).
296
FIGURE 2. petri-net of SCDAS structure

297
Extent of system central~zat~on (0) - in the sense of definition and

analytical relation ~n Cv~rkun (1982).
The unified scale (0,10), where value 10 represents the excellence and
value 0 the unacceptable, was selected.
T.1.
TABLE OF ASPIRATIONS
E CRITERIA C1 C2 C3
j +
4-k
DECISION
MAKERS
p Tr 0
D1
y' 5 7 0
Y" 7 9 @)
y' 3 6,5 6
D2
-"
y 6 Q 9
-y'
D3
0 (0 2
-"
y 0 7 4
D4
y' G) 5 ~~
-Y" G) CD CD
D5
y' 7 8 8
-"
y 8 9 9
THE ABSOLUTE fl y' 6 4 9

DIFFERENCE fly " 6 3,5 9
THE MEAN y' 5 6,16 5,33

ASPIRATION -Y" 7 8,33 7,33
LEVELS
.DISAGREEMENT 2,685 2,5 5,29
INDICATOR
THE UNIFIED 0-10 0-10 0-10
SCALE
A NATURAL. 0-1 O-Tr 0-1
SCALE
T.2.
TABLE OF ATTRIBUTE-ALTERNATIVE ASSIGNEMENTS,INITIAL SETTING OF ASPIRATION LEVELS AND VALUES OF ACHIEVEMENT FUNCTIONS
8(Y,y',y")
j+
k = 1 k = 2 k = 3 k = 4 k = 5
-I- •
~TERNATIVES
~ 7,1j / 7V/ / c=O c=1 c=O c=1 c=O c=1 c=O c=1 c=O c=1
Al (Fig. 3a) y 1,95 7,98 6,7 -6,1 -3,5 -7,6 3 -7,3
A2 (Fig. 3b) Y 5,55 9,65 6,7 -4,4 2,3 -3,1 13,6 -2,1
A3 (Fig.3c) y 4,15 9,35 5,1 -4,4 -1,5 -4,9 11,6 -4,1

I\J
<D
ex>
-I-k y' 5 7
D1 -If
CD
y 7 9
®
y' 3 6,5 6
D2 -If
y 6 I~,~ 9
y' 2
D3 -If
0 0
Y (9) 7 4
y'
D4 -If
CD 5 €,oy
y CD CD 0)
y' 7 8 8
D5 -If
Y 8 9 9
'---
299
Step 3: InitiaZ setting of aspi~ation ZepeZs. During this step, the co~
mmittee members (denoted here by ~ndex k) should be asked to define two

reference
_ points:
_ the reservation point Y~k
J and aspiration point YJ~k
where yjk < yjk' where j denotes the number of attribute. The assigned
aspir"ation levels are contained in Table T .1. In this case the concre-
te values of the aspiration levels y' and y" are determined hypotheti-
cally, as an ilustrat~on of a possible wide disagreement between commi-
ttee members.
Aggregated aspiration level is, in general case:
(13)
A
where kj' k j denote indices of extreme value assignements (circles in
Table T.2.); thus
(14)
y"
The absolute difference of opinions is calculated then according to the

relations:
(15)
under the assumption that the indicies of assigned aspiration levels

are sorted according to decreasing values of aspirations value.
Disagreement indicator is then calculated according to the formula:
DI. (16)
J
where:
'V
Yjk = Yjk + y-J"'k) /2
(-, (17)
All the calculated values are indicated in Table T.l. A high value of
the indicator (DI j >l) implies in principle a discussion between commi-
ttee members and possible correction of th7 assigned aspiration levels.
Occasionally (as is the case here with 0 criterion) the nature of the
criterion implies an apriori the possibility of divergent approaches
to the allocation of aspiration levels.
300
Step 4: Initial survey of alternatives. Three alternatives considered

are shown graphically in Fig. 3.
--"
--"
--"
0)
"
~
0< K1 !: 1
K2~O
-" K1 K2"
--" [[J
n =12(20) r=6(10)
'l\
~
JIIJJ0/S=2
r=6(10)
b)
0< K, ~,
0< K ~ 1
n=12(20)
C)
FIGURE 3. The three alternative structures (topologies)

of a integrated process control system
301
Step 5: ALternative-attribute assignement. Th~s step includes alloca-

tion of attribute values for each of alternative solutions. In this
case, the selected criteria are quantitative by their nature, i.e. ana-
lytically described by formulae derived from queueing theory (see Vla-
cic and Matic, (1986». However, some of parameters of these formulae,
such as mean arrival rates and loading factors, must be assessed empi-
rically. For this reason, we consider here several examples of the
scope of plant automatization.
This scope, which can be characterized by numbers of measuring,
open and closed control loops, type and number of final control elemen-
ts, number and type of signalling and alarm statuses, etc, must be re-
presented by examples drawn from actual industrial plants automatizati-
on and be based on the experience in mastering practical complexities
of such systems. Without describing details of such analysis (see Vla-
cic (1986»~ we give only examples of some actual and one hypothetical
case in Table T.3. The test point "A" was obtained by summarizing the
automatization degree of a 32 MW block; the test points "B" I "0" and
"E" was formed on the basis of performed automatization of an actual
110 MW, 230 MW and 300 MW blocks respectivelly - Energoinvest (1979-85).
Only the test point "e" was hypothetically formed. The differences bet-
ween the test pOints are not only due to the different power of the
thermo-power block but also due to the actually applied automatization
degree. It should be pOinted out that the test point "0" is a case of
with the highest, actually carried out automatization degree.
By choosing the parameters for calculations as implied by the test ca-
ses from T.3. and calculating the values of the attributes or criteria
defined in Step 2, the following qualitative results can be observed
(see Vlacic, Matic (1986»:
increased complexity of the process control system structure does not
lead to a significant increase of the number of control loops which
can be included in the system .
• increased complexity of the system structure does not practically in-
crease its applicability from the aspect of the system response time.
Increased centralization degree increases the controlability of the
system but decreases the degree of automatization covered by the sy-
stem.
These qualitative results help us in a reasonable assignement of attri-
bute values for each alternative. The three alternatives, as specified
in Fig. 3., are considered to represent three possible structures of
automation system for a plant of the type corresponding to the test
302
case "0" from T.3. In order to reduce the uncertainty of attribute va-
lues assignement, we consider here two values. The first value
y'(i,j,k) = y' (i,j) is obtained directly from analytical calculations
for the attribute j and alternative i (independently from the committee
member'k) for parameters resulting from the test case "0". The second
value y"(i,j,k) = y"(i,j) is obtained under the assumption that loading
factors were increased, as compared to the test case "0", by 5%. These
values are shown in Table T.4.
A final value for attribute assignement is obtained by taking the
mean:
y(i,j,k) y(i,j) 0.5(y'(i,j) + y"(i,j» (18)
T.3.
~19~~~'~~">T
NUMBER OF
A B C D E
CLOSED-LOOP 16 50 60 80 90
NUMBER OF
120 260 440 700 750
TRANSMITERS
NUMBER OF
SIGNALLING AND 70 180 310 400 450
ALARMS
NUMBER OF
FINAL CONTROL 44 230 500 800 950
ELEMENTS
T.4.
ALLOCATED ATTRIBUTES VALUES
ALTERNATIVES P Tr <')
PROJECTED +5% PROJECTED +5% PROJECTED +5%
Al 2.5 1.4 8.6 7.3 6.7 6.7
A2 5.8 5.3 9.7 9.6 6.7 6.7

A3 4.3 4.0 9.4 9.3 5.1 5.1
303
step 6: Individual ranking. By apply~ng the ~ormula (12), we can compute

achievement function values under the individual aspiration levels of
each committee member, for each alternative with its assigned attribute
values; the alternatives are then ranked according to these achievement
values. We skip here the details of this analysis.
Step 7. Individual analysis. Assuming that the committee members are

satisfied with individual rankings and it is not necessary to return to
one of the earlier steps, we proceed to Step 8.
Step 8. Group ranking. Table T.S. contains the individual ranking lists
of alternative solutions formed on the basis of achievement function
values in Step 6.
T.S.
TABLE OF INDIVIDUAL AND FINAL DECISIONS
s(y,y' ,y") INDIVIDUAL

+k AND FINAL
ALTERNATIVES
DECISION RANKS
MAKERS A1 A2 A3
D1 -6,1 -4,4 -4,4 A2 , A3 , A1
D2 -3,S -2,3 -l,S A2 , A3 , A1
D3 -7,6 -3,1 -4,9 A2 , A3 , A1
D4 3 13,6 11,6 A2 , A3 , A1
DS -7,3 -2,1 -4,1 A2 , A3 , Al
AGGREGATE -4,3 1,26 -0,66 A2 , A3 , A1
The aggregated achievement function value for each alternative

solution was calculated in Table T.S. as an average of individual achi-
evement values corresponding to each committee member. These aggregated
achievements are then used as the basis of judgemental evaluation of
the composite upper-level attribute "dynamic characteristics of the sy-
stem structure".
304
6. UPPER LEVEL AGGREGATrON: THE EVALUATrON OF

THE MULTICRrTERIA MODEL FOR THE JUDGEMENT
OF OVERALL PERFORMANCE OF A DISTRIBUTED
PROCESS CONTROL SYSTEM
We continue here the analysis under the assumption that it is ne-

cessary to select process control system whose automatization level is
corresponding to the test point "D" from Table T.3. (a thermo-power
block, capacity of lx230 MW with control at the functional group levels).
Step 1: Defining the prooedure, K = 5,vk = 1 for all k = 1, ••. K is ass-

umed as before.
Step 2: Setection and soating of attributes. The following criteria ha-

ve been selected:
dynamic characteristics of system structure, globally treated as a

qualitative criterion; the global evaluation is performed on the ba-
sis of aggregation of quantitative lower-level attributes.
flexibility of the system structure and uniformity of the algorithms

of the system components, defined as a qualitative criterion and eva-
luated on the basis of data concerning the functions of the control
station, whereby flexibility is higher if the algorithmic homogenoui-
ty of system components is smaller.
completeness of system functions, which can be defined as a quantita-

tive criterion evaluated on the basis of the percentage of fulfilment
of required system functions, but is eventually assessed judgementa-
lly.
design and functional configuration defined as a qualitative criteri-

on, explicitely dependent on the subjective perception of the decisi-
on maker.
redundancy and safety of the system, as a qualitative criterion based

on quantitative data concerning realisation redundancy (functional
and nonfunctional) and the safety degree in the remote control of po-
sitioning of final control elements.
discontinuity of the technological process, defined as a qualitative

criterion, evaluated on the basis of the automatization degree of
start-up procedures and shut-down operations of the technological
plant.
unification of control equipment in the plant as a qualitative crite-

rion based on quantitative data (by the number of various type of
305
final control elements, measurement sensors, etc).
documentation quality, defined as a qualitative criterion, based on

quantitative data concerning the documentation of the system.
Since all the selected criteria or attributes are qualitative,

for the purpose of their homogenous scaling it is necessary to define
for each one of them what is understood by qualification ~excelent~,
and what under the qualification ~unacceptable~.
For all these attributes, we assume that ~unacceptable~ is measured by

o and ~excelent~ by 10 points on assessment scale. However, the meaning
of these qualification is not necessarily clear for all attributes. For
example, unification of control equipment might be understood by some
decision makers as excellent, if it is high; other decision makers might
prefer diversity of control equipment. This can be resolved in the next
step.
Step 3: Setting aspiration levels. Decision makers who prefer diversity

in control equipment would specify for this attribute very low aspira-
tion level.
Step 4: Initial survey of alternatives. The three diferent process con-

trol systems are considered as alternatives, with following individual
characteristics:
Alternative solution CS 1 (Control System - 1)
The system structure corresponds to the alternative AI, fig. 3.
Functionaly, system components have heterogenous algorithms designed

for measurements, closed-loop control and for logical-program control.
When compared to user requirements, system functions are almost com-

plete.
Design and functional structuring is not good.
Redundancy and safety of the system are completely satisfying user

requirements.
Measures for handling discontinuity of process flow are complete.
Documentation is complete.
There is no unification of control equipment in the plant.
Alternative solution CS 2 (Control System ~ 2)
The system structure corresponds to the alternative A2 , Fig. 3.
Functionaly, system components have homogenous algorithms, designed

only either for closed-loop control or measurement or logical-program
T.6.
TABLE OF ATTRIBUTE-ALTERNATIVE ASSIGNEMENTS,INITIAL SETTING OF ASPIRATION LEVELS AND VALUES OF ACHIEVEMENT FUNCTIONS
CRITERIA C1 C2 C3 C4 C5 C6 C7 C8 s(y,y' ,y")

j+
k = 1 k = 2 k = 3 k .. 4 k = 5
H I
ALTERNATIVES DK FS PF PS RS NT DO UN c=O cool c=O c=l c=O cal c=O c=l c=O c=l I
y' 4 9 5 5 9 9 8 1 I
Al - CS 1 -6,6 -3,0 -7,5 -3,4 -9,0 -4,5 -9,0 -4,3 -6,0 -2,9
II
y 5 10 6 6 10 10 9 2
y' 9 1 5 9 9 1 7 4
A2 - CS 2 II
-8,5 -3,9 -8,0 -3,7 -9,0 -4,4 -8,8 -4,4 -9,0 -4,4
y 10 2 6 10 lO 2 8 5
y' 8 6 10 1 1 6 1 9
A3 - CS 3 .-8,5 -4,1 -8,3 -3,9 -9,0 -4,6 -9,0 -4,5 -8,7 -4,2
II
y 9 7 10 2 2 7 2 10 c.>
~
+k y' 7
8 7 8 G) 3 CD 3
D1
-II -(8)
y. 9 8 7 5 G) 4
0
y' 7 5 5 6 4
D2
-II
CD 8 ®
y 9 8 7 9 8 7 7
CD
y' 6 9 6 @ @ @ I@
D3 -II
0
Y 8 9 6 10 @ @ (!p)
CD
y' 9 9 @ 6 6 8 10
D4 _II
CD
y 9 9 9 @ 7 9 10
CD
y' @ § 9 .8 8 9 (0
D5 -II
CD
y 160) (Fa) I(ro) 9 fa) (3) 10 !(l)
307
control actions.
When compared to user requirements, system functions are incomplete.
Design and functional structuring are user-oriented.
Redundancy and safety of the system are completely satisfying user

requirements.
Procedures for automatic start-up and shut-down are not realised.
Documentation is incomplete.
Plant control equipment unification is incomplete.
Alternative solution CS 3 (Control System - 3)
• The system structure corresponds to the alternative A3 , Fig. 3.
T.7.
TABLE OF ,INDIVIDUAL AND FINAL DECISIONS
s(y,y' ,y") INDIVIDUAL

+k ALTERNATIVES AND FINAL
DECISION RANKS
MAKERS CS 1 CS 2 CS 3
c = 0 -6,6 -8,S -8,S CS 1 ,CS 2 ,CS 3

D1
c = 1 -3,0 -3,9 -4,1 CS 1 ,cs 2 ,CS 3
c = 0 -7,S -8,0 -8,3 CS 1 'CS 2 'CS 3

D2
c = 1 -3,4 -3,7 -3,9 Cs 1 ,CS 2 ,CS 3
c = 0 -9,0 -9,0 -9,0 CS 1 ,CS 2 ,CS 3

D3
c = 1 -4,S -4,4 -4,6 CS 2 , CS 1,CS 3
c = 0 -9,0 -8,8 -9,0 Cs 2 ,CS 1 'CS 3

D4
c = 1 -4,3 -4,4 -4,S CS 1 ,CS 2 ,CS 3
c = 0 -6,0 -9,0 -8,7 CS 1 'CS 3 'CS 2

DS
c = 1 -2,9 -4,4 -4,2 CS 1 'CS 3 'CS 2
c = 0 -7,62 -8,66 -8,7 Cs 1 ,Cs 2 ,CS 3

AGGREGATE
c = 1 -3,62 -4,16 -4,26 Cs 1 ,CS 2 'CS 3
308
Logical - program control functions are covered by special type of

system~s components.
When compared to user requirements, sy~tem functions are complete.
Design and functional structuring are not user-oriented.
Redundancy and safety of the system are not supported.
procedures for automatic start-up and shut-down of automated plant

practically are not realised.
Documentation is unacceptable.
Unification of control equipment in the plant is complete.
Step 5: Aloeeation of attributes. On the basis of the description of

alternative solutions, values for attributes have been appropriately
allocated. Because of exploratory nature of this study, uniform values
for all committee members were assumed (the decision makers did not ha-
ve the oportunity of correcting the assessment of attributes).
Step 6, 7: Individual ranking and analysis. As before, we assume that

it is not necessary to return to one of previous steps.
Step 8: Group ranking. Table T.7. shows individual ranking list of al-
ternative solutions according to achievement function values and an
aggregated ranking based on averaging achievement across the committee.
7. CONCLUSIONS
The obtained results show that the applied methodology of multi-

criteria decision making leads towards a satisfactory solution, even if
explicitelly divergent values of aspiration levels were assumed for the
purpose of ilustration. Further work on this example will include se-
ssions with decision makers - who will discuss and correct their aspi-
rations and make their own corrections of attribute assessments for all
alternatives. Computerized version of SCDAS decision support system
might be also applied.
309
8. REFERENCES
USHPKYH, A.A. (1982). OCHOS~ CHHT83a CTPYKTYP~ cnomHHX CHCT8M. HaYKa,

1982. MocKsa.
Debreu, G. (1959). Theory of value (Wiley, New York) •
Dreyfus, S.E. (1984). Beyond rationality (in M. Grauer, M. Thompson,
A.P. Wierzbicki, editors: Plural Rationality and Interactive De-
cision Process, Proceedings, Sopron 1984, Springer Verlag, Berlin).
Energoinvest (1979-84). Main Projects of Process Control Instrumentati-
on for 4x30 Mw, 1x110 Mw, 1x230 Mw and 1x300 Mw Power Plant Units
(in Serbo-Croat), Sarajevo, Yugoslavia.
Fishburn, P.C. (1970). Utility theory for decision making (Wiley, New
York) .
Gorecki, H., G. Dobrowolski, T. Rys, M. Wiecek, and M. Zebrowski (1984).
Decision support based on the skeleton method - the HG package
(in M. Grauer, M. Thompson, A.P. Wierzbicki, editors: Plural Ra-
tionality and Interactive Decision Processes, Proceedings, Sopron
1984, Springer Verlag, Berlin).
Grauer, M., A. Lewandowski and A.P. Wierzbicki (1982). DIDAS - theory,
implementation and experience (in M. Grauer and A.P. Wierzbicki,
editors: Interactive Decision Analysis, Springer Verlag, Berlin).
Keeney, R.I., and H. Raiffa (1976). Decisions with multiple objectives:
preferences and value trade-offs (Wiley, New York) .
Lewandowski, A., S. Johnson, and A.P. Wierzbicki (1985). A selection
committee decision support system: implementation, tutorial exam-
ple and users manual. Mimeograph, International Institute for
Applied Systems Analysis, Laxenburg, Austria; to be presented at
the MCDM Conference in Kyoto, Japan, August 1986.
March, J.G., and H.A. Simon (1958). Organizations (Wiley, New York).
Nakayama, H.J. (1984). On the components in interactive multiobjective
programming methods (in M. Grauer, M. Thompson, A.P. Wierzbicki,
editors: Plural Rationality and Interactive Decision Processes,
Proceedings, Sopron 1984, Springer Verlag, Berlin).
Raiffa, H. (1982). The art and science of negotiations (Harvard Univer-
sity Press, Cambridge, Massachusetts).
Sakawa, M. (1983). Interactive fuzzy decision making for multiobjective
nonlinear optimization (in M. Grauer and A.P. Wierzbicki, editors:
Interactive Decision Analysis, Springer Verlag, Berlin).
Simon, H.A. (1957). Models of man. Macmillan, New York.
Simon, H.A. (1958). Administrative behavior. MacMillan, New York.
Steuer, R.E. and E.U. Choo (1983). An interactive weighted Chebyshev
procedure for multiple objective programming. Mathematical Pro-
gramming, 26, 326-344.
Tietz, R. (1983). Aspiration levels in bargaining and economic decision
making. Lecture notes in Economics and Mathematical Systems, 213,
Springer Verlag, Berlin.
Tietz. R. (1984). On the structure, stabilization and accuracy of the
decision process (in M. Grauer, M. Thompson, A.P. Wierzbicki,
editors: Plural Rationality and Interactive Decision Processes,
Proceedings, Sopron 1984, Springer Verlag, Berlin).
Vlacic, Lj., and B. Matic (1984). Multicriteria analysis of dynamic
properties of hierarchical distributed process control system
310
structure, The First European Workshop on RTC-LSS, Patras Univer-

sity, 1984, Greece.
Vlacic, Lj., and B. Matic (1985). The selection of hierarchical micro-
computer-based process control system employing the multicriteria
approach. proceedings of the 11-th IMACS World Congress, August
1985.
Vlacic, Lj., and B. Matic (1986). A multicriteria decision support mo-
del for evaluation of process control system performance. To
appear in Ruschitzka M. editor: IMACS 1985 Transactions on Scien-
tific Computation, Vol. 2, Computer Systems: Performance and Si-
mulation, North-Holland.
Vlacic, Lj. (1986). Multicriteria Approach to Topology Determination of
Distributed Systems for Complex Control of Thermo-Power and Pet-
rochemical Process. Doctoral disertation (in Serbo-Croation),
University of Sarajevo, Yugoslavia.
Weistroffer, H.R. (1984). A combined over- and under-achievement pro-
gramming approach to multiple objectives decision making. Large
Scale Systems, 7, 47-58.
Wierzbicki, A.P. (1977). Basic properties of scalarizing functions for
multiobjective optimization. Mathematische Operations forschung
und Statistik, Optimization, 8, 55-60.
Wierzbicki, A.P. (1978). On the use of penalty functions in multiobjec-
tive optimization. In Oettli, W., Steffens, F., et al., editors:
Proceedings of the III-rd Symposium on Operational Research, Uni-
versitat Mannheim. Athenaum.
Wierzbicki, A.P. (1979). A methodological guide to multiobjective opti-
mization. WP-79-122, International Institute for Applied Systems
AnalYSiS, Laxenburg, Austria; also in Proceedings of the 9-th
IFIP Conference on Optimization Techniques, Warsaw 1980.
Wierzbicki, A.P. (1982). A mathematical basis for satisficing decision
making. Mathematical Modelling, 3, 391-405.
Wierzbicki, A.P. (1984). Negotiations and mediation in conflicts, I:
The role of mathematical approaches and methods. In Chesnut, H.
et al., editors: Supplemental Ways of Improving International
Stability, Pergamon Press, Oxford.
Wierzbicki, A.P. (1985). Negotiation and mediation in conflicts, II:
Plural rationality and interactive decision processes (in M. Gra-
uer, M. Thompson, A.P. Wierzbicki, editors: Plural Rationality
and Interactive Decision Processes, Proceedings, Sop ron 1984,
Springer Verlag, Berlin).
Wierzbicki, A.P. (1986). On the completeness and constructivenes of pa-
rametric characterizations to vector optimization problems. To
appear OR-Spektrum.
A FLEXIBLE MODEL FOR MULTI -OBJECTIVE OPTIMIZATION
H. Roland Weistroffer
School of Business
Virginia Commonwealth University
Richmond, Virginia 23284-0001, USA
Introduction
Decision problems often involve a multitude of objectives (or decision

criteria) which are incommensurable and possibly conflicting with one another. At
times, a decision-maker may not be able to express an explicit preference ordering
for the various objectives. In fact, he may find it difficult to state the objec-
tives, or to distinguish between objectives and constraints. For example, in
designing a new car model, the resistance to impact may be viewed as an objective to
be maximized, or as a constraint fixed by law.
Decision problems" that are not fully structured cannot be solved by traditional
optimization techniques, since there is no optimum solution in the mathematical
sense. Rather, a compromise solution with respect to the multiple objectives is
sought, specifically, that compromise solution which is most satisfactory to the
decision maker (i.e. there should not exist any other possible solution which the
decision maker may find preferable). When a decision maker is unable to communicate
his preference ordering explicitly, an iterative technique is desired to solicit the
preference information from the decision maker in a form in which he is capable of
giving it (based largely on his intuition). This information is used to partially
structure the problem (perhaps producing a preliminary "best" solution). The
analyst then goes back to the decision maker to extract more preference information
and to allow him to modify the information given previously.
Mathematically, decision problems of this type can be represented by the
following multi-objective optimization (or vector optimization) problem:
maximize [f 1 (x), f 2 (x), ••• ,f k (x»)T

subject to g1(x)
.?. 0, (1)
where x = [xl' x2' "" x n ) ERn is the vector of decision variables, each
xERn representing a possible solution to the problem, the f i : Rn~ R, i = 1, " ' , k,
are the objective functions, and the gj: Rn 7 R, j = 1, " ' , m, are the constraint
functions.
312
The best compromise solution sought, would have to be feasible (i.e., satisfy
the given constraints) and efficient (or non-dominated or Pareto optimum), that is,
there should be no other feasible solution at which all the objective functions
achieve equal or better values and at which at least one objective function achieves
a better value. An iterative and interactive technique can be used to determine the
best compromise solution within the efficient and feasible solution set.
The model and technique proposed in this paper is based on the reference point
approach briefly described in the following section.
The Reference Point Approach
The reference point approach (Wierzbicki 1982, Lewandowski and Grauer 1982,
Grauer 1983) basically works as follows: the decision maker provides reference
values, i.e., objective function values that he considers acceptable for the
decision problem, without knowing whether these values are realizable or not. A
computer program is then used to minimize the under-achievements with respect to
those reference values that were picked too high, and to maximize the over-
achievements with respect to those reference values that were picked too low, i.e.
minimize si ~ fi(x), i = 1, 2, ••• , k, (2)
where the reference values si may be larger or smaller than the corresponding
objective function values fi (x). The minimization is carried out via some order
preserving scalarizing function (or surrogate objective function) and subject to the
given constraints gj (x) 2. 0, j = 1, 2, ••• , m. The resultant solution and the
corresponding objective function values are shown to the decision maker, who may now
decide to raise or lower some of the reference values, based on new insight as to
what is realizable. A new solution is found, and the process continues until the
decision maker feels that the solution obtained is the best compromise solution (or
very close to it).
A special case of this approach arises when the reference values are con-
sistently larger than the corresponding actually achieved objective function
values. The reference values may then be viewed as targets or goals or desired
values. After finding a solution by minimizing the under-achievements, the decision
maker is asked to lower some of the desired values to a more realistic level, and
the process is repeated as before. The approach can now be regarded as a general-
ized form of goal programming. This approach lets the decision maker set his
initial expectations very high (i.e. specify rather optimistic desired values), and
then iteratively and interactively induces him to lower these expectations to real-
istic levels. The interactive goal programming method of Weistroffer (1983) is
based on this approach. STEM (Benayoun, de Montgolfier, Tergny and Laritchev 1971)
also follows the philosophy of decreasing expectations, but does not ask the
decision maker to reduce any desired values, but rather to change some of the
objective functions into additional constraints.
313
It may be argued that psychologically it is preferable to start out with low

expectations. and then increase these to the maximum achievable. rather than to
start out with high expectations and then haVing to cut back. This is accomplished
by another special case of the reference point approach. the case where the refer-
ence values are consistently lower than the actually achieved objective function
values. The reference values may then be viewed as minimum achievement levels or
required values. (2) then in fact means that the distance between the required
values and the corresponding objective function values is maximized. After finding
such a solution. the decision maker is invited to increase some of the required
values (but not to exceed the corresponding achieved objective function values. and
the process is repeated until further increase of the required values is not
possible or until the decision maker is satisfied with the achieved solution. The
over-achievement programming method of Weistroffer (1982) follows this approach.
The method of satisfactory goals of Benson (!975) follows the philosophy of
increasing requirements. but does not make use of a scalarizing function to maximize
the over-achievements. Instead. each objective function is maximized individually.
subject to the other objective functions meeting the corresponding required values.
The decreasing e~pectations approach and the increasing requirements approach
can be combined into a bracketing approach. where the decision maker specifies both
high reference values (desired values) and low reference values (required values)
and an order preserving scalarizing function is used to find a compromise solution
within this bracket. which maximizes the over-achievements with respect to the
required values and minimizes the under-achievements with respect to the desired
values. The decision maker may then increase some of the required values or
decrease some of the desired values. thus reducing the size of the bracket (or
box). The process is repeated until a satisfactory solution is found or until the
bracket becomes so small that the differences between the solutions within it are no
longer significant. This bracketing approach is followed by the combined over- and
under-achievement programming method of Weistroffer (1984a). The interactive
multiple-goal programming method of Spronk (1981) may also be viewed as following
the bracketing philosophy. however. it has several other features and it may also be
viewed as following the increasing expectations philosophy.
Alternative Proble. Formulation
In the introduction to this paper it was pointed out that the decision maker
may not always be able to distinguish between objectives and constraints. For that
reason. the following alternative problem formulation to (1) is proposed:
maximize [f 1(x). f 2 (x) • •••• fk(x)]T
for fi(x) < di • i = 1. 2 • •••• k • (3)
subject to fi (x) ~ r1' i = 1. 2 ••••• k.
where the new k is equal to k + m in (1). Objective functions and constraint

314
functions are identical in this formulation, both are depicted by fi (x), and are
referred to as criterion functions. Under the usual assumptions that the objective
functions are continuous and that the feasible set is compact, (3) is equivalent to
(1): The compactness and continuity conditions imply the existence of upper and
lower bounds for the objective functions and the d i and r i may be set equal to those
bounds. The d i and ri for the constraint functions may be set to zero.
The d i can be viewed as desired values and the ri can be viewed as required
values.
The Proposed Model
The alternative problem formulation presented above suggests the simultaneous

use of desired values and required values and thus the bracketing approach becomes
an obvious candidate to tackle the decision problem.
The scalarizing function used by Weistroffer (1984a) however, has the dis-
advantage that i f it is used without additional weighting factors (to deemphasize
the influence of the required values and to emphasize the influence of the desired
values), it sometimes over-achieves some of the desired values at the cost of a
larger under-achievement for some of the other desired values, rather than staying
within the specified bracket. An improved scalarizing function is therefore pro-
posed, which does not allow this to happen. An earlier version of this new
scalarizing function is described in Weistroffer (1984b).
An important characteristic of any decision aid is flexibility. The decision
maker should not be coerced into providing information with which he is not com-
fortable. It should not be necessary for him to provide the required values and the
desired values for every decision criterion. Furthermore, he should be allowed to
change these values, if new insights provided to him result in a modification of his
intuitive preference structure. Also, the information provided on the importance of
an objective to exceed the required value and achieve a value "close" to the desired
value should be open to review at every iteration. Thus, a former constraint may
change into a full fledged objective and vice versa, or the degree of its importance
may change. The proposed iterative technique exhibits such flexibility.
Use of scalarizing functions of the following form is proposed:
(4)
where[xl' x2' "., x n 1 = xeRn is the vector of decision variables, each element of
Rn representing a specific solution, fj: Rn ~ R, j = 1, 2, "" k are the criterion
functions, d j , j = 1, 2, " ' , k are the corresponding desired values, rj' j = 1, 2,
.", k are the corresponding required values, WjE(O, 11 expresses the importance of
getting close to the desired value d j rather than merely meet the required value rj
for each criterion j E{1, 2, ...
, k} (Wj determines to what degree a criterion
function is to be considered an objective or a constraint). The scalarizing
315
function described in Weistroffer (l984b) is the special case where Wj _ 1.

Constraints in this earlier approach are handled by letting d j : rj for all j
corresponding to constraints.
Since the decision maker need not specify required values and/or desired values
for all criteria. the d j not specified are set equal to the ~ values f;. the
maximum values for each fj(x). j 0 . 2 ••••• k}. when all the other criterion
functions are ignored. Similarly. the unspecified rj may be set to the pessimistic
values f*j' which are defined as the smallest values that the fj(x) attain at any of
the ideal solutions (where an ideal solution is defined as a solution in Rn at which
at least one fj(x) achiveves its ideal value). Whereas the ideal values are upper
bounds for the criterion functions. it must be pointed out that the pessimistic
values are not lower bounds for the criterion functions on the efficient solution
set. and that the most satisfactory solution may well be one at which some of the
criterion functions take on lower values than their respective pessimistic values
(Weistroffer 1985).
As long as d j ~ fj(x). j = 1 ••••• k. the proposed scalarizing function is
monotone increasing with respect to any increase in any of the criterion function
values. Thus. for sufficiently large desired values. maximizing the proposed
scalarizing function will always produce an efficient solution: If x* is not
efficient. then it is possible to increase at least one of the criterion functions
without decreasing any of the others. which means that the scalarizing function (4)
is not maximum at x*. Furthermore. the resultant solution will always satisfy all
of the required values. provided such a solution exists. since the scalarizing
function tends to - 00 as any fj(x) approaches its required value. After being
presented with the obtained solution. the decision maker may modify some or all of
the rj • dj and/or Wj' j = 1. 2 ••••• k. To ensure feasibility (once this is
established for the initially specified required values) the rj should always be
kept smaller than the actually achieved corresponding criterion function values. In
general. the achieved function values will be contained in the hyperdimensional box
(or bracket) specified by the required and desired values. The importance-
parameters. Wj' will influence the position of these function values within the box.
If the decision maker is reasonably consistent with respect to the modi-
fications at each iteration (he is allowed to be erratic at the beginning of the
procedure. as long as there is a long-term trend towards preference consistency).
the method should converge towards the most satisfactory solution. The procedure
may be stopped any time that the decision maker feels that no more significant
improvement is possible.
Some Remarks
Even though the identical treatment of objectives and constraints is advan-

tageous from a theoretical point of view as well as for many real world situations.
there may be other situations where the explicit designation of some constraints is
316
desirable. From a numerical point of view, some experimentation with the wi values
for strict constraints may be required. The author is currently working on
designing software to implement the ideas described in this paper.
References
1. R. Benayoun, J. de Montgolfier, T. Tergny, and O. Laritchev (1971). Linear

programming with multiple objective functions: step method (STEM). Mathe-
matical Programming, 1, 366-375.
2. R. G. Benson (1975). Interactive mUltiple criteria optimization using satis-
factory goals. Doctoral thesis, University of Iowa, Iowa City.
3. M. Grauer (1983). Reference point optimization - the nonlinear case. In P.
Hansen (editor), Essays and Surveys on Multiple Criteria Decision Making,
Springer-Verlag, Berlin, 126-135.
4. A. Lewandowski and M. Grauer (1982). The reference point optimization approach-
methods of efficient implementation. Working paper WP-82-019 , International
Institute for Applied Systems Analysis, Laxenburg, Austria. .
5. J. Spronk (1981). Interactive Multiple Goal Programming. Martinus Nijhoff,
Boston.
6. H. R. Weistroffer (1982). Multiple criteria decision making with interactive
over-achievement programming. Operations Research Letters, 1 (6), 241-245.
7. H. R. Weistroffer (1983). An interactive goal programming method for non-linear
multiple-criteria decision-making problems. Computers and Operations Research,
10(4), 311-320.
8. H. R. Weistroffer (1984a). A combined over- and under-achievement programming
approach to multiple objectives decision making. Large Scale Systems, 7, 47-58.
9. H. R. Weistroffer (1984b). Multiple objectives optimization by inter-actively
narrowing bounds on the most preferred solution. Proceedings of the 1984 IEEE
International Conference on Systems, Man and Cybernetics, Halifax, Nova Scotia,
Canada, October 10-12.
10. H. R. Weistroffer (1985). Careful usage of pessimistic values is needed in
multiple objectives optimization. Operations Research Letters 4(1), 23-25.
11. A. P. Wierzbicki (1982). A mathematical basis for satisficing decision-
making. Mathematical Modeling 3, 391-405.
FOUNDATION OF ,
EFFECTIVE GOAL SETTING'
P. L. YU
and
1. S. Chien
School of Business
University of Kansas
Lawrence, Kansas 66045
Abstract
Problems of effective goal sett'ing can be formulated into complex multiple

criteria optimal control systems in which the control variables are (i) selecting
measurable goal f\f[lcti~ns, (ii) setting goal achievement levels and (iii) deter-
mining effective supportive systems; and the objectives are (I) to maximize the
attention time allocation to the work assigned and (.II) to maximize the efficiency
and effectiveness (or ability) of work performance. Utilizing the concepts of
human behavior mechanism and habitual domains, we offer a general principle of
effective goal setting as to select control variables which can create favorable
working condit:i:ons and relati.ve high levels of charges and confidence on the per-
formers so that they can allocate as much time as pos'sible to perform the desired
job-related operations in a most efficient and effective way. We give a detailed
discussion of the applications of this principle o.ver various classes' of prob-
lems. Some emp'irically 'known result'S on goal setting, which in general are con-
sistent with and can be explained by the principles, are also sketched.
1. Introduction
In Multiple Criteria Decisien Analysis, 'we are usually given a set of goal
functions or criteria and a set of alternatives., with which we try to find the
most "valuable" alternative according to some preference defined on the outcome
space. The most valuable alternative is usually the one which maximizes some
value function defined on the outcome space. It can be an element of some nondom-
ina ted solutions or some satisfying solutions. (See [18, 21] and those quoted
therein for instances.) One of the common properties in this kind of approach is
that the outcome space and the alternative set are fixed. However, many decision
1. This research has been supported by NSF of U.S.A. Grant No. IST-8418863.
318
problems which are involved with multiple criteria do not necessarily have a fixed
set of outcome space and alternative sets. It is well known, and has been stud-
ied, that if one would set a little more difficult goal, his performance results
usually could be better (See Section 5)-. That is, by stretching their ability a
little bit human beings can achieve more. Thus, the outcome space can be
stretched a little bit to contain some preset goal levels. It i.s also well known
that our performance tends to respond differently -to different sets of goals.
Therefore, the selection of the goal functions can be very important in achieving
the overall performance over some period of time.
The purpose of this article is to investigate the foundation for effective
goal settings as to maximize the overall performance over a period of time. The
goals are treated as variable, rather than predetermined. The problem, however,
is still closely related to multiple criteria decision making problems (MCDM). It
simply involves more complexity in the setup than ordinary MCDM.
In order to facilitate our discussion, let us consider the following athlete

developme.~ problem first.
Mr. John Coach is a track and field coach at a high school. One day he finds
a very well built athletic student, Tom Student. Tom can run one hundred meters
in abou-t eleven seconds. The coach. is ambitious, and when he spotted this stu-
dent, he immediately thought that Tom Student could be an Q1ympic star. He wanted
to fulfill his mission by training and developing Tom Student as to make him an
Olympic star, which became John's overall objective. The following are some im-
portant questions which need to be addressed:
(i) How can John Coach implant the idea and inspire Tom Student to set a goal
to become an Olympic star and be willing to go through the hardships of tough
training? How can~this drive for becoming an Olympic star be continued without
interruption or dissipation?
(ii) Which set of subgoa1s, such as the speed measured in seconds, the
streamlining of movement, the vitality, the time of training everyday etc., are
most important in improving Tom Student's performance?
(iii) How specific does the goal setting need to be to maximize the improve-
ment of performance? For instance, take speed as an example. Should the coach
specify that Tom Student should achieve 10.5 seconds by one week or one month, or
should the coach just tell him to try his best, without specifically setting a
specific goal level to achieve.
(iv) Should the coach continuously assist, supervise and encourage the stu-
dent in his training and performance or just leave him alone and tell him to try
his best?
(v) Should the coach create an environment or barrier that would allow Tom
to concentrate his efforts on improving his ability to run? (Otherwise, problems
319
such as Tom's girlfriends, parents, or financial problems, etc., may disturb his
attention and drive for improvement.)
(vi) Should the Coach help Tom develop a habit far training himself to a-
chieve his goals as scheduled with 100% concentration in order to become an Olym-
pic star?
In this paper, we shall discuss and answer the above questions and offer a
general principle for effective goal setting. The principle is not only appli-
cable to the above athlete problem, but also to many other decision problems or
situations in which proper goal setting can be used to improve overall perfor-
mance, including the problems of assembly line workers., receptionists, managers,
researchers and corporate presidents. The principle can be used for self goal
setting to improve self-performance and be used for goal setting for other persons
to improve their performance.
In order to describe and illustrate the foundation of effective goal setting,
we shall briefly review some fundamentals of human behavior mechanisms and sketch
their relation to the problem of effective goal set'ting in Section 2. In Section
3, we shall briefly review the concept of habitual domains and use the concept to
discuss the processes af implanting, accepting and habituating of ideas in goal
setting.
In Section 4, we lay down a general principle of effective goal setting as a
complex mUltiple criteria optimal control system. Using habitual domain concepts,
in Section 4.1, we first classify decision problems into four classes: simple
routine, mixed routine, fuzzy and challenging. The general principle is then
described in Section 4.2, utilizing the concepts of behavior mechanism discussed
in Section 2. Further discussion on the control parameters of goal setting and
supportive systems are given respectively in Section 4.3 and 4.4. Finally, the
principle is mathematically formulated in Section 4.5.
There is rich literature in goal setting contributed by organization behav-
iorists and industrial psychologists. For instances, see [I, 3-16] and those
quoted therein. Because of their emphasis on statistical empirical studies (a
habitual domain), they primarily focus on simple routine or mixed routine pro-
blems, not including fuzzy or challenging problems. Strictly speaking, their
results cannot be used for a broad general class of goal setting problems. How-
ever, as their results are, in general, supporting the principle described, we
shall briefly sketch their empirical study results in Section 5. Thus, the reader
can use our principle to explain their results as verification exercises.
2. A Behavior Mechanism
In order to understand the functioning of goal setting on behavior and per-

formance, we shall first briefly outline a behavior mechanism and state its rela-
320
tionship to effective goal setting. FIGwchart: 1 illustrates the micro dynamic

mechanism of behavior. Details concerning the mechanism can be found in [17] or
Chapter 9 of [18]. For our purposes, the main aoneepts 'are sketched here as fol-
lows:
-..
co
c:
r
l-
e
c 51
-,(14)
-c;
~
Il e e I
I- : :
c I
u
I c:CI !:!
.....
.2 - -I
I> ..
-
e
en
>.
.r:.
I C ~I
Ci)
I ~ ell
...,
I tl ~ I
Q..
(13)
1< -gl
1_ <J
.. en (12)
Ci)
~
.....I
~
......----,---.... External
Experience/ Being
Reinforcement Observed
Flowchart 1 Behavior Mechanism
(I) Each individual is endowed with an internal information processing and

problem solving capacity. Like a "super computer," the brain encodes,
stores and retrieves ideas and messages, using the pattern circuits of
excited neural cells. The stored messages can be dynamically restructured
so as to most efficiently retrieve needed relevant information. Using
association and analogy, the brain interprets arriving inputs based on the
stored ~ry or learned knowledge. (See Box 1; also Hypotheses 1-4 of
[17,18] for details.) The arriving events can come from self-suggestion
(Box 10), physiological mtmltoring (Box 11) and/or external inputs (Boxes
321
12, 13 and 7).
(II) Each human being has a set of goals or equilibrium points to reach and
maintain (goal setting). A list of gods is given in Table 1 (p. 8). The
status or stat·e of each goal is constant.1y (conS'ciously or unconsciously)
monitored (state valuation). When a significant "unfavorable" deviation
between the perceived value and the ideal value exists, each goal can pro-
duce a "charge" (a tension or urge). The totality of the charges produced
by all goals which form a hierarchical structure is called charge struc-
ture. This charge structure can change rapidly with time, information
inputs, and one's psychological state and conditien. (See Boxes 2-5; also
Hypotheses 5-6 of [17,18].) Note that when a goal (e.g., health) is main-
tained at the tolerable interval of the ideal, without other effort, the
goal may produce very little or no charge and will be neglected.
Table 1. A Structure of Goal Functions
(i) Survival and Security: physiological health (proper blood

pressure,. body temperature and balance of biochemical states);
proper level and quality of air, water, food, heat, clothes,
shelter and mobility; safety; acquisition of money and other
economic goods;
(ii) Perpetuation of the Species: sexual activities; giving birth

to the next generation; family love, health and welfare;
(iii) Self-Importance Feeling: self-respect and self-esteem; esteem

and respect from others; power and dominance; recognition and
prestige; achievement; creativity; superiority; accumulation of
money and wealth; giving and accepting sympathy and protective-
ness;
(iv) Social Approval: esteem and respect from others; friendship;

affiliation with (desired) groups; conformity with group ideol-
ogies, beliefs, attitudes and behaviors; giving and accepting
sympathy and protectiveness;
(v) Sensuous Gratification: sexual; visual; auditory; smell;

taste; tactile;
(vi) Cognitive Consistency and Curiosity: consistency in thinking

and opinions; exploring and acquiring knowledge, truth, beauty
and religion;
(vii) Self-Actualization: ability to accept and depend on the self,

to cease from identifying with others, to rely on one's own
standards, to aspire to the "ego-ideal" and to detach oneself
from social demands and customs when desirable.
(III) Attention Allocation (Box 6) is defined as a human's time allocation of his

internal information processing and problem solving capacity over various
322
events and activities. The time could be measured in milliseconds. The

purpose of attention is tn release the charges in the most efficient way.
These ways involve: (a) actio1l1il ·aDd discharges when solutions are obtained
(Boxes 8-9); (b) acquiring external information (Boxes 7, 12, and 13); or
(c) self-suggestion for internal thinlti;ng, justification and rationaliza-
tion (Box 10). All of these functions feed back to the internal informa-
tion processing center. (~ee Hypothesis 6 of [17,18].)
(IV) Each event or problem can be associated with a set of goals. Its signifi-
cance can be defined in terms of the remaining charge structures when the
event is removed. According to lexicographical ordering, the significance
of events can be compared. The most 'significant event or problem will
command one's attention. As charge structures rapidly change, the atten-
tion can switch rapidly from one event to otller events. (See Hypothesis 6
of [17,18].)
(V) When there is a set of alternatives for discharge, the one which can most
reduce the current charge structures will be ~hosen for discharge. This is
called the least resistance principle. The r.emaining charge is a measure
of resistance to the total discharge. (See Hypothesis 7 of [17,18].) Note
that (iv) and (v) are similar. Both of them are complex dynamic optimiza-
tion problems for our brain and mind.
(VI) When the existing set of alternatives cannot adequately reduce the level of
charge structures, people have two possible behavior tendencies (Box 14):
(A) Active problem solving (without changing the ideal value of the goals,
the charge is not released until the ideal is reached), and (B) Avoidance
justification (using self-suggestion to lower the ideal value or signifi-
cance of the problem as to reduce the charge level). Note that active
problem solving transforms high level of charge into force, ~ or ~
vat ion; but avoidance justification simply diffuses or dissipates the char-
ges. In the latter case, charges are not equal to force, drive or motiva-
tion.
(VII) All functions/components of the flow chart are interconnected. Through
time, they interact with one another. For instance, once an action is
taken (Box 9), say publishing an article, the event and its consequence
will be observed and registered in the decision maker's memory (Box 1) and
also will likely be observed and interpreted by other parties (Boxes 12 and
13) which ·may, in turn, react upon the original decision maker (Boxes 7 and
11). (See Hypothesis 8 in [17,18] for information inputs.)
Remark 2.1 According to the model, the folloWing are some

necessary conditions for goal setting to be effective: ti) the goal must be set
so that it can produce a high level of charges as to command the attention al10ca-
323
til>n; (ii:) the subject should have sufficient ability and confidence that the
goals can be obtained in the due time, which induces active problem solving rather
than avoidance justification; and (iii) the supportive systems (including reward-
ing systems and monitoring systems) and the environment should be such that the
easiest way to release the high level of charges is to perform the desired
job-related works to reach the preset goals.
In our at'hlete's problem, if John (loachcannot inspire Tom Student to a high
level of charges or burning desire ,to become an Olympic star, or if Tom Student
cannot mass enough confidence that he can become a star in due time, or if Tom
Student is in an environment in which other more serious events keep on bothering
him (such as, diseases and/or health problems, 'disputes with girlfriends, serious
financial disasters, etc.), then we cannot expect that Tom S,tudent can contin-
uously put his attention to athletic training to become an Olympic star. Even if
he could occasionally think about ,the event. because of other priorities or avoid-
ance justification he could not mass enough effort and time to go through the
training to achieve the goal of becoming an Olympic star.
3. Habitual Domsins
The dynamic mechanism, Although changing with time and with one's psycho-
logical states, can stabilize and, unless extraordinary events occur, each indi-
vidual can have stable habitual patterns for processing information. This obser-
vation is captured by the concept of habitual domains.
It has been recognized that each human being has habitual ways to respond to
stimuli. These are sometimes called conditioned or programmed behaviors. For
example, each individual has habitual ways of eating, dressing, and speaking.
Some habitually emphasize economical gains, while others habitually emphasize
social reputation. Some are habitually persistent in their goal pursuit, while
others habitually change their objectives. Some are habitually optimistic and see
the po'sitive side, while others are habitually pessimistic and see the negative
side. Some habitually pay attention to details; while others only to generali-
ties.
These habitual ways of perceiving, thinking, responding and acting can be

captured by the concept of habitual domains (HD). Flow Char.t 1 illustrates that
HD involve self-suggestions, external information inputs, physiological monitor-
ing, goal setting, state valuation, charge structures, attention and discharges.
They also involve encoding, storing, retrieving and interpretation mechanisms.
When a particular aspect or function is emphasized, i t will be designated as "HD
on that function." Thus, HD on self~suggestton, HD on charge structures, HD on
attention, etc. all follow. When the responses to a particular event are of in-
terest, we can designate them as "HD on the responses to that event." Thus, HD on
324
job seeking, house purchasing, dealing with friends, etc. also follow. In the
following discussion, BD are used without specifying 'the corresponding functions
or events, as only the general properties of HD will be discussed.
3.1 Elements of Habitual Domains
A habitual domain at time t, denoted by HOt, is defined as the collection

of the following:
(1) PDt (potential domain): The collection of ideas/actions
that can be potentially activated at time t;
(ii) ADt (actual domain): The set of ideas/actions that is

actually activated at time t;
(iii) APt (activation probability): The activation probability

or confidence structure at time t which indicates the possibility for
a subset of ideas/actions in PDt to be in AD t ;
(iv) R(It,Ot) (reachable domain): The reachable (attainable) set

of ideas/actions from the initial set of ideas It through the set of
operators Qt.
R(It,Ot) shall be described more extensively shortly. Observe that ADt
PDt· The relationship between ADt and PDt is similar to that between the realized
value and the sampling space of a random variable. The set ADt will be affected
by the charge structures and attention at time t. The probability or confidence
level (APt) for a set of ideas/actions to be activated will depend on how strongly
the ideas/actions have been encoded and how easily the ideas/actions can be re-
trieved from memory storage. It is also closely related to the set of initially
activated ideas around the time t.
Now consider reachable domains. Through self-suggestion and/or external
information inputs, one idea or a set of ideas can be used to stimulate or gener-
ate other ideas. This suggests that a set of operators exists, defined on the
subsets of PDt' which generate ideas in PDt from the subsets of PDt. As an exam-
ple, suppose that one is interested in investing a fixed amount of cash in stocks
A and B. The concept (the operator) that any portfolio (a convex combination) of
A and B would also be of interest will expand the alternative set of A and B into
the set of all convex combinations of A and B. Note, the operators are also ele-
ments of PDt.
Let It be a set of ideas with It CPD t , and let 0t be a set of operators
which generate ideas in PDt from subsets of PDt. Then R(It,Ot)' the reachable
~, is the set of ideas/actions that can be reached (or attained) from It and
325
Qt. Hore precisely, R(It,Ot) is the set- of ideas/actions that can be cumulatively
generated by any sequence of operators from the set of 0t which act on It and the
resulting ideas/actions frem the operations. As an example, let It = {O,l} and 0t
- {+} ("+" is the ordinary mathematical addition symbol). Then R(It,Ot) is the
set of all nonnegative integers. If 0t .. {+,-}, then R(It,Ot) is the set of all
integers. In the athlete problem, the idea of becoming an Olympic star can gener-
ate "glory, pride-, wealth, hard' work, discipline ••• " as a subset of the reachable
domain.
Potential and reachable domains are a1ese1y related. We say that {It,Ot}
is a generator of PDt iff PDt" R(It,Ot)' and {It,Ot} is a basis for PDt iff
{It ,Ot} is a generator of PDt and no strict subset of {It ,Ot} is a generator of
PDt·
Finally we observe that HOt' as defined, is dynamic and changes with
time. All four elements of HOt can evolve with time: this makes analysis of HD t
fairly difficult. Fortunately, it can be shown that the HOt can reach a stable
state. Thus, human behaviors are still, to a certain degree, predictable.
In [2,181, it is shown that PDt and APt depend on the activation of pat-
terns of neural circuits. Under suitable conditions, these patterns will satisfy
a system of differential equations with stable steady solutions. Thus, PDt and
APt can reach their stable states unless extraordinary events arrive continuously
(which rarely is the case).
Once PDt and APt reach their stable states, one can expect ADt to occur
with some regularity. In such stable states of HD t , one ean expect habitual ways
of thinking, responding, and Teacting to stimuli and events to occur. Thus, per-
sonality, attitude, and conditioned or programmed behavior will be formed for each
individual. Such formation has a great impact on decision making styles and opti-
mization theories. The existence of habitual domains has significant implications
for high stake decision problems, optimal solutions, gaming and conflict dissolu-
tion, career management and leadership. For the details see [2,18-201. We shall
discuss only those related to effective goal setting.
3.2 Cores of Habitual Domains
Understanding the HOs of ourself and others is very essential for effective
goal setting. For instance, in the athlete problem (I), if John Coach wants to
successfully implant a burning desire (high level of charge) in Tom Student's mind
to become an Olympic star, he must first explore Tom's HO. What are the active
ideas and interests in Tom's mind? What are his CUTrent goals for life? How much
confidence does he have in himself and in the coach's suggestions? What are his
channels of external information? Without a careful study, the implanting may
fail. For example, suppose that Tom's father is a successful owner of a restaur-
326
ant whom Tom respects very highly. Tom wants to be a millionaire by getting into
the restaurant business after his graduation. He does not have high hopes and
interest in becoming a sport star because he does not believe that he could become
a star and thus a milli@naire~ If the 'coaCh doea not know the above, he may just
try to persuade Tom by saying how glortous and fulfilling it wiLl be to become an
Olympic star without suggesting -to Tom that an Oly.mpi~ star can also become a
millionaire and without building up the confidence for Tom himself to believe that
he could be a star. If this is the case, we see that the coach's probability of
success cannot be too great. If the coach knows Tom's HD~ he may visit his father
and layout the glorious picture in which Tom is an Olympic star. He may show the
father why Tom can be a star and hint that many Ol"}'1llpic stars did become million-
aires and have happy lives. Once the fathe.r is convinced., he could talk to Tom.
Through his father's channel, Tom may build up more confidence in himself and more
interest in becoming a star.
Note that newly arrived messages or concepts ean be rejected quickly if
they are not consistent with the existing HD and the subject does not have an
"open minded" operator in his MD. Thi's is due to the least resistance principle
(it is easier to reject than to aecept, because it requires encoding, association
and integration to accept the new ~oncepts and to maintain logical consistency).
In fact, it always takes time to implant new ideas and have the ideas accepted.
In studying effective goal .setting, the concept of the cores of HDs become
very important. Roughly speaking, the core of the HD is the set of ideas or con-
'cepts that are very strongly implanted in the HD and can be easily retrieved to
interpret the arriving events or to perform the needed work. As a core is a con-
cept on how strongly the ideas are encoded and how easily retrieved, while a gen-
erator or basis is a concept on those ideas that can potentially activated, we
notice that the core can usually just be a proper subset of a generator or a ba-
sis. For example, when we talk about relatives and friends, the set of {the par-
ents, ·the wife or husband, the sons and daughters, the business colleagues, the
neighbors} can 'be a core; ~ut the set cannot be a generator or a basis for all
re'latives and friends. In the athlete problem, when talking about future ambi-
tion, {a millionaire, restaurant} can be the core of Tom's current HD; but the set
cannot be a generator or a basis for all potential ideas about future ambition.
New ideas can be implanted into HD and become elements of the core.
In order to make the concept of core precise. let I be a given idea and
let be a particular event or problem under consideration.
Define the activation density of idea I in event E at time t by:
atU, E ) = P,robt-[I activates IEgets attention] (1)
which is the conditional probability for I to be activated given that our atten-
tion is paid to E at time t.
Note that the activation density is a function of time and
327
o .. 8<t (I, E) .. 1. Now we can defh.ethea-core of lID with respect

to E at time t, denoted by Ct ( E, a) to be the set of ideas or
concepts that bave act1:vation dens:it.y exceeding or equal to a. That is,
Ct ( E, a) .. {I I at(I, E) > a}. (2)
Note that if I e Ct ( E, a), with high value of a, (say a > 0.9), then I
is an active and very frequently retrieved idea. when E catches our attention.
Also note the Ct(E, a) C Ct(E, ~,) if a > ~.
In our lat'er discussion, unless otherwise specified, whenever we talk about
the core of HD we shall mean a-core with high value of a.
Note that for a routine work such as assembly, there is a preferred or
efficient sequence of operations. For effe'ctiw goal setting, we want this pre-
ferred 'sequence to become the core of the HD of the workers. In fact, in effect-
ive goal se.tting for any problem, we want. to help search the preferred sequence of
operations and make the sequence a major part of the eore of lID. This will be
discussed in the next sections.
For convenience, given lID for a 'particular problem E, we shall use the
following terminology.
By implanting ~n idea I into HD we mean putting idea I, perhaps subtly,
into the subject's lID. At the time of implanting, the idea I is usually new or
fairly new to the subject. That is, the activation density of I in E, at(I, E),
is zero or very low. As time goes by, the newly implanted idea can be rejected or
aceepted by HD. By accepting the idea I in lID, we shall mean that the idea I
becomes an active element of lID with at(I, E) > O. As I is repetitively used, its
activati~n density will become larger and can eventually become an element of the
core of lID. Thus when the problem E is attended, idea I will be habitually or
aatomaticallyactivated. A set of ideas r = {I1 , ••• ,I k} is said to become the
habituation of event or problem E if r becomes the core of HD whenever E is
attended.
As an example, in our athlete problem, let event E be the future career and
I be "it is great to be an Olympic star". After being implanted, idea "I" can
gradually be accepted by Tom Student. It can eventually 'become the habituation of
"the future career" for Tom. That is, whenever "the future career" is mentioned,
Tom would think of "it is great to be an Olympic star".
4. Foundation of Effective Goal Setting
Utilizing those concepts described in the previous sections, we shall first

in Section 4.1, classify decision problems or situations. The classification can
help us to better understand the later discussion. In section 4.2, we lay down
the principles of effective goal setting. In Sections 4.3 we discuss the impact
of the parameters in goal setting on performance, and in Section 4.4 we outline
328
some impact of the pa~ameters in supportive systems.
4.1. A classification of Decision Problems
Depending on the regularity and availability that ideas, concepts and ope-
rations are needed to successfully perform jobs or solve problems, we can roughly
classify the decision problems into four classes.
Class I. Simple Routine Problems. In this class, the set of ideas, con-
cepts and operations which are needed to successfully solve problems is almost
fixed and well ordered. Assembly line work, typiug, sleeping, eating, driving,
etc. can belong to this class. With proper training, most 'people can acquire this
set of ideas, concepts and operations and incorporate them into the core of HD for
solving their problems. Thus, habituation of the order of operations becomes an
important part in performing and solving the problems.
Class II. Mixed Routine Problems. To successfully execute or solve this
class of problems, one needs to solve variable combinations of simple routine
problems. The work of repa'lr shops, receptionists, waitresses, clerks, and regu-
lar dinner prepara~ion belong to this class. In this class, the set of ideas,
concepts and operations needed to successfully solve the individual component
problems are known, and can be acquired and habituated as the core of HD by proper
training. However" the best order to perform the job may not be fixed. There are
also time allocation problems over the individual component problems. One notices
that the problems of this class are more complex than those of Class I. The core
of HB in this class, in general, is an a-core with a smaller than that of Class
1.
Class III. Fuzzy Problems. There are no sure set of ideas, concepts and
operations that can guarantee the success of a solution, even though the set can
be fuzzily known. Sales problems, advertisement problems, ordinary research and
development problems, simple conflict problems with friends, problems of preparing
for an important dinner party, etc. are some examples of this class of problems.
There may exist a set of ideas, concepts and operations which are commonly but
fuzzily known to be good for solving the problems. This set of ideas may be ac-
quired and habituated through training and learning as to become a large propor-
'tion of the core of HD (i.e. a-core with high value of a) for solving the prob-
lems. However, this set alone is usually not adequate to surely and successfully
solve the problems. There is always a need of new ideas or concepts, which may be
acquired from the potential domain by active thinking and association, to success-
fully solve the problems. In 'general, a-core, with a large a, of the decision
maker's HD is not adequate for solving the problems. In fact, rigid and inflex-
ible HD (i.e. the a-core is almost fixed even if we lower the value of a) may
pTove to be detrimental to solving the fuzzy problems.
Class IV. Challenging Problems. These problems cannot be successfully
solved by a-core of HD, no matter how low we reduce the value of a. These kinds
329
of problems include innovative research and development· problems which challenge

the existing technical assumptions, .arket restructuring problems, complex con-
flict resolution problems, traumatic disasters., etc. The problems can be solved
only by expanding and restructuring our HD. Fixed mindedness (fixed HD) usually
becomes a major resisting force for solving challenging problems. In order to
expand our RDs we need to go through the process of implanting, accepting and
habituating.
In the athlete problem, if Sohn Coach has never had the experience of
coaching an Olympic star before, then transforming Tom Student into a star can be
a challenging problem for him. After he implants enthusiasm in Tom Student to
work all the way towards becoming a star, the training and supportive encourage-
ment may prove to be a fuzzy problem to him, at least .at .the very beginning. Once
both of them become good friends,.knowing, trusting and respecting each other, the
training problem may gradually become a mixed routine problem, etc. Notice that
through the process., the RDs of John and Tom are gradually changed and restruct-
ured as to become compatible and habitual.
Finally, we note that each one of us has solved some of the above four
classes of problems, and someone's routine problems may be others' fuzzy or chal-
lenging problems. The above classification depends on each individual's RD, which
results from his own experience and learning.
4.2 A General Principle of Effective Goal Setting
In this section we shall describe a general principle of effective goal

setting. The detailed applications of this prinCiple, depending on the RDs of the
performers and the kind of problems they are facing, will be explored in the next
two subsections.
Utilizing the concepts of Section 2 and 3, we can state the general princi-
ple of effective goal setting as follows:
"To select goal functions and their achievement levels, and to set up pro-
per supportive systems to create favorable working conditions, and relatively high
levels of charges and confidence on the performers so that a continuous allocation
of as much time as possible to perform the desired job-related operations in the
most efficient and effective way is the least resistance method for the performers
to release their charges (thus obtain their satisfactions)".
Note that the principle describe·s a complex optimal control system in which
the control variables are goal setting (selecting goals and their achievement
levels) and suppor.tive systems; the state variables are working conditions, charge
structures and confidence, which may be only partially observable; and the object-
ives are to maximize the attention allocation of time to job-related works, to
maximize the efficiency/effectiveness of performing the works to release the per-
330
formers' charge., and, pe1:haps, to matimize the favorability of the working en-
vironment. A mathematical formulation of this 'complex optimal control system will
be described in Section 4.5, after the preparat'i'on of Sections 4.3 - 4.4.
Let us first integrate the above principle with the behavior mechanism as
depicted in Flow Chart 2. In the flow chart we assume that physiological condi-
tions and external information inputs, except supportive systems, have been neu-
tralized and do not create high levels of charges or dis.turbances on the perfor-
mers. Thus they are put in the insignificant sub-box "others" of Box (4).
,---
I ,..-~-I-..., ..-~-I-...,(2)
Supportive
Systems
I
I
I
I
I
,
I
r - - - - - - , (7)
I
I Working
I r - - - -.......- - - - , ( 4 )
Environment
I
r----t---..------i
I I Others
,, L----'------r---'
Confidence
Training,
I Assistance,
I r---- Abi 11 ty
& Supervising
: .. ---1---,----1...----,
t- I Avoldonce
, Justification
1..------'----.,r----....
(9)
Drive and Monitoring
Performance
Flow Chart 2.
Note that in Flow Chart 2, the boxes in the right column (i.e. Boxes (7) -
(10» are components of the supportive system; while those in the left column
(Le. Boxes (1) - (6» are the modified "behavior mechanism". Let us explain the
principle using Flow Chart 2 as follows:
Boxes (1) - (2). Goal Setting and State Valuation. Although the func-
tioning of these two boxes are as in Flow Chart 1, they are subject to the influ-
ence and control of our purposeful goal setting and supportive system in addition
to physiological monitoring and other external information inputs. In Box (1),
"goal setting" has been used as a purposeful control on job performance as well as
331
its general meaning of setting the ideal value of living goals beyond those rela-
ted to job performance (1. e. including t-hose listed in Table 1.)
Boxes (3) - (4). Charge Structures and Attention. In order to obtain
attention allocation, the purposeful goal setting must be able to create the rela-
tively highest level of charges on the performer's chaTge structures. That is,
the charge of "not reaching the performance goals" is so much higher than that of
"reaching the performance goals", that no life goal functions other than the per-
formance goals can create a more significant charge. When the performance goals
can create a sufficiently high level of Charge, while all other life goals cannot
do so, the performers can have a burning desire to achieve the set performance
goals. To achieve this objective, we need to (A) create the rosy and promising
picture of the state when the performance goals ,are accomplished (positive re-
wards); (B) let the performers realize the sad and depressing state when the per-
formance goals are ~ accomplished (negative rewards or punishment); (C) set up
barriers so that job-unrelated messages or disturbances would not reach the per-
formers and create a high level of charge; and (D) develop good working environ-
ment in which the performers enjoy their work and in whi~h accidents are unlikely
to occur.
Note that in Box (4), there is a sub-box, Attention to Others, which in-
cludes physiological monitoring and external information inputs unrelated to job
performance. The more attention' paid to "others", the less attention paid to
job-related work. It may represent a leakage of time and effort with respect to
performance. Note that 100% of attention over a iong period of time to job--
related work is impossible, because men as living systems at least need to eat,
sleep, etc. to release the high level of charges cr.eated by physiological goal
functions (see [17 _,18] for further discussion).
Boxes (5) - (6). Positive Problem Solving or Avoidance Justification, and
Drive and Performance. As mentioned in Section 2, positive problem solving is to
work hard to accomplish the preset goal values, while avoidance justification is
to lower the ideal goal value or the significance of the problems. Thus, after
attention, the attitude of positive problem solving will produce "drive" and per-
formance; while that of avoidance justification will usually not. Besides person-
ality, confidence, hope and the ability to accomplish the ideal goal value in an
acceptable time horizon plays an important role in deciding which attitude will be
taken. Using supportive systems, especially training, as'sistance and supervising
(Box 8), to increase the performer's confidence and ability is an important means
to enhance the attitude of positive problem solving and to improve the perfor-
mance.
Boxes (7) - (10). Supportive Systems These four boxes are some important
devices for supportive systems. They help create or modulate charge structures as
to make performers pay more atteution to job-related problems, cultivate a posi-
332
tive problem solving attitude, improve coafidence and ability to accomplish the
ideal goal value, and to actually achieve the set goal value. We shall further
explore the impact of these supportive devices on performance in Section 4.4.
Note that by utilizing Flow Chart 2, one eould desc~ibe a dynamic system
for the problem of effective goal setting (see Section 4.5.) ~ith proper assump-
tions such as additivity with respect to time, the dynamic system can even be
expressed in terms of di:fferential or integral equations. To avoid detraction, we
shall not stop here to do so.
4.3. Parameters in Goal Setting
4.3.1 Selection of Goals
As we want to create a high level of charge and maximize the time allocated
to job-related work, when selecting goals we shall pay attention to performers'
habitual domains. To be effective, the selected goal should be able to strike
resonance on the performers' HDs. If the performer is striving for survival and
needs money, then .the goals directly related to monetary reward will catch his
attention. On the other hand, if the performer is loaded with money and looking
for social recognition, then the goals related to social recognition may catch his
attention. If the performer has already reached a mature state of self-actualiza-
tion and can fai.rly easily detach himself from the attraction of money and social
recognition, then the goals related to independence and freedom of the self may
catch his attention.
In Table 1, we list a fairly inclusive collection of goal functions that
may affect our behavior and charge structures. One way to visualize the work of
these goal functions is to open a telephone directory or to think of various in-
dustries and businesses which try to satisfy some peoples' needs generated by
these goal functions. However, to transform the goal functions listed in Table 1
into a set of vivid goals that can strike resonance in peoples' HDs, we need to
actively study the interested peoples' HDs as well as use our ability of associa-
tion and creativity. For instance, "salary" may be more striking than "accumula-
tion of wealth"; "firing from job" than "self-importance feeling"; "sex" than
"giving birth to the next generation", etc. The billboard which says "Driving Is
Like Making Love, Don't Rush to Finish It" may be more effective to people than
one which says "Don't Speed". We shall leave these to the reader's creative imag-
ination.
In addition, to be effective the selected goals must be clear, easily mea-
surable and closely related to the overall performance that we are looking for.
In the athlete problem, the time needed for 100 meters is clear, easily measurable
and directly related the overall performance; while the streamlining of movement
333
is directly related but less clear and more difficult to measure even with the aid
of video tapes. The clear and measurable goals make monitoring and feedback eas-
ier, which can help performers' "state valuat:ioo" to cr.eate charges (Box (3» and
to enhance the confidence and positive problem solving (Box (5».
Finally, we note that for Class I-II problema. as the set of ideas, con-
cepts and operations for successfully executing the jobs is known and almost
fixed, the selection of clear and easily measl1'l'able goals is lDUch easier than that
of Class III and IV. In the case that no clear and measurable goals can be found,
one may try "I will spend two hours in the morning en some section of the problem,
two hours in the afternoon on other section of the problem" (Le. we use time
allocation as goals; "two hours" is arbitrary, just for an example).
Through implanting, accepting and habituation, the selected goals will have
impact on our behavior and overall performance. As time passes, it may be a good
idea to review whether the habituated goa'ls are still effective for overall per-
formance.
4.3.2 Goal Levels
Once the set of goals are selected, we need to specify at what level and
within what time frame (minutes, hours, days, weeks, etc.) in order to achieve
optimal performance.
The more specific and difficult the level, within the acceptable range, the
more likely is the setting to create a high level of charge and obtain the atten-
tion allocation. As a consequence, more drive and performance will be produced.
In the extreme, in Class I of simple routine problems, industrial engineers have
used motion and time studies to set the "optimal" level of production outcome for
each worker, which has proved to be very successful in so called "scientific man-
agement".
Note that the high level of goals can create a high level of charges, which
may be short-lived if the workers do not have the confidence and ability to a-
chieve it (through avoidance justification (Box (5». Proper training, assistance
and progress report are some important means for maintaining the confidence to
achieve the set goal levels. Again through implanting, training, accepting and
habituating, the performers of Class I-II (simple routine - mixed routine prob-
lems) can come to accept and commit themselves to achieving the set goal levels
and improve their proficiency in job performance.
For the problems in Class III-IV (fuzzy and challenging problems) the time
to achieve the set goals may be highly uncertain which may thwart the confidence
to continue the work. In addition to a good support system, the following may be
helpful in maintaining the confidence and drive to achieve the goals: (i) decom-
posing the overall project into a number of subprojects and trying to complete the
334
subprojects one at 3 time; (i~) using self-suggestion to see the rosy and promis
ing picture of completing the project to reinforce the burning desire (charge) to
achieve the goals; and (iii) caLling for belp from experts and other useful sour-
ces. In the athlete problem, the coach can set ~e goal level progressively,
starting with 10.8 seconds, then 10.6, 10.• 4., 10.2 secous, ••• etc.; he and the
student can renew and reinforce the rosy picture of becoming an Olympic star to
encourage themselves.
Finally, in ~etting the goal level for other people to perform, partici-
pation of the workers or performers may be important. ~his is especially so when
the problems are of Class III and IV (fuzzy and challenging). Participation al-
lows us to better see the HDs of performers and bet-ter under·stand the problems.
In addition, because of participati-on., the perfoT1llerc may obtain a feeling of
importance and be more willing to .commit themselves to achieve the set goal le-
vels. However, for simple or lIIi:xed routine prob-Iems (Class I or II), as the set
of ideas, concepts and operations needed for successful execution of the job is
known, more objec·tive assessment for .achievement level's. can be obtained (for in-
stance, motion and time studies) which can be used to set the goal level effect-
ively. In these cases, the impact .Q!I overall performance by worker participation
in setting goal level msy not be significant.
4.4 Parameters in Supportive Systems
4.4.1 Working Environment Controls (Box 7)
The following are some devices:

(i) Layout. This can affect the production efficiency and the ability for
the performers to achieve the preset goal levels. This is indeed an important
topic in motion and time ·studies for increasing productivity.
(i~) Safety. Danger-free working environments, and accident prevention
and ·treatment are the major concern. Dangers and accidents can create high levels
of charges and drain.the energy and time from performing job-related work. To
maintain safety in working place is therefore very important for productivity.
(iii) Barrier. Proper insulation of the performers from unnecessary in-
terruption from outside world can prevent charges from the outside world and leak-
age of attention to job-unrelated events.
(iv) Atmosphere. A peaceful but competitive atmosphere in which the per-
formers can excel can increase the ability and confidence to reach and surpass the
preset goal levels.
Note that all of the above four devices are helpful for problems of all
Classes (I-IV). However, the degree of rigidity msy vary. In general in the
higher classes (III or IV), certain degree of flexibility is needed to stimulate
335
creativity for problem solving.
4.4.2. Training, Assistance and Supervising (Box 8)
Training and assistance can increase. the perforaers. ability and confidence
to achieve the preset goal levels; while supervising ean, in addition, create
their charges and attention to job-related work- Note that training involves
implanting, accepting and habituating processea of ideas, concepts and operations,
which is very important for the beginners to achieve the preset goal levels, es-
pecially those doing simple and mixed routine works. Assistance is usually needed
when unusual problems occur in the routine work. Timely assistance can reduce the
performers' frustration and leakage of attention to work, in addition to increas-
ing their ability to cope with unusual problems.
4.4.3. Results Monitoring and Feedback (Box 9)
Human beings have an innate need for external information to fulfill their
life goals. Information on their performance and rewards are important to ·them
for assuring themselves how'well they have been doing (related to self-important
feeling and social approval category of Table 1). This information is an import-
ant input to State Valuation (Box ~2», which can affect charge structures, atten-
tion, confidence, drive and overall performsnce.
In the athlete problem, keeping record o~ the p~gress of Tom Student and
showing it to him may provide a main thrust to increase his drive for further
improvement. Note that more detailed monitoring and feedback could also improve
the training and ability to perform.
In fuzzy and challenging problems (Classes III-IV), the monitor and feed-
back may be more difficult. People may be periodically self-assessing as to as-
certain the progress and situation of the problems.
4.4.4. Reward Systems
We work to fulfill our life goals (Table 1). The reward of the work is
thus vital in providing the charge, attention and drive to achieve the preset goal
levels. Usually, the larger the reward, the larger the charge and drive to accom-
plish the goals.
Note that the reward systems should be broadly considered. These include
monetary payment, authority, power, position and promotion. Not only positive
rewards but negative rewards need to be considered. To be effective, the reward
systems need to create for the performers a large difference between the charge of
~ achieving the goals and that of achieving the goals. The large difference,
which signifies the importance of achieving the goals, can command the attention
336
allocation to perform the work. HoweveT, i f there is no pToper training, support,

and confidence. the attention can be sbort and the d~ve needed for the work may
not sustain. This problem can be severe especially for fuzzy and challenging
problems. In the athlete ·problem. if Tom Student has no high regard for the re-
ward of becoming an Olympic sta~. or has no confidence in obtaining it. then most
likely he could not generate the continued high level of drive necessary to become
an Olympic star.
4.5 A Complex Optimal Control System

Wi~h the above discussion. we could summarize the principles of effective
goal setting into the following 'complex optimal control systems. (Refer to Flow
Chart 2 and Section 4.2).
Let Ut .. (u1:1:">" .upt ), Nt E: U CRP be the 'vector control variable after
p goal functions are selected. (recall from' Section 4..:3 that selection of goal
functions is a part of goal setting);
V t ,. (vl1:'••••• v qt ). V t E: VCRq" be the vector control variable for the
q-supportive parameters (See Section 4.4).
N'ete that U and V represent the sets in which the respective control para-
meters can be selected.
Now for the state variables let
Xt (xl t ' •••• xit ) be the working conditions.
Y .. (Ylt' ····yjt) be the charge structures.
t
Z .. (ZIt' •••• zpt) be the confidence level to achieve the set goal levels
t
with Zkt .. 1 or 0 being 100% or 0% confidence to achieve the kth set
goal level.
bet t be the external environmental inputs.
E: Then.
Xt+t.t Ft (X t •Yt·Z t ; Ut·V t ; E: t ) (3)
Gt (Xt·Yt·Z t ; Ut·V t ; E: t ) (4)
Yt+t. t
Zt+t. t Ht (X t •Yt·Z t ; Ut·V t ; E: t ) (5)
With suitable assumptions. (3) - (5) may formulated into differential equa-
tions. We shall leave this to the reader.
For the objective functions. let the attention allocation at time t
be
(6)
with
if at time t attention is paid to the work
otherwise
Note that the total attention time allocation to the work during the time
337
interval [O,T] is given by
At = J
T
at(X t , Yt , Zt; Ut , Vt;£t) dt (7)
°
Now let the efficiency and effectiveness in perf~rming the work be repre-
sented by
Et ....lIt = Et(Xt,Yt,Zt; Ut,V t , £t) (8)
Then the principle of effective goal setting for the time horizon [O,T] can be
speci~ied as (i) selecting p clear and measurable goals and determining their
achievement levels, Ut £ U; and (ii) selecting t~e .supportive systems Vt £ V, so
that the total attent-ion time a-llocation to the work "'r and the efficiency and
effectiveness to perform the work ET are maximized, where ~ and ET satisfy the
dynamics of (3) - (8).
Many solution concepts can be utilized for solving the above two criteria
problems. For instances see [18,21].
Note that the favorability of the workingenvirO'l:lment, which may be impor-
tant to fuzzy and challenging problems (Classes III and IV), can be similarly
formulated as an objec~ive function. We shall leave it to the reader.
Also, instead of two criteria,
and ET , maximization, we may consider
~
only one criterion by aggregating the two criteria into one. For instance, one
may simply consider the performance index of
T
PT ~ at • Et(Xt'Yt,Zt;Ut,Vt;£t)dt
when proper assumptions are made to make the above integral meaningful. Again, we
shall leave this to the reader for further exploration.
Finally supportive systems can be costly. The reader may want also to
consider cost in the formulation.
5. Some Empirically Known Results
In this section we shall sketch some empirically known results of goal

setting. As mentioned before, these known results, due to the need for statis-
tical studies, are limited to simple routine or mixed routine problems. The rea-
der may find it useful to verify the principle described in the previous section
by using it to explain the empirical results.
5.1 Goal Acceptance
Goal acceptance is known to be a prerequisite for goal setting to affect

task performance. The study by Erez and Zidon [5), using 140 technicians and
engineers attending technical training courses of a research and development or-
338
ganization as subjects to perform the task called completing a perceptual speed

test which required them to determine how many digits or letters in a row were the
same as the circled one to the left of each row, found that the relationship be-
tween goal difficulty and task performance is linear and PQsitive when goals are
accepted; while it is linear and negative .whea they are rejected. Goals are
viewed as having significance for task performance only to the extent that they
are accepted by the individual. A goal which is rejected is almost impossible to
regulate task performance. Internally accepted goals, rather than externally set
goals, determine the effect on task performanee. When a goal is perceived to be
impossible, it will simply be rejected by the individual.
5.2 Goal Difficulty
Assuming sufficient ability, Locke and Bryan's research in 1966 [10], using
70 university students to perform the Complex Computation Task, found that there
was a positive relationship between goal difficulty and task performance. Impli-
citly, it was assumed that the goal is internally accepted by the individual. In
the 1981 review, Locke, Shaw, Saari and Latham [12], concluded that, since 1969,
48 studies partly or wholly supported the hypothesis that hard goals lead to bet-
ter performance than medium or easy goals. 9 studies failed to support it. Ia
general, it is valid to say that difficult, challenging, but realistic goals will
enhance task performance. Although we know the positive relationship between goal
difficulty and task performance, we are not sure whether the performance will
increase at an increasing, decreasing, or constant rate with the increase in the
degree of goal difficulty.
5.3 Goal Specificity
By asking 10 male and 10 female paid college student volunteers from Uni-
versity of Maryland to perform a simple addition task, Bryan and Locke [1], con-
cluded that specific goals can serve to raise the motivation of subjects to a
higher level. The assigning of specific and hard but realistic goals to the sub-
jects raised the task performance and favored the development of more positive
attitudes toward the task. Dossett, Latham and Mitchell [3], with female clerical
personnel as subjects solving simple arithmetic problems, found that specific
goals led to higher performance than did the "do best" goals. Similar results
were' found in Locke [9] and Locke and Bryan [10] studies, by using college stu-
dents as subjects to do simple arithmetic problems. In their 1981 review [12],
Locke, Shaw, Saari and Latham stated that individuals with specific and hard or
challenging goals outperform individuals with specific easy goals, "do best"
goals, or no goals. When goals are set in quantitative terms, they tend to be
clearer, and more specific to the individual.
339
5.4 Knowledge of Results
Knowledge «>f re'sults has long been an important variable in goal setting
theory. Many studies have been done to determine its effect on task perform-
ance. Cummings, Schwab, and Rosen'.s [3] field experiment with college students as
subjects performing simple addition task showed that correct knowledge of results
(KR) increased goal level significantly above that generated by no KR. Dossett,
Latham and Mitchell [4], through asking 60 female clerical personnel to perform
simple addition, concluded that, with goal difficulty held constant, there was no
significant effect found for KR on task performance. Similar results were found
in Locke [9] and Locke and Bryan's studies [10]. Both studies used college stu-
dents to perform simple arithmetic work. However, given specific, hard goals, a
significant goal effect was found between KR and no KR subjects. In conclusion,
[12] stated that KR appears to be necessary if goals are to improve performance.
KR achieves its motivational effects through the incorporating of goal setting and
that in the absence of goal setting KR alone is not sufficient to improve perfor-
mance.
5.5 Rewards
60 male and female subjects, 17 to 19 years of age, were hired to work 5

hours a week in a simulated company to perform the job of working on programmed
texts designed to teach introductory principles of electr±city. It was found that
money can be a powerful motivating factor for task performance by Terborg [16].
Money could affect the level at which goals are set or the level at which inten-
tions are established; it might create more spontaneous goal setting than would
occur with no money; and it might affect the individual's degree of goal commit-
ment. Money itself might even become a goal for an individual to attain. Mone-
tary rewards have independent results on task performance [16]. If money alone is
to 'affect task performance, the amounts involved have to be large rather than
small [12].
5.6 Participative vs. Assigned Goals
Using 60 female clerical personnel to perform simple arithmetic work, Dos-

sett, Latham and Mitchell's study [4] indicated that, with a constant level of
goal difficulty, there was no significant difference between assigned and partici-
pative goals on task performance. But goal attainment was higher under the
assigned condition. Individual differences exerted influence on attaining goals
with high-esteem individuals who received KR having attained goals more often than
those with low-esteem when goals were set participatively. A field experiment
340
conducted by Ivancevich [8], using sales personnel as subjects to attend a sales

training, found that both the participative and assigned goals enhanced task per-
formance and that participatively set goals were not superior to those assigned.
If the goal difficulty was held constant, then again, no. significant differences
between assigned goals and parttci~ativeLy set goals were found [12]. However,
participation may lead to higher goal commitment, and to the acceptance of harder
goals.
5.7 Supportiveness
A'longitudinal field stuuy of two medi~sized organizations which have

implemented MBO programs by Ivancevich [6] f.ound that an active participation role
by management in the design and "implementation of "1IIIlnagement by objectives" (MBO)
can have "a significant impact on improving the job satisfaction of managers. In
another study [7] involving the diagnosis and implementation of an MBO program for
a manufacturing corporati.0n in its six plants, Ivancevich stated that MBO being
supported by the company's president had some influence on the participants. A
factor analysis of a questionnaire administered to 292 pulpwood producers indica-
ted that goal setting was correlated with high productivity and a low number of
injuries only when it is accompanied by supervision (Ronan and Latham (15]). [12]
also concluded that supportiveness is an important factor for goal setting to
affect task performance. Supportiveness can lead to the setting of higher goals
and greater goal acceptance or commitment.
5.8 Ability, Etc.
The experiment conducted by Locke, Fredrick, and Bobko [11] by using col-
lege undergraduates as subjects to perform the job of giving uses for common ob-
jects indicated that ability, self-efficacy, goals, and task strategies were all
related to task performance. No reliable individual difference factors, other
than ability, have emerged in the goal-setting literature, probably because most
of the studies have used assigned goals [12]. Other things being equal, greater
ability should lead to better task performance.
5.9 Effective Time
In the study of two medium-sized organizations which adopted MBO programs,

Ivancevich [6] found that, twenty months later, since MBO had been introduced, its
impact reduced to be insignificant compared to before the program. He suggested
that some form of reinforcement is necessary for MBO to be effective over a longer
time period. In another study [7] conducted in six plants of a manufacturing
corporation producing machine parts for other industrial users, he found similar
results and confirmed the necessity of reinforcement for MBO's long-term ef-
341
fects. There is probably an optimal time span for the setting of goals that may
depend on both the individual and the task situation [121.
6. Conclusion
We have offered a general princ~ple of effective gQal setting for various

classes of decision problems. Many research problema remain to be explored. For
instance, from a theoretical and mathematical point of ~ew, by restricting our-
selves to a specific class of problema and situations, with proper assumptions
more concrete results can be specified or statts·tically tested. And from a prac-
tical point of view, the principle can be applied to our d~ily life. The applica-
tion is an art. The more we practice, the more effective we will be. The reader
may find it beneficial to do so.
References
1. Bryan, J. F., and Locke, E. A. "Goal Setting As A Means Of Increasing

Motivation," Journal of Applied Psychology, 1967, 51, PP. 274-277.
2. Chan. S. J., and Yu, P. L., "Stable Habitual Domains: Existence and
Implications," Journal of Mathematical Analysis and Applications, Vol.
110, No.2, 1985, PP. 469-482.
3. Cummings, L. L., Schwab, D. P., and Rosen, M., "·Performance and Knowledge
Of Results As Determinants Of Goal Setting," Journal of Applied Psy-
chology, 1971, 55, PP. 526-530.
4. Dossett, D. L., Latham, G. P., and Mitchell, T. R., "Effects of Assigned

Versus Participatively Set Goals, Knowledge of Results, and Individual
Differences on Employee Behavior When Goal Difficulty Is Held Con-
stant," Journal of Applied Psychology, 1979, 64, PP. 291-298.
5. Erez, M. and Zidon, 1., "Effects of Goal Acceptance on the Relationship

of Goal Difficulty to Performance," Journal of Applied Psychology,
1984, 69, PP. 69-78.
6. Ivancevich, J. M., "A Longitudianl Assessment of Management by Objec-

tives," Administrative Science Quarterly, 1972, 17, PP. 126-138 ••
7. Ivancevich, J. M., "Changes in Performance in a Management by Objectives

Program, Administrative Science Quarterly, 1974, 19, PP. 563-574.
8. Ivancevich, J. M., "Effects of Goal Settiag on Performance and Job Satis-

faction," Journal of Applied Psychology, 1976, 61, PP. 605-612.
9. Locke, E. A., "Moti.vational Effects of ttnowledge of Results: Knowledge

or Goal Setting.," Journal. of Applied Psychology, 1967, 51,
PP. 324-329.
10. Locke, E. A., and kyan, J. F., "The Effects of Goal-Setting, Rule-Learn-
ing, and Knowledge of Score on Performance," American Journal of Psy-
chology, 1966, 79, PP. 451-457.
342
ll. Locke, E. A., Frederick~ E., and Bobko, P., "Effect of Self-Efficacy,
Goals, and Task Strategies. on.Task Performance,"Journal of Aplied
Psychology, 1984, 69, PP. 241-251.
12, Locke, E. A., Shaw, K. N., Saari, L. M., and Latham, G. R., "Goal Setting
and Task Performa'llce: 1969-19Se.~" Psychological Bulletin, 1981, 90,
PP. 125-152.
13. Mackenzie, K. D.~ Organi'zatiortal Design, the Organizational Audit and

Analysis Techn~logy, Ablex Publishing Cor.~ Norwood, New Jersey, 1986.
14. Miner, J. B., Theories of Organizational Behavior, The Dryden Press,

Hinsdale, Iliinois, 1980.
15. Ronan, W. W., and Latham, G. P., "Effects of Goal Setting and Supervision
on Worker Behavior in an Industrial Situation," Journal of Applied
Psychology, 1973, 58, 302-307.
16. Terborg, J. R. "Tlte Motivati'Onal Components en Goal Setting," Journal of

Applied Psychology, 1976, 61, PP. 613-621.
17. Yu, P. L., "Behavior Bases and Habitual Domains of Hu1JIaU Decision/-
Behavior - An Integration of Psychology, Optimi2stion Theory and Com-
mon Wisdom," International J-ournal .of Systems, Measurement and Deci-
sions, 1981, 1, PP. 39-62.
18. Yu, P. L. 'Multiple Criteria Decision Making: Concepts, Techniques and

Extensions, Plenum, New York, New York, 1985.
19. Yu, P. L., Behavior Mechanism and strategical becisions, the Foundation
and Applications of Knowing Yourself and Others, Academic Sinica
Press, Institute of Economics, Taipei, Taiwan, 1985.
20. Yu, P. L., and Huang, S. D., Behavior Bases and Habitual Domain Analysis
- Theory and Applications, (to be published).
21. Zeleny, M., Multiple Criteria Decision Making~_~cGraw Hill, New York, New
York, 1982.
IV. ENGINEERING APPLICATIONS
MULTICRITERIA OPTIMIZATION
PROCEDURES IN APPLICATION
ON STRUCTURAL MECHANICS SYSTEMS
by
Hans A, Eschenauer
Research Laboratory for Applied Structural OptImization
at the
Institute of Mechanics and Control Engineering
University of Siegen FR GERMANY
J
SUMMARY: This paper reports on investigations carried out

in the field of vee~o4 op~~m~za~~o~ and its practical applica-
tion on a structural mechanics system. The main reason is that
in addition to the minimization of costs for developing and
manufacturing machines and plants, some other objectives such
as shape accuracy, reliability and others are playing an
important role as well. These problems can be formulated as
"Optimization Problems with Multiple Objectives" (Vector-
Optimization, PARETO-Optimization). Most of the objectives are
non-linear and besides that competing. Thus, they do not lead
to one or several solutions for the optimum but rather to a
"functional-efficient" solution set, Le. the decision maker
is able to 'select out of this set the most efficient eomp4o-
m~~e ~otu~~o~. Moreover, the vector optimization problem is
transformed into a scalar substitute problem by means of a
preference function. This so-called optimization strategy is
an important part of the modett~~g. For carrying out the trans-
formation, several preference functions, e.g. objective
weighting, distance functions, constraint oriented trans-
formation (Trade-off Method), and Min-Max-Formulation have
been analysed and tested. The investigations showed that the
efficiency of the single preference function is problem-
dependent and furthermore dependent on the adapting on special
optimization algorithms (Methods of Mathematical Programming).
Within the scope of this contribution the application of
multicriteria optimization techniques in structural mechanics
systems is shown, e.g. the shape optimization of a special
shell structure (conveyer belt drum).
1. INTRODUCTION
By using the prinCiples of minimum potential energy respectively

complementary energy, the fundamental equations of the boundary-
value problems in the theory of elasticity can be found. As we
know for an elastic body in equ~l~b~~um, the displacements in
the body will be such that the potential energy is a minimum
and the stress components will be such that the complementary
energy is a minimum. One can show that, by means of the methods
in the calculus of variation, the governing differential equa-
tions and the boundary conditions can be derived as a conse-
quence of these minimal principles.
346
A further important application of minimal principles is the

use of approach functions. Instead of solving the governing
differential equations together with the boundary conditions,
a task often mathematically very difficult, we may interpret
the problem as such a one seeking the functions which satisfy
the boundary conditions and minimize the potential energy rr or
the complementary energy rr*. Two well-known variational methods
are the approach methods by Lord RAYLEIGH, W. RITZ and B.G.
GALERKIN [1], [2]. G.W. von LEIBNITZ (1646-1716) and L. EULER
(1707-1783) also established the mathematical tool for finding
the extreme values of given functions by introducing the infin~
tesimal calculus. Thus, the principles of mechanics allowed the
formulation and solution of classical problems by means of the
calculus of variation. Examples of these are the "curve of the
shortest falling time" ("brachistochrone") and the isoperi-
metric problems investigated by JACOB BERNOULLI (1655-1705)
and DAVID BERNOULLI (1700-1782). Another interesting example
is the finding of the smallest resistance of a body of revol-
ution by I. NEWTON (1643-1727). With the principle of least
efficiency and the integral principle, J.L. de LAGRANGE
(1736-1813) and W.R. HAMILTON (1805-1865) contributed to the
perfection of the calculus of variation which still is the
basis for various optimization methods. J.L. de LAGRANGE,
T. CLAUSEN and de SAINT-VENANT, for example, concentrated on
finding the optimal design of one-dimensional supporting
structures under various loads. Typical examples are the bar
exposed to buckling respectively torsion or the cantilever
beam for which optimal cross-sectional shapes were found by
means of the calculus of variation. For this purpose, opti-
mality criteria are derived as necessary conditions, e.g.
EULER's equations in the case of unconstraint problems. If
constraints are considered, LAGRANGE's multiplier method is
applied; it corresponds to the solution of an isoperimetric
problem [3]. In the field of applied optimization many publi-
cations during the last decades have been based on variational
principles, among others [4]. In recent papers about Mathemat-
ical Programming Algorithms the optimality criteria can be
integrated into optimization procedures very efficiently.
347
However, these mathematical optimization methods are very

seldom applied in industrial practice. There are certainly
a number of reasons for this, e.g. the general opinion that
optimization has always been a major task of an engineer. But
the "optimization method of a practitioner" is nothing more
than a "trial and error"-method which often enough is both
ineffective and expensive. There are also some other reasons
explaining why the optimization theory is not frequently
applied in engineering environment. One is the practitioner's
demand for reliable, efficient and robust methods which can
be applied easily and problem-independently. Experience shows
that this demand can hardly be met. Thus, it is hard to close
the gap between theory and practice. This problem will remain
a major research objective as products and plants simply call
for best possible designs in the future.
The following aspects show the necessity of introducing opti-

mization procedures into the practical design phase:
1) Increasing the quality and quantity of products and plants
and at the same time reducing costs and thereby being
competitive.
2) Fulfilling the permanently increasing specification demands
as well as considering reliability and security proofs,
observing severe pollution regulations and saving energy
and raw materials.
3) Introducing inevitable rationalization measures in devel-
opment and design offices (CAD, CAE) in order to save more
time for the staff to work creatively.
As demanded in a multitude of applications the optimal layouts

of constructions require special attention if specified more
than one objective respectively criteria corresponding to the
aspects 1) and 2). Such an optimization problem for mutt~pte
objectives is also called vecton on potyopt~m~zat~on (mutt~
object~ve on mutt~cn~ten~a opt~m~zat~on). With reference to
v. PARETO (1848-1923), the French-Italian economist and
sociologist, who established an optimality concept in the
348
field of economics based on a multitude of objectives, i.e.

on the permanent conflict of interests and antagonisms in
social life, this is also called Pa~eto-opt~m~zat~on [5].
The application of vector optimization on problems of struc-

tural mechanics or technology in general took quite a long
time. It was W.STADLER [6], [7], who in the 1970's for the
first time referred to the scientific application of Pa~eto'~
opt~mal~ty concept, and who published several papers especially

on natural shapes. From end of the seventies, vector optimiza-
tion has been more and more integrated into problems of
Optimal Design in the papers of a number of scientists (e.g. [8],
[4], [9], [13], [10], [11], [12], [14]).
EnVironmentol conditions
Loadings
{ Objectives Main fl (K)
Various Demands Side f, (K). .. f. (xl
IIII Sh ap e conditions
:\
Constraints 9. (~) sO
Feasibility
Study
,,,..,. ;'''";';"~
Global behavior
Project -Compliance
I--
/
Definition Phase with the most
important
demands
I Appl ication: Highly Accurate

Focussing Systems
Development
\
Phase Local behavior
-Compliance with
I
certain demands
1/
for single
components
(subsystems)
Design Phase
(Subsystems)
Appl ication: Shape
Optimization
I
Manufacturing
Erection
Tesling
349
2. MATHEMATICAL FUNDAMENTALS
The objective of structural optimization is to select the

values of the design variables x i (i=1, ••• ,n) under consideration
of various constraints in such a way that an objective function
f=f(~) receives an extreme value. This reads in the abbreviated
form:
Min {f (~) (2.1)

x E IR n
with
IR set of real numbers,
f objective function,
XElRn vector of n design variables,
~ .vector of p inequality constraints,
h vector of q equality constraints (e.g.

system equations for the determination
of stresses and deformations)
X:= {x € IRn : g(~) = Q, ~(~) ~ Q}

"feasible" domain where ~ has to be inter-
preted for each individual component.
The problems of structural optimization are that the objective

function and the constraints are commonly non-linear functions
of the design variable vector xERn where the continuity of the
functionals as well as their derivatives are assumed as given.
For dealing with a pnoblem w~~h mult~ple objee~~ve~ a design

variable vector ~ renders the m eomponen~~ of the objective
function vector as small as possible and fulfills all constraints.
As a modification from (2.1) this vector optimization problem
(VOP) reads as follows:
"Min" h (x) (2.2)

x E IRn
350
Vector optimization problems have difficulty in finding

appropriate solutions in order to consider the multiple
objectives in a given way. A characteristic of such optimiza-
tion problems with multiple objective functions is the appear-
ance of an objee~~ve eon61~e~, i.e. none of the feasible
solutions allows the simultaneous optimal satisfaction of all
the objectives, respectively the individual solutions of each
single objective function differ. Con~equen~ly, ~he ~ubjee~
06 vee~on op~~m~za~~on ~~ ~o maRe ~~a~emen~~ 60n objee~~ve

eon61~e~~ .
When dealing with various fields of vector optimization the

following definitions are of importance:
Definition 1: Efficient Vectors (Fig. 2.1)
A vector x* E X is efficient with regard to X exactly if no

vector x E X of the following quality exists:
xi < xi* for all i E (1, ... ,n)
x.1 < x.*

1
for at least one i E (1, ... ,n) .
This is called:
"K-Efficiency" (by Koopmans)

"PARETO-Optimality" (by PARETO)
.;.;
~-
,'"
,
I I
I
/ ax
1/
'/
Ug. '2.. 1 ;
,/ E66~e~en~ vee~on
.- '"
"" and eolian.
=--...Bcr--'~~ .:: -: ~ __
351
Definition 2: Efficient Collar
The vectors efficient with regard to X constitute the efficient

collar of X:
ax* := {x* E X; x* is efficient with regard to X}
Definition 3: Mappings in n-dimensional space
Let Rn resp. Rm be the n- resp. m-dimensional EUCLIDean vector

spaces. A system of m real functions fj E F 1 n in n variables
x 1 ' ••• , x n ' j = 1 , ••• ,m ,
x c: Rn -+ R
fj {(X 1 '···,Xn ) -+
Yj f j (x 1 ,···,xn )
with the common domain of definition X c: Rn can be considered

as a mapping of one part of Rn into one part of Rm:
To each ~ E X exactly one y E Y is attached, but each y E Y

can be the map of several x E X. The set of all mappings is
n
termed Fm:
The single components f 1 ..• f m of f are called coordinate

functions of the mapping !.
CHARNES and COOPER have formulated a further efficiency theorem

referring to the coordinate functions f 1 , ••. ,fm of the mapping !.
Definition 4: Functional-efficiency of PARETO-optimality
A vector x* E X is then and only then functional-efficient or

PARETO-optimal for the problem (1.2), if there is no vector
x E X with the characteristics:
352
f. (X) ~_ f. (x*) for all i E {1, ••• ,m} and

J - J - (2.3)
f. (X) < f. (x*) for at least one i E {1, •.• ,m}
J - J -
For all vectors being not PARETO-optimal the value of at least

one objective function fj can be reduced without increasing the
functional values of the other components. Fig. 2.2 shows a
mapping of the two-dimensional design space X into the objective
function space or the criterion space Y where the PARETO-optimal
solutions lie on the curve section AB.
Criterion SRoce
Fig. 2.2: Mapping 06 a 6ea~ible de~ign ~pace

into the c~ite~ion ~pace
Solutions of non-linear vector optimization problems can be

found in different ways. By defining so-called substitute
problems, these are normally reduced to scalar optimization
problems. This means a compromise solution x
out of the
complete solution set X* can be found, whereby
- X* is the set
of all x* as in (2.3).
Definition 5: Substitute problem and preference function

The problem
Min p[f(~)]
(2.4a)
xEX
353
is then transformed into a substitute problem if one x E X*

exists with the characteristics:
p[! (g)] = Min p[! <.~)] • (2.4b)

x EX
The function p is called pne&enenee &unet~on or ~ub~t~tute
objeet~ve &unet~on [8], [15].
It is obviously important to prove that the solutions of all

substitute problems g are PARETO-optimal or functional-efficient
with respect to X and the set of objective functions f 1 , ••. ,f m,
i.e. that a point y = !(g) actually lies on the efficient
boundary Cly* [6], [1 6] .
3. ESTABLISHING AN OPTIMIZATION PROCEDURE
An optimization procedure, applicable to all problems and at

the same time efficient, will be very difficult to realize.
Super-computers and vector computers may help here in future,
but even then the circle of those who use computers with such
a high efficiency will remain relatively small. Of course, the
establishing of architectures of program systems for the opti-
mization of component parts and structures should aim at a
certain universality in terms of application to various types
of constructions. Such a conception which was established at
the Research Laboratory For Applied Structural Optimization
of the University of Siegen shall be described shortly. The
partition of the optimization process into three subproblems
provides the premises for the development and establishment of
the software system SAPOP (~tructural ~nalysis ~rogram and
Qptimization ~rocedure) which stands out on account of its
modular structure and defined sectional quantities.
The main program SAPOP prepares the data exchange between the
optimization algorithms and the structural analysis. The actual
optimization process is coordinated by the selected optimiza-
tion algorithms. The respective structural analysis program is
constantly integrated into the computations being made in order
354
to analyse the current designs with regard to objective function

values and constraints. The essential characteristic of this
program structure is its relatively unproblematic exchange of
program components (modular techniques). In addition, it is
possible to include further subprograms necessary for the
treatment of special design problems into the program library.
Optimization
Algorilhms
Model of
actual design .r.
SEOLI 2
SEJlML I--=-J!_""I OptimlZotlon
SLIP40 Modelling
~ I--'~-:-i.-g~-:-i-I FUNC
t:JLe!2l".
GREGA
LPNLP
SONP
EXIBEM
E.S!l
OptimizatIon Pos'~ond Pre Structural
Modellmg Processing Analysis
F~9. 3.1: Block d~a9~am 06 th~ S06twa~~-Sy~t~m SAPOP
3.1 Optimization Algorithms

Most solution algorithms are iterative, i.e. a starting vector
~O and successive application of the algorithm will yield
respectively "improved" vectors ~1'~2' ••.• This iteration

sequence for l~n~a~ optimization problems is finite in length.
Thus, the solution x* is exactly reached after a finite number
of steps, conv~~9~nc~ towa~d~ th~ ~olut~on po~nt can only be
expected for non-l~n~a~ problems~ the process is finished when
a pOint "sufficiently close" to the solution is reached.
In recent years, mathematicians have developed fairly efficient

and reliable algorithms with regard to mathematical efficiency
355
and numerical precision. Nevertheless, certain problems still

remain unsolved. In order to find out convergence character-
istics of the algorithms, the functions to be optimized need
to fulfill certain mathematical characteristics like convexity,
continuity and differentiability, which is usually presupposed
without a proof. Some of the algorithms we used are shown in
Fig. 3.1 and described in detail in some papers, e.g. [14].
The algorithms are subdivided into the large class of Mathemat-
ical Programming Methods on the one hand, and that of the
optimality criteria methods on the other. For the set-up of our
optimization procedure, we applied various methods of the first
class.
One of the op~~m~za~~on pnoeedune~ LPNLP was developed by PIERRE

and LOWE [17]. It is based on an augmen~ed LAGRANGE-6une~~on
by which equality and inequality constraints can be integrated.
The unconstraint optimization problem can then be solved by
means of a Quasi-NEWTON-Method.
There are certain disadvantages in applying the LAGRANGE-

function directly as an objective function for unconstraint
problems because the KUHN-TUCKER conditions are not sufficient.
Even if it holds VL(~*,~*,~*) = Q, V2 L* is not necessarily
positive definite. Therefore, the LAGRANGE-function is expanded
with "penalty terms":
L (x,a., e,w)
a----
with
q 2
P1 = 1: [h.(x)]
i=1 ~-
P2 = 1: [g . (x) ]
J -
2
C
a
= {j ej > o} (3.1)
J'€C a
P3 1: [g.(x)]
J -
2
Cb = {j I ej o und g. >
J - o} .
j€C b
The optimization problem is solved in a sequence of

unconstraint optimizations and correction phases.
356
a) Minimization phase
The augmented LAGRANGE-function La(~' uk, ~k, ~k) is to be

minimized. This is done by successively finding appropriate
search directions and a one-dimensional search for the minimum.
Starting with xk, an iterative method for the determination of

x k+ 1.~s ~n
. t ro d-
uce d were
h x k+1.~s t h e so l '
ut~on 0 f t h e f ree
minimization problem.
Then, a number of search directions s. is produced, and the

-~
respective one-dimensional optimization problem is solved:
L
a
(A~)
~
= Min {L
A. a
(x. + A.S.)} •
-~ ~-~
~
Thus, the i-th iteration step yields:
x'+ 1 = -~ * s .•
x. + A.~
-~ -~
La(A i ) is approached by quadratic interpolation, and Ai* and

L are determined for this. After a definite number of one-
a (A~)
~
dimensional optimizations one receives
x
k+ 1 x* with La(~*,~k,~k,~k) = Min {La(~,~k,~k,~k)}
xEIRn
The generation of search directions is based on the NEWTON-

method derived from TAYLOR's expansion of the objective
function around x.:
-~
T
L (x) RJ L (x.) + 'VL (x. )
a - a -~ a-~
For 'VL (x)

a-
o it follows
x = x. - H.- 1 'VL (x.) (3.2)

-~ =~ a -~
with HESSE's matrix ~i' This yields the search direction
-1
s. - H. 'VL (x.) (3.3)
-~ =~ a -~
Since the expenditure for calculating the inverse HESSE matrix

is too high, ~i-1 is approximated by Ei step by step. Again,
357
this is done by the formula developed by DAVIDON-FLETCHER-

POWELL (DFP) [18]. Any positive definite, symmetric matrix,
e.g. the unit matrix ~, can be taken as the starting matrix ~O.
b) Correction Phase
The minimization phase yields
(3.4)
If the solution point x~+1 is to equal the desired optimum,

the LAGRANGE-multiplie;s a~, ~~ must equal the optimal values
*
~ ,
*
~ • At the same time
-~
~
-
becomes sufficiently large due to
the penalty character, so that it must hold:
VL(~*,~*,~*) VL (x*,a*,~*,w) = 0 .
a - - - -
The objective of the correction phase is to select a~+1 and

~+ 1
~ in such a way that
L(~ ~+1,~~+1,_r.:t~+1) (~+1 ,a,

~ r.:t~ ~)
I-' :: L
a -x - -I-' ,w
-
is fulfilled. This leads to the modification rules for the
LAGRANGE-multipliers.
If equality constraints are given, it holds
a.~ ~+1 = a.~ k + 2 w k h i (~ k+1) , i=1,2, ••• ,q.

1
For ineguality constraints one receives
with j E C "
a"
~ k k+1
{
0 for ~. + 2 w2 g. (x ) ~ 0
~. ~+1 = J J -
J ~.k + 2 w2 ~ gj(~ k+1 ) others;
J
with j E Cb :
{
0 for g. (x ~+ 1 ) ~ 0
~. ~+1 J -
J ~ ~+1
2 w3 g. (x ) others.
J -
358
Ini tializin9_
~,~,~,~
Minimization p-hase
!';Ik+ 1~ Min La ( !,;, ~ Ikl, ~(kl , ~(kl)
~
Correction p-hase
Determining of the values for
~(k+1~ Q(k+1l ,~(k+1l
Flg. 3.2: Block dlag~am 6o~ LPNLP
After the determination of the new LAGRANGE-multipliers the

weighting factors are changed:
{
w. for wi wimax
k+1 lmax
w. i=1,2,3.
1 k others
YW i
As a further efficient algorithm one should mention the Method

of Sequential Linearization (SEQLI),described in detail in [14]
and applied for various practical problems.
3.2 Structural Analysis - Mathematical-Mechanical Modelling
The starting pOint of every structural optimization is to

figure out one or several "first guesses" that are as suitable
as possible. Thus, creativity on the part of the designer will
be a future demand as well. Special concentration will have to
be dedicated to the transformation of a given, real structure
into a mathematical-mechanical model including a reasonably
applied structural analysis. Thl~ 6l~~t ~tep mu~t be taken
359
veny cane6ully becau~e a ~u66~c~ently good ne~ult 06 an opt~

m~zat~on computat~on e~~ent~ally depend~ on the qual~ty 06 the
mathemat~cal-mechan~cal model.
We distinguish between four different mathematical-mechanical

models for structural analysis:
1) Ordinary Differential Equation Models
The shell of revolution serves as an example of these. The

fundamental equations of the bending theory of the shell of
revolution with symmetrical load can be combined in an ordinary
differential equation system of the first order, the so-called
state or system equations [19]:
dy dy
~ Y + ~ - + c (3.4)
dcp dcp
with
T
Y (rN cpcp , rQ cp , rMcpcp , u, w, X) state vector,
c load vector,
~, ~ matrices.
It is, however, also possible to transform (3.4) into a

differential equation of higher order.
A typical example of this is the spherical shell with a
differential equation of fourth order concerning the bending
angle X resp. the shearing force Qcp [19]:
LLX + 4]..1.4 X = 0 (3.5a)
resp. LLQ
cp
+ 4]..1.4 Q
cp o (3. 5b)
with L( ) = ( )" + cotcp ( ) I - cotcp ( MEISSNER-

opera t or, ( I~d)
dcp
decay factor .
These models, particularly the state equations (3.4) have been

used for numerous applications, e.g. [20], [21].
360
2) Difference Equation Models

By introducing various types of difference quotients differen-
tial equations can be transformed into difference equations.
If, for example, the first derivatives in (3.4) are substi-
tuted by the front difference quotient
d
dcp ~i-1
(3.6)
then all of the state variables at the point i can be expressed

by values at the point i-1. In the case of a shell we call such
an element an elongated twill. After modifying (3.4) with (3.6)
(see [19]) we finally obtain
~i = ~i ~i-1 • (3.7)
Here, the ~i are the modified state vectors and ~i the transfer
matrix. With r~gard to the condition of compatibility, whereby
the same forces and moments must be transferred and the same
deformations must exist when two elongated twills have a
common intersection point, the transfer procedure between
boundary i=O and i=n of a shell of revolution is
~n = ~l. ~i ~O =~ ~O (3.8)
This transfer procedure has been successfully used for numerous

optimization procedures, e.g. [20], [21].
3) Partial Differential Equation Models
As in 1), these equations can also be expressed in the form of

a system of lower order or as an individual equation of higher
order. A typical example is the following coupled system of
partial differential equations concerning the deformation
variables u, v, w, ~x' ~y. It was used as the structural
analysis for optimizing a rectangular sandwich panel with
anisotropic layers [23].
361
B
11
u'" + B
22
V· + (B
12
+2B
33
) (u' +v") + K11 W'"
x + K22~Y' +
- p
(3.9)
.,
B 22 v + B33 v " +(B33+B12)~' + K22Wy + K33 W; +(K 33 +K 12 ) WX -
- S22(W y +w) = 0 .
4) Algebraic (Non-Difference) Equation Models
In a great number of cases the Finite Element-Models or the

Boundary Element-Models are used for structural analysis within
the frame of optimization procedures. In the stiffness method
the whole continuum resp. the classified total structure is
divided into small (finite) elements, and the state of dis-
placements for each element is described by a polynominal
approach with free parameters. The best approximation to the
real field of displacements is achieved if the free parameters
(here: the displacements of the intersections of the single
elements) are determined in such a way that the variation of
the total energy disappears (minimal principle see Chapter 1).
The node displacements of the nodes determined by this pro-

cedure yield the field of displacements:
~(~) d - f 0 (3.10)
with ~(~) total stiffness matrix ,

d displacement vector ,
f loading vector .
After having calculated the displacements~strains and stresses
can be determined by the constitutive equations.
362
A great number of shells and frame structures were optimized

using FE-Models (see for example [21]).
3.3 Optimization Modelling - strategies
From an engineer's point of view, optimization modelling is of

essential importance. It is defined as the determination of
structural and design variables as well as the establishment
of objective and constraint functions. For some types of prob-
lems, this will call for an interdisciplinary exchange of ideas
between various fields.
For optimization modelling and for defining a strategy in the

sense of vector optimization, all terms relevant for the opti-
mization are listed and integrated into the main program via
so-called problem programs. They work as a link between opti-
mization algorithms and structural analysis. Thus, this part
includes all pnoblem ~peeinie information in contrast to 3.1
and 3.2 which are pnoblem independent within the program
system. Figure 3.1 shows the arrangement of these three parts
of the optimization procedure in a block diagram.
Finding Pareto-optimal solutions belongs to a strategy of the

optimization modelling in the following if there is a problem
with multiple objectives.
Meanwhile, a number of publications have dealt with various

methods for transforming vector optimization problems into
substitute problems (among others [8], [14], [16], [27]). As
the problem dependence of the various methods may be highly
relevant, it is one of the aims to test their efficiency and
their preference behaviour on typical structures.
1) Objeetive weight~ng is obviously one of the most relevant

substitute models of vector optimization problems. It permits
formulating the preference independent of the individual
minima; it is also guaranteed that for convex problems all
points will lie on the efficient boundary. The preference
363
function is determined here by the total sum of the single

objective functions f 1 , .•• ,fm provided with the weighting
factors w1 ' ••• ,wm:
m
L [w J' f. (x) ] (3.11)
j=1 ] -
where m
o < Wj < 1 , L w.
J
j=1
In economics this so-called bene61t-model has already been

applied for quite a long time [15], [24]. Objective weighting
presents a scalarization of the vecto~ problem.
2) The dl~tanee 6unetlon~ also lead to a scalarization of the

vector problem. If a decision maker expresses a so-called
- - - T
demand-level vector y = (Y1""'Ym) , with the objective
function value to be achieved in the best possible way corres-
ponding in structural Optimization to a set of assumed spec-
ification values or demands for the single objective functions,
the respective substitute problem is:
m
:= ( L If. (x) _ Y.l r ) 1/r (3.12)
j=1 J - J
< r < 00, x E Rn
where the variation of r aims at various interpretations of the

"distance" between the demand level y and the functional-
efficient solution. In any case, the selection of an appropri-
ate distance function is designed to achieve the components of
the vector y in the best possible way.
The following dl~tanee 6unetlon~ are most frequently applied:

m
r = 1 : p[!(~)] = r f . (x) - y. (3.13a)
j=1 J - ]
-
(f.(x) - y.) 2
r = 2 (3.13b)
] - ]
EUCLIDian metric,
r ... 00 Max I f.(x) -

] -
y.J I, (3.13c)
j=1,m
TSCHEBYSCHEFF metric.
364
It should be noticed that the distance functions can cause the

following disadvantages [16]:
- the selection of "wrong" demand levels y will lead to non-
efficient solutions,
- the selection of "correct" or "valid" demand levels y
requires the knowledge about the individual minima fj of the
m objective functions f.(x), (j=1, ••• ,m) which is not easy
J -
to achieve with non-convex problems.
3) The eon~t4a~nt o~~ented t~an~6o~mat~on (t~ade-066-method)

is a further successful method for transforming a vector
optimization problem into a scalar substitute problem. Here
one of the objective functions is minimized and the others are
restricted upwards:
p[i(~)] = f1 (~) (3.14)

with
f. (x) ~
J - -
Y'J j=1, ••• ,m
Thus, f1 is called ma~n objeet~ve, and f 2 , ••• ,f m are called

~eeonda~y objeet~ve~. The given problem can be interpreted in
such a way that when minimizing f 1 , the other components have

to achieve at least the values Yml. The main objective func-
tion f1 should be formulated by an objective function for which
an a-priori estimation by an upper limit Y1 is practically not
available.
4) Besides the preference functions described above, the

m~n-max-6o~mu~at~on plays a very important role in solving
substitute problems. It is based on the minimization of rela-
tive deviations of the single objective functions from the
respective individual minimum [10], [25].
If the extremes fj are established separately for each objec-

tive function (criterion), the desired solution is the variable
~ with the smallest value of the relative deviations of all
objective functions. The scalar substitute problem according

to the m~n-max-6o~mu~at~on can be defined as follows:
365
p[!(~)] := Max [z.(x)] (3.15a)

J -
j=1,m
where
fj (~) - fj
z. (x) f. > 0 (3.15b)
J - J
fj
j=1, ••• ,m.
For convex problems, the solution x of (3.15) is PARETO-

optimal resp. functional-efficient. It is also called min-
max-optimum as it yields the "best" possible compromise
solution under observance of all objective functions with equal
priority. In [4], a reasonable modification of (2.19) is given
for practical computations. It consists of the minimization of
a new variable ~ (comparable to slack variables) while simul-
taneously considering additional constraints:
z. (x) - 13 ~- 0 x E IR n + 1 , (3.16)
J -
j=1, •.• ,m
The equation (3.16) is profitable especially for non-linear

optimization problems effectively applying inequality con-
straints in the optimization algorithms (e.g. methods of
sequential linearization) [26].
The min-max-formulation described above yields the compromise

solution ~ considering all objective functions with equal
priority. But if the single objectives have to meet a special
order or if the complete functional-efficient solution set X*
is of great importance for the decision maker, the min-max-
formulations can be modified or extended as follows:
a) Min-Max-Formulation with Objective Weighting
The introduction of dimensionless weighting factors w. > 0

J
transforms the substitute problem (3.15) into:
Max [w. z. (x) ] (3.17)

j=1 fm J J-
where z. (x) denotes the same relative deviation as in (3.15).

J -
The weighting factors describe the priority of the single
objective functions. Thus, it is possible to select definite
366
compromise solutions from random fields of functional-

efficient sets. Moreover, the variation of Wj allows
the complete solution set to be established ..
An appropriate modified form is also valid for (3.16):
P [_f (_x)] = f3 A W. Z • (x) - f3 < 0 x E IRn + 1 , (3.18)

J J - =
j=1, ••• ,m
Fig. 3.3a shows the geometric interpretation of (3.18) for the

two-dimensional case. It is obvious that the dependence of the
ratio w1 /w 2 of the two weighting factors results in different
compromise solutions describing th8 whole functional-efficient
boundary.
b) Min-max-formulations by selecting a demand-level vector
If the definition of the relative deviations in (3.15b) is not

based on the individual minima fj but on the given comp~nents
Yj of the demand-level vector with the characteristics Yj =f j,
we get analogous substitute problems to (3.17) and (3.18).
However, this problem formulation does not guarantee that in
the solution pOint ~ att inequality constraints become active,
i.e. that they can be regarded as equality constraints. Even
if all inequality constraints become active, the solution
vector x lies on the intersection of the line in space with
the functional-efficient solution set X* . The difference to
w, 1
Wi=:r
Y, I, I, I, I,
F~g. 3.3: Min-Max-Op~um fio~ two object{ve fiunct{on6 und~ cOn6~d~on

ofi ~fi6~en;t wugh:ttng fiaao~ ~e..R.a.;UOn6 and demand tevu veaOM y.
367
the formulation mentioned before is shown in Fig. 3.3b. If the

line going through the point of the vector yand being defined
by the ratio w1 /w 2 intersects the functional-efficient
boundary, the intersection point is also the compromise sol-
ution. If there is no intersection point, the point for f1 or
f2 is the solution referring to the ratio between w1 /w 2 • The
special selection of a demand-level vector i = Q with a
simultaneous renunciation of the division through Yj within
the relative deviation z. (x) yields a further modification of
J -
the min-max-formulation
Max [w. f J. (~) ] (3.19)

j=1,m J
This min-max-formulation is applied very frequently in
practice [4], [26].
4. APPLICATION ON FINDING THE OPTIMAL LAYOUT OF A SHELL

STRUCTURE
In numerous examples taken from the fields of antenna and
telescope construction, aircraft and space technology,
constructions made of composite material, crane construction
etc., the optimization procedure for the layouts of component
parts regarding multiple objectives resp. criteria described
here was tested successfully. Publications on this topic can
be found in [6], further references are given there. As an
example, the procedure for finding the optimal shape design
of a component in conveyer technology, a so-called conveyer
belt drum, is dealt with in the following.
4.1 Problem Formulation

In order to describe curved supporting structures (shells),
the GAUSSian surface parameters Ea of the middle surface and
the coordinate ~ perpendicular to the middle surface have to
be selected [22]. The shape function B describing the shape of
of the shell then consists of a part of the middle surface
E(Ea ) and a part 21 t(E a ) ~3(Ea) .ln normal direction (Fig. 4.1):
368
(4. 1 )
where t(E a ) describes the wall thickness distribution.
The objective is to find the optimal shape function
a
R* E [" := {~(E ) : IR 2 -+
3
IR } ,
for which the objective function F : [" -+ IR takes a minimum,

resp. it is to formulate:
Min F(~) F (~*) (4.2a)

RE[,
_---tl~_~2
middle
surface
Fig. 4. 1: Ve~e~iption 06 the ~hape 06 a ~hett
With ~ := {~: 1R3 -+ IR, ~ arbitrary real-valued function}

and Pk' q£ E ~, the feasible domain of variation is determined
by the constraint operators Hk , G£ : [" -+ ~ with
Hk~ Pk k=1 , ••• ,m H (4.2b)
G£~ q£ £= 1 , •.. ,mG
and the bounds E [':

~1 ' ~2
~1 < -R < -u
R (4.2c)
369
Here, the desired shape functions approximated by means B are

of approach functions R with free parameters x. If x E Rn
and r := {B
(Ea,~) : R2 x Rn + R 3 } describes ~he sp~ce of the
approach functions, the shape functions R E r can be approxi-
mated by the function RE r [14]:
(4.3)
With a d~~ect method a set of shape functions with free

parameters is put into the problem formulation in order to
transform it into a parameter optimization problem. This
parameter optimization problem can be solved then by means of
corresponding optimization algorithms [26]. Fig. 4.2 shows
how a shape optimization problem with multiple objectives
is transformed into a scalar substitute problem to find
out an optimal compromise solution x.
Shape Optimization Probtem

"Min" E{B) B E 1r2
optimal
Hk B = Pk · k =1. ...• mH => functional sp ace rt
Gl B S ql · I =1•. .• mG
B Bu
I
~ ~
I'll
I Approach funcl,on
I B{1;·.~)
Vector Optimization Problem

"Min" .!.{~) ~ E IR n.
optimal
hk (~) = 0 · k =1 .. .. ,m h => solution s et X*
9 l (~) ~ 0 • I =1. ...• mg
~l S P :s. ~u
I IOphm,,"',on Strategy
L pI .!.{~))
I
~
Scalar Optimization Problem
Min pl.!.{~)J ~ E IR"
optimal
hk{~) = 0 . k =1 .. ... mh => compromise solution &.
9l (~) sO. I =1 .. ..• mg
!5. l ~ ~ :s ~u
F~a. 4.2: T~an~60~mat~on 06 a ~hape opt~m~zat~on

p~obtem w~th mutt~pte object~ve~ ~nto a
~cata~ ~ub~t~tute p~obtem
370
4.2 Optimal Shape Design of a Conveyer Belt Drum

Efficient conveyers are necessary for extensive soil shifting
operations in open mining. Here, large rubber-belt conveyers
could prove to work successfully. According to Fig. 4.3.a, the
conveyer belt drum, the track supporting roller, and the belt
are essential components of the belt conveyers. A conveyer
belt drum consists of the cylindrical drum shell (1) and the
bottom (2) (Fig. 4.3.b). Drum bottom and shaft (4) are connected
by the clamping ring (3).
The development in this field is characterized by permanently

increasing demands on conveying capacity, conveying track, and
operation safety which leads to enlarged distances between the
axes and the conveyer belt width.
(1) Cylincrical drum shell

(2) Drum bottom
(3) Clamping ring

(4) Shaft
(~
F~g. 4.3: Ske~eh 06 ~he bel~ eonveye~

a) eomple~e ~y~~em
b) eomveye~ be~ eyl~nde~
As the objectives can be sufficiently realized by essentially

larger tension forces, the stresses in the belt drums are
inevitably enlarged too. It was attempted to reduce these
stresses by extending the wall thickness and by implementing
ribs. But these measures often led to an extreme increase in
weight, so that damages and failures could not be avoided.
371
The optimal shape design of a conveyer belt drum was treated

with the direct method of shape optimization. On the basis of
a given middle surface contour r(~), the optimal wall thickness
distribution t(~) had to be determined in a way that the
objectives of optimization "minimal weight W" and "minimal
reference stress "were fulfilled in the best possible
0
rmax
way. Constraints are given as stress indications. The desired
function is approximated by a partially linear approach
t(~)
T
function with x (t 1 ,t 2 ,t 3 ,t 4 ) (Fig. 4 . 4) :
t1 - t2
,
t(~,~) = t1 - (~-~1 )
- ~2
~1 ~ ~ < ~2
~1
t3 - t4
t(~,~) t3 - (~-~ 1 )
- ~4
~2 ~ ~ ~ <!l3 (4.4)
~3
n;
t(~,~) t4 < <
~3 ~
"2
p(x.~)
T= 650 kN
r
... p(x.~) = Po + P, cos ~ +
1----- 1000 ----~--_L400-l +p2cos2~
Fig. 4.4: Conveye~ belt d~um unde~ d~um 6o~ee4 a4

loadJ.J
The re-transformation into a scalar substitute problem is

carried out by a min-max-strategy. A sequential linearization
method is used as optimization algorithm. This chapter also
presents the transfer matrices method applied for the
structural analysis (see [6]) .
372
The functional-efficient solution set (Fig. 4.5) illustrates

the influence of different preference concepts on the shape
of the optimized conveyer belt drum. As the optimization
computations showed, the min-max-formulation has turned out to
be an efficient optimization strategy. It is optimally adapted
to the sequential linear method by its linear substitute
function and the additional inequality constraints.
60+---
50 +----\l----j--
40 +----+,~-t-
,----,----,
I I
20
10
G
o ,----l---4---I---+--9-j.0--1~00::-;(kNl
Fig. 4.5: Funetional-ennieient ~olution~ on the

eonveye~ belt d~um
The example of the conveyer belt drum is also used for testing
various optimization algorithms combined with strategies. It
is generally not possible to give an optimization procedure
and an optimization strategy in such a way that they definitely
lead to a good result in all cases. For this reason some
combinations for particular cases of application, among others
the conveyer belt drum, were evaluated via quality criteria
in numerous investigations [6]. The criteria consider four
different aspects:
optimality,
velocity,
quality,
security.
373
The single aspects receive a point rating reaching from P=1

(best valuation) to p=o (worst valuation). In analogy to the
objective weighting they are finally additively combined to
the quality criterion Q with the help of weighting factors wi:
Q = L. 1w.P.
.1.
i=1, ••• ,4 L w.1. = 1 • (4.5)
1.
For determining the quality Q of an optimization procedure,

the following weighting factors are selected here: one puts up
w1=O.3 for optimality, w2 =O.2 for velocity, w3 =O.4 for quality
and w4 =O.1 for security which is regarded as less important.
Thus, the quality criterion which cali maximally be reached is
standardized to Q=1.
COT MM OW OF COT MM OW OF
a) Objee~~ve~: - We~gh~ b) Obj~e~~ve~: - We~gh~

- Ve6lee~~oYL - S~Jte~~e~
f~g. 4.6: Valua~~oYL 06 Op~~m~za~~oYL PJtoeeduJte~

COYLveyeJt Bel~ VJtum
Fig. 4.6a and 4.6b illustrate the quality of respective

combinations for the optimization procedure. They give the
efficiency of the procedure for an optimization and clearly
show that the efficiency of an optimization procedure is very
much dependent on the problem. Even if only the objective
functions for the component part differ, the figures show
374
marked differences. Considering the choice of the weighting

factors for the single quality criteria case a) shows only
slight differences of the total quality. The optimization
model but also the selection of the single criteria and the
weighting factors decisively influence the total criterion.
In order to receive generally valid statements about the
selection of the optimization procedure, numerous investiga-
tions are still necessary. In every case of application it is
necessary to adapt the valuation criteria to one's own demands
and to the demands of the problem being calculated.
5. CONCLUSION
The software package SAPOP was developed as an optimization
procedure on the basis of the three columns "Structural
Analysis", "Optimization Algorithm", and "Optimization Model-
ling". SAPOP connects problem-independent Optimization and
Structural Analysis Methods by a so-called problem program.
An essential part of the problem program is the transformation
procedure in order to transform a Vector-Optimization-Problem
into a Substitute Scalar Problem (SSP). The application and
tests of the efficiency of the developed optimization methods
on a special task of Structural Mechanics, e.g. shape opti-
mization of a shell structure (conveyer belt drum) is shown.
Finally, several combinations of Optimization Strategies and
Optimization Algorithms have been tested by establishing
valuation criteria exemplified by a conveyor belt drum. This
is the first step towards establishing a knowledge based
system for an expert system.
APPRECIATION
The author wishes to express his appreciation and thanks to
Miss Baerbel Huette who typed this chapter with great patience.
375
REFERENCES
[1] SZABO, I.: Hohere Technische Mechanik, 5. Auflage.

Berlin, Gottingen, Heidelberg: Springer 1972.
[2] SZABO, I.: Geschichte der mechanischen Prinzipien.

Basel, Boston, stuttgart: Birkhauser 1979.
[3] CLAUSEN, T.: Uber die Form Architektonischer Saulen.

Melanges Mathematiques et Astronomiques I (1849-1853),
279-294.
[4] BENDS¢E, M.P.; OLHOFF, N.; TAYLOR, J.E.: A Variational

Formulation for Multicriteria Structural Optimization.
Danish Center for Applied Mathematics and Mechanics.
Technical University of Denmark. Report n. 258, 1983.
[5] PARETO, V.: Manual of Political Economy. Ubersetzung von

A.S. Schwier der franzosischen Ausgabe von 1927. The
Mac Millan Press Ltd., London-Basingslohe, 1971.
[6] STADLER, W.: Preference Optimality and Application of

PARETO-Optimality. In: MARZOLLO/LEITMANN (Editors):
Multicriterion Decision Making. CISM Courses and Lectures;
Berlin, Heidelberg, New York: Springer, 1975.
[7] STADLER, W.: Multicriteria Optimization in Mechanics (A

Survey). Appl. Mech. Review, 37, 3 (1984), 227-286.
[8] BAIER, H.: Mathematische Programmierung zur Optimierung

von Tragwerken insbesondere bei mehrfachen Zielen.
Dissertation, TH Darmstadt, 1978.
[9] KOSKI, J.: Truss Optimization with Vector Criterion.

Tampere University of Technology Publication, Tampere 1979.
[10] OSYCZKA, A.: Multicriterion Optimization in Engineering.

New York, Chichester, Brisbane, Toronto: John Wiley, 1984.
[11] RADFORD, A.D.; GERO, J.S.; ROSENMAN, M.A.; BALACHANDRAN,

M.: Pareto Optimization as a Computer-Aided Design Tool.
In: GERO, J.S. (Editor): Optimization in computer-aided
design. North-Holland Amsterdam, New York, Oxford:
Elsevier Science Publishing Company 1984.
[12] DUCKSTEIN, L.: Multiobjective Optimization in Structural

Design: The Model Choice Problem. In: ATREK, GALLAGUER,
RAGSDELL, ZIENKIEWICZ (Editors): New Directions in
Optimum Structural Design. Chichester, New York, Brisbane,
Toronto, Singapore: J. Wiley & Sons, 1984.
[13] KOSKI, J.: Multicriterion Optimization in Structural

Design. In: ATREK, GALLAGUER, RAGSDELL, ZIENKIEWICZ
(Editors): New Directions in Optimum Structural Design:
Chichester, New York, Brisbane, Toronto, Singapore.
J. Wiley & Sons, 1984.
376
[14] ESCHENAUER, H.: Numerical and Experimental Investigations

on Structural Optimization of Engineering Designs.
DFG-Report of the Research Laboratory for Applied Struc-
tural Optimization, University of Siegen, May 1986.
[15] FANDEL, G.: optimale Entscheidung bei mehrfacher Ziel-

setzung. Berlin, Heidelberg, New York: Springer, Lecture
Notes in Economics and Mathematical Systems 76, 1972.
[16] SATTLER, H.-J.: Ersatzprobleme fUr Vektoroptimierungs-

aufgaben und ihre Anwendung in der Strukturmechanik.
Dissertation, Universitat-GH Siegen, 1982.
[17] PIERRE, D.A.; LOWE, M.J.: Mathematical Programming via

Augmented Lagrangian. London, Amsterdam, Ontario, Sydney,
Tokyo: Addison Wesley, 1975.
[18] FLETCHER, R.; POWELL, M.J.D.: A Rapidly Convergent Descent

Method for Minimization. Computer Journal, Vol. 7 (1963)
163-168.
[19] SCHNELL, W.; ESCHENAUER, H.: Elastizitatstheorie II,

Schalen •. Mannheim, Wien, ZUrich: B.I.-Wissenschaftsver-
lag, 1984.
[20] ESCHENAUER, H.; SATTLER, H.-Fr.: Optimierung von Schalen-

bauteilen fUr mehrere Ziele. ZAMM 63, T404-T406, (1983).
[21] ESCHENAUER, H.; KNEPPE, G.: Application of Optimization

Procedures on the Design of Various Shell Structures.
In: System Modelling and Optimization, ed. by THOFT-
CHRISTENSEN. Berlin, Heidelberg, New York, Tokyo:
Springer, 1984.
[22] ESCHENAUER, H.; SCHNELL, W.: Elastizitatstheorie I,

Grundlagen, Scheiben und Platten. Mannheim, Wien, ZUrich:
B.I.-Wissenschaftsverlag, 2. ed., 1986.
[23] ESCHENAUER, H.; FUCHS, W.: Optimum Design einer punkt-

f5rmig belasteten Sandwichplatte mit orthotropen CFK-
Deckschichten. ZAMM 65, (1985), T305-T306.
[24] DUCK, W.: Optimierung unter mehreren Zielen. Braun-

schweig: Vieweg, 1979.
[25] ESCHENAUER, H.; KNEPPE, G.: Min-Max-Formulierungen als

Strategie in der Gestaltsoptimierung. ZAMM 66, (1986),
T344-T345.
[26] KNEPPE, G.: Direkte L5sungsstrategien zur Gestaltsoptimie-

rung von Flachentragwerken. Dissertation, Universitat-GH
Siegen, 1985.
[27] JAHN, J.: Scalarization in Multiobjective Optimization.

TH Darmstadt, FB Mathematik, Preprint-Nr. 847, 1984.
V, RELATED TOPICS
THE EFFICIENCY OF A METHOD OF FEASIBLE DIRECTIONS
FOR SOLVING VARIATIONAL INEQUALITIES
H. Kirsten and R. Tichatschke
Sektion Mathematik/Physik, Padagogische Hochschule Halle
Postfach 763, DDR-4050 Halle
An algorithm of the type of feasible directions is

described efficiently solving extremal problems which
arise by discretization of variational inequalities.
Using the maximum principle, the convergence of the
method and the independence of the iteration number from
the discretization parameter is shown. A comparison with
other methods confirms the efficiency of the algorithm.
O. Introduction
In this paper the numerical solution of elliptic variational inequali-

ties and the corresponding e>:tremal problems are considered. There are
many problems in mathematical physics as in the theory of electricity,
of elasto-plastic'media and in the hydrodynamics which can be formu-
lated as variational inequalities (cf.[5], [6], [13]). The solution of
the underlying extremal problem in infinite dimensional function
spaces as well as the approximation of this problem was already stu-
died in detail (cf.[2],[9],[10]).
The numerical treatment of such specific high-dimensional nonlinear
programming problems with general numerical methods of nonlinear
(convex) programming is not efficient in every case. Therefore, it is
necessary to develope solution methods for these special classes of
programming problems (cf.[1],[7],[11],[12],[19]). But it seems to be
difficult to make efficiency estimations for these methods with
respect to the discretization parameters.
In a previous paper [16J a regularized method of feasible directions
was described for an e)ttremal problem which arises by the performed
obstacle problem as a special type of variational inequality. Certain
properties of monotonicity of the iterates and the convergence of this
method can be proved by using explicitly the maximum principle which
holds for the underlying boundary value problem.
In the present paper it is shown that the iteration number for solving
stationary variational inequalities by the help of this type of
regularized method of feasible directions is bounded above for
arbitrary refinement of discretization and this boundary is
independent of the discretization parameter.
380
1. The Problem under Consideration
The following elliptic variational inequality is considered:

ue:K I a(u,v - u) 2: (f,v - u) 'f vEK (1. 1)
with
a(v,w) ...
r
J (L
m QV ()w
k ( x ) - - + q(x)v(x)w(x) }dx ,
i=l ()XjaXi
B
(f,v) .. ) f(x)v(x)dx ,
B
1
K" (VEW 2 (B) : v(x) 2: f(x) XE B, v i r " g}.
m
The set Be R, m 2: 1, denotes a bounded domain with piecewise smooth
boundary r .. aB and the functions involved are assumed to be suffi-
ciently smooth to have a corresponding regular solution u of problem
(1.1) (cf.C13]). In this case the solution of the variational inequa-
lity (1.1) is equivalent to the solution of the extremal problem
ueK J (u) ... min ( J (v) I v E K )- (1. 2)
with
1
J(v) .. -a(v,v) - (f,v).
2
1
If u E W (B) is a solution of (1.1> (resp. of (1.2», then the domain B
2
decomposes into the following two s\.\bsets:
.. = f(x)
..
B ( X E B u(x) } (Coincidence set) ,
B
+
... { x E B U ()I ) ). t(x) } B \ B
and the set r0 .. B 1\

-+
B is called the free boundary.
Since in B the inequality f(x) S. 0 holds (cf.C18J>, i.e. the power f

works at the direction of the obstacle 'f, it is assumed that this
inequality is satisfied in the whole domain B.
If the solution u of problem (1.1) is sufficiently smooth, then the

equation
+
LCu(x)] .. f(x) xeB
is available, where L denotes the elliptic differential operator
LCu(x)] .. - r:m -a ou(x)

(k(x)---) + q(x)u(x)
i=l OXj OXj
For this elliptic differential operator the maximum principle

381
2
(cf.[17J) is valid. i.e. let u e. C (B) f'I C(B) be a function satisfying
the following two inequalities.
L[u(x)J~~O >: E B and x E r ,

then
u(x) ~~ a
The general way for solving infinite extremal problems (1.2), however,
is their approximation by finite programming problems. In this con-
crete case the approximation is done by the method of finite elements,
1
where the appro>:imation of the space W (B) is convenient by using
2
linear elements (cf.[9J, [10J).
Let B a Finite-Element-Model of B which arises from a regular trian-
h
gulation of the domain B, then the following notations are used:
i
x - vertices of the triangulation B , N - number of vertices, w (x) -
h I
basic functions of the Finite-Element-Space V - span(w , ••• ,w ) and h
h 1 N
denotes the discretization parameter. Since the finite dimensional
N
function space V . is isomorph with the Euclidean space R the follo-
h
wing identification is possiblel
T 1 N T N
V t V v := (v , •••• v ) .. (v (x ), ••• ,v (x » E R •
h h 1 N h h
In this way the approximation of the extremal problem (1.2) leads to a

ccmve>: quadratic optimization problem:
LI E K ,J (u) .. min { J (v) : v E K } • (1.3)
h N N
with
1
J(v) .. -(v,Av) - (f,v)
2
and
N
'" { vER
K
N
v
i
2: f.
1
i E. I, v g.
i '" 1
i E Ir } , (1. 4)
respectively
Ir } .
N N
K '" { vER L v w (x) .::: f(>: ) )( E B v '" g.1 ~ (1. 5)
N i-1 i h'
Analogously the following notations are used:
f(>:
i
)
'" f 1 .. 9 (>{
i
) . g. ,
1
(f ,w.
1
)
'" f 1. , f . (f , •••• f )
1 N
T
,
a(w . ,w.
1 J
) . a
i j
A '" {a }
ij i ,j"'l
N
. {i ,2, •••• N} , Ir '" {i E I : >:

i
E r } .
The assumptic:m f(>() :,0 for all )( E B leads now t.O the discrete version
f (0 i:i, ••• ,N (1.6)
i -
382
In dependence on the chosen set K by (1.4) (resp. by (1.5» the

N
problem (1.3) is a finite or a semi-infinite optimization problem. For
the construction of an approximate solution of (1.2) the finite ap-
proach (1.3), (1.4) is already sufficent. The solution u of (1.3),
h
(1.4) converges to the solution u of (1.2) with the power O(h)
(cf.[10J).
The semi-infinite problem (1.3),(1.5) is in the following used to
investigate the efficiency of the hereafter described method, particu-
larly by using imbedding techniques the independence of the number of
iterations from the discretization can be proved.
The discrete version of the maximum principle which arises by the
Finite-Element-Model B is the following (cf. [21], [23]) :
h
If Au ~ Av and u
i
S. v
i
i £. Ir , then u ~ v .
The solution u of (1.3) satisfies the necessary and sufficient opti-
h
mality conditions which have the following forml
( Au) 2: fiE I \ I r
h i i
(1. 7)
«Au ) f ) (u 'f ) =0 i £ 1\ Ir
h i i ih i
2. A Method of Feasible Directions
For the computational solution of the convex programming problem (1.3)

with respect to linear restrictions an algorithm of the type of feasi-
ble directions is used (cf.[14J,[15J,[16J), i.e. at the iteration point
k k
v E K a feasible direction r has to be calculated and at this
N k~f
direction the ne>,t iteration point v has to be computed.
In many applications of the method of feasible directions the direc-
k
tion search problem is linearized at the iteration point v and there-
fore the locally best descent direction is used (cf.[3J,[SJ,[22J). In
these cases it can be proved that the method converges sublinearly.
Here is described a method of the type of feasible directions which is
suitable not only for the solution of the finite problem (1.3), (1.4)
but also for the solution of the semi-infinite problem (1.3), (1.5).
Both methods differ only from the computation of the step lengths,
since in the semi-infinite case infinitely many restrictions have to
be talcen into consideration.
The direction search problems are the same in both methods, therefore
383
both methods have the properties which are in the following described.
To ensure "good" convergence order of this method a direction search
problem is considered which determines the global best descent direc-
tion. Therefore, the solution of the direction search program at the
k
iteration point v is determined by the direction
k "'k k
r := v v
"'Ie
where v is the solution of the following linear equation system:
"'k k
(Av ) f i ; It.(v )
i i
"'Ie Ie k
v == v i 4i I t. (v )
i
"'k
v g1 i e Ir
i
I: I:
with I£(v ) = { i £ I \ Ir : v
i
'f
i
;5;, E. }.
Remark 2.1.: System (2.1) is a linear equation system of the dimension

I:
N - card(Ir",V Ie.(v » which arises by fixing those components of
the vel(tor v, where boundary conditions gi are given or where
a given £ - precision with the obstacle is attained.
Theorem 2.1.: (cf.t16])
Ie "'I: Ie I: "'k
The direction r := v - v at the point v £ K, where v is
N
generated by (2.1), is feasible and is the global best descent
at the linear manifold
'"
K
N
(v
I:
) == { v£R
N
v
i
= v
k
i
i E It. (v
k
) , v
i
.. gi i t Ir } .
k
The direction r is the local best descent direction of the regula-
rized functional
1 1 -1
J A-1 (v) = -(v,Av> -1 - (f,v>A-1 = -(v,v> - <A f,v>
2 A 2
k
on the linear manifold K (v). Because of the positive definiteness of
N
the matrix A the norms <.,.) and <.,.> _1 are spectral equivalent.
A
Now, in order to compute the iteration point
1:+1 I: I: k
v := v + t r
384
k
the step-size t is determined as maximal as possible, i.e.
k+1 k k k k
J(v ) .. min (J(v + tr ) t 2: 0, v + tr E K } (2.2)
N
Theorem 2.2.: (cf.t16])
For solving the problem (1.3),(1.4) with the direction search

k
(2.1) the step-size t is given by
k k k k
t .. min {1, min {(f - v )/r r < 0 }} (2.3)
k i i i i
i;I£(v)
Remark 2.2.: In order to solve problem (1.3),(1.5) the step-size is

determined analogously:
k k k k
t .. min ( i , min k{(f(x) - v (x»/r (x) r (x) < 0 }} (2.4)
i;BE.(v )
where
r
k
(x) .. I: r w (x) ...
N
i=l
k
i i
1::
i=l
N "'k
(v
i
- v
k
i
)w (x)
i
BE(v
k
) .. h
k
(x f; B : v (x) - f(x) < f. } .
Because of the construction of the algorithm the following properties
k
are satisfied for the iteration points v •
Lemllla 2.1.1 (cf.t14])
k
For the sequence (v ) the following relations are ensured:
k+1 k
(1) J(v ) < J(v) "fk,
k+1 k
IE(v ) ::> I E (v ) 'f k,
k+1 k k+1 k+i

<iii>If1f.(v )=IE(v),thenv =v "fi2:2,
k k+1
<i v) If I (v ) ... I (u ), then v .. u
o 0 h h
385
k
Because of the monotonous increase of the £- active index set I£(v )
the convergence of this method of feasible directions is not guaran-
teed for any feasible starting point. By using the discrete maximum
principle it is possible to prove some properties of monotonicity of
k "'k
the sequences (v) and (v ) which guarantee the convergence of this
algorithm.
o
Let the starti ng point v ~ K be chosen as solution of the linear
N
equation system
0
(Av ) 0 i E I \ Ir
i
-
(2.5)
0
v g. i " Ir
i 1
with
'"
g. 2: g. £ Ir
1 1
Because of the fact that the matrix A is the discrete analogon of the
elliptical differ.ential operator L problem (2.5) describes a discre-
tized homogenuous differential equation with boundary conditions of
o
first kind. Therefore, v always exists.
Theorem 2.3.1
o
If the starting point v is chosen by (2.5), the following monoto-
nic properties are satisfied.
k k-l
(1) (Av ) 2: f i E I \ ( Ir v I £ (v », V k ~ 1,
i i
(i 1> v
k
2: 1.1
h
¥ k ~ 0 .
Proof. The proof is given with the help of complete induction with
respect to the iteration steps. Because of the linearity of the opera-
o
tor A and because of the assumptions on the starting point v the
following estimations are true:
(1.1
h i
) - V
0
.. g1 v
0
i g. - g.
1 1
.. 0
i " Ir ,
(1.1
h i
) - V
0
i
.. f
i
- v
0
i
i f
i
f
i
.. 0 i E I
0
(1.1
h
)
(A(u
h
- v
0
»
i
(Au
h i
) - (Av
0
)
i
(Au )
h i
.. f
i
i 0
i , I \ (I r v I (1.1 ».
o h
386
Hence, the discrete maximum prinziple ensures
o
(u) - (v) i 0 i f. I \ (I r v I (u »
h i i o h
and
o o
u - v i 0 i.e. u i v
h h
Now, for the first iterate the following equality is obtained:
1 o
(Av) '" (A (t (v
0 "'0
- v ) + v »
0 0 0 "'0 0
.. t (Av ) + (1-t ) (Av ) ...
i i i i
(2.6)
o "'0 0 0
... t (Av) ... t f i .If i e I \ (I r v I e. (v ».
o
Consider (1.6) and t f. [O,lJ then it follows
1 o
(Av) ! f i ( I \ (I r v I E. (v ».
i i
1
Setting v .. u - v the following inequality is true
h
o
v .. (u ) - v ... gi - (g + t (g - g » ..
i h i i i i i
o
(l-t ) (g - g ) i 0 for i f. I r
i i
o 0
v (u) - v ... (u) - v i 0 for i e I < (v ).
i hi i hi i ..
o
For i E 1\ (I r v If(v » follows from (1.7) and (2.6)
1 o 0
(Av) ... (Au) - (Av) .. f - t f = (1-t )f So 0 •
i h i i i i
Therefore, by using the discrete maximum principle

o 1
vi So 0 E 1 \ (I r v
for i IE(v », hence v S. 0, i.e. u h So v
Now, let the statement of the theorem be true for step \;:-1:
\;:-1 \;:-2 \;:-1

(Av ) i ! f i ¥ i E I \ (I r U I E. (v » and v ! uh •
Using the induction assumption one can conclude
\;: \;:-1 "'\;:-1 \;:-1 \;:-1 \;:-1 "'k-1
( Av) .. (A (t (v - v ) + v ) ) .. t ( Av +
i i i
k-l \;:-1 \;:-1 k-1 \;:-1
+ (1-t ) (Av "" t f + (1-t ) (Av )!
i i
\;:-1 k-1
! t f + ( 1-t )f f
i i '" i
k'-2
for i E I \ (I r v Ie.(v ». (2.7)
387
k-l k-2
Lemma 2.1. provides I£(v ) ~ I£(v ) and therefore
k k-1
(Av ) l f for i E I \ (I r lJ I t (v »
i
k
Now, consider v LI
h
- v the relations
v S. 0 i E Ir and v ~~ 0
i i
are satisfied again.
With (1.7) and (2.7) it follows
k k
(Av) (Au) - (Av ) f - (Av ) s.f ·-f =0
h i i i i i
k-1
for (I \ (I r v I € (v ».
Again, the discrete maximum principle leads to

k-l k
v .: 0
i -
i E I \ Or U Ie(v ». Hence, v S. 0 such that LI
h
S. v is
obtained. This completes the proof.
Remark 2.3.: The proof of Theorem 2.3. implies
k k-1 k-1
(Av) S. (Av ) ¥ i (I \ (I r lJ I f. (v ) ), 'rf k
k k-1
This property leads to the inequality v ~ v -'rf k.
Furthermore, the following relation with respect to the total
step length holds:
k k k
(Av) = t (v ) f "'I i £ I \ (I r U Ie (v », .!f k ,
I( k+l
this leads to t(v ) S. t(v ) "'I k.
k
Therefore, the total step length t(v ) can be used as a measure
for the convergence of the method.
(:I
Remark 2.4.: If v is chosen as in (2.5) it can be shown in the same
"'k
manner that the sequence (v ) has the following property:
"'k "'k+1 "'k

v ~~v ¥I( and v S. LI + £e
h
T
with e - (1,1, ••• ,1) •
Because of the finiteness of the index set I and because of Lemma

2.1. (ii> there exists an index k for which
o
388
UsiAg the arguments of Theorem 2.3. and Remark 2.4. the following
statement about the convergence of the algorithm can be proved:
Theorem 2.4.1
o
Let the starting point v be chosen as in (2.5). Then there
exists an index k with
o
and a i u u i Se ,
h
i.e. the algorithm co~verges after a finite number of steps to the
E - optimal solution u of problem (1.3).
Remark 2.5.: A convenient stopping criteria of the algorithm is the

relation
k
t L 1 - £ (for a given £),
s s
k
since the condition t - 1 is equivalent to
ko +1 "'ko
v '" v and
Now, the complete algorithms of the regularized method of feasible

di recti ons worles as follows:
Algorithm AI (for the computation of problem (1.3) ,(1.4»
o o
91: Choose v ~ K (Av) '" a
N i
put k :'" 0, choose £ > 0, £ > O.

s
k "'k k
921 Determine r := v - v with
"'k k
(Av ) f i ; I E (v ) ,
i i
"'k k Ie
v '" v i E I£(v ) ,
i i
"'k
v g. i ~ I r
i 1
389
931 Determine
t
I:
.- min { 1 , min
i , I t (v
{(op
i
- v
I:
i
)/r
I:
i
r
I:
i
<a )} .
941 Put v
1:+1
.= v
I:
+ t r
I: I:
, I: 1= k+1.
I: I:
951 If t ~ 1 - f stop 1 v i s f - opti mal, otherwi se go to 92.
51
Algorithm B: (for the computation of problem (1.3),(1.5»

The steps 91,92,84 and 85 are the same as in Algorithm A, only in
step 83 formula (2.3) is to be replaced by formula (2.4).
Because of the fact that in the step-size problem (2.4) the restric-
tion set of problem (2.3) is included, the step length by using Algo-
rithm B can not be larger then by using Algorithm A assuming the
k
direction r is the same in both cases. Therefore, the following
connection between the total step length of both algorithms can
be provedl
Lemllla 2.2.: (cf.[14])
I: I:
(v) and (v) denote the sequences which are generated by Algo-
A B
rithm A and B respectively.
I: k k+l 1:+1
If t(v ) ~ t(v ) then t(v ) ~ t(v )
A B A B
o 0
Because of the fact that both algorithms start with t(v ) .. t(v ) .. 0
A B
k 10
and terminates with t(v 0) .. t(v ) = 1, the faster convergence of
A B
Algorithm A in face of Algorithm B can be concluded from Lemma 2.2 ••
3. The Efficiency of the Method
The monotonic increase of the set of E - active restrictions at every

iteration step leads immediately to an upper bound of the iteration
number. This bound is given by the number of discretization points
lying in the £ - coincidence set of problem (1.1> or (1.2) respective-
-m
ly, and increases with power O(h ), if the discretization of B will
be refined. Therefore, it. seems to be possible to have a strong in-
rease of the iteration number by reason of refinement of the discre-
390
tization and the method would be impracticable.

However, it can be shown that the iteration number does not depend on
the discretization constant h and the method works efficiently. There-
fore, the behaviour of the iterates at different discretizations is
developed. From this point of view it can be shown that independently
of the discretization always a certain minimal increase of the £-
coincidence set is attainable. By using the property that a better
approximation of the coincidence set leads to a higher convergence of
this method, the existence of an upper bound for the iterations inde-
pendent of the discretization h will be proved in the following.
The index h denotes the discretization parameter of problem (1.3).
Furthermore let h' < h be, with i' and i the indices of the
discretization points are denoted which belong to the discretiza-
tion h' or h respectively. First of all the behaviour of the
iterates for different discretizations is considered:
Theorem 3.1.:
Denote I I • I I 00 the di screte Max i mum norm and I et Ie and I be two

arbitrary iteration indices, with
k I
II v
h
- v
h'
I 100 (3.1)
and
"'k "'1
II v
h
- v
h'
I
I
I
'00 (3.2)
Then the following estimations are true:
k I
(i) for the step length: t - t
h h'
k+1 1+1 c hl'1
(ii) for the iterates: Ilv v h ' 1100 So 4
h
where ~ .. min {p,q},
i.e., if the previous iterates are closely related then the follo-
wing iterates do it too.
Proof:
i
(i) For the simplicity it is assumed that B ~ B then every x £ B
h' h h
also belongs to B • Denote
h'
k
t .. «v k ) f )/«v ) - (v
"'k
) )
i h i i h i h i
respectively
391
1 1 "'1
t == «v ) 'f ) / ( (v ) - (v ) ),
i' h' i ' i' h' i" h' i '
i i '
the step length at the point x respectively at x
Then the step length of the method is determined by
k 1
t '" mi n { t liE I } respectively t = min { t i'E 1')0.
h i h' i'
k 1 1
Without loss of generality let t ~ t be and t is determined at
h h' h'
the step i'. Obviously the following inequality holds:
o
k 1
t - t S. t - t (3.3)
h h' io i~
The comparison of the step lengths for an arbitrary i and

attached i' gives
1 k 1 1 k "'1
t - t -[(v ) (v ) - (v ) f - (v ) (v ) +
ii' n h I h' i ' h' I ' i h i h' i'
"'1 k 1 k
+ (v h') i ' of i - (v ) (v ) + (v ) f +
h i h' i ' h i i
";k 1 "'k
+ (v ) (v ) - (v ) 'f ]
h i h' i ' h i I
with
k "'Ie 1 "'1
n := [(v ) - (v ) ] [ (v ) - (v ) ]
h i h i h' i' h' i '
Since 'f = 'f , i t follows

Ii'
1 k 1 "'1 "'k
t
i
- t
i '
'" -n { [ (v h) i - (v h') i ,] fi + [ (v h') i' - (v h) I] fi -
k "'1 "'k 1
- (v ) (v ) + (v ) (v ) )0 ..
h I h' i' h i h' i '
1 k 1 "'1 "'k
'" - {[(v ) - (v ) ].pi + [(v ) - (v ) ] f +
n h i h'i' h'l' hi i
"'k 1 "'1 1 "'1 1
+ (v ) (v ) - (v ) (v ) + (v ) (v )
h i h' i ' h' i ' h' I ' h' i' h' i '
k "'1
- (v ) (v ) )0 ==
h i h' I '
1 k 1 "'1 "'k
n
{[(v
h i
) - (v
h' i '
) ] f.1 + [(v
h' I '
) - (v
h
)
i ] 'fi +
"'k "'1 1 1 k "'1

+ [(v
h i
) - (v )
h' i'
](v )
h' I '
+ [(v )
h' i '
- (v )
h I
](v )
h' i'
).
1 "'1
Because of the continuity of the functions~, v and v In B there
exists a constant c Independently of h such that
392
c 1 c "'1 c
l'f I ~ -
i 2
I (v
h' i '
) ~
2
and I (v
h'
)
i
~ -2
Moreover, because of the construction of the search direction we have
k "'k k
(v ) - (v ) (r ) > 0
h i h i h i
and
1 "'1 1
(v ) - (v ) .. - (r ) >0
h' i ' h' i ' h' i '
such that
A
there exists a constant c independently of hand h' with
lin ~ c. Therefore,
I( 1 "'1 "'k
It - t ~ cc I (v ) - (v ) I+ cc I (v ) - (v ) I
i i ' h i h' i' h' i ' h i
and together with the assumptions (3.1> and (3.2) the estimation
A P .., q
It t ~ ccc h + ccc h 5. c h""
i i ' 1 2 3
holds for an arbitrary iteration index. Therefore, immediatly from

(3.3) it is clear that
Itk _ t 1 I ~ c h.M
h h' 3
(ii) Because of the construction of the iterates the following compa-

rison is possible:
k+l 1+1 k k "'k k 1 1 "'1

v
h
- v
h'
=v + t (v - v ) - ( v
h'
+ t
h'
(v
h'
- v
h'
) ]
h h h h
I( 1 k "'k 1 "'1 1 1 k k
== v - v
h'
+ t v - t
h'
v
h'
+ t
h'
v
h'
- t v
h h h h h
k 1 k "'k k "'1 k "'1 1 "'1
=v h
- v
h'
+ t
h
v
h
- t v
h h'
+ t v
h'
- t
h'
v
h'
+
h
1 1 1 k 1 k k k
+ t v - t v + t v - t v
h' h' h' h h' h h h
k k "'I( "'1 1 k 1
=v h
- v
h'
+ t
h
(v
h
- v
h'
) + v
h'
(t
h
- t
h'
) +
1 1 k k 1 k
+ t
h'
(v
h'
- v
h
) + v
h
(t
h'
- t
h
)
The step lengths are bounded and the concerned functions are conti-
nuous over B, therefore a constant c exists independently of h such
that
k 1 k 1
It ~ c , It :c;. c: 'I Ilv
h
: I s. c: and I Iv II 00 ~ c.
h h' 00 h'
393
Thus the estimation
k+ 1 1+1 k 1 k 1 "'1 "'k

Ilv - v II ~ c( It - t I + Ilv - v II + Ilv - v II",,)
h h' 00 h h' h h' 00 h' h
is true and together with (3.1),(3.2) and part (i) of the theorem the
proof is complete;
P q
I I v k+1 - v 1+1 I I 00 ~ c (c h I" + c h + c h ) ~
c hl>'
h h' 3 1 2 4
Remark 3.1.1 The assumptions (3.1) and (3.2) are never of restricted
kind. The meaning of condition (3.1) is that the iteration solu-
tions constructed by the algorithm are closely related and Theorem
3.1. shows that the following iteratws have the same property.
Condition (3.2) is a demand for the approximation of an elliptic
boundary value problem. Hence the existence of q is guaranteed.
k
Remark 3.2.1 The proof was given for the case that v is constructed
k
by Algorithm A~ If v is generated by Algorithm B in the proof it
is taken into consideration that the step lengths t and t a r e
ii'
to calculate not only at the discretization points but also at an
arbitrary point Xi B. But for the step length at the point x I: B
the same estimation is true as at the discretization point.
Because of the small deviation of the iterates it is also true that in

the corresponding coincidence sets a small difference between the
iterates appeare •• Therefore, it can be proved that for a suffiCiently
sma1l discretization h the e. - activity of certain restrictions is
also secured for every discretization parameter h' which is smaller
then h.
The()rem 3.2.1
Let k and 1 be two iteration indices for which
k-1 1-1 k
Ilv - v and t (v ) < 1 •
h h' h
Then for sufficient small h >0 there exists a coincidence set
k A
k k
I Eo (v ) ;6 0 , I E(v ) c I E (v )
h h h
such that I E (v
k
h
) c I E (v
1
h'
) ¥ h' < h .
394
k k
Proof. Since t(v ) < 1 the coincidence set I£(v ) is not empty and
h h
k k
an index i ~ 1£ (v ) exists for which (v )
o h h io
Now, the following set is defined:
k k ...
,
A
I £ (v ) { i E I I (v )
'Pi S. £.
). with 0 < E. < E. (3.4)
h h i
k
A k
Since i E I £ (v ) the set It(v ) is not empty. Using Theorem 3.1. , an
index r 0 h
= min (p,q)
h
(q analogous to (3.2» exists such that
k I k I JA
Ilv
h
- v
h'
II
GO
S. ChI' i. e. I (v
h i
) - (v )
h' i '
S. ch
A
k
Thus, for i E I £. (v ) the inequality
h
(v
I
h' i '
) - 'f
i '
S. (v
k
)
h i
+ ch
,All
- 'P S. E + ch
i
JA
S. f.
is true, if Es..E. - chf'\ •

But if h is chosen sufficiently small, the condition 0 < £. < f is
performed.
I k
Therefore, the relation i E IE(v ) holds for arbitrary i£I~(v ), i. e.
h' ~ h
1
C I < (v ).
" h'
Remark 3.3: The statement of Theorem 3.2. is also true for the case
k
that v is generated by Algorithm B. Therefore, the index set
A
k
IE,(V ) is to replace by a set
h
... k k
BE (v ) = ( )C ~ B I V (x) - 'P(x) S. E. ). , (3.5)
h h
i.e., there exist for sufficiently small h a set

... k A
k k
BE (v h) ¢ 121, B£(v ) C BE (v h) such that
h
k I
BE (v h) c Be(v ) ¥h'<h.
h'
The existence of such a minimum growth set (3.4) (resp. (3.5» is of

decisive importance for the proof that the number of iterations is
independent of the discretization. For the proof of this independence
it is necessary to Imow how the different E. - enlargement works to the
step length.
395
Lemma 3. t.1 (cf. [14])
Starting the Algorithm A (resp. B) at the step k with the diffe-

rent enlargements E- and E.2 E. < E. ) , then
1 1 2
k k
t ( E ) S. t ( E. ) holds for the step lengths.
1 2
The following lemma describes the variation of the total step length
if the discretization parameter is modified.
Lemma 3.2.1 (cf.[14J)
k
Let v and v be generated by Algorithm Band let
h h'
k 1
t(v ) 5. t(v h' ) be.
h
Then, at the next iteration step the inequality
k+1 1+1
t(v 5. t(v holds.
h h'
The behaviour of the algorithm is the following: a larger total step

length is evident with a larger coincidence set. Lemma 3.2. shows, if
one coincidence set includes the other, then this property holds at
the next iteration step too. This property leads to the statement that
the number of i t.erati ons for sol vi ng probl em (1. 3) can not increase
arbitrarily large, moreover, it is bounded above by a constant which
is independent of the discretization. In order to do that Algorithm B
k
has to be modified in the following manner: at the k-th step let v be
determined, by Theorem 3.1. there exists a set
k k k ..... k k
I.(v );Ii 0, Ir(v ) c I.(v ) such that I.(v ) c IE,(v ) ¥h'( h.
Co h c. h co h c. h h'
Ie Ie
Now, the algorithm continues with Ie(v ) .'" Ie (v ), i.e. , at the
h h
k Ie
iteration step k the set Ie(v h) is replaced by the subset Ie(v h)'
The iteration sequence which is generated by this modification is
"Ie
denoted by (v). This modified algorithm stops after finite
h
Ie
number of steps (infact of (3.4) the modified method with IE. (v )
h
is interpretable as an algorithm with t - enlargement).
396
Lemma 3.3.:
"'k
The sequence (v ) which is generated by the modified Algorithm B
h
converges to an £ - optimal solution of problem (1.3) after a
finite number of steps, i.e., there exists an index k such that
o
and 0 s. u u
h
S. £e •
Now, Lemma 3.3. guarantees that the iteration number of Algorithm A is

bounded above
by the constant k. If k (h) denotes the number of
o 0
iterations of Algorithm A for solving problem (1.3) with the discreti-
zation h, the following statement can be proved:
Theorem 3.3.:
For every discretization the iteration number of Algorithm A is

bounded above with a constant k , i.e., there exists a constant k
o 0
which is ind.ependent of the discretization h such that
I( (h) S. k .., h •
o 0
Ie
Proof. For arbitrary h' the sequence ( v ) , generated by Algorithm A,
h'
is compared with (v
h
..... k
h
) . Since
) with f. < £ can be chosen as in
I£(v
k A
(3.4) , Lemma 3.1- and 2.2 provide the following estimation for the
step length at the O-th step:
o 0 0 ..... 0
t (v ) 2: t (v ).
h' h
1 A1
This leads to the total step length t(v ) 2: t(v ).
h' h
k ..... k
Now, at the step Ie let t(v ) 2: t(v ) be. As an immediate consequen-
h' h
ce of Lemma 3.2. and 2.2.
k+1 ..... 1e+1

t(v ) 2: t(v )
h' h
holds.
Ie Ale
Therefore, t(v ) 2: t(v ) holds for all k and particular for Ie the
h h 0
inequality
Ie ..... Ie
t(v 0 ) 2: t(v 0) =1
h' h
can be concluded from Lemma 3.3 ••

397
Hence, Algorithm A terminates at latest at the step k , i.e.

o
k (h') ~ I( •
o 0
The behaviour of the regularized method of feasible directions is

distinguished by the fact: the greater the coincidence set of problem
(1.1) or (1.2) with respect to B the smaller is the dimension of the
linear equation system (2.1) which is to be solved at every step.
Conversely, if the coincidence set is relatively small with respect to
B the direction search problem generates a "global" descent and the
optimal solution is computable in few iteration steps.
4. Numerical Results
The theoretical pr'oof of the independence of the iteration number from

the discretization can be verified by practical computations too.
For instance, the following Torsion Problem is considered
(cf.[10], Chapter 3.7.4):
2 2 2 0 .. 5
B { X E R r :c (>: + x) < 0.5 } •
1 2
L[u] '" - 6 u , f(x) - -10 o • f(x) '" - (0. 5 - r) •
o
The starting point is v (x) '" O. the domain B is discretized by a
regular triangulation with the grid constant h.
Table below contains the iteration number k for the computation of
problem (1.3). (1.4) with Algorithm A for several discretization
parameters h (€ '" 0.001).
The iteration numbers for other numerical methods solving this problem
with the same precision are described in Table 2.
Table 1:
h 0.1 0.05 0.025 0.0125

N 109 397 1:::;19 6109
k 6 6 6 6
398
Table 21
penalty manifold relaxation

h N method subprogr. (w= 1.9)
0.1 109 4 3 74
0.05 397 5 3 76
0.025 1519 8 6 80
0.0125 6109 9 10 97
Literatur
[1] AHN, B.H.: Iterative methods for linear complementarity prob-

lems with upper bounds,
Math. Programming 26 (1983) 295-315.
[2] AUBIN, J.P.: Approximation of variational inequalities,

in: Functional Analysis and Optimization, Ed. E.R. CAIANIELLO,
Acad. Press, New York, 1966, 7-14.
[3] CHANEY, R.W.: On the rate of convergence of some feasible

directions algorithms,
JOTA 20 (1976) 297-313.
[4] CIARLET, P.G.: Discrete maximum principle for finite

difference operators,
Aequat. Math. 4 (1970) 338-352.
[5] DUVAUT, G., J.L. LIONS: Les inequations en mecanique et en

physique,
DLlnod, Paris, 1972.
[6] EKELAND,I., R.TEMAMI Convex analysis and variational problems,
North-Holland Publ.Co., Amsterdam-Oxford 1976.
[7] FAGE, D.M.: The indicatri>: method in free bOLlndary problems,

Numerische Mathematik 38 (1981) 39-52 and 255-261.
[8] FLETCHER, R., M.J.D. POWEL: A rapidly convergent method for
minimization,
Computing J. 6 (1963) 163-168.
[9] GLOWINSKI, R.: Numerical methods for nonlinear variational

problems,
Springer Verlag, New York, 1984.
[10] GLOWINSKI, R., J.L. LIONS, R. TREMOLIERES: Analyse numerique

des inequations variationelles,
Dunod, Paris, 1976.
399
[llJ HACKBUSCH, W., H.D. MITTELMANN: On multi-grid methods for

variational inequalities,
Numerische Mathematik 42 (1983) 65-76.
(12) HARTWIG, K.H.I Numerische Loesung allgemeiner linearer ellip-

tischer Variationsungleichungen mittels Di fferenzenverfahren ,
Diss. A, TH Karl-Marx-Stadt, 1984.
[13J KINDERLEHRER, D., G.STAMPACCHIA: An introduction to variatio-

nal inequalities and their applications,
Acad. Press, New York, 1980.
[14) KIRSTEN, H.I Numerische Loesung elliptischer Variationsunglei-
chungen mit Verfahren der zulaessigen Richtungen,
Diss. A, TH Karl-Marx-Stadt, 1985.
[15J KIRSTEN, H., R. TICHATSCHKE: Ueber ein Verfahren der zulaessi-
gen Richtungen zur Loesung des Hindernisproblems (russ.),
Optimizatija (Novosibirsk) 34 (51), (1984) 88-100.
[16J KIRSTEN, H., R. TICHATSCHKE: On a method of feasible direc-

tions for solving variational inequalities,
Mathem. Operationsforschung und Statistik, 8er. Optimization,
16 (1985) 4, 535-546.
[17) LADVZENSKAJA,O.A., N.N. URAL'ZEWA: Lineare und quasilineare

Gleichungen elliptischen Typs (russ.),
Nauka, Moskau, 1973.
[18J LIONS, J.L.I Quelques methodes de resolution des problemes aux

limites non lineaires,
Dunod, Paris, 1969.
[19J MANDEL, J.: A multilevel iterative method for symmetric posi-

tive definite linear complementary problems,
Appl. Math. and Optim. 11 (1984) 77-95.
[20] MICHLIN, 8.G.: Lehrgang der mathematischen Physik,

Alcademie Verlag, Berlin 1972.
[21J NATTERER, F., B. WERNER: Verallgemeinerung des Maximumprinzips
fuer den Laplaceoperator,
Numerische Mathematik 22 (1974), 149-156.
[22J TICHATSCHKE, R.: Lineare semi-infinite Optimierungsaufgaben

und ihre Anwendung in der Approximationstheorie,
Wissenschaftliche Schriftenreihe, TH Karl-Marx-Stadt,
Heft 4, 1981.
[23J TICHATSCHKE, R.: Semi-infinite programming problems,

Mathem. control theory, Banach-Center-Publ., vol.14, 543-554,
PWN - Polish Scientific Publishers, Warsaw 1985.
[24J WERNER, B.: Monotonie und finite Elemente bei elliptischen
Differentialgleichungen,
18NM 27, Birkhaeuser Verlag 1975.
BIVARIATIONAL BOUNDING METHODS
P.D. Robinson
College of Science, Sultan Qaboos university,
PO Box 32486 Al Khoud, Sultanate of Oman.
and
P.K. Yuen
School of Mathematical Sciences,
University of Bradford, Bradford, U.K.
I. INTRODUCTION
A problem of interest in optimizaLion theory is that of bounding

the inner product <g,u> of an arbitrary vector 9 with u, the solution-
vector of an equation Fu = 0 in the inner-product space. In many
situations a suitable choice of 9 can lead to bounds on the solution u
itself. We describe here an approach based on a bivariational approx-
imation to <g,u>, namely
J(V,U) = -<V,FU> + <g,U>. (1 )
For simplicity we assume that the space is real and the inner product
is symmetric. The square-norm notation aun2 = <U,U> is employed.
2. LINEAR EQUATIONS WITH POSITIVE OPERATORS
Let u be the solution (assumed unknown) of a linear equation

Au - f = 0, (2 )
where the operator A has an adjoint A* so that for suitable arbitrary
V and U:
<V,AU> = <A*V,U>. (3)
Suppose that A is positive in the sense that for arbitrary U
<U,AU> = <U,t(A+A*)U> = <U,A*U> ~ b<U,U>, b>O. (4 )
Then it is possible to derive the bounds

J + ~b-l(s+C) ~ <g,u> ~ J + ~b-l(S-C), (5 )
where
J J(V,U) -<V,AU> + <V,f> + <g,U>, (6)
S S(V,U) <AU - f, A*V - g>, (7)
C C(V,U) = jAU - fll IA*V - gil ( 8)

The bivariational approximation J(V,U) on which these bounds are
based is associated with the pair of equations
401
Au - f = 0, A*v - g = 0, (9)
(v also assumed unknown), and we note that
J(v,u) = <g,u>. (10)
If
U = u + ou, V = v + OV (11)
then
J(V,U) - <g,u> = -<ov,Aou> (12)
and as the right-hand side of (12) has indefinite sign then J(V,U)
does not provide a direct bound for <g,u>. However if the trial
vectors U and V are constrained to satisfy
t(AU-f) + t- 1 (A*V-9) = tAou + t- 1A*OV = 0, t>O, (13)
and the resulting functional J is designated J+, then
J+ -<g,u> = t- 2 <ov,A*OV> ~ 0. (14)
Similarly if the constraint
t(AU-f) - t- 1 (A*V-9) = tABu - t- 1A*Bv = 0, t > 0, (15)
specifies J_ , then,
J_ -<g,u> = -t- 2 <ov,A*OV> S 0. (16)
Thus we have the constrained bounds
J+(V,U) ~ <g,u> ~ J_(V,U), (17)
which may be difficult to implement directly as the inverses of A
and A* are assumed unknown.
The canonical structure of these bounds is exhibited by working
in terms of the new vector-variables:
X+Y = tU, x+y = tu, X-Y = t- 1v, x-y = t- 1v, p+q = tf, p-q = t- 1 g (18)
and the operators
L = i(A+A*), M = l(A-A*), (19)
these latter denoting the self-adjoint and skew-self-adjoint parts of
A. The equation-pair (9) becomes
Lx + My - P = 0, Ly + Mx - q = 0, (20)
and we have
J(V,U) = I(X,Y) = -<X,LX> + <Y,LY> - 2<X,MY> + 2<X,p> - 2<Y,q>. (21)
Since from (4)
<Z,LZ> ~ b<Z,Z>, b>O, (22)
the functional I(X,Y) is convex in Y and concave in X. It thus
possesses the saddle property characteristic of dual variational
bounding principles [~]. In fact
J+ = 1+ = I(X,Y): pI/~X = °<=> LX + MY p = 0, (23)
and
J_= I = I(X,Y): aI/~Y °<=> LY + MX - q = 0. (24)
402
Specifically
I+ = <Y,LY> - 2<Y,q> + «MY-p), L- 1 (MY-P»
I(X,Y) + «LX+MY-p), L- 1 (LX+MY-P», X arbitrary, (25)
and similarly
rot .... -1 tJ
I_ - I(X,Y) - «LY+MX-q), L (LY+MX-q», Y arbitrary. (26)
1
Should L- be known, then (25) and (26) furnish direct upper and lower
bounds on <g,u>. Otherwise they can be weakened by use of the
inequality
(27)
Transformation back to the original vector-variables, and optimization
with respect to the disposable parameter t, finally yields the cons-
traint-free result stated in (5) - (8) above [~].
It is interesting to note that in the especially simple case when
A is self-adjoint and when g = f the bounds in (5) reduce to the result
R(U) + b-111AU - fll2 ~ <f,u> ~ R(U) (28)
involving the Rayleigh functional
R(U) = -<U,AU> + 2<f,U> (29)
of which (6) is'a generalization.
If a weaker condition than (4), namely
II AU U ~ 1311 UII, ~ > 0, (30)
is assumed, then it is possible to derive the weaker bounds
J + j ~-1 C ~ <g,u> ~ J - l~-l C (31)
by direct Schwarz inequality methods [~].
Bounds on other functionals of u can also be derived in some
situations. For example if
<U,(A ± B)U> ~ E±<U,U>, (32)
then we can show that
J + 1:. IIAU - f) + (A*V-2BU)\l2 ~ <u,Bu> ~ J -.2:. II (AU-f) (A*V-2BU)U 2
4E 4E+ (33 )
where
,.,
J - <V,AU> + <V,f> + <U,BU>. (34)
3. NONLINEAR EQUATIONS
Results based on the bivariational approximation (1) are

available for classes of non-linear equations
Fu = 0; (35)
these are obtained by Schwarz inequality methods, and generalize
(31) [i,2,~]. For example, if for all suitable U1 and U2
403
"FU 1 - FU 2 " ~ cn U1 - U20, c > 0, (36)

and
UFU1 - FU 2 - A(U 1 - U2 )0 s t AUU 1 - u202, A> 0, (37)
then we have the result
J + C ~ <g,u> ~ J - ""C (38)
where
(39)
Here A is an arbitrary linear operator, which can be chosen for ease of
computation [2,~]. The G~teaux derivative choice
A = F' (U) (40)
ensures that the bounds are bivariational [~]. The underlying framework
for bounds of this nature has been explored in detail [lJ.
4. SOME APPLICATIONS
Apart from their direct use in estimating quantities <g,u> of

intrinsic interest, these methods can be adapted to find optimal
approximations to the solution-vectors of various problems. Recently
they have been employed successfully with integral equations [~,2'~]
and with integro-differential equations. For an integral equation of
the form
J
~
u(x) + k(x,y)u(y)dy fIx)
(41)
a
set in a space with inner product
J
~
<V,U> = V(x)U(x)dx , (42)
we can take
g(x) = k(x' ,x) (43)
giving
<g,u> = f(x') - u(x'), (44)
so that bounds on <g,u> yield pointwise points on u(x'). This
generalizes previous work for integral equations with symmetric kernels
[~]. Integro-differential equations can be tackled in a similar way by
first converting them into equations like (41) using appropriate Green's
functions.
404
To illustrate the high accuracy which is obtainable with single-

parameter trial vectors, consider the Volterra integro-differential
problem
Jx 2u(y)dy =
]
1
u'(x) + t X2, 0 < X < 1,
x (45)
u(O) =0
for which no analytical solution seems available. The equivalent inte-
gral equation is
x 1
u(x) + i J z2dzf u(y)dy = -x
1 3
3
, o ~ x ~ 1 , (46)
o z
and the relevant constant in (4) is b = 1. The trivial choice of trial

vectors U = 0, V = 0 leads to the bounds
=.....£ x 3 + __1_ x 7 + 1 x 3{ 1 _ 2 x}i

u± ( X)
144 336 - 12./7 9" n ' (47)
and an optimized choice of the form

U (x) = cf (x) ., V(x' , x) = dk (x' , x) (48)
yields positive bounds
x' 0.2 0.4 0.6 0.8 1.0
2.55915625x10- 3 2.0482022x10- 2 6.925523x10- 2 0.1649823 0.325632

2.55915621x10- 3 2.0482020x10- 2 6.925522x10- 2 0.1649817 0.325628.
The accuracy here is much better than can be obtained using a con-
ventional 200-step finite-difference Simpson-quadrature approach.
For problems where u is an n-vector with components (u1 ,u 2 ' •••
un)' we can take g = (0,0 ••• 1 .•• 0), so that <g,u> is any desired com-
ponent of u. Thus in principle bounds can be obtained on solutions of
difference equations, or on solutions of discretized differential
problems.
Adaptation of bivariational methods to programming problems
involving inequality constraints is possible [~].
~5
REFERENCES
1. Sewell, M.J. and Noble, B., Proc. Roy. Soc. (London) A361,
293-324 (1978).
2. Robinson, P.D. and Yuen, P.K., S.I.A.M. J. Numer. Anal.
23, 1230-1240 (1976).
3. Barnsley, M.F. and Robinson, P.D., Proc. Roy. Soc. (Edinburgh)
75A, 109-118 (1976).
4. Barnsley, M.F. and Robinson, P.D., J. lnst. Maths. Applics.,
9, 485-504 (1977).
5. pack,-D.C., Cole, R.J. and Mika, J., I.M.A.J. Appl. Maths.,
32, 253-266 (1984).
6. Robinson, P.D. and Yuen, P.K., Bivariational Methods for
Hammerstein Integral Equations, submitted for publication.
7. Robinson, P.D. and Barnsley, M.F., S.I.A.M. J. Numer. Anal.
16, 135-144 (1979).
Vol. 211: P. van den Heuvel. The Stability of a Macroeconomic Vol. 237: Misspecification Analysis. Proceedings, 1983. Edited by
System with Quantity Constraints. VII. 169 pages. 1983. T. K. Dijkstra. V, 129 pages. 1984.
Vol. 212: R. Sato and T. NOno. Invariance Principles and the Vol. 238: W; Domschke, A. Drexl, Location and Layout Planning.
Structure of Technology. V. 94 pages. 1983. IV, 134 pages_ 1985.
Vol. 213: Aspiration Levels in Bargaining and Economic Decision Vol. 239: Microeconomic Models of Housing Markets. Edited by
Making. Proceedings. 1982. Edited by R. Tietz. VIII. 406 pages. 1983. K. Stahl. VII, 197 pages. 1985.
Vol. 214: M. Faber. H. Niemes und G. Stephan. Entropie. Umwelt- Vol. 240: Contributions to Operations Research. Proceedings,1984.
schutz und Rohstoffverbrauch. IX. 181 Seiten.1983. Edited by K. Neumann and D. Pallaschke. V, 190 pages. 1985.
Vol. 241: U. Wittmann, Das Konzept rationaler Preiserwartungen.

Vol. 215: Semi-Infinite Programming and Applications. Proceedings.
XI, 310 Seiten. 1985.
1981. Edited by A. V. Fiacco and K.O. Kortanek. XI. 322 pages. 1983.
Vol. 242: Decision Making with Multiple Objectives. Proceedings,
Vol. 216: H. H. MUlier. Fiscal Policies in a General Equilibrium Model
1984. Edited by Y. Y. Haimes and V. Chankong. XI, 571 pages. 1985.
with Persistent Unemployment VI. 92 pages. 1983.
Vol. 243: Integer Programming and Related Areas. A Classified
Vol. 217: Ch. Grootaert, The Relation Between Final Demand and
Bibliography 1981-1984. Edited by R. von Randow. XX, 386 pages.
Income Distribution. XIV. 105 pages. 1983.
1985.
Vol. 218: P. van Loon. A Dynamic Theory of the Firm: Production.
Finance and Investment VII. 191 pages. 1983. Vol. 244: Advances in Equilibrium Theory. Proceedings,1984. Edited
by C. D. Aliprantis, O. Burkinshaw and N.J. Rothman. II, 235 pages.
Vol. 219: E. van Damme. Refinements ofthe Nash Equilibrium Concept 1985.
VI. 151 pages. 1983. Vol. 245: J.E.M. Wilhelm, Arbitrage Theory. VII, 114 pages. 1985.
Vol. 220: M. Aoki. Notes on Economic Time Series Analysis: System Vol. 246: P. W. Otter, Dynamic Feature Space Modelling, Filtering
Theoretic Perspectives. IX. 249 pages. 1983. and Self-Tuning Control of Stochastic Systems. XIV,177 pages.1985.
Vol. 221: S. Nakamura. An Inter-Industry Translog Model of Prices Vol. 247: Optimization and Discrete Choice in Urban System~.
and Technical Change for the West German Economy. XIV. 290 Proceedings, 1983. Edited by B. G. Hutchinson, P. Nijkamp and
pages. 1984. M. Batty. VI, 371 pages. 1985.
Vol. 222: P. Meier. Energy Systems Analysis for Developing Countries.
Vol. 248: Plural Rationality and Interactive Decision Processes. Pro-
VI. 344 pages. 1984. ceedings. 1984. Edited by M. Grauer, M. Thompson and A.P.
Vol. 223: W. Trockel. Market Demand. VIII. 205 pages. 1984. Woerzbicki. VI, 354 pages. 1985.
Vol. 224: M. Kiy. Ein disaggregiertes Prognosesystem fUr die Bundes- Vol. 249: Spatial Price Equilibrium: Advances in Theory, Ccmputation
and Application. Proceedings, 1984. Edited by P. T. Harker. VII, 277
republik Deutschland. XVIII. 276 Seiten. 1984.
pages. 1985.
Vol. 225: T. R. von Ungern-Sternberg. Zur Analyse von Markten mit
Vol. 250: M. Roubens, Ph. Vincke, Preference Modelling. VIII.
unvollstandiger Nachfragerinformation. IX. 125 Seiten. 1984
94 pages. 1985.
Vol. 226: Selected Topics in Operations Research and Mathematical Vol. 251: Input-Oulput Modeling. Proceedings, 1984. Edited by
Economics. Proceedings. 1983. Edited by G. Hammer and D. A. Smyshlyaev. VI, 261 pages. 1985.
Pallaschke. IX. 478 pages. 1984.
Vol. 252: A. Birolini, On the Use of Stochastic Processes in Modeling
Vol. 227: Risk and Capital. Proceedings. 1983. Edited by G. Bam- Reliability Problems. VI, 105 pages. 1985.
berg and K. Spremann. VII. 306 pages. 1984.
Vol. 253: C. Wothagen, Economic Th~ry and Intemational Trade in
Vol. 228: NonlinearModels of Fluctuating Growth. Proceedings.1983. Natural Exhaustible Resources. VI. 172 pages. 1985.
Edited by R. M. Goodwin. M. Kruger and A. Vercelli. XVII. 277 pages.
1984. Vol. 2,54: S. Mulier, Arbitrage Pricing of Ccntingent Claims. VIII, 151
pages. 1985.
Vol. 229: Interactive Decision Analysis. Proceedings. 1983. Edited
by M. Grauer and A.P. Wierzbicki. VIII. 269 pages. 1984. Vol. 255: Nondifferentiable Optimization: Motivations and Applica-
tions. Proceedings, 1984. Edited by V. F. Demyanov and D. Paiissch-
Vol. 230: Macro-Economic Planning with Conflicting Goals. Proceed- ke. VI, 350 pages. 1985.
ings. 1982. Edited by M. Despontin. P. Nijkamp and J. Spronk. VI. Vol. 256: Convexity and Duality in Optimization. Proceedings, 1984.
297 pages. 1984. Edited by J. Ponstein. V, 142 pages. 1985.
Vol. 231: G. F. Newell. The M/M/= Service System with Ranked Ser- Vol. 257: Dynamics of Macrosystems. Proceedings, 1984. Edited by
vers in Heavy Traffic. XI. 126 pages. 1984. J.-P. Aubin, D. Saari and K. Sigmund. VI, 280 pages. 1985.
Vol. 232: l. Bauwens. Bayesian Full Information Analysis of Vol. 258: H. Funke, Eine allgemeine Theorie der Polypol- und
Simultaneous Equation Models Using Integration by Monte Carlo. Oligopolpreisbildung. III, 237 pages. 1985.
VI. 114 pages. 1984.
Vol. 259: Infinite Programming. Proceedings, 1984. Edited by
Vol. 233: G. Wagenhals. The World Copper Market XI. 190 pages. E.J. Anderson and A.B. Philpott. XIV, 244 pages. 1985.
1984.
Vol. 260: H.-J. Kruse, Degeneracy Graphs and the Neighbourhood
Vol. 234: B. C. Eaves. A Course in Triangulations for Solving Problem. VIII, 128 pages. 1986.
Equations with Deformations. III. 302 pages. 1984.
Vol. 261: Th.R.Gulledge, Jr., N.K. Womer, The Economics of Made-
Vol. 235: Stochastic Models in ReliabilityTheory. Proceedings. 1984. to-Order Production. VI, 134 pages. 1986.
Edited by S. Osaki and Y. Hatoyama. VII. 212 pages. 1984. Vol. 262: H. U. Buhl, A Neo-Classical Theory of Distribution and
Wealth. V. 148 paQltS. 1986.
Vol. 236: G. Gandolfo. P.C. Padoan. A Disequilibrium Model of
Real and Financial Accumulation in an Open Economy. VI,172 pages. Vol. 263: M. Schafer, Resource Extraction and Market Structure. XI,
1984. 154 pages. 1986.
Vol. 264: Models of Economic Dynamics. Proceedings, 1983. Ed~ed Vol. 293: D. Batten, J. Casti, B. Johansson (Eds.), Economic Evolu-
by H.F. Sonnenschein. VII, 212 pages. 1986. tion and Structural Adjustment. Proceedings, 1985. VI, 382 pages.
1987.
Vol. 265: Dynamic Games and Applications in Economics. Ed~ed by
T. 8a§ar. IX, 288 pages. 1986. Vol. 294: J. Jahn, W. Krsbs (Eds.), Recent Advances and Historical
Developmeni of Vector Optimization. VII, 405 pages. 1987.
Vol. 266: Multi-Stage Production Planning and Inventory Control.
Edited by S. Axsater, Ch. Schneeweiss and E. Silver. V, 264 pages.
1986.
Vol. 267: R. Bemelmans, The Capacity Aspect of Inventories. IX, 165
pages. 1986.
Vol. 268: V. Firchau, Information Evaluation in Cap~al Markets. VII, 103
pages. 1986.
Vol. 269: A. Borglin, H. Keiding, Optimality in Infinite Horizon Econo-

mies. VI, 180 pages. 1986.
Vol. 270: Technological Change, Employment and Spatial Dynamics.
Proceedings 1985. Edited by P. Nijkamp. VII, 466 pages. 1986.
Vol. 271: C. Hildreth, The Cowles Commission in Chicago, 1939-
1955. V, 176 pages. 1986.
Vol. 272: G. Clemenz, Credit Markets with Asymmetric Information.
VIII, 212 pages. 1986.
Vol. 273: Large-Scale Modelling and Interactive Decision Analysis.
Proceedings, 1985. Edited by G. Fandel, M. Grauer, A. Kurzhanski
and A.P. Wierzbicki. VII, 363 pages. 1986.
Vol. 274: W. K. Klein Haneveld, Duality in Stochastic Linear and
Dynamic Programming. VII, 295 pages. 1986.
Vol. 275: Competition, Instability, and Nonlinear Cycles. Proceedings,
1985. Ed~ed by W. Semmler. XII, 340 pages. 1986.
Vol. 276: M.R. Baye, D.A. Black, Consumer Behavior, Cost of Living
Measures, and the Income Tax. VII, 119 pages. 1986.
Vol. 277: Studies in Austrian Capital Theory, Investment and Time.
Edited by M. Faber. VI, 317 pages. 1986.
Vol. 278: W. E. Diewert, The Measurement of the Economic Benefits of
Infrastructure SeNices. V, 202 pages. 1986.
Vol. 279: H.-J. Buttler, G. Frei and B. Schips, Estimation of Disequi-
librium Models. VI, 114 pages. 1986.
Vol. 280: H. T. Lau, Combinatorial Heuristic Algorithms w~h
FORTRAN. VII, 126 pages. 1986.
Vol. 281: Ch.-L. Hwang, M.-J. Lin, Group Decision Making under
Multiple Criteria. XI, 400 pages. 1987.
Vol. 282: K. ScMtkowski, More Test Examples for Nonlinear Pro-
gramming Codes. V, 261 pages. 1987.
Vol. 283: G. Gabisch, H.-W. lorenz, Business Cycle Theory. VII,
229 pages. 1987.
Vol. 284: H. L(jtkepohl, Forecasting Aggregated Vector ARMA
Processes. X, 323 pages. 1987.
Vol. 285: Toward Interactive and Intelligent Decision Support
Systems. Volume 1. Proceedings, 1986. Edited by Y. Sawaragi,
K. Inoue and H. Nakayama. XII, 445 pages. 1987.
Vol. 286: Toward Interactive and Intelligent Decision Support
Systems. Volume 2. Proceedings, 1986. Edited by Y. Sawaragi,
K. Inoue and H. Nakayama. XII, 450 pages. 1987.
Vol. 287: Dynamical Systems. Proceedings, 1985. Edited by A. B.
Kurzhanski and K. Sigmund. VI, 215 pages. 1987.
Vol. 288: G.D. Rudebusch, The Estimation of Macroeconomic Dis-
equilibrium Models with Regime Classification Information. VII, 128
pages. 1987.
Vol. 289: B. R. Meijboom, Planning in Decentralized Firms. X, 168
pages. 1987.
Vol. 290: D.A. Carlson, A. Haurie, Infinite Horizon Optimal Control.
XI, 254 pages. 1987.
Vol. 291: N. Takahashi, Design of Adaptive Organizations. VI, 140
pages. 1987.
Vol. 292: I. Tchijov, L. Tomaszewicz (Eds.), Input-Output Modeling.
Proceedings, 1985. VI, 195 pages. 1987.
M.Aoki
State Space Modeling of Time Series

1986. 76 figures. X, 314 pages. fSBN 3-540-17256-4
Contents: Introduction. - The Notion of State. - Representation of Time
Series. - State Space and ARMA Representation.- Properties of State Space
Models. - Innovation Processes. - Kalman Filters. - State Vectors and Opti-
mality Measures. - Computation of System Matrices. - Approximated
Model and Error Analysis. - Numerical Examples. - Appendices. -
References. - Subject Index.
G.Gandolfo
International Economics
1986. 133 figures. XXVI, 804 pages. ISBN 3-540-16707-2
Contents: The Pure Theory ofInternational Trade: Introduction. The clas-
sical (Ricardo-Torrens) theory of comparative costs. The neoclassical theory
of international trade. The Heckscher-Ohlin model. Tariffs, protection,
economic integration. International trade and economic growth. Some refi-
nements ofthe orthodox theory. The "new" theories of international trade.
Neo-Ricardian theories of international trade. - Bibliography.
International Monetary Theory: The foreign exchange market. Balance of
payments and national accounts. The role of the exchange rate in the
adjustment process in a partial equilibrium framework. The role of income
changes in the adjustment process. The absorption approach and interac-
tions between exchange-rate and income in the adjustment process. Money
and other assets in the adjustment process under fixed exchange rates.
Money and other assets in the adjustment process under flexible exchange
rates. International capital movements and other problems. FIxed versus
flexible exchange rates. International liquidity and international fmancial
markets. The problem of the integration between the pure theory of inter-
national trade and international monetary economics. - Bibliography. -
Name Index. - Subject Index.
W. Whitmore, Jr.
Aggregate Economic Choice

1986.40 figures. X, 427 pages. ISBN 3-540-16162-7
Contents: Conventional Macro Models and Aggregate Economic Choice. -
The Basic Structure. - The Household Sector in the Simple Model. - The
Nonfmancial Business Sector in the Simple Model. - National Product and
the Dynamic Properties of the Simple Model. - Government Activity and
Aggregate Economic Choice. - The Government Sector in the Expanded
Model. - Household Sector Behavior in the Expanded Model. - The Non-
Springer-Verlag fmancial Business Sector in the Expanded Model. - The Private Fmancial
Berlin Heidelberg New York Sector. - The Central Bank. - The Complete Model of a Closed Economy.
London Paris Tokyo - Model of a Large Open Economy. - Conclusion.

CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler

Uploaded by

Copyright:

Available Formats

You might also like

CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler

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

CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler

Uploaded by

Copyright:

Available Formats

Lecture Notes in Economics and Mathematical Systems

continuation on page 407

J. Jahn W. Krabs (Eds.)

H.Albach M. Beckmann (Managing Editor) P. Dhrymes

Prof. Dr. M. Beckmann

Prof. Dr. W. Krelle

Prof. Dr. Johannes Jahn

Prof. Dr. Werner Krabs

ISBN-13: 978-3-540-18215-3 e-ISBN-13: 978-3-642-46618-2

This volume contains refereed contributions to the "International

Johannes Jahn Werner Krabs

II. MATHEMATICAL THEORY

S. DOLECKI and C. MALIVERT:

K.-H. ELSTER and A. GOPFERT:

III. GOAL SETTING AND DECISION MAKING

Lj. VLACIC, A. WIERZBICKI and B. MATIC:

P.L. YU and I.S. CHIEN:

IV. ENGINEERING APPLICATIONS

H. KIRSTEN and R. TICHATSCHKE:

P.D. ROBINSON and P.K. YUEN:

riulticriteria optimization is what one attempts to carry out in

~Vhat follows is a biographical treat~ent of these authors and their

As with most mathematical disciplines, multicriteria optimization

The subject of this paper is the evolution of the following

Let Q (open) c nn and introduce the inequality constraints

the equality constraints

with corresponding functional constraint set

x = {XEJRn : XEQ, f(x):<> 0, h(x) = OJ.

The criterion functions are

gi(') : X - R, i = 1,2, ... ,N,

with corresponding criteria map

The attainable set is

Y = g (X) = {y EJRN : y = g (x), x EX} .

x :<> y iff x. :<> Yi' ViE I = {1,2, ... ,n};

x < Y iff x :<> Y, and x t y;

x« Y iff x 1 < y i ' ViE I.

In a broad sense, multicriteria optimization is nothing more than

induce such a preference by means of a mapping ~(.) D .... JR., with

Figure 1. Decisions with one Desideratum

is used to induce a preference ~ on D via the mapping ~(.). In the

An analogous approach is used in multicriteria optimization - one

latter multicriterion approach han become the accepted one in the

Generally, the fundamental purpose of a preference on a decision

If the preference is induced by means of a mapping g(.) = (g1 (.),

g(d) ~ g(d) ==l> g(d) - g(d).

In order to make use of traditional single criterion optimization

x;iy <F;> W(x) sW(y),

x, Y ElRN Thus, W(.) may again be viewed as a utility function for ~

Ideally, maximizing W(g (d» with respect to dE D then generates the

These are the essentials of multicriteria optimization. ~1any have

Vector optimization is an inherent part of economic equilibrium

A second contributing branch was the inception of the theory of

In 1951, Koopmans introduced the efficient point set in the context

Hathematically, the subject can be said to have had its origin in

As indicated, multicriteria optimization arose naturally in

The following papers present a selective history of multicriteria

3. Welfare Theory and Scalarization

Welfare theory traditionally deals with competing economic agents,