Studies in Fuzziness and Soft Computing

Susanne Saminger-Platz
Radko Mesiar Editors

On Logical,
Algebraic, and
Probabilistic
Aspects of Fuzzy
Set Theory
Studies in Fuzziness and Soft Computing

Volume 336

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
About this Series

The series “Studies in Fuzziness and Soft Computing” contains publications on

various topics in the area of soft computing, which include fuzzy sets, rough sets,
neural networks, evolutionary computation, probabilistic and evidential reasoning,
multi-valued logic, and related fields. The publications within “Studies in
Fuzziness and Soft Computing” are primarily monographs and edited volumes.
They cover significant recent developments in the field, both of a foundational and
applicable character. An important feature of the series is its short publication time
and world-wide distribution. This permits a rapid and broad dissemination of
research results.

More information about this series at http://www.springer.com/series/2941

Susanne Saminger-Platz Radko Mesiar
•

Editors

On Logical, Algebraic,
and Probabilistic Aspects
of Fuzzy Set Theory

123
Editors
Susanne Saminger-Platz Radko Mesiar
Department of Knowledge-Based Department of Mathematics and Descriptive
Mathematical Systems Geometry
Johannes Kepler University Linz Slovak University of Technology
Linz Bratislava
Austria Slovakia

ISSN 1434-9922 ISSN 1860-0808 (electronic)

Studies in Fuzziness and Soft Computing
ISBN 978-3-319-28807-9 ISBN 978-3-319-28808-6 (eBook)
DOI 10.1007/978-3-319-28808-6

Library of Congress Control Number: 2015960233

© Springer International Publishing Switzerland 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by SpringerNature

The registered company is Springer International Publishing AG Switzerland
Dedicated to Erich Peter Klement

Copyright OÖN/Weihbold
Preface

Erich Peter Klement got interested in fuzzy set theory already in the 1970s while
being a young Assistant Professor at Johannes Kepler University in Linz, Austria.
In 1979, he stayed with Lotfi A. Zadeh at Berkeley University as a Visiting
Research Associate, several other research visits to universities in Europe and the
United States followed. It was also in 1979 when, together with Ulrich Höhle and
Robert Lowen, he first organized the “International Seminar on Fuzzy Set Theory”
in Linz. At that time, he could not know that there would be more than 35 seminars
to follow, well-established and widely known to the scientific community as the
“Linz Seminars on Fuzzy Set Theory”.
Organizing and hosting the seminars for so many years has not only been a great
service to the community but also had a big impact on the evolvement of science
dedicated to this field. The philosophy of the seminar has always been to encourage
critical discussions on mathematical aspects of fuzzy set theory by bringing together
researchers from different fields. Those, sometimes even controversial, discussions
have helped to develop a common understanding of the treatment of fuzzy sets and
fuzzy logic. And it happened more than once, in particular after the fall of the Iron
Curtain, that researchers from different countries first met in person on the occasion
of one of the Linz seminars.
It has also been during the early 1990s that Peter Klement founded the Fuzzy
Logic Laboratorium Linz Hagenberg (FLLL). Through industrial, applied and basic
research projects, he provided a working place and research perspectives for
(young) colleagues from different countries and disciplines. As the head of the
FLLL, and also the Department of Knowledge-Based Mathematical Systems at
Johannes Kepler University, he has hosted numerous international researchers
within, but also independently of, many research actions such as CEEPUS and
COST encouraging again discussions on the theory and application of fuzzy set
theory and beyond.
Besides his activities for the scientific community, Peter Klement has always
been active as a researcher himself. His early research interests have mainly been
devoted to (fuzzy) measures and integrals as several of his articles from the 1980s

vii
viii Preface

prove. Fuzzy measures have also been the background of his close collaboration
with the Dan Butnariu leading to the publication of the joint monograph entitled
“Triangular Norm-Based Measures and Games with Fuzzy Coalitions” published in
1993.
In 1992, as a result of the first visit of Radko Mesiar and Endre Pap to Linz with
an original intention to work on fuzzy measures and integrals, another long time
research cooperation, namely on the triangular norms and triangular conorms was
established resulting in a lot of journal articles, but in particular in the publication
of the joint monograph entitled “Triangular Norms” in 2000. By the intensified
work on triangular norms, Peter Klement’s attention had also been drawn to
copulas so that since the early years of the new millennium also copulas and
quasi-copulas had appeared in the titles of his articles, as well as topics related
to aggregation functions leading back to his original research interests in
(generalized) integration.
It is therefore not by chance that the current edited volume reflects and covers
several aspects of Peter Klement’s research acitivities. Among the authors one can
find former Ph.D. students, former colleagues from the Department of
Knowledge-Based Mathematical Systems, as well as colleagues, co-authors, friends
of Peter Klement with more than 30 years of experience in fuzzy set theory. Some
of the chapters included reflect personal views on traditional topics of the Linz
seminar—some of which containing even controversial aspects and fostering a
discussion on the mathematics behind. Other chapters deal with deep mathematical
theory of the algebraic and logical foundations of fuzzy set theory and fuzzy logic.
Several chapters approach topics related to Peter Klement’s personal research
interests in copulas, measures and integrals, as well as aggregation problems.

We briefly summarize the single chapters included in this volume:

Siegfried Gottwald has contributed to chapter discussing the main developments
in the field of mathematically oriented fuzzy logics and how they found their
representation over the years in the Linz Seminars on Fuzzy Set Theory. Let us
acknowledge that Siegfried Gottwald had been a regular and active participant to
the Linz Seminars since 1990 and we are deeply sorrow that he passed away before
the finalization of this edited volume.
Enric Trillas has provided a very individual view on fuzzy sets and their personal
and scientific perception since their introduction by Lotfi A. Zadeh in his seminal
paper in 1965.
In his contribution “Modules in the Category Sup” Ulrich Höhle explains basic
properties of left modules on unital quantales with perspectives towards fuzzy set
theory and contributes to the clarification of mathematical, in particular the alge-
braic, basis of fuzzy set theory inside mathematics.
Daniele Mundici elaborates in his chapter a geometric approach to MV-algebras
and relates algebraic aspects to the basis of fuzzy resp. many-valued logics.
Francesc Esteva and Lluis Godo discuss the equational characterization of
continuous t-norms being an indispensable tool for modelling the semantic inter-
pretation of the intersection in fuzzy logics in narrow sense.
Preface ix

Also Thomas Vetterlein and Milan Petrík focus on the semantics of fuzzy logics
by discussing two different ways of investigating totally ordered monoids as an
interpretation of the conjunction in fuzzy logics.
Andrea Mesiarová-Zemánková’s chapter provides a characterization of the
structure of uninorms with continuous diagonal functions. Uninorms may be seen
as generalizations of t-norms and t-conorms, as they are associative and commu-
tative increasing operations on the unit interval whose neutral element can be, in
contrast to t-norms, respectively t-conorms, any interior element of the unit interval
and allow to model also bipolar behaviour in aggregation problems.
Humberto Bustince, Edurne Barrenechea, Miguel Pagola and Javier Fernandez
provide an overview on concepts of overlap and grouping functions generalizing
ideas of connectives from fuzzy set theory for the aggregation of information in
fuzzy classification systems.
Fabrizio Durante and Elisa Perrone in their chapter focus on asymmetric copulas
and their application in the design of experiments. The importance of copulas stems
from Sklar’s theorem clarifying that the dependence of a multivariate distribution
function of its univariate marginal distributions is, in case of continuity, completely
captured by a unique copula. The asymmetry of a copula therefore reflects the
non-exchangeability of the underlying random variables.
Carlo Sempi elaborates in his chapter the relationship between copulas and
stochastic processes, in particular the Brownian motion.
Anna Kolesárová and Andrea Stupňanová discuss extensions of capacities to
n-ary aggregation functions with relationships to the discrete Choquet and Sugeno
integral stressing the role of n-ary copulas when generalizing Lovász and Owen
extensions.
Ronald R. Yager approaches a more recent problem in aggregation, namely the
problem of multi-source information fusion by using measure representations. The
concepts of assurance and opportunity in the measure framework are also discussed.
Michel Grabisch focusses on bases and transforms of set functions on a finite set.
The basic duality between bases and invertible linear transforms is established,
covering, among others, the case of the Moebius transform, the Fourier transform
and interaction transforms.
Siegfried Weber in his chapter deals with conditioning for Boolean subsets,
indicator functions and fuzzy subsets. It introduces and discusses two types of
iteration.
Endre Pap discusses the integration of multivalued functions from additive to
arbitrary non-negative set functions. In particular, a set-valued Gould-type integral
of multifunctions is introduced and discussed.
Our special thanks go to our authors for their willingness to contribute to this
comprehensive volume. And we hope that the readers will enjoy reading all or part
of the chapters.
We congratulate Peter Klement for his scientific achievements and we are
thankful for the support he has given to us and to the scientific community in fuzzy
set theory throughout so many years. We are happy to witness that, although being
retired from being a university professor, he still enjoys being an active researcher.
x Preface

We wish him all the best, in particular healthiness, for pursuing his goals in the
future.
We have been supported by our universities, the Johannes Kepler University in
Linz and the Slovak University of Technology in Bratislava. We also gratefully
acknowledge the support of the grants APVV-14-0013 and the support in the
framework of the Technologie-Transfer-Förderung Wi-2014-200710/3KX/Kai
of the Upper Austrian Government, as well as the encouragement and the help
of Prof. Janusz Kacprzyk for the preparation of this edited volume.

Linz, Bratislava Susanne Saminger-Platz

November 2015 Radko Mesiar
Contents

Fuzzy Logic and the Linz Seminar: Themes and Some Personal
Reminiscences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Siegfried Gottwald
How I Saw, and How I See Fuzzy Sets. . . . . . . . . . . . . . . . . . . . . . . . . 13
Enric Trillas
Modules in the Category Sup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Ulrich Höhle
A Geometric Approach to MV-Algebras . . . . . . . . . . . . . . . . . . . . . . . 57
Daniele Mundici
On the Equational Characterization of Continuous t-Norms . . . . . . . . . 71
Francesc Esteva and Lluís Godo
The Semantics of Fuzzy Logics: Two Approaches
to Finite Tomonoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Thomas Vetterlein and Milan Petrík
Structure of Uninorms with Continuous Diagonal Functions . . . . . . . . . 109
Andrea Mesiarová-Zemánková
The Notions of Overlap and Grouping Functions . . . . . . . . . . . . . . . . . 137
Humberto Bustince, Edurne Barrenechea, Miguel Pagola
and Javier Fernandez
Asymmetric Copulas and Their Application in Design
of Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Fabrizio Durante and Elisa Perrone
Copulæ of Processes Related to the Brownian Motion:
A Brief Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Carlo Sempi

xi
xii Contents

Extensions of Capacities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

Anna Kolesárová and Andrea Stupňanová
Multi-source Information Fusion Using Measure Representations . . . . . 199
Ronald R. Yager
Bases and Transforms of Set Functions . . . . . . . . . . . . . . . . . . . . . . . . 215
Michel Grabisch
Conditioning for Boolean Subsets, Indicator Functions
and Fuzzy Subsets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Siegfried Weber
Multivalued Functions Integration: from Additive to Arbitrary
Non-negative Set Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Endre Pap

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

Fuzzy Logic and the Linz Seminar:
Themes and Some Personal Reminiscences

Siegfried Gottwald

Abstract The paper discusses concisely the main developments in the field of
mathematically oriented fuzzy logics and how they found their representation over
the years in the Linz Seminars on Fuzzy Set Theory.

1 Introduction

The Linz Seminar on Fuzzy Set Theory, first organized by Peter Klement in 1979,
soon became famous among mathematicians interested in fuzzy set topics, pure as
well as applied ones.
Personally, I first met Peter Klement in 1983 at the Polish Symposium on Interval
& Fuzzy Mathematics [1] in Poznań, and then again in 1985 at an International
Workshop on Fuzzy Sets Applications [2] in Eisenach. We soon had a quite friendly
relationship—and he told me that I’d have a standing invitation to the Linz Seminar as
soon as I was able to “tunnel” the iron curtain. This, however, proved to be impossible
till 1989. So I first attended the Linz Seminar in 1990—and ever since with only a
few exceptions. And already in 1990 Peter asked me kindly to join the Program
Committee, what I accepted with pleasure.

2 Logic and Fuzzy Sets—The Early Years

As I had graduated in 1969 at Leipzig University under the supervision of Dieter

Klaua, from the very beginning it was clear to me that fuzzy sets should be dis-
cussed within the framework of many-valued, particularly Łukasiewicz-like logics.
Throughout the 1970s I did this, particularly in my Habilitation Thesis (published

S. Gottwald (B)
Universität Leipzig, Abt. Logik Am Institut Für Philosophie, Beethovenstr. 15,
04107 Leipzig, Germany
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 1

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_1
2 S. Gottwald

as) [3–5], using (only) Łukasiewicz logic—but with a strong feeling for the fact that
the whole approach would naturally work within a more general logical framework.
It was the contact and cooperation with E. Czogała from Gliwice in the beginning
1980s, and soon also with his graduate student W. Pedrycz, and particularly a stay
of W. Pedrycz at Delft University of Technology, that brought to me the information
of the t-norm framework. And it was this cooperation together with a 1983 Summer
School on fuzzy set matters in Bulgaria where I learned much about the importance
of fuzzy relation equations for fuzzy modeling, and of the solution approach by
E. Sanchez.
As a result I started to consider t-norm based logics [6] and fuzzy set theory in
this realm [7], and I generalized also solvability considerations for systems of fuzzy
relation equations to this framework [8, 9].
From the Linz circle then I learned that the Linz Seminar was at its early meetings
instrumental in establishing the fundamental role of the t-norms as suitable candidates
for generalized, interactive intersections of fuzzy sets, and hence of generalized,
non-idempotent conjunction connectives for systems of many-valued logics. Core
members of the Linz Seminar crew had contributed to the discussions around t-
norms, e.g. at the 2nd Linz Seminar of 1980 with t-norm related contributions by
D. Dubois and E.P. Klement.
Additionally, what I learned from the Linz Seminars in the early 1990s was the
central role that algebraic considerations, even category theoretic ones, should play
in the fuzzy sets context to structure the theory in a suitable way. And I realized
with pleasure the open-mindedness of most of the Linz participants regarding a large
diversity of mathematical topics.
Besides the t-norm topic, themes from formal logic have been considered only
occasionally in the early years of the Linz Seminar: U. Cerutti discussed in 1981
category theoretic aspects, L. Valverde problems of generalized connectives in 1981
and 1982. Logical operators had also been the topic of L. Kohout in 1984 and 1985.
A whole day was devoted to logic, for the first time, in 1988, with U. Höhle and
L. Kohout as speakers.
As times went on and the mathematical discussions in the fuzzy sets field became
more mature, the Linz Seminar was for the first time mainly devoted to a particular
topic in 1988: Measures and Integrals, followed by the topic of Applications of
Category Theory to Fuzzy Subsets in 1989.

3 A New Time—Open Borders in Europe

With the breakdown of the iron curtain in 1989 immediately a much wider audience
was reached by and interested in the Linz Seminar, and new funding possibilities
became available which Peter Klement was able to use in a quite effective manner.
Devoted mainly to Applications of Logical and Algebraic Aspects of Fuzzy Rela-
tions the 1990 seminar considerably enlarged its number of participants and joined
researchers from the former East and West. With my personal focus on logic, let me
Fuzzy Logic and the Linz Seminar: Themes and Some Personal Reminiscences 3

mention F. Esteva, D. Mundici, and V. Novák, who for the first time attended the
Linz Seminar in 1990.
Since, logical aspects of fuzzy sets and mathematical aspects of suitable non-
classical logics have quite regularly been in the focus of the Linz Seminar. This
was e.g. the case of the Linz Seminars of 1992 on Non-Classical Logics and Their
Applications, of 1996 on Fuzzy Sets, Logics, and Artificial Intelligence, and again of
1997 on Enriched Lattice Structures for Many-Valued and Fuzzy Logics. These last
two seminars did have P. Hájek among the participants—who, in those years, was
in the main phase of his research for the logic BL of continuous t-norms presented
1998 in the seminal monograph [10].
This series of logic related Linz Seminars continued in 2000 with the topic of
Mathematical Aspects of Non-Classical Logics and Fuzzy Inference, in 2005 with
Fuzzy Logics and Related Structures, in 2010 with Lattice-Valued Logic and its
Applications, and in 2014 with Graded Logical Approaches and their Applications.
But also the Linz Seminars in the years 1993, 2004, and 2009 had logic as one of the
fields of particular interest.
Thus, from the mid-1990 the Linz Seminar well mirrored the fact of increasingly
strong relationships between fuzzy logics and mathematics.

4 Core Topics of the Development

What happened in fuzzy logic in that time, from a mathematical point of view?

4.1 T-Norm Based Logics

As already mentioned, it was clear from the mid-1980s that t-norm based logics
offer a suitable framework for fuzzy set theory, and to identify the membership
degrees with the truth degrees of such logics. And the standard understanding of
the membership degrees of fuzzy sets also supported the choice of the truth degree
1 as the only designated one. There were, thus, two prominent particular cases for
such logics: the infinite-valued versions L∞ of the Łukasiewicz and G∞ of the Gödel
families of many-valued logics, cf. [11].
The context which was given by the t-norms as truth degree functions for non-
idempotent conjunction connectives and offered, however, some different pathways
for the introduction of further related connectives. Thus, t-conorms became natural
candidates to determine non-idempotent disjunction connectives. Furthermore the
fuzzy community agreed to have idempotent conjunction ∧ and disjunction ∨ con-
nectives with min, max, respectively, as truth degree functions, too.
There was, yet, no such standard choice neither for a negation nor for an impli-
cation connective. The matter is, however, not too difficult regarding candidates for
generalized, many-valued negation functions n. Standard properties should only be
4 S. Gottwald

n(0) = 1 and n(1) = 0 together with antitonicity. It seems, essentially, to be a matter

of choice whether one additionally likes to have e.g. involutiveness n(n(x)) = x, or
strict antitonicity, or continuity.
This situation triggered research which started at the one hand from de Morgan
algebras to study their algebraic properties, and on the other hand from functional
equations which reflect central logical laws of classical logic and study solutions in
the context of de Morgan algebras. This area of research was and is carried through
essentially by Spanish researchers like E. Trillas or L. Valverde, and has only occa-
sionally been presented at the Linz Seminars (e.g. in 1981, 1982, and 1987). But these
considerations have often been restricted to implication free fragments of proposi-
tional languages.
For suitably generalized implication connectives the matter proved to be more
intricate.
The context of conjunction, disjunction, and negation connectives—provided by
t-norms, t-conorms, and suitable negation functions—offers various possibilities to
define implication connectives like in classical logic. The easiest way is to define
an implication → by ϕ → ψ =d f ¬ϕ ∨ ψ with the connectives ∨, ¬ characterized
by a t-conorm and an involutive negation function, respectively. Such implication
connectives are called S-implications.1 The standard implication of the Łukasiewicz
logic L∞ has such a characterization.
But in the area of fuzzy relation equations another type of implication connective
proved to be important to describe the inclusion-maximal solutions: so called R-
implications2 characterized for a t-norm ∗ by the adjointness condition

a ∗ b ≤ c iff a ≤ b ⇒ c . (1)

The standard implication of the Gödel logic G∞ has such a characterization. In this
case one has ∗ = min and the R-implication coincides with the relative pseudo-
complement in the complete lattice ([0, 1], min, max, 0, 1).
Interestingly the standard implication of the Łukasiewicz logic L∞ is also an
R-implication. And because of the connections between R-implications and fuzzy
relation equations, those R-implications had also been used in the previously men-
tioned naive approaches toward t-norm based logics and related fuzzy set theoretic
developments [6, 9].
Nevertheless it seems to be a kind of competitive situation between t-norm based
logics with R-implications, and such ones with S-implications. And indeed, the latter
situation was studied in [12]. However, the mainstream development has focussed
on t-norm based logics with R-implications. It seems that the crucial point for this
decision was the fact that for R-implications the rule of detachment has a simple and
convincing argument for its correctness in the fact that the formula

1 This comes from the fact that in early phases of these considerations t-conorms also had been
discussed under the name “S-norms”.
2 The name derives from the algebraic operation of residuation.
Fuzzy Logic and the Linz Seminar: Themes and Some Personal Reminiscences 5

ϕ & (ϕ ⇒ ψ) ⇒ ψ (2)

is logically valid. With ⇒ read as an S-implication, however, this formula fails to be

logically valid.
The t-norm based residuated logics had in the beginning 1990s been developed
only on an intuitive level: there existed axiomatizations only for two particular cases:
Łukasiewicz logic L∞ and Gödel logic G∞ . What was additionally known was the
fact that the choice of an R-implication, i.e. the acceptance of the adjointness condi-
tion (1) forced the t-norm T involved in this condition to be left-continuous, i.e. to
have all its unary parameterizations Ta (x) = T (a, x) as left-continuous functions.
A first breakthrough came from U. Höhle who gave in [13, 14] an adequate axiom-
atization of a logic ML which was characterized by an algebraic semantics constituted
by the class of all integral commutative residuated lattice ordered monoids, and who
claimed that this should be the formalization of fuzzy logic. Important results on
these algebraic structures U. Höhle had presented at the 1992 Linz Seminar. These
integral commutative residuated lattice ordered monoids proved to be an important
specification of the commutative lattice ordered semigroups which Goguen [15] had
proposed as suitable algebraic structures for membership degrees of fuzzy sets.
In the beginning 1990s also P. Hájek got interested in the topic of logics related to
fuzzy sets. I remember that, after a colloquium talk he had given at Leipzig University,
he asked me with reference to the German language forerunner [16] of [11] whether
I had ever thought about an approach toward a product-based infinite-valued logic
similar to C.C. Chang’s approach [17] toward (a completeness proof for) Łukasiewicz
logic L∞ using MV-algebras.
Such a product logic Π , i.e. a residuated t-norm based logic with the arithmetic
product as basic t-norm, was presented in [18]. Its algebraic characterization by the
class of all product algebras proved to be prototypical for P. Hájek’s later approach
toward the logic BL of all continuous t-norms.
Because nobody saw a possibility to axiomatize residuated t-norm based logics in
general, the idea of P. Hájek was to axiomatize the common logic of all continuous
t-norms. P. Hájek’s restriction to continuous t-norms came from his conviction that
for application only continuous t-norms should be relevant.
For this approach it was substantial that an algebraic characterization of the con-
tinuity of t-norms by the divisibility condition

a ∧ b = a ∗ (a ⇒ b) (3)

was known from [14]. And it was equally important that P. Hájek restricted the class
of integral commutative residuated lattice ordered monoids to those ones which
additionally satisfied the prelinearity condition

(a ⇒ b) ∨ (b ⇒ a) = 1 , (4)

called algebraic strong de Morgan law in [14]. Since then, these structures are known
as BL-algebras.
6 S. Gottwald

In the Linz Seminars of 1996 and 1997 P. Hájek presented core ideas and results
of his approach.

4.2 Graded Notions of Consequence

Another stream in the area of fuzzy sets related logics started in 1979 with J. Pavelka’s
discussion [19] of many-valued propositional logics with graded notions of conse-
quence, i.e. of logics which allowed to consider syntactic as well as semantic conse-
quence hulls of fuzzy sets of formulas. The general context was the one offered by
Goguen’s commutative lattice ordered semigroups with residuation added.
The semantic consequence hull Cn∗|= (Σ) of a fuzzy set Σ of formulas is defined
in a rather standard way with reference to models of Σ. In this context, a [0, 1]-
evaluation e is called a model of a fuzzy set Σ of formulas iff for each formula ϕ
one has
membership degree of ϕ in Σ ≤ e(ϕ) (5)

for the truth degree e(ϕ) of ϕ under e.

Similar as in classical model theory one can connect with each [0, 1]-evaluation
e a fuzzified theory Th(e) of e in choosing for each formula ϕ the truth degree e(ϕ)
as the membership degree of ϕ in Th(e). These notions allow to define in a natural
way 
Cn∗|= (Σ) =df {Th(e) | e model of Σ} . (6)

A corresponding syntax for such a graded notion of consequence has to cope with
membership degrees. Therefore the propositional language of [19] was enriched with
truth degree constants for each possible membership/truth degree, i.e. with constants
for each real out of [0, 1]. Furthermore, derivations with a fuzzy set Σ of premisses
have to treat formulas and degrees in parallel, and so have to act inference rules.
As a result, each derivation is a derivation of a formula ϕ to some degree a, the
proof degree of ϕ from Σ for that particular derivation. Because formulas may have
different derivations, they may have different proof degrees in this context. The
provability degree prΣ (ϕ) of ϕ from Σ then is the supremum of all possible proof
degrees of ϕ from Σ.
And the syntactic consequence hull Cn∗ (Σ) of a fuzzy set Σ of formulas is
defined by the condition

membership degree of ϕ in Cn∗ (Σ) =df prΣ (ϕ) . (7)

A general completeness theorem then is the statement Cn∗|= (Σ) = Cn∗ (Σ) and
could be proved by J. Pavelka [19] only for Łukasiewicz logic L∞ as basic system.
The reason is that the completeness proof needs the continuity of the residuation
Fuzzy Logic and the Linz Seminar: Themes and Some Personal Reminiscences 7

operation, and only the Łukasiewicz t-norm and its isomorphic versions have con-
tinuous R-implications [20].
This line of research was extended to first-order logic by V. Novák [21]. He
attended the Linz Seminar, beginning in 1990, quite often and presented various of
his theoretical achievements, e.g. about model theoretic results, but also applications
e.g. to problems in theoretical linguistics like the modeling of intermediate quantifiers
[22].
Also in this first-order case a general completeness theorem Cn∗|= (Σ) = Cn∗ (Σ)
can be proved only for Łukasiewicz logic L∞ as basic system–the proof again needs
the continuity of the R-implication.
The whole problem of graded notions of consequence can also be treated more
algebraically, having in mind that for classical logic there is a strong relationship
between consequence relations and closure operators in the class of all sets of for-
mulas.
Such a treatment needs the reference to closure operators in the class of all
fuzzy sets of formulas. They are defined in the standard way, i.e. via increasing-
ness, monotonicity, and idempotency and need only the reference to the (binary)
inclusion relation for fuzzy sets. To a large extent this approach was studied by G.
Gerla [23] who attended the Linz Seminar e.g. in 1996 and presented basic ideas of
this approach.
However, this more algebraic way toward graded notions of consequence does
not give more general results: a completeness theorem results again only for the case
of Łukasiewicz logic L∞ as basic system.

4.3 Lattice-Valued Structures and Category Theory

The focus of the considerations in the field of mathematical fuzzy logics toward
residuated lattice-ordered monoids forced also other investigations into the theory
of fuzzy sets with membership degrees in such structures or in similar ones.
The approaches toward lattice-valued mathematics brought together earlier inves-
tigations on fuzzy topologies and on category theoretic treatments of fuzzy set matters
which had been main topics e.g. of the 1989 Linz Seminar on Applications of Cat-
egory Theory to Fuzzy Subsets [24]. They have repeatedly been core topics for the
Linz Seminars, e.g. in 1993, 1997, 2004, and 2010, and they somehow culminated
in the 2012 Seminar on Enriched Category Theory and Related Topics documented
in [25]. Enriched categories have hom-sets which themselves are structured, e.g. as
residuated lattices. Enriched category theory can be used to understand the underly-
ing structure of fuzzy set theory as a particular monoidal closed category. From this
point of view, the grading membership functions become enriched presheaves.
The understanding of fuzzy sets as particular presheaves works, however, also
in the simpler context of usual categories. Also this point of view was quite often
discussed in the Linz Seminar and gave rise to M-valued sets as early as in [26],
and in this context to the understanding of the graded self-identity as measuring an
extent of existence [27, 28].
8 S. Gottwald

This notion of M-valued set touches a further important problem for many-valued
and fuzzy logics: graded identities. Such an M-valued set is an ordered pair A =
(|A|, δ A ) consisting of a crisp set |A| and a graded, i.e. M-valued (local) equality
relation δ A satisfying
(E1): δ A (x, y) ≤ δ A (x, x) ∧ δ A (y, y) , strictness
(E2): δ A (x, y) = δ A (y, x) , symmetry
(E3): δ A (x, y) ∗ (δ A (y, y) → δ A (y, z)) ≤ δ A (x, z) . transitivity
The degree of self-identity δ A (x, x) is understood as degree of extent for x in A, or
also as degree of existence, as was done for the Heyting algebra valued case in [29].
The problem of graded identities is a general problem in many-valued first-order
logics. It is not yet solved and has important philosophical aspects which, from this
author’s point of view, are still quite incompletely understood.
For the case of the finitely-valued Łukasiewicz logics it had already in 1958 by
H. Thiele [30] been shown that the usual Leibniz principle of substitutivity of equals
forces the equality relation to be crisp. A weakening of that principle, however, allows
for graded identities, as shown by this author in [31], cf. also [11].
Also for fuzzy sets a graded identity defined with respect to a graded membership
predicate ε by

x ≡ y =df ∀z(zεx → zεy) & ∀z(zεy → zεx) (8)

appears quite naturally and is used e.g. in [7, 32].

But there is also a competing possibility. P. Hájek [33] offered, expanding an early
approach [34] by Th. Skolem, an “almost naive” axiomatic set theory in the realm
of Łukasiewicz logic, having unrestricted comprehension as its sole axiom schema.
This Cantor-Łukasiewicz fuzzy set theory CŁ gets two identity relations, extensional
equality ≈ and Leibniz equality = defined by

x ≈ y =df ∀z(zεx ↔ zεy) , (9)

x = y =df ∀u(xεu ↔ yεu) . (10)

From those identity relations, Leibniz equality is crisp and extensional equality really
graded.
To a large extent these results can also be generalized to the setting with the logic
MTL of all left-continuous t-norms as background logic, as discussed in [35].
This topic of graded identities occasionally occurred during the Linz Seminars,
but never was a core topic there. It is not only a mathematical problem.
Fuzzy Logic and the Linz Seminar: Themes and Some Personal Reminiscences 9

5 Concluding Remarks

Of course, it is nearly impossible to cover with such a series like that of the Linz
Seminars all the main developments. However, if one tries to find rather important
gaps, only proof theory comes into the focus. For long, proof theory–understood here
as the field of sequent and tableau calculi–for many-valued and fuzzy logics was not
a well developed and really active field. Besides some isolated research results often
restricted to finitely-valued logics, explained e.g. in [11], prior to the end of the 1990s
there had not been published important papers.
The situation changed as hypersequents, i.e. finite multisets of sequents as intro-
duced by A. Avron [36], became known as a useful tool for fuzzy logics. M. Baaz et
al. [37] indicate the first steps into this new field, and the monograph [38] documents
the amount of work done in the first decade of this century.
There are other areas, related to mathematical fuzzy logics but more application
oriented, which have not been discussed here, but have been present at the Linz
Seminars to at least some degree. Here I have in mind the field of approximate
reasoning, and particularly possibility theory, which was often discussed in Linz
Seminars devoted to more applied fields of mathematics.
Summing up and reconsidering all the 35 Linz Seminars on Fuzzy Set Theory from
1979 till 2014, all of them locally well organized by Peter Klement, it is remarkable
that most of the essential topics in the development of mathematically oriented fuzzy
logics have been presented and actively supported by those seminars.

References

1. Albrycht, J., Wiśniewski, H. (eds.): Proceedings of the Polish Symposium on Interval & Fuzzy
Mathematics, Poznań. Wydawn Politech Pozn (1983)
2. Bocklisch, S., Orlovski, S., Peschel, M., Nishiwaki, Y. (eds.): Fuzzy sets applications, method-
ological approaches, and results. In: Mathematische Forschung, vol. 30. Akademie-Verlag,
Berlin (1986)
3. Gottwald, S.: A cumulative system of fuzzy sets. In: Marek, W., Srebrny, M., Zarach, A. (eds.)
Set Theory Hierarchy Theory, Mem. Tribute A. Mostowski, Bierutowice (1975) Lecture Notes
In Mathematics, vol. 537, pp. 109–119. Springer, Berlin (1976)
4. Gottwald, S.: Set theory for fuzzy sets of higher level. Fuzzy Sets Syst. 2, 125–151 (1979)
5. Gottwald, S.: Fuzzy uniqueness of fuzzy mappings. Fuzzy Sets Syst. 3, 49–74 (1980)
6. Gottwald, S.: T -Normen und ϕ-Operatoren als Wahrheitswertfunktionen mehrwertiger Junk-
toren. In: Wechsung, G. (ed.) Frege Conference 1984 (Schwerin, 1984). Mathematical
Research, vol. 20, pp. 121–128. Akademie-Verlag, Berlin (1984)
7. Gottwald, S.: Fuzzy set theory with t-norms and ϕ-operators. In: Di Nola, A., Ventre, A.G.S.
(eds.) The Mathematics of Fuzzy Systems. Interdisciplinary Systems Research, vol. 88, pp.
143–195. TÜV Rheinland, Köln (Cologne) (1986)
8. Gottwald, S.: Generalized solvability criteria for fuzzy equations. Fuzzy Sets Syst. 17, 285–296
(1985)
9. Gottwald, S.: Characterizations of the solvability of fuzzy equations. Elektron. Informationsver-
arbeitung Kybernetik 22, 67–91 (1986)
10 S. Gottwald

10. Hájek, P.: Metamathematics of Fuzzy Logic. In: Trends in Logic, vol. 4. Kluwer Acad. Publ,
Dordrecht (1998)
11. Gottwald, S.: A Treatise on Many-Valued Logics. In: Studies in Logic and Computation, vol.
9. Research Studies Press, Baldock (2001)
12. Butnariu, D., Klement, E.P., Zafrany, S.: On triangular norm-based propositional fuzzy logics.
Fuzzy Sets Syst. 69, 241–255 (1995)
13. Höhle, U.: Monoidal logic. In: Kruse, R., Gebhard, J., Palm, R. (eds.) Fuzzy Systems in
Computer Science. Artificial Intelligence, pp. 233–243. Verlag Vieweg, Wiesbaden (1994)
14. Höhle, U.: Commutative, residuated l-monoids. In: Höhle, U., Klement, E.P. (eds.) Non-
Classical Logics and Their Applications to Fuzzy Subsets. Theory and Decision Library Series
B, vol. 32, pp. 53–106. Kluwer Acad. Publ., Dordrecht (1995)
15. Goguen, J.A.: The logic of inexact concepts. Synthese 19, 325–373 (1968–69)
16. Gottwald, S.: Mehrwertige Logik. Logica Nova. Akademie-Verlag, Berlin (1989)
17. Chang, C.C.: Algebraic analysis of many valued logics. Trans. Am. Math. Soc. 88, 476–490
(1958)
18. Hájek, P., Lluís, G., Francesc, E.: A complete many-valued logic with product-conjunction.
Arch. Math. Log. 35, 191–208 (1996)
19. Pavelka, J.: On fuzzy logic. I–III. Zeitschr. math. Logik Grundl. Math. 25, 45–52, 119–134,
447–464 (1979)
20. Jayaram, B.: On the continuity of residuals of triangular norms. Nonlinear Anal. 72, 1010–1018
(2010)
21. Novák, V.: On the syntactico-semantical completeness of first-order fuzzy logic. I: Syntax and
semantics. II: Main results. Kybernetika 26, 47–66, 134–154 (1990)
22. Novák, V.: A formal theory of intermediate quantifiers. Fuzzy Sets Syst. 159(10), 1229–1246
(2008)
23. Gerla, G.: Fuzzy logic. Mathematical Tools for Approximate Reasoning. In: Trends in Logic,
vol. 11. Kluwer Academic Publishers (2001)
24. Rodabaugh, S.E., Klement, E.P., Höhle, U. (eds.): Applications of Category Theory to Fuzzy
Subsets. Kluwer Acad. Publ, Dordrecht (1992)
25. Höhle, U., Klement, E.P. (eds.): Fuzzy Sets and Systems, vol. 256 (2014)
26. Höhle, U.: M-valued sets and sheaves over integral commutative C L-monoids. In: Rodabaugh,
S.E., et al. (eds.) Applications of Category Theory to Fuzzy Subsets, Theory and Decision
Library Series B, vol. 14, pp. 34–72. Kluwer Acad. Publ., Dordrecht (1992)
27. Höhle, U.: Many valued logic and sheaf theory. Sci. Math. Japon. 68(3), 417–433 (2008)
28. Höhle, U.: Many-valued equalities and their representations. In: Klement, E.P., Mesiar, R.
(eds.) Logical, Algebraic, Analytic, and Probabilistic Aspects of Triangular Norms, pp. 301–
319. Elsevier, Dordrecht (2005)
29. Scott, D.S.: Identity and existence in intuitionistic logic. In: Fourman, M.P., Mulvey, C.J.,
Scott, D.S. (eds.) Applications of Sheaves. Lecture Notes in Mathematics, vol. 753, pp. 660–
696. Springer, New York (1979)
30. Thiele, H.: Theorie der endlichwertigen Łukasiewiczschen Prädikatenkalküle der ersten Stufe.
Zeitschr. math. Logik Grundl. Math 4, 108–142 (1958)
31. Gottwald, S.: A generalized Łukasiewicz-style identity logic. In: de Alcantara, L.P. (ed.) Math-
ematical Logic and Formal Systems. Lecture Notes Pure Applied Mathematics, vol. 94, pp.
183–195. Marcel Dekker, New York (1985)
32. Gottwald, S.: Fuzzy Sets and Fuzzy Logic. Artificial Intelligence. Verlag Vieweg, Wiesbaden,
and Tecnea, Toulouse (1993)
33. Hájek, P.: On equality and natural numbers in Cantor-Łukasiewicz set theory. Log. J. IGPL
21(3), 91–100 (2013)
34. Skolem, Th.: Bemerkungen zum Komprehensionsaxiom. Zeitschr. math. Logik Grundl. Math.
3, 1–17 (1957)
35. Běhounek, L., Haniková, Z.: Set theory and arithmetic in fuzzy logic. In: Montagna, F. (ed.) Petr
Hájek on Mathematical Fuzzy Logic. Outstanding Contributions to Logic, vol. 6, pp. 63–89.
Springer, Switzerland (2015)
Fuzzy Logic and the Linz Seminar: Themes and Some Personal Reminiscences 11

36. Arnon, A.: Hypersequents, logical consequence and intermediate logics for concurrency. Ann.
Math. Log. AI 4, 225–248 (1991)
37. Baaz, M., Ciabattoni, A., Fermüller, C., Veith, H. (eds.): Proof theory of fuzzy logics: urquhart’s
C and related logics. In: Mathematical Foundations of Computer Science. Lecture Notes in
Computer Science, vol. 1450, pp. 203–212. Springer, Berlin (1998)
38. Metcalfe, G., Olivetti, N., Gabbay, D.: Proof theory for product logics. Neural Netw. World
13, 549–558 (2003)
How I Saw, and How I See Fuzzy Sets

Enric Trillas

Abstract This paper does not pretend a ‘technical’ presentation of a particular topic
with an exhausting list of references; it just would like to contain some reflections
of the author concerning how he sees, or better, he wishes, the future of current
fuzzy logic that, in his view and at the risk of stagnation, cannot lie on any kind of
‘logicism’ but on ‘scienticism’.

1 Introduction

It can be said that as it was originally introduced by Zadeh [16], fuzzy set theory
mainly deals with two important linguistic phenomena, imprecision and non-random
uncertainty, and that fuzzy sets can be applied, among others, to the study of dynam-
ical systems whose behavior can be described by sets of imprecise linguistic rules,
and to the random uncertainty associated to some linguistic statements [7, 13]. For
instance, the theory of possibility can deal with non-random uncertainty, fuzzy con-
trol with dynamical systems, and fuzzy probability with random fuzzy events.
The ground of fuzzy set theory lies in the, historically not surprising, fact that
predicates acting in a universe of discourse generate linguistic collectives in it; col-
lectives [14], except when they degenerate in just a single classical set, are cloudy
linguistic entities neither well known, nor easy to specify virtual or ‘thought’ entities
whose appearances, or states, are just membership functions, fuzzy sets allowing to
see their projections inside the fog of ordinary language. Hence, fuzzy sets can be
seen as a starting point for the currently non existing scientific study of linguistic
collectives. In sum and grossly speaking, fuzzy sets deal with ordinary language;
they are mathematical entities contextually reflecting collectives, and modeled by
their membership functions. They meant to pass from an old world of exact thinking

To Professor Peter Klement, with deep affection.

E. Trillas (B)
European Centre for Soft Computing, Mieres, Asturias, Spain
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 13

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_2
14 E. Trillas

represented by sets, to a new world of approximate thinking represented by them.

The future of fuzzy sets can be seen around a new mathematical study of ordinary,
or common sense, reasoning in which the central idea is, instead of ‘deducing’ from
precise premises reflecting totally known information even if not fully describing
something, to that of ‘conjecturing’ [9] from imprecise premises reflecting informa-
tion partially known and also able to reach creative conclusions. That is, to increase
the informative content of the premises or previous information; in sum, to be in touch
with creativity.
Fuzzy sets have to do with both the representation of information, and to how new
one can be obtained by just a ‘previous thinking’ as it is done always in searching
for a new aspect of a problem, and that, latter on, should be either formalized, or
checked against some reality to acquire the status of ‘new’ knowledge. Of course,
in these processes of conjecturing those of deducing, abducing, and also lucubrating
are included [11].
Without no doubt, it can also be said that the idea of fuzzy sets was born in
the ‘cultural’ neighborhood of cybernetics, where analogical computers [1] were
seriously taken into account. Fuzzy sets can be seen indeed as ‘analogical entities’
in contrast to the ‘digital crisp sets’ and, since most of the human knowledge is
essentially analogical, it is not at all surprising that fuzzy sets can be suitable for
representing, at least, expert knowledge. In fact, the first application of fuzzy sets
to the control of machines, introduced in 1972 by the late Abe Mamdani [5], can
be considered as a method for the management of imprecise expert knowledge, and
who knows if, in a future, and provided analogical quantum computers [15] were
actually constructed, fuzzy sets will not play some new role in their functioning. If
from a philosophical and scientific point of view fuzzy sets are but measures, from
a technological one they are just analogical tool constructs representing knowledge.

2 My First Steps into Fuzzy Logic

I entered into fuzzy logic by chance. It was through an interview, in a French news-
paper, with the late professor Arnold Kaufmann in which he spoke on his then recently
appeared book ‘Ensembles flous’. The subject interested me since I was doing my
research work on the probabilistic metrics introduced by Karl Menger, and knew his
paper entitled ‘Ensembles flous’, a new concept that he translated into English by
hazy sets. I bought Kaufmann’s book, read a good part of it, and to some extent I
was actually disappointed; my first glance at fuzzy sets make me to believe that they
were just a simple generalization of sets. Nevertheless, since some of the examples
in the book called upon my attention and made me curious, I decided to read the
1965 paper where fuzzy sets were originally introduced, whose title is ‘Fuzzy Sets’
and was written by Zadeh [16]. Before reading this paper, and as a consequence of
both my mathematical formation, and the reading of Menger’s paper, I was unable
to see fuzzy sets unlinked with probability; indeed, hazy sets represent something
like the probability of an element belonging to a set.
How I Saw, and How I See Fuzzy Sets 15

But the reading of Zadeh’s paper suddenly changed my view. The subject was
towards representing the multitude of imprecise predicates of which language is full;
it was, for me, the first mathematical model for taking into account the imprecision
that, permeating language, affects ordinary reasoning with concepts that are not
definable like those managed in mathematics by ‘if and only if’ conditions, but only
describable from their use in several contexts as they appear in dictionaries. It was
for me a new land to be explored, and I was captivated as I see the possibility of
building up mathematical models of common reasoning. I decide to start with such
an exploration! At the end, in that time I was unhappy with the ‘bourbakism’ of
which mathematics was full in Spain, and I was also worried by the giving up that,
from time ago, logicians kept on ordinary reasoning.
Although the only references in the first Zadeh’s paper on fuzzy sets are the purely
mathematical books by Birkhoff, Halmos, and Kleene, and since from very young
I kept a deep interest in Bertrand Russell’s philosophical writings, I remembered
the Russell’s paper ‘On Vagueness’ and believed that there should be some links
between fuzzy sets and vagueness. This idea conducted me to the 1972 paper by
Aldo De Luca and Settimo Termini, where they established the then new concept of
a ‘fuzzy entropy’, and I thought it is nothing else than a measure of the vagueness or,
by duality of its classicality, booleanity, or crispness, the linguistic label of a fuzzy
set presents. This idea make me to think that ‘fuzziness’ is just a restriction of the
vagueness of a predicate whenever it can be represented by a fuzzy set, and this was
for me a challenging philosophical idea that, many years ahead, conducted me to see
that fuzzy sets are nothing else than measures of the meaning of predicates. My first
papers on fuzzy sets dealt, between 1976 and 1978, with trying to find functionally
expressible mathematical formulas able to represent fuzzy entropies, but different
and more general than the unique logarithmic fuzzy entropy shown by De Luca and
Termini in his paper. In addition, I also tried to relate them with the Sugeno’s fuzzy
integral since, in the meantime, I was acquainted with Michio Sugeno in Toulouse,
and with his 1974 Ph.D. Thesis.
Early after these worries, I began to be interested in the subject of fuzzy connec-
tives and fuzzy inference. For the first I was essentially motivated by the fact that
fuzzy connectives can’t show the same properties in all contexts, and that distrib-
utivity is, for instance, a very constraining and crisp property. What conducted me
in such direction were the papers by Bellman and Giertz [2], and that on negation
by Lowen [4]. My dedication to Probabilistic Metric Spaces, that bring me to know
Bert Schweizer and Able Sklar after meeting Karl Menger in Chicago, introduced me
to solve Functional Equations, and I got the idea of characterizing the (continuous)
strong negations by just solving an easy functional equation. Since I was familiar
with Schweizer and Sklar’s t-norms (a restriction by adding associativity to those
introduced by Menger [6]), I can introduce in fuzzy logic these ordered semi-groups
in the unit interval. Finally, the fact that as examples of his Compositional Rule
of Inference, Zadeh did show some that were non-preserving the classical Modus
Ponens when the input is just the antecedent of the rule, I tried to study this ‘Modus’
in fuzzy logic by formulating it as Hardegree did in Orthomodular lattices [3].
16 E. Trillas

To end this section, that corresponds with the time in which I met for the first
time professor Peter Klement and in Barcelona, let me remember again that, as
another consequence of Menger’s work, but due to the idea of mathematically
modeling the breaking of synonymous chains, I introduced in fuzzy logic the
T-indistinguishabilities, or T-equivalences, allowing to relating such problem with
that of Poincaré concerning the physical continuum. I would like to say that Menger’s
trace in fuzzy logic or, at least, in my contributions to it, is certainly of some relevance.

3 Zadeh’s Fuzzy Sets Are but Measures of Meaning

For a lot of time after 1965, the mathematical nature of fuzzy sets in relation with
the meaning of their linguistic label, was not clearly explained. They were simply
viewed as membership functions generalizing the characteristic function of crisp sets
and, supposedly, representing its meaning in the universe of discourse but without
counting with a meaning’s operational description [8]. If philosophers largely debated
on the meaning of ‘meaning’, they never attended the representation of meaning, and
it lacked a scientific study that today can be considered started with the work of Zadeh,
and in a form close to the Wittgenstein of the ‘Philosophical Investigations’, when
he states that almost always ‘the meaning of a word is its use in language’. How
can even if not defined, the use or management in language of a linguistic label be
mathematically described?
If P is a linguistic label, or predicate, acting in a universe of discourse X through
the elemental statements ‘x is P’, for a suitably management of P the two binary
relations in X , that empirically come from linguistic perception, from its use,
• x = P y ⇔ x shows the property named P equally than y shows it ⇔ x is equally
P than y,
• x ≤ P y ⇔ x is less P than y,
should be known [2, 14]. When both relations ≤ P and = P do coincide, it is said that
the use of P in X is precise, rigid, or crisp, = P is an equivalence, and X is partitioned
in the equivalence classes in the quotient set X/ =, [x] = {y ∈ X ; y = P x}. Instead
and when ≤ P == P that, provided it can be supposed = P =≤ P ∩ ≤−1 P , implies that it
is not ≤ P ⊆ ≤−1
P , it is said that the use of P in X is imprecise, flexible, or fuzzy. In any
case, the graph (X, ≤ P ) represents the qualitative, or primary, meaning of P in X .
In this way, the previously amorphous universe of discourse X , is softly structured
thanks to the use of P in it. The simple and usual act of ‘speaking’ on a property
recognizable in the elements of X , endows X with the arcs of this graph; an idea
corresponding with the intuitive one that rational speech tries to introduce some kind
of ‘ordering’ in the universe of discourse, also corresponding to the establishment of
some necessary link between ordering and understanding. Nevertheless, the graph
does not exhaust the ‘full meaning’ of P in X , and when it is ≤ P = ∅, it can be said
that P is metaphysically used in X , that P is metaphysical, or meaningless in X
[11]. Notice that it is thanks to the relation ≤ P that can be seen the variability of the
How I Saw, and How I See Fuzzy Sets 17

property named P along the elements of X ; that it is ≤ P == P , is what permits to

say that the use of P is imprecise in X .
If P is not metaphysically used in X , that is, if ≤ P = ∅, then a measure of the
extent of P in X , is a mapping µ P : X → [0, 1], such that

(1) x ≤ P y ⇒ µ P (x) ≤ µ P (y)

(2) z maximal for ≤ P ⇒ µ P (z) = 1
(3) z minimal for ≤ P ⇒ µ P (z) = 0.

Once the graph (X, ≤ P ) is known it can be said that P is measurable in X , and once
a measure µ P is known that it is effectively measurable in X [14].
The three former properties are not sufficient, in general, to specify a measure,
but there is only a single one if the predicate is precise; to specify a measure either
more information on the use of P, or to establish a reasonable hypothesis on it, is
necessary.
In any case, each measure µ P is the membership function of a fuzzy set in X
labeled P. Fuzzy sets are defined by the measures of the extent up to which the
elements in X are P show the property named P; shortly speaking it can be said
that fuzzy sets are measures of meaning, like probabilities are measures of random
uncertainty, and fuzzy entropies are measures of fuzziness. It should be noticed that
each quantity (X, ≤ P , µ P ) represents a good enough knowledge on the meaning of
P in X for its scientific consideration; it can be said that such quantities are the
typically scientific domestication of meaning [10], and can offer a new perspective
for studying both fuzzy sets and fuzzy logic.
It should be noticed that if the use of P in X is rigid, it is x = P y ⇔ x ≤ P y &
y ≤ P x ⇒ µ P (x) = µ P (y), and hence, µ P is constant in the classes modulo P, the
only values µ P can take are 0 or 1, µ−1 P (1) is the crisp subset specified by P in X ,
µ−1P (0) its classical complement, and one of them can be empty.
In praxis, a fuzzy set is designed by means of the information on its linguistic
label that is available and that, most of the times, is not the full relation ≤ P , but a
part of it; there are cases in which obtaining ≤ P can be very difficult. Hence and very
often, neither it is always ≤ P completely known, nor it can be stated that the designed
membership function µ∗P is truly a measure, but some unknown approximation to it.
Consequently, the designer cannot work with ≤ P but only with the total order defined
by x ≤µ∗P y ⇔ µ∗P (x) ≤ µ∗P (y), called the working meaning of P in X . Provided
µ∗P were actually a measure or, at least, it can be supposed it verifies property (1),
and since then, x ≤ P y ⇒ µ∗P (x) ≤ µ∗P (y) ⇔ x ≤µ∗P y, implies ≤ P ⊆ ≤µ∗P , that is,
the working meaning extends the qualitative meaning of P. The act of measuring P,
modifies its qualitative meaning by adding more arcs to it [14].
Notice that since in most cases the relation ≤ P has not a total, or linear, character,
it cannot coincide with the linear orders ≤µ P . When there is coincidence, it is said
that the measure perfectly reflects the qualitative meaning of P. It is easy to proof
that, provided ≤ P is reflexive and transitive, then = P is an equivalence relation,
and that the mapping C : X → X/ = P , assigning to each x the equivalence class
[x] = C(x), verifies x ≤ P y ⇔ [x] ≤∗P [y] ⇔ C(x) ≤∗P C(y). Hence, not only C
18 E. Trillas

perfectly reflects the qualitative meaning of P in X , but this idea opens the door for
defining ‘qualitative measures’ of a predicate by taking, instead of the unit interval,
some non-numerical posets with which more possibilities of perfectly reflecting the
qualitative meaning could appear.

4 On the Other Types of Fuzzy Sets

In those cases in which the measure does not perfectly reflect the qualitative meaning,
and since in science is not at all rare to manage measures with complex values, it could
be suitable to substitute the real interval [0, 1] by the complex one {a + bi; a, b ∈
[0, 1]}, the complex circle, endowed with the usual partial order a1 + b1 i ≤ a2 +
b2 i ⇔ a1 ≤ a2 & b1 ≤ b1 , and with analogous properties [14] to the former (1),
(2), and (3). This substitution cannot guarantee that a complex-valued measure will
perfectly reflect the qualitative meaning, but just that it can offer more possibilities for
it, since the working order will be not linear. This substitution that can be equivalently
seen by taking an interval-valued measure, just changing a + bi by the interval [a, b],
and corresponding to a particular type of the so-called type-2 fuzzy sets reflecting
that the value of the measure carries with the uncertainty coming from only being
sure that it is in the interval [a, b].
Analogously, and instead of the real or the complex unit intervals, it can be taken
the set [0, 1][0,1] , of the fuzzy sets in the unit interval (type-2 fuzzy sets) that contains
images isomorphic to both the unit interval and the complex unit interval, and for
those cases in which the only that can be asserted is that the value of the measure is,
for instance, either ‘around 0.7’, or ‘high’ [12]. In this form, all the types of fuzzy
sets currently considered, are integrated thanks to the quantities, either numerical or
functional, representing the meaning of its linguistic label.
The full meaning of a linguistic label P is not unique, but it is actually context-
dependent and purpose-driven. Each quantity (X, ≤ P , µ P ), real, complex or fuzzy
valued, is obtained through what the designer can know, in a given context, of the
use, action or behavior of P in X , or through some reasonable hypothesis he could
be able to make on such behavior. This last is the often considered case in most
applications, in which the real-valued measure, the membership function, is supposed
to be trapezoidal, or just triangular.
Once seen that the membership functions of fuzzy sets mean nothing else than a
‘measure of the meaning’ of its linguistic label, it can be remembered the famous
words of Lord Kelvin shortened to ‘If you cannot measure it, it is not science’.
There are, notwithstanding and at least, two aspects introducing important differences
between Lord Kelvin’s times and ours. In the first place, it is the fact that if, lets it
say, science is essentially concerned with matter and energy, fuzzy set theory is
concerned with knowledge and information, and directly related with the so called
Information Technologies. In a second place, in Lord Kelvin’s science there were
and are known systematic procedures and laboratory methods, to measure the basic
parameters of the studied things, but now and for what concerns, for instance, the
How I Saw, and How I See Fuzzy Sets 19

design of membership functions, the situation is different and more linked to some
analogy with virtual objects, than with physically real objects. It is not the same to
study the chemical composition of an organic product, or the movement of a star, than
to study the meaning of a written piece, or the control of a machine whose behavior
is known by the knowledge of the experts in their functioning and once linguistically
described. Nevertheless, this is the kind of problems currently worrying Artificial
Intelligence.

5 The Evoluation of Fuzzy Logic

Anyway, the evolution of fuzzy logic towards Zadeh’s Computing with Words and
Perceptions, CwW for short [17], is conducting towards the mathematical represen-
tation of statements larger and more complex than the more or less simple rules
considered in fuzzy control [8]. This will mean to face with the necessity of con-
sidering different ways of expressing conditional statements, and the already known
linguistic connectives ‘and’, ‘or’, ‘not’, etc., since there is not a universal form of
expressing them in language, like it is in classical logic and set theory, but respectively
represented in fuzzy logic by residuated implications, S-implications, conjunctive
implications, t-norms, t-conorms, negation functions, etc. For all that there are a lot
of mathematical models facilitating to fuzzy logic a remarkable armamentarium for
the representation of statements, and for doing deductive inferences with them, but
what is not yet clear enough are the linguistic subjects to which such armamentarium
is applicable, and to which is not. For instance, fuzzy logic only considers function-
ally expressible connectives, but no suitable criteria are known for recognizing this
hypothesis in concrete cases, and, analogously, the use of non strong but continuous
negations is not yet spread into fuzzy logic applications to represent language. Even
more, almost always the used connectives are min, max, prod, and 1-id; there exists
a big separation between what is employed by practitioners of fuzzy logic, and what
is kept in the theoretical armamentarium generated by mathematicians.
In sum, it seems that fuzzy logic is approaching the time in which it should face a
turning point. The great subjects fuzzy logic deals with are linguistic imprecision and
non-random uncertainty, not to say anything on the very important but scientifically
almost pending subjects of ambiguity, the presence of multiple meanings, and com-
mon sense non-deductive reasoning [11] with imprecise, non-randomly uncertain,
and ambiguous words.
The only way to properly afford it is, in the author’s view, the transformation of
fuzzy logic in a kind of ‘physics’ of imprecision, non-random uncertainty and ambi-
guity. That is, in a new experimental science that, based in Natural Language, can
count with mathematical models able to give important parameters to be experimen-
tally computed at each case once their can be found in the same study of language, and
not by abstract mathematical thought considerations. What is needed is to transform
the study of language from a logic one in a scientific one.
20 E. Trillas

When fuzzy logic was initially developed in the past Century’s seventies and
eighties, almost the only back referents for its study were classical and multiple-
valued logics, but now it should be centered in Natural Language. If current fuzzy
logic already meant an important progress in the way asked by John von Neumann
of introducing mathematical analysis in the study of those subjects without a just
‘yes’ or ‘not’ hypothesis for its validity, it can be the right moment to go a step ahead
and turning towards the Artificial Intelligence’s ‘Gordian Knot’ of trying to reach
computers thinking like people usually do.

6 Conclusion

Up to some point, and although many papers of a mathematical character, even with
some of them of a true mathematical quality, are being continuously published in
the setting of ‘theoretic fuzzy logic’, its evolution seems to be actually stagnated
because of some moving away of what is the essence of fuzzy logic. By one side,
those papers remain practically unknown or, at least, not considered, for those who
devote their efforts to the applications of fuzzy logic, and by the other the motivation
of their authors is almost always purely abstract; in them, it rarely appears a ‘real
fuzzy problem’ to which either their results could be applied to, or just it can be
suggested by the paper’s content. It seems as if in current fuzzy logic it were two
streams, that of mathematicians and that of engineers, but fuzzy logic should be an
integrated study of what is ‘fuzzy’ and, in principle, that practitioners ignore the
obtained mathematical results, marks a limit in their capability of designing fuzzy
systems. There is, perhaps, some kind of isolation between both types of researchers,
the most relevant of the ones not mixed with the most relevant of the others. This and
to some extent, goes against the cross-fertilization of both groups and can contribute
to the closing of the first in their own mathematical interest. Anyway, and in the
last years, I hopefully heard on mixed groups working in some specific projects.
Notwithstanding, and as far as I know, such projects are on very specific topics not
directly related with CwW.
In the author’s view, and by looking at what is fuzzy and what is for its study, the
great challenge for the best continuation of theoretic fuzzy logic lies in the problems
that are in the back of the new Zadeh’s ‘Computing with Words and Perceptions’,
where the problems that were essential for the introduction of fuzzy sets could acquire
all their relevance as clearly dealing with Natural Language’s complex phrases, and
with the non-deductive varieties of Commonsense Reasoning. Nevertheless, the nat-
ural and dynamic characters of both language and reasoning, seems to suggest that
a new, and scientific, study of them cannot be completely afforded by only counting
with the abstract reasoning reached through mathematical theorems that only can be
successfully applied provided all what is being supposed for their proofs is actually
verified in a concrete and actual situation. Something that is, usually, very difficult
to check as, it happens, for instance, when trying to use an S-implication function
How I Saw, and How I See Fuzzy Sets 21

for linguistic rules in which the representation of the negation of their antecedents is
actually unknown.
It is in the thought of the author that the main subjects of ‘fuzzy logic’ or, by
extension CwW, are both the representation and technical management of the impre-
cision and the uncertainty pervading natural language and commonsense reasoning
in non-trivial statements. In some cases, for instance, the meaning of the components
of a large statement is only captured after having captured the full meaning of the
full statement, something different of what is done in logic where always it is done
by departing from the meaning of the components.
To afford those subjects it seems recommendable to face them as they are, natural
phenomena of which, and in addition, we have a scarce knowledge that, notwith-
standing, should be increased by the only way it can be followed for any natural
phenomena, namely, by experimenting in controlled forms as it is typical of science.
It is with the conjunction of experimentation and mathematical modeling how mea-
surable parameters can be obtained and deep conclusions attained. Science always
needs to count with suitable frames for representing what it deals with, thanks to
which some mathematical models could be established and that, at its turn, facili-
tates some numerical parameters necessary to going on with more experimentation.
À la Popper, research is always an un-ended quest.
A new experimental science dealing with linguistic imprecision and uncertainty,
both random and not random, seems to appear in the horizon and into the complex
knitting of language. It is an enterprise that jointly with, and close to, fully knowing
the brain’s functioning, could contribute to capture what is rationality by going
far from old metaphysical ideas, and by means of the single way mankind has for
acquiring safe knowledge, the scientific method.
Would young researchers in the XXI Century devote their efforts to such a chal-
lenging enterprise!

Acknowledgments This paper is partially funded by the ‘Foundation for the Advancement of Soft
Computing’, Mieres (Asturias), Spain.

References

1. Basáñez, L., Batle, N., Ferraté, G., Grané, J., Trillas, E.: A first approach to sigma-transform.
J. Math. Anal. Appl. 92(1), 224–233 (1983)
2. Bellman, R., Giertz, M.: On the analytic formalism of fuzzy sets. Inf. Sci. 5, 149–156 (1973)
3. Hardegree, G.M.: The conditional in quantum logic. Synthese 29, 63–80 (1974)
4. Lowen, R.: On fuzzy complements. Inf. Sci. 14, 107–113 (1978)
5. Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic controller.
Int. J. Hum. Comput. Stud. 7(1), 1–13 (1975)
6. Menger, K.: Statistical metries. Proc. Nat. Acad. Sci. USA 28(12), 535–537 (1942)
7. Nguyen, H.T., Walker, E.A.: A First Course in Fuzzy Logic. Chapman & Hall, Boca Raton
(2000)
8. Trillas, E., Guadarrama, S.: Fuzzy representations need a careful design. Int. J. Gen. Sys. 39(3),
329–346 (2010)
22 E. Trillas

9. Trillas, E.: A model for ‘crisp reasoning’ with fuzzy sets. Int. J. Intell. Syst. 27, 859–872 (2012)
10. Trillas, E.: En defensa del razonamiento creativo, Universidad Pública de Navarra (2014)
11. Trillas, E.: Razonamiento; significado, incertidumbre y borrosidad; Ed. Upna, Pamplona (2015)
12. Trillas, E.: An algebraic model of reasoning to support Zadeh’s CwW (2015)
13. Trillas, E., Eciolaza, L.: Fuzzy logic: an introductory course for engineering students. Springer
(2015)
14. Trillas, E., Termini, S., Moraga, C.: A naïve way of looking at fuzzy sets. Fuzzy Sets Syst. In
press (2015)
15. Williams, C.P., Clearwater, S.H.: Explorations in Quantum Computing. Springer, NY (1988)
16. Zadeh, L.A.: Fuzzy sets. Inf. Control 8, 338–353 (1965)
17. Zadeh, L.A.: Computing with Words: Principal Concepts and Ideas. Springer, Berlin (2012)
Modules in the Category Sup

Ulrich Höhle

Abstract This chapter explains basic properties of left modules on unital quantales
with the perspective towards fuzzy set theory. Typical constructions such as the
fuzzy power set, Zadeh’s forward operator or binary operations defined according
to Zadeh’s extension principle are constructions in the symmetric monoidal closed
category of complete lattices and join preserving maps. Moreover, involutive left
modules play a significant role in the representation theory of C ∗ -algebras.

1 Introduction

The motivation of this chapter is to make a contribution to the mathematical founda-

tions of fuzzy set theory and to describe the place where fuzzy set theory is residing
inside mathematics. Let Sup be the category of complete lattices and join preserv-
ing maps. Our thesis is that module theory in Sup is the algebraic basis of fuzzy set
theory. We justify this thesis by the following observations:
• In Zadeh’s pioneering paper on fuzzy sets (cf. [30]), the real unit interval provided
with the bounded sum appears as underlying mathematical structure. Since the
bounded sum is the t-conorm of the Łukasiewicz arithmetic conjunction,

α ∗ β = max(α + β − 1, 0), α, β ∈ [0, 1],

L.A. Zadeh has tacitly used the unital quantale given by the canonical M V -algebra
as algebraic basis.
• Let Q be a unital quantale and X be a set. Then the Q-valued (i.e. Q-fuzzy) power
set of X is the free left Q-module generated by X . This result appears for the
first time in [12] and describes the fuzzy power set by a universal property in the
language of module theory in Sup.

U. Höhle (B)
Fachbereich C Mathematik und Naturwissenschaften, Bergische Universität,
Wuppertal, Germany
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 23

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_3
24 U. Höhle

• Left Q-modules and complete Q-valued (i.e. Q-fuzzy) ordered sets are equivalent
concepts—a result which goes back to I. Stubbe in a more general context given
by quantaloid enriched categories (cf. [29]).
• Binary operations constructed according to Zadeh’s extension principle are the
tensor product of the respective Minkowski multiplications with the multiplication
of the underlying unital quantale (cf. [8]).
Because of the previous observations we believe that it is meaningful to put together
various important aspects of module theory in Sup with the perspective towards
fuzzy set theory. In this sense this paper is a survey on some basic properties of left
modules on unital quantales.
We begin with the construction of the tensor product in Sup which goes back
to Z. Shmuely 1974 (cf. [27]). Since there exist various approaches to the tensor
product (cf. [3, 12]), we prefer here the understanding of the tensor product as
“function space”—i.e. as a complete lattice of join reversing maps (cf. Proposition
1 on p. 6 in [12]). Subsequently, we view unital quantales as monoids in Sup and
develop the usual module theory in Sup including complete many-valued ordered
sets which represent a typical phenomenon of left-modules in Sup. In this context
it is interesting to realize that involutive left modules play a significant role in the
representation theory of C ∗ -algebras (cf. [13, 14, 16, 24]).

2 The Tensor Product in the Category Sup

Let Sup be the category of complete lattices and arbitrary join preserving maps. It
is well known that Sup is a complete and cocomplete category. In particular, regular
quotients of complete lattices L can be identified with closure operators on L—these
c
are isotone self-maps L − → L provided with the properties 1 L ≤ c and c ◦ c ≤ c.
With regard to algebraic considerations the most important property of Sup is the
existence of a tensor product transforming Sup into a symmetric monoidal closed
category (cf. [17]). The purpose of this section is to recall these fundamental proper-
ties and to explain the role of the tensor product for the construction of fuzzy power
sets.
f
Let L and M be complete lattices.
 Amap L − → M is join reversing if for all
subsets A of L the relation f ( A) = f (A) holds. Obviously, join reversing
f
maps L − → M are always part of a Galois connection ( f, g) between L and M (cf.
g
[4, 7]) where the second polarity M −
→ L is determined by

g(m) = { ∈ L | m ≤ f ()} (1)

f
Therefore we denote the set of all join reserving maps L −
→ M by G(L , M) and
consider the following partial order on G(L , M):
Modules in the Category Sup 25

f1 ≤ f2 ⇔ ∀ ∈ L : f 1 () ≤ f 2 (). (2)

Obviously, (G(L , M), ≤) is a complete lattice in which meets (but in general not
joins) are computed pointwisely. In this context, the formula (1) determines an order
isomorphism between G(L , M) and G(M, L).
b
Let N be a further complete lattice. Referring to [3] a map L × M −→ N is a
bimorphism (in Sup) iff b is join preserving in each variable separately—i.e. for
all m ∈ M and  ∈ L the correspondences  → b(, m) and m → b(, m) are join
preserving.

Example 1 Let L and M be complete lattices. For every pair (, m) ∈ L × M we

f (,m)
construct a join reversing map L −−−→ M as follows (cf. Lemma 1.5 in [27])
⎧ ⎫
⎪
⎪ ⊥, z ≤ , ⎪
⎪
⎨ ⎬
f (,m) (z) = m, z ≤ , z = ⊥, z ∈ L. (3)
⎪
⎪ ⎪
⎪
⎩ ⎭
, z = ⊥,

where ⊥ () denotes the respective universal lower (upper) universal bound in L

and M. Obviously, the following properties hold for all ,  ∈ L and m, m ∈ M:
• f (⊥,m) and f (,⊥) coincide with the universal lower bound in G(L , M).
• If  = ⊥ and m = ⊥, then f (,m) ≤ f ( ,m ) iff  ≤  and m ≤ m .
β
On this background we introduce a bimorphism L × M −
→ G(L , M) by:

β(, m) = f (,m) , (, m) ∈ L × M. (4)

It is evident that β is join preserving in its second argument. In order to verify that β is
also join preserving in its first argument, it is sufficient to consider a non empty subset
{i | i ∈ I } of L. Then f ∈ G(L , M) is an upper bound of A := {β(i , m) | i ∈ I }
iff m ≤ f (i ) holds
  for all i ∈ I . 
Since f is join reversing, f is an upper bound
of A iff m ≤ f i . Hence, β i , m is the smallest upper bound of A—i.e.
 
i∈I i∈I
β(i , m) = β i , m .
i∈I i∈I

Definition 1 Let L and M be complete lattices. A pair (β, X ) is called the tensor
β
product of L and M, if X is a complete lattice and L × M −
→ X is a bimorphism
such that the following universal property holds:
b
For every bimorphism L × M −
→ N there exists a unique join preserving map
h
X−
→ N making the following diagram commutative:
26 U. Höhle

β
L×M / X
??
??
? (5)
b ??
h
 
N

In the special case of Sup the next theorem is a version of the Corollary of
Proposition 5 in [3].
Theorem 1 For every pair (L , M) of complete lattices L and M the tensor product
(β, X ) exists and is unique up to an isomorphism in the sense of Sup.

Proof The uniqueness of the tensor product is an immediate corollary from its uni-
versal property.
 In order
to verify the existence of the tensor product, we consider
the pair β, G(L , M) constructed in Example 1.
b f
Let L × M − → N be a bimorphism. Since every join reversing map L −
→ M is the
join of {β(, f ()) |  ∈ L} in G(L , M)—i.e.

f = β(, f ()), (6)
∈L

h
there exists at most a join preserving map G(L , M) −
→ N making the diagram
β
L×M / G(L , M)
JJJ
JJJ (7)
JJJ h
b
$ 
N

commutative.
h
In order to establish the existence of a join preserving map G(L , M) −
→ N s.t. the
h
diagram (7) commutes, we proceed as follows. We define a map G(L , M) −
→ N by
 
h( f ) = b(, m), A ⊆ L × M, β(, m) = f. (8)
(,m)∈A (,m)∈A

Provided that we can show that h is well defined, then formula (8) shows immediately
that h is join preserving and makes the diagram (7) commutative. Therefore we only
 h is well defined. For this purpose we choose a further subset B of L × M
verify that
with β(, m) = f . Since N is complete, it is sufficient to verify that the
(,m)∈B
respective sets of upper bounds of {b(, m) | (, m) ∈ A} and {b(, m) | (, m) ∈ B}
coincide.
First we apply the property that b is join preserving in each variable separately
fz
and define a join reversing map L −→ M for all z ∈ N as follows

f z () = {m ∈ M | b(, m) ≤ z},  ∈ L .
Modules in the Category Sup 27

If z ∈ N be an upper bound of {b(, m) | (, m) ∈ A}, then m ≤ f z () holds for all
(, m) ∈ A. Since f is the join of {β(, m) | (, m) ∈ A}, we obtain f ≤ f z . Since
f is also the join of {β(, m) | (, m) ∈ B}, the relation

m ≤ f z (), (, m) ∈ B (9)

follows. Because of b(, f z ()) ≤ z we infer from (9) that z is also an upper
bound of {b(, m) | (, m) ∈ B}. Interchanging the role of A and B the assertion
follows. 

Because of the previous theorem we fix the following notation and terminology.
The tensor product of two complete lattices L and M is denoted by L ⊗ M. By abuse
β
of notation the bimorphism L × M − → L ⊗ M is also denoted by ⊗, and instead of
β(, m) we also write  ⊗ m. Elements of L ⊗ M are called tensors, and tensors of
the special type  ⊗ m are called elementary tensors. It follows immediately from (6)
that every tensor is a join of an appropriate family of elementary tensors. Sometimes,
in the case of  = ⊥ and m = ⊥, the following equivalence is useful:

 ⊗ m ≤  ⊗ m ⇐⇒  ≤  and m ≤ m .

Let L be a complete lattice. We show that the L-fuzzy power set is the tensor
product of L with the ordinary power set. For this purpose we choose a set X and
consider the partial ordering ≤ on L X defined by:

g1 ≤ g2 ⇔ ∀x ∈ X : g1 (x) ≤ g2 (x).

Then L X is a complete lattice. In the literature on fuzzy sets L X is called the L-fuzzy
or L-valued power set of X .

Theorem 2 ([8]) Let L be a complete lattice, X be a set, and let P(X ) be the ordinary
ΦL
power set of X . There exists an order isomorphism L X −−→ L ⊗ P(X ) defined by:

Φ L (g) = g(x) ⊗ {x}, g ∈ L X .
x∈X

Proof It is easily seen that every element g ∈ L X can be identified with a join revers-
Fg
ing map P(X ) −→ L

Fg (A) = g(x), A ∈ P(X ),
x∈A

ΦL 
and vice versa. Hence the map L X −−→ G L , P(X ) = L ⊗ P(X ) defined by

[Φ L (g)]() = {x ∈ X |  ≤ g(x)}, ∈L (10)

28 U. Höhle

is bijective and isotone. Since the inverse map of Φ L has the form

[Φ L−1 ( f )](x) = { ∈ L | x ∈ f ()}, f ∈ G(L , P(X )),

Φ L is an order isomorphism. The proof is complete, if we can show that Φ L (g) is

the smallest upper bound of

A := {g(x) ⊗ {x} | x ∈ X }.

An element k ∈ L ⊗ P(X ) is an upper bound of A iff {x} ⊆ k(g(x)) holds for all
x ∈ X iff {x ∈ X |  ≤ g(x)} ⊆ k() holds for all  ∈ L. Hence Φ L (g) (cf. (10)) is
the smallest upper bound of A. 

It follows from the universal property of the tensor product that the tensor prod-
uct in Sup induces a bifunctor (also denoted by ⊗) from Sup × Sup to Sup. In
h1 h2
particular, the tensor product of join preserving maps L 1 −→ L 2 and M1 −→ M2 is
determined by the commutativity of the following diagram:

⊗
L 1 × M1 / L 1 ⊗ M1

h 1 ×h 2 h 1 ⊗h 2
 
L 2 × M2 / L 2 ⊗ M2
⊗

Before we show that ⊗ induces a symmetric monoidal closed structure on Sup

we apply the tensor product of join preserving maps to the construction of variable-
basis fuzzy power set operators (cf. [25]). For this purpose we recall the following
notation. Let P : Set → Set be the covariant power set functor—i.e.
ϕ
P(X ) = P(X ), X−
→ Y, P(ϕ)(A) = ϕ(A), A ∈ P(X ).

h ϕ
Further, let L −
→ M be a join preserving map. Then for any map X −
→ Y the variable-
→
(ϕ,h)
basis power set forward operator L X −−−−→ M Y has the form (cf. 3.9(6) in [25]):

[(ϕ, h)→ (g)](y) = {h ◦ g(x) | ϕ(x) = y}, y ∈ Y.

ϕ h
Corollary 1 Let X − → Y be a map and L −
→ M be a join preserving map. Then the
following diagram is commutative:
(ϕ,h)→
LX / MY

ΦL ΦM
 
L ⊗ P(X ) / M ⊗ P(Y )
h⊗P(ϕ)
Modules in the Category Sup 29

Proof Referring to Theorem 2 we obtain the following relation for all g ∈ L X :

 
h ⊗ P(ϕ) Φ L (g) = h(g(x)) ⊗ {ϕ(x)}
x∈X

= {h(g(x)) | ϕ(x) = y} ⊗ {y}
y∈Y

= Φ M (ϕ, h)→ (g) . 

Because of the previous corollary the variable-basis power set forward operator
coincides with the tensor product of the traditional power set forward operator (i.e. the
operator taking direct images) with the join preserving map performing the change
of basis. In this sense the tensor product of Sup plays a fundamental role in the
mathematical foundations of fuzzy set theory.
The next important observation is the property that for every complete lattice M
the endofunctor F : Sup → Sup defined by

h
F(L) = L ⊗ M, L1 −
→ L 2 , F(h) = h ⊗ 1 M

has a right adjoint functor G. This result follows from the general constructions in
[3]. But we give here a direct proof.
h
For complete lattices M and N the set [M, N ] of all join preserving maps M −
→N
is a complete lattice w.r.t. the pointwisely defined order. Then G : Sup → Sup is
defined by:

k
G(N ) = [M, N ], N1 −
→ N2 , [G(k)](h) = k ◦ h, h ∈ [M, N1 ].
ev N
Further, the evaluation map [M, N ] × M −−→ N with ev N (h, m) = h(m) is a bimor-
phism. Because of the universal property of the tensor product there exists a unique
εN
join preserving map [M, N ] ⊗ M −→ N making the diagram

⊗
[M, N ] × M / [M, N ] ⊗ M
OOO
OOO
OOO εN
ev N OOO 
'
N

commutative. It is not difficult to see that ε = (ε N ) N ∈|Sup| is a natural transformation

from F ◦ G to idSup .
h
Theorem 3 Let L , M and N be complete lattices and L ⊗ M −
→ N be a join pre-
h
serving map. Then there exists a unique join preserving map L −−→ [M, N ] making
the following diagram commutative:
30 U. Höhle

h⊗1 M
L⊗M / [M, N ] ⊗ M
OOO
OOO
OOO εN
OOO
h OOO 
'
N

bh h
Proof The bimorphism L × M −→ N corresponding to L ⊗ M −
→ N is given by

bh (, m) = h( ⊗ m), (, m) ∈ L × M.

Then we conclude from the commutativity of the following diagram

h×1 M
L×M / [M, N ] × N
77JJ
77 JJJ tt 
77 JJJ⊗ ⊗ tt 
tt 
t
77 JJJ tt 
77 JJ t tt 
77 J$ ztt 
77 h⊗1 M
/ [M, N ] ⊗ M  
bh 7
L⊗M

ev N
77 
77 
77 h ε N 
77 
7   
N / N
1N

that h is unique and given by:

[h()](m) = bh (, m),  ∈ L , m ∈ M. (11)

Since joins in [M, N ] are computed pointwisely, the map h is in fact join
preserving. 
h
The previous theorem motivates the following terminology. Let L ⊗ M −
→ N be
h
a join preserving map. Then the join preserving map L −−→ [M, N ] determined by
(11) is called the monoidal adjoint map of h.
The correspondence h → h induces an order isomorphism from
[L ⊗ M, N ] onto [L , [M, N ]] (Proposition 5 in [3]). As a corollary of this fact
we obtain the associativity of the tensor product.

Corollary 2 Let L , M and N be complete lattices. There exists a unique order

aL M N
isomorphism (L ⊗ M) ⊗ N −−−→ L ⊗ (M ⊗ N ) satisfying the following property
for all (, m, n) ∈ L × M × N :

a L M N ( ⊗ m) ⊗ n =  ⊗ (m ⊗ n). (12)
Modules in the Category Sup 31

Proof By X op we denote the dual lattice of X . Hence the tensor product L ⊗ M has
the form L ⊗ M = [L , M op ]op .
If we choose ( ⊗ m) ⊗ n ∈ (L ⊗ M) ⊗ N = [L ⊗ M, N op ]op , then it is not
difficult to conclude from (11) that the relation ( ⊗ m) ⊗ n =  ⊗ (m ⊗ n)
holds. Hence the formation of taking monoidal adjoint maps is the desired order
isomorphism from (L ⊗ M) ⊗ N = [L ⊗ M, N op ]op to [L , [M, N op ]]op =
L ⊗ (M ⊗ N ). 

It follows immediately from (12) that a L M N is a component of a natural iso-

morphism making the pentagonal diagram (cf. [15, 17]) commutative. Hence the
bifunctor determined by the tensor product is associative.

Comment The associativity of the tensor product appears already as Theorem 1.7
in [27], but without an explicit statement of property (12).

The commutativity of the tensor product—i.e. the symmetry c of the bifunctor

⊗—goes back to Z. Shmuely 1974 (cf. [27, Theorem1.5, Lemma1.5]).

Lemma 1 Let L and M be complete lattices. Then there exist a unique order
cL M
isomorphism L ⊗ M −−→ M ⊗ L satisfying the following condition for all pairs
(, m) ∈ L × M:

c L M ( ⊗ m) = m ⊗ . (13)

Proof We fix  ∈ L and m ∈ M with  = ⊥ and m = ⊥. Because of the relation


{ ∈ L | m ≤ ( ⊗ m)( )} = (m ⊗ )(m ), m ∈ M

the pair (l ⊗ m, m ⊗ ) is a Galois connection between L and M. Hence the

correspondence determined by (1) is the desired order isomorphism c L M
satisfying (13). 

The next lemma explains that the 2-chain is the unit object w.r.t. the tensor product
in Sup (cf. [12, p. 6, Proposition 2(ii)]).

Lemma 2 Let L and M be complete lattices and let 1l = {0, 1} be the lattice with two
lM rL
different elements. Then there exist order isomorphisms 1l ⊗ M −
→ M and L ⊗ 1l −
→
L such that the following diagrams are commutative:

(L ⊗ 1l) ⊗ M
a X 1lY
/ L ⊗ (1l ⊗ M) L ⊗ 1l
c L1l
/ 1l ⊗ L
OOO ??
OO' oo ? 
r ⊗1
L M
wooo1 ⊗l
L M r L ?   l L
(14)
L⊗M L

In the case of L = M = 1l the relation r1l = l1l holds.

32 U. Höhle

Proof Let 0 be the bottom element of two element lattice 1l = {0, 1}. For f ∈ L ⊗ 1l
and g ∈ 1l ⊗ M we define 0 ∈ L and m 0 ∈ M by

0 = { ∈ L | f () = 1} and y0 = g(1)

and observe f = 0 ⊗ 1 and g = 1 ⊗ m 0 . Hence the desired order isomorphisms r X

and lY are defined by r L ( ⊗ 1) =  and lY (1 ⊗ m) = m. It is easily seen that the
diagrams in (14) are commutative and the property r1l = l1l holds. 

We can summarize the results of Theorem 3, Corollary 2 and Lemmas 1, 2 as

follows.
Fact. The septuple (Sup, ⊗, a, c, l, r, 1l) is a symmetric monoidal closed category
(cf. [15, 17]).

3 Unital Quantales as Monoids in Sup

∗
A pair (X, ∗) is a prequantale (cf. [26]) if X is a complete lattice and X × X −
→ X is
a bimorphism of Sup. Instead of ∗(x, y) we also write x ∗ y for x, y ∈ X . If (X, ∗)
is a prequantale, then ∗ is called the multiplication of X .
Because of the universal property of the tensor product in Sup the multiplication
∗ of a prequantale X can be identified with a binary operation in the sense of Sup—

i.e. a join preserving map X ⊗ X −→ X s.t. ∗ =  ◦ ⊗ holds. Hence prequantales
and magmas in Sup are equivalent concepts.
Notation If ∗ is the multiplication of a prequantale X with its corresponding binary
operation , then instead of (x ⊗ y) we also write x  y with x, y ∈ X . Since
x ∗ y and x  y coincide, it will depend on the context which kind of notation we
will prefer.
h
A homomorphism between prequantales is a join preserving map X − → Y which
also preserves the respective multiplications—i.e. h(x ∗ z) = h(x) ∗ h(z). Hence a
h
homomorphism (X, ∗) −
→ (Y, ∗) is characterized by the commutativity of the fol-
lowing diagram:
X⊗X
h⊗h
/ Y ⊗Y
 
 
 
 
 
 
X /Y
h
Modules in the Category Sup 33

Since (Sup, ⊗, a, c, l, r, 1l) is bi-closed (cf. Fact and [15]) and Sup is cocomplete,
we conclude from the free algebra algorithm (cf. [1] and p. 186 in [2]) that the
algorithm converges for the endofunctor T : Sup → Sup determined by

h
T(X ) = X ⊗ X, X−
→ Y, T(h) = h ⊗ h.

Hence every complete lattice generates a free magma (i.e. free prequantale) (cf. [6]).
A prequantale (X, ∗) is a quantale (cf. [26]) if ∗ is associative—i.e. (x ∗ y) ∗ z =
x ∗ (y ∗ z). A quantale is unital, if ∗ has a unit e—i.e. x ∗ e = e ∗ x for all x ∈ X .
A homomorphism h is unital, if h preserves the respective units—i.e. h(e) = e.
Let (X, ∗, e) be a unital quantale. If we identify the multiplication with its corre-
sponding binary operation and the unit with the join preserving map
1l = {0, 1} −
→ X sending 1 to e, then (X, , e) is a monoid in the sense of the sym-
metric monoidal category (Sup, ⊗, a, c, l, r, 1l) (cf. Corollary 2, Lemmas 1 and 2).
Since (Sup, ⊗, a, c, l, r, 1l) is bi-closed and Sup is cocomplete, every complete lat-
tice generates a free monoid (i.e. a free unital quantale) (cf. Theorem 2 on p. 172
in [17]).

Example 2 Let IN0 be the set of natural numbers including 0, and let 1l be the unit
object of the tensor product in Sup. Then the free unital quantale generated by 1l
coincides with the power set P(IN0 ) of IN0 equipped with the Minkowski addition
—i.e.
A  B = {a + b | a ∈ A, b ∈ B}, A, B ∈ P(IN0 ).
η1l
In fact, if 1l = {0, 1} −→ P(IN0 ) is the join preserving embedding with η1l (1) =
{1} (and η1l (0) = ∅), then for every further unital quantale (X, ∗, e) and for every
h
further join preserving map 1l −
→ X (e.g. h(1) = x) there exists a unique unital

h
homomorphism P(IN0 ) −→ X with the property h  ◦ η1l = h. In particular, h  is
given by: 
h  (A) = {x n | n ∈ A}, A ∈ P(IN0 )

where x 0 = e and x n = x ∗ x n−1 for n ∈ IN.

As always in monoidal categories the tensor product of monoids exist. In

(Sup, ⊗, a, c, l, r, 1l) the situation is as follows. Let (X 1 , ∗1 , e1 ) and (X 2 , ∗2 , e2 )
be unital quantales. First we infer from the associativity and commutativity of the
tensor product that there exists a unique order isomorphism

Φ
(X 1 ⊗ X 2 ) ⊗ (X 1 ⊗ X 2 ) −
→ (X 1 ⊗ X 1 ) ⊗ (X 2 ⊗ X 2 )

satisfying the condition Φ (x1 ⊗ x2 ) ⊗ (y1 ⊗ y2 ) = (x1 ⊗ y1 ) ⊗ (x2 ⊗ y2 ) for all
1 2
x1 , y1 ∈ X 1 and x2 , y2 ∈ X 2 . Hence, if X 1 ⊗ X 1 −−→ X 1 and X 2 ⊗ X 2 −−→ X 2
denote the binary operations corresponding to ∗1 and ∗2 , then there exists a binary
34 U. Höhle

operation on X 1 ⊗ X 2 defined by the composition of Φ with the tensor prod-

uct of 1 and 2 —i.e. (1 ⊗ 2 ) ◦ Φ. Obviously, the corresponding bimorphism

(X 1 ⊗ X 2 ) × (X 1 ⊗ X 2 ) −
→ X 1 ⊗ X 2 is uniquely determined by the property

(x1 ⊗ x2 )  (y1 ⊗ y2 ) = (x1 ∗1 y1 ) ⊗ (x2 ∗2 y2 ), x1 , y1 ∈ X 1 , x2 , y2 ∈ X 2 .

(15)
The associativity of  follows immediately from the associativity of ∗1 and ∗2 and
the fact that every tensor is the join of an appropriate family of elementary tensors.
Hence the triple (X 1 ⊗ X 2 , , e1 ⊗ e2 ) is a unital quantale and is called the tensor
product of (X 1 , ∗1 , e1 ) and (X 2 , ∗2 , .e2 ).
A special case of the previous situation is the construction of binary operations
on many-valued power sets according to Zadeh’s extension principle (cf. [8]) where
L.A. Zadeh considers only the special case ∗ = ∧ in his original papers [31–33].

Example 3 Let (X, ·, e) be a monoid in Set and P(X ) be the power set of X provided
with the set inclusion as partial order and the Minkowski multiplication  w.r.t. the
multiplication · on X . Then (P(X ), , {e}) is a unital quantale.
Let (Q, ∗, e) be a further unital quantale. The binary operation  on the Q-valued
power set Q X defined according to Zadeh’s extension principle is given by:

(g1  g2 )(x) = g1 (x1 ) ∗ g2 (x2 ), g1 , g2 ∈ Q X , x ∈ X. (16)
x1 ·x2 =x

If we now identify the tensor product Q ⊗ P(X ) with Q X (cf. Theorem 2), then 
coincides with the tensor product of ∗ with . In fact, because of (15) the relation

  
(g1  g2 )(x) ⊗ {x} = g1 (x1 ) ∗ g2 (x2 ) ⊗ {x1 · x2 }
x∈X x1 ,x2 ∈X
  
= g1 (x1 ) ∗ g(x2 ) ⊗ {x1 }  {x2 }
x1 ,x2 ∈X
  
= g1 (x1 ) ⊗ {x1 }  g(x2 ) ⊗ {x2 }
x1 ,x2 ∈X
 
= g1 (x) ⊗ {x}  g2 (x) ⊗ {x} .
x∈X x∈X

holds. Hence the identification of the multiplication  (defined according to Zadeh’s

extension principle) with the tensor product  of the respective multiplications follows
from Theorem 2.

For further details on the tensor product of unital quantales the reader is referred
to [8].
Modules in the Category Sup 35

A unital quantale (X, ∗, e) is called a unital quantale with involution or briefly

ι
involutive if there exists an involutive anti-automorphism X − → X —i.e. ι is an order
preserving involution on X satisfying the property ι(x ∗ y) = ι(y) ∗ ι(x) for all
x, y ∈ X . Instead of ι(x) we also write x . An element x ∈ X is self-adjoint if x = x
holds. The unit and the universal bounds are always self-adjoint.
h
A unital homomorphism X − → Y between involutive and unital quantales is invo-
lutive if h(x ) = h(x) holds for all x ∈ X .
The next lemma shows that every order preserving involution on a complete lattice
X can be extended to an involutive anti-automorphism on the free unital quantale
(X  , ∗, e ) generated by X .

Lemma 3 Let X be a complete lattice and (X  , ∗, e ) be the free unital quantale

ηX
generated by X with the corresponding embedding X −→ X  . For every order pre-

serving involution X − → X there exists a unique involutive anti-automorphism ι on
(X , ∗, e ) satisfying the condition ι ◦ η X = η X ◦ ι.
 

Proof Since the transposition of the multiplication preserves the associativity law, the
triple (X  , ∗op , e ) with x ∗op y = y ∗ x is again a unital quantale. Then ι is given
η X ◦ι ι
by the extension of X −−→ X  to a unique unital homomorphism (X  , ∗, e ) −
→
(X  , ∗op , e ). 

Because of the previous lemma there exist a plenty of involutive and unital quan-
tales. We finish this section with two prominent examples. (cf. [21]).

Example 4 Let A be a unital C ∗ -algebra with involution ∗ (cf. Sect. 6.3), and let
Max(A) be the set of all closed linear subspaces of A. The adjoint of a closed,
linear subspace M has the form M = {a ∗ | a ∈ M}. If M and N are closed, linear
 then the product M ∗ N is given by the closure of the set of all finite
subspaces,
sums ai · bi with ai ∈ M and bi ∈ N —i.e. M ∗ N = M N . Hence the restriction
of ∗ to closed ideals coincides with the usual ideal multiplication. In particular,
(M ∗ N ) = N ∗ M holds for all M, N ∈ Max(A).
On Max(A) we consider the partial order determined by the set inclusion. Then
Max(A) is a complete lattice, and the multiplication ∗ of closed linear subspaces is
join preserving in each variable separately. Hence (Max(A), ∗, ) is an involutive and
unital quantale. It is well known that Max(A) characterizes A up to an ∗-isomorphism
(cf. Theorem 3.3 in [16]). Sometimes Max(A) is also called the spectrum of A.

Example 5 A complete De Morgan algebra is a complete lattice L equipped with

an order reversing involution 0 —this means that the self-map λ −→ λ0 of L is an
involution on L satisfying the condition

λ1 ≤ λ2 =⇒ λ02 ≤ λ01 , λ1 , λ2 ∈ L .

Hence λ −→ λ0 is an anti-automorphism of L.
36 U. Höhle

We show that every complete De Morgan algebra (L , 0 ) gives rise to an involutive

and unital quantale. For this purpose we consider the complete lattice [L , L] of
all join preserving self-maps of L (cf. Sect. 2) provided with the composition as
multiplication—i.e.
f 1 ◦ f 2 (λ) = f 1 ( f 2 (λ)), λ ∈ L .

Then ([L , L], ◦, 1 L ) is a unital quantale. Further, the order reversing involution 0 on
L induces an order preserving involution on [L , L] by:
 0
f (λ) = f  (λ0 ) , λ ∈ L

where f  is the right adjoint map of f —i.e.


f  (μ) = {λ ∈ L | f (λ) ≤ μ}, μ ∈ L .

Since adjoint situations can be composed, the relation ( f 1 ◦ f 2 ) = f 2 ◦ f 1 follows.

Hence ([L , L], ◦, 1 L , ) is an involutive and unital quantale. (cf. [20], Example (5)
in [22]).

4 Modules on Unital Quantales

In any monoidal category C the concept of left-modules is available—i.e. objects

of C provided with a left action w.r.t. a monoid in C (cf. [17]). Here we recall the
axioms of a left-module in the special case that the monoidal category coincides with
(Sup, ⊗, a, c, , r, 1l) (cf. Fact in Sect. 2).
Let (Q, ∗, e) be a unital quantale with the corresponding binary operation

Q ⊗ Q −→ Q determined by the multiplication ∗. A left action on a complete lattice

M w.r.t. (Q, ∗, e) is a join preserving map Q ⊗ M − → M such that the following
diagrams commute (cf. [17]):
⊗1 M
(Q ⊗ Q) ⊗ M / Q⊗M

aQQM  (17)
 
Q ⊗ (Q ⊗ M) / Q⊗M / M
1Q ⊗ 

1l ⊗ M / Q⊗ M
e⊗1 M
?? 
??
?? 
?? 
??   (18)
M ?? 
?? 
 
M
Modules in the Category Sup 37

A pair (M, ) is a left Q-module if M is a complete lattice and  is a left action on

M w.r.t. (Q, ∗, e).
Let (α, x) −→ (α ⊗ x) be the bimorphism Q × M − → M corresponding to the
left action  on M. Then instead of (α ⊗ x) we also write α  x.
Since every tensor is the join of an appropriate family of elementary tensors, the
commutativity of the diagrams (17) and (18) is equivalent to the following axioms:
(M1) If α, β ∈ Q and m ∈ M, then α  (β  m) = (α ∗ β)  m.
(M2) If e denotes the unit of Q, then e  m = m for all m ∈ M.

Proposition 1 Let Q be a unital quantale, M be a complete lattice, and let [M, M]

be the unital quantale of all join preserving self-maps of M provided with the compo-
ev M
sition as multiplication. Further, let [M, M] × M −−→ M be the evaluation map—
i.e. ev M ( f, m) = f (m). Then there exists a bijective map between the set of all
left actions  on M in the sense of Sup and the set of all unital homomorphisms
h
Q−
→ [M, M] making the following diagram commutative:

Q⊗ M
h⊗1 M
/ [M, M] ⊗ M o ⊗ [M, M] × M
OOO
OOO ooo
OOO εM
ooooo
 OOOO o ev M
O'  woooo
M

where ε M is the join preserving map determined by the evaluation map ev M .

Proof Since Sup is monoidal closed (cf. Theorem 3), the homomorphism h coincides
with the monoidal adjoint of the left-action . The fact that h preserves the algebraic
structure follows from a chase of diagrams or by a simple calculation using the
axioms (M1) and (M2) directly. 

The previous proposition says that left Q-modules M can simply be characterized
by unital homomorphisms from Q to [M, M].
Before we continue, we first give a simple example of a left Q-module.
f
Example 6 Let X be a set and Q X be the set of all maps X −
→ Q provided with the
partial order defined pointwisely—i.e.

f ≤ g ⇔ f (x) ≤ g(x) for all x ∈ X.

Then Q X is the Q-valued power set of X (cf. Sect. 2), and the left action  on Q X
is determined by

(α  f )(x) = α ∗ f (x), α ∈ Q, f ∈ Q X , x ∈ X. (19)

Hence Q X is a left Q-module.

38 U. Höhle

h
Left Q-module homomorphisms are join preserving maps M − → N preserving the
respective left actions—i.e. the commutativity of the following diagram:
1Q ⊗h
Q⊗ M / Q⊗ N
 
 
 
 
 
 

M / N
h

As usually left Q-modules and left Q-module homomorphisms form a category

which we denote by Mod(Q).
It is well known that the forgetful functor U1 : Mod(Q) → Sup has a left adjoint
functor F1 : Sup → Mod(Q) sending a complete lattice X to Q ⊗ X where the
left action  on Q ⊗ X is defined by the commutativity of the following diagram
(cf. [17]):
a−1
Q ⊗ (Q ⊗ X )
QQX
/ (Q ⊗ Q) ⊗ X

⊗1 X (20)


$ 
Q⊗ X

Since in [17] the commutativity of the diagrams (17) and (18) has not been verified
in the general case of monoidal categories, we insert here this information in the
special case of Sup for the convenience of the reader.
Lemma 4 Let X be a complete lattice. Then the join preserving map

Q ⊗ (Q ⊗ X ) −
→ Q ⊗ X determined by (20) is a left action on Q ⊗ X .

Proof Since tensors in Q ⊗ X are joins of elementary tensors, we restrict our interest
to elementary tensors and choose x ∈ X and α, β, γ ∈ Q. Since  is associative, we
obtain:

α  (β  (γ ⊗ x)) = α  (β  γ) ⊗ x
= (α  (β  γ)) ⊗ x
= ((α  β)  γ) ⊗ x
= (α  β)  (γ ⊗ x).

Hence (M1) is verified. The axiom (M2) is evident. 

Modules in the Category Sup 39

For α ∈ Q and f ∈ Q ⊗ X the expression α  f is explicitly given by

Addition. 
α f = (α ∗ β) ⊗ f (β) where we have made use of (6).
β∈Q

A combination of Lemma 4 with Example 6 leads to the following corollary of

Theorem 2.

Corollary 3 ([12]) Let X be a set and P(X ) be the power set of X . Then the
ΦQ
order isomorphism Q X −−→ Q ⊗ P(X ) specified in Theorem 2 is a left Q-module
isomorphism.

Proof We maintain the notation from Example 6 and Lemma 4. In order to ver-
ify the relation α  ΦQ (g) = ΦQ (α  g) for all α ∈ Q and g ∈ Q X we refer to
Theorem 2 and obtain:

α  ΦQ (g) = α  g(x) ⊗ {x}
x∈X

= α  (g(x) ⊗ {x})
x∈X

= (α  g(x)) ⊗ {x}
x∈X

= (α ∗ g(x)) ⊗ {x}
x∈X
= ΦQ (α  g). 

η XM
Theorem 4 Let X be a complete lattice, (Q, ∗, e) be a unital quantale and X −−→
Q ⊗ X be the join preserving map determined by η XM (x) = e ⊗ x for all x ∈ X . Then
h
for every left Q-module N and for every join preserving map X −
→ N there exists a

h
unique left Q-module homomorphism Q ⊗ X −→ N making the following diagram
commutative:
ηM
X OO
X
/ Q⊗ X
OOO
OOO
OOO h (21)
h OOO 
'
N

Proof (a) (Uniqueness). Let h  be a left Q-module homomorphism making the dia-
gram (21) commutative. Then for all x ∈ X the relation h  (e ⊗ x) = h(x) follows.
Because of (20) (cf. Lemma 4) we obtain for γ ∈ Q and x ∈ X :

h  (γ ⊗ x) = h  ((γ ∗ e) ⊗ x) = h  (γ  (e ⊗ x)) = γ  h  (e ⊗ x) = γ  h(x).

40 U. Höhle

Since every tensor in Q ⊗ X is a join of elementary tensors, h  is uniquely deter-

mined by the commutativity of the diagram (21).

(b) (Existence). Let  be the left action on N . We define a join preserving map
h
Q ⊗ X −→ N by
h  =  ◦ (1Q ⊗ h). (22)

Then for every α ∈ Q and for every elementary tensor γ ⊗ x the relation

h  (α  (γ ⊗ x)) = h  ((α  γ) ⊗ x)

= [ ◦ (1Q ⊗ h)] (α  γ) ⊗ x
= (α  γ)  h(x)
= α  (γ  h(x))
= α  h  (γ ⊗ x).

holds. Since h  is join preserving and every tensor is a join of elementary tensors, h 
is obviously a left Q-module homomorphism.
Finally, since e is the unit of Q, we conclude from (22) that the relation

h  (e ⊗ x) = e  h(x) = h(x)

holds for all x ∈ X . Hence h  makes the diagram (21) commutative. 

Referring to Lemma 4 and Theorem 4 there exists a functor

F1 : Sup −
→ Mod(Q)

acting on objects and morphisms as follows:

ϕ
F1 (X ) = Q ⊗ X and for X −
→ Y, F1 (ϕ) = (ηYM ◦ ϕ) . (23)

Corollary 4 The functor F1 is left adjoint to U1 .

Proof The assertion follows immediately from Theorem 4. 

Corollary 5 The forgetful functor from U2 : Mod(Q) → Set has a left adjoint.

Proof Let X be a set, P(X ) be the ordinary power set of X , and let L be a complete
f
lattice. Since every map X −
→ L has a unique extension to a join preserving map

f
P(X ) −→ L with 
f  (A) = f (x), A ∈ P(X ),
x∈A
Modules in the Category Sup 41

the forgetful functor U0 : Sup → Set has a left adjoint functor F0 : Set → Sup
which sends a set X to its power set P(X ). Since adjoint situations compose, we
conclude from Corollary 4 that F2 = F1 ◦ F0 is left adjoint to U2 = U0 ◦ U1 . 

It follows from Corollaries 3, 4 and 5 that the monad corresponding to the adjoint
situation F2 −−| U2 is the Q-valued power set monad (PQ , η Q , μQ ) (on Set) where
ϕ 
PQ (X ) = Q X ,X− → Y, [PQ (ϕ)(g)](y) = {g(x) | ϕ(x) = y}, y ∈ Y,

−1 e, z = x
ηQ
X (x) = ΦQ (e ⊗ {x}) = 1{x} , 1{x} (z) = ⊥, z  = x, z, x ∈ X,

[μQ G(g) ∗ g(x), x ∈ X, G ∈ QQ .
X
X (G)](x) =
g∈Q X

The Eilenberg-Moore category of the Q-valued power set monad is isomorphic

to Mod(Q) (cf. [28]). The proof of this result is a generalization of the standard proof
that Sup is isomorphic to the Eilenberg-Moore category of the ordinary power set
monad (see Example I.5.15 in [18]). What is important here for us is the special
relationship between algebra and fuzzy set theory expressed by following
statement:

The Q-valued power set Q X is the free left Q-module generated by the set X .
This observation goes back to Joyal and Tierney (cf. p. 10 in [12]) and means
from a mathematical point of view that fuzzy set theory is module theory on unital
quantales. The next section is a confirmation of this insight.
At the end of this section we would like to draw the attention of the reader to the
interesting fact that submodules of Q X play a strategic role in the study of stratified
Q-valued topological spaces (cf. p. 180 in [9]).

5 Complete Q-Valued Order Sets and Left Q-Modules

First we recall the axioms of a many-valued preorder (cf. [5, 11, 23]). Let (Q, ∗, e)
p
be a unital quantale and X be a set. A map X × X − → Q is a Q-valued preorder
(Q-preorder for short) if p satisfies the following properties for all x, y, z ∈ X :
e ≤ p(x, x), (Reflexivity)
p(x, y) ∗ p(y, z) ≤ p(x, z). (Transitivity)
Every Q-preorder p has an underlying ordinary preorder ≤ defined by

x ≤ y ⇔ e ≤ p(x, y). (24)

42 U. Höhle

A Q-preorder p is skeletal or antisymmetric iff its underlying preorder is antisym-

metric. An antisymmetric Q-preorder is also called a Q-valued order (or Q-order
for short).
A pair (X, p) is a Q-preordered set if X is a set and p is a Q-preorder on X . The
same applies to Q-orders. It is well known that Q-preordered sets are Q-enriched
categories where Q is viewed as a monoidal biclosed category (cf. [15]).
h
Let (X, p) and (Y, q) be Q-preordered sets. A Q-homomorphism X −
→ Y is a
map satisfying the following condition

p(x, z) ≤ q(h(x), h(z))

for all x, z ∈ X . Obviously, Q-homomorphisms are always isotone w.r.t. the under-
lying preorders. The class of Q-preordered sets and the class of Q-homomorphisms
form a category denoted by Pre(Q).
Let (X, p) be a Q-preordered set. A covariant Q-presheaf on (X, p) is a Q-fuzzy
f
set X −
→ Q which is right-extensional—i.e.

f (x) ∗ p(x, y) ≤ f (y), x, y ∈ X.

On the set P(X, p) of all covariant Q-presheaves on (X, p) we introduce a Q-order

d
P(X, p) × P(X, p) −
→ Q as follows. First, we recall the left-implication  of Q

αβ = {γ ∈ Q | γ ∗ β ≤ α}, α, β ∈ Q.

Then  is an antisymmetric Q-preorder on Q and its underlying partial order coin-

cides with the dual order of Q. Now the Q-order d on P(X, p) is defined by

d( f, g) = f (x)  g(x), f, g ∈ P(X, p), (25)
x∈X


Hence P(X, p), d is a Q-ordered set.
ξ
Lemma 5 Let (X, p) be a Q-preordered
 set and P(X, p) −
→ X be a Q-homomor-
phism satisfying the condition ξ p(x, ) = x for all x ∈ X . Then p is antisym-
ξ 
metric, and every further Q-homomorphism P(X, p) − → X with ξ p(x, ) = x
coincides with ξ—i.e. ξ = ξ.

Proof Let us choose x, y ∈ X with e ≤ p(x, y) and e ≤ p(y, x). Then we conclude
from the transitivity of p that p(x, z) = p(y, z) holds for all z ∈ X . Hence x =
ξ( p(x, )) = ξ( p(y, )) = y follows—i.e. p is antisymmetric.
Modules in the Category Sup 43

Moreover, since every covariant Q-presheaf is right-extensional, the relation


d( f, p(x, )) = f (z)  p(x, z) = f (x)
z∈X

holds for all x ∈ X . Hence we obtain:

 
f (x) ≤ p ξ( f ), ξ( p(x, )) = p ξ( f ), x , x ∈ X. (26)

ξ 
If P(X, p) −→ X is a further Q-homomorphism with ξ p(x, ) = x for all x ∈ X ,
then we infer from (26) that ξ( f ) ≤ ξ( f ) holds where ≤ is the underlying order in
(X, p). Interchanging now the role of ξ and ξ the relation ξ = ξ follows from the
antisymmetry of ≤. 

Motivated by Lemma 5 we introduce the following terminology.

Definition 2 A triple (X, p, ξ) is a called a complete Q-ordered set if (X, p) is a

ξ
Q-ordered set and P(X, p) −
→ X is a Q-homomorphism provided with the property:

ξ p(x, ) = x, x ∈ X. (27)

Since ξ is unique, ξ is also called the formation of arbitrary meets in (X, p).

Theorem 5 Let M be a left Q-module with the left action . There exists a Q-
preorder p on M provided with the following properties:

(i) p(x, y) = {α ∈ Q | α  y ≤ x}, x, y ∈ M.
ξ
(ii) The map P(X, p) −
→ X defined by

ξ( f ) = f (x)  x, f ∈ P(M, p). (28)
x∈M

is a Q-homomorphism and satisfies (27)—i.e. (M, p, ξ) is a complete Q-ordered

set.

Proof We define p by (i) and show that p is a Q-preorder. Because of (M2) the
reflexivity of p is evident. With regard to the transitivity of p we use (M1) and
observe:

p(x, y) ∗ p(y, z)  z = p(x, y)  ( p(y, z)  z) ≤ p(x, y)  y ≤ x.

Hence p(x, y) ∗ p(y, z) ≤ p(x, z) follows. Since the underlying preorder of p coin-
cides with the dual order of M, p is even antisymmetric—i.e. p is Q-order on M.
In order to verify (ii) we proceed
  Because of p(x, z)  z ≤ x it follows
as follows.

immediately from (28) that ξ p(x, ) = p(x, z)  z = x holds for all x ∈ X .
z∈M
44 U. Höhle

Hence ξ satisfies (27). It remains to show that ξ is a Q-homomorphism. For this

purpose we choose f, g ∈ P(M, p) and observe:

d( f, g)  ξ(g) = d( f, g)  (g(x)  x)
x∈X
   
= f (x)  g(x) ∗ g(x)  x
x∈X

≤ f (x)  x
x∈X
= ξ( f ).

Hence d( f, g) ≤ p(ξ( f ), ξ(g)) follows—i.e. ξ is in fact a Q-homomorphism. 

Corollary 6 Let X be a set. Then the complete Q-ordered set (Q X , p, ξ) induced

by the free left Q-module Q X has the following form:

p( f, g) = f (x)  g(x), f, g ∈ Q X ,
x∈X
ξ 
P(Q X , p) −
→ Q X , ξ(F)(x) = F( f ) ∗ f (x), F ∈ P(Q X , p).
f ∈Q X

Proof Let  be the left action on Q X determined by (19). Then

 
f (x)  g(x) = {α ∈ Q | α  g ≤ f }
x∈X

follows. Hence the assertion follows immediately from Theorem 5, (28) and (19). 

Theorem 5 shows that every left Q-module gives rise to a complete

Q-ordered set. In the following considerations we show that also the converse holds.
For this purpose we first complete the object function (X, p) → P(X, p) to an end-
ofunctor P of Pre(Q):

h P(h)
(X, p) −
→ (Y, q), P(X, p) −−→ P(Y, q),

[P(h)(g)](y) = g(x) ∗ q(h(x), y), y ∈ Y.
x∈X

Further, we need a natural transformation μ : P ◦ P → P determined by

μ(X, p)
μ = (μ(X, p) )(X, p)∈|Pre(Q)| , where P(P(X, p), d) −−−→ P(X, p),
 
[μ(X, p) (F)](x) = F( f ) ∗ f (x), F ∈ P P(X, p), d , x ∈ X.
f ∈P(X, p)
Modules in the Category Sup 45

Lemma 6 Let (X, p, ξ) be a complete Q-ordered set. Then the following diagram
is commutative:
P(ξ)
P(P(X, p), d) / P(X, p)

μ(X, p) ξ (29)
 
P(X, p) / X
ξ

Proof Let ≤op be the dual partial order of the underlying partial order of p—i.e.

x ≤op y ⇔ e ≤ p(y, x).

Since ξ is a Q-homomorphism, for f, g ∈ P(X, p) we derive the following impli-

cation from the definition of d:

f ≤ g ⇒ ξ( f ) ≤op ξ(g). (30)

Further, because of the right-extensionality of covariant presheaves f ∈ P(X, p) the

relation  
f (x) = d f, p(x, ) = f (z)  d(x, z).
z∈X

holds for all x ∈ X . Now we apply (27) and again the property that ξ is a
Q-homomorphism and obtain

f (x) ≤ p(ξ( f ), x) f ∈ P(X, p), x ∈ X (31)


which implies μ(X, p) (F) ≤ P(ξ)(F) for all F ∈ P P(X, p), d . Because of (30)
the following relation holds:
 
ξ μ(X, p) (F) ≤op ξ P(ξ)(F) . (32)


On the other hand, for all f 0 ∈ P(X, p) we observe F( f 0 ) ≤ d μ(X, p) (F), f 0 .
Hence the relation
   
F( f 0 ) ∗ p ξ( f 0 ), z ≤ p ξ(μ(X, p) (F)), ξ( f 0 ) ∗ p ξ( f 0 ), z ≤ p ξ(μ(X, p) (F)), z

is valid—i.e. P(ξ)(F) ≤ p ξ(μ(X, p) (F)), . Now we apply (27) and (30) and
obtain:
 
ξ P(ξ)(F) ≤op ξ μ(X, p) (F) . (33)
46 U. Höhle
 
Finally, since ≤op is antisymmetric, ξ P(ξ)(F) = ξ μ(X, p) (F) follows from (32)
and (33). Thus the diagram (29) is commutative. 

Theorem 6 Let (X, p, ξ) be a complete Q-ordered set and ≤op be the dual order
) is a complete (ordinary)
w.r.t. the underlying partial order of p. Then (X, ≤op
lattice, and the map (α, x) −→ α · x = ξ α ∗ p(x, ) is a bimorphism. Moreover,

the join preserving map Q ⊗ X −
→ X determined by

α  x = ξ α ∗ p(x, ) , α ∈ Q, x ∈ X (34)

is a left action on X —i.e. (X, ) is a left Q-module, and the following relation holds:

ξ(g) = g(x)  x, g ∈ P(X, p). (35)
x∈X

Proof Let ⊥ be the universal lower bound of P(X, p) in the sense of the dual order
of the underlying partial order of d—i.e.

⊥(z) ≤ g(z), g ∈ P(X, p), z ∈ X

where ≤ is the partial order on Q. Because of (30) the element ξ(⊥) is the universal
 ≤ ). If A is a non empty subset of Xop, then it follows immediately
lower bound of (X, op

from (30) that ξ p(x, ) is the join of A w.r.t. ≤ —this means that (X, ≤op )
x∈A
is a complete lattice.
·
(a) We define a map Q × X −
→ X by

α · x = ξ α ∗ p(x, ) , α ∈ Q, x ∈ X (36)

and show that · is a bimorphism (in Sup). Because of the definition of · the relation
⊥ · x = ⊥ is evident for all x ∈ X . On the other hand, let us consider the univer-
sal lower bound ⊥ = ξ(⊥) in (X, ≤op ). Then we define a right-extensional map
F
P(X, p) −
→ Q by

F( f ) = α ∗ d(⊥, f ), f ∈ P(X, p), α ∈ Q.

Obviously, μ(X, p) (F) coincides with the universal lower bound in P(X, p)—i.e.
μ(X, p) (F) = ⊥. Further, we observe:

[P(ξ)(F)](z) = α ∗ d(⊥, f ) ∗ p(ξ( f ), z) = α ∗ p(ξ(⊥), z), z ∈ X.
f ∈P(X, p)

Hence the relation


⊥ = ξ ◦ μ(X, p) (F) = ξ ◦ P(ξ)(F) = ξ α ∗ p(ξ(⊥), ) = α · ξ(⊥) = α · ⊥
Modules in the Category Sup 47

follows from Lemma 6 and (36).

Further, for any non empty subset {gi | i ∈ I } of P(X, p) we define an element
F of P(P(X, p), d) by:

F( f ) = d(gi , f ), f ∈ P(X, p).
i∈I

Referring again to Lemma 6 we obtain:

 
ξ( gi ) = ξ(μ(X, p) (F)) = ξ(P(ξ)(F)) = ξ p(ξ(gi ), ) .
i∈I i∈I

Hence ξ is join preserving w.r.t. ≤op .  

Because of the previous observation the relation ( αi ) · x = αi · x follows
i∈I i∈I
immediately from the definition
 of ·. Now we consider a non empty subset {xi | i ∈ I }
of X and define g = p(xi , ). Hence ξ(g) is the join of {xi | i ∈ I } w.r.t. ≤op .
i∈I
Further, for α ∈ Q we define F ∈ P(P(X, p), d) by:

F( f ) = α ∗ d(g, f ), f ∈ P(X, p).

Obviously, μ(X, p) (F) = α ∗ g holds. Now we apply Lemma 6 and use again the fact
that ξ is join preserving:
 
α · xi = ξ α ∗ p(xi , )
i∈I i∈I
= ξ(α ∗ g)
= ξ(μ(X, p) (F))
= ξ(P(ξ)(F))

= ξ α ∗ p(ξ(g), )

= α · ( xi ).
i∈I

Hence we have verified that · is join preserving in each variable separately w.r.t. ≤op .

(b) Because of the universal property of the tensor product there exists a unique join

preserving map Q ⊗ X − → X making the diagram
⊗
Q⊗ X / Q⊗ X
OOO
OOO
O
· OOO 

'
X
48 U. Höhle

commutative. Obviously,  satisfies (34) because of (36). Therfore we only show

that  is a left action on (X, ≤op ).
Because of (27) and (34) the axiom (M2) is evident. In order to verify (M1) we
fix α, β ∈ Q and x ∈ X . Then we define F ∈ P(P(X, p), d) as follows:

g(z) = α ∗ p(x, z), z ∈ X, F( f ) = β ∗ d(g, f ), f ∈ P(X, p).

Obviously, μ(X, p) (F) = (β ∗ α) ∗ p(x, ) and ξ(g) = α  x hold. Now we apply

again Lemma 6 and obtain:

(β ∗ α)  x = ξ(μ(X, p) (F)) = ξ(P(ξ)(F)) = ξ(β ∗ p(ξ(g), )) = β  (α  x).

(c) In order to verify (35) we fix g ∈ P(X, p) and use its right-extensionality—i.e.

g(z) = g(x) ∗ p(x, z), z ∈ X.
x∈X

Hence (35) follows form (34) and the property that ξ is join preserving. 

From Theorems 5 and 6 it follows that left Q-modules and complete Q-ordered
sets are equivalent concepts—a result which goes back to Stubbe in the more general
context of quantaloid enriched categories (cf. [29], see also Remark 5.6 in [10]).
In this context, complete Q-ordered sets emphasizes the many-valued (i.e. enriched
categorical) aspect, while left Q-modules refer to the algebraic properties of this
theory. Since in the fuzzy community pre-singletons of the form α ∗ p(x, ) are
viewed as Q-fuzzy points, it is important to realize that the left action of α on x in
the sense of Sup means the join of α ∗ p(x, ) w.r.t. the dual order determined by
p (cf. Theorem 5(i)).
Finally, we mention the fact that complete Q-ordered sets coincide with algebras
of the monad of covariant Q-presheaves (cf. Remark 5.7 in [10]).

6 Left Q-Modules on Involutive Quantales

Let Q be an involutive and unital quantale with unit e and an order preserving
involution . Further, let (M, ) be a left Q-module, and M op be the complete lat-
tice provided with the dual order of M. It is easily seen that the right implication

Q × M op −→ M op defined by

αm = {n ∈ M | α · n ≤ m}, α ∈ Q, m ∈ M

is a bimorphism. Referring to the universal property of the tensor product in Sup the
right implication  and the order preserving involution on Q induce a left action 
on M op determined as follows

α  m = α  m, α ∈ Q, m ∈ M (37)
Modules in the Category Sup 49

In fact, the following relations are an immediate corollary of (M1) and (M2):

e  m = e  m = e  m = m,
α  (β  m) = α  (β  m) = (β ∗ α )  m = (α ∗ β)  m.

Hence M op is a left Q-module w.r.t. the left action defined by (37).

h
Proposition 2 Let Q be an involutive and unital quantale and X − → Y be a left
Q-module homomorphism. If h  is the right adjoint of h and the complete lattices
X op and Y op are provided with the respective left actions according to (37), then
h
Y op −→ X op is again a left Q-module homomorphism.

Proof Let us choose α ∈ Q, x ∈ X and y ∈ Y . Then the following chain of equiv-

alences hold:

x ≤ α  h  (y) ⇔ α  x ≤ h  (y) ⇔ h(α  x) = α  h(x) ≤ y .

Hence x ≤ α  h  (y) ⇔ h(x) ≤ α  y ⇔ x ≤ h  (α  y) follows—i.e.

α  h (y) = h  (α  y).



If Q is an involutive and unital quantale, then we conclude from the previous

proposition that the category Mod(Q) is self-dual.

6.1 Two-Forms in Sup

Let ⊗ be the tensor product and 1l be the unit object in Sup (cf. Sect. 2). A 2-form
ϕ
on a complete lattice M is a join preserving map M ⊗ M − → 1l. A 2-form on M is
symmetric if the diagram
cM M
M⊗M / M⊗M
OOO
OOO ϕ
O
ϕ OOO 
'
1l

is commutative where c M M is determined by Lemma 1.

Since Sup is monoidal closed (cf. Theorem 3), every 2-form on M can be identified
ϕ
with its monoidal adjoint map M −−→ [M, 1l]. A symmetric 2-form ϕ on M is faithful
if its monoidal adjoint is a monomorphism in Sup—this means an injective and join
preserving map.
50 U. Höhle

In order to give a characterization of 2-forms we recall the simple fact that for
h
every complete lattice X a join preserving map X −
→ 1l can be identified with a
unique element z ∈ X —i.e.

0, x ≤ z,
h(x) =
1, x ≤ z.

ϕ
Hence by definition of the tensor product (cf. Sect. 2) a 2-form M ⊗ M −
→ 1l can be
f
identified with a join reversing map M −
→ M. In particular, a 2-form ϕ is symmetric
iff the chain of equivalences

n ≤ f (m) ⇔ m⊗n ≤ f ⇔ n⊗m ≤ f ⇔ m ≤ f (n)

holds for all m, n ∈ M—this means that ϕ is symmetric iff the corresponding join
reversing map f satisfies the following condition

m ≤ f ( f (m)), m ∈ M. (38)

Since f is antitone and ≤ is antisymmetric, it is interesting to see that (38) also

implies the subsequent relation

f (m) = f ( f ( f (m))), m ∈ M. (39)

Further, the monoidal adjoint map ϕ of a 2-form ϕ can be characterized by its
corresponding join reversing map f as follows:

[ϕ(m)](n) = 0 ⇔ ϕ(m ⊗ n) = 0 ⇔ m⊗n ≤ f ⇔ n ≤ f (m)


0, n ≤ f (m),
i.e. [ϕ(m)](n) = n ∈ M.
1, n ≤ f (m),

Because of (39) a symmetric 2-form ϕ on M is faithful iff the corresponding join

reversing map f is an involution. Hence we have established the following important
result due to Resende [24].

Proposition 3 Let M be a complete lattice. There exists a bijective map between the
set of all symmetric and faithful 2-forms ϕ on M and the set of all order reversing
involutions 0 on M such that the condition

ϕ(m ⊗ n) = 0 ⇔ n ≤ m0

holds for all m, n ∈ M.

Modules in the Category Sup 51

Because of the previous proposition we gain the important understanding that

the lattice-theoretic concept of a complete De Morgan algebra (cf. Example 5) can
be formulated entirely in terms of categorical data provided by Sup. In this sense
order reversing involutions on complete lattices are not only related to non-classical
negations, but they play also an important geometric role expressed by their associated
orthogonality relation:

m ⊥n ⇔ n ≤ m0.

6.2 Involutive Left Q-Modules

Let (Q, ∗, e, ) be a unital quantale with involution. We enrich left Q-modules by

symmetric 2-forms (cf. [24]).

Definition 3 A pair (M, ϕ) is called an involutive left Q-module if M is left

Q-module and ϕ is a symmetric 2-form on M such that the relation

ϕ (α  x) ⊗ y = ϕ(x ⊗ (α  y) (40)

holds for all α ∈ Q and x, y ∈ M where  is the left action on M.

The next theorem is a refinement of Proposition 1.

Theorem 7 Let (M, 0 ) be a complete De Morgan algebra and [M, M] be the invo-
lutive and unital quantale of all join preserving self-maps of M (cf. Example 5).
Further, let ϕ be the faithful and symmetric 2-form on M corresponding to the order
reversing involution α → α0 . Then there exists a bijective map between the set of
all left actions  on M with the property that (M, ϕ) is an involutive left Q-module
h
and the set of all involutive and unital homomorphisms Q −
→ [M, M] such that the
following diagram is commutative:

Q⊗M
h⊗1 M
/ [M, M] ⊗ M
OOO
OOO
OO εM
 OOO
O' 
M

where ε M is the join preserving map corresponding to the evaluation map ev M .

Proof Let  be a left action on M. Referring to the proof of Proposition 1 it is suffi-

cient to show that  satisfies (40) iff the monoidal adjoint map of  is an involutive
(unital) homomorphism.
52 U. Höhle

(a) Let us assume that  fulfills (40) and h is the monoidal adjoint of . We maintain
the notation of Example 5 and obtain the following chain of equivalences for all
α ∈ Q and m, n ∈ M:

m ≤ [h(α) ](n 0 ) ⇔ [h(α)](m) ≤ n 0

⇔ α  m ≤ n0

⇔ ϕ (α  m) ⊗ n = 0

⇔ ϕ m ⊗ (α  n) = 0
⇔ α  n ≤ m 0
⇔ [h(α )](n) ≤ m 0
 0
⇔ m ≤ [h(α ](n) .

Hence [h(α) ](n) = [h(α ](n) follows for all n ∈ M.

h
(b) Let us assume that Q − → [M, M] is an involutive and unital homomorphism.
Since the left action  on M is determined by:

α  m = [h(α)](m), α ∈ Q, m ∈ M, (41)

we obtain:

ϕ (α  m) ⊗ n = 0 ⇔ α  m ≤ n 0
⇔ [h(α)](m) ≤ n 0
⇔ m ≤ [h(α) ](n 0 )
⇔ [h(α) ](n) ≤ m 0
⇔ [h(α )](n) ≤ m 0
⇔ α  n ≤ m 0

⇔ ϕ m ⊗ (α  n) = 0.

Hence (M, ϕ) is an involutive left Q-module w.r.t. the left action defined in (41). 

Morphisms between involutive left Q-modules (M, ϕ) and (N , ψ) are left Q-

h
module homomorphisms M − → N which also preserve the respective symmetric
2-forms—i.e. the commutativity of the following diagram:

M⊗M
h⊗h
/ N⊗N
OOO
OOO
O
ϕ OOO 
ψ
'
1l
Modules in the Category Sup 53

h
In the case of faithful and symmetric 2-forms a left Q-module homomorphism M − →
N is a morphism iff h is orthogonal—i.e. h preserves and reflects the respective
orthogonality relations—i.e. m 1 ⊥ m 2 ⇔ h(m 1 )⊥ h(m 2 ) for all m 1 , m 2 ∈ M.

6.3 Representations of C ∗ -Algebras and Involutive

Left Modules
For the convenience of the reader we recall the axioms of a C ∗ -algebra. A Banach
algebra A = (A, +, ·) with unit e (cf. [13, 19]) is a C ∗ -algebra with unit iff A is
provided with a conjugate-linear map a → a ∗ of A into itself satisfying the following
conditions:
(C1) (a ∗ )∗ = a for all a ∈ A.
(C2) (a · b)∗ = b∗ · a ∗ for all a, b ∈ A.
(C3)  a ∗ · a  =  a 2 for all a ∈ A.
Sometimes the conjugate-linear map a → a ∗ is called the involution of A. In this
context the condition (C3) is also known as C ∗ -property.
Because of (C1), (C3) and the submultiplicativity of the norm the involution
a → a ∗ is always an isometry.
Morphisms between C ∗ -algebras are ∗-homomorphisms—these are algebra
π
homomorphisms A − → B with the property π(a ∗ ) = π(a)∗ for all a ∈ A. Hence
∗-homomorphisms are algebra homomorphisms preserving the corresponding invo-
lutions. It follows from the spectral theory of self-adjoint elements of A and the
C ∗ -property that a ∗-homomorphism π satisfies always the condition

 π(a)  ≤  a , a ∈ A.

Hence ∗-homomorphisms are continuous.

Example 7 ([19]) Let H be a Hilbert space and L(H) be the Banach algebra of
T
all bounded and linear operators H − → H. Then L(H) is a C ∗ -algebra w.r.t. to the
involution given by the formation of adjoint operators—i.e.

T (x), y = x, T ∗ (y), x, y ∈ H.

In the next definition we summarize basic properties of representations of

C ∗ -algebras (cf. [13, 14]).

Definition 4 (a) A representation of a unital C ∗ -algebra A = (A, +, ·) is a pair

(H, π) where H is a Hilbert space and π is a ∗-homomorphism from A to the
π
C ∗ -algebra of all bounded and linear operators on H—i.e. A − → L(H).
(b) A representation (π, H) of A is called cyclic if there exists a vector x ∈ H
with x = 0 s.t. the closure of {π(a)(x) | a ∈ A} coincides with H. In this context x
is termed a cyclic vector for π.
54 U. Höhle

(c) A representation (π, H) of A is irreducible iff every non-trivial closed linear

subspace U of H being invariant under π(A) (i.e. {π(a)(x) | a ∈ A, x ∈ U } ⊆ U )
coincides with H.
The aim of the following considerations is to show that every representation of
a unital C ∗ -algebra A induces an involutive left Max(A)-module with faithful and
symmetric 2-form where Max(A) is the spectrum of A (cf. Example 4). The next
theorem is due to Mulvey and Pelletier [21].
Theorem 8 Let (π, H) be a representation of a unital C ∗ -algebra A. Further, let
P(H) be the complete De Morgan algebra of all closed linear subspaces of H
provided with the orthogonal complement as order reversing involution 0 , and let
[P(H), P(H)] be the involutive and unital quantale of all join preserving self-maps
of P(H) (cf. Example 5). Then π induces an involutive and unital homomorphism
hπ
Max(A) −→ [P(H), P(H)] by
  
[h π (I )](U ) = top. closure lin.hull π(a)(x) | a ∈ I, x ∈ U (42)

where I ∈ Max(A) and U ∈ P(H).

Because of Theorems 7 and 8 every representation (π, H) of a unital C ∗ -algebra
A induces a left action π on P(H) in the sense of Sup determined by:
  
I π U = top. closure lin.hull π(a)(x) | a ∈ I, x ∈ U . (43)

In this context the faithful and symmetric 2-form ϕ corresponds to the orthogo-
nal complementation in P(H). The pair (P(H), π ) is also called the canonical
involutive left Max(A)-module associated with the representation (π, H) of A.
The next results are due to Kruml and Resende [16] showing that involutive left
Q-modules play a significant role in the theory of operator algebras.
Theorem 9 (a) Let (π, H) be a representation of a unital C ∗ -algebra A = (A, +, ·)
and (P(H), π ) be the canonical involutive left Max(A)-module associated with
(π, H).
(a) A vector x ∈ H with x = 0 is cyclic for π iff  π x =  where x is the
1-dimensional subspace of H generated by x.
(b) The representation (π, H) is irreducible iff every atom a of P(H) is a generator
of P(H)—i.e. P(H) = {I π a | I ∈ Max(A)}.
(c) Two representations (π1 , H1 ) and (π2 , H2 ) of A are equivalent (i.e. there exists
T
a unitary transformation H1 − → H2 s.t. π2 (a) = T ◦ π1 (a) ◦ T ∗ holds for all
a ∈ A) iff the respective canonical involutive left Max(A)-modules associated
with (π1 , H) and (π2 , H) are isomorphic.
It is worthwhile to note that the previous results make use of fundamental princi-
ples of the theory of C ∗ -algebras (see e.g. Proposition 4.5.3 in [13], Theorems 10.2.7
and 10.2.10 in [14]).
Modules in the Category Sup 55

References

1. Adámek, J.: Free algebras and automata realization in the language of categories. Comment
Math. Univ. Carol. 15, 589–602 (1974)
2. Adámek, J., Trnková, V.: Automata and Algebras in Categories. Kluwer Academic Publishers,
Dordrecht (1990)
3. Banaschewski, B., Nelson, E.: Tensor products and bimorphisms. Cand. Math. Bull. 19, 385–
402 (1976)
4. Birkhoff, G.: Lattice Theory, Colloquium Publications, vol. 25, 3rd edn., eighth printing. Amer-
ican Mathematical Society, Rhode Island (1995)
5. Denniston, J.T., Melton, A., Rodabaugh, S.E.: Enriched categories and many-valued preorders:
categorical, semantical and topological perspectives. Fuzzy Sets Syst. 256, 4–56 (2014)
6. Eklund, P., Höhle, U., Kortelainen, J.: A survey on the categorical term construction with
applications. Fuzzy Sets and Syst. doi:10.1016/j.fss.2015.07.003
7. Galatos, N., Jipsen, P., Kowalski, T., Ono, H.: Residuated Lattices: An Algebraic Glimpse at
Substructural Logics, Studies in Logic, vol. 151. Elsevier, Amsterdam (2007)
8. Gutiérrez García, J., Höhle, U., Kubiak, T.: Tensor products in Sup and their application in
constructing quantales (submitted)
9. Höhle, U.: Many Valued Topology and Its Applications. Kluwer Academic Publishers, Boston
(2001)
10. Höhle, U.: Categorical foundations of topology with applications to quantaloid enriched topo-
logical spaces. Fuzzy Sets Syst. 256, 166–210 (2014)
11. Höhle, U.: Many-valued preorders I: the basis of many-valued mathematics. In: Magdalena, L.,
et al. (eds.) Enric Trillas: A Passion for Fuzzy Sets, Studies in Fuzziness and Soft Computing,
vol. 322, pp. 125–150. Springer, Heidelberg (2015)
12. Joyal, A., Tierney, M.: An Extension of the Galois Theory of Grothendieck, Memoirs of the
American Mathematical Society, vol. 51, Number 309. American Mathematical Society (1984)
13. Kadison, R.V., Ringrose, J.R.: Fundamentals of the Theory of Operator Algebras, Volume I
Elementary Theory, Graduate Studies in Mathematics Volume 15. American Mathematical
Society (1997)
14. Kadison, R.V., Ringrose, J.R.: Fundamentals of the Theory of Operator Algebras, Volume
II Adavanced Theory, Graduate Studies in Mathematics Volume 16. American Mathematical
Society (1997)
15. Kelly, G.M.: Basic Concepts of Enriched Category Theory, London Mathematical Society
Lecture Notes Series 64. Cambridge University Press (1982)
16. Kruml, D., Resende, P.: On quantales that classify C ∗ -algebras. Cahiers Topol. Géom. Différ.
Catég. 45, 287–296 (2004)
17. Mac Lane, S.: Categories for the Working Mathematician, 2nd edn. Springer (1998)
18. Manes, E.G.: Algebraic Theories. Springer, New York (1976)
19. Meise, R., Vogt, D.: Introduction to Functional Analysis, Oxford Gruaduate Texts in Mathe-
matics. Oxford University Press (1997)
20. Mulvey, C.J., Pelletier, J.W.: A quantisation of the calculus of relations, CMS Proceedings,
vol. 13, pp. 345–360. American Mathematical Society, Providence (1992)
21. Mulvey, C.J., Pelletier, J.W.: On the quantisation of points. J. Pure Appl. Algebra 159, 231–295
(2001)
22. Pelletier, J.W., Rosický, J.: Simple involutive quantales. J. Algebra 195, 367–386 (1987)
23. Pu, Q., Zhang, D.: Preordered sets valued in a G L-monoid. Fuzzy Sets Syst. 187, 1–32 (2012)
24. Resende, P.: Sup-lattice 2-forms and quantales. J. Algebra 276, 143–167 (2004)
25. Rodabaugh, S.E.: Powerset operator foundations for poslat fuzzy theories and topologies. In:
Höhle, U., Rodabaugh, S.E. (eds.) Logic, Topology, Theory, Measure, Mathematics of Fuzzy
Sets, pp. 91–116. Kluwer Academic Publishers (1999)
26. Rosenthal, K.I.: Quantales and Their Applications, Pitman Research Notes in Mathematics,
vol. 234. Longman Scientific Technical, Longman House, Burnt Mill, Harlow (1990)
56 U. Höhle

27. Shmuely, Z.: The structure of Galois connections. Pac. J. Math. 54, 209–225 (1974)
28. Solovyov, S.A.: Powerset operator foundation for catalg fuzzy set theories. Iran. J. Fuzzy Syst.
8, 1–46 (2011)
29. Stubbe, I.: Categorical structures enriched in a quantaloid tensored and cotensored categories.
Theory Appl. Categ. 16, 283–306 (2006)
30. Zadeh, L.A.: Fuzzy sets. Inf. Control 8, 338–353 (1965)
31. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning
I. Inf. Sci. 8, 119–249 (1975)
32. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning
II. Inf. Sci. 8, 301–357 (1975)
33. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning
III. Inf. Sci. 9, 43–80 (1975)
A Geometric Approach to MV-Algebras

Daniele Mundici

Abstract Markov unrecognizability theorem puts an end to the classical program

of equipping any combinatorial manifold M with a computable set IM of invariants
such that a manifold N is homeomorphic to M iff IM = IN . To make sense of
the statement of the theorem manifolds are replaced by finite strings of symbols
for triangulated rational polyhedra, and homeomorphisms are understood as rational
PL-homeomorphisms. Thus, objects and arrows undergo a radical transformation—
and yet with no essential loss of generality for the original recognition problem. A
further restriction on the arrows arises if one views the recognizability problem from
the viewpoint of algorithmic complexity theory: here one must take into account
the amount of information needed to specify rational polyhedra. We are thus left
with the category of rational polyhedra (objects) with integer PL-maps (arrows). A
new geometry arises, where the affine group over the integers takes on the same
role as the isometry group does in euclidean space. Differently from the category
of rational polyhedra with rational PL-maps, a wealth of new geometric computable
invariants emerges in this new category. We discuss in particular the rational measure
of rational polyhedra. Its role and applicability is amplified by the duality between
rational polyhedra and finitely presented MV-algebras.

1 Where Do the Łukasiewicz Axioms Come From?

Boolean logic L2 deals with {0, 1}-observables/events. For instance, in the reduction
of the colorability problem to the boolean satisfiability problem, given a graph G
and a palette of k colors, the basic observable “the first vertex of G gets the third color”
is coded by a variable X13 and every composite observable (such as “each vertex of G

D. Mundici (B)
Department of Mathematics and Computer Science, University of Florence,
Florence, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 57

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_4
58 D. Mundici

gets precisely one among the k available colors”) is coded by a boolean combination
of the Xij in such a way that the k-colorability of G amounts to the satisfiability of
a suitable boolean formula φG in the basic observables Xij . The faithfulness of the
map G → φG is accounted for by saying that G is k-colorable iff φG is satisfiable.
The efficiency of this map follows from its being computable in polynomial time.
Now most observables in physics, as well as most random variables in real life,
are not {0, 1}-valued; measurements are not infinitely precise and their outcome can
only be given by specifying a real number together with an error interval. Since phys-
ical laws are formulated in terms of relations between real-valued quantities rather
than relations between intervals, errors are implicitly taken care of by assuming that
observables are continuous: continuity ensures that small errors in the measurement
of the basic observables have small effects on the evaluation of compound observ-
ables.
For any bounded observable O one may rescale the measurement unit in such
a way that the result of any measurement of O fits into the real interval [0, 1].
Once a [0, 1]-valued logic L is chosen to deal with [0, 1]-valued observables as
boolean logic L2 does for {0, 1}-observables, compound observables are formalized
in L by applying continuous connectives to the output (i.e., the truth-value) of basic
observables: the latter are the “variables” of L.
The functional completeness of boolean logic ensures that all n-variable boolean
functions are obtainable from the variables via the boolean connectives. By contrast, a
brute force counting argument shows that no [0, 1]-valued logic L can be functionally
complete, and hence one must judiciously select the most appropriate connectives
for L.
If, as is often the case, L is defined in terms of a consequence relation and the
L-consequence relation is formulated via Modus Ponens (MP), then inevitably L
must be equipped with an “implication” operation ⇒L : [0, 1]2 → [0, 1]. By the
above discussion, ⇒L must be continuous. If ⇒L is to be (minimally) reminiscent
of boolean implication then the order of premises is irrelevant, and for any two
truth-values x, y ∈ [0, 1], x ⇒L y equals 1 precisely when x ≤ y.
Elementary as they are, these three conditions characterize the implication x →Ł∞
y = min(1, 1 − x + y) of the Łukasiewicz infinite-valued calculus Ł∞ , [19]:

Lemma 1 For any map ⇒ : [0, 1]2 → [0, 1] the following conditions are equiva-
lent:
(i) ⇒ is continuous, x ⇒ (y ⇒ z) = y ⇒ (x ⇒ z), and (1 = x ⇒ y iff
x ≤ y).
(ii) There exists precisely one increasing bijection φ of [0, 1] onto [0, 1] such that
x ⇒ y = φ −1 (φ(x) →Ł∞ φ(y)) for all x, y ∈ [0, 1].
A Geometric Approach to MV-Algebras 59

Proof This is known as the Smets-Magrez Theorem [31]. Some conditions assumed
in [31] are redundant (see Fodor and Roubens [16, Theorem 1.15] and Baczyński
[1]. Also see [2]). Related results had been obtained earlier by Trillas and Valverde
in [33, Theorem 3.4]. 

Theorem 1 Let I = ([0, 1], 0, 1, ⇒, ¬) be the unit real interval equipped with a
binary operation ⇒ satisfying the three conditions in Lemma 1(i), and with the
derived operation ¬x = x ⇒ 0. Then the algebra I has no nontrivial congruences
and satisfies the following axioms:
(i) 1⇒x=x
(ii) (x ⇒ y) ⇒ ((y ⇒ z) ⇒ (x ⇒ z)) = 1
(iii) ((x ⇒ y) ⇒ y) = ((y ⇒ x) ⇒ x)
(iv) (¬x ⇒ ¬y) ⇒ (y ⇒ x) = 1.
Conversely, every algebra B = ([0, 1], 0, 1, ⇒, ¬) without nontrivial congruences
and satisfying axioms (i)–(iv) is isomorphic to an algebra I  = ([0, 1], 0, 1, ⇒ , ¬ ),
where ⇒ satisfies the three conditions in Lemma 1(i), and ¬ x = x ⇒ 0.

Proof The map φ of Lemma 1(ii) is an isomorphism between I and the standard
Wajsberg algebra. The converse statement follows from [11, 3.5] together with the
well known fact that the standard MV-algebra [0, 1]MV = ([0, 1], 0, 1, ¬, ⊕) (where
¬x = 1 − x and x ⊕ y = min(1, x + y)) is not isomorphic to any of its proper sub-
algebras, [11, 7.2.6]. 

Equations (i)–(iv) characterize Wajsberg algebras—i.e., MV-algebras up to term

equivalence, [11]. In this way, MV-algebras (= HSP([0, 1]MV )) can be introduced
in the fastest possible way. Interpreted as tautologies, equations (i)–(iv) amount to
the Łukasiewicz axioms [19]. Thus MV-algebras can be redefined as the algebras of
the only [0, 1]-valued logic whose Modus Ponens relation is framed in terms of an
implication connective satisfying the three elementary conditions of Lemma 1(i). A
related approach to Łukasiewicz logic is given by the following result:

Proposition 1 ([23]) Let ([0, 1], ∨, ∧, ∗, ⇒, 0, 1) be a residuated lattice in which

∨ and ∧ are the natural max and min operations. If the map ⇒ : [0, 1]2 → [0, 1] is
continuous then ∗ and ⇒ are the Łukasiewicz conjunction x ∗ y = max(0, x + y − 1)
and implication x ⇒ y = min(1, 1 − x + y).

In the language of t-norms (i.e., commutative associative monotone binary oper-

ations on [0, 1] having 1 as their neutral element, [20]) the above results show that
among all continuous t-norms, Łukasiewicz conjunction is the only one yielding a
logic with a continuous implication connective.
Our next aim in this paper is to approach MV-algebras from an entirely dif-
ferent viewpoint, starting from Markov’s celebrated unrecognizability theorem for
manifolds.
60 D. Mundici

2 A New Geometry: Rational Polyhedra with Integer

PL-Maps

You’re browsing, let us imagine, in a music shop, and come across a box of faded pianola
rolls. One of them bears an illegible title, and you unroll the first foot or two, to see if you
can recognize the work from the pattern of holes in the paper. Are there four beats in the bar,
or only three? Does the piece begin on the tonic, or some other note? Eventually you decide
that the only way of finding out is to buy the roll, take it home, and play it on the pianola.
Within seconds your ears have told you what your eyes were quite unable to make out – that
you are now the proud possessor of a piano arrangement of “Colonel Bogey”.

Longuet-Higgins, H. C. (1979). “Review Lecture: The Perception of Music”. Proceedings

of the Royal Society B: Biological Sciences 205 (1160) page 307.

A similar situation occurs in the recognition of geometrical figures P, Q. How

can we effectively determine that P is a tetrahedron up to homeomorphism? How
can we prove that P is not homeomorphic to Q? One should first note that for the
statement of the problem to make sense, P and Q must be presented as finite strings of
symbols. Not all presentations are equally good. The evolution of notational systems
for the natural numbers shows that notations allowing more efficient computations
supersede less efficient notations.
To code a combinatorial manifold P by a finite string of symbols, one usually
proceeds as follows: (i) first equips P with a triangulation Δ0 , (ii) next replaces P
by the underlying set of a suitable linearized counterpart Δ of Δ0 , and (iii) finally
assumes that each simplex in Δ has rational vertices. In this way, P becomes a rational
polyhedron P in euclidean space Rn , i.e., a finite union of simplexes S1 , . . . , Sk ⊆ Rn
with rational vertices. P need not be convex, nor connected [32].
The original recognition problem for combinatorial manifolds P and Q has been
transformed into an essentially equivalent problem for rational polyhedra, but P
and Q are disfigured into finite unions of rational simplexes, (like music is dis-
figured into score bars) and homeomorphisms are now replaced by rational PL-
homeomorphisms, i.e., invertible PL-maps φ such that every linear piece of both φ
and its inverse has rational coefficients.
By definition, a triangulation is rational if so are all its simplexes. Rational poly-
hedra are the same as underlying sets (supports) of rational triangulations [32]. Thus
they provide the following precise formulation of Markov theorem:

Theorem 2 (A.A. Markov, 1958, see [17, 30]) No Turing-computable procedure

can decide if two rational polyhedra P and Q are rationally PL-homeomorphic.

This result puts an end to the time-honored program of equipping every combi-
natorial manifold with a computable set of invariants sufficient to recognize home-
omorphic objects. The program was successful for curves and surfaces but fails for
higher-dimensional manifolds. To investigate the computability of homeomorphism,
manifolds are replaced by rational polyhedra, and homeomorphisms are replaced by
rational PL-homeomorphisms. In this way—at the very least—the recognizability
A Geometric Approach to MV-Algebras 61

problem becomes recursively enumerable: some Turing machine can effectively enu-
merate all pairs of rationally PL-homeomorphic rational polyhedra.

Are rational PL-maps the only reasonable arrows for rational polyhedra?

As problem instances in computability theory are coded by finite strings of sym-

bols, in algorithmic complexity theory the length of these input strings are related
to the time needed to compute the output. Accordingly, in any category of rational
polyhedra where space complexity is to have a role, it is natural to assume that invert-
ible arrows between two rational polyhedra P and Q preserve the space complexity
of the strings representing P and Q.
The following is a precise definition: For every point y = (y1 , . . . , yn ) ∈ Qn let
us denote by den(y) the least common denominator of the coordinates of y. We say
that den(y) is the denominator of y.
The vector ỹ = (den(y) · y1 , . . . , den(y) · yn , den(y)) ∈ Zn+1 is called the
homogeneous correspondent of y. Given two rational polyhedra P ⊆ [0, 1]n and
Q ⊆ [0, 1]m , a rational PL-homeomorphism η of P onto Q is said to be a Z-
homeomorphism if den(x) = den(η(x)) for each rational point x ∈ P. Equivalently,
[27], each linear piece of both η and η−1 has integer coefficients. (The number of
linear pieces of η is always finite.)
At the end of the day we are left with a category of rational polyhedra where
arrows are given by integer PL-maps, for short Z-maps, i.e., piecewise linear maps
ζ : P → Q such that every linear piece of ζ has integer coefficients, [27]. Then
Z-homeomorphisms coincide with those invertible maps η from a rational polyhe-
dron P ⊆ Rn onto a rational polyhedron Q ⊆ Rm such that both η and its inverse
are Z-maps. A new geometry arises, where the affine group over the integers has the
same role as that of the isometry group in euclidean space, [9].
Differently from the category of rational polyhedra with rational PL-maps, when
rational PL-maps are specialized to integer PL-maps a wealth of new geometric
computable invariants for any rational polyhedron P emerges: the number nd of
points of denominator d lying in P, d = 1, 2, . . . ; the number of simplexes in the
smallest regular triangulation of P (see below for the definition of regularity); the
smallest n such that P is Z-embeddable into Rn with preservation of denominators;
the rational volume of P (to be defined later on in this paper).

These invariants make the Z-homeomorphism of rational polyhedra more eas-

ily recognizable than rational PL-homeomorphism, just like the music we listen
to is better recognizable than the music we see coded on a pianola roll. Modulo
the dualities described in the next section, both finitely presented MV-algebras
and unital -groups inherit these invariants—although the latter need not be
immediately apparent within the purely algebraic framework.
62 D. Mundici

The counterpart of Markov’s unrecognizability theorem for this new category of

rational polyhedra is still open: it is not known whether the Z-homeomorphism of
rational polyhedra is a decidable problem.

3 The Rational Measure of Rational Polyhedra

In this section, we introduce the Z-homeomorphism invariant length, area, volume,…

of rational polyhedra.
Following [27], for any triangulation ∇ of P we denote by ∇ max the set of maximal
simplexes in ∇. 
For all i = 0, 1, 2, . . . we let P(i) = {T ∈ ∇ max | dim(T ) = i}, and we say that
P(i) is the i-dimensional part of P. P(i) is a (possibly empty) polyhedron and does not
depend on the triangulation ∇ of P. If P(i) is nonempty, then it is an i-dimensional
polyhedron. The j-dimensional part of such P(i) is empty iff j = i.
An m-simplex U = conv(w0 , . . . , wm ) ⊆ [0, 1]n is said to be regular (unimodular,
in [26]) if it is rational and the set of integer vectors {w̃0 , . . . , w̃m } (the homogeneous
correspondents of w0 , . . . , wm ) can be extended to a basis of the free abelian group
Zn+1 . A simplicial complex is said to be a regular triangulation (of its support) if
all its simplexes are regular. Regular triangulations are the affine counterparts of the
regular fans of toric algebraic geometry, [14, 36] (the “nonsingular fans” of [29]).
For every regular m-simplex T = conv(v0 , . . . , vm ) ⊆ Rn , m = 0, . . . , n, we use
the abbreviation den(T ) = den(v0 ) · · · den(vm ). Then the rational measure λ(T ) of
a regular k-simplex T in Rn is given by λ(T ) = (k! den(T ))−1 .
For any rational polyhedron P ⊆ Rn and regular triangulation Δ of P, therational
measure λ(i) Δ (P) of the i-dimensional part P
(i)
of P is given by λ(i) Δ (P) = {λ(S) |
dim(S) = i, S ∈ Δ }, where the sum equals zero if there are no maximal i-
max

simplexes in Δ.
The following result ensures that (i) every rational polyhedron has a regular tri-
angulation, (ii) the rational measure is independent of the chosen triangulation, and
(iii) is invariant under Z-homeomorphisms:
Theorem 3 ([26]) Let P be a rational polyhedron in Rn . We then have:

(a) P is the support of a regular triangulation.

(b) For i = 0, 1, . . . , n, and arbitrary regular triangulations Δ, ∇ of a rational
polyhedron P ⊆ Rn , we have the identity λ(i) (i)
Δ (P) = λ∇ (P). Thus we can write
λi (P) instead of λ(i)
Δ (P), and call λi (P) the i-dimensional rational measure of P.
(c) λi (P) is invariant under Z-homeomorphisms.

Proof (a) [28, Lemma 2.1]. The proof relies upon toric desingularization, [14, VI,
8.5], [29, pp. 23, 31].
A Geometric Approach to MV-Algebras 63

(b) [28, Theorem 2.3]. The proof follows from the Morelli-Włodarckzyk solution
of the weak Oda conjecture, [24, 36].
(c) [28, Theorem 1.1]. The proof follows from the De Concini-Procesi theorem
on elimination of points of indeterminacy in toric varieties, [29, p. 39]. 

Thus Z-homeomorphisms preserve not only the topological properties but also
the rational measure of rational polyhedra.

The rational measure has the following characterization:

Theorem 4 (a) Let L (n) and H (n) respectively denote n-dimensional Lebesgue and
Hausdorff measure, [13, 15]. Let P (n) denote the set of all rational polyhedra in Rn .
Then for each n = 1, 2, . . . and d = 0, 1, . . . , the map λd : P (n) → R≥0 has the
following properties, for all P, Q ∈ P (n) :

(i) (Invariance) If P = γ (Q) for some map γ belonging to the n-dimensional affine
group over the integers, then λd (P) = λd (Q).
(ii) (Valuation) λd (∅) = 0, λd (P) = λd (P(d) ), and the restriction of λd to the set
of all rational polyhedra P, Q in Rn having dimension at most d is a valuation:
λd (P) + λd (Q) = λd (P ∪ Q) + λd (P ∩ Q).
(iii) (Conservativity) For any P ∈ P (n) let (P, 0) = {(x, 0) ∈ Rn+1 | x ∈ P}. Then
λd (P) = λd (P, 0).
(iv) (Pyramid) For k = 1, . . . , n, if conv(v0 , . . . , vk ) is a regular k-simplex in Rn
with v0 ∈ Zn then λk (conv(v0 , . . . , vk )) = λk−1 (conv(v1 , . . . , vk ))/k.
(v) (Normalization) Let j = 1, . . . , n. Suppose the set B = {w1 , . . . , wj } ⊆ Zn is
part of a basis of the free
 abelian group Zn . Let the closed parallelepiped
 PB ⊆
j
R be defined by PB = x ∈ R | x = i=1 γi wi , 0 ≤ γi ≤ 1 . Then λj (PB ) =
n n

1.
(vi) (Proportionality) Let A be an m-dimensional rational affine subspace of Rn for
some m = 0, . . . , n. Then there is a constant κA > 0, only depending on A, such
that λm (Q) = κA · H (m) (Q) for every rational m-simplex Q ⊆ A.

∗∗∗

(b) The six properties above uniquely characterize the rational measures λo , . . . , λn ,
among all maps from P (n) to R≥0 , for each n = 1, 2, . . ..

Proof (a) [28, 4.2].

(b) [28, 8.2]. 
64 D. Mundici

4 Enter -Groups, Unital -Groups, and MV-Algebras

An -group is an (always abelian) group equipped with a translation invariant lattice

order. Baker and Beynon proved the following duality theorem:

Theorem 5 ([3–5]) The category of rational polyhedra with rational PL-maps is

dually equivalent to finitely presented -groups with their homomorphisms.

From the effectiveness of this duality we get the following equivalent reformula-
tion of Markov theorem:

Theorem 6 ([17]) The isomorphism problem for finitely presented -groups is

Turing-undecidable.

A unital -group is an -group with a distinguished positive archimedean ele-

ment. Just as -groups originate as a modern formalization of classical euclidean
magnitudes, unital -groups also take care of the (archimedean property of the) unit
of measurement. While the archimedean property is undefinable in first-order logic,
the following result yields an equational counterpart of unital -groups:

Theorem 7 ([25]) There is a categorical equivalence Γ between unital -groups

and MV-algebras.

Among others, this result allows us to speak of “finitely presented” unital -

groups, as the Γ -correspondents of finitely presented MV-algebras—which turn out
to coincide with finitely presented unital -groups in the sense of Gabriel and Ulmer,
[10, Remark 5.10], [22, Lemma 3.1].
Finitely presented MV-algebras and unital -groups have the following geometric
counterpart:

Theorem 8 ([21, 27]) The category of rational polyhedra with Z-maps is dually
equivalent to finitely presented MV-algebras with their homomorphisms. The duality
sends each rational polyhedron P ⊆ [0, 1]n to the MV-algebra M (P) = {f |`P | f ∈
M ([0, 1]n )}, the symbol “ |` ” denoting restriction.

Combining Γ with this duality we have:

Theorem 9 The category of rational polyhedra with Z-maps is dually equivalent to

finitely presented unital -groups with their unital -homomorphisms.

Summing up :

rational polyhedra with rational PL-maps rational polyhedra with integer PL-maps
=
finitely presented -groups finitely presented unital -groups
A Geometric Approach to MV-Algebras 65

5 Applying the Rational Measure to Projective

MV-Algebras

As a particular case of a general definition, an MV-algebra A is projective if whenever

ψ : B → C is a surjective homomorphism and φ : A → C is a homomorphism, there
is a homomorphism θ : A → B such that φ = ψ ◦ θ .
Finitely generated projective MV-algebras are an interesting subclass of finitely
presented MV-algebras: among others, they clarify such notions as exactness and
admissibility in the proof-theory of Łukasiewicz logic, [6, Sect. 4.5].
While Baker and Beynon [3–5] showed that an -group G is finitely generated
projective iff it is finitely presented, the situation is different for unital -groups and
MV-algebras. As shown by the following result, in combination with Theorem 8,
being finitely generated projective is a much stricter condition than being finitely
presented.

Theorem 10 ([8]) Let A be an n-generator projective MV-algebra. Then A is iso-

morphic to the MV-algebra M (P) obtained by restricting to P the functions of
M ([0, 1]n ), for some set P satisfying the following conditions:
(i) P is a rational polyhedron in [0, 1]n containing a vertex of the cube [0, 1]n ;
(ii) P is contractible;
(iii) For every regular triangulation Δ of P and maximal simplex T of Δ, the greatest
common divisor of the denominators of the vertices of T is equal to 1.

Through a further excursion in algebraic topology [18, 34, 35], Cabrer [7] has
recently shown that conditions (i)–(iii) are also sufficient for M (P) to be isomorphic
to an n-generator projective MV-algebra.
Property (iii) is known as the “strong regularity” of P, equivalently, its “anchored-
ness”, [58]. It is equivalent to asking that the affine hull of T contains an integer point
of Rn .
A folklore general result in universal algebra is to the effect that an n-generator
MV-algebra A is projective iff it is isomorphic to a retract R of the free n-generator
MV-algebra M ([0, 1]n ) of McNaughton functions over the unit n-cube [0, 1]n . Stated
otherwise, there is a retraction (idempotent endomorphism) ρ of M ([0, 1]n ) onto
R∼= A. Let us consider the following innocent looking problem:

Problem 1 What is the number of retractions of M ([0, 1]n ) onto R?

Note that this number is ≥1 precisely because A is projective. The answer is given
by Theorem 11 below, whose statement is surprisingly simple–although the proof
uses the rational measure of rational polyhedral in a fairly sophisticated way.
66 D. Mundici

For every finitely generated projective MV-algebra C we define the index ι(C) as
ι(C) = sup{number of retractions of M ([0, 1]n ) onto C  }, where n is the smallest
number of generators of C, and C  ranges over arbitrary retracts of M ([0, 1]n )
isomorphic to C.
An easy verification shows that ι(M ([0, 1]n )) = 1 for all n = 1, 2, . . . .
For every n-generator MV-algebra B, the construction introduced in [27, Corollary
4.18], yields a canonical (Yosida) homeomorphism of the maximal spectral space
μB onto a closed subset M of [0, 1]n . If M = cl(int(M)) then following Kuratowski,
[12, p. 20] we unambiguously say that μB is a closed domain in [0, 1]n .

Theorem 11 (L.M. Cabrer, D.M.) Let A be a finitely generated projective MV-

algebra. Let n be the smallest number of generators of A. Then the index of A is finite
iff the maximal spectral space of A is a closed domain in [0, 1]n .

The proof uses Theorem 10, along with the properties of the rational measure of
the maximal spectral space of A, (Theorem 3).
Last, but not least, another interesting application of the rational measure is in the
classification of orbits of affine subspaces of Rn under the action of the n-dimensional
affine group over the integers. This uses the orbit classification of [9] as a preliminary
step.

6 Appendix: Recent Applications of MV-Algebras

(a Selection)

As we have seen, MV-algebra theory heavily draws from algebraic topology and
toric geometry. Conversely, the book [27] shows that MV-algebras have many appli-
cations to diverse areas of mathematics. Here is a selection of recent developments
subsequent to the publication of [27]:

• Riesz spaces, [45, 48, 72]

• Differential geometry, [41, 42, 45, 69]
• Algebraic geometry, [39, 40]
• Categories, duality, sheafs, [10, 21, 37, 46, 47, 56, 65, 67]
• Semirings, tropical and idempotent mathematics, [37, 38, 49, 50, 55, 59]
• Probability, [52, 75]
• Games, [60–64, 66]
• Multisets, [46, 71]
• Semantics of Łukasiewicz logic, [68, 70]
• Proof-theory of Łukasiewicz logic, [6, 43, 58]
• Modal logic, Belief, [53, 54, 57, 64]
• Quantum structures, [51, 73, 74, 76, 77]
• Topological groups, [78]
• Discrete dynamical systems, [9]
• Interval Algebras, [44].
A Geometric Approach to MV-Algebras 67

All these interactions between algebraic, geometric, measure-theoretic, logic-

algorithmic notions are typical of mathematics. The latter is pervaded by functors
that connect one part with another, and transfer information, as blood circulation
does in a living body. In this paper we have just seen the action of functors on finitely
presented MV-algebras and rational polyhedra.

Acknowledgments I am grateful to my friend Peter Klement, whose many papers and books
[20, and references therein] taught me the importance of t-norms, and whose kind hospitality at
Magdalena Bildungshaus allowed me to get in contact with a community of mathematicians—of
which he has been for decades one of the focal points—involved in all aspects of fuzzy logic.

References

1. Baczyński, M.: Residual implications revisited. Notes Smets-Magrez Theorem, Fuzzy Sets
Syst. 145, 267–277 (2004)
2. Baczyński, M., Jayaram, B.: (S, N)- and R-implications: a state-of-the-art survey. Fuzzy Sets
Syst. 159, 1836–1859 (2008)
3. Baker, K.A.: Free vector lattices. Can. J. Math. 20, 58–66 (1968)
4. Beynon, W.M.: On rational subdivisions of polyhedra with rational vertices. Can. J. Math. 29,
238–242 (1977)
5. Beynon, W.M.: Applications of duality in the theory of finitely generated lattice-ordered abelian
groups. Can. J. Math. 29, 243–254 (1977)
6. Cabrer, L.M.: Simplicial geometry of unital lattice-ordered abelian groups. Forum Math. 27,
1309–1344 (2015). doi:10.1515/forum-2011-0131
7. Cabrer, L.M.: Rational simplicial geometry and projective lattice-ordered abelian groups.
arXiv:1405.7118v1 [math.RA] 28 May 2014
8. Cabrer, L.M., Mundici, D.: Rational polyhedra and projective lattice-ordered abelian groups
with order unit. Commun. Contemp. Math. 14(3), 1250017 (20 pages) (2012). doi:10.1142/
S0219199712500174
9. Cabrer, L.M., Mundici, D.: Classifying orbits of the affine group over the integers, to appear
in Ergodic Theory Dyn. Syst. doi:10.1017/etds.2015.45
10. Caramello, O., Russo, A.C.: The Morita-equivalence between MV-algebras and lattice-ordered
abelian groups with strong unit. J. Algebra 422, 752–787 (2015)
11. Cignoli, R., D’Ottaviano, I.M.L., Mundici, D.: Algebraic Foundations of Many-Valued Rea-
soning, Trends in Logic, vol. 7. Kluwer, Dordrecht (2000)
12. Engelking, R.: General Topology, Revised and completed edition, Sigma Series in Pure Math-
ematics, vol. 6. Heldermann Verlag, Berlin (1989)
13. Evans, L.C., Gariepy, R.F.: Measure Theory and Fine Properties of Functions. CRC Press,
Boca Raton (1992)
14. Ewald, G.: Combinatorial Convexity and Algebraic Geometry. Springer, New York (1996)
15. Federer, H.: Geometric Measure Theory. Springer, New York (1969)
16. Fodor, J., Roubens, M.: Fuzzy Preference Modeling and Multicriteria Decision Support. Kluwer
Academic Publishers, Dordrecht (1994)
17. Glass, A.M.W., Madden, J.J.: The word problem versus the isomorphism problem. J. Lond.
Math. Soc. (2), 30, 53–61 (1984)
18. Hatcher, A., Algebraic Topology. Cambridge University Press (2001)
19. Łukasiewicz, J., Tarski, A.: Untersuchungen über den Aussagenkalkül, Comptes Rendus des
séances de la Société des Sciences et des Lettres de Varsovie, Classe III, 23, pp. 30–50 (1930).
English translation: Investigations into the Sentential Calculus, Chapter IV. In: A. Tarski, Logic,
Semantics, Metamathematics. Clarendon Press, Oxford (1956). Reprinted: Hackett, Indianapo-
lis (1983)
68 D. Mundici

20. Klement, E.P., Mesiar, R., Pap, E.: Triangular Norms. Kluwer, Dordrecht (2000)
21. Marra, V., Spada, L.: Duality, projectivity, and unification in Łukasiewicz logic and MV-
algebras. Ann. Pure Appl. Logic 164, 192–210 (2013)
22. Marra, V., Spada, L.: Two isomorphism criteria for directed colimits. arXiv:1312.0432v1, 2
Dec 2013
23. Menu, J., Pavelka, J.: A note on tensor products on the unit interval. Comment. Math. Univ.
Carol. 17, 71–83 (1976)
24. Morelli, R.: The birational geometry of toric varieties. J. Algebraic Geom. 5, 751–782 (1996)
25. Mundici, D.: Interpretation of AF C ∗ -algebras in Łukasiewicz sentential calculus. J. Funct.
Anal. 65, 15–63 (1986)
26. Mundici, D.: The Haar theorem for lattice-ordered abelian groups with order-unit. Discret.
Contin. Dyn. Syst. 21, 537–549 (2008)
27. Mundici, D.: Advanced Łukasiewicz calculus and MV-algebras. Trends in Logic, vol. 35.
Springer, Berlin (2011)
28. Mundici, D.: Invariant measure under the affine group over Z, combinatorics. Probab. Comput.
23, 248–268 (2014)
29. Oda, T.: Convex bodies and algebraic geometry. Convex Bodies and Algebraic Geometry.
Springer, New York (1988)
30. Shtan’ko, M.A.: Markov’s theorem and algorithmically non-recognizable combinatorial man-
ifolds, Izvestiya RAN. Ser. Math. 68, 207–224 (2004)
31. Smets, P., Magrez, P.: Implication in fuzzy logic. Int. J. Approx. Reason. 1, 327–347 (1987)
32. Stallings, J.R.: Lectures on Polyhedral Topology. Tata Institute of Fundamental Research,
Mumbay (1967)
33. Trillas, E., Valverde, L.: On implication and indistinguishability in the setting of fuzzy logic.
In: Kacprzyk, J., Yager, R.R. (eds.) Management Decision Support Systems using Fuzzy Sets
and Possibility Theory, pp. 198–212. Technical University Rhineland, Cologne (1985)
34. Whitehead, J.H.C.: On subdivisions of complexes. Math. Proc. Camb. Philos. Soc. 31, 69–75
(1935)
35. Whitehead, J.H.C.: Simplicial spaces, nuclei and m-groups. Proc. Lond. Math. Soc. 45, 243–
327 (1939)
36. Włodarczyk, J.: Decompositions of birational toric maps in blow-ups and blow-downs. Trans.
Am. Math. Soc. 349, 373–411 (1997)

Additional Recent Literature Cited in Section 5

37. Belluce, L.P., Di Nola, A., Ferraioli, A.R.: MV-semirings and their sheaf representations. Order
30, 165–179 (2013). doi:10.1007/s11083-011-9234-0
38. Belluce, L.P., Di Nola, A., Ferraioli, A.R.: Ideals of MV-semirings and MV-algebras. In: Litvi-
nov, G.L., Sergeev, S.N. (eds.) Tropical and Idempotent Mathematics and Applications. Con-
temporary Mathematics, vol. 616, pp. 59–76 (2014)
39. Belluce, L.P., Di Nola, A., Lenzi, G.: On generalizing the Nullstellensatz for MV-algebras. J.
Logic Comput. 25, 701–707 (2015). doi:10.1093/logcom/exu042
40. Belluce, L.P., Di Nola, A., Lenzi, G.: Algebraic geometry for MV-algebras. J. Symb. Logic
79(4), 1061–1091 (2014)
41. Busaniche, M., Mundici, D.: Bouligand-Severi tangents in MV-algebras. Revista Matemática
Iberoamericana 30(1), 191–201 (2014)
42. Cabrer, L.M.: Bouligand-Severi k-tangents and strongly semisimple MV-algebras. J. Algebra
404, 271–283 (2014)
43. Cabrer, L.M.: Exact Unification. arXiv:1410.5583v1 [math.LO] 21 Oct 2014
44. Cabrer, L.M., Mundici, D.: Interval MV-algebras and generalizations. Int. J. Approx. Reason.
55, 1623–1642 (2014)
A Geometric Approach to MV-Algebras 69

45. Cabrer, L.M., Mundici, D.M.: Severi-Bouligand tangents, Frenet frames and Riesz spaces.
Adv. Appl. Math. 64, 1–20 (2015)
46. Cignoli, R., Marra, V.: Stone duality for real-valued multisets. Forum Math. 24, 1317–1331
(2012)
47. Di Nola, A., Ferraioli, A.R., Lenzi, G.: Algebraically closed MV-algebras and their sheaf
representation. Ann. Pure Appl. Logic 164, 349–355 (2013)
48. Di Nola, A., Leustean, I.: Łukasiewicz logic and Riesz spaces. Soft Comput. 18, 2349–2363
(2014). doi:10.1007/s00500-014-1348-z
49. Di Nola, A., Russo, C.: Semiring and semimodule issues in MV-algebras. Comm. Algebra 41,
1017–1048 (2013)
50. Di Nola, A., Russo, C.: MV-semirings as a new perspective on mathematical fuzzy set theory:
a survey. arXiv:1102.1999v4, 14 Nov 2014
51. Dvurečenskij, A.: Quantum structures versus partially ordered groups. Int. J. Theor. Phys.
doi:10.1007/s10773-014-2479-9
52. Fedel, M., Keimel, K., Montagna, F., Roth, W.: Imprecise probabilities, bets and functional
analytic methods in Łukasiewicz logic. Forum Math. 25, 405–441 (2013). doi:10.1515/FORM.
2011.123
53. Flaminio, T., Godo, L., Kroupa, T.: Belief functions on MV-algebras of fuzzy sets: an overview.
In: Torra, V., Narukawa, Y., Sugeno, M. (eds.) Non-Additive Measures, Studies in Fuzziness
and Soft Computing, vol. 310, pp. 173–200. Springer (2014)
54. Flaminio, T., Godo, L., Hosni, H.: Coherence in the aggregate: a betting method for belief
functions on many-valued events. Int. J. Approx. Reason. 58, 71–86 (2015). doi:10.1016/j.ijar.
2015.01.001
55. Gavalec, M., Nemcová, Z., Sergeev, S.: Tropical linear algebra with the Łukasiewicz T-norm.
Fuzzy Sets Syst. 276, 131–148 (2015). doi:10.1016/j.fss.2014.11.008
56. Gehrke, M., van Gool, S.J., Marra, V.: Sheaf representations of MV-algebras and lattice-ordered
abelian groups via duality. J. Algebra 417, 290–332 (2014)
57. Hansoul, G., Teheux, B.: Extending Łukasiewicz logics with a modality: algebraic approach
to relational semantics. Stud. Logica 101, 505–545 (2013)
58. Jeřábek, E.E.: The complexity of admissible rules of Łukasiewicz logic. J. Logic Comput. 23,
693–705 (2013)
59. Kala, V.: Lattice-ordered abelian groups finitely generated as semirings, to appear in the J.
Commut. Algebra. arXiv:1502.01651
60. Kroupa, T.: Core of coalition games on MV-algebras. J. Logic Comput. 21, 479–492 (2011)
61. Kroupa, T.: A generalized Möbius transform of games on MV-algebras and its application to a
Cimmino-type algorithm for the core, optimization theory and related topics. Contemp. Math.
568, 139–158 (2012)
62. Kroupa, T.: States in Łukasiewicz logic correspond to probabilities of rational polyhedra. Int.
J. Approx. Reason. 53, 435–446 (2012)
63. Kroupa, T., Majer, O.: Optimal strategic reasoning with McNaughton functions. Int. J. Approx.
Reason. 55, 1458–1468 (2014)
64. Kroupa, T., Teheux, B.: Modal extension of Łukasiewicz logic for reasoning about coalitional
power. arXiv:1411.6452v1, 24 Nov 2014
65. Lawson, M.V., Scott, P.: AF inverse monoids and the structure of countable MV-algebras.
arXiv:1408.1231v2, 13 Oct 2014
66. Marchioni, E., Woolridge, M.: Łukasiewicz games, In: Huhns (eds.) Proceedings of the 13th
International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2014),
Paris, France, pp. 837–844 (2014)
67. Marra, V., Spada, L.: The dual adjunction between MV-algebras and Tychonoff spaces. Studia
Logica, special issue in memoriam Leo Esakia 100, 253–278 (2012)
68. Mundici, D.: The differential semantics of Łukasiewicz syntactic consequence, Chapter 7. In:
Montagna, F. (ed.) Petr Hájek on Mathematical Fuzzy Logic, Outstanding Contributions, vol. 6,
pp. 143–157. Springer International Publishing Switzerland (2015). doi:10.1007/978-3-319-
06233-4
70 D. Mundici

69. Mundici, D., Pedrini, A.: The Euler characteristic and valuations on MV-algebras. Math. Slo-
vaca 64, 563–570 (2014). doi:10.2478/s12175-014-0226-6
70. Mundici, D., Picardi, C.: Faulty sets of Boolean formulas and Łukasiewicz logic.J. Logic
Comput., Adv. Access published Dec 8 (2014). doi:10.1093/logcom/exu073
71. Nganou, J.B.: Profinite MV-algebras and multisets. Order, doi:10.1007/s11083-014-9345-5
72. Pedrini, A.: The Euler characteristic of a polyhedron as a valuation on its coordinate vector
lattice. arXiv:1209.3248v1, 14 Sep 2012
73. Pulmannová, S.: Representations of MV-algebras by Hilbert-space effects. Int. J. Theoret. Phys.
52, 2163–2170 (2013)
74. Pulmannová, S., Vinceková, E.: MV-pairs and state operators. Fuzzy Sets Syst. 260, 62–76
(2015)
75. Riečan, B.: Variation on a Poincaré theorem. Fuzzy Sets Syst. 232, 39–45 (2013)
76. Xie, Y., Li, Y., Yang, A.: The pasting construction for effect algebras. Math. Slovaca 64, 1051–
1074 (2014). doi:10.2478/s12175-014-0258-y
77. Shang, Y., Lu, X., Lu, R.: Computing power of turing machines in the framework of unsharp
quantum logic. Theoret. Comput. Sci. 598, 2–14 (2015). doi:10.1016/j.tcs.2014.12.015
78. Weber, H.: On topological MV-algebras and topological -groups. Topology Appl. 159, 3392–
3395 (2012)
On the Equational Characterization
of Continuous t-Norms

Francesc Esteva and Lluís Godo

Abstract A (continuous) t-norm is called equationally definable when the corre-

sponding standard BL-algebra [0, 1]∗ defined by ∗ and its residuum is the only (up to
isomorphism) standard BL-algebra that generates the same variety V ar ([0, 1]∗ ). In
this chapter we check that a continuous t-norm ∗ is equationally definable if and only
if the t-norm is a finite ordinal sum of copies of the three basic continuous t-norms,
i.e. Łukasiewicz, Gödel and Product t-norms.

1 Introduction

A core constituent of fuzzy logic in narrow sense [15], from where the discipline
of Mathematical fuzzy logic has been intensively developed in the last two decades
[5, 10, 11, 14], is the family of residuated many-valued logical calculi with truth
values on the real unit interval [0, 1], and with min, max, a (left-continuous) t-norm
∗ and its residuum →∗ as basic truth functions, interpreting respectively the lattice
meet and joint connectives, a strong conjunction and its adjoint implication. These
logics are also known as t-norm based fuzzy logics.
In this framework, Hájek introduced in [11, 12] the so-called Basic Fuzzy logic,
BL for short, to capture the 1-tautologies common to all many-valued calculi in [0, 1]
defined by a continuous t-norm and its residuum, as proved in [4]. Thus, BL is in
fact a common sublogic of three well-known fuzzy logics: Łukasiewicz’s infinitely-
valued logic, Gödel’s infinitely-valued logic and Product logic, corresponding to the
three basic t-norms, i.e. Łukasiewicz, minimum and product t-norms.

F. Esteva · L. Godo (B)

Artificial Intelligence Research Institute (IIIA - CSIC), Campus UAB,
08193 Bellaterra, Spain
e-mail: [email protected]
F. Esteva
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 71

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_5
72 F. Esteva and L. Godo

The variety of BL-algebras constitutes the algebraic semantics of Hájek’s BL,

which is generated by the so-called standard BL-algebras [0, 1]∗ , that is, the
BL-algebras defined on the real unit interval [0, 1], and that in turn are induced
by continuous t-norms ∗ and their residuum →∗ . Some subvarieties of BL generated
by a single standard BL-chain [0, 1]∗ are well-known, in particular the subvarieties
of MV algebras, Gödel algebras and Product algebras, the algebraic counterparts of
Łukasiewicz, Gödel and Product logics respectively. These varieties are respectively
generated by the standard algebras defined by Łukasiewicz, minimum and product t-
norms, and are fully described and equationally characterized in the literature. A step
further was done in [8], where all varieties V ar ([0, 1]∗ ) of BL-algebras generated
by a single standard BL-chain [0, 1]∗ was proved to be finitely axiomatizable.
Then the question arises of whether such an axiomatization of V ar ([0, 1]∗ )
(i.e. a set of equations) univocally characterizes ∗ itself, in the sense of whether
[0, 1]∗ is the only (up to isomorphism) standard BL-algebra that generates the same
variety V ar ([0, 1]∗ ). When this is so, we say that ∗ is equationally definable.
As a rather direct consequence of results in [8], in this short note, and after
introducing some needed preliminaries, we check in Sect. 3 that a continuous t-norm
is equationally definable if and only if the t-norm is a finite ordinal sum of the
three basic continuous t-norms, while in Sect. 4 we show how to effectively find a
set of equations of V ar ([0, 1]∗ ) for an arbitrary equationally definable continuous
t-norm ∗.

2 Preliminaries

We start with some elementary and well-known definitions and results about t-norms,
just for the sake of the paper being self-contained. A t-norm is a binary operation
on [0, 1] that is commutative, associative, non-decreasing (monotone) in both vari-
ables and that have 0 as absorbent and 1 as unity. A t-norm is continuous if it is
continuous as real function of two variables. The three basic continuous t-norms are
minimum (min), product (the usual product of reals, ) and Łukasiewicz (denoted
∗ L and defined by x ∗ L y = max(0, x + y − 1)). The greatest and smallest continu-
ous t-norms are the minimum and the Łukasiewicz t-norms respectively, i.e., for all
continuous t-norm ∗ and for all x, y ∈ [0, 1], we have x ∗ L y ≤ x ∗ y ≤ min(x, y).
The following are some basic results on continuous t-norms, see e.g. [13] for
further details and results:
• Any continuous t-norm is an ordinal sum of (possibly infinitely-many) copies1 of
the minimum, product and Łukasiewicz t-norms.
• A t-norm ∗ is continuous if and only if it satisfies the divisibility condition: for all
x, y ∈ [0, 1] with x > y there exists z ∈ [0, 1] such that y = x ∗ z.

1 If
we allow for at most a countable number of degenerated components with a single idempotent
element.
On the Equational Characterization of Continuous t-Norms 73

• Each left-continuous t-norm ∗ uniquely defines a binary operation →∗ , called

the residuum of ∗, that satisfies the following condition: for all x, y, z ∈ [0, 1],
x ∗ y ≤ z if and only if x ≤ y →∗ z (residuation or adjunction condition).
• The residuum →∗ of a left-continuous t-norm ∗ is actually defined as x →∗ y =
max{z ∈ [0, 1] : x ∗ z ≤ y} (residuated implication).
• A left-continuous t-norm ∗ is continuous if and only if the following equation is
satisfied: for all x, y ∈ [0, 1], x ∗ (x →∗ y) = min(x, y) (Divisibility equation).
On the oher hand, it is also well known that the algebraic counterpart of Hájek’s
BL logic [11] is given by the variety of BL-algebras, i.e. algebraic structures A =
(A, ∧, ∨, ∗, →, 0, 1) satisfying:
• (A, ∧, ∨, 0, 1) is a bounded distributive lattice,
• (A, ∗, 1) is a commutative monoid with unit 1,
• ∗ and → form and adjoint pair, i.e. they satisfy the residuation condition: for all
x, y, z ∈ A, x ∗ y ≤ z if and only if x ≤ y → z,
• Prelinearity: for all x, y ∈ A, (x → y) ∨ (y → x) = 1,
• Divisibility: for all x, y ∈ A, x ∗ (x → y) = x ∧ y.
In other words, BL-algebras are a subclass of residuated lattices, namely, the class
of bounded, commutative, integral residuated lattices further satisfying pre-linearity
and divisibility.
A standard BL-chain is a BL-algebra defined over the real unit interval [0, 1]. It
is easy to prove that:
• A continuous t-norm and its residuum defines a standard BL-chain,
• Each standard BL-chains is defined by a continuous t-norm and its residuum.
The last items shows that there is a bijection between continuous t-norms and
standard BL-chains. From now on, we will denote by [0, 1]∗ the BL-algebra
([0, 1], min, max, ∗, →∗ , 0, 1) defined by a continuous t-norm ∗ and its residuum.
The ordinal sum representation for continuous t-norms extends to an ordinal sum
representation for standard BL-chains in the obvious way, the only new thing to
consider is the definition of the residuum over the whole ordinal sum in terms of
the residuum over each component. Using a similar representation for BL-chains, in
[4] it was proved that the logic BL is complete with respect to the class of standard
BL-chains, or in other words, that the whole variety of BL-algebras is generated by
the class of standard BL-chains.
A related class of algebraic structures is that of hoops. In what follows we introduce
some basic definitions and results about hoops and the decomposition theorem for
BL-chains as ordinal sums of hoops that we will use in the next section, see [2, 3, 9]
for more details.

Definition 1 A hoop is an algebraic structure A = (A, ∗, →, 1) such that:

• ∗ is a binary commutative operation with unit 1, i.e. x ∗ y = y ∗ x and 1 ∗ x = x
for all x, y ∈ A
74 F. Esteva and L. Godo

• → is a binary operation satisfying:

– for all x ∈ A, x → x = 1,
– for all x, y, z ∈ A, (x ∗ y) → z = x → (y → z),
– for all x, y ∈ A, x ∗ (x → y) = y ∗ (y → x).
The associated order relation is defined by: x ≤ y if x → y = 1.
A basic hoop is a hoop satisfying the following condition:
• ((x → y) → z) ∗ (y → x) → z) → z = 1
A Wajsberg hoop is a hoop satisfying the following condition:
• for all x, y ∈ A, (x → y) → y = (y → x) → x.
A cancellative hoop is a hoop such that:
• for all x, y, z ∈ A, x ∗ y ≤ x ∗ z implies that y ≤ z.

From this definition, one can check the following facts and properties:
(i) ≤ as defined above is indeed an ordering and 1 is maximal
(ii) ∗ is associative
(iii) ∗ is monotonically increasing w.r.t. ≤: x ≤ y implies x ∗ z ≤ y ∗ z
(iv) (∗, →) is an adjoint pair: x → y ≤ z iff x ∗ y ≤ z
(v) x ∗ (x → y) ≤ y
(vi) 1→x =x
Furthermore, regarding the classes of basic, Wajsberg and cancellative hoops, the
following relationship among them hold: every Wajsberg hoop is basic and each
cancellative hoop is Wajsberg (hence basic as well). Note that hoops have an greatest
element, but they may lack a least element. A hoop A = (A, ∗, →, 1) is called
bounded if (A, ≤) has a least element. Then it turns out that cancellative hoops
coincide with unbounded Wajsberg hoops, while bounded Wajsberg hoops coincide
with MV-algebras.
Prominent examples of Wajsberg hoops are the following:
• 2, defined on a set of two elements {a, 1}, that is in fact a two-element Boolean
algebra.
• Ł = ([0, 1], ∗Ł , →Ł , 1), the (bounded) Wajsberg hoop defined over [0, 1] by the
Łukasiewicz t-norm and its residuum.
• C = ((0, 1], , → , 1), the (unbounded) cancellative hoop defined over (0, 1] by
the product t-norm and its residuum.
A similar construction to the ordinal sums for t-norms and BL-chains can be also
defined for hoops.

Definition 2 Let (I, ≤) be a totally ordered set, and for all i ∈ I let Ai = (Ai , ∗i ,
→i , 1) be a hoop such thatAi ∩ A j = {1}
 for every j
= i. Then the ordinal sum of
this family is the structure i∈I Ai = ( i∈I Ai , ∗, →, 1), where the operations are
defined as follows:
On the Equational Characterization of Continuous t-Norms 75
⎧
⎪
⎨x ∗i y if x, y ∈ Ai ,
x ∗ y := x if x ∈ Ai \{1}, y ∈ A j , and i < j,
⎪
⎩
y if y ∈ Ai \{1}, x ∈ A j , and i < j.
⎧
⎪
⎨x →i y if x, y ∈ Ai ,
x → y := y if x ∈ Ai , y ∈ A j , and i > j,
⎪
⎩
1 otherwise.

Notice that in an ordinal sum of hoops, the greatest element is common to all
the hoops and to the ordinal sum as well. For instance, the product standard chain
[0, 1]Π = ([0, 1], , → , 0, 1), viewed as a bounded hoop, can be decomposed as
the ordinal sum of 2 and C, i.e. [0, 1]Π = 2 ⊕ C. Actually, in [1] the authors prove
that any BL-chain, viewed as a bounded basic hoop, can be decomposed as an ordinal
sum of linearly ordered Wajsberg hoops. Restricted to standard BL-chains, this result
amounts to say that any standard BL-chain, as a hoop, can be decomposed as an
ordinal sum of (suitably arranged) copies of the Wajsberg hoops 2, C and Ł. In this
way, besides viewing the standard product algebra as the ordinal sum of 2 plus C, we
can understand the standard Gödel chain as being isomorphic to the ordinal sum of
continuum many of copies of 2 (one for each element of a Gödel component), while
the standard Łukasiewicz chain [0, 1]Ł = ([0, 1], ∗Ł , →Ł , 0, 1) coincides with Ł as
hoop.
As already mentioned, regarding the ordinal sums of hoops just defined, one can
notice that the main difference with respect to the ordinal sum of BL-chains is that
the top elements of the components are identified with the top element of the ordinal
sum. Therefore, for instance, when considering the decomposition of a BL-chain as
an ordinal sum of Wajsberg hoops (2, C or Ł in the case of standard BL-chains),
the top of any component is the top of the ordinal sum, and given two consecutive
components, the bottom (if it exists) of the second component is not in the first
component. Notice also that the decomposition of any standard BL-chain as ordinal
sum of hoops has always a first component that is either 2 (if it is an SBL-chain2 ) or
Ł otherwise.
Finally recall that a set of equations determine a variety (or equational class) of
algebraic structures. By inspecting their definition, it is clear that the classes of hoops,
basic hoops and Wajsberg hoops are indeed varieties. The class of cancellative hoops
turns out to be a variety as well, since the condition used in Definition 2 can be shown
to be equivalent to the validity of the equation x = y → (y ∗ x).
Thus it is interesting to know how the varieties generated by the main three
prominent Wajsberg hoops, 2, C and Ł, are related to each other. To do so we consider
the following three terms:
• eŁ (x) = (x → x 2 ) ∨ ((x → x 3 ) → x 2 )
• eC (x) = (x → x 2 )
• e2 (x) = (x → x 3 ) → x 2

2 That is, a standard BL-chain defined by an strict continuous t-norm.

76 F. Esteva and L. Godo

where x n stands for x∗ . n. . ∗x. An easy computation shows that the equation
eŁ (x) = 1 is valid in 2 and C and not in Ł, eC (x) = 1 is a valid equation in 2
and neither in C nor in Ł, and finally, the equation e2 (x) = 1 is valid in C and neither
in 2 nor in Ł.
Therefore, it is clear that C and Ł do not belong to variety of hoops V ar (2)
generated by 2, while 2 and Ł do not belong to the variety V ar (C) generated by C.
On the other hand, it is easy to check that both 2 and C belong to the variety of
hoops V ar (Ł) generated by Ł, since 2 is a subhoop of Ł and C is a subhoop of the
well-known Chang algebra, which is an MV-algebra, and thus belongs to V ar (Ł).
Summarising, we have

2, C ∈ V ar (Ł), C, Ł ∈
/ V ar (2), 2, Ł ∈
/ V ar (C),

and thus, the following strict inclusions among varieties hold:

V ar (2) ⊂ V ar (Ł), V ar (C) ⊂ V ar (Ł).

3 Characterization of Standard BL-Chains

that Are Equationally Definable

Let us denote by [0, 1]∗ either the standard BL-chain, or its corresponding hoop when
no confusion exists, defined over [0, 1] by a continuous t-norm ∗ and its residuum →∗ .
The goal of this section is to characterize those continuous t-norms ∗ that admit an
equational characterization in the sense that the variety V ar ([0, 1]∗ ) is uniquely
generated by [0, 1]∗ , that is, for any other standard BL-chain [0, 1]◦ with ◦ being a
t-norm non isomorphic to ∗, V ar ([0, 1]∗ )
= V ar ([0, 1]◦ ). In such a case, we can
say that the set of equations defining V ar ([0, 1]∗ ) characterize ∗.
Actually, generalizing the well-known Mostert and Shields representation theo-
rem of continuous t-norms, Hájek showed in [12] that every standard BL-chain [0, 1]∗
can be isomorphically decomposed as an ordinal sum (over a bounded ordered index
set) of Gödel, Łukasiewicz and Product BL-chain components. However, as hoops,
each Gödel BL-chain is isomorphic to an ordinal sum of (possibly infinite) copies
of 2, while Łukasiewicz and Product components on a closed real interval are iso-
morphic to Ł and Π = 2 ⊕ C respectively. Then any standard BL-chain, as a hoop,
will be isomorphic to a (possibly infinite) ordinal sum of Wajsberg hoops Ł, C and 2.
The following definition and proposition are particular cases of more general
definitions and results given in [8], and therefore here we only state them without
proofs.

Definition 3 (i) We will denote by Fin the set of ordinal sums (as hoops) of
finitely-many copies of Ł, 2 and C, and whose first component is either Ł or 2.
On the Equational Characterization of Continuous t-Norms 77

(ii) Let A be a standard

 BL-chain whose decomposition as ordinal sum of hoops
is
 A = A 0 ⊕ ( i∈I Ai ). Then Fin(A) is the set of all finite ordinal sums
B
i=0,...,n i of Wajsberg hoops satisfying the following conditions:
• Each Bi is either 2, C or Ł,
• B0 is either 2 or Ł,
• There are components A0 < A1 < · · · < An of A such that for every i =
0, . . . , n: (i) if Bi = Ł then Ai is isomorphic to Ł; (ii) if Bi = C, then Ai is
isomorphic either to C or to Ł; and (iii) if Bi = 2, then Ai is isomorphic either
to 2 or to Ł.

Example 1 Consider the standard BL-chain A = G ⊕ Ł ⊕ Π . Then, for instance,

2 ⊕ Ł and 2 ⊕ 2 ⊕ Ł ⊕ C are in Fin(A), while neither Ł ⊕ B for any B ∈ Fin, nor
2 ⊕ Ł ⊕ Ł are in Fin(A).

As shown next, the set of Fin([0, 1]∗ ) of BL-chains univocally determines the
variety V ([0, 1]∗ ) induced by the t-norm ∗.

Proposition 1 (c.f. Theorem 3.9 of [8]) Let [0, 1]∗ , [0, 1]◦ be two standard
BL-chains. Then V ar ([0, 1]∗ ) ⊆ V ar ([0, 1]◦ ) if, and only if, Fin([0, 1]∗ ) ⊆
Fin([0, 1]◦ ). Hence, V ar ([0, 1]∗ ) = V ar ([0, 1]◦ ) if, and only if, Fin([0, 1]∗ ) =
Fin([0, 1]◦ ).

Notation convention: In the following, given two continuous t-norms ∗ and ◦, we

will write ∗ ≡ ◦ to denote that they isomorphic in the usual sense of t-norms, that
is, when there exists an increasing bijection f : [0, 1] → [0, 1] such that, for any
x, y ∈ [0, 1], x ◦ y = f −1 ( f (x) ∗ f (y)).
The following lemma is straightforward to check.

Lemma 1 If ∗ and ◦ are two continuous t-norms such that both [0, 1]∗ and [0, 1]◦
have a finite ordinal sum decomposition in terms of BL-components, then ∗ ≡ ◦ if,
and only if, they have the same decomposition,

From the above proposition and lemma, the characterization of the equationally
definable standard BL-chains follows.

Proposition 2 A continuous t-norm ∗ admits an equational characterization if, and

only if, the corresponding standard BL-chain [0, 1]∗ can be decomposed as an ordinal
sum with finitely-many copies of components Ł, G and Π .

Proof First we prove that for a continuous t-norm ∗ whose decomposition as ordinal
has a finite number of components, V ar ([0, 1]◦ ) = V ar ([0, 1]∗ ) if and only if ◦ ≡
∗ (the components of their decomposition as ordinal sums are the same). By the
previous proposition, this is equivalent to prove that if ◦ is a continuous t-norm such
that ◦
≡ ∗, then Fin(◦)
= Fin(∗). We prove this claim by cases, adapting a more
general proof in [8]:
78 F. Esteva and L. Godo

• If the decomposition of [0, 1]◦ has more components than the decomposition of
[0, 1]∗ then it is evident that there exist BL-chains in Fin(◦) that are not in Fin(∗).
For example let ◦ be a continuous t-norm obtained as Ł ⊕ G, and let ∗ be a contin-
uous t-norm obtained as Ł ⊕ Π ⊕ G. Then it is clear that 2 ⊕ C ∈ Fin([0, 1]∗ )
but 2 ⊕ C ∈ / Fin([0, 1]◦ ).
• An analogous reasoning proves the statement when the decomposition of [0, 1]◦
has more components than the decomposition of [0, 1]∗ .
• If the number of components of the decomposition [0, 1]∗ and [0, 1]◦ is the same,
then they need to differ in some component and thus we can find BL-chains that are
in Fin(∗) and not in Fin(◦) and viceversa. For example, let ◦ be the continuous
t-norm obtained as Ł ⊕ Ł ⊕ G and let ∗ be the continuous t-norm obtained as
Ł ⊕ Π ⊕ G. Then we have that 2 ⊕ C ∈ Fin([0, 1]∗ ) but 2 ⊕ C ∈ / Fin([0, 1]◦ ),
while Ł ⊕ Ł ∈ Fin([0, 1]◦ ) and Ł ⊕ Ł ∈ / Fin([0, 1]∗ ).
In the case the decomposition of [0, 1]∗ has infinitely many components, it is easy
to prove that there exist infinitely-many continuous t-norms ◦ such that ∗
≡ ◦ but
Fin([0, 1]∗ ) = Fin([0, 1]◦ ). We do not formally prove the statement but we give
some examples:

• If the decomposition of [0, 1]∗ consists of an infinite number of Łukasiewicz

components Ł, then any other standard BL-chain [0, 1]◦ whose decomposition
begins with an Ł component and contains infinitely many Łukasiewicz components
together with (finitely or infinitely many) components Π or G, defines the same
variety, namely, the full variety of BL-algebras, see [1].
• If the decomposition of [0, 1]∗ begins with a 2 component and contains an infinite
number of Łukasiewicz components, then any other standard BL-chain [0, 1]◦
whose decomposition begins with a 2 component and contains infinitely many
Łukasiewicz components together with (finitely or infinitely many) components
Π or G, defines the same variety, namely, the full variety of SBL-algebras, see [1].

4 How to Find a Set of Equations of an Equationally

Definable t-Norm

After identifying in the last section which t-norms are equationally definable, in this
section we show how to find an effective set of equations for each of them, again rely-
ing in results from [8]. It has to be remarked that the equations actually characterise
the variety generated by the standard algebra [0, 1]∗ for a given equationally definable
t-norm ∗, and hence the equations will involve not only the operation corresponding
to the t-norm but the operation corresponding to its residuum as well.
First we introduce an equation that will have a key role in axiomatizing the varieties
V ([0, 1]∗ ).
Definition 4 Let A be a BL-chain whose decomposition
 as ordinal sum of Wajsberg
hoops has finitely many components, i.e., A = i=0,1,...,n Ai . Then we will denote
On the Equational Characterization of Continuous t-Norms 79

by e A the following equation on n + 1 variables,

⎡ ⎤

 

⎣ ((xi+1 → xi ) → xi ) ∗ (¬¬x0 → x0 ) → xi ⎦ ∨ eiA (xi ) = 1
i=0,...,n−1 i=0,...,n i=0,...,n
(e A )

where eiA (x) = eŁ (x) if Ai = Ł, eiA (x) = eC (x) if Ai = C, and eiA (x) = e2 (x) if
Ai = 2.

Notation convention: for the sake of a simpler notation, from now on we will use
Fin(∗) and V ar (∗) to respectively denote Fin([0, 1]∗ ) and V ar ([0, 1]∗ ).

Lemma 2 Let ∗ be a continuous t-norm whose corresponding standard BL-chain

has a decomposition as ordinal sum with finitely many components Ł, Π and G, and
let A ∈ Fin. Then e A is valid in all BL-chains B ∈ Fin(∗) if and only if A ∈
/ Fin(∗).

And from this result, we can prove the following equational characterization as a
particular case of a more general result in [8, Theorem 5.2].

Proposition 3 Let ∗ be a continuous t-norm whose corresponding standard

BL-chain [0, 1]∗ has a decomposition as ordinal sum with finitely many components
Ł, Π and G. Then,
V ar (∗) is axiomatized by the set of equations AX (∗) = {e B : B ∈ Fin(∗⊥ )},
where Fin(∗⊥ ) = Fin\Fin(∗).

Note that AX (∗) may contain an infinite number of equations. However we can
do it better. Actually, one can show that one needs only a finite subset of AX (∗) to
axiomatize V ar (∗). Indeed, it is only necessary to keep from Fin(∗⊥ ) only those BL-
chains that are minimal in the following sense. Define an ordering relation in the set
Fin as follows: for all A, B ∈ Fin, define A  B if A ∈ V ar (B). And denote by
Min(∗⊥ ) the minimal elements of Fin(∗⊥ ) with respect to the order . It is then
clear that it is enough to consider the set of equations corresponding to the BL-chains
of Min(∗⊥ ), and moreover, it can be shown that Min(∗⊥ ) is always finite, and hence
that V ar (∗) can be axiomatized by a finite set of equations.

Proposition 4 Let ∗ be a continuous t-norm whose decomposition as ordinal sum

of t-norms has finitely many components. Then:
(i) The set Min(∗⊥ ) is finite.
(ii) V ar (∗) is axiomatized by the finite set of equations

AX min (∗) = {e B : B ∈ Min(∗⊥ )}.

Following [8], given an arbitrary continuous t-norm ∗ and its decomposition as

ordinal sum of Ł, G and Π components, an algorithmic procedure to find the set
Min(∗⊥ ) can be given. The idea to find the minimal elements of Fin which are not
80 F. Esteva and L. Godo

2
L

2 2 2 C 2 L

2 C 2 2 C C 2 C L

Fig. 1 Analysis for ∗ = G ⊕ Ł

in Fin(∗) is to iteratively checking ordinal sums from Fin of increasing length (1,
2, 3, etc.). At a given step i, a given current ordinal sum B of length i is checked
whether there is another non-discarded ordinal sum B of length ≤i such that B  B.
If so, the current ordinal sum is discarded for further analysis at step i + 1. Otherwise
B ∈ Min(∗⊥ ) only if B is checked to not belong to Fin(∗). At next step i + 1, only
those non-discarded ordinal sums at step i are expanded with a new component, and
the procedure starts over. This iterative procedure ends in a finite number of steps.
We exemplify this procedure with two examples.

Example 2 Consider a continuous t-norm ∗ isomorphic to G ⊕ Ł. The above itera-

tive procedure, depicted in Fig. 1 as a spanning tree, yields:

Min(∗⊥ ) = {Ł, 2 ⊕ C ⊕ 2, 2 ⊕ C ⊕ C}.

Example 3 Consider a continuous t-norm ∗ isomorphic to G ⊕ Ł ⊕ Π ⊕ Ł. The

above iterative procedure, depicted in Fig. 2, yields:

Min(∗⊥ ) = {Ł, 2 ⊕ C ⊕ Ł ⊕ 2, 2 ⊕ C ⊕ Ł ⊕ C}.

Therefore using the result of the previous proposition, we automatically have a

finite set of equations AX min (∗) univocally characterising ∗, since the only contin-
uous t-norm algebra (up to isomorphism) belonging to V ar (∗) is [0, 1]∗ itself.
Dedication
This short note is dedicated to Peter Klement in the occasion of his retirement. We
are deeply indebted to Peter, not only for his outstanding and numerous scientific
On the Equational Characterization of Continuous t-Norms 81

2
L

2 2 2 C 2 L

2 C 2 2 C C 2 C L

2 C L 2 2 C L C 2 C L L

Fig. 2 Analysis for ∗ = G ⊕ Ł ⊕ Π ⊕ Ł

contributions to the field of fuzzy logic, but also for his incredible task of fostering
the exchange of ideas and the collaboration among researchers in our community,
mainly (but not only) through his Linz Seminars on Fuzzy Set Theory since 1979.
Congratulations Peter!

Acknowledgments The authors have been partially supported by the Spanish MINECO project
EdeTRI TIN2012-39348-C02-01.

References

1. Aglianó, P., Montagna, F.: Varieties of BL-algebras I: general properties. J. Pure Appl. Algebra
181, 105–129 (2003)
2. Aglianó, P., Ferreirim, I.M.A., Montagna, F.: Basic hoops: an algebraic study of continuous
t-norms. Stud. Logica 87(1), 73–98 (2007)
3. Blok, W.J., Ferreirim, I.M.A.: On the structure of hoops. Algebra Univers. 43, 233–257 (2000)
4. Cignoli, R., Esteva, F., Godo, L., Torrens, A.: Basic logic is the logic of continuous t-norms
and their residua. Soft Comput. 4, 106–112 (2000)
5. Cintula, P., Hájek, P., Noguera, C. (eds.): Handbook of Mathematical Fuzzy Logic (in 2 vol-
umes), Studies in Logic, Mathematical Logic and Foundations, vols. 37 and 38. College Pub-
lications, London (2011)
6. Di Nola, A., Esteva, F., Garcia, P., Godo, L., Sessa, S.: Subvarieties of BL-algebras generated
by single-component chains. Arch. Math. Logic 41, 673–685 (2002)
7. Di Nola, A., Lettieri, A.: Equational characterization of all varieties of MV-algebras. J. Algebra
221, 463–474 (1999)
8. Esteva, F., Godo, L., Montagna, F.: Equational characterization of the subvarieties of BL gen-
erated by t-norm algebras. Stud. Logica 76(2), 161–200 (2004)
82 F. Esteva and L. Godo

9. Ferreirim, I.M.A.: On varieties and quasivarieties of hoops and their reducts. Thesis, University
of Illinois at Chicago (1992)
10. Gottwald, S.: A Traitise on Multiple-valued Logics. Studies in Logic and Computation.
Research Studies Press, Baldock (2001)
11. Hájek, P.: Metamathematics of Fuzzy Logic. Trends in Logic, vol. 4. Studia Logica Library,
Kluwer, Dordercht (1998)
12. Hájek, P.: Basic logic and BL-algebras. Soft Comput. 2, 124–128 (1998)
13. Klement, P., Mesiar, R., Pap, E.: Triangular Norms. Trends in Logic, vol. 8, Studia Logica
Library. Kluwer, Dordrecht (2000)
14. Novák, V., Perfilieva, I., Močkoř, J.: Mathematical Principles of Fuzzy Logic. Kluwer, Boston
(1999)
15. Zadeh, L.A.: Preface. In: Marks-II, R.J. (ed.) Fuzzy Logic Technology and Applications. IEEE
Technical Activities Board (1994)
The Semantics of Fuzzy Logics:
Two Approaches to Finite Tomonoids

Thomas Vetterlein and Milan Petrík

Abstract Fuzzy logic generalises classical logic; in addition to the latter’s truth
values “false” and “true”, the former allows also intermediary truth degrees. The
conjunction is, accordingly, interpreted by an operation acting on a chain, making
the set of truth degrees into a totally ordered monoid. We present in this chapter two
different ways of investigating this type of algebras. We restrict to the finite case.

1 Introduction

The idea on which fuzzy logic is built is best understood in relationship with the
canonical way in which reasoning is formalised: with classical propositional logic.
The latter is the logic of “false” and “true” and propositions are evaluated in this
two-element set. Among the connectives we find the logical “and”, “or”, and “not”,
interpreted in the well-known way. In addition to the two classical truth values,
fuzzy logic uses intermediary degrees of truth [14]. Usually, “false” and “true” are
identified with the real numbers 0 and 1, respectively; the remaining real numbers
serve as further truth degrees and may express relative tendencies.
The difficulty of this approach is that there is no straightforward way to tell how
the logical connectives should be interpreted. We rather have to make a decision, for
instance, about the interpretation of the conjunction. Different interpretations will
in general lead to different logics. As a consequence of this situation, fuzzy logic
has in fact emerged as a family of many-valued logics, each of which may bring its
own challenges. According to a common agreement, the binary operation on the real

T. Vetterlein (B)
Department of Knowledge-Based Mathematical Systems,
Johannes Kepler University, Linz, Austria
e-mail: [email protected]
M. Petrík
Department of Mathematics Faculty of Engineering, Czech University
of Life Sciences, Prague, Czech Republic
e-mail: [email protected]
© Springer International Publishing Switzerland 2016 83
S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_6
84 T. Vetterlein and M. Petrík

unit interval taken for this purpose should be a t-norm: associative, commutative,
possessing 1 as an identity, and monotone in each argument. If the set of truth values
is not taken to be an uncountable set but for instance a finite chain, the operation
should still fulfil the same algebraic conditions. It is natural to assume that the chain
of truth degrees is a negative, commutative totally ordered monoid.
The present work is to be seen among the efforts of classifying these algebraic
structures. A considerable amount of work has been done on this topic during recent
years. In line with the given background, residuation has usually been additionally
assumed and MTL-algebras were considered [8, 22]. Our paper [29] is devoted to
MTL-algebras based on the real unit interval. For residuated lattices in general, see [3,
11]. MTL-chains fulfilling certain additional properties were considered in several
works as well. For instance, MTL-chains with the weak cancellation property are
the topic of [21] as well as [15]. Idempotent residuated chains are studied in [4].
The paper [16] deals with finite MTL-chains and their relationship to Abelian totally
ordered groups.
The present chapter is devoted to the finite case. The tomonoids considered are
assumed to be either finite, or at least to be finitely generated. We present two different
approaches based on our work [24, 28], respectively. We provide an introduction to
the main ideas; further details can be found in the indicated papers. We note that in
[27] two further approaches to the structures under consideration are offered.

2 Totally Ordered Monoids

We investigate in this chapter the following structures [5, 9, 12, 13, 23, 30].
Definition 1 An algebra (L; +, 0) is a monoid if (i) + is an associative binary
operation and (ii) 0 is an identity for +. A monoid (L; +, 0) is called commutative
if + is commutative.
A partial order ≤ on a monoid L is called compatible if, for any a, b, c, d ∈ L,
a ≤ b and c ≤ d imply a + c ≤ b + d. A structure (L; ≤, +, 0) such that (L; +, 0)
is a monoid and ≤ is a compatible total order on L is called a totally ordered monoid,
or tomonoid for short.
Moreover, a tomonoid (L; ≤, +, 0) is called commutative if so is its monoidal
reduct. L is called positive if 0 is the bottom element. L is called finitely generated
if L, as a monoid, is generated by finitely many elements.
For instance, let [0,1] be the real unit interval and let ⊕ : [0,1]2 → [0,1] be a
t-conorm, that is, associative, commutative, behaving neutrally w.r.t. 0, and monotone
in each argument [20]. Then ([0,1] ; ≤, ⊕, 0) is commutative, positive tomonoid.
Similarly, let L ⊂ [0,1] be a finite subset of [0,1] containing 0 and 1 and let ⊕ : L 2 →
L be a discrete t-conorm [7]. This is equivalent to say that (L; ≤, ⊕, 0) is a finite,
commutative, positive tomonoid.
We have written tomonoids in the additive way; alternatively, we may deal with
the dual structures. In this case, the order is reversed and the multiplicative notation is
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 85

used. In particular, the monoidal operation is then denoted by a product-like symbol

and the monoidal identity by 1. The aforementioned examples would in this case
become tomonoids based on a t-norm or a discrete t-norm, respectively. The choice
of order and notation is not solely a matter of taste. In many-valued logics, a larger
value corresponds to a higher degree of presence and hence the multiplicative notation
is common. In the context of free monoids, in contrast, the additive notation is
predominant. Within the present chapter, both possibilities will be made use of.
A tomonoid consisting of the monoidal identity alone is called trivial. We will
tacitly assume throughout this paper that all tomonoids are non-trivial. A set of
generators of a (non-trivial) tomonoid L will be understood to be a non-empty, finite
set of elements distinct from 0 that generate L as a monoid.
Congruences of tomonoids are defined as follows; cf. [9]. Recall that a subset C
of a poset is called convex if a, c ∈ C and a ≤ b ≤ c imply b ∈ C.

Definition 2 Let (L; ≤, +, 0) be a tomonoid. A tomonoid congruence on L is a

congruence ≈ of L as a monoid such that all ≈-classes are convex. On the quotient
L ≈ , we then denote the operation induced by + again by + and, for a, b ∈ L, we
let a ≈ ≤ b ≈ if a ≈ b or a < b.

We immediately check that this definition is as intended.

Lemma 1 Let ≈ be a tomonoid congruence on a tomonoid (L; ≤, +, 0). Then the

quotient (L ≈ ; ≤, +, 0 ≈ ) is a tomonoid again. Furthermore, if L is commutative,
positive, finitely generated, then so is L ≈ , respectively.

It is difficult to classify the congruences of tomonoids. There are, however, certain

special types that allow an easy description. For instance, an ideal of a commutative,
positive tomonoid induces a congruence in a natural way [3]. For a discussion of this
type of congruences, see, e.g., [29]. Moreover, there is an order-theoretic analogue
of a Rees quotient; this type of congruences will be central in the second part of this
chapter.

3 Representation of Tomonoids by Direction Cones

The first part of the present chapter is devoted to finitely generated, positive, com-
mutative tomonoids; we will write “fg.p.c. tomonoids” for short. In particular, the
finite, positive, commutative tomonoids, which correspond to the so-called discrete
t-norms [7], are included in the discussion.
We investigate a particular way of representing such tomonoids. We are guided
by the following ideas. First of all, any monoid can be identified with a congruence
on a free monoid. Similarly, we may describe tomonoids by what we call monomial
preorders. Second, the order of totally ordered Abelian groups is characterised by
their cone. We introduce for tomonoids an analogous object; the so-called direction
86 T. Vetterlein and M. Petrík

cones are certain subsets of Zn that describe tomonoids and each fg.p.c. tomonoid is
a quotient of a tomonoid arising in this way.
The results of this section originate from the paper [28], to which we refer for
further details. A continuation of this work, in which the finite case is especially
emphasised, can be found in [26].

3.1 Congruences and Monomial Preorders

Free commutative monoids play a central role in what follows. We identify the
free commutative monoid over n ≥ 1 elements with Nn . The addition is defined
pointwise and the identity is 0̄ = (0, . . . , 0), the n-tuple consisting of zeros only. We
also define u i = (0, . . . , 0, 1, 0, . . . , 0), “1” being at the i-th position. Clearly then,
U (Nn ) = {u 1 , . . . , u n } is a set of generators of Nn .
We endow Nn with the componentwise natural order. That is, for (a1 , . . . , an ),
(b1 , . . . , bn ) ∈ Nn , we put

(a1 , . . . , an )  (b1 , . . . , bn ) if a1 ≤ b1 , . . . , an ≤ bn . (1)

Clearly,  is a lattice order on (Nn ; +, 0̄) and  is compatible with the addition.

Fg.p.c. tomonoids can be conveniently described on the basis of the free commu-
tative monoid Nn as follows.
We call a reflexive and transitive binary relation  on a set A a preorder. We write
a ≺ b if a  b but not b  a. Any preorder  gives rise to an equivalence relation ≈,
called its symmetrisation, where a ≈ b if a  b and b  a. We call the equivalence
class of some a w.r.t. ≈ a -class and we denote it by a  . The preorder  induces
on the quotient A  a partial order, which we denote by  again.
We call a preorder  total if a  b or b  a for any pair a, b ∈ A. Moreover, we
call  positive if 0 ≺ a for all a = 0. Finally, if  is defined on a monoid (L; +, 0),
we call  compatible if a  b implies a + c  b + c.
In computational mathematics, the notion “monomial ordering” refers to compat-
ible, positive, total orders on Nn ; see, e.g., [6]. Analogously, we call a preorder  on
Nn monomial if  is compatible, positive, and total. The significance of monomial
preorders becomes clear in the following proposition.

Proposition 1 Let  be a monomial preorder on (Nn ; +, 0̄). Then its symmetrisa-

tion is a monoid congruence whose classes are convex and such that 0̄  = {0̄}.
Moreover, (Nn  ; , +, {0̄}) is a fg.p.c. tomonoid.
Conversely, let (L; ≤, +, 0) be a fg.p.c. tomonoid; assume that the n ≥ 1 elements
g1 , . . . , gn ∈ L\{0} generate L. Let ι : Nn → L be the surjective monoid homomor-
phism determined by ι(u i ) = gi , i = 1, . . . , n. For a, b ∈ Nn define

a  b if ι(a) ≤ ι(b). (2)

The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 87

Then  is a monomial preorder of Nn , and ι induces an isomorphism between

(Nn  ; , +, {0̄}) and (L; ≤, +, 0).
Proof Let  be a monomial preorder on Nn . Then, for a, b, c, d ∈ Nn , a ≈ c and b ≈
d imply a + b ≈ c + d by the compatibility of ; hence ≈ is a monoid congruence.
As  is also positive,  extends , and it follows that the -classes are convex.
Again by the positivity, the -class of 0̄ consists of 0̄ alone.
As  is compatible, the partial order  induced on Nn  is compatible as well;
that is, (Nn  ; , +, 0̄  ) is a commutative pomonoid. Since, for any a, b ∈ Nn ,
a  b or b  a, Nn  is actually a tomonoid. Moreover, since 0̄ ≺ a for any a ∈
Nn \{0̄}, Nn  is a positive, commutative tomonoid, which is generated by the
finitely many elements u 1  , . . . , u n  .
For the second part, assume that (L; ≤, +, 0) is a fg.p.c. tomonoid and g1 , . . . ,
gn ∈ L\{0} generate L as a monoid. Let furthermore ι : Nn → L be as indicated
and let  be defined by (2). By construction,  is transitive and reflexive, that
is, a preorder.  is compatible because so is ≤ and ι is a monoid homomorphism.
Moreover,  is positive because L is positive and hence ι(a) ≤ 0 holds only if a = 0̄.
Hence  is a monomial preorder. Finally, for a, b ∈ Nn , we have a ≈ b if and only
if a  b and b  a if and only if ι(a) = ι(b); hence ι induces an isomorphism as
claimed.
We conclude that any monomial preorder  on Nn gives rise to a fg.p.c. tomonoid
L. We call L in this case the tomonoid represented by .
Proposition 1 also states that, up to isomorphism, any fg.p.c. tomonoid L arises
in this way from a monomial preorder. In other words, describing fg.p.c. tomonoids
can be done by describing monomial preorders. This is what we will do in the sequel.

3.2 Tomonoids Arising from Totally Ordered Abelian Groups

The positive cones of totally ordered Abelian groups give rise to typical examples
of fg.p.c. tomonoids. We will discuss these examples in some detail because they
motivate our way of representing fg.p.c. tomonoids in general.
Definition 3 Let (G; ≤, +, 0) be a totally ordered Abelian group and let G + = {g ∈
G : g ≥ 0} be its positive cone. Assume that G is generated by g1 , . . . , gn ∈ G + \{0},
where n ≥ 1. Let L be the submonoid of G generated by g1 , . . . , gn and let L be
endowed with the total order inherited from G, with the group addition, and with the
constant 0. Then we call (L; ≤, +, 0) a group cone tomonoid.
Clearly, a group cone tomonoid is a fg.p.c. tomonoid. Note that in general we do
not deal with the whole positive cone of a totally ordered Abelian group. In fact, the
latter is in general not finitely generated even if the group is.
Group cone tomonoids are characterised by the following condition. We say that a
fg.p.c. tomonoid L is cancellative if, for all a, b, c ∈ L, a + c = b + c implies a = b.
Note that in this case, for all a, b, c ∈ L, a ≤ b is equivalent to a + c ≤ b + c.
88 T. Vetterlein and M. Petrík

Proposition 2 A fg.p.c. tomonoid (L; ≤, +, 0) is a group cone tomonoid if and only

if it is cancellative.

Proof The “only if” part follows from the construction of a group cone tomonoid.
To see the “if” part, let L be cancellative. Let G be the group consisting of the
differences of elements of L; see, e.g., [10, Chap. II.2]. Viewing L as a subset of G, we
introduce a total order on G as follows: for a, b, c, d ∈ L, we define a − b ≤ c − d
if a + d ≤ b + c in L. Then (G; ≤, +, 0) is a totally ordered Abelian group, and
(L; ≤, +, 0) is a subtomonoid of (G + ; ≤, +, 0). The assertion follows.

Group cone tomonoids correspond by Proposition 1 to particular monomial pre-

orders. We call a preorder  on Nn cancellative if, for any a, b, c ∈ Nn , a  b is
equivalent to a + c  b + c.

Proposition 3 Let the fg.p.c. tomonoid L be represented by the monomial preorder

 on Nn . Then L is a group cone tomonoid if and only if  is cancellative.

Proof Let L be a group cone tomonoid. Then (Nn  ; , +, {0̄}) is cancella-

tive by Proposition 2. Thus, for a, b, c ∈ Nn , we have a  b iff a   b  iff
a  + c   b  + c  iff a + c   b + c  iff a + c  b + c, that is,  is
cancellative.
Conversely, let  be cancellative. Then (Nn  ; , +, {0̄}) is a cancellative fg.p.c.
tomonoid and hence, by Proposition 2, a group cone tomonoid.

Recall next that the order of a partially ordered Abelian group (G; ≤, +, 0) is
uniquely determined by its positive cone G + . In fact, for any g, h ∈ G, g ≤ h if
and only if h − g ∈ G + . We may also view the positive cone of a partially ordered
group as the set of all differences of elements g and h such that g ≤ h; indeed,
G + = {h − g : g, h ∈ G such that g ≤ h}.
We may use the same object to describe group cone tomonoids. We denote by
(Zn ; +, 0̄) the free Abelian group generated by n ≥ 1 elements. Furthermore,  will
be the partial order on Zn defined according to (1): for a, b ∈ Zn , we put a  b if
a + c = b for some c ∈ Nn . Then (Zn ; , +, 0̄) is a lattice-ordered group.

Definition 4 Let  be a cancellative monomial preorder on Nn . Then the set

P = {b − a ∈ Zn : a, b ∈ Nn such that a  b}

is called the positive cone of .

A positive cone determines the preorder from which it is defined as in the case of
groups.

Lemma 2 Let P ⊆ Zn be the positive cone of the cancellative monomial preorder

 on Nn . Then we have:
(GO) For any a, b ∈ Nn , a  b if and only if b − a ∈ P.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 89

Proof By definition, a  b implies b − a ∈ P.

Conversely, let b − a ∈ P. Then there are c, d ∈ Nn such that c  d and d − c =
b − a. It follows a + d = b + c  b + d and hence a  b.

By Lemma 2, we have for any cancellative monomial preorder 

P = {z ∈ Zn : a  b for some a, b ∈ Nn such that z = b − a}

= {z ∈ Zn : a  b for all a, b ∈ Nn such that z = b − a}. (3)

The positive cones of partially ordered Abelian groups possess an intrinsic charac-
terisation: they are exactly the cancellative commutative monoids such that a + b = 0
implies a = b = 0 [10]. The positive cones of cancellative monomial preorders can
be described in a similar way.

Theorem 1 A set P ⊆ Zn is the positive cone of a cancellative monomial preorder

on Nn if and only if the following conditions are fulfilled:
(GC1) If z ∈ Nn , then z ∈ P. Moreover, if z ∈ Nn \ {0̄}, then −z ∈
/ P.
(GC2) P is closed under addition.
(GC3) For any z ∈ Zn , at least one of z ∈ P or −z ∈ P holds.
In this case, P = P , where  is given by condition (GO) above.

Proof Let  be a cancellative monomial preorder on Nn . Clearly, 0 ∈ P then.

Furthermore, any z ∈ Nn \{0̄} is in P because 0̄  z holds by the positivity of .
Assume that also −z ∈ P . Then there is a b ∈ Nn such that b + z  b and hence
by the cancellativity z  0, in contradiction to the positivity of . (GC1) is shown.
For a, b, c, d ∈ Nn , a  b and c  d implies a + c  b + c  b + d. We conclude
that if b − a, d − c ∈ P , also (b − a) + (d − c) = (b + d) − (a + c) ∈ P . This
shows (GC2).
For a, b ∈ Nn , at least one of a  b or b  a holds because  is total. (GC3)
follows as well.
Let now P ⊆ Zn fulfil (GC1)–(GC3). For a, b ∈ Nn , let a  b if b − a ∈ P. We
claim that  is a cancellative monomial preorder. As 0 ∈ P by (GC1),  is reflexive.
By (GC2),  is transitive. Hence  is a preorder.  is total by (GC3) and posi-
tive by (GC1). Finally, by construction, a  b is equivalent to a + c  b + c; the
compatibility and cancellativity of  follows.
It remains to show that P is actually the positive cone P of . By Lemma 2, we
have that, for any a, b ∈ Nn , b − a ∈ P if and only if a  b. But by construction,
a  b if and only if b − a ∈ P. Hence P = P .
Finally, if P ⊆ Zn is the positive cone of any cancellative monomial preorder ,
then  is by Lemma 2 uniquely determined by (GO). The last statement follows.
90 T. Vetterlein and M. Petrík

3.3 Direction Cones

Positive cones describe cancellative fg.p.c. tomonoids. In this section we will gener-
alise this notion to cover a wider class of tomonoids. In this case we will not obtain
a strict correlation, but we will be led to a Galois correspondence.
Let  be a monomial preorder on Nn . If  is cancellative, then for any a, b ∈ Nn
the question of whether or not a  b holds depends only on the difference z = b − a:
we have a  b if and only if c  d for any other pair c, d ∈ Nn such that z = d − c.
In fact, the positive cone P consists of these differences; a  b if and only if
b − a ∈ P .
In general, the question of whether or not we have a  b does not depend on the
difference b − a alone. For instance, it may be the case that a + c  b + c holds for
some c ∈ Nn but not a  b. However, let z ∈ Zn . Then the following lemma implies
that still at least one of following possibilities applies: a  b for all a, b ∈ Nn such
that b − a = z, or b  a for all a, b ∈ Nn such that b − a = z.

Lemma 3 Let z ∈ Zn . Then there is a unique pair a, b ∈ Nn such that z = b − a

and, for any c, d ∈ Nn such that z = d − c, we have c = a + t and d = b + t for
some t ∈ Nn .

Proof Put a = −z ∨ 0̄ and b = z ∨ 0̄. Then z = b − a. Moreover, if c, d ∈ Nn such

that d − c = z, we have c  0̄ and c = d − z  −z, thus c  a; similarly, d  b. As
b − a = d − c, the differences c − a and d − b coincide and hence c = a + t and
d = b + t for some t ∈ Nn . The uniqueness of a, b follows from the -minimality.

Let a, b ∈ Nn be associated with z ∈ Zn according to Lemma 3. Inspecting the

proof, we see that b is simply the positive part of z ∈ Zn , and a is its (negated)
negative part. Let us define

z + = z ∨ 0̄,
z − = −z ∨ 0̄.

Then we have
z = z+ − z−

and any other pair of elements of Nn whose difference is z arises from z + and z − by
adding a t ∈ Nn .
For a compatible preorder  on Nn , the obvious consequence is the following.
Let z ∈ Zn . If z −  z + , we conclude from Lemma 3 and the compatibility of  that
a  b actually holds for any pair a, b ∈ Nn such that b − a = z. Thus, intuitively,
we may view any z ∈ Zn such that z −  z + as being “positively directed”; for, in
this case we have a  a + z for any a ∈ Nn such that a + z ∈ Nn . Our viewpoint is
reflected in the following definition.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 91

Definition 5 Let  be a monomial preorder on Nn . Then the set

C = {z ∈ Zn : z −  z + }

is called the direction cone of .

By Lemma 3 we then have

C = {z ∈ Zn : a  b for all a, b ∈ Nn such that z = b − a}. (4)

The natural question is now if there is a characterisation of direction cones similar

to the case of positive cones. Comparing with (3), we see that the direction cone of a
cancellative monomial preorder is its positive cone. In the general case, we conclude
from the positivity of  that condition (GC1) for positive cones applies here as well,
and from the totality of  also condition (GC3) is immediate: for each z ∈ Zn , at
least one of z or −z is in C .
However, a direction cone does not in general fulfil condition (GC2), that is, it is
not necessarily closed under addition. The following notion can be used instead. We
call a k-tuple (x1 , . . . , xk ), k ≥ 2, of elements of Zn addable if

(x1 + · · · + xk )− + x1 + · · · + xi  0̄ (5)

for all i = 0, . . . , k. Note that for addability the order matters.

Lemma 4 The direction cone of a monomial preorder on Nn is a set C ⊆ Zn fulfilling

the following conditions:
(C1) Let z ∈ Nn . Then z ∈ C and, if z = 0̄, −z ∈ / C.
(C2) Let (x1 , . . . , xk ), k ≥ 2, be an addable k-tuple of elements of C. Then x1 +
· · · + xk ∈ C.
(C3) Let z ∈ Zn . Then z ∈ C or −z ∈ C.

Proof (C1) We have Nn ⊆ C because  is positive. Assume that −z ∈ C, where

z ∈ Nn . Then z = (−z)−  (−z)+ = 0̄ and the positivity of  implies z = 0̄.
Recall next that, by (4), a  b for any a, b ∈ Nn such that b − a ∈ C.
To see (C2), let (x1 , . . . , xk ) be as indicated, and put z = x1 + · · · + xk . Then
z − , z − + x1 , . . . , z − + x1 + · · · + xk ∈ Nn . By assumption, x1 , . . . , xk ∈ C; thus
z −  z − + x1  . . .  z − + x1 + · · · + xk = z − + z = z + .
(C3) holds because  is total.

Our next aim is to show that conditions (C1)–(C3) characterise direction cones.
A preorder gives rise to a direction cone, which fulfils (C1)–(C3). Conversely, we
can assign a preorder to a set fulfilling (C1)–(C3).

Definition 6 Let C ⊆ Zn fulfil (C1)–(C3). Let C be the smallest preorder on Nn

such that
92 T. Vetterlein and M. Petrík

(O) a C b for any a, b ∈ Nn such that b − a ∈ C.

Then we call C the monomial preorder induced by C.
In other words, for a subset C of Zn fulfilling (C1)–(C3) and a, b ∈ Nn , we
have a C b if and only if there are k ≥ 1 elements z 1 , . . . , z k ∈ C such that
a, a + z 1 , a + z 1 + z 2 , . . . , a + z 1 + · · · + z k  0̄ and a + z 1 + · · · + z k = b. We
note that this is not the same as to say that b − a is a sum of elements of C.
Lemma 5 Let C ⊆ Zn fulfil (C1)–(C3). Then C , the monomial preorder induced
by C, is in fact a monomial preorder.
Proof By construction, C is a preorder, and by (C3), C is total. It is furthermore
clear that C is compatible with the addition.
Assume next that, for some a ∈ Nn , a C 0̄ holds according to the prescription
(O). Then a = 0̄ by (C1). It follows that 0̄ ≺C a for all a ∈ Nn \{0̄}, that is, C is
positive. This completes the proof that C is a monomial preorder.
Theorem 2 A set C ⊆ Zn is the direction cone of a monomial preorder if and only
if C fulfils (C1)–(C3). In this case, C is the direction cone of C .
Proof A direction cone fulfils (C1)–(C3) by Lemma 4.
Conversely, let C fulfil (C1)–(C3). Let C be the induced preorder. By Lemma 5,
C is a monomial preorder.
It remains to show that CC , the direction cone of C , coincides with C, that is,
for z ∈ Zn , z − C z + if and only if z ∈ C. The “if” part holds by construction. For
the “only if” part, assume that z − C z + = z − + z. Then z = x1 + · · · + xk for some
x1 , . . . , xk ∈ C such that z − + x1 + · · · xi  0̄ for i = 0, . . . , k. Then (x1 , . . . , xk )
is addable, hence z ∈ C by (C2).
In the sequel, when speaking about direction cones without reference to a mono-
mial preorder, we mean a subset of Zn that fulfils the conditions (C1)–(C3).
A direction cone induces a preorder. As seen next, any preorder contains a preorder
arising in this way.
Theorem 3 Let  be a monomial preorder. Then  extends C , the monomial
preorder induced by the direction cone of .
Moreover, the direction cone of C is C again.
Proof Let a, b ∈ Nn and assume that a C b holds according to the prescription
(O). Then b − a ∈ C , that is, z −  z + , where z = b − a. In view of Lemma 3, it
follows a  b. We conclude that C ⊆ .
The second part holds by Theorem 2.
We apply the shown facts to tomonoids.
Definition 7 Let C ⊆ Zn be a direction cone. Then we call the tomonoid represented
by C a cone tomonoid.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 93

Theorem 4 Each fg.p.c. tomonoid L is the quotient of a cone tomonoid.

Proof This follows from Theorem 3.

3.4 A Galois Connection

We have seen that there is a mutual correspondence between monomial preorders and
direction cones. This correspondence is not one-to-one, some monomial preorders
are proper extensions of those that are induced by direction cones. However, we can
established a Galois correspondence between the two sets.
Let us fix an n ≥ 1. Let P be the set of all monomial preorders on Nn and let C
be the set of all direction cones in Zn . We partially order the two sets by means of
the set-theoretic inclusion. We then readily check that the two mappings

P → C ,  → C ,
C → P, C → C

are order-preserving. The mappings are not one-to-one; in fact, the former is surjec-
tive but not injective, and the latter is injective but not surjective. From Theorems
2 and 3 we conclude what happens when applying the mappings successively: any
 ∈ P is an extension of C ; and any C ∈ C is equal to CC . Hence there is the
following Galois connection between P and C : for any  ∈ P and C ∈ C ,

C ⊆  if and only if C ⊆ C .

3.5 Example

We conclude by presenting an example illustrating the results of this section. Let L

be the 9-element fg.p.c. tomonoid specified as follows. Let L be generated by its two
elements a and b and assume that

0 < a < b < 2a < a + b < 2b < 3a <

2a + b = a + 2b = 4a < 2a + 2b = 3a + b = 5a = 3b

and that the last indicated element is the top element. In accordance with Proposi-
tion 1, let ι : N2 → L be the surjective monoid homomorphism such that ι((1, 0)) = a
and ι((0, 1)) = b, and endow N2 with the preorder  according to (2). Then we have

(0, 0) ≺ (1, 0) ≺ (0, 1) ≺ (2, 0) ≺ (1, 1) ≺ (0, 2) ≺

(3, 0) ≺ (2, 1) ≈ (1, 2) ≈ (4, 0) ≺ (m, n),
94 T. Vetterlein and M. Petrík

.. .. .. .. .. ..
. . . . . .

(0, 3) (1, 3) (2, 3) (3, 3) (4, 3) (5, 3) ...

(0, 2) (1, 2) (2, 2) (3, 2) (4, 2) (5, 2) ...

(0, 1) (1, 1) (2, 1) (3, 1) (4, 1) (5, 1) ...

(0, 0) (1, 0) (2, 0) (3, 0) (4, 0) (5, 0) ...

Fig. 1 The example tomonoid L. The simple arrows indicate the immediate-successor relation
w.r.t. ; the double arrows indicate -equivalence

where (m, n) is any of the remaining elements of N2 . A graphical representation of

(L; ≤, +, 0) can be found in Fig. 1.
According to Definition 5, the direction cone is

C = {( p, q) ∈ Z2 : (− p ∨ 0, −q ∨ 0)  ( p ∨ 0, q ∨ 0)}
= {( p, q) ∈ Z2 : p, q ≥ 0} ∪
{(−2, 2), (−1, 1), (−1, 2), (2, −1), (3, −2), (3, −1), (4, −2), (4, −1)} ∪
{( p, q) ∈ Z2 : p ≤ 0 and q ≥ 3} ∪
{( p, q) ∈ Z2 : p ≥ 5 and q ≤ 0}.

This set is depicted in Fig. 2.

Finally, we calculate C , the preorder representing a cone tomonoid whose
quotient is L. The preorder C can most easily be read off directly from Fig. 1.
Namely, we collect the order relations that hold between elements of the form (m, 0)
and (0, n), where m, n ≥ 1; then we translate and concatenate them. The result is
depicted in Fig. 3. From C , we get L by requiring the elements (2, 1), (1, 2), and
(4, 0) of N2 to be equivalent.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 95

.. .. .. .. .. .. .. .. .. .. .. .. .. ..
. . . . . . . . . . . . . .
··· ···
··· ···
··· ···
··· ···
···
···

···
···
···
···
···
···
.. .. ..
. . .

Fig. 2 The direction cone C of the monomial preorder  representing L. Each element of C is
depicted as a circle in the Z2 plane

.. .. .. .. .. ..
. . . . . .

(0, 3) (1, 3) (2, 3) (3, 3) (4, 3) (5, 3) ...

(0, 2) (1, 2) (2, 2) (3, 2) (4, 2) (5, 2) ...

(0, 1) (1, 1) (2, 1) (3, 1) (4, 1) (5, 1) ...

(0, 0) (1, 0) (2, 0) (3, 0) (4, 0) (5, 0) ...

Fig. 3 The cone tomonoid represented by C , whose quotient is L

96 T. Vetterlein and M. Petrík

4 One-Element Rees Coextensions of Finite

Negative Tomonoids

In the second part of this chapter, we develop a much different point of view on
tomonoids. To begin with, we switch to the dual order and the multiplicative notation,
as is common in fuzzy logic.
We will again assume the property called “positive” in the previous part. In the
present context, however, positivity means that 1 is the top element; accordingly, we
will refer to this property as “negative”. Furthermore, we will restrict to the finite
case. Finally, our considerations do not rely on the commutativity of the monoidal
product and hence we will not assume this condition here.
We shall write “f.n.” for “finite, negative”. That is, a f.n. tomonoid is a structure
(L; ≤, , 1) such that (L; , 1) is a finite monoid and ≤ is a compatible total order
whose top element is 1.
Our aim is to describe the construction of f.n. tomonoids in a step-by-step manner.
The main idea is the following. Let (L; ≤, , 1) be a non-trivial f.n. tomonoid and let
0 and α be its smallest and second smallest element, respectively. Then the identifica-
tion of 0 and α is a tomonoid congruence and the quotient is by one element smaller
than L. Continuing in the same way, we get a sequence of tomonoids that ends with
the trivial one. It seems then natural to ask how to generate such a sequence in the
reversed order. That is, given an f.n. tomonoid L, how can we determine all those
f.n. tomonoids L̄ that are by one element larger and such that the identification of
their smallest two elements leads back to L? This is in fact the question that we will
answer. We will provide a practical method of determining from L systematically
all tomonoids L̄ of the indicated type.
The results of the present section are due to [24], where further details can be
found. For a more general approach to the extension of partially ordered monoids,
see, e.g., [18, 19].

4.1 Rees Congruences

Consider a negative tomonoid (L; ≤, , 1) and let q be one of its elements. Then
Iq = {a ∈ L : a ≤ q} is an ideal of L, seen as a monoid. Indeed, by the negativity of
L, a ≤ q implies a  b ≤ q and b  a ≤ q for any b ∈ L. Consequently, we may
form the Rees quotient of the monoid L by Iq ; see, e.g., [17]. Its elements may be
identified with the elements that are not in Iq as well as one further element, usually
denoted by 0. Obviously, this monoid congruence has only convex classes and hence
it is a tomonoid congruence; cf., e.g., [9].

Definition 8 Let (L; ≤, , 1) be a f.n. tomonoid and let q ∈ L. For a, b ∈ L, let

a ≈q b if a = b or a, b ≤ q. Then we call ≈q the Rees congruence by q. We denote
the quotient by L/q and call it the Rees quotient of L by q.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 97

Moreover, we call L a Rees coextension of L/q. We call L a one-element Rees

coextension, or simply a one-element coextension, if L is non-trivial and q is the
atom of L.

For a finite chain L, we will denote by 0 the bottom element and we write

L ∗ = L \ {0}.

If L has at least two elements, we furthermore call the second smallest element of L
the atom of L. We will use in the sequel the symbol α to denote the atom.
Given a non-trivial f.n. tomonoid L, then its Rees quotient L/α by its atom α arises
from L by the identification of the smallest two elements. L is in this case a one-
element coextension of L/α. Our aim is to determine all one-element coextensions
of a f.n. tomonoid. We will then obviously be in the position to construct, starting
from the trivial tomonoid, successively all f.n. tomonoids.

4.2 Tomonoid Partitions

A binary operation  on a set A gives rise to a partition of A × A: the blocks of the

partition are the subsets of all those pairs that are mapped by  to the same value.
The blocks are commonly referred to as the level sets of . This partition, together
with the assignment that associates with each block the respective element of A,
specifies  uniquely.
The representation of binary operations based on level sets was first applied to
the theory of tomonoids in [25]. We note that it comes along with the possibility of
representing tomonoids within two dimensions only.

Definition 9 Let (L; ≤, , 1) be a tomonoid. We define, for any (a, b), (c, d) ∈ L 2 ,

(a, b) ∼ (c, d) if a  b = c  d.

We call ∼ the level equivalence of L.

Based on the level equivalence of a tomonoid L, we will endow the set L 2 with a
first-order structure as follows.

Definition 10 Let ≤ be a total order on a set L and let 1 ∈ L. We denote the com-
ponentwise order on L 2 by , that is, we put

(a, b)  (c, d) if a ≤ c and b ≤ d

98 T. Vetterlein and M. Petrík

for a, b, c, d ∈ L. Moreover, let ∼ be an equivalence relation on L 2 such that the

following conditions hold:

(P1) For any a, b, c, d, e, f ∈ L, if (1, e) ∼ (a, b)  (c, d) ∼ (1, f ), then

e ≤ f.
(P2) For any (a, b) ∈ L 2 , there is exactly one c ∈ L such that (a, b) ∼ (1, c) ∼
(c, 1).
(P3) For any a, b, c, d, e ∈ L, (a, b) ∼ (d, 1) and (b, c) ∼ (1, e) imply (d, c) ∼
(a, e).
We then call the structure (L 2 ; , ∼, (1,1)) a tomonoid partition.

Proposition 4 Let (L; ≤, , 1) be a tomonoid and let ∼ be the level equivalence of

L. Then (L 2 ; , ∼, (1,1)) is a tomonoid partition.

Proof Let a, b, c, d ∈ L. By the compatibility of ≤ with , we have that (a, b) 

(c, d) implies a  b ≤ c  d. (P1) follows. Moreover, as 1 is the monoidal identity,
we have that (a, b) ∼ (c, 1) iff (a, b) ∼ (1, c) iff a  b = c. Hence also (P2) holds.
Finally, (P3) is implied by the associativity of .

By Proposition 4, each tomonoid L gives rise to a tomonoid partition; we will

speak about the tomonoid partition associated with L.
We next see that there is a converse of Proposition 4. We will use the following
simplified notation. When L is a chain and 1 ∈ L, we will identify the elements of the
form (1, c) ∈ L 2 , where c ∈ L, with c. It will be clear from the context if c denotes
an element of L or of L 2 . For instance, if ∼ is an equivalence relation on L 2 , then
(a, b) ∼ c means (a, b) ∼ (1, c). Similarly, the ∼-class of some c ∈ L is meant to
be the ∼-class containing (1, c).

Proposition 5 Let (L 2 ; , ∼, (1,1)) be a tomonoid partition. Let ≤ be the underly-

ing total order of L. Moreover, for any a, b ∈ L, let

a  b = the unique c such that (a, b) ∼ c. (6)

Then (L; ≤, , 1) is the unique tomonoid such that (L 2 ; , ∼, (1,1)) is its associated
tomonoid partition.

Proof By assumption, L is totally ordered and  is the induced componentwise order

on L 2 . Evidently,  determines the total order ≤ on L uniquely. It is furthermore
clear from (P2) that  can be defined by (6).
For a ∈ L, we have 1  a = a by construction and a  1 = 1  a by (P2). Fur-
thermore, (P2) and (P3) imply the associativity of . Thus (L; , 1) is a monoid. Let
a ≤ b. Then (a, c)  (b, c), and we conclude from (P1) that a  c ≤ b  c. Simi-
larly, we see that c  a ≤ c  b. Thus ≤ is compatible with  and (L; ≤, , 1) is
a tomonoid. It is clear that ∼ is the level equivalence of L and we conclude that
(L 2 ; , ∼, (1,1)) is its associated tomonoid partition.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 99

0 t u v w x y z 1
0 t u v w x y z 1 1
0 0 t u u v w x z z
0 0 0 t t u u v y y
0 0 0 t t u u v x x
0 0 0 0 0 t t u w w
0 0 0 0 0 t t u v v
0 0 0 0 0 0 0 t u u
0 0 0 0 0 0 0 0 t t
0 0 0 0 0 0 0 0 0 0

Fig. 4 A tomonoid partition associated with an eight-element negative tomonoid L. Rows and
columns of the array correspond to the elements of L, thus each square in the array corresponds
to a pair (a, b) ∈ L 2 , where a is the row index and b is the column index. In order to represent ∼,
we have indicated in each square (a, b) the product of a and b in L; two squares are ∼-equivalent
iff they contain the same symbol. For instance, the ∼-class of u comprises even elements and the
∼-class of 1 just one

Let (L 2 ; , ∼, (1,1)) be associated to another tomonoid (L  ; ≤ ,  , 1 ). Then,

by the way in which a tomonoid partition is constructed from a tomonoid, L  = L,
≤ = ≤, and 1 = 1. Furthermore, if for some a, b, c ∈ L we have a  b = c, then
(a, b) ∼ (1, c) and hence a  b = c. We conclude  = .
By Propositions 4 and 5, tomonoids and tomonoid partitions are in a one-to-one
correspondence. We will present our results in the sequel mostly with reference to
the latter, that is, with reference to tomonoid partitions.
Let us next devote some remarks to the geometric interpretation of the conditions
(P1)–(P3) in Definition 10. Let L be a tomonoid. Then L is a chain and hence L 2 can
be viewed as a square array. For elements (a, b), (c, d) ∈ L 2 , (a, b)  (c, d) means
that (a, b) is left underneath (c, d). Moreover, for negative tomonoids, 1 is the top
element; in this case, (1, 1) is located in the upper right corner of L 2 . See Fig. 4 for
an illustration.
In order to interpret (P1)–(P3), let us view the level equivalence of L as a partition
of L 2 . Condition (P2) has probably the most straightforward meaning. By (P2), each
block contains exactly one element of the form (1, c), c ∈ L. That is, we may index
the blocks by the elements of the line indexed by 1. Furthermore, (c, 1) and (1, c)
are for each c ∈ L in the same block and hence a similar statement holds also for the
column indexed by 1.
By the identification of the blocks with the line indexed by 1, the blocks are totally
ordered. Condition (P1) says that the componentwise order on L 2 is in accordance
with this total order. Namely, when moving from any element of a block to the right
or upwards, we arrive at a block indexed by a larger element.
Condition (P3), which accounts for the associativity, possesses an appealing geo-
metric interpretation as well. An illustration is given in Fig. 5. Here, we assume that
1 is the top element of L. Within the square array representing L 2 , consider two
100 T. Vetterlein and M. Petrík

e c b
0 t u v w x y z 1
0 t u v w x y z 1 1
0 0 t u u v w x z z b
0 0 0 t t u u v y y a
0 0 0 t t u u v x x
0 0 0 0 0 t t u w w
0 0 0 0 0 t t u v v d
0 0 0 0 0 0 0 t u u
0 0 0 0 0 0 0 0 t t
0 0 0 0 0 0 0 0 0 0

Fig. 5 The “Reidemeister” condition (P3). A (connected or broken) bold line between two elements
of the array indicates level equivalence. By (P3), the equivalences of the pairs connected by a solid
line imply the equivalence of the pair connected by a broken line

rectangles such that one hits the upper edge and the other one hits the right edge.
Assume that the upper left, upper right, and lower right vertices of these rectangles
are in the same blocks, respectively. By (P3), then also the remaining pair, consisting
of the lower left vertices, is in the same block. A related property is known from the
field of web geometry and called the “Reidemeister condition” [1, 2].
We conclude the subsection with a characterisation of those tomonoid partitions
in which we are actually interested: the finite, negative ones. The slightly optimised
characterisation will be useful in subsequent proofs.

Proposition 6 Let (L; ≤) be a finite and at least two-element chain with the top
element 1. Let 0 be the bottom element of L. Then (L 2 ; , ∼, (1,1)) is a tomonoid
partition if and only if (P1), (P2), and the following condition hold:

(P3’) For any a, b, c, d, e ∈ L \ {0, 1}, (a, b) ∼ d and (b, c) ∼ e imply (d, c) ∼
(a, e).
In this case, (L 2 ; , ∼, (1,1)) is finite and negative.

Proof The “only if” part is clear by definition.

To see the “if” part, let (L 2 ; , ∼, (1,1)) fulfil (P1), (P2), and (P3’). We next
show that the negativity criterion of Lemma 6(i) holds:
() (a, b) ∼ (1, c) implies c ≤ a and c ≤ b.
Indeed, in this case (c, 1) ∼ (1, c) ∼ (a, b)  (a, 1) by (P2) and the fact that 1 is
the top element. Hence, by (P1), c ≤ a. Similarly, we see that c ≤ b.
It remains to prove (P3). Let a, b, c, d, e ∈ L be such that (a, b) ∼ d and (b, c) ∼
e. We have to show (d, c) ∼ (a, e) if one of the five elements equals 0 or 1. We
consider certain cases only, the remaining ones are seen similarly.
Let a = 1. Then (1, b) ∼ (1, d), hence b = d by (P2), and it follows (d, c) =
(b, c) ∼ (1, e) = (a, e).
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 101

Let d = 1. Then (a, b) ∼ (1, 1), and by (), we conclude a = b = 1. From

(b, c) ∼ e it follows e = c. Hence (d, c) = (a, e).
Note next that, for any f ∈ L, ( f, 0) ∼ 0. This follows again from ().
Let a = 0. Then (a, b) = (0, b) ∼ 0 and hence d = 0. Hence (d, c) = (0, c) ∼
0 ∼ (0, e) = (a, e).
Let d = 0. Then (d, c) = (0, c) ∼ 0. From (b, c) ∼ e, it follows by () that e ≤
b. Hence (1, 0) ∼ (0, 0)  (a, e)  (a, b) ∼ (1, 0) and we conclude from (P1) that
(a, e) ∼ 0. In particular, (a, e) ∼ (d, c).

4.3 Properties and Constructions for Tomonoid Partitions

We have seen that tomonoids and tomonoid partitions are in a one-to-one corre-
spondence. Consequently, we can apply properties, constructions, etc. defined for
tomonoids to tomonoid partitions as well. We establish in this subsection a few of
such correspondences.
For convenience, we will apply to tomonoid partitions the same notions as to
tomonoids. For instance, a tomonoid partition will be called negative if the corre-
sponding tomonoid is negative.

Lemma 6 Let (L 2 ; , ∼, (1,1)) be a tomonoid partition.

(i) The following statements are pairwise equivalent:
• L 2 is negative.
• (1, 1) is the top element of L 2 .
• The ∼-class of any c ∈ L is contained in {(a, b) ∈ L 2 : a, b ≥ c}.
(ii) The following statements are equivalent:
• L 2 is commutative.
• (a, b) ∼ (b, a) for any a, b ∈ L.

A further property considered in the sequel is Archimedeanicity. In what follows,

we write a n for the n-fold product a  · · ·  a.

Definition 11 We call a negative tomonoid Archimedean if, for any a ≤ b < 1, there
is an n ≥ 1 such that bn ≤ a.

Note that negative tomonoids with at most two elements are trivially Archimedean.
Archimedean f.n. tomonoid partitions are characterised as follows.

Lemma 7 Let (L 2 ; , ∼, (1,1)) be a f.n. tomonoid partition. The following state-

ments are pairwise equivalent:
• L 2 is Archimedean.
• (b, a)  (1, a) for any a ∈ L  and b < 1.
• (a, b)  (a, 1) for any a ∈ L  and b < 1.
102 T. Vetterlein and M. Petrík

Proof Let (L; ≤, , 1) be the corresponding f.n. tomonoid and let 0 be the bottom
element of L. W.l.o.g., we can assume 0 = 1. We show that (i) and (ii) are equivalent.
The equivalence of (i) and (iii) is seen similarly.
Assume that (ii) holds. By the negativity of L, we have b  a < a for all a = 0 and
b < 1. Let a < 1. Then, for any n ≥ 1, either a n+1 < a n or a n = 0. As L is finite, the
latter possibility applies for a sufficiently large n. It follows that L is Archimedean.
Assume that (ii) does not hold. Let a = 0 and b < 1 such that b  a = a. As L is
negative, we then have a ≤ b and it follows bn ≥ bn−1  a = a > 0 for any n ≥ 2.
Hence L cannot be Archimedean.

We next see how Rees quotients are formed in our framework.

Proposition 7 Let (L 2 ; , ∼, (1,1)) be a negative tomonoid partition and let q ∈ L.

Let L q = {a ∈ L : a > q} ∪ ˙ {0}, where 0 is a new element, and endow L q with the
total order extending the total order on {a ∈ L : a > q} such that 0 is the bottom
element. Then, for each c ∈ L q  , the ∼-class of c is contained in (L q  )2 . Let ∼q be the
equivalence relation on L q 2 whose classes are the ∼-classes of each c ∈ L q  as well
as the subset of L q 2 containing the remaining elements. Then (L q 2 ; , ∼q , (1,1)) is
the Rees quotient of L 2 by q.

Proof Let (L; ≤, , 1) be the corresponding negative tomonoid. Let q be the binary
operation on L q such that (L q ; ≤, q , 1) is (under the obvious identifications) the
Rees quotient of L by q. Let (L q 2 ; , ∼q , (1,1)) be the associated tomonoid partition.
Let a, b, c ∈ L such that c > q and (a, b) ∼ c. Then a, b ≥ c by Lemma 6(i) and
consequently a, b > q. We conclude that the ∼-class of each c ∈ L q  is contained
in (L q  )2 .
We have to show ∼q = ∼q . Let a, b, c ∈ L q such that c = 0. Then (a, b) ∼q c iff
a q b = c iff a  b = c iff (a, b) ∼ c. Hence the ∼q -class of each c ∈ L q  coincides
with the ∼-class of c. There is only one further ∼q -class, the ∼q -class of 0, which
consequently consists of all elements of L q 2 not belonging to the ∼-class of any
c ∈ L q .

We may interpret Proposition 7 once again geometrically. Let L 2 be a finite nega-

tive tomonoid partition and let q be an element of the underlying tomonoid L. Then
the Rees quotient by q arises from the partition on L 2 by removing all columns and
rows indexed by elements ≤ q and by adding instead a single new column from left
and a single new row from below. Moreover, all elements that originally belonged
to a class of some a ≤ q are joined into a single class, which is the class of the new
zero. In contrast, the classes of elements strictly larger than q remain unchanged.
Figure 6 shows the chain obtained from a eight-element tomonoid by applying
this procedure repeatedly to the respective atom.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 103

0 u v w x y z 1 0 v w x y z 1 0 w x y z 1
0 u v w x y z 1 1 0 v w x y z 1 1 0 w x y z 1 1
0 0 u u v w x z z 0 0 0 v w x z z 0 0 0 w x z z
0 0 0 0 u u v y y 0 0 0 0 0 v y y 0 0 0 0 0 y y
0 0 0 0 u u v x x 0 0 0 0 0 v x x 0 0 0 0 0 x x
0 0 0 0 0 0 u w w 0 0 0 0 0 0 w w 0 0 0 0 0 w w
0 0 0 0 0 0 u v v 0 0 0 0 0 0 v v 0 0 0 0 0 0 0
0 0 0 0 0 0 0 u u 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0

0 x y z 1 0 y z 1 0 z 1 0 1 1
0 x y z 1 1 0 y z 1 1 0 z 1 1 0 1 1 1 1
0 0 0 x z z 0 0 0 z z 0 0 z z 0 0 0
0 0 0 0 y y 0 0 0 y y 0 0 0 0
0 0 0 0 x x 0 0 0 0 0
0 0 0 0 0 0

Fig. 6 Starting from the eight-element tomonoid shown in Fig. 4, the successive formation of Rees
quotients by the atom leads eventually to the trivial tomonoid

4.4 One-Element Coextensions

Based on the level-set representation, we will in this subsection provide a systematic

description of all one-element coextensions of a finite, negative tomonoid. We will
restrict to the Archimedean case; for the general case we refer to [24]. That is, we
will determine the coextensions of Archimedean f.n. tomonoids that are Archimedean
again.
We will proceed, roughly, as follows. We start from a tomonoid partition, seen as
a partitioned square array; cf. Fig. 4. We enlarge the sides of this square by one ele-
ment, doubling the lowest row and left-most column. We determine the equivalence
relation ∼ ¯ that makes the enlarged square into a tomonoid partition in two steps.
We first determine an intermediate equivalence relation ∼, ˙ called the ramification.
∼˙ has a universal property: the level equivalence of any Archimedean one-element
coextension extends ∼. ˙ Second, we choose the final equivalence relation ∼, ¯ merging
certain ∼-classes
˙ such that the part of the square containing the classes of the new
tomonoid’s bottom element and atom is divided up into exactly two ∼-classes.
¯
For a chain (L; ≤), let us define L̄ = L  ∪ ˙ {0, α}, where 0, α are new elements,
and let us endow L̄ with the total order extending the total order on L  such that
0 < α < a for all a ∈ L  . We call ( L̄; ≤) the zero doubling extension of L.
Furthermore, let (L; ≤, , 1) be a f.n. tomonoid. We will assume that any one-
element coextension of L is of the form ( L̄; ≤, , ¯ 1). In particular, the intersection
¯
of L and L̄ is exactly L and a  b = a  b whenever a, b, a  b ∈ L  .

104 T. Vetterlein and M. Petrík

Definition 12 Let (L 2 ; , ∼, (1,1)) be an Archimedean f.n. tomonoid partition. Let

˙ {0, α} be the zero doubling extension of L. We define
L̄ = L  ∪

P = {(a, b) ∈ L̄ 2 : a, b ∈ L ∗ and there is a c ∈ L  such that (a, b) ∼ c},

Q = L̄ 2 \ P. (7)

Let ∼
˙ be the smallest equivalence relation on L̄ 2 such that the following conditions
hold:

(E1) For any (a, b), (c, d) ∈ P such that (a, b) ∼ (c, d), we have (a, b) ∼
˙ (c, d).
(E2) For any (a, b), (b, c) ∈ P and d, e ∈ L  such that (d, c), (a, e) ∈ Q, (a, b) ∼
d, and (b, c) ∼ e, we have (d, c) ∼˙ (a, e).
(E3) For any a, b, c, e ∈ L  such that (a, b) ∈ Q, (b, c) ∼ e, and c < 1, we have
(a, e) ∼
˙ 0.
Moreover, for any a, b, c, d ∈ L  such that (b, c) ∈ Q, (a, b) ∼ d, and a < 1,
we have (d, c) ∼
˙ 0.
(E4) We have (0, 1) ∼
˙ (1, 0) ∼
˙ (α, b) ∼˙ (b, α) for any b < 1, and (α, 1) ∼
˙ (1, α).
Moreover, for any (a, b), (c, d) ∈ Q such that (a, b)  (c, d) ∼˙ 0, we have
(a, b) ∼
˙ 0.
Then we call the structure ( L̄ 2 ; , ∼,
˙ (1,1)) the ramification of (L 2 ; , ∼, (1,1)).

A few remarks might help to clarify the meaning of Definition 12. Let the tomonoid
partition (L 2 ; , ∼, (1,1)) be given. The subset P of L̄ 2 consists of all pairs (a, b) ∈
L  2 whose product in L is not the bottom element. That is, P is the union of the
∼-classes of all c ∈ L  and this union lies in L  2 . We note that P is an upwards
closed subset of L̄ 2 and, consequently, its complement Q is a downward closed subset
of L̄ 2 .
The intermediate equivalence relation ∼ ˙ is determined by successive application
of conditions (E1)–(E4). We observe that ∼-equivalences
˙ involving elements of P
are required by condition (E1) only. In fact, all the ∼-classes contained in P are
∼-classes
˙ as well.
The ∼-classes
˙ contained in Q are determined by conditions (E2)–(E4). In fact,
each prescription contained in (E2) and (E3) is of the form that certain ∼-equivalences
imply that a certain pair of elements of Q is ∼-equivalent.
˙ Finally, (E4) prescribes
that the ∼-class
˙ of 0 is downward closed. We remark that Q contains the ∼-class ˙
of the bottom element 0, the ∼-class
˙ of the atom α, and possibly further ∼-classes,
˙
which contain neither (1, c) nor (c, 1) for any c ∈ L̄.
In the sequel, for two equivalence relations ∼1 and ∼2 on a set A, we say that
∼1 is coarser than ∼2 if ∼2 ⊆ ∼1 . In other words, ∼1 coarser than ∼2 if and only if
each ∼1 -class is a union of ∼2 -classes.

Lemma 8 Let (L 2 ; , ∼, (1,1)) be an Archimedean f.n. tomonoid partition and let

( L̄ 2 ; , ∼,
¯ (1,1)) be an Archimedean one-element coextension of L 2 . Furthermore,
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 105

let ( L̄ 2 ; , ∼,
˙ (1,1)) be the ramification of L 2 . Then ∼
¯ is coarser than ∼ ˙ and the
following holds: the ∼-class
¯ of each c ∈ L  coincides with the ∼-class
˙ of c, the
∼-class
¯ of 0 is downward closed, and each ∼-class
¯ contains exactly one element of
the form (1, c) for some c ∈ L̄.

Proof Let (L; ≤, , 1) and ( L̄; ≤, , ¯ 1), where L̄ = L  ∪ ˙ {0, α}, be the two
tomonoids in question.
As noted above, condition (E1) requires ∼-equivalences
˙ only between elements
of P and the remaining conditions require ∼-equivalences
˙ only between elements
of Q. Furthermore, P is the union of the ∼-classes of all c ∈ L  . By (E1), these
∼-classes are also ∼-classes.
˙ Moreover, by Proposition 7, each ∼-class of a c ∈ L 
is a ∼-class.
¯ We conclude that the ∼-class
¯ of each c ∈ L  coincides with the ∼-class
˙
of c and P is the union of these subsets.
We next check that any two elements that are ∼-equivalent
˙ according to one of the
conditions (E2)–(E4) are also ∼-equivalent.
¯ Since ∼˙ is, by assumption, the smallest
equivalence relation with the indicated properties, it will then follow that ∼ ˙ ⊆ ∼. ¯
Ad (E2): Let (a, b), (b, c) ∈ P, d, e ∈ L  , (a, b) ∼ d, and (b, c) ∼ e. Then
a, b, c ∈ L  , hence a ¯ b = a  b = d and b  ¯ c = b  c = e. Consequently, d  ¯
c = (a  ¯ b) ¯ c=a ¯ (b ¯ c) = a ¯ e, that is (d, c) ∼ ¯ (a, e).
Ad (E3): Let a, b, c, e ∈ L  , (a, b) ∈ Q, (b, c) ∼ e, and c < 1. Then a  ¯ b≤
α and hence a  ¯ e=a ¯ (b ¯ c) = (a ¯ b)  ¯ c≤α ¯ c. As L is assumed to be
Archimedean, α is the atom of L̄, and c < 1, we conclude α  ¯ c = 0. Hence (a, e) ∼ ¯
0. Similarly, we argue for the second part of (E3).
Ad (E4): As L is Archimedean, we have, for any b < 1, 0  ¯ 1=1 ¯ 0=
α ¯ b=b ¯ α = 0 by Lemma 7 and hence (0, 1) ∼ ¯ (1, 0) ∼¯ (α, b) ∼¯ (b, α). Fur-
thermore, we have (α, 1) ∼ ¯ (1, α). Finally, let (a, b), (c, d) ∈ Q and assume (a, b) 
(c, d) ∼ ¯ 0. Then a ¯ b≤c ¯ d = 0 and thus (a, b) ∼ ¯ 0 as well.
It is finally clear that the ∼-class
¯ of 0 is downward closed. The last statement
holds by condition (P2) of a tomonoid partition.

The following theorem is the main result of this section.

Theorem 5 Let (L 2 ; , ∼, (1,1)) be an Archimedean f.n. tomonoid partition and

let ( L̄ 2 ; , ∼,
˙ (1,1)) be the ramification of L 2 . Let ∼
¯ be an equivalence relation on
L that is coarser than ∼
2
˙ and such that the following holds: the ∼-class
¯ of each
c ∈ L  coincides with the ∼-class
˙ of c, the ∼-class
¯ of 0 is downward closed, and
each ∼-class
¯ contains exactly one element of the form (1, c) for some c ∈ L̄. Then
( L̄ 2 ; , ∼,
¯ (1,1)) is an Archimedean one-element coextension of L 2 .
Moreover, all Archimedean one-element coextensions of L 2 arise in this way.

Proof P, defined by (7), is the union of the ∼-classes of all c ∈ L  . As we have

seen in the proof of Lemma 8, these subsets of P are also ∼-classes.
˙ Recall also
that P is upwards closed and Q = L̄ 2 \ P is downward closed.
By (E4), we have (1, 0) ∼˙ (0, 1) and (1, α) ∼
˙ (α, 1). We claim that (1, 0) 
˙ (1, α).
Indeed, (E1), (E2), and (E3) involve only elements (a, b) such that a, b ∈ L  . Hence,
none of these prescriptions involves the elements (1, α) or (α, 1). Moreover, by (E4),
106 T. Vetterlein and M. Petrík

the elements (a, 0) and (0, a) for any a as well as (a, α) and (α, a) for any a = 1
belong to the ∼-class
˙ of (1, 0). Again, (1, α) and (α, 1) are not concerned. Finally,
the ∼-class
˙ of (1, 0) is a downward closed set. Also this prescription has no effect on
(1, α) or (α, 1) because there is no element in Q that is larger than (1, α) or (α, 1).
We conclude that {(1, α), (α, 1)} is an own ∼-class
˙ and our claim is shown.
Let now ∼ ¯ ⊇∼ ˙ be as indicated. Note that, by what we have seen so far, at least
one such equivalence relation exists. In accordance with Proposition 6, we will verify
(P1), (P2), and (P3’).
We have shown that (1, c) ∼ ¯ (c, 1) for all c ∈ L̄. By construction, ∼
¯ fulfils (P2).
Furthermore, the ∼-class
¯ of 0 is downward closed and Q, which is the union of the
∼-classes
¯ of 0 and α, is downward closed as well. We conclude that (P1) holds for
∼.
¯
It remains to show that ∼ ¯ fulfils (P3’). Let a, b, c, d, e ∈ L \ {0, 1} such that
(a, b) ∼¯ d and (b, c) ∼ ¯ e. We distinguish the following cases.
Case 1. Let d, e ∈ L  . Then (a, b) ∼ d and (b, c) ∼ e. As ∼ fulfils (P3), we have
(d, c) ∼ (a, e). In particular, it follows that (d, c) ∈ P iff (a, e) ∈ P. If (d, c) and
(a, e) are both in P, we have (d, c) ∼ ¯ (a, e) because the ∼-classes
˙ contained in P
are ∼-classes
¯ as well. If (d, c) and (a, e) are both in Q, we have (d, c) ∼ ˙ (a, e) by
(E2) and consequently also (d, c) ∼ ¯ (a, e), because ∼ ¯ extends ∼.
˙
Case 2. Let d = α and e ∈ L  . Then (d, c) ∼ ˙ 0 by (E4). Furthermore, we have
a ∈ L  by (E4), b, c ∈ L  because (b, c) ∈ P, (a, b) ∈ Q, and (b, c) ∼ e. It follows
(a, e) ∼
˙ 0 by (E3). Consequently, (d, c) ∼ ¯ 0∼ ¯ (a, e).
Case 3. Let d ∈ L  and e = α. We argue similarly to Case 2.
Case 4. Let d = e = α. Then (d, c) ∼ ˙ (a, e) ∼˙ 0 by (E4) and consequently also
(d, c) ∼¯ (a, e).
By Proposition 6, ( L̄ 2 ; , ∼,
¯ (1,1)) is a f.n. tomonoid partition, which is moreover
Archimedean by (E4) and Lemma 7. It is finally clear from Proposition 7 that the
Rees quotient of L̄ 2 by the atom α is L 2 .
The final statement follows from Lemma 8.

Let us summarise our construction and add some remarks. In order to determine the
one-element coextensions of a f.n. tomonoid L, we start from its associated tomonoid
partition (L 2 ; , ∼, (1,1)). We first determine its ramification ( L̄ 2 ; , ∼,
˙ (1,1))
according to Definition 12. This is done by means of the conditions (E1)–(E4);
note that these prescriptions are largely independent, it is not necessary to apply
them in a recursive way. To obtain, second, a coextension of the desired type, the set
Z = (1, 0) ∼¯ , i.e. the ∼-class
¯ of the bottom element, is chosen according to Theo-
rem 5. This is done as simple as follows: Z is a union of ∼-classes
˙ contained in Q
including (1, 0) ∼˙ but excluding the ∼-class
˙ {(1, α), (α, 1)}, and Z is downward
closed. Thus, to determine a specific one-element coextension, all we have to do is
to select an arbitrary set of ∼-classes
˙ different from {(α, 1), (1, α)} and Z will then
be the smallest downward closed set containing them.
Note that one possible choice is Z = Q \ {(α, 1), (1, α)}. This means that the
explained procedure always leads to a result, that is, every Archimedean, finite,
negative tomonoid has at least one Archimedean one-element coextension.
The Semantics of Fuzzy Logics: Two Approaches to Finite Tomonoids 107

Also in the general case, it is interesting that the explained procedure never requires
revisions. At no place decisions are required that lead to an impossible situation, we
may always proceed to end up with a coextension as desired.

Acknowledgments The support of the first author by the Austrian Science Fund (FWF): project
I 1923-N25 (New perspectives on residuated posets) and the support of the second author by the
Czech Science Foundation under Project 15-07724Y are gratefully acknowledged.

References

1. Aczél, J.: Quasigroups, nets and nomograms. Adv. Math. 1, 383–450 (1965)
2. Blaschke, W., Bol, G.: Geometrie der Gewebe, topologische Fragen der Differentialgeometrie
(in German). Springer, Berlin (1939)
3. Blount, K., Tsinakis, C.: The structure of residuated lattices. Int. J. Algebra Comput. 13, 437–
461 (2003)
4. Chen, W., Zhao, X.: The structure of idempotent residuated chains. Czech. Math. J. 59, 453–479
(2009)
5. Clifford, A.H., Preston, G.B.: The Algebraic Theory of Semigroups, vol. 1. American Mathe-
matical Society, Providence (1961)
6. Cox, D., Little, J., O’Shea, D.: Ideals, varieties, and algorithms. An introduction to Computa-
tional Algebraic Geometry and Commutative Algebra, 3rd edn. Springer, New York (2007)
7. De Baets, B., Mesiar, R.: Discrete triangular norms. In: Rodabaugh, S.E., et al. (Eds.), Topo-
logical and Algebraic Structures in Fuzzy Sets. A Handbook of Recent Developments in the
Mathematics of Fuzzy Sets, pp. 389–400. Kluwer Academic Publishers, Dordrecht (2003)
8. Esteva, F., Godo, L.: Monoidal t-norm based logic: towards a logic for left-continuous t-norms.
Fuzzy Sets Syst. 124, 271–288 (2001)
9. Evans, K., Konikoff, M., Madden, J.J., Mathis, R., Whipple, G.: Totally ordered commutative
monoids. Semigroup Forum 62, 249–278 (2001)
10. Fuchs, L.: Partially Ordered Algebraic Systems. Pergamon Press, Oxford (1963)
11. Galatos, N., Jipsen, P., Kowalski, T., Ono, H.: Residuated lattices. An Algebraic Glimpse at
Substructural Logics. Elsevier, Amsterdam (2007)
12. Grillet, P.A.: Semigroups. An Introduction to the Structure Theory. Marcel Dekker, New York
(1995)
13. Grillet, P.A.: Commutative Semigroups. Kluwer Academic Publishers, Dordrecht (2001)
14. Hájek, P.: Metamathematics of Fuzzy Logic. Kluwer Academic Publisher, Dordrecht (1998)
15. Horčík, R.: Structure of commutative cancellative integral residuated lattices on (0, 1]. Algebra
Univers. 57, 303–332 (2007)
16. Horčík, R.: On the structure of finite integral commutative residuated chains. J. Log. Comput.
21, 717–728 (2011)
17. Howie, J.M.: An Introduction to Semigroup Theory. Academic Press, London (1976)
18. Hulin, A.J.: Extensions of ordered semigroups. Czech. Math. J. 26, 1–12 (1976)
19. Kehayopulu, N., Tsingelis, M.: Ideal extensions of ordered semigroups. Commun. Algebra 31,
4939–4969 (2003)
20. Klement, E.P., Mesiar, R., Pap, E.: Triangular Norms. Kluwer Academic Publishers, Dordrecht
(2000)
21. Montagna, F., Noguera, C., Horčík, R.: On weakly cancellative fuzzy logics. J. Log. Comput.
16, 423–450 (2006)
22. Noguera, C., Esteva, F., Godo, L.: Generalized continuous and left-continuous t-norms arising
from algebraic semantics for fuzzy logics. Inf. Sci. 180, 1354–1372 (2010)
23. Petrich, M.: Introduction to Semigroups. Charles E. Merrill Publishing Company, Columbus
(1973)
108 T. Vetterlein and M. Petrík

24. Petrík, M., Vetterlein, Th.: Rees coextensions of finite, negative tomonoids. J. Log. Comput.
(to appear)
25. Petrík, M., Sarkoci, P.: Associativity of triangular norms characterized by the geometry of their
level sets. Fuzzy Sets Syst. 202, 100–109 (2012)
26. Vetterlein, Th.: A Representation of Finite, Positive, Commutative Tomonoids. www.flll.jku.
at/sites/default/files/u24/Direction-f-cones.pdf
27. Vetterlein, Th.: Algebraic semantics: the structure of chains. In: Cintula, P., Fermüller, C.,
Noguera, C. (eds.) Handbook of Mathematical Fuzzy Logic, vol. 3 (to appear)
28. Vetterlein, Th.: On positive commutative tomonoids. Algebra Univ. (to appear)
29. Vetterlein, Th: Totally ordered monoids based on triangular norms. Commun. Algebra 43,
2643–2679 (2015)
30. Ya, E.: Gabovich, Fully ordered semigroups and their applications. Russ. Math. Surv. 31,
147–216 (1976)
Structure of Uninorms with Continuous
Diagonal Functions

Andrea Mesiarová-Zemánková

Abstract The structure of uninorms which are continuous on some special parts
of the unit square is discussed. After a summary of partial results achieved in the
characterization of uninorms with continuous underlying t-norm and t-conorm in
the past years, a full characterization of these uninorms is described. Representation
theorems based on the set of discontinuity points of such a uninorm and the ordinal
sum construction for semigroups are presented. Further generalizations yield uni-
norms with continuous diagonal functions. Several results related to uninorms with
continuous diagonals are investigated. Further generalizations are also discussed.

1 Introduction

Uninorms (originally called uni-norms, see [49]) are functions that on the one hand
generalize both t-norms and t-conorms, and on the other hand allow bipolar behaviour
[50]. Moreover, uninorms linearly transformed to the interval [−1, 1] are just bipolar
t-conorms (see [31]). A binary function is a uninorm if it is commutative, associative,
non-decreasing in each variable and has a neutral element e ∈ [0, 1]. Therefore the
class of uninorms covers also the class of t-norms (for which e = 1) and the class
of t-conorms (for which e = 0). To distinguish uninorms which are not t-norms
or t-conorms, in later works authors assume that a uninorm has a neutral element
e ∈ ]0, 1[. Such uninorms are also called proper.
Due to the associativity the n-ary form of any uninorm is uniquely given and thus
it can be extended to an aggregation function working on n∈N [0, 1]n .
In this chapter we focus on uninorms with continuous diagonals. At first we focus
on uninorms with continuous underlying functions (for the definition of underlying
functions see Sect. 2). Afterwards we will search for conditions under which uni-
norms with continuous diagonals have continuous underlying functions. The chapter
is organized as follows. In Sect. 2 we will give a historical overview of results related

A. Mesiarová-Zemánková (B)
Mathematical Institute, Slovak Academy of Sciences, Bratislava, Slovakia
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 109

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_7
110 A. Mesiarová-Zemánková

to the characterization of uninorms with continuous underlying functions. In Sect. 3

we will show a full characterization of uninorms with continuous underlying func-
tions. The historical overview of results related to t-norms with continuous diagonals
can be found in Sect. 4, where we will discuss uninorms with continuous diagonals
and some further generalizations. We describe our conclusions and further perspec-
tives in Sect. 5.

2 Historical Overview

First results on the structure of uninorms are due to Yager and Rybalov [49] and
Fodor et al. [12]. As first observation we can introduce the fact that for a uninorm
with the neutral element e ∈ [0, 1] there is U (a, 0) = 0 for all a ≤ e, and U (a, 1) = 1
for all a ≥ e. This observation was then extended in [12] where it was shown that
for each uninorm U the restriction of U to [0, e]2 is a t-norm on [0, e]2 , i.e., a linear
transformation of some t-norm TU on [0, 1]2 and the restriction of U to [e, 1]2 is a
t-conorm on [e, 1]2 , i.e., a linear transformation of some t-conorm SU on [0, 1]2 . We
will call TU (SU ) an underlying t-norm (t-conorm) of U , and together TU and SU
will be called the underlying functions of U . If e = 1 (e = 0) then the underlying
t-conorm (t-norm) is not defined and the underlying t-norm (t-conorm) is the uninorm
itself. Moreover,

min(x, y) ≤ U (x, y) ≤ max(x, y)

for all (x, y) ∈ [0, e] × [e, 1] ∪ [e, 1] × [0, e]. Thus each uninorm has a conjunctive
behaviour if all inputs are below the neutral element, a disjunctive behaviour if all
inputs are above the neutral element, and an averaging behaviour on the reminder.
For every uninorm the value a = U (0, 1) is the annihilator of U and it holds a ∈
{0, 1} (see [12]). If U (0, 1) = 0 the uninorm U is called conjunctive (andlike) and if
U (0, 1) = 1 the uninorm U is called disjunctive (orlike). Moreover, if U : [0, 1]2 −→
[0, 1] is a continuous uninorm then e = 1 or e = 0, i.e., U is either a continuous t-
norm, or a continuous t-conorm (see [22]). This means that there is no uninorm
with a neutral element e ∈ ]0, 1[ continuous on the whole unit square. In fact for
every uninorm either the function u 1 : [0, 1] −→ [0, 1] or the function u 0 : [0, 1] −→
[0, 1] is not continuous, where u 1 (x) = U (1, x) and u 0 (x) = U (0, x) for x ∈ [0, 1].
This follows from the fact shown in [12], that if u 1 and u 0 are both continuous except
at x = e then U (0, 1) = 0 implies U (x, 1) = x for all x ∈ [0, e[, and U (0, 1) = 1
implies U (x, 0) = x for all x ∈ ]e, 1] This yields the following theorem.

Theorem 1 ([12]) Suppose that U : [0, 1]2 −→ [0, 1] is a uninorm with neutral
element e ∈ ]0, 1[ and both functions x → U (x, 1) and x → U (x, 0) (x ∈ [0, 1])
are continuous except at the point x = e. Then U is given by one of the following
forms. If U (0, 1) = 0 then
Structure of Uninorms with Continuous Diagonal Functions 111
⎧
⎪
⎨e · T ( e , e ),
x y
if (x, y) ∈ [0, e]2 ,
Umin (x, y) = e + (1 − e) · S( x−e , y−e
), if (x, y) ∈ [e, 1]2 ,
⎪
⎩
1−e 1−e
min(x, y), otherwise,

and if U (0, 1) = 1 then

⎧
⎪
⎨e · T ( e , e ),
x y
if (x, y) ∈ [0, e]2 ,
Umax (x, y) = e + (1 − e) · S( x−e , y−e
), if (x, y) ∈ [e, 1]2 ,
⎪
⎩
1−e 1−e
max(x, y), otherwise.

In both formulas T is a t-norm and S is a t-conorm.

On the other hand, from [22] we know that if T : [0, 1]2 −→ [0, 1] is a t-norm
and S : [0, 1]2 −→ [0, 1] is a t-conorm, then for any e ∈ [0, 1] the two functions
Umin , Umax : [0, 1]2 −→ [0, 1] given by
⎧
⎪
⎨e · T ( e , e )
x y
if (x, y) ∈ [0, e]2 ,
Umin (x, y) = e + (1 − e) · S( x−e , y−e
) if (x, y) ∈ [e, 1]2 ,
⎪
⎩
1−e 1−e
min(x, y) otherwise

and
⎧
⎪
⎨e · T ( e , e )
x y
if (x, y) ∈ [0, e]2 ,
Umax (x, y) = e + (1 − e) · S( x−e , y−e
) if (x, y) ∈ [e, 1]2 ,
⎪
⎩
1−e 1−e
max(x, y) otherwise

are uninorms. We will denote the set of all uninorms of the first type by Umin and of
the second type by Umax .
In the literature we can find several aggregation functions that generalize uni-
norms. In [26] Mas et al. introduced left and right uninorms, where the commuta-
tivity of uninorms was relaxed and a left uninorm possesses a left neutral element
and a right uninorm possesses a right neutral element. A uninorm without commu-
tativity was called a pseudo-uninorm in [48]. By removing the associativity and the
commutativity from the axioms of uninorms Liu introduced in [23] the concept of
semi-uninorms (on a complete lattice) and Su et al. [46] introduced the concept of
left and right semi-uninorms (on a complete lattice). Uninorms on a finite totally
ordered set were studied in [25]. By replacing the neutral element of a uninorm by
an n-neutral element we obtain n-uninorms that were studied in [2].
It is evident that the more we generalize the uninorm functions (on [0, 1]2 ), the
more complicated their structure will be. Thus in the clarification of the structure of
uninorms we should take an opposite approach and start from some special subclasses
of uninorms with less complicated structure.
From the observations made above we see that the structure of t-norms and
t-conorms plays a major role in the investigation of the structure of uninorms. Since
112 A. Mesiarová-Zemánková

t-norms and t-conorms are dual to each other, also their structure is similar. Let
us focus for a while on the class of t-norms. Although the class of left-continuous
t-norms was not yet fully characterized, and several peculiar examples of t-norms
belong to this class, the class of continuous t-norms (t-conorms) has a simple char-
acterization. Each continuous t-norm is an ordinal sum of continuous Archimedean
t-norms, and each continuous Archimedean t-norm is generated by a continuous
additive generator. Note that a t-norm T (t-conorm S) is called Archimedean if for
all x, y ∈ ]0, 1[ there exists an n ∈ N such that x T(n) < y, (x S(n) > y) where

x T(n) = T (x, T (x, . . .)), x S(n) = S(x, S(x, . . .))

 
 

n-times n-times

(see [19]). A continuous t-norm (t-conorm) is Archimedean if and only if it has

only trivial idempotent elements 0 and 1. A continuous Archimedean t-norm T
(t-conorm S) is either strict, i.e., strictly increasing on ]0, 1]2 (on [0, 1[2 ), or nilpotent,
i.e., there exists (x, y) ∈ [0, 1]2 such that T (x, y) = 0 (S(x, y) = 1).
The simplicity of the structure of continuous t-norms and t-conorms encouraged
several authors to believe that the structure of uninorms with continuous underlying
functions could inherit a similar easy characterization. As we will see later, the
structure of uninorms with continuous underlying functions is similar to the structure
of continuous t-norms in a number of aspects, however, there are some peculiarities
that are different. Anyhow, in both cases we rely on two main construction methods:
the ordinal sum construction and the construction based on an additive generator.
Recall an original definition of an ordinal sum of semigroups by Clifford [5].

Theorem 2 Let A = ∅ be a totally ordered set and (G α )α∈A with G α = (X α , ∗α ) be

a family of semigroups. Assume that for all α, β ∈ A with α < β the sets X α and X β
are either disjoint or that X α ∩ X β = {xα,β }, where xα,β is both the neutral element
of G β and where for each γ ∈ A with α < γ < β we have
of G α and the annihilator
X γ = {xα,β }. Put X = X α and define the binary operation ∗ on X by
α∈A

⎧
⎪
⎨x ∗α y if (x, y) ∈ X α × X α ,
x∗y= x if (x, y) ∈ X α × X β and α < β,
⎪
⎩
y if (x, y) ∈ X α × X β and α > β.

Then G = (X, ∗) is a semigroup. The semigroup G is commutative if and only if for

each α ∈ A the semigroup G α is commutative.

Each continuous t-norm has a closed set of idempotent points I (see [19]) and
thus [0, 1] \ I is equal to a union of open disjoint subintervals ]ak , bk [ of [0, 1]
which define supports for summands in the ordinal sum representation, i.e., the
corresponding semigroup is defined on [ak , bk [. If a t-norm is isomorphic to a strict
t-norm on [ak , bk ]2 we can further divide this summand into an ordinal sum of
semigroups defined on {ak } and on ]ak , bk [, however, if T is on [ak , bk ]2 isomorphic
Structure of Uninorms with Continuous Diagonal Functions 113

min SU1 max SU2

a a

TU1 min TU2 max

Fig. 1 The uninorm U1 (left) and the uninorm U2 (right) from Remark 1. The bold lines denote
the points of discontinuity of U1 and U2

to a nilpotent t-norm this is not possible. In the case of t-norms the distinction between
{ak } with ]ak , bk [ and [ak , bk [ plays no role, however, as we will see, in the case of
uninorms this will be a crucial observation.
Remark 1 Assume U1 ∈ Umin and U2 ∈ Umax with respective neutral elements
e1 , e2 ∈ ]0, 1[. Then U1 and U2 are ordinal sums of semigroups in the sense of Clif-
ford. In the first case we have three semigroups G a = ([0, e1 [, TU1 ),
G b = (]e1 , 1], SU1 ), G c = ({e1 }, min), A1 = {a, b, c} and the order on the set
A1 is given by a < b < c. In the second case we have again three semigroups
G a = ([0, e2 [, TU2 ), G b = (]e2 , 1], SU2 ), G c = ({e2 }, max), A2 = {a, b, c} and the
order on the set A2 is given by b < a < c. We can see the uninorms U1 and U2 in
Fig. 1.
In [12] we can find another example of uninorms that are related to ordinal sum
of semigroups, namely pseudo-continuous uninorms. A uninorm U with neutral ele-
ment e is called pseudo-continuous on [0, 1]2 if and only if U is continuous on the
set [0, 1]2 \ {(x, y) | x = e or y = e}, i.e., U is continuous on the unit square except
the segments [(e, 0), (e, 1)] and [(0, e), (1, e)]. For a pseudo-continuous conjunc-
tive uninorm U with neutral element e ∈ ]0, 1[ we have U ∈ Umin , where TU and
SU are continuous in the open unit square and thus they are either continuous, or
obtained from an ordinal sum of a continuous t-norms (t-conorms) and a t-subnorm
(t-superconorm). More details on uninorms such that TU and SU are continuous in
the open unit square can be found in Sect. 4.2. Similarly, for a pseudo-continuous
disjunctive uninorm U with neutral element e ∈ ]0, 1[ we have U ∈ Umax and TU
and SU are continuous in the open unit square.
So far we know that each continuous t-norm is equal to an ordinal sum of
Archimedean t-norms. Furthermore, each Archimedean t-norm possesses a contin-
uous additive generator.
Proposition 1 Let t : [0, 1] −→ [0, ∞] be a continuous strictly decreasing function
such that t (1) = 0. Then the binary operation T : [0, 1]2 −→ [0, 1] given by

T (x, y) = t −1 (min(t (0), t (x) + t (y)))

is a continuous t-norm. The function t is called an additive generator of T .

114 A. Mesiarová-Zemánková

Extending the concept of additive generators also to the class of uninorms Fodor et
al. defined representable uninorms in [12] (which were before independently studied
as associative compensatory operators in [18]). A uninorm U : [0, 1]2 −→ [0, 1]
is called representable if there exists a continuous, strictly increasing function
h : [0, 1] −→ [−∞, ∞], h(0) = −∞, h(1) = ∞ such that

U (x, y) = h −1 (h(x) + h(y)).

The function h is called an additive generator of the uninorm U . Note that if we relax
the strict monotonicity of h then the neutral element of the generated function will be
lost. If we relax conditions h(0) = −∞, h(1) = ∞ then if h(0) = 0, h(1) > 0 we
obtain a t-conorm, if h(1) = 0, h(0) < 0 we obtain a t-norm, if h(0) > 0, h(1) > 0
we obtain a t-subnorm (see Sect. 4.2) and if h(0) < 0, h(1) < 0 we obtain a t-
superconorm. In the case that h(0) < 0, h(1) > 0 the associativity will be lost.
Further, a uninorm U : [0, 1]2 −→ [0, 1] is representable if and only if U is con-
tinuous everywhere on the unit square except for the two points (0, 1) and (1, 0)
([1, 6, 11, 31, 42]), i.e., U is almost continuous. This implies that if U is almost
continuous then it is strictly increasing on the open unit square, i.e., TU and SU are
a strict t-norm and a strict t-conorm. On the other hand, from [14] it follows that if
T and S are a strict t-norm and a strict t-conorm then there exists a representable
uninorm U such that TU = T and SU = S.
In our summary we have already mentioned almost continuous uninorms, i.e.,
uninorms continuous on [0, 1] \ {(0, 1), (1, 0)}, pseudo-continuous uninorms, i.e.,
uninorm continuous on [0, 1]2 \ {(x, y) | x = e or y = e}, and now let us focus
on uninorms continuous on ]0, 1[2 . In [14] we can find several interesting results about
this class of uninorms. First, if U is a uninorm with the neutral element e ∈ ]0, 1[ and
there exists a u ∈ [0, e[ such that U (x, y) = x for all x ∈ ]u, e[ and y ∈ ]e, 1[, then
U is not continuous in ]0, 1[2 . Further, these uninorms have a very clear structure.
Theorem 3 ([14]) Assume that U : [0, 1]2 −→ [0, 1] is a uninorm with the neutral
element e ∈ ]0, 1[ and U is continuous on ]0, 1[2 . Then U can be represented as
(i) or (ii) :
(i) ⎧
⎪
⎪ e · TU ( xe , ey ), if x, y ∈ [0, u],
⎪
⎪
⎪
⎪ h −1 (h(x) + h(y)), if x, y ∈ ]u, 1[,
⎪
⎪
⎨x, if x ∈ [0, u], y ∈ ]u, 1[,
U (x, y) =
⎪
⎪ x, if x ∈ [0, λ[, y = 1,
⎪
⎪
⎪
⎪ if x ∈ ]λ, 1], y = 1,
⎪
⎪
1,
⎩
x or 1, if x = λ, y = 1,

where, u ∈ [0, e[, λ ∈ [0, u], U (λ, λ) = λ, function h : [u, 1] −→ [−∞, ∞]

is continuous, strictly increasing, and h(u) = −∞, h(e) = 0, h(1) = ∞. Note
that the values of U on the remaining parts of [0, 1]2 are defined in such a way
that U is commutative.
Structure of Uninorms with Continuous Diagonal Functions 115

(ii) ⎧
⎪
⎪e + (1 − e) · SU ( x−e , y−e
), if x, y ∈ [v, 1],
⎪
⎪
1−e 1−e
⎪
⎪ −1
r (r (x) + r (y)), if x, y ∈ ]0, v[,
⎪
⎪
⎨x, if x ∈ [v, 1], y ∈ ]0, v[,
U (x, y) =
⎪
⎪ x, if x ∈ ]ω, 1], y = 0,
⎪
⎪
⎪
⎪0, if x ∈ [0, ω[, y = 0,
⎪
⎪
⎩
x or 0, if x = ω, y = 0,

where, v ∈ ]e, 1], ω ∈ [v, 1], U (ω, ω) = ω, function r : [0, v] −→ [−∞, ∞] is con-
tinuous, strictly increasing, and r (0) = −∞, r (e) = 0, r (v) = ∞.

Remark 2 Assume a uninorm with the neutral element e ∈ ]0, 1[, continuous on
]0, 1[2 , which has the form (i) with λ > 0. This uninorm is an ordinal sum of semi-
groups in the sense of Clifford. Here we have five semigroups G a = ([0, λ[, U ),
G b = ({λ}, U ), G c = (]λ, u[, U ), G d = ([u, 1[, U ), and G f = ({1}, U ), with
A1 = {a, b, c, d, f } and the order on the set A1 is given by a < b < f < c < d
if U (1, λ) = λ and by a < f < b < c < d if U (1, λ) = 1. Similarly, if U has the
form (ii) with ω < 1 then it is an ordinal sum of semigroups in the sense of Clifford.
Here we have five semigroups G a = ({0}, U ), G b = (]0, v], U ), G c = (]v, ω[, U ),
G d = ({ω}, U ), and G f = (]ω, 1], U ), with A2 = {a, b, c, d, f } and the order on the
set A2 is given by f < d < a < c < b if U (0, ω) = ω and by f < a < d < c < b
if U (0, ω) = 0.

Another important subclass of uninorms are idempotent uninorms, i.e., uninorms

where U (x, x) = x for all x ∈ [0, 1]. In the case of t-norms and t-conorms there is
only one idempotent t-norm—the minimum, and only one idempotent t-conorm—
the maximum. Therefore idempotent uninorms are uniquely given and continuous
on [0, e]2 ∪ [e, 1]2 . Idempotent uninorms were studied in several papers and for our
purposes we will recall results of [3, 24, 41] (for more information see the references
in these papers). From [3] we see that every idempotent uninorm U is internal, i.e.,
U (x, y) ∈ {x, y} holds for all (x, y) ∈ [0, 1]2 . Further, idempotent uninorms that
are left-continuous, or right-continuous were characterized in [3]. The complete
characterization of idempotent uninorms from [24] was later corrected in [41]. In the
following a non-increasing function g : [0, 1] −→ [0, 1] is called Id-symmetrical if
its completed graph Fg is Id-symmetrical, i.e., (x, y) ∈ Fg if and only if (y, x) ∈ Fg .
Note that a completed graph was defined in [41] as follows: let g : [0, 1] −→ [0, 1]
be any decreasing function and let G be the graph of g, that is

G = {(x, g(x)) | x ∈ [0, 1]};

for any point of discontinuity s of g, let s − and s + be the corresponding lateral limits.
Then, we define the completed graph of g, denoted by Fg , as the set obtained from
G by adding the vertical segments in any discontinuity point s, from s − to s + .
116 A. Mesiarová-Zemánková

Theorem 4 Consider e ∈ ]0, 1[. The following items are equivalent:

(i) U is an idempotent uninorm with neutral element e.
(ii) There exists a decreasing, Id-symmetrical function g : [0, 1] −→ [0, 1] with fixed
point e such that U is for all (x, y) ∈ [0, 1]2 given by
⎧
⎪
⎨min(x, y) if y < g(x) or y = g(x), x < g(g(x)),
U (x, y) = max(x, y) if y > g(x) or y = g(x), x > g(g(x)),
⎪
⎩
x or y if y = g(x), x = g(g(x)),

being commutative on the set of points (x, g(x)) such that x = g(g(x)).
For more details we recommend [41]. Idempotent uninorms on finite ordinal scales
were studied in [4].
A uninorm U which is internal on A(e) = [0, e] × [e, 1] ∪ [e, 1] × [0, e], i.e.,
U (x, y) ∈ {x, y} for all (x, y) ∈ A(e) is called locally internal on A(e). For uni-
norms locally internal on A(e) we have a result from [8](see also [7]) which shows
that if U : [0, 1]2 −→ [0, 1] is a uninorm locally internal on A(e) with neutral ele-
ment e ∈ ]0, 1[ then there exists a non-increasing function g : [0, 1] −→ [0, 1] with
fixed point e, such that inf{y ∈ [0, 1] | g(y) = g(x)} ≤ g(g(x)) ≤ sup{y ∈ [0, 1] |
g(y) = g(x)} for all x ∈ [0, 1], g(x) = 0 for all x > g(0), g(x) = 1 for all x < g(1),
and for all (x, y) ∈ A(e) there is
⎧
⎪
⎨min(x, y) if y < g(x), or y = g(x), x < g(g(x)),
U (x, y) = max(x, y) if y > g(x), or x = g(x), x > g(g(x)),
⎪
⎩
x or y if y = g(x) and x = g(g(x)).

Further, U |[0,e]2 is an ordinal sum of t-norm summands defined on intervals [ai , bi ],

i ∈ A1 such that ]ai , bi [ ⊂ [0, e] \ {g(x) | x ∈ [e, 1]} for all i ∈ A1 , and U |[e,1]2 is
an ordinal sum of t-conorm summands defined on intervals [ci , di ], i ∈ A2 such that
]ci , di [ ⊂ [e, 1] \ {g(x) | x ∈ [0, e]} for all i ∈ A2 .
On the other hand, satisfaction of all these conditions is not enough to ensure that
U is associative. Both necessary and sufficient conditions for a uninorm to be locally
internal on A(e) can be found in [9].
Finally, we recall two additional important classes of uninorms: uninorms with
underlying t-norm and t-conorm given as ordinal sums and Archimedean uninorms,
i.e., uninorms where both the underlying t-norm as well as the underlying t-conorm
are Archimedean. From [10] we can obtain the following result: let U be a uninorm,
and a, b, c, d ∈ [0, 1], a ≤ b ≤ e ≤ c ≤ d be such that b is the neutral element of
U |[a,b]2 and c is the neutral element of U |[c,d]2 ; then the set ([a, b] ∪ [c, d])2 is closed
under U . Several other results from [10] were extended in [34], where it was also
shown that if the underlying t-norm and t-conorm of a uninorm U are continuous then
for all idempotent points a, b, c, d ∈ [0, 1], a ≤ b ≤ e ≤ c ≤ d, of a uninorm U the
set ([a, b[∪{U (b, c)}∪]c, d])2 is closed under U . This shows that in the investigation
of uninorms with continuous underlying t-norm and t-conorm given as ordinal sums,
Structure of Uninorms with Continuous Diagonal Functions 117

the class of Archimedean uninorms with continuous underlying functions plays an

indispensable role. Further, for a uninorm with continuous underlying functions for
every idempotent point a ∈ [0, 1] of U we have U (x, y) ∈ {x, y} for all (x, y) ∈
{a} × [0, 1] ∪ [0, 1] × {a}. Thus if either TU = min, or SU = max then U is locally
internal on A(e).
Let us now focus on the class of Archimedean uninorms with continuous underly-
ing functions. Recall that a continuous Archimedean t-norm (t-conorm) is either strict
or nilpotent. First let us focus on the case when both TU and SU are strict. The main
result from [11] (compare also [40]) says that for a uninorm U : [0, 1]2 −→ [0, 1]
with the neutral element e ∈ ]0, 1[ such that both TU and SU are strict one of the
following three statements hold:
(i) U ∈ Umin ,
(ii) U ∈ Umax ,
(iii) U is representable.
Here, if we assume any (x0 , y0 ) ∈ ]0, e[ × ]e, 1[ ∪ ]e, 1[ × ]0, e[, then if
U (x0 , y0 ) = min(x0 , y0 ) we have U ∈ Umin , if U (x0 , y0 ) = max(x0 , y0 ) then we
have U ∈ Umax , and if min(x0 , y0 ) < U (x0 , y0 ) < max(x0 , y0 ) then U is repre-
sentable. This result was later corrected in [21] where it was observed that if
U (x0 , y0 ) = min(x0 , y0 ) (U (x0 , y0 ) = max(x0 , y0 )) for some (x0 , y0 ) ∈ ]0, e[ ×
]e, 1[ ∪, ]e, 1[ × ]0, e[ the uninorm U does not necessarily belong to Umin (Umax )
since it can differ on the boundary of the unit square. More precisely, the following
result was shown.

Theorem 5 ([21]) Let U : [0, 1]2 −→ [0, 1] be a uninorm with the neutral element
e ∈ ]0, 1[ such that both TU and SU are strict then one of the following seven state-
ments holds:
(i) U ∈ Umin ,
(ii) ⎧
⎪
⎪ e · TU ( xe , ey ) if (x, y) ∈ [0, e]2 ,
⎪
⎨ e + (1 − e) · SU ( x−e , y−e
) if (x, y) ∈ [e, 1]2 ,
U (x, y) = 1−e 1−e
⎪1
⎪ if x = 1 or y = 1,
⎪
⎩
min(x, y) otherwise,

(iii)
⎧
⎪
⎪ e · TU ( xe , ey ) if (x, y) ∈ [0, e]2 ,
⎪
⎨e + (1 − e) · S ( x−e , y−e
U 1−e ) if (x, y) ∈ [e, 1]2 ,
U (x, y) = 1−e
⎪1
⎪ if x = 1, y > 0 or y = 1, x > 0,
⎪
⎩
min(x, y) otherwise,

(iv) U ∈ Umax ,
118 A. Mesiarová-Zemánková

(v) ⎧
⎪
⎪e · TU ( xe , ey ) if (x, y) ∈ [0, e]2 ,
⎪
⎨e + (1 − e) · S ( x−e , y−e
U 1−e ) if (x, y) ∈ [e, 1]2 ,
U (x, y) = 1−e
⎪
⎪0 if x = 0 or y = 0,
⎪
⎩
max(x, y) otherwise,

(vi)
⎧
⎪
⎪ e · TU ( xe , ey ) if (x, y) ∈ [0, e]2 ,
⎪
⎨e + (1 − e) · S ( x−e , y−e
U 1−e ) if (x,y) ∈ [e, 1]2 ,
U (x, y) = 1−e
⎪
⎪ 0 if x = 0, y < 1 or y = 0, x < 1,
⎪
⎩
max(x, y) otherwise,

(vii) U is representable.

As we mentioned before, if U ∈ Umin , or U ∈ Umax , then U is an ordinal sum

in the sense of Clifford. Similarly, six of the previous seven forms of a uninorm
with strict TU and SU are ordinal sums in the sense of Clifford. Since both TU and
SU are strict we can divide a semigroup acting on [0, e[ into two semigroups acting
on {0} and ]0, e[ and similarly we can divide semigroup acting on ]e, 1] into two
semigroups acting on and ]e, 1[ and {1}. By changing the order of these semigroups
in the ordinal sum construction we can obtain all six cases mentioned above, except
the case when U is representable (see also [35]).
Now assume that both TU and SU are nilpotent. In such a case we know that the
semigroup acting on [0, e[ (]e, 1]) cannot be further divided. In [21] we can find the
following result.

Theorem 6 ([21]) Let U : [0, 1] −→ [0, 1]2 be a uninorm with the neutral element
e ∈ ]0, 1[ such that both TU and SU are nilpotent. Then either one of the following
two statements holds:
(i) U ∈ Umin ,
(ii) U ∈ Umax .

If we focus on a general Archimedean uninorm with continuous underlying oper-

ations, from [21, 34, 35] we can see that if TU is strict and SU is nilpotent (TU is
nilpotent and SU is strict) then U can have a form (i), (iv), or (v) ((i), (ii), or (iv))
from Theorem 5.
Now we have presented all important results that form a basis for the characteriza-
tion of uninorms with continuous underlying t-norm and t-conorm. In the following
section we will show this characterization.
Structure of Uninorms with Continuous Diagonal Functions 119

3 Characterization of Uninorms with Continuous

Underlying Functions

As a first step towards the characterization of uninorms with continuous underlying

t-norm and t-conorm we should recall the ordinal sum construction for uninorms
which was introduced in [32]. For any 0 ≤ a ≤ b < c ≤ d ≤ 1, v ∈ [b, c], and
a uninorm U with the neutral element e ∈ [0, 1] we will use a transformation
f : [0, 1] −→ [a, b[ ∪ {v} ∪ ]c, d] given by
⎧
⎪
⎨(b − a) · e + a if x ∈ [0, e[,
x

f (x) = v if x = e, (1)
⎪
⎩ (1−x)(d−c)
d − (1−e) otherwise.

Then f is linear on [0, e[ and on ]e, 1] and thus it is a piece-wise linear isomorphism
of [0, 1] to ([a, b[ ∪ {v} ∪ ]c, d]) and if U : [0, 1] −→ [0, 1]2 is a uninorm then a
binary function Uva,b,c,d : ([a, b[ ∪ {v} ∪ ]c, d])2 −→ ([a, b[ ∪ {v} ∪ ]c, d]) given by

Uva,b,c,d (x, y) = f (U ( f −1 (x), f −1 (y))) (2)

is a uninorm on ([a, b[ ∪ {v} ∪ ]c, d])2 . Note that we assume that a = b (c = d) if and
only if e = 0 (e = 1). The function f is piece-wise linear, however, more generally
we can use any increasing isomorphic transformation.
If U1 and U2 are uninorms, with respective neutral elements e1 , e2 , then for 0 ≤
a < b < c < d ≤ 1 we have

(U1 )a,b,c,d
v (x, y) = (U2 )a,b,c,d
v (x, y)

if and only if U1 (x, y) = φ−1 (U2 (φ(x), φ(y))), where φ : [0, 1] −→ [0, 1] is a
strictly increasing isomorphism with φ(e1 ) = e2 which is linear on [0, e1 ] and on
[e1 , 1]. Similar result can be obtain for the case when a = b (c = d), however, then
only the corresponding parts of uninorms are isomorphic.
Now we have the following result
Proposition 2 ([32]) Assume e ∈ [0, 1]. Let K be an index set which is finite or
countably infinite and let (]ak , bk [)k∈K be a disjoint
 system of open subintervals
(which can be also empty) of [0, e], such that k∈K [ak , bk ] = [0, e]. Similarly, let
(]ck , dk [)k∈K be a disjoint
 system of open subintervals (which can be also empty)
of [e, 1], such that k∈K [ck , dk ] = [e, 1]. Let further these two systems be anti-
comonotone, i.e., bk ≤ ai if and only if ck ≥ di for all i, k ∈ K . Assume a family of
uninorms (Uk )k∈K on [0, 1]2 such that if both ]ak , bk [ and ]ck , dk [ are non-empty then
Uk is a proper uninorm, otherwise if ]ak , bk [ is non-empty then Uk is a t-norm and if
]ck , dk [ is non-empty then Uk is a t-conorm. If both ]ak , bk [ and ]ck , dk [ are empty then
ak = bk = ak1 = bk1 and ck = dk = ck1 = dk1 does not hold for any k1 ∈ K , k = k1 ,
and here only the value Uk (0, 1) is interesting. Denote K ∗ = {k ∈ K |]ak , bk [= ∅}
120 A. Mesiarová-Zemánková

and K ∗ = {k ∈ K |]ck , dk [ = ∅}. Further, let B = {bk | k ∈ K } \ {ak | k ∈ K ∗ } and

C = {ck | k ∈ K } \ {dk | k ∈ K ∗ }. We define a function n : B −→ B ∪ C given for
all bk ∈ B by 
bk if Uk (1, 0) = 0,
n(bk ) =
ck else.

Let the ordinal sum U e = (ak , bk , ck , dk , Uk  | k ∈ K )e be given by

⎧
⎪
⎪ y if x = e,
⎪
⎪
⎪
⎪ x if y = e,
⎪
⎪
⎪
⎪ (U ak ,bk ,ck ,dk
k )vk if (x, y) ∈ ([ak , bk [∪]ck , dk ])2 ,
⎪
⎪
⎪
⎪
⎪
⎪ x if y ∈ [bk , ck ], x ∈ [ak , dk ] \ [bk , ck ],
⎪
⎪
⎪
⎪ y if x ∈ [bk , ck ], y ∈ [ak , dk ] \ [bk , ck ],
⎪
⎪
⎪
⎪min(x, y) if (x, y) ∈ [bk , ck ]2 \ (]bk , ck [2 ∪{(bk , ck ), (ck , bk )}),
⎪
⎪
⎨
where bk ∈ B, ck ∈ C, x + y < ck + bk ,
U e (x, y) =
⎪
⎪max(x, y) if (x, y) ∈ [bk , ck ]2 \ (]bk , ck [2 ∪{(bk , ck ), (ck , bk )}),
⎪
⎪
⎪
⎪ where bk ∈ B, ck ∈ C, x + y > ck + bk ,
⎪
⎪
⎪
⎪n(bk ) if (x, y) = (bk , ck )or (x, y) = (ck , bk ), bk ∈ B, ck ∈ C,
⎪
⎪
⎪
⎪
⎪
⎪min(x, y) if (x, y) ∈ {bk } × [bk , ck ] ∪ [bk , ck ] × {bk }
⎪
⎪
⎪
⎪ and bk ∈ B, ck ∈ / C,
⎪
⎪
⎪
⎪ if (x, y) ∈ {ck } × [bk , ck ] ∪ [bk , ck ] × {ck }
⎪
⎪max(x, y)
⎩
and bk ∈ / B, ck ∈ C,

where vk = ck (vk = bk ) if there exists an i ∈ K such that bk = ai and Ui is disjunctive

(conjunctive) and vk = n(bk ) if bk ∈ B, ck ∈ C, vk = bk if bk ∈ B, ck ∈ / C, vk = ck if
bk ∈/ B, ck ∈ C, and (Uk )avkk ,bk ,ck ,dk is given by the formula (2). Then U e is a uninorm.
Inthe case of t-norms (t-conorms)
 ordinal sum can be defined also on such intervals
that k∈K [ak , bk ] = [0, 1] ( k∈K [ck , dk ] = [0, 1]). In such a case the remaining
parts
 of the unit square are simply filled in by the min (max). In the case of uninorms, if
k∈K [a k , bk ]  = [0, e] ( k∈K [ck , dk ]  = [e, 1]) the remaining parts of the unit square
should be filled in by internal uninorms, which are, however, not unique and therefore
we have to specify them.
If for a summand ak , bk , ck , dk , Uk  for some k ∈ K we have ak = bk and ck = dk
we say that this summand is empty. If for a summand ak , bk , ck , dk , Uk  for some
k ∈ K we have ak = bk and ck = dk we say that this summand is complete. We
say that an ordinal sum U e = (ak , bk , ck , dk , Uk  | k ∈ K )e is complete when all its
summands are complete.
If all summands used in this ordinal sum construction have continuous underlying
t-norm and t-conorm also U e will have continuous underlying t-norm and t-conorm.
Thus this ordinal sum construction can be used to construct uninorms with continuous
Structure of Uninorms with Continuous Diagonal Functions 121

underlying t-norm and t-conorm. From [33] we know that a uninorm U on [0, 1]2
is a complete ordinal sum of representable uninorms if and only if there exists a
continuous strictly decreasing function r : [0, 1] −→ [0, 1] with r (0) = 1, r (e) = e
and r (1) = 0 such that U is continuous on [0, 1] \ {(x, r (x)) | x ∈ [0, 1]} and U has
countably many idempotent points.
On the other hand, not every uninorm with continuous underlying t-norm and
t-conorm can be obtained as an ordinal sum of uninorms with continuous underlying
t-norm and t-conorm. Recall Theorem 5: here all cases except when U is representable
are ordinal sums in the sense of Clifford, however, only cases (i) and (iv) are ordinal
sums of uninorms. Let us see where this difference is hidden. For a detailed proof
we recommend [35].
We see that the ordinal sum of uninorms is based on the intervals of the form
[a, b[ ∪ ]c, d]. However, if the corresponding t-norm (t-conorm) that acts on [a, b]2
([c, d]2 ) is strict we can divide the semigroup acting on [a, b[ (]c, d]) into semigroups
acting on {a} and ]a, b[ (]c, d[ and {d}). By changing the order of these semigroups
in the ordinal sum construction (in the sense of Clifford) we can then obtain uni-
norms that cannot be obtained as an ordinal sum of uninorms, although they are
Archimedean. Thus in the case of Archimedean uninorms the problem is always on
the border of the unit square. In the case of t-norms (t-conorms) on the border of the
unit square we always have T = min (S = max), however, in the case of uninorms
with continuous underlying functions it is a mixture of min and max.
Recall that for a uninorm with continuous underlying t-norm and t-conorm U is
internal on {q} × [0, 1], where q ∈ [0, 1] is an idempotent point of U . After some
computations we can see that each uninorm with continuous underlying t-norm and
t-conorm is equal to an ordinal sum of uninorms with continuous underlying t-norm
and t-conorm on the set ([0, 1] \ IU )2 , where IU is the set of idempotent points of U .
Thus the difference can appear only in points (x, y) ∈ [0, 1]2 such that at least one
of x and y is idempotent.
Summarizing, we need a more general construction than the ordinal sum con-
struction that would allow semigroups defined on singletons. However, in order to
keep monotonicity, the order of these singletons cannot be changed arbitrarily, but it
depends on the other summands in the ordinal sum construction. As a first condition,
in [35] it was shown that for every idempotent point q of U the point of change
where U (x, q) = min(x, q) changes to U (x, q) = max(x, q), i.e., such a p ∈ [0, 1]
that U (x, q) = min(x, q) for all x < p and U (x, q) = max(x, q) for all x > p, is
unique and it is an idempotent point of U . Assume now that U with continuous
underlying t-norm and t-conorm is not equal to an ordinal sum of uninorms. This
means that for an ordinal sum of uninorms V such that U = V on ([0, 1] \ IU )2 there
exists an idempotent point q ∈ [0, 1] of U such that U (q, x) = V (q, x) for some
x ∈ [0, 1]. Assume that q > e (the other case is analogical). Then both restrictions
of U and V on [0, q]2 are uninorms on [0, q]2 . Let us transform these restrictions
to [0, 1]2 linearly. We obtain two uninorms U ∗ and V ∗ that are not equal on the
boundary of the unit square and V ∗ is an ordinal sum of uninorms. Now we will
recall two examples of such ordinal sums of uninorms.
122 A. Mesiarová-Zemánková

Recall Theorem 3 case (i). Here U is an ordinal sum of uninorms only if

U (1, x) = x for all x ∈ [0, u]. In such a case U is an ordinal sum of a representable
uninorm, which corresponds to a complete summand, and several non-complete sum-
mands that together represent the t-norm on [0, u]2 . Otherwise we have a separate
semigroup ({1}, U ) and values U (1, x) for x ∈ [0, 1] determine the order on the set
A. We should stress that we can separate {1} from [u, 1] since U is representable
on [u, 1]2 and thus U is a strict t-conorm on [e, 1]2 . Thus we see that if we have
an ordinal sum V ∗ of a representable uninorm and several non-complete t-norm
summands we can obtain different uninorms that coincides with V ∗ on ]0, 1[2 by
shifting the point of change, where V ∗ (x, 1) = 1 changes to V ∗ (x, 1) = x, through
all idempotent points contained in [0, u].
On the other hand, assume an ordinal sum V ∗ of a uninorm on [a, b] for
some 0 < a ≤ e ≤ b < 1 and a representable uninorm on [0, a[ ∪ ]b, 1]. Then since
V ∗ (x, 1) = 1 for all x ∈ ]0, a[ and V ∗ (x, 0) = 0 for all x ∈ ]b, 1[ and V ∗ (0, 1) is
determined by the corresponding representable uninorm, we see that due to the
monotonicity the values V ∗ (x, 1) and V ∗ (x, 0) are uniquely determined for all
x ∈ [0, 1]. Thus we cannot obtain any different uninorm by redefining the values
on the border of the unit square. From these two examples we see that the difference
can occur only if a non-complete summand appears. These observations can be sum-
marized into the construction method called the extended ordinal sum of uninorms
(compare [35]). First note that a continuous t-norm T : [0, 1]2 −→ [0, 1] (t-conorm
S : [0, 1]2 −→ [0, 1]) is called c-strict if T (x, y) ∈ ]0, 1[ (S(x, y) ∈ ]0, 1[ ) for all
(x, y) ∈ ]0, 1[2 . In the other case T (S) will be called c-nilpotent.

Proposition 3 Let U e : [0, 1]2 −→ [0, 1] be a uninorm such that there is U e =

(ak , bk , ck , dk , Uk  | k ∈ K )e , where all conditions of Proposition 2 are satisfied.
Denote G = {bk | k ∈ K , ak = bk = e, U e (bk , ck ) = bk }, H = {ck | k ∈ K , ck =
dk = e, U e (bk , ck ) = ck }, and for x ∈ G denote G x = {k ∈ K | bk = x}, for x ∈ H
denote Hx = {k ∈ K | ck = x}. Let G ∗∗ x be the closure of the set {ck | k ∈ G x } and
denote G ∗x = G ∗∗ x \ {d i }i∈K , and let H ∗∗
x be the closure of the set {bk | k ∈ Hx } and
∗ ∗∗
denote Hx = Hx \ {ai }i∈K . Further, for k ∈ G x , x ∈ G denote

{{ck }, [ck , dk [, [ck , dk ]} if SUk is c-strict,
Fk =
{{ck }, [ck , dk ]} if SUk is c-nilpotent,

for c ∈ G ∗x denote

{[0, c]} if c = inf{ck | k ∈ G x },
Fc∗ =
{[0, c[, [0, c]} else,

and for k ∈ Hx , x ∈ H denote


{∅, {ak }, [ak , bk [} if TUk is c-strict,
Jk =
{∅, [ak , bk [} if TUk is c-nilpotent,
Structure of Uninorms with Continuous Diagonal Functions 123

and for c ∈ Hx∗ denote


{[0, b[} if b = sup{bk | k ∈ Hx },
Jb∗ =
{[0, b[, [0, b]} else.

With the convention S ∪ {S1 , S2 } = {S ∪ S1 , S ∪ S2 } let

  
g : G −→ ( ([0, ck [ ∪ Fk ) ∪ Fc∗ )
x∈G k∈G x c∈G ∗x

be a function such that

 
g(x) ∈ ([0, ck [ ∪ Fk ) ∪ Fc∗
k∈G x c∈G ∗x

and let
  
h : H −→ ( ([0, ak [ ∪ Jk ) ∪ Jb∗ )
x∈H k∈Hx b∈Hx∗

be a function such that

 
h(x) ∈ ([0, ak [ ∪ Jk ) ∪ Jb∗ ,
k∈Hx b∈Hx∗

where for all x ∈ G, y ∈ H there is y ∈ g(x) if and only if x ∈ h(y).

Then the binary function V e : [0, 1]2 −→ [0, 1] given by
⎧
⎪
⎪U e (x, y) if (x, y) ∈ ([0, 1] \ (G ∪ H ))2 ,
⎪
⎪
⎪
⎪ if x ∈ G, y ∈ g(x), or y ∈ G, x ∈ g(y),
⎨min(x, y)
V (x, y) = max(x, y)
e
if x ∈ G, y ∈/ g(x), or y ∈ G, x ∈ / g(y),
⎪
⎪
⎪
⎪min(x, y) if x ∈ H, y ∈ h(x), or y ∈ H, x ∈ h(y),
⎪
⎪
⎩max(x, y) if x ∈ H, y ∈ / h(x), or y ∈ H, x ∈ / h(y)

is a uninorm, which will be called an extended ordinal sum of uninorms. We write

V e = (ak , bk , ck , dk , Uk  | k ∈ K )e .

Example 1 Assume an ordinal sum uninorm U e : [0, 1]2 −→ [0, 1] such that U e =
(0, e, e, e, T , 0, 0, e, b, C1 , 0, 0, b, 1, C2 )e , for some b, e ∈ [0, 1], 0 < e <
b < 1 (see Fig. 2) and a t-norm T and t-conorms C1 , C2 . Assume that C1 is
c-strict and C2 is c-nilpotent. Since T is a t-norm we have U e (0, e) = 0 and thus
we have G = {0}, H = ∅ for sets G, H from the previous proposition. If we denote
the summands respectively as 1, 2, 3 for K = {1, 2, 3} we get G 0 = {2, 3}. Further,
G ∗∗ ∗
x = {e, b}, G x = ∅,
124 A. Mesiarová-Zemánková

Fig. 2 The ordinal sum

uninorm U e from Example 1
max max C2∗

max C1∗ max

T∗ max max

F2 = {{e}, [e, b[, [e, b]},

F3 = {{b}, [b, 1]}.

The function g is defined only in one point 0 and its range is the set

{[0, e], [0, b[, [0, b], [0, 1]}.

Thus if V e = (0, e, e, e, T , 0, 0, e, b, C1 , 0, 0, b, 1, C2 )e , is an extended ordi-

nal sum we have V e = U e if g(0) = [0, e]. Further we can define three other different
extended ordinal sums by respectively selecting a different value/interval for g(0).
It is evident that V e and U e may differ only on {0} × [0, 1] ∪ [0, 1] × {0}. A sketch
of a more complicated example can be seen on Fig. 3.
If all summands in the extended ordinal sum of uninorms are uninorms with
continuous underlying functions also the extended ordinal sum will be a uninorm
with continuous underlying functions. Now we can present an opposite result which
will complete the characterization of uninorms with continuous underlying functions
via the extended ordinal sum construction.
Theorem 7 Let U : [0, 1]2 −→ [0, 1] be a uninorm with continuous underlying
t-norm and t-conorm, with the neutral element e ∈ [0, 1]. Then U = V e , where
V e = (ak , bk , ck , dk , Uk  | k ∈ K )e is an extended ordinal sum of uninorms for some
systems (]ak , bk [)k∈K and (]ck , dk [)k∈K satisfying all conditions of Proposition 3,
where for all k ∈ K the uninorm Uk is either internal (including the minimum t-norm
and the maximum t-conorm), or representable (including continuous Archimedean
t-norms and t-conorms).
The proof of this theorem follows from [35, Proposition12].
From the above result we see that each uninorm with continuous underlying
functions can be decomposed via the extended ordinal sum construction into internal
and Archimedean uninorms. Since these uninorms were already characterized (see
Structure of Uninorms with Continuous Diagonal Functions 125

Fig. 3 Sketch of a uninorm

m+1 m+1
which is an ordinal sum with max

m + 1 summands. The
summands 1 and m + 1 are m
q
complete, the others are qq
non-complete. The rounded
max 4
area (the line in the center)
designates the place where 3
the ordinal sum construction
and the extended ordinal sum 2
 max
construction can differ
min

max
1

 

m+1  
m+1
min

the previous section) we have now a detailed knowledge about structure of uninorms
with continuous underlying functions.
The major role in the decomposition of a uninorm with continuous underlying
functions to an extended ordinal sum of uninorms plays its characterizing multi-
function, which is yet another possibility how to characterize uninorms with con-
tinuous underlying t-norm and t-conorm. This characterizing multi-function in fact
covers the set of points of discontinuity of such a uninorm U . We will now show
results from [34] which characterize the set of points of discontinuity of a uninorm
with continuous underlying functions.
The first interesting result shows that a uninorm with continuous underlying func-
tions is either left-continuous or right continuous (or continuous) in each point from
[0, 1]2 . Next we need the definition of a multi-function.
Definition 1 A mapping p : X −→ P(Y ) is called a multi-function if for every
x ∈ X it assigns a subset of Y , i.e., p(x) ⊆ Y . A multi-function p is called
(i) non-increasing if for all x1 , x2 ∈ X, x1 < x2 there is p(x1 ) ≥ p(x2 ), i.e., for all
y1 ∈ p(x1 ) and all y2 ∈ p(x2 ) we have y1 ≥ y2 and thus Card ( p(x1 ) ∩ p(x2 )) ≤
1,
(ii) symmetric if y ∈ p(x) if and only if x ∈ p(y).
The graph of a multi-function p will be denoted by G( p), i.e., (x, y) ∈ G( p) if and
only if y ∈ p(x).
A symmetric multi-function p : [0, 1] −→ P([0, 1]) is surjective, i.e., for all y ∈ Y
there exists an x ∈ X such that y ∈ p(x), if and only if we have p(x) = ∅ for
all x ∈ X . The graph of a symmetric, surjective, non-increasing multi-function
126 A. Mesiarová-Zemánková

p : [0, 1] −→ P([0, 1]) is a connected line. For any uninorm with continuous under-
lying functions we denote A = inf{x | U (x, 0) > 0}, B = sup{x | U (x, 1) < 1}
and let a, d ∈ [0, 1] be such that U (x, y) = e for some y ∈ [0, 1] if and only if
x ∈ ]a, d[. Then either A = 1, B = 0, or A = 1, B = 0, or A = 1, B = 0. Note
that if A = 1, B = 0, then U is non-continuous in (B, 1), if A = 1, B = 0, then U
is non-continuous in (0, A), and if A = 1, B = 0 then U is non-continuous in (0, 1).
Further, we have 0 ≤ B ≤ a ≤ e ≤ d ≤ A ≤ 1. Now we can introduce the result
describing the set of points of discontinuity of a uninorm with continuous underly-
ing functions via a multi-function. Note that in the results on idempotent uninorms
and uninorms locally internal on A(e) introduced in the previous section we can see
examples of such a multi-function (see also Fig. 1).

Theorem 8 Let U : [0, 1]2 −→ [0, 1] be a uninorm with continuous underlying

t-norm and t-conorm. Then there exists a symmetric, surjective, non-increasing multi-
function r on [0, 1]2 such that U is continuous on [0, 1]2 \ R, where R = G(r ).

The corresponding multi-function r : [0, 1] −→ P([0, 1]) is given by

⎧
⎪
⎪ {1} if x ∈ ]0, B[,
⎪
⎪
⎪
⎪ {0} if x ∈ ]A, 1[,
⎪
⎪
⎨[0, B] if x = 1,
r (x) =
⎪
⎪[A, 1] if x = 0,
⎪
⎪
⎪
⎪{y | U (x, y) = e} if x ∈ ]a, d[,
⎪
⎪
⎩
{y | (x, y) ∈ R ∗ } otherwise,

where R ∗ = {(x, y) ∈ [0, 1]2 |U is non-continuous in (x,y)}.

Note that U need not to be non-continuous in all points of R. In fact, U is continu-
ous in all points from {x} × [0, 1] for all x ∈ [0, B[ ∪ ]a, d[ ∪ ]A, 1]. The symmetric
non-increasing multi-function from the previous theorem need not to be unique. The
differences can appear on ]a, d[. However, if we require additionally that U (x, y) = e
implies (x, y) ∈ G(r ) for all (x, y) ∈ [0, 1]2 , such a multi-function is uniquely given
and such a multi-function is called the characterizing multi-function of a uninorm U
with continuous underlying functions.
In the following example we will show that the existence of the symmetric, sur-
jective, non-increasing characterizing multi-function for a uninorm need not imply
that the uninorm has continuous underlying functions.

Example 2 Let U : [0, 1]2 −→ [0, 1] be given by

⎧
⎪
⎪ 0 if max(x, y) < e,
⎪
⎨x if y = e,
U (x, y) =
⎪y
⎪ if x = e,
⎪
⎩
max(x, y) otherwise.
Structure of Uninorms with Continuous Diagonal Functions 127

max max max max

0 max 0 max

Fig. 4 The uninorm U from Example 2. The bold lines denote the points of discontinuity of U
(left) and the characterizing multi-function r of U (right)

Then U ∈ Umax is a uninorm, where the underlying t-norm is the drastic product and
the underlying t-conorm is the maximum. This uninorm is non-continuous in points
from {e} × [0, e] ∪ [0, e] × {e}. Thus the corresponding multi-function is given by
(see Fig. 4) ⎧
⎪
⎪ [e, 1] if x = 0,
⎪
⎨e if x ∈ ]0, e[,
r (x) =
⎪
⎪ [0, e] if x = e,
⎪
⎩
0 otherwise.

Since U (x, y) = e implies x = y = e we see that U is continuous on [0, 1]2 \ R,

where R = G(r ) and r is a symmetric, surjective, non-increasing multi-function
such that U (x, y) = e implies (x, y) ∈ R. However, the drastic product t-norm is
not continuous and thus U does not have continuous underlying functions.

Although the existence of the symmetric, surjective, non-increasing characteriz-

ing multi-function for a uninorm does not mean that the underlying functions are
continuous, from the above example we can observe that in points from {e} × ]0, e[
the uninorm U is neither left- nor right-continuous. Therefore we get the following.

Theorem 9 ([34]) Let U : [0, 1]2 −→ [0, 1] be a uninorm which is continuous on

[0, 1]2 \ R, where R = G(r ) and r is a symmetric, surjective, non-increasing multi-
function such that U (x, y) = e implies (x, y) ∈ R. Then U has continuous under-
lying functions if and only if in each point (x, y) ∈ [0, 1]2 the uninorm U is either
left-continuous or right-continuous (or continuous).

Assume a uninorm U with continuous underlying functions and its characterizing

multi-function. The graph of this characterizing multi-function can be decomposed
into maximal segments which are either horizontal, vertical, or strictly decreasing,
128 A. Mesiarová-Zemánková

and border points of these segments are always idempotent elements of U . Then
each horizonal and vertical segment corresponds to a non-complete summand, i.e.,
a continuous t-norm or t-conorm, in the decomposition by means of the extended
ordinal sum construction. These t-norms (t-conorms) need not be Archimedean, but
could be decomposable to further non-complete summands. Moreover, each strictly
decreasing segment corresponds to an ordinal sum of complete summands, which
are representable or idempotent uninorms (in the case of idempotent uninorms their
characterizing multi-function should be strictly decreasing in this situation). Thus
by combination of decomposition into components based on idempotent points of
U and on maximal segments of its characterizing multi-function we can couple
corresponding idempotent points and obtain a full decomposition via the extended
ordinal sum construction (see [35]).

4 Uninorms with Continuous Diagonals and Further

Generalizations

In the previous section we have shown a full characterization of uninorms with

continuous underlying t-norm and t-conorm. In this section we will try to answer
the question how much we can weaken the conditions in order to get the same
characterization as in the previous sections. In other words, we will study uninorms
continuous on some parts of the unit interval and see under which conditions we
will obtain a uninorm with continuous underlying t-norm and t-conorm. In the first
part of this section we will focus on uninorms continuous on the diagonal and in the
second part we will discuss further generalizations.

4.1 Uninorms Continuous on Diagonal

Recall that for a binary operation O : [0, 1]2 −→ [0, 1] its corresponding diagonal
function d O : [0, 1] −→ [0, 1] is given by d O (x) = O(x, x) for all x ∈ [0, 1]. As
we have seen above, in the structure of uninorms a major role is played by t-norms
and t-conorms, and this holds especially when we discuss areas around the diagonal.
Therefore in the investigation of uninorms with continuous diagonals we have to first
focus on t-norms (t-conorms) with continuous diagonals. As t-norms and t-conorms
are dual all results on t-norms with continuous diagonals can be immediately obtained
also for t-conorms.
The first step in the study of t-norms with continuous diagonals was an open
problem from [43]: whether a t-norm with continuous diagonal function have to be
continuous. It is obvious that each continuous t-norm has a continuous diagonal and
the construction of all continuous t-norms with the given continuous diagonal was
described in [47] (see also [17, 27]). For the opposite problem a negative answer
Structure of Uninorms with Continuous Diagonal Functions 129

was given by Krause [20] (see also [19, 29, 44]). The Krause t-norm has a fractal-
like structure, it is Archimedean and continuous on the diagonal, however, it is not
right-continuous and it is not left-continuous on the border of the unit square.
Thus in general a t-norm with a continuous diagonal need not be continuous.
However, there are several results showing under which conditions such a t-norm is
continuous. We will start with Archimedean t-norms. In [37] continuous extensions
of t-norms known on [a, b]2 , for 0 ≤ a ≤ b ≤ 1, to the whole unit square were
investigated. From this paper we can obtain the following results:
(i) Let T be a t-norm continuous on [a, 1]2 , such that T (x, x) < x for all x ∈
[a, 1[. Then T is on [a, 1]2 conditionally cancellative, i.e., T (x, y) = T (x, z) >
T (a, a) implies y = z for all x, y, z ∈ [a, 1].
(ii) Let T be an Archimedean t-norm continuous and cancellative on [a, 1]2 , such
that T has a continuous diagonal. Then T is continuous.
(iii) Let T be an Archimedean t-norm continuous and conditionally cancellative,
but not cancellative, on [a, 1]2 . Then T is continuous.
Thus we see that for an Archimedean t-norm the continuity on [a, 1]2 for arbitrary
0 < a < 1 and continuity of the diagonal implies continuity on the whole unit square.
Moreover, such a t-norm is uniquely determined by its values on [a, 1]2 and its
diagonal function. Note that if we relax the Archimedean property then an ordinal
sum of the Krause t-norm on [0, a] for any 0 < a < 1 will yield a non-continuous
t-norm with continuous diagonal which is continuous on [a, 1]2 . Further, in the
following example we will see that the continuity of the diagonal is not implied by
the Archimedean property and continuity on [a, 1]2 .

Example 3 ([38]) Assume a t-norm T∗ : [0, 1]2 −→ [0, 1] given by

⎧
⎪
⎨x · y if x · y ≥ 41 ,
T∗ (x, y) = min(x, y) if max(x, y) = 1, min(x, y) < 41 ,
⎪
⎩
0 otherwise.

Then T∗ is an Archimedean t-norm which is continuous on [ 21 , 1]2 . However, T∗ is

not continuous on the diagonal. Note that T∗ is the weakest t-norm that coincide with
the product t-norm on [ 21 , 1]2 .

We can summarize the above results in the following Corollary.

Corollary 1 Let U : [0, 1]2 −→ [0, 1] be an Archimedean uninorm which is con-

tinuous on the diagonal. Then U has continuous underlying functions if and only if it
is continuous on [a, e]2 and [e, b]2 for some a, b ∈ [0, 1] with 0 ≤ a < e < b ≤ 1.

From the above result we see that, for example, if for a uninorm U the TU is the
Krause t-norm and SU is the t-conorm dual to the Krause t-norm then there are no
such a, b ∈ [0, 1], 0 ≤ a < e < b ≤ 1, that U it is continuous on [a, e]2 and [e, b]2 .
130 A. Mesiarová-Zemánková

In general, we have the following result for both t-norms and t-conorms.

Corollary 2 ([36]) A t-norm T : [0, 1]2 −→ [0, 1] (t-conorm S : [0, 1]2 −→ [0, 1])
is continuous if and only if it is continuous on the diagonal and for each idempotent
point x ∈ ]0, 1] (x ∈ [0, 1[) there exists a δx > 0 such that T (S) is continuous in
(x, y) for all y ∈ [x − δx , x] (y ∈ [x, x + δx ]).

From this corollary we can conclude that a uninorm U : [0, 1]2 −→ [0, 1] has
continuous underlying functions if and only if it is continuous on the diagonal and
for each idempotent point x ∈ ]0, 1[ there exists a δx > 0 such that if x ≤ e then U
is continuous in (x, y) for all y ∈ [x − δx , x], and if y ≥ e then U is continuous in
(x, y) for all y ∈ [x, x + δx ]. For more details we recommend [36]. If we again take a
uninorm U such that TU is the Krause t-norm and SU is the t-conorm dual to the Krause
t-norm then there is no such δ > 0 that U is continuous on {e} × [e − δ, e + δ].

4.2 Further Generalizations

In this subsection we will first focus on the case when we have no information about
the continuity of the diagonal of the given uninorm. In this case let us recall a result
from [13] which shows that if a cancellative Archimedean t-norm is (left-)continuous
in the point (1, 1) then it is isomorphic with the product t-norm, i.e., it is continuous
and strict. This means that if a uninorm U is cancellative on ]0, e]2 ∪ [e, 1[2 and
continuous in (e, e) then U has continuous underlying functions and thus it has a
structure described in the previous section.
Moreover, in [36] it was shown that an Archimedean t-norm which is border-
continuous (i.e., each point from the boundary of the unit square is the point of
continuity of T ) is also continuous. This means that an Archimedean uninorm which
is left-continuous in all points from [0, e] × {e} and right-continuous in all points
from [e, 1] × {e} has continuous underlying functions.
From these two results we see that the Archimedean property can ‘spread’ the
continuity from some part of the unit square to the whole unit square. For both
results, the Archimedean property of the corresponding uninorm is crucial. In the first
case recall an example of a non-Archimedean, left-continuous, cancellative t-norm
([28, 45]) which is given by T (x, y) = z = 1
2z i
for all (x, y) ∈ ]0, 1]2 , where
1 1 i∈N
x= 2 xi
, y= 2 yi
, and {xi }i∈N , {yi }i∈N , {z i }i∈N are increasing sequences of
i∈N i∈N
natural numbers, and z i = xi + yi − i. For the second result it is enough to take an
ordinal sum of the Krause t-norm on [0, a] for some a ∈ ]0, 1[. Then this ordinal
sum is border continuous, non-Archimedean and non-continuous.
In the final part of this chapter we will focus on uninorms that are continuous on
[0, e[2 ∪ ]e, 1]2 . Assume such a uninorm U and focus on TU (similar observations
can be obtained for SU by duality). We will now show that TU can be obtained
Structure of Uninorms with Continuous Diagonal Functions 131

from a continuous t-subnorm ([16]) by redefining its values on the border of the
unit square. First, let us recall that a binary operation M : [0, 1]2 −→ [0, 1] is a
t-subnorm if it is commutative, associative, non-decreasing in both variables and
M(x, y) ≤ min(x, y) for all (x, y) ∈ [0, 1]2 . From each t-subnorm we can obtain a
t-norm T given by T (x, y) = M(x, y) for all (x, y) ∈ [0, 1[2 , T (x, y) = min(x, y)
otherwise. This process is called lifting of a t-subnorm to a t-norm. Then T = M
if and only if M is a t-norm. Vice versa in [15] a border-continuous projection
MT : [0, 1]2 −→ [0, 1] of a t-norm was defined by

T (x, y) if (x, y) ∈ [0, 1[2 ,
MT (x, y) = − −
T (x , y ) if max(x, y) = 1.

The idea of this border-continuous projection was to obtain a reverse process to the
lifting of a t-subnorm to a t-norm. However, such a border-continuous projection
need not to be monotone.
Example 4 Let T : [0, 1]2 −→ [0, 1] be given by
⎧
⎪
⎨min(x, y) if max(x, y) = 1,
T (x, y) = 23 (x + y) − 5
if (x, y) ∈ [ 21 , 1]2 , max(x, y) < 1,
⎪
⎩
6
0 otherwise.

Then MT ( 21 , 1) = 0 and MT ( 21 , 78 ) = 1
12
.
Thus the proper definition of a border-continuous projection is rather MT (x, y) =
T (x, y) if (x, y) ∈ [0, 1[2 , MT (x, y) = T (x − , y) if x = 1, y < 1, MT (x, y) =
T (x, y − ) if y = 1, x < 1, and MT (x, y) = T (x − , y − ) if x = y = 1. In this case
a border-continuous projection of a t-norm is commutative, bounded by minimum
and non-decreasing in both variables. Note that a border-continuous projection is
not border-continuous in the sense that each point from the border of the unit square
is a point of continuity of MT , but in the sense that MT (x, 1− ) = MT (x, 1), while
the function MT (·, 1) can be non-continuous. However, if we take the t-norm from
Example 4 then MT is not associative, i.e., not a t-subnorm. Indeed, in this case

2
(x + y) − 5
if (x, y) ∈ [ 21 , 1]2 ,
MT (x, y) = 3 6
0 otherwise,

and MT (MT (1, 1), x) = 2·x3

− 21 for all x ≥ 43 , however, MT (1, MT (1, x)) = 0 for
all x < 1. Thus we have to characterize all t-norms such that their border-continuous
projection is a t-subnorm. This problem was solved in [39]. It is evident that if T is
border-continuous then MT = T . Further we have the following result.
Proposition 4 ([39]) For a t-norm T : [0, 1]2 −→ [0, 1] its border-continuous pro-
jection MT : [0, 1]2 −→ [0, 1] is a t-subnorm if and only if the following two condi-
tions are satisfied:
132 A. Mesiarová-Zemánková

(i) for all x, y ∈ [0, 1[ either T (u 0 , x) = lim − T (u, x) for some u 0 ∈ [0, 1[, or
u−→1
T (a, y) = lim − T (v, y), where a = lim − T (u, x),
v−→a u−→1
(ii) either lim − T (u, u) = 1, or T (u 0 , v0 ) = lim − T (u, u) for some u 0 , v0 ∈ [0, 1[,
u−→1 u−→1
or for all x ∈ [0, 1[ there is T (b, x) = lim − T (v, x), where b = lim − T (u, u).
v−→b u−→1

As an easy corollary of the previous result we see that if T is left-continuous on

[0, 1[2 then MT is a t-subnorm: this is for example the case of the Krause t-norm.
Since we focus on t-norms TU continuous on [0, 1[2 then MTU is always a continu-
ous t-subnorm and a similar result can be obtained also for SU . From [30] we know that
a continuous t-subnorm is an ordinal sum of continuous Archimedean t-norms and a
continuous Archimedean t-subnorm. The set of continuous Archimedean t-subnorms
can be divided into three parts: continuous cancellative t-subnorms, continuous nilpo-
tent t-subnorms, continuous t-subnorms with no nilpotent element which are not
cancellative. Here continuous nilpotent t-subnorms are such that they posses a nilpo-
tent element x ∈ ]0, 1] with M(x, x) = 0. Although the structure of uninorms such
that MTU (M SU ) is not cancellative is quite complicated, in the case that MTU (M SU )
is a continuous cancellative t-subnorm (t-superconorm) several results similar as in
the case of uninorms with strict underlying functions can be shown. Note that a
continuous cancellative t-subnorm (t-superconorm) is always Archimedean.

Proposition 5 ([39]) Let U : [0, 1]2 −→ [0, 1] be a uninorm with neutral ele-
ment e ∈ ]0, 1[, such that MTU (MCU ) is a continuous cancellative t-subnorm
(t-superconorm). Then U has one of the seven forms from Theorem 5.

We can use this result to characterize also non-Archimedean uninorms continuous

on [0, e[2 ∪]e, 1]2 , such that their underlying t-norm (t-conorm) is an ordinal sum
of continuous Archimedean t-norms (t-conorms) and the last (the first) cancellative
continuous t-subnorm (t-superconorm).
We will denote by Ulcc be the set of all uninorms continuous on [0, e[2 ∪ ]e, 1]2
such that there exists an idempotent point b0 ∈ [0, 1], b0 < e, such that U is contin-
uous and cancellative on ]b0 , e[2 , and by Ur cc be the set of all uninorms continuous
on [0, e[2 ∪ ]e, 1]2 such that there exists an idempotent point c0 ∈ [0, 1], c0 > e,
such that U is continuous and cancellative on ]e, c0 [2 . In [39] it was shown that if
U ∈ Ulcc ∩ Ur cc then U is an extended ordinal sum of a uninorm with continuous
underlying functions on [0, b0 [ ∪ {U (b0 , c0 )} ∪ ]c0 , 1] and an Archimedean uninorm
cancellative and continuous on ]0, e[2 ∪]e, 1[2 acting on [b0 , c0 ].
Further, if U ∈ Ulcc and e = inf{c ∈ ]e, 1] | c is an idempotent point} then U is
an extended ordinal sum of a uninorm with continuous underlying functions on
[0, b0 ]∪]e, 1] and an Archimedean t-norm cancellative and continuous on ]0, 1[2
acting on [b0 , e].
Structure of Uninorms with Continuous Diagonal Functions 133

Finally, if U ∈ Ur cc and e = sup{b ∈ [0, e[| b is an idempotent point} then U is

an extended ordinal sum of a uninorm with continuous underlying functions on
[0, e[ ∪ [c0 , 1] and an Archimedean t-conorm cancellative and continuous on ]0, 1[2
acting on [e, c0 ].

5 Conclusions and Further Perspectives

We have discussed uninorms with continuous underlying functions, shown several

partial results that were achieved, and then demonstrated their representation via the
extended ordinal sum construction. We have also characterized the set on which such
a uninorm is discontinuous. We have further shown how these results can be used
also for uninorms for which only parts of the underlying functions are known.
We have also discussed a generalization of this problem, i.e., the case when a
uninorm is continuous on [0, e[2 ∪ ]e, 1]2 . We have shown the characterization of
such uninorms in the case when the underlying functions are related to a continuous
cancellative t-subnorm and a continuous cancellative t-superconorm, respectively. An
open problem for future research is the full characterization of uninorms continuous
on [0, e[2 ∪ ]e, 1]2 .
We have also shown the characterization of all t-norms such that their border-
continuous projection is associative, i.e., a t-subnorm.
As we have seen in Sect. 2, recently a number of papers was focused on uni-
norms on complete lattices. Therefore future research on uninorms will be related to
uninorms on more abstract scales.

Acknowledgments This work was supported by grants VEGA 2/0049/14, APVV-0178-11 and
Program Fellowship of SAS.

References

1. Aczél, J.: Lectures on Functional Equations and their Applications. Academic Press, New York
(1966)
2. Akella, P.: Structure of n-uninorms. Fuzzy Sets Syst. 158, 1631–1651 (2007)
3. De Baets, B.: Idempotent uninorms. Eur. J. Oper. Res. 118, 631–642 (1998)
4. De Baets, B., Fodor, J., Ruiz, D., Torrens, J.: Idempotent uninorms on finite ordinal scales. Int.
J. Uncertain. Fuzziness, Knowl.-Based Syst. 107, 1–14 (2009)
5. Clifford, A.H.: Naturally totally ordered commutative semigroups. Am. J. Math. 76, 631–646
(1954)
6. Dombi, J.: A general class of fuzzy operators, the De Morgan class of fuzzy operators and
fuzziness measures induced by fuzzy operator. Fuzzy Sets Syst. 8, 149–163 (1982)
7. Drewniak, J., Drygaś, P.: On a class of uninorms. Int. J. Uncertain. Fuzziness, Knowl.-Based
Syst. 10, 5–10 (2002)
8. Drygaś, P.: Discussion of the structure of uninorms. Kybernetika 41, 213–226 (2005)
9. Drygaś, P.: On monotonic operations which are locally internal on some subset of their domain.
Proc. EUSFLAT Conf. 2, 185–191 (2007)
134 A. Mesiarová-Zemánková

10. Drygaś, P.: On properties of uninorms with underlying t-norm and t-conorm given as ordinal
sums. Fuzzy Sets Syst. 161, 149–157 (2010)
11. Fodor, J., De Baets, B.: A single-point characterization of representable uninorms. Fuzzy Sets
Syst. 202, 89–99 (2012)
12. Fodor, J., Yager, R.R., Rybalov, A.: Structure of uninorms. Int. J. Uncertain. Fuzziness, Knowl.-
Based Syst. 5, 411–427 (1997)
13. Hájek, P.: Observations on the monoidal t-norm logic. Fuzzy Sets Syst. 132, 107–112 (2002)
14. Hu, S., Li, Z.: The structure of continuous uninorms. Fuzzy Sets Syst. 124, 43–52 (2001)
15. Jayaram, B., Baczyński, M., Mesiar, R.: R-implications and the exchange principle: a complete
characterization. In: Galichet, S., Montero, J., Mauris, G. (eds.) Proceedings of EUSFLAT-2011
and LFA-2011, pp. 223–229. Aix-les-Bains, France (2011)
16. Jenei, S.: A note on the ordinal sum theorem and its consequence for the construction of
triangular norms. Fuzzy Sets Syst. 126, 199–205 (2002)
17. Kimberling, C.: On a class of associative functions. Publ. Math. Debrecen 20, 21–39 (1973)
18. Klement, E.P., Mesiar, R., Pap, E.: On the relationship of associative compensatory operators to
triangular norms and conorms. Int. J. Uncertainty, Fuzziness, Knowl.-Based Syst. 4, 129–144
(1996)
19. Klement, E.P., Mesiar, R., Pap, E.: Triangular Norms. Kluwer Academic Publishers, Dordrecht
(2000)
20. Krause, G.M.: The Devil’s Terraces: a Discontinuous Associative Function, personal commu-
nication (2015)
21. Li, G., Liu, H.W., Fodor, J.: Single-point characterization of uninorms with nilpotent underlying
t-norm and t-conorm. Int. J. Unc. Fuzz. Knowl. Based Syst. 22, 591–604 (2014)
22. Li, Y.M., Shi, Z.K.: Remarks on uninorm aggregation operators. Fuzzy Sets Syst. 114, 377–380
(2000)
23. Liu, H.W.: Semi-uninorms and implications on a complete lattice. Fuzzy Sets Syst. 191, 72–82
(2012)
24. Martín, J., Mayor, G., Torrens, J.: On locally internal monotonic operations. Fuzzy Sets Syst.
137, 27–42 (2003)
25. Mas, M., Mayor, G., Torrens, J.: t-operators and uninorms on a finite totally ordered set. Int. J.
Intell. Syst. 14, 909–922 (1999)
26. Mas, M., Monserrat, M., Torrens, J.: On left and right uninorms. Int. J. Uncertainty, Fuzziness
Knowl.-Based Syst. 9(4), 491–507 (2001)
27. Mesiar, R., Navara, M.: Diagonals of continuous triangular norms. Fuzzy Sets Syst. 104, 35–41
(1999)
28. Mesiar, R.: Triangular norms—an overview. In: Reusch, B., Temme, K.H. (eds.) Computational
Intelligence in Theory and Practice, pp. 35–54. Physica-Verlag, Heidelberg (2001)
29. Mesiarová, A.: Wild T-norms. J. Electr. Eng. 12/s, 36–40 (2000)
30. Mesiarová, A.: Continuous triangular subnorms. Fuzzy Sets Syst. 142, 75–83 (2004)
31. Mesiarová-Zemánková, A.: Multi-polar t-conorms and uninorms. Inf. Sci. 301, 227–240 (2015)
32. Mesiarová-Zemánková, A.: Ordinal sum of uninorms and generalized uninorms, Int. J. Approx-
imate Reasoning, under Rev. (2015)
33. Mesiarová-Zemánková, A.: Ordinal sums of representable uninorms, Fuzzy Sets Syst., under
Rev. (2015)
34. Mesiarová-Zemánková., A.: Characterization of uninorms with continuous underlying t-norm
and t-conorm by their set of discontinuity points. IEEE Trans. Fuzzy Syst., under Rev. (2015)
35. Mesiarová-Zemánková., A.: Characterization of uninorms with continuous underlying t-norm
and t-conorm by means of the ordinal sum construction. Int. J. Approximate Reasoning, under
Rev. (2015)
36. Mesiarová-Zemánková., A.: T-norms and t-conorms continuous around diagonals. Fuzzy Sets
Syst., under Rev. (2015)
37. Mesiarová-Zemánková A.: Continuous completions of triangular norms known on a subregion
of the unit interval. Fuzzy Sets Syst., under Rev. (2015)
Structure of Uninorms with Continuous Diagonal Functions 135

38. Mesiarová-Zemánková A.: Extremal completions of triangular norms known on a subregion

of the unit interval. In: Torra, V., Narukawa, Y. (eds.), Proceedings MDAI 2015 Conference,
LNAI 9321, pp. 21–32. Springer (2015)
39. Mesiarová-Zemánková A.: Uninorms continuous on [0, e[2 ∪]e, 1]2 . Inf. Sci., under Rev. (2015)
40. Petrík, M., Mesiar, R.: On the structure of special classes of uninorms. Fuzzy Sets Syst. 240,
22–38 (2014)
41. Ruiz-Aguilera, D., Torrens, J., De Baets, B., Fodor, J.: Some remarks on the characterization
of idempotent uninorms. In: Hüllermeier, E., Kruse, R., Hoffmann, F. (eds.) Computational
Intelligence for Knowledge-Based Systems Design, Proceedings of the 13th IPMU 2010 Con-
ference, LNAI 6178, pp. 425–434. Springer, Berlin (2010)
42. Ruiz, D., Torrens, J.: Distributivity and conditional distributivity of a uninorm and a continuous
t-conorm. IEEE Trans. Fuzzy Syst. 14(2), 180–190 (2006)
43. Schweizer, B., Sklar, A.: Probabilistic Metric Spaces. North-Holland, New York (1983)
44. Smutná, D.: Non-Continuous t-norms with continuous diagonal. J. Electr. Eng. 12/s, 51–53
(2000)
45. Smutná, D.: On a peculiar t-norm. BUSEFAL 75, 60–67 (1998)
46. Su, Y., Wang, Z., Tang, K.: Left and right semi-uninorms on a complete lattice. Kybernetika
49(6), 948–961 (2013)
47. Tkadlec, J.: Triangular norms with continuous diagonals. Tatra Mt. Math. Publ. 16, 187–195
(1999)
48. Wang, Z., Fang, J.X.: Residual operations of left and right uninorms on a complete lattice.
Fuzzy Sets Syst. 160, 22–31 (2009)
49. Yager, R.R., Rybalov, A.: Uninorm aggregation operators. Fuzzy Sets Syst. 80, 111–120 (1996)
50. Yager, R.R., Rybalov, A.: Bipolar aggregation using the uninorms. Fuzzy Optim. Decis. Making
10, 59–70 (2011)
The Notions of Overlap and Grouping
Functions

Humberto Bustince, Edurne Barrenechea, Miguel Pagola

and Javier Fernandez

Abstract In this work, we make a review of the concepts of overlap and grouping
functions, mainly from a theoretical point of view. In particular, we summarize some
of the most relevant works that have been published in recent years about this topic.

1 Introduction

In many situations it is necessary to assign a given element or object to one out of

several available classes. If a clear boundary among the classes does not exist, it may
be difficult to carry out such an assignation. Even more, the classes may be fuzzy in
nature, and experts may realize that elements are simply in between several classes
(see, e.g., [3]). But it may be also the case that even if the boundaries are clear, we
do not have enough precision about them [26]. In any of these situations the concept
of overlap arises (see [5–7, 14, 15, 32]).
The concept of overlap as a bivariate aggregation operator was first introduced in
[9] to measure the degree of overlap of an object in a fuzzy classification system with
two classes, with an eye specially kept in image processing applications [22]. How-
ever, overlap function have has been succesfully applied to many other situations,
when it is necessary to know the degree of overlap of objects in two-class classifi-
cation systems, as the image segmentation problem described in [22] (in which it is

H. Bustince (B) · E. Barrenechea · M. Pagola · J. Fernandez

Departamento of Automática y Computación and with the Institute of Smart Cities,
Universidad Publica de Navarra, 31006 Navarra, Spain
e-mail: [email protected]
E. Barrenechea
e-mail: [email protected]
M. Pagola
e-mail: [email protected]
J. Fernandez
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 137

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_8
138 H. Bustince et al.

necessary to discriminate between object and background) or in the framework of

preference relations [10].
Mathematically speaking, overlap functions are just a particular instance of bivari-
ate, continuous aggregation functions [11]; that is, increasing functions which are
defined in the unit square and which fulfill appropriate boundary conditions. It is
therefore worth to consider the possible relationships between overlap functions and
other well-known examples of aggregation functions, as t-norms, copulas, semicop-
ulas or quasi-copulas. Notice that although we, in principle, consider that overlap
functions are defined in the unit square, this is not an essential requirement, and other
domains may be considered too.
Overlap functions can also be extended to situations where more than two classes
are involved, but associativity is not a natural requirement [18]. This is the case, for
instance, for classification problems which involve rules of the type:

Rule R j : If x1 is 
A j1 and . . . and xn is 
A jn
(1)
then Class = C j with RW j ,

then in the inference procedure (see [19–21, 28]) an aggregation

An (μ 
A j1 (x p1 ), . . . , μ 
A jn (x pn ))

is commonly used, where x p = (x p1 , . . . , x pn ) is a new example to be classified. In

fact, in this kind of procedures, the product t-norm, is usually considered in order
to carry out such aggregation. However, in some of these situations associativity is
not a natural assumption. This fact leads to the extension of the concept of overlap
function to the n-dimensional setting.
Besides, and closely related to the notion of overlap function, it also appears the
concept of a grouping function. In this case, instead of being interested in determining
whether an input falls into more than one class, we focus on determining up to what
extent it belongs to at least one of the considered classes. The main interest of this
notion of grouping function lies in the fact that the class of overlap functions may be
obtained by duality with respect to any strong negation from the class of grouping
functions, and vice-versa. This implies, as a relevant consequence, that overlap and
grouping functions can be used to replace t-norms and t-conorms in those situations
where associativity is not a natural requirement. This is the case, for instance and
under appropriate circumstances, of preferences structures, see [10].
In this chapter we present an overview of the main concepts and results related to
overlap and grouping, following the papers [4, 9, 10, 13, 18, 22, 25]. The structure
of the chapter is as follows. We start recalling some preliminary notions and results.
In Sect. 3 we introduce the notion of a bivariate overlap functions. In Sect. 4 we
discuss additive generators for overlap functions. In Sect. 5 we recall the concept of
grouping functions and its relation with overlap functions. Section 6 is devoted to
the n-dimensional extension of overlap and grouping functions. We finish with some
conclusions and references.
The Notions of Overlap and Grouping Functions 139

2 Preliminaries

We start recalling some concepts and results which are well-known, in order to fix
notations.We start with the concept of automorphism of the unit interval.

Definition 1 An automorphism of the unit interval is any continuous and strictly

increasing function ϕ : [0, 1] → [0, 1] such that ϕ(0) = 0 and ϕ(1) = 1.

The definition of aggregation functions is a very well-known one. Here we follow

the approach and definitions given in [8, 11, 17, 24] (but see also [2, 12]).

Definition 2 An aggregation function of dimension n (n-ary aggregation func-

tion) is an increasing mapping M : [0, 1]n → [0, 1] such that M(0, . . . , 0) = 0 and
M(1, . . . , 1) = 1.

For the particular case of bivariate aggregation functions we remind the following
definitions.

Definition 3 Let M be a bivariate aggregation function.

(i) M is said to be symmetric if M(x, y) = M(y, x) for any x, y ∈ [0, 1].
(ii) M is said to be associative if M(M(x, y), z) = M(x, M(y, z)) for any x, y, z ∈
[0, 1].

One of the most relevant example of aggregation function is provided by t-norms.

These function play a key role, for instance, to model conjunctions in fuzzy logics
or intersections in fuzzy set theory.

Definition 4 A triangular norm (t-norm for short) is an associative, symmetric

bivariate aggregation function T : [0, 1]2 → [0, 1] such that T (1, x) = x for all
x ∈ [0, 1]. A strictly increasing (in ]0, 1]2 ) continuous t-norm T is called a strict
t-norm.

A particular type of continuous t-norms are Archimedean t-norms (see, e.g., [23]).

Definition 5 A continuous t-norm T is said to be Archimedean if T (x, x) < x for

all x ∈]0, 1[.
It is worth to note that, although this is not the usual definition of Archimedean
t-norm that can be found in the literature, both are equivalent when dealing with
continuous t-norms [23]. It is also important to recall that every strict t-norm (i.e.,
any continuous and strictly increasing t-norm) is necessarily Archimedean. In fact,
any strict t-norm can be obtained perturbing the product t-norm TP (x, y) = x y by
means of an appropriate automorphism (see [23, 29]).
140 H. Bustince et al.

Theorem 1 A t-norm T is strict if and only if there exists an automorphism ϕ of the

unit interval such that

T (x, y) = ϕ −1 (ϕ(x)ϕ(y)), x, y ∈ [0, 1].

It is possible to classify t-norms using the notion of ordinal sum, that we recall now.

Definition 6 Suppose that {[am , bm ]} is a countable family of non-overlapping,

closed, non-trivial, proper subintervals of [0, 1], To each [am , bm ] in the family asso-
ciate a t-norm Tm . The ordinal sum of the family {([am , bm ], Tm )} is the mapping
T : [0, 1]2 → [0, 1] given by

am + (bm − am )Tm ( bx−a m
m −am
, by−a m
m −am
) if (x, y) ∈ [am , bm ]2
T (x, y) =
min(x, y) otherwise

Each Tm is called a summand.

By means of these ordinal sums, t-norms can be classified as follows [16].

Theorem 2 Assume that T is a continuous t-norm. Then, one of the following three
cases is valid for T :
1. T (x, y) = min(x, y);
2. T is Archimedean;
3. there exists a family {([am , bm ], Tm )} such that T is the ordinal sum of this family
and each Tm is a continuous Archimedean t-norm.

A general study on t-norms and their properties may be found, for instance, in
[1, 15, 16, 23, 29].

2.1 Copulas, Semicopulas and Quasi-copulas

Definition 7 A mapping S : [0, 1]2 → [0, 1] is called a semicopula if it is nonde-

creasing in each coordinate and 1 is its neutral element, i.e., S(x, 1) = S(1, x) = x
for all x ∈ [0, 1].

Definition 8 A quasi-copula is a semicopula Q which is also a 1-Lipschitz function

(with respect to the L 1 -norm).

Definition 9 A copula is a semicopula C which is 2-increasing, i.e.,

C(x, y) + C(x  , y  ) − C(x  , y) − C(x, y  ) ≥ 0 for all 0 ≤ x ≤ x  ≤ 1, 0 ≤ y ≤ y  ≤ 1.

Observe that, as stated previously, each copula is a quasi-copula. More generally,

all copulas, semicopulas and quasi-copulas are aggregation functions.
The Notions of Overlap and Grouping Functions 141

3 Overlap Functions

In this section, we mainly follows the developments in [9]. As we have already said,
overlap functions are a particular instance of bivariate aggregation functions. The
formal definition reads as follows.

Definition 10 [9] A mapping G O : [0, 1]2 → [0, 1] is an overlap function if it sat-

isfies the following conditions:
(G O 1). G O is symmetric.
(G O 2). G O (x, y) = 0 if and only if x y = 0.
(G O 3). G O (x, y) = 1 if and only if x y = 1.
(G O 4). G O is non-decreasing.
(G O 5). G O is continuous.

There are many possible examples of overlap functions. For instance, G O (x, y) =
min(x, y) or G O (x, y) = x p y p for p > 0. Note that, in the latter example the result-
ing overlap function is not associative unless p = 1, whereas, if p = 1, we recover
the product, which is also a t-norm (as it is also the case of the minimum).
Let’s denote by O the set of all overlap functions. Then:

Theorem 3 (O, ≤O ) with the ordering ≤O defined for G 1 , G 2 ∈ O by

G 1 ≤O G 2 if and only if G 1 (x, y) ≤ G 2 (x, y)

for all x, y ∈ [0, 1], is a lattice.

The lattice (O, ≤O ) is not complete (no top neither bottom elements, for example).
On the other hand, it is closed with respect to appropriate aggregation functions. In
particular, it holds that:

Theorem 4 Let M : [0, 1] × [0, 1] → [0, 1] be a mapping. For G 1 , G 2 ∈ O, define

the mapping M (G 1 , G 2 ) : [0, 1] × [0, 1] → [0, 1] as

M (G 1 , G 2 )(x, y) = M(G 1 (x, y), G 2 (x, y)) for all x, y ∈ [0, 1] .

Then, M (G 1 , G 2 ) ∈ O for any G 1 , G 2 ∈ O if and only if there is a continuous

aggregation function M ∗ : [0, 1] × [0, 1] → [0, 1] with no zero divisors and such
that also its dual (M ∗ )d (that is, the mapping (M ∗ )d (x, y) = 1 − M ∗ (1 − x, 1 − y))
has no zero divisors (i.e., if M ∗ (x, y) = 1 then necessarily either x = 1 or y = 1)
so that M| E = M ∗ | E , where E =]0, 1[2 ∪{(0, 0), (1, 1)}.

An important consequence of the previous result is the convexity of the class of

overlap functions.
142 H. Bustince et al.

Corollary 1 Let G 1 , . . . , G m be overlap functions and

w1 , . . . , wm be non nega-
tive weights with wi = 1. Then the convex sum G = wi G i is also an overlap
function.

Overlap functions can be characterized in terms of rational expressions as follows.

Theorem 5 The mapping G O : [0, 1]2 → [0, 1] is an overlap function if and only if

f (x, y)
G O (x, y) =
f (x, y) + h(x, y)

for some f, h : [0, 1]2 → [0, 1] such that

1. f and h are symmetric;
2. f is non decreasing and h is non increasing;
3. f (x, y) = 0 if and only if x y = 0;
4. h(x, y) = 0 if and only if x y = 1;
5. f and h are continuous functions.
√
Example 1 Take f (x, y) = x y and h(x, y) = max(1 − x, 1 − y), then we have
that by the construction given in Theorem 5 we get an overlap function
√
xy
G O (x, y) = √ .
x y + max(1 − x, 1 − y)

Corollary 2 In the setting of Theorem 5, G O (x, x) = x for some x ∈ (0, 1) if and

only if
x
f (x, x) = h(x, x) .
1−x

The rational expressions which are provided in Theorem 5 need not be unique.
However, the lack of uniqueness allows us to recover a whole family of overlap
functions as follows.

Corollary 3 Let f and h be two functions in the setting of the previous theorem.
Then, for k1 , k2 ∈]0, ∞[, the mappings

f k1 (x, y)
G kO1 ,k2 (x, y) =
f k1 (x,y) + h k2 (x, y)

define a parametric family of overlap functions.

Corollary 4 In the same setting of Theorem 5, let us assume that G O can be

expressed in two different ways:

f 1 (x, y) f 2 (x, y)
G O (x, y) = =
f 1 (x, y) + h 1 (x, y) f 2 (x, y) + h 2 (x, y)
The Notions of Overlap and Grouping Functions 143

for any x, y ∈ [0, 1] and let M be a bivariate continuous aggregation function that
is homogeneous of order one. Then, if we define f (x, y) = M( f 1 (x, y), f 2 (x, y))
and h(x, y) = M(h 1 (x, y), h 2 (x, y)) it also holds that
f (x, y)
G O (x, y) = .
f (x, y) + h(x, y)

3.1 Overlap Functions and t-Norms

Overlap functions do not require associativity in their definition. If the associativity

property is required, we recover a t-norm.
Theorem 6 Let G O be and associative overlap function. Then G O is a t-norm.
Besides, Theorem 2 of classification of t-norms can be extended to cover those
t-norms which are also overlap functions as follows.
Theorem 7 If a t-norm T is an overlap function, then T belongs to one of the
following three types:
(1) T = TM ;
(2) T is strict;
(3) T is the ordinal sum of the family {([am , bm ], Tm )}, with all the Tm continuous
Archimedean and such that if for some m 0 am 0 = 0, then necessarily Tm 0 is a strict
t-norm.
Example 2 1. In the construction of the following overlap function we use item
(3) of Theorem 7 taking as t-norm the product, (which is strict, continuous and
Archimedean), for the corresponding interval [0, 0.5].

2x y if (x, y) ∈ [0, 0.5]2
G O (x, y) =
min(x, y) otherwise

2. In the construction of the following overlap function we take the product and the
Lukasiewicz t-norms (see p. 84 of [23]). Nevertheless, in this overlap function
we do not consider any interval of the type [0, bm ].
⎧
⎨ 0.1 + 2.5(x − 0.1)(y − 0.1) if (x, y) ∈ [0.1, 0.5]2
G O (x, y) = 0.7 + max(x + y − 1.6, 0) if (x, y) ∈ [0.7, 0.9]2
⎩
min(x, y) otherwise

3. The following t-norm satisfies all the properties required to overlap func-
tions, except (G O 2). This is due to the fact that in [0, 0.25]2 we consider the
Lukasiewicz t-norm which is continuous and Archimedean but not strict.

max(x + y − 0.25, 0) if (x, y) ∈ [0, 0.25]2
T (x, y) =
min(x, y) otherwise
144 H. Bustince et al.

3.2 Overlap Functions and Semicopulas, Quasi-copulas

and Copulas

As a first result we have the following.

Proposition 1 Let S be a symmetric semicopula. Then S is an overlap function if
and only if S is continuous and has not zero divisors.
Corollary 5 Let Q be a symmetric quasicopula without zero divisors. Then Q is
also an overlap function.
It is also possible to recover copulas from overlap functions.
Theorem 8 Let G O be an overlap function being homogeneous of order k + 1, with
k ∈ [0, 1]. Suppose that there exists e ∈ [0, 1] such that G O (x, e) = G O (e, x) = x
for all x ∈ [0, 1], that is, that G O has a neutral element e. Then G O is also a copula.
Notice that the family (min(x k y, x y k ) for k ∈ [0, 1] is the so called Cuadras-Augé
family of copulas [27].

4 Additive Generators of Overlap Functions

As we have already discussed, there exists a close relation between overlap functions
and other relevant classes of operators such as t-norms. Moreover, in the same way
as it is done for the latter, it is possible to study the use additive generators for overlap
functions [13].
We start recalling the notion of pseudo-inverse of a given function.
Definition 11 [31] Let f : [a, b] → [c, d] be an increasing or decreasing function.1
function f (−1) : [c, d] → [a, b] defined by
⎧
⎨ sup{x ∈ [a, b] | f (x) < y} if f (a) < f (b),
f (−1) (y) = sup{x ∈ [a, b] | f (x) > y} if f (a) > f (b), (2)
⎩
a if f (a) = f (b)

is called the pseudo-inverse of f .

Let’s denote the range or image of a function f : A → B by Ran( f ). Note that, if
a function f : [a, b] → [c, d] is increasing (decreasing) then f (−1) is also increasing
(decreasing). If f is strictly increasing (decreasing) then f (−1) is continuous, f (−1) ◦
f = I d[a,b] and f ◦ f (−1) (x) = x if and only if x ∈ Ran( f ).
Our approach here to additive generators for overlap functions follows is based
on the work by Viceník [30, 31]. We start with some auxiliary lemmas.

1 Inthis paper, an increasing (decreasing) function does not need to be strictly increasing (decreas-
ing).
The Notions of Overlap and Grouping Functions 145

Lemma 1 Let θ : [0, 1] → [0, ∞] be a decreasing function such that

1. θ (x) + θ (y) ∈ Ran(θ ), for x, y ∈ [0, 1] and
2. if θ (x) = θ (0) then x = 0.
Then θ (x) + θ (y) ≥ θ (0) if and only if x = 0 or y = 0.

Lemma 2 Consider functions θ : [0, 1] → [0, ∞] and ϑ : [0, ∞] → [0, 1] such

that, for each x0 ∈ [0, 1], if it holds that

ϑ(θ (x)) = x0 if and only if x = x0 , (3)

then θ (x) = θ (x0 ) if and only if x = x0 .

Then we arrive at the main result regarding additive generators of overlap

functions.

Theorem 9 Let θ : [0, 1] → [0, ∞] and ϑ : [0, ∞] → [0, 1] be continuous and

decreasing functions such that
1. θ (x) + θ (y) ∈ Ran(θ ), for x, y ∈ [0, 1] ;
2. ϑ(θ (x)) = 0 if and only x = 0;
3. ϑ(θ (x)) = 1 if and only x = 1;
4. θ (x) + θ (y) = θ (1) if and only x = 1 and y = 1.
Then, the function Oθ,ϑ : [0, 1]2 → [0, 1], defined by

Oθ,ϑ (x, y) = ϑ(θ (x) + θ (y)), (4)

is an overlap function.

Corollary 6 Let θ : [0, 1] → [0, ∞] and ϑ : [0, ∞] → [0, 1] be continuous and

decreasing functions such that
1. θ (x) = ∞ if and only if x = 0;
2. θ (x) = 0 if and only if x = 1;
3. ϑ(x) = 1 if and only if x = 0;
4. ϑ(x) = 0 if and only if x = ∞.
Then, the function Oθ,ϑ : [0, 1]2 → [0, 1], defined by

Oθ,ϑ (x, y) = ϑ(θ (x) + θ (y)), (5)

is an overlap function.

Proposition 2 Let θ : [0, 1] → [0, ∞] and ϑ : [0, ∞] → [0, 1] be continuous and

decreasing functions such that
146 H. Bustince et al.

1. ϑ(x) = 1 if and only if x = 0;

2. ϑ(x) = 0 if and only if x = ∞;
3. 0 ∈ Ran(θ );
4. Oθ,ϑ (x, y) = ϑ(θ (x) + θ (y)) is an overlap function.
Then, the following conditions also hold:
5. θ (x) = ∞ if and only if x = 0;
6. θ (x) = 0 if and only if x = 1;

(θ, ϑ) is called an additive generator pair of the overlap function Oθ,ϑ , and Oθ,ϑ
is said to be additively generated by the pair (θ, ϑ).

Example 3 Consider the functions θ : [0, 1] → [0, ∞] and ϑ : [0, ∞] → [0, 1],
defined, respectively by:

−2 ln x if x = 0
θ (x) =
∞ if x = 0

and 
e−x if x = ∞
ϑ(x) =
0 if x = ∞,

which are continuous and decreasing functions, satisfying the conditions 1–4 of
Corollary 6. Then, whenever x = 0 and y = 0, one has that:

Oθ,ϑ (x, y) = ϑ(θ (x) + θ (y)) = e−(−2 ln x−2 ln y) = eln x

2 2
y
= x 2 y2.

Otherwise, if x = 0, it holds that

Oθ,ϑ (0, y) = ϑ(θ (0) + θ (y)) = ϑ(∞ + θ (y)) = 0,

and, similarly, if y = 0, then Oθ,ϑ (x, 0) = 0. It follows that

Oθ,ϑ (x, y) = x 2 y 2 ,

and so we recover ae non associative overlap function for which 1 is not a neutral
element.

Corollary 7 Considering the same conditions as in Theorem 9, whenever ϑ = θ (−1)

then Oθ,ϑ is a positive t-norm (i.e., without divisors of zero).

Theorem 10 Let G O : [0, 1]2 → [0, 1] be an overlap function having 1 as neutral

element. Then, if G O is additively generated by a pair (θ, ϑ), with θ : [0, 1] →
[0, ∞] and ϑ : [0, ∞] → [0, 1] satisfying the conditions of Theorem 9, then G O is
associative.

The following result is straight:

The Notions of Overlap and Grouping Functions 147

Corollary 8 Let G O : [0, 1]2 → [0, 1] be an overlap function additively generated

by a pair (θ, ϑ). G O is a t-norm if and only if 1 is a neutral element of G O .

Notice that whenever T is a positive continuous t-norm (that is, an overlap func-
tion) that is additively generated by a function t : [0, 1] → [0, ∞], then it is also
additively generated by a pair (θ, ϑ) in the sense of Theorem 9, where θ = t and
ϑ = t (−1) , and vice-versa.
A more detailed study on additive generators for overlap functions can be
found in [13].
Now let’s recall the definition of pseudo-automorphism.

Definition 12 A function F : [0, 1] → [0, 1] is said to be a pseudo-automorphism

if the following conditions hold:
(PA1) F is increasing;
(PA2) F is continuous;
(PA3) F (x) = 1 if and only if x = 1;
(PA3) F (x) = 0 if and only if x = 0.
An automorphism ϕ : [0, 1] → [0, 1] is a strictly increasing pseudo-automorphism.

Then we have the following important result

Theorem 11 Let ϕ : [0, 1] → [0, 1] be an automorphism and T : [0, 1]2 → [0, 1]

be a t-norm. The function Oϕ,T : [0, 1]2 → [0, 1], defined by

Oϕ,T (x, y) = ϕ(T (x, y)), (6)

is an overlap function if and only if T is positive and continuous.

Oϕ,T is called the overlap obtained by the distortion of the t-norm T by the
automorphism ϕ, or the overlap function obtained by a (ϕ, T )-distortion.
However, it is important to remark that not every overlap is obtained as a distortion
of a t-norm.

Proposition 3 There exists an overlap function G O : [0, 1]2 → [0, 1] which is not
a (F , T )-distortion of a positive and continuous t-norm T : [0, 1]2 → [0, 1] for
any automorphism F .

In the same way, it is also true that not every overlap function G O : [0, 1]2 →
[0, 1] can be obtained by means of a (F , T )-distortion, for some pseudo-automor-
phism F : [0, 1] → [0, 1] and some positive and continuous t-norm T : [0, 1]2 →
[0, 1].
148 H. Bustince et al.

5 Overlap Functions and Grouping Functions

The notion of grouping function [10] arises in the applied field when it is necessary
to measure up to what extent a given element belongs to at least one of two given
classes. The formal definition reads as follows.

Definition 13 A mapping G G : [0, 1]2 → [0, 1] is a grouping function if it satisfies

the following conditions:
(G G 1) G G is symmetric;
(G G 2) G G (x, y) = 0 if and only if x = y = 0;
(G G 3) G G (x, y) = 1 if and only if x = 1 or y = 1;
(G G 4) G G is non-decreasing;
(G G 5) G G is continuous.

There exists a close link between overlap and grouping functions. In fact, both
concepts are dual to each other.

Definition 14 [22] Let G be an overlap function (resp. a grouping function) and let
n 1 and n 2 be two continuous negation operators such that:
1. n 1 (x) = 0 (resp. n 2 (x) = 0) if and only if x = 1, and
2. n 1 (x) = 1 (resp. n 2 (x) = 1) if and only if x = 0
Then, the operator
n 1 ,n 2
G (x, y) = n 1 (G(n 2 (x), n 2 (y)))

is called the dual grouping (resp. overlap) of G with respect to n 1 and n 2 .

In fact, every grouping function can be obtained as the dual through a strong
negation of an overlap function. Besides, note that this duality implies that many of
the results we have discussed for overlap functions can be translated straightforwardly
for grouping functions just making use of duality. So, in some sense, it seems that
overlap and grouping functions are related in the same way as t-norms are t-conorms
are. However, the drop of associativity in the definition of both overlap and grouping
functions gives raise to dramatic differences from the case of t-norms and t-conorms.
In particular, it does not hold that overlap functions are always smaller than or equal
to grouping functions.
√ √
Example 4 Consider the overlap function given by G O (x, y) = min( x, y) and
the grouping function given by G G (x, y) = max(x 2 , y 2 ). Then:
√ √
G O (0.5, 0.25) = min( 0.5, 0.25) > max((0.5)2 , (0.25)2 ) = G G (0.5, 0.25)

So in this case G O  G G .
The Notions of Overlap and Grouping Functions 149

The problem of when an overlap function is smaller than or equal to a grouping

function is very relevant from the point of view of applications and was deeply
analyzed in [25], and we now summarize the results from that paper.
The first important concept at this regard is that of f-bounding.

Definition 15 Let f be a mapping from [0, 1] to [0, 1]. An overlap function G O

(resp. a grouping function G G ) is called f -bound if the equality G O (x, 1) = f (x)
(resp. G G (x, 0) = f (x)) holds for all x ∈ [0, 1]

We denote by O f (G f ) the class of f -bound overlap functions (grouping func-

tions). Note that the class of overlap functions (resp.grouping functions) can be
obtained as the union of these classes O f (resp. G f ).
Then the following result is straight.

Proposition 4 Let G be either an f -bound overlap or an f -bound grouping. Then:

1. f is continuous,
2. f is increasing,
3. f (x) = 0 if and only if x = 0,
4. f (x) = 1 if and only if x = 1.

Let Ω denote the set of mappings satisfying the four properties given in
Proposition 4. Then

Proposition 5 Let G O and G G be an overlap and a grouping function, respectively,

and f ∈ Ω. Then:
1. if G O ∈ O f , then G O (x, y) ≤ min{ f (x), f (y)} for all x, y ∈ [0, 1].
2. if G G ∈ G f , then G G (x, y) ≥ max{ f (x), f (y)}) for all x, y ∈ [0, 1].

The relevance of the notion of f -bounding is clear in the following result.

Proposition 6 Let G O and G G be an f 1 -overlap function and an f 2 -grouping func-

tion, respectively, with f 1 , f 2 ∈ Ω. If f 1 ≤ f 2 , then G O (x, y) ≤ G G (x, y) for all
x, y ∈ [0, 1].

Corollary 9 Let G O be an overlap function and G G a grouping function. If both,

G O and G G , are f-bound then G O ≤ G G .

However, Proposition 6 provides a necessary but not sufficient condition to ensure

that a given overlap function is less than or equal to a given grouping function.
Nevertheless, the following result holds.

Proposition 7 Let G O be an overlap function (resp. let G G be a grouping function),

then for each mapping f ∈ Ω there exists G G ∈ G f (resp. G O ∈ O f ) such that
G O ≤ GG.
150 H. Bustince et al.

Besides, the choice of a bounding function imposes some restrictions to the size
of the corresponding overlap and grouping functions.
Proposition 8 Let f ∈ Ω, then
1. the operator G O (x, y) = f (min(x, y)) is the greatest f -bound overlap and
2. the operator G G (x, y) = f (max(x, y)) is the least f -bound grouping.
It is relevant to note that there is neither least element in O f nor greatest ele-
ment in G f .
Proposition 9 Let f ∈ Ω, then the infimum of O f is:
⎧
 ⎨ f (x) if y = 1
G O (x, y) = f (y) if x = 1
⎩
G O ∈O f 0 Otherwise

and the supremum of G f is:

⎧
⎨ f (x) if y = 0
G G (x, y) = f (y) if x = 0
⎩
G G ∈G f 1 Otherwise

Corollary 10 Let f ∈ Ω. Then the set O f (resp. G f ) does not have a least element
(resp. a greatest element).
Another interesting property of O f and G f is that they are dense as posets.
Proposition 10 Let G 1 and G 2 be two overlaps in O f (resp. two groupings in
G f ) such that G 1 < G 2 . Then there exists G ∈ O f (resp. G ∈ G f ) such that G 1 <
G < G2.
The following definition establishes a restriction on negations used in the dual
construction in order to maintain the f -bound condition.
Let’s recall now the notion of f -duality.
Definition 16 Let G be an overlap (resp. grouping), f ∈ Ω an automorphism on
the unit interval (i.e., a bijective mapping), n a bijective negation2 and let us consider
the negations:
• n 1 (x) = f (n(x)) and
• n 2 (x) = f −1 (n −1 (x)),
Then the operator
n, f

G (x, y) = n 1 G(n 2 (x), n 2 (y))

is called the f -dual grouping (resp. f -dual overlap) of G with respect to n.

2 This kind of negation is called in the Literature strict.

The Notions of Overlap and Grouping Functions 151

Proposition 11 Let f ∈ Ω, G ∈ O f (resp. G ∈ G f ) be a bijection and n a bijective

negation, then the f -dual grouping function (resp. overlap function) of G w.r.t. n is
f -bound.

From Proposition 11, we can ensure that any f-bound overlap functions is less
than or equal to any of its f-dual grouping functions.
n, f
Corollary 11 Let f ∈ Ω, G O ∈ O f and let G O be an f -dual grouping function
n, f
of G O . Then the inequality G O ≤ G O holds.

And reciprocally:
n, f
Corollary 12 Let f ∈ Ω, G G ∈ G f and let G G be an f -dual overlap function
n, f
of G G . Then the inequality G G ≤ G G holds.

In [25] this study is made deeper by considering also the so-called f -diagonal
overlap and grouping functions, that is, overlap and grouping functions whose values
at the diagonal are given by a fixed function f .

6 n-Dimensional Overlap and Grouping Functions

Since overlap functions are not assumed to be associative, its extensions to the n-
dimensional case with n > 2 requires of a specific definition. From now on, we
follow the developments in [18].

Definition 17 An n-dimensional aggregation function G O : [0, 1]n −→ [0, 1] is an

n-dimensional overlap function if and only if:
1. G O is symmetric.

n
2. G O (x1 , . . . , xn ) = 0 if and only if xi = 0.
i=1
3. G O (x1 , . . . , xn ) = 1 if and only if xi = 1 for all i ∈ {1, . . . , n}.
4. G O is increasing.
5. G O is continuous.

Note that, taking into account this definition, an object c that belongs to three
classes C1 , C2 and C3 with degrees x1 = 1, x2 = 1 and x3 = 0.3 will not have the
maximum degree of overlap since condition (3) of the previous definition is not
satisfied. Even more, if the degrees are x1 = 1, x2 = 1 and x3 = 0, from the second
condition we will conclude that the n-dimensional degree of overlapping of this
object into the classification system given by the classes C1 , C2 and C3 will be zero.
This is the reason why this first extension of the original idea of overlap proposed
in [9] has been called n-dimensional overlap. Let us observe that this definition is
closely related with the idea of intersection of n classes.
152 H. Bustince et al.

Example 5 It is easy to see that the following aggregation functions are n-dimensio-
nal overlap functions:
p
p
1. The minimum powered by p. G O (x1 , . . . , xn ) = min {xi } = min {xi } with
1≤i≤n 1≤i≤n
p > 0.

n
1
2. The geometric mean. G O (x1 , . . . , xn ) = ( xi ) n .
i=1 n
xi
3. The Einstein product aggregation operator. E P(x1 , . . . , xn ) = ni=1
1 + i=1 (1 − xi )
π 
n
4. The sinus induced overlap G O (x1 , . . . , xn ) = sin ( xi ) p with p > 0.
2 i=1

The characterization results already introduced for bivariate overlap functions

may be extended in a straight way for n-dimensional overlap functions. In particular,
we remark the following result.

Proposition 12 Let An : [0, 1]n −→ [0, 1] be an aggregation function. If An is aver-

aging, then An is an n-dimensional overlap function if and only if it is symmetric,
continuous, has zero as absorbing element and satisfies An (x, 1, . . . , 1) = 1 for any
x = 1.

In fact, We have the following theorem.

Theorem 12 Let G 1 , . . . , G m be n-dimensional overlap functions and let M :

[0, 1]m −→ [0, 1] be a continuous and symmetric aggregation function such that
if M(x) = 0 then xi = 0 for some i and M(x) = 1 only if xi = 1 for some i. Then
the aggregation function G : [0, 1]n −→ [0, 1] defined as G(x) = M(G 1 (x), . . . ,
G m (x)) is an n-dimensional overlap function.

Remark 1 Notice that, since any averaging aggregation function M being symmetric
and continuous satisfies the conditions of the previous Theorem, it is possible to
conclude that any continuous symmetric averaging aggregation of n-dimensional
overlap functions is also an n-dimensional overlap function.

As an illustrative case, consider that of OWA operators.

Proposition 13 Let W = (w1 , . . . , wn ) ∈ [0, 1]n be a weighting vector. The follow-
ing statements are equivalent:
1. The OWA operator defined by the weighting vector W is an n-dimensional overlap
function.
2. wn = 1
Also for the case of Kolmogorov-Nagumo means the following result holds.
The Notions of Overlap and Grouping Functions 153

Proposition 14 Let f : [0, 1] → [−∞, 0] be a continuous increasing bijection, that

is, f :]0, 1] →] − ∞, 0] is an increasing bijection such that lim x→0 f (x) = −∞,
so by abuse of notation we define f (0) = −∞. Then the function: G O (x1 , . . . , xn ) =
f −1 ( f (x1 ) + · · · + f (xn )) is an n-dimensional overlap function.

This corresponds to the n-ary extension of a strict t-norm generated by an additive

generator f . Note than an analogous extension may be done for grouping functions.

Definition 18 An n-dimensional function

G G : [0, 1]n −→ [0, 1]

is an n-dimensional grouping function if and only if it satisfies the following

conditions:
1. G G is symmetric.
2. G G (x) = 0 if and only if xi = 0, for all i = 1, . . . , n.
3. G G (x) = 1 if and only if there exist i ∈ {1, . . . , n} with xi = 1.
4. G G is non-decreasing.
5. G G is continuous.

Again, some particular continuous t-conorms (their n-ary forms) and their convex
combinations are prototypical examples of n-ary grouping functions.

Example 6 The following aggregation functions are examples of n-dimensional

grouping functions:
p
• The maximum powered by p. G G (x1 , . . . , xn ) = max {xi } with p > 0.
1≤i≤n
n
i=1 x i
• The Einstein sum aggregation operator. E S(x1 , . . . , xn ) =  n
1 + i=1 xi
Theorem 13 Let G O be an n-dimensional overlap function and let n 1 and n 2 be
two continuous negation operators such that:
1. n 1 (x) = 0 (resp. n 2 (x) = 0) if and only if x = 1, and
2. n 1 (x) = 1 (resp. n 2 (x) = 1) if and only if x = 0
Then the function G : [0, 1]n −→ [0, 1] defined as

G(x1 , . . . , xn ) = n 1 (G O (n 2 (x1 ), . . . , n 2 (xn ))

is an n-dimensional grouping aggregation function.

On the other hand, it is possible to build an n-dimensional overlap function from

a grouping aggregation function and a negation function using duality, as in the case
of bivariate overlap and grouping functions.
154 H. Bustince et al.

Theorem 14 Let G G be an n-dimensional grouping function and let n 1 and n 2 be

two continuous negation operators such that:
1. n 1 (x) = 0 (resp. n 2 (x) = 0) if and only if x = 1, and
2. n 1 (x) = 1 (resp. n 2 (x) = 1) if and only if x = 0.
Then the function G : [0, 1]n −→ [0, 1] defined as

G(x1 , . . . , xn ) = n 1 (G G (n 2 (x1 ), . . . , n 2 (xn )))

is an n-dimensional overlap aggregation function.

Theorem 15 Let G 1 , . . . G m be n-dimensional grouping functions and let M :

[0, 1]m −→ [0, 1] be a continuous aggregation function such that M(x) = 0 if and
only if xi = 0 for some i and M(x) = 1 if and only if xi = 1 for all i. Then the aggre-
gation function G : [0, 1]n −→ [0, 1] defined as G(x) = M(G 1 (x), . . . , G m (x)) is
an n-dimensional grouping function.

Corollary 13 Let G 1 , . . . , G m be n-dimensional grouping functions and let w1 , . . . ,

m  m
wm be nonnegative weights with wi = 1. Then the convex sum G(x) = wi
i=1 i=1
G i (x) is also an n-dimensional grouping function.

A more detailed analysis of n-dimensional overlap and grouping functions can be

found in [18].

7 Conclusions

In this chapter we have made a review of the main definitions and results related
to overlap and grouping functions. These functions have shown themselves very
useful in applications where associativity is not a natural requirement. Although a
detailed explanation of such applications would require a too large amount of space,
it is worth to mention the use of overlap functions in decision making to define
preference structures [10], in classification to get generalizations of fuzzy rule based
algorithms which improve the results of those based on the use t-norms [28], or
in image processing [22]. In all these cases it is worth to mention that classical
algorithms may be improved with the use of these overlap and grouping functions.

Acknowledgments The authors have been supported by project TIN2013-40765-P.

The Notions of Overlap and Grouping Functions 155

References

1. Alsina, C., Frank, M.J., Schweizer, B.: Associative Functions. Triangular Norms and Copulas.
World Scientific, Hackensack (2006)
2. Amo, A., Montero, J., Molina, E.: Representation of consistent recursive rules. Eur. J. Oper.
Res. 130, 29–53 (2001)
3. Amo, A., Montero, J., Biging, G., Cutello, V.: Fuzzy classification systems. Eur. J. Oper. Res.
156, 459–507 (2004)
4. Bedregal, B., Dimuro, G.P., Bustince, H., Barrenechea, E.: New results on overlap and grouping
functions. Inf. Sci. 249, 148–170 (2013)
5. Bustince, H., Barrenechea, E., Pagola, M., et al.: Weak fuzzy S-subsethood measures. Overlap
index. Int. J. Uncertainty Fuzziness Knowl.-Based Syst. 14(5), 537–560 (2006)
6. Bustince, H., Mohedano, V., Barrenechea, E., et al.: Definition and construction of fuzzy DI-
subsethood measures. Inf. Sci. 176(21), 3190–3231 (2006)
7. Bustince, H., Pagola, M., Barrenechea, E.: Construction of fuzzy indices from fuzzy DI-
subsethood measures: application to the global comparison of images. Inf. Sci. 177(3), 906–929
(2007)
8. Bustince, H., Montero, J., Barrenechea, E., Pagola, M.: Semiautoduality in a restricted family
of aggregation operators. Fuzzy Sets Syst. 158(12), 1360–1377 (2007)
9. Bustince, H., Fernandez, J., Mesiar, R., Montero, J., Orduna, R.: Overlap functions. Nonlin.
Anal.-Theory Meth. Appl. 72, 1488–1499 (2010)
10. Bustince, H., Pagola, M., Mesiar, R., Hullermeier, E., Herrera, F., Montero, J., Orduna, R.:
Grouping, overlap, and generalized bientropic functions for fuzzy modeling of pairwise com-
parisons. IEEE Trans. Fuzzy Syst. 20, 405–415 (2012)
11. Calvo, T., Kolesárová, A., Komorníkova, M., Mesiar, R.: Aggregation operators: properties,
classes and construction methods. In: Aggregation Operators New Trends and Applications.
Physica-Verlag, Heidelberg (2002)
12. Cutello, V., Montero, J.: Recursive connective rules. Int. J. Intell. Syst. 14, 3–20 (1999)
13. Dimuro, G.P., Bedregal, B., Bustince, H., Asiaín, M.J., Mesiar, R.: On additive generators of
overlap functions. Fuzzy Sets Syst. (in press) doi:10.1016/j.fss.2015.02.008
14. Dubois, D., Koning, J.L.: Social choice axioms for fuzzy set aggregation. Fuzzy Sets Syst. 58,
339–342 (1991)
15. Dubois, D., Ostasiewicz, W., Prade, H.: Fuzzy Sets: History and Basic Notions. In: Fundamen-
tals of Fuzzy Sets. Kluwer, Boston (2000)
16. Fodor J., Roubens M., Fuzzy preference modelling and multicriteria decision support. In:
Theory and Decision Library, Kluwer Academic Publishers (1994)
17. Gómez, D., Montero, J.: A discussion on aggregation operators. Kybernetika 40, 107–120
(2004)
18. Gómez, D., Rodríguez, J.T., Montero, J., Bustince, H., Barrenechea, E.: n-dimensional overlap
functions. Fuzzy Sets Syst. (in press). doi:10.1016/j.fss.2014.11.023
19. Hosseini, M.S., Eftekhari-Moghadam, A.M.: Fuzzy rule-based reasoning approach for event
detection and annotation of broadcast soccer video. Appl. Soft Comput. 13(2), 846–866 (2013)
20. Ishibuchi, H., Nakashima, T., Nii, M.: Classification and Modeling with Linguistic Information
Granules: Advanced Approaches to Linguistic Data Mining. Springer, Berlin (2004)
21. Ishibuchi, H., Yamamoto, T.: Rule weight specification in fuzzy rule-based classification sys-
tems. IEEE Trans. Fuzzy Syst. 13, 428–435 (2005)
22. Jurio, A., Bustince, H., Pagola, M., Pradera, A., Yager, R.: Some properties of overlap and
grouping functions and their application to image thresholding. Fuzzy Sets Syst. 229, 69–90
(2013)
23. Klement, E.P., Mesiar, R., Pap, E.: Triangular Norms, Trends in Logic, Studia Logica Library,
vol. 8. Kluwer Academic Publishers, Dordrecht (2000)
24. Klir, G.J., Folger, T.A.: Fuzzy Sets. Uncertainty and Information, Prentice Hall, Englewood
Cliffs (1988)
156 H. Bustince et al.

25. Madrid, N., Burusco, A., Bustince, H., Fernandez, J., Perfilieva, I.: Upper bounding overlaps
by groupings. Fuzzy Sets Syst. 264, 76–99 (2015)
26. Montero, J., Gomez, D., Bustince, H.: On the relevance of some families of fuzzy sets. Fuzzy
Sets Syst. 158, 2429–2442 (2007)
27. Nelsen, R.B.: An introduction to Copulas. Lecture Notes in Statistics, vol. 139. Springer, New
York (1999)
28. Sanz, J., Galar, M., Jurio, A., Brugos, A., Pagola, M., Bustince, H.: Medical diagnosis of
cardiovascular diseases using an interval-valued fuzzy rule-based classification system. Appl.
Soft Comput. J. 20, 103–111 (2014)
29. Schweizer, B., Sklar, A.: Probabilistic Metric Spaces. North-Holland, Amsterdam (1983)
30. Viceník, P.: Additive generators of non-continuous triangular norms, Topological an Algebraic
Structures in Fuzzy Sets. Kluwer (2003)
31. Viceník, P.: Additive generators of associative functions. Fuzzy Sets Syst. 153, 137–160 (2005)
32. Zadeh, L.A.: Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Syst. 1, 3–28 (1978)
Asymmetric Copulas and Their Application
in Design of Experiments

Fabrizio Durante and Elisa Perrone

Abstract We present an overview on definitions and properties of asymmetric

copulas, i.e. copulas whose values are not invariant under any permutation of their
arguments. In particular, we review an axiomatic approach in the definition of a
measure of asymmetry (non–exchangeability) for copulas, starting with the seminal
contributions by Klement and Mesiar [45] and Nelsen [56]. Then we discuss how
asymmetric copulas may be useful also in the optimal design of experiments and
how they may provide additional insights into these problems.

1 Introduction

The problem of describing relationships among random variables has attracted a lot
of attention during the years, especially for inference and prediction. Nowadays, it
can benefit of the use of copulas, which provide a convenient tool to construct and
estimate a multivariate stochastic model. In fact, following a copula approach, we
may determine the joint probability law of a random vector X = (X 1 , . . . , X d ) in two
steps: first, we fix the marginal behavior of each component X i , then we describe the
relationships among the X i ’s by means of a suitable copula, which is a multivariate
distribution function whose univariate marginals are uniformly distributed on [0, 1].
For more details about copulas, see, for instance, [28, 41, 55, 62].
In order to determine possible models for a random vector X, it is then important
to have at disposal a variety of copulas that may cover different situations that arise
in practice. To this end, several investigations in the literature considered the pro-
blem of determining families of copulas and studying their properties with particular
emphasis on their range of association (as measured by Spearman’s correlation coef-

F. Durante
Faculty of Economics and Management, Free University of Bozen–Bolzano, Bolzano, Italy
e-mail: [email protected]
E. Perrone (B)
Institute for Applied Statistics, Johannes Kepler University Linz, Linz, Austria
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 157

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_9
158 F. Durante and E. Perrone

ficient and Kendall’s tau, for instance), tail dependence and so on. For an overview
on different constructions, we refer the reader to [41, 55].
Interestingly, many of these constructions overlap with families of triangular
norms (see, for instance, [46, 62]). In fact, in dimension d = 2, copulas can be seen
as binary operations on [0, 1] that are supermodular and have neutral element 1. In
particular, when they are associative and commutative, they form an interesting class
of triangular norms, known as Archimedean copulas (see, for instance, [33, 53]).
Archimedean copulas have become quite popular in applications because of sev-
eral interesting properties that make them tractable especially for inferential pur-
poses (see, for instance, [35]). However, their main practical limitation is that they
are exchangeable, i.e. the values assumed by the copula does not change under any
permutation of its arguments.
To get more flexibility, a popular approach is to consider (multivariate) models that
exploit bivariate Archimedean (or, generally, exchangeable) copulas for determining
the marginal or conditional dependence and, hence, combining such pieces into
an high–dimension framework. Approaches of such type include vine–pair copula
constructions [7], factor copula models [20, 49, 52] or nested Archimedean copulas.
Another possibility is to find methods for providing constructions that, starting
with a given (exchangeable) copula C, produce another copula C  (usually, with an
additional number of parameters) that need not be exchangeable. See, for instance,
Khoudraji’s device [31, 43] and related extensions [11, 50, 51].
It should also be stressed that some classes of (semi–)parametric bivariate copulas
can explicitly handle the non–exchangeable case. Consider, for instance, extreme–
value and Archimax copulas [5, 6, 47], Liouville copulas [54], copulas that are
invariant under (univariate) truncation [17, 18], asymmetric semilinear copulas [8]
among others.
Methods and tools for coping with non–exchangeable (asymmetric) copulas have
become quite popular in the recent years; see, for instance, the excellent overview
provided in [36]. In this contribution, we shortly review some of the recent litera-
ture on this topic by emphasizing earlier contributions provided in [45, 56], which
have suggested possible ways to quantify non–exchangeability in copula models.
Then, we consider a novel application of (asymmetric) copulas in optimal designs of
experiments and show how their use may provide a convenient tool in this framework
too.

2 A Glimpse of Non–exchangeable Copulas

We start by recalling the basic definition of exchangeability in dimension 2.

Definition 1 A (bivariate) copula C is said to be exchangeable (or symmetric) if

C(u, v) = C(v, u) for all (u, v) ∈ [0, 1]2 .
Asymmetric Copulas and Their Application in Design of Experiments 159

In particular, if (U, V ) is a random pair distributed according to an exchangeable

copula C, then
P(V ≤ v | U ≤ u) = P(U ≤ v | V ≤ u).

Thus, the conditional distribution of (V | U ≤ u) is equal to the conditional distrib-

ution of (U | V ≤ u). In particular, it implies that non–exchangeability occurs when
there is a causality relationship between U and V .
Below, we illustrate several practical examples where non–exchangeability may
appear.
• In the study of financial time series, it is often the case that high losses in one
(large) market often imply high losses in other (smaller) markets, while the oppo-
site direction rarely appears. This phenomenon can be interpreted in terms of
financial contagion that should be captured with copulas that should describe this
asymmetric behavior. See, for instance, [15, 16, 40].
• In reliability theory, the failure of a system can depend (at least) on two charac-
teristics: the usage history and the age. As such, the warranty policy for certain
types of products specifies the limits of coverage in terms of both age and usage.
However, it is often the case that, in warranty claim data analysis, the dependence
between these two features is related to a non–exchangeable copula, as addressed
for instance in [68].
• In environmental applications, consider, for instance, a spatial analysis of a river
basin where data are collected at different gauge stations. Then, due to the physical
and geographic structure of the river networks (for instance, one river can be a
tributary of another one), non–exchangeable dependencies may occur at various
levels. For more details, see for instance [1, 36].

Remark 1 In fuzzy logic (in a broad sense), instead, the adoption of a symmetric
function (e.g., triangular norm) as conjunctive operation is sometimes considered
quite strong. Therefore, several investigations have stressed the importance of intro-
ducing also asymmetric (i.e. non-commutative) conjunctions [3, 19, 30, 37, 42].

Graphically, exchangeability can be interpreted in the sense that the level sets
associated with an exchangeable copula C are symmetric with respect to the main
diagonal of the unit square (see Fig. 1). Analogously, the 3D plot of an exchangeable
copula (which is a surface) is symmetric with respect to the vertical plane passing
through the line {x = y}.
These graphical interpretations also suggest possible ways to transform an
exchangeable copula into another copula that does not share this property. Examples
are illustrated below.

Example 1 Let C be an exchangeable copula that is different from the Fréchet-

Hoeffding upper bound copula M2 (u, v) = min{u, v}. Then C can be modified to a
non–exchangeable copula by means of a suitable patchwork construction [12, 13,
26]. Without loss of generality, let R be a rectangle contained in {(u, v) ∈ [0, 1]2 : u ≥
v} such that VC (R) > 0. Then, one can construct another copula C  whose induced
160 F. Durante and E. Perrone

1.0
0.9

0.8
0.8

0.7

0.6
0.6

0.5

v
0.4
0.4

0.3

0.2
0.2

0.1

0.0
0.0 0.2 0.4 0.6 0.8 1.0
u

Fig. 1 Level curves of a Clayton copula with parameter α = 2

measure μC  coincides with μC on the Borel sets of [0, 1]2 \R, while the probability
mass is spread in a different way on R (see Fig. 2). The copula C  obtained in this
way is, in general, non–exchangeable.
For instance, consider the independence copula Π2 (u, v) = uv and let a ∈ [0, 21 ].
Consider the rectangular patchwork

 = ([0, a] × [1 − a, 1], C1 )Π ,

C

 is not exchangeable and its expression is given by:

where C1 = Π . Then, C
⎧  
⎨a 2 C1 u , v − a + au, (u, v) ∈ [0, a] × [1 − a, 1],
⎪
 v) =
C(u, a a
⎪
⎩uv, otherwise,

where a = 1 − a.

1 1

R TU

TL

0 1 0 1

Fig. 2 Graphical illustration of a rectangular (left) and diagonal (right) patchwork construction
Asymmetric Copulas and Their Application in Design of Experiments 161

Example 2 Let C be an exchangeable copula, C = M2 . For every t ∈ [0, 1], let

δC (t) := C(t, t) be the diagonal section of C. In view of the results in [12], there
exists another exchangeable copula C1 = C whose diagonal section coincides with
C. Moreover, as a consequence of [22, Theorem 1], the following non–exchangeable
copula can be given:

 v) = C(u, v), (u, v) ∈ TL ,
C(u,
C1 (u, v), (u, v) ∈ TU ,

where TL = {(u, v) ∈ [0, 1]2 : u ≥ v}, TU = [0, 1]2 \TL . This type of construction is
known as diagonal patchwork of two copulas (see also [25, 57]).
Remark 2 In previous examples, we start with a copula C = M2 , However, if C =
M2 it can be easily transformed into a non–exchangeable copula by means of a push–
forward of its induced measure under a suitable measure–preserving transformation.
For more details, see [27, 67].
Example 1 can be also used to check that, for every exchangeable copula C = M2 ,
a non–exchangeable copula C  exists that is sufficiently close to C (in the L ∞ norm).
Instead, given a non–exchangeable copula, there is no general way to approximate
it via an exchangeable one.
At a more abstract level, the sets of exchangeable and non–exchangeable copulas
are quite different, as the following general results hold:
• Exchangeable copulas form a closed set in the class C of all copulas endowed
with the L ∞ norm, i.e. a sequence of exchangeable copulas uniformly converges
to an exchangeable copula. In particular, exchangeable copulas are not dense in
C , while non-exchangeable copulas are so.
• In some sense (namely, in the sense of Baire category), the class of exchangeable
copulas is a small set (i.e. a set of first category) in C endowed with the L ∞ norm,
while a typical copula is non-exchangeable (see [14] for more details).
These facts also provide additional motivations to consider in depth non–exchange-
able copulas.

3 Measures of Non–exchangeability

As noted in [56], “in a sense, the relationship between exchangeability and non–
exchangeability is analogous to the relationship between independence and depen-
dence” for random variables. In particular, this analogy could be exploited to
construct a measure of non–exchangeability, mimicking the axiomatic approach by
Rényi [61]. Following this viewpoint, Klement et al. [21] have introduced a set of
axioms for a measure of non–exchangeability for identically distributed and contin-
uous random variables, which turned out to depend only on the associated copula.
This definition is reported below.
162 F. Durante and E. Perrone

Definition 2 A function μ : C → R+ is a measure of non–exchangeability for C if

it satisfies the following properties:
• μ(C) ≤ K for some K ∈ R+ and for all C ∈ C ;
• μ(C) = 0 if, and only if, C is symmetric;
• μ(C) = μ(C t ) for every C ∈ C , with C t (u, v) = C(v, u);
• μ(C) = μ(Ĉ) for every C ∈ C , with C
survival copula related to C;
n n
• If C, (Cn )n∈N ∈ C , Cn → C (pointwise) implies μ(Cn ) → μ(C).

The most popular measures of non–exchangeability can be derived from the L p

distance d p in C ( p ∈ [1, +∞]), given by
 1  1 1/ p
d p (A, B) := |A(u, v) − B(u, v)| du dv p
,
0 0

when p is in [1, +∞[, and, for p = +∞ (see [56]),

d∞ (A, B) := max |A(u, v) − B(u, v)|.

(u,v)∈[0,1]2

In particular, it was proved in [21] the following result.

Theorem 1 For every p ∈ [1, +∞], μ p : C → R+ defined by μ p (C) := d p (C, C t )
is a measure of non–exchangeability.
Moreover, given μ p defined as above, there exists a constant K p ∈ R+ and (at
least) a copula C p ∈ C such that

μ p (C p ) = K p and ∀C ∈ C μ p (C p ) ≥ μ p (C).

Such C p is called maximally non-exchangeable copula with respect to μ p . As a

consequence, we may always suppose that μ p takes values on [0, 1].
It was proved in [45, 56] that

∀C ∈ C max |C(u, v) − C(v, u)| ≤ 13 .

(u,v)∈[0,1]2

Moreover, the copulas that are maximally non–exchangeable with respect to μ∞

have been derived in [45, 56] and their support is depicted in Fig. 3.
 of Example 1 is given by
Example 3 The measure of non-exchageability for C

 = 3a 2
μ+∞ (C) max |C1 (x, y) − x y| .
(x,y)∈[0,1]2

Maximum asymmetry for such a C  is, hence, obtained when C1 belongs to {W, M}.
 3a 2
For such a case, μ+∞ (C) = 4 .
Asymmetric Copulas and Their Application in Design of Experiments 163

1 1

2/3

1/3 1/3

0 2/3 1 0 1/3 2/3 1

Fig. 3 Support of maximally non-exchangeable copulas with respect to μ∞

Remark 3 In [4], the authors determined best possible lower and upper copulas C
under the constraint that, for a fixed t, μ∞ (C) = t. Maximally non–exchangeable
copulas in d dimensions (d ≥ 2) have been provided, instead, in [38]. An alternative
measure of non–exchangeability for copulas has been proposed in [64].

There is a strong nexus between non-exchangeability measures and dependence

measures, as considered also in [56, Sect. 3]. Consider, for instance, the Schweizer–
Wolff measure of dependence given by

κ(C) = 4 max |C(u, v) − uv|.

(u,v)∈[0,1]2

For more details, see [63]. It is easily shown that, for every C ∈ C ,

3
μ∞ (C) ≤ κ(C).
2
Moreover, if we consider other (qualitative) dependence notions like positive quad-
rant dependence, stochastic increasingness, etc., we may notice that:
• positive (resp. negative) dependence plays in favor of exchangeability;
• stronger positive (resp. negative) dependence implies smaller non–exchangeability.
Results of this type can be found, for instance, in [23, 24].
Various possible techniques for measuring asymmetry in copula models have been
also employed in developing statistical tests, as done for instance in [34, 36] (see
also [48, 60]), or to test special copula-based stochastic processes [2].
164 F. Durante and E. Perrone

4 Optimal Design with Non–exchangeable Copulas

As largely discussed in the introduction, copulas are substantially used in several

applied areas. A quite new domain is related to the design of the experiments, an
area where their exploitation has been just recently taken into account (see [9, 58]).
Optimal experimental design is a statistical area mainly applied to environmental
problems, clinical trials testing and industrial procedures. All these applications
relate to the challenging problem of describing phenomena that are generally non-
symmetric. However, in design of experiments, the asymmetry of a specific situation
has usually been reflected by a different behavior of the margins, only. Moreover,
the possible impact of the dependence structure on the design has received little
attention.
A first step in including this latter aspect has been made in [58], where a copula–
based approach has been introduced. In particular, through several examples it has
been shown how taking into account a dependence between the random variables
involved in the phenomenon leads to relevant changes on the obtained design. Never-
theless, it has not yet been considered how the impact of the asymmetry of the copula
influences the optimal design. In the following, we will investigate this latter aspect
by assuming a bivariate stochastic model with identically distributed margins, but
(eventually) asymmetric dependence structures. Specifically, the examined asymme-
try that will be considered is controlled by adding parameters to an exchangeable
copula. Therefore, its impact can be measured by comparing the designs obtained
for the symmetric model with the ones obtained for the asymmetric one. Such a
study might be of interest because it highlights the robustness of the obtained design
against various dependency scenarios.
In the following, we first introduce some basics about experimental design. Then,
we illustrate the impact of the asymmetrization on the designs for different values of
the model parameters.

4.1 Optimal Experimental Design: Some Basics

From now on, we shall consider a vector xT = (x1 , . . . , xr ) ∈ X of control variables,

where X ⊂ Rr is a compact set. We focus directly on the bivariate case.
The results of the observations and of the expectations in a regression experiment
are the vectors:
y(x) = (y1 (x), y2 (x)),

E[Y(x)] = E[(Y1 , Y2 )] = η(x, β) = (η1 (x, β), η2 (x, β)),

where β = (β1 , . . . , βk ) is an unknown parameter vector to be estimated and ηi (i =

1, 2) are known functions. We denote by FYi (yi (x, β)) the margins of each Yi for i ∈
{1, 2}. According to Sklar’s theorem [66], we assume that the dependence between
Asymmetric Copulas and Their Application in Design of Experiments 165

Y1 and Y2 is modeled by a copula function Cα , depending on α = {α1 , . . . , αl }, an

unknown (copula) parameter vector. Hence, the joint model can be expressed in the
form
Cα (FY1 (y1 (x, β)), FY2 (y2 (x, β))).

The Fisher Information Matrix for a single observation, i.e., m(x, γ ), is a (k + l) ×

(k + l) matrix whose elements are given by
  ∂2 
∂2
E − log Cα (FY1 (y1 (x, β)), FY2 (y2 (x, β))) (1)
∂γi ∂γ j ∂ y1 ∂ y2

where γ = {γ1 , . . . , γk+l } = {β1 , . . . , βk , α1 , . . . , αl }.

The aim of design theory is to quantify the amount of information on both sets of
parameters α and β, respectively, from the regression experiment embodied in the
Fisher Information Matrix.
For N independent observations at x1 , . . . , x N , the corresponding Information
matrix is


N 
N  
x1 . . . x N
M(ξ, γ ) = wi m(xi , γ ), wi = 1 and ξ = .
w1 . . . w N
i=1 i=1

The approximate design theory is concerned with finding ξ ∗ (γ ) such that it maxi-
mizes some scalar function φ(M(ξ, γ )), i.e., the so-called design criterion. Here-
inafter, we consider only D-optimality, i.e., φ(M) = log det M, provided M is non-
singular.
The formulation of a Kiefer-Wolfowitz type equivalence relation (see [44]) is the
cornerstone of the theoretical foundation of optimal design. The following theorem
of such type is a generalized version of a result formulated in [39] and based on the
findings in [65]. A detailed proof of this result can be found in [58].

Theorem 2 For a local parameter vector (γ̄ ), the following properties are equiva-
lent:
• ξ ∗ is D-optimal;
• tr [M(ξ ∗ , γ̄ )−1 m(x, γ̄ )] ≤ (k + l), ∀x ∈ X ;
• ξ ∗ minimize max tr [M(ξ ∗ , γ̄ )−1 m(x, γ̄ )], over all ξ ∈ Ξ , where Ξ is the design
x∈X
space.

Theorem 2 allows one to implement standard design algorithms such as of the

Fedorov-Wynn type (see [29, 69]). It also provides simple checks for D-optimality
through the maxima of d(x, ξ ∗ ) = tr [M(ξ ∗ , γ̄ )−1 m(x, γ̄ )], which is usually called
sensitivity function. The sensitivity function plays an important role in the convex
design theory. In fact, its maxima determine the location of the points that are the
most informative with respect to the optimality criterion.
The next definition is important for the comparison of two different designs.
166 F. Durante and E. Perrone

Definition 3 Let (k + l) be the number of the model parameters. The ratio

 1/(k+l)
|M(ξ, γ )|
D(ξ, ξ  ) = (2)
|M(ξ  , γ )|

is called D-Efficiency of the design ξ with respect to the design ξ  .

As it is now evident from previous definitions and results, the optimal designs
depend upon the trend model structure and the chosen copula. Furthermore, such
designs might also be influenced by the unknown parameter values for γ through
the induced nonlinearities (see [59]). Thence, we are resorting to localized designs
around the values γ̄ .

4.2 A Model for Binary Outcomes

In this section we present an example with potential applications in clinical trials.

Let us formally introduce the model. We assume a bivariate binary response
(Yi1 , Yi2 ), i = 1, . . . , n, with four possible outcomes {(0, 0), (0, 1), (1, 0), (1, 1)}
where 1 usually represents a success and 0 a failure of a given treatment. We denote
the joint probabilities of Y1 and Y2 by p y1 ,y2 = P(Y1 = y1 , Y2 = y2 ) where (y1 , y2 ) ∈
{0, 1}2 . In a clinical trial context, Y1 and Y2 could represent, for instance, efficacy
and toxicity of a tested drug (see [9, 10]).
Now, define

p11 = Cα (π1 , π2 ), p10 = π1 − p11 ,

(3)
p01 = π2 − p11 , p00 = 1 − π1 − π2 + p11 ,

where π1 and π2 are the marginal probabilities of success and Cα is a given copula.
Following [39], we assume that
 
πi
log = βi1 + βi2 x, i = 1, 2 (4)
1 − πi

with x ∈ [0, 10]. As shown in [10], in such a model the Fisher information matrix
for a single observation can be written as:
 
∂p T 1 ∂p
m(x, γ ) = P −1 + eeT , (5)
∂γ 1 − p11 − p10 − p01 ∂γ

where p = ( p11 , p10 , p01 ), P = diag(p) and e = (1, 1, 1)T .

Since our focus is on the effects of asymmetric dependence, here we consider
π1 = π2 , i.e., the same marginal behavior of Y1 and Y2 . Specifically, we restrict the
Asymmetric Copulas and Their Application in Design of Experiments 167

parameters related to the margins to β1 = (β11 , β12 ) with ‘localized’ initial values
β̄ 1 = [−1, 1].
We start with an exchangeable model, where the probability of success p11 is
given by
p11 = FY1 ,Y2 (π1 , π1 ; α1 ) = C(π1 , π1 ; α1 ),

with C being an exchangeable copula. The idea is to take into account a transforma-
tion of the copula C in order to obtain an asymmetric dependence for the random
vector (Y1 , Y2 ). In practice, this is obtained by bringing two additional parameters
into the model that may induce the asymmetry. Then a natural way how to inquire the
effect of the applied transformation is simply to compare the designs obtained for C
and those obtained for the asymmetric version of C, by calculating the corresponding
D-efficiencies.
The copula transformation we take into account consists of modifying a given
exchangeable copula C = Cα1 , with parameter α1 , into the copula C̃ = C̃α1 ,α2 ,α3
defined, for every (u, v) ∈ [0, 1]2 , by

C̃(u, v) = u α2 vα3 Cα1 (u 1−α2 , v1−α3 ), (6)

where α2 , α3 ∈ [0, 1]. For α2 = α3 , C̃ is non–exchangeable. Transformation (6) is the

well-known Khoudraji’s asymmetrization, described in [43]. Notice that the depen-
dence in C̃ of (6) is limited since, as shown in [32],

(1 − α2 )(1 − α3 )
τ (C̃) ≤ =: τmax (C̃).
(1 − α2 ) + (1 − α3 ) − (1 − α2 )(1 − α3 )

It is worth to stress that, while in the symmetric model the estimated parameter vector
is (β11 , β12 , α1 ), in the asymmetric case we bring the focus on the asymmetrization
by considering as vector of the estimated parameters (β11 , β12 , α1 , α2 , α3 ).
In our example, the symmetric copula Cα1 is assumed belonging to the Clay-
ton family and the following scheme is carried on. For some values of Kendall’s τ
for the copula Cα1 of (3), we find the D-Efficiency of the optimal design ξ , which
corresponds to the exchangeable Clayton copula with the given τ , with respect to
the optimal design ξ̄ ∗ , which corresponds to the asymmetric model of Eq. (6) with
(suitable) parameters (α˜1 , α˜2 , α˜3 ) in order to ensure that τ (C) = τ (C̃). The losses
in D-efficiency (in percentage) are hence used to quantify the impact of the asym-
metrization on the design.
In this small numerical illustration we assume that τ ∈ {0.10, 0.25, 0.50}, while
(α˜2 , α˜3 ) may vary in {0, 0.2, 0.4, 0.6, 0.8} (without loss of generality, we consider
α2 > α3 ). However, it should be noted that some combinations of (α˜2 , α˜3 ) may not
provide the given values of Kendall’s measure of association.
Table 1 shows the losses in D-Efficiency for various localized parameter vectors.
Analyzing the results, one may notice that the losses are generally quite substantial.
Such high losses prove that the symmetric model and the asymmetric one provide
reasonably different optimal designs. An evidence of the difference between the two
168 F. Durante and E. Perrone

Table 1 Losses in D-efficiency (in percentage) between Clayton and asymmetric Clayton models
τ α1 (α˜2 , α˜3 ) α˜1 τmax (C̃(α˜1 ,α˜2 ,α˜3 ) ) Loss in D–efficiency (in %)
0.50 2.00 (0.2, 0.0) 3.5 0.80 43.99
0.25 0.66 (0.2, 0.0) 0.9 0.80 49.49
0.10 0.22 (0.2, 0.0) 0.3 0.80 55.70
0.50 2.00 (0.4, 0.0) 11 0.60 75.70
0.25 0.66 (0.4, 0.0) 1.5 0.60 51.74
0.10 0.22 (0.4, 0.0) 0.4 0.60 56.27
0.25 0.66 (0.6, 0.0) 3.6 0.40 63.80
0.10 0.22 (0.6, 0.0) 0.7 0.40 57.90
0.10 0.22 (0.8, 0.0) 2.2 0.20 67.46
0.25 0.66 (0.4, 0.2) 2.0 0.52 51.98
0.10 0.22 (0.4, 0.2) 0.5 0.52 56.52
0.25 0.66 (0.6, 0.2) 5.3 0.36 64.86
0.10 0.22 (0.6, 0.2) 0.9 0.36 58.07
0.10 0.22 (0.8, 0.2) 3.5 0.19 70.37
0.25 0.66 (0.6, 0.4) 9.5 0.31 65.83
0.10 0.22 (0.6, 0.4) 1.4 0.31 57.42

obtained designs can also be seen by simply comparing Fig. 4a or b with Fig. 5. In fact,
the optimal design found for the asymmetric model has three support points, while
the one related to the symmetric model just has two support points. Moreover, Table 1
also indicates that the losses depend upon the chosen transformation. As a matter
of fact, by looking at the values corresponding to the same fixed τ , a considerable
increase of loss relative to stronger asymmetries can be observed. This aspect is
also observable from Fig. 4, where the different geometry of two optimal designs
corresponding to two distinct asymmetrizations is shown.

(a) (b)
4 4

3 3

2 2

1 1

x x

Fig. 4 Sensitivity functions (continuous lines) and design weights (bars) for asymmetric Clayton
with (α˜1 , α˜2 , α˜3 ) = (11, 0.4, 0) (Fig. a) and (α˜1 , α˜2 , α˜3 ) = (3.5, 0.2, 0) (Fig. b), respectively
Asymmetric Copulas and Their Application in Design of Experiments 169

Fig. 5 Sensitivity function 3

(continuous line) and design
weights (bars) for symmetric
Clayton with τ = 0.5 2

5 Conclusions

This paper provides an overview on basics and motivations of asymmetric copu-

las. The work deals with both theoretical and practical aspects of the usage of such
functions. With the aim of highlighting the importance of asymmetric copula mod-
els in applications, an example on optimal experimental design is carried on. The
study is conducted by comparing designs obtained for an initially symmetric model
with optimal designs for asymmetric transformations of the original (symmetric)
model. Results for stronger and weaker asymmetrizations are reported. Overall, in
the presented illustration, symmetric and asymmetric dependence structures affect
the optimal design in a different way. Moreover, the effect of the transformation
seems to increase for stronger asymmetrizations. As a conclusion, this work shows
that, even by assuming equal marginal behavior, the non-exchangeability of a spe-
cific phenomenon may be caught by the dependence structure thanks to the use of
flexible copula models.

Acknowledgments This article is devoted to Prof. Erich Peter Klement on the occasion of his
retirement.
The first author has been supported by the Faculty of Economics and Management of Free University
of Bozen-Bolzano, Italy, via the project “Model Uncertainty and Dependence”.
The second author would like to thank Prof. Werner Müller, for his support and useful suggestions.
She was supported by the project ANR-2011-IS01-001-01 “DESIRE” and Austrian Science Fund
(FWF) I 883-N18.

References

1. Bacigal, T., Jágr, V., Mesiar, R.: Non-exchangeable random variables, Archimax copulas and
their fitting to real data. Kybernetika (Prague) 47(4), 519–531 (2011)
2. Beare, B.K., Seo, J.: Time irreversible copula-based Markov models. Econom. Theory 30(5),
923–960 (2014)
3. Běhounek, L., Bodenhofer, U., Cintula, P., Saminger-Platz, S., Sarkoci, P.: Graded dominance
and related graded properties of fuzzy connectives. Fuzzy Sets Syst. 262, 78–101 (2015)
4. Beliakov, G., De Baets, B., De Meyer, H., Nelsen, R.B., Úbeda-Flores, M.: Best-possible
bounds on the set of copulas with given degree of non-exchangeability. J. Math. Anal. Appl.
417(1), 451–468 (2014)
170 F. Durante and E. Perrone

5. Capéraà, P., Fougères, A.L., Genest, C.: Bivariate distributions with given extreme value attrac-
tor. J. Multivar. Anal. 72(1), 30–49 (2000)
6. Charpentier, A., Fougères, A.L., Genest, C., Nešlehová, J.G.: Multivariate Archimax copulas.
J. Multivar. Anal. 126, 118–136 (2014)
7. Czado, C.: Pair-copula constructions of multivariate copulas. In: Jaworski, P., Durante, F.,
Härdle, W., Rychlik, T. (eds.) Copula Theory and Its Applications. Lecture Notes in Statistics—
Proceedings, vol. 198, pp. 93–109. Springer, Berlin (2010)
8. De Baets, B., De Meyer, H., Mesiar, R.: Asymmetric semilinear copulas. Kybernetika (Prague)
43(2), 221–233 (2007)
9. Denman, N.G., McGree, J.M., Eccleston, J.A., Duffull, S.B.: Design of experiments for bivari-
ate binary responses modelled by Copula functions. Comput. Stat. Data Anal. 55(4), 1509–1520
(2011)
10. Dragalin, V., Fedorov, V.: Adaptive designs for dose-finding based on efficacytoxicity response.
J. Stat. Plan. Infer. 136(6), 1800–1823 (2006)
11. Durante, F.: Construction of non-exchangeable bivariate distribution functions. Stat. Pap. 50(2),
383–391 (2009)
12. Durante, F., Fernández-Sánchez, J.: On the classes of copulas and quasi-copulas with a given
diagonal section. Int. J. Uncertain. Fuzziness Knowl.-Based Syst. 19(1), 1–10 (2011)
13. Durante, F., Fernández-Sánchez, J., Sempi, C.: Multivariate patchwork copulas: a unified
approach with applications to partial comonotonicity. Insur. Math. Econ. 53, 897–905 (2013)
14. Durante, F., Fernández-Sánchez, J., Trutschnig, W.: Baire category results for exchangeable
copulas. Fuzzy Sets Syst. 284, 146–151 (2016)
15. Durante, F., Foscolo, E., Jaworski, P., Wang, H.: A spatial contagion measure for financial time
series. Expert Syst. Appl. 41(8), 4023–4034 (2014)
16. Durante, F., Jaworski, P.: Spatial contagion between financial markets: a copula-based approach.
Appl. Stoch. Models Bus. Ind. 26(5), 551–564 (2010)
17. Durante, F., Jaworski, P.: Invariant dependence structure under univariate truncation. Statistics
46(2), 263–277 (2012)
18. Durante, F., Jaworski, P., Mesiar, R.: Invariant dependence structures and Archimedean copulas.
Stat. Probab. Lett. 81(12), 1995–2003 (2011)
19. Durante, F., Klement, E., Mesiar, R., Sempi, C.: Conjunctors and their residual implicators:
characterizations and construction methods. Mediterr. J. Math. 4(3), 343–356 (2007)
20. Durante, F., Klement, E., Quesada-Molina, J., Sarkoci, P.: Remarks on two product-like con-
structions for copulas. Kybernetika (Prague) 43(2), 235–244 (2007)
21. Durante, F., Klement, E., Sempi, C., Úbeda-Flores, M.: Measures of non-exchangeability for
bivariate random vectors. Stat. Pap. 51(3), 687–699 (2010)
22. Durante, F., Kolesárová, A., Mesiar, R., Sempi, C.: Copulas with given diagonal sections: novel
constructions and applications. Int. J. Uncertain. Fuzziness Knowl.-Based Syst. 15(4), 397–410
(2007)
23. Durante, F., Papini, P.L.: Componentwise concave copulas and their asymmetry. Kybernetika
(Prague) 45(6), 1003–1011 (2009)
24. Durante, F., Papini, P.L.: Non-exchangeability of negatively dependent random variables.
Metrika 71(2), 139–149 (2010)
25. Durante, F., Rodríguez-Lallena, J.A., Úbeda-Flores, M.: New constructions of diagonal patch-
work copulas. Inf. Sci. 179(19), 3383–3391 (2009)
26. Durante, F., Saminger-Platz, S., Sarkoci, P.: Rectangular patchwork for bivariate copulas and
tail dependence. Comm. Stat. Theory Methods 38(15), 2515–2527 (2009)
27. Durante, F., Sarkoci, P., Sempi, C.: Shuffles of copulas. J. Math. Anal. Appl. 352(2), 914–921
(2009)
28. Durante, F., Sempi, C.: Principles of Copula Theory. CRC/Chapman and Hall, Boca Raton
(2015)
29. Fedorov, V.V.: The design of experiments in the multiresponse case. Theory Probab. Appl.
16(2), 323–332 (1971)
Asymmetric Copulas and Their Application in Design of Experiments 171

30. Fodor, J.C., Keresztfalvi, T.: Nonstandard conjunctions and implications in fuzzy logic. Int. J.
Approx. Reason. 12(2), 69–84 (1995)
31. Genest, C., Ghoudi, K., Rivest, L.P.: Understanding relationships using copulas, by E. Frees
and E. Valdez, January 1998. N. Am. Actuar. J. 2(3), 143–149 (1998)
32. Genest, C., Kojadinovic, I., Nešlehová, J., Yan, J.: A goodness-of-fit test for bivariate extreme-
value copulas. Bernoulli 17(1), 253–275 (2011)
33. Genest, C., MacKay, R.J.: Copules archimédiennes et familles de lois bidimensionnelles dont
les marges sont données. Can. J. Stat. 14(2), 145–159 (1986)
34. Genest, C., Nešlehová, J., Quessy, J.F.: Tests of symmetry for bivariate copulas. Ann. Inst. Stat.
Math. 64(4), 811–834 (2012)
35. Genest, C., Nešlehová, J., Ziegel, J.: Inference in multivariate Archimedean copula models.
TEST 20(2), 223–256 (2011)
36. Genest, C., Nešlehová, J.: Assessing and modeling asymmetry in bivariate continuous data. In:
Jaworski, P., Durante, F., Härdle, W. (eds.) Copulae in Mathematical and Quantitative Finance.
Lecture Notes in Statistics, pp. 91–114. Springer, Berlin (2013)
37. Hájek, P., Mesiar, R.: On copulas, quasicopulas and fuzzy logic. Soft Comput. 12(12), 123–
1243 (2008)
38. Harder, M., Stadtmüller, U.: Maximal non-exchangeability in dimension d. J. Multivar. Anal.
124, 31–41 (2014)
39. Heise, M.A., Myers, R.H.: Optimal designs for bivariate logistic regression. Biometrics 52(2),
613–624 (1996)
40. Jaworski, P., Pitera, M.: On spatial contagion and multivariate GARCH models. Appl. Stoch.
Models Bus. Ind. 30, 303–327 (2014)
41. Joe, H.: Dependence Modeling with Copulas. Chapman and Hall/CRC, London (2014)
42. Kawaguchi, M.F., Miyakoshi, M.: Composite fuzzy relational equations with noncommutative
conjunctions. Inf. Sci. 110(1–2), 113–125 (1998)
43. Khoudraji, A.: Contributions à l’étude des copules et à la modélisation des valeurs extrêmes
bivariées. Ph.D. thesis, Université de Laval, Québec (Canada) (1995)
44. Kiefer, J., Wolfowitz, J.: The equivalence of two extremum problems. Can. J. Math. 12, 363–366
(1960)
45. Klement, E.P., Mesiar, R.: How non-symmetric can a copula be? Comment. Math. Univ. Car-
olin. 47(1), 141–148 (2006)
46. Klement, E.P., Mesiar, R., Pap, E.: Triangular norms. Trends in Logic-Studia Logica Library,
vol. 8. Kluwer Academic Publishers, Dordrecht (2000)
47. Klement, E.P., Mesiar, R., Pap, E.: Archimax copulas and invariance under transformations.
C. R. Math. Acad. Sci. Paris 340(10), 755–758 (2005)
48. Kojadinovic, I., Yan, J.: A non-parametric test of exchangeability for extreme-value and left-tail
decreasing bivariate copulas. Scand. J. Stat. 39(3), 480–496 (2012)
49. Krupskii, P., Joe, H.: Factor copula models for multivariate data. J. Multivar. Anal. 120, 85–101
(2013)
50. Liebscher, E.: Construction of asymmetric multivariate copulas. J. Multivar. Anal. 99(10),
2234–2250 (2008)
51. Liebscher, E.: Erratum to Construction of asymmetric multivariate copulas. J. Multivar. Anal.
102(4), 869–870 (2011)
52. Mazo, G., Girard, S., Forbes, F.: A flexible and tractable class of one-factor copulas. Stat.
Comput., in press (2015)
53. McNeil, A.J., Nešlehová, J.: Multivariate Archimedean copulas, d-monotone functions and
1 -norm symmetric distributions. Ann. Stat. 37(5B), 3059–3097 (2009)
54. McNeil, A.J., Nešlehová, J.: From Archimedean to Liouville copulas. J. Multivar. Anal. 101(8),
1772–1790 (2010)
55. Nelsen, R.B.: An Introduction to Copulas. Springer Series in Statistics, 2nd edn. Springer, New
York (2006)
56. Nelsen, R.B.: Extremes of nonexchangeability. Stat. Pap. 48(2), 329–336 (2007)
172 F. Durante and E. Perrone

57. Nelsen, R.B., Quesada-Molina, J.J., Rodríguez-Lallena, J.A., Úbeda-Flores, M.: On the con-
struction of copulas and quasi-copulas with given diagonal sections. Insur. Math. Econom.
42(2), 473–483 (2008)
58. Perrone, E., Müller, W.G.: Optimal design for copula models. Statistics, in press (2015).
doi:10.1080/02331888.2015.1111892
59. Pronzato, L., Pázman, A.: Design of Eperiments in Nonlinear Models. Springer Lecture Notes
in Statistics, vol. 212 (2013)
60. Quessy, J.F., Bahraoui, T.: Graphical and formal statistical tools for assessing the symmetry of
bivariate copulas. Can. J. Stat. 41(4), 637–656 (2013)
61. Rényi, A.: On measures of dependence. Acta Math. Acad. Sci. Hungar. 10, 441–451 (1959)
62. Schweizer, B., Sklar, A.: North-Holland Series in Probability and Applied Mathematics. Prob-
abilistic metric spaces. North-Holland Publishing Co., New York (1983)
63. Schweizer, B., Wolff, E.F.: On nonparametric measures of dependence for random variables.
Ann. Stat. 9(4), 879–885 (1981)
64. Siburg, K.F., Stoimenov, P.A.: Symmetry of functions and exchangeability of random variables.
Stat. Pap. 52(1), 1–15 (2011)
65. Silvey, S.D.: Optimal Design (Science Paperbacks). Chapman and Hall (1980)
66. Sklar, A.: Fonctions de répartition à n dimensions et leurs marges. Publications de l’Institut de
Statistique de Paris 8, 229–231 (1959)
67. Trutschnig, W., Fernández Sánchez, J.: Some results on shuffles of two-dimensional copulas.
J. Stat. Plan. Infer. 143(2), 251–260 (2013)
68. Wu, S.: Construction of asymmetric copulas and its application in two-dimensional reliability
modelling. Eur. J. Oper. Res. 238, 476–485 (2014)
69. Wynn, H.P.: The sequential generation of D-optimum experimental designs. Ann. Math. Stat.
41(5), 1655–1664 (1970)
Copulæ of Processes Related to the Brownian
Motion: A Brief Survey

Carlo Sempi

Abstract The copulas of a few stochastic processes related to the Brownian motion
are derived; specifically, if (Xt ) is one such process, the copula of the pair (Xs , Xt ) is
determined for s < t.

1 Introduction

The study of the copulas of stochastic processes is rapidly becoming important.

However, not all aspects of stochastic processes can be dealt with through copulas.
In general, one may safely say that, given a stochastic process (Xt ) and n times
t1 < t2 < . . . < tn , the distribution function(=d.f.) of the random variables Xt1 , …,
Xtn can in principle be amenable to calculation via copulas. Here we survey known
results on the deduction of copulas Cs,t of pairs (Xs , Xt ) with s < t where (Xt )t≥0 is
a stochastic process related to the Brownian motion, in a sense to be made precise
in the following. Also, following [6], we establish the copula of a Brownian motion
(Bt ) and its supremum St . Although the present paper is a survey a few new results
are presented, notably in Sects. 3 and 4. Of the growing literature on these aspects
only the items that are strictly relevant for the present chapter will be quoted.
We briefly recall the definition of a (bivariate) copula and the main properties that
will be needed.

Definition 1 A copula is a function C : I2 → I, where I = [0, 1] such that:

(a) for every u ∈ I, C(u, 0) = C(0, u) = 0 and C(u, 1) = C(1, u) = u;
(b) C is 2–increasing: for all u, u , v and v in I with u ≤ u and v ≤ v

C(u , v ) − C(u, v ) − C(u , v) + C(u, v) ≥ 0 .

C. Sempi (B)
Dipartimento di Matematica e Fisica “Ennio De Giorgi”,
Università del Salento, Lecce 73100, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 173

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_10
174 C. Sempi

In other words, a copula C is (the restriction to the unit square I2 of) a d.f. that
concentrates all the probability mass on I2 and that has uniform margins. The main
result is provided by Sklar’s theorem [7].

Theorem 1 Let a random vector X = (X, Y ) be given on a probability space

(Ω, F , P), let H(x, y) := P(X ≤ x, Y ≤ y) be the joint d.f. of X, and let F(x) =
P(X ≤ x) and G(y) = P(Y ≤ y) be its marginals. Then there exists a copula C = CX
such that, for every point (x, y) ∈ R2 ,

H(x, y) = C (F(x), G(y)) . (1)

If the marginals F and G are continuous, then the copula C is uniquely defined.

Under the assumptions of Theorem 1, if F and G are continuous, then there exists
a unique copula C associated with X that is determined, for all (u, v) ∈ [0,1[2 , via
the formula
 
C(u, v) = H F (−1) (u), G(−1) (v) , (2)

where F (−1) (t) := inf{x ∈ R : F(x) ≥ t} is the right–continuous quasi–inverse of F;

similarly for G(−1) . When both F and G are continuous one may speak of the copula
of the random vector X and denote it by CX .

Theorem 2 ([1]) In the probability space (Ω, F , P) let the continuous random
variables X and Y have d.f.’s F and G and copula C. Then, one has a.e.
 
P(X ≤ x | Y )(ω) = E 1{X≤x} | Y (ω) = ∂2 C (FX (x), FY (Y (ω))) , (3)
 
P(Y ≤ y | X)(ω) = E 1{Y ≤y} | X (ω) = ∂1 C(FX (X(ω)), FY (y)) . (4)

Here we have set

∂f (s, t) ∂f (s, t)
∂1 f (s, t) := , and ∂2 f (s, t) := .
∂s ∂t
Theorem 3 Let X and Y be continuous random variables defined on the probability
space (Ω, F , P) and consider the continuous mappings

f : Ran X → R and g : Ran Y → R .

(a) If both f and g are strictly increasing, then, for all (u, v) ∈ I2 ,

Cf (X),g(Y ) (u, v) = CXY (u, v) ;

Copulæ of Processes Related to the Brownian Motion: A Brief Survey 175

(b) if both f and g are strictly decreasing, then, for all (u, v) ∈ I2 ,

Cf (X),g(Y ) (u, v) = C σ1 σ2 (u, v) = u + v − 1 + CXY (1 − u, 1 − v) .

We refer to [4] and to the forthcoming monograph [2] for the properties of copulas
and to [3, 5, 8] for those of stochastic processes.

2 The Copula of a Brownian Motion

With the exception of the last remark, the results of this section were established
in [1].
Given a standard Brownian motion (=BM, for short) (Bt )t≥0 on the probabilitiy
space (Ω, F , P), we wish to find the copula Cs,tB
of the pair (Bs , Bt ) with s < t.
The transition probabilities for the BM (Bt ) are given, for s ∈ ]0, t[ and x and y
in R, by  
y−x
P(x, s; y, t) := P(Bt ≤ y | Bs = x) =  √ ,
t−s

where  denotes the distribution function of the standard normal law N(0, 1). By
Theorem 2 one has
P(x, s; y, t) = ∂1 Cs,t
B
(Fs (x), Ft (y)) ,

where Ft is the d.f. of Bt . Therefore

 Fs (x)
B
Cs,t (Fs (x), Ft (y)) = ∂1 Cs,t
B
(w, Ft (y)) dw
0
 Fs (x)  
y−x
=  √ dw . (5)
0 t−s

As is well known (see, e.g., [3] or [5]) the d.f. Ft of Bt is given, for t > 0, by
 
x
Ft (x) =  √ .
t

This is both continuous and strictly increasing. As a consequence the copula of

(Bs , Bt ) is uniquely determined; moreover, Ft has an inverse, which will be needed
shortly and which is easily calculated
√
Ft−1 (u) = t −1 (u) (u ∈ ] 0, 1[) .
176 C. Sempi

B
Replacing this result into (5) yields the copula Cs,t
  √ −1 √ 
u
t  (v) − s −1 (w)
B
Cs,t (u, v) =  √ dw . (6)
0 t−s

This copula has partial derivatives given, for (u, v) ∈ ]0, 1[2 by
 √ −1 √ 
t  (v) − s −1 (u)
B
=
∂1 Cs,t (u, v) √ , (7)
t−s
  u  √ √ 
1 t t −1 (v) − s −1 (w)
∂2 Cs,t
B
(u, v) =  −1  ϕ √ dw ,
ϕ  (v) t−s 0 t−s

where ϕ is the density of the standard normal law N(0, 1). A further derivation of
B
(7) provides the density of Cs,t
 √ √ 
t 1 t −1 (v) − s −1 (u)
B
cs,t (u, v) =  ϕ √ . (8)
t − s ϕ  (v)
−1 t−s

Notice that, because of Theorem 3(a), the copula of the Brownian bridge Bt∗ =
Bt − t B1 coincides with the copula (6) of the Brownian motion (Bt ).

3 The Copula of the Geometric Brownian Motion

A geometric Brownian motion (Xt ) satisfies the stochastic integral Eq. (see, e.g., [3])
 t  t
Xt = X0 + μ Xs ds + ν Xs dBs , (9)
0 0

where μ ∈ R, ν > 0, X0 = 0. The solution of (9) is given by

  
1 2
Xt = X0 exp μ − ν t + ν Bt .
2

Thus if X0 > 0, then Xt is a strictly increasing function of Bt , Xt = ft (Bt ), where

  
1 2
ft (x) := X0 exp μ− ν t+νx .
2

As a consequence of Theorem 3 (a) the copula Cs,t GB+

of the random vector (Xs , Xt )
with s ∈ ]0, t[ coincides with the copula Cs,t of the BM given by (6) and has the same
B

density (8).
Copulæ of Processes Related to the Brownian Motion: A Brief Survey 177

If, on the other hand, X0 < 0, then Xt is a strictly decreasing function of Bt ,

Xt = ft (Bt ), with the same function ft as above, which is now strictly decreasing.
Therefore, by Theorem 3(b) the copula of (Xs , Xt ) is given by
GB−
Cs,t (u, v) = u + v − 1 + Cs,t
B
(1 − u, 1 − v)
 1−u  √ −1 √ 
t  (1 − v) − s −1 (w)
=u+v−1+  √ dw , (10)
0 t−s

which differs from the copula of Eq. (6). When X0 < 0, the density of Cs,t
GB−
(u, v) is
given by
 √ √ 
t 1 t −1 (1 − v) − s −1 (1 − u)
GB−
cs,t (u, v) =  ϕ √ .
t − s ϕ −1 (1 − v) t−s

4 The Copula of the Ornstein–Uhlenbeck Process

The Ornstein–Uhlenbeck process Ut solves the stochastic differential equation

√
dUt = −Ut dt + 2 dBt , (11)

where (Bt ) is a Brownian motion. The connexion between the Ornstein–Uhlenbeck

process (Ut ) and the Brownian motion (Bt ) is given by

Ut = 2−t Be2t , (12)

so that the distribution function FtOU of Ut is

   
FtOU (x) = P (Ut ≤ x) = P e−t Be2t ≤ x = P Be2t ≤ et x
 t 
e x
= √ = (x) .
e2t

Thus
 OU −1
Ft (u) = −1 (u) .

The joint distribution function of (Us , Ut ) is given, for s ∈ ]0, t[ and for x and y in
R by
 OU 
OU
Fs,t (x, y) = P(Us ≤ x, Ut ≤ y) = Cs,t
OU
Fs (x), FtOU (y)
= Cs,t
OU
((x), (y)) , (13)
178 C. Sempi

so that
 −1 
OU
Cs,t (u, v) = Fs,t
OU
 (u), −1 (v) .

On the other hand, one has, in view of Eq. (5),

 
OU
Fs,t (x, y) = P(Us ≤ x, Ut ≤ y) = P e−s Be2s ≤ x, e−t Be2t ≤ y
   
= P Be2s ≤ es x, Be2t ≤ et y = CeB2s ,e2t FeB2s (es x), FeB2t (et y)
  s   t 
e x e x
= Ce2s ,e2t  √
B
, √ = CeB2s ,e2t ((x), (y)) . (14)
e2s e2t

By comparing Eqs. (13) and (14), one obtains

  
u
et −1 (v) − es −1 (w)
OU
Cs,t (u, v) =  √ dw , (15)
0 e2t − e2s

which is the expression of the copula of the Ornstein–Uhlenbeck process.

Its density is given by
 
et et −1 (v) − es −1 (u)
OU
cs,t (u, v) =√ ϕ √ .
e2t − e2s e2t − e2s

5 The Supremum of a Browian Motion

The supremum (St ) of a Brownian motion (Bt ) is defined by

St := sup{Bs : s ≤ t} .

By recourse to the reflection principle, see, e.g., [5, Sect. 1.3], one has
 
y
FSt (y) = 2  √ − 1 ;
t

since  is strictly increasing, FSt has an inverse given by

 
√ v+1
FS−1 (v) = t −1 . (16)
t
2

Let Ht denote the joint distribution function of (Bt , St ); for x ≤ y, one has, because
of the continuity of both Bt and St and as a consequence of Eq. (13.4) in [5],
Copulæ of Processes Related to the Brownian Motion: A Brief Survey 179

Ht (x, y) = P (Bt ≤ x, St ≤ y) = P (Bt ≤ x) − P (Bt ≤ x, St ≥ y)

 
x
=  √ − P (Bt ≤ y − (y − x), St ≥ y)
t
 
x
=  √ − P (Bt ≥ 2y − x)
t
    
x 2y − x
= √ − 1− √
t t
   
x x − 2y
= √ − √ , (17)
t t

while, if x > y, then

 
y
Ht (x, y) = P(St ≤ y) = 2  √ − 1 . (18)
t

Then the unique copula CtBS of the random vector (Bt , St ) can be obtained by
 
CtBS (u, v) = Ht FB−1
t
(u), FS−1
t
(v) .

From this, one has, if u ≤ (v + 1)/2, or, equivalently, if FB−1

t
(u) ≤ FS−1
t
(v),
 
√ 1 −1
CtBS (u, v) =  t √  (u)
t
    
1 √ −1 √ v+1
− √ t  (u) − 2 t −1
t 2
  
v+1
= u −  −1 (u) − 2 −1 . (19)
2

If, on the other hand, u > (v + 1)/2, then

  
√ 1 −1 v + 1
CtBS (u, v) = 2 t√  −1 = v.
t 2

Notice that in either case the result does not depend on time: the copula of Bt and
St is independent of t and is given by

  
u −  −1 (u) − 2 −1 v+1 , u ≤ (v + 1)/2 ,
C (u, v) =
BS 2
v, u ≥ (v + 1)/2 .
180 C. Sempi

References

1. Darsow, W.F., Nguyen, B.E., Olsen, T.: Copulas and Markov processes. Illinois J. Math. 36,
600–642 (1992)
2. Durante, F., Sempi, C.: Principles of Copula Theory. Chapman & Hall/CRC, Boca Raton (2015)
3. Karatzas, I., Shreve, S.E.: Brownian Motion and Stochastic Calculus, 2nd edn. Springer, Berlin
(1991)
4. Nelsen, R.B.: An Introduction to Copulas, 2nd edn. Springer, New York (2006)
5. Rogers, L.C.G., Williams, D.: Diffusions, Markov processes and martingales. In: Foundations,
2nd Edn. Cambridge University Press (2000)
6. Schmitz, V.: Copulas and stochastic processes. Ph.D. Dissertation, Rheinische-Westfälische
Technische Hochschule, Aachen, Germany (2003)
7. Sklar, A.: Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Statist. Univ. Paris
8, 229–231 (1959)
8. Stirzaker, D.: Stochastic processes and models. Oxford University Press (2005)
Extensions of Capacities

Anna Kolesárová and Andrea Stupňanová

Abstract We study extensions of capacities on N = {1, . . . , n} to n-ary aggregation

functions acting on [0, 1]n . Besides recalling the universal integral based approaches
following the ideas of Klement et al. and generalizations of the Lovász and Owen
extensions, we also present some new approaches to extending such capacities which
are based on a generalization of the formulas for the discrete Choquet and Sugeno
integrals.

1 Introduction

In decision-making processes crisp alternatives are characterized by {0, 1}-valued

score vectors. Given a set N = {1, . . . , n} of criteria, crisp alternatives are described
by score vectors x = (x1 , . . . , xn ) ∈ {0, 1}n and evaluated by Boolean utility func-
tions. A Boolean (normed) utility function is a pseudo-Boolean nondecreasing
function u : {0, 1}n → [0, 1] satisfying the properties u(0) = u(0, . . . , 0) = 0 and
u(1) = u(1, . . . , 1) = 1. When criteria are evaluated in the graded scale [0, 1], there
is a need to extend Boolean utility functions to utility functions acting on score vec-
tors x ∈ [0, 1]n . Utility functions U : [0, 1]n →[0, 1] should also be nondecreasing
to satisfy the Pareto principle. Boolean utility functions u : {0, 1}n → [0, 1] can be
identified with capacities on the set N and normed utility functions U : [0, 1]n →
[0, 1] are in a one-to-one correspondence with n-ary aggregation functions on the
interval [0, 1]. Recall that a capacity m on the set N , [9, 30], is a nondecreasing set
function m : 2 N → [0, 1] with the properties m(∅) = 0 and m(N ) = 1. The corre-
sponding m and u are linked by the relation m(K ) = u(1 K ), K ⊆ N , where 1 K is

A. Kolesárová (B)
Faculty of Chemical and Food Technology, Slovak University of Technology in Bratislava,
Radlinského 9, 812 37 Bratislava 1, Slovakia
e-mail: [email protected]
A. Stupňanová
Faculty of Civil Engineering, Slovak University of Technology in Bratislava,
Radlinského 11, 810 05 Bratislava 1, Slovakia
e-mail: [email protected]
© Springer International Publishing Switzerland 2016 181
S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_11
182 A. Kolesárová and A. Stupňanová

the indicator of the set K . An n-ary aggregation function (n ∈ N, n ≥ 2) on the inter-

val [0, 1] is a function A : [0, 1]n →[0, 1] which is nondecreasing in each variable
and satisfies the boundary conditions A(0) = 0 and A(1) = 1, see [1, 3, 8]. Instead
of extensions of Boolean utility functions to utility functions one can investigate
extensions of capacities to aggregation functions, i.e., for a given capacity m to look
for aggregation functions A satisfying the property A(1 K ) = m(K ) for all K ⊆ N .
Throughout this paper, the set of all capacities on N will be denoted by Mn and the
set of all n-ary aggregation functions on [0, 1] by An . The aim of this contribution is
to discuss extensions of capacities to n-ary functions acting on [0, 1]n . Note that, in
general, these extensions need not be monotone and their range need not be included
in [0, 1]. In such cases, we will look for the constraints ensuring that for any capacity
m on N the corresponding extension is an aggregation function, i.e., it can be seen
as a normed utility function satisfying the Pareto principle.
There are several kinds of integrals on N which are based on capacities. Though
these integrals are monotone, they are not always extensions of the corresponding
capacities. For example, the concave integral introduced by Lehrer [16] is an exten-
sion of a capacity m only if m is totally balanced. In particular, if m is supermod-
ular then the concave integral coincides with the Choquet integral [5, 7]. Similarly,
in general, the convex integral [22] does not extend the corresponding capacity. It
coincides with the Choquet integral if and only if m is a submodular capacity. The
Pan-integral of Yang and Klir [31], see also [30], which is based on a semiring
(R+ , ⊕, ), is an extension of the considered capacity m only in some particular
cases, namely if m is ⊕-superadditive, i.e., if m(K ∪ L) ≥ m(K ) ⊕ m(L) whenever
K , L ⊂ N , K ∩ L = ∅. On the other hand, universal integrals on [0, 1], introduced
by Klement et al. in [12], can serve as a positive example of a method how to extend
a capacity m : 2 N → {0, 1} into an aggregation function A : [0, 1]n → [0, 1] so that
m(K ) = A(1 K ) for any K ⊆ N .
The contribution is organized as follows. In the next section, we recall and exem-
plify universal integrals and bring several non-classical integrals. In particular, we
recall extremal semicopula-based integrals and hierarchical classes of copula-based
integrals. We also recall a related extension method proposed by Klement et al. in
[11]. In Sect. 3, we discuss the Möbius transform-based extension method general-
izing the classical Lovász and Owen extensions [17, 26] and an extension method
based on the possibilistic Möbius transform. In Sect. 4, we show new types of exten-
sion methods based on a generalization of formulas for the discrete Choquet and
Sugeno integrals.

2 Universal Integrals on [0, 1]

A function ⊗ : [0, 1]2 → [0, 1] is called a semicopula [6] if it is monotone and

e = 1 is its neutral element, i.e., x ⊗ 1 = 1 ⊗ x = x for each x ∈ [0, 1]. Observe
that semicopulas are not required to be associative or symmetric.
Extensions of Capacities 183

In what follows, we recall the concept of a universal integral on [0, 1], which was
introduced by Klement et al. in [12], as a common framework covering the Choquet,
Shilkret and Sugeno integrals.

Definition 1 A mapping I : (Mn × [0, 1]n ) → [0, 1] is called a universal inte-
n∈N
gral on [0, 1] whenever it satisfies the following axioms:
(UI1) For each fixed n ∈ N, I |Mn × [0, 1]n is nondecreasing in both components.
(UI2) There is a semicopula ⊗ : [0, 1]2 → [0, 1] such that for any n ∈ N, m ∈ Mn ,
c ∈ [0, 1] and K ⊆ N ,
I (m, c · 1 K ) = c ⊗ m(K ).

(UI3) For any (m 1 , x1 ), (m 2 , x2 ) ∈ (Mn × [0, 1]n ) such that m 1 ({i ∈ N1 | x1,i
n∈N
≥ t}) = m 2 ({ j ∈ N2 | x2, j ≥ t}) it holds that

I (m 1 , x1 ) = I (m 2 , x2 ).

By the axiom (UI2), it holds that

I (m, 1 K ) = 1 ⊗ m(K ) = m(K ),

i.e., universal integrals always extend the considered capacities. The monotonicity
of I (m, ·) is guaranteed due to (UI1), i.e., I (m, ·) is an aggregation function.
Now, we recall three basic universal integrals.
• The Choquet integral was introduced in 1953 in [5] as follows

1
Ch m (x) = m({i ∈ N | xi ≥ t})dt. (1)
0

There are three other equivalent formulas for the discrete Choquet integral. The
first one is based on the Möbius transform Mm : 2 N → R of a capacity m, given by

Mm (K ) = (−1)card(K \L) m(L), (2)
L⊆K

and was introduced in [4]:


Ch m (x) = Mm (K ) min{xi | i ∈ K }. (3)
K ⊆N

Note that the right-hand side of this formula is also known as the Lovász extension
of m [17].
184 A. Kolesárová and A. Stupňanová

Taking into account a geometrical meaning of (1), we can write the following
equivalent formulas for the discrete Choquet integral:


n
   
Ch m (x) = x(i) − x(i−1) m K (i) (4)
i=1

and

n
    
Ch m (x) = x(i) m K (i) − m K (i+1) , (5)
i=1

where (·) : N → N is a permutation such that x(1) ≤ · · · ≤ x(n) and K (i) = {(i), . . . ,
(n)}, i = 1, . . . , n. By convention, in (4), x(0) = 0 and in (5), K (n+1) = ∅. Note that if
there are some ties between the input values, there exist more permutations satisfying
the above constraints, but both formulas (4) and (5) always give the same value of
Ch m (x).

• The Shilkret integral was introduced in 1971 in [27], originally for maxitive mea-
sures only, as follows

Sh m (x) = sup {t · m({i ∈ N | xi ≥ t}) | t ∈ [0, 1]} , (6)

and, equivalently, it can be written as

   
Sh m (x) = sup x(i) · m K (i) | i ∈ N . (7)

• The Sugeno integral was introduced in 1974 in [29] (in Japanese published in
1972) as follows

Su m (x) = sup{min{t, m({i ∈ N | xi ≥ t})} | t ∈ [0, 1]}. (8)

Equivalently, it holds that

    
Su m (x) = sup min x(i) , m K (i) | i ∈ N . (9)

Formally, both the Shilkret and Sugeno integrals can be seen as particular exam-
ples of the smallest universal integral on [0, 1] related to a semicopula ⊗, which was
introduced in [12] by the formula

I⊗ (m, x) = sup {t ⊗ m({i ∈ N | xi ≥ t}) | t ∈ [0, 1]} , (10)

which can also be written as

   
I⊗ (m, x) = sup x(i) ⊗ m K (i) | i ∈ N . (11)
Extensions of Capacities 185

Namely, the Shilkret integral can be obtained for the semicopula ⊗ = Π , Π (x, y) =
x y, and the Sugeno integral for ⊗ = Min, Min(x, y) = min{x, y}.
Now, let us consider copulas [25] as special types of semicopulas. Recall that
C : [0, 1]2 → [0, 1] is a copula if it is a supermodular semicopula, i.e., if for all
x, y ∈ [0, 1]2 it holds that

C(x ∨ y) + C(x ∧ y) ≥ C(x) + C(y).

Clearly, for ⊗ = C, the formula (10) gives

IC (m, x) = sup {C(t, m({i ∈ N | xi ≥ t}) | t ∈ [0, 1]} , (12)

which can also be written as

    
IC (m, x) = sup C x(i) , m K (i) | i ∈ N .

A special class of universal integrals on [0, 1] based on copulas was introduced

in [12], see also [11], as follows:
 
I[C] (m, x) = PC (x, y) ∈ [0, 1]2 | y < m({i ∈ N | xi ≥ x . (13)

Here PC is a probability measure on Borel subsets of the unite square [0, 1]2
generated by the probabilities of the rectangles [0, x] × [0, y], PC ([0, x] × [0, y]) =
C(x, y). Consequently, using the notation as in (4) and (5), we can write the formula
(13) in the form


n
    
I[C] (m, x) = C x(i) , m(K (i) ) − C x(i−1) , m(K (i) ) (14)
i=1

and also as


n
    
I[C] (m, x) = C x(i) , m(K (i) ) − C x(i) , m(K (i+1) ) . (15)
i=1

Clearly, if C is the standard product copula Π , which models the independence of

random variables, then (14) turns into (4) and (15) turns into (5). Hence, I[Π] is the
Choquet integral, i.e., I[Π] (m, x) = Ch m (x) for any m ∈ Mn and x ∈ [0, 1]n , n ∈ N.
Similarly, one can show that the comonotone dependence copula Min generates
the Sugeno integral, i.e., for any m ∈ Mn and x ∈ [0, 1]n , n ∈ N, I[Min] (m, x) =
Su m (x).
In [13], a family of general copula-based integrals was proposed, compare also
[23]. We adopt this proposal to the framework of discrete universal integrals as
follows:
186 A. Kolesárová and A. Stupňanová

Definition 2 Let n ∈ N and let C : [0, 1]2 → [0, 1] be a fixed copula. The (n, C)-
universal integral on [0, 1],
 
IC(n) : Mk × [0, 1]k → [0, 1],
k∈N

is given by
⎧ ⎛ ⎛ ⎛⎧ ⎫⎞ ⎞
⎨ n i ⎨  i ⎬
IC(n) (m, x) = sup ⎝C ⎝ a j , m ⎝ p ∈ {1, . . . , k} | x p ≥ a j ⎠⎠
⎩ ⎩ ⎭
i=1 j=0 j=0
⎛ ⎛⎧ ⎫⎞⎞⎞⎫

i−1 ⎨ i ⎬ ⎬
− C⎝ a j , m ⎝ p ∈ {1, . . . , k} | x p ≥ a j ⎠⎠⎠ ,
⎩ ⎭ ⎭
j=0 j=0
(16)
n
where a0 = 0, a1 , . . . , an ≥ 0 and j=1 a j ≤ 1.

Remark 1 Note that if n = 1, for an arbitrary semicopula S : [0, 1]2 → [0, 1], the
functional I S(1) = I S given by (16) is the (weakest) universal integral linked to S,
compare (10) and (11). However, I S(2) does not satisfy the axiom of the monotonicity
of universal integrals, in general. To ensure this, S should be supermodular, i.e., a
copula.

Example 1 Consider n = 3 and the uniform capacity m ∈ M3 , m(E) = card(E)

3
. Let
S : [0, 1]2 → [0, 1] be a semicopula such that
       
1 2 2 1 2 2 1 1 1
S , =S , =S , = , and S , = 0.
3 3 3 3 3 3 3 3 3

For example, the function

  
1 1
S(x, y) = med Min(x, y), W (x, y), max x − , y − ,
3 3

where W : [0, 1]2 → [0, 1], W (x, y) = max{0, x + y − 1}, is the Fréchet-Hoeffding
lower bound (the weakest copula), satisfies the given conditions. Then
  
2 2 2 2
I S(2) m, , , = ,
3 3 3 3
        
1 2 2 1 2 2 1 2 1 1 1 1
I S(2) m, , , = S ,1 + S , −S , = + − = ,
3 3 3 3 3 3 3 3 3 3 3 3
        
1 1 2 1 2 1 1 1 1 1 2
I S(2) m, , , = S ,1 + S , −S , = + −0= ,
3 3 3 3 3 3 3 3 3 3 3
Extensions of Capacities 187

which shows that the functional I S(2) (m, ·) is not monotone.

Proposition 1 Let C : [0, 1]2 → [0, 1] be a copula. Then

(i) IC = IC(1) ≤ IC(2) ≤ · · · ≤ IC(n) ≤ · · · ≤ I[C] .
(ii) For a fixed n, for any m ∈ Mn and any copula C, the integrals IC(n) and I[C]
coincide.

The integrals IC(n) , n ∈ N, can be seen as lower approximations of the integral

I[C] . In a similar way, we can introduce upper approximations of I[C] .

Definition 3 Let n ∈ N and let C : [0, 1]2 → [0, 1] be a fixed copula. The (C, n)-
universal integral on [0, 1],
 
C
I(n) : Mk × [0, 1]k → [0, 1],
k∈N

is given by
⎧ ⎛ ⎛ ⎛⎧ ⎫⎞⎞
⎨ n i−1 ⎨ i−1 ⎬
C
I(n) (m, x) = inf ⎝C ⎝ a j , m ⎝ p ∈ {1, . . . , k} | x p > a j ⎠⎠
⎩ ⎩ ⎭
i=1 j=0 j=0
⎛ ⎛⎧ ⎫⎞⎞⎞⎫

i−1 ⎨ 
i−1 ⎬ ⎬
− C⎝ a j , m ⎝ p ∈ {1, . . . , k} | x p > a j ⎠⎠⎠ ,
⎩ ⎭ ⎭
j=0 j=0
(17)
n
where a0 = 0, a1 , . . . , an ≥ 0 and max{x1 , . . . , xn } ≤ j=1 a j ≤ 1.
C
Observe that for any copula C it holds that I(1) (m, x) = (max xi , m({i ∈ N | xi >
0})).

Proposition 2 Let C : [0, 1]2 → [0, 1] be a copula. Then

C
(i) I(1) ≥ I(2)
C
≥ · · · ≥ I(n)
C
≥ · · · ≥ I[C] .
(ii) For a fixed n, for any m ∈ Mn and any copula C, the integrals I(n)
C
and I[C]
coincide.

Summarizing Propositions 1 and 2, we get the following conclusion.

Corollary 1 For a fixed n, for any m ∈ Mn , copula C and x ∈ [0, 1]n , it holds that

IC (m, x) = IC(1) (m, x) ≤ IC(2) (m, x) ≤ · · · ≤ IC(n) (m, x) = I[C] (m, x)

= I(n)
C
(m, x) ≤ · · · ≤ I(2)
C
(m, x) ≤ I(1)
C
(m, x).
188 A. Kolesárová and A. Stupňanová

Example 2 Let n = 3 and let m ∈ M3 be a uniform capacity, m(E) = card(E)

and
  3
x = 13 , 23 , 1 . Then for the product copula Π we have

4
IΠ (m, x) = Sh m (x) = IΠ(1) (m, x) = ,
9
5
IΠ(2) (m, x) = ,
9
6
IΠ(3) (m, x) = I(3)
Π
(m, x) = Ch m (x) = ,
9
Π 7
I(2) (m, x) = ,
9
Π 9
I(1) (m, x) = = 1.
9
For the Fréchet-Hoeffding lower bound W we get:

1
IW (m, x) = IW(1) (m, x) = ,
3
2
IW(2) (m, x) = ,
3
3
IW(3) (m, x) = I[W ] (m, x) = I(3)
W
(m, x) = I(2)
W
(m, x) = I(1)
W
(m, x) = = 1,
3
and for the Fréchet-Hoeffding upper bound Min:

(1) 2
I Min (m, x) = Su m (x) = I Min (m, x) = · · · = I(2)
Min
(m, x) = ,
3
Min
I(1) (m, x) = 1.

3 Möbius Transform-Based Extensions of Capacities

The formula (3) can be seen as an extension of the capacity m constructed by means
of the Möbius transform of m and the aggregation function Min, which inspired us
to propose a generalization of this approach. We briefly recall the main idea and the
results from [14].
Let x = (x1 , . . . , xn ) ∈ [0, 1]n be any input n-tuple. To each subset K ⊆ N we
assign an n-tuple x K = (u 1 , . . . , u n ), where

xi if i ∈ K ,
ui =
1 otherwise.
Extensions of Capacities 189

Clearly, x∅ = (1, . . . , 1) = 1 and x N = x. Let A be an n-ary aggregation function,

m a capacity on N and Mm its Möbius transform. Let us define the function Fm,A :
[0, 1]n → R by 
Fm,A (x1 , . . . , xn ) = Mm (K ) A(x K ). (18)
K ⊆N

Note that, in general, there is a difference between formulas (3) and (18). While in (3)
the aggregation function Min, Min(x1 , . . . , xk ) = min{x1 , . . . , xk }, is considered as
a k-ary aggregation function for k = 1, . . . , n, in (18) only a fixed n-ary aggregation
function A is applied. The k-tuples corresponding to the index sets K ⊆ N with
|K | = k are always completed into n-tuples by setting to 1 the components with the
indices in N \K . Moreover, formula (3) defines an extended aggregation function
[8], i.e., it can be applied for any arity n. However, for the n-ary aggregation function
A = Min for which e = 1 is a neutral element, both formulas coincide, i.e., Fm,Min
given by (18) is the Lovász extension of m [8, 17, 20]. Similarly, for the product
aggregationfunction A  = Π , Π (x1 , . . . , xn ) = x1 · x2 · . . . · xn , (18) gives the same
values as Mm (K ) xi , i.e., Fm,Π is the so-called Owen extension of m [26].
K ⊆N i∈K
In general, the function Fm,A defined by (18) is neither an extension of m nor an
aggregation function.
Let us first characterize all aggregation functions A ∈ An such that for each m ∈
Mn , Fm,A is an extension of m.

Theorem 1 Let A ∈ An . For each m ∈ Mn , the function Fm,A defined by (18) is an

extension of m if and only if A is an aggregation function with zero annihilator.

As mentioned above, our aim is to characterize all aggregation functions A ∈ An

with the property that, for all m ∈ Mn , the function Fm,A is an aggregation function
extending m. Let us first focus on the binary case:
Let A ∈ A2 be a binary aggregation function with zero annihilator and m ∈ M2 .
If m is determined by the values a, b ∈ [0, 1], m({1}) = a, m({2}) = b, then

Fm,A (x, y) = a A(x, 1) + b A(1, y) + (1 − a − b)A(x, y). (19)

For characterizing the functions A ∈ A2 satisfying our requirement, the notion of

2-dimensional quasi-copula (quasi-copula for short) is needed. Recall that a function
Q : [0, 1]2 → [0, 1] is called a quasi-copula if it is a 1-Lipschitz semicopula, i.e., a
semicopula satisfying for all x, x  , y, y  ∈ [0, 1] the property

|Q(x  , y  ) − Q(x, y)| ≤ |x  − x| + |y  − y|.

Note that each copula C is also a quasi-copula, but in general, the opposite claim is
not true.
190 A. Kolesárová and A. Stupňanová

Theorem 2 Let A ∈ A2 . For each m ∈ M2 , the function Fm,A given by (18) is an

aggregation function extending m if and only if for each (x, y) ∈ [0, 1]2 it holds that

A(x, y) = Q( f (x), g(y)), (20)

where Q is a quasi-copula and f, g are nondecreasing [0, 1] → [0, 1] functions

with the properties f (0) = g(0) = 0, f (1) = g(1) = 1.

Corollary 2 For aggregation functions A ∈ A2 given by A(x, y) = Q( f (x), g(y))

with Q, f and g satisfying Theorem 2, it holds that

Fm,A (x, y) = Fm,Q ( f (x), g(y)).

Example 3 Consider the Hamacher product (with parameter 0) C H0 : [0, 1]2 →

xy
[0, 1] given by C H0 (x, y) = x+y−x y
whenever (x, y) = (0, 0) (and C H0 (0, 0) =
0). Note that C H0 is a copula, i.e., also a quasi-copula. Define the functions
f, g : [0, 1] → [0, 1] by f (x) = g(x) = 1+x
2x
. Then, by Theorem 2, we obtain the
aggregation function A,

2x y
A(x, y) = C H0 ( f (x), g(y)) = ,
x+y

i.e., the standard harmonic mean.

Let m be any capacity in M2 , with m({1}) = a, m({2}) = b, a, b ∈ [0, 1].
Then, using the formula (19), see also Corollary 2, we get the aggregation function
Fm,A : [0, 1]2 → [0, 1],

2ax 2by 2x y
Fm,A (x, y) = + + (1 − a − b) ,
x +1 y+1 x+y

which extends the capacity m.

Before characterizing all n-ary aggregation functions A ∈ An satisfying our

requirement, let us recall that for an n-ary aggregation function A the A-volume
of an n-box [a, b] in [0, 1]n , [a, b] = [a1 , b1 ] × · · · × [an , bn ], is defined by

V A ([a, b]) = (−1)α(c) A(c),

where the sum is taken over all vertices c = (c1 , . . . , cn ) of the n-box [a, b] (i.e.,
each ck is equal to either ak or bk ), and α(c) is the number of indices k  s such that
ck = ak .
For n ≥ 2, we have the following general result.
Extensions of Capacities 191

Theorem 3 For a given aggregation function A ∈ An , n ≥ 2, the following claims

are equivalent.
(i) For each fuzzy measure m ∈ Mn , the function Fm,A is an aggregation function
extending m.
(ii) A is an aggregation function with zero annihilator and for each [a, b] ⊆ [0, 1]n
such that {0, 1} ∩ {a1 , . . . , an , b1 , . . . , bn } = ∅, the A-volume V A ([a, b]) is non-
negative.

For example, all n-copulas [25, 28] are suitable aggregation functions for our
construction. Recall that n-copulas are defined as functions C : [0, 1]n → [0, 1]
satisfying
(C1) the boundary conditions:
if 0 ∈ {x1 , . . . , xn } then C(x1 , . . . , xn ) = 0,
C(1, . . . , 1, x j , 1, . . . , 1) = x j for each j = 1, . . . , n and each x j ∈ [0, 1],
(C2) the n-increasing property:
VC ([a, b]) ≥ 0 for each n-box [a, b] in [0, 1]n .
It is easy to see that the aggregation functions described in the following proposition
also have zero annihilator and that the volumes of all n-boxes in [0, 1]n are non-
negative.

Proposition 3 Let C be an n-copula, f i : [0, 1] → [0, 1], i = 1, . . . , n, nonde-

creasing functions such that for each i, f i (0) = 0, f i (1) = 1. Then the function
A : [0, 1]n → [0, 1] defined by

A(x1 , . . . , xn ) = C ( f 1 (x1 ), . . . , f n (xn )) ,

is an n-ary aggregation function such that, for all m ∈ Mn , the function Fm,A is an
aggregation function extending m.

However, there are aggregation functions A which are neither copulas nor obtained
by a distortion of some copula as in Proposition 3, and in spite of that, for all m ∈ Mn ,
Fm,A is an aggregation function extending m, see the following example.

Example 4 Consider the function A : [0, 1]3 → [0, 1], given by

A(x, y, z) = x yz min(1, x + y + z).

The function A is a ternary aggregation function with zero annihilator. After a quite
tedious computations one obtains that, for each m ∈ M3 , Fm,A is an aggregation func-
tion extending m. However, A is not a copula, because, e.g., for a = (0.3, 0.3, 0.3)
and b = (0.35, 0.35, 0.35) the A-volume of the corresponding 3-box is V A ([a, b]) =
−0.0019 < 0, i.e., the 3-increasing property of A fails. This example can be gener-
alized to any n > 3.
192 A. Kolesárová and A. Stupňanová

Finally, let us mention that in [15] we also studied extension methods based on
the possibilistic Möbius transform Mm∨ of a capacity m, where Mm∨ : 2 N → [0, 1] is
given by ⎧
⎨0 if m(K ) = m(L)
Mm∨ (K ) = for some L  K , (21)
⎩
m(K ) otherwise.

More details on the possibilistic Möbius transform can be found, e.g., in [19, 21].
For a semicopula ⊗ : [0, 1]2 → [0, 1], a capacity m ∈ Mn and an aggregation
∨,⊗
function A ∈ An , let Fm,A : [0, 1]n → [0, 1] be a function given by

∨,⊗
  
Fm,A (x) = Mm∨ (K ) ⊗ A(x K ) , (22)
K ⊆N

where the meaning of x K is the same as in (18).

∨,⊗
Note that the function Fm,A can also be written as

∨,⊗

Fm,A (x) = (m(K ) ⊗ A(x K )). (23)
K ⊆N

∨,⊗
Theorem 4 Let A ∈ An , n ≥ 2. Then the function Fm,A is for each m ∈ Mn and
each semicopula ⊗ an aggregation function extending m if and only if A is an
aggregation function with zero annihilator.

Let us note that

• Taking ⊗ = ∧, where ∧(x, y) = min{x, y}, and A = Min, we obtain
∨,∧
Fm,Min (x) = Su m (x),

∨,∧
i.e., Fm,Min is the Sugeno integral with respect to the capacity m (compare with
the relation of Fm,Min and the Choquet integral Ch m ).
∨,Π
• In general, for any n ≥ 2, Fm,Min leads to the Shilkret integral Sh m .
∨,⊗
• Similarly, for any semicopula ⊗, the function Fm,Min coincides with the weakest
universal integral I⊗ based on ⊗ with respect to the capacity m, see (10).

4 Modifications of the Choquet and Sugeno Integrals

4.1 Fusion Function-Based Discrete Choquet-Like Integral

Inspired by the formula (4) for the discrete Choquet integral, we proposed its fusion
function-based modification [24], substituting the product in (4) by a binary fusion
Extensions of Capacities 193

function. Note that a binary fusion function is any function F : [0, 1]2 → [0, 1],
see [2].
Definition 4 ([24]) Let F : [0, 1]2 → [0, 1] be a fusion function satisfying for
each y ∈ [0, 1] the property F(0, y) = 0. Then the functional CmF : [0, 1]n → [0, n]
given by

n
 
CmF (x) = F x(i) − x(i−1) , m(K (i) ) . (24)
i=1

is called a fusion function-based discrete Choquet-like integral (F-based discrete

Choquet-like integral for short).
The meaning of symbols in (24) is the same as in (4). Note that due to the property
F(0, y) = 0 valid for each y ∈ [0, 1], the functional CmF is defined correctly, see [24].
Therefore, we will always work with F ∈ F0 , where

F0 = {F : [0, 1]2 → [0, 1] | ∀y ∈ [0, 1], F(0, y) = 0}.

It is clear that for F = Π , for any capacity m ∈ Mn , CmΠ is equal to the Choquet
integral, CmΠ = Ch m .
However, in general, functions CmF need not be monotone. Moreover, their range
need not be contained in [0, 1], see the following example.
Example 5 Consider the function
⎧
⎨ 0 if x = 0,
F(x, y) = y if x = 1,
⎩
1 otherwise.

Then, for a uniform capacity m ∈ Mn , m(K ) = cardn (K ) , and any x ∈]0, 1]n , such that
card({x1 , . . . , xn }) = n, we have CmF (x) = n and CmF (1 K ) = m(K ) for each K ⊆ N .
Considering, for example, n = 3 and the points (0, 0, 0), (0.2, 0.3, 0.7) and (1, 1, 1),
we get CmF (0, 0, 0) = 0, CmF (0.2, 0.3, 0.7) = 3 and CmF (1, 1, 1) = 1, which shows
that the monotonicity of CmF is violated and Ran(CmF )  [0, 1]. Note that CmF is an
extension of m for any m ∈ Mn .
A deep study of properties of functionals CmF can be found in [24]. In this paper we
only focus on the properties of fusion functions F which are necessary and sufficient
for CmF to be an aggregation function and an extension of m for any m ∈ Mn .
Proposition 4 Let F : [0, 1]2 → [0, 1] be a fusion function. Then, for any capac-
ity m ∈ Mn , CmF extends m, i.e., CmF (1 K ) = m(K ) for each K ⊆ N , if and only if
F(1, y) = y for each y ∈ [0, 1].
Now, we fix n = 2. Then any capacity m ∈ M2 is completely determined by values
a, b ∈ [0, 1], where a = m({1}) and b = m({2}). For any F ∈ F0 and m ∈ M2 it
holds that
194 A. Kolesárová and A. Stupňanová

F(x, 1) + F(y − x, b) if x ≤ y,
CmF (x, y) = (25)
F(y, 1) + F(x − y, a) otherwise.

Note that always CmF (0, 0) = 0 and CmF (1, 1) = 1 whenever F(1, 1) = 1.
Let us denote by F(2) the class of all binary fusion functions F ∈ F0 such that for
each m ∈ M2 , CmF is a binary aggregation function. The following theorem provides
a characterization of all F ∈ F(2) .
Theorem 5 Let F ∈ F0 . Then F ∈ F(2) if and only if F(x, 1) = x for all x ∈ [0, 1]
and the function F(·, y) : [0, 1] → [0, 1] is increasing and 1-Lipschitz for each y ∈
[0, 1], i.e.,
0 ≤ F(x2 , y) − F(x1 , y) ≤ x2 − x1

whenever 0 ≤ x1 ≤ x2 ≤ 1.
Note that if CmF is an aggregation function for any m ∈ M2 , then CmF is idempotent
and translation invariant [24].
Before giving an example, we recall that each quasi-copula Q satisfies the con-
straints of Theorem 5, hence Q ∈ F(2) .
Example 6 (i) Consider the greatest quasi-copula Min and a capacity m ∈ M2 ,
with m({1}) = a, m({2}) = b. Then

min{y, x + b} if x ≤ y,
CmMin (x, y) =
min{x, y + a} otherwise
= max {min{y, x + b}, min{x, y + a}} ,

see Fig. 1.
(ii) Consider the smallest quasi-copula W and m ∈ M2 , determined by a, b ∈ [0, 1]
as in (i). Then

max{x, y + b − 1} if x ≤ y,
CmW (x, y) =
max{y, x + a − 1} otherwise
= min {max{x, y + b − 1}, max{y, x + a − 1}} ,

see Fig. 2.
In Theorem 5, we have characterized the class F(2) of all binary fusion functions
F ∈ F0 such that the F-based discrete Choquet-like integral CmF defined on [0, 1]2
is a binary aggregation for any m ∈ M2 . Now, we will characterize the class F(n) ,
n > 2, of all binary fusion functions F ∈ F0 for which CmF defined on [0, 1]n is an
n-ary aggregation function for any m ∈ Mn .
Theorem 6 Let F ∈ F0 . For any n > 2, F ∈ F(n) if and only if F ∈ G , where G
is the set of all F ∈ F0 which are given by F(x, y) = x f (y), for some increasing
function f : [0, 1] → [0, 1] satisfying f (1) = 1.
Extensions of Capacities 195

Fig. 1 2D illustration of 1
CmMin from Example 6(i) for 0.9
a = 1/2, b = 1/3 0.8 x+b
0.7 y
0.6
b
0.5
0.4
0.3 x
0.2
y+a
0.1
0
0 0.2 0.4 0.6 0.8 1
a

Fig. 2 2D illustration of CmW 1

from Example 6(ii) for 0.9 y+b−1
a = 1/2, b = 1/3 0.8
0.7
0.6 x
1−b 0.5
0.4
0.3 y
0.2 x+a−1
0.1
0
0 0.2 0.4 0.6 0.8 1
1−a

Remark 2 In the case of Theorem 6, if F ∈ F(n) , then CmF = Ch f (m) , where Ch f (m)
is the Choquet integral with respect to the capacity f (m) ∈ Mn given by

0 if K = ∅,
f (m)(K ) =
f (m(K )) otherwise.

Indeed, if F ∈ F(n) = G , then F(x, y) = x f (y), and


n
   
n
    
CmF (x) = F x(i) − x(i−1) , m K (i) = x(i) − x(i−1) f m K (i)
i=1 i=1
= Ch f (m) (x).

Summarizing, we have proved that for any n > 2, an F-based discrete Choquet-
like integral CmF is an n-ary aggregation function for any m ∈ Mn if and only if
it is the Choquet integral with respect to the capacity m distorted by a function f
generating F, i.e., CmF = Ch f (m) .
In our approach, we have generalized the formula (4) of the discrete Choquet
integral. Note that the formula (5) can be generalized in a similar way. For more
details we refer to [10].
196 A. Kolesárová and A. Stupňanová

4.2 Fusion Function-Based Discrete Sugeno-Like Integral

The formula (9) for the discrete Sugeno integral Su m : [0, 1]n → [0, 1] can be writ-
ten as
n
  
Su m (x) = min x(i) , m K (i) .
i=1

Inspired by this formula, for any fusion function F : [0, 1]2 → [0, 1], we can define
the function Su mF : [0, 1]n → [0, 1] by the formula


n
  
Su mF (x) = F x(i) , m K (i) , (26)
i=1

compare with [18].

Example 7 (i) For fusion functions F given by F(x, y) = x λ y, with λ ∈]0, ∞[,
the formula (26) gives
 
Su mF (x) = Sh m x1λ , . . . , xnλ .
 
(ii) If F(x, y) = min x λ , y , λ ∈]0, ∞[, then
 
Su mF (x) = Su m x1λ , . . . , xnλ .

Clearly, for λ = 1, in (i) we obtain F = Π and Su Π m (x) = Sh m (x) for each x ∈

[0, 1]n (compare with (7)), i.e., our approach also covers the Shilkret integral.

We first show a sufficient condition for F ensuring that Su mF is a pre-aggregation

function for any capacity m. A function A : [0, 1]n → [0, 1] is said to be a pre-
aggregation function [18] if it satisfies the boundary conditions of aggregation func-
tions A(0) = 0 and A(1) = 1, and if there is at least one direction r = (r1 , . . . , rn ) ∈
[0, 1]n , r = 0, in which A is nondecreasing, i.e.,

A(x1 + cr1 , . . . , xn + crn ) ≥ A(x1 , . . . , xn )

for each (x1 , . . . , xn ) ∈ [0, 1]n and c > 0 such that (x1 + cr1 , . . . , xn + crn ) ∈ [0, 1]n .
Note that aggregation functions are pre-aggregation functions which are nondecreas-
ing in the direction of any vector r ∈ [0, 1]n \ {0}.

Proposition 5 Let F : [0, 1]2 → [0, 1] be a function increasing in the first variable
and let for each y ∈ [0, 1], F(0, y) = 0 and F(1, 1) = 1. Then SmF defined in (26) is
a pre-aggregation function for any capacity m ∈ Mn .

Example 8 Consider the function F : [0, 1]2 → [0, 1], F(x, y) = x|2y − 1|. Note
that F is a proper pre-aggregation function (not an aggregation function) which
Extensions of Capacities 197

satisfies the constraints of Proposition 5, and thus, for any m ∈ Mn , the function

n   
SmF : [0, 1]n → [0, 1], SmF (x) = F x(i) , m A(i) is a pre-aggregation function
i=1
(even an aggregation function). For example, for n = 2, m({1}) = 1/3, m({2}) =
3/4, we get 
x ∨ 2y i f x ≤ y,
Sm (x, y) =
F
y ∨ x3 i f x > y.

Note that any function F satisfying the constraints of Proposition 5 is, in fact,
a binary (1, 0)-nondecreasing pre-aggregation function which satisfies F(0, y) = 0
for each y ∈ [0, 1].
It is not difficult to check that Su mF , given by (26), extends an arbitrary capacity m
whenever F(1, y) = y for each y ∈ [0, 1]. Moreover, the monotonicity of F ensures
the monotonicity of Su mF , independently of m. If F is an aggregation function, so is
CmF . Summarizing these facts, we obtain the following result.

Theorem 7 Let F : [0, 1]2 → [0, 1] be an aggregation function such that F(0, y) =
0 and F(1, y) = y for each y ∈ [0, 1]. Then, for any n ≥ 2 and m ∈ Mn , the func-
tion Su mF : [0, 1]n → [0, 1] defined by (26), is an aggregation function extending the
capacity m.

Acknowledgments This work was supported by the project APVV–14–0013.

References

1. Beliakov, G., Pradera, A., Calvo, T.: Aggregation Functions: A Guide for Practitioners.
Springer, Heidelberg (2007)
2. Bustince, H., Fernandez, J., Kolesárová, A., Mesiar, R.: Directional monotonicity of fusion
functions. Eur. J. Oper. Res. 244, 300–308 (2015)
3. Calvo, T., Kolesárová, A., Komorníková, M., Mesiar, R.: Aggregation operators: properties,
classes and construction methods. In: Calvo, T., Mayor, G., Mesiar, R. (eds.) Aggregation
Operators. New Trends and Applications, pp. 3–107. Physica-Verlag, Heidelberg (2002)
4. Chateauneuf, A., Jaffray, J.Y.: Some characterizations of lower probabilities and other
monotone capacities through the use of Möbius inversion. Math. Soc. Sci. 17, 263–283 (1989)
5. Choquet, G.: Theory of capacities. Ann. Inst. Fourier 5, 131–295 (1953)
6. Durante, F., Sempi, C.: Semicopulae. Kybernetika 41, 315–328 (2005)
7. Grabisch, M.: Fuzzy integral in multicriteria decision making. Fuzzy Sets Syst. 69, 279–298
(1995)
8. Grabisch, M., Marichal, J.-L., Mesiar, R., Pap, E.: Aggregation Functions. Cambridge Univer-
sity Press, Cambridge (2009)
9. Grabisch, M., Murofushi, T., Sugeno, M.: Fuzzy measures and integrals. Theory and Applica-
tions. Physica Verlag, Heidelberg (2000)
10. Horanská, Ľ., Šipošová, A.: A note on a generalization of the Choquet integral. In: Proceedings
of Uncertainty Modelling 2015, STU, Bratislava (2015)
11. Klement, E.P., Mesiar, R., Pap, E.: Measure-based aggregation operators. Fuzzy Sets Syst.
142(1), 3–14 (2004)
198 A. Kolesárová and A. Stupňanová

12. Klement, E.P., Mesiar, R., Pap, E.: A universal integral as common frame for Choquet and
Sugeno integral. IEEE Trans. Fuzzy Syst. 18, 178–187 (2010)
13. Klement, E.P., Mesiar, R., Spizzichino, F., Stupňanová, A.: Universal integrals based on cop-
ulas. Fuzzy Optim. Decis. Making 13, 273–289 (2014)
14. Kolesárová, A., Stupňanová, A., Beganová, J.: Aggregation-based extensions of fuzzy mea-
sures. Fuzzy Sets Syst. 194, 1–14 (2012)
15. Kolesárová, A., Stupňanová, A.: On some extensions methods for normed utility functions. In:
Proceedings AGOP’2011, pp. 169–174. Benevento, Italy (2011)
16. Lehrer, E.: A new integral for capacities. Econ. Theor. 39, 157–176 (2009)
17. Lovász, L.: Submodular function and convexity. In: Mathematical Programming: The State of
the Art, pp. 235–257. Springer, Berlin (1983)
18. Lucca, G., Sanz, J.A., Pereira Dimuro, G., Bedregal, B., Mesiar, R., Kolesárová, A., Bustince,
H.: Pre-aggregation functions: construction and an application. IEEE Trans. Fuzzy Syst. (2015),
accepted for publication
19. Marichal, J.-L., Mathonet, P., Tousset, E.: Mesures floues définies sur une échelle ordinale.
Working paper (1996)
20. Marichal, J.-L.: Aggregation of interacting criteria by means of the discrete Choquet integral. In:
Calvo, T., Mayor, G., Mesiar, R. (eds.) Aggregation Operators. New Trends and Applications,
pp. 224–244. Physica-Verlag, Heidelberg (2002)
21. Mesiar, R.: k-Order pan-additive discrete fuzzy measures. In: 7th IFSA World Congress, pp.
488–490. Prague (1997)
22. Mesiar, R., Li, J., Pap, E.: Superdecomposition integral. Fuzzy Sets Syst. 259, 3–11 (2015)
23. Mesiar, R., Stupňanová, A.: Decomposition integrals. Int. J. Approximate Reasoning 54, 1252–
1259 (2013)
24. Mesiar, R., Kolesárová, A., Bustince, H., Dimuro, G.P., Bedregal, B.: Fusion functions based
Choquet-like integrals. Submitted (2015)
25. Nelsen, R.B.: An Introduction to Copulas, 2nd edn. Springer, New York (2006)
26. Owen, G.: Multilinear extensions of games. In: The Shapley value. In: Roth, A.E. (ed.) Essays
in Honour of Lloyd S. Shapley, pp. 139–151. Cambridge University Press (1988)
27. Shilkret, N.: Maxitive measure and integration. Indag. Math. 33, 109–116 (1971)
28. Sklar, A.: Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Stat. Univ. Paris
8, 229–231 (1959)
29. Sugeno, M.: Theory of fuzzy integrals and its applications. PhD thesis, Tokyo Institute of
Technology (1974)
30. Wang, Z., Klir, G.J.: Fuzzy Measure Theory. Plenum Press, New York (1992)
31. Yang, Q.: The pan integral on the fuzzy measure space. Fuzzy Math. 3, 107–114 (1985). (in
Chinese)
Multi-source Information Fusion Using
Measure Representations

Ronald R. Yager

Abstract We first look at the issue of representing information about an uncertain

variable using a measure. We focus on some notable measures that can be used. We
discuss the role of aggregation functions in the task of combining measures to form
new measures. We look at this in the framework of multi-source information fusion.
We focus on the fusion of probabilistic and possibilistic information and discuss
its role in hard-soft information fusion. We look at some characterizing features
associated with measures used to represent uncertain values of variables. We discuss
the concepts of assurance and opportunity that play a role in the process of answering
questions using information obtained from a measure.

1 Introduction

Intelligent decision-making must take advantage of various types of available infor-

mation, much of which has some degree of uncertainty. Some of the most important
sources of information are statistical data, physical sensors and human observa-
tion. Here we are interested in the problem of multi-source uncertain information
fusion [10]. We note that statistical data as well physical sensor provided information
generally have a probabilistic type of uncertainty whereas the linguistic information
provided by humans typically introduces a possibilistic type of uncertainty [19].
We note that many different structures have been suggested for the representation
of uncertain information [8, 9, 13]. In order to provide a unified framework for
the representation of different types of uncertain information we use a set measure
approach for the representation of uncertain information. We discuss a set measure
representation of uncertain information. Here we shall we focus on measures over
finite universes. In the multi-source fusion problem, we have a collection of pieces of
information that must be fused based on expert provided instructions on how to fuse
these pieces of information. Generally these instructions can involve a combination of

R. Yager (B)
Machine Intelligence Institute, Iona College, New Rochelle, NY 10801, USA
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 199

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_12
200 R.R. Yager

linguistically and mathematically expressed directions. We began to look at aggrega-

tion functions [1, 5] as a way for fusing this information. Some of the most important
sources of information are statistical data, physical sensors and human observation.
We note that statistical data as well physical sensor provided information generally
have a probabilistic type of uncertainty whereas the linguistic information provided
by humans typically introduces a possibilistic type of uncertainty [19]. The combi-
nation of these two types of information is referred to as hard-soft fusion [10] and
generally involves the fusion of probabilistic and possibilistic measures. Two charac-
terizing features of measures that are used to represent uncertain information are the
entropy and attitudinal character of a measure, we briefly investigate these concepts.
We discuss the ideas of assurance and opportunity [20] that play a fundamental role
in the task of answering questions using information represented with a measure.

2 Representing Uncertain Information Using Set Measures

A mapping μ : 2 X → [0, 1] is called a set function. We now introduce a special

type of set function called a fuzzy measure, it is also referred to as a monotonic set
measure [6, 12, 14].
Definition 1 A fuzzy measure on X is a mapping μ : 2 X → [0, 1] such that
1. μ(∅) = 0,
2. μ(X ) = 1,
3. μ(A) ≥ μ(B) if B ⊆ A.
We can easily see that for any subsets A and B we have

μ(A ∩ B) ≤ min(μ(A), μ(B)),

μ(A ∪ B) ≥ max(μ(A), μ(B)).

A set function we can associate with any measure is its dual. We define the dual of
μ as the set function μ̂ : X → [0, 1] defined such that μ̂(A) = 1 − μ(A) where A is
the complement of A. We can easily show that if μ is a measure on X then μ̂ is also
ˆ
a measure on X . We note that the dual of the dual is the original measure, μ̂(A) =
1 − μ̂(A) = 1 − (1 − μ(A)) = μ(A). Thus a measure and its dual are unique pairs.
We say an element x ∈ X is irrelevant with respect to μ if for all A we have
μ(A ∪ {x}) = μ(A). We observe that if x is irrelevant then μ({x}) = 0. If we define
E −x = X \ {x} and if x is irrelevant then μ(E −x ) = 1. We see this follows since
X = E −x ∪ {x} and hence μ(X ) = μ(E −x ∪ {x}) = μ(E −x ). We must emphasize
that μ({x}) = 0 does not necessarily make x irrelevant. We note that if x is irrelevant
to μ it is also irrelevant to its dual μ̂.
A fuzzy measure provides a very general structure for the representation of our
knowledge about an uncertain variable. Let V be a variable taking its value in the
space X . In using μ to express our knowledge about V we provide the following
Multi-source Information Fusion Using Measure Representations 201

interpretation. For any subset B of X we have that μ(B) indicates our anticipation
that the value of V lies in B. We see that μ(∅) = 0 reflects the fact that the value of
V will not be in the null set. The property μ(X ) = 1, called the normality condition,
indicates the fact that we are completely confident that the value of V lies in X .
Finally the monotonicity of μ reflects the fact that you cannot be more confident of
finding the value V in the set B than in a set that contains B. Here we say that the set
or event A happens if the value of V lies in A, thus μ(A) is seen as our anticipation
that the event A happens.
We shall use the expression V is μ to denote the situation where knowledge
about V is carried by the set measure μ. In the following we shall restrict our-
selves to the case where X is finite. Let us look at the representation of some types
of knowledge using this representation. Consider first the case of certainty where
we know that the value V is exactly x1 . We can express this using a fuzzy mea-
sure μ such that μ(B) = 1 if x1 ∈ B and μ(B) = 0 if x1 ∈ / B. We refer to this as
a Dirac measure focused at x1 . Consider now the case of probabilistic uncertainty

where we have Pr ob(xi ) = pi . In this case we define μ such that μ(B) = xi ∈B pi .
Here we see that μ({xi }) = pi . Thus we see that the probability measure has the
property of additivity. Thus the probability measure is a special fuzzy measure.
In this case we refer to μ(B) by the more specific name of probability n of B,
μ(B) = Pr ob(B). Since μ(X ) = Pr ob(X ) = 1 here we require i=1 pi = 1. For
this measure μ(A ∪ B) = μ(A) + μ(B) − μ(A ∩ B). We see here if A ∩ B = ∅
then μ(A ∪ B) = μ(A) + μ(B). We note that this measure requires only knowl-
edge of the pi to completely determine μ(A) for all A. A special case of this
is where all pi are the same, here pi = 1/n where n is the cardinality of X .
Another important type of uncertainty representable in this framework is possibilis-
tic uncertainty [4, 21]. We recall that usually this type of uncertainty is obtained
from a linguistic description of our knowledge. Here we associate with each xi a
value πi ∈ [0, 1] called the possibility of xi . In this case the associated measure
μ is called a possibility measure and is defined as μ(A) = maxxi ∈A πi . The con-
dition that μ(X ) = 1 requires that at least one πi = 1. We see for this measure
that μ(A ∪ B) = max(μ(A), μ(B)) for any A and B. We note that the Dirac mea-
sure is the only measure that is both a probability and possibility measure. Another
class of measures are decomposable measures. These generalize the possibility mea-
sure. Here if S is any t-conorm [7] then if μ({xi }) = ai we have μ(A) = Sxi ∈A (ai ).
A requirement here is that μ(X ) = Sxi ∈X (ai ) = 1. We note that for these mea-
sure if A ∩ B = ∅, then μ(A ∪ B) = S(μ(A), μ(B)). Since max is a t-conorm we
see that the possibility measure is an example of these decomposable measures.
Another notable type of measure is one in which μ(A ∩ B) = min(μ(A), μ(B))
these are referred to as certainty or necessity measures [2]. In this case we refer
to μ(A) as the necessity of A and use the notation N ec(A) instead of μ(A). We
note that if we let Fi = X \ {xi } then we can completely define this measure using
μ(Fi ). Here for any A, μ(A) = min xi ∈A (μ(Fi )). We shall refer to μ(Fi ) = βi . Since
μ(∅) = min x j ∈X (β j ) then we see that fact that μ(∅) = 0 requires at least one β j = 0.
The case where all β j are equal implies all β j = 0. We note that a necessity mea-
sure is the dual of a possibility measure. We also note that these necessity measures
202 R.R. Yager

are notable in that while for all measures μ(A ∩ B) ≤ min(μ(A), μ(B)), the neces-
sity measure attains this minimum, μ(A ∩ B) = min(μ(A), μ(B)). A generalization
of the necessity measure can be obtained by defining μ(A ∩ B) = T (μ(A), μ(B))
where T is a t-norm [7]. Here again all we need are n values for μ(Fi ) = βi . Another
important example of fuzzy measures are the cardinality based uncertainty measures.
Here our anticipation that the value of the variable lies in a particular subset just
depends on the number of elements in the subset. No distinction is made between
the individual outcomes elements. Let ai be a collection of parameters such that
0 = a0 ≤ a1 ≤ · · · ≤ an = 1. Let us denote |B| as the cardinality of the set B, the
number of elements in B. We now define μ(B) = a|B| . Here then for all subsets
with the same number of elements we have the same of anticipation of finding the
value in it. We can show for this measure that no elements are impossible. A useful
parameter associated with these measures is wi = ai − ai−1 which indicates the ben-
efit of adding the ith element to a set. Two important examples of cardinality based
uncertainty measures are the optimistic measure μ∗ and the pessimistic measure μ∗ .
For μ∗ we have μ∗ (A) = 1 for A = ∅. and μ∗ (∅) = 0 and for μ∗ (A) = 0 for A = X
and μ∗ (X ) = 1. For μ∗ we have a j = 1 for all j ≥ 1 and a0 = 0 and for μ∗ we have
a j = 0 for all j < n and an = 1. Another example of cardinality-based measure is
one where ai = i/n. We can use the wi to provide a measure of optimism associated
with a cardinality based measure μ on X [15]. If |X | = n then

n
(n − j)w j
O p(μ) = .
j=1
n−1

We see when μ is μ∗ that O p(μ∗ ) = 1 and when μ = μ∗ that O p(μ) = 0. If μ is

such that ai = i/n for all i, then O p(μ) = 0.5.

3 Measures Derived from Other Measures

Assume μ1 is some fuzzy measure on X and F : [0, 1] → [0, 1] is such that F(0) =
0, F(1) = 1, and F(a) ≥ F(b) if a > b then it can be shown that μ(A) = F(μ1 (A))
is a fuzzy measure [19].
Another class of derived measures are those composed from other measures.
Let μ1 and μ2 be two fuzzy measures on X . We can show that the set function μ
defined such that μ(A) = μ1 (A)μ2 (A) for all A ⊆ X is also a measure. We see that
μ(∅) = μ1 (∅)μ2 (∅), μ(X ) = μ1 (X )μ2 (X ) and if A ⊇ B then μ1 (A) ≥ μ1 (B) and
μ2 (A) ≥ μ2 (B) and hence μ(A) = μ1 (A)μ2 (A) ≥ μ1 (B)μ2 (B) = μ(B). Thus μ
as defined above is a fuzzy measure.
q This result can easily be extended to the fusion
of q fuzzy measure μ(A) = i=1 μi (A).
In the following we shall provide a generalization of this result, which shall form
the basis of our approach to the fusion of multi-source information.
Multi-source Information Fusion Using Measure Representations 203

Definition 2 An aggregation function G is a function of the integer q > 1 arguments

G : [0, 1]q → [0, 1] having the properties G(0, 0, . . . , 0) = 0, G(1, 1, . . . , 1) = 1
and G(a1 , . . . aq ) ≥ G(b1 , . . . , bq ) if all a j ≥ b j .

Theorem 1 Assume for j = 1 to q that μ j are a collection of fuzzy measures on X .

Then μ defined such that for all A ⊆ X

μ(A) = G(μ1 (A), . . . , μq (A))

is a fuzzy measure.
Proof 1. μ(∅) = G(μ1 (∅), . . . , μq (∅)) = G(0, 0, . . . , 0) = 0,
2. μ(X ) = G(μ1 (X ), . . . , μq (X )) = G(1, . . . , 1) = 1,
3. We have μ(A) = G(μ1 (A), . . . , μq (A)) and μ(B) = G(μ1 (B), . . . , μq (B)).
With B ⊆ A, since μ j (A) ≥ μ j (B) for all j, then μ(A) ≥ μ(B).
In situations where we have a finite space and the μ j are simple measures the
calculation of the aggregate measure is relatively easily. We shall refer to these mea-
sures as quasi-simple. Let X = {x1 , . . . , xq }. Assume we have q simple measures,
μ j for j = 1 to q. For j = 1 to q and i = l to n let αi j be the parameters associated
with these measures. Thus here μk (A) just depends on αik for xi ∈ A. Let G be
some aggregation function then with μ(A) = G(μ1 (A), . . . , μq (A)) the calculation
of μ(A) just depends on the values of μ j (A) which in turn just depends on the para-
meter αi j . The key point here is that for a measure defined using the aggregation of
simple measures, the calculation of the measure of a set is not complex.

4 On the Fusion of Information from Multiple Sources

In the problem of fusing information from multiple sources we have a collection of n

sources each of which is providing information about the variable V . Here we shall
assume each of these pieces of information can be expressed in terms of a fuzzy
measure. Thus our information is a collection V is μ j for j = 1 to q where each μ j
is a fuzzy measure defined on the domain X of V .
In addition we must have some expert provided instructions on how to fuse these
pieces of information so as to obtain a unified view of the value of V . The basis
of this expert provided knowledge can be very diverse. It can be based on a human
expert’s practical experience in processing multiple-sourced information. It can be
based on some formal data mining technology. Most generally these instructions can
involve a combination of linguistically and mathematical expressed directions. A
fundamental task in multi-source information fusion (MSIF) is the translation of these
instructions into formal operations that can be applied. The task of operationalizing
these expert provided instructions is generally a very complex one, it often involves a
tradeoff between precisely following the instructions and functionality, translating the
instruction into implementable operations. Here the capacity of Zadeh’s paradigm
204 R.R. Yager

of computing with words [22, 23] can become very useful for translating these
instructions into formal operations.
The type of aggregation operators previously introduced provides a very useful
tool for implementing a wide body of expert provided instructions for fusing mul-
tiple pieces of information. One of our interests here is to look at the use of these
aggregation operators for the fusion of information expressed via fuzzy measures.
We shall be particularly concerned with probability and possibility type information
as they represent two very important classes of provided information. We note that
possibilistic information often arises from a linguistic description of the value of
some variable. An example of this is information such as the house is close. Here
close is a linguistic term that can be expressed using fuzzy sets which in turn induces
a possibility distribution on the variable “the distance of the house to the river.”
Probabilistic information often appears because it provides an effective model to
represent the accuracy of physical sensing devices. It should be emphasized that the
use of probability is not a reflection of randomness in the variable of interest but a
reflection of the lack of accuracy of the measuring device.

5 Conjuncting Possibility and Probability Measures:

Hard-Soft Fusion

In the following we shall consider the conjunction “anding” fusion of probabilistic

and possibilistic information. The t-norm aggregation operator provides an aggrega-
tion implementation of an anding or conjunction of multiple pieces of information [1].
This use of “and” usually reflects an instruction to find a fused value that simultane-
ously satisfies all the multiple pieces of information. Assume we have two pieces of
information, V is μ1 and V is μ2 . Consider their fusion V is μ where V is μ = V is
μ1 and V is μ2 . Here μ is a fuzzy measure defined such that for any subset A of the
domain X of V we have
μ(A) = T (μ1 (A), μ2 (A))

where T is a t-norm.
Let us look at this for some notable cases of μ1 and μ2 . Consider first the case
where V is μ1 , indicates the value of V is exactly x1 and V is μ2 , indicates the value
of V is exactly x2 . Here μ1 has μ1 (A) = 1 for all A s.t. x1 ∈ A and μ1 (A) = 0 for
all A s.t. x1 ∈
/ A. μ2 has similar information, μ2 (A) = 1 if x2 ∈ A and μ2 (A) = 0 if
x2 ∈/ A. In this case we get as our fused value μ(A) = 1 if {x1 , x2 } ⊆ A and μ(A) = 0
otherwise. Here we have an anticipation of one of finding the value of V in any set
containing both x1 and x2 . It is interesting to note that in the case where we have
two sources giving very different values for the variable we provide a fused solution
that tried to unify these two conflicting values rather then reporting a conflict. This
is useful, because for some decisions the action taken when V is either x1 or x2 is
the same so a solution need not differentiate. In a more general situation, if we have
Multi-source Information Fusion Using Measure Representations 205

q sources each saying that V has the value x j then our fused function μ would be
such that μ(A) = 1 if {x1 , . . . , x j } ⊆ A and μ(A) = 0 otherwise. Again here we see
a kind of unification of the different pieces of information.
We can now consider the case where μ1 is as above but where μ2 is a possibility
distribution with μ2 ({x j }) = π j and μ2 (A) = maxx j ∈A π j . In this case we have for
any subset A with x1 ∈ A that μ(A) = maxx j ∈A π j and μ(A) = 0 if x1 ∈ / A. In
particular we note that μ({x1 }) = π1 and μ({x j }) = 0 for j = 1. We see that μ is
not a pure possibility distribution. We note that all the preceding results hold for any
t-norm since at least one of the arguments was always with 1 or 0.
We again consider the case of two pieces of information, V is μ1 and V is μ2
where μ1 indicates that V is exactly x1 while μ2 is any arbitrary measure. Here
we get for the fusion of the conjunction of these that V is μ where μ is such that
μ(A) = 0 if x1 ∈ / A and μ(A) = μ2 (A) if x1 ∈ A.
An interesting special case of this occurs when μ2 is a probability distribution
with pi the probability of xi . In this case we see that while μ({x1 }) = p1 we have
μ({x j }) = 0 for j = 1. Furthermore for any A such that  x1 ∈ / A we get μ(A) = 0
while for any A such that x1 ∈ A we have μ(A) = x j ∈A p j . We see μ is not
a probability measure if p1 < 1. As indicated earlier μ(A) is our anticipation of
finding a value V in the subset A. We observe that μ(A) = 0 or μ(A) ≥ p1 .
We now consider the conjunction in the case where μ1 is a probability distribution
and μ2 is a possibility distribution. In this case μ(A) = T (μ1 (A), μ2 (A)) where
μ1 (A) = x j ∈A p j and μ2 (A) = maxxi ∈A πi . We observe here that for any t-norm
T we get A = {x j } that μ({x j }) = T ( p j , π j ).
Let us look at this fusion of probabilistic and possibilistic information for some
notable examples of t-norms. We first consider the product t-norm. In this case
μ({x j }) = p j · π j and more generally
⎛ ⎞

μ(A) = μ1 (A) · μ2 (A) = ⎝ p j ⎠ · maxx j ∈A π j = Prob(A) · Poss(A).
x j ∈A

We further observe that the product t-norm leads to a nice formulation where
⎛ ⎞
 

μ(A) = ⎝ p j ⎠ · maxx j ∈A π j = p j · maxx j ∈A π j .

x j ∈A x j ∈A

Thus we see that each of the p j is multiplied by the maximum possibility in A.

Furthermore we can express this as
 

μ(A) = p j (π j + Δ j ) = pjπj + pjΔj

x j ∈A x j ∈A
206 R.R. Yager

where Δ j = π A∗ − π j with π A∗ = maxxk ∈A πk . With p j π j = μ({x j }) we see that we

can express 
μ(A) = (μ({x j }) + p j Δ j ).
x j ∈A

We see also if à = A ∪ {x j+1 } then

μ( Ã) = μ(A) + μ({x j+1 }) + p j+1 Δ j , if π j+1 ≤ π A∗

μ( Ã) = μ(A) + μ({x j+1 }) + π j+1 − π A∗ , if π j+1 > π A∗ .

We further observe that if A has one x j such that π j = 1 then μ( Ã) = μ(A) + p j+1 .

Example 1 Assume X = {x1 , x2 , x3 } and let μ1 be a probability distribution such

that p1 = 0.7, p2 = 0.2 and p3 = 0.1. Let μ2 be a possibility distribution such that
π1 = 0.3, π2 = 1 and π3 = 0.8. In this case we have for any subset A of X that
μ(A) = μ1 (A) · μ2 (A) = Prob(A) · Poss(A):

μ({x1 }) = 0.7 · 0.3 = 0.21 μ({x2 }) = 0.2 · 1 = 0.2

μ({x3 }) = 0.1 · 0.8 = 0.8 μ({x1 , x2 }) = 0.9 · 1 = 0.9
μ({x1 , x3 }) = 0.8 · 0.8 = 0.64 μ({x2 , x3 }) = 0.3 · 1 = 0.3
μ({x1 , x2 , x3 }) = 1 · 1 = 1

We see here that we have introduced a new class of measures, which we shall refer
to as P2 measures. Let us understand this measure. This measure is characterized
by two sets of parameters
n P = [ p1 , . . . , pn ] and  = [π1 , . . 
. , πn ] where all pi and
πi ∈ [0, 1] and i=1 pi = 1 and max j π j = 1. Here μ(A) = x j ∈A pi · maxx j ∈A π j
and in particular μ({x j }) = π j p j .
We note that this approach can be easily generalized to the case where we have
more than two sources of information some of which are possibilistic and some
probabilistic. Thus here if V is μi are a collection of q pieces of information q about V
then their fusion under the preceding “anding” imperative is μ(A) = i=1 μi (A).

6 Characterizing Features: Entropy and Attitude

Here we now describe some tools useful in the characterization of fuzzy measures
that are used to model uncertain variables. These are particularly useful in compar-
ing fuzzy measures with respect to properties such as their overall uncertainty. We
refer the reader to [16–18] more details on these. In probability theory an impor-
tant tool is the Shannon entropy. With this tool we are able to indicate the overall
uncertainty of a probability distribution. Assume we have a probability distribu-
tion on the space X = {x1 , . . . , xn } where p j is the probability of x j . Here p j is
Multi-source Information Fusion Using Measure Representations 207

capturing the distinction in our belief about each of the elements, p j is the support
for x j as the outcome.
 We recall the Shannon entropy associated with this distribution
is H (P) = − j p j ln( p j ). The Shannon entropy quantifies the overall uncertainty
associated with the probability distribution. In [16] we extended this idea to fuzzy
measures. In order to extend this to a general measure μ we first introduce the idea
of the Shapley index [11]. For any x j ∈ X we define its Shapley index S j as


n−1 
Sj = (rk (μ(K ∪ {x j }) − μ(K ))).
k=0 K ⊂F j
|K |=k

In the above K is a subset of cardinality |K |, F j = X \ {x j } and rk = (n−k−1)!k! n!

. It
can
n be shown that for any fuzzy measure μ it is always the case that S j ∈ [0, 1] and
j=1 S j = 1.
In some sense the S j are reflecting the information about the distinction between
the anticipation of each of x j . As a matter of fact it can be that if μ is a probability
measure then S j = p j . Thus the Shapley index is capturing information about the
distinction in the elements.
Using these Shapley values in [16] we extended  the idea of entropy to a measure
μ by defining the entropy of μ as H (μ) = − j S j ln(S j ). Since Si = pi for a
probability measure this definition is compatible with the classic entropy. We note
that if μ is a possibility measure and the xi are indexed such that μ({x j }) ≥ μ({xi })
j
for j > i then S j = i=1 μ({xi })−μ({x
n+1−i
i−1 })
.
A special interesting case is that of the cardinality based measure. For this measure
it can be shown that S j = 1/n for all x j . Here then there is no distinction between
the x j and the entropy is the largest, H (μ) = ln(n).
If μ is a measure with Shapley indices S j and Shapley entropy
H (μ) = − j S j ln(S j ) then if μ̂ is the dual of μ we now show that its Shapley
indices Sˆj = S j and hence H (μ̂) = H (μ), duals have the same entropy. For the dual
we have
n−1 
Sˆj = (rk (μ̂(K ∪ {x j }) − μ̂(K )))
k=0 k∈F j ,
|K |=k

where F j = X \ {x j }. Since μ̂(A) = 1 − μ(A) then

μ̂(K ∪ {x j }) − μ̂(K ) = (μ(K ) − μ(K ∪ {x j })).

We observe that K ∪ {x j } = K \ {x j } and hence we get

μ̂(K ∪ {x j }) − μ̂(K ) = μ(K ) − μ(K \ {x j })

and therefore Sˆj = S j .

208 R.R. Yager

Now we consider another characterizing feature of a fuzzy measure used to convey

information about an uncertain variable which is called the attitudinal character of
the fuzzy measure [18]. Consider the two cardinality based measures μ∗ and μ∗ .
We recall μ∗ (A) = 1 for all A = ∅ and μ∗ (∅) = 0 while μ∗ (A) = 0 for all A = X
and μ(X ) = 1. As these are both cardinality-based measures they both have no
information distinguishing the elements in X regarding their being the value of V .
We recall that μ(A) is the anticipation of finding the value of V in A. We see that
these two measures are dealing with our lack of information in very different ways,
μ∗ deals with the complete lack of knowledge in a very optimistic way, it anticipates
finding the value of V in any non-null subset. On the other hand μ∗ deals with the
complete lack of knowledge in a very pessimistic way it doesnt anticipate finding
the value of V in any subset except X . We see that these two measures display polar
attitudes about the anticipation of finding V when faced with lack of information.
Another cardinality based measure, μ(A) = card(A)
n
, falls between these two extremes
and is more neutral.
The preceding fuzzy measures have illustrated different attitudes about the nature
of uncertainty. In the following we introduce a characterization of a fuzzy measure
that allows us to quantify these differing attitudes. We begin by introducing the
cardinality index of a fuzzy measure [18].
Definition 3 Let μ be a measure on X where the cardinality of X , |X | = n. For
k = 0 to n − 1 we define Ck as
 
 
C k = λk (μ(K ∪ {x}) − μ(K ))
all K x ∈K
/
|K |=k

(n−k−1)!k!
where λk = n!
. We call Ck the kth cardinality index.
We see that Ck measures the average gain in anticipation we get in going from a
subset
of cardinality k to one of cardinality k + 1. In [18] it was shown that Ck ∈ [0, 1]
n−1
and k=0 Ck = 1.
Yager [18] used the cardinality index to define a characteristic of a fuzzy measure
called its attitudinal character. The attitudinal character of a measure m is defined as

1 
n−1
AC(μ) = (n + k + 1)Ck .
n − 1 k=0

n−1
We note that since k=0 Ck = 1 then we can express the attitudinal character as
n−1
kCk
AC(μ) = 1 − k=0
.
n−1

In [18] it was shown that AC(μ) ∈ [0, 1]. Furthermore the large values of AC
indicate a more optimistic type of measure while the small values indicate a more
Multi-source Information Fusion Using Measure Representations 209

pessimistic type of measure, AC(μ) = 1 being the most optimistic and AC(μ) =
0 being the most pessimistic. Thus the attitudinal character is providing a scale
characterizing a measures degree of optimism/pessimism about its anticipation of
finding the variables value in a given subset.
In [18] we obtained the cardinality index for cardinality-based measures and
probabilistic type measures. We recall for a cardinality based measure μ we have a
set 0 ≤ a0 ≤ ai ≤ an = 1 such that μ(E) = a|E| , where |E| is the cardinality of E.
We showed that for this type of measure the cardinality index is

Ck = ak+1 − ak for k = 0 to n − 1.

A few notable examples of this are worth pointing out here. In the case of μ∗ where
μ∗ (E) = 1 for all E = ∅ and μ∗ (∅) = 0 it was shown that C0 = 1 and Ck = 0 for
all k = 0. In the case μ∗ where μ∗ (E) = 0 for E = X and μ(X ) = 1 then Cn−1 = 1
and Ck = 0 for other k. In the case where a j = nj then Ck = n1 for all k = 0 to n − 1.
Another situation studied in [18] was the additive or probabilistic uncertainty.
Here we associate with each xi a value pi ∈ [0, 1] so that they sum to one. In [18] it
was shown that independent of the value of the pi the cardinality index for all k is
always Ck = n1 . Thus the cardinality index doesn’t distinguish between probability
distributions, all probability distributions have the same cardinalityindex.
n−1
kC
Let us now consider the attitudinal character, AC(μ) = 1 − k=0 n−1
k
, of these
∗
measures. First we consider μ , here C0 = 1 and all other Ck = 0 and hence
AC(μ∗ ) = 1. This is as expected since μ∗ is the most optimistic measure. On the
other hand for μ∗ since Cn−1 = 1 and all other Ck = 0 we have AC(μ∗ ) = 0. Again
this is as expected since μ∗ is very pessimistic. Consider now the probabilistic case
where Ck = n1 for all k. Here AC(μ) = 0.5, this has a neutral value of 0.5. This is
also the case of the cardinality based measure with a j = nj .
We shall investigate the relationship between the cardinality index of a measure
μ and the cardinality index of its dual μ̂. We recall for μ,
 
C k = λk ( (μ(K ∪ {x}) − μ(K ))),
all K x ∈K
/
|K |=k

it is the average increase in anticipation in going for sets of cardinality k to k + 1. If

μ̂ is the dual then
 
Ĉk = λk ( (μ̂(K ∪ {x}) − μ̂(K ))).
all K x ∈K
/
|K |=k

Since μ̂ is the dual of μ then μ̂(E) = 1 − μ(E). In this case we see that μ̂(K ) =
1 − μ(K ) and μ̂(K ∪ {x}) = 1 − μ(K ∪ {x}) and there then
 
(μ̂(K ∪ {x}) − μ̂(K )) = (μ(K ) − μ(K ∪ {x})).
x ∈K
/ x ∈K
/
210 R.R. Yager

We note that K = X \ K and K ∪ {x} = X \ (K ∪ {x}) = K \ {x}. Thus we get that

μ̂(K ∪ {x}) − μ̂(K ) = μ(K ) − μ(K \ {x}).

If |K | = k then |K | = n − k and |K \ {x}| = n − k − 1. Here then we see that Ĉk is

the average increase in anticipation of μ in going from sets of cardinality n − k − 1
to n − k which is what we denoted as Cn−k−1 . Thus we see that Ĉk = Ck−n−1 . Using
this we can be relate the attitudinal characters of μ and μ̂.
The following important theorem relates the attitudinal characters of dual
measures.
Theorem 2 AC(μ̂) = 1 − AC(μ).

7 Assurance and Opportunity of Outcomes

Here we shall discuss the concepts of assurance and opportunity which play an
important role in the process of answering questions using information obtained
from the fusion of measures. Consider now we have a variable V with domain X .
Assume our knowledge about the value of V is expressed using a measure μ on X .
Thus for any subset A of X the value μ(A) indicates our anticipation of A occurring,
that is of finding the value of V in the set A. Does a value μ(A) = 1 assure us that
A will occur? Let us look at this in more detail. Consider the measure μ∗ (A) = 1
for all A = ∅. In the case we have μ∗ (A) = 1. However in this case we also have
μ∗ (A) = 1 so here we have just as strong an anticipation that A will not occur. In
order for us to be assured that A will occur we have to anticipate A will occur and
also anticipate that A will not occur. The degree to which our anticipation that A will
not occur can be measured by 1 − μ(A). However we note that this is the dual of
μ, μ̂(A). Using this we introduce a set function called the assurance of A which we
define as λ(A) = μ(A) ∧ μ̂(A) [20]. We easily see that λ is a measure. One thing
we note about this measure of assurance is that λ(A) ≤ μ(A). Thus the measure of
assurance of a set is never larger than its measure of anticipation. We also observe
that λ(A) ≤ μ̂(A).
Consider the two measures μ and μ̂ its dual. We have λμ (A) = μ(A) ∧ μ̂(A) and
ˆ
λμ̂ (A) = μ̂(A) ∧ μ̂(A). ˆ
However since μ̂(A) = μ(A) we have for measures that are
duals we have λμ (A) = λμ̂ (A). However it is not the case that μ̂(A) = μ(A).
Closely related to the concept of assurance is the concept of opportunity. We define
this as the set function Ψ where Ψ (A) = μ(A) ∨ μ̂(A) and we refer to Ψ (A) as the
opportunity associated with A. It essentially measures the opportunity of V lying in
A. It is clear this Ψ (A) ≥ λ(A). We can easily show that Ψ is itself a measure. We note
that Ψ (A) ≥ μ(A) and Ψ (A) ≥ μ̂(A). As is the case for the measure of assurance
we see that if μ and μ̂ are duals then Ψμ (A) = Ψμ̂ (A). We indicate that Ψ (A) is seen
as the opportunity that the value of V will lie in A. We note that Ψ (A) can be seen
as an optimistic view of the occurrence of A and λ(A) as a pessimistic view.
Multi-source Information Fusion Using Measure Representations 211

Actually we see a duality relationship between the measures Ψ and λ. Consider

the dual of Ψ (A), (Ψˆ )(A) = λ(A).). Thus the concepts of opportunity and assurance
are duals. However there exists a very important special relationship between these
duals, Ψ (A) ≥ λ(A). In the following we shall look at Ψ and λ for some special
cases of μ.
As we previously noted the same piece of information can sometimes be expressed
in different ways using a measure representation. Consider the situation in which I
know nothing about the value of the variable V other than it must be in the set X .
There exists two contrary ways I can express this information. One is optimistic with
μ∗ and the other pessimistically with μ∗ . The optimistic way has μ∗ (A) = 1 for all
A = ∅ and only μ∗ (∅) = 0. The pessimistic method has μ∗ (A) = 0 for A = X and
μ∗ (X ) = 1. We also emphasize that μ∗ and μ∗ are duals. We see μ̂∗ (X ) = 1 − μ∗ (A)
however since μ∗ (B) = 1 for any B = ∅, then μ̂∗ (A) = 0 for all A = X which is
μ∗ . As we shall subsequently see that use of λ and Ψ provides unification whereby
we have the same information for both.
Consider first μ∗ where μ∗ (A) = 1 for all A = ∅. In the case μ̂∗ (A) =
1 − μ∗ (A) = 0 for all A = X . Here then while λ∗ (X ) = 1 we see that for all A = X

λ∗ (A) = μ∗ (A) ∧ μ̂∗ (A) = 0.

Thus in this case while μ∗ (A) = 1 for all A = ∅ we have λ∗ (A) = 0 for all A = X .
Thus this case of complete lack of information about the value of V allows for no
assurance about anything except that V lies in its domain X .
Consider now μ∗ , we see μ∗ (A) = 0 for all A = X and hence λ∗ (A) = 0 for
A = X . Thus here again in this case of complete lack of knowledge about a value
of V we get λ(A) = 0. Thus here while μ∗ (A) and μ∗ (A) have different values for
most A we have λ∗ (A) = λ∗ (A) for all A.
We further observe since Ψ is the dual of λ then in the cases of μ∗ and μ∗ we have
that Ψ ∗ (A) = Ψ∗ (A) and furthermore there are such that Ψ ∗ (A) = Ψ∗ (A) = 1 for
all A = ∅. Thus here while we have no assurance that V lies in A we have complete
opportunity for it to lie in A.
Thus these measures μ∗ and μ∗ have the same values for λ and Ψ . Since these
both correspond to lack of any information we can refer to these as λ? and Ψ? . In the
preceding we have shown Ψ? (A) = 1 for all A = ∅ and ψ? (∅) = 0 and λ? (A) = 0
for all A = X . Here we see except for the extremes of X and ∅ the values of Ψ? (A)
and λ? (A) are as far apart as they can possibly be. Here we see for each A we have
an opportunity but no assurance for finding V .
Let us now consider nμ is a probability measure. We recall her
the case where
μ({xi }) = pi , μ(A) = xi ∈A pi and i=1 pi = 1. Consider now the dual of this
measure. In this case
 
μ̂(A) = 1 − μ(A) = 1 − pi = pi = μ(A).
xi ∈A xi ∈A
212 R.R. Yager

Thus in the case of a probability measure the measure and its dual are the same.
We furthermore see it is also equal to the assurance and opportunity measures

λ(A) = μ̂(A) ∧ μ(A) = μ(A) ∧ μ(A) = μ(A),

Ψ (A) = μ̂(A) ∨ μ(A) = μ(A) ∨ μ(A) = μ(A).

Then here we Ψ (A) = λ(A) = μ(A). We note that this fact holds for any self-dual
fuzzy measure, some times called participation measures.
This is an important and very special property of the probability distribution.
There is no difference between any of the measures. The measures of opportunity,
assurance and anticipation are all the same. In the situation μ(A) is often referred to
as the probability of A.
An important special case of probability measure is one in which all p j = n1 .
In this case μ(A) = card(A)
n
. In some situations this special case has been used to
represent lack of knowledge. However we emphasize this formulation implies some
knowledge about V . In particular, it at least assumes that the measure of all the
elements μ({x j }), are the same.
Consider now the measure μ in the case where we know that V = x ∗ . Here
we recall μ(B) = 1 if x ∗ ∈ B and μ(B) = 0 for x ∗ ∈ / B. Consider the dual of this
μ̂(B) = 1 − μ(B). We see μ(B) = 1 if x ∗ ∈ / B that is if x ∗ ∈
/ B. Also we see that
μ(B) = 0 if x ∗ ∈/ B that is if x ∗ ∈ B. Thus μ̂(B) = 1 if x ∗ ∈ B. We see that in this
case μ̂ and μare the same and hence

λ(A) = μ̂(A) ∧ μ(A) = μ(A) = μ̂(A) ∨ μ(A) = Ψ (A).

Thus if x ∗ ∈ A we get λ(A) = 1 and x ∈ / A we get λ(A) = 0. Actually this is a

special probability measure where P(x ∗ ) = 1 and all other are zero, it is a Dirac
measure.
We now consider the possibility measure. We recall for these measures μ({xi }) =
πi and μ(A ∪ B) = max(μ(A), μ(B)). Here μ(A) = maxxi ∈A (πi ). We note that at
least one πi = 1. The dual of a possibility measure μ̂(A) is defined as μ̂(A) =
1 − μ(A) and hence we get

μ̂(A) = 1 − maxxi ∈A (πi ).

The duals of possibility measures have a very interesting property,

μ̂(A ∩ B) = min(μ̂(A), μ̂(B)).

Measures having this property are referred to as certainty measures [2]. Since μ
and μ̂ are unique pairs we refer to μ̂ as the associated certainty measure. Certainty
measures are also referred to as necessity measures
A very important and special relationship exists between a possibility measure
and its dual certainty measure. If μ is a possibility measure then for any set A we
Multi-source Information Fusion Using Measure Representations 213

have μ(A) ≥ μ̂(A). Let us see that this is true. Here we have μ(A) = maxxi ∈A (πi )
and μ̂(A) = 1 − maxxi ∈A (πi ). We note that at least one of the πi = 1. If A has one
of the xi with πi = 1 then μ(A) = 1 and μ(A) ≥ μ̂(A). If there exists no element
with πi = 1 contained in A then at least one such element must be in A and hence
maxxi ∈A (πi ) = 1 and therefore μ̂(A) = 0 and again we have μ(A) ≥ μ̂(A). An
important implication of this is that for possibility measures we always have

λ(A) = μ̂(A) ∧ μ(A) = μ̂(A).

Thus here the measure of assurance of A is equal to the dual measure of A, the
certainty of A. Furthermore in this case of possibility measures we always have

Ψ (A) = μ̂(A) ∨ μ(A) = μ(A).

Thus here Ψ (A) is always the measure of possibility.

We want to emphasize that not all measures have this clear relationship between
μ(A) and μ̂(A). Thus here in the case of a possibility measure since λ(A) = μ̂(A)
then Ψ (A) = μ(A) ≥ μ̂(A) = λ(A). This relationship is greatly used in possibility
theory [3].

8 Conclusion

We first looked at the issue of representing information about an uncertain variable

using a measure. We focused on some notable measures that can be used. We dis-
cussed the role of aggregation functions in the task of combining measures to form,
new measures. We looked at this in the framework of multi-source information fusion.
We focused on the fusion of probabilistic and possibilistic information and discussed
its role in hard-soft information fusion. We looked at some characterizing features
associated with measures used to represent uncertain values of variables. We dis-
cussed the concepts of assurance and opportunity that play a role in the process of
answering questions using information obtained from a measure.

References

1. Beliakov, G., Pradera, A., Calvo, T.: Aggregation Functions: A Guide for Practitioners.
Springer, Heidelberg (2007)
2. Dubois, D., Prade, H.: Necessity measures and the resolution principle. IEEE Trans. Syst. Man
Cybern. 17, 474–478 (1987)
3. Dubois, D., Prade, H.: Possibility Theory: An Approach to Computerized Processing of Uncer-
tainty. Plenum Press, New York (1988)
214 R.R. Yager

4. Dubois, D., Prade, H.: Formal representations of uncertainty. In: Bouyssou, D., Dubois, D.,
Pirlot, M., Prade, H. (eds.) Decision Making Process: Concepts and Methods. Wiley, Hoboken
(2010)
5. Grabisch, M., Marichal, J.-L., Mesiar, R., Pap, E.: Aggregation Functions. Cambridge Univer-
sity Press, Cambridge (2009)
6. Klement, E.P.: A theory of fuzzy measures: a survey. In: Gupta, M.M., Sanchez, E. (eds.) Fuzzy
Information and Decision Processes, pp. 59–66. North-Holland, Amsterdam (1982)
7. Klement, E.P., Mesiar, R., Pap, E.: Triangular Norms. Kluwer Academic Publishers, Dordrecht
(2000)
8. Klir, G.J.: Uncertainty and Information. Wiley, New York (2006)
9. Liu, L., Yager, R.R.: Classic works of the Dempster-Shafer theory of belief functions: an
introduction. In: Yager, R.R., Liu, L. (eds.) Classic Works of the Dempster-Shafer Theory of
Belief Functions, pp. 1–34. Springer, Heidelberg (2008)
10. Llinas, J., Nagi, R., Hall, D.L., Lavery, J.: A multi-disciplinary university research initiative in
hard and soft information fusion: Overview, research strategies and initial results. In: Proceed-
ings of the 13th International Conference on Information Fusion (Fusion 2010). Edinburgh,
UK, Unpaginated (2010)
11. Shapley, L.S.: A value for n-person games. In: Kuhn, H.W., Tucker, A.W. (eds.) Contributions
to Game Theory, pp. 307–317. Princeton University Press, Princeton (1953)
12. Sugeno, M.: Fuzzy measures and fuzzy integrals: a survey. In: Gupta, M.M., Saridis, G.N.,
Gaines, B.R. (eds.) Fuzzy Automata and Decision Process, pp. 89–102. North-Holland, Ams-
terdam (1997)
13. Walley, P., Fine, T.: Toward a frequentist theory of upper and lower probability. Ann. Stat. 10,
741–761 (1982)
14. Wang, Z., Yang, R., Leung, K.S.: Nonlinear Integrals and Their Applications in Data Mining.
World Scientific, Singapore (2010)
15. Yager, R.R.: On ordered weighted averaging aggregation operators in multi-criteria decision
making. IEEE Trans. Syst. Man Cybern. 18, 183–190 (1988)
16. Yager, R.R.: On the entropy of fuzzy measures. IEEE Trans. Fuzzy Sets Syst. 8, 453–461
(2000)
17. Yager, R.R.: Measuring the information and character of a fuzzy measure. In: Proceedings
of the Joint 9th IFSA World Congress and the 20th NAFIPS International Conference, pp.
1718–1722. Vancouver (2001)
18. Yager, R.R.: On the cardinality index and attitudinal character of fuzzy measures. Int. J. Gen.
Syst. 31, 303–329 (2002)
19. Yager, R.R.: A measure based approach to the fusion of possibilistic and probabilistic uncer-
tainty. Fuzzy Optim. Decis. Making 10, 91–113 (2011)
20. Yager, R.R.: Measures of assurance and opportunity in modeling uncertain information. Int. J.
Intell. Syst. 27, 776–796 (2012)
21. Zadeh, L.A.: Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Syst. 1, 3–28 (1978)
22. Zadeh, L.A.: From computing with numbers to computing with words–From manipulation of
measurements to manipulations of perceptions. IEEE Trans. Circuits Syst. 45, 105–119 (1999)
23. Zadeh, L.A.: Generalized theory of uncertainty (GTU)-principal concepts and ideas. Comput.
Stat. Data Anal. 51, 15–46 (2006)
Bases and Transforms of Set Functions

Michel Grabisch

Abstract The chapter studies the vector space of set functions on a finite set X , which
can be alternatively seen as pseudo-Boolean functions, and including as a special
cases games. We present several bases (unanimity games, Walsh and parity functions)
and make an emphasis on the Fourier transform. Then we establish the basic duality
between bases and invertible linear transform (e.g., the Möbius transform, the Fourier
transform and interaction transforms). We apply it to solve the well-known inverse
problem in cooperative game theory (find all games with same Shapley value), and
to find various equivalent expressions of the Choquet integral.

1 Introduction

Set functions on a finite set X are of fundamental usage in many areas of discrete
mathematics, e.g., cooperative game theory [20], combinatorial optimization [11],
decision making [14], computer sciences [6], and more generally operations research
[16], where in the latter domain, they are more often encountered under the form
of pseudo-Boolean functions. Specific domains focus on specific subclasses of set
functions, e.g., game theory uses set functions vanishing on the empty set (these
are characteristic functions of transferable utility games, which are simply called
“games”), decision theory needs games which are monotone with respect to inclusion
(called capacities), while combinatorial optimization often deals with submodular
games.
An interesting feature of set functions and games is that they form a vector space
of dimension 2|X | (2|X |−1 for games). Most often, this feature is ignored, although
clearly one can take advantage of the concepts and techniques of linear algebra
when dealing with set functions and games. In particular the notion of basis is of
importance. The best-known basis is perhaps the basis of unanimity games (this is the

M. Grabisch (B)
Paris School of Economics, University of Paris I, 106-112, Bd de L’Hôpital,
75013 Paris, France
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 215

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_13
216 M. Grabisch

usual name given in game theory; they are closely related to the incidence functions in
combinatorics, see [1]), although each domain has its prefered bases. For example, in
computer sciences mostly the basis of parity functions is used, essentially because the
encoding of sets is done by −1, +1, rather than by 0, 1. The usage of a particular basis
induces a particular representation of set functions, viewed as a transform, which is
by definition linear and invertible. For example, the representation through the basis
of unanimity games is the Möbius transform (or Möbius inverse), widely used in
combinatorics and well known in decision making and game theory (known under
the name of Harsanyi dividends [17] in the latter domain), while the representation
through parity functions is the Fourier transform, which has many applications in
computer sciences.
As far as the author can see from the literature, the duality between bases and
transforms (i.e., the representation of set functions into a given basis) has never been
exploited nor even remarked. A systematic exploitation of this fact can lead to the
discovery of new bases and transforms, as well as an easy solution to the so called
inverse problem in game theory: find all games having the same Shapley value (and
similar ones). Also, it permits to get several different expressions of linear operators
on games or set functions, like the Choquet integral.
The aim of this chapter is to bring a survey on the above mentioned issues.
Section 2 gives a brief account on set functions and pseudo-Boolean functions.
Section 3 describes the best-known bases (unanimity games, Walsh functions and
parity functions). Section 4 is about the Fourier transform and its properties, while
Sect. 5 explains the fundamental duality between bases and transforms, and gives
many examples. Sections 6 and 7 are applications of this duality principle to the
solution of the above mentioned inverse problem and to the finding of equivalent
expressions of the Choquet integral.

2 Set Functions and Pseudo-Boolean Functions

In the whole chapter, we consider a finite universe X , with |X | = n. Occasionally,

we will use the notation [n] = {1, . . . , n}.
A set function on X is a mapping ξ : 2 X → R. A game is a set function v vanishing
on the empty set: v(∅) = 0.
X
Clearly, the set R2 of set functions on X is a 2n -dimensional vector space. We
X
introduce on R2 the following scalar product:

1 
ξ, ξ   = ξ(S)ξ  (S).
2n S⊆X

There is another vision of set functions, namely the pseudo-Boolean functions

[16], noting that any subset A of X can be encoded by its characteristic function 1 A .
Formally, a pseudo-Boolean function is a mapping f : {0, 1}n → R. The equivalence
Bases and Transforms of Set Functions 217

between pseudo-Boolean functions and set functions can be seen through the coding
function 1 : 2 X → {0, 1}n defined by A → 1 A , with 1 A (i) = 1 if and only if i ∈ A.
Then
ξ f = f ◦ 1, f ξ = ξ ◦ 1−1

where ξ f denotes the set function associated to f , and f ξ is the pseudo-Boolean

function associated to ξ. If follows that the set of pseudo-Boolean functions of n
variables is a 2n -dimensional vector space, with scalar product

1 
 f, f   = f (x) f  (x).
2n x∈{0,1}n

Pseudo-Boolean functions were obtained from set functions by coding subsets by

1 and 0. It is noteworhty that other encodings are possible, for example using −1, 1
instead of 0, 1. We will come back to this encoding later and shows that it is quite
useful.

3 Bases of Set Functions and Pseudo-Boolean Functions

3.1 Unanimity Games

X
Perhaps the best known basis of R2 is the basis of the so-called unanimity games.
For any nonempty subset S ⊆ X , the unanimity game centered on S is the game
defined by 
1, if T ⊇ S
u S (T ) = .
0, otherwise

Defining the set function u ∅ (S) = 1 for every S ⊆ X , it is well known that {u S } S∈2 X
X
is a basis for R2 . It is also well known that the coordinates of ξ in this basis are the
Möbius transform coefficients:

ξ= m ξ (S)u S
S∈2 X

with 
m ξ (S) = (−1)|S\T | ξ(T ).
T ⊆S

The Möbius transform (or Möbius inverse) is a well-known tool in combinatorics

since the work of [21] (see also [1, 3], etc.).
218 M. Grabisch

A drawback of the basis of unanimity games is that it is not orthogonal w.r.t. the
above scalar product, as it is easy to see even with n = 2:

1
u {1} , u {2}  = (u {1} ({1})u {2} ({1}) + u {1} ({2})u {2} ({2})
4
1
+ u {1} ({1, 2})u {2} ({1, 2}) = = 0.
4
In the formalism
 of pseudo-Boolean functions, unanimity games u S correspond
to monomials i∈S xi . Hence we have
 
f = m f (S) xi
S⊆[n] i∈S

where m f is the Möbius transform of ξ f (this slight abuse of notation should not be
confusing).

3.2 Walsh Functions

Another basis of pseudo-Boolean functions is the basis of Walsh functions, which

are monomials defined by

wT (x) = (2xi − 1) (T ⊆ [n], x ∈ {0, 1}n )
i∈T

or, in set function notation:

wT (S) = (−1)|T \S| (S, T ∈ 2 X ). (1)

It can be shown that the Walsh functions {wT }T ⊆[n] form an orthornomal basis of the
pseudo-Boolean functions:

w S , wT  = 1 iff S = T, and 0 otherwise.

It is important to note thatletting z i = 2xi − 1 ∈ {−1, +1}, the Walsh functions

reduce to the monomials i∈T z i : hence the Walsh functions are obtained when
subsets are encoded by −1, 1 instead of 0, 1. We see that this simple change makes
the basis orthonormal. It can be shown that the coordinates of a pseudo-Boolean
function into this basis are given by

   m f (S)   1 f
f (x) = w T (x) = I (T )wT (x) (2)
T ⊆[n] S⊇T
2|S| T ⊆[n]
2|T | B
Bases and Transforms of Set Functions 219

f
where IB is the Banzhaf interaction transform defined in terms of the Möbius trans-
form by
f
  1 |S\T |
IB (T ) = m f (S) (T ∈ 2 X )
S⊇T
2

(this transform will be properly introduced later).

As a historical remark, we note that the original Walsh functions [26] are rather
different in their definition:
∞
Wk (x) = (−1) j=0 k j x j+1
(k ∈ N0 , x ∈ [0, 1]), (3)

with k = k0 + k1 2 + k2 22 + · · · km 2m , ki ∈ {0, 1} for all i, and x = x1 2−1 + x2 2−2 +

x3 2−3 + · · · , xi ∈ {0, 1} for all i, the binary representations of k and x. They form an
orthonormal basis of the set of square integrable functions on [0, 1]. The connection
with our Walsh functions is that the latter have a discretized domain
1 2 3 2n − 1
0, n
, n, n,...,
2 2 2 2n
of 2n points, corresponding to the 2n subsets of [n]. More precisely, w S (x) corre-
sponds to Wk (x  ) such that S and k have same binary coding, and

x1 = 1 − x1 , . . . , xn = 1 − xn , and x j = 0 for j > n.

3.3 Parity Functions

Another family of functions related to the Walsh functions are the parity functions.
The parity function associated to S ⊆ [n] is the function

χ S (x) = (−1)1S ·x = (−1) i∈S xi
(x ∈ {0, 1}n ). (4)

Its name comes from the fact that it takes only values −1 and +1, depending on
whether there is an odd or even number of elements of coordinates of x equal to 1
in S. It expression as a set function is

χ S (T ) = (−1)|S∩T | (S, T ∈ 2 X ). (5)

Up to a recoding by ε(1) = 0 and ε(−1) = 1, the parity functions are the Walsh
functions:
 
w S (z) = z i = (−1) i∈S ε(zi ) = χ S (ε(z)) (z ∈ {−1, 1}n ).
i∈S
220 M. Grabisch

Consequently, they form another orthonormal basis of the vector space of

pseudo-Boolean functions. The interest of parity functions is that they lead to the
well-known Fourier transform, to which the next section is devoted.

4 The Fourier Transform

In the basis of parity functions, it can be shown that any pseudo-Boolean function f
is expressed by 
f =
f (S)χ S
S⊆[n]

where the coordinates of f in this basis, denoted by

f (S), are given by

1 
f (S) =  f, χ S  = n (−1)1S ·x f (x) (S ⊆ [n])
2 x∈{0,1}n

or, in terms of set functions,

1 
ξ(S) = n (−1)|S∩T | ξ(T ) (S ⊆ [n]). (6)
2 T ⊆[n]

The set of coefficients

f (S), S ⊆ [n], is called the Fourier transform of f , and is
widely used in computer sciences (see, e.g., a survey in [6], as well as [19]).
We show now some properties of the Fourier transform, and to this end, we
introduce some additional notions and notation. We may consider x ∈ {0, 1}n as a
random variable with uniform distribution. Then, the expected value and the variance
of a pseudo-Boolean function f are

1 
E[ f ] = f (x)
2n x∈{0,1}n


Var[ f ] = E ( f − E[ f ])2 = E[ f 2 ] − E2 [ f ].

The convolution product of two pseudo-Boolean functions f, g is defined by

1 
( f ∗ g)(x) = f (x ⊕ y)g(y) (x ∈ {0, 1}n )
2n y∈{0,1}n

where ⊕ denotes the coordinatewise binary addition:

1 ⊕ 1 = 0 = 0 ⊕ 0, 1 ⊕ 0 = 0 ⊕ 1 = 1.
Bases and Transforms of Set Functions 221

In terms of set functions, we obtain the following expression:

1 
( f ∗ g)(S) = f (SΔT )g(T ) (S ∈ 2 X ).
2n T ⊆[n]

The properties of the Fourier transform are gathered in the next theorem.
Theorem 1 Let f, g be two pseudo-Boolean functions. The following holds.

(i) f (0) = f (S);
S⊆[n]

(ii)
f (∅) = E[ f ];

(iii) (Parseval’s identity)  f 2 = S⊆[n]

f 2 (S);

(iv)
f 2 (S) = Var[ f ];
S∈2[n] \{∅}

(v) f is constant if and only if

f (S) = 0 for all S = ∅;

(vi) (
f ∗ g)(S) = g (S) for all S ∈ 2[n] .
f (S)

The name “Fourier transform” comes from the work of Fourier on the repre-
sentation of integrable functions. The Fourier transform of a function, viewed as
a function of time, gives its frequency representation. Exactly the same results as
those in Theorem 1 hold for the original Fourier transform, which explains its name
in computer sciences. However, it must be noted that all the properties, except the last
one on convolution, are direct consequences of the orthonormality of the basis. To
the opinion of the author, this transform should be rather called the Walsh transform,
since the definition of the Walsh function is an infinite version of the parity function
used here, as a comparison of (3) and (4) reveals.

5 Bases and Linear Transforms on Set Functions

The previous sections have introduced various bases on set functions and pseudo-
Boolean functions: the unanimity games, the Walsh functions and the parity func-
tions. As a by-product, two fundamental notions in combinatorics and computer
sciences have appeared, namely the Möbius transform and the Fourier transform.
The name “transform” intuitively means (in particular by reference to well-known
transforms used in mathematics for the analysis of real-valued functions: the (origi-
nal) Fourier transform and the Laplace transform, mainly) a representation in another
domain, but equivalent to the original one, i.e., the main desirable characteristic of
the transform is that it should be invertible.
222 M. Grabisch

Although it comes as an evidence from elementary considerations in linear alge-

bra, there is a duality between bases and linear invertible transforms, which to the
knowledge of the author, has never been exploited nor remarked. The aim of this
section is to explain this duality and to apply it, first to the known bases and trans-
forms so as to obtain new bases and transforms, and second to some well-known
inverse problem and representation of integrals w.r.t. games.
We define a transform on the set of set functions on X as a linear invertible
mapping Ψ : R2 → R2 , with ξ → Ψ ξ . The following discussion is made easier
X X

if one consider set functions ξ as row vectors and use matrix notation. To a basis
S ) S∈2 X , we make correspond the matrix B = [b S ] of row vectors b S . Hence ξ =
(b
S∈2 X w S b S = w B is the expression of ξ in this basis. The following lemma gives
the exact equivalence between bases and transforms.
Lemma 1 [10] For every basis B, there is a (unique) transform Ψ such that for any
X
ξ ∈ R2 , 
ξ= Ψ ξ (S)b S , (7)
S∈2 X


whose inverse Ψ −1 is given by ξ → (Ψ −1 )ξ = S∈2 X ξ(S)b S = ξ B.
Conversely, to any transform Ψ corresponds a unique basis B such that (7) holds,
given by b S = (Ψ −1 )δS .
In the above lemma, δ S denotes the Dirac set function defined by

1, if S = T
δ S (T ) = (S ∈ 2 X ).
0, otherwise

We apply this result on a number of commonly used bases and transforms.

(i) The Möbius transform is, as we have already noticed, related to the basis of
unanimity games:
 
ξ(S) = m ξ (T )u T (S) = m ξ (T ), (S ⊆ X ),
T ∈2 X T ⊆S

with 
m ξ (S) = (−1)|S\T | ξ(T ).
T ⊆S

(ii) The co-Möbius transform ([15], a.k.a. commonality function [23]) is defined
by:
 
m̌ ξ (S) = (−1)n−|T | ξ(T ) = (−1)|T | ξ(X \ T ) (S ∈ 2 X ).
T ⊇X \S T ⊆S
Bases and Transforms of Set Functions 223

Its inverse relation is


ξ(S) = (−1)|T | m̌ ξ (T ).
T ⊆X \S

By Lemma 1, the associated basis is

 
(−1)|T | if S ∩ T = ∅
ǔ T (S) = (−1)|B| δT (B) =
0 otherwise.
B⊆X \S

(iii) The (Shapley) interaction transform [12] is defined by

 (n − t − s)!t! 
I ξ (S) = (−1)|S\L| ξ(T ∪ L),
T ⊆X \S
(n − s + 1)! L⊆S

and the inverse relation is given by

 |K |
ξ(S) = β|S∩K | I ξ (K ),
K ⊆X

where
k 
 k
βkl = Bl− j (k ≤ l),
j=0
j

and B0 , B1 , . . . are the Bernoulli numbers. The first values of βkl are given in
Table 1.
The associated basis {bTI }T ∈2 X is

|T |
bTI (S) = β|T ∩S| (S ∈ 2 X ).

Table 1 The coefficients βkl k \l 0 1 2 3 4

0 1 − 21 1
6 0 − 30
1

1 1
2 −3
1 1
6 − 30
1

2 1
6 −6
1 2
15
3 0 − 30
1

4 − 30
1
224 M. Grabisch

(iv) The Banzhaf interaction transform [22] is defined by

 1 n−s 
ξ
IB (S) = (−1)|S\K | ξ(K ),
2 K ⊆X

with inverse relation

  1 k
(IB−1 )ξ (S) = (−1)|K \S| ξ(K ). (8)
K ⊆X
2

The associated basis {bTI B }T ∈2 X is

  1 k  1 |T |
bTI B (S) = (−1)|K \S| δT (K ) = (−1)|T \S| .
K ⊆X
2 2

(v) The Fourier interaction transform: as it was already explained in Sect. 4, the
Fourier transform, as given by (6), corresponds to the basis of parity functions,
given by (5).
The relations between the Fourier transform and the Möbius and Banzhaf
transforms are given as follows:
 1

ξ(S) = (−1)|S| m ξ (K ). (9)
2 k
K ⊇S

 −1 s

ξ(S) =
ξ
IB (S). (10)
2
(vi) The Walsh basis: This basis is defined by (1). Let us recover the correspond-
ing transform ξ → W ξ by using Lemma 1, which we already gave in (2). By
Lemma 1, the inverse transform is immediate:

(W −1 )ξ (S) = ξ(T )(−1)|T \S| .
T ⊆X

The direct transform is obtained by solving the linear system


ξ(S) = W ξ (T )(−1)|T \S| (S ∈ 2 X ),
T ⊆X

or by simply noticing that wT (S) = 2|T | bTI B (S), which from

 ξ

ξ(S) = I B (T )bTI B (S) = W ξ (T )wT (S)
T ⊆X T ⊆X
Bases and Transforms of Set Functions 225

yields the components of W ξ as

 1 |T |
ξ
W ξ (T ) = I B (T ) (T ∈ 2 X ).
2
We recover formula (2). Note that the Fourier and Walsh bases are related as
follows:

χT (S) = χ S (T ) = (−1)|S∩T | = (−1)|S\(X \T )| = w S (X \ T ).

Also, from (10), we find

F ξ (S) = (−1)s W ξ (S) (S ∈ 2 X ). (11)

(vii) The Yokote basis: (see [27, 28]) it is a basis of the set of games, which is
defined by

1, if |S ∩ T | = 1
bTY (S) = (S ∈ 2 X \ ∅). (12)
0, otherwise

Any game v reads in this basis


v= Y v (T )bTY (13)
T ∈2 X \∅

where the coordinates Y v (S) define the Yokote transform Y . We give now Y v
in terms of m v and v, as well as the inverse relations:

m v (S) = |S|(−1)|S|+1 Y v (K ) (∅ = S ⊆ X ). (14)
K ⊇S

 1
Y v (S) = (−1)|S|+1 m v (K ) (∅ = S ⊆ X ). (15)
K ⊇S
|K |

 (n − s − l)!(s + l − 1)!
Y v (S) = (−1)|S∩L|+1 v(L). (16)
L⊆X
n!

Table 2 summarizes the correspondence between bases and transforms.

 226

Table 2 Correspondence between bases and transforms

Transform Basis

 1, if S ⊇ T
Möbius m ξ (S) = (−1)|S\T | ξ(T ) u T (S) =
T ⊆S
0, otherwise

 (−1)|T | , if S ∩ T = ∅
ξ n−|T |
co-Möbius m̌ (S) = (−1) ξ(T ) ǔ T (S) =
T ⊇X \S
0, otherwise

ξ  1, if S ∩ T = ∅
|S|+1 n−|T |
Conjugate unanimity games U (S) = (−1) (−1) ξ(T ) u T (S) =
T ⊇X \S
0, otherwise
|T |
Shapley interaction I ξ (S) = bTI (S) = β|T ∩S|
 |X \ (S ∪ K )|!|K \ S|!
(−1)|S\K | ξ(K )
(n − s + 1)!
K ⊆X
 1 n−s   1 |T |
ξ
Banzhaf interaction IB (S) = (−1)|S\K | ξ(K ) bTI B (S) = (−1)|T \S|
2 2
K ⊆X
1 
Fourier ξ(S) = n (−1)|S∩K | ξ(K ) χT (S) = (−1)|S∩T |
2
K ⊆X
1 
Walsh W ξ (S) = n (−1)|S\K | ξ(K ) wT (S) = (−1)|T \S|
2
K ⊆X 
v 1, if |S ∩ T | = 1
Yokote (S = ∅) Y (S) = bTY (S) =
 (n − s − l)!(s + l − 1)! 0, otherwise
(−1)|S∩L|+1 v(L)
n!
L⊆X
M. Grabisch
Bases and Transforms of Set Functions 227

6 The Inverse Problem for Linear Values

X
In cooperative game theory, a linear value is a linear mapping Φ : R2 → R X assign-
ing to any game v a n-dim vector Φ(v), representing a sharing of v(X ) among all
players (elements
 of X ). For this reason a value most often satisfies efficiency, in
the sense that i∈X Φi (v) = v(X ). The best-known values are the Shapley value
[24] and the Banzhaf value [2]. They are both linear and their definition amounts to
considering the respective interaction transforms for singletons, i.e.:

ΦiSh (v) = I v ({i}), ΦiB (v) = I Bv ({i}), (i ∈ X ). (17)

Hence, the interaction and Banzhaf interaction transforms can be seen as extensions
of these values.
The duality between transforms and bases permits to easily solve the so-called
“inverse problem”, well-known in game theory (see, e.g., [8, 9, 18, 28]): given a
game v on X , find all games v  having the same Shapley value (or any other linear
value: Banzhaf, egalitarian, etc.), i.e., Φ Sh (v) = Φ Sh (v  ).
Considering a linear value Φ and a game v, finding all games v  such that
Φ(v) = Φ(v  ) amounts by linearity to solving Φ(v − v  ) = 0, i.e., v − v  ∈ ker(Φ).
Hence the solution of the inverse problem reduces to finding the kernel of the linear
operator Φ.
The kernel is easily found if there exists a transform Ψ extending the linear value
Φ, exactly as the interaction transform extends the Shapley value (see (17)). Indeed,
the kernel is just the space spanned by the vectors f S of the corresponding basis
with |S| > 1. We illustrate this method with the Shapley value. For any game v, its
epxression in the basis induced by the interaction transform is:
  
v= I v (S)b SI = ΦiSh (v)b{i}
I
+ I v (S)b SI ,
S∈2 X i∈X |S|>1

which implies 
v ∈ ker(Φ Sh ) ⇐⇒ v= I v (S)b SI
|S|>1

i.e.,  
ker(Φ Sh ) = λ S b SI | λ S ∈ R .
|S|>1

In the case where |X | = 3, we obtain, using Table 1:

1
v(∅) = λ∅ + (λ12 + λ13 + λ23 )
6
1 1 1 1
v(1) = λ∅ − λ12 − λ13 + λ23 + λ123
3 3 6 6
228 M. Grabisch

1 1 1 1
v(2) = λ∅ − λ12 + λ13 − λ23 + λ123
3 6 3 6
1 1 1 1
v(3) = λ∅ + λ12 − λ13 − λ23 + λ123
6 3 3 6
1 1 1 1
v(12) = λ∅ + λ12 − λ13 − λ23 − λ123
6 3 3 6
1 1 1 1
v(13) = λ∅ − λ12 + λ13 − λ23 − λ123
3 6 3 6
1 1 1 1
v(23) = λ∅ − λ12 − λ13 + λ23 − λ123
3 3 6 6
1
v(123) = λ∅ + (λ12 + λ13 + λ23 ),
6
where λ S ∈ R for every S ⊆ {1, 2, 3}.
Let us give a second illustrative example with the Banzhaf value. Consider the
following problem: Given a n-dim vector y, find all games v s.t. Φ B (v) = y. The set
of solutions is simply the set of games of the form

v = v y + w, with w ∈ ker(Φ B ),

and v y is any game s.t. Φ B (v y ) = y. Since the Banzhaf interaction transform gener-
alizes the Banzhaf index, we have

ker(Φ B ) = Sp{bTIB , |T | > 1}

with bTIB (S) = (1/2)|T | (−1)|T \S| , and Sp denotes the space spanned by the vectors.
Now, v y can be obtained as the inverse transform of the game w defined by w({i}) =
yi for all i ∈ X , and w(S) = 0 otherwise. This yields by (8):

1   
v y (S) = yi − yi .
2 i∈S i ∈S
/

In the case where there is no known transform which extends the linear value
under consideration, a general method is given by [10], consisting in the following
steps:
X
(i) Select a basis E = {e1 , . . . , ek } of the range Φ(R2 ).
X
(ii) Find set functions b1 , . . . , bk ∈ R2 such that Φ(bi ) = ei , i = 1, . . . , k.
(iii) Complete the independent set {b1 , . . . , bk } to form a basis B = {b1 , . . . , b2n }
X
of R2 .
( j) ( j)
(iv) Compute the coordinates 1 , . . . , k of Φ(b j ) in E for j = k + 1, . . . , 2n .
 ( j)
(v) Compute bΦj = b j − i=1 k
i bi for j = k + 1, . . . , 2n .
Φ
Finally, the basis of the kernel is {bk+1 , . . . , b2Φn }.
Bases and Transforms of Set Functions 229

7 Alternative Expressions of the Choquet Integral

A second application of Lemma 1 is to obtain various equivalent expressions of the

Choquet integral. Let v be a game on X , and f : X → R+ a real-valued nonnegative
mapping. The Choquet integral [5] of f w.r.t. v is defined by

 ∞
f dv = v({x ∈ X | f (x) ≥ α}) dα (18)
0

which yields in the discrete case (|X | = n):

 
n
f dv = ( f σ(i) − f σ(i−1) )v({σ(i), . . . σ(n)})
i=1

where σ is a permutation on [n] such that f σ(1) ≤ f σ(2) ≤ · · · ≤ f σ(n) , and f σ(0) = 0.

Remark 1 Usually, the Choquet integral is defined w.r.t. a capacity, that is, a game
which is monotone with respect to set inclusion. However, it is not possible to extend
the definition to arbitrary set functions ξ, i.e., such that ξ(∅) = 0. Indeed, if ξ(∅) =
0, it is easily seen from (18) that the integral becomes unbounded.

 The Choquetintegral ispositively homogeneous but not additive in general, i.e.,

( f + g) dv = f dv + g dv. However, an interesting feature of this integral is
that it is linear w.r.t. the game:
  

f d(v + αv ) = f dv + α f dv  .

Expressing v in some basis, it is then possible to get the expression of the Choquet
integral w.r.t. the corresponding transform.
Let Ψ be a linear invertible transform, and {bΨA } A∈2 X be the corresponding basis of
set functions. Due to Remark 1, one has to be careful because many bases are com-
posed of set functions which are not games. Therefore, some adaptation is necessary.
From a basis {bΨA } A∈2 X , we build a basis of games {bΨ
A } A∈2 X \{∅} as follows:

bΨS (T ), if T = ∅
bΨ
S (T ) = (S ∈ 2 X \ {∅}).
0, otherwise

Then, from the linearity of the integral, for every f ∈ R X and every game v,
     
v
f dv = fd Ψ (A)bΨ
A = Ψ v (A) f dbΨ
A .
∅= A⊆X ∅= A⊆X
230 M. Grabisch

It is therefore sufficient to compute f dbΨ
A for every A ⊆ X , A  = ∅. One obtains
the following expressions, for the main bases:
• For the Möbius transform:  
f du A = fi (19)
i∈A

• For the co-Möbius transform:

 
f dǔ A = (−1)|A|+1 fi .
i∈A

• For the Fourier transform:

 |A|

f dχA = f σ(n) + 2 (−1) j f i j
j=1

with A = {i 1 , . . . , i |A| } and f i1  · · ·  f i|A| .

Equation (19) is well-known and was first proved by [4] (also by [25]), extending a
result of [7]).
There is no simple expression for the case of the interaction transform, although
such an expression exists and has been obtained through a different method [13]:
   
v+
f dv = B|K | I (A ∪ K ) fi
A⊆X K ⊆X \A i∈A
  
−
+ (−1)|A|+1 B|K | I v (A ∪ K ) fi (20)
∅= A∈2 X K ⊆X \A i∈A

where 
v+ I v (A), if I v (A) > 0
I (A) = ,
0, otherwise

v− I v (A), if I v (A) < 0
I (A) = (A ∈ 2 X ).
0, otherwise

References

1. Aigner, M.: Combinatorial Theory. Springer (1979)

2. Banzhaf, J.: Weighted voting doesn’t work: a mathematical analysis. Rutgers Law Rev. 19,
317–343 (1965)
3. Berge, C.: Principles of Combinatorics. Academic Press (1971)
Bases and Transforms of Set Functions 231

4. Chateauneuf, A., Jaffray, J.-Y.: Some characterizations of lower probabilities and other
monotone capacities through the use of Möbius inversion. Math. Soc. Sci. 17, 263–283 (1989)
5. Choquet, G.: Theory of capacities. Annales de l’Institut Fourier 5, 131–295 (1953)
6. de Wolf, R.: A brief introduction to Fourier analysis on the Boolean cube. Theory Comput.
Lib. Graduate Surv. 1, 1–20 (2008)
7. Dempster, A.P.: Upper and lower probabilities induced by a multivalued mapping. Ann. Math.
Stat. 38, 325–339 (1967)
8. Dragan, I.: The potential basis and the weighted Shapley value. Libertas Mathematica 11,
139–150 (1991)
9. Dragan, I.: The least square values and the Shapley value for cooperative TU games. TOP 14,
61–73 (2006)
10. Faigle, U., Grabisch, M.: Bases and linear transforms of TU-games and cooperation systems.
Int. J. Game Theory, to appear
11. Fujishige, S.: Submodular functions and optimization. In: Annals of Discrete Mathematics,
vol. 58, 2nd edn. Elsevier, Amsterdam (2005)
12. Grabisch, M.: k-Order additive discrete fuzzy measures and their representation. Fuzzy Sets
Syst. 92, 167–189 (1997)
13. Grabisch, M., Labreuche, Ch.: The symmetric and asymmetric Choquet integrals on finite
spaces for decision making. Stat. Pap. 43, 37–52 (2002)
14. Grabisch, M., Labreuche, Ch.: A decade of application of the Choquet and Sugeno integrals
in multi-criteria decision aid. Ann. Oper. Res. 175, 247–286 (2010). doi:10.1007/s10479-009-
0655-8
15. Grabisch, M., Marichal, J.-L., Roubens, M.: Equivalent representations of set functions. Math.
Oper. Res. 25(2), 157–178 (2000)
16. Hammer, P.L., Rudeanu, S.: Boolean Methods in Operations Research and Related Areas.
Springer (1968)
17. Harsanyi, J.C.: A simplified bargaining model for the n-person cooperative game. Int. Econ.
Rev. 4, 194–220 (1963)
18. Kleinberg, N.L., Weiss, J.H.: Equivalent n-person games and the null space of the Shapley
value. Math. Oper. Res. 10(2), 233–243 (1985)
19. O’Donnell, R.: Analysis of Boolean functions, draft 2.0, ch. 1–3 (2007). http://www.cs.cmu.
edu/~odonnell11/boolean-analysis
20. Peleg, B., Sudhölter, P.: Introduction to the Theory of Cooperative Games. Kluwer Academic
Publisher (2003)
21. Rota, G.C.: On the foundations of combinatorial theory I. Theory of Möbius functions.
Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete. 2, 340–368 (1964)
22. Roubens, M.: Interaction between criteria and definition of weights in MCDA problems. In:
44th Meeting of the European Working Group “Multicriteria Aid for Decisions”. Brussels,
Belgium (1996)
23. Shafer, G.: A Mathematical Theory of Evidence. Princeton University Press (1976)
24. Shapley, L.S.: A value for n-person games. In: Kuhn, H.W., Tucker, A.W. (eds.) Contributions
to the Theory of Games, Vol. II, number 28 in Annals of Mathematics Studies, pp. 307–317.
Princeton University Press (1953)
25. Walley, P.: Coherent lower (and upper) probabilities. Technical Report 22, University of War-
vick, Coventry (1981)
26. Walsh, J.: A closed set of normal orthogonal functions. Am. J. Math. 45, 5–24 (1923)
27. Yokote, K.: Weak addition invariance and axiomatization of the weighted Shapley value. Int.
J. Game Theory 44, 275–293 (2015)
28. Yokote, K., Funaki, Y., Kamijo, Y.: Linear basis to the Shapley value. Technical report, Waseda
Economic Working Paper Series (2013)
Conditioning for Boolean Subsets, Indicator
Functions and Fuzzy Subsets

Siegfried Weber

Abstract This chapter deals with measure-free conditioning. It starts with the mean
value based definition of conditional fuzzy subsets which again gives a fuzzy subset.
Applying this general construction to indicator functions, it is proved that these
conditionals form an MV-algebra and that this is isomorphic to the already known
MV-algebra of the interval based conditional Boolean subsets. In the following,
the problem of iteration is completely solved with the result that there are exactly
two types of iteration, called the blurred resp. the sharper one, which remain in the
corresponding MV-algebras. Moreover, the general concept of conditional operators
plays a significant role. Finally, the problem of extending an uncertainty measure is
discussed.

1 Introduction

The first step in “measure-free conditioning” consists in the construction of condi-

tional events “a given b” as well-defined elements of some structured set such that
the (unconditional) events a are described by the conditional events “a given the sure
event”. In a second step, the uncertainty of such conditional events is expressed by
elements of the real unit interval as values of a suitable measure.
Whereas the author’s last papers [19–21] treated problems mainly from the second
step, the present paper concentrates on the first step. In the introduction of our joint
paper [11] with Ulrich Höhle we observed that “… the iteration of measure-free
conditioning is still an open problem.” In [11] we gave a partial solution for the quite
general situation of events from a Girard algebra. In the present paper we will give a
complete solution to this problem for the classical situation of events from a Boolean
algebra B of subsets of an universe Ω, where conditioning operators, as introduced
in [11], will play an essential role.

S. Weber (B)
Institut Für Mathematik, Universität Mainz, FB 08, Mainz, Germany
e-mail: [email protected]

© Springer International Publishing Switzerland 2016 233

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_14
234 S. Weber

Another motivation for the present paper came from several talks at the Linz Sem-
inar about “a new axiomatic approach”, where the conditional events are introduced
by three-valued “generalized indicator functions”

(∗) 1A|B = 1 · 1A∩B + u · 1Bc + 0 · 1Ac ∩B , A, B ∈ B, with some u ∈ [0, 1] ,

where the “undetermined value” u is not a constant number but is considered as

value u = t(A | B) of a “conditional uncertainty measure”, see [5, 6]. Therefore, this
interesting approach is not measure-free. Contrary to this, in the present paper the
measure-free version of the right side in (∗) will be considered, i.e. where u is a
constant. We will see that u = (0 | 0) should be a self-complemented element of
the MV-chain [0, 1] and, therefore, u = 21 . It will be shown that these conditional
indicator functions form an MV-algebra, where the underlying partial ordering results
to be equivalent to the interval ordering used for the interval based definition of
conditional Boolean subsets

(CB) (A  B) = [A ∩ B , B → A] for A, B ∈ B,

denoting this set by B̃. This lattice-interval approach has been treated in detail in a lot
of publications, see e.g. [9, 11, 17] and the references therein. In this approach the
“conditional Boolean event” (A  B) is defined as the set of all possible conditional
candidates between the Boolean events “A and B” resp. “if B then A” as the extreme
candidates. In contrast to this approach, it is well-known that there is no reasonable
way to define a conditional as some Boolean subset (A | B) between these two extreme
Boolean subsets. On the other hand, this alternative approach really is applicable to
indicator functions. It seems that this approach has not yet treated systematically.
This will be done in the present paper, including the problem of iteration.
More concretely, we start with the general construction, introduced in [18], of
conditional f uzzy sets defined pointwisely by

(CF) (ϕ | ψ) = C(ϕ ∧ ψ , ψ → ϕ) ∈ F for fuzzy subsets ϕ, ψ ∈ F = [0, 1]Ω ,

based on a mean value function C on the unit interval [0, 1] which is compatible
with the complement in [0, 1]. The properties of a mean value function C are very
natural in order to choose some fuzzy subset between the extreme fuzzy subsets.
The additional property of compatibility guarantees that C generates a conditioning
operator | on F. Now, applying this construction to indicator functions, we prove that
the subset F1 of the conditional indicator functions

1
(CI) (1A | 1B ) = 1 · 1A∩B + · 1Bc + 0 · 1Ac ∩B
2

is an MV-algebra and closed under iteration if and only if C = C1 given by C1 (0, 21 ) =

1
2
or C = C2 given by C2 (0, 21 ) = 0. Because of these values we call the first type
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 235

“blurred iteration” and the second type “sharper iteration”. Furthermore, we prove
that any conditioning operator | on F1 is generated by exactly one of these two
compatible mean value functions.
Moreover, it follows that i : F1 → B̃ is an MV-algebra isomorphism between
the two types (CI) resp. (CB) of conditionals. We use this result to obtain two
corresponding conditioning operators ˜| on B̃ by

((A1  B1 ) ˜| (A2  B2 )) = i ◦ ((1A1 | 1B1 ) | (1A2 | 1B2 )) .

Particularly, it follows that the conditional Boolean subsets (A  B) = ({A} ˜| {B})

are recovered as values of the conditioning operators ˜| applied to singletons. This
approach to the iteration process in B̃ leads to the same result as in [11] although is
very different, but it shows that the two special compatible mean value functions on
B̃ given in [11] are the only two and are obtained as

C̃k ( i ◦ ϕ , i ◦ ψ ) = i ◦ Ck (ϕ, ψ) for ϕ, ψ ∈ F1 , ϕ ≤ ψ , k = 1, 2 .

The main results of the present paper were presented in the author’s talk at the
34th Linz Seminar on Fuzzy Set Theory, 2013.
The paper is organized as follows. In Sect. 2 we put together the basic prerequisites,
referring to MV-algebras, mean value functions, conditioning operators and the con-
ditional fuzzy subsets (CF) in F. Section 3 treats the conditional indicator functions
(CI) in F1 which form an MV-algebra (Theorem 1). Section 4 treats the conditional
Boolean subsets (CB) in B̃ (Theorem 2) and the MV-algebra isomorphism between
F1 and B̃ (Theorem 3). Section 5 is dedicated to the iteration processes in both F1
(Theorem 4) and B̃ (Theorem 5). Finally, in Sect. 6 we present the basic topics from
the second step of measure-free conditioning, i.e. referring to uncertainty measures:
the general Definition 5, the Theorem 6 for F1 and B̃, some examples and remarks,
including the Remark 9 for the alternative approach (∗).

2 Basic Definitions and Results

The basic structure needed in the paper is that of an MV-algebra which we introduce
as follows.
Definition 1 (MV-algebras)
(i) A set L is called a residuated lattice if it is equipped with the two structures of a
bounded lattice (L, ≤, ∧, ∨) with universal upper (resp. lower) bound 1 (resp.
0) and a commutative monoid (semigroup with 1 as unit) (L, ), such that there
exist all residuals b → a given by the residuation property

cb ≤ a ⇐⇒ c ≤ b → a.
236 S. Weber

For each b ∈ L a residual complement can be defined by

b = b → 0.

(ii) Furthermore, L is called an MV-algebra if the lattice-join can be expressed by

residuals by means of the additional property

(MV) b ∨ a = (b → a) → a.

For each a, b ∈ L a dual semigroup operation (with 0 as unit) can be defined by

a  b = (a  b ) .

Sometimes we will use the notation (L, , , ) instead of L, dealing with

explicitely the MV-operations in this order.
The values a  b of the dual operation in (ii) will be interpreted as “unions”. If
the “intersection” satisfies a  b = 0 we call them “disjoint unions” and write a ˙ b,
for short.
In previous papers, e.g. in [21], the author used the name commutative residuated
lattice ordered semigroup with zero from [1] rather than the name residuated lattice,
see [15], the references therein and the book [8].
The operations denoted here by , ,  play the role of the MV-algebra operations
−
·, , + originally used by Chang in [3] resp. , ¬, ⊕ used e.g. in the book [4] as
general reference to MV-algebras. See also the biographical Remark 6.5 in [21].
In previous papers we needed the structure of a Girard algebra, which is a resid-
uated lattice where the residual complementation is idempotent, i.e. a structure
”between” residuated lattices and MV-algebras, see [11, 21] and the biographical
Remark 6.2 therein.

Remark 1 (Boolean algebras) It is well known that an MV-algebra L is a Boolean

algebra if and only if  = ∧ .

In the present paper we will often tacitly use the following well-known properties:

Proposition 1 (Additional properties in MV-algebras)

(i) The residual complement has the involution property

b = b.

(ii) The residuals can be expressed as

b → a = b  a = b ˙ (a ∧ b).
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 237

(iii) The lattice-meet can be expressed by the semigroup operations and the residual
by means of the divisibility property

b ∧ a = b  (b → a) = b  (b  a).

The following well-known result is the starting point of the present paper.

Lemma 1 (The MV-algebra of fuzzy subsets) The set

F = [0, 1]Ω = {ϕ : Ω → [0, 1]}

of all fuzzy subsets ϕ of an universe Ω has the structure of an MV-algebra, inherited

pointwisely from the standard MV-(algebra-)chain ([0, 1], , ¬, ⊕) for the values
ϕ(ω) by

a  b = (a + b − 1) ∨ 0 , ¬a = 1 − a , a ⊕ b = (a + b) ∧ 1 ,

where  , ¬ , ⊕ denote (here and in the following) the MV-operations “intersec-

tion”, “complement”, “union” in [0, 1] resp. F.

In the following we deal with the basics about mean value functions which will
play a crucial role in our presentation of conditioning.

Definition 2 (Mean value functions on MV-algebras) A compatible mean value

function C on an MV-algebra L = (L, , , ) is
(i) a mean value function on L, i.e. a map C defined for all a, b ∈ L with a ≤ b
and values in L, which is isotone in both arguments and idempotent,
(ii) which is compatible with the complement in L, i.e. satisfying (C(a, b)) =
C(b , a ).

Remark 2 (Compatible mean value functions) The additional property that a mean
value function C on an MV-algebra L is compatible implies that (C(0, 1)) = C(0, 1),
i.e. in L has to exist a self-complemented element which can serve as value C(0, 1).

Example 1 (Compatible mean value functions on [0,1]) The standard MV-chain

[0, 1] has 21 as the (unique) self-complemented element and admits a lot of compatible
mean value functions Ck . In the following we deal with four of these to which we will
refer in the following sections. Particularly, the first two will be essential in Sect. 5,
where the special value Ck (0, 21 ) will play an important role. We denote by
⎧ ⎫
⎪
⎨ b if b < 2 ⎪
1
⎬
(i) C1 (a, b) = 21 if a ≤ 21 ≤ b with C1 (0, 21 ) = 21 ,
⎪
⎩ a if 1 < a ⎪ ⎭
a 2

if (a, b) = (0, 1)
(ii) C2 (a, b) = 1+a−b
with C2 (0, 21 ) = 0 ,
1
2
if (a, b) = (0, 1)
238 S. Weber

(iii) C3 (a, b) = a+b

2
with C3 (0, 21 ) = 1
4
,

(iv) C4 (a, b) = b
1+b−a
with C4 (0, 21 ) = 1
3
.
In contrast to these four, the following mean value function, needed in Sect. 6, is not
compatible:
(v) C5 (a, b) = (1 − t) · a + t · b for some t ∈ [0, 1], t = 1
2
.

In the last part of this section, we present one concept of conditioning, first the
general definition introduced in [18], then the connection with the axiomatic approach
via conditioning operators introduced in [10] and, finally, the application to fuzzy
subsets. For more details see also [11], where this concept is extended to the more
general structure of Girard algebras.

Definition 3 (Conditional elements in an MV-algebra) Let L be an MV-algebra with

a self-complemented element and let C be any compatible mean value function on L.
Then conditional elements “a given b” are defined by

(C) (a | b) = C(a ∧ b , b → a) ∈ L for a, b ∈ L.

Proposition 2 (Conditional elements and conditioning operator in an MV-algebra)

Let L = (L, , , ) be an MV-algebra with a self-complemented element. Then:
(i) Each compatible mean value function C on L “generates” a conditioning oper-
ator | on L, i.e. a binary operation |: L × L −→ L with values (a | b) given
by (C), which satisfies
(C1) (a | 1) = a,
(C2) (a | b) = (a ∧ b | b),
(C3) a1 ≤ a2 ⇒ (a1 | b) ≤ (a2 | b),
(C4) b1 ≤ b2 and a ∧ b2 ≤ a ∧ b1 ⇒ (a | b2 ) ≤ (a | b1 ),
(C5) (a | b) = (a  b | b).
Particularly, conditioning only with lower bound 0 and upper bound 1 in L
leads to

(0 | 1) = C(0, 0) = 0 , (0 | 0) = (0 | 0) = C(0, 1) , (1 | 1) = C(1, 1) = 1 .

(ii) The compatible mean value function C can be recovered by

C(a, b) = (a | b → a) for a ≤ b.

(iii) Vice versa, given any conditioning operator | on L, i.e. fulfilling the axioms
(C1) − (C5) from (i), then there exists a compatible mean value function C,
given by (ii) if and only if | satisfies the additional condition

(cmvf ) a1 ≤ a2 ⇒ (a1 ∧ b | b → a1 ) ≤ (a2 ∧ b | b → a2 ).

Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 239

Corollary 1 (Conditional fuzzy subsets) Let F = [0, 1]Ω be the MV-algebra of fuzzy
subsets of a universe Ω. Then each compatible mean value function C on [0, 1] leads
to a conditioning operator | on F, where its values are the “conditional fuzzy subsets”
given pointwise by

(CF) (ϕ | ψ) = C(ϕ ∧ ψ , ψ → ϕ) ∈ F for ϕ, ψ ∈ F.

3 The MV-Subalgebra F1 of Conditional Indicator

Functions

Notation 1 Let B = (B, ∩,c , ∪) be a Boolean (MV-)algebra of (crisp) subsets of

the universe Ω. Then we will denote by

F0 = {1A : A ∈ B} resp. F1 = {(1A | 1B ) : 1A , 1B ∈ F0 }

the subsets of F containing all indicator functions resp. all conditional indicator
functions.

Remark 3 It is well known that F0 is a Boolean (MV-)subalgebra of F such that

F0 ∼
= B is a Boolean algebra isomorphism:

1A  1B = 1A∩B , ¬1A = 1Ac , 1A ⊕ 1B = 1A∪B .

Applying the general construction (CF) of conditioning fuzzy subsets in F to

indicator functions requires only the properties C(0, 0) = 0, C(1, 1) = 1 of any mean
value function C on [0, 1] and C(0, 1) = 21 of any compatible one and, therefore,
leads directly to the following

Proposition 3 (Conditional indicator functions) The elements of F1 can be written

as
1 1
(CI) (1A | 1B ) = 1 · 1A∩B + · 1Bc + 0 · 1Ac ∩B = 1A∩B + · 1Bc .
2 2
These “conditional indicator functions” do not depend on the choice of the compat-
ible mean value function C on [0, 1], in other words, only the values (1 | 1) = 1,
(0 | 1) = 0 and (1 | 0) = (0 | 0) = 21 are needed.

It follows from (CI) that the conditioning operator | on F cannot be restricted to a

conditioning operator on F0 . The question, if this is possible on F1 , will be answered
in Sect. 5. As a first step in this direction we need the following

Theorem 1 (The MV-subalgebra of conditional indicator functions) The subset F1

of all conditional indicator functions is an MV-subalgebra of F with respect to
240 S. Weber

(i) (1A1 | 1B1 )  (1A2 | 1B2 ) = (1D | 1E ) with

D = A1 ∩ B1 ∩ A2 ∩ B2 and E = D ∪ (Ac1 ∩ B1 ) ∪ (Ac2 ∩ B2 ) ∪ (B1c ∩ B2c )

(ii) ¬(1A | 1B ) = (1Ac | 1B ) ,

(iii) (1A1 | 1B1 ) ⊕ (1A2 | 1B2 ) = (1F | 1G ) with

F = (A1 ∩ B1 ) ∪ (A2 ∩ B2 ) ∪ (B1c ∩ B2c ) and G = F ∪ (B1 ∩ B2 ) .
Clearly, F0 is a Boolean (MV-)subalgebra of F1 .

Proof (i) is obtained by applying the conditionals (· | ·) in the form (CI) and the
values a  b for a, b ∈ {0, 21 , 1} :

(1A1 | 1B1 )  (1A2 | 1B2 )

= (1 · 1A1 ∩B1 + 1
2
· 1B1c + 0 · 1Ac1 ∩B1 )  (1 · 1A2 ∩B2 + 1
2
· 1B2c + 0 · 1Ac2 ∩B2 )
= 1 · 1A1 ∩B1 ∩A2 ∩B2 + 1
2
· 1(A1 ∩B1 ∩B2c )∪(B1c ∩A2 ∩B2 ) + 0 · 1(Ac1 ∩B1 )∪(Ac2 ∩B2 )∪(B1c ∩B2c ) .

From this we obtain D directly and E = ((A1 ∩ B1 ∩ B2c ) ∪ (B1c ∩ A2 ∩ B2 ))c can
be rewritten into the given form, where D ⊆ E.
Formula (ii) follows directly:

¬(1A | 1B ) = ¬(1 · 1A∩B + 1

2
· 1Bc + 0 · 1Ac ∩B )
= 0 · 1A∩B + 1
2
· 1Bc + 1 · 1Ac ∩B = (1Ac | 1B ).

In analogy to (i) it follows (iii):

(1A1 | 1B1 ) ⊕ (1A2 | 1B2 )

= 1 · 1(A1 ∩B1 )∪(A2 ∩B2 )∪(B1c ∩B2c ) + 1
2
· 1(Ac1 ∩B1 ∩B2c )∪(B1c ∩Ac2 ∩B2 ) + 0 · 1Ac1 ∩B1 ∩Ac2 ∩B2 .

From this we obtain F directly and G = ((Ac1 ∩ B1 ∩ B2c ) ∪ (B1c ∩ Ac2 ∩ B2 ))c can
be rewritten into the given form, where F ⊆ G. 

Remark 4 (Measure-free conditioning) The form (CI) has a long history, for which
we refer to the book [9] and the references therein, where instead of 21 often the
value u but also other symbols as e.g. ? in [7] are used for this third “undefined”
or “undetermined” value. Some proposals for connecting these, sometimes named
“generalized indicator functions”, were given, very different among themselves and
from those in the preceding theorem. But all proposals correspond to “measure-free
conditioning”.
A completely different concept was proposed in [5], where the undetermined value
u is not a constant, i.e. is no longer the same number for all “conditional events”.
This concept will briefly be discussed in Remark 9.
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 241

4 The Isomorphism Between F1 and the MV-Algebra B̃

of Conditional Boolean Subsets

In Sect. 2, a conditional element (a | b) of elements a, b in an MV-algebra L is

introduced as some mean value C(a ∧ b, b → a) between the “natural conditional
candidates” a ∧ b and b → a. But this requires that L has a self-complemented
element.
Therefore, this “mean value approach” is not applicable for a Boolean algebra L.
A solution is to take all candidates between the two mentioned above. This leads to
the “interval approach” of the following

Definition 4 (Conditional Boolean events) The conditional Boolean event “a given

b” of two (crisp) events a, b in any Boolean algebra L = (L, ∧, , ∨) is defined as
the lattice interval

(a  b) = [ a ∧ b , b → a ] = [ a ∧ b , b ∨ a ].

The set of all conditional Boolean events of events in L will be denoted by L̃ .

The author get to know this construction during the talk [12], for details see the
book [9] and the references therein. An extension to MV-algebras was presented in
the author’s talk [16], for details see [17, 18]. In [11] this concept has been extended
to Girard algebras.

Remark 5 The conditional events are in a one-to-one correspondence to intervals

via
[a, c] = (a  c ∨ a).

Therefore, in the following we will alternatively take the conditional events or the
intervals as elements of L̃.

Lemma 2 (The MV-algebra of conditional Boolean events) Let L̃ be the set of con-
ditional Boolean events from the preceding definition. Then the following assertions
hold:
(I1) (a  1) = [a, a] = {a}, i.e. L̃ extends L,
(I2) (a  b) = (a ∧ b  b).
Furthermore, L̃ is a lattice with respect to the partial ordering for intervals

[a, c] ≤ [d, f ] if and only if a ≤ d , c ≤ f

where, therefore, (1  1) is the upper resp. (0  1) the lower universal bound. The
monotonicity properties follow from the lattice structure:
242 S. Weber

(I3) a1 ≤ a2 ⇒ (a1  b) ≤ (a2  b),

(I4) b1 ≤ b2 and a ∧ b2 ≤ a ∧ b1 ⇒ (a  b2 ) ≤ (a  b1 ).
Finally, L̃ is an MV-algebra with respect to
(a  b)  (c  d)
= (a ∧ b ∧ c ∧ d  (a ∧ b ∧ c ∧ d) ∨ (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  )),

(I5) (a  b) = (a  b),

(a  b)  (c  d)
= ((a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  )  (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  ) ∨ (b ∧ d)).

Proof The result was proved in [11] more generally for any MV-algebra L, but only
in terms of intervals. Rewriting these intervals into conditional events lead to the
given formulae for the MV-algebra operations as follows.
In order to obtain the first result apply, in this order, the definition of (·  ·), the
definition of  from Theorem 2.3 in [11], the preceding remark and known properties
in a Boolean algebra:

(a  b)  (c  d) = [ a ∧ b , b ∨ a ]  [ c ∧ d , d  ∨ c ]
= [ (a ∧ b) ∧ (c ∧ d) , ((a ∧ b) ∧ (d  ∨ c)) ∨ ((b ∨ a) ∧ (c ∧ d)) ]
= ( a ∧ b ∧ c ∧ d  ( (a ∨ b ∨ (d ∧ c )) ∧ ((b ∧ a ) ∨ c ∨ d  ) ) ∨ (a ∧ b ∧ c ∧ d) )
= ( a ∧ b ∧ c ∧ d  ( (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  ) ∨ (a ∧ b ∧ c ∧ d) ).

By analogous steps (I5) follows:

(a  b) = [ a ∧ b , b ∨ a ] = [ (b ∨ a) , (a ∧ b) ] = [ b ∧ a , a ∨ b ]

= ( b ∧ a  (a ∨ b ) ∨ (b ∧ a ) ) = (b ∧ a  b) = (a  b).

Analogously and using also the former formulae and in the final step the property
(I2) the third formula follows:

(a  b)  (c  d) = ( (a  b)  (c  d) ) = ( (a  b)  (c  d) )

= ( a ∧ b ∧ c ∧ d  (a ∧ b ∧ c ∧ d) ∨ (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  ) )
= ( a ∨ b ∨ c ∨ d   (a ∧ b ∧ c ∧ d) ∨ (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  ) )
= ( (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  )  (b ∧ d) ∨ (a ∧ b) ∨ (c ∧ d) ∨ (b ∧ d  ) ). 

The properties (I2) − (I5) for the interval based conditionals (a  b) are com-
pletely analogous to (C2) − (C5) for the mean value based conditionals (a | b),
whereas the difference between (I1) and (C1) means simply that the original ele-
ments are embedded into the conditionals in the former construction and are special
conditionals in the latter one.
The preceding result was already proved in [9], mainly in Theorem 3 of Sect. 3,
which can be seen after rewriting the operations given there.
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 243

As already mentioned within the proof, in [11] the result was generalized to each
MV-algebra L where the set L̃ results to be a Girard algebra.
Structures L which are more general than an MV-algebra can also be extended to
the set L̃ of intervals. On the one hand, the Main Theorem 3.1 from [11] extends a
Girard algebra L to a Girard algebra L̃. On the other hand, Theorem 15 from [15]
deals with an “interval-valued residuated lattice” (IVRL) L̃, where the underlying
“base lattice” L results to be a residuated lattice. The particular case for α = 0 in
Theorem 15 from [15] corresponds to the Main Theorem 3.1 from [11]. But in both
situations, intervals cannot be rewritten into conditional events.

Now let B = (B, ∩,c , ∪) be a Boolean (MV-)algebra of (crisp) subsets of the

universe Ω and let B̃ be the corresponding set of conditional Boolean subsets

(CB) (A  B) = [A ∩ B , B → A] = [A ∩ B , Bc ∪ A] for A, B ∈ B.

Then the result of the preceding lemma can be rewritten into the following

Theorem 2 (The MV-algebra of conditional Boolean subsets) For a Boolean alge-

bra B of subsets of the universe Ω, the set B̃ is an MV-algebra with respect to
(i) (A1  B1 )  (A2  B2 ) = (D  E),
(ii) (A  B) = (Ac  B),
(iii) (A1  B1 )  (A2  B2 ) = (F  G),
where D, E resp. F, G result to be the same as for the MV-algebra F1 in Sect. 3.

As a summary of the preceding theorem and the above mentioned corresponding

result in Sect. 3 we obtain explicitely the following (compare with our Remarks 2.5
and 3.3 in [11])

Theorem 3 (The MV-algebra isomorphism between F1 and B̃ ) It follows that F1 ∼

=
B̃ is an MV-algebra isomorphism between the two MV-algebras of “conditionals”,
i.e. of
(CI)conditional indicator functions (1A | 1B ) = 1 · 1A∩B + 21 · 1Bc + 0 · 1Ac ∩B
∈ F1 and
(CB) conditional Boolean subsets (A  B) = [ A ∩ B , Bc ∪ A ] ∈ B̃ ,
for A, B ∈ B, where both partial orderings are equivalent:
 
A1 ∩ B1 ⊆ A2 ∩ B2
(1A1 | 1B1 ) ≤ (1A2 | 1B2 ) ⇔ ⇔ (A1  B1 ) ≤ (A2  B2 ).
B1 → A1 ⊆ B2 → A2

Furthermore, the MV-algebra isomorphism F1 ∼ = B̃ extends the Boolean alge-

bra isomorphism F0 ∼
= B because of (1A | 1Ω ) = 1A ∈ F0 and (A  Ω) = {A},
A ∈ B.
244 S. Weber

Proof Only the assertion referring to the orderings has not yet proved explicitely. In
order to establish the first equivalence, it follows from the left side that A1 ∩ B1 ⊆
A2 ∩ B2 (from “1 ≤ 1”) and Ac2 ∩ B2 ⊆ Ac1 ∩ B1 (from “0 ≤ 0”) and, therefore, also
B1 → A1 ⊆ B2 → A2 , i.e. the direction “⇒” is proved. The other direction in the
first equivalence can easily be obtained by checking the possible cases. The second
equivalence is precisely the definition of the ordering for intervals. 
For the author it was a little surprising when he realized that the (very natural)
ordering in F1 is equivalent to the interval ordering in B̃ in such a direct way, that
this can be seen as a motivation for considering the interval ordering as the adequate
ordering.

5 Iteration of Conditioning

The conditional indicator functions are fuzzy subsets and, therefore, the general
construction (CF) from Sect. 2 can be applied and leads to the following
Lemma 3 (Iteration of conditional indicator functions) The iterated conditional
indicator functions have the form

((1A1 | 1B1 ) | (1A2 | 1B2 )) = 1 · 1A1 ∩B1 ∩A2 ∩B2 + (1 − c) · 1(A1 ∪B1c )∩B2c
1
+ · 1(B1c ∪Ac2 )∩B2 + c · 1Ac1 ∩B1 ∩B2c + 0 · 1Ac1 ∩B1 ∩A2 ∩B2
2

with c = (0 | 21 ) = C(0, 21 ) ∈ [0, 21 ] and, therefore, 1 − c = ( 21 | 21 ) = C( 21 , 1) ∈

[ 21 , 1] .
Proof The announced formula follows from the values

(1 | 1) = C(1, 1) = 1 , (1 | 21 ) = ( 21 | 21 ) = C( 21 , 1) = 1 − c ,
(1 | 0) = ( 21 | 0) = (0 | 0) = C(0, 1) = 21 , ( 21 | 1) = C( 21 , 21 ) = 21 ,
(0 | 21 ) = C(0, 21 ) = c , (0 | 1) = C(0, 0) = 0 . 

As a direct consequence of the preceding lemma we obtain the following

Theorem 4 (Iteration within F1 ) Iteration of conditional indicator functions remains
in F1 , i.e. the conditioning operator | on F can be restricted to a conditioning oper-
ator on F1 ,
if and only if (0 | 21 ) = C(0, 21 ) ∈ {0, 21 } .
Therefore, denoting by C1 resp. C2 any compatible mean value function on [0, 1]
with
1 1 1
C1 (0, ) = resp. C2 (0, ) = 0 ,
2 2 2
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 245

these two mean value functions lead to the following two types of iteration, namely
the

“blurred iteration” for C1 : ((1A1 | 1B1 ) | (1A2 | 1B2 )) = (1A1 | 1B1 ∩A2 ∩B2 ) ,

“sharper iteration” for C2 : ((1A1 | 1B1 ) | (1A2 | 1B2 )) = (1A1 ∪B1c | 1(B1 ∩A2 )∪B2c ) .

Moreover, these two types of iterations can be obtained from any conditioning oper-
ator | on F1 , i.e. fulfilling the conditions (C1) − (C5), with the additional property
(0 | 21 ) = 21 for the blurred resp. (0 | 21 ) = 0 for the sharper iteration.

Proof The formulae follow from the preceding lemma and the property (C2).
On the one hand, the value c = 21 leads to the blurred iteration

(1A1 ∩B1 ∩A2 ∩B2 | 1B1 ∩A2 ∩B2 ) = (1A1 | 1B1 ∩A2 ∩B2 ).

On the other hand, the value c = 0 leads to the sharper iteration

(1(A1 ∩B1 ∩A2 )∪(A1 ∩B2c )∪(B1c ∩B2c ) | 1(B1 ∩A2 )∪B2c ) = (1A1 ∪B1c | 1(B1 ∩A2 )∪B2c ).

Finally, we have to prove the additional condition (cmvf ) for any conditioning oper-
ator | on F1 . For that purpose, let ϕ1 = (1A1 | 1B1 ), ϕ2 = (1A2 | 1B2 ), ψ = (1A | 1B )
with ϕ1 ≤ ϕ2 . Then

(ϕ1 ∧ ψ | ψ → ϕ1 )
= (1 · 1A1 ∩B1 ∩A∩B + 1
2
· 1(A1 ∩Bc )∪(B1c ∩A)∪(B1c ∩Bc ) + 0 · 1(Ac1 ∩B1 )∪(Ac ∩B) |
1 · 1(Ac ∩B)∪(A1 ∩B1 )∪(B1c ∩Bc ) + 1
2
· 1(Ac1 ∩B1 ∩Bc )∪(B1c ∩A∩B) + 0 · 1Ac1 ∩B1 ∩A∩B ).

For (0 | 21 ) = 1
2
this is equal to

(1A1 ∩B1 ∩A∩B | 1(A1 ∩B1 ∩B)∪(Ac ∩B) ) ≤ (1A2 ∩B2 ∩A∩B | 1(A2 ∩B2 ∩B)∪(Ac ∩B) )
= (ϕ2 ∧ ψ | ψ → ϕ2 ),

where “≤” follows from

A1 ∩ B1 ∩ A ∩ B ⊆ A2 ∩ B2 ∩ A ∩ B and
((A1 ∩ B1 ∩ B) ∪ (Ac ∩ B)) → (A1 ∩ B1 ∩ A ∩ B) = B → A

which does not depend on k = 1.

For (0 | 21 ) = 0 it follows analogously

(1(A1 ∪B1c )∩A∩B | 1((A1 ∪B1c ∪Ac )∩B)∪(Ac1 ∩B1 ∩Bc ) )

≤ (1(A2 ∪B2c )∩A∩B | 1((A2 ∪B2c ∪Ac )∩B)∪(Ac2 ∩B2 ∩Bc ) )
246 S. Weber

because of
(A1 ∪ B1c ) ∩ A ∩ B = (B1 → A1 ) ∩ A ∩ B ⊆ . . . and

(((A1 ∪ B1c ∪ Ac ) ∩ B) ∪ (Ac1 ∩ B1 ∩ Bc )) → ((A1 ∪ B1c ) ∩ A ∩ B)

= ((B1 → A1 ) ∩ Bc ) ∪ (A ∩ B) ⊆ . . .

Let us observe that for the blurred iteration only the part A1 ∩ B1 ⊆ A2 ∩ B2 of
the inequality ϕ1 ≤ ϕ2 was needed, whereas for the sharper iteration only the other
part B1 → A1 ⊆ B2 → A2 was needed. 

In the preceding theorem and in the following, the second type we call “sharper
iteration” because the conditionals with the undetermined value as condition have
the “sharp values” (0 | 21 ) = 0 and (1 | 21 ) = ( 21 | 21 ) = 1, whereas the first type we
call “blurred iteration” because all these conditionals have the “blurred value” 21 .
In the following two corollaries we will specify the result of the preceding theorem
for two special situations, which will be taken up and briefly discussed at the end of
this section.

Corollary 2 (Iteration conditioned on an indicator function) In the special case of

crisp indicator function in the condition, both types of iteration lead to the same
result, i.e. for
((1A1 | 1B1 ) | 1A2 ) = (1A1 | 1B1 ∩A2 ) .

Corollary 3 (Iteration with equal indicator function as conditions) In the special

case of equal indicator function in the conditions, both types of iteration lead to
different results, i.e. for the

blurred iteration: ((1A1 | 1B ) | (1A2 | 1B )) = (1A1 | 1A2 ∩B ) ,

sharper iteration: ((1A1 | 1B ) | (1A2 | 1B )) = (1A1 ∪Bc | 1A2 ∪Bc ) .

Now we will prove the result of an iteration process for conditional Boolean
subsets using essentially the isomorphism between these and the conditional indicator
functions. Although the result can also be obtained as a particular case of the more
general situation of Sect. 5 in [11], the formulation of the following theorem and
its proof are very different from [11] and show cleary the analogy of both iteration
processes and that, furthermore, also in the following setting there are exactly two
types of iteration.

Theorem 5 (Iteration within B̃ ) Let ((1A1 | 1B1 ) | (1A2 | 1B2 )) ∈ F1 be any of the
two types of iteration of conditional indicator functions from the preceding theo-
rem and let i : F1 −→ B̃ be the isomorphism from Sect. 4. Then there exist two
conditioning operators ˜| on B̃ given by

((A1  B1 ) ˜| (A2  B2 )) = i ◦ ((1A1 | 1B1 ) | (1A2 | 1B2 )) ,

Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 247

which result to be the corresponding blurred resp. sharper iteration in B̃ :

1 1 1
((A1  B1 ) ˜| (A2  B2 )) = (A1  B1 ∩ A2 ∩ B2 ) for (0 | ) = C1 (0, ) = ,
2 2 2
1 1
((A1  B1 ) ˜| (A2  B2 )) = (A1 ∪ B1c  (B1 ∩ A2 ) ∪ B2c ) for (0 | ) = C2 (0, ) = 0 .
2 2

Moreover, the two conditioning operators ˜| are generated by two compatible mean
value functions C̃k on B̃ given by

C̃k ( i ◦ ϕ , i ◦ ψ ) = i ◦ Ck (ϕ, ψ) for ϕ, ψ ∈ F1 , ϕ ≤ ψ , k = 1, 2 ,

which result to be
C̃1 ( [α, β] , [γ , δ] ) = [α, δ] for C1 ,

C˜2 ( [α, β] , [γ , δ] ) = [β ∩ γ , β ∪ γ ] for C2 ,

written the arguments of C̃k as intervals [α, β] ≤ [γ , δ] of elements α ⊆ β, γ ⊆ δ

in B.

Proof The assertion that both types of ˜| are conditioning operators on B̃ generated by
some compatible mean value function follows immediately from the facts that both
types of | are conditioning operators on F1 fulfilling (cmvf ) and the isomorphism i
between these two MV-algebras. Naturally, it is needed that i is an isomorphism with
respect to the partial orderings, the lattice operations and the MV-algebra operations.
This will be illustrated in the following only for two properties, namely for (C1):

((A1  B1 ) ˜| (Ω  Ω)) = i ◦ ((1A1 | 1B1 ) | (1Ω | 1Ω )) = i ◦ (1A1 | 1B1 ) = (A1  B1 ),

and for (C5):

((A1  B1 ) ˜| (A2  B2 )) = (i ◦ ((1A1 | 1B1 ) | (1A2 | 1B2 )))

= i ◦ (¬((1A1 | 1B1 ) | (1A2 | 1B2 )))
= i ◦ (¬(1A1 | 1B1 )  (1A2 | 1B2 ) | (1A2 | 1B2 ))
= ((A1  B1 )  (A2  B2 ) ˜| (A2  B2 )).

Therefore, it follows by the definitions of ˜| and | that C̃k ( i ◦ ϕ , i ◦ ψ ) = i ◦ Ck (ϕ, ψ)

for

i ◦ ϕ = (A1  B1 ) ∧ (A2  B2 ) = [α, β] ≤ i ◦ ψ = (A2  B2 ) → (A1  B1 ) = [γ , δ].

248 S. Weber

Vice versa, given ϕ, ψ ∈ F1 with ϕ ≤ ψ, then there exist [α, β] ≤ [γ , δ] in B̃ such

that
i ◦ ϕ = [α, β] , i ◦ ψ = [γ , δ].

Finally, the last two formulae are obtained if we start with

C̃k ([α, β] , [γ , δ]) = C̃k ((α  β c ∪ α) , (γ  δ c ∪ γ )) = . . .

= i ◦ ((1α | 1β c ∪α ) | (1α∪(β∩γ c )∪δc | 1α∪(β∩γ c )∪(β c ∩γ )∪δc ))

and apply the formulae for the blurred and sharper iteration from the preceding
theorem. 

In [10] the formula for the blurred iteration has been proposed as one concrete
example, in [11] the formulae for both types of iteration appear as concrete examples.
But here we have shown that there do not exist more. We will compare these two
types of iteration in the following

Remark 6 (Comparison between blurred and sharper iteration in B̃) Let us write
the results of the two types of iteration from the preceding theorem as intervals:

( (A1  B1 ) ˜| (A2  B2 ) ) = [ αk , βk ] ,

where k = 1 resp. k = 2 correspond to the blurred resp. sharper iteration. Then the
following assertions can be established as an exercise. It follows that
(i) in general, α1 ⊆ α2 ⊆ β2 ⊆ β1 , i.e. [α2 , β2 ] ⊆ [α1 , β1 ],
(ii) particularly, [α2 , β2 ] = [α1 , β1 ] if and only if B2 = Ω, where [αk , βk ] = (A1 
B1 ∩ A2 ),
(iii) α2 = β2 if and only if B2 ⊆ B1 ∩ A2 , where {α2 } = (B1 → A1  Ω) ⊆ (A1 
B2 ) = [α1 , β1 ].

Part (i) of the preceding remark shows that the two types of iteration lead to
conditional Boolean subsets which are not comparable with respect to the partial
ordering ≤ but with respect to the set-inclusion ⊆ in B̃. Moreover, this relation ⊆
can be interpreted as “sharper than” resp. ⊇ as “more blurred than”, because a sharper
interval has less Boolean subsets than a more blurred interval. In this sense, [α2 , β2 ]
as result of the sharper iteration is “sharper than” [α1 , β1 ] as result of the blurred
iteration.
In the former papers [10, 11, 20, 21] we used the identification of an interval [α, β]
with the ordered pair (α, β) of its endpoints α, β in an MV-algebra L. Therefore, the
singletons {α} = [α, α] are identified with the pairs (α, α) and we will refer to as
elements of the “diagonal” of L̃ in both cases. This will be used in the following

Remark and Notation 1 (Conditioning operator on B̃ restricted to its diagonal)

The conditioning operator ˜| on B̃ from the preceding theorem applied to elements
 Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 249

of the diagonal of B̃ permits to recover the conditional Boolean subsets via the
property
(D) ({A} ˜| {B}) = (A  B).

Particularly, it follows that ({Ω}˜|{Ω}) = (Ω  Ω) = {Ω} and ({∅}˜|{Ω}) = (∅ 

Ω) = {∅} are in the diagonal, but not the self-complemented element in B̃:
({∅}˜|{∅}) = (∅  ∅) = B.
Moreover, property (D) permits to rewrite the iterates of conditional Boolean subsets
from the preceding theorem also into the form

(({A1 }˜|{B1 }) ˜| ({A2 }˜|{B2 })) using the conditioning operator ˜| on B̃,

which result to be completely analogous to the iterates of conditional indicator func-

tions

((1A1 | 1B1 ) | (1A2 | 1B2 )) using the conditioning operator | on F1 .

Also the results of both iteration procedures are completely analogous.

Therefore, for both we use for short the following

formal notation of iteration: ((A1 | B1 ) | (A2 | B2 )) using here | as formal symbol.

Particularly, ({A}˜|{Ω}) = {A} and (1A | 1Ω ) = 1A both are written for short as
(A | Ω) = A.

Now, using the preceding notation, we can rewrite the results of the two corollaries
from the beginning of this section not only as iterates in F1 but at the same time as
iterates in B̃.

Remark 7 (Blurred vs. sharper iterations) For the blurred iterations the following
three special iterations lead to the same result.
(i) ((A | B) | D) = (A | B ∩ D),
(ii) (A | (B | D)) = (A | B ∩ D),
(iii) ((A | D) | (B | D)) = (A | B ∩ D).
For the sharper iterations property (i) remains valid, but the other two lead to the
following different results.
(ii)* (A | (B | D)) = (A | D → B),
(iii)* ((A | D) | (B | D)) = (D → A | D → B).
These tree special iterations were treated and discussed by many authors but mostly
in an ad hoc manner.
One of the references in which the first two are systematically discussed is [7]
at the end of Sect. 2.4. There the authors justified property (i) by the “convention
? | 0 = ? = ? | 1”. This corresponds precisely to the two properties C(0, 1) = 21 =
250 S. Weber

C( 21 , 21 ) valid for any compatible mean value function C on [0, 1]. Calabrese ([2])
took (i) as definition for ((A | B) | D).
In the following the authors made out clearly the two possible meanings for
(A | (B | D)), the first justified by “1 | ? = ? = 0 | ?” which corresponds to C1 and
leads to property (ii) and the second by “1 | ? = 1 and 0 | ? = 0” which corresponds
to C2 and leads to property (ii)* taken by Calabrese as definition for (A | (B | D)).
Based on these arguments the authors of [7] proposed (but without additional jus-
tification) two possible meanings for the general iteration ((A | B) | (D | E)). As
first definition for this they proposed (A | B ∩ D ∩ E) and called it “associative
definition” due to the equality ((A | B) | D) = (A | (B | D)). This is indeed our
blurred iteration corresponding to the choice C1 . But as second definition they
proposed (A | B ∩ (E → D)) which is quite different from our sharper iteration
(B → A | E → (B ∩ D)) we derived from the choice C2 .
For the third special iteration, Copeland (see e.g. the book [13]) took property (iii)
as definition for ((A | D) | (B | D)) which corresponds to our blurred iteration. It
seems that property (iii)*, which corresponds to the sharper iteration, has not been
considered in the literature.
Finally, let us observe that the iteration process proposed in [9], Sect. 8.1, is
completely different because its result does not remain in B̃ but is an interval of
elements of B̃.

6 Uncertainty Measures of Conditionals

Definition 5 (Uncertainty measures on MV-algebras) Let L = (L, , , ) be an

MV-algebra. Then we use the following notations.
(i) A function m : L → [0, 1] will be called an uncertainty measure if it satisfies
the following conditions:
(M1) m(0) = 0, m(1) = 1 (boundary conditions),
(M2) a ≤ b ⇒ m(a) ≤ m(b) (isotonicity).
(ii) An uncertainty measure m on L will be called, resp.
(M3) compatible (with the complement) if m(a ) = 1 − m(a) ,
(A) additive if m(a ˙ b) = m(a) + m(b) for all disjoint unions.
As in the preceding sections we use also here the same symbols 0 resp. 1 for the
universal lower resp. upper bound in the lattice L as well as for the real numbers in
the unit interval [0, 1].
The additivity of measures on MV-algebras has a clear meaning in analogy to
the additivity on Boolean algebras, for some details we refer to the biographical
Remark 6.7 in [21] and the references therein, particularly [14]. In contrast to this,
the additivity in more general structures is not so clear, for the case of Girard algebras
we refer to [21], its biographical Remark 6.8 and the references therein. Particularly,
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 251

in the process of extending the additivity on a finite MV-algebra L to L̃, there appears
the compatible mean value function M from the example (iv) in Sect. 2, see also one
of the following examples.
Returning now to the MV-algebras of conditionals, we will present the corre-
sponding results in the following

Theorem 6 (Uncertainty measures on B̃ and on F1 ) Let B = (B, ∩,c , ∪) be a

Boolean algebra of (crisp) subsets of the universe Ω and let B̃ resp. F1 be the cor-
responding MV-algebras of conditional Boolean subsets (A  B) resp. conditional
indicator functions (1A | 1B ). Then:
(i) Any uncertainty measure μ on B can be extended to uncertainty measures

μ̃ on B̃, given by μ̃(A  B) = M(μ(A ∩ B) , μ(Bc ∪˙ (A ∩ B))) ,

which are “generated by” mean value functions M on [0, 1], resp.

m on F1 , given by m = μ̃ ◦ i , i.e. m(1A | 1B ) = μ̃(A  B) ,

with the isomorphism i : F1 −→ B̃ from Sect. 4. Particularly, it follows that

m(1A ) = m(1A | 1Ω ) = μ̃(A  Ω) = μ(A) for all A ∈ B.

(ii) If, furthermore, M is a compatible mean value function and μ is a compatible

(uncertainty) measure, then μ̃ and m are compatible measures. Moreover, it
follows that

1
m(1A | 1B ) = μ̃(A  B) =for μ(B) = 0,
2
1
particularly m(1∅ | 1∅ ) = μ̃(∅  ∅) = .
2
(iii) An additive (probability) measure μ has a unique extension to an additive
measure μ̃ resp. m given by

1
m(1A | 1B ) = μ̃(A  B) = μ(A ∩ B) + · μ(Bc ) = (1A | 1B ) dμ ,
2 Ω

where μ̃ is generated by
x+y
M(x, y) = .
2
from the example (iii) in Sect. 2.
252 S. Weber

Proof (i) The boundary conditions (M1) follow directly:

m(1∅ | 1Ω ) = μ̃(∅  Ω) = M(μ(∅), μ(∅)) = 0 ,

m(1Ω | 1Ω ) = μ̃(Ω  Ω) = M(μ(Ω), μ(Ω)) = 1.

The isotonicity condition (M2) follows from the isotonicity of μ and M and the
isomorphism i.
(ii) Also the compatibility condition is transfered from M and μ to μ̃ and m:

m(¬(1A | 1B )) = μ̃((A  B) ) = μ̃(Ac  B)

= M(μ(Ac ∩ B), μ(Bc ∪ Ac ))
= 1 − M(1 − μ(Bc ∪ Ac ), 1 − μ(Ac ∩ B))
= 1 − M(μ(A ∩ B), μ(Bc ∪ A))
= 1 − μ̃(A  B) = 1 − m(1A | 1B ).

(iii) was proved in [10] for μ̃, in a more general context. 

Remark 8 (The classical conditional probability) Another extension of an additive

measure μ on B is generated by the compatible mean value function M from the
example (ii) in Sect. 2, i.e. given by

x 1
M(x, y) = for (x, y) = (0, 1), M(0, 1) = ,
1+x−y 2

and leads to the classical “conditional probability”

μ(A ∩ B)
m(1A | 1B ) = μ̃(A  B) = μB (A) for μ(B) > 0, where μB (A) = .
μ(B)

Particularly, it follows that μ̃(B  B) = 1 for μ(B) > 0 and μ̃(B  B) = 21 for
μ(B) = 0.
By part (ii) of the preceding theorem this extension is a compatible measure. By part
(iii) it is not additive, but only “levelwise additive”:

˙ 1A2 | 1B ) = μ̃(A1 ∪˙ A2  B)
m(1A1 ⊕
= μ̃(A1  B) + μ̃(A2  B) (but only) for μ(B) > 0.

Example 2 (A further compatible measure extension) As already mentioned at the

beginning of this section, a further compatible measure extension μ̃ of an additive
measure μ on B is generated by the compatible mean value function M from the
example (iv) in Sect. 2, i.e. given by
y
M(x, y) = ,
1+y−x
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 253

and has, therefore, the following form

μ(A ∩ B) + μ(Bc )
m(1A | 1B ) = μ̃(A  B) = .
1 + μ(Bc )

But the extensions μ̃ resp. m are neither additive nor levelwise additive.

Example 3 (A non-compatible measure extension) Extensions μ̃ resp. m of an addi-

tive measure μ on B which are neither compatible (therefore not additive) nor lev-
elwise additive are those, where μ̃ is generated by the non-compatible mean value
function M from the example (v) in Sect. 2, i.e. given by

1
M(x, y) = (1 − t) · x + t · y with some t ∈ [0, 1], t = .
2
These have the following form

m(1A | 1B ) = μ̃(A  B) = μ(A ∩ B) + t · μ(Bc ) .

Remark 9 (The alternative model of PACS) The definition (CI) of conditional

indicator functions (1A | 1B ) in Sect. 3 requieres only the values C(0, 0) = 0,
C(1, 1) = 1 of any mean value function C as “determined truth values” and
C(0, 1) = 21 of any compatible C as “undetermined truth value”. If we drop the
assumption of compatibility and set C(0, 1) = t = 21 , we are led to the two expres-
sions
1 · 1A∩B + t · 1Bc + 0 · 1Ac ∩B for (1A | 1B ) ,

1 · 1A∩B + (1 − t) · 1Bc + 0 · 1Ac ∩B for ¬(1Ac | 1B ) ,

which should be equal because of the property (C5) in F1 , i.e. ¬(1A | 1B ) =

(1Ac | 1B ). But this property, generally accepted as a reasonable one, implies that
t cannot be a constant.
Therefore, instead of (CI) we are led to the alternative expression

(∗) 1A|B = 1 · 1A∩B + t(A | B) · 1Bc + 0 · 1Ac ∩B with some t(A | B) ∈ [0, 1] ,

where we used the notation 1A|B from [5] instead of our notation (1A | 1B ). Really,
in spite of the very similar expressions (CI) and (∗), the difference between them
is essential. While (CI) is a “measure-free” definition of conditionals, the approach
based on (∗) requires some knowledge or assumptions for the undetermined values
t(A | B). For instance, if we assume that the (reasonable) property 1 − 1A|B = 1Ac |B
holds then, applying (∗) to both 1A|B and 1Ac |B , it follows that t(Ac | B) = 1 − t
(A | B).
Indeed, several authors take (∗) as definition for a “truth-value of the conditional event
A | B” and show that t(A | B) fulfills the axioms of a general conditional probability,
254 S. Weber

see the above mentioned paper [5] and the references therein. In [6], models of
“Partial Algebraic Conditional Spaces” over a Boolean algebra B of subsets of the
universe Ω are axiomatically introduced and it is shown that they are in a one-to-one
correspondence to “De Finetti-Popper conditional probabilities”.
Now, if μ is a probability (measure) on the Boolean algebra B then it follows from
(∗) that

(∗∗) m(1A|B ) = 1A|B dμ = μ(A ∩ B) + t(A, B) · μ(Bc ) ,
Ω

compare with the result of the preceding example. On the other hand, it follows from
(∗∗) for 0 < μ(B) < 1 that

μ(A ∩ B) μ(A ∩ B)
m(1A|B ) = if and only if t(A | B) = ,
μ(B) μ(B)

i.e. in this setting we have the classical conditional probability t(A | B) = μB (A) =
m(1A|B ).

7 Conclusions

All authors treating measure-free conditioning agree that a conditional element inter-
preted as “a given b” for the (unconditional) elements a, b from some structured set
L should be defined using the elements interpreted as “a and b” resp. “b implies a”.
For an MV-algebra L there are two reasonable ways:
In the interval approach, a conditional element is defined as

(a  b) = [a ∧ b , b → a] ∈ L̃ for a, b ∈ L,

i.e. as the set of all elements between a ∧ b and b → a. Particularly, it follows that
(0  0) = L.
In the mean value approach, a conditional element is defined as

(a | b) = C(a ∧ b , b → a) ∈ L for a, b ∈ L,

i.e. as some element between a ∧ b and b → a. It is crucial that this approach requires
that L should have a self-complemented element u which results to be (0 | 0) = u.
Therefore, for a Boolean algebra L only the first approach is applicable. But for
a Boolean algebra L = B of subsets of a universe Ω, we can use the one-to-one
correspondence between Boolean subsets A ∈ B and its indicator functions 1A as
special fuzzy subsets and then apply the second approach. The present paper contains
three main results:
Conditioning for Boolean Subsets, Indicator Functions and Fuzzy Subsets 255

• The conditionals (1A | 1B ) form an MV-subalgebra F1 of the MV-algebra F of

fuzzy subsets, see Theorem 1.
• There is an MV-algebra isomorphism between F1 and B̃, see Theorem 3.
• There are exactly two types of iterations within F1 and, because of the isomor-
phism, also within B̃, see Theorems 4 resp. 5.

Acknowledgments I am very grateful to Peter Klement, as organizator and motor of his Linz
Seminars where I received a lot of stimulations since my first participation in 1983, and as friend
from our common time working together and in private occations.

References

1. Birkhoff, G.: Lattice Theory. Providence, RI (1960)

2. Calabrese, P.: An algebraic synthesis of the foundations of logic and probability. Inf. Sci. 42,
187–237 (1987)
3. Chang, C.C.: Algebraic analysis of many valued logics. Trans. Am. Math. Soc. 88, 467–490
(1958)
4. Cignoli, R.L.O., D’Ottaviano, I.M.L., Mundici, D.: Algebraic Foundations of Many-valued
Reasoning. Kluwer Academic Publishers, Dordrecht (2000)
5. Coletti, G., Scozzafava, R.: From conditional events to conditional measures: a new axiomatic
approach. Ann. Math. Artif. Intell. 32, 373–392 (2001)
6. Di Nola, A., Scozzafava, R.: Partial algebraic conditional spaces. Int. J. Uncertainty, Fuzziness
Knowl.-Based Syst. 12, 781–789 (2004)
7. Dubois, D., Prade, H.: Conditioning, non-monotonic logic and non-standard uncertainty mod-
els. In: Goodman, I.R., Gupta, M.M., Nguyen, H.T., Rogers, G.S. (eds.) Conditional Logic in
Expert Systems, pp. 115–158. North-Holland, Amsterdam (1991)
8. Galatos, N., Jipsen, P., Kowalski, T., Ono, H.: Residuated lattices: an algebraic glimpse at
substructural logics. Studies in Logic and the Foundations of Mathematics 151, Elsevier (2007)
9. Goodman, I.R., Nguyen, H.T., Walker, E.A.: Conditional Inference and Logic for Intelligent
Systems—A Theory of Measure-free Conditioning. North-Holland, Amsterdam (1991)
10. Höhle, U., Weber, S.: Uncertainty measures, realizations and entropies. In: Goutsias, J., Mahler,
R.P.S., Nguyen, H.T. (eds.) Random Sets: Theory and Applications, pp. 259–295. Springer,
Berlin (1997)
11. Höhle, U., Weber, S.: On conditioning operators. In: Höhle, U., Rodabaugh, S. (eds.) Mathe-
matics of Fuzzy Sets—Logic, Topology and Measure Theory, pp. 653–673. Kluwer Academic
Publishers, Dordrecht (1999)
12. Nguyen, H.T.: On representation and combinability of uncertainty. In: Abstracts of the 2nd
IFSA Congress, Tokyo (1987)
13. Pfanzagl, J.: Theory of Measurement, 2nd edn. Physica, Würzburg, Wien (1971)
14. Riečan, B., Mundici, D.: Probability on MV-algebras. In: Pap, E. (ed.) Handbook of Measure
Theory, vol. 2, pp. 869–909. North-Holland, Amsterdam (2002)
15. Van Gasse, B., Cornelis, C., Deschrijver, G., Kerre, E.E.: A characterization of interval-valued
residuated lattices. Int. J. Approximate Reasoning 49, 478–487 (2008)
16. Weber, S.: On conditional measures and events. In: Abstracts of the 9th Linz Seminar on Fuzzy
Set Theory (1987)
17. Weber, S.: Conditioning on MV-algebras and additive measures. I. Fuzzy Sets Syst. 92, 241–250
(1997)
18. Weber, S.: Conditioning on MV-algebras and additive measures, further results. In: Dubois,
D., Prade, H., Klement, E.P. (eds.) Fuzzy Sets, Logics and Reasoning about Knowledge, pp.
175–199. Kluwer Academic Publishers, Dordrecht (1999)
256 S. Weber

19. Weber, S.: Uncertainty measures—problems concerning additivity. Fuzzy Sets Syst. 160, 371–
383 (2009)
20. Weber, S.: A complete characterization of all weakly additive measures and of all valuations
on the canonical extension of any finite MV-chain. Fuzzy Sets Syst. 161, 1350–1367 (2010)
21. Weber, S.: Measure-free conditioning and extensions of additive measures on finite MV-
algebras. Fuzzy Sets Syst. 161, 2479–2504 (2010)
Multivalued Functions Integration:
from Additive to Arbitrary Non-negative
Set Function

Endre Pap

Abstract It is given a short overview of some integrals of multifunctions based

on additive measures, as strong, Aumann and Aumann-Gould integrals. It is con-
sidered also a multi-valued Choquet integral based on a multisubmeasure. Then it
is introduced a set-valued Gould type integral of multifunctions with values in the
family of all nonempty bounded subsets of a real Banach space X and with respect
to an arbitrary non-negative set function. There are given some basic properties of
the integrable multifunctions, and some continuity properties of the multimeasure
induced by set-valued integral.

1 Introduction

Theory of multifunctions, i.e., set-valued maps, correspondences, etc., is important

field of investigations as theoretical and practical applications [1, 2, 4, 5, 9–11, 20,
21, 32, 42, 55, 60]. It allows one to take into account the multiplicity of possible
choices, the lack of information and/or the uncertainty in a lot of situations ranging
from Optimal Control to Economic Theory, see [3, 31, 60]. In particular, measur-
able multifunctions, i.e., set-valued random variables, random sets, are investigated
in probability and statistics, with many applications, see first papers [38, 54]. Var-
ious types of integrals for multifunctions have many applications in mathematical
economics, theory of control, probabilities. Integrals of multifunctions can be used
as an aggregation tool when dealing with a large amount of information fusing and
with data mining problems such as programming and classification. In processes of
subjective evaluation, for instance, the integral of a multifunction can be a tool in
synthetic evaluation of the quality of a given object, when the score function may

E. Pap (B)
Singidunum University, Danijelova 32, 11000 Belgrade, Serbia
e-mail: [email protected]
E. Pap
Óbuda University, Becsi út 92, Budapest 1034, Hungary

© Springer International Publishing Switzerland 2016 257

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6_15
258 E. Pap

be set-valued i.e., for each quality factor there exists a multiple score or a set of
estimations.
A generalization of the classical integrals is related to their extension on non-
additive measures. For non-additive integrals see [8, 37, 39, 46, 47, 50]. Inte-
grals of multifunctions have been defined in different ways: Aumann method
[4, 7, 10], by the Rådström-Hörmander [53] embedding theorem [17], via Pettis
method [11, 19, 58, 59], using (as in Dunford [20]) sequences of multifunctions
[13, 16, 42], using finite or countable sums that generalize the Riemann sums
[5, 6, 10, 33], via Choquet or Sugeno integrals [14, 15, 41, 45, 49, 58–63], as
extension of pseudo-integral [28]. For survey on multivalued integrals of Aumann
and Debreu types see [55]. The Gould integral was defined in [27] via finite sums for
real functions relative to a finitely additive vector measure μ. Different generaliza-
tions of the Gould integral were introduced and studied in [22–25, 51, 56] for real
functions and in [52] for multifunctions relative to a finitely additive vector measure
(via Aumann method). In this paper we define and study another set-valued Gould
type integral of multifunctions (taking values in the space of all nonempty bounded
subsets of a real Banach space) based on an arbitrary non-negative set function.
We define the integral as a limit of a net of finite integral sums. Many important
properties of the integral work with an arbitrary non-negative set function μ without
supplementary conditions on μ.
This paper is organized as follows. In Sect. 2 we collect some facts about families
of subsets of Banach space, set functions and general multimeasures. Section 3 deals
with the integration of strongly measurable multifunctions and, more particularly,
with those that can be approximated by simple measurable multifunctions in the
sense of the Hausdorff distance, the Aumann integral, whose construction is based on
integrable selections and which has become the most popular set-valued integral, and
Aumann-Gould integral. In Sect. 4 is presented a multi-valued Choquet integral based
on multisubmeasure. In Sect. 5 we define a Gould type integral for multifunctions
relative to an arbitrary non-negative set function and point out some relationships
between integrability and total measurability. There are given some basic properties
of the integrable multifunctions, and some continuity properties of the multimeasure
induced by set-valued integral. Section 6 is for the conclusion.

2 Preliminaries

2.1 Families of Subsets of Banach Space

Let T be an abstract nonempty set, P(T ) the family of all subsets of T , A an algebra
of subsets of T, (X,  · ) a real Banach space with the metric d induced by its norm,
P0 (X ) the family of all nonempty subsets of X , Pb (X ) the family of all nonempty
bounded subsets of X , Pbc (X ) the family of all nonempty bounded convex subsets
of X , Cb (X ) the family of all nonempty bounded closed subsets of X , Cbc (X ) the
Multivalued Functions Integration: from Additive … 259

family of all nonempty bounded closed convex subsets of X and Kc (X ) the family
of all nonempty compact convex subsets of X .
For every A, B ∈ P0 (X ) and every α ∈ R, let

A + B = {x + y | x ∈ A, y ∈ B},
α A = {αx | x ∈ A}.

We denote by A the closure of A with respect to the topology induced by the norm
of X . Let + be the Minkowski addition on P0 (X ), i.e.,

A + B = A + B (A, B ∈ P0 (X )).

Let h be the Hausdorff metric given by

 
h(A, B) = max sup d(x, B), sup d(x, A) (A, B ∈ P0 (X )),
x∈A x∈B

and d(x, B) = inf d(x, y). (Cb (X ), h) and (Kc (X ), h) are complete metric spaces,
y∈B
see [32]. We denote |A| = h(A, {0}), for every A ∈ P0 (X ), where 0 is the origin of
X.
Relations of the Hausdorff metric h with respect to the operations + and +
are given by the following inequalities. If A, B, C, D, Ai , Bi ∈ P0 (X ), for every
i ∈ {1, . . . , n} and n ∈ N, then
 n 
 
n 
n
h Ai , Bi  h(Ai , Bi ),
i=1 i=1 i=1

h(α A + β B, γ A + δ B)  |α − γ | · |A| + |β − δ| · |B| (α, β, γ , δ ∈ R).

Definition 1 A partition of T is a finite family π = {Ai | i = 1, . . . , n} ⊂ A such


n
that Ai ∩ A j = ∅, i = j and Ai = T.
i=1

Then, as usual, we have that for two partitions π = {Ai | i = 1, . . . n} and π  =

{B j | j = 1, . . . m} of T we say that π  is finer than π , denoted π   π , if for every
j = 1, . . . , m, there exists i j = 1, . . . , n, so that B j ⊆ Ai j . The common refinement
of two partitions π = {Ai | i = 1, . . . n} and π  = {B j | j = 1, . . . m} is the partition

π ∧ π  = {Ai ∩ B j | i = 1, . . . , n; j = 1, . . . , m}.

We denote by P the class of all partitions of T and if A ∈ A is fixed, by P A we

denote the class of all partitions of A.
260 E. Pap

2.2 Set Functions

We have by [46].
Definition 2 Let μ : A → [0, ∞] be a non-negative set function with μ(∅) = 0.
We say that μ is:
(i) monotone if μ(A)  μ(B), for every A, B ∈ A , with A ⊆ B;
(ii) subadditive if μ(A ∪ B)  μ(A) + μ(B), for every A, B ∈ A , with A ∩
B = ∅;
(iii) submeasure if μ is monotone and subadditive;
∞
(iv) σ -subadditive if μ(A)  μ(An ), for every sequence of pairwise disjoint
n=1

∞
sets (An )n∈N ⊂ A , with A = An ∈ A ;
n=1
(v) finitely additive if μ(A ∪ B) = μ(A) + μ(B) for every disjoint A, B ∈ A ;
n
(vi) measure if lim μ(Ak ) = μ(A), for every sequence of pairwise disjoint
n→∞ k=1

∞
sets (An )n∈N ⊂ A , with A = An ∈ A ;
n=1
(vii) increasing convergent if lim μ(An ) = μ(A), for every increasing sequence
n→∞
∞
of sets (An )n∈N ⊂ A (i.e. An ⊂ An+1 , for every n ∈ N), with ∪ An = A ∈ A
n=1
(denoted by An  A);
(viii) decreasing convergent if lim μ(An ) = μ(A), for every decreasing sequence
n→∞
∞
of sets (An )n∈N ⊂ A , i.e. An+1 ⊂ An , for every n ∈ N, with ∩ An = A ∈ A
n=1
(denoted by An  A), and μ(A1 ) < ∞;
(ix) order-continuous (shortly, o-continuous) if lim μ(An ) = 0, for every decreas-
n→∞
ing sequence of sets (An )n∈N ⊂ A , with An  ∅.;
(x) exhaustive if lim μ(An ) = 0, for every sequence of pairwise disjoint sets
n→∞
(An )n∈N ⊂ A .
We introduce two types of variations of set function based on [46].
Definition 3 We consider the following set functions associated to an arbitrary set
function μ : A → [0, ∞] with μ(∅) = 0:
(i) μ (the disjoint variation of μ) defined, for every A ∈ A , by


n
μ(A) = sup μ(Bi ) ,
i=1

where the supremum is considered over all finite partitions {Bi }i=1 n
of A,
with Bi ∈ A , for every i ∈ {1, . . . , n}. μ is said to be of finite variation if
μ(T ) < ∞.
Multivalued Functions Integration: from Additive … 261

(ii) 
μ defined, for every A ⊆ T , by

μ(A) = inf{μ(B) | A ⊆ B, B ∈ A }.


The properties of set functions μ and μ̃ are investigated in details in [46]. If we

deal with Kc (X )-valued multifunctions, then the Minkowski addition + becomes
the usual addition of sets +, i.e., A + B = {x + y|x ∈ A, y ∈ B}).

Definition 4 Let μ : A → [0, ∞] be an arbitrary set function. We say that a prop-

erty (P) holds μ -almost everywhere (briefly, μ-ae) if there exists A ∈ P(T ), with

μ(A) = 0, such that the property (P) is valid on T \A.

Definition 5 Let ν, μ : A → [0, ∞] be two arbitrary set functions. We say that ν

is absolutely continuous with respect to μ if for every ε > 0, there is δ > 0, such
that for any A ∈ A , with μ(A) < δ, we have ν(A) < ε (denoted by ν  μ).

Definition 6 Let f, f n : T → R, n ∈ N. The sequence ( f n )n∈N is said to be:

(i) convergent in μ-measure if for every ε > 0, {t ∈ T ; | f n (t) − f (t)| > ε} ∈ A
μ
and lim μ({t ∈ T ; | f n (t) − f (t)| > ε}) = 0 (denoted by f n −
→ f ),
n→∞
(ii) convergent almost everywhere if there exists A ∈ A , with μ(A) = 0, such that
ae
lim f n (t) = f (t) for every t ∈ T \A (denoted by f n −
→ f ).
n→∞

For more details on convergences see [40].

2.3 General Multimeasures

The theory of set-valued measures (multimeasures) whose values are subsets of some
Banach space, under additivity condition, was investigated in [2, 12, 18, 26]. Some
of the motivations were the applications to mathematical economics and to statistics.
To extend the notion of σ -additivity, there are at least three possible definitions
depending on the series summability concept considered in the space of closed sets,
see [30].
Here we shall consider general multimeasures with some additional properties.
We have by [22].
Definition 7 Let M : A → P0 (X ) be a set multifunction, with M(∅) = {0}. M is
said to be:
(i) monotone if M(A) ⊆ M(B), for every A, B ∈ A , with A ⊆ B;
(ii) additive multimeasure if M(A ∪ B) = M(A) + M(B), ∀A, B ∈ A , A ∩ B =
∅;
(iii) absolutely continuous with respect to μ if for every ε > 0, there is δ > 0
such that for every A ∈ A with μ(A) < δ, we have |M(A)| < ε (denoted by
M  μ);
262 E. Pap

(iv) increasing convergent if lim h(M(An ), M(A)) = 0, for every increasing

n→∞

∞
sequence of sets (An )n∈N ⊂ A , with An = A ∈ A ;
n=0
(v) decreasing convergent if lim h(M(An ), M(A)) = 0, for every decreasing
n→∞

∞
sequence of sets (An )n∈N ⊂ A , with An = A ∈ A , and M(A1 ) < ∞;
n=0
(vi) order-continuous (shortly, o-continuous) if lim |M(An )| = 0, for every
n→∞
decreasing sequence of sets (An )n∈N ⊂ A , with An  ∅;
(vii) exhaustive if lim |M(An )| = 0, for every sequence of pairwise disjoint sets
n→∞
(An )n∈N ⊂ A ;  

n
(viii) h-multimeasure if lim h M(Ak ), M(A) = 0, for every sequence of
n→∞ k=0

∞
pairwise disjoint sets (An )n∈N ⊂ A , with A = An ∈ A ;
n=0
(ix) of finite variation if M(T ) < ∞, where M is defined for every A ∈ A by
 n 

M(A) = sup |M(Bi )| ,
i=1

where the supremum is considered over all finite partitions {Bi }i=1
n
of A, with
Bi ∈ A , ∀i ∈ {1, . . . , n};
(x) null-additive if for every A, B ∈ A with M(A) = {0} we have M(A ∪ B) =
M(B);
(xi) weakly null-additive if for every A, B ∈ A with M(A) = M(B) = {0} we
have M(A ∪ B) = {0}.

Remark 1 The theory of random sets is closely related to multimeasures, especially

when X is finite dimensional, see [35, 43, 44]. There are lot of important probabilistic
and statistical applications, e.g., to Stereology and to Image Processing.

3 Some Multivalued Additive Integrals

3.1 Strong Integral

Let (T, A , p) be a probability space. A multifunction is a map, defined on T, whose

values are subsets of some given set. In this section, we shall restrict our attention
to the space Cb (X ) of closed bounded subsets of X, endowed with the topology
τ H generated by the Hausdorff metric h. Further, we consider the Borel σ -field
B(Cb (X ), τ H ) generated by the τ H -open subsets of Cb (X ). A multifunction F :
T → Cb (X ) is said to be strongly A -measurable (or simply, strongly measurable)
Multivalued Functions Integration: from Additive … 263

if, for every member W of B(Cb , τ H ), one has F −1 (W ) ∈ A . Multifunctions that

enjoy some measurability property are also called “random sets”.
We start by defining the set-valued integral of a simple multifunction, i.e., of a
multifunction F assuming only a finite number of values, see [30].
Definition 8 Let {A1 , . . . , Ak } be an A -measurable partition of T and let F : T →
Cb (X ) be a measurable nmultifunction taking on the value K i ∈ Cb for any t ∈ Ai (i =
1, . . . , k), i.e., F = i=1 1 Ai K i , where 1 Ai denotes the characteristic function of Ai .
Then the integral (or expectation) of F is the member of Cb defined by


n
E(F) = p(Ai )K i (1)
i=1

When X is finite dimensional, the closure operation is not necessary. Given a sub-
space C  of Cb , we denote by L 1 (C  , A ) the class of strongly A -measurable
multifunctions with values in C  such that E||F|| <, where E is the integral given
by (1), and by S (C  , A ) the subclass of L 1 (C  , A ) whose members are strongly
A -measurable simple multifunctions.
Now, we construct the integral of a strongly measurable multifunction F : T →
C  , where C  is a τ H -separable subspace of Ccb . Without restriction, we assume
that C  is τ H -closed, and stable under the Minkowski addition and multiplication by
positive scalar.
Definition 9 For a strongly A -measurable simple multifunction F : T → C  , we
define the map Φ : S (C  , A ) → C  by

Φ(F) = E(F) (2)

where E(F) is defined by (1).

Now, we can extend the integral on the space of multifunctions whose members
can be approximated by simple functions. We denote by Lh1 (C , A ) the subclass of
L 1 (C  , A ) of those F that can be approximated by simple multifunctions, i.e., such
that one can find a sequence (Fn )n∈N in S (C  , A ) such that

lim h(F(t), Fn (t)) = 0 p-a.e.

n→∞

We have by [29].
Theorem 1 The map Φ : S (C  , A ) → C  given by (2) is extended to a map Φ̃
from L 1 (C  , A ) into C  (called strong integral) with the following properties

(i) Φ̃(F + G) = Φ̃(F) + Φ̃(G).

(ii) Φ̃(a F) = a Φ̃(F), for a  0.
(iii) h(Φ̃(F), Φ̃(G))  Φ̃(h(F, G)), specially ||E(F)||  E||F||.
264 E. Pap

Remark 2 The preceding approach for defining the set-valued integral is explicit, in
that it starts with simple multifunctions and uses only elementary operations such as
the Minkowski addition and the scalar multiplication. An alternative approach for
constructing the set-valued conditional expectation for a compact convex multifunc-
tion is given in [17].

3.2 Aumann Integral

Let (T, A , p) be a probability space and B be a sub-σ -field of A . By L 0 (T, B, p; X )

we denote the space of all (classes of) measurable functions from (T, B) into
(X, B(X )). For every F ∈ M (C (X )) (the set of closed valued multifunctions) we
define

S (F, A ) = f ∈ L 0 (T, B , p; X ) | f (t) ∈ F(t), for p-almost every t ∈ dom(F) .

L 0 (T, A , p; X ) endowed with the topology of convergence in probability is a metriz-

able topological vector space, provided one identify two functions that coincide p-
almost everywhere. Since a sequence converging in probability admits an almost
everywhere converging subsequence, then for any sub-σ -field B of A , the set
S (F, B) is closed in L 0 (T, B, p; X ). We denote by L 1 (T, A , p; X ) the subspace
of L 0 (T, A , p; X ), whose members are Bochner integrable. Given a sub-σ -field
B of A and a multifunction F, we define the following L 1 (X )-closed subset of
L 1 (T, A , p; X )

S 1 (F, B ) = f ∈ L 1 (T, A , p; X ) | f (t) ∈ F(t), for p-almost every t ∈ dom(F) .

The following notion of integral for multifunctions was introduced by Aumann [4].

Definition 10 For any measurable multifunction F and any sub-σ -field B of A ,

the set-valued (Aumann) integral of F over T, with respect to B, is denoted by
I (F, B) and defined by
 
I (F, B) = f dp | f ∈ S (F, B) .
1
T

The relation between strong and Aumann integral is given in the following theorem.
Theorem 2 Let F ∈ M (C (X )). If we assume in addition that F is integrable, then
the following property hold. If F is integrably bounded and takes on its values in a
τ H -separable subspace of Cbc (X ), then one has

Φ̃(F) = I (F)

where Φ̃(F) denotes the integral as defined in Sect. 3.1 (Theorem 1).
Multivalued Functions Integration: from Additive … 265

3.3 Aumann-Gould Type Integral

The Gould integral was introduced in [27] using finite sums for real functions with
respect to a finitely additive vector measure μ. If μ is a countably additive vector
measure of finite variation, then the Gould integral coincides with the Dunford inte-
gral and the Gelfand-Pettis integral. Different generalizations of the Gould integral
were introduced and studied in [22–25, 51, 56]. We consider in this section a Gould
integral of multifunction with respect to a finitely additive multimeasure.

Definition 11 If μ : A → X is a finitely additive measure, we say that μ is a selector

of a multimeasure M : A → P0 (X ) if μ(A) ∈ M(A), for every A ∈ A . We denote
by S M the family of all selectors of M.

We have by [52].

Definition 12 Let M : A → Pwkc (X ) (nonempty weakly compact convex subsets

of X ) be a finitely additive multimeasure of bounded variation and f : T → R a
bounded function. We say that f is Gould integrable with respect  to M if there
exists a nonempty weakly compact convex subset of X, denoted by T f d M, which
satisfies the condition that for every ε > 0 there exists a partition πε of T such that,
for every partition π = {E i }i=0
n
finer than πε and for every choice of si ∈ E i , we have
 n  

h f (si )M(E i ), f dM < ε.
i=0 T


The set T f d M is called the M-Gould integral of f.

Definition 13 If A ∈ A , the set

  
f
(G) f dμ | μ ∈ S M
A

is called
 the Aumann-Gould integral of the function f : T → R on A, where
(G) f dμ denotes  of f with respect to the vector measure μ. We
 the Gould integral
denote it by (AG) A f d M. If (AG) A f d M = ∅, we say that f is Aumann-Gould
integrable on A.

Theorem 3 If M is an additive multimeasure, (T, A , M) is a complete measurable

space and the Banach space X has the Radon-Nikodym property, then any M-Gould-
integrable function f is Aumann-Gould-integrable and, moreover,
 
f d M = (AG) f d M, A ∈A.
A A
266 E. Pap

4 A Multi-valued Choquet Integral Based

on a Multisubmeasure

Jang et al. [34] have introduced a multi-valued Choquet integral for multifunction
taking values in the class of all closed, nonempty sets of the interval [0, ∞[ based
on a nonnegative monotone set function, for which the selectors are real Choquet
integrals.
We present in this section another multivalued Choquet integral, this time for non-
negative functions with respect to multisubmeasures taking values in Kc ([0, ∞[).
We can associate to M : A → Kc ([0, ∞[) two nonnegative set functions: μ1 (A) =
inf M(A), and μ2 (A) = sup M(A), for every A ∈ A , such that M(A) = [μ1 (A),
μ2 (A)], for every A ∈ A , and set of selectors for M is nonempty since contains the
submeasures μ1 and μ2 . Let be f : T → [0, ∞[ a measurable function. We denote
C
by S M ( f ) the set of all monotone set functions μ : A → [0, ∞[ which are selectors
of the multisubmeasure M with respect to which the function f is Choquet integrable.
Using a similar procedure as in Sect. 3.3 for Aumann-Gould integral we introduce
the following notion by [57].

Definition 14 Let M : A → Kc ([0, ∞[) be a multisubmeasure, f : T → [0, ∞[

a measurable function and A ∈ A . The Aumann-Choquet integral (shortly (AC)-
integral) of f on A with respect to the multisubmeasure M is the set
   
(AC) f d M = (C) f dμ | μ ∈ S M
C
(f) ,
A A

where (C) A f dμ is the usual Choquet
 integral. A function f is said to be Aumann-
Choquet integrable on A if (AC) A f d M = ∅.

We denote by L AC (M) the set of all Aumann-Choquet integrable functions with

respect to M on A.

Theorem 4 The multifunction


N (A) = (AC) f d M (A ∈ A ),
A

has the following properties.

(i) N is a monotone set multifunction.
(ii) If f ∈ L AC (M) and μ is weakly null-additive, then the multifunction defined
by N is weakly null-additive.
(iii) If f ∈ L AC (M) and μ is null-additive, then the multifunction defined by N is
null-additive.
Multivalued Functions Integration: from Additive … 267

5 Gould Type Integral of Multifunctions Based

on Arbitrary Nonegative Set Function

We present here some recent results without proofs, which can be found in [48].

5.1 Definition and Basic Properties

In this section we introduce a Gould type set-valued integral for Pb (X )-valued

multifunctions with respect to a non-negative set function and point out some rela-
tionships between total measurability and integrability. Suppose μ : A → [0, ∞[
is a non-negative set function, with μ(∅) = 0. We introduce total measurability and
Gould type integrability for multifunctions relative to a set function.

Definition 15 Let F : T → P0 (X ) be a multifunction. F is said to be  μ-totally

measurable (on T ) if for every ε > 0 there exists a partition πε = {Ai | i = 0, . . . , n}
μ(A0 ) < ε, and sup h(F(t), F(s)) < ε, for every i ∈ {1, . . . , n}.
of T such that 
t,s∈Ai
F is said to be 
μ-totally measurable on B ∈ A if the restriction F| B of F to B is
μ-totally measurable on (B, A B , μ B ).


Remark 3 (i) If F is  μ-totally measurable on T , then F is μ-totally measurable

on every A ∈ A .
(ii) Let F : T → Cc (R), F(t) = [ f (t), g(t)], ∀t ∈ T, where f, g : T → R are real
functions, such that f  g. Then F is 
μ-totally measurable if and only if f and
g are both 
μ-totally measurable. This result follows by the equality

h([x, y], [z, t]) = max{|x − z|, |y − t|} (x, y, z, t ∈ R, x  y, z  t).

For every multifunction F : T → Pb (X ) we denote


n
σ F,μ (π ) = F(ti )μ(Ai ) = μ(A1 )F(t1 ) + · · · + μ(An )F(tn )
i=1

for every partition π = {Ai | i = 1, . . . , n} of T and every ti ∈ Ai , i ∈ {1, . . . , n}.

If there is no doubt, we shall denote σ F,μ (π ) shortly by σ (π ).

Definition 16 A multifunction F : T → Pb (X ) is said to be μ-integrable (on T ) if

the net (σ (π ))π∈(P,) is convergent in (Cb (X ), h), where P, the set of all partitions
of T , is ordered by the order relation “” given in Sect. 2.1. If (σ (π ))π∈(P,) is
convergent, then its limit is called the integral of F on T with respect to μ, denoted by

(G) F dμ.
T
268 E. Pap

F is said to be μ-integrable on B ∈ A if the restriction F| B of F to B is μ-integrable

on (B, A B , μ B ).

We easily obtain by the Definition 16.


Proposition 1 (i) If (G) T F dμ exists, it is unique.
(ii) F is μ-integrable on T if and only if there exists a set I ∈ Cb (X ) such that
for every ε > 0, there exists a partition πε of T , such that for every other
partition of T , π = {Ai | i = 1 . . . , n}, with π  πε and every choice of points
ti ∈ Ai , i ∈ {1, . . . , n}, we have h(σ (π ), I ) < ε.

(iii) If F : T → Kc (X ) is Kc (X )-valued, then (G) T F dμ ∈ Kc (X ).
(iv) If μ(T ) = 0, then F is μ-integrable on T and (G) T F dμ = {0}.

A multifunction F : T → P0 (X ) is called bounded if there exists C  0 such

that |F(t)|  C, ∀t ∈ T. If F : T → P0 (X ) is bounded, then F is Pb (X )-valued.
The converse is not true.
Example 1 Let F : [0, ∞[→ Pb ([0, ∞[) be defined by F(t) = [t, t + 1] for
every t ∈ [0, ∞[. Then F is Pb ([0, ∞[)-valued, but F is not bounded since
sup |F(t)| = ∞.
t∈[0,∞[

5.2 Gould µ-Integrability of Multifunction

Theorem 5 Suppose F : T → Pb (X ) is a μ-integrable multifunction. Then

(i) for every A ∈ A , F is μ-integrable on A;
(ii) F is μ-integrable on A ∈ A if and only if F1 A is μ-integrable on T.

In the following, we present some results concerning the relation between 

μ-total
measurability and μ-integrability.

Theorem 6 Suppose μ : T → [0, ∞[ is finitely additive. If F : T → Pb (X ) is a

bounded 
μ-totally measurable multifunction, then F is μ-integrable.

Example 2 (i) Suppose μ is finitely additive. If F(t) = E ∈ Cbc (X ), for every

t ∈ T , then F is μ-integrable and

(G) F dμ = μ(T )E.
T


n
(ii) Generally, if F = E i · 1 Ai , E i ∈ Cbc (X ), i ∈ {1, . . . , n}, {Ai }i=1
n
⊂ A is
i=1
a partition of T and 1 Ai is the characteristic function of Ai , for every i ∈
{1, . . . , n}, then F is μ-integrable and
Multivalued Functions Integration: from Additive … 269

 
n
(G) F dμ = E i μ(Ai ).
T i=1

(iii) Let F : T → Pkc (R) be defined by F(t) = [ f 1 (t), f 2 (t)], ∀t ∈ T , where

f 1 , f 2 : T → R are real functions such that f 1  f 2 . Then F is Gould
μ-integrable if and only if f 1 and f 2 are both Gould μ-integrable as usual
functions. In this case we obtain
   
(G) F dμ = f 1 dμ, f 2 dμ .
T T T

Particularly, if f 1 = f 2 = f, then
  
(G) F dμ = f dμ .
T T

We have by [46].
Definition 17 If μ : A → [0, ∞[ is a non-negative set function, with μ(∅) = 0,
then a set A ∈ A is said to be an atom of μ if μ(A) > 0 and for every B ∈ A , with
B ⊆ A, we have μ(B) = 0 or μ(A\B) = 0.

Theorem 7 Let μ : A → [0, ∞[ be a submeasure of finite variation and F : T →

Pb (X ) a 
μ-totally measurable bounded multifunction. Then F is Gould μ-integrable
on every atom of μ.

5.3 Further Properties of Integrable Multifunctions

In this section we list some general properties of the integral introduced in Definition
16, based on [48]. Suppose μ : A → [0, ∞[ is a non-negative set function, with
μ(∅) = 0.

Theorem 8 Let F, G : T → Pb (X ) be Gould μ-integrable multifunctions. Then

we have the following statements.
(i) F + G is Gould μ-integrable and
  
(G) (F + G) dμ = (G) F dμ+ (G) G dμ.
T T T

(ii) For every α ∈ R a multifunction α F is Gould μ-integrable on T and

 
(G) α F dμ = α · (G) Fdμ.
T T
270 E. Pap

(iii) For every α  0 a multifunction F is Gould αμ-integrable on T and

 
(G) F d(αμ) = α · (G) F dμ.
T T

(iv) For F(t) ⊆ G(t), for every t ∈ T, we have

 
F dμ ⊆ G dμ.
T T

(v) Related Hausdorff metric we have

   
h (G) F dμ, (G) G dμ  sup h(F(t), G(t)) · μ(T ).
T T t∈T

(vi) Specially, related to absolute value we obtain

  
 
(G) F dμ  sup |F(t)| · μ(T ).
 
T t∈T

Theorem 9 Suppose μ1 , μ2 : A → [0, ∞[ are non-negative set functions, with

μ1 (∅) = μ2 (∅) = 0. Then we have the following statements.
(i) If F : T → Pbc (X ) is both μ1 -integrable and μ2 -integrable and μ : A →
[0, ∞[ is defined by μ(A) = μ1 (A) + μ2 (A), for every A ∈ A , then F is μ-
integrable and
  
(G) F d(μ1 + μ2 ) = (G) Fdμ1 + (G) F dμ2 .
T T T

(ii) Let μ1  μ2 , and F : T → Pb (R) be defined by F(t) = [0, f (t)], where f :

T → [0, ∞[ is a real function. If F is both μ1 -integrable and μ2 -integrable, then
 
(G) F dμ1 ⊆ (G) F dμ2 .
T T

5.4 Set Multifunction Induced by Set-Valued Integral

Suppose μ : A → [0, ∞[ is a non-negative set function with μ(∅) = 0 and F : T →

Pb (X ) is a μ-integrable multifunction. In this section, we present several properties
concerning the set multifunction M : A → Cb (X ) defined by

M(A) = (G) Fdμ, (A ∈ A ). (3)
A
Multivalued Functions Integration: from Additive … 271

Theorem 10 Let be B, C ∈ A , with B ∩ C = ∅. If F : T → Pb (X ) is μ-integrable

both on B and on C, then F is μ-integrable on B ∪ C and moreover
  
M(A ∪ B) = (G) F dμ = (G) F dμ+ (G) F dμ = M(A) + M(B).
B∪C B C

We list some basic properties of set multifunction M defined by (3).

Theorem 11 Let F : T → Pb (X ) be a μ-integrable multifunction
 and the set mul-
tifunction M : A → Cb (X ), defined by M(A) = (G) A F dμ, for every A ∈ A .
Then the following hold
(i) M is an additive multimeasure.
(ii) M  μ.
(iii) If μ is of finite variation, then M is of finite variation.
(iv) If μ is o-continuous (exhaustive respectively), then M is also o-continuous
(exhaustive respectively).
For monotone set function μ we obtain the following theorem.

Theorem 12 Suppose μ : A → [0, ∞[ is monotone. Let F, G : T → Pb (X ) be

 F is μ-integrable
bounded multifunctions on T such that  on T and F = G μ-ae.
Then G is μ-integrable on T and (G) T F dμ = (G) T G dμ.

Theorem 13 Suppose μ : A → [0, ∞[ is a submeasure of finite variation. Let F :

T → Pb (X ) be abounded μ-integrable multifunction and M : A → Cb (X ) defined
by M(A) = (G) A Fdμ, ∀A ∈ A . Then the following properties hold
(i) If μ is o-continuous (increasing convergent, decreasing convergent respectively),
then the same is M.
(ii) If μ is σ -subadditive, then M is an h-multimeasure.

For special multifunctions (interval valued) we can prove some additional properties.

Theorem 14 Suppose μ : A → [0, ∞[ is monotone. Let F : T → Pb (R) be

defined by F(t) = [0, f (t)], ∀t ∈ T, where f : T → [0, ∞[ is a real function and
let B, C ∈ A be such that B ⊆ C. If F is μ-integrable both on B and on C, then
 
(G) Fdμ ⊆ (G) Fdμ.
B C

Theorem 15 Let F : T → Pb (R) be a μ-integrable multifunction defined by

F(t) = [0, f (t)], ∀t ∈ T, where f : T → [0, ∞[ is a bounded  μ-totally measur-
able function. If μ is finitely additive and of finite variation, then

M(A) = (G) f dμ = |M(A)| (A ∈ A ).
A
272 E. Pap

6 Conclusion

It is given a short overview of basic integrals of multifunctions with respect to addi-

tive set functions, as strong integral, Aumann integral, Aumann-Gould integral. Then
it is presented an extension of Choquet integral based on multisubmeasure. It is intro-
duced a set-valued Gould type integral of Pb (X )-valued multifunctions with respect
to an arbitrary non-negative set function. There are presented important properties of
this integral. In our future works we shall investigate the relationships among this set-
valued Gould type integral and other set-valued integrals of Dunford, Pettis, Choquet,
Sugeno, Birkhoff, McShane, Henstock-Kurzweil. In this spirit it will be investigated
the extension of universal integral, introduced in [36], for multifunctions.

Acknowledgments This research was supported by the grant MNPRS 174009 and by the project
“Mathematical models of intelligent systems and their applications” which was supported by the
Provincial Secretariat for Science and Technological Development of Vojvodina.

References

1. Apreutesei, G.: Cauchy nets and convergent nets on semilinear topological spaces. Topology
Appl. 159, 2922–2931 (2012)
2. Artstein, Z.: Set-valued measures. Trans. Amer. Math. Soc. 165 (1972)
3. Aubin, J.P., Frankowska, H.: Set-Valued Analysis. Birkhäuser, Boston (1990)
4. Aumann, R.J.: Integrals of set-valued maps. J. Math. Anal. Appl. 12, 1–12 (1965)
5. Boccuto, A., Sambucini, A.R.: A McShane integral for multifunctions. J. Concr. Appl. Math.
2(4), 307–325 (2004)
6. Bongiorno, B., Pfeffer, W.F., Thomson, B.S.: A full descriptive definition of the gage integral.
Canad. Math. Bull. 39(4), 390–401 (1996)
7. Brink, H.E., Maritz, P.: Integration of multifunctions with respect to a multimeasure. Glasnik
Math. 35, 313–334 (2000)
8. Cao, Y.: Aggregating multiple classification results using Choquet integral for financial distress
early warning. Expert Syst. Appl. 38(7), 8285–8292 (2011)
9. Caponetti, D., Di Piazza, L., Kadets, V.: Description of the limit set of Henstock-Kurzweil
integral sums of vector-valued functions. J. Math. Anal. Appl. 421, 1151–1162 (2015)
10. Cascales, B., Rodriguez, J.: Birkhoff integral for multi-valued functions. J. Math. Anal. Appl.
297, 540–560 (2004)
11. Cascales, B., Kadets, V., Rodriguez, J.: The Pettis integral for multivalued functions via single-
valued functions. J. Math. Anal. Appl. 332, 1–10 (2007)
12. Costé, A.: Sur les multimesures valeurs fermes bornes d’un espace de Banach. C.R. Acad. Sci.
Paris 280, 567–570 (1975)
13. Croitoru, A.: An integral for multifunctions with respect to a multimeasure. An. Şt. Univ. “Al.
I. Cuza” Iaşi 49, 95–106 (2003)
14. Croitoru, A.: Fuzzy integral of measurable multifunctions. Iran. J. Fuzzy Syst. 9(4), 133–140
(2012)
15. Croitoru, A.: Strong integral of multifunctions relative to a monotone measure. Fuzzy Sets
Syst. 244, 20–33 (2014)
16. Croitoru, A., Godet-Thobie, C.: Set-valued integration in seminorm. I. Annals of University of
Craiova, Mathematics and Computer Science Series, vol. 33, pp. 16–25 (2006)
17. Debreu, G.: Integration of correspondences. In: Proceedings 5th Berkely Symposium on Math-
ematical Statistics and Probability II, Part. I, pp. 351–372 (1967)
Multivalued Functions Integration: from Additive … 273

18. Debreu, G., Schmeidler, D.: The Radon-Nikodym derivative of a correspondence. In: Proceed-
ings of the Sixth Berkeley Symposium of Mathematical Statistics and Probability (1971)
19. Di Piazza, L., Musial, K.: Set-valued Kurzweil-Henstock-Pettis integral. Set-Valued Anal. 13,
167–179 (2005)
20. Dunford, N., Schwartz, J.: Linear Operators I. General Theory. Interscience, New York (1958)
21. Frankowska, H.: An open mapping principle for set-valued map. J. Math. Anal. Appl. 127,
172–180 (1987)
22. Gavriluţ, A.: On some properties of the Gould type integral with respect to a multisubmeasure.
An. Şt. Univ. “Al. I. Cuza” Iaşi 52(1), 177–194 (2006)
23. Gavriluţ, A.: Fuzzy Gould integrability on atoms. Iran. J. Fuzy Syst. 8(3), 113–124 (2011)
24. Gavriluţ, A.: The general Gould type integral with respect to a multisubmeasure. Math. Slovaca
60(3), 289–318 (2010)
25. Gavriluţ, A., Iosif, A.E., Croitoru, A.: The Gould integral in Banach lattices. Positivity 19,
65–82 (2015)
26. Godet-Thobie, C.: Some results about multimeasures and their selections. In: Measure Theory
at Oberwolfach 1979, Lecture Notes in Mathematics, vol. 794, pp. 112–116. Springer (1980)
27. Gould, G.G.: Integration over vector-valued measures. Proc. London Math. Soc. 15, 193–205
(1965)
28. Grbić, T., S̆tajner-Papuga, I., S̆trboja, M.: An approach to pseudo-integration of set-valued
functions. Inf. Sci. 181(11), 2278–2292 (2011)
29. Hess, C.: Conditional expectation and martingales of random sets. J. Pattern Recognit. 32,
1543–1567 (1999)
30. Hess, C.: Set-valued integration and set-valued probability theory: an overview. in [47],
617–673
31. Hildebrand, W.: Core and Equilibria of a Large Economy. Princeton University Press (1974)
32. Hu, S., Papageorgiou, N.S.: Handbook of Multivalued Analysis, vol. I. Kluwer Academic
Publishers, Dordrecht (1997)
33. Hukuhara, M.: Integration des applications mesurables dont la valuer est un compact convexe.
Funkcialaj Ekvacioj 10, 205–223 (1967)
34. Jang, L.C., Kwon, J.S.: On the representation of Choquet integrals of set-valued functions, and
null sets. Fuzzy Sets Syst. 112, 233–239 (2000)
35. Kendall, M.G., Moran, P.A.P.: Geometrical Probability. Charles Griffin, London (1963)
36. Klement, E.P., Mesiar, R., Pap, E.: A universal integral as common frame for Choquet and
Sugeno integral. IEEE Trans. Fuzzy Syst. 18(1), 178–187 (2000)
37. Klement, E.P., Mesiar, R., Li, J., Pap, E.: Integrals based on monotone set functions. Fuzzy
Sets Syst. 281, 88–102 (2015)
38. Kudo, H.: Dependent experiments and sufficient statistics. Natur. Sci. Rep. Ochanomizu Univ.
4, 151–163 (1954)
39. Kuncová, K., Malý, J.: Non-absolutely convergent integrals in metric spaces. J. Math. Anal.
Appl. 401, 578–600 (2013)
40. Li, J., Mesiar, R., Pap, E., Klement, E.P.: Convergence theorems for monotone measures. Fuzzy
Sets Syst. 281, 103–127 (2015)
41. Liu, W.L., Song, X.Q., Zhang, Q.Z., Zhang, S.B.: (T) Fuzzy integral of multi-dimensional
function with respect to multi-valued measure. Iran. J. Fuzzy Syst. 9(3), 111–126 (2012)
42. Martellotti, A., Sambucini, A.R.: A Radon-Nikodym theorem for multimeasures. Atti Sem.
Mat. Fis. Univ. Modena 42, 579–599 (1994)
43. Matheron, G.: Random Sets and Integral Geometry. Wiley (1975)
44. Molchanov, I.S.: Limit Theorems for Unions of Random Closed Sets. Lecture Notes in Math-
ematics, vol. 1561. Springer (1993)
45. Narukawa, Y., Torra, V.: Multidimensional generalized fuzzy integral. Fuzzy Sets Syst. 160,
802–815 (2009)
46. Pap, E.: Null-Additive Set Functions. Kluwer Academic Publishers, Dordrecht (1995)
47. Pap, E.: Handbook of Measure Theory. Elsevier (2002)
274 E. Pap

48. Pap, E., Gavriluţ, A., Croitoru, A.: Gould type integral of multifunctions relative to non-negative
set functions (submitted)
49. Park, C.K.: Set-valued Choquet-Pettis integrals. Korean J. Math. 20(4), 381–393 (2012)
50. Pham, T.D., Brandl, M., Nguyen, N.D., Nguyen, T.V.: Fuzzy measure of multiple risk factors
in the prediction of osteoporotic fractures. In: Proceedings of the 9-th WSEAS International
Conference on Fuzzy Systems (FS’08), pp. 171–177 (2008)
51. Precupanu, A., Croitoru, A.: A Gould type integral with respect to a multimeasure, I/I. An. Şt.
Univ. “Al. I. Cuza” Iaşi 48, 165–200/ 49 (2003); 183–207 (2002)
52. Precupanu, A.M., Satco, B.: The Aumann-Gould integral. Medit. J. Math. 5, 429–441 (2008)
53. Rådström, H.: An embedding theorem for spaces of convex sets. Proc. Amer. Math. Soc. 3,
151–158 (1952)
54. Robbins, H.E.: On the measure of random set I. Ann. Math. Stat. 15, 70–74 (1944)
55. Sambucini, A.R.: A survey on multivalued integration. Atti Sem. Mat. Fis. Univ. Modena 50,
53–63 (2002)
56. Satco, B.: A Vitali type theorem for the set-valued Gould integral. An. Şt. Univ. “Al. I. Cuza”
Iaşi 51, 191–200 (2005)
57. Sofian-Boca, F-N.: A multi-valued Choquet integral with respect to a multisubmeasure. An.
Şt. Univ. “Al. I. Cuza” Iaşi 61, 1, 129–152 (2015)
58. Stamate, C.: Vector fuzzy integral. Recent Advances in Neural Network. Fuzzy Systems and
Evolutionary Computing, pp. 221–224 (2010)
59. Stamate, C., Croitoru, A.: Non-linear integrals, properties and relationships. Recent Advances
in Telecommunications, Signals and Systems (Proceedings of NOLASC’13), WSEAS Press,
118–123 (2013)
60. Tversky, A., Kahneman, D.: Advances in prospect theory: commulative representation of uncer-
tainty. J. Risk Uncertain. 5, 297–323 (1992)
61. Zhang, D., Wang, Z.: On set-valued fuzzy integrals. Fuzzy Sets Syst. 56, 237–241 (1993)
62. Zhang, D., Guo, C.: Fuzzy integrals of set-valued mappings and fuzzy mappings. Fuzzy Sets
Syst. 75, 103–109 (1995)
63. Zhang, D., Guo, C., Liu, D.: Set-valued Choquet integrals revisited. Fuzzy Sets Syst. 147,
475–485 (2004)
Author Index

B M
Barrenechea, Edurne, 137 Mesiarová-Zemánková, Andrea, 109
Bustince, Humberto, 137 Mundici, Daniele, 57

D P
Durante, Fabrizio, 157 Pagola, Miguel, 137
Pap, Endre, 257
Perrone, Elisa, 157
Petrík, Milan, 83
E
Esteva, Francesc, 71
S
Sempi, Carlo, 173
F Stupňanová, Andrea, 181
Fernandez, Javier, 137

T
G Trillas, Enric, 13
Godo, Lluís, 71
Gottwald, Siegfried, 1
Grabisch, Michel, 215 V
Vetterlein, Thomas, 83

H W
Höhle, Ulrich, 23 Weber, Siegfried, 233

K Y
Kolesárová, Anna, 181 Yager, Ronald R., 199

© Springer International Publishing Switzerland 2016 275

S. Saminger-Platz and R. Mesiar (eds.), On Logical, Algebraic, and Probabilistic
Aspects of Fuzzy Set Theory, Studies in Fuzziness and Soft Computing 336,
DOI 10.1007/978-3-319-28808-6