Fuzzy Logic An Introductory Course For Eng - Enric Trillas

Studies in Fuzziness and Soft Computing
Enric Trillas
Luka Eciolaza
Fuzzy Logic
An Introductory Course for Engineering
Students
Volume 320
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

Enric Trillas Luka Eciolaza
•
Fuzzy Logic
An Introductory Course for Engineering
Students
123
Enric Trillas Luka Eciolaza
European Centre for Soft Computing European Centre for Soft Computing
Mieres, Asturias Mieres, Asturias
Spain Spain
ISSN 1434-9922 ISSN 1860-0808 (electronic)

ISBN 978-3-319-14202-9 ISBN 978-3-319-14203-6 (eBook)
DOI 10.1007/978-3-319-14203-6
Library of Congress Control Number: 2014958577
Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
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
Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
This book is dedicated to Professor Lotfi
A. Zadeh, as a humble homage of the authors
in the 50 years of his paper ‘Fuzzy Sets’!
Preface
This book was thought as a non-conventional first course textbook in Fuzzy Logic
for engineers ending with an introduction to one of the most fruitful topics arisen
from it, Fuzzy Control. It is from the teaching’s strategy of the authors, summarized
by “Nothing can substitute the own homework of the student” from which it comes
its non-conventional character, partially manifested by the ‘continuous’ form of
presenting the considered topics by joining theoretical explanations and examples,
and not always following the typically mathematical style of ‘theorem-corolaries’.
Behind this strategy is the opinion that, at the university level, students and
professors ought to learn jointly, students do not wait to receive everything from the
professor’s lectures, but should read more than a single recommended textbook.
Consequently, this book is neither a manual with recipes to be uncritically applied,
nor it is directed to those that can be only interested in mathematical subtleties. The
reader should be aware that fuzzy logic is the study and computational management
of imprecision and non-random uncertainty, both with the highest accuracy and
precision possible at each case, that fuzzy logic is not fuzzy in itself.
Each university course requires a particular teaching tactic that not only depends
on the number of lecturing hours, but on the aim of the course and on the audience’s
characteristics. In particular, additional tutorials supplied by the professor are
essential for a good learning process. Tutorials in which other forms of considering
the course’s topics and more sophisticated problems can be proposed. This is at the
own hands of the professor.
The book just presents some basic mathematical models for fuzzy logic but
without the intention to just subordinate it to mathematics. Fuzzy logic is neither a
part of mathematics, nor even of logic, like Physics is not so. Notwithstanding,
what is paramount is the importance and usefulness of mathematical models in
experimental sciences and technology, as well as in computer science and computer
technology and, in particular, in Soft Computing, where fuzzy logic plays a pivotal
role. But the suitability of such models only can come from the success of its testing
against some reality, for instance, in true applications; applications play in the
techno-scientific world an analogous role to that of experimentation in natural
sciences. For instance, if the branch called ‘Fuzzy Control’ served not as a direct
vii
viii Preface
justification of fuzzy logic, the success fuzzy logic has in control applications can be
seen as a kind of experimentation to show its usefulness in the study of dynamical
systems linguistically described by systems of imprecise rules. Fuzzy logic is much
more than what is in this introductory textbook; its applications spread along many
domains of science and technology.
The reader should be early acquainted with the fact that, differently from clas-
sical bi-valuate logic, almost all in fuzzy logic are context-dependent and purpose-
driven. For instance, when representing a system of imprecise linguistic rules in
fuzzy terms, all the predicates, connectives, and conditionals in the rules should be
specified accordingly with its contextual meaning and the type of inference, for-
wards or backwards, to be done. In fuzzy logic not only everything is a matter of
degree, but its practice requires the art of designing systems by means of the
available theoretic armamentarium. The student should be conscious that a mistake
in the design process can conduct to solve a different problem than the targeted one.
It should never be forgotten that it is most important in research to not stop
questioning (Albert Einstein). Posing good questions (Isaac Rabi) whose answers
can result in fertile ones (Karl Menger) is what reveals the relevance of a researcher.
Contents
1 On the Roots of Fuzzy Sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 A Genesis of Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 L-Degree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Opposite, Negate, and Middle . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.1 Antonyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.2 Negations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.3 Antonyms and Negations . . . . . . . . . . . . . . . . . . . . . . . 17
1.2.4 Medium Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 AND/OR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.1 AND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.2 OR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4 Qualified, Modified, and Constrained Predicates . . . . . . . . . . . . 22
1.4.1 Qualified Predicates . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.4.2 Linguistic Modifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4.3 Constrained Predicates . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4.4 Group Meaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4.5 Synonims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5 Linguistic Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5.1 Fuzzy Partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.6 A Note on Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.6.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2 Algebras of Fuzzy Sets . . . . . . . . . . . . .... . . . . . . . . . . . . . . . . . 35

2.1 Introduction. . . . . . . . . . . . . . . . . .... . . . . . . . . . . . . . . . . . 35
2.1.1 Cartesian Product. . . . . . . . .... . . . . . . . . . . . . . . . . . 35
2.1.2 Extension Principle . . . . . . .... . . . . . . . . . . . . . . . . . 36
2.1.3 Preservation of the Classical Case . . . . . . . . . . . . . . . . . 37
2.1.4 Resolution . . . . . . . . . . . . .... . . . . . . . . . . . . . . . . . 38
ix
x Contents
2.2 The Concept of an ‘Algebra of Fuzzy Sets’. . . . . . . . . . . . . . . . 41

2.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.2 Algebras of Fuzzy Sets. . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.3 Non-contradiction and Excluded-Middle . . . . . . . . . . . . . 47
2.2.4 Decomposable Algebras . . . . . . . . . . . . . . . . . . . . . . . . 50
2.2.5 Standard Algebras of Fuzzy Sets . . . . . . . . . . . . . . . . . . 52
2.2.6 Strong Negations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.2.7 Continuous T-Norms and T-Conorms . . . . . . . . . . . . . . . 59
2.2.8 Laws of Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2.9 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.3 On Aggregating Imprecise Information . . . . . . . . . . . . . . . . . . . 77
2.3.1 Aggregation Functions . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.3.2 Ordered Weighted Means . . . . . . . . . . . . . . . . . . . . . . . 79
2.3.3 More on Aggregations . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.3.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3 Reasoning and Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.1 What Does It Mean “Logic”? . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.1.1 Logic and Consequence Operators . . . . . . . . . . . . . . . . . 83
3.1.2 Conjecturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2 Reasoning with Conditionals: Representation . . . . . . . . . . . . . . . 87
3.2.1 What is a Conditional? . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.2.2 The Case of Boolean Algebras . . . . . . . . . . . . . . . . . . . 88
3.2.3 Fuzzy Conditionals . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.3 Short Note on Other Modes of Reasoning . . . . . . . . . . . . . . . . . 98
3.4 Inference with Fuzzy Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4.1 Finite Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.4.2 Inference with Several Rules . . . . . . . . . . . . . . . . . . . . . 103
3.4.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.5 Deffuzification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.6 Rules and Conjectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.7 Two Final Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4 Fuzzy Relations . . . . . . . . . . . . . . . . . . . . . . . . .............. 117

4.1 What Is a Fuzzy Relation? . . . . . . . . . . . . . .............. 117
4.2 How to Compose Fuzzy Relations? . . . . . . . .............. 119
4.3 Which Relevant Properties Do Have a Fuzzy
Binary Relation? . . . . . . . . . . . . . . . . . . . . .............. 121
4.4 The Concept of T-State. . . . . . . . . . . . . . . . .............. 124
4.5 Fuzzy relations and -cuts . . . . . . . . . . . . . .............. 125
Contents xi
5 T-Preorders and T-Indistinguishabilities . . . . . . . . . . . . . . . . . . . . 131

5.1 Which Is the Aim of This Section? . . . . . . . . . . . . . . . . . . . . . 131
5.2 The Characterization of T-Preorders . . . . . . . . . . . . . . . . . . . . . 133
5.3 The Characterization of T-Indistinguishabilities . . . . . . . . . . . . . 134
6 Fuzzy Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

6.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.2 Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.2.1 Operations with Fuzzy Numbers . . . . . . . . . . . . . . . . . . 145
6.2.2 Operations with Triangular Fuzzy Numbers. . . . . . . . . . . 146
6.2.3 Note. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3 A Note on the Lattice of Fuzzy Numbers . . . . . . . . . . . . . . . . . 151
6.3.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.4 A Note on Fuzzy Quantifiers . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.4.1 Quantified Fuzzy Statements . . . . . . . . . . . . . . . . . . . . . 156
7 Fuzzy Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

7.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.2 The Concept of a Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.3 Types of Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.4 -Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.5 Measures of Possibility and Necessity. . . . . . . . . . . . . . . . . . . . 164
7.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.7 Probability, Possibility and Necessity . . . . . . . . . . . . . . . . . . . . 170
7.8 Probability of Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8 An Introduction to Fuzzy Control . . . . . . . . . . . . . . . . . . . . . . . . . 175

8.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.1.1 Note. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
8.2 Revising Conditional and Implications in Fuzzy Control . . . . . . . 179
8.2.1 Inference from Imprecise Rules . . . . . . . . . . . . . . . . . . . 180
8.2.2 Takagi-Sugeno of Order 1. . . . . . . . . . . . . . . . . . . . . . . 189
8.3 Control of Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . 190
8.3.1 State-Space Representation . . . . . . . . . . . . . . . . . . . . . . 190
8.3.2 Takagi-Sugeno Models for Control
of Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.3.3 Stability Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8.3.4 Parallel Distributed Compensation . . . . . . . . . . . . . . . . . 195
8.3.5 Piecewise Bilinear Model . . . . . . . . . . . . . . . . . . . . . . . 197
8.3.6 Vertex Placement Principle . . . . . . . . . . . . . . . . . . . . . . 199
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Chapter 1
On the Roots of Fuzzy Sets
As fuzzy sets were introduced by Zadeh in 1965, they were born closely linked with
imprecise predicates, that is, with names of non-precisely defined classes of objects.
Even more, most of the applications of Zadeh’s ideas are made with properties
the objects do verify in some degree between the two classical extremes 0 and 1,
respectively. Because of that, it is not at all odd to introduce fuzzy sets from some
considerations on how predicates are used in language. We will follow Wittgenstein’s
statement “The meaning of a word is its use in language”.
1.1 A Genesis of Fuzzy Sets
Usually, isolated words mean nothing. To mean something, words are to be used
in a given context and in a known way. Words do serve to describe perceptions, to
translate reasoning, and to show the reasons for judgements.
For example, what it is meant by the predicate unleaty? It is impossible to answer
this question, since nobody has never heard something like “this is unleaty”, “such
is leaty”, etc. Neither unleaty, nor leaty, are English words, nobody has used them
and they don’t appear in English dictionaries. Meaning is inherited by predicates P
only after being used, in some ground, by means of elemental statements ‘x is P’.
By now, unleaty has no meaning.
Words are introduced in a language by using them in concrete ways. For example,
the Spanish word ‘madre’ comes from the latin ‘mater’ that, at its turn, came from one
in older indo-european languages, but always used to name someone’s mother. Later
on, the word could take, by analogy, new meanings as it is, for example, ‘mother
country’, ‘goodmother’, ‘mother in law’, ‘mother of all wars’, ‘mother Nature’,
‘mother-of-pearl’, etc. The predicate leaty neither appears in English, nor in Spanish,
French, German, …, because it was never before used to name a property of the
elements in some class, like it is with tall, young, middle-aged, with people, high
with buildings or mountains, or heavy with metals, for example. It is only through
its use that predicates do acquire ‘meaning’.
© Springer International Publishing Switzerland 2015 1
E. Trillas and L. Eciolaza, Fuzzy Logic, Studies in Fuzziness
and Soft Computing 320, DOI 10.1007/978-3-319-14203-6_1
2 1 On the Roots of Fuzzy Sets
What is it meant by the predicate odd? In principle, it depends on where it is

used. With natural numbers n, ‘n is odd’ is used accordingly with the mathematical
definition/rule ‘n is odd if and only if once divided by 2 the rest is 1’. With people,
things, or situations, odd does coincide with the meaning of ‘strange’, ‘separated’,
and ‘not often’. In the first context, the predicate odd is precise or crisp, since natural
numbers are only odd or not odd, but in other contexts the predicate is imprecise or
fuzzy, since for example, ‘this is an odd book’, or ‘that is an odd event’, could admit
degrees depending on to what extent the book or the event can be qualified as odd.
When the predicate is imprecise there is not a perfect classification of the objects to
which it refers to.
Only after a predicate acquires meaning, concepts like ‘tallness’, ‘oddity’,
‘heaviness’, etc., appear in the corresponding language. Predicates do appear in
language after being used, and only after being used they can evolve in new con-
texts and give birth to concepts. Like it happened with ‘high’ and ‘highness’, and
‘royal highness’. Concepts like uncertainty come from a mother-predicate as it is
‘uncertain’ in this case.
Notice that there are not natural numbers that can be qualified as ‘very odd’, or
‘more or less odd’, but there are buildings that can be called to be so. If the predicate
is crisp, that is, it names just an either yes, or not, property, the application to it of a
linguistic modifier needs to be newly defined, but if the predicate is imprecise it does
not, since people immediately understand what it is meant by, for instance, ‘very odd’.
In what follows we will only deal with predicates P, the name of a property on
a previously given set X = {x, y, z, . . .}, and considering the use of P through the
elemental statements ‘x is P’, for all x in X , and accepting (à la Wittgenstein) that
the meaning of P is its use in the current language. Hence, the first problem to tackle
is placed by the following question, How the use, or meaning, of P on X can be
mathematically represented or modeled?
First, and to distinguish between two statements ‘x is P’ and ‘y is P’, for x = y
in X , let us suppose it is possible to decide when is ‘x is less P than y’, where x
shows the property named P less than y shows it.
Let us call P the relation in X given by
x P y ⇔ x is less P than y,
and suppose P is a preorder (enjoys the reflexive and transitive properties). That is,
• x P x, for all x in X
• If x P y, and y P z, then x P z.
The preordered set (X, P ) reflects the organization P induces in X , and the
preorder P is the primary use of P in X . The relation −1P is defined by ‘x P
−1
y ⇔ y P x’. Note that the relation P is, usually, empirically perceived; it could
also be called the perceptive meaning of P in X .
We will say that ‘x is equally P than y’ whenever x P y and y P x, and write
x = P y, with (= P ) = ( P ∩ −1 P ). Since, obviously,
1.1 A Genesis of Fuzzy Sets 3
• x = P x, for all x in X
• x =P y ⇔ y =P x
• x = P y, and y = P z imply x = P z,
the relation = P is an equivalence in X , and gives the quotient set X/ = P , of the
classes of equally-P elements. The predicate P is semi-rigid in X if X/ = P consists
in a finite number of classes. Of course, all predicates on a finite X are semi-rigid.
Example 1.1.1 Let it be X = {x1 , . . . , x5 }, and P a predicate inducing the preorder
given by the matrix with entries

1, if xi P x j
entr y (i, j) =
0, otherwise,
that is, ⎛ ⎞
1 0 1 0 1
⎜0 1 0 1 0⎟
⎜ ⎟
[ P ] = ⎜
⎜1 0 1 0 1⎟⎟
⎝0 1 0 1 0⎠
1 0 1 0 1
The quotient set X/ = P has the two classes {x1 , x3 , x5 } and {x2 , x4 }.
If, in the same X , it is Q with primary use defined by
xi Q x j ⇔ i j,
it results X/ = Q with the 5 classes {x1 }, {x2 }, {x3 }, {x4 }, {x5 }.

Example 1.1.2 In X = [0, 10], consider the predicate P = around five, with P
defined by
• If x, y ∈ [0, 5], x P y ⇔ x y
• If x, y ∈ (5, 10], x P y ⇔ y x
• If x ∈ [0, 5], y ∈ (5, 10], x and y are not P -comparable.
Obviously, P is a preorder, and

(x y)&(y x) : x = y
x =P y ⇔ ⇔ x = y,
(y x)&(x y) : y = x
that is, = P is the equality. Hence, X/ = P = {{x}; x ∈ [0, 10]}, and P is not semi-
rigid.
1.1.1 L-Degree
Let us suppose P in X given by its primary use P , and let (L , ) be a poset. An

L-degree, or L-measure, for P in X is a function μ P : X → L, such that if x P y,
then μ P (x) μ P (y)- in the order of the poset. The idea behind this definition is
that μ P (x) ∈ L evaluates up to which extent x is P, up to which extent x verifies the
property named by P. It can be written,
Degree up to which x is P = μ P (x) ∈ L ,
and said that the primary use of P is gradable in (L , ). If x = P y, then

μ P (x) = μ P (y), as is easily proven. Hence, μ P is constant in the equivalence’s
classes modulo = P .
Once μ P is known, it can be defined the relation μ P in X by
x μ P y ⇔ μ P (x) μ P (y),
with which it is P ⊂ μ P . When P = μ P , μ P perfectly reflects the primary use

of P in X .
Relation μ P is a preorder since it is obviously reflexive and transitive. The pair
( P , μ P ) reflects a use of P in X , and once P and (L , ) are fixed, there can
exist again several uses of P in X depending on the L-degree μ P . L-degrees are also
called L-x5.
When L = [0, 1], endowed with the linear order of the unit interval, the L-sets
are known as fuzzy sets. In this case, the triplet (X, P , μ P ) is a numerical quantity,
where of course, given P and X , neither the relation P , nor the measure μ P are in
general unique.
When L = {a + ib; 0 a 1 & 0 b 1} = C is endowed with the partial
order
a1 + ib1 a2 + ib2 ⇔ a1 a2 & b1 b2 ,
L-sets are ‘complex fuzzy sets’. Notice that (C, ) is isomorphic with the set of the
sub-intervals [a, b] ⊆ [0, 1], once partially ordered by
[a1 , b1 ] [a2 , b2 ] ⇔ a1 a2 & b1 b2 .
When it is only known that the degree up to which “x is P” belongs to some sub-
interval [a, b], it can be taken μ P (x) = a + ib. In science and technology, complex
quantities are not at all rare.
Remark 1.1.3 Provided L = [0, 1], the relation μ P is linear or total, since given
x, y in X , it is either μ P (x) μ P (y), or μ P (x) ≥ μ P (y); that is, either x μ P y,
or y μ P x.
Remark 1.1.4 The relation P is not always linear, there can exist elements x, y in X
such that it is neither x P y, nor y P x, that is, elements which are not comparable
under P . Of course, in the cases in which P is not linear and L = [0, 1], it cannot
be P = μ P , the degree cannot perfectly reflect the primary use of P, and the
relation μ P enlarges the primary use P with all the links between elements in X
contained in P − μ P , provided this difference-set is non empty.
Remark 1.1.5 When (L , ) is not linear, as it is with L = C the unit complex

interval, it is also not sure that μ P can coincide with the primary meaning, but
being not linear there can exist more possibilities for the coincidence.
Remark 1.1.6 Since in fuzzy logic it is always only considered the membership
function and not the primary meaning serving for its design, the working scientist
should not confuse the two relations μ P and P . For this reason, μ P can be called
the working-meaning of P in X , without forgetting that it can be bigger than the
primary meaning.
1.1.2 Fuzzy Sets
Once the use ( P , μ P ) of P in X is given for the poset (L , ), it will be said that
the new object P defined by
∼
• x ∈r P , if and only if r = μ P (x), for r ∈ L,
∼
• P = Q , if and only if μ P = μ Q
∼ ∼
is the L-set labeled P. The set L X = {μ; μ : X → L}, is usually and abusively
called the set of all L-sets in X , since it contains all the possible degrees in L.
Notice that a predicate P could give many L-sets, in dependance on which poset
(L , ), and which function μ P ∈ L X , are chosen.
From now on, it will be supposed that (l, ) has a minimum element α (α
r, ∀r ∈ L), and a maximum element ω(r ω, ∀r ∈ L). With L 0 = {α, ω}, (L 0 , )
is a poset isomorphic to ({0, 1}, ).
Since a classical subset A of X is characterized by its membership function

1, if x∈A
μ A (x) =
0, if x∈
/ A,
the subset A can be viewed as an L-set, whose function only takes the values α or ω,
that is,
ω, if x ∈ A
μ A (x) =
α, if x ∈ / A.
Hence, the classical subsets of X are nothing else than the functions in L 0X , a set
included in L X . Classical subsets are limiting or particular case of L-sets. They are
degenerate L-sets.
Notice that if the classical subset A is labeled by a crisp predicate P, this predicate
is semi-rigid since X/ = P has, at most, two classes. Crisp predicates are rigid.
Very often μ P is perceptually designed, that is, designed from what is perceived
or known on the concrete use of P in X . It can be thought that such design is purely
subjective, in the sense of being made just by what the designer believes on the use
of P. But this would not be the case, except if the designer proceeds in a non rational
way. The designer should try to be as most sure as possible on the correction of his/
her/its perceptions.
In a lot of cases, mainly in the applications, predicates P in X result indirectly

evaluable in ([0, 1], ) thanks to a numerical characteristic Ch P of the current use
of P in X . Such characteristic allows to translate ‘X is P’ into ‘Ch P (x) is Q’, with
an adequate predicate Q on the range of the function Ch P : X → R. For example,
if X is a population and P = old, it could be the case that Ch old = numerical age =
Age, in which case ‘x is old’ can be translated into ‘Age(x) is big’, with a modeling
of big according with the current use of old, and provided this predicate only depends
in the subject’s age. In this cases, once the designer is sure that Ch P and Q are good
enough for the case, and also that the order in the numerical interval where Ch P
ranges is adequate to model the order P by the order Q , he/she/it should again
be sure that the degrees μ Q (Ch P (x)) agree with the expected degrees μ P (x).
A design’s process never can arrive to something “exact”, but approximate and, if
possible, keeping everything under some bounds. For example, by taking the degrees
μ Q (Ch P (x)) in intervals (a(x), b(x)), where a(x) is the minimum of the acceptable
values for μ Q (Ch P (x)), and b(x) the maximum of them. This can allow to take
μ Q (Ch P (x)) as, for instance, an average of a(x) and b(x), or as the complex number
a(x) + ib(x). For example, if m x is the middle point of the interval (a(x), b(x)),
and the confidence that the value is between a(x) and m x can be quantified in a
coefficient a1 , and that of being between m x and b(x) by a2 , it can be taken the
average μ Q (Ch P (x)) = [a1 · a(x) + a2 · b(x)]/a1 + a2 .
A rational design should be carefully made by taking into account all the available
information, or knowledge, in the use of P in X , as well as of that induced on Q in
its numerical universe. If the use of P in X is not known, it is impossible to design
neither μ Q (Ch P ), nor μ P , nor to accept μ P (x) = μ Q (Ch P (x)), for all x ∈ X .
As it was said, if [x] is a class in X/ = P , μ P is constant in it. Let us denote by vx
the value of μ P in the class [x].
Provided P is semi-rigid with at most two classes in X/ = P , then either X/ = P =
{[x]}, or X/ = P = {[x], [y]}. In the first case, μ P only has a single value v ∈ L.
In the second case, μ P has at most two values, and, for obvious reasons, we will
only consider the situation where this values vx , v y are different. When,
• X/ = P = {[x]}, and either μ P (x) = α for all x in X , or μ P (x) = ω for all x in
X , it results P = ∅ in the first case, and P = X in the second. In both cases, P
∼ ∼
is a rigid or binary predicate in X .
• X/ = P = {[x], [y]}, and vx = v y , P is of the type

∼
In the particular case {vx , v y } = {α, ω}, P is of the type
{x}
{y}
1 2 3 4 X
that is, P is the classical or crisp subset [x] of X , and P is a rigid or binary predicate
∼
in X .
Let us show some examples.
Example 1.1.7 With P = around five, and P as shown in the example [2 in 1.2],
the function
⎧
⎨ 0, if x ∈ [0, 4] ∪ [6, 10]
μ P (x) = x − 4, if x ∈ [4, 5]
⎩
6−x if x ∈ [5, 6],
whose graphics is
verifies: x P y ⇒ μ P (x) μ P (y), and hence μ P is an [0, 1]−degree for P in

[0, 10], that originates a fuzzy set P of the numbers in [0, 10] that are around five.
∼
Obviously, this degree does not perfectly reflect the primary use of P.
Example 1.1.8 Consider P = big in X = [0, 10]. Let us show several possible
degrees for it, after agreeing that ‘x P y if and only if x y, in the linear order of
[0, 10]’.
A [0, 1]-degree μ P is any function X = [0, 10] → [0, 1], such that
If x y, then μ P (x) μ P (y),
that is, any non-decreasing function (of which there are many). We can also agree
that μ P (0) = 0, and μ P (10) = 1. With this, it is clear that all degrees for big will
show some family resemblance.
Once fixed (L , ) = ([0, 1], ), and P = , the different uses of big only
depend on which function μ P is chosen to reflect the meaning of the predicate in
[0, 10]. Of course, it is P = μ P if and only if function μ P is strictly non-decreasing,
as it is the case either with μ P (x) = x/10, or with μ P (x) = x 2 /100. Provided μ P
is not strictly non-decreasing as, for example, with
since 6 P 4 (strictly), but μ P (6) = μ P (4) = 0.5, μ P does not perfectly reflect the
primary use of big in [0, 10]. In the same way, the crisp degree

1, if x > 8
μ P (x) =
0, otherwise,
does not perfectly reflect the primary use of big when translated into ‘after eight’.
Another possible model for μ P is
⎧
⎨ 0, if x ∈ [0, 2]
μ P (x) = 6 ,
x−2
if x ∈ [2, 8]
⎩
1 if x ∈ [8, 10],
with graphic.
All these models are linear, with the exception os μ P (x) = x 2 /100, that is
quadratic. Another quadratic models are given by μ(1)
P = ( 6 ) , and μ P =
x−2 2 (2)
(1)
1 − μ P (10 − x) = 1 − ( 8−x
6 ) , with graphics
2
Finally, the following two models are hyperbolic,

⎧
⎪
⎨ 0, if x ∈ [0, 2]
μ(3)
P (x) = 7
· x−1 ,
x−2
if x ∈ [2, 8]
⎪
⎩
6
1 if x ∈ [8, 10],
⎧
⎪
⎨ 0, if x ∈ [0, 2]
(4)
μ P (x) = − 16 ( x−9
7
+ 1), if x ∈ [2, 8]
⎪
⎩
1 if x ∈ [8, 10],
with graphics.
Example 1.1.9 The predicate old, once numerically characterized by Ch old = Age,
can be translated into the interval [0, 120], in years, by μold (x) = μbig (Age(x)),
and a linear model for it can be
⎧
⎪
⎨ 0, if 0 Age(x) 45
μold (x) = μbig (Age(x)) = 30 ,
x−45
if 45 Age(x) 75
⎪
⎩
1 if 75 Age(x) 120.
with graphic.
µold µbig Age
In cases like this one, it is important to notice that the function μold depends on
the values 45 and 75, as well as on the form of the curve in the sub-interval [45, 75].
It can be supposed, for example, that the above function is supplied by a person
in the range of the fifties, but that one in the seventies would design μold as the
curve
with μold increasing quadratically between the 50 and the 80 years.

Analogously, a person in the twenties could design μold as
Remark 1.1.10 There are different models for the uses of the same predicate P in X ,
and such uses are reflected in the corresponding models μ P in [0, 1] X . It is because of
this that it is actually important the process of designing the membership functions.
Example 1.1.11 Analogously to the case of big, the predicate A5 = around five in
[0, 10], can have non-linear but quadratic models, as the one given by
⎧
⎨ 0, if x ∈ [0, 4] ∪ [6, 10]
μ A5 (x) = (x − 4)2 , if x ∈ [4, 5]
⎩
(6 − x)2 if x ∈ [5, 6],
whose graphics is
Remark 1.1.12 Since each P in a set X can have different degrees μ P , at each
particular case the meaning of P should be well captured to not represent it by a
mistaken function that will translate a different use of the predicate.
Remark 1.1.13 Given μ P , and the L−set P , the degree is also called the membership
∼
function of the L-set. At this respect,
• x ∈α P , is classically written x ∈
/ P
∼ ∼
• x ∈ω P , is classically written x ∈ P .
∼ ∼
1.2 Opposite, Negate, and Middle
Very often the meaning of a predicate P is not captured without simultaneously

capturing one of its opposites a P (a for antonym, a synonym of opposite). How
can I recognize that John is young without the possibility of recognizing that Peter
is old? Can someone recognize that a person is tall but not that other person is
short?
The mastering of perception-based predicates shows this kind of polarity: we
jointly learn the meaning of P and some of its opposites a P. Even more, without
knowing how to use young and old it is not possible to know how to use middle-
aged, that is equivalent to ‘not young and not old’. The same could be said with
respect to warm that is equivalent with ‘not cold and not hot’, in relation with
water’s temperature. Composite predicates of this type are very frequent, for example
medium, actually equivalent to ‘not big and not small’.
It should be noticed that a ‘middle’ predicate only exists with imprecise predicates,
but not with precise ones. For example, in the set of natural numbers, if P = even,
it is a P = odd, and a(odd) = even, thus (not even) and (not odd) = odd and even,
but for no n it can be stated ‘n is odd and even’.
Let us remark that, although P and a P are linguistic terms, not P is not a linguistic
term. For example, in all dictionary we will find poor and rich, but neither not poor nor
not rich. The negate of P, not P, is more a logical concept than a linguistic one. Our
current problem is how to find the uses of a P(a P , μa P ), and not P(not P , μnot P ),
given a use ( P , μ P ) of P.
1.2.1 Antonyms
Concerning every opposite a P of P, this opposition is translated by
a P = −1
P
since ‘x is less a P than y’ should be equivalent to ‘y is less P than x’.

Hence, a(a P) = a−1P = (−1 P )
−1 = , that reflects a(a P) = P. For
P
example, with P = tall, it is a P = short and a(a P) = tall.
This property of a P shows a way for obtaining μa P once μ P is known. Let it
A : X → X be a symmetry on X , that is a function such that
• If x P y, then A(y) P A(x)
• A ◦ A = id X ,
and, once μ P : X → L is known, take μa P (x) = μ P (A(x)), for all x in X , that is,
μa P = μ P ◦ A.
Function μa P = μ P ◦ A is a degree for a P, since:
• x a P y ⇔ y P x ⇒ A(x) P A(y) ⇒ μ P (A(x)) μ P (A(y)),
1.2 Opposite, Negate, and Middle 13
and verifies,
• μa(a P) = μ(a P) ◦ A = (μ P ◦ A) ◦ A = μ P ◦ (A ◦ A) = μ P ◦ id X = μ P .
Then, for all symmetry A in X , we have an opposite for P. For example, in X =
[0, 10] with μbig (x) = 10
x
, and A(x) = 10 − x, it is μbig (10 − x) = 10−x
10 = 1 − 10 ,
x
and with a(big) = small, it can be said μsmall (x) = 1 − 10 .x
If big is represented by
⎧
⎨ 0, if x ∈ [0, 4]
(x−4)
μbig (x) = 4 , if x ∈ [4, 8]
⎩
1, if x ∈ [8, 10],
with the same symmetry A(x) = 10 − x, it results

⎧
⎨ 0, if x ∈ [6, 10]
(6−x)
μsmall (x) = μbig (10 − x) = 4 , if x ∈ [2, 6]
⎩
1, if x ∈ [0, 2],
graphically,
It is easy to prove that A(x) = 10 10−x

10+x is also a symmetry in [0, 10], since it
verifies,
• If x y, then A(y) A(x),
• A(A(x)) = x, for all x in [0, 10]
Hence, another opposite of big with μbig (x) = x, is μa big (x) = μbig (A(x)) =
A(x) = 10 10−x10+x . It gives a different representation for small, μsmall (x) = 10 10+x
10−x
in [0, 1].
The last two examples show a serious trouble. There are no points x ∈ X such
that it is simultaneously μ P (x) = 0 and μa P (x) = 0. The pairs of opposites (P, a P)
for which there is a region in X such that both μ P and μa P take the value 0, called
neutral region, are called regular opposites. Hence, the two above pairs (big, small)
are not regular. Nevertheless, if big is represented by

0, if 0x 7
μbig (x) = (x−7)
3 , if 7 x 10,
with A(x) = 10 − x, results

0, if 3 x 10
μsmall (x) = μbig (10 − x) = (3−x)
3 , if 0 x 3,
with graphics
that shows the neutral region (3, 7). This pair is regular.
1.2.2 Negations
Let it P be a predicate in X , and P = notP its negate. The only we can say about
the relation between P and P is that it is P ⊂ −1
P , since
If x is less P than y, then y is less not P than x,
or, equivalently, P ⊂ −1

P . We can also easily agree that,
• If μ P (x) = α, then μ P (x) = ω

• If μ P (x) = ω, then μ P (x) = α.
Let it N : L → L be a function such that
1. If a b, then N (b) N (a),
2. N (α) = ω, and N (ω) = α,
with such a function N , it is μ P = N ◦ μ P an L-degree for P , since
x P y ⇒ y P x ⇒ μ P (y) μ P (x)
⇒ N (μ P (x)) N (μ P (y)) ⇔ μ P (x) μ P (y).
Hence, given an L-degree μ P of P in X , with each function N verifying (1)

and (2), we get the L-degree μ P = N ◦ μ P . Such functions N are called negation
functions.
Provided the negation function does verify
3. N ◦ N = id L ,
then
μ(P ) (x) = N (μ P (x)) = N (N (μ P (x)))

= (N ◦ N )(μ P (x)) = id L (μ P (x)) = μ P (x),
for all x in X , or μ(P ) = μ P .

Functions N verifying (1), (2), and (3) are called strong negations, and are almost
the only used with fuzzy sets, for the following reason. With L = [0, 1], all strong
negations are, obviously, continuous, hence, if μ P is continuous also μ P is such,
and if μ P has some discontinuities, μ P has the same discontinuities. That is, strong
negations do not add discontinuities to those of μ P .
If L = [0, 1], there is a family of strong negations widely used in fuzzy set theory,
the so-called Sugeno’s negations:
1−a
Nλ (a) = , with λ > −1, for all a ∈ [0, 1].
1 + λa
For example, N0 (a) = 1 − a, N1 (a) = 1−a

1+a , N−0.5 = 1−0.5a ,
1−a
N2 (a) = 1−a
1+2a , etc.
Since obviously,
N λ1 N λ2 ⇔ λ 1 λ2 ,
it results:
• If λ ∈ (−1, 0], then Nλ N0
• If λ ∈ (0, +∞], then N0 < Nλ ,
graphically
μ
Nλ , λ ∈ (−1, 0)
N0
N λ , λ ∈ (0, + ∞)
Notice that, provided N is in the Sugeno’s family of strong negations, it is enough

to know a concrete pair of numbers (a, Nλ (a)) to compute the corresponding λ. For
example,
• If Nλ (0.5) = 0.5, it results λ = 0
• If Nλ (0.7) = 0.4, it results λ = − 14 5
• If Nλ (0.4) = 0.5, it results λ = 21 .

Example 1.2.1 With μ A5 as defined in 1.1.6, and with N0 , it is

⎧
⎪
⎨ 1, if x ∈ [0, 4] ∪ [6, 10]
μnot A5 (x) = 1 − μ A5 (x) 5 − x, if x ∈ [4, 5]
⎪
⎩
x −5 if x ∈ [5, 6],
with graphics.
Example 1.2.2 With μsmall as defined in the second case of 2.2, and with N1 it is
⎧
⎪
⎨ 1, if 6 x 10
1 − μsmall (x)
μnot small (x) = N1 (μsmall (x)) = = 10−x
x−2
, if 2x 6
1 + μsmall (x) ⎪⎩
0 if 0 x 2,
whose graphics are
Example 1.2.3 With μbig as defined in the third case of 2.2, and with N0 , it is

1, if 0x 7
μnot big (x) =
3 , 7 x 10,
10−x
if
with graphics
1.2.3 Antonyms and Negations
With pairs of antonyms (P, a P), it should always be taken into account that con-
ditional statements like “If the bottle is empty, then it is not full”, conduct to the
inequality μa P μnot P , with P = f ull, showing that not P could be taken as the
biggest antonym of P. It is not often the case in which a P = not P, practically it
only happens when aP is such that there is not any linguistic term a P in the language.
When a P and not P are not coincidental, it is said that a P is a strict antonym of P.
When modeling μa P = μ P ◦ A, and μnot P = N ◦ μ P , with a symmetry A of X
and a strong negation N in [0, 1], it results the condition of coherence:
μP ◦ A N ◦ μP ,
between A and N , that should be always verified. If A is known, N should be chosen

to satisfy this coherence’s condition, and if what is known is N , then A should be
chosen to verify such condition.
Example 1.2.4 With N0 (a) = 1 − a, and A(x) = 10 − x in X = [0, 10], if

⎧
⎨ 0, if x ∈ [0, 4]
μbig (x) = 4 ,
x−4
if ∈ [4, 8]
⎩
1, if x ∈ [8, 10],
results ⎧
⎨ 0, if x ∈ [6, 10]
μsmall (x) = μbig (10 − x) = 4 ,
6−x
if ∈ [2, 6]
⎩
1, if x ∈ [0, 2],
⎧
⎨ 1, if x ∈ [0, 4]
μnot big (x) = 1 − μbig (x) = 4 ,
8−x
if ∈ [4, 8]
⎩
0, if x ∈ [8, 10],
whose graphics show that the pair (big, small) is coherent, since μsmall μnot big .
Example 1.2.5 In [0, 10] take μbig (x) = 10

x
and μnot big (x) = 1−μbig (x) = 1− 10
x
.
Which symmetries A : [0, 10] → [0, 10] can be taken for having μsmall = μbig ◦ A?
From the coherence’s condition with N0 , follows
A(x) x
μsmall (x) = μbig (A(x)) = μnot big (x) = 1 −
10 10
hence, A(x) 10 − x is the condition A must satisfy. For example,
• If A1 (x) = 10 − x, it results μsmall (x) = 1 − 10x
= μnot big (x), a non-regular
case.
• If A2 (x) = 10 · 10−x
10+x , for which A2 (x) 10 − x, it results μsmall (x) = μbig (10 ·
10−x
10+x ) = 10−x
10+x , with graphics
also showing coherence.

Example 1.2.6 With the same μbig (x) = 10 x
in the previous example, and μsmall =
μbig (10−x), which Sugeno’s strong negation Nλ can be used for having μnot big (x) =
Nλ (μbig (x))?
It should be,
10 − x x 10 − x
μsmall (x) = μbig (10 − x) = Nλ (μbig (x)) = Nλ ( ) =
10 10 10 + λx
that means λx 0. Hence, λ ∈ (−1, 0]. For example,

• With Nλ = N0 , there is coherence
• With Nλ = N−0.5 , or Nλ (a) = 2(1−a)
2−a , there is coherence
• With Nλ = N1 , there is no coherence.
1.2.4 Medium Term
Given a regular pair of antonyms (P, a P), the middle or medium term of them, is
the predicate MP = not P and Not aP, as it was said in Sect. 1.2.1. Although we
don’t yet studied the diverse models for the conjunction ‘and’, by the moment let us
take the model given by the operation minimum (= min) in [0, 1]. Then,
μ M P = min(μnot P (x), μnot a P (x)).
For example, with P = small in [0, 10], N0 , and A(x) = 10 − x, is:
μ small
μ big
μ not small μ not big
6 8
The triplets (P, M P, a P) play an important role in the applications of fuzzy sets.
1.3 AND/OR
Let P, Q be two predicates in X . Consider the new predicates ‘P and Q’, and ‘P or
Q’, used by means of:
• ‘x is P and Q’ ⇔ ‘x is P’ and ‘x is Q’
• ‘x is P or Q’ ⇔ Not (Not ‘x is P’ and Not ‘x is Q’) ⇔ Not (‘x is P ’ and ’ x is
Q ’) ⇔ x is ‘(P and Q )’,
with respective primary uses P and Q ⊂ P ∩ Q , and P or Q . Take L-degrees
μP , μQ .
1.3.1 AND
Given (L , ), let ∗ : L × L → L be an operation, verifying the properties

• a b, c d ⇒ a ∗ c b ∗ d
• a ∗ c a, a ∗ c c,
Then, μ P (x) ∗ μ Q (x) is an L-degree for P and Q in X , since:
x P and Q y ⇒ x P y and x Q y ⇒ μ P (x) μ P (y) and μ Q (x)
μ Q (y) ⇒ μ P (x) ∗ μ Q (x) μ P (y) ∗ μ Q (y).
Hence, defining μ P (x) ∗ μ Q (x) = μ P and Q (x), an L-degree for ‘P and Q’ in
X is obtained, and the operation ∗ can be called an and-operation or a conjunction.
Notice that it is,
μ P and Q (x) μ P (x), and μ P and Q (x) μ Q (x).
In the case (L , ) is a lattice (see Sect. 1.6) with the minimum operation · = min,
and ∗ ·, it implies
μ P and Q (x) μ P (x) · μ Q (x).
Then, if (L , ·, +, ) is a lattice, then the and-operation is at least ∗ = ·(min), with

which we have the L-degree
μ P and Q (x) = μ P (x) · μ Q (x), ∀x ∈ X.
1.3.2 OR
If ∗ is an and-operation, and N : L → L is a strong negation, define
a ⊕ b = N (N (a) ∗ N (b)), for all a, b ∈ L .
Since a b, c d ⇒ N (b) N (a), N (d) N (c) ⇒ N (b) ∗ N (d)

N (a) ∗ N (c) ⇒ N (N (a) ∗ N (c)) N (N (b) ∗ N (d)), it results a ⊕ c b ⊕ d.
1.3 AND/OR 21
Analogously, from N (a) ∗ N (b) N (a), it follows a N (N (a) ∗ N (b)) = a ⊕ b,

and b a ⊕ b, for all a, b.
Then, μ P (x) ⊕ μ Q (x) is an L-degree for P or Q, since (remember that it is
P ⊂ −1
P ):
x P or Q y ⇔ x (P and Q ) y ⇔ y P and Q x ⇒ μ P and Q (y)
μ P and Q (x) ⇔ μ P (y) ∗ μ Q (y) μ P (x) ∗ μ Q (x) ⇒ N (μ P (y)) ∗ N (μ Q (y))
N (μ P (x)) ∗ N (μ Q (x)) ⇒ N (N (μ P (x)) ∗ N (μ Q (x)) N (N (μ P (y)) ∗ N (μ Q (y))
⇔ μ P (x) ⊕ μ Q (x) μ P (y) ⊕ μ Q (y).
Hence, μ P (x) ⊕ μ Q (x) can be taken as an L-degree for ‘P or Q’ in X, and the
operation ⊕ can be called an or-operation or a disjunction. Notice that
μ P (x) μ P or Q (x), μ Q (x) μ P or Q (x).
In the case (L , ·, +) is a lattice with the maximum operation + = max, and + ∗,

it implies μ P (x) + μ Q (x) μ P or Q (x).
Then, if (L , ·, +, ) is a lattice, at least it is the operation ⊕ = +(max), with
which we have the L-degree
μ P or Q (x) = μ P (x) + μ Q (x), ∀x ∈ X.
With such degree it holds the ‘duality’ law
μ P or Q = μP and Q .
Remark 1.3.1 The lattice operation ·(+) is not necessarily the only operation ∗(⊕),
that can exist. In the case (L , ) is not a lattice for ·(+), there can also exist other
operations ∗ and ⊕. For example, L = [0, 1] is not a lattice with ∗ = pr od, but
a b, c d ⇒ pr od(a, c) pr od(b, d),
and pr od(a, b) a, pr od(a, b) b. Hence, pr od can be eventually used to

model the use of and. Analogously, ⊕ = pr od ∗ (a, b) = 1 − pr od(1 − a, 1 −
b) = a + b − a · b verifies a b, c d ⇒ pr od ∗ (a, c) pr od ∗ (b, d), and
a pr od ∗ (a, b), b pr od ∗ (a, b). Hence, pr od ∗ can be eventually used to model
the use of or.
Remark 1.3.2 The existence of operations ∗ and ⊕ in L, warrants the existence of

L-degrees for and, or, respectively.
Remark 1.3.3 Since a ∗ b (a ∗ b) ⊕ (a ∗ b), and a ∗ b a, a ∗ b b it follows
(a ∗ b) ⊕ (a ∗ b) a ⊕ b,
and a ∗ b a ⊕ b, for all a, b in L, and all pair of operations ∗ and ⊕.

Remark 1.3.4 It should be noticed that what has been presented is sufficient but not
necessary for the representation of μ P and Q and μ P or Q b means of μ P and μ Q .
Given P, if with M for medium, MP = Not P and Not aP = P and (aP) ,

with a negation function N for not, and a symmetry A for the opposite, and ∗ for
and, results
μ M P (x) = μ P and (a P) (x) = μ P (x) ∗ μ(a P) (x) = N (μ P (x)) ∗ N (μ P (A(x))),
for all x in X .
1.4 Qualified, Modified, and Constrained Predicates
1.4.1 Qualified Predicates
Let P be a predicate on X , with L-degree μ P : X → L, and τ a predicate on

μ P (X ) ⊂ L, with L-degree μτ : μ P (X ) → L. Suppose ⊂ τ , and consider the
qualified predicate ‘P is τ ’,
‘x is (P is τ )’ := x is P is τ ,
provided ∅ = P is τ ⊂ P . On these conditions,
μ P is τ = μτ ◦ μ P
is an L-degree for ‘P is τ ’ in X , since:
x P is τ y ⇒ x P y ⇒ μ P (x) μ P (y) ⇒ μ P (x) τ μ P (y)

⇒ μτ (μ P (x)) μτ (μ P (y)),
that is, (μτ ◦ μ P )(x) (μτ ◦ μ P )(y).

For example, with L = [0, 1], P = small in [0, 10], with P = −1 , and
μ P (x) = 1 − 10x
, if τ = large in [0, 1] is with τ = , and

0, if 0 x 0.5
μτ (x) =
1, if 0.5 x 1,
it results
0, if 5 x < 10
μτ ◦ μ P (x) =
1, if 0 x < 5,
that allow the interpretation of small is large as less than five.

1.4 Qualified, Modified, and Constrained Predicates 23
Take τ = tr ue on [0, 1], with τ = , and the degree μτ as a non-decreasing

function [0, 1] → [0, 1], such that μτ (0) = 0, μτ (1) = 1, once accepting that ‘0 is
τ ’ is false, and ‘1 is τ ’ is true. Taking μτ (x) = x, as it is usual in fuzzy logic, it is
Degree up to which x is P is true = μτ (μ P (x)) = μ P (x), for all x in X,
that allows to accept, as it is usual,
Degree of true of x is P = μ P (x).
1.4.2 Linguistic Modifiers
Linguistic modifiers or linguistic hedges, m, are adverbs acting on P just in the con-
catenated form m P. For example, with m = very and P = tall, it is m P = very tall.
A characteristic that linguistically distinguishes imprecise predicates from pre-
cise ones, is that in the first case and once P and m are given, mP is immediately
understood. If P is precise (for example, P = even in the set of natural numbers),
mP needs of a new definition to be understood (what it means very even?). Modifiers
only modify, but do not change abruptly imprecise predicates.
If P in X is with the use ( P , μ P ), and m in μ P (x) ⊂ L is with ⊂ m and
μm , provided m P ⊂ P , it can be taken the degree
μm P = μm ◦ μ P ,
since, x m P y ⇒ x P y ⇒ μ P (x) μ P (y) ⇒ μ P (x) m μ P (y) ⇒

μm (μ P (x)) μm (μ P (y)), or (μm ◦ μ P )(x) (μm ◦ μ P )(y).
Among linguistic modifiers there are two specially interesting types:
• Expansive modifiers, verifying idμ P (x) μm ,
• Contractive modifiers, verifying μm idμ P (x) .
With the expansive, it results
idμ P (x) (μ P (x)) = μ P (x) μm (μ P (x)) = μm P (x) : μ P (x) μm P (x),

for all x in X.
With the contractive, it results
μm P (x) = μm (μ P (x)) idμ P (x) (μ P (x)) = μ P (x) : μm P (x) μ P (x),

for all x in X.
This is what happens in L = [0, 1] with the Zadeh’s old definitions,

√
μmor e or less (a) = a, μver y (a) = a 2 .
1.4.3 Constrained Predicates
Let P and Q predicates in X and Y , respectively, with uses ( P , μ P ), ( Q , μ Q ).

Each relation ∅ = R(P, Q):
(x is P, y is Q) ∈ R(P, Q),
allows to define the constrained predicate Q|P = ‘Q if P’, in X × Y , given by
(x, y) ∈ Q|P ⇔ (x is P, y is Q) ∈ R(P, Q).
An example is given by the interpretation
(x, y) ∈ Q|P ⇔ ‘If x is P, then y is Q ⇔ ‘x is P implies y is Q .
Provided Q|P induces a preorder Q|P in X × Y , and that there is an L-degree

μ Q|P : X × Y → L ,
(x1 , y1 ) Q|P (x2 , y2 ) ⇒ μ Q|P (x1 , y1 ) μ Q|P (x2 , y2 ),
it could be studied how to express μ Q|P by means of μ P and μ Q .

Notice that there are several possibilities for obtaining Q|P from both P and
Q , i.e.,
Q|P = P × Q , Q|P = −1 P × Q , etc.
The degree μ Q|P is said to be decomposable, or functionally expressible, if there

is an operation J : L × L → L , such that
μ Q|P (x, y) = J (μ P (x), μ Q (y)),
for all (x, y) ∈ X × Y , and it again remains to be tested that μ Q|P is actually a
L-degree for Q|P. For example,
• If Q|P = Q × P , and J is non-decreasing in both variables, it is
(x1 , y1 ) Q|P (x2 , y2 ) ⇔ x1 P x2 , y1 Q y2 ⇒ μ P (x1 ) μ P (x2 ), and
μ Q (y1 ) μ Q (y2 ) ⇒
J (μ P (x1 ), μ Q (y1 )) J (μ P (x2 ), μ Q (y2 )),
or
μ Q|P (x1 , y1 ) μ Q|P (x2 , y2 ).
• If Q|P = −1 −1
P × Q , and J is decreasing in both variables, it also follows the
same conclusion,
• If Q|P = −1
P × Q , and J is decreasing in its first variable, and non-decreasing
in the second, it also follows the same conclusion.
Etc.
Remark 1.4.1 The decomposability or functional expressibility of μ Q|P , μnot P ,

μa P , μ P and Q , and μ P or Q , is not a general property of the L-degrees of the predi-
cates Q|P, not P, a P, P and Q, and P or Q. What has been shown at such respect
with functions J, N , A, ∗, and ⊕, respectively, is just to be taken as examples of
the existence of L-degrees. Although in the applications of fuzzy logic is currently
accepted that all these predicates are functionally expressible, that is, expressed
through numerical functions
J, ∗, ⊕ : [0, 1] × [0, 1] → [0, 1], N : [0, 1] → [0, 1], A : X → X,
it should not be considered that this is always the case.

Remark 1.4.2 Q|P is an example of a relational predicate, that is, a predicate R on
X × Y such that (x, y) ∈ R, with x ∈ X , and y ∈ Y . For example, R = larger,
implies, around, etc. Of course, once either x or y are fixed, what results is a predicate
(unary) in Y or X , respectively, as it is with ‘x is around y’, if X = Y = [0, 10],
where with y = 5 it results the unary predicate around five in [0, 10].
Relational, or binary, predicates can be either precise or imprecise. In the first
case, they originate a crisp subset of X × Y defined by

ω, if (x, y) ∈ R
μ R (x, y) =
α, otherwise.
In the second, they originate an L-set in X × Y defined by
μ R (x, y) = Degree in L up to which it is (x, y) ∈ R,
once an L-degree for R is known.

Remark 1.4.3 In the case L = [0, 1], functions J : [0, 1] × [0, 1] → [0, 1] allowing
to represent μ Q|P by J ◦ (μ P × μ Q ), are called fuzzy relations, and if the predicate
Q|P interprets a rule, these relations are called fuzzy conditionals.
1.4.4 Group Meaning
The meaning of words is not fixed for all people and all context. For example,
in a dinner with three commensals the deliciousness of the dessert plates could
easily result in three different orderings of such plates. Since language is a social
phenomenon, also meaning is such, and it is possible to talk on the meaning of
predicates for a group of people in, of course, a given context.
For a group of people G = { p1 , . . . , pm }, a predicate P on X can show m primary
meanings P,i , 1 i m. Since

m
( P,i ) = P,G
i=1
is not empty (all P,i are reflexive), it can be taken:
Primar y meaning o f P on X f or the gr oup G = P,G .
Notice that provided all P,i are preorders, P,G is also a preorder.
If m L − degr ees μ(i)P are known for each primary meaning P,i , since
• x = P,G y ⇔ x = P,1 y & . . . & x = P,m y,
• x P,G y ⇔ x P,1 y & . . . & x P,m y,
for each function : L m → L, non-decreasing in each place i between 1 and m
(for example, if a b then (a, x2 , . . . , xm ) (b, x2 , . . . , xm )), or aggregation
function, it results
• x P,G y ⇒ (μ(1) (m) (1) (m)
P (x), . . . , μ P (x)) (μ P (y), . . . , μ P (y)),
that allows to take
(1) (m)
μGP (X ) = (μ P (x), . . . , μ P (x)), for all x ∈ X,
as an aggregate L − degr ee of P on X for the group G. The meaning for G results

from aggregating its people’s meanings.
1.4.5 Synonims
In the language, synonymy is a complex problem whose roots are possibly to be

searched for in the apparition of new facts or concepts for which there is not yet a
word for their designation. Then, what is sometimes done is to designate the new
fact/concept by means of an old word whose meaning is considered, for some rea-
sons, similar to that of the new fact/concept. That is, for example, that in which
the old word was already used in situations judged similar to those where the new
fact/concept appears/applies.
Synonymy is related with some kind of similarity or proximity of meaning but
here we will only try to present some previous treats of it.
Let P be a predicate on X with P , and Q a predicate on Y with Q . If there
exists a bijective function u : X → Y such that,
• x1 P x2 ⇔ u(x1 ) Q u(x2 ),
predicates P and Q are u-primary-synonyms. Notice that when X = Y , with u = id X ,
what results is that P and Q are id X -primary synonyms, or primary synonyms for
short, if and only if P = Q , that is, if and only if
Primary meaning of P on X = Primary meaning of Q on X
If P and Q are id X -primary synonyms, it is said that they are exact or perfect
synonyms when μ P = μ Q , and it results ( P , μ P ) = ( Q , μ Q ).
For example, if P = small on X = [0, 1] is with P = −1 (the reverse linear

order on the real line), and Q = short on Y = [0, 10] is with Q = −1 (also the
reverse linear order on the real line), u(x) = 10x gives
• x −1 y ⇔ 10x −1 10y
taking P ∩ −1 −1
P and Q ∩ Q equal to the identity (=) on the real line. Then,
small and short can be considered a pair of u-primary synonyms.
If P acts on X with an L − degr ee μ P , Q acts on Y and is a u-primary synonym
of P, from
y1 Q y2 ⇔ u −1 (y1 ) P u −1 (y2 ) ⇒ μ P (u −1 (y1 )) μ P (u −1 (y2 )),
it follows that
μ Q = μ P ◦ u −1
is an L − degr ee for Q. In this situation it is
y1 μ Q y2 ⇔ u −1 (y1 ) μ P u −1 (y2 ),
or
x1 μ P x2 ⇔ u(x1 ) μ Q u(x2 ),
that are equivalent to

μ Q = μ P ◦(u × u).
For example, with the before mentioned predicates short and small, it is
μshor t (y) = μsmall (y/10)
for all y in [0, 10], and results
y1 μ Q y2 ⇔ y1 /10 μ P y2 /10.
Remark 1.4.4 Whenever P and Q are u-synonyms, it could be stated that

“P means Q”.
Remark 1.4.5 The definition of primary meaning is just a formal one trying to
approach an important aspect of the meaning of linguistic predicates, when act-
ing on a given universe of discourse. The same can be said about the definition of
u-primary synonyms with which it does not hold, in general, that a pair of linguistic
synonyms are necessarily u−primary synonyms. Anyway, what can be said is that
Q is a migration of P to the universe Y .
Remark 1.4.6 In some way, the current meaning of a predicate, the form in which it
is used today in the plane language, partially inherits its past meanings.
It is the evolution of the use of P in different universes that produces the

descriptions of P that are found in dictionaries. Such current meaning, or mean-
ings, always never can be described by a ‘if and only if’ definition like in the case of
P = odd in the set of positive integers, but just by some more or less clear descrip-
tion, as it is with the same word ‘odd’ when applied either to social situations, or to
people. Most of the words appearing in dictionaries are imprecise and, hence, not
definable by precise terms. For this kind of reasons, dictionaries contain several short
descriptions for most of the words in them.
The current meaning of P is also influenced with the simultaneous use of its
synonyms and antonyms, as well as by the more or less different uses of P some
groups of people did, for saying nothing of the influences coming from other natural
languages. Of course, there are some words that evanish since their universe of
application disappears; words are ‘born’ when applied to concrete universes, and
are ‘dying’ when these universes disappear, as it happened, for instance, with some
old words that were use in the wide middle age’s world of cavalry. Language is a
dynamical system, and their dynamicity implies that it cannot be well studied by
only employing the methods that are typical of classical logic. Fuzzy logic could be
a mathematical and practical tool for that study in the cases where ambiguity can be
left aside.
1.5 Linguistic Variables
Linguistic variables are basic tools in most application of fuzzy sets in the technol-
ogy’s field. They do mainly appear when linguistically describing the behavior of the
physical variables of a system. A linguistic variable explicits a concept by (linguis-
tically) granulating some elemental components of it, by showing the perceptually
distinguishable shades that are relevant for the corresponding application.
A linguistic variable LV is formed after considering
1. Its principal predicate, P
2. One of the opposites of P, a P
3. Some linguistic modifiers m 1 , . . . , m n ,
and by adding:
4. Its negate (not P), or the middle-predicate (MP), or P and Q, or not m 1 P, or P
and m 2 a P, . . .
Then LV, is called the linguistic variable generated by P, and reflects the linguistic
granulation perceived for the concept. For example,
• LV = Age, is Age = { young, old, middle-aged, not old, not very young, …}
• LV = Truth, is Truth = { true, false, very false, not very true, …}
• LV = Temperature, is Temperature = { cold, hot, warm, not cold, not very hot, …}
• LV = Size, is Size = {large, small, medium, very large, …}
• LV = Height, is
1.5 Linguistic Variables 29
– For buildings, Height = {low, high, medium, very high, not very low, …}
– For people, Height = {tall, short, medium, very tall, more or less short, …}
• LV = Speed, is Speed = {fast, slow, very slow, more or less fast, not fast, …}
Although the number of terms in a Linguistic Variable can be large, usually it is
comprised between 5 and 9 (7 ± 2) since, in a lot of cases, less than 5 shades is poor
and more than 9 is excessive. Anyway in a good number of applications there only
appear the three terms P, aP, and MP.
Usually, in the applications, the variables range in the set of real numbers and,
because of this, the predicate P acts in some interval of the real line. For example,
the linguistic variable ‘Temperature’ in the interval between
−10, and 50 degrees Celsius, is often represented by only the three terms μcold , μhot ,
μwar m , with warm = not cold and not hot.
Analogously, ‘Height’ for people, can be represented in [0, 2] meters by the lin-
guistic variable with the four terms μtall , μshor t , μver y shor t , μmor e or less tall in the
following figure,
1.5.1 Fuzzy Partition
It is sometimes useful in the applications that, once ordered in some sequence, the
fuzzy sets in a linguistic variable L V = {μ0 , μ1 , . . . , μn } do form what is called a
fuzzy partition (or a unit’s partition), that is, verifying

n
μ j (x) = 1, ∀x ∈ X.
j=0
This definition is a direct generalization of what happens in the classical case. If

X = A0 ∪ A1 ∪ · · · ∪ An , with Ai ∩ A j = ∅ for i = j, is a classical partition of X ,
it is μ A0 (x) + μ A1 (x) + · · · + μ An (x) = 1, since for each x ∈ X there is just one A j
such that x ∈ A j , but x ∈ / Ai , for i = j, that is,
μ A j (x) = 1, and μ Ai (x) = 0 if i = j,

n
that implies j=0 μ A j (x) = 1. Let us show three examples.
Example 1.5.1 In X = [0, 4], take

1 − x, if 0 x 1,
μ0 (x) =
0 if 1 x 4,
⎧
⎪ 0, if 0 x j − 1,
⎪
⎪
⎨x +1− j if j − 1 x j,
μ j (x) = for 1 j 3
⎪
⎪ j +1−x if j x j + 1,
⎪
⎩
1, if j + 1 x 4,

0, if 0 x 3,
μ4 (x) =
x −3 if 3 x 4,
Graphically,
Obviously,
4
• If 0 x 1, j=0 μ j (x) = μ0 (x) + μ1 (x) = 1 − x + x = 1

• If 1 j 3, j x j + 1, 4j=0 μ j (x) = μ j (x) + μ j+1 (x) = ( j + 1 − x) +
(x + 1 − j − 1) = 1.
• If 3 x 4, 4j=0 μ j (x) = μ3 (x) + μ4 (x) = 4 − x + x − 3 = 1
Hence {μ0 , μ1 , . . . , μn } is a fuzzy partition of [0, 4]. Notice that each μ j can be
labeled by the predicate around j= A j , that is μ j = μ A j .
1.5 Linguistic Variables 31
Example 1.5.2 In X = [0, 10], take

⎧x

⎨ 4, if 0 x 4,
1 − x4 , if 0 x 4,
μ0 (x) = , μ1 (x) = 1, if 4x 6 ,
0, if 4 x 10 ⎩ 10−x
4 , if 6 x 10

0, if 0 x 6,
μ2 (x) =
4 , 6 x 10.
x−6
if
Graphically
Since, obviously, μ0 (x) + μ1 (x) + μ2 (x) = 1 for all x ∈ [0, 10], {μ0 , μ1 , μ2 } is
a fuzzy partition of the interval [0, 10].
Example 1.5.3 By proceeding in the same way that in last example, it is easy to
prove that it is
{μcold , μhot , μwar m },
a fuzzy partition of the interval [−10, 50].
Notice that, with the strong negation N0 = 1−id, if {μ0 , μ1 , μ2 } is a fuzzy partition
of X , it is
μ1 (x) + μ2 (x) = 1 − μ0 (x), or μ1 + μ2 = μ0
μ0 (x) + μ1 (x) = 1 − μ2 (x), or μ0 + μ1 = μ2
μ0 (x) + μ2 (x) = 1 − μ1 (x), or μ0 + μ2 = μ1
The complement (with N0 ) of each element in a fuzzy partition, is just the addition
of the other elements in the partition.
1.6 A Note on Lattices
1. An algebraic structure (L , ) is a partially ordered set, or poset, if the relation

⊆ L × L is a partial order, that is, enjoys the properties:
• Reflexive, a a for all a in L.
• Transitive, (a b) & (b c) ⇒ (a c).
• Anti-symmetric, (a b) & (b a) ⇔ a = b.
Provided (L , ) only verifies the reflexive and the transitive properties, is called
a preordered set.
2. An algebraic structure (L , ·, +) is a lattice, provided the binary operations · and
+ verify,
• Both operations are commutative and associative, that is, a · b = b · a, a + b =
b + a, a · (b · c) = (a · b) · c, a + (b + c) = (a + b) + c, for all a, b, c in L.
• Both operations are idempotent, that is, a · a = a, a + a = a, for all a in L.
Without proof, let’s state what follows:
(a) It is easy to prove that the relation a b ⇔ a · b = a, is a partial order called

the ‘natural order of the lattice’, and that it is equivalent to a b ⇔ a+b = b.
(b) Operation + verifies: a+b is the lowest element in (L , ) such that a a+b,
and b a + b. That is, in the natural order, there is no element in L that is
both between a and a + b, and between b and a + b. Operation · verifies: a · b
is the greatest element in (L , ) such that a · b a, and a · b b. That is, in
the natural order, there is no element in L between a · b and a, and between
a · b and b, in the order .
(c) If a b, then a · c b · c, and a + c b + c, for all c in L.
3. A lattice is bounded if there are elements 0 and 1 in L, such that 0 a 1, for

all a in L and called, respectively, the minimum and the maximum of the lattice.
These elements verify:
0 · a = 0, 1 · a = a, 0 + a = a, and 1 + a = 1, for all a in L. That is, 0 is
absorbent for ·, and neutral for +, and 1 is neutral for ·, and absorbent for +.
4. In a bounded lattice, a mapping : L → L is a negation provided: 0 = 1, 1 = 0,
a b ⇒ b a , and a = (a ) = a, for all a in L.
5. A bounded lattice with a negation is dual, provided the operations · and + are
linked by (a · b) = a + b , for all a, b in L. This expression is equivalent to
(a + b) = a · b , since it immediately follows from (a · b ) = a + b = a + b.
6. A lattice is distributive, provided: a + b · c = (a + b) · (a + c), and a · (b + c) =
a · b + a · c.
Provided it is dual, a bounded distributive lattice with a negation is a De Morgan
algebra. A bounded lattice with a negation such that it is dual, and: a · a = 0
(law of non-contradiction), and a + a = 1 (law of excluded-middle), is an Ortho-
lattice. Notice that both laws are equivalent; for instance, a · a = 0 ⇒ (a · a ) =
1 ⇒ a + a = 1.
An Ortho-lattice such that it holds the property
a b ⇔ b = a + a · b,
is called an Ortho-modular lattice. Distributive ortho-lattices, are called Boolean
algebras. Notice that in them:
• a b ⇒ a + a · b = (a + a ) · (a + b) = a + b = b
• b = a + a · b ≥ a ⇒ a b,
that is, all Boolean algebras are Ortho-modular lattices, but not reciprocally.
1.6 A Note on Lattices 33
It is obvious that a De Morgan algebra (L , ·, +, ) verifying the laws of

non-contradiction and excluded-middle is a Boolean algebra. In all De Morgan
algebra there is included the Boolean algebra whose elements are the Boolean
elements of the algebra, that is, that given by the set {a ∈ L; a · a = 0} =
{a ∈ L; a + a = 1}, a set that is never empty since at least contains 0 and 1.
1.6.1 Examples
1. A good example of a Boolean algebra is the power-set P(X ) = {A; A ⊆ X },

whose elements are the subsets of a set X . In this case, the operations are:

• the intersection
of subsets is ·
• the union of subsets is +
• the complement of subsets (c ), is ,
and 0 is the empty set ∅, and 1 is the full set X .
Power-sets are typical instances Boolean algebras. In fact, any Boolean algebra
is isomorphic to a power-set.
2. The unit interval [0, 1], endowed with the operations: · = min, + = max, and
= 1 − id, is a De Morgan algebra, but not a Boolean one, since, for instance,
min(a, 1 − a) = 0 ⇔ a = 0, or a = 1. The Boolean elements of this algebra

are just 0 and 1.
3. The set of functions [0, 1] X = {μ; μ : X → [0, 1]}, endowed with the operations
given by (μ·σ)(x) = min(μ(x), σ(x)), (μ+σ)(x) = max(μ(x), σ(x)), μ (x) =
1 − μ(x), for all x in X , is a De Morgan algebra, whose Boolean elements are the
functions μ ∈ {0, 1} X . Hence, this algebra is not a Boolean one.
4. The set whose elements are all the vector subspaces of R 3 constitute an Ortho-
modular lattice, once endowed with the operations:
• The intersection of two subspaces (·),
• The minimum subspace that contains two subspaces (+),
• The subspace that is orthogonal to a subspace ( ).
For instance, if it is < u, v, w > = R 3 , that is, these three vectors are an orthog-
onal basis of R 3 , it is:
• < u > · < v > = 0, the null subspace, with 0 = (0, 0, 0),
• < u > + < w > = < u, w >, the plane given by the two orthogonal vectors
u an w,
• < v > = < u, w >, the same plane.
Chapter 2
Algebras of Fuzzy Sets
2.1 Introduction
From now on it will be only considered the case in which (L , ) = ([0, 1], ), that
is, of Zadeh’s fuzzy sets, with predicates P in X known through a degree μ P : X →
[0, 1], and without knowing, necessarily, its primary use P . The set of all fuzzy sets
in X , [0, 1] X , will be also denoted by F(X ). In this case, the preorder μ P is linear,
or total, since for all x, y in X it is either μ P (x) μ P (y), or μ P (y) μ P (x), that
is, it is either x μ P y or y μ P x for all x, y in X . Hence, μ P rarely will perfectly
reflect the primary use of P in X , since P is usually not linear.
In the case in which X is finite, X = {x1 , . . . , xn }, the fuzzy sets μ ∈ [0, 1] X ,
will be represented by
μ = μ(x1 )/x1 + μ(x2 )/x2 + · · · + μ(xn )/xn ,
with the convention that if some term μ(x j )/x j does not appear, is that it is μ(x j ) = 0.
For example, with X = {1, 2, 3, 4}, the expression
μ = 0.5/x1 + 0.7/x2 + 1/x4 ,
refers to the fuzzy set in X given by μ(x1 ) = 0.5, μ(x2 ) = 0.7, μ(x3 ) = 0, μ(x4 ) = 1.
Analogously, the fuzzy set μ = N0 ◦ μ (N0 = 1 − id) is
μ = 0.5/x1 + 0.3/x2 + 1/x3 ,
2.1.1 Cartesian Product
If A, B are crisp subsets in X and Y , respectively, that is, A ∈ P(X ) and B ∈ P(X ),
its cartesian product A × B = {(a, b); a ∈ A, b ∈ B} ⊂ X × Y , is with the
36 2 Algebras of Fuzzy Sets
membership function μ A×B : X × Y → {0, 1}, given by
μ A×B (x, y) = min(μ A (x), μ B (y)
for all x ∈ X , y ∈ Y . It is μ A×B (x, y) = 1 ⇔ μ A (x) = μ B (y) = 1.

In the same vein, if μ ∈ F(X ), and σ ∈ F(Y ), the cartesian product μ × σ is
defined by directly generalizing the classical case:
μ × σ = min ◦ (μ, σ).
For example, with X = {x1 , x2 , x3 }, Y = {y1 , y2 }, and
μ = 1/x1 + 0.8/x2 , σ = 0.9/y1 + 0.7/y2 ,
it is
μ × σ = 0.9/(x1 , y1 ) + 0.7/(x1 , y2 ) + 0.8/(x2 , y1 ) + 0.7/(x2 , y2 ),
with (μ × σ)(x3 , y1 ) = (μ × σ)(x3 , y2 ) = 0.

y
With μbig (x) = x5 if x ∈ [0, 5] and μsmall (y) = 1 − 7 if y ∈ [0, 7], it is
x y
(μbig × μsmall )(x, y) = min( , 1 − ),
5 7
the representation of the cartesian product as a surface contained in the cube [0, 5] ×
[0, 7] × [0, 1].
Of course, if μ = μ A ∈ {0, 1} X , σ = μ B ∈ {0, 1} X , it is not only μ × σ ∈ {0, 1} X
but μ × σ = μ A×B .
2.1.2 Extension Principle
If f : X → Y is a mapping and A is a crisp subset of X , A ⊂ X , it is f (A) = {y ∈

Y ; f (a) = y, a ∈ A} the f-image of A in Y. Notice that

1, if it exists x ∈ A such that f (x) = y,
μ f (A) (y) = sup{μ A (x); f (x) = y} =
0, otherwise.
With these f-image, the mapping f : X → Y is extended to the respective power

sets by

f : P(X ) → P(Y ), A → f (A).
2.1 Introduction 37
In the same vein, given a mapping f : X → Y , it can be extended to the fuzzy

power sets F(X ), F(Y ) by

f : F(X ) → F(Y )

f (μ)(y) = sup{μ(x); f (x) = y}, for all y ∈ Y,
and f is known as the ‘extension’ of f to the fuzzy parts, and the definition as the
Zadeh’s Extension Principle.
For example, if f : [0, 10] → [0, 1], is given by f (x) = 1 − 10 x
, the
fuzzy set μ(x) = 10 in [0, 1][0,10] extends to the fuzzy set in [0, 1],
x
f (μ)(y) =
sup{μ(x); f (x) = y} = sup{ 10x
; 1 − 10x
= y} = 1 − y, for all y ∈ [0, 1].
If X = {1, 2, 3, 4}, Y = {a, b, c}, the mapping f : X → Y such that
f (1) = f (2) = a, f (3) = f (4) = b,
extends the fuzzy set μ = 1/1 + 0.4/2 + 1/3 + 0.7/4 in F(X ), to the fuzzy set
f (μ)
in F(Y ) with values,

f (μ)(a) = max{μ(x); x ∈ f −1 (a)} = max{μ(1), μ(2)} = max(1, 0.4) = 1

f (μ)(b) = max{μ(3), μ(4)} = 1

f (μ)(c) = 0, since f −1 (c) = ∅.
Hence,

f (μ) = 1/a + 1/b,
that corresponds to the crisp subset {a, b} of Y .

Notice that if μ = μ A ∈ {0, 1} X , it is
f (μ A ) = μ f (A) , that is not only a crisp sub-
set of Y , but coincides with the classical extension f (A) of A. Nevertheless, as it is
shown by the above example, it can happen that f (μ) ∈ P(Y ) with μ ∈ F(X )−P(X ).
2.1.3 Preservation of the Classical Case
Like with the cartesian product and with the extension principle, all operations with
fuzzy sets must reproduce, when the data are crisp, the corresponding result obtained
in the crisp theory. This is the principle of preservation of the classical case, that is
forced by the will, and the necessity, of including all ‘the classical’ as a particular
case of the algebras of fuzzy sets.
To illustrate this preservation’s principle, let us show a negative example. With
X = [0, 1], and all μ ∈ [0, 1][0,1] , the function
μ∗ (x) = 1 − μ(1 − x),

verifies:
• μ∗0 (x) = 1 − μ0 (1 − x) = 1: μ∗0 = μ1
• μ∗1 (x) = 1 − μ1 (1 − x) = 1 − 1 = 0: μ∗1 = μ0
• μ σ ⇒ 1 − σ(1 − x) 1 − μ(1 − x) ⇒ σ ∗ μ∗
• μ∗∗ (x) = 1 − μ∗ (1 − x) = 1 − [1 − μ(x)] = μ(x): μ∗∗ = μ.
Hence, it could seem that the function μ → μ∗ can be taken as a “strong negation”
for the fuzzy sets in [0, 1], but it is not the case. Notice that if μ ∈ {0, 1}[0,1] , then
it should be also μ∗ ∈ {0, 1}[0,1] , that is, if μ represents a classical subset of [0, 1],
also μ∗ should represent not only a classical subset but precisely the complement of
μ. But with A = [0, 21 ] ⊂ X ,

1, 0 x 0.5,
μ A (x) =
0, 0.5 < x 1,
follows,

1, 0.5 < x 1, 0, 0.5 < x 1,
μ∗A (x) = 1 − μ A (1 − x) = 1 − =
0, 0 x 0.5, 1, 0 x 0.5
that represents the subset [0, 0.5], but not Ac = (0.5, 1]. The unary operation ∗
violates the preservation principle, and hence it cannot be taken into account to
negate fuzzy sets.
2.1.4 Resolution
Let us denote by μr the constant fuzzy sets in [0, 1] X , μr (x) = r , for r ∈ [0, 1]
and all x ∈ X . Notice that in {0, 1} X there are only the “constants” μ0 and μ1 , that
correspond to the sets ∅ and X , respectively.
Given μ ∈ [0, 1] X , let us denote by μ(r ) the fuzzy (crisp) set

(r ) 1, if r μ(x),
μ (x) =
0, otherwise,
for all r ∈ [0, 1], and by [μ(r ) ] the corresponding classical subset {x ∈ X ; r μ(x)}.
These sets are called the r-cuts of μ and it is always [μ(0) ] = X .
For example, in the following figures are shown, respectively, the constant fuzzy
set μ0.5 , and the 0.3-cut of two different fuzzy sets.
2.1 Introduction 39
Notice that when μ A ∈ {0, 1} X , it is

• [μ(0)
A ] = X , since for all x ∈ X it is 0 μ(x)
(r )
• If r > 0, [μ A ] = A, since for all x ∈ A it is 0 < r 1 = μ A (x),
then, the only r-cuts of a crisp subset A of X are X and A.
If r s, since s μ(x) implies r μ(x), it results [μ(s) ] ⊂ [μ(r ) ]: r-cuts are
decreasing when their indices increase.
Let us show an example with X = {1, 2, 3, 4, 5} and μ = 0.8/1+0.6/2+0.7/3+
1/4 + 1/5, where the only significative values for the r -cuts are 0.6, 0.7, 0.8, and 1:
• [μ(0.6) ] = {1, 2, 3, 4, 5} = X
• [μ(0.7) ] = {1, 3, 4, 5}
• [μ(0.8) ] = {1, 4, 5}
• [μ(1) ] = {4, 5}.
It is clear that 0.6 0.7 0.8 1, and [μ(1) ] ⊂ [μ(0.8) ] ⊂ [μ(0.7) ] ⊂ [μ(0.6) ].
Theorem 2.1.1 (Theorem of resolution) For all μ ∈ [0, 1] X , is μ(x) = max{r ∈
[0, 1]; min(μr (x), μ(r ) (x)}, for all x ∈ X .
Proof
max min(μr (x), μ(r ) (x)) = max min(r, μ(r ) (x))

0r 1 0r 1

1, if r μ(x), r, if r μ(x),
= max min r, = max = μ(x).
0r 1 0, otherwise, 0r 1 0, otherwise,

Example 2.1.2 With X and μ in the last example, it is
μ(x) = max(min(0.6, μ(0.6) (x)), min(0.7, μ(0.7) (x)),

min(0.8, μ(0.8) (x)), min(1, μ(1) (x))),
and, for instance,

μ(1) = max(min(0.6, 1), min(0.7, 1), min(0.8, 1), min(1, 0)) = 0.8
μ(4) = max(min(0.6, 1), min(0.7, 1), min(0.8, 1), min(1, 1)) = 1
etc.
Example 2.1.3 With X = {1, 2, 3}, take μ : X × X → [0, 1], given by μ(i, j) =
min(i, j)
3 . This fuzzy set in X × X can be represented either by the matrix
⎛1 1 1 ⎞
3 3 3
⎜ ⎟
⎝ 1
3
2
3
2
3 ⎠,
1 2
3 3 1
or by the graph
1/3 2/3
1/3
1 2
1/3
2/3
1/3 2/3
1/3
2 1
Since the matrices of μ(1) , μ( 3 ) , and μ( 3 ) are respectively
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
000 000 111
⎝0 0 0⎠,⎝0 1 1⎠,⎝1 1 1⎠,
001 011 111
it results⎛ ⎛ ⎛ ⎞⎞ ⎛ ⎛ ⎞⎞ ⎛ ⎛ ⎞⎞⎞
000 000 111
⎜ ⎜ ⎜ ⎟⎟ ⎜ ⎜ ⎟⎟ ⎜ ⎜ ⎟⎟⎟
max ⎝min ⎝1, ⎝ 0 0 0 ⎠⎠ , min ⎝ 23 , ⎝ 0 1 1 ⎠⎠ , min ⎝ 13 , ⎝ 1 1 1 ⎠⎠⎠
001 011 111
⎛⎛ ⎞ ⎛ ⎞ ⎛ 1 1 1 ⎞⎞ ⎛ 1 1 1 ⎞
000 0 0 0 3 3 3 3 3 3
⎜⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎟ ⎜ ⎟
= max ⎝⎝ 0 0 0 ⎠ , ⎝ 0 23 23 ⎠ , ⎝ 13 13 13 ⎠⎠ = ⎝ 31 23 23 ⎠ ,
001 0 23 23 1 1 1
3 3 3
1 2
3 3 1
accordingly with the theorem of resolution.
2.2 The Concept of an ‘Algebra of Fuzzy Sets’ 41
2.2 The Concept of an ‘Algebra of Fuzzy Sets’
2.2.1 Introduction
Functions
μ ∈ F(X ) = [0, 1] X ,
will be labeled only when it is some predicate P in X such that μ P = μ, and it is

obvious that it could be the fact of having μ P = μ Q = μ R · · · = μ, in which case the
predicates P, Q, R, . . . are exact synonyms in X . Notwithstanding there are much
more functions in [0, 1] X than predicates in X , and given a not previously labeled
μ ∈ [0, 1] X , it can be ‘artificially’ introduced the predicate M (= μ) such that,
Degree up to which x is M = μ(x), for all x in X.
• Notice that F(X ) will be taken as ‘ordered’ (partially) by means of the binary
pointwise relation
μ σ ⇔ μ(x) σ(x), for all x ∈ X,
that induces the pointwise identity
μ = σ ⇔ μ σ and σ μ ⇔ μ(x) = σ(x), for all x ∈ X.
The pointwise relation is also called the ‘inclusion’, and μ σ denoted by

‘μ is included in σ’. It enjoys the laws reflexive, antisymmetric and transitive.
It will be always considered that F(X ) denotes, at least, the structure
([0, 1] X ; ; =). Observe that if μ A , μ B ∈ {0, 1} X , that is, A and B are in P(X ),
then it follows
μ A μ B ⇔ A ⊂ B; μ A = μ B ⇔ A = B,
and
x ∈ A ⇔ μ A (x) = 1; x∈
/ A ⇔ μ A (x) = 0.
The classical symbol ∈ is the fuzzy symbol ∈, and ∈

/ is ∈.
1 0
For example, with the fuzzy set P given by function μ in the next figure it is x ∈
/ P
∼ ∼
if 0 x 3, and 7 x 10, but x ∈ P if 4 x 6, and x ∈ P , if
∼ μ(x) ∼
x ∈ (3, 4) ∪ (6, 7) with 0 < μ(x) < 1. If x = 3.5, since it is μ(x) = x − 3, when
x ∈ (3, 4), it is 3.5 ∈ P .
0.5 ∼
• The height of μ ∈ F(X ) is H (μ) = Sup μ(x) = sup μ. In the last example, it is
x∈X
H (μ) = 1. In the finite example
μ = 0.7/x1 + 0.9/x2 + 0.7/x3 ,
in X = {x1 , x2 , x3 , x4 }, it is H (μ) = 0.9. In a case like
it is H (μ) = 1, although there is not any x ∈ R such that μ(x) = 1. If there is

some x ∈ X such that μ(x) = 1, it is said that μ is a normalized fuzzy set.
• In the case X is finite, X = {x1 , . . . , xn }, the number

n
|μ| = μ(xi )
i=1
sigma-count, of μ, a name coming from the fact that if

is the crisp-cardinality, or
n
A ⊆ X has p elements it is i=1 μ(xi ) = p. Obviously, μ∅ = μ0 , gives |μ0 | = 0,
μ X = μ1 , gives |μ1 | = n, and, μ σ implies |μ| |σ|.
Remark 2.2.1 The pointwise definition of fuzzy sets inclusion implies that, for exam-
ple, the fuzzy sets
μ = 0.7/x1 + 0.8/x2 + 1/x3 + 0.7/x4

σ = 0.70001/x1 + 0.7/x2 + 1/x3 + 0.6/x4
in X , do not verify μ σ although it is σ(x2 ) < μ(x2 ), σ(x3 ) = μ(x3 ), σ(x4 ) <
μ(x4 ), but σ(x1 ) > μ(x1 ), with σ(x1 ) − μ(x1 ) = 0.00001. Pointwise ‘inclusion’ is
strongly affected by very small variations of the membership values. Actually, it is
not a flexible, or fuzzy, concept, but a crisp one.
Because of this, it could be preferable to take the inclusion of fuzzy sets as an
gradable concept r (r ∈ [0, 1]), and a used definition of which is
| min(μ, σ)|
μ r σ ⇔ r,
|μ|
n
with | min(μ, σ)| = i=1 min(μ(xi ), σ(xi )).
In last example, it is | min(μ, σ)| = 0.7 + 0.7 + 1 + 0.6 = 3, |μ| = 0.7 + 0.8 +
1 + 0.7 = 3.2, and r = 3/3.2 = 0.9375 ≈ 0.94. That is, μ σ : μ ‘is almost
0.9375
included in’ σ.
Since |σ| = 0.70001 + 0.7 + 1 + 0.6 = 3.00001, it is | min(μ,σ)|
|σ| = 0.9999, or
σ 0.9999 μ. That is, σ is more included in μ, than μ is included in σ!
Remark 2.2.2 Of course, if μ σ, it is min(μ, σ) = μ, and r = 1, that is,
μ σ ⇒ μ σ.
1
Nevertheless, since it is only

min(μ(xi ), σ(xi )) min( μ(xi ), σ(xi )),
from μ σ (or | min(μ, σ)| |μ|) it does not necessarily follow μ σ.

1
Let us show an example with crisp subsets. If X = {1, 2, 3, 4, 5, 6, 7}, and A =
{1, 3, 5, 7}, B = {1, 3, 5, 6}, it is

7
7
7
min(μ A (i), μ B (i)) = 3, μ B (i) = 4, μ A (i) = 4.
i=1 i=1 i=1
hence
μ A 3 μ B , or A ⊂ 3 B
4 4
μ B 3 μ A , or B ⊂ 3 A
4 4
2.2.2 Algebras of Fuzzy Sets
Once F(X ) = ([0, 1] X ; ; =) is taken, a general algebra of fuzzy sets comes from
endowing F(X ) with three operations:
1. : [0, 1] X → [0, 1] X ,
2. · : [0, 1] X × [0, 1] X → [0, 1] X ,
3. + : [0, 1] X × [0, 1] X → [0, 1] X ,
respectively called the complement μ of μ, the intersection μ · σ of ‘μ and σ’, and
the union μ + σ of ‘μ or σ’. Then ([0, 1] X ; ; =; ·; +; ), is called an algebra of
fuzzy sets, provided the following laws do hold:
(a) If μ σ, then γ · μ γ · σ, and μ · γ σ · γ, for all γ ∈ [0, 1] X

(b) If μ σ, then μ + γ σ + γ, and γ + μ γ + σ for all γ ∈ [0, 1] X
(c) If μ σ, then σ μ
(d) For any μ ∈ [0, 1] X , μ · μ1 = μ1 · μ = μ, μ + μ0 = μ0 + μ = μ
(e) For all μ A , μ B ∈ {0, 1} X , μA = μ Ac , μ A · μ B = μ A∩B , μ A + μ B = μ A∪B
(preservation of the classical case).
Remark 2.2.3 It is not difficult to prove that no general algebra of fuzzy sets is a
Boolean algebra. The proof comes from the fact that to be a Boolean algebra would
imply μ · μ = μ0 and μ + μ = μ1 for all μ ∈ [0, 1] X , and consists in finding some
μ for which these equalities are not satisfied.
Remark 2.2.4 It is immediate to prove that μ · μ0 = μ0 · μ = μ0 , μ + μ1 = μ1 + μ =

μ1 for all μ ∈ [0, 1] X
Remark 2.2.5 Notice that the laws μ · σ = σ · μ (commutativity of the intersection),

μ + σ = σ + μ (commutativity of the union), and μ = μ (involution of the
complement) are not supposed to be always verified. Nor it is supposed that the
algebras ([0, 1] X ·, +, ) are dual ones, that is, the so-called De Morgan laws,
(μ + σ) = μ · σ , (μ · σ) = μ + σ ,
are not supposed to hold in general. It is neither supported μ · μ = μ, and σ + σ = σ,

nor μ · (σ · λ) = (μ · σ) · λ and μ + (σ + λ) = (μ + σ) + λ, nor that · and + are
associative, μ · (σ · λ) = (μ · σ) · λ, and μ + (σ + λ) = (μ + σ) + λ.
Theorem 2.2.6 For any μ, σ ∈ [0, 1] X : μ · σ min(μ, σ) max(μ, σ) μ + σ.
Proof From μ μ1 (μ(x) 1, for all x ∈ X ), follows μ · σ μ1 · σ = σ. From

σ μ1 , follows μ · σ μ1 · μ = μ. Thus,
(μ · σ)(x) σ(x), (μ · σ)(x) μ(x)

⇒ (μ · σ)(x) min(μ(x), σ(x)) = min(μ, σ)(x),
or μ · σ min(μ, σ). Hence, the operation min is the greatest possible intersection
of fuzzy sets. Analogously, μ0 μ, μ0 σ ⇒ μ0 + σ = σ μ + σ, μ + μ0 = μ
μ + σ, and max(μ, σ) μ + σ: The operation max is the smallest possible union of
fuzzy sets.
Obviously, for all μ, σ ∈ [0, 1] X : μ · σ μ μ + σ, μ · σ σ μ + σ.

Theorem 2.2.7 An operation ∗ : [0, 1] X × [0, 1] X → [0, 1] X is called idempotent
if and only if μ ∗ μ = μ, for all μ ∈ [0, 1] X
• The intersection · is idempotent if and only if · = min
• The union + is idempotent if and only if + = max
Proof The operations min, and max are obviously idempotent. Let us show that if ·
is idempotent, it must be · = min. By Theorem 2.2.6, it is always μ · σ min(μ, σ),
and the idempotency of · implies
min(μ, σ) · min(μ, σ) = min(μ, σ).
But min(μ, σ) μ, min(μ, σ) σ, imply min(μ, σ) · min(μ, σ) μ · σ, that is
min(μ, σ) μ · σ,
and, by Theorem 2.2.6, min(μ, σ) = μ · σ. A similar proof applies to + and max.

Theorem 2.2.8 (Absortion laws)
• μ · (μ + σ) = μ holds for all μ, σ ∈ [0, 1] X ⇔ · = min
• μ + (μ · σ) = μ holds for all μ, σ ∈ [0, 1] X ⇔ + = max
Proof If · = min, the formula min(μ, μ + σ) = μ does hold, since μ μ + σ.
Provided it is always μ · (μ+σ) = μ, taking σ = μ0 follows μ · (μ+μ0 ) = μ · μ = μ,
that holds if and only if · = min. If + = max, the formula max(μ, μ · σ) = μ does
hold since μ · σ μ. Provided it is always μ + (μ · σ) = μ, taking σ = μ1 follows
μ + (μ · μ1 ) = μ + μ = μ, that holds if and only if + = max.
Theorem 2.2.9 (Duality, or De Morgan laws) Provided the complement is invo-
lutive ((μ ) = μ = μ, for all μ ∈ [0, 1] X ), the algebra of fuzzy sets ([0, 1] X ,
min, max, ) is a dual algebra.
Proof If is involutive, from μ + σ = (μ · σ) it follows μ · σ = (μ + σ) since
μ + σ = μ + σ = (μ · σ ) , hence (μ + σ) = μ · σ . The converse is proven in
the same way. And the two De Morgan laws
(μ · σ) = μ + σ , (μ + σ) = μ · σ
result equivalent. Then, it is enough to prove max(μ, σ) = (min(μ , σ )) , for all
μ, σ in [0, 1] X . Since
min(μ , σ ) μ , min(μ , σ ) σ ,
it follows
μ = μ (min(μ , σ )) , σ = σ (min(μ , σ )) ,
and
max(μ, σ) (min(μ , σ )) (2.1)
On the other hand,
μ max(μ, σ) ⇒ (max(μ, σ)) μ

σ max(μ, σ) ⇒ (max(μ, σ)) σ
imply (max(μ, σ)) min(μ , σ ), or
(min(μ , σ )) max(μ, σ). (2.2)
Now, from (2.2) and (2.1), follows the result.
Theorem 2.2.10 (Kleene’s Law) In all general algebra of fuzzy sets it holds the law
μ · μ σ + σ ,
for all μ, σ in [0, 1] X .
Proof We have to prove that, for any x ∈ X , it is (μ · μ )(x) (σ + σ )(x). But

it only can be either μ(x) σ(x), or σ(x) μ(x) for each x ∈ X . In the firs
case, it is (μ · μ )(x) min(μ(x), μ (x)) μ(x) σ(x) max(σ(x), σ (x)) =
(max(σ, σ )(x) (σ + σ )(x). In the second case, it is μ (x) σ (x), and
(μ · μ )(x) min(μ(x), μ (x)) μ (x) σ (x) max(σ(x), σ (x)) =
(max(σ, σ )(x) (σ + σ )(x). Notice that provided μ and σ were crisp sets, the
Kleene’s law is reduced to μ0 μ1 .
Remark 2.2.11 Concerning duality, Theorem 2.2.9 only states that the algebra given
by the triplets (min, max, ), with involutive, are dual algebras. But they are not the
only dual algebras. For example, with · = product,
(μ · σ)(x) = μ(x) · σ(x), ∀x ∈ X,
it is easy to proof that taking μ (x) = 1 − μ(x), and
(μ + σ)(x) = 1 − (1 − μ(x))(1 − σ(x)) = μ(x) + σ(x) − μ(x) · σ(x),
it is ([0, 1] X , ·, +, ) an algebra of fuzzy sets that since it is
μ + σ = (μ · σ ) ,
is a dual algebra. Nevertheless, with μ (x) = 1 − μ(x), (μ · σ)(x) = μ(x) · σ(x),

and (μ + σ)(x) = max(μ(x), σ(x)), we get an algebra that is not dual since
(μ · σ ) (x) = μ(x) + σ(x) − μ(x) · σ(x)
does not coincides with max(μ(x), σ(x)), as it is easy to see.
Remark 2.2.12 It is easy to prove that, for each algebra of fuzzy sets ([0, 1] X , ·, +, ),
the operation
μ + σ = (μ · σ ) ,
gives the new algebra ([0, 1] X , ·, + , ). If the complement is involutive (μ = μ),
then (μ + σ) = μ · σ .
Analogously, with the operation μ · σ = (μ + σ ) , one has the new algebra
([0, 1] X , · , + , ) and, if is involutive, (μ · σ) = μ + σ .
2.2.3 Non-contradiction and Excluded-Middle
An statement is self-contradictory whenever entails its negation. For example, the

only classical set that is self-contradictory is the empty one:
A ⊆ Ac ⇒ A ∩ A ⊆ A ∩ Ac ⇒ A ⊆ ∅ ⇒ A = ∅.
Perhaps, this is the reason of the difficulty children do have on accepting that ∅ is a
set!
Within an algebra of fuzzy sets there are many self-contradictory fuzzy sets. For
example, with N = 1 − id it is
1
μ μ ⇔ μ(x) 1 − μ(x) ⇔ μ(x) , ∀x ∈ X,
2
hence: μ is self-contradictory if and only if μ μ 1 . Analogously, with the strong
2
negation N (x) = 1−x
1+x , it is
1 − μ(x)
μ μ ⇔ μ(x) ⇔ μ(x)2 + 2μ(x) − 1 0
1 + μ(x)
√
⇔ μ(x) 2 − 1, ∀x ∈ X,
that is, μ is self-contradictory if and only id μ μ√2−1 .

Notice that 1/2 is the fixed-point
√ of the strong negation N = 1 − id (1 − n =
n ⇔ n = 1/2), and that √ 2 − 1 is the fixed-point of the strong negation
N = 1−id
1+id ( 1+n = n ⇔ n = 2 − 1).
1−n
Notice that if μ P μa P , since it is always supposed that μa P μnot P , it follows

μ P μnot P , and μ P is self-contradictory.
The classical principle of non-contradiction, “it is impossible to have both an
statement and its negation”, could be interpreted as “P and not P is impossible”, or
“P and not P is self-contradictory”. All general algebras of fuzzy sets do verify the
principle of non-contradiction once stated in this form.
Theorem 2.2.13 If ([0, 1] X , ·, +, ) is an algebra of fuzzy sets, it holds the principle

of non-contradiction stated by: μ · μ (μ · μ ) for all μ ∈ [0, 1] X . That is, for all
μ ∈ [0, 1] X , μ · μ is self-contradictory.
Proof It is μ · μ min(μ, μ ) μ, and μ · μ min(μ, μ ) μ ; from the

first inequality it follows μ (μ · μ ) , and then with the second follows μ · μ
(μ · μ ) .
Notice that no additional hypotheses on the connective ·, and the complement ,

are needed for the proof of this theorem. In the algebras of fuzzy sets the non-
contradiction principle is a theorem: the algebra’s axioms imply the principle. It
is not true, as it is sometimes stated, that fuzzy sets do not verify the principle of
non-contradiction in which science is based.
The classical principle of Excluded-Middle, “It is always P or not P”, can be
interpreted as “Not (P or Not P) is a self-contradiction”and it is verified by all algebra
of fuzzy sets.
Theorem 2.2.14 If ([0, 1] X , ·, +, ) is an algebra of fuzzy sets, it holds the principle
of Excluded-Middle stated by: (μ + μ ) ((μ + μ ) ) for all μ ∈ [0, 1] X . That is,
for all μ ∈ [0, 1] X , (μ + μ ) is self-contradictory.
Proof It is,
• μ max(μ, μ ) μ + μ ⇒ (μ + μ ) μ ⇒ (μ ) ((μ + μ ) )
• μ max(μ, μ ) μ + μ ⇒ (μ + μ ) (μ )
then (μ + μ ) ((μ + μ ) ) .
Notice that no additional hypotheses on the connective +, and the complement ,

are needed for the proof of this theorem: In the algebra of fuzzy sets the excluded-
middle principle is a theorem. In conclusion,
In all algebra of fuzzy sets ([0, 1] X , ·, +, ),
• The logic principles of non-contradiction and excluded-middle are theorems, once
stated through the concept of self-contradiction.
A very different situation appears if these two principles are stated as it is currently
done within logic and classical set theory, that is, by stating
• “P and not P” is false
• “P or not P” is true,
or,
• There is no x in X such that “x is P and x is not P”
• For all x in X it is “x is P or x is not P”
translated into
• (μ P · μnot P )(x) = 0, for all x in X
• (μ P + μnot P )(x) = 1, for all x in X
that corresponds to “solve” the equations with fuzzy sets,
μ · μ = μ0 , μ + μ = μ1 ,
that is, to find for which intersections · and which unions +, these equations do hold.
Of course, they do not hold in all cases, for example, with N = 1 − id,
• If · = min, it is not always min(μ(x), 1 − μ(x)) = 0,
• If + = max, it is not always max(μ(x), 1 − μ(x)) = 1,
• If · = W, (W (a, b) = max(0, a + b − 1)) it is W (a, 1 − a) = max(0, a + 1 −
a − 1)) = 0, and W (μ(x), 1 − μ(x))) = 0 for all x in X
• If + = W ∗ , (W ∗ (a, b) = min(1, a + b)), it is W ∗ (μ(x), 1 − μ(x))) = 1 for all x
in X .
That is, there are algebras of fuzzy sets where this forms of non-contradiction
or excluded-middle hold, and algebras where this principles do not jointly hold.
In the algebras with the triplet (min, max, 1 − id) do not hold both principles, in
the algebras with (W, max, 1 − id) it holds the principle of non-contradiction but
not that of excluded-middle, in the algebras with (min, W ∗ , 1 − id) it holds the
excluded-middle but not the principle of non-contradiction, and in the algebras with
(W, W ∗ , 1 − id) both principles hold.
Remark 2.2.15 Let us show that with μ ∈ {0, 1} X it is always μ · μ = μ0 and

μ + μ = μ1 . If μ ∈ {0, 1} X , denote A = {x ∈ X ; μ(x) = 1}. Obviously, μ = μ A
and μ = μ Ac , hence
• (μ · μ ) = μ A · μ Ac = μ A∩Ac = μ∅ = μ0
• (μ + μ ) = μ A + μ Ac = μ A∪Ac = μ X = μ1 .
Remark 2.2.16 Results in Theorems 2.2.13 and 2.2.14 challenge the usual statement
that in fuzzy sets the basic principles of Non-contradiction and Excluded-middle fail.
A statement that could conduct to believe that fuzzy set algebras are not properly
grounded in a solid ground.
The fact is, notwithstanding, that these two principles were established before
the current ways of considering the problems of logic and, of course, before the
nomenclature of set theory. In set theory (or Boolean algebras), A ∩ Ac = ∅ and
A ∩ Ac ⊂ (A ∩ Ac )c are equivalent formulas since, as it was said, it is
B = ∅ ⇔ B ⊂ Bc,
an equivalence only verified in the setting of ortholattices (of which Boolean algebras
are a particular case), but that does not hold on weaker algebraic structures like it
is, for example, the case of the above defined algebras of fuzzy sets. Let us call
‘restricted’ the principles stated by μ · μ = μ0 and μ + μ = μ1 .
2.2.4 Decomposable Algebras
Definition 2.2.17 An operation with fuzzy sets ∗ : [0, 1] X × [0, 1] X → [0, 1] X

is decomposable, or functionally expressible, if it exists a numerical operation
∗:
[0, 1] × [0, 1] → [0, 1], such that
(μ ∗ σ)(x) =
∗(μ(x), σ(x)),
for all μ, σ in [0, 1] X and all x in X . Of course, by this formula, a numerical operation

∗ allows to define an operation ∗ for fuzzy sets.
For example, the operation min in [0, 1] X is decomposable since, by definition, it is
(min(μ, σ))(x) = min(μ(x), σ(x))
for all μ, σ in [0, 1] X and all x in X .
Definition 2.2.18 A function f : [0, 1] X → [0, 1] X is decomposable, or function-

ally expressible, if it exists a numerical function
f : [0, 1] → [0, 1], such that
( f (μ))(x) =
f (μ(x)),
for all μ in [0, 1] X and all x in X .
For example, the function defined by
μ (x) = 1 − μ(x),
is decomposable because of N0 = 1 − id gives μ (x) = N0 (μ(x)). With X = [0, 1],

the function defined by μ∗ (x) = 1 − μ(1 − x) is not decomposable, since if it were
such, that is, if there is N : [0, 1] → [0, 1] such that 1 − μ(1 − x) = N (μ(x)), it
suffices to take μ, σ ∈ [0, 1] X such that
• μ(0) = 0, μ(1) = 1 ⇒ N (0) = 1 − μ(1 − 0) = 1 − μ(1) = 0, N (1) =
1 − μ(1 − 1) = 1 − μ(0) = 1
• σ(0) = 1, σ(1) = 0 ⇒ N (0) = 1 − σ(1 − 0) = 1 − σ(1) = 1, N (1) =
1 − σ(1 − 1) = 1 − σ(0) = 0
that is absurd.
The algebras of fuzzy sets ([0, 1] X , ·, +, ), can be
Decomposable if the three operation ·, +, are decomposable
Partially decomposable if at least one of the three operations ·, +, is decom-
posable
Non decomposable if no one of the three operations is decomposable
In what follows we will only deal with decomposable algebras, that is, such that:
There are three functions F : [0, 1]×[0, 1] → [0, 1], G : [0, 1]×[0, 1] → [0, 1],
and N : [0, 1] → [0, 1] with which
(μ · σ)(x) = F(μ(x), σ(x)), (μ + σ)(x) = G(μ(x), σ(x)), μ (x) = N (μ(x)),
for all μ, σ ∈ [0, 1] X , and all x ∈ X . For short, μ · σ = F ◦ (μ × σ), μ + σ =

G ◦ (μ × σ), μ = N ◦ μ.
In these cases, instead of ([0, 1] X , ·, +, ) it is written ([0, 1] X , F, G, N ).
The laws verified by the triplet (·, +, ) force analogous laws for the triplet
(F, G, N ). For example, linked to the axioms, it is
(a) If a b, then F(a, c) F(b, c), F(c, a) F(c, b), for all c ∈ [0, 1]
(b) If a b, then G(a, c) G(b, c), G(c, a) G(c, b), for all c ∈ [0, 1]
(c) If a b, then N (b) N (a)
(d) F(1, a) = F(a, 1) = a, G(0, a) = G(a, 0) = a
(e) N (0) = 1, N (1) = 0, F(0, a) = F(a, 0) = 0, G(1, a) = G(a, 1) = 1
and linked to the Theorems 2.2.6–2.2.13 it is
• F(a, b) min(a, b) max(a, b) G(a, b) for all a, b in [0, 1]. In particular, it
is F G.
• F is idempotent (F(a, a) = a, for all a in [0, 1]), if and only if F = min.
• G is idempotent (G(a, a) = a, for all a in [0, 1]), if and only if G = max.
• It is F(a, G(a, b)) = a, for all a, b in [0, 1], if and only if F = min.
• It is G(a, F(a, b)) = a, for all a, b in [0, 1], if and only if G = max.
• A triplet (F, G, N ) is called dual, or De Morgan triplet, if F = N ◦ G ◦ (N × N ),
or, G = N ◦ F ◦ (N × N ), that is, F(a, b) = N (G(N (a), N (b)), or G(a, b) =
N (F(N (a), N (b)), for all a, b in [0, 1]. Notice that, in this case, it is enough to
know N and F to have G, or N and G to have F.
• If N is involutive (N (N (a)) = a, for all a ∈ [0, 1] or N 2 = N ), the triplet
(min, max, N ) is a dual one.
• All triplet (F, G, N ) verifies F(a, N (a)) G(b, N (b)), for all a, b in [0, 1],
• It is F(a, N (a)) N (F(a, N (a))), for all a in [0, 1]
• It is N (F(a, N (a))) N (N (F(a, N (a)))), for all a in [0, 1]
• Given N involutive and F, and denoting by G N the dual of F respect to N ,
G N (a, b) = N (F(N (a), N (b))), it follows G N (N (a), a) N (G N (N (a), a)),
that with
μ + σ = G N ◦ (μ × σ),
gives μ + μ (μ + μ) .

• It is always F(a, N (a)) G(b, N (b)), for all a, b in [0, 1].
Remark 2.2.19 With classical sets, from A∩ B ⊆ A∪ B it results (A∪ B)∪(A∩ B) =

A∪ B. In a decomposable theory, to have the analogous law (μ+σ)+(μ · σ) = μ+σ
it should be G(G(a, b), F(a, b)) = G(a, b), for all a, b in [0, 1], that is verified if,
for example, G = max, F = min, or G = W ∗ , F = W
Remark 2.2.20 Let us show an example of a ‘union’ that is not-decomposable. Define

max(μ(x), σ(x)), if μ or σ are in{0, 1} X (crisp)
(μ + σ)(x) =
max(H (μ), H (σ)), otherwise.
It is easy to show that this operation verifies the laws b, d and e in Sect. 2.1.1. Hence,
it is a union for fuzzy sets that, in addition, is commutative. It is not idempotent,
since if μ ∈ [0, 1] X − {0, 1} X , it is (μ + μ)(x) = H (μ) = μ(x). It does not exist a
function G : [0, 1] × [0, 1] → [0, 1] such that
(μ + σ)(x) = G(μ(x), σ(x))
for all x ∈ X and all μ, σ[0, 1] X . Indeed, let us suppose that such a G does exist,
and take μ = μ0.5 . Then
• With σ = μ0 , it is (μ + σ)(x) = max( 21 , 0) = 21 . Hence G( 21 , 0) = 21 .
• With σ(x) = x, is (μ + σ)(x) = max(H (μ), H (σ)) = max( 21 , 1) = 1, and
(μ + σ)(0) = 1 = G( 21 , 0), that is absurd.
To have a not-decomposable ‘intersection’, it is enough to define, with μ = 1 − μ,
the dual operation,
(μ · σ)(x) = [(μ + σ )] (x) = 1 − (μ + σ )(x)

min(μ(x), σ(x)), if μ or σ are in{0, 1} X
=
max(H (μ ), H (σ )), otherwise.
2.2.5 Standard Algebras of Fuzzy Sets
An standard algebra of fuzzy sets is a decomposable algebra of fuzzy sets such that:
1. μ · σ = σ · μ, for all μ, σ in [0, 1] X (· is commutative)
2. μ + σ = σ + μ, for all μ, σ in [0, 1] X (+ is commutative)
3. μ · (σ · λ) = (μ · σ) · λ, for all μ, σ, λ in [0, 1] X (· is associative)
4. μ + (σ + λ) = (μ + σ) + λ, for all μ, σ, λ in [0, 1] X (+ is associative)
5. μ = μ, for all μ in [0, 1] X ( is involutive).
Hence, writing
μ · σ = T ◦ (μ × σ), μ + σ = S ◦ (μ × σ), μ = N ◦ μ,
functions T, S : [0, 1] × [0, 1] → [0, 1] and N : [0, 1] → [0, 1], in addition to the
corresponding general properties stated before, must verify
• T is commutative, S is commutative
• T is associative, S is associative
• N is involutive,
that is:
• T (a, b) = T (b, a), S(a, b) = S(b, a), for all a, b in [0, 1]
• T (a, T (b, c)) = T (T (a, b), c), S(a, S(b, c)) = S(S(a, b), c), for all a, b, c in
[0, 1]
• N (N (a)) = a, for all a in [0, 1], or N ◦ N = id, or N = N −1 .
Functions T and S are called t-norms and t-conorms, respectively. Functions N
are strong negations. Hence, ([0, 1], T, ) is an ordered semigroup with neutral 1,
and absorbent 0, and ([0, 1], S, ) is also an ordered semigroup but with neutral 0
and absorbent 1. Since N (1) = 0, it seems that this two kind of ordered semigroups
should show some character of duality. This duality goes in the way:
• If T is a t-norm, TN = N ◦ S ◦ (N × N ) is a t-conorm
• If S is a t-conorm, S N = N ◦ S ◦ (N × N ) is a t-norm
that are easy to prove. Of course, from Sect. 2.1.4,
• If T is a t-norm, T min, and min is a t-norm
• If S is a t-conorm, max S, and max is a t-conorm
Hence, for all t-norm T and all t-conorm S:
T min max S,
in particular, T S.1 Even more, the function

⎧
⎨ b, if a = 1
min(a, b), if a = 1 or b = 1
Z (a, b) = a, if b = 1 =
⎩ 0, otherwise,
0, otherwise,
is obviously a t-norm such that Z T for all t-norm T. Consequently,

⎧
⎨ b, if a = 0
Z ∗ (a, b) = 1 − Z (1 − a, 1 − b) = a, if b = 0
⎩
1, otherwise,

max(a, b), if a = 0 or b = 0
=
1, otherwise,
is a t-conorm such that S Z ∗ for all t-conorm S. Hence, for all t-norm T and all
t-conorm S,
Z T min max S Z ∗ .
1 Notice that T S mean T (a, b) S(a, b), for all (a, b) ∈ [0, 1] × [0, 1].
Notice that min(max) is a continuous t-norm (t-conorm), but Z (Z ∗ ) is a discontinuous

t-norm (t-conorm). The operations in [0, 1] given by
• T pr od (a, b) = pr od (a, b) = a · b
• TW (a, b) = W (a, b) = max(0, a + b − 1) = (max(0, Sum − 1))(a, b),
are also continuous t-norms. Then, the dual operations,
∗ (a, b) = 1 − T
• T pr od pr od (1 − a, 1 − b) = 1 − (1 − a) · (1 − b) = a + b − a · b =
(Sum − pr od)(a, b)
• W ∗ (a, b) = 1 − W (1 − a, 1 − b) = min(1, a + b) = min(1, Sum(a, b))
are continuous t-conorms. Since it is easy to prove that
W T pr od min,
∗
it follows max T pr ∗
od W , and
∗ ∗
Z W T pr od min max T pr od W .
Remark 2.2.21 Since Z (0.5, 0.5) = 0, t-norm Z has zero-divisors. Analogously,

from
W (a, b) = 0 ⇔ max(0, a + b − 1) = 0 ⇔ a + b 1,
it follows that t-norm W has zero-divisors, for example, W (0.5, 0.4) = 0; t-norms
min and T pr od do not have zero-divisors:
• T pr od (a, b) = 0 ⇔ a = 0, or b = 0
• min(a, b) = 0 ⇔ a = 0, or b = 0
Proposition 2.2.22 The only idempotent t-norm, i.e. T (a, a) = a, for all a ∈ [0, 1],
is T = min.
Proof If T is idempotent, min(a, b) = T (min(a, b), min(a, b)) T (a, b) since

min(a, b) a, and min(a, b) b. Hence, min(a, b) T (a, b) min(a, b)
implies T = min.
Proposition 2.2.23 The only idempotent t-conorm, i.e. S(a, a) = a, for all a ∈
[0, 1], is S = max.
Proof If S is idempotent, max(a, b) = S(max(a, b), max(a, b)) ≥ S(a, b), since
max(a, b) ≥ a, and max(a, b) ≥ b. Hence, max(a, b) ≥ S(a, b) ≥ max(a, b)
implies S = max.
Remark 2.2.24 t-norms can be continuous, like min, T pr od , and W , or discontinuous,

like Z . They can have zero-divisors, like W and Z , or not like min and T pr od . They
can have all elements in [0, 1] idempotent (only T = min), only have the idempotents
0 and 1 (like T pr od and W ), or have some idempotents different from 0 and 1. In

any case, since it is always T (0, 0) = 0 and T (1, 1) = 1, 0 and 1 are idempotent
elements for all t-norms.
Remark 2.2.25 Analogous considerations can be made for t-conorms. There are
∗
discontinuous t-conorms like Z ∗ , and continuous ones like T pr ∗
od and W . The only
for which all elements in [0, 1] are idempotent is S = max. Since S(0, 0) = 0
and S(1, 1) = 1, 0 and 1 are always idempotent, and there are t-conorms that only
have these two idempotents (like T pr∗ ∗
od and W ), as well as those that have some
idempotents different from 0,1. There are t-conorms without one-divisors, like max
∗ , and t-conorms with one-divisors like W ∗ , for example, W ∗ (0.5, 0.5) =
and T pr od
min(1, 1) = 1.
Remark 2.2.26 There is not a characterization theorem for all t-norms (t-conorms),
but it is a characterization of the continuous t-norms (t-conorms) that will be pre-
sented by means of the following, and easy to prove, results:
• If ϕ : [0, 1] → [0, 1] verifies, (1) If x y, then ϕ(x) ϕ(y), (2) ϕ is bijective,
(3) ϕ(0) = 0, ϕ(1) = 1 (ϕ is an order-automorphism of the ordered interval
([0, 1], )), and T is a t-norm, then Tϕ = ϕ−1 ◦ T ◦ (ϕ × ϕ) is also a t-norm.
Given T , the set {Tϕ ; ϕ an order-automorphism} is called the family of T .
• T is a continuous t-norm if and only if all t-norms Tϕ are continuous.
• If S is a t-conorm, then Sϕ = ϕ−1 ◦ S ◦ (ϕ × ϕ) is also a t-conorm, and S
is continuous if and only if all t-conorms Sϕ are continuous, the set {Sϕ ; ϕ an
order-automorphism} is called the family of S.
In particular,
• The family of T = min, is reduced to the only t-norm min, since ϕ−1 (min(ϕ(a),
ϕ(b)) = min(ϕ−1 (ϕ(a)), ϕ−1 (ϕ(b))) = min(a, b)
• The family of T pr od contains all continuous t-norms of the form prodϕ (a, b) =
ϕ−1 (ϕ(a).(ϕ(b)).
• The family of W contains all t-norms of the form Wϕ (a, b) = ϕ−1 (W (ϕ(a),
ϕ(b)) = ϕ−1 (max(0, ϕ(a) + ϕ(b) − 1)), and all of them are continuous t-norms.
Notice that no t-norm in the family { pr odϕ } has zero-divisors, since pr odϕ (a, b) =
0 ⇔ ϕ(a) · ϕ(b) = 0 ⇔ ϕ(a) = 0 or ϕ(b) = 0 ⇔ a = 0, or b = 0. Instead
all t-norms Wϕ have zero-divisors, since Wϕ (a, b) = 0 ⇔ max(0, ϕ(a) + ϕ(b) −
1) ⇔ ϕ(a) + ϕ(b) 1. Of course, neither t-norms pr odϕ , nor Wϕ , have more
idempotents than 0 and 1:
• a = Wϕ (a, a) = ϕ−1 (max(0, 2ϕ(a) − 1)) ⇔ ϕ(a) = max((0, 2ϕ(a) − 1) ⇔
ϕ(a) = 0 or ϕ(a) = 1 or a = 0 or a = 1.
• a = pr odϕ (a, a) = ϕ−1 (ϕ(a).(ϕ(a)) ⇔ ϕ(a) = ϕ(a).(ϕ(a) ⇔ ϕ(a) = 0 or
ϕ(a) = 1 or a = 0 or a = 1.
Analogously,
• The family of S = max, only contains this t-conorm.

∗
• The family of T pr od contains all t-conorms of the form
pr odϕ∗ (a, a) = ϕ−1 ( pr od ∗ (ϕ(a), ϕ(b))) = ϕ−1 (ϕ(a) + ϕ(b) − ϕ(a) · ϕ(b))
• The family of W ∗ contains all t-conorms of the form
W ∗ (a, a) = ϕ−1 (W ∗ (ϕ(a), ϕ(b)) = ϕ−1 (min(1, ϕ(a) + ϕ(b)))
Remark 2.2.27 The order-automorphism ϕ plays the role of a functional parameter.

By taking, ϕ(x) = x r , it follows, for example,

Wϕ (a, b) = r
max((0, a r + br − 1), Wϕ∗ (a, b) = r
min((1, a r + br )
giving a family of t-norms (t-conorms) depending on the numerical parameter r > 0.

Notice that with ϕ(x) = x r ,
√
r r
Pr odϕ (a, b) = a · br = a · b = Pr od(a, b),
but
√
Pr odϕ∗ (a, b) = ϕ−1 (ϕ(a) + ϕ(b) − ϕ(a) · ϕ(b)) = a + br − a r · br .
r r
2.2.6 Strong Negations
As it was said before, an strong negation is a function N : [0, 1] → [0, 1] such that
• N (0) = 1
• If a b, then N (b) N (a)
• N (N (a)) = a, for all a ∈ [0, 1], or N 2 = id.
Notice that N 2 = id is equivalent to N = N −1 , that shows N is a continuous
function: It is N (1) = N (N (0)) = 0, and if a < b it should be N (b) < N (a) since
N (b) = N (a) would imply N (N (b)) = N (N (a)), or a = b. Hence, N is strictly
decreasing.
Since N is continuous, the equation N (x) = x has solutions, but there is only
one. Suppose N (x1 ) = x1 and N (x2 ) = x2 . Either x1 x2 , or x2 < x1 . In the
first case, it follows N (x2 ) N (x1 ), or x2 x1 , and x1 = x2 . In the second case,
N (x1 ) < N (x2 ), or x1 < x2 , that is absurd. Then, each strong negation has a single
fixed point N (n) = n, in the open interval (0, 1), since N (0) = 1, N (1) = 0 show
that 0 and 1 are not fixed points.
Remark 2.2.28 In the classical case (a Boolean algebra L, or a power set P(X )), the
transformation
F : P(X ) → P(X ), F(A) = Ac ,
has no fixed points, since A = Ac implies A ∩ A = A ∩ Ac , or A = ∅, and ∅c = X .

Nevertheless, with fuzzy sets
F : [0, 1] X → [0, 1] X , F(μ) = μ = N ◦ μ,
the equation μ = μ , N (μ(x)) = μ(x) for all x in X , has the only solution μ(x) = n
for all x in X , that is μ = μn , with n = N (n) the fix point of N .
In the fuzzy case, that mapping F shows a kind of symmetry that is not in the
crisp case.
An order-automorphism of the ordered unit interval ([0, 1], ) ϕ : [0, 1] →
[0, 1], verifies by definition,
• ϕ(0) = 0, ϕ(1) = 1
• If a < b, then ϕ(a) < ϕ(b).
Hence, ϕ is continuous, and its inverse function ϕ−1 verifies,
• ϕ−1 (0) = 0, ϕ−1 (1) = 1
• If a < b, then ϕ−1 (a) < ϕ−1 (b).
Let us denote by Nϕ the function Nϕ : [0, 1] → [0, 1] defined by
Nϕ (a) = ϕ−1 (1 − ϕ(a)), for all a ∈ [0, 1]
Proposition 2.2.29 Nϕ is a strong negation.
Proof It is Nϕ (0) = ϕ−1 (1 − ϕ(0)) = ϕ−1 (1) = 1. In addition, if a b, it follows

1 − ϕ(b) 1 − ϕ(a), and ϕ−1 (1 − ϕ(b)) ϕ−1 (1 − ϕ(a)), or Nϕ (b) Nϕ (a).
Finally,
Nϕ (Nϕ (a)) = Nϕ (ϕ−1 (1 − ϕ(a))) = ϕ−1 (1 − ϕ(ϕ−1 (1 − ϕ(a))))

= ϕ−1 (1 − 1 + ϕ(a)) = ϕ−1 (ϕ(a)) = a
Theorem 2.2.30 If N is a strong negation, there exist order-automorphisms ϕ such

that N = Nϕ .
Proof Let it be n = N (n) ∈ (0, 1) the fixed point of N , and consider an strictly
non-decreasing function h : [0, n] → [0, 21 ] such that h(0) = 0 and h(n) = 21 . With
h define the function ϕ : [0, 1] → [0, 1] by

h(x), if x ∈ [0, n]
ϕ(x) =
1 − h(N (x)), if x ∈ (n, 1].
This function ϕ is, obviously, continuous, strictly increasing,2 and verifies ϕ(0) =
h(0) = 0, ϕ(1) = 1 − h(N (1)) = 1 − h(0) = 1. Then
• If x ∈ [0, n), or N (x) ∈ (n, 1], ϕ(N (x)) = 1 − h(x) = 1 − ϕ(x), and N (x) =
ϕ−1 (1 − ϕ(x)).
• If x = n, N (n) = n = h −1 ( 21 ) = ϕ−1 ( 21 ), or N (n) = ϕ−1 (1 − ϕ(x)
• If x ∈ (n, 1], or N (x) ∈ [0, n), ϕ(N (x)) + ϕ(x) = h(N (x)) + 1 − h(N (x)) = 1
In conclusion, N (x) = ϕ−1 (1 − ϕ(x)), for all x in [0, 1], or N = Nϕ .
Notice that the proof of last theorem shows clearly that the order-automorphism
ϕ such that N = Nϕ is not unique. Notice also that with ϕ = id[0,1] it follows
N (x) = 1 − x, the fundamental strong negation, with which it results N = Nϕ =
ϕ−1 ◦ (1 − id[0,1] ) ◦ ϕ = ϕ−1 ◦ N ◦ ϕ, that is, all strong negations belong to the
family of N0 (x) = 1 − x. Nevertheless, in all cases it is n = ϕ−1 ( 21 ) the fixed point
of Nϕ . √
If ϕ(x) = x 2 , it results Nϕ (x) = 1 − x 2 , called the circular negation. If ϕ(x) =
−1 −1 −1 1−x
1+x , or ϕ (x) = 2−x , it follows Nϕ (x) = ϕ (1 − 1+x ) = ϕ ( 1+x ) = 1+3x ,
2x x 2x 1−x
that is the strong negation N3 of the before mentioned Sugeno’s negations.

With ϕ(x) = λ1 ln(1 + λx α ), λ > −1, α > 0, it follows the bi-parametric family
α 1
Nϕ (x) = ( 1+λx
1−x
α ) α , where with α = 1 it is obtained the Sugeno’s family of strong
negations Nλ = 1+λx
1−x
(λ > −1) that only depends on one single parameter.
Remark 2.2.31 The only linear strong negation N is N = N0 , since from N (a) =
αa + β, with N (0) = 1 = β and N (1) = 0 = α + 1, follows α = −1 and
N (a) = 1 − a.
Remark 2.2.32 The only “rational” strong negations N of the form N (x) = ax+b cx+d ,
a, b different of 0, are those Nλ (λ > −1) in the Sugeno’s family. It follows from:
• N (0) = 1 = db , or d = b
c+d , or a = −b
• N (1) = 0 = a+b
that gives
−bx + b b(1 − x) 1−x
N (x) = = = .
cx + b cx + b 1 + bc x
To have 0 N (x) 1, it should be 1 − x 1 + bc . But −1 = bc implies N (x) = 1,

that is not an strong negation. Hence it is −1 < bc , and with λ = bc , it follows
N (x) = 1+λx
1−x
= Nλ (x), with −1 < λ.
2 For (x < y) is evident that ϕ(x) < ϕ(y) if either x, y ∈ [0, n], or x, y ∈ (n, 1]. If x ∈ [0, n],
y ∈ (n, 1] and x < y, since h(x) + h(N (x)) < 1, it is h(x) < 1 − h(N (x)), or ϕ(x) < ϕ(y).
2.2.7 Continuous T-Norms and T-Conorms
As it was said, the only t-norm that is idempotent for all a in [0, 1], is T = min, and
the t-norms in { pr od} ∪ {W } only have the idempotents 0 and 1. As it was also said,
there are t-norms with several (but not all) idempotent elements. For example, the
function
1 1
+ W (3x − 1, 3y − 1), if (x, y) ∈ [ 13 , 23 ]2
T (x, y) = 3 3
min(x, y), otherwise ,
0.8
min
0.6
2/3
0.4 W
0.2 1/3
0
1
1/3 2/3
0.5
0.8 1
0 0.4 0.6
0 0.2
that as it is easy to prove is a t-norm, verifies

• T (x, x) = min(x, x) = x, if x ∈ / [ 13 , 23 ]
• T ( 13 , 13 ) = 13 + 13 W (0, 0) = 13 + 13 .0 = 1
3
• T ( 23 , 3)
2
= 1
3 + 3 W (1, 1) = 3 + 3
1 1 1
= 2
3
• T ( 21 , 2)
1
= 1
3 + 3 W ( 3 − 1, 3 − 1)
1 2 2
= 1
3 + 13 W ( 21 , 21 ) = 1
3 + 13 .0 = 1
3 = 1
2
• etc.
that is, all elements in [0, 1] − [ 13 , 23 ], as well 13 and 23 are idempotent for T , and the
elements in ( 13 , 23 ) are not idempotent.
Look that an analogous result is obtained when changing W by prod in the above
expression of T . Without proof it follows the theorem that completely characterizes
all continuous t-norms.
Theorem 2.2.33 T is a continuous t-norm if and only if,
1. T = min, T is in the family of min
2. T = pr odϕ , T is in the family of prod
3. T = Wϕ , T is in the family of W
4. There exist an index set (finite or countable infinite), a family of pairwise disjoint
open intervals in [0, 1], {(a1 , b1 ); i ∈ I }, and a family of t-norms Ti ∈ { pr od} ∪
{W }(i ∈ I ), such that

ai + (bi − ai )Ti ( bx−a i
, y−ai ),
i −ai bi −ai
if (x, y) ∈ [ai , bi ]2
T (x, y) =
min(x, y), otherwise,
for any x, y in [0, 1.

The continuous t-norm of the fourth type are called ordinal-sums of the continuous
t-norms Ti ∈ { pr od} ∪ {W }.
Remark 2.2.34 Why the names t-norm and t-conorm? The “t” comes from “triangu-
lar”, because these functions were introduced to formalize the triangular property of
probabilistic distances, i.e. distances whose values are something like the probability
that the numerical distance between two points is less than a given number. They
were introduced by Karl Menger with the name triangular-norms without considering
associativity.
The name of t-conorm refers to the duality with a t-norm, since S is a t-conorm
if and only if 1 − S(1 − x, 1 − y) is a t-norm. In general, it should be pointed out
that, for each strong negation N , S is a t-conorm if and only if N ◦ S ◦ (N × N ) is a
t-norm.
Theorem 2.2.35 S is a continuous t-conorm if and only if,
1. S = max, S is in the family of max
2. S = pr odϕ∗ , S is in the family of prod∗
3. S = Wϕ∗ , S is in the family of W ∗
4. There exist an index set (finite or countable infinite), a family of pairwise dis-
joint open intervals of [0, 1], {(a1 , b1 ); i ∈ I }, and a family of t-conorms
Si ∈ { pr od ∗ } ∪ {W ∗ }(i ∈ I ), such that

ai + (bi − ai )Si ( bx−a i
, y−ai ),
i −ai bi −ai
if (x, y) ∈ [ai , bi ]2
S(x, y) =
max(x, y), otherwise,
for any x, y in [0, 1].

Remark 2.2.36 Notice that with both ordinal-sums of t-norms and of t-conorms,
provided it is [0, 1] = [0, b1 ] ∪ [b1 , b2 ] · · · ∪ [bn−1 , bn ] ∪ [bn , 1], a finite partition
of the unit interval [0, 1], like, for example
the only idempotent elements are b1 , b2 , b3 , b4 , etc., as well as 0 and 1, that is, the
points giving the partition of [0, 1].
Remark 2.2.37 Although currently only continuous t-norms in {min}∪{ pr od}∪{W }

are taken into account in both theoretic fuzzy logic and its applications, it should
be pointed out that provided there are, at least, two statements ‘x is P’ and ‘x is
Q’ such that μ P (x), μ Q (x) ∈
/ {0, 1} and μ P and P = μ P , μ Q and Q = μ Q , the only
possibility for representing μ · σ = T ◦ (μ × σ), is by taking as continuous t-norm T
an ordinal-sum with the single interval (min(μ P (x), μ Q (x)), max(μ P (x), μ Q (x))).
Remark 2.2.38 Which t-norms are strictly non-decreasing in the sense that if 0 <
a < b < 1, then T (a, c) < T (b, c) for all c ∈ [0, 1]?
• If T = min, the answer is negative. For example, 0.3 < 0.5, but min(0.2, 0.3) =
min(0.2, 0.5) = 0.2
• If T = Wϕ , the answer is also negative. For example,0.3 < 0.5, but W (0.2, 0.3) =
W (0.2, 0.5) = 0
• If T = prodϕ , the answer is positive, since: a < b ⇒ ϕ(a) < ϕ(b) ⇒ ϕ(a) ·
ϕ(c) < ϕ(b) · ϕ(c) ⇒ ϕ−1 (ϕ(a) · ϕ(c))) < ϕ−1 (ϕ(b) · ϕ(c))), or pr odϕ (a, c) <
pr odϕ (b, c), because ϕ(c) ∈ (0, 1].
• If T is an ordinal-sum, it can’t be strictly non-decreasing because of the values it
takes with min.
Analogously, the only t-conorms that are strictly non-decreasing are those in
{ pr odϕ∗ }.
2.2.8 Laws of Fuzzy Sets
As it was said, in all standard algebras ([0, 1] X , T, S, N ) of fuzzy sets the triplet
(T, S, N ) share the following common properties:
1. T and S are commutative and associative
2. 1 is neutral for T , and 0 is neutral for S
3. 0 is absorbent for T , and 1 is absorbent for S
4. For all T and S, it is T min < max S
5. Each T (S) is non decreasing in the two variables
6. N is involutive, strictly decreasing and such that N (0) = 1,
a list of properties that gives some basic laws for fuzzy sets in the standard algebras,
like
• μ · σ = σ · μ, μ + σ = σ + μ,
• μ + (σ + λ) = (μ + σ) + λ = (σ + μ) + λ = λ + (σ + μ)
• μ · μ1 = μ, μ + μ0 = μ, μ + μ1 = μ1 , μ · μ0 = μ0
• If μ σ, then μ · λ σ · λ, and λ + μ σ + λ
• etc.
Anyway, a lot of laws typical of classical sets are not always valid in all standard
algebras of fuzzy sets. For example, (P(X ), ∩, ∪, c ) is a Boolean algebra and no
one ([0, 1] X , T, S, N ) is a Boolean algebra. In particular, (P(X ), ∩, ∪) is a lattice
and the only standard algebra that is a lattice is that with T = min and S = max.
Let us study in which standard algebras some laws of crisp sets do hold.
2.2.8.1 Distributive Laws
With classical sets it always do hold the two distributive laws

1. A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C)
2. A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C)
and the question is for which triplets (T, S, N ) do hold the corresponding laws with
fuzzy sets
1. μ · (σ + λ) = μ · σ + μ · λ,
2. μ + (σ · λ) = (μ + σ) · (μ + λ).
This questions correspond to solve the functional equations in the unknowns T and S:
T (a, S(b, c)) = S(T (a, b), T (a, c)) (2.3)
S(a, T (b, c)) = T (S(a, b), S(a, c)) (2.4)
for all a, b, c in [0, 1].
Lemma 2.2.39 Equation (2.3) does hold if and only if S = max.
Proof With b = c = 1, is T (a, S(1, 1)) = S(T (a, 1), T (a, 1)) or a = S(a, a). That
is S = max.
Provided S = max, the equation is T (a, max(b, c)) = max(T (a, b), T (a, c)). If
either b c or c b, it is immediate to check its validity for all t-norm T.
Lemma 2.2.40 Equation (2.4) does hold if and only if T = min.
Proof With b = c = 0, is S(a, 0) = a = T (a, a). That is T = min.

Provided T = min, the equation is S(a, min(b, c)) = min(S(a, b), S(a, c)). If
either b c or c b, it is immediate to check its validity for all t-conorm T.
Hence,
• In all standard algebras with (T, max), it holds μ · (σ + λ) = μ · σ + μ · λ,
• In all standard algebras with (min, S), it holds μ + (σ · λ) = (μ + σ) · (μ + λ)
• The two distributive laws (2.3) and (2.4) do jointly hold if and only if T = min
and S = max.
2.2.8.2 De Morgan Laws
With classical sets it always do hold the De Morgan, or duality, laws

1. A ∪ B = (Ac ∩ B c )c
2. A ∩ B = (Ac ∪ B c )c
showing that one of the two operators ∩, ∪ can be defined by the other and the
complementation. With fuzzy sets in standard algebras, these laws are
1. μ · σ = (μ + σ ) , or (μ · σ) = μ + σ
2. μ + σ = (μ · σ ) , or (μ + σ) = μ · σ ,
that correspond to the functional equations in the unknowns T , S and N
• T (a, b) = N (S(N (a), N (b)))
• S(a, b) = N (T (N (a), T (b)))
for all a, b in [0, 1].
Obviously, law (1) does hold if and only if T = N ◦ S ◦ (N × N ) and law (2)
does hold if and only if S = N ◦ T ◦ (N × N ), two formulas that are equivalent
since, for example, from the first it follows (with N 2 = id) N ◦ T = S ◦ (N × N )
or N ◦ S = T ◦ (N × N ), that is, the second formula.
Hence, the two De Morgan laws hold in an algebra given by the triplet (T, S, N )
if and only if T = N ◦ S ◦ (N × N ) , that is, T and S are N -dual.
2.2.8.3 Restricted Non-contradiction Principle µ · µ = µ0
With classical sets it always holds A ∩ Ac = ∅. With fuzzy sets, when is it
μ · μ = μ0 ?
The equation to be solved is
T (a, N (a)) = 0, for all a ∈ [0, 1]
in the unknowns T and N .

Theorem 2.2.41 If T is a continuous t-norm, and N is a strong negation, it is
T (a, N (a)) = 0 for all a ∈ [0, 1], if and only if T = Wϕ and N Nϕ .
Proof With the fixed point n ∈ (0, 1) of N , it follows T (n, n) = 0, that is, T has zero-
divisors. Hence, T = Wϕ , and Wϕ (a, N (a)) = ϕ−1 (max(0, ϕ(a)+ϕ(N (a))−1)) =
0, or max(0, ϕ(a) + ϕ(N (a)) − 1) = 0, or ϕ(a) + ϕ(N (a)) − 1 0, that implies
ϕ(N (a)) 1 − ϕ(a), or N (a) ϕ−1 (1 − ϕ(a)) = Nϕ (a), for all a ∈ [0, 1]. Hence
N Nϕ . The reciprocal is a simple calculation.
Then, the (restricted) non-contradiction principle μ · μ = μ0 holds if and only

if T = Wϕ and N Nϕ , for any order-automorphism ϕ and any t-conorm S. For
example, it holds (with ϕ = id) if T = W , S = max, N = N0 , and it does not hold
provided T = min, or T = pr odϕ .
2.2.8.4 Restricted Excluded-Middle Principle µ + µ = µ1
With classical sets it always holds A ∪ Ac = X . When is it μ + μ = μ1 for fuzzy

sets? When it does hold the equation S(a, N (a)) = 1 for all a ∈ [0, 1]?
Theorem 2.2.42 If S is a continuous t-conorm, and N is a strong negation, it is
S(a, N (a)) = 1 for all a ∈ [0, 1], if and only if S = Wψ∗ and Nψ N .
Proof With N (n) = n ∈ (0, 1), it follows S(n, n) = 1. That is S = Wψ∗ , and 1 =
Wψ∗ (a, n(a)) = ψ −1 (min(1, ψ(a) + ψ(N (a)))), or 1 = min(1, ψ(a) + ψ(N (a))).
Hence, 1 ψ(a)+ψ(N (a)), or Nψ (a) = ψ −1 (1−ψ(a) N (a)). That is, Nψ N .
The reciprocal is a simple calculation.
Then, the (restricted) excluded-middle principle μ + μ = μ1 holds if and only

if S = Wψ∗ and Nψ N , for any order automorphism ψ and any t-norm T . For
example, it holds (with ψ = id) if S = W ∗ , T = min, N = N0 , but it does not hold
provided S = max or S = pr od ∗ .
2.2.8.5 Both Restricted Principles of Non-contradiction

and Excluded-Middle
From last theorems it immediately follows that,

Theorem 2.2.43 In a standard algebra of fuzzy sets with a triplet (T, S, N ), it holds
μ · μ = μ0 and μ + μ = μ1 if and only if T = Wϕ , S = Wψ∗ , and Nψ N Nϕ .
In particular, they hold if T = W , S = W ∗ , and N = N0 , or with ϕ(x) = x 2 and

ψ(x) = x, they hold with the triplet:

T (x, y) = max(0, x 2 + y 2 − 1), S(x, y) = min(1, x + y),

1 − x N (x) 1 − x 2,
for all x, y in [0, 1].

Of course, with ϕ = ψ, the principles hold with Wϕ , Wϕ∗ , and N = Nϕ .
2.2.8.6 Laws of Absorption
With classical sets, the absorption laws A ∩ (A ∪ B) = A, and A ∪ (A ∩ B) = A,

always hold. With fuzzy sets in standard algebras, the formulas and respective equa-
tions
• μ · (μ + σ) = μ, T (a, S(a, b)) = a
• μ + (μ · σ) = μ, S(a, T (a, b)) = a
must be studied to find for which algebras these laws do hold.
Lemma 2.2.44 If T and S are, respectively, a t-norm and a t-conorm, it is

T (a, S(a, b)) = a for all a, b in [0, 1] if and only if T = min.
Proof If T = min, since a S(a, b), it follows min(a, S(a, b)) = a. With b = 0,
the equation gives T (a, a) = a, and = min.
Lemma 2.2.45 If T and S are, respectively, a t-norm and a continuous t-conorm, it

is S(a, T (a, b)) = a for all a, b in [0, 1] if and only if S = max.
Proof Since T (a, b) a, it follows max(a, T (a, b)) = a. With b = 1, the equation
gives S(a, a) = a, and = max.
Hence,
• The law μ · (μ + σ) = μ, holds for all S and T = min
• The law μ + (μ · σ) = μ, holds for all T and S = max
• The two laws hold jointly if and only if T = min and S = max.
2.2.8.7 The Law of von Neumann
With classical sets it always holds the law of von Neumann, or law of the perfect
repartition,
A = (A ∩ B) ∪ (A ∩ B c ),
that follows from A = A ∩ X = A ∩ (B ∪ B c ) = (A ∩ B) ∪ (A ∩ B c ), and generalizes

that of the excluded-middle since A = X implies X = (B ∩ X )∪(X ∩ B c ) = B ∪ B c .
From that law, by duality it follows Ac = (A ∩ B)c ∩ (A ∩ B c )c = (Ac ∪ B c ) ∩
(A ∪ B), that is
c
A = (A ∪ B) ∩ (A ∪ B c ),
a law that generalizes that of non-contradiction since A = ∅ implies ∅ = B ∩ B c .

With fuzzy sets, the question is the validity of the laws
μ = μ · σ + μ · σ , μ = (μ + σ) · (μ + σ ),
or of the functional equations
a = S(T (a, b), T (a, N (b))), a = T (S(a, b), S(a, N (b)))
Lemma 2.2.46 The equation a = S(T (a, b), T (a, N (b))) holds if and only if T =
pr odϕ , S = Wϕ∗ , N = Nϕ .
Proof If T = pr odϕ , S = Wϕ∗ , N = Nϕ , it is S(T (a, b), T (a, N (b))) =

Wϕ∗ ( pr odϕ (a, b), pr odϕ (a, Nϕ (b))) = ϕ−1 (W ∗ (ϕ( pr odϕ (a, b)), ϕ( pr odϕ (a, Nϕ
(b))))) = ϕ−1 (W ∗ (ϕ(a)·ϕ(b), ϕ(a)·ϕ(Nϕ (b)))) = ϕ−1 (min(1, ϕ(a)·ϕ(b)+ϕ(a)·
ϕ(Nϕ (b)))) = ϕ−1 (min(1, ϕ(a) · ϕ(b) + ϕ(a) · (1 − ϕ(b)) = ϕ−1 (min(1, ϕ(a) ·
ϕ(b) + ϕ(a) − ϕ(a) · ϕ(b))) = ϕ−1 (min(1, ϕ(a)) = a.
The proof of the reciprocal will be avoided since it is technically complex. Let us
say only that a = S(T (a, b), T (a, N (b))) gives, with a = 1, 1 = S(b, N (b)), that
implies S = W ∗ and Nϕ N .
It can be also proven that a = T (S(a, b), S(a, N (b))) if and only if T = Wϕ ,
S = pr odϕ∗ , N = Nϕ . Notice only that a = 0 gives T (b, N (b)) = 0, or T = Wϕ
and N = Nϕ .
Notice that the verification of von Neumann’s law require, in the case of fuzzy
sets, non-dual theories, like those given by the triplets ( pr odϕ , Wϕ∗ , Nϕ ), and
(Wϕ , pr odϕ∗ , Nϕ ).
2.2.8.8 Which Standard Algebra Is Closer to a Boolean Algebra?
The results in last section can be summarized in the following Table 2.1.
Hence, the algebras with the triplets (min, max, N ) are the ones that preserve more
structural Boolean properties. Indeed, these algebras preserve all the basic Boolean
laws except those of non-contradiction and excluded-middle. They are distributive
pseudo-complemented lattices, that is, De Morgan algebras that, in addition and like
all algebras of fuzzy sets, verify the law of Kleene,
T (a, N (a)) S(b, N (b)),
for all a, b in [0, 1]. The algebras given by the triplets (min, max, N ) are De Morgan-
Kleene algebras.
2.2.8.9 Last Comments
It can be considered, in addition to the structural Boolean laws, the cases that can be
derived from them, for example,
(A ∩ B c )c = B ∪ (Ac ∩ B c ),
Table 2.1 Basic boolean properties

T S N
Lattice Min Max All
Identity
T (a, 1) = a, T (a, 0) = 0 All – –
S(a, 0) = a, S(a, 1) = 1 – All –
Commutativity
T (a, b) = T (b, a) All – –
S(a, b) = S(b, a) – All –
Associativity
T (a, T (b, c)) = T (T (a, b), c) All – –
S(a, S(b, c)) = S(S(a, b), c) – All –
Involution
N (N (a)) = a – – All
B1. Idempotency
T (a, a) = a Min – –
S(a, a) = a – Max –
Distributivity
T (a, S(b, c)) = S(T (a, b), T (a, c)) All Max –
S(a, T (b, c)) = T (S(a, b), S(a, c)) Min All –
Absorption
T (a, S(a, b)) = a Min All –
S(a, T (a, b)) = a All Max –
Non-contradiction
T (a, N (a)) = 0 Wϕ – N Nϕ
Excluded-middle
S(a, N (a)) = 1 – Wϕ∗ N ≥ Nϕ
De Morgan’s laws
N (T (a, b)) = S(N (a), N (b)) T =N◦S◦N×N
N (S(a, b)) = T (N (a), N (b)) T =N◦S◦N×N
that follows from B ∪(Ac ∩ B c ) = (B ∪ Ac )∩(B ∪ B c ) = (B ∪ Ac )∩ X = B ∪ Ac =

(A ∩ B c )c . This law, in fuzzy set theory, is translated by
(μ · σ ) = σ + μ · σ
or
N (T (a, N (b))) = S(b, T (N (a), N (b))),
that holds with T = pr odϕ , S = Wϕ∗ , N = Nϕ .

Nevertheless, not all derived law has solution within the standard algebras of fuzzy
sets, as it is the case with (A ∪ A) ∩ (A ∩ Ac ) = ∅ (or A ∩ ∅ = ∅), since for
μ · μ + μ · μ = μ0 , []
there are no triplets (T, S, N ) for which it can hold S(T (a, a), T (a, N (a))) = 0 for
all a in [0, 1].
In the same vein, there are some laws that have solutions when different t-norms,
t-conorms and strong negations are considered. For example, (μ + μ) · (μ · μ ) = μ0 ,
that comes from (A ∪ A) ∩ (A ∩ Ac ) = ∅, translated in the form T1 (S(a, a),
T2 (a, N (a))) = 0, has infinite solutions like, for example, with an strong nega-
tion N , such that N N0 , T1 = min, T2 = W and any t-conorm S, since
min(S(a, a), T2 (a, N (a)) = T2 (a, N (a)) = W (a, N (a)) = max(0, a + N (a) −
1) = 0, because of T2 (a, N (a)) a S(a, a), and N (a) 1 − a, or
a + N (a) − 1 0.
Another case is given by the classical (derived) laws
A ∩ (Ac ∪ B) = A ∩ B, A ∪ (Ac ∩ B) = A ∪ B,
and the corresponding ‘possible’ fuzzy laws
μ · (μ + σ) = μ · σ, μ + (μ · σ) = μ + σ,
which functional equations
T1 (a, S(N (a), b)) = T2 (a, b), S1 (a, T (N (a), b)) = S2 (a, b),
do not have solutions with T1 = T2 and S1 = S2 , respectively, but that with N = N0 ,

W and W ∗ do verify
• W (a, W ∗ (1 − a, b)) = max(0, min(a, b)) = min(a, b)
• W ∗ (a, W (1 − a, b)) = min(1, max(a, b)) = max(a, b)
that is, they have the solutions (T1 = W, S = W ∗ , T2 = min) and (S1 = W ∗ , T =
W, S2 = max), respectively. Thus, it is possible to consider more complex algebras
of fuzzy sets by means of n-tuples of the type (T1 , . . . , Tm ; S1 , . . . , Sr ; N1 , . . . , N p ).
Notwithstanding, there are more derived laws than [] that have no solutions
neither in standard algebras, nor with different t-norms, t-conorms, or different strong
negations. The fact that no standard algebra of fuzzy sets is a Boolean algebra, makes
impossible to simultaneously deal in such algebras with all formulas that are valid
with classical sets.
2.2.9 Examples
Example 2.2.47 In a scale between 10 and 50 ◦ C, the label ‘cold’ referred to

temperature, is graduated by
⎧
⎨ 1, if 10 x 15
μC (x) = 10 ,
25−x
if 15 x 25
⎩
0, if 25 x 50.
Which one of the following linguistic labels: cold, hot, warm, more or less cold,
more or less warm, is the more adequate for the temperatures of 20, 21 and 22 ◦ C?
Solution. With C = cold, it is μC (20) = 10 5
= 0.5 and μmor e or less C (20) =
√
0.5 = 0.71. To obtain H = hot, we can compute μ H as the opposite of μC :
⎧
⎨ 1, if 45 x 50
μ H (x) = μC (50 + 10 − x) = μC (60 − x) = 10 ,
x−35
if 35 x 45
⎩
0, if 10 x 35
graphically
By defining, as it is usual, warm = not cold and not hot, that is
μw = μcold · μhot
with · represented by min, and by N0 ,
μw (x) = min(1 − μcold (x), 1 − μhot (x)),
for all x ∈ [10, 50], it results
Then:
• x = 20, gives μc (20) = 5
10 = 0.5, μ H (20) = 0, μW (20) = 5
10 = 0.5,
• x = 21, gives μc (21) = 4
10 = 0.4, μ H (21) = 0, μW (21) = 6
10 = 0.6,
• x = 22, gives μc (22) = 3
10 = 0.3, μ H (22) = 0, μW (22) = 7
10 = 0.7,
and
√
• μmor e or less cold (20) = √ μc (20) = 0.71,
μmor e or less war m (20) = μW (20) = 0.71
• μmor e or less ;cold (21) = 0.63, μmor e or less ;war m (21) = 0.77,
• μmor e or less ;cold (22) = 0.55, μmor e or less ;war m (22) = 0.84.
Hence
• The more adequate linguistic label for x = 20, cannot be decided but it could be
either ‘not cold’, or ‘not warm’. Since, it is not hot at all, we can take ‘not cold’.
• For x = 21, is ‘mol warm’ (mol = more or less)
• For x = 22, is ‘mol warm’
Example 2.2.48 On the age of a person p, it is known that
37 Age( p) 41,
and neither Age( p) 32, nor 43 Age( p). What can be said on the degree up to
which it could be Age( p) = 35, and Age( p) = 42?
Solution. What is unknown is the variation of Age( p) between 32 and 37, as well
as between 41 and 43. Since Age varies continuously, we can suppose there are two
functions
f : [32, 37] → [0, 1], g : [41, 43] → [0, 1]
such that f (32) = 0, f (37) = 1, g(41) = 1, g(43) = 0, with f strictly non-

decreasing, and g strictly decreasing. Then, we can define μ Age( p) : [0, 100] →
[0, 1], by
⎧
⎪
⎪ 0, if x ∈ [0, 32] ∪ [43, 100]
⎨
1, if x ∈ [37, 41]
μ Age( p) (x) =
⎪
⎪ f (x), if x ∈ [32, 37]
⎩
g(x), if x ∈ [41, 43]
and μ Age( p) (35) = f (35), μ Age( p) (42) = g(42). Graphically
f
g Age ( p )
32 37 41 43
To determine f and g more information is needed, but in the absence of it, we

5 , and g(x) = 2 , with which
can decide to take the linear models f (x) = x−32 43−x
3 1
μ Age( p) (35) = , μ Age( p) (42) = .
5 2
As it will be seen later on, 0.6 is the possibility that Age( p) = 35, and 0.5 that
of Age( p) = 42. Hence, it seems a little bit more possible that it be ‘Age( p) = 42’
than ‘ Age( p) = 35’.
Example 2.2.49 Knowing that Height(John) = 175 cm, and Height(Peter) = 180 cm,
consider the two statements:
p = It is false that John is not very tall or is more or less short
q = It is false that Peter is not very tall or is more or less short.
which is more true?
Solution. Both statements can be written by
Is false that x is P,
with P = ‘(not very tall) or (more or less short)’.
Hence

μ P (x) = S(μver y tall (x), μnot shor t (x)) = S(N (μtall (x)2 ), μtall (A(x)) ),
with a continuous t-conorm S, a strong negation N , and a symmetry A on X , provided

x varies in a scale of heights.
What should be compared are the two values N (μ P (175)) and N (μ P (180)), and
for that it is needed to know μtall . Let us take
⎧
⎨ 0, if x ∈ [0, 150]
μtall (x) = strictly non decreasing , if x ∈ [150, 190]
⎩
1, if x ∈ [190, 210]
with, perhaps, μtall (x) = 0.025x −3.75, x ∈ [150, 190], if we need to have numbers.
Hence, with A(x) = 210 − x, it is A(175) = 210 − 175 = 35, and μtall (35) = 0,
as well as A(180) = 220 − 180 = 30, and μtall (30) = 0, because of that
μ P (175) = S(N (μtall (175)), 0) = N (μtall (175)2 )
μ P (180) = S(N (μtall (180)), 0) = N (μtall (180)2 ).
Since μtall is strictly non-decreasing between 150 and 190, it is μtall (175) <
μtall (180), and N (μtall (180)2 ) < N (μtall (175)2 ). Finally,
N (N (μtall (175)2 )) < N (N (μtall (180)2 )), or μtall (175)2 < μtall (180)2 ,
and q is strictly more true than p.

only a form for μtall , as well as

Notice that it is not needed to fix S and N , but √
to accept that μver y P (x) = μ P (x)2 , μmol P (x) = μ P (x), and A(x) = 210 − x.
This last hypotheses is perfectly reasonable since μtall is non-decreasing, and then
the order μtall is just the order of [0, 210].
Provided we need to know up to which numerical degree p and q do hold, we can
use the linear function in the figure,
and fix N = N0 . It results

degree up to which p is true = 1 − (1 − μ2tall (175)) = μ2tall (175) = 0.391
degree up to which q is true = μ2tall (180) = 0.563,
that shows how is q more true than p.
Example 2.2.50 It is known that the algebra with which fuzzy sets must be combined
should satisfy the laws
μ + μ · σ = μ, μ · (μ + σ) = μ,
as well as that the negation is linear, Determine the triplet (T, S, N ) and, with X =
[0, 10], and ⎧
⎨ 0, if 0 x 5
x
μ(x) = , σ(x) = x−5 , if 5 x 8
10 ⎩ x−6
6
4 , if 8 x 10,
compute μ · σ, μ + σ, and μ + σ.
Solution. The first law of absorption μ + μ · σ = μ, implies S = max for
any T . The second law of absorption μ · (μ + σ) = μ, implies T = min for
any S. Hence (T, S) = (min, max), and the only linear N is N = N0 . Hence,
(T, S, N ) = (min, max, 1 − id). With the graphics of μ and σ in the figure,
μ′ μ
7 80:9 :
it follows μ · σ = min(μ, σ) = σ, μ+σ = max(μ, σ) = μ, and μ +σ = (1−μ)+σ

is the pointed curve. That is,

μ (x), if x ∈ [0, 6.87]
(μ + σ)(x) =
σ(x), if x ∈ [6.87, 10].
Notice that x = 6.87 comes from the equation 1 − x

10 = x−5
6 .
Example 2.2.51 Consider the Sugeno’s family of strong negations Nλ (x) =

1+λx (λ > −1). If −1 < λ1 < λ2 , it follows 1 + λ1 x < 1 + λ2 x, or 1+λ2 x < 1+λ1 x ,
1−x 1 1
that is, Nλ2 (x) < Nλ1 (x). Hence,

If λ 0 : N0 Nλ
If 0 λ : Nλ N0
Nλ ,λ ∈(−1,0)
N0
Nλ ,λ ∈(0, + ∞)
Compare the graphics of μ = N0 ◦ μ and μ = N1 ◦ μ, in a figure, if

⎧
⎪
⎪ 0, if x ∈ [0, 3] ∪ [7, 10]
⎨
1, if x ∈ [4, 6]
μ(x) =
⎪
⎪ x − 3, if x ∈ [3, 4]
⎩
7 − x, if x ∈ [6, 7]
Solution. It is ⎧
⎪ 1, if x ∈ [0, 3] ∪ [0, 4]
⎪
⎪
⎨ 0, if x ∈ [4, 6]
μ (x) = 4−x
⎪
⎪ x , if x ∈ [3, 4]
⎪
⎩ x−6
x , if x ∈ [6, 7],
and μ1 (x) = N1 (μ(x)) = 1−μ(x)

1+μ(x) , or
⎧
⎪
⎪ 0, if x ∈ [0, 3] ∪ [7, 10]
⎪
⎪
⎨ 1, if x ∈ [4, 5]
μ1 (x) =
⎪
⎪ x−2 ,
4−x
if x ∈ [3, 4]
⎪
⎪
⎩ x−6 , if x ∈ [6, 7]
8−x
Hence,
Look that N1 (3.2) = 0. 6, N1 (3.5) = 0.

3, N1 (3.4) = 0.43, N1 (3.6) = 0.25,

N1 (3.8) = 0.1, but N0 (3.2) = 0.8, N0 (3.5) = 0.5, N0 (3.6) = 0.4, N0 (3.4) =
0.6, , N0 (3.6) = 0.4, N0 (3.8) = 0.2.
Example 2.2.52 The negation is linear, and the standard algebra must verify the
law μ = μ · σ + μ · σ . Determine the triplet (T, S, N ), and with μ(x) = 10
x
(in
X = [0, 10]) and ⎧
⎪
⎨ 1, if x ∈ [0, 5]
σ(x) = 2 ,
7−x
if x ∈ [5, 7]
⎪
⎩
0, if x ∈ [7, 10]
check that μ · σ + μ · σ = μ, μ · σ + μ · σ = σ, and (μ · σ ) = σ + μ · σ .

Solution. From N linear, N = N0 , it follows that we can take ϕ = id, and
from μ · σ + μ · σ = μ it follows T = prodϕ , S = Wϕ∗ , and N = Nϕ . Hence
(T, S, N ) = (prod, W ∗ , N ).
⎧
⎨0

Since μ (x) = 1 − 10 , and σ (x) =
x x−5
, it follows
⎩ 2
1
⎧ x ⎧ ⎧
⎨ 1 − 10
x
⎪
⎨ 10 ⎪
⎨0 ⎪
μσ(x) = x(7−x)
, μσ (x) = x(x−5)
, μ σ(x) = 2 (1 − 10 )
7−x x
.
⎪
⎩
20 ⎪
⎩
20 ⎪
⎩
x 0
0 10
Hence,
⎧ ∗ x ⎫
⎨ W ( 10 , 0) = 10
x
⎪ ⎪
⎬ x
(μ · σ + μ · σ )(x) = W ∗ ( x(7−x) x(x−5)
20 , 20 ) =
x = = μ(x),
⎪
⎩ ∗
10 ⎪
⎭ 10
W (0, 10x
) = 10
x
⎧ ∗ x ⎫
⎨ W ( 10 , 1 − 10 ) = 1
x
⎪ ⎪
⎬
(σ · μ + σ · μ )(x) = W ∗ ( x(7−x)
20 , 2 (1 −
7−x
10 )
x
= 7−x = σ(x).
⎪
⎩ ∗
2 ⎪
⎭
W (0, 0) = 0
Finally, since,
⎧ ⎧
⎨1 ⎪
⎨0
(μ · σ ) (x) = 1 − (μ · σ )(x) = 1 − x(x−5) , and (μ · σ )(x) = x ) x−5 ,
(1 − 10
⎩ x
20 ⎪
⎩
2
1 − 10 1− x 10
it results
⎧ ∗
⎨ W (1, 0) = 1
⎪

(σ + (μ · σ ))(x) = W ∗ ( 7−x
2 , 2 (1 − 10 ) = 1 −
x−5 x x(x−5)
= (μ · σ ) (x) .
⎪
⎩ ∗
20
W (0, 1 − 10
x
) = 1 − 10
x
Example 2.2.53 Predicate F = high fever refers to the interval [37, 42] in a clini-
cal thermometer, in which the values {37, 37.5, 38, . . . , 41.5, 42} are significative.
Asking an expert one obtains the following fuzzy set
μ F = 0.3/38.5 + 0.5/39 + 0.7/39.5 + 0.8/40

+ 0.9/40.5 + 1/41 + 1/41.5 + 1/42
where it is clear that 0/37 + 0/37.5 + 0/38 is avoided since this values of the body’s
temperature are not significative for F. With all that, give the membership function
of P = very high fever, Q = more or less high fever, R = low fever, S = not high
fever.
√
Solution. With the usual definition μver y F = μ2F , μmol F = + μ F , μlow F =
μ F (37 + 42 − x) = μ F (79 − x), μnot F = 1 − μ F , it results:
• μ P = 0.09/38.5 + 0.25/39 + 0.49/39.5 + 0.64/40 + 0.81/40.5 + 1/41 +
1/41.5 + 1/42.
• μ Q = 0.55/38.5 + 0.7/39 + 0.84/39.5 + 0.89/40 + 0.95/40.5 + 1/41 +
1/41.5 + 1/42.
• μ R = 1/37 + 1/37.5 + 1/38 + 0.9/38.5 + 0.8/39.5 + 0.7/39.5 +
0.5/40.5 + 0.3/40.5.
• μ S = 1/37 + 1/37.5 + 1/38 + 0.7/38.5 + 0.5/39.5 + 0.3/39.5 +
0.2/40.5 + 0.1/40.5.
Notice the incoherence produced by μ S = μnot F μlow F = μ R . An incoherence
showing that it cannot be taken the representation μnot F = 1 − μ F , but some
μnot F = N ◦ μ F with N ≥ N0 .
For example, if N (x) = 1−0.9x
1−x
, it is
μ S = 1/37 + 1/37.5 + 1/38 + +0.96/38.5 + 0.91/39.5 + 0.81/39.5 +
0.71/40 + 0.53/40.5 + 1/41 + 1/41.5 + 1/42, showing μlow F μnot F .
Look that
μ F & low F = 0.3/38.5 + 0.5/39.5 + 0.7/39.5 + 0.5/40 + 0.3/40.5
provided μ F & low F = min(μ F , μlow F ).
Example 2.2.54 Describe in fuzzy terms, the statements

p = John is young and around forty,
q = John is old or around forty.
Solution. The solution will come after representing the predicates P = young,
aP = Old, A40 = around forty, in a scale of 0–100 years. The general forms of μ P ,
μa P , and μ A40 , are
with μa P (x) = μ P (100 − x) since μ P is non-decreasing. Once these functions were

established accordingly with the current use of P and A40, it will be:
Degree( p) = T (μ P (x), μ A40 (x)), Degree(q) = S(μa P (x), μ A40 (x)),
with convenient continuous t-norm T and t-conorm S. This formulas are the descrip-
tion of p and q in fuzzy terms.
For example, if a = 20, b = 50, μ P = 50−x 30 , if 20 x 50, and 40 − a =
30, 40 + b = 50, with μ A40 piece-wise linear, T = min, S = max, with
⎧
⎪
⎨ 0, if 0 x 50
μa P (x) = μ P (100 − x) = 30 ,
x−50
if 50 x 80
⎪
⎩
1, if 80 x 100,
the graphics is
The slashed function describes p, and the continuous one describes q. Of course,
Degree( p) min(μ P (x), μ A40 (x)) μ P (35) = 0.5.
Remark 2.2.55 To select T and S, the following points could be taken into account,
• It could be perfectly the case that ‘John is young and not young’ with a positive
degree. Hence, the laws of Non-contradiction can be avoided, and T ∈ / {W }.
• Since it is reasonable to accept that ‘John is old’ and ‘John is old’ does coincide
with ‘John is old’, and ‘John is young’ or ‘John is young’ does coincide with
‘John is young’ , we can decide to take either T = min, S = max, or T and S as
ordinal-sums.
• Provided idempotency is avoided i.e., ‘John is young’ and ‘John is young’ does
coincide with ‘John is very young’, instead of T = min, we can take T = prod,
that is more interactively than min. In this case, because it does not seem that
duality should be avoided, we could take √
S = Prod∗ , and
√
then
∗
μver y young or mol old (45) =Prod ( 36 , 6 ) = 6 (1 + 36 ) = 0.5636,
1 6 1 35 6
√
that is greater than the value 6
6
= 0.408 obtained with = max.
Example 2.2.56 In X = [0, 10] the predicate P = big is represented by μ(x) = 10 x

.
In which points in [0, 10] is the degree of ‘big’ less than that of ‘not big’?
Solution. Given μ, the problem is to find for which x ∈ X it is μ(x) μ (x) =
Nϕ (μ(x)) = ϕ−1 (1 − ϕ(μ(x))), that is, μ(x) ϕ−1 ( 21 ). Then,
• If N = N0 , x
10 1− x
10 , or x 5.
x
1− 10 √
• If N = x
N1 , 10 x
1+ 10
, or x 2 + 20x − 100 0, that means x 10( 2 − 1) =
4.142
x
1− 10 √
• If N = N2 , x
, or x 2 + 10x − 50 0, that means x 75 − 5 = 3.66
10 1+ 2x
10
Hence,
• If N = N0 , the set is [0, 5], and the threshold (of selfcontradiction) of big is 5.
• If N = N1 , the set is [0, 4.142], and the threshold is 4.142
• If N = N2 , the set is [0, 3.66], and the threshold is 3.66
Notice that changing big by not big, the thresholds do remain but the sets are, respec-
tively, [5, 10], [4.142, 10], and [3.66, 10].
2.3 On Aggregating Imprecise Information
The kind of problems this section will deal with are like the following. An exam is
corrected by three referees R1 , R2 , R3 , each
3one with a different weight of strongness
W (Ri ) ∈ [0, 1], 1 i 3, such that i=1 W (Ri ) = 1. Each referee assigns a
numerical qualification pi ∈ [0, 10] to the exam delivered by a given student. How
these qualification can be “aggregated” to obtain final qualification for the student’s
exam? A recognized usual way of doing it is by the weighted mean:
1 p1 p2 p3
Q= · W (R1 ) + · W (R2 ) + · W (R3 ),
10 10 10 10
pi
with 10 ∈ [0, 10]. For example, if W = (0.5, 0.3, 0.2) and P = (7, 6, 5), it follows
1
Q = 0.7 × 0.5 + 0.6 × 0.3 + 0.2 × 0.5 = 0.63
10
that implies Q = 6.3. Provided the three referees have the same weight, it is W =
( 13 , 13 , 13 ), and then, Q = p1 · 13 + p2 · 13 + p3 · 13 = p1 + p32 + p3 = 7+6+5
3 = 6, is
just the arithmetic mean of the three qualifications.
Another way of obtaining the final qualification, this time by ignoring the referee’s
character, is by the geometric mean
√ √
Q= p1 · p2 · p3 = 7 × 6 × 5 = 5.94,
3 3
showing
√ that in a problem with p1 = p2 = 10, p3 = 0, it results Q =
3
10 × 10 × 0 = 0, when the arithmetic mean is 20
3 = 6.67.
2.3.1 Aggregation Functions
Most of these problems are “represented” by the so-called Aggregation Functions,

that is, functions
A : [0, 1]n → [0, 1],
such that
1. A is continuous in all variables
2. A(0, . . . , 0) = 0, and A(1, . . . , 1) = 1
3. If x1 y1 , . . . , xn yn , then A(x1 , . . . , xn ) A(y1 , . . . , yn ).
Sometimes it is said that A is an n-dimensional aggregation function. Continuous
t-norms and continuous t-conorms are 2-dimensional aggregation functions.
Of the many types of aggregation functions, a particular and important type are
the quasi-linear means,
n

M(x1 , . . . , xn ) = f −1 pi . f (xi )
i=1

n
with ( p1 , . . . , pn ) in [0, 1], verifying pi = 1, and f : [0, 1] → R, continuous,
i=1
one-to-one, and monotonic. Function f is called the generator of M.
Notice that if f is the identity f (x) = x, we get the weighted means:

n
M(x1 , . . . , xn ) = pi · xi ,
i=1
that with p1 = 1
n (1 i n) is the arithmetic mean
2.3 On Aggregating Imprecise Information 79
1
n
M(x1 , . . . , xn ) = xi ,
n
i=1
and with f (x) = − log x, and p1 = 1

n (1 i n) is the geometric mean
√
M(x1 , . . . , xn ) = n
p1 · p2 · · · pn .
1
With f (x) = x α (α > 0), is f −1 (x) = x α , and with p1 = 1
n, it is obtained the
family of quasi-linear means,
1
x1α + · · · + xnα α
Mα (x1 , . . . , xn ) = .
n
In particular, with α = 1, is M1 the arithmetic mean, and with α = −1, it follows

n
M−1 (x1 , . . . , xn ) = (provided x1 , . . . , xn = 0),
1
x1 + ··· + 1
xn
called Harmonic Mean. As it is easy to prove,

√
lim Mα (x1 , . . . , xn ) = n
x1 · · · xn
α→0
lim Mα (x1 , . . . , xn ) = max(x1 , . . . , xn )

α→∞
lim Mα (x1 , . . . , xn ) = min(x1 , . . . , xn ).

α→−∞
2.3.2 Ordered Weighted Means
It is said that M : [0, 1]n → [0, 1] is a mean, when M is continuous, monotonic,

and verifies:
min(x1 , . . . , xn ) M(x1 , . . . , xn ) max(x1 , . . . , xn ).
Since, min(0, . . . , 0) M(0, . . . , 0) max(0, . . . , 0) = 0, it results M(0, . . . ,

0) = 0. Since, min(1, . . . , 1) M(1, . . . , 1) max(1, . . . , 1) = 1, it results
M(1, . . . , 1) = 1. Hence, of course, quasi-linear means are means, but there are
more of such means. An important and useful example are the Ordered Weighted
Means (OWA). Its definition is the following:
O : [0, 1]n → [0, 1] is an OWA, if O(x1 , . . . , xn ) is obtained under the process,

n
• Select weights p1 , . . . , pn in [0, 1], such that pi = 1.
i=1
• Permute the n-pla (x1 , . . . , xn ), to the n-pla (x1∗ , . . . , xn∗ ) such that x1∗ · · · xn∗ )

n
• O(x1 , . . . , xn ) = pi · xi∗ .
i=1
For example, if n = 2,
O(x1 , x2 ) = p1 · min(x1 , x2 ) + p2 · max(x1 , x2 ), with p1 + p2 = 1.
If n = 4, and the weights are (0.2, 0.4, 0.3, 0.1), it is
O(0.2, 0.5, 0.7, 0.3) = O(0.2, 0.3, 0.5, 0.7)

= 0.2 × 0.2 + 0.4 × 0.3 + 0.3 × 0.5 + 0.1 × 0.7 = 0.74.
2.3.3 More on Aggregations
Because they are associative, continuous t-norms and continuous t-conorms can be
extended to n-dimensional aggregation functions. For example, with n = 3,
T (x1 , x2 , x3 ) = T (x1 , T (x2 , x3 )) = T (T (x1 , x2 ), x3 )) = · · ·
S(x1 , x2 , x3 ) = S(x1 , S(x2 , x3 )) = S(S(x1 , x2 ), x3 )) = · · ·
Nevertheless, not all aggregation functions are associative. For example, if M

is the arithmetic mean, M(x1 , M(x2 , x3 )) = 2x1 +x42 +x3 , but M(M(x1 , x2 ), x3 )) =
x1 +x2 +2x3
4 . Concerning means, the only associative are min, and max.
In general, Aggregation Functions are not commutative. For example, a
2-dimensional quasi-linear mean
M(x1 , x2 ) = f −1 ( p1 f (x1 ) + p2 f (x2 )), p1 + p2 = 1,
is commutative if and only if p1 = p2 = 21 . Arithmetic and geometric means are

commutative, but weighted means in general are not.
If T is a continuous t-norm, and S a continuous t-conorm, the function
A(x1 , x2 ) = p1 T (x1 , x2 ) + p2 S(x1 , x2 ), p1 + p2 = 1
is an aggregation function that, since T min max S, in general is not a

mean. The only exception is with T = min, and S = max, as it was said before. For
example,
2.3 On Aggregating Imprecise Information 81
• A(x1 , x2 ) = 0.7x1 .x2 + 0.3W ∗ (x1 , x2 )

• A(x1 , x2 ) = 0.6 min(x1 , x2 ) + 0.4(x1 + x2 − x1 .x2 )
• A(x1 , x2 ) = 0.6W (x1 , x2 ) + 0.4 max(x1 , x2 ),
are aggregation functions.
2.3.4 Examples
The pointwise aggregation of classical sets is not, in general, a classical set, but a
fuzzy one. For example, the arithmetic mean verifies
1
M(0, 0) = 0, M(0, 1) = M(1, 0) = , M(1, 1) = 1
2
and, if A, B are crisp subsets, M(A, B) is not a crisp subset if given by M(μ A , μ B )
(x) = M(μ A (x), μ B (x)). On the contrary, with the geometric mean G, it is
G(0, 0) = G(0, 1) = G(1, 0) = 0, G(1, 1) = 1,
and G(A, B) is a crisp set.

In all cases, if μ ∈ [0, 1] X , σ ∈ [0, 1]Y , and A is an aggregation function, then
A(μ, σ)(x, y) = A(μ(x), σ(y)),
for all x ∈ X, y ∈ Y , is a fuzzy set A(μ, σ) ∈ [0, 1] X ×Y called the aggregation of μ

and σ. When X = Y it could be defined the fuzzy set A(μ, σ) ∈ [0, 1] X ,
A(μ, σ)(x) = A(μ(x), σ(x)), for all x ∈ X .
Example 2.3.1 If X = {1, 2, 3, 4, 5}, and μ = 0.6/1 + 0.7/2 + 0.5/3 + 1/4, σ =

0.9/1 + 0.5/3 + 0.7/4 + 0.8/5, compute M(μ, σ), G(μ, σ), and O(μ, σ) with O
the OWA with weights p1 = 0.4, p2 = 0.6.
Solution.
M(μ, σ) = 0.75/1 + 0.35/2 + 0.5/3 + 0.85/4 + 0.4/5
G(μ, σ) = 0.735/1 + 0/2 + 0.5/3 + 0.837/4 + 0/5
O(μ, σ) = (0.4 × 0.6 + 0.6 × 0.9)/1 + (0.4 × 0 + 0.6 × 0.7)/2 + (0.4 ×
0.5 + 0.6 × 0.5)/3 + (0.4 × 0.7 + 0.6 × 1)/4 + (0.4 × 0 + 0.6 × 0.8)/5 =
0.72/1 + 0.42/2 + 0.5/3 + 0.88/4 + 0.48/5.
Notice that G(μ, σ) M(μ, σ), but that neither G(μ, σ) and O(μ, σ), nor M(μ, σ)
and O(μ, σ), are order-comparable.
Example 2.3.2 A linguistic variable has the fuzzy values H = high, S = short and
M = medium, with H represented by
⎧
⎨ 0, if 0 x 5
μ H (x) = 1, if 7 x 10
⎩ x−5
2 , if 5 x 7.
In the two suppositions, M = H · S , and that M is the aggregation of H and S

under the weighted mean A(x1 , x2 ) = 0.3x1 + 0.7x2 , compute μ M .
Solution. Since S is an antonym of H , and μ H is monotonic, μ S (x) = μ H (10 − x),
that is ⎧
⎨ 1, if 0 x 3
μ S (x) = 0, if 5 x 10
⎩ 5−x
2 , if 3 x 5.
The solution for the first supposition appears in the following sequence of figures
(with · = min, = 1 − id).
The solution for the second supposition is

⎧
⎪
⎪ 0.7, if 0 x 3
⎪
⎨ 0.7(5−x) , if 3 x 5
μ H (x) = μ A(H,S) (x) = 2
⎪
⎪
0.3(x−5)
, if 5 x 7
⎪
⎩ 2
0.3, if 7 x 10,
graphically,
0.7
M
0.3
3 5 7
Chapter 3
Reasoning and Fuzzy Logic
3.1 What Does It Mean “Logic”?
The question is, in fact, a philosophical one whose discussion does not correspond
to this text, and that received a lot of comments and discussions by philosophers.
Instead of such question, there is the more particular,what is a logic?, that can be
answered not philosophicaly but in terms of the mathematical definition of what is a
consequence’s operator. A definition that corresponds to an abstraction of the term
“deduction”.
3.1.1 Logic and Consequence Operators
A logic is a triplet (X, A, C), where A ⊂ P(X ) is a family of parts of X , and

C : A → A is a mapping verifying
1. For all P ∈ A, P ⊂ C(P).
2. If P, Q ∈ A, and P ⊂ Q, then C(P) ⊂ C(Q).
3. For all P ∈ A, C(C(P)) ⊂ C(P).
It can be said that the pair (A, C) defines a logic in the set X .
From 1. and 2., it follows C(P) ⊂ C(C(P)), and from 3. results

3 . For all P ∈ A, C(C(P)) = C(P), or C 2 = C.
Hence, a consequence’s operator is one that is extensive (1.), monotonic (2.), and a
closure (3 .). These operators were introduced by the logician Alfred Tarski in the
thirties of 20th century, and are called to be compact provided for all P ∈ A, it exists
a finite set { p1 , . . . , pn } ⊂ P such that
4. C(P) = C{ p1 , . . . , pn }.

84 3 Reasoning and Fuzzy Logic
It is clear that if P is finite, C is compact. This happens, for example if X is finite.

A consequence’s operator C is consistent when ‘q ∈ C(P) ⇒ q ∈ / C(P)’.
Example 3.1.1 In a finite lattice (X, ·, +; 0, 1), let’s consider A = P0 (X ) = {P =
{ p1 , . . . , pn } ⊂ X, p∧ = p1 · · · pn = 0}, and the operator
Cons : P0 (X ) → P0 (X ), defined by Cons(P) = {q ∈ X ; p∧ q},
with the partial order given by x y ⇔ x · y = x. Cons is a consequence’s

operator:
1. Since p∧ pi , it is pi ∈ Cons( p), for all pi ∈ P. That is, P ⊂ Cons(P).
2. If P = { p1 , . . . , pn } ⊂ { p1 , . . . , pn , pn+1 , . . . , pm } = Q, since q∧ = p1 · · · pn ·
pn+1 · · · pm p1 · · · pn = p∧ , if q ∈ Cons(P), from p∧ q, it follows
q∧ q, and q ∈ Cons(Q). Hence, P ⊂ Q implies Cons(P) ⊂ Cons(Q).
3. Obviously, Min Cons(P) = p∧ , hence, Cons(P) ∈ P0 (X ) since Min Cons
(P) = 0. Then, if q ∈ Cons(Cons(P)), or Min Cons(P) q, or p∧ q, it
follows q ∈ Cons(P). Hence, Cons(Cons(P)) ⊂ Cons(P).
In any finite lattice (X, ·, +; 0, 1), it can be considered the logic (X, P0 (X ), Cons),
and it can be proved that if the lattice is endowed with a complement such that
(X, ·, +, ; 0, 1) is a Boolean algebra, any operator of consequences C : P0 (X ) →
P0 (X ) verifies C ⊂ Cons, that is C(P) ⊂ Cons(P), for all P ∈ P0 (X ). Cons is
the biggest operator of consequences in a Boolean algebra with A = P0 (X ).
The set of premises P is consistent, if p∧ is not self contradictory ( p∧ p∧ ). If

it were p∧ p∧ , it will be also p∧ ∈ Cons(P), that is absurd if Cons is consistent.
In this cases, q ∈ Cons(P) and q ∈ Cons(P), or p∧ q and p∧ q (or q p∧ )

implies p∧ p∧ , that is absurd. Thus, q ∈ Cons(P) does imply q ∈ / Cons(P),
and the operator Cons is consistent.
Remark 3.1.2 Instead of a lattice, let us take the set [0, 1] X endowed with a fuzzy
intersection
μ∧ = μ1 · · · μn = T ◦ (μ1 × · · · × μn )
(T a continuous t-norm), the partial order μ σ ⇔ μ(x) σ (x), for all x ∈ X , and
the empty set μ0 = μ∅ . Take the set P0 ([0, 1] X ) that consists of the finite subsets
P = {μ1 , . . . , μn } ⊂ [0, 1] X such that μ∧ = μ0 . The definition
Cons(P) = {σ ∈ [0, 1] X ; μ∧ σ },
allows the same result as in the case before.
There is a, perhaps alternative, way of constructing a logic in a set X . It follows

from the following results.
• If (X, A, C) is a logic in X , the binary relation ‘x C y ⇔ y ∈ C({x})’, defined
if {x} ∈ A, is a preorder.
3.1 What Does It Mean “Logic”? 85
Proof Since {x} ⊂ C({x}), it is x C x, for all {x} ∈ A. If x C y and y C z,

it is y ∈ C({x}), and z ∈ C({y}), hence from {y} ⊂ C({x}) follows C({y}) ⊂
C(C({x})) = {x}, and z ∈ C({x}), or x C z.
Notice that the preorder C is defined only with the pairs (x, y) ∈ X × X such
that {x} ∈ A and {y} ∈ A.
• Given a set X and A ⊂ P(X ), if is a preorder in X such that
If x ∈ P, and x y, it is y ∈ Q, for some Q ∈ A,
the operator C : A → A, defined by C(P) = {y ∈ X ; ∃x ∈ P & x y},
verifies,
1. ∀x ∈ P, it is x x, then x ∈ C(P) : P ⊂ C(P)
2. If P ⊂ Q, and y ∈ C(P), there is x ∈ P such that x y. Since x ∈ Q, it is
y ∈ C(Q). Hence, C(P) ⊂ C(Q).
3. If y ∈ C(C(P)), there is x ∈ C(P), such that x y. But there is also z ∈ P
such that z x. Hence z y, and y ∈ C(P). That is, C(C(P)) ⊂ C(P).
4. The preorder C coincides with the initial , since x C y ⇔ y ∈
C({x}) ⇔ x y.
Example 3.1.3 Let it be the set X = {1, 2, 3, 4, 5} endowed with the preorder
1 2 3 4 5
It is, for example C({1, 5}) = {1, 5, 2, 3, 4}, C({4, 5}) = {4, 5}, C({2, 3, 4}) =
{2, 3, 4, 5}, C({1, 2, 4, 5}) = {1, 2, 4, 5, 3} = X , C({5}) = {5}, and C({3}) =
{3, 4, 5}.
Hence, it is possible to identify a logic in a set with a preordering of it.
3.1.2 Conjecturing
A process to pass from a set of premises P = {μ1 , . . . , μn } to a ‘conclusion’ σ

is a conclusive reasoning, that is sometimes symbolized by P σ . A conclusive
reasoning P σ is deductive if there exists either an operator of consequences C,
or a preorder , such that P σ is equivalent to σ ∈ C(P), or to μi σ for all
μi ∈ P.
Anyway, not all conclusive reasonings are deductive. In common reasoning if
there is the possibility of stating P σ and defining C(P) = {σ ; P σ }, the axiom
of monotonicity is not always verified by such C, but it could verify
• If P ⊂ Q, then C(Q) ⊂ C(P), C is anti-monotonic.

• If P ⊂ Q, then C(P) and C(Q) are not comparable, that is, it is neither C(P) ⊂
C(Q), nor C(Q) ⊂ C(P), C is non-monotonic.
Most common conclusive reasonings are not deductive, but of a conjectural type, that
is, the conclusion σ is provisionally accepted because, simply, it is non contradictory
with all or with part of the information. It should be noticed that the lost of the
monotonicity of C motivates the lost of the transitive property of the preorder C .
We will say that P σ is a conjectural kind of conclusive reasoning if there exists
an operator of consequences C, such that
P σ ⇔ σ ∈
/ C(P) ⇔ N ◦ σ ∈
/ C(P),
for some strong negation N . Analogously, if what we have is a preorder (instead

of C), σ is a conjecture of the information contained in P, when
P σ ⇔ ∀μ ∈ P : μ σ .
In what follows, we will only take into account the operator Cons(P) = {σ ; μ∧
σ }, provided μ∧ = μ0 . Consequently, the set of the conjectures that, through
Cons, is associated to any finite set P = {μ1 , . . . , μn } of premises such that
μ∧ = μ1 · · · μn = T ◦ (μ1 × . . . × μn ) = μ0 , for some continuous t-norm T , is
Con j (P) = {σ ∈ [0, 1] X ; μ∧ σ }c ,
with σ = N ◦ σ , for some strong negation N .

Notice that it is necessary to take a pair of connectives (T, N ) to define Con j (P).
Remark 3.1.4 Instead of μ∧ = μ0 , in what follows we will suppose that μ∧ is not

self-contradictory, that is μ∧ μ∧ . With this hypothesis, if σ ∈ Cons(P) ⇔ μ∧
σ , it can’t be μ∧ σ because of σ μ∧ implies μ∧ μ∧ . Hence, it should
be μ∧ σ or σ ∈ Con j (P). That is, Cons(P) ⊂ Con j (P). In addition; Cons is
consistent. Notice that μ0 ∈ / Con j (P).
In fact, μ∧ μ∧ is more general than μ∧ = μ0 , since μ∧ = μ0 implies
μ∧ = μ0 μ1 = μ0 = μ∧ . Hence, if it is μ∧ μ∧ , it is also μ∧ = μ0 .
With this change,
Con j (P) = Cons(P) ∪ H yp(P) ∪ Sp(P),
where
• H yp(P) = {σ ∈ Con j (P); μ0 < σ < μ∧ }, is the set of hypotheses of P.
• Sp(P) = {σ ∈ Con j (P); μ∧ N C σ }, is the set of speculations of P, where NC
means non-comparable under ,
and verifying
3.1 What Does It Mean “Logic”? 87
Cons(P) ∩ H yp(P) = Cons(P) ∩ Sp(P) = H yp(P) ∩ Sp(P) = ∅.
Notice that
• Consequences follow (in the partial order ) from μ∧ : all the premises explain
the consequences.
• Hypotheses explain all the premises, since σ < μ∧ μi , all the premises follow
from each hypothesis.
• Speculations, are the conjectures for which it is neither μ∧ σ , nor σ < μ∧ .
With all that, processes,
• P σ , with σ ∈ Con j (P), is a guessing, or conjectural reasoning
• P σ , with σ ∈ Cons(P), is a deduction, or deductive reasoning
• P σ , with σ ∈ H yp(P), is an abduction, or abductive reasoning
• P σ , with σ ∈ Sp(P), is an speculation, or speculative reasoning
Concerning this types of conclusive reasonings, it should be pointed out what follows.
1. The set Con j (P) is not always in A = P0 (X ). Hence, it can’t be taken as a set
of premises, it has not sense to consider Con j (Con j (P)), or Cons(Con j (P)).
2. If P ⊂ Q, it is Con j (Q) ⊂ Con j (P), because if σ ∈ Con j (Q), from I n f Q
I n f P, if I n f Q σ , it is I n f P σ . Hence, if σ ∈ Con j (Q) it is σ ∈
Con j (P), and Con j is anti-monotonic.
3. The sets H yp(P) and Sp(P) are not always in P0 (X ). Hence, they can’t be taken
as sets of premises.
4. If P ⊂ Q, it is H yp(Q) ⊂ H yp(P), since σ ∈ H yp(Q), or μ0 < σ < I n f Q
implies μ0 < σ < I n f P. Hence, the operator H yp is anti-monotonic.
5. Concerning the operator Sp, if P ⊂ Q it can be Sp(P) Sp(Q), and Sp(Q)
Sp(P). Hence, Sp is a non-monotonic operator.
3.2 Reasoning with Conditionals: Representation
3.2.1 What is a Conditional?
A conditional is an statement of the form ‘If a, then b’ := a → b, with two previous

statements a, b. For example, “If it is raining, then I take an umbrella”, where a =
It is raining, b = I take an umbrella, or “If the food is well cooked and well served,
and the wine is of good quality, then the tip will be higher than usual”, wit a = The
food is well cooked and well served, and the wine is of good quality, b = The tip
will be higher than usual.
Notice that the first example is a crisp conditional, but the second is an impre-
cise one. In what follows we will take into account the representation of imprecise
conditionals of the type
If x is P, then y is Q,
where x ∈ X , y ∈ Y , P is a predicate (precise or imprecise) in X , and Q is a predicate
(precise or imprecise) in Y . For example, with X = [0, 1], Y = [0, 10],
If x is small, then y is big,
or,
If x is small and y is big, then z is not small,
with x, y ∈ [0, 1] and z ∈ [0, 10].
What it means to represent a conditional statement like “If x is P, then y is Q”?
It means to translate it in fuzzy terms. For example, ‘x is P’ and ‘y is Q’ will be
translated by μ P (x) and μ Q (y) with adequate fuzzy sets μ P , μ Q ∈ [0, 1] X , adequate
in the sense that they capture the use of P on X and Q on Y .
But how to represent the full statement “If x is P, then y is Q”:= μ P (x) →
μ Q (y)? It is always done, in fuzzy logic, by means of a function J : [0, 1]×[0, 1] →
[0, 1], such that
μ P (x) → μ Q (y) = J (μ P (x), μ Q (y)) ∈ [0, 1]
for all x ∈ X , and y ∈ Y . But, which function J can be taken? It depends on the
‘meaning’ of the conditional statement, and this requires to look at what happens in
general.
Remark 3.2.1 Imprecise conditionals are very useful in ordinary life, for instance,
• If the turn is not so far, and the car’s speed is not high, then press softly the break,
• If the the food was well-cooked and of quality, and the service was good, the the
tip should be higher than 15 %.
Both in common life and in technology, a lot of imprecise conditionals (rules) are
considered. The need for its representation will be obvious in the next section.
3.2.2 The Case of Boolean Algebras
Let us consider, in the first place, the case in which the statements are crisp and
belong to a Boolean algebra (B, ·, +, ; 0, 1), where ·, +, stand, respectively, for
the intersection (and), the union (or), and the complement (negation, not), 0 is the
minimum element, and 1 is the maximum. As it is well known, this Boolean algebra
is naturally ordered by means of the partial order
a b ⇔ a · b = a ⇔ a + b = b ⇔ b = a + a · b.
Representing ‘If a, then b’ by a → b, what is important is that from the set of

premises P = {a, a → b} should follow the statement b as a logical consequence.
3.2 Reasoning with Conditionals: Representation 89
For this goal, it should be a · (a → b) = 0, and a · (a → b) b. In a boolean

algebra it is
a · z b ⇔ z a + b,
because of:
1. a · z b ⇒ a + a · z = a + z a + b, and since z a + z, follows z a + b
2. z a + b ⇒ a · z a · (a + b) = a · b b.
Then, a · (a → b) b ⇔ a → b a + b.
A conditional function is a mapping →: B × B → B, such that a · (a → b) b
for all a, b ∈ B. Hence, in a Boolean algebra, the biggest conditional is a + b, the
so-called material conditional, and any smaller function is also a conditional. For
example, from
a · b b a + b, a a + b
it follows that a → b = a · b, a → b = b, a → b = a , are conditionals.

Analogously, from
a · b + a · b a + b,
it also follows that a → b = a · b + a · b is a conditional. Different ways of writing

a + b in a boolean algebra, are
a · (b + b ) + a · b, a + a · b, b + a · b
since a · (b + b ) + a · b = a + a · b = (a + a ) · (a + b) = a + b, and
b + a · b = (b + a ) · (b + b ) = a + b.
Notice that from z 1 a + b, z 2 a + b, follows z 1 · z 2 (a + b) · (a + b) =
a + b, z 1 + z 2 (a + b) + (a + b) = a + b, hence, the union and the intersection

of conditionals is also a conditional. For example, a · b + a + b = a + b is obviously

a conditional.
There are two-variable functions a → b such that a → b a + b, but are not
expressible as a single formula with the connectives , ·, + as the before considered
cases. For example,
a · b, if a · b = 0
a→b=
a + b, if a · b = 0,

a · b, if a · b = 0
verifies (a → b) · a = = a · b b, that is, is
a · (a + b) = a · b, if a · b = 0,
a conditional. Analogously,

1, if a b
a→b=
b, otherwise,

a, if a b
verifies (a → b) · a = b, that is, is a conditional.
a · b, otherwise ,
Remark 3.2.2 If the Boolean algebra B is complete, that is, for any A ⊂ B, A = ∅,
it exists Sup A ∈ B, then
Sup{z ∈ B; a · z b} = Sup{z ∈ B; z a + b} = a + b.
Remark 3.2.3 The character of conditional of a +b is exclusive of Boolean algebras.

That is, in any ortholattice, the validity of a · (a + b) b, for all a, b, forces the
ortholattice to be a Boolean algebra.
Remark 3.2.4 a → b a +b, is a property that only holds in Boolean algebras, that
is, in ortholattices the equivalence a · z b ⇔ z a + b, is not valid. It only holds
in Boolean algebras. For example, in orthomodular lattices, both a →1 b = a +a ·b,
and a →2 b = b + a · b (that verify a →2 b = b →1 a ), are conditionals, but is
neither a →1 b a →2 b nor a →2 b a →1 b.
The conditional a →1 b = a + a · b is called the Sasaki hook, and a →2 b =
b + a · b is the Dishkant hook, and, of course, only in Boolean algebras are both

coincidental with a + b. The Sasaki and the Dishkant hooks are used as models for
the conditional statements in the reasoning in Quantum Logic.
Remark 3.2.5 The scheme of Modus Ponens
If a, then b
a
b,
corresponds to forwards reasoning, that is, goes from the antecedent a to the con-
sequent b thanks to the conditional a → b, through a · (a → b) b. Backwards
reasoning goes from the consequent to the antecedent (also thanks to a → b), ad it
is modeled by the Modus Tollens scheme.
If a, then b
not b
not a,
that is translated by b · (a → b) a ⇔ a → b (b ) + a = a + b. Thus,

in Boolean algebras, a → b = a + b, also allows backwards reasoning, provided
b · (a → b) = b · (a + b) = a · b = 0, or a + b = 1. Nevertheless, although
b · (a · b) = 0 a , it is clear that the conjunctive conditional a → b = a · b does
not allow backwards reasoning since b · (a → b) = 0.
3.2.3 Fuzzy Conditionals
Let us return to the case of fuzzy logic, that is, to a conditional linguistic expression,
or rule, like ‘If x is P, then y is Q’, represented in fuzzy terms by
(μ P → μ Q )(x, y) = J (μ P (x), μ Q (y)),
for all x ∈ X , y ∈ Y . The problem is which function J : [0, 1] × [0, 1] → [0, 1],
is to be taken at each case if, of course, it gives a conditional. That is, if from the
premises {‘x is P’, ‘If x is P, then y is Q’} follows ‘y is Q’ as a logical consequence.
Formally speaking, it should exist a continous t-norm T0 such that
T0 (μ P (x), J (μ P (x), μ Q (y))) μ Q (y)
for all x ∈ X , y ∈ Y . This condition, that should hold for any μ P (x) ∈ [0, 1], and
any μ Q (y) ∈ [0, 1], conducts to the inequality
T0 (a, J (a, b)) b
for all a, b in [0, 1], J should verify to represent conditional statements. It is called
the Modus Ponens Inequality, since it allows the scheme of reasoning
If x is P, then y is Q
x is P
y is Q
called the scheme of Modus Ponens. J is called a T0 -conditional function (shortly,

T0 -conditional).
There is a theorem, whose proof will be omitted, showing that being T0 a contin-
uous t-norm, it is
T0 (a, J (a, b)) b ⇔ J (a, b) JT0 (a, b) = sup{z ∈ [0, 1]; T0 (z, a) b}.
Hence, for each T0 , the greatest T0 -conditional is the function JT0 , since, it verifies
the MP-inequality T0 (a, JT0 (a, b)) = min(a, b) b.
Remark 3.2.6 For reasons that will be latter on presented, T -conditionals JT are
called R-implications (R shorting residuated). They come directly from the Boolean
equation a + b = sup{z; a · z b}.
If μ, σ ∈ {0, 1} X , take (μ → σ )(x, y) = JT (μ(x), σ (y)) = sup{z ∈

[0, 1]; T (z, μ(x)) σ (y)}. If μ(x) ∈ {0, 1}, σ (y) ∈ {0, 1}, it is
⎧
⎪ JT (0, 0) = 1
⎪
⎨
JT (0, 1) = 1
JT (μ(x), σ (y)) =
⎪
⎪ JT (1, 0) = 0
⎩
JT (1, 1) = 1,
that coincides with the values of max(1 − μ(x), σ (y)). That is, all R-implications
do coincide with the Boolean material conditional μ + σ in the case that μ and σ
are crisp sets. R-implications generalize the material conditional. Of course, this will
happen with any J such that
J (0, 0) = 1, J (0, 1) = 1, J (1, 0) = 0, J (1, 1) = 1
Example 3.2.7 The immediate generalization of the Boolean conditional a → b =

a + b, is given by (μ + σ )(x, y) = μ (x) + σ (y), that is, by J (a, b) = S(N (a), b),
for all a, b in [0, 1]. These operators are called S-implications (S shortens ‘strong’).
With,
• S = max, N = N0 , is J (a, b) = max(1 − a, b), called the Kleene-Diennes
conditional.
• S = pr od ∗ , N = N0 , is J (a, b) = 1−a + ab, called the Reichenbach conditional.
• S = W ∗ , N = N0 , is J (a, b) = min(1, 1 − a + b), called the Łukasiewicz
conditional.
Notice that the MP inequality T0 (S(N (a), b)) a, is verified on the last three cases
with T0 = W :
• W (a, max(1 − a, b)) = max(0, a + b − 1) = W (a, b) b
• W (a, 1 − a + a · b) = max(0, a · b) = a · b b
• W (a, min(1, 1 − a + b)) = max(0, min(a, b)) = min(a, b) b,
hence, the three cases are W-conditionals.
Example 3.2.8 Let us see how is JT , when T is, respectively, the continuous t-norm
min, pr odϕ , Wϕ .

1, if a b
• T = min, Jmin (a, b) = sup{z ∈ [0, 1]; min(z, a) b} = (G ödel
b, if a > b
implication).
T = pr odϕ , JT (a, b) = sup{z ∈ [0, 1]; ϕ(a).ϕ(z) ϕ(b)} =
•
1, if a b
ϕ(b) (Goguen implication).
ϕ −1 ( ϕ(a) ), if a > b
• T = Wϕ , JT (a, b) = sup{z ∈ [0, 1]; ϕ −1 (W (ϕ(a), ϕ(z)) b} = ϕ −1 (min(1, 1−
ϕ(a) + ϕ(b))) (Łukasiewicz implication).
Since each JT is a T -conditional, G ödel’s is a min-conditional, Goguen’s are prodϕ -
conditionals, and Łukasiewicz’s are Wϕ -conditionals. Notice that the S-implications
of the form
Wϕ∗ (Nϕ (a), b) = ϕ −1 (W ∗ (ϕ(Nϕ (a)), ϕ(b))) = ϕ −1 (W ∗ (1 − ϕ(a), ϕ(b))) =
ϕ −1 (min(1, 1 − ϕ(a) + ϕ(b))),
are exactly the Łukasiewicz’s R-implications: the only R-implications that are
S-implications are the Łukasiewicz’s ones. If, for instance, it were

1, if a b
Jmin (a, b) = S(N (a), b) =
b, if a > b
it will result
1, if N (a) b
S(a, b) =
b, if N (a) > b
a function that is not a t-conorm, since S(a, 0) = 0 = a, if a > 0. Hence, Jmin is not
an S-implication. An analogous reasoning shows that J pr odϕ are not S-implications.
Example 3.2.9 The protoform μ → σ = μ + μ · σ (coming from the Sasaki hook),

gives
J1 (a, b) = S(N (a), T (a, b)),
with S a continuous t-conorm, T a continous t-norm, an N an strong negation. These

function are called Q-conditionals (Q for Quantum). For example,
• S = max, T = min, N = N0 , is J1 (a, b) = max(1−a, min(a, b)), is the so-called
Early-Zadeh operator
• S = max, T = prod, N = N0 , is J1 (a, b) = max(1 − a, ab)
• S = prod ∗ , T = min, N = N0 , is J1 (a, b) = 1 − a + a 2 b
• S = W ∗ , T = W , N = N0 , is J1 (a, b) = max(1 − a, b), that coincides with the
Kleene-Diennes implication
• S = W ∗ , T = prod, N = N0 , is J1 (a, b) = 1 − a + ab, that coincides with the
Reichenbach implication
• S = W ∗ , T = min, N = N0 , is J1 (a, b) = min(1, 1 − a + b), that coincides with
the Łukasiewicz implication
With which t-norm T0 do verify the MP inequality these Q-operators? For instance,
• W (a, max(1 − a, min(a, b)) = max(0, a + min(a, b) − 1) = W (a, min(a, b))
min(a, b) b
• W (a, max(1 − a, a · b) = max(0, a + a · b) − 1) = W (a, a · b) a · b b
• W (a, 1 − a + a 2 b) = max(0, a 2 , b) = a 2 b b
• W (a, max(1 − a, b)) b (as it is proven before)
• W (a, 1 − a + ab) b (as it is proven before)
Example 3.2.10 The protoform μ → σ = σ + μ · σ (coming from the Dishkant

hook), gives the D-operators:
J2 (a, b) = S(b, T (N (a), N (b))),
with which J2 (N (b), N (a)) = S(N (a), T (a, b)) = J1 (a, b) or, equivalently,
J2 (a, b) = J1 (N (b), N (a)): D-operators are the contrasymmetricals of Q-operators.
Hence, it can be repeated all that has been said for J1 . For example,
If S = max, T = min, N = N0 , it is J2 (a, b) = J1 (1 − a, 1 − b) =
max(b, min(1 − b, 1 − a)) = max(b, 1 − max(a, b)), that verifies
W (a, max(b, 1 − max(a, b))) = max(0, max(a

+ b − 1, a − max(a, b))) =
W (a, b), if b a or b > a and b > 21
b. It is a W-conditional.
0, if b > a and b 21
Example 3.2.11 The protoform μ → σ = μ · σ (coming from the classical con-

junctive conditional a → b = a · b), gives
J (a, b) = T (a, b),
functions with the inconvenience of the property J (a, b) = J (b, a), but verifying,
T0 (a, J (a, b)) = T0 (a, T (a, b)) T (a, b) min(a, b) b
that is, all of them are conditionals for any t-norm T0 and in particular, for the greatest
of them. They are always taken as min-conditionals. For example,
• If T = min, J (a, b) = min(a, b), is called the Mamdani conditional
• If T = prod ϕ , J (a, b) = ϕ −1 (ϕ(a) · ϕ(b)), are called Larsen conditionals
• T = Wϕ is never used, since it can be J (a, b) = 0 with a > 0 and b > 0.
For example, with ϕ(x) = x(1+x)
(an order automorphism), it is ϕ −1 (x) =
√ 2
8x+1−1
2 , and
a(1 + a) b(1 + b) ab(1 + a)(1 + b)

J (a, b) = ϕ −1 (ϕ(a) · ϕ(b)) = ϕ −1 ( . ) = ϕ −1 ( )
2 2 4
√
ab(1 + a)(1 + b) + 1 − 1
=
2
that, of course, is a min-conditional.
Remark 3.2.12 A way of avoiding the undesirable symmetry J (a, b) = J (b, a) in

the case of Mamdani-Larsen min-conditionals, is taking
J (a, b) = T (a r , bs ),
with real numbers r, s with 1 > s. Then T0 (a, T (a r , bs )) min(min(a r , bs ), bs )

bs b, and J is a min-conditional.
Example 3.2.13 Once given a conditional statement ‘If x is P, then y is Q’, and
represented ‘x is P’ by μ P (x), and ‘y is Q’ by μ Q (y), it remains to be understood
what it is meant by the ‘statement’ μ P (x) → μ Q (y). It could be, or not to be,
μ P (x) → μ Q (y) = (μ P → μ Q )(x, y), with μ P → μ Q a fuzzy set in X × Y ,
identified with some expression involving the connectives and (·), or (+), not ( ).
In the affirmative case, it is said that μ P → μ Q is expressible in material form, for
example, μ P → μ Q = μP +μ Q , or μ P → μ Q = μP +μ P ·μ Q , etc. These material
forms are called protoforms; for instance, the protoform of the Kleene-Diennes fuzzy
conditional is a + b, and that of Mamdani fuzzy conditional is a · b.
If μ P → μ Q does not correspond with a protoform, one can try to represent it
by means of an R-implication, that is, by Jmin or by some J pr odϕ , since all the JWϕ
do correspond to a protoform a + b, or μP + μ Q , with + represented by Wϕ∗ and
by Nϕ .
In addition, there is a problem that should be taken into account when representing
μ P → μ Q . The problem is the following. Suppose that we know μ P → μ Q
should be represented by a function J that is a min-conditional, but that we are
not able to decide a protoform and we take Jmin . Since J Jmin , we will reach
the biggest possible output. This should be known. Analogously, if J should be a
prod-conditional, from J J pr od , follows the same comment.
Example 3.2.14 Let’s stop for a while in the above mentioned concept of implication
function, a concept that comes directly from the properties shown by the Boolean
conditional a → b = a + b, whose truth value is usually represented by
v(a → b) = max(1 − v(a), v(b)).
Look that, if b1 b2 , it follows a + b1 a + b2 , or a → b1 a → b2 , if a1 a2 ,

is a2 a1 , and a2 → b a1 → b. Since v(a) and v(b) only take the values {0, 1},
v(a → b) shows the truth-table.
v(a) v(b) v(a → b)

0 0 1
0 1 1
1 0 0
1 1 1
Because of that, a function J : [0, 1]×[0, 1] → [0, 1] is called a fuzzy implication

function provided:
1. J is decreasing in its first variable, and non-decreasing in its second one
2. J (0, 0) = J (0, 1) = J (1, 1) = 1, J (1, 0) = 0
Obviously, S-implications and R-implications are fuzzy implication functions, but
Q and D operators are not always so, since, for instance
JS (0.4, 0.3) = max(1 − 0.4, min(0.4, 0.3)) = 0.6
J D (0.3, 0.1) = max(0.1, min(0.7, 0.9)) = 0.7
J D (0.3, 0.4) = max(0.4, min(0.7, 0.6)) = 0.6,
that is, for instance 0.1 < 0.4 but J D (0.3, 0.1) > J D (0.3, 0.4).
Analogously, J (a, b) = T (a, b) are not implication functions, since J (0, 0) =

T (0, 0) = 0, and a1 a2 ⇒ J (a1 , b) J (a2 , b). Also Sasaki operators are
not, for example, 0.4 < 0.8, but JS (0.4, 0.9) = max(1 − 0.4, min(0.4, 0.9)) =
max(0.6, 0.4) = 0.6 < 0.8 = JS (0.8, 0.9), that is, JS is non-decreasing in the
second variable.
To represent conditional statements as they appear in language, the concept of
implication function is sometimes excessive. What is needed are just T-conditionals,
except in the cases where more properties are necessary to represent the meaning of
the conditional statements.
Remark 3.2.15 There is a question that is not independent of the representation J

for the conditional statement. The question is: What are we going to do with J ?
Which is the purpose for using J ?
We are to make an inference that, in principle, could be forwards
{μ, μ → σ } σ,
or backwards,
{σ , μ → σ } μ .
The first type of inference corresponds to search for the solutions of μ·(μ → σ ) σ ,
that is,for J and T0 such that,
T0 (μ(x), J (μ(x), σ (y)) σ (y); ∀x, y ∈ X. (∗ )
The second type corresponds to search for the solutions of σ · (μ → σ ) μ , that

is, for J , T1 and N such that,
T1 (N (σ (y)), J (μ(x), σ (y)) N (μ(x)); ∀x, y ∈ X. (∗∗ )
Hence, given J , we need to know T0 such that T0 (a, J (a, b)) b, for forward
inference, and given J and N , we need to know T1 such that T1 (N (b), J (a, b))
N (a), for backwards inference.
Notice that the two t-norms in (∗ ) and (∗∗ ) are not necessarily coincidental. For
example, given J (a, b) = max(1 − a, b), with N = N0 , can we do backwards
inference? To answer this question we just need to know if there is a continuous
t-norm T1 such that T1 (1 − b, max(0, max(1 − a, b)) 1 − a, with a = 1 it results
T1 (1 − b, b) = 0, and T1 = W . Then, since,
W (1 − b, max(1 − a, b)) = max(0,
1 − b + max(1 − a, b)− 1)
1 − a − b, if a + b 1
= max(0, max(1 − a − b, 0)) = 1 − a, because
0, if a + b > 1
b 0 implies −b 0, and 1 − a − b 1 − a. Finally, the answer is: Yes, with
T1 = W . It can also be done backwards inference in the following cases,

1 − b, if a b
1. With Jmin , since W (1−b, Jmin (a, b)) = 1−a.
W (1 − b, b) = 0, if a > b

1 − b, if a b
2. With J pr od , since W (1 − b, J pr od (a, b)) = 1 − a.
a , if a > b
b 1−b
3. With JW , since W (1 − b, min(1, 1 − a + b)) = min(1 − b, 1 − a) 1 − a.
4. With J (a, b) = 1 − a + ab, since W (1 − b, min(1, 1 − a + ab)) =
(1 − b)(1 − a) 1 − a.
5. With J (a, b) = max(1−a, min(a, b)), since W (1−b, max(1−a, min(a, b)))
W (1 − b, max(1 − a, b)) 1 − a.
6. With J (a, b) = pr od ∗ (1 − a, a · b) = 1 − a + a 2 b, since W (1 − b, pr od ∗ (1 − a,
a · b)) W (1 − b, pr od ∗ (1 − a, b)) = W (1 − b, 1 − a + a · b) 1 − a.
Nevertheless, the case J (a, b) = T (a, b) is, actually, negative. To have T1 (1 −

b, T (a, b)) 1−a, it is necessary (take a = 1) that T1 (1−b, b) = 0, that is, T1 = W ,
but it is W (1 − b, T (a, b)) = max(0, T (a, b) − b) = 0, since T (a, b) b implies
T (a, b) − b 0. Hence, by one side from W (1 − b, T (a, b)) = 0 1 − a, it seems
that backwards inference is possible. But, given the scheme: “σ , μ → σ : μ ”, what
results is σ · (μ → σ ) = μ0 , that forces Con j ({σ , μ → σ }) = ∅. In conclusion,
Mamdani-Larsen conditionals don’t allow backwards inference.
Remark 3.2.16 It should be pointed out that except Jmin , J pr odϕ , and J (a, b) =
T (a, b), most of the functions J are T0 -conditionals for T0 = W , and almost all
do also verify backwards inference also with T0 = W . And the t-norms in the
Łukasiewicz’s family show the disturbing problem of having zero-divisors!
Remark 3.2.17 The name R-implication, or residuated implication, comes from the
idea of ‘residuum’ that clearly appear in the case of J pr od when
b
If a > b, then J pr od (a, b) = .
a
Remark 3.2.18 In the same vein under which it was proven that R-implications J
with T = Wϕ are not S-implications, it is easy to show that they are not expressible
in material protoform, that is, by an expression with logical connectives. Take the
perhaps more general material protoform μ · (σ + σ ) + μ · σ . Is it possible that
JT0 (a, b) = S1 (T1 (N1 (a), S2 (b, N2 (b)), T2 (a, b))),
for T0 = Wϕ , S1 and S2 continuous t-conorms, T1 , T2 continuous t-norms, and

N1 , N2 strong negations? With b = 0, it follows
JT0 (a, 0) = sup{z ∈ [0, 1]; T (z, a) 0} = 0
S1 (T1 (N1 (a), S2 (0, 1)), 0) = T1 (N1 (a), 1) = N1 (a),
or, S1 (N1 (a), 0) = 0. This means that S1 is not a t-conorm. Hence, the decision of
representing an R-implication can’t be taken from a material protoform interpretation
of it.
3.3 Short Note on Other Modes of Reasoning
The mode of reasoning given by the scheme
{μ, μ → σ } σ
is classically called Modus Ponendo Ponens (from the Latin, mode of starting the
truth (of σ ) by placing the truth (of μ)), or, for short Modus Ponens. That given by
the scheme {σ , μ → σ } μ is classically called Modus Tollendo Tollens (from
the Latin, mode of stating the falsity (of μ) by placing the falsity (of σ )), or, for
short Modus Tollens. They correspond to what we called forwards and backwards
reasoning. But there are again other modes of reasoning that can be considered, for
example,
• Modus Tollendo Ponens, given by the scheme {μ , μ+σ } σ and also called Mode
of Disjunctive Reasoning, and classically proven by μ ·(μ+σ ) = μ·μ +μ ·σ =
μ · σ μ (in a Boolean algebra).
• Modus Ponendo Tollens, given by the scheme {μ, (μ·σ ) } σ , classically proven
by μ · (μ · σ ) = μ · (μ + σ ) = μ · σ σ (in a Boolean algebra).
• Constructive Dilemma, given by the scheme {μ + λ, μ → σ, λ → η} σ + η,
classically proven by (μ + λ) · (μ → σ ) · (λ → η) = (μ + λ) · (μ + σ ) · (λ + η) =
μσ λ +ηλμ +ημσ +ηλσ σ +η (in a Boolean algebra where a → b = a +b).
• Destructive Dilemma, given by the scheme {μ + σ , λ → μ, η → σ, } λ + σ ,
classically proven by (μ + σ )(λ + μ)(η + σ ) = λ (μ · η + μ σ ) + η (σ · λ +
μ · σ ) λ + η (in a Boolean algebra where a → b = a + b).
What in the fuzzy case? For example, in the case of the Disjunctive Mode we
need to find all the possibilities for μ · (μ + σ ) σ , that is, to solve the functional
equation
T (N (a), S(a, b)) b
for all a, b in [0, 1], in the three variables T, S, N . With b = 0, it follows

T (N (a), a) = 0, or T = Wϕ , N Nϕ . Taking N = Nϕ it results Wϕ (Nϕ (a), S
(a, b)) = ϕ −1 (max(0, 1 − ϕ(a) + ϕ(S(a, b)) − 1))
= ϕ −1 (max(0, +ϕ(S(a, b)) − ϕ(a))) b, implying ϕ(S(a, b)) − ϕ(a) ϕ(b)
or ϕ(S(a, b)) ϕ(a) + ϕ(b). Hence, S(a, b) ϕ −1 (min(1, ϕ(a) + ϕ(b))) =
W ∗ (a, b). For example, it can be taken S = Wϕ∗ or S = max.
• Wϕ (Nϕ (a), max(a, b)) = ϕ −1 (max(0, ϕ(b) − ϕ(a))) ϕ −1 (ϕ(b)) = b
• Wϕ (Nϕ (a), Wϕ∗ (a, b)) = ϕ −1 (min(1 − ϕ(a), ϕ(b))) ϕ −1 (ϕ(b)) = b
Hence, the disjunctive mode can be used in, for example, the cases
(Wϕ , Nϕ , max) and (Wϕ , Nϕ , Wϕ∗ ).
3.3 Short Note on Other Modes of Reasoning 99
Remark 3.3.1 It should be pointed out that the Modus Ponendo Tollens (MPT) can
be reduced, in the case of duality, to the disjunctive mode by means of the change
μ = α , σ = β , in which case since (μ · σ ) = μ + σ it follows μ · (μ · σ ) =
μ · (μ + σ ) = α · (α + β) β = σ . Hence, it holds with the triplet (Wϕ , Nϕ , Wϕ∗ ).
3.4 Inference with Fuzzy Rules
A central topic fuzzy logic deals with are non-rigid, dynamic systems involving
‘variables’ x1 , . . . , xn , y taking values in, respectively, universes X 1 , . . . , X n , Y ,
and constrained by imprecise rules ri of the type
If x1 is P1i , and x2 is P2i ,…, and xn is Pni , then y is Q i (1 i n),
with predicates P ji (1 j n) in X j , and Q i in Y .
Let’s consider the simplest case with two variables x ∈ X, y ∈ Y , constrained by

a single rule If x is P, then y is Q.
When observing the system (x, y), the variable x in the rule’s antecedent not
always will show ‘x is P’, but ‘x is P ∗ ’ with P ∗ some predicate slightly modificate
from P. For example, if P = shor t, it could be P ∗ = ver y shor t, almost shor t,
etc. A concrete example is
Rule: If tomatoes are red, they are ripe
Observation: Tomatoes are very red
Conclusion: Tomatoes are very ripe,
where P = r ed, Q = ri pe, P ∗ = ver y r ed, and Q ∗ = ver y ri pe. Hence, the
corresponding scheme of forwards reasoning is
Rule: If x is P, then y is Q
Observation: x is P ∗
Consequence: y is Q ∗
where P, Q, P ∗ are known, and Q ∗ is unknown. This scheme is the Generalized
Modus Ponens (GMP), and to find Q ∗ through fuzzy artillery it is needed to have the
representations
• μ P ∈ [0, 1] X , of P.
• μ Q ∈ [0, 1]Y , of Q.
• μ P ∗ ∈ [0, 1] X , of P ∗ .
• (μ P → μ Q )(x, y) = J (μ P (x), μ Q (y)), with a convenient T -conditional J , for
the rule.
From this representations should follow a representation of Q ∗ , that is, μ Q ∗ ∈

[0, 1]Y , by taking into account that it should be a logical consequence of the set of
premises {μ P ∗ , μ P → μ Q } by the Generalized Modus Ponens (GMP). That is, μ Q ∗
does verify
0 = T0 (μ P ∗ (x), J (μ P (x), μ Q (y))) μ Q ∗ (y), (3.1)
for all x ∈ X, y ∈ Y , and a continuous t-norm T0 verifying
0 = T0 (μ P (x), J (μ P (x), μ Q (y))) μ Q (y),
for all x ∈ X, y ∈ Y , stating that if P ∗ = P, then Q ∗ = Q. This is done by

preserving the Modus Ponens (MP) in the occasion in which ‘x is P’ is observed.
The fuzzy set μ P ∗ is called the input
P* Q*
and μ Q ∗ is the output.

Obviously,
μ Q ∗ (y) = sup T0 (μ P ∗ (x), J (μ P (x), μ Q (y))), ∀y ∈ Y,

x∈X
is the greatest function verifying the GMP (3.1). This formula is known as the Com-
positional Rule of Fuzzy Inference (CRI, for short), and was introduced by Lotfi A.
Zadeh as the output fuzzy logic considers in the systems that are described by rules. It
is not to be forgotten that T0 is the continuous t-norm that makes J a T0 -conditional.
Sometimes, the input is just numerical, crisp in the form ‘x = x0 ’, that is, ‘x is
P ∗ ’ is ‘x is x0 ’, or ‘x ∈ {x0 }’ and then μ P ∗ = μ{x0 } , with

1, if x = x0
μ{x0 } (x) =
0, if x = x0 .
In this case,
μ Q ∗ (y) = sup T0 (μ{x0 } (x), J (μ P (x), μ Q (y))) = J (μ P (x0 ), μ Q (y)), ∀y ∈ Y,

x∈X
is a simpler expression that does not force to compute supx∈X .

Sometimes, in addition to the input, the rule’s consequent μ Q is also numerical,
that is ‘y is Q’ is ‘y = y0 ’, or ‘y is y0 ’ , or ‘y ∈ {y0 }’. In this case
3.4 Inference with Fuzzy Rules 101

1, if y = y0
μ Q (y) = μ{y0 } (y) = ,
0, if y = y0
and the output is

J (μ P (x0 ), 1), if y = y0
μ Q ∗ (y) = J (μ P (x0 ), μ{y0 } (y)) =
J (μ P (x0 ), 0), if y = y0 .
Example 3.4.1 Take X = [0, 1], Y = [0, 10], and the rule ‘If x is small, then y is
big’, with the observation that ‘x is big’ and J (a, b) = max(1 − a, b). With the
y
membership functions μ P (x) = 1 − x, μ Q (y) = 10 , μ P ∗ (x) = x, it results
y y
μ Q ∗ (y) = sup W (x, max(x, )) = W (1, max(1, )) = 1,
x∈[0,1] 10 10
or, μ Q ∗ = μ1 , that means Q ∗ = all.
Example 3.4.2 With the same rule of last example and the input x0 = 0.5, it is
y
μ Q ∗ (y) = max(0.5, max(0.5, )), ∀y ∈ [0, 1]
10
graphically,
Q*
Example 3.4.3 With the rule ‘If x is small, then y = 8’, and the input x = 0.5, is

1, if y = 8
μ Q ∗ (y) = max(0.5, μ{8} (y)) = ,
0.5, if y = 8
graphically,
Remark 3.4.4 In general, it is not easy to assign a name to the functional output μ Q ∗ ,
that is, to express Q ∗ linguistically. In Example 3.4.2, it could be said Q ∗ = big
after 5 and constantly equal to 0.5 before 5. In Example 3.4.3, it could be said Q ∗ =
almost always 0.5.
3.4.1 Finite Case
Let’s consider the particular case where both universes X and Y are finite sets. If
X = {x1 , . . . , xn }, Y = {y1 , . . . , ym },
fuzzy sets μ P , μ Q and μ P ∗ are of the forms:

• μ P = r1 /x1 + · · · + rn /xn , meaning μ P (xi ) = ri , 1 i n
• μ Q = s1 /y1 + · · · + sn /ym , meaning μ Q (y j ) = si , 1 j m
• μ P ∗ = r1∗ /x1 + · · · + rn∗ /xn , meaning μ P ∗ (xi ) = ri∗ , 1 i n
Provided J is an adequate T0 -conditional to represent the rule ‘If x is P, then y is

Q’ ((μ P → μ Q )(x, y) = J (μ P (x), μ Q (y))), it is
(μ P → μ Q )(xi , y j ) = J (μ P (xi ), μ Q (y j )) = J (ri , s j )

= ai j , 1 i n, 1 j m.
With all that,
μ Q ∗ (y j ) = sup T0 (μ P ∗ (xi ), J (μ P (xi ), μ Q (y j )) = max T0 (ri∗ , J (ri , s j ))

1i n 1i n
= max T0 (ri∗ , ai j ), 1 i n,
1i n
since in the finite case the sup is just max.
Calling μ Q ∗ = s1∗ /y1 + . . . + sm∗ /ym , it results
s ∗j = max T0 (ri∗ , ai j ), 1 j m,
1i n
showing an special ‘composition’ ⎛ of the matrices

⎞
a11 a12 . . . a1m
⎜ a21 a22 . . . a2m ⎟
⎜ ⎟
(r1∗ , . . . , rn∗ ) = [μ P ∗ ] and ⎜ . . .. ⎟ = [J ],
⎝ .. .. . ⎠
an1 an2 . . . anm

in which the elements of the classical product of matrices (rows by columns) ri∗ ·
1i n
ai j are substituted by max T0 (ri∗ , ai j ).
1i n
This composition is called the max-T0 product of matrices, instead of the classical
sum-prod composition. Hence,
[μ Q ∗ ] = (s1∗ , . . . , sm∗ ) = [μ P ∗ ] ⊗ [J ],
gives the CRI’s output.
Example 3.4.5 With μ P = 0.7/x1 + 0.8/x2 + 1/x3 , μ Q = 0.9/y1 + 0.6/y2 +

0.8/y4 , μ P ∗ = 0.6/x1 + 0.7/x2 + 1/x3 , and J (a, b) = min(1, 1 − a + b), follows:
a11 = J (0.7, 0.9) = min(1, 1 − 0.7 + 0.9) = 1; a12 = J (0.7, 0.6) = 0.9; a13 =
J (0.7, 0) = 0.3; a14 = J (0.7, 0.8) = 1
a21 = J (0.8, 0.9) = 1; a22 = J (0.8, 0.6) = 0.8; a23 = J (0.8, 0) = 0.2; a24 =
J (0.8, 0.8) = 1
a31 = J (1, 0.9) = 0.9; a32 = J (1, 0.6) = 0.6; a33 = J (1, 0) = 0; a34 =
J (1, 0.8) = 0.8.
Hence, ⎛ ⎞
1 0.9 0.3 1
[μ Q ∗ ] = (0.6 0.7 1) ⊗ ⎝ 1 0.8 0.2 1 ⎠ = (0.9 1 0.9 0.8),
0.9 0.6 0 0.8
since: (max(W (0.6, 1), W (0.7, 1), W (1, 0.9)), max(W (0.6, 0.9), W (0.7, 0.8),
W (1, 0.6)), max(W (0.6, 0.3), W (0.7, 0.2), W (1, 0)), max(W (0.6, 1), W (0.7, 1),
W (1, 0.8))) = (max(0.6, 0.7, 0.9), max(1, 1, 0.6), max(0.9, 0.9, 0),
max(0.6, 0.7, 0.8)) = (0.9 1 0.9 0.8). That is
μ Q ∗ = 0.9/y1 + 1/y2 + 0.9/y3 + 0.8/y4 .
In the case μ P is interpreted P = more or less big, μ Q is interpreted Q = not

very big, and μ P ∗ is interpreted P ∗ = medium, it is possible to agree on Q ∗ = more
or less big.
3.4.2 Inference with Several Rules
Actually, there are no systems described by a single rule. What to do when a system
is described by, at least, two rules? With, for example
• r1: If x is P1 , then y is Q 1
• r2: If x is P2 , then y is Q 2 ,
an input μ P ∗ gives
• with r1, the output μ Q ∗1 , obtained by using CRI

• with r2, the output μ Q ∗2 , obtained by using CRI.
Provided the firing of the rules corresponds to ‘fire r1 or fire r2’, then the final
output is given by
μ Q ∗ = max(μ Q ∗1 , μ Q ∗2 ),
and analogously for more than two rules. For example, in the case of p rules
r1 , . . . , r p , the result will be
μ Q ∗ = max(μ Q ∗1 , . . . , μ Q ∗p ),
where μ Q i∗ (1 i p) is the output obtained with the rule ri and the input μ P ∗ .
Example 3.4.6 With X = [0, 1], Y = [0, 10], consider the rules
• If x is big, then y = 2
• If x is small, then y = 8
• If x is ar ound 0.5, then y = 6,
and the input x0 = 0.4. Which is the final output of this system if the rules are
represented by J (a, b) = a · b?

0.4, if y = 2
• μ Q ∗1 (y) = J (μ B (0.4), μ{2} (y)) = 0.4.μ{2} (y) =
0, if y = 2

0.6, if y = 8
• μ Q ∗2 (y) = J (μs (0.4), μ{8} (y)) = 0.6.μ{2} (y) =
0, if y = 8
• μ Q ∗3 (y) = J (μ A0.5 (0.4), μ{6} (y)) = μ A0.5 (0.4).μ{6} (y)

μ A0.5 (0.4), if y = 6
=
0, if y = 6,
with μ B (x) = x, and μ S (x) = 1 − x. Taking as μ A0.5 the triangular function

and since the left side equation is y = x−0.35
0.15 , it is μ A0.5 (0.4) = 0. 3. Hence,

μ Q ∗3 (y) = 0. 3, if y = 6
0, if y = 6,
Finally
⎧
⎪
⎪ 0.4, if y = 2
⎪
⎨
μ Q ∗ (y) = max(μ Q ∗1 (y), μ Q ∗2 (y), μ Q ∗3 (y)) = 0. 3, if y = 6
⎪
⎪ if y = 8
⎪ 0.6,
⎩
0, otherwise
the output is given by the graphics,
Q*
In many applications, the output should be converted into a single numerical

value: it should be ‘defuzzified’. In this cases with numerical input and numerical
rule’s consequents (the most used in fuzzy control), such number is easyly obtained
by averaging the values of μ Q ∗ , in the form

2 × 0.4 + 6 × 0. 3 + 8 × 0.6 7.5998
= = 5.6999 5.7
1.3333
0.4 + 0. 3 + 0.6
Hence, the numerical output that corresponds to the input x0 = 0.4, is y0 = 5.7.
Remark 3.4.7 Notice that once a system of rules linguistically describing the
behavior of a system is given, and where the consequents of the rules are numerical,
at each numerical input x0 in X does correspond a numerical output y0 in Y . In that
way, a function CRI: X → Y is defined. As it will be later on commented, were
the system’s behaviour previously known by a continuous function f : X → Y , the
function CRI approaches, under some additional conditions, the function f .
Remark 3.4.8 Look how important is to properly select the T-conditionals repre-
senting the rules.
Given the rule ‘If x is small, then y is big’, with X = Y = [0, 1], and μ S (x) =
1 − x, μ B (y) = y, J (a, b) = max(1 − a, b),
it follows J (μ S (x), μ B (y)) = max(x, y), that could be interpreted as ‘x is big or y
is big’.
With J (a, b) = min(1, 1 − a + b), it follows J (μ S (x), μ B (y)) = min(1, x + y) =
W ∗ (x, y), also interpretable as ‘x is big or y is big’. But with J (a, b) = min(a, b),
is J (μ S (x), μ B (y)) = min(1 − x, y), interpreted as ‘x is small and y is big’.
3.4.3 Examples
Example 3.4.9 Rule ‘If x is big, then y = 0.8’, with x, y in [0, 1], and the observation
x ∈ [0.4, 0.6]. Hence:

x, if y = 0.8
J (μ B (x), μ{0.8} (y)) = x · μ{0.8} (y) =
0, if y = 0.8

0.6, if y = 0.8
Then, μ Q ∗ (y) = Sup min(μ[0.4,0.6] (x), xμ{0.8} (y)) =
x∈[0,1] 0, if y = 0.8,
since,
Example 3.4.10 Rule: ‘If x is big, then y is small’, with the same observation as that
in the last example and with μ B (x) = x, μ S (y) = 1 − y, and J (a, b) = min(a, b),
follows:
μ Q ∗ (y) = Sup min(μ[0.4,0.6] (x), min(x, 1 − y)) = Sup min(min(μ[0.4,0.6] (x),
x∈[0,1] x∈[0,1]
x), 1 − y) = Sup min(x, 1 − y) = min(0.6, 1 − y).
x∈[0,1]
Example 3.4.11 X = {1, 2, 3}, Y = {6, 7}. Rule: ‘If x is around 2, then y = 6’, and
μ P ∗ (x) = 0.6/1 + 0.9/2 + 0.7/3, with ⎛ J (a, ⎞
b) = ab (Larsen). It results
0.5 0
(μ Q ∗ (6), μ Q ∗ (7)) = (0.6 0.9 0.7) ⊗ ⎝ 1 0 ⎠ = (0.9 0.7), that is μ Q ∗ = 0.9/6 +
0.5 0
0/7.
Example 3.4.12 With x ∈ [0, 1], and y ∈ [0, 1], consider

• r1:‘If x is big, then y is small’, represented by J1 (a, b) = ab
• r2:‘If x is very small, then y is very big’, represented by J2 (a, b) = min(a, b)
Consider x 0 = 0.4, μ B (x) = x, μ S (y) = 1 − y, μV S (x) = (1 − x)2 , μV B (y) = y 2 .
Then
• J1 (μ B (x), μ S (y)) = x(1 − y)

• J2 (μV S (x), μV B (y)) = min((1 − x)2 , y 2 ) = [min(1 − x, y)]2 .
Hence,
• μ Q ∗1 (x) = J1 (0.4, 1 − y) = 0.4 · (1 − y),

0.36, if y 0.6
• μ Q ∗2 (x) = J2 ((1 − 0.4)2 , y 2 ) = (min(0.6, y))2 =
y 2 , if y < 0.6,
whose graphics are,
µ Q*
µ Q1* µ Q2*
Finally, ⎧
⎨ 0.4(1 − y), if 0 y 0.463
μ Q ∗ (y) = max(μ Q ∗1 (y), μ Q ∗2 (y)) = y 2 , if 0.463 y 0.6
⎩
0.36, if 0.6 y 1.
Example 3.4.13 Let’s find the function CRI:X → Y , in the case with X = [0, 1],
Y = [0, 1], and
• r1: If x is small, then y = 9
• r2: If x is big, then y = 2,

1 − x, if y = 9
it follows μ Q ∗1 (y) = (1 − x)μ{9} (y) =
0, if y = 9,

x, if y = 2
μ Q ∗2 (y) = xμ{2} (y) =
0, if y = 2,
and
⎧
⎨ x, if y = 2
μ Q ∗ (y) = max(μ Q ∗1 (y), μ Q ∗2 (y)) = 1 − x, if y = 9 that gives,
⎩
0, otherwise
2x + 9(1 − x)
CRI(x) = = 9 − 7x,
x +1−x
as the “theoretical” (linear) behavior of the system (x, y). For each x0 ∈ X , the value
CRI(x0 ) ∈ Y is the defuzzified value that corresponds to x0 .
Remark 3.4.14 Systems of fuzzy rules behave as universal approximators. This

means the following. Suppose a system (x, y), with x ∈ [a, b], y ∈ [c, d], that
behave by following the continuous function f (x) = y. For each ε > 0,there is
always a system of fuzzy rules and a defuzzification method for the output, giving a
function CRI:X → Y such that
| f (x) − CRI(x)| < ε, for all x ∈ [a, b]
This theorem (whose proof is here avoided) is simply an existential one, since
there is no general method for obtaing neither a fuzzy representation of the system
of rules, not the defuzzifiction method. It simply shows that it is possible to find a
CRI approaching enough well f for all points in [a, b].
3.5 Deffuzification
How to defuzzify non discrete outputs μ Q ∗ ? Let us proceed with two examples
without computational difficulties.
1st Example. Rules,
• r1: If x is big, then y is small
• r2: If x is small, then y is big
with X = [0, 1], and Y = [0, 10]. Take,
y y
μ B (x) = x, μ S (y) = 1 − , μ S (x) = 1 − x, μ B (y) = ,
10 10
and J (a, b) = min(a, b) -Mamdani-. Notice that, with the observation x 0 = 0.5,
y y
μ Q ∗1 (y) = min(0.5, 1 − ), μ Q ∗2 (y) = min(1 − 0.5, ).
10 10
3.5 Deffuzification 109
y y y
Then, μ Q ∗ (y) = max(min(0.5, 1 − 10 ), min(0.5, 10 )) = max(0.5, min(1 − 10 ,
y y y
10 )) = 0.5, since min(1 − ,
10 10 ) 0.5.
The area below μ Q ∗ (y) = 0.5, is A = 0.5 × 10 = 5 square units. Hence, a way to
defuzzify μ Q ∗ consists of searching the center of area, that is, a point y0 ∈ [0, 10]
such that
y0 y0
A
0.5dy = = 2.5, or dy = 5.
2
0 0
y
Hence, [y]00 = y0 = 5. The defuzzified value corresponding to x0 = 0.5, is y0 = 5.
The method, when the conditional is Mamdani, is graphically reflected as follows.
2nd Example. Identical to the first example, but with the input x0 = 0.3. It is
y y
μ Q ∗1 (y) = min(0.3, 1 − ), μ Q ∗2 (y) = min(1 − 0.3, ).
10 10
Hence,
⎧
⎨ 0.3, if 0 y 3
y
μ Q ∗ (y) = 10 , if 3 y 7
⎩
0.7, if 7 y 10
that is graphically find as follows:

The area below μ Q ∗ is A = rectangle(1) + rectangle(2) + triangle (3) = 0.3 × 7 +

3 × 0.7 + 4×0.4
2 = 5. Since the area of the rectangle with base [0, 3] is 0.3 × 3 = 0.9,
it is y0 > 3. Hence,
3 y0
y
0.3dy + dy = A/2 = 2.5.
10
0 3
y0 y0
Then 3 × 0.3 + 1
10 3 ydy = 2.5, or 3 ydy = 10(2.5 − 0.9) = 16. Thus,
y 2 y0
[ ] = 16 ⇒ y02 − 9 = 32 ⇒ y02 = 41 ⇒ y0 = 6.4.
2 3
The defuzzified value that corresponds to x0 = 0.3, is y0 = 6.4.
Defuzzifying with the centre of the area we obtained an output for all values x0
in [0, 1]. Let’s see it by means of the function CRI with defuzzification made by the
centre of the area.
1. The graphics, for any input x0 1/2, is
The area below μ Q ∗ is
(1 − x0 − x0 )(10(1 − x0 ) − 10x0 )
A = 10(1 − x0 )x0 + (10 − 10(1 − x0 ))(1 − x0 ) +
2
3.5 Deffuzification 111
then,
2 A = 20(1 − x0 )x0 + 20(1 − x0 )x0 + (1 − 2x0 )(10 − 20x0 ),
and
A = 20x0 (1 − x0 ) + 5 − 20x0 + 20x02 = 5.
Hence,
y0
y
10x02 + dy = 2.5 ⇒ y0 = 50 − 100x02 ,
10
10x0
and

C R I (x) = 50 − 100x 2 , if x 1/2.
√
For example, C R I (0.5) = 0.5, and C R I (0.3) = 41 = 6.4, as it was shown. It
is also C R I (0.1) = 7.
√
Last formula, C R I (x) = 20 − 100x 2 , gives real values provided 20−100x 2
0, or x 2 1/5. Since, it is x 1/5, that implies x 2 1/4 1/2, and it follows
that the formula is useful for all x ∈ [0, 1] such that x 1/2.
2. For any input x0 1/2, the graphic is
and the area below μ Q ∗ is
(10x0 − 10(1 − x0 ))(x0 − 1 + x0 )

A = 10x0 (1 − x0 ) + (10 − 10x0 )x0 +
2
= 20x0 (1 − x0 ) + (10x0 − 5)(2x0 − 1) = 5.
Hence, A/2 = 2.5, and
y0 y0
y
10(1 − x0 ) +2
dy = 2.5, or 100(1 − x0 )2 + ydy = 2.5,
10
10(1−x0 ) 10(1−x0 )

giving y02 = 50 − 100(1 − x0 ), or y0 = 50 − 100(1 − x0 )2 , that gives real
values provided
√ 50 − 100(1 − x0 )2 0, equivalent to 1/2 (1 − x0 )2 , or to
x0 1 − 1/2. Since, 1/2 x, it is (1 − x)2 1/4 1/2. Then,

C R I (x) = 50 − 100(1 − x)2 , if 1/2 x.
For example,√ √ √
CRI(0.7) = 41 = 6.4, CRI(0.8) = 46 = 6.78, CRI(0.9) = 49 = 7.
3. Finally, with defuzzification by the centre of area, is:

50(1 − 2x ), if x 1/2
2
CRI(x) =
50(1 − 2(1 − x) ), if x 1/2.
2
√ √
Notice that√CRI(0) = 50, CRI(0.5) = 50(1 − 20.52 ) = 25 = 5, and
CRI(1) = 50.
The graphic of CRI is
3.6 Rules and Conjectures
As it was said before, the output is a logical consequence of the premises given by
the input and the rule. Notwithstanding, the situation is different if, taking the rule as
defining the system, only the input is considered as a premise. But, before to consider
this question, let us consider what happens when there is more than a single rule.
1. If μ Q ∗1 , μ Q ∗2 ∈ Cons({μ P ∗ }), from μ P ∗ max(μ Q ∗1 , μ Q ∗2 ) = μ Q ∗ , follows
μ Q ∗ ∈ Cons({μ P ∗ }).
2. If μ Q ∗1 or μ Q ∗2 is a conjecture of {μ P ∗ }, then μ Q ∗ ∈ Con j ({μ P ∗ }). The proof
follows in this way, provided it is, for instance, μ Q ∗1 ∈ Con j ({μ P ∗ }),
• μ Q ∗ = max(μ Q ∗1 , μ Q ∗2 ) ⇒ μ Q ∗1 μ Q ∗ , μ Q ∗2 μ Q ∗ ⇒ μQ ∗ μQ ∗ , μQ ∗
1
μQ ∗ ⇒ μQ ∗ min(μQ ∗ , μQ ∗ ).
2 1 2
• If μ P ∗ μQ ∗ , then μ P ∗ μQ ∗ , that is absurd. Hence, it is μ P ∗ μQ ∗ , and
1
μ Q ∗ ∈ Con j ({μ P ∗ }).
3.6 Rules and Conjectures 113
In conclusion
• If all the partial outputs μ Q ∗1 , μ Q ∗2 , ..., μ Q ∗n , are consequences of the input μ P ∗ ,
also the final output μ Q ∗ is a consequence of μ P ∗ .
• If at least one of the partial outputs μ Q i∗ (1 i n) is just a conjecture of the
input μ P ∗ , also the final output μ Q ∗ is a conjecture of μ P ∗ .
Nevertheless, it is not usual that μ Q ∗ results to be a consequence of the single input
μ P ∗ . Let’s us introduce a necessary and sufficient condition for it in the particular
case in which there is only one rule represented by J (a, b) = ab (Larsen).
Let it be “If x is P, then y is Q” (x, y ∈ X ), and (μ P → μ Q )(x, y) = μ P (x)μ Q (y),

with the input x = x0 . Then,
μ Q ∗ (y) = μ P ∗ (x0 )μ Q (y), ∀y ∈ X.
Provided μ{x0 } = μ0 , to have μ{x0 } (y) μ Q ∗ (y) = μ P ∗ (x0 )μ Q (y), it is neces-

sary that, with y = x0 , 1 μ P ∗ (x0 )μ Q ∗ (x0 ) or 1 = μ P ∗ (x0 ) = μ Q ∗ (x0 ). Hence,
μ Q ∗ ∈ Cons({μ{x0 } }) implies μ P ∗ (x0 ) = μ Q ∗ (x0 ) = 1.
Provided μ P ∗ (x0 ) = μ Q ∗ (x0 ) = 1, from μ Q ∗ (y) = μ P ∗ (x0 )μ Q (y), follows

μ Q ∗ (y) = μ Q (y), for all y ∈ X , and,
• If y = x0 , μ{x0 } (x0 ) = 1 = μ Q ∗ (x0 )
• If y = x0 , μ{x0 } (y) = 0 μ Q ∗ (y),
that is
μ{x0 } (y) μ Q ∗ (y), for all y ∈ X.
Hence, in this particular case, the necessary and sufficient condition for being μ Q ∗ ∈
Cons({μ{x0 } }) is that μ P ∗ (x0 ) = μ Q ∗ (x0 ) = 1. Nevertheless, what happens in
most of the cases is that μ Q ∗ ∈ Con j ({μ{x0 } }), with μ Q ∗ ∈ Sp({μ{x0 } }), or μ Q ∗ ∈
H yp({μ{x0 } }).
3.7 Two Final Examples
Let’s show an example in which the output is a speculation of the input and other in
which the output is a hypothesis.
Example. Rule, “If x is small, then y is big”, with X = Y = [0, 10] and J (a, b) = ab
y
(Larsen), with μ S (x) = 1 − 10x
, μ B (y) = 10 , and x0 = 5. It is
5 y y
μ Q ∗ (y) = J (1 − , )= ,
10 10 20
and from the graphic
{5}
it is clear that μ Q ∗ is not comparable with μ{5} (μ Q ∗ NC μ{5} ), and that μ{5}
μQ ∗ = 1 − μ Q ∗ . Hence,
μ Q ∗ ∈ Con j ({μ{5} }), and namely μ Q ∗ ∈ Sp({μ{5} }).
Notice that μ Q ∗ (5) = 1

4 = 1.
Example. Rule, “If x is big, then y is very big”, with X = Y = [0, 10] and the
observation “x is constantly equal to 0.8” for all x ∈ [0, 1]. Taking
μ B (x) = x, μv B (y) = y 2 , J (a, b) = min(a, b) (Mamdani),
follows
μ Q ∗ (y) = min(μ P ∗ (0.8), μv B (y)) = min(0.8, y 2 ).
Graphically,
√
since y 2 = 0.8 means y = 0.8. Hence, it is μ0 = μ Q ∗ μ P ∗ , and μ P ∗ = μQ ∗ =
1 − μ Q ∗ , that imply μ Q ∗ ∈ Con j ({μ P ∗ }), and namely μ Q ∗ ∈ H yp({μ P ∗ }).
Notice that this second example contains the observation that the input is a con-
stant.
Last Remark
For any continuous t-norm T , the function μ : Y → [0, 1], defined by
μ(y) = Sup T (μ P ∗ (x), J (μ P (x), μ Q (y))), ∀y ∈ Y,

x∈X
does verify
3.7 Two Final Examples 115
T (μ P ∗ (x), J (μ P (x), μ Q (y))) μ(y), ∀y ∈ Y, x ∈ X,
that is, μ ∈ Cons({μ P ∗ , μ P → μ Q }). Nevertheless, if T = T0 , the continuous

t-norm for which J is a T0 -conditional, that is, such that
T0 (μ P (x), J (μ P (x), μ Q (y))) μ(y) =, ∀y ∈ Y, x ∈ X,
it could be that when μ P ∗ = μ P , then μ = μ Q . A undesiderable situation, because

fuzzy logic must contain all classical cases.
For example, with the rule ‘If x is small, then y is big’ (X = Y = [0, 1]),
and J (a, b) = max(1 − a, b) that is a W -conditional, taking μ S (x) = 1 − x and
μ B (y) = y, follows:
• With T = W , μ(y) = Sup W ((1 − x), max(x, y)) = Sup (0, y − x) = y =
x∈[0,1] x∈[0,1]
μ B (y).
• With T = pr od, μ(y) = Sup (1 − x) max(x, y) = Sup max((1 − x)x, (1 −
x∈[0,1] x∈[0,1]
x)y) = y = max(1/4, y/2), not coincidental with μ B .
• with T = min, μ(y) = Sup min((1 − x), max(x, y)) = 1, or μ = μ1 , also not
x∈[0,1]
coincidental with μ B .
Hence, although with any continuous t-norm T , an output μ is obtained, if this T

does not make J a T -conditional it is not sure that P = P ∗ implies Q = Q ∗ . For
this reason, it is necessary to take T0 with CRI!
Chapter 4
Fuzzy Relations
4.1 What Is a Fuzzy Relation?
A predicate R on a cartesian product X 1 × · · · × X n is called a relational (n-ary)

predicate. For example, if X 1 = X 2 = [0, 10], R = close to, ‘(x, y) is R’, or ‘x is
close to y’, is a relational binary predicate.
Analogously, if X 1 = X 2 = London, R = lives in the same borough, or ‘x lives
in the same borough than y’, is a relational binary predicate.
A fuzzy relation in X 1 × · · · × X n is any function μ : X 1 × · · · × X n → [0, 1].
If interpreting μ R (x1 , . . . , xn ) =‘degree up to which (x1 , . . . , xn ) is in R ’, it is said

that μ R represents the n-ary relational relation R’.
Any rule ‘If x is P, then y is Q’ defines the binary predicate Q/P in X ×Y given by
(x, y) is Q/P ⇔ If x is P, then y is Q,
whose representation, or membership function of the corresponding fuzzy set Q/P

is given by
μ Q/P (x, y) = (μ P → μ Q )(x, y) = J (μ P (x), μ Q (y)),
once a T-conditional J adapted to the meaning of Q/P is selected. A fuzzy relation

μ R is nothing else than a fuzzy set R in X 1 × · · · × X n .

For example, if R = ‘close to’, is represented by
μ R (x, y) = max(0, k|x − y|), for all x, y ∈ [0, 1],

118 4 Fuzzy Relations
with k ∈ (0, 1) a parameter chosen at each particular case, it is μ R (0, 0) = 0,

μ R (1, 1) = 0, μ R (0, 1) = μ R (1, 0) = max(0, k) = k, μ R (1/2, 1) = μ R (1, 1/2) =
max(0, k/2) = k/2, etc., with the graphic,
When the sets X 1 , . . . , X n are finite, μ R is reduced to a matrix. For example if

X 1 = {x1 , . . . , x p }, and X 2 = {y1 , . . . , yq }, then
μ R (xi , y j ) = ri j , 1 i n, 1 j m, or,
μ R = r11 /(x1 , y1 ) + · · · + rnm /(xn , ym ),
that gives the n × m matrix

⎛ ⎞
r11 r12 . . . r1m
⎜ r21 r22 . . . r2m ⎟
⎜ ⎟
[R] = ⎜ . .. .. ⎟ .
⎝ .. . . ⎠
rn1 rn2 . . . rnm
In the finite case there is again another representation of a fuzzy relation by means
of a directed graph.
For example, if X 1 = {x1 , x2 } and X 2 = {y1 , y2 , y3 }, the fuzzy
0.5 0.7 1
relation [R] = , corresponds to the directed graph
0.8 0 0.8
4.2 How to Compose Fuzzy Relations? 119
0.5 y1
0.8
x1
0.7
y2
x2 1
0.8
y3
4.2 How to Compose Fuzzy Relations?
Given two fuzzy relations μ : X × Y → [0, 1], and σ : Y × Z → [0, 1], how can
we obtain a relation λ : X × Z → [0, 1] through μ and σ? To solve this problem,
there is the Sup − T product of fuzzy relations, given by
λ(x, z) = Sup T (μ(x, y), σ(y, z)), for all (x, z) ∈ X × Z ,

y∈Y
a formula that, in the finite case X = {x1 , . . . , xn }, Y = {y1 , . . . , ym }, Z =

{z 1 , . . . , z p }, reduces to,
λ(xi , z j ) = Max T (μ(xi , yk ), σ(yk , z j )).

1k m
Provided [μ] = (rik ), [σ] = (sk j ), then
ti j = λ(xi , z j ) = Max T (ri,k , sk j ), 1 i n, 1 j p,

1k m
giving the matrix [λ] = (ti j ) as the Max-T product, or composition, of the matrices
[rik ] and [sk j ], that was introduced before by: (ti j ) = (rik ) ⊗T (sk j ).
Example 4.2.1 1 Let
• X = { p1 , . . . , p4 }, a set of patients
• Y = {s1 , s2 , s3 }, a set of symptoms
• Z = {d1 , . . . , d5 }, a set of deceases,
and the fuzzy relation σ
⎛ ⎞
0.7 0 0 0.3 0.6
[σ] = ⎝ 0.5 0.5 0.8 0.4 0 ⎠
0 0.7 0.2 0.9 0
showing the medical knowledge of how strongly each symptom is associated with a
decease. Suppose also that, by examining the patients, the doctors conclude the matrix
⎛ ⎞
0 0.3 0.4
⎜ 0.2 0.5 0.3 ⎟
[μ] = ⎜
⎝ 0.8 0
⎟
0 ⎠
0.7 0.7 0.9
that describes numerically how strongly the symptoms are manifested in the patients.
Then,
[λ] = [μ] ⊗min [σ]
is the matrix expressing the association patients/deceases, and facilitates a medical

diagnose. That is,
⎛ ⎞
0 0.3 0.4 ⎛ ⎞
⎜ 0.2 0.5 0.3 ⎟ 0.7 0 0 0.3 0.6
[λ] = ⎜⎝ 0.8 0
⎟ ⊗ ⎝ 0.5 0.5 0.8 0.4 0 ⎠
0 ⎠ min
0 0.7 0.2 0.9 0
0.7 0.7 0.9
⎛ ⎞
0.3 0.4 0.3 0.4 0
⎜ 0.5 0.5 0.5 0.4 0.2 ⎟
=⎜ ⎝ 0.7 0
⎟,
0 0.3 0.6 ⎠
0.7 0.7 0.7 0.9 0.6
where, for instance,
t43 = max(min(0.7, 0), min(0.7, 0.8), min(0.9, 0.2)) = max(0, 0.7, 0.2) = 0.7.
The matrix [λ] results from a mixing between knowledge and observation.
1 From, Fuzzy Set Theory, by G.J. Klir, U.H. St. Clair, B. Yuan, Prentice/Hall, N.J., 1997.
4.2 How to Compose Fuzzy Relations? 121
Remark 4.2.2 1. As it is easy to prove, it is
([μ] ⊗T [σ])t = [σ]t ⊗T [μ]t ,
with the matrices [σ]t , [μ]t , defined by σ t (x, y) = σ(y, x), μt (x, y) = μ(y, x),
giving
([μ]t )t = [μ].
The matrix [μ]t is the transposed of [μ].

2. In general, the max-T composition is associative, but not commutative. That is, if
the compositions [μ] ⊗T ([σ] ⊗T [λ]), and ([μ] ⊗T [σ]) ⊗T [λ], are possible it is,
([μ] ⊗T [σ]) ⊗T [λ] = [μ] ⊗T ([σ] ⊗T [λ]),
but, in general, [μ] ⊗T [σ] = [σ] ⊗T [μ].
4.3 Which Relevant Properties Do Have a Fuzzy Binary

Relation?
The most relevant properties of a fuzzy relation μ : X × X → [0, 1] are the following,
1. Reflexive property, μ(x, x) = 1, for all x ∈ X .
2. Symmetric property, μ(x, y) = μ(y, x), for all x, y ∈ X , implies x = y.
3. Antisymmetric property, μ(x, y) > 0, μ(y, x) > 0 implies x = y.
4. T-transitive property T (μ(x, y), μ(y, z)) μ(x, z), for all x, y, z ∈ X , and
some continuous t-norm T .
In the finite case, for what concerns properties reflexive and symmetric, the matrix
[μ] = (ti j ) shows the respective properties,
1 . It is tii = 1, for all 1 i n, that is, the main diagonal of [μ] is constituted
by n numbers equal to 1.
2 . It is ti j = t ji , for all 1 i, j n, that is, the elements of [μ] are placed
symmetrically with respect to the main diagonal.
For example, the matrix

1 0.7
is reflexive, but not symmetric,
0.6 1
and the matrix

⎛ ⎞
1 0.6 0.7
⎝ 0.6 0 0.9 ⎠ is symmetric, but not reflexive.
0.7 0.9 0.5
For what concerns the antisymmetric property, ti j > 0 and t ji > 0, implies i = j.
For example, the matrix ⎛ ⎞
1 0 0.7 0
⎜ 0.6 1 0 0.7 ⎟
⎜ ⎟
⎝ 0 0.5 1 0.8 ⎠
0.7 0 0 1
is antisymmetric.
Let’s define the binary relation, [μ] [σ], between n × n matrices if ti j si j

for all 1 i, j n, provided [μ] = (ti j ), [σ] = (si j ). With such definition,
• [μ] reflects a T -transitive fuzzy relation μ, if and only if, [μ] ⊗T [μ] [μ].
The proof is as follows.
a. If [μ] is T -transitive, from
T (μ(xi , x j ), μ(x j , xk )) μ(xi , xk ),
or T (ti j , t jk ) tik , it is Max T (ti j , t jk ) tik . That is, [μ] ⊗T [μ] [μ].
1 j n
b. If [μ] ⊗T [μ] [μ], or Max T (ti j , t jk ) tik , follows T (ti j , t jk ) tik , for all
1 j n
1 i, j n. That is, μ is T -transitive.
• If μ is reflexive and T -transitive, it is [μ] ⊗T [μ] = [μ]. since
tik = T (1, tik ) = T (tii , tik ) Max T (ti j , t jk ) tik ,

1 j n
implies
tik = Max T (ti j , t jk ), or [μ] = [μ] ⊗T [μ].
1 j n
Remark 4.3.1 The definitions given in this section contain the case of the corre-
sponding classical crisp definitions,
• If R ⊂ X × X is a classical reflexive relation in X , its membership function μ R
reflects (x, x) ∈ R for all x ∈ X , by μ R (x, x) = 1.
• If R ⊂ X × X is a classical symmetric relation in X,
(x, y) ∈ R ⇔ (y, x) ∈ R is reflected by μ(x, y) = μ(y, x).
• If R ⊂ X × X is antisymmetric,
(x, y) ∈ R & (y, x) ∈ R ⇔ x = y
is reflected by μ R (x, y) = μ R (y, x) = 1(>0) ⇒ x = y.

4.3 Which Relevant Properties Do Have a Fuzzy Binary Relation? 123
• If R is transitive, (x, y) ∈ R & (y, z) ∈ R ⇒ (x, z) ∈ R, is reflected by

μ R (x, y) = μ R (y, z) = 1 ⇒ μ R (x, z) = 1, that implies,
T (μ R (x, y), μ R (y, z)) μ R (x, z),
for all t-norms T . Notice that if μ R (x, y) = 0 or μ R (y, z) = 0, then, for example,
T (μ R (x, y), μ R (y, z)) = T (0, μ R (y, z)) = 0 μ R (x, z).
Hence, for all x, y, z in X , and any continuous t-norm T , is
T (μ R (x, y), μ R (y, z)) μ R (x, z).
that reflects equationally the transitivity of R.
Example 4.3.2 1. The matrix

⎛ ⎞
1 1/8 2/8
[μ] = ⎝ 1/8 1 3/8 ⎠
2/8 3/8 1
is reflexive and symmetric (fuzzy similarity). In addition,
[μ] ⊗ pr od [μ] = [μ],
that is, [μ] is prod-transitive. Notice that,

⎛ ⎞
1 1/8 2/8
[μ] ⊗min [μ] = ⎝ 1/8 1 3/8 ⎠ = [μ],
2/8 2/8 1
and [μ] is not min-transitive. Of course, since W pr od, [μ] is also W -

transitive. Notice that this last matrix is reflexive, non symmetric, but min-
transitive, since
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 1/8 2/8 1 1/8 2/8 1 2/8 2/8
⎝ 1/8 1 3/8 ⎠ ⊗min ⎝ 1/8 1 3/8 ⎠ = ⎝ 2/8 1 3/8 ⎠ ,
2/8 2/8 1 2/8 2/8 1 2/8 2/8 1
hence, this 3 × 3-matrix reflects a min-preorder and, consequently, a T -preorder

for all t-norm T .
2. The before mentioned fuzzy relation μ(x, y) = max(0, 1 − k|x − y|) is not only
reflexive and symmetric, but also W -transitive, as it can be proved by distinguish-
ing the four cases: k1 ≥ |x − y|, k1 ≥ |y −z|; k1 < |x − y|, k1 < |y −z|; k1 ≥ |x − y|,
k < |y − z|; and k < |x − y|, k ≥ |y − z|.
1 1 1
Remark 4.3.3 If the fuzzy relation μ R is T-transitive, and T1 is a t-norm such that
T1 T , then μ R is also T1 -transitive, since
T1 (μ R (x, y), μ R (y, z)) T (μ R (x, y), μ R (y, z)) μ R (x, z),
for all x, y, z.
4.4 The Concept of T-State
Given a fuzzy relation μ : X × X → [0, 1] and a continuous t-norm T , a fuzzy set

σ : X → [0, 1] is a T -state of μ, if
T (σ(x), μ(x, y)) σ(y), ∀(x, y) ∈ X × X.
All constant fuzzy sets μk (x) = k for all x ∈ X , and k ∈ [0, 1], are T-states of
any fuzzy relation μ : X × X → [0, 1]: T (μk (x), μ(x, y)) μk (x) = k = μk (y).
For example, μ0 = μ∅ and μ1 = μ X , are always T -states. Hence, the set T (μ) of
all T -states σ of μ is never empty. From now on, in general we will only refer to non
constant T -states σ.
Given a fuzzy relation μ : X × Y → [0, 1], once y ∈ Y is fixed, we can define

the fuzzy set μ y : X × X → [0, 1], defined by,
μ y (x) = μ(x, y), for all x ∈ X.
When X, Y are finite sets, μ y is the y-column of the matrix [μ].

If μ : X × X → [0, 1] is a symmetric and T -transitive fuzzy relation, from
T (μ(x, y), μ(y, z)) μ(x, z), for all x, y, z in X,
follows T (μ(y, x), μ(y, z)) μ(z, x), or
T (μx (y), μ(y, z)) μx (z),
that is, μx is a T -state of μ.
For example, with X = {x1 , x2 , x3 } and μ : X × X → [0, 1] given by

⎛ ⎞
1 1/8 2/8
[μ] = ⎝ 1/8 1 3/8 ⎠
2/8 3/8 1
4.4 The Concept of T-State 125
It is μx1 = 1/x1 + 1/8/x2 + 2/8/x3 , and it results,
μx1 (x1 )μ(x1 , y) = μ(x1 , y) = μx1 (y).
It is also μx2 = 1/8/x1 + 1/x2 + 2/8/x3 , and
1
μx2 (x1 )μ(x1 , y) = μ(x1 , y),
8
showing,
• 8 μ(x 1 , x 1 )
1
= 8 = μx2 (x 1 )
1
• 8 μ(x 1 , x 2 )
1
= 1
82
< 18 = μx2 (x2 )
• 8 μ(x 1 , x 3 )
1
= 2
82
< 38 = μx2 (x3 ),
etc. That is, the three fuzzy sets μx1 , μx2 , μx3 are pr od-states of μ.
Remark 4.4.1 When the fuzzy relation μ represents a conditional statement Q/P
(a fuzzy rule ‘If x is P, then y is Q’), the T -states of μ are among the fuzzy sets
verifying the Modus Ponens with respect to the continuous t-norm T .
4.5 Fuzzy relations and α-cuts
Given a fuzzy relation μ : X × X → [0, 1], the α-cuts of μ are the classical (crisp)
relations μ(α) defined by,
μ(α) = {(x, y) ∈ X × X ; μ(x, y) ≥ α},
for all α ∈ [0, 1]. Obviously, μ(0) = X × X , and if α1 ≥ α2 , it is μ(α1 ) μ(α2 ) .

• μ is symmetric, if and only if all its α-cuts are symmetric (classical) relations.
μ(x, y) = μ(y, x) ⇔ (x, y) ∈ μ(α) and (y, x) ∈ μ(α) .
• μ is reflexive, if and only if all α-cuts are reflexive (classical) relations.
μ(x, x) = 1 ⇔ (x, x) ∈ μ(α) , since α 1.
• If μ is antisymmetric, all the α-cuts are antisymmetric (crisp) relations,
If μ(x, y) ≥ α > 0, and μ(y, x) ≥ α > 0, it is x = y.
• If μ is a T -transitive fuzzy relation,
(x, y) ∈ μ(α) , and (y, z) ∈ μ(α) , implies (x, z) ∈ μ(T (α,α)) ,

since,
μ(x, y) ≥ α, μ(y, z) ≥ α ⇒ T (α, α) T (μ(x, y)μ(y, z)) μ(x, z).
Hence, only if μ is min-transitive, it is sure that all α-cuts are classical preorders,
and μ results decomposed in the family of preorders {μα ; α ∈ (0, 1)}.
Example 4.5.1 With X = {1, 2, 3, 4}, the matrix

⎛ ⎞
1 0.6 1 0.6
⎜ 0.3 1 0.3 0.3 ⎟
[μ] = ⎜ ⎟
⎝ 1 0.6 1 0.6 ⎠ ,
0.4 0.8 0.4 1
is obviously reflexive but not symmetric, and verifies [μ] ⊗min [μ] = [μ]. Hence, μ
is a min-preorder. Its different α-cuts are
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 0 1 0 1 0 1 0 1 1 1 1
⎜0 1 0 0⎟ ⎜0 1 0 0⎟ ⎜0 1 0 0⎟
[μ(1) ] = ⎜ ⎟ ⎜ ⎟ ⎜
⎝ 1 0 1 0 ⎠ , [μ(0.8) ] = ⎝ 1 0 1 0 ⎠ , [μ(0.6) ] = ⎝ 1 1 1 1 ⎠ ,
⎟
0 0 0 1 0 1 0 1 0 1 0 1
⎛ ⎞ ⎛ ⎞
1 1 1 1 1 1 1 1
⎜0 1 0 0⎟ ⎜1 1 1 1⎟
[μ(0.4) ] = ⎜ ⎟ ⎜
⎝ 1 1 1 1 ⎠ , [μ(0.3) ] = ⎝ 1 1 1 1 ⎠ ,
⎟
1 1 1 1 1 1 1 1
that give the classical (α) preorders, that follows:
• (1) : 1 (1) 1, 2 (1) 2, 3 (1) 3, 4 (1) 4, 1 (1) 3, 3 (1) 1.
• (0.8) : 1 (0.8) 1, 2 (0.8) 2, 3 (0.8) 3, 4 (0.8) 4, 1 (0.8) 3, 3 (0.8)
1, 4 (0.8) 2.
• (0.6) : 1 (0.6) 1, . . . , 4 (0.6) 4, 1 (0.6) 3, 3 (0.6) 1, 4 (0.6) 2, 1 (0.6)
2, 1 (0.6) 4, 3 (0.6) 2, 3 (0.6) 4.
• (0.4) : 1 (0.4) 1, . . . , 4 (0.4) 4, 1 (0.4) 3, . . . , 4 (0.4) 1, 4 (0.4) 3.
• (0.3) : 1 (0.3) 1, . . . , 4 (0.3) 3, 2 (0.3) 1, . . . , 4 (0.3) 2,
and, since, 0.3 0.4 0.6 0.8 1, verify (1) ⊂(0.8) ⊂(0.6) ⊂(0.4) ⊂(0.3) .
Example 4.5.2 The fuzzy relation μ : X × X → [0, 1], with X = {1, 2, 3, 4, 5, 6},
given by
⎛ ⎞
1 0.2 1 0.6 0.2 0.6
⎜ 0.2 1 0.2 0.2 0.8 0.2 ⎟
⎜ ⎟
⎜ 1 0.2 1 0.6 0.2 0.6 ⎟
[μ] = ⎜
⎜ 0.6 0.2 0.6 1 0.2 0.8 ⎟
⎟
⎜ ⎟
⎝ 0.2 0.8 0.2 0.2 1 0.8 ⎠
0.6 0.2 0.6 0.8 0.8 1
4.5 Fuzzy relations and α-cuts 127
is reflexive, symmetrical and min-transitive, since [μ] ⊗min [μ] = [μ]. Hence, all the
α-cuts are classical equivalence relations, each one defining a partition of X . The
different values of α are 0.2, 0.6, 0.8 and 1 (levels of crispness of μ), and it is easy
to see that the corresponding partitions π0.2 , π0.6 , π0.8 , and π1 , can be located as the
partition tree:
1 1 3 4 6 2 5
0.8 1 3 4 6 2 5
0.6 1 3 4 6 2 5
0.2 1 3 4 6 2 5
This tree is called the fuzzy quotient of X by μ.

Example 4.5.3 The fuzzy relation μ : {1, 2, . . . , 6} → [0, 1] given by
⎛ ⎞
1 0.8 0.2 0.6 0.6 0.4
⎜0 1 0 0 0.6 0 ⎟
⎜ ⎟
⎜0 0 1 0 0.5 0 ⎟
[μ] = ⎜
⎜0 0
⎟
⎜ 0 1 0.6 0.4 ⎟⎟
⎝0 0 0 0 1 0 ⎠
0 0 0 0 0 1
is reflexive, antisymmetric and min-transitive. Hence, is a fuzzy ordering whose

α-cuts should be crisp partial orderings. Namely,
⎛ ⎞
1 0 0 0 0 0 5 6
⎜0 1 0 0 0 0⎟
⎜ ⎟
⎜0 0 1 0 0 0⎟
1. [μ(1) ] = ⎜ ⎟
⎜ 0 0 0 1 0 0 ⎟, with partial order 2 3 4
⎜ ⎟
⎝0 0 0 0 1 0⎠
0 0 0 0 0 1 1
that only connects the pairs (1, 1), (2, 2), . . . (6, 6), and is called the disconnected
partial order.
5 6
⎛ ⎞
1 1 0 0 0 0
⎜0 1 0 0 0 0⎟
⎜ ⎟
⎜0 0 1 0 0 0⎟
2. [μ(0.8) ] = ⎜
⎜0
⎟ 2 3 4
⎜ 0 0 1 0 0⎟⎟
⎝0 0 0 0 1 0⎠
0 0 0 0 0 1
1
5 6
⎛ ⎞
1 1 0 1 1 0
⎜0 1 0 0 1 0⎟
⎜ ⎟
⎜0 0 1 0 0 0⎟
3. [μ(0.6) ] = ⎜
⎜0
⎟ 2 3 4
⎜ 0 0 1 1 0⎟⎟
⎝0 0 0 0 1 0⎠
0 0 0 0 0 1
1
5 6
⎛ ⎞
1 1 0 1 1 0
⎜0 1 0 0 1 0⎟
⎜ ⎟
⎜0 0 1 0 1 0⎟
4. [μ(0.5) ] = ⎜
⎜0
⎟ 2 3 4
⎜ 0 0 1 1 0⎟⎟
⎝0 0 0 0 1 0⎠
0 0 0 0 0 1
1
5 6
⎛ ⎞
1 1 0 1 1 1
⎜0 1 0 0 1 0⎟
⎜ ⎟
⎜0 0 1 0 1 0⎟
5. [μ(0.4) ] = ⎜
⎜0
⎟ 2 3 4
⎜ 0 0 1 1 0⎟⎟
⎝0 0 0 0 1 0⎠
0 0 0 0 0 1
1
4.5 Fuzzy relations and α-cuts 129
All this crisp partial orderings come from the fuzzy ordering given by [μ], and
which valued directed graph is
5 6
0.6
0.5 0.4
2 3 4
0.2
0.8 0.6
where some arrows are avoided because of the min-transitivity of μ. For instance,
the degree between 1 and 5, is
μ(1, 5) = Max min(μ(1, k), μ(k, 5)) = max(min(0.8, 0.6),

1k 6
min(0.2, 0.5)) = 0.6,
and the strengh of the order between 1 and 6, is
μ(1, 6) = Max min(μ(1, k), μ(k, 6)) = min(0.6, 0.4) = 0.4.

1k 6
Of course, μ(1, 1) = μ(2, 2) = . . . = μ(6, 6) = 1.

Chapter 5
T-Preorders and T-Indistinguishabilities
5.1 Which Is the Aim of This Section?
To characterize the T-Preorder and the T-indistinguishability relations by means

of particular classes of them, and showing ways of constructing T-preorders and
T-indistinguishabilities.
Given a fuzzy relation on X, μ : X × X → [0, 1], with the three properties

1. Reflexivity, μ(x, x) = 1, for all x ∈ X
2. Symmetry, μ(x, y) = μ(y, x), for all (x, y) ∈ X × X
3. T-transitivity, T (μ(x, y), μ(y, z)) ≤ μ(x, z), for all x, y, z in X, and a continuous
t-norm T,
it can be named
• T-Preorders, those μ verifying 1 and 3.
• T-Indistinguishabilities, those μ verifying 1, 2, and 3.
• Similarities, those μ verifying 1 and 2.
A particular class of T-Preorders is given by the known operators (R-implications)
JT (a, b) = Sup{z ∈ [0, 1]; T (z, a) ≤ b},
and the corresponding class of T-indistinguishabilities is given by the operators
ET (a, b) = Min(JT (a, b), JT (b, a)),
shortly written ET = min(JT , JTδ ), with JTδ (a, b) = JT (b, a). It is obvious that
JT (a, a) = 1, and that JT (a, b) = JT (b, a),
as well as that ET (a, a) = 1, and ET (a, b) = ET (b, a). What it is not so obvious is
that relations JT are T-Transitive:
132 5 T-Preorders and T-Indistinguishabilities
T (JT (a, b), JT (b, c)) ≤ JT (a, c) for all a, b, c in [0, 1]
To avoid some difficult technicalities, we will exemplify this general result in the
particular case JW (a, b) = min(1, 1 − a + b):
W (JW (a, b), JW (b, c)) = W (min(1, 1 − a + b), min(1, 1 − b + c))

= max(0, min(0, b − a) + min(1, 1 − b + c))

min(1, 1 − b + c) ≤ (1, 1 − a + c), if a > b
=
max(0, b − a + min(1, 1 − b + c)) ≤ min(1, 1 − a + c), if a ≤ b
≤ JW (a, c).
Then, with the T-transitivity of JT , it is
T (ET (a, b), ET (b, c)) = T (min(JT (a, b), JT (b, a)), min(JT (b, c), JT (c, b)))
≤ T (JT (a, b), J(b, c)) = ET (a, c),
hence, all relations ET are T-transitive. Then,

– All relations JT are T-Preorders, and
– All relations ET are T-Indistinguishabilities
• If {Ri; i ∈ I} is a collection of T-Preorders, Inf i∈I Ri (x, y) = R(x, y), is also a

T-Preorder.
Obviously, R(x, x) = Inf i∈I Ri (x, x) = 1, and
T (R(a, b), R(b, c)) = T (Inf Ri (a, b), Inf Ri (b, c))
i∈I i∈I
≤ Inf T (Ri (a, b), Ri (b, c))
i∈I
≤ Inf Ri (a, c) = R(a, c),
i∈I
since T is a continuous t-norm.

• If {Ei ; i ∈ I} is a collection of T-indistinguishabilities, Inf i∈I Ei (x, y) = E(x, y) is
also a T-indistinguishability.
Obviously,
E(a, a) = Inf Ei (a, a) = 1,
i∈I
and
E(a, b) = Inf Ei (a, b) = Inf Ei (b, a) = E(b, a).
i∈I i∈I
5.1 Which Is the Aim of This Section? 133
Finally,
T (E(a, b), E(b, c)) = T (Inf Ei (a, b), Inf Ei (b, c))
i∈I i∈I
≤ Inf T (Ei (a, b), Ei (b, c))
i∈I
≤ Inf Ei (a, c) = E(a, c)
i∈I
since T is a continuous t-norm.
5.2 The Characterization of T-Preorders
If σ ∈ [0, 1]X , define JTσ (x, y) = JT (σ(x), σ(y)). Obviously JTσ is a T-Preorder in
X × X, and if F ⊂ [0, 1]X it is also
Inf JTσ
σ∈F
a T-Preorder.
Given a fuzzy relation μ : X × X → [0, 1], consider the set T (μ) of its T-states,
that is the fuzzy sets σ : X → [0, 1], such that
T (σ(x), μ(x, y)) ≤ σ(y), for all x, y ∈ X.
As it is known this last inequality is equivalent to
μ(x, y) ≤ JT (σ(x), σ(y)) = JTσ (x, y),
thus,
σ ∈ T (μ) ⇔ μ ≤ JTσ
Hence,
σ ∈ T (μ) ⇔ μ ≤ Inf JTσ ,
σ∈T (μ)
and it is clear that if μ = Inf σ∈T (μ) JTσ , μ is a T-Preorder.

Avoiding some technical difficulties in the proof of the converse statement, let us
state
• μ is a T-Preorder if and only if μ = Inf σ∈T (μ) JTσ ,
a result characterizing the structure of all T-preorders.
For example,
• T = min, all min-preorders are of the form,

σ 1, if σ(x) ≤ σ(y)
μ(x, y) = Inf Jmin = Inf
σ∈T (μ) σ∈T (μ) σ(y), if σ(x) > σ(y)
• T = prodϕ , all prodϕ -preorders are of the form,

σ 1, if σ(x) ≤ σ(y)
μ(x, y) = Inf Jprod = Inf −1 ϕ(y)
σ∈T (μ)
ϕ
σ∈T (μ) ϕ ( ϕ(x) ), if σ(x) > σ(y)
• T = Wϕ , all Wϕ -preorders are of the form

σ
μ(x, y) = Inf JW (x, y) = Inf min(1, 1 − σ(x) + σ(y))
σ∈T (μ) σ∈T (μ)
Remark 5.2.1 As an immediate consequence of all that has been said, for any fam-
ily of fuzzy sets F ⊂ [0, 1]X , the T-preorder Inf σ∈F JTα has the elements in F as
T-states. If μ(x, y) = Inf σ∈F JTσ (x, y), it follows μ ≤ JTσ , ∀σ ∈ F, or, equivalently
σ ∈ T (μ), ∀σ ∈ F, or F ⊂ T (μ). For example, with X = [0, 1] and the two functions
σ1 (x) = x, σ2 (x) = 1 − x, it is
μ(x, y) = min(JTσ1 (x, y), JTσ2 (x, y))
a T-preorder. With T = W ,
μ(x, y) = min(min(1, 1 − σ1 (x) + σ1 (y)), min(1, 1 − σ2 (x) + σ2 (y)))

= min(1, 1 − x + y, 1 + x − y),
is a W-preorder.
5.3 The Characterization of T-Indistinguishabilities
Let us consider, the T-indistinguishabilities
ETσ (x, y) = min(JTσ (x, y), JTσ (y, x)) = min(JT (σ(x), σ(y)), JT (σ(y), σ(x))).
If μ is a fuzzy relation in [0, 1]X×X , consider T (μ). Obviously
μ(x, y) ≤ Inf ETσ (x, y), for all x, y in X.

σ∈T (μ)
5.3 The Characterization of T-Indistinguishabilities 135
In the same vein that in the case of T-Preorders,

• μ is a T-indistinguishability if and only if μ = Inf σ∈T (μ) TTσ .
For example,

1, if σ(x) ≤ σ(y) 1, if σ(y) ≤ σ(x)
• T = min, μ(x, y) = Inf min ,
σ∈T (μ) σ(y), if σ(x) > σ(y) σ(x), if σ(y) > σ(x)

1, if σ(x) = σ(y)
= Inf
σ∈T (μ) min(σ(x), σ(y)), otherwise

1, if σ(x) = σ(y)
= Inf min(σ(x), σ(y)), otherwise
σ∈T (μ)

1, if σ(x) ≤ σ(y) 1, if σ(y) ≤ σ(x)
• T = prod, μ(x · y) = Inf min σ(y) , σ(x)
σ∈T (μ) σ(x) if σ(x) > σ(y)
, σ(y) , if σ(y) > σ(x)

= Inf
1,
if σ(x) = σ(y)
σ(y) σ(x)
σ∈T (μ) min σ(x) , σ(y) , otherwise
⎧
⎨ 1,
if σ(x) = σ(y)
= σ(y) σ(x)
⎩ Inf min σ(x) , σ(y) , otherwise
σ∈T (μ)
• T = W , μ(x, y) = Inf min(min(1, 1 − σ(x) + σ(y), 1 − σ(y) + σ(x)))

σ∈T (μ)
= Inf min(1, 1 − max(σ(x) − σ(y), σ(y) − σ(x)))
σ∈T (μ)
= Inf min(1, 1 − |σ(x) + σ(y)|)
σ∈T (μ)
= Inf (1 − |σ(x) − σ(y)|)
σ∈T (μ)
Remark 5.3.1 Like in the case of T-Preorders, for any family of functions F ⊂
[0, 1]X , the T-indistinguishability Inf σ∈F ETσ has the elements in F as T-states. Notice
that with μ(x, y) = Inf σ∈F ETσ (x, y) = Inf σ∈F min(JTσ (x, y), JTσ (y, x)), it follows
μ(x, y) ≤ JTσ (x, y), for all x, y in X, that is equivalent to σ ∈ T (μ). Hence,
∀σ ∈ F ⇒ σ ∈ T (μ), or F ⊂ T (μ).
With X = [0, 1] and the two functions σ1 (x) = x, σ2 (x) = 1 − x, it is

μ(x, y) = Min(ETσ1 (x, y), ETσ2 (x, y)),
a T-indistinguishability.
• With T = W , is
μ(x, y) = Min(1 − |σ1 (x) − σ1 (y)|, 1 − |σ2 (x) − σ2 (y)|)
= Min(1 − |x − y|, 1 − |y − x|) = 1 − |x − y|.
Notice that with σ1 (x) = x, σ2 (x) = x 2 , results
μ(x, y) = Min(1 − |x − y|, 1 − |x 2 − y2 |),
that is, μ(x, y) = 1 − |x − y|, provided x + y ≤ 1, and μ(x, y) = 1 − |x 2 − y2 | if

x + y > 1.
• With T = prod, results
⎧
⎪ 1, x = y
1, x ≤ y 1, y≤x ⎨
y
μ(x, y) = Min , = x, y < x .
y
x, x>y 1−y ⎪
1−x , y>x ⎩ 1−y
1−x , y > x
• With T = min, results

⎧
⎪ x=y
1, x ≤ y 1, y≤y ⎨ 1,
μ(x, y) = min , = y, x = y ≤ 1/2
y, x > y 1 − y, y > x ⎪
⎩
1 − y, x = y > 1/2
Example 5.3.2 A finite example with X = {1, 2, 3, 4}. Take σ1 (x) = x

4, σ2 (x) =
1 − 4x , and T = W . It is
⎛ ⎞
1 1 1 1
σ1 ⎜ 3/4 1 1 1 ⎟
[JW ]=⎜
⎝ 1/2 3/4 1 1 ⎠
⎟
1/4 1/2 3/4 1
Since, f.e.,

σ1 1 2 σ1 3 1
JW (1, 2) = min 1, 1 − + = 1, JW (3, 1) = min 1, 1 − + = 1/2,
4 4 4 4

σ1 1 σ1 3
JW (4, 1) = min 1, 1 − 1 + = 1/4, JW (3, 4) = min 1, 1 − + 1 = 1, etc.
4 4
It is also ⎛ ⎞
1 3/4 1/2 1/4
σ2 ⎜ 1 1 3/4 1/2 ⎟
⎜
[JW ] = ⎝ ⎟
1/2 1 1 3/4 ⎠
1 1 1 1
since, f.e.,

σ2 3 1 σ2 1 3
JW (1, 2) = min 1, 1 − + = 3/4, JW (3, 1) = min 1, 1 − + = 1/2,
4 2 4 4

σ2 3 σ2 1
JW (4, 1) = min 1, 1 − 0 + = 1, JW (3, 4) = min 1, 1 − + 0 = 3/4, etc.
4 4
Hence,
⎛ ⎞
1 3/4 1/2 1/4
⎜ 3/4 1 3/4 1/2 ⎟
σ1 σ2 ⎜ ⎟
[EW ] = min([JW ], [JW ]) = ⎜ ⎟
⎝ 1/2 3/4 1 3/4 ⎠
1/4 1/2 3/4 1
is a W-indistinguishability on X, with the directed graph
1/4 3/4
1/2
1 1/2 3
3/4
3/4
2
Example 5.3.3 With an intersection μ · σ = T ◦ (μ × σ), (T a continuous t-norm)

define
ET (μ, σ) = Sup (μ · σ).
Obviously, ET (μ, σ) = ET (σ, μ), since T is a commutative operation. To have,
ET (μ, μ) = 1, or Sup T ◦ (μ × μ) = 1,
since T ◦ (μ × μ) ≤ μ, it should be Sup μ = 1, that is, since X is a finite set, μ should

be a normalized fuzzy set, a fuzzy set for which it exists x0 ∈ X such that μ(x0 ) = 1.
From now on let’s consider the set N (X) = {μ ∈ [0, 1]X ; Sup μ = 1}. The
mapping
ET : N (X) × N (X) → [0, 1],
is reflexive and symmetric, that is, ET is a similarity in N (X). Obviously, ET ≤

Emin , for all continuous t-norm T. Since,
T (ET (μ, σ), ET (σ, λ)) = T (Sup T (μ, σ), Sup T (σ, λ))
= Sup T (T (μ, σ), T (σ, λ))
≤ Sup T (μ, λ) = ET (μ, λ),
it results that, in addition, ET is a T ∗ -indistinguishability for all continuous t-norm

T ∗ such that T ∗ ≤ T . In particular, Emin is not only a min-indistinguishability but a
T-indistinguishability for all continuous t-norm T.
For example with X = {1, 2, 3, 4}, and the two fuzzy sets
μ = 1|1 + 0.4|2 + 0.8|3 + 0.7|4, σ = 0.6|1 + 0.5|2 + 0.8|3 + 1|4
it follows
μ · σ = 0.6|1 + T (0.4, 0.5)|2 + T (0.8, 0.8)|3 + 0.7|4,
and
• Emin (μ, σ) = 0.8, since min(0.8, 0.8) = 0.8, min(0.4, 0.5) = 0.4
• Eprod (μ, σ) = 0.64, since prod(0.8, 0.8) = 0.64, prod(0.4, 0.5) = 0.2
• EW (μ, σ) = 0.6, since W (0.8, 0.8) = 0.6, W (0.4, 0.5) = 0
Remark 5.3.4 Provided T = min or T = prodϕ , if ET (μ, σ) > 0, and ET (σ, α) > 0,
it is
0 < T (ET (μ, σ), ET (σ, α)) ≤ ET (μ, α),
and ET (μ, α) > 0. In these cases, ET is said to be strictly transitive.

Nevertheless, in the case in which a fuzzy relation E : X × X → [0, 1] is Wϕ -
transitive, from
E(x, y) > 0 and E(y, z) > 0,
what follows is that, from the equivalence given by the existence of r > 0 such that
E(x, y) > r, E(y, z) > r, it is
Wϕ (r, r) = ϕ−1 (max(0, 2ϕ(r) − 1)) ≤ Wϕ (E(x, y), E(y, z)) ≤ E(x, z).
Hence, to have 0 < E(x, z), it is necessary that 0 < Wϕ (r, r), that is,
0 < max(0, 2ϕ(r) − 1), or r > ϕ−1 (0.5).
For example, if√it is T = W , it should be r > 0.5, and if T = Wϕ with ϕ(x) = x 2 it

should be r > 0.5 = 0.7071. In the case of the fuzzy relation
n
i−1 min(μ(xi ), σ(xi ))
E(μ, σ) = ,
max( ni−1 μ(xi ), ni−1 σ(xi ))
with X = {x1 , . . . , xn }, it results Wϕ -transitive for ϕ(x) = x 2 . Thus, if
0.71 < E(μ, σ), and 0.71 < E(σ, α)

it follows 0 < E(μ, α), since Wϕ (0.71, 0.71) = max(0, 2 × 0.712 − 1) = 0.09.
Chapter 6
Fuzzy Arithmetic
6.1 Introduction
As it was explained before, any operation ∗ : R × R → R, can be extended to one

: [0, 1]R × [0, 1]R → [0, 1]R , by means of the extension principle:
(μ σ)(t) = Sup min(μ(x), σ(y)).

t=x∗y
This extension includes the crisp subsets A ⊂ R, since μA ∈ {0, 1}R ⊂ [0, 1]R .
For example, with A1 = {1, 2, 3}, and A2 = {1, 3, 5}, and only taking into account
the numbers in N ⊂ R, it is
(μA1 ⊕ μA2 )(t) = Sup min(μA1 (x), μA2 (y)), with t, x, y in N,

t=x+y
and + the addition of natural numbers. Since x + y ∈ {2, 4, 5, 3, 5, 7, 8}, it results

μA2 ⊕ μA2 = μ{2,3,4,5,6,7,8} , that is A1 ⊕ A2 = {2, 3, 4, 5, 6, 7, 8}.
In the same vein, if A1 = [a, b], and A2 = [c, d] are intervals of the real line R,
it results
(μA1 ⊕ μA2 )(t) = Sup min(μ[a,b] (x), μ[c,d] (y))

t=x+y
= μ[a+c,b+d] , or [a, b] ⊕ [c, d] = [a + c, b + d].
Analogously, it results
• [a, b] [c, d] = [a − d, b − c]
• [a, b] ⊗ [c, d] = [min(ad, ac,bd, bc),
max(ad,ac, bd,bc)]
• If 0 ∈ ÷ [c, d] = min c , d , c , d , max c , d , c , d
/ [c, d], [a, b] a a b b a a b b
.

142 6 Fuzzy Arithmetic
For example,
• [7, 8] ⊕ [−1, 9] = [6, 17]
• [7, 8] [−1, 9] = [−2, 9]
• [3, 4]
[2, 2] = [6, 8]
• ÷ [1, 2] = [2, 10]
[4, 10]
• 2 ⊕ [7, 8] = [2, 2] ⊕ [7, 8] = [9, 10]
• 2
[7, 8] = [2, 2]
[7, 8] = [min(14, 16, 14, 16), max(14, 16)] = [14, 16]
7 8 7 8 7
÷ 2 = [7, 8]
• [7, 8] ÷ [2, 2] = min 2 , 2 , max 2 , 2 = 2, 4
In short, and accordingly with the ‘preservation of the classical case’, through
the extension principle both the numerical and the interval arithmetics are preserved.
What follows is a yet larger arithmetic with fuzzy concepts.
For all μ ∈ [0, 1]R , and all r ∈ [0, 1], it can be computed that:
• (r ⊕ μ)(t) = (μr ⊕ μ)(t) = Sup min(μr (x), μ(y))
t=x+y

1, μ(y), if x = r μ(y), x = r
= Sup min = Sup
t=x+y 0, μ(y), if x = r t=x+y 0, x = r
= μ(t − r).
Hence ((−1) ⊕ μ)(t) = μ(t − 1), (μ1 ⊕ μ)(t) = μ(t − 1), (μ 0 ⊕ μ)(t) = μ(t).
• (r
μ)(t) = (μr
μ)(t) = Sup min(μr (x), μ(t)) = μ rt , if r = 0. Hence,
t=x·y
1
r
μ (t) = μ(rt), (1
μ)(t) = μ(t), (μ1
μ)(t) = μ(t).
• (μ0
μ)(t) = Sup min(μ0 (x), μ(t)) = 0 = μ0 (t).
t=x·y
1
μ t , if t = 0
• μ1 (t) = Sup min(1, μ(y)) = Supμ(y) =
t= x t= x 0, if t = 0
y y
Example 6.1.1 Given the fuzzy set
compute 2
μ and 2 ⊕ μ.
It is (2
μ)(t) = μ(t/2), and (2 ⊕ μ)(t) = μ(t − 2). Hence,
6.1 Introduction 143
It should be pointed out that, since the cartesian product μ × σ of μ ∈ [0, 1]X and
σ ∈ [0, 1]Y , is defined by
μ × σ(x, y) = min(μ(x), σ(y)), μ × σ ∈ [0, 1]X×Y ,
the operations , with ∗ ∈ {+, −, ·, :}, can also be represented in the form
μ σ(t) = Sup (μ × σ)(x, y),

t=x∗y
that in the finite case are simply,
μ σ(t) = Max (μ × σ)(x, y).

t=x∗y
Example 6.1.2 If X = {1, 2, 3}, and μ = 0.7|1 + 0.9|2 + 1|3, σ = 0.8|1 + 0.9|3,
since X × X = {(1, 1), (1, 2), . . . , (3, 2), (3, 3)}, it results
μ × σ = 0.7|(1, 1) + 0|(1, 2) + 0.7|(1, 3) + 0.8|(2, 1)

+ 0|(2, 2) + 0.9|(2, 3) + 0.8|(3, 1) + 0|(3, 2) + 0.9|(3, 3).
Since t = x + y will take the values 1 + 1 = 2, 1 + 2 = 3, 1 + 3 = 4, 2 + 1 = 3,

2 + 2 = 4, 2 + 3 = 5, 3 + 1 = 4, 3 + 2 = 5,3 + 3 = 6, that is t ∈ {2, 3, 4, 5, 6}, it
results:
μ ⊕ σ(2) = 0.7, μ ⊕ σ(3) = max(0, 0.8) = 0.8, μ ⊕ σ(4) = max(0.7, 0.8, 0) = 0.8,
μ ⊕ σ(5) = max(0.9, 0) = 0.9, μ ⊕ σ(6) = 0.9
Hence,
μ ⊕ σ = 0.7|2 + 0.8|3 + 0.8|4 + 0.9|5 + 0.9|6.
For the product μ

σ, it is t = x · y ∈ {1, 2, 3, 4, 6, 9}, and
μ · σ = 0.7|1 + 0.8|2 + 0.8|3 + 0|4 + 0.9|6 + 0.9|9.

6.2 Fuzzy Numbers
In the usual scientific computation √is rather unusual to consider exact numbers, as in
the case of taking π ≈ 3.1416, or 2 ≈√ 1.4142. In some cases, an interval containing
the number is considered, for example, 2 ∈ [1.412, 1.413]. Fuzzy numbers are just
a “fuzzification” of this last idea. For example, the fuzzy number “approximately 2”
can be taken either as the fuzzy set,
1
0 1 1.9 2 2.1
or as the fuzzy set.

1
0 1 1.5 1.9 2 2.1 2.5 3
That kind of fuzzy sets are known as fuzzy numbers. Their definition is the fol-
lowing.
Definition 6.2.1 A fuzzy set μ ∈ [0, 1]R is a fuzzy number, provided there is a
closed interval [a, b] ∈ R, such that
1. μ(x) = 1, for all x ∈ [a, b]
2. It exists a function L : (−∞, a) → [0, 1], that is continuous and non-decreasing,
such that μ(x) = L(x), for all x ∈ (−∞, a).
3. It exists a function R : (b, +∞) → [0, 1], that is continuous and decreasing, such
that μ(x) = R(x), for all x ∈ (b, +∞).
6.2 Fuzzy Numbers 145
The following are examples of fuzzy numbers.
6.2.1 Operations with Fuzzy Numbers
It can be proven that, if ∗ ∈ {+, −, ×, :}, and if μ, σ are fuzzy numbers, then μ σ
is also a fuzzy number. For example, with the fuzzy number,
⎧
⎨ x − 2, if x ∈ (2, 3)
μ3 (x) = 4 − x, if x ∈ (3, 4)
⎩
0, otherwise,
for which [a, b] = [3, 3] = {3}, L(x) = x − 2, R(x) = 4 − x, it results

⎧ t−4
⎨ 2 , if t ∈ [4, 6]
⎪
(μ3 ⊕ μ3 )(t) = Sup min(μ3 (x), μ3 (y)) = 2 ,
8−t
if t ∈ [6, 8]
t=x+y ⎪
⎩
0, otherwise
since for t < 4, or x+y < 4, and for t > 8, or x+y > 8, it should be (μ3 ⊕μ3 )(t) = 0;
for t = 6, or x + y = 6, it should be (μ3 ⊕ μ3 )(6) = 1, and:
• For t ∈ [4, 6], L is the segment joining (4, 0) and (6, 1), that is

x y 1
x−4
0 = 4 0 1 = −x + 2y + 4, or y =
6 1 1 2
• For t ∈ [6, 8], R is the segment joining (6, 1) and (8, 0), that is

x y 1
8−x
0 = 6 1 1 = x + 2y − 8, or y =
8 0 1 2
Graphically
Notice that the result is also a fuzzy number μ6 , such that the amplitude of it
([8, 4], 8 − 4 = 4) is twice than that of the initial μ3 ([2, 4], 4 − 2 = 2). When
operating with fuzzy numbers, the amplitude grows. If there is uncertainty about the
value 3, the uncertainty about 3 + 3 is twice, and such uncertainty will grow each
time we ‘add’ more numbers.
6.2.2 Operations with Triangular Fuzzy Numbers
Let’s show a systematic way of operating with triangular fuzzy numbers, that is,
those represented by ⎧
⎨ L(x), if x ∈ (a, n)
μn (x) = R(x), if x ∈ (n, b)
⎩
0, otherwise,
with L and R linear functions, and μn (n) = 1. Graphically,

Namely, take
and
Obviously (μn ⊕ μm )(n + m) = 1; if t ≤ m + n − (α + β), (μn ⊕ μm )(t) = 0;

if t ≥ m + n + (α + β), (μn ⊕ μm )(t) = 0. Hence, the problem is reduced to com-
pute the values of μn ⊕ μm in the two intervals (m + n − (α + β), m + n), and
(m + n, m + n + (α + β)). It can be done as follows:
1. If x ∈ (n − α, n), and y ∈ (m − β, m), t = x + y ∈ (n + m − (α + β), n + m), and,
y+m
since μn (x) = 1 + x−n
α , μm (y) = 1 + β , it follows (with μn (x) = μm (y) = z):
x = αz + n − α, y = βz + m − β, and
t = x + y = (α + β)z + n + m − α − β,
giving
t − (n + m) + (α + β)
z= ,
α+β
that is
t − (n + m) + (α + β)
(μn ⊕ μm )(t) = , if t ∈ (m + n − (α + β), m + n).
α+β
2. If x ∈ (n, n + α), and y ∈ (m, m + β), t ∈ (n + m, n + m + (α + β)), and like

in (1), it is z = t−(n+m)+(α+β)
α+β , that is
t − (n + m) − (α + β)
(μn ⊕ μm )(t) = , if t ∈ (n + m, m + n − (α + β)).
α+β
3. Graphically, with μn ⊕ μm = μn+m

For example, μ3 with α = 1, and μ8 with β = 2, give μ3 ⊕ μ8 = μ11 with

α + β = 3, that is the triangle on [11 − 3, 11 + 3] = [8, 14].
the substraction μn μm = μn [(−1)

μm ], with
Let’s take into account
((−1)
μm )(t) = μm −1 t
= μm (−t), that means
with −μm = (−1)

μm . The fuzzy number −μm is obtained by translating μm
symmetrically at the other side of x = 0. For example, with
⎧
⎨ x − 2, if x ∈ (2, 3)
μ3 (x) = 4 − x, if x ∈ (3, 4)
⎩
0, otherwise,
it results
⎧ t+2 ⎫
⎨ 2 ,
⎪ if x ∈ (−2, 0) ⎪
⎬
(μ3 μ3 )(t) = 2−t
2 , if x ∈ (0, 2)
⎪
⎩ ⎪
⎭
0, otherwise.
= μ0 (t), with the interval [a, b] = [−2, +2].
For what concerns the product μn

μm of two triangular fuzzy numbers, the
method is the same of the sum, but remembering that since (μn
μm )(t) =
Sup min(μn (s), μm (y)), the value t will be reached by the product x · y. For example,
t=x·y
with μ3 as below, let’s compute μ3
μ3 . The process is as follows.
1. If t ≤ 4, and x · y = t, either x ≤ 2, or y ≤ 2. Hence, (μ3

μ3 )(t) = 0.
Analogously, if t ≥ 16, (μ3
μ3 )(t) = 0.
Obviously, (μ3
μ3 )(9) = 1.
2. Hence, L will be defined in [4, 9], and R in [9, 16].
3. If x, y ∈ [2, 3], it is x · y ∈ [4, 9], hence,
L(t) = Sup min(x − 2, y − 2) = α ⇒ α = x − 2, α = y − 2

t=x·y
√
⇒ x = α + 2, y = α + 2 ⇒ t = x · y = (α + 2)2 ⇒ α = t − 2.
4. If x, y ∈ [3, 4], it is x · y ∈ [9, 16], hence,
R(t) = Sup min(4 − x, 4 − y) = α ⇒ α = 4 − x, α = 4 − y

t=x·y
√
⇒ x = 4 − α, y = 4 − α ⇒ t = (4 − α)2 ⇒ α = 4 − t.
Finally, ⎧√ ⎫
⎨ t −√2, if t ∈ [4, 9] ⎬
(μ3
μ3 )(t) = 4 − t, if t ∈ [9, 16] = μ9 (t).
⎩ ⎭
0, otherwise.
Graphically
Remark 6.2.2 It should be pointed out that, in the case of the product of triangular
fuzzy numbers, the result is not a triangular (linear) fuzzy number, since functions L
and R are not linear. In addition the interval [a, b] is not symmetrical. Look that for
μ9 is [4, 16] whose mid point is not 9 but 10.
Let’s now consider the quotient

÷ of fuzzy numbers defined by
÷ μm )(t) = Sup min(μn (x), μm (y)),

(μn
t= xy

μm 1x , if x=0
or μn ÷ μm = μn
μ , with μ
1 1
(x) = , and
m m 0, if x = 0

m = 1.
1 1
μm
÷ μ3 , with μ3 as in Sect. 6.2.2. It will be:
Let’s compute the example μ3
⎧ 1
⎪ 4 − 1y , if y ∈ 4, 3
1
⎨ 1 1
1
μ3 (y) = 1y − 2, if y ∈ 3, 2
⎪
⎩
0, otherwise.
Hence, for μ3
÷ μ3 = μ3
1
μ3 , it will be (μ3
÷ μ3 )(1) = 1, and
α+2
• L, α = 4 − 1
y =x−2⇒t =x·y = 4−α ⇒α= 4t−2
t+1
• R, α = 4 − x = 1
y −2⇒t =x·y = 4−α
α+2 ⇒α= 4−2t
t+1 .
That is ⎧ 4t−2 ⎫
⎨ t+1 ,
⎪ if t ∈ [ 21 , 1] ⎪
⎬
(μ3
÷ μ3 )(t) =
t+1 ,
4−2t
if t ∈ [1, 2] = μ1 (t)
⎪
⎩ ⎪
⎭
0, otherwise
Graphically,
Remark 6.2.3 Obviously, like in the product’s case, the result is not a linear-
triangular fuzzy number, and the basic interval [a, b] is not symmetrical respect
to the point mn .
6.2.3 Note
As it is easy to see, if μ, σ ∈ [0, 1]R , and ∗ ∈ {+, −, ·, :}, it follows the relation
between the corresponding α-cuts:
(μ σ)α = μα σα ,
with the particular supposition that when ∗ = :, it should be required that 0 ∈

/ μα for
all α ∈ (0, 1]. Notice that it also follows

μσ = α ∧ (μ ∗ σ)α = α ∧ [μα σα ].
α∈[0,1] α∈[0,1]
It is for this equality that the before hand computations were made. For example,
with
⎧ x+1 ⎧ x−1
⎨ 2 ,
⎪ if t ∈ (−1, 1) ⎨ 2 ,
⎪ if t ∈ (1, 3]
(μ1 )(x) = 3−x
2 , if t ∈ (1, 3) , (μ3 )(x) = 5−x
2 , if t ∈ (3, 5)
⎪
⎩ ⎪
⎩
0, otherwise 0, otherwise
the corresponding α-cuts are
(μ1 )α = [2α − 1, 3 − 2α] , (μ3 )α = [2α + 1, 5 − 2α],
with which
[μ1 ⊕ μ3 ]α = [4α, 8 − 4α], for α ∈ (0, 1].
Hence,
⎧x
⎨ 4,
⎪ if t ∈ (0, 4]
(μ1 ⊕ μ3 )(t) = 4 ,
8−x
if t ∈ (4, 8] = μ4 (t)
⎪
⎩
0, otherwise
6.3 A Note on the Lattice of Fuzzy Numbers
As it is well known, (R, min, max) is a distributive lattice that come from the totally
ordered set (R, ≤). The order ≤ is definable from the lattice operations min, max by
a ≤ b ⇔ a = min(a, b) ⇔ b = max(a, b).
In addition, with a = 1 − a, it is min(a, b) = (max(a , b )) , and max(a, b) =

(min(a , b )) . Let’s extend these operations to the set R∗ of all fuzzy numbers, by
(μ
∧ σ)(t) = Sup (μ × σ)(x, y), and (μ
∨ σ)(t) = Sup (μ × σ)(x, y).
t=min(x,y) t=max(x,y)
Without the proof, let’s state:
Theorem 6.3.1 (R∗ ,

∧,
∨ ) is a distributive lattice, that means:
• ∧ σ, μ
μ ∗
∨σ ∈ R
• μ
∧σ = σ ∧ μ, μ
∧μ = μ
• μ
∨σ = σ ∨ μ, μ
∨μ = μ
• ∧ (μ
μ ∨ σ) = μ, μ ∧ σ) = μ
∨ (μ
• ∧ (σ
μ ∧ σ)
∨ λ) = (μ ∧ λ)
∨ (μ
• μ ∧ λ) = (μ
∨ (σ ∧ (μ
∨ σ) ∨ λ),
for all μ, σ, λ in R∗ .
Hence, R∗ can be endowed with the partial order given by
μ ≤∗ σ ⇔ μ
∧σ = μ ⇔ μ
∨ σ = σ.
6.3.1 Example
With X = {1, 2, 3}, and μ = 0.8|1 + 0.7|2 + 1|3, σ = 0.9|1 + 1|2 + 0.6|3, compute
∧ σ, μ
μ ∨ σ.
Since, t = min(x, y) and t = max(x, y) belong to {1, 2, 3}, it results:
∧ σ)(t) =
• (μ Max [min(μ(x), σ(y))], and:
t=min(x,y)
– (μ
∧ σ)(1) = max(min(μ(1), σ(1)), min(μ(1), σ(2)), min(μ(2), σ(1)),
min(μ(1), σ(3)), min(μ(3), σ(1)) = max(min(0.8, 0.9), min(0.8, 1),
min(0.7, 0.9), min(0.8, 0.6), min(1, 0.9)) = max(0.8, 0.8, 0.7, 0.6, 0.9) = 0.9
∧ σ)(2) = max(min(μ(2), σ(3)), min(μ(3), σ(2)))
– (μ
= max(min(0.7, 0.6), min(1, 1)) = max(0.6, 1) = 1
∧ σ)(3) = min(μ(3), σ(3)) = min(1, 0.6) = 0.6
– (μ
that is μ
∧ σ = 0.9|1 + 1|2 + 0.6|3.
This fuzzy set is different from μ · σ = 0.8|1 + 0.7|2 + 0.6|3, with the t-norm min.
∨ σ)(t) = Max [min(μ(x), σ(y))], and:
• (μ
t=max(x,y)
– (μ
∨ σ)(1) = min(μ(1), σ(1)) = min(0.8, 0.9) = 0.8
∨ σ)(2) = max(min(μ(1), σ(2)), min(μ(2), σ(1)))
– (μ
= max(min(0.8, 1), min(0.7, 0.9)) = max(0.8, 0.7) = 0.8
6.3 A Note on the Lattice of Fuzzy Numbers 153
– (μ
∨ σ)(3) = max(min(μ(1), σ(3)), min(μ(3), σ(1)), min(μ(2), σ(3)),
min(μ(3), σ(2)), min(μ(3), σ(3))) = max(min(0.8, 0.6), min(1, 0.9),
min(0.7, 0.6), min(1, 1), min(1, 0.6)) = 1,
that is μ
∨ σ = 0.8|1 + 0.8|2 + 1|3.
This fuzzy set is different from μ + σ = 0.9|1 + 1|2 + 1|3, with + the t-conorm
max.
Remark 6.3.2 It is easy to check that, although it is not μ ≤ σ pointwise, it is

∧ σ ≤∗ μ
μ ∨ σ.
Remark 6.3.3 Since t = min(x, y) means
• t = x, if x ≤ y,
• t = y, if y ≤ x,
it is immediate that
(μ
∧ σ)(t) = max[Sup min(μ(x), σ(t)), Sup min(μ(t), σ(y))],
t≤x t≤y
∧ σ, given μ and σ. Analogously, since t =

a formula that facilitates to obtain μ
max(x, y) means
• t = x, if y ≤ x,
• t = y, if y ≥ x,
there is a similar formula for μ
∨ σ.
For example, in the case of μ and σ given in the figure
it results:
1. If t ≤ a1 , (μ
∧ σ)(t) = 0
2. If a2 ≤ t, (μ ∧ σ)(t) = 0
3. If a1 ≤ t ≤ b1 , (μ
∧ σ)(t) = μ(t)
4. If b1 ≤ t ≤ c, (μ ∧ σ)(t) = σ(t)
5. If c ≤ t ≤ a2 , (μ
∧ σ)(t) = μ(t).
∧ σ = μ, and μ ≤∗ σ, although it is not μ ≤ σ.
Hence, μ
6.4 A Note on Fuzzy Quantifiers
In classical logic, only two quantifiers are considered. The universal quantifier ∀ (for
all), and ∃ (it exists, or for some), the existential quantifier, with the addition of ∃!
(exists only one). For example, given a sequence of real numbers (an ), it is said that
the real number a is its limit, when
∀ > 0, ∃k ∈ N, ∀n ∈ N : [(n > k) → (|an − a| < )].
Analogously, a function f : [a, b] → R is bounded, when
∃M ∈ R+ , ∀x ∈ [a, b] : [|f (x)| < M].
The importance of these two quantifiers to clearly write mathematical expressions

does not need to be stressed. Nevertheless, both in arithmetic computing and in natural
language more quantifiers are needed and used.
Examples of arithmetical quantifiers are the percentages. For example

• The 85 % of the employees are married
• Between the 40 and the 70 % of the employees are single.
For example, if it is known that
The 35% of the employees are married
The 25% of the married employees are young
What can be said on the employees that are young?
A question which answer is, obviously, 35 × 25 = 875, that is, at least the 8.75 %
of the employees are young.
Another example is
Between 15 and 25 employees are married
Between 5 and 10 married employees are young
What can be said on the employees that are young?
in which what matters is the length of the two intervals [15, 25] (l = 10), and
[5, 10] (l = 5), that give at least, the interval [5, 10 + (10 − 5)] = [5, 15] of
employees that are young.
In natural language imprecise quantifiers like ‘about five’, ‘about half’, ‘most’,
etc., appear and can be represented by means of fuzzy numbers. For example,
6.4 A Note on Fuzzy Quantifiers 155
These fuzzy quantifiers are of two main types

• Absolute quantifiers, when are fuzzy numbers in R (independent of the cardinality
of the universe of discourse)
• Relative quantifiers, when are fuzzy numbers in [0, 1] (dependent of the cardinality
of the universe of discourse)
Of course, the “interval” arithmetical quantifiers belong to the class of absolute
(fuzzy) quantifiers as a (crisp) particular case, and the “percentage” arithmetical
quantifiers belong to the class of relative (fuzzy) quantifiers as a (crisp) particular
case.
The important problem is to compute the degree of validity, or truth, of statements
with fuzzy quantifiers, like for example:
• In a given class, there are about three students whose fluency in English is low,
that firstly should be represented in fuzzy terms. To do that, call μA3 the fuzzy quan-
tifiers, and μLF the degree of low English-fluency, and rewrite the above statement
as
• There are μA3 , μLF ’s.
Then, the degree of validity can be taken as
t = μA3 (|μLF |), with |μLF | the cardinality of μF .
For example, with

and the scores of LF of the (supposedly) five students {1, 2, 3, 4, 5} in the class,
given by
μLF = 0|1 + 0|2 + 0.75|3 + 1|4 + 0.5|5,
it is |μL | = 1 + 0.75 + 0.5 = 2.25, and
t = μA3 (2.25) = 2.25 − 2 = 0.25,
since the equation of the line joining the points (2, 0) and (3, 1), in the figure of μA3 ,
is y = x − 2.
6.4.1 Quantified Fuzzy Statements
Another, more general, kind of quantified fuzzy statements, is
There are Q ∈ X, such that “F1 (x)is P1 ”, . . . , “Fn (x) is Pn ”,
with X the universe of discourse, Fi : X → Fi (X) ⊂ R, 1 ≤ i ≤ n, and Pi a

predicate in Fi (X), 1 ≤ i ≤ n. For example,
There are about 6 employees in the company that are young and whose
computer skills are high.
where X = {x1 , . . . , xn } is the set of employees, Q = about 6, F1 = age, and F2 =

computer skills. These statement can be compressed to the form
There are QH1 ’s H2 ’s ≡ There are Q(H1 and H2 ),
with H1 (x) = μP1 (H1 (x)), H2 (x) = μP2 (H2 (x)), for all x ∈ X, and that correspond
with the rewritting:
There are about 6 employees that are young and with high computer skills,
n
of the given statement. Finally, with Z = |H1 ∩ H2 | = i=1 min(μP1 (F1 (xi )),
μP2 (F2 (xi ))), and Q(Z) = Q(|H1 ∩ H2 |), results the more compressed form
Z is Q
6.4 A Note on Fuzzy Quantifiers 157
corresponding to:
The number of employees that are young and with high computer skills, is
about 6,
and that gives the truth-value t = Q(Z)

The compression of fuzzy quantified statements is essential for its representation
and for computing its truth-value.
Let’s consider the case with relative quantifiers. The most simple case is
• Among the x ∈ X, Q are such that “F(x) is P”,
for example,
• Among the company’s employees almost all are young.
This statement can be compressed to
• Q are H’s, with H(x) = μP (F(x)),
corresponding to
• Almost all company’s employees are young.
Finally, the last compression can be done in the form:
Z is Q,
|H|
with Z = |X| , that gives the truth value t = Q(Z).
Example 6.4.1 (A problem of inference) Given n statements of the compressed form

“Zi is Qi ”, 1 ≤ i ≤ n with Qi absolute or relative quantifiers, which statement of the
form “Zi is Q” can be inferred?
For example,
There are about 10 workers in the establishment
About half of the establishment workers are women
We search Q such that: There are Q women in the stablishment
Now, it is Q1 = about 10, Q2 = about half, X1 = set of workers, X2 = set of

women ⊂ X1 . Then the syllogism can be stated by
Z1 is Q1
Z2 is Q2
where Z1 = |X1 |, Z2 = |X|X 1 ∪X2 |

1|
= |X2|
|X1 | . Hence the conclusion is “Z is Q”,
with Z = |X2 |, and it rests to compute Q.
A rule often used to obtain Q is the following. If its exists f : Rn → R such that
Z = f (Z1 , . . . , Zn ), take Q = f (Q1 , . . . , Qn ). In the example, it is Z1 ·Z2 = |X2 | = Z,
hence Q = Q1 · Q2 , and the inference is
Z1 · Z2 is Q1 · Q2 , with μQ1 · μQ2 = min(μQ1 , μQ2 ).
As a final example,
Most of the workers are young
About half of the young workers are women
Most × About half workers are young women.
Chapter 7
Fuzzy Measures
7.1 Introduction
Fuzzy sets not only appear by representing imprecise predicates, but also partial or
incomplete information. This is the case, for example, of a function X taking values
in [0, 10], of which it is only known that
5 X 7, 8 X , and X 2.
This information can be represented by means of a fuzzy set μ X ∈ [0, 10] , by

taking:
• μ X (x) = 1 ⇔ X takes the value x
• μ X (x) = 0 ⇔ X don’t takes the value x
• μ X (x) = a · x + b ⇔ It is unknown if X can take the value x,
and the above information resumed by:
⎧
⎪ 0, if x ∈ [0, 2] ∪ [8, 10]
⎪
⎪
⎨ x−23 , if x ∈ [2, 5]
μ X (x) = 1, x ∈ [5, 7]
⎪
⎪ if
⎪
⎩ 8 − x, if x ∈ [7, 8]
In these cases, the questions of the uncertainty concerning the questions,

• X ∈ A for A ⊂ [0, 10]
• X ∈ A, for μ A ∈ [0, 10] X ,
r
arise and suggests the problem of its measuring. In this section, several “measures”
of that kind of uncertainty, including probabilities when applicable, are introduced
and studied.
The area of a polygon is a measure of its extensional size, and the length of a
segment is a measure of its longitudinal size. Everybody knows several examples of
160 7 Fuzzy Measures
measures, as it is, for example, the number of apples in a basket. But, how can the
concept of measure be formalized?
7.2 The Concept of a Measure
Given a set X , and provided

• F ⊂ [0, 1] X is a family of fuzzy subsets of X , such that μ0 , μ1 ∈ F, and that is
a preorder in F translating a qualitative binary relation between the elements in F,
• (L , ) is a preordered set with first element 0,
we will say that m : F → L is a (L , )-measure, whenever
1. m(μ0 ) = 0
2. If μ σ, then m(μ) m(σ).
Example 7.2.1 Take F = [0, 1] X , (L , ) = ([0, 1], ) the unit interval with the
partial linear order of the real line, and the qualitative relation “μ is less fuzzy than
σ”, translated by the so-called sharpened order S (=) defined by

μ(x) σ(y), if σ(x) 1/2
μ S σ ⇔
σ(x) μ(x), if σ(x) > 1/2,
that is a reflexive, transitive and antisymmetric, crisp relation. The fuzzy set μ0.5 is
the highest one, and all crisp sets μ ∈ {0, 1} X are minimals in (F, S ). Any mapping,
m : [0, 1] X → [0, 1] such that
• If μ is crisp, then m(μ) = 0
• m(μ0.5 ) = 1
• If μ S σ, then m(μ) m(σ),
is a ([0, 1], )-measure since m(μ0 ) = 0 because of μ0 ∈ {0, 1} X . These measures
are called measures of fuzziness, or fuzzy entropies.
If X = {x1 , . . . , xn } is finite, the following mappings are examples of fuzzy
entropies:
1. m(μ) = 1 − 2 max |μ(xi ) − 21 |
1i n
2. m(μ) = 2 max μ(xi ) · (1 − μ(xi ))
1i n
n
3. m(μ) = i=1 σ(μ(xi )), with σ(x) = x ln x − (1 − x) ln(1 − x) (logarithmic
entropy).
1 n 1, if μ(x) > 0.5
4. m(μ) = 2n 1=1 |μ(xi ) − μCμ (xi )|, with μCμ (x) = , the
0, if μ(x) 0.5
closest crisp set to μ (linear index of fuzziness)

n
5. m(μ) = 2n 1
i=1 (μ(x i ) − μCμ (x i )) (quadratic index of fuzziness).
2
7.2 The Concept of a Measure 161
Remark 7.2.2 1. For some specific problems, measures of fuzziness are selected
verifying the additional property of symmetry:
• For some negation N , it is m(μ) = m(N ◦ μ) = m(μ ).
For example, with N = 1 − id, measures 1, 2, 3, 4 do verify this property of
symmetry m(μ) = m(1 − μ).
2. With each measure of fuzziness m, it can be defined a measure of booleanity

1 − m, that, obviously verifies:
• the value of 1 − m is 1 for the crip sets.
• the value of 1 − m is 0 for μ1/2 .
• 1 − m is non-increasing with respect to the order S .
Example 7.2.3 Take F ∈ [0, 1] X , with the partial pointwise order ‘μ σ ⇔ μ(x)
σ(x), for all x ∈ X ’ (and such that μ0 , μ1 ∈ F). A mapping m : [0, 1] X → [0, 1] is
a fuzzy measure provided m verifies:
1. m(μ0 ) = 0
2. m(μ1 ) = 1
3. If μ σ, then m(μ) m(σ).
When F ∈ {0, 1} X ≈ P(X ), a fuzzy measure is defined by

1. m(∅) = 0
2. m(X ) = 1
3. If A ⊂ B, then m(A) m(B).

For example, if X is a finite set {x1 , . . . , xn }, and |μ| = xi ∈X μ(xi ), the crisp
cardinality of μ, then the function m(μ) = |μ| n , is a fuzzy measure.
Remember, that since μ · σ = T ◦ (μ × σ), μ + σ = S ◦ (μ × σ), it is
μ · σ μ, μ · σ σ, μ μ + σ, σ μ + σ.
Provided μ · σ, μ + σ ∈ F, for all fuzzy measure m, is:
m(μ · σ) m(μ), m(μ · σ) m(σ), m(μ) m(μ + σ), m(σ) m(μ + σ),
and
m(μ · σ) min(m(μ), m(σ)) max(m(μ), m(σ)) m(μ + σ).
In the particular case in which μ, σ ∈ {0, 1} X , it results
m(A ∩ B) min(m(A), m(B)) max(m(A), m(B)) m(A ∪ B),
for all A, B ∈ P(X ). Notice that fuzzy measures can be applied to both [0, 1] X and
P(X ).
7.3 Types of Measures
Given a triplet (X, F, m), where m is a fuzzy measure, if for some negation (N ) and
some union +(S), is
1. When μ σ , then m(μ + σ) m(μ) + m(σ), m is sub-additive
2. When μ σ , then m(μ + σ) m(μ) + m(σ), m is super-additive,
and when m is both sub-additive and super-additive, that is
When μ σ , then m(μ + σ) = m(μ) + m(σ), m is additive.
This classification (once completed with those measures that are neither sub-
additive, nor super-additive), in the case in which μ, σ ∈ {0, 1} X , particularizes
to:
• If A ∩ B = ∅, and m(A ∪ B) m(A) + m(B), m is sub-additive
• If A ∩ B = ∅, and m(A ∪ B) m(A) + m(B), m is super-additive
• If A ∩ B = ∅, and m(A ∪ B) = m(A) + m(B), m is additive.
|A|
Example 7.3.1 The measure m(A) = n , in a finite set X = {x1 , . . . , xn }, is addi-
tive.
7.4 λ-Measures
With F = P(X ), m λ : P(X ) → [0, 1] is called a Sugeno’s λ-measure if, with

λ > −1, it is:
1. m λ (∅) = 0
2. m λ (X ) = 1
3. If A ∩ B = ∅, m λ (A ∪ B) = m λ (A) + m λ (B) + λm λ (A)m λ (B).
Theorem 7.4.1 All mapping m λ is, actually, a fuzzy measure.
Proof What lacks to be proven is that A ⊂ B implies m λ (A) m λ (B). Since

A ⊂ B ⇔ B = A ∪ (AC ∩ B), with A ∩ (AC ∩ B) = ∅, it follows m λ (B) =
m λ (AC ∩ B) + λm λ (A)m λ (AC ∩ B) = m λ (A)[1 + λm λ (AC ∩ B)] + m λ (AC ∩ B).
From λ > −1, it follows 1 + λm λ (AC ∩ B) > 1 − m λ (AC ∩ B), and m λ (B) >
m λ (A)[1 − m λ (AC ∩ B)] = m λ (A) − m λ (A)m λ (AC ∩ B) > m λ (A).
1−m λ (A)
Theorem 7.4.2 m λ (AC ) = 1+λm λ (A) , for all A ∈ P(X ).
Proof From A ∩ AC = ∅, follows m λ (X ) = 1 = m λ (A ∪ AC ) = m λ (A) +

m λ (AC ) + λm λ (A)m λ (AC ) = m λ (AC )[1 + λm λ (A)] + m λ (A).
7.4 λ-Measures 163
Notice, that it is m λ (AC ) = Nλ (m λ (A)), with the Sugeno’s negation Nλ (x) =

1+λx (λ > −1).
1−x
Remarks 7.4.3 • With the t-conorm Sλ (x, y) = x +y+λx y, it follows m λ (A∪ B) =

Sλ (m λ (A), m λ (B)), if A ∩ B = ∅.
• When λ = 0, m 0 is just a probability defined in P(X ) since it results m λ (A ∪ B) =
m λ (A) + m λ (B) when A ∩ B = ∅, that is, m 0 is an additive measure. In addition,
since N0 (x) = 1 − x, it is m 0 (AC ) = 1 − m 0 (A).
• Notice that the axioms required for a λ-measure do not individuate a single one
of them. For example, with λ = 0 what is obtained is the set of all probabilities
on P(X ).
• If λ ∈ (−1, 0), it results
A ∩ B = ∅ : m λ (A ∪ B) m λ (A) + m λ (B),
that is, if λ ∈ (−1, 0), all the corresponding λ-measures are sub-additive. As it
is easy to prove, if λ ∈ (0, +∞), m λ is super-additive.
• As it is well known, if X is a finite set X = {x1 , . . . , xn }, all probabilities m 0 :
P(X ) → [0, 1] are defined by choosing n numbers m 0 ({xi }) ∈ [0, 1], 1 i n,
n
verifying i=1 m 0 ({xi }) = 1, because 1 = m 0 ({x1 , . . . , xn }) = m 0 (x1 ) + . . . +
m 0 (xn ). Something analogous happens with λ-measures when X = {x1 , . . . , xn }.
For example, if X = {x1 , x2 , x3 }, it follows
1 = m λ (X ) = m λ ({x1 , x2 , x3 })
= m λ ({x1 , x2 } ∪ {x3 }) = m λ ({x1 , x2 }) + m λ (x3 ) + λm λ ({x1 , x2 })m λ (x3 )

3
= m λ (xi ) + λ m λ (xi )m λ (x j ) + λ2 m λ (x1 )m λ (x2 )m λ (x3 ).
i=1 1=i< j=3
and for each λ ∈ (−1, +∞), the values m λ (xi ), 1 i 3, are

to be taken veri-
3
fying this equation that, of course, with λ = 0 reduces to 1 = i=1 m 0 (xi ).
In the case that, for example, is m(x1 ) = 0, follows
1 = m λ (x2 )+m λ (x3 )+λ[m λ (x2 )m λ (x3 )] = m λ (x2 )+m λ (x3 )+λm λ (x2 )m λ (x3 )
1−m λ (x2 )
that is: m λ (x3 ) = 1+λm λ (x2 ) . With m λ (x 2 ) = 0.7, results m λ (x 3 ) = 1+0.3λ . With
0.3
λ = 1: m 1 (x3 ) = 1.03 = 0.29. That is: a measure m 1 is defined in X = {x 1 , x 2 , x 3 },

0.3
by m 1 (x1 ) = 0, m 1 (x2 ) = 0.7, m 1 (x3 ) = 0.29. Notice that, since m 1 is not a

3
probability, it is i=1 m(xi ) = 0.99 < 1.
7.5 Measures of Possibility and Necessity
Let it be F ⊂ P(X ) a Boolean algebra of subsets of X . A mapping π : F → [0, 1]

is called a measure of possibility, provided:
• π(∅) = ∅
• π(X ) = 1
• π(A ∪ B) = max(π(A), π(B)), for all A, B in F.
Notice that the last property does not require A ∩ B = ∅. Actually, any of these
mappings are fuzzy measures, since:
A ⊂ B ⇔ A ∪ B = B: π(B) = max(π(A), π(B)) π(A), or π(A) π(B).
From max(π(A), π(B)) π(A) + π(B), it follows π(A ∪ B) π(A) + π(B)

even if A ∩ B = ∅. Hence, possibility measures are sub-additive.
From A∪ AC = X , it is 1 = max(π(A), π(AC )) π(A)+π(AC ), or 1−π(A)
π(AC ).
Obviously,
π(A1 ∪A2 ∪. . .∪An ) = max(π(A1 ), π(A2 ), . . . , π(An )), for all A1 , . . . , An in F.
Hence, if X = {x1 , . . . , xn } is a finite set, to have a possibility measure π, its

values π(xi ) do verify:
1 = π(X ) = max(π(x1 ), π(x2 ), . . . , π(xn )),
forcing that some of the values π(xi ) should equal 1. For example, if X = {x1 , x2 , x3 },
the three values π(x1 ) = 0, π(x2 ) = 0.5, π(x3 ) = 1, define a particular mea-
sure of possibility on P(X ). It is, for example, π({x1 , x2 }) = max(0, 0.5) = 0.5,
π({x1 , x3 }) = max(0, 1) = 1, etc.
Remark 7.5.1 Instead of a family F of crisp sets no problem arises in considering

a family of fuzzy sets. Measures of possibility can be applied to fuzzy sets with the
only changes of A ∩ B = ∅ by μ σ , and A ∩ B = ∅ by μ σ . The only caution
is to use the connectives min, max to preserve distributivity.
Theorem 7.5.2 For each μ ∈ [0, 1] X such that Sup μ = 1, the mapping πμ : F →
[0, 1] defined by
πμ (A) = Sup min(μ(x), μ A (x)), A ∈ F,
x∈X
is a possibility measure.
Proof πμ (∅) = Sup min(μ(x), 0) = 0. πμ (X ) = Sup min(μ(x), 1) =

x∈X x∈X
Sup μ(x) = 1. Finally, since μ A∪B (x) = max(μ A (x), μ B (x)) for all x ∈ X :
x∈X
7.5 Measures of Possibility and Necessity 165
πμ (A ∪ B) = Sup min(μ(x), μ A∪B (x)) = Sup min(μ(x), max(μ A (x), μ B (x)))

x∈X x∈X
= Sup max(min(μ(x), μ A (x)), min(μ(x), μ B (x)))
x∈X
= max(Sup min(μ(x), μ A (x)), Sup min(μ(x), μ B (x)))
x∈X x∈X
= max(πμ (A), πμ (B)).
It is said that μ is the possibility distribution of πμ , a possibility measure that can

be considered as the one “conditioned” by μ. Thus, each fuzzy set μ ∈ [0, 1] X such
that Sup μ = 1 (μ is never self-contradictory) induces the possibility measure πμ .
Provided X = {x1 , . . . , xn } is a finite set, it results
πμ (A) = Max min(μ(xi ), μ A (xi )).

1i n
Hence, all non self-contradictory fuzzy sets, those μ such that Sup μ = 1, can be
viewed as possibility distributions.
Examples 7.5.3 1. If X = {x1 , x2 , x3 } and μ = 1|x1 + 0.6|x2 + 0.8|x3 , follows

πμ (X ) = Max(min(1, μ A (x1 )), min(0.6, μ A (x2 )), min(0.8, μ A (x3 )) that, with
A = {x2 , x3 } gives πμ (A) = max(0, 0.6, 0.8) = 0.8, and with A =
{x1 , x3 } gives πμ (A) = max(min(1, 1), min(0.6, 0), min(0.8, 1)) = max
(1, 0, 0.8) = 1.
2. If X = [0, 10], A = [5, 8], with μ in the next figure, it is πμ

([5, 8]) =
0, x ∈ [0, 5] ∪ [8, 10]
Sup min(μ(x), μ[5,18] (x)) = Sup = 1.
x∈[0,10] x∈[0,10] μ(x), x ∈ [5, 18]
A mapping N : F → [0, 1] is a measure of necessity provided

• N (∅) = 0
• N (X ) = 1
• N (A ∩ B) = min(N (A), N (B)), for all A, B in F.
Since A ⊂ B ⇔ A ∩ B = A, it is N (A) = min(N (A), N (B)) N (B).

Hence, all these mappings are actually fuzzy measures. From min(N (A), N (B))
π(A)+π(B), it follows N (A∩ B) π(A)+π(B), Then 0 = N (∅) = N (A∩ AC ) =
min(N (A), N (AC )), that implies:
Either N (A) = 0, or N (AC ) = 0, and N (A) + N (AC ) 1, or

N (AC ) 1 − N (A).
Obviously, if A1 , A2 , . . . , An ∈ F : N (A1 ∩ A2 ∩ . . . ∩ An ) = min(N (A1 ),

N (A2 ), . . . , N (An )). Then if X = {x1 , . . . , xn }, to define a necessity measure it is
needed to take N (x1 ), N (x2 ), . . . , N (x3 ) such that: 0 = min(N (x1 ), N (x2 ), . . . ,
N (xn )), an equality that forces some of the values N (xi ) to be 0.
Theorem 7.5.4 Given a possibility measure π : F → [0, 1], the function Nπ (A) =
1 − π(AC ), for all A ∈ F, is a necessity measure.
Proof Nπ (∅) = 1 − π(X ) = 0. Nπ (X ) = 1 − π(∅) = 1. Nπ (A ∩ B) =

1 − π(AC ∪ B C ) = 1 − max(π(AC ), π(B C )) = min(1 − π(AC ), 1 − π(B C )) =
min(Nπ (A), Nπ (B)).
Theorem 7.5.5 Given a necessity measure N : F → [0, 1], the function π N (A) =
1 − N (AC ), for all A ∈ F is a possibility measure.
Proof π N (∅) = 1 − N (X ) = 0. π N (X ) = 1 − N (∅) = 1. π N (A ∪ B) = 1 −

N (AC ∩ B C ) = 1 − min(N (AC ), N (B C )) = max(π N (A), π N (B)).
The pairs (π, Nπ ) and (N , π N ) are called dual-pairs of possibility/necessity mea-

sures. Notice that with them it can be read:
N ecessit y o f A = “not the possibilit y o f not A”

Possibilit y o f A = “not the necessit y o f not A”.
Remark 7.5.6 If π = πμ , the corresponding Nπμ is given by
Nπμ (A) = 1−πμ (A) = 1−Sup min(μ(x), μ A (x)) = I n f max(1−μ(x), μ AC (x)),

x∈X x∈X
that, in the finite case X = {x1 , . . . , xn }, is
Nπμ (A) = min max(1 − μ(xi ), μ AC (xi )).

1i n
Theorem 7.5.7 For all dual pair (π, N ) is:

1. If N (A) > 0, then π(A) = 1
2. If π(A) < 1, then N (A) = 0
7.5 Measures of Possibility and Necessity 167
Proof
1 = max(π(A), π(AC )), and 0 = min(N (A), N (AC )),
follows that if N (A) > 0, then N (AC ) = 0, and π(A) = 1 − N (AC ) = 1. If

π(A) < 1, then π(AC ) = 1 and N (A) = 1 − π(AC ) = 0
Remark 7.5.8 Although the proof will not be presented, let’s show the following
important notice. In the case X is finite, for any possibility measure π it exists a (non
unique!) fuzzy set μ ∈ [0, 1] X with Sup μ = 1 such that π = πμ . In the finite case,
all possibility measures come from possibility distributions.
7.6 Examples
Example 7.6.1 On the age of a person p it is only available the incomplete informa-
tion given by
1. 37 Age( p) 41
2. It is neither Age( p) 32, nor Age( p) 43.
What can be said about the possibility and the necessity of “Age( p) 42”,
“Age( p) 40”, and “Age( p) 33”?
Solution
The available incomplete information can be represented by the following possi-
bility distribution μ:
Hence,
• πμ (Age( p) 42) = πμ ([42, 100]) = Sup min(μ(x), μ[42,100] (x)) =
x∈[0,100]
Sup μ(x) = μ(42) : ∈ (0, 1). Hence Nπμ (Age( p) 42) = 0.
x∈[42,100]
The value μ(42) can be computed as follows. The segment between (41, 1) and
(43, 0), verifies

x y 1
43 − x
0 = 41 1 1 = x + 2y − 43 ⇒ y = ,
43 0 1 2
hence, μ(42) = 43−42

2 = 0.5. The possibility of “Age( p) 42” is 0.5, and its
necessity is 0.
• πμ (Age( p) 40) = πμ ([0, 40]) = μ(40) = 1, and Nπμ (Age( p) 40) =
Nπμ ([0, 40]) = 1 − πμ ((40, 100]) = 1 − Sup μ(x) = 1 − μ(40) = 0.
x∈(40,100]
The possibility of “Age( p) 42” is 1, and its necessity is 0.
• πμ (Age( p) 33) = πμ ([33, 100]) = 1, and Nπμ (Age( p) 33) = Nπμ
([33, 100]) = 1 − πμ ([0, 33)) = 1 − μ(33). This value can be computed by:

x y 1
x − 32 1
0 = 32 0 1 = −x + 5y + 32 ⇒ y = , and μ(33) = .
37 1 1 5 5
That is, Nπμ (Age( p) 33) = 1 − 1

5 = 45 . The possibility of “Age( p) 33” is
1, and its necessity is 0.8.
Example 7.6.2 It is only known that a function F : X → [0, 1] doesn’t take any
value below 1/4 but takes values above 3/4. What can be said on the possibility and
necessity of the imprecise statements “F is small”, and “F is not small”?
Solution
The available but incomplete information can be translated by the possibility
distribution μ
7.6 Examples 169
Let’s take μsmall (x) = 1 − x, μnot small (x) = x. It is:
• πμ (F is small) = Sup min(μ(x), μsmall (x)) = Sup min(μ(x), 1 − x) =

x∈[0,1] x∈[0,1]
1/2 ⇒ Nπμ (F is small) = 1 − πμ (F is not small)
• πμ (F is not small) = Sup min(μ(x), μnot small (x)) = Sup min(μ(x), x) =
x∈[0,1] x∈[0,1]
1 ⇒ Nπμ (F is small) = 1 − πμ (F is small).
Hence:
– πμ (F is small) = 0.5, and Nπμ (F is small) = 1 − 1 = 0.
– πμ (F is not small) = 1, and Nπμ (F is not small) = 1 − 0.5 = 0.5.
Remark 7.6.3 Notice that this example is, like the following, not with questions
related to precise or crisp sets, but to impreciseness (fuzzy sets). Although Possibility
Theory is introduced with crisp sets, it is also applicable to fuzzy sets within the theory
given by the triplet (min, max, 1 − id).
Example 7.6.4 John is a member of a community where the predicate Y = young

is used following
⎧
⎨ 1, if x 30
μY (x) = 0, if x > 40
⎩ 40−x
10 , if 30 < x 40
Find the possibility and the necessity of the statement “John is around 35 years
old”.
Solution. The graphics of μY and μ P , with P = Ar ound 35, are
Hence, πμ (John is around 35 years old) = Sup x∈[0,100] min(μY (x), μ P (x)) =
2
3 , since it does correspond with the intersection of μY and μ P , that is, of the straight
10 , y = 5 . It results x = 3 and y = 3 . Since
lines respectively given by y = 40−x x−3 100 2
πμ (John is around 35 years old) < 1, it results Nπμ (John is around 35 yearsold)
= 0.
7.7 Probability, Possibility and Necessity
In the case where F is a Boolean sub-algebra of P(X ), also probabilities p : F →

[0, 1] can be taken into account. Once a dual-pair (N , π) is given, it appears the
problem of which probabilities are consistent with (N , π).
The property N (A) + N (AC ) 1, equivalent to N (A) 1 − N (AC ) = π(A),
shows that for all A ∈ F it is
N (A) π(A),
provided the pair (N , π) is dual. A probability p is said consistent with the dual pair
(N , π) if
N (A) p(A) π(A), for all A ∈ F.
In this hypothesis, if N (A) > 0, it is also p(A) > 0, and π(A) > 0, and if
π(A) = 0 it is N (A) = p(A) = 0. That is,
• If something is just a little bit necessary, it is probable and possible.

• If something is not possible at all, it is neither necessary nor probable.
Analogously, if p(A) > 0 it is π(A) > 0 although it could be N (A) = 0.
That is,
• If something is just a little bit probable, it is possible, but not necessarily necessary.
Notice that from N (A) p(A) π(A), or equivalently from 1 − π(AC )

p(A) π(A), follows 1 − π(A) p(AC ) π(AC ).
Example 7.7.1 Let is X = {1, 2, 3}, and μ = 0.7|1 + 1|2 + 0.5|3. Then, with
(A)
πμ (A) = Sup min(μ(i), μ A (i)), and Nμ = 1 − πμ (A), we get:
i∈X
1. πμ (1) = μ(1) = 0.7, πμ (2) = μ(2) = 1, πμ (3) = μ(3) = 0.5.
πμ ({1, 2}) = max(μ(1), μ(2)) = 1, πμ ({1, 3}) = max(0.7, 0.5) = 0.7,
πμ ({2, 3}) = 1, πμ {X } = 1.
2. Nμ (1) = 1 − πμ ({2, 3}) = 1 − 1 = 0, Nμ (2) = 1 − πμ ({1, 3}) = 1 − 0.7 = 0.3,
Nμ (3) = 1 − πμ ({1, 2}) = 0
Nμ ({1, 2}) = 1 − πμ (3) = 0.5, Nμ ({1, 3}) = 1 − πμ (2) = 0, Nμ ({2, 3}) =
1 − πμ (1) = 0.3, Nμ (X ) = 1.
Hence, the consistent probabilities are given by triplets p(1), p(2), p(3) in [0, 1]
such that p(1) + p(2) + p(3) = 1, and verifying:
0 p(1) 0.7, 0.3 p(2) 1, 0 p(3) 0.5.

7.7 Probability, Possibility and Necessity 171
For instance, with p(1) = 0.4, p(2) = 0.5, p(3) = 0.1, we have a consistent
probability, as well as with p(1) = 0.5, p(2) = 0.3, p(3) = 0.2. But the probability
given by p(1) = 0.6, p(2) = 0.2, p(3) = 0.2 is not consistent, because of p(2) <
0.3. With it there is an element whose probability is smaller than its necessity. In the
same vein, the probability given by the triplet p(1) = 0.1, p(2) = 0.3, p(3) = 0.6, is
also non-consistent because one of the probabilities is greater than the corresponding
possibility.
7.8 Probability of Fuzzy Sets
Let’s shortly formalize the classical concept of probability. In a universe X , let

F ⊂ P(X ) be a Boolean algebra of parts of X , that is,
• If A, B ∈ F ⇒ A ∩ B, A ∪ B, AC , B ⊂ ∈ F.
• ∅ ∈ F, X ∈ F.
It is said that p : F → [0, 1] is a probability in (X, F), provided

• p(∅) = 0
• If A ∩ B = 0, then p(A ∪ B) = p(A) + p(B).
Theorem 7.8.1 1. p(AC ) = 1 − p(A), for all A ∈ F.

2. If A ⊂ B, then p(A) p(B) (that is p is a measure)
3. p(X ) = 1.
4. p(A ∪ B) + p(A ∩ B) = p(A) + p(B), for all A, B ∈ F.
Proof Items (1) and (2) just follow from the fact that p is a 0-measure. p(X ) =
p(∅⊂ ) = 1 − p(∅) = 1. Finally, since
A ∪ B = (A ∩ B) ∪ (B − A) ∪ (A − B),
with (A ∩ B) ∩ (B − A) = ∅, (A ∩ B) ∩ (A − B) = ∅, and (B − A) ∩ (A − B) = ∅,
follows
p(A ∪ B) = p(A ∩ B) + p(B − A) + p(A − B).
But, from A = (A − B) ∪ (A ∩ B), and B = (B − A) ∪ (A ∩ B), it also follows

(since the unions are disjunct):
• p(A) = p(A − B) + p(A ∩ B) ⇒ p(A − B) = p(A) − p(A ∩ B)
• p(B) = p(B − A) + p(A ∩ B) ⇒ p(B − B) = p(B) − p(A ∩ B).
Hence, p(A ∪ B) = p(A ∩ B) + p(A) − p(A ∩ B) + p(B) − p(A ∩ B) =
p(A) + p(B) − p(A ∩ B).
Of special importance for the applications are the probabilities defined in the real
line R, with F = B the so-called Borel’s algebra, given by all the unions, comple-
ments and intersections of the open, closed, semi-open, and semi-closed intervals of
R. Then, if A ∈ B, the probability of A is defined by the Lebesgue-Stieltjes integral

p(A) = dP = μ A (x)d x = E(μ A ),
A R
that is, as the mathematical expectation of μ A . Then, if μ ∈ [0, 1]R is Borel-

measurable, it can be analogously defined

p(μ) = E(μ) = μ(d)d x
R
Obviously, p(μ0 ) = E(μ0 ) = 0, p(μ1 ) = E(μ1 ) = 1, and if μ σ follows

p(μ) p(σ). In addition, with μ · σ = min ◦ (μ × σ), μ + σ = max ◦ (μ × σ), it is
p(μ + σ) + p(μ · σ) = p(μ) + p(σ),
that implies: If μ · σ = μ0 , then p(μ + σ) = p(μ) + p(σ).
Example 7.8.2 Which is the probability of the fuzzy set μ (fuzzy event) given by
Solution

10 · p(μ) = [0,10] μd x = [3,7] μd x = [3,4] μ1 d x + [4,6] 1d x + [6,7] μ2 d x =

2 + [3,4] (x − 3)d x + [6,7] (7 − x)d x = 2 + 21 + 21 = 3. Then p(μ) = 10
3
= 0.3.
7.8 Probability of Fuzzy Sets 173
Example 7.8.3 Which is the probability of the fuzzy event μ
Solution
5 6
10· p(μ) = [0,10] μd x = [4,5] μ1 d x + [5,6] μ2 d x = 4 (4−x)d x + 5 (6−x)d x =
− 21 + 23 = 1. Then p(μ) = 0.1.
Example 7.8.4 Which is the probability of the fuzzy event μ:
σ μ
0 1/4 3/4 1
Which is the conditional probability p(σ/μ)?

Solution
3/4 1 3/4
p(μ) = [0,10] μd x = 1/4 μ1 d x + 3/4 1d x = 4 + 1/4 (2x − 2 )d x = 4
1 1 1
+ 1
4 = 21 .
p(σ·μ)
p(σ|μ) = p(μ) = 2 p(σ · μ), with σ · μ = min ◦(σ × μ), given by
1/2 1 7/16
Then p(σ · μ) = 1/4 (4 − x)d x + 1/2 xd x = 7
16 , and p(σ|μ) = 1/2 = 78 .
Chapter 8
An Introduction to Fuzzy Control
8.1 Introduction
Probably one of the most successful developments of fuzzy reasoning, from the
industrial point of view, is the design of fuzzy control systems, also called linguistic
control systems, or simpler, the applications of fuzzy controllers.
A fuzzy control system is based on a set of fuzzy “if-then” rules of behavior
that consider the kind of stimuli from the environment, that the system will receive,
meanwhile at a given time, the values of these stimuli represent the facts, that the
rules have to consider to offer proper actions.
As it has been proved from its origins, fuzzy control should be useful in situations
where (a) There is no acceptable mathematical model for the plant, (b) There are
experienced human operators who can satisfactorily control the plant and provide
qualitative control rules in terms of vague and fuzzy sentences, and (c) In applica-
tions where there is a large uncertainty or unknown variation in plant parameters and
structures. In this cases, fuzzy control can be considered as a model-free approach
and it does not require a mathematical model of the objective plant. It is referred to
as a knowledge-based control approach, and it makes an effective use of all available
information related to the system, from sensors which provide numerical measure-
ments of key variables to human experts who provide linguistic descriptions about
the system and control instructions.
However, currently fuzzy control approach is mainly devoted to model-based
methods. On the one hand, there are cases in which operators cannot precisely tell
their action in a particular situation, or simply, operator’s control may not be always
optimal with respect to some performance objective. In this sense, identification for
obtaining fuzzy models from process data is very important.
On the other hand, although the uncertainty in dynamic systems can be handled
using appropriate control actions (e.g., their sensitivity to external disturbances and
parameter changes can be reduced stabilizing an unstable system, or the systems
dynamic behavior could be modified speeding up a slow system), the application
of control action over a system can potentially destabilize stable plants. Thus,

176 8 An Introduction to Fuzzy Control
stability analysis is a major issue in control design. A model-based fuzzy control

approach allows performing system stability, performance and robustness analysis.
However, this approach also implies the use of mathematically involved and rather
non-transparent techniques to ensure robust performance.
When dealing with the control of nonlinear systems, a conventional approach to
design controllers is the derivation of fuzzy models from given nonlinear system
equations. Dynamic systems are modeled in the state space framework, using a state
transition model, which describes the evolution of the states over time. Fuzzy logic
provides a simple and straightforward way to decompose the task of modeling and
control design into a group of local tasks, which tend to be easy to handle. In the
end, fuzzy logic also provides the mechanism to blend these local tasks together to
deliver the overall model and control design.
Advances in modern control have made available a large number of powerful
design tools, specially in the case of linear control systems. These tools range from
state space optimal control to the robust control paradigms. In the case of Takagi-
Sugeno (TS) fuzzy models, a local linear system description is used for each rule,
which then uses a control methodology to fully take advantage of the advances of
modern control theory. From a theoretical point of view, model-based fuzzy control
can meet established rigor and systematic control synthesis of conventional control
theory in terms of stability analysis, systematic design procedures, incorporation of
performance specifications, robustness and optimality.
A fuzzy controller has the general structure shown in the following figure. Three
main blocks may be distinguished: a data base, a rule base and a processing unit.
Data Rule
Base Base
IN OUT
Inference
Fuzzification Defuzzification
Engine
The data base contains all information needed to specify a particular configuration
of the fuzzy controller. It will have information related to the number, shape and dis-
tribution of the fuzzy sets specifying the meaning of the linguistic terms of each
linguistic variable involved in the application. The data base also contains infor-
mation related to the kind of crisp-fuzzy and fuzzy-crisp conversions, and the type
of operations for numerical calculations of the conjunction or disjunction premises.
Similarly with respect to the availability of operations for the implication, that real-
izes the “if-then” connective of the rules, as well as for the operation that computes
the aggregation of the conclusions of several rules that might be simultaneously
activated by the prevailing conditions of the environment.
The rule base contains the set of rules that will govern the behavior of the controller.
Through the rule base the task of modeling and control design is decomposed into a
group of local tasks, which tend to be easy to handle.
The processing unit cares for the compatibility of data (input and output interfaces)
and for the execution of the rules (inference engine), which is mostly done through
pointwise numerical calculations of implications.
As it has been mentioned in previous chapters, the basic rules of reasoning used
in classical logic are the modus ponens for forward reasoning and modus tollens for
backwards reasoning. The symbolic expression of the modus ponens,
A B
A
B,
means that if the rule “if A then B” is given and the event A is observed, then the
event B should also be observed.
These processes are referred to in the literature as inference. As it has been
presented in this book, in the case of fuzzy logic, a generalization of modus ponens
is used, based on fuzzy sets. Given a universal set X , a fuzzy set A on X is defined by
its membership function μ A : X → [0, 1] and for all x ∈ X , μ A (x) gives the degree
of membership of x to A, or the degree with which x fulfills the concept represented
by A. Fuzzy sets have a semantic role: they represent the way in which a statement
“x is A” is used in a given context.
The generalized modus ponens used in fuzzy control may be given in its simplest
expression as
A B
A
B ,
where A, A∗ , B and B ∗ are fuzzy sets, A and A∗ are defined on a same universe, but
they are not necessarily equal. Similarly for B and B ∗ . The meaning in this case is
the following: given a rule “if A then B” and observing an event A∗ which is similar
to A, an event B ∗ is expected, which should also be similar to B.
To allow more specified situations to define a rule base in a multivariate system,
in the expression A → B, A may stand for a set of conditions that have to be fulfilled
at the same time. The formal representation is a conjunction of fuzzy premises.
For the computation of conjunctions, operations belonging to the class of trian-
gular norms, or simply t-norms, are used. If T : [0, 1]2 → [0, 1] is a t-norm, then it
is non-decreasing, associative, commutative and has 1 as identity. The other needed
operation is the “then” in “if A then B”. In the case of fuzzy logic this operation is
not only an extension of the classical implication and its characterization has been
thoroughly studied.
Example 8.1.1 The following example is presented at a “phenomenological” level

to allow an intuitive understanding of the main issues. In this case a heating system
has to be controlled. The system is based on warm water circulating at constant speed
and passing through well distributed heating panels. The water temperature should
be proportional to the heating demand. The following rules reflect the knowledge of
an experienced operator of the heating system:
R1: IF the external temperature is freezing,

THEN the heating demand is large
R2: IF the external temperature is cold,
THEN the heating demand is average
R3: IF the external temperature is medium,
THEN the heating demand is low
It becomes apparent that the meaning of the predicates freezing, cold and medium
is different if the system is intended for Ottawa, Madrid or Dakar, since they would
not be used in exactly the same way. Similarly, the meaning of a large, average or low
heating demand (in a KW scale) is different for a system meant to heat an office room,
a conference hall for 200 people or a 25 stores office building. The relative ordering
of the terms large, average and low will certainly be the same, and their shapes will
probably be the same, in all three mentioned cases. Considering one instance of the
problem as illustrated in the figure where it is assumed that the external temperature
is 6 ◦ C.
freezing cold medium low av. high
R1
low av. high
R2
R3
-15 6 15
It is fairly obvious that if the external temperature is 6 ◦ C, the first rule does
not apply, since 6 ◦ is not considered to be freezing, but “more cold than medium”.
The degree of satisfaction of the premises or conditions stated in the “if” part of
rules 2 and 3 will affect the strength of the corresponding conclusions. This will be
specified by the “then” operation. Rules 2 and 3 are activated to a certain degree and
give proportional suggestions for action. These have to be combined into one single
action by means of an aggregation operation. From the many aggregation operations
that may be used for this purpose, the pointwise maximum is usually the first choice.
The aggregated result is also shown in the figure.
Since it was assumed that finally the water temperature should be proportional to
the heating demand, the fuzzy set representing the aggregation of activated heating
demands has to be converted into a real value through the deffuzzification. This
process will imply an information loss, since it is analogous to representing a signal
with only one coefficient of its Fourier power spectrum. However, experimental
results have shown that “approximating” a fuzzy set by the abscise of its gravity
center or the abscise of its center of area leads to an adequate control performance.
8.1.1 Note
This chapter is divided in two main parts that on the one hand pretends to put fuzzy
control in the context of the book, and on the other hand introduces the present
situation in the area of fuzzy control.
The first part makes a revision of the fundamentals of approximated reasoning
explained throughout the book. A condensed view of how to apply the presented
theory from a fuzzy control perspective is given.
The second part outlines a glimpse of the state of the art of fuzzy control, where
fuzzy plant models and the procedures to obtain them are introduced. In the area
of fuzzy control there has been a shift from its original motivation of interpretable
fuzzy systems (systems that emulate human control strategy and are easy to use and
understand) towards a much more rigorous analysis where mainstream (nonlinear)
control criteria are considered. The reader must be aware that in order to fully under-
stand this part, he/she should probably refer to more specialized books in the topic
as this section only pretends to present a very general view of the problem.
8.2 Revising Conditional and Implications in Fuzzy Control
The success of fuzzy logic mainly resides in the representation of elementary

statements “x is P” (x ∈ X and P a precise or imprecise predicate or linguistic
label on X ) by a function μ P : X → [0, 1], in the hypothesis that μ P (x) is the
degree up to which “x is P”, or x verifies the property named P. In the same view,
a rule “If x is P, then y is Q” (x ∈ X, y ∈ Y ) is represented by a function of the
variables μ P and μ Q , by
R(μ P , μ Q )(x, y),
a number in [0, 1] once P, Q, x and y are fixed.

Functions R can be or can be not functionally expressible, that is, it can exist or
it cannot exist a numerical function J : [0, 1] × [0, 1] → [0, 1] such that
R(μ P , μ Q )(x, y) = J (μ P (x), μ Q (y))
for all x ∈ X , y ∈ Y . Fuzzy logic works mainly within the positive supposition,
and several families of such functions J exist. Such numerical functions are called
conditional functions, and their diverse types are derived from the linguistic meaning
of the conditional phrase (the rule) “If x is P, then y is Q”, that is, from its use in
the universe X × Y at each particular problem.
Even if it is some repetition of what was earlier presented, let’s summarize the
steps that are necessary.
8.2.1 Inference from Imprecise Rules
If a dynamic systems is considered, with input variables (x1 , . . . , xn ) and output

variable y, whose behavior is known by m imprecise rules Ri :
R1 : If x1 is P11 and x2 is P12 · · · and xn is P1n , then y is Q 1

R2 : If x1 is P21 and x2 is P22 · · · and xn is P2n , then y is Q 2
...
Rm : If x1 is Pm1 and x2 is Pm2 · · · and xn is Pmn , then y is Q m
The question is the following: If we observe that the input variables are in the
“states” “x1 is P1∗ ”, “x2 is P2∗ ”, . . . , “xn is Pn∗ ”, respectively, what can be inferred
for variable y? That is, supposing that it lies in the “state” “y is Q ∗ ”, what is Q ∗ ?
Without loss of generality let us consider only the case n = m = 2:
R1 : If x1 is P11 and x2 is P12 , then y is Q1
R2 : If x1 is P21 and x2 is P22 , then y is Q2
x1 is P1 and x2 is P2
y is Q , what is Q ?
where x1 ∈ X 1 , x2 ∈ X 2 , y ∈ Y , and let us represent it in fuzzy logic by means of

the functions R1∗ and R2∗ . This leads to:
for convenient continuous t-norms T1 , T2 and convenient implication functions J1 ,

J2 . Convenient, in the sense of adequate to the use of the conditional phrases R1 and
R2 relative to the problem under consideration within the given system S.
To find a solution to this problem, the problem of linguistic control, that is a part
of what can be called intelligent systems control, we need to pass throughout several
steps.
First Step: Functions J
Implication functions J : [0, 1] × [0, 1] → [0, 1] are obtained through the interpre-
tation and representation of the rule’s use, that is, from the actual meaning of these
conditional phrases.Thus, different types of protoform exist
a → b ⇔ J (a, b)
and are used as introduced previously in the book. For example:

8.2 Revising Conditional and Implications in Fuzzy Control 181
• S-implications: a → b = a + b and J (a, b) = S(N (a), b) for all a, b in [0, 1],

with some strong negation function N and some continuous t-conorm S.
• R-implications: a → b = a + b and J (a, b) = Sup {z ∈ [0, 1]; T (z, a) b } =
JT (a, b)) for all a, b in [0, 1] and some continuous t-norm T .
• Q-implications: a → b = a + a · b and J belongs to the family J (a, b) =
S(N (a), T (a, b).
• ML-implications: a → b = a · b and J is in the family of functions J (a, b) =
T (a, b).
In fuzzy control mostly ML-implications are considered. Since T is a continu-

ous t-norm, it is T = Min, or T = Prod ϕ = ϕ−1 ◦ Prod ◦ (ϕ × ϕ), or
T = Wϕ = ϕ−1 ◦ W ◦ (ϕ × ϕ), with ϕ an order-automorphism of the unit
interval ([0, 1], ), Prod(x, y) = x · y and W (x, y) = Max(0, x + y − 1). Of
course, T can be also an ordinal-sum, but these t-norms have never been considered
in fuzzy logic. Hence, a ML-implication belongs to the types: JM (a, b) = Min(a, b),
JL (a, b) = Prod ϕ (a, b) and JW (a, b) = Wϕ (a, b). Only in the third type we can
have J (a, b) = 0 with a = 0 and b = 0, since it is Wϕ (a, b) = 0 whenever
ϕ(a) + ϕ(b) 1, and as in fuzzy control it is desirable not only that a = 0 implies
J (a, b) = 0 but also that a = 0 and b = 0 imply J (a, b) = 0, the third type is rarely
used and only JM (a, b) = Min(a, b) (Mamdani implication) and JL (a, b) = a · b
(Larsen implication) are almost always considered. Notice that for all S, and Q
implication functions it is: J (0, b) = 1.
Second Step: Modus Ponens
Rules are used in our problem to infer μ Q ∗ , and this inference requires that when the
states of the input variables x1 , . . . , xn are exactly those appearing in the antecedent
part of one of the m rules, say rule number i, then μ Q ∗ should be the consequent μ Q i
of this rule. That is, each rule should satisfy the meta-rule of Modus Ponens:
This meta-rule is satisfied when there is a continuous t-norm T1 such that
T1 (μ P ∗ (x), R(μ P , μ Q )(x, y)) μ Q ∗ (y),
for all x ∈ X , y ∈ Y . When R(μ P , μ Q )(x, y) = J (μ P (x), μ Q (y)), the last

inequation is
T1 (μ P ∗ (x), J (μ P (x), μ Q (y)) μ Q ∗ (y),
for all x ∈ X , y ∈ Y .
Hence, for each type of implication function J we need to know which T1 allows
the verification of the Modus Ponens inequality:
T1 (a, J (a, b)) b, for all a, b ∈ [0, 1].

For example, with an S-implication, since T1 (a, S(N (a), b)) b implies
(with b = 0) T1 (a, N (a)) = 0, it should be T1 = Wϕ for some automorphism
ϕ of ([0, 1], ).
With R-implications JT , since
T (a, JT (a, b)) = Min(a, b) b,
the same T in JT allows to have that inequality.

With Q-implications, since
T1 (a, S(N (a), T (a, b))) b
also implies T1 (a, N (a)) = 0 for all a ∈ [0, 1], it should be also T1 = Wϕ .
Concerning ML-implications, since
T1 (a, T (a, b)) Min(a, Min(a; b)) = Min(a, b) b,
because both T1 Min and T Min, the Modus Ponens inequality is verified for
all t-norms T1 and, hence, for T1 = Min (the biggest t-norm).
If T1 verifies T1 (a, J (a, b)) b, because of the well-known result that for
left-continuous t-norms T1 , T1 (a, t) b is equivalent to t JT1 (a, b), it results
that the inequality is equivalent to J (a, b) JT1 (a, b). Hence, among the func-
tions J verifying the Modus Ponens inequality with a continuous t-norm T1 , the
R-implication JT1 is the biggest one and, consequently, T1 (a, JT1 (a, b)) is closer to
b than T1 (a, J (a, b)). In particular, it is

1 if a b
J M (a, b) = Min(a, b) JMin (a, b) =
b if a > b
and
1 if a b
JL (a, b) = a · b JMin (a, b) JProd (a, b) = b
a if a > b
(since a · b b).
Third Step: Zadeh’s Compositional Rule of Inference CRI
Once the rule “If x is P, then y is Q” is represented by J (μ P (x), μ Q (y)), and a
continuous t-norm T1 such that J JT1 is known, the inference:
is obtained by the Zadeh’s Compositional Rule of Inference (CRI):
μ Q ∗ (y) = Sup T1 (μ P ∗ (x), J (μ P (x), μ Q (y))), for all y ∈ Y.

x∈X
It should be pointed out that Zadeh’s CRI is not a “result” but a meta-rule. It is a
“directive” allowing to reach a solution to our problem, and it should be noticed
that when P ∗ = P it is not in general Q ∗ = Q. For example, in the case of ML-
implications it is:
μ Q ∗ (y) = SupMin(μ P (x), T (μ P (x), μ Q (y)))

x∈X
Sup T (μ P (x), μ Q (y)) = T (Sup μ P (x), μ Q (y)) = μ Q (y),
x∈X x∈X
provided that Sup μ P = 1, and because of T (μ P (x), μ Q (y)) μ Q (y) and T is

continuous. But, for example, if T = Min, Sup μ P = 0.9 and Sup μ Q = 1, then
μ Q ∗ (y) = Min(0.9, μ Q (y)) = μ Q (y). Notice that for all the cases in which μ P is
normalized (μ P (x0 ) = 1 for some x0 ∈ X ), Mamdani-Larsen implications do verify
μ Q ∗ = μ Q whenever μ P ∗ = μ P .
Fourth Step: Numerical Input
This is the case in which μ P ∗ is exactly x = x0 or x ∈ {x0 }. That is “x is P ∗ ” is the
statement “x is x0 ” and hence

1 if x = x0
μ P (x) =
∗
0 if x = x0
In that case,
μ Q ∗ (y) = SupT1 (μ P j∗ (x), J (μ P (x), μ Q (y))) = J (μ P (x0 ), μ Q (y)),

x∈X
for all y ∈ Y .
For example, let J be a ML-implication, then μ Q ∗ (y) = T (μ P (x0 ), μ Q (y)).
If X = [0, 10], Y = [0, 1], P = close to 4, Q = big, with uses as shown in the
following figure and moreover x0 = 3.5, with J (a, b) = Min(a, b), then μ Q ∗ (y) =
Min(μ P (3.5), μ Q (y)) = Min(0.5, μ Q (y)), with μ P (x) = x − 3 between 3 and 4.
1 1
μP μQ
0.5
0
3 4 5 10 0 0.5 1
Hence, the graphic of the output μ Q ∗ is the one shown figure

1
0.5
μ Q∗
0 0.5 1
when x0 = 3.5.
Fifth Step: Numerical Consequent
This is the case in which μ Q is y = y0 or y ∈ {y0 }. That is “y is Q” corresponds to
“y is y0 ”, and

1 if y = y0
μ Q (y) = μ y0 =
0 if y = y0
In this case:

J (μ P (x), 1) if y = y0
J (μ P (x), μ y0 (y)) =
J (μ P (x), 0) if y = y0 ,
and the output μ Q ∗ depends on the values of J, namely, on J (a, 1) and J (a, 0).
Notice that if:
• J is an S-implication, J (a, 1) = 1; J (a, 0) = N (a).
• J is an Q-implication, J (a, 1) = S(N (a), a); J (a, 0) = N (a).
• J is an R-implication, J (a, 1) = 1; J (a, 0) = Sup{z ∈ [0, 1]; T (z, a) = 0}.
• J is an ML-implication, J (a, 1) = a; J (a, 0) = 0.
When J is an ML-implication, with J (a, 1) = a and J (a, 0) = 0:
μ Q ∗ (y) = Sup Min(μ P ∗ (x), J (μ P (x), μ y0 (y)))

x∈X

Sup Min(μ P ∗ (x), μ P (x)), if y = y0
= μ Q ∗ = x∈X
0, if y = y0 ,
Notice that if the input is also numerical, i.e. x = x0 then the output is:

Sup μ P ∗ (x) if y = y0 and x = x0
μ Q∗ = x∈X
0 if y = y0 and x = x0 .
Sixth Step: Several Rules with Numerical Input

The problem is now
By using the CRI, R1 with the input “x is P ∗ ” gives an output “y is Q ∗1 ”, with

μ Q ∗1 its membership function and R2 with the input “x is P ∗ ” gives “y is Q ∗2 ” with
μ Q ∗2 its membership function. The total output, “y is Q ∗ ”, corresponds to the idea
(y is Q ∗1 ) or (y is Q ∗2 ), and translating this ‘or ’ by means of the lowest t-conorm its
value can be obtained by μ Q ∗ (y) = Max(μ Q ∗1 (y), μ Q ∗2 (y)).
For example, the Mandani’s method consist in taking J (a, b) = Min(a, b) and
the Larsen’s method, in taking J (a, b) = a · b. Let us consider in X = [0, 10],
Y = [0, 1] the problem:
R1 : If x is close-to 4, then y is big

R2 : If x is small, then y is small
x = 3.5
and let us find μ Q ∗ using both methods by supposing “close-to 4” as in the former
example, μbig (y) = y, μsmall (x) = 1 − 10
x
, and μsmall (y) = 1 − y.
∗ ∗
The outputs Q 1 , Q 2 are:
IF μ close-to 4 THEN μ big
1 1
μ Q∗1
R1: 0.5
0
3 4 5 10 0 0.5 1
IF μ small THEN μ small

1 1
0.65 μ Q∗2
R2:
0 1
3.5 10 0 0.35
• With Mamdani’s method: μ Q ∗1 (y) = Min(μclose-to 4 (3.5), μbig (y)) = Min(0.5, y)

(see the figure above, upper right corner)
μ Q ∗2 (y) = Min(μsmall (3.5), μsmall (y)) = Min(0.65, 1 − y) (see lower right corner
of previous figure)
Hence, μ Q ∗ (y) = Max(μ Q ∗1 (y), μ Q ∗2 (y)) = Max(Min(0.5, y), Min(0.65, 1 − y))
(as shown in following figure).
1
0.65 μ Q∗
0.5
0 0.35 0.5 1
• With Larsen’s Method: μ Q ∗1 (y) = 0.5y (left of following figure)

μ Q ∗2 (y) = 0.65(1 − y) (middle of following figure)
Hence, μ Q ∗ (y) = Max(0.5y, 0.65(1 − y)) (right of following figure)
μ big μ small
1 1 1
0.65 0.65
0.5 μQ* 0.5

μ Q *2
μ Q *1
0 0.5 1 0 0.35 1 0 0.56 1
Seventh Step: A More Complex Example with Numerical Inputs and Conse-
quences
Let us consider the case:
With Larsen’s method and translating the and in the antecedents also by T = Prod.
• Rule R1 is represented by JL (T (μ P11 (x1 ), μ P12 (x2 )), μ y1 (y))

μ P11 (x1 ) · μ P12 (x2 ), y = y1
= μ P11 (x1 ) · μ P12 (x2 ) · μ y1 (y) =
0, y = y1
• Rule R2 is represented by JL (T (μ P21 (x1 ), μ P22 (x2 )), μ y2 (y))

μ P21 (x1 ) · μ P22 (x2 ), y = y2
= μ P21 (x1 ) · μ P22 (x2 ) · μ y2 (y) =
0, y = y2
Hence, the corresponding outputs under the CRI are:

μ P11 (x1∗ ) · μ P12 (x2∗ ), y = y1
μ Q ∗1 (y) =
0, y = y1

μ P21 (x1∗ ) · μ P22 (x2∗ ), y = y2
μ Q ∗2 (y) =
0, y = y2
Consequently:
⎧
⎨ μ P11 (x1∗ ) · μ P12 (x2∗ ), y = y1
μ Q∗ = Max(μ Q 1 (y), μ Q 2 (y)) μ P21 (x1∗ ) · μ P22 (x2∗ ),
∗ ∗ y = y2
⎩
0, otherwise
Eight Step: Defuzzification

Zadeh’s CRI gives an output function μ Q ∗ , but what it is frequently needed, mainly
in control, is an output number as an “order” to be executed by the system. Hence,
this step consists in compacting in the best possible way, in a single real number,
the information on the system’s behavior contained in μ Q ∗ . That is, the goal is to
defuzzify μ Q ∗ .
In the applications the most interesting cases are those in which, respectively,
either μ Q ∗ is a non-null continuous function, or it is a non-null function at only a
finite number of points in Y .
In the first case and among the diverse methods that have been suggested in the
literature, that known as “center of gravity” of the area below μ Q ∗ as well as the one
known as “center of area” are perhaps the most used ones. In the second case, if for
example, ⎧
⎪
⎪ α1 , if y = y1
⎪
⎪
⎨ α2 , if y = y2
μ Q ∗ (y) = · · ·
⎪
⎪
⎪
⎪ αn , if y = yn
⎩
0, otherwise
the most popular method of defuzzification is that consisting in taking the weighted
mean
α1 · y1 + α2 · y2 + · · · + αn · yn
μ Q∗ =
α1 + α1 + · · · + αn
The case of m rules with numerical consequents and numerical inputs, with
Larsen’s implication function and with defuzzification by the weighted mean, is the
basis of the so-called Takagi-Sugeno methods of fuzzy inference of orders 1, 2, 3, . . .

etc.
Example 8.2.1 In the example shown in sixth step where the output μ Q ∗ is obtained
through Mamdani, the area below μ Q ∗ is easily computed by 0.35 × 0.15 +
0.15 × 0.15
2 + 0.5 × 1 = 0.564. Hence, the center of area is a point y0 ∈ (0, 1)
such that, the areas to the left and to the right of y0 are equal, i.e.:
y0
0.564
= 0.282 = μ Q ∗ (y)dy
2
0
0.35 y0
= 0.65dy + (1 − y)dy
0 0.35
y0
= 0.228 + y0 − 0.35 − ydy
0.35
as the line joining the points (0.35, 0.65) and (0.5, 0.5) is z = 1 − y. Hence:
y0
y0 − ydy = 0.282 − 0.228 + 0.35 = 0.404
0.35

y0
y2 y02 0.352
y0 − = y0 − − = 0.404
2 0.35 2 2
gives: y02 − 2y0 + 0.686 = 0, with positive root y0 = 0.43919.
Example 8.2.2 In the case of non-null function at two number of points y1 , y2

(a two rule system) where weighted mean is taken as the output value y0 :
μ P11 (x1∗ )μ P12 (x2∗ 2)y1 + μ P21 (x1∗ )μ P22 (x2∗ )y2
y0 =
μ P11 (x1∗ )μ P12 (x2∗ ) + μ P21 (x1∗ )μ P22 (x2∗ )
Provided that X 1 = X 2 = [0, 1], Y = [0, 10], μ P11 (x1 ) = x1 , μ P12 (x2 ) = 1 − x2 ,
y1 = 6, μ P21 (x1 ) = 1 − x1 , μ P22 (x2 ) = x2 , y2 = 4, x1∗ = 0.3 and x2∗ = 0.7, the
calculation will be:
0.3 × (1 − 0.7) × 6 + (1 − 0.3) × 0.7 × 4 0.54 + 1.96 2.5
y0 = = = = 4.31
0.3 × (1 − 0.7) + (1 − 0.3) × 0.7 0.09 + 0.49 0.58
8.2.2 Takagi-Sugeno of Order 1
The method of Takagi-Sugeno of order n is an immediate generalization of the last

example with numerical input, numerical consequents, ML-implication and defuzzi-
fication by a weighted mean. It is the case in which in place of consequents y = yi ,
y is taken as a polynomial of degree n in the n variables x1 , x2 , . . . , xn and with the
given coefficients, appearing in the rules’ antecedents. For the sake of brevity, and
without any loss of generality, let us consider the case of m = 2 rules, with n = 2
variables:
Let us shorten y = α1 x1 + α2 x2 + α3 by q1 , and y = β1 x1 + β2 x2 + β3 by q2 .

with q1∗ = α1 x1∗ + α2 x2∗ + α3 and q2∗ = β1 x1∗ + β2 x2∗ + β3 . It follows:

1, if x1 = x1∗ and x2 = x2∗
μ x1∗ x2∗ (x1 , x2 ) =
0, otherwise

1, if q1
μq1 (y) =
0, otherwise

1, if q2
μq2 (y) =
0, otherwise
As rule R1 is represented by J1 = μ P11 (x1 ) · μ P12 (x2 ) · μq1 (y) and rule R2 by
J2 = μ P21 (x1 ) · μ P22 (x2 ) · μq2 (y), it follows:
μ Q ∗1 (y) = Sup Min(μx1∗ x2∗ (x1 , x2 ), μ P11 (x1 ) · μ P12 (x2 ) · μq1 (y))
x∈X, y∈Y

μ P11 (x1∗ )μ P12 (x2∗ ), if q1
= μ P11 (x1∗ )μ P12 (x2∗ )μq1 (y) =
0, otherwise

μ P21 (x1∗ )μ P22 (x2∗ ), if q2
μ Q ∗2 (y) = μ P21 (x1∗ )μ P22 (x2∗ )μq2 (y) =
0, otherwise
Hence,
⎧
⎨ μ P11 (x1∗ )μ P12 (x2∗ ), if q1
μ Q ∗ (y) = Max(μ Q ∗1 (y), μ Q ∗2 (y)) = μ P21 (x1∗ )μ P22 (x2∗ ), if q2
⎩
0, otherwise
And finally:
μ P11 (x1∗ )μ P12 (x2∗ )q1∗ + μ P21 (x1∗ )μ P22 (x2∗ )q2∗
y0 =
μ P11 (x1∗ )μ P12 (x2∗ ) + μ P21 (x1∗ )μ P22 (x2∗ )
8.3 Control of Nonlinear Systems
8.3.1 State-Space Representation
The state-space representation of a dynamical system is used in control engineering

in order to completely describe its behavior by first-order differential equations of the
involved state variables. Input variables u(t) = [u 1 (t), . . . , u p (t)]T , output variables
y(t) = [y1 (t), . . . , ym (t)]T and state variables x(t) = [x1 (t), . . . , xn (t)]T are all
needed to completely define any dynamical system. The state variables of a system
constitute the smallest set of variables which completely determine the state of a
system and in practice these variables usually correspond with physical magnitudes.
Thus, if the state of any system and its inputs are known at any given time, the future
behavior of the system will be determined by its state space representation. The state
variables are not always measurable and/or observable.
The following equation is a generic state-space representation of a nonlinear
system.
ẋ(t) = f (x(t)) + g(x(t))u(t)
y(t) = h(x(t))
If the dynamical system is linear and time invariant, the differential and algebraic
equations may be written in a matrix form, such as,

ẋ(t) = Ax(t) + Bu(t)
y(t) = C x(t)
Example 8.3.1 Taken a mechanical system of a mass with a spring and a damper as
an example,
y(t)
k
u
m
where an external force is the only input u(t) to the system, and the position y(t) is
the output. The physical equation of this single input single output (SISO) system is
8.3 Control of Nonlinear Systems 191
m ÿ + b ẏ + ky = u
which represents a second order linear system. In this case the state variables are
defined as
x1 (t) = y(t)
x2 (t) = ẏ(t)
and so the state space representation is,

ẋ1 = x2
ẋ2 = − mk x1 − b
m x2 + 1
mu
The output equation is defined by
y(t) = x1 (t).
In a matrix form, the state equation can be rewritten as,

ẋ1 0 1 x1 0
= + 1 u,
ẋ2 − mk − mb x2 m
similarly the output equation is

x1
y= 10 .
x2
Both state and output equations form the matrix state space representation of the
system
ẋ = Ax + Bu
y = Cx
with

0 1 0
A= , B= , C= 10 .
− mk − mb 1
m
8.3.2 Takagi-Sugeno Models for Control of Nonlinear Systems
The so-called Takagi-Sugeno (TS) fuzzy model is particularly useful for the control
of nonlinear systems. The TS fuzzy model is described by fuzzy “if-then” rules which
represent local linear input-output relations of a nonlinear system in a state-space
form. The overall fuzzy model of the system is achieved by fuzzy “blending” of the
linear systems. A TS fuzzy model representing a continuous dynamical system has
the following form,
Rule i:
IF z 1 (t) is Mi1 and · · · and z p (t) is Mi p

ẋ(t) = Ai x(t) + Bi u(t)
THEN
y(t) = Ci x(t)
for i = 1, 2, . . . , r .
Here, Mi j is the fuzzy set and r is the number of model rules; x(t) ∈ R n is the state
vector, u(t) ∈ R m is the input vector, y(t) ∈ R q is the output vector, Ai ∈ R n×n ,
Bi ∈ R n×m , and Ci ∈ R q×n ; z(t) = {z 1 (t), . . . , z p (t)} are known premise variables
that may be functions of the state variables and/or external disturbances.
Each linear consequent equation represented by Ai x(t) + Bi u(t) is called a “sub-
system”.
ẋ(t) = ri=1 h i (z(t)){Ai x(t) + Bi u(t)}
y(t) = ri=1 h i (z(t)) Ci x(t)
where,
wi (z(t))
h i (z(t)) = r ,
i=1 wi (z(t))

p
wi (z(t)) = Mi j (z(t))
j=1
for all t. The term Mi j (z(t)) is the grade of membership of z j (t) in Mi j . We have
r
i=1 h i (z(t)) = 1, and h i (z(t)) 0, for i = 1, 2, . . . , r and all t.
Thanks to the normalized membership functions, the linear dynamic TS model
is in fact a convex combination of local linear models. This property facilitates the
stability analysis of the fuzzy system. Fuzzy systems are universal function approx-
imators and hence can be used to model a wide class of processes.
Example 8.3.2 Consider the following nonlinear dynamic system
⎧
⎨ ẋ1 = −x1 + x1 x2
ẋ2 = x1 − 3x2
⎩
y = x1
with x1 , x2 ∈ [−1, 1]. This system can be exactly represented, using the sector
nonlinearity approach. In this model, the scheduling variable z 1 is chosen as x2 , the
fuzzy sets are M11 = ‘around − 1’, M12 = ‘around 1’, and the corresponding
membership functions are h 11 = (1 − z 1 )/2 and h 12 = (1 + z 1 )/2, respectively. The
following TS fuzzy system with linear consequents is the exact representation of the
original model,
R1 :
IF z 1 is around − 1
⎧
⎨ −2 0
ẋ = x
THEN 1 −3
⎩
y = x1
R2 :
IF z 1 is around 1
⎧
⎨ 00
ẋ = x
THEN 1 −3
⎩
y = x1
It can be easily seen that this is an exact representationof the nonlinear system
2
in the compact set x1 , x2 ∈ [−1, 1], when expressions i=1 h i (z(t)) Ai x(t) and
2
h
i=1 i (z(t)) C i x(t) are developed,

1 − x2 −2 0 x1 1 + x2 0 0 x1 −x1 + x1 x2
+ =
2 1 −3 x2 2 1 −3 x2 x1 − 3x2
1 − x2 1 + x2
x1 + x1 = x1 = y.
2 2
The main idea of the controller design based on TS fuzzy systems is to derive
each control rule so as to compensate each subsystem or rule of the fuzzy system.
So, for every subsystem
ẋ = Ai x + Bi u
a control signal of the form
u = −Fi x
is selected, resulting in a closed-loop system formed by a set of subsystems such as,
ẋ = (Ai − Bi Fi )x.
The design procedure is conceptually simple and natural. The stability analysis and
control design problems can be reduced to linear matrix inequality (LMI) problems.
8.3.3 Stability Analysis
TS fuzzy model is a convex combination of linear models. This structure facili-

tates stability analysis and observer design by using effective algorithms based on
Lyapunov functions and linear matrix inequalities (LMI).
Talking about the Lyapunov function V (x) of a system, is in a way equivalent
to talk about the potential energy of the system. For a system to be stable in the
sense of Lyapunov, it must remain within a bounded region “close-enough” to an
equilibrium point starting from any initial state, and when external input is u(t) = 0.
Similarly, asymptotic stability means that the system, not only remains close, but
must eventually converge to an equilibrium point. Any system in a given initial state
must have a positive potential energy V (x) 0 (V (x) = 0 in an equilibrium point),
and to ensure that it eventually converges to an equilibrium point, the derivative of
this Lyapunov function must be negative V̇ (x) = dt d
V (x) 0. In a sense, it must
be proven that the energy of the system eventually dissipates taking the system to an
equilibrium point, when no external input is acting over the system (u(t) = 0).
The basic stability analysis considers the following quadratic Lyapunov function:
V (x) = x T (t)P x(t),
with P = P T > 0. If this Lyapunov function is considered, its derivative along the
trajectories of the TS fuzzy subsystem of the type ẋ = Ai x is
V̇ (x) = ẋ T (t)P + P ẋ(t) = AiT P + PAi
and so:
Theorem 8.3.3 The equilibrium of a TS fuzzy system is globally asymptotically
stable if there exists a common positive definite matrix P such that
AiT P + PAi < 0,
for i = 1, 2, . . . , r .
That is, a common P has to exist for all subsystems. This theorem reduces to the
Lyapunov stability theorem for continuous linear systems when r = 1. If there exists
a P > 0 such that V (x(t)) proves the stability of the system, it is also said to be
quadratically stable as V (x(t)) is called a quadratic Lyapunov function.
The theorem presents a sufficient condition for the quadratic stability of the TS
system, and this is a common P problem that can be solved efficiently via convex
optimization techniques for LMIs. For systems and control, the LMI optimization is
particularly useful due to the fact that a wide variety of system and control problems
can be recast as LMI problems. Apart from a few special cases these problems do
not have analytical solutions. However, through the LMI framework they can be
efficiently solved numerically in all cases. Therefore, recasting a control problem as
an LMI problem is equivalent to finding a “solution” to the original problem.
8.3.4 Parallel Distributed Compensation
The parallel distributed compensation (PDC) offers a procedure to design a fuzzy

controller from a given TS fuzzy model. To realize the PDC, a controlled nonlinear
system is first represented by a TS fuzzy model. Each control rule is designed from
the corresponding rule of a TS fuzzy model, and the designed fuzzy controller shares
the same fuzzy sets with the fuzzy model in the premise parts.
The fuzzy controller will have the following form,
Control Rule i:
IF z 1 (t) is Mi1 and · · · and z p (t) is Mi p ,

THEN u(t) = −Fi x(t),
for i = 1, 2, . . . , r .
The fuzzy control rules have a linear controller (state feedback laws in this case)
in the consequent parts. The overall fuzzy controller is represented by

r
u(t) = − h i (z(t)) Fi x(t).
i=1
The fuzzy controller design consists on determining the local feedback gains Fi in
the consequent parts. With PDC we have a simple and natural procedure to handle
nonlinear control systems. Other nonlinear control techniques require special and
rather involved knowledge.
The overall closed-loop system using the PDC controller method is,

r
r
ẋ(t) = h i (z(t))h j (z(t)){Ai − Bi F j }x(t)}.
i=1 j=1
For this case, the following sufficient condition for stability can be obtained:
Theorem 8.3.4 The equilibrium of a fuzzy control system is globally asymptotically
stable if there exists a common positive definite matrix P such that
{Ai − Bi F j }T P + P{Ai − Bi F j } < 0
for h i (z(t)) · h j (z(t)) = 0, ∀t, i, j = 1, 2, . . . , r .

A control problem in the framework of the TS fuzzy model and PDC design,
targets the design of a controller that ensures the stability of the closed-loop system.
Example 8.3.5 Considering the balancing of an inverted pendulum on a cart. The
motion equations of the pendulum are described by,
ẋ1 (t) = x2 (t)

gsin(x1 (t)) − amlx22 (t)sin(2x1 (t))/2 − acos(x1 (t))u(t)

ẋ2 (t) =
4l/3 − amlcos2 (x1 (t))
where x1 (t) is the angle of the pendulum, x2 (t) is the angular velocity; g is the
gravity, m is the mass of the pendulum, M is the mass of the cart, l is the length of
the pendulum, u(t) is the force applied to the cart and a = 1/(m + M).
A fuzzy model which approximates the dynamics of the nonlinear plant for the
range x1 ∈ (−π/2, π/2) can be realized by the following two rules,
R1 : IF x1 is about 0 THEN ẋ(t) = A1 x(t) + B1 u(t)
R2 : IF |x1 | is about π/2 THEN ẋ(t) = A2 x(t) + B2 u(t)
where,

0 1 0
A1 = g B1 = g
4l/3−aml 0 − 4l/3−aml

0 1 0
A2 = 2g B2 = aβ
π(4l/3−amlβ 2 )
0 − 4l/3−aml
and β = cos(88◦ ). The membership functions are
R2 R1
x1
-90 0 90
Through the PDC control, the stability of the system is guaranteed for example
with
F1 = [−120.67 − 22.67] ,
F2 = [−2551.6 − 764]
where it is obtained

3.62 0.62
P= .
0.62 0.28
The extension of the pendulum control to the full circle x1 ∈ [−π, π] work space,
would be achieved similarly, only by adding two more rules to the TS fuzzy system.
8.3.5 Piecewise Bilinear Model
Recently, Piecewise Bilinear (PB) model is being used for control purpose. The
PB model is a fully parametric model to represent Linear/Nonlinear systems. The
obtained model is built on piecewise rectangular regions, and each region is defined
by four vertices partitioning the state space. As the conventional nonlinear system
control based on TS fuzzy model represents a connection of linear state-space models
by sector nonlinearity, the PB model represents a convex combination of the vertices
defining piecewise regions.
In this approach, bilinear functions are used to regionally approximate any given
function. A bilinear function is a nonlinear function of the form y = a + bx1 +
cx2 + d x1 x2 , where any four points in the three dimensional space are spanned with
a bi-affine plane.
PB model has a good general approximation capability and it has a continuous
crossing over the piecewise regions. Its interpretability, simplicity and visibility facil-
itates the realization of controllers in industrial applications. A local error does not
trigger a global error and its interpolation nature generates robust outputs. A draw-
back of the PB model is that the stability analysis based on Lyapunov is difficult as
bilinear matrix inequalities (BMI) must be solved.
If a general case of an affine two-dimensional nonlinear control system is consid-
ered, ⎧
⎨ ẋ1 = f 1 (x1 , x2 )
ẋ2 = f 2 (x1 , x2 ) + g(x1 , x2 ) · u
⎩
y = h(x1 , x2 )
where u is the input. For the PB representation of a state-space equation, a coordinate

vector d(σ, τ ) of the state space and a rectangle Ri j must be defined as,
d(i, j) ≡ (d1 (i), d2 ( j))T

Ri j ≡ [d1 (i), d1 (i + 1)] × [d2 ( j), d2 ( j + 1)]
where i ∈ (1, . . . , n 1 ) and j ∈ (1, . . . , n 2 ) are integers, and d1 (i) < d1 (i + 1),
d2 ( j) < d2 ( j + 1). The PB models are formed by matrices of size (n 1 × n 2 ), where
n 1 and n 2 represent the number of partitions of dimension x1 and x2 respectively.
Each value in the matrix is referred to as a vertex in the PB model. The operational
region of the system is divided into (n 1 − 1 × n 2 − 1) piecewise regions that are
analyzed independently.
The PB model is a particular case of TS systems, as it is derived from a set of

fuzzy if-then rules with singleton consequents such that
IF x is W στ , THEN ẋ is f (σ, τ )
which in a two-dimensional case, x ∈ R2 is a state vector, W στ = (w1σ (x1 ), w2τ (x2 ))T
is a membership function vector, f (σ, τ ) = ( f 1 (σ, τ ), f 2 (σ, τ ))T ∈ R is a singleton
consequent vector, and σ, τ ∈ Z are integers (1 σ n 1 , 1 τ n 2 ) defined by,
σ(x1 ) = d1 (max(i)) where d1 (i) x1 ,
τ (x2 ) = d2 (max( j)) where d2 ( j) x2 .
For x ∈ Rστ , the PB models that approximates a general nonlinear function is

expressed as, ⎧
⎪ τ
σ+1 +1
⎪
⎪ f 1 (x1 , x2 ) =
j
w1i (x1 )w2 (x2 ) f 1 (i, j),
⎪
⎪
⎪
⎪ i=σ j=τ
⎪
⎪
⎪
⎪ τ
σ+1 +1
⎪
⎪ f (x , x ) =
j
w1i (x1 )w2 (x2 ) f 2 (i, j),
⎨ 2 1 2
i=σ j=τ
⎪
⎪ τ
σ+1 +1
⎪
⎪ g(x1 , x2 ) =
j
w1i (x1 )w2 (x2 )g(i, j),
⎪
⎪
⎪
⎪ i=σ j=τ
⎪
⎪
⎪
⎪ τ
σ+1 +1
⎪
⎩ h(x1 , x2 ) =
j
w1i (x1 )w2 (x2 )h(i, j),
i=σ j=τ
where ⎧ σ
⎪
⎪ w (x1 ) = 1 − α,
⎪ 1σ+1
⎨ w1 (x1 ) = α,
⎪ τ
⎪ w2 (x2 ) = 1 − β,
⎪
⎩ τ +1
w2 (x2 ) = β,
and
x1 − d1 (σ)
α=
d1 (σ + 1) − d1 (σ)
x2 − d2 (τ )
β=
d2 (τ + 1) − d2 (τ )
j
in which case w1i , w2 ∈ [0, 1].
In every region of the PB models, i.e.: f 1 (x1 , x2 ), the values are computed through
bilinear interpolation of the corresponding four vertexes as shown in the figure.
Note that the approximation is made by only using the values of a nonlinear
function at the vertexes of Ri j ’s.
8.3.6 Vertex Placement Principle
Generally speaking, if a plant model P and a desired (closed-loop) plant model DP

are given, we may have a strategy to design a Controller C by setting a relation
C = DP − P. This strategy works for linear systems but does not work in general for
nonlinear systems as we know. However if nonlinear systems are modeled with PB
systems, we can take this strategy by applying Vertex Placement Principle (VPP).
This is a general principle to design a LUT-controller to utilize the characteristics
of the PB model of an objective plant. Given a partition of the sate space into sub-
regions, the property of a PB model can be completely described by the values of a
PB model at the vertexes of regions. Therefore, to design a LUT-controller means to
assign the values of a table at the vertex positions. VPP guides us to assign the values
of a controller at the vertex positions based on the values of a PB model describing
the nonlinear plant. In this sense, VPP is similar to the idea of Pole Placement in
linear systems.
Considering an nonlinear plant (P) type

⎧
⎨ ẋ1 = x2
τ
σ+1 +1
j
⎩ ẋ2 = f 2 p (x1 , x2 ) + u = w1i (x1 )w2 (x2 ) f 2P (i, j) + u
i=σ j=τ
if a controller σ+1 τ +1
j
u = C(x1 , x2 ) = w1i (x1 )w2 (x2 )C(i, j)
i=σ j=τ
is considered, the closed-loop CP will be described as

ẋ1 = x2
ẋ2 = f 2cp (x1 , x2 ) = f 2 p (x1 , x2 ) + C(x1 , x2 )
In order to make the closed-loop behavior of the system be approximated to the

performance of a desired plant DP ( f 2cp = f 2d p ), the controller can be directly
computed as,
u = C(x1 , x2 ) = f 2d p (x1 , x2 ) − f 2 p (x1 , x2 ).
Example 8.3.6 Considering the so called “Van der Pol” oscillator, which is a non-
linear plant expressed as,

x˙1 = x2
x˙2 = −x1 + (1 − x12 )x2 + u
If the operational region of the systems is selected as −4 x1 4 and −4

x2 4. Thus, the partition of the whole region into piecewise regions is done
through d1 = d2 = (−4, −2, 0, 2, 4). The piecewise bilinear model representation
of the f 2 p is the following (Table 8.1).
Table 8.1 PB representation x1 \x2 d2 = −4 d2 = −2 d2 = 0 d2 = 2 d2 = 4

of f 2 p
d1 = −4 64 34 4 −26 −56
d1 = −2 14 8 2 −4 −10
d1 =0 −4 −2 0 2 4
d1 =2 10 4 −2 −8 −14
d1 =4 56 26 −4 −34 −64
The following figures show both the original surface of the Van der Pol system
and it’s piecewise bilinear representation for the vertex positions d1 and d2 .
The goal is to design a controller C so that the closed-loop system behaves like
the linear plant represented as,

x˙1 = x2
x˙2 = ax1 + bx2 + u
where a = −1 and b = −3. The PB representation of f 2d p for the linear plant is the
following (Table 8.2).
With the controller calculated as u = f 2d p − f 2 p , the following PB model of the
controller C is obtained (Table 8.3).
Table 8.2 PB representation x1 \x2 d2 = −4 d2 = −2 d2 = 0 d2 = 2 d2 = 4

of f 2dp
d1 = −4 16 10 4 −2 −8
d1 = −2 14 8 2 −4 −10
d1 =0 12 6 0 −6 −12
d1 =2 10 4 −2 −8 −14
d1 =4 8 2 −4 −10 −16
Table 8.3 PB model of C x1 \x2 d2 = −4 d2 = −2 d2 = 0 d2 = 2 d2 = 4

d1 = −4 –48 –24 0 24 48
d1 = −2 0 0 0 0 0
d1 =0 16 8 0 –8 –16
d1 =2 0 0 0 0 0
d1 =4 –48 –24 0 24 48
Next figures show both the controller C and the resulting closed-loop plant
surfaces respectively.
Bibliography
1. B. Bede, Mathematics of Fuzzy Sets and Fuzzy Logic (Springer, London, 2013)
2. R. Belohlavek, Fuzzy Relational Systems: Foundations and Principles (Kluwer Academic
Publishers, New York, 2002)
3. G. Böhme, Fuzzy-Logik: Einführung in die algebraischen und logischen Grundlagen (Springer,
London, 1993)
4. B. Bouchon-Meunier, La logique floue (PUF, Paris, 2003)
5. S.P. Boyd et al., Linear Matrix Inequalities in System and Control Theory (SIAM, Philadelphia,
1994)
6. B. Demant, Fuzzy-theorie oder die faszination des vagen (Vieweg, Brauschweig, Wiesbaden,
1993)
7. D. Dubois et al., Fuzzy Sets: History and Basic Notions (Springer, London, 2000)
8. M. Grabisch et al., Fuzzy Measures and Integrals: Theory and Applications (Springer, New
York, 2000)
9. G. Klir, B. Yuan, Fuzzy Sets and Fuzzy Logic, vol 4 (Prentice Hall, New Jersey, 1995)
10. G.J. Klir, U. St Clair, B. Yuan, Fuzzy Set Theory: Foundations and Applications (Prentice-Hall,
Upper Saddle River, 1997)
11. Z. Lendek et al., Stability Analysis and Nonlinear Observer Design Using Takagi-Sugeno Fuzzy
Models (Springer, London, 2010)
12. B. Liu, Theory and Practice of Uncertain Programming. (Springer, London, 2009)
13. R. Lowen, Fuzzy set theory (Basic Concepts, Techniques and Bibliography, Dordrecht, 1996)
14. R. López de Mántaras, Approximate Reasoning Models (Prentice Hall PTR, Upper Saddle
River, 1990)
15. D. McNeill, Fuzzy Logic: The Revolutionary Computer Technology that is Changing Our World
(Simon and Schuster, New York, 1994)
16. G. Metcalfe, N. Olivetti, D.M. Gabbay, Proof Theory for Fuzzy Logics, vol 36 (Springer,
London, 2008)
17. H.T. Nguyen, E.A. Walker, A First Course in Fuzzy Logic (CRC Press, Boca Raton, 2005)
18. A. Syropoulos, Theory of Fuzzy Computation (Springer, New York, 2014)
19. K. Tanaka, An Introduction to Fuzzy Logic for Practical Applications (Springer, New York,
1997)
20. K. Tanaka, H.O. Wang, Fuzzy Control Systems Design and Analysis: A Linear Matrix Inequality
Approach (Wiley, Danvers, 2004)
21. E. Trillas, Conjuntos Borrosos (Vicens-Vives, Barcelona, 1980)
22. E. Trillas, C. Alsina, J.-M. Terricabras, Introducción a la lógica borrosa. (Ariel Barcelona,
Spain, 1995)
and Soft Computing 320, DOI 10.1007/978-3-319-14203-6
204 Bibliography
23. E. Turunen, E. Turunen, Mathematics Behind Fuzzy Logic (Physica-Verlag, Heidelberg, 1999)
24. X. Wang, D. Ruan, E.E. Kerre, Mathematics of Fuzziness: Basic Issues (Springer, Heidelberg,
2009)
25. J. Yen, R. Langari, Fuzzy Logic: Intelligence, Control, and Information (Prentice-Hall, Upper
Saddle River, 1998)
26. L.A. Zadeh, Fuzzy Sets and Applications: Selected Papers (Wiley, New York, 1987)
27. L.A. Zadeh, Computing with Words: Principal Concepts and Ideas (Springer, Berlin, 2012)
28. L.A. Zadeh, B. Yuan, G.J. Klir, Fuzzy sets, fuzzy logic, and fuzzy systems: selected papers by
Lotfi A (World Scientific, Zadeh, 1996)
29. H.J. Zimmermann, Fuzzy Set Theory and Its Applications, Second Revised Edition. (Springer,
New York, 1992)

Fuzzy Logic An Introductory Course For Eng - Enric Trillas

Uploaded by

Copyright:

Available Formats

Fuzzy Logic An Introductory Course For Eng - Enric Trillas

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fuzzy Logic An Introductory Course For Eng - Enric Trillas

Uploaded by

Copyright:

Available Formats

Studies in Fuzziness and Soft Computing

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

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

ISSN 1434-9922 ISSN 1860-0808 (electronic)

Library of Congress Control Number: 2014958577

Springer Cham Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

1 On the Roots of Fuzzy Sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Algebras of Fuzzy Sets . . . . . . . . . . . . .... . . . . . . . . . . . . . . . . . 35

2.2 The Concept of an ‘Algebra of Fuzzy Sets’. . . . . . . . . . . . . . . . 41

3 Reasoning and Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4 Fuzzy Relations . . . . . . . . . . . . . . . . . . . . . . . . .............. 117

5 T-Preorders and T-Indistinguishabilities . . . . . . . . . . . . . . . . . . . . 131

6 Fuzzy Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

7 Fuzzy Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

8 An Introduction to Fuzzy Control . . . . . . . . . . . . . . . . . . . . . . . . . 175

1.1 A Genesis of Fuzzy Sets

What is it meant by the predicate odd? In principle, it depends on where it is

it results X/ = Q with the 5 classes {x1 }, {x2 }, {x3 }, {x4 }, {x5 }.

Let us suppose P in X given by its primary use P , and let (L , ) be a poset. An

Degree up to which x is P = μ P (x) ∈ L ,

and said that the primary use of P is gradable in (L , ). If x = P y, then

with which it is P ⊂ μ P . When P = μ P , μ P perfectly reflects the primary use

[a1 , b1 ] [a2 , b2 ] ⇔ a1 a2 & b1 b2 .

Remark 1.1.5 When (L , ) is not linear, as it is with L = C the unit complex

1.1.2 Fuzzy Sets

In a lot of cases, mainly in the applications, predicates P in X result indirectly

• X/ = P = {[x], [y]}, and vx = v y , P is of the type

In the particular case {vx , v y } = {α, ω}, P is of the type

verifies: x P y ⇒ μ P (x) μ P (y), and hence μ P is an [0, 1]−degree for P in

If x y, then μ P (x) μ P (y),

Finally, the following two models are hyperbolic,

µold µbig Age

with μold increasing quadratically between the 50 and the 80 years.

Analogously, a person in the twenties could design μold as

1.2 Opposite, Negate, and Middle

Very often the meaning of a predicate P is not captured without simultaneously

Concerning every opposite a P of P, this opposition is translated by

since ‘x is less a P than y’ should be equivalent to ‘y is less P than x’.

and with a(big) = small, it can be said μsmall (x) = 1 − 10 .x

with the same symmetry A(x) = 10 − x, it results

It is easy to prove that A(x) = 10 10−x

with A(x) = 10 − x, results

If x is less P than y, then y is less not P than x,

or, equivalently, P ⊂ −1

• If μ P (x) = α, then μ P  (x) = ω

Hence, given an L-degree μ P of P in X , with each function N verifying (1)

μ(P  ) (x) = N (μ P  (x)) = N (N (μ P (x)))

for all x in X , or μ(P  ) = μ P .

For example, N0 (a) = 1 − a, N1 (a) = 1−a

Notice that, provided N is in the Sugeno’s family of strong negations, it is enough

• If Nλ (0.4) = 0.5, it results λ = 21 .

Example 1.2.1 With μ A5 as defined in 1.1.6, and with N0 , it is

whose graphics are

1.2.3 Antonyms and Negations

between A and N , that should be always verified. If A is known, N should be chosen

Example 1.2.4 With N0 (a) = 1 − a, and A(x) = 10 − x in X = [0, 10], if

Example 1.2.5 In [0, 10] take μbig (x) = 10

• If μ P (x) = α, then μ P (x) = ω

μ(P ) (x) = N (μ P (x)) = N (N (μ P (x)))

for all x in X , or μ(P ) = μ P .

μ P (x) μ P or Q (x), μ Q (x) μ P or Q (x).

Given P, if with M for medium, MP = Not P and Not aP = P and (aP) ,

(x, y) ∈ Q|P ⇔ ‘If x is P, then y is Q ⇔ ‘x is P implies y is Q .