BOOK

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/260794375
Fuzzy Sets and Fuzzy Logic. The Foundations of Application—from a

Mathematical Point of View
Book · February 1993

DOI: 10.1007/978-3-322-86812-1
CITATIONS READS
269 2,131
1 author:
Siegfried Gottwald
University of Leipzig
178 PUBLICATIONS 3,159 CITATIONS
SEE PROFILE
All content following this page was uploaded by Siegfried Gottwald on 17 March 2014.
The user has requested enhancement of the downloaded file.

6earen
*T1
D
@
lsnu
e tuoJJ
e\el^Jo lqod lBcllBtueqlsl^tr
- uorlucrlddy
Jo suollspunoc eqJ
rIEoT '{zznt puu
slas llzznt 'Eurllapouronrl
'('c1a 'qceeds 'saE€rur'
fV
eEpelmoul'Eurssecord aEpslr
'Jnoh"qaq go uorls
1ua6r11a1ur
plB^UloCpeygEols
-
Die DeutscheBibliothek ClP-Einheitsaufnahme
Gottwald, Siegfried: - Braunschweig

F;;i sets anJn , zy logic/ Siegfüed Gottwald'
; Wiesbaden: Vieweg, 1993
(Artifi cial Intelligence)
rsBN 3-523-05311-9
and
son
thu
a,w
AMSSubjectClassification:03E72,03852,04A72,68N17'68Po5,93C42'94D05
too
Wiesbaden
Verlag Vieweg, P.O' Box 5829,D-6200 fieL
frot
DistributionrightsforFrancebyTeknea,Toulouse,Marseille,Barcelona.
itr
tior
All rights reserved 1993
mbH, Braunschweig/IViesbaden' reft
o Friedr. vieweg & Sohnverlagsgesellschaft
rea
Publishing Group International'
vieweg is a subsidiary company of the Bertelsmann IF.
An
stored in a retrieval system
ma
No part of the publication may be reproduced,
ortransmitted,mechanical,photocopyingorotherwise'withoutprior
permission of the copyright holder' ati
ma
tht
Coverdesign:L. Markgraf,Wiesbaden-
chi
Lengerich
print.a ani boundby iengericherHandelsdruckerei' of
Printedon acid-freePaper fuz
Printedin GermanY
is
of
ISSN0940-699
rs BN 3 -5 2 8 -0 5 3 1 l -9
Preface
Quite a long time after its inception in about 1965, the field of fuzzy sets
and fuzzy logic was ranked to be some exotic field of research because of
its seeming tendency away from “hard” mathematical modelling and thus –
as often supposed by mathematically inclined outsiders – completely away
from an acceptable mathematically based theory.
The very recent success with even consumer products involving fuzzy
tools may not only cause the rapidly growing interest of the engineering
community in this field, but may also become responsible for more extended
interest from mathematicians and computer scientists.
The term “fuzzy logic” has undergone an almost continuous change from
merely referring to tools from many-valued logic appearing in considerations
of fuzzy sets and sometimes of a generalized switching theory, through the
field of fuzzy logic control with its limited use of nonstandard reasoning
by interpreting implication-like formulations of the rule oriented if. . . then
- type as fuzzy relations between the values of fuzzy variables, up to the
presently prevailing understanding as a theory of approximate reasoning.
Therefore it may not be without some interest to give here a reconsidera-
tion of a line of research which tends toward a combination of mathematical,
in particular set theoretical and logical tools which are related to problems
of a foundational character with problems which have arisen out of applica-
tional, engineering discussions about modelling strategies in the fuzzy field,
especially in fuzzy (logic) control. The unifying background idea is the use of
notions and methods from many-valued logics in the discussion of problems
which, in the end, are related to fuzzy sets applications.
vi
The author’s lasting interest in this field arose out of his early work on the
set theoretic foundation for fuzzy sets, cf. Gottwald (1971, 1971a, 1974,
1976, 1979, 1980) and the historical reminiscences in Gottwald (1984c),
but was essentially stimulated since the late 1970s by fruitful contacts with
(mainly Polish) control engineering people, among whom Witold Pedrycz
was the most influential.
The present monograph gives a rethought, revised and reorganized, as
well as rewritten and extended collection of the author’s essential results
obtained since that time and within, as well as via, this cooperation – results
which directly and indirectly are related to fuzzy relation equations and fuzzy
control.
Chapter 1 is essentially based on the research reported in Gottwald
(1984), but contains also ideas of Gottwald (1986, 1986a). Chapter 2 is
based on Gottwald (1986a), extends those results partly with results from
Gottwald (1974, 1986) and also includes unpublished material announced
in Gottwald (1991). Chapter 3 is based on Gottwald (1983, 1986, 1992),
as well as on parts of Gottwald/Pedrycz (1986, 1988). Chapter 4 com-
bines results from Gottwald (1984a, 1986b) and Gottwald/Pedrycz
(1986) with newly written (and yet unpublished) parts. The final chapter 5
has Gottwald/Pedrycz (1986, 1986a, 1988) as its most essential sources.
The final writing of this book was initiated and decisively made possible
by a grant from the SEL-Stiftung (Stuttgart) that enabled the author to
stay for the period of writing away from his home university of Leipzig at
the Computer Science Department of the Technische Hochschule Darmstadt
within the stimulating atmosphere of the FG Intellektik, whose chairman
Wolfgang Bibel had the final responsibility for the author’s excellent working
conditions.
To all these institutions and those – named as well as unnamed – people
the author is exceptionally grateful. He also is grateful to his wife and fam-
ily for their stimulating interest, as well as for their success in leaving him
untouched by almost all the small problems of everyday life.
Darmstadt / Leipzig,
Summer 1992.
Contents
1 Logical Preliminaries 1
1.1 Basic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 t-norms and Φ-operators . . . . . . . . . . . . . . . . . . . . . 5
1.3 t-norm based connectives . . . . . . . . . . . . . . . . . . . . 18
2 Basic fuzzy set theory 37

2.1 Set algebra for fuzzy sets . . . . . . . . . . . . . . . . . . . . 37
2.2 Fuzzy relations . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.3 The full image under a relation . . . . . . . . . . . . . . . . . 63
2.4 Special types of fuzzy relations . . . . . . . . . . . . . . . . . 67
2.5 Graded properties of fuzzy relations . . . . . . . . . . . . . . 84
3 Set equations with fuzzy sets 97

3.1 Fuzzy equations and some of their applications . . . . . . . . 97
3.2 Solvability of fuzzy relational equations . . . . . . . . . . . . 100
3.3 Solvability of fuzzy arithmetical equations . . . . . . . . . . . 103
3.4 Solvability of systems of fuzzy equations . . . . . . . . . . . . 106
3.5 Solvability degrees and approximate solutions . . . . . . . . . 112
3.6 Towards more difficult equations . . . . . . . . . . . . . . . . 128
4 Fuzzy controllers 135

4.1 The construction of fuzzy controllers . . . . . . . . . . . . . . 135
4.2 The problem of interaction . . . . . . . . . . . . . . . . . . . 144
4.2.1 General results on interactivity . . . . . . . . . . . . . 145
4.2.2 Coding rules by A ×t B . . . . . . . . . . . . . . . . . 149
4.2.3 Coding rules by (A ×t1 B) ∪t2 (A ×t1 Y ) . . . . . . . . 152
4.2.4 Coding rules by (A ×t1 Y ) ∪t2 (X ×t1 B) . . . . . . . . 157
4.3 Manipulation of fuzzy data . . . . . . . . . . . . . . . . . . . 159
vii
viii CONTENTS
5 Methodological issues 171

5.1 Comparison of fuzzy sets . . . . . . . . . . . . . . . . . . . . . 171
5.2 Approximate solutions . . . . . . . . . . . . . . . . . . . . . . 177
5.3 Fuzzy equations for processing fuzzy data . . . . . . . . . . . 183
5.4 Evaluation of fuzzy models . . . . . . . . . . . . . . . . . . . 188
5.4.1 Evaluating fuzzy models through fuzzy integrals . . . 188
5.4.2 Evaluating fuzzy models using probabilistic ideas . . . 192
5.4.3 Consequences for fuzzy modelling strategies . . . . . . 195
5.5 Controllability and predictability . . . . . . . . . . . . . . . . 198
Bibliography 207
Index 216
0 CONTENTS
Chapter 1
Logical Preliminaries
1.1 Basic notions

A fuzzy set A over some universe of discourse X (which itself is a classical set)
is characterized by its membership function µA : X → [0, 1] and often called
fuzzy subset of X . The fuzzy sets will be identified with their membership
functions such that the class IF (X ) of all fuzzy subsets of the universe of
discourse X becomes the class [0, 1]X of all functions f with dom (f ) = X
and rg (f ) ⊆ [0, 1].
Fuzzy sets are thus generalized characteristic functions. In essentially
the same way as the values χM (a) of the characteristic function χM of a
classical, “crisp” set M code the truth and falsity of the predicate “a ∈ M”
such that
½
1, iff a ∈ M, i.e. iff “a ∈ M” is true
χM (a) =
0, iff a ∈
/ M, i.e. iff “a ∈ M” is false,
the membership degrees of fuzzy sets are considered as generalized truth
values, i.e. as truth degrees of a suitable many-valued logic L.
The many-valued logic L which we will use has the real interval [0,1]
as its set of truth degrees and is supposed to be semantically based, i.e. we
assume that the connectives and quantifiers of the language of L will be
defined via truth functions and truth functional conditions, and we assume
that the truth degree 1 is the only (positively) designated one. We will not
look at problems of axiomatization here.
The truth functions of many-valued logic which are chosen to define e.g.
a conjunction, a disjunction, or a negation are usually supposed to behave
in such a way that they assume a value out of {0, 1} if all their arguments
2 CHAPTER 1. LOGICAL PRELIMINARIES
are from {0, 1} – and additionally to coincide under such circumstances

with the binary truth functions of conjunction, disjunction, and negation of
classical (propositional) logic. But other, more nonstandard truth functions
are possible and have been considered; cf. Rescher (1969), Gottwald
(1989).
Many-valued logic as a separate field of logical investigations essentially
developed out of the work of the Polish logician J. L Ã ukasiewicz and was
extensively developed in Poland in the 1920s; cf. L Ã ukasiewicz/Tarski
(1930), L Ã ukasiewicz (1970). Since then, the real unit interval [0,1] has
traditionally been a standard choice for the set of truth degrees in many-
valued logic. And the functions min, max are the truth functions of standard
generalizations of conjunction and disjunction.
As usual we denote with ∧ the (generalized) conjunction with truth func-
tion min : [0, 1]2 → [0, 1] and denote with ∨ the (generalized) disjunction
with truth function max : [0, 1]2 → [0, 1]. Additionally a kind of standard
(generalized) negation ¬ with truth function nL (x) = 1 − x usually is con-
sidered.
Yet, whole classes of further truth functions for conjunctions, disjunc-
tions, implications, and negations seem to be of interest. Therefore, as con-
nectives we furthermore suppose to have conjunctions ∧t defined by given
t-norms t and implication operators →t also defined by given t-norms t
(but via so-called Φ-operators connected with the t-norms t); additionally
we connect with each t-conorm st a disjunction operator ∨t and with every
negation function n a negation operator ∼n . (Those types of truth functions
we have referred to will be defined in the next section.)
As far as quantifiers are concerned, we take one for many-valued univer-
sal quantification (denoted: ∀) and another one for many-valued existential
quantification (denoted: ∃).
The well-formed formulas of our language are supposed to be defined in
the usual way. And, assuming that the actual values of all free variables of a
well-formed formula H always are determined by the context, we write [[H]]
for the truth degree of this formula H.
For the quantifiers we have independent of the t-norms under consider-
ation
[[∀xH(x)]] = inf [[H(x/a)]], (1.1)
a∈X
[[∃xH(x)]] = sup [[H(x/a)]] (1.2)
a∈X
where obviously the “substitution notation” H(x/a) does mean that the free
1.1. BASIC NOTIONS 3
variable x of H has to be given the value a from the universe X .

As in classical logic it is often helpful to have besides those unrestricted
quantifiers also the restricted ones, usually written ∀C xH(x) and ∃C xH(x).
Here C indicates some unary predicate or some (crisp) set, and accord-
ingly those restricted quantifications are understood as ∀x(C(x) → H(x))
or ∃x(C(x) ∧ H(x)) for a predicate C and as ∀x(x ∈ C → H(x)) or
∃x(x ∈ C ∧ H(x)) for a set C.
In many-valued logic one has to be more careful in that translation. Prin-
cipally one ought to distinguish two cases: C marking a classical, two-valued
restriction (realized as a unary “crisp” predicate or a crisp set) or C marking
a “many-valued restriction”. It seems that for the latter case a good intuitive
understanding is lacking such that in such a case one should prefer to work
with unrestricted quantifications. But if C is a crisp restriction now taken
as a crisp set, a quite natural understanding of restricted quantifications is
at hand:
[[∀C xH(x)]] = inf [[H(x/a)]], (1.3)

a∈C
[[∃C xH(x)]] = sup[[H(x/a)]]. (1.4)
a∈C
(If C means a “crisp” predicate, substitute C(a) for a ∈ C.)

The language of L is thus supposed to be a first order one. It is not
necessary to have all the details of this language fixed in advance. We will
– from time to time, if necessary – change some details without causing
confusion by this; but generally, for the set theoretic applications which
later on become of central importance, we suppose there are two sorts of
variables availabe:
– lower case Latin letters like a, b, c, . . . , x, y, z as constants and variables
for elements of the universe of discourse X ;
– upper case Latin letters like A, B, C, . . . , X, Y, Z as constants and vari-
ables for fuzzy subsets of some universe of discourse X .
The basic binary predicate symbol that denotes the membership relation
of elements of the universe of discourse X with respect to some fuzzy subset
of X will be “ ε ” and the atomic formulas built up with it are of type
“a ε A”. Their truth degrees correspond to the membership degrees of the
fuzzy sets, such that for each fuzzy set A ∈ IF (X ) and each x ∈ X one has
the fundamental equation
[[x ε A]] =def µA (x). (1.5)

If necessary, we will add variables and constants for fuzzy relations too.
Furthermore, we need predicate symbols for some many-valued generaliza-
tions of usual set-theoretic relations. These predicate symbols will have al-
most the same graphical shape for fuzzy sets as have the analogous ones for
crisp sets; by use of ε they will normally be introduced through definitions
in the same style as is usually done in classical set theory.
In a few cases we are interested to have (propositional) constants for the
truth degree 1, i.e. for any logically valid expression, as well as for the truth
degree 0. We use > as a constant for the truth degree 1 and ⊥ as a constant
for the truth degree 0.
Our choice of truth degree 1 as the only designated one has as a conse-
quence that we consider a formula H as (logically) valid, denoted: |= H, iff
H has (for all values of its free variables) truth degree 1:
|= H iff [[H]] = 1. (1.6)
And a further consequence of this choice is that a rule of inference

H1 , . . . , Hn
H
with the premises H1 , . . . , Hn and the conclusion H is correct iff H has truth
degree 1 in each situation in which all the premisses H1 , . . . , Hn have truth
degree 1.
Obvious examples of correct rules of inference are the rule of generaliza-
tion
H(x)
(1.7)
∀xH(x)
and the rule of particularization
H(x/a)
(1.8)
∃xH(x)
for first order many-valued logic. The correctness proofs here consist in the
simple remarks that
if |= H(x) then |= ∀xH(x)
as well as
if |= H(x/a) then |= ∃xH(x)

1.2. T-NORMS AND Φ-OPERATORS 5
hold true for every well-formed formula H with x among its free variables
and any element a of the corresponding universe.
Of course, the rule of generalization also holds true for each possible
restricted universal quantification (1.3) and the rule of particularization for
all those restricted existential quantifications (1.4) of the form “∃C x” with
a ∈ C.
To avoid too many parentheses, the usual kind of ranking of the binding
strength of the logical constants will be used: the (restricted as well as un-
restricted) quantifiers bind stronger than the negations, which bind stronger
than the conjunctions and disjunctions, and those connectives finally bind
stronger than any implication or biimplication connective.
In our classical metalanguage we use ⇒ for (material) implication, ⇔
V W
for biimplication, and sometimes for universal quantification and for
existential quantification.
1.2 t-norms and Φ-operators

Definition 1.1 (a) A binary operation t in the real interval [0, 1] is a
t-norm iff it is
(T1) associative and commutative;
(T2) non-decreasing in the first – and hence in each – argument;
(T3) has 1 as neutral element, i.e. u t 1 = u for each u ∈ [0, 1].
(b) A binary operation s in [0, 1] is a t-conorm iff it is

(S1) associative and commutative;
(S2) non-decreasing in each argument;
(S3) has 0 as neutral element, i.e. u s 0 = u for each u ∈ [0, 1].
The t-norms – a common shorthand term for “triangular norms” – have

been widely used in investigations into probabilistic metric spaces, cf. e.g.
Schweizer/Sklar (1961, 1983), Ling (1965), Frank (1979). From
those investigations and also independent of them, it can be argued that the
t-norms are suitable candidates for conjunctions in many-valued logic, and
the t-conorms candidates for disjunctions. This point of view is explained
in the next section.
Concerning notation we always will feel free, for t-norms as well as t-
conorms and later on for Φ-operators too, to use them either in prefix or
in infix notation, i.e. to take them either as two-place functions in [0,1] or

as binary operations in [0,1]. That means for a t-norm t that one always
has: u t v = t(u, v), and that the type of notation is changed without further
comment according to which notation seems to be best in the particular
context.
In Table 1.1 we will now list some t-norms which are usually main ex-
amples or which are members of well known families of t-norms.
t-norm t formula for t(u, v)

tG min{u, v}
tL max{0, u + v − 1}
tP u·v

 min{u, v} for u = 1
tD or v = 1

0 otherwise
uv
tH
γ, γ≥0
γ + (1 − γ) · (u + v − uv)
tY
p, p>0 1 − min{((1 − u)p + (1 − v)p )1/p , 1}
1
tD
λ, λ>0 1−
1+ (( 1−u
u )
λ + ( 1−v λ 1/λ
v ) )
u + v − 1 + λuv
tW
λ , λ > −1 max{0, }
1+λ
Table 1.1: Examples of t-norms1
Generally, the t-norms and the t-conorms can be reduced one to another.
1
The last four families of t-norms each aim to cover a whole “interval” of possible t-
norms between borderline cases given by the extremal values of the respective parameters.
They have been introduced by Hamacher (1978), Yager (1980), Dombi (1982) and
Weber (1983). The upper index in each case points to the name of the defining author.
Pictorial representations of those t-norms and even more families of t-norms are presented
in Mizumoto (1989).
For, given any t-norm t, by

st (u, v) =def 1 − t(1 − u, 1 − v) (1.9)
a t-conorm st is defined; and given any t-conorm s by
ts (u, v) =def 1 − s(1 − u, 1 − v) (1.10)
a t-norm ts is defined. Furthermore, it is easy to see that this is a 1-1
correspondence between t-norms and t-conorms because of
(t)s = t, (s)ts = s (1.11)
t
for each t-norm t, t-conorm s.
As a consequence of (1.11) one can either focus attention on the t-norms
or on the t-conorms: in any case it is not necessary to treat both types of
functions à par. We have decided here to take the t-norms to be of greater
importance. Hence, in the following all the t-conorms s will be considered
as t-conorms of the form st defined via some t-norm t.
For the t-norms, as specific real functions, one has the natural pointwise
partial ordering of functions which as a reflexive ordering is defined by
t1 <
= t2 =def t1 (u, v) ≤ t2 (u, v) for all u, v ∈ [0, 1] (1.12)
for any t-norms t1 , t2 and as an irreflexive ordering by
t1 < t2 =def t1 <
= t2 and t1 6= t2 . (1.13)
Because from the defining properties (T2), (T3) one has 0 ≤ u t v ≤ u t 1 = u
as well as 0 ≤ u t v ≤ 1 t v = v for any t-norm t and all u, v ∈ [0, 1], that
means all together that
tD < <
= t = tG for each t-norm t. (1.14)
Thus for the first four examples of t-norms from Table 1.1 there holds fur-
thermore
tD < tL < tP < tG (1.15)
as can be seen almost directly.
The same partial orderings (1.12), (1.13) also apply to t-conorms. And
for t-conorms their orderings are dual to those for the corresponding t-norms
in the sense that for all t-norms t1 , t2 one has
st1 < <
= st2 ⇔ t2 = t1 (1.16)
and hence also
st1 < st2 ⇔ t2 < t1 .
Therefore it suffices to consider these orderings for t-norms only.

Of course, equations (1.9) and (1.10) correspond to the usual de Mor-
gan laws which connect conjunction and disjunction: the essential point
is that here the function x 7→ 1 − x has to be interpreted as describing a
generalized negation, but really it is the truth function of the “standard”
negation ¬.
This function x 7→ 1 − x is not the only candidate for a generalized
negation; but it is the most commonly considered one. Yet, e.g. Weber
(1983) and Klement (1982) discuss other negation functions too. Following
Weber (1983) we give
Definition 1.2 A unary operation n in the real interval [0, 1] is a negation

function iff the following hold true
(N1) n(0) = 1, n(1) = 0,
(N2) n is non-increasing;
additionally such a negation function will be called strict iff
(N3) n is decreasing and continuous,
finally n is an involution iff n is a strict negation function and fulfills
(N4) n(n(u)) = u for each u ∈ [0, 1].
Clearly, the function x 7→ 1 − x is an involution. Furthermore, the

connections of t-norms and t-conorms established in equations (1.9), (1.10)
can be generalized to any strict negation function. Weber (1983) proved
that for any t-norm t and strict negation function n by
sn,t (u, v) = n−1 (t(n(u), n(v))) (1.17)
a t-conorm is defined; and that for any t-conorm s and strict negation func-
tion n by
tn,s (u, v) = n−1 (s(n(u), n(v))) (1.18)
a t-norm is defined. For involutions n in this way a 1-1 correspondence is

established between t-norms and t-conorms.
For the many-valued connectives having these t-norms, t-conorms or
negation functions as truth functions, clearly equations (1.17) and (1.18)
validate corresponding deMorgan laws.
t st (u, v)
tG max{u, v}
tL min{1, u + v}
tP u+v−u·v
½
tD max{u, v} for u · v = 0
1 otherwise
u + v − uv − (1 − γ)uv
tH
γ, γ≥0
1 − (1 − γ)uv
tY
p, p>0 min{(up + v p )1/p , 1}
1
tD
λ, λ>0 u λ v λ −1/λ
1+ (( 1−u ) + ( 1−v ) )
tW
λ , λ > −1 min{u + v + λuv, 1}
Table 1.2: Examples of t-norms and their t-conorms
In Table 1.2 we have listed for the t-norms of Table 1.1 the t-conorms
related to them via (1.9).
Besides the truth functions for conjunction, disjunction and negation in
many-valued logic we will have to consider truth functions for implications.
That role will be played by the Φ-operators. In section 1.3 we will discuss
logically valid formulas of a language of many-valued logic which has just
these types of truth functions interpreting their propositional connectives.
The results there will indicate that the t-norms are suitable candidates for
generalized, i.e. many-valued conjunction operators and that the Φ-operators
are equally suitable candidates for many-valued implication operators.
Definition 1.3 A binary operation ϕ in the real unit interval [0, 1] is called
Φ-operator (connected with a given t-norm t) iff for all u, v, w ∈ [0, 1] the
following hold true
(Φ1) v ≤ w ⇒ u ϕ v ≤ u ϕ w;
(Φ2) u t (uϕv) ≤ v;
(Φ3) v ≤ u ϕ (u t v).
The Φ-operators have been introduced by Pedrycz (1982) to describe

solutions of fuzzy equations which are formulated using the t-norm which the
Φ-operator is connected with. They generalize the α-operation of Sanchez
(1976, 1977), which now becomes the special case of the Φ-operator con-
nected with the t-norm min.
As a side remark let us mention that the definition of Φ-operators as
given by Pedrycz (1982) used the condition
max{u ϕ v, u ϕ w} ≤ u ϕ max{v, w}
which obviously is equivalent to our (Φ1).

The character of the connection between a Φ-operator and its corre-
sponding t-norm is not made completely clear by (Φ1) to (Φ3). It was in
Gottwald (1984, 1986) where this connection was proved in full generality.
The results are given in the following propositions 1.1 and 1.3.
Proposition 1.1 (Representation Lemma) If ϕ is a Φ-operator which

is connected to the t-norm t, then for all u, v ∈ [0, 1] it holds true that
u ϕ v = sup{w | u t w ≤ v}. (1.19)
Proof. By (Φ2) we immediately get
u ϕ v ≤ sup{w | v t w ≤ v}.
But in case u ϕ v < sup{w | v t w ≤ v} there would exist w0 such that u ϕ w0

and u t w0 ≤ v, therefore
w0 ≤ u ϕ (u t w0 ) ≤ u ϕ v < w0 ,
which is a contradiction. Thus our representation lemma is proved. QED
Therefore, one Φ-operator at most is connected to each t-norm, which in

case of existence can be defined using formula (1.19).
Furthermore, the supremum in (1.19) really is a maximum because by
(Φ2) we have u t (u ϕ v) ≤ v and thus
u t sup{w | u t w ≤ v} = u t (u ϕ v) ≤ v.
Hence it is always even

u ϕ v = max{w | u t w ≤ v}.
But what about the existence of a Φ-operator for a given t-norm? Does
such a Φ-operator always exist, i.e. is the binary operation ϕ defined via
(1.19) a Φ-operator for each t-norm? To formulate the answer we need an
additional property of t-norms, their left continuity or their lower semicon-
tinuity (in each of the arguments).
Because of the commutativity (T1) of the t-norms it is enough to consider
these properties for the first argument only. The left continuity of a t-
norm t is then, of course, the condition that for all u0 , v0 ∈ [0, 1] and all
convergent sequences (ui )i≥1 of points from [0, 1] with limi→∞ ui = u0 and
always ui < u0 one has
lim (u0 t v0 ) = u0 t ( lim ui ).
i→∞ i→∞
And the lower semicontinuity of t means that for each u0 , v0 ∈ [0, 1] and each
² > 0 there is a δ > 0 such that u t v0 > u0 t v0 − ² for all u ∈ (u0 − δ, u0 ].
Both these notions coincide for t-norms.
Proposition 1.2 A t-norm is left continuous iff it is lower semicontinuous.
Proof. More general, for each monotonically non-decreasing function the

lower semicontinuity and the left continuity coincide. Hence suppose that f
is monotonically non-decreasing.
Assume first that f is lower semicontinuous, x ∈ dom (f ) and (xi )i<∞
a converging sequence of points xi ∈ dom (f ) with xi < x for all i. Then
limi→∞ f (xi ) exists and one has limi→∞ f (xi ) ≤ f (x). In the case that
limi→∞ f (xi ) < f (x) let 2² = f (x) − limi→∞ f (xi ). Then in each neighbour-
hood of x there is some xk with f (xk ) < f (x) − ² contradicting the lower
semicontinuity of f . Thus limi→∞ f (xi ) = f (x) and thus f is left continuous
(at x and hence in the whole domain).
Now suppose that f is left continuous. Consider x0 ∈ dom (f ) and some
² > 0. Then there exists some δ > 0 such that |f (x0 ) − f (x)| < ² for all
x ∈ dom (f ) with x0 − x < δ. By monotonicity of f hence f (x0 ) − f (x) < ²,
i.e. f (x) > f (x0 ) − ² and thus f is lower semicontinuous (at x0 and hence in
dom (f )).
Because each t-norm by (T2) is monotonically non-decreasing in each
argument the lower semicontinuity of a t-norm (in one and hence both of its
arguments) coincides with their left continuity. QED
In the following we write

LSC(t) for t is lower semicontiuous.
The characterization of the lower semicontinuity for t-norms we most often
need later on is a simple corollary of the last proof and reads:
LSC(t) ⇔ (1.20)
u t (sup vξ ) = sup(u t vξ ) for all u, (vξ )ξ∈Ξ from [0, 1].
ξ∈Ξ ξ∈Ξ
In a few cases it is not the lower semicontinuity of a t-norm t that is

needed, but the upper semicontinuity2 or, which again is the same, the right
continuity of t. (This equivalence can be proven as in the last proof.) We
write USC(t) if t is upper semicontinuous. Again, the crucial property we
usually need in case USC(t) is that
USC(t) ⇔ (1.21)
u t (inf vξ ) = inf (u t vξ ) for all u, (vξ )ξ∈Ξ from [0, 1].
ξ∈Ξ ξ∈Ξ
Now we are able to solve the problem of the existence of a Φ-operator

with respect to a t-norm t.
Proposition 1.3 (Existence Lemma) For a t-norm t there exists a Φ-

operator ϕ connected with t iff t is lower semicontinuous.
Proof. Assume LSC(t). Let ϕ be defined by (1.19). Obviously this binary

operator ϕ in [0,1] is non-decreasing in the second argument, i.e. fulfills (Φ1).
By LSC(t) one has always
u t (u ϕ v) = u t (sup{w | u t w ≤ v})
= sup{u t w | u zt w ≤ v} ≤ v
and thus (Φ2). And for (Φ3) one immediately has
u ϕ (u t v) = sup{w | u t w ≤ u t v} ≥ v.
Now suppose that LSC(t) fails. Then there exist in [0, 1] real numbers
u0 , v0 and a family (vξ )ξ∈Ξ of such numbers with
v0 = sup(u0 t vξ ) < u0 t (sup(vξ )).
ξ∈Ξ ξ∈Ξ
2
The upper semicontinuity of t means that for each u0 , v0 ∈ [0, 1] and each ² > 0 there
is a δ > 0 such that u t v0 < u0 t v0 + ² for all u ∈ [u0 , u0 + δ).
If we assume that there exists a Φ-operator ϕ connected to t then we have

from the representation lemma that
u0 t (u0 ϕ v0 ) = u0 t sup{w | u0 t w ≤ v0 } ≥ u0 t (sup(vξ )) > v0 ,
ξ∈Ξ
contradicting (Φ2). Hence in that case where LSC(t) fails there does not
exist a Φ-operator connected with t. QED
Definition 1.4 We denote the uniquely determined Φ-operator which is

connected via (1.19) with each lower semicontinuous t-norm t by ϕt , i.e.
we put
u ϕt v = sup{w | u t w ≤ v}. (1.22)
In Table 1.3 we give a list of Φ-operators which correspond via (1.22)

to t-norms of Table 1.1. Here the Φ-operator ϕtL is the truth function of
the well known L Ã ukasiewicz implication of many-valued logic and ϕtG is
related to another such implication operator, often named after K. Gödel3 .
From formulas (1.19), (1.22) it is easy to prove some fundamental prop-
erties of every Φ-operator.
Proposition 1.4 For each binary operator ϕ defined by formula (1.19) it

holds true for all u, v, w:
(i) u ≤ v ⇒ u ϕ w ≥ v ϕ w;
(ii) u ≤ v ⇒ u ϕ v = 1;
(iii) 1 ϕ v = v;
(iv) v ≤ u ϕ v.
Proof. Obvious from (1.19).
Proposition 1.5 For every t-norm t with property LSC(t) one has for all
u, v, w:
(i) u ϕt (v ϕt (u t v)) = 1;
(ii) (u t v) ϕt w ≤ u ϕt (v ϕt w);
(iii) u ϕt v ≤ (u t w) ϕt (v t w).
3
The implication operator of many-valued logic with that truth function was introduced
in Gödel (1932) in connection with his studies on intuitionistic logic. There he also used
a negation which was related to his implication by just the same relation we shall use later
on in definition 1.8 in general to introduce a negation function for each t-norm.
t ϕt (u, v)
½
1 for u ≤ v
tG
v for u > v
tL min{1, 1 − u + v}
½
tP min{1, v/u} for u 6= 0
1 for u = 0
v + (γ − 1)(1 − u)v
tH
γ, γ≥0 for u > v 4
u + (γ − 1)(1 − u)v
tY
p, p>0 1 − ((1 − v)p − (1 − u)p )1/p for u > v
1
tD
λ, λ>0 for u > v
1+ (( 1−v λ − ( 1−u λ 1/λ
v ) u ) )
1 − u + v + λv
tW
λ , λ > −1 for u > v
1 + λu
Table 1.3: Examples of t-norms and their Φ-operators 5
Proof. (i) is obvious from (Φ2) and proposition 1.4 (ii). To get (ii) we
observe that by (1.22) it is enough to prove
u t ((u t v) ϕt w) ≤ v ϕt w,
and for this again by (1.22) it is enough to prove
v t (u t ((u t v) ϕt w)) ≤ w
which is true because of (Φ2) and the associativity and commutativity of t.
For (iii) it is, again by (1.22), enough to prove
u t s ≤ v ⇒ (u t t) t s ≤ v t t
which is obvious by associativity and monotonicity of t, i.e. by conditions
(T1), (T2). QED
4
Here and in the following lines in case u ≤ v the result is always 1 as in the first line.
5
Of course here the drastic product tD has to fail because tD is not lower semicontin-
uous, i.e. not left continuous for each point (1, v0 ) of [0, 1]2 with v0 < 1.
Proposition 1.6 Suppose LSC(t). Then it holds true for all u, v, w ∈ [0, 1]:
(i) u ≤ v ⇔ u ϕt v = 1;
(ii) (u t v) ϕt w = u ϕt (v ϕt w);
(iii) (u ϕt v) t (v ϕt w) ≤ u ϕt w.
Proof. (i) Because of proposition 1.4 (i) only (⇐) has to be proved. Hence
suppose
1 = u ϕt v = sup{w | u t w ≤ v}.
Then by (T2) we get
w < 1 ⇒ utw ≤ v
and hence
u = u t 1 = u t sup{w | w < 1} = sup{u t w | w < 1} ≤ v.
(ii) Because of proposition 1.5 (ii) it is enough to prove
u ϕt (v ϕt w) ≤ (u t v) ϕt w.
And by (1.22) for this inequalitiy only
u ts ≤ v ϕt w ⇒ (u t v) t s ≤ w
has to be proved. Hence assume that u ts ≤ v ϕt w; in this case we have
(u t v) t s = (v t u) t s ≤ v t (v ϕt w) ≤ w
by (T1), (T2) and (Φ2).

(iii) By LSC(t) and formula (1.22) we get in a straightforward manner
(u ϕt v) t (v ϕt w) = sup{s t t | u t s ≤ v and v t t ≤ w}
≤ sup{r | u t r ≤ w} = u ϕt w
because of the fact that one has
u t (s t t) = (u t s) t t ≤ v t t ≤ w
from u t s ≤ v and v t t ≤ w. QED
If we even suppose that our t-norms are continuous functions, then the
following two interesting additional results become provable.
Proposition 1.7 For each continuous t-norm t the following hold true
(i) u t (u ϕt v) = u ∧ v,
(ii) v ϕt (u t (u ϕt v)) = v ϕt u.
Proof. (i) We have u t (u ϕt v) ≤ v for every u, v and u t u ≤ u for every u.

From this and
_
u t (u ϕt v) = {u t u | u t u ≤ v}
for continuous t, (i) follows.
(ii) Immediately from (i) we have
v ϕt (u t (u ϕt v)) = v ϕt (u ∧ v) = v ϕt u
which is obvious for u ≤ v, but if v < u we have 1 = v ϕt v = v ϕt u. QED
Using the Φ-operators we are able to introduce a special kind of negation

function with respect to each t-norm t with the property LSC(t) by the
following definition.
Definition 1.5 Suppose LSC(t). Then for each u ∈ [0, 1] let be

nt (u) =def ϕt (u, 0).
Corollary 1.8 For each t-norm t with the property LSC(t) the function nt :
[0, 1] → [0, 1] is a negation function.
Proof. Obviously nt (0) = 1 and nt (1) = 0 hold true. And from proposition
1.1 it follows that each such function nt is non-increasing. QED
In general, however, nt need not be a strict negation function or even

an involution.
There are some additional properties in the class of t-norms which deserve
to be mentioned for the question of the choice of a specific t-norm.
Definition 1.6 A t-norm t is called idempotent iff u t u = u for all u ∈

[0, 1]; and it is called interactive iff there exist some u0 , v0 ∈ [0, 1] such that
u0 t v0 ∈/ {u0 , v0 }. This terminology is used for t-conorms in the same sense.
Corollary 1.9 (i) A t-norm t is idempotent iff its corresponding t-conorm

st is idempotent.
(ii) A t-norm t is interactive iff its corresponding t-conorm st is inter-
active.
Proof. Both claims are immediate consequences of the corresponding defi-

nitions. (For (ii) the simpler proof results if one proves the equivalent claim:
t is non-interactive iff st is non-interactive.) QED
It is, therefore, enough to discuss t-norms with respect to interactivity

and idempotency. The same results then hold true for t-conorms too.
Proposition 1.10 (i) The only idempotent t-norm is the t-norm t = min.
(ii) The only non-interactive t-norm is the t-norm t = min.
Proof. (i) Suppose that t is an idempotent t-norm. Consider any u, v ∈ [0, 1]

and suppose u ≤ v. Then one has
u = utu ≤ utv ≤ ut1 = u
by (T2), (T3) and the idempotency of t. Therefore u t v = min{u, v} in this
case. And for v ≤ u the same result holds true by symmetry.
(ii) Assume that t is non-interactive. Then for each u ∈ [0, 1] one has
u t u ∈ {u} and thus u t u = u. Hence t is idempotent and therefore t = min
by (i). QED
Proposition 1.11 (i) If a t-norm t is distributive with respect to its corre-

sponding t-conorm st then t = min.
(ii) If a t-conorm st is distributive with respect to its corresponding t-
norm t then t = min.
Proof. (i) Assume that this kind of distributivity holds, i.e. that for all
u, v, w ∈ [0, 1]
u t (v st w) = (u t v) st (u t w)
holds true. Taking v = w = 1 gives
u = u st u for all u ∈ [0, 1],
i.e. the idempotency of st . By corollary 1.9 then t is also idempotent and
hence t = min.
(ii) Now the distributivity condition is that for all u, v, w ∈ [0, 1] there
holds true
u st (v t w) = (u st v) t (u st w).
This time choosing v = w = 0 gives

u = u t u for all u ∈ [0, 1],
hence again the idempotency of t and thus t = min. QED
Distributivity of a t-norm with respect to its t-conorm or of a t-conorm

with respect to its t-norm is one aspect of distributivity between generalized
conjunction and disjunction connectives only. Other ones relate e.g. different
t-norms, or a t-norm with a t-conorm of another t-norm etc. We will not (try
to) discuss this topic in full generality but mention only some very special
cases in which one of the involved t-norms is t = tG = min. Remembering
the notational convention (1.26), by abuse of language we will also write
here ∧ = tG = min and ∨ = stG = max.
Proposition 1.12 For all u, v, w ∈ [0, 1] and any t-norm t there hold true
the distributive laws
(i) u t (v ∧ w) = (u t v) ∧ (u t w),
(ii) u t (v ∨ w) = (u t v) ∨ (u t w),
(iii) u st (v ∧ w) = (u st v) ∧ (u st w),
(iv) u st (v ∨ w) = (u st v) ∨ (u st w)
as well as the subdistributive laws
(v) u ∧ (v t w) ≥ (u ∧ v) t (u ∧ w),
(vi) u ∨ (v t w) ≥ (u ∨ v) t (u ∨ w),
(vii) u ∧ (v st w) ≤ (u ∧ v) st (u ∧ w),
(viii) u ∨ (v st w) ≤ (u ∨ v) st (u ∨ w).
Proof. Claims (i) – (iv) are direct consequences of the definitions of t-norms
and their t-conorms. The claims (v) – (viii) follow by applying the respective
distributivity result from (i) – (iv) to the right side of that subdistributivity
and then using simple inequalities for t-norms and t-conorms. QED
1.3 t-norm based connectives

As was already mentioned, the t-norms are considered as truth functions of
generalized conjunction operators in a suitable many-valued logic, and the
1.3. T-NORM BASED CONNECTIVES 19
Φ-operators shall be considered as truth functions of generalized implica-

tion operators. In the present section we will support this point of view by
presenting a series of formulas (both propositional and first order) involving
these new, generalized connectives which under their intended interpreta-
tions become (logically) valid. These formulas, all of them generalizations
of well known logically valid formulas of classical logic, mark the extent to
which t-norms and Φ-operators may count as suitable generalizations of the
corresponding classical, two-valued truth functions. Nevertheless, the for-
mulas discussed further on are not only used to support this point of view,
but, besides this reason, also have their independent value within our later
developments and will be referred to from different proofs and calculations
there.
Throughout this section, H, G, F (sometimes with indices) are used to
denote well formed formulas of our (enriched) language of many-valued logic.
Definition 1.7 For each t-norm t we denote by ∧t the conjunction con-
nective of the language of our system L of many-valued logic that has this
t-norm t as its truth function, similarly we use in the language of L for each
t-norm t a disjunction connective ∨t with truth function st . And for each
t-norm t which fulfills LSC(t) by →t we denote that implication connective
of the language of our system L of many-valued logic that has the Φ-operator
ϕt as its truth function.
Hence we enrich the language of the system L with as many new t-norm
based conjunction, disjunction and implication connectives as seem to be
appropriate in each case. In practice that will usually mean including one
or two t-norms into the considerations and hence adding one or two such t-
norm based connectives of each kind to the language of L. Because of (1.11),
with these notational conventions one also has for any given t-conorm a
disjunction connective with this t-conorm as truth function. Hence if H is
e.g. the conjunction H1 ∧t H2 , we have
[[H1 ∧t H2 ]] = t([[H1 ]], [[H2 ]]) = [[H1 ]] t [[H2 ]]. (1.23)
Correspondingly we have for the semantic interpretation of the implication
connective →t the simple fact that
[[H1 →t H2 ]] = ϕt ([[H1 ]], [[H2 ]]). (1.24)
Sometimes we also use a biimplication operator ↔t connected with a
t-norm t and defined by
H1 ↔t H2 =def (H1 →t H2 ) ∧t (H2 →t H2 ). (1.25)
As we later on occasionally have to discuss separately the strongest t-

norm tG = min, as well as the L Ã ukasiewicz t-norm tL and their corre-
sponding t-conorms, we use as shorthand notations
∧ for ∧min and ∨ for ∨min , (1.26)
& for ∧tL and ] for ∨tL , (1.27)
and by abuse of language from time to time those symbols are also used for
those t-norms and t-conorms themselves.
Proposition 1.13 For each t-norm t which has the property LSC(t) and all
H1 , H2 one has
|= H1 →t H2 iff [[H1 ]] ≤ [[H2 ]], (1.28)
|= H1 ↔t H2 iff [[H1 ]] = [[H2 ]]. (1.29)
Proof. The first claim is a restatement of proposition 1.6 (i) using (1.6), the
second one a consequence of the first using: u t v = 1 ⇔ u = v = 1. QED
Therefore all the logically valid implications of our generalized language

are inequalities between truth degrees “in disguise”, and the logically valid
biimplications accordingly represent equalities for truth degrees.
For simplicity of notation we use the arrows → and ↔ here also without
subscripts: in that case we assume that a subscript “t1 ” has been deleted
that refers to any t-norm t1 with property LSC(t1 ).
Proposition 1.14 For any t-norm t the following expressions are logically
valid:
(i) |= H ∧t > ↔ H and |= H ∧t ⊥ ↔ ⊥,
(ii) |= H ∨t > ↔ > and |= H ∨t ⊥ ↔ H,
(iii) |= H1 ∧t H2 ↔ H2 ∧t H1 ,
(iv) |= H1 ∨t H2 ↔ H2 ∨t H1 ,
(v) |= H1 ∧t (H2 ∧t H3 ) ↔ (H1 ∧t H2 ) ∧t H3 ,
(vi) |= H1 ∨t (H2 ∨t H3 ) ↔ (H1 ∨t H2 ) ∨t H3 .
Proof. Those statements immediately follow from properties (T1) to (T3)

of any t-norm and properties (S1) to (S3) of any t-conorm. QED
Therefore the t-norm based conjunctions and disjunctions are already

by definition – modulo logical validity – commutative as well as associative
operations. As was already seen for t-norms and t-conorms distributivity

properties only hold true for special cases. As an immediate consequence
of the results of proposition 1.12 we mention the following two propositions
without further proof.
Proposition 1.15 For any t-norm t the following equivalences are logically
valid:
(i) |= H1 ∧t (H2 ∧ H3 ) ↔ (H1 ∧t H2 ) ∧ (H1 ∧t H3 ),

(ii) |= H1 ∧t (H2 ∨ H3 ) ↔ (H1 ∧t H2 ) ∨ (H1 ∧t H3 ),
(iii) |= H1 ∨t (H2 ∧ H3 ) ↔ (H1 ∨t H2 ) ∧ (H1 ∨t H3 ),
(iv) |= H1 ∨t (H2 ∨ H3 ) ↔ (H1 ∨t H2 ) ∨ (H1 ∨t H3 ).
Proposition 1.16 For any t-norm t the following implications are logically
valid:
(i) |= (H1 ∧ H2 ) ∧t (H1 ∧ H3 ) → H1 ∧ (H2 ∧t H3 ),

(ii) |= (H1 ∨ H2 ) ∧t (H1 ∨ H3 ) → H1 ∨ (H2 ∧t H3 ),
(iii) |= H1 ∧ (H2 ∨t H3 ) → (H1 ∧ H2 ) ∨t (H1 ∧ H3 ),
(iv) |= H1 ∨ (H2 ∨t H3 ) → (H1 ∨ H2 ) ∨t (H1 ∨ H3 ).
Proposition 1.17 For any t-norm t with property LSC(t) the following ex-
pressions are logically valid:
(i) |= H ∧t (H →t G) → G,
(ii) |= G → (H →t H ∧t G),
(iii) |= H → (G →t H),
(iv) |= (H1 →t H2 ) ∧t G → (H1 →t H2 ∧t G).
Proof. The claims (i) and (ii) follow immediately from the defining proper-
ties (Φ2), (Φ3) of the Φ-operators. (iii) is proposition 1.4 (iv). And to get
(iv) it is enough to have
(u ϕt v) t w ≤ u ϕt (v t w) (1.30)
for any u, v, w ∈ [0, 1]. To see that this last inequality holds true one starts
from
(u ϕt v) t w = w t sup{s | u t s ≤ v} = sup{w t s | u t s ≤ v}.

But with u t s ≤ v one has u t (w t s) = (u t s) t w ≤ v t w and hence
{w t s | u t s ≤ v} ⊆ {z | u t z ≤ v t w}.
Therefore one gets
(u ϕt v) t w ≤ sup{z | u t z ≤ v t w} = u ϕt (v t w)
and thus has (1.30) and hence (iv). QED
The monotonicities of the t-norms and the Φ-operators which were stated
in (T2), (Φ1) and proposition 1.4 (i) can directly be rewritten as the impli-
cations (which, yet, are classical ones, i.e. formulated in the metalanguage):
(a) if |= H1 → H2 then |= H1 ∧t G → H2 ∧t G,
(b) if |= H1 → H2 then |= (G →t H1 ) → (G →t H2 ),
(c) if |= H1 → H2 then |= (H2 →t G) → (H1 →t G).
But it is also possible to generalize them to inequalities between truth de-

grees, that means to logically valid “many-valued” implications.
As in classical logic one has a direct connection between the result of
proposition 1.17 (i) and the inference rule of modus ponens: because of
proposition 1.17 (i) the rule of detachment
H, H →t G
(modus ponens)
G
is correct, for, having given well formed formulas H and H →t G such
that |= H and |= H →t G, then by proposition 1.17 (i) immediately |= G.
Hence any application of the modus ponens as a rule of inference leads from
logically valid formulas again to a logically valid formula.
As a side remark let us mention that each one of the metatheoretical
implications (a),. . . ,(c) which we have just mentioned also provides the cor-
rectness proof for some rule of inference; e.g. because of (b) the rule of
inference
H1 →t H2
(G →t H1 ) →t (G →t H2 )
is correct. Furthermore, each logically valid many-valued implication of the
form
|= H1 ∧t . . . ∧t Hn →t H
yields the correctness of the rule of inference

H1 , . . . , Hn
.
H
Therefore, as corollaries of the following propositions we get the correctness
of a whole series of rules of inference but will not mention this fact in those
future cases separately.
(i) |= (H1 →t H2 ) → (H1 ∧t G →t H2 ∧t G),

(ii) |= (H1 →t H2 ) → ((G →t H1 ) →t (G →t H2 )),
(iii) |= (H1 →t H2 ) → ((H2 →t G) →t (H1 →t G)).
Proof. (i) By the monotonicity condition (T2) one has for each t-norm t
u t z ≤ v ⇒ (u t w) t z = (u t z) t w ≤ v t w
and therefore using the representation lemma
u ϕt v ≤ (u t w) ϕt (v t w).
But that is precisely the statement in its truth functional form.

(ii) What has to be proven is
u ϕt v ≤ (w ϕt u) ϕt (w ϕt v),
that means by the representation lemma
sup{z | u t z ≤ v} ≤ sup{x | (w ϕt u) t z ≤ w ϕt v}.
Therefore the main point is to prove the claim that always
u t z ≤ v ⇒ (w ϕt u) t z ≤ w ϕt v
holds true. Thus assume u t z ≤ v. Then using (1.30) one gets
(w ϕt u) t z ≤ w ϕt (u t z) ≤ w ϕt v
and the proof of (ii) is accomplished.

(iii) In the present case one has to prove that always
u ϕt v ≤ (v ϕt w) ϕt (u ϕt w),
and gets as in part (ii) via the representation lemma as a sufficient condition
for that inequality to hold true to have
u t z ≤ v ⇒ (v ϕt w) t z ≤ u ϕt w
for all u, v, w, z ∈ [0, 1]. Thus assume u t z ≤ v. By LSC(t) and this assump-
tion one has
(v ϕt w) t z = z t sup{x | v t x ≤ w}
= sup{z t x | v t x ≤ w}
≤ sup{z t x | u t z t x ≤ w}
and hence
(v ϕt w) t z ≤ sup{y | u t y ≤ w} = u ϕt w,
i.e. the sufficiency condition. QED
(i) |= H1 →t (H2 →t G) ↔ (H1 ∧t H2 →t G),
(ii) |= (H1 →t H2 ) ∧t (H2 →t H3 ) → (H1 →t H3 ),
(iii) |= (H1 →t G1 ) ∧t (H2 →t G2 ) → (H1 ∧t G1 →t H2 ∧t G2 ),
(iv) |= (H1 →t G1 ) ∧ (H2 →t G2 ) → (H1 ∧ G1 →t H2 ∧ G2 ),
(v) |= (H1 →t G1 ) ∧ (H2 →t G2 ) → (H1 ∨ G1 →t H2 ∨ G2 ).
And in the case that the t-norm t subdistributes over its t-conorm in the
sense that
u t (v st w) ≤ (u t v)st (u t w)
holds true for all u, v, w ∈ [0, 1] then one also has
(vi) |= (H1 →t G1 ) ∧t (H2 →t G2 ) → (H1 ∨t G1 →t H2 ∨t G2 ).
Proof. (i) It is enough to prove for all u, v, w ∈ [0, 1] that always

u ϕt (v ϕt w) = (u t v) ϕt w,
which means by the representation lemma and (T1) to prove
sup{z | u t z ≤ v ϕt w} = sup{z | v t u t z ≤ w}.
To get this equality it is sufficient to always have that
u t z ≤ v ϕt w ⇔ v t u t z ≤ w.
But if u t z ≤ v ϕt w one has by (T2)
v t (u t z) ≤ v t sup{s | v t s ≤ w} = sup{v t s | v t s ≤ w} ≤ w;
and conversely in the case that v t u t z ≤ w one has by (Φ3)
u t z ≤ v ϕt (v t (u t z)) ≤ v ϕt w.
(ii) It is enough to prove for all u, v, w ∈ [0, 1] that always
(u ϕt v) t (v ϕt w) ≤ (u ϕt w).
But using again the representation lemma one has
sup{z t (v ϕt w) | u t z ≤ v} = sup{z t s | u t z ≤ v and v t s ≤ w}

≤ sup{r | u t r ≤ w}
because one has u t (z t s) ≤ w immediately from u t z ≤ v and v t s ≤ w.

Thus (ii) follows.
(iii) From (1.30) one gets
(u ϕt v) t (w ϕt z) ≤ u ϕt (v t (w ϕt z))
≤ u ϕt (w t (v ϕt z)) = (u t w) ϕt (v t z)
as in the proof of (i) and thus also has (iii).

(iv) What we have to prove for any u1 , u2 , v1 , v2 ∈ [0, 1] is that
min{u1 ϕt v1 , u2 ϕt v2 } ≤ (min{u1 , u2 }) ϕt (min{v1 , v2 }).
Without loss of generality we may suppose that u1 ≤ u2 . Then we have to

prove the simpler inequality
min{u1 ϕt v1 , u2 ϕt v2 } ≤ u1 ϕt (min{v1 , v2 }). (1.31)
But if v1 ≤ v2 , the right hand side of (1.31) becomes u1 ϕt v1 and thus is

one of the terms over which on the left hand side the minimum has to be
taken – hence (1.31) holds true in this case. And in the case that v2 < v1 ,
then one has u1 t w ≤ u2 t w for any w ∈ [0, 1] and hence u2 ϕt v2 ≤ u1 ϕt v1 .
Therefore (1.31) becomes
u2 ϕt v2 ≤ u1 ϕt v2
which obviously holds true because of u1 ≤ u2 .

(v) What we now have to prove for any u1 , u2 , v1 , v2 ∈ [0, 1] is
min{u1 ϕt v1 , u2 ϕt v2 } ≤ (max{u1 , u2 }) ϕt (max{v1 , v2 }).
Without loss of generality we may suppose that u1 ϕt v1 ≤ u2 ϕt v2 . If then

for some w ∈ [0, 1] one has u1 t w ≤ v1 , so also u2 t w ≤ v2 ; because otherwise
in case v2 < u2 t w one would have w > u2 ϕt v2 and thus u1 ϕt v1 > u2 ϕt v2 ,
a contradiction.
Therefore we now have min{u1 ϕt v1 , u2 ϕt v2 } = u1 ϕt v1 and further-
more
u1 ϕt v1 = sup{w | u1 t w ≤ v1 }
≤ sup {w | max{u1 t w, u2 t w} ≤ max{v1 , v2 }}
= sup {w | max{u1 , u2 } t w ≤ max{v1 , v2 }}
= (max{u1 , u2 }) ϕt (max{v1 , v2 }),
hence all is revealed.

(vi) This time we have to prove for any u1 , u2 , v1 , v2 ∈ [0, 1] that
(u1 ϕt v1 ) t (u2 ϕt v2 ) ≤ (u1 st u2 ) ϕt (v1 st v2 ).
Because of LSC(t) we first have
(u1 ϕt v1 ) t (u2 ϕt v2 )
= sup{w | u1 t w ≤ v1 } t sup{z | u2 t z ≤ v2 }
= sup {w t sup{z | u2 t z ≤ v2 } | u1 t w ≤ v1 }
= sup { sup{w t z | u2 t z ≤ v2 } | u1 t w ≤ v1 }
= sup{w t z | u1 t w ≤ v1 and u2 t z ≤ v2 }
≤ sup{w t z | (u1 t w) st (u2 t z) ≤ v1 st v2 }.
By our subdistributivity assumption we now get
(u1 t w) st (u2 t z) ≥ (u1 st u2 ) t (w t z)
for all w, z ∈ [0, 1], for we may, by symmetry, additionally suppose that
w ≥ z and then have the following sequence of inequalities
(u1 t w) st (u2 t z) ≥ (u1 t z) st (u2 t z)

≥ (u1 st u2 ) t z
≥ (u1 st u2 ) t (w t z).
Thus we have
(u1 st u2 ) ϕt (v1 st v2 ) = sup{q | (u1 st u2 ) t q ≤ v1 st v2 }

≥ sup{w t z | (u1 st u2 ) t (w t z) ≤ v1 st v2 }
≥ sup{w t z | (u1 t w) st (u2 t z) ≤ v1 st v2 }.
All these inequalities together prove the inequality we had to establish in

this case. QED
The standard connections (1.9), (1.10) between t-norms and t-conorms

can now be restated as general de Morgan laws for ∧t and ∨t with respect
to the (standard) negation ¬.
valid:
(i) |= ¬(H1 ∧t H2 ) ↔ (¬H1 ∨t ¬H2 ),

(ii) |= ¬(H1 ∨t H2 ) ↔ (¬H1 ∧t ¬H2 ).
Proof. Obvious.
The negation functions nt which were introduced in definition 1.5 can

now also be used to consider t-norm based negations.
Definition 1.8 For any t-norm t which has the property LSC(t) let −t be
that (unary) negation connective which has nt as its truth function, that
means always put
[[−t H]] = nt ([[H]]).
Proposition 1.21 For any t-norm t with property LSC(t) and all formulas
H it holds true
|= −t H ↔ (H →t ⊥).
Proof. Obvious by comparing the respective truth functions. QED
This way of connecting a negation operator with an implication operator

was for example used by K. Gödel (1932) in his studies on intuitionistic
logic, and was also present in the work of the L
Ã ukasiewicz group in their
development of many-valued logic; cf. LÃ ukasiewicz/Tarski (1930).
It is interesting to note that for the L

Ã ukasiewicz conjunction & with
truth function tL the negation −tL is nothing other than the standard nega-
tion ¬. And furthermore in that special case one also has
|= (H →tL G) ⇔ (¬H ∨tL G).
Proposition 1.22 For any t-norm t with the property LSC(t) the following
expressions are logically valid:
(i) |= H → −t −t H,
(ii) |= −t H ↔ −t −t −t H.
Proof. (i) is an immediate corollary of the representation lemma and the

property LSC(t).
(ii) Because of (i) one only has to show that
((u ϕt 0) ϕt 0) ϕt 0 ≤ u ϕt 0 (1.32)
for all u ∈ [0, 1]. By the representation lemma and LSC(t) one has
((u ϕt 0) ϕt 0) ϕt 0 = sup{w | sup{v | (u ϕt 0) t v ≤ 0} t w ≤ 0}
= sup{w | sup{v t w | (u ϕt 0) t v ≤ 0} ≤ 0}.
Now suppose that w0 ∈ [0, 1] is one of the values which belong to the set
whose supremum has to be considered here, i.e. is chosen such that
sup{v t w0 | (u ϕt 0) t v ≤ 0} ≤ 0.
Then for each v ∈ [0, 1]
sup{z t v | u t z ≤ 0} = (u ϕt 0) t v ≤ 0 ⇒ v t w0 ≤ 0,
which means that for all z, v ∈ [0, 1] one has
(u t z ≤ 0 ⇒ z t v ≤ 0) ⇒ v t w0 ≤ 0. (1.33)
Assuming
w0 > u ϕt 0 = sup{z | u t z ≤ 0}
then gives w0 > 0 and u t w0 > 0, i.e. (u t w0 ≤ 0 ⇒ w0 t v ≤ 0) holds true
for any v and hence v t w0 ≤ 0 for each v by (1.33). For v = 1 this gives
w0 = 1 t w0 ≤ 0 and thus a contradiction. Hence w0 ≤ u ϕt 0 and thus (1.32)
by the choice of w0 . QED
Proposition 1.23 For any t-norm t with the property LSC(t) the following
expressions are logically valid:
(i) |= −t (H ∧t −t H),
(ii) |= H →t (−t H →t G),
(iii) |= (H →t −t G) ↔ (G →t −t H),
(iv) |= (H →t G) → (−t G →t −t H),
(v) |= (H →t G) ∧t (H →t −t G) → −t (H ∧t H).
Proof. (i) immediately follows from u t (u ϕt 0) ≤ 0, which is an instance of

(Φ2). (ii) follows from (i) using proposition 1.19 (i) together with 0 ϕt v = 1
for any v.
(iii) can be split into two problems, the first one to prove
|= (H →t −t G) → (G →t −t H),
or equivalently to prove
|= (H →t −t G) →t (G →t −t H), (1.34)
and the second one to prove that the reverse implication from (1.34) is valid
too. Because of the symmetry of both problems it is enough only to prove
(1.34). Using proposition 1.19 (i) the problem (1.34) becomes to prove
|= (H →t −t G) ∧t G →t −t H, (1.35)
but with proposition 1.17 (iv) one has
[[(H →t −t G) ∧t G]] ≤ [[H →t −t G ∧t G]] ≤ [[H →t ⊥]] = [[−t H]]
from definition 1.8, which is (1.35). Hence (1.34) is proven.

(iv) can equivalently be stated in the form (1.34) but with G and −t G
exchanged. Thus the same calculations as in (iii) yield the proof.
Finally (v) is an easy consequence of proposition 1.19 (iii) using (i),
i.e.u̇sing [[G ∧t −t G]] = 0. QED
Corollary 1.24 For the t-norm t = tL contraposition holds true in the more
traditional form
|= (H1 →tL H2 ) ↔ (¬H2 →tL ¬H1 ),

and also for each t-norm t with the property LSC(t), for which additionally
there holds true |= (−t −t H ↔ H), even
|= (H1 →t H2 ) ↔ (−t H2 →t −t H1 )
holds true.
Proof. Because of |= (¬¬H ↔ H) both cases can be proven by the same
argument: simply substitute ¬G resp. −t G for G in proposition 1.23 (iii)
and then delete the double negations. QED
The foregoing discussions of the present section were all concerned with
propositional many-valued logic. That is by far the most important topic
for t-norm based logical operators. One could also try, of course, to extend
these t-norm-based considerations to the realm of quantifiers and to define
new t-norm based universal quantifiers which for interpretations with finite
universes reduce to iterations of a t-norm-based conjunction connective ∧t ,
as well as to extend them to new t-norm-based existential quantifiers, which
under the same restrictions concerning the interpretations reduce to an it-
eration of a t-conorm-based disjunction connective. But at present it seems
that such a generalization strategy would not produce interesting quanti-
fiers. One of the main difficulties, it seems, comes from the fact that all
the t-norms besides t = min are not idempotent and that therefore the
quantifiers based on such non-idempotent t-norms behave quite differently
in different interpretations depending on the cardinalities of the respective
universes.
In any case, finite iterations of these t-norm-based connectives, mainly
conjunction connectives ∧t and disjunction connectives ∨t , can be used in
the usual way through inductive generalizations. Only in a few situations
later will on the need arise for infinitary generalizations of such proposi-
tional connectives. In those cases another way round shall be taken instead
of introducing new quantifiers: infinitely long expressions (with a definite
length) will then be used. This method proves to be a suitable substitute to
use instead of giving the above mentioned generalization of the quantifiers.
Thus we restrict the considerations to those cases which combine the
usual quantifiers of many-valued first order logic as defined in (1.1), (1.2),
and in (1.3), (1.4) in their bounded form, with t-norm based propositional
connectives.
For the LÃ ukasiewicz negation −tL = ¬ one immedately has generaliza-
tions of the well known de Morgan laws connecting those quantifiers in
classical logic.
Proposition 1.25 The following expressions are logically valid:

(i) |= ¬∀xH(x) ↔ ∃x¬H(x),
(ii) |= ¬∃xH(x) ↔ ∀x¬H(x),
(iii) |= ∀xH(x) ↔ ¬∃x¬H(x),
(iv) |= ∃xH(x) ↔ ¬∀x¬H(x).
Proof. Straightforward from the corresponding definitions.
valid:
(i) |= ∀xH1 (x) ∧t ∀xH2 (x) → ∀x(H1 (x) ∧t H2 (x)),
(ii) |= ∃x(H1 (x) ∧t H2 (x)) → ∃xH1 (x) ∧t ∃xH2 (x),
(iii) |= ∀xH1 (x) ∨t ∀xH2 (x) → ∀x(H1 (x) ∨t H2 (x)),
(iv) |= ∃x(H1 (x) ∨t H2 (x)) → ∃xH1 (x) ∨t ∃xH2 (x).
Proof. (i) For each a ∈ X one has [[∀xHi (x)]] ≤ [[Hi (a)]] for i = 1, 2 and
thus
[[∀xH1 (x) ∧t ∀xH2 (x)]] ≤ [[H1 (a) ∧t H2 (a)]].
Hence for any variable “y” which does not appear (either at all or at least
not free) in “∀xH1 (x) ∧t ∀xH2 (x)” one has
[[∀xH1 (x) ∧t ∀xH2 (x)]] ≤ inf [[H1 (y) ∧t H2 (y)]].
y∈X
And this is exactly (i).

(ii) is proven along the same lines but starting from the inequalities
[[Hi (a)]] ≤ [[∃xHi (x)]].
(iii) as well as (iv) follow via corollary 1.24 from (ii) and (i) respectively,
and the propositions 1.25 and 1.20 by simple, direct transformations. QED
pressions are logically valid, if additionally G does not contain the variable
x free:
(i) |= ∀x(G →t H(x)) ↔ (G →t ∀xH(x)),
(ii) |= ∀x(H(x) →t G) ↔ (∃xH(x) →t G),
(iii) |= ∃x(G →t H(x)) ↔ (G →t ∃xH(x)),
(iv) |= ∃x(H(x) →t G) ↔ (∀xH(x) →t G).
Proof. We prove only (i) and (ii) because the rest can be proven with
exactly the same methods.
(i) To get the needed equality
[[∀x(G →t H(x))]] = [[G →t ∀xH(x)]]
for the truth degrees we show two corresponding inequalities. By the defi-
nition (1.1) of [[∀ . . .]] as an infimum one has
[[∀x(G →t H(x))]] ≤ [[G →t H(a)]]
for each a ∈ X and thus
|= ∀x(G →t H(x)) →t (G →t H(a)).
Using proposition 1.19 (i) that gives
|= ∀x(G →t H(x)) ∧t G →t H(a)
and hence
[[∀x(G →t H(x)) ∧t G]] ≤ [[H(a)]],
which means for any variable “y” which does not appear in the well-formed
formula “∀x(G →t H(x)) ∧t G”:
[[∀x(G →t H(x)) ∧t G]] ≤ inf [[H(y)]] = [[∀xH(x)]].
y∈X
Therefore
|= ∀x(G →t H(x)) ∧t G →t ∀xH(x)
and again by proposition 1.19 (i)
|= ∀x(G →t H(x)) →t (G →t ∀xH(x)).
But that means just
[[∀x(G →t H(x))]] ≤ [[G →t ∀xH(x)]].
To get the reverse inequality one may start from the fact that one has
[[G →t ∀xH(x)]] ≤ [[G →t H(a)]]
for each a ∈ X and thus
[[G →t ∀xH(x)]] ≤ inf [[G →t H(y)]].
y∈X
But this is already the inequality that is needed.

(ii) As in the proof of claim (i) one has to prove the equation
[[∀x(H(x) →t G)]] = [[∃xH(x) →t G]].
The fact that [[∃xH(x)]] ≥ [[H(a)]] for each a ∈ X and that therefore, be-
cause of proposition 1.4 (i), [[∃xH(x) →t G]] ≤ [[H(a) →t G]] holds true
immediately leads to
[[∃xH(x) →t G]] ≤ inf [[H(y) →t G]] = [[∀x(H(x) →t G)]].

y∈X
Thus the first inequality is proven.

For the other one we use the notation g = [[G]] and h(x) = [[H(x)]] and
have via representation lemma to prove that
inf (h(x) ϕt g) ≤ (sup h(x)) ϕt g = sup{w | (sup h(x)) t w ≤ g}.

x∈X x∈X x∈X
But this results from the fact that the leftmost term itself is one of the values
over which the supremum has to be taken in the rightmost term, for one has:
(sup h(x)) t inf (h(x) ϕt g) ≤ inf ((sup h(y)) t (h(x) ϕt g))

x∈X x∈X x∈X y∈X
= inf sup (h(y) t (h(x) ϕt g)) ≤ sup (h(y) t (h(y) ϕt g)) ≤ g.
x∈X y∈X y∈X
Hence the second inequality is proven too. QED
pressions are logically valid, if additionally G does not contain the variable
x free:
(i) |= ∃x(G ∧t H(x)) ↔ G ∧t ∃xH(x),

(ii) |= ∀x(H1 (x) →t H2 (x)) ∧t ∃xH1 (x) → ∃xH2 (x).
Proof. With the notation g = [[G]] and h(x) = [[H(x)]] as in the last proof,
for (i) one has to prove
sup (g t h(x)) = g t (sup h(x)),

x∈X x∈X
and this is exactly the condition LSC(t). Thus (i) holds true simply by
assumption. And for (ii) one has
[[∀x(H1 (x) →t H2 (x)) ∧t ∃xH1 (x)]]

= [[∃x(∀y(H1 (y) →t H2 (y)) ∧t H1 (x))]]
≤ [[∃x((H1 (x) →t H2 (x)) ∧t H1 (x))]]
≤ [[∃xH2 (x)]]
using (i) together with (Φ2), and that is all that has to be shown. QED
The most interesting point with item (i) of the last proposition is, that
the condition LSC(t) means exactly the logical validity of that expression,
i.e. means that as in classical logic an existential quantifier can be “moved
into” a conjunction (now: t-norm based) in case only one of the conjuncts
contains the quantified variable. From the logical perspective it is, besides
the definability of a corresponding implication operator, the need for this
property which makes the assumption LSC(t) often necessary.
What has been lacking until now with regard to such logically valid for-
mulas which give equivalences describing possibilities for the distribution of
quantifiers over suitable conjunctions and implications is the case of univer-
sal quantification of a t-norm based conjunction. Indeed, that needs another
type of assumption.
Proposition 1.29 For any t-norm t with property USC(t) the following ex-
pression is logically valid, if additionally G does not contain the variable x
free:
|= ∀x(G ∧t H(x)) ↔ G ∧t ∀xH(x).
Proof. Straightforward calculations as in the last proofs. QED
Therefore the closest analogies with the situation in classical logic are
obtained from the assumption of the continuity of the t-norm t which is
behind the connectives under consideration. But at present it does not seem
to be clear whether this assumption of continuity is too strong, or whether
one should accept it in every case. In other words, it is an open problem
which discontinuous t-norms (if any) may be of special interest in the realm
of fuzzy sets theory. The drastic product tD is the most well known example
of a discontinuous t-norm, but unfortunately (?) it has the property USC(tD )
and lacks having the property LSC(tD ), thus there is no Φ-operator, i.e. no
implication operator intimately connected to tD .
As the last topic essentially involving the quantifiers, let us reconsider
the problem of proposition 1.25 and take a look at the relations between any
of the t-norm based negations −t and the quantifiers.
(i) |= ∃x −t H(x) ↔ −t ∀xH(x),
(ii) |= −t ∃xH(x) ↔ ∀x −t H(x).
Proof. By definition 1.8 and proposition 1.27 (iv) one has

[[∃x −t H]] = [[∃x(H →t ⊥)]] = [[∀xH →t ⊥]] = [[−t ∀xH]]
and hence (i). To get (ii) one has to refer to proposition 1.27 (ii) instead.
QED
As a means for the formulation of some later results we also need the
following notions and notations.
Definition 1.9 Suppose H1 , H2 , H3 , . . . is a sequence of well-formed formu-

las and Θ a well-formed formula. Then let be
1
Y 1
^
Hi =def H1 , Hi =def H1 ,
i=1 i=1
n+1
Y n
Y n+1
^ 1
^
Hi =def ( Hi ) ∧t Hn+1 , Hi =def ( Hi ) ∧ Hn+1 ,
i=1 i=1 i=1 i=1
for each integer n ≥ 1 and additionally

n
Y
[Θ]n =def Θ.
i=1
We have omitted the explicite reference to the t-norm t with our finite it-
Q
eration of ∧t and with the exponential notation [Θ]n . We assume that this
causes no difficulties as we do not discuss different t-norms simultaneously
(besides t = min, which has its own notation).
What we especially need in chapter 3 is the following result concern-
ing the exchangeability of many-valued existential quantification with the
exponential notation of that last definition.
Proposition 1.31 Suppose that the t-norm t has property LSC(t). For each
integer n ≥ 1 and each well-formed formula H(X) of our many-valued lan-
guage for fuzzy sets theory there holds true
[[∃X([H(X)]n )]] = [[[∃XH(X)]n ]].
Proof. Obviously one has [[H(X)]] ≤ [[∃XH(X)]] and therefore also the
inequality [[[H(X)]n ]] ≤ [[[∃XH(X)]n ]], such that
[[∃X([H(X)]n )]] ≤ [[[∃XH(X)]n ]].
But we also have

n
Y
n
[[[∃XH(X)] ]] = [[ (∃XH(X))]]
i=1
n
Y
= sup{ [[ H(Xi /Ai )]] | A1 , . . . , An ∈ IF (X )}
i=1
because of LSC(t). And because

n
Y
[[ H(Xi /Ai )]] ≤ max [[H(Xi /Ai )n ]]
1≤i≤n
i=1
we now get
[[[∃XH(X)]n ]] ≤ sup { [[[H(X/A)]n ]] | A ∈ IF (X )} = [[∃X([H(X)]n )]].
Thus the proof is complete. QED

Chapter 2
Basic fuzzy set theory
2.1 Set algebra for fuzzy sets

It is an old problem of fuzzy set theory, already stressed in Zadeh (1965),
that we do not have really convincing arguments for something like a “right
choice” of our connectives, i.e. of our operations describing the union and in-
tersection of fuzzy sets. Of course, the problem in the past has been discussed
from different points of view, cf. e.g. Bellman/Giertz (1973), Yager
(1979), Gottwald (1979a) and Giles (1976), (1979), Thole/Zimmer-
mann/Zysno (1979), Zimmermann/Zysno (1980), Hamacher (1978)
too. Besides those motivational discussions for the choice of “right” con-
nectives for fuzzy set operations there is another mainstream in the current
research work: to accept a broad class of possible candidates for fuzzy set
operations, to discuss them all in parallel, and to choose concrete ones for
concrete applications. In this sense different families of operations have been
discussed by for example Yager (1980), Dombi (1982), Weber (1983),
all of which proved to be special cases of the t-norms of Schweizer/Sklar
(1961) which we discussed extensively in section 1.2. Hence, if we do not
change the set [0,1] of generalized membership degrees, these t-norms and
their duals, the t-conorms (which are sometimes called s-norms too), seem
to require attention. As is well known, the t-norms generalize the usual
conjunction operator and hence define intersection operations for fuzzy sets;
correspondingly the t-conorms generalize the usual disjunction operator and
thus define union operations for fuzzy sets. Additionally, in the following we
have to consider generalized complementation operations for fuzzy sets, and
we will use the Φ-operators as generalizations of the implication operator
38 CHAPTER 2. BASIC FUZZY SET THEORY
and hence for defining generalized inclusion relations.

As usual, every fuzzy set A on a given universe of discourse X is uniquely
characterized by its membership function µA : X → [0, 1] and furthermore
will be identified with this membership function. That means that defining
a fuzzy set A (on X ) is to define a function µA : X → [0, 1].
The approach which is presented here will use to a great extent the
language of (a suitable system of) a many-valued logic L and, from this
perspective, add two more points to a simple treatment of fuzzy set theory:
(i) we use the generalized – i.e. two-place many-valued – membership
predicate ε and interpret the membership degrees µA (x) as truth degrees,
i.e. we put (cf. (1.5))
µA (x) = truth value of x ε A = [[x ε A]],
in such a way making obvious on the formal level the strong and far-reaching
analogies of classical and fuzzy set theory;
(ii) we extend these analogies further by introducing a suitable many-
valued generalization of the very useful classical abstraction terms {x | . . .}
for “the set of all x such that . . . ”; these generalized abstraction terms will
prove as useful for the definition of fuzzy sets as the corresponding classical
abstraction terms do for the definition of crisp sets.
In this book we will neither consider the problem of axiomatizability
or axiomatization of the set of all valid formulae, nor discuss if and how
to base fuzzy set theory on some system of axioms. Instead, we prefer a
“naive” approach which presupposes that we have – for every given universe
of discourse X – a sufficiently good intuitive understanding of the class of
fuzzy subsets of X .
Additionally, this approach shall be restricted to such fuzzy sets which
would constitute the fuzzy sets of the first level inside a full “cumulative”,
set theoretical hierarchy of fuzzy sets based on some given class of urele-
ments. In set theoretical terms that means that our fuzzy sets shall only
be fuzzy sets of “urelements”. Different approaches toward full cumulative
hierarchies of fuzzy sets like those ones of Chapin (1974/75), Gottwald
(1979), Zhang (1980), or Weidner (1981) are not relevant for our present
purposes.1
To introduce, as announced, the generalized abstraction terms, we now
give the following
1
Of course there are no extra assumptions concerning the nature of the elements of our
universes of discourse. Therefore they themselves may be fuzzy sets, but this fact will not
become recognized by our formalism.
2.1. SET ALGEBRA FOR FUZZY SETS 39
Definition 2.1 For each universe of discourse X and each formula H(x) of
our set theoretic language of many-valued logic we denote by {x ∈ X k H(x)}
or simply by {x k H(x)} that fuzzy set A on X whose membership function
µA is characterized by: µA (a) = [[H(a)]] for each a ∈ X ; i.e. {x ∈ X k H(x)}
has the characteristic property
[[a ε {x ∈ X k H(x)}]] = [[H(x/a)]] for all a ∈ X .
Furthermore, we sometimes have to use a simple many-valued translation

of usual identity, which is defined as
½
. 1, if a = b
[[a = b]] = (2.1)
0 otherwise
and will also be allowed to be used in the case that instead of a, b some terms
(of a well defined kind) appear.
Definition 2.2 For every t-norm t, t-conorm st and negation function n

let for any fuzzy sets A, B be their (t-norm-based) intersection and union,
as well as their complement:
A ∩t B =def {x k x ε A ∧t x ε B},
A ∪t B =def {x k x ε A ∨t x ε B},
CCn A =def {x k ∼n (x ε A)}.
For the special negation functions nt , t any t-norm with the property LSC(t),
and ntL , a separate, simpler notation shall be used:
CCt A =def {x k ∼n (x ε A)} and CCA =def {x k ¬(x ε A)}.

t
Thus, our generalized abstraction terms {x k H(x)} allow for definitional
formulas which on the formal level are strongly analogous to the formulas
which in classical set theory usually describe the crisp notions corresponding
to the fuzzy ones introduced here. Furthermore these far-reaching formal
analogies with the crisp case may often be observed in the formulations of
the results as well as in the proofs — and they are a severe indication of
the (at least methodological) usefulness of our approach toward fuzziness
through (the language of) many-valued logic.
To occasionally simplify notation we follow the common usage of writing
in the case of t = tG = min
∩ for ∩tG and ∪ for ∪tG . (2.2)

Also, as usual, by IF (X ) the class

IF (X ) =def [0, 1]X
of all fuzzy subsets of X will be denoted. Here, IF (X ) is the set power (or the
direct power in algebraic terms) of the set of generalized membership (i.e.:
truth) degrees with respect to the universe of discourse as exponent. It is well
known that this point of view can be extended to include the consideration
of set algebraic operations for fuzzy sets too. That means, having given any
families (ti )i∈I of t-norms, (stj )j∈J of t-conorms, and (nk )k∈K of negation
functions, the algebraic structure
F(X ) = hIF (X ), (∪ti )i∈I , (∩tj )j∈J , (CCnk )k∈K i (2.3)
of fuzzy subsets of X is the direct power Π X of the structure
Π =def h[0, 1], (ti )i∈I , (stj )j∈J , (nk )k∈K i
of generalized membership degrees. The same remark holds true for any
kind of L-fuzzy sets too, but as we are interested in t-norms we have to
restrict ourselves here to the consideration of the set [0,1] of generalized
membership degrees. But, obviously, our principal idea to treat fuzzy set
theory with the help of (the language of) a suitable many-valued logic works
in the same way too for other kinds of structures of generalized membership
degrees than those we are actually discussing here.
If one speaks on the structures Π and F(X ) one has to use terms referring
to the elements of those structures. In our case those terms for the structure
Π are exactly our expressions of the language of many-valued logic which
are built up from variables for truth degrees – or expressions “x ε A” with
x a variable for elements of X and (the letter) A a symbol for elements of
IF (X ) – with the help of the connectives (i.e. operation symbols for truth
functions) ∧ti , ∨tj , ∼nk ; and the terms for the structure F(X ) are exactly
our expressions of the extended set theoretic language which are built up
from variables and constants for fuzzy sets with the help of the corresponding
operation symbols ∩ti , ∪tj , CCnk introduced in definition 2.2.
By a basic Horn formula we now understand a finite disjunction (in
the classical metalanguage!) of at most one term equation and negations of
term equations otherwise. Thus simple examples of basic Horn formulas
with respect to Π are (with p, q as [propositional] variables for generalized
truth degrees, i.e. for elements of [0,1])
p ∧ti q = q ∧ti p,
p ∧ti q 6= p ∧ti (q∨tj ∼nk q),
and corresponding basic Horn formulas for F(X ) are e.g.
A ∩ti B = B ∩ti A,
A ∩ti B 6= A ∩ti (B ∪tj C
Cnk B).
Furthermore, by a Horn formula we mean such a formula which is built up

from basic Horn formulas (with respect to the same structure) using con-
junction, existential and universal quantification (of classical metalanguage).
Referring to a well known theorem in classical model theory, cf. e.g.
Chang/Keisler (1973), we now are able to state many set algebraic laws
for fuzzy sets at once.
Theorem 2.1 Suppose that H is any Horn formula (with respect to F(X )).
Then the structure F(X ) is a model for H, i.e. H holds true for fuzzy sub-
sets of X , if the Horn formula Ĥ corresponding to H but referring to the
structure Π (i.e. Ĥ is built up from H by changing ∩t to ∧t , ∪t to ∨t ,
CCn to ∼n and by an appropriate change of variables too) holds true in the
structure Π.
Proof. (cf. Chang/Keisler (1973), pp. 326 ff.).
As an immediate consequence of this theorem we get the commutativity

and associativity of all unions ∪t and all intersections ∩t for fuzzy sets
from the corresponding properties of t-conorms and t-norms because these
properties can be simply formulated by Horn formulas as e.g.
^
(A ∩t B = B ∩t A),
A,B
^
(A ∩t (B ∩t C) = (A ∩t B) ∩t C)
A,B,C
for the intersections.

If we add the natural ordering relation ≤ to the structure Π we get in
the direct power F(X ) an inclusion relation characterized by
A⊂B iff [[a ε A]] ≤ [[a ε B]] for all a ∈ X (2.4)
which already was defined by Zadeh (1965) and which is a partial ordering
in the class IF (X ) of all fuzzy subsets of X .
It is interesting to note that our theorem can be extended to include these
ordering relations too. For this purpose we have to extend what is meant
by a basic Horn formula: this is now not only a disjunction of negated

or unnegated term equations but a disjunction of – negated or unnegated
– term equations or term inequalities τ10 ≤ τ20 and τ1 ⊂ τ2 for terms τ1 , τ2
relating to Π and terms τ10 , τ20 relating to the structure F(X ) of (2.3). Hence,
for example, the transitivity of the ordering relation ≤ can be described by
a Horn formula and therefore our theorem yields
A⊂B and B ⊂ C ⇒ A ⊂ C
for any fuzzy sets A, B, C.
Unfortunately, as we have to refer in our theorem to the classical met-
alanguage, the binary relation ⊂ for fuzzy sets is a two-valued relation, i.e.
not itself a graded one. Thus one would only have to compare the fuzzy
set with their graded membership relation by a two-valued implication re-
lation. Instead, from the intuitive idea of an unsharp boundary for fuzzy
sets it would be more natural to have an “unsharp” inclusion relation too
for the fuzzy sets. Fortunately, by using our language of many-valued logic
it is quite simple and natural to introduce such a graded, i.e. “fuzzified” or
many-valued inclusion relation again in strong formal analogy with the usual
definition of the inclusion relation. To do this we need only the possibility
of universal quantification in our language with many truth degrees and we
need a generalized, i.e. many-valued implication operator. But, according to
the results of section 1.3 the Φ-operators ϕt are suitable candidates for such
implication operators. Hence we give the
Definition 2.3 For any fuzzy sets A, B and any t-norm t with property
LSC(t) let
A⊂=t B =def ∀x(x ε A →t x ε B),
A ≡t B =def A⊂ ⊂
=t B ∧t B =t A.
The truth degree [[A ⊂
=t B]] is a degree of containment of A in B and the
truth degree [[A ≡t B]] is a degree of equality for the fuzzy sets A, B.
To see how these definitions work let us look at the t-norms tG and tL .
For a readable formulation of the following results we use besides the support
supp (A) = {x ∈ X | µA (x) > 0} = {x ∈ X | [[x ε A]] 6= 0}
of A ∈ IF (X ) for any fuzzy sets A, B ∈ IF (X ) also the crisp sets
{A > B} =def {x ∈ X | [[x ε A]] > [[x ε B]]}, (2.5)
{A 6= B} =def {x ∈ X | [[x ε A]] 6= [[x ε B]]}. (2.6)
These sets generalize the support in the sense that
supp (A) = {A > ∅} = {A 6= ∅}
holds true for each A ∈ IF (X ).

Straightforward calculations give for t = tG = min the result2
[[A ⊂
=tG B]] = x∈{A>B}
inf [[x ε B]] (2.7)
with the corollary
supp (A) \ supp (B) 6= ∅ ⇒ [[A ⊂

= B]] = 0. (2.8)
tG
Then immediately one also has by definition 2.3
[[A ≡tG B]] = inf min{[[x ε A]], [[x ε B]]}

x∈{A6=B}
= inf [[x ε A ∩ B]] (2.9)
x∈{A6=B}
now with the corollary
supp (A) 6= supp (B) ⇒ [[A ≡tG B]] = 0 (2.10)
which indicates that ≡tG is quite a strong fuzzified equality.

The other case t = tL again by elementary calculations first gives
[[A ⊂
=tL B]] = 1 − sup ([[x ε A]] − [[x ε B]]) (2.11)
x∈{A>B}
and therefore with the auxiliary notation
∆(A, B) =def sup ([[x ε A]] − [[x ε B]])

x∈{A>B}
quite directly
[[A ≡tL B]] = max{0, 1 − (∆(A, B) + ∆(B, A))}. (2.12)
This time, contrary to the results (2.8) and (2.10), both of the claims
[[A ⊂
=tL B]] 6= 0 and supp (A) 6⊆ supp (B)
2
We consider, as usual, the infimum over the empty set as =1 and the supremum over
the empty set as =0.
as well as both of the claims

[[A ≡tL B]] 6= 0 and supp (A) 6= supp (B),
can be true at once.
To some extent therefore ⊂ =tL and ≡tL better meet the intuition be-
hind the fuzzified inclusion and equality than ⊂=tG and ≡tG , namely the
intuition that “small” deviations from the “true”, i.e. complete inclusion or
equality do not completely falsify the generalized inclusion or equality. In
addition, deviations from supp (A) ⊆ supp (B) and supp (A) = supp (B)
should surely count as “small” as long as combined with small differences in
the membership degrees over the “critical” regions supp (A) \ supp (B) and
supp (A) 4 supp (B).3
As in chapter 1 we prefer to formulate the further results of this chapter

in the form of the logical validity of formulas. If it is equivalent to formulate
such a result as an inequality or as an equation for truth degrees we, as
before, shall write → and ↔ without index instead of →t0 and ↔t0 in all
those cases where t0 is any t-norm subject only to the restriction to have the
property LSC(t0 ). But we extend this notation and allow also for ⊂ = and
≡ without index – with the same understanding that this is shorthand for
⊂ and ≡ with a t-norm t0 with property LSC(t0 ).
=t0 t0
The second one of the graded relations defined in definition 2.3 is a kind
of many-valued identity relation and will be considered later on. First let us
note that because of proposition 1.6 (i) obviously
A ⊂ B ⇔ |= A ⊂ ⊂
=t B ⇔ |= A = B (2.13)
and therefore also
A = B ⇔ |= A ≡t B ⇔ |= A ≡ B (2.14)
Hence ⊂ =t is a suitable generalization of ⊂; a fact which once again supports
the point of view that treating fuzziness by (the language of) many-valued
logic is a useful doing. And ≡t is at least nicely related to the usual identity
relation for fuzzy sets.4 Furthermore, for the many-valued inclusion relations
3
4 denotes here the symmetric difference of crisp sets.
4
In proposition 2.4 and later on we shall see that ≡t has many properties which are
natural generalizations of typical properties of the usual (two valued) identity relation.
Indeed, ≡t is an interesting candidate for a generalized identity of many-valued logic.
This topic is discussed in more detail from the theoretical perspective of a many-valued
first-order logic with identity in Gottwald (1989).
⊂ suitable many-valued versions of the fundamental properties of inclusion

=t
for crisp sets hold true too.
Proposition 2.2 For every t-norm t with property LSC(t) and all fuzzy sets
A, B, C there hold true
(i) |= A ⊂
=t A,
(ii) |= A ⊂
=t B → A ∩t C ⊂ =t B ∩t C,
(iii) ⊂
|= A = B ∧t B = C →t A ⊂
⊂
t t =t C,
(iv) |= A ⊂
=t B → CCt B ⊂=t C
Ct A,
and if additionally st1 is any t-conorm such that t distributes over st1 we
too have
(v) |= A ⊂ ⊂
=t B → A ∪t1 C =t B ∪t1 C.
Proof. (i) is obvious. For (ii) we have to use the corresponding definitions
2.2 and 2.3 and the fact that
[[H1 →t H2 ]] ≤ [[H1 ∧t G →t H2 ∧t G]]
which was proven in proposition 1.18 (i).

(iii) By definition 2.2 we have for the truth degrees
[[A ⊂ ⊂
=t B ∧t B =t C]]
= [[∀x(x ε A →t x ε B) ∧t ∀x(x ε B →t x ε C)]]
≤ [[∀x((x ε A →t x ε B) ∧t (x ε B →t x ε C))]]
using the fact that according to proposition 1.26 (i)
|= ∀xH1 ∧t ∀xH2 → ∀x(H1 ∧t H2 )
holds true for all formulas H1 , H2 and all t-norms with the property LSC(t).
We furthermore have
|= (H1 →t H2 ) ∧t (H2 →t H3 ) →t (H1 →t H3 )
for all formulas H1 , H2 , H3 ; cf. proposition 1.19 (ii). Thus we get
[[A ⊂ ⊂ ⊂
=t B ∧t B =t C]] ≤ [[A =t C]]
which immediately yields our statement.
(iv) becomes, by the same kind of argument as in (ii) and (iii), a straight-
forward consequence of
[[H1 →t H2 ]] ≤ [[−t H2 →t −t H1 ]]
and this is true in the case of LSC(t) because of proposition 1.23 (iv).
To prove (v) we look as in case (ii) for a proof of
[[A ⊂ ⊂
=t B]] ≤ [[A ∪t1 C =t B ∪t1 C]]
which is by definition of ⊂
=t easily established if one is able to show
[[x ε A →t x ε B]] ≤ [[x ε A ∨t1 x ε C →t x ε B ∨t1 x ε C]]
Hence we look for a condition of t, t1 which guarantees for all formulas
H1 , H2 , G that
[[H1 →t H2 ]] ≤ [[H1 ∨t1 G →t H2 ∨t1 G]],
and this means looking for such a condition that yields
sup{s | u t s ≤ v} ≤ sup{t | (u st1 w) t t ≤ v st1 w} (2.15)
Now, if t distributes over st1 , i.e. if always
(u st1 w) t t = (u t t) st1 (w t t),
we get immediately always
u t s ≤ v ⇒ (u st1 w) t s ≤ v st1 w
and hence (2.15). Therefore (v) is established. QED
Remark. The present proof shows for example that the distributivity of t
over st1 is only a sufficient condition for the monotonicity property
|= A ⊂ ⊂
=t B →t A ∪t1 C =t B ∪t1 C,
a subdistributivity condition like
(u st1 w) t v ≤ (u t v) st1 (w t v) (2.16)
or simply the property that
(u st1 w) t v ≤ u st1 (w t v) (2.17)
(for all u, v, w ∈ [0, 1] of course) are sufficient conditions too.
Furthermore, this proof indicates that (iv) can also be generalized. Con-
sidering any negation function n and looking for the truth of
|= A ⊂ ⊂
=t B → CCt B =t CCt A, (2.18)
our line of arguments calls for the contraposition condition
u ϕt v ≤ n(v) ϕt n(u)
to hold true as a sufficient condition for (2.18).
Furthermore, it is possible in a standard way to extend the monotonicity
results to for example
|= A ⊂ ⊂ ⊂
=t B ∧t C =t D → A ∩t C =t B ∩t D
for the case of the intersection ∩t . Besides the monotonicity result of propo-
sition P3.4 (ii) one simply has to use the transitivity result of proposition
P3.4 (iii).
Surely, as indicated in this remark, the results of proposition 2.2 can be

extended in various ways. But this will not be our problem here, because
we mainly intend to show how to work in fuzzy set theory with t-norms, t-
conorms and Φ-operators, and how to do this in the context of many-valued
logic. Therefore we only add two further special monotonicity results which
we will later on have to refer to and then look at the generalized identity
relations ≡t of definition 2.3 and their fundamental properties.
Proposition 2.3 For every t-norm t with property LSC(t) and all fuzzy sets
A, B, C there hold true
(i) |= A ⊂ ⊂ ⊂
=t B ∧ C =t D → A ∩ C =t B ∩ D,
(ii) |= A ⊂ ⊂ ⊂
=t B ∧ C =t D → A ∪ C =t B ∪ D.
Proof. (i) is an immediate consequence of proposition 1.19 (iv) and (ii) one
of proposition 1.19 (v). QED
Proposition 2.4 Suppose LSC(t). Then for all fuzzy sets A, B, C there hold
true
(i) |= A ≡t A,
(ii) |= A ≡t B → B ≡t A,
(iii) |= A ≡t B ∧t B ≡t C →t A ≡t C,
(iv) |= A ≡t B → A ∩t C ≡t B ∩t C,
(v) |= A ≡t B → CCt A ≡t C
Ct B,
and if the t-conorm st1 is such that t distributes over st1 one also has
(vi) |= A ≡t B → A ∪t1 C ≡t B ∪t1 C.
Proof. (i) and (ii) are obvious.

(iii) is a simple consequence of the transitivity of ⊂
=t , cf. proposition 2.2
(ii), the definition 2.3 of ≡t , and of the fact that in case LSC(t)
|= (H1 →t G1 ) ∧t (H2 →t G2 ) → (H1 ∧t H2 →t G1 ∧t G2 ) (2.19)
holds true for all formulas H1 , H2 , G1 , G2 ; cf. proposition 1.19 (iii).

(iv) by definition of ≡t means
|= (A ⊂ ⊂
=t B ∧t B =t A) →
→ (A ∩t C ⊂ ⊂
=t B ∩t C ∧t B ∩t C =t A ∩t C)
and hence is a consequence of proposition 2.2 (iii) via (2.19).

Finally, (v) and (vi) can be established in exactly the same way from
propositions 2.2 (v) and 2.2 (iv) respectively. QED
For the formulation of further results we need some additional notions.

At first these are the empty fuzzy set, fuzzy singletons, and the universal
fuzzy set, and later on the intersection and the union of a (crisp) family of
fuzzy sets.
Definition 2.4 For any t-norm t let be the universal set X with respect to
the universe of discourse X
.
X =def {x ∈ X k x = x},
and for each truth degree u 6= 0 the u-universal set with respect to X
.
X [u] =def {x ∈ X k x = x ∧t u}.
The empty fuzzy set ∅ (with respect to X ) is
∅ =def CCX.
And for each a ∈ X and any truth degree u the (fuzzy) u-singleton of a is
the fuzzy set
.
hhaiiu =def {x k x = a ∧t u}.
By abuse of notation we have, in the definitions of the u-universal set

X [u]and the u-singletons, also taken the truth degree u as the constant of
our language denoting just this degree. And we have chosen the same symbol
for the empty fuzzy set as for the usual, i.e. crisp empty set. This will not
cause any problems and has the advantage of simplicity.
De facto we have that X [1] = X and that each fuzzy singleton hhaiiu is
independent of the t-norm which is used in this definition because obviously
always
[[b ε X [u] ]] = u,
½
u, if b = a
[[b ε hhaiiu ]] =
0 otherwise
holds true and additionally
[[b ε ∅]] = 0.
Furthermore, hhaii0 = ∅ is always the case. Obviously too it holds true that
|= ∀x(x ε X),
|= ∀x ∼n (x ε ∅)
for any negation function n.
Proposition 2.5 For each fuzzy set A and every t-norm t with the property
LSC(t) there hold true
(i) |= A ≡t ∅ ↔ ∀x −t (x ε A),
(ii) |= −t ∃x(x ε A) ↔ A ≡t ∅,
(iii) |= A ≡t X ↔ ∀x(x ε A).
Proof. (i) is obvious from the corresponding definitions and proposition 1.4
(ii). (ii) follows from (i) and the fact that
|= −t ∃xH ↔ ∀x −t H
holds true; cf. proposition 1.30 (ii). Finally, (iii) again is a simple conse-
quence of the corresponding definitions and proposition 1.4 (iii). QED
The definitions of ∅ and X are quite natural ones from the point of view
of set theory. With the same resulting objects we could have chosen a more
algebraic approach: if one includes the truth degrees 0 and 1 as nullary

operation in the structure Π then also the direct power F(X ) has to be
extended with two nullary operations, and these are exactly ∅ and X. As
our theorem 2.1 (on the transfer of the truth of Horn formulas) covers this
case too, we can also get a series of results this way, e.g. for all fuzzy sets A
∅ ⊂ A and A ⊂ X,
or in our many-valued language
|= ∅ ⊂
=t A and |= A ⊂
=t X
for every t-norm t. From the fact that
p ∧t (−t p) = ⊥
holds true in Π because of (Φ2) we also get in this way immediately
A ∩t CCt A = ∅,
as those formulas are (basic) Horn formulas. Thus we have in our many
valued language
|= A ∩t CCt A ≡t ∅.
But, again, more and stronger results are provable than by this reference
only to our transfer theorem 2.1.
Proposition 2.6 For all t-norms t with property LSC(t), t-conorms st1 and
fuzzy sets A, B there hold true
(i) |= A ∩t B ≡t X → A ≡t X ∧ B ≡t X,
(ii) |= A ≡t X ∧t B ≡t X → A ∩t B ≡t X,
(iii) |= A ∪t1 B ≡t ∅ → A ≡t ∅ ∧ B ≡t ∅,
and if t, t1 have the weak de Morgan property that Π is a model of
[[(−t p) ∧t (−t q)]] ≤ [[−t (p ∨t1 q)]]
then it also holds true that
(iv) |= A ≡t ∅ ∧t B ≡t ∅ → A ∪t1 B ≡t ∅.
Proof. Let us, because of the additional assumption with regard to (iv),
start by proving (iii) and (iv). Then (i) and (ii) can be proven analogously
without this additional assumption. The details for those cases are then
easily supplied and hence omitted.
(iii) We immediately have
[[A ∪t1 B ≡t ∅]] = [[∀x −t (x ε A ∨t1 x ε B)]]
≤ [[∀x −t (x ε A)]] = [[A ≡t ∅]]
from monotonicity properties of ∨t1 and −t . Hence by commutativity of
∪t1 :
[[A ∪t1 B ≡t ∅]] ≤ min{[[A ≡t ∅]], [[B ≡t ∅]]},
and hence the result which was stated.
Point (iv) can be analogously derived using
[[A ≡t ∅ ∧t B ≡t ∅]] = [[∀x −t (x ε A) ∧t ∀x −t (x ε B)]]
≤ [[∀x(−t (x ε A) ∧t −t (x ε B))]],
which holds true according to proposition 1.26 (i), and hence we are able to
continue this estimation with
[[A ≡t ∅ ∧t B ≡t ∅]] ≤ [[∀x −t (x ε A ∨t1 x ε B)]]
because of the assumption on t, t1 stated additionally for (iv); thus
[[A ≡t ∅ ∧t B ≡t ∅]] ≤ [[A ∪t1 B ≡t ∅]]
and hence our statement. QED
In order to now define the “big”, i.e. infinitary unions and intersections of
families of fuzzy sets we will discuss a slight extionsion of our many-valued
language: we have to introduce infinitary conjunctions and disjunctions. The
finite iterations of each one of the conjunction operators ∧t and of each one
of the disjunction operators ∨t are routine and will not be discussed here.
What is much more problematic is the infinitary version in general: but we
avoid almost all difficulties in restricting ourselves to the t-norm min and
the t-conorm max only.
Having given any (crisp) index set I and for each i ∈ I a formula Hi we
introduce as new formulas for infinitary conjunctions and disjunctions
^
^ _
_
Hi and Hi .
i∈I i∈I
Those new formulas may be used in building more complex ones as we have
so far done with the more simple ones. The truth values of those formulas
are given as
^
^
[[ Hi ]] =def inf [[Hi ]],
i∈I i∈I
_
_
[[ Hi ]] =def sup[[Hi ]].
i∈I i∈I
V
V W
W
From these definitions it is obvious that i∈I and i∈I will act much like
the quantifiers ∀ and ∃. But we will not prove a detailled list of logically
valid formulas including such infinitary parts – if in the following we have
to use some special results we will prove them when needed.
Definition 2.5 For each familiy (Ai )i∈I of fuzzy sets let
\ ^
^
Ai =def {x k (x ε Ai )},
i∈I
i∈I
[ _
_
Ai =def {x k (x ε Ai )}.
i∈I
i∈I
As usual, the generalized union can also be used in our fuzzy case to
describe every fuzzy set as a union of fuzzy singletons:
[
A= hhaii[[a ε A]] . (2.20)
a∈X
Yet this representation is sometimes too complicated because all the fuzzy
singletons hhaii0 are inessential. First, as we mentioned immediately after
definition 2.4, hhaii0 = ∅ always applies and on the other hand the definition
S
2.5 generally allows one to disregard in a generalized union i∈I Ai all those
Ak with Ak = ∅. Hence instead of (2.20) we simply get
[
A= hhaii[[a ε A]] , (2.21)
a∈supp(A)
which is exactly the same as in Zadeh’s (1975) notation the representation

of a fuzzy set as a sum or integral of singletons, i.e. is the same as each one
of the equalities
X Z
A= µA (a)/a or A= µA (a)/a.
a∈supp(A) a∈supp(A)
Proposition 2.7 For all fuzzy sets (Ai )i∈I , B and arbitrary t-norms t with
the property LSC(t) and t-conorms st1 there hold true:
\
(i) |= Ai ⊂
=t Ak ,
i∈I
[
(ii) |= Ak ⊂
=t Ai ,
i∈I
^
^ \
(iii) |= (B ⊂ ⊂
=t Ai ) ↔ B =t Ai ,
i∈I
i∈I
^
^ [
(iv) |= (Ai ⊂
=t B) ↔ Ai ⊂
=t B,
i∈I
i∈I
and without restrictions concerning the involved t-norms there hold true:
\ \
(v) |= Ai ∩t B ⊂
= (Ai ∩t B),
i∈I i∈I
\ \
(vi) |= Ai ∪t1 B ⊂
= (Ai ∪t1 B),
i∈I i∈I
[ [
(vii) |= (Ai ∩t B) ⊂
= Ai ∩t B,
i∈I i∈I
[ [
(viii) |= (Ai ∪t1 B) ⊂
= Ai ∪t1 B.
i∈I i∈I
Proof. (i) and (ii) are obvious. For (iii) we observe that
^
^
[[ (B ⊂
=t Ai )]] = inf inf [[x ε B →t x ε Ai ]]
i∈I i∈I x∈X
^
^
= inf [[x ε B →t (x ε Ai )]]
x∈X i∈I
according to proposition 1.27 (i). This immediately gives our result. (To
be precise, we have to mention that the formulation of proposition 1.27 (i)
gives the corresponding result for the quantifier ∀; but the proof of that
V
V
result works exactly the same way for the present infinitary conjunction
too.)
To prove (iv) via an equation of suitable truth degrees we proceed anal-
ogously but with reference to proposition 1.27 (ii).
Furthermore, (v) to (viii) all are simple consequences of the monotonicity
of t-norms and t-conorms. The details can be omitted. QED
Proposition 2.8 For all fuzzy sets (Ai )i∈I , B and arbitrary t-norms t and
t-conorms st1 there hold true:
(i) in the case of USC(t)
\ \
|= Ai ∩t B ≡ (Ai ∩t B),
i∈I i∈I
(ii) in the case of USC(st1 )

\ \
|= Ai ∪t1 B ≡ (Ai ∪t1 B),
i∈I i∈I
(iii) in the case of LSC(t)

[ [
|= Ai ∩t B ≡ (Ai ∩t B),
i∈I i∈I
(iv) in the case of LSC(st1 )

[ [
|= Ai ∪t1 B ≡ (Ai ∪t1 B).
i∈I i∈I
Proof. By straightforward calculations from the monotonicity of t and st1

as mentioned in the last proof. The semicontinuity assumptions in every
case guarantee the equality of truth degrees of the left and right sides of the
≡t -equations. QED
Principally with the same elementary steps as for proving these distribu-
tivity properties, one is able to prove the general commutativity and asso-
ciativity of those infinitary intersections and unions of fuzzy sets, as well as
their general distributivity with respect to one another. For formulations of
those laws in the crisp case we refer to Klaua (1964) or any other textbook
of set theory which tends toward a more formalized presentation of the topic
in the language of first order logic; in the fuzzy case here the laws have to
be formulated in the same way. We do not give the details which, in some
sense, are routine.
Instead we went to the consideration of cartesian products of fuzzy sets.
The problem here is – as later on with fuzzy (binary) relations again – that
the cartesian product of two fuzzy subsets of X need not be a fuzzy subset
of X : it depends on X and on whether X is closed with respect to (usual)
ordered pairing.
Fortunately, for the following discussions of fuzzy cartesian products we

do not need to have the universe of discourse X closed under ordered pairing:
we can always work with fuzzy subsets of different universes of discourse.
First let us define the fuzzy cartesian products using the usual notion of
an ordered pair (a, b) of elements a, b of the universe of discourse X .
Definition 2.6 For all fuzzy sets A, B and each t-norm t let
A ×t B =def {(x, y) k x ε A ∧t y ε B}.
Obviously we have as an extended meaning of ⊂

=t and ≡t for cartesian
products:
A ×t B ⊂
=t1 C ×t D =def
∀x∀y(x ε A ∧t y ε B →t1 x ε C ∧t y ε D), (2.22)
A ×t B ≡t1 C ×t D =def
(A ×t B ⊂ ⊂
=t1 C ×t D) ∧t1 (C ×t D =t1 A ×t B) (2.23)
for all fuzzy sets A, B, C, D and all t-norms t, t1 where, of course, t1 has the
property LSC(t).
Set algebraic equations involving fuzzy cartesian products easily result
from identities for t-norms, t-conorms and negation functions via our transfer
theorem 2.1. To illustrate the method of proof we give some simple examples.
Proposition 2.9 For t-norms t, t1 , t-conorms st2 and fuzzy sets A, B, C

there hold true
(i) if t distributes over t1 :
|= A ×t (B ∩t1 C) ≡ (A ×t B) ∩t1 (A ×t C),
(ii) if t distributes over st2 :
|= A ×t (B ∪t2 C) ≡ (A ×t B) ∪t2 (A ×t C).
Proof. We have by definitions and distributivity of t over t1
A ×t (B ∩t1 C) = {(x, y) k x ε A ∧t (y ε B ∧t1 y ε C)}

= {(x, y) k (x ε A ∧t y ε B) ∧t1 (x ε A ∧t y ε C)}
= (A ×t B) ∩t1 (A ×t C)
and hence (i). In the same way the proof of (ii) can proceed. QED
We only mention the fact that the fuzzy cartesian product is commuta-
tive and associative – via suitable canonical isomorphisms as in the classical
case. Further equalities for fuzzy cartesian products are easily provable by
the method of the last proof. We will not discuss this topic more extensively.
Instead, we look for proofs of some monotonicity properties and some sub-
stitutivity results for fuzzy identities.
Proposition 2.10 Let t be a t-norm with property LSC(t), A, B, C and all

(Ai )i∈I fuzzy sets. There then hold true
(i) |= A ⊂ ⊂
=t B → A ×t C =t B ×t C,
(ii) |= A ≡t B → A ×t C ≡t B ×t C,
(iii) |= A ≡t ∅ ∨ B ≡t ∅ → A ×t B ≡t ∅,
which can be sharpened for any t-conorm st1 which has the de Morgan
property
[[−t (u ∧t v)]] = [[(−t u) ∨t1 (−t v)]]
for all truth degrees u, v to:
(iv) |= A ×t B ≡t ∅ → A ≡t ∅ ∨t1 B ≡t ∅.
With respect to the infinitary unions and intersections in each case one has
[ [
(v) |= (Ai ×t B) ⊂
= Ai ×t B,
i∈I i∈I
\ \
(vi) |= Ai ×t B ⊂
= (Ai ×t B),
i∈I i∈I
with, assuming LSC(t), the sharpening

[ [
(vii) |= (Ai ×t B) ≡ Ai ×t B,
i∈I i∈I
and with, assuming USC(t), the sharpening

\ \
(viii) |= Ai ×t B ≡ (Ai ×t B).
i∈I i∈I
We omit the proof. But all those statements are analogous to earlier
ones relating to for example ∩t instead of ×t . In every case, too, the proofs
needed here are analogous to those earlier proofs which we have presented
in detail or sketched with hints for the essential steps.
The part of set algebra for fuzzy sets that was presented here will be
representative enough to allow the reader to extend it easily by himself in
any of the directions he is interested in.
Therefore, we finish our listing of interesting results of set algebra. Only a
slight modification of the construction yielding the fuzzy cartesian products
will be mentioned. The point is that the ordered pair (x, y) is a very special
object among those which can be built up from x, y. Hence we may assume
to have given some binary – or also n-ary, but we will not consider this
only slightly more general case – function f on the universe of discourse X .
Additionally we assume that each one of the values f (a, b) can be described
by a term of our language. To extend our abstraction terms to the case
of fuzzy sets with objects f (a, b) as “elements” we define for all formulas
H(x, y):
.
{f (x, y) k H(x, y)} =def {z k ∃x∃y(H(x, y) ∧t z = f (x, y))} (2.24)
for each t-norm t. To avoid any inessential complication let us additionally

assume that always f (x, y) ∈ X , i.e. that f is a binary operation in X . Now
we can formulate the famous extension principle of Zadeh (1975) for any
t-norm in a very simple manner.
Definition 2.7 (Extension Principle) Suppose that a binary operation ∗

in the universe of discourse X is given and some t-norm t chosen. This
operation is extended to a binary operation ∗t for fuzzy sets from IF (X ) by
defining
A ∗t B =def {a ∗ b k a ε A ∧t b ε B}
for all fuzzy sets A, B ∈ IF (X ).
Corollary 2.11 Take the operations ∗ and ∗t as in definition 2.7 and con-
sider C = A ∗t B for A, B ∈ IF (X ). It then holds true
µC (z) = sup µA (x) t µB (y) for all z ∈ X .

x,y∈X
z=x∗y
Proof. By (2.24) and the definitions of the quantifier ∃ and the class term
notation for fuzzy sets, the extension principle immediately yields
.
µC (z) = sup [[x ε A ∧t b ε B]] t [[z = x ∗ y]].
x,y∈X
Therefore
µC (z) = sup [[x ε A ∧t b ε B]]
x,y∈X
z=x∗y
.
because always [[z = x ∗ y]] ∈ {0, 1} and only those pairs (x, y) need to be
.
considered here, for which [[z = x ∗ y]] = 1, i.e. for which z = x ∗ y is the
case.5 Now
[[x ε A ∧t b ε B]] = µA (x) t µB (y)
gives the result. QED
2.2 Fuzzy relations

Intuitively, a fuzzy relation is a fuzzy set of ordered pairs, i.e. such a fuzzy
set whose support is a subset of a suitable crisp cartesian product. That
means, we intuitively identify relations with binary ones. This will be enough
generality for the present purposes. Of course, mathematically it is obvious
that in the same way one may consider n-ary fuzzy relations for any n ≥ 2.
From the formal point of view we have different possibilities to treat
fuzzy relations, e.g. we may
– assume that our universe of discourse X consists of ordered pairs only,
such that each fuzzy subset of X is a fuzzy relation;
– suppose that the universe of discourse X is closed with respect to

ordered pairing, i.e., we may suppose that for a, b ∈ X we have (a, b) ∈
X too;
– consider parallel with the universe of discourse X the set X × X as a

second universe and distinguish fuzzy sets as fuzzy subsets of X from
fuzzy relations as fuzzy subsets of X × X .
We will essentially follow this third approach here. The first one would
allow for a too restricted treatment of fuzzy relations only; and the second
approach in principle would force us to take as X a whole universe for clas-
sical set theory (or “nearly” such a whole universe): a situation which we
are interested in avoiding here.
5
It is essential for this argumentation that as usual the supremum over the empty set
is taken to be =1.
2.2. FUZZY RELATIONS 59
We have to extend our language to treat fuzzy relations. First we add a

third sort of symbols: those for fuzzy relations. We use the upper case latin
letters
R, S, T (possibly with subscripts or primes)
as symbols for fuzzy relations. As in this chapter we do not have to quantify
over fuzzy relations, these symbols will act as constants for fuzzy relations.
Of course, as before we will avoid using the letters R, S, T as symbols of
fuzzy sets. But, again as before in connection with cartesian products of
fuzzy sets, we extend in a straightforward way our use of generalized – i.e.
many-valued – abstraction terms to the case of fuzzy relations too. And the
same will be done with the notation for the set algebraic operations; i.e. we
now take for example
R ∩t S =def {(x, y) k (x, y) ε R ∧t (x, y) ε S}
and in the same manner put
R⊂ =t S =def ∀x∀y((x, y) ε R →t (x, y) ε S).
It is straightforward to see that all the results on inclusion and intersec-
tion of fuzzy sets, formulated and proved in the last section, also hold true
(with slightly modified proofs only) for fuzzy relations too. Thus it is not
necessary to write down these results for fuzzy relations again; nevertheless
we will use them in the following. So we start by first defining some specific
notions of relation theory.
Definition 2.8 For each t-norm t and all fuzzy relations R, S let be the
domain, the range, the inverse relation of R and the (relational) product of
R, S:
dom (R) =def {x k ∃y((x, y) ε R)},
rg (R) =def {y k ∃x((x, y) ε R)},
−1
R =def {(x, y) k (y, x) ε R},
R ◦t S =def {(x, y) k ∃z((x, z) ε R ∧t (z, y) ε S)}.
Proposition 2.12 For all fuzzy relations R, S one has
(i) dom (R−1 ) = rg (R) and rg (R−1 ) = dom (R),
(ii) ⊂
|= dom (R ◦t S) = dom (R),
(iii) |= rg (R ◦t S) ⊂
= rg (R),
(iv) (R−1 )−1 = R,
(v) (R ◦t S)−1 = S −1 ◦t R−1 .
Proof. All the results can be proved by routine methods; thus e.g. (v)
essentially follows from the definition of the inverse fuzzy relation R−1 and
the commutativity of ∧t by
(R ◦t S)−1 = {(x, y) k ∃z((y, z) ε R ∧t (z, x) ε S)}
= {(x, y) k ∃z((x, z) ε S −1 ∧t (z, y) ε R−1 )}
= S −1 ◦t R−1 .
And for (ii), for example, we have to show
|= ∀x(x ε dom (R ◦t S) → x ε dom (R))
which goes to show that
|= ∀x(∃y((x, y) ε R ◦t S) → ∃z((x, z) ε R)), (2.25)
but because of
[[(x, y) ε R ◦t S]] = [[∃z((x, z) ε R ∧t (z, y) ε S)]]
we have by monotonicity of the t-norms
|= ∃y((x, y) ε R ◦t S) → ∃y∃z((x, z) ε R)
and hence (2.25) by first dropping the empty quantification ∃y in the succe-
dent and then universally quantifying. QED
Proposition 2.13 If LSC(t), or if X is finite, or also if R, S, T have only

a finite number of membership degrees, then it holds true that
R ◦t (S ◦t T ) = (R ◦t S) ◦t T.
Proof. The way of proving this equation is routine. The essential point is
to prove that the formulas
∃u((x, u) ε R ∧t ∃v((u, v) ε S ∧t (v, y) ε T ))
and
∃v(∃u((x, u) ε R ∧t (u, v) ε S) ∧t (v, y) ε T )
have the same truth degrees. To get that result we have – besides using the
associativity of ∧t – first to “move the quantifier ∃v outside” a conjunction
2.2. FUZZY RELATIONS 61
and then to “move ∃u inside” a corresponding conjunction. As stated in

proposition 1.28 (i) this is possible without a change of truth degrees in the
case of LSC(t). Obviously, these changes of quantifiers are unproblematic in
those cases where the existential quantifiers ∃u, ∃v correspond to the opera-
tions of taking a maximum of truth degrees, which happens if only a finite
number of truth degrees are possible for the formulas following ∃u or ∃v,
and this happens e.g. for finite universes of discourse X and for those fuzzy
relations R, S, T with only a finite number of membership degrees really ap-
pearing, i.e. for those membership functions µR , µS , µT of fuzzy relations
which have, as usual functions, finite ranges. QED
Unfortunately, in general we cannot avoid the assumption LSC(t) in this

proposition. To get a counterexample in the case that LSC(t) fails, let us
choose t = tD = “drastic product” and as universe of discourse X = IN =
“set of natural numbers (without zero!)”. Consider the fuzzy relations
∞
[ ∞
[
R = hh(1, 5)ii0.9 S= hh(5, n)ii(1− 1 ) , T = hh(n, 1)ii1 .
n
n=1 n=1
Writing R ◦D S instead of R ◦tD S we get successively
S ◦D T = hh(5, 1)ii1 , R ◦D (S ◦D T ) = hh(1, 1)ii0.9
and on the other side
R ◦D S = ∅, (R ◦D S) ◦D T = ∅.6
As a further remark let us add here that the assumptions on the finiteness
of X or of the set of membership degrees of R, S, T may also in the following
propositions be added instead of LSC(t). However, for simplicity we mention
only the case of LSC(t).
6
By the way, if we consider additionally the t-conorm ∨ = max, then we have for all
u, v, w ∈ [0, 1]
(u ∨ v) tD w = (u tD w) ∨ (v tD w)
and hence property (1) of Ovchinnikov (1981). Therefore our example too is a coun-
terexample to the unrestricted associativity of the composition of fuzzy relations (as stated
erroneously in Ovchinnikov (1981; p. 172)).
Proposition 2.14 Consider t-norms t and t1 such that t distributes over

t1 , and a t-conorm st2 such that t also distributes over st2 . It then holds
true that
(i) |= R ◦t (S ∩t1 T ) ⊂
= (R ◦t S) ∩t1 (R ◦t T ),
(ii) ⊂
|= R ◦t (S ∪t2 T ) = (R ◦t S) ∪t2 (R ◦t T ).
Proof. By distributivity of t over t1 we get

R ◦t (S ∩t1 T )
= {(x, y) k ∃z((x, z) ε R ∧t ((z, y) ε S ∧t1 (z, y) ε T ))}
= {(x, y) k ∃z(((x, z) ε R ∧t (z, y) ε S) ∧t1 ((x, z) ε R ∧t (z, y) ε T ))}
which gives (i) by the definition of the union of fuzzy relations and the fact
that
[[∃z(H1 ∧t1 H2 )]] ≤ [[∃zH1 ∧t1 ∃zH2 ]]
for all formulas H1 , H2 and t-norms t1 .
(ii) is proved by corresponding calculations now using the fact that
[[∃z(H1 ∨t2 H2 )]] ≤ [[∃zH1 ∨t2 ∃zH2 ]]
for all formulas H1 , H2 and t-conorms st2 . QED
If we take t = t1 in this proposition, the condition of the distributivity

of t over t equivalently reduces to the idempotency of t. For, clearly, if t is
idempotent, then t distributes over itself, i.e.
u t (v t w) = (u t v) t (u t w) (2.26)
always holds true; on the other hand if (2.26) always holds true then put
v = w = 1 which gives u = u t u and hence the idempotency of t which
finally means t = min.
Proposition 2.15 For any t-norm t with property LSC(t) the generalized
monotonicity of the relational product of fuzzy relations holds true in the
forms
(i) |= R ⊂ ⊂
=t S → R ◦t T =t S ◦t T,
(ii) |= S ⊂ ⊂
=t T → R ◦t S =t R ◦t T,
(iii) |= R1 ⊂ ⊂ ⊂
=t R2 ∧t S1 =t S2 →t R1 ◦t S1 =t R2 ◦t S2 .
2.3. THE FULL IMAGE UNDER A RELATION 63
Proof. We consider only (i), case (ii) is analogous. Suppose LSC(t). Then
we have
|= R ⊂
=t S →
∀x∀y∀z((x, y) ε R ∧t (y, z) ε T →t (x, y) ε S ∧t (y, z) ε T )
because of proposition 1.18 (i) one always has
|= (H1 →t H2 ) → (H1 ∧t G →t H2 ∧t G),
furthermore obviously
|= R ⊂
=t S → ∀x∀y∀z((x, y) ε R ∧t (y, z) ε T →t (x, z) ε S ◦t T )
and therefore, using the fact that in the case of LSC(t)
|= ∀x∀y∀z((x, y) ε R ∧t (y, z) ε T →t (x, z) ε S ◦t T )

→ ∀x∀z(∃y((x, y) ε R ∧t (y, z) ε T ) →t (x, z) ε S ◦t T )
(cf. proposition 1.27 (ii)), we get by transitivity of →
|= R ⊂
=t S → ∀x∀z((x, y) ε R ◦t T →t (x, z) ε S ◦t T )
which had to be proved.
To get (iii) we have to use (i), (ii) together with the transitivity of fuzzy
inclusion for fuzzy relations:
|= R ⊂ ⊂ ⊂
=t S ∧t S =t T →t R =t T
which is nothing else than a special case of the corresponding result for fuzzy
sets in general. QED
2.3 The full image under a relation

In classical set theory, having given a set of ordered pairs, one always has
two ways of looking at it: one may consider this set as a (binary) relation or
as a (unary) mapping.7 Depending on that point of view one is interested
in different properties of and operations with sets of ordered pairs.
7
We follow the usage to speak of a mapping also in the case that there is no uniqueness
with respect to the second argument, mappings are thus “multiple valued” functions. We
reserve the name of a “function” as usual for the unique mappings.
The same situation, of course, appears with fuzzy sets of ordered pairs.
So far we have looked at them as fuzzy relations. For the purpose of the
present book this perspective remains the essential one. Nevertheless, there
is one notion from the realm of the mapping-related ones that has to be
considered together with fuzzy relations: the full image of a (fuzzy) set
under a (fuzzy) relation. And the background for this interest is in the later
use of the so-called “compositional rule of inference” in connection with our
discussions of fuzzy controllers.
From a more set theoretical point of view the notions and facts from
the mapping-related world are as essential and interesting as those from
the relation-related world. This also holds true in the field of fuzzy sets.
Moreover, in the mapping-related world of fuzzy sets there arise completely
new problems, e.g. at first giving a satisfactory definition of uniqueness that
is nontrivial in the sense that it refers to an equality predicate that allows
for truth degrees different from the degrees 1 and 0.8
Definition 2.9 For each t-norm t, every fuzzy relation R ∈ IF (X × Y) and

all fuzzy sets A ∈ IF (X ) and B ∈ IF (Y), let
R00 A =def {y k ∃x(x ε A ∧t (x, y) ε R)},
and in the case that t has the property LSC(t) also
R ↓ B =def {x k ∀y((x, y) ε R →t y ε B)}.
Obviously, R00 A is a fuzzified version of the full image of a set under a

mapping, i.e. under a binary relation. Simultaneously, R00 A corresponds to
what Bellman/Zadeh (1970) called “conditioned fuzzy set”. And R ↓ A
corresponds to the ° α -operation of Sanchez (1984). In some (weak) sense
this operation R ↓ B is dual to taking the full image R00 A of A under R.
As a side remark let us mention that this fuzzified full image of a fuzzy
set under a fuzzy relation could be used to define the range and hence, cf.
proposition 2.12 (i), also the domain of a fuzzy relation. The reason is that
rg(R) = R00 (X [1] ) (2.27)
holds true for any fuzzy relation R ∈ IF (X × Y).
Usually the t-norm used in the definition of R00 A is clear from the context
and the same one which is (essentially) used there. Thus we shall not mention
it explicitely in our notation.
8
For work in this direction the interested reader is referred to Gottwald (1980) as one
possible approach.
2.3. THE FULL IMAGE UNDER A RELATION 65
Proposition 2.16 For all fuzzy sets A, B ∈ IF (X ) and each fuzzy relation
R ∈ IF (X × Y) there hold true
(i) |= R00 (A ∪ B) ≡ R00 A ∪ R00 B,

(ii) |= R00 (A ∩ B) ⊂ 00 00
= R A ∩ R B;
in the case that R00 .. is defined using the t-norm tG = min it also holds true
(iii) |= R00 A ∪t R00 B ⊂ 00

= (R ∪t R) (A ∪t B).
Proof. (i) By definition 2.9 and the distributivity of any t-norm t over ∨
according to proposition 1.12 one has
R00 A ∪ R00 B = {y k ∃x(x ε A ∧t (x, y) ε R) ∨ ∃x(x ε B ∧t (x, y) ε R)}

= {y k ∃x((x ε A ∧t (x, y) ε R) ∨ (x ε B ∧t (x, y) ε R))}
= {y k ∃x((x ε A ∨ x ε B) ∧t (x, y) ε R)}
= R00 (A ∪ B).
(ii) follows by the same type of calculations but now using proposition
1.26 (ii) for the step from the first to the second line, giving the (nonfuzzy)
inclusion “⊃” instead of equality.
(iii) With reference to proposition 1.26 (iv) the calculations to determine
R00 A ∪t R00 B can start as in case (i) and proceed as in (ii), but now with an
inclusion “⊂” resulting, and have later on to use the distributivity of any
t-conorm over ∧ and that therefore one has
|= (H1 ∧ H3 ) ∨t (H2 ∧ H3 ) ↔ (H1 ∨t H2 ) ∧ (H3 ∨t H3 )
for all expressions H1 , H2 , H3 . The rest is routine matter. QED
The inspection of the proofs for the claims (i) and (ii) of this propo-
sition indicates a further possibility of generalization: under some suitable
assumptions ∪ and ∩ can be replaced by some ∪t or some ∩t respectively.
We state the result without further proof.
Proposition 2.17 Consider fuzzy sets A, B ∈ IF (X ) and a fuzzy relation

R ∈ IF (X × Y). Suppose that the definition of R00 .. uses a t-norm t1 and
that another t-norm t is considered too.
(i) If t1 distributes over st then it holds true
|= R00 (A ∪t B) ⊂ 00 00
= R A ∪t R B.
(ii) If t1 distributes over t or if at least the subdistributivity condition

u t1 (v t w) ≤ (u t1 v) t (u t1 w) holds true for all u, v, w ∈ [0, 1], then one has
|= R00 (A ∩t B) ⊂ 00 00
= R A ∩t R B.
For the following considerations let us suppose for simplicity that X = Y

and thus all fuzzy sets are from IF (X ) and all fuzzy relations from IF (X ×X ).
Proposition 2.18 Suppose LSC(t). Then it holds true for all fuzzy sets
A, B and all fuzzy relations R, S that
(i) |= A ⊂
=t B → R00 A ⊂ 00
=t R B,
(ii) |= R ⊂
=t S → R00 A ⊂ 00
=t S A,
(iii) |= A ⊂
=t B→R↓A⊂ =t R ↓ B,
(iv) |= R ⊂
=t S→S↓A⊂ =t R ↓ A.
Proof. (i) Because of proposition 1.13 the logical validity of a generalized

implication is equivalent to an inequality of the corresponding truth degrees.
What has to be shown therefore is that
[[A ⊂ 00 ⊂ 00
=t B]] ≤ [[R A =t R B]].
We start from the right hand side of this inequality and have by the corre-
sponding definitions
[[R00 A ⊂ 00
=t R B]]
= [[∀x(x ε R00 A →t x ε R00 B)]]
= [[∀x(∃y(y ε A ∧t (y, x) ε R) →t ∃z(z ε B ∧t (z, x) ε R))]]
and using proposition 1.27 (ii) furthermore
[[R00 A ⊂ 00
=t R B]]
= [[∀x∀y((y ε A ∧t (y, x) ε R) →t ∃z(z ε B ∧t (z, x) ε R))]].
Now from (1.2) together with (Φ1) and proposition 1.18 (i) we successively
get
[[R00 A ⊂ 00
=t R B]]
≥ [[∀x∀y((y ε A ∧t (y, x) ε R) →t (y ε B ∧t (y, x) ε R))]]
≥ [[∀x∀y(y ε A →t y ε B)]]
2.4. SPECIAL TYPES OF FUZZY RELATIONS 67
which by dropping the empty quantification “∀x” just means
[[R00 A ⊂ 00 ⊂
=t R B]] ≥ [[A =t B]].
The same kind of calculations yield (ii). We omit these details.

(iii) As before it has to be shown that the inequality
[[A ⊂ ⊂
=t B]] ≤ [[R ↓ A =t R ↓ B]]
holds true. But now the corresponding definitions lead to the following
estimations
[[R ↓ A ⊂=t R ↓ B]]

= [[∀x(∀y((x, y) ε R →t y ε A) →t ∀z((x, z) ε R →t z ε B))]]
= [[∀x∀z(∀y((x, y) ε R →t y ε A) →t ((x, z) ε R →t z ε B))]]
≥ [[∀x∀z(((x, z) ε R →t z ε A) →t ((x, z) ε R →t z ε B))]]
≥ [[∀x∀z(z ε A) →t z ε B)]]
which use the antimonotonicity of ϕt in the first argument, i.e. proposition

1.4 (i), as well as proposition 1.18 (ii).
Finally (iv) can be proven along the same line of arguments. But now
instead of proposition 1.18 (ii) reference has to be made to proposition 1.18
(iii) which causes the reversal in the order of appearance of R and S on both
sides of the arrow → in the formula. QED
2.4 Special types of fuzzy relations

In order to be able to discuss some special types of fuzzy relations in the fol-
lowing, we first have to introduce some specific properties of fuzzy relations.
In this section these shall always be crisp properties, i.e. properties which a
fuzzy relation either has or does not have. Different lists of such properties
can be found in for example Gottwald (1974), Dubois/Prade (1980),
Ovchinnikov (1981), and Chakraborty/Das (1983). As we only con-
sider binary relations, we will change our notation slightly and write – as for
usual binary relations – simply xRy for the truth degree (x, y) ε R, i.e. we
use [[xRy]] = µR (x, y).
Definition 2.10 A fuzzy relation R is (with respect to a given t-norm t and

a negation function n)
(i) reflexive (in X ) iff |= ∀x(xRx),
(ii) irreflexive iff |= ∀x ∼n (xRx),
(iii) t-transitive iff |= ∀x∀y∀z(xRy ∧t yRz →t xRz),
(iv) symmetric iff |= ∀x∀y(xRy →t yRx),
.
(v) t-antisymmetric iff |= ∀x∀y(xRy ∧t yRx →t x = y),
(vi) t-asymmetric iff |= ∀x∀y ∼n (xRy ∧t yRx).
Remark. By the usual truth condition (1.6) for logical validity and defini-
tion (1.1) of the many-valued universal quantifier obviously the irreflexivity
condition becomes: [[xRx]] = 1 for each x ∈ X . Therefore the definition 2.10
(ii) of irreflexivity and the definition 2.10 (vi) of t-asymmetry are indepen-
dent of the choice of a specific negation function.
As an additional consequence the symmetry condition (iv) becomes that:
[[xRy →t yRx]] = 1 for all x, y ∈ X . As long as one supposes that the t-
norm t which is involved here has the property LSC(t), the condition reads:
[[xRy]] ≤ [[yRx]] = 1 for all x, y ∈ X . Thus the definition 2.10 (iv) is also
independent of the choice of a specific lower semicontinuous t-norm.
As usual these properties may also be characterized using the set theo-
retic operations for fuzzy relations. To do this we need a kind of fuzzified
diagonal. Let the fuzzy diagonal of X be the fuzzy relation
[ .
∆X =def hh(a, a)ii1 = {(x, y) ∈ X × X k x = y}. (2.28)
a∈X
Proposition 2.19 For each fuzzy relation R it holds true that

(i) R reflexive (in X ) iff |= ∆X ⊂= R,
(ii) R irreflexive iff |= R ∩t ∆X ≡ ∅,
(iii) R t-transitive iff |= R ◦t R ⊂
=t R
iff |= R ◦t R ⊂
= R,
(iv) R symmetric ⊂ −1
iff |= R = R ,
(v) R t-antisymmetric iff |= R ∩t R−1 ⊂ = ∆X ,
(vi) R t-asymmetric iff |= R ∩t R−1 ≡ ∅.
Proof. Obvious.
First we look for a fuzzification of equivalence relations. Such fuzzified

equivalence relations have been discussed, for example by Klaua (1970),
Zadeh (1971), Ovchinnikov (1981) and Chakraborty/Das (1983a).

We will give only the most basic facts.
Definition 2.11 A fuzzy equivalence relation R (in X ) is such a fuzzy

relation which is reflexive (in X ), symmetric and t-transitive. For each
fuzzy equivalence relation R and each b ∈ X the fuzzy R-equivalence class of
b is
hbiR =def {x k xRb}.
By the symmetry of each fuzzy equivalence relation it is immediately
hbiR = {x k bRx}
and using the notion of the full image from definition 2.9 we hence have
hbiR = R00 hhbii1
for the fuzzy R-equivalence classes.
Proposition 2.20 For each fuzzy equivalence relation R, each t-norm t with
the property LSC(t) and all a, b ∈ X , there hold true:
(i) |= a ε haiR ,
(ii) |= b ε haiR ↔t aRb,
(iii) |= haiR ≡t hbiR ↔t aRb,
(iv) |= ∃x(x ε haiR ∩t hbiR ) ↔t aRb.
Proof. (ii) is simply a restatement of the definition of haiR ; and (i) follows
from this definition by reflexivity of R.
(iii) From definition of ≡t and from (ii) one immediately has
|= haiR ≡t hbiR →t (aRa ↔t aRb)
and hence by reflexivity of R
|= haiR ≡t hbiR →t aRb;
furthermore by transitivity of R and symmetry
|= aRb →t (xRa →t xRb),

|= aRb →t (xRb →t xRa).
Therefore proposition 1.6 (i) immediately gives
|= aRb →t ((xRa →t xRb) ∧ (xRb →t xRa)).
Because of the linear ordering of the set of truth degrees and of proposition
1.4 (ii) and condition (T3) for all formulas H, G one has
[[(H →t G) ∧ (G →t H)]] = [[(H →t G) ∧t (G →t H)]]
which in the present case gives
|= aRb →t ((xRa ↔t xRb)
and hence, e.g. using proposition 1.27 (i),
|= aRb →t haiR ≡t hbiR .
Thus (iii) is proved.

(iv) By (i) and (ii) it holds true that
|= aRb →t b ε haiR ∩t hbiR
and hence of course as part of (iv) already
|= aRb →t ∃x(x ε haiR ∩t hbiR ). (2.29)
Otherwise we also have
|= x ε haiR ∩t hbiR →t xRa ∧t xRb
and thus together with t-transitivity and symmetry of R
|= x ε haiR ∩t hbiR →t aRb
which together with LSC(t) gives
|= ∃x(x ε haiR ∩t hbiR →t aRb). (2.30)
Thus finally (iv) is proved by (2.29) and (2.30). QED
Our next concern will be the connection of fuzzy equivalence relations

with generalized partitions of the universe of discourse.
Definition 2.12 A fuzzy partition of (the universe of discourse) X is such

a crisp (!) class Z of fuzzy subsets of X such that
[
(Z1) A=X
A∈Z
and for all A, B ∈ Z:
(Z2) |= ∃x(x ε A ∩t B) →t A ≡t B.
Proposition 2.21 Suppose LSC(t).

(i) If R is a fuzzy equivalence relation (in the universe X ), then the set
of R-equivalence classes
Z/R = {haiR | a ∈ X }
is a fuzzy partition of X .
(ii) If Z is a fuzzy partition of X , then a fuzzy equivalence relation (in
X ) is defined by
_
_ _
_
RZ = {(a, b) k (a ε A ∧t b ε B ∧t A ≡t B)}.
A∈Z B∈Z
Proof. (i) Because from proposition 2.20 (i) one has [[a ε haiR ]] = 1 for each
a ∈ X , we have that R has property (Z1). Property (Z2) is a straightforward
consequence of proposition 2.20 (iii) and (iv).
(ii) Obviously RZ is symmetric. For the reflexivity of RZ one has to
consider
_
_
[[(a, a) ε RZ ]] = [[ (a ε A ∧t a ε B ∧t A ≡t B)]]
A,B∈Z
_
_
= [[ (a ε A ∧t a ε A)]]
A∈Z
and hence to prove that
sup [[a ε A ∧t a ε A]] = 1. (2.31)

A∈Z
But for (2.31) it is necessary and sufficient to have
sup [[a ε A]] = 1

A∈Z
and that is precisely condition (Z1).

In order to prove the t-transitivity of RZ consider any a, b, c ∈ X . Then
[[(a, b) ε RZ ∧t (b, c) ε RZ ]] =
__
[[ (a ε A ∧t b ε B ∧t A ≡t B ∧t
A,B,C,D∈Z
∧t b ε D ∧t c ε C ∧t D ≡t C)]].
By (Z2) we easily get
[[b ε B ∧t b ε D]] ≤ [[B ≡t D]]
and from this inequalitiy together with
[[A ≡t B ∧t B ≡t D ∧t D ≡t C]] ≤ [[A ≡t C]]
by t-transitivity of ≡t and (T2), i.e. monotonicity of the t-norm t,

_
_
[[(a, b) ε RZ ∧t (b, c) ε RZ ]] ≤ [[ (a ε A ∧t c ε C ∧t A ≡t C)]]
A,C∈Z
≤ [[(a, c) ε R/Z]]
Hence the fuzzy relation R is transitive too. QED
Our approach to fuzzy equivalence relations – or: (fuzzy) similarity re-

lations using the terminology of Zadeh (1971) – is a straightforward gen-
eralization of the corresponding approaches of Zadeh (1971) and of Klaua
(1970). The approach gives a fuzzification of the notion of equivalence rela-
tion in a given universe of discourse. What classically comes to the same is to
consider an equivalence relation in some crisp set: one only has to interpret
that crisp set as the universe of discourse.
Unfortunately, for fuzzy sets the situation is not so simple. It would
be nice to have not only a theory of fuzzy equivalence relations in a given
universe of discourse, but also a theory of fuzzy equivalence relations in a
given fuzzy set A, in the sense that the fuzzy equivalence relation in A is
connected with any kind of (generalized – i.e. fuzzy) partition of A. Yet, so
far there has been no success in this problem.
As the next topic, let us discuss the construction of the transitive hull
of a given fuzzy relation. As with fuzzy equivalence relations, the approach
is mainly along the same lines as Zadeh (1971), but as before we get some
more general results.
Definition 2.13 For fuzzy relations R and any given t-norm t we put re-
cursively for all natural numbers n
R1 =def R,
n+1
R =def Rn ◦t R,
and take as the t-transitive hull of R the fuzzy relation
∞
[
Tt (R) =def Rn .
n=1
Proposition 2.22 For each fuzzy relation R we have

(i) |= R ⊂
= Tt (R),
(ii) R t-transitive ⇒ R = Tt (R),
and if the t-norm t has property LSC(t)
(iii) Tt (R) is t-transitive,
(iv) Tt (Tt (R)) = Tt (R).
Proof. (i) is obvious. (ii) hence is proved if we show |= Tt (R) ⊂ = R for t-

transitive fuzzy relations R. But if R is t-transitive, i.e. if |= R ◦t R ⊂
= R, we
3 2 ⊂
have by R = R ◦t R and by the transitivity of = immediately from this
t-transitivity: |= R3 ⊂ n ⊂
= R. Recursively we get |= R = R for each integer
⊂
n ≥ 1. Therefore obviously |= Tt (R) = R and (ii) is proved.
For (iii) let m, n be any natural numbers. Using LSC(t) and the definition
of Rm in a straightforward way, one gets
[[(x, y) ε Rn ∧t (y, z) ε Rm ]] ≤ [[(x, z) ε Rn+m ]]
and thus
|= (x, y) ε Rn ∧t (y, z) ε Rm →t (x, z) ε Tt (R).
Now taking suprema on all n ≥ 1 and on all m ≥ 1 gives by LSC(t)
|= (x, y) ε Tt (R) ∧t (y, z) ε Tt (R) →t (x, z) ε Tt (R).
i.e. the t-transitivity of Tt (R). By (ii) and (iii) now (iv) is obvious. QED
We can prove some relatively weak results on the monotonicity of the Tt -

operator for the general case of any t-norm. More interesting monotonicity
properties are provable in the case that t = min.
Proposition 2.23 For fuzzy relations R, S we have:
(i) if S is t-transitive and |= R ⊂ ⊂

= S, then |= Tt (R) = S;
and with the additional assumption LSC(t)
(ii) if |= R ⊂ ⊂
=t S, then |= Tt (R) =t Tt (S).
Proof. If S is t-transitive then |= S n ⊂

= S for each n ≥ 1 as mentioned in
the foregoing proof. Furthermore
|= R2 ⊂
= R ◦t S and |= R ◦t S ⊂
=S
2
by |= R ⊂
= S and the monotonicity of ◦t . Hence together that gives
|= R2 ⊂
= S;
and in the same way more generally
|= Rn ⊂
=S
for each n ≥ 1. Therefore obviously
|= Tt (R) ⊂
=S
and hence (i).

By proposition 2.22 (i) we have |= R ⊂ ⊂
= Tt (S) from |= R = S. Because
of the t-transitivity of the fuzzy relation Tt (S) the result (ii) follows from
part (i). QED
This proposition discusses the situation that, either with or without the
t-transitivity of the fuzzy relation S, one has that the graded inclusion
“R ⊂=t S” has truth degree 1. Then the statement is that some other in-
clusions have truth degree 1 too. More general and thus more interesting
would be to have instead for example inequalities between the degrees to
which these fuzzy inclusion relations hold true. Such results will only now
be established for a special case.9
9
At present it is an open problem whether the results of the next proposition can be
extended to other t-norms besides t = min or not.
Proposition 2.24 For the t-norm t = tG = min one has for all fuzzy
relations R, S in the case that S is tG -transitive:
(i) |= R ⊂ ⊂
=tG S → TtG (R) =tG S,
and without any transitivity restraint for S the generalized monotonicity
(ii) |= R ⊂ ⊂
=tG S → TtG (R) =tG TtG (S).
Proof. (i) Suppose S to be tG -transitive, i.e. to be min-transitive. Then
one has (always deleting the index tG ):
|= R ⊂ 2 ⊂
= S →R = R◦S
|= R ⊂
= S →R◦S = S
⊂ 2
and hence together with |= S 2 ⊂= S:

|= R ⊂ 2 ⊂
= S → R = S.
Continuing this line of reasoning gives
|= R ⊂
=S→R =S
n ⊂
for each n ≥ 1. From this one gets together with

∞
[
inf [[R ⊂
n
= S]] = [[ Rn ⊂
= S]]
n≥1
n=1
the fact that
|= R ⊂ ⊂
= S → T (R) = S.
(ii) Again deleting all indices tG we have from proposition 2.22 (i)
|= R ⊂ ⊂ ⊂
= S → R = S ∧ S = T (S)
and hence
|= R ⊂ ⊂
= S → R = T (S).
Now part (i) of our proposition gives because of the transitivity of T (S)
|= R ⊂ ⊂
= T (S) → T (R) = T (S)
and thus
|= R ⊂ ⊂
= S → T (R) = T (S)
altogether. QED
As a third topic let us discuss fuzzy ordering relations. Again following

Zadeh (1971) and generalizing Gottwald (1974) we give the following
Definition 2.14 A fuzzy relation which is reflexive and t-transitive is called

fuzzy preordering; and a reflexive, t-transitive, and t-antisymmetric fuzzy
relation is called fuzzy partial ordering.
For fuzzy relations and for crisp relations it is enough to discuss the
reflexive case: in a simple way one can change from reflexive fuzzy ordering
relations to corresponding irreflexive ones in essentially the same manner as
in classical set theory.
Definition 2.15 For each t-norm t and each fuzzy relation R ∈ IF (X × X )

let
R+ =def R ∪t ∆X ,
−
R =def R ∩t C
C∆X .
Corollary 2.25 For each t-norm t and each fuzzy relation R one has
(i) R− is irreflexive,
(ii) R+ is reflexive.
Proof. Obvious.
Proposition 2.26 Let R be any fuzzy relation and t any t-norm with the
property LSC(t). Then one has
(i) If R is t-transitive then R+ is t-transitive too.
(ii) If R− is t-transitive then R is t-transitive too.
(iii) If R is t-transitive and t-antisymmetric then R− is t-transitive.
(iv) If R is t-antisymmetric then R− is t-asymmetric.
Proof. Straightforward from the corresponding definitions.
For fuzzy ordering relations these results give the above-mentioned simple
connection between the reflexive and the irreflexive case. To state that result
we adopt here the
Definition 2.16 A fuzzy relation R ∈ IF (X × X ) is an irreflexive fuzzy

partial ordering iff R is an irreflexive, t-transitive and t-asymmetric relation.
Proposition 2.27 For each fuzzy relation R and any t-norm t with property
LSC(t) one has:
(i) If R is a fuzzy partial ordering then R− is an irreflexive fuzzy partial

ordering.
(ii) If R is an irreflexive fuzzy partial ordering then R+ is a fuzzy partial

ordering.
The proof is by straightforward calculations and will not be given in

detail. But because of this result we confine ourselves to the case of reflexive
fuzzy ordering relations.
To formulate some of our following results we refer to the terminology of
for example ring theory and say that a t-norm t has zero divisors iff there
exist u, v ∈ (0, 1] such that u t v = 0; in such a case u, v itself are called zero
divisors; cf. also Ovchinnikov (1981).
Obviously there are t-norms which have zero divisors. The L Ã ukasiewicz
t-norm tL is an example: because of
tL (u, v) = max{0, u + v − 1}
a pair (u0 , v0 ) with 0 < u0 , v0 ≤ 1 is a pair of zero divisors of tL iff u0 +v0 ≤ 1.

In addition, the t-norm tP is an example of a t-norm which does not have
zero divisors.
Furthermore, the ordering relation (1.12) of t-norms gives additional in-
formation concerning t-norms with zero divisors. One easily recognizes that
for all t-norms t1 , t2
t1 <
= t2 and t2 has zero divisors ⇒ t1 has zero divisors (2.32)
holds true. Therefore also
t1 <
= t2 and t1 is without zero divisors
⇒ t2 is without zero divisors (2.33)
All together that gives for each t-norm t:
t<= tL ⇒ t has zero divisors, (2.34)

tP <
= t ⇒ t is without zero divisors. (2.35)
Sometimes the considerations are restricted to Archimedean t-norms,

cf. Weber (1983). A t-norm t is called Archimedean iff t is a continuous
function and for each 0 6= u ∈ [0, 1] one has u t u < u. Each Archimedean
t-norm t has a generating function f : [0, 1] → [0, ∞] which is decreasing
and continuous with f (1) = 0 such that always
u t v = f (−1) (f (u) + f (v)).
Here f (−1) : [0, ∞] → [0, 1] is the pseudo-inverse of f defined as

½
f (−1) (y) =def f −1 (y), if y ∈ rg(f ),
0 otherwise.
Thus one has
u t v = 0 ⇔ f (u) + f (v) ≥ f (0).
That means that for each Archimedean t-norm t with generating function f
one has
t has zero divisors ⇔

f (0) ≤ f (u0 ) + f (v0 ) for some u0 , v0 ∈ (0, 1].
Occasionally the subclass of strict t-norms is also discussed. A t-norm is

called strict iff it is Archimedean and strictly monotonously increasing in
[0, 1]. Then obviously one has for each t-norm t:
t strict ⇒ t without zero divisors.
Proposition 2.28 If R is a fuzzy partial ordering (preordering) then the

crisp relation
R≥1 = {(x, y) ∈ supp(R) | [[xRy]] = 1}
is a partial ordering (preordering) in the usual sense.
Proof. Straightforward from definitions.
Unfortunately a corresponding result for supp (R) instead of the kernel

R≥1 holds true only with restrictions on the t-norm used in the formulation
of transitivity and antisymmetry properties.
Proposition 2.29 If the t-norm t has zero divisors and X at least three
elements, then there exists a fuzzy partial ordering R0 such that the usual
crisp relation supp(R0 ) is not transitive.
Proof. Consider a, b, c ∈ X and take zero divisors u, v ∈ (0, 1] such that

utv = 0. Then
R0 = hh(a, b)iiu ∪ hh(b, c)iiv ∪ ∆X
is a fuzzy partial ordering, but the crisp relation supp (R0 ) = {(a, b), (b, c)}
is not transitive (in the usual sense). QED
Additionally we mention that already in the case that card (X ) ≥ 2 and

t has zero divisors, a fuzzy partial ordering may be found whose support is
not an antisymmetric relation (in the usual sense).
Proposition 2.30 Suppose that the t-norm t does not have zero divisors.
Then for each fuzzy partial ordering (preordering) R the support supp(R) is
a crisp partial ordering (preordering) in the usual sense.
Proof. Straightforward from definitions.
Proposition 2.31 (i) If R, S are fuzzy partial orderings (preorderings) then

for each t-norm t the fuzzy relation R ∩t S too is a fuzzy partial ordering
(preordering).
(ii) Assume LSC(t). Consider any sequence (Ri )i<α of type α, α any
ordinal number, of fuzzy partial orderings (preorderings) such that for all
indices i, j < α:
if i < j then |= Ri ⊂
=t Rj .
Then the fuzzy relation
α
[
R̄ = Ri
i=1
is a fuzzy partial ordering (preordering) too.
Proof. (i) The reflexivity of R ∩t S is obvious. For the t-transitivity we

have
[[x(R ∩t S)y ∧t y(R ∩t S)z]]

= [[xRy ∧t xSy ∧t yRz ∧t ySz]]
≤ [[xRz ∧t xSz]] = [[x(R ∩t S)z]]
by the t-transitivity of R and S, and thus the t-transitivity of R ∩t S results.

If both R, S are t-antisymmetric the t-antisymmetry of R ∩t S follows in the
same way.
(ii) Obviously R̄ is reflexive. Furthermore we have by LSC(t) for all
x, y, z ∈ X :
[[xR̄y ∧t y R̄z]] = (sup[[xRi y]]) t (sup[[yRj z]])

i<α i<α
= sup [[xRi y ∧t yRj z]]
i,j<α
≤ sup [[xRmax{i,j} z]] = [[xR̄z]]
i,j<α
and hence the t-transitivity of R̄. Again, if all fuzzy relations Ri are t-
antisymmetric then the t-antisymmetry of R̄ follows by corresponding cal-
culations. QED
Proposition 2.32 If R is a fuzzy preordering then the fuzzy relation
Q =def {(x, y) k xRy ∧t yRx} (2.36)
is a fuzzy equivalence relation.
Proof. Reflexivity and symmetry of Q are obvious by definition (2.36) and

the reflexivity of R. And the t-transitivity of Q is a simple consequence of
the t-transitivity of R. QED
Proposition 2.33 Let R be a fuzzy preordering and define the fuzzy equiv-
alence relation Q as in (2.36). Then by
R̂ =def {(haiQ , hbiQ ) k aRb}
a fuzzy relation is defined in the quotient set X /Q which is a fuzzy partial

ordering.
Proof. First we have to show that R̂ is correctly defined, i.e. that this
definition does not really depend on the elements a, b which describe the
fuzzy equivalence classes haiQ , hbiQ .
Hence consider a, a0 such that haiQ = ha0 iQ . Then by [[a ε haiQ ]] = 1 one
has [[a ε ha0 iQ ]] = 1, i.e. [[aQa0 ]] = 1. Hence immediately [[aRa0 ]] = [[a0 Ra]] =
1. Now we get for every x
[[aRx]] = [[a0 Ra ∧t aRx]] ≤ [[a0 Rx]] = [[aRa0 ∧t a0 Rx]] ≤ [[aRx]]
by the t-transitivity of R and thus [[aRx]] = [[a0 Rx]]. Therefore
[[haiQ R̂hbiQ ]] = [[aRb]] = [[a0 Rb]] = [[ha0 iQ R̂hbiQ ]].
In the same way the independence of the choice of b is established.

The reflexivity of R is obvious. The t-transitivity results from the fact
that for all fuzzy equivalence classes haiQ , hbiQ , hciQ :
[[haiQ R̂hbiQ ∧t hbiQ R̂hciQ ]]

= [[aRb ∧t bRc]] ≤ [[haiQ R̂hciQ ]].
Thus the antisymmetry of R̂ remains. But immediately

.
[[haiQ R̂hbiQ ∧t hbiQ R̂haiQ ]] = [[aRb ∧t bRa]] ≤ [[a = b]]
and obviously
. .
[[a = b]] ≤ [[haiQ = hbiQ ]].
Thus R̂ is a fuzzy partial ordering. QED
As a side remark let us remember that according to proposition 2.20 (iii)

in case LSC(t) we have
|= haiQ ≡t hbiQ ↔t aQb
i.e. in the present situation
|= haiQ ≡t hbiQ ↔t aRb ∧t bRa.
If we assume that our universe of discourse X is already a class of fuzzy

sets, which means that our fuzzy subsets of X and our fuzzy relations in
X are fuzzy sets of higher level in the terminology of Gottwald (1979),
then it seems reasonable to discuss another, suitably changed version of
antisymmetry of such a fuzzy relation R characterized through the definition
|= ∀x∀y(xRy ∧t yRx →t x ≡t y).

If in that case we repeat the construction of R̂ from proposition 2.33 we find

that this fuzzy relation R̂ is antisymmetric in this new sense too: we have
[[haiQ R̂hbiQ ∧t hbiQ R̂haiQ ]] = [[aRb ∧t bRa]] ≤ [[haiQ ≡t hbiQ ]]
and hence
|= ∀x∀y(hxiQ R̂hyiQ ∧t hyiQ R̂hxiQ →t hxiQ ≡t hyiQ ).
As our last topic in this section we discuss a fuzzification of the well-

known Szpilrajn’s theorem on the extension of a partial ordering to a
linear ordering. We get here a generalization of the fuzzy version of this
theorem as proved in Zadeh (1971).
Definition 2.17 Given any t-conorm st1 we denote a fuzzy relation R as

st1 -linear iff
|= ∀x∀y(xRy ∨t1 yRx),
and we denote R as weakly linear iff supp(R) is a linear relation in the usual
sense. Additionally we call R a st1 -linear (weakly linear) fuzzy ordering iff
R is a fuzzy partial ordering and also st1 -linear (weakly linear).
Proposition 2.34 Suppose that the t-norm t does not have zero divisors.
Then for every fuzzy partial ordering R there exists a st1 -linear fuzzy or-
dering SR (st1 any t-conorm) such that |= R ⊂ =t SR . If the fuzzy partial
ordering R has the further property that there exists u ∈ (0, 1] such that for
all a, b ∈ X :
(a, b) ∈ supp(R) ⇒ [[aRb]] ≥ u,
then there exists a weakly linear fuzzy ordering SfR such that |= R ⊂ e
=t SR and
for all a, b ∈ X also
(a, b) ∈ supp(R) ⇒ [[aRb]] = [[aSfR b]].
Proof. By proposition 2.30 under the present assumptions supp (R) is a

partial ordering (in the usual sense). By the classical Szpilrajn’s theo-
∗ which extends supp (R).
rem hence there exists an usual linear ordering SR
Define the fuzzy relation SR by
½ ∗y
1 iff xSR
[[(x, y) ε SR ]] =
0 otherwise.
Obviously SR is a fuzzy partial ordering which furthermore is st1 -linear for

each t-conorm st1 because of the usual linearity of SR∗ . By the way, the
∗
construction of SR guarantees SR = supp (SR ).
Of course, by the construction of SR we get
|= R ⊂
= SR
simply from supp (R) ⊆ SR ∗ = supp (S ). Yet, again by this construction,

R
all the ordered pairs (a, b) ∈ supp (R) which belong to R with a membership
degree different from 1 have a strictly greater membership degree in SR than
in R.
Assuming the existence of some u ∈ (0, 1] such that
(a, b) ∈ supp (R) ⇒ [[aRb]] ≥ u (2.37)
for all a, b ∈ X , we define an auxiliary fuzzy relation Ru by
[[(x, y) ε Ru ]] = max{[[(x, y) ε R]], u}
and put
SfR = SR ∩t Ru
(which gives the same fuzzy relation SfR for each t-norm t by construction
of SR ). Obviously again
|= R ⊂ f
= SR .
But now additionally
(a, b) ∈ supp (R) ⇒ [[aRb]] = [[aSfR b]];
and again SfR is a fuzzy partial ordering which now, yet, is weakly linear by
∗ . In order to prove the t-transitivity and t-antisymmetry
the linearity of SR
of SfR we essentially have to use (2.37) and the fact that
(a, b) ∈ supp (SfR ) \ supp (R) ⇒ [[aSfR b]] = u.
The details again are a routine matter. QED

2.5 Graded properties of fuzzy relations

The properties of fuzzy relations which were introduced in the previous def-
inition 2.10 are crisp in the sense that a fuzzy relation R either has one
of those properties or does not have that property. Therefore the situation
here is comparable with that we had before with the Zadeh’s crisp inclusion
(2.4) for fuzzy sets. And as we in that case were interested in generalizing
that inclusion ⊂ to a graded one by introducing the fuzzified inclusions ⊂ =t
in definition 2.3, we now again look for possibilities to generalize the rela-
tion properties of definition 2.10 in such a way that fuzzified, i.e. graded
properties result.
The fact that we always intended to write down our notions in the lan-
guage of (a suitable) many-valued logic now opens the door to get such
graded properties in a canonical manner out of those introduced in defini-
tion 2.10. The crucial point is that definition 2.10 used the logical validity
of suitable well-formed formulas to define the relation properties. Therefore
deleting this demand of logical validity and retaining the characteristic well-
formed formulas immediately opens a way to get graded properties of the
type we are looking for.
In the formal sense these graded properties themselves are to be repre-
sented by many-valued predicates of our language of many-valued logic.
Definition 2.18 For each fuzzy relation R and any t-norm t with property
LSC(t) let
Ref l(R) =def ∀x(xRx),
Irref lt (R) =def ∀x −t (xRx),
T ranst (R) =def ∀x∀y∀z(xRy ∧t yRz →t xRz),
Symmt (R) =def ∀x∀y(xRy →t yRx),
.
Antisymmt (R) =def ∀x∀y(xRy ∧t yRx →t x = y),
Asymmt (R) =def ∀x∀y −t (xRy ∧t yRx).
To get a simpler approach, in comparison with definition 2.10, we here

dispense with the use of a separate negation function n to define the pred-
icates Irref lt and Asymmt . On the other hand this change from crisp to
graded properties also causes a dependency on the involved t-norm for the
properties of irreflexivity and symmetry.
Immediately one has e.g. the result that
|= T ranst (R) ⇔ R t-transitive
2.5. GRADED PROPERTIES OF FUZZY RELATIONS 85
and the corresponding results for all the other graded properties with respect
to their crisp versions
As in the case of the crisp versions of relation properties, a characteriza-
tion of these graded properties with genuinely set theoretic notions is now
also possible.
Proposition 2.35 For each fuzzy relation R it holds true that

|= Ref l(R) ↔ ∆X ⊂ = R,
|= Irref lt (R) ↔ R ∩t ∆X ≡t ∅,
|= T ranst (R) ↔ R ◦t R ⊂
=t R,
|= Symmt (R) ↔ R ⊂ −1
=t R ,
|= Antisymmt (R) ↔ R ∩t R−1 ⊂ =t ∆X ,
−1
|= Asymmt (R) ↔ R ∩t R ≡t ∅.
Proof. Straightforward from the respective definitions, additionally using

proposition 2.5 (i) for the cases of the predicates Irref lt and Asymmt .
QED
In general, this “fuzzification” of properties of fuzzy relations is noth-

ing completely new. Before we essentially did the same with our definition
2.3 which introduced graded versions ⊂ =t and ≡t of inclusion and equality
for fuzzy sets. The small difference is only that there the generalization to
graded relations concerned binary relations, i.e. binary predicates for fuzzy
sets, and now we are concerned with generalized properties, i.e. unary pred-
icates for fuzzy relations.
As a relatively simple and very well-known example of a not completely
elementary special type of fuzzy relations, let us further on take a look at
fuzzy ordering relations which – for the crisp case – we already discussed in
the previous section. Now a graded notion of fuzzy partial ordering shall
be introduced. First we focus on the reflexive case.
Definition 2.19 Let R ∈ IF (X × X ) be a fuzzy relation and t any t-norm

with property LSC(t). Put
F P Ot (R) =def Ref l(R) ∧t T ranst (R) ∧t Antisymmt (R).
Obviously by LSC(t) and basic properties of t-norms one has

|= F P Ot (R) ⇔ R fuzzy partial ordering.
Corollary 2.36 Let R ∈ IF (X × X ) be a fuzzy relation and t any t-norm

with property LSC(t). Then
|= F P Ot (R) ↔ ∆X ⊂ ⊂ −1
=t R ∧t R ◦t R =t R ∧t R ∩t R ≡t ∅.
Proof. Obvious by proposition 2.35.
Let us first consider the usual duality between reflexive and irreflexive
fuzzy ordering relations established by changing from R to R+ = R ∪t ∆X
and from R to R− = R \t ∆X = {(x, y) k (x, y) ε R ∧t −t (x, y) ε ∆X }, i.e.
to R− = R ∩t CC∆X , cf. definition 2.15 and proposition 2.27.
Proposition 2.37 Let R ∈ IF (X × X ) be a fuzzy relation and t any t-norm

with the property LSC(t). Then there hold true
(i) |= Ref l(R ∪t ∆X ),
(ii) |= Irref lt (R \t ∆X ),
(iii) |= T ranst (R) →t T ranst (R ∪t ∆X ),
(iv) |= T ranst (R \t ∆X ) →t T ranst (R).
Proof. (i) and (ii) are obvious. For (iii) one has to prove the inequality
[[∀x, y, z(xRy ∧t yRz →t xRz)]] ≤
[[∀x, y, z(xR+ y ∧t yR+ z →t xR+ z)]]
for R+ = R ∪t ∆X . To get this we first show the inequality
[[aRb ∧t bRc →t aRc]] ≤ [[aR+ b ∧t bR+ c →t aR+ c]] (2.38)
for arbitrary a, b, c ∈ X .
In the case that a = b holds true it remains to be shown that
[[aRa ∧t aRc →t aRc]] ≤ [[aR+ a ∧t aR+ c →t aR+ c]].
But this obviously holds true because both these implications have truth
degree 1: in each of them the truth degree of the antecedent is not greater
than the truth degree of the succedent.
In the case that b = c holds true, essentially the same argument gives
(2.38).
Thus the case a 6= b 6= c remains to be considered. Now, however, one
has
[[aRb]] = [[aR+ b]], [[bRc]] = [[bR+ c]], [[aRc]] ≤ [[aR+ c]]
and thus again (2.38).

Therefore (2.38) generally holds true. But then obviously also
[[∀x, y, z(xRy ∧t yRz →t xRz)]] ≤ [[aR+ b ∧t bR+ c →t aR+ c]]
holds true. Here the left-hand side of the inequality is independent of the
choice of a, b, c. That means that even the infimum (with respect to a, b, c) of
the right-hand side is not smaller than the left-hand side, but this is exactly
(iii) according to definition 2.18.
Finally (iv) follows along the same line of argument as (iii). QED
Unfortunately the transitivity property cannot be transfered from R to

R \t ∆X in a manner directly analogous to (iii). However, that transfer
becomes possible with an additional assumption.

with the property LSC(t). Then one has
|= T ranst (R) ∧t Antisymmt (R) → T ranst (R \t ∆X ).
Proof. As in part (iii) of the last proof it is enough to show that for
R− = R \t ∆X = R ∩t CC∆X and all a, b, c ∈ X
[[T ranst (R) ∧t Antisymmt (R)]] ≤ [[aR− b ∧t bR− c →t aR− c]]. (2.39)
If c 6= a holds true, then [[aR− c]] = [[aRc]]. By |= R− ⊂

=t R hence
[[aRb ∧t bRc →t aRc]] ≤ [[aR− b ∧t bR− c →t aRc]]
and thus even
[[T ranst (R)]] ≤ [[aR− b ∧t bR− c →t aRc]]
holds true. So (2.39) is established in this case.

Now assume c = a. Then [[aR− c]] = 0 and (2.39) becomes
[[T ranst (R) ∧t Antisymmt (R)]] ≤ [[−t (aR− b ∧t bR− a)]]. (2.40)
In the case that a = b one has [[aR− b]] = 0 and thus [[−t (aR− b∧t bR− a)]] = 1,
.
which immediately gives (2.40). And in the case that a 6= b one has [[a =
b]] = 0 and thus
.
[[−t (aR− b ∧t bR− a)]] = [[aR− b ∧t bR− a →t a = b]],
hence
.
[[Antisymmt (R)]] ≤ [[aR− b ∧t bR− a →t a = b]]
and therefore (2.40). Thus (2.39) holds true in any case. QED
Similar transfer results hold true regarding antisymmetry and asymme-

try.

with property LSC(t). Then there hold true
(i) |= Antisymmt (R) → Asymmt (R \t ∆X ),

(ii) |= Asymmt (R \t ∆X ) ↔ Antisymmt (R \t ∆X ),
(iii) |= Asymmt (R) → Antisymmt (R ∪t ∆X ).
Proof. The same type of estimations as in the proofs of the last two propo-
sitions yield the results. QED
By and large, the basic links for a generalization of the usual duality of
reflexive and irreflexive partial orderings seem to be established with those
results. What remains to be given is a definition of graded irreflexive fuzzy
partial orderings. Caused by the process of grading this definition is not as
obvious as in the reflexive case. The background difficulty is the well-known
fact for classical irreflexive partial orderings that they can either be defined
by the requirements of irreflexivity and transitivity or – equivalently – by
those of irreflexivity, transitivity, and asymmetry.
In the present situation therefore we are first interested whether transi-
tivity together with irreflexivity also imply asymmetry in our more general
context. Indeed we have

with property LSC(t). Then it holds true that
|= T ranst (R) ∧t Irref lt (R) → Asymmt (R).
Proof. As in the foregoing proofs it is enough to prove for any a, b ∈ X
|= −t aRa ∧t ∀z(aRb ∧t bRz →t aRz) →t (aRb →t −t bRa). (2.41)

Using proposition 1.19 (i) twice and also proposition 1.27 (i) one finds that
(2.41) is equivalent to
|= −t aRa ∧t aRb ∧t (aRb →t ∀z(bRz →t aRz)) →t −t bRa.
But now using propositions 1.17 (i) and 1.23 (iv) together with the anti-
monotonicity of the Φ-operators in the first argument one has for the truth
degree of this expression
[[−t aRa ∧t aRb ∧t (aRb →t ∀z(bRz →t aRz)) →t −t bRa]]

≥ [[−t aRa ∧t ∀z(bRz →t aRz) →t −t bRa]]
= [[∀z(bRz →t aRz) →t (−t aRa →t −t bRa)]]
≥ [[∀z(bRz →t aRz) →t (bRa →t aRa)]]
= 1
and thus (2.41). QED
Therefore we decide to base irreflexive partial orderings only on transi-

tivity and irreflexivity.

with property LSC(t). Then we put
F P Ot∗ (R) =def T ranst (R) ∧t Irref lt (R).
With this definition all the notions and preliminary results are completed
which we need to state the generalized connection between the reflexive and
irreflexive graded fuzzy partial ordering relations.

with property LSC(t). Then there hold true
(i) |= T ranst (R) ∧t Antisymmt (R) → F P Ot∗ (R \t ∆X ),

(ii) |= F P Ot (R) → F P Ot∗ (R \t ∆X ),
(iii) |= F P Ot∗ (R) ∧t Asymmt (R) → F P Ot (R ∪t ∆X ).
Proof. Because of proposition 2.37 (ii) claim (i) is a direct consequence of

proposition 2.38. Of course, (ii) is a simple corollary of (i). And (iii) results
from Proposition 2.37 (i) together with proposition 2.37 (iii) and proposition
2.39 (iii). QED
The last claim of this proposition could be read as indicating that our
definition 2.20 was the wrong choice. But considering instead of F P Ot∗ the
many-valued predicate
F P Ot∗∗ (R) =def T ranst (R) ∧t Irref lt (R) ∧t Asymmt (R) (2.42)
one indeed would be able to prove
|= F P Ot∗∗ (R) → F P Ot (R ∪t ∆X ),
yet instead of proposition 2.41 (ii) then only
|= F P Ot (R) ∧t Antisymmt (R) → F P Ot∗∗ (R \t ∆X )
would be provable.
All together that means that neither definition 2.20 nor (2.42) gives the
simple results one has for crisp partial orderings. Only a restriction of the
t-norm under consideration allows for a simplification of proposition 2.41
(ii).
Corollary 2.42 For each fuzzy relation R ∈ IF (X × X ) and the t-norm

t = tG = min one has
|= F P Ot∗G (R) → F P OtG (R ∪ ∆X ).
Proof. In this special case one has
[[F P Ot∗G (R)]] = [[F P Ot∗G (R) ∧ F P Ot∗G (R)]]

≤ [[F P Ot∗G (R) ∧ AsymmtG (R)]]
and thus the result immediately follows from proposition 2.41 (ii). QED
Besides the ordering relations the equivalence relations are of fundamen-

tal importance in crisp mathematics. Like ordering relations we already
introduced fuzzy equivalence relations in the last section, cf. definition 2.11.
Here we intend to have a graded property of being an equivalence relation.

with property LSC(t). Then we define
Eqrelt (R) =def Ref l(R) ∧t Symmt (R) ∧t T ranst (R).

We do not intend to present here an extended theory of graded fuzzy

equivalence relations. That is to the same degree a future research require-
ment as a more elaborated theory of graded fuzzy orderings. But we shall
consider one topic, the fuzzification of the well-known fact that any two dis-
tinct equivalence classes of a (crisp) equivalence relation are disjoint sets.
There was a first type of fuzzification, for fuzzy relations which “really”, i.e.
to truth degree 1, are fuzzy equivalence relations, presented in proposition
2.20 (iv), a result that can equivalently be written as
|= haiR ∩t hbiR ≡t ∅ ↔ −t aRb
or as
|= haiR ∩t hbiR ≡
6 t ∅ ↔ aRb.
Further partial fuzzifications of this result read as follows.

with property LSC(t). Then there hold true for any a, b ∈ X
(i) |= T ranst (R) → (aRb →t haiR ⊂=t hbiR ),

(ii) ⊂
|= Ref l(R) → (haiR = hbiR →t aRb),
t
(iii) |= T ranst (R) ∧t Symmt (R) →
(haiR ∩t hbiR 6≡t ∅ →t aRb).
Proof. (i) By the corresponding definitions and propositions 1.19 (i) and
1.17 (i), one has the following inequalities for truth degrees:
[[T ranst (R) ∧t aRb]] = [[∀x, y, z(xRy ∧t yRz →t xRz) ∧t aRb]]

≤ [[∀x(aRb ∧t (xRa ∧t aRb →t xRb)]]
= [[∀x(aRb∧t )(aRbtot (xRa →t xRb)))]]
≤ [[∀x(xRa →t xRb))]]
= [[haiR ⊂
=t hbiR ]].
Hence (i) immediately follows again using proposition 1.19 (i).
(ii) With essentially the same logical background as in the foregoing proof
of claim (i) one gets (ii) by using proposition 1.19 (i) together with
[[Ref l(R) ∧t haiR ⊂

=t hbiR ]] ≤ [[aRa ∧t (aRa →t aRb)]]
≤ [[aRb]].
(iii) According to proposition 2.5 (ii) one has

[[haiR ∩t hbiR ≡
6 t ∅]] = [[∃x(x ε haiR ∩t hbiR )]]
= [[∃x(xRa ∧t xRb)]].
Therefore with propositions 1.28 (i) and 1.17 (i) one gets as an estimation
of truth degrees:
[[T ranst (R) ∧t Symmt (R) ∧t ∃x(xRa ∧t xRb)]]
= [[∃x(T ranst (R) ∧t Symmt (R) ∧t xRa ∧t xRb)]]
≤ [[∃x(T ranst (R) ∧t (xRa →t aRx) ∧t xRa ∧t xRb)]]
≤ [[∃x(T ranst (R) ∧t aRx ∧t xRb)]]
≤ [[∃x((aRx ∧t xRb →t aRb) ∧t aRx ∧t xRb)]]
≤ [[∃x(aRb)]] = [[aRb]].
Now once again the application of proposition 1.19 (i) gives the result. QED
Corollary 2.44 Under the same conditions as in proposition 2.42 one has
(i) |= Ref l(R) ∧ T ranst (R) → (haiR ⊂
=t hbiR ↔t aRb),
(ii) |= T ranst (R) ∧t Symmt (R) ∧t −t aRb →
haiR ∩t hbiR ≡t ∅,
2
(iii) |= [T ranst (R)] → (aRb ∧t bRa →t haiR ≡t hbiR ),
(iv) |= T ranst (R) →
(aRb ∧ bRa →t haiR ⊂ ⊂
=t hbiR ∧ hbiR =t haiR ).
Proof. (i) results via proposition 1.19 (iv) from proposition 2.43 (i) and (ii)
using the fact that [[H1 ↔t H2 ]] = [[(H1 →t H2 ) ∧ (H2 →t H1 )]] always holds
true.
(ii) We start from the succedent of the implication in the previous propo-
sition 2.43 (iii) and have by propositions 1.22 (i) and 1.19 (ii)
|= (haiR ∩t hbiR ≡
6 t ∅ →t aRb)
→ (−t (haiR ∩t hbiR ≡t ∅) →t −t −t aRb)
and by proposition 1.23 (iii) also
|= (−t (haiR ∩t hbiR ≡t ∅) →t −t −t aRb)
→ (−t aRb →t −t −t (haiR ∩t hbiR ≡t ∅)).
Therefore from propositions 2.5 (ii) and 2.43 (iii) by an iterated application
of (the consequence rule corresponding to) proposition 1.19 (ii) one finds
|= T ranst (R) ∧t Symmt (R) → −t −t −t ∃x(x ε haiR ∩t hbiR ).
With proposition 1.22 (ii) that means
|= T ranst (R) ∧t Symmt (R) → −t ∃x(x ε haiR ∩t hbiR )
which gives the final
|= T ranst (R) ∧t Symmt (R) → haiR ∩t hbiR ≡t ∅
by again applying proposition 2.5 (ii).

(iii) Let H(a, b) be the formula aRb →t haiR ⊂ =t hbiR . Taking the
formula T ranst (R) → H(a, b), whose logical validity is stated in proposition
2.43 (i), together with the formula T ranst (R) → H(b, a) which results
from the earlier one by interchanging a and b, that gives via proposition 1.19
(iii)
|= [T ranst (R)]2 → (H(a, b) ∧t H(b, a)). (2.43)
Now both formulas H(a, b) and H(b, a) themselves are implications. Thus
again applying proposition 1.19 (iii) to the succedent of (2.43) finally gives
our claim (iii) according to the definition of ≡t .
(iv) follows by the same arguments, but refering to proposition 1.19 (iv)
instead of 1.19 (iii) and thus having
|= T ranst (R) → (H(a, b) ∧ H(b, a))
instead of (2.43). QED
For the t-norm t = tG = min even a little more becomes provable.
Corollary 2.45 For each fuzzy relation R ∈ IF (X × X ) and all a, b ∈ X

there hold true
(i) |= T ranstG (R) → (aRb →tG haiR ≡tG hbiR )),

(ii) |= T ranstG (R) ∧ SymmtG →
(haiR ∩ hbiR ≡tG haiR ∩ {x k aRb}).
Proof. (i) results immediately from corollary 2.44 (ii) by the idempotency
of the t-norm tG .
(ii) With regard to proposition 2.43 (iii) we have the following sequence
of transformations of the truth degree of the respective succedent:
[[haiR ∩ hbiR ≡
6 tG ∅ →tG aRb]]
= [[∃x(x ε haiR ∩ hbiR ) →tG aRb]]
= [[∀x(x ε haiR ∧ hbiR →tG aRb)]]
= [[∀x(x ε haiR ∧ hbiR →tG aRb ∧ x ε haiR )]]
whose last transformation rests on simple properties of the min-operator tG

and its corresponding Φ-operator ϕG .
Using the fact that for each x
[[aRb]] = [[x ε {y k aRb}]] (2.44)
and changing the bound variable y of the generalized class term {y k aRb}
into x gives
6 tG ∅ →tG aRb]] = [[haiR ∩ hbiR ⊂

[[haiR ∩ hbiR ≡ =tG haiR ∩ {x k aRb}]]
and thus all together
|= T ranstG (R) ∧ SymmtG →

(haiR ∩ hbiR ⊂
=tG haiR ∩ {x k aRb}). (2.45)
On the other hand starting from proposition 2.43 (i) we have the following
sequence of transformations of the truth degree of the respective succedent:
[[aRb →tG haiR ⊂

=tG hbiR ]]
= [[aRb →tG ∀x(x ε haiR →tG x ε hbiR )]]
= [[∀x(aRb ∧ x ε haiR →tG x ε hbiR )]]
= [[∀x(aRb ∧ x ε haiR →tG x ε haiR ∧ x ε hbiR )]]
and thus, again using (2.44),
|= T ranstG (R) → (haiR ∩ {x k aRb} ⊂

=tG haiR ∩ hbiR ). (2.46)
Combining (2.45) and (2.46) via proposition 1.19 (iv) then gives the result.
QED
Having for a fuzzy relation treated their property to be a fuzzy partial

ordering relation, as well as their property to be a fuzzy equivalence relation
both as graded predicates, it remains, with respect to the topics discussed in
the previous section, to take a look at the fuzzified transitive hull of a fuzzy
relation and to see which results on those t-transitive hulls were not given
in a “truly fuzzified” form there. This obviously concerns proposition 2.23
and proposition 2.22 (ii).
Therefore let us try to generalize the last mentioned result too. We
immediately have
|= R ≡t Tt (R) ↔ Tt (R) ⊂ =t R
because of proposition 2.22 (i). Furthermore we have already proved the
results
|= Ref l(R) → ∆X ⊂ =t R,
|= > → R ⊂ =t R,
|= T ranst (R) → R2 ⊂ =t R
the second of which, however, was previously given as |= R ⊂ =t R. Thus we
have according to proposition 1.19 (iv)
|= Ref l(R) ∧ > ∧ T ranst (R) → (∆X ⊂ ⊂ 2 ⊂
=t R ∧ R =t R ∧ R =t R)
and hence because of proposition 2.3 (i)
2
[
|= Ref l(R) ∧ T ranst (R) → Ri ⊂
=t R. (2.47)
i=0
From R3 = R2 ◦t R according to definition 2.13 together with proposition

2.15 (i), i.e. the monotonicity of ◦t , we find
|= T ranst (R) → R3 ⊂ =t R
2
and thus
|= [T ranst (R)]2 → R3 ⊂ =t R. (2.48)
Unfortunately this antecedent [T ranst (R)]2 cannot be simplified in such a

way that (2.48) holds true with T ranst (R) instead of [T ranst (R)]2 . Thus
we only get from (2.47) and (2.48)
3
[
|= Ref l(R) ∧ [T ranst (R)]2 → Ri ⊂
=t R. (2.49)
i=0
using proposition 1.19 (iv) and the fact that
[[T ranst (R) ∧ [T ranst (R)]2 ]] = [[[T ranst (R)]2 ]].
Of course, this transition from (2.47) to (2.49) can be iterated using the
fact that (2.48) holds true more generally in the form
|= [T ranst (R)]n → Rn+1 ⊂

=t R.
Therefore one has for each integer n > 1
n+1
[
|= Ref l(R) ∧ [T ranst (R)]n → Ri ⊂
=t R. (2.50)
i=0
To finally find the result for Tt (R) we are looking for, we use the symbolic
∞
Q
expression [T ranst (R)]n with the truth degree
i=0
∞
Y
[[ [T ranst (R)]n ]] =def lim [[[T ranst (R)]n ]]
n→∞
i=0
and get out of (2.50) according to proposition 2.7 (iv) that

∞
Y
|= Ref l(R) ∧ [T ranst (R)]n → Tt (R) ⊂
=t R
i=0
and thus
∞
Y
|= Ref l(R) ∧ [T ranst (R)]n → Tt (R) ≡t R.
i=0
∞
Q
Unfortunately, caused by the fact that the expression [T ranst (R)]n
i=0
appears as part of the antecedent, this result seems to be quite week. Only
in the special case t = tG does this infinite iteration of the t-conjunction
disappear and give
|= Ref l(R) ∧ T ranstG (R) → TtG (R) ≡tG R.

Chapter 3
Set equations with fuzzy sets
3.1 Fuzzy equations and some of their applications

The notion of the fuzzy equation itself has not so far been defined mathemat-
ically. Essentially, in the present literature, three types of fuzzy equations
have been discussed. The starting point was in Sanchez (1974, 1976) where
the author considered fuzzy relation(al) equations of the type
R ◦t X = S (3.1)
with t = tG = min or, more generally, any intersection operation of a com-

plete Brouwerian lattice and R, S given fuzzy relations. The problem was
to find the greatest solution X; the motivation of the problem originated
from medical applications of fuzziness; cf. Sanchez (1974, 1977) and also
Kaufmann (1977).
A second type of fuzzy equation, also sometimes called fuzzy relation(al)
equations and by abuse of notation written R ◦t A = B or also A ◦t R = B,
is equations of the form
R00 A = B. (3.2)
Again, Sanchez (1978) was the first one who discussed equations of this
type, but in the special form
R00 A = A
which he called “eigen fuzzy sets equations”. The fuzzy relation R was
supposed to be given, and an eigen fuzzy set A as a solution had to be
determined. More precisely, Sanchez (1978) looked for the greatest solution
98 CHAPTER 3. SET EQUATIONS WITH FUZZY SETS
of R00 A = A because, of course, A = ∅ is a trivial solution for each R. His

motivation, again, came from medical applications.
Fuzzy equations of the kind (3.2) have since found a much wider field
of application. In automatic control, for example, the fuzzy controllers first
considered by Mamdani (1974, 1976) turned out to be very useful. In the
simplest case such a fuzzy controller has one linguistic input variable and
one linguistic output variable. The values of these linguistic variables are
fuzzy subsets of corresponding input and output spaces. And the fuzzy
controller transforms an input fuzzy set, i.e. some given special value of the
input linguistic variable, into an output fuzzy set. Hence, a fuzzy controller
may be viewed as a fuzzy relation R connecting input A with output B by
a relationship of the form (3.2).
Usually, such a fuzzy controller is supposed to be constructed from a
finite list of control rules
Ai |=
⇒ Bi , i = 1, . . . , n (3.3)
connecting values Ai of the input variable with values Bi of the output

variable. Hence, the fuzzy controller to be constructed is intended to be
described by (and indeed may be identified with) a solution R to the system
R00 Ai = Bi , i = 1, . . . , n (3.4)
of fuzzy equations.
Originally, Mamdani used for R the fuzzy union R̂ of the fuzzy carte-
sian products Ai × Bi . But this fuzzy relation R̂ need not be a solution
of system (3.4) of fuzzy equations as was mentioned for example by Czo-
gala/Pedrycz (1981); and conditions which guarantee that this fuzzy rela-
tion R is a solution of (3.4) seem to be quite strong, cf. chapter 4. Therefore,
the preferable way to construct a fuzzy controller R realizing the system of
control rules (3.3) is to choose R as a solution to the system (3.4) of set
equations – or at least as some kind of approximate solution to it.
Besides these applications to fuzzy controllers, fuzzy equations of type
(3.2) have been discussed in connection with for example the identification
of fuzzy systems, prediction in fuzzy systems, and in fuzzy decision-making,
cf. for example Pedrycz (1983, 1983a).
All these applicational ideas caused a growing interest in the solution
of fuzzy equations of type (3.2), in algorithms to construct such solutions,
in structural properties of the set of all solutions, and in the solution of
systems like (3.4); cf. for example Czogala/Drewniak/Pedrycz (1982),
3.1. FUZZY EQUATIONS AND SOME OF THEIR APPLICATIONS 99
diNola/Sessa (1983), Pedrycz (1983a), Pedrycz/Czogala/Hirota

(1984), Gottwald (1984b), diNola/Pedrycz/Sessa (1985), and the
book diNola/ Sessa/Pedrycz/Sanchez (1989).
A third type of fuzzy equation is of interest in the field of fuzzy arith-
metic, i.e. for the consideration of arithmetical operations for fuzzy numbers.
These operations for fuzzy numbers, i.e. for fuzzy subsets of the real line, are
defined via the extension principle from the usual arithmetical operations.
They are, of course, generalizations of the operations of interval arithmetic,
and as in interval arithmetic, the extended subtraction is no longer the in-
verse operation of extended addition, i.e. for fuzzy numbers A, B the fuzzy
equation
A +t X = B (3.5)
in general does not have the solution X = B −t A. Therefore the problem

arises to characterize those equations (3.5) which have a solution – e.g. by
conditions for A, B – and to determine the set of solutions.
For the existence of direct – additive or multiplicative – inverses this
problem was discussed by Mizumoto/Tanaka (1979) yielding very strong
conditions: only such fuzzy numbers have inverses which are crisp single-
tons. Therefore, Yager (1980a) discussed approximate solutions to fuzzy
equations and evaluated the quality of approximation with linguistic truth
values. Later on Sanchez (1984) proved an interesting solvability criterion.
And it is this criterion that will be generalized here and also extended to
types (3.1), (3.2) of fuzzy equations in such a way that it does not only
characterize the fuzzy equations which have solutions – but also in the other
cases gives an upper bound measuring the quality of the best possible ap-
proximate solutions. Yet, we will measure the quality of an approximate
solution not with a linguistic truth value, i.e. a fuzzy subset of [0,1], as was
done by Yager (1980a), but our index of solvability will be a generalized
truth value, i.e. a real number of the interval [0,1].
In the following we will refer to fuzzy equations of types (3.1), (3.2) as
fuzzy relational equations because of the involvement of fuzzy relations, and
to fuzzy equations of type
A ∗t X = B
as fuzzy arithmetical equations because of their (possible) relationship to

fuzzy numbers.
3.2 Solvability of fuzzy relational equations

We have first to introduce an additional operation for fuzzy relations that,
in a suitable sense, is dual to the relational product R ◦t S of definition 2.8.
Definition 3.1 For all fuzzy relations R, S ∈ IF (X × X ) and each t-norm

t with property LSC(t) let
R ¦t S =def {(x, y) k ∀z((x, z) ε R →t (z, y) ε S)}.
Additionally we need a new operation, besides the fuzzy cartesian prod-
uct, that gives a fuzzy relation in IF (X ×Y) out of two fuzzy sets from IF (X )
and IF (Y).
Definition 3.2 For all fuzzy sets A ∈ IF (X ) and B ∈ IF (Y) and each t-
norm t with property LSC(t) let
A ¤t B =def {(x, y) k x ε A →t y ε B}.
Proposition 3.1 For all fuzzy sets A, B and fuzzy relations R, S and each
t-norm t with property LSC(t) there hold true
(i) |= R00 (R ↓ B) ⊂
=t B,
00 ⊂
(ii) |= (A ¤t B) A = B,
t
(iii) |= R ◦t (R ¦t S) ⊂=t S,
(iv) |= (R ¦t S −1 )−1 ◦t R ⊂
=t S.
Proof. (i) For every point a ∈ X of our universe of discourse we have
[[a ε R00 (R ↓ B)]] = [[∃x(x ε R ↓ B ∧t (x, a) ε R)]]
= [[∃x((x, a) ε R ∧t ∀z((x, z) ε R →t z ε B))]]
≤ [[∃x((x, a) ε R ∧t ((x, a) ε R →t a ε B))]]
≤ [[∃x(a ε B)]] = [[a ε B]]
by definition (1.1) of many-valued generalization, the monotonicity (T2) of
every t-norm, and proposition 1.17 (i). Our result follows by the rule of
generalization and proposition 1.13.
All the other facts will be proved along the same lines of argument: to get
the analogous inequalities use corresponding definitions, delete many-valued
generalization, and reduce à la proposition 1.17 (i). Thus we can omit the
details. QED
3.2. SOLVABILITY OF FUZZY RELATIONAL EQUATIONS 101
Proposition 3.2 For all fuzzy sets A, B and fuzzy relations R, S, T and
each t-norm t with property LSC(t) there hold true
(i) |= R00 A ⊂ ⊂
=t B →t A =t R ↓ B,
(ii) |= R00 A ⊂ ⊂
=t B →t R =t A ¤t B,
(iii) |= R ◦t S ⊂ ⊂ −1
=t T →t S =t R ¦t T,
(iv) |= R ◦t S ⊂ ⊂ −1 −1
=t T →t R =t (S ¦t T ) .
Proof. (i) The following estimation proves the result:
[[A ⊂
=t R ↓ B]] = [[∀x(x ε A →t x ε R ↓ B)]]
= [[∀x∀y(x ε A →t ((x, y) ε R →t y ε B))]]
≥ [[∀x∀y(x ε A ∧t (x, y) ε R →t y ε B)]]
≥ [[∀y(∃x(x ε A ∧t (x, y) ε R) →t y ε B)]]
= [[∀y(y ε R00 A →t y ε B)]]
= [[R00 A ⊂
=t B]].
Here, besides the corresponding definitions, the propositions 1.27 (i), 1.19
(i), and 1.27 (ii) have been successively used .
Just the same line of argument is also successful for parts (ii) to (iv).
Thus again the details can be omitted. QED
Corollary 3.3 For all fuzzy sets A, B and fuzzy relations R, S, T and each
t-norm t with the property LSC(t) the following facts hold true:
(i) Every solution of equation R00 X ≡t B is a fuzzy subset of R ↓ B.
(ii) Every solution of equation X 00 A ≡t B is a fuzzy subset of A ¦t B.
(iii) Every solution of equation R ◦t X ≡t S is a fuzzy subset of R−1 ¦t S.
(iv) Every solution of equation X◦t R ≡t S is a fuzzy subset of (R¦t S −1 )−1 .
Proof. Suppose A to be a solution of the equation R00 X ≡t B. Then

of course [[R00 A ≡t B]] = 1 and therefore also [[R00 A ⊂=t B]] = 1, i.e. by
⊂
proposition 3.2 (i): [[A = R ↓ B]] = 1. Thus (i) is proved.
t
The other claims follow in the same manner. QED
As a side remark we have to note that we have misused the term “fuzzy
subset” here a little, otherwise in this book a fuzzy subset of some (crisp!)
set X will be taken to mean any function from X into [0,1]. But here by a
fuzzy subset A of a fuzzy (!) set B there is meant such a fuzzy set A for
which |= A ⊂=t B holds true. This difference will always become clear from
the context, and we hope not to cause any misunderstanding.
Now we are able to state and prove the main results for (single) fuzzy
relational equations.
Theorem 3.4 Suppose that t is a t-norm with property LSC(t). Then for
all fuzzy sets A, B and fuzzy relations R, S there hold true
(i) |= ∃X(R00 X ≡t B) ↔t R00 (R ↓ B) ≡t B,
(ii) |= ∃X(X 00 A ≡t B) ↔t (A ¤t B)00 A ≡t B,
(iii) |= ∃X(R ◦t X ≡t S) ↔t R ◦t (R−1 ¦t S) ≡t S,
(iv) |= ∃X(X ◦t R ≡t S) ↔t (R ¦t S −1 )−1 ◦t R ≡t S.
Proof. (i) By proposition 1.13 we have to prove the equality

[[∃X(R00 X ≡t B)]] = [[R00 (R ↓ B) ≡t B]]
for these truth degrees. Obviously, from the definition (1.2) of the existential
quantifier and the fact that R ↓ B ∈ IF (X ), we immediately have
[[∃X(R00 X ≡t B)]] ≥ [[R00 (R ↓ B) ≡t B]]
Thus the reverse inequality is what really has to be proved. But, from
proposition 3.1 (i) and definition 2.3 one has
[[R00 (R ↓ B) ≡t B]] = [[B ⊂ 00
=t R (R ↓ B)]]
and hence only to prove
[[∃X(R00 X ≡t B)]] ≥ [[B ⊂ 00
=t R (R ↓ B)]].
For this it is enough to show for each fuzzy set A
[[R00 A ≡t B]] ≥ [[B ⊂ 00
=t R (R ↓ B)]].
Using definition 2.3 and proposition 3.2 (ii) one has
[[R00 A ≡t B]] = [[R00 A ⊂ ⊂ 00
=t B ∧t B =t R A]]
≤ [[A ⊂ ⊂ 00
=t R ↓ B ∧t B =t R A]]
3.3. SOLVABILITY OF FUZZY ARITHMETICAL EQUATIONS 103
and furthermore by the monotonicity of the full image and the transitivity
of the fuzzy inclusion ⊂=t , i.e. by propositions 2.18 (i) and 2.2 (iii) one is
able to continue this estimation as
[[R00 A ≡t B]] ≤ [[R00 A ⊂ 00 ⊂ 00
=t R (R ↓ B) ∧t B =t R A]]
≤ [[B ⊂ 00
=t R (R ↓ B)]].
Thus (i) is proved.
The other results formulated in this theorem can be proved in the same
way. All the facts necessary for carrying out these proofs have been formu-
lated in propositions 2.2, 2.18, 2.15, 3.1, and 3.2. Hence, the rest of this
proof is now again routine. QED
Corollary 3.5 For all fuzzy sets A, B and fuzzy relations R, S, T and each
t-norm t with the property LSC(t) the following facts hold true:
(i) Equation R00 X = B has a solution iff R00 (R ↓ B) = B.
(ii) Equation X 00 A = B has a solution iff (A ¦t B)00 A = B.
(iii) Equation R ◦t X = S has a solution iff R ◦t (R−1 ¦t S) = S.
(iv) Equation X ◦t R = S has a solution iff (R ¦t S −1 )−1 ◦t R = S.
Proof. (i) If R00 (R ↓ B) = B then R ↓ B is a solution to equation R00 X = B;

and then it is the greatest one by corollary 3.3 (i). If otherwise equation
R”X = B has a solution, say A, then R00 A = B, i.e. [[R00 A ≡t B]] = 1 and
hence [[∃X(R00 X ≡t B)]] = 1, thus R00 (R ↓ B) ≡t B by proposition 3.4 (i)
and this means R00 (R ↓ B) = B because of (2.14).
(ii) to (iv) follow by corresponding arguments. QED
3.3 Solvability of fuzzy arithmetical equations

Now we suppose that in the universe of discourse X a binary operation ∗ is
given. This operation is extended to a binary operation, denoted ∗t for a
given t-norm t, in the class of all fuzzy subsets of X by the extension princi-
ple. Additionally, a companion operation ˜ ∗t is needed which in the present
case is analogous to the earlier operations ¤t , ¦t , as well as ↓ introduced in
definitions 3.2, 3.1, and 2.9. (For the special case that t = tG = min this
operation was introduced by Sanchez (1984).)
Definition 3.3 For all fuzzy sets A, B and each t-norm t with the property
LSC(t) put
B ˜∗t A =def {y k ∀x(x ε A →t x ∗ y ε B)}.
Proposition 3.6 For all fuzzy sets A, B, C and each t-norm t with the prop-
erty LSC(t) there hold true
(i) |= B ⊂ ⊂
=t C →t A ∗t B =t A ∗t C,
(ii) |= A ∗t B ⊂ ⊂
=t C →t B =t C ˜
∗t A,
(iii) ∗t A) ⊂
|= A ∗t (B ˜ =t B.
Proof. (i) To get the result we prove for the corresponding truth degrees
the inequality
[[B ⊂ ⊂
=t C]] ≤ [[A ∗t B =t A ∗t C]].
Immediately from the extension principle together with (2.24) one has
[[A ∗t B ⊂
=t A ∗t C]] = [[∀x(x ε A ∗t B →t x ε A ∗t C)]]
.
= [[∀x(∃y∃z(y ε A ∧t z ε B ∧t x = y ∗ z) →t x ε A ∗t C)]]
.
= [[∀x∀y∀z(y ε A ∧t z ε B ∧t x = y ∗ z →t x ε A ∗t C)]].
With the same trick as in the proof of proposition 2.18 (i) for eliminating the
existential quantifier that comes into this formula by substituting x ε A ∗t C
.
through ∃u∃v(u ε A ∧t v ε C ∧t x = u ∗ v) we get furthermore
[[A ∗t B ⊂=t A ∗t C]]

≥ [[∀x∀y∀z(z ε B ∧t Θ(x, y, z) →t z ε C ∧t Θ(x, y, z)]]
.
with formula Θ(x, y, z) : y ε A ∧t x = y ∗ z. Now proposition 1.18 (i) and the
dropping of the empty quantifiers allows us to continue the estimation
[[A ∗t B ⊂ ⊂
=t A ∗t C]] ≥ [[∀z(z ε B →t z ε C]] = [[B =t C]]
and thus to finish the proof of (i).
(ii) The corresponding inequality for truth values will now also be proved.
It is
[[B ⊂
=t C ˜∗t A]] = [[∀x(x ε B →t ∀y(y ε A →t y ∗ x ε C))]]
= [[∀x∀y(x ε B ∧t y ε A →t y ∗ x ε C)]]
3.3. SOLVABILITY OF FUZZY ARITHMETICAL EQUATIONS 105
which because of [[x ε B ∧t y ε A]] ≤ [[y ∗ x ε A ∗t B]] and the antimonotonicity

of ϕt in the first argument can be continued as
[[B ⊂
=t C ˜∗t A]] ≥ [[∀x∀y(y ∗ x ε A ∗t B →t y ∗ x ε C)]]
≥ [[∀z(z ε A ∗t B →t z ε C)]]
= [[A ∗t B ⊂ =t C]].
Finally, (iii) is proved by simply applying the corresponding definitions
together with proposition 1.17 (i). QED
Theorem 3.7 For each t-norm t with the property LSC(t) and all fuzzy sets
A, B there holds true
|= ∃X(A ∗t X ≡t B) ↔t A ∗t (B ˜
∗t A) ≡t B.
Proof. Obviously
[[A ∗t (B ˜∗t A) ≡t B]] ≤ [[∃X(A ∗t X ≡t B)]]
and by proposition 3.6 (iii) also
[[A ∗t (B ˜∗t A) ≡t B]] = [[B ⊂
=t A ∗t (B ˜ ∗t A)]].
Therefore, it remains to be proved
[[∃X(A ∗t X ≡t B)]] ≤ [[B ⊂ =t A ∗t (B ˜∗t A)]],
i.e. for every fuzzy set C that
[[A ∗t C ≡t B]] ≤ [[B ⊂=t A ∗t (B ˜∗t A)]].
And indeed
[[A ∗t C ≡t B]] = [[A ∗t C ⊂ ⊂
=t B ∧t B =t A ∗t C]]
≤ [[C ⊂ =t B ˜∗t A ∧t B ⊂ =t A ∗t C]]
⊂
≤ [[A ∗t C = A ∗t (B ˜ ∗t A) ∧t B ⊂
t =t A ∗t C]]
≤ [[B ⊂ =t A ∗t (B ˜∗t A)]]
by the previous results. QED
For completeness we should, for non-commutative operations ∗, also con-

sider equations of the type
X ∗t A ≡t B
but by considering in X the operation a]b = b ∗ a we may reduce this case
to that which is considered in theorem 3.7.
Proposition 3.8 Suppose that the t-norm t has the property LSC(t). Then
equation A ∗t X = B has a solution iff A ∗t (B ˜∗t A) = B; and if equation
A ∗t X = B has a solution, then B ˜
∗t A is the greatest one.
Proof. By the same argument that was used in the proofs of corollaries 3.5
and 3.3. QED
The result of this corollary 3.8 was, for the t-norm t = min, first proved
by Sanchez (1984) and the starting point for the present generalizations.
3.4 Solvability of systems of fuzzy equations

So far we have only discussed the solvability of single fuzzy equations, de-
spite the fact that systems of fuzzy relational equations like (3.4) are essential
means for the construction of fuzzy controllers. Moreover, the research liter-
ature till now has not paid much attention to this topic – Gottwald (1984b)
seems to be an exception. Here there are given for systems of fuzzy equa-
tions, of each one of the types considered before, inequalities for the truth
degrees of the formula of our language of many-valued logic that states the
existence of a solution to the system.
Unfortunately, we have not found equations characterizing these truth
degrees as in theorems 3.4 and 3.7, but only inequalities. Nevertheless we
find lower bounds – and in a certain weak sense also upper bounds.
Theorem 3.9 Suppose that the t-norm t has the property LSC(t). Consider
finite sequences (Ai )1≤i≤n and (Bi )1≤i≤n of fuzzy sets as well as (Ri )1≤i≤n
and (Si )1≤i≤n of fuzzy relations. Then there hold true
(i) in the case of a system of fuzzy equations with given fuzzy relations
and given full images that has to be solved for an argument fuzzy set:
n
Y
[[(∃X (Ri00 X ≡t Bi ))n ]]
i=1
n
Y n
\
≤ [[ (Ri00 ( (Rj ↓ Bj )) ≡t Bi )]]
i=1 j=1
n
Y
≤ [[∃X (Ri00 X ≡t Bi )]]
i=1
3.4. SOLVABILITY OF SYSTEMS OF FUZZY EQUATIONS 107
(ii) in the case of a system of fuzzy equations with given input fuzzy sets
and given full images that has to be solved for a corresponding fuzzy
relation:
n
Y
[[(∃X (X 00 Ai ≡t Bi ))n ]]
i=1
n
Y n
\
≤ [[ (( (Aj ¤t Bj ))00 Ai ≡t Bi )]]
i=1 j=1
n
Y
≤ [[∃X (X 00 Ai ≡t Bi )]]
i=1
(iii) in the case of a system of fuzzy equations with given relational products
and given first factors that has to be solved for the second factor:
n
Y
[[(∃X (Ri ◦t X ≡t Si ))n ]]
i=1
n
Y n
\
≤ [[ (Ri ◦t (Rj −1 ¦t Sj ) ≡t Si )]]
i=1 j=1
Yn
≤ [[∃X (Ri ◦t X ≡t Si )]]
i=1
(iv) in the case of a system of fuzzy equations with given relational products
and given second factors that has to be solved for the first factor:
n
Y
[[(∃X (X ◦t Ri ≡t Si ))n ]]
i=1
n
Y n
\
≤ [[ (( (Rj ¦t Sj −1 )−1 ) ◦t Ri ≡t Si )]]
i=1 j=1
n
Y
≤ [[∃X (X ◦t Ri ≡t Si )]]
i=1
(v) in the case of a system of “linear” fuzzy arithmetical equations with

given first operands and given results that has to be solved for the second
operand:
n
Y
[[(∃X (Ai ∗t X ≡t Bi ))n ]]
i=1
n
Y n
\
≤ [[ (Ai ∗t ( (Bj ˜
∗t Aj )) ≡t Bi )]]
i=1 j=1
n
Y
≤ [[∃X (Ai ∗t X ≡t Bi )]]
i=1
Proof. As before the proofs run parallel for each one of these cases. We
therefore give only the derivation for case (ii) because of its close connection
with fuzzy controllers.
Because of proposition 1.31 one has
n
Y
[[(∃X (X 00 Ai ≡t Bi ))n ]]
i=1
n
Y
= sup{ [[( (T 00 Ai ≡t Bi ))n ]] | T ∈ IF (X × X )},
i=1
hence for the first inequality stated in (ii) it has only to be proved
n
Y n
Y n
\
[[( (T 00 Ai ≡t Bi ))n ]] ≤ [[ (( (Aj ¤t Bj ))00 Ai ≡t Bi )]] (3.6)
i=1 i=1 j=1
for any T ∈ IF (X × X ). For the sake of brevity let be

n
\
D= (Aj ¤t Bj ).
j=1
This gives the following sequence of inequalities:

n
Y
[[( (T 00 Ai ≡t Bi ))n ]]
i=1
n
Y
= [[ (T 00 Ai ⊂ ⊂ 00
=t Bi ∧t Bi =t T Ai )]]
i,j=1
Y n n
Y
= [[ (T 00 Ai ⊂
=t Bi ) ∧t (Bi ⊂ 00
=t T Ai )]]
i,j=1 i,j=1
n
Y n
Y
≤ [[ (T ⊂
=t Ai ¤t Bi ) ∧t (Bi ⊂ 00
=t T Ai )]]
i,j=1 i,j=1
Y n ^ n n
Y
≤ [[ ( (T ⊂
=t Ai ¤t Bi )) ∧t (Bi ⊂ 00
=t T Ai )]]
j=1 i=1 i,j=1
because of proposition 3.2 (ii) and the fact that xty ≤ min(x, y) for every
t-norm t and all x, y ∈ [0, 1], which because of proposition 2.7 (iii) gives
furthermore
n
Y n
Y
≤ [[ (T ⊂
=t D) ∧t (Bi ⊂ 00
=t T Ai )]]
j=1 i,j=1
Yn
≤ [[ (Bi ⊂ 00 ⊂
=t T Ai ∧t T =t tD)]]
i=1
Yn
≤ [[ (Bi ⊂ 00 00 ⊂ 00
=t T Ai ∧t T Ai =t tD Ai )]]
i=1
Yn
≤ [[ (Bi ⊂ 00
=t D Ai )]],
i=1
where now successively xty ≤ x for each t-norm t and all x, y ∈ [0, 1],
proposition 2.18 (ii), and the generalized transitivity of the fuzzy inclusion
⊂ , i.e. proposition 2.2 (iii) have been used. From proposition 2.7 (i) we
=t
get for every 1 ≤ i ≤ n
|= D ⊂
=t Ai ¤t Bi
and therefore by propositions 2.18 (ii), 3.1 (ii), and again the generalized
transitivity of the fuzzy inclusion ⊂
=t immediately
|= D00 Ai ⊂
=t Bi .
This gives
n
Y
[[ ⊂ Bi )]] = 1
(D00 Ai =
t
i=1
and hence
n
Y n
Y
[[ (Bi ⊂ 00
=t D Ai )]] = [[ (D00 Ai ≡t Bi )]].
i=1 i=1
So, finally, inequality (3.6) is proved. And for the full proof of (ii) only
n
Y n
Y
00
[[ (D Ai ≡t Bi )]] ≤ [[∃X (X 00 Ai ≡t Bi )]]
i=1 i=1
remains. But this is obvious. Thus, the proof of (ii) is complete. QED
Because of their relationship with fuzzy controllers and with fuzzy arith-
metic, in the following corollaries we discuss only systems of fuzzy equations
of the type
X 00 Ai ≡t Bi , 1≤i≤n (3.7)
and of the type
Ai ∗ X ≡t Bi , 1 ≤ i ≤ n. (3.8)
Corollary 3.10 Consider only the t-norm t = tG = min and omit all in-
dices that refer to this t-norm. Then we have for the systems (3.7), (3.8) of
fuzzy equations the solvability criteria:
(i) in the case of a system of fuzzy equations with given input fuzzy sets
and given full images that has to be solved for a corresponding fuzzy
relation:
n
^ n
^ n
\
|= ∃X (X 00 Ai ≡t Bi ) ↔ (( (Aj ¤t Bj ))00 Ai ≡t Bi );
i=1 i=1 j=1
(ii) in the case of a system of “linear” fuzzy arithmetical equations with

given first operands and given results that has to be solved for the second
operand:
n
^ n
^ n
\
|= ∃X (Ai ∗t X ≡t Bi ) ↔ (Ai ∗t ( (Bj ˜
∗t Aj )) ≡t Bi ).
i=1 i=1 j=1
Proof. For the t-norm t = min we obviously have [[Θn ]] = [[Θ]] for each
well-formed formula Θ, hence (i) and (ii) are only reformulations of theorem
3.9 (ii) and 3.9 (v) respectively for the present case. QED
Corollary 3.11 Suppose that the t-norm t has property LSC(t). Then there
hold true:
T
(i) System (3.7) has a solution iff ni=1 (Ai ¤t Bi ) is a solution to this
T
system; and if (3.7) has a solution then ni=1 (Ai ¤t Bi ) is the greatest
one.
T
(ii) System (3.8) has a solution iff ni=1 (Bi ˜
∗t Ai ) is a solution to this sys-
T
tem; and if (3.8) has a solution then ni=1 (Bi ˜ ∗t Ai ) is the greatest
one.
Proof. We prove only (i) because the proof of (ii) is similar. Put D =
Tn
i=1 (Ai ¤t Bi ). If D is a solution of (3.7), then (3.7) obviously has a solution.
Hence suppose that (3.7) has a solution, say R. Then for each 1 ≤ i ≤ N
[[R00 Ai ≡t Bi ]] = 1
and thus by proposition 3.2 (ii)
n
^
[[ (R ⊂
=t Ai ¤t Bi )]] = 1
i=1
which by proposition 2.7 (iii) gives [[R ⊂

=t D]] = 1, i.e. the fuzzy set R is a
(fuzzy) subset of the fuzzy set D. But also
n
Y
[[ (R00 Ai ≡t Bi )]] = 1
i=1
and therefore
n
Y n
Y
[[(∃X (X 00 Ai ≡t Bi ))n ]] = [[∃X (X 00 Ai ≡t Bi )]] = 1.
i=1 i=1
Now by theorem 3.9 (ii) one has
n
Y
[[ (D00 Ai ≡t Bi )]] = 1
i=1
which means that D is a solution to (3.7). Hence (i) is proved. QED

3.5 Solvability degrees and approximate solutions

In a general setting, in referring to a fuzzy (relation) equation one quite
often has in mind an equation describing a relationship between fuzzy sets
in two (possibly different) “spaces”, i.e. universes of discourse. Such a form
of relationship is supposed to be represented by a fuzzy relation between the
elements of these “spaces”, i.e. over the (crisp) cartesian product of those
universes of discourse. More precisely, consider two fuzzy sets A ∈ IF (X )
and B ∈ IF (Y), as well as a fuzzy relation R ∈ IF (X × Y). Then a fuzzy
relation equation can be written down in a general form as
Θ(R, A) = B (3.9)
where Θ is a suitable operator producing a fuzzy set B out of a fuzzy set A

and a fuzzy relation R. With respect to our previous discussion the actual
situation is a quite direct generalization of equations of type (3.2). And the
case of a system
Θ(R, Ai ) = Bi , i = 1, . . . , n (3.10)
of fuzzy relation equations fits into these considerations as well.

Even more general, of course, is to consider Θ as an operator (of some
finite arity) which maps fuzzy sets and relations onto fuzzy sets or fuzzy
relations – and for which some of the arguments have to be determined. We
will not discuss the problem of fuzzy equations in this generality.
The following considerations are mainly restricted to the case of fuzzy
relation equations of type (3.2) and of systems (3.4) of such equations, i.e. to
special cases of (3.9) and (3.10). They can be and sometimes will be extended
to other types of fuzzy relation equations, as well as to fuzzy arithmetical
equations without principal difficulties. The last section has shown how far-
reaching the analogies are which one has in treating these “parallel” cases
guided by mainly the same ideas. Hence we essentially avoid that kind of
“parallelization” here.
Our strategy in the present chapter toward discussing the solvability
behaviour of fuzzy (relation) equations is essentially that one of a “many-
valued translation” — we changed, with respect to some t-norm t, from
traditional equations like (3.9): B = Θ(R, A) which are to be solved for R
or A to their many-valued counterparts B ≡t Θ(R, A) which – for lower
semicontinuous t-norms t – have the property that always
[[Θ(R, A) ≡t B]] = 1 ⇔ Θ(R, A) = B.

3.5. SOLVABILITY DEGREES AND APPROXIMATE SOLUTIONS 113
What we achieved to different levels of satisfaction were characterizations

of the truth degrees [[∃X(Θ(R, X) ≡t B)]] and [[∃X(Θ(X, A) ≡t B)]] which
did not involve the variable1 X, i.e. which are built up using only the “given
data” R, B and A, B respectively.
For systems of fuzzy equations the situation, in the last section, was al-
most the same: only the truth degree to be determined now was for example
n
Q
of the form [[∃X (Θ(X, Ai ) ≡t Bi )]], i.e. it was taken as the truth degree
i=1
of the sentence2 : “The system of fuzzy equations
Θ(X, Ai ) ≡t Bi , i = 1, . . . , n
has a solution.”
Instead of directly discussing the problem of solvability of fuzzy rela-
tion equations (3.9) or of systems (3.10) of such equations we consider the
truth degrees which we just mentioned as solvability degrees indicating the
solvability behaviour of our (systems of) fuzzy equations.
Definition 3.4 For each one of the fuzzy (relation) equations (3.9) and of
the systems (3.10) of such equations, which are supposed to be solved with
respect to the fuzzy relation R, their solvability degree is the truth degree
ξ0 =def [[∃X(Θ(X, A) ≡t B)]] (3.11)
and in case of a system of equations

n
Y
ξ =def [[∃X (Θ(X, Ai ) ≡t Bi )]]. (3.12)
i=1
As an immediate consequence we have the following corollary. To for-

mulate it we introduce in analogy with definition 1.9 the finite iteration T
of the t-norm t in the following way for any u1 , . . . , un+1 ∈ [0, 1]:
1 n+1 n
T ui =def u1, T ui =def ( T ui) t un+1. (3.13)
i=1 i=1 i=1
1
In some sense that variable changes its character in those formulas: in the first one it
is a variable for a fuzzy subset of an universe of discourse X and in the second one for a
fuzzy relation, that means for a fuzzy subset of an universe X × Y. But this distinction is
clear from the context and thus there is no need to express it in the notation too.
2
Of course, this sentence has to be read as being formulated in the language of many-
valued logic.
Corollary 3.12 For all A, Ai ∈ IF (X ) and B, Bi ∈ IF (Y) the solvability

degree ξ0 of equation (3.9) and the solvability degree ξ of the system (3.10)
of equations are
ξ0 = sup{ [[Θ(R, A) ≡t B]] | R ∈ IF (X × Y)}
n
ξ = sup{ T [[Θ(R, Ai) ≡t Bi]] | R ∈ IF (X × Y)}.
i=1
What, now, is the relation between the solvability of equations and the
value of the solvability degree? In one direction there is quite a simple
connection.
Proposition 3.13 If a fuzzy equation (3.9) or a system (3.10) of such equa-

tions has a solution, then its solvability degree is = 1.
Proof. Obvious from proposition 2.4 (i).
Thus, a solvability degree < 1 indicates that there is no solution.

For the converse case that the solvability degree is =1 only a partial
solution is at hand.
Suppose that a t-norm t is given such that there is a finite set L ⊆ [0, 1]
of truth degrees with 0, 1 ∈ L, and which is closed with respect to the
operations min, max and u 7→ 1 − u, as well as with respect to t. Then t can
be supposed to be a continuous t-norm and L is even closed with respect
to the operation ϕt . Let us call such a finite set L a (finite) t-clan. By
IF L (X ) we denote the class of all L-fuzzy subsets of X .
Given a continuous t-norm t and a finite t-clan L, for all A, Ai ∈ IF L (X )
and B, Bi ∈ IF L (Y) we also consider the relative solvability degrees
(L)
ξ0 =def sup{ [[Θ(R, A) ≡t B]] | R ∈ IF L (X × Y)} (3.14)
n
ξ (L) =def sup{ T [[Θ(R, Ai) ≡t Bi]] | R ∈ IF L(X × Y)}. (3.15)
i=1
Of course, using bounded quantification and writing R = IF L (X × Y)

one has
(L)
ξ0 = [[∃R X(Θ(X, A) ≡t B)]]
n
Y
ξ (L) = [[∃R X (Θ(X, Ai ) ≡t Bi )]].
i=1
Proposition 3.14 Suppose that L is a finite t-clan with respect to a con-

tinuous t-norm t. Then one has
(L)
ξ0 ≤ ξ0 and ξ (L) ≤ ξ
for all such (systems of ) fuzzy equations with A, Ai ∈ IF L (X ) and B, Bi ∈

IF L (Y).
Proof. The “full” solvability degrees ξ0 , ξ are, according to corollary 3.12,

suprema over bigger classes of fuzzy relations than the relative solvability
degrees, and hence not smaller numbers than those degrees. QED
Proposition 3.15 Suppose that L is a finite t-clan with respect to a con-

(L)
tinuous t-norm t and that ξ0 = 1 resp. ξ (L) = 1 for some (system of ) fuzzy
equation(s). Then that fuzzy equation or that system of fuzzy equations has
a solution.
Proof. Because of the finiteness of the t-clan L there exists an L-fuzzy

n
relation R0 such that [[Θ(R0 , A) ≡t B]] = 1 resp. T [[Θ(R0, Ai) ≡t Bi]] = 1.
i=1
In the case of a single equation that means Θ(R, A) = B by (2.14), and for
a system of fuzzy equations that first means [[Θ(R, Ai ) ≡t Bi ]] = 1 for each
i = 1, . . . , n by (T2), (T3), and thus also that always Θ(R, Ai ) = Bi . QED
From a theoretical point of view it is unsatisfactory to have these re-

strictions to the case of finite t-clans. Indeed, the guess is that the last
proposition holds true without this assumption. Yet, in practice this restric-
tion to finite t-clans is not such a severe one because quite often one considers
only finite universes of discourse and also t-norms for which a suitable subset
k
Lm of [0, 1] of the kind Lm = { m−1 | 0 ≤ k < m}, i.e. of equidistant rational
numbers is a t-clan.
The interest in the solvability degree of (systems of) fuzzy equations does
not only come from this connection between the existence of solutions and
the value of this degree being =1. There is another aspect that deserves to
be taken into consideration: a connection between the solvability index and
the approximate solvability of (systems of) fuzzy equations.
The background for this new aspect is that the generalized, “fuzzified”
identity relations ≡t are intimately related to distances between fuzzy sets.
The intuition behind this relation comes from the interpretation of ≡t

as a graded measure of equality of fuzzy sets and of their indistinguishabil-
ity. The negation of such an indistinguishability relation ≡t should hence
be a kind of graded distinguishability and thus (perhaps) even a kind of
“distance”. And that really is the case.
In the following considerations we refer to the notion of a metric in IF (X ).
As usual, a twoplace function % from IF (X ) into the non-negative reals IR+
is a metric iff for all A, B, C ∈ IF (X ) the following conditions hold true:
(M1) %(A, B) = 0 iff A = B, (identity property)

(M2) %(A, B) = %(B, A), (symmetry)
(M3) %(A, C) + %(C, B) ≥ %(A, B). (triangle inequality)
Sometimes the identity condition (M1) is weakened to the condition
(M1p ) %(A, A) = 0.
By a pseudo-metric % then a function is meant that fulfills conditions (M1p ),

(M2), (M3).
Definition 3.5 For each left continuous t-norm t, the binary t-distinguish-
ability function %t is that function on IF (X ) for which it is
%t (A, B) =def 1 − [[A ≡t B]] (3.16)
for all A, B ∈ IF (X ).
Obviously, for each t-distinguishability function %t one has for all A, B ∈

IF (X ):
%t (A, B) = [[¬(A ≡t B)]].
For fuzzy sets A, B ∈ X in the case of the t-norm t = tG = min one gets
by simple calculations the corresponding distinguishability function %G =
%tG
%G = sup (1 − [[x ε A ∩ B]])

x∈X
[[x ε A]]6=[[x ε B]]
= sup ([[x ε C
C(A ∩ B)]])
x∈X
[[x ε A]]6=[[x ε B]]
and in the case of the t-norm t = tL one gets after some elementary trans-
formations the distinguishability function %L = %tL as
n
%L (A, B) = min 1, max{0, sup ([[x ε A]] − [[x ε B]])}+
x∈X
o
+ max{0, sup ([[x ε B]] − [[x ε A]])} .
x∈X
This function %L (A, B) is loosely related to the Čebyšev distance of the

membership functions µA , µB defined as
dC (µA , µB ) = sup |µA (x) − µB (x)|
x∈X
in the sense that one always has

dC (µA , µB ) ≤ %L (A, B) ≤ 2 · dC (µA , µB )
and especially
³ ´
|= A ⊂
=tL B ⇒ %L (A, B) = dC (µA , µB )
which means, using the original crisp implication relation ⊂ for fuzzy sets
as mentioned in (2.4),
A ⊂ B ⇒ %L (A, B) = dC (µA , µB ).
The problem now is to find a necessary and sufficient condition for t to
yield via (3.16) a metric with properties (M1),. . . , (M3) as distinguishability
function %t .
As simple consequences of the definitions 2.4 of fuzzy singletons and 2.3
of fuzzified inclusion ⊂
=t and identity ≡t , first note that for all a ∈ X and
u, v ∈ [0, 1] one has
[[hhaiiu ⊂
=t hhaiiv ]] = u ϕt v = sup{w | u t w ≤ v} (3.17)
and hence especially for u = 1:
[[hhaii1 ⊂
=t hhaiiv ]] = v. (3.18)
Now we can formulate and prove our characterization result.
Theorem 3.16 Suppose t is a left continuous t-norm. Then the function

%t of (3.16) is a metric in IF (X ) iff t >
= tL , i.e. iff for all u, v ∈ [0, 1]:
max{0, u + v − 1} ≤ u t v.
Proof. Obviously, (M2) holds true because ≡t is a symmetric relation.

Condition (M1) by the corresponding definitions means
%t (A, B) = 0 ⇔ [[A ⊂ ⊂
=t B]] = 1 and [[B =t A]] = 1
for all A, B ∈ IF (X ). Additionally one has
[[A ⊂
=t B]] = 1 ⇔ [[x ε A →t x ε B]] = 1 for all x ∈ X
and furthermore for all u, v ∈ [0, 1]:
u ϕt v = 1 ⇔ sup{w | u t w ≤ v} = 1.
Therefore proposition 1.6 (i) yields that
[[A ⊂
=t B]] ⇔ [[x ε A]] ≤ [[x ε B]] = 1 for all x ∈ X
and hence also
[[A ≡t B]] = 1 ⇔ [[x ε A]] = [[x ε B]] for all x ∈ X

⇔ A = B.
To also get the triangle inequality (M3) one may start from the sup-t-
transitivity of ≡t , cf. proposition 2.4 (iii), which can be written as
[[A ≡t B ∧t B ≡t C]] ≤ [[A ≡t C]]
or equivalently as
%t (A, C) ≤ 1 − [[A ≡t B]] t [[B ≡t C]]. (3.19)
Then one has
%t (A, B) + %t (B, C) = 1 − ([[A ≡t B]] + [[B ≡t C]] − 1)

≥ 1 − tL ([[A ≡t B]], [[B ≡t C]])
≥ 1 − t([[A ≡t B]], [[B ≡t C]])
≥ %t (A, C)
by (3.19), provided that t > = tL .

>
Thus, if t = tL holds true, %t is a metric.
Now assume that t > = tL is not the case, i.e. that there exist membership
degrees u0 , v0 ∈ [0, 1] such that
u0 t v0 < u0 tL v0 = max{0, u0 = v0 − 1}.

Then consider points a, b ∈ X with a 6= b and let be
A0 =def hhaii1 ∪ hhbiiv0 ,

B0 =def hhaiiu0 ∪ hhbiiv0 ,
C0 =def hhaiiu0 ∪ hhbii1
unions of suitable fuzzy singletons of a and b. By straightforward calculations

one gets via (3.17), (3.18):
[[A0 ≡t B0 ]] = [[hhaii1 ⊂
=t hhaiiu0 ]] = u0 ,
[[B0 ≡t C0 ]] = [[hhbii1 ⊂
=t hhbiiv0 ]] = v0 ,
[[A0 ≡t C0 ]] = [[hhaii1 ⊂ ⊂
=t hhaiiu0 ]] t [[hhbii1 =t hhbiiv0 ]] = u0 t v0 .
Hence one has that
[[A0 ≡t B0 ∧t B0 ≡t C0 ]] = [[A0 ≡t C0 ]] = u0 t v0
< u0 tL v0 = [[A0 ≡t B0 & b0 ≡t C0 ]]
and therefore
%t (A0 , B0 ) + %t (B0 , C0 ) = 1 − ([[A0 ≡t B0 ]] + [[B0 ≡t C0 ]] − 1)

= 1 − [[A0 ≡t B0 & B0 ≡t C0 ]]
< 1 − [[A0 ≡t C0 ]] = %t (A0 , C0 ),
i.e. the triangle inequality does not hold true in this case.
Thus, if t >
= tL does not hold true for %t , the triangle inequality fails
and hence %t is not a metric. QED
A close look at the way our last theorem 3.16 was stated and proven
shows that it is enough to have a function ϕ introduced by definition (1.19)
for a given t-norm t. That means to some extent we can avoid supposing
that this function ϕ is a Φ-operator.
Formally, as we already mentioned in chapter 1, this definition does not
need the assumption of the left continuity of t. Thus, perhaps for the present
purposes of getting a metric for fuzzy sets, this assumption is not necessary
or can be weakened.
Indeed, our arguments to establish that %t has properties (M2) and (M3)
do not use the left continuity of t. Thus, the crucial point is the proof of
(M1). And for that case the following proposition holds true.
Proposition 3.17 For %t (A, B) =def 1 − [[A ≡t B]] there holds true
%t (A, B) = 0 ⇔ A = B for all A, B ∈ IF (X )
iff each function tu =def λv(u t v), u ∈ [0, 1] is (left) continuous at v = 1.
Proof. By the corresponding definitions one has
%t (A, B) = 0 ⇔ [[A ≡t B]] = 1

⇔ [[A ⊂ ⊂
=t B ∧t B =t A]] = 1
⇔ [[A ⊂ ⊂
=t B]] = 1 and [[B =t A]] = 1
and furthermore
[[A ⊂
=t B]] = 1 ⇔ x∈X
inf [[x ε A →t x ε B]] = 1
⇔ [[x ε A →t x ε B]] = 1 for all x ∈ X .
No left continuity of t is used here. Using only (1.19) now one gets
[[u →t v]] = 1 ⇔ sup{w | u t w ≤ v} = 1
for all u, v ∈ [0, 1]. All together that means
%t (A, B) = 0 ⇔
[[x ε A →t x ε B]] and [[x ε B →t x ε A]] = 1 for all x ∈ X (3.20)
and the equivalence of the following two conditions:
%t (A, B) = 0 ⇔ A = B
and
A=B ⇔ (3.21)
[[x ε A →t x ε B]] = 1 and [[x ε B →t x ε A]] = 1 for all x ∈ X .
It is obvious that
u ≤ v ⇒ [[u ϕt v]] = 1
holds true also without reference to the left continuity of t. Hence part (⇒)
of (3.21) is obvious too. But what about the reverse implication
[[u ϕt v]] = 1 ⇒ u ≤ v,
i.e. what about
u > v ⇒ [[u ϕt v]] < 1 ?
Assume that u0 > v0 and [[u0 ϕt v0 ]] = sup{w | uo t w ≤ v0 } < 1 holds true

for some u0 , v0 ∈ [0, 1]. Then there is a w0 < 1 such that u0 t w0 > v0 and,
of course, v0 < 1 hold true. Hence, if u0 > v0 and [[u0 ϕt v0 ]] = 1 hold true,
then the function tu0 = λv(u0 t v) is not left continuous at point v = 1.
Indeed otherwise one would have sup{w | u0 t w ≤ v0 } = 1 and therefore
v0 ≥ sup{u0 t w | u0 t w ≤ v0 }
= u0 t sup{w | u0 t w ≤ v0 } = u0 t 1 = u0 .
On the other hand, assuming that for each u ∈ [0, 1] the function tu =
λv(u t v) is left continuous at v = 1 gives for all u0 , v0 ∈ [0, 1]:
u0 ϕt v0 = sup{w | u0 t w ≤ v0 } = 1
⇒ u0 = u0 t 1 = u + 0 t sup{w | u0 t w ≤ v0 }
= sup{u0 t w | u0 t w ≤ v0 } ≤ v0 ,
i.e.
[[u0 ϕt v0 ]] = 1 ⇒ u0 ≤ v0 .
Hence assuming that all tu are left continuous at v = 1 gives
[[x ε A →t x ε B]] = 1 and [[x ε B →t x ε A]] = 1

⇒ [[x ε A]] ≤ [[x ε B]] and [[x ε B]] ≤ [[x ε A]]
for all x ∈ X and thus also the (⇐)-part of (3.21). QED
This proposition therefore shows that the continuity assumption in the

theorem 3.16 can be weakened.
Furthermore, by the reflexivity of ≡t it is obvious that condition (M1p ),
which instead of (M1) is used to define pseudo-metrics, always holds true for
%t . Thus the following results hold true too.
Corollary 3.18 %t is a metric iff t > = tL and each one of the functions
t u = λv(u t v) is left continuous at the point v = 1.
Corollary 3.19 %t is a pseudo-metric iff t >

= tL .
Unfortunately, however, it is not clear whether these corollaries are as

interesting as the theorem. The crucial point here is that in the case of a
t-norm t which is not left continuous, for ϕt the property
u t (u ϕt v) ≤ v for all u, v ∈ [0, 1] (3.22)
and thus (Φ2) fails, as was shown in the proof of the Existence Lemma.
But (Φ2), one of the properties which in chapter 1 were used to define Φ-
operators, means that the generalized conjunction ∧t and the generalized
implication →t together fulfill a kind of generalized modus ponens: the truth
degree of a well-formed formula H2 is never smaller than the truth degree of
the ∧t -conjunction of formulas H1 and H1 →t H2 .
So we have the situation that the greater generality of those corollaries
with respect to theorem 3.16 rests on the assumption that the modus ponens
condition (3.22) is not essential for the context the corollaries are used in.
Of course, the distances %t are bounded : %t (A, B) ≤ 1 always applies.
For each fuzzy equation (3.9): Θ(X, A) = B and each fuzzy relation R0
thus the truth degree
ξ0 (R0 ) =def [[Θ(R0 , A) ≡t B]]
not only evaluates the truth of the sentence “R0 is a solution of equation
Θ(X, A) = B”, i.e. measures the degree to which R0 is a solution of the
equation, but at the same time the degree
η0 (R0 ) =def 1 − ξ0 (R0 ),
i.e. the degree
η0 (R0 ) = %t (Θ(R0 , A), B) = [[¬(Θ(X, A) ≡t B)]]
evaluates the distance in between, i.e. the difference of the fuzzy sets Θ(R0 , A)
and B, and thus characterizes how “far” R0 is from a true solution.
For systems (3.10) of fuzzy equations the situation is the same. For any
fuzzy relation R0 the degree
n
Y
ξ(R0 ) =def [[ (Θ(R0 , Ai ) ≡t Bi )]]
i=1
evaluates to what extent R0 is a solution of the system (3.10). Hence also

in the present case the degree
n
Y
η(R0 ) =def 1 − ξ(R0 ) = [[¬( (Θ(R0 , Ai ) ≡t Bi ))]]
i=1
characterizes how “far” R0 is from a true solution to the system of fuzzy

equations.
The reference to a true solution is not essential for these considerations.
Thus the degrees η0 (R0 ) and η(R0 ) can be considered as measuring the
distance between the left-hand side(s) of (3.9) and (3.10) “produced” by
taking X = R0 and their “ideal” values, i.e. the right-hand sides of (3.9)
and (3.10) respectively. Vanishing distances give η0 (R0 ) = 0 or η(R0 ) = 0
and thus ξ0 (R0 ) = 1 or ξ(R0 ) = 1. On the other hand, the worst cases are
η0 (R0 ) = 1 and η(R0 ) = 1 and thus ξ0 (R0 ) = 0 and ξ(R0 ) = 0.
That means that the degrees ξ0 (R0 ) in the case of a single equation (3.9)
and ξ(R0 ) in the case of a system (3.10) of fuzzy equations measure a kind
of approximation quality of R0 with respect to some (possibly nonexistent)
ideal solution. Therefore the solvability degrees
ξ0 = sup ξ0 (R) and ξ = sup ξ(R),
R R
with the supremum in both cases having taken over all possible fuzzy re-
lations R, are upper bounds for the quality of approximate solvability of
(3.9) and (3.10) and hence the solvability degrees measure the best possi-
ble approximate solvability of (systems of) fuzzy equations in sense of the
distances %t .
The situation is especially simple in cases of single fuzzy equations. Re-
considering from the present point of view our earlier results in theorems
3.4 and 3.7 we already determined there the best possible approximate so-
lutions. Those results, under this new perspective, are once again presented
in Table 3.1.
fuzzy equation best possible approximate solution

R00 X = B R↓B
X 00 A = B A ¤t B
R ◦t X = S R−1 ¦t S
X ◦t R = S (R ¦t S −1 )−1
A ∗t X = B B˜∗t A
Table 3.1: Approximate solutions of some fuzzy equations
For systems of fuzzy equations the situation is more difficult. Without

restrictions concerning the t-norm which always is considered to be given,
theorem 3.9 only gives inequalities. But if one restricts the considerations
to the t-norm t = tG = min, then one can extend the results of corollary
3.10 to all types of equations in Table 3.1 in an (almost) uniform way and
has the following
Proposition 3.20 Suppose t = tG = min. For each type of fuzzy equations

in Table 3.1 then there exist best possible approximate solutions to systems
of such fuzzy equations; and in each case the intersection of all best possible
approximate solutions, as given in Table 3.1, of the single equations of the
system is such a best possible approximate solution to the whole system.
Surely, in the definitions of solvability degrees ξ, ξ0 also other kinds of

“distances” of fuzzy sets apart from our ≡t -based (pseudo-) metrics %t could
have been used. We have not discussed such possibilities so far in detail. The
reference to ≡t has the advantage that in the proofs of characterizations
of the solvability degrees (3.11), (3.12), our above methods are available
which are quite straightforward generalizations (to many-valued logic) of
usual methods from classical set theory and logic.
As proved in theorem 3.4 for lower semicontinuous t-norms t, the solv-
ability degree (3.11) can be characterized as the value
ξ0 = [[(A ¤t B)00 A ≡t B]]
where the fuzzy relation A ¤t B is defined according to definition 3.2.

From proposition 3.1 (ii) we have for the solvability degree ξ0 of the
equation (3.2)
^ _
ξ0 = ([[y ε B]] ϕt ([[x ε A]] t ([[x ε A]] ϕt [[y ε B]]))) (3.23)
y∈Y x∈X
by definition 2.3 and well known properties of Φ-operators and t-norms.

The situation is a bit more complicated for systems (3.4) of fuzzy relation
equations: in theorem 3.9, only lower and upper bounds for the solvability
degree (3.12) could be given in general, and a full characterization only in
the case that t = min. For systems (3.4) of fuzzy relation equations the
result was
n n
\ n
ξ≥ T [[(( (Aj ¤t Bj ))00 Ai ≡t Bi )]] ≥ Tξ
i=1 j=1 i=1
for lower semicontinuous t-norms t. Hence, by definition 3.2 and the mono-
tonicity of ◦t with respect to the inclusion of fuzzy sets we obtain
n ^ _ ^
ξ ≥ T ([[y ε Bi ]] ϕt ([[x ε Ai ]] t ([[x ε Aj ]] ϕt [[y ε Bj ]])))
i = 1 y∈Y x∈X 1≤j≤n
n
≥ T ξ. (3.24)
i=1
To get simpler formulas we now suppose that our t-norms t are contin-
uous. Then using proposition 1.7 we get a much simpler formulation of the
solvability degree (3.11) than earlier in (3.23). Remember for this that it is
_
hgt (A) = [[x ε A]] = [[∃x(x ε A)]]. (3.25)
x∈X
Proposition 3.21 For continuous t-norms t one has for the solvability de-
gree ξ0 of fuzzy relation equation (3.2)
ξ0 = hgt (B) ϕt hgt (A).
Proof. From (3.23) and the monotonicity of ϕt in the second argument we

get
^ _
ξ0 = ([[y ε B]] ϕt ([[x ε A]] t ([[x ε A]] ϕt [[y ε B]])))
y∈Y x∈X
and hence by proposition 1.7 (ii) and monotonicity

^ _
ξ0 = ([[y ε B]] ϕt [[x ε A]])).
y∈Y x∈X
Now, using results of chapter 1 we get by continuity of t

_ _
ξ0 = ( [[y ε B]]) ϕt ( [[x ε A]])
y∈Y x∈X
which is what is to be proved according to (3.25). QED
Corollary 3.22 For each continuous t-norm t for the solvability degree ξ0
of equation (3.2) there hold
(i) hgt (A) ≤ ξ0 ≤ 1,
(ii) hgt (A) = 1 ⇒ ξ0 = 1,
(iii) hgt (B) = 1 ⇒ ξ0 = hgt (A).
Proof. Obvious.
Remember inequalities (3.24). Put simply j = i in the minimum on all

1 ≤ j ≤ n. From monotonicity of t and of ϕt in the second argument then
we get for the solvability degree ξ of (3.12)
n n ^ _
Tξ≤ T ([[y ε Bi ]] ϕt ([[x ε Ai ]] t ([[x ε Ai ]] ϕt [[y ε Bi ]])))
i=1 i = 1 y∈Y x∈X
which by the proof of proposition 3.21 becomes

n n
T ξ ≤ T (hgt (Bi) ϕt hgt (Ai)).
i=1 i=1
Hence we have proved the following proposition.
Proposition 3.23 Let t be a continuous t-norm. Denote by ξi the solvabil-

ity degree of the i-th equation of system (3.4). Then for the solvability degree
ξ of the system (3.4) one has
n n
T ξ≤ T ξi.
i=1 i=1
Furthermore, from (3.24) and corresponding results of chapter 1 we get

immediately for this solvability degree
n ^ _ ^
ξ≥ T ([[y ε Bi ]] ϕt ([[x ε Ai ]] t ([[x ε Aj ]] ϕt [[y ε Bj ]]))).
i = 1 y∈Y x∈X 1≤j≤n
Now we can use u ϕt v ≥ v to reduce the term [[x ε Aj ]] ϕt [[y ε Bj (y)]] to

[[y ε Bj ]] and “move back” the supremum on x ∈ X to the second argument
of the Φ-operator and finally to the first argument of the following t-norm
term. This gives
Proposition 3.24 Let t be a continuous t-norm. Then,

n ^ n
\
ξ≥ T ([[y ε Bi ]] ϕt (hgt (Ai ) t ([[y ε Bj ]])))
i = 1 y∈Y j=1
holds for the solvability degree ξ of system (3.4) of fuzzy equations.

Unfortunately, hence, for systems of fuzzy relation equations we get only

bounds for the solvability degree, while for single fuzzy relation equations
we obtained a full characterization.
Another difference seems worth mentioning for applications. If we assume
that the input data are described by normal fuzzy sets, i.e. by such ones
whose height is =1, then proposition 3.21 always gives solvability degree =1
for the single fuzzy relation equation. Therefore proposition 3.23 also gives
only ξ ≤ 1 for the solvability degree of system (3.4). But from proposition
3.24 we get a simpler formula for a lower bound of ξ in this case. To write
down this formula, we have to use the fuzzified inclusion ⊂=t .
Corollary 3.25 Let t be a continuous t-norm. Suppose that the fuzzy sets
Ai in the system (3.4) are normal fuzzy sets for all indices i. Then
n n
\
ξ≥ T [[Bi ⊂=t Bj ]].
i=1 j=1
Proof. Straightforward.
One has even better results for the t-norm min. In this special case
n
obviously T . . . = min . . . and hence (3.24) is an equality, i.e. it gives
1≤i≤n
i=1
a characterization of the solvability degree of system (3.4) in terms of the
input and output fuzzy sets. Nevertheless, even the bounds which result
from propositions 3.23 and 3.24 are interesting in this case.
Corollary 3.26 Consider the t-norm min. Then

^ n
[ n
\ ^
min{ ξi , [[ Bi ⊂
=t Bj ]]} ≤ ξ ≤ ξi .
1≤i≤n i=1 j=1 1≤i≤n
Proof. For t = tG = min the corresponding Φ-operator ϕG (cf. table 2.3)

distributes over this t-norm. Hence we get from proposition 3.24 now
^ ^ n
\
ξ≥ min{[[y ε Bi ]] ϕG hgt (Ai ), [[y ε Bi ]]ϕG [[y ε Bj ]]}.
1≤i≤n y∈Y j=1
From this, elementary calculations give

^ ^ n
[ n
\
ξ ≥ min{ (hgt (Bi ) ϕG hgt (Ai )), ([[y ε Bi ]] ϕG [[y ε Bj ]])}
1≤i≤n y∈Y i=1 j=1
which proves the lower bound of ξ by proposition 3.21 and definition 2.3.
The upper bound immediately results from proposition 3.23. QED
3.6 Towards more difficult equations

For this section we adopt the notation A ◦t R = B instead of our earlier
R00 A = B, and additionally we use Zadeh’s original membership function
notation µA (x), . . . instead of our [[x ε A]]. The reason is that in this way a
simpler comparison can be made with parts of the research literature in the
engineering field.
Types of fuzzy relational equations studied so far and which are of spe-
cial significance for applications in the field of fuzzy modelling include the
following:
– fuzzy relational equations with sup-t-composition
B = A ◦t R, (3.26)
µB (y) = sup (µA (x) t µR (x, y)), (3.27)
x∈X
which is our type (3.2) of equations and which is strongly connected

with the “compositional rule of inference” of Zadeh (1973), cf. also
chapter 4;
– fuzzy relational equations with inf-s-composition
B = A ¦t R, (3.28)
µB (y) = inf (µA (x) st µR (x, y)), (3.29)
x∈X
which in a suitable sense is dual to the foregoing case;

– the adjoint fuzzy relational equations of Pedrycz (1985) with inf-ϕ-
composition
B = A ¤t R, (3.30)
µB (y) = inf (µA (x) ϕt µR (x, y)). (3.31)
x∈X
Besides these types, which might be viewed as beingof a basic nature, one
can consider some types of a more complex form which sometimes are formed
on the basis of the types given above. Of such possibilities we recall two:
3.6. TOWARDS MORE DIFFICULT EQUATIONS 129
– a convex combination of equations of types (3.26) and (3.28) as pro-

posed in Ohsato/Sekiguchi (1983)
B = λ · (A ◦t R1 ) + (1 − λ) · (A ¦s R2 ), (3.32)
t1
µB (y) = λ(y) · (sup (µA (x) st µR1 (x, y))) +
x∈X
+(1 − λ(y)) · inf (µA (x) st1 µR2 (x, y)) (3.33)
x∈X
with λ : Y → [0, 1] and the t-norm t, as well as the t-conorm st1 , in

Ohsato/Sekiguchi (1983) taken as max, min only;
– fuzzy relational equations with an equality operator as discussed in

diNola/Pedrycz/Sessa (1988)
B = A 1t R, (3.34)
µB (y) = sup (µA (x) ϕt µR (x, y) ∧ (µR (x, y) ϕt µA (x))), (3.35)
x∈X
For all the basic forms of equations (3.28) – (3.31) the family of solu-
tions, if solutions do exist at all, has been characterized and their extremal
(maximal or minimal) elements have been obtained. Moreover, for systems
Bi = Ai ◦t R, i = 1, . . . , n (3.36)
Bi = Ai ¦t R, i = 1, . . . , n (3.37)
Bi = Ai ¤t R, i = 1, . . . , n (3.38)
of such equations the relevant results are also available – mainly under the
condition that they do have an exact solution.
To have an overall picture of the results they are collected in Table 3.2.
Concerning the notation used in this table we have to add two explana-
tions. First, by R0 , R00 , R000 we denote the sets of solutions of the systems
(3.36), (3.37), (3.38) respectively of fuzzy relational equations. Secondly, as
a dual to the Φ-operator (1.19) we use the sometimes so-called β-operator
defined for all u, v ∈ [0, 1] as:
u βt v =def inf{w ∈ [0, 1] | u st w ≥ v}; (3.39)
and in an analogous manner as the operator ¤t was introduced in definition

3.2 with reference to the implication connective →t , i.e. to the Φ-operator ϕt ,
Type of Solution3 to a Solution3 to a

Interpretation
equation single equation system of equations
the (fuzzified) full im-
age of a fuzzy set under Tn
B = A ◦t R, a fuzzy relation R̂ =
R̂ = A ¤t B i=1 (Ai¤t Bi )
sup-t- or for each point b of Tn 0
composition R̂ = sup R0 R̂ = i=1 sup Ri
B there is a point a of
A such that b can be
reached from point a
B = A ¦t R, dual to the first case in Sn

sense that Ř = A ¢t B Ř = i=1 (Ai ¢t Bi )
inf-s- Sn 00
00 Ř =
composition B = A ¦t R Ř = inf R i=1 inf R
iff B̄ = Ā ◦t R̄
Sn
B = A ¤t R, each point b of B can R̃ = A ×t B R̃ = i=1 (Ai ×t Bi )
inf-ϕ- be reached from every 000
Sn 000
R̃ = inf R R̃ = i=1 inf R
composition point a of A
Table 3.2: Basic types of fuzzy relation equations for control applications
now a “dual” operator ¢t for fuzzy sets is introduced such that C := A ¢t B

is characterized by the membership function
C := A ¢t B : µC (x, y) =def µA (x) βt µB (y).
At present however, for the general case it seems quite difficult to give
simple and easy-to-check conditions for the solvability of a system of equa-
tions. The discussion in chapter 4 concerning sufficiency conditions for the
non-interactivity of systems of control rules gives an indication of the prob-
lems one is confronted with in looking for simple conditions here. This was
one of the reasons for the discussion of degrees of solvability for example in
Gottwald (1986) and in this chapter. Actually we will present only some
conditions for the solvability of single equations; cf. Table 3.3. The case of
systems of such fuzzy relational equations has not yet been considered in
detail.
As it becomes quite clear from later discussions concerning applications
to fuzzy control and fuzzy modelling in general, the results for example of
Table 3.2 have a significant value only in the case that solutions really exist,
i.e. that not only “approximate” solutions (in some suitable sense of that
Type of equation Condition

V W
B = A ◦t R (µA (x) ≥ µB (y))
y∈Y x∈X
V W
B = A ¦t R (µA (x) ≤ µB (y))
y∈Y x∈X
no restrictive condition in the case that t is

B = A ¤t R
strictly monotonic in both arguments
Table 3.3: Necessary and sufficient conditions that a single equation

has a solution.
word) exist. If this true solvability is not the case – and it is this more
uncomfortable situation one usually meets in practice – then because of
the (present) lack of an extended mathematical theory of such (systems of)
equations, the user has to think about other ways of overcoming the problem
of the nonexistence of (true) solutions.
A simple, and perhaps for the practitioner the most obvious, way out is
to use the formulas which describe solutions – in the case of solvability –
even if a solution to the system of fuzzy relation equations to be considered
does not exist – and then to check the quality of the “approximate solution”
derived in this way.
But having taken this point of view one can move one step further: in-
stead of having proven a formula to give a solution in the case of solvability,
one can start from a formula which one guesses to describe a solution –
of course, if there are some acceptable reasons for such a guess. And, in-
deed, for some classes of fuzzy relational equations such acceptable guesses
are available. To present some basic ones let us distinguish for relational
equations Θ(R, A) = B two different types.
Definition 3.6 A fuzzy equation Θ(R, A) = B will be said to be of sup-type

in the case that one has
µB (y) = µΘ (A, R)(y) = sup Γ (µA (x), µR (x, y)) (3.40)
x∈X
where the term Γ is built up using the membership degrees µA (a), µR (a, b)
and combining them for example by a t-norm, a Φ-operator or some suitable
3
in case of solvability
other kinds of “simple” operators; and such an equation will be said to be of

inf-type in the case that one has
µB (y) = µΘ (A, R)(y) = inf Γ (µA (x), µR (x, y)) (3.41)

x∈X
with the term Γ chosen accordingly.
To show the influence of this distinction on the structure of the set of so-
lutions of these equations we will consider the following facts which should be
taken into account for the discussions of true and of approximate solutions.
Fact 1: The union of any two solutions of a fuzzy relational equation of

sup-type is again a solution to this equation, and hence this equation
has a greatest solution.
Fact 2: The intersection of any two solutions of a fuzzy relational equation

of inf-type is again a solution to this equation, and hence this equation
has a smallest solution.
Fact 3: Any system of fuzzy relational equations of sup-type has in the case
of solvability as the greatest solution the intersection of all the greatest
solutions of its single equations.
Fact 4: Any system of fuzzy relational equations of inf-type has in the case
of solvability as the smallest solution the union of all the smallest
solutions of its single equations.
It is interesting to mention that these types of behaviour can also be

found with some mixed types of fuzzy relational equations. To look at an
example we refer to the convex combination form (3.32), (3.33) of fuzzy
relational equations discussed by Ohsato/Sekiguchi (1983,1985) for A ◦t
R1 as sup-min-composition and A ¦t R2 as inf-max-composition. Indeed,
in Ohsato/Sekiguchi (1985) a solvability behaviour of such equations is
proven which combines Facts 1 and 2: the type (3.32) of equations has an
extremal solution (Ř1 , R̂2 ) in the sense that R10 ⊆ Ř1 and R20 ⊆ R̂2 for all
solutions (R10 , R20 ) to this equation. Extending these results one can prove
that for solutions (R10 , R20 ) and (R100 , R200 ) of this equation also (R10 ∪ R100 , R20 ∩
R200 ) is a solution. And extending Facts 3 and 4 one can prove that an
extremal solution of a system of such equations (3.32) is determined – in
the case of its existence – in its “first coordinate” as the intersection of the
“first coordinates” of the extremal solutions of the single equations, and
in its “second coordinate” as the union of the “second coordinates” of the

extremal solutions of the single equations.
What we have achieved in this chapter is to give characterizations of
the solvability of some simple types of fuzzy equations and systems of such
equations whose usefulness in applications is already established. Using the
solvability degrees of a fuzzy equation or of a system of such equations a char-
acterization became possible to which degree – measured by the (pseudo-)
metric %t connected with ≡t – the fuzzy equation resp. the system of fuzzy
equations has a solution. At the same time, this solvability degree gives in-
formation about best possible approximate solutions if there does not exist
any real solution.
For single fuzzy equations this solvability degree can be calculated from
the results proved in this chapter. For systems of fuzzy equations at least
bounds for this solvability degree may be calculated: a lower bound was
explicitely given in theorem 3.4, but from the fact that this lower bound
also is an upper bound for the “n-th power” of the solvability degree it
follows that each truth value whose “n-th power” is not smaller than this
lower bound itself is an upper bound of the solvability degree.
Further research should test the usefulness of these solvability degrees for
practical applications of fuzziness. Some discussions relating to this point
will be given later on in chapter 5.
Chapter 4
Fuzzy controllers
4.1 The construction of fuzzy controllers

In very general terms, a fuzzy controller is some device that is intended to
modelize some roughly known or roughly described process. Used together
with the real process in feedback mode it may really act as a control device,
and used in feedforward mode it can be used for predicting the process
behaviour. In a certain sense thus a fuzzy controller realizes some vaguely
known, vaguely described, or intrinsically rough “algorithm”.1
The usual textbook examples of such vague “algorithms” comprise things
like: cooking recipes; instructions for driving a car; rules governing the be-
haviour of a human operator who controls some chemical process; medical
diagnosis etc.
All these examples and consequently the fuzzy controller too are sup-
posed to be rule-based, i.e. constituted by a (finite) set of control rules. That
is an idea the fuzzy controller has in common with a widespread naive, as
well as formalized, understanding of algorithmic methods, but more general
also with most methods of knowledge transfer and knowledge representation.
The background idea is simply that the essential structure of such knowledge
can be split into small portions of the type: if there is some situation then
something is to be expected or has to be done.
That means that all those rule-based kinds of knowledge representation
1
Of course here we refer to an intuitive, everyday understanding of the general notion
of “algorithm”; it is not the formalized, mathematical notion of algorithm that should be
the background idea. But, surely, there are (loose) connections between the formal and
the present informal notion of algorithm.
136 CHAPTER 4. FUZZY CONTROLLERS
refer to input as well as to output information. Usually the concrete input

and output informations for each single case are seen as specific values of
input or output variables. The sets of possible values of each such (input or
output) variable are supposed to be determined in advance and very often
considered as possible numerical values with respect to some scaling. But in
general they are simply abstract sets of objects.
This last-mentioned fact is the background for a nice theoretical simpli-
fication. In any case, namely, one can suppose to have only one input and
one output variable. For, if one has initially for example n input variables
α1 , . . . , αn with respective universes V1 , . . . , Vn of possible values, one can
introduce a new variable ᾱ, often denoted ᾱ = (α1 , . . . , αn ), whose set of
possible values is the cartesian product V1 × · · · × Vn , i.e. whose values are n-
tuples of values of α1 , . . . , αn . And if one has m output variables β1 , . . . , βm ,
one either considers one new, more complex output variable β̄ = (β1 , . . . , βm )
or – normally much better – one takes into account m separate one-output
fuzzy controllers which work parallel. Therefore in the following it is always
enough to discuss the case of one input and one output variable.
The choice of values for the input and output variables is one of the
topics with a very specialized understanding in the case of fuzzy controllers:
all these input and output variables are taken as linguistic variables,2 i.e.
their values are fuzzy subsets of suitable universes of discourse. Some specific
interpretations are tied with this choice of “linguistic values”. The first one,
which was essential for naming these variables “linguistic” ones, is that it
is supposed that these linguistic values are represented by common words
from everyday language (like: high speed, hot water, very heavy rain etc.).
These words are understood as meaning suitable fuzzy sets. The second
one of these specific interpretations is that these fuzzy sets have to be fuzzy
subsets of some suitable universe of discourse – and that the members of
this suitable universe of discourse are considered as possible “true” values
of some crisp, i.e. traditional variable loosely connected with the linguistic
variable (like: speed, temperature of water, amount of rain per area and time
etc.). And, finally, as a third and very essential interpretation one assumes
that the “linguistic” values of the linguistic variable code unsharp, vague
information about the possible “true” value of the traditional variable tied
with the linguistic one.
Finally, crucial and constitutive for the idea of a fuzzy controller is that
2
The notion of linguistic variable was introduced by Zadeh (1975); cf. also Zimmer-
mann (1985), Bandemer/Gottwald (1989).
4.1. THE CONSTRUCTION OF FUZZY CONTROLLERS 137
the control rules (and hence the fuzzy controller) shall also become applicable
in situations, i.e. for values of the input variable which are not explicitly used
in the (premises of the) control rules.
A fuzzy controller thus constitutes a connection, a relation between the
values of the input and the output variables. Formally, hence, one intends
to construct a fuzzy controller as a fuzzy relation (in the precise sense of
section 2.2), even to identify it with such a fuzzy relation, and to get this
relation out of the system of control rules.3
There is no uniform way to construct fuzzy controllers. But at present
three strongly connected methods are used in most cases. To discuss them in
more detail the input variable now will be denoted α, the output variable β.
Both are linguistic variables; their universes of discourse (for the “linguistic”
values) shall be X for α and Y for β.
The starting point in each case is a finite list of control rules
if value(α) = Ai then value(β) = Bi , i = 1, . . . , n, (4.1)
the generating family that constitutes the fuzzy controller.

Because we suppose to have one fixed input and one fixed output variable
we can follow common usage to take
Ai |=
⇒ Bi , i = 1, . . . , n (4.2)
as a shorthand notation for the generating family of control rules (4.1).

The pioneering approaches to fuzzy (logic) control given in Mamdani
(1974, 1976) and Mamdani/Assilian (1975) on the theoretical side and in
Holmblad/Østergaard (1982) with the first industrial application seem-
ingly followed quite different strategies. Both started from a finite list (4.1)
of control rules; but then Mamdani wrote down almost directly a fuzzy re-
lation – and Holmblad/Østergaard used the idea that a specific input
value of the input variable4 would activate each control rule to a certain de-
gree, and then this control rule would “act to that degree”, i.e., contribute
with its specific output to the total output according to this degree of acti-
vation.
3
The terminology is not really unique up to now. Usually a fuzzy controller is considered
as a fuzzy relation. In this sense a system of control rules isn’t yet a controller, but
constitutes one. And a (hardware or software) device actually having such a fuzzy relation
implemented and thus realizing it then is a – practical – realization of a fuzzy controller.
4
Holmblad/Østergaard (1982) really did consider three independent input vari-
ables; but that is inessential here according to our foregoing remarks.
First, let us have a closer look at this method of activation degrees. Given
the degree of activation actj (A) of the j-th control rule Aj |= ⇒ Bj with re-
spect to the input value A, the output Bj of that j-th rule is “weighted”, i.e.
suitably modified into a new fuzzy set Bjw depending on the degree actj (A).
At the end all the modified, weighted outputs “together”, i.e. suitably su-
perposed yield the final, total output B for the given input A.
Obviously, there are a lot of decisions to be made about the formal, math-
ematical character of the operations vaguely mentioned in this description
of the approach through degrees of activation. The essential ones concern:
• the definition of the degree actj (A) of activation of the j-th control
rule for the input A;
• the definition of the “weighted” output Bjw of the j-th rule if activated
to degree actj (A);
• the method of superposition of the weighted outputs Bjw of all the

(activated) control rules.
The degree of activation actj (A) is, from the intuitive understanding of
this process of activation, measuring the coincidence of the actual input A
and the “rule input” Aj . If under the possibilistic reading both fuzzy sets A
and Aj are taken as possibility distributions for the “true” input value it is
intuitively appealing to take
actj (A) =def hgt (A ∩t Aj ) (4.3)
with respect to a suitable t-norm t. Holmblad/Østergaard (1982) take

here t = min, such that they have
actj (A) = sup min{µA (x), µAj (x)}. (4.4)

x∈X
The special degree of activation (4.3) can be understood as the degree of

possibility to which control rule j may be the “correct” one. In some sense,
thus, this is an optimistic view concerning the activation of control rule j
by input A. A more pessimistic view may be represented for example by
choosing
act∗j (A) =def [[A ≡t Aj ]] (4.5)
or some other “degree of equality” of A and Aj . This idea, however, seems to

be very pessimistic because (4.5) may be read as meaning that control rule
j is activated (to a corresponding degree) just if the fuzzy input corresponds

exactly to the situation covered by this j-th control rule. It is more suitable,
perhaps, to approach the matter from the point of view that control rule
j should become activated (to a corresponding degree) if the fuzzy input
corresponds to a situation that is covered by the j-th rule. Then instead of
(4.5) one would define
act?j (A) = [[A ⊂

=t Aj ]] = [[∀x(x ε A →t x ε Aj )]]. (4.6)
But these “pessimistic” points of view have not till now been used for any
realized fuzzy controllers. (Also, the theoretical consequences of those “pes-
simistic” approaches need further investigations.)
Accepting the possibilistic point of view a point y ∈ Y of the universe of
discourse of the output variable β is the possible “true” output value if it is
the “true” output of the j-th control rule and if that rule is activated. This
idea, written down in our language for fuzzy sets, means for the “weighted”
outputs Bjw :
[[y ε Bjw ]] = [[y ε Bj ∧t actj (A)]]. (4.7)
Here Holmblad/Østergaard (1982) take t = tP = product and thus

have in the usual notation with membership functions:
µBjw (y) = actj (A) · µBj (y) for all y ∈ Y. (4.8)
Of course, other t-norms can do the job as well. Sometimes t = min is taken
instead of (4.8).5
But it seems that for the modification of the output fuzzy sets the choice
(4.8), i.e. t = tP in (4.7), is intuitively often more appealing then the choice
t = min. For choosing t = min causes a modification of Bj that can be
visualized as a “cutting at level actj (A)” in the sense that for all x0 ∈ X with
µBj (x0 ) > actj (A) one obtains through this modification µBjw (x0 ) = actj (A)
and for all other x ∈ X their membership degrees with respect to Bj and Bjw
coincide. With t = tP instead the membership function of Bjw has a kind of
similarity with that of Bj in the sense of some “proportional reduction” of
all membership degrees.
As we mentioned in connection with (4.3) and (4.5), (4.6), for the defini-
tion of the degrees of activation, as well as for the definition of the weighted
5
In the current literature as e.g. Zimmermann (1985) one then often speaks of the
max-dot resp. max-min inference method; the term “max” in these names refers to the
method of superposition we discuss below – cf. (4.9) and the remarks that follow there.
outputs Bjw , there exist conflicting, dual points of view connected with ei-
ther the “optimistic” or the “pessimistic” perspective. Further variants for
defining the weighted outputs will only be discussed later on, cf. (4.15).
The final point in this method of activation degrees is to make a su-
perposition of the weighted outputs of the single (activated) control rules.6
From the “optimistic” perspective of possibilistic understanding it is natural
to collect, i.e. to sum up all the “local” control information provided by the
outputs of the single control rules according to their degrees of activation,
i.e. to take the final (fuzzy) output B ∈ IF (Y) as
n
[
B =def Biw . (4.9)
i=1
Together with formulas (4.3) and (4.7) that gives for each y ∈ Y:7
n
_
|= y ε B ↔ (y ε Bi ∧t1 ∃x(x ε A ∧t x ε Ai )) (4.10)
i=1
what in the case of LSC(t1 ) can be written as
n
_
|= y ε B ↔ ∃x ((x ε A ∧t x ε Ai ) ∧t1 y ε Bi ) (4.11)
i=1
n
W
with as the finite iteration of the max-disjunction.
i=1
Assuming furthermore t1 = t together with LSC(t) formula (4.11) can
be transformed into the simpler form
n
_
|= y ε B ↔ ∃x (x ε A ∧t x ε Ai ∧t y ε Bi )
i=1
n
_
↔ ∃x(x ε A ∧t (x ε Ai ∧t y ε Bi ))
i=1
n
[
↔ ∃x(x ε A ∧t (x, y) ε (Ai ×t Bi )). (4.12)
i=1
6
It is inessential here to restrict the considerations to the activated rules only. If one
takes into account all the rules, i.e. also those rules with degree of activation = 0, then
because of (4.7) those rules contribute the empty fuzzy set ∅ as weighted output to that
superposition (4.9) – and thus in any case do not influence the actual output at all. And
precisely that last point will be of leading importance later on when discussing other
variants of weighted outputs for the “pessimistic” case.
7
Here as already in section 1.3 we assume that the biimplication ↔ is based on some
lower semicontinuous t-norm t0 : it is not necessary to be more specific because |= H1 ↔ H2
is equivalent to [[H1 ]] = [[H2 ]] independent of the specific choice of t0 .
With the set theoretic notion of the full image R00 A of a fuzzy set A under
a fuzzy relation R, cf. definition 2.9, formula (4.12) means
n
[
B=( (Ai ×t Bi ))00 A. (4.13)
i=1
On the more pessimistic side one tends to take instead of (4.9) for the su-
perposition of the single weighted outputs Bjw the idea that the final control
output of the set of control rules should be chosen in such a way that it is
proposed by all relevant, i.e. activated rules. This idea, hence, corresponds
to taking the fuzzy output as
n
\
B =def Biw . (4.14)
i=1
But now, with (4.7) any control rule which is not activated at all, i.e. for
which one has actj (A) = 0 as degree of activation, would contribute Bjw = ∅
to the intersection (4.14) and hence cause the fuzzy output to become B = ∅.
That, obviously, is counterintuitive. The idea, thus, leading to (4.14) has
to be combined with another choice of the weighting process too, i.e. with
another definition of Bjw instead of (4.7). One possibility is to take a new
weighted output Bj? such that
[[y ε Bj? ]] =def [[act?j (A) →t1 y ε Bj ]] (4.15)
which needs the assumption LSC(t1 ). That approach, together with (4.14)
n
V
and now with as the finite iteration of the min-conjunction, gives
i=1
n
^
|= y ε B ↔ (act?i (A) →t1 y ε Bi ). (4.16)
i=1
With a degree of activation according to (4.3) then one has

n
^
|= y ε B ↔ (∃x(x ε A ∧t x ε Ai ) →t1 y ε Bi )
i=1
^n
↔ ∀x(x ε A ∧t x ε Ai →t1 y ε Bi ) (4.17)
i=1
n
^
↔ ∀x (x ε A ∧t x ε Ai →t1 y ε Bi ) (4.18)
i=1
using proposition 1.27 (ii). Assuming furthermore t1 = t together with

LSC(t) yields by proposition 1.19 (i)
n
^
|= y ε B ↔ ∀x (x ε A →t (x ε Ai →t y ε Bi ))
i=1
n
^
↔ ∀x(x ε A →t (x ε Ai →t1 y ε Bi )). (4.19)
i=1
With the degree of activation (4.6) instead one has

n
^
|= y ε B ↔ (∀x(x ε A →t x ε Ai ) →t1 y ε Bi )
i=1
^n
↔ ∃x((x ε A →t x ε Ai ) →t1 y ε Bi ) (4.20)
i=1
which, even for t = t1 seem to be hard to transform into a much simpler

and more appealing form like (4.19).
Mamdani’s way of approach toward fuzzy control in Mamdani (1974,
1976) immediately aimed at the transformation of each control rule Ai |= ⇒ Bi
into a fuzzy relation Υ(Ai |= ⇒ Bi ) in X × Y coding that rule and at the
“collection” of all those codes of the single control rules into a fuzzy relation
R, the fuzzy controller. He used as the final input-output-behaviour of the
fuzzy controller R the approach of taking
B := R00 A (4.21)
as the output of the fuzzy controller R for the input A, thus using an earlier
proposal by Zadeh (1973).8
Writing as usual Ri for the code of control rule i, i.e. taking
Ri =def Υ(Ai |=
⇒ Bi ), (4.22)
Mamdani’s way of “collecting” all these coding relations Ri into one fuzzy
relation R was to “sum them up” in the sense of taking their union (cf.
definition 2.5) and thus having
n
[ n
[
R= Υ(Ai |=
⇒ Bi ) = Ri (4.23)
i=1 i=1
8
Zadeh (1973) did not use the name “full image of A under R” for (4.21) but spoke of
the “compositional rule of inference” and wrote A ◦ R for our R00 A.
independent of the different ways Ri = Υ(Ai |=

⇒ Bi ) may have been chosen.
There is no general rule governing the coding procedure for the control
rules. But there is a kind of agreement that ideally these codes Ri should
have the property that one has to have
Ri00 Ai = Bi (4.24)
for each i = 1, . . . , n. Furthermore, there is often the background idea that
the relation Ri has to reflect in some sense the fact that in its original, full-
length form (4.1) the control rules are seen as fuzzy implications. Thus one
often constructs Ri = Υ(Ai |= ⇒ Bi ) in a manner which is analogous to well
known possibilities one has in two-valued classical propositional logic to ex-
press the implication connective using for example the conjunction, negation,
and disjunction connectives. Of course, there are many such possibilities if
one does not restrict oneself to the shortest and simplest ones.
Prefered candidates for ways to build the relations Ri are often:
Υ(Ai |=
⇒ Bi ) = A × B, (4.25)
Υ(Ai |=
⇒ Bi ) = (A × B) ∪ (A × Y ), (4.26)
⇒ Bi ) = (A × Y ) ∪ (X × B).
Υ(Ai |= (4.27)
For many more such proposals cf. Mizumoto (1982), Mizumoto/Zimmer-
mann (1982).
Mamdani originally had chosen version (4.25) thereby meeting condition
(4.24) for each control rule.
If one compares (4.23) with (4.13) one immediately recognizes that there
is a very close connection between the approaches by Holmblad/Øster-
gaard (1982) and by Mamdani (1976), notwithstanding the fact that both
these approaches quite naturally generalize into different directions.
The third method of approach toward fuzzy controllers, besides those
by Holmblad/Østergaard and by Mamdani, also intends to consider a
fuzzy controller as a fuzzy relation R which is constituted by a set (4.1) of
control rules. And that approach too accepts the input-output-behaviour
(4.21) in the sense of the “compositional rule of inference” .
As a side remark let us mention that, of course, one is not forced to
accept that approach through (4.21) and definition 2.9, i.e. to accept the
“compositional rule of inference” as the decisive way to convert an input into
an output. But at present there does not exist any other reasonable proposal
here – besides, in some sense, the direct approach of Holmblad/Øster-
gaard which, however, in special cases like (4.13) reduces to this method
too.
The essential idea now is that this type of input-output-behaviour of the

fuzzy controller R should prove to be a generalization of the input-output-
relation (4.24) immmediately pretended by the control rules, i.e. that from
input Ai the output Bi should be produced by R, i.e.
R00 Ai = Bi , for all i = 1, . . . , n (4.28)
has to be the case.

In the same way as (4.24), for each i = 1, . . . , n, can be understood as
a condition which the fuzzy relation R has to meet for the fuzzy sets Ai , Bi
given by the i-th control rule, the condition (4.28) can be seen as constitutive
for the fuzzy controller R. In other words, as (4.24), for each i = 1, . . . , n,
can be considered as an equation which has to be solved with respect to
an unknown fuzzy relation Ri , the condition (4.28) can be considered as a
system of equations9 which has to be solved with respect to an unknown
fuzzy relation R.
This third kind of approach has stimulated many studies concerning solu-
tions and sets of solutions of such relation equations, as well as of systems of
them, cf. diNola/Sessa/Pedrycz/Sanchez (1989), and is also the back-
ground for further discussion in this book.
4.2 The problem of interaction

The second one (4.23) of these ways to construct a fuzzy controller out of the
fuzzy relations Υ(Ai |=
⇒ Bi ) which code its single control rules, sometimes
is not in accordance with (4.28), i.e. with the idea that (4.24) ought also
to hold true for each control rule in the case that instead of Ri the fuzzy
controller R is taken into account. This possibility was seen quite early
among control engineers, cf. e.g. Czogala/Pedrycz (1981). That means
that it may happen that there finally is a control rule Ak |=
⇒ Bk such that
Bk 6= R00 Ak (4.29)
is the case. But that means that the final fuzzy controller R from (4.23)
sometimes may act in a different way than was intended in the set of control
rules and the coding procedure. As we will say for this situation: the control
rules may interact.
9
Because of the specific form of these equations they usually are denoted as (fuzzy)
relation(al) equations.
4.2. THE PROBLEM OF INTERACTION 145
4.2.1 General results on interactivity

Definition 4.1 A generating family of control rules Ai |= ⇒ Bi , i ∈ I =
{1, . . . , n}, for a fuzzy controller R is called interactive iff there exists some
index k ∈ I such that Bk 6= R00 Ak ; otherwise the generating family of control
rules is called non-interactive.
As in this definition we will, in the present section, always use the nota-
tion I = {1, . . . , n} for simplicity of notation.
Because, according to definition 2.3, the condition of equality for fuzzy
sets can be split into two conditions of being subsets of one another, the
problem of non-interactivity of the families of generating rules can be split
too.
Definition 4.2 Suppose that R is a fuzzy controller to realize the family of

control rules Ai |=
⇒ Bi . We say that R has the superset property w. r. t. the
generating family (4.2) iff
R00 Ai ⊃ Bi for all i ∈ I , (4.30)
and we say that R has the subset property w. r. t. this family iff
R00 Ai ⊂ Bi for all i ∈ I (4.31)
hold true.
We always assume in the following considerations that the fuzzy con-

troller R was constructed from a family (4.1) of control rules.
Proposition 4.1 (i) A sufficient condition that the superset property holds
true for this fuzzy controller R is that for all i ∈ I and all y ∈ Y there exists
some k ∈ I and some x0 ∈ X such that
[[y ε Bi ]] ≤ [[(x0 ε Ai ∧t ((x0 , y) ε Υ(Ak |=

⇒ Bk )))]].
If X is a finite set, this condition is necessary too.

(ii) A necessary and sufficient condition that the subset property holds
true for this fuzzy controller R is that for all i, j ∈ I, all x ∈ X and all y ∈ Y
one has
[[y ε Bi ]] ≥ [[x ε Ai ∧t ((x, y) ε Υ(Aj |=

⇒ Bj ))]].
Proof. By the corresponding definitions we immediately have that for all

y ∈ Y:
[[y ε R00 A]]
= [[∃x(x ε A ∧t ∃I k((x, y) ε Υ(Ak |=
⇒ Bk )))]]
= [[∃x∃I k(x ε A ∧t ((x, y) ε Υ(Ak |=
⇒ Bk )))]].
Therefore the superset property means that for all i ∈ I and y ∈ Y
[[y ε Bi ]] ≤ [[∃x∃I j(x ε Ai ∧t ((x, y) ε Υ(Aj |=
⇒ Bj )))]]
has to be the case; and the subset property means that for all i ∈ I and
y ∈ Y:
[[y ε Bi ]] ≥ [[∃x∃I k(x ε Ai ∧t ((x, y) ε Υ(Aj |=
⇒ Bj )))]]
has to be the case.
Because ∃ means taking the supremum of the corresponding truth de-
grees, and the generating families of control rules are always supposed to be
finite, we have that the superset property means that for all i ∈ I and all
y ∈ Y there exists some k ∈ I such that
[[y ε Bi ]] ≤ [[∃x(x ε Ai ∧t ((x, y) ε Υ(Ak |=
⇒ Bk )))]] (4.32)
= sup [[x ε Ai ∧t ((x, y) ε Υ(Ak |=
⇒ Bk ))]]
x∈X
has to be the case, and correspondingly the subset property means that for
all i, j ∈ I and all y ∈ Y
[[y ε Bi ]] ≥ [[∃x(x ε Ai ∧t ((x, y) ε Υ(Aj |=
⇒ Bj )))]] (4.33)
= sup [[x ε Ai ∧t ((x, y) ε Υ(Aj |=
⇒ Bj ))]]
x∈X
has to be the case. Now it is obvious that the condition given in (ii) is
necessary and sufficient for the subset property to hold true.
Of course the existence of some x0 ∈ X such that
[[y ε Bi ]] ≤ [[(x0 ε Ai ∧t ((x0 , y) ε Υ(Ak |=
⇒ Bk )))]] (4.34)
is the case is a sufficient condition for (4.32) to hold true. The crucial
point for the difficulties with necessary conditions for (4.32) is that there
the supremum is on the right hand side of the ≤-condition. But for a finite
universe of discourse X the supremum becomes a maximum and the difficulty
disappears. Hence (i) is proved too. QED
Unfortunately, the conditions we obtained in this proposition are not very

handy ones. Therefore we will look for simpler necessary or sufficient condi-
tions for the superset, as well as the subset condition. Without restriction
to special cases how to choose the “local” fuzzy relations Ri = Υ(Ai |=⇒ Bi )
one has the following corollary.
Corollary 4.2 (i) A necessary condition for the superset property of R is

^
(hgt (Bi ) ≤ hgt (Ai )). (4.35)
i∈I
(ii) A sufficient condition for the superset property of R is

^
⇒ Bi )00 Ai ).
(Bi ⊂ Υ(Ai |= (4.36)
i∈I
(iii) A necessary and sufficient condition for the subset property of R is

^
⇒ Bi )00 Aj ⊂ Bj ).
(Υ(Ai |= (4.37)
i,j∈I
Sometimes it is useful to have not only absolute conditions concerning the

superset and the subset property, but also relative ones. For that purpose
we give the following definition.
Definition 4.3 Consider a control rule Ai |= ⇒ Bi and two of its realizations

Ri1 , Ri2 as a fuzzy relation. Then we call Ri1 a stronger realization (of that
rule) as Ri2 and Ri2 a weaker realization (of that rule) as Ri1 iff Ri1 ⊂ Ri2
holds true. In the same manner, if R1 and R2 are fuzzy relations which
realize a given set (4.1) of control rules for a fuzzy controller, we call R1 a
stronger realization (of (4.1)) as R2 or R2 a weaker realization (of (4.1)) as
R1 iff R1 ⊂ R2 .
Corollary 4.3 Let R1 , R2 be two realizations of a set of control rules (4.1)

as a fuzzy relation.
(i) If R1 is a stronger realization as R2 and R1 has the superset property,
then R2 has the superset property too.
(ii) If R1 is a weaker realization as R2 and R1 has the subset property,
then R2 has the subset property too.
Proof. Both claims are immediate consequences of the monotonicity, propo-

sition 2.18(ii), of the full image of a fuzzy set under a fuzzy relation with
respect to the inclusion of the relations involved and of the two foregoing
definitions. QED
Of course, if the coding relation R of a fuzzy controller which is con-

stituted by a set (4.1) of control rules is built up in Mamdani’s way (4.9),
(4.25) out of the coding relations Ri = Ai ×Bi of the single rules of (4.1), and
if one has two coding procedures for the single rules such that each relation
Ri1 produced by the first coding procedure is a weaker version than the fuzzy
relation Ri2 produced by the second coding procedure, then R1 is a weaker
version than R2 . But what is essential here is not the specific way in which
one “collects” the individual information Ri to get the full coding relation
R for the fuzzy controller, but only a certain monotonicity property again.
Proposition 4.4 Suppose there are two procedures to obtain for the control
rules Ai |=
⇒ Bi of a fuzzy controller with generating family (4.1) their real-
izations Ri1 , Ri2 as fuzzy relations. Suppose additionally that the method by
which the complete coding relation R for a fuzzy controller is constructed out
of the local relations Ri is monotonuous with respect to inclusion. Then R1
is a stronger realization of a set of control rules (4.2) if for each of the rules
⇒ Bi the coding fuzzy relation Ri1 is a stronger realization of that control
Ai |=
rule than the coding relation Ri2 .
Proof. If Ri1 is always a stronger realization than Ri2 , by definition 4.3
that means that Ri1 ⊂ Ri2 is always the case. Therfore, by the monotonic
dependence of each Rk of its constituents Rik , k = 1, 2 one immediately has
that R1 ⊂ R2 . QED
The essential property of monotonic dependence of the coding relation

R of a fuzzy controller from its local constituents, the coding relations Ri of
the single control rules, is not only fulfilled for Mamdani’s method of taking
for R the union of all the Ri = Ai × Bi , but for example also for the method
(4.14) to take as R the intersection of suitable Ri .
In the case of the coding procedures (4.25) to (4.27) for all fuzzy sets
A ∈ IF (X ), B ∈ IF (Y) we have the inclusions
A × B ⊂ (A × B) ∪ (A × Y ) ⊂ (A × Y ) ∪ (X × B) (4.38)
and therefore, if these coding procedures are used within the collecting strat-
egy (4.9), coding procedure (4.25) is the strongest one and coding procedure
(4.27) the weakest one of these three.
Thus, in (4.38) the superset property is transmitted “from left to right”

and the subset property “from right to left”.
To get more concrete results we now will consider as special cases those
where the coding procedure for the single control rules is given by (4.25) to
(4.27). Additionally, in these three cases we suppose that the codes of the
single control rules are collected together via (4.23) to give the fuzzy rela-
tion R which codes the whole fuzzy controller. Remember for the following
considerations that the inequalities which appeared in proposition (4.1) as
necessary and sufficient conditions for the superset or subset property have
on their right sides a reference to a conjunction connective, i.e. to a t-norm.
This t-norm will now always be denoted t without an index. It has to be
distinguished from further t-norms which will be considered and which come
into consideration by the special coding procedures for the control rules. In
order to be not too specific here we intend to allow that in principle the
t-norm t and these other ones are different.
4.2.2 Coding rules by Υ(A |=

⇒ B) = A ×t B
We assume that the coding procedure for the single control rules is simply to
take the cartesian product of input and output of the rule. For that cartesian
product it is supposed that it refers to a t-norm t1 .
Proposition 4.5 Sufficient conditions for R to have the superset property

are:
(i) for each t-norm t:
hgt(Aj ) = 1 for all j ∈ I;
(ii) if t = t1 = min:
hgt(Bj ) ≤ hgt(Aj ) for all j ∈ I.
Proof. By (4.32) we have as a necessary and sufficient condition for the

superset property in the present case that for all i ∈ I and all y ∈ Y there
exists some k ∈ I such that
[[y ε Bi ]]
≤ sup [[(x ε Ai ∧t ((x, y) ε (Ak ×t1 Bk )))]]
x∈X
≤ sup [[x ε Ai ∧t (x ε Ak ∧t1 y ε Bk )]]. (4.39)
x∈X
Now choosing k = i one gets as a sufficient condition

[[y ε Bi ]] ≤ sup [[(x ε Ai ∧t (x ε Ai ∧t1 y ε Bi )]], (4.40)
x∈X
which because of hgt (Ai ) = supx∈X [[x ε Ai ]] and the left continuity of t and
of t1 can be rewritten as
[[y ε Bi ]] ≤ hgt (Ai ) t (hgt (Ai ) t1 [[y ε Bi ]]). (4.41)
Thus it is obvious that condition (i) is a sufficient one for (4.41) to hold true
for all i ∈ I and all y ∈ Y, and hence also for (4.39) to hold true for all i ∈ I
and all y ∈ Y.
Furthermore, if we assume that t = t1 = min is the case, (4.41) can
essentially be simplified and is equivalent to the fact that
[[y ε Bi ]] ≤ min{hgt (Ai ), min{hgt (Ai ), [[y ε Bi ]]}}
≤ min{hgt (Ai ), [[y ε Bi ]]} (4.42)
holds true for all i ∈ I and all y ∈ Y. But this means that (ii) is also a
sufficient condition for R to have the superset property. QED
Perhaps at first sight condition (i) of this proposition appears strong,

but it is quite reasonable in most applications of fuzzy controllers. The rea-
son is simply that usually the typical input values Ai for the control rules
characterize some kinds of “standard situations” for the “true” values of the
input variable. Therefore these input fuzzy sets Ai are usually normal and
quite often even unimodal. In the specific situation with the min-operator
as the only t-norm under consideration, by condition (ii) an even weaker as-
sumption already suffices, which additionally is easy to check and to handle.
Hence, the superset property of a fuzzy controller is quite a weak condition
which is easy to meet in applications.
Proposition 4.6 Sufficient conditions for R to have the subset property are:
(i) without any restriction concerning the t-norm involved:

Ai ∩t Aj = ∅ for all i, j ∈ I, i 6= j;
(ii) if u t (v t1 w) ≤ (u t v) t1 w for all u, v, w, ∈ [0, 1]:
hgt(Bi ) t1 hgt (Ai ∩t Aj ) = 0 for all i ∈ I, i 6= j;
(iii) if t = tD is the drastic product
hgt(Ai ∩ Aj ) < 1 for all i, j ∈ I, i 6= j.

subset property in the present case that for all i, j ∈ I and all y ∈ Y
[[y ε Bi ]]
≥ sup [[x ε Ai ∧t ((x, y) ε (Aj ×t1 Bj ))]]
x∈X
≥ sup [[x ε Ai ∧t (x ε Aj ∧t1 y ε Bj )]]. (4.43)
x∈X
has to be the case. Using the left continuity of the t-norms involved here
one gets as in the last proof as an equivalent condition to (4.43) that for all
i, j ∈ I and all y ∈ Y
[[y ε Bi ]] ≥ hgt (Ai ) ∧t (hgt (Aj ) ∧t1 [[y ε Bj ]]) (4.44)
has to be the case. In the case that i = j inequality (4.44) obviously always
holds true. But for i 6= j the only reasonable way to get a sufficient condition
for (4.44) and hence for (4.43) is to make sure that the right-hand side of
inequality (4.44) is always zero. Because of
hgt (Ai ) ∧t hgt (Aj ) ≥ hgt (Ai ) ∧t (hgt (Aj ) ∧t1 [[y ε Bj ]])
for all i, j ∈ I and all y ∈ Y, the sufficiency of condition (i) immediately
follows.
The inequalities
(hgt (Ai ) ∧t hgt (Aj )) ∧t1 hgt (Bj )
≥ hgt (Ai ) ∧t (hgt (Aj ) ∧t1 hgt (Bj ))
≥ hgt (Ai ) ∧t (hgt (Aj ) ∧t1 [[y ε Bj ]]),
the first one of them a consequence of the specific assumption of (ii), together
with (4.44) prove that (ii) is also a sufficient condition for the subset property
of R.
Finally, condition (iii) is a sufficient one because of the equivalence:
Ai ∩t Aj = ∅ ⇔ hgt (Ai ∩ Aj ) < 1 for t = tD
for all fuzzy sets Ai , Aj , and of the fact that in the case of Ai ∩t Aj = ∅ one
has
[[x ε Ai ∧t (x ε Aj ∧t1 y ε Bj )]] = 0
independent of [[y ε Bj ]]. Here the left continuity is not used in the arguments.
QED
Condition (i) of this proposition is simply the pairwise disjointness of

the family of input fuzzy sets Ai of the fuzzy controller under consideration.
Caused by the possible choice of an interactive t-norm for t this does not
necessarily mean the pairwise disjointness of the supports supp (Ai ) of these
input fuzzy sets, but nevertheless is a quite strong demand. Condition (iii) is
essentially the same condition, only reformulated using specific properties of
the drastic product. Condition (ii), finally, is some weakening of this pairwise
disjointness guided by the height, i.e. the “amount of subnormality” of Bi ,
but does not look very manageable. Hence, the sufficient conditions this last
proposition gives for the subset property are relatively strong ones. And
the chosen way of proof indicates in addition that much simpler sufficient
conditions for the subset property of R seem not – or at least not easily –
available. That means the subset property for R is a much stronger demand
than the superset property.
4.2.3 ⇒ B) = (A ×t1 B) ∪t2 (A ×t1 Y )

Coding rules by Υ(A |=
We now assume that the coding procedure for the single control rules is more
involved. The main idea, however, is near to case 1. The coding there of the
control rules (Ai |=
⇒ Bi ) by the cartesian product Ai ×t1 Bi can be seen as
a kind of coding of a fuzzy implication
if input = Ai then output = Bi . (4.45)
Now the i-th control rule is not read as precisely this fuzzy implication but
instead as the fuzzy implication
if input = Ai then output = Bi else output = anything
with the else-clause furthermore understood as saying
if input = not-Ai then output = anything (4.46)
and being disjunctively appended to (4.45). Then the following conditions

result.

are:
(i) in any case
hgt(Aj ) = 1 for all j ∈ I;

(ii) for each i ∈ I there exists some k ∈ I such that
Ai ≥1 ∩ (X \ supp(Ak )) 6= ∅;
(iii) if t1 = t2 = tL or t1 = t2 = min, then

³ [ ´
hgt(Bi ) ≤ hgt Ai ∩t Aj for all j ∈ I;
j∈I
and assuming additionally that t = min one also gets

(iv) if either t = t1 = t2 = min or t = min together with t1 = t2 = tL ,
then
hgt(Bj ) ≤ hgt(Aj ) for all j ∈ I.

superset property in the present case that for all i ∈ I and all y ∈ Y there
exists some k ∈ I such that

⇒ Bk )))]]
= sup [[x ε Ai ∧t ((x, y) ε (Ak ×t1 Bk ) ∪t2 (Ak ×t1 Y )]]
x∈X
= sup [[x ε Ai ∧t ((x ε Ak ∧t1 y ε Bk ) ∨t2 x ε Ak )]]. (4.47)
x∈X
Using [[x ε Ak ]] ≥ 0 together with the monotonicity of ∨t2 and ∧t and with
condition (S3) in the definition 1.1 of t-conorms, we get as a sufficient con-
dition
[[y ε Bi ]] ≤ sup [[x ε Ai ∧t (x ε Ak ∧t1 y ε Bk )]], (4.48)

x∈X
which is exactly condition (4.39). Thus (i) is also a sufficient condition in

this case.
Furthermore, it is surely sufficient for (4.47) to have for each i ∈ I and
y∈Y
sup [[x ε Ai ∧t ((x ε Ak ∧t1 y ε Bk ) ∨t2 x ε Ak )]] = 1 (4.49)

x∈X
for a suitable k ∈ I. Therefore it is sufficient too to always find x ∈ X and

k ∈ I such that
[[x ε Ai ]] = [[x ε Ak ]] = 1.
But this precisely is condition (ii).

Now suppose that t1 = t2 = tL . Then one has that for all u, v ∈ [0, 1]
the equality
(u&v) ] u = u ∨ v
holds true where u = 1 − u is used. Thus, (4.47) becomes the form: for all
i ∈ I and all y ∈ Y there exists some k ∈ I such that
[[y ε Bi ]] ≤ sup [[x ε Ai ∧t (x ε Ak ∨ y ε Bk )]]

x∈X
= sup [[(x ε Ai ∩t Ak ) ∨ (x ε Ai ∧t y ε Bk )]]
x∈X
= max{sup [[x ε Ai ∩t Ak ]], sup [[x ε Ai ∧t y ε Bk ]]}.
x∈X x∈X
By the left continuity of t this last form can be written as
[[y ε Bi ]] ≤ max{hgt (Ai ∩t Ak ), hgt (Ai ) t [[y ε Bk ]]} (4.50)

and immediately proves the sufficiency of the condition that for each i ∈ I
there exists some k ∈ I such that
hgt (Bi ) ≤ hgt (Ai ∩t Ak ).
But now it is a routine matter to verify the sufficiency of condition (iii) for
the present case.
Assuming furthermore that t = min, (4.50) becomes the equivalent form:
for all i ∈ I and all y ∈ Y there exists some k ∈ I such that
[[y ε Bi ]] ≤ max{hgt (Ai ∩ Ak ), min{hgt (Ai ), [[y ε Bk ]]}}.

Therefore, to get (4.47) true, it now suffices always to have
[[y ε Bi ]] ≤ min{hgt (Ai ), [[y ε Bk ]]} for some k ∈ I, (4.51)

and obviously for (4.51) one now has condition (iv) as a sufficient one.
Finally, assume t1 = t2 = min. In that case (4.47) and (4.50) become:
[[y ε Bi ]] ≤ sup [[x ε Ai ∧t ((x ε Ak ∧ y ε Bk ) ∨ x ε Ak )]]

x∈X
= sup [[(x ε Ai ∧t (x ε Ak ∧ y ε Bk )) ∨ (x ε Ai ∧t x ε Ak )]]
x∈X
= max{hgt (Ai ∩t Ak ), sup [[x ε Ai ∧t (x ε Ak ∧ y ε Bk )]]}
x∈X
To have this condition true it suffices if for all i ∈ I one has
hgt (Bi ) ≤ hgt (Ai ∩t Ak ) for some k ∈ I,
which easily gives condition (iii) for the present case. Again now assuming
furthermore that also t = min is the case yields from (4.47) the condition:
[[y ε Bi ]] ≤ sup [[(x ε Ai ∩ Ak ∧ y ε Bk ) ∨ (x ε Ai ∩ Ak )]]

x∈X
= max{min{hgt (Ai ∩ Ak ), [[y ε Bk ]]}, hgt (Ai ∩ Ak )}.
For this inequality it suffices that for each i ∈ I:
hgt (Bi ) ≤ hgt (Ai ∩ Ak ) for some k ∈ I.
Thus, (iv) is also a sufficient condition in this situation. QED
As in the previous subsection 4.2.2 now the sufficient conditions for the
superset property of R are not hard to meet by a fuzzy controller either,
and in general conditions (i) to (iii) seem to be much more important than
condition (iv) in a normal application. The situation we met in that subsec-
tion 4.2.2, that sufficient conditions for the subset property of R are much
more difficult to find and to have realized, is repeated for the present situa-
tion. Furthermore, we now have to restrict ourselves to some special cases
for the choice of the t-norms t1 , t2 to find results which do not look extremely
unmanageable. But, the following sufficient conditions for the subset prop-
erty are also quite strong: so for example in items (i) as well as (ii) of the
following proposition one has to fulfill two simultaneous demands, the first
of them formulating a kind of “almost pairwise disjointness” of the input
fuzzy sets, and the second one demanding some “considerable overlap” of
the output fuzzy sets. That indicates that the coding procedure we are ac-
tually discussing for the single control rules is not well suited for the subset
property of R.
Proposition 4.8 Sufficient conditions for R to have the subset property are:
(i) if one assumes t1 = t2 = tL , then
1 − hgt(Bi ) ≥ hgt(Ai ∩t Aj ) for all i, j ∈ I

together with
Y [hgt(Ai )] ∩t Bj ⊂ Bi for all i, j ∈ I;
(ii) if one assumes t = t1 = t2 = min, then
1 − hgt(Bi ) ≥ hgt(Ai ∩ Aj ) for all i, j ∈ I
together with
Y [hgt(Ai ∩Aj )] ∩ Bj ⊂ Bi for all i, j ∈ I.
Proof. Without any restrictions concerning the t-norms involved one has
by (4.33) as a necessary and sufficient condition for the subset property of
R that for all i, j ∈ I and all y ∈ Y
[[y ε Bi ]] ≥ [[∃x(x ε Ai ∧t ((x, y) ε Υ(Aj |=
⇒ Bj )))]]
= sup [[x ε Ai ∧t ((x, y) ε (Aj ×t1 Bj ) ∪t2 (Aj ×t1 Y )]]
x∈X
= sup [[x ε Ai ∧t ((x ε Aj ∧t1 y ε Bj ) ∨t2 x ε Aj )]]. (4.52)
x∈X
For t1 = t2 = tL this becomes as in the last proof

[[y ε Bi ]] ≥ sup [[x ε Ai ∧t (x ε Aj ∨ y ε Bj )]]
x∈X
= sup [[(x ε Ai ∩t Aj ) ∨ (x ε Ai ∧t y ε Bj )]]
x∈X
= max{hgt (Ai ∩t Aj ), hgt (Ai ) ∧t [[y ε Bj ]]}. (4.53)
Thus a sufficient condition results if both values, the maximum of which is
taken in (4.53), separately are no greater than the left-hand side [[y ε Bi ]]. For
hgt (Ai ∩t Aj ) this worst case scenario demands that one has 1 − hgt (Bi ) ≥
hgt (Ai ∩t Aj ) for all i, j ∈ I. And in the other case the demand reads
[[y ε Bi ]] ≥ hgt (Ai ) ∧t [[y ε Bj ]], which holds true for Y [hgt (Ai )] ∩t Bj ⊂ Bi .
Thus we have proven case (i).
Starting from t = t1 = t2 = min one gets the condition that for all
[[y ε Bi ]] ≥ sup [[(x ε Ai ∩ Aj ) ∨ (x ε Ai ∩ Aj ∧ y ε Bj )]]
x∈X
= max{hgt (Ai ∩ Aj ), min{hgt (Ai ∩ Aj ), [[y ε Bj ]]}}.
The same arguments used to infer conditions (i) from (4.53) now give con-
ditions (ii). QED
4.2.4 Coding rules by Υ(A |=

⇒ B) = (A ×t1 Y ) ∪t2 (X ×t1 B)
The coding procedure for the single control rules now can be understood as
reading the fuzzy implication (4.45) as
either input = A or output = B
which is closely analogous to the standard way of classical propositional logic
to express implication by disjunction and negation.

are:
(i) without restrictions concerning the t-norms involved:
hgt(Ai ) = 1 for all i ∈ I;
(ii) in the case that the t-norm t is distributive over the t-conorm st2 ,
thus for example in the case that t2 = min,
[
hgt(Bi ) ≤ hgt(Ai ∩t Aj ) for all i ∈ I;
j∈I
(iii) in the case that t = t2 = tL : for each i ∈ I there exists some k ∈ I

such that
Ai ≥hgt(Bi ) ∩ (Ai ∩tL Aj )≥hgt(Bi ) 6= ∅.
Proof. By (4.32) we also have as a necessary and sufficient condition for

the superset property in the present case that for all i ∈ I and all y ∈ Y
there exists some k ∈ I such that
⇒ Bk )))]]
= sup [[x ε Ai ∧t ((x, y) ε (Ak ×t1 Y ) ∪t2 (X ×t1 Bk )]]
x∈X
= sup [[x ε Ai ∧t (x ε Ak ∨t2 y ε Bk )]]. (4.54)
x∈X
As a side remark let us mention that the t-norm t1 does not appear any
more in the final formula (4.54), hence the choice of t1 is inessential for the
present coding procedure of the single control rules.
Using in (4.54) the monotonicity of st2 together with [[x ε Ak ]] ≥ 0 gives
as a sufficient condition that for all i ∈ I one has
[[y ε Bi ]] ≤ sup [[x ε Ai ∧t y ε Bk ]]
x∈X
= hgt (Ai ) t [[y ε Bk ]]
for some k ∈ I. Hence (i) is a sufficient condition.

With the distributivity assumption of (ii) one gets from (4.54) the con-
dition that for all i ∈ I and all y ∈ Y there exists some k ∈ I such that
[[y ε Bi ]] ≤ sup [[(x ε Ai ∧t x ε Ak ) ∨t2 (x ε Ai ∧t y ε Bk )]]

x∈X
= sup [[(x ε Ai ∩t Ak ) ∨t2 . . .)]]
x∈X
and thus immediately condition (ii) because of
hgt (Ai ∩t Ak ) ≤ sup [[(x ε Ai ∩t Ak ) ∨t2 . . .)]]

x∈X
for all i, k ∈ I.
And in the case that t = t2 = tL , using directly the definitions of tL , stL
given in Tables 1.1 and 1.2 one has from (4.54) the condition that for all
i ∈ I and all y ∈ Y there exists some k ∈ I such that
[[y ε Bi ]] ≤ sup max{0, [[x ε Ai ]] + min{[[x ε Ak ]] + [[y ε Bk ]], 1} − 1}

x∈X
= sup max{0, min{[[x ε Ai ]] + [[x ε Ak ]] − 1 + [[y ε Bk ]], [[x ε Ai ]]}}
x∈X
= max{0, sup min{[[x ε Ai & x ε Ak ]] + [[y ε Bk ]], [[x ε Ai ]]}}. (4.55)
x∈X
To have this condition true it is surely sufficient that the final term of (4.55)
is no smaller than hgt (Bi ). And condition (iii) of our proposition yields just
this. QED
Proposition 4.10 Sufficient conditions for R to have the subset property

are:
(i) if one assumes t2 = min then
1 − hgt(Bi ) ≥ hgt(Ai ∩t Aj ) for all i, j ∈ I
together with
Y [hgt(Ai )] ∩t Bj ⊂ Bi for all i, j ∈ I;
(ii) if one assumes t = t2 = tL then
1 − hgt(Bi ) ≥ hgt(Ai ) for all i, j ∈ I.

4.3. MANIPULATION OF FUZZY DATA 159
Proof. Without any restrictions concerning the t-norms involved one now
has by (4.33) as a necessary and sufficient condition for the subset property
of R that for all i, j ∈ I and all y ∈ Y
[[y ε Bi ]] ≥ sup [[x ε Ai ∧t (x ε Aj ∨t2 y ε Bj )]]. (4.56)
x∈X
Starting from (4.56) and using t2 = min one gets the condition that for
all i, j ∈ I and all y ∈ Y
[[y ε Bi ]] ≥ sup [[(x ε Ai ∩t Aj ) ∨ (x ε Ai ∧t y ε Bj )]]
x∈X
= max{hgt (Ai ∩t Aj ), hgt (Ai ) t [[y ε Bj ]]}.
To get a sufficient condition out of these inequalities we again take the worst
case scenario for the left hand side and try to have the last-mentioned right-
hand side ≤ 1 − hgt (Bi ), i.e. to have both terms, the maximum is taken
over there, to be ≤ 1 − hgt (Bi ). This immediately gives both parts of (i).
Finally, in the case that t = t2 = tL the same calculations as in the last
proof give, from (4.56), the necessary and sufficient condition that for all
[[y ε Bi ]] ≥ max{0, sup min{[[x ε Ai & x ε Ak ]] + [[y ε Bk ]], [[x ε Ai ]]}}.
x∈X
This obviously is equivalent to the simpler condition that for all i, j ∈ I and
all y ∈ Y one has
[[y ε Bi ]] ≥ min{[[x ε Ai & x ε Aj ]] + [[y ε Bj ]], [[x ε Ai ]]}.
The same worst case scenario as just before, now applied with respect to the
second term within the min-operator, immediately gives (ii). QED
4.3 Solvability degrees and the manipulation of

fuzzy data
Let us start with the idea that the fuzzy data we are considering for the
input and output values in our fuzzy model themselves to some extent are
given only approximately. This, of course, may be caused by quite different
reasons. One such possibility, sometimes realized in engineering practice, is
that one has a relatively clear idea of higher membership degrees, but much
less information concerning small membership degrees. Another possibility
is that some noise – with small values – makes the smaller membership
degrees (especially) uncertain. A third possibility may be realized in case
one first has a linguistic model of some process and then has to “translate”
the linguistic values into fuzzy sets: in this case one also has some degree of
freedom for the actual choice of these fuzzy inputs and outputs.
One of the types of approaches often considered in the case of such a kind
of uncertainty concerning the fuzzy “data” is to change from usual fuzzy sets
to fuzzy sets of type 2 or perhaps only to interval valued fuzzy sets. We shall
not follow this strategy.
Instead here we assume that some threshold level α is given such that
membership degrees smaller than α are less reliable than those greater than
this threshold α.
To discuss the influence of such a threshold level on the existence of
solutions of fuzzy relation equations and of systems of such equations let us
first introduce some additional notation.
Let α ∈ [0, 1] be a threshold level. For a given fuzzy set A ∈ IF (X ) define
Aα =def A ∪ X [α] , (4.57)
i.e. put for the membership degrees
½
µA (x) if µA (x) ≥ α
µAα (x) = max{µA (x), α} =
α if µA (x) < α,
and correspondingly define
[
Aα =def A ∩ hhaii1 , (4.58)
a∈A≥α
which means for the membership degrees

½
µA (x) if µA (x) ≥ α
µAα (x) =
0 if µA (x) < α.
Both fuzzy sets Aα , Aα are borderline cases of those fuzzy sets which cor-
respond to the given fuzzy set A according to the threshold level α and its
interpretation.
In the following, we use the threshold level to change our fuzzy relation
equations. Instead of equations of the type
A ◦t R = B (4.59)
we consider equations of the kinds
Aα ◦t R = B α and Aα ◦t R = Bα . (4.60)
In the same way we change the systems (4.28) of fuzzy relation equations.
To discuss the change in solvability behaviour connected with these changes
of fuzzy relation equations we consider the corresponding solvability degrees.
Before coming back to systems of fuzzy relation equations let us first
consider the case that we will – according to our discussions related to the
threshold level – change the input and output data A, B of fuzzy relation
equations (4.59) to Aα , B α or to Aα , Bα . Let us denote the corresponding
solvability degrees by ξ(α) and by ξ 0 (α), i.e. according to proposition 3.21
we consider
ξ(α) =def hgt (B α ) ϕt hgt (Aα ), (4.61)

0
ξ (α) =def hgt (Bα ) ϕt hgt (Aα ). (4.62)
Our problem will be to consider the relations of the threshold depending

relative solvability degrees ξ(α), ξ 0 (α) to the “absolute” solvability degree ξ0 ,
i.e. to consider the change in the solvability behaviour of (4.59) caused by
the changes of input and output data motivated by the threshold level α.
Proposition 4.11 Let t be a continuous t-norm and α, β threshold levels.

Then for the corresponding solvability degrees of equation (4.59) there hold
true
(i) α ≤ ξ(α) ≤ 1,
(ii) ξ0 ≤ ξ(α),
(iii) ξ(0) = ξ0 and ξ(1) = 1,
(iv) α ≤ β ⇒ ξ(α) ≤ ξ(β).
Proof. (i) Obviously ξ(α) ≤ 1. And the worst case, i.e. smallest value of
ξ(α) is given, according to (4.61), in the case that hgt (B α ) = 1 and, because
of hgt (Aα ) ≥ α, that hgt (Aα ) = α. But then ξ(α) = 1 ϕt α = α results.
(ii) Immediately one has hgt (Aα ) = max{hgt (A), α} and analogously
for B α . Hence
hgt (B) ≤ hgt (A) ⇒ hgt (B α ) ≤ hgt (Aα ),

hgt (B), hgt (A) ≤ α ⇒ hgt (B α ) = hgt (Aα ),
i.e. ξ(α) = 1 in both cases, and finally
hgt (B) > max{hgt (A), α} ⇒

hgt (B α ) = hgt (B) and hgt (Aα ) ≥ hgt (A)
such that ξ0 ≤ ξ(α) in this case, by monotonicity of ϕt in the second argu-

ment. Hence ξ0 ≤ ξ(α) in general.
(iii) Of course, A0 = A and B 0 = B, thus ξ(0) = ξ0 . And if we consider
the 1-universal fuzzy subset X [1] of the universe of discourse X whose mem-
bership degree is =1 in each point of that universe, then A1 = X [1] , as well
as B 1 = Y [1] , and hence ξ(1) = 1 because of X [1] ◦t (X [1] ×t Y [1] ) = Y [1] and
the fact that X [1] ×t Y [1] is the 1-universal fuzzy subset of the universe of
discourse X × Y.
(iv) Interpret Aα , B α as the original input and output data for the fuzzy
relation equation (4.59) which will be modified according to a threshold level
β. Then from α ≤ β one gets (Aα )β = Aβ and (B α )β = B β . Now apply (ii).
QED
Thus we know that the solvability behaviour of equation (4.59) becomes

better and better, i.e. the solvability degree becomes greater and greater
with growing threshold level α.
From an intuitive point of view it seems that the idea connected with
the introduction of threshold level α may not only be realized by the data
transformation
A 7→ Aα and B 7→ B α (4.63)
but equally well by the corresponding data transformation
A 7→ Aα and B 7→ Bα (4.64)
However, this is not the case. For example the solvability degree ξ 0 (α) from
(4.62) has not the monotonicity property analogous to proposition 4.11 (iv).
To show this, we consider a concrete example for two different kinds of t-
norms.
Let us consider the following two cases and assume that A, B are fuzzy
singletons with
hgt (A) = 0.7 and hgt (B) = 0.9.
Case 1. Take as t-norm t = tG = min with the corresponding Φ-operator

ϕG satisfying according to Table 1.3
½
1, if u ≤ v
u ϕG v = (4.65)
v, if u > v.
Case 2. Take as t-norm t = tL = & as given in Table 1.1 with the corre-
sponding Φ-operator ϕL satisfying according to Table 1.3
u ϕL v = min{1, 1 − u + v},
which is the well-known L

Ã ukasiewicz implication operator of many-
valued logic.
The values of ξ 0 (α) for different threshold levels are collected in the fol-
lowing Table 4.1.
Threshold Case 1 Case 2

α ≤ 0.7 ξ 0 (α) = 0.7 ξ 0 (α) = 0.8
0.7 < α ≤ 0.9 ξ 0 (α) =0 ξ 0 (α) = 0.1
α > 0.9 ξ 0 (α) = 1 ξ 0 (α) = 1
Table 4.1: Relative solvability degrees with respect to the data

transformation (4.64)
Of course, smaller t-norms than t = & will give greater values of ξ 0 (α),
but the qualitative behaviour will be still the same: a much smaller value
of ξ 0 (α) for 0.7 < α ≤ 0.9 as for α ≤ 0.7 and α > 0.9. And it is this effect
that makes (for the author’s feeling) the relative solvability degree ξ 0 (α)
inappropriate for the present purposes.
Therefore, for systems (4.28) of fuzzy relation equations we will discuss
only the data transformations of kind (4.63) and the corresponding changes
ξ 7→ ξ(α) of solvability degrees.
Finally, let us discuss the influence that has a data transformation à la
(4.63) for the solvability degree of system (4.28). We denote by ξ(α) now
the solvability degree of the system
Aαi ◦t R = Biα , i = 1, . . . , n,
i.e. the (relative) solvability degree of the “transformed system” of (4.28)

with respect to the threshold level α. Unfortunately, again, the lack of a
direct characterization of the systems solvability degree ξ allows us to prove
only some of the properties that correspond to the results of the previous
proposition 4.11.
Proposition 4.12 Let t be a continuous t-norm, α any threshold level.

Then we have for the solvability degrees of the modified systems (4.28) of
fuzzy equations:
(i) ξ(0) = 0 and ξ(1) = 1,

n
(ii) T (α t α) ≤ ξ(α) ≤ 1,
i=1
n
(iii) if all input fuzzy sets are normal then T α ≤ ξ(α).
i=1
Proof. (i) and ξ(α) ≤ 1 in (ii) are obvious. From proposition 3.24 we get
n ^ n
ξ(α) ≥ T (1 ϕt (α t α)) = T (α t α)
i = 1 y∈Y i=1
which completes (ii). From proposition 3.24 we also get
n ^ n
ξ(α) ≥ T (1 ϕt (1 t α)) = Tα
i = 1 y∈Y i=1
supposing the extra assumption of (iii). Thus also (iii). QED
Again, the case t = ∧ = min with its direct characterization of the

solvability degree allows us to prove more.
Proposition 4.13 Let t = ∧ and α, β be any threshold levels. Then there

hold true
(i) α ≤ ξ(α) ≤ 1,
(ii) ξ ≤ ξ(α),
(iii) α ≤ β ⇒ ξ(α) ≤ ξ(β).
Proof. (i) comes directly from proposition 4.12 (ii).

For (ii) remember formula (3.24). As a first remark note that always
µAj (x) ϕG µBj (y) ≤ µAαj (x) ϕG µBjα (y) (4.66)

because for µAj (x) ≤ α we have the value 1 on the right hand side, and in
case µAj (x) > α we have (4.66) because of µBj (y) ≤ µBjα (y). Furthermore
because of α ≤ µBjα (y) and (4.65): α ≤ µAαj (x) ϕG µBjα (y). Hence
_ n ^ o
min µAi (x), (µAj (x) ϕG µBj (y))
x∈X 1≤j≤n
_ n ^ o
≤ min µAαi (x), (µAαj (x) ϕG µBjα (y))
x∈X 1≤j≤n
and also
_ n ^ o
α≤ min µAαi (x), (µAαj (x) ϕG µBjα (y)) .
x∈X 1≤j≤n
Thus by the same arguments we used to establish (4.66) now we get for each
index i
_ n ^ o
µBi (y) ϕG min µAi (x), (µAj (x) ϕG µBj (y))
x∈X 1≤j≤n
_ n ^ o
≤ µBi (y) ϕG min µAαi (x), (µAαj (x) ϕG µBjα (y))
x∈X 1≤j≤n
and from this (ii).

Finally, (iii) is a consequence of (ii) in the same way as in proposition
4.11 item (iv) was a consequence of item (ii) there: simply use the same
argument. QED
For “small” values of the threshold level α one can assume that the data
transformation (4.63) – as well as the data transformation (4.64) – preserves
the most significant parts of the shapes of the membership functions, because
only membership degrees lower than α are changed. Therefore, in a calcula-
tion of a nonfuzzy value for the control output y0 ∈ Y, for example according
to the method of the centre of gravity as introduced in Mamdani (1976), one
should neglect all the values of the membership functions Biα smaller than
α, i.e. one should take in the case of an output value B := R00 A = A ◦ R
with respect to a concrete input value A:
X . X
y0 = y · µB (y) µB (y) (4.67)
{y∈Y|µB (y)>α} {y∈Y|µB (y)>α}
in the case of finite universes of discourse X , Y, and one should – together

with the assumptions of the existence of the integrals to be considered – take
Z . Z
y0 = y · µB (y) dy µB (y) dy. (4.68)
{y∈Y|µB (y)>α} {y∈Y|µB (y)>α}
for continuous universes of discourse X , Y.

The method for the handling of fuzzy data we just have discussed, shall
now be illustrated by a numerical example of an industrial-oriented system
from Mamdani/Assilian (1975). Here we sketch only the background of
the technical system, referring for details to the original paper.
The linguistic rules describe a strategy of control for a steam engine. The
number of control rules is 15, while the knowledge base contains statements
of the following type:
if input = Ai (= Ei × CEi )
then output = Bi , i = 1, . . . , 15. (4.69)
Here in the i-th rule the input value Ai = Ei × CEi is the fuzzy cartesian
product of the fuzzy description of the error (Ei ) and the change of error
(CEi ) with respect to the intended state of the steam engine, that means
µAi (x, y) = µEi (x) ∧ µCEi (y),
and Bi is the fuzzy set of control of that i-th rule.

From the rules (4.69) the fuzzy relation R is computed according to the
Mamdani strategy, i.e. according to formula (4.23), and therefore one has
15
[ 15
[
R= (Ai × Bi ) = ((Ei × CEi ) × Bi ).
i=1 i=1
We consider only the t-norm t = min and get with the inference mecha-
nism given by (Ei × CEi ) ◦ R for the system of fuzzy relation equations
corresponding to (4.69) the solvability degree
15
^
ξ(α) = ξi (α)
i=1
where ξi (α) denotes the solvability degree of the fuzzy relation equation for
the i-th rule. The results are shown in Figure 4.1. Thereby α takes discrete
values varying from 0.0 to 1.0 with a discretization step of 0.1.
ξ(α)
6
1.0 pprp p p p p p p prp p p p p p p p rp p p p p p p p rp p p p p p p p prp p p p p p p p rp p p p p p p p prp p p p p p p r
pp p
pp
pp p
pp
pp p
p
0.5 pp
pp p
pp
pp prp
ppp pp
ppp p prp p
prp p p - α
0.5 1.0
Figure 4.1: Solvability degree ξ(α) of the Mamdani/Assilian-system

(4.69) as a function of the threshold level α.
Note that the value of the threshold level equal to 0.3 still yields the value
1 for the solvability index ξ(0.3). Moving down with threshold α we obtain
lower values of this solvability index for the entire set of control rules. But
the information which the (relative) solvability degree of the whole system
provides is quite global. What is lacking with this degree ξ(α) is information
about the detailed behaviour of the single rules. Thus one also has to look
for the (relative) solvability degrees of the single rules. Below in Table 4.2
we summarize for the Mamdani/Assilian-system the rules which have the
lowest values of the solvability index ξi (α).
Of course, decreasing values of α produce an ever-extending list of control
rules that shows lower values of their degrees of solvability, in the present
example especially the rules no. 6, 12, 15 – numbered according to the details
in Mamdani/Assilian (1975) – are critical with respect to their (relative)
solvability degrees.
Low values of the solvability degrees of some control rules are confirmed
by another construction that evaluates their credibility in sense of a certainty
factor CF , cf. Pedrycz (1985). Then in this example for the 12th control
rule this certainty factor has the value CF12 = 0. The rules with numbers
6, 13, 14, 15 have values of the certainty factor CF which are below 0.5.
Indeed, in the above construction for α = 0.2, one correspondingly has a
number of control rule

α
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0.2 .2 .2 .2
0.1 .1 .9 .1 .1 .1 .1
0 .5 0 .5 .9 0 0 0 0
Table 4.2: Relative solvability degrees ξi (α) of the

Mamdani/Assilian-controller (4.69)10
subset of three control rules: no. 6, 12, 15, for which the values ξi (0.2) do
not exceed 0.2.
For the threshold value α0 = 0.3 some of the output fuzzy sets Biα , i.e.
of the fuzzy control advices, have been displayed in Figure 4.2.
µ(ui )
1.0 6 bp r
pp B1 , B1α B11 , B11 α ¢
pp ¢
pp ¢
bp ¢r
pp ¤
pp ¤
pp ¤
pp ¤
pp ¤
0.5 pp
p ¤
bp p r¤
pp p p p ¡£
r r r ppp p p rbp p p p p p prbp p p p p p p rpbp p p p p p pbrp p p p p p p brp p p p p p p pbrp p p p p p p rpbp p p p p p prbp p p p p p p rb¡p p p p p£p p pbp p p p p p p bp p p p p p b
pp
pp £
pp £
pp
bp p r£
ppp ¡
p
r r r r rbp p p p p p p prbp p p p p p p brp p p p p p p pbrp p p p p p p rbp p p p p p p prbp p p p p p p ¡ bp p p p p p p pbp p p p p p p bp p p p p p p pbp p p p p p p b
u1 u3 u5 u7 u9 u11 u13 u15
Figure 4.2: Fuzzy output sets Bi and Biα for i = 1, 11 and α = 0.3
The value α0 = 0.3 is low enough such that the changes Ai 7→ A0.3 i and
0.3
Bi 7→ Bi , which according to our method will yield a completely solvable
system of relation equations, i.e. a non-interactive fuzzy controller out of the
10
Each empty entry means the relative solvability degree =1.
system (4.69) of control rules, will not hide the structure of the relation of
the model. This structure can essentially be seen in the output fuzzy sets
Bi – and Figure 4.2 explains that the modified outputs Bi0.3 also present the
essentials of that structure.
The approach towards the manipulation of fuzzy data that we have con-
sidered in this section enables us to use the formal apparatus from fuzzy
relational equations in an interesting way for the evaluation of fuzzy models.
It is remarkable to underline the fact that two mechanisms, the mechanism
for combining the fuzzy pieces of evidence and that one which plays with
inference with the use of fuzzy information, should be discussed simultane-
ously, and they are significantly associated. The concrete way of combining
fuzzy data leads to a certain way for inferring fuzzy consequents. These
facts result from a direct association of the composition operator ◦t and the
corresponding Φ-operator. The first is used in the implementation of the
inference schema, while the latter is joined with the particularization of a
proper way of combining the fuzzy premises. This way of thinking gives an
alternative in comparison to commonly proposed schemata which come from
the compositional rule of inference as stated by Zadeh.
Moreover the introduction of the solvability degree of the system of fuzzy
relational equations makes it possible to express the property of “easiness”
of solving that system. It provides an opportunity to look at appropriate
solutions with regard to the level of threshold. We have discussed a way
of changing the data set that makes it possible to render the set of fuzzy
equations solvable, and moreover to measure the deformation of fuzzy data
used in this process.
One additional remark of a general nature. A global identification pro-
cedure, despite the nature of the factor of uncertainty taken into account,
consists of three steps:
(a) the selection of the structure of the model proposed (i. e. inputs,
outputs, order of the model, etc.);
(b) the estimation of model parameters;
(c) the test of the model with respect to the data set provided (usually it
is the same one as applied in the second step).
For comparison, let us look at the identification task in the presence of ran-
domness. Then the model is constructed with the help of statistical methods.
The first step relies on a choice of a “reasonable” type of function express-
ing a dependence between input and output variables. Further, unknown
parameters of the model (the function) are estimated. Note, moreover, that
at this stage the values of the parameters are computed and a level of their
uncertainty is also expressed via relevant confidence intervals. Hence, even
if in the real world situation the parameters of a model are nonrandom, the
estimation procedure involves that they become viewed as random variables.
Finally, a consistent degree of the model is commonly performed by means of
F-statistics. And if this stage provides a negative result, the entire procedure
is repeated, changing the structure of the model.
The identification procedure which (at least implicitely) is proposed with
the ideas of this section fits this general schema. The structure of a fuzzy
model – of the type of a fuzzy controller – is established selecting the form
of the fuzzy relation equations, viz. the type of the composition operation
together with the t-norm involved in it, i.e. is established with the control
rules and their translations. The parameter estimation for the model then
proceeds through our methods for discussing the solvability of (systems of)
fuzzy relation equations. Here, the solvability degrees serve as indices indi-
cating how credible the structure of the fuzzy relation is which constitutes
the fuzzy controller and hence how good the fuzzy model is. Additionally,
in the methodological perspective this (relative) solvability degree has an
analogous role like a confidence interval: if ξ(α) is lower than a borderline
value ξ ∗ (α) for a certain α – then repeat the identification procedure, i.e.
“adjust” the fuzzy model.
The range of the applicability of this specialized discussion via relative
solvability degrees is closely related to those systems where a tool for ma-
nipulating fuzzy data is needed, e.g. expert systems, robotics, and pattern
recognition.
Chapter 5
Some methodological issues

of fuzzy modelling
5.1 The comparison of fuzzy sets and the inverse

problem
Describing fuzzy models by their input-output behaviour is a very conve-
nient tool in fuzzy modelling. This approach immediately leads to fuzzy
relational equations – usually to systems of them – and to the necessity to
solve them. Solving such systems of equation is, as the topics discussed in
the previous chapters 3 and 4 indicate, a complicated task, not only for nu-
merical reasons but also because of the fact that in the most essential cases
in general only approximate solutions exist. Therefore, a detailed discussion
of the solvability behaviour of systems of fuzzy relational equations and of
factors which influence or change this behaviour is highly interesting, not
only from a theoretical perspective but for very practical reasons of fuzzy
model building too.
The choice of the fuzzy operators used in the equations, i.e. the choice of
the t-norms in the definitions of R00 A and R ◦t S, is one point for discussion.
Another is the change from sup-min composition or inf-max composition,
which are used within the method of getting the fuzzy output values out of
the fuzzy input values, i.e. are used in the compositional rule of inference, to
other, perhaps more complicated rules like those ones in (3.37) and (3.38)
or even some convex combination of such simple kinds of compositions as
carried out in the work of Ohsato/Sekiguchi (1983, 1985) and discussed
in section 3.6.
172 CHAPTER 5. METHODOLOGICAL ISSUES
In all these cases extremal solutions of systems are – given the solvability
of the system under consideration – available from corresponding extremal
solutions of the single equations. Additionally, in interesting special cases of
the non-solvability of such systems, these constructions give the best possi-
ble approximate solutions – and in general at least nearly the best possible
ones. Hence, a “brute-force” methodology can be taken into account which
starts by considering such a suitable combination of solutions to the single
equations of the system as an a priori approximate solution of the whole
system. If doing so, it becomes essential to discuss the quality of an ap-
proximate solution. Such a discussion can be provided by using global, but
also by using local solvability (or rather, solution) indices. Such indices can
also be used to define analogues to confidence intervals (with respect to each
point of the corresponding universe of discourse) for the actual approximate
solution. Such an approach provides means for evaluating the quality of the
actual approximate solution at hand, of the fuzzy model, and hence also for
ways to look for improved models.
Therefore we now first reconsider the problem of the comparison of two
fuzzy sets. Let C, D be fuzzy sets defined over the same universe of discourse
X , i.e. C, D : X → [0, 1]. The following question here is the essential one:
to what extent C, D are the same fuzzy set? Or to put it into other words:
how distinct are the fuzzy sets C, D?
At the beginning let us recall some main fashions of comparison that
have previously been discussed.
(A) The most direct approach is to measure some kind of distance of the
membership functions of C and D, i.e. to calculate a distance value d(C, D),
cf. Kaufmann (1975) as well as our distances %t of definition 3.5. In general,
some Minkowski-type distance is quite often employed:
³Z ´1/p
d(C, D) = |µC (x) − µD (x)|p dx , p ≥ 1,
X
where we assume the existence of all integrals appearing (which may be
guaranteed by restrictions on the types of the membership functions) or
have, for finite universes of discourse X , to read them as sums.
As special cases for p = 1 one has the Hamming distance, and p = 2
yields the usual Euclidean distance.
Other distances which seem to be equally well suited for fuzzy sets are
Čebyšev-type distances of the form
³ ´1/p
d(C, D) = sup |µC (x) − µD (x)|p , p ≥ 1,
X
5.1. COMPARISON OF FUZZY SETS 173
with the well-known Čebyšev metric as the special case p = 1.

If the specified distance is equal to zero, we may consider C, D as indis-
tinguishable but, of course, for another choice of a distance function this dis-
tance may differ from zero. Note that this way of expressing equality comes
almost directly from the well-known concepts of mathematical analysis, and
hence does not produce any difficulties in understanding its mathematical
background. On the other hand, from the applicational point of view the
choice of a proper form of this distance is not obvious, and this requires
deeper investigation – usually not of a purely mathematical character.
(B) A second way of comparison of two fuzzy sets is closely related to

the so-called possibility measure proposed by Zadeh (1978). Two indices
have been put into discussion:
– possibility of C with respect to D, Poss(C|D), given by

_
Poss(C|D) = (µC (x) t µD (x))
x∈X
= [[∃x(x ε C ∧t x ε D)]] = hgt (C ∩t D), (5.1)
– certainty of C with respect to D, Cert(C|D), given by

^
Cert(C|D) = (µC (x) st (1 − µD (x)))
x∈X
= [[∀x(¬x ε D ∨t x ε C)]] (5.2)
W V
with t any t-norm, st the corresponding t-conorm, for supremum, and
for infimum. As a side remark let us mention that in the original formulation
Zadeh specified the t-norm as t = min, and the t-conorm st as maximum.
The possibility of C with respect to D expresses a degree of overlapping
of C and D, whereas Cert(C|D) is connected with measuring a degree of
containment of D in C; but Cert(C|D) is not the inclusion degree [[D ⊂ =t C]]
because in (5.2) instead of the implication p →t q, crucial for defining the
fuzzified inclusion ⊂
=t , a (classically and also for t = & = tL – but not in
general – equivalent) combination ¬p ∨t q is used.
Remembering that t-norms and their t-conorms are conjugated by the
formula (1.9), one obtains
Cert(C|D) = 1 − Poss(C
CC|D) = 1 − hgt (CCC ∩ D).
Now the equality of two fuzzy sets C and D may be evaluated by taking the
minimum of degrees Cert(D|C) and Cert(C|D); this index of comparison
has been discussed in correspondence to some aspects of matching of fuzzy
data.
(C) A third approach for the comparison of fuzzy sets uses our previous
more set theoretically oriented tools, especially the fuzzified identity ≡t
of definition 2.3. Modelling the involved connectives for implication and
conjunction in our fuzzy set theoretic setting via any residuation operator,
i.e. any Φ-operator ϕ for implication and any (left) continuous t-norm t for
conjunction, in addition to our previous equality degree [[C ≡t D]] also some
local degree to which two fuzzy sets C and D are equal each other at a point
a ∈ X is given as a number kC = Dk(a) defined e.g. in analogy with our
definition 2.3 of ≡t , but deleting the universal quantifier ∀ there, i.e. the
inf-operator. That means we take this local degree of equality as the value
kC = Dk(a) =def (µC (a) ϕt µD (a)) t (µD (a) ϕt µC (a)) (5.3)
which is nothing other then the truth degree
kC = Dk(a) = [[a ε C ↔t a ε D]].
The simple fact that always µC (a) ≤ µD (a) or µD (a) ≤ µC (a) is the case, i.e.
that always µC (a) ϕt µD (a) = 1 or µD (a) ϕt µC (a) = 1, allows to simplify
(5.3) to
kC = Dk(a) = (µC (a) ϕt µD (a)) ∧ (µD (a) ϕt µC (a)). (5.4)
Thus one has, together with the Φ-operator ϕt , only to take the min-
operator instead of the t-norm t to finally find kC = Dk(a).
Going one step further, nevertheless one usually also likes to have a
number expressing a (unique) degree to which C and D are equal to each
other in a global sense. For this one has to aggregate the partial evaluations
(5.3) and (5.4) of equality over the whole space X .
There is, however, no unique way to perform these tasks. Perhaps the
most preferred ways in the engineering community are to take either a so-
called optimistic or a pessimistic form of aggregation.
In the optimistic case one prefers to modelize the degree of the statement
“C and D are equal to each other” by the maximal value of kC = Dk(x)
over X , viz.
kC = Dk =def sup kC = Dk(x). (5.5)
x∈X
5.1. COMPARISON OF FUZZY SETS 175
Note that the sup-operator here corresponds to the existential quantifier ∃;

so one essentially takes only one selected element in X in which C and D
are locally equal at the highest degree. It seems reasonable to look at this
“optimistic” global degree as indicating some degree of possibility that C, D
may be equal.
On the opposite, i.e. pessimistic pole, one often uses the formula
kC = Dk =def inf kC = Dk(x). (5.6)

x∈X
This degree of equality has a “pessimistic character” because obviously this

degree (5.6) indicates the worst case of all the local degrees (5.3). Here as al-
ways the inf-operator corresponds to the universal quantifier ∀ and therefore
(5.6) can be rewritten as
kC = Dk = [[∀x(x ε C ↔t x ε D)]].
Therefore according to proposition 1.26 (i) this means
kC = Dk = [[∀x((x ε C →t x ε D) ∧t (x ε D →t x ε C))]]
≥ [[∀x(x ε C →t x ε D) ∧t ∀x(x ε D →t x ε C))]]
= [[C ⊂ ⊂
=t D ∧t D =t C]] = [[C ≡t D]]
and thus is closely analogous to our previous equality degree.
Generalized quantifiers in the sense of the standard model theory of
mathematical logic like “almost everywhere” instead of these classical “some-
times” or “always” versions, or even fuzzy quantifiers like “most”, “not too
few” applied to the localized equality values of C, D would give global evalu-
ations in between the optimistic and the pessimistic point of view. Another
possibility to get an intermediate value of global equality not as likely to
cause overestimation or understimation as (5.5) or (5.6) is to take for exam-
ple the average value of all the local equality degrees, i.e. to consider
X
kC = Dk =def kC = Dk(x) / card (X ) (5.7)
x∈X
in the case of a finite universe of discourse X or of at least a finite support of

the fuzzy set C ∪D, and to consider correspondingly the normalized integral,
assuming P -integrability of the function kC = Dk : X → IR,
Z
1
kC = Dk =def kC = Dk(x) dP
card (X [1] ) X
in the case of a continuous, finitely P -measurable universe of discourse X or

of at least a finitely P -measurable support supp (C ∪ D).
In the sequel we will also utilize some characteristics that can be deduced
from histograms of the values of the local equality degrees (5.3) computed
for each element of the universe of discourse.
An interesting problem that arises in relation to such a local equality
degree (5.3) is related to the so-called inverse problem, cf. Pedrycz (1990a).
In concise formulation we can state it as follows:
for a given fuzzy set C and a certain value v determine such
a fuzzy set D for which the inequality
kC = Dk(x) ≥ v (5.8)
holds true for all points x of the universe of discourse.

In other words, one is interested in getting a fuzzy set D such that C and D
are equal to each other to a degree locally not lower than v. (The equality
of C and D again here is considered in the pointwise sense).
Omitting a detailed discussion which has been performed in Pedrycz
(1990a), it is worth recalling only a few main results. At the very beginning
notice that (5.8) is usually not uniquely solvable. Instead, one can indicate
a certain range of the unit interval in which the values of the membership
function of D may be taken to fulfill (5.8). Thus, the resulting fuzzy set D
can in fact be interpreted as an interval-valued fuzzy set, a so-called Φ-fuzzy
set, cf. Sambuc (1975).
It is remarkable to note that replacing the inequality sign in (5.8) by
equality may cause there to be no solution. Thus, our formula (5.8) forms
an appropriate problem statement.
A few facts now will be summarized to give properties of the fuzzy sets
forming a solution to (5.8). We start by rewriting (5.3) in a slightly mod-
ified form utilizing properties of the Φ-operator ϕt , which starting from a
continuous t-norm t is given by (1.19). Straightforward calculations yield

 µC (x) ϕt µD (x), if µD (x) < µC (x)
kC = Dk(x) = µD (x) ϕt µC (x), if µD (x) > µC (x) (5.9)

1, if µD (x) = µC (x).
Let us denote by D∗v (a) the set of values of the membership function µD (x)
at point a satisfying (5.8), i.e.
D∗v (a) =def {µD (a) ∈ [0, 1] | v ≤ kC = Dk(a)}.
5.2. APPROXIMATE SOLUTIONS 177
Thus, if we put v = 0 then obviously µD (a) ∈ [0, 1] without restriction,

i.e. µD may take any values between 0 and 1; therefore D∗v (a) = [0, 1].
On the opposite side, putting v = 1 reduces D∗v (a) to exactly one point:
D∗v (a) = {µC (a)}. For intermediate situations, viz. for v ∈ (0, 1), the set
D∗v (a) is derived by solving the two inequalities
µC (a) ϕt µD (a) ≥ v, (5.10)

µD (a) ϕt µC (a) ≥ v. (5.11)
Solving the first of them we get a subinterval [d1 , µC (a)] while the second
leads us to a subinterval [µC (a), d2 ] where d1 and d2 are determined by
solving (5.10) and (5.11) respectively. Finally, D∗v (a) is nothing other than
the union of both those intervals:
D∗v (a) = [d1 , µC (a)] ∪ [µC (a), d2 ] = [d1 , d2 ]. (5.12)
Definition 5.1 The set (5.12) of solutions D∗v (a) will be called equality
interval or confidence interval for the fuzzy set D (at point a) induced by v
for the given fuzzy set C.
This inverse construction seems to be of considerable use in determining

resulting fuzzy sets for fuzzy models constructed in the framework of fuzzy
relational equations. In order to produce a more detailed picture of what
goes on in this area and what form of solution can be obtained there, we
reconsider this topic in the last part of section 5.4.
5.2 Approximate solutions for systems of fuzzy re-

lational equations
As it becomes quite clear from the previous discussions the results for exam-
ple of Table 3.2 have a significant value for applications only in the case that
solutions really exist, i.e. that not only approximate solutions exist. If this
true solvability is not the case – and it is this more uncomfortable situation
one usually meets in practice1 – then one has to think about other ways to
overcome the problem of the nonexistence of (true) solutions.
1
The situation here with the fuzzy data set (Ai , Bi ), i = 1, . . . , n, is quite analogous
to the usual situation one has with nonfuzzy data e.g. from a sequence of measurements:
theoretically, perhaps, they should be points on some curve, but in practice this only
approximately happens to be the case.
A quite simple, and perhaps for the practitioner the most obvious, way
out is to use the formulas which describe solutions – in the case of solvability
– of systems of fuzzy equations like
Θ(R, Ai ) = Bi , i = 1, . . . , n
even if there does not exist a solution to the system of fuzzy relational
equations to be considered – and then to check the quality of the approx-
imate solution obtained in this way. If this quality is not too bad, i.e. if
results/outputs B i given for the corresponding data/inputs Ai by the ap-
proximate solution are not “too far” from the results/outputs Bi one intends
to get from the set of data (Ai , Bi ) that constitutes the intended model and
hence determines the considered system of fuzzy relational equations, in such
a case at least an approximate fuzzy model is realized and can be used within
some quality bounds.
Taking such a path is completely in accordance with fuzzy modelling,
because any fuzzy model aims at a rough, i.e. approximate description of
real processes, situations, and the like.
Recapitulating the central results of chapter 3, which are of course cen-
tral from the present point of view, the point already made in section 3.5 was
that we found best possible approximate solutions for single equations, cf.
theorems 3.4, 3.7, and also found suitable candidates for “nearly best pos-
sible” approximate solutions to systems of fuzzy equations in theorem 3.9.
Section 3.6 added, with the results mentioned there in Facts 1 to Fact 4,
cf. page 132f., some further generalizations for more abstract types of fuzzy
equations.
These results are all together now the background for some kinds of
reasonable guesses we have in mind concerning the determination of approx-
imate solutions.
Claim 1: For systems of fuzzy relational equations of sup-type consider the

intersection of all the greatest solutions of the single equations (or of
good approximations for them) as a suitable approximate solution for
the system.
Claim 2: For systems of fuzzy relational equations of inf-type consider the

union of all the smallest solutions of the single equations (or of good
approximations for them) as a suitable approximate solution to the
system.
Claim 3: And for “combined” types of equations consider a suitable “com-

bined strategy” to provide a reasonable guess for an approximate so-
lution.
Applying this kind of “brute-force” strategy, in every case one needs some
accompanying discussion of the (expected) quality of such an approximate
solution. Of course, besides this “brute-force” approach towards solving
systems of fuzzy relational equations there are other, competing or supple-
mentary ways to treat this problem of approximately “solving” such systems.
Bearing in mind that in general no solution will exist, it is of special
interest to discuss ways of making any such system of fuzzy relational equa-
tions “better solvable” up to a level of approximate solvability acceptable by
the intended model-builder or user. In general, two ways of approach may
be distinguished here. The first is closely related to such changes in the data
set that might make a set of equations “more” solvable in the sense of the
existence of better approximate solutions. The second is oriented toward the
modification of the structure of the equations, for example by taking another
form of equation, or simply by extending the structure of the equations, e.g.
by adding some new variables (often called explanatory ones). We list some
representative methods coming from these two groups and refer to the liter-
ature where more detailed descriptions may be found. We also recall some
underlying ideas supporting these methods.
While considering the first possible stream of modification we have to
pay attention to the fact that again at least two main groups might be
recognized: in the first we simply skip some elements of the data set (in fact
this deletion modifies it), whereas in the second all the elements of the data
set are modified.
1. In Gottwald/Pedrycz (1985) a way of measuring the consis-

tency of pairs of the data set, say (Ai , Bi ) and (Aj , Bj ), has been proposed.
Roughly speaking, the procedure works as follows: A suitable consistency
index has to be considered that takes into account for pairs (Ai0 |= ⇒ Bi0 ) and
(Aj0 |=
⇒ Bj0 ) of control rules an absolute difference between the possibility of
Ai0 versus Aj0 and the possibility of Bi0 versus Bj0 . If for the two specified
indices i0 and j0 a low value is obtained (e.g. almost equal to zero) then one
pair of this data set is expected to be a candidate for being removed from
the entire data set.
The simple idea behind this approach is that it is taken as reasonable to
suppose that “nearly equal” input data should not be combined with “quite
different” output data.
A further variant of such a type of approach may be first to split the data
set into some clusters of control rules which within each cluster are similar
to each other. Then either
a simpler set of control rules can be taken into account with some “cen-
tres” of those clusters – or each cluster can be transformed into one rule,
but with probabilistic sets as input and output values.
2. A different approach was proposed in Gottwald/Pedrycz (1986)

and reconsidered in the previous chapter, section 4.3. There all the fuzzy
sets Ai ∈ IF X are modified by replacing them by their modified version:
Ai 7→ Aαi = Ai ∪ X [α]
where α is a threshold level supposed to be provided by an inspection of the

concrete applicational situation. The basic idea behind this data transfor-
mation (4.63) is that the smallest values of the membership functions of Ai
are irrelevant: for example they may be very sensitive to existing noise and
therefore they should be masked by an appropriately chosen value of the
threshold level α. As proved in that section, the solvability degree discussed
there is an increasing function of α. Then by observing its dependency of
the threshold α, one can take a suitable value of α which on one hand is low
enough not to disturb the data set at a significant level, and on the other
hand can lead to a significant increment of the solvability degree.
3. Yet another approach follows again a different way. Here the fuzzy
sets Ai are replaced by “sharpened” versions; more precisely, instead of the
original (input) data Ai , i = 1, . . . , n, one takes pairwise disjoint families of
fuzzy sets (A0i )1≤i≤n , i.e. such which fulfill the condition
A0i ∩t A0j = ∅ for all 1 ≤ i < j ≤ n. (5.13)
In this case, if for each pair (A01 , Bi ) there exists a solution to the corre-
sponding equation, then one is sure that there exists a solution for the entire
system of equations. The background facts here are the results of section 4.2
which indicate that the pairwise disjointness of the “input” data (A0i )1≤i≤n
is always a sufficient condition for the non-interactivity of that system (seen
as a realization of a system of control rules). And non-interactivity of course
guarantees solvability.
4. A proposal formulated by Wagenknecht/Hartmann (1986, 1986a)

deals with modifications of the fuzzy sets Bi forming the right-hand parts of
the equations. Instead of using the fuzzy sets Bi , so-called fuzzy sets with
tolerances have been utilized. By a fuzzy set with a tolerance they essentially
mean an interval-valued fuzzy set. At present, the tolerances are attached in
a heuristic fashion to reach a situation where the entire set of equations has
a solution. Nevertheless, the tolerances should be adjusted by a user and
then there is no security that they are not taken too broadly or that their
choice does not guarantee the existence of a solution.
5. Yet another way has been discussed in Li (1985/86), where a certain

construction has been designed to perturbate the data set and to reach a
state where the set of equations has a solution.
6. In Hirota/Pedrycz (1983) the use of the concept of probabilistic

set has been discussed. In this application, a probabilistic set is used to
mean a certain class of fuzzy sets that are similar to each other. As a conse-
quence, the data set is split into classes, i.e. it is structured. Afterwards it is
attempted to solve a reduced set of equations just by taking representatives
of each class, i.e. prototypes for each class.
7. Instead of measuring either the consistency of data pairs (Ai0 , Bi0 ),

(Aj0 , Bj0 ) as mentioned in point 1. or discussing the degrees of equality of
the intended and the realized output (for some given input) as we did in
earlier chapters, as well as in section 5.1, one could start from the data set
of pairs (Ai , Bi ), i = 1, . . . , n, together with some given distance d for fuzzy
subsets of the output space.
Now, looking in the traditional mathematical style for an optimization
problem, one can pose and discuss the problem to find such a fuzzy relation
R, i.e. to find a (card (X ) · card (Y))-matrix R, for which
n
X n
X
d(Bi , R00 Ai ) = d(Bi , Ai ◦t R) ⇒ min !
i=1 i=1
under the constraints that
µR (a, b) ∈ [0, 1] for all a ∈ X , b ∈ Y.
This is a problem with card (X ) · card (Y) variables and thus will rather
quickly present dimensionality problems because in interesting applications
the input and output spaces card (X ), card (Y) will be finite sets which are
not too big but which also are not very small finite sets.
Therefore, perhaps, another way round could prove to be more efficient:

given such a distance function d, try to express it (at least approximately)
as a distance function %t and then apply the “logical” methods of chapter 3.
Recapitulating this first stream of methods we can observe that they

modify the data set to make the corresponding system of equations “more
solvable” just by discovering a structure in it or in one of its “slightly”
modified forms – modified for example by the elimination of “outliers” in
the data set.
Discussing the second main stream of investigation, it is worth recalling
some ideas which are of interest for it.
A) First, keeping in mind that now the structure of the equation is

supposed to be fixed (viz. the type of composition) and moreover the fuzzy
sets Ai are unchanged, one can make some parametric studies either for
modifying some t-norm or Φ-operator appearing in the equation, or look for
a suitable value of the parameter in any parametrized class of, for example,
t-norms. This may lead to higher accordance of the data provided and the
fuzzy sets generated by the equations. Unfortunately, till now almost any
method for a systematic search within the whole infinite set of t-norms or t-
conorms has been lacking. Some partial search is nevertheless possible if one
restricts the considerations, for example, to one of the parametrized families
of t-norms mentioned in Table 1.1.
B) Contrary to the previous approach which might be seen as a paramet-

ric adjustment, one is now interested in considering different structures of
the fuzzy relational equations. Restricting ourselves to those already listed,
we have four ways when fitting the basic types of equations together with
the convex combination of Ohsato/Sekiguchi (1983), which always has a
solution in the case of a suitable choice of the convex combinator λ, a (fuzzy)
parameter controlling the influence of the “simple” types of equations which
are combined.
C) A third way that allows one to reach the main goal, i.e. to get a
solvable system of equations, is to add a new fuzzy variable which “separates”
the fuzzy sets A1 , A2 , . . . , An , cf. Pedrycz (1988, 1990). In its spirit this
way is similar to the method of Gottwald/Pedrycz (1986) but now no
5.3. FUZZY EQUATIONS FOR PROCESSING FUZZY DATA 183
input fuzzy set A1 , . . . , An is modified. To give a straightforward explanation

just take a system of equations
Ai ◦t R = Bi , i = 1, . . . , n. (5.14)
If Ri 6= ∅ for each of the solution sets Ri of the i-th equation, then to make
the system solvable it is enough to consider a new universe of discourse
Z = {z1 , . . . , zn } and to define fuzzy singletons Z1 , . . . , Zn as Zi = hhzi ii1 ,
i.e. by
½
1, if i = j
µZi (zj ) =def (5.15)
0 otherwise.
Then this new fuzzy variable defined in Z allows one to be sure that the
system of equations
(Ai × Zi ) ◦t R = Bi , i = 1, . . . , n (5.16)
has a solution. Note, however, that now R is defined in the cartesian product
of the universes X , Y, Z. In other words, the enriched structure (5.16) of
(5.14) is sufficient to have the system of equations solvable. The new universe
of discourse Z used here is a “space of explanation” and obviously in close
analogy to additional variables discussed for example in statistical models.
The weakest point of this approach is perhaps the special choice of the
separating fuzzy sets Zi chosen as fuzzy singletons – which therefore have
no linguistic meaning, i.e. usually may not reflect any linguistic concept
connected with the process whose control is intended to be realized using the
control rules Ai |=⇒ Bi . Nevertheless, one can search for a way to fuzzify these
singletons by just remembering not to exceed a certain overlap, measured for
example as the height of the intersections Zi ∩ Zj , between these auxiliary
fuzzy sets. For a more formal discussion see Pedrycz (1988).
Until now we have presented ideas how a degree of solvability of a given

system of equations can be enlarged; we have not discussed how the property
of (partial) lack of solvability influences the precision of the results obtained
by means of the fuzzy model constructed in such a way. This forms the topic
of section 5.4.
5.3 Fuzzy equations for processing fuzzy data

Bearing in mind the results obtained in section 4.3 with regard to data
transformations, let us take a look at their role in the processing of fuzzy
data. The problem may actually be formulated in the following way.
Some real world system shall be given that has to be modelled. Further-
more it is assumed that there are some fuzzy pieces of evidence associated
with this system which may be described by a collection of conditional state-
ments:
if Ai then Bi , i = 1, . . . , n.
The notion of system is considered here in a broad sense of the word. Thus
a system may be treated as a dynamic system with an unknown structure as
in Dubois/Prade (1980), Pedrycz (1982), Zadeh (1971a) or as a global
knowledge base (explanatory data base) for the manipulation of common-
sense knowledge as in Zadeh (1983, 1984). The given data Ai , Bi are sup-
posed to be linguistic terms, i.e. linguistic values of suitable (input and out-
put) variables, describing ambiguous facts available at hand. These linguistic
terms are expressed as fuzzy subsets Ai : X → [0, 1] and Bi : Y → [0, 1] of
suitable universes of discourse.
In order to process fuzzy data one has to build a model of the system
for which the above-mentioned data are relevant. That means that one has
a “reality layer” and a “modelling layer” which have to correspond to each
other sufficiently well in the relevant aspects of the problem.
In the modelling layer a surface model structure combining the fuzzy data
Ai and Bi in the sense of the fuzzy relational equations (4.28) is assumed.
The fuzzy relation R of the model results from a rule-based description of the
system behaviour from which the fuzzy data Ai , Bi is available. Of course,
not all the data pairs (Ai , Bi ) for i = 1, . . . , n may satisfy the system of
equations which fixes the model as supposed above. This is due to several
reasons:
• the structure of the model does not perfectly follow the structure of
the system which is indeed fuzzy; for instance, we have no extra infor-
mation on logical connectives – viewed as the concrete t-norm – of the
model;
• the data standing at our disposal are not unbiased – they may perhaps
be corrupted by some noise. (Note that the problem of their true
characterization is not as evident as in signal analysis performed in
for example communication theory for signals with additive Gaussian
noise.)
The approach we follow here is in modifying the data by imposing some
threshold level. The idea is that lower values of the membership functions
of Ai and Bi are less reliable than the highest ones. Thus, the threshold
level α is used to convert the original data set, i.e. the pairs (Ai , Bi ) for all
i = 1, . . . , n into a modified one (Aαi , Biα ), i = 1, . . . , n according to the data
transformation (4.63).
The fuzzy relation of the model is then calculated according to theorem
3.9 (ii), i.e. as
n
\
R= (Aαi ¤t Biα )
i=1
while the solvability degree ξ(α) of the system of equations is chosen ac-
cording to (3.12) but derived from the system of fuzzy relation equations
modified according to the manipulation procedure (4.63).
Any discussion of the value α of threshold level should embrace two facts:
(i) higher values of α allow us to achieve higher values of the solvability

index ξ(α);
(ii) higher values of α and hence more extended modifications of the origi-
nal data set lead to more rough models of the original process and less
transparent model relations.
Fact (i) has been proved in previous sections. Fact (ii) comes from the
property of Φ-operators that the values of the membership function of R,
calculated according to (5.3), are not smaller than α. It implies in turn that
the structure of the relation of the original model becomes partly hidden.
Of course, for α = 1 there is no remaining structure – the membership
function of R becomes simply equal to (X [1] × Y [1] ) and hence corresponds
to the linguistic term “unknown”. If one agrees to accept the lower values
of threshold α, a certain structure of the model appears but its adequacy
(measured via the solvability degree) decreases. All in all, increasing values
of the threshold cause improvements in the internal adequacy of the model
R in the sense of its behaviour in accordance with the constituting rules, but
the ability of recognition of the structure of the fuzzy model falls.
Some illustration of the behaviour of ξ(α) is shown in Figure 5.1.
Case 5.1a illustrates an ideal situation where the entire collection of the
data perfectly fits the equation of the model. A situation lying almost on the
opposite pole is shown in Figure 5.1b. Now a slight change of the threshold
α down the value 1.0 gives the value zero for the solvability degree. Case 5.1c
represents an intermediate situation where the structure of the model may
ξ(α) ξ(α)
16 p
p
16
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
p
p
p
p
p
p
p
p
p
p
p-α α
À
-
0 1 0 1
Figure 5.1a Figure 5.1b
ξ(α)
6
1 »((pp
!»
! p
´ p
, p
p
½ p
½ p
´ p
´ p
© p
» © p
ÃÃ » p-α
0 1
Figure 5.1c
Figure 5.1: Examples of solvability degrees ξ(α) as

functions of the threshold level α
be discovered to a certain level without drastic reduction in the value of the

solvability degree.
In such a sense ξ(α) is interpreted as a quality index of the fuzzy model
that is going to be constructed. The model characterized by the solvability
degree of Figure 5.1a is perfect, while for Figure 5.1b it is (almost) meaning-
less. This remark opens a new way of comparison of two fuzzy models: we
say that a first one M1 is superior to a second one M2 , M1 Â M2 , iff there
holds
ξ 1 (α) ≥ ξ 2 (α) for all α ∈ [0, 1],
where the superscript attached to the solvability degree refers to the model.
Now, for example, one can search for an appropriate t-norm to be used in
the model equations which leads to a high value of ξ(α).
Let us consider a certain policy that leads to some “slightest” modifica-

tion of all the fuzzy data pairs (Ai , Bi ), i = 1, . . . , n and which is sufficient
to achieve the value =1 for the solvability degree. Denote by α0 the smallest
value of the threshold level for which the solvability degree equals one:
α0 = min{α ∈ [0, 1] | ξ(α) = 1}.
Solve the system of fuzzy relational equations
Aαi 0 ◦t R = Biα0 , i = 1, . . . , n
with respect to the fuzzy relation R. This yields the fuzzy relation R∗ given
as
n
\
R∗ = (Aαi 0 ¤t Biα0 ). (5.17)
i=1
Now convert all the fuzzy data Bi to Bi∗ according to the following formula:
Bi∗ = Ai ◦t R∗ , i = 1, . . . , n.
This in turn gives a modified set of pairs of fuzzy data (Ai , Bi∗ ), i = 1, . . . , n
whose corresponding system of model equations has value =1 of its solvability
degree.
Being equipped with the fuzzy model in the form of a system of fuzzy
relational equations or the fuzzy relation R determined by such a system
and its evaluation by ξ(α), the determination of the fuzzy consequence, i.e.
the fuzzy output B for a given fuzzy input A may then be performed in the
following steps:
– evaluation of the model quality via its expected outputs and the solv-
ability degree ξ,
– modification of the result of this composition according to (4.63) with

respect to the threshold level α0 , where α0 may be the lowest value of
the threshold level that gives an accetable value of the relative solv-
ability degree ξ(α),
– determination of the corresponding “approximate model” R∗ ,
– calculation of the composition of the input A with the “approximate”

fuzzy relation R∗ .
It is remarkable that in a situation in which α0 is high enough, a fuzzy

set of the form B α0 becomes nearly meaningless (too “fuzzy”). Moreover,
one cannot modify the result of the composition of A and R∗ , but in such
a case it should be underlined that the values of the membership function
which are below α0 should be neglected, and we are unable to treat them as
significant.
The effect of deformation of B into B α0 may additionally be measured
by some energy measure G of the fuzzy set ∆B α0 defined as
µ∆B α0 (y) =def µB α0 (y) − µB (y) for all y ∈ Y
preferably some energy measure G with G(∅) = 0 but whose values are 6= 0
for fuzzy singletons.
5.4 Evaluating the quality of fuzzy models

In the previous sections we have summarized among other things how fuzzy
relational equations can be solved in an exact manner, or at least in an
approximate fashion. Now we intend to move one more step forward and
to discuss further ways in which the precision of a solution of fuzzy relation
equations and also of other fuzzy models for given sets of fuzzy data may
be expressed quantitatively. Immediately one concludes that this might
have primordial influence on the application of constructed fuzzy relational
equations. In the case of a broad utilization of such fuzzy models one likes
to have at hand a clear answer what quality of results can be expected if
the fuzzy model has been chosen – and also how to compare different fuzzy
models for the same set of fuzzy data or the same process.
The most essential and fruitful idea such evaluations and comparisons
shall be based upon is quite often some kind of “performance index” which
is (mainly) intended to be used for relative evaluations of fuzzy models.
5.4.1 Evaluating fuzzy models through fuzzy integrals

The “localization” (5.3) of the generalized equality considerations we men-
tioned in section 5.1 shall now also be used in another way.
Suppose for simplicity that the universe of discourse Y of the output
fuzzy sets Bi of some fuzzy controller with a family (4.1) of control rules is
a finite set: Y = {b1 , . . . , bm }, and consider fuzzy sets C, D ∈ IF (Y). Let be
for each index j = 1, . . . , m :
γjC,D =def kC = Dk(bj )
5.4. EVALUATION OF FUZZY MODELS 189
according to (5.3) and (5.4). That means with γjC,D we “measure” pointwise
the coincidence of the membership degrees of C and D getting for example
γjC,D = 1 in the case that µC (bj ) = µD (bj ).
Suppose additionally that there has been determined some fuzzy relation
R as a – perhaps only approximate – realization of the fuzzy controller
constituted by (4.1). Because R may only approximately realize the control
rules (4.1) let additionally be
B i = R00 Ai = Ai ◦t R
the real output of R for the input Ai . To simplify notation put
γji =def γjBi ,B i = kBi = B i k(bj ). (5.18)

Now combine those indices for all control rules into
n
^
Γj =def γji (5.19)
i=1
and unite all these local comparison indices into the evaluation vector
Γ =def (Γ1 , . . . , Γm ). (5.20)
Instead of looking at Γ as a vector of local comparison indices, one can
also view Γ as the fuzzy subset of Y of all those points where the fuzzy
model fits in well with the fuzzy data set.
The essential idea now is to use this vector or fuzzy set Γ to evaluate the
global property
Fuzzy model R well represents the fuzzy data set over
(5.21)
the whole output space Y.
The reference to formula (5.4) in the definition of the present indices
γij has as an immediate consequence an inequality for the indices Γj : an
inequality which relates to the dependence of those indices on the t-norm
which defines the Φ-operator which is involved in formula (5.18), i.e. in
(t)
formula (5.4). Writing for the moment Γj in the case that the definition
(5.19) is (implicitely) referring back to the t-norm t, we immediately have
for any left continuous t-norms t1 , t2 that
(t1 ) (t )
t1 <
= t2 ⇒ Γj ≤ Γj 2
(min)
and thus that Γj conveys the most pessimistic, i.e. most restrictive eval-
uation of the (localized) equality of fuzzy sets.
If Γ = 1 is the vector with all its components =1 then one may assume
that the model R of the process under consideration completely fits the data
set (4.1). If Γ 6= 1 a degree of fitness of the model, i.e. of the fuzzy controller
R will now be expressed through a fuzzy integral.
Two different fuzzy models R1 , R2 of the same set of fuzzy data, e.g. the
same system (4.1) of control rules, can be compared with respect to property
(5.21) in an almost obvious way if their respective evaluation vectors Γ 1 , Γ 2
of local equality degrees of realized and expected outputs chosen according
to (5.20) are comparable:
If for two fuzzy models R1 , R2 of the same set of fuzzy data for their
evaluation vectors Γ 1 , Γ 2 one has Γ 1 > 2 1 2
= Γ , i.e. Γj ≥ Γj for all j = 1, . . . , m,
then the fuzzy model R1 is the better one of the two.
Unfortunately, however, this situation of comparable evaluation vectors
Γ 1, Γ 2is an exception. Usually these vectors are incomparable with respect
to their natural componentwise partial ordering. This fact, as well as the
pointwise construction of Γ with respect to the points of the output space Y,
suggests realizing a partial evaluation of the quality of fuzzy models relative
to each point of Y, i.e. considering instead of the global property (5.21) their
local version
Fuzzy model R well represents the fuzzy data set at

(5.22)
the point bk of the output space Y.
In the spirit of the Gestalt principle, cf. for example Corge (1975), it is
obvious that the evaluation of the global property (5.21) cannot be deduced
by simple, perhaps even linear aggregation of the partial evaluations of the
model given using (5.22). This leads to the idea of considering some fuzzy
measure and the fuzzy integral defined by it as plausible tools for formulating
the global evaluation of the fuzzy model out of its local evaluations.
By a fuzzy measure G over the universe of discourse Y, as defined in
Sugeno (1974, 1977), a real valued function G : IP (Y) → [0, 1] over the
power set IP (Y) of Y is meant which has the properties
(FM1) G(∅) = 0 and G(Y) = 1,
(FM2) for all B1 , B2 ⊆ Y
B1 ⊆ B2 ⇒ G(B1 ) ≤ G(B2 ),
(FM3) for all ⊆-monotonic sequences (Bn )n≥1 of subsets of Y
lim G(Bn ) = G( lim Bn ).

n→∞ n→∞
The second of these properties, the monotonicity of the fuzzy measure G, is

of crucial interest here.2
The fuzzy integral based on such a fuzzy measure G is defined for any3
function H : Y → [0, 1], i.e. any fuzzy subset H of Y and any (crisp) subset
B of Y by the formula
Z
H(x) ◦ G(.) =def sup min{α, G(B ∩ H ≥α )}. (5.23)
B α∈[0,1]
The idea now is that a fuzzy measure G “measures” with its values G(B)
to what extent one can judge the quality of a fuzzy model with respect
to the global property (5.21) out of the local variants (5.22), i.e. from the
knowledge of the localized information Γj of (5.19) for points in B. And the
fuzzy integral provides this global evaluation from the local ones.
Assuming that one has the relevant information Γj on the local behaviour
of a fuzzy model for all points bj of a subset B of the output space Y and also
the fuzzy measure G available then an evaluation of the fuzzy model, based
on that partial information and using the fuzzy integral, can be provided by
the index
Z
Λ(B, Γ ) =def Γ (x) ◦ G(.) (5.24)
B
if we write Γ (bj ) = Γj in this formula for ease of notation. Of course, the

ideal situation for obtaining the global evaluation of the fuzzy model is to
have B = Y. In that case formula (5.23) becomes simpler and the fuzzy
model is then evaluated by the index
Z
ΛΓ =def Γ (x) ◦ G(.) = sup min{α, G(Γ ≥α )}.
Y α∈[0,1]
2
This monotonicity property replaces the stronger additivity property of the usual mea-
sures. By the way, let us note that the continuity property (FM3) is of course inessential
for fuzzy measures over finite sets.
3
In the case of an infinite universe of discourse Y the matter becomes a little more
difficult. As with traditional measures and integrals, one has to restrict the definition of
the measure to some suitable σ-algebra of subsets of Y and is only able to integrate over
functions H : Y → [0, 1] which are measurable with respect to that σ-algebra.
The monotonicity (FM2) of the fuzzy measure now reflects the fact that
the knowledge of more local information always adds to the global evalua-
tion of the fuzzy model. Additionally, however, the fuzzy measure allows
the “splitting” of the output space Y into regions of points whose local
evaluations of the fuzzy model are relatively more essential for the global
evaluation than points of other regions. That means that one becomes able
to distinguish points of the output space which are of “central importance”
for the global evaluation of the fuzzy model from such ones which are not
so important for the global evaluation. In practical applications the points
of central importance may then be such ones of a region in Y which is cru-
cial for the overall behaviour of the fuzzy model because the model has to
act quite exactly there – and other regions which allow for a rougher model
behaviour can have a lower significance for the global evaluation.
Nevertheless, the information provided by the index Λ(B, Γ ) from (5.24)
is quite incomplete in the case that B 6= Y because one is unable to distin-
guish the lack of information about the local behaviour of the fuzzy model
from the bad quality of this model. Therefore it seems preferable to switch
from this index (5.24) to another index Ξ, which itself is an ordered pair
consisting of the value G(B) of the “weighted” portion of available informa-
tion on the one hand, and of the normed fuzzy integral (5.24) on the other
hand:
³ Z
Λ(B, Γ ) ´ ³ 1 ´
Ξ(B, Γ ) =def G(B), = G(B), Γ (x) ◦ G(.) .
G(B) G(B) B
As in the case of the global or local indices Γ and Γj of (5.20), (5.19)

this index Ξ(B, Γ ) does not have an absolute meaning either – it is “only” a
tool to compare different fuzzy models for the same fuzzy data or the same
process.
5.4.2 Evaluating fuzzy models using probabilistic ideas

A close and transparent analogy can be found in general modeling principles
with the aid of statistical means. There it is a necessary step to validate
a model and to express the precision attached to it. Most simply, usually,
one tests the model with respect to its relevance, and for example an F-test
is frequently used. Moreover, the model is equipped with confidence curves
associated with the equations of the model. This way of thinking has a long
tradition and all its stages are commonly accepted and used in the field of
statistical modelling, cf. for example Box/Jenkins (1970).
The same does not hold true for the fuzzy set approach. Mainly only
the first of these steps is merely solved; unfortunately we cannot give a
quantitative characterization of the quality (relevance) of a model. A few
approaches try to tackle this problem, but in a qualitative way. The proposal
formulated here represents an attempt to express and measure the relevance
of a model and in its consequences leads to the formation of fuzzy sets of a
complex character, namely interval-valued fuzzy sets or fuzzy sets of type 2,
i.e. fuzzy sets whose membership degrees are intervals or fuzzy subsets of
[0, 1]; cf. Mizumoto/Tanaka (1976), Sambuc (1975).
In order to provide a clear description of the approach, it will be given in
an algorithmic form. This gives a concise presentation and allows potential
users to have this idea ready for use. Additionally we include some comments
to explain the consecutive steps.
Having at hand a data set represented by pairs (Ai , Bi ), i = 1, . . . , n
of fuzzy sets a fuzzy relation is to be constructed to (approximately) solve
the corresponding fuzzy relational equations. To focus attention we restrict
ourselves to one specified form of the equations, e.g. to (3.26), (3.27). Of
course, the procedure described in the following applies to any form of equa-
tions as indicated in sections 3.5 and 3.6. The fuzzy relation R is assumed
to be obtained via any suitable method.
Let us consider the fuzzy sets B i resulting from the fuzzy sets Ai com-
posed by R, i.e.: B i = Ai ◦t R. If the fuzzy model – here the fuzzy relation R
– is perfect, which almost never occurs in practice, then for every i = 1, . . . , n
we have B i = Bi . Since this case is not realistic, we may present a global
evaluation of the model by computing the (global) equality degree B i ≡t Bi
for each pair (B i , Bi ). But here we are more interested to take the local
point of view, i.e. to look “how equal” the fuzzy sets B i and Bi are separate
at each point of the universe of discourse Y. Thus we shall refer to the local
degrees of equality (5.3) and (5.4). Fixing some b ∈ Y we get for i = 1, . . . , n
a sequence of reals indicating how closely the membership degrees µBi (b)
and µB i (b) are equal to each other. Now, instead of looking for a single
number expressing en block a similarity of realized and intended outputs
such as (5.5), (5.6), we build a kind of empirical distribution function of
some equality degree using as a sample the values
kBi = B i k(b) for all i = 1, . . . , n. (5.25)
Let us denote this function by F (w; b) with w ∈ [0, 1]. By definition this
distribution function is given as

1
F (w; b) =def · card ({i ∈ {1, . . . , n} | kBi = B i k(b) ≤ w}). (5.26)
n
In virtue of (5.26), F (w; b) is monotonous in the first argument with F (1; b) =
1. We interpret F (w; b) as expressing the probability that the local equality
index at b does not exceed the specified value w in any case of the application
of the model.4 Now, let us further introduce the value
p = 1 − F (w; b) (5.27)
which will be considered as the probability that our local equality degree at
b ∈ Y attains values greater than w. Rewriting (5.27) as
Prob{(equality degree at b) > w} = p (5.28)
we recognize a direct dependence between this value of probability p and

the length of a kind of confidence interval induced by the value w. For
consistency of notation denote this relationship (5.28) by Q(w; b) = p. Sum-
marizing we have that p is a nonincreasing function of w.
Obviously, for any fuzzy set B and w = 0 we get a confidence interval of
length 1, while for w = 1 the length of the confidence interval is reduced to
zero.
From (5.27) we observe that with increasing values of w we have de-
creasing values of p. Hence this implies that if p increases, the length of the
confidence interval increases as well. Of course, one dislikes accepting confi-
dence intervals which are too broad since this conveys no useful information.
On the other hand, confidence intervals which are too narrow cannot be
accepted due to the significantly low values of probability attached to them.
From these remarks one can deduce a rationale how to choose a suitable
value of w. We want to achieve two contradictory goals. One is to get as high
a value of probability as possible. The second is to have confidence intervals
which are narrow enough. Denote by lb (w) the length of this confidence
interval. Hence, to fulfill both requirements we consider as a performance
index the product
Q(w; b) · (1 − lb (w)). (5.29)

4
Of course, this idea presupposes that all the control rules (A |=
⇒ B) are of the same
importance or applied with the same basic probability. That, surely, is the omly reasonable
presupposition for a first approach. Nevertheless, a more sophisticated approach should
allow one to dispense with this presupposition.
Since the first factor is a nonincreasing function of w and the second one is
a nondecreasing one, therefore there is at least one point w0 for which (5.29)
achieves a maximum value:
Q(w0 ; b) · (1 − lb (w0 )) = sup Q(w; b) · (1 − lb (w)). (5.30)

w∈[0,1]
This value w0 might be utilized for the further construction of an interval-

valued fuzzy set.
This gradual construction of a value w0 = w0 (b) is repeated at every
point of the universe Y; thus we get a function w0 with values w0 (b).
In this way the fuzzy model is characterized not only by the fuzzy relation
R, but also by a set of functions Q(w; b) – or at least a set of numbers w0 (b),
b ∈ Y, is attached.
For fuzzy relational equations we proceed accordingly. For a given fuzzy
set A compute B = A ◦t R – or having B at hand solve the inverse problem.
In both cases a suitable value of w is taken based on a prespecified probability
p or of extremal nature as w0 . In the direct mode this strategy yields an
interval-valued fuzzy set [B− , B+ ] that gives an impression how precise the
fuzzy set is which is “produced” by the fuzzy model behind the system of
equations.
For the two ways just described we can give the following interpretations:
– in the first situation we say that the genuine fuzzy set, which appears
in a system when A is given as input, has at each point b ∈ Y a
membership degree with respect to the output lying between µB− (b)
and µB+ (b);
– for the second construction we have a fuzzy set [B− , B+ ] with highest
specificity, thus the bounds µB− (b) and µB+ (b) determine a region in
[0, 1] in which one is sure to get a value of the membership at a high
value of probability and, simultaneously, this interval is chosen to be
narrow enough for the actual purposes.
5.4.3 Consequences for fuzzy modelling strategies

Without going deeply into various applications where fuzzy relational equa-
tions are discussed we will rather focus on methodological aspects which
arise while studying the introduced approach.
For this, note again that having an experimental data set in the form of
pairs (Ai , Bi ), two models can be constructed and their use may depend on
some specifity of the problem discussed. For short let us refer to them as
direct and backward use of the intended model. And let us assume that, as
before, the data set of pairs (Ai , Bi ) determines a system of fuzzy relation
equations to be solved to determine the intended model R.
As a schematic illustration of the following ideas, viz. to take two fuzzy
models for direct and backward mode of utilization, cf. Figure 5.2.
A fuzzy model1 for determination [B− , B+ ]

- direct - of the -
mode of use confidence interval
[A− , A+ ] determination fuzzy model2 for B

¾ of the ¾ backward ¾
confidence interval mode of use
Figure 5.2: Direct and backward use of fuzzy models
In the direct mode of using the model, possessing the fuzzy relation R
to tie A and B, viz. B is tied with A via R, we can find the response of the
model (system) for any given A straightforwardly.
But in the case that the fuzzy relation has not been determined or that
the system of fuzzy relation equations is not solvable and has only approx-
imate solutions the discussion becomes more involved. Thus let us suppose
that we only have some confidence intervals indicating the model quality.
Following the previously discussed schema of computation of the output we
get an interval-valued fuzzy set [B− , B+ ] which expresses bounds in which
the response of the system is contained – e.g. with a prescribed level of prob-
ability. Afterwards, if one is interested in describing successive states of the
system which has a dynamics governed by the constituting fuzzy data, fur-
ther iterations might be performed. Then the first output is the input for the
second step of the iteration. In other words, A in the next step is replaced
by [B− , B+ ]. For B− and B+ , separately, the corresponding interval-valued
set is built by taking the minimum of the two lower bounds of them and the
maximum of their two upper bounds, respectively, as resulting bounds. Ob-
viously, the width of the intervals of this interval-valued fuzzy set is broader
than the two original ones on which it is based.
The resulting effect is in agreement with our intuition: the result can
never be more precise than the arguments taken into account; cf. as a related
paper concerning this topic e.g. Czogala/Gottwald/Pedrycz (1982).
The so-called backward mode of the utilization of the fuzzy model cor-
responds to all the questions related to A if B is known. Of course, the
preferable situation is that R has already been computed. For instance, to
answer the question which input A – if any – leads to the output B, we have
to solve what is called the inverse problem. Originally, cf. Sanchez (1976),
assuming a nonempty set of solutions, one immediately started to consider
a fuzzy set A fulfilling the model equation, cf. Table 3.2.
But bearing in mind that in most cases the solvability of the basic system
of model equations which correspond to the constituting fuzzy data is lacking
and one thus has to look at approximate solutions, and that additionally one
is confronted with the aspect of model precision in general, it may be useful
to reformulate the problem thus:
Find A such that the fuzzy set resulting from A is covered by

[B− , B+ ], i.e. its membership values are within the bounds
of this characteristic interval-valued set [B− , B+ ].
However, even the solution of this problem might be too tedious a task.
Therefore we can think about a different model just establishing a re-
lation between B and A which is a little different. The model just derived
may answer the question which interval-valued fuzzy set [A− , A+ ] is obtained
for a specified output fuzzy set B. Surely, the “new” fuzzy relation can be
different from the previous one, and the precision of the new fuzzy model
might also be different.
As was mentioned previously, here we do not intend to discuss detailed
applications.
What remains to be discussed is how the analysis coming from this for-
ward and backward use of models can be enriched by imposing imprecision
attached to the interval-valued fuzzy sets. It seems that the forward and
backward strategy can transparently be utilized for reasoning schemes in
knowledge-based systems with mechanisms managing uncertainty in-built
in them. In such a case the models correspond to modelling directions of
reasoning.
5.5 Controllability and predictability in fuzzy

models
For any process to be modelled by scientific means, it is of interest to know
about its controllability, i.e. to know to what extent it is possible to influence
and govern – by external choice of some initial or boundary conditions –
or even to (completely) control its output. Of course, processes with a
very low or even zero level of controllability are without interest for (e.g.
computerized) control and for (standard) technical applications in general.
In the same way it is of interest for such processes to have information
about their long-range behaviour in time – with respect to either a continuous
or a discrete time structure – in order to be able to predict the process
behaviour for some longer time interval.
Ideally both these problems should be solved directly for the process
under consideration. But for quite different reasons this is not always pos-
sible. A kind of way out then is to look at some suitably good model of
the process and to discuss these problems of process behaviour not directly
for the process but for the model instead. Of course, the results of such an
approach have to be interpreted very carefully. But if there is independent
information about the model quality, the detour through the model may be
an interesting alternative for the direct study of the process.
The process model in particular may be a fuzzy model based on some
fuzzy (e.g. linguistic) data concerning the (qualitative) process behaviour.
How then can one approach these problems of controllability and predictabil-
ity? Of course, differing from the usual situation one has in mind when dis-
cussing fuzzy controllers, here we suppose that the fuzzy model is a model for
the process under consideration in the sense that it models the process in the
feedforward manner. The standard situation with fuzzy control, on the con-
trary, is the “inverse” one: the fuzzy controller is usually supposed to act in
the feedback manner with respect to the process and thus already supposes
the controllability of the process under consideration and even information
on how to control that process.
As usual in our approach we suppose that the fuzzy model is constituted
by a fuzzy relation R and that the model’s inputs and outputs are connected
via the “compositional rule of inference”, i.e. via fuzzy relation equations
of type (3.2). Therefore our fuzzy model connects with any input data
A ∈ IF (X ) the output B ∈ IF (Y) given by
B = R00 A = A ◦t R. (5.31)
5.5. CONTROLLABILITY AND PREDICTABILITY 199
For the controllability purpose we now make a further restriction: instead

of discussing the controllability of the process we discuss the controllability of
the model5 . Therefore the fuzzy output B of the model has to be interpreted
as formulating a goal to be reached by presenting the model with a suitable
input A. In the language of fuzzy equations: now equation (5.31) has to
be solved with respect to A for given data R and B (and of course a given
t-norm t).
From theorem 3.4 (i) we know that in the case of the solvability of an
equation of type
B = R00 X = X ◦t R (5.32)
the greatest solution (with respect to the partial ordering ⊂

=t in IF (X )) is
given by
X = R ↓ B = {x k ∀y((x, y) ε R →t y ε B)}.
Furthermore the discussion in section 3.5 adds the information that the
fuzzy set R ↓ B is even the best possible approximate solution of the fuzzy
equation (5.32) if there does not exist a solution at all.
Thus independent of the solvability of equation (5.32), for any given goal
B the fuzzy set A = R ↓ B can be taken as an optimal input value for
the fuzzy model to reach the goal B at least approximately. Of course, the
model output really reached with the input A = R ↓ B will be the fuzzy set
B = R00 (R ↓ B) = (R ↓ B) ◦t R. (5.33)
Again, therefore, the problem arises of comparing the fuzzy sets B and B.
And again there is the global approach towards that problem which was
discussed in the context of the fuzzified equality relation ≡t in chapter 3,
as well as the localized approach of section 5.4.1 with its idea of a pointwise
comparison of B and B over the output space Y of the fuzzy model R.
In the case of the global approach one looks for the truth degree of the
formula B ≡t B, i.e. for the truth degree
δB =def [[B ≡t R00 (R ↓ B)]]. (5.34)

5
Of course, with this decision we accept a – perhaps essential – weakening of the
problem of the controllability of the process. The feeling is that this restriction is not
too severe if the model has sufficiently high quality. Nevertheless, in any case one needs
extra discussion linking the controllability behaviour of the model with the corresponding
behaviour of the process.
Using proposition 3.1 (i) one has, assuming LSC(t),
|= B ≡t R00 (R ↓ B) ↔ B ⊂ 00
=t R (R ↓ B) (5.35)
and thus also
|= B ≡t R00 (R ↓ B) → B ⊂
=t rg(R) (5.36)
by the simple fact that
|= R00 (R ↓ B) ⊂
=t rg(R).
Thus there is
δB ≤ [[B ⊂
=t rg(R)]] = y∈Y
inf sup (µB (y) ϕt µR (x, y))
x∈X
which gives quite a simple upper bound of the global degree δB to which the
goal B can be reached by the fuzzy model R.
The localized approach, for simplicity, will again suppose that the output
space Y = {b1 , . . . , bm } is a finite set. Then, like in formula (5.20), the actual
evaluation vector Γ = (Γ , . . . , Γm ) will have the components
Γj =def kB = R00 (R ↓ B)k(bj )
which means, by the corresponding definitions together with (5.35) and

(5.36), that one has
Γj = [[bj ε B →t bj ε R00 (R ↓ B)]] ≤ [[bj ε B →t bj ε rg(R)]]
and which yields for each j = 1, . . . , m the upper bound
Γj ≤ sup (µB (bj ) ϕt µR (x, bj )).

x∈X
As in section 5.4.1 the localized approach now opens the way for another
kind of globalization of the local indices Γj which evaluate the controllability
of R with respect to the goal B: guided by a fuzzy measure G : IP (Y) → [0, 1]
over the output space Y with the fuzzy integral (5.24) one gets another global
controllability index
Z
ΛB =def Γ (x) ◦ G(.) = sup min{α, G(Γ ≥α )} (5.37)
Y α∈[0,1]
for the fuzzy model R with respect to the goal B.

The advantage of this second type of global index (5.37) over the first one
(5.34) is again the fact that the fuzzy measure G may distinguish regions
in the output space Y which are more sensitive for the evaluation of the
prediction quality of R with respect to the goal B than other ones.
In both cases, however, the evaluation of the prediction quality happens
only with respect to a fixed goal B. If one has in mind only some (very) few
goals one is interested in reaching by the fuzzy model R this situation may
be acceptable. In general, however, one should take into account a second
globalization procedure here: a globalization of the prediction quality with
respect to “all” the possible goals. In a very abstract sense the class IF (Y)
of all fuzzy subsets of the output space Y is the class of all possible goals.
Again now not all these possible goals in IF (Y) will have equal importance.
Thus here too a (second) fuzzy measure G : IP (IP (Y)) → [0, 1] can be taken
into consideration to evaluate the different possible goals with respect to
their importance for the real control actions one expects. Using this fuzzy
measure G the globalization of the evaluation of the controllability of the
fuzzy model R with respect to all the possible goals can be given in one of
the forms
Z Z
ΛB ◦ G(.) or δB ◦ G(.),
Z Z
Z = IP (Y) the class of all possible goals, depending on which global evalu-
ation of the controllability quality of R with respect to a fixed goal B the
approach is based.
Once again, as already done in section 5.4.1, let us however mention that
all those evaluation indices (localized or globalized) for the graded controlla-
bility property of a fuzzy model R have essentially only a relative meaning:
they are helpful for comparing either the relative (global) controllability be-
haviour of different fuzzy models, or the controllability behaviour of one
fuzzy model with respect to different goals either in the global or in the lo-
calized sense. Their absolute meaning does not usually provide a high level
of information.
Besides the controllability problem, we also mentioned initially the pre-

dictability problem with respect to a fuzzy model R. To have a closer look
at it, we now assume that the fuzzy model under consideration describes the
change of an actual situation from one point in time to the next one. That
means that we suppose to have a discrete time structure, and also that the
input and output spaces of the fuzzy model coincide: Y = X . Otherwise,
in the case that Y = 6 X , the successive outputs through time of the fuzzy
model R are determined by the successive inputs the model becomes con-
fronted with – thus all the “prediction” would consist in getting knowledge
of the input sequence, but this obviously in general is independent of the
fuzzy model under consideration. Thus our restrictions are well motivated.
Given one input fuzzy set A ∈ X the sequence of consecutive output fuzzy
sets of the process under consideration is denoted by B1 , B2 , B3 , . . . and the
corresponding sequence of consecutive output fuzzy sets of the fuzzy model
R by B 1 , B 2 , B 3 , . . .. That means for the fuzzy model that we consider the
sequence
B 1 = R00 A,
B 2 = R00 B 1 = R00 (R00 A) = (R2 )00 A,
B 3 = R00 B 2 = R00 ((R2 )00 A) = (R3 )00 A,
..
.
B n = . . . = (Rn−1 )00 A
with the notation Rn for the iterated relational product as introduced in
definition 2.13. This sequence B 1 , B 2 , B 3 , . . . of successive outputs of the
fuzzy model R generated by the initial input A is the exact prediction for
the models behaviour. The predictability for the model is hence not a serious
problem.
Thus here we again look at the mutual relation of the sequences of suc-
cessive outputs of the process on the one hand and of the fuzzy model on
the other hand, both starting from the same initial “input situation”. What
we would like to evaluate now is the property
The fuzzy model R predicts well the systems long range
(5.38)
behaviour.
Suppose that the accuracy of the fuzzy model R is characterized as before
in (5.20) by a vector Γ of indices Γj localizing the model quality according
to the points of the (finite) input and output space X = {b1 , . . . , bm }. This
vector may be given independently of the present considerations.
Using this vector Γ together with the first output B 1 of the fuzzy model R
and remembering the idea of “confidence intervals” mentioned in sections 5.1
and 5.4.2, we are able to define upper and lower bounds (B1 )∗ and (B1 )∗ for
the first process output B1 . Quite a simple idea is to determine these both
bounds pointwise through possible deviations of the membership degrees of
B 1 and B1 .
Having in mind the fact that the localizing indices Γj measure some local
degree of coincidence of membership degrees and that the negation of such a
coincidence degree should be something like a degree for deviation of these
membership degrees it is reasonable to use as membership degrees for the
fuzzy sets (B1 )∗ and (B1 )∗ the values
µ(B1 )∗ (bj ) = min{1, µB 1 (bj ) + (1 − Γj )}, (5.39)
µ(B1 )∗ (bj ) = max{0, µB 1 (bj ) − (1 − Γj )}. (5.40)
In a modified notation more in the traditional style of formal logic, these
formulas can be rewritten using the connectives &, ] of the L Ã ukasiewicz
many-valued logic as
|= bj ε (B1 )∗ ↔ bj ε B 1 ] (¬Γj ) ↔ (Γj →L bj ε B 1 ), (5.41)
|= bj ε (B1 )∗ ↔ Γj & bj ε B 1 . (5.42)
These last mentioned formulations indicate a possible further generalization
which will not be discussed here: instead of the L Ã ukasiewicz conjunction &
and implication →L some left continuous t-norm t and its corresponding
Φ-operator ϕt could be taken. The crucial point with such a generalization,
however, seems to be that one first has to check in which sense other t-norms
than tL adapt the idea of “confidence intervals” which was constitutive for
the approach through (5.39), (5.40).
Unfortunately, however, this rewriting does not yield “nice” formulas.
Therefore we additionally look for a more set theoretical shape for (5.39)
and (5.40). To reach that goal we have to reconsider the evaluation vector
Γ = (Γ1 , . . . , Γm ) of the local comparison indices. The idea leading to the
introduction of such an evaluation vector Γ in (5.20) was to connect with
each point bj ∈ Y its local comparison index Γj ∈ [0, 1], therefore Γ is
nothing other than a fuzzy subset of the input and output space X = Y.
For the membership degrees one obviously has
µΓ (bj ) = Γj .
Rereading (5.41), (5.42) with this interpretation of Γ in mind immediately
gives the simple representations
(B1 )∗ = CCΓ ∪tL B 1 = Γ ¤tL B 1 and (B1 )∗ = Γ ∩tL B 1 . (5.43)
∗
With these fuzzy sets (B1 ) and (B1 )∗ we now have an upper and a lower
bound for the process output B1 to the input A in the sense that we expect
to have
|= (B1 )∗ ⊂ ⊂
=t B1 =t (B1 ) .
∗
(5.44)
Of course, both bounds (B1 )∗ , (B1 )∗ have to be functions simply of the input
A (and the fuzzy model R). This, together with the idea that the construc-
tion of these bounds for B1 will become suitably iterated to also discuss
bounds for the further process outputs B2 , B3 , . . . forces us to introduce an
even more flexible notation. Hence we define for any A ∈ IF (X )
UR (A) =def CCΓ ∪tL R00 A, (5.45)

LR (A) =def Γ ∩tL R00 A. (5.46)
Independent of the idea behind (5.44) one immediately has for each fuzzy
set A ∈ IF (X ):
|= LR (A) ⊂
=t UR (A). (5.47)
These upper and lower bounds (5.45) and (5.46) provide the formal basis
for the problem of the long range prediction of the process behaviour. In
Gottwald/Pedrycz (1986a) a worst case discussion used them in the
sense that the bounds for the next step were taken in such a way that the
new lower bound was the minimum of the lower bounds of the model outputs
for both previous bounds and the new upper bound was the maximum of
the upper bounds of the same model outputs. This strategy, however, was
unnecessarily difficult and can be simplified just by using some monotonicity
properties of the operators UR , LR which generalize property (5.47).
Proposition 5.1 For each fuzzy model R ∈ IF (X ×X ), all fuzzy sets A, B ∈

IF (X ) and each t-norm t with property LSC(t) there hold true
(i) if |= A ⊂
=t B then |= UR (A) ⊂
=t UR (B),
(ii) if |= A ⊂
=t B then ⊂
|= LR (A) = LR (B).
t
Proof. Straightforward from the corresponding definitions or, even better,
by reference to theorem 2.1. QED
Compared with the usual style of results on fuzzy sets we intended to

prove in chapter 2, these monotonicity results are weak. Nevertheless, they
suffice for the present considerations on the prediction problem. But indeed,
they can be generalized. Unfortunately,howevert, in order to do so we need
some cumbersome assumptions, resulting from a reference to proposition 2.2
in the proof of these generalizations. It was mainly for this reason that the
more restricted proposition 5.1 was presented first.
Proposition 5.2 For each fuzzy model R ∈ IF (X ×X ), all fuzzy sets A, B ∈

IF (X ) and each t-norm t with property LSC(t) there hold true
(i) |= A ⊂ ⊂
=t B → LR (A) =t LR (B),
and if the t-norm t also distributes over the L Ã ukasiewicz disjunction ] or
fulfills one of the conditions (2.16), (2.17) with respect to st1 = ] there also
holds true
(ii) |= A ⊂ ⊂
=t B → UR (A) =t UR (B).
Proof. (i) By proposition 2.18 (i) one has
|= A ⊂ 00 ⊂
=t B → R A =t R B
00
and by proposition 2.2 (ii) then also

|= R00 A ⊂ 00 00 ⊂
=t R B → CCΓ ∪tL R A =t C CΓ ∪tL R00 B.
Using the transitivity of ⊂ =t together this yields (i).
(ii) is proven in the same manner but with reference to proposition 2.2
(iv) instead of proposition 2.2 (ii). And this reference together with the
remark following proposition 2.2, causes the additional assumptions for (ii).
QED
After this theoretical side remark concerning the monotonicity of the op-
erators UR , LR , let us return to the prediction problem. Already the (weaker)
monotonicity results of proposition 5.1 allows one to simplify the previously
mentioned discussion of Gottwald/Pedrycz (1986a) equivalently in con-
sidering as upper bounds for the process outputs B1 , B2 , B3 , . . . the succes-
sive iterations
U1 (A) = U(A),
U2 (A) = U(U(A)),
U3 (A) = U(U2 (A)),
..
.
of the upper bound operator (5.45) and as lower bounds of the process
outputs the corresponding iterations of the lower bound operator (5.46):
L1 (A) = L(A),
L2 (A) = L(L(A)),
L3 (A) = L(L2 (A)),
..
.
To get now information on the prediction quality of the fuzzy model R

(with respect to the initial input A) for a fixed, say k-step time horizon, we
compare the bounds Uk (A) and Lk (A) in the localized manner of sections 5.1
and 5.4.1. That means we regard for each point bj ∈ X = {b1 , . . . , bm } the
index
ηjk =def kUk (A) = Lk (A)k(bj )
which can be seen as defining a fuzzy subset η k of X whose membership

degrees are always µηk (bj ) = ηjk .
Again now a globalization is possible of these local evaluations of the
prediction quality of R for the time horizon k. As before we suppose that
a fuzzy measure G over X is given such that using the fuzzy integral the
globalization happens with respect to this fuzzy measure G and yields the
index
Z
Ψ(R, k) =def η k (x) ◦ G(.)
X
of k-step predictability.
Taken in an absolute sense, for a fixed fuzzy model R this index Ψ(R, k)
simply indicates the decline in prediction quality with a growing length of
the time horizon. That is nothing which is new or of deeper interest and
only a kind of formal confirmation of the well-known fact that the prediction
quality decreases with longer prediction intervals. The real interest in the
index Ψ(R, k) therefore is again in its use to compare different fuzzy models
of the same process with respect to their prediction behaviour.
Bibliography
Bandemer, H. and S. Gottwald

(1989) Einführung in Fuzzy-Methoden. Theorie und Anwendung un-
scharfer Mengen. Berlin (Akademie-Verlag). [3rd ed. 1992]
Bellman, R. and M. Giertz
(1973) On the analytic formalism of the theory of fuzzy sets. Infor-
mation Sci. 5, 149 – 156.
Box, G. E. P. and G. M. Jenkins
(1970) Time Series Analysis. Forecasting and Control. San Francisco
(Holden Day).
Chakraborty, M. K. and M. Das
(1983) Studies in fuzzy relations over fuzzy subsets. Fuzzy Sets Syst.
9, 79 – 89.
(1983a) On fuzzy equivalence. I, II. Fuzzy Sets Syst. 11, 185 – 193 and
299 – 307.
Chang, C. C. and H. J. Keisler
(1973) Model Theory. Amsterdam (North-Holland Publ. Comp.).
Chapin, E. W.
(1974/75) Set-valued set theory. I, II. Notre Dame J. Formal Logic 15,
614 – 634; 16, 255 – 267.
Corge, Ch.
(1975) Elements d’Informatique. Informatique et Demarche de
l’Esprit. Paris (Larousse).
Czogala, E.; J. Drewniak and W. Pedrycz
(1982) Fuzzy relation equations on a finite set. Fuzzy Sets Syst. 7, 89
– 101.
208 BIBLIOGRAPHY
Czogala, E.; S. Gottwald and W. Pedrycz

(1982) Aspects for the evaluation of decision situations. In: Fuzzy In-
formation and Decision Processes (M. M. Gupta, E. Sanchez,
eds.), Amsterdam (North-Holland Publ. Comp.), 41 – 49.
Czogala, E. and W. Pedrycz
(1981) Some problems concerning the construction of algorithms of
decision-making in fuzzy systems. Intern. J. Man-Machine
Stud. 15, 201 – 211.
Di Nola, A.; W. Pedrycz and S. Sessa
(1985) On measures of fuzziness of solutions of fuzzy relation equa-
tions with generalized connectives. J. Math. Anal. Appl. 121,
443 – 453.
Di Nola, A. and S. Sessa
(1983) On the set of solutions of composite fuzzy relation equations.
Fuzzy Sets Syst. 9, 275 – 285.
Di Nola, A.; S. Sessa; W. Pedrycz and E. Sanchez
(1989) Fuzzy Relation Equations and Their Applications to Knowl-
edge Engineering. Theory and Decision Libr., ser. D, Dor-
drecht (Kluwer Acad. Publ.).
Dombi, J.
(1982) A general class of fuzzy operators, the de Morgan class of fuzzy
operators and fuzziness measures induced by fuzzy operators.
Dubois, D. and H. Prade
(1980) Fuzzy Sets and Systems: Theory and Applications. New York
(Academic Press).
Frank, H. J.
(1979) On the simultaneous associativity of F (x, y) and x + y −
F (x, y). Aequat. Math. 19, 194 – 226.
Giles, R.
(1976) L
Ã ukasiewicz logic and fuzzy set theory. Intern. J. Man-Machine
Stud. 8, 313 – 327.
(1979) A formal system for fuzzy reasoning. Fuzzy Sets Syst. 2, 233
– 257.
BIBLIOGRAPHY 209
Gödel, K.
(1932) Zum intuitionistischen Aussagenkalkül. Anzeiger Akad. Wiss.
Wien, math.-naturwiss. Kl., 69, 65 – 66.
Gottwald, S.
(1971) Zahlbereichskonstruktionen in einer mehrwertigen Mengen-
lehre.
Ztschr. math. Logik Grundl. Math. 17, 145 – 188.
(1971a) Elementare Inhalts- und Maßtheorie in einer mehrwertigen
Mengenlehre. Math. Nachr. 50, 27 – 68.
(1974) Mehrwertige Anordnungsrelationen in klassischen Mengen.
Math.
Nachr. 63, 205 – 212.
(1976) A cumulative system of fuzzy sets. In: Set Theory and Hi-
erarchy Theory. Memorial Tribute A. Mostowski, Bierutowice
1975 (A. Zarach et al., eds.), Lecture Notes Math., vol. 537,
Berlin (Springer), 109 – 119.
(1979) Set theory for fuzzy sets of higher level. Fuzzy Sets Syst. 2,
125 – 151.
(1979a) Eine Anwendungsvariante der mehrwertigen Logik. Wiss.
Ztschr. KMU Leipzig, Ges.- u. Sprachwiss. R., 28, 303 – 312.
(1980) Fuzzy uniqueness of fuzzy mappings. Fuzzy Sets Syst. 3, 49 –
74.
(1983) Generalization of some results of Elie Sanchez. BUSEFAL
(Laborat. LSI, Univ. Paul Sabatier, Toulouse), no. 16, 54 –
60.
(1984) T-Normen und ϕ-Operatoren als Wahrheitswertfunktionen
mehrwertiger Junktoren. In: Frege Conference 1984, Proc. In-
tern. Conf. Schwerin Sept. 10–14, 1984 (G. Wechsung, ed.),
Math. Research, vol. 20, Berlin (Akademie-Verlag), 121 – 128.
(1984a) Criteria for non-interactivity of fuzzy logic controller rules.
In: Large Scale Systems: Theory and Applications, Proc. 3rd
IFAC/IFORS Symp. Warsaw 1983 (A. Straszak, ed.), Oxford
(Pergamon Press), 229 – 233.
(1984b) On the existence of solutions of systems of fuzzy equations.
(1984c) Fuzzy set theory. Some aspects of the early development. In:
Aspects of Vagueness (H.-J. Skala, S. Termini, E. Trillas, eds.),
Theory and Decision Libr., vol. 39, Dordrecht (Reidel), 13 –
29.
(1986) Characterizations of the solvability of fuzzy equations. Elek-
tron. Informationsverarb. Kybernet. EIK 22, 67 – 91.
210 BIBLIOGRAPHY
(1986a) Fuzzy set theory with t-norms and ϕ-operators. In: The Math-
ematics of Fuzzy Systems (A. Di Nola, A. G. S. Ventre, eds.),
Interdisciplinary Systems Res., vol. 88, Köln (TÜV Rhein-
land), 143 – 195.
(1986b) On some theoretical problems concerning the construction of
fuzzy controllers. In: Fuzzy Sets Applications, Methodological
Approaches, and Results (St. F. Bocklisch et al., eds.), Math.
Research, vol. 30, Berlin (Akademie-Verlag), 45 – 55.
(1989) Mehrwertige Logik. Eine Einführung in Theorie und Anwen-
dungen. Berlin (Akademie-Verlag).
(1990) Some observations and problems connected with fuzzy relation
equations. Fasciculi Mathematici, Nr. 19, Poznan (Polytech.
Poznan. Inst. Math.), 87 – 92.
(1991) Fuzzified fuzzy relations. In: Proc. IFSA ’91 Brussels (R.
Lowen, M. Roubens, eds.), vol.: Mathematics, Brussels (Vrije
Univ. Brussels), 82 – 86.
(1992) On t-norms which are related to distances of fuzzy sets. BUSE-
FAL, no. 50, 25 - 30.
Gottwald, S. and W. Pedrycz
(1985) Analysis and synthesis of fuzzy controller. Problems Control
Inform. Theory 14, 33 – 45.
(1986) Solvability of fuzzy relational equations and manipulation of
fuzzy data. Fuzzy Sets Syst. 18, 1 – 21.
(1986a) On the suitability of fuzzy models: an evaluation through
fuzzy integrals. Intern. J. Man-Machine Stud. 24, 141 – 151.
(1988) On the methodology of solving fuzzy relational equations and
its impact on fuzzy modelling. In: Fuzzy Logic in Knowledge-
Based Systems, Decision and Control (M. M. Gupta, T. Ya-
makawa, eds.), Amsterdam (North-Holland Publ. Comp.), 197
– 210.
Hamacher, H.
(1978) ber logische Aggregationen nicht-binär explizierter Entschei-
dungskriterien. Frankfurt/Main (Rita G. Fischer Verlag).
Hirota, K. and W. Pedrycz
(1983) Analysis and synthesis of fuzzy systems by the use of proba-
bilistic sets. Fuzzy Sets Syst. 10, 1 – 13.
Holmblad, L. P. and J. J. Østergaard
(1982) Control of a cement kiln by fuzzy logic. In: Fuzzy Informa-
tion and Decision Processes (M. M. Gupta, E. Sanchez, eds.),
Amsterdam (North-Holland Publ. Comp.), 389 – 399.
BIBLIOGRAPHY 211
Kaufmann, A.
(1975) Introduction to the Theory of Fuzzy Subsets. Vol. 1: Funda-
mental Theoretical Elements. New York (Academic Press).
(1977) Introduction à la théorie des sous-ensembles flous. A l’usage
des
ingénieurs. t. 4: Compléments et nouvelles applications. Paris
(Masson).
Klaua, D.
(1964) Allgemeine Mengenlehre. Berlin (Akademie-Verlag).
(1970) Stetige Gleichmächtigkeiten kontinuierlich-wertiger Mengen.
Monatsber. Deut. Akad. Wiss. Berlin 12, 749 – 758.
Klement, E. P.
(1982) Construction of fuzzy σ-algebras using triangular norms. J.
Math. Anal. Appl. 85, 543 – 565.
Li H.-X.
(1985/86) Fuzzy perturbation analysis. Part I: Directional perturbation.
Fuzzy Sets Syst. 17, 189 – 197;
Part II: Undirectional perturbation, ibid. 19, 165 – 175.
Ling, C. H.
(1965) Representation of associative functions. Publ. Math. Debrecen
12, 182 – 212.
L
Ã ukasiewicz, J.
(1970) Selected Works (L. Borkowski, ed.). Amsterdam (North-
Holland Publ. Comp.).
L
Ã ukasiewicz, J. and A. Tarski
(1930) Untersuchungen über den Aussagenkalkül. Comptes Rendus
Soc. Sci. et Lettr. Varsovie, cl. III, 23, 30 – 50.
Mamdani, E. H.
(1974) Application of fuzzy algorithms for the control of a simple
dynamic plant. Proc. IEEE 121, 1585 – 1588.
(1976) Advances in the linguistic synthesis of fuzzy controllers. Intern.
J. Man-Machine Stud. 8, 669 – 678.
Mamdani, E. H. and S. Assilian
(1975) An experiment in linguistic synthesis with a fuzzy logic con-
troller. Intern. J. Man-Machine Stud. 7, 1 – 13. [cf. also Mam-
dani/Gaines (1981)]
212 BIBLIOGRAPHY
Mamdani, E. H. and B. R. Gaines, eds.

(1981) Fuzzy Reasoning and Its Applications. New York (Academic
Press).
Mizumoto, M.
(1982) Fuzzy inference using max-∧ . -composition in the compositional
rule of inference. In: Approximate Reasoning in Decision Anal-
ysis (M. M. Gupta, E. Sanchez, eds.), Amsterdam (North-
Holland Publ. Comp.), 67 – 76.
(1989) Pictorial representations of fuzzy connectives, part I: Cases of
t-norms, t-conorms and averaging operators. Fuzzy Sets Syst.
31, 217 – 242.
Mizumoto, M. and K. Tanaka
(1976) Some properties of fuzzy sets of type 2. Information and Con-
trol 31, 312 – 340.
(1979) Some properties of fuzzy numbers. In: Advances in Fuzzy Set
Theory and Applications (M. M. Gupta, R. K. Ragade, R. R.
Yager, eds.), Amsterdam (North-Holland Publ. Comp.), 153 –
164.
Mizumoto, M. and H.-J. Zimmermann
(1982) Comparison of fuzzy reasoning methods. Fuzzy Sets Syst. 8,
253 – 283.
Ohsato, A. and T. Sekiguchi
(1983) Convexly combined form of fuzzy relational equations and
its application to knowledge representation. In: Proc. Intern.
Conf. Systems, Man and Cybernet., Bombay 1983/84; vol. 1,
Bombay – New Delhi (IEEE India Council), 294 – 299.
(1985) Maximin solution of the convexly combined form of composite
fuzzy relation equations [in Japanese]. Transact. Soc. Instru-
ment and Control Engineers (Japan) 21, 423 – 428.
Ovchinnikov, S. V.
(1981) Structure of fuzzy binary relations. Fuzzy Sets Syst. 6, 169 –
195.
Pedrycz, W.
(1982) Fuzzy control and fuzzy systems. Dept. Math., Delft Univ. of
Technology, Report 82 14.
(1983) Fuzzy relational equations with generalized connectives and
their applications. Fuzzy Sets Syst. 10, 185 – 201.
(1983a) Numerical and applicational aspects of fuzzy relational equa-
tions. Fuzzy Sets Syst. 11, 1 – 18.
BIBLIOGRAPHY 213
(1985) Applications of fuzzy relational equations for methods of rea-

soning in presence of fuzzy data. Fuzzy Sets Syst. 16, 163 –
175.
(1988) Approximate solutions of fuzzy relational equations. Fuzzy
Sets Syst. 28, 183 – 202.
(1989) Fuzzy Control and Fuzzy Systems. Taunton – New York (Re-
search Stud. Press – Wiley).
(1990) Relevancy of fuzzy models. Information Sci. 52, 285 – 302.
(1990a) Direct and inverse problem in comparison of fuzzy data. Fuzzy
Sets Syst. 34, 223 – 235.
Pedrycz, W.; E. Czogala and K. Hirota
(1984) Some remarks on the identification problem in fuzzy systems.
Rescher, N.
(1969) Many-Valued Logic. New York (McGraw-Hill).
Rodabaugh, E. S.; E. P. Klement and U. Höhle, eds.
(1992) Applications of Category Theory to Fuzzy Subsets. Dordrecht
(Kluwer Acad. Publ.).
Sanchez, E.
(1974) Equations de relations floues. Thèse de Doctorat, Faculté de
Médecine de Marseille.
(1976) Resolution of composite fuzzy relation equations. Information
and Control 30, 38 – 48.
(1977) Solutions in composite fuzzy relation equations: application
to medical diagnosis in Brouwerian logic. In: Fuzzy Automata
and Decision Processes (M. M. Gupta, G. N. Saridis. B. R.
Gaines, eds.), Amsterdam (North-Holland Publ. Comp.), 221
– 234.
(1978) Resolution of eigen fuzzy sets equations. Fuzzy Sets Syst. 1,
69 – 74.
(1984) Solution of fuzzy equations with extended operations. Fuzzy
Sets Syst. 12, 237 – 248.
Schweizer, B. and A. Sklar
(1961) Associative functions and statistical triangle inequalities.
Publ. Math. Debrecen 8, 169 – 186.
(1983) Probabilistic Metric Spaces. Amsterdam (North-Holland Publ.
Comp.).
214 BIBLIOGRAPHY
Sugeno, M.
(1974) Theory of Fuzzy Integral and Its Applications. Ph. D. Thesis,
Tokyo Inst. of Technology, Tokyo.
(1977) Fuzzy measures and fuzzy integrals: a survey. In: Fuzzy
Automata and Decision Processes (M. M. Gupta, G. N.
Saridis, B. N. Gaines, eds.), Amsterdam (North-Holland Publ.
Comp.), 89 – 102.
Thole, U.; H.-J. Zimmermann and P. Zysno
(1979) On the suitability of minimum and product operators for the
intersection of fuzzy sets. Fuzzy Sets Syst. 2, 167 – 180.
Wagenknecht, M. and K. Hartmann
(1986) On the solution of direct and inverse problems for fuzzy equa-
tion systems with tolerances. In: Fuzzy Sets Applications,
Methodological Approaches, and Results (St. Bocklisch et al.,
eds.); Math. Research, vol. 30, Berlin (Akademie-Verlag), 37
– 44.
(1986a) Fuzzy modelling with tolerances. Fuzzy Sets Syst. 20, 325 –
332.
Weber, S.
(1983) A general concept of fuzzy connectives, negations and impli-
cations based on t-norms and t-conorms. Fuzzy Sets Syst. 11,
115 – 134.
Weidner, A. J.
(1981) Fuzzy sets and Boolean-valued universes. Fuzzy Sets Syst. 6,
61 – 72.
Yager, R. R.
(1979) A measurement-informational discussion of fuzzy union and
fuzzy intersection. Intern. J. Man-Machine Stud. 11, 189 –
200.
(1980) On a general class of fuzzy connectives. Fuzzy Sets Syst. 4,
235 – 242.
Zadeh, L. A.
(1965) Fuzzy sets. Information and Control 8, 338 – 353.
(1971) Similarity relations and fuzzy orderings. Information Sci. 3,
159 – 176.
(1971a) Toward a theory of fuzzy systems. In: Aspects of Network and
System Theory (R. E. Kalman, N. de Claris, eds.), New York
(Holt, Rinehart and Winston), 469 – 490.
BIBLIOGRAPHY 215
(1973) Outline of a new approach to the analysis of complex systems

and decision processes. IEEE Trans. Systems, Man and Cy-
bernet. SMC-3, 28 – 44 [cf. also (1987)].
(1975) The concept of a linguistic variable and its application to ap-
proximate reasoning. I. - III. Information Sci. 8, 199 – 250 and
301 – 357; 9, 43 – 80. [cf. also (1987)]
(1978) Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets
Syst. 1, 3 – 28. [cf. also (1987)]
(1983) The role of fuzzy logic in the management of uncertainty in
expert systems. Fuzzy Sets Syst. 11, 199 – 227. [cf. also (1987)]
(1984) A theory of commonsense knowledge. In: Aspects of Vagueness
(H. J. Skala, S. Termini, E. Trillas, eds.), Dordrecht (Reidel),
257 – 295. [cf. also (1987)]
(1987) Fuzzy Sets and Applications. Selected Papers. (R. R. Yager et
al., eds.), New York (Wiley).
Zhang, J.-W.
(1980) A unified treatment of fuzzy set theory and Boolean-valued set
theory – fuzzy set structures and normal fuzzy set structures.
J. Math. Anal. Appl. 76, 297 – 301.
Zimmermann, H.-J.
(1985) Fuzzy Set Theory – and Its Applications. Dordrecht (Kluwer
- Nijhoff). [2nd ed. 1991]
Zimmermann, H.-J. and P. Zysno
(1980) Latent connectives in human decision making. Fuzzy Sets Syst.
4, 37 – 51.
Index
A disjunction . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
abstraction term infinitary . . . . . . . . . . . . . . . . . . . . . . 51
generalized . . . . . . . . . . . . . . . . . . . . . 38 t-norm based . . . . . . . . . . . . . . . . . . 19
activation . . . . . . . . . . . . . . . . . . . . . . . . . . 139 distinguishability function . . . . . . . . . 116
approximate solution . . . . . . . . . . . . . . 178 distributive laws . . . . . . . . . . . . . . . . . . . . 18
best possible . . . . . . . . . . . . . . . . . . 172 domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
quality of an . . . . . . . . . . . . . . . . . . 172
approximate solvability . . . . . . . . . . . . 122
approximation quality . . . . . . . . . . . . . 122
E
empty fuzzy set . . . . . . . . . . . . . . . . . . . . . 48
equality interval . . . . . . . . . . . . . . . . . . . 177
B equation
β-operator . . . . . . . . . . . . . . . . . . . . . . . . . 129 of inf-type . . . . . . . . . . . . . . . . . . . . 178
brute-force strategy . . . . . . . . . . . . . . . . 179 of sup-type . . . . . . . . . . . . . . . . . . . 178
eigen fuzzy set . . . . . . . . . . . . . . . . . 97
fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 97
C fuzzy arithmetical . . . . . . . . . 99, 103
cartesian product fuzzy relational . . . . . . . . 97, 99, 100
fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 55 of inf-type . . . . . . . . . . . . . . . . . . . . 131
certainty . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 of sup-type . . . . . . . . . . . . . . . . . . . 131
clan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 equivalence class
code (of a control rule) . . . . . . . . . . . . 142 fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 69
complement . . . . . . . . . . . . . . . . . . . . . . . . . 39 equivalence relation
compositional rule of inference 142, 143 fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 69
conjunction . . . . . . . . . . . . . . . . . . . . . . . . . . 2 evaluation vector . . . . . . . . . . . . . . . . . . 189
infinitary . . . . . . . . . . . . . . . . . . . . . . 51 existence lemma . . . . . . . . . . . . . . . . . . . . 12
t-norm based . . . . . . . . . . . . . . . . . . 19 extension principle . . . . . . . . . . . . . . . . . . 57
control rule . . . . . . . . . . . . . . . .98, 135, 137
crisp set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
F
full image . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
D fuzzy cartesian product . . . . . . . . . . . . . 55
de Morgan law . . . . . . . . . . . . . . . . . . 27, 31 fuzzy controller . . . . . . . . . . . . 98, 135, 137
degree of activation . . . . . . . . . . . . . . . . 137 fuzzy equation . . . . . . . . . . . . . . . . . . . . . . 97
degree of containment . . . . . . . . . . . . . . . 42 fuzzy equivalence relation . . . . . . . . . . . 69
degree of equality . . . . . . . . . . . . . . . . . . . 42 graded . . . . . . . . . . . . . . . . . . . . . . . . . 91
local . . . . . . . . . . . . . . . . . . . . . . . . . . 174 fuzzy integral . . . . . . . . . . . . . . . . . . . . . . 191
degree of solvability . . . . . . . . . . . . . . . . 113 fuzzy measure . . . . . . . . . . . . . . . . . . . . . 190
diagonal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 fuzzy partial ordering . . . . . . . . . . . . . . . 76
INDEX 217
graded . . . . . . . . . . . . . . . . . . . . . . . . . 85 metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

fuzzy preordering . . . . . . . . . . . . . . . . . . . 76 modus ponens . . . . . . . . . . . . . . . . . . . . . . 22
fuzzy relation . . . . . . . . . . . . . . . . . . . . . . . 58
fuzzy set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
conditioned . . . . . . . . . . . . . . . . . . . . 64 N
empty . . . . . . . . . . . . . . . . . . . . . . . . . 48 negation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
universal . . . . . . . . . . . . . . . . . . . . . . . 48 t-norm based . . . . . . . . . . . . . . . . . . 27
fuzzy subset . . . . . . . . . . . . . . . . . . . . . . . . . .1 negation function . . . . . . . . . . . . . . . . . 8, 16
strict . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
G O
generating family . . . . . . . . . . . . . . . . . . 137
operator
graded properties . . . . . . . . . . . . . . . . . . . 84
β- . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Φ- . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
H
Horn formula . . . . . . . . . . . . . . . . . . . . . . . 41
basic . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
P
partial ordering
fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 76
I graded fuzzy . . . . . . . . . . . . . . . . . . . 85
identity relation irreflexive fuzzy . . . . . . . . . . . . . . . . 76
many-valued . . . . . . . . . . . . . . . . . . . 44 partition
implication fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 71
t-norm based . . . . . . . . . . . . . . . . . . 19 Φ-operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
inclusion relation . . . . . . . . . . . . . . . . . . . 41 possibility . . . . . . . . . . . . . . . . . . . . . . . . . 173
inclusion relation preordering
many-valued . . . . . . . . . . . . . . . . . . . 42 fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 76
inference product
compositional rule of . . . . . . . . . . 142 cartesian . . . . . . . . . . . . . . . . . . . . . . . 55
interactive . . . . . . . . . . . . . . . . . . . . . . . . . 145 relational . . . . . . . . . . . . . . . . . . . . . . 59
intersection properties
of a family . . . . . . . . . . . . . . . . . . . . . 51 graded . . . . . . . . . . . . . . . . . . . . . . . . . 84
infinitary . . . . . . . . . . . . . . . . . . . . . . 51 pseudo-metric . . . . . . . . . . . . . . . . . . . . . .116
t-norm based . . . . . . . . . . . . . . . . . . 39
inverse problem . . . . . . . . . . . . . . . . . . . . 176
inverse relation . . . . . . . . . . . . . . . . . . . . . 59 Q
involution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 quantifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
restricted . . . . . . . . . . . . . . . . . . . . . . . 3
L R
left continuity . . . . . . . . . . . . . . . . . . . . . . . 11
linguistic variable . . . . . . . . . . . . . . . . . . 136 range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
realization (of a rule)
stronger . . . . . . . . . . . . . . . . . . . . . . 147
M weaker . . . . . . . . . . . . . . . . . . . . . . . . 147
membership function . . . . . . . . . . . . . . . . . 1 relation
membership predicate t-antisymmetric . . . . . . . . . . . . . . . . 68
generalized . . . . . . . . . . . . . . . . . . . . . 38 t-asymmetric . . . . . . . . . . . . . . . . . . 68
many-valued . . . . . . . . . . . . . . . . . . . 38 fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 58
View publication stats
218 INDEX
inverse . . . . . . . . . . . . . . . . . . . . . . . . . 59 T
irreflexive . . . . . . . . . . . . . . . . . . . . . . 68 t-clan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
linear . . . . . . . . . . . . . . . . . . . . . . . . . . 82 t-conorm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
reflexive . . . . . . . . . . . . . . . . . . . . . . . 68 t-norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
similarity . . . . . . . . . . . . . . . . . . . . . . 72 Archimedean . . . . . . . . . . . . . . . . . . . 77
symmetric . . . . . . . . . . . . . . . . . . . . . 68 idempotent . . . . . . . . . . . . . . . . . . . . 16
t-transitive . . . . . . . . . . . . . . . . . . . . . 68 interactive . . . . . . . . . . . . . . . . . . . . . 16
weakly linear . . . . . . . . . . . . . . . . . . .82 strict . . . . . . . . . . . . . . . . . . . . . . . . . . 78
relational equation . . . . . . . . . . . . . . . . . . 97 threshold level . . . . . . . . . . . . . . . . . . . . . 160
representation lemma . . . . . . . . . . . . . . . 10 transitive hull . . . . . . . . . . . . . . . . . . . . . . . 73
right continuity . . . . . . . . . . . . . . . . . . . . . 12 truth degree . . . . . . . . . . . . . . . . . . . . . . . 1, 2
rule of detachment . . . . . . . . . . . . . . . . . . 22 designated . . . . . . . . . . . . . . . . . . . . 1, 4
rule of generalization . . . . . . . . . . . . . . . . . 4
rule of inference . . . . . . . . . . . . . . . . . . . . . . 4
compositional . . . . . . . . . . . . . . . . . 143 U
correct . . . . . . . . . . . . . . . . . . . . . . . . . . 4 union
rule of particularization . . . . . . . . . . . . . . 4 of a family . . . . . . . . . . . . . . . . . . . . . 51
infinitary . . . . . . . . . . . . . . . . . . . . . . 51
t-norm based . . . . . . . . . . . . . . . . . . 39
S universal set . . . . . . . . . . . . . . . . . . . . . . . . 48
semicontinuity u- . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
lower . . . . . . . . . . . . . . . . . . . . . . . . . . 11 universe of discourse . . . . . . . . . . . . . . . . . 1
upper . . . . . . . . . . . . . . . . . . . . . . . . . . 12
similarity relation . . . . . . . . . . . . . . . . . . . 72
singleton
V
valid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
fuzzy . . . . . . . . . . . . . . . . . . . . . . . . . . 48
logically . . . . . . . . . . . . . . . . . . . . . . . . . 4
u- . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
solvability degree . . . . . . . . . . . . . . . . . . 113
relative . . . . . . . . . . . . . . . . . . 114, 161 Z
subdistributive laws . . . . . . . . . . . . . . . . . 18 zero divisors . . . . . . . . . . . . . . . . . . . . . . . . 77
subset property . . . . . . . . . . . . . . . . . . . . 145
superset property . . . . . . . . . . . . . . . . . . 145
support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Szpilrajn’s theorem . . . . . . . . . . . . . . . . . 82

BOOK

Uploaded by

Copyright:

Available Formats

BOOK

Uploaded by

Document Information

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

BOOK

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Fuzzy Sets and Fuzzy Logic. The Foundations of Application—from a

Book · February 1993

The user has requested enhancement of the downloaded file.

Gottwald, Siegfried: - Braunschweig

2 Basic fuzzy set theory 37

3 Set equations with fuzzy sets 97

4 Fuzzy controllers 135

5 Methodological issues 171

1.1 Basic notions

are from {0, 1} – and additionally to coincide under such circumstances

variable x of H has to be given the value a from the universe X .

[[∀C xH(x)]] = inf [[H(x/a)]], (1.3)

(If C means a “crisp” predicate, substitute C(a) for a ∈ C.)

[[x ε A]] =def µA (x). (1.5)

|= H iff [[H]] = 1. (1.6)

And a further consequence of this choice is that a rule of inference

if |= H(x) then |= ∀xH(x)

if |= H(x/a) then |= ∃xH(x)

1.2 t-norms and Φ-operators

(b) A binary operation s in [0, 1] is a t-conorm iff it is

The t-norms – a common shorthand term for “triangular norms” – have

in infix notation, i.e. to take them either as two-place functions in [0,1] or

t-norm t formula for t(u, v)

Table 1.1: Examples of t-norms1

For, given any t-norm t, by

and hence also

st1 < st2 ⇔ t2 < t1 .

Therefore it suffices to consider these orderings for t-norms only.

Definition 1.2 A unary operation n in the real interval [0, 1] is a negation

Clearly, the function x 7→ 1 − x is an involution. Furthermore, the

sn,t (u, v) = n−1 (t(n(u), n(v))) (1.17)

tn,s (u, v) = n−1 (s(n(u), n(v))) (1.18)

a t-norm is defined. For involutions n in this way a 1-1 correspondence is

Table 1.2: Examples of t-norms and their t-conorms

The Φ-operators have been introduced by Pedrycz (1982) to describe

which obviously is equivalent to our (Φ1).

Proposition 1.1 (Representation Lemma) If ϕ is a Φ-operator which

u ϕ v = sup{w | u t w ≤ v}. (1.19)

Proof. By (Φ2) we immediately get

But in case u ϕ v < sup{w | v t w ≤ v} there would exist w0 such that u ϕ w0

which is a contradiction. Thus our representation lemma is proved. QED

Therefore, one Φ-operator at most is connected to each t-norm, which in

Hence it is always even

Proposition 1.2 A t-norm is left continuous iff it is lower semicontinuous.

Proof. More general, for each monotonically non-decreasing function the

In the following we write

In a few cases it is not the lower semicontinuity of a t-norm t that is

Now we are able to solve the problem of the existence of a Φ-operator

Proposition 1.3 (Existence Lemma) For a t-norm t there exists a Φ-

Proof. Assume LSC(t). Let ϕ be defined by (1.19). Obviously this binary

If we assume that there exists a Φ-operator ϕ connected to t then we have

Definition 1.4 We denote the uniquely determined Φ-operator which is

In Table 1.3 we give a list of Φ-operators which correspond via (1.22)

Proposition 1.4 For each binary operator ϕ defined by formula (1.19) it

Proof. Obvious from (1.19).

Table 1.3: Examples of t-norms and their Φ-operators 5

Then by (T2) we get

u = u t 1 = u t sup{w | w < 1} = sup{u t w | w < 1} ≤ v.

(ii) Because of proposition 1.5 (ii) it is enough to prove