v

Preface
some great text
Petr Cintula, Petr H ajek, Carles Noguera
the editors
vii
The list of authors
LIBOR B EHOUNEK
Institute of Computer Science
Academy of Sciences of the Czech Republic
Pod Vod arenskou v e z 2
182 07 Prague 8, Czech Republic
Email: behounek@cs.cas.cz
MANUELA BUSANICHE
Instituto de Matem atica Aplicada del Litoral
FIQ, CONICET-UNL
Guemes 3450
S3000GLN-Santa Fe, Argentina
Email: mbusaniche@santafe-conicet.gov.ar
PETR CINTULA
Institute of Computer Science
Academy of Sciences of the Czech Republic
Pod Vod arenskou v e z 2
182 07 Prague 8, Czech Republic
Email: cintula@cs.cas.cz
PETR H AJEK
Institute of Computer Science
Academy of Sciences of the Czech Republic
Pod Vod arenskou v e z 2
182 07 Prague 8, Czech Republic
Email: hajek@cs.cas.cz
ROSTISLAV HOR CIK
Institute of Computer Science
Academy of Sciences of the Czech Republic
Pod Vod arenskou v e z 2
182 07 Prague 8, Czech Republic
Email: horcik@cs.cas.cz
GEORGE METCALFE
Mathematical Institute
University of Bern
Sidlerstrasse 5
CH-3012 Bern, Switzerland
Email: george.metcalfe@math.unibe.ch
viii
FRANCO MONTAGNA
Department of Mathematics and Computer Science,
Pian dei Mantellini 44,
53100 Siena, Italy
Email: montagna@unisi.it
CARLES NOGUERA
Articial Intelligence Research Institute (IIIA), CSIC
Campus de la Universitat Aut ` onoma de Barcelona s/n
08193 Bellaterra, Catalonia, Spain
Email: cnoguera@iiia.csic.es
CONTENTS
Preface v
The list of authors vii
Volume 1
I Introduction to Mathematical Fuzzy Logic
I Libor B ehounek, Petr Cintula, and Petr H ajek 1
1 Propositional logics of continuous t-norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Standard semantics of t-norm fuzzy logics . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Axiomatic systems for logics of continuous t-norms. . . . . . . . . . . . . . . . . . . 13
1.3 Algebraic semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2 Variations of basic propositional fuzzy logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1 Discarding axioms or rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2 Adding new connectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Discarding connectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.4 Adding axioms or rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Families of fuzzy logics in the logical landscape. . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.1 Fuzzy logics among substructural logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Core and -core fuzzy logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3 Fuzzy logics as algebraically implicative semilinear logics . . . . . . . . . . . . . 63
4 Metamathematics of propositional fuzzy logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1 Completeness theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2 Functional representation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Proof theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4 Computational complexity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5 Predicate fuzzy logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.1 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3 The identity and sorts of objects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.4 Axiomatic theories over fuzzy logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.5 Higher-order fuzzy logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
x
II A General Framework for Mathematical Fuzzy Logic
II Petr Cintula and Carles Noguera 103
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
2 Weakly implicative logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
2.1 Basic notions and a rst completeness theorem . . . . . . . . . . . . . . . . . . . . . . . 106
2.2 Weakly implicative logics and a second completeness theorem . . . . . . . . . 112
2.3 Advanced semantics and a third completeness theorem. . . . . . . . . . . . . . . . 116
2.4 Algebraically implicative logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2.5 Substructural logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
2.6 Deduction Theorems and Proof by Cases in substructural logics . . . . . . . . 137
2.7 Generalized disjunctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3 Semilinear logics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
3.1 Basic denitions, properties, and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
3.2 Disjunction and semilinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
3.3 Strengthening completeness: densely ordered chains . . . . . . . . . . . . . . . . . . 166
3.4 Strengthening completeness: arbitrary classes of chains . . . . . . . . . . . . . . . 171
4 First-order predicate semilinear logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
4.1 Basic syntactic and semantic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
4.2 Axiomatic systems and soundness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
4.3 Predicate substructural logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.4 Completeness theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
4.5 -Henkin theories, Skolemization, and witnessed semantics. . . . . . . . . . . . 197
5 Historical remarks and further reading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
III Proof Theory for Mathematical Fuzzy Logic
III George Metcalfe 211
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
2 A Gentzen systems primer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
2.1 Rules, systems, derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
2.2 A proof system for lattices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
2.3 A sequent calculus for classical logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
2.4 Substructural logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
3 From sequents to hypersequents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
3.1 The core set of rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
3.2 Adding structural rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
3.3 Soundness and completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
4 Eliminations and applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
4.1 Cut elimination. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
4.2 Density elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
5 The fundamental fuzzy logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
5.1 G odel logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
5.2 ukasiewicz logic and Giless game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
xi
5.3 Product logic and related systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
5.4 Uniform systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
6 Quantiers and modalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
6.1 First-order quantiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
6.2 Propositional quantiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
6.3 Modalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
7 Historical remarks and further reading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
IV Algebraic Semantics: Semilinear FL-Algebras
IV Rostislav Hor ck 287
1 Substructural logics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
2 Algebraic preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
2.1 Universal algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
2.2 Residuated maps and Galois connections. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
3 FL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
3.1 Examples of FL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
3.2 Involutive FL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
3.3 A bottom and a top element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
3.4 Algebraizability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
3.5 Congruences and lters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
3.6 Semilinear varieties of FL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
3.7 Nuclei and conuclei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
3.8 Dedekind-MacNeille completion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
4 Completeness with respect to distinguished semantics . . . . . . . . . . . . . . . . . . . . . 322
4.1 Integral semilinear varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
4.2 Non-integral semilinear varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
5 Subvariety lattice. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
5.1 General facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
5.2 Finitely many atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
5.3 Continuum many atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
6 Historical remarks and further reading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
V H ajeks BL Logic and BL-Algebras
V Manuela Busaniche and Franco Montagna 359
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
2 Denitions and basic properties of BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
2.1 Continuous t-norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
2.2 Hoops and BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
2.3 BL-algebras as residuated lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
2.4 BL-chains and completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
2.5 Important subvarieties of BL-algebras. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
xii
2.6 Implicative lters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
2.7 Regular and dense elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
3 Decomposition theorems for BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
3.1 Ordinal sums. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
3.2 Decomposition of BL-chains into irreducible hoops . . . . . . . . . . . . . . . . . . . 373
3.3 Decomposition into regular and dense elements. . . . . . . . . . . . . . . . . . . . . . . 379
3.4 Poset products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
3.5 BL-algebras as subdirect poset products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
3.6 BL-algebras as MV-product algebras with a conucleus . . . . . . . . . . . . . . . . 384
4 Classes generating the variety of BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
4.1 General facts about ordinal sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
4.2 Generation by nite members . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
4.3 Generation by t-norm BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
4.4 Generation by single BL-chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
4.5 Generating the variety of SBL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
4.6 Generating classes and completeness theorems . . . . . . . . . . . . . . . . . . . . . . . 395
5 Varieties of BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
5.1 Varieties generated by t-norm algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
5.2 Varieties generated by regular BL-chains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
5.3 Canonical representation of regular BL-chains . . . . . . . . . . . . . . . . . . . . . . . . 404
5.4 Axiomatizing the variety generated by a nite BL-chain . . . . . . . . . . . . . . . 409
5.5 Axiomatizing the variety generated by an arbitrary nite BL-algebra. . . . 412
6 Amalgamation, interpolation and Beths property in BL. . . . . . . . . . . . . . . . . . . . 413
6.1 General facts about interpolation and Beths property . . . . . . . . . . . . . . . . . 413
6.2 Amalgamation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
6.3 Deductive interpolation and Beths property, and strong amalgamation . . 422
6.4 Craig interpolation and implicative Beth property . . . . . . . . . . . . . . . . . . . . . 424
7 Completions of BL-algebras. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
7.1 Negative results about completions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
7.2 k-subvarieties and classication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
7.3 Completions in k-subvarieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
7.4 Canonical extensions of BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434
7.5 Canonical extensions of k-subvarieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
8 Further investigations of BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
8.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
8.2 Generalizations of the notion of BL-algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 442
8.3 Implicative subreducts of BL-algebras and of their generalizations. . . . . . 445
8.4 Further results about BL-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
9 Historical remarks and further reading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
INDEX 455
Chapter I Introduction to Mathematical
Fuzzy Logic
LIBOR B

EHOUNEK, PETR CINTULA, AND PETR H

AJEK
This chapter provides an introduction to the eld of mathematical fuzzy logic, giving
an overview of its basic notions and results.
Similarly as in other branches of logic, formal systems of fuzzy logic can be strati-
ed according to the complexity of their formal means. In this way we can distinguish
propositional, rst-order, higher-order, modal, etc., fuzzy logics, which closely parallel
their classical counterparts. In the rst section of this chapter we shall start with the
simplest layer of propositional fuzzy logics, whose language consists of propositions
compounded by propositional connectives, and focus on the paradigmatic case of logics
based on continuous t-norms.
In Section 2 we indicate the directions in which the apparatus of fuzzy logic intro-
duced in the rst section can be extended. We present a rich landscape of mathematical
fuzzy logic, inhabited by dozens of different propositional logical systems and their
classes, with complex interrelations and interesting metamathematical properties.
The multitude of fuzzy logics introduced in the rst two sections calls for a general
unifying (meta)theory. Section 3 rst studies the position of fuzzy logics in the logical
landscapei.e., their relationship to well-known broader families of propositional logics
(such as substructural or algebraizable) as well as to particular prominent non-classical
logics. Then the section briey surveys characteristic properties of fuzzy logics, both
those shared with the aforementioned broad classes of logics and those particular to
fuzzy logic.
Section 4 then surveys basic metamathematical properties of propositional fuzzy
logics. Finally, Section 5 presents an introduction to fuzzy logic of the rst (and higher)
order.
In this chapter we adopt a didactic stance: we proceed from simple concrete promi-
nent fuzzy logics to their numerous generalizations and nally to an abstract theory of
mathematical fuzzy logic. This naturally gives rise to a certain redundancy and repeti-
tions in denitions and theorems, but we feel that this price is necessary for upholding
Comenius maxim from the simple to the complex.
Particular topics in mathematical fuzzy logic are explained in detail in subsequent
chapters of this Handbook, to which this introduction frequently refers. All proofs of
theorems (except one) are omitted in this Introduction; we refer the reader to the spe-
cialized chapters of this Handbook and the literature listed in the References.
2 Libor B ehounek, Petr Cintula, and Petr H ajek
1 Propositional logics of continuous t-norms
Any non-classical propositional logic has many facets, which can roughly be divided
between syntactical ones (e.g., Hilbert or Gentzen axiomatic systems and their induced
sets of theorems and provability relations) and semantical ones (e.g., general or intended
algebraic semantics, game-theoretical semantics, relational semantics, etc., and their
induced set of tautologies and semantical consequence relations). The need to formulate
and prove general theorems on classes of logics forces us to choose one of these facets
for a formal denition of logic as a mathematical object. Usually any of these facets
determines the other ones uniquely (in some xed setting), thus the selection is mainly a
matter of preference. Here we have decided to identify logics (as mathematical objects)
with consequence relations, presented either syntactically (as the provability relation of
some axiomatic system) or semantically (as the consequence relation of some particular
semantics).
In this section we shall rst present the traditional, standard semantics of a certain
prominent class of fuzzy logics (namely, those based on continuous t-norms), and dene
a consequence relation induced by this semantics; then we shall proceed to the axiomatic
systems corresponding to this consequence relation. As we shall see, a straightforward
denition of the semantic consequence relation would lack certain desirable metamath-
ematical properties, including axiomatizability by means of nitary derivation rules.
In order to keep the correspondence between the syntactic and semantic facets of these
logics, the denition of the semantic consequence relation will therefore have to be mod-
ied to ensure its nitarity. After presenting axiomatic systems for prominent logics of
continuous t-norms, we shall briey hint at some of their metamathematical properties,
which will be treated in more detail in Section 4 of this chapter as well as in further chap-
ters of this Handbook. In subsequent sections we shall then introduce broader classes of
fuzzy logics and describe their properties in more detail.
1.1 Standard semantics of t-norm fuzzy logics
Mathematical fuzzy logic generalizes bivalent Boolean logic to larger systems L
of truth values, typically the real unit interval [0, 1]. Thenow classicalexposition
of [83] starts with certain natural constraints on the semantics of conjunction and other
propositional connectives, which are devised in such a way as to give rise to a well-
designed propositional and predicate calculi. These constraints on the semantics of
propositional connectives and the [0, 1]-valued semantics (with respect to which many
important systems of mathematical fuzzy logic are sound and complete) distinguishes
mathematical fuzzy logic from the study of other many-valued logics.
The most fundamental assumption of (mainstream) mathematical fuzzy logic is that
of truth-functionality of all propositional connectives. That is, each n-ary propositional
connective c is semantically interpreted by a function F
c
: L
n
L; the truth value of a
formula c(
1
, . . . ,
n
) is then dened as F
c
(x
1
, . . . , x
n
), where x
i
is the truth value of
the subformula
i
, for each i 1, . . . , n. In other words, the truth value of a formula
only depends on the truth values of its subformulae (from which it can be calculated by
the truth functions F
c
of the connectives),
1
independently of the meaning, structure, or
1
Observe that the truth functions F
c
: L
n
L generalize truth tables of two-valued Boolean logic, as the
Chapter I: Introduction to Mathematical Fuzzy Logic 3
other characteristics of the subformulae.
Truth-functionality is one of the design choices employed in (mainstream) mathe-
matical fuzzy logic. In other words, mathematical fuzzy logic as developed here denes
the meaning of complex formulae truth-functionally, and studies mathematical proper-
ties of the resulting truth-functional systems. Though it is certainly possible to develop
many-valued logics that are not truth-functional with respect to [0, 1] (or another in-
tended system of truth values),
2
it turns out that the truth-functional systems described
here are mathematically rich and demonstrably applicable to many real-life problems.
This by no means controverts the suitability and applicability of non-truth-functional
many-valued logics for many purposes and applications (including logical analysis of
natural language).
Another design choice selects the real unit interval [0, 1] as the intended (or stan-
dard) system of truth values. Consequently, the primary intended application of formal
fuzzy logic is to propositions that can be assigned a numerical quantity (normalized to
the unit interval), taken as their truth value in a semantical model. Nevertheless, the
inference laws of fuzzy logic generalize over a class of admissible assignments of truth
values, and so abstract from particular truth values. The rules of fuzzy logic are thus ap-
plicable to gradual propositions even in cases when their particular truth values cannot
be determined.
3
The truth values 0 and 1 are meant to represent, respectively, the (analogues of)
classical truth values false and true (to the full degree). The values between 0 and 1
represent intermediate grades of partial truth that can be assigned to propositions. Partial
truth is here understood as a technical term, referring just to the graded quality assigned
to propositions that is studied in formal fuzzy logic. Its philosophical interpretation is
left open here, as it is irrelevant to the mathematical study of the formal systems.
The usual order of reals is understood as representing the logical strength of
propositions, which decreases from 0 to 1: the larger the truth value of a proposition,
the truer (in the technical sense) is the proposition taken to be. This endows the standard
system [0, 1] of truth values with the structure of a complete linearly ordered lattice.
Further design choices of propositional fuzzy logic restrict possible truth-functions
of the logical connectives to those satisfying certain natural constraints. Following the
account of [83], we start with conditions on the truth function of conjunction (&). The
following conditions are required of : [0, 1]
2
[0, 1], for all x, y, z [0, 1]:
Commutativity: x y = y x.
Associativity: (x y) z = x (y z).
Monotonicity: if x x
t
, then x y x
t
y.
Unit: x 1 = x.
Continuity: is continuous on [0, 1]
2
.
latter can be regarded as functions F
c
: 0, 1
n
0, 1.
2
Such systems have been developed e.g. in [154, Part III], [71], or [5].
3
As is often the case with natural-language predicates: even though it is impossible to determine whether a
man of height 1.80 m is tall to degree 0.7 or 0.8, the laws of fuzzy logic are still applicable to the predicate tall.
4 Libor B ehounek, Petr Cintula, and Petr H ajek
Commutativity and associativity embody the idea that the truth value of a conjunction
does not depend on the order or bracketing of the conjoined propositions. Monotonicity
expresses the intuition that increasing the truth value of a conjunct should not decrease
the truth value of their conjunction. The neutrality of 1 is motivated by its interpretation
as full truth (conjunction with full truth should not change the truth value of a propo-
sition). The identity x 0 = 0, corresponding to the interpretation of 0 as full falsity,
already follows from the above conditions (see Theorem 1.1.5(1)). The condition of
continuity formalizes the intuitive idea that an innitesimal change of the truth value of
a conjunct should not radically change the truth value of the conjunction.
We could add further conditions required of (e.g., idempotence), but it has proved
useful to stop here, as the above conditions already yield a rich and interesting theory
and further conditions would be too limiting.
4
It turns out that functions satisfying the
rst four conditions from the above list have previously been studied in the theory of
probabilistic metric spaces under the name triangular norms (or t-norms for short):
DEFINITION 1.1.1. A binary function : [0, 1]
2
[0, 1] is a t-norm (or triangular
norm) if it is commutative, associative, monotone, and 1 is its unit element.
The above conditions on the interpretation of conjunction thus amount to the re-
quirement that is a continuous t-norm. An extensive study of the properties of t-norms
and continuous t-norms can be found in [123]. Here we only recall such properties
that are needed for the development of mathematical fuzzy logic (for proofs see [123]
or [83]).
DEFINITION 1.1.2. Let be a t-norm and x [0, 1]. Then we dene x
0
= 1 and
x
n+1
= x x
n
for each n N.
We say that x is an idempotent element of if x x = x.
We say that x is a nilpotent element of if there is a natural number n such that
x
n
= 0.
DEFINITION 1.1.3. We say that a t-norm is Archimedean if for each x, y (0, 1)
there is n N such that x
n
y. A continuous Archimedean t-norm is called strict if
0 is its only nilpotent element; otherwise it is called nilpotent.
We shall say that a t-norm is isomorphic to a binary operation : [a, b]
2
[a, b] if
there is a strictly increasing function f : [0, 1] [a, b] such that f(x y) = f(x) f(y)
for all x, y [0, 1].
DEFINITION 1.1.4. Let T be a system of triples (
i
, a
i
, b
i
) indexed by i I such that

i
is a t-norm for each i I and (a
i
, b
i
) [0, 1] are mutually disjoint open intervals.
Then the t-norm dened for each x, y [0, 1] as
x y =
_
a
i
+ (b
i
a
i
)
_
xa
i
b
i
a
i

i
ya
i
b
i
a
i
_
if x, y (a
i
, b
i
) for some i I
min(x, y) otherwise
4
For instance, the additional requirement of idempotence (x x = x for all x [0, 1]) would already
limit to a single possible function, namely the minimum: x y = min(x, y). However, there are contexts
in applications of mathematical fuzzy logic where interpretations of conjunction other than the minimum are
appropriate.
Chapter I: Introduction to Mathematical Fuzzy Logic 5
is called the ordinal sum of the t-norms
i
on the intervals (a
i
, b
i
) and denoted by

T.
If T is a nite system of triples (
1
, a
1
, b
1
), . . . , (
n
, a
n
, b
n
) and a
1
< b
1
< a
2
<
b
2
< < a
n
< b
n
, then the ordinal sum

T can also be written as

1

n
,
provided a
i
, b
i
are known from the context or are irrelevant (as all such ordinal sums
are isomorphic). In general, T induces an order on I by letting i j iff a
i
< a
j
;
in this sense we shall speak of the rst, last, second, etc. component of the ordinal sum

T (if they exist in the order ).

Without loss of generality, we shall assume further on that the union

iI
(a
i
, b
i
) of
intervals from T is dense in [0, 1] (this can always be achieved by adding to T all triples
(
G
, a
j
, b
j
) such that (a
j
, b
j
) is a maximal open interval of idempotents and
G
is the
minimum t-norm of Example 1.1.6 below).
THEOREM 1.1.5. Let be a t-norm. Then:
1. 0 x = 0 for all x [0, 1].
2. The set of nilpotent elements of is an interval [0, a) or [0, a] for some a [0, 1].
3. If x is an idempotent of , then x a = min(x, a) for any a [0, 1] and a b = a
for any a, b [0, 1] such that a x b.
4. The t-norm is continuous iff it is continuous in one variable, i.e., iff f
x
(y) =
x y is continuous for each x [0, 1]; an analogous claim is valid for lower
semi-continuity (also called left-continuity) and upper semi-continuity (or right-
continuity).
EXAMPLE 1.1.6. The following three are prominent examples of continuous t-norms:
The minimum, also called the G odel t-norm: x
G
y = min(x, y)
The product, also called the product t-norm: x

y = x y
The ukasiewicz t-norm: x

y = max(x + y 1, 0)
In subscripts and ordinal sums, we shall often write just G, , and instead of
G
,

,
and

, respectively.
The minimum is the only idempotent t-norm. The ukasiewicz and product t-norms
are prototypical (and unique modulo isomorphism, by Theorem 1.1.7(3) below) rep-
resentatives of, respectively, nilpotent and strict Archimedian t-norms. Moreover, all
continuous t-norms can be decomposed into ordinal sums of isomorphic copies of these
t-norms (the MostertShields theorem, see Theorem 1.1.7(5) below). The following
theorem summarizes the most important properties of continuous t-norms:
THEOREM 1.1.7. Let be a continuous t-norm. Then:
1. is Archimedean iff 0 and 1 are its only idempotents.
2. If is Archimedean, then it is nilpotent iff each x [0, 1) is a nilpotent element
of .
6 Libor B ehounek, Petr Cintula, and Petr H ajek
3. All continuous Archimedean t-norms are isomorphic to either the ukasiewicz t-
norm

(nilpotent t-norms) or to the product t-norm

(strict t-norms).
4. The ukasiewicz t-norm

is isomorphic to the product t-norm undercut in any

a (0, 1), i.e., to the function x y = min(a, x y) on [a, 1]
2
.
5. The set of idempotents of is a closed subset of [0, 1]. Its complement is there-
fore a countable union of disjoint open intervals (a
i
, b
i
), for i N. Since the
interval (a
i
, b
i
) contains no idempotents of , the restriction of to (a
i
, b
i
) is iso-
morphic to a continuous Archimedean t-norm (either strict or nilpotent), and so is
isomorphic either to

or

. If there is no i N such that x, y (a

i
, b
i
), then
xy = min(x, y) by Theorem 1.1.5(3). Consequently, every continuous t-norm is
isomorphic to an ordinal sum of the three t-norms from Example 1.1.6. Since also
any ordinal sum of continuous t-norms is a continuous t-norm, a t-norm is con-
tinuous iff it is isomorphic to an ordinal sum of
G
,

, and

(MostertShields
theorem).
The suitability of continuous t-norms for the truth functions of fuzzy conjunction
is afrmed by the following property, which provides a natural candidate for the truth
functions of implication.
THEOREM 1.1.8. For any continuous t-norm there is a unique binary operation

on [0, 1] such that for all x, y, z [0, 1],

z x y iff z x

y. (1)
The operation

is called the residuum (or residual) of . The following theorem

lists basic properties of the residua of continuous t-norms:
THEOREM 1.1.9. Let be a continuous t-norm,

its residuum. Then for any x, y

[0, 1]:
1. x

y = maxz [ x z y
2. x

is a right adjoint to the functor x in the lattice [0, 1] taken as a poset

category
3. x

y = 1 iff x y
4. 0

y = 1
5. 1

y = y
6. min(x, y) = x (x

y)
7. max(x, y) = min((x

y)

y, (y

x)

x)
Theorem 1.1.9(1) provides an explicit denition of

, implicitly dened by the

residuation condition (1) of Theorem 1.1.8. A residuated couple and

equips the
Chapter I: Introduction to Mathematical Fuzzy Logic 7
interval [0, 1] with the structure of a (complete divisible linear bounded integral com-
mutative) residuated lattice (see Denition 1.3.1). Because of the adjunction stated by
Theorem 1.1.9(2), the residuated pair ,

is often called an adjoint pair.

By Theorem 1.1.9(1), the residuum is the pointwise largest function such that x
(x

y) y; or equivalently by Theorem 1.1.9(3), such that x (x

y)

y equals
1 for all x, y [0, 1]. The latter condition can be understood as an internalized fuzzy
version of the modus ponens rule (if implication is interpreted as

and, in accordance
with previous motivation, interprets conjunction and 1 represents full truth). This
justies adopting

for the truth function of implication (), as it is the logically

weakest function that makes the fuzzy modus ponens generally valid (i.e., fully true for
any truth values of fuzzy propositions).
By Theorem 1.1.9(3), the full truth of fuzzy implication internalizes the ordering of
truth values by logical strength.
5
Theorem 1.1.9(4), which represents a fuzzy version
of ex falso quodlibet, then corresponds to the designation of full falsity as the logi-
cally strongest value, while Theorem 1.1.9(4) conrms full truth as a logically neutral
assumption.
EXAMPLE 1.1.10. The residua of the three prominent continuous t-norms
G
,

, and

are, respectively, the functions

G
(G odel implication),

(product implication,
also known as Goguen implication), and

(ukasiewicz implication), which have the

following values for x > y:
x
G
y = y
x

y = y/x
x

y = 1 x + y
(For x y, all residua evaluate to 1 by Theorem 1.1.9(3).) Observe that of these three,
only

is continuous, while

is discontinuous at the point 0, 0 and

G
at the
points x, x for x < 1.
By Theorems 1.1.9(6)(7), the operations of minimum and maximum are denable
in terms of and

. Since the operation of maximum extends the bivalent truth-table

for classical disjunction on 0, 1, it is a suitable truth-function for fuzzy disjunction ().
The operation of minimum, which like extends the classical truth-table for conjunction,
becomes a second candidate for the truth-function of fuzzy conjunction. Since it is
anyway denable from and

, it can be safely included in the propositional language

of fuzzy logic, playing the r ole of another conjunctive connective (). Fuzzy logics will
thus consider two conjunctions, one realized by a continuous t-norm (called strong or
residuated conjunction) and the other realized by the minimum (called weak, lattice, or
minimum conjunction). The two conjunctions coincide for the idempotent t-norm
G
,
but differ for all other t-norms; there are in fact deeper reasons for the presence of two
5
Notice that fully true fuzzy implication preserves the quality of being true at least to the degree of its
antecedent. This motivates Fonts [65] denition of truth degrees (as opposed to truth values) as intervals
[, 1], for any truth value [0, 1]. Theorem 1.1.9(3) then embodies the fact that fully true fuzzy implication
preserves truth degrees of propositions, and that a logically stronger proposition (i.e., one with the smaller truth
value) always implies a logically weaker one.
8 Libor B ehounek, Petr Cintula, and Petr H ajek
conjunctions in fuzzy logic (see Section 3.1). In analogy to strong and weak conjunction,
strong disjunction () can be introduced in some systems of fuzzy logic (see 2.2.2 or
Chapter VI). The disjunction interpreted by the maximum is then called weak (or lattice)
disjunction. The terms min-conjunction and max-disjunction are also used for the two
lattice connectives.
The truth-function of the equivalence connective () can suitably be dened as the
bi-residuum

of :
x

y = min(x

y, y

x), (2)
for all x, y [0, 1]. It is irrelevant whether weak or strong conjunction is used in (2), as
(x

y) (y

x) = min(x

y, y

x) for all x, y [0, 1], by Theorems 1.1.9(3)

and 1.1.9(5).
Fuzzy negation () can conveniently be realized by the truth function

: [0, 1]
[0, 1] dened as

x = x

0, which represents a fuzzy version of the reductio ad

absurdum. To distinguish it from an additional involutive negation introduced in some
systems of fuzzy logic (see Section 2.2.2), this negation is sometimes called residual
negation.
EXAMPLE 1.1.11. Residual negations for the three prominent continuous t-norms from
Example 1.1.6 come out as follows, for any x [0, 1]:

x = 1 x

G
x =

x =
_
1 for x = 0
0 for x > 0
Observe that

is continuous and involutive (i.e.,

x = x for all x), while

G
is
bivalent and coincides with

. The bivalent negation

G
=

is called strict (or

G odel) negation.
The function

in fact coincides with

G
(i.e., is strict), unless has a rst com-
ponent isomorphic to

on [0, a] for some a (0, 1], in which case

is isomorphic
to

on [0, a] and

x = 0 for x a. All other continuous t-norms (i.e., those whose

rst component is isomorphic to either
G
or

, and those that have no rst component

in the ordinal sum) have strict residual negation.
In this way, every continuous t-norm determines a family of truth functions ,

, min, max,

, which are natural candidates for the semantics of basic propo-

sitional connectives &, , , , , of fuzzy logic. These functions equip the interval
[0, 1] with an algebraic structure that can be used for a standard denition of algebraic
semantics for fuzzy logic. We shall call them t-algebras and denote them by [0, 1]

(see
Denition 1.1.12 below); later we show that t-algebras are special examples of the so-
called BL-algebras (see Denition 1.3.1). Notice that the constants 0 and 1 are added to
the signature of [0, 1]

, since 0 is used in the denition of

, and 1 is anyway denable

as

0 for any by Theorem 1.1.9(4); they will respectively represent the propositional
constants 0 for full falsity and 1 for full truth.
Chapter I: Introduction to Mathematical Fuzzy Logic 9
DEFINITION 1.1.12. For a continuous t-norm , we dene the t-algebra of as the
algebra
[0, 1]

= [0, 1], ,

, min, max, 0, 1,
where

is the residuum of and min, max are the minimum and maximum in the
usual order of reals. The operations

and

are dened in each t-algebra by setting

x

y = min(x

y, y

x) and

x = x

0 for all x, y [0, 1].

By convention, for any set K of continuous t-norms we denote the corresponding
set of t-algebras by K, and vice versa.
Now we can dene the syntax and standard semantics of logics based on continuous
t-norms as follows:
DEFINITION 1.1.13 (Syntax and standard semantics of logics of continuous t-norms).
The language / of the propositional fuzzy logic L
K
of the set K of continuous t-norms
consists of the propositional variables p, q, r, . . . , the binary propositional connectives
& (strong conjunction), (implication), (weak conjunction), (weak disjunction),
and (equivalence), the unary propositional connective (negation), and the propo-
sitional constants 1 (truth) and 0 (falsity).
The formulae of / are formed as usual; by convention, and and have the low-
est and unary connective(s) the highest priority. The (denumerable) set of propositional
variables of / will be denoted by Var and the set of all formulae of / by Fm
1
.
We shall use Greek letters , , , . . . for formulae and , , . . . for sets of formu-
lae. We shall also use the following abbreviations:
0

df
1 and
n+1

df

n
&, for
all n N.
A [0, 1]-evaluation of propositional variables is a mapping e: Var [0, 1]. For
any continuous t-norm , the evaluation e of propositional variables extends uniquely to
the -evaluation e

: Fm
1
[0, 1] of all formulae by the following recursive denition
(Tarski conditions), for any proposition variable p and any formulae , :
e

(p) = e(p)
e

( & ) = e

() e

()
e

( ) = e

()

()
e

( ) = min(e

(), e

())
e

( ) = max(e

(), e

())
e

( ) = e

()

()
e

() =

(e

())
e

(0) = 0
e

(1) = 1
10 Libor B ehounek, Petr Cintula, and Petr H ajek
By Theorem 1.1.9 and the denitions of

and

, we can obviously note that:

e

( ) = e

( & ( ))
e

( ) = e

((( ) ) (( ) ))
e

( ) = e

(( ) & ( ))
e

() = e

(( 0))
e

(1) = e

(0)
Thus the only necessary connectives in any logic of continuous t-norms are &, , and 0;
the remaining ones are denable using the denitions above.
It can be observed that given a continuous t-norm , some formulae obtain the truth
value 1 for any -evaluation of propositional variables; these can be regarded as the
tautologies of the fuzzy logic based on the continuous t-norm . We shall call such
formulae -tautologies. Some formulae turn out to be -tautologies for any continuous
t-norm in a given set K; these are tautologies of the logic L
K
:
DEFINITION 1.1.14. Let K be a set of continuous t-norms. A formula is a tautology
of the logic L
K
if e

() = 1 for any K and any -evaluation e

.
We shall use the expressions K-tautology, K-tautology, and L
K
-tautology as syn-
onyms for tautology of L
K
. We shall also write just -tautology if K = . If K is the
set of all continuous t-norms, then K-tautologies are also referred to as t-tautologies.
The semantic consequence relation of L
K
is the relation between sets of formulae
and formulae which preserves the truth value 1 (representing full truth) from premises
to conclusions under each evaluation of propositions in each t-algebra from K.
DEFINITION 1.1.15. A -evaluation e

is called a -model of a set of formulae if

e

() = 1 for all . A formula is a semantic consequence of in L

K
(written
[=
K
) if for each K, all -models e

of are -models of . The relation

[=
K
is called the semantical consequence relation of L
K
.
Sometimes we shall write [=
K
instead of [=
K
, and just [=

if K = .
In classical logic, the set of tautologies straightforwardly encodes the semantic
consequence relation, via the compactness and deduction theorems valid for this log-
ics (i.e., [= iff there is a nite subset
1
, . . . ,
n
of such that the formula

1

n
is a tautology). We shall see, however, that most fuzzy logics lack
this suitable form of the deduction theorem, and the semantical consequence given by
the standard (intended) semantics is seldom nitary (i.e., axiomatizable by deduction
rules with nitely many premises). Thus in many fuzzy logics, the relationship between
the set of tautologies and the consequence relation is not straightforward. In fact, in
many fuzzy logics, the set of tautologies does not determine the consequence relation,
as shown by the following example:
EXAMPLE 1.1.16. For a given set K of continuous t-norms, an alternative semantic
Chapter I: Introduction to Mathematical Fuzzy Logic 11
consequence relation [=
t
K
can be dened in the following way [20]:
6
[=
t
K
iff for each K and each -evaluation e

holds inf

() e

().
It can be easily shown that for each we have [=
K
iff [=
t
K
, but , [=
K
0
whereas , ,[=
K
0.
On the other hand, the consequence relation does determine the set of tautologies,
as obviously is a K-tautology if [=
K
. It is therefore more appropriate to identify
fuzzy logics (as mathematical objects) with their logical consequence relations rather
than just their sets of tautologies.
Unfortunately, as hinted above, the consequence relation [=
K
cannot be, in general,
axiomatized in a nitary way (i.e., by deduction rules with nitely many premises), as
follows from this counterexample:
EXAMPLE 1.1.17. Let be either or . We consider a theory = q p
n
[ n N.
We shall show that [=

p q, but for no nite

t
holds:
t
[=

p q.
We present the proof for = ; the proof for = is analogous. The rst claim is
proved by contradiction: assume that there is a -evaluation e such that e(p q) < 1
and e(q p
n
) = 1 for each n. Then obviously e(q) > 0, e(p) < 1, and e(q) e(p)
n
,
which is clearly impossible (since inf
nN
e(p)
n
= 0). The second claim: let m be the
maximal n such that q p
n

t
. Let e(q) = a and e(p) =
n

a, for arbitrary
a (0, 1). Then clearly e(p q) =
n

a < 1 and e(q p

n
) = 1 for each n m.
Even though it is not in general possible to axiomatize the consequence relation [=
K
by rules with nitely many premises, it turns out that its nitary companion [=
n
K
, i.e.,
the relation dened as
[=
n
K
iff there is a nite set
t
such that
t
[=
K
,
is nitely axiomatizable for any set K of t-algebras, with modus ponens (from and
infer ) as the only deduction rule (see the next subsection). Note that for nite
sets of premises the relations [=
K
and [=
n
K
coincide, thus in particular the tautologies of
these two consequence relations are always the same. On the other hand, the previous
example demonstrates that these two relations differ in general. In fact, even more can
be proven:
THEOREM1.1.18 ( [51, 104]). Let Kbe a set of t-algebras. Then [=
K
and [=
n
K
coincide
iff K = [0, 1]
G
.
In order to have the correspondence between the consequence relation and its ax-
iomatic presentation, by the logic of K we shall, therefore, call the nitary consequence
relation [=
n
K
:
6
Note, however, that this consequence relation has rather bad properties: e.g., the modus ponens rule is not
satised by [=

K
, unless K = G.
12 Libor B ehounek, Petr Cintula, and Petr H ajek
DEFINITION 1.1.19. The logic of a set of continuous t-norms K (or equivalently the
logic of K) will be identied with the nitary consequence relation [=
n
K
and denoted
by L
K
.
The logics of

, and
G
are, respectively, called ukasiewicz, product, and
G odel (fuzzy) logic and denoted by , , and G.
7
The logic of all continuous t-norms is
called basic logic and denoted by BL.
8
When a misunderstanding could arise, the logic
of continuous t-norms should better be called basic fuzzy logic or H ajeks basic logic.
The logic of all continuous t-norms whose residual negation is strict is called SBL (for
strict basic logic).
The innitary consequence relation [=
K
can be called the innitary logic of K, with
the particular cases of innitary ukasiewicz, G odel, product and (strict) basic logic.
For the logic L
K
, the t-algebras of K are called the standard L
K
-algebras. (For
general L
K
-algebras see Section 1.3.)
Obviously, the (innitary) logics (and so the sets of tautologies as well) of isomor-
phic continuous t-norms coincide, and similarly for sets of continuous t-norms differing
only by isomorphism. The converse, however, is not true, as for instance the logic BL
of all continuous t-norms coincides with the logic of any single t-norm which is an or-
dinal sum containing innitely many isomorphic copies of

having a rst component

isomorphic to

.
Thus there are sets K and K
t
of continuous t-norms such that K ,= K
t
, but L
K
=
L
K
; thus both K and K
t
are the sets of standard L
K
-algebras. The notion of standard
algebra, i.e., that of intended semantics of a given logic, is clearly a matter of choice that
is essentially ad hoc; therefore we will leave this discrepancy unresolved.
Note that for L , , G we have that L = L

if and only is isomorphic to

L
. Therefore, we postulate that the t-algebras [0, 1]

, [0, 1]

, and [0, 1]
G
(being the
intended semantics of ukasiewicz, product, and G odel logic) are the only algebras to
be called, respectively, the standard product algebra, the standard MV-algebra,
9
and the
standard G odel algebra.
10
Finally we postulate that all t-algebras [0, 1]

are standard
BL-algebras (or standard SBL-algebras if has strict residual negation).
Although the logics G, , and are incomparable in strength (see Section 1.2 for
examples of formulae in which they differ), it is not the case for all logics L

of particular
continuous t-norms. The following theorem displays some of the relationships between
the logics L
K
, for K a set of continuous t-norms (the rst part was proved in [104], the
remaining ones are folklore):
7
Two of these logics had been known before the advent of mathematical fuzzy logic. The logic was intro-
duced in 1930 by ukasiewicz and Tarski [129] (its three-valued variant already in 1920 by ukasiewicz [128])
and studied by Hay [106], Rose [162], and others. The logic G was implicitly dened in G odels 1932 pa-
per [74] and extensively studied esp. by Dummett [47] and Horn [115]. Hence they have been known in
the literature as the innite-valued (propositional) logic of ukasiewicz and Dummetts (or G odelDummetts)
propositional logic.
8
At least three other logics are also called basic logic in the literature; two related to intuitionistic
logic [165, 174] and one to relevant logics [164]
9
Algebras for ukasiewicz logic are traditionally called MV-algebras, where MV stands for many-
valued (since ukasiewicz logic was considered a paradigmatic example of many-valued logic).
10
This exception actually does not make a difference in the case of the standard G odel algebra, as G = L
K
if and only if K =
G
, and so [0, 1]
G
is the only standard G odel algebra by either denition.
Chapter I: Introduction to Mathematical Fuzzy Logic 13
Figure 1. The prominent logics of continuous t-norms and their position with respect
to classical logic (Bool) and intuitionistic logic (Int). Note that similar Hasse diagrams
in this chapter only capture the relative strength of the logics (increasing upwards), and
not the lattice relationships between the logics: e.g., SBL is not the intersection of the
logics G and .
THEOREM 1.1.20. Let K, K
t
be sets of continuous t-norms and ,
t
continuous t-
norms. Then the following properties hold:
1. There is a nite set of continuous t-norms

K such that L
K
= L
K
.
2. If K K
t
, then L
K
L
K
.
3. L

4. If has innitely many idempotents, then L

G.
5. BL L
K
, and L
K
or L
K
or L
K
G.
1.2 Axiomatic systems for logics of continuous t-norms
Now we shall deal with the axiomatic facet of fuzzy logics, following the exposition
of [83]. First let us recall some standard denitions pertaining to the notion of (nitary)
Hilbert-style calculus:
DEFINITION 1.2.1. A Hilbert-style calculus (or axiomatic system) is given by a set of
axioms and a set of derivation rules. Axioms are selected formulae in a given language.
Derivation rules are pairs consisting of a nite set of formulae (called the premises of
the rule) and a single formula (called the conclusion of the rule).
The axioms and derivation rules of a Hilbert-style calculus are usually given in
the form of schematai.e., with formulae containing placeholders to be replaced by
arbitrary formulae of the language. Particular axioms and derivation rules of the calculus
are then instances of the schema resulting from such a replacement. We say that an
axiomatization is nite if it has nitely many schemata of axioms and derivation rules.
We say that a calculus L
t
is an axiomatic extension of L if both have the same language
and L
t
arises by adding some set of axiom schemata (but no deduction rules) to L.
14 Libor B ehounek, Petr Cintula, and Petr H ajek
DEFINITION 1.2.2. A proof of a formula from a theory
11
T in a given Hilbert-style
calculus L is a nite sequence of formulae whose last member is and whose every
member is either (i) an axiom of the calculus, (ii) an element of T, or (iii) is derived
from previous members of the sequence by a derivation rule of the calculus (i.e., is the
conclusion of some derivation rule whose all premises are among the predecessors of
the formula in the sequence).
If there is a proof of fromT in L, we say that is provable fromT in the calculus L
(written T
L
). If
L
, we say that is a theorem of the axiomatic system L
(written just
L
). The relation
L
, i.e., the set of all pairs T, such that is
provable from T in L, is called the provability relation of the calculus L.
It turns out (see [29, 58, 83, 104]) that the logic BL of each set K of continuous
t-norms, there is a nite axiomatic system such that the logic L
K
(regarded here as
a nitary consequence relation, cf. Def. 1.1.19) coincides with the provability relation
given by this axiomatic system. An explicit formulation of this claim for the logics BL,
SBL, , , and G will be given in Theorem 1.2.4.
Recall that the language of BL consists of the primitive binary connectives and &,
and the truth constant 0. Further it contains the following derived connectives dened as:

df
& ( )

df
(( ) ) (( ) )

df
( ) & ( )

df
0
1
df
0
In this section we shall show that the following set of axioms,
12
(BL1) ( ) (( ) ( ))
(BL4) & ( ) & ( )
(BL5a) ( & ) ( ( )
(BL5b) ( ( ) ( & )
(BL6) (( ) ) ((( ) ) )
(BL7) 0
together with the deduction rule of modus ponens
(MP) from and infer ,
is a sound and complete axiomatization of the logic BL. Before we formulate this claim
formally, we shall discuss the r oles of these axioms and give axiomatic systems for some
of the stronger logics introduced in the previous section.
11
A theory is just a set of formulae. In the literature the term theory is sometimes used for sets of formulae
that are closed under the provability relation of a given logic. In this chapter we opted for the simpler meaning,
but in the in next one we use the latter.
12
Note that in fact we use axiomatic schemata, i.e., for an arbitrary formula , the formula 0 is an
instance of the axiom schema (BL7), and similarly for other axioms.
Chapter I: Introduction to Mathematical Fuzzy Logic 15
The axiom (BL1), also called sufxing, ensures the transitivity of implication (cf. the
relationship between and the ordering of truth values: see Theorem 1.1.9(3) and Sec-
tion 3.3). The axiom (BL4) ensures the commutativity of the dened min-conjunction ;
it is also called the axiom of divisibility, as together with the other axioms and deni-
tions it ensures the divisibility condition in BL-algebras (see Denition 1.3.1). The
mutually converse implications (BL5a) and (BL5b) express the residuation condition
(cf. Theorem 1.1.8(1) and Denition 1.3.1). The axiom (BL6), which is equivalent to
( ) ( ), expresses the property of prelinearity (cf. Denition 1.3.1). Fi-
nally, the axiom (BL7) is the ex falso quodlibet law. This is a minimal independent set of
axioms [26]. The numbering of axioms is due to the original numbering in [83], which
included two more axioms (later proved redundant, see [26]):
(BL2) &
(BL3) & &
The logics G, , and are axiomatized by adding some of the following axioms to
the axiomatic system of BL:
(G) &
()
() (( & ) )
Thus, G = BL+(G); = BL+(); and = BL+(). The axiom (G) expresses the
idempotence of the residuated conjunction (note that the converse implication is a spe-
cial case of (BL2)), and so its coincidence with the minimum conjunction. The axiom
() of double negation expresses the involutiveness of negation (cf. Example 1.1.11); its
converse implication is provable already in BL: see theorem (T
BL
38) on page 18. Prod-
uct logic can equivalently be axiomatized by adding the following two axioms to BL:
(S) ( )
(
S
) (( & & ) ( ))
The axiom (S) expresses the strictness of residual negation (Example 1.1.11). Adding
just the axiom (S) to BL yields the logic SBL of continuous t-norms with strict residual
negation (see Denition 1.1.19). The logic SBL extends BL and is extended by both
and Gas well as by all other logics L
K
where K is a set of t-norms with strict negations,
i.e., t-norms which are not of the form (cf. Example 1.1.11 and Theorem 1.1.20). In
SBL, the axiom (
S
) expresses the cancellativity of the t-norm: provided is non-zero
(which in SBL is expressed by ), it can be canceled from & & .
The logics of other continuous t-norms (or even sets of continuous t-norms) can
also be axiomatized by adding certain nite sets of axioms to BL; however, since most
of these axioms are not very intuitive, we do not present them here; they can be found
in [58]. Extending BL (or any logic of continuous t-norms) by the law of excluded
middle,
(LEM) ,
16 Libor B ehounek, Petr Cintula, and Petr H ajek
already yields classical logic.
The completeness theorem for logics of continuous t-norms can be formulated in
a compact way by stating the equality of the logic (dened semantically as nitary se-
mantic consequence relations, see Denition 1.1.19) and the corresponding syntactic
provability relation (of Denition 1.2.2):
THEOREM 1.2.3. Let L be any of , G, , BL, and SBL. Then for every theory and
formula holds:

L
iff [=
n
L
.
An expanded (traditional) formulation of this so-called nite strong standard com-
pleteness theorem for the logics of continuous t-norms is as follows:
THEOREM 1.2.4 (Finite strong standard completeness). Let L be one of the , G, ,
BL, SBL. Then the following are equivalent for any nite theory and a formula :
1.
L

2. e() = 1 for each standard L-algebra [0, 1]

and any -model e of .

Notice that Theorem 1.1.18 tells us that the restriction to nite theories can only
be dropped in the case of G odel logic. The fact that
L
implies the second claim
of Theorem 1.2.4 (usually called the soundness of the axiomatic system of L w.r.t. its
standard semantics) is easily obtained by simple computation in [0,1]. We will not show
the proof of the converse implication (usually called the completeness of the axiomatic
system of L w.r.t. its standard semantics) here;
13
for particular logics it can be found in
Chapter V, Chapter VI, Chapter VII of this Handbook. Nevertheless, in the next sub-
section we will show the proof of a weaker claim, namely the completeness w.r.t. the
(broader) class of linearly ordered L-algebras.
Let us now list some important formulae provable in the logics we have just dened,
and comment on their r ole and importance. We start with theorems of BL (which by
Theorem 1.2.4 coincide with t-tautologies) and list them with their customary names.
Most of them are proved (and others are easily derivable from those proved) in [83].
First, the logic BL proves the following simple properties of implication:
(T
BL
1) reexivity
(T
BL
2) ( ) weakening
(T
BL
3) ( ( )) ( ( )) exchange
(T
BL
4) ( ) (( ) ( )) prexing
(T
BL
5) (( ) ) assertion
13
For BL this was proved in [29] by showing the redundancy of two additional axioms needed in H ajeks
completeness proof [82]; for SBL in [29]; for ukasiewicz logic in [106] (a weaker result, showing only the
coincidence of tautologies and theorems of ukasiewicz logic, was proved before independently in [24, 163]);
for G odel logic in [47]; and for product logic in [99].
Chapter I: Introduction to Mathematical Fuzzy Logic 17
We continue with the properties of the residuated conjunction &. Note the rst
two theorem were originally included among the axioms of BL (see page 15), and are
therefore often referred to as (BL2) and (BL3):
(T
BL
6) & (&-form of) weakening
(T
BL
7) & & commutativity of &
(T
BL
8) ( ) ( & & )) monotonicity of &
(T
BL
9) ( & ) & & ( & ) associativity of &
(T
BL
10) ( & )
Next we deal with properties of min-conjunction and max-disjunction . Note
that the minimum conjunction differs from the residuated one &, i.a., by its idempo-
tence. Also note that the commutativity of is stated in axiom BL4.
(T
BL
11) &
(T
BL
12)
(T
BL
13) idempotency of
(T
BL
14) ( ) ( ) associativity of
(T
BL
15) ( ) ( )) monotonicity of
(T
BL
16) ( ) ( )
Theorems (T
BL
11) and (T
BL
10) allow us to obtain (by a double application of
modus ponens) a derived deduction rule of BL called -adjunction: ,
BL

(thus in particular: ,
BL
).
(T
BL
17)
(T
BL
18) idempotency of
(T
BL
19) commutativity of
(T
BL
20) ( ) ( ) associativity of
(T
BL
21) ( ) ( )) monotonicity of
(T
BL
22) ( ) ( )
(T
BL
23) ( ) ( ) prelinearity
Furthermore we list several forms of distributivity laws between the connectives:
(T
BL
24) ( ) ( ) ( ) distributivity of over
(T
BL
25) ( ) ( ) ( ) distributivity of over
(T
BL
26) ( ) ( ) ( ) distributivity of over
(T
BL
27) ( ) ( ) ( ) distributivity of over
(T
BL
28) & ( ) ( & ) ( & ) distributivity of & over
18 Libor B ehounek, Petr Cintula, and Petr H ajek
(T
BL
29) & ( ) ( & ) ( & ) distributivity of & over
(T
BL
30) ( ) ( ) ( ) distributivity of over
(T
BL
31) ( ) ( ) ( ) distributivity of over
The following properties of truth constants 0 and 1 are provable in BL:
(T
BL
32) & 1 1 is the unit of
(T
BL
33) & 0 0 0 is the annihilator of
(T
BL
34) (1 ) push and pop
Finally we give some BL-provable properties of residual negation. Note that the
properties of negation are weaker than those known from the classical logic: in fact,
they more resemble those of intuitionistic logic; later we show that BL indeed does not
prove their classical variants.
(T
BL
35) ( ) ex falso quodlibet
(T
BL
36) ( &) law of non-contradiction
(T
BL
37) ( ) ( ) antitonicity of
(T
BL
38) weak double negation law
(T
BL
39) ( ) ( ) weak De Morgan law
(T
BL
40) ( ) ( ) weak De Morgan law
Observe that the monotonicity properties proved for all the connectives can be used
to prove the corresponding properties of congruence w.r.t. equivalencee.g.:
(T
BL
41) ( ) (( ) ( )) right-congruence of
(T
BL
42) ( ) (( ) ( )) left-congruence of
(T
BL
43) ( ) (( & ) ( & )) congruence of &
Thus proceeding inductively (and using modus ponens) we obtain the following the-
orem, which later plays a crucial r ole in the proof of the completeness theorem (see
Denition 1.3.8).
THEOREM 1.2.5. Let be a formula and let
t
be a formula resulting from by
replacing some occurrences of its subformula by a formula . Then:

BL

t
. (Cong)
The theorem holds analogously in all extensions of BL. By Theorem 1.2.5, we
also obtain: ,
BL

t
. Provably equivalent formulae can thus be freely inter-
substituted in the context of any formula without changing its provability.
Now let us show some theorems of the prominent axiomatic extensions of BL:
THEOREM 1.2.6. The logic SBL can be equivalently axiomatized by adding any of the
following (SBL-provable) formulae to the axioms of BL:
Chapter I: Introduction to Mathematical Fuzzy Logic 19
(T
SBL
1) weak law of excluded middle
(T
SBL
2) ( )
(T
SBL
3) ( & ) ( )
EXAMPLE 1.2.7. The axiom (S), and thus (by the previous theorem) also the formulae
(T
SBL
1)(T
SBL
3), are not provable in ukasiewicz logic. Indeed, consider any evalua-
tion e such that e() =
1
2
: then e

() =
1
2
, thus e

(( )) =
1
2
,= 1.
THEOREM 1.2.8. ukasiewicz logic can be equivalently axiomatized by adding any of
the following -provable formulae to the axioms of BL:
(T

1) ( ) ( ) converse contraposition law

(T

2) (( ) ) (( ) ) Wajsberg axiom
(T

3) & ( ) denability of &

(T

4) ( ) ( &) denability of
(T

5) ( ) De Morgan law
(T

6) ( ) De Morgan law
EXAMPLE 1.2.9. The axiom (), and thus (by the previous theorem) also the formulae
(T

1)(T

6), are provable neither in G odel logic nor product logic (and also not in
SBL and BL). Indeed, consider any evaluation e() =
1
2
. In both G and we obtain
e() = 1, thus e( ) =
1
2
,= 1.
Notice that the theorems (T

3)(T

4) allow us to further reduce the set of basic

connectives in ukasiewicz logic, to either , 0, or &, , or , . The last one
is the set of connectives used by ukasiewicz and Tarski in their original axiomatiza-
tion of ukasiewicz logic (see, e.g., [83]); they used an equivalent system of axioms,
composed of (BL1), (T
BL
2), (T

1), (T

2), and modus ponens.

So far we have seen four notable axiomatic extensions of BL, namely, , G, , and
SBL, shown by Theorem 1.2.4 to be the logics of certain (sets of) continuous t-norms.
Clearly not all axiomatic extensions of BL are t-norm based in the latter sense: for
example, even though classical (Boolean) logic is the axiomatic extension of BL by the
law of excluded middle, it is not complete w.r.t. any set of t-algebras, since for
any -evaluation e in any t-algebra [0, 1]

, e( ) = 1 iff e() = 0 or e() = 1.

All axiomatic extensions of BL enjoy a so-called local deduction theorem (for a
proof see Chapter II):
THEOREM 1.2.10 (Local deduction theorem). Let L be an axiomatic extension of BL.
Then for each theory and formulae , it holds:
T,
L
iff there is a natural n such that T
n
.
The word local in the name of the theorem refers to the dependence of the multiplic-
ity n on the right-hand side of the equivalence on the formula . An axiomatic extension
L of BL enjoys the classical deduction theorem (T,
L
iff T
L
) if and
20 Libor B ehounek, Petr Cintula, and Petr H ajek
only if L extends G odel logic. Thus (see Theorem 1.1.20) G odel logic is the only logic
of continuous t-norms which enjoys the classical deduction theorem.
The following two properties of axiomatic extensions of BL, the proof by cases
property and the semilinearity property, were implicitly proved (though not mentioned
by these names) in [83]. While the former is widely used and studied in the general
logical literature, the latter was rst explicitly formulated in [34] and called there the
prelinearity property; later in [41] it was renamed to its current name (for reasons for
this change see the comments preceding Denition 3.3.9). The semilinearity property
plays a crucial r ole in the proof of the so-called linear completeness of fuzzy logics (see
Theorem 1.3.12), and arguably in whole Mathematical Fuzzy Logic (cf. [13]); therefore
we sketch its proof here.
THEOREM 1.2.11. Let L be an axiomatic extension of BL. Then L satises the follow-
ing two properties, for each theory and formulae , , :
1. The proof by cases property (PCP):
,
L
whenever ,
L
and ,
L

2. The semilinearity property (SLP):

L
whenever ,
L
and ,
L

Proof. The rst claimis a straightforward application of the local deduction theoremand
theorems of BL: Fromthe assumptions we obtain some mand n such that
L

n

and
L

m
, and so by (T
BL
27) we have
L

n

m
. If we show that
BL proves: ()
n+m

n

m
the proof is done. Using (T
BL
29) and associativity
and commutativity of we obtain ()
n+m

_
m+n
i=0

i
&
m+ni
. Observe that for
each i we have
i
&
m+ni

n
or
i
&
m+ni

m
(using (BL2) and (T
BL
8)),
Theorems (T
BL
17) and (T
BL
29) complete the proof.
The second claim is then a simple corollary of theorem (T
BL
23), ( )
( ).
It can be proved that if ,
L
, then there is a theory
t
such that
t
,
L
and

t
,
L
for each /
t
.
14
Note than in classical logic such a theory
t
would be
complete (i.e., for each formula , either of would be provable from). In fuzzy
logic, however, only a weaker statement can be proved:
L
and
L

for each , .
15
The theories satisfying this property have in [83] and subsequent papers
been called complete theories, since the r ole they play in the proof of a completeness
theorem for fuzzy logics is similar to that of complete theories in the completeness proof
for classical logic. However, this name can be misleading, as even though this property
entails completeness (in the above sense of or for each ) in classical
logic, in fuzzy logics it differs from this notion of completeness. Thus in recent papers,
14
The proof consists in a simple application of Zorns lemma, see Chapter II.
15
The proof is easy: Assume otherwise, then from the maximality of

we get

,
L
and

,
L
thus the semilinearity property will give us

L
a contradiction.
Chapter I: Introduction to Mathematical Fuzzy Logic 21
the term linear theory has been coined for this notion (for the rationale of the term
linear see Lemma 1.3.9).
Besides linear theories, we also dene the notion of prime theory, which is well
known, e.g., from the study of super-intuitionistic logics.
DEFINITION 1.2.12 ( [34]). We say that a theory is
Linear in L if for each , it holds that
L
or
L
.
Prime in L if for each , it holds that
L
implies
L
or
L
.
It can be shown [175] that the notions of prime and linear theory coincide in any
axiomatic extension of BL (but they differ, e.g., in intuitionistic logic where, due to the
disjunction property, the set of theorems is a prime, but clearly not linear theory). Later
we will see that this equality is one of the characteristic properties of fuzzy logics (in
a certain setting, see Chapter II for details). As hinted above, the following theorem
of [83] is crucial for the proof of the so-called linear completeness of fuzzy logics,
described in the next Section 1.3.
THEOREM 1.2.13 (Linear / prime extension principle). Let L be an axiomatic extension
of BL. Let be a theory and a formula such that ,
L
. Then there is a linear
(prime) theory
t
such that
t
,
L
.
1.3 Algebraic semantics
In this subsection we introduce a more general algebraic semantics of the logics we
have considered so far. It will be, in a certain specic sense, the most general algebraic
semantics and so we can easily prove it is also a complete semantics. The proof of the so-
called general completeness, together with its strengthening, the linear completeness, is
the crucial rst step in the proof of the standard completeness mentioned in the previous
section. Unlike in the standard case, we are able to showthat the semantical consequence
relations given by general and linear semantics are nitary, and the axiomatic systems
presented in the previous subsection are complete w.r.t. these semantics for all (possibly
innite) sets of formulae.
General algebras for the logic BL are called BL-algebras [83]. The notion of a
BL-algebra can be dened in several equivalent ways; the denition given below puts
BL-algebras in the context of the well known and deeply studied class of residuated
lattices (see, e.g., [67]). For simplicity we use the same symbols for the connectives of
L and the operations in the algebras, although we keep using the alternative symbols,
introduced in Section 1.1, for realizations of connectives in the standard semantics. If
necessary, they can be disambiguated by superscripting the name of the algebra to the
operation symbol.
DEFINITION 1.3.1 ( [67]). A bounded integral commutative residuated lattice, a bicrl
for short, is an algebra A = A, &, , , , 0, 1 such that:
1. A, , , 0, 1 is a bounded lattice
2. A, &, 1 is a commutative monoid
22 Libor B ehounek, Petr Cintula, and Petr H ajek
3. is the residuum of &, i.e., for each x, y, z A holds: x&y z iff x y z.
The class of all bicrls will be denoted by B"C1'. The operations and are
dened in each bicrl by setting for all x, y [0, 1]:
x y = (x y) (y x)
x = x 0
We say that a bicrl Ais
Linearly ordered if the order induced by its lattice reduct is total
Prelinear if it satises the identity (x y) (y x) = 1 for all x, y A
Divisible if it satises the identity x y = x & (x y) for all x, y A.
Divisible prelinear bicrls are simply called BL-algebras and their class is denoted
by B'; linearly ordered BL-algebras are called BL-chains.
In earlier literature on mathematical fuzzy logic (prominently in [83], where the reader
can also nd the proofs of the following theorems), bicrls were called simply residuated
lattices.
If we assume that an algebra A = A, &, , , , 0, 1 satises conditions 1 and 2
of the denition of bicrl, then the residuation condition 3 is equivalent to the following
pair of identities:
(x & (x z y)) z = z
(x (x & y z)) y = y,
for all x, y, z A. Thus both B"C1' and B' are varieties of algebras. Before we show
the relation of the just dened class of BL-algebras and the logic BL, let us give some
of its basic properties and dene its important subvarieties related to other fuzzy logics
introduced in the previous section.
THEOREM 1.3.2. Let Abe bicrl. Then:
1. x y iff x y = 1
2. x = y iff x y = 1
It is very easy to check that every t-algebra is a BL-algebra; in fact, even the con-
verse claim can be proved:
THEOREM 1.3.3. A bicrl A = [0, 1], &, , min, max, 0, 1 is a BL-algebra if and
only if & is a continuous t-norm and is its residuum.
For any bicrl Awe dene the notion of A-evaluation and A-model in the same way
as in the case of t-algebras (see Denitions 1.1.131.1.15). The next theorem shows that
BL-algebras form a sound semantics of the logic BL.
Chapter I: Introduction to Mathematical Fuzzy Logic 23
THEOREM 1.3.4 (Soundness). For any theory and a formula such that
BL

we have e() = 1 for each BL-algebra Aand any A-model e of .
In fact even more can be proved, namely that the BL-algebras are the maximal sound
semantics of BL among all algebras with the same signature (see Section 3.3 for details).
In particular, if A is a bicrl, then A is a BL-algebra if and only if for any theory and
a formula such that
BL
holds that e() = 1 for each A-model e of .
DEFINITION 1.3.5. A BL-algebra Ais called:
An SBL-algebra if it satises the identity x x = 0 for all x A
An MV-algebra
16
if it satises the identity x = x for all x A
A product algebra if it satises the identity x ((x x & y) y) = 1 for all
x, y A
A G odel algebra if it satises the identity x & x = x for all x A.
The corresponding varieties of algebras will be denoted by SB', MV, P and G.
Notice that in the denition of product algebras we simply use the dening ax-
iom (), and put it always equal to 1. Analogously the dening identities of SBL-, MV-
and G odel algebras could be equivalently replaced by (x x) = 1, x x = 1,
and x x & x = 1 respectively, i.e., by the axioms (S), (), and (G). Conversely, each
subvariety of BL-algebras determines an axiomatic extension of BL. We give a general
denition:
DEFINITION 1.3.6. Let L be an axiomatic extension of BL. We dene a subvariety V
L
of BL-algebras as those satisfying, for each L-theorem , the identity = 1 for all
values of propositional variables. The algebras from V
L
are called L-algebras.
Let L be an axiomatic extension of BL by a set of axioms /; then we can simply
prove (by induction on the complexity of the proof) that Ais an L-algebra iff it satises,
for each /, the identity = 1 for all values of propositional letters. In particular,
for the logics from the previous subsection we have: SB' = V
SBL
, MV = V

, P = V

,
and G = V
G
. Consequently, we will sometimes write ' instead of V
L
.
THEOREM1.3.7 (General completeness). Let L be an axiomatic extension of BL. Then
the following are equivalent for each theory and formula :
1.
L

2. e() = 1 for each L-algebra Aand any A-model e of .
16
As mentioned in footnote 9 on p. 12, we speak about MV-algebras rather than ukasiewicz algebras for
the historical reasons. Strictly speaking the MV-algebras are usually presented in a different signature, but can
be shown termwise equivalent to our denition of MV-algebras, cast as a subvariety of B"C1'. MV-algebras
can also be, termwise equivalently, presented as Wajsberg algebras, with only one binary connective and
the truth constant 0. (Recall that we have seen in Theorem 1.2.8 that in ukasiewicz logic & can be dened
and the remaining connectives of bicrl are denable already in BL.)
24 Libor B ehounek, Petr Cintula, and Petr H ajek
Below we give a hint of the proof of this fundamental theorem. First we need one
important denition and a lemma.
DEFINITION 1.3.8 (LindenbaumTarski algebra). Let L be an axiomatic extension
of BL and a theory. For every formula , we dene the set
[]

= [
L
.
The LindenbaumTarski algebra of the theory , denoted by LT

, is the algebra with

the domain []

[ a formula of BL and the operations dened as follows:

[]

&
LT

[]

= [ & ]

[]

LT

[]

= [ ]

[]

LT

[]

= [ ]

[]

LT

[]

= [ ]

0
LT

= [0]

1
LT

= [1]

The soundness of the latter denition of operations follows from Theorem 1.2.5.
The proof of the following lemma is straightforward.
LEMMA 1.3.9. Let L be an axiomatic extension of BL and a theory. Then:
1. LT

is an L-algebra
2. LT

is an L-chain iff is a linear theory

3. The LT

-evaluation e() = []

is an LT

-model of .
Proof of Theorem 1.3.7. One implication is Theorem 1.3.4. We prove the converse im-
plication counterpositively: assume that ,
L
and consider the LT

-evaluation
e() = []

. We know that e is an LT

-model of and obviously e() ,= 1

LT

(as otherwise 1, and so a contradiction).

Next we move to a characteristic topic of mathematical fuzzy logic: completeness
w.r.t. linearly ordered algebras of truth values. Since all t-algebras are indeed linear, we
have already shown a form of such completeness; however, recall that we have done so
for nite theories only. Historically speaking, there are two possible approaches: one
using the notion of subdirect product and another using the notion of linear theory. The
next chapter will study the mutual relationship of these two approaches and show their
equivalence; here we give the main theorems of both approaches.
DEFINITION 1.3.10. We say that a bicrl A is representable as a subdirect product of
the family of bicrls A
i
[ i I if there is an injective homomorphism from A into
the direct product

iI
A
i
such that for every i I, the composition of with the i-th
projection,
i
, is surjective. In this case, is called a subdirect representation of A;
it is called nite if I is nite.
A bicrl Ais (nitely) subdirectly irreducible if for every (nite) subdirect represen-
tation with a family A
i
[ i I there is i I such that
i
is an isomorphism.
Chapter I: Introduction to Mathematical Fuzzy Logic 25
THEOREM 1.3.11 (Linear subdirect representability). Let L be a axiomatic extension
of BL. Then any L-algebra is a subdirect product of a set of L-chains.
THEOREM 1.3.12 (Linear completeness). Let L be an axiomatic extension of BL. Then
the following are equivalent for each theory and formula :
1.
L

2. e() = 1 for each L-algebra Aand A-model e of
3. e() = 1 for each L-chain Aand A-model e of .
Proof. We can either (straightforwardly) use the previous theorem to show that 3. im-
plies 2., or we can counterpositively prove that 3. implies 1.: We start as in the proof of
Theorem 1.3.7, and using Theorem 1.2.13 we obtain a linear theory
t
such that

t
,
L
. From Lemma 1.3.9 we know that LT

is an L-chain, and again the LT

-
evaluation e() = []

is an LT

-model of such that obviously e() ,= 1

LT

.
Theorem 1.3.11 can be strengthened in the following way:
THEOREM1.3.13 ( [145]). Let L be an axiomatic extension of BL. Then any L-algebra
is a nitely subdirectly irreducible iff it is an L-chain.
By a well known algebraic fact, the class of subvarieties of B' as well as the set
of axiomatic extensions of BL form lattices. In Denition 1.3.6 we have introduced the
mapping V
L
that assigns to each axiomatic extension of BL the corresponding subva-
riety of B'. This mapping is in fact a dual isomorphism of these two lattices, whose
inverse
V
assigns to V the extension of BL by the set of axioms /
V
= [
the identity = is valid in V. This fact is formalized in the next theorem, which for
BL-algebras is folklore; in the general framework of algebraizable logic it was proved
by of Blok and Pigozzi in [18] (see Section 3.3, and Theorem 3.3.8 in particular, for
more details on this topic).
THEOREM 1.3.14. Let L, L
t
be axiomatic extensions of BL and let U, U
t
be subvari-
eties of B'. Then:
1. U = V

U
and L =
V
L
2. U U
t
=
U

U
and L L
t
= V
L
V
L
.
2 Variations of basic propositional fuzzy logics
In Section 1 we introduced basic systems of fuzzy logic based on continuous t-
norms. In this section we shall indicate the directions in which the apparatus of fuzzy
logic can be extended. The principal ways of altering the logics of continuous t-norms
are the following:
Discarding axioms or rules of a fuzzy logic, enlarging thus the class of its models
Adding new connectives, thus increasing the expressive power of the logic
26 Libor B ehounek, Petr Cintula, and Petr H ajek
Discarding connectives present in a fuzzy logic and passing to its fragments
Adding axioms or rules to obtain stronger logics with more specic models.
These types of alterations can further be combinede.g., some axioms can rst be dis-
carded and some others (possibly weaker or incomparably strong) added back, connec-
tives can be added to a previously weakened logic, etc. Variations of the above types
generate a rich landscape of mathematical fuzzy logic, inhabited by dozens of different
logical systems and their classes, with complex interrelations and interesting metamath-
ematical properties.
In the subsections of the present section, we shall briey introduce the main fuzzy
logics resulting from these modications of the logics of continuous t-norms. The way
of extending propositional fuzzy logics to predicate fuzzy logics of the rst or higher
order is postponed to Section 5.
The new fuzzy logics will mainly be introduced by their standard semantics, or by
syntactic manipulation with their language or axiomatic system. In order to avoid many
repetitive denitions, we shall employ the following terminological convention (for a
general formal denition of L-algebras and other notions mentioned in this convention
see Section 3.3):
CONVENTION 2.0.1. For all newly introduced logics L, the corresponding algebras
for which L is sound will be called L-algebras. A terminological exception are modica-
tions of ukasiewicz logic, where (subscripted) MV- is used instead of - (cf. footnote 9
on p. 12).
More precisely, we say that Ais an L-algebra if
L
implies that any A-model
of all formulae from is an A-model of , where an evaluation e is an A-model of if
e() = 1 unless said otherwise.
17
We say that an L-algebra Ais:
An L-chain if the order determined by the lattice connectives of L is linear
18
A real-valued L-algebra if its lattice reduct is the real interval, either closed [0, 1]
or half-open (0, 1], with the usual ordering of reals
A (nitely) subdirectly irreducible L-algebra if it cannot be obtained as a non-
trivial (nite) subdirect product of a family of L-algebras (for a formal denition
in the case of BL-algebras see Denition 1.3.10).
In the previous section we have seen that particular real-valued L-algebras may be
specied as the standard L-algebras. These algebras are the intended semantics of the
logic. By default, all real-valued algebras will be considered standard. In some cases,
however, we may want to select a narrower class of real-valued algebras as standard: for
instance, if all real-valued L-algebras are mutually isomorphic (as is the case, e.g., in G,
, and , see p. 12), we pick one particular representative.
17
A different specication of A-models is employed, e.g., in Section 2.1.2. As in the previous section, the
class of L-algebras will usually admit some explicit (often equational) description.
18
If the logic L does not possess lattice connectives, the order can be dened by means of implication, see
Section 3.3.
Chapter I: Introduction to Mathematical Fuzzy Logic 27
CONVENTION 2.0.2 (Standard algebras). Unless specied otherwise, all (and only)
real-valued L-algebras are the standard L-algebras.
The logic need not be complete with respect to its standard algebras;
19
if it is, we
speak about the standard completeness of the fuzzy logic. As we have seen in Sec-
tions 1.11.2, the logics and enjoy standard completeness for nite theories only;
G odel logic, on the other hand, enjoys standard completeness even for innite theories.
This motivates the distinction made in the following convention.
CONVENTION 2.0.3 (Standard completeness). We say that L enjoys nite strong stan-
dard completeness if the following conditions are equivalent for each formula and
each nite theory T:
T
L

For each standard L-algebra Aand each A-model e of T, e is an A-model of .
If the equivalence holds for all theories, we say that L enjoys (innitary-)strong standard
completeness, and we speak just about (weak) standard completeness if the equivalence
holds for T = .
All logics L that will be introduced in this section enjoy several important prop-
erties that have been discussed in the previous section for logic BL and its axiomatic
extensions:
The derivability of the congruence rule (cf. Theorem 1.2.5)
The proof by cases property PCP and the semilinearity property SLP (cf. Theo-
rem 1.2.11)
The linear (or prime) extension property (cf. Theorem 1.2.13)
The general completeness theorem (i.e., completeness w.r.t. all L-algebras, cf.
Theorem 1.3.7)
The linear completeness theorem (i.e., completeness w.r.t. all L-chains, cf. Theo-
rem 1.3.12)
The linear subdirect decomposition property (cf. Theorems 1.3.11 and 1.3.13).
Therefore we are not going to mention these properties repeatedly, and will omit stating
that each of the upcoming logics possesses them. On the other hand, the logics will differ
in what form of the (local) deduction theorem (cf. Theorem 1.2.10) holds for them; also
the form of standard completeness (see Convention 2.0.3) will change from logic to
logic.
Finally, let us remark that the derivability of the congruence rule and general com-
pleteness are properties possessed by a broad range of logics (namely the so-called
19
Incompleteness w.r.t. the intended semantics is not unusual in logic: cf., e.g., the essential incompleteness
of Peano arithmetics w.r.t. the standard model of natural numbers.
28 Libor B ehounek, Petr Cintula, and Petr H ajek
weakly implicative logics, which include, i.a., intuitionistic logic or normal modal log-
ics). The remaining properties of the above list, on the other hand, are characteristic of
fuzzy logics: namely, they are satised by the vast majority of logics studied under the
name fuzzy logic in the literature, and rarely satised by a logic not commonly cast as
fuzzy logic. This remark will be made more precise in Section 3 and Chapter II.
2.1 Discarding axioms or rules
Even though the conditions adopted in Section 1.1 for propositional connectives,
with the ensuing axioms of Section 1.2, are reasonable assumptions on generalized ver-
sions of classical propositional connectives, not all of them are necessary for generating
a meaningful system of truth-functional fuzzy logic. In this subsection we shall describe
several systems arising from dropping some of the properties of logics of continuous
t-norms. The algebraic semantics of logics introduced in this subsection is thoroughly
studied in Chapter IV.
2.1.1 Logics of left-continuous t-norms
In Section 1.1 we assumed that the t-norm representing conjunction is continuous.
This assumption ensured the existence of a unique residuum (see Theorem 1.1.8), result-
ing in a good interplay between conjunction and implication. It turns out, however, that
continuity is unnecessarily strong a condition for the existence of a unique residuum, the
minimal condition for residuation being just the left-continuity of the t-norm.
Recall that a unary function is left-continuous (or lower-semicontinuous) if and only
if it commutes with suprema: sup
xa
f(x) = f(a). By a left-continuous t-norm we
mean a t-norm that is left-continuous in either argument (see [123]). Unlike for con-
tinuous t-norms, no characterization similar to Theorem 1.1.7(5) (the Mostert-Shields
Theorem) is known for left-continuous t-norms. Prominent examples of left-continuous
t-norms that are not continuous are the (weak) nilpotent minimum t-norms
NM(n)
:
EXAMPLE 2.1.1 ( [53]). An order-reversing function n: [0, 1] [0, 1] with n(n(x))
x for all x [0, 1] and n(1) = 0 will be called a weak negation. Given a weak negation
n, the weak nilpotent minimum t-norm
WNM(n)
is dened as follows:
x
WNM(n)
y =
_
0 if x n(y)
min(x, y) otherwise.
(3)
Each
WNM(n)
is nilpotent and left-continuous, but not (right-)continuous.
A weak negation n that is is involutive, i.e., n(n(x)) = x for all x [0, 1], will be
called a strong negation. If n is a strong negation, then
WNM(n)
is called the nilpotent
minimum t-norm pertaining to n and can be denoted by
NM(n)
. For the standard
involution n(x) = 1 x, the nilpotent minimum
NM(n)
is called the standard nilpotent
minimum t-norm and denoted by
NM
. Historically, this t-norm was the rst known
example of left-continuous, but not continuous t-norm [64].
Theorem 1.1.8 holds for left-continuous t-norms as well, and ensures the unique ex-
istence of its residuum satisfying the condition (1) of the Theorem. Also Theorem 1.1.9
holds equally well for left-continuous t-norms, except the claim 6 (the denability of
Chapter I: Introduction to Mathematical Fuzzy Logic 29
min in terms of and ): due to its failure, the minimum conjunction has to be in-
cluded among primitive connectives of the logic of left-continuous t-norms.
EXAMPLE 2.1.2 ( [53]). The residuum
WNM(n)
of any weak nilpotent minimum t-
norm
WNM(n)
comes out as
x
WNM(n)
y =
_
1 if x y
max(n(x), y) otherwise.
The residual negation
WNM(n)
pertaining to the weak nilpotent minimum t-norm

WNM(n)
coincides with n.
The denition of standard semantics for the logics of left-continuous t-norms can
thus run along the same lines as for the logics of continuous t-norms described in
Section 1.1. The notions of tautologicity and (nitary) consequence relation with re-
spect to any set of left-continuous t-norms (or the logic of the set of left-continuous t-
norms) can be dened in the same way as for continuous t-norms (see Denitions 1.1.15
and 1.1.19; the notational conventions introduced in the denitions will be extended to
left-continuous t-norms as well).
The logic of all left-continuous t-norms is called monoidal t-norm logic, or MTL,
and was introduced in [53]. The primitive connectives of the logic MTL are &, , ,
and 0. Its derived connectives , , , and 1 are dened in the same way as in BL
(see Section 1.2). Also the axioms and rules of MTL are the same as those of BL, only
the divisibility axiom (BL4) is replaced by the following three axioms describing the
properties of :
(MTL4a)
(MTL4b)
(MTL4c) & ( )
Clearly, the logic BL extends MTL by the converse of (MTL4c). The logic MTL shares
many metamathematical properties with the logic BL: besides those mentioned in the
introduction to this section, also all formulae (T
BL
1)(T
BL
43) are theorems of MTL as
well, and MTL enjoys the same variant of local deduction theorem (cf. Theorem 1.2.10).
MTL-algebras (i.e., the algebras for which the logic MTL is sound) can be char-
acterized as prelinear bounded integral commutative residuated lattices (see Deni-
tion 1.3.1);
20
the class of MTL-algebras is thus a variety. Unlike in BL, where the
standard completeness could be proved only for nite theories, in MTL it can be proved
for all theories [117];
21
i.e., the logic MTL enjoys the innitary-strong standard com-
pleteness.
20
This class of algebras was independently introduced in [63] under the name weak-BL-algebras.
21
This difference is caused by the fact that for the class of all left-continuous t-norms we have [=
T
=
[=
n
T
, like in the case of the minimum t-norm, but unlike all other continuous t-norms (see Theorem 1.1.18).
30 Libor B ehounek, Petr Cintula, and Petr H ajek
2.1.2 Uninorm fuzzy logics
Another way of relaxing the conditions on t-norm fuzzy logics is dropping the re-
quirement that the unit element of the operation representing conjunction coincides with
the largest element of the lattice of truth values. This leads to a generalization of t-norms,
called the uninorms:
DEFINITION 2.1.3 ( [179]). A binary function : [0, 1]
2
[0, 1] is a uninorm if it is
commutative, associative, monotone, and has a unit element e [0, 1]. The uninorm
is conjunctive if 0 1 = 0, and disjunctive if 0 1 = 1.
All uninorms are either conjunctive or disjunctive. T-norms are uninorms with the
unit element e = 1. The unique residuum

that satises the residuation condition (1)

of Theorem 1.1.8 and is explicitly given by Theorem 1.1.9(1) exists if and only if is
left-continuous and conjunctive.
All continuous conjunctive uninorms are already continuous t-norms. Prominent ex-
amples of left-continuous conjunctive uninorms that are not t-norms are the idempotent
left-continuous conjunctive uninorms characterized by the following theorem:
THEOREM 2.1.4 ( [43]). The binary operation : [0, 1]
2
[0, 1] is an idempotent
left-continuous conjunctive uninorm with the unit element e (0, 1] if and only if there
is a weak negation (see Example 2.1.1) n: [0, 1] [0, 1] such that n(e) = e and
x y =
_
min(x, y) if y n(x)
max(x, y) otherwise
(4)
The uninorm given by (4) will be denoted by
IU(n)
, or just
IU
if n(x) = 1 x.
The existence of unique residua makes it possible to construct the semantics of
propositional calculus based on left-continuous conjunctive uninorms along similar lines
as done for (left-)continuous t-norms in Section 1.1; the rst paper studying logics of
uninorms is [133]. The primitive binary connectives of uninorm logics are &, , ,
and , realized respectively by a left-continuous conjunctive uninorm , its residuum

, and the lattice operations of minimum and maximum. Since the unit element e of
need not equal 1, two different primitive truth constants 1 and are distinguished,
the former represented by e and the latter by 1. Similarly there are two different falsity-
related primitive truth constants 0 and ,
22
the latter realized by 0 and the former by any
xed element f [0, 1]. Negation is dened as 0 and equivalence as
( ) ( ).
In logics based on t-norms, only the largest truth value 1 (which coincides with the
unit element e of conjunction) is considered as representing the full truth of a proposition
(i.e., is the designated truth value in the denitions of consequence and tautologicity).
In logics based on left-continuous conjunctive uninorms (where e 1), all truth values
x e are considered as representing the full truth of propositions.
23
Consequently,
22
The symbols t and f are also often used in the literature for 1 and 0, respectively.
23
Earlier than in the context of mathematical fuzzy logic, degrees of full truth were discussed by Casari
in [23]. Note, however, that the understanding of the values e and f in mathematical fuzzy logic differs from
Chapter I: Introduction to Mathematical Fuzzy Logic 31
the formula is dened to be a tautology w.r.t. a set K of uninorms if v

() e
(i.e., if v

() is a designated truth value) for each valuation v in each uninorm algebra

[0, 1]

= [0, 1], ,

, min, max, 0, 1, e, f for K. Similarly, the consequence

relation [=
K
is dened by transmission of the designatedness property v

() e from
the set of premises to the conclusion in each valuation v

for each K. Formally,

we say that v is an -model of iff v

() e for each , and then proceed as in

Denition 1.1.15. The notion of nitary consequence relation w.r.t. a set K of uninorms,
or the logic of K, is dened analogously to Denition 1.1.19.
Thus, while t-norm fuzzy logics recognize degrees of truth between the full truth
e = 1 and the full falsity f = 0, uninorm logics furthermore recognize degrees of full
truth x e and degrees of full falsity x f. The truth constant thus denotes the
largest degree of full truth, while 1 denotes the smallest degree of full truth. Similarly 0
denotes the largest and the smallest full falsity. Observe that since no assumption is
made on the value of f, uninorm logics leave room for dialetheism: if f e, then the
truth values x [e, f] represent propositions that are both fully true and fully false.
Unlike in t-norm logics, in uninorm logics is not generally weaker than &: if
x, y e, then on the contrary holds x y (x y) x y. In uninorm logics it is
therefore preferable to use the names lattice and residual conjunction instead of weak
and strong conjunction. Another consequence of this fact is that the -adjunction rule
(which is derivable in BL, see the comments after (T
BL
16) in Section 1.2) has to be
added as a primitive rule of UL.
The logic UL of all left-continuous conjunctive uninorms can be axiomatized by the
following axioms:
(UL1)
(UL2) ( ) (( ) ( ))
(UL3) ( ( )) ( ( ))
(UL4) ( & ) ( ( ))
(UL5a,b) ,
(UL6) ( ) ( ) ( )
(UL7a,b) ( ), ( )
(UL8) ( ) ( ) ( )
(UL9) (1 )
(UL10a,b) ,
(UL11) (( ) 1) (( ) 1)
with the denitions
( ) ( )
0
that used in engineering fuzzy methods and the theory of aggregation operators. In the latter areas, only the
values 0 and 1 are understood, respectively, as full membership and full non-membership, while the value e
just separates values with different aggregating behavior.
32 Libor B ehounek, Petr Cintula, and Petr H ajek
and the derivation rules of modus ponens and -adjunction:
(MP) From and infer
(-Adj) From and infer .
Further logics of various sets of left-continuous conjunctive uninorms will be intro-
duced in Section 2.4. The logic MTL of left-continuous t-norms extends the logic UL
of left-continuous conjunctive uninorms by the axiom stating that 0 is the smallest and
1 the largest truth value:
(W) (0 ) ( 1)
Unlike MTL, the logic UL differs more radically from BL: e.g., the formulae (T
BL
6),
(T
BL
2), and (T
BL
11) are not theorems of UL, and the local deduction theorem of BL
(cf. Theorem 1.2.10) is not valid for UL. However we can prove a modied variant of
the deduction theorem:
T,
L
iff there is natural n such that T ( 1)
n
.
The logic UL enjoys the innitary-strong standard completeness theorem.
The algebras for which the logic UL is sound (i.e., UL-algebras) can be character-
ized as pointed semilinear
24
commutative bounded residuated lattices:
DEFINITION 2.1.5 ( [67, 133]). A bounded pointed commutative residuated lattice is
an algebra A = A, &, , , , 0, 1, , such that:
1. A, , , , is a bounded lattice
2. A, &, 1 is a commutative monoid
3. is the residuum of &, i.e., for each x, y, z A holds: x&y z iff x y z.
A bounded pointed commutative residuated lattice A is a UL-algebra if it satises the
identity ((x y) 1) ((y x) 1) = 1 for all x, y A.
2.1.3 Fuzzy logics with non-commutative conjunction
Now we shall survey fuzzy logics that omit the assumption that strong conjunction
is commutative, i.e., that x y = y x for all truth values x and y. The residuation
condition (1) of Theorem 1.1.8 then has two non-equivalent variants:
z x y iff z x y
x z y iff z x y.
24
A class of algebras is called semilinear (a term representable is sometimes used instead) if all its subdi-
rectly irreducible members (cf. Denition 1.3.10) are linearly ordered. Note that in BL- and MTL-algebras,
the axiom (x y) (y x) = 1 called prelinearity was sufcient to enforce semilinearity (which can
equivalently be formulated as the subdirect representation theorem, cf. Theorem 1.3.11). In UL and psMTL
r
(introduced in Section 2.1.3 below) this is no longer the case, and more complex axioms are needed to ensure
semilinearity (cf. Denition 2.1.5 and footnote 26 on p. 34). See Section 3.3 for more details and Chapter II
for a thorough exposition of this phenomenon.
Chapter I: Introduction to Mathematical Fuzzy Logic 33
Each of them corresponds to a distinct residuum, respectively denoted by / and and
called the left and right residuum. The residuation condition then reads:
x y z iff x z / y iff y x z. (5)
The general non-commutative case of residuated lattices (cf. Denitions 1.3.1 and 2.1.5)
is thus dened as follows:
DEFINITION 2.1.6 (Residuated lattice [67]). A residuated lattice is a lattice equipped
with a monotone monoidal operation (with a unit e) and a pair of operations /,
satisfying the residuation condition (5).
The various subclasses of general residuated lattices (e.g., bounded, integral, linear,
etc.) are dened analogously to the commutative case (see Section 1.3 and Chapter IV).
Clearly is commutative iff / and coincide, in which case they can be denoted by
as usual. In [0, 1], or any residuated lattice, the two residua can be explicitly dened
as follows (cf. Theorem 1.1.9(1)):
z / y = maxx [ x y z
x z = maxy [ x y z
The residua / and are the pointwise largest (i.e., logically weakest) functions validating
the two versions of internalized modus ponens:
(z / y) y z (6)
x (x z) z (7)
Integral bounded residuated lattices on [0, 1] are given by left-continuous pseudo-
t-norms ( [63]), i.e., a binary operation : [0, 1]
2
[0, 1] which is associative, non-
decreasing in both arguments, has 1 as both left and right unit element, and is left-
continuous in both arguments.
EXAMPLE 2.1.7. Let 0 < a < b < 1 and let
x y =
_
0 if x a and y b
min(x, y) otherwise
This is a non-commutative pseudo-t-norm which is left-continuous in both arguments.
The existence of the left and right residua of left-continuous pseudo-t-norms makes
it possible to dene the propositional logic of pseudo-t-norms along the lines of Sec-
tion 1.1, with the main difference of having two implications: represented by /, and
represented by .
25
The ensuing logic of left-continuous pseudo-t-norms has been de-
25
In the literature on substructural logics, the implications are usually denoted by the same signs /, \ as
the residua. In the literature on non-commutative fuzzy logic, /, are often denoted by (with swapped
arguments) and \, by . Here we use the signs , suggested by L.N. Stout, since besides indicating the
side of conjoining the antecedent in (6)(7) they also mark the direction of the implication from the antecedent
to the succedent.
34 Libor B ehounek, Petr Cintula, and Petr H ajek
noted by psMTL
r
and called representable pseudo-MTL.
26
It can be axiomatized as
follows:
(psMTL
r
1a,b) ( ) (( ) ( )), (( ) ( )) ( )
(psMTL
r
2a,b) & , &
(psMTL
r
3)
(psMTL
r
4a)
(psMTL
r
4b)
(psMTL
r
4c,d) & ( ) , ( ) &
(psMTL
r
5a,b) ( ( ) ( & )), (( ) ) ( & )
(psMTL
r
6a) (( ) ) ((( ) ) )
(psMTL
r
6b) ( ( ( ))) ( ( ))
(psMTL
r
7) 0
(psMTL
r
8a,b) ( ) ( & ( ) ), ( ) ( ( ) & )
with the denitions
(( ) ) ( ( ))
( ) ( )
( ) ( )
and the derivation rules
(MP) from and infer
(Imp a,b) from infer , from infer
The logic psMTL
r
also enjoys a variant of the local deduction theorem, but we do not
give it here due to its complexity (see Chapter II).
The algebras for which the logic psMTL
r
is sound are called representable psMTL-
algebras, or psMTL
r
-algebras; they can be characterized as semilinear bounded integral
residuated lattices (see Denition 2.1.6). Their lattice reduct is [0, 1] if and only if is
a left-continuous pseudo-t-norm (then they are called standard psMTL
r
-algebras). The
innitary-strong standard completeness theorem holds for psMTL
r
[118].
A similar non-commutative generalization can be done with other logics. In this
way we obtain, e.g., the logic psBL
r
that extends psMTL
r
by the converse of the axioms
(psMTL
r
4c,d),
27
or the logic psUL of semilinear bounded residuated lattices. In both
26
The superscript
r
and the adjective representable come from the historical fact that an ill-motivated
weaker logic lacking the axioms (psMTL
r
8a,b) was originally named psMTL. The latter logic was intro-
duced in [88], following the terminology for previously introduced logic psBL [87], whose name was in turn
based on previously dened corresponding algebraic structures [46], originally motivated by non-commutative
version of MV-algebras [69]. The logic psMTL does not enjoy standard completeness nor linear subdirect
representation (i.e., semilinearity, cf. footnote 24 on p. 32). The representability means recovering the latter
property in psMTL
r
by adding the axioms (psMTL
r
8a,b) to the original psMTL. The fact that these axioms
ensure this property was explicitly proved in [126]; it is also a corollary of theharacterization of representable
residuated lattices given in [119]. A systematic name FL

w
for psMTL
r
will be introduced in Section 3.
27
Consequently, psBL
r
-algebras are psMTL
r
-algebras satisfying the identity x (x\ y) = (y / x) x =
x y.
Chapter I: Introduction to Mathematical Fuzzy Logic 35
Figure 2. Relative position of fuzzy logics introduced in Section 2.1.
cases, however, the standard completeness is lost: in psBL
r
it is simply because any con-
tinuous pseudo-t-normis already a continuous t-norm[63] and it can be shown [176] that
psUL is not the logic of residuated pseudo-uninorms either (for details see Chapter IV).
2.2 Adding new connectives
We shall now survey another direction in varying the apparatus of propositional
fuzzy logics, namely expanding its expressive power by adding new primitive connec-
tives. It can be observed that truth functions denable by means of the basic propo-
sitional connectives (&, , , , , , and 0) of the logic MTL in a given stan-
dard MTL-algebra form just a limited subset of all truth functions [0, 1]
n
[0, 1]
(or A
n
A in a general MTL-algebra A). The same is true about other fuzzy logics
treated in previous sections (for characterizations of truth functions representable by for-
mulae in various fuzzy logics see Section 4.2). Various important propositional concepts
are actually not expressible by these basic connectives, and if they are to be available in
a system of fuzzy logic, new primitive connectives with suitable truth functions have to
be added. The logics introduced in this subsection are studied in detail in Chapter VIII.
2.2.1 The Delta connective
One of the concepts that is expressible neither in MTL nor in other logics of left-
continuous t-norms (except some of their nitely valued variants, see Section 2.4.3)
is the notion of full truth of a proposition, formally represented by a truth function
: A A such that (x) = 1
A
iff x = 1
A
for all x A (clearly such a function is
not denable in the standard MV-algebra). This deciency can be remedied by adding
a new primitive connective ,
28
with the standard and linear semantics given as:
(x) =
_
1
A
if x = 1
A
0
A
otherwise.
(8)
28
The connective rst appeared in [142] in the context of intuitionistic and G odel logics, and was extensively
studied by M. Baaz; therefore it is often called Baaz Delta. Its axiomatization was generalized for other fuzzy
logics in [83].
36 Libor B ehounek, Petr Cintula, and Petr H ajek
Tarski conditions (see Denition 1.1.13) are then extended by the clause
e

() = (e

()).
Let us x (for the rest of this subsection) an axiomatic extension L of MTL. A
fuzzy logic L expanded by this connective is called L with and denoted by L
Z
. It
can be axiomatized by adding the following axioms and rule to the axiomatic system for
the logic L:
(1)
(2) ( )
(3)
(4)
(5) ( ) ( )
and the rule of -necessitation,
(-Nec) from infer .
Linear L
Z
-algebras are linear L-algebras expanded by the operation dened
by (8) above.
29
The usual linear and general completeness theorems hold for L
Z
, as
well as the linear subdirect representation theorem. In consequence of these properties,
L
Z
extends L conservatively. The form of the deduction theorem for L
Z
is different
from that for L (see Theorem 1.2.10):
THEOREM 2.2.1 (-deduction theorem). Let L be an axiomatic extension of MTL.
Then for any set of formulae T and formulae , of L
Z
, the following equivalence
holds:
T,
L

iff T
L

The standard L
Z
-algebras are dened (cf. Convention 2.0.2) as those L
Z
-algebras
whose lattice reduct is [0, 1], with the exception of
Z
and
Z
, where only the single
t-algebras [0, 1]

resp. [0, 1]

expanded by are considered standard. The standard

L
Z
-algebras are thus the standard L-algebras expanded by the operation dened by
(8). The logic L
Z
is nitely (resp., innitary-) strong standard complete if and only if L
is nitely (resp., innitary-) strong standard complete [35].
Besides the notion of full truth, also the ordering of truth values is internalized in
L
Z
, as (x y) = 1 iff x y. The connective makes it possible to interpret
classical logic in L
Z
by prexing each propositional variable by . Moreover, the logic
G
Z
can be interpreted in any L
Z
(so already in MTL
Z
), as G odel conjunction is the
lattice conjunction of MTL
Z
and G odel implication is denable in MTL
Z
by setting:

G

df
( ) .
29
The semantics of . in non-linear L

algebras, however, need not be that given by (8): rather, general

L

-algebras are L-algebras B expanded by the operation dened as (8) in every component Aof the linear
subdirect representation of B.
Chapter I: Introduction to Mathematical Fuzzy Logic 37
2.2.2 Fuzzy logics with additional involutive negation
Another frequent expansion of basic t-norm logics is by adding an extra unary con-
nective which is order-reversing (i.e., if x y then x y for all truth values x, y)
and involutive (i.e., x = x for all truth values x). The connective is usually called
involutive negation, or simply involution. The expansion is done in a similar way as in
the case of .
Let us again x (for the rest of this subsection) an axiomatic extension L of MTL
or MTL
Z
. The logic L expanded by involutive negation is called L with involution
and denoted by L

.
30
In logics with involution we depart from Convention 2.0.2 and
call standard L

-algebras only the expansions of standard L-algebras by the standard

involutive negation x = 1 x. Thus there are just single standard

- and G

-
algebras, and the standard MV

-algebra is (modulo the signature) identical with the

standard MV-algebra [0, 1]

, as in the former both negations ( and ) coincide.

The logic L

is axiomatized by adding the following axiom and rule to the logic L:

()
(Rev-) From ( ) infer ( )
In logics with , the rule (Rev-) can equivalently be replaced by the axiom
(Rev-) ( ) ( )
If ( ) is a theorem of L (i.e., the residual negation of L is strict, as, e.g., in
G, , or SBL; cf. Example 1.1.11), then (Rev-) can be replaced by the following two
axioms and rule [30]:
(Rev-S1) ( ) ( )
(Rev-S2)
(Rev-S3) From infer
In such logics, the connective is denable by setting . Thus if L has
strict residual negation, then L

includes L
Z
; in particular, G
Z
G

and
Z

. Furthermore, in this case the logic L

enjoys the same deduction theorem as L

Z
(cf. Theorem 2.2.1)
A (form of) standard completeness of L

can be proved for a broad class of t-norm

logics satisfying certain criteria (for details see [62] or Chapter VIII), among them the
logics MTL

, BL

, SBL

, and G

. As in the case of L
Z
, the logic L

is nitely
(resp., innitary-) strong standard complete if and only if L is nitely (resp., innitary-)
strong standard complete.
30
As in the case of Baaz Delta, the additional involutive negation rst appeared in [142] in the context of
intuitionistic and G odel logics. Later is was studied in [56] in logics L extending SBL (in which the connective
., needed for the rule (Rev-.) below, becomes denable in L

). The following generalization [62] required

that the logic L extend MTL

; the logic called MTL

in [62] thus in fact denoted the logic MTL

. We
are changing the terminology here to a more systematic one, since the .can be dispensed with by using the
rule (Rev-); cf. Chapter II. Note also that this generalization also allows us to add an additional involution
to a logic in which already the residual negation is involutive, e.g., IMTL or .
38 Libor B ehounek, Petr Cintula, and Petr H ajek
The logics

and

fail to be standard complete, i.e., complete w.r.t. the (single)

standard MV-algebra [0, 1]

resp. -algebra [0, 1]

expanded by the single standard

involution 1 x.
31
Nevertheless, they do enjoy what is sometimes called the semi-
standard completeness [56], i.e., completeness w.r.t. the (single) standard MV or -
algebra expanded by arbitrary involutive negation on [0, 1]. Interestingly, both these
logics are also complete w.r.t. the class of all real-valued MV or -algebras expanded
by the standard involutive negation 1 x (this is a simple corollary of the fact that
all involutive negations on [0, 1] are isomorphic). Thus in

and

we can, roughly
speaking, x either the standard t-norm or the standard negation, but not both at once.
Clearly the logic of the standard MV

-algebra is the axiomatic extension of

by the
axiom . The case of

is more involved and interesting; we will comment

on it further in Section 2.2.3. Axiomatic extensions of

were studied in detail in [40]

and [105].
An important feature of the logics L

is the denability of the so-called strong

disjunction, i.e., a connective dual to strong conjunction:
( &)
The standard semantics of on [0, 1] is that of a t-conorm:
DEFINITION 2.2.2 ( [123]). A binary function S: [0, 1]
2
[0, 1] is called a triangular
conorm (a t-conorm for short) if it is commutative, associative, monotone, and 0 is its
neutral element.
T-norms and t-conorms are mutually dual via the standard involution 1 x: every
t-norm T has the dual t-conorm S
T
(x, y) = 1 T(1 x, 1 y), and every t-conorm S
has the dual t-norm T
S
(x, y) = 1 S(1 x, 1 y), with S
T
S
= S and T
S
T
= T.
EXAMPLE 2.2.3. The t-conorms dual to the three prominent continuous t-norms come
out as follows:
The t-conorm dual to

is the bounded sum x

y = max(x + y, 1).
The t-conorm dual to

is the probabilistic sum x

y = x + y xy.
The t-conorm dual to
G
is the maximum. Consequently, due to the standard
completeness of G

, strong disjunction coincides with lattice disjunction

in G

.
The presence of two disjunctions alongside two conjunctions in fuzzy logics is nat-
ural in view of their relation to substructural logics (see Section 3.1). Various concepts
can be expressed by means of strong disjunction, including the so-called S-implication,
I
S
(x, y) = x y, often encountered in applied fuzzy logic, or the Q-implication,
I
Q
(x, y) = x (x &y), related to the implication used in quantum logic. De Morgan
laws with hold not only for & and (by the denition of ), but also for and in
MTL

.
31
This failure is obvious in the case of

, where standard completeness would entail that coincides with

residual negation, which clearly is not the case; the failure in

is shown in [56].
Chapter I: Introduction to Mathematical Fuzzy Logic 39
2.2.3 Fuzzy logics with intermediate truth constants
Another way of enriching the expressive power of a fuzzy logic is expansion by
truth constants (i.e., nullary connectives) other than 0 and 1. If a new truth constant is
intended to have a xed standard semantics r [0, 1], it is denoted by r; in this case,
Convention 2.0.2 is strengthened so as to admit as standard only those algebras that
indeed interpret r as r.
EXAMPLE 2.2.4. ukasiewicz logic can easily be expanded by a truth constant with
the standard semantics of one half: since

0.5 = 0.5, the axiom 0.5 0.5 forces

the truth constant 0.5 to be realized by 0.5 in the standard MV-algebra [0, 1]

. In the
general semantics, 0.5 is realized by the xed point of . Only those MV-algebras
that do have this xed point can be expanded to models of with 0.5. The expansion
therefore excludes, i.a., the two-element (Boolean) MV-algebra 0, 1, and so ensures
the many-valuedness of the logic.
An important logic is the expansion of ukasiewicz logic by truth constants for all
rational numbers in [0, 1]. The idea of using truth constants denoting truth degrees in
the language of fuzzy logic goes back to Pavelka [154], who used truth constants for all
reals from [0, 1]. Later it turned out, though, that for the main results to work well it is
sufcient to introduce truth constants just for the rational numbers of [0, 1], which does
not force the language to be uncountable.
DEFINITION 2.2.5 ( [83]). The rational Pavelka logic, denoted by RPL, is the expan-
sion of ukasiewicz logic by the truth constants r for each rational r [0, 1] and the
bookkeeping axioms for each rational r, s [0, 1]:
r & s r s
(r s) r s.
The standard RPL-algebra is the expansion of the standard MV-algebra [0, 1]

with each
r interpreted as r.
The logic RPL is a conservative extension of ukasiewicz logic and possesses the
same deduction theorem. It has the nite strong standard completeness. Moreover,
RPL enjoys the following property called Pavelka-style completeness for arbitrary sets
of formulae:
THEOREM2.2.6 (Pavelka-style completeness of RPL, [83]). Let T be a set of formulae
(possibly innite) and a formula of RPL. Let the provability degree of in the theory
T be the real number
[[
T
= supr [ T
RPL
r
and the truth degree ||
T
of in T be the inmum of truth degrees of in all standard
models of T. Then
[[
T
= ||
T
.
The continuity of is essential for Pavelka-style completeness; in expansions by ratio-
nal truth constants of t-norm logics other than ukasiewicz, additional innitary rules
40 Libor B ehounek, Petr Cintula, and Petr H ajek
have to be added in order for these logics to enjoy this style of completeness (see
[33, 56, 57] for details). It is, nevertheless, possible to consider just the usual style
of completeness for these logics; see e.g. [51, 59, 166] or Chapter VIII for a detailed
exposition of known results.
Finally, Pavelka-style logics are closely related to the so-called fuzzy logics with
evaluated syntax, which also incorporate the truth values of formulae directly into the
syntax of the logic. An evaluated formula is a pair r, , which expresses the fact that
the truth value of is at least r. The rules of fuzzy logics with evaluated syntax operate
on evaluated formulae: e.g., the evaluated rule of modus ponens has the form:
From r, and s, infer r

s, .
The logic with evaluated syntax based on ukasiewicz logic, denoted by
Ev
, can be in-
terpreted in RPL, by translating an evaluated formula r, as the RPL-formula r .
The variant
Ev
based on the logic introduced below has also been considered.
See [150] for a comprehensive treatment of these logics and further references.
2.2.4 Fuzzy logics with multiple sets of t-norm connectives
As stated above, one of the main reasons for expanding t-norm logics by additional
connectives is the fact that only a limited set of truth functions is in general available in
the logic of any particular left-continuous t-norm (cf. Section 4.2). Thus, for example,
ukasiewicz logic only possesses connectives interpreted by additive arithmetical oper-
ations, while in product logic we are only in possession of multiplicative connectives.
A solution to the need of possessing a fuller arithmetic power over truth degrees is to
combine connectives pertaining to particular left-continuous t-norms in one logic.
Several logics of this kind, differing in expressive power, have been described in the
literature. Most of them add connectives pertaining to the product t-normto ukasiewicz
logic. Adding just the product conjunction to or
Z
, with various strength of axioms,
leads to logics P and P
t
(possibly with , see [112])
32
and the propositional logic
of TakeutiTitani [170]. These logics have remarkable logical properties: the logic P
is not even weakly standard complete (see Convention 2.0.3) w.r.t. the standard MV-
algebra expanded by the connective &

interpreted as

. The logic P
t
, which ex-
tends P by the deduction rule ( &

) does have the nite strong standard

completeness, but the additional rule cannot be replaced by axioms. Consequently, P
t
lacks the deduction theorem (cf. Theorem 1.2.10) and the class of P
t
-algebras is not a
variety of algebras.
Despite these interesting properties we will not describe these logics in detail here
and will rather concentrate on the (much better behaving) logic , which arises from
adding both product conjunction and implication to ukasiewicz logic, and its expansion

1
2
by the truth constant
1
2
(cf. Example 2.2.4).
33
The logic
1
2
is one of the expres-
sively strongest fuzzy logics studied in the literature; it includes many other t-norm
32
The algebras for these logics are studied under the name PMV-algebras, see, e.g., [44, 137, 141].
33
A more systematic denotation for the connective represented by the value
1
2
in the standard semantics
would be
1
2
, with the expansion of by this connective denoted by
1
2
. For typographical reasons,
however, we shall follow the tradition of denoting the constant simply by
1
2
and the logic by
1
2
.
Chapter I: Introduction to Mathematical Fuzzy Logic 41
fuzzy logics and contains a broad class of denable connectives, while still possessing
good metamathematical properties. In this sense, the logic
1
2
can be viewed as an
over-arching system for a large class of t-norm fuzzy logics.
DEFINITION 2.2.7 ( [31, 52, 57]). The primitive connectives of the logic are the
truth constant 0, ukasiewicz implication

, product implication

, and product
conjunction &

. Furthermore we dene the following derived connectives of :

0
1

&

&

&

)
(

G
(

)
and

) (

) for , , G. The axioms and rules of the

logic are the following:
() the axioms and rules of (with -subscripted connectives)
() the axioms and rules of (with -subscripted connectives)
(-Nec) from infer
(

a,b) (

), (

)
(Distr) &

( )

( &

) ( &

)
The standard -algebra [0, 1]

is the standard MV-algebra [0, 1]

expanded by
the product t-norm and its residuum.
In the standard -algebra [0, 1]

, the connectives subscripted by , , or G have

the standard semantics of the corresponding t-norm logics. The connectives and ,
denable already in , have the standard semantics of bounded addition and subtraction.
The tautologicity and the (nitary) consequence relation are dened along the usual
lines (cf. Section 1.1), with 1 as the only designated truth value. The -algebras,
i.e., general algebras for which the logic is sound (see Convention 2.0.1), can be
characterized as MV-algebras expanded by the product adjoint pair (so that the appro-
priate reduct is a -algebra) and satisfying the distributivity axiom (Distr). The logic
enjoys only the nite strong standard completeness theorem [57].
Due to the presence of the basic connectives of all the three logics , , and G as
well as and an involutive negation

, the logic contains the logics

Z
,

, and
G

(thus also
Z
, G
Z
, and all weaker logics). It can moreover be shown to extend ,
G

, and

(with or without ) conservatively. Several alternative axiomatizations of

42 Libor B ehounek, Petr Cintula, and Petr H ajek
are known; for instance, it can be viewed as an axiomatic extension of

by the
axiom (BL1) for ukasiewicz implication dened as

( & ( ))
[30], or by the axiom (BL3) for ukasiewicz conjunction dened as &

&
( ) [173], where the connectives of

are taken for the product connectives

and

of .
The logic can further be expanded by the truth constant
1
2
(see footnote 33 on
p. 40) satisfying the axiom
1
2

1
2
, with the standard semantics 0.5 (cf. Exam-
ple 2.2.4). This not only excludes the two-valued -algebra, in which the connectives
coincide with the classical bivalent ones, but also increases signicantly the expressive
power of the logic, as all rational truth constants r are denable and their bookkeeping
axioms are provable in
1
2
(cf. Denition 2.2.5);
1
2
thus contains the logic RPL.
The metamathematical properties of mentioned in the previous paragraphs hold for

1
2
as well.
The logic
1
2
thus contains connectives corresponding in standard semantics to
all basic arithmetical operations (namely, , , , , &

, and

) and comparison
relations (namely, (

) and (

)) as well as all rational numbers in

[0, 1]. Consequently, a broad class of truth functions on [0, 1] is representable in the
standard semantics of
1
2
(see Section 4.2). The
1
2
-representable functions include
not only the truth functions of basic connectives of the three salient continuous t-norms
, , and G, but also of all their nite ordinal sums:
THEOREM 2.2.8 ( [32]). Let be a nite ordinal sum of

, and
G
on rational
subintervals of [0, 1] and

be its residuum. Then there are connectives &

and

denable in
1
2
with the standard semantics of and

. Moreover, if is provable
in the logic L

of , then the formula

obtained by replacing the connectives & and

of L

by the dened connectives &

and

of
1
2
is provable in
1
2
.
The logic
1
2
thus contains the logics of all nite ordinal sums of the basic continuous
t-norms, and also many particular left-continuous t-norms (e.g., the nilpotent minimum
of Example 2.1.1) and uninorms (see [130] for details).
The logics and
1
2
(as well as the logics P and P
t
mentioned in the be-
ginning of this subsection, either with or without ) can be extended by rational truth
constants (which in
1
2
are already denable) and appropriate innitary rules to ob-
tain their Pavelka-style extensions (containing RPL and contained in the Pavelka-style
extension R of ) that enjoy Pavelka-style completeness (cf. Theorem 2.2.6).
2.3 Discarding connectives
Various fragments of fuzzy logics in restricted languages have been studied. The
most important ones are certain natural expansions of the logic BCK (for which see,
e.g., [116, 152] or [67, 2.3.2]). An extensive study of these fragments is given in [38].
In this subsection we restrict ourselves to the axiomatic extensions of MTL introduced
in the previous sections and to languages containing implication (as implication-less
fragments of our fuzzy logics are essentially classical, see [1]) and a subset of the con-
nectives &, , , 0.
Let / be a sublanguage of , &, , , 0 and L an axiomatic extension of MTL.
The /-fragment of L, denoted by L

/, is the logic in the language / such that

L

1

Chapter I: Introduction to Mathematical Fuzzy Logic 43
Figure 3. Relative positions of prominent logics introduced in Section 2.2. Thin lines
indicate extensions in the same language (or just with addition of denable connectives),
while thick lines indicate expansions by connectives not denable in the weaker logic.
iff
L
, for every Fm
1
. Algebraically speaking, the /-reduct of an
algebra A is the algebra A

/ with the same domain and operations from /; i.e., we

just forget the operations not in /. An /-subreduct of A is just a subalgebra of the
/-reduct of A. For a logic L in language / and a connective c /, the logic L

(/c)
is sometimes called the c-free fragment of L, and an analogous convention is used for
(sub)reducts of algebras.
2.3.1 Falsity-free fuzzy logics
Here we shall briey discuss 0-free fragments of fuzzy logics. L-algebras for 0-free
t-norm fuzzy logics L fall within the class of algebras known as hoops. Hoops were
introduced in [21] and studied, e.g., in [17, 61]. The seminal paper on hoops in the
context of fuzzy logic is [55], followed by [2]; see these two papers (where all results
mentioned in this section could be found) or Chapter V for detailed references.
DEFINITION 2.3.1. A structure H = (H, &, , 1) is a hoop if & is a commutative
operation on H with the unit 1 and is a binary operation satisfying
44 Libor B ehounek, Petr Cintula, and Petr H ajek
x x = 1
x & (x y) = y & (y x)
x (y z) = (x & y) z
for all x, y, z H. Dene x y iff x y = 1. A hoop is
Prelinear
34
if (x y) z ((y x) z) z) for all x, y, z H
Wajsberg if (x y) y = (y x) x for all x, y H
Cancellative if x & y x
t
& y implies x x
t
for all x, x
t
, y H.
One can prove that, in each hoop, is an ordering, & is associative and non-decreasing,
is the residuum of &, and 1 is the largest element. Hoops can be characterized as
divisible integral commutative residuated lattices (see Denition 1.3.1).
Prelinear hoops are precisely the 0-free (or falsity-free) subreducts (i.e., subalgebras
of 0-free reducts) of BL-algebras, and Wajsberg hoops are falsity-free subreducts of
Wajsberg algebras.
35
Note that each cancellative hoop is Wajsberg and each Wajsberg
hoop is prelinear. On the other hand, each unbounded Wajsberg hoop is cancellative
and bounded Wajsberg hoops are 0-free reducts of Wajsberg algebras. Linearly ordered
Wajsberg hoops play an important r ole in the description of the structure of BL-chains,
which can be decomposed into an ordinal sum of linearly ordered Wajsberg hoops in
a similar (though slightly different) manner as can continuous t-norms be decomposed
into an ordinal sum of the three basic continuous t-norms by Theorem 1.1.7(5); see
Chapter V for details.
The logic BLH of prelinear hoops, or basic hoop logic, has the axioms (BL1)
(BL6) of the logic BL and the rule of modus ponens; thus only the last axiom (BL7)
that speaks of 0 is deleted (see Section 1.2). Similarly the logic MTLH is the logic with
the axioms of MTL except the last one (speaking of 0) and modus ponens. The logic
GHextends BLHby the the axiom (G) of G odel logic. The logic Hof Wajsberg hoops
extends BLH by the axiom
(WH) (( ) ) (( ) ).
The logic H extends BLH by the following three axioms:
(H1) ( & ) (( ( )) )
(H2) (( ) ) & ( & & ) & ( & & ) ( )
(H3) (( ) (( ) )) (( ) ) (( ) ).
34
The term basic hoop has been used in the literature instead [55]. However, that terminology is rather
confusing, as the condition is equivalent to the prelinearity condition from the denition of BL- and MTL-
algebras, while the divisibility condition (which needs to be added to MTL-algebras to obtain BL-algebras
the algebras for basic logic) is in fact satised in all hoops.
35
Recall that Wajsberg algebras are termwise equivalent to MV-algebras, see footnote 16 on p. 23.
Chapter I: Introduction to Mathematical Fuzzy Logic 45
The logic of cancellative hoops, or cancellative hoop logic CHL, extends the logic BHL
by the following axiom:
(CH) ( & ) .
The logic CHL is not a 0-free fragment of any logic that proves 0 , since expanding
CHL by 0 and the axiom 0 makes the resulting logic, BL + (CH), inconsistent (as
easily shown by instantiating (CH) by 0 for and ). Product logic is not an extension of
CHL, as (CH) is not valid in (one cannot cancel by 0); however, there is the following
connection between CHL and :
THEOREM 2.3.2. Let all propositional variables occurring in be among p
1
, . . . , p
n
.
Then CHL proves iff proves p
1
. . . p
n
.
Moreover it can be shown that:
The logic BLH is the 0-free fragment of both BL and SBL.
The logic MTLH is the 0-free fragment of MTL, IMTL, and SMTL (introduced
below in Section 2.4.1).
The logic H (or GH or H, resp.) is the 0-free fragment of (or G or , resp.).
The rst two claims are perhaps surprising, as we have different logics which share the
same 0-free fragment; this shows that these logics really differ only in the properties of
negation.
The algebras for the logics MTLH, GH, and H are respectively called prelinear
semihoops, G odel hoops, and product hoops, and they are 0-free subreducts of the cor-
responding MTL-, G-, and -algebras. Standard (semi-)hoops are just all real-valued
(semi-)hoops.
36
The denition of standard hoops for stronger hoop logics is more com-
plex:
The standard cancellative hoop is the positive part of the standard product algebra
(i.e., the half-open interval (0, 1] with the standard product operations).
The standard G odel, Wajsberg, and product hoop is just the 0-free reduct of
[0, 1]
G
, [0, 1]

, and [0, 1]

, respectively.
The logics MTLH and GH enjoy innitary-strong standard completeness, whereas the
logic BLH, H, H, and CHL only enjoy nite strong standard completeness.
What about deleting both the truth constant for falsity and the commutativity of
conjunction from a t-norm fuzzy logic? Algebras of this kind have been studied under
the name pseudohoops. A logic generalizing both the logic psMTL and the hoop logic
MTLH was introduced under the name ea logic. See [70] for pseudohoops and [91]
for eas.
36
Recall that in Convention 2.0.1 we allowed the real-unit interval to be half-open.
46 Libor B ehounek, Petr Cintula, and Petr H ajek
2.3.2 Discarding other connectives
In this subsection we deal with the remaining fragments of prominent axiomatic ex-
tensions of MTL. All results of this subsection appear in [38], where further references
to the original sources of some particular results can be found. Our rst goal is to present
explicit axiomatic systems for these fragments. In order to be able to formulate many
such results at once we introduce the following notion (particularized to our setting).
DEFINITION 2.3.3. Let L be an axiomatic extension of MTL and let /be an axiomatic
system for L. We say that / is strongly separable if for each propositional language
/ &, , , 0, the /-fragment of L is axiomatized by the axioms and
rules from / that contain the connectives from / only.
The axiomatic system MTL
s
, with modus ponens as the only deduction rule and
axioms listed below, is a strongly separable axiomatic system of MTL:
37
(MTL
s
1) ( ) (( ) ( ))
(MTL
s
2) ( )
(MTL
s
3) ( ( )) ( ( ))
(MTL
s
6) (( ) ) ((( ) ) )
(MTL
s
5a) ( ( )) (& )
(MTL
s
5b) (& ) ( ( ))
(MTL
s
7) 0
(MTL
s
4a) ( ) (( ) ( ))
(MTL
s
4b)
(MTL
s
4c)
(MTL
s
4a) ( ) (( ) ( ))
(MTL
s
4b)
(MTL
s
4c)
To obtain strongly separable axiomatic systems for prominent axiomatic extensions
of MTL, we need implicational forms of their characteristic axioms:
(Div) (( ) ( )) (( ) ( )) divisibility
(Waj) (( ) ) (( ) ) Wajsberg axiom
(Contr) ( ( )) ( ) contraction
(SBL) ( ) strictness
37
The numbering of axioms follows the numbers of corresponding axioms in the axiomatic systems of BL
and MTL introduced in Section 1.2 and 2.1; the axioms are grouped according to the connectives involved.
Note that the rst three axioms with modus ponens constitute the axiomatic system of the well-known purely
implicational logic BCK (see, e.g., [67, 2.3.2]). Its extension by the axiom (MTL
s
6) is the implicational
fragment of MTL, called fuzzy BCK, or FBCK, in [38].
Chapter I: Introduction to Mathematical Fuzzy Logic 47
Logic Axioms added to MTL
s
BL (Div)
SBL (Div), (SBL)
(Div), (Waj)
G (Div), (Contr)
(Div), (SBL), (H1)(H3)
Table 1. Strongly separable axiomatic systems for some prominent extensions of MTL
Fragment: , , , , , & , &, , &,
MTL
BL, SBL, ,
G &, & ,
,
Table 2. Denability of connectives in fragments without 0
Also recall the axioms (H1)(H3) of product hoop logic fromSection 2.3.1. Strongly
separable axiomatic systems for prominent extensions of MTL are listed in Table 1.
We dene standard L

/-algebras as just the /-reducts of standard L-algebras. Stan-

dard completeness theorems are then obviously inherited by the fragments: if L enjoys
innitary- (resp. nite) strong standard completeness, then also L

/ enjoys innitary-
(resp. nite) strong standard completeness. Thus we can, for instance, say that MTL

is the logic of the residua of left-continuous t-norms.

Next let us present some results on the denability of connectives in prominent
fragments. We say that a connective c is denable in L

/ if there is an /-formula
such that c(p
1
, . . . , p
n
) (p
1
, . . . , p
m
) is a theorem of L. Note that the dening
formula can have a different number of variables, as witnessed, e.g., by the denability
of 1 in the -fragment of MTL: indeed 1 (p p) is a theorem of MTL. Let us
list some known positive results:
is denable in MTL

, &, by ( & ( )) ( & ( ))

is denable in MTL

, by (( ) ) (( ) )
is denable in BL

, & by & ( )
is denable in

by ( )
& is denable in

, 0 by ( )
& is denable in G

, by .
Of course a connective denable in a logic in a certain language is denable in all
stronger logics and/or bigger languages. The denability of connectives in fragments
48 Libor B ehounek, Petr Cintula, and Petr H ajek
Fragment: 0, , , , , , & , &, , &,
MTL
BL, SBL, ,
G &, & ,
&, , &, &, & ,
Table 3. Denability of connectives in fragments with 0
of prominent extensions of MTL is summarized in Tables 2 and 3, which list all of the
connectives &, , , 0 that are denable in each fragment (omitting those which are al-
ready present in the fragments language). The constant 0 is not denable in any of the
fragments listed (the empty column for the language , &, , is therefore omitted in
Table 2).
2.4 Adding axioms or rules
Another way of varying fuzzy logics is by strengthening it by additional axioms or
rules, thereby narrowing down the class of its algebraic models. In some cases, adding
an axiom or rule leads to a previously introduced logic (e.g., MTL plus the axiom of
idempotence of conjunction yields the same logic as BL plus the same axiom, namely
G odel logic). In other cases, however, a new fuzzy logic is obtained (e.g., MTL plus the
axiom of double negation yields the logic IMTL, which is weaker than an analogous
extension of BL, i.e., ukasiewicz logic).
We shall discuss two main kinds of such axiomatic strengthening, namely imposing
some restrictions on the behavior of logical connectives (esp. negation or conjunction)
and limiting the set of truth values to nite cardinalities. We shall introduce main fuzzy
logics arising by these kinds of strengthening, and briey discuss their properties and
mutual relationships.
2.4.1 Special properties of negation
Axioms or rules added to a given fuzzy logic can enforce special properties of resid-
ual negation. Prominent extensions of this kind are the logics IMTL, IUL, SMTL,
and SBL, which enforce either involutiveness or strictness (see Example 1.1.11) of resid-
ual negation.
The involutiveness of residual negation can be ensured in any extension of the uni-
norm logic UL by adding the axiom of double negation
() .
The converse implication is provable in all extensions of UL (in fact, in all
logics of pointed commutative residuated lattices, as (( 0) 0) follows by
modus ponens from residuation). Adding this axiom to BL yields ukasiewicz logic
(see Section 1.2). The extensions by () of MTL and UL are called, respectively,
IMTL [53] and IUL [133] (where the I stands for involutive).
38
The logic IMTL en-
38
The logic IMTL was implicitly studied already in [23] together with its algebras (called there m-z-
Chapter I: Introduction to Mathematical Fuzzy Logic 49
joys innitary-strong standard completeness, i.e., completeness w.r.t. all standard MTL-
algebras [0, 1]

with involutive residual negation [50]; the standard completeness of IUL

is an open problem.
Like in and the logics L

, a strong disjunction is denable in IMTL, by

( & ). Notice, however, that only a limited set of t-conorms is obtained by this
denition as the standard semantics of in IMTL, since the t-norm in standard IMTL-
algebras has to be such as to yield involutive residual negation.
The logic IMTL shares many features with ukasiewicz logic, which is its extension
by the divisibility axiom (BL4). Like in ukasiewicz logic (see Theorem 1.2.8), the
implication and strong conjunction are mutually inter-denable and the contraposition
law holds in IMTL. It can be shown that IMTL is equivalently axiomatized by adding
any of the following IMTL-provable formulae: (T

1), (T

3), (T

4), (T

5), or (T

6) of
Theorem 1.2.8. Note, however, that the Wajsberg axiom (T

2) is not provable in IMTL,

and adding it to MTL in fact yields ukasiewicz logic.
The strictness of residual negation (i.e., the fact that is a pseudocomplement) can
be enforced in any extension of MTL by adding the axiom
(S) ( ).
Analogously to the logic SBL = BL + (S) introduced in Section 1, we can dene the
logic SMTL as the extension of MTL by (S). Like SBL (which extends SMTL by
the divisibility axiom (BL4)), also SMTL can equivalently be axiomatized by any of
the formulae (T
SBL
1), (T
SBL
2), or (T
SBL
3) of Theorem 1.2.6. These theorems restrict
the behavior of negation and conjunction in SMTL-chains. In particular, the theorem
(T
SBL
1), , expresses the bivalence of negationi.e., the fact that is in each
SMTL-chain evaluated either to 0 or 1. The theorem (T
SBL
3), ( & ) ( ),
entails that SMTL-chains are exactly those MTL-chains that have no zero divisors (i.e.,
non-zero elements x, y such that x y = 0).
Adding the axiom (S) to IMTL, or the axiom () to SMTL, already makes the
logic classical. Recall, however, from Section 2.2.2 that an additional involutive nega-
tion can be naturally added to any logic containing SMTL, by the axioms (), (Rev-
S1), (Rev-S2), and (Rev-S3), and that is denable in SMTL

by . Strict
negation
S
is, on the other hand, denable already in MTL
Z
by
S
. Both
strict and involutive negation are also present in logics containing SMTL

or IMTL
Z
(incl., e.g.,
Z
or ). Thus, even though it is not possible to have a fuzzy negation that
is both involutive and strict, one can have both negations in richer fuzzy logics.
2.4.2 Special properties of conjunction
Various properties of conjunction & correspond to special axioms added to usual
systems of fuzzy logics. For instance, the idempotence of & corresponds to the validity
of the axiom
(C) &
in any extension of the uninorm logic UL. Adding the axiom (C) to MTL (or any logic
between MTL and G, e.g., BL, SBL, or SMTL) yields G odel logic; adding (C) to IMTL
pregroups).
50 Libor B ehounek, Petr Cintula, and Petr H ajek
makes the logic classical. Adding the axiom to the uninorm logic UL yields the logic
called UML, or uninorm mingle logic. UML is the logic of idempotent uninorms, char-
acterized by Theorem 2.1.4. Extending the logic UML further by the axioms 0 0
and 0 1 yields the logic IUML of the single idempotent uninorm
IU
introduced in
Theorem 2.1.4; in both cases, even the innitary-strong standard completeness of the
logic can be proved [133]. Notice that while (C) entails the coincidence of & and in
all t-norm logics, this is not in general so in uninorm logics, as idempotent residuated
uninorms need not coincide with the minimum.
The axiom (C) can be called the axiom of contraction, as it expresses the fact that
any conjunction & . . . & of the same conjuncts can be contracted to a single .
39
Writing the axiom (C) as
2
suggests a natural generalization
(C
n
)
n

n1
for any n 2, called n-contraction.
40
Semantically, the validity of (C
n
) is equivalent
to the condition that x
n1
is an idempotent element of & for each x. For any axiomatic
extension L of MTL we dene the logic C
n
L as L + (C
n
). Clearly, (C
n
) implies (C
m
)
in MTL for all m n; thus C
m
L C
n
L if m n. In the logic C
n
L, a global bound
can obviously be given in the deduction theorem:
,
C
n
L
iff
C
n
L

n1
.
Since C
2
MTL is G odel logic, the logics C
n
MTL are intermediary between MTL and G.
It can be proved that C
n+1
MTL _ C
n
MTL and that MTL =

n
C
n
MTL, and sim-
ilarly for C
n
IMTL, C
n
BL, and C
n
. For each n, the logics C
n
MTL and C
n
IMTL
are innitary-strong standard complete [27]. This is not the case with C
n
BL and C
n
,
though, as the only standard C
n
BL-chain is [0, 1]
G
and the only C
n
-chains are MV-
chains with less than n elements.
Let us say that an MTL-algebra A is n-contractive if (C
n
) is valid in A. It can be
observed that, for instance, every nite MTL-algebra is n-contractive for some n. All
weak nilpotent minimum t-algebras (see Example 2.1.1) are 3-contractive and standard
G odel algebra is clearly weak nilpotent minimum t-algebra. Therefore the logic WNM
of weak nilpotent minima, introduced in [53], is intermediary between C
3
MTL and
C
2
MTL = G; it can be easily shown that both inclusions are proper. The logic WNM
can be axiomatized by adding the axiom
41
(WNM) ( & ) ( & )
to MTL. The innitary-strong standard completeness of WNM can be proved [53].
39
In substructural logics, usually (e.g., in [67]) just the implication & is called the axiom of
contraction (as it corresponds to the structural rule of contraction, cf. Section 3.1), while the converse impli-
cation & is called the axiom of expansion. Nevertheless, since the latter is a theorem of MTL, the
distinction does not matter in MTL or stronger logics.
40
Again, in substructural logics the termn-contraction is used just for the implication
n+1

n
and the
property (C
n
) is (e.g., in [67]) called n-potence. In the context of ukasiewicz logic, n-contraction has been
studied in [157]; the systematic study of n-contraction in mathematical fuzzy logic has originated with [27].
41
Observe that its rst disjunct represents the nilpotent part and the second disjunct the idempotent part of
a weak nilpotent minimum t-norm.
Chapter I: Introduction to Mathematical Fuzzy Logic 51
The logic NM of nilpotent minima (i.e., weak nilpotent minima given by an invo-
lutive negation, see Example 2.1.1), introduced in [53], can be shown to extend WNM
by the axiom (), ; thus NM is an extension (also proper) of C
3
IMTL. All
NM-algebras on the real unit interval [0, 1] are mutually isomorphic. Because of this
fact, only the t-algebra [0, 1]
NM
of the standard nilpotent minimum t-norm
NM
(given
by the standard involution 1 x) is called the standard NM-algebra.
42
The innitary-
strong standard completeness theorem can be proved for the logic NM [53].
The logic NM expanded by is in a close relationship to G odel logic with involu-
tion, as the standard connectives of NM
Z
and G

turn out to be denable in each other

by mutually inverse translations. Thus, due to the standard completeness of both logics,
NM
Z
and G

have the same logical strength and can be considered notational variants
of each other.
Another property of & that can be axiomatically enforced in extensions of MTL
is n-nilpotence, i.e., the identity x
n
= 0 for each x ,= 1. For any given n 2, the
(n 1)-nilpotence of & is ensured by the axiom
(S
n
)
n1
.
For any extension L of MTL, the logic L + (S
n
) is denoted by S
n
L. These logics
were studied in [113]; the axioms (S
n
) were studied under the name n-excluded middle
in [124] in the context of substructural logics, and rst brought to mathematical fuzzy
logic in [73]). The name (unrelated to SBL and SMTL) and the numbering of the
axioms (S
n
) is motivated by the relation (described below) between the logics S
n
MTL
and C
n
MTL. Clearly S
m
L S
n
L for m n, as MTL proves (S
n
) (S
m
) if n m.
The axiom (S
2
) is the law of excluded middle, thus all logics S
2
L coincide with classical
logic.
Recall that an algebra A is simple if it has only trivial congruences (i.e., its only
congruences are A
2
and the identity on A), and semisimple if it is a subdirect product
of simple algebras. S
n
MTL-algebras can be characterized as n-contractive semisimple
MTL-algebras. Consequently, S
n
L extends C
n
L for each n; it can be shown that for
L = MTL the inclusions are strict for any n 3, but on the other hand, S
n
= C
n

for all n.
It can be observed that S
3
MTL-chains are just those in which & is the so-called
drastic product, i.e., x &y = 0 for all x, y < 1. Since the drastic product on [0, 1] is not
left-continuous, there are no standard S
3
MTL-algebras. Generally it can be shown that
every simple n-contractive MTL-chain must have a co-atom (i.e., the largest element
smaller than 1); thus there are no standard S
n
MTL-algebras for any n. Consequently,
none of the logics S
n
MTL (nor S
n
IMTL) can enjoy any kind of standard completeness.
Earlier we have seen that the logic IMTL arises by adding to MTL the characteristic
axiom () of ukasiewicz logic over BL (cf. Section 1.2), while extending MTL by
the characteristic axiom (G) = (C
2
) of G odel logic over BL already yields G odel logic
itself. We can now ask what logic arises by adding the characteristic axiom of product
logic over BL,
42
Cf. the similarly motivated departure from Convention 2.0.2 in the cases of ukasiewicz and product
logic.
52 Libor B ehounek, Petr Cintula, and Petr H ajek
() (( & ) ),
to MTL. The resulting logic MTL = MTL + () is the logic of cancellative
MTL-algebras (i.e, those validating cancelation by non-zero elements: if z ,= 0 and
x z = y z, then x = y, for all x, y, and z), just like = BL + () is the logic
of cancellative BL-algebras. This logic was introduced in [86]. Also just like over BL
(see Section 1.2), the axiom () can equivalently be replaced by the two axioms (S) and
(); thus MTL = MTL + (S) + (
S
), too, and the logic is intermediary between
SMTL and (both inclusions are strict). The logic MTL is nitely strong standard
complete [108] (though not innitary-strong standard complete [110]).
The logic SBL contains both G and , since it is based on a common property of

G
and

, namely the strictness of their residual negation (see Example 1.1.11), but it
turns out to be strictly weaker that the intersection of these two logics. Similarly there is
a common property of

and

, related to the fact that

is isomorphic to truncated

(see Theorem 1.1.7(4)), namely their cancellativity restricted to elements with non-zero
conjunction. This property is called weak cancellativity:
DEFINITION 2.4.1. An MTL-chain A = L, , , , , 0, 1 is weakly cancellative if
x z = y z ,= 0 implies x = y for all x, y, z A.
Weak cancellativity is in MTL characterized by the following axiom:
(WC) ( & ) (( & ) ).
For any extension of MTL, the logic L +(WC) is denoted by WCL; the logics WCBL
and WCMTL were introduced in [140]. The following facts about the logics WCL can
be shown [140, 145]:
The logic WCBL is exactly the intersection of and .
S
n
L = C
n
L for any axiomatic extension L of WCMTL.
Adding (WC) to IMTL already yields ukasiewicz logic.
The logics WCBL and WCMTL enjoy nite strong standard completeness.
Note that the rst two claims entail the already mentioned fact that S
n
= C
n
.
2.4.3 Finitely-valued fuzzy logics
Let the logic L be an axiomatic extension of MTL and n 1 a natural number. We
dene the logic L
n
as the extension of L by the axiom
(n)
n

i=1
(
i1

i
).
It can be easily demonstrated that an L-chain A is an L
n
-chain if and only if A has at
most n elements: just observe that the axiom (n) is not satised by an A-evaluation e
iff e(
0
) > e(
1
) > > e(
n
), thus it holds for all evaluations and all formulae
i
iff there are at most n values in A.
Clearly L
n
cannot enjoy standard completeness, but it satises a bunch of other
interesting properties [35]:
Chapter I: Introduction to Mathematical Fuzzy Logic 53
Figure 4. Relative positions of prominent logics introduced in Section 2.4.
Strong niteness: There is a nite set K of nite algebras (namely, the set of all
L
n
-chains on subsets of 1, . . . , n) such that
L
n
= [=
K
.
Tabularity: There is a nite L
n
-algebra F such that
L
n
= [=
F
.
Finite embeddability property: Every nite subset of any L
n
-algebra can be
partially embedded into a nite L
n
-algebra.
Note that L
n
extends C
n
L (see Section 2.4.2). Let us now briey discuss the situation
in particular logics. First let us note that the only nite MTL-chains (so a fortiori
the only nite product chains) are two-element Boolean algebras. Consequently, the
only nite MTL-algebras are nite Boolean algebras, and so there are no non-trivial
(i.e., non-classical) nitely-valued MTL (nor product) logics. Actually even more can
be shown: the only proper extensions (not necessarily axiomatic!) of product logic are
either classical or inconsistent.
Further let us observe that for each n there is only one (up to isomorphism) n-valued
MV and G-chain. Let us denote these algebras by MV
n
and G
n
, respectively, and by

n
and G
n
denote the corresponding logics semantically induced by these algebras.
First let us describe the situation in nitely-valued G odel logics [76]:
G
n
= G
n
and G =

n=1
G
n
.
G
n
G
m
iff m n.
For each proper axiomatic extension L of G there is n such that L = G
n
.
The situation in nitely-valued ukasiewicz logics is more complex. While it is
still true that =

n=1

n
, it is not true that
n

m
iff m n. For instance,
54 Libor B ehounek, Petr Cintula, and Petr H ajek
Figure 5. Relative positions of further prominent logics of Section 2.4.
(2p)
4
4(p
2
) and ((p p)
3
) are tautologies of
4
, though not of
3
(where
p
n

&
n
i=1
p as usual and np

n
i=1
p). Therefore clearly
n
,=
n
. The (non-
linear) ordering by strength of the logics
n
given by the single nite MV-chain MV
n
is described by the following theorem which in [129] is attributed to Lindenbaum:
THEOREM 2.4.2.
n

m
iff n 1 divides m1, for any m, n 2.
Unlike in G odel logic, there are proper axiomatic extensions of ukasiewicz logic that
differ from [=
K
for any set K of nite MV-chains (e.g. the logic of the so-called Chang
MV-algebra, see Chapter VI).
3 Families of fuzzy logics in the logical landscape
In previous sections we introduced numerous prominent members of the broad fam-
ily of logical systems studied in mathematical fuzzy logic, their axiomatic systems, and
their general, linear, and standard algebraic semantics. The multitude of fuzzy logics
calls for a general unifying (meta)theory. Indeed, many metamathematical properties,
such as general and linear completeness theorems or the linear subdirect representation
theorem, can be proved generally for large classes of logics delimited by simple (e.g.,
syntactic) criteria.
In this section we shall study the position of fuzzy logics in the logical landscape
i.e., their relationship to well-known broader families of propositional logics (such as
substructural or algebraizable) as well as to particular prominent non-classical logics.
We shall briey survey characteristic properties of fuzzy logics, both those shared with
the mentioned broad classes of logics and those particular to fuzzy logic.
In the rst subsection we show the position of fuzzy logics among substructural
logics. Then, in Section 3.2, we shall describe two important classes of fuzzy logics,
Chapter I: Introduction to Mathematical Fuzzy Logic 55
namely the core and -core fuzzy logics. These classes will play an important r ole
in the remaining sections of this Chapter as well as in some subsequent chapters. Fi-
nally, in Section 3.3, we shall recall some basic notions of Abstract Algebraic Logic,
particularized for our needs (this theory is covered in much more detail in Chapter II).
3.1 Fuzzy logics among substructural logics
In this subsection we show that fuzzy logics introduced in Section 2.1 can be seen
as the logics of (suitable classes of) pointed residuated lattices
43
(see Denitions 1.3.1,
2.1.5, and 2.1.6).
(Pointed) residuated lattices form the algebraic semantics for so-called (intuitionis-
tic) substructural logics. Substructural logics provide a unifying framework for several
kinds of logics, such as relevance logics, variants of Girards linear logic, the Lambek
calculus, the logic BCK, etc. In a specic sense (suggested by Ono in [151]), substruc-
tural logics can be delimited as the logics of varieties of residuated lattices. As we have
seen, many fuzzy logics fall within this delimitation, and can thus be seen as a special
kind of substructural logics. Substructural logics thus form a neighborhood of fuzzy log-
ics in the landscape of non-classical logics. In this section we shall briey introduce the
class of (intuitionistic) substructural logics, indicate the position of fuzzy logics within
this family, and describe the relationship of fuzzy logics to other substructural logics.
For details on substructural logics, including the relationship of some prominent fuzzy
logics to this class, see esp. [54, 67, 151, 160].
The logic of all (pointed) residuated lattices (with elements x 1 taken as desig-
nated in the denition of logical consequence) is called the full Lambek calculus FL.
The name comes from the fact that its conjunctionimplication fragment is the well-
known Lambek Calculus (an important tool in the study of categorial grammars). The
full Lambek calculus is an expansion of the Lambek calculus to the full language, con-
taining also the lattice connectives, the dened connectives of negation and equivalence,
and the propositional constants for truth and falsity.
Proof-theoretically, the logic FL arises by removing the structural rules of exchange,
weakening, and contraction from the Gentzen-style calculus LJ of intuitionistic logic.
44
In the absence of structural rules, certain equally well motivated variants of the opera-
tional rules for propositional connectives become non-equivalent, and thus dene differ-
ent connectives. This explains why in substructural logics (and consequently in fuzzy
logics as their special kind) propositional connectives known from intuitionistic or clas-
sical logic naturally split into pairs of different connectives (that have to be included in
the general signature of a pointed residuated lattice). In particular, implication and nega-
43
I.e., residuated lattices with an additional, arbitrarily interpreted nullary connective 0. The prex pointed
is often omitted, as the classes of algebras differ only in signature. The term FL-algebras is sometimes used
instead.
44
Recall that Gentzen-style sequent calculi have two kinds of rules: operational rules for introduction
of propositional connectives, and structural rules for manipulation with whole formulae in sequents. The
Gentzen-style calculi LJ and LK for intuitionistic resp. classical logic have (besides the indispensable rule
introducing the axiom sequents and the eliminable rule of cut) exactly the structural rules of exchange, weak-
ening, and contraction; the logic FL is obtained by removing all these three structural rules from LJ. For
more details on Gentzen-style calculi LJ and LK, their structural rules, and substructural logics in general,
see, e.g., [67, 151, 153, 160].
56 Libor B ehounek, Petr Cintula, and Petr H ajek
tion split into two variants (left and right) in the absence of exchange (cf. the connectives
and in the non-commutative fuzzy logics of Section 2.1.3, and the corresponding
negations 0 and 0); the constants for truth and falsity split each into two in the ab-
sence of weakening (cf. the distinction between 1, 0 and , in uninorm fuzzy logics,
Section 2.1.2); and conjunction splits into two in the absence of contraction or weak-
ening (cf. the presence of lattice conjunction and residuated conjunction & in t-norm
fuzzy logics), as does disjunction ( vs. ) in contraction-free logics with involutive
negation.
45
The full propositional language of the logics of residuated lattices (includ-
ing fuzzy logics) is therefore assumed to contain all of these connectives.
46
Under the
presence of the appropriate structural rules (or the equivalent axioms), these connectives
collapse into the single variants known from classical (or intuitionistic) logic.
The basic substructural logics are obtained by extending FL by a subset of the fol-
lowing Hilbert-style axioms, which correspond to the Gentzen-style structural rules of
exchange, weakening, and contraction of LJ (so the resulting logics arise by removing
just some of the three structural rules from LJ), and the law of double negation (which
corresponds to starting from the calculus LK for classical logic instead of LJ for intu-
itionistic logic):
(E) & &
(W) ( 1) (0 )
(C) &
() ((0 ) ( 0)) ((( 0) 0) )
The algebraic properties of pointed residuated lattices corresponding to these axioms
are, respectively: commutativity, boundedness and integrality,
47
square-increasingness
(also know as superidempotence), and the classicality of residual negation
48
(see Def-
initions 1.3.1, 2.1.5, and 2.1.6).
45
The rules for disjunction affected by the absence of contraction operate with two or more formulae on the
right-hand side of the sequent, which is forbidden in the calculus LJ. The split of disjunction thus does not
occur in contraction-free logics based on LJ, and only occurs in contraction-free logics based on the calculus
LK for classical logic (or equivalently, if the double negation law is added to LJ). The corresponding rules
for , on the other hand, operate on the left-hand side, upon which no restriction is imposed in LJ; the split of
conjunction thus occurs in all contraction-free logics of residuated lattices, including all t-norm and uninorm
fuzzy logics (except for G odel logic, which is contractive).
46
In the context of substructural logics, the residual conjunction &and its dual disjunction are often called
fusion and ssion, respectively, or alternatively, multiplicative, group, parallel, or intensional conjunction
and disjunction. The connectives and are often called additive, lattice, comparative, or extensional
conjunction and disjunction. The names strong and weak conjunction or disjunction, common in t-norm
fuzzy logics, are not suitable in the absence of weakening, since & is not generally stronger than without
weakening.
47
The rst conjunct of (W) corresponds algebraically to the integrality of a pointed residuated lattice and
proof-theoretically to the rule of weakening on the left-hand side of sequents; similarly the second corresponds
to boundedness and the weakening on the right (restricted to a single formula in LJ). Often (see, e.g., [68]),
the conjuncts are discussed separately and denoted by i and o in the subscripts indicating extensions of FL.
48
The rst conjunct of () corresponds algebraically to the cyclicity of residual negation (i.e., 0 / x =
x \ 0), and the second to its involutiveness (which in general needs to be formulated as 0 / (x \ 0) = x
and (0 / x) \ 0 = x, but in the presence of the rst condition can be formulated in our way). Note that in
commutative substructural logics, () can be replaced by (( 0) 0) .
Chapter I: Introduction to Mathematical Fuzzy Logic 57
FL
FL
c
FL
e
FL
w
FL
ec
FL
ew
Int
CFL
CFL
c
LL
CFL
w
R
ND
ALL
Bool
Figure 6. Basic intuitionistic and classical substructural logics.
The logic CFL = FL+() is called the classical full Lambek calculus. The name
refers to the fact that proof-theoretically it can be dened by removing the three struc-
tural rules from the calculus LK of classical logic. Extending FL or CFL by a subset of
the axioms (E), (W), (C) yields the basic (intuitionistic or classical, resp.) substruc-
tural logics, systematically denoted by subscripting FL resp. CFL by corresponding
lowercase letters. The logic FL
ew
is also known as H ohles monoidal logic [107]; CFL
e
as Girards (multiplicativeadditive) linear logic [72] (without exponentials and additive
constants), LL; CFL
ew
as afne linear logic [172], or Grishins logic [80]; FL
e(w)
as
intuitionistic (afne) linear logic I(A)LL; and CFL
ec
as Meyers relevance logic R
ND
(or R minus distribution) or LR (for lattice R) [48]. Since (E) is provable in FL
cw
,
the logics FL
cw
= FL
ecw
coincide with intuitionistic logic Int and CFL
cw
= CFL
ecw
with classical logic Bool; all other logics FL
x
and CFL
x
are mutually different. The
relationships between the 14 basic substructural logics are depicted in Figure 6.
Among the logics of residuated lattices, fuzzy logics introduced in Section 2 are
distinguished by the property of semilinearity, i.e., completeness w.r.t. a class of linearly
ordered pointed residuated lattices. The main scope of mathematical fuzzy logic thus
can be delimited as the study of intuitionistic substructural semilinear logics, or the
logics of linearly ordered pointed residuated lattices.
49
Indeed, many important fuzzy logics arise as semilinear extensions of basic sub-
structural logics, i.e., as the logics of linearly ordered FL
x
- or CFL
x
-algebras [54, 133].
For instance, the logic MTL turns out to be the logic of linear FL
ew
-algebras; UL of
linear FL
e
-algebras; psMTL
r
of linear FL
w
-algebras; IMTL of linear CFL
ew
-algebras;
49
The name deductive fuzzy logics was proposed for this class of logics in [10], based on Onos formulation
in [151] suggesting that residuation gives logics a deductive face. The connection of ukasiewicz logic
to other prominent substructural logics has rst been pointed out in [23], and fuzzy logics have been rmly
established as members of the family of substructural logics in [54, 153]. A formal delimitation of the class of
fuzzy logics by the semilinearity property was proposed and advocated in [14]; cf. Section 3.3.
58 Libor B ehounek, Petr Cintula, and Petr H ajek
and G of linear FL
cw
-algebras (i.e., linear Heyting algebras). The restriction of the al-
gebraic semantics to linear algebras will systematically be denoted by the superscript .
We thus have the following identities:
FL

w
= psMTL
r
FL

e
= UL CFL

e
= IUL
FL

ec
= UML
FL

ew
= MTL CFL

ew
= IMTL
FL

cw
= G CFL

cw
= Bool
Other important fuzzy logics arise as extensions of these fundamental fuzzy logics by
special axioms or rules (e.g., SMTL = FL

ew
+ (S), IUML = CFL
ec
+ (0 1), etc.),
expansions by additional connectives with appropriate axioms and rules (e.g., , log-
ics with or etc.), fragments discarding some connectives (e.g., hoop fuzzy logics),
logics dened by a combination of these methods, and similar modications. More-
over, several fuzzy logics that have not yet been thoroughly investigated are obtained
in this way, e.g., FL

or FL

c
. Also the well-known substructural logic RM (relevance
with mingle) is itself semilinear (i.e., complete w.r.t. linearly ordered algebras), and so
belongs to the family of fuzzy logics in this sense.
Most prototypical fuzzy logics (including all t-norm fuzzy logics) are semilinear
extensions of FL
ew
. The position of these fuzzy logics among other substructural logics
is indicated in Figure 7. For axiomatic extensions of FL
ew
, semilinearity is equivalent to
the axiom( )( ). In axiomatic extensions of FL
e
(which include uninorm
logics), semilinearity is equivalent to (( ) 1) (( ) 1). More details on
the relationship between fuzzy and substructural logics can be found in Section 4.3 and
Chapter II, Chapter IV, and Chapter III
3.2 Core and -core fuzzy logics
The two classes of fuzzy logics we are going to introduce in this subsection are
not very broad from the general perspective of the whole logical landscape: in fact,
they do not even cover the majority of fuzzy logics introduced in the previous section.
Nevertheless, they do cover the most studied ones: the absolute majority of papers on
mathematical fuzzy logic actually study logics from these two classes, and the study
of other fuzzy logics has started only recently. The classes of logics were introduced
in [98] in order to provide a common framework to the study of rst-order fuzzy logics;
later they played a similar r ole in the general study of completeness of fuzzy logics w.r.t.
distinguished semantics in [35] (see Sections 5 and 4.1). The rough idea is to dene
a class of logics that share most desirable properties with MTL, and which could be
delimited in a simple syntactic way.
As we have seen in the previous section, we need some exibility as regards both
propositional languages and logics. Therefore, for the sake of reference and in order to
x terminology in a way convenient for this section, we shall start with some standard
general denitions and conventions. (For a detailed treatment of the general theory of
logical calculi see, e.g., [42, 178].)
Chapter I: Introduction to Mathematical Fuzzy Logic 59
FL
ew
= IALL
Int CFL
ew
= ALL
Bool
MTL
G IMTL
superintuitionistic
linear
fuzzy
(C)
(C)
(C)
(C)
()
()
()
()

Figure 7. The position of t-norm fuzzy logics among substructural logics
DEFINITION 3.2.1. A (propositional) language is a pair / = Conn
1
, Ar
1
, where
Conn
1
is a well-ordered countable set of (propositional) connectives and Ar
1
is a
function assigning a natural number to each element of Conn
1
. The number Ar
1
(c) is
called the arity of c Conn
1
. We shall write c, n / as a shorthand for c Conn
1
and Ar
1
(c) = n. Nullary connectives are also called truth-constants.
The set Form
1
of (propositional) formulae in the language / over the xed de-
numerable set Var of (propositional) variables is the smallest set containing Var, the
truth constants of /, and closed under the connectives from / (that is, c(
1
, . . . ,
n
)
Form
1
whenever
1
, . . . ,
n
Form
1
and c, n /).
An /-substitution is a mapping : Form
1
Form
1
commuting with the connec-
tives of / (i.e., (c(
1
, . . . ,
n
)) = c((
1
), . . . , (
n
)) for each c, n /).
By a logic (in the language /) we mean a substitution-invariant Tarski consequence
relation over /; i.e., a relation
L
T(Form
1
)Form
1
that satises, for each ,
Form
1
, each , Form
1
, and each /-substitution , the following conditions:
1. If , then
L

2. If
L
for each , and
L
, then
L

3. If
L
, then ()
L
().
We shall often use just L as a synonym for
L
. A logic L is nitary if for each
Form
1
and Form
1
such that
L
there exists a nite
t
such that

t

L
.
Note that the cardinality restrictions on the sets / and Var are assumed just for
60 Libor B ehounek, Petr Cintula, and Petr H ajek
simplicity. Clearly for each nitary
50
logic in the sense of this denition there is a
Hilbert-style calculus (cf. Denition 1.2.1) such that the relation of provability (cf. Def-
inition 1.2.2) in this calculus coincides with the logic; all such calculi will be called
axiomatic systems (or presentations) of the logic in question.
EXAMPLE 3.2.2. All logics of (sets of) continuous t-norms (see Denition 1.1.19) are
nitary logics in the sense of Denition 3.2.1. The innitary logics of (sets of) continuous
t-norms (i.e., the semantical consequence relations [=
K
of Denition 1.1.15) are also
logics in the sense of this denition, but they are not nitary (with the exception of
K =
G
, see Theorem 1.1.18). Also all logics introduced in Section 2 are examples
of nitary logics.
DEFINITION 3.2.3. We say that a logic L
t
in the language /
t
is an expansion of a logic
L in the language / /
t
if
L
implies
L
for each Form
1
. We
say that the expansion is conservative if the converse implication holds as well. We say
that the expansion is axiomatic if L
t
can be axiomatized by adding some axioms (but no
rules) to some axiomatic system of L. We use the term extension instead of expansion if
/ = /
t
.
Now we are able to dene the core and -core fuzzy logics.
DEFINITION 3.2.4. We say that a nitary logic L in a language / is a core fuzzy logic,
if:
1. L expands MTL.
2. For all /-formulae , , the following holds:

L

t
, (Cong)
where
t
is a formula resulting from by replacing some occurrences of its sub-
formula by a formula .
3. L has the Local Deduction Theorem, i.e., for each set of /-formulae T ,
holds:
T,
L
iff there is n N such that T
L

n
. (/TT )
EXAMPLE 3.2.5. The following logics introduced in the previous sections are core
fuzzy logics: BL, MTL, IMTL, MTL, NM, WNM, WCMTL, SBL, ukasiewicz,
product, and G odel logic; the n-valued variants of these logics; extensions of these
logics by the axiom (C
n
) or (S
n
); and some logics in expanded languages: the logic
P, Rational Pavelka Logic, and expansions of all mentioned logics by truth constants.
Note that classical logic is a core fuzzy logic as well, but intuitionistic logic is not.
Many important fuzzy logics, incl., e.g., the logics MTL
Z
and , however, fall outside
the class of core fuzzy logics. Therefore we introduce a second, analogously dened,
class:
50
The denition of a Hilbert-style calculus can be modied to cover innitary logics as well, but we shall
not need it in this Chapter.
Chapter I: Introduction to Mathematical Fuzzy Logic 61
DEFINITION 3.2.6. We say that a nitary logic L in a language / is a -core fuzzy
logic, if:
1. L expands MTL
Z
2. For all /-formulae , , the following holds:

L

t
, (Cong)
where
t
is a formula resulting from by replacing some occurrences of its sub-
formula by a formula .
3. L satises the -Deduction Theorem, i.e., for each set of /-formulae T ,
holds:
T,
L
iff T
L
. (TT
Z
)
EXAMPLE 3.2.7. The following logics introduced in the previous sections are -core
fuzzy logics: extensions of all core fuzzy logics by ; some fuzzy logics with additional
involutive negation (SMTL

, SBL

, G

); and the logics and

1
2
.
The only consistent logic which is both core and -core is classical logic. There-
fore all -core logics are examples of logics which are not core fuzzy logics. Of course
all logics weaker than MTL are neither core nor -core (including, e.g., fragments of
MTL, the uninorm logic UL, the non-commutative fuzzy logic psMTL
r
, etc.). How-
ever, there are natural examples of such logics even among expansions of MTL:
51
EXAMPLE 3.2.8. The following logics introduced in the previous sections are neither
core nor -core fuzzy logics (because they neither enjoy /TT nor expand MTL
Z
):
some fuzzy logics with additional involutive negation (MTL

, BL

); and the
logic P
t
.
The following theorem gives an alternative denition of (-)core fuzzy logics. (It
is a direct consequence of [34, Corollary 8 and Theorem 6].)
THEOREM 3.2.9. Let L be an expansion of MTL (respectively, of MTL
Z
) that satis-
es the condition (Cong). Then L is a core (resp., -core) fuzzy logic if and only if it is
an axiomatic expansion of MTL (MTL
Z
, resp.).
The notion of L-algebra can be generally dened for (-)core fuzzy logics as fol-
lows:
DEFINITION 3.2.10. Let L be a core fuzzy logic and let C be the set of connec-
tives of L that are not present in MTL. An L-algebra is a structure A = A, &,
, , , 0, 1, (c)
cC
such that A, &, , , , 0, 1 is an MTL-algebra, and for every
axiom of L, the identity e() = 1 holds under all A-evaluations e.
Analogously we dene L-algebras for -core fuzzy logics. An L-algebra is called
an L-chain if its MTL-reduct is an MTL-chain. The class of all L-algebras is denoted
by '.
51
We have not seen (and will not see) any logic that does not satisfy (Cong).
62 Libor B ehounek, Petr Cintula, and Petr H ajek
The following theorem collects the basic properties of (-)core fuzzy logics. We
can see that these logics share many important logical and algebraic properties with the
logics BL and MTL.
THEOREM 3.2.11 ( [35, 98]). Let L be a (-)core fuzzy logic. Then:
1. ' is a variety of algebras.
2. The lattice of nitary extensions of L is dually isomorphic to the lattice of sub-
quasivarieties of '.
3. The lattice of axiomatic extensions of L is dually isomorphic to the lattice of
subvarieties of '.
4. L satises the proof by cases property PCP (cf. Theorem 1.2.11): for each theory
and formulae , , ,
,
L
whenever ,
L
and ,
L
.
5. L satises the semilinearity property SLP (cf. Theorem 1.2.11): for each theory
and formulae , , ,

L
whenever ,
L
and ,
L
.
6. For each theory and each formula such that ,
L
there is a linear (or
equivalently, prime) theory
t
such that
t
,
L
(cf. Theorem 1.2.13).
7. Every L-algebra is representable as a subdirect product of L-chains (cf. Theo-
rem 1.3.11).
8. The class of nitely subdirectly irreducible L-algebras coincides with the class of
L-chains (cf. Theorem 1.3.13).
9. The following conditions (cf. Theorems 1.3.7 and 1.3.12) are equivalent for every
theory and a formula :

L

e() = 1 for each L-algebra Aand any A-model e of
e() = 1 for each L-chain Aand any A-model e of .
Using the axioms of , we can determine the semantics of in L-chains:
THEOREM 3.2.12. Let L be a -core fuzzy logic and Aan L-chain. Then
A
x = 1
A
if x = 1
A
, and
A
x = 0
A
otherwise.
For each core fuzzy logic L we can dene the corresponding -fuzzy logic L
Z
resulting from L in the same way as MTL
Z
from MTL. The following two results are
straightforward corollaries of the previous two theorems.
Chapter I: Introduction to Mathematical Fuzzy Logic 63
THEOREM 3.2.13. For every core fuzzy logic L, the logic L
Z
is a conservative expan-
sion of L.
THEOREM 3.2.14. Let L be a nitary expansion of MTL
Z
satisfying (Cong). Then L
is a -core fuzzy logic if and only if L is strongly complete w.r.t. L-chains (see the last
condition in Theorem 3.2.11).
3.3 Fuzzy logics as algebraically implicative semilinear logics
The class of core fuzzy logics is quite broad, but it still does not cover all fuzzy
logics mentioned in the Section 2, mainly because these logics are weaker than MTL.
As we have seen in the previous subsection, there is a large family of such logics that
is extensively studied in the literature. Here we shall briey introduce fundamentals of
an abstract theory of propositional fuzzy logics; for a detailed exposition and references
see Chapter II of this Handbook.
DEFINITION 3.3.1 ( [34, 158]). A logic L in the language / is a weakly implicative
logic if / contains a binary connective such that:

L

,
L

,
L

,
L
c(
1
, . . .
i
, , . . . ,
n
) c(
1
, . . .
i
, , . . . ,
n
),
for every c, n / and i < n.
A weakly implicative logic is Rasiowa-implicative if

L
.
Note that the last condition in the denition of weakly implicative logics gives us
that for each set of formulae T , , it holds:
,
L
() (). (Cong)
EXAMPLE 3.3.2. All logics mentioned so far in this chapter are weakly implicative,
including all substructural logics. (In non-commutative logics, the r ole of can be
played by both and .)
The logic FL
w
and all its extensions, including all logics of (sets of) (left-)continuous
t-norms, as well as all ()-core fuzzy logics are Rasiowa-implicative, and so are all
fragments of these logics that contain implication (see Section 2.3).
It can be shown that Rasiowa-implicative logics share an important common feature
of core fuzzy logics: namely that in their natural algebraic semantics there is always just
a single element which is designated (i.e., is regarded as fully true in the denition
of the logics consequence relation), and this element can be dened as the one satis-
fying the equation x = 1. This is no longer true in UL (which is clearly not Rasiowa-
implicative), where the designated truth values are those bigger than 1; nevertheless, we
64 Libor B ehounek, Petr Cintula, and Petr H ajek
can still dene these truth values as those satisfying the equation x 1 = 1. To en-
compass also logics like UL together with their natural algebraic semantics we dene
a broader class of algebraically implicative logics where the designated truth values are
equationally denable. Let us note here that nitary algebraically implicative logics are
exactly those weakly implicative logics which are algebraizable in the sense of Blok and
Pigozzi [18].
Let us now formalize this condition and express it in a purely syntactic way.
52
Ob-
serve that the logics and their algebraic semantics we have studied throughout this chap-
ter have the following two properties: (i) if the logic proves and is designated,
then so is , and (ii) x y and y x are both designated iff x = y.
DEFINITION 3.3.3. A weakly implicative logic L is algebraically implicative if there is
a pair of formulae
L
(p),
L
(p) of single variable p such that for each formula holds:

L

L
()
L
(),

L

L
()
L
(),

L
()
L
(),
L
()
L
()
L
.
EXAMPLE 3.3.4. Every Rasiowa-implicative logic is algebraically implicative, via the
pair
L
(p) = p and
L
(p) = p p. (In fact, any theorem of L with at most one
variable can play the r ole of
L
e.g., the truth constant 1).
The Full Lambek logic FL and all its extensions (including, e.g., the uninorm logic
UL) are algebraically implicative, via the pair
L
(p) = p 1 and
L
(p) = 1 (or

L
(p) = p 1 and
L
(p) = p).
DEFINITION 3.3.5 (/-algebras). Let / be a propositional language. An /-algebra is
an algebra A = A, c
A

c,n)1
with the signature n
c,n)1
.
An A-evaluation is a mapping e: Form
1
A that commutes with the connectives
from / (i.e., e(c
A
(
1
, . . . ,
n
)) = c
A
(e(
1
), . . . , e(
n
)) for each c, n /.
The next denition embodies the slogan that
L
(p) =
L
(p) denes the designated
(fully true) truth values (cf. also Example 3.3.4).
DEFINITION 3.3.6 (Designated elements and models). Let L be an algebraically im-
plicative logic and Aan /-algebra. We dene the set D
A
of designated elements in A
as:
D
A
= x [
L
(x) =
L
(x)
We say that an A-evaluation e is A-model of a theory if e() D
A
for each .
Note that e() D
A
iff e(
L
()) = e(
L
()). In a more general setting there is
an abstract notion of logical matrix consisting of an algebra with a set of designated el-
ements. In algebraically implicative logic we assume that the designated set is uniquely
determined by an equation in the algebra; in the general setting, the situation is more
complex and we leave its study to Chapter II.
52
We provide a simplied account, assuming denability by a single pair of formulae. For the proper
denition see [41] or Chapter II.
Chapter I: Introduction to Mathematical Fuzzy Logic 65
DEFINITION 3.3.7 (Algebraic semantics). Let L be an algebraically implicative logic
and A an /-algebra. We say that A is an L-algebra if for each theory , formula ,
and elements x, y A it holds that:
If
L
then any A-model of is also an A-model of
If x
A
y D
A
and y
A
x D
A
, then x = y.
The class of all L-algebras will be denoted by '.
We could easily show that our general denition of ' for logics studied in this
chapter coincides with the particular explicit denitions provided while dening those
logics in the previous sections. It can be shown that it is sufcient to check the rst
condition for axioms and deduction rules of some (any) presentation of L. Let us now
sample some abstract variants of theorems shown for particular logics in the previous
sections. We start with results valid for all algebraically implicative logics, fuzzy or not.
THEOREM 3.3.8 ( [18, 41]). Let L be an algebraically implicative logic. Then:
1. Every extension of L is an algebraically implicative logic
2. The following conditions are equivalent for every theory and formula :

L

For each L-algebra Aand each A-model e of holds that e is an A-model
of
3. If L is nitary, then ' is a quasivariety of algebras
4. If L is nitary, then the lattice of nitary extensions of L is dually isomorphic to
the lattice of subquasivarieties of '
5. The lattice of axiomatic extensions of L is dually isomorphic to the lattice of
relative subvarieties of '
6. The relation
A
dened as x
A
y iff x
A
y D
A
(i.e., x is less than or
equal to y whenever the implication x
A
y is fully true) is an ordering on A
Note that the antisymmetry of
A
follows from the second condition in the deni-
tion of L-algebras. Also note that the majority of algebras related to logics introduced in
the previous sections were ordered (usually lattice-ordered) and that this order coincides
with the just dened order
A
. We say that an L-algebra is linearly ordered (or that it
is an L-chain) if
A
is a total order.
Let us now focus on fuzzy logics. Observe all logics that we have called fuzzy in
this chapter are complete w.r.t. their linearly ordered algebras.
53
Following the tradition
of Universal Algebra to call a class of algebras semi-X if its subdirectly irreducible
members have the property X (see the next theorem), we shall call such logics semilin-
ear.
53
In [14] it is argued that all and only such logics should be called fuzzy logics, since this property is
shared by a vast majority of logics studied in the literature under this name and does not apply to most logics
that are generally not labeled as fuzzy (e.g., intuitionistic logic); here we opt for a more neutral stance.
66 Libor B ehounek, Petr Cintula, and Petr H ajek
DEFINITION 3.3.9 (Semilinear logics, [41]). An algebraically implicative logic L is
semilinear if the following conditions are equivalent: for each theory and each for-
mula :

L

For each L-chain Aand each A-model e of holds that e is an A-model of
A theory T is linear in L if for each pair of formulae , holds T
L
or
T
L
. In logic with in the language we also dene: a theory T is prime in L
if for each pair of formulae , holds T
L
implies T
L
or T
L
.
The next central theorem shows that many properties proved separately in the liter-
ature on mathematical fuzzy logic are in fact instances of a general theorem.
THEOREM 3.3.10 ( [41]). Let L be a nitary algebraically implicative logic. Then the
following conditions are equivalent:
L is semilinear
L has SLP, i.e., for each theory and formulae , , holds:
if ,
L
and ,
L
, then
L

For every theory and every formula such that ,
L
, there is a linear theory

t
such that
t
,
L

Each L-algebra is a subdirect product of L-chains
L-chains are exactly the relatively nitely subdirectly irreducible L-algebras.
54
If the language of L contains a connective such that
,
L
and ,
L
,
we can equivalently add:
55
L proves ( ) ( ) and has PCP, i.e., for each theory and formulae
, , holds:
if ,
L
and ,
L
, then ,
L

L proves ( ) ( ), and for every theory and every formula such
that ,
L
, there is a prime theory
t
such that
t
,
L

54
I.e., they cannot be decomposed as a non-trivial subdirect product of L-algebras. The restriction to L-
algebras holds automatically if ' is a variety, because in that case each component of a subdirect product is a
homomorphic image of an L-algebra, and therefore an L-algebra.
55
The validity of the two rules required of is only necessary for proving that semilinearity implies the
following three notions; the converse direction holds generally.
Chapter I: Introduction to Mathematical Fuzzy Logic 67
L proves ( ) ( ), and for every nite theory , a formula such
that
L
, and a propositional variable p not occurring in and we have:
p [
L
p.
One of the problems frequently studied in the literature is how, given a substructural
logic L, to axiomatize the logic given by L-chains.
THEOREM 3.3.11 ( [41]). Let L be a nitary algebraically implicative logic. Let fur-
thermore the logic L

be dened as:
L
iff for each L-chain and each A-model e
of holds that e is an A-model of . Then L

is the least semilinear logic extending L.

The next theorem shows that if the language of L contains , we can easily nd an
axiomatization of the logic L

.
THEOREM 3.3.12 ( [41]). Let L be a nitary algebraically implicative logic in a lan-
guage containing , and let / be one of its axiomatic systems. Assume further that
,
L
and ,
L
.
Then the logic L

is axiomatized by / plus the axiom ( ) ( ) plus the

rules of the form p [ T p, for each rule T from /, where p is a
variable that does not occur in T and .
4 Metamathematics of propositional fuzzy logics
In this section we survey the basic metamathematical properties of fuzzy logics.
We restrict ourselves to (-)core fuzzy logics (see Section 3.2), even though most of
the denitions can be formulated and most of the results proven in weaker logics (e.g.,
non-commutative, non-integral, with restricted language, etc.).
The rst subsection deals with (different) forms of completeness theorems with re-
spect to some distinguished semantics. Note that we have already seen real-chain and
standard semantics; besides those, we will also study hyperreal-, rational- and nite-
chain semantics of fuzzy logics. The results in this subsection are of two kinds: (i)
general, i.e., for arbitrary semantics and/or arbitrary (-)core fuzzy logic; and (ii) par-
ticular, for concrete semantics and/or (-)core fuzzy logic. The general results are
from [35]; see Chapter II for generalization of these results results to a much wider
context. Known particular results are also summarized in [35]; see the corresponding
chapters of this Handbook for proofs of some of these results and detailed references.
The second subsection deals with functional representation, i.e., the question as to
which real-valued functions are expressible by formulae of prominent (-)core fuzzy
logic. This question is connected to a prominent algebraic question regarding the de-
scription of the free algebra in a given variety. We will not go into details here; see
Chapter IX for a detailed exposition, proofs, and references.
The third subsection briey introduces the proof theory of certain fuzzy logics, by
means of (hyper)-sequent Gentzen systems. The whole Chapter III of this Handbook
is dedicated to elaboration of this important aspect of mathematical fuzzy logic. The
nal subsection studies fuzzy logics from the viewpoint of computational complexity.
In particular, the complexity of important sets of formulae (tautologies, satisable for-
mulae) given by prominent (-)core fuzzy logic is determined in the subsection. Again,
68 Libor B ehounek, Petr Cintula, and Petr H ajek
the whole Chapter X is dedicated to this topic; an interested reader can nd detailed
exposition, proofs, and references there.
4.1 Completeness theorems
Let us x a (-)core logic L. For a class K of L-chains we dene the seman-
tical consequence w.r.t. K analogously to the case of continuous t-norms (cf. Deni-
tion 1.1.15):
[=
K
iff e() = 1 for each A K and each A-model e of .
DEFINITION 4.1.1 (Completeness properties). Let K be a class of L-chains. We say
that L has the property of:
(Innitary) strong K-completeness, SKC for short, if for every set of formulae
holds:
L
if and only if [=
K
.
Finite strong K-completeness, FSKC for short, if for every nite set of formulae
holds:
L
if and only if [=
K
.
K-completeness, KC for short, if for every formula holds:
L
if and only if
[=
K
.
Now we give three theorems characterizing the just dened properties in a purely
algebraic fashion. First we need one important denition.
DEFINITION 4.1.2 (Partial embeddability). Given two algebras A and B of the same
language /, we say that a nite subset X of Ais partially embeddable into Bif there is a
one-to-one mapping f : X B such that for each c, n / and each a
1
, . . . , a
n
X
satisfying c
A
(a
1
, . . . , a
n
) X, f(c
A
(a
1
, . . . , a
n
)) = c
B
(f(a
1
), . . . , f(a
n
)).
A class K of algebras is partially embeddable into a class K
t
if every nite subset of
every member of K is partially embeddable into a member of K
t
.
THEOREM 4.1.3. Let K be a class of L-chains. Then:
1. L has KC if and only if ' = V(K), where V(K) is the variety generated by K.
2. L has FSKCif and only if ' = Q(K), where Q(K) is the quasi-variety generated
by K.
3. L has SKC if and only if ' = ISP
-f
(K), where ISP
-f
denotes the closure
under isomorphic images, subalgebras, and the operator of reduced products over
countably complete lters.
THEOREM 4.1.4. Let K be a class of L-chains. Then the following conditions are
equivalent:
1. L has SKC.
2. Every countable L-chain is embeddable into a member of K.
Chapter I: Introduction to Mathematical Fuzzy Logic 69
3. Every countable subdirectly irreducible L-chain is embeddable into a member
of K.
THEOREM 4.1.5. Let K be a class of L-chains. Then the following conditions are
equivalent:
1. L has FSKC.
2. Every L-chain is embeddable into an ultraproduct of members of K.
If furthermore the language of L is nite, we can equivalently add:
3. Every L-chain is partially embeddable into K.
4. Every countable subdirectly irreducible L-chain is partially embeddable into K.
The condition 4 is a sufcient condition for the FSKC of L, but is not necessary, as
shown in [35, Example 3.10].
Next we shall give some relationships between nite strong and innitary strong
K-completeness:
THEOREM 4.1.6. Let K be a class of L-chains. Assume that L has FSKC. Then:
1. L has SKC iff [=
K
is nitary.
2. L has SKC if K is closed under ultraproducts.
3. L has SP
U
(K)C, where P
U
denotes the closure under ultraproducts.
We can also prove that a -core fuzzy logic has KC if and only if it has FSKC.
Also for a core fuzzy logic L it can be shown that: L has the SKC (resp. FSKC) with
respect to a class of L-chains K if and only if L
Z
has the SK
Z
C (resp. FSK
Z
C), where
K
Z
is the class of -expansions of chains in K.
Now we move to particular distinguished semantics. Let us x a (-)core logic L.
We dene the following classes of L-chains:
Real-valued L-chains, !: those whose lattice reduct is [0, 1] with the usual order.
Rational-valued L-chains, Q: those whose lattice reduct is [0, 1]
Q
with the usual
order.
Hyperreal-valued L-chains, !

: those whose lattice reduct is any ultrapower of

[0, 1] with the usual order.
Finite-valued L-chains, T: those chains which are nite.
Recall that by conventions of Section 2, standard L-chains are usually a (conven-
tionally distinguished) subclass of real-valued L-chains (often all of them, cf. Conven-
tion 2.0.2). Note also that the class of hyperreal-valued L-chains !

contains the class

of real-valued L-chains ! (just take the ultrapower over a principal ultralter). We
could furthermore consider strict hyperreal-valued L-chains, which arise by removing
70 Libor B ehounek, Petr Cintula, and Petr H ajek
Figure 8. Real-chain, rational-chain, and hyperreal-chain completeness properties.
the real-valued ones from !

; however, it can be shown that all the three complete-

ness properties remain unchanged by this modication. In fact, even the rational-valued
and hyperreal-valued completeness properties coincide: all known relationships of these
properties are depicted in Figure 8.
The completeness w.r.t. nite chains has already been mentioned in Section 2.4.3,
where we have dened the n-valued extension of a given fuzzy logic by means of the
axiom (n). This axiom in fact characterizes the strong nite-chain completeness, as
the following can be proved to be equivalent:
1. L enjoys STC
2. All L-chains are nite
3. There is a natural number n such that the length of each L-chain is less or equal
than n
4. There is a natural number n such that
L
_
i<n
(x
i
x
i+1
).
The weaker forms of completeness w.r.t. nite chains correspond to some well-
known algebraic properties, namely:
L enjoys TC if and only if the variety of L-algebras enjoys FMP (the nite model
property).
L enjoys FSTC if and only if the variety of L-algebras enjoys SFMP (the strong
nite model property). Moreover, if the language of L is nite, these properties
are also equivalent to FEP (the nite embeddability property) for the variety of
L-algebras.
The known completeness results for fuzzy logics introduced in this chapter can be
summarized as follows:
The following logics enjoy S!C (and thus also SQC and S!

C): MTL, IMTL,

SMTL, G, WNM, NM, C
n
MTL, and C
n
IMTL.
The following logics enjoy FS!C (and thus also SQC and S!

C), but not S!C:

WCMTL, MTL, BL, SBL, , and .
Chapter I: Introduction to Mathematical Fuzzy Logic 71
The logic

introduced in [45] is an example of logic enjoying SQCbut not even

!C.

is the axiomatic extension of BL by the axiom

( ) (( & ) ).
The only logics enjoying STC introduced in this chapter are the nite-valued
variants of G
n
and
n
.
The following logics enjoy FSTC, but not STC: MTL, IMTL, SMTL, WNM,
C
n
MTL, C
n
IMTL, BL, SBL, , G, and NM.
The following logics do not enjoy TC: WCMTL, MTL, and .
4.2 Functional representation
It is well known that all Boolean functions (i.e., functions from0, 1
n
to 0, 1) are
expressible by formulae of classical logic. If we replace the set 0, 1 by [0, 1], we are
in the realm of fuzzy logic and a natural question is which functions from[0, 1]
n
to [0, 1]
are expressible by a formula of a particular fuzzy logic. Clearly, due to the cardinality
reasons we cannot expect the analog of the classical results (there are uncountably many
functions and only countably many formulae).
We shall restrict ourselves to (-)core fuzzy logics that enjoy standard complete-
ness w.r.t. a unique standard algebra. For this section let L be one of the following logics:
, G, , RPL, P
t
, their extensions by , and the logics and
1
2
. By [0, 1]
L
we
denote the corresponding standard algebra.
DEFINITION 4.2.1. Let f be a function f : [0, 1]
n
[0, 1] and (v
1
, . . . , v
n
) a for-
mula of a logic L. We say that the function f is represented by the formula in L (or
that is a representation of f) if e() = f(e(v
1
), e(v
2
), . . . , e(v
n
)) for each [0, 1]
L
-
evaluation e.
The functional representation of the logic L is the class of all functions from any
power of [0, 1] into [0, 1] that are representable by some formula of L.
The goal of this subsection is to summarize known characterizations of functional
representation of aforementioned logics.
56
See Chapter IX of this Handbook for detailed
exposition of this area, including more results, proofs, and references.
The rst theorem of this kind is McNaughtons celebrated result:
THEOREM 4.2.2 ( [131, 144]). The functional representation of ukasiewicz logic is
the class of McNaughton functions, i.e., continuous piecewise linear functions with in-
teger coefcients.
Moving to the logic RPL would change littleonly the absolute coefcients of
the linear functions could now be rational. The description for product logic is more
complex. First let us dene:
56
One can observe that functional representation is not (strictly speaking) a property of a logic, but of an
algebra; the standard completeness allow us to perform this slight abuse of language.
72 Libor B ehounek, Petr Cintula, and Petr H ajek
DEFINITION 4.2.3. Let n be a natural number. Then a set S [0, 1]
n
is a region of
positivity if there is M i [ 1 i n such that S = x
1
, . . . , x
n
[0, 1]
n
[
x
i
> 0 iff i M.
THEOREM 4.2.4 ( [36]). The functional representation of product logic is the class of
functions which, restricted to any region of positivity are either identically equal to 0 or
continuous piecewise monomial.
For a (rather complex) characterization of the functional representation of G odel
logic we refer the reader to Chapter IX.
DEFINITION 4.2.5. A subset S of [0, 1]
n
is Q-semialgebraic if it is a Boolean combi-
nation of sets of the form
x
1
, . . . , x
n
[0, 1]
n
[ P(x
1
, . . . , x
n
) > 0
for polynomials P with integer coefcients. If all of the polynomials are linear, then S
is linear Q-semialgebraic.
It can be easily shown that the domain of each linear piece in a McNaughton func-
tion is linear Q-semialgebraic. In Theorem 4.2.2, the latter property was ensured by the
continuity; however, if the non-continuous Baaz delta operation is added, the condition
has to be explicitly stated.
THEOREM 4.2.6. The functional representation of
Z
is the set of piecewise linear
functions with integer coefcients, with the pieces having linear Q-semialgebraic do-
mains.
57
As we have seen, the functions expressible in expansions of ukasiewicz logic
are piecewise functions determined by three parameters: (i) the requirement or non-
requirement of continuity, (ii) the form of functions describing the pieces, and (iii)
the description of domains of the pieces. Table 4 summarizes known results (mainly
from [141]). The column Contin. indicates whether the continuity of the functions
is required or not. The column Domains indicates whether the domains are all Q-
semialgebraic sets, or just the linear ones.
4.3 Proof theory
Proof theory is an important branch of mathematical logic (see [22] for a detailed
and representative introduction). It deals with analysis of proofs (in a given logic) and
the study of calculi designed to satisfy certain convenient properties. Such calculi (e.g.,
the Gentzen-style sequent calculi) are in many respects better behaved (e.g., as regards
the complexity of proofs) than the Hilbert-style calculi employed in this chapter so far.
Unlike general Hilbert-style calculi, which trivially exist for all logics (understood as
57
In more detail, the functions f : [0, 1]
n
[0, 1] for which there is a nite partition of [0, 1]
n
such that
each block of the partition is a linear Q-semi-algebraic set and f restricted to each block is a linear function
with integer coefcients.
58
The well-known PierceBirkhoff conjecture [120] is equivalent to the supposition that the functional
representation of P

is the set of continuous piecewise polynomial functions with integer coefcients.

Chapter I: Introduction to Mathematical Fuzzy Logic 73
Logic Contin. Domains Pieces
yes linear linear functions with integer coefcients

Z
no linear linear functions with integer coefcients
RPL yes linear linear integer coefcients and a rational shift
P
t
yes ? ?
58
P
t
Z
no all polynomials with integer coefcients
no all fractions of polynomials with integer coefcients

1
2
no all fractions of polyn. w. int. coeff., f[0, 1
n
] 0, 1
Table 4. Functional representations of prominent fuzzy logics
consequence relations), the construction of calculi satisfying some required properties
can be a non-trivial task, in many cases even demonstrably impossible.
Proof theory for fuzzy logics (with the exception of G odel logic, whose position as
both fuzzy and intermediate logic makes it a special case) has seen a rapid development
in recent years. The eld has been started by Metcalfe [132] in the early 2000s; a
milestone is the monograph [136]. In this Handbook, the proof theory of fuzzy logics is
treated in detail in Chapter III. Here we present just a basic idea and one motivational
example.
Recall that Gentzen-style calculi have two kinds of rules:
Operational rules for introduction of propositional connectives (into formulae),
and
Structural rules for manipulation with premises and conclusions (as whole for-
mulae).
The rules operate on sequents, which are pairs of sequences of formulae, usually written
as
1
, . . . ,
n

1
, . . . ,
m
. Capital Greek letters (esp. , , , ) will represent
sequences of formulae (including the empty sequence) in sequent schemata. The rule
(R) indicating the possibility of deriving a sequent S fromsequents S
1
, . . . , S
k
is written
in the following form:
(R)
S
1
. . . S
k
S
.
A sequent calculus is a set of (schematic) rules. A derivation in the calculus is a (nite)
tree with nodes evaluated by sequents which are derived by the rules of the calculus from
their immediate predecessors. We say that the sequent S is derivable from S
1
, . . . , S
n
in the calculus if there is a derivation with leaves evaluated by S
1
, . . . , S
n
and the root
by S; if moreover n = 0, we say that S is provable in the calculus.
Recall the Full Lambek logic with exchange and weakening, FL
ew
, and its classi-
cal (i.e., involutive) version CFL
ew
, introduced in Section 3.1. The following example
gives Gentzen-style calculi for these logics.
EXAMPLE 4.3.1. The calculus GCFL
ew
has the following operational rules for the
connectives , &, , , , 0 and 1:
74 Libor B ehounek, Petr Cintula, and Petr H ajek
(0 L)
0,
(1 R)
, 1
C,
(L)
,
for C ,
, ,
(R)
,
, ,
(&L)
& ,

1

1
,
2

2
,
(&R)

1
,
2

1
,
2
, &
, ,
(L)
,
, C
(R)
,
for C ,
,
1

1
,
2

2
(L)
,
1
,
2

1
,
2
, ,
(R)
,
, ,
(L)
, , ,
, ,
(R)
,
Moreover, the calculus GCFL
ew
has the following structural rules:
(Ax)

, ,
(Cut)
, ,
, , ,
(E-L)
, , ,
, , ,
(E-R)
, , ,

(W-L)
,

(W-R)
,
The abbreviated labels of the structural rules stand for Axiom, Exchange, and Weakening,
(left or right). Note that the structural rules of Contraction
, ,
(C-L)
,
, ,
(C-R)
,
are missing as CFL
ew
is a substructural logic. By removing further structural rules of
Exchange and Weakening we would obtain calculi for the the weaker logics CFL
(e)(w)
.
The calculus GFL
ew
has the same rules as GCFL
ew
, with the restriction that the
length of the right-hand sequence of any sequent in a derivation is at most 1. (Conse-
quently of the structural rules, only (E-L), (W-L), and (C-L), and the rule (W-R) for
sequents of the form occur in GFL
ew
.)
GCFL
ew
is a calculus for the logic CFL
ew
in the following sense: the sequent

1
, . . . ,
n

1
, . . . ,
m
is provable in GCFL
ew
iff its interpretation
1
&. . .&
n

1
. . .
m
is a theorem of CFL
ew
. Analogously, GFL
ew
is a calculus for FL
ew
.
59
59
For m = 0, i.e., the empty right-hand side of the sequent, we use the interpretation
1
&. . . &
n
0;
analogously for n = 0 we use the interpretation 1
1

m
. Note that in the case of MTL each
sequent has at most one formula on the right-hand side, thus the connective is not needed in its interpretation.
Chapter I: Introduction to Mathematical Fuzzy Logic 75
One of the central topics in proof-theory is the question of redundance of the (Cut)
rule in a given Gentzen-style calculus. Note that all remaining rules enjoy the so-called
subformula property: i.e., the formulae in premises of rules are subformulae of those in
the conclusion (thus, roughly speaking, any backward step in the proof-tree reduces
the complexity of formulae occurring in the labeling sequents in a very transparent way,
which often provides a decision procedure for a given logic). Cut-free calculi enjoying
subformula property are often called analytic.
THEOREM 4.3.2 (Cut elimination). The calculi GFL
ew
and GCFL
ew
enjoy cut elimi-
nation; i.e., removing the rule (Cut) does not change the set of provable sequents.
A natural question is whether we can strengthen the two Gentzen-style calculi
(preferably preserving the cut elimination property) to obtain proof systems for the semi-
linear versions of the logics FL
ew
, and CFL
ew
, i.e., the logics MTL and IMTL. In
Hilbert-style calculi, the transition is provided by adding the prelinearity axiom (
) ( ). However, for Gentzen-style calculi with the cut elimination property
(and the interpretation of sequents as above), this question was answered negatively by
Ciabattoni, Galatos, and Terui:
THEOREM 4.3.3 ( [28]). If the sequents
1
, . . . ,
n

1
, . . . ,
m
are interpreted as

1
&. . . &
n

1
. . .
m
, then there is no Gentzen system for IMTL extending
GCFL
ew
with structural rules (analogously for MTL).
One way to overcome this problem is to change the syntactical framework from
sequents to the so-called hypersequents. Hypersequents were originally introduced by
Avron as a proof-theoretic framework for the Relevance-Mingle logic RM [3]. Later
the hypersequent calculi were generalized to G odel logic [4] and other fuzzy logics [66,
133136]. As an example of a hypersequent calculi for fuzzy logics, we present here a
hypersequent Gentzen-style calculus for the logic IMTL (originally introduced in [6]).
A hypersequent is a nite multiset of sequents written as:

1

1
[
2

2
[ [
n

n
A hypersequent version of the sequent rule (R) has the form
(H-R)
1 [ S
1
. . . 1 [ S
k
1 [ S
,
where S
1
, . . . , S
k
/S is an instance of (R), and 1 is a hypersequent variable.
DEFINITION 4.3.4. The hypersequent calculus GIMTL consists of the hypersequent
versions of all rules of GCFL
ew
, plus the following rules:
(EW)
1
1 [ (
(EC)
1 [ ( [ (
1 [ (
(Com)
1 [
1
,
1

1
,
1
1 [
2
,
2

2
,
2
1 [
1
,
2

1
,
2
[
1
,
2

1
,
2
The abbreviated labels of these rules stand for External Weakening, External Contrac-
tion, and Communication.
76 Libor B ehounek, Petr Cintula, and Petr H ajek
THEOREM 4.3.5. For a sequent S =
1
, . . . ,
n

1
, . . . ,
m
we set
I(S) =
1
& . . . &
n

1
. . .
m
;
for a hypersequent 1 = S
1
[ [ S
n
we set
I(1) = I(S
1
) I(S
2
) . . . I(S
n
).
Then the hypersequent 1 is derivable in GIMTL if and only if I(1) is a theorem of
IMTL. Furthermore, the (Cut) rule is eliminable from GIMTL.
4.4 Computational complexity
In this section we deal with computational complexity issues in (-)core fuzzy
logics. Unlike in classical logic, the sets of tautologies and satisable formulae do not
in fuzzy logics determine each other. This is caused by the fact that e() ,= 1 does not
imply e() = 1, but only e() > 0. Consequently, in addition to the sets of tautologies
and 1-satisable formulae, also the sets of positive tautologies and positively satisable
set of formulae are studied. For simplicity, in this section we consider only the rst four
notions and only with respect to the standard semantics. Furthermore we x a (-)core
logic enjoying the nite strong standard completeness (thus in particular we have that
its standard tautologies coincide with its theorems). Chapter X studies these notions
relativized to other important classes of algebras and for logics other than (-)core, and
also tackles other complexity problems encountered in mathematical fuzzy logic (e.g.,
the complexity of the provability relation of a propositional fuzzy logic, or even of the
universal fragment of the rst-order theory of particular classes of algebras).
DEFINITION 4.4.1. Let L be a (-)core fuzzy logic enjoying the nite strong standard
completeness. Then we dene the following sets of formulae:
SAT
pos
(L) if there is a standard L-algebra A and an A-evaluation e such
that e() > 0.
SAT(L) if there is a standard L-algebra Aand an A-evaluation e such that
e() = 1.
TAUT
pos
(L) if for each standard L-algebra A and each A-evaluation e
holds e() > 0.
TAUT(L) if for each standard L-algebra A and each A-evaluation e holds
e() = 1.
Although these sets of formulae are not as tightly related as their classical analogs,
some interrelations can still be proved (especially for stronger fuzzy logics):
LEMMA 4.4.2. In general we can prove:
TAUT
pos
(L) iff , SAT(L)
SAT
pos
(L) iff , TAUT(L)
Chapter I: Introduction to Mathematical Fuzzy Logic 77
Logic SAT(L), SAT
pos
(L) TAUT(L), TAUT
pos
(L) reference
NP-c coNP-c [143]
BL NP-c coNP-c [8]
G NP-c coNP-c [7]
NP-c coNP-c [7]
SMTL, MTL NP-c decidable, coNP-hard [109] [114]
MTL, IMTL NP-hard decidable, coNP-hard [19] [114]
,
1
2
PSPACE PSPACE [103]
Table 5. Standard computational complexity of prominent fuzzy logics
If the logic L expands IMTL, we can also prove the converse:
60
SAT(L) iff , TAUT
pos
(L)
TAUT(L) iff , SAT
pos
(L)
If the logic L extends SMTL, we can also prove:
SAT
pos
(L) iff SAT(L) iff is classically satisable
TAUT
pos
(L) iff TAUT(L) iff is classical tautology
TAUT
pos
(L) iff / SAT
pos
(L) iff is classical tautology
Finally, if the logic L is -core, then we can also prove:
SAT(L) iff SAT
pos
(L)
TAUT(L) iff TAUT
pos
(L)
Despite the more complex denitions involved, the computational complexity of
these problems for prominent fuzzy logics does not differ much from their classical
counterparts: see Table 5 for a selection of known and unknown results on standard
computational complexity.
5 Predicate fuzzy logics
In this section we survey basic facts about predicate fuzzy logics. We restrict our-
selves to (-)core fuzzy logics (see Section 3.2), even though most of the denitions
and theorems can be formulated and proven in weaker logicse.g., non-commutative,
non-integral, with restricted language, etc.as well: for more general formulations and
results see Chapter II. The text of this subsection is loosely based on the survey pa-
per [37]; basic information with full proofs for the logics of continuous t-norms is found
in the monograph [83].
5.1 Syntax
In the following let L be a xed (-)core fuzzy logic in a propositional language /.
The language of the rst-order fuzzy logic is dened in the same way as in classical
rst-order logic. In order to x notation and terminology we give an explicit denition:
60
It can also be proved in expansions of MTL

, for instead of .
78 Libor B ehounek, Petr Cintula, and Petr H ajek
DEFINITION 5.1.1. A predicate language T is a triple Pred
1
, Func
1
, Ar
1
, where
Pred
1
is a non-empty set of predicate symbols, Func
1
is a set (disjoint with Pred
1
)
of function symbols, and Ar
1
is the arity function, assigning to each predicate or func-
tion symbol a natural number called the arity of the symbol. The function symbols F
with Ar
1
(F) = 0 are called object constants. The predicates symbols P for which
Ar
1
(P) = 0 are called truth constants.
61
T-terms and (atomic) T-formulae of a given predicate language are dened as in
classical logic (note that the notion of formula also depends on propositional connectives
in /). A T-theory is a set of T-formulae. The notions of free occurrence of a variable,
substitutability, open formula, and closed formula (or, synonymously, sentence) are de-
ned in the same way as in classical logic. Unlike in classical logic, in fuzzy logics
without involutive negation the quantiers and are not mutually denable, so the
primitive language of L has to contain both of them.
There are several variants of a rst-order extension of a propositional fuzzy logic L
that can be dened. Here we shall introduce the rst-order logics L
m
and L (of models
over general resp. linear algebras); later we shall extend the family by considering the
logics L
w
and L
s
of, respectively, witnessed and standard models. The axiomatic
systems of the logics L
m
and L are dened as follows:
DEFINITION 5.1.2. Let L be a (-)core fuzzy logic and T a rst-order language. The
logic L
m
has the following axioms:
(P) Instances of the axioms of L (with T-formulae substituted for
propositional variables)
(1) (x)(x) (t), where the T-term t is substitutable for x in
(1) (t) (x)(x), where the T-term t is substitutable for x in
(2) (x)( ) ( (x)), where x is not free in
(2) (x)( ) ((x) ), where x is not free in
The deduction rules of L
m
are those of L and the rule of generalization:
(Gen) From infer (x).
The logic L is the extension of L
m
by the axiom:
(3) (x)( ) (x), where x is not free in
The notions of proof and provability are in fuzzy logics dened in the same way as
in classical logic. The fact that the formula is provable in L
m
from a theory T will
be denoted by T
L
m , and analogously for L; in a xed context we can write just
T .
61
The r oles of nullary predicates of 1 and nullary connectives of / are analogous, even though the values
of the former are only xed under a given interpretation of the predicate language, while the values of the
latter are xed under all such interpretations. The ambiguity of the term truth constant (see Denition 3.2.1
and Section 2.2.3) is thus a harmless abuse of language.
Chapter I: Introduction to Mathematical Fuzzy Logic 79
Logic -den. (3)-elim.
G, , WNM (and weaker logics) No No
IMTL, NM Yes No
G

, SBL

Yes ?
(and stronger logics) Yes Yes
Table 6. -denability and (3)-eliminability in prominent rst-order fuzzy logics L
A general theory of rst-order non-classical logics was rst given by Helena Ra-
siowa in [158]. Her rst-order extension of a given Rasiowa-implicative logic (see Def-
inition 3.3.1) was axiomatized analogously to L
m
, only the axioms (2) and (2) were
replaced by the corresponding rules. It can be shown that in the context of ()-core
fuzzy logic these two axiomatizations coincide (for the proof of an even more general
formulation of this claimsee Chapter II). The superscript m stands for minimal, since
L
m
is, in a sense, the weakest rst-order extension of L: as will be seen in Section 5.2,
L
m
is sound and complete w.r.t. rst-order models built over arbitrary L-algebras.
However, the axioms of L
m
are not strong enough to ensure the completeness
w.r.t. rst-order models over linear L-algebrasi.e., the linear completeness theorem,
common to all fuzzy logics. This is why it is needed to add the axiom (3), which is
valid in all models over linear L-algebras (though not generally in models over arbitrary
L-algebras) and ensures the linear completeness theorem for the resulting logic L.
62
This makes L the natural rst-order extension of a given ()-core fuzzy logic L.
Consequently, this rst-order logic is denoted as L with no superscript, though its
more systematic denotation would be L

. Finally let us note that in the context of

mathematical fuzzy logic, the logics L
m
were rediscovered by P. H ajek in [84], denoted
there by L

.
The following two notions distinguish logics for which the axiomatic systems of
L
m
and L can be simplied.
DEFINITION 5.1.3. Let L be a (-)core fuzzy logic. We say that:
The logic L has (3)-eliminability if axiom(3) is redundant, i.e., if L
m
= L.
The logic L has -denability if there is a (denable) unary connective in the
language of L such that
L
(x) (x).
If a unary connective such that
L
and
L

(i.e., is an involutive negation) is denable in the language of a (-)core logic L,
then L has -denability. If L has -denability, then the axioms (1) and (2)
are redundant. Table 6 contains known facts on -denability and (3)-eliminability in
particular t-norm based logics.
Let us list some important theorems that are provable in all logics L
m
. For their
proofs in MTL or BL see [53, 83]; their proofs in a weaker setting can be found in
Chapter II.
62
The fact was rst observed for G odel logic by Horn in [115].
80 Libor B ehounek, Petr Cintula, and Petr H ajek
THEOREM 5.1.4. Let L be a (-)core fuzzy logic and T a predicate language. Let ,
, be T-formulae, x a variable not free in , and x
t
a variable not occurring in .
The following T-formulae are then theorems of L
m
:
(T1) (x)
(T2) (x)
(T3) (x)(x) (x
t
)(x
t
)
(T4) (x)(x) (x
t
)(x
t
)
(T5) (x)(y) (y)(x)
(T6) (x)(y) (y)(x)
(T7) (x)( ) ((x) (x))
(T8) (x)( ) ((x) (x))
(T9) ( (x)) (x)( )
(T10) ((x) ) (x)( )
(T11) (x)( ) ( (x))
(T12) (x)( ) ((x) )
(T13) (x) (x) (x)( )
(T14) (x)( ) (x) (x)
(T15) (x) (x)( )
(T16) (x)( ) (x)
(T17) (x)( & ) (x) &
(T18) (x)(
n
) ((x))
n
(T19) (x) (x)
(T20) (x) (x)
The implication converse to (T11) is provable in
m
and , but not in G. The
implications converse to (T12) and (T19) are provable in
m
, but neither in G
nor .
Finally, the formula (x) (x)( ) is a theorem of L for any L.
Some syntactic metatheorems known from propositional fuzzy logics hold analo-
gously for rst-order fuzzy logics:
63
THEOREM 5.1.5. Let T be a predicate language.
1. Let L be a -core fuzzy logic, a T-formula and
t
a T-formula obtained from
by replacing some occurrences by . Then

t
.
(The congruence property for L
m
and L)
63
For details see Chapter II or [98].
Chapter I: Introduction to Mathematical Fuzzy Logic 81
2. Let T be a theory, (x) a T-formula, and c a constant not occurring in T .
Then T (c) iff T (x). (The constants theorem for L
m
and L)
3. Let L be a core fuzzy logic. Then for each theory T and sentences , , the
following holds: T, iff there is natural n such that T
n
. (The local
deduction theorem for L
m
and L)
4. Let L be a -core fuzzy logic. Then for each T-theory T and T-sentences , ,
the following holds: T, iff T . (The -deduction theorem for
L
m
and L)
5. Consequently, let L be (-)core fuzzy logic, T a T-theory, and , , T-senten-
ces. Then the following metatheorems hold for L:
(a) If T, and T, , then T, . (The proof by cases property)
(b) If T, and T, , then T . (The semilinearity
property)
The following theorem demonstrates that in (-)core fuzzy logics one can conser-
vatively introduce Skolem functions in a similar manner as in classical logic. (The def-
inition of conservative extension in rst-order fuzzy logics is analogous to the classical
denition.)
THEOREM 5.1.6. Let L be a core fuzzy logic, T a predicate language, (y, x
1
, . . . , x
n
)
a T-formula, T a T-theory such that T
L
(y)(y, x
1
, . . . , x
n
), and F

a function
symbol of arity n not present in T. Then the (T F

)-theory
T
t
= T (F

(x
1
, . . . , x
n
), x
1
, . . . , x
n
)
is a conservative extension of T over L.
The analogous theorem holds, with one small modication, for -core fuzzy logics.
THEOREM 5.1.7. Let L be a -core fuzzy logic, T a predicate language, (y, x
1
, . . . ,
x
n
) a T-formula, T a T-theory such that T
L
(y)(y, x
1
, . . . , x
n
), and F

a
function symbol of arity n not present in T. Then the (T F

)-theory
T
t
= T (F

(x
1
, . . . , x
n
), x
1
, . . . , x
n
)
is a conservative extension of T over L.
5.2 Semantics
In this subsection, we shall introduce the general and linear semantics of predi-
cate fuzzy logics, corresponding to the axiomatic systems L
m
and L, as well as the
logics L
s
and L
w
, corresponding to the (more restrictive) semantics of standard and
witnessed predicate models.
82 Libor B ehounek, Petr Cintula, and Petr H ajek
5.2.1 Basic denitions and completeness theorems
Now we shall look at the semantics of rst-order fuzzy logics. To simplify the for-
mulation of upcoming denitions let us x: a (-)core fuzzy logic L in a propositional
language /, a predicate language T = (Pred, Func, Ar), and an L-algebra B.
DEFINITION 5.2.1. A B-structure M for the predicate language T has the form:
M = M, (P
M
)
PPred
, (F
M
)
FFunc
, where M is a non-empty domain; for each n-
ary predicate symbol P Pred, P
M
is an n-ary fuzzy relation on M, i.e., a function
M
n
B (identied with an element of B if n = 0); for each n-ary function symbol
F Func, F
M
is a function M
n
M (identied with an element of M if n = 0).
Let Mbe a B-structure for T. An M-evaluation of the object variables is a mapping
v which assigns an element from M to each object variable. Let v be an M-evaluation,
x a variable, and a M. Then by v[xa] we denote the M-evaluation such that
v[xa](x) = a and v[xa](y) = v(y) for each object variable y different from x.
Let Mbe a B-structure for T and v an M-evaluation. We dene the values of terms
and the truth values of formulae in Mfor an evaluation v recursively as follows:
|x|
B
M,v
= v(x)
|F(t
1
, . . . , t
n
)|
B
M,v
= F
M
(|t
1
|
B
M,v
, . . . , |t
n
|
B
M,v
) for F Func
|P(t
1
, . . . , t
n
)|
B
M,v
= P
M
(|t
1
|
B
M,v
, . . . , |t
n
|
B
M,v
) for P Pred
|c(
1
, . . . ,
n
)|
B
M,v
= c
B
(|
1
|
B
M,v
, . . . , |
n
|
B
M,v
) for c /
|(x)|
B
M,v
= inf||
B
M,v[xa]
[ a M
|(x)|
B
M,v
= sup||
B
M,v[xa]
[ a M
If the inmum or supremum does not exist, we take the truth value of the quantied
formula as undened. We say that the B-structure M is safe if ||
B
M,v
is dened for
each T-formula and each M-evaluation v.
We shall write:
|(a
1
, . . . , a
n
)|
B,M)
instead of |(x
1
, . . . , x
n
)|
B
M,v
if v(x
i
) = a
i
for all i
n
B, M [= [v] if ||
B
M,v
= 1
B, M [= if B, M [= [v] for each M-evaluation v. When B is known
from the context, we write M [= only.
B [= if B, M [= for each safe B-structure M (we also say that is a
B-tautology).
Let M be a safe B-structure for T and T a T-theory. Then M is called a B-model
of T if B, M [= for each T.
Observe that models are safe structures (by denition). As obviously each safe B-
structure is a B-model of the empty theory, we shall use the term model for both models
Chapter I: Introduction to Mathematical Fuzzy Logic 83
and safe structures in the rest of the text. By a slight abuse of language we use the term
model also for the pair B, M. We say that B, M is an -model (of T) whenever Bis
linearly ordered. Thus by the phrase for each (-)model B, M (of T) we mean for
each (linear) L-algebra B and each safe B-model M(of T). The syntax and semantics
of rst-order fuzzy logics are bound together by the following completeness theorems:
64
THEOREM 5.2.2 (General strong completeness for L
m
). Let L be a (-)core fuzzy
logic, T a predicate language, T a T-theory, and a T-formula. Then the following
are equivalent:
T
L
m
B, M [= for each model B, M of the theory T.
THEOREM 5.2.3 (Linear strong completeness for L). Let L be a (-)core fuzzy logic,
T a predicate language, T a T-theory, and a T-formula. Then the following are
equivalent:
T
L

B, M [= for each -model B, M of the theory T.
5.2.2 Standard semantics
Besides the general and linear semantics, also the intended (or standard) [0, 1]-
semantics can be considered for rst-order fuzzy logics.
DEFINITION 5.2.4. We shall say that B, M is an s-model (of T) if B, M is a
model (of T) and B is a standard L-algebra.
65
Let L be a (-)core fuzzy logic. If the equivalence
T iff B, M [= for each s-model B, M of the theory T
holds for:
T = , we say that L enjoys (weak) standard completeness
All nite theories T, we say that L enjoys nite strong standard completeness
All theories T, we say that L enjoys innitary strong standard completeness.
The next theoremsummarizes known results on the standard completeness of promi-
nent (-)core fuzzy logics. Their proofs are scattered in the literature; see [53, 83] for
the proofs for the main core fuzzy logics, and the survey paper [35] for more information
and detailed references.
64
For the proofs of Theorems 5.2.2 and 5.2.3 see [98]. Instances of Theorem 5.2.3 for various (.-)core
fuzzy logics were originally proved separately (usually for countable predicate languages only): the proofs for
prominent core fuzzy logics can be found in [53, 83].
65
For the notion of standard L-algebra see Denition 1.1.19, Convention 2.0.2, and some exceptions to this
convention in Section 2.
84 Libor B ehounek, Petr Cintula, and Petr H ajek
THEOREM 5.2.5. The logics L are innitary-strong standard complete if L is any
of the following logics: MTL, SMTL, IMTL, WNM, NM, G (all of them with or
without ), and G

.
The logics L are not even weakly standard complete if L is any of the following
logics: BL, SBL, , , MTL, WCMTL
66
(all of them with or without ), SBL

,
and .
If the logic L does not enjoy standard completeness, the standard rst-order logic
L
s
of s-models over L can be introduced. It can be shown, for instance, that the stan-
dard ukasiewicz rst-order logic
s
is not nitary, but can be axiomatized by means
of an innitary rule (see [106]).
Let us mention Pavelka logic RPL in this context.
67
Clearly RPL is a core fuzzy
logic, so its predicate version RPL is strongly complete w.r.t. all safe interpretations
over all RPL-chains. Furthermore, the logic RPL extends ukasiewicz predicate
fuzzy logic conservatively [102]. Now let us turn to its standard semantics: con-
sider only models over the standard RPL-algebra (i.e., s-models). Then the denitions
of provability degree and truth degree can be introduced in the same manner as in the
propositional case: let T be a theory and a formula; the provability degree [[
T
is the
supremum of all r Q such that T proves r (i.e., that is at least r-true); and the
truth degree ||
T
is the inmum of the truth degrees of in all (standard) models of T.
Then the Pavelka-style completeness holds true: let T be a theory over RPL and a
formula; then [[
T
= ||
T
. Notice, however, that the Pavelka-style completeness of
the logic RPL does not entail its standard completeness: we only know that a standard
tautology has the provability degree 1, i.e., that for each r < 1 we can prove r .
A usual method to prove the failure of standard completeness of a given logic is to
show that the set of its standard tautologies is not recursively enumerable, and therefore
it cannot coincide with the set of its theorems. Determining the position in the arith-
metical hierarchy (see e.g. [161]) of prominent sets of formulae (such as the tautologies
of a given logic) is an important eld of study in mathematical logic. Here we just
briey mention a few results related to fuzzy logics: for a full treatment of the arith-
metical complexity of fuzzy logics see Chapter XI; some important papers on the topic
are [85, 89, 90, 138, 139].
First let us introduce some prominent sets of formulae given by a fuzzy logic L:
DEFINITION 5.2.6. Let be a sentence of L. We say that is
A general (resp., standard) tautology of L if ||
B,M)
= 1 for each - (resp.,
s-) model B, M
Generally (resp., standardly) satisable in L if ||
B,M)
= 1 for some - (resp.,
s-) model B, M.
The sets of general and standard tautologies and generally and standardly satisable
sentences will be denoted, respectively, by genTAUT, stTAUT, genSAT, and stSAT.
66
There is a general result in [140] that any logic between WCMTL and is not standard complete.
67
See Section 2.2.3 and [83, 154].
Chapter I: Introduction to Mathematical Fuzzy Logic 85
G BL
stTAUT
1
-c
2
-c N-A N-A
stSAT
1
-c
1
-c N-A N-A
Table 7. Arithmetical complexity of standard semantics
For illustration, let us state the results for four predicate logics: BL, , G,
and . For each of them, the set of general tautologies is
1
-complete (thus they are
recursively axiomatizable, but undecidable) and the set of generally satisable formulae
is
1
-complete. For the arithmetical complexity of their standard semantics see Table 7
(where -c stands for -complete and N-A for non-arithmetical). It can be seen
that as far as standard semantics is concerned, the four logics differ drastically with
respect to the degree of undecidability.
5.2.3 Witnessed models
Recall from Denition 5.2.1 that the truth degree of an existentially quantied for-
mula is dened as the supremum of the truth degrees of its instances. The supremum
may, though, be larger than the truth value of any instance. If, nevertheless, the supre-
mum of the truth values is achieved by some instance (i.e., is the maximum of the truth
degrees of instances), we call the quantied formula witnessed (and similarly for uni-
versally quantied formulae):
DEFINITION 5.2.7. A formula (x) with free variables y
1
, . . . , y
n
is witnessed in
B, M if for each evaluation a
1
, . . . , a
n
M of y
1
, . . . , y
n
there is an element b M
such that
|(x)(x, a
1
, . . . , a
n
)|
B,M)
= |(b, a
1
, . . . , a
n
)|
B,M)
,
and similarly for (x). We call a model B, M witnessed if each formula beginning
with a quantier is witnessed in B, M.
Let L be a core fuzzy logic. We dene the logic L
w
as the extension of L by the
axiom schemata:
68
(C) (x)((x) (y)(y))
(C) (x)((y)(y) (x))
THEOREM 5.2.8 (Witnessed completeness). Let L be a core predicate fuzzy logic, T
a predicate language, T a theory, and a formula. Then T
L
w iff B, M [=
for each witnessed -model B, M of the theory T.
Note that the only rst-order logic L

of a continuous t-norm that proves (C)

and (C) is ukasiewicz logic, which is thus strongly complete with respect to witnessed
models. For product logic, we have the following:
68
References for witnessed semantics of fuzzy logic are [9496, 98].
86 Libor B ehounek, Petr Cintula, and Petr H ajek
DEFINITION 5.2.9. Call a formula (x) with free variables y
1
, . . . , y
n
weakly wit-
nessed in B, M if for each evaluation a
1
, . . . , a
n
M of y
1
, . . . , y
n
such that
|(x)(x, a
1
, . . . , a
n
)|
B,M)
,= 0
there is an element b M such that
|(x)(x, a
1
, . . . , a
n
)|
B,M)
= |(b, a
1
, . . . , a
n
)|
B,M)
.
Call a model B, M weakly witnessed if each formula beginning with an existential
quantier is witnessed and each formula beginning with a universal quantier is weakly
witnessed in B, M.
The analog of Theorem 5.2.8 can be proved for with respect to weakly witnessed
models [127].
5.3 The identity and sorts of objects
Further possible extensions of rst-order fuzzy logics regard the identity (or equal-
ity) and sorts of objects.
5.3.1 The identity predicate
The identity (or equality) of objects can be included in the logical vocabulary of any
rst-order fuzzy logic as a newprimitive binary predicate =, with the intended semantics
of the crisp identity on the domain of a rst-order model M; i.e., |x = y|
B
M,v
= 1 iff
v(x) = v(y) and |x = y|
B
M,v
= 0 iff v(x) ,= v(y). This intended behavior of =,
however, clearly cannot be ensured by any set of rst-order axioms.
69
In any extension
of rst-order MTL, the following axioms approximate the intended behavior of = by
allowing only those non-identical objects that are logically indistinguishable in the given
predicate language (i.e., that share all properties expressible in the predicate language
in all models to the same degrees) to be fully identical:
(Re
=
) x = x
(LP) x = y ((x) (y))
(Crisp
=
) x = y (x = y),
for any (which can contain other free variables besides x) and y substitutable for x
in (x). The axiom (Re
=
) expresses the fact that any object is identical with itself,
and the axiom schema (LP) expresses one direction of the Leibniz principle of identity,
namely the fact that identical objects have the same properties. The axiom (Crisp
=
)
of the crispness of = is redundant in the extensions of MTL
Z

m
or
m
, where it is
provable from the schema (LP). For any (-)core fuzzy logic L the rst-order fuzzy
logic L extended by the axioms (Re
=
), (LP), and (Crisp
=
) will be called L with
identity or L
=
(and similarly for the logics L
m
=
, L
s
=
, and L
w
=
).
69
Like in classical rst-order logic, this is because for each model Min which = is realized as the crisp
identity of objects we can easily construct a model M

satisfying the same formulae, in which = does not

coincide with the identity of objects. (Just replicate any element a of Mand dene all predicates and functors
identically on the new copies as they were on a.)
Chapter I: Introduction to Mathematical Fuzzy Logic 87
Even though in L

=
, the predicate = need not be interpreted as the crisp identity,
it is always a crisp equivalence relation, as L
m
=
proves its symmetry and transitivity.
Since in any model Mof L

=
all formulae are by (LP) congruent with =, factorizing
Mby this crisp equivalence relation yields a model M
t
that satises the same formulae
as Mto the same degrees and in which = has the intended semantics (i.e., is realized as
the crisp identity of objects).
The identity predicate enables, i.a., elimination of function symbols (including the
Skolem functions, see Theorems 5.1.6 and 5.1.7) and (crisp) sorts of variables (intro-
duced later in this section) in (-)core fuzzy logics in the same way as in classical
logic. For fuzzy logics with identity see [37].
5.3.2 Sorts of objects
To enable a smooth work in axiomatic theories over rst-order fuzzy logics, similar
to the practice of classical mathematics, it is expedient to introduce sorts of objects in
the predicate language. We shall only consider crisp sorts of objects, as quantication
restricted to fuzzy domains is rather problematic in contraction-free fuzzy logics. On
the other hand, we shall allow for subsumption of sorts, formalizing the situation that all
objects of one sort are also objects of another sort (e.g., the fact that all natural numbers
are also rational numbers, which in turn are also real numbers). This kind of sorts for
predicate fuzzy logics was introduced in [12].
Let us x some weakly implicative fuzzy logic L. We shall present a multi-sorted
variant of the logic L, but mutatis mutandis the denitions apply to L
m
, L
s
, L
w
, or
any combination thereof. We shall only give the differences from the single-sorted case,
not repeating the denitions that are the same for multi- and single-sorted L.
DEFINITION 5.3.1. A multi-sorted predicate language is a quintuple
T = Sort, _, Pred, Func, Ar,
where Sort is a non-empty set of sorts, _ is an ordering on Sort (indicating the sub-
sumption of sorts), Pred is a non-empty set of predicate symbols, Func is a set (dis-
joint with Pred) of function symbols, and Ar is the arity function that assigns to each
predicate symbol a nite sequence of elements of Sort and to each function symbol a
non-empty nite sequence of elements of Sort.
The length [Ar(P)[ of the sequence Ar(P) is called the arity of the predicate sym-
bol P and the number [Ar(F)[ 1 is called the arity of the function symbol F. If
Ar(F) = s, then the function F is called an individual constant of sort s. If s
1
_ s
2
,
we say that the sort s
2
subsumes the sort s
1
.
For each s Sort there is an innite countable set of individual variables of sort s,
denoted by x
s
, y
s
, . . . The notion of T-term is dened recursively as follows:
All individual variables of each sort s Sort are T-terms of sort s.
If t
1
, . . . , t
n
are T-terms of the respective sorts s
1
, . . . , s
n
Sort and F is a
function symbol such that Ar(F) = s
t
1
, . . . , s
t
n
, s
n+1
, where s
i
_ s
t
i
for all
i n, then F(t
1
, . . . , t
n
) is a T-term of sort s
n+1
.
88 Libor B ehounek, Petr Cintula, and Petr H ajek
Only the expressions that arise by a nite iteration of the previous two rules are
T-terms.
If t
1
, . . . , t
n
are T-terms of respective sorts s
1
, . . . , s
n
and P a predicate symbol
such that Ar(P) = s
t
1
, . . . , s
t
n
, where s
i
_ s
t
i
for all i n, then P(t
1
, . . . , t
n
) is an
atomic T-formula.
The notions of T-formula, T-theory, bound and free variable, and T-sentence are
dened analogously to the single-sorted case. A term t of sort s is substitutable for the
variable x
s
in a formula (x
s
, . . . ) if s _ s and no variable occurring in t is bounded
in (t, . . . ).
In multi-sorted rst-order logics with the identity predicate we furthermore assume
that for each sort s Sort there is a predicate symbol =
s
Pred with Ar(=
s
) =
s, s. (The predicate =
s
is usually denoted just by =, as s is determined by the sort of
its arguments.)
Let us x a multi-sorted predicate language T = Sort, _, Pred, Func, Ar. The
semantics of multi-sorted rst-order fuzzy logic for this language is dened as follows:
DEFINITION 5.3.2. A B-structure Mfor the predicate language T has the form M =
(M
s
)
sSort
, (P
M
)
PPred
, (F
M
)
FFunc
, where:
For each s Sort, M
s
is a non-empty set and M
s
M
s
iff s _ s
t
;
For each predicate symbol P Pred such that Ar(P) = s
1
, . . . , s
n
, P
M
is a
function
_
n
i=1
M
s
i
_
B (identied with an element of B if n = 0); and
For each function symbol F Func such that Ar(F) = s
1
, . . . , s
n
, s
n+1
, F
M
is a function
_
n
i=1
M
s
i
_
M
s
n+1
(identied with an element of M
s
1
if n = 0).
In multi-sorted rst-order logics with identity it is furthermore assumed that if
s _ s
t
, then the realizations of =
s
and =
s
coincide on the domain of s.
An M-evaluation of the object variables in a B-structure M is a mapping that
assigns to each variable of sort s an element from M
s
. The denitions of the value of a
term, the truth value of a formula, a safe structure, a B-tautology, and a B-model of a
theory are the same as in the single-sorted case.
Notice that the domains of the sorts of variables are assumed to be crisp sets. This
requirement is analogous to single-sorted models, whose domains are also assumed to
be crisp.
70
Due to this analogy, most metamathematical properties of single-sorted rst-
order fuzzy logics directly translate for multi-sorted languages as well. In particular, the
axioms, and consequently the provable formulae, of rst-order fuzzy logics for multi-
sorted languages are the same as in single-sorted rst-order fuzzy logics (see Deni-
tion 5.1.2), only with the appropriate restrictions on the sorts of terms and variables.
Also many metatheorems on rst-order fuzzy logics (such as the completeness Theo-
rems 5.2.2 and 5.2.3) can be straightforwardly generalized for multi-sorted languages.
70
One reason for this design choice is the fact that quantication over a fuzzy domain has rather problematic
properties (unless the membership degrees of elements in the domain are idempotent w.r.t. conjunction).
Chapter I: Introduction to Mathematical Fuzzy Logic 89
Sorts of variables can also be conservatively introduced (by crisp formulae) and elimi-
nated (in logics with =) in rst-order fuzzy logics in the manner fully similar to that in
classical rst-order logic (cf. [97]).
5.4 Axiomatic theories over fuzzy logic
First-order fuzzy logics are strong enough to support non-trivial axiomatic mathe-
matical theories. We shall briey survey the development of some kinds of arithmetic
and set theory over fuzzy logic. The reader is assumed to know basic facts about classi-
cal Peano arithmetic PA and ZermeloFraenkel set theory ZF.
5.4.1 Arithmetic over fuzzy logics
Clearly, over classical logic one cannot extend PA by a truth predicate Tr and the
dequotation axiom schema Tr(), where is a name (the G odel number) of the
sentence . This is due to the Russell paradox: using G odel diagonalization method we
could construct a formula such that Tr(), hence satisfying Tr() Tr(),
which is contradictory over classical logic: a sentence cannot be equivalent to its own
negation. But it need not be contradictory over ukasiewicz logic: the sentence may
have the truth value one half.
An arithmetical theory with a truth predicate and the dequotation schema has in-
deed been introduced and shown to be consistent over ukasiewicz logic.
71
The theory
PATr over (or RPL) contains the predicate of equality and the functional predi-
cates of successor, addition, and multiplication, all of them postulated crisp; a (fuzzy)
unary predicate Tr; the axioms of Robinsons arithmetic; the schema of induction for all
formulae (including those containing the predicate Tr) formulated as a deduction rule;
and the dequotation schema as above. Roughly speaking, PATr is thus an extension of
crisp PA by a fuzzy truth predicate Tr. The following are the main results on the theory:
THEOREM 5.4.1 ( [101, 159, 180]).
1. The theory PATr is consistent: there is a crisp (classical) model of PA expand-
able to a [0, 1]

-model of PATr.
2. The standard crisp model of PA (the structure of natural numbers) cannot be
expanded to a model of PATrthe theory is -inconsistent.
3. The extension of PATr with the axioms saying that (arithmetized) Tr commutes
with connectives is inconsistent.
Also a weak variant FQ

of Robinsons arithmetic Q was studied over rst-order

fuzzy logic [81, 93]. In FQ

, the ternary predicates of addition and multiplication

are not assumed to be total functions. The underlying logic is fuzzy (BL or similar).
A variant of G odels rst incompleteness theorem can be proved even for this weak
theory: each axiomatizable consistent extension of FQ

is incomplete; also the essential

undecidability of FQ

can be proved.
71
By H ajek, Paris, and Shepherdson in [101]; for alternative proofs of Theorem 5.4.1(2) see [159] and [180].
90 Libor B ehounek, Petr Cintula, and Petr H ajek
5.4.2 Axiomatic fuzzy set theory
Now let us turn to set theory. First, there is a fuzzy set theory in the style of
ZermeloFraenkel set theory. Recall that the universe of the classical ZermeloFraenkel
set theory ZF is the union of the hierarchy of sets constructed from the empty set by the
transnitely iterated operation of power set. We shall present here an axiomatic system
FST of fuzzy set theory over the logic BL
Z
due to H ajek and Hanikov a introduced
in [100].
72
DEFINITION 5.4.2. FSTis a rst-order theory over the logic BL
Z
with (crisp) equal-
ity = and the binary predicate . Below are listed the axioms of the theory FST, for
any formula in which z is not free and any formula (both and can contain addi-
tional free variables besides those explicitly written in the axioms). The formulation of
the axioms employs the predicates Crisp (crispness of sets) and (fuzzy subsethood),
dened as Crisp(x) (q)(q x q / x), where q / x abbreviates (q x), and
x y (q)(q x q y), and the operations of union , singleton x, and the
empty set , which can be introduced as the Skolem functions (see Theorem 5.1.7) of the
corresponding axioms.
(extensionality) x = y ((x y) &(y x))
(empty set) (x)(y)(y x)
(pair) (z)(u)(u z (u = x u = y))
(union) (z)(u)(u z (z)(u y & y x))
(weak power) (z)(u)(u z (u x))
(innity) (z)( z & (x)(x z (x x z)))
(separation) (z)(u)(u z (u x & (u, x)))
(collection) (z)[(u)(u x (v)(u, v))
(u)(u x (v)(v z & (u, v)))]
(-induction) ((y)(y x (y)) (x)) (x)(x)
(support) (z)(Crisp(z) &(x z))
In FST, one can dene the predicate HCrisp of hereditary crispness in a straight-
forward way. Let HC be the class of all hereditarily crisp sets. It can be shown that FST
proves all axioms of ZF with all quantiers restricted to HC. The class HC is thus an
inner model of ZF (over classical logic) in FST (over BL
Z
). In classical ZFC, each
completely ordered BL-chain determines a natural model of FST, containing with each
fuzzy set all its fuzzy subsets. For more details on FST see [100].
In fuzzy logic it is moreover possible to develop set theory of a different style,
namely a fuzzy version of Cantors nave set theory with unrestricted comprehension.
Recall that in classical logic, unrestricted comprehension is contradictory by Russells
72
H ajek and Hanikov as work follows previous versions of axiomatic fuzzy set theory over a variant and an
expansion of G odel logic by Takeuti and Titani [169, 170], which in turn follows the works on intuitionistic
ZF-style set theory [79, 156]. ZF-style universes of fuzzy sets over ukasiewicz logic had been semantically
studied by Klaua and Gottwald [75, 78, 121, 122]. Acomprehensive survey of various approaches to axiomatic
fuzzy set theory is Gottwalds [77].
Chapter I: Introduction to Mathematical Fuzzy Logic 91
paradox: if x = y [ y / y then x x x / x. However, it is known that
for the derivation of a contradiction from the latter formula in classical or intuitionistic
logic, the rule of contraction (see Section 3.1) is necessary: in some contraction-free
logics (namely, BCK and certain variants of linear logic), the unrestricted compre-
hension schema is known to be consistent [80, 155, 171]. It has been conjectured by
Skolem [167, 168] that the full comprehension schema is consistent over ukasiewicz
logic. (Notice that the formula x x x / x can be satised in [0, 1]

if the
truth value of x x is one half.) Partial consistency results (with comprehension re-
stricted to formulae of limited quantier complexity) are due to Skolem, Chang, and
Fenstad [25, 60, 168].
A set theory over with full comprehension (called Cantorukasiewicz set the-
ory C)
73
can be dened as follows: the language contains a single binary predicate
and the comprehension terms x [ (x) for each formula (possibly containing
other free variables besides x); the only extralogical axioms are the instances of the
comprehension schema
u x [ (x) (u)
for each formula . If consistent (see below), the theory shows markedly different be-
havior from ZF-style fuzzy set theories. For instance, the assumption that the exten-
sional and intensional (or Leibniz) equalities
x =
e
y (z)(z x z y)
x = y (z)(x z y z)
coincide is contradictory in C. Various results (esp. on natural numbers, some of them
negative) have been proved in C [92, 181]. However, no natural classical model of
C has been found and the consistency status of the theory is (as of 2011) unknown.
74
Nevertheless, even if C turns out to be contradictory, the full comprehension schema
has been proved to be consistent at least over IMTL (Terui 2011, pers. comm.), in which
many theorems of C still hold and which shows similar metamathematical features
as C.
5.5 Higher-order fuzzy logic
Two systems of higher-order extension of rst-order fuzzy logic have been devel-
oped: fuzzy Church-style simple type theory FTT and Henkin-style higher-order fuzzy
logic, also known as Fuzzy Class Theory FCT.
5.5.1 Fuzzy type theory
Fuzzy type theory FTT was originally (in [146]) dened over the logic IMTL
Z
;
later it has been generalized to several other logics, including
Z
, BL
Z
, and [147].
Being a Church-style type theory, it is not formulated in the language of rst-order fuzzy
73
In [92] and subsequent papers, the theory was denoted by C
0
, while C denoted a contradictory ex-
tension of C
0
. A corresponding theory over standard rst-order ukasiewicz logic
s
is denoted by H
in [177, 181].
74
In 2010, Terui (pers. comm.) found what appears to be a serious gap in Whites [177] proof of the relative
consistency w.r.t. ZFC.
92 Libor B ehounek, Petr Cintula, and Petr H ajek
logic, but in the typed -calculus (for a standard reference see [9]). We shall therefore
provide only a brief sketch of the theory here, and refer the interested reader to [146] for
the full description.
The system of types in FTT is dened over two primitive types, (for elements)
and o (for truth degrees), in the usual way (for each types , there is a complex type
() for crisp functions from the domain of to the domain of ). The language of
FTT contains an innite countable set of variables x

for each type and the special

constants C
(oo)o
, D
oo
, and E
(o)
for each type . Formulae (i.e., -terms) of type
are dened as usual in typed -calculus.
The semantics of FTT interprets the type as a (non-empty) set M

(of atomic
objects), the type o as a complete linearly ordered IMTL
Z
-algebra M
o
, and each com-
plex type as a subset of the set M
M

of all (crisp) functions from M

to M

. The
constants C
(oo)o
, D
oo
, and E
(oo)o
are interpreted, respectively, as , , and in the
IMTL
Z
-algebra M
o
, the constant E
(o)
as a (primitive) fuzzy equality
75
on M

, and
the constants E
(o())()
as fuzzy equalities E

dened recursively as E

(f, g) =
_
xM

(f(x), g(x)) for each f, g M

. The functions f M

are further-
more required to be weakly extensional w.r.t. E

and E

, i.e., if E

(x, y) = 1 then
E

(f(x), f(y)) = 1 for all x, y M

. The interpretation of -terms is dened as

usual in typed -calculus. A general model is an interpretation such that each formula
is assigned a value (which is a non-trivial condition since M

can be a proper subset

of M
M

).
The axiomatic system of FTT is given by 16 formulae of type o and two inference
rules (of intersubstitutivity of equivalent formulae and -necessitation). The notions
of provability, theory, and model are then dened analogously as in rst-order (fuzzy)
logic. FTT extends the rst-order logic IMTL
Z
, is non-trivial (i.e., has non-bivalent
models), and is sound w.r.t. general models sketched above. It enjoys the -deduction
theorem and the following form of completeness: a theory over FTT is consistent iff it
has a general model.
FTT can serve as a background framework for more specialized axiomatic theories
(formulated as sets of FTT-formulae of type o). It has mostly been used for model-
ing linguistic phenomena connected with the agenda of fuzzy logic in broader sense,
such as generalized quantiers [149], evaluative linguistic expressions [148]. This en-
deavor includes the development of necessary fuzzy-mathematical backgrounde.g.,
fuzzy measures and integrals [49].
5.5.2 Fuzzy Class Theory
Fuzzy Class Theory FCT was originally introduced in [12] over the logics and

1
2
, but can in fact be dened over any extension of MTL
Z
(see, e.g., [11]). It can
be characterized as Henkin-style higher-order fuzzy logic, or Russell-style simple fuzzy
type theory.
Let L be a -core fuzzy logic. Fuzzy Class Theory (FCT) over L, or Henkin-style
higher-order fuzzy logic L, is a theory in multi-sorted rst-order fuzzy logic L with
75
I.e., a function E: M
2

M
o
such that (i) Exx = 1, (ii) Exy = Eyx, (iii) Exy Eyz Exz, and
(iv) Exy = 1 iff x = y, for all x, y, z M
o
. Such functions are also known as separated (or unimodal)
fuzzy similarities.
Chapter I: Introduction to Mathematical Fuzzy Logic 93
identity. The language of FCT has the sorts of variables for:
Atomic objects (lowercase letters x, y, . . . )
Fuzzy classes of atomic objects (uppercase letters A, B, . . . )
Fuzzy classes of fuzzy classes of atomic objects (calligraphic letters /, B, . . . )
Etc., in general fuzzy classes of the n-th order (X
(n)
, Y
(n)
, . . . )
Each of these sorts subsumes sorts for tuples of all arities k N. Variables of order n
and arity k will be denoted by X
(n,k)
, Y
(n,k)
, . . .
Besides the (crisp) identity predicate = on each sort, the language of FCT contains:
The membership predicate between objects of sorts
(n)
and
(n+1,1)
Class terms x [ of sort
(n+1,1)
, for any variable x of sort
(n)
and any for-
mula
Terms x
1
, . . . , x
k
of sort
(n,k)
for k-tuples of individuals x
1
, . . . , x
k
of sort
(n,1)
FCT has the following axioms over multi-sorted rst-order logic L, for all formulae
(x, . . . ), variables of all compatible sorts, and all k N:
The tuple-identity axioms: x
1
, . . . , x
k
= y
1
, . . . , y
k
x
1
= y
1
& . . . &
x
k
= y
k
The comprehension axioms: y x [ (x) (y)
The extensionality axioms: (x)(x A x B) A = B
The axioms of extensionality express the stipulation that fuzzy sets (of any order) are
determined by their membership functions. The axioms of comprehension ensure that
each property denable in the language of the theory delimits a fuzzy set x [ of
the appropriate order.
76
The intended models of FCT are determined by a (linear) L-algebra L and a crisp
universe of discourse U. The variables of sort
(n,k)
range over the set U
n,k
= U
k
n,1
and the universal variables of sort
(n)
over U
n
=

kN
U
n,k
, where U
0,1
= U and
U
n+1,1
= L
U
n
for each n N. For x U
n
and A U
n+1,1
, the value of membership
predicate x A is dened simply as the functional value A(x); the tuples and identity
are interpreted in the obvious way. An intended model of FCT thus consists of fuzzy
sets of all nite orders and arities, represented by L-valued membership functions. FCT
is sound w.r.t. its intended models, and can thus be viewed as an (incomplete, but suf-
ciently strong) axiomatization of Zadehs notion of fuzzy set (of all orders and arities).
77
Since the theory of intended models of FCT is not recursively axiomatizable (for
the same reasons as in classical higher-order fuzzy logic, namely the denability of true
76
Alternatively, FCT can be formulated with the comprehension axioms (A).(x A (x)). The
(eliminable) comprehension terms then arise as Skolem functions (see Section 5.3) of these axioms.
77
Intended models over standard L-algebras are therefore called Zadeh models in [12] and subsequent pa-
pers.
94 Libor B ehounek, Petr Cintula, and Petr H ajek
arithmetic), FCT is only complete w.r.t. its general models, in which U
n+1,1
can be a
proper subset of L
U
n
(closed under all denable operations) and U
n
a proper superset
of

kN
U
n,k
. Thus, not all fuzzy subsets of U
n
(i.e., functions U
n
L) may be
represented as a fuzzy class (i.e., an element of U
n+1,1
) in a general model of FCT;
78
nevertheless, the axioms of comprehension ensure that at least all fuzzy subsets of U
n
denable in the language of the theory are represented by elements of U
n+1,1
.
The axiomatization of fuzzy sets in FCT turns out to be strong enough to formal-
ize all usual notions of fuzzy set theory and other disciplines of fuzzy mathematics.
By means of , all notions denable in classical higher-order logic can in fact be rep-
resented in FCT extended by appropriate denitions [12, 39]; all usual structures of
classical mathematics (e.g., natural or real numbers) are thus available in the framework
of FCT. Since moreover the language of FCT is similar to that of classical math-
ematics, FCT was proposed (in [13]) for a foundational theory for logic-based fuzzy
mathematics. Several disciplines of fuzzy mathematics have thereafter been developed
in the foundational framework of FCT, including elementary theory of fuzzy sets [12],
fuzzy relations [11, 15], fuzzy topology [16, 125], fuzzy interval arithmetic [111], etc.
Acknowledgments
Thanks are due to Francesco Paoli and Llus Godo for comments on an earlier draft
of this Chapter, to Eva Posp silov a for redrawing the pictures, and to Martina Da nkov a
for generating the 3D graphs of t-norms and their residua. All authors acknowledge
the support of the grant ICC/08/E018 of the Czech Science Foundation (part of ESF
Eurocores project LogiCCC FP006 LoMoReVI) and the Institutional Research Plan
AV0Z10300504.
BIBLIOGRAPHY
[1] Rom` a J. Adillon,
`
Angel Garca-Cerda na, and Ventura Verd u. On three implication-less fragments of
t-norm based fuzzy logics. Fuzzy Sets and Systems, 158(23):25752590, 2007.
[2] Paolo Aglian` o, Isabel Maria Andr e Ferreirim, and Franco Montagna. Basic hoops: An algebraic study
of continuous t-norms. Studia Logica, 87(1):7398, 2007.
[3] Arnon Avron. A constructive analysis of RM. Journal of Symbolic Logic, 52(4):939951, 1987.
[4] Arnon Avron. Hypersequents, logical consequence and intermediate logics for concurrency. Annals of
Mathematics and Articial Intelligence, 4(34):225248, 1991.
[5] Arnon Avron and Iddo Lev. Non-deterministic multiple-valued structures. Journal of Logic and Com-
putation, 15(3):241261, 2005.
[6] Matthias Baaz, Agata Ciabattoni, and Franco Montagna. Analytic calculi for monoidal t-norm based
logic. Fundamenta Informaticae, 59(4):315332, 2004.
[7] Matthias Baaz, Petr H ajek, Jan Kraj cek, and David

Svejda. Embedding logics into product logic.
Studia Logica, 61(1):3547, 1998.
[8] Matthias Baaz, Petr H ajek, Franco Montagna, and Helmut Veith. Complexity of t-tautologies. Annals
of Pure and Applied Logic, 113(13):311, 2002.
[9] Henk P. Barendregt. The Lambda Calculus, Its Syntax and Semantics, volume 103 of Studies in Logic
and the Foundations of Mathematics. Elsevier, 1984.
[10] Libor B ehounek. On the difference between traditional and deductive fuzzy logic. Fuzzy Sets and
Systems, 159(10):11531164, 2008.
78
It can be observed that if the underlying logic L is weaker than classical logic, then FCT also has crisp
models, isomorphic to general models of classical Henkin-style higher-order logic. If necessary, an additional
axiom such as (A)(x A (x A)) can be adopted to ensure the fuzziness of all its models.
Chapter I: Introduction to Mathematical Fuzzy Logic 95
[11] Libor B ehounek, Ulrich Bodenhofer, and Petr Cintula. Relations in Fuzzy Class Theory: Initial steps.
Fuzzy Sets and Systems, 159(14):17291772, 2008.
[12] Libor B ehounek and Petr Cintula. Fuzzy class theory. Fuzzy Sets and Systems, 154(1):3455, 2005.
[13] Libor B ehounek and Petr Cintula. From fuzzy logic to fuzzy mathematics: A methodological mani-
festo. Fuzzy Sets and Systems, 157(5):642646, 2006.
[14] Libor B ehounek and Petr Cintula. Fuzzy logics as the logics of chains. Fuzzy Sets and Systems,
157(5):604610, 2006.
[15] Libor B ehounek and Martina Da nkov a. Relational compositions in Fuzzy Class Theory. Fuzzy Sets
and Systems, 160(8):10051036, 2009.
[16] Libor B ehounek and Tom a s Kroupa. Topology in Fuzzy Class Theory: Basic notions. In Patricia Melin,
Oscar Castillo, Lluis T. Aguilar, Janusz Kacprzyk, and Witold Pedrycz, editors, Foundations of Fuzzy
Logic and Soft Computing, volume 4529 of Lecture Notes in Articial Intelligence, pages 513522.
Springer, Berlin etc., 2007.
[17] Willem J. Blok and Isabel Maria Andr e Ferreirim. On the structure of hoops. Algebra Universalis,
43(23):233257, 2000.
[18] Willem J. Blok and Don L. Pigozzi. Algebraizable Logics, volume 396 of Memoirs of the American
Mathematical Society. American Mathematical Society, Providence, RI, 1989. Freely downloadable
from http://orion.math.iastate.edu/dpigozzi/.
[19] Willem J. Blok and Clint J. van Alten. The nite embeddability property for residuated lattices, pocrims
and BCK-algebras. Algebra Universalis, 48(3):253271, 2002.
[20] F elix Bou, Francesc Esteva, Josep Maria Font,
`
Angel Gil, Llus Godo, Antoni Torrens, and Ventura
Verd u. Logics preserving degrees of truth from varieties of residuated lattices. Journal of Logic and
Computation, 19(6):10311069, 2009.
[21] J.R. B uchi and Owens T.M. Complemented monoids and hoops. Unpublished manuscript.
[22] Sam Buss, editor. Handbook of Proof Theory. Kluwer, 1998.
[23] Ettore Casari. Comparative logics and Abelian -groups. In Logic Colloquium 88 (Padova, 1988),
volume 127 of Studies in Logic and the Foundations of Mathematics, pages 161190. North-Holland,
Amsterdam, 1989.
[24] Chen Chung Chang. A new proof of the completeness of the ukasiewicz axioms. Transactions of the
American Mathematical Society, 93(1):7480, 1959.
[25] Chen Chung Chang. The axiom of comprehension in innite valued logic. Mathematica Scandinavica,
13:930, 1963.
[26] Karel Chvalovsk y. On the independence of axioms in BL and MTL. To appear in Fuzzy Sets and
Systems.
[27] Agata Ciabattoni, Francesc Esteva, and Llus Godo. T-norm based logics with n-contraction. Neural
Network World, 12(5):453460, 2002.
[28] Agata Ciabattoni, Nikolaos Galatos, and Kazushige Terui. Fromaxioms to analytic rules in nonclassical
logics. In Proceedings of the 23rd Annual IEEE Symposium on Logic in Computer Science (LICS),
pages 229240, 2008.
[29] Roberto Cignoli, Francesc Esteva, Llus Godo, and Antoni Torrens. Basic fuzzy logic is the logic of
continuous t-norms and their residua. Soft Computing, 4(2):106112, 2000.
[30] Petr Cintula. An alternative approach to the logic. Neural Network World, 11(6):561572, 2001.
[31] Petr Cintula. The and
1
2
propositional and predicate logics. Fuzzy Sets and Systems,
124(3):289302, 2001.
[32] Petr Cintula. Advances in the and
1
2
logics. Archive for Mathematical Logic, 42(5):449468,
2003.
[33] Petr Cintula. From Fuzzy Logic to Fuzzy Mathematics. PhD thesis, Czech Technical University, Faculty
of Nuclear Sciences and Physical Egineering, Prague, 2005.
[34] Petr Cintula. Weakly implicative (fuzzy) logics I: Basic properties. Archive for Mathematical Logic,
45(6):673704, 2006.
[35] Petr Cintula, Francesc Esteva, Joan Gispert, Llus Godo, Franco Montagna, and Carles Noguera. Dis-
tinguished algebraic semantics for t-norm based fuzzy logics: Methods and algebraic equivalencies.
Annals of Pure and Applied Logic, 160(1):5381, 2009.
[36] Petr Cintula and Brunella Gerla. Semi-normal forms and functional representation of product fuzzy
logic. Fuzzy Sets and Systems, 143(1):89110, 2004.
[37] Petr Cintula and Petr H ajek. Triangular norm predicate fuzzy logics. Fuzzy Sets and Systems,
161(3):311346, 2010.
[38] Petr Cintula, Petr H ajek, and Rostislav Hor ck. Formal systems of fuzzy logic and their fragments.
Annals of Pure and Applied Logic, 150(13):4065, 2007.
96 Libor B ehounek, Petr Cintula, and Petr H ajek
[39] Petr Cintula and Rostislav Hor ck. Fuzzy Class Theory: Some advanced topics. In Martin

St epni cka,
Vil em Nov ak, and Ulrich Bodenhofer, editors, New Dimensions in Fuzzy Logic and Related Technolo-
gies. Proceedings of the 5th EUSFLAT Conference, volume I, pages 137144, Ostrava, 2007. University
of Ostrava.
[40] Petr Cintula, Erich Peter Klement, Radko Mesiar, and Mirko Navara. Residuated logics based on strict
t-norms with an involutive negation. Mathematical Logic Quarterly, 52(3):269282, 2006.
[41] Petr Cintula and Carles Noguera. Implicational (semilinear) logics I: A new hierarchy. Archive for
Mathematical Logic, 49(4):417446, 2010.
[42] Janusz Czelakowski. Protoalgebraic Logics, volume 10 of Trends in Logic. Kluwer, Dordrecht, 2001.
[43] Bernard De Baets. Idempotent uninorms. European Journal of Operational Research, 118(3):631642,
1999.
[44] Antonio Di Nola and Anatolij Dvure censkij. Product MV-algebras. Multiple-Valued Logic, 6(1
2):193215, 2001.
[45] Antonio Di Nola, Francesc Esteva, Pere Garcia, Llus Godo, and Salvatore Sessa. Subvarieties of
BL-algebras generated by single-component chains. Archive for Mathematical Logic, 41(7):673685,
2002.
[46] Antonio Di Nola, George Georgescu, and Afrodita Iorgulescu. Pseudo-BL-algebras, part III. Multiple-
Valued Logic, 8(56):673714, 717750, 2002.
[47] Michael Dummett. A propositional calculus with denumerable matrix. Journal of Symbolic Logic,
24(2):97106, 1959.
[48] J. Michael Dunn and Greg Restall. Relevance logic. In Dov Gabbay and Franz Guenther, editors,
Handbook of Philosophical Logic, volume 6, pages 1136. Kluwer, second edition, 2002.
[49] Antonn Dvor ak and Michal Hol capek. L-fuzzy quantiers of type 1) determined by fuzzy measures.
Fuzzy Sets and Systems, 160(23):34253452, 2009.
[50] Francesc Esteva, Joan Gispert, Llus Godo, and Franco Montagna. On the standard and rational com-
pleteness of some axiomatic extensions of the monoidal t-norm logic. Studia Logica, 71(2):199226,
2002.
[51] Francesc Esteva, Joan Gispert, Llus Godo, and Carles Noguera. Adding truth-constants to logics of
continuous t-norms: Axiomatization and completeness results. Fuzzy Sets and Systems, 158(6):597
618, 2007.
[52] Francesc Esteva and Llus Godo. Putting together ukasiewicz and product logic. Mathware and Soft
Computing, 6(23):219234, 1999.
[53] Francesc Esteva and Llus Godo. Monoidal t-norm based logic: Towards a logic for left-continuous
t-norms. Fuzzy Sets and Systems, 124(3):271288, 2001.
[54] Francesc Esteva, Llus Godo, and
`
Angel Garca-Cerda na. On the hierarchy of t-norm based residuated
fuzzy logics. In Melvin Chris Fitting and Ewa Orlowska, editors, Beyond Two: Theory and Applications
of Multiple-Valued Logic, volume 114 of Studies in Fuzziness and Soft Computing, pages 251272.
Physica-Verlag, Heidelberg, 2003.
[55] Francesc Esteva, Llus Godo, Petr H ajek, and Franco Montagna. Hoops and fuzzy logic. Journal of
Logic and Computation, 13(4):532555, 2003.
[56] Francesc Esteva, Llus Godo, Petr H ajek, and Mirko Navara. Residuated fuzzy logics with an involutive
negation. Archive for Mathematical Logic, 39(2):103124, 2000.
[57] Francesc Esteva, Llus Godo, and Franco Montagna. The and
1
2
logics: Two complete fuzzy
systems joining ukasiewicz and product logics. Archive for Mathematical Logic, 40(1):3967, 2001.
[58] Francesc Esteva, Llus Godo, and Franco Montagna. Axiomatization of any residuated fuzzy logic
dened by a continuous t-norm. In Taner Bilgic, Bernard De Baets, and Okays Kaynak, editors, Fuzzy
Sets and Systems IFSA 2003, volume 2715 of Lecture Notes in Computer Science, pages 172179.
Springer, Berlin/Heidelberg, 2003.
[59] Francesc Esteva, Llus Godo, and Carles Noguera. Expanding the propositional logic of a t-norm with
truth-constants: Completeness results for rational semantics. Soft Computing, 14(3):273284, 2010.
[60] Jens Erik Fenstad. On the consistency of the axiom of comprehension in the ukasiewicz innite valued
logic. Mathematica Scandinavica, 14:6574, 1964.
[61] Isabel Maria Andr e Ferreirim. On Varieties and Quasivarieties of Hoops and Their Reducts. PhD
thesis, University of Illinois at Chicago, Chicago, 1992.
[62] Tommaso Flaminio and Enrico Marchioni. T-norm based logics with an independent involutive nega-
tion. Fuzzy Sets and Systems, 157(4):31253144, 2006.
[63] Paul Flondor, George Georgescu, and Afrodita Iorgulescu. Pseudo t-norms and pseudo-BL-algebras.
Soft Computing, 5(5):355371, 2001.
Chapter I: Introduction to Mathematical Fuzzy Logic 97
[64] J anos Fodor. Nilpotent minimum and related connectives for fuzzy logic. In Proc. FUZZ-IEEE 1995,
pages 20772082, 1995.
[65] Josep Maria Font. Taking degrees of truth seriously. Studia Logica, 91(3):383406, 2009.
[66] Dov M. Gabbay and George Metcalfe. Fuzzy logics based on [0, 1)-continuous uninorms. Archive for
Mathematical Logic, 46(6):425469, 2007.
[67] Nikolaos Galatos, Peter Jipsen, Tomasz Kowalski, and Hiroakira Ono. Residuated Lattices: An Alge-
braic Glimpse at Substructural Logics, volume 151 of Studies in Logic and the Foundations of Mathe-
matics. Elsevier, Amsterdam, 2007.
[68] Nikolaos Galatos and Hiroakira Ono. Algebraization, parametrized local deduction theorem and inter-
polation for substructural logics over FL. Studia Logica, 83(13):279308, 2006.
[69] George Georgescu and Afrodita Iorgulescu. Pseudo-MV algebras. Multiple-Valued Logic, 6(12):95
135, 2001.
[70] George Georgescu, Laurent iu Leustean, and Viorel Preoteasa. Pseudo-hoops. Journal of Multiple-
Valued Logic and Soft Computing, 11(12):153184, 2005.
[71] Giangiacomo Gerla. Fuzzy LogicMathematical Tool for Approximate Reasoning, volume 11 of
Trends in Logic. Kluwer and Plenum Press, New York, 2001.
[72] Jean-Yves Girard. Linear logic. Theoretical Computer Science, 50(1):1102, 1987.
[73] Joan Gispert and Antoni Torrens. Axiomatic extensions of IMT3 logic. Studia Logica, 81(3):311324,
2005.
[74] Kurt G odel. Zum intuitionistischen Aussagenkalk ul. Anzieger Akademie der Wissenschaften Wien,
69:6566, 1932.
[75] Siegfried Gottwald. Set theory for fuzzy sets of higher level. Fuzzy Sets and Systems, 2(2):125151,
1979.
[76] Siegfried Gottwald. Fuzzy Sets and Fuzzy Logic: Foundations of Applicationfrom a Mathematical
Point of View. Vieweg, Wiesbaden, 1993.
[77] Siegfried Gottwald. Universes of fuzzy sets and axiomatizations of fuzzy set theory. Part I: Model-
based and axiomatic approaches. Studia Logica, 82(2):211144, 2006.
[78] Siegfried Gottwald. An early approach toward graded identity and graded membership in set theory.
Fuzzy Sets and Systems, 161(18):23692379, 2010.
[79] Robin J. Grayson. Heyting-valued models for intuitionistic set theory. In M. Fourman, C. Mulvey, and
D.S. Scott, editors, Application of Sheaves, volume 743 of Lecture Notes in Computer Science, pages
402414. Springer, Berlin, 1979.
[80] Vyacheslav Nikolaevich Grishin. Predicate and set-theoretic calculi based on logic without contrac-
tions. Mathematics of the USSRIzvestiya, 18:4159, 1982.
[81] Petr H ajek. Towards metamathematics of weak arithmetic over fuzzy logic. To appear in Logic Journal
of IGPL, doi:10.1093/jigpal/jzp091.
[82] Petr H ajek. Basic fuzzy logic and BL-algebras. Soft Computing, 2(3):124128, 1998.
[83] Petr H ajek. Metamathematics of Fuzzy Logic, volume 4 of Trends in Logic. Kluwer, Dordrecht, 1998.
[84] Petr H ajek. Function symbols in fuzzy logic. In Proceedings of the EastWest Fuzzy Colloquium, pages
28, Zittau/G orlitz, 2000. IPM.
[85] Petr H ajek. Fuzzy logic and arithmetical hierarchy III. Studia Logica, 68(1):129142, 2001.
[86] Petr H ajek. Observations on the monoidal t-norm logic. Fuzzy Sets and Systems, 132(1):107112,
2002.
[87] Petr H ajek. Fuzzy logics with noncommutative conjunctions. Journal of Logic and Computation,
13(4):469479, 2003.
[88] Petr H ajek. Observations on non-commutative fuzzy logic. Soft Computing, 8(1):3843, 2003.
[89] Petr H ajek. Fuzzy logic and arithmetical hierarchy IV. In Vincent F. Hendricks, Fabian Neuhaus,
Stig Andur Pedersen, Uwe Schefer, and Heinrich Wansing, editors, First-Order Logic Revised, pages
107115. Logos Verlag, Berlin, 2004.
[90] Petr H ajek. Arithmetical complexity of fuzzy predicate logicsa survey. Soft Computing, 9(12):935
941, 2005.
[91] Petr H ajek. Fleas and fuzzy logic. Journal of Multiple-Valued Logic and Soft Computing, 11(12):137
152, 2005.
[92] Petr H ajek. On arithmetic in Cantor-ukasiewicz fuzzy set theory. Archive for Mathematical Logic,
44(6):763782, 2005.
[93] Petr H ajek. Mathematical fuzzy logic and natural numbers. Fundamenta Informaticae, 81(13):155
163, 2007.
[94] Petr H ajek. On witnessed models in fuzzy logic. Mathematical Logic Quarterly, 53(1):6677, 2007.
98 Libor B ehounek, Petr Cintula, and Petr H ajek
[95] Petr H ajek. On witnessed models in fuzzy logic II. Mathematical Logic Quarterly, 53(6):610615,
2007.
[96] Petr H ajek. On witnessed models in fuzzy logic III witnessed g odel logics. Mathematical Logic
Quarterly, 56(2):171174, 2010.
[97] Petr H ajek. Comments on interpretability and decidability in fuzzy logic. Journal of Logic and Com-
putation, 21(5):823828, 2011.
[98] Petr H ajek and Petr Cintula. On theories and models in fuzzy predicate logics. Journal of Symbolic
Logic, 71(3):863880, 2006.
[99] Petr H ajek, Llus Godo, and Francesc Esteva. A complete many-valued logic with product conjunction.
Archive for Mathematical Logic, 35(3):191208, 1996.
[100] Petr H ajek and Zuzana Hanikov a. A development of set theory in fuzzy logic. In Melvin Chris Fitting
and Ewa Orlowska, editors, Beyond Two: Theory and Applications of Multiple-Valued Logic, volume
114 of Studies in Fuzziness and Soft Computing, pages 273285. Physica-Verlag, Heidelberg, 2003.
[101] Petr H ajek, Jeff Paris, and John C. Shepherdson. The liar paradox and fuzzy logic. Journal of Symbolic
Logic, 65(1):339346, 2000.
[102] Petr H ajek, Jeff Paris, and John C. Shepherdson. Rational Pavelka logic is a conservative extension of
ukasiewicz logic. Journal of Symbolic Logic, 65(2):669682, 2000.
[103] Petr H ajek and Sauro Tulipani. Complexity of fuzzy probability logics. Fundamenta Informaticae,
45(3):207213, 2001.
[104] Zuzana Hanikov a. Varieties generated by standard BL-algebras. Submitted.
[105] Zuzana Hanikov a and Petr Savick y. Distinguishing standard SBL-algebras with involutive negations
by propositional formulas. Mathematical Logic Quarterly, 54(6):579596, 2008.
[106] Louise Schmir Hay. Axiomatization of the innite-valued predicate calculus. Journal of Symbolic
Logic, 28(1):7786, 1963.
[107] Ulrich H ohle. Commutative, residuated l-monoids. In Ulrich H ohle and Erich Peter Klement, editors,
Non-Classical Logics and Their Applications to Fuzzy Subsets, pages 53106. Kluwer, Dordrecht,
1995.
[108] Rostislav Hor ck. Standard completeness theorem for MTL. Archive for Mathematical Logic,
44(4):413424, 2005.
[109] Rostislav Hor ck. Decidability of cancellative extension of monoidal t-norm based logic. Logic Journal
of the Interest Group of Pure and Applied Logic, 14(6):827843, 2006.
[110] Rostislav Hor ck. On the failure of standard completeness in MTL for innite theories. Fuzzy Sets
and Systems, 158(6):619624, 2007.
[111] Rostislav Hor ck. Solution of a system of linear equations with fuzzy numbers. Fuzzy Sets and Systems,
159(14):17881810, 2008.
[112] Rostislav Hor ck and Petr Cintula. Product ukasiewicz logic. Archive for Mathematical Logic,
43(4):477503, 2004.
[113] Rostislav Hor ck, Carles Noguera, and Milan Petrk. On n-contractive fuzzy logics. Mathematical
Logic Quarterly, 53(3):268288, 2007.
[114] Rostislav Hor ck and Kazushige Terui. Disjunction property and complexity of substructural logics.
Theoretical Computer Science, 412(31):39924006, 2011.
[115] Alfred Horn. Logic with truth values in a linearly ordered Heyting algebras. Journal of Symbolic Logic,
34(3):395408, 1969.
[116] Y. Imai and Kiyoshi Is eki. On axiomatic systems of propositional calculi XIV. Proceedings Japan
Academy, 42:1922, 1966.
[117] S andor Jenei and Franco Montagna. A proof of standard completeness for Esteva and Godos logic
MTL. Studia Logica, 70(2):183192, 2002.
[118] S andor Jenei and Franco Montagna. A proof of standard completeness for non-commutative monoidal
t-norm logic. Neural Network World, 13(5):481489, 2003.
[119] Peter Jipsen and Constantine Tsinakis. A survey of residuated lattices. In J. Martinez, editor, Ordered
Algebraic Structures, pages 1956. Kluwer Academic Publishers, Dordrecht, 2002.
[120] Heldermann Keimel. Some trends in lattice-ordered groups and rings. In K. A. Baker and R. Wille,
editors, Lattice Theory and its Applications. In Celebration of Garrett Birkhoff s 80th Birthday., vol-
ume 23 of Research and Exposition in Mathematics, pages 131161. Heldermann, Darmstadt, 1995.
[121] Dieter Klaua.

Uber einen Ansatz zur mehrwertigen Mengenlehre. Monatsb. Deutsch. Akad. Wiss.
Berlin, 7:859867, 1965.
[122] Dieter Klaua.

Uber einen zweiten Ansatz zur mehrwertigen Mengenlehre. Monatsb. Deutsch. Akad.
Wiss. Berlin, 8:782802, 1966.
Chapter I: Introduction to Mathematical Fuzzy Logic 99
[123] Erich Peter Klement, Radko Mesiar, and Endre Pap. Triangular Norms, volume 8 of Trends in Logic.
Kluwer, Dordrecht, 2000.
[124] Tomasz Kowalski. Semisimplicity, EDPC and discriminator varieties of residuated lattices. Studia
Logica, 77(2):255265, 2004.
[125] Tom a s Kroupa. Filters in Fuzzy Class Theory. Fuzzy Sets and Systems, 159(14):17731787, 2008.
[126] Jan K uhr. Pseudo-BL algebras and PRl-monoids. Mathematica Bohemica, 128:199208, 2003.
[127] Michael C. Laskowski and Shirin Malekpour. Provability in predicate product logic. Archive for
Mathematical Logic, 46(56):365378, 2007.
[128] Jan ukasiewicz. O logice tr ojwarto sciowej (On three-valued logic). Ruch lozoczny, 5:170171,
1920.
[129] Jan ukasiewicz and Alfred Tarski. Untersuchungen uber den Aussagenkalk ul. Comptes Rendus des
S eances de la Soci et e des Sciences et des Lettres de Varsovie, cl. III, 23(iii):3050, 1930.
[130] Enrico Marchioni and Franco Montagna. On triangular norms and uninorms denable in
1
2
. Inter-
national Journal of Approximate Reasoning, 47(2):179201, 2008.
[131] Robert McNaughton. A theorem about innite-valued sentential logic. Journal of Symbolic Logic,
16(1):113, 1951.
[132] George Metcalfe. Proof Theory for Propositional Fuzzy Logics. PhD thesis, Department of Computer
Science, Kings College, London, 2004.
[133] George Metcalfe and Franco Montagna. Substructural fuzzy logics. Journal of Symbolic Logic,
72(3):834864, 2007.
[134] George Metcalfe, Nicola Olivetti, and Dov Gabbay. Analytic proof calculi for product logics. Archive
for Mathematical Logic, 43(7):859889, 2004.
[135] George Metcalfe, Nicola Olivetti, and Dov Gabbay. Sequent and hypersequent calculi for Abelian and
ukasiewicz logics. ACM Transactions on Computational Logic, 6(3):578613, 2005.
[136] George Metcalfe, Nicola Olivetti, and Dov M. Gabbay. Proof Theory for Fuzzy Logics, volume 36 of
Applied Logic Series. Springer, 2008.
[137] Franco Montagna. An algebraic approach to propositional fuzzy logic. Journal of Logic, Language
and Information, 9:91124, 2000.
[138] Franco Montagna. Three complexity problems in quantied fuzzy logic. Studia Logica, 68(1):143152,
2001.
[139] Franco Montagna. On the predicate logics of continuous t-normBL-algebras. Archive for Mathematical
Logic, 44(1):97114, 2005.
[140] Franco Montagna, Carles Noguera, and Rostislav Hor ck. On weakly cancellative fuzzy logics. Journal
of Logic and Computation, 16(4):423450, 2006.
[141] Franco Montagna and Giovanni Panti. Adding structure to MV-algebras. Journal of Pure and Applied
Algebra, 164(3):365387, 2001.
[142] Ant onio A. Monteiro. Sur les alg` ebres de Heyting sym etriques. Portugaliae Mathematica, 39(14):1
239, 1980.
[143] Daniele Mundici. Satisability in many-valued sentential logic is NP-complete. Theoretical Computer
Science, 52(12):145153, 1987.
[144] Daniele Mundici. A constructive proof of McNaughtons Theorem in innite-valued logics. Journal of
Symbolic Logic, 59(2):596602, 1994.
[145] Carles Noguera. Algebraic Study of Axiomatic Extensions of Triangular Norm Based Fuzzy Logics,
volume 27 of Monograes de lInstitut dInvestigaci o en Intellig` encia Articial. Consell Superior
dInvestigacions Cientques, Barcelona, 2007.
[146] Vil em Nov ak. On fuzzy type theory. Fuzzy Sets and Systems, 149(2):235273, 2004.
[147] Vil em Nov ak. Fuzzy type theory as higher order fuzzy logic. In Proc. The Sixth International Confer-
ence on Intelligent Technologies, pages 2126, Bankgok, 2005.
[148] Vil emNov ak. Fuzzy logic theory of evaluating expressions and comparative quantiers. In Proceedings
of 11th IPMU Conference, volume 2, pages 15721579, Paris, 2006. Edition EDK.
[149] Vil em Nov ak. A formal theory of intermediate quantiers. Fuzzy Sets and Systems, 159(10):1229
1246, 2008.
[150] Vil em Nov ak, Irina Perlieva, and Jir Mo ckor. Mathematical Principles of Fuzzy Logic. Kluwer,
Dordrecht, 2000.
[151] Hiroakira Ono. Substructural logics and residuated latticesan introduction. In Vincent F. Hendricks
and Jacek Malinowski, editors, 50 Years of Studia Logica, volume 21 of Trends in Logic, pages 193
228. Kluwer, Dordrecht, 2003.
[152] Hiroakira Ono and Yuichi Komori. Logic without the contraction rule. Journal of Symbolic Logic,
50(1):169201, 1985.
100 Libor B ehounek, Petr Cintula, and Petr H ajek
[153] Francesco Paoli. Substructural Logics: A Primer, volume 13 of Trends in Logic. Kluwer, Dordrecht,
2002.
[154] Jan Pavelka. On fuzzy logic I, II, III. Zeitschrift f ur Mathematische Logik und Grundlagen der Mathe-
matik, 25:4552, 119134, 447464, 1979.
[155] Uwe Petersen. Logic without contraction as based on inclusion and unrestricted abstraction. Studia
Logica, 64(3):365403, 2000.
[156] William C. Powell. Extending G odels negative interpretation to ZF. Journal of Symbolic Logic,
40(2):221229, 1975.
[157] Andreja Prijatelj. Bounded contraction and Gentzen-style formulation of ukasiewicz logics. Studia
Logica, 57(23):437456, 1996.
[158] Helena Rasiowa. An Algebraic Approach to Non-Classical Logics. North-Holland, Amsterdam, 1974.
[159] Greg Restall. Arithmetic and truth in ukasiewiczs innitely valued logic. Logique et Analyse, 36:25
38, 1995.
[160] Greg Restall. An Introduction to Substructural Logics. Routledge, New York, 2000.
[161] Hartley Rogers, Jr. Theory of Recursive Functions and Effective Computability. McGraw-Hill, 1967.
[162] Alan Rose. Formalisation du calcul propositionnel implicatif a
0
valeurs de ukasiewicz. Comptes
Rendus Hebdomadaire des S eances de lAcad emie des Sciences, Paris, 243:11831185, 1956.
[163] Alan Rose and J.Barkley Rosser. Fragments of many-valued statement calculi. Transactions of the
American Mathematical Society, 87(1):153, 1958.
[164] Richard Routley, Val Plumwood, Robert K. Meyer, and Ross T. Brady. Relevant Logics and their Rivals.
Part I. Ridgeview Publishing Co., Atascadero, CA, 1982. The Basic Philosophical and Semantical
Theory.
[165] Giovanni Sambin, Giulia Battilotti, and Claudia Faggian. Basic logic: Reection, symmetry, visibility.
Journal of Symbolic Logic, 65(3):9791013, 2000.
[166] Petr Savick y, Roberto Cignoli, Francesc Esteva, Llus Godo, and Carles Noguera. On product logic
with truth constants. Journal of Logic and Computation, 16(2):205225, 2006.
[167] Thoralf Skolem. Bemerkungen zum Komprehensionsaxiom. Zeitschrift f ur Mathematische Logik und
Grundlagen der Mathematik, 3(15):117, 1957.
[168] Thoralf Skolem. A set theory based on a certain 3-valued logic. Mathematica Scandinavica, 8:7180,
1960.
[169] Gaisi Takeuti and Satoko Titani. Intuitionistic fuzzy logic and intuitionistic fuzzy set theory. Journal
of Symbolic Logic, 49(3):851866, 1984.
[170] Gaisi Takeuti and Satoko Titani. Fuzzy logic and fuzzy set theory. Archive for Mathematical Logic,
32(1):132, 1992.
[171] Kazushige Terui. Light afne set theory: A naive set theory of polynomial time. Studia Logica,
77(1):940, 2004.
[172] Anne S. Troelstra. Lectures on Linear Logic, volume 29 of CSLI Lecture Notes. ????, ????, 1992.
[173] Thomas Vetterlein. Partial algebras for ukasiewicz logic and its extensions. Archive for Mathematical
Logic, 44(7):913933, 2005.
[174] Albert Visser. A propositional logic with explicit xed points. Studia Logica, 40(2):155175, 1981.
[175] San-Min Wang and Petr Cintula. Logics with disjunction and proof by cases. Archive for Mathematical
Logic, 47(5):435446, 2008.
[176] Sanmin Wang and Bin Zhao. HpsUL is not the logic of pseudo-uninorms and their residua. Logic
Journal of the Interest Group of Pure and Applied Logic, 17(4):413419, 2009.
[177] Richard B. White. The consistency of the axiom of comprehension in the innite-valued predicate logic
of ukasiewicz. Journal of Philosophical Logic, 8(1):509534, 1979.
[178] Ryszard W ojcicki. Theory of Logical Calculi, volume 199 of Synthese Library. Kluwer Academic
Publishers, Dordrecht/Boston/London, 1988.
[179] Ronald R. Yager and Alexander Rybalov. Uninorm aggregation operators. Fuzzy Sets and Systems,
80(1):111120, 1996.
[180] Shunsuke Yatabe. A note on H ajek, Paris and Shepherdsons theorem. Logic Journal of the Interest
Group of Pure and Applied Logic, 13(2):261266, 2005.
[181] Shunsuke Yatabe. Distinguishing non-standard natural numbers in a set theory within ukasiewicz
logic. Archive for Mathematical Logic, 46(34):281287, 2007.
LIBOR B EHOUNEK, PETR CINTULA, AND PETR H AJEK
Institute of Computer Science
Academy of Sciences of the Czech Republic
Chapter I: Introduction to Mathematical Fuzzy Logic 101
Pod Vod arenskou v e z 2
182 07 Prague 8, Czech Republic
Email: behounek, cintula, hajek@cs.cas.cz
Chapter II
A General Framework for Mathematical Fuzzy
Logic
PETR CINTULA AND CARLES NOGUERA
1 Introduction
Mathematical Fuzzy Logic was born in the last decade of the XXth century as a sys-
tematical study of a particular kind of systems of non-classical many-valued logic with
the works of Baaz, Cignoli, Esteva, Godo, Gottwald, H ajek, Montagna, Mundici, Nov ak
and others (see e.g. [3, 14, 4952, 57, 58, 69, 72]). Because of their motivation in the
theory of fuzzy sets, the rst studied systems were those that admitted a semantics based
on particular well-known continuous t-norms: ukasiewicz, product and minimum t-
norm, which respectively corresponded to ukasiewicz, Product and G odel-Dummett
many-valued logics. The rst comprehensive attempt at systematization of the studies
on these logics was H ajeks celebrated monograph [53] published in 1998. This book
studied both propositional and rst-order formalisms for these logics and set the agenda
for the area by considering all the usual issues in Mathematical Logic for these specic
systems, including their algebraic semantics, proof theory, decidability and computa-
tional aspects, and applications. Moreover, in order to provide a common ground for the
three aforementioned systems, the monograph presented Basic fuzzy Logic BL, which
was conjectured (and later proved [15]) to be complete with respect to the semantics of
all continuous t-norms, and was hence a common base logic that could be axiomatically
extended to the three of them.
The main outcome of the monograph was that, by setting solid logical foundations,
it gave rise to a ourishing new eld of study, as witnessed by the prolic literature since
1998 (surveyed, though not exhaustively, by Chapter I of this Handbook), in which an
increasing number of researchers have contributed by proposing a growing collection
of systems of fuzzy logics obtained by modifying the dening conditions of BL and its
three main extensions. For instance, the divisibility condition of BL was removed in the
logic MTL [33] which is complete with respect to the semantics of all left-continuous
t-norms [64], many axiomatic extensions of MTL were studied (see e.g. [31, 71]), nega-
tion was removed when considering fuzzy logics based on hoops [34], commutativity
of t-norms was disregarded in [54], and t-norms were replaced by uninorms in [68].
On the other hand, logics with a higher expressive power were introduced by consid-
ering expanded real-valued algebras (with projection , involution , truth-constants,
etc., see e.g. [17, 32, 35, 36, 38]). Coherently with their initial motivations, the propo-
104 Petr Cintula and Carles Noguera
nents of all these systems have always borne in mind an intended (so called standard)
semantics based on real-valued algebras, and tried to show soundness and complete-
ness of the logics with respect to them. However, in recent works fuzzy logics have
started emancipating from the real-valued algebras as the only intended semantics by
considering systems complete with respect to rational, nite or hyperreal linearly or-
dered algebras [19, 31, 37, 39, 70].
When dealing with this huge variety of fuzzy logics, and in order to avoid a useless
repetition of analogous results and proofs, one may want to have some tools to prove
general results that apply not only to a particular logic, but to a whole class of logics. To
some extent this has been achieved by means of the notions of core and -core fuzzy
logics from [56] that have provided a useful framework for some papers such as [19, 70].
However, those classes contain only axiomatic expansions of MTL and MTL
Z
logics,
so they do not cover the aforementioned weaker systems. Therefore, we need to look
for a more general framework able to cope with all known examples and with other new
logics that may arise in the near future.
In doing so, one certainly needs some intuition about the class of objects one would
like to mathematically determine, namely some intuition of what are the minimal prop-
erties that should be required for a logic to be fuzzy. The evolution outlined above shows
that almost no property of these systems is essential as they have been step-by-step dis-
regarded. Nevertheless, there is one that has remained untouched so far: completeness
with respect to a semantics based on linearly ordered algebras. It actually corresponds
to the main thesis of [6] that defends that fuzzy logics are the logics of chains. Such a
claim must be read as a methodological statement, pointing at a roughly dened class of
logics, rather than a precise mathematical description of what fuzzy logics are (or should
be), for there could be many different ways in which a logic might enjoy a complete se-
mantics based on chains.
On the other hand, Algebraic Logic is the branch of Mathematical Logic that studies
logical systems by giving them a semantics based on some particular kind of algebraic
structures. The development we have just outlined shows how Algebraic Logic has been
fruitfully applied to fuzzy logics, and it has also been very useful in many other families
of non-classical logics. Moreover, in the last decades, it has evolved to a more abstract
discipline, Abstract Algebraic Logic, which aims at understanding the various ways in
which a logical system can be endowed with an algebraic semantics and developing
methods and results to deal with broad classes of those systems (see the survey [41] or
the comprehensive monographs [8, 25, 40, 88]). Therefore, it is a reasonable candidate
to provide the general framework we were looking for.
The aim of this chapter is to present a marriage of Mathematical Fuzzy Logic and
(Abstract) Algebraic Logic. In other words, we want to use the notions and techniques
from the latter to create a new framework where we can develop in a natural way a par-
ticular technical notion corresponding to the intuition of fuzzy logics as the logics of
chains. Since the order relation in the algebraic counterparts of fuzzy logics is typically
determined by an implication connective, we will present our framework in the context
of weakly implicative logics (rstly introduced in [18]) which generalize the well-known
class of implicative logics studied by Rasiowa in [76]. These logics enjoy an implica-
tion connective such that for any algebra Ain the algebraic semantics one can dene
Chapter II: A General Framework for Mathematical Fuzzy Logic 105
an order relation by setting for any pair of elements a, b in A: a b iff a
A
b F,
where F is the subset of designated elements of the algebra representing truth (in typical
examples F = 1
A
or F = a A [ a
A
1
A
= 1
A
). This allows to characterize
fuzzy logics, in this context, as those which are complete with respect to the class of al-
gebras where implication denes a linear ordering or, equivalently, as those logics whose
nitely subdirectly irreducible algebraic models are linearly ordered by the implication.
We call them weakly implicative semilinear logics inspired by the tradition in Universal
Algebra of calling a class of algebras semiX whenever that their sudirectly irreducible
members are X. We choose the term semilinear instead of fuzzy in spite of the fact
that a rst step towards the general denition we are offering here had been done by the
rst author in [18], when he introduced the class under the name weakly implicative fuzzy
logics, because the term fuzzy is probably too heavily charged with many conicting
potential meanings. It needs to be stressed that by this mathematical denition we do not
expect to capture the whole intuitive notion of arbitrary fuzzy logic. Even if we would
agree that the linear ordering in the semantics is crucial for a formal logic to be fuzzy,
there might still be several other ways in which a logic might have a complete semantics
somehow based on chains (see e.g. [9, 10]). But still, the notion of weakly implicative
semilinear logic will be able to include (and provide a useful mathematical framework
for) almost all the prominent examples of fuzzy logics known so far and exclude non-
classical logics which are usually not recognized as fuzzy logics in the Logic community.
The chapter is structured as follows. In Section 2 we introduce the necessary notions
from (Abstract) Algebraic Logic, the denition of weakly implicative logic and some
renements thereof and provide three increasingly stronger completeness theorems for
them. Moreover, we present a very general notion of substructural logics as a particular
family of weakly implicative logics, discuss their syntactical properties and deduction
theorems, and we conclude with a rather general study of disjunction connectives. Sec-
tion 3 presents and studies the main notion of this chapter: semilinearity. It characterizes
semilinear logics in terms of properties of lters and properties of disjunctions and gives
methods to axiomatize semilinear logics. Section 4 studies rst-order predicate systems
built over weakly implicative semilinear logics. It gives axiomatizations, completeness
theorems, and a general process of Skolemization. We conclude with Section 5 provid-
ing historical remarks to understand the genesis of the ideas and results presented in this
chapter and many bibliographical references for further studies in related topics.
2 Weakly implicative logics
This section provides the general basis for the framework presented in this chapter.
Subsection 2.1 gives the most elementary necessary syntactical and semantical notions
and proves completeness of all logics with respect to the class of their models. Sub-
section 2.2 introduces the notion of weakly implicative logics and proves their com-
pleteness with respect to the class of their reduced models. Subsection 2.3 introduces
other semantical notions, including relatively subdirectly irreducible models (RSI), and
proves completeness of weakly implicative logics with respect to the class of their RSI
reduced models. Subsection 2.4 studies the class of algebraically implicative logics, i.e.
those weakly implicative logics enjoying a stronger link with their algebraic semantics.
106 Petr Cintula and Carles Noguera
Subsection 2.5 studies particular kind of algebraically implicative logics: a wide class
of substructural logics based on the non-associative version of Full Lambek logic. Sub-
section 2.6 proves local forms of deduction theorems for associative substructural logics
and shows their relation with proof by cases properties. Finally, Subsection 2.7 consid-
ers proof by cases properties in the framework of a general notion of disjunction and
gives some characterizations that will be very useful in the rest of the chapter.
2.1 Basic notions and a rst completeness theorem
In this preliminary subsection we give the most basic syntactic and semantic notions
we need for a general framework to study propositional logics and we prove a rst
completeness theorem for them.
DEFINITION 2.1.1 (Language). A propositional language / is a countable type, i.e.
a function ar : C
1
N, where C
1
is a countable set of symbols called connectives,
giving for each one its arity. Nullary connectives are also called truth-constants. We
write c, n / whenever c C
1
and ar(c) = n.
The restriction to countable languages is necessary for very few results and simpli-
es the formulation of many others. The same holds for the following restriction of the
cardinality of the set of propositional variables. Note that, in particular, all the notions
and results of this subsection do not rely on these restrictions.
DEFINITION 2.1.2 (Formula). Let Var be a xed innite countable set of symbols
called (propositional) variables. The set Fm
1
of (propositional) formulae in a propo-
sitional language / is the least set containing Var and closed under connectives of /,
i.e., for each c, n / and every
1
, . . . ,
n
Fm
1
, c(
1
, . . . ,
n
) is a formula.
In what follows, beginning with the next denition, it will be convenient to identify
Fm
1
with the domain of the absolutely free algebra Fm
1
of type / and generators
Var.
1
The variables will be usually denoted by lower case Latin letters p, q, r, . . . . The
formulae will be usually denoted by lower-case Greek letters , , , . . . and their sets
by upper-case ones , , , . . . . The set of all sequences (including innite sequences)
of formulae is denoted by Fm

1
.
DEFINITION 2.1.3 (Substitution). Let / be a propositional language. An /-substitut-
ion is an endomorphism on the algebra Fm
1
, i.e. a mapping : Fm
1
Fm
1
, such
that (c(
1
, . . . ,
n
)) = c((
1
), . . . , (
n
)) holds for each c, n / and every

1
, . . . ,
n
Fm
1
.
Since an /-substitution is a mapping whose domain is a free /-algebra, it is fully
determined by its values on the generators (propositional variables).
DEFINITION 2.1.4 (Consecution). A consecution
2
in a propositional language / is a
pair , , where Fm
1
.
1
Recall that Fm
L
has the domain Fm
L
and operations: c
Fm
L
(
1
, . . . ,
n
) = c(
1
, . . . ,
n
).
2
The term consecution is taken from [1] (the term sequent is sometimes used instead).
Chapter II: A General Framework for Mathematical Fuzzy Logic 107
Instead of , we write . To simplify matters we will identify a formula
with the consecution of the form . Clearly, each subset A of the set of all
consecutions can be understood as a relation between sets of formulae and formulae.
We will use an inx notation and write

instead of A.
DEFINITION 2.1.5 (Logic). Let / be a propositional language. A set L of consecutions
in /is called a logic in the language /when it satises the following conditions for each
Fm
1
:
If , then
L
(Reexivity)
If
L
for each and
L
, then
L
(Cut)
If
L
, then []
L
() for each /-substitution (Structurality)
A logic L is called inconsistent if L is the set of all consecutions.
Observe that reexivity implies that any logic is non-empty and together with cut it
entails the following monotonicity condition:
If
L
and , then
L
(Monotonicity)
Notice the difference between (denoting an object) and
L
(stating
the fact L). When the language or logic are known from the context we omit
the parameters L or /; the same convention will be followed in any other case indexed
by L or /. Moreover, instead of , , and we respectively
write just , , , , and . Finally, we write instead of
for each and instead of and . The formulae such that
are called theorems of the logic.
It is easy to observe that the intersection of an arbitrary class of logics in the same
language is a logic as well. Let us introduce the notion of theory. The importance of this
notion will become apparent later when we introduce Lindenbaum matrices.
DEFINITION 2.1.6 (Theory). A theory of a logic L is a set of formulae T such that if
T
L
then T. By Th(L) we denote the set of all theories of L.
Theories are sometimes also called deductively closed sets of formulae and will be
usually denoted by upper case Latin letters T, S, R, . . . . Notice that for each set of
formulae, the set Th
L
() = [
L
belongs to Th(L). Observe that Th
L
() is
the least theory containing ; we call it the theory generated by . Note that the set of
theorems of L equals to Th
L
() and thus it is a subset of any theory T of the logic L.
Now we introduce the notion of axiomatic system as the same kind of objects as log-
ics, i.e. sets of consecutions closed under substitutions; this will simplify the formulation
of some upcoming results.
DEFINITION 2.1.7 (Axiomatic system). Let / be a propositional language. An ax-
iomatic system /o in the language / is a set /o of consecutions closed under arbi-
trary substitutions. The elements of /o of the form are called axioms if = ,
nitary deduction rules if is a nite set, and innitary deduction rules otherwise. An
axiomatic system is said to be nitary if all its deduction rules are nitary.
108 Petr Cintula and Carles Noguera
Notice that the convention we have made above identifying the consecution
with the formula , allows to call an axiom of the axiomatic system. Of course,
each axiomatic system can also be seen as a collection of schemata (by a schema we
understand a consecution and all its substitution instances).
DEFINITION 2.1.8 (Proof). Let / be a propositional language and /o an axiomatic
system in /. A proof of a formula from a set of formulae in /o is a well-founded
tree (with no innitely-long branch) labeled by formulae such that
its root is labeled by and leaves by axioms of /o or elements of ,
if a node is labeled by and ,= is the set of labels of its preceding nodes,
then /o.
We write
,S
if there is a proof of from in /o.
Observe that formal proofs can be seen as well-founded relations (with leaves as
minimal elements and the root as a maximum), thus we can prove facts about formu-
lae by induction over the complexity of their formal proofs. Notice that a deduction
rule
1
,
2
, . . . gives a way to construct a proof of from if we know the
proofs of
1
,
2
, . . . from : we just glue them together in a single tree using the rule

1
,
2
, . . . . In contrast, the meta-rule: from
1
,
2
, . . . obtain
only tells us that if there are proofs of
1
,
2
, . . . from , , . . ., then there is a proof of
from as well, though it gives no hint to its construction. We could say that rules are
inferences between formulae, whereas meta-rules are in fact inferences between conse-
cutions. We will see prominent examples of meta-rules at the end of this section and in
the next one.
LEMMA 2.1.9. Let / be a propositional language and /o an axiomatic system in /.
Then
,S
is the least logic containing /o.
Proof. Obviously
,S
is a logic and /o
,S
. We prove that for each logic L, if
/o L, then
,S
L. Assume that
,S
, i.e. there is a proof of from . By
induction over the complexity of the proof we can show that for each formula which
labels some node in the proof we have
L
, and hence in particular
L
.
DEFINITION 2.1.10 (Presentation, nitary logic). Let / be a propositional language,
/o an axiomatic system in /, and L a logic in /. We say that /o is an axiomatic
system for (or a presentation of) the logic L if L =
,S
. A logic is said to be nitary if
it has some nitary presentation.
Observe that each logic has a presentation, for L understood as an axiomatic system
is a presentation of the logic L itself (due to Lemma 2.1.9). Next we show that our
denition of nitary logics is equivalent to the usual one:
LEMMA 2.1.11. Let L be a logic. Then L is nitary iff for each set of formulae
we have: if
L
, then there is a nite
t
such that
t

L
.
Chapter II: A General Framework for Mathematical Fuzzy Logic 109
Proof. Assume that L is nitary. Then, by denition, it has a nitary presentation /o.
Observe that proofs in a nitary axiomatic system are always nite (because by deni-
tion the tree has no innite branches and, because of nitarity, each node has nitely
many preceding nodes, thus by K onigs Lemma the tree is nite). This gives the impli-
cation from left to right. The reverse direction is straightforward.
Observe that in the nitary case we can represent the tree as a linear sequence of
formulae, obtaining thus the usual notion of nite proof.
DEFINITION 2.1.12 (Finitary companion). The nitary companion of a logic L is the
logic T((L) dened as:
JC(L)
iff there is a nite subset
0
such that
0

L
.
Note that the nitary companion of a logic L is the strongest nitary logic contained
in L and it is naturally axiomatized by the set of all nitary consecutions provable in L.
DEFINITION 2.1.13 (Expansion). Let /
1
/
2
be propositional languages, L
i
a logic
in /
i
, and o a set of consecutions in /
2
. We say that L
2
is
the expansion of L
1
by o if it is the weakest logic in /
2
containing L
1
and o, i.e.
the logic axiomatized by all /
2
-substitutional instances of consecutions from o
added to any presentation of L
1
,
an expansion of L
1
if L
1
L
2
, i.e. it is the expansion of L
1
by o, for some set of
consecutions o,
an axiomatic expansion of L
1
if it is an expansion obtained by adding a set of
formulae,
a conservative expansion of L
1
if it is an expansion and for each consecution
in /
1
we have that
L
2
entails
L
1
.
If /
1
= /
2
, we use extension instead expansion.
3
Next we introduce the necessary basic semantical notions. Let us x a propositional
language /. The logics in this language are given a semantical interpretation by means
of the notion of logical matrix, which is a pair formed by an /-algebra (which interprets
the formulae capitalizing on the fact that / can also be seen as an algebraic language)
and a lter (a subset of designated elements in the domain of the algebra which gives a
notion of truth for the logic):
DEFINITION 2.1.14 (Logical matrix). An /-matrix is a pair A = A, F where A is
an /-algebra called the algebraic reduct of A, and F is a subset of A called the lter
of A. The elements of F are called designated elements of A.
A matrix is trivial if F = A. A matrix is nite if its underlying algebra has a nite
domain. The matrices where A = Fm
1
are called Lindenbaum matrices.
3
Observe that any conservative extension of any logic is just the logic itself.
110 Petr Cintula and Carles Noguera
DEFINITION 2.1.15 (Evaluation). Let A be an /-algebra. An A-evaluation is a ho-
momorphism from Fm
1
to A, i.e. a mapping e: Fm
1
A, such that for each
c, n / and each n-tuple of formulae
1
, . . . ,
n
we have: e(c(
1
, . . . ,
n
)) =
c
A
(e(
1
), . . . , e(
n
)).
As in the case of substitutions, since an A-evaluation is a mapping whose domain
is a free /-algebra, it is fully determined by its values on the generators (proposi-
tional variables). By e[pa] we denote the evaluation obtained from e by assigning
the element a A to the variable p and leaving the values of remaining variables un-
changed. For a formula build from variables p
1
, . . . , p
n
, an algebra A, elements
a
1
, . . . , a
n
A and an A-evaluation e such that e(p
i
) = a
i
, we write
A
(a
1
, . . . , a
n
)
instead of e((p
1
, . . . , p
n
)). Given a matrix A = A, F and A-evaluation e, we will
also call e an A-evaluation.
DEFINITION 2.1.16 (Semantical consequence). A formula is a semantical conse-
quence of a set of formulae w.r.t. a class K of /-matrices if for each A, F K and
each A-evaluation e, we have e() F whenever e[] F; we denote it by [=
K
.
We write [=
A
instead of [=
A]
. Obviously, [=
K
is a set a of consecutions, but more
can be proved (the second claim will be generalized in Proposition 2.3.13):
LEMMA 2.1.17. Let K a class of /-matrices. Then [=
K
is a logic in /. Furthermore if
K is a nite class of nite matrices, then the logic [=
K
is nitary.
Proof. We need to check the three properties in the denition of logic. The rst one
is obvious. To show the second one x A, F K and an A-evaluation e such that
e[] F. Then clearly e() F for each , i.e. e[] F, and so e() F.
The nal condition: x A, F and e as before and assume that e([]) F. Since
e
t
= e is an A-evaluation and e
t
[] F, we obtain e(()) = e
t
() F.
The second claim: if we prove it for K = A, F the proof is done by observing
that: (1) [=
KL
= [=
K
[=
L
, and (2) the intersection of two nitary logics is nitary.
Assume that
t
,[=
K
for each nite
t
and we want to show that ,[=
K
.
Let us consider the nite set A endowed with discrete topology and its power A
Var
with product (=weak) topology. Both spaces are compact (the rst one trivially and
the second one due to Tychonoff theorem). Clearly each evaluation e can be identied
with an element of A
Var
and vice versa. For each formula we dene a mapping
H

: A
Var
A as H

(e) = e(). It can be easily shown that these mappings are

continuous, thus (H

)
1
[F] is a closed set and so is the set (H

)
1
[F](H

)
1
[AF]
(i.e. the set of evaluations with satises the formula but not the formula ). Let us
now consider the system of closed sets (H

)
1
(H

)
1
[A F]) [ . This is
clearly a centered system (the intersection of any nite subsystem given by a set
t
is
non-empty, because it contains any evaluation which witnesses that
t
,[=
K
). Thus,
due to the compactness of A
Var
, the intersection of the whole system is non-empty and
the proof is done (because any element of this intersection is an evaluation satisfying the
set but not the formula ).
DEFINITION 2.1.18 (L-matrix). Let L be a logic in / and Aan /-matrix. We say that
Ais an L-matrix if L [=
A
. We denote the class of L-matrices by MOD(L).
Chapter II: A General Framework for Mathematical Fuzzy Logic 111
Observe that for each presentation /o of a logic L we have: A MOD(L) iff
/o [=
A
(one direction is obvious, the second one is Lemma 2.1.9).
LEMMA 2.1.19. Let L be a logic in / and a mapping g : A B be a homomorphism
of /-algebras A, B. Then:
A, g
1
[G] MOD(L), whenever B, G MOD(L),
B, g[F] MOD(L), whenever A, F MOD(L) and g is surjective and
g(x) g[F] implies x F.
Proof. The rst claim is straightforward: assume that
L
and e[] g
1
[G] for
some A-evaluation e. Thus, g[e[]] G which, since g e is a B-evaluation and
B, G MOD(L), implies that g(e()) G, i.e. e() g
1
[G].
The second claim: assume that
L
and for a B-evaluation f it is the case that
f[] g[F]. Let us dene an A-evaluation e by setting e(v) = a for some a such that
g(a) = f(v) (such a has to exist because g is surjective). Next we show by induction
that f() = g(e()). The base is trivial. Let us assume that = c(
1
, . . . ,
n
). Then:
f(c(
1
, . . . ,
n
)) = c
B
(f(
1
), . . . , f(
n
))
= c
B
(g(e(
1
)), . . . , g(e(
n
)))
= g(c
A
(e(
1
), . . . , e(
n
)))
= g(e(c(
1
, . . . ,
n
)).
From g[e[]] = f[] g(F) we obtain e[] F. Thus e() F and so nally
f() = g(e()) g[F].
DEFINITION 2.1.20 (Logical lter). Given a logic L in /and an /-algebra A, a subset
F Ais an L-lter if A, F MOD(L). By Ti
L
(A) we denote the set of all L-lters
over A.
Observe that A Ti
L
(A) and Ti
L
(A) is closed under arbitrary intersections. This
means that Ti
L
(A) is a closure system (we will deal with them in detail in Subsec-
tion 2.3) which allows us to endow it with a (complete) lattice structure:
DEFINITION 2.1.21 (Generated lters and lattice of logical lters). Let L be a logic
in / and A an /-algebra. Given a set X A, the logical lter generated by X is
Fi
A
L
(X) =

F Ti
L
(A) [ X F. Ti
L
(A) is given a lattice structure by dening
for any F, G Ti
L
(A), F G = F G and F G = Fi
A
L
(F G).
Moreover, some results in Subsection 2.7 will need the following denition:
DEFINITION 2.1.22 (Filter-distributivity). A logic L is lter-distributive if for each
/-algebra, the lattice Ti
L
(A) is distributive.
The elements of a lter generated by a set are characterized in the next proposition
by means of the notion of proof in algebra. It consists in generalizing to any algebra the
notion of proof introduced in Denition 2.1.8 for the algebra of formulae.
112 Petr Cintula and Carles Noguera
PROPOSITION 2.1.23 (Proof in algebra). Let L be a logic, /o one of its presenta-
tions, A an /-algebra, and X a A. Let us dene a set V
,S
T(A) A as
e[], e() [ e is an A-evaluation and /o.
4
Then a Fi
A
(X) iff there is
a well-founded tree (called proof of a from X) labeled by elements of A such that
its root is labeled by a, and leaves are labeled by elements x such that x X
or , x V
,S
,
if a node is labeled by x and Z ,= is the set of labels of its preceding nodes,
then Z, x V
,S
.
Proof. Let D(X) be the set of elements of A for which there exists a proof from X.
We can easily show that /o [=
A,D(X))
. Indeed, assume that /o and
h[] D(X) for some evaluation h. Then for each x h[] there is a proof from X
and, since h[], h[] V
,S
, we can connect these proofs so that they will forma proof
of h(). Thus D(X) Ti
L
(A) and, since X D(X), we obtain Fi(X) D(X). To
prove the converse direction consider x D(X) and notice that for each y appearing
in its proof we can easily prove inductively that y Fi(X) (because Fi(X) is closed
under all the rules of L, in particular those in /o).
Next we show that the lters of Lindenbaum matrices can be nicely characterized.
PROPOSITION 2.1.24. For any logic L in a language /, Ti
L
(Fm
L
) = Th(L).
Proof. Let Ti
L
(Fm
L
), i.e. if
L
then for each Fm
1
-evaluation e we have
e() whenever e[] . Therefore, in the particular case where the evaluation e
is the identity and = , we obtain Th(L).
Next assume that T Th(L),
L
and e is an Fm
1
-evaluation such that
e[] T, thus also T
L
e[]. By structurality, also e[]
L
e(), and thus also
T
L
e(). Since T is a theory, we have e() T.
We close the subsection observing that the notions introduced so far are enough to
obtain a rst completeness theorem for any logic.
THEOREM 2.1.25 (Completeness w.r.t. all models). Let L be a logic. Then for each set
of formulae and each formula the following holds:
L
iff [=
MOD(L)
.
Proof. Soundness is obvious. For the reverse direction assume that -
L
and dene
T = Th
L
(). We know that Fm
1
, T MOD(L) and then the identity mapping is
the Fm
1
, T-evaluation we need to show that ,[=
MOD(L)
.
2.2 Weakly implicative logics and a second completeness theorem
In this subsection we rst introduce the main dening notion for the framework of
this chapter: the class of weakly implicative logics. Then we use the notions of Leibniz
congruence and reduced model to prove a second completeness theorem. Although
these notions can be introduced in general for any propositional logic and completeness
with respect to its reduced models can be proved in general, we will restrict to weakly
implicative logics for the sake of simplicity.
4
Note that if A = Fm
L
, then V
AS
= ,S.
Chapter II: A General Framework for Mathematical Fuzzy Logic 113
DEFINITION 2.2.1 (Weakly implicative logic). Let L be a logic in a language /. We
say that L is a weakly implicative logic if there is a binary connective (primitive or
denable by a formula of two variables in language /) such that:
(R)
L

(MP) ,
L

(T) ,
L

(sCng) ,
L
c(
1
, . . . ,
i
, , . . . ,
n
) c(
1
, . . . ,
i
, , . . . ,
n
)
for each c, n / and each 0 i < n.
The acronyms respectively stand for reexivity, modus ponens, transitivity and
symmetrized congruence. The connective is called a weak implication of L. There
could be, in principle, several different weak implications in a given logic. For the sake
of simpler notation we will avoid indexing many of the upcoming notions with a weak
implication by assuming from now on that each language comes with a xed binary
(primitive or derivable) connective , such that if a logic in this language is weakly
implicative, then is one of its weak implications and all notions are dened w.r.t. this
particular implication. We will call the principal implication of the logic. In some
rare cases, when we may need to speak at once about notions corresponding to different
weak implications, we will index these notions by their corresponding weak implication
to avoid any confusion.
EXAMPLE 2.2.2. In classical logic the usual connectives of implication and equiva-
lence, and , are both actually weak implications in our sense, but observe they have
a very different logical behavior (for instance, only the former satises ).
More generally, all connectives in the various logics mentioned in Chapter I are weak
implications. Therefore, all these logics are examples of weakly implicative logics.
Now we consider the properties of the symmetrization of a weak implication in
a logic L. Given a pair of formulae , , we use the expression to denote the
set of formulae , (recall that, according to previous conventions, by

L
we mean that
L
and
L
). We can easily show that
behaves like a congruence.
THEOREM 2.2.3 (Congruence Property). Let L be a weakly implicative logic and
, , formulae. Then:

L
,

L
,
,
L
,

L
,
where is obtained from by replacing some occurrences of in by .
By using the last part for =
t
we obtain an important corollary:
114 Petr Cintula and Carles Noguera
COROLLARY 2.2.4. Let and
t
be two weak implications in a logic L. Then:

L

t
.
Therefore, if we had two different weak implications in a logic, their symmetriza-
tions would behave exactly in the same way as far as provability is concerned. Now,
aiming to obtain a ner complete semantics for weakly implicative logics, we introduce
some further semantic notions.
DEFINITION 2.2.5 (Leibniz congruence). Let A = A, F be an L-matrix for a weakly
implicative logic L. The matrix preorder
A
of Ais dened as a
A
b iff a
A
b F.
Further we dene the Leibniz congruence
A
(F) of A as a, b
A
(F) iff a
A
b
and b
A
a.
DEFINITION 2.2.6 (Logical congruence). A logical congruence in a matrix A, F is
a congruence of A compatible with F, i.e. such that for each a, b A if a F and
a, b , then b F.
THEOREM 2.2.7 (Characterization of Leibniz congruence). Let L be a weakly implica-
tive logic and A = A, F MOD(L). Then:

A
is a preorder,

A
(F) is the largest logical congruence of A, and
a, b
A
(F) if, and only if, for each formula and each A-evaluation e it is
the case that e[pa]() F iff e[pb]() F.
Proof. The fact that
A
is a preorder follows from(R) and (T).
A
(F) is a congruence
because of (sCng), and it is logical because of (MP). To see that it is the largest one,
assume that is a logical congruence of A and a, b . Since a, a , we have
a
A
a, a
A
b , and a
A
a F hence, by compatibility a
A
b F;
analogously b
A
a F, hence a, b
A
(F), and so we have
A
(F).
Final claim: one direction is a straightforward corollary of (sCng). The converse
direction: consider the formula p q and the evaluation e(q) = b. Then we obtain that
a
A
b F iff b
A
b F. Thus a
A
b and, since using the evaluation e(q) = a
we can prove b
A
a, the proof is done.
DEFINITION 2.2.8 (Reduced matrix, MOD

(L), ALG

(L)). Let L be a weakly im-

plicative logic. An L-matrix A = A, F is said to be reduced if
A
(F) is the identity
relation Id
A
. The class of all reduced models of L is denoted by MOD

(L), and the

class of algebraic reducts of matrices of MOD

(L) is denoted by ALG

(L). The
members of ALG

(L) are called L-algebras.

Observe that a reduced model of a logic is non-trivial if, and only if, its algebraic
reduct has more than one element. We could alternatively dene reduced matrices as
those whose matrix preorder is an order. The next lemma shows us how to turn any
model into a reduced one. Let A = A, F be a matrix, we use the following notation:
[a]
F
= b A [ a, b
A
(F), [F] = [a]
F
[ a F, and A

= A/
A
(F), [F].
Chapter II: A General Framework for Mathematical Fuzzy Logic 115
LEMMA 2.2.9. Let L be a weakly implicative logic and A = A, F MOD(L).
Then:
1. A

MOD(L).
2. [a]
F

A
[b]
F
iff a
A
b F, for every a, b A.
3. A

MOD

(L).
Proof. 1. Clearly []
F
is a surjective homomorphism from A onto A/
A
(F). By
Lemma 2.1.19, all we need to show is: [a]
F
[F] implies a F. The assumption
gives us [a]
F
= [b]
F
for some b F. Then a, b
A
(F) and, since
A
(F) is
a logical congruence, we obtain a F.
2. [a]
F

A
[b]
F
iff [a]
F

A/
A
(F)
[b]
F
[F] iff [a
A
b]
F
[F] iff a
A
b F.
3. [a]
F

A
[b]
F
and [b]
F

A
[a]
F
entail a, b
A
(F) and so [a]
F
= [b]
F
.
DEFINITION 2.2.10 (Leibniz operator). Let L be a weakly implicative logic in a lan-
guage /, and A be an /-algebra. The Leibniz operator associated to A is the function
giving for each F Ti
L
(A) the Leibniz congruence
A
(F).
PROPOSITION 2.2.11. Let L be a weakly implicative logic L and A an /-algebra.
Then
1.
A
is monotone (i.e. if F G then
A
(F)
A
(G)),
2.
A
commutes with inverse images by homomorphisms, that is, for every
/-algebra B, every homomorphism h: A B and every F Ti
L
(B),

A
(h
1
[F]) = h
1
[
B
(F)] = a, b [ h(a), h(b)
B
(F),
3.
A
[Ti
L
(A)] = Con
ALG

(L)
(A), where by Con
ALG

(L)
(A) we denote the
set ordered by inclusion
5
of congruences of Agiving a quotient in ALG

(L).
Proof. The proofs of the rst two claims are easy (by Lemma 2.1.19 h
1
[F] is indeed a
lter on A). To prove the third rst one observe that
A
[Ti
L
(A)] Con
ALG

(L)
(A)
(due to the Lemma 2.2.9). To show the second direction assume Con
ALG

(L)
(A).
We know that A/ ALG

(L), i.e. A/, F

0
MOD

(L) for some lter F

0
. Let
k be the canonical mapping from A onto A/, we dene F = k
1
[F
0
] and, again by
Lemma 2.1.19, we know that F Ti
L
(A). To complete the proof just observe that

A
(F) =
A
(k
1
[F
0
]) = k
1
[
A/
(F
0
)] = k
1
[Id
A/
] = .
Next we introduce the well-known notion of Lindenbaum-Tarski matrices, in the
traditional way as it is usually done in the literature, and show how they are related to
reduced matrices.
DEFINITION 2.2.12 (Lindenbaum-Tarski matrix). Let L be a weakly implicative logic
in / and T Th(L). For every formula , we dene the set
[]
T
= Fm
1
[ T.
5
Later, after Proposition 2.3.13, we show that Con
ALG

(L)
(A) is actually a lattice.
116 Petr Cintula and Carles Noguera
The Lindenbaum-Tarski matrix with respect to the logic L and the theory T, LindT
T
, is
the /-matrix where the designated set is []
T
[ T, and the /-algebra has the do-
main []
T
[ Fm
1
and operations c
LindT
T
([
1
]
T
, . . . [
n
]
T
) = [c(
1
, . . .
n
)]
T
.
Clearly, for every T Th(L), the matrix LindT
T
coincides with Fm
1
, T

.
Now we are ready to prove the main result of this subsection.
THEOREM 2.2.13 (Completeness w.r.t. reduced models). Let L be a weakly implicative
logic. Then for any set of formulae and any formula the following holds:
L
iff
[=
MOD

(L)
.
Proof. Soundness is obvious. For the reverse direction, let T be the theory generated
by ; clearly LindT
T
MOD

(L) (Lemma 2.2.9). Consider a LindT

T
-evaluation e
dened as e() = []
T
and observe that e[] e[T] [T]. Thus from [=
MOD

(L)

we obtain that []
T
= e() [T] and thus T
L
and so nally
L
.
The proof shows how the theorem can be strengthened: every weakly implicative
logic is complete w.r.t. the class of Lindenbaum-Tarski matrices.
2.3 Advanced semantics and a third completeness theorem
In this subsection, after recalling some further knowledge about closure systems,
closure operators and logical matrices, we obtain a third completeness theorem for ni-
tary weakly implicative logics.
DEFINITION 2.3.1 (Closure system). A closure system over a set A is a collection
of subsets ( T(A) closed under arbitrary intersections and such that A (. The
elements of ( are called closed sets.
For example, we have seen that given a logic L and A, F MOD(L), Ti
L
(A)
is a closure system over A; in particular Th(L) is a closure system over Fm
1
.
DEFINITION 2.3.2 (Closure operator). Given a set A, a closure operator over A is a
mapping C: T(A) T(A) such that for every X, Y A:
1. X C(X),
2. C(X) = C(C(X)),
3. if X Y , then C(X) C(Y ).
Every closure operator C denes a closure system: X A [ C(X) = X. Con-
versely, given a closure system ( over A, one can dene a closure operator as follows:
C(X) =

Y ( [ X Y . This gives a one-to-one correspondence between closure

operators and systems. The closure operator associated to the closure systemTh(L) will
be denoted as Th
L
, analogously the one associated to Ti
L
(A) will be denoted as Fi
A
L
;
as usual, we omit the parameters when known from the context.
A closure operator C is nitary if for every X A, C(X) =

C(Y ) [ Y X
and Y is nite. A closure system ( is called inductive if it is closed under unions of
upwards directed families (i.e. families T ,= such that for every A, B T, there is
C T such that A B C).
Chapter II: A General Framework for Mathematical Fuzzy Logic 117
THEOREM 2.3.3 (Schmidt Theorem). A closure operator C is nitary if, and only if,
its associated closure system ( is inductive.
Proof. Assume that C is nitary and take an upwards directed family T (. It sufces
to show that C(

T)

T. Take any a C(

T). By nitarity, a C(a

1
, . . . , a
n
)
for some a
1
, . . . , a
n


T. Since T is upwards directed, there exists a T
0
T such
that a
1
, . . . , a
n
T
0
and, hence, a C(T
0
) = T
0
T, and so a

T. Conversely,
assume that ( is inductive, take any X A and consider the family T = C(F) [
F X nite. Since T is clearly upwards directed we have

T (, and therefore

T = C(X).
Note that the nitarity of a logic L is equivalent to the nitarity of the corresponding
closure operator Th
L
. The next corollary is the rst example we meet in this chapter of
the so-called transfer theorems: theorems which transfer a given property of a logic L
(understood as the closure operator/system over the set of formulae) to the analogous
property of closure operator/system of all L-lters over any algebra.
COROLLARY 2.3.4 (Transfer theorem for nitarity). Given a logic L in a language /,
the following conditions are equivalent:
1. L is nitary,
2. Fi
A
L
is a nitary closure operator for any /-algebra A,
3. Ti
L
(A) is an inductive closure system for any /-algebra A.
Proof. The equivalence of the last two claims is established by the previous theorem.
It is clear that 2 implies 1 by taking A = Fm
1
. Let us see that 1 implies 3. Take an
upwards directed family T Ti
L
(A) and dene F =

T. We need to show that

F Ti
L
(A). Assume that
L
and e is an A-evaluation such that e[] F.
Since L is nitary, there is a nite set
0
such that
0

L
. Then, since the
family is upwards directed, there has to be an F
0
T such that e[
0
] F
0
and so
e() F
0
F.
A useful notion in the theory of closure systems is that of base, which is a distin-
guished family of closed sets allowing to describe all closed sets of the system.
DEFINITION 2.3.5 (Base). A base of a closure system ( over A is any B ( satisfying
one of the following equivalent conditions:
1. ( is the nest closure system containing B,
2. for every T ( A, there is a T B such that T =

T,
3. for every T ( A, T =

B B [ T B,
4. for every Y ( and a A Y there is Z B such that Y Z and a / Z.
118 Petr Cintula and Carles Noguera
DEFINITION 2.3.6 (Maximal w.r.t. an element, saturated, and (nitely) -irreducible
closed sets). An element X of a closure system ( over A is called
maximal w.r.t. an element a if it is a maximal element of the set Y ( [ a / Y
w.r.t. the order given by inclusion,
saturated if it is maximal w.r.t. some element a,
(nitely) -irreducible if for each (nite non-empty) set } ( such that
X =

Y
Y , there is Y } such that X = Y .
Note that the set A is nitely -irreducible but is not -irreducible, because it is
the intersection of the empty set. Also observe that nite--irreducibility of X can be
equivalently dened by the following condition: for each Y
1
, Y
2
( such that X =
Y
1
Y
2
we have X = Y
1
or X = Y
2
.
PROPOSITION 2.3.7. Let ( be a closure system over a set A and T (. Then, T is
saturated if, and only if, T is -irreducible.
Proof. Assume that T is not -irreducible, i.e. there is a family T
i
[ i I ( such
that T =

iI
T
i
and T _ T
i
for every i I. Therefore, for every i I we can
choose b
i
T
i
T, and thus T _ C(T, b
i
) T
i
; this gives: T =

C(T, b
i
) [ i I
and hence T =

C(T, b) [ b / T. Assume, in search of a contradiction, that T is

maximal w.r.t. an element a A. Then for every b / T, we have T _ C(T, b), and the
maximality implies a C(T, b). Thus a

C(T, b) [ b / T = T; a contradiction.
Conversely, assume that T is -irreducible. Clearly, T _

C(T, b) [ b / T and
thus there is a

C(T, b) [ b / T T, which means that T is maximal w.r.t. a.
Indeed: if T
t
( and T _ T
t
, then there is b T
t
T, and thus a C(T, b) T
t
.
The next lemma allows to prove that in nitary closure systems, the -irreducible
sets always form a base. This will be later used, as a particular consequence, to obtain a
rened completeness theorem for weakly implicative logics.
LEMMA 2.3.8 (Abstract Lindenbaum Lemma). Let C be a nitary closure operator
and ( its corresponding closure system. If T ( and a / T, then there is T
t
( such
that T T
t
and T
t
is maximal with respect to a.
Proof. The proof is an easy application of Zorns Lemma. Observe that the set / =
S ( [ T S, a / S is clearly non-empty because T /. Take any chain
S
i
[ i I /. By Schmidt Theorem

iI
S
i
( and it is obvious that it contains
T and it does not have a as an element, hence

iI
S
i
/ and it is an upper bound
of the chain. By Zorns Lemma / has some maximal element T
t
which satises the
desired property.
COROLLARY 2.3.9. Let C be a nitary closure operator and ( its associated closure
system. Then the class of -irreducible (i.e. saturated) sets of ( forms a base of (.
On the other hand, it is also interesting to consider the maximal closed sets in a
closure system:
Chapter II: A General Framework for Mathematical Fuzzy Logic 119
DEFINITION 2.3.10 (Maximal closed set). Let ( be a closure system over A. A closed
set T ( A is called maximal or maximally consistent if it is a maximal element in
( A with respect to the order given by inclusion.
The following characterization is straightforward:
PROPOSITION 2.3.11. Let ( be a closure system over A and T ( A. The
following are equivalent:
1. T is maximally consistent,
2. T is maximal w.r.t. every a A T.
As another consequence of the abstract Lindenbaum Lemma one can show that
every closed set can be extended to a maximally consistent one:
PROPOSITION 2.3.12. Let ( be a nitary closure system over A with an inconsistent
element (i.e. an element a A such that C(a) = A). Then every T ( A can be
extended to a maximally consistent T
t
(.
Proof. Since a does not belong to T (otherwise we would have T = A), by the Lin-
denbaum Lemma, there is T
t
T maximal w.r.t. a. Then T
t
is actually maximally
consistent. Indeed, if T
t
_ T
tt
, then a T
tt
and thus T
tt
= A.
However, this last result does not entail that maximally consistent sets form a base.
Although it is well-known that this is the case in classical logic, it is not generally true.
For instance, the basic fuzzy logic BL introduced in the previous chapter provides a
counterexample.
Now we introduce some further necessary notions on matrix theory. Observe that
an /-matrix A, F can be regarded as a rst-order structure in the equality-free predi-
cate language with function symbols from / and a unique unary predicate symbol, with
domain A, function symbols interpreted as the operations of A, and the predicate inter-
preted by F. Fromthis perspective, one can dene the usual notions of substructure (now
called submatrix), homomorphism (if a F
1
, then h(a) F
2
), strict homomorphism
(a F
1
iff h(a) F
2
), isomorphism, direct product, reduced product and ultraproduct
for matrices. Given a class of matrices K, we will denote by S(K), H(K), H
S
(K),
I(K), P(K), P
R
(K) and P
U
(K) the closure of K under the mentioned operations. An-
other special operation on the classes of matrices we will need later is the operator of
reduced products over countably complete lters (i.e. lters closed under countable in-
tersections) which we will denote as P
-f
. Note that obviously P(K) P
-f
(K). It
should also be noted that a bijective matrix homomorphism is not necessarily an iso-
morphism (because its inverse need not be a matrix homomorphism). An embedding of
matrices is an injective strict homomorphism.
Fromthe results in [25] one can obtain the following properties about the behavior of
these operators on models and reduced models (observe that the third claim generalizes
Lemma 2.1.17):
120 Petr Cintula and Carles Noguera
PROPOSITION 2.3.13. Let L be a weakly implicative logic. Then:
1. SP(MOD(L)) MOD(L),
2. SP
-f
(MOD

(L)) MOD

(L),
3. If K MOD(L), P
U
I(K) I(K), and L = [=
K
, then L is nitary.
4. P
U
(MOD

(L)) MOD

(L) iff L is nitary.

As a consequence we obtain that, for every weakly implicative logic L, ALG

(L)
is closed under subalgebras and direct products; moreover for every /-algebra A the
set of relative congruences, Con
ALG

(L)
(A) is a complete lattice w.r.t. the inclusion
order (indeed, given a family A Con
ALG

(L)
(A) the quotient of Aby

A embeds
into the direct product of the quotients of A by the elements of A and hence, since
ALG

(A) is closed under S and P, we are done).

The notion of subdirect product from Universal Algebra is also generalized to ma-
trices. A matrix A is said to be representable as a subdirect product of the family of
matrices A
i
[ i I if there is an embedding homomorphism fromAinto the direct
product

iI
A
i
such that for every i I, the composition of with the i-th projection,

i
, is a surjective homomorphism. In this case, is called a subdirect representation,
and it is called nite if I is nite.
Let L be a logic and K MOD

(L). By P
SD
(K) we denote the closure of K
under subdirect products. Anon-trivial matrix A Kis (nitely) subdirectly irreducible
relative to Kif for every (nite non-empty) subdirect representation of Awith a family
A
i
[ i I K there is i I such that
i
is an isomorphism. The class of all
(nitely) subdirectly irreducible matrices relative to K is denoted as K
R(F)SI
. Of course
K
RSI
K
RFSI
. When K = MOD

(L) these classes are characterized in the following

way:
THEOREM2.3.14 (Characterization of RSI and RFSI reduced models). Given a weakly
implicative logic L and A = A, F MOD

(L), we have:
1. A MOD

(L)
RSI
if, and only if, F is -irreducible in Ti
L
(A).
2. A MOD

(L)
RFSI
if, and only if, F is nitely -irreducible in Ti
L
(A).
Proof. Let us rst solve the case when Ais a trivial reduced matrix, i.e., F = A = a.
Recall that in this case, F is nitely -irreducible but not -irreducible in Ti
L
(A).
Obviously, A MOD

(L)
RFSI
and A / MOD

(L)
RSI
, because the product of the
empty system of matrices is a trivial matrix.
We write only the proof for the rst claim (the second one is completely analo-
gous). Suppose that A MOD

(L)
RSI
and, in search of a contradiction, that F is not
-irreducible in Ti
L
(A), i.e. F =

iI
F
i
where F _ F
i
Ti
L
(A) for every i I.
We use these lters to dene reduced matrices A
i
= A, F
i

MOD

(L) and show

that : A

iI
A
i
, dened as (a) = [a]
F
i
[ i I, is a subdirect representa-
tion of A. The homomorphisms
i
are indeed surjective, is clearly is a strict
homomorphism, so it remains to show the injectivity of . Assume that a ,= b. Then,
Chapter II: A General Framework for Mathematical Fuzzy Logic 121
since A is reduced, we can (without loss of generality) assume that a
A
b / F and
so a
A
b / F
i
for some i I. Thus [a]
F
i
,= [b]
F
i
and so (a) ,= (b). Since
A MOD

(L)
RSI
, there must be j I such that
j
is an isomorphism. Assume
now that a F
j
, this implies
j
((a)) = [a]
F
j
[F
j
] and, since
j
is isomorphism
it is a also a strict homomorphism of A and A
j
and so a F, and hence F
j
= Fa
contradiction.
We prove the converse direction contrapositively: assume that A / MOD

(L)
RSI
,
i.e., there is a family of reduced models of the logic A
i
= A
i
, F
i
[ i I and a
subdirect representation : A

iI
A
i
where no projection gives an isomorphism.
This will allow us to dene a collection of lters giving a decomposition of F. Indeed,
take

F
i
= (
i
)
1
[F
i
] and so by Lemma 2.1.19

F
i
Ti
L
(A). Due to the strictness
of we have F =

iI

F
i
. If F =

F
j
for some j I, then
j
would be an
isomorphism, contradicting the hypothesis. Therefore F is not -irreducible in Ti
L
(A).
THEOREM2.3.15 (Subdirect representation). If L is a nitary weakly implicative logic,
then MOD

(L) = P
SD
(MOD

(L)
RSI
), thus in particular every matrix in MOD

(L)
is representable as a subdirect product of matrices in MOD

(L)
RSI
.
Proof. One inclusion is relatively easy: P
SD
(MOD

(L)
RSI
) SP(MOD

(L))
SP
-f
(MOD

(L)) MOD

(L), (the last inclusion is due to claim 2 in Propo-

sition 2.3.13, the others are trivial). To prove the converse inclusion consider A =
A, F MOD

(L). By Corollary 2.3.4, Fi

A
L
is nitary and, by Corollary 2.3.9,
there exists a family F
i
[ i I of -irreducible lters such that F =

iI
F
i
.
Take A
i
= A, F
i

. So we have a subdirect representation : A

iI
A
i
taking
(a) = [a]
i
[ i I for every a A. Because F
i
is -irreducible, Theorem 2.3.14
tells us that A, F
i

MOD

(L)
RSI
for every i I .
As a consequence of this theorem and Theorem 2.2.13 we obtain a third complete-
ness theorem (this time restricted to nitary logics):
THEOREM2.3.16 (Completeness w.r.t. RSI reduced models). Let L be a nitary weakly
implicative logic. Then
L
= [=
MOD

(L)
RSI
.
2.4 Algebraically implicative logics
In this subsection we consider the relation between weakly implicative logics and
the equational consequence on their corresponding classes of algebras. Let us x a
propositional language /.
DEFINITION 2.4.1 (Equation). An equation in the language / is a formal expression
of the form , where , Fm
1
.
DEFINITION 2.4.2 (Equational consequence). We say that an equation is a
consequence of a set of equations w.r.t. a class Kof /-algebras if for each A Kand
each A-evaluation e we have e() = e() whenever e() = e() for each ;
when this is the case, we denote it by [=
K
.
122 Petr Cintula and Carles Noguera
Given any weakly implicative logic L, the equational consequence given by the class
of L-algebras can be translated into the logic in the following way:
PROPOSITION 2.4.3. Let L be a weakly implicative logic and a set of
equations. Then [=
ALG

(L)
iff [
L
.
Proof. We show the proof of one implication, the proof of the converse one is similar.
Assume [=
ALG

(L)
. To check that [
L
it
is enough (due to the completeness theorem 2.2.13) to check the equivalent semantical
statement [ [=
MOD

(L)
. Take any A, F MOD

(L)
and an A-evaluation v satisfying the premises, i.e. for every we have
v()
A
v(), v()
A
v() F, and hence (since the matrix is reduced) v() =
v(). By the assumption (using that A ALG

(L)) we know that v() = v() and

thus v()
A
v(), v()
A
v() F.
However, if we want to obtain a better connection between the logic and the equa-
tional consequence enjoying also a translation from the former to the latter, we need to
restrict to a special subclass of weakly implicative logics.
DEFINITION 2.4.4 (Algebraically implicative logic). We say that a logic L is alge-
braically implicative if it is weakly implicative and there is a set of equations c in one
variable such that for each A = A, F MOD

(L) and each a A holds: a F

if, and only if,
A
(a) =
A
(a) for every c. In this case, ALG

(L) is called
the equivalent algebraic semantics of L.
For a set of formulae, by c[] we denote the set

c() [ of equations.
THEOREM 2.4.5 (Characterizations of algebraically implicative logics). Given any
weakly implicative logic L, the following are equivalent:
1. L is algebraically implicative.
2. There is a set of equations c in one variable such that
(Alg) p
L
(p) (p) [ c.
3. There is a set of equations c in one variable such that:
For every Fm
1
,
L
iff c[] [=
ALG

(L)
c(),
p q [=
ALG

(L)
c[p q] and c[p q] [=
ALG

(L)
p q.
4. For every /-algebra A, the Leibniz operator
A
is a lattice isomorphism from
Ti
L
(A) to Con
ALG

(L)
(A).
5. For every A, F MOD

(L), F is the least L-lter on A.

In the rst three items the sets c can be taken the same.
Chapter II: A General Framework for Mathematical Fuzzy Logic 123
Proof. First, we prove the equivalence of the rst three claims, then the equivalence of
the last two claims, and nally we prove the implications 14 and 42.
21: It follows immediately from the completeness theorem and the denition of alge-
braically implicative logic.
13: The rst condition again is easily checked by using the completeness theorem and
the denition of algebraically implicative logic. To prove p q [=
ALG

(L)
c[p q]
take A, F MOD

(L) and an evaluation e on A such that e(p) = e(q). Then

e(p)
A
e(q) F (by (R)) and so
A
(e(p)
A
e(q)) =
A
(e(p)
A
e(q)) for every
c, and the same for the reverse implication. To prove c[p q] [=
ALG

(L)
p q take A, F MOD

(L) and an evaluation e on A satisfying the equations in

the premises. Then e(p)
A
e(q), e(q)
A
e(p) F, i.e. e(p), e(q)
A
(F) and,
since the matrix is reduced e(p) = e(q).
32: We want to check p
L
(p) (p) [ c. By the rst condition in
3, this is equivalent to a (double) equational consequence w.r.t. ALG

(L) which, using

the second condition in 3 becomes a trivial statement.
45: Just observe that
A
(F) = Id
A
and use the isomorphism of
A
.
54: Recall that for every /-algebra A, Con
ALG

(L)
(A) is a complete lattice (see
the comments after Proposition 2.3.13). From Proposition 2.2.11 we know that
A
is
surjective and it preserves meets. We show that it is one-to-one. Suppose
A
(F) =

A
(G) for some F, G Ti
L
(A). Then, by the assumption 5, F/
A
(F) is the least
L-lter on A/
A
(F), and G/
A
(G) is the least L-lter on A/
A
(G) = A/
A
(F),
thus: F/
A
(F) = G/
A
(G). Now take any a F. [a]
F
F/
A
(F) = G/
A
(G),
and, since
A
(F) =
A
(G), [a]
F
= [a]
G
, and thus [a]
G
G/
A
(G), which gives
a G. By symmetry, we have F = G. We show now that
A
is order-reecting: if

A
(F)
A
(G) then
A
(F G) =
A
(F)
A
(G) =
A
(F), so F G = F, by
injectivity, and thus F G. Therefore,
A
is an order-preserving and order-reecting
bijection, and hence it is a lattice isomorphism.
14: We rst show that
A
is one-one. Suppose
A
(F) =
A
(G) for some F, G
Ti
L
(A). Given any a A, we have the following chain of equivalencies: a F iff
[a]
F
F/
A
(F) iff ([a]
F
) = ([a]
F
) for every c iff ([a]
G
) = ([a]
G
)
for every c iff [a]
G
G/
A
(G) iff a G. In a very similar way we can
check that it is order-reecting. From Proposition 2.2.11 we know that
A
(F) is onto
and order-preserving, and thus it is a lattice isomorphism.
42: First we prove T = Th
L
( [ ,
Fm
(T)) for every T Th(L).
Dene T
t
= Th
L
( [ ,
Fm
(T)). On the one hand, T
t
T, so by
monotonicity
Fm
(T
t
)
Fm
(T). On the other hand, if ,
Fm
(T), then
T
t
, so ,
Fm
(T
t
). Therefore, we have
Fm
(T
t
) =
Fm
(T) and,
by injectivity, T = T
t
. Thus, in particular we have shown that
p [ ,
Fm
(Th
L
(p)).
Let be the substitution mapping all variables to p. Then
p () () [ ,
Fm
(Th
L
(p)).
Therefore the set c(p) = () () [ ,
Fm
(Th
L
(p)) clearly satises the
condition (Alg).
124 Petr Cintula and Carles Noguera
Observe that, due to Corollary 2.2.4, the denition of algebraically implicative log-
ics is intrinsic because it does not depend on the chosen implication.
EXAMPLE 2.4.6. In many cases of interest, one equation is enough to satisfy condition
(Alg). For instance, classical logic and, in general, all the expansions of MTL men-
tioned in the previous chapter are algebraically implicative by using the set p 1,
and UL is algebraically implicative by using p 1 1.
PROPOSITION 2.4.7. If L is a nitary algebraically implicative logic, then ALG

(L)
is a quasivariety and the set c can be taken nite.
Proof. To prove that ALG

(L) is a quasivariety (i.e. a quasiequational class of alge-

bras) it is enough to take an arbitrary /-algebra Asuch that all the quasiequations valid
in ALG

(L) hold in A, and prove that then A ALG

(L). Dene the lter of A

as the set F = a A [
A
(a) =
A
(a) for every c. Let us see that
A, F MOD

(L). Suppose that

L
. By nitarity, there is a nite
0
such
that
0

L
. Assume that, for some evaluation e on A, e[] F. By the rst condition
in part 3 of the previous theorem we have c[
0
] [=
ALG

(L)
c(), which can be seen
as a quasiequation and hence also valid in A. On the other hand we have e[
0
] F,
therefore e() F. Finally, the second condition in part 3 of the previous theorem
implies that the matrix is reduced.
The fact that the set c can be taken nite is a straightforward corollary of claim 2 of
Theorem 2.4.5.
Analogously to the convention for weakly implicative logics, which always come
with a xed principal implication, for each algebraically implicative logic we x the de-
fault set of equations providing algebraicity and denote it as c (in fact, this set is unique
up to interderivability in ALG

(L)). Observe that these equations can be identied

with their corresponding pairs of formulae; then we call them algebraizing pairs. The
fact that lters in reduced matrices can be dened by equations has several interesting
straightforward consequences.
PROPOSITION 2.4.8. Let L be an algebraically implicative logic, A, B ALG

(L),
and A, F MOD

(L). Then:
1. F G for any G Ti
L
(A).
2. If A, G MOD

(L) then F = G, i.e. Ais the algebraic reduct of a unique

reduced matrix.
3. A mapping h: A B is a homomorphism of algebras from Ato B iff it is a
homomorphism between the corresponding matrices.
4. A mapping h: A B is an embedding of algebras from Ato B iff it is a
one-to-one strict homomorphism between the corresponding matrices.
5. A ALG

(L)
R(F)SI
iff A, F MOD

(L)
R(F)SI
.
Chapter II: A General Framework for Mathematical Fuzzy Logic 125
Finally, we close our excursion to Abstract Algebraic Logic by introducing two
special subclasses of algebraically implicative logics.
DEFINITION 2.4.9 (Rasiowa-implicative and regularly implicative logics). We say that
a logic L is Rasiowa-implicative if it is weakly implicative and
(W)
L

We use the term regularly implicative if L satises only this weaker condition:
(Reg) ,
L

PROPOSITION 2.4.10. A weakly implicative logic L is regularly implicative iff all the
lters of the matrices in MOD

(L) are singletons.

Proof. Elementary check.
PROPOSITION 2.4.11. Each Rasiowa-implicative logic is regularly implicative and
each regularly implicative logic is algebraically implicative.
Proof. The rst claim is obvious. For the second one it sufces to check that any reg-
ularly implicative logic satises the condition (Alg) (see Theorem 2.4.5) for the set of
equations c(p) = p p p.
The four classes of implicative logics that we have dened (weakly, algebraically,
regularly, Rasiowa-implicative) are mutually distinct. Indeed:
The equivalence fragment of classical logic is a regularly implicative but not
Rasiowa-implicative logic (to be more precise it is easy to see that the equivalence
connective does not satisfy the condition (W); moreover, in [22] it is proved
that no weak implication satisfying this condition is denable in this logic).
The uninorm logic UL is algebraically, but not regularly, implicative (because of
Proposition 2.4.10).
The logic BCI is weakly, but not algebraically, implicative (see [8]).
2.5 Substructural logics
In this section we introduce an important broad family of weakly implicative logics,
the substructural logics, starting froma very weak one which we call SL. We present this
basic logic in an implicit way as the least logic with certain desired behavior of connec-
tives. Then we dene substructural logics as expansions of the corresponding fragment
of SL. We will study some syntactical and semantical properties, and algebraization
of these logics. Then we will be able to identify them among the substructural logics
studied under this label in the literature; indeed we will show that SL actually coincides
with the bounded non-associative Full Lambek logic.
The language /
SL
consists of the connectives listed in Table 1, i.e. most of the usual
connectives in substructural logics (we will comment on the names and r ole played by
these connectives after the next denition). When writing formulae in this language
126 Petr Cintula and Carles Noguera
Symbol Arity Name Alternative names
2 principal implication right residuum
& 2 residuated conjunction fusion, multiplicative/strong conj.
2 dual implication left residuum
2 lattice protoconjunction additive/weak/lattice conjunction
2 lattice protodisjunction additive/weak/lattice disjunction
1 0 verum multiplicative truth, unit
0 0 falsum multiplicative falsum
0 top additive/lattice truth
0 bottom additive/lattice falsum
Table 1. The language of SL
Consecution symbol Name
( ) & (Res) residuation
( ) ( ) (E

) -exchange
(symm) symmetry
(1) lower bound
(2) lower bound
, (3) inmality
(1) upper bound
(2) upper bound
, (3) supremality
1 (Push) push
1 (Pop) pop
(Veq) verum ex quolibet
(Efq) ex falso quodlibet
Table 2. Consecutions for SL
(or its fragments) we will assume that the increasing binding order is: rst &, then
, , and nally , . For the sake of consistency with the general convention
in this chapter that every logic comes with a xed principal implication , we keep on
using this notation along with as the dual implication (soon we will prove a duality
theorem that shows that the choice between the principal and the dual implication is in
a way arbitrary). When identifying SL with the bounded non-associative Full Lambek
logic we will also show how our notation relates to the standard one for substructural
logics in the literature which uses and / instead of and , graphically denoting that
these implications are respectively the right and left residua of the conjunction &.
DEFINITION 2.5.1 (The logic SL). SL is the weakest weakly implicative logic in the
language /
SL
satisfying the consecutions from Table 2.
Chapter II: A General Framework for Mathematical Fuzzy Logic 127
Observe that SL is a weakly implicative logic and is its principal implication.
Even though we do not explicitly postulate any additional properties of , we will see
in Proposition 2.5.5 that its interplay with other connectives entails some rather strong
properties usually possessed by implications in known (non-)classical logics. The con-
nective & is a residuated conjunction whose r ole could be described as aggregation of
premises in a chain of implications as shown by residuation rules (Res). In fact, it
must be noted that the order of arguments in the formulation of (Res) is arbitrary (for
any connective & we could always dene its transposition &
t
as &
t
= & );
we have decided to formulate it in this way to have a more straightforward connection
with a stronger axiomatic formulation of (Res) which is equivalent to associativity (see
Theorem 2.5.7). While (Res) allows us to aggregate premises, (E

) allows us to swap
them but at the price of replacing the inner occurence of the principal implication by its
dual version (the rule (symm) ensures that can be seen as another principal im-
plication in SL interderivable with ). However, we cannot replace (E

) by a simpler
form involving only one implication, because it would entail commutativity of &(which
can be refuted by a simple semantic counterexample). On the other hand, the semantics
of these connectives is quite simple. Indeed, if we x &
A
in any reduced SL-matrix A
then
A
has to be its right residuum and
A
the left residuum (see part 8 of Proposi-
tion 2.5.10) and both
A
and
A
dene the same matrix order
A
. For more details
on residuated structures and their logics see [44].
The remaining binary connectives are easily understood: the rules for and en-
sure that these connectives correspond to the operations of inmum and supremum in
the lattice order given by the principal implication. Note however that we do not call
them conjunction and disjunction in Table 1 but add the prex proto. The reason
is that these rules are not enough to enforce by themselves a proper behavior of these
connectives:
1. in the case of , the adjunction rule ,
SL
, essential in the intended
behavior of conjunctions, holds due to the presence of the truth constant 1 and
fails in the least weakly implicative logic satisfying all consecutions from Table 2
but (Push) and (Pop),
2. in the case of , this protodisjunction does not enjoy the Proof by Cases Property
in SL: , and , entail , (in Section 2.6 we will see
how to recover this property in some extensions of SL and Section 2.7 studies its
characterizations and consequences).
The meaning of and and their dening rules is self-explanatory as maximum
and minimum elements of the order induced by the principal implication. The r ole of 1
is to be the least designated truth value. Finally, the r ole of 0, although its value is left
unspecied (note that there is no consecution involving 0 among those in Table 2), is to
dene negations by 0 and 0.
Of course we could immediately design one specic axiomatic system for SL (con-
sisting of reexivity, transitivity, modus ponens, the congruence rules for all connectives
and consecutions from Table 2). Later (Theorem 2.5.13) we will present a more natural
axiomatic system for SL. The idea behind our denition of SL, and behind the conven-
tion for substructural logics that we will introduce soon, is to pick a short list of rules
128 Petr Cintula and Carles Noguera
that connectives must satisfy to have the minimal usual behavior in substructural logics.
Moreover, as we will soon see (Proposition 2.5.4), the connectives are uniquely deter-
mined by these rules. The axiomatic system mentioned above allows us to prove quite
easily the following duality theorem:
DEFINITION 2.5.2 (Mirror image). Given a formula of /
SL
its mirror image
t
is
obtained by replacing in all occurrences of by , and vice versa, and by replacing
all subformulae of the form & by &. The mirror image of a set T of formulae of
/
SL
is T
t
=
t
[ T.
THEOREM 2.5.3 (Duality Theorem). For each set of formulae T of /
SL
we have:
T
SL
iff T
t

SL

t
.
Proof. We show only one direction (the second one immediately follows from the fact
that (
t
)
t
= ). We prove the claim for axioms and rules from the axiomatic system
described in the last paragraph above with formulae replaced by variables and then, by
structurality and the notion of proof, we are done.
The case of (symm) is trivial. From p q
SL
p & r q & r and p q
SL
r & p r & q we obtain p q, q p
SL
p & r q & r and p q, q p
SL
r & p r & q. And so we have the mirror version of congruence for &. The mirror
versions of congruence rules for both implications are proved analogously.
Next observe: ( )
SL
( )
SL
( )
SL
( ) and ( )
SL
( )
SL
( )
SL
&
SL
& , thus also mirror versions of (E

) and (Res) are proved.

Let T be any of the remaining rules. We know that neither & nor appears in
any formula fromT and all of these formulas are either variables or contain ex-
actly once and as principal connective. Thus the rule (symm) gives us straightforwardly
the mirror versions of these rules.
The following proposition shows that connectives are uniquely determined by the
rules we have introduced.
PROPOSITION 2.5.4. Let L be a weakly implicative expansion of SL with the same
principal implication and c a connective of /
SL
0. Suppose that c is a connective
(primitive or denable) of L of the same arity as c obeying the rules for c in Table 2.
Then the two connectives are equivalent in L, i.e.
L
c c , or
L
c c,
according to the arity of c.
Proof. The only non-trivial case is for . From
L
( ) ( ) we use
(E

) to obtain
L
(( ) ). Using (E

) this time for gives us

L
( ) ( ). The second implication is completely analogous.
PROPOSITION 2.5.5. The following are derivable in SL:
(P
SL
1) (( ) )
(P
SL
2) & ( )
Chapter II: A General Framework for Mathematical Fuzzy Logic 129
(P
SL
3) (( ) )
(P
SL
4) ( )
(P
SL
5) ( ) ( ) (Sufxing)
(P
SL
6) ( ) ( ) (Prexing)
(P
SL
7) ( & )
(P
SL
8) & &
(P
SL
9) & &
(P
SL
10)
1

1
,
2

2

1
&
2

1
&
2
(P
SL
11) ,
(P
SL
12) ( ) ( ) ( )
(P
SL
13) ( ) ( ) ( )
(P
SL
14) 1
(P
SL
15) 1 ( )
(P
SL
16) (1 )
(P
SL
17) & 1
(P
SL
18) 1 &
(P
SL
19) ( )
(P
SL
20) & ( ) & &
(P
SL
21) ( ) & & &
(P
SL
22) ( 1) & ( 1) 1
(P
SL
23) ( 1) & ( 1) 1
(P
SL
24) ( ) 1 ( 1 1)
For , , SL also proves:
(C

)
(I

)
(A

) ( ) ( )
Proof. The proof of the second part is straightforward. We give hints of the proofs of
the more complicated statements in the rst part:
(P
SL
1) From ( ) ( ) using (E

).
(P
SL
4) From (P
SL
1) using (MP) and (symm).
(P
SL
5) From , (( ) ) (( ) ) using (E

).
(P
SL
6) From (P
SL
2) we obtain & ( ) and (Res) completes
the proof.
(P
SL
8) From ( & ) we obtain ( & ) and (Res)
completes the proof.
130 Petr Cintula and Carles Noguera
(P
SL
9) Use the previous claim together with the duality theorem and (symm) twice.
(P
SL
11) 1 and 1 and thus , 1 . The rest is trivial.
(P
SL
12) Using (1), (P
SL
2) (P
SL
8) we prove & (( ) ( ) ) and
analogously also & (( ) ( ) ), (3) and (Res) complete
the proof.
(P
SL
13) First, we obtain (( ) ) (( ) ( ) ) from
(1) and the dual of (P
SL
5). Then from (( ) ) (P
SL
1) we
obtain (( ) ( ) ). Analogously we can prove that
(( )( ) ). Finally, (3) and (E

) complete the proof.

(P
SL
16) From 1 ( ) using (E

) we obtain (1 ). The converse

implication follows from (P
SL
4) and (P
SL
14).
(P
SL
20) From(1) (or (2) respectively) and (P
SL
8) we obtain: & &()
and & & ( ), and so (3) completes the proof of one implica-
tion. The converse one: from (Res) and (1) (or (2) respectively) we obtain
( & &) and ( & &). (3) and (Res)
complete the proof.
(P
SL
21) Analogous (or using duality theorem and (symm)).
(P
SL
22) We apply (P
SL
8) to 1 1 and obtain ( 1) & ( 1) ( 1) & 1
and (P
SL
17) completes the proof.
Next we study some notable extensions of SL.
DEFINITION 2.5.6. Let us consider the following consecutions:
a
1
& ( & ) ( & ) & re-associate to the left
a
2
( & ) & & ( & ) re-associate to the right
e ( ) ( ) exchange
c ( ) contraction
i ( ) left weakening
o 0 right weakening
Given any X a
1
, a
2
, e, c, i, o, and any weakly implicative logic L in a sufciently
expressive language, by L
X
we denote the expansion of L by X. If both a
1
and a
2
are
in X we replace them by the symbol a. Analogously if both i and o are in X we replace
them by the symbol w.
The next theorem shows how the axioms of exchange, contraction and weakening
can be described as properties of &. Moreover, it also shows under which conditions the
conjunction & is associative. It turns out that both halves of associativity are equivalent
to other interesting logical laws, usually resulting from strengthening rules of SL into an
axiomatic form.
THEOREM 2.5.7. Each one of the following axiomatic extensions of SL is axiomatized
by any one of the corresponding indicated rules:
Chapter II: A General Framework for Mathematical Fuzzy Logic 131
SL
a
1
1. ( & ) ( ( ))
2. ( ) (( ) ( ))
3. ( ( )) ( ( ))
SL
a
2
1. ( ( )) ( & )
2. ( ( )) ( ( ))
SL
e
1. & &
2. ( ) ( )
3. ( ) ( )
SL
c
1. &
2. &
SL
i
1. &
2.
3. 1
4. &
5. &
Proof. Recall that axioms are regarded as rules with empty set of hypotheses. The
claims of this theorem will be proved by showing (a chain of) implications of the form
the extension of SL by the rule x derives the rule y ([xy] in symbols), where x and y
are either the names of the rules or numbers denoting the formulae in question.
SL
a
1
[12] From &( ) we obtain ( ) ( &( ) ) (by
Sufxing). Thus 1 (in the form ( &( ) ) (( ) ( )))
and transitivity complete the proof.
[23] From (( ) ) (an instance of (P
SL
3)) and 2 we obtain
(( ( )) ( )). (E

) nishes the proof.

[31] ( & ) ( & ) and so & (( & ) )
by (E

). Thus ( (( & ) )), thus by 3 and transitivity

(( & ) ( )). (E

) nishes the proof.

[1a
1
] From (&) & (&) & we get (& (&) &)
by (Res); by 1 and transitivity we obtain ( ( (&) &)); (Res)
used twice completes the proof.
[a
1
2] & ( ) thus by Sufxing ( ) ( & ( ) );
(Res) gives us ( & ( )) & ( ) and so using a
1
we obtain
& (( ) & ( )) ; (Res) completes the proof.
132 Petr Cintula and Carles Noguera
SL
a
2
[1a
2
] From & ( & ) & ( & ) by (Res) used twice we obtain
( ( &( &))); by 1 and transitivity we obtain (&
& ( & )); (Res) completes the proof.
[21] From ( ( )) ( ( )) and (E

) we obtain
(( ( )) ( )). Thus also ( (( ( )) ))
(by 2 and transitivity). Using (Res) and (E

) complete the proof.

[a
2
2] From ( ( )) ( ( )) and (E

) we obtain
(( ( )) ( )) and so by the (Res) used twice we have
&(( ( )) &) . Thus also ( &( ( ))) & by
a
2
. Using (Res) and (E

) we obtain & ( ( )) ( ); (Res)

completes the proof.
SL
e
[1e] is obvious using (Res); [e2] follows from Proposition 2.5.4. To prove
[23] we start with (P
SL
1): (( ) ), then 2 and (E

) complete the
proof. To prove [31] we start with (P
SL
7) and by 3 we obtain ( &),
then (E

) and (Res) complete the proof.

SL
c
[1c] is obvious using (Res); for [c2] observe that from (1) and (2) be
obtain ( ) & ( ) & by (P
SL
10) and so (Res) and c complete the
proof. The nal claim [21] follows easily using (I

).
SL
i
The proofs of [1i], [i2], [23], [31], [34], and [43] are almost straightfor-
ward. To conclude the proof observe that from the fact that [i4] we obtain [i5]
using (3); the nal implication [5i] is trivial.
Now we can easily obtain the duality theorem (cf. Theorem 2.5.3) for notable ex-
tensions of SL. Recall that by
t
we denote the mirror image of .
THEOREM 2.5.8 (Duality theorem for SL
X
). Let X a, e, c, i, o. Then for each set
of formulae T we have:
T
SL
X
iff T
t

SL
X

t
.
Based on the logic SL we introduce now a general notion of substructural logic. By
doing so we do not expect to encompass all logics that may have been labeled in this
manner in the literature, but we only intend to introduce a broad class of substructural
logics in the framework of weakly implicative logics to which our methods will usefully
apply. We could achieve a greater level of generality by means of a more complex, and
probably less natural, denition, however we think the following convention is broad
enough for the purposes of the present text.
CONVENTION 2.5.9 ((Associative) substructural logic). A weakly implicative logic in
a language / is substructural if it is an expansion of the / /
SL
-fragment of SL. A
substructural logic is associative if it expands the / /
SL
-fragment of SL
a
.
Note that this convention clearly covers many well-known systems such as BCK
and BCI, all fuzzy logics introduced in Chapter I, intuitionistic and classical logic.
Chapter II: A General Framework for Mathematical Fuzzy Logic 133
Later, in Theorem 2.5.13, we identify SL with a well studied logical system: a bounded
non-associative variant of Full Lambek logic FL [44]. Then it will become apparent that
our convention also covers most logics referred to as substructural logics in the litera-
ture. In particular it covers all substructural logics over FL as deeply studied in [44]
(axiomatic extensions of FL) or substructural logics as logics of residuated structures
as proposed in the nal remarks (Section 6) of [74] (fragments of axiomatic extensions
of FL).
Our design choices make the denition of substructural logic a normative one in the
sense that, when using a traditional symbol for a connective of a given logic, we are pos-
tulating that it must at least satisfy the logical rules derivable in SL. This may have some
unexpected consequences. For instance, the logic BCK

of BCK-semilattices [62] in
the language , is not a substructural logic in the sense of our convention (it does
not satisfy (P
SL
12) from Proposition 2.5.5); however, if we formulated it in the language
, , then it would indeed satisfy the convention (because then the only /
SL
connec-
tive present in its language, namely , behaves as it should). Other examples which
illustrate this situation are Avrons logics RMI
min
and RMI, excluded from our notion
because they do not satisfy (P
SL
11).
The previously proved syntactical properties of SL and its prominent extensions
(Proposition 2.5.5 and Theorem 2.5.7) clearly hold for all substructural logics in a suf-
ciently expressive language. Let us list several further observations on substructural
logics (L stands for an arbitrary weakly implicative substructural logic in a sufciently
expressive language):
In L
o
the truth constants 0 and coincide (
SL
o
0) using Proposition 2.5.4.
In L
i
the truth constants 1 and coincide (
SL
i
1) using Proposition 2.5.4,
and furthermore also
SL
i
1 ( ).
L
ae
is axiomatized (relative to L) by axiom ( ( )) ( ( )).
L
ac
proves ( ( )) ( ) (but, in general, this axiom as not
sufcient to axiomatize L
ac
relative to L).
L is Rasiowa-implicative iff it proves i (and thus all these logics are algebraically
implicative). Furthermore, in Rasiowa-implicative substructural logics we prove
1 ( ) and so 1 can be viewed as a dened connective.
The following are some basic semantic properties of the connectives in substructural
logics, which can be easily checked.
PROPOSITION 2.5.10. Let L be a substructural logic in a sufciently expressive lan-
guage and A = A, F MOD

(L). Then:
1. 1
A
= min

A
F.
2.
A
= max

A
A and
A
F.
3.
A
= min

A
A and
A
/ F if Ais not trivial.
134 Petr Cintula and Carles Noguera
4.
A
is a -semilattice order.
5.
A
is a -semilattice order.
6.
A
is antitone in the rst argument and monotone in the second one w.r.t.
A
.
7. &
A
is monotone in both arguments w.r.t.
A
and 1
A
is its unit.
8. For every x, y, z A, x &
A
y
A
z iff y
A
x
A
z iff x
A
y
A
z.
9. For every x, y, z A, x
A
y = maxz [ x &
A
z
A
y.
10. For every x, y, z A, x
A
y = maxz [ z &
A
x
A
y.
11. For every x, y, z A, x &
A
y = minz [ y
A
x
A
z =
= minz [ x
A
y
A
z.
Proof. All claims are easily checked. The eighth item follows from (Res) and the dual-
ity theorem, and the last three items follow from this one.
THEOREM 2.5.11. Any substructural logic with or in its language is algebraically
implicative.
Proof. The proof would be simpler if we further assumed that 1 is in the language of
our logic, because then the algebraizing pair would be simply 1, 1 (or 1, ),
which can be proved in a straightforward way or using the previous proposition. Let us
give a more general proof not assuming the presence of 1.
Assume that our logic has in its language; we show that ( ) ,
is an algebraizing pair (in the second case we would analogously prove that claim for
( ) , ). Trivially ( ) ( ) and ( )
( ) (because ( ) ). The second direction: clearly ( )
( ) ( ) and so ( ) ( ) (as obviously
( ) ).
Our next aim is to identify our basic logic SL and its extensions SL
X
among well-
known substructural logics. Since many substructural logics are introduced in the liter-
ature in their unbounded form, we need the following convention:
CONVENTION 2.5.12. Let L be a weakly implicative logic. The logic called bounded
L, denoted as L

, is the expansion of L in the language with the additional truth con-

stant satisfying the axiom .
Reputably, the most prominent substructural logic is the Full Lambek Calculus [44],
denoted as FL. FL is formulated in a variant of our language /
SL
, which we will call
here /
FL
, with truth constants 0 and 1, two implication connectives and / (in the
commutative extensions of FL, denoted as FL
e
, these two connectives coincide and then
they are denoted by the symbol ), residuated conjunction &,
6
and lattice connectives
, . An axiomatic system for FL (taken from [44, Fig. 2.10]) is presented in Table 4.
6
There is a usual convention in the papers on this logic and its variants to omit the symbol &and write just
instead of &.
Chapter II: A General Framework for Mathematical Fuzzy Logic 135
, ( )
( ) ( ) ( ) ( )
(( / ) ) ( ) ( / ) /
( ) ( ) ( ) ,
()()() (/)(/)(/)
( ) ( )
1 1 ( ) (1 )
Table 3. The axiomatic system for the non-associative Full Lambek logic
Another important substructural logic is a non-associative variant of Full Lambek
Calculus, also in the language /
FL
, given by the axiomatic system in Table 3 (presented
in [46, Fig. 5]). The name and the symbol for this logic are not yet settled and thus we
avoid any explicit reference.
Before we show how these two logics are related with SL we provide a transla-
tion between their languages. We actually present two possible translations (the duality
theorem ensures the validity of the following theorem no matter which one we use).
7
/
FL
notation direct translation indirect translation
&

/
THEOREM 2.5.13. The logic SL
a
is the termwise equivalent to bounded Full Lam-
bek logic using of the translations given above. Analogously, the logic SL is termwise
equivalent to bounded non-associative Full Lambek logic.
Proof. All the axioms and rules of (non-associative) Full Lambek logic are among the
consecutions in Table 2, those proved in Proposition 2.5.5 or are equivalent to associa-
tivity as shown in Theorem 2.5.7). We leave the converse direction as an exercise.
Therefore, all the prominent extensions of SL
a
that we have considered here are
the bounded versions of the well-known extensions of FL studied in the mainstream
literature on substructural logics, namely for every X e, c, i, o SL
a,X
coincides
with the bounded version of FL
X
(modulo language translation). The logics FL
X
are
called basic substructural logics in [44].
Since in any substructural logic extending SL
e
only one implication is needed, we
obtain a simplied axiomatic system for FL
e
in Table 5 (taken from [44, Fig. 2.9]). One
can easily observe that the axiomatic system for FL
ew
can be simplied by taking the
one for FL
e
and replacing (&) with & and removing the axioms (1),
(1 ), and the rule (adj
u
).
7
This independence from the choice of the translation also ensures that any fragment of (non-associative)
Full Lambek logic containing at least / or \ is substructural in the sense of Convention 2.5.9.
136 Petr Cintula and Carles Noguera
(id
l
) (identity)
(pf
l
) ()[()()] (prexing)
(as
ll
) [(/)] (assertion)
(a) [()/][(/)] (associativity)
(&/) [()/](/) (fusion divisions)
(&) [( 1)( 1)]( ) (fusion conjunction)
() ( ) (conjunction division)
() ( ) (conjunction division)
() [() ()][( )] (division conjunction)
() ( ) (division disjunction)
() ( ) (division disjunction)
() [() ()][( )] (disjunction division)
(&) () (division fusion)
(&) [()]() (fusion division)
(1) 1 (unit)
(1) 1() (unit division)
(1) (1) (division unit)
(mp
l
) , (modus ponens)
(adj
u
) 1 (adjunction unit)
(pn
l
) (product normality)
(pn
r
) / (product normality)
Table 4. An axiomatic system for FL
(id) (identity)
(pf) ( ) (( ) ( )) (prexing)
(per) ( ( )) ( ( )) (permutation)
(&) [( 1)( 1)] ( ) (fusion conjunction)
( ) ( ) (conjunction implication)
( ) ( ) (conjunction implication)
( ) [( ) ( )] [ ( )] (implication conjunction)
( ) ( ) (implication disjunction)
( ) ( ) (implication disjunction)
( ) [( ) ( )] [( ) ] (disjunction implication)
( &) ( ) (division fusion)
(& ) [ ( )] ( ) (fusion implication)
(1) 1 (unit)
(1 ) 1 ( ) (unit implication)
(mp) , (modus ponens)
(adj
u
) 1 (adjunction unit)
Table 5. An axiomatic system for FL
e
Chapter II: A General Framework for Mathematical Fuzzy Logic 137
2.6 Deduction Theorems and Proof by Cases in substructural logics
In this subsection we deal with various forms of deduction theorems and use them
to obtain proof by cases properties for prominent substructural logics. Recall that we
work with a xed set of propositional variables Var. A -formula is built using Var and
a xed propositional variable not occurring in Var, a -substitution is dened for the
extended language as expected. These notions will play a technical r ole in the upcoming
denitions. Let be a formula, be a -formula, and a -substitution dened as
() = and p = p for p Var. By () we denote the formula (in the original set
of variables) .
DEFINITION 2.6.1. Given a set of -formulae we dene the set

of -formulae as
the smallest set such that

,
()

for each and each

.
If L has & and 1 in its language we dene the set () of -formulae as the smallest
set of formulae containing 1 and closed under &.
Note that the elements of () can be uniquely described by nite trees labeled by
elements of 1.
8
DEFINITION 2.6.2 ((Almost) (MP)-based logic, basic deduction terms). Let bDT be
a set of -formulae. A substructural logic L is almost (MP)-based w.r.t. the set of basic
deduction terms bDT if:
The set bDT is closed under all -substitutions such that () = .
L has a presentation where the only deduction rules are modus ponens and those
from () [ Fm
1
, bDT.
For each bDT and each formulae , , there exist
1
,
2
bDT such that:

L

1
( ) (
2
() ()).
L is called (MP)-based if it admits the empty set as a set of basic deduction terms.
Note that the described axiomatic system is indeed closed under all substitutions.
FL
e
is almost (MP)-based with bDT = 1 (recall the axiomatic system in Table 5
and (P
SL
24)), while FL
ew
is (MP)-based. Also notice that any axiomatic extension of
an almost (MP)-based logic is almost (MP)-based too. Finally observe that
L
()
for each (bDT

), and, if L is (MP)-based (i.e. has bDT = ), then bDT

= .
8
A problem could arise here if already contains a conjunction of some of its elements; then there are (at
least) two trees representing this formula as conjunction of elements of . Clearly that has to be a tree con-
taining all the possible tree-representation as subtrees; we take this maximal one as the unique representation.
138 Petr Cintula and Carles Noguera
THEOREM 2.6.3 (Almost-Implicational Deduction Theorem for almost (MP)-based
logics). Let L be a substructural logic with & and 1 in the language, and assume that
it is almost (MP)-based with a set of basic deduction terms bDT. Then for each set
, of formulae the following holds:
,
L
iff
L
() for some (bDT

).
Proof. To prove the right-to-left direction just recall
L
() for any (bDT

)
and use modus ponens. To prove the left-to-right direction we show that for each
in the proof of from the assumptions there is

(bDT

) such that

L

() . If = we set

= ; if or it is an axiom we set

= 1.
Assume that is obtained by modus ponens from and . By induction
hypothesis, we have
L

() and
L

() ( ). From the
former we derive
L
( ) (

() ), and so, by using the latter,

L

() (

() ), and thus
L

() &

() , and it sufces to set

&

to complete the proof.

Assume that is obtained by a rule of the form , where = () for some
bDT. By the induction hypothesis, we know that
L

() . We rst
observe one simple claim and prove another two:
Claim 1: for each bDT and for every formulae , there is

bDT such that

L

() ().
Claim 2: For each bDT and formulae , , there exist
1
,
2
bDT such that:

L

1
() &
2
() ( & ).
Proof of Claim 2: From the assumption of the theorem we obtain:

2
( & ) (
1
() ( & )) for some
1
,

2
bDT.
Using Claim 1 for =

2
and the fact that
L
( & ) we obtain:

L

2
()

2
( & ) for some
2
bDT.
The rest of the proof is simple.
Claim 3: For each bDT, (bDT

), and formula there exists

(bDT

)
such that:

L

() (()).
Proof of Claim 3: We proceed by induction via the depth of the tree representing .
If bDT

or = 1 the proof is done by setting

= () or

= 1 respectively.
Next assume that =
1
&
2
for some
1
,
2
(bDT

). By Claim 2 we obtain

1
,
2
bDT such that
L

1
(
1
()) &
2
(
2
()) (
1
() &
2
()). Then by the
induction assumption we obtain

1
,

2
(bDT

) such that
L

1
()
1
(
1
())
and
L

2
()
2
(
2
()). Setting

=

1
&

2
completes the proof using (P
SL
10).
The rest of the proof of the theorem: Recall that we have
L

() . There-
fore via Claim 1 we obtain

bDT such that
L

(

()) (). Claim 3 gives

us

such that
L

() .
Chapter II: A General Framework for Mathematical Fuzzy Logic 139
DEFINITION 2.6.4 (Almost-Implicational Deduction Theorem, deduction terms). Let
DT be a set of -formulae. A logic L has the Almost-Implicational Deduction Theorem
w.r.t. a set of deduction terms DT, if for each set , of formulae:
,
L
iff
L
() for some DT.
The previous theorem says that all almost (MP)-based logics enjoy the Almost-
Implicational Deduction Theorem with DT = (bDT

). We improve this result in two

ways: rst we notice under which assumption we can simplify the set of deduction terms,
and second we show that the condition that L is almost (MP)-based in Theorem 2.6.3 is
in fact necessary.
THEOREM 2.6.5. Let L be a substructural with & and 1 in the language satisfying the
Almost-Implicational Deduction Theorem w.r.t. a set DT.
L has the Almost-Implicational Deduction Theorem w.r.t. a set DT
t
DT if
and only if for every DT and every formula there is DT
t
such that

L
() ().
If L is nitary, then L is almost (MP)-based with the set
bDT = [ DT, a -substitution such that () = .
Proof. The proof of the right-to-left direction of the rst claim is straightforward. The
converse one is also easy: from
L
() () we obtain (using Almost-Implicational
Deduction Theorem w.r.t. DT)
L
() and so (using Almost-Implicational Deduc-
tion Theorem w.r.t. DT
t
) we obtain
L
() () for some DT
t
.
For the proof of the second claim, let us dene the logic L
t
axiomatized by all the
theorems of L, modus ponens and the rules () [ Fm
1
, bDT (note
that this set is closed under substitutions). From
L
() () and the right-to-left
direction of the deduction theorem we obtain that
L
(), as L is substructural,
it has modus ponens and so L
t
L. Assume that
L
. Due to the nitarity we
have
1
, . . .
n

L
for
i
. By repeatedly using the left-to-right direction of
the deduction theorem we obtain
L

1
(
1
) (
n
(
n
) ) ) for
i
DT.
Thus we obviously have
1
, . . .
n

L
. The last dening condition of a set of basic
deduction terms is easily obtained by a double application of the deduction theorem to
,
L
().
Let us recall the standard notation
n
=
n1
& (where
0
= 1). Note that in
associative substructural logics the bracketing in
n
is irrelevant.
COROLLARY 2.6.6 (Implicational Deduction Theorem for associative (MP)-based
logics). Let L be an associative substructural logic with & and 1 in the language. Then:
L is (MP)-based iff L is nitary and for each set , of formulae the following
holds:
,
L
iff
L

n
for some n 0.
140 Petr Cintula and Carles Noguera
We have seen that FL
ew
is an example of (MP)-based logic, thus we have just
shown that it enjoys this form of deduction theorem (and obviously the same holds for
its axiomatic extensions). In contrast, we can use the previous theorem to show that
FL
e
is not (MP)-based: indeed,
FL
e
1 would entail provability of the formula

n
1 for some n which can be refuted by a simple semantical counterexample.
On the other hand, since FL
e
is associative and
FL
e
(p 1) (p 1) 1, we know
that for each (bDT

FL
e
) there is some n such that
FL
e
(p 1)
n
, therefore
for each set , of formulae the following holds:
,
FL
e
iff
FL
e
( 1)
n
for some n 0.
The situation in FL is more complicated. First we need to introduce the notion of
conjugate.
DEFINITION2.6.7 (Left, right and iterated conjugates). Given formula , we dene left
and right conjugates w.r.t. as

() = ( &)1 and

() = ( &)1.
An iterated conjugate is a formula of the form () =

1
(

n
()) ), where
each

i
is either

i
or

i
.
A formula of the form () where is a left, right, or iterated conjugate is called
left, right, or iterated (resp.) conjugate of .
THEOREM 2.6.8 (Almost-Implicational Deduction Theorem for FL). The logic FL is
almost (MP)-based with the set of basic deduction terms

(),

() [ Fm
1
.
Therefore, for every set , of formulae the following holds:
,
FL
iff
FL
() for some conjunction of iterated conjugates.
Proof. Let bDT
FL
be the set of all left and right conjugates. Note that this set is closed
under every substitution for which () = and that the following derivations are
valid in FL:
1
()
1
() 1,

(),

(). Therefore the product

normality and adjunction unit rules could be equivalently replaced by the rules ()
for bDT
FL
. To complete the proof that bDT
FL
is a set of basic deductive terms of
FL, it is enough to prove the following:

( ) (

()

()),

( ) (

()

()).
The proof is heavily based on associativity; we use its variant forms introduced in
Theorem 2.5.7. First we prove ( ) ( & & ): from (P
SL
7) in the
form ( ( & )) and prexing get ( ) ( ( & )) and so
associativity nishes the proof. Now we prove the rst claim:
a ( ) (( ) ( )) mirror of associativity
b ( ( )) ( & ) mirror of associativity
c ( ) ( & & ) proved above
Chapter II: A General Framework for Mathematical Fuzzy Logic 141
d & (( ) & ) c and (E

)
e ( & ) [ (( ) & )] d and mirror of (P
SL
6)
f ( & ) [ & ( ) & ] e and an instance of b
g ( & ) [( & ( )) ( & )] f and an instance of a
h ( & ( )) [( & ) ( & )] g and (E

)
i ( & ( )) 1 [( & ) 1 ( & ) 1] h and
(P
SL
24) twice
Proof of the second claim:
a
t
( ) ( & & ) mirror of c
b
t
(( ) )) (P
SL
1)
c
t
(( ) & & ) a
t
and an instance of b
t
d
t
( ) & ( & ) c
t
and (E

)
e
t
( ( ) & ) [ ( & )] d
t
and (P
SL
6)
f
t
( ( ) & ) ( & & ) e
t
and associativity
g
t
( ( ) & ) [( & ) ( & )] f
t
and associativity
h
t
( ( ) & ) 1 [( & ) 1 ( & ) 1] g
t
and
(P
SL
24) twice
Interestingly enough, the deductions theorems studied in this section yield a con-
nection with a variant of the classical proof by cases property. Recall that classical logic
enjoys this meta-rule:
, ,
,
We will see now how a similar property can be obtained for almost (MP)-based sub-
structural logics with a more complex form of disjunction built from their sets of basic
deduction terms.
THEOREM 2.6.9 (Proof by Cases Property). Let L be a substructural logic with & and
1 in its language and assume that it is almost (MP)-based with a set of basic deduction
terms bDT such that
for each bDT

we have
L
() 1 for any formula ,
there is
0
bDT such that
L

0
() for any formula .
142 Petr Cintula and Carles Noguera
Then the following meta-rule is valid in L:
, ,
() () [ , bDT

Proof. We start by showing that L enjoys the Almost-Implicational Deduction Theorem
w.r.t. the set DT = (bDT

) [
L
() 1. We use Theorem 2.6.5: for
each (bDT

) we consider

(bDT

) resulting from understood as a tree by

replacing the node-labels by
0
(); then using (P
SL
8) and (P
SL
9) and the assumptions
of this theorem we can easily show that for each formula we have
L

() ()
and
L

() 1. Note that for all DT we can easily prove
L
() & and

L
& () .
Next, assume that ,
L
and ,
L
. Fromthe Almost-Implicational Deduc-
tion Theorem we obtain

DT such that
L

() and
L

()
and so
L

()

() . The proof is done by showing by induction over the

sum of the depths of the trees representing

that:
() () [ , bDT

()

().
The base of induction (when

bDT

) is trivial. For the induction step assume

that

=
1
&
2
. Using (P
SL
20), (P
SL
21), (1), (2), and (3) we obtain the following
chain of implications:
(

()
1
()) & (

()
2
())
[

() &

()] [

() &
2
()] [
1
() &

()] [
1
() &
2
()]

()

()

() [
1
() &
2
()]

()

().
The induction assumption used for

()
1
() and

()
2
() together with
(P
SL
7) completes the proof.
COROLLARY 2.6.10 (Proof by Cases Property for logics with weakening). Let L be
a substructural Rasiowa-implicative logic with & and 1 in its language. Assume that
it is almost (MP)-based with a set of basic deduction terms bDT. Then the following
meta-rule is valid in L:
, ,
() () [ , bDT

Proof. Let us dene bDT

= bDT and note that bDT

satises the conditions

of the previous theorem. As clearly, bDT

= bDT

the proof is done.

As a corollary we obtain the proof by cases property for three prominent substruc-
tural logics (note that in the case of FL
e
we can simplify the form given by the theorem
above using the provability of
FL
e
(p 1) (p 1) 1).
Chapter II: A General Framework for Mathematical Fuzzy Logic 143
COROLLARY 2.6.11 (Proof by cases in extensions of FL). The following meta-rule is
valid in FL:
, ,

1
()
1
() [
1
,
2
iterated conjugates
The following meta-rule is valid in FL
e
:
, ,
, ( 1) ( 1)
The following meta-rule is valid in FL
ew
:
, ,
,
2.7 Generalized disjunctions
As we have seen in the previous subsection, substructural logics may retain a form
of the classical proof by cases property at the price of using a more complex disjunction.
For instance, in the case of FL we have used a rather complicated set
1
()
2
() [

1
,
2
iterated conjugates of innitely many formulae involving two variables and pa-
rameters. The proof by cases property will play an important r ole in the following
sections where we study the interplay of disjunctions and implications (in particular,
disjunctions will be used to provide a powerful characterization of semilinear implica-
tions and, moreover, as we will see in Section 4, they are crucial for rst-order logics).
In order to prepare the ground for that, in this subsection we provide an abstract analysis
of disjunction connectives general enough to cover their possible complicated forms, as
the one we have seen in FL. Although a number of results we prove in this section hold
in general, for the sake of simplicity here we will mostly restrict ourselves to the case of
nitary logics which will allow us to provide easier proofs. We will indicate in which
results the nitarity assumption is not actually used.
DEFINITION 2.7.1 (Notation for generalized disjunctions). Let (p, q,

r ) be a set
of formulae in two variables p, q and a sequence (possibly empty, nite or innite) of
further variables

r called parameters. We dene:
=
_
(, ,

) [

Fm

1
.
Given sets
1
,
2
Fm
1
,
1
2
denotes the set

[
1
,
2
.
When there are no parameters in the set (p, q) and it is a singleton, we write
instead of .
CONVENTION2.7.2 (p-protodisjunction,protodisjunction). A parameterized set of for-
mulae (p, q,

r ) will be called a p-protodisjunction in L whenever it satises:
(PD)
L
and
L
.
If has no parameters we drop the prex p-.
144 Petr Cintula and Carles Noguera
This convention does not dene an interesting notion on its own because, actually,
any theorem (or set of theorems) in two variables of a given logic would be a protodis-
junction in this logic; we only introduce it as a useful means to shorten the formulation
of many upcoming denitions and results. In contrast, requiring the property of proof
by cases results in more interesting notions of disjunction.
DEFINITION 2.7.3 (p-disjunction, disjunction). Given a logic L, a (p-)protodisjunction
is called a (p-)disjunction (in L) whenever it satises the Proof by Cases Property,
PCP for short:
,
L
,
L

,
L

.
Observe that if is a disjunction in a logic it remains a disjunction in all its ax-
iomatic extensions and in its fragments containing the connectives used in . Later we
will give sufcient and necessary conditions for the preservation of the Proof by Cases
Property in expansions. Interestingly enough, all p-disjunctions in a given logic are
interderivable as we can easily prove:
LEMMA 2.7.4. Let L be a logic and ,
t
parameterized sets of formulae. Assume
that is a p-disjunction in L. Then:
t
is a p-disjunction in L iff
L

t
.
Thus the notion of p-disjunction is intrinsic for a given logic and in the upcom-
ing denition it does not matter which p-disjunction we choose since all of them are
interderivable.
DEFINITION 2.7.5 (p-disjunctional logic, disjunctional logic, disjunctive logic). Let L
be a logic. We say that L is a (p-)disjunctional logic if it has a (p-)disjunction. We use
the term disjunctive instead if the disjunction is just a single parameter-free formula.
As we will see in Example 2.7.9, in substructural logics the connective need not be
a disjunction in the technical sense just dened. Therefore, we introduce the following
terminology:
DEFINITION 2.7.6 (Lattice-disjunctive logic). Let L be weakly implicative disjunctive
logic L with principal implication and disjunction . Then L is lattice disjunctive if:
(1)
L

(2)
L

(3) ,
L

Thus a substructural logic with in its language is lattice-disjunctive iff satises
PCP. Note that if a logic L satises the conditions (1)(3) for two different (prim-
itive or derivable) connectives and
t
, then we can easily prove a stronger version of
Lemma 2.7.4 namely:
L

t
. Also note that if L is a lattice-disjunctive
logic, then for each A MOD

(L), the algebra A,

A
is a join-semilattice with
semi-lattice order
A
.
Clearly the classes of lattice-disjunctive, disjunctive, disjunctional and p-disjunc-
tional logics form an increasing chain under inclusion. Next, we present several ex-
amples of lattice-disjunctive logics to show that all these inclusions are proper, thus
demonstrating the non-triviality of our hierarchy.
Chapter II: A General Framework for Mathematical Fuzzy Logic 145
LEMMA 2.7.7. Any (MP)-based substructural Rasiowa-implicative logic with &and 1
in its language (e.g. any axiomatic extension of FL
ew
) is a lattice-disjunctive logic.
Proof. The PCP for these logics is shown in Corollary 2.6.10 by using the connective
and hence they are lattice-disjunctive.
THEOREM 2.7.8. There is a p-disjunctional logic which is not disjunctional, a disjunc-
tional logic which is not disjunctive, and a disjunctive substructural logic which is not
lattice-disjunctive.
The proof of this theorem is established by the following three examples.
EXAMPLE 2.7.9. The logic FL
e
is disjunctive but not lattice-disjunctive. In Corol-
lary 2.6.11 we have seen that (p, q) = (p 1) (q 1) satises the PCP. Since it
clearly satises (PD) too, it is a p-disjunction.
Assume now that the lattice connective satises the PCP. From 1 we
easily get (1). As also (1), we could use the PCP of to obtain
(1). Consider the FL
e
-matrix with the domain , a, b, 1, ; designated
set 1, ; the lattice connectives dened in the way that the elements form the non-
distributive lattice diamond where is the minimum element and is the maximum;
residual conjunction: x & 1 = 1 & x = x and x & y = x y for x, y ,= 1; and
implication: x y = maxz [ z & x y. Then, we reach a contradiction from these
simple observations: a b = but (a 1) b = b = b.
EXAMPLE 2.7.10. The implicational fragment of G odel-Dummett logic G

(intro-
duced in Chapter I) is disjunctional but not disjunctive. First we show that the set
= ( ) , ( ) is a disjunction. Since G is an axiomatic
extension of FL
ew
it satises: ( ) and ( ) and so
satises (PD). Now observe that , , ( ) , ( ) and
as we assume that , thus , , and so by the deduction theorem
, ( ) . Analogously we prove , ( ) and as
(( ) ) ((( ) ) ) is a theorem of G odel-Dummett logic we
obtain , as needed.
Assume that (p, q) is a disjunction. As a consequence of the standard complete-
ness theorem for G, we know that G

is complete with respect to the matrix A whose

universe is the real unit interval [0, 1], the lter is 1 and the only operation is:
a
A
b =
_
1 if a b,
b otherwise.
By Lemma 2.7.4, the formula (p, q) and the set are mutually derivable in
G

. As in G odel logic we have: (( ) ) (( ) ), we can

use the deduction theorem (Theorem 2.6.3) we obtain that (p, q) is interpreted in A
as the function maximum. So, in particular, for every a, b [0, 1) we have
A
(a, b) =
maxa, b. We show by an innite descent argument that this is impossible. Since is
the only connective in the language, the formula must be (p, q) = (p, q) (p, q).
Take any a, b [0, 1). If a b,
A
(a, b) =
A
(a, b)
A

A
(a, b) = b, which implies
146 Petr Cintula and Carles Noguera

A
(a, b) = b. Analogously, if a > b, we have
A
(a, b) = a. Thus, (p, q) would
be a strictly shorter formula with the same property. Following this line of reasoning
we would derive that for each a, b [0, 1) we have either a
A
b = maxa, b or
b
A
a = maxa, ba contradiction.
EXAMPLE 2.7.11. The implicational fragment of intuitionistic logic IPC

is p-dis-
junctional but not disjunctional. The fact that IPC

is p-disjunctional follows from

Theorem 2.7.20 (the lter-distributivity of this logic was proved in [28]). Assume, for
contradiction, that a set is a disjunction in IPC

. Thus by Corollary 2.7.17 it also a

disjunction in the full intuitionistic logic IPC. From Lemma 2.7.7 we know that IPC is
lattice-disjunctive and so, by Lemma 2.7.4, we have pq
L
pq. Using nitarity, the
presence of the lattice conjunction in the language of IPC and the deduction theorem
we obtain a formula
t
of two variables p, q built using only implication and lattice
conjunction such that
IPC
p
t
q p qwhich is known to be impossible (see
e.g. [67]).
On the other hand, we have seen in Corollary 2.6.11 that in FL the set (p, q, r) =

1
(p)
2
(q) [
1
,
2
iterated conjugates satises the PCP. As it clearly satises
also (PD) it is a p-disjunction. Sato in [79, Proposition 6.9] showed that there is no
protodisjunction in FL which would satisfy the PCP, i.e. that FL is not disjunctive.
Thanks to the presence of the lattice conjunction in FL we could derive that not even a
nite set of formulae would sufce to dene a disjunction in FL, but, unfortunately, we
are not able to extend this result to the non-existence of disjunctions dened by innite
sets. However we conjecture:
CONJECTURE 2.7.12. The logic FL is another example of a p-disjunctional logic
which is not disjunctional.
The Proof by Cases Property implies other interesting syntactical properties that a
disjunction is expected to satisfy: commutativity, idempotency, and associativity.
9
The
proof of the next lemma is straightforward.
LEMMA 2.7.13. If L is a logic and is a p-protodisjunction satisfying the PCP, then
it also satises the following conditions:
(C

)
L

(I

)
L

(A

) ()
L
()
Next we provide a characterization of (p-)disjunctions, essentially by showing what
needs to be added to the properties of the previous lemma to obtain a (p-)disjunction.
DEFINITION 2.7.14 (-form). Let be a p-protodisjunction and R = be a
consecution. We dene the -form of R, denoted by R

, as the set of consecutions

[ Fm
1
and .
9
Observe that commutativity, idempotency, and associativity are typically also satised by conjunction,
whereas (PD) and the PCP are typically satised only by disjunction connectives.
Chapter II: A General Framework for Mathematical Fuzzy Logic 147
THEOREM 2.7.15 (Syntactic characterization of p-disjunctions). Let L be a nitary
logic with a presentation /o and a (p-)protodisjunction. Then the following are
equivalent:
1. is a (p-)disjunction,
2. enjoys the strong Proof by Cases Property, sPCP for short,
,
1

L
,
2

L

,
1

2

L

,
3. satises (C

), (I

), and R

L for each R L,
4. satises (C

), (I

), and R

L for each R /o.

Proof. We start by showing the equivalence of the rst three properties, and then we
complete the proof by showing the implication 43 (the implication (34 is trivial).
12: we proceed by using induction. First observe that due the nitary of L we can
assume that both
1
,
2
are nite. Call a pair ,
1

L
and ,
2

L
a situation;
dene the complexity of a situation as a pair n, m where n and m are respectively the
cardinalities of
1

2
and
2

1
. We show by induction on k = n +m that in each
situation we obtain ,
1

2

L
.
First assume k 2. If n = 0, i.e.
1

2
, we obtain
1

1

1

2
and
since ,
1

1

L

1
the proof is done. The proof for m = 0 is analogous. If
n = m = 1 we use the PCP. The induction step: consider a situation with complexity
n, m, where n + m > 2. We can assume without loss of generality that n 2, take
a formula
1

2
and dene
t
1
=
1
. We know that ,
t
1
,
L
and
,
2

L
. Thus we also know that ,
t
1
,
L
and ,
t
1
,
2

L
; notice that
the complexity of this situation is 1, m and so we can use the induction assumption to
obtain ,
t
1
,
2

L
.
Thus we have the situation ,
2
,
t
1

L
and ,
2
,
2

L
(the second
claim is trivial); the complexity of this situation is n
t
, m
t
, where n
t
n 1 and
m
t
m, and so by the induction assumption we obtain ,
2
,
t
1

2

L
(which
is exactly what we wanted).
23: from
L
we obtain and
L
using (PD). By (PD) we also obtain

L
and the sPCP completes the proof.
31: assume than ,
L
and ,
L
. Using the assumption we obtain
,
L
and ,
L
. Using (C

) and (I

) we obtain
, ,
L
. By (PD) we known that
L
and
L
and so
the proof is done.
43: assume that
L
and we show
L
for each formula and
each appearing in the proof of from . If or is an axiom, the proof is trivial.
Now assume that R = is the deduction rule we use to get . From the induction
assumption we have
L
. As we know that R

L the proof is done.

REMARK 2.7.16. Notice that the equivalence of conditions 2, 3, and 4 could be es-
tablished even without the assumption of nitarity. Moreover, any of these conditions
implies in general condition 1.
148 Petr Cintula and Carles Noguera
In Example 2.7.9 we have seen a connective which clearly satises (PD), (C

),
(I

), and (A

) but it is not a disjunction. This demonstrates that the condition R

L
for each R L (resp. R /o) is necessary in parts 3 and 4 of the previous theorem (in
fact to prove this we have shown that the -form of the rule (adj
u
) is not valid in FL
e
).
As a corollary of the previous theorem we can answer the question when remains
a p-disjunction in an expansion of a given logic.
COROLLARY 2.7.17. Let be a p-disjunction in a nitary logic L
1
and let L
2
be an
expansion of L
1
by a set ( of nitary consecutions. Then is a p-disjunction in L
2
iff
R

L
2
for each R (. In particular, is a p-disjunction in any axiomatic expansion
of L
1
.
Proof. The left-to-right direction is a straightforward application of the previous theo-
rem. For the reverse direction take a presentation /o of L
1
. We know that L
2
has a
presentation /o
t
= [] [ is an /
2
-substitution, /o (. Thus
we need to prove that for each /o ( and for each /
2
-substitution we
have ([] )

L
2
, i.e. for each /
2
-formula , each (p, q, r
1
, . . . , r
n
) and
each sequence
1
, . . . ,
n
of /
2
-formulae we have []
L
2
(, ,
1
, . . . ,
n
).
If (, this is the assumption; we solve the remaining case.
Consider any enumeration of the propositional variables such that p
0
= q, p
i
= r
i
,
and /
1
-substitutions ,
1
and /
2
-substitution dened as:
p
i
= p
i+n+1
,

1
p
i
= p
in1
for i > n and p
i
otherwise,
p
i
= (p
in1
) for i > n, p
i
=
i
for 1 i n and p
0
= .
Observe that
1
= and = . From /o we get [] /o
and because ([] )

L
1
L
2
we obtain: []q
L
2
(, q, r
1
, . . . , r
n
) and
so [[]q]
L
2
(, q, r
1
, . . . , r
n
). Because obviously, (, q, r
1
, . . . , r
n
) =
(, ,
1
, . . . ,
n
), if we prove [[]q] [] the proof is done. It is enough
to observe that the formulae in [[]q] are of the form

(, ,
1
, . . . ,
k
)
for some ,

(p, q, r
1
, . . . , r
k
) and a sequence of /
2
-formulae
1
, . . . ,
k
.
Next, we prove that PCP also enjoys a transfer theorem (recall the commentary
before Corollary 2.3.4).
THEOREM 2.7.18 (Transfer theorem for PCP). Let L be a nitary logic with a presen-
tation /o and a (p-)protodisjunction. Then the following are equivalent:
1. is a (p-)disjunction,
2. Fi(Y ) Fi(Z) = Fi(Y
A
Z) for each /-algebra Aand each Y, Z A,
3. Fi(X, x) Fi(X, y) = Fi(X, x
A
y) for each /-algebra Aand each
X x, y A.
Chapter II: A General Framework for Mathematical Fuzzy Logic 149
Proof. 12: The inclusion Fi(X
A
Y ) Fi(X) Fi(Y ) follows easily from (PD).
To prove the converse one, we start by showing that for each x Fi(X) we have
x
A
y Fi(X
A
y) for each y. If x Fi(X), then (due to Proposition 2.1.23) there is
a proof of x from X in some presentation /o of L. We show that z
A
y Fi(X
A
y)
for each z labeling any node of that proof, i.e. for each (p, q, r
1
, . . . , r
n
) and each
sequence u
1
, . . . , u
n
of elements of A we have
A
(z, y, u
1
, . . . , u
n
) Fi(X
A
y).
It is trivial if z X. Otherwise there is a set Z of labels of the preceding nodes
(possible empty), a consecution /o, and an evaluation h, such that h[] = Z
and h() = z. Without loss of generality we could assume that variables q, r
1
, . . . , r
n
do not occur
10
in and so we can set h(q) = y and h(r
i
) = u
i
for every i
1, . . . , n. Thus h[q] Z
A
y Fi(X
A
y) (the last inclusion follows from the
induction assumption). From Theorem 2.7.15 we know that q
L
(, q, r
1
, . . . , r
n
)
and so
A
(z, y, u
1
, . . . , u
n
) = h((, q, r
1
, . . . , r
n
)) Fi(X
A
y).
Now we can nally prove that Fi(X) Fi(Y ) Fi(X
A
Y ). If z Fi(X) then
by the just proved claim for each y Y holds: z
A
y Fi(X
A
y) and so, by (C

),
y
A
z Fi(X
A
y). This can be more compactly written as: Y
A
z Fi(X
A
Y ).
Analogously we obtain z
A
z Fi(Y
A
z) from z Fi(Y ). Thus z Fi(Y
A
z)
(by (I

)) and so z Fi(X
A
Y ).
23: One inclusion again easily follows from (PD). To prove the other one we
use 2 to obtain Fi(X, x) Fi(X, y) Fi(X
A
X, X
A
y, x
A
X, x
A
y) and (PD)
completes the proof.
The nal implication is trivial.
REMARK 2.7.19. We have actually proved transfer of sPCP in all (not necessarily
nitary) logics. Furthermore the equivalence of conditions 1 and 3 (but not 2) holds in
all (not necessarily nitary) weakly implicative logics [21]. Later (in Theorem 3.2.13)
we will prove equivalence of all these three conditions (and those from Theorem 2.7.15)
for a special subclass of (not necessarily nitary) weakly implicative logics.
The following theorem gives an intrinsic characterization of p-disjunctional logics
in terms of (lter-)distributivity (recall Denition 2.1.22), as opposed to Theorem 2.7.15
characterizing particular p-disjunctions.
THEOREM 2.7.20 (Characterization of p-disjunctional logics). For any nitary weakly
implicative logic L the following are equivalent:
1. L is p-disjunctional,
2. L is lter-distributive,
3. the lattice Th(L) is distributive.
10
We could dene a new suitable with the same properties using a Hilbert-hotel style argument:
consider any enumeration of the propositional variables such that p
0
= q, p
i
= r
i
, a substitution (p
i
) =
p
i+n+1
, and any evaluation h

such that h

(p) = h(p). Then [] is the consecution we need:

indeed [] ,S, h

[[]] = Z, and h

() = z. Note that here we have used our assumption that

the axiomatic systems are closed under substitutions.
150 Petr Cintula and Carles Noguera
Proof. 12: Given a p-disjunction and F, G, H Ti
L
(A), we can write the follow-
ing chain of equations (the non-trivial ones are due to part 2 of Theorem 2.7.18):
F (G H) = Fi(F) Fi(G H)
= Fi(F(G H))
= Fi((FG) (FH))
= Fi(FG) Fi(FH)
= (Fi(F) Fi(G)) (Fi(F) Fi(H))
= (F G) (F H).
23: Trivial.
31: Dene = Th
L
(p) Th
L
(q). Given a pair of formulae , , take a surjec-
tive substitution such that p = and q = . If we knewthat Th
L
(p)Th
L
(q) =
Th
L
([pq]), then we could write the following chain of equations demonstrating that
is a p-disjunction (we use the denition of join in Th(L), the distributivity of Th(L),
the denition of , the fact mentioned above, and we observe that, due to the surjectivity
of , we have [pq] = ):
Th
L
(, ) Th
L
(, ) = (Th
L
() Th
L
()) (Th
L
() Th
L
())
= Th
L
() (Th
L
() Th
L
())
= Th
L
() Th
L
([pq])
= Th
L
() Th
L
()
= Th
L
(, ).
The rest of the proof is dedicated to proving Th
L
(p) Th
L
(q) = Th
L
([pq]).
We dene a theory Y =
1
[Th
L
()] (clearly Y Th(L) by Lemma 2.1.19), an in-
terval in the lattice of theories [Y, Fm
1
] = T Th(L) [ Y T, and mappings
: [Y, Fm
1
] Th(L) and
1
: Th(L) [Y, Fm
1
] dened as (T) = [T] and

1
(T) =
1
[T]. Clearly
1
[T] [Y, Fm
1
] for each T Th(L) (by Lemma 2.1.19
and the fact that Th
L
() T). To prove that (T) Th(L) for each T [Y, Fm
1
],
taking into account Lemma 2.1.19, all we need to show is () [T] implies T:
the assumption gives us = for some T, therefore
L
, hence
T (because Y and Y T), and thus by (MP) we obtain T.
Claim 1: is an isomorphism. Clearly both and
1
are order-preserving, thus
all we need to show is
1
((T)) = T and (
1
(S)) = S for each T [Y, Fm
1
]
and S Th(L). The two non-trivial inclusions are
1
((T)) T (which follows
from the already proved fact: () [T] implies T) and (
1
(S)) S (which
follows from the surjectivity of ).
Claim 2: Th
L
([]) = (Y Th
L
()) for each set of formulae . The rst
inclusion follows from: [] (Y Th
L
()) and (Y Th
L
()) Th(L).
The second inclusion: if (Y Th
L
()), then = and Y,
L
. Thus
[Y ], []
L
(), i.e., []
L
.
Now we can nish the proof of the theorem by a series of equations (we use Claim 2,
Chapter II: A General Framework for Mathematical Fuzzy Logic 151
Claim 1, distributivity of Th(L), and Claim 2 again):
Th
L
(p) Th
L
(q) = (Y Th
L
(p)) (Y Th
L
(q))
= ((Y Th
L
(p)) (Y Th
L
(q)))
= (Y (Th
L
(p) Th
L
(q)))
= Th
L
([Th
L
(p) Th
L
(q)])
= Th
L
([pq]).
Next, we introduce the notion of -prime lter by generalizing the classical notion
of prime lter in Boolean algebras.
DEFINITION 2.7.21 (Prime lter). Let L = /,
L
be a logic, a (possibly parame-
terized) set of formulae in two variables, A an /-algebra, and F Ti
L
(A). Then, F
is called -prime if for every a, b A, a
A
b F iff a F or b F.
Notice that when denes a disjunction connective , the previous denition gives
just the usual notion of prime lter.
DEFINITION 2.7.22 (Prime Extension Property). A logic L has the prime extension
property, PEP for short, with respect to a set if -prime theories form a base of the
closure system Th(L).
In the parameter-free case, we have the following characterizations of disjunctions
in terms of prime lters and their properties:
THEOREM 2.7.23 (Characterizations of disjunctions). Let L be a nitary logic and
a protodisjunction. Then the following are equivalent:
1. is a disjunction.
2. For every /-algebra Aand every F Ti
L
(A), F is nitely -irreducible iff it is
-prime.
3. For every A, F MOD

(L), A, F MOD

(L)
RFSI
iff F is -prime.
4. For every Fm
1
,
L
iff [=
A,F)MOD

(L) ] F is -prime]
.
5. For every /-algebra A, -prime lters form a base of Ti
L
(A).
6. L has the PEP w.r.t. .
Proof. 12: Consider any F Ti
L
(A), assume rst that F is not -prime, i.e. there
are x / F and y / F such that x
A
y F (the second implication in the denition
of prime lter holds always due to (PD)). Thus from Theorem 2.7.15 we know that
F = Fi(F, x
A
y) = Fi(F, x) Fi(F, y), i.e. F is the intersection of two strictly
bigger lters. Next assume that F is nitely reducible, i.e. F = F
1
F
2
and F _ F
i
.
Let us consider a
i
F
i
F. Thus, by (PD), we know that a
1

A
a
2
F
i
and so
a
1

A
a
2
F, i.e. F is not -prime.
The implication 23 is a direct consequence of Theorem 2.3.14 and 34 follows
directly from Theorem 2.3.16.
152 Petr Cintula and Carles Noguera
41: We will show that has the PCP. Assume that , -
L
. Thus
there is an A, F MOD

(L) where F is -prime and an A-evaluation e such

that e[, ] F and e() , F. Since is parameter-free, we obtain e()e() =
e[] F, and since F is -prime, we have e() F or e() F and so , -
L

or , -
L
.
We have established the equivalence of the rst four claims. To complete the proof
we rst observe that 25 follows immediately from the fact that (nitely) -irreducible
lters form a base of Ti
L
(A) (Corollary 2.3.9). The implication 56 is trivial. We
complete the proof by showing 61: assume that , -
L
, then using the PEP
there is a -prime theory
t
such that
t
-
L
. Thus
t

L
or
t

L
.
And so clearly , -
L
or , -
L
.
REMARK 2.7.24. Notice that the hypothesis in the previous theorem about the absence
of parameters in has only been used to show 41. In fact, in the general case,
condition 1 (is a p-disjunction) still implies all the rest, and conditions 1, 5, 6 are still
equivalent.
We consider now the problem of nding the minimal p-disjunctional extension of a
given logic. Observe, on the one hand, that if is a p-disjunction in a family of logics, it
remains a p-disjunction in their intersection. On the other hand, any is a p-disjunction
in the inconsistent logic. Thus, the following denition is sound:
DEFINITION 2.7.25 (Logic L

). Let L be a nitary logic and a p-protodisjunction.

We denote by L

the least logic extending L where is a p-disjunction.

PROPOSITION 2.7.26. Let L be a nitary logic and a p-protodisjunction. Then L

is nitary and it is the intersection of all nitary extensions of L where has the PCP.
Proof. Recall the notion of nitary companion of a logic S, denoted as T((S), which
is the largest nitary logic contained in S. Thus, since L is nitary, we know that L
T((L

) L

. If we show that has the PCP in T((L

), we obtain T((L

) = L

and hence L

is nitary. Actually, one can easily show in general that if has the PCP
in S, then it has the PCP in T((S) as well.
Moreover, in the parameter-free case, we can easily present a simple axiomatization
and a complete semantics for L

.
THEOREM 2.7.27 (Axiomatization of L

). Let L be a logic with a nitary presentation

/o and let be a p-protodisjunction satisfying (C

), (I

), and (A

). Then L

is
axiomatized by /o

[ R /o.
Proof. Let

L denote the logic axiomatized by /o
t
= /o

[ R /o (note
that this set is closed under all substitutions because we assume to be parameter-free).
First observe that for each R /o
t
we have R

L (obviously if R /o
t
, otherwise
we use (A

)) hence we can use Theorem 2.7.15 to obtain that has the PCP in

L and
thus L

L. On the other hand clearly /o L

and thus also R

for each
R /o (by Theorem 2.7.15 which we can use because Proposition 2.7.26 tells us that
L

is nitary) and so

L L

.
Chapter II: A General Framework for Mathematical Fuzzy Logic 153
THEOREM 2.7.28 (Semantics of L

). Let L be a nitary logic and a protodisjunc-

tion. Dene MOD

p
(L) = A, F MOD

(L) [ F is -prime. Then:

L
= [=
MOD

p
(L)
.
Proof. We prove that [=
MOD

p
(L)
is the least extension of L where is a disjunction.
It is clear that it is an extension of L and that is a disjunction there (because of claim
4 in Theorem 2.7.23 and the fact that MOD

(L

) MOD

(L)). Finally, assume

that L
t
is another extension of L where is a disjunction and [=
MOD

p
(L)
, then
[=
MOD

p
(L

)
, and so
L
.
Let us recall that matrices can be regarded as rst-order structures where the lter
corresponds to a unary predicate F, i.e. all atomic formulae in the corresponding clas-
sical rst-order language are of the form F() where is a formula. A set of positive
clauses ( =
_

C
F() [ C ( is said to be valid in a matrix M = A, F, writ-
ten as M [= (, if for each C ( and each M-evaluation e there is a
C
such that
e() F. A positive universal class of matrices is the collection of all models of a set
of universal closures of positive clauses.
11
The next theorem presents an axiomatization,
by means of a p-disjunction, of any logic given by a positive universal class of matrices.
THEOREM 2.7.29 (Axiomatization of the logic of a positive universal class of matri-
ces). Let L be a nitary logic with a p-disjunction and ( a set of positive clauses.
Then:
[=
BMOD

(L) ] B]=C]
= L +
_

1
[

1
F() (.
Proof. Let us rst denote the formula
1
as C

for a clause C =
_
1
F() (
and observe that for each matrix M = A, F we have: if M as a rst-order structure
satises C, then M as a matrix satises the propositional formula C

. Moreover, if F
is -prime, the reverse implication holds as well.
Let us denote the left-hand side logic as L
h
and the right-hand side one as L
a
.
Clearly L L
h
and due to the observation above also
L
h C

for each C (. Thus

L
a
L
h
.
The converse direction L
h
L
a
will be proven counterpositively. Assume that
there is a set of formulae such that: ,
L
a . Since is a p-disjunction in
L
a
(by Corollary 2.7.17), we know that there is A = A, F MOD

(L
a
) where
F is -prime (by Theorem 2.7.23) such that ,[=
A
. If we show that A [= (, the
proof is done. Assume that A ,[= C for some C (. Then, by the observation at the
beginning of the proof, Aas matrix would not satisfy the propositional formula C

a
contradiction.
As a consequence we obtain a general method to axiomatize the intersection of
axiomatic extensions of a common base logic by means of a generalized disjunction.
11
Positive universal classes are usually dened as the collection of all models of a set of positive universal
formulae, i.e. the universal closures of formulae built from atoms using conjunction and disjunction. Clearly
each formula of this kind can be written as the universal closure of a conjunction of positive clauses and so
its generated positive universal class is just the positive universal class generated by the collection of these
positive clauses.
154 Petr Cintula and Carles Noguera
THEOREM 2.7.30 (Axiomatization of intersections of axiomatic extensions). Let L be
a nitary logic with a p-disjunction , and let L
1
, L
2
be axiomatic extensions of L
respectively given by the sets of axioms /
1
and /
2
(without loss of generality we can
assume that /
1
and /
2
share no propositional variables). Then:
L
1
L
2
= L +
_
[ /
1
, /
2
.
Proof. It is easy to see that L
1
L
2
= [=
MOD

(L
1
)MOD

(L
2
)
. Consider the set of
positive clauses ( = F() F() [ /
1
, /
2
.
If we show that MOD

(L
1
) MOD

(L
2
) = B MOD

(L) [ B [= (,
the proof is done by Theorem 2.7.29. One inclusion is trivial. We prove the con-
verse one counterpositively: consider A MOD

(L) such that A / MOD

(L
1
)
MOD

(L
2
), i.e. there is
i
/
i
such that ,[=
A

i
; consider evaluations e
i
witnessing
those facts. As
1
and
2
do not share any propositional variable there is an evaluation
e witnessing both facts (e is dened as e
i
in the variables occurring in
i
and arbitrarily
elsewhere). This evaluation also shows that A ,[= F(
1
) F(
2
).
As another consequence of Theorem 2.7.29 the following example shows that in
some cases one can easily axiomatize the logic dened by linearly ordered models of
a given logic in terms of disjunction. Logics complete with respect to linearly ordered
matrices are the central topic of the next section.
EXAMPLE 2.7.31. Since is a disjunction in FL
ew
and any matrix MMOD(FL
ew
)
is linearly ordered iff M [= F( ) F( ), we can apply Theorem 2.7.29
and obtain:
[=
BMOD

(FL
ew
) ] B is linearly ordered]
= FL
ew
+ ( ) ( ).
3 Semilinear logics
This section is devoted to the central topic of the chapter: logics complete with
respect to linearly ordered matrices, which we call semilinear. We aim to encompass by
this notion the vast majority of fuzzy logics in the literature. Always in the context of
weakly implicative logics, in the rst subsection we provide the technical denition of
semilinear logic and some auxiliary notions which allowfor an approach to a large extent
analogous to that we have followed for disjunctions. In the second subsection, we study
the strong interplay between semilinear implications and disjunctions and obtain several
interesting consequences. In the last subsection we focus on completeness properties
with respect to ner semantics, i.e. distinguished subclasses of linearly ordered matrices.
3.1 Basic denitions, properties, and examples
We want to study a general notion of logic complete with respect to a semantics
of linearly ordered matrices. A natural design choice is to restrict to the context of
weakly implicative logics presented in the previous section because in these logics the
implication connective induces a preorder in the matrices, which is actually an order
relation in reduced models.
Chapter II: A General Framework for Mathematical Fuzzy Logic 155
DEFINITION 3.1.1 (Linear lter and linear model). Let L be a weakly implicative logic.
Take any non-trivial A = A, F MOD(L). The lter F is called linear if
A
is a
total preorder, i.e. for every a, b A, a
A
b F or b
A
a F. Furthermore, we
say that A is a linearly ordered model (or just a linear model) if
A
is a linear order
(equivalently: F is linear and A is reduced). We denote the class of all linear models
as MOD

(L).
Now, based on linear models, we can introduce the central concept of this section:
DEFINITION 3.1.2 (Semilinear implication, semilinear logic). Let L be a weakly im-
plicative logic. We say that is a semilinear implication if the linear models it denes
are a complete semantics for L, i.e.
L
= [=
MOD

(L)
. A weakly implicative logic is
semilinear if it has a semilinear implication.
Observe that the class of linear models and the notion of semilinearity thereof are
not intrinsically dened for a given logic: they depend on which possible implication
has been chosen as principal. For instance, in classical logic both and are weak
implications, but only makes the logic semilinear (linear models w.r.t. to are the
trivial model and that based on the two-element Boolean algebra, while the only linear
model w.r.t. is the trivial one).
The following simple lemma has an important corollary and will be used later to
provide some useful counterexamples.
LEMMA 3.1.3. Let L be a weakly implicative logic, Aan /-algebra, and F Ti
L
(A)
a linear lter. Then the set [F, A] = G Ti
L
(A) [ F G is linearly ordered by
inclusion.
12
Proof. Assume that there are two incomparable lters G
1
, G
2
[F, A] and take ele-
ments a
1
G
1
G
2
and a
2
G
2
G
1
. Assume (without a loss of generality) that
a
1

A,F)
a
2
. Thus also a
1

A
a
2
F G
1
and so by (MP) also a
2
G
1
a
contradiction.
PROPOSITION 3.1.4. Let L be a weakly implicative logic. Then, all linear lters are
nitely -irreducible, and thus MOD

(L) MOD

(L)
RFSI
.
Proof. If Ais an /-algebra and F Ti
L
(A) is a linear lter, by the previous lemma we
know that [F, A] is linearly ordered by inclusion. Assume that F = G
1
G
2
, for some
G
1
, G
2
Ti
L
(A). Then we must have G
1
G
2
, and hence F = G
1
, or G
2
G
1
, and
so F = G
2
; therefore F is nitely -irreducible. The second claim follows immediately
from Theorem 2.3.14.
In particular, linear theories are linear lters over the algebra Fm
1
, and they are
nitely -irreducible. Recall (from Corollary 2.3.9) that in nitary logics the nitely
-irreducible theories form a base of the closure system Th(L). Thus, an interesting
question is to determine under which conditions linear theories form a base of Th(L).
12
Observe that if L is algebraically implicative and A, F) MOD

(L), then Ti
L
(A) = [F, A].
156 Petr Cintula and Carles Noguera
We can characterize (see [18]) it by means of a generalization of the so-called prelin-
earity property which we rename to semilinearity property. This change of terminol-
ogy follows the tradition from Universal Algebra of calling a class of algebras semiX
whenever its subdirectly irreducible members have the property X because, indeed, as
will see in Theorem 3.1.8, nitary semilinear logics are characterized as those where all
subdirectly irreducible models are linearly ordered.
DEFINITION 3.1.5 (Linear Extension Property, Semilinearity Property). We say that a
weakly implicative logic L has the
Linear Extension Property LEP if linear theories form a base of Th(L), i.e. for
every theory T Th(L) and every formula Fm
1
T, there is a linear theory
T
t
T such that / T
t
.
Semilinearity Property SLP if the following meta-rule is valid:
,
L
,
L


L

Next we prove a transfer theorem for the SLP. Recall our standing assumption that
the set Var of propositional variables is denumerable.
THEOREM 3.1.6 (Transfer of SLP). Assume that a weakly implicative logic L satises
the SLP. Then for each /-algebra Aand each set X a, b A the following holds:
Fi(X, a b) Fi(X, b a) = Fi(X).
Proof. To prove the non-trivial direction we show that for each t / Fi(X) we have
t / Fi(X, a b) or t / Fi(X, b a). We distinguish two cases based on the
cardinality of A.
1) Firstly assume that A is countable. We can assume that the set Var of proposi-
tional variables contains (or is equal to) the set v
z
[ z A (where v
z
,= v
w
whenever
z ,= w). Consider the following set of formulae:
= v
z
[ z Fi(X)
_
c,n)1
c(v
z
1
, . . . v
z
n
) v
c
A
(z
1
,...z
n
)
[ z
i
A.
Clearly, -
L
v
t
(because A, Fi(X) MOD(L) and for the A-evaluation e(v
z
) = z:
e[] Fi(X) and e(v
t
) , Fi(X)). Thus by the SLP we have , v
a
v
b
-
L
v
t
or
, v
b
v
a
-
L
v
t
. Assume (without loss of generality) the former case and denote
T
t
= Th
L
(, v
a
v
b
). We show that the mapping h: A Fm
1
/T
t
dened as
h(z) = [v
z
]
T
is a homomorphism by a simple chain of equalities:
h(c
A
(z
1
, . . . , z
n
)) = [v
c
A
(z
1
,...,z
n
)
]
T

= [c(v
z
1
, . . . v
z
n
)]
T

= c
Fm
L
/T

([v
z
1
]
T
, . . . [v
z
n
]
T
)
= c
Fm
L
/T

(h(z
1
), . . . h(z
n
)).
Chapter II: A General Framework for Mathematical Fuzzy Logic 157
Thus F = h
1
([T
t
]) Ti
L
(A) and, since clearly X a b F and t , F, we
have established that t / Fi(X, a b).
2) Secondly assume that A is uncountable. We introduce a new set of propositional
variables
13
Var
t
= v
z
[ z A; we can safely assume that it contains the original
set Var. We dene a new logic L
t
in the language /
t
which has the same connectives
as / and variables from Var
t
. If we show that this logic has the SLP we can repeat the
constructions from the rst part of this proof. From our assumption we know that there
is a presentation /o of L such that each of its rules has countably many premises.
Let us dene /o
t
= [X] () [ X /o and is an /
t
-substitution
and L
t
=
,S
. Observe that
L
iff there is a countable set
t
such that

t

L
(clearly any proof in /o
t
has countably many leaves, because all of its rules
have countably many premises). Next observe that L
t
is a conservative expansion of L
(consider the substitution sending all variables from Var to themselves and the rest to
a xed p Var, take any proof of from in /o
t
and observe that the same tree with
labels replaced by is a proof of from in L).
We show that L
t
has the SLP: assume that ,
L
and ,
L
.
There is a countable subset
t
such that
t
,
L
and
t
,
L
.
Consider the set Var
0
of variables occurring in
t
, , and a bijection g on the set
Var
t
such that the image Var
0
is a subset of Var (such bijection clearly exists). Thus
for the /
t
-substitution induced by g exists an inverse substitution
1
and [
t
]
, , Fm
1
. Clearly also [
t
],
L
and [
t
],
L

. Using the fact that L

t
expands L conservatively, we obtain [
t
],
L

and [
t
],
L
. From the SLP of L we know that [
t
]
L
and
[
t
]
L
so by structurality for the inverse substitution
1
also
t

L
.
THEOREM 3.1.7 (Properties of semilinear logics). Let L be a semilinear logic in lan-
guage /. Then:
1. L has the LEP,
2. L has the SLP,
3. L has the transferred SLP, i.e. Fi(X, a b) Fi(X, b a) = Fi(X) for each
/-algebra Aand each set X a, b A,
4. linear lters coincide with nitely -irreducible ones in each /-algebra,
5. MOD

(L)
RFSI
= MOD

(L),
6. MOD

(L)
RSI
MOD

(L).
Proof. If T -
L
, then there is a B = A, F MOD

(L) and a B-evaluation e such

that e[T] F and e() , F. We dene T
t
= e
1
[F]. Obviously T
t
is a theory, T T
t
13
Notice that this set of variables is not countable, so it does not satisfy the cardinality restriction that we
have assumed from the beginning of the chapter for the sake of simplicity. However, an inspection of the
relevant parts of the general theory of logical calculi that we have introduced so far shows that everything
needed for this proof would work as well without that restriction. Therefore, in this proof we can violate our
general assumption without problems.
158 Petr Cintula and Carles Noguera
and T
t
-
L
. Since
B
is a linear order, e()
B
e() or e()
B
e() for each
and . Thus either e( ) F or e( ) F, i.e. T
t
or T
t
.
The second claim: if T -
L
, then (using the LEP) there is a linear theory T
t
T,
such that T
t
-
L
. Assume (without a loss of generality) that T
t

L
, then
obviously T, -
L
.
The third claim: follows from the SLP using the transfer theorem proved above.
The fourth claim: let A be an /-algebra. One direction is Proposition 3.1.4. The
other one obviously holds for F = A and otherwise follows from the previous claim
counterpositively: assume that there are a, b A such that a b / F and b a / F,
i.e. F _ Fi(F, a b) and F _ Fi(F, b a). Therefore we obtain F = Fi(F) =
Fi(F, a b) Fi(F, b a) and so F is nitely -reducible.
The fth claim: it is an easy corollary of the previous claim and Theorem 2.3.14.
The nal claim is a trivial consequence of the previous one.
Observe that, in fact, only the LEP has been proved directly from semilinearity;
all the remaining claims of the theorem above have been shown from their direct pre-
decessors. Now, an obvious question is when these claims are equivalent. First notice
that claim 5 tells us that semilinear logics are complete w.r.t. MOD

(L)
RFSI
, a known
property of nitary logics established in Theorem 2.3.16, where even more is shown:
completeness w.r.t. MOD

(L)
RSI
, which is exactly the property needed for the nal
implication in the proof (namely, that claim 6 implies semilinearity). However, this
property is rather obscure, and hence we choose to formulate the following characteri-
zation theorem in terms of nitarity.
THEOREM 3.1.8 (Characterization of semilinear logics). Let L be a weakly implicative
logic. The following are equivalent:
1. L is semilinear,
2. L has the LEP.
Furthermore, if L is nitary the list of equivalences can be expanded with:
3. L has the SLP,
4. L has the transferred SLP,
5. linear lters coincide with nitely -irreducible ones in each /-algebra,
6. MOD

(L)
RFSI
= MOD

(L),
7. MOD

(L)
RSI
MOD

(L).
Proof. All we have to do is to prove two implications. One of them, 7 implies 1, is an
immediate consequence of Theorem 2.3.16; we show the nal one: 2 implies 1. Assume
that -
L
, let T be the theory generated by and T
t
T a linear theory such that
T
t
-
L
. From part 3 of Lemma 2.2.9 we know LindT
T
MOD

(L) and its second

part trivially entails that [T
t
] is a linear lter, i.e. LindT
T
MOD

(L). The rest of

the proof is the same as the proof of the completeness theorem 2.2.13.
Chapter II: A General Framework for Mathematical Fuzzy Logic 159
The previous theorems have several interesting and important corollaries. The rst
one uses the trivial observation that , and for regular implications
also , . Thus, by the SLP, we derive .
COROLLARY 3.1.9. Every regularly implicative semilinear logic is also Rasiowa-
implicative.
Another interesting corollary is obtained by observing that the LEP of a logic is
preserved in all its axiomatic extensions.
COROLLARY 3.1.10. All axiomatic extensions of a semilinear logic are semilinear too.
This corollary will be particularly useful when presenting a large class of weakly
implicative logics which are not semilinear no matter which weak implication we might
take as principal.
It is quite easy to show that a given logic is not semilinear for a xed principal
implication. Consider e.g. intuitionistic logic with its usual implication: the well-known
fact that the linear Heyting algebras do not generate the variety of Heyting algebras
does the job. The next example uses again our characterization theorem (together with
Lemma 3.1.3) to show much more:
EXAMPLE 3.1.11. Intuitionistic logic is not semilinear w.r.t. any principal implication.
Proof. We provide two alternative proofs of this fact. First a very simple ad hoc one,
and then a more sophisticated proof using the machinery introduced in the present chap-
ter which has the advantage of providing a general method to show undenability of
semilinear implications in many logics.
1. Let IPC be the intuitionistic propositional logic. Assume that
t
is a semilin-
ear implication in IPC and we show that p
t
q
IPC
p q (where is
the usual implication of intuitionistic logic), which entails that is a semilinear
implicationa contradiction. One direction is simple: from p, p
t
q
IPC
q we
obtain (using Deduction Theorem) p
t
q
IPC
p q. The reverse direction:
using the rst direction we obtain q
t
p, p q
IPC
q p. Since, trivially,
q
t
p, p q
IPC
p q and the symmetrizations of any pair of weak implica-
tions are interderivable (Corollary 2.2.4), we obtain q
t
p, p q
IPC
p
t
q.
Now, using the following trivial fact p
t
q, p q
IPC
p
t
q and the SLP,
we conclude that p q
IPC
p
t
q.
2. IPC can be presented as FL
ewc
, and hence it is a substructural logic in the scope
of Theorem 2.5.11. In fact, it is a Rasiowa-implicative logic and MOD

(IPC) =
A, 1
A
[ A HA, where HA is the variety of Heyting algebras. The
isomorphism (see Theorem 2.4.5) between lters and congruences in any Heyt-
ing algebra A tells us that 1
A
is -irreducible in Ti
IPC
(A) if, and only if,
the identity relation is -irreducible in Con(A), i.e. Ais subdirectly irreducible.
Thus, MOD

(IPC)
RSI
= A, 1
A
[ A HA
SI
. Assume now, in search of
a contradiction, that
t
is a weak implication in IPC, MOD

(IPC) is the class

160 Petr Cintula and Carles Noguera
of its linear models and IPC is complete w.r.t. MOD

(IPC). By Theorem 3.1.8,

we have A, 1
A
[ A HA
SI
MOD

(IPC). Now, it is sufcient to

consider a subdirectly irreducible Heyting algebra where the natural lattice order
is not linear (it is well-known that these algebras exist) and it will have two in-
comparable lters (IPC-lters are known to be the same as lattice lters over the
Heyting algebra). Then Lemma 3.1.3 gives the contradiction.
Combining this example and the previous corollary we obtain:
PROPOSITION 3.1.12. If L is a substructural logic that can be axiomatically extended
to IPC, then it is not semilinear w.r.t. any principal implication.
Prominent logics falling under the scope of the previous proposition are the follow-
ing: SL
X
and FL
X
for any X e, c, i, o. On the other hand, observe that the second
proof in Example 3.1.11 can be used in many other weakly implicative nitary logics L:
all one needs is to nd a member of MOD

(L)
RFSI
whose algebra admits two incom-
parable logical lters. For instance, consider now the variety V of pointed residuated
lattices generated by the symmetric rotation (see this construction e.g. in [47, 63]) of all
Heyting algebras. Clearly, its corresponding logic has an involutive negation, that is, it
proves . Reasoning exactly in the same way as before, we can show that this
logic is not semilinear w.r.t. any principal implication and thus the same holds for any
substructural logic whose algebraic semantics contains V. A particular case of that is
Girards Linear logic (without exponentials).
At the end of this subsection we present another corollary of Theorem 3.1.8 that
shows that semilinearity of implications is preserved under intersections of logics and
discuss some of its consequences.
COROLLARY 3.1.13. The intersection of a family of semilinear logics in the same
language is a semilinear logic.
Proof. Let J be a family of semilinear logics and

L its intersection. We show that

L has
the LEP. Let T be a theory in

L and , T, i.e. T -

L
. Thus there has to be a logic
L J such that T -
L
, i.e. , Th
L
(T). Thus by the LEP of L there is a linear theory
T
t
in L such that T
t
Th
L
(T) T and , T
t
. Since T
t
is also a theory in

L, the
proof is done.
On the other hand, the inconsistent logic is trivially semilinear. Thus, the following
denition is sound:
DEFINITION 3.1.14 (Logic L

). Given a weakly implicative logic L, we denote by L

the least semilinear logic extending L.

In the next subsection we will give a method to axiomatize L

. However, it is very
simple to determine a complete semantics for this logic, as described in the following
straightforward proposition.
PROPOSITION 3.1.15. Let L be a weakly implicative logic. Then L

= [=
MOD

(L)
and MOD

(L

) = MOD

(L).
Chapter II: A General Framework for Mathematical Fuzzy Logic 161
Moreover, nitarity is preserved when taking the least semilinear extension:
PROPOSITION 3.1.16. If L is a nitary weakly implicative logic, then so is L

.
Proof. Recall the notion of nitary companion of a logic S, denoted as T((S), which
is the largest nitary logic contained in S. Thus, since L is nitary, we know that L
T((L

) L

. If we show that T((L

) is semilinear, we obtain T((L

) = L

and hence
L

is nitary. Actually, one can easily show, by checking that the SLP is preserved, that
semilinearity is preserved in general when taking the nitary companion of a logic.
Note that the proof of the previous theorem also says that if L is nitary, then L

is
the intersection of all its nitary semilinear extensions.
3.2 Disjunction and semilinearity
In this subsection we consider the relationships between p-disjunctions, semilinear
implications and their related properties. In particular, provided that a couple of sim-
ple syntactic conditions are satised, we will see that a logic is p-disjunctional iff it is
semilinear, we will obtain axiomatizations for L

logics, and new characterizations of

p-disjunctions.
We start by demonstrating that for a p-disjunctional logic we can very easily ax-
iomatize its least semilinear extension (in particular, we show that it is an axiomatic
extension).
THEOREM 3.2.1 (Axiomatization of L

). Let L be a nitary p-disjunctional weakly

implicative logic. Then L

is the extension of L with the axiom(s):

(P

)
L
( )( ).
Proof. Using Proposition 3.1.15 we know that L

= [=
MOD

(L)
. The proof is com-
pleted by Theorem 2.7.29; we only need to observe that a matrix A MOD

(L) iff
A [= P, where P is the positive clause F( ) F( ).
The axiom(s) (P

) is (are) called the prelinearity axiom(s). Observe that these

axioms holds in a semilinear logic for an arbitrary p-protodisjunction (using SLP to

L
( )( ) and
L
( )( )). As we have seen
in Subsection 2.7, substructural logics over FL are typical examples of nitary weakly
implicative p-disjunctional logics. This allows us to achieve some well-known results
(collected in Table 6) as corollaries of the theorem above.
Later in this subsection (Example 3.2.7) we will see alternative axiomatizations of
FL

e
and in the next chapter we will substantially simplify the axiomatic system for FL

.
A natural question is how to axiomatize the least semilinear extensions of logics which
are not p-disjunctional or where the p-disjunction is unknown. Of course, a rst idea is
to choose a suitable p-protodisjunction and extend this logic into L

and then proceed

by the previous theorem. But it is not so simple: how would we know that L

? In
order to overcome this problem, we introduce a pair of consecutions which play a kind
of dual r ole to (P

):
(MP

) ,
L
and ,
L

162 Petr Cintula and Carles Noguera
L axiom(s) needed to axiomatize L

FL
1
( )
2
( ) [
1
,
2
iterated conjugates
FL
e
(( ) 1) (( ) 1)
FL
ew
( ) ( )
IPC

[( ) ( )] ( ), [( ) ( )] ( )
Table 6. Axiomatization of notable substructural semilinear logics
PROPOSITION 3.2.2. (MP

) is satised in:
any logic for any satisfying the PCP.
any substructural (not necessarily lattice-disjunctive!) logic for = .
Proof. The rst claim is simple (from , and , ). To prove the
second one observe that any substructural logic proves: .
The introduced consecutions (P

) and (MP

) are indeed natural binding condi-

tions for implication and disjunction, as shown by the next lemma and theorem.
LEMMA 3.2.3. Let L be a weakly implicative logic in /, a p-protodisjunction, and
Aan /-algebra.
If L fulls (MP

), then each linear lter in Ais -prime.

If L fulls (P

), then each -prime lter in Ais linear.

Proof. The rst claim: assume that F is linear (a
A
b F or b
A
a F) and
a
A
b F. Thus from (MP

) we obtain that b F or a F.
The second claim: assume that F is not linear, i.e. there are elements a, b such that
x = a
A
b , F and y = b
A
a , F. x
A
y = (a
A
b)
A
(b
A
a) F
because L satises (P

), thus F is not -prime.

THEOREM 3.2.4 (Interplay of p-disjunctions and semilinearity). Let L be a nitary
weakly implicative logic. The following are equivalent:
L is p-disjunctional and satises (P

),
L is semilinear and satises (MP

).
Thus in particular:
1. Let L be a weakly implicative logic satisfying (P

) and (MP

). Then, L is
semilinear iff L is p-disjunctional.
2. Let L be a p-disjunctional weakly implicative logic. Then, L is semilinear iff L
satises (P

).
Chapter II: A General Framework for Mathematical Fuzzy Logic 163
3. Let L be a semilinear logic and a p-protodisjunction. Then, L is p-disjunctional
iff L satises (MP

).
Proof. Top-to-bottom implication: we know that each nitary p-disjunctional logic sat-
ises (MP

) and has the PEP (Theorem 2.7.23) and, since from (P

) we know that
-prime theories are linear, we obtain the LEP and the proof is done. The second
direction is analogous.
Notice that claim 1 of this theorem provides many additional characterizations of
semilinearity by the means of Theorems 2.7.15 and 2.7.23 in a broad class nitary logics
satisfying (P

) and (MP

). Similarly, notice that claims 2 and 3 reduce, in huge classes

of logics, the validity of a meta-rule, SLP or PCP, to the validity of a simple rule, (P

)
or (MP

).
This theorem has two interesting corollaries. First, we can use Corollary 2.7.17 to
extend Corollary 3.1.10 from axiomatic extensions to axiomatic expansions.
COROLLARY 3.2.5. Let L
1
be a p-disjunctional weakly implicative logic and let L
2
be
an axiomatic expansion which is weakly implicative with the same principal implication.
If L
1
is semilinear, then so is L
2
.
Second, we can return to our original goal of providing an axiomatization for L

.
Recall that by L

we denote the weakest logic extending L where is a p-disjunction

(see Theorem 2.7.27 for an axiomatization of this logic).
COROLLARY 3.2.6. Let L be a nitary weakly implicative logic and a p-proto-
disjunction satisfying (MP

). Then L

is the extension of L

by (P

).
Proof. Since L

+ (P

) is an axiomatic extension of L

, remains a p-disjunction
there (by Corollary 2.7.17). Thus, by Theorem 3.2.4, it is a semilinear logic.
Let L
t
be a nitary semilinear extension of L. Clearly L
t
satises (MP

) as well
and thus by Theorem 3.2.4 it is a p-disjunctional logic. Thus L

L
t
and, since the
Theorem 3.2.4 also tells that L
t
satises (P

), the proof is done.

The restriction of this corollary to p-disjunctional logics (which, of course, satisfy
(MP

) and L

= L) gives an alternative proof of Theorem 3.2.1. As we have seen,

substructural logics with in the language form a big class of logics satisfying (MP

).
Let us summarize some consequences of the previous claims for these logics and add
some more interesting ones. In the beginning of this subsection we have seen one way
to axiomatize L

: identify a good p-disjunction and add the prelinearity axioms for this
p-disjunction. The previous corollary provides an alternative way for substructural log-
ics with in the language which produces less elegant axiomatizations, but can be seen
as more robust, because it does not require to identify a p-disjunction in L. We simply
extend L into L

(just by adding the -forms of all rules, see Theorem 2.7.29) and then
we add prelinearity written using .
EXAMPLE 3.2.7. FL

e
is the extension of FL
e
by the rule ( 1) and the
axiom ( )( ). The -form of modus ponens, , ( ) ,
already holds in FL
e
, which is not difcult to prove.
164 Petr Cintula and Carles Noguera
LEMMA 3.2.8. Let L be a lattice-disjunctive substructural logic. Then the following
are equivalent:
(P

)
L
( ) ( )
(lin

)
L
( ) ( ) ( )
(lin

)
L
( ) ( ) ( )
Proof. We prove the equivalence of the rst two claims; the equivalence of the rst and
the third ones is prove analogously. First recall that
L
and so

L
( ) ( ). The proof is completed by (1) and the PCP.
The other direction: from we obtain ( ) ( ). The
rest is simple.
PROPOSITION 3.2.9. Let L be a nitary substructural logic with in its language.
Then:
L is semilinear iff it is lattice-disjunctive and satises (P

),
L

is the extension of L

by any of these axioms: (P

), (lin

) or (lin

).
The next proposition is a generalization of Example 2.7.10.
PROPOSITION 3.2.10. Let L be a nitary Rasiowa-implicative substructural semilin-
ear logic. Then the set = (p q) q, (q p) p is a disjunction.
Proof. We can easily show that is a protodisjunction because p
L
(p q) q
((P
SL
4) of Proposition 2.5.5) and q
L
(p q) q (W).
Next we show that satises the PCP. Assume that ,
L
and , . Thus
clearly , ,
L
and so , ,
L
; analogously for .
The SLP completes the proof.
Notice that if the logic from the previous proposition would contain in the lan-
guage, we would obtain (p q) q, (q p) p p q. A question is
whether we could internalize this equivalence. First observe that
t
= (p q) q,
(q p) p would be a disjunction as well. Then we can prove:
PROPOSITION 3.2.11. Let L be a Rasiowa-implicative substructural semilinear logic.
Then:

L
[( ) ] [( ) ],

L
[ & ( )] [ & ( )].
Furthermore the logic L extends SL
e
iff

L
[( ) ] [( ) ].
Proof. The rst claim: Left-to-right direction is simple. The converse one is based on a
simple observation:
L
[( ) ] , a consequence of (symm) and
(P
SL
4) of Proposition 2.5.5.
Chapter II: A General Framework for Mathematical Fuzzy Logic 165
The second claim: clearly & ( ) and & ( ) , thus
[&( )] [&( )] ; the rest is simple. The converse direction: assume
that . Thus (&( )) and so (&( ))(&( )).
The rest easily follows from the SLP.
One direction of the third claim trivially follows from the rst claim. To prove the
converse one, observe that the assumption [( ) ] [( ) ]
entails: (( ) ). Then we obtain [(( ) ) ( )]
[ ( )] (using (Sf)). The proof is completed by another instance of (Sf),
namely: ( ) (( ) ) ( ).
The following proposition is a straightforward corollary of the fact that semilinear
logics are complete with respect to linearly ordered matrices, whose algebraic reducts
are clearly distributive lattices.
PROPOSITION 3.2.12. Let L be a substructural semilinear logic with and in its
language. Then the , -reduct of Ais a distributive lattice for each A ALG

(L).
We end this subsection by using the characterization theorems 2.7.15 and 2.7.23,
and Remark 2.7.24, to demonstrate many equivalent conditions for a connective to be
a disjunction. These theorems were formulated for nitary logics; now we show that in
the case of semilinear logics we can drop this condition.
THEOREM 3.2.13 (Characterizations of (p-)disjunctions in (possibly innitary) semi-
linear logics). Let L be a semilinear logic with a presentation /o and p-protodisjunction
. Then the following are equivalent:
1. L has the PCP w.r.t. ,
2. L proves (MP

),
3. L has the PEP w.r.t. ,
4. L has the sPCP w.r.t. ,
5. satises (C

), (I

), and R

L for each R L,
6. satises (C

), (I

), and R

L for each R /o,

7. Fi(Y ) Fi(Z) = Fi(Y
A
Z) for each /-algebra Aand each Y, Z A
8. Fi(X, x) Fi(X, y) = Fi(X, x
A
y) for each /-algebra Aand each
X x, y A.
Moreover, if is parameter-free, we can add:
9. For every /-algebra Aand every F Ti
L
(A), F is nitely -irreducible iff it is
-prime,
10. for every A, F MOD

(L), A, F MOD

(L)
RFSI
iff F is -prime,
11. for every Fm
1
,
L
if, and only if, [=
MOD

p
(L)
.
166 Petr Cintula and Carles Noguera
Proof. Recall that L satises (P

). The implications 12 and 23 are trivial (recall

that each semilinear logic has the LEP).
34: assume that ,
1

2
-
L
, then using the PEP there is a -prime theory
T
1

2
such that T -
L
. If
1
T we know that ,
1
-
L
. Assume
otherwise, i.e. there is
1
T. Then, since
2
T and T is -prime, we
obtain
2
T and so ,
2
-
L
.
From Remarks 2.7.16 and 2.7.19 we know that conditions 4, 5, 6, and 7 are equiva-
lent and because implications 78 and 81 are trivial, we have established the equiv-
alence of the rst eight claims.
19: On one hand, we know that linear and -prime lters coincide; on the other
hand, from Proposition 3.1.4, we know that linear lters are nitely -irreducible. From
8 we have that Fi(F, x
A
y) Fi(F, y
A
x) = Fi(F, (x
A
y)
A
(y
A
x)) =
F and so we obtain that nitely -irreducible lters are linear.
910: direct consequence of Theorem 2.3.14.
1011: we know that L is complete w.r.t. MOD

(L), thus by Proposition 3.1.4 it

is also complete w.r.t. MOD

(L)
RFSI
, and the claim follows from 10.
The nal implication 111 is proved in the same way as the corresponding impli-
cation in Theorem 2.7.23.
COROLLARY 3.2.14. Let L
1
be a semilinear logic with a p-protodisjunction which
satises (MP

) and L
2
a nitary weakly implicative expansion of L
1
by a set of conse-
cutions (. Then, L
2
is semilinear if, and only if, R

L
2
for each R (.
Proof. Observe from the assumption we obtain that L
i
satises (MP

) and (P

) for
i = 1, 2 and R

L
1
for each R L
1
.
Because L
2
is semilinear we can use Theorem 3.2.13 to complete the proof of one
direction. The converse direction: we know that L
2
has an axiomatic system closed
under the formation of -forms of its rules. Therefore, since it is nitary, we know that
it is p-disjunctional (Theorem 2.7.15) and so Theorem 3.2.4 completes the proof.
EXAMPLE 3.2.15. We show that the logic MTL
Z
(introduced in Chapter I) is semilin-
ear. Recall that this logic is an expansion of MTL (i.e. the logic FL

ew
) by a new unary
connective , by adding the deduction rule
MTL

and some additional axioms

(see Chapter I) which ensure that it remains weakly implicative with the same principal
implication. Clearly the logics MTL and MTL
Z
satisfy the conditions of the previous
corollary, so if we show that p
MTL

p the proof is done.

Using the deduction rule we obtain p
MTL

( p) and then it is enough

to recall that ( ) and are among the axioms of MTL
Z
.
3.3 Strengthening completeness: densely ordered chains
We have proposed semilinear logics as a useful mathematical notion to encompass
and study most examples of fuzzy logics by characterizing them as the logics of lin-
early ordered matrices. However, a genuine item in the agenda of fuzzy logics is that
of looking for ner complete semantics based on some particular kind of linearly or-
dered models such as standard models based on the real unit interval, or models dened
on the rational unit interval, or models over nite chains. In this section we focus on
Chapter II: A General Framework for Mathematical Fuzzy Logic 167
the semantics of densely ordered linear matrices, a common feature of both real- and
rational-valued ones. We will characterize completeness with respect to this semantics
by means of a special kind of lter and some meta-rules, in a similar fashion to what we
have already done for disjunctions and semilinear implications.
DEFINITION 3.3.1 (Dense lter). Let L be a weakly implicative logic and A = A, F
an L-matrix. Then F is called a dense lter if F is linear and for every a, b A such
that
14
a <
A
b there is z A such that a <
A
z and z <
A
b.
A matrix A is called a dense linear matrix if it is reduced and F is dense (equiv-
alently: if
A
is a dense order). The class of all dense linear L-models is denoted as
MOD

(L).
The dense extension property DEP will be dened analogously as the LEP and the
PEP but with some non-trivial changes. Like in the case of disjunctions, where we char-
acterize (in nitary logics) the dening meta-rule PCP by some suitable lter extension
principle, we start with the meta-rule DP which was already introduced in the literature.
Thus, our goal is to provide a corresponding lter extension principle. The problem is
that DP is not structural because it refers to an unused propositional variable. That is
the reason why we will be forced to formulate the DEP only in Lindenbaum matrices,
and not for theories but for some particular sets of formulae. These denitions will still
allow us to obtain a nice interplay between a lter extension principle, a completeness
property, and a logical meta-rule, as in the previous cases.
DEFINITION 3.3.2 (Density Property). Let L be a weakly implicative logic and a
p-protodisjunction. We say that L has the Density Property DP with respect to if for
any set of formulae , , and any variable p not occurring in , , the
following holds: if
L
( p)(p ), then
L
( ).
DEFINITION 3.3.3 (Dense Extension Property). Let L be a weakly implicative logic.
We say that L has the Dense Extension Property DEP if every set of formulae such
that -
L
and there are innitely many variables not occurring in can be extended
into a dense theory T such that T -
L
.
In order to prove the characterization of dense completeness in terms of the DEP,
we need the following technical lemma.
LEMMA 3.3.4. Let L be a weakly implicative logic, A MOD

(L), T a theory,
and a formula. If T ,[=
A
, then there is a countable submatrix A
t
of A such that
A
t
MOD

(L) and T ,[=

A
.
Proof. Clearly Ais non-trivial and so it is innite (because its matrix order is dense).
Let e be any evaluation witnessing T ,[=
A
; we dene K as the subset of A containing
valuations assigned to all subformulae of formulae from T by e. Clearly K is
countable. We dene two sequences K
i
of countable subsets of Aand A
i
of submatrices
of Aas: K
1
= K and for i > 0:
14
By a <
A
b we understand a
A
b and b ,
A
a; note that we are not assuming that Ais reduced, so this
convention is not trivial.
168 Petr Cintula and Carles Noguera
A
i
is the submatrix generated by K
i
. (Clearly each A
i
is countable.)
K
i+1
is any countable dense subset of A containing A
i
.
15
Clearly A
i
is a directed family of reduced matrices and so their union A
t
is in
MOD

(L) (see [25, Theorem 0.7.2.]). Obviously A

t
is a countable submatrix of A
such that A
t
MOD

(L) (because of the construction) and T ,[=

A
(the evaluation
e does the job because K A
t
).
THEOREM 3.3.5 (Characterization of dense completeness). Let L be a weakly implica-
tive logic. Then,
L
= [=
MOD

(L)
iff L has the DEP.
Proof. Right-to-left: we repeat the usual completeness proof via constructing the appro-
priate Lindenbaum-Tarski matrix with an interesting twist to overcome the restrictions
of the DEP.
Let us consider a set of formulae such that -
L
. Let us enumerate
the propositional variables and dene substitutions and
t
by setting: (v
i
) = v
2i
;

t
(v
2i
) = v
i
,
t
(v
2i+1
) = v
i
for each i 0. Observe that
t
= for any formula
. Thus also [] -
L
(otherwise, by structurality,
t
[]
L

t
, i.e.
L
a
contradiction). Notice that there are innitely many variables not occurring in [] and
so we can use the DEP to obtain a dense theory T such that T [] and T -
L
.
Take the matrix A = LindT
T
= Fm
1
/
L
(T), [T], observe that A MOD

(L),
and consider the A-evaluation e() = []
T
. We know that e[T] [T] and e() / [T].
Let us now consider the A-evaluation e
t
() = e() and observe that e
t
() =
e() / [T]. As [] T, we obtain that e
t
[] = e[[]] e[T] [T]. Thus, we
obtain T ,[=
A
.
Left-to-right: consider a set of formulae with innitely many unused variables
and a formula such that -
L
. We can use our assumption to obtain a dense linear
L-matrix A = A, F and an A-evaluation e such that e[] F and e() , F. Without
loss of generality we can assume that A is countable (due to previous lemma). Let us
consider, for any a A, a variable v
a
not occurring in (such variables exist).
Further consider an A-evaluation e
t
such that e
t
(p) = e(p) for variables in and
e
t
(v
a
) = a for a A.
Consider the set of formulae T = [ e
t
() F. Clearly T is a theory, T ,
and , T; it remains to be shown that T is dense in L. Linearity is simple (for each
and , clearly e
t
()
A
e
t
() F or e
t
()
A
e
t
() F). Observe that
<
Fm,T)
iff e
t
() <
A
e
t
(). In this case, since A is dense, there is a A such
that e
t
() <
A
a = e
t
(v
a
) <
A
e
t
(). Thus <
Fm,T)
v
a
and v
a
<
Fm,T)
.
Combining this result with the characterization of semilinearity in terms of LEP
in Theorem 3.1.8 we obtain:
COROLLARY 3.3.6. Let L be a weakly implicative logic. Then, if L has the DEP then
it has the LEP.
15
Consider any two elements a, b A
i
such that a < b and there is no element of A
i
between a and b.
There has to be a set X
a,b
A [a, b], X
a,b
isomorphic to Q. Then we construct A
i+1
with the desired
properties simply by adding all such sets to A
i
.
Chapter II: A General Framework for Mathematical Fuzzy Logic 169
We consider now the relation of the DEP with the density property DP.
LEMMA 3.3.7. Let L be a weakly implicative logic with a p-protodisjunction satis-
fying the (MP

). Then, the DEP implies the DP.

Proof. Assume that -
L
( ) and
L
( p)(p ) for some
variable p not occurring in , , , . From the rst assumption we know that there
is a formula (( )) such that -
L
. Thus there is a dense linear
L-matrix A = A, F and an A-evaluation e such that e[] F and e() , F.
Clearly e() , e() (otherwise e( ) F and so e() F) and so e() < e()
(because A is linear). Since A is dense, there is an element e() < a < e(). Take
evaluation e
t
(v) = a for v = p and e(v) otherwise (clearly e
t
[] F). Then for the
elements
1
= e
t
( p) and
2
= e
t
(p ) we have e
t
(
i
) , F. Notice that also
e
t
() , F (otherwise e() F). Using Lemma 3.2.3 we know that F is also a -prime
lter and so e
t
(
1
)e
t
(
2
)e
t
() , F. Therefore, -
L
( p)(p )a
contradiction.
For nitary logics, and in the absence of parameters in , we can prove the equiva-
lence of the DEP and the DP, and thus give the main result of this subsection: the syn-
tactical characterization of completeness with respect to dense linear models by means
of a meta-rule.
THEOREM 3.3.8 (Characterization of dense completeness). Let L be a nitary semilin-
ear disjunctional logic. Then the following are equivalent:
1.
L
= [=
MOD

(L)
.
2. L has the DEP.
3. L has the DP.
Proof. It is enough to prove that 3 implies 2. Consider a set of formulae such that
-
L
and there are innitely many variables not occurring in . Let us enumerate all
pairs of formulae. We introduce two sequences of sets of formulae
i
and A
i
such that

i
-
L
A
i
. We start with
0
= and A
0
= . Given any i > 0 consider the following
cases:
If
i
,
i

i
-
L
A
i
, then we dene
i+1
=
i

i

i
and A
i+1
= A
i
.
If
i
,
i

i

L
A
i
, then we dene
i+1
=
i

i

i
and A
i+1
=
A
i
(
i
p)(p
i
) for some variable p not occurring in
i
A
i

i
,
i

(since there are innitely many variables not occurring in , we can nd in each
step some unused one; notice that this would be no longer true if would be
parameterized).
Now we prove by induction that
i
-
L
A
i
for every i. When i = 0 it is true
by assumption. For the induction step, if we have proceeded by the rst case, clearly

i+1
-
L
A
i+1
. Otherwise, assume, by the way of contradiction, that
i+1

L
A
i+1
, i.e.
that
i
,
i

i

L
A
i+1
. As clearly also
i
,
i

i

L
A
i+1
(using the assumption
170 Petr Cintula and Carles Noguera
of the second case and properties of protodisjunction) we can use the SLP to obtain

i

L
A
i+1
. Thus we also have
i

L
A
i
(
i

i
) using DP. Finally observe that
from
i
, A
i

L
A
i
and
i
,
i

i

L
A
i
we can obtain
i
, A
i
(
i

i
)
L
A
i
via the syntactical characterization of p-disjunctions in Theorem 2.7.15. Putting this
together we obtain
i

L
A
i
a contradiction with the induction hypotheses.
Dene T as the L-theory generated by the union of all
i
s. First observe that
T -
L
A
i
for each i (otherwise by nitarity there would be some j such that
j

L
A
i
and so clearly
maxi,j]

L
A
maxi,j]
a contradiction).
Thus T -
L
and from the construction it follows that T is linear. Now assume
that T -
L

i

i
, then we had to proceed via the second case in the construction
(otherwise already
i+1

L

i

i
) thus also T -
L

i
p and T -
L
p
i
(because otherwise T
L
A
i+1
).
Note that the premises of the previous theorem are fullled by any substructural
semilinear logic with in its language. Analogously to the case of L

and L

, we con-
sider now the problem of nding the weakest extension of a logic enjoying completeness
with respect to dense linear models.
LEMMA 3.3.9. Let J be a family of weakly implicative logics in the same language and

L its intersection. If every logic of J has the DEP, then so has

L.
Proof. Let be a set of formulae with innitely many variables not occurring in it and
a formula such that -

L
. Thus there has to be a logic L J such that -
L
.
Thus by the DEP of L there is a dense L-theory T and , T. Since clearly T is
also an

L-theory, the proof is done.
This, together with the fact that any weakly implicative logic has at least one exten-
sion which is complete with respect to its dense linear models (namely the inconsistent
logic), gives the following result:
THEOREM 3.3.10 (The logic L

and its semantics). Let L be a weakly implicative

logic. Then there is the weakest logic extending L which is complete w.r.t. its dense
linear models. Let us denote this logic as L

.
The proofs of the following two results run parallel to those of their analogues in
previous sections.
PROPOSITION 3.3.11. Let L be a weakly implicative logic. Then
L
= [=
MOD

(L)
and MOD

(L

) = MOD

(L).
PROPOSITION 3.3.12. Let L be a nitary weakly implicative logic. Then L

is nitary.
THEOREM 3.3.13 (L

in nitary logics). Let L be a nitary semilinear disjunctional

logic. Then, L

is equal to the intersection of all its extensions satisfying the DP iff this
intersection is nitary and semilinear.
Proof. Let us denote that intersection as

L. One direction is a simple consequence
of Corollary 3.3.6. The converse direction: clearly

L enjoys the DP and is disjunc-
tional (due to Theorem 3.2.4). Thus by Theorem 3.3.8 it has the DEP and so L

L.
Chapter II: A General Framework for Mathematical Fuzzy Logic 171
Lemma 3.3.7 tells us that each extension of L with the DEP has also the DP, thus

L L

.
These results simplify and give a new insight into an approach used in the fuzzy
logic literature to prove dense completeness (presented in Section 4.2 of Chapter III).
Indeed, this approach starts from a suitable proof-theoretic description of a logic L,
which then is extended into a proof-system for the intersection of all extensions of L
satisfying the DP just by adding DP as a rule (in the proof-theoretic terms, not as we
understand rules here). This rule is then shown to be eliminable (using analogs of the
well-known cut-elimination techniques). Thus we can conclude by the previous theorem
that L = L

(or equality of their nite derivations, or equality of their sets of theorems,

depending on how the elimination of DP was done), and hence the original logic is
complete w.r.t. its dense linear models (of course, our general theory is not helpful in
this last step, because here one needs to use specic properties of the logic in question).
3.4 Strengthening completeness: arbitrary classes of chains
We move now to arbitrary semantics of linearly ordered matrices. So far we have
only considered completeness as the equality of two logics, i.e. two structural conse-
quence relations, one syntactically presented, the other semantically dened. However,
it is usual in the literature to study weaker notions of completeness: equality of the
sets of theorems and equality of nite sequents. We formalize this by introducing three
types of completeness properties according to the cardinality of the sets of premises in
the derivations. Later we will obtain characterizations for these properties and relations
between them.
DEFINITION 3.4.1 (Completeness properties). Let L be a weakly implicative semilin-
ear logic and K MOD

(L). We say that L has the property of:

strong K-completeness, SKC for short, when for every set of formulae :

L
if, and only if, [=
K
.
nite strong K-completeness, FSKC for short, when for every nite set of formu-
lae :
L
if, and only if, [=
K
.
K-completeness, KC for short, when for every formula :
L
if, and only if,
[=
K
.
Of course, the SKC implies the FSKC, and the FSKC implies the KC. Our next
aim is to prove characterizations of these properties that will allow, for particular choices
of semilinear logics and classes of linearly ordered models, either to show or to falsify
the corresponding completeness properties. Although a more general approach is possi-
ble, here such characterizations will be obtained by using the algebraizability condition
of the logics.
In reduced matrices of algebraically implicative logics, as have seen, the lters are
equationally denable, and so each reduced matrix is uniquely determined by its alge-
braic reduct. Thus, by a slight abuse of language, we will use the symbol [=
K
(for a
class of L-chains K) not only for the equational consequence but also for the semantical
172 Petr Cintula and Carles Noguera
consequence on the set of formulae given by the corresponding class of matrices. The
confusion cannot happen as it is always clear from the context whether we speak about
formulae or equations (see, e.g., the next theorem).
CONVENTION 3.4.2. From now on, until the end of this subsection, we assume that L
is an algebraically implicative semilinear logic with a principal implication and K a
class of L-chains.
First, we can obtain the following equivalent algebraic properties for each type of
completeness:
THEOREM 3.4.3 (Algebraic characterization of completeness properties). The follow-
ing hold:
1. L has the KC if, and only if, V(ALG

(L)) = V(K).
2. L has the FSKC if, and only if, Q(ALG

(L)) = Q(K).
3. L has the SKC if, and only if, ALG

(L) = ISP
-f
(K).
Proof. Let us prove the rst claim. For the left-to-right direction take an arbitrary equa-
tion . Then: [=
ALG

(L)
iff
L
iff [=
K
iff [=
K
.
Therefore ALG

(L) and Ksatisfy the same equations and hence they generate the same
variety. The other direction is straightforward.
The remaining points are proved analogously using that quasivarieties are charac-
terized by quasiequations, and the classes closed under the operator ISP
-f
are charac-
terized by generalized quasiequations with countably many premises (we can omit this
operator on the left side of the equation because that ALG

(L) is closed under ISP

-f
,
see Proposition 2.3.13).
Observe that in the rst claim of the previous theorem we could have written only
H(ALG

(L)) instead of V(ALG

(L)) and if ALG

(L) is a quasivariety (e.g. if L is

nitary) we could write just ALG

(L) instead of Q(ALG

(L)) in the second claim.

Other useful characterizations of completeness properties are obtained in terms of
embeddability. To present them, we need rst one denition and one lemma.
DEFINITION 3.4.4 (Directed set of formulae). A set of formulae is directed if for
each , there is such that both and are provable in L (we
call an upper bound of and ).
LEMMA 3.4.5. Assume that L is nitary and has the SKC. Then for every set of for-
mulae and every directed set of formulae the following are equivalent:
-
L
for each .
there is a matrix A, F MOD

(L) with A K and an A-evaluation e such

that e[] F and e[] F = .
Chapter II: A General Framework for Mathematical Fuzzy Logic 173
Proof. One direction is obvious. For the other one, rst assume that there exists a propo-
sitional variable v not appearing in . Dene the set
t
= v [ .
We show that
t
-
L
v by the way of contradiction. Assume that
t

L
v. Thus there
are nite sets

and

such that

v [


L
v. Let
denote an upper bound of the formulae in

. Since -
L
, we know that there is a
matrix A, F MOD(L) and an evaluation e such that e[] F and e() / F.
We dene the evaluation e
t
as e
t
(p) = e(p) for each p ,= v and e
t
(v) = e(). Clearly,
e
t
[

v [

] F and e
t
(v) / Fa contradiction.
Now, by the SKC, there are B, G MOD

(L) with B K and e such that

e[
t
] G and e(v) / G. Thus e[] G = (if e() G for some then, since
e[
t
] G, we would obtain e(v) Ga contradiction).
Assume now that uses all propositional variables. In this case we consider
an enumeration of all propositional variables v
n
[ n N and take the substitutions
,
t
dened by (v
n
) = v
n+1
and
t
(v
n+1
) = v
n
(and
t
(v
0
) = v
0
); note that
t

is the identity. Observe that [] -
L
() for each (indeed, if []
L
(), by
structurality we would have
t
[[]]
L

t
(), i.e.
L
). The variable v
0
does not
appear in [] [], so we can apply the previous reasoning to these sets an obtain a
model and an evaluation e as desired.
THEOREM 3.4.6 (Characterization of strong completeness). Assume that L is nitary
and lattice-disjunctive. Then the following are equivalent:
1. L has the SKC.
2. Every non-trivial countable member of ALG

(L)
RFSI
is embeddable into some
member of K.
3. Every countable member of ALG

(L)
RSI
is embeddable into some member of K.
Proof. 12: Take a countable A ALG

(L)
RFSI
and let F be its lter. Consider a
set of pairwise different variables v
a
[ a A (we can do it because A is countable)
and the following sets of formulae:
= c(v
a
1
, . . . , v
a
n
) v
c
A
(a
1
,...,a
n
)
[ c, n / and a
1
, . . . , a
n
A,
= v
a
1
. . . v
a
n
[ n N and a
1
, . . . a
n
A F.
Clearly is directed and -
L
for each . Indeed, take the A-evaluation
e(v
a
) = a; we have e[] F but a
1
. . . a
n
/ F (otherwise, since F is prime,
a
i
F for some ia contradiction).
Nowwe use the SKCand Lemma 3.4.5 to obtain B, G MOD

(L) with B K
and a B-evaluation e such that e[] G and e() / G for each . Consider the
mapping f : A B dened as f(a) = e(v
a
). It is clear that f is a homomorphism from
Ato B. We show that it is one-one: take a, b A such that a ,= b and assume, without
loss of generality, a
A
b / F. Therefore, f(a)
B
f(b) = e(v
a
)
B
e(v
b
) =
e(v
a
A
b
) / G and thus f(a) ,= f(b).
23: Obvious.
174 Petr Cintula and Carles Noguera
31: Suppose that for some and we have -
L
. Then, since L is nitary,
by Theorem 2.3.16, there are A, F MOD

(L)
RSI
and e such that e[] F and
e() / F. Let B be the countable subalgebra of Agenerated by e[Fm
1
]. Consider the
submatrix B, B F MOD

(L). B is not necessarily subdirectly irreducible but

it is representable as a subdirect product of a family of C
i
[ i I ALG

(L)
RSI
;
let G
i
be their corresponding lters and let be the representation homomorphism.
It is clear that e[] B F and e() / B F. There is some j I such that
(
j
)(e()) / G
j
. C
j
is a countable member of ALG

(L)
RSI
, so by the assumption
there is a matrix C, G MOD

(L) with C K and an embedding f : C

j
C,
and hence, using this model and the evaluation f
j
e, we obtain ,[=
K
.
DEFINITION 3.4.7 (Partial embeddability). Given two algebras A and B of the same
language /, we say a nite subset X of A is partially embeddable into B if there is a
one-to-one mapping f : X B such that for each c, n / and each a
1
, . . . , a
n
X
satisfying c
A
(a
1
, . . . , a
n
) X, f(c
A
(a
1
, . . . , a
n
)) = c
B
(f(a
1
), . . . , f(a
n
)).
A class K of algebras is partially embeddable into a class K
t
if every nite subset of
every member of K is partially embeddable into a member of K
t
.
THEOREM 3.4.8 (Characterization of nite strong completeness). Assume that L is a
nitary and lattice-disjunctive logic with a nite language /. Then the following are
equivalent:
1. L has the FSKC.
2. Every countable non-trivial member of ALG

(L)
RFSI
is partially embeddable
into K.
3. Every non-trivial member of ALG

(L)
RFSI
is partially embeddable into K.
4. Every member of ALG

(L)
RSI
is partially embeddable into K.
5. Every countable member of ALG

(L)
RSI
is partially embeddable into K.
Proof. The implications 34 and 45 are trivial; 51 is proved analogously to the
implication 31 of Theorem 3.4.6.
12: Take a countable A ALG

(L)
RFSI
, with lter F, and a nite set B A.
Dene the set B
t
= B a
A
b [ a, b B. Consider a set of pairwise different
variables v
a
[ a A, a formula =
_
aB

\F
v
a
, and the following set of formulae
(notice a difference between this set and the set from the proof of Theorem 3.4.6):
=c(v
a
1
, . . . , v
a
n
)v
c
A
(a
1
,...,a
n
)
[ c, n/ and a
1
, . . . , a
n
, c
A
(a
1
, . . . , a
n
)B
t
.
Observe that is nite and -
L
(use the A-evaluation e(v
a
) = a). Thus, by the
FSKC, there is C K, with lter G, and a C-evaluation e such that e[] G and
e() / G. Dene a mapping f : B C as f(a) = e(v
a
). Obviously f is a partial
homomorphism. We show that f is one-to-one. Take a, b B such that a ,= b. We
know that a
A
b B; also, without loss of generality, we can assume a
A
b / F.
Thus, f(a)
C
f(b) = e(v
a
)
C
e(v
b
) = e(v
a
A
b
) / G (because e() / G) and so
f(a) ,= f(b).
Chapter II: A General Framework for Mathematical Fuzzy Logic 175
23: Take any A ALG

(L)
RFSI
, a nite X A and consider the countable
subalgebra B A generated by X. It is enough to prove that B is nitely subdirectly
irreducible relative to ALG

(L). Let F, G be the lters such that A, F, B, G

MOD

(L) (we know that G = B F and, because the logic is algebraically implica-
tive, Fi
A
(G) = F). Suppose, by the way of contradiction, that G = G
1
G
2
for some
G
1
, G
2
Ti
L
(B) such that G _ G
1
, G
2
. Take b
i
G
i
G. Observe that b
1
, b
2
/ F
and thus F _ Fi
A
(G
i
). By Theorem 2.7.15 we have: G
1
G
2
= Fi
B
(G
1
G
2
) =
G F. Finally we obtain: Fi
A
(G
1
) Fi
A
(G
2
) = Fi
A
(G
1
G
2
) F, which implies
that F is not nitely -irreduciblea contradiction.
REMARK 3.4.9. Notice that the implications from 2, 3, 4, or 5 to 1 hold also for innite
languages, whereas the converse ones do not (as shown by the following example).
EXAMPLE 3.4.10. Consider the language / resulting from /
0
= &, , , , 0, 1
by adding a denumerable set C = c
n
[ n N of new 0-ary connectives. Let G
C
be
the conservative expansion of G odel-Dummett logic in this language with no additional
axioms or rules. It is a semilinear Rasiowa-implicative logic (in fact, it is a core fuzzy
logic; see Chapter I). Let G
C
denote its equivalent algebraic semantics, which in fact is
the variety of G odel algebras with innitely many constants arbitrarily interpreted. Now
consider the subclass !
1
of algebras fromG
C
dened over [0, 1] in which all constants,
except for a nite number, are interpreted as 1.
Consider any nite set such that -
G
C
. Then also -
G
, where
we understand the new constants just as propositional variables. Thus by the strong
standard completeness of G odel-Dummett logic, there is a [0, 1]
G
-evaluation e such that
e[] 1 and e() < 1. We construct a G
C
-algebra A resulting from [0, 1]
G
by
setting c
A
n
= e(c
n
) for every c
n
occurring in and c
A
n
= 1 otherwise. Notice that
e can be viewed as A-evaluation and, since A !
1
(because contains only
nitely many constants) we obtain, ,[=
1
1
. Thus we have shown that the FS!
1
C
holds for G
C
.
On the other hand, let us by [0, 1]
0
denote the G odel algebra on [0, 1] with all new
constants interpreted as 0. Clearly, any partial subalgebra of [0, 1]
0
containing 0 does
not partially embed into any algebra in !
1
.
Nevertheless, we can give the following characterization for the FSKC that holds
even for logics in innite languages without disjunction.
THEOREM 3.4.11 (Characterization of nite strong completeness). If L is nitary, then
the following are equivalent:
1. L satises the FSKC.
2. Every L-chain belongs to ISP
U
(K).
Proof. 12: if L satises the FSKC then, by Theorem 3.4.3, ALG

(L) = Q(K). It
follows from [26, Lemma 1.5] that every relative nitely subdirectly irreducible member
of Q(K) (i.e. each L-chain) belongs to ISP
U
(K).
21: if every L-chain belongs to ISP
U
(K), since every L-algebra is representable
as subdirect product of L-chains we have that
176 Petr Cintula and Carles Noguera
ALG

(L) P
SD
(ISP
U
(K)) Q(K) ALG

(L).
Therefore by Theorem 3.4.3, L has the FSKC.
We know that the SKC means that L and [=
K
coincide; we can also formulate the
FSKC in a similar manner:
PROPOSITION 3.4.12. Assume that L is nitary. Then L has the FSKC if and only if L
is the nitary companion of [=
K
.
Proof. The direction from right to left is obvious. Assume that L has the FSKC and
take L
t
= T(([=
K
), the nitary companion of [=
K
. Then we have:
L
iff there is
a nite
t
such that
t
[=
K
iff there is a nite
t
such that
t

L
(by the
FSKC) iff
L
(by nitarity of L).
COROLLARY 3.4.13. Assume that L is nitary and [=
K
is nitary too (e.g. P
U
I(K)
I(K)). Then L has the SKC if and only if L has the FSKC.
COROLLARY 3.4.14. Assume that L is nitary and enjoys the FSKC. Then L has the
SP
U
(K)C.
Nowwe showthat the failure of completeness properties is inherited by conservative
expansions.
PROPOSITION3.4.15. Let L
t
be a conservative expansion of L, K
t
a class of L
t
-chains
and K the class of their L-reducts. Then:
If L
t
enjoys the K
t
C, then L enjoys the KC.
If L
t
enjoys the FSK
t
C, then L enjoys the FSKC.
If L
t
enjoys the SK
t
C, then L enjoys the SKC.
Proof. All the implications are proved in a similar way. Let us prove as an example the
rst one. We want to show that L has the KC and we do it contrapositively: assume
,
L
. Since L
t
is a conservative expansion of L, we also have ,
L
and so by the K
t
C
we obtain ,[=
A
for some A
t
K
t
. Thus also ,[=
A
for the reduct A of A
t
. As
A K we obtain ,[=
K
.
An interesting semantics for which we can apply the characterization of strong com-
pleteness is that formed by nite chains:
PROPOSITION 3.4.16. Assume that L is nitary and lattice-disjunctive and let us by T
denote the class of all nite L-chains. Then the following are equivalent:
1. L enjoys the STC.
2. All L-chains are nite.
3. There exists n N such each L-chain has at most n elements.
Chapter II: A General Framework for Mathematical Fuzzy Logic 177
4. There exists n N such that
L
_
i<n
(x
i
x
i+1
).
Proof. 12: From Theorem 3.4.6 we know that every countable L-chain is embeddable
into some member of T, thus there are not innite countable L-chains and so by the
downward L owenheim-Skolem theorem there are no innite chains.
23: If all the algebras in ALG

(L) are nite then there must a bound for their

length, because otherwise by means of an ultraproduct we could build an innite one.
31: Trivial.
34: Take an arbitrary L-chain A, with lter F, and elements a
0
, . . . a
n
A.
Since A has at most n elements it is impossible that a
0
> a
1
> . . . > a
n
, thus there is
some k such that a
k
a
k+1
, i.e. a
k

A
a
k+1
F, and hence it satises the formula.
Since the logic is complete w.r.t. chains, the proof is done.
42: Suppose that there is an L-chain A, with lter F and elements a
0
, . . . a
n
A
such that a
0
> a
1
> . . . > a
n
. Then a
i

A
a
i+1
/ F, for every i < n, and as F is
-prime we know that ,[=
A
_
i<n
(x
i
x
i+1
).
COROLLARY 3.4.17. For a nitary lattice-disjunctive logic L and a natural number
n, the axiomatic extension L
n
obtained by adding the schema
_
i<n
(x
i
x
i+1
), is a
semilinear logic which is strongly complete with respect the L-chains of length less than
or equal to n.
Finally, as other examples of meaningful semantics based on some particular class
of chains, we consider chains dened over intervals of real or rational numbers. Com-
pleteness with respect these kind of semantics has been a traditional item in the agenda
of fuzzy logics (giving rise to some of the so-called standard completeness theorems).
DEFINITION 3.4.18 (Real and rational semantics). The class ! MOD

(L) is de-
ned as: A ! if the domain of A is the closed, open, or semi-open real unit interval
and
A
is the usual order on reals. The class Q MOD

(L) is analogously dened

requiring rational unit intervals as domains.
THEOREM 3.4.19 (Relations between real and rational completeness properties). Let
L be a nitary logic.
1. L has the FSQC iff it has the SQC.
2. If L has the !C, then it has the QC.
3. If L has the FS!C, then it has the SQC.
Proof. The three claims are proved in a similar way by using the downward L owenheim-
Skolem Theorem. Let us show the rst one as an example. If L has the FSQC, then
it also has the SP
U
(Q)C by Corollary 3.4.14. Assume that -
L
. Then there is
A P
U
(Q), with lter F, and an evaluation e such that e[] F and e() / F. It is
clear that Ais a densely ordered chain. Consider the countable set S = e(p) [ p propo-
sitional variable in . By the downward L owenheim-Skolem Theorem (consid-
ering algebras as rst-order structures), there exists a countable elementary substructure
B A, such that S B. Therefore, B is also densely ordered and hence isomorphic
to an element of Q,
16
and so we have a counterexample showing the SQC.
16
We use the well-known fact that any two countable (bounded) dense orders are isomorphic.
178 Petr Cintula and Carles Noguera
Observe that, since we are restricted to logics in countable languages, the comple-
teness properties with respect to Q are, in fact, equivalent to completeness properties
with respect to the whole class of densely ordered linear models. Indeed, when we have
an evaluation over a densely ordered linear model providing a counterexample to some
derivation, we can apply the downward L owenheim-Skolem Theorem to the (countable)
subalgebra generated by the image of all formulae by the evaluation and obtain a ra-
tional countermodel. In particular, the SQC turns out to be the completeness property
characterized in Theorem 3.3.8.
4 First-order predicate semilinear logics
In this section we introduce the basics of the theory of rst-order predicate semi-
linear logics. We will see that for each semilinear logic L one can dene two natural
predicate logics: L
m
, the minimal one (because it is complete w.r.t. all matrices), and
its strengthening L (complete w.r.t. linearly ordered matrices). Interestingly enough,
and unlike what happens in the propositional case, these two logics need not coincide
(see Example 4.1.18).
Our goal is to nd axiomatizations for both logics. We will show that L
m
can be
nicely axiomatized for nearly all weakly implicative logics (in fact we could axiomatize
it for all of them but at price of increased complexity). However, to axiomatize L we
need to restrict to logics with a reasonable disjunction connective. Thus we make the
following convention:
CONVENTION 4.0.1. In order to simplify the formulation of upcoming denitions and
results, let us assume from now on that each logic L has the following properties:
L is a weakly implicative semilinear logic with principal implication .
L has a p-protodisjunction satisfying (MP

) (and thus all the properties from

Theorem 3.2.13, like the PCP and the PEP).
The language of L contains a truth constant 1 satisfying the consecutions
1 .
Typical (though not exhaustive) examples of logics satisfying the above conditions are
substructural semilinear logics which either have both 1 and in their language or are
Rasiowa-implicative.
Notice that for any L satisfying this convention and any L-matrix A = A, F it
is the case that 1
A
= min

A
F. It is important to add that all the results about the
minimal predicate logic L
m
proved in this section would also hold for logics satisfying
only the third condition of the above convention; however to minimize the complexity
of this section, we prefer to keep all the assumptions from the beginning.
In the rst subsection we deal with basic syntactic and semantic notions. Notice that
our restriction to the class of logics above (in particular the third condition) is used for
the rst time in Example 4.1.15 to demonstrate the soundness of generalization rule. The
other two assumptions are used in Example 4.1.17 to show the soundness of the -form
Chapter II: A General Framework for Mathematical Fuzzy Logic 179
of this rule in the semantics based on chains. The second subsection presents axiomati-
zations of both predicate logics and proves their soundness. The third subsection shows
alternative simpler axiomatizations in predicate substructural logics and other syntacti-
cal properties of these logics like the local deduction theorem. The fourth subsection
contains the proof of completeness of both kinds of predicate logics with respect to the
presented axiomatizations by means of a generalization of the Henkin-style proof used
for classical and some non-classical rst-order logics. Finally, the fth subsection stud-
ies a notion of Skolemization for our rst-order logics and considers the semantics of
witnessed models.
4.1 Basic syntactic and semantic notions
The following denitions are absolutely standard; we present them for the readers
convenience. Let us x a propositional language / and a logic L.
DEFINITION 4.1.1 (Predicate language). A predicate language T is a triple P, F, ar,
where P is a non-empty set of predicate symbols, F is a set of function symbols, and ar
is a function assigning to each predicate and function symbol a natural number called
the arity of the symbol. The functions f for which ar(f) = 0 are called object constants.
The predicates P for which ar(P) = 0 are called truth constants.
Let us further x a predicate language T = P, F, ar and a denumerable set V
whose elements are called object variables.
DEFINITION 4.1.2 (Term). The set of T-terms is the minimum set X such that:
V X, and
if t
1
, . . . , t
n
X and f is an n-ary function symbol, then f(t
1
, . . . , t
n
) X.
DEFINITION 4.1.3 (Formulae). An atomic T-formula in any expression P(t
1
, . . . , t
n
)
where P is an n-ary predicate symbol and t
1
, . . . , t
n
are T-terms. Atomic T-formulae
and nullary logical connectives of / are called atomic /, T-formulae. The set of
/, T-formulae is the minimum set X such that:
X contains the atomic /, T-formulae,
X is closed under logical connectives of /, and
if X and x is an object variable, then (x), (x) X.
CONVENTION 4.1.4. We speak about T-formulae if the propositional language is
clear from the context and we speak about terms and formulae if both the proposi-
tional and the predicate languages are clear from the context. The same convention
will be used for any other notion dened in this section parameterized by propositional
or predicate languages.
Given a set of formulae , we denote by T

the predicate language containing

exactly the predicate and function symbols that occur in formulae of .
180 Petr Cintula and Carles Noguera
DEFINITION 4.1.5 (Bound and free variables, closed terms and formulae). An occur-
rence of a variable x in a formula is bound if it is in the scope of some quantier
over x; otherwise it is called a free occurrence.
A variable is free in a formula if it has a free occurrence in . A term is closed if
it contains no variables. A formula is closed if it has no free variables; closed formulae
are also known as sentences.
CONVENTION 4.1.6. Instead of
1
, . . . ,
n
(where
i
s are terms or formulae and n is
arbitrary or xed by the context) we shall sometimes write just

.
Unless stated otherwise, by the notation (z) we signify that all free variables of
are among those in the vector of pairwise different object variables z.
If (x
1
, . . . , x
n
, z ) is a formula and we replace all free occurrences of x
i
s in by
terms t
i
, we denote the resulting formula in the context simply by (t
1
, . . . , t
n
, z ).
DEFINITION 4.1.7 (Substitutability). A term t is substitutable for the object variable x
in a formula (x, z ) iff no occurrence of any variable occurring in t is bound in (t, z )
unless it was already bound in (x, z ).
DEFINITION 4.1.8 (Theory). A theory T is a pair T, , where T is a predicate
language and is a set of T-formulae. A theory is called closed if all its formulae
are sentences.
For convenience we sometimes identify the theory T and its set of formulae and
say that T is a T-theory to indicate that its language is T (see e.g. Denition 4.1.12).
DEFINITION4.1.9 (Structure). A T-structure Sis a pair A, S where AMOD

(L)
and S has the form S, P
S

PP
, f
S

fF
, where S is a non-empty domain; P
S
is an
n-ary fuzzy relation, i.e. a function S
n
A, for each n-ary predicate symbol P P
with n 1 and an element of A if P is a truth constant; f
S
is a function S
n
S for
each n-ary function symbol f F with n 1 and an element of S if f is an object
constant.
Sometimes, S is called an A-structure for T and we write P
S
instead of P
S
.
DEFINITION 4.1.10 (Evaluation). Let S be a T-structure. An S-evaluation of the
object variables is a mapping v which assigns to each variable an element from S.
Let v be an S-evaluation, x a variable, and a S. Then v[xa] is an S-evaluation
such that v[xa](x) = a and v[xa](y) = v(y) for each object variable y ,= x.
DEFINITION 4.1.11 (Truth denition). Let S = A, S be a T-structure and v an
S-evaluation. We dene values of the terms and truth values of the formulae in Sfor an
evaluation v as:
|x|
S
v
= v(x),
|f(t
1
, . . . , t
n
)|
S
v
= f
S
(|t
1
|
S
v
, . . . , |t
n
|
S
v
), for f F
|P(t
1
, . . . , t
n
)|
S
v
= P
S
(|t
1
|
S
v
, . . . , |t
n
|
S
v
), for P P
|c(
1
, . . . ,
n
)|
S
v
= c
A
(|
1
|
S
v
, . . . , |
n
|
S
v
), for c /
|(x)|
S
v
= inf

A||
S
v[xa]
[ a S,
|(x)|
S
v
= sup

A||
S
v[xa]
[ a S.
Chapter II: A General Framework for Mathematical Fuzzy Logic 181
If the inmum or supremum does not exist, we take its value as undened. We say that
S is safe iff ||
S
v
is dened for each T-formula and each S-evaluation v.
We set the following useful denotations for a structure S = A, F, S. We write
|(a
1
, . . . , a
n
)|
S
instead of |(x
1
, . . . , x
n
)|
S
v
for v(x
i
) = a
i
.
S [= [v] if ||
S
v
F.
S [= if S [= [v] for each S-evaluation v.
S [= if S [= for each .
We sometimes write ||
A
S,v
instead of ||
S
v
to keep the traditional notation from the
literature.
DEFINITION 4.1.12 (Model). Let T be a T-theory and K MOD

(L). A T-struc-
ture M= A, M is called a K-model of T if it is safe, A K, and M[= T.
We use just the term A-model instead of A-model and we also use this term
for its safe A-structure. When the logic L is known from the context, we just use the
terms model and -model instead of MOD

(L)-model and MOD

(L)-model.
Notice that as each theory comes with its xed predicate language we do not need to
specify the language of Mwhen we say that it is a model of the theory T. By a slight
abuse of language we will use the term model instead of safe T-structure, when the
language is clear from the context.
DEFINITION 4.1.13 (Consequence relation). Let K MOD

(L). A T-formula
is a semantical (sentential) consequence of a T-theory T w.r.t. the class K, in symbols
T [=
K
, if for each K-model Mof T we have M[= .
Note that both in the denition of model and semantical consequence, the language
of the theory T plays a minor r ole; basically they could be formulated just for sets of
formulae. Indeed we can prove the following:
PROPOSITION 4.1.14. Let K MOD

(L) and a set of T-formulae. Then

the following are equivalent:
1. T
t
, [=
K
for all T
t
T,
2. T
t
, [=
K
for some T
t
T,
3. T, [=
K
.
Proof. If we show that for each T
t
T we have: T, [=
K
iff T
t
, [=
K
the
proof is done. One direction is simple as any T-reduct of a K-model of T
t
, is clearly
a K-model of T, .
To prove the second one we need to showthat for any safe T-structure M= A, M
we can construct a safe T
t
-structure M
t
= A, M
t
such that Mis its T-reduct. The
idea is clear: take M, dene the interpretations of additional function and predicate
182 Petr Cintula and Carles Noguera
symbols in an arbitrary way and check that the resulting structure is still safe. Let s
be an arbitrary element from M, an arbitrary T-sentence and a = ||
M
. For any
functional symbol f from T
t
T we dene f
M
(x) = s and for any predicate symbol
P from T
t
T we set P
M
(x) = a. Clearly Mis the T-reduct of M
t
and if we show
that M
t
is safe the proof is done.
Assume (without loss of generality) that there is a T
t
-formula (x, y) and a se-
quence of elements r from M such that inf|(r, r)|
M

[ r M does not exist. Let z

be a variable different fromx and y. Then for any T
t
-termt(x, y) we construct a T-term

t(x, y, z) by replacing each subterm of t(x, y) given by some f T

t
T by the variable
z. Clearly for each u, u from M we have: |t(u, u)|
M

=
_
_
t(u, u, s)
_
_
M
. Let further
(x, y, z) be the T-formula resulting from (x, y) by replacing:
each atomic subformula given by a predicate symbol P T
t
T by the sen-
tence ,
each term t by the term

t.
Notice that |(r, r)|
M

= | (r, r, s)|
M
for each r, r S. Therefore, the inmum
of | (r, r, s)|
M
[ r M does not exist, i.e. Mis not safea contradiction.
Now we give a series of examples. The rst two demonstrate that we need to have
unit in the language for the validity of the well-known generalization rule; the other two
show that in rst-order semilinear logics, unlike in the propositional case, the conse-
quence relations [=
MOD

(L)
and [=
MOD

(L)
need not coincide.
EXAMPLE 4.1.15. We show that for any L: [=
MOD

(L)
(x). Consider a model
M = A, F, M of and an M-evaluation e. We know that ||
M
e[xa]
F for
each a M. Since L satises Convention 4.0.1, we know from Proposition 2.5.10 that
inf
A
||
M
e[xa]
[ a M inf
A
F = 1 F (the rst inmum exists because Mis
safe).
EXAMPLE 4.1.16. Consider the logic L given by the three-valued reduced matrix with
domain a, b, , lter a, b and dened as:
a b
a a a
a a
b b
Clearly, L is a weakly implicative logic and it is not difcult to build a model showing
that ,[=
MOD

(L)
(x).
EXAMPLE 4.1.17. We show that for any lattice-disjunctive logic L: [=
MOD

(L)
((x)) whenever x is not free in . Consider an -model M of and an
M-evaluation e. If M [= [e] we are done. Assume that M ,[= [e], then also M ,[=
[e[xa]] for each a M (because x is not free in !). Since the lter in the matrix
is -prime, we know that M [= [e[xa]]; the rest of the proof is the same as in
Example 4.1.15.
Chapter II: A General Framework for Mathematical Fuzzy Logic 183
EXAMPLE 4.1.18. [=
MOD

(G)
is different from [=
MOD

(G)
for the G odel-Dummett
logic G and so they also differ for any logic weaker than G. Indeed, from the previous
example we know [=
MOD

(G)
((x)) and we show that ,[=
MOD

(G)
((x)). Consider = P(x) for a unary predicate P and = c a nullary predicate.
Take the G-algebra Awhose domain is 0, 1/n [ n N; the lattice operations are
given by the lattice order: elements different from are ordered as usual, 0 1,
is incomparable with all the other elements; residual conjunction &
A
=
A
; and
implication
A
is its residuum (c
A
1 = 0
A
c = c
A
c = 1, c
A
0 =
c
A
1/n = 1/2, and 1
A
c = 1/n
A
c = c, and dened as usual for the
remaining values). It is easy to check that the equation (x y) (y x) 1 is
valid in A, so we have a non-linearly ordered G-algebra. Now take N as the domain of a
rst-order structure S and interpret c
S
= and P
S
(n) = 1/n, and we have the desired
counterexample.
At the end of this subsection we introduce two special kinds of models, exhaustive
and fully-named ones.
DEFINITION 4.1.19 (Exhaustive model). Let M = A, M be a model for T. We
dene the set
A
exhM
= ||
M
v
[ a T-formula and v an M-evaluation
and say that Mis exhaustive if A = A
exhM
.
Intuitively, Mis exhaustive if Aonly contains the necessary values to interpret rst-
order formulae of the language. The following straightforward proposition shows that
for any model we can always restrict to its exhaustive submodel.
PROPOSITION 4.1.20. Let M= A, F, M be a model. Then:
There is a subalgebra A
exhM
of Awith domain A
exhM
.
There is a submatrix A
exhM
= A
exhM
, F A
exhM
of A, F.
If we denote by M
exh
the model A
exhM
, M
t
, where P
M
= P
M
and F
M
=
F
M
for each predicate symbol P and functional symbol F, then M
exh
is exhaus-
tive and for each formula and each M-evaluation v holds:
||
M
v
= ||
M
exh
v
.
DEFINITION 4.1.21 (Fully named model). Let Mbe a model. We say that Mis fully
named if for each m M there is a closed term t, such that t
M
= m.
4.2 Axiomatic systems and soundness
As we have seen above, the two natural semantical consequence relations we have
introduced for rst-order semilinear logics may be different in general. The goal of this
subsection is to propose axiomatizations for both of them and show their basic properties
including their soundness, their completeness will be proved later, in Subsection 4.4.
184 Petr Cintula and Carles Noguera
DEFINITION 4.2.1 (Predicate logics L
m
and L). Let L be a logic in /. The minimal
predicate logic over L (in a predicate language T), denoted as L
m
, is the logic dened
by the following axiomatic system:
(P) the rules resulting from the consecutions of L by substituting
propositional variables by /, T-formulae,
(1)
L
m (x)(x, z) (t, z), where t is substitutable for x in ,
(1)
L
m (t, z) (x)(x, z), where t is substitutable for x in ,
(2)
L
m (x), where x is not free in ,
(2)
L
m (x) , where x is not free in .
Further, we dene the predicate logic over L (in a predicate language T), denoted
as L, as the extension of L
m
by the following two rules:
(2)

( )
L
( (x)), where x is neither free in nor in ,
(2)

( )
L
((x) ), where x is neither free in nor in .
CONVENTION 4.2.2. Many results and denitions in this section are valid for both
logics L
m
and L. To simplify matters, when a denition or a theorem does not specif-
ically mention a particular predicate logic we mean that it holds for both of them.
Notice that we have omitted the propositional language / in the symbol for the
predicate logics over L for it is always that of L. Omitting the symbol for the predicate
language could be more confusing. In order to avoid possible problems, we rst dene
the notion of proof from a T-theory T in the (minimal) predicate logic over L in a
predicate language T in the same way we did it in the propositional case, denoting it
by means of . We can obtain the analog of Proposition 4.1.14 either as a consequence
of the completeness theorem or by a direct syntactical proof (as in classical logic).
17
Finally, observe that there is no need to mention the used p-disjunction in the symbol for
L, because we know that all p-disjunctions are mutually derivable.
PROPOSITION 4.2.3. Let and a set of T-formulae. Then the following are
equivalent:
1. T
t
, for all T
t
T,
2. T
t
, for some T
t
T,
3. T, .
PROPOSITION 4.2.4. Let /o be a presentation of L, then the group of rules (P) in
the axiomatization of L
m
can be equivalently replaced by
17
A hint of the proof in the nitary case: consider a proof of from1

, ), i.e. a sequence of 1

-formulae.
We can transform any element of this proof by the following process: replace any term f(s), where f , 1,
by an unused variable, and replace any atomic subformula Q(s), where Q is an n-ary symbol not in 1, by
an arbitrary 1-formula (x) with n free variables. It can be seen that the resulting sequence of formulae is a
proof of from 1, ).
Chapter II: A General Framework for Mathematical Fuzzy Logic 185
(P
,S
) the rules resulting from the rules of /o by the substitution of the
propositional variables by /, T-formulae
Furthermore, axioms (2) and (2) are redundant in the axiomatization of L.
Proof. The proof of the rst claim is straightforward. We prove the second one: using
(PD) we know that ( )( (x)), thus from (2)

we obtain
( (x))( (x)). The idempotency of completes the proof of
(2). The proof of (2) is analogous.
Later we will see simpler axiomatizations of L
m
and L for particular choices of
a logic L.
PROPOSITION 4.2.5. The following consecutions are provable:
(0) (x),
(T1) (x) (x),
(T2) (x) (x),
(T3) (x) if x is not free in ,
(T4) (x) if x is not free in ,
(T5) (x)(x, z) (x
t
)(x
t
, z) if x
t
does not occur in (x, z),
(T6) (x)(x, z) (x
t
)(x
t
, z) if x
t
does not occur in (x, z),
(T7) (x)(y) (y)(x),
(T8) (x)(y) (y)(x).
Proof. The proof of generalization (0) is a simple corollary of rule (2) used for
= 1. We show the proof of odd claims (for ); the proofs for are analogous.
(T1) Using (1) and (T) we obtain (x) . Rule (2) completes the
proof.
(T3) One implication is axiom (1). To prove the second one starts from (R) in the
form and the rule (2) completes the proof.
(T5) Clearly (x)(x, z) (x
t
, z) by (1) (x
t
is clearly substitutable for x in
). Rule (2) completes the proof of one implication (x
t
is clearly not free in
(x)(x, z)). The proof of the second implication is symmetric.
(T7) From (1) and (T1) we obtain (x)(y)(x, y, z) (x)(x, y, z). Rule (2)
completes the proof of one implication. The proof of the second one is symmetric.
Observe that the condition x
t
does not occur in (x, z) is unnecessarily strong
and could be replaced by x
t
is both substitutable for x and not free in (x, z) and x is
both substitutable for x
t
and not free in (x
t
, z).
From the group of the rules (P) and rules (T1), (T2) we obtain (by induction on
the complexity of the formula ):
THEOREM 4.2.6 (Congruence Property). Let , , be sentences. Then:
186 Petr Cintula and Carles Noguera


,
Further assume that is a formula and is obtained from by replacing some
occurrences by . Then
.
The next straightforward proposition shows that we can restrict our attention to
sentences, as usual in rst-order logics.
PROPOSITION4.2.7. Let be arbitrary formulae. We denote by the universal
closure of (i.e. if x
1
, . . . , x
n
are the free variables in , then = (x
1
) . . . (x
n
)),
and by the set of universal closures of all formulae in . Then: iff .
The following theorem shows that free variables behave as constants naming arbi-
trary elements.
THEOREM 4.2.8 (Constants Theorem). Let (x, z) set of formulae and c a
constant not occurring in (x, z). Then (c, z) iff (x, z).
Proof. The right-to-left direction follows easily from (0) and (1). Assume that
(c, z). Let the sequence
1
, . . . ,
n
be a proof of (c, z) from (assuming for
simplicity that the logic is nitary; for the general case the proof is analogous). Let
y be a variable different from x and not occurring in the formulae
1
, . . . ,
n
. We
denote by S
y
c
(
i
) the substitution in
i
of each occurrence of c by y. We will show that
S
y
c
(
0
), . . . , S
y
c
(
n
) is a proof of S
y
c
(
n
) = (y, z) from
0
=
1
, . . . ,
n
.
This will end the proof because then
0
(y)(y, z) (by (0)), hence
0
(x, z)
(by (1)), so nally (x, z).
If
i
, then S
y
c
(
i
) =
i

0
, because y does not occur in
0
. If
i
results
from a rule in (P), then the same holds for S
y
c
(
i
) because the substitution preserves the
propositional structure of formulae. Assume that
i
= (v)(v, z) (t, z), where t
is substitutable for v in . Then S
y
c
(
i
) = (v)S
y
c
((v, z)) S
y
c
((t, z)), where t is
substitutable for v in S
y
c
(), is still an instance of (1). The case of (1) is analogous.
Assume that
i
results from an application of the rule (2), i.e.
j
= , for some
j < i, and
i
= (v), where v is not free in . Then S
y
c
(
j
) = S
y
c
() S
y
c
()
and S
y
c
(
i
) = S
y
c
() (v)S
y
c
(), where v is not free in S
y
c
(), so the formula still
results from an application of (2). The remaining rules are analogously checked.
The next lemma and its two corollaries show that the p-disjunction retains some
good properties in the rst-order logic L, such as closure under -forms and the PCP.
LEMMA 4.2.9. For each set of formulae , such that
L
we have
L
for each sentence .
Chapter II: A General Framework for Mathematical Fuzzy Logic 187
Proof. We show
L
for each appearing in the proof of from . If
or it is an axiom, the proof is trivial. Now assume that
t

L
is the inference rule
we use to obtain . From the induction assumption we have
L

t
. Since

t

L
(for propositional rules due the PCP of L, for rst-order rules due to
our denition of the axiomatic system for L), the proof of this claim is done.
COROLLARY 4.2.10. The following consecution is provable in L:
(0)

L
(x), where x is not free in .
Proof. Let (x, y), (y) be formulae, x a variable not among y, and c constants not
occurring in those formulae. By the previous lemma we obtain: (x, c)(c)
L
(x)((x, c))(c). Since (x, y)(y)
L
(x, c)(c) (using (0) and (1)),
we obtain (x, y)(y)
L
(x)((x, c))(c). The Constants Theorem completes
the proof.
COROLLARY 4.2.11. L enjoys the PCP and the SLP, i.e. for each T-theory T and
T-sentences , , and holds:
T,
L
T,
L

T,
L

T,
L
T,
L

T
L

Proof. Now assume that
1

L
and
2

L
. Using the Lemma 4.2.9 we obtain
that
2

L
and
1

L
for each
2
and so
1

2

L

2
.
Using (I

) and (C

) we obtain
1

2

L
, and so we have the PCP.
To prove the SLP we start from T,
L
and T,
L
and by the
PCP we obtain T, ( )( )
L
. Knowing that L satises the (P

) we
obtain T
L
.
We leave the proof of the soundness of both our logics with respect to their intended
semantics as an exercise for the reader. Recall that in Example 4.1.18 we have seen that

L
[=
MOD

(L)
need not be true in general.
THEOREM 4.2.12 (Soundness of rst-order logics). For any logic L the following
holds:

L
m [=
MOD

(L)

L
[=
MOD

(L)
4.3 Predicate substructural logics
In this subsection we focus on the predicate logics over substructural logics. We will
see that the axiomatic systems corresponding to their predicate logics can be presented
in a simpler way. Recall that, according to Convention 2.5.9, a propositional weakly
implicative logic L in a language / is a substructural logic if L is an expansion of the
/ /
SL
-fragment of SL. Thus in particular all /-consecutions provable in SL are
provable in L too. Unfortunately the situation is more complicated in the predicate
case. For instance, if L is the -fragment of SL, we do not know whether L
m
is the
-fragment of SL
m
. Therefore e.g. from the fact that the upcoming formula (2
t
) is
188 Petr Cintula and Carles Noguera
a theorem of SL
m
we cannot infer that it is a theorem of L
m
even though its only
propositional connective is .
Therefore we are not going to prove our claims just for SL and assume that they
will transfer to the proper fragments, but we formulate the forthcoming theorems for
the smallest fragments in which we can express their proofs.
18
For simplicity, we will
however tacitly assume that whenever we formulate some claim in relation with some
logic, this logics has at least the necessary connectives to express the claim. We start
with the rst-order version of the Duality Theorem 2.5.8 (the denition of mirror image
for predicate formulae is the natural extension of Denition 2.5.2). Its proof is straight-
forward: all new predicate axioms and rules have a principal implication which can
be easily replaced by using the rule (symm).
THEOREM 4.3.1 (Duality Theorem). Let , / /
SL
and L the /-fragment
of SL
X
for X a, e, c, i, o. For any T-theory T and any T-formula :
T iff T
t

t
.
PROPOSITION 4.3.2. Let L be a substructural logic. Let , , be formulae and x a
variable not free in . The following hold:
(T9)
L
m ( (x)) (x)( ),
(T10)
L
m ((x) ) (x)( ),
(T11)
L
m (x)( ) ( (x)),
(T12)
L
m (x)( ) ((x) ),
(T13)
L
m (x) (x) (x)( ),
(T14)
L
m (x)( ) (x) (x),
(T15)
L
m (x) (x)( ),
(T16)
L
m (x)( ) (x) .
If & is in the language, then we also have:
(2
t
)
L
m (x)( ) ( (x)),
(T1
t
)
L
m (x)( ) ((x) (x)).
If is in the language, then we also have:
(2
t
)
L
m (x)( ) ((x) ),
(T2
t
)
L
m (x)( ) ((x) (x)),
(T17)
L
m (x)( & ) (x) & .
If is in the language, then we also have:
(3)
L
(x)( ) (x) ,
(3)
L
(x) (x)( ).
18
Of course, the resulting language restriction can be seen only as an upper bound, because one might
always expect to nd a proof in smaller language, ideally using only the language of the claim itself.
Chapter II: A General Framework for Mathematical Fuzzy Logic 189
Proof. The proofs of the rst four statements are simple: use (1) or (1), then prexing
or sufxing, and then (2) or (2). The proofs of the left-to-right directions in the
second four statements are also simple: use (1) or (1), monotonicity of or , and
then (2) or (2). Let us show the proof of right-to-left direction of (T13) ((T14) is
fully analogous): from we get (x)( ) using (1) and so by (2)
also (x)( ) (x), analogously for and then rule 3 completes the proof.
(2
t
) from (1) we get (x)( ) ( ) and so & (x)( ) .
Using the rule (2) and residuation again completes the proof.
(T1
t
) from (1), sufxing, and prexing we get (x)( ) ((x) ), (2)
and (2
t
) complete the proof.
(2
t
) from (1) we get (x)( ) ( ) and so ((x)( ) ).
Using the rule (2) and (E)

again completes the proof.

(T2
t
) the proof is analogous to the proof of (T1
t
).
(T17) from(1) we get & (x)&. (2) completes the proof of one direction.
The reverse one: we start with ( &) then apply (2), and then (T2
t
)
completes the proof.
(3) from (x)( ) we get ((x)( ) ) ((x)( ) )
(by (lin

)). Using (0)

we obtain (x)((x)() )((x)() ),

and so by (2
t
) we obtain ((x)( ) (x)) ((x)( ) ). The
rest is simple.
(3) From (x)( ) we get [ (x)( )] [ (x)( )]
(by (lin

)). Using (0)

we obtain (x)[ (x)()] [ (x)()],

and so by (2
t
) we obtain [(x) (x)()][ (x)()]. Therefore:
(x) (x)( ).
Notice that, as corollaries of (T13) and (T14), we easily obtain the provability of
(x) (x)( ) and (x)( ) (x) , for x not free in .
REMARK 4.3.3. Notice that the quantication theory in rst-order substructural logics
is almost classical. In fact, it is much closer to the intuitionistic one as the only two
unprovable quantier shifts for implication are those which are also unprovable in intu-
itionistic logic. More formally, let L be any logic that can be expanded to intuitionistic
logic; then due to the soundness we know that:
(T11
r
) -
L
m ( (x)) (x)( ),
(T12
r
) -
L
m ((x) ) (x)( ).
Next we show how to simplify the axiomatic systems of our predicate logics. The
previous proposition showed us a wide class of logics satisfying the precondition of the
following theorem.
190 Petr Cintula and Carles Noguera
THEOREM 4.3.4 (Simpler axiomatization of L
m
). Let us assume that the logic L
m
proves (2
t
) and (2
t
). Then L
m
can be alternatively axiomatized by (P) and the
following:
(1)
L
m (x)(x, z) (t, z), where t is substitutable for x in ,
(1)
L
m (t, z) (x)(x, z), where t is substitutable for x in ,
(2
t
)
L
m (x)( ) ( (x)), where x is not free in ,
(2
t
)
L
m (x)( ) ((x) ), where x is not free in ,
(0)
L
m (x).
Furthermore, an alternative axiomatization of L can be obtained extending this
system with:
(0)

L
((x)), where x is not free in .
Proof. The rst part of the claim is trivial. One direction of the second claim follows
from Corollary 4.2.10. To prove the second direction observe that the logic just dened
satises an analog of Lemma 4.2.9 and Constants Theorem, and thus we can prove
(2)

and (2)

in the same way we have proved (0)

in Corollary 4.2.10.
As a consequence we obtain the next theorem, for which we need a rather rich
language. We formulate it for expansions of SL but, in fact, the presence of , &, ,
and 1 would sufce. Recall that we are restricted to semilinear logics in this whole
section.
THEOREM 4.3.5 (Axiomatization of rst-order substructural logics). Let L be an ex-
pansion of SL. Then L can be alternatively axiomatized by (P) and the following:
(1)
L
(x)(x, z) (t, z), where t is substitutable for x in ,
(1)
L
(t, z) (x)(x, z), where t is substitutable for x in ,
(2
t
)
L
(x)( ) ( (x)), where x is not free in ,
(2
t
)
L
(x)( ) ((x) ), where x is not free in ,
(3)
L
(x)( ) (x) , where x is not free in ,
(0)
L
(x).
Of course in both previous theorems in the propositional part of the axiomatic sys-
tem of L
m
and L we can replace (P) by (P
,S
) where /o is an arbitrary axiomatic
system for L (see Proposition 4.2.4). Let us show that that in the case of ukasiewicz
logic both predicate logics coincide:
COROLLARY 4.3.6. =
m
.
Proof. It is enough to show that
m
proves (3). From ( ) (( ) )
and (T1
t
) we obtain (x)( ) (x)(( ) ). Now, again by (T1
t
),
we have (x)(( ) ) ((x)( ) (x)). By (T9) and sufxing,
((x)( ) (x)) (( (x)) (x)), and so nally we obtain
(( (x)) (x)) (x) . Transitivity ends the proof.
Chapter II: A General Framework for Mathematical Fuzzy Logic 191
The next proposition, which can be seen as quantier shift of over the dened
unary connectives
n
, is crucial in the proof of Corollary 4.3.10 which will play an im-
portant r ole in Subsection 4.5.
PROPOSITION 4.3.7. Let L be an associative substructural logic with in its lan-
guage. Then:
(T18)
L
m (x)(
n
) ((x))
n
.
Proof. The left-to-right direction is simple: using (1) n-times and monotonicity of
& we obtain
n
((x))
n
; (2) completes the proof.
19
We prove the converse
direction for n = 2; the proof for n > 2 is analogous. First observe the provability of
the propositional formula
&
2

2
(from we obtain & & and so &
2

2
; we obtain the
same from and hence the SLP completes the proof). Assume that x is free in
(otherwise the proof is trivial) and no other variables are free in (this assumption
only simplies the notation, the proof for a formula with more free variables would
essentially be the same). Choose a variable y which does not occur in , then clearly
(x) & (y)
2
(x)
2
(y) and so by (1), the properties of and (T6) (in the
form: (x)
2
(y)
2
(y)) we get (x) & (y) (x)
2
. Thus by (2) we obtain
(y)((x) & (y)) (x)
2
and so by (T17) and (T6) we have (x) & (x)
(x)
2
. We just repeat the last three steps to complete the proof.
Our next aim is to prove a form of local deduction theorem for predicate substruc-
tural logics. Let T be a predicate language and DT be a set of -formulae (i.e. propo-
sitional formulae in language / built from the normal set of variables enhanced with a
new distinguished variable ; see the beginning of Subsection 2.6). By DT
1
we denote
the set of formulae resulting from any -formula from DT by replacing all its proposi-
tional variables other than by arbitrary T-sentences. Note that elements of DT
1
are
not T-formulae, but if we substitute all occurrences of by a T-sentence we get another
T-sentence.
THEOREM 4.3.8 (Local Deduction Theorem for L
m
). Let L be a substructural logic
with &, and 1 in its language. Let T be a predicate language. Assume that L is
almost (MP)-based with a set of basic deductive terms bDT. Then for each T-theory
T, T-formula and T-sentence , we have:
T,
L
m iff T
L
m () for some (bDT

)
1
.
Proof. The observation that
L
m () for each (bDT

)
1
completes the
proof of the right-to-left direction. To prove the converse one we rst observe (using
Theorem 4.3.4, Proposition 4.3.2 and the comments after Theorem 4.3.5) that L
m
can
be axiomatized using modus ponens, rules of the form
L
m () for (bDT

)
1
,
and
L
m (x). The proof runs along the lines of the proof of Theorem 2.6.3. The
19
Notice that the proof of the left-to-right direction does not use the associativity assumption.
192 Petr Cintula and Carles Noguera
induction base and the induction steps for all the rules except
L
m (x) are done in
the same way as in the propositional case. Let us deal with the remaining one, i.e. assume
that = (x). From the induction assumption there has to be

(bDT

)
1
such
that T
L
m

() . Using (2) we obtain T

L
m

() (x) and so setting

completes the proof.

Using Theorem 4.3.5 we can easily prove the analog of the theorem above for L.
THEOREM 4.3.9 (Local Deduction Theorem for L). Let L be a substructural logic
expanding SL. Let T be a predicate language. Assume that L is almost (MP)-based
with the set of basic deductive terms bDT. Then for each T-theory T, T-formula and
T-sentence , we have:
T,
L
iff T
L
() for some (bDT

)
1
.
As a corollary we obtain the following claim for L which will be useful in Subsec-
tion 4.5.
COROLLARY 4.3.10. Let L be an axiomatic expansion of FL

e
. Then for each predicate
language T, each T-theory T, each T-formula (x) and any constant c , T holds that
T (c) is a conservative expansion (in the logic L) of T (x)(x).
Proof. Assume that T (c)
L
. Then, by the local deduction theorem, there
is n such that T
L
((c) 1)
n
. Thus by the Constants Theorem and (2)
we obtain T
L
(x)((x) 1)
n
. Using (T18) and (3) we obtain T
L
((x)((x) 1))
n
and T
L
((x)(x) 1)
n
. Local deduction theorem
completes the proof.
4.4 Completeness theorem
In this subsection we will show that the axiomatic systems L
m
and L are re-
spectively presentations of the semantically dened rst-order logics [=
MOD

(L)
and
[=
MOD

(L)
, i.e. we will prove two completeness theorems by showing that the reverse
inclusions in Theorem 4.2.12 hold as well. To this end we need the notions of linear and
-Henkin theory. Again we proceed for both logics at once.
DEFINITION 4.4.1 (Linear and -Henkin theories). Let T be a predicate language. A
T-theory T is
linear if for each pair of T-sentences and we have T or T .
-Henkin if for each T-formula such that T - (x)(x) there is an object
constant c in T such that T - (c).
Note that the quantier (x) could be omitted from the denition. Next we intro-
duce the notions of Lindenbaum-Tarski algebra and canonical model of a theory T.
DEFINITION 4.4.2 (Lindenbaum-Tarski algebra). Let be a T-sentence and T a
T-theory. We dene
[]
T
= [ a T-sentence and T .
Chapter II: A General Framework for Mathematical Fuzzy Logic 193
The Lindenbaum-Tarski matrix of T, denoted by LindT
T
, has the domain L
T
=
[]
T
[ a T-sentence, operations:
c
LindT
T
([
1
]
T
, . . . , [
n
]
T
) = [c(
1
, . . . ,
n
)]
T
(for each n-ary connective c of L and each T-sentences
1
, . . . ,
n
), and the lter
[T] = []
T
[ a T-sentence and T .
The denition is sound due to the congruence property of proved in Theo-
rem 4.2.6. The Linbenbaum-Tarski matrix of a theory T will allow us to dene a model
of T, the so-called canonical model, where the formulae not provable from T are not
valid. The following proposition shows that the matrix belongs to the corresponding
classes with respect to which we want to prove completeness. Next proposition (and to
a large extent its proof) is analogous to Lemma 2.2.9 for propositional logics.
PROPOSITION 4.4.3. Let T a T-theory. Then:
1. []
T

LindT
T
[]
T
iff T .
2. LindT
T
MOD

(L).
3. LindT
T
MOD

(L) iff T is linear.

Proof. We show these three claims for the minimal logic; the proofs for L are com-
pletely analogous. Part 1 is proved by the following chain of simple equivalencies:
[]
T

LindT
T
[]
T
iff []
T

LindT
T
[]
T
[T] iff [ ]
T
[T] iff T
L
m .
If we show that LindT
T
MOD(L), the proofs of parts 2 and 3 trivially follow.
Assume that
L
and let us x a LindT
T
-evaluation e such that e[] [T] and
we need to show that e() [T]. Let us dene a mapping from propositional for-
mulae to /, T-sentences by induction over the complexity of formulae: (v) e(v)
(arbitrarily for each propositional variable v) and (c(
1
, . . . ,
n
)) = c(
1
, . . .
n
)
for each n-ary connective c and propositional formulae
1
, . . . ,
n
. Further we show
that for each propositional formula we obtain []
T
= e() by induction: for vari-
ables it is clear, now assume that = c(
1
, . . . ,
n
), we obtain [c(
1
, . . . ,
n
)]
T
=
[c(
1
, . . .
n
)]
T
= c
LindT
T
([
1
]
T
, . . . , [
n
]
T
) = c
LindT
T
(e(
1
), . . . , e(
n
)) =
e(c(
1
, . . . ,
n
)).
Since e[] [T], we have T []. From
L
we obtain []
L
m
(due to the group (P) of rules in the axiomatization of L
m
). Taken together, we have
T
L
m and so e() = [()]
T
[T].
LEMMA 4.4.4. Let T be a -Henkin T-theory. Then for any T-formula with only
one free variable x holds:
[(x)]
T
= inf

LindT
T
[(c)]
T
[ c C,
[(x)]
T
= sup

LindT
T
[(c)]
T
[ c C,
where C is the set of all closed T-terms.
194 Petr Cintula and Carles Noguera
Proof. We prove only the rst claim for the proof of the second one is completely anal-
ogous. It is simple to see that [(x)]
T
is a lower bound: from axiom (1) and part 1 of
the previous proposition we obtain [(x)]
T

LindT
T
[(c)]
T
for all terms c C.
Assume that []
T
,
LindT
T
[(x)]
T
. Without loss of generality we assume that
x is not free in (because by (T5) we know that [(x)]
T
= [(y)]
T
if y does not
occur in (x)). Thus T - (x) and so T - (x) (by rule (2)) and
T - (x)( (x)) (by rule (0)). By the -Henkin property of T we obtain a
constant d C such that T - (d). Thus nally []
T
,
LindT
T
[(d)]
T
, i.e. []
T
is not a lower bound of [(c)]
T
[ c C.
DEFINITION 4.4.5 (Canonical model). Let T a -Henkin T-theory. The canonical
model of T, denoted by CM
T
, is the T-structure LindT
T
, S where the domain of S
consists of the closed T-terms, and
f
S
(t
1
, . . . , t
n
) = f(t
1
, . . . , t
n
) for each n-ary function symbol f T,
P
S
(t
1
, . . . , t
n
) = [P(t
1
, . . . , t
n
)]
T
for each n-ary predicate symbol P T.
Nowwe can easily prove the following proposition which shows that CM
T
is indeed
a T-model of T:
PROPOSITION 4.4.6. Let T be a -Henkin T-theory. Then for each T-sentence :
1. ||
CM
T
= []
T
.
2. CM
T
[= iff T .
Thus CM
T
is an exhaustive, fully-named model of T and furthermore T is linear iff
CM
T
is an -model of T.
The following two theorems are crucial for the completeness proofs of our two
logics. We give the proof of the rst one only. The second one is much more complicated
and we postpone its proof right after the completeness theorem.
THEOREM 4.4.7. Let T be a predicate language and T a T-theory such that
T -
L
m . Then there is a predicate language T
t
T and a -Henkin T
t
-theory
(in L
m
) T
t
T such that T
t
-
L
m .
Proof. Let T
t
be an expansion of T by countably many new object constants, and take
T
t
= T
t
, T. Take any T
t
-formula (x), such that T
t
-
L
m (x)(x). Thus T
t
-
L
m
(x) and so T
t
-
L
m (c) for any c not occurring in T
t
(because T
t
contains
just T-formulae and is a nite object there always is such c T
t
and so we can use
Constants Theorem).
THEOREM 4.4.8. Let L be a nitary logic, T be a predicate language, and T a
T-theory such that T -
L
. Then there is a predicate language T
t
T and a linear
-Henkin T
t
-theory (in L) T
t
T such that T
t
-
L
.
Chapter II: A General Framework for Mathematical Fuzzy Logic 195
The proof of the next two theorems is straightforward: soundness was already es-
tablished and completeness is a corollary of Proposition 4.4.6 and Theorem 4.4.7 or
Theorem 4.4.8 respectively.
THEOREM 4.4.9 (Completeness theorem for L
m
). Let L be a logic and T a
T-theory. Then the following are equivalent:
T
L
m ,
T [=
MOD

(L)
,
There is a predicate language T
t
T such that M[= for each exhaustive,
fully named, model Mof T
t
, T.
THEOREM4.4.10 (Completeness theoremfor L). Let L be a nitary logic and T
a T-theory. Then the following are equivalent:
T
L
,
T [=
MOD

(L)
,
There is a predicate language T
t
T such that M[= for each exhaustive,
fully named, -model Mof T
t
, T.
The rest of this subsection is devoted to the promised proof of Theorem 4.4.8. To
this end, we rst need to prepare some notions and prove a crucial lemma. From now
on we work in the logic L:
DEFINITION 4.4.11 (Restricted Henkin theory). Let T T
t
be predicate languages.
A T
t
-theory T is T--Henkin if for each T-sentence (x) such that T - (x)(x) there
is a constant c T
t
such that T - (c).
Notice that when T
t
= T we obtain the already dened (without the prex T)
notion of -Henkin theory. Recall that T S means that T for each S and so
by T - S we mean that there is S such that T - .
CONVENTION 4.4.12. Let be a set of T-theories and T a T-theory. We write T '
whenever T - S for each S .
DEFINITION 4.4.13 (Deductively directed set of theories). A set of T-theories is
deductively directed if for each T, S there is R such that T R and S R;
we call R an upper bound of T and S.
We are now ready to prove the Fundamental Lemma which will have Theorem 4.4.8
as a corollary. The level of generality of our result, dealing with logics with arbitrary
p-disjunctions, forces us to use the technical complication of dealing with deductively
directed sets of theories . Theorem 4.4.8 will be an application starting from the par-
ticular case when = .
LEMMA 4.4.14 (Fundamental Lemma). Let L be a nitary logic, T a T-theory and
a deductively directed set of closed T-theories such that T ' . Then:
196 Petr Cintula and Carles Noguera
1. there is a predicate language T
t
T, a T
t
-theory T
t
T, and a deductively
directed set of closed T
t
-theories
t
, such that
T
t
'
t
,
each theory S T
t
in arbitrary language is T--Henkin whenever S '
t
,
2. there is a linear T-theory T
t
T such that T
t
' .
Proof. 1. We construct the extensions by a transnite recursion. The language T
t
is the
expansion of T by new constants c

[ < [[T[[. We also enumerate all T-formulae

with one free variable by ordinal numbers as

for < [[T[[. We construct T

t
-theories
T

and sets of closed T

t
-theories

such that T

and

for each ,
T

'

, and

is deductively directed. Let T

0
= T,
0
= , and observe that they
full our conditions.
For each [[T[[ we dene: T
<
=

<
T

and
<
=

<

. Notice that
from the induction assumption we obtain that T
<
'
<
(due to the nitarity) and

<
is deductively directed. We distinguish two possibilities:
(H1) If T
<
R(x)

(x) for some R

<
, then we dene T

= T
<

(x)

(x) and

=
<
.
(H2) Otherwise we dene T

= T
<
and

=
<
R

(c

) [ R
<
.
We show that our conditions are met no matter which possibility occurred.
(H1)

is obviously deductively directed. Assume for contradiction that T

= T
<

(x)

(x) R
t
for some R
t

. We take an upper bound

R of R and R
t
and notice that T
<
(x)

(x)

R and T
<
R

R. Thus using the
sPCP (which can be proved from the PCP as in the propositional case) we obtain
T
<
R(x)

(x)

R and so T
<


R. Since

R
<
we have a
contradiction with T
<
'
<
.
(H2) Assume by the way of contradiction that T

= T
<
R for some R

. From
the induction assumption we know that T
<
- R for each R
<
and so R has
to be of the form R
t

(c

) for some R
t

<
. Since c

does not appear in

T
<

<
, we can use Constants Theorem to obtain T

R
t

(x), and also

T

R
t
(x)

(x) (by (0)

)), a contradiction with the fact that we are in

the case (H2). To show that

is deductively directed we distinguish four cases:

rst if both R, R
t

<
then they have an upper bound already in
<
. Second
assume that R
<
and R
t
= S

(c

) for some S
<
. Let

R
<
be the upper bound of R and S. Thus

R

(c

is an upper bound of
R (trivially) and R
t
(by the sPCP and the trivial fact that

(c

) S

(c

)).
The nal two cases are analogous.
Now take T
t
= T
<]]1]]
and
t
=
<]]1]]
. Thus by the induction assumption
T
t
'
t
. Let now S be any theory such that T
t
S and S '
t
. We show that
S is T--Henkin. Clearly for each < [[T[[ if S - (x)

(x), then we must have

Chapter II: A General Framework for Mathematical Fuzzy Logic 197
used case (H2) (otherwise T

(x)

(x) and so S (x)

(x)). If S

(c

),
then S R

(c

) for any R
<
. Since we have used case (H2), we know that
R

(c

a contradiction with S '

t
.
2. We say that T is maximally consistent w.r.t. if T ' and for each , T
there is R such that T, R. By Zorns Lemma be obtain a theory T
t
T which
is maximally consistent w.r.t. . We only have to show that T
t
is linear. Assume that
/ T
t
and / T
t
. Thus there are R, S such that T
t
, R and
T
t
, S; consider an upper bound

R of R and S and using the SLP we obtain
that T
t


Ra contradiction.
Proof of Theorem 4.4.8. We construct our extension by induction over N. Take T
0
= T
and
0
= , T
0
= T. We construct predicate languages T
i
, T
i
-theories T
i
, and
deductively directed sets
i
of closed T
i
-theories, such that T
i1
T
i
, T
i1
T
i
,

i1

i
, and T
i
'
i
. Observe that the theory T
0
, set
0
, and language T
0
satisfy
T
0
'
0
. The induction step: we use part 1 of Lemma 4.4.14 for T
i
, T
i
,
i
, and dene
their successors as T
t
i
, T
t
i
, and
t
i
(the lemma assures us that our required conditions are
fullled). Then we dene T
t
=

T
i
[ i N, the T
t
-theory

T =

T
i
[ i N, and

t
=

i
[ i N. Finally, we use part 2 of Lemma 4.4.14 for T
t
,

T, and
t
and
dene T
t
as

T
t
.
Obviously T
t
is linear, T
i
T
t
, and T
t
'
i
for each i (thus in particular T
t
- ).
From part 1 of Lemma 4.4.14 and the denition of T
t
we obtain that T
t
is a T
i
--Henkin
T
t
-theory for each i, and so it is a -Henkin T
t
-theory.
REMARK 4.4.15. Notice that we have proved more: the maximal consistency of T with
respect to .
4.5 -Henkin theories, Skolemization, and witnessed semantics
In this subsection we will deal with rst-order logics L only. Our goal is twofold.
First, we study a notion of Skolemization for these logics which, provided that the prop-
erty in Corollary 4.3.10 is satised, allows to erase existential quantiers in a formula
by conservatively adding new functional symbols. Second, we deal with the particu-
lar stronger semantics of witnessed models, i.e. models where the truth value of each
quantied formula coincides with the truth-value of some of its instances. We show that
any logic L admitting Skolemization can be axiomatically extended to a logic enjoying
completeness w.r.t. witnessed models.
We start by introducing the notion of -Henkin theory, dual to the already intro-
duced (and classically equivalent) notion of -Henkin theory. It will be convenient to
restrict its validity to a class of formulae, which later will be determined by some
particular syntactical property (e.g. those starting with the connective , or formulae
satisfying excluded middle, or just all formulae). At the start, however, we need not
assume anything. In the extreme case could be just a single formula. Thus, let us x
a class of formulae of arbitrary languages.
DEFINITION 4.5.1 (-Henkin theory). Let T T
t
be predicate languages. We say
that a T
t
-theory T is:
198 Petr Cintula and Carles Noguera
-T--Henkin if for each T-formula (x) such that T (x)(x) there is
a constant c T
t
and T (c).
-Henkin if it is -Henkin and -T
t
--Henkin.
Henkin if it is -Henkin and is the class of all formulae.
DEFINITION 4.5.2 (preSkolem logic). We say that L is -preSkolem if T (c)
is a conservative expansion of T (x)(x) for each language T, each T-theory T,
each T-formula (x) and any constant c , T.
Again, if is the class of all formulae we drop the prex -.
EXAMPLE 4.5.3. In Corollary 4.3.10 we have seen that each predicate logic over an
axiomatic expansion of FL

e
is preSkolem. This includes all core fuzzy logics introduced
in Chapter I. We show additional examples of preSkolem logics based on -core fuzzy
logics.
Let L be a -core fuzzy logic, and be a class of all formulae of the form .
We show that L is -preSkolem. Let us rst recall that L enjoys the global deduction
theorem: , iff . Next assume that T (c)
L
. Then
by the deduction theorem T
L
(c) and so T
L
(c) . Thus by
the Constants Theorem and (2) we obtain T
L
(x)((x)) and so by modus
ponens, we have: T, (x)((x))
L
.
LEMMA 4.5.4 (Fundamental Lemma). Let L a -preSkolem predicate logic, T a
T-theory, and a deductively directed set of closed T-theories such that T ' . Then
there is T
t
T and a T
t
-theory T
t
T such that
T
t
' ,
each theory S T
t
in arbitrary language is -T--Henkin whenever S ' .
Proof. We construct our expansion by transnite recursion as in the proof of the rst
part of Lemma 4.4.14. Let

be the set of all T-formulae of the form (x) . We
expand our predicate language with new constants c

[ < [[

[[ and enumerate all

formulae from

by ordinal numbers as

.
We construct theories T

such that T

for and T

' . Let T
0
= T
and observe that it fulls our condition. For each we dene the set T
<
=

<
T

.
Notice that from the induction assumption and nitarity we obtain that T
<
' . We
distinguish two possibilities:
(W1) If T
<
(x)

(x) ' , we dene T

= T
<

(c).
(W2) Otherwise we dene T

= T
<
.
In the case (W1) we use the fact that T
<

(c

) is a conservative expansion of
T
<
(x)

(x) (because L is -preSkolem) to obtain T

' . In the case (W2)

we obtain it trivially.
Chapter II: A General Framework for Mathematical Fuzzy Logic 199
Take T
t
= T
<]]

]]
and observe that clearly T
t
' . Let S be a theory in an
arbitrary language such that T
t
S and S ' . We show that S is -T--Henkin. If
S (x)

(x) then we used case (W1) (fromT

<
(x)

(x) Rfor some R

we would obtain S R, a contradiction). Thus T

(c

) and so S

(c

).
THEOREM 4.5.5. The following are equivalent:
1. L is -preSkolem.
2. For each T-theory T, such that T - there is T
t
T and a linear -Henkin
T
t
-theory T
t
T and T
t
- .
Proof. Assume that L is -preSkolem and T - , some T-formulae T . We
construct our extension by induction over N. Take T
0
= T and
0
= , T
0
= T.
We construct theories
i
and T
i
, and predicate languages T
i
such that T
i
is a T
i
-theory,

i
is a directed set of T
i
-sentences, T
i
'
i
, and T
i
T
j
, T
i
T
j
and
i

j
for
i j. Observe that the theory T
0
, the set
0
and the language T
0
full these conditions.
The induction step:
If i is odd: use part 1 of Lemma 4.4.14 for T
i
, T
i
,
i
, and dene their successors
as T
t
i
, T
t
i
, and
t
i
.
If i is even: use Lemma 4.5.4 for T
i
, T
i
,
i
, and dene their successors as T
t
i
, T
t
i
,
and
i
.
Now we dene T
t
=

T
i
[ i N,

T =

T
i
[ i N, and
t
=

i
[ i N.
Finally, we use part 2 of Lemma 4.4.14 for T
t
,

T, and
t
and dene T
t
as

T
t
.
Obviously T
t
is linear, T
i
T
t
, and T
t
'
1

i
for each i. Thus from part 1 of
Lemma 4.4.14 and part 2 of Lemma 4.5.4 and the denition of T
t
we obtain that T
t
is
-Henkin.
Let us now prove the converse direction. Take T
1
= T (c) and T
2
= T
(x)(x). We show that T
2
- implies T
1
- for each formula (assuming
that c does not appear in T , ). We know that there is T
t
T and a -Henkin
T
t
-theory T
t
T
2
such that T
t
- . Since T
t
(x)(x), there is a T
t
-constant c such
that T
t
(c). Thus in any model Mof T
t
holds: M[= (c) and since CM
T
,[= the
proof is done.
Now we are ready to prove that the preSkolem property allows (in fact, it is equiva-
lent) to perform the usual process of Skolemization, i.e. introducing functional symbols
to take care of existential quantiers under the scope of universal quantiers. To this
end, we need a further technical restriction on the classes .
DEFINITION 4.5.6 (Term-closed classes). A class of formulae is term-closed if for
each formula (x, y) , language T, and each sequence of closed T-terms

t, we
have (x,

t) .
Typical examples term-closed classes are the class of all formulae, the class of all
formulae starting with (see Example 4.5.3), or the class of all provably classical
formulae (i.e. formulae such that
L
, assuming L expands FL
ew
).
200 Petr Cintula and Carles Noguera
THEOREM 4.5.7 (Skolemization). Let be a term-closes class of formulae. Then the
following are equivalent:
1. L is -preSkolem.
2. T (y)(f

(y), y) is a conservative expansion of T (y)(x)(x, y)

for each language T, each T-theory T, each T-formula (x, y) and any
functional symbol f

, T of a proper arity.
Proof. The reverse direction is trivial. The proof of the direct one is analogous to the
proof of the second part in Theorem 4.5.5. We denote T (y)(f

(y), y) as T
1
and T (y)(x)(x, y) as T
2
. We show that T
2
- implies T
1
- for each
formula . By Theorem 4.5.5 we know that there is T
t
T and a -Henkin T
t
-theory
T
t
T
2
such that T
t
- , and hence CM
t
T
,[= . For each sequence

t of closed
T
t
-terms T
t
(x)(x,

t) (by (1)) and hence there is a T

t
-constant c

t
such that T
t

(c

t
,

t) (we know that (x,

t) because is term-closed). Since c

t
is an element of
the domain of CM
t
T
, we can dene a model Mby expanding CM
t
T
with one functional
symbol dened as: (f

)
M
(

t) = c

t
. Since, for each T
t
-formula, obviously, M [= iff
CM
t
T
[= , we obtain: M [= T and M ,[= . Also clearly M [= (y)(f

(y), y), and

thus the proof is done.
Our next aims are to consider witnessed models as a meaningful semantics for rst-
order semilinear logics and axiomatize the logic complete with respect to them. They
are dened as those models where, resembling structures for rst-order classical logic,
every quantier is realized by some particular element of the domain.
DEFINITION 4.5.8 (Witnessed model). Let Q be either or . We call a T-formula
(x, y) Q-witnessed in an -model M in language T if for each a M there in an
element b M such that
|(Qx)(x, a)|
M
= |(b, a)|
M
.
Given a set of T-formulae, we call a -model M in language T -Q-witnessed if
each formula fromis Q-witnessed in M. Finally, we omit the prex Q- if the formula
(-model) is both - and -witnessed; we also omit the prex - if is the set of all
T-formulae.
DEFINITION 4.5.9 (Witnessing axioms). Given a class of formulae, we dene the
following classes
W

= (x)((y)(y, z) (x, z)) [

W

= (x)((x, z) (y)(y, z)) [

Finally, we dene W

= W

.
THEOREM 4.5.10 (Completeness w.r.t. witnessed models). Let be a term-closed
class of formulae and Q be the symbol , , or the empty sequence. Let L be W
Q

-
preSkolem. Then for each T and the following are equivalent:
Chapter II: A General Framework for Mathematical Fuzzy Logic 201
W
Q

, T .
M[= for each -Q-witnessed -model Mof T.
Proof. One direction is simple, just observe that the axioms fromW
Q

are obviously tau-

tologies in -Q-witnessed models. To prove the converse one, assume that W

, T - .
Consider a W
Q

-Henkin theory T
t
T which T
t
- (such a theory exists due to The-
orem 4.5.5). If we show that the canonical model CM
T
is -Q-witnessed, the proof is
done. Let us assume that Q = and take (x, y) and a sequence

t of elements
of the domain of the canonical model, i.e. closed terms. We know that (x,

t)
(since is term-closed) and thus T
t
(x)((y)(y,

t) (x,

t)), thus (because T

t
is W
Q

-Henkin) there has to be a constant c such that T

t
(y)(y,

t) (c,

t) and
so [(y)(y,

t)]
T
= [(c,

t)]
T
.
DEFINITION 4.5.11 (Witnessed extension). Let L be a logic. We dene the witnessed
predicate logic over L, denoted as, L
w
as the extension of L by the following witness-
ing axioms:
(x)((y)(y, z) (x, z))
(x)((x, z) (y)(y, z))
COROLLARY 4.5.12. Let L be a preSkolem logic. Then for each T, the following
are equivalent:
T
L
w .
M[= for each witnessed -model Mof T.
EXAMPLE 4.5.13. As prominent examples we check the validity of the witnessing
axioms in the rst-order versions of the three main logics based on continuous t-norms:
1. ukasiewicz logic proves both witnessing axioms, i.e.
w
= . Let us prove
them. For the rst one it is enough to show that ( (x)) (x)(
), where x is not free in , is provable (the axiom follows by taking =
(y)(y, z) and = (x, z)). proves the following implications: (x)(
) (x)( & ), (x)( & ) & (x), & (x) (
(x)), ( (x)) ( (x)). By transitivity and contra-
position we are done. Similarly, the other axiom follows from ((x) )
(x)( ), where x is not free in using ((x) ) ( (x)),
( (x)) ( (x)), ( (x)) (x)( ),
(x)( ) (x)( ).
2. Product logic proves only one of the witnessing axioms. Indeed, it can be se-
mantically shown that ( (x)) (x)( ) is a tautology (an easy
application of the fact that the implication in Product logic is left-continuous).
For the second one, we can build a counterexample for (x)(P(x) (y)P(y)).
Consider a model over the standard chain [0, 1]

such that N is its domain and

|P(n)| [ n N is a strictly decreasing sequence converging to 0.
202 Petr Cintula and Carles Noguera
3. In G odel logic both witnessing axioms fail. Indeed, for the second one we can use
the same counterexample as above (over the standard G odel chain, of course). For
the rst one, we give a counterexample to the formula (x)((y)P(y) P(x)).
Again, consider a model over the standard chain [0, 1]
G
, such that N is its domain
and |P(n)| [ n N is any strictly increasing sequence converging to r < 1.
Then |(x)((y)P(y) P(x))| = sup
nN
(sup
mN
|P(m)| |P(n)|) =
sup
nN
(r |P(n)|) = sup
nN
|P(n)| = r < 1.
5 Historical remarks and further reading
Most of the basic notions used in this chapter come from the eld of Algebraic
Logic. This discipline was born in the XIXth century with the pioneering works of
Boole, De Morgan, Pierce and others on classical logic, and has evolved into the study
(with a heavy use of tools from Universal Algebra; see e.g. [11]) of classes of algebras
providing semantics for non-classical logics. Typically, the connection between a propo-
sitional logic and a class of algebras is obtained by means of the Lindenbaum-Tarski
method. The attempts to generalize this method have given rise to Abstract Algebraic
Logic (AAL) as a natural evolution of the eld aiming to understand the process by
which a class of algebras can be associated to an arbitrary logic. Our presentation has
strongly capitalized on the notions and methods from AAL too. Readers interested in
this area and its history are referred to the survey [41] and the comprehensive mono-
graphs [25, 40, 88].
Section 2
The rst four subsections of this section can be seen as a short introduction to
AAL particularized, for didactic reasons, to the framework of weakly implicative logics
(WIL). The notion of consequence operator was introduced by Tarski in [83] and the
condition of structurality (invariance under substitutions) was added by o s and Suszko
in [66]. Mainstream AAL has been extensively developed by the Polish school in the
paradigm dened by this notion (see [25, 88]). Wojcicki introduced reduced matrices
and reduced matrix models in [87] and implicitly obtained the corresponding complete-
ness theorems. Schmidt Theorem (Theorem 2.3.3) is from [80]. The third bullet item of
Theorem 2.2.7 motivates the name Leibniz congruence (resembling Leibnizs princi-
ple of identity of indiscernibles, this result shows that a pair of formulae are congruent
iff they are indistinguishable in the matrix model). The name appeared rst in Blok and
Pigozzis paper [7]. The characterization in the third bullet item of Theorem 2.2.7 holds
for arbitrary logics and can be deduced e.g. from Maltsevs Lemma (see Lemma V.3.1
of [11]). Most results in Subsection 2.3 follow from Section 3.7 in [88], in particular, the
subdirect decomposition theorem for reduced matrix models of a nitary logic and com-
plete w.r.t. RSI reduced matrix models. Subsection 2.4 owes much to Blok and Pigozzis
1989 memoir [8].
Weakly implicative logics were rst introduced in [18] as a generalization of Ra-
siowas notion of implicative logics (see [76]), which we have called here Rasiowa-
implicative logics. Matrices for WIL coincide with the class of the so-called prestandard
matrices (see [29]), whereas the ordered matrices coincide with the so-called stan-
Chapter II: A General Framework for Mathematical Fuzzy Logic 203
dard matrices. Weakly implicative logics have been generalized in [22] to weakly p-
implicational logics, in a pure AAL fashion, by considering a generalized notion of im-
plication which, instead of being a binary connective, can be dened by sets of (possibly
parameterized) formulae. This paper also studies the position of weakly implicative log-
ics in the so-called Leibniz hierarchy: weakly implicative logics are a proper subclass
of nitely equivalential logics (and therefore of protoalgebraic logics), our algebraically
implicative logics are exactly those weakly implicative logics which are algebraizable
and, if furthermore they are nitary, they are algebraizable in the sense of Blok and
Pigozzi [8]. All the notions (except that of WIL) and results appearing in these four sub-
sections can either be found in [25] or are particularizations of notions and results from
this book to the context of weakly implicative logics. With a few exceptions,
20
all the
results proved for (nitary) weakly implicative logics hold for arbitrary (nitary) logics.
Subsections 2.5 and 2.6 are dedicated to substructural logics. These logics can
be roughly dened as those logical systems such that, when presented by means of
Gentzen-style calculus, lack some of the so-called structural rules: exchange, weak-
ening, contraction (see e.g. [75, 78, 81]). As such, this area covers a wide variety of
systems independently developed since mid XXth century, including relevant logics [1],
Linear logic [48], Lambek calculus [65], fuzzy logics presented in the previous chap-
ter, and other many-valued logics like monoidal logic [60]. In the last two decades,
Algebraic Logic has developed a uniform approach to substructural logics as the logics
of residuated lattices, i.e. propositional logics algebraizable in the sense of [8] whose
equivalent algebraic semantics is a class of residuated lattices
21
(most results are col-
lected in the recent monograph [44], where the weakest considered logic is FL). The
notion of substructural logic we have proposed here extends this approach by consid-
ering the weaker base logic SL from [46] and allowing for well-behaved expansions
and fragments. The denitions in these subsections are original, though most of the re-
sults (especially the algebraic variants of the proved rules) are folklore of the theory of
residuated structures. The axiomatic systems for prominent existing substructural logics
SL, FL, and FL
e
have been taken from [46] and [44]. Also the deduction theorems for
FL and its main axiomatic extensions were already known (see e.g. [44, 45]), but our
proofs of these theorems (using the notion of almost (MP)-based logic) and showing
their relation with proof by cases properties are novelties of the present chapter.
The nal subsection of Section 2 studies the notions of generalized disjunction de-
ned using the proof by cases property. The different variants of generalized disjunc-
tions had already been considered in the framework of Abstract Algebraic Logic in sev-
eral works (see e.g. [2325, 30, 40, 42, 8486]). Our approach follows that of [25],
where its wide generality is achieved by allowing a parameterized set of formulae in-
stead of a single formula p q, which gives rise to the notion of p-disjunction, in our
terminology. We give a more systematic account, introducing classes of logics based on
the properties of the disjunction they possess, including a new class of lattice-disjunctive
logics (logics where is interpreted as the supremum of the order given by implication),
20
The exceptions are: the rst two claims in Proposition 2.2.11, the second claim of Proposition 2.3.13, and
Theorem 2.4.5.
21
Then, classical and intuitionistic logics, although enjoying the structural rules, are included in the family
of substructural logics as extreme cases.
204 Petr Cintula and Carles Noguera
show their separations (with one open problem, see Conjecture 2.7.12). The proofs of
the Theorems 2.7.15, 2.7.20, and 2.7.23 are based on the proofs of corresponding the-
orems in [25, 2.5.1] (some of these results were known before, for detailed references
see [25]), however or formulation of these theorems differs substantially from those
in [25]. For a detailed exposition of relation of the results in this subsection and previ-
ous works on disjunction see [21] where we generalize the results for (p-)disjunctions in
nitary logics (or semilinear logics in Subsection 3.2) presented here to arbitrary logics
(non-necessarily nitary or semilinear). Finally, Theorem 2.7.29, where we show how
to use (appropriate) disjunction to axiomatize positive universal classes of reduced ma-
trices, was inspired by the paper [43] where the author proved (directly and without the
notion of disjunction) a particular version of our result for the substructural logic FL.
22
Section 3
The notions and results of the rst subsection do not have many direct predecessors:
the notion of weakly implicative semilinear logic was introduced by Cintula in [18],
under the name weakly implicative fuzzy logic. The present authors have general-
ized this class of logics to the context of weakly p-implicational (protoalgebraic) logics
in [22] (where all the results of this subsection with exception of Theorem 3.1.6, which
is new here, can be found). This paper introduced an important terminological change:
the term fuzzy (overloaded by other meanings) from [18] was replaced by (a more
neutral) term semilinear. This term was introduced by Olson and Raftery in [73] in
the context of residuated lattices to describe the varieties whose subdirectly irreducible
members are linear (following the tradition from Universal Algebra of calling a class
of algebras semiX whenever its subdirectly irreducible members have the property
X). The spiritual predecessors of this work are far more numerous: from the beginning
of Mathematical Fuzzy Logic it was clear that there are numerous logical systems de-
serving to be studied as fuzzy logics, thus already H ajek in his seminal monograph [53]
considered not one, or even a few logics, but all axiomatic extensions of his Basic Fuzzy
Logic. For the description of the way from extensions of BL to semilinear logics see the
introduction of this chapter.
The second subsection studies the interplay between semilinearity and p-disjunct-
ions. The importance of disjunction was well-known to the community of mathematical
fuzzy logicians since the inception of fuzzy logics (viz. the axiom of prelinearity or
its crucial r ole in rst-order fuzzy logics). Several known important results involving
disjunction (e.g. axiomatizations of important fuzzy logics) are obtained as corollaries
of general theorems proved here. The rst abstract study of this interplay was carried
out in [86, Section 6] (where a less general version of Theorem 3.2.4 and its corollaries
was proved) and will be generalized to weakly p-implicational logics in a forthcoming
paper.
The nal two subsections study rened completeness properties (w.r.t. distinguished
classes of algebras), which has always been the central topic in fuzzy logic literature
since the very beginning, taking in account the original motivation of fuzzy logics as
many-valued systems taking truth values from the real unit interval. At rst (in Subsec-
22
Inspecting his proof, one can notice that its main ingredient can be seen as the demonstration that a
particular set is a generalized disjunction in FL.
Chapter II: A General Framework for Mathematical Fuzzy Logic 205
tion 3.3) we have concentrated on semantics given by dense chains. The crucial density
rule originally appeared in [82] in a much more specic context, then was generalized to
a wide class of fuzzy logics in [68], and nally it has been studied in [13] in a very gen-
eral context of hypersequent calculi; however, the level of generality of this nal study
is clearly incomparable with ours (we subsume the rst two). The nal subsection gen-
eralizes the results of [19] (for (-)core fuzzy logics) to either algebraically implicative
logic or, more specically, lattice-disjunctive logics.
Section 4
Our approach to predicate logics follows that of Rasiowa from [76], where she gen-
eralizes the Rasiowa-Sikorski-style intuitionistic rst-order predicate logic [77] to the
class of logics we call Rasiowa-implicative logics. It starts from a propositional logic,
which enjoys an equivalent algebraic semantics, and an implication connective dening
an order relation on the algebras that allows to interpret the existential (resp. univer-
sal) quantication of a formula as the supremum (resp. inmum) of the values of its
instances. The main result is the completeness w.r.t. all algebras of truth values. We can
easily generalize this approach to weakly implicative logics, but in order to obtain com-
pleteness w.r.t. linearly ordered algebras for semilinear logics (as in the propositional
case) some additional axioms are necessary. Historically speaking the rst such example
is G odel-Dummett rst-order predicate logic axiomatized in [61] (relative to the logic
axiomatized in Rasiowas way) by adding the axiom (3): (x)( ) (x)
(for x not free in ).
23
Using this axiom many other rst-order logics complete w.r.t.
their linearly ordered algebras were axiomatized (see e.g. [16, 33, 53]). The next step
forward was [56], where the proof of completeness was not only generalized to arbi-
trary languages (previous proofs were restricted to countable languages) but performed
uniformly for all (-)core fuzzy logics (identifying the crucial r ole of disjunction in
the process). Another noteworthy case is [20] where the rst-order implicational frag-
ment of MTL was axiomatized, using (as was later observed) axiom(3) written for the
generalized (!) disjunction of MTL.
All subsections of this section, except for the last one, are dedicated to formalizing
these ideas. The last subsection studies the so-called -preSkolem logics, where we
can prove a general form of Skolemization (already studied before in fuzzy logics, e.g.
in [4]) and completeness w.r.t. witnessed models (i.e. models where the truth value of
each quantied formula coincides with the truth-value of some of its instances). These
models were rst considered in [55] in the context of ukasiewicz logic (see also [12]
for a weaker notion useful in product logics). Our completeness theorem generalizes
the one proved in [56] for core fuzzy logics to all preSkolem logics (which include all
semilinear axiomatic expansions of the uninorm logic UL).
Other current trends in the research on rst-order fuzzy logics that have not been
covered here include the study of particular rst-order G odel logics (see e.g. [5]), the
development of a model theory for fuzzy logics (see [19, 27, 56]) and works on descrip-
tion fuzzy logics (see [2, 55]).
23
In fact the standard ukasiewicz rst-order logic was axiomatized earlier in [59], but this is a peculiar
case for two reasons: rst in this logic the semantics given by all algebras and by chains coincide, and second
the actually axiomatized logic in this paper is the rst-order innitary logic of the standard MV-algebra.
206 Petr Cintula and Carles Noguera
Acknowledgments
We want to thank Vincenzo Marra, James Raftery,
`
Angel Garca Cerdana, and Ven-
tura Verd u for their reading of preliminary versions of the chapter and their helpful
comments and corrections. Petr Cintula has been supported by the grant ICC/08/E018
of the Czech Science Foundation (part of ESF Eurocores project LogiCCC FP006 Lo-
MoReVI) and the Institutional Research Plan AV0Z10300504. Carles Noguera ac-
knowledges support of the research contract Juan de la Cierva JCI-2009-05453, the
MICINN project TASSAT (TIN2010-20967-C04-01), the ESF Eurocores-LogICCC /
MICINN project FFI2008-03126-E/FILO, and the FP7-PEOPLE-2009-IRSES project
MaToMUVI (PIRSES-GA-2009- 247584).
BIBLIOGRAPHY
[1] Alan Ross Anderson and Nuel D. Belnap. Entailment: The Logic of Relevance and Necessity, volume 1.
Princeton University Press, Princeton, 1975.
[2] Eva Armengol and
`
Angel Garc`a-Cerda na. Fuzzy description logics and t-norm based fuzzy logics.
International Journal of Approximate Reasoning, 51(6):632655, 2002.
[3] Matthias Baaz. Innite-valued G odel logic with 0-1-projections and relativisations. In Petr H ajek, editor,
G odel96: Logical Foundations of Mathematics, Computer Science, and Physics, volume 6 of Lecture
Notes in Logic, pages 2333. Springer-Verlag, Brno, 1996.
[4] Matthias Baaz and George Metcalfe. Herbrands theorem, Skolemization, and proof systems for rst-
order ukasiewicz logic. Journal of Logic and Computation, 20(1):3554, 2010.
[5] Matthias Baaz, Norbert Preining, and Richard Zach. First-order G odel logics. Annals of Pure and
Applied Logic, 147(12):2347, 2007.
[6] Libor B ehounek and Petr Cintula. Fuzzy logics as the logics of chains. Fuzzy Sets and Systems,
157(5):604610, 2006.
[7] Willem J. Blok and Don L. Pigozzi. Protoalgebraic logics. Studia Logica, 45:337369, 1986.
[8] Willem J. Blok and Don L. Pigozzi. Algebraizable Logics, volume 396 of Memoirs of the American
Mathematical Society. American Mathematical Society, Providence, RI, 1989. Freely downloadable
from http://orion.math.iastate.edu/dpigozzi/.
[9] F elix Bou, Francesc Esteva, Josep Maria Font,
`
Angel Gil, Llus Godo, Antoni Torrens, and Ventura
Verd u. Logics preserving degrees of truth from varieties of residuated lattices. Journal of Logic and
Computation, 19(6):10311069, 2009.
[10] F elix Bou, Francesc Esteva, Llus Godo, and Ricardo O. Rodrguez. On the minimum many-valued
modal logic over a nite residuated lattice. Journal of Logic and Computation, 21(5):739790, 2011.
[11] Stanley Burris and H.P. Sankappanavar. A Course in Universal Algebra, volume 78 of Graduate Texts
in Mathematics. Springer-Verlag, 1981.
[12] Marco Cerami and Francesc Esteva. Strict core fuzzy logics and quasi-witnessed models. To appear in
Archive for Mathematical Logic.
[13] Agata Ciabattoni and George Metcalfe. Density elimination. Theoretical Computer Science, 403(1
2):328346, 2008.
[14] Roberto Cignoli, Itala M.L. DOttaviano, and Daniele Mundici. Algebraic Foundations of Many-Valued
Reasoning, volume 7 of Trends in Logic. Kluwer, Dordrecht, 1999.
[15] Roberto Cignoli, Francesc Esteva, Llus Godo, and Antoni Torrens. Basic fuzzy logic is the logic of
continuous t-norms and their residua. Soft Computing, 4(2):106112, 2000.
[16] Petr Cintula. The and
1
2
propositional and predicate logics. Fuzzy Sets and Systems, 124(3):289
302, 2001.
[17] Petr Cintula. From Fuzzy Logic to Fuzzy Mathematics. PhD thesis, Czech Technical University, Faculty
of Nuclear Sciences and Physical Egineering, Prague, 2005.
[18] Petr Cintula. Weakly implicative (fuzzy) logics I: Basic properties. Archive for Mathematical Logic,
45(6):673704, 2006.
[19] Petr Cintula, Francesc Esteva, Joan Gispert, Llus Godo, Franco Montagna, and Carles Noguera. Distin-
guished algebraic semantics for t-norm based fuzzy logics: Methods and algebraic equivalencies. Annals
of Pure and Applied Logic, 160(1):5381, 2009.
Chapter II: A General Framework for Mathematical Fuzzy Logic 207
[20] Petr Cintula, Petr H ajek, and Rostislav Hor ck. Formal systems of fuzzy logic and their fragments.
Annals of Pure and Applied Logic, 150(13):4065, 2007.
[21] Petr Cintula and Carles Noguera. The proof by cases property and its variants in structural consequence
relations. Submitted.
[22] Petr Cintula and Carles Noguera. Implicational (semilinear) logics I: A new hierarchy. Archive for
Mathematical Logic, 49(4):417446, 2010.
[23] Janusz Czelakowski. Logical matrices, primitive satisfaction and nitely based logics. Studia Logica,
42(1):89104, 1983.
[24] Janusz Czelakowski. Remarks on nitely based logics. In Proceedings of the Logic Colloquium 1983.
Vol. 1. Models and Sets, volume 1103 of Lecture Notes in Mathematics, pages 147168. Springer, Berlin,
1984.
[25] Janusz Czelakowski. Protoalgebraic Logics, volume 10 of Trends in Logic. Kluwer, Dordrecht, 2001.
[26] Janusz Czelakowski and Wieslaw Dziobiak. Congruence distributive quasivarieties whose nitely sub-
directly irreducible members form a universal class. Algebra Universalis, 27(1):128149, 1990.
[27] Pilar Dellunde. Preserving mappings in fuzzy predicate logics. To appear in Journal of Logic and
Computation.
[28] Antonio Diego. Sur les algebres de Hilbert. Collection de logique Math ematique. Ser. A. (Ed Hermann),
21:152, 1966.
[29] J. Michael Dunn and Gary M. Hardegree. Algebraic Methods in Philosophical Logic, volume 41 of
Oxford Logic Guides. Oxford University Press, Oxford, 2001.
[30] Wojciech Dzik. On the content of lattices of logics part 1: The representation theorem for lattices of
logics. Reports on Mathematical Logic, 13:1728, 1981.
[31] Francesc Esteva, Joan Gispert, Llus Godo, and Franco Montagna. On the standard and rational com-
pleteness of some axiomatic extensions of the monoidal t-norm logic. Studia Logica, 71(2):199226,
2002.
[32] Francesc Esteva, Joan Gispert, Llus Godo, and Carles Noguera. Adding truth-constants to logics of
continuous t-norms: Axiomatization and completeness results. Fuzzy Sets and Systems, 158(6):597
618, 2007.
[33] Francesc Esteva and Llus Godo. Monoidal t-norm based logic: Towards a logic for left-continuous
t-norms. Fuzzy Sets and Systems, 124(3):271288, 2001.
[34] Francesc Esteva, Llus Godo, Petr H ajek, and Franco Montagna. Hoops and fuzzy logic. Journal of
Logic and Computation, 13(4):532555, 2003.
[35] Francesc Esteva, Llus Godo, and Franco Montagna. The and
1
2
logics: Two complete fuzzy
systems joining ukasiewicz and product logics. Archive for Mathematical Logic, 40(1):3967, 2001.
[36] Francesc Esteva, Llus Godo, and Carles Noguera. First-order t-norm based fuzzy logics with truth-
constants: Distinguished semantics and completeness properties. Annals of Pure and Applied Logic,
161(2):185202, 2009.
[37] Francesc Esteva, Llus Godo, and Carles Noguera. Expanding the propositional logic of a t-norm with
truth-constants: Completeness results for rational semantics. Soft Computing, 14(3):273284, 2010.
[38] Francesc Esteva, Llus Godo, and Carles Noguera. On expansions of WNM t-norm based logics with
truth-constants. Fuzzy Sets and Systems, 161(3):347368, 2010.
[39] Tommaso Flaminio. Strong non-standard completeness for fuzzy logics. Soft Computing, 12(4):321
333, 2008.
[40] Josep Maria Font and Ramon Jansana. A General Algebraic Semantics for Sentential Logics, volume 7
of Lecture Notes in Logic. Association for Symbolic Logic, Ithaca, NY, 2 edition, 2009. Freely down-
loadable from http://www.mat.ub.edu/

jmfont/publs.html.
[41] Josep Maria Font, Ramon Jansana, and Don L. Pigozzi. A survey of Abstract Algebraic Logic. Studia
Logica, 74(12, Special Issue on Abstract Algebraic Logic II):1397, 2003.
[42] Josep Maria Font and Ventura Verd u. Algebraic logic for classical conjunction and disjunction. Studia
Logica, 50(34):391419, 1991.
[43] Nikolaos Galatos. Equational bases for joins of residuated-lattice varieties. Studia Logica, 76(2):227
240, 2004.
[44] Nikolaos Galatos, Peter Jipsen, Tomasz Kowalski, and Hiroakira Ono. Residuated Lattices: An Alge-
braic Glimpse at Substructural Logics, volume 151 of Studies in Logic and the Foundations of Mathe-
matics. Elsevier, Amsterdam, 2007.
[45] Nikolaos Galatos and Hiroakira Ono. Algebraization, parametrized local deduction theorem and inter-
polation for substructural logics over FL. Studia Logica, 83(13):279308, 2006.
[46] Nikolaos Galatos and Hiroakira Ono. Cut elimination and strong separation for substructural logics: An
algebraic approach. Annals of Pure and Applied Logic, 161(9):10971133, 2010.
208 Petr Cintula and Carles Noguera
[47] Nikolaos Galatos and James G. Raftery. Adding involution to residuated structures. Studia Logica,
77(2):181207, 2004.
[48] Jean-Yves Girard. Linear logic. Theoretical Computer Science, 50(1):1102, 1987.
[49] Siegfried Gottwald. Fuzzy Sets and Fuzzy Logic: Foundations of Applicationfrom a Mathematical
Point of View. Vieweg, Wiesbaden, 1993.
[50] Petr H ajek. Fuzzy logic and arithmetical hierarchy. Fuzzy Sets and Systems, 73(3):359363, 1995.
[51] Petr H ajek. Fuzzy logic and arithmetical hierarchy II. Studia Logica, 58(1):129141, 1997.
[52] Petr H ajek. Basic fuzzy logic and BL-algebras. Soft Computing, 2(3):124128, 1998.
[53] Petr H ajek. Metamathematics of Fuzzy Logic, volume 4 of Trends in Logic. Kluwer, Dordrecht, 1998.
[54] Petr H ajek. Fuzzy logics with noncommutative conjunctions. Journal of Logic and Computation,
13(4):469479, 2003.
[55] Petr H ajek. Making fuzzy description logic more general. Fuzzy Sets and Systems, 154(1):115, 2005.
[56] Petr H ajek and Petr Cintula. On theories and models in fuzzy predicate logics. Journal of Symbolic
Logic, 71(3):863880, 2006.
[57] Petr H ajek, Llus Godo, and Francesc Esteva. Fuzzy logic and probability. In Proceedings of the 11th
Annual Conference on Uncertainty in Articial Intelligence UAI 95, pages 237244, Montreal, 1995.
[58] Petr H ajek, Llus Godo, and Francesc Esteva. A complete many-valued logic with product conjunction.
Archive for Mathematical Logic, 35(3):191208, 1996.
[59] Louise Schmir Hay. Axiomatization of the innite-valued predicate calculus. Journal of Symbolic Logic,
28(1):7786, 1963.
[60] Ulrich H ohle. Commutative, residuated l-monoids. In Ulrich H ohle and Erich Peter Klement, editors,
Non-Classical Logics and Their Applications to Fuzzy Subsets, pages 53106. Kluwer, Dordrecht, 1995.
[61] Alfred Horn. Logic with truth values in a linearly ordered Heyting algebras. Journal of Symbolic Logic,
34(3):395408, 1969.
[62] Pawe Idziak. Lattice operations in BCK-algebras. Mathematica Japonica, 29(6):839846, 1982.
[63] S andor Jenei. On the structure of rotation-invariant semigroups. Archive for Mathematical Logic,
42(5):489514, 2003.
[64] S andor Jenei and Franco Montagna. A proof of standard completeness for Esteva and Godos logic
MTL. Studia Logica, 70(2):183192, 2002.
[65] Joachim Lambek. The mathematics of sentence structure. American Mathematical Monthly, 65:154
170, 1958.
[66] Jerzy o s and Roman Suszko. Remarks on sentential logics. Indagationes Mathematicae, 20:177183,
1958.
[67] J.C.C. McKinsey. Proof of the independence of the primitive symbols of Heytings calculus of proposi-
tions. Journal of Symbolic Logic, 4(4):155158, 1939.
[68] George Metcalfe and Franco Montagna. Substructural fuzzy logics. Journal of Symbolic Logic,
72(3):834864, 2007.
[69] Franco Montagna. An algebraic approach to propositional fuzzy logic. Journal of Logic, Language and
Information, 9:91124, 2000.
[70] Franco Montagna and Carles Noguera. Arithmetical complexity of rst-order predicate fuzzy logics
over distinguished semantics. Journal of Logic and Computation, 20(2):399424, 2010.
[71] Carles Noguera. Algebraic Study of Axiomatic Extensions of Triangular Norm Based Fuzzy Logics,
volume 27 of Monograes de lInstitut dInvestigaci o en Intellig` encia Articial. Consell Superior
dInvestigacions Cientques, Barcelona, 2007.
[72] Vil em Nov ak. On the syntactico-semantical completeness of rst-order fuzzy logic part I (syntax and
semantic), part II (main results). Kybernetika, 26:4766, 134154, 1990.
[73] Jeffrey S. Olson and James G. Raftery. Positive Sugihara monoids. Algebra Universalis, 57(1):7599,
2007.
[74] Hiroakira Ono. Substructural logics and residuated latticesan introduction. In Vincent F. Hendricks
and Jacek Malinowski, editors, 50 Years of Studia Logica, volume 21 of Trends in Logic, pages 193228.
Kluwer, Dordrecht, 2003.
[75] Francesco Paoli. Substructural Logics: A Primer, volume 13 of Trends in Logic. Kluwer, Dordrecht,
2002.
[76] Helena Rasiowa. An Algebraic Approach to Non-Classical Logics. North-Holland, Amsterdam, 1974.
[77] Helena Rasiowa and Roman Sikorski. The Mathematics of Metamathematics. Panstwowe Wydawnictwo
Naukowe, Warsaw, 1963.
[78] Greg Restall. An Introduction to Substructural Logics. Routledge, New York, 2000.
[79] Kentaro Sato. Proper semantics for substructural logics, from a stalker theoretic point of view. Studia
Logica, 88(2):295324, 2008.
Chapter II: A General Framework for Mathematical Fuzzy Logic 209
[80] J urgen Schmidt.

Uber die Rolle der transniten Schlu?weisen in einer allgemeinen Idealtheorie. Math-
ematische Nachrichten, 7:165182, 1952.
[81] Peter Schroeder-Heister and Kosta Dosen, editors. Substructural Logics, volume 2 of Studies in Logic
and Computation. Oxford University Press, Oxford, 1994.
[82] Gaisi Takeuti and Satoko Titani. Intuitionistic fuzzy logic and intuitionistic fuzzy set theory. Journal of
Symbolic Logic, 49(3):851866, 1984.
[83] Alfred Tarski.

Uber einige fundamentale Begri?e der Metamathematik. C. R. Soci et e des Sciences et
Letters Varsovie, cl. III, 23:2229, 1930.
[84] Antoni Torrens and Ventura Verd u. Distributivity and irreducibility in closure systems. Technical report,
Faculty of Mathematics, University of Barcelona, Barcelona, 1982.
[85] Ventura Verd u. L` ogiques distributives i booleanes. Stochastica, 3:97108, 1979.
[86] San-Min Wang and Petr Cintula. Logics with disjunction and proof by cases. Archive for Mathematical
Logic, 47(5):435446, 2008.
[87] Ryszard W ojcicki. Matrix approach in the methodology of sentential calculi. Studia Logica, 32:737,
1973.
[88] Ryszard W ojcicki. Theory of Logical Calculi, volume 199 of Synthese Library. Kluwer Academic
Publishers, Dordrecht/Boston/London, 1988.
PETR CINTULA
Institute of Computer Science
Academy of Sciences of the Czech Republic
Pod Vod arenskou v e z 2
182 07 Prague 8, Czech Republic
Email: cintula@cs.cas.cz
CARLES NOGUERA
Articial Intelligence Research Institute (IIIA), CSIC
Campus de la Universitat Aut ` onoma de Barcelona s/n
08193 Bellaterra, Catalonia, Spain
Email: cnoguera@iiia.csic.es
Chapter III
Proof Theory for Mathematical Fuzzy Logic
GEORGE METCALFE
1 Introduction
As clearly demonstrated in other chapters of this handbook, fuzzy logics are moti-
vated rst and foremost by semantic considerations: in particular, by the goal of repre-
senting and reasoning about truth degrees. However, as logics, they evidently deal also
with the notion of proof. This much is apparent in Hilbert-style axiom systems, which
(with enough patience) will churn out each theorem of the logic using a few simple rules
and a list of axiom schema. But while Hilbert systems are a convenient formalism for
presenting logics corresponding to classes of algebras, they are not so exible when it
comes to searching for, analyzing, and reasoning about proofs. At each step in a Hilbert
system proof, the next axiom or instance of a rule like modus ponens should be guessed.
Much better are proof systems with more restrictions on how to proceed: ideally, sys-
tems where proofs are analytic, built from the raw material (subformulas) of the formula
to be proved. Such systems and their applications will be the subject of this chapter.
Surprisingly perhaps, most fuzzy logics investigated in the literature have a very
natural proof-theoretic formulation. They occur (alongside intuitionistic logic, relevant
logics, linear logic, Lambek calculus, etc.) as substructural logics in the framework of
Gentzen systems. Typically, such systems gain exibility by not dealing directly with
formulas, but rather with sequents: ordered pairs of sequences (or sets or multisets) of
formulas. So-called sequent calculi provide a natural home for a range of logics from
linguistics, philosophy, computer science, and mathematics, as well as corresponding
directly to some well studied classes of algebras. However, for fuzzy logics, sequents
are not quite enough. A step up in complexity is required to hypersequents: multisets
of sequents interpreted as disjunctions. Proof systems for many fuzzy logics are then
obtained simply by transferring sequent calculi to the hypersequent level and adding an
extra rule (or two). For example, a hypersequent calculus for G odel logic is obtained
by allowing hypersequent contexts in Gentzens sequent calculus for intuitionistic logic
and adding a rule permitting communication between sequents.
The goal of this chapter, however, is not only to show that fuzzy logics can be
presented proof-theoretically, but also to show why proof theory matters for this eld.
What can be done with analytic proof systems that cannot be achieved with a purely
algebraic approach? For many substructural logics, a standard answer would be that
these systems are essential for establishing algorithmic properties such as decidability,
212 George Metcalfe
complexity, and interpolation. In particular, the only known proofs of decidability for
the full Lambek calculus (which extends even to the rst-order level) or, equivalently,
the equational theory of (pointed) residuated lattices, make use of Gentzen systems.
For hypersequents, the situation is not so straightforward, however. Decidability and
complexity results can be obtained in certain cases, but sometimes the mere existence
of an analytic calculus is of no help whatsoever. So why care about developing proof
theory for fuzzy logics?
This chapter concentrates on two important applications. First, proof theory pro-
vides a novel approach to tackling the central problem in mathematical fuzzy logic of
standard completeness: showing that a logic is complete with respect to the intended
fuzzy semantics, or, equivalently, showing that a variety of algebras is generated by cer-
tain distinguished members. Roughly, the idea is to add a special density rule to a
logic that guarantees standard completeness and then to show proof-theoretically that it
can be eliminated. For example, this method provides the only known proof of the stan-
dard completeness of uninorm logic, or, equivalently, the generation of the variety of
semilinear bounded pointed commutative residuated lattices by its members with lattice
reduct [0, 1]. The second key application of proof theory considered here is the extension
of propositional fuzzy logics to the rst-order level, a step that is somewhat problematic
algebraically but completely natural for Gentzen systems. The proof-theoretic presen-
tation facilitates investigation of key topics such as Herbrands theorem and Skolemiza-
tion, as well as providing a means for tackling (fragments of) non recursively axiom-
atizable cases such as rst-order ukasiewicz logic. Also discussed are the (hitherto
proof-theoretically rather underdeveloped) topics of propositional fuzzy logics extended
with modalities and propositional quantiers.
The focus in this chapter will be on concepts and ideas, avoiding the temptation to
generalize or over-complicate, but important techniques and landmark results will be
presented in detail. Proofs, or at least proof sketches, are provided except in cases bear-
ing signicant similarity to preceding results or involving considerable technicalities.
Nevertheless, due to size limitations, several interesting topics have been omitted. In
particular, only commutative innite-valued logics are considered here, avoiding well-
developed but rather distinct approaches to nite-valued logics and similar but more
intricate presentations of noncommutative fuzzy logics. We also limit our attention to
Gentzen systems. Other proof frameworks such as natural deduction, tableaux, resolu-
tion, and display logic, to name just a few, each have their own distinct advantages and
disadvantages. However, not only do Gentzen systems appear to be the most natural
and useful formalism in the context of fuzzy logics, but also ideas and results from this
domain can typically be transferred to other frameworks. For further details, we recom-
mend consulting the monograph [53], and the historical remarks collected at the end of
the chapter.
2 A Gentzen systems primer
In this section, we introduce and begin to explore some of the main ideas underly-
ing Gentzen systems, in particular, how and when they can be developed and what they
might be good for. We focus on three case studieslattices, classical logic, and sub-
Chapter III: Proof Theory for Mathematical Fuzzy Logic 213
structural logicsto showcase the diversity of the framework, and end with the obvious
question: how can we extend these methods to fuzzy logics?
214 George Metcalfe
2.1 Rules, systems, derivations
Let us start with some general denitions that will serve us well in what follows.
First, note that the objects of proof systems can be more complicated entities than for-
mulas: equations, sequents, hypersequents, sequents of sequents, and so on. We there-
fore assume only that these objects constitute a set of structures. An (inference) rule
(r) for is then a set of ordered pairs , a called inferences where is a -
nite (possibly empty) set of premises and a is the conclusion, typically written for
= a
1
, . . . , a
n
as a
1
, . . . , a
n
/ a or
a
1
... a
n
a
. An inference with no premises is called
an axiom. Usually, axioms and rules are dened via schema that use meta-variables to
stand for arbitrary members of or sets based on such as multisets of members of .
A proof system C is an ordered pair , R consisting of a set of structures and
a set R of rules for . A C-derivation d of a from is a nite tree
1
of
height h(d) labelled by members of such that: (1) a labels the root; (2) for each node
labelled a
0
, either a
0
or its child nodes are labelled a
1
, . . . , a
n
and a
1
, . . . , a
n
/ a
0
is an instance of a rule of C. If there is a C-derivation d of a from , then a is said
to be C-derivable from , written d;
C
a or simply
C
a. More generally, a
rule is C-derivable if for each instance a
1
, . . . , a
n
/ a of the rule, a is C-derivable from
a
1
, . . . , a
n
. Finally, a rule is C-admissible if for each instance a
1
, . . . , a
n
/ a of the
rule, whenever
C
a
i
for i = 1 . . . n, then
C
a.
Despite the generality of these denitions, in practice we deal exclusively with struc-
tures built from formulas. Recall that a propositional language / consists of a set of
symbols (connectives) with xed arities and that the set of formulas Fm
1
is the domain
of the absolutely free algebra Fm
1
of / over a xed countably innite set of variables.
We use the (subscripted) symbols p, q, r and , , to denote arbitrary variables and
formulas, respectively, and dene cp(), the complexity of , to be the number of oc-
currences of connectives in . Connectives of arity 0 are called constants, while both
variables and constants are called atoms or atomic formulas.
For simplicity, we mostly restrict our attention in what follows to the language /
p
with binary connectives , , &, and constants 0, 1, , , dening also =
def
0, =
def
,
0
= 1, and
n+1
= &
n
for n N. Note
only that in general our methods apply uniformly to fragments of /
p
, and in particular
to the languages obtained by dropping the constants , (e.g., for FL
e
-algebras) or
, , 0 (e.g., for commutative residuated lattices). A useful example of a proof sys-
tem HMAILL for the set of formulas Fm
1
p
, corresponding to multiplicative additive
intuitionistic linear logic, is presented in Figure 1.
2.2 A proof system for lattices
Let us begin with a simple but relevant case study. Recall that the class 'A of
lattices can be dened in at least two ways. Order-theoretically, a lattice is a partially
ordered set L, such that for all x, y L, both the meet xy and join xy exist in L.
1
Recall that a nite tree is a nite poset P, ) with a distinguished element x
0
P called the root such
that x
0
x for all x P and y P [ y x, ) is linearly ordered for all x P. The members
of P are called nodes and each node x such that y P [ x < y = is called a leaf. For each leaf
x, the set y P [ y x is called a branch. A node x is a child of a parent node y if y < x and
z P [ y < z < x = . The height of the tree is sup[y P [ y x[ [ x P.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 215
(B) ( ) (( ) ( ))
(C) ( ( )) ( ( ))
(I)
(&1) ( ( &))
(&2) ( ( )) (( &) )
(11) 1
(12) (1 )
(1) ( )
(2) ( )
(3) (( ) ( )) ( ( ))
(1) ( )
(2) ( )
(3) (( ) ( )) (( ) )
()
()

(MP)


(ADJ)
Figure 1. The Hilbert system HMAILL
Algebraically, a lattice is an algebra L, , for a language /
l
with binary operations
and such that dening x y iff (if and only if) x y = x ensures that L, is
a lattice with meet and join in the order-theoretic sense. Alternatively, lattices are
algebras L, , satisfying the equations:
p (q r) (p q) r p (q r) (p q) r
p q q p p q q p
p (p q) p p (p q) p.
It would be convenient, however, to have also an algorithmic presentation of 'A: a
method for determining when an arbitrary lattice equation holds in all lattices. Let us
therefore restrict our attention to very simple axioms, inequations of the form
(where stands for ), and dene rules that decompose formulas by
introducing operation symbols on either side of the inequation. We also include (for
now) a cut rule corresponding to the transitivity of the relation . The resulting proof
system GLat is displayed in Figure 2, where rules are presented schematically using
, , (subscripted) to stand for arbitrary lattice formulas.
EXAMPLE 2.2.1. Derivations in GLat are nite trees labelled with inequations; e.g.,
q q
(ID)
p q q
()
2
p p
(ID)
p q p
()
1
p q q p
()
p p
(ID)
p p
(ID)
p p p
()
establish commutativity and half of the idempotency law for .
216 George Metcalfe
Axioms Cut rule

(ID)


(CUT)
Left logical rules Right logical rules

1

2

()
1

1

1

2
()
1

1

2

()
2

2

1

2
()
2

1

2

1

2

()

1

2

1

2
()
Figure 2. The proof system GLat
As intended, GLat really is a proof system for 'A. One direction is easy to check.
Clearly the axioms hold in all lattices, and it is not hard to see that if the premises of an
instance of a rule hold in all lattices, then so does the conclusion. Hence, by a simple
induction on the height of a derivation,
GLat
implies [=
LAT
.
2
For
the opposite direction, we dene (as for Hilbert systems) a Lindenbaum-Tarski (or free)
algebra. First we introduce a binary relation on Fm
1
l
by stipulating that
, i (
GLat
and
GLat
).
This relation is reexive by (ID), symmetric by denition, and transitive by (CUT),
and therefore an equivalence relation on Fm
1
l
. In fact, it is a congruence on the lat-
tice formula algebra Fm
1
l
. For example, if
1
,
1
and
2
,
2
, then

1

2
,
1

2
as established by the derivations:
.
.
.

1

1

1

2

1
()
1
.
.
.

2

2

1

2

2
()
2

1

2

1

2
()
.
.
.

1

1

1

2

1
()
1
.
.
.

2

2

1

2

2
()
2

1

2

1

2
()
Moreover, the quotient algebra Fm
1
l
/ is a lattice. For example, is commutative in
virtue of the derivation:

(ID)

()
2

(ID)

()
1

()
2
For a class of algebras K of the same language / and a set of equations for /, recall that
[=
K
means that for each A K and homomorphism e: Fm
L
A, whenever ker e, also
ker e. By convention, we write [=
K
if = and [=
A
if K = A.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 217
Observe also that in the algebra Fm
1
l
/, for any , Fm
1
l
:
/ / i / / = /
i ( )/ = /
i ,
i
GLat
and
GLat

i
GLat
.
Suppose that [=
LAT
. Then in particular, e() e() in Fm
1
l
/ where e
is the natural homomorphism e: Fm
1
l
Fm
1
l
/ dened by e() = /. So
/ / in Fm
1
l
/, and by the above reasoning,
GLat
. We have
therefore established:
THEOREM 2.2.2.
GLat
iff [=
LAT
.
Although proof systems can have any axioms and rules whatsoever, they are often
designed carefully to have special properties. In particular, notice that every rule of
GLat except (CUT) has the so-called subformula property. Namely, every formula
occurring in the premises of an instance of such a rule occurs as a subformula of a for-
mula in the conclusion. Such rules are often also called analytic since they decompose
(reading upwards) formulas into their constituent subformulas. Derivations involving
only these rules are much easier to investigate. In particular, we can see easily that start-
ing with any inequation and applying the rules of GLat except (CUT) backwards, we
always terminate with inequations containing only variables.
It would be nice therefore to restrict our attention to cut-free derivations in GLat
that make no use of the cut rule. This is possible so long as we are concerned only with
inequations derivable from the empty set of inequations. For any proof system GL with
a cut rule (CUT), let us x GL

to be GL without (CUT). We will say that GL admits

cut elimination if there is an algorithm which given a derivation of a structure from in
GL, outputs a derivation of the structure from in GL

.
THEOREM 2.2.3. GLat admits cut elimination.
Proof. It is enough to give a constructive proof of the following:
Claim: if d
1

GLat
and d
2

GLat
, then
GLat
.
We prove the claim by induction on h(d
1
) + h(d
2
). First, let us consider the case when
h(d
1
) = 0 or h(d
2
) = 0. Suppose without loss of generality that d
1
consists of an
instance of the rule (ID). Then = , so clearly
GLat
using the derivation d
2
.
For the inductive step there are two cases.
First, suppose that one of the derivations d
1
and d
2
ends with an instance of a rule
that decomposes or , respectively. For example, suppose that =
1

2
and d
1
ends with
.
.
.

1

.
.
.

1

2

()
218 George Metcalfe
Let d
11
and d
12
be the derivations of
1
and
2
, respectively. Since
h(d
11
) < h(d
1
) and h(d
12
) < h(d
1
), by two applications of the induction hypoth-
esis,
GLat

1
and
GLat

2
. But then by an application of (),
also
GLat

1

2
as required.
Second, suppose that both derivations end with rules that decompose . Consider
the case where =
1

2
. Then d
1
and d
2
end with
.
.
.

1
.
.
.

2

1

2
()
.
.
.

1

2

()
i
where i 1, 2. Let d
1i
be the derivation ending with
i
and let d
2i
be the
derivation ending with
i
. Since h(d
1i
) < h(d
1
) and h(d
2i
) < h(d
2
), by the
induction hypothesis,
GLat
, as required.
Hence, as argued above, we obtain a decision procedure for checking whether an
equation holds in all lattices.
COROLLARY 2.2.4. The equational theory of lattices is decidable.
Let us also mention that we obtain immediately the admissibility of certain rules
for lattices (or, equivalently, validity of universal formulas in free lattices), including
Whitmans condition:
[=
LAT

1

2

1

2
= [=
LAT

1

2

1
or [=
LAT

1

2

2
or
[=
LAT

1

1

2
or [=
LAT

2

1

2
.
Just consider the last possible application of a rule in any derivation of an inequation

1

2

1

2
in GLat

.
2.3 A sequent calculus for classical logic
Inequations are suitable structures for reasoning about lattices, but most logics and
classes of algebras requires something a bit more complicated. Consider, for example,
the following inequation corresponding to the distributivity law:
(p q) r (p r) (q r).
Quite rightly, this inequation is not derivable in GLat. The only step that we could take
(working backwards) in a cut-free derivation would be to apply ()
1
, ()
2
, ()
1
,
or ()
2
and it is easy to see that none of the resulting inequations are derivable for
these cases. So what rule(s) might we add to obtain distributive lattices? Somehow we
need to represent the on the left of inequations and perhaps also the on the right.
The solution of Gentzen to this problem was to consider sequences of formulas di-
vided by a sequent arrow, the left sequence representing a conjunction of formulas, the
right, a disjunction. However, since we deal only with commutative logics in this chap-
ter, we will make use of multisets rather than sequences of formulas.
3
Finite multisets
3
Formally, a multiset over a set Ais an ordered pair A, f) where f is a function f : A N, called nite if
x A [ f(x) > 0 is nite. For multisets A, f
1
) and A, f
2
), A, f
1
) A, f
2
) = A, f) with f(x) =
f
1
(x) +f
2
(x); A, f
1
) A, f
2
) iff f
1
(x) f
2
(x) for all x A; x A, f) iff x A and f(x) > 0.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 219
are typically written using set-like notation with brackets [ and ] (rather than and )
where elements can be repeated, e.g., [a, a, b, b, b]. Arbitrary nite multisets of formulas
will be denoted by (subscripted) upper case Greek letters , , , . We often write
, and , to denote the multiset sums and [], respectively, for the
multiset [], and a blank space for the empty multiset []. Also, for any multiset , we
dene
0
= [] and
n+1
=
n
for all n N.
An /-sequent S is then an ordered pair of nite multisets of /-formulas, written

1
, . . . ,
n

1
, . . . ,
m
called single-conclusion if m 1, and atomic if
1
, . . . ,
n
,
1
, . . . ,
m
are atoms.
Intuitively, we might understand S as
1
and . . . and
n
implies
1
or . . . or
m
.
A sequent calculus for classical logic (or Boolean algebras) in the language /
c
with
binary connectives , , and constants , is presented in Figure 3. As in GLat,
there are very simple axioms, a cut rule, and rules that introduce connectives on both
sides of the sequent arrow. Also present are structural rules that do not decompose
individual formulas but rather manipulate the structure of the sequents themselves. In
particular, the weakening rules (WL) and (WR) allow any formula to be added on the
left or right of sequents and correspond for lattices to the identities and
, while the contraction rules (CL) and (CR) allow two occurrences of a
formula to be contracted into one and correspond for lattices to the identities
and . These rules suggest reading the , on the left of sequents as and on
the right of sequents as ; that is, a sequent can be interpreted as the formula
_

_
where
_
[] = and
_
[] = .
EXAMPLE 2.3.1. Let us take a look at a derivation in GCL of Peirces law:

(ID)
,
(WR)
,
()

(ID)
( ) ,
()
( )
(CR)
(( ) )
()
EXAMPLE 2.3.2. The cut rule is closely related to the modus ponens rule used in
Hilbert systems. Indeed we can simulate (MP) in systems with (CUT) as follows:



(ID)

(ID)
,
()

(CUT)

(CUT)
Soundness with respect to classical logicwe will write [=
CL
to mean that is
classically validfollows easily from the validity of the axioms and the preservation of
validity from premises to conclusion of the rules. Completeness for GCL can be estab-
lished making use of the Lindenbaum-Tarski construction explained above for lattices,
and a more involved proof of cut elimination then implies the following result:
220 George Metcalfe
Axioms Cut rule

(ID)

1
,
1

2
,
2

1
,
2

1
,
2
(CUT)
Left logical rules Right logical rules
,
()
,
()

1
,
1

2
,
2

1
,
2
,
1
,
2
()
, ,
,
()
,
,
()
1
,
,
()
1
,
,
()
2
,
,
()
2
, ,
,
()
, ,
,
()
Left structural rules Right structural rules

,
(WL)

,
(WR)
, ,
,
(CL)
, ,
,
(CR)
Figure 3. The sequent calculus GCL
THEOREM 2.3.3.
GCL
iff [=
CL
_

_
.
However, let us sketch here a more direct semantic completeness proof for GCL

.
Consider these alternative conjunction, disjunction, and implication rules:
, ,
,
()

, ,
,
()

, ,
,
()

Each of these rules is easily derivable in GCL

using the corresponding logical rules

and the structural rules. E.g., for ()
t
we have the GCL

-derivation:
, ,
, ,
()
2
, ,
()
1
,
(CL)
Chapter III: Proof Theory for Mathematical Fuzzy Logic 221
Indeed, the converse also holds: namely, the rules ()
1
, ()
2
, ()
1
, ()
2
, and
() are derivable using ()
t
, ()
t
, and ()
t
and the structural rules. E.g., for
()
1
and ()
2
, we have the derivations:
,
, ,
(WL)
,
()

,
, ,
(WL)
,
()

Moreover, the rules ()

t
, ()
t
, and ()
t
, together with the rules (), (),
(), (), and () have a useful invertibility property. Namely, for any in-
stance

1

1
. . .
n

n

of one of these rules:
[=
CL

i [=
CL

i
for i = 1 . . . n.
This property allows us to prove
GCL
iff [=
CL
_

_
by induction
on the number of occurrences of connectives in . For the base case, we note
that a sequent containing only propositional variables is classically valid iff
there is a common variable on both sides of the sequent: is then derivable
using just weakening rules and (ID). For the inductive step, we choose an occurrence
of a connective and apply the corresponding rule and the induction hypothesis. E.g.,
suppose that =
t
[ ]. Then using the aforementioned property of ()
t
,
together with the induction hypothesis:
[=
CL
_

t
[ ]
_
i [=
CL
_

t
[, ]
_

i
GCL

t
[, ] .
But
GCL

t
[, ] implies
GCL

t
[] since ()
t
is derivable,
and
GCL

t
[ ] implies [=
CL
_

t
[ ]
_
by soundness, so
we have the desired equivalence.
2.4 Substructural logics
A remarkable feature of GCL is that it is transformed into a calculus for intuition-
istic logic by applying one simple restriction. Let the set of inferences of a sequent
rule (r) with single-conclusion premises and conclusion be called the single-conclusion
version of (r). The sequent calculus GIL for intuitionistic logic then consists of the
single-conclusion versions of the rules of GCL, noting that GIL has no right contrac-
tion rules, and right weakening is conned to premises with empty succedents (blocking,
for instance, the derivation of Peirces law in Example 2.3.1). In this sense, intuitionistic
logic provides the rst example of a substructural logic: intuitively, a logic that lacks
some of the structural rules of classical logic.
Further substructural logics are obtained by removing weakening or contraction
rules from GIL and GCL. In such cases, there also arise further natural choices for
222 George Metcalfe
connectives; in particular (inspired perhaps by the above completeness proof for GCL

)
we can introduce rules for an alternative conjunction connective &
, ,
, &
(&)

1
,
1

2
,
2

1
,
2
& ,
1
,
2
(&)
and for alternative truth constants 1 and 0

, 1
(1)
1
(1)
0
(0)

0,
(0)
In GCL and GIL it is possible to prove the sequents (& ), ( &),
(1 ), ( 1), (0 ), and ( 0), but in logics lacking structural rules, this
might no longer be the case. If we add all of these extra rules to GCL without weakening
and contraction, we obtain a calculus GMALL, and taking single-conclusion versions,
we obtain a calculus GMAILL corresponding to the Hilbert system HMAILL of Fig-
ure 1. Removing the rules for and gives a calculus for FL
e
-algebras, and removing
also the rules for 0 gives a calculus for commutative residuated lattices.
4
All these cal-
culi admit cut elimination and have proved extremely useful (indeed essential) in tack-
ling questions such as decidability, complexity, interpolation, etc. for the corresponding
logics and classes of algebras.
One of the central themes of this chapter is that proof theory reveals fuzzy logics to
be rst-class substructural logics. How is this to be understood? Roughly, the expression
substructural refers to the fact that such logics, which all live in a certain sense below
the surface of classical logic, fail to admit one or more classically valid structural rules.
Most convincingly, logics dened by sequent calculi obtained by removing weakening
or contraction rules from GCL or GIL may be deemed substructural, although even
in these cases, further logical rules may be added to capture connectives that split in
the absence of structural rules. Other structural rules, such as weaker versions of
weakening or contraction, may also be added, giving a family of logics characterized by
cut-free sequent calculi. On the other hand, there are classes of logics (e.g., most relevant
logics and, as we will see, most fuzzy logics) that almost t into this framework but
require more exible formalisms than sequents. More perplexing still, there are closely
related logics (and classes of algebras) lacking structural rules for which no reasonable
cut-free calculus is known. Whether these logics should or should not be considered
substructural is perhaps simply a matter of taste.
3 From sequents to hypersequents
We can do a lot with sequents; in particular, we can develop analytic proof systems
for a broad range of substructural and other non-classical logics. However, for fuzzy log-
ics, sequents are not quite enough. To see why, consider a potential cut-free derivation
4
Recall that a commutative residuated lattice is an algebra A = A, , , &, , 1) such that A, , )
is a lattice, A, &, 1) is a commutative monoid, and x & y z iff x y z for all x, y, z A. Pointed
commutative residuated lattices or FL
e
-algebras add an extra nullary operation 0 to the signature, while
bounded FL
e
-algebras add also nullary operations and so that A, , , , ) is a bounded lattice.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 223
of a sequent corresponding to the prelinearity axiom
(p q) (q p).
A cut-free derivation of this in GCL requires both contraction and weakening rules. If
one of these is unavailable, then we can either tinker with different sequent calculie.g.,
taking non-standard (different to those of GMALL) rules for connectives or interpreta-
tions of sequents (this approach is taken for some logics in Section 5)or extend the
sequent framework.
Take another look at (p q) (q p). To work further on this sequent, it
would be helpful to represent the premises of ()
1
and ()
2
together as
p q or q p.
We would then be able to operate on p q and q p in parallel. With this in mind, let
us dene an /-hypersequent to be a nite multiset of /-sequents, written

1

1
[ . . . [
n

n
called single-conclusion if each
i

i
for i 1, . . . , n is single-conclusion and
atomic if each
i

i
for i 1, . . . , n is atomic. We will denote arbitrary hyperse-
quents by the (subscripted) symbols ( and 1, writing ( [ 1and ( [ S (for a sequent S)
to denote the multiset unions ( 1 and ( [S], respectively.
3.1 The core set of rules
Although we now have new structures to play with, we can still use essentially the
same logical rules. Supposing that we are dealing with some xed language, let the
hypersequent version of a (single-conclusion) sequent rule (r) be the set of inferences
dened by the schema
( [ S
1
. . . ( [ S
n
( [ S
where S
1
, . . . , S
n
/ S is an instance of (r) and ( is any (single-conclusion) hyperse-
quent. The single-conclusion version of a hypersequent rule is the set of inferences of
the rule with single-conclusion premises and conclusion.
The core set of hypersequent rules displayed in Figure 4 contains hypersequent ver-
sions of the rules of the sequent calculus GMALL described in the previous section
together with special structural rules operating at the level of sequents: external weak-
ening (EW) and external contraction (EC). Essentially these extra rules characterize [
as an additive disjunction: (EW) adds sequents and (EC) removes them. They provide
greater exibility for recording choices in a derivation. In particular, the rules for on
the right and those for on the left can be combined as follows:
( [ , [ ,
( [ ,
()
( [ , [ ,
( [ ,
()
Let GL be any calculus with (EW) and (EC). Then () and () are derivable in
GL extended with ()
1
, ()
2
, ()
1
, and ()
2
, and vice versa: these rules are
224 George Metcalfe
derivable in GL extended with () and (). E.g., to show that () is derivable
using ()
1
and ()
2
, we have:
( [ , [ ,
( [ , [ ,
()
2
( [ , [ ,
()
1
( [ ,
(EC)
On the other hand, it is easily shown that adding (EW) and (EC) to the hypersequent
version of a sequent calculus has no effect on which sequents are derivable. To get
more, we will need to add rules that allow sequents to interact.
For convenience, let us call a hypersequent calculus a core rule set extension if it
contains at least the core rules displayed in Figure 4 or their single-conclusion versions.
We will consider a great variety of core rule set extensions, but let us rst investigate
some useful general properties of the logical rules. For a proof systemC, a rule is said to
be C-invertible if for each instance a
1
, . . . , a
n
/ a of the rule, whenever a is C-derivable,
also a
i
is C-derivable for each i 1, . . . , n. For example, consider an instance of the
rule (&) and any core rule set extension GL. If the conclusion (( [ , & )
is GL-derivable, then the premise is also GL-derivable as shown below:
( [
(ID)
( [
(ID)
( [ , &
(&)
( [ , &
( [ , ,
(CUT)
So (&) is GL-invertible. Indeed, this property holds also for several other logical
rules, often either for the left rules or right rules for a connective, but not both. Moreover,
if GL contains the core rule set (not just the single-conclusion versions), both the left
and right rules for and the right rule for of the following GL-derivable logical rules
are GL-invertible:
( [
1
,
1
( [
2
,
2
( [
1
,
2
,
1
,
2
()
( [ , ,
( [ ,
()
( [ ,
( [ ,
()
( [ ,
( [ ,
()
LEMMA 3.1.1. The rules (&), (), (0), (0), (1), (1), (), (),
(), (), and () are GL-invertible for any core rule set extension GL, and if
GL is not restricted to the single-conclusion versions of the core rules, then also ()
and () are GL-invertible.
Structural rules manipulate sequents and formulas with no regard to their internal
composition. If we x the logical rules of our systems, then it is these manipulations
that give each calculus its distinctive properties. Here we consider important examples
of two kinds of structural rules: internal rules that manipulate formulas within individ-
ual sequents, and external rules that manipulate whole hypersequents. We also intro-
duce along the way many different calculi for fuzzy logics, collecting these denitions
together for the readers convenience in Table 1 on page 231.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 225
Axioms Cut rule
G |
(ID)
G |
1
,
1
G |
2
,
2
G |
1
,
2

1
,
2
(CUT)
External weakening External contraction
G
G | H
(EW)
G | H | H
G | H
(EC)
Left logical rules Right logical rules
G | ,
()
G | ,
()
G |
G | , 1
(1)
G | 1
(1)
G | 0
(0)
G |
G | 0,
(0)
G |
1
,
1
G |
2
,
2
G |
1
,
2
,
1
,
2
()
G | , ,
G | ,
()
G | , ,
G | , &
(&)
G |
1
,
1
G |
2
,
2
G |
1
,
2
&,
1
,
2
(&)
G | ,
G | ,
()
1
G | ,
G | ,
()
1
G | ,
G | ,
()
2
G | ,
G | ,
()
2
G | , G | ,
G | ,
()
G | , G | ,
G | ,
()
Figure 4. The core rule set
It is often helpful (in particular, for tting derivations onto the page) to use rules
combined with applications of (EW) and (EC). We will denote such combinations with
the superscript

. For example, we might make use of a version of (CUT) where the
context side-hypersequents are added rather than merged:
(
1
[
1
,
1
(
2
[
2
,
2
(
1
[ (
2
[
1
,
2

1
,
2
(CUT)

226 George Metcalfe

The core rules, (CUT) excepted, have the subformula property: the only formulas
occurring in the premises are subformulas of formulas occurring in the conclusion. This
ensures that each derivation using only these rules can be viewed (reading upwards) as a
decomposition of formulas in which no new material is added. In fact we can decompose
to such an extent that (ID) can be restricted to instances containing only variables.
LEMMA 3.1.2. Let GL consist of the logical rules for a sublanguage / of /
p
and
instances of (ID) of the form (p p) where p is any variable. Then
GL
for all
Fm
1
.
Proof. We proceed by induction on cp(). If is a variable p, then ( ) is a suitable
instance of (ID). If is a constant, then the cases of and follow immediately using
() and (), respectively, and if is 0 or 1, then we have GL-derivations:
1
(1)
1 1
(1)
0
(0)
0 0
(0)
Now suppose that is for , &, , . By the induction hypothesis twice,

GL
and
GL
, and we can construct derivations using the left and right
rules for . E.g., for and , we have:

,
()

()


()
1


()
2

()
3.2 Adding structural rules
Just as extending GMALL or the full Lambek calculus with structural rules denes
sequent calculi for other substructural logics, so adding hypersequent versions of these
same rules to the core rule set can give calculi for a range of fuzzy (and other) logics. In-
deed, now we have a wider choice since we can also add structural rules that manipulate
more than one sequent.
Communication
For Hilbert systems, it is the prelinearity and distributivity axioms that are key for char-
acterizing linearity. For Gentzen systems, it is the following communication rule, so-
called because formulas are communicated between different sequents:
( [
1
,
1

1
,
1
( [
2
,
2

2
,
2
( [
1
,
2

1
,
2
[
1
,
2

1
,
2
(COM)
EXAMPLE 3.2.1. The best way to understand communication is by returning to the
tricky prelinearity axioms ( ) ( ). The following derivation uses (COM)
and (EC) (implicit in the derived rule ()) but no other structural rules:
Chapter III: Proof Theory for Mathematical Fuzzy Logic 227

(ID)

(ID)
[
(COM)
[
()
[
()
( ) ( )
()
Notice that the hypersequent ( [ ) two lines down might be read as just a
hypersequent translation of the formula ( ) ( ).
EXAMPLE 3.2.2. Consider this derivation of a helpful property of disjunction:
[
(ID)

(ID)

(ID)
[
(COM)
[
(ID)
[
()
[
()
We can also use the communication rule to derive distributivity axioms, where the top
right hypersequent is derived as above:

(ID)
( )
()
1

(ID)
( )
()
1
|
| ( )
()
2
| ( )
()

( ) | ( )
()
2
( ) | ( )
()

( ) ( ) ( )
()
( ( )) (( ) ( ))
()
The addition of (COM) gains us a couple of helpful extra invertibilities to add to
those in Lemma 3.1.1. Consider the dened rule (). In any core rule set extension,
we can derive the premise of an instance of this rule from its conclusion as follows
[ ( [ ,
( [ [ ,
(CUT)

( [ ,
( [ , [ ,
(CUT)

where the leftmost hypersequent is derived as in Example 3.2.2.

LEMMA 3.2.3. () and () are GL-invertible for any core rule set extension GL
with (COM).
A close relative of communication, the so-called split rule, divides one sequent in
the premise into two sequents in the conclusion.
( [
1
,
2

1
,
2
( [
1

1
[
2

2
(SPLIT)
228 George Metcalfe
EXAMPLE 3.2.4. Just as the communication rule corresponds to the prelinearity and
distributivity axioms, so the split rule corresponds to the law of excluded middle:

(ID)
[
(SPLIT)
[
()

()
We can now identify Gentzen systems for some elementary fuzzy logics, letting
GIUL (for involutive uninorm logic) consist of the core set of rules together with (COM),
and GUL (for uninorm logic) consist of the single-conclusion versions of these rules.
Weakening
Weakening axioms (or integrality at the algebraic level) are a common feature of fuzzy
logics based on t-norms. In Gentzen systems, weakening rules typically come in pairs,
introducing new formulas on the left and right of sequents. Here, for simplicity, we
combine these left and right rules into one:
( [
( [ , ,
(W)
The systemGIMTL (for involutive monoidal t-norm logic) is GIUL extended with (W),
and GMTL for (monoidal t-norm logic) is GUL extended with the single-conclusion
version of (W).
EXAMPLE 3.2.5. (W) allows us to prove that 0 implies every formula and that 1 is
implied by every formula:
1
(ID)
1
(W)
1
()
0
(0)
0
(W)
0
()
A weaker version of the weakening rule mixes two sequents into one; the empty
sequent can also be viewed as a nullary mix of no formulas:
( [
1

1
( [
2

2
( [
1
,
2

1
,
2
(MIX)
( [
(EMP)
EXAMPLE 3.2.6. (MIX) and (EMP) connect the constants 1 and 0. The former allows
us to prove (not unreasonably) that 0 is less true than 1; the latter (more bizarrely)
allows us to prove the converse:

(EMP)
1
(1)
1 0
(0)
1 0
()
0
(0)
1
(1)
0 1
(MIX)
0 1
()
Moreover, (COM) is derivable using (SPLIT) and (MIX) (and so is redundant in Gentzen
systems with these rules), while (SPLIT) is derivable using (COM) and (EMP).
Chapter III: Proof Theory for Mathematical Fuzzy Logic 229
Contraction
The other (with weakening) core structural rules for Gentzen systems are the so-called
contraction rules, which reduce the number of occurrences of a formula in a sequent.
Typically, we encounter rules that contract a single formula on the left or right. However,
as for weakening, we will consider here the more general version:
( [ , , , ,
( [ , ,
(C)
EXAMPLE 3.2.7. (C) helps us to derive standard contraction axioms as follows:

(ID)

(ID)
, &
(&)
&
(C)
( & )
()
Adding the single-conclusion version of (C) to GMTL gives the hypersequent cal-
culus GG (for G odel logic). (Or, consider the hypersequent version of a calculus for
intuitionistic logic plus (EW), (EC), and (COM).) Adding (C) to the core rule set plus
weakening (W) gives a calculus for classical logic.
The situation for calculi without weakening is more complicated. Adding (C) and
(MIX) to the core rule set gives a calculus for the non-distributive R-mingle logic RM
ND
.
A calculus GRM (which does prove distributivity) for R-mingle is dened at the hy-
persequent level by adding (C) and (MIX) to GIUL. Adding also (EMP) then gives a
calculus GIUML (for involutive uninorm mingle logic).
EXAMPLE 3.2.8. Sometimes structural rules interact in unexpected ways; e.g., in the
presence of (W) and (SPLIT), we can derive the contraction rule (C):
( [ , , , ,
( [ [ , ,
(SPLIT)
( [ , , [ , ,
(W)
( [ , ,
(EC)
Indeed, GMTL extended with the single-conclusion version of (SPLIT) is a single-
conclusion hypersequent calculus for classical logic.
A more complicated rule, which contracts formulas selectively, is the mingle rule
where (as the name suggests) elements from two sequents are combined into one:
( [
1
, ,
1
( [
2
, ,
2
( [
1
,
2
, ,
1
,
2
(MGL)
The calculus GUML (for uninorm mingle logic) is GUL extended with the single-
conclusion versions of (MGL) and (C).
230 George Metcalfe
EXAMPLE 3.2.9. Mingle axioms are derivable in calculi with (MGL) as follows:

(ID)

(ID)
,
(MGL)
&
(&)
( & )
()
The contraction rule (C) can also be generalized as follows:
( [ ,
n
1

n
1
, . . . ( [ ,
n
n1

n
n1
,
( [ ,
1
, . . . ,
n1

1
, . . . ,
n1
,
(C
n
) n = 2, 3, . . .
Adding (C
n
) (n 2) to GIMTL and its single-conclusion version to GMTL gives
calculi GIMTL
n
and GMTL
n
, respectively, for involutive n-contractive monoidal t-
norm logic and n-contractive monoidal t-norm logic.
Other forms of contraction include the contraction of a whole sequent; in particular,
adding the following single-conclusion rule to GMTL gives a calculus GSMTL for
strict monoidal t-norm logic:
( [ ,
( [
(SC
2
)
EXAMPLE 3.2.10. Using (SC
2
), we can derive non-contradiction axioms:

(ID)
,
()
,
()
2
,
()
1

(SC
2
)
( )
()
Other forms of contraction may get rather complicated. For example, a calculus
GWNM for weak nilpotent minimum logic is obtained by adding the following single-
conclusion rule to GMTL:
( [
1
,
2
,
2

( [
1
,
1
,
2

( [
1
,
2
,
2

( [
1
,
1
,
2

( [
1
,
2
[
1
,
2

(WN)
On the other hand, a calculus for nilpotent minimum logic GNM requires only the ad-
dition to GIMTL of the rule:
( [ , , , , ( [ , , , ,
( [ , ,
(N)
3.3 Soundness and completeness
So far we have introduced many systems (collected in Table 1), but other than their
names, there has been little to indicate that these calculi correspond to fuzzy logics. Here
Chapter III: Proof Theory for Mathematical Fuzzy Logic 231
Calculus Rules Single-Conclusion
GUL core rules + (COM) yes
GIUL core rules + (COM) no
GMTL GUL + (W) yes
GIMTL GIUL + (W) no
GMTL
n
GMTL + (C
n
) yes
GIMTL
n
GIMTL + (C
n
) no
GSMTL GMTL + (SC
2
) yes
GWNM GMTL + (WN) yes
GNM GIMTL + (N) no
GG GMTL + (C) yes
GUML GUL + (C) + (MGL) yes
GRM GIUL + (C) + (MIX) no
GIUML GRM + (EMP) no
Table 1. Some hypersequent calculi
we show that this is indeed the case, preferring for reasons of familiarity to connect rules
and hypersequent calculi with axioms and Hilbert systems rather than with equations and
classes of algebras, noting, however, that the relationship between Hilbert systems and
corresponding classes of algebras is established in other chapters of the handbook.
Our rst task will therefore be to dene a standard translation of sequents and
hypersequents into formulas:
I( ) =
def
&
I(S
1
[ . . . [ S
n
) =
def
I(S
1
) . . . I(S
n
)
where [
1
, . . . ,
n
] =
def
(
1
. . .
n
) for &, , &[] =
def
1, and [] =
def
0.
EXAMPLE 3.3.1. Consider a hypersequent
( = (, [ , [ , , ).
To nd the standard interpretation of (, we rst interpret the sequents
I(, ) = ( & ( ))
I( , ) = 1 ( )
I(, , ) = ( & & ) 0
and then take the disjunction, to get
I(() = (( & ( )) ) (1 ( )) (( & & ) 0).
Showing that a Gentzen system GL is sound and complete with respect to a Hilbert
system HL consists of showing that
GL
( iff
HL
I((). Both directions involve rela-
tively straightforward inductions on the height of derivations. Let us take as our starting
232 George Metcalfe
point the hypersequent calculus GUL and the Hilbert system HUL obtained by extend-
ing HMAILL (see Figure 1) with the prelinearity and distributivity axioms
(PRL) ( ) ( )
(DIS) ( ( )) (( ) ( )) .
To show (inductively) that
GUL
( implies
HUL
I((), we have the rather onerous
task of establishing for each instance (
1
. . . (
n
/( of a rule of GUL, that
HUL
I((
i
)
for i = 1 . . . n implies
HUL
I((). Let us just consider one of the harder cases: a
single-conclusion instance of the communication rule. Suppose that

HUL
I(( [
1
,
1

1
) and
HUL
I(( [
2
,
2

2
).
We want to show

HUL
I(( [
1
,
2

1
[
2
,
1

2
).
For convenience, let us dene

1
= I(
1

1
),
2
= I(
2

2
),
1
= &
1
,
2
= &
2
, = I(().
Suppose then that

HUL
(
1

1
) and
HUL
(
2

2
) .
Making use of the following instances of the (B) axiom schema
(
1

1
) ((
1

2
) (
1

2
))
and (
2

2
) ((
2

1
) (
2

1
))
we obtain (using also various HMAILL-derivable formulas)

HUL
((
1

2
) (
1

2
)) and
HUL
((
2

1
) (
2

1
)).
So using the adjunction rule (ADJ) and the distributivity axiom schema (DIS)

HUL
(((
1

2
) (
1

2
)) ((
2

1
) (
2

1
))) .
Making use of the axioms of HMAILL for and , we obtain

HUL
(((
1

2
) (
2

1
)) ((
1

2
) (
2

1
)))
and, since (
1

2
) (
2

1
) is an instance of (PRL), nally

HUL
((
2

1
) (
1

2
)) .
To show that
HUL
I(() implies
GUL
(, we observe rst that for each axiom of
HUL, the sequent ( ) is GUL-derivable. In particular, the cases of (PRL) and (DIS)
are established in Examples 3.2.1 and 3.2.2, respectively. Moreover, GUL-derivability
Chapter III: Proof Theory for Mathematical Fuzzy Logic 233
is preserved by (MP) (see Example 2.3.2) and (ADJ) (using the () rule). Hence an
induction on the height of a derivation in HUL establishes the intermediate result that

HUL
I(() implies
GUL
I((). Suppose then that

GUL
(&
1

1
) . . . (&
n

n
).
By the GUL-invertibility of () and () (Lemmas 3.1.1 and 3.2.3), we obtain

GUL
&
1

1
[ . . . [ &
n

n
.
But then using the GUL-invertibility of (&), (1), and (0) (Lemma 3.1.1)

GUL

1

1
[ . . . [
n

n
.
So
HUL
I(() implies
GUL
( as required.
A very similar proof establishes that the calculus GIUL (the core rule set plus
(COM)) is sound and complete with respect to the Hilbert system HIUL obtained by
adding to HUL the involution axioms
(INV) .
Moreover, from an analysis of the above reasoning, we obtain sufcient conditions for
an extension of GUL or GIUL to be sound and complete with respect to an extension
of HUL or HIUL. Let us say that a hypersequent rule (r) and axiom schema / are
L-matching if
1. For every instance (
1
, . . . , (
n
/ ( of (r):

HL+,
I((
i
) for i = 1 . . . n implies
HL+,
I(().
2.
GL+(r)
for every axiom of /.
Examples of matching rules and axiom schema are displayed in Table 2.
THEOREM 3.3.2. Let L be UL or IUL and suppose that (r
i
) and /
i
are L-matching
for i = 1 . . . n. Then

GL+(r
1
)+...+(r
n
)
( i
HL+,
1
+...+,
n
I(().
4 Eliminations and applications
The hypersequent calculi dened in the previous section provide a uniform and nat-
ural presentation of a wide range of fuzzy logics. But are they useful? Proof search in
Hilbert systems is hindered by the need to guess formulas and as premises
when applying modus ponens. However, the same situation seems to occur for Gentzen
systems: we have to guess which formula to use when applying (CUT). If we could do
without the cut rule, then we could just apply rules where formulas in the premises are
subformulas of formulas in the conclusion. This subformula property (or analyticity
234 George Metcalfe
Structural Rule Matching Axiom Schema
[
[ , ,
(W) ( 1) (0 )
[
1

1
[
2

2
[
1
,
2

1
,
2
(MIX) 0 1
[
1
,
2

1
,
2
[
1

1
[
2

2
(SPLIT)
[
(EMP)
1 0
[ , , , ,
[ , ,
(C) ( &)
[
1
, ,
1
[
2
, ,
2
[
1
,
2
, ,
1
,
2
(MGL) ( &)
[ ,
n
1

n
1
, . . . [ ,
n
n1

n
n1
,
[ ,
1
, . . . ,
n1

1
, . . . ,
n1
,
(C
n
)
n1

n
[
1
,
2
,
2

[
1
,
1
,
2

[
1
,
2
,
2

[
1
,
1
,
2

[
1
,
2
[
1
,
2

(WN)
( &) (( ) ( &))
[ , , , , [ , , , ,
[ , ,
(N)
((
2
) (()
2
))
( )
[ , ,
[
(SC
2
) ( )
Table 2. Matching rules and axioms
of the cut-free calculus) can be useful for, among other things, establishing decidabil-
ity, complexity, interpolation, and conservative extension results. Here we show that
for calculi satisfying certain properties we can algorithmically transform derivations in
the calculus with (CUT) into derivations without this rule; that is, we eliminate (CUT)
from derivations. We also provide similar elimination results for the so-called density
rule. Since the hypersequent calculi extended with the density rule are complete with
respect to dense linearly ordered algebras, density elimination implies the same comple-
teness result for the original calculi and provides the key step for standard completeness
Chapter III: Proof Theory for Mathematical Fuzzy Logic 235
proofs for the corresponding fuzzy logics.
236 George Metcalfe
4.1 Cut elimination
Let us start with an example. Suppose that we want to eliminate an application of
(CUT) froman otherwise cut-free derivation in the single-conclusion calculus GMAILL:
.
.
.
,
.
.
.

,
(CUT)
The cut formula occurs on the left in one premise, and on the right in the other.
A natural strategy for eliminating this application of (CUT) is to look more carefully at
the derivations of these premises. If one of the premises is an instance of (ID), then it
must be ( ) and the other premise must be (, ). Otherwise, we have
two possibilities. The rst is that one of the premises ends with an application of a rule
where is not decomposed, e.g., letting =
1

2
[ ]:
.
.
.

1
, ,
.
.
.

1
,
2
, ,
()
.
.
.

1
,
2
, ,
(CUT)
In this case, we can push the cut upwards in the derivation to get:
.
.
.

1
, ,
.
.
.

1
, ,
(CUT)
.
.
.

1
,
2
, ,
()
That is, we have a derivation where the left premise in the new application of (CUT) has
a shorter derivation than the application in the original derivation.
The second possibility is that is decomposed in the last application of a rule in
both premises, e.g., with =
1

2
and = :
.
.
.

1

.
.
.

2
,

1
,
2
,
()
.
.
.
,

()

1
,
2
,
(CUT)
In this case we rearrange our derivation in a different way: we replace the application of
(CUT) with applications of (CUT) with cut formulas and :
.
.
.

2
,
.
.
.
,

2
, ,
(CUT)
.
.
.

1
,
2
,
(CUT)
We now have two applications of (CUT) but with cut formulas of a smaller complexity
than the original application.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 237
This procedure, formalized using a double induction on cut formula complexity and
the combined height of derivations of the premises, eliminates applications of (CUT)
for many sequent calculi. However, it encounters a problem with rules that contract
formulas in one or more of the premises. Consider the following situation:
.
.
.
, ,
,
(C)
.
.
.

,
(CUT)
In this case we need to perform several cuts at once, using a rule something like:
, []
n

,
n

For hypersequent calculi, the situation is further complicated by the fact that whole
sequents may be contracted using (EC). This means that a cut formula occurring in the
premises of an application of (CUT) may appear in several sequents in a hypersequent
higher up in the derivation, e.g.
.
.
.
, [ ,
,
(EC)
.
.
.

,
(CUT)
To cope with this situation, we should eliminate even more general versions of (CUT)
that perform multiple cuts in different sequents. This method of cut elimination applies
to a broad class of hypersequent calculi satisfying a substitutivity condition that allows
cuts to be pushed upwards in derivations.
To explain the general method, we make use of some denitions and notational
conveniences. First, let us assume that , , m, n, i, j, k always denote natural numbers
and , , , multisets of formulas, recalling that
0
= [] and
n+1
=
n
. S
denotes a sequent and (, 1 hypersequents, and we let [(
i
]
n
i=1
denote the hypersequent
(
1
[ . . . [ (
n
and (
i

n
i=1
denote the set of hypersequents (
1
, . . . , (
n
.
A marked hypersequent is a hypersequent with exactly one occurrence of a formula
distinguished, written (( [ , ) or (( [ , ). For a hypersequent ( and
a marked hypersequent 1, we dene the set CUT((, 1) of results of applying (CUT)
multiple times as follows:
(1) If does not occur in

n
i=1

i
where
( = [
i
, []

i

i
]
n
i=1
and 1 = (1
t
[ , ),
then the set CUT((, 1) contains, for 0
i

i
(i = 1 . . . n),
1
t
[ [
i
,

i
, []

i
,
i
]
n
i=1
.
238 George Metcalfe
(2) If does not occur in

n
i=1

i
where
( = [
i
[]

i
,
i
]
n
i=1
and 1 = (1
t
[ , ),
then the set CUT((, 1) contains for 0
i

i
(i = 1 . . . n),
1
t
[ [
i
,

i
[]

i
,

i
,
i
]
n
i=1
.
One of the crucial steps for our cut elimination method will be shifting (multiple)
applications of (CUT) upwards over applications of other rules. We dene a (structural)
rule (r) to be substitutive if for any
1. instance (
1
, . . . , (
n
/ ( of (r)
2. marked hypersequent 1 (single-conclusion if (r) is single-conclusion)
3. (
t
CUT((, 1)
there exist (
t
i
CUT((
i
, 1) for i = 1 . . . n such that
(
t
1
. . . (
t
n
(
t
is an instance of (r).
The name substitutive is apt because the condition implies that substituting occur-
rences of with on the left and on the right, in both the conclusion of a rule
instance and suitably in its premises, gives another instance of the rule. In particular,
it is easy to see that the structural rules introduced in Table 2 are substitutive, since (1)
each multiset variable , , , , . . . and hypersequent variable ( occurs at most once
in the conclusion, (2) every multiset or hypersequent variable occurring in a premise
occurs in the conclusion, and (3) multiset variables are paired so that, e.g., in the single-
conclusion case, any multiset variable occurring on the right in the conclusion occurs
always with another multiset variable on the left. However, for example, the following
anti-contraction rule is not substitutive:
( [ , ,
( [ , , , ,
Just consider an instance of the form (p / p, p ) and a marked hypersequent
(q p). Neither (p / p, q ) nor (q / p, q ) is an instance of the rule.
Note, moreover, that each core logical rule (r) is not substitutive but rather almost
substitutive in the sense that the rule obtained by removing the principal formula from
the conclusion of instances and its decomposed parts from their premises, is substitutive.
Before embarking on our general cut elimination proof, let us review some of the
main ideas. First, note that it is enough just to consider an uppermost application of
(CUT) in a derivation. If we can remove such an application without introducing any
new applications, then we can eliminate applications of (CUT) one by one. So, recall-
ing that GL

is the calculus GL without (CUT), to establish cut elimination for GL

it is enough to show constructively that if the premises of an instance of (CUT) are
GL

-derivable, then the conclusion is GL

-derivable. Following our earlier discussion,

we show more generally that if two hypersequents ( and 1 are derivable, the rst with
a marked occurrence of , then any hypersequent in CUT((, 1) is derivable.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 239
THEOREM 4.1.1. Cut elimination holds for any extension of GUL with substitutive
single-conclusion rules or extension of GIUL with substitutive rules.
Proof. Let GL be an extension of GUL with substitutive single-conclusion rules or an
extension of GIUL with substitutive rules. As observed above, it sufces to show that
an uppermost application of (CUT) in any GL-derivation (i.e., where the premises
are GL

-derivable) can be eliminated without introducing new applications of (CUT).

Hence it is enough to prove that for any hypersequent ( and marked hypersequent 1
with marked formula :
If d

GL
( and d
)

GL
1, then
GL
(
t
for all (
t
CUT((, 1).
We prove the claim by a triple induction on the lexicographically ordered triple
cp(), e(d
)
), h(d

)
where e(d) =
_
0 if d ends with a logical rule applied to a marked principal formula
1 otherwise.
We begin by considering the last application of a rule (r) in d

. If (r) is (ID), then (

is of the form ((
t
[ ). So every member of CUT((, 1) is of the form (1
t
[ ()
or (1
t
[ 1), and the claim follows using (EW). Otherwise, there are two cases:
(a) The application of (r) is of the form:
(
1
. . . (
n
(
and the principal formula (if there is one) is not an occurrence of on the opposite side
to the marked occurrence in 1. Pick (
t
CUT((, 1). By the substitutivity of (r) or
almost-substitutivity if (r) is a logical rule, there exist (
t
i
CUT((
i
, 1) for i = 1 . . . n
such that
(
t
1
. . . (
t
n
(
t
is an instance of (r).
But, by the induction hypothesis,
GL
(
t
i
for i = 1 . . . n, so also
GL
(
t
as required.
(b) The application of (r) is of the form:
(
1
. . . (
n
(
t
[ , []

or
(
1
. . . (
n
(
t
[ []

,
where the principal formula is an occurrence of on the opposite side to the marked
occurrence in 1, and , or , , respectively. Pick (
)
CUT((, 1) where 1
is of the form, respectively,
1
t
[ , or 1
t
[ , .
The only tricky case (others follow as in case (a) using almost-substitutivity) is when
(
)
is of the form
1
t
[ (
tt
[ ,

,
240 George Metcalfe
with (
tt
CUT((
t
, 1). Then we also have a member of CUT((, 1) of the form,
respectively,
1
t
[ (
tt
[ ,
1
,
1
, or 1
t
[ (
tt
[ ,
1
,
1
, .
Since (r) is almost-substitutive, there exist (
t
i
CUT((
i
, 1) for i = 1 . . . n so that we
have an instance of (r) of the form, respectively,
(
t
1
. . . (
t
n
1
t
[ (
tt
[ ,
1
,
1
,
or
(
t
1
. . . (
t
n
1
t
[ (
tt
[ ,
1
,
1
,
Moreover, by the induction hypothesis,
GL
(
t
i
for i = 1 . . . n, so we have a derivation
d ending with such an application of (r).
Now we consider two subcases:
1. e(d
)
) = 1: i.e., d
)
does not end with the application of a logical rule to the
marked occurrence of . Mark the remaining occurrence of on the left or right
as appropriate in d and remove the marking of in d
)
. So e(d) = 0 and
cp(), e(d), h(d
)
) < cp(), e(d
)
), h(d

).
Hence by the induction hypothesis and a further application of (EC),
GL
(
)
.
2. e(d
)
) = 0: i.e., d
)
ends with the application of a logical rule to the marked
occurrence of , and is of the form, respectively,
1
1
. . . 1
m
1
t
[ ,
or
1
1
. . . 1
m
1
t
[ ,
Then (
)
= (1
t
[ (
tt
[ ,

, ) is derivable from (
t
1
, . . . , (
t
n
, and
1
1
, . . . , 1
m
by cuts on subformulas
1
, . . . ,
k
of where
cp(
i
), e(d
)
), h(d) < cp(), e(d
)
), h(d

) for i = 1 . . . k.
It follows, using (EW) and the induction hypothesis (several times perhaps), that

GL
(
)
. For example, suppose that GL is a single-conclusion calculus, =
, and d ends with
1
t
[ (
tt
[
1
,

1
t
[ (
tt
[
2
,
1
,
1
t
[ (
tt
[
1
,
2
,
1
,
Then d
)
ends with
1
t
[ ,
1
t
[
and we can apply (EW) and the induction hypothesis once to obtain

GL
1
t
[ (
tt
[
1
,
+1

and again to obtain

GL
1
t
[ (
tt
[
1
,
2
,

.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 241
One immediate consequence of this cut elimination result is that cut-free deriva-
tions in the considered calculi have the subformula property: any formula occurring in
such a derivation must occur as a subformula in the derived hypersequent. The subfor-
mula property is crucial not only for developing automated reasoning methods based on
Gentzen systems, but also for more theoretical applications. Notice for example that the
empty sequent (which leads to inconsistency in logics with weakening) can, by the
subformula property, only be derivable if it is an axiom of the calculus. More generally,
a hypersequent ( is derivable iff it is derivable when the logical rules are restricted to
those for connectives occurring in (.
Let us make this last claim more precise. A system C
1
for a set of structures
1
is a conservative extension of a system C
2
for
2

1
if for all a
2
:
C
1
a iff

C
2
a. In this case, C
2
is often called the
2
-fragment of C
1
. By cut elimination, any
extension GL of GUL with substitutive single-conclusion rules or GIUL with substitu-
tive rules is a conservative extension of the calculus GL with a restricted set of logical
rules. However, this is cheating slightly. Although rules for may not be available
in the restricted calculus, we still have plenty of rules for the external disjunction [:
namely, (EW), (EC), and (COM). It is therefore more interesting to ask if we can obtain
conservative extension results for Hilbert systems, where only formulas are involved in
derivations. Or, put another way, can we nd axiomatizations for fragments of Hilbert
systems for fuzzy logics? Let us consider a pertinent example: the implicational frag-
ment of monoidal t-norm logic MTL.
The Hilbert system BCK consists of the modus ponens rule (MP) and axiom schema:
(B) ( ) (( ) ( )) (transitivity)
(C) ( ( )) ( ( )) (permutation)
(K) ( ) (weakening).
Let HMTL

be BCK extended with the axiom schema

(IPRL) (( ) ) ((( ) ) )
and let GMTL

be GMTL with the logical rules restricted to () and ().

THEOREM 4.1.2. HMTL is a conservative extension of HMTL

.
Proof. Suppose that
HMTL
for some implicational formula . Then making use of
Theorems 3.3.2 and 4.1.1,
GMTL
. So it remains to show that
GMTL

implies
HMTL
. We dene an interpretation of single-conclusion hypersequents,
parameterized by the variable q, that avoids any mention of connectives other than :
I
q
(
1
, . . . ,
n
) =
def

1
. . .
n

I
q
(
1
, . . . ,
n
) =
def

1
. . .
n
q
I
q
(S
1
[ . . . [ S
n
) =
def
(I
q
(S
1
) q) . . . (I
q
(S
n
) q) q
where
1
. . .
n
is short for
1
(
2
(. . . (
n
) . . .). Observe
now that it sufces to prove the following:
If
GMTL
(, then
HMTL
I
q
(() for any variable q not occurring in (.
242 George Metcalfe
If
GMTL
, then, by the above claim,
HMTL
( q) q for some q not
occurring in . But then substituting for q in the HMTL

-derivation, we obtain

HMTL
( ) and hence also
HMTL
.
We prove the claim by induction on the height of a GMTL

-derivation of (. For
the base case, ( is of the form (
t
[ and we simply note that
HMTL

((( ) q) q) for any formula . The inductive step involves a number of
tedious Hilbert derivations. For rules of the form (( [ S
1
) . . . (( [ S
n
) / (( [ S), it is
sufcient to show that:

HMTL
I
q
(S
1
) . . . I
q
(S
n
) I
q
(S).
For example, for (), letting
1
= [
1
, . . . ,
k
] and = I
q
(
2
), we have a
HMTL

-derivation of
(
1
. . .
k

1
) (
2
) (
1
. . .
k
(
1

2
) ).
For cases not of this form, we proceed a little differently. E.g., for (EC), suppose that

GMTL
( [ S [ S. Then by the induction hypothesis, for some suitable
1
, . . . ,
n
:

HMTL
(
1
q) . . . (
n
q) (I
q
(S) q) (I
q
(S) q) q.
Substituting (
1
q) . . . (
n
q) (I
q
(S) q) q for q in the above
derivation and simplifying, we obtain as required

HMTL
(
1
q) . . . (
n
q) (I
q
(S) q) q.
Similar results can be obtained for Hilbert systems for other logics with weakening
such as G odel logic and involutive monoidal t-norm logic. However, in the case of
weakening-free logics, we can no longer simulate the connective using or the other
multiplicative connectives. A particularly interesting open question is to provide an
axiomatization (if one exists) of the implicational fragment of HUL. (Could this perhaps
be just the implicational fragment of HMAILL?)
For sequent calculi, cut elimination is often a key tool for establishing the decidabil-
ity of derivability of sequents in the calculus and hence also of derivability of formulas in
the corresponding Hilbert system (or indeed, of the equational theory of the correspond-
ing class of algebras). As an easy example, notice that proof search in GMAILL

, pro-
ceeding by applying rules backwards, must terminate since the sum of the complexities
of formulas in sequents decreases from conclusion to premises in any instance of a rule
of this calculus. Hence GMAILL-derivability (also HMAILL-derivability of formulas
and the equational theory of FL
e
-algebras) is decidable. The same argument works for
GMAILL with weakening and related calculi, but fails in the presence of contraction
rules where premises can have greater complexity than the conclusion. Sometimes, e.g.,
for fragments of the relevance logic R, this can be dealt with by using restricted rules
and some kind of loop-checking mechanism. For hypersequent calculi, however, the
issue is further complicated by the presence of the external contraction rule (EC) which
duplicates whole sequents. Nevertheless, for calculi with both internal and external con-
traction rules, such as GG, GIUML, and GUML, we can again obtain terminating proof
search fairly easily.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 243
Let us dene a 3-sequent to be any sequent such that each formula occurs
no more than three times in and no more than three times in , and a 3-hypersequent
to be a hypersequent consisting only of 3-sequents and containing no more than three
copies of the same sequent. Now consider the set of formulas F occurring in some 3-
hypersequent. There is only a nite number of 3-sequents containing subformulas of the
formulas in F, and hence also only a nite number of 3-hypersequents built from such
3-sequents. It is straightforward, using cut elimination, to prove that if GL is GUML,
GIUML, or GG, then a 3-hypersequent is GL-derivable iff it has a GL-derivation in
which only 3-hypersequents occur. Decidability of derivability of hypersequents in these
calculi then follows by restricting to derivations in which only 3-hypersequents occur
and using loop-checking (i.e., checking that the premises of a rule instance have not
already occurred lower in the derivation).
THEOREM 4.1.3. Derivability in GUML, GIUML, and GG is decidable.
Note that the logics MTL, IMTL, and SMTL (and hence also the derivability of hyper-
sequents in GMTL, GIMTL, and GSMTL) have been proved decidable using algebraic
methods, while the decidability of the logics UL and IUL (and their corresponding cal-
culi) is still open.
4.2 Density elimination
We turn our attention now to an interesting relative of the cut rule, the hypersequent
version of a density rule introduced by Takeuti and Titani to axiomatize rst-order
G odel logic:
( [
1
p,
1
[
2
, p
2
( [
1
,
2

1
,
2
(DENSITY)
where p does not occur in (,
1
,
2
,
1
, or
2
.
Note that (DENSITY) does not have the subformula property and cannot be presented
schematically without the additional condition on the variable p. Roughly speaking, the
rule expresses the density of the set of truth values for the logic. Consider an instance
( p [ p ) / . Understanding [ as a classical disjunction and as , we
can read such an instance contrapositively as if > , then > p > for some p.
For a hypersequent calculus GL, let GL
D
be GL extended with (DENSITY). Our
goal in this section will be to show that for certain (families of) hypersequent calculi
GL, the extended calculus GL
D
admits density elimination; in other words, we can add
(DENSITY), but we dont really need it.
EXAMPLE 4.2.1. Adding (DENSITY) can have a dramatic effect. If GL is a calcu-
lus with (ID), (COM), and (C) (without single-conclusion restrictions), then the empty
sequent is derivable in GL
D
as follows:
p p
(ID)
p p
(ID)
p, p [ p, p
(COM)
p, p [ p
(C)
p [ p
(C)

(DENSITY)
244 George Metcalfe
In particular, adding (DENSITY) to a hypersequent calculus for classical logic gives
inconsistency. Moreover, (DENSITY) cannot be eliminated from the calculus GRM
D
since the empty sequent is derivable in GRM
D
but not in GRM.
Our density elimination method will proceedlike cut eliminationby removing
applications of the rule which are uppermost in a derivation. Suppose that we have a
derivation d ending with
.
.
.
p [ , p
,
(DENSITY)
The idea of our proof is to replace occurrences of p in d in an asymmetric way: with
if p occurs on the left, and with on the left and on the right, if p occurs on the right.
What we get is not quite a derivation, but still a nite tree labelled with hypersequents,
now ending with
.
.
.
, [ ,
,
The last step in this not-quite-a-derivation is an application of (EC). The applications of
logical rules and most structural rules appearing in the original derivation are preserved
by the replacement due to substitutivity or almost-substitutivity. Where the derivation
potentially breaks down is in rules like (COM) where ps can occur in premises on both
the left and the right. For example, suppose that d ends with
p p
(ID)
.
.
.

t
,

t
p [ , p
(COM)
.
.
.
p [ , p
,
(DENSITY)
If we replace ps as suggested, we get
,
.
.
.

t
,

t
, [ ,
(COM)
.
.
.
, [ ,
,
(EC)
But now we have a missing part of the derivation: the sub-derivation of (, ),
which was what we wanted to prove in the rst place. However, notice that in this case,
Chapter III: Proof Theory for Mathematical Fuzzy Logic 245
we can simply replace the application of (COM) with an application of (EW) and remove
the occurrence of (, ) as a premise. More generally, we are able to use (CUT)
and cut elimination to repair such derivations.
Let us assume for now that we deal only with single-conclusion hypersequents.
Moreover, we consider hypersequents where the variable p occurs only in a limited
fashion: only as an atom and not on both the left and right in the same sequent. More
precisely, a hypersequent ( is called p-regular if it is of the form
[
i
p]
n
i=1
[ [
j
, [p]

j

j
]
m
j=1
where p does not occur in
1
, . . . ,
n
,
1
, . . . ,
m
,
1
, . . . ,
m
.
We also need a way of distinguishing the occurrences of the variable p introduced by
the density rule. A double-p-marked hypersequent has just two occurrences of p, both
marked: one on the left in a sequent and one on the right in another sequent, written
1 [ p [ , p .
The idea is to combine a p-regular hypersequent ( with a double-p-marked hypersequent
(1 [ p [ , p ), essentially by applying (CUT) exhaustively to ( with
(1 [ p) and (1 [ , p ) with cut formula p. Suppose that
1. ( = [
i
p]
n
i=1
[ [
j
, [p]

j

j
]
m
j=1
is a p-regular hypersequent
2. 1
p
= (1 [ p [ , p ) is a double-p-marked hypersequent.
Then DEN((, 1
p
) = (1 [ [
i
, ]
n
i=1
[ [
j
,

j

j
]
m
j=1
).
EXAMPLE 4.2.2. Consider the p-regular and double-p-marked hypersequents
( = (q r p [ q p [ r q, p, p r) and 1
p
= (s p [ s q, p r).
Then DEN((, 1
p
) = (q r, s q r [ q, s q r [ r q, s, s r).
Note that a double-p-marked hypersequent 1
p
= (1 [ p [ , p ) is always
p-regular and
DEN(1
p
, 1
p
) = (1 [ 1 [ , [ , ).
Now let us apply these ideas to a concrete family of calculi. We will call a hypersequent
rule local if it is the hypersequent version of a sequent rule, substitutive, and for each of
its instances, any variable occurring on the left (right) in a premise must occur on the left
(right) in the conclusion. In other words, local rules do not shift variables from one side
of a sequent to the other. The structural rules (W), (C), (MIX), (EMP), and (C
n
) (n 2)
are all local, but not, for example
( [
( [
THEOREM 4.2.3. Let GL be any extension of GMTL with local single-conclusion
rules. Then GL
D
admits density elimination.
246 George Metcalfe
Proof. As for cut elimination, it is enough to consider the uppermost applications of
(DENSITY). More generally, we prove that for any p-regular hypersequent ( and double-
p-marked hypersequent 1
p
= (1 [ p [ , p ):
If d
GL
( and d
t

GL
1
p
, then
GL
DEN((, 1
p
).
To see that this sufces, consider an uppermost application of (DENSITY) with premise
( = ((
t
[ p [ , p ). Suppose that
GL
(. Then by cut elimination,

GL
( and it follows from the claim with 1
p
= ((
t
[ p [ , p ) that

GL
(
t
[ (
t
[ , [ , . So by (EC),
GL
(
t
[ , as required.
We prove the claim by induction on h(d). Consider the last rule (r) applied in d.
If (r) is (ID) and ( = ((
t
[ ) (where cannot be p since ( is p-regular), then
DEN((, 1
p
) = (1
t
[ ) for some 1
t
and is derivable by (ID). If (r) is (EC) or
(EW), then the claim follows by applying the induction hypothesis and (r). Otherwise:
Suppose that (r) is a rule other than (ID), (EC), (EW), or (COM), and d ends with
(
t
[ S
1
. . . (
t
[ S
n
(
t
[ S
(r)
Note that ((
t
[ S) is p-regular by assumption. Also (r) is local so occurrences of
p cannot switch sides in a sequent from the premises to the conclusion. Hence
((
t
[ S
1
), . . . , ((
t
[ S
n
) are all p-regular, and by the induction hypothesis:

GL
DEN(((
t
[ S
i
), 1
p
) for i = 1 . . . n.
But DEN(((
t
[ S), 1
p
) is the result of multiple applications of (CUT) between
(1 [ (
t
[ S) and ( p) and (, p ). Since (r) is substitutive or almost-
substitutive and p cannot occur in the premises of an instance of (r) with no p in
the conclusion, we obtain an instance of (r)
DEN(((
t
[ S
1
), 1
p
) . . . DEN(((
t
[ S
n
), 1
p
)
DEN(((
t
[ S), 1
p
)
So
GL
DEN(((
t
[ S), 1
p
) as required.
Suppose now that (r) is (COM). If both premises are p-regular, then the claim
follows by applying the induction hypothesis to the premises and using (COM).
For example, suppose that d ends with:
(
t
[
1
,
1
p (
t
[
2
,
2
, [p]
k

(
t
[
1
,
2
, [p]
k
[
1
,
2
p
(COM)
By the induction hypothesis twice:

GL
DEN((
t
, 1
p
) [
1
,
1
, and
GL
DEN((
t
, 1
p
) [
2
,
2
,
k
.
Hence by (COM), as required:

GL
DEN((
t
, 1
p
) [
1
,
2
,
k
[
1
,
2
, .
Chapter III: Proof Theory for Mathematical Fuzzy Logic 247
Suppose then that one of the premises is not p-regular and d ends with:
(
t
[
1
,
1
, [p]
m+1
p (
t
[
2
, [p]
k
,
2

(
t
[
1
,
2
, [p]
k+m+1
[
1
,
2
p
(COM)
Let (
1
= DEN((
t
, 1
p
). Then by the induction hypothesis:
d
1

GL
(
1
[
2
,
k
,
2
.
Our aim is to nd a derivation for

GL
(
1
[
1
,
2
,
k+m+1
[
1
,
2
, .
Consider the GL

-derivation d
t
of (1 [ p [ , p ). We can substitute
&
2
(recalling that &[] = 1) for p in this derivation to get
d
2

GL
1 [ &
2
[ , &
2
.
Let d
3
be the (easy) derivation of (1 [
2
&
2
) using (&), (1), and
(ID), and let d
t
2
be the derivation
.
.
. d
2
1 [ &
2
[ , &
2

1 [
m+1
&
2
[ , &
2

(W)
1 [
m+1
&
2
[
1
, &
2
,
(W)
.
.
. d
3
1 [
2
&
2
1 [
m+1
&
2
[
1
,
2
,
(CUT)
(
1
[
m+1
&
2
[
1
,
2
,
(EW)
Also, let d
t
1
be the derivation
.
.
. d
1
(
1
[
2
,
k
,
2

.
.
.
(&) or (1)
(
1
[
2
,
k
, &
2

(&) or (1)
(
1
[
1
,
2
,
k
, &
2

(W)
Finally, putting these pieces together, we obtain the required derivation:
.
.
. d
t
1
(
1
[
1
,
2
,
k
, &
2

.
.
. d
t
2
(
1
[
m+1
&
2
[
1
,
2
,
(
1
[
1
,
2
,
k+m+1
[
1
,
2
,
(CUT)

COROLLARY 4.2.4. GMTL

D
, GSMTL
D
, GG
D
, and GMTL
D
n
(n 3) admit density
elimination.
248 George Metcalfe
While for single-conclusion calculi with weakening, density elimination goes hand-
in-hand with cut elimination, calculi that lack weakening and/or are multiple-conclusion,
are more difcult to deal with. Density elimination has been established for several
such calculi, in particular, GIMTL for involutive monoidal t-norm logic and GUL for
uninorm logic (or semilinear bounded FL
e
-algebras), making use of tailored induction
hypotheses. However, no method is yet known for GIUL.
THEOREM 4.2.5. GIMTL
D
, GNM
D
, GUL
D
, GUML
D
, and GIUML
D
admit density
elimination.
The key application of density elimination is to establish standard completeness results
for fuzzy logics, or, equivalently, to show that certain varieties of semilinear bounded
FL
e
-algebras are generated (as quasivarieties) by their members with lattice reduct [0, 1].
More precisely, density elimination can be used to show that a fuzzy logic is complete
with respect to a class of linearly and densely ordered bounded FL
e
-algebras (so-called
rational completeness); completeness with respect to a class of bounded FL
e
-algebras
with lattice reduct [0, 1] is then established algebraically using a Dedekind-MacNeille
completion procedure.
Recall (e.g., from Chapters II and IV) that with any axiomatic extension HL of
HUL, we can associate a variety ' of bounded FL
e
-algebras, having linearly ordered
members (chains) '"N('), in such a way that for all T Fm
1
p
:
T
HL
i 1 [ T [=
L
1 i 1 [ T [=
LIN(L)
1 .
Let us now dene |EN(') to be the class of all dense linearly ordered members of ',
and say that HL has the Density Property if for any T , , Fm
1
p
and variable
p not occurring in T , , :
T
HL
( p) (p ) T
HL
( ) .
A proof of the following lemma (in a more general setting) may be found in Theo-
rem 3.3.8 of Chapter II.
LEMMA 4.2.6. For any axiomatic extension HL of HUL with the density property:
T
HL
i 1 [ T [=
DEN(L)
1 .
Suppose now that there is a hypersequent calculus GL that is sound and complete
with respect to HL such that GL
D
admits density elimination. By the local deduction
theorem for axiomatic extensions of HUL (also established in a more general setting
in Section 2.6 of Chapter II), if T
HL
( p) (p ) , then for some

= (
1
1) & . . . & (
m
1) with
1
, . . . ,
m
T

HL

(( p) (p ) )
and, using some derivabilities in HUL

HL
((

& ) p) (p (

)) (

).
Chapter III: Proof Theory for Mathematical Fuzzy Logic 249
But then by the completeness of GL

GL

, p [ p,

and using rst the density rule and then density elimination for GL
D

GL

, ,

.
Finally, using the soundness of GL and some derivabilities in HUL

HL
(

&

) (( ) )
and so by the other direction of the local deduction theorem
T
HL
( ) .
Hence HL has the density property, and we obtain, using Lemma 4.2.6, Corollary 4.2.4,
and Theorem 4.2.5:
THEOREM 4.2.7. If L is MTL, SMTL, MTL
n
(n 3), GG, IMTL, NM, UL, UML,
or IUML, then
T
HL
i 1 [ T [=
DEN(L)
1 .
5 The fundamental fuzzy logics
We turn our attention in this section to developing proof theory for some of the
most renownedor in H ajeks parlance, fundamentalfuzzy logics: G odel logic G,
ukasiewicz logic , and product logic P. For G odel logic, we introduce and exploit
some interesting alternatives to the hypersequent calculus GG. For ukasiewicz logic
and product logic, where the standard approach described in the last two sections fails,
we obtain calculi by changing the interpretation of hypersequents and introducing new
logical rules. Finally, we consider a generalization of hypersequents for which these
three logics have a uniform proof-theoretic presentation.
5.1 G odel logic
Let us rst refresh our memory of G odel logic G in the language /
c
with binary
connectives , , and constants , . The standard semantics of G is characterized
by the G odel t-norm min and its residuum
G
, dened on the real unit interval [0, 1] by
x
G
y =
_
y if x > y
1 otherwise.
More precisely, a G-valuation is a function v : Fm
1
c
[0, 1] satisfying
v() = 0
v() = 1
v( ) = min(v(), v())
v( ) = max(v(), v())
v( ) = v()
G
v().
A formula is G-valid, written [=
G
, iff v() = 1 for all G-valuations v.
250 George Metcalfe
, ,
,
()
G
, ,
, ( )
(())
, ,
,
()
, , ,
, ( )
(())
, ,
, ( )
()
, ,
, ( )
()
, ,
( ) ,
()
, ,
( ),
()
, ,
( ),
(())
, , ,
( ) ,
(())
Figure 5. Sequent decomposition rules for G odel logic
The hypersequent calculus GG is an elegant and informative presentation of G odel
logic, an extension both of GMTL (for monoidal t-normlogic) with contraction, and of a
hypersequent version of Gentzens LJ (for intuitionistic logic) with communication. Cut
elimination for GG provides an easy proof of decidability, and, via density elimination,
a more complicated proof of standard completeness. However, even with loop-checking,
which gives termination of the rules (read upwards), GG is not particularly efcient for
theorem proving. External contraction can double the size of hypersequents, and since
not all the logical rules of the calculus are invertible, backtracking is required for proof
search. Below we consider two calculi that address these issues, starting with a system
in Gentzens original sequent framework.
Since sequents are not as exible as hypersequents, we remove the restriction to
single-conclusion sequents and dene (recalling that
_
[] =
def
and
_
[] =
def
):
I
G
( ) =
def

.
We also make use of more complicated rules for connectives that decompose formulas
into formulas with a smaller complexity, assuming for convenience that =
def
(( ) ) (( ) ) since the number of rules required increases
exponentially with the number of connectives.
Let us call a sequent atomic implicational if it contains only atoms and implications
of the forma b where a and b are atoms. Then it is easily seen that (read upwards) the
rules dened in Figure 5 reduce any sequent to a set of atomic implicational sequents.
Moreover, by simple arithmetic, these rules are sound and invertible with respect to
G-validity (i.e., the conclusion of each rule instance is G-valid iff all the premises are
G-valid). So we can check if a sequent is G-valid by applying the decomposition rules
(upwards) exhaustively, then checking the G-validity of the resulting atomic implica-
tional sequents. For this last step, we may also give a more immediately meaningful
presentation of G-validity.
An inequality is an ordered triple a b where a and b are atoms and <, . A
set of inequalities o is G-valid, written [=
G
o, iff for all G-valuations v, v(a) v(b) for
Chapter III: Proof Theory for Mathematical Fuzzy Logic 251
some a b o. Moreover, we dene a set of inequalities Iqs( ) for each atomic
implicational sequent by
(a b) Iqs( ) if (a b) ( b) Iqs( ) if b
(a < b) Iqs( ) if (b a) (a < ) Iqs( ) if a .
LEMMA 5.1.1. [=
G
I
G
( ) iff [=
G
Iqs( ).
Proof. Making use of the standard deduction theorem for G, [=
G
I
G
( ) iff for
every G-valuation v: either v() < 1 for some or v() = 1 for some . So
[=
G
I
G
( ) iff either v(a b) = 1 for some (a b) , v(b) = 1 for some
b , v(b a) < 1 for some (b a) , or v(a) < 1 for some a . Hence
[=
G
I
G
( ) iff v(a) v(b) for some (a b) Iqs( ) as required.
Examples of G-valid sets of inequalities include
(p q), (q < r), (r s), (s < p) ( < p), (p < q), (q r).
In the rst case we have a sequence of inequalities beginning and ending with p, and in
the second we have a sequence beginning with and involving a non-strict inequality.
This chain-like form is a common feature of all G-valid sets of inequalities.
LEMMA 5.1.2. A set of inequalities o is G-valid iff there exists (a
i

i
a
i+1
) o for
i = 1 . . . n such that either (1) a
1
= a
n+1
or a
1
= or a
n+1
= , where
i
is for
some i 1, . . . , n, or (2) a
1
= and a
n+1
= .
Proof. o is clearly G-valid if conditions (1) or (2) are met. For the other direction, we
proceed by induction on the number of different variables occurring in o. Note rst that
if one of a a, a , a, or < occurs in o, then we are done. This takes
care of the cases with at most one variable. Otherwise, we x a variable q occurring in
o, and dene
o
<
=
def
a < b [ a < q, q < b o, a ,= q, b ,= q
o

=
def
a b [ a
1
q, q
2
b o, a ,= q, b ,= q,
1
,
2

o
t
=
def
a b o [ a ,= q, b ,= q o
<
o

.
o
t
has fewer different variables than o. So if o
t
is G-valid, then applying the induction
hypothesis to o
t
, we have (a
i

i
a
i+1
) o
t
for i = 1 . . . n, satisfying either (1) or
(2) above. But then easily by replacing the inequalities a
i

i
a
i+1
that occur in o
<
or
o

appropriately by a
i

t
q and q
tt
a
i+1
, we get that (1) or (2) holds for o. Hence
it is sufcient to show that o
t
is G-valid. Suppose otherwise, i.e., that there exists a
G-valuation v such that v(a) v(b) does not hold for any a b o
t
. We show for a
contradiction that o is not G-valid. Let
x = minv(a) [ a q o and y = maxv(b) [ q b o.
Note rst that x y. Otherwise it follows that for some a, b, we have a
1
q, q
2
b o
and v(a) < v(b). But then (a b) o
t
so v(a) v(b), a contradiction. So there are
252 George Metcalfe
Axioms
, ,
(IDW)
,
()
,
()
Decomposition rules: Figure 5.
Atomic rules
, b a,
, a b
()
a
, a b
a b,
()
a
, b a a b,
a b,
(LIN)
Figure 6. The sequent calculus GG
g
two cases. If x > y, then we change v so that x > v(q) > y. For any (a q) or (q b)
in o, we have v(a) x > v(q) > y v(b). Hence o is not G-valid, a contradiction.
Now suppose that x = y and let us change v so that v(q) = x. We must have atoms a
0
,
b
0
such that (a
0
< q) and (q < b
0
) are in o and v(a
0
) = v(b
0
) = v(q). Consider any
(a
1
q) or (q
2
b) in o. Since (a
1
b
0
) and (a
0

2
b) are in o
t
, v(a)
1
v(q) = v(b
0
)
and v(a
0
) = v(q)
2
v(b) cannot hold. So o is again not G-valid, a contradiction.
The sequent calculus GG
g
displayed in Figure 6 extends the decomposition rules
for G with rules for dealing with atomic implicational sequents. The rule ()
a
is the
usual classical implication left rule restricted to atoms, while ()
a
combines applica-
tions of weakening and implication right rules. The rule (LIN) is what really extends the
calculus beyond intuitionistic logic and characterizes the linearity of the truth values.
EXAMPLE 5.1.3. We illustrate GG
g
with the following derivation:
r r, q p
(IDW)
q r, p q p q, r, q p
(IDW)
q r p q, r, q p
(LIN)
(p q) r r, q p
(())
THEOREM 5.1.4.
GG
g
S iff [=
G
I
G
(S).
Proof. The left to right direction proceeds as usual by induction on the height of a proof
of S in GG
g
. For the right to left direction, suppose that [=
G
I
G
( ). We can
assume that is atomic implicational. Let
t
= [a b [ b a ]. Observe
that [=
G
I
G
(
t
) and is derivable from
t
using repeated applications
of (LIN). Hence there exists a sequence of inequalities (a
i

i
a
i+1
) Iqs(
t
) for
i = 1 . . . n satisfying either (1) or (2) from Lemma 5.1.2. Moreover, we can assume that

i
is for at most one i. Otherwise, replacing any one of the two or more occurrences
of with < gives a sequence of inequalities that still satises either (1) or (2). If a b
is replaced with a < b where (a b) , then (b a) . Also if b , then
removing b still gives a sequence of inequalities satisfying (1) or (2).
We use the sequence to consider the form of the sequents. There are several cases.
As a rst example, suppose that contains an atom a
1
and a
1
< . . . < a
k
< .
Then the sequent is of the form (
tt
, a
k
, a
k
a
k1
, . . . , a
2
a
1
a
1
,
tt
) which is
easily derived using ()
a
and (IDW). Suppose now that contains an implication
a
1
a
2
and we have a sequence a
1
a
2
< . . . < a
k
< a
1
. Then the sequent is of
the form (
tt
, a
1
a
k
, . . . , a
3
a
2
a
1
a
2
,
tt
) which is easily derived using
()
a
, ()
a
, and (IDW). Other cases are very similar.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 253
Axioms: all G-valid atomic sequents of relations.
Logical rules
G | |
G |
()
G | G |
G |
()
G | G |
G |
()
G | |
G |
()
G | < G | <
G | <
(<)
G | | < G | <
G | <
(<)
G | | < G |
G |
()
G | |
G |
()
Figure 7. The sequent of relations calculus GG
r
Notice that unlike the completeness proof for GG, the above proof is entirely se-
mantic. This allows us to sketch a simple proof of standard completeness for a given
Hilbert system HG for G odel logic. Suppose that [=
G
. Then by Theorem 5.1.4,

GG
g
. But it is usually (depending on the Hilbert system) easy to show that GG
g
is sound with respect to HG. So
HG
.
Finally, there is a way to have the best of both worlds: invertible logical rules that
feature just one principal connective at a time. We allow two types of sequents, cor-
responding intuitively to and <, and treat sets of these sequents where exactly one
formula occurs on each side. More precisely, a sequent of relations S is a set of ordered
triples, written

1

1

1
[ . . . [
n

n

n
,
where
i
,
i
Fm
1
c
and
i
<, for i = 1 . . . n. S is G-valid, written [=
G
S, iff
for all G-valuations v, v(
i
)
i
v(
i
) for some i 1, . . . , n.
A sequent of relations calculus GG
r
for G odel logic in the language /
c
is displayed
in Figure 7, where the axioms are dened simply to be all valid atomic (containing only
atoms) sequents of relations.
EXAMPLE 5.1.5. We illustrate GG
r
with the following derivation:
p q [ q [ p [ q < p p q [ q [ p [ q p
p q [ q [ p q p [ p
()
p q [ q [ (p q) p
()
p q [ (p q) p
()
(p q) ((p q) p)
()
Observe that the uppermost sequents of relations of this derivation are G-valid since for
any G-valuation v, always v(p) v(q) or v(q) < v(p).
Note that the logical rules preserve G-validity in both directions, so we can reduce
the G-validity of a sequent of relations to the G-validity of a set of atomic sequents of
relations, and hence obtain:
254 George Metcalfe
THEOREM 5.1.6.
GG
r
S iff [=
G
S.
Further rules may also be added, along similar lines to GG
g
, to deal directly with proving
the G-validity of atomic sequents of relations.
5.2 ukasiewicz logic and Giless game
Just as G odel logic G may be considered the logic of order, so ukasiewicz logic
can be viewed as the logic of magnitude. In this logic, size matters. Whereas G is
based on the (only) idempotent t-norm min, ukasiewicz logic is based on the nilpotent
Archimedean t-norm x y = max(0, x + y 1). For simplicity, let us again use
a restricted (but fully expressive) language, /

, with one binary connective and a

constant , dening
=
def
=
def

& =
def
( ) =
def

=
def
& ( ) =
def
( ) .
An -valuation is a function v : Fm
1

[0, 1] such that v() = 0 and

v( ) = min(1, 1 v() + v())
where the valuations of the dened connectives emerge as expected as:
v() = 1 v() v() = 1
v( & ) = max(0, v() + v() 1) v( ) = min(1, v() + v())
v( ) = min(v(), v()) v( ) = max(v(), v()).
A formula is -valid, written [=

, iff v() = 1 for all -valuations v.

Unlike G odel logic, we have not as yet encountered any sequent or hypersequent
calculi for ukasiewicz logic. The most desirable solution would be just to add some
further structural rules to the calculus GIMTL. However, so far no such rules have been
discovered, or indeed are expected. Instead, we take a different approach: we interpret
sequents outside of the language of . Rather than give a direct interpretation of sequents
and hypersequents as formulas, we dene a semantic criterion for validity.
For a multiset of formulas , we x

() =
def
1 +

[v() 1 [ ]
and dene for each hypersequent (
[=

( iff for all -valuations v:

v

()
v

() for some ( ) (.
This denition may seem a little strange. However, notice that for a single-conclusion
hypersequent (:
[=

( iff [=

I(().
In particular, [=

iff [=

. Moreover, as we will see shortly, there exists a quite

natural interpretation of hypersequents for ukasiewicz logic based on a simple two-
player dialogue game.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 255
Axioms
G |
(ID)
G |
(EMP)
G | ,
()

Structural rules:
G
G | H
(EW)
G | H | H
G | H
(EC)
G |
G | ,
(W)
G |
1
,
2

1
,
2
G |
1

1
|
2

2
(SPLIT)
G |
1

1
G |
2

2
G |
1
,
2

1
,
2
(MIX)
Logical rules
G | , ,
G | ,
()
A
G | G | , ,
G | ,
()

Figure 8. The hypersequent calculus G

The hypersequent calculus G is presented in Figure 8. There are some key differ-
ences in comparison with calculi of previous sections. First, but just for convenience,
the calculus is cut-free. More importantly, the logical rules for are non-standard, the
left rule has just one premise while the right rule has two. Observe, nevertheless, that
the standard implication rule () is derivable for the single-conclusion case:
( [
(EMP)
( [
(W)
( [ ,
( [
()

Note also that the standard rules for and can be added, while the derived initial
hypersequents for are of the form (( [ ), that is, the single-conclusion version
of (). On the other hand, the appropriate rules (derived and simplied) for the
dened connective & are non-standard:
( [ , , [ ,
( [ & ,
(&)

( [ , , ( [ ,
( [ , &
(&)

EXAMPLE 5.2.1. Consider the following derivation in G:

(ID)

(ID)
, ,
(MIX)
,
()
A

(ID)

(ID)
, ,
(MIX)
, , ,
(W)
, ,
()

( ) ,
()
A
( ) ( )
()
(( ) ) (( ) )
()
THEOREM 5.2.2. [=

( iff
G
(.
256 George Metcalfe
Proof. The right-to-left direction is established as usual by an induction on the height
of a derivation of ( in G. For the left-to-right direction, we make use of the following
G-derivable rules:
( [ [ , ,
( [ ,
()

( [ ( [ , ,
( [ ,
()

Applying these rules backwards, we get that every -valid hypersequent is G-derivable
from -valid atomic hypersequents. It therefore sufces to prove the theorem for the
case where ( is atomic. We proceed by induction on the number of distinct propositional
variables occurring on the left hand side of sequents in (. Suppose that there are none.
It follows that only occurs on the left of sequents. We claim that there must exist a
sequent where the number of occurrences of on the left is greater than or equal to the
number of formulas on the right. If not, then dening an -valuation where all variables
take the value 0, we obtain a contradiction. Hence we can easily derive ( using (EW),
(W), (MIX), and ()

.
Otherwise, we pick a variable q occurring on the left of one of the sequents of (.
If q occurs on both sides in the same sequent, then we apply (MIX) and (ID) backwards
to remove it, noting that the new hypersequent is also -valid. Next, we use (EC) and
(SPLIT) backwards to multiply sequents, giving (for some ) a hypersequent
(
t
= ((
0
[ [
i
, [q]

i
]
n
i=1
[ [
j
[q]

,
j
]
m
j=1
)
where q does not occur in (
0
,
i
,
i
,
j
, or
j
for i = 1 . . . n and j = 1 . . . m.
Observe that
G
( if
G
(
t
. Also [=

(
t
. Let us now dene
1 = ((
0
[ [
i
,
j

j
,
i
]
j=1...m
i=1...n
[ [
i

i
]
n
i=1
[ [
j
[q]

,
j
]
m
j=1
).
Clearly 1 contains fewer distinct variables occurring on the left of sequents. Also
(
t
is derivable from 1. Reasoning backwards, we apply (EC) and (SPLIT) to (
t
to
combine sequents of the form (
i
, [q]

i
) and (
j
[q]

,
j
) into one:
(
i
,
j
, [q]

[q]

,
i
,
j
). Then we apply (MIX) and (ID) backwards to remove the
balanced occurrences of q, and (W) backwards to (
i
, [q]

i
) to get (
i

i
).
Hence it is sufcient to show that 1 is -valid, since then by the induction hypothesis

G
1. Suppose otherwise for a contradiction, i.e., that there exists an -valuation v
such that
v

() >
v

() for all ( ) 1. We dene

x = max(
v

(
i
)
v

(
i
) [ 1 i n )
y = min(
v

(
j
)
v

(
j
) [ 1 j m 0).
Notice rst that x < y. Otherwise,
v

(
i
) +
v

(
j
)
v

(
j
) +
v

(
i
) for some i, j:
a contradiction. Now change v so that x < (v(q) 1) < y, noting that v(q) (0, 1).
Then for i = 1 . . . n and j = 1 . . . m:

(
i
)
v

(
i
) < (v(q) 1) and (v(q) 1) <
v

(
j
)
v

(
j
).
Hence, rearranging,
v

(
i
[q]

) >
v

(
i
) and
v

(
j
) >
v

(
j
[q]

). But then
,[=

(
t
, a contradiction.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 257
Axioms

(ID)

(EMP)
,
()

Structural rules

n

(SC
n
) n 2

,
(W)

1

1

2

2

1
,
2

1
,
2
(MIX)
Logical rules
, , ,
,
()
s

, ,
,
()

Figure 9. The sequent calculus G

s
A nice by-product of this completeness proof is a syntactic proof of the decidability
of validity in ukasiewicz logic. A hypersequent is -valid iff it is derivable by applying
the rules ()

and ()

exhaustively and then iteratively removing occurrences

of each variable as described above. Hence we have a terminating procedure to decide
whether any given formula is or is not -valid.
Dening a hypersequent calculus for requires ingenuity, but it does not come as a
complete surprise. The existence of a sequent calculus for on the other hand is quite
unexpected. The key idea here is to represent differences between sequents occurring in
a hypersequent using implicational formulas. Consider again the rule ()

. Using
the fact that v() + v( ) = v() + v( ) for all -valuations v, we replace
this rule at the sequent level with
, , ,
,
()
s

Notice that the premise (, , , ) is derivable both from (, , )

using (W), and from ( ) using ()
s

again together with the sequent rules

(MIX), (ID), and (W). Intuitively, we think of (, , , ) as representing
both sequents. The combination of such sequents, performed by the split rule in the
hypersequent calculus, is achieved here by the sequent contraction rules (SC
n
) for n 2.
The resulting calculus G
s
is displayed in Figure 9.
EXAMPLE 5.2.3. Compare the following G
s
-derivation with the G-derivation of
the same sequent given in Example 5.2.1, recalling that the standard single-conclusion
implication right rule () is derivable using ()

, (W), and (EMP):

(ID)

(ID)

(ID)
, ,
(MIX)
( ), , ,
(W)
( ) ,
()
s

( ) , , ,
(MIX)
( ) ,
()
s

( ) ( )
()
(( ) ) (( ) )
()
258 George Metcalfe
The completeness proof for G
s
relies on a (quite complicated) translation of sequent
derivations into derivations in the hypersequent calculus G, and is omitted here.
THEOREM 5.2.4. [=

S iff
G
s
S.
We conclude our discussion of ukasiewicz logic with a dialogue game introduced
by Robin Giles in the 1970s (for details see the historical remarks at the end of the
chapter) that has an intriguing connection to the hypersequent calculus described above.
Suppose that two players-me and you-agree to pay 1$ to their opponent for every
false statement that they make. An elementary state of the game consists of a multiset of
atoms [a
1
, . . . , a
m
] asserted by you and a multiset of atoms [b
1
, . . . , b
n
] asserted by me,
denoted [a
1
, . . . , a
m
b
1
, . . . , b
n
]. Each atom a is read as the (repeatable) elementary
(yes/no) experiment E
a
yields a positive result. If a is either always true or false, the
setting is classical. The situation is more interesting, however, if the same experiment
can yield different results when repeated. In this case, for every run of the game, a
xed risk value q [0, 1] is associated with each variable q, where = 1. The risk
associated with a multiset of atoms is then [a
1
, . . . , a
m
] =
def
a
1
+. . .+a
m
. That is,
my risk corresponds to the amount of money that I expect to have to pay to you according
to the results of the experiments associated with the atoms that I assert. Hence, for an
elementary state [a
1
, . . . , a
m
b
1
, . . . , b
n
], the condition a
1
, . . . , a
m
b
1
, . . . , b
n

expresses that I do not expect any loss (but possibly some gain).
More generally, a dialogue state (d-state) of the game is denoted [ ] where
and are nite multisets of formulas currently asserted by you and me, respectively.
Non-atomic formulas in and may then be decomposed via the following dialogue
rule for implication:
(R

) If I assert , then whenever you attack this statement by assert-

ing , I must assert also . (And vice versa, i.e., for the roles of me
and you switched.)
This rule reects the idea that implication is characterized by the principle that asserting
if , then obliges the assertion also of if the opponent in a dialogue asserts .
However, a player may also choose never to attack the opponents assertion of .
A round with initiator P (me or you) and respondent Q (the other player) is therefore a
transition from one d-state to another successor d-state that consists of two moves:
1. P picks one of the formulas asserted by Q.
2. P either attacks this formula by asserting or grants the formula, meaning that
P declares never to attack that particular occurrence. In the rst case, Q must
respond immediately by asserting ; in the second, no action is required from Q.
Each formula can be attacked or granted at most once. Hence the formula chosen by
the initiator is removed from the d-state. In denoting strategies as trees, below, we use
intermediary states or, for short, i-states to reect the initiators choice of the formula
that gets attacked or granted. Intermediary states are exactly the same as the preceding d-
state, except that the formula chosen by the initiator is marked (denoted by underlining).
Chapter III: Proof Theory for Mathematical Fuzzy Logic 259
To dene games precisely, we also need to know whose turn it is to initiate the next
round for each non-elementary d-state. Formally, we dene a regulation as a function
that assigns to each non-elementary d-state either the label Y, for you initiate the next
round, or the label I, for I initiate the next round, indicating that the possible runs of
the game are constrained accordingly. A regulation is consistent if the label Y (or I) is
only assigned to d-states where such an initiating move is possible, i.e., where there is a
formula among my (or your) currently asserted formulas. The correct label for
an i-state is determined by the label of the immediately preceding d-state.
A game form G([ ], ) is a tree of states with the initial d-state [ ] as root,
where the successor nodes to any state S are the states that result from legal moves at S
according to the consistent regulation . In particular, the leaf nodes of G([ ], )
are the reachable elementary states. A game consists of a game form G([ ], )
together with a risk assignment , and a run of the game is a branch of G([ ], ).
In other words, a run consists of a sequence of alternating d-states and i-states, obtained
from successive moves as described above, beginning with the initial d-state [ ] and
ending in an elementary state.
EXAMPLE 5.2.5. Consider the d-state [p q a b, c d]. If it is my turn to initiate
the next round (indicated by
I
), then I can only either attack or grant your statement
p q. Accordingly, there are two ways that runs of the game can continue:
[p q a b, c d]
I
[p q a b, c d]
I
[q p, a b, c d]
[p q a b, c d]
I
[p q a b, c d]
I
[ a b, c d].
If it is your turn to move, then there are two successor i-states and hence four possible
continuations depending on (1) which of my two statements you choose, and (2) whether
you decide to attack or grant the chosen formula:
[p q a b, c d]
Y
[p q a b, c d]
Y
[p q, a b, c d]
[p q a b, c d]
Y
[p q a b, c d]
Y
[p q c d]
[p q a b, c d]
Y
[p q a b, c d]
Y
[p q, c a b, d]
[p q a b, c d]
Y
[p q a b, c d]
Y
[p q a b].
Suppose that a run of G([ ], ) with risk assignment ends with the elemen-
tary state [a
1
, . . . , a
m
b
1
, . . . , b
n
]. We say that I win in that run if I do not expect any
loss of money resulting from betting on results of the corresponding elementary experi-
ments: more formally, I win if a
1
, . . . , a
m
b
1
, . . . , b
n
. In general, a game strategy
for a particular player is a function from states to states that determines every choice
260 George Metcalfe
of a legal move by that player but leaves all choices of the other player open. In our
context this means that a strategy for me is obtained from a game form by (iteratively
from the root) deleting all but one successor of every state labelled I. A strategy for a
game form G([ ], ) is called a winning strategy (for me) for a risk assignment
if a
1
, . . . , a
m
b
1
, . . . , b
n
holds for each of its leaf nodes [a
1
, . . . , a
m
b
1
, . . . , b
n
].
The following result expresses the key relationship between -validity and winning
strategies in the game (see the historical remarks for references):
THEOREM 5.2.6. A formula is -valid iff I have a winning strategy for the game
G([ ], ) with any risk assignment , where is an arbitrary consistent regulation.
To connect Giless game to cut-free proofs in an appropriate calculus for ukasiewicz
logic, we must abstract from particular risk assignments and look rather at disjunctive
(winning) strategies that arise when disjunctions of states instead of single game states
are considered. Let us use D = S
1
_
. . .
_
S
n
to denote a state disjunction. Since the
order of its component states (ordinary game states) S
1
, . . . , S
n
is irrelevant, a disjunc-
tive state may be viewed as a multiset of states. A disjunctive strategy for D respecting
a regulation is a tree of state disjunctions with root D where the successor nodes are
in principle determined in the same way as for ordinary strategies. However, we also
allow for the possibility of duplicating a component of a state in order to let disjunctive
strategies for a player P record different ways of proceeding for P in identical compo-
nent states. More precisely, there are two kinds of non-leaf nodes D = S
1
_
. . .
_
S
n
in a disjunctive strategy for P:
1. Playing nodes, focused on some component S
i
of D. The successor nodes are
like those for S
i
in ordinary strategies, except for the presence of additional com-
ponents (that remain unchanged). I.e., if, according to , it is Ps turn to play at
S
i
, then there is a single successor state disjunction where the component S
i
of D
is replaced by some S
t
i
corresponding to some move of P. If the opposing player
Q is to move at S
i
, then all possible moves of Q determine the successor nodes
of D where S
i
is replaced by a state obtained by the corresponding move.
2. Duplicating nodes, where the single successor node is obtained by duplicating one
of the components in D.
A disjunctive winning strategy (for me) for the family of instances of Giless game based
on G([ ], ) is a disjunctive strategy where in each leaf node for every risk assign-
ment there is at least one component, i.e., elementary d-state [a
1
, . . . , a
m
b
1
, . . . , b
n
]
such that a
1
, . . . , a
m
b
1
, . . . , b
n
.
As should be fairly obvious, d-states may be understood as sequents, and state dis-
junctions as hypersequents. Furthermore, a cursory analysis of disjunctive strategies re-
veals that (ignoring marking of formulas and the different notation) these are derivations
of hypersequents from atomic hypersequents using (EC), (EW), ()

, the derived
rule ()

used in the proof of Theorem 5.2.2, and the redundant rule (representing
the introduction of intermediate states)
( . . . (
(
(R)
Chapter III: Proof Theory for Mathematical Fuzzy Logic 261
Moreover, if a disjunctive strategy is winning, then the atomic hypersequents corre-
sponding to the leaf nodes are -valid. It is not the case that any derivation from-valid
hypersequents using (EC), (EW), ()

, ()

, and (R) corresponds directly to a

disjunctive winning strategy, but given such a derivation, a disjunctive winning strategy
can be constructed for any consistent regulation.
5.3 Product logic and related systems
Product logic P, the third fundamental fuzzy logic, possesses features of both of its
more famous siblings. Roughly speaking, it behaves like on the interval (0, 1], and like
G at 0. Let us work again with a more restricted language /
b
with binary connectives &
and , and a constant , dening
=
def
=
def
& ( )
=
def
=
def
(( ) ) (( ) ).
P-valuations are then functions v : Fm
1
b
[0, 1] such that v() = 0 and
v( & ) = v() v() v( ) =
_
v()/v() if v() > v()
1 otherwise.
A formula is P-valid, written [=
P
, iff v() = 1 for all P-valuations v.
In developing Gentzen systems for P, the problems are similar to those encountered
for . It does not seem to be possible to obtain a calculus by adding structural rules to
a standard calculus like GMTL. Rather, as for , we make use of a non-standard in-
terpretation for sequentsthis time within the language of the logicand then develop
tailored logical rules for that interpretation.
I
P
( ) =
def
& &
I
P
(S
1
[ . . . [ S
n
) =
def
I
P
(S
1
) . . . I
P
(S
n
).
Hence, letting
v
P
() =
def

v() where
v
P
([]) = 1, we have
[=
P
I
P
(() i for all valuations v,
v
P
()
v
P
() for some ( ) (.
The hypersequent calculus GPbased on this interpretation is displayed in Figure 10,
where is used here simply as a convenient abbreviation for .
EXAMPLE 5.3.1. We illustrate the calculus with the following derivation, noting that
the single-conclusion rule () is derivable in GP exactly as in G:

(ID)

(ID)
, ,
(MIX)
& ,
(&)
, & ,
(W)

(ID)
,
()
P
, ( & )
()
P
( ( & ))
()
(( ( & )) )
()
262 George Metcalfe
Axioms
G |
(ID)
G |
(EMP)
G | ,
()
Structural rules:
G
G | H
(EW)
G | H | H
G | H
(EC)
G |
G | ,
(W)
G |
1
,
2

1
,
2
G |
1

1
|
2

2
(SPLIT)
G |
1

1
G |
2

2
G |
1
,
2

1
,
2
(MIX)
Logical rules
G | , ,
G | , &
(&)
G | , ,
G | &,
(&)
A
G |
G | ,
()
P
G | G | , ,
G | ,
()

G | , G | , ,
G | ,
()
P
Figure 10. The hypersequent calculus GP
Soundness is proved in the usual way. For completeness, we proceed semantically.
First note that the following rule is GP-derivable using (EC), (W), ()
P
, and (EW):
( [ , ( [ [ , ,
( [ ,
()
i
P
Now let GP
+
be GP extended with the following sound and invertible (with respect to
P-validity) decomposition rules for negated formulas:
( [ , [
( [ , ( )
()
( [ , ( [ ,
( [ , ( & )
(&)
Completeness for GP
+
is shown by rst using invertible logical rules and the extra rules
for negation to reduce the P-validity of hypersequents to the P-validity of hypersequents
containing atoms and possibly also negated variables on the left of sequents, and then
showing that such hypersequents are derivable using a similar argument to that employed
for G in Theorem 5.2.2. Finally, completeness for GP is obtained by showing (induc-
tively) that applications of the extra rules can be eliminated from derivations in GP
+
.
THEOREM 5.3.2. [=
P
I
P
(() iff
GP
(.
Similarly to the case of G odel logic, GP can be used to establish standard complete-
ness for a given axiomatization HP of product logic. If [=
P
, then by the completeness
of GP,
GP
. But it is usually (depending on the Hilbert system) straightforward to
show that GPis sound with respect to HP. Hence
HP
. Moreover, following the same
ideas as for ukasiewicz logic, decidability of validity in product logic can be deduced
from the completeness proof for GP
+
and a sequent calculus GP
s
can be dened.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 263
Some related fuzzy logics, not as important as the three just covered, but interesting
nonetheless, can be tackled using similar techniques. Recall that cancellative hoop logic
CHL, the logic of cancellative hoops, emerges by removing from the language of P
and restricting P-valuations to the half-open interval (0, 1]. A for this logic is obtained
by removing the rule ()

from G and adding the rules for & of GP. Further

examples are cross ratio logic CRL, the logic based on the cross-ratio uninorm, and
Abelian logic A, the logic of Abelian lattice-ordered groups.
5.4 Uniform systems
The calculi for the fundamental logics dened above are nice in many respects, but
they do seem rather arbitrary. For each logic we have simply picked rules that work.
In contrast, the systems of the previous sections have the same logical rules and dif-
fer only at the structural level. Here we dene a similar framework for , G, and P.
That is, we dene uniform logical rules and distinguish the logics using only structural
rules. The price we pay is more structure: hypersequents are generalized to relational
hypersequents with two kinds of sequents.
We will use a language /
f
with binary connectives , &, , and constants ,
that is adequate for , G, and P, and indeed any logic based on a residuated t-norm. A
relational hypersequent (r-hypersequent) is a nite multiset of ordered triples, written

1

1

1
[ . . . [
n

n

n
where
i
<, and
i
and
i
are nite multisets of formulas for i = 1 . . . n. Note
that a hypersequent can be treated as an r-hypersequent with just one relation symbol,
and that a sequent of relations can be treated as an r-hypersequent where all multisets

i
,
i
contain exactly one formula.
Validity for r-hypersequents is dened for each logic individually, understanding
[ as before as a meta-level disjunction, where < and denote inequalities between
combinations (different for each logic) of truth values of formulas. The symbols < and
therefore have two uses: a syntactic one as part of an r-hypersequent, and a semantic
one as an inequality holding between mathematical expressions. Often we will use to
stand uniformly for or < (in either sense).
For L , G, P, we dene
v
L
([]) = 1 and for ,= []:

() = 1 +

[v() 1 [ ]

v
G
() = min[v() [ ]

v
P
() =

[v() [ ].
An r-hypersequent ( is said to be L-valid, written [=
L
(, iff for each L-valuation v,

v
L
()
v
L
() for some ( ) (.
Notice immediately that for any formula and L , G, P:
[=
L
iff [=
L
.
Hence we can express that a single formula is L-valid, as well as other relationships such
as [=
L
< that cannot be expressed using the L-validity of a formula.
264 George Metcalfe
[ , , [ [ [ <
[ ,
()
[ [ , , [
[ ,
()
[ , [ , ,
[ &,
(&)
[ , , [ ,
[ , &
(&)
[ , [ ,
[ ,
()
[ , [ ,
[ ,
()
[ , [ ,
[ ,
()
[ , [ ,
[ ,
()
Figure 11. Uniform relational hypersequent rules
EXAMPLE 5.4.1. ( = (r r, q [ p, q < p) is -valid since for any -valuation v:
v(q) = 1
v

[r] = v(r) = 1 + (v(r) 1) + (v(q) 1) =

v

[r, q]
v(q) < 1
v

[p, q] = 1 + (v(p) 1) + (v(q) 1) < v(p) =

v

[p].
However, ( is not L-valid for L G, P, since if v(p) = v(q) = 0 and v(r) > 0, then

v
L
[r] = v(r) > 0 =
v
L
[r, q] and
v
L
[p, q] = 0 = v(p) =
v
L
[p].
Our reward for this greater exibility, both in the structures and their interpretations,
is the set of uniform rules displayed in Figure 11 (recalling that is uniformly either or
< in each instance of a rule). Notice that the rules for , , and have the subformula
property, but the rules for & do not ( appears in the premises and possibly not the
conclusion). This is another cost of uniformity. In the cases of G and P, we could
remove the right premise of (&), and ( , ) in the premise of (&), while for
we could make do with just rules for .
The uniform rules are sound and invertible for , G, and P (i.e., for L , G, P,
the conclusion of any rule instance is L-valid iff all the premises are L-valid). Moreover,
it is easy to see that for each rule instance, the multiset complexity of the premises is
strictly less than the multiset complexity of the conclusion. Hence a sound and complete
calculus for L where L , G, P consists of the uniform rules extended with (as
initial r-hypersequents) all L-valid atomic r-hypersequents. Indeed, it can be shown
that the sets of L-valid atomic r-hypersequents , G, and P are polynomial time, and
by revising the uniform rules slightly, the systems can then be used to establish co-NP
completeness for the sets of valid formulas for these logics. Alternatively, structural
rulesdifferent for each logiccan be given for dealing with atomic r-hypersequents.
Details may be found in the references given at the end of this chapter.
6 Quantiers and modalities
In this section, we consider proof-theoretic methods for three extensions of fuzzy
logics beyond the propositional level: rst-order quantiers, propositional quantiers,
and modalities.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 265
6.1 First-order quantiers
Surprisingly perhaps, many of the methods and results for propositional logics trans-
fer unscathed to the rst-order level. We can extend the hypersequent calculi described
in Sections 3 and 4 with rules for the universal quantier and existential quantier
, and obtain completeness with respect to appropriate Hilbert systems or classes of
algebras. Moreover, although the logics are undecidable, analogues of Herbrands the-
orem and Skolemizationtwin pillars of theorem proving in classical logiccan be
established for their prenex fragments. In the case of the non recursively axiomatizable
rst-order ukasiewicz logic, the results are necessarily weaker; however, an alternative
approximate Herbrand theorem can be obtained, leading in turn to a cut-free hyperse-
quent calculus with an innitary rule.
Let us suppose that we have already dened a hypersequent calculus GL over the
usual propositional language /
p
and that ' is a xed countable rst-order language with
function symbols, relation symbols, and variables. To simplify matters, we make a syn-
tactic distinction between bound variables, denoted by x, y and free variables, denoted
by a, b, and assume that only bound variables are quantied and only free variables
occur freely. Terms are denoted by t, and rst-order formulas (as in the propositional
case) by , , , writing ( a) to denote that the free variables of are among those in
a = a
1
, . . . , a
n
and (

t) to denote distinguished occurrences of terms

t = t
1
, . . . , t
n
in
. We then dene GL to consist of all instances of rule schema of GL with '-formulas
together with the following quantier rules:
( [ , (t)
( [ , (x)(x)
()
( [ (a),
( [ (x)(x),
()
( [ , (a)
( [ , (x)(x)
()
( [ (t),
( [ (x)(x),
()
where a does not occur in the lower hypersequent in () or ().
Note that unlike in the usual quantier axioms, here we do not require that the term t is
substitutable for x in () and (): no free variable in t can be bound in (t) since
free variables and bound variables are distinguished.
EXAMPLE 6.1.1. We can use these quantier rules and the core rules for to give
sequent derivations of standard existential quantier axioms (noting that, by denition,
x does not occur freely in ):
(a) (a)
(ID)

(ID)
(a) , (a)
()
(x)( ), (a)
()
(x)( ), (x)
()
(x)( ) (x)
()
(x)( ) ((x) )
()
Furthermore, we can also derive the so-called shifting law of quantiers axioms.
Recall that () and () are derived rules of GUL, and, from Example 3.2.2, that
266 George Metcalfe

GUL
[ . Then we have a derivation in GUL (noting again
that, by denition, x does not occur freely in ):
(a) [ (a) (a)
(a) [ (x)( ) (a)
()
(x)( ) [ (x)( ) (a)
()
(x)( ) [ (x)( ) (x)
()
(x)( ) (x)
()
(x)( ) ( (x))
()
Such axioms are derivable in calculi for rst-order classical logic, but not in calculi for
rst-order intuitionistic logic or many other rst-order substructural logics.
Soundness and completeness are established for rst-order Gentzen systems with
respect to Hilbert systems similarly to the propositional case by showing that the new
quantier rules preserve derivability in the Hilbert system, and that the extra axioms
are derivable and the new rules preserve derivability in the hypersequent calculus. Cut
elimination also follows the same pattern as the propositional case. To take care of sub-
stituting variables, however, we denote a hypersequent containing a distinguished free
variable a by ((a), and make use of the following lemma, proved by a straightforward
induction on the height of a derivation.
LEMMA 6.1.2. Let GL be an extension of GUL with substitutive single-conclusion
rules or of GIUL with substitutive rules. If d; (
1
(a), . . . , (
n
(a)
GL
((a) and t is a
term with variables not occurring in d, then d
t
; (
1
(t), . . . , (
n
(t)
GL
((t) for some
derivation d
t
with h(d
t
) = h(d).
THEOREM 6.1.3. Let GL be an extension of GUL with substitutive single-conclusion
rules or of GIUL with substitutive rules. Then cut elimination holds for GL.
Proof. As in Theorem 4.1.1, it is sufcient to prove that for any hypersequent ( and
hypersequent 1 with marked formula :
If d

GL
( and d
)

GL
1, then
GL
(
t
for all (
t
CUT((, 1).
We prove the claim as before by induction on the lexicographically ordered triple
cp(), e(d
)
), h(d

)
assuming, using Lemma 6.1.2, that any new free variables introduced (upwards) by
() or () in d

(d
)
) do not occur in d
)
(d

).
If d

ends with a rule application where the principal formula is not an occurrence
of on the opposite side to d
)
, then (as in the propositional case), we can make use of
the almost-substitutivity of the rule and apply the induction hypothesis. The assumption
that new free variables are distinct in d

from those in d
)
and vice versa ensures that
this also works for the quantier rules. Let us therefore assumesince propositional
connectives are treated in the proof of Theorem 4.1.1that is of the form (x)(x)
and d

ends with:
Chapter III: Proof Theory for Mathematical Fuzzy Logic 267
(
t
[ , [(x)(x)]
1
, (t)
(
t
[ , [(x)(x)]

or
(
t
[ (a), [(x)(x)]
1
,
(
t
[ [(x)(x)]

,
where , and 1 is of the form, respectively
1
t
[ (x)(x), or 1
t
[ , (x)(x) .
Let (
)
CUT((, 1). The only tricky case is when (
)
is of the form (1
t
[ (
tt
[
,

, ) where (
tt
CUT((
t
, 1). But then also
either
1
t
[ (
tt
[ ,
1
, (t)
1
,
1
t
[ (
tt
[ ,
1
, (x)(x)
1
,
or
1
t
[ (
tt
[ ,
1
(a),
1
,
1
t
[ (
tt
[ ,
1
(x)(x),
1
,
is an instance of the appropriate rule. Moreover, by the induction hypothesis, the premise
is derivable so we have a derivation d ending with such a rule application.
If e(d
)
) = 1, i.e., d
)
does not end with the application of a logical rule to the
marked occurrence of , then we proceed as in the propositional case. Suppose therefore
that e(d
)
) = 0: i.e., d
)
ends with an application of () or () to the marked
occurrence of , and is of the form:
1
t
[ (a),
1
t
[ (x)(x),
or
1
t
[ , (t)
1
t
[ , (x)(x)
By Lemma 6.1.2, there is a derivation of the same height of (1
t
[ (t), ) or
(1
t
[ (
tt
[ ,
1
(t),
1
, ). But then by the induction hypothesis, since
cp((t)) < cp((x)(x)), using a further application of (EC),
GL
(
)
.
It follows from cut elimination that GL is a conservative extensions of GL. Just
observe that any GL-derivable propositional hypersequent has a cut-free derivation in
GL which does not involve any quantiers. It is also straightforward to extend den-
sity elimination to rst-order Gentzen systems: the quantier rules are treated just like
the rules for other connectives. Moreover, density elimination can be used to establish
standard completeness results for rst-order fuzzy logics following essentially the same
procedure as for propositional fuzzy logics described in Section 4.
Cut elimination will not help us with decidability; indeed, all these logics are un-
decidable. However, we can use these results instead to prove versions of Herbrands
theorem and Skolemization for their prenex fragments. Recall that a prenex formula is a
rst-order formula with all the quantiers at the front. We can show that the validity of
a prenex formula is equivalent to the validity of a set of propositional formulas. The key
technical tool for establishing this result is a mid-hypersequent theorem (an analogue
of Gentzens mid-sequent theorem) for a calculus GL, stating that any GL-derivable
268 George Metcalfe
prenex formula has a GL-derivation where the propositional connective inferences pre-
cede all quantier inferences. It then follows that there exist hypersequents in the deriva-
tion with no propositional connective inference below or quantier inference above.
For clarity, we focus on the case of GUL, and to save us some effort, perform ma-
nipulations on a slight variant of this calculus. Let GUL

be the cut-free hypersequent

calculus GUL

with (ID) restricted to instances containing only variables, and ()

and () replaced with
G |
1
,
1
G |
2
,
2
G |
1
,
1
|
2
,
2
()

G |
1
G |
2

G |
1
|
2

()

These rules are easily derived in GUL using (COM), and we leave it as an (easy)
exercise to show that
GUL
( iff
GUL
(.
THEOREM 6.1.4. Let ( be a single-conclusion hypersequent containing only prenex
formulas. If d
GUL
(, then d
t

GUL
( for some derivation d
t
where no proposi-
tional inference is below a quantier inference.
Proof. Let the order o(d) of a derivation d be the multiset containing the lengths (car-
dinalities) of any sub-branches of d (i.e., subpaths of branches of d) that start with a
propositional connective inference and end with a quantier inference. Then it is suf-
cient to establish the following:
If d
GUL
(, then d
t

GUL
( for some derivation d
t
where o(d
t
) = [].
We proceed by induction on o(d) using the standard multiset ordering <. The base case
where o(d) = [] is immediate. For the inductive step we have a number of possibilities.
(1) Suppose that a quantier inference occurs directly above a propositional connective
inference. Then we can rearrange the derivation so that the quantier rule application
is below the propositional connective rule application, and use the induction hypothesis.
The only tricky cases are where () appears above ()

, or () above ()

.
Let us consider the former, i.e., d ends with:
.
.
. d
1
1 [
1
, (t),
1
1 [
1
, (x)(x),
1
()
.
.
. d
2
1 [
2
,
2
1 [
1
, (x)(x),
1
[
2
,
2
()

We can replace this with the following derivation d

t
:
.
.
. d
1
1 [
1
, (t),
1
.
.
. d
2
1 [
2
,
2
1 [
1
, (t),
1
[
2
,
2
()

1 [
1
, (x)(x),
1
[
2
,
2
()
Since o(d
t
) < o(d) the induction hypothesis can be applied.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 269
If the previous case does not occur, then there must be a sub-branch with structural
inferences occurring between the propositional connective inference and quantier in-
ferences. We note without proof that applications of (EC) can be pushed downwards
over the other structural rules in derivations, and (EW) can be pushed upwards. If a
quantier inference is directly above (EW) or (COM), then we can move the application
of the structural rule above the quantier inference. In the (most complicated) case of
(COM) and (), we have the following situation:
.
.
. d
1
1 [
1
, (t),
1

1
,
1
1 [
1
, (x)(x),
1

1
,
1
()
.
.
. d
2
1 [
2
,
2

2
,
2
1 [
1
, (x)(x),
2

1
,
2
[
1
,
2

1
,
2
(COM)
We can replace this with the following derivation d
t
:
.
.
. d
1
1 [
1
, (t),
1

1
,
1
.
.
. d
2
1 [
2
,
2

2
,
2
1 [
1
, (t),
2

1
,
2
[
1
,
2

1
,
2
(COM)
1 [
1
, (x)(x),
2

1
,
2
[
1
,
2

1
,
2
()
Since o(d
t
) < o(d) the induction hypothesis can be applied. The nal possibility is that
an application of (EC) occurs directly above a logical connective inference. Again we
are always able to push the relevant rule application upwards.
For a formula , let C

, F

, and P

be, respectively, the constants, non-nullary

function symbols, and predicate symbols occurring in , adding a constant to C

if it is
empty. The Herbrand universe of is dened as H() =

n=0
H
n
() where
H
0
() = C

H
n+1
() = H
n
()f(t
1
, . . . , t
k
) [ t
1
, . . . , t
k
H
n
() and f F

with arity k.
Suppose now that
GUL
( x)( x) where is quantier-free. Then by Theo-
rem 6.1.4, we have a GUL

-derivation of ( ( x)( x)) where no propositional

inference follows a quantier inference. So there must be a GUL

-derivable mid-
hypersequent ( (

t
1
) [ . . . [ (

t
n
)) for some

t
1
, . . . ,

t
n
H(), and hence:
THEOREM 6.1.5.
GUL
( x)( x) iff
GUL

_
n
i=1
(

t
i
) for some terms

t
1
, . . . ,

t
n
H().
Herbrands theorem allows us to reduce the validity problem for an existential for-
mula to the validity of formulas that are (essentially) propositional. We can also make
use of this theorem (as in classical logic) to do the same for prenex formulas. The idea
here is to remove universal quantiers iteratively, replacing the variables that they bind
with terms consisting of a new function symbol with variables bound by preceding ex-
istential quantiers. This process is called Skolemization for fuzzy logics, although it
is worth noting that for classical logic, the usual process involves removing existential
quantiers and preserving satisability.
270 George Metcalfe
Let be a prenex formula and assume harmlessly that the ith occurrence of is
labelled
i
and that no function symbol f
i
of any arity occurs in . Then the Skolem
form
S
of is dened by induction as follows:
1. If is of the form ( x)( x) where is quantier-free, then
S
is ( x)( x).
2. If is of the form ( x)(
i
y)( x, y), then
S
is (( x)( x, f
i
( x)))
S
.
EXAMPLE 6.1.6. Consider the prenex formula = (x)(y)(z)(u)p(x, y, z, u).
Skolemizing in two steps, we obtain that
S
= (x)(z)p(x, f(x), z, g(x, z)).
Focussing again for clarity on the case of GUL, note that
GUL

S
follows inductively since
GUL
( x)(y)( x, y) ( x)( x, f( x)). The other
direction does not hold in general, even in classical logic. However, we can show instead
the weaker claim that
GUL

S
implies
GUL
.
LEMMA 6.1.7. Let ( x)( x) be the Skolem form of a prenex formula , and

t
1
, . . . ,

t
n
H(( x)( x)). Then is derivable from (

t
1
) [ . . . [ (

t
n
) using (EW),
(EC), (), and (), where in (), any variable-free term not occurring in the
conclusion may be used in the premise.
Hence if ( x)( x) is the Skolem form of a prenex formula and, moreover,
GUL

_
n
i=1
(

t
i
) for some

t
1
, . . . ,

t
n
H(), then by Lemma 6.1.7, ( ) is derivable
from ( (

t
1
) [ . . . [ (

t
n
)) using (EW), (EC), (), and (). So
GUL
as
required.
THEOREM 6.1.8. Let ( x)( x) be the Skolem form of a prenex formula . Then the
following are equivalent:
(1)
GUL

(2)
GUL
( x)( x)
(3)
GUL

_
n
i=1
(

t
i
) for some

t
1
, . . . ,

t
n
H().
The rst-order situation for ukasiewicz logic and its relatives P and CHL is more
complicated. For these logics, the set of valid formulas is not recursively enumerable
(for P, not even arithmetical), so nite sets of axiom and rule schema cannot be enough.
Let us take a closer look at this problem for rst-order ukasiewicz logic in the
language with connectives , , , and . Our rst observation is that the Herbrand
theorem cannot hold for this logic. Note that [=

(x)p(x) (y)p(y). So using

quantier-shifting equivalences of
[=

(y)(x)(p(x) p(y))
and by the easy direction of Skolemization
[=

(y)(p(f(y)) p(y)).
Chapter III: Proof Theory for Mathematical Fuzzy Logic 271
If the Herbrand theorem did hold for , then for some constant c and n N
+
:
[=

i=1
(p(f
i
(c)) p(f
i1
(c)))
where f
0
(c) = c and f
i+1
(c) = f(f
i
(c)) for all i N. But now we can dene a
structure such that the value of p(f
i
(c)) is i/n for i = 0 . . . n, a contradiction. So the
Herbrand theorem must fail.
Take another look at the formula
_
n
i=1
(p(f
i
(c)) p(f
i1
(c))), however. Al-
though this is not a valid formula of , it comes within one nth of being one. Observe
that for any r
0
, r
1
, . . . , r
n
[0, 1]:
min
i1,...,n]
r
i1
r
i
1/n and so also max
i1,...,n]
1 r
i1
+ r
i
1 1/n.
Let us write for r [0, 1]:
[=
>r

iff takes a value > r in all structures.

Then for any r < 1 1/n:
[=
>r

i=1
(p(f
i
(c)) p(f
i1
(c))).
I.e., we have Herbrand approximations of (y)(p(f(y)) p(y)) taking values arbi-
trarily close to 1. This illustrates a more general phenomenon, captured by the following
approximate Herbrand theorem:
THEOREM 6.1.9. [=

( x)( x) iff for all r < 1:

[=
>r

i=1
(

t
i
) for some

t
1
, . . . ,

t
n
H().
This approximate Herbrand theorem has a nice corollary. Let = ( x)( y)( x, y)
where is both quantier-free and function-free. Then [=

iff [=

( y)( c, y) for
some new constants c. Let C be the (nite) set of constants occurring in ( y)( c, y),
adding one if the set is empty. Using the previous theorem:
[=

iff for all r < 1: [=

>r

_
n
i=1
( c,

t
i
) for some

t
1
, . . . ,

t
n
C
iff [=

dC
( c,

d).
But the validity problem for propositional ukasiewicz logic is decidable, so the validity
problem for function-free formulas ( x)( y)( x, y) and also the validity problem for
function-free one-bound-variable formulas of are decidable.
Now we can use the approximate Herbrand theorem to establish Skolemization for
. Since any formula has an equivalent prenex formula, both Skolemization and the
approximate Herbrand theorem hold for all formulas of .
272 George Metcalfe
THEOREM 6.1.10. Let be a formula and (Q y)( y) an equivalent prenex form for .
Let ( x)
F
( x) be the Skolem form of (Q y)( y). Then the following are equivalent:
(1) [=

(2) [=

(Q y)( y)
(3) [=

( x)
F
( x)
(4) For all r < 1: [=
>r

_
n
i=1

F
(

t
i
) for some

t
1
, . . . ,

t
n
H(
F
).
Now consider the system G, obtained by extending G with the standard rules
for the universal and existential quantiers. It is easily seen that the standard quantier
axioms are derivable; e.g.
(a) (a)
(ID)

(ID)
(a), , (a)
(MIX)
(a), (a)
()
A
(x)( ), (a)
()
(x)( ), (x)
()
(x)( ) (x)
()
(x)( ) ( (x))
()
G is sound with respect to the standard semantics for ; i.e., if d
G
, then
[=

. The other direction cannot hold. On the other hand, G extended with the
cut rule (CUT) is complete with respect to the algebraic semantics dened in terms of
MV-algebras. Unfortunately, cut elimination fails for G + (CUT); consider, e.g.
(z)p(z) (z)p(z)
(ID)
(z)p(z) (z)p(z)
()
p(a) p(a)
(ID)
(z)p(z) p(a)
()
p(a) p(a)
(ID)
p(b) p(b)
(ID)
(z)p(z) p(b)
()
(z)p(z), p(a) p(b), p(a)
(MIX)
(z)p(z) p(a) p(b), p(a)
()

(z)p(z) (y)(p(a) p(y)), p(a)

()
(z)p(z) (x)(y)(p(x) p(y)), p(a)
()
(z)p(z) (x)(y)(p(x) p(y)), (z)p(z)
()
(z)p(z) (z)p(z) (x)(y)(p(x) p(y))
()
A
(x)(y)(p(x) p(y))
(CUT)
No G-derivation exists for this sequent. A simple induction shows that in such a
derivation there would be a branch where any hypersequent ( satises that ( contains a
sequent with an occurrence of a subformula of (x)(y)(p(x) p(y)) on the right that
does not occur on the left in any sequent in (. But then the branch could not contain an
initial sequent, a contradiction.
To obtain a complete calculus the full rst-order ukasiewicz logic, we need an
innitary rule. Suppose that [=

( x)( x). Then for all k N

+
:
G
[]
k
where
=
_
n
i=1

F
(

t
i
) for some

t
1
, . . . ,

t
n
H(
F
). So using Lemma 6.1.7 k times, we can
apply the quantier and structural rules of G to derive ( []
k
) from ( []
k
).
Chapter III: Proof Theory for Mathematical Fuzzy Logic 273
THEOREM 6.1.11. Let be a prenex formula. Then [=

iff is derivable in G
extended with the rule:
[]
k
for all k N
+

EXAMPLE 6.1.12. Consider our earlier problematic formula (x)(y)(p(x) p(y)).
To derive this in our calculus we would have to perform an innite number of derivations
in G. E.g., in the case where n = 2, we have:
p(b) p(c)
()

p(b)
()

p(c)
()

p(b) p(b)
(ID)
, p(b) p(b), p(c)
(MIX)
p(b), p(b) p(c)
()

, p(a) p(b), p(b) p(c)

(W)
p(a) p(b), p(b) p(c)
()

p(a) p(b), (y)(p(b) p(y))

()
p(a) p(b), (x)(y)(p(x) p(y))
()
(y)(p(a) p(y)), (x)(y)(p(x) p(y))
()
(x)(y)(p(x) p(y)), (x)(y)(p(x) p(y))
()
6.2 Propositional quantiers
The addition of propositional quantiers p and pintuitivelyfor all/some propo-
sitions pcan be useful for expressing properties of truth value sets. For example,
rather than expressing density using a rule, we might use density axioms
(p)(( p) (p )) ( )
where p does not occur in or . More generally, this extra exibility can be used to
express topological properties of sets of real numbers (the intended semantics of a logic)
such as the existence of limit points or successor elements.
In classical logic, propositional quantiers are little more than a notational conve-
nience: (p)(p) and (p)(p) are interpreted by inma and suprema over the truth
value set 0, 1 and can therefore be dened as just () () and () ().
However, in non-classical logics propositional quantiers can increase the expressive
power of the language quite considerably, an important example being quantied in-
tuitionistic logic. For fuzzy logics, we will consider only proof-theoretic methods for
the already tricky case of quantied G odel logic. Let us make use of the language
/
c
= , , , , and add to the denition of a formula the condition that if is
a formula and p is a variable, then (p) and (p) are formulas. As in the rst-order
case, we distinguish between bound variables, denoted p, q, and free variables, denoted
a, b. We also write ( a) to denote that the free variables of are among those in a,
using () to denote distinguished occurrences of a formula in .
G-valuations are extended to such formulas as follows, where for a valuation v and
[0, 1], v[/p] is dened by v[/p](q) = if q = p and v[/p](q) = v(q) otherwise:
v((p)) = supv[/p]() [ [0, 1]
v((p)) = infv[/p]() [ [0, 1].
A formula is QG-valid, written [=
QG
, if v() = 1 for all G-valuations v.
274 George Metcalfe
A Hilbert system HQG for QG is obtained as an extension of an axiomatization for
G odel logic with the axioms and rules:
(Q) () (q)(q)
(Q) (q)(q) ()
(Q) (q)( ) ( (q)) q not occurring in
(QD) (q)(( q) (q )) ( ) q not occurring in or
(a)
(q)(q)
(Q)
(a)
(q)(q)
(Q)
where a does not occur in or .
It is straightforward to check that HQG is sound with respect to QG-validity. How-
ever, the crucial result for HQG is the following (we refer to the historical remarks for
references to the rather involved proof):
THEOREM 6.2.1. HQG admits quantier elimination: for every formula there is a
quantier-free formula whose variables occur in such that
HQG
.
Using the completeness of the ordinary axiomatization for G odel logic, [=
QG
iff

HQG
for any quantier-free formula . Hence, combining this with quantier elimi-
nation, we obtain
HQG
iff [=
QG
. In fact it follows from the proof of Theorem 6.2.1
that this last result holds for HQG even when the in (Q) and (Q) are restricted to
quantier-free formulas. Moreover, quantier elimination provides an easy proof of the
interpolation property for G odel logic.
COROLLARY 6.2.2. G odel logic G admits interpolation, i.e., if [=
G
, then
[=
G
and [=
G
for some formula whose variables occur in both
and .
Proof. Suppose that [=
G
( a,

b) (

b, c) where a and

b and

b and c are the distinct
variables occurring in and , respectively. Then easily [=
QG
( a,

b) ( p)( p,

b)
and [=
QG
( p)( p,

b) (

b, c). But then by quantier elimination, there exists a

propositional formula (

b) such that [=
QG
( p)( p,

b) (

b) as required.
A hypersequent calculus for quantied G odel logic is obtained by adding rules for
propositional quantiers to the calculus GG (where substitutions for variables are re-
stricted to quantier-free formulas) and, crucially in this case, the density rule. That is,
GQG consists of GG extended with:
( [ , ()
( [ , (p)(p)
()
Q
( [ (a)
( [ (p)(p)
()
Q
( [ , (a)
( [ , (p)(p)
()
Q
( [ ()
( [ (p)(p)
()
Q
( [
1
, a [
2
a
( [
1
,
2

(DENSITY)
where is quantier-free in ()
Q
and ()
Q
, and a does not occur
in the premises of ()
Q
or ()
Q
or the conclusion of (DENSITY).
Chapter III: Proof Theory for Mathematical Fuzzy Logic 275
EXAMPLE 6.2.3. We illustrate this calculus with a derivation of the density axioms (for
space reasons, omitting the easy derivations of , a a and a, a ):
a, a
, a a a , a
a , a [ a, a
(COM)
a , a
a , a [ ( a) (a ), a
()

( a) (a ), a [ ( a) (a ), a
()

( a) (a ), a [ (q)(( q) (q )), a
()
Q
(q)(( q) (q )), a [ (q)(( q) (q )), a
()
Q
(q)(( q) (q )), (q)(( q) (q )),
(DENSITY)
(q)(( q) (q )),
(C)
(q)(( q) (q ))
()
(q)(( q) (q )) ( )
()
Notice that the contraction rule (C) is essential for this derivation. To nd axioma-
tizations of quantied fuzzy logics lacking contraction, alternative density axioms are
required, or the presence of the density rule itself.
Soundness and completeness proofs for GQG follow the usual pattern. It is easy to
see that the quantier rules preserve validity, so an induction on the height of a deriva-
tion gives soundness, while completeness follows as before from the fact that the extra
axioms of HQG (restricting the in (Q) and (Q) to quantier-free formulas) are
derivable, and the extra rules are admissible.
THEOREM 6.2.4.
GQG
( iff
QG
I(().
Cut elimination also follows the same pattern as described above but here there are
a couple of important extra points to consider.
THEOREM 6.2.5. Cut elimination holds for GQG.
Proof. As in Theorem 4.1.1, it sufces to prove that for any hypersequent ( and hyper-
sequent 1 with marked formula :
If d

GQG
( and d
)

GQG
1, then
GQG
CUT((, 1).
However, this time we have to add an extra parameter to our induction hypothesis to cope
with the fact that the complexity of the cut formula treated can increase when stepping,
e.g., from (p)(p) to (). We let q() be the number of occurrences of quantiers in
, and prove the claim by induction on the lexicographically ordered quadruple:
q(), cp(), e(d
)
), h(d

).
Notice rst that we can assume, similarly to the rst-order case, that new variables
introduced by the density rule are completely new, i.e., do not occur elsewhere in the
derivations of ( and 1. Given this assumption, cases involving the density rule proceed
in the same way as for other structural rules. Where we really need the extra parameter
276 George Metcalfe
in the induction hypothesis is the case where both branches end with a quantier rule
applied to an occurrence of . E.g., for = (p)(p), we might have:
.
.
.
(
t
[ , (), [(p)(p)]
n1

(
t
[ , [(p)(p)]
n

()
Q
.
.
.
1
t
[ (a)
1
t
[ (p)(p)
()
Q
Consider a member of CUT((, 1) of the form (the only tricky case):
(
tt
[ ,
n

where (
tt
CUT((
t
, 1). Then by the induction hypothesis:

GQG
(
tt
[ , (),
n1

We can substitute for a in the derivation of 1
t
[ (a) (an easy induction)
to obtain a derivation of 1
t
[ (). But q(()) < q((p)(p)). So by the
induction hypothesis again and (EC),
GQG
(
tt
[ ,
n
as required.
Of course, density elimination, which holds for GG and GG, does not hold for GQG
since the density rule is needed to prove the (QD) axioms.
6.3 Modalities
Modal logics and their relatives such as description logics occupy an important
realm between propositional and rst-order classical logic. Ideally, they are expressive
enough to capture and reason about notions such as time, space, knowledge, etc., but
also have good computational properties such as decidability and reasonable complexity
bounds. Combining modal notions with fuzziness in the form of modal fuzzy logics
has been explored in a number of contexts such as representing and reasoning about
fuzzy beliefs, spatial reasoning in the presence of vagueness, fuzzy similarity measures,
and fuzzy description logics. General approaches have also been proposed for dealing
with many-valued modal logics, although most of the axiomatization and decidability
results so far have been limited to the nite-valued case. References are provided in the
historical remarks at the end of this chapter.
Proof theoretic methods for modal fuzzy logics have not as yet been developed with
any great degree of generality or uniformity, an exception to this being the case where
a modal operator is interpreted as a truth stresser; e.g., is interpreted as is
denitely true, very true, more true than false, etc. For such modalities, where the
intended interpretation of is a unary function on [0, 1] (or some other linearly ordered
set), we can extend the language /
p
with the single unary operator and add to our
hypersequent calculi rules such as:
( [
( [
()
( [ ,
( [ ,
()
( [
( [
()
where [
1
, . . . ,
n
] stands for [
1
, . . . ,
n
].
Chapter III: Proof Theory for Mathematical Fuzzy Logic 277
These rules are just hypersequent versions of rules familiar from sequent calculi for
the modal logics K, KT, K4, and S4. The standard necessitation rule for modal logics
corresponds to instances of () or () where and ( are empty, and the usual K
axioms are derivable using the core implication rules and () as follows:

(ID)

(ID)
,
()
( ),
()
( )
()
( ) ( )
()
Obviously T axioms of the form and 4 axioms of the form are
derivable using () and (), respectively. However, all is not quite what it seems
here. Shifting modality axioms of the form ( ) ( ), which are
not derivable even in the modal logic S4, are derivable here using () together with the
usual logical rules, (COM), (EW), and (EC) (the top hypersequent being derived as in
Example 3.2.2):
[
( ) [ ( )
()
( )
()
( ) ( )
()
In fact, the derivability of these axioms is needed for the calculi to be complete with
respect to linearly ordered algebras.
The behaviour of can be characterized further by adding modal versions of
structural rules such as:
G | , ,
G | ,
(C)

G |
G | ,
(W)

G |
1
,
G |
1
,
2
|
(SPLIT)

The structural rules (C)

and (W)

are hypersequent versions of rules used for the

exponential ! in linear logic, and may be used to provide embeddings of fuzzy logics into
other fuzzy logics extended with a suitable modality. The rule (SPLIT)

, on the other
hand, is used to provide proof systems for fuzzy logics extended with the globalization
modality dened as x = 1 if x = 1, x = 0 otherwise.
Given a hypersequent calculus GL, we can dene, for example, the systems
GLK
r
= GL + () GLKT
r
= GLK
r
+ ()
GLK4
r
= GLK
r
+ () GLS4
r
= GLK
r
+ () + ()
GL

= GLS4
r
+ (SPLIT)

GL! = GLS4
r
+ (C)

+ (W)

where
r
denotes that these systems are complete with respect to linearly ordered alge-
bras. Moreover, if GL is an extension of GUL with substitutive single-conclusion rules
or extension of GIUL with substitutive rules, then the extended calculi also admit cut
elimination.
278 George Metcalfe
In general, however, modal fuzzy logics are based not solely on functions on [0, 1],
but rather on some variant of the Kripke frames and models appearing in classical modal
logics. Here, the challenge is to prove that a proposed Hilbert or Gentzen system is com-
plete with respect to the given semantics. Since the only proof-theoretic results in this
direction have been based on G odel logic, we conne our attention to this case, making
use of a language /
.
that extends /
c
with the unary operators and . Following gen-
eral approaches described in the literature, G odel modal logics are generalizations of the
modal logic K where connectives behave locally at individual worlds as in G odel logic.
In particular, GK and GK
F
are dened as modal G odel fuzzy logics based on standard
Kripke frames and Kripke frames with fuzzy accessibility relations, respectively.
We dene a fuzzy Kripke frame to be a pair F = W, R where W is a non-empty
set of worlds and R: W W [0, 1] is a binary fuzzy accessibility relation on W. If
Rxy 0, 1 for all x, y W, then Ris called crisp and F is called simply a (standard)
Kripke frame. In this case, we may consider R W
2
and write Rxy or (x, y) R to
mean Rxy = 1. Next, a Kripke model for GK
F
is a 3-tuple K = W, R, V where
W, R is a fuzzy Kripke frame and V : Fm
1

W [0, 1] is a mapping, called a

GK
F
-valuation, satisfying
V (, x) = 1
V (, x) = 0
V ( , x) = V (, x)
G
V (, x)
V ( , x) = min(V (, x), V (, x))
V ( , x) = max(V (, x), V (, x))
V (, x) = infRxy
G
V (, y) [ y W
V (, x) = supmin(V (, y), Rxy) [ y W.
A Kripke model for GK satises the extra condition that W, R is a standard Kripke
frame. In this case, the conditions for and may also be read as:
V (, x) = inf(1 V (, y) [ Rxy)
V (, x) = sup(0 V (, y) [ Rxy).
A formula is valid in W, R, V if V (, x) = 1 for all x W, and L-valid for
L GK, GK
F
, written [=
L
, if is valid in all Kripke models W, R, V for L.
Developing proof-theoretic methods for the full languages of these logics remains
an open problem, as is nding axiomatizations and establishing decidability. Instead, let
us consider the box and diamond fragments of GKand GK
F
based on restrictions of
the language /
.
to the single modality sublanguages /

and /
.
, respectively. Proof-
theoretic methods have been used to establish decidability and complexity results for
these fragments, as well as completeness for the following axiomatizations:
HGK

is HG extended with
(K

) ( ) ( )
(Z

(NEC)

Chapter III: Proof Theory for Mathematical Fuzzy Logic 279

HGK
.
is HG extended with
(K
.
) ( ) ( )
(Z
.
)
(F
.
)
( )
( )
(NEC)
.
HGK
F
.
is HGK
.
with (NEC)
.
replaced by


(NEC)

.
The diamond fragment of GK
F
has the nite model property, and so is decidable, but
this is not the case either for the diamond fragment of GK or the box fragment of GK,
which coincides with the box fragment of GK
F
.
To provide a uniform proof-theoretic presentation of these fragments, we turn to the
sequent of relations framework introduced in the previous section. Let us dene SG to
consist of the logical rules of Figure 7 extended with the following rules:
( [
(ID)
( [ [ <
(
(CS)
(
( [
(EW)
( [ [
( [
(WL)
( [ [
( [
(WR)
( [
1

1

1
[
2

2

2
[
1

2
( [
1

1

1
[
2

2

2
[
2

1

1
( [
1

1

1
[
2

2

2
(COM)
We adopt the following helpful notation for sets of relations:
[
1
, . . . ,
n
] =
def

1
[ . . . [
n

[] =
def

[] < =
def

[
1
, . . . ,
m
] =
def

1
[ . . . [
m
[] =
def

< [] =
def
.
Note that we always restrict expressions to cases where either or has at most
one element. We remark, moreover, that and can be considered as sets of formulas
rather than multisets without changing the meaning of the notation.
Now we dene sequent of relations calculi for fragments of G odel modal logics:
SGK

consists of SG (for /

) extended with:
[
( [ [
()
SGK
.
consists of SG (for /
.
) extended with:
[ < [
( [ [ < [
()
280 George Metcalfe
SGK
F
.
consists of SG (for /
.
) extended with:
[ <
( [ [ <
()

EXAMPLE 6.3.1. Consider the following derivation of a sequent of relations corre-

sponding to the axiom (p q) (p q):
p q [ q < p [ p p
(ID)
p q [ q < p [ q q
(ID)
p q [ q < p
(COM)
p q [ q [ q < p
(EW)
p q [ q q
(ID)
p q [ p q q
()
p q [ (p q) q
()
(p q) p q
()
(p q) p q [ p q
(EW)
(p q) (p q)
()
Soundness is relatively straightforward to establish for these calculi but complete-
ness is rather involved. We note merely that decidability is an immediate consequence of
completeness since for any given sequent of relations there is a nite number of sequent
of relations that can occur in a derivation of it in one of these calculi. Moreover, with
a little more analysis, the calculi can be used to show that each of the fragments is in
PSPACE and hence PSPACE-complete.
We note nally that in the case of the box fragment, a hypersequent calculus GGK

admitting cut elimination is obtained by extending GG with the rule

[
( [ [
()
EXAMPLE 6.3.2. The (Z

) axioms are derivable in GGK

as follows:
p p
(ID)
p, p
()
p p
(ID)
p, p
()
p, p [ p, p
(COM)
p, p [ p
(C)
p [ p
(C)
p [ p
()
p [p
()
p [ p
()
p [ p
()
p [ p p
(W)
p p [ p p
(W)
p p
(EC)
p p
()
Chapter III: Proof Theory for Mathematical Fuzzy Logic 281
7 Historical remarks and further reading
For the history and development of mathematical fuzzy logic, we refer the reader to
other chapters of this handbook and to the monographs [41, 44, 53]. For structural proof
theory, the crucial event, both historically and scientically, arrived in the 1930s with
Gentzens introduction of (and cut elimination proofs for) the sequent calculi LK and LJ
for rst-order classical logic and intuitionistic logic, respectively, to which can be traced
in particular the rst proofs of Theorems 2.2.2, 2.2.3, and 2.3.3 [38]. Gentzen made use
of LJ and LK as tools for investigating the consistency of mathematical theories, and
indeed these calculi continue to play a central role in the eld of proof theory. See, e.g.,
the handbook [22] and monographs [64, 67] for more details.
So-called Gentzen systems for substructural and other non-classical logics were
subsequently developed in a wide range of contexts. In linguistics, a sequent calculus
was dened by Lambek in the 1950s to model the assignment of types such as adjec-
tive or verb phrase to strings of words in natural language [45]. Since the order (of
words or types) as well as multiplicity is crucial in this case, sequents are (as they were
for Gentzen) ordered pairs of sequences of formulas, and the calculus lacks not only
weakening and contraction rules but also exchange rules for permuting formulas in se-
quents. A second important source of substructural logics was the family of relevance
logics developed for philosophical reasons by Anderson and Belnap and co-workers
from the 1960s onwards (see in particular [3, 4]), where relevance is mirrored proof-
theoretically by dropping weakening rules. Conversely, logics without contraction were
introduced by Grishin in the 1980s [42] with the aim of avoiding set-theoretic paradoxes
involving the comprehension principle, and investigated extensively by Ono and Komori
in [58]. Girards linear logic [40], introduced in 1987 to model computational processes,
drops both weakening and contraction rules but allows these to be recovered for certain
formulas using special modal operators ! and ?. Comprehensive introductions to sequent
calculi and other features of substructural logics and their algebraic semantics may be
found in [37, 59, 62].
Hypersequents were formally introduced by Avron in 1987 as a suitable proof-
theoretic framework for the relevant-mingle logic RM [6]; similar structures also ap-
peared independently in a calculus dened by Pottinger for the modal logic S5 [60].
Avron subsequently introduced hypersequent calculi for three-valued ukasiewicz logic
and, crucially for the study of fuzzy logics, G odel logic in [7]. These ideas were taken
up, extended, and generalized in various directions by researchers clustered around the
logic group at the Technical University of Vienna. In particular, Baaz and Zach ex-
tended Avrons calculus to rst-order G odel logic in 2000, using this system to establish
the mid-hypersequent theorem and hence the Herbrand theorem for the prenex fragment
of the logic [18]. These authors also provided a rst syntactic elimination of the density
rule, previously used by Takeuti and Titani to axiomatize rst-order G odel logic (intu-
itionistic fuzzy logic) in [65]. Further results on Herbrands theorem and Skolemization
in G odel logic may be found in [9], and on the extension of G odel logic with proposi-
tional quantiers (in particular proofs of Theorems 6.2.1, 6.2.4, and 6.2.5) in [12, 17]
(see also [10, 16]).
282 George Metcalfe
The proof-theoretic approach to fuzzy logics was extended to Godo and Estevas
monoidal t-norm logic MTL by Baaz, Ciabattoni, and Montagna in 2004 [11]; a hy-
persequent calculus is obtained by dropping contraction from the calculus for G odel
logic. Complementing this picture, hypersequent calculi were dened also for other log-
ics with weakening, including IMTL, SMTL, MTL
n
, and IMTL
n
(n 3) by various
authors in [25, 27, 28]. Around the same time, Metcalfe extended this picture further
by introducing uninorm logic UL, obtained proof-theoretically by dropping weakening
rules from the calculus for MTL, and related logics such as IUL, UML, and IUML [46]
(see also [36, 49]). These works improved on Avrons original cut elimination proofs
(based on a rather complicated history method) by developing a Sch utte-Tait-style
approach, that eliminates largest cuts in derivations. An alternative cut elimination by
substitutions method was also dened by Ciabattoni in [24] and used to obtain uniform
proofs for single-conclusion hypersequent calculi. The related uniform proof of Theo-
rem 4.1.1 may be found (in a more general context allowing alternative rules for con-
nectives) together with example applications of cut elimination such as Theorems 4.1.2
and 4.1.3 in the monograph [53]. These approaches contributed, moreover, to a grow-
ing literature on syntactic and semantic classications of proof calculi for which the
cut rule is admissibleincluding, in particular, an algebraic method developed for se-
quent calculi by Okada and Terui in [56, 57, 66]. This latter approach has in turn led
to the introduction by Ciabattoni, Galatos, and Terui in 2008 [29] of an algorithm that
transforms axiom schema in certain syntactic classes into either sequent or hypersequent
(depending on the class) single-conclusion rules preserving cut elimination (extended to
multiple-conclusion calculi by Ciabattoni, Strassburger and Terui in [33]). As an ex-
ample of the algorithm, this paper introduced a calculus for weak nilpotent minimum
logic WNM. A calculus for nilpotent minimum logic NM was obtained as an extension
of the hypersequent version of a calculus for constructive logic with strong negation by
Metcalfe in [48].
As noted above, elimination of a hypersequent version of Takeuti and Titanis den-
sity rule was rst obtained for rst-order G odel logic by Baaz and Zach in 2000 using a
Gentzen-style proof that shifts applications of the rule upwards in derivations [18]. This
method was extended to a wide range of hypersequent calculi by Metcalfe and Montagna
in 2007 and used to establish standard completeness for propositional fuzzy logics [49].
The more elegant density elimination by substitutions method described in this chap-
ter (Theorems 4.2.3 and 4.2.5), where applications of the density rule are removed by
making suitable substitutions, was introduced by Ciabattoni and Metcalfe in [31] and
used to provide classications of standard complete (rst-order) fuzzy logics.
Earlier pre-hypersequent work on proof theory for fuzzy logics focussed mostly
on G odel logic and ukasiewicz logic. A rst (rather complicated) sequent calculus
for the former was dened by Sonobe in 1975 [63], and improved versions (terminating
and contraction-free) were later developed by Avellone, Ferrari, and Miglioli [5], and
Dyckhoff [34]. The sequent calculus presented here (and Theorem5.1.4) appears in [53],
adapted from calculi dened by Avron and Konikowska in 2001 [8], while the sequent of
relations approach (including a proof of Theorem 5.1.6), which applies to a wider class
of projective logics, was developed by Baaz and Ferm uller in the 1999 paper [13]. For
ukasiewicz logic, the rst calculi introduced made use either of the cut rule [1, 61], ex-
Chapter III: Proof Theory for Mathematical Fuzzy Logic 283
tra syntax [43, 69], or a translation into nite-valued logics [2]. The elegant analytic se-
quent and hypersequent calculi dened above were developed by Metcalfe, Olivetti, and
Gabbay in 2005 [52], alongside calculi for Meyer and Slaneys Abelian logic, the logic
of lattice-ordered Abelian groups [54] (see also [30, 47]). Similar calculi for product
logic and cancellative hoop logic were introduced by the same authors in [51]. The rela-
tionship of derivations in G to (disjunctive) strategies for Giless game (introduced by
Giles in the 1970s [39]) is explored in detail by Fermu uller and Metcalfe in [35], while
the uniform r-hypersequent approach for G odel logic, ukasiewicz logic, and product
logic was developed by Ciabattoni, Ferm uller, and Metcalfe in [26].
More details of the general proof-theoretic approach to rst-order fuzzy logics de-
scribed in Section 6.1, in particular proofs of Theorems 6.1.3, 6.1.4, 6.1.5, and 6.1.8,
may be found in the monograph [53], generalizing earlier results given in the context
of rst-order G odel logic [18] and monoidal t-norm logic [11]. The proof-theoretic re-
sults described here for rst-order ukasiewicz logic are taken from the papers of Baaz
and Metcalfe [14, 15]. Modal fuzzy logics and the related topic of fuzzy description
logics are currently the subject of intensive investigation, in particular by researchers in
Barcelona (see, e.g., [19, 20]). The general proof-theoretic approach for adding modal-
ities to hypersequent calculi that results in modal fuzzy logics complete with respect to
chains is described in [32], while the particular case of fragments of G odel modal logics
(introduced in [23]) is considered in [50].
Finally, note that no Gentzen system has been presented in this chapter for H ajeks
basic logic BL, the logic of continuous t-norms. Basic logic is one of the most important
and widely studied fuzzy logics, and from an algebraic perspective, is rather natural: the
variety of BL-algebras is generated not only by all standard (continuous t-norm based)
BL-algebras, but even by just one such algebra. In fact this algebra provides one route
to dening a calculus of sorts. Information about the positioning of the valuations of the
formula can be encoded by structural features of the calculus and ultimately the question
of the validity of a formula in BL can be reduced (as in ukasiewicz logic) to solving
linear programming problems. This approach, developed in [21, 55], does not provide
an elegant calculus for BL but does at least give a reasonable algorithm for deciding
questions of validity in this logic. An alternative approach by Vetterlein [68], based on
relational hypersequents, gives an elegant calculus for the product-free fragment of BL,
but requires the addition of an extra modal operator in order to capture the full logic.
Acknowledgements
The author acknowledges support from Marie Curie Reintegration grant PIRG06-
GA-2009-256492 and Swiss National Science Foundation grant 20002 129507, and
would also like to thank Agata Ciabattoni and Kazushige Terui for their very helpful
comments and suggestions.
284 George Metcalfe
BIBLIOGRAPHY
[1] Alan Adamson and Robin Giles. A game-based formal system for

. Studia Logica, 38(1):4973,

1979.
[2] Stefano Aguzzoli and Agata Ciabattoni. Finiteness in innite-valued logic. Journal of Logic, Language
and Information, 9(1):529, 2000.
[3] Alan Ross Anderson and Nuel D. Belnap. Entailment: The Logic of Relevance and Necessity, volume 1.
Princeton University Press, Princeton, 1975.
[4] Alan Ross Anderson, Nuel D. Belnap, and J. Michael Dunn. Entailment: The Logic of Relevance and
Necessity, volume 2. Princeton University Press, Princeton, 1992.
[5] Alessandro Avellone, Mauro Ferrari, and Pierangelo Miglioli. Duplication-free tableau calculi and re-
lated cut-free sequent calculi for the interpolable propositional intermediate logics. Logic Journal of the
Interest Group of Pure and Applied Logic, 7(4):447480, 1999.
[6] Arnon Avron. A constructive analysis of RM. Journal of Symbolic Logic, 52(4):939951, 1987.
[7] Arnon Avron. Hypersequents, logical consequence and intermediate logics for concurrency. Annals of
Mathematics and Articial Intelligence, 4(34):225248, 1991.
[8] Arnon Avron and Beata Konikowska. Decomposition proof systems for G odel-Dummett logics. Studia
Logica, 69(2):197219, 2001.
[9] Matthias Baaz, Agata Ciabattoni, and Christian Georg Ferm uller. Herbrand theorem for prenex G odel
logic and its consequences for theorem proving. In Robert Nieuwenhuis and Andrei Voronkov, editors,
Logic for Programming, Articial Intelligence, and Reasoning, volume 2250/2001 of Lecture Notes in
Computer Science, pages 201216, 2001.
[10] Matthias Baaz, Agata Ciabattoni, and Christian Georg Ferm uller. Hypersequent calculi for G odel logics:
A survey. Journal of Logic and Computation, 13(6):835861, 2003.
[11] Matthias Baaz, Agata Ciabattoni, and Franco Montagna. Analytic calculi for monoidal t-norm based
logic. Fundamenta Informaticae, 59(4):315332, 2004.
[12] Matthias Baaz, Christian G. Ferm uller, and Helmut Veith. An analytic calculus for quantied propo-
sitional G odel logic. In Proceedings of TABLEAUX 2000, volume 1847 of Lecture Notes in Computer
Science, pages 112126. Springer, 2000.
[13] Matthias Baaz and Christian Georg Ferm uller. Analytic calculi for projective logics. In N.V. Murray,
editor, Automated Reasoning with Analytic Tableaux and Related Methods, TABLEAUX99, volume
1617 of Lecture Notes in Computer Science, pages 3650. Springer, Saratoga Springs, 1999.
[14] Matthias Baaz and George Metcalfe. Herbrand theorems and Skolemization for prenex fuzzy logics. In
Proceedings of CiE 2008, volume 5028 of Lecture Notes in Computer Science, pages 2231. Springer,
2008.
[15] Matthias Baaz and George Metcalfe. Herbrands theorem, Skolemization, and proof systems for rst-
order ukasiewicz logic. Journal of Logic and Computation, 20(1):3554, 2010.
[16] Matthias Baaz, Norbert Preining, and Richard Zach. First-order G odel logics. Annals of Pure and
Applied Logic, 147(12):2347, 2007.
[17] Matthias Baaz and Helmut Veith. An axiomatization of quantied propositional G odel logic using the
Takeuti-Titani rule. In Logic Colloquium 1999, volume 13 of Lecture Notes in Logic, pages 91104,
2000.
[18] Matthias Baaz and Richard Zach. Hypersequents and the proof theory of intuitionistic fuzzy logic. In
Peter G. Clote and Helmut Schwichtenberg, editors, Proceedings of 14th CSL Workshop, volume 1862
of Lecture Notes in Computer Science, pages 187201, Berlin, 2000. Springer-Verlag.
[19] F elix Bou, Francesc Esteva, Josep Maria Font,
`
Angel Gil, Llus Godo, Antoni Torrens, and Ventura
Verd u. Logics preserving degrees of truth from varieties of residuated lattices. Journal of Logic and
Computation, 19(6):10311069, 2009.
[20] F elix Bou, Francesc Esteva, Llus Godo, and Ricardo O. Rodrguez. On the minimum many-valued
modal logic over a nite residuated lattice. Journal of Logic and Computation, 21(5):739790, 2011.
[21] Simone Bova and Franco Montagna. Proof search in H ajeks Basic Logic. ACM Transactions on Com-
putational Logic, 9(3), 2008.
[22] Sam Buss, editor. Handbook of Proof Theory. Kluwer, 1998.
[23] Xavier Caicedo and Ricardo O. Rodrguez. Sandard G odel modal logics. Studia Logica, 94(2):189214,
2010.
[24] Agata Ciabattoni. Automated generation of analytic calculi for logics with linearity. In Jerzy
Marcinkowski and Andrzej Tarlecki, editors, Proceedings of 18th CSL Workshop, volume 3210 of Lec-
ture Notes in Computer Science, pages 503517, Berlin, 2000. Springer-Verlag.
Chapter III: Proof Theory for Mathematical Fuzzy Logic 285
[25] Agata Ciabattoni, Francesc Esteva, and Llus Godo. T-norm based logics with n-contraction. Neural
Network World, 12(5):453460, 2002.
[26] Agata Ciabattoni, Christian G. Ferm uller, and George Metcalfe. Uniform rules and dialogue games for
fuzzy logics. In Proceedings of LPAR 2004, volume 3452 of Lecture Notes in Articial Intelligence,
pages 496510. Springer, 2005.
[27] Agata Ciabattoni and Mauro Ferrari. Hypersequent calculi for some intermediate logics with bounded
Kripke models. Journal of Logic and Computation, 11(2):283294, 2001.
[28] Agata Ciabattoni, Dov M. Gabbay, and Nicola Olivetti. Cut-free proof systems for logics of weak
excluded middle. Soft Computing, 2(4):147156, 1999.
[29] Agata Ciabattoni, Nikolaos Galatos, and Kazushige Terui. From axioms to analytic rules in nonclassical
logics. In Proceedings of the 23rd Annual IEEE Symposium on Logic in Computer Science (LICS), pages
229240, 2008.
[30] Agata Ciabattoni and George Metcalfe. Bounded ukasiewicz logics. In M. Cialdea Mayer and F. Pirri,
editors, Proceedings of TABLEAUX 2003, volume 2796 of Lecture Notes in Computer Science, pages
3248. Springer, 2003.
[31] Agata Ciabattoni and George Metcalfe. Density elimination. Theoretical Computer Science, 403(1
2):328346, 2008.
[32] Agata Ciabattoni, George Metcalfe, and Franco Montagna. Algebraic and proof-theoretic characteriza-
tions of truth stressers for MTL and its extensions. Fuzzy Sets and Systems, 161(3):369389, 2010.
[33] Agata Ciabattoni, Lutz Strassburger, and Kazushige Terui. Expanding the realm of systematic proof
theory. In Erich Gr adel and Reinhard Kahle, editors, Proceedings of 23rd CSL Workshop, volume 5771
of Lecture Notes in Computer Science, pages 163178, Berlin, 2009. Springer.
[34] Roy Dyckhoff. A deterministic terminating sequent calculus for G odel-Dummett logic. Logic Journal
of the Interest Group of Pure and Applied Logic, 7(3):319326, 1999.
[35] Christian G. Ferm uller and George Metcalfe. Giless game and proof theory for ukasiewicz logic.
Studia Logica, 92(1):2761, 2009.
[36] Dov M. Gabbay and George Metcalfe. Fuzzy logics based on [0, 1)-continuous uninorms. Archive for
Mathematical Logic, 46(6):425469, 2007.
[37] Nikolaos Galatos, Peter Jipsen, Tomasz Kowalski, and Hiroakira Ono. Residuated Lattices: An Alge-
braic Glimpse at Substructural Logics, volume 151 of Studies in Logic and the Foundations of Mathe-
matics. Elsevier, Amsterdam, 2007.
[38] Gerhard Gentzen. Untersuchungen uber das Logische Schlieen I. & II. Mathematische Zeitschrift,
39(1):176210, 405431, 1935.
[39] Robin Giles. ukasiewicz logic and fuzzy set theory. International Journal of ManMachine Studies,
8(3):313327, 1976.
[40] Jean-Yves Girard. Linear logic. Theoretical Computer Science, 50(1):1102, 1987.
[41] Siegfried Gottwald. A Treatise on Many-Valued Logics, volume 9 of Studies in Logic and Computation.
Research Studies Press, Baldock, 2001.
[42] Vyacheslav Nikolaevich Grishin. Predicate and set-theoretic calculi based on logic without contractions.
Mathematics of the USSRIzvestiya, 18:4159, 1982.
[43] Reiner H ahnle. Automated Deduction in Multiple-Valued Logics, volume 10 of International Series of
Monographs on Computer Science. Oxford University Press, 1993.
[44] Petr H ajek. Metamathematics of Fuzzy Logic, volume 4 of Trends in Logic. Kluwer, Dordrecht, 1998.
[45] Joachim Lambek. The mathematics of sentence structure. American Mathematical Monthly, 65:154
170, 1958.
[46] George Metcalfe. Proof Theory for Propositional Fuzzy Logics. PhD thesis, Department of Computer
Science, Kings College, London, 2004.
[47] George Metcalfe. Proof calculi for Casaris comparative logics. Journal of Logic and Computation,
16(4):405422, 2006.
[48] George Metcalfe. A sequent calculus for constructive logic with strong negation as a substructural logic.
Bulletin of the Section of Logic, 38(1):17, 2009.
[49] George Metcalfe and Franco Montagna. Substructural fuzzy logics. Journal of Symbolic Logic,
72(3):834864, 2007.
[50] George Metcalfe and Nicola Olivetti. Towards a proof theory of G odel modal logics. Logical Methods
in Computer Science, 7:127, 2011.
[51] George Metcalfe, Nicola Olivetti, and Dov Gabbay. Analytic proof calculi for product logics. Archive
for Mathematical Logic, 43(7):859889, 2004.
[52] George Metcalfe, Nicola Olivetti, and Dov Gabbay. Sequent and hypersequent calculi for Abelian and
ukasiewicz logics. ACM Transactions on Computational Logic, 6(3):578613, 2005.
286 George Metcalfe
[53] George Metcalfe, Nicola Olivetti, and Dov M. Gabbay. Proof Theory for Fuzzy Logics, volume 36 of
Applied Logic Series. Springer, 2008.
[54] Robert K. Meyer and John K. Slaney. Abelian logic from A to Z. In Graham Priest, Richard Routley,
and Jean Norman, editors, Paraconsistent Logic: Essays on the Inconsistent, pages 245288. Philosophia
Verlag, 1989.
[55] Franco Montagna, G. Michele Pinna, and Elisa B.P. Tiezzi. A tableau calculus for H ajeks logic BL.
Journal of Logic and Computation, 13(2):241259, 2003.
[56] Mitsuhiro Okada. Phase semantics for higher order completeness, cut-elimination and normalization
proofs. Electronic Notes in Theoretical Computer Science, 3, 1996.
[57] Mitsuhiro Okada and Kazushige Terui. The nite model property for various fragments of intuitionistic
linear logic. Journal of Symbolic Logic, 64(2):790802, 1999.
[58] Hiroakira Ono and Yuichi Komori. Logic without the contraction rule. Journal of Symbolic Logic,
50(1):169201, 1985.
[59] Francesco Paoli. Substructural Logics: A Primer, volume 13 of Trends in Logic. Kluwer, Dordrecht,
2002.
[60] Garrel Pottinger. Uniform, cut-free formulations of T, S4 and S5 (abstract). Journal of Symbolic Logic,
48(3):900, 1983.
[61] Andreja Prijatelj. Bounded contraction and Gentzen-style formulation of ukasiewicz logics. Studia
Logica, 57(23):437456, 1996.
[62] Greg Restall. An Introduction to Substructural Logics. Routledge, New York, 2000.
[63] O. Sonobe. A Gentzen-type formulation of some intermediate propositional logics. Journal of Tsuda
College, 7:714, 1975.
[64] Gaisi Takeuti. Proof Theory, volume 81 of Studies in Logic and the Foundations of Mathematics. North
Holland, Amsterdam, 2nd edition, 1987.
[65] Gaisi Takeuti and Satoko Titani. Intuitionistic fuzzy logic and intuitionistic fuzzy set theory. Journal of
Symbolic Logic, 49(3):851866, 1984.
[66] Kazushige Terui. Which structural rules admit cut elimination? an algebraic criterion. Journal of
Symbolic Logic, 72(3):738754, 2007.
[67] Anne S. Troelstra and Helmut Schwichtenberg. Basic Proof Theory. Cambridge University Press, 2
edition, 2000.
[68] Thomas Vetterlein. Analytic calculi for logics of ordinal multiples of standard t-norms. Journal of Logic
and Computation, 18(1):3557, 2008.
[69] Hubert Wagner. A new resolution calculus for the innite-valued propositional logic of ukasiewicz. In
Proceedings of FTP 98, Technical Report E1852-GS-981, pages 234243. TU Wien, Austria, 1998.
GEORGE METCALFE
Mathematical Institute
University of Bern
Sidlerstrasse 5
CH-3012 Bern, Switzerland
Email: george.metcalfe@math.unibe.ch
Chapter IV
Algebraic Semantics: Semilinear FL-Algebras
ROSTISLAV HOR

IK
1 Substructural logics
Fuzzy logics were originally motivated by semantical considerations as logics whose
intended set of truth values is the real unit interval [0, 1]. Later this was relaxed and [0, 1]
was replaced by an arbitrary linearly ordered set. Thus by a fuzzy logic we mean a logic
complete with respect to a class of linearly ordered algebras. Since the notion of a fuzzy
logic has not been precisely established, we will call such a logic semilinear. The term
semilinear comes from the fact that the algebraic semantics for such a logic contains also
algebras which are not linearly ordered but their building blocks (subdirectly irreducible
algebras) are linearly ordered.
In order to discuss an algebraic semantics for semilinear logics, we have to spec-
ify precisely which logics we are going to study. We will put semilinear logics into
a framework of substructural logics because most of the semilinear logics, studied so
far, belong to the hierarchy of substructural logics. Note that in Chapter II semilinear
logics are studied in a more general framework of weakly implicative logics. To keep
the text within a reasonable size, we have to make several design choices. First, we
will restrict ourselves only to logics in the full language of substructural logics. Thus
we will not mention results on various fragments. Second, as a base substructural logic
we consider the full Lambek calculus. This logic was dened in Chapter II as a weakly
implicative logic and its Hilbert style calculus was given. Here we present FL by means
of a Gentzen sequent system (see also Chapter III). Then semilinear logics appear as
axiomatic extensions of FL.
Given a set Q, we denote its powerset by T(Q). Recall that a consequence relation
on a set Q is a binary relation T(Q) Q such that for every X Y x, z Q
we have
X x for all x X,
if X y for all y Y and Y z, then X z.
Given a set of formulas Fm in a language /, a logic L for us is a consequence relation

L
T(Fm) Fm on the set of formulas Fm which is substitution invariant, i.e.,
if
L
, then []
L
[] for all substitutions in the language /.
288 Rostislav Hor ck
, ,
(cut)
, ,
,
(1)
, 1,

(0)
0
, ,
()
, ,
, ,
()
, ,

()


()


()

, , , ,
()
, ,
, , ,
( )
, ,

()
,
, ,
()
, , ,
,
()

, ,
(/)
, /, ,
,
(/)
/
Figure 1. The rules of the sequent calculus for the logic FL.
The language /
FL
of the logic FL consists of a countable set of propositional vari-
ables, binary connectives , , , , /, and constants 0, 1. The binary connectives are re-
spectively called lattice conjunction, lattice disjunction, fusion, left and right implication
(or division). Thus from now on Fm denotes the set of formulas in the language /
FL
.
In order to dene
FL
, we have to say what a sequent is and what it means that a
sequent is provable from a set of sequents. A sequent is an ordered pair , where
is a sequence of formulas (could be also empty) and a formula or the empty sequence.
A sequent is called initial if it has a form of one of the following sequents:
1 0
where Fm. Given a set of sequents S and a sequent s, we say that s is provable
from S in the sequent calculus for FL if one of the following conditions is satised:
s S,
s is initial,
s can be obtained from S and the initial sequents by application of nitely many
rules from Figure 1, where symbols , , denote sequences of formulas and
, , formulas.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 289

()

()
( ) 1
()
( ) 1
Figure 2. The proof of (( )) 1 from .
Now we are ready to dene
FL
. Let Fm. Then
FL
holds iff
the sequent is provable from the set of sequents [ . Analogously
we say that is provable from (also proves ) in the logic FL if
FL
. We
will illustrate the above notions on an example. Let and be formulas. Then Figure 2
shows that
FL
(( )) 1.
Let be a set of formulas understood as axiom schemata. An axiomatic extension
FL + of FL by the set of axiom schemata is the logic dened by the same sequent
calculus as FL enriched by new initial sequents
() [ and is a substitution in the language /
FL
.
The consequence relation
FL+
is dened analogously as
FL
using this sequent calcu-
lus. Many of the substructural logics (among them also semilinear logics) can be viewed
as axiomatic extensions of FL. Thus in this text we will call a logic L substructural if L
is an axiomatic extension of FL.
Table 1 shows a few basic axiom schemata together with their names. Let S
e, c, i, o. We denote by FL
S
the axiomatic extension of FL by axiom schemata from
S, e.g. FL
ci
denotes axiomatic extension of FL by (c) and (i). Note that (w) is an
abbreviation for (i) together with (o). Thus we write for instance FL
ew
instead of FL
eio
.
The logics FL
S
are called basic substructural logics.
Name Axiom schema(ta)
(e) ( )( )
(c) ( )
(i) 1
(o) 0
(w) 1, 0
Table 1. Axioms schemata of basic substructural logics.
REMARK 1.0.1. We have introduced the basic substructural logics FL
S
as axiomatic
extensions of FL. In proof theory they are usually equivalently presented as extensions
of FL by combinations of corresponding structural rules of exchange (e), contraction
(c), left weakening (i) and right weakening (o), see Figure 3. Thus FL
ecw
is nothing else
but intuitionistic logic.
290 Rostislav Hor ck
, , ,
(e)
, , ,
, , ,
(c)
, ,
,
(i)
, ,

(o)

Figure 3. The structural rules of exchange (e), contraction (c), left weakening (i) and
right weakening (o).
FL
FL
ew
FL
i
FL
ei
FL
ec
FL
e
FL
c
FL
eo
FL
eco
FL
co
FL
o
FL
w
FL
ci
= FL
eci
FL
cw
= FL
ecw
Figure 4. Basic substructural logics.
Substructural logics can be partially ordered by their strength. More precisely, this
partial order is just the inclusion ordering, i.e., a logic L
2
is stronger than a logic L
1
if
L
1

L
2
. In fact, this order is a complete lattice order, i.e., the substructural
logics form a complete lattice (FL). The join
_
iI
L
i
of a collection of substructural
logics L
i
[ i I, where L
i
= FL +
i
for some
i
Fm, is the substructural
logic axiomatized by the union of
i
s, i.e.,
_
iI
L
i
= FL +

iI

i
. A description
of meets is not so easy but it can be done as shown in Chapter II. Further, note that
FL is the bottom element of (FL). The top element is the inconsistent logic which
proves everything. Close to the top element is located classical logic which is one of the
maximally consistent substructural logics. Figure 4 depicts the ordering from (FL)
for basic substructural logics. Note that FL
ci
= FL
eci
and FL
cw
= FL
ecw
since (e) can
be proved in the presence of (c) and (i).
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 291
As we mentioned at the beginning we are interested in substructural logics which
are complete with respect to a class of linearly ordered algebras, so-called semilinear
logics. Since this is mainly a semantical notion, we postpone a formal denition of
semilinear logics until we dene algebraic semantics for substructural logics.
2 Algebraic preliminaries
We have to recall several denitions and results. First, we recall basic facts from
universal algebra. Second, we recall the notion of a residuated map which is an essential
ingredient of our algebraic semantics for substructural logics. Finally, a closely related
notion of a Galois connection is discussed. The sets of natural numbers, integers, rational
and real numbers are denoted N, Z, Q, R respectively.
2.1 Universal algebra
We assume the reader is familiar with the basics of universal algebra, most of which
can be found in [6]. Thus we will recall only a necessary minimum to present the
results in subsequent sections. Let A be an algebra for an algebraic language /, i.e.,
A = A, f
A
[ f / is a set A endowed with a set of operations indexed by the
connectives from/ such that corresponding to each n-ary connective f there is an n-ary
operation f
A
: A
n
A. Given a term t in the language /, the corresponding term
function on A is denoted t
A
or shortly t if A is clear from the context. We will be
interested mainly in the case when / = /
FL
. Thus the set of terms will be typically just
the set of formulas Fm in the language /
FL
.
Let us x an algebraic language /. In the rest of this section all algebras will
be algebras for the language /. Having an algebra A, we can dene the equational
consequence relation [=
A
. Recall that an evaluation into A is a homomorphism from
the term algebra (i.e., the absolutely free algebra) into A determined uniquely by the
images of variables. Let E = t
i
= s
i
[ i I be a set of identities and t = s an
identity. Then
E [=
A
t = s iff for all evaluations e we have e(t) = e(s)
whenever e(t
i
) = e(s
i
) for all i I.
If E = then we write only [=
A
t = s instead of [=
A
t = s. The denition of [=
A
can be extended to an arbitrary class of algebras K, by E [=
K
t = s iff E [=
A
t = s
holds for every A K.
Having dened the consequence relation in an algebra, we may dene an equational
class of algebras. Given a set of identities E = t
i
= s
i
[ i I, the equational class
dened by E is the class of all algebras A such that [=
A
t
i
= s
i
holds for all i I.
Similarly, one can dene a quasi-equational class of algebras. Given a set of quasi-
identities Q, the quasi-equational class dened by Q is the class of all algebras A such
that for every quasi-identity
t
1
= s
1
and and t
n
= s
n
implies t = s
from Q we have t
i
= s
i
[ i = 1, . . . , n [=
A
t = s.
292 Rostislav Hor ck
Equational and quasi-equational classes can be characterized by means of class op-
erators. Let K be a class of algebras. Then we dene I(K), H(K), S(K), P(K) and
P
U
(K) to be, respectively, the class of all algebras isomorphic to some member of K,
the class of all homomorphic images of members from K, the class of all subalgebras of
members fromK, the class of all direct products of members from K and the class of all
ultraproducts of members fromK. The class Kis called a variety if it is closed under H,
S and P. If K is closed under I, S, P, P
U
and contains a trivial algebra, then K is said
to be a quasivariety. It is known that the smallest variety (resp. quasivariety) containing
a class K is the class HSP(K) (resp. ISPP
U
(K)). That is why we shortly denote the
compositions of operators HSP and ISPP
U
respectively V and Q.
It is very well known due to Birkhoff that varieties are precisely equational classes.
An analogous result for quasivarieties was proved by Malcev.
THEOREM 2.1.1. Let K be a class of algebras. Then the following hold:
1. K is a variety (i.e., K = V(K)) iff it is an equational class.
2. K is a quasivariety (i.e., K = Q(K)) iff it is a quasi-equational class.
Another important concept from universal algebra is that of a subdirect product and
a subdirectly irreducible algebra.
DEFINITION 2.1.2. An algebra A is a subdirect product of a family A
i
[ i I of
algebras if the following hold:
1. Ais a subalgebra of

iI
A
i
,
2.
i
(A) = A
i
for all i I, where
i
denotes the projection to the i-th component.
Given an algebra A, a family A
i
[ i I of algebras and an embedding f :
A

iI
A
i
, we say the f is subdirect if f[A] is a subdirect product of A
i
[ i I.
DEFINITION 2.1.3. An algebra A is said to be subdirectly irreducible if it is non-
trivial and for every subdirect embedding f : A

iI
A
i
there is i I such that

i
f : A A
i
is an isomorphism.
Subdirectly irreducible algebras can be characterized by means of their congruence
lattices. Let A be a nontrivial algebra and Con(A) its congruence lattice. Then A
is subdirectly irreducible iff the bottom element of Con(A) is completely meet-
irreducible. Recall that an element a in a lattice is called completely meet-irreducible if
a =
_
iI
a
i
implies a = a
i
for some i I.
THEOREM 2.1.4. A nontrivial algebra Ais subdirectly irreducible iff is completely
meet irreducible, i.e., Con(A) has a minimum.
The minimum of Con(A) from the previous theorem is called the monolith
of A. Subdirectly irreducible algebras are important because they are building blocks of
any algebra. Precisely, we have the following theorem.
THEOREM 2.1.5. Every nontrivial algebra A is isomorphic to a subdirect product of
subdirectly irreducible algebras (which are homomorphic images of A).
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 293
An important consequence of the previous theorem is that the class of subdirectly
irreducible algebras inside a variety denes the same consequence relation as the whole
variety. Let K be a variety. The class of its subdirectly irreducible members is denoted
K
SI
. We start with a lemma which is easy to prove.
LEMMA 2.1.6. Let K be a class of algebras. Then [=
O(K)
= [=
K
for any operator
O I, S, P.
PROPOSITION 2.1.7. Let K be a variety. Then [=
K
= [=
K
SI
.
Proof. Clearly [=
K
[=
K
SI
since K K
SI
. Conversely, assume E [=
K
SI
t = s.
We have to show that E [=
K
t = s. Let A K. By Theorem 2.1.5 the algebra A
is isomorphic to a subalgebra of

iI
A
i
where A
i
s are subdirectly irreducible and
homomorphic images of A. Since varieties are closed under homomorphic images, we
must have A
i
K
SI
for all i I. The rest follows by Lemma 2.1.6.
The next theorem is known as J onssons lemma. Recall that a variety ' is called
congruence distributive if for all A ' the congruence lattice Con(A) is distributive.
THEOREM 2.1.8. Let V(K) be a congruence distributive variety generated by a class
K and A V(K). If Ais subdirectly irreducible, then A HSP
U
(K).
It is well known that varieties whose members have a lattice reduct are congruence
distributive. This is important for us since most of our algebras will have a lattice reduct
so that the above theorem applies to them.
Finally, we need to recall several denitions on partial subalgebras and partial em-
beddings.
DEFINITION 2.1.9. Let A = A, f
A
i
[ i I be an algebra and , = G A. The
partial subalgebra Gof Ais the partial algebra G = G, f
G
i
[ i I, where for every
n-ary operation f
i
and a
1
, . . . , a
n
G we have
f
G
i
(a
1
, . . . , a
n
) =
_
f
A
i
(a
1
, . . . , a
n
) if f
A
i
(a
1
, . . . , a
n
) G,
undened if f
A
i
(a
1
, . . . , a
n
) , G.
Clearly, every usual subalgebra is also a partial subalgebra whose operations are
dened everywhere.
DEFINITION 2.1.10. Let A = A, f
A
i
[ i I, B = B, f
B
i
[ i I be algebras
of the same type and G a partial subalgebra of A. A one-to-one map h: G B is
called a partial embedding if it preserves all existing operations on G, i.e., for every
n-ary operation f
i
and a
1
, . . . , a
n
G if f
G
i
(a
1
, . . . , a
m
) is dened then
h(f
G
i
(a
1
, . . . , a
m
)) = f
B
i
(h(a
1
), . . . , h(a
m
)).
Let G be a partial subalgebra of an algebra A and f : G B a partial embedding
from Gto an algebra B. Observe that the restriction of f to a nonempty subset G
t
G
gives a partial embedding from the partial subalgebra G
t
to B. In particular, if G = A
(i.e., f is the usual embedding) then the restriction of f to G
t
gives a partial embedding.
294 Rostislav Hor ck
DEFINITION 2.1.11. Let KA be a class of algebras of the same type. We say that
Ais partially embeddable into K if for every nite partial subalgebra Gof Athere is a
partial embedding f : G B for some B K.
2.2 Residuated maps and Galois connections
For details on the notions and results from this section see [5, 11, 14]. Before we
introduce the notion of a residuated map, we recall several common notions from order
theory. Let P = P, be a partially ordered set (shortly poset). A subset X P is
called a downset if X is closed downwards, i.e., for all x, y P we have x X and
y x implies y X. Analogously, Y P is an upset if Y is closed upwards. The
sets of all downsets and upsets of P form posets ordered by inclusion. The poset of all
downsets (resp. upsets) of P is denoted T(P) (resp. |(P)).
Given a subset S P we can nd the smallest downset S containing S as the
intersection of all downsets containing S, i.e.,
S =

X T(P) [ S X.
We can also describe S by means of principal downsets. By a principal downset we
mean a downset of the form x = y P [ y x for some x P. We shortly
write x instead of x. Then S =

xS
x. Dually we can dene the smallest upset
S containing a subset S and a principal upset x.
Other well-known useful notions are those of a closure and an interior operator. A
map : P P is called a closure operator on P if it satises the following conditions:
is order-preserving, i.e., x y implies (x) (y),
is expanding, i.e., x (x), and
idempotent, i.e., ((x)) = (x).
Dually, an interior operator on P is a map : P P which is
order-preserving,
contracting, i.e., (x) x, and
idempotent.
Elements in [P] are called -closed or just closed if the closure operator is clear from
the context. Similarly, elements from [P] are called -open or shortly open. A subset
of closed elements B [P] (resp. open elements B [P]) is called a basis if every
element in [P] (resp. [P]) can be expressed as a meet (resp. join) of elements from B.
Let be a closure or interior operator on P. Its image [P] forms a subposet [P] of
the poset P. Let be a closure operator and an interior operator on P. The closure and
interior operators are completely determined by their images [P] and [P]. Namely,
for x P,
(x) = mina [P] [ x a, (x) = maxa [P] [ a x. (1)
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 295
Conversely any subsets C, O P induce respectively a closure operator and an inte-
rior operator if the following minima and maxima exist for every x P:
(x) = mina C [ x a, (x) = maxa O [ a x. (2)
The posets [P] and [P] are called respectively closure and interior systems of P.
The following denition of a residuated map will be crucial when we dene the
algebraic semantics for substructural logics since the operation interpreting the fusion
is in some sense residuated as we will see later.
DEFINITION 2.2.1. Let P and Q be posets. A map f : P Q is called residuated if
there exists a map f

: Q P such that for all p P and q Q we have

f(p) q iff p f

(q).
In the above case, we say that f and f

form a residuated pair. The map f

is called a
residual of f.
Note that f and f

form an adjunction between posets P and Qwhen we view them

as categories, i.e., objects are the elements of the poset and morphisms are given by the
order relation.
The next proposition lists well-known properties of residuated maps.
PROPOSITION 2.2.2. Let P and Q be posets and f : P Q a residuated map with
its residual f

. Then the following hold.

1. f, f

are monotone,
2. f

(q) = maxp P [ f(p) q,

3. f(p) = minq Q [ p f

(q),
4. f preserves arbitrary existing joins and f

arbitrary existing meets, i.e., for any

X P and Y Qwe have f(
_
X) =
_
xX
f(x) and f

(
_
Y ) =
_
yY
f

(y)
if
_
X and
_
Y exist.
The above proposition shows that residuated maps preserve arbitrary existing joins
and their residuals arbitrary existing meets. This can be further strengthened if P and
Qare complete lattices as is shown in the next proposition.
PROPOSITION 2.2.3. Let P and Q be complete lattices. Then a map f : P Q is
residuated iff f preserves arbitrary joins. Dually, a map f

: Q P is a residual of a
map f : P Q iff f

preserves arbitrary meets.

A notion closely related to residuated maps is that of a Galois connection.
DEFINITION 2.2.4. Let P and Q be posets. The maps

: P Q and

: Q P
form a Galois connection if for all p P and q Q we have
q p

iff p q

.
The maps

and

are called polarities of the Galois connection.
296 Rostislav Hor ck
Observe that polarities of a Galois connection between posets P and Qform also a
residuated pair between P and Q

where Q

denotes the dual poset of Q. Namely,

is a residuated map and

is its residual.
The following are well-known properties of Galois connections. Note that we com-
pose the polarities from left to right since they are written as superscripts of arguments.
PROPOSITION 2.2.5. Let

be a Galois connection between posets P and Q. Then

the following hold:
1. The maps

and

are both order-reversing.
2. The compositions

: P P and

: Q Q are both closure operators.
3. We have

=

and

=

.
Let A, B be sets. We denote their powersets respectively T(A) and T(B). Typi-
cal examples of Galois connections from the literature are Galois connections between
T(A) and T(B) induced by a binary relation R A B. The polarities in this case
are given by upper bounds and lower bounds with respect to the relation R. We are
going to discuss these examples in more details because we will need them later. How-
ever, we present them in a more general setting. Namely, we will replace the sets A, B
by posets. Clearly, this is more general since each set can be viewed as a discrete poset
(i.e., a set ordered by the identity relation). Further, note that T(A) = T(A) = |(A)
for a discrete poset A.
Let A, B be two posets and R A B a relation satisfying for all a, x A and
b, y B the following implication:
x a and a R b and b y implies x R y. (3)
We call a binary relation R satisfying (3) a poset relation. The name comes from cate-
gory theory since poset relations are just the usual relations in the category of posets.
One can use R in order to dene a Galois connection between T(A) and |(A)
given by maps

: T(A) |(B) and

: |(B) T(A). The maps

and

are
dened as follows:
X

= y B [ (x X)(x R y),
Y

= x A [ (y Y )(x R y).
Observe that by (3) the image X

is an upset in B and Y

is a downset in A. To
see that

,

dene a Galois connection, consider X T(A) and Y |(B). Then

Y X

in fact means that every element y Y satises the dening condition (x

X)(x R y). Thus validity of Y X

is equivalent to the validity of the following

rst-order formula:
(y Y )(x X)(x R y).
Since the validity of the above formula does not depend on the order of quantiers, it
is immediate that the validity of this formula is also equivalent to validity of X Y

.
Consequently, we obtain the following lemma.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 297
LEMMA 2.2.6. The pair of maps

: T(A) |(B) and

: |(B) T(A) denes a
Galois connection, i.e.,
Y X

iff X Y

.
It follows from Proposition 2.2.5 that the composition

: T(A) T(A) is a
closure operator. We denote this closure operator
R
where the subscript R refers to
the relation R dening

and

. One can easily nd a basis for
R
as is shown in the
following lemma.
LEMMA 2.2.7. The collection (y)

[ y B, where y is a principal upset from

|(B), forms a basis for
R
.
Proof. Note that for each upset Y |(B) the downset Y

is
R
-closed since Y

=
Y

by Proposition 2.2.5. In particular, for each y B the downset (y)

is
R
-closed.
Moreover, we have
Y

= x A [ (y Y )(x R y) =

yY
x A [ x R y =

yY
(y)

.
The last equality follows from (3).
3 FL-algebras
This chapter introduces an algebraic semantics for the logic FL, namely the class
of so-called FL-algebras. We will work only with algebras for the language /
FL
(and
its fragments) consisting of binary connectives , , /, , and constants 0, 1. We call
, , / respectively multiplication, left and right division. The absolutely free algebra for
this language is the term algebra Fm = Fm, , , /, , , 0, 1. When writing terms
in this language we will assume in the absence of parentheses that is performed rst
followed by , / and nally , . We write a
n
as a shortcut for a a a (n times),
where a
0
= 1. We often write ab instead of a b.
We start with denitions of simple structures adding more and more conditions so
that we nally obtain the denition of an FL-algebra. An algebra A = A, with a
binary operation is called a groupoid. If the multiplication is associative then A is
said to be a semigroup. If there is a partial order on A and the multiplication is
order-preserving in both arguments (i.e., x y implies zx zy and xz yz for
all x, y, z A), then A = A, , is called a partially ordered groupoid if A, is
a groupoid and a partially ordered semigroup if A, is a semigroup. Further, if this
partial order is in fact a lattice order and the multiplication distributes over nite joins
(i.e., x (y z) = x y x z and (y z) x = y xz x hold for all x, y, z A), then
A = A, , is called a lattice ordered groupoid (resp. lattice ordered semigroup).
Let A = A, , be a partially ordered semigroup. If there is a neutral element
1 for the multiplication (i.e., 1 x = x = x 1 holds for all x A), then A =
A, , 1, is said to be a partially ordered monoid (shortly pomonoid). Analogously
as for semigroups we dene a lattice ordered monoid (shortly -monoid) as a pomonoid
whose partial order is a lattice order and x(yz) = xyxz and (yz)x = yxzx.
298 Rostislav Hor ck
Next we will illustrate the impact the theory of residuated maps has on the properties
of A. Let A = A, , be a partially ordered groupoid whose operation is residuated
component-wise, i.e., for every a A the unary maps l
a
(x) = a x and r
a
(x) = x a
are residuated. Using the residuals l

a
and r

a
of l
a
and r
a
, we can dene new binary
operations , / on Aas follows:
xz = l

x
(z) and z/y = r

y
(z).
Then the operations , and / are connected by the following property which we call
the residuation property:
x y z i y xz i x z/y.
The above condition is in fact often used as the dening condition for residuated binary
maps. Precisely, a binary map : AA Ais said to be residuated if there exist binary
operations : AA A, /: AA A satisfying the residuation property. Thus we
have the following denition.
DEFINITION 3.0.1. A residuated partially ordered groupoid is an algebraic structure
A = A, , , /, such that A, is a poset and the residuation property
x y z i y xz i x z/y.
is satised for all x, y, z A.
Note that in the previous denition we do not claim that A, , is a partially
ordered groupoid because it is not necessary. It follows immediately fromthe residuation
property. Indeed, let x, y, z A and assume that x y. Then yz yz implies
y yz/z by the residuation property. Thus we have x yz/z. Using the residuation
property again, we obtain xz yz. Similarly one can prove zx zy. It is also easy
to prove that both divisions are order-preserving in the numerator and order-reversing in
the denominator.
Properties of unary residuated maps can be easily transferred also to the binary case.
We can apply Proposition 2.2.2 to get the following proposition.
PROPOSITION 3.0.2. Let A be a residuated groupoid and a, b, c A. Then the fol-
lowing hold:
1. ac = maxb A [ ab c and c/b = maxa A [ ab c.
2. a b = minc A [ b ac = minc A [ a c/b.
3. Multiplication distributes over any existing join, i.e., if
_
X and
_
Y exist for
X, Y A, then so does
_
xX,yY
xy, and
(

X) (

Y ) =

xX,yY
xy.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 299
4. Divisions preserve all existing meets in the numerator and convert all existing
joins in the denominator to meets, i.e., if
_
X and
_
Y exist for X, Y A, then
for any z A the following equalities hold (in particular the right-hand sides
exist):
z(

Y ) =

yY
zy, (

Y )/z =

yY
y/z,
(

X)z =

xX
xz, z/(

X) =

xX
z/x.
Residuated lattice ordered semigroups, residuated pomonoids, and residuated -
monoids are dened analogously as residuated partially ordered groupoids. Observe
that the residuation property implies that the multiplication distributes over any existing
join. Thus having a residuated pomonoid A whose partial order is a lattice order, it
follows immediately that Ais a residuated -monoid.
Now we are ready to dene FL-algebras which form a complete algebraic semantics
for the logic FL as we will see later.
DEFINITION 3.0.3. An algebra A = A, , , /, , , 0, 1 is an FL-algebra if
A, , 1 is a monoid,
A, , is a lattice,
a b c iff b ac iff a c/b, for all a, b, c A,
0 is an arbitrary element of A.
In other words, an FL-algebra A is a residuated -monoid endowed with a constant 0.
If the lattice order is linear then we call A an FL-chain. The class of all FL-algebras
is denoted F'.
Although the class of FL-algebras is not dened only by identities, it can be done as
is shown in the following theorem. This theorem uses, instead of identities, inequalities.
However, since every FL-algebra has a lattice reduct, we can formally consider each
inequality x y as an identity x = x y or y = x y.
THEOREM 3.0.4 ([4]). An algebra A = A, , , /, , , 0, 1 is an FL-algebra iff it
satises the equations dening monoids, the equations dening lattices and the following
identities:
1. x (xz y) z,
2. (y z/x) x z,
3. y x(x y z),
4. y (z y x)/x.
Therefore, F' is a variety.
300 Rostislav Hor ck
Moreover, F' is a congruence distributive variety. This follows from the fact that
each FL-algebra has a lattice reduct. Thus we can show the congruence distributivity
for F' by the same majority term as for lattices (see [6]).
For the readers convenience the following lemma provides a list of basic properties
of FL-algebras which follow easily from Denition 3.0.3.
LEMMA 3.0.5. The following identities hold in any FL-algebra.
1. (x/y)y x and y(yx) x,
2. (x/y)/z = x/zy and z(yx) = yzx,
3. x(y/z) = (xy)/z,
4. x/1 = x = 1x,
5. 1 xx and 1 x/x,
6. (z/y)(y/x) z/x and (xy)(yz) xz.
An FL-algebra Ais commutative if the multiplication is commutative, i.e., xy = yx
holds in A. We refer to commutative FL-algebras as FL
e
-algebras. Observe that in
FL
e
-algebras the third condition in Denition 3.0.3 forces the equality of ab and b/a.
In this case we denote them as a b.
We call an FL-algebra A integral if 1 is a top element, i.e., A satises the identity
x 1. Integral FL-algebras are called shortly FL
i
-algebras. Similarly, FL
o
-algebras
are FL-algebras where 0 is the bottom element, i.e., they satisfy 0 x. FL-algebras
where the identity x x
2
holds are called contractive or sometimes square-increasing.
We refer to them as FL
c
-algebras. The contractivity of an FL-algebra A implies that
the multiplication on its negative elements coincides with the meet. An element a A
is called negative (resp. positive) if a 1 (resp. a 1). Let x, y A be negative
elements. Then xy x1 = x and similarly xy y. Thus xy x y. Consequently,
x y (x y)
2
xy x y.
This implies that all contractive integral FL-algebras are commutative.
Observe that the notation FL
e
-algebras, FL
i
-algebras, FL
o
-algebras and FL
c
-algeb-
ras resembles the notation for basic substructural logics. This is not a coincidence be-
cause these algebras will serve as equivalent algebraic semantics for the corresponding
substructural logics. Hence we use an analogous notation for names of FL-algebras as
for the basic substructural logics. More precisely, let S e, c, i, o. Then FL
S
-algebras
are FL-algebras satisfying the extra properties listed in S (see Table 2).
In FL-algebras the constant 0 allows us to dene the following two unary operations
a = a0 and a = 0/a which are called respectively left and right negation. When
writing terms containing the negations, we assume that the negations are performed
rst in the absence of parentheses. Note that the following equivalence holds in any
FL-algebra A:
x y i y x.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 301
The equivalence above follows immediately from the fact that the multiplication in
FL-algebras is residuated. It says that , form a Galois connection. Thus , are
both order-reversing mappings. The next lemma lists basic properties of the negations.
LEMMA 3.0.6. Let Abe an FL-algebra and x A. Then the following hold:
1. 1 = 0 = 1,
2. x(x) 0 and (x)x 0,
3. x x and x x,
4. x = x and x = x,
5. xy = (yx) and y/x = (xy).
An FL-algebra A is said to be involutive if the identities x = x and x = x
hold in A. Consequently, 0 = 1 = 1 = 1 = 0. Moreover, , become
order-reversing bijections. Indeed, x = y implies x = x = y = y. Similarly,
x = y implies x = y. Observe also that the identity x 1 is equivalent to 0 x in
involutive FL-algebras. Indeed, let A be an involutive FL-algebra and a A. Assume
that the identity x 1 holds in A. Then a 1. Thus 0 = 1 a = a.
Conversely, suppose that 0 x holds. Then 0 x. Thus x = x 0 =
1. Consequently, the classes of involutive FL
i
-algebras, involutive FL
o
-algebras and
involutive FL
w
-algebras are the same.
An FL-algebra A is called cyclic if x = x. This happens in particular when A
is commutative because = / holds in FL
e
-algebras. Thus, every FL
e
-algebra is cyclic
but the converse need not be true.
Another important class of FL-algebras is the class of residuated lattices. An FL-
algebra is called a residuated lattice (or shortly RL-algebra) if it satises the identity
1 = 0. Residuated lattices are usually dened as the 0-free reducts of FL-algebras.
However, we choose the above denition in order to capture FL-algebras and residuated
lattices within one variety of algebras of the same type.
Since RL-algebras forma subvariety of FL-algebras, any denition for FL-algebras
applies also to RL-algebras. Thus we can dene RL
S
-algebras for S e, c, i, o in the
same way as FL
S
-algebras. However, note that the only RL
o
-algebra is the trivial one.
Indeed, 1 has to be a bottom element in any RL
o
-algebra A. Then it follows that A is
trivial by the following lemma.
LEMMA 3.0.7. Let A be a nontrivial FL-algebra. Then there is a strictly negative
element a A, i.e., a < 1.
Proof. Since A is nontrivial, there is an element b A such that b ,= 1. If 1 , b then
a = b1 < 1. If 1 < b then we take a = b1. Clearly we have a 11 = 1. Moreover,
a < 1; otherwise b = b a = b (b1) 1.
All the properties of FL-algebras and RL-algebras we mentioned so far are sum-
marized in Table 2. Note that classes of FL-algebras and RL-algebras are denoted
302 Rostislav Hor ck
Adjective Dening equations Subvariety of F' Subvariety of 1'
Commutative (e) x y = y x F'
e
1'
e
Integral (i) x 1 F'
i
1'
i
Contractive (c) x x
2
F'
c
1'
c
Bounded (o) 0 x F'
o
trivial variety
Integral bounded (w) 0 x 1 F'
w
trivial variety
Involutive x = x = x "nF' "n1'
Cyclic x = x CyF' Cy1'
Table 2. Subvarieties of FL-algebras and residuated lattices.
by the blackboard bold letters, e.g., F'
ew
denotes the variety of FL
ew
-algebras or
"nF'
w
stands for the variety of involutive FL
w
-algebras. Having an FL-algebra A =
A, , , /, , , 0, 1, one can change the interpretation of 0 to 1 in order to dene a
corresponding RL-algebra A
r
= A, , , /, , , 1, 1. We will omit the double 1 in the
signature above, i.e., we write just A
r
= A, , , /, , , 1.
3.1 Examples of FL-algebras
The variety of FL-algebras encompasses a lot of well-known classes of algebras
which were dened and also studied independently. Among these classes are for in-
stance Heyting algebras which can be viewed as FL
cw
-algebras. Moreover, involutive
Heyting algebras are just Boolean algebras so that "nF'
cw
is nothing but the variety of
Boolean algebras.
Another well-known class of algebras which can be viewed as FL-algebras are lat-
tice ordered groups (shortly -groups). The signature of an -group Gis usually written
as G, , , ,
1
, 1. However, -groups are term equivalent to RL-algebras satisfying
an extra identity x(x1) = 1. The term equivalence translates x
1
to x1 and con-
versely xy to x
1
y and x/y to xy
1
. Let G be an -group viewed as an RL-algebra.
Since RL-algebras are FL-algebras satisfying 1 = 0, we have x = x1 = x
1
in G.
Analogously x = 1/x = x
1
. Consequently, x = x. Moreover, x = x =
(x
1
)
1
= x. Thus -groups in fact form a subvariety of cyclic involutive RL-algebras.
There are also examples of FL-algebras coming from fuzzy logic. Recall that a
t-norm is a binary operation on the real unit interval [0, 1] which is associative, com-
mutative, monotone and 1 is its neutral element. Moreover, if a (
_
Y ) =
_
yY
a y
for any a Y [0, 1], is said to be left-continuous. In such a case we also have
(
_
Y ) a =
_
yY
y a by commutativity. Assume that is left-continuous. Then
is residuated component-wise (see Proposition 2.2.3 and the discussion above Deni-
tion 3.0.1). Consequently, there is a binary operation

on [0, 1] such that

x y z i y x

z.
Summing up, we have a residuated commutative pomonoid [0, 1], , 1, which is lin-
early ordered with a top element 1 and a bottom element 0. Thus the algebra [0, 1]

=
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 303
[0, 1], ,

, , , 0, 1 is just an FL
ew
-chain. These algebras are also known as stan-
dard MTL-algebras. The word standard refers to the fact that the universe of these
algebras is the intended set of truth-degrees [0, 1]. Later we will see that the semilinear
extension of FL
ew
is complete with respect to FL
ew
-chains on [0, 1].
The above denitions works also for the case when is in addition continuous (in the
usual way). Then the corresponding FL
ew
-chain [0, 1]

is in fact a standard BL-algebra

(for more details on BL-algebras see Chapter V).
3.2 Involutive FL-algebras
We need to recall several properties of involutive FL-algebras which will be use-
ful in the sequel. The most important property is that in involutive FL-algebras the
divisions are denable by means of multiplication and negations. Also conversely, the
multiplication is denable by divisions and negations.
LEMMA 3.2.1. Let Abe an involutive FL-algebra. Then we have
1. x y = (yx) = (y/x),
2. xy = [(y)x] and y/x = [x(y)].
Proof. Let x, y, z A. For the rst part we have the following chain of equivalences:
xy z = z iff xy(z) 0 iff y(z) x iff z yx iff (yx) z.
Then we can prove xy (yx) by substituting (yx) for z and (yx) xy
by substituting xy for z. The proof of xy = (y/x) is analogous.
To prove the second part we can employ the following chain of equivalences:
z xy iff xz y iff (y)xz 0 iff (y)x z iff z [(y)x].
Again by substituting respectively xy and [(y)x] for z, we obtain xy = [(y)x].
The proof of the other equality is analogous.
Considering the previous lemma, it is natural to ask whether we can dene invo-
lutive FL-algebras in a restricted language replacing either the division operations or
the multiplication by negations , . Both possibilities can be done. However, we will
focus only on the second possibility because we will need it later.
A division lattice P = P, , , , / is a lattice endowed with two division opera-
tions, which satisfy the following condition:
y xz i x z/y. (4)
A division lattice P is said to be associative if the identity x(y/z) = (xy)/z holds
in P. Further, we say that P admits a unit if there is an element 1 P such that
1x = x = x/1.
An involutive pair on a lattice P is a pair of order-reversing operations ,
satisfying the double negation law, i.e., x = x = x. We call a division lattice
involutive if it is equipped with an involutive pair , and the operations , are
compatible with the divisions, i.e., the following version of contraposition holds:
yx = y/x. (5)
304 Rostislav Hor ck
THEOREM 3.2.2 ([18]). Involutive FL-algebras are term equivalent to associative in-
volutive division lattices that admit a unit.
The term equivalence in the above theorem is proved via the following translations.
Having an involutive FL-algebra, one can dene
x = x0 and x = 0/x.
Conversely, having an associative involutive division lattice that admits a unit 1, one can
dene
x y = (yx) = (y/x) and 0 = 1 = 1.
3.3 A bottom and a top element
An FL-algebra A need not possess a bottom element, e.g. -groups are typical ex-
amples of FL-algebras without a bottom element. Nevertheless, if A has a bottom
element then it has to have also a top element since = is a top element
of A. It is a natural question whether every FL-algebra A can be extended to a lower
bounded one. The answer to this question is afrmative.
Let A = A, , , /, , , 0, 1 be an FL-algebra. We describe howto extend Ato an
FL-algebra A

having a bottom element . Let A

= A, with A, =
. The order on A

extends the order on A by setting a for every a A. The

multiplication and divisions are extended as follows:
x = x = ,
x = x =
_
if x A ,
if x = ,
x = /x = ,
x = x/ = ,
x = x/ =
_
if x A ,
if x = ,
x = /x =
_
if x A ,
if x = .
Then it can be shown that A

= A

, , , /, , , 0, 1 is an FL-algebra such that Ais

a subalgebra of A

. Thus we have the following proposition.

PROPOSITION 3.3.1 ([21]). Every FL-algebra Acan be embedded in lower bounded
FL-algebra A

. Moreover, the construction of A

preserves the following properties

of A: commutativity, contraction and linear order.
3.4 Algebraizability
In this section we are going to present the most important consequences of the fact
that the logic FL is algebraizable. Nevertheless, we will not dene the notion of alge-
braizability precisely (for details see [1] and Chapter II). Roughly speaking, a logic L is
algebraizable if there is a class of algebras ' (so-called equivalent algebraic semantics)
satisfying the following conditions:
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 305
1. There is a translation from formulas in the language of L to identities in the lan-
guage of members from'. Conversely, there is also a translation from identities
to formulas.
2. Let be a set of formulas in the language of L. Then

L
i [] [=
L
(). (6)
3. Let s = t be an identity in the language of '. Then
s = t [=
L
((s = t)) and ((s = t)) [=
L
s = t. (7)
In order to show that the logic FL is algebraizable, one has to rst dene the transla-
tions and . Let Fm and s = t an identity in the language /
FL
. Then () is the
identity 1 = 1 (or shortly 1 ) and (s = t) is the formula (st) (ts). Observe
that the formulas Fm are just terms in the language /
FL
so the above denition makes
sense. Then one has to prove that and satisfy the conditions from the denition of
an algebraizable logic. This was done in [19]. Further details can be found in Chapter II
where it is proved that FL is an algebraizable weakly implicative logic. Such logics are
called algebraically implicative in Chapter II.
THEOREM 3.4.1. The translations and satisfy (6) and (7) for L = FL and ' = F'.
So the logic FL is algebraizable and its equivalent algebraic semantics is the variety F'.
The fact that FL is algebraizable has several important consequences which follow
from general results on algebraizable logics. First, there is a dual lattice isomorphism V
between the lattice of substructural logics (FL) and the subvariety lattice (F') of
FL-algebras. Given a substructural logic L = FL+, we dene V(L) as the subvariety
of FL-algebras satisfying identities from []. Conversely, given a subvariety ' of
FL-algebras axiomatized relatively to F' by a set of identities E, the inverse V
1
(')
is the axiomatic extension L = FL + [E]. Second, we have the strong completeness
theorem not only for FL but for any axiomatic extension of FL (i.e., any substructural
logic).
THEOREM 3.4.2. Let Fm and Es = t a set of identities in the language
/
FL
. Then the following hold:
1. For a substructural logic L we have

L
i 1 [ [=
V(L)
1 .
2. For a subvariety ' of FL-algebras we have
E [=
L
s = t i (uv) (vu) [ u = v E
V
1
(L)
(st) (ts).
306 Rostislav Hor ck
3.5 Congruences and lters
In this section we are going to show that congruences in an FL-algebra A can be
characterized by certain subsets of A which are called lters. More precisely, we are
going to show that lters form a lattice which is isomorphic to Con(A).
Before we dene the notion of a lter, we have to introduce conjugates. Let A be
an algebra for the language /
FL
. Given a, x A, we dene the left conjugate and
the right conjugate of x with respect to a respectively as
a
(x) = (axa) 1 and

a
(x) = (ax/a) 1. An iterated conjugate of x with respect to a
1
, . . . , a
n
A is a
composition of the form
a
1
(
a
2
(
a
n
(x) )), where
a
i

a
i
,
a
i
for every i.
Let X A. Then the set of all iterated conjugates of elements fromX is denoted (X).
If X = x then (x) is abbreviated as (x). A subset X A is called normal if it
is closed under all iterated conjugates, i.e., (X) X. Given a subset X A we also
dene its closure under multiplication
(X) = x
1
x
n
A [ x
i
X, n 1 1.
The left and right conjugates were dened for any algebra Afor the language /
FL
.
If, in addition, A is an FL-algebra, then
1
(x) =
1
(x) = x 1 since 1(x1) =
(1x)/1 = x hold in any FL-algebra. Furthermore,
a
(1) =
a
(1) = 1 because 1 aa
and 1 a/a (see Lemma 3.0.5).
DEFINITION 3.5.1. Let Abe an FL-algebra and F A. Then F is a lter of Aif the
following hold:
(u) 1 F
(up) F is an upset, i.e., F = F,
(p) F is closed under multiplication, i.e., (F) F,
(n) F is normal, i.e., (F) F.
It is easy to see that (n) can be simplied when A is commutative. Namely, the
condition (F) F can be replaced by F 1 F, where F 1 = x 1 [ x F.
Indeed, in the commutative case we have
(axa) 1 = (aax) 1 x 1.
If Ais in addition integral (i.e., an FL
ei
-algebra), then (n) is redundant.
The notion of a lter was not discovered only by investigating congruences of FL-
algebras but it comes from the logic itself. Namely lters of an FL-algebra A are just
subsets closed under all deductions; a subset F A is closed under all deductions if for
all Fm and all evaluations e: Fm A we have the following implication:

FL
and e[] F implies e() F.
For further details on this see the notion of a logical lter in Chapter II.
Let Abe an FL-algebra. Observe that the collection Fi(A) of all lters of Aform a
poset Fi(A) = Fi(A), ordered by inclusion. It is in fact a complete lattice because
it is easy to see that lters are closed under arbitrary intersections. Since Fi(A) is a
complete lattice, it makes sense to dene a least lter F(X) containing a subset X A
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 307
as the intersection of all lters F such that X F. We also say that F(X) is the lter
generated by X. It is a natural question whether we can describe elements of F(X) by
means of elements from X. The answer is the content of the following theorem.
THEOREM 3.5.2. Let A be an FL-algebra and X A. Then F(X) = (X).
In other words, y F(X) iff there are iterated conjugates
1
, . . . ,
n
and elements
x
1
, . . . , x
n
X such that
1
(x
1
)
n
(x
n
) y. If A is commutative then F(X) =
(X 1).
Now we can dene the isomorphism between Fi(A) and Con(A) for an FL-alge-
bra A. Let Con(A) and F Fi(A). We dene two maps F
c
: Con(A) Fi(A)
and
f
: Fi(A) Con(A) as follows:
F
c
() = (1/),

f
(F) = a, b [ ab, ba F.
Note that
f
(F) is the Leibniz congruence of the matrix A, F (see Chapter II). Now
we are ready for the promised theorem on lters and congruences.
THEOREM 3.5.3. Let A be an FL-algebra. Then the lattice Fi(A) is isomorphic to
the congruence lattice Con(A) via the mutually inverse maps
f
and F
c
.
Let A be an FL-algebra and F Fi(A). The above theorem justies the conven-
tion of writing A/F instead of A/
f
(F). Analogously, a congruence class of a A
with respect to
f
(F) is denoted a/F.
3.6 Semilinear varieties of FL-algebras
Now we have all the necessary background to introduce semilinear substructural
logics. The notion of a semilinear logic is semantical since it is dened by a property
of the corresponding algebraic semantics. Roughly, a semilinear logic should be a logic
which is sound and complete with respect to a class of linearly ordered algebras. We
will make this denition precise for substructural logics. By algebraization we know
that every substructural logic L has an equivalent algebraic semantics V(L). Moreover,
the consequence relation
L
can be equivalently translated to the consequence relation
[=
V(L)
, see Theorem 3.4.2. In particular, L is sound and complete with respect to the
class of algebras V(L). Inside the variety V(L) we can identify linearly ordered algebras
and use them in the denition of a semilinear substructural logic. More precisely, for a
variety of FL-algebras ', let us denote the class of all FL-chains from ' by '
C
.
DEFINITION 3.6.1. Let L be a substructural logic and ' its equivalent algebraic se-
mantics. Then L is semilinear if [=
L
= [=
L
C
. In that case, ' is called a semilinear
variety.
Concerning the above denition, we also refer to Chapter II where semilinear logics
are dened as weakly implicative logics complete with respect to the class of its linear
models. Thus semilinear logics in Chapter II are understood in a broader context than
here because every substructural logic is weakly implicative but not vice versa.
308 Rostislav Hor ck
It follows immediately from the denition that semilinear varieties are generated
by their chains since we have [=
L
s = t iff [=
L
C
s = t for any identity s = t. One
can prove also the converse implication saying that varieties generated by chains are
semilinear. To see this, let C be a class of FL-chains and K = V(C). By Theorem 2.1.8
K
SI
HSP
U
(C). Furthermore, ultraproducts, subalgebras and homomorphic images
of chains are again chains. Thus we have
K
SI
HSP
U
(C) K
C
,
showing that subdirectly irreducible members of Khave to be chains. Then using Propo-
sition 2.1.7, we obtain
[=
K
[=
K
C
[=
K
SI
= [=
K
.
Hence [=
K
= [=
K
C
, i.e., K is semilinear. Summing up the above ideas, we obtain the
following proposition.
PROPOSITION 3.6.2. Let ' be a variety of FL-algebras. Then the following are equiv-
alent:
' is semilinear,
' = V('
C
),
'
SI
'
C
.
Given a variety ' of FL-algebras, we can nd the least variety containing '
C
. We
denote this variety '

. The class '

is clearly the variety generated by '

C
, i.e., '

=
V('
C
). Thus '

is a semilinear variety by Proposition 3.6.2. Moreover, '

is the
greatest semilinear variety below '. Indeed, assume that '

K ' for a semilinear

variety K. Since K is generated by chains which belong to ', we must have K
C
'
C
.
Hence K '

. Let L and L

be the substructural logics corresponding respectively

to the varieties ' and '

via the dual lattice isomorphism between substructural logics

and subvarieties of FL-algebras. Then it follows from the discussion above that L

is the
weakest semilinear substructural logic above L. A reader may want to consult Chapter II
where the logic L

is also discussed. Note that '

C
is denoted MOD

(L) in Chapter II.

It is a natural question how to axiomatize the variety '

(resp. logic L

). In fact, it is
sufcient to axiomatize F'

because '

= F'

'. In order to axiomatize F'

, it suf-
ces by Propositions 2.1.7 and 3.6.2 to nd identities expressing the fact that subdirectly
irreducible members in the variety given by these identities are linearly ordered.
We start with a useful lemma characterizing intersections of principal lters (i.e., l-
ters generated by singletons). Let Abe any algebra for the language /
FL
and X, Y A.
Then we dene XY = xy [ x X, y Y . If A = Fmthen 1 = X denotes the
set of identities 1 = [ X. Recall also that (X) denotes the set of all iterated
conjugates of elements from X.
LEMMA 3.6.3. Let Abe an FL-algebra and x, y A. Then
F(x) F(y) = F((x) (y)).
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 309
Proof. (): Since F((x)(y)) is the least lter containing (x)(y), it is sufcient
to prove that a b F(x) F(y) for all a (x) and b (y). Clearly, a F(x)
since lters are closed under conjugations. Thus a a b F(x) because F(x) is an
upset. Similarly, a b F(y).
(): Let z F(x) F(y). By Theorem 3.5.2 there are k, l N, a (x) and
b (y) such that a
k
z and b
l
z. Consequently, a
k
b
l
z. Since multiplication
distributes over joins, we get
(a b)
k+l
=

f : 1,...,k+l]a,b]
f(1) f(k + l).
As every f(1) f(k + l) contains either at least k-times a or at least l-times b, it
follows that (a b)
k+l
a
k
b
l
(note that a, b 1 because any conjugate is less than
or equal to 1). Thus (a b)
k+l
z. This means that z belongs to the lter generated by
a b (x) (y).
REMARK 3.6.4. Let p, q Fm be propositional variables. It is shown in Chapter II
that the set of formulas (p) (q) forms a p-disjunction. The above lemma shows that
this p-disjunction satises a weak form of the Proof by Cases Property.
Using Lemma 3.6.3, we can prove that a subdirectly irreducible FL-algebra A is
linearly ordered iff it satises the following set of identities:
1 = ((x y)y) ((x y)x), (8)
where x, y Fm are variables and ((x y)y), ((x y)x) are sets of iterated
conjugates in the term algebra Fm.
To see that the identities (8) gives an axiomatization for F'

, observe rst that in

any FL-algebra Awe have x y iff x y y iff 1 (x y)y. Thus an FL-algebra
Ais linearly ordered iff it satises the following universal sentence:
(x)(y)(1 (x y)y or 1 (x y)x).
Further, observe that in any FL-algebra A we have 1 x implies (x) = 1 for any
iterated conjugate . Indeed, let a A. Then 1 x implies a xa. Thus 1 = axa1
=
a
(x). Similarly, 1 =
a
(x). Thus also 1 = (x) for any iterated conjugate .
Consequently, every FL-chain Asatises the universal sentence
(x)(y)(1 =
1
((x y)y) or 1 =
2
((x y)x)),
for arbitrary iterated conjugates
1
,
2
. Then it is easy to see that every FL-chain Ahas
to satisfy all the identities (8), i.e., the identities:
1 =
1
((x y)y)
2
((x y)x) [
1
,
2
iterated conjugates.
Conversely, assume that A is a subdirectly irreducible FL-algebra satisfying (8).
We want to show that for any x, y A either 1 (x y)y or 1 (x y)x. By
Lemma 3.6.3 we have
F((x y)y) F((x y)x) = F(((x y)y) ((x y)x)) = 1.
310 Rostislav Hor ck
Since A is subdirectly irreducible, the least lter 1 has to be a completely meet-
irreducible element of Fi(A) (see Theorems 2.1.4 and 3.5.3). Thus one of the lters
F((x y)y), F((x y)x) has to be 1. Consequently, either 1 (x y)y or
1 (x y)x.
The axiomatization of F'

given by identities (8) is not nite. Nevertheless, the

following theorem shows that we can improve it so that the resulting axiomatization is
nite. Note also that Chapter II provides another innite axiomatization of F'

using
the prelinearity axioms similar to (8).
THEOREM 3.6.5. The class F'

of semilinear FL-algebras is axiomatized by the fol-

lowing 4-variable identity:
1 =
a
((x y)y)
b
((x y)x). (9)
Proof. It is clear that an FL-algebra satisfying 1 =
1
((x y)y)
2
((x y)x) for
all iterated conjugates
1
,
2
satises in particular 1 =
a
((x y)y)
b
((x y)x).
Conversely, let Abe an FL-algebra satisfying 1 =
a
((x y)y)
b
((x y)x).
Observe that the following quasi-identities hold in A:
c d = 1
a
(c) d = 1, c d = 1 c
b
(d) = 1.
Indeed, assume that c d = 1. Then c, d 1. If we substitute d for x, c for y and 1 for
b in (9), we get
1 =
a
((d c)c)
1
((d c)d) =
a
(c) (d 1) =
a
(c) d.
Analogously we can derive the second quasi-identity. Finally, applying these quasi-
identities on the identity 1 =
a
((x y)y)
b
((x y)x), we can derive 1 =

1
((x y)y)
2
((x y)x) for any iterated conjugates
1
,
2
.
The above theorem suggests the following denition.
DEFINITION 3.6.6. An FL-algebra is called semilinear if it satises the identity
1 =
a
((x y)y)
b
((x y)x).
The axiomatization of F'

given in Theorem 3.6.5 can be further simplied if we

replace the variety F' by the variety F'
e
of commutative FL-algebras. Then we can
get rid of the conjugates in (9). Indeed, let A be a semilinear FL
e
-algebra. Then A
satises (9), in particular for a = b = 1 we have
1 = ((x y)y 1) ((x y)x 1). (10)
Conversely, assume that A is an FL
e
-algebra satisfying identity (10). For all a, x A
we have
a
(x) = axa 1 = aax 1 x 1. Similarly, we have
b
(y) y 1 for
b, y A. Using this observation we can obtain the following chain of inequalities:
1 = ((x y)y 1) ((x y)x 1)
a
((x y)y)
b
((x y)x) 1.
Thus the identity (9) holds in A, i.e., Ais semilinear.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 311
The identity (10) can be simplied a little bit more if one observes that (xy)y =
xy yy and (x y)x = xx yx. Since 1 xx and 1 yy, we obtain
(x y)y 1 = xy 1 and (x y)x = yx 1. Thus the next theorem follows.
THEOREM 3.6.7. The variety F'

e
is axiomatized relatively to F'
e
by the identity
1 = (yx 1) (xy 1). (11)
If we want to axiomatize semilinear FL
ei
-algebras relatively to F'
ei
we can even
drop the meets with 1 in (11) because 1 is the top element in FL
ei
-algebras.
THEOREM 3.6.8. The variety F'

ei
is axiomatized relatively to F'
ei
by the identity
1 = yx xy. (12)
As we already mentioned, most fuzzy logics are in fact semilinear substructural
logics. However, many of them were introduced independently. Thus they appear in
the literature under different names than used here. The same is true also for corre-
sponding algebras of truth values. Table 3 provides a translation table between names of
semilinear FL-algebras used here and their original names.
Our name Original name
FL

ew
-algebras MTL-algebras
FL

cw
-algebras G odel algebras
FL

w
-algebras psMTL
r
-algebras
Involutive FL

ew
-algebras IMTL-algebras
Involutive FL

cw
-algebras Boolean algebras
FL

e
-algebras (with ,) UL-algebras
Involutive FL

e
-algebras (with ,) IUL-algebras
RL

ei
-algebras prelinear semihoops
Table 3. Our names versus original names of semilinear FL-algebras.
3.7 Nuclei and conuclei
Nuclei and conuclei belong to the most useful constructions on FL-algebras (as
we will see also in this chapter). They are in fact closure and interior operators on
FL-algebras which preserve their semigroup structure in a lax way. A closure operator
on an FL-algebra A = A, , , /, , , 0, 1 is called a nucleus if for all x, y A we
have
(x) (y) (x y).
On the other hand, an interior operator on A is said to be a conucleus
1
if for all
x, y A we have
(x) (y) (x y) and (1) = 1.
1
In the literature there is also a slightly more general denition of a conucleus where the condition (1) =
1 is replaced by a weaker condition (1)(x) = (x) = (x)(1).
312 Rostislav Hor ck
Recall that closure and interior operators are fully determined by their images. The same
is true also for nuclei and conuclei.
LEMMA 3.7.1. Let A, B be FL-algebras, a closure operator on Aand an interior
operator on B. Then we have the following:
1. is a nucleus on Aiff x/y, yx [A] for all x [A], y A.
2. is a conucleus on B iff [B] forms a submonoid of B.
Proof. Suppose that is a nucleus. Let x [A] and y A. We have to show that x/y
is -closed, i.e., (x/y) = x/y. Clearly x/y (x/y) because is a closure operator.
To see the other inequality (x/y) x/y, note that
(x/y)y (x/y)(y) ((x/y)y) (x) = x.
Thus (x/y) x/y by the residuation property. Similarly one can prove that yx is
-closed. Conversely, suppose that x/y, yx [A] for all x [A], y A. We
have to show that (x)(y) (xy). We have xy (xy) since is a closure
operator. Thus x (xy)/y. Using the assumption on (xy)/y, we get (x)
(xy)/y. Consequently, y (x)(xy). Again using the assumption, we obtain
(y) (x)(xy). Thus (x)(y) (xy).
To see the second part of the lemma, note that the inequality (x)(y) (xy)
is equivalent to the equation ((x)(y)) = (x)(y), which says that [B] is closed
under multiplication. Indeed, suppose that (x)(y) (xy) holds. Since is idem-
potent and contracting, we have
(x)(y) = ((x))((y)) ((x)(y)) (x)(y).
Conversely, assume that ((x)(y)) = (x)(y) holds. Then (x) x and (y) y
implies (x)(y) xy. Thus we have
(x)(y) = ((x)(y)) (xy).
Consequently, satises (x)(y) (xy) iff [B] is a subsemigroup. Thus the claim
follows because 1 [B].
The characterization of nuclei by means of their images given in Lemma 3.7.1(1)
can be further improved in the sense that it is sufcient to check the condition only for
some elements. Recall that a subset D of a lattice Ais called join-dense if every element
a A is equal to a (possibly innite) join of elements from D.
LEMMA 3.7.2. Let Abe an FL-algebra and a closure operator on Awith a basis B.
Further assume that D Ais a join-dense set in A. Then is a nucleus iff the following
condition holds:
b/d and db are -closed for all d D and b B. (13)
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 313
Proof. By Lemma 3.7.1 the closure operator is a nucleus iff x/y, yx are -closed for
all x [A], y A. Clearly, this condition implies (13) since D A and B [A].
Conversely, assume that (13) holds and consider x [A] and y A. Since B is a
basis of we have x =
_
X for some X B. Similarly, y =
_
Y for some Y D
because D is join-dense. Consequently, we have yx =
_
Y
_
X =
_
dY
_
bX
db
by Proposition 3.0.2. By our assumption dbs are -closed, hence yx is -closed
as well because an arbitrary meet of -closed elements is again a -closed element.
Analogously one can show that x/y is -closed.
Let A = A, , , /, , , 0, 1 be an FL-algebra and a nucleus on A. Then the
algebra
[A] = [A],

, , /, ,

, (0), (1),
where the operations

and

are dened by
x

y = (x y),
x

y = (x y),
is called the -retraction (or a nuclear retraction) of A. Dually, if is a conucleus on
Athen the algebra
[A] = [A], ,

, /

, , (0), 1,
where the operations

, /

and

are dened by
x

y = (xy),
x/

y = (x/y),
x

y = (x y),
is called the -contraction (or a conuclear contraction) of A.
THEOREM 3.7.3 ([18]). Let A, B be FL-algebras, a nucleus on A, and a conu-
cleus on B. Then the -retraction [A] and the -contraction [B] are FL-algebras.
The notions of -retraction and -contraction for FL-chains can be simplied be-
cause any subset of a chain is closed under nite meets and joins. Namely, we have

= and

= . Indeed, let Abe an FL-chain, a nucleus on Aand x, y [A].

Then either x y or y x. Without any loss of generality assume x y. Then
(x y) = (y) = y = x y. The equality

= can be proved analogously.

To see how the above notions works, we will present examples of a conucleus and
a nucleus. Let A be an FL-algebra and : A A a map dened by (x) = x 1.
It is easy to see that is an interior operator. Moreover, is a conucleus since [A]
forms a submonoid of A. Indeed, if x, y [A] then x, y 1. Thus xy 1, i.e.,
(xy) = xy. Consequently, [A] is an FL-algebra. It is even an FL
i
-algebra as 1 is
clearly a top element of [A]. The algebra [A] is often called the negative cone of A
and denoted A

.
314 Rostislav Hor ck
Now, let B be an FL
i
-algebra and b B. Then the map : B B dened by
(x) = x b is clearly a closure operator. Moreover, is a nucleus since
(x)(y) = (x b)(y b) = xy xb by b
2
xy b = (xy).
The last inequality holds since B is integral thus xb, by, b
2
are all less than or equal to
b. Consequently, [B] is an FL-algebra which clearly remains integral.
One can also combine the above mentioned conucleus with the nucleus . Assume
that B = [A]. Then the combination gives an FL
i
-algebra [[A]] whose universe is
just the interval [b, 1] in the lattice reduct of A.
We have seen that nuclei and conuclei can be used to construct new FL-algebras. It
is also important to know which properties are preserved by these constructions. Let A
be an FL-algebra. We present two ways how to show that some property is preserved
by nuclei (resp. conuclei).
LEMMA 3.7.4. Nuclei are homomorphisms with respect to the language , , 0, 1
and images of nuclei are subalgebras with respect to the language , /, . On the
other hand, conuclei are homomorphisms with respect to the language , 0, 1 and
subalgebras with respect to the language , , 1.
Proof. The claims on subalgebras follow immediately from the denition of a nuclear
retraction and a conuclear contraction.
Let A be an FL-algebra and a nucleus on A. We claim that : A [A] is a
, , 0, 1-homomorphism. The map preserves 0 and 1 because (0) and (1) are the
corresponding constants in [A]. Let x, y A. We show that (xy) = (x)

(y) =
((x)(y)). Obviously (xy) ((x)(y)) because is expanding and monotone.
Conversely, we have (x)(y) (xy), thus ((x)(y)) ((xy)) = (xy).
Further, we show that (xy) = (x)

(y) = ((x) (y)). Again the inequality

(x y) ((x) (y)) is obvious since is expanding and monotone. Conversely,
we have (x), (y) (x y) and thus also (x) (y) (x y). Consequently,
((x) (y)) ((x y)) = (x y).
Finally, let A be an FL-algebra and a conucleus on A. We will show that is a
, 0, 1-homomorphism. The preservation of 0, 1 is obvious. Let x, y A. It remains
to check that (x y) = (x)

(y) = ((x) (y)). Since is contracting and

monotone, we have ((x) (y)) (x y). Conversely, (x y) (x), (y)
and thus also (x y) (x) (y). Consequently, (x y) = ((x y))
((x) (y)).
Since homomorphic images preserve identities and substructures preserve universal
sentences, it follows from the previous lemma that nuclei preserve all identities in the
language , , 0, 1 and all universal sentences in the language , /, . On the other
hand, conuclei preserve all identities in the language , 0, 1 and universal sentences
in the language , , 1.
Although nuclei and conuclei are not homomorphisms of FL-algebras, they can be
at least partial homomorphisms in some cases. The following two lemmas present such
cases which we will need in the sequel. Let A be an FL-algebra. The rst lemma
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 315
shows that a conucleus on A is always a partial embedding of the partial subalgebra
of -open elements into the -contraction [A].
LEMMA 3.7.5. Let A be an FL-algebra and a conucleus on A. Further, let G be
a partial subalgebra of A dened on a set of -open elements, i.e., G [A]. Then
restricted to G is a partial embedding from Gto the -contraction [A].
Proof. First note that the restriction of to G is just the identity map. Thus it is one-
to-one. Second, preserves , 0, and 1 by Lemma 3.7.4. We show that preserves
and . Recall that and are computed in [A] as in A because [A] is a subalgebra
of A with respect to the language , , 1. Let , and a, b, a b G. Then
(a) = a, (b) = b and (a b) = a b. Thus we have
(a b) = a b = (a) (b).
Finally, we show that preserves the divisions. Let a, b G. Since (a) = a and
(b) = b, we have
(ab) = ((a)(b)) = (a)

(b).
Similarly, (a/b) = (a)/

(b) for all a, b G.

Let A, B be FL-algebras and a nucleus on B. The second lemma shows that if
we are able to embed A into B via f : A B in such a way that f[A] [B], then f
becomes an embedding of Ainto the -retraction [B]. Nevertheless, the next lemma is
a bit more general than this. It is sufcient to assume that f is only a partial embedding
of a partial subalgebra of A into B. Moreover, we will formulate the lemma in such a
way that it works also for restricted languages. This will be helpful later. For instance,
we will use the following lemma in order to show that f is an embedding with respect to
all operations but . Then we can concentrate only on the preservation of which does
not follow from this general lemma.
LEMMA 3.7.6. Let A, B be FL-algebras and a nucleus on B. Further, let A
t
be a
reduct of A, G a partial subalgebra of A
t
and f : G B a partial embedding such
that f[G] [B]. Then f is a partial embedding from G to the corresponding reduct
of [B].
Proof. Since behaves like the identity map on [B], the composition f equals
f. Thus we have to show that f : G [B] preserves all the operations from
G. If the operation is one of , , 0, 1, then f preserves it because is a , , 0, 1-
homomorphismby Lemma 3.7.4. Let , /, be an operation fromGand a, b G
such that a b G. Recall that is computed in [B] in the same way as in B. Thus
we have to show that (f(a b)) = (f(a)) (f(b)). Since f[G] [B], this
equality transforms to f(a b) = f(a) f(b) which holds because f preserves by our
assumption.
316 Rostislav Hor ck
3.8 Dedekind-MacNeille completion
Completions play an important role in substructural logics. Given an FL-algebra A,
by a completion of A we mean a pair B, f where B is an FL-algebra whose lattice
reduct is complete and f : A B is an embedding. We will be mainly interested in
completions of FL-chains since it will be a crucial ingredient when proving the comple-
teness of semilinear logics with respect to the [0, 1]-valued semantics. There are many
completions of a given FL-algebra A. Among them the most important is the so-called
Dedekind-MacNeille completion because it preserves all existing meets and joins, i.e.,
f[
_
S] =
_
f[S] and f[
_
S] =
_
f[S] for any S A. The completions having the
latter property are said to be regular.
In what follows we will show how to construct the Dedekind-MacNeille completion
of an FL-algebra A into which A is embeddable and whose lattice reduct is the well-
known Dedekind-MacNeille completion of the lattice reduct of A. We will proceed
in two steps. First, we will consider the poset T(A) of all downsets of A ordered
by inclusion. In fact, T(A) forms a complete lattice. One can even extend the other
operations of the FL-algebra A to T(A) so that T(A) becomes an FL-algebra. Then
one could expect that the mapping x x is an embedding but this is not the case
because this map need not preserve nite joins. Thus we have to continue with a second
step which selects from T(A) only some of the downsets by means of a nucleus.
Let A = A, , 0, 1, be a pointed pomonoid, i.e., A, , 1, is a pomonoid and
0 A. One can dene an FL-algebra on the set T(A) of all downsets of A. The set
T(A) itself forms a complete lattice since downsets are closed under arbitrary intersec-
tions and unions. We can extend the monoid operation from Ato T(A) as follows:
X Y = u A [ (x X)(y Y )(u xy).
Note that X Y is in fact the downset generated by the complex product
X Y = x y [ x X, y Y .
To simplify the notation we write X y instead of X y and similarly x Y instead of
x Y .
We will see that is residuated with the divisions dened for X, Y, Z T(A) as
follows:
XZ = y A [ (x X)(z Z)(xy z) = y A [ X y Z,
Z/Y = x A [ (y Y )(z Z)(xy z) = x A [ x Y Z.
It is easy to see from the denition that both XZ and Z/Y are downsets. The algebra
T(A) = T(A), , , /, , , 0, 1 is called the downset monoid of A. The next
lemma shows that the downset monoid forms an FL-algebra.
LEMMA 3.8.1. The algebra T(A) = T(A), , , /, , , 0, 1 is an FL-algebra.
Proof. The multiplication is clearly associative since is, i.e., for U, V, W T(A)
we have
(U V ) W = U (V W) = z A [ (u U)(v V )(w W)(z u v w).
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 317
To see that 1 is a neutral element for , note the following:
U 1 = z A [ (u U)(z u) = U.
The last equality holds because U is a downset. Similarly we have 1 U = U.
Finally, we have to check that is residuated, i.e.,
X Y Z iff Y XZ iff X Z/Y
for all X, Y, Z T(A). Since XY = (X Y ) and Z is a downset, it follows that X
Y Z holds iff X Y Z, i.e., the rst-order sentence (x X)(y Y )(z Z)
(xy z) is valid. Taking into account that the validity of this sentence does not depend
on the order of universal quantiers, this sentence is equivalent to Y XZ and also
to X Z/Y .
Note that the construction of the downset monoid T(A) preserves some basic prop-
erties of A. Namely, it is easy to see that T(A) remains commutative if A is. Also, if
A is integral (i.e., it satises x 1), then T(A) is an FL
i
-algebra since 1 = A. If 0
is a bottom element of Athen 0 = 0 is a bottom element of T(A), i.e., T(A) is an
FL
o
-algebra. The FL-algebra T(A) is contractive if x x
2
holds in A. Indeed, given
any x X A, we have x x
2
. Thus x (X X), i.e., X X X. Further,
T(A) is clearly a chain if A is linearly ordered. Finally, it is obvious that T(A) is an
RL-algebra if 1 = 0 holds in A. Summing up, we have the following lemma.
LEMMA 3.8.2. Let Abe a pointed pomonoid. The following properties of Aare shared
by T(A):
1. commutativity xy = yx,
2. integrality x 1,
3. contractivity x x
2
,
4. 0 x,
5. linear order,
6. 1 = 0.
Let A be an FL-algebra. It has a reduct which forms a pointed partially ordered
monoid. Hence Lemma 3.8.1 applies to A. Since the lattice reduct of T(A) is complete,
it is natural ask whether T(A) could be the rst component of a completion T(A), f
of A. A natural choice for the embedding f would be f(x) = x. However, this f
need not preserve nite joins. Thus we have to modify T(A) so that f becomes also
a join-preserving map. Nevertheless, if A is an FL-chain, then f is an embedding and
T(A), f is a completion of A. However, this completion need not be regular because
T(A) could contain a downset of A which has a supremum in A. Summing up, we
have at least the following proposition.
318 Rostislav Hor ck
PROPOSITION 3.8.3. Let A be an FL-algebra. Then the map f : A T(A) dened
by f(x) = x is an embedding of the -free reduct of Ainto the -free reduct of T(A).
If, in addition, Ais linearly ordered then f is a homomorphism of FL-algebras.
Proof. Since x = y implies x = y, f has to be one-to-one. The map f clearly
preserves the constants 0 and 1. Further, f preserves multiplication because x y =
(x y) = (xy). Indeed, it sufces to note that x y has a maximum xy. It is
straightforward to check that (xy) = xy. Finally, we will show that f preserves
the divisions. We have
xy = z A [ x z y = z A [ xz y = (xy).
Similarly one can show x/y = (x/y).
To prove the additional part, let x, y A. Then x y or y x. Without any loss
of generality suppose that x y. Then (x y) = y = x y.
The next step in the Dedekind-MacNeille completion of an FL-algebra A is to
choose only some of the downsets from T(A) so that the map f(x) = x becomes
join-preserving. This can be done by a suitable nucleus. Let A = A, , 0, 1, be a
pointed pomonoid and B = B, a poset. Recall that every poset relation R AB
denes a closure operator
R
=

on T(A) (see Lemma 2.2.6 and the discussion
below it). We will characterize which of the poset relations R dene a nucleus, i.e., a
closure operator satisfying
R
(X)
R
(Y )
R
(X Y ) for all X, Y T(A).
First, note that x [ x A is a join-dense set in T(A). Indeed, we have X =

xX
x for any downset X T(A). Thus Lemma 3.7.2 can be reformulated as
follows.
LEMMA 3.8.4. Let A = A, , 0, 1, be a pointed pomonoid and a closure operator
on T(A) with a basis B. Then the following are equivalent:
1. is a nucleus on T(A).
2. C/x and xC are -closed for all x A and C B.
Let A = A, , 0, 1, be a pointed pomonoid and B = B, a poset. A poset
relation R A B is said to be nuclear if for all x, y A and z B, there exist
subsets xz, zy B such that
x y R z iff y R xz iff x R zy,
where y R xz means that y is in the relation with all the elements from xz. The
meaning of x R zy is dened analogously.
The poset relations which dene a nucleus on T(A) are exactly those which are
nuclear. Before we prove it, observe that the following hold for a poset relation R
AB, Z T(A), x A and z B:
xZ = y A [ (z Z)(xy z) = y A [ xy Z.
(z)

= a A [ (u z)(a R u) = a A [ a R z.
x(z)

= y A [ xy R z.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 319
LEMMA 3.8.5. Let A = A, , 0, 1, be a pointed pomonoid, B = B, a poset
and R A B a poset relation. The closure operator
R
on T(A) is a nucleus iff R
is nuclear.
Proof. By Lemma 2.2.7, the collection B = (z)

[ z B forms a basis of
R
.
So by Lemma 3.8.4
R
is a nucleus iff x(z)

and (z)

/x are
R
-closed for all
x A and z B. A downset x(z)

is
R
-closed iff there is a subset A B such
that x(z)

A where A = (c)

[ c xz for some subset xz B. Thus

x(z)

is
R
-closed iff the following equivalence holds:
y x(z)

iff y (c)

for all c xz.

This can be equivalently transformed to
xy R z iff y R c for all c xz.
The last equivalence is nothing but the part of the denition of a nuclear poset relation.
One can prove analogously that (z)

/x is
R
-closed iff R satises
xy R z iff x R zy.
Let A be an FL-algebra. Now we can employ Lemma 3.8.5 in order to dene a
suitable nucleus on T(A) for the construction of the Dedekind-MacNeille completion
of A. Consider the lattice order AA on A. This is clearly a poset relation which
is nuclear since there are singletons xz = xz and zy = z/y for all x, y, z A
such that
xy z i y xz i x z/y.
Thus

is a nucleus on T(A). The resulting FL-algebra

[T(A)] is called the

Dedekind-MacNeille completion of A. It is clear that the lattice reduct of

[T(A)]
is complete because the lattice reduct of T(A) is complete and the image of any clo-
sure operator on a complete lattice is complete as well. Note also that the lattice
reduct of

[T(A)] is the well-known Dedekind-MacNeille completion of the lattice

reduct of A because it consists of all downsets which are stable under taking all upper
bounds and then all lower bounds. It only remains to show that there is an embedding
f : A

[T(A)].
THEOREM 3.8.6. Every FL-algebra Acan be embedded into its Dedekind-MacNeille
completion

[T(A)].
Proof. Consider the map f : A T(A) dened by f(x) = x. By Lemma 3.8.3 f is
an embedding of the -free reduct of Ainto T(A). In order to invoke Lemma 3.7.6 we
have to show that every x is

-closed, i.e., (x)

= x. Since the relation giving

the Galois connection

,

is the order on A, we have

(x)

= a A [ a x = x,
(x)

= a A [ a x = x.
320 Rostislav Hor ck
Thus f : A

[T(A)] is an embedding of the -free reduct of A into the -free

reduct of

[T(A)] by Lemma 3.7.6. It remains to check that f preserves also . We

have to show that
(x y) = f(x y) = f(x)

f(y) =

(x y) = (x y)

.
Note that
(x y)

= u A [ u x, u y = (x y).
Thus (x y)

= (x y).
The applicability of this theorem depends also on whether the algebra

[T(A)]
remains in the same variety from which A was taken. The next theorem shows that
some properties are preserved by the construction of

[T(A)].
THEOREM 3.8.7. Let S e, c, i, o. Then the varieties F'
S
and 1'
S
are both closed
under the Dedekind-MacNeille completion. The same is true also for the classes of all
FL
S
-chains and RL
S
-chains.
Proof. The FL-algebra

[T(A)] is an FL
S
-algebra (resp. RL
S
-algebra), if A is. In-
deed, using Lemma 3.8.2, T(A) is an FL
S
-algebra (resp. RL
S
-algebra), if A is. Since
any nucleus preserves properties from S and also the identity 1 = 0 (see Lemma 3.7.4),

[T(A)] is an FL
S
-algebra (resp. RL
S
-algebra) as well. The rest of the theorem fol-
lows from the fact that a nuclear retraction of a chain is again a chain.
Although the above theorem holds also for RL
S
-algebras, note that if o S then
1'
S
is the trivial variety containing only the trivial algebra. Thus the above theorem
is trivially valid in this case. Similar observations can be made also in the following
theorems within this section.
Unfortunately, the Dedekind-MacNeille completion does not preserve the identity
axiomatizing the semilinear FL-algebras. Thus

[T(A)] need not be an FL

-algebra
if A is. However, we can use the subdirect representation theorem in order to obtain at
least some completion, although this need not be regular.
THEOREM 3.8.8. Let S e, c, i, o. Then every FL

S
-algebra (resp. RL

S
-algebra)
can be embedded into a complete FL

S
-algebra (resp. RL

S
-algebra).
Proof. Let Abe an FL

S
-algebra (resp. RL

S
-algebra). By Theorem 2.1.5 Acan be em-
bedded into

iI
A
i
where A
i
s are subdirectly irreducible and homomorphic images
of A. Thus all A
i
s belong to the semilinear variety generated by A. Moreover, all A
i
s
have to be FL
S
-chains (resp. RL
S
-chains) by Proposition 3.6.2. Using Theorem 3.8.6,
each A
i
can be embedded into its Dedekind-MacNeille completion A
i
. Thus A is em-
beddable into

iI
A
i
which clearly has a complete lattice reduct.
Another question is whether the Dedekind-MacNeille completion of an involutive
FL-algebra remains involutive. We have seen that properties like commutativity, inte-
grality etc. were preserved stepwise in the construction of

[T(A)], i.e., if A is an
FL
S
-algebra for S e, c, i, o then also T(A) and

[T(A)] are FL
S
-algebras. On
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 321
the other hand, if A is an involutive FL-algebra, T(A) need not be involutive. Never-
theless, after applying

, the resulting completion becomes involutive. Let X T(A).

We dene two sets X

= x [ x X and X

= x [ x X. The following
lemma shows how the negations X and X in T(A) (thus also in

[T(A)]) are
related to the sets X

and X

.
LEMMA 3.8.9. Let Abe an involutive FL-algebra. Then we have for all X T(A):
X = X

= X

and X = X

= X

.
Proof. We have
X = X0
= y A [ (x X)(xy 0)
= y A [ (x X)(y x)
= X

.
Also
X = y A [ (x X)(xy 0)
= y A [ (x X)(x y)
= y A [ y X

= y A [ y X

= X

.
The second part of the lemma is proved analogously.
Then it is easy to see if A is involutive so it

[T(A)] (because for any

-closed
downset X we have X = X

= X

= X). Similarly, X = X.
THEOREM 3.8.10. Let S e, c, w and A be an involutive FL
S
-algebra (resp.
RL
S
-algebra). Then its Dedekind-MacNeille completion

[T(A)] is an involutive
FL
S
-algebra (resp. RL
S
-algebra) as well.
Again note that for RL
w
-algebras the above theorem holds trivially because there is
only one RL
w
-algebra, namely the trivial one.
Theorem 3.8.6 can be also used in order to derive a structural characterization for
FL-algebras. Namely, each FL-algebra A is embeddable into a nuclear retraction
of a downset monoid T(M) for a pointed pomonoid M (namely M is the pointed-
pomonoid reduct of A). Thus if A is an FL
S
-algebra, then M satises the properties
from S (i.e., commutativity, contractivity, integrality and zero-boundedness). Moreover,
1 = 0 holds in M if Ais an RL
S
-algebra. Thus we obtain the following theorem.
THEOREM 3.8.11. Let S e, c, i, o. Then the variety F'
S
(resp. 1'
S
) equals the
class of subalgebras of nuclear retractions of downset monoids T(M), where M is a
pointed pomonoid (resp. pomonoid) satisfying the properties from S.
In addition, the class of FL
S
-chains (resp. RL
S
-chains) equals the class of subalge-
bras of nuclear retractions of downset monoids T(M), where M is a linearly ordered
pointed pomonoid (resp. linearly ordered pomonoid) satisfying the properties from S.
322 Rostislav Hor ck
4 Completeness with respect to distinguished semantics
Each semilinear variety ' of FL-algebras denes its corresponding consequence
relation [=
L
. In this section we are going to discuss whether it is possible in some cases
to replace the class of algebras ' by a smaller class K ' in such a way that [=
K
= [=
L
.
It may happen that it is not possible to have this equality for a given K. In that case we
can consider whether this equality holds when the relations [=
K
, [=
L
are restricted to
certain subsets.
Now we dene more precisely what we have meant above by the restriction of [=
K
and [=
L
. We distinguish the following kinds of completeness properties of 'with respect
to K:
1. Strong completeness (SKC): [=
L
= [=
K
.
2. Finite strong completeness (FSKC): for every nite set of identities E s = t
we have E [=
L
s = t iff E [=
K
s = t. In other words, ' is generated as a
quasivariety by K.
3. Completeness (KC): for every identity s = t we have [=
L
s = t iff [=
K
s = t. In
other words, ' is generated as a variety by K.
Clearly SKC implies FSKC which implies KC. However, none of the reverse implica-
tions holds generally.
We will be interested mainly in the rst two completeness properties. Let us recall
how to prove these properties from Chapter II. We always have [=
L
[=
K
because
K '. Moreover, [=
L
= [=
L
C
since ' is semilinear, where '
C
denotes the class of
all chains from '. Thus in order to prove SKC for ', it sufces to show [=
K
[=
L
C
.
Assume that E ,[=
L
C
s = t. Then there is a nontrivial FL-chain A from '
C
and an
evaluation e: Fm A satisfying all identities from E but not s = t. We may assume
that Ais countable since our language is countable. More precisely, we can take just the
countable subalgebra of Agenerated by the countable set
G = e() A [ is a subterm occurring in E s = t.
Consequently, if we prove that each countable nontrivial member of '
C
is embeddable
into a member Bof Kvia an embedding f : A B, we are done since the composition
f e gives an evaluation into B which satises all the premises from E but not the
conclusion s = t. Thus E ,[=
K
s = t.
THEOREM 4.0.1. Let ' be a semilinear variety of FL-algebras and K a subclass
of '. If every countable nontrivial chain from ' is embeddable into a member of K
then ' has the strong completeness property with respect to K.
In order to prove that ' has FSKC, we have to modify the above method a little bit.
Again, we assume that E ,[=
L
C
s = t but now for a nite set of identities E s = t.
Thus there is a nontrivial FL-chain Afrom'
C
and an evaluation e: Fm Asatisfying
all identities from E but not s = t. Again we suppose that A is generated by the set
G which is now nite. In order to obtain a counterexample in K, it sufces to nd a
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 323
partial embedding f : G B from the partial subalgebra G of A into a member B of
K. Then one can dene an evaluation e
t
: Fm B by setting e
t
(p) = f(e(p)), if p is a
variable occurring in E s = t, and e
t
(p) is arbitrary, if p is a variable not occurring
in E s = t. It is easy to see that the resulting evaluation e
t
satises all the identities
from E but not s = t. Thus E ,[=
K
s = t. Summing up, if we want to prove that ' has
FSKC, it is sufcient to show that every countable nontrivial chain A '
C
is partially
embeddable into K (see Denition 2.1.11).
THEOREM 4.0.2. Let ' be a semilinear variety of FL-algebras and K a subclass
of '. If every nontrivial chain from ' is partially embeddable into K then ' has the
nite strong completeness property with respect to K.
When proving completeness properties for a semilinear variety ', we will focus on
the following subclasses of ':
1. consists of all countably innite dense chains from ',
2. 1 is the class of all complete chains C from ' containing a countably innite
subset S dense in C (i.e., if x, y C such that x < y then there is s S such
that x < s < y),
3. F is the class of all nite chains from '.
The reason why we are interested in the above-mentioned types of classes comes from
the following considerations. The motivation for the second type comes fromfuzzy logic
since fuzzy logics are usually understood as logics whose intended set of truth values
is the real unit interval [0, 1]. This corresponds exactly to the second type because each
complete chain, containing a countably innite subset dense in it, is order-isomorphic
to [0, 1]. Thus each algebra from 1 is isomorphic to an algebra whose universe is [0, 1].
Note however that [0, 1] plays here only the role of an order type. Thus its bounds 0, 1
need not correspond with the interpretations of constants 0, 1 from /
FL
, i.e., there are
FL-chains A = [0, 1], , , /, , , 0
A
, 1
A
such that 0 < 0
A
and 1
A
< 1. Neverthe-
less, if A is an FL
i
-chain, then 1
A
= 1. Analogously, 0
A
= 0 if A is an FL
o
-chain.
Moreover, we have both if Ais an FL
w
-chain.
On the other hand, the rst type corresponds to the semantics whose set of truth
values forms an interval in the set of rational numbers. The rst type will serve as an
intermediate step when proving a completeness property with respect to the second type.
Finally, the last type is motivated mainly by an application point of view, namely if '
has FSFC then the set of quasi-identities valid in ' is decidable. Similarly, SFC implies
decidability of the equational theory of '. Note that FSFC is usually called the nite
embeddability property (FEP) in the literature.
Concerning the structure of the following text on the completeness properties, we
will split it into two parts. The rst part contains positive results on completeness proper-
ties for integral semilinear varieties (i.e., semilinear varieties containing FL
i
-algebras).
The second part deals with non-integral varieties where we have mainly negative results.
324 Rostislav Hor ck
4.1 Integral semilinear varieties
4.1.1 Completeness with respect to countably innite dense chains and
[0, 1]-valued semantics
Let A, be a chain and a, b A. We denote the fact that a is a subcover of b as
a b, i.e., a b holds iff a < b and there is no c A such that a < c < b. A chain
A, is said to be dense if a b does not hold for any a, b A. Note that the trivial
one element chain is dense by this denition.
Let i S e, c, i, o and F'

S
the corresponding variety of semilinear FL
S
-
algebras. First, we will focus on SC for F'

S
. It is obvious that each chain from is
isomorphic to an algebra whose universe is either the set Q[0, 1] or Q(0, 1] depending
on the existence of a bottom element.
By Theorem 4.0.1 in order to show SC for F'

S
, it is sufcient to prove that
each countable nontrivial FL
S
-chain Acan be embedded into a countably innite dense
FL
S
-chain D. Suppose that we have an FL
S
-chain A = A, , , /, , , 0
A
, 1
A
which
is countable and nontrivial. If Ais dense then A has to be innite because Ais nontriv-
ial. Thus we can set D = A. If Ais not dense then there is at least one element a which
has a subcover a
t
. As we want to extend Aso that it becomes dense, we have to ll for
each such element a the gap between a and a
t
by a countable dense chain. This can be
done by pasting a copy of rational numbers (namely Q (0, 1)) into the gap between
a and a
t
(see Figure 5). Formally we can dene the set D as the following subset of
A(Q (0, 1]):
D = a, 1 [ a Aa, q [ q Q(0, 1) and (a
t
A) such that a
t
a. (14)
Then the lexicographic order
lex
on D is a dense linear order and 1
A
, 1 is a top
element. If 0
A
is a bottom element of Athen 0
A
, 1 is a bottom element of D because
0
A
has no subcover in this case. Moreover, the subset A1 D is order-isomorphic
to A.
Observe that we can dene two operators on the chain D = D,
lex
whose image
is A1, namely a closure operator and an interior operator dened as follows:
(a, q) = a, 1, (15)
(a, q) =
_
a, 1 if q = 1,
a
t
, 1 if q < 1 and a
t
a.
(16)
Note that both operators behave like the identity map on A1. Their behavior on the
remaining elements is depicted in Figure 5.
It is easy to see that is a closure operator on D = D,
lex
. To see that is
an interior operator on D, note that is idempotent and contracting. Thus it sufces to
show that is monotone. Assume that a, q
lex
b, p. If p = 1 then (a, q)
lex
a, q
lex
b, 1 = (b, 1). If p < 1 then there is b
t
A such that b
t
b. Now
there are two cases. First, if a < b then a b
t
. Consequently, (a, q)
lex
a, q
lex
b
t
, 1 = (b, p). Second, if a = b then q p and so (a, q) = b
t
, 1 = (b, p). Thus
is really an interior operator. Summing up, if we identify A with A 1, we obtain
following general lemma.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 325
a
t
a
=
(x)
(x)
x
Q (0, 1)
A D
Figure 5. Filling gaps.
LEMMA 4.1.1. Let A = A, , , 0, 1 be a countable nontrivial chain with a top ele-
ment 1. Then Acan be extended to a countably innite dense chain D = D, , , 0, 1
with a top element 1. Moreover, 0 is a bottom element of D if 0 is a bottom element
of A. There are also a closure operator and an interior operator on D such that
A = [A] = [A].
We have the dense chain D. The next step is to extend the multiplication on the
FL
S
-chain Ato D. In the light of Lemma 4.1.1 we will do this more generally, namely
we do not restrict ourselves to the particular set D dened by (14) but we will consider a
more abstract setting where we have a chain D = D, , , 0, 1 together with constants
0, 1 endowed with a closure operator and an interior operator such that [D] = [D]
and [D] forms an FL
S
-chain A = A, , , /, , , 0, 1, i.e., A = [D]. Note that in
this setting 1 has to be a top element in D. Indeed, if there would be x D such that 1 <
x then 1 < (x) Awhich would mean that Ais not integral. Also, if 0 is a bottom ele-
ment of Athen 0 has to be a bottom element in Dbecause x < 0 implies 0 > (x) A.
The rst natural idea to extend the multiplication from A to the whole of D, is to
dene the multiplication on Das x y = (x) (y), i.e., an element x which does not
belong to A behaves like (x). This works only partially since 1 x = (1) (x) =
1(x) = (x), i.e., 1 is not a neutral element of . However, we still have the following:
LEMMA 4.1.2. Let C = C, , be a chain, a closure operator on C and an
interior operator on C such that [C] = [C]. Further assume that the image [C]
forms a residuated lattice-ordered semigroup [C] = [C], , , /, , . Then the
algebra C = C, ,
C
, /
C
, , is a residuated lattice-ordered semigroup, where
x y = (x) (y), x/
C
y = (x)/(y), x
C
y = (x)(y).
Moreover, C is commutative if [C] is.
326 Rostislav Hor ck
Proof. The operation is clearly associative since has this property. Thus C is a
semigroup. Moreover, C is clearly commutative if [C] is. Thus it sufces to prove that
C is residuated. Suppose that x y = (x) (y) z. Since (x) (y) is -open, we
have (x) (y) = ((x) (y)) (z). Consequently, x (x) (z)/(y) =
z/
C
y. Conversely, suppose that x z/
C
y = (z)/(y). Since (z)/(y) is -closed,
we have (x) ((z)/(y)) = (z)/(y). Consequently, x y = (x) (y)
(z) z. Analogously for the left division. Thus /
C
and
C
are the residuals of .
Applying Lemma 4.1.2 to our dense chain D, we obtain a residuated lattice-ordered
semigroup D = D, ,
D
, /
D
, , . However, as we mentioned above, is not a
monoid operation because 1 is not a neutral element. In particular, 1 x = (x) x,
i.e., the result of 1 x could be greater than we need. Thus we have to further modify .
Note that the chain D also forms an FL
ci
-chain D

= D, , , , , 0, 1, where
x y =
_
1 if x y,
y otherwise.
(17)
In fact, the minimum operation on D is the greatest among all integral -monoid
operations on D. Thus it seems to be natural to lessen the values of by a combination
with . Precisely, we dene a new operation on D, which will have all the necessary
properties, as follows:
x y = (x y) x y.
The next two lemmas show that has the desired properties. In the following lemma
when we have a groupoid operation on a semilattice we adopt a convention that the
groupoid operation binds stronger than the semilattice operation.
LEMMA 4.1.3. Let M, , , 1 be an algebra such that M, , 1 is a meet-semilattice
with a top element 1 and M, a semigroup satisfying:
1. the operation distributes over , i.e., x (y z) = x y x z, (y z) x =
y x z x,
2. 1 x x = x 1 x = x.
Then M, , 1 is a monoid, where x y = x y x y, and is commutative if is.
Proof. First, we check that 1 is a neutral element. We have 1 x = (1 x) 1 x = x.
Analogously x 1 = x. Second we prove that is associative. We have
(x y) z = (x y x y) z
= (x y x y) z x y x y z
= x y z x z y z x y x y z.
Similarly,
x (y z) = x (y z y z)
= x (y z y z) x y z y z
= x y z x y x z x y z y z.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 327
Since is commutative, we get (x y) z = x (y z). Consequently is a monoid
operation on M. The last claim about commutativity follows immediately from the
denition of .
LEMMA 4.1.4. Let i S e, c, i, o, D

= D, , , , , 0, 1 an FL
ci
-chain
and a closure operator on D

such that [D] = [D], , , /, , , 0, 1 is an FL

S
-
chain. Further, assume that is an interior operator on D

such that [D] = [D].

Then the algebra D = D, ,
D
, /
D
, , , 0, 1 is an FL
S
-chain where
x y = (x) (y) x y,
x
D
y = (x)(y) x y,
x/
D
y = (x)/(y) y x.
Moreover, the operations ,
D
, /
D
, when restricted to [D], coincide respectively with
, , /, i.e., [D] is a subalgebra of D.
Proof. Let denote the operation on D dened by xy = (x) (y). By Lemma 4.1.2
this operation is associative and monotone because it is residuated. Since D is a chain,
the monotonicity of is equivalent to its distributivity over . Moreover, since 1 [D],
we have (1) = 1. Thus 1x = (1)(x) = 1(x) = (x) x. Similarly, x1 x.
Consequently, Lemma 4.1.3 applies to since x y = x y xy. Thus is a monoid
operation on D.
Further, it follows from Lemma 4.1.2 that is residuated with the divisions given by
(x)(y) and (x)/(y). We have to show that the join combination of these divisions
with gives the divisions for . We have the following chain of equivalences:
x y = x y x y z i x y z or x y z
i x (z)/(y) or x y z
i x (z)/(y) y z = z/
D
y.
Analogously we obtain x y z iff y x
D
z. Thus D forms an FL-chain. It is even
an FL
S
-chain because 1 is the top element and 0 is the bottom element of D if it is a
bottom element of [D] (see the discussion after Lemma 4.1.1). Further, D is clearly
commutative if [D] is. Finally, if [D] is contractive (i.e., x y = (x) (y)) then
x y = (x) (y) x y = x y.
Thus D is contractive as well.
Finally, we have to show that the restrictions of operations ,
D
, /
D
on [D] be-
have like , , /. Since [D] forms an FL
i
-chain, we have (x) (y) (x) (y).
Thus
(x) (y) = (x) (y) (x) (y) = (x) (y).
For the right division we have
(x)/
D
(y) = ((x))/((y)) (y) (x) = (x)/(y) (y) (x).
328 Rostislav Hor ck
Using again that [D] forms an FL
i
-chain, we obtain
((y) (x)) (y) ((y) (x)) (y) = (y) ((y) (x)) (x).
Thus by residuation (y) (x) (x)/(y). Consequently, we have (x)/
D
(y) =
(x)/(y). Similarly, we can prove (x)
D
(y) = (x)(y).
Now using Lemma 4.1.1 together with Lemma 4.1.4 we can prove that every count-
able nontrivial FL
S
-chain A(where i S e, c, i, o) is embeddable into a count-
ably innite dense FL
S
-chain D. Moreover, it is clear that the construction of D pre-
serves the identity 1 = 0. Thus the next theorem follows by Theorem 4.0.1.
THEOREM 4.1.5. Let i S e, c, i, o. Then the variety F'

S
has the strong
completeness property with respect to the class of all countably innite dense FL
S
-
chains. The same is true also for the variety 1'

S
if i S e, c, i.
Thus we have proved SC for F'

S
. Now it is easy to extend this result to S1C us-
ing the Dedekind-MacNeille completion. Let Abe a countably innite dense FL
S
-chain
for S e, c, i, o. Then Aembeds into its Dedekind-MacNeille completion A
t
which
is an FL
S
-chain as well (see Theorems 3.8.6 and 3.8.7). Then A
t
is order-isomorphic
to [0, 1] because A
t
is a complete chain containing a countably innite subset A dense
in A
t
(the density follows from the well-known properties of the Dedekind-MacNeille
completion). The same procedure can be done also with any countably innite dense
RL
S
-chain for S e, c, i. Thus we get the following theorem.
THEOREM 4.1.6. Let i S e, c, i, o. Then the variety F'

S
has the strong
completeness property with respect to the class of all FL
S
-chains on [0, 1]. The same is
true also for the variety 1'

S
if i S e, c, i.
4.1.2 Involutive algebras
So far we have not discussed the completeness properties for involutive FL-algebras.
We still assume that our involutive FL-algebras are integral. Recall that involutive
FL
i
-algebras are in fact involutive FL
w
-algebras since integrality in this case implies
that 0 is a bottom element. Hence, in the remaining part of this section we will work
with involutive FL
S
-algebras where w S e, w. The contractivity (c) is omitted
since the only involutive FL
cw
-chain is the two element Boolean algebra. Thus the class
of involutive FL

cw
-algebras is the variety of Boolean algebras which has neither C
nor 1C. It also makes no sense to deal with involutive RL
w
-algebras because there is
only the trivial RL
w
-algebra.
Before we prove that the variety "nF'

S
enjoys SC, we present a general lemma
showing how to modify an FL
w
-chain so that it becomes involutive.
LEMMA 4.1.7. Let A = A, , , /, , , 0, 1 be an FL
w
-chain and , an involu-
tive pair on Asatisfying yx = yx for all x, y A, where
xy = xy x, xy = x/y y.
Then A
i
= A, , , , , , 0, 1 is an involutive FL
w
-chain, where xy = (yx).
Moreover, if Ais commutative and = then A
i
is commutative as well.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 329
Proof. Note that and in this lemma do not denote the negations in A. Further
observe that 1 = 0 = 1 because and are order-reversing bijections, 1 is a top
element and 0 a bottom element. Using Theorem 3.2.2, it sufces to show that A
i
is an
associative involutive division lattice whose unit is 1.
First, we show that (4) holds for , . Since Ais a chain, we have
y xz = xz x i y xz or y x
i x z/y or x y
i x z/y y = zy.
Second, using the assumption that , satises the contraposition law yx =
yx, we can prove the associativity for and . Observe that in any FL-chain the
following distributive laws holds:
a(b c) = ab ac, (b c)/a = b/a c/a.
Thus we have the following chain of equations:
x(zy) = x(z/y y) x = x(z/y) xy x = x(z/y) xy
= (xz)/y xy = (xz)/y x/y y = (xz x)/y y = (xz)y.
Finally, we check that 1 is the unit for and . We have
1x = 1x 1 = x 0 = x.
Similarly x1 = x.
To see the moreover part, observe that if A is commutative then xy = y/x for all
x, y A. Assuming = , we obtain xy = xy x = y/x x = yx. Thus
A
i
is commutative because by the contraposition law yx = yx we have
x y = (yx) = (yx) = (yx) = (xy) = y x.
Let A = A, , , /, , , 0, 1 be a nontrivial countable involutive FL
S
-chain where
w S e, w and , the involutive pair given by the negations on A. To
prove SC for "nF'

S
, we have to embed A into a countably innite dense involutive
FL
S
-chain (see Theorem 4.0.1). We know from the previous section (see Lemmas 4.1.1
and 4.1.4) that A can be embedded into a countably innite dense FL
S
-chain D =
D, ,
D
, /
D
, , , 0, 1 endowed with a closure operator and an interior operator
such that A = [D] = [D]. However, Dneed not be involutive. Namely, the negations
in D are computed as follows (note that (0) = 0 = (0) and x 0 = 0 for x > 0
and 0 0 = 1):

D
x = x
D
0 = (x)(0) x 0 = (x) x 0 = (x),

D
x = 0/
D
x = (0)/(x) x 0 = (x) x 0 = (x).
Consequently, the negations do not form an involutive pair because

D
x = ((x)) = (x) = (x) =
D

D
x.
330 Rostislav Hor ck
1
0
0
a b c 1
Figure 6. The involutive modication of in D. The bold line denotes
D
and the
dashed line its involutive modication
D
i
.
The second equality follows from the fact that [D] is closed under all operations from
A, in particular under . An example of this situation is depicted in Figure 6, where
0, a, b, c, 1 is the set of -closed elements.
In what follows we will modify both divisions
D
, /
D
so that the corresponding
negations become involutive and their restrictions on the -closed elements will behave
like the original negations from A (see Figure 6). Recall that the dense chain D was
dened in (14) as the set
D = a, 1 [ a A a, q [ q Q (0, 1) and (a
t
A) such that a
t
a.
Inside D we identify A with A 1 which is the image of the closure operator
and the interior operator on D(see (15) and (16)). First, we show that it is possible to
construct an involutive pair on Dwhose restrictions to the -closed and -open elements
behave like the original negations , from A.
LEMMA 4.1.8. Let
D
i
and
D
i
be unary operations on D dened as follows:

D
i
a, q =
_
a, 1 if q = 1,
a
t
, 1 q if q < 1,

D
i
a, q =
_
a, 1 if q = 1,
a
t
, 1 q if q < 1,
where a
t
is the unique subcover of a. Then the pair
D
i
,
D
i
is an involutive pair
such that
D
i
a, 1 = a, 1 and
D
i
a, 1 = a, 1. Moreover,
D
i
=
D
i
if
= .
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 331
Proof. Let a, q b, r. We show that
D
i
is order-reversing, i.e.,
D
i
b, r

D
i
a, q. If a = b then q r. Consequently, 1 r 1 q. Thus
D
i
b, r

D
i
a, q holds in this case. Now suppose that a < b. Depending on the values of q, r
we have several cases. If q < 1 then a
t
< b and a
t
< b
t
if the subcover b
t
of b exists.
Thus a
t
> b and also a
t
> b
t
if b
t
exists. Consequently,
D
i
b, r
D
i
a, q.
If q = r = 1 then a > b and the desired inequality again holds. The last possible
case is q = 1 and r < 1. If a < b
t
then the desired inequality follows as before. If
a = b
t
then
D
i
b, r = a, 1 r < a, 1 =
D
i
a, q. Similarly, one can prove
that
D
i
is order-reversing.
Finally, we check that
D
i

D
i
a, q = a, q. The proof of
D
i

D
i
a, q =
a, q is analogous. If q = 1 then it clearly holds since
D
i

D
i
a, 1 = a, 1 =
a, q. Suppose that q < 1. Then a has a unique subcover a
t
, i.e., a
t
a. Since is an
order-reversing bijection, we have x y iff y x. Thus a a
t
, i.e., a is the
unique subcover of a
t
. Consequently, we obtain

D
i

D
i
a, q = a
t
, 1 q = a, 1 (1 q) = a, q.
The moreover part of the lemma is obvious.
Lemma 4.1.8 implies that there is an involutive pair
D
i
,
D
i
on the countably
innite dense FL
S
-chain D such that on -closed and -open elements coincides with
the original negations fromA; recall the denitions of and from (15) and (16). Thus
we have

D
i
(x) = (x),
D
i
(x) = (x), (18)

D
i
(x) = (x),
D
i
(x) = (x). (19)
Now we want to use the involutive pair
D
i
,
D
i
from the previous lemma in
order to make D an involutive FL
S
-chain. In the light of Lemma 4.1.7 it is sufcient
to show that the involutive pair
D
i
,
D
i
satises the contraposition law y
D
i
x =

D
i
yx, where
xy = x
D
y
D
i
x, xy = x/
D
y
D
i
y.
LEMMA 4.1.9. We have
D
i
(x) = (
D
i
x) and
D
i
(x) = (
D
i
x).
Proof. Since x (x), we have
D
i
(x)
D
i
x. Thus
D
i
(x) (
D
i
x)
because
D
i
(x) = (x) is -open. On the other hand, we have (
D
i
x)
D
i
x.
Hence x
D
i
(
D
i
x). Since
D
i
(
D
i
x) = (
D
i
x) is -closed, we obtain
(x)
D
i
(
D
i
x). Consequently, (
D
i
x)
D
i
(x). The other equality is
proved analogously.
LEMMA 4.1.10. The involutive pair
D
i
,
D
i
satises the contraposition law
y
D
i
x =
D
i
yx.
332 Rostislav Hor ck
Proof. Recall from Lemma 4.1.4 that
y
D
x = (y)(x) y x, y/
D
x = (y)/(x) x y,
where is the residual of from D

. Thus we have
y
D
i
x = (y)(
D
i
x) (y
D
i
x)
D
i
y,

D
i
yx = (
D
i
y)/(x) (x
D
i
y)
D
i
x.
Since y
D
i
x iff x
D
i
y, we have 1 y
D
i
x iff 1 x
D
i
y.
Thus y
D
i
x and
D
i
yx both equal 1 if y
D
i
x holds. Assume that y >
D
i
x.
Then also x >
D
i
y. Consequently, y
D
i
x =
D
i
x and x
D
i
y =
D
i
y.
Thus it sufces to show that (y)(
D
i
x) = (
D
i
y)/(x). By Lemma 4.1.9 and
the equations (18), (19) we obtain
(y)(
D
i
x) = (y)
D
i
(x) = (y)(x) = (y)/(x) = (
D
i
y)/(x).
The last but one equality follows from the fact that , forms an involutive pair on
A = [D]. Thus , has to satisfy the contraposition law yx = y/x.
Now using Lemma 4.1.7 together with Lemma 4.1.10, we can modify the countably
innite dense FL
S
-chain D so that D
i
= D, , , , , , 0, 1 becomes an involutive
FL
S
-chain where
xy = x
D
y
D
i
x, xy = x/
D
y
D
i
y, x y =
D
i
(y
D
i
x).
Now it remains to prove that A is a subalgebra of D
i
. Recall that A = [D]. We
will show that the restrictions of , , to [D] coincide respectively with , /, . Let
(x), (y) [D]. Since A is a subalgebra of D (see Lemma 4.1.4), we obtain using
also (18) the following:
(x)(y) = (x)
D
(y)
D
i
(x) = (x)(y) (x) = (x)(y).
The last equality holds because in any FL
w
-algebra we have xy x0 = x. Simi-
larly, we can prove (x)(y) = (x)/(y). Finally, since A is involutive, the multi-
plication in Ais denable by , and as u v = (vu). Thus using the fact that
(x) and (y)(x) are -closed, we get using Lemma 3.2.1 and (18) the following
chain of equations:
(x) (y) =
D
i
((y)
D
i
(x)) = ((y)(x)) = (x) (y).
Consequently, Ais a subalgebra of D
i
and the next theorem follows.
THEOREM 4.1.11. Let w S e, w. Then the variety "nF'

S
has the strong
completeness property with respect to the class of all countably innite dense involutive
FL
S
-chains.
Using again the Dedekind-MacNeille completion (see Theorem 3.8.10), one can
easily derive the following theorem.
THEOREM 4.1.12. Let w S e, w. Then the variety "nF'

S
has the strong
completeness property with respect to the class of all FL
S
-chains on [0, 1].
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 333
4.1.3 Completeness with respect to nite algebras
Now we will turn our attention to the completeness properties with respect to the
class of nite algebras. We will focus on the nite strong completeness property because
one usually cannot hope for the strong completeness property in this case.
Before we start, we have to recall several results on dual well partial orders. Let P
be a poset. We say that P is a dual well partial order (shortly dwpo) if P contains
no innite ascending chain and no innite anti-chain. Dual well partial orders can be
characterized by means of good sequences. A sequence p = p
i
P [ i N is said to
be good if there are natural numbers n < m such that p
n
p
m
.
LEMMA 4.1.13 ([36]). A poset P is a dwpo iff every innite sequence of elements from
P is good.
The next two lemmas show how to construct new dwpos from given dwpos.
LEMMA 4.1.14. The direct product of two dwpos is a dwpo.
Let P = P, be a poset. By P

we denote the set of all nite sequences of

elements from P, i.e., P

forms a free monoid generated by P. One can introduce a

partial order _ on P

as follows:
a
0
, . . . , a
m
_ b
0
, . . . , b
n
iff there exist i
0
, . . . , i
n
such that
0 i
0
< i
1
< < i
n
m and a
i
k
b
k
for k = 0, . . . , n.
The following lemma is just a reformulation of the well-known result by Higman [24].
LEMMA 4.1.15 (Higmans lemma). Let P = P, be a dwpo. Then P

= P

, _
is a dwpo as well.
For us it will be important to know that images of dwpos, which are linearly ordered,
are dually well ordered as is proved in the following lemma.
LEMMA 4.1.16. Let P be a dwpo, Q a chain and f : P Q an order-preserving
surjective map. Then Qis dually well ordered.
Proof. Assume not. Then there is an ascending sequence q
0
< q
1
< q
2
< of
elements from Q. Take any sequence p = p
i
P [ i N such that f(p
i
) = q
i
. Since
f is order-preserving, we have p
n
, p
m
for any n < m. Thus p is not good which is
not possible by Lemma 4.1.13.
We have presented all the necessary results on dwpos. Let i S e, c, i, o.
We are going to show that F'

S
enjoys FSFC. In order to invoke Theorem 4.0.2 we
have to show that every nontrivial FL
S
-chain A is partially embeddable into the class
of all nite FL
S
-chains, i.e., for every nite partial subalgebra G of A we have to nd
a partial embedding from G into a nite FL
S
-chain B. We can assume without any
loss of generality that Ahas a bottom element . If not, then one can embed Ainto its
Dedekind-MacNeille completion which has a bottom element because its lattice reduct
is complete. We may also assume that G contains , 0, and 1 because the restriction of
a partial embedding f : G , 0, 1 B to G is a partial embedding as well.
334 Rostislav Hor ck
Let M be the submonoid of Agenerated by G. Then M is clearly linearly ordered
and contains , 0, 1. Thus M = M, , 0, 1, is a pointed pomonoid with the bottom
element . Using Higmans lemma, we can prove the following lemma.
LEMMA 4.1.17. The pointed pomonoid M is dually well ordered.
Proof. The set G is nite and linearly ordered by the order on A. Thus G forms
a dwpo. By Higmans lemma G

, _ is a dwpo as well. Moreover, G

is a free
monoid. Thus there is a surjective monoid homomorphism h: G

M such that
h(a
0
, . . . , a
m
) = a
0
a
m
. If we show that h is order-preserving, then the claim
follows by Lemma 4.1.16. Suppose that a
0
, . . . , a
m
_ b
0
, . . . , b
n
, i.e., there are
i
0
, . . . , i
n
such that 0 i
0
< i
1
< < i
n
m and a
i
k
b
k
. Since h is a monoid
homomorphism and Ais integral, we have
h(a
0
, . . . , a
m
) = a
0
a
m
a
i
0
a
i
n
b
0
b
n
= h(b
0
, . . . , b
n
).
It follows from Lemma 4.1.17 that M is the image of a conucleus on A. Indeed,
M induces an interior operator (x) = maxa M [ a x. The maximum always
exists because M is dually well ordered and contains the bottom element . Since M is
a submonoid of A, is a conucleus by Lemma 3.7.1. Consequently, M forms an FL-
algebra, namely the -contraction [A] = [A], ,

, /

, , , 0, 1, where [A] =
M. Note that the constant 0 is not modied in [A] because 0 M. The meet need
not be modied as well because Ais linearly ordered. Indeed, let x, y [A]. Without
any loss of generality we may assume that x y. Thus (x y) = (x) = x = x y.
Moreover, [A] is an FL
S
-chain by Lemma 3.7.4.
Now using Lemma 3.7.5, there is a partial embedding f : G [A] from G into
[A] because G M = [A]. The next step is to nd a nite subset B [A] which
forms an image of a nucleus on [A]. Let B be the subset of [A] dened as follows:
B = a

c/

b [A] [ a, b [A], c G =
_
cG
a

c/

b [A] [ a, b [A].
In the above denition we are writing terms of the form a

c/

b without parentheses
because (ac)/b = a(c/b) holds in any FL-algebra.
LEMMA 4.1.18. The subset B is nite.
Proof. Clearly, since G is nite, it is sufcient to prove that B
c
= a

c/

b [A] [
a, b [A] is nite for a xed c G. Assume that B
c
is not nite. Since [A] = M
is dually well ordered, B
c
has to contain an innite descending chain a
0

c/

b
0
>
a
1

c/

b
1
> . This chain denes a sequence p = a
i
, b
i
M
2
[ i N. By
Lemma 4.1.14 the direct product M
2
, forms a dwpo. Thus p has to be good, i.e.,
there are n < m such that a
n
, b
n
a
m
, b
m
. Consequently, we have
a
m
(a
n

c/

b
n
)b
m
a
n
(a
n

c/

b
n
)b
n
c.
Hence by residuation we obtain a
n

c/

b
n
a
m

c/

b
m
; a contradiction with the
fact that a
0

c/

b
0
> a
1

c/

b
1
> is a descending chain.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 335
Now we have to show that the set B is the image of a nucleus . First, note that
G B because g = 1

g/

1 for every g G. In particular, 1 B. Thus B induces

a closure operator (x) = mina B [ x a. The minimum always exists because
B is nite and 1 B is the top element. In order to invoke Lemma 3.7.1 we have to
show that x/

y, y

x B for x B and y [A]. Let x B and y [A]. Then

x = a

c/

b for c G and a, b [A]. Thus x/

y = ((a

c)/

b)/

y = a

c/

yb
by Lemma 3.0.5. Consequently, x/

y B. Analogously, y

x B. Thus B forms an
FL-algebra, namely the -retraction [[A]]. Moreover, [[A]] remains an FL
S
-chain
by Lemma 3.7.4.
Since G B, Lemma 3.7.6 implies that the partial embedding f : G [A] is in
fact a partial embedding of G into [[A]]. Thus the next theorem follows. Note that
the identity 1 = 0 is preserved by and . Thus we get FSFC also for 1'

S
.
THEOREM 4.1.19. Let i S e, c, i, o. The variety F'

S
(resp. 1'

S
) has the
nite strong completeness property with respect to the class of all nite FL
S
-chains
(resp. RL
S
-chains).
4.1.4 Cyclic involutive algebras
Now we are going to discuss whether the proof of Theorem 4.1.19 can be used
also for involutive FL

i
-algebras. Recall that involutive FL
i
-chains are in fact involutive
FL
w
-chains and there is only one nontrivial involutive FL
cw
-chain, namely the two
element Boolean algebra 2. Since the variety of Boolean algebras (i.e., the semilinear
variety "nF'

cw
) obviously enjoys the strong completeness property with respect to nite
chains (SFC), we will focus only on involutive FL
S
-algebras for w S e, w.
Let A be a nontrivial involutive FL
S
-chain where w S e, w. The con-
struction of the nite FL
S
-chain [[A]] from the proof of Theorem 4.1.19 need not
preserve the double negation law x = x = x. Thus the above method cannot
be used in order to prove that the variety of involutive FL

S
-algebras has FSFC. Nev-
ertheless the method works at least for cyclic involutive FL

S
-algebras. Thus assume
further that A is also cyclic, i.e., it satises x0 = x = x = 0/x. Let G be a nite
partial subalgebra of A. Without any loss of generality we will assume that 0, 1 G
(observe that 0 is a bottom element) and G is closed under the negations , . This
does not affect niteness of G because A is cyclic (i.e., x = x) and involutive (i.e.,
x = x = x).
We start with a general lemma on cyclic FL-algebras. Let Abe an FL-algebra. We
call an element c A involutive if c = c = c. The next lemma shows that
having an involutive element c in a cyclic FL-algebra A, one can nd other involutive
elements in A.
LEMMA 4.1.20. Let Abe a cyclic FL-algebra and c A an involutive element. Then
ac/b is involutive as well for all a, b A.
Proof. Since A is cyclic (i.e., = ), it sufces to prove (ac/b) = ac/b. We
rst show that ac/b = (b(c)a). Using Lemma 3.0.6, we obtain
(b(c)a) = ((c)a)/b = ((c)a)/b = a(c)/b = ac/b.
336 Rostislav Hor ck
Consequently, again using Lemma 3.0.6 and the latter equation, we have
(ac/b) = (b(c)a) = (b(c)a) = ac/b.
In the same way as in the proof of Theorem 4.1.22 we construct the nite algebra
[[A]], where [A] is the submonoid of Agenerated by G and
[[A]] = a

c/

b [A] [ a, b [A], c G.
Moreover, we know from the proof of Theorem 4.1.19 that there is a partial embedding
of Ginto [[A]]. Thus it sufces to prove that [[A]] is cyclic and involutive.
LEMMA 4.1.21. The nite FL
S
-chain [[A]] is cyclic and involutive.
Proof. First, observe that 0 G [[A]]. Thus [[A]] is a subalgebra of [A]
not only with respect to the language , , / (see Lemma 3.7.4) but also , , /, 0.
Thus the negations in [[A]] are just the restrictions of the negations from [A]. Con-
sequently, in order to check that [[A]] is cyclic and involutive, it sufces to show
that [A] is cyclic and [[A]] contains only involutive elements from [A]. Let

denote the negations in [A] (the symbols , stand for the negations in A). The
cyclicity of [A] is easy because for x [A] we have

x = x

0 = (x0) = (x) = (x) = (0/x) = 0/

x =

x.
Thus [A] is cyclic. Consequently, [[A]] is cyclic.
To see that [[A]] is involutive, note that every c G satises c = c because
Ais involutive. Since G is closed under negations and all elements from G are -open,
we have

c = (c) = c. Consequently, we obtain

c = (c) = (c) = c.
Thus every c Gis involutive in [A]. Consequently, since every element from[[A]]
is of the form a

c/

b for c G and a, b [A], all the elements from [[A]] are

involutive in [A] by Lemma 4.1.20. Thus [[A]] is involutive.
THEOREM 4.1.22. Let w S e, w. The variety Cy"nF'

S
of semilinear cyclic
involutive FL
S
-algebras has the nite strong completeness property with respect to the
class of all nite cyclic involutive FL
S
-chains.
4.2 Non-integral semilinear varieties
Now we turn our attention to the semilinear varieties of FL-algebras which are not
integral (i.e., they contain also non-integral FL-algebras). Unlike the previous part on
integral varieties, we will present mainly negative results on completeness properties
showing that most of the non-integral varieties we are discussing here do not enjoy
the completeness properties with respect to the previously considered classes of alge-
bras. As before we start with the completeness properties with respect to the class of all
countably innite dense chains and chains on [0, 1].
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 337
4.2.1 Completeness with respect to countably innite dense chains and
[0, 1]-valued semantics
It turns out that the density of an FL-chain A forces A to satisfy an extra identity
which is not generally valid in FL-algebras. In particular, there are non-integral FL-
chains where this identity does not hold.
LEMMA 4.2.1. Every dense FL-chain Asatises the following identity
1 (zy) ((xy)(1/(zx))). (20)
Proof. If z y then 1 zy and the identity is clearly valid. Thus assume y < z which
is equivalent to zy < 1. Then we have to show that 1 (xy)(1/(zx)) which, by
the residuation property, is equivalent to (xy) (zx) 1. We will prove the latter
inequality by reductio ad absurdum. Let a = xy and b = zx. To get a contradiction,
assume that a b > 1. Observe that by Lemma 3.0.5 we have b a = (zx) (xy)
zy < 1. This means that one of a, b is strictly greater than 1 and the other one strictly
less. Without any loss of generality suppose that a > 1 and b < 1. Observe that b
cannot be a bottom element because u = u = for all u A. Thus the set
M = u A [ u < b is not empty. Moreover, since A is dense, we get
_
M = b.
Consequently, 1 < a b = a
_
M =
_
uM
(a u). Thus there is u
t
M such that
1 a u
t
. Now we have the following chain of inequalities:
b (a b) = (b a) b b b (a u
t
) b (a b).
Thus b (a u
t
) = b. On the other hand, (b a) u
t
u
t
< b which is a contradiction
since is associative.
The above lemma shows that every dense chain has to satisfy the identity 1
(zy) ((xy)(1/(zx))). In particular, it holds in every FL-chain whose universe is
[0, 1]. However, there is a four element FL-chain where this identity is not valid as is
shown in the following lemma.
LEMMA 4.2.2. There is an RL
c
-chain A
4
and an FL
co
-chain B
4
where the following
equation does not hold:
1 (zy) ((xy)(1/(zx))).
Proof. Let A
4
= , a, 1, ordered by < a < 1 < (see Figure 7). The
multiplication is dened as follows:
a 1

a a a
1 a 1
a
It is straightforward to check that is associative. Moreover, it is clearly residuated
as A
4
is nite. Thus A
4
= A
4
, , , /, , , 1 is an RL-chain. Moreover, A
4
is an
338 Rostislav Hor ck

a
1

A
4
= 0
B
4
a
1

B
4

a
1
b
c
0
d

C
8
Figure 7. The algebras A
4
, B
4
and C
8
.
RL
c
-chain because A
4
is idempotent, i.e., x = x
2
holds. Let x = , y = a and z = 1.
Then zy = a, xy = a and zx = . Consequently,
(zy) ((xy)(1/(zx))) = a (a(1/)) = a (a) = a = a < 1.
Thus ,[=
A
1 (zy) ((xy)(1/(zx))). The FL
co
-chain B
4
can be easily obtain
from A
4
by interpreting the constant 0 as (see Figure 7).
Let S c, o. Using Lemmas 4.2.1 and 4.2.2, it is clear that the variety F'

S
of
semilinear FL
S
-algebras enjoy neither C nor 1C. The same holds for 1'

and 1'

c
.
Thus we obtain the following theorem.
THEOREM 4.2.3. Let S c, o. Then the varieties F'

S
, 1'

, and 1'

c
enjoy neither
C nor 1C.
The previous theorem states that some semilinear varieties do not have C and 1C.
Consequently, they do not enjoy also SC, FSC, S1C and FS1C.
The method by which Theorem 4.2.3 was proved, suggests that a similar approach
can be used also for other semilinear varieties. It is sufcient for a given variety to nd
an algebra where (20) is not valid. One can even use a computer system for searching
such counterexamples. The next lemma presents a cyclic involutive FL
c
-chain which
was found using the computer program Mace4.
LEMMA 4.2.4. There is a cyclic involutive FL
c
-chain C
8
where the following equation
does not hold:
1 (zy) ((xy)(1/(zx))).
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 339
Proof. Let C
8
= , a, 1, b, c, 0, d, ordered by < a < 1 < b < c < 0 < d <
(see Figure 7). The multiplication is dened as follows:
a 1 b c 0 d

a a a a c 0 0
1 a 1 b c 0 d
b b b b c d d
c c c c
0 c 0
d c d

Then the divisions can be computed using Proposition 3.0.2.
We will check that (20) is not valid in C
8
. Let x = y = c and z = 0. Then
zy = zx = 0c = a and xy = cc = b. Consequently,
(zy) ((xy)(1/(zx))) = a (b(1/a)) = a (b1) = a = a < 1.
Thus ,[=
C
8
1 (zy) ((xy)(1/(zx))).
It immediately follows from Lemma 4.2.4 that an analogue of Theorem 4.2.3 holds
also for cyclic involutive FL

c
-algebras and also for semilinear varieties above it.
THEOREM 4.2.5. Let S = or S = c. Then the varieties "nF'

S
and Cy"nF'

S
enjoy neither C nor 1C.
Note that the counterexamples given in Lemmas 4.2.2 and 4.2.4 are not commuta-
tive. Thus it is natural to ask whether we can obtain some of the completeness prop-
erties at least in the commutative case. The answer to this question is afrmative. It
was proved in [32] that F'

e
enjoys SC and S1C. However, the result was proved
proof-theoretically using the hypersequent calculus for semilinear FL
e
. So far there is
no algebraic proof of this fact. Thus we present the next theorem without a proof.
THEOREM 4.2.6. The variety F'

e
has the strong completeness property with respect
to the class of all countably innite dense FL
e
-chains and also FL
e
-chains on [0, 1].
Let us remark that the above theorem was proved not precisely for semilinear FL
e
-
algebras but for semilinear FL
e
-algebras with an expanded language by constants
and . These constants are interpreted by a bottom and a top element, thus all FL
e
-
algebras in this expanded language are bounded. The reason why we could claim The-
orem 4.2.6 follows from Proposition 3.3.1 from which it follows that every FL
e
-chain
can be extended to a bounded FL
e
-chain.
Concerning the completeness properties for non-integral semilinear varieties with
respect to the classes of countably innite dense chains and chains on [0, 1], there remain
still unsolved cases. For instance, the situation is not known for the following semilinear
varieties: F'

ec
, 1'

ec
, "nF'

e
and "nF'

ec
.
340 Rostislav Hor ck
4.2.2 Completeness with respect to nite algebras
Negative results can be obtained also for the completeness properties with respect to
the class of nite chains. Namely, we will show that a bunch of non-integral semilinear
varieties of FL-algebras do not enjoy FSFC.
Consider the following quasi-identity saying that every positive right-invertible ele-
ment x has to be equal to 1:
1 x and x y = 1 implies x = 1. (21)
This quasi-identity holds in any nite FL-algebra A. Indeed, let x, y A such that
1 x and xy = 1. We have x x
2
x
3
because x is positive. Further,
since A is nite, there is n N such that x
n+1
= x
n
. Multiplying this equation by
y
n
from the right-hand side, we obtain x = 1. Thus every variety ' of FL-algebras
containing a member with a strictly positive invertible element cannot enjoy the nite
strong completeness property with respect to the class F of all nite members of '
because 1 x, x y = 1 [=
F
x = 1 but 1 x, x y = 1 ,[=
L
x = 1. Using this
observation, we can prove the following theorem.
THEOREM 4.2.7. Let S e, o. Then the variety F'

S
does not have the nite strong
completeness property with respect to the class of all nite FL
S
-chains. The same is true
also for "nF'

S
, Cy"nF'

S
, 1'

S
, "n1'

S
, and Cy"n1'

S
if S e.
Proof. Consider the additive -group Z of integers. It forms a cyclic involutive RL
e
-
chain. The involutive RL
e
-chain Z does not satisfy (21) because every positive integer
is invertible. Thus the claim follows for F'

S
, "nF'

S
, Cy"nF'

S
, 1'

S
, "n1'

S
, and
Cy"n1'

S
if S e.
If o S then we can extend Z to a bounded FL
eo
-chain Z

by Proposition 3.3.1
interpreting the constant 0 by the bottom element . Since Z

possesses a strictly
positive invertible element, the claim follows for F'

S
also for S e, o.
Thus the above-mentioned varieties are not generated by nite members as quasi-
varieties. On the other hand, it is not known whether the above varieties are generated
by their nite members, i.e., if they enjoy at least FC. Also for c S c, o we do
not know whether F'

S
satises some of the completeness properties with respect to the
class of all nite FL
S
-chains.
On the other hand, we have at least some positive results. It turns out that contraction
together with commutativity is sufcient to prove FSFC. It was proved in [40] that
the varieties F'
ec
and F'
eco
enjoy FSFC. The same was proved also for "nF'
ec
and
"nF'
eco
in [37]. Although it is not mentioned in the above-mentioned papers explicitly,
the results easily extend also to the semilinear case.
THEOREM 4.2.8. The varieties F'

ec
, F'

eco
, "nF'

ec
, and "nF'

eco
enjoy the nite
strong completeness property with respect to the class of all nite chains.
5 Subvariety lattice
In this section we will study the structure of the subvariety lattice of FL

-algebras.
Its structure is very complex and it is not very well known. Even at the very bottom we
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 341
have continuum many atoms. Thus we focus here only on the cardinalities of atoms in
the subvariety lattices of varieties F'

S
for S e, c, i, o and 1'

S
for S e, c, i.
Note that due to algebraizability the cardinality of atoms in the subvariety lattice of F'

S
(resp. 1'

S
) corresponds to the cardinality of maximally consistent axiomatic extensions
of FL

S
(resp. RL

S
).
5.1 General facts
Given a variety V of FL-algebras, we denote its subvariety lattice (V). Recall
Theorem 2.1.1 saying that varieties of algebras are exactly equational classes. Since
FL-algebras are dened over a nite language, there are countably many equations over
this language. Due to this fact, it follows that (F') can have cardinality at most 2

0
.
There are in fact two types of results we are going to prove. The rst one are re-
sults saying that a subvariety lattice contains continuum many atoms. Such results are
proved by constructing 2

0
algebras generating pair-wise different atoms in the subva-
riety lattice. The second type of results shows that there are only nitely many atoms
in a subvariety lattice ('). This can be proved by showing that each nontrivial vari-
ety V (') contains an algebra generating one of the nitely many atoms. In both
cases we need a criterion whether a given algebra A generates an atom in a subvariety
lattice. A summary of the results we are going to prove can be found in Figures 9, 10.
The cardinalities of atoms are written in the boxes next to the corresponding semilinear
varieties. It is clear that if a variety has a subvariety lattice with 2

0
atoms then the same
is true also for all varieties above it.
A nontrivial algebra A is said to be strictly simple if it lacks nontrivial proper sub-
algebras and congruences. By proper subalgebra of A we mean here a subalgebra B
which is not isomorphic to A. Note that this notion differs from the usual one saying that
a subalgebra B of A is proper if B _ A. Strictly simple algebras are good candidates
for generators of an atom. Let A be a strictly simple FL-algebra. Since the variety of
FL-algebras is congruence distributive, we can use J onssons Lemma to show that sub-
directly irreducible members in V(A) are contained in HSP
U
(A). We will mention
two special cases in which strictly simple algebras generate an atom.
First, if A is nite then P
U
(A) contains just A because any ultrapower of a nite
structure is isomorphic to the original structure. Thus each subdirectly irreducible mem-
ber has to belong to HS(A) in this case. Since A has no proper nontrivial subalgebras
and congruences, we get that A is the only subdirectly irreducible algebra in V(A)
which means that V(A) has to be an atom (recall that every variety is generated by its
subdirectly irreducible members). Thus we have the following lemma.
LEMMA 5.1.1. Let A be a nite nontrivial strictly simple FL-algebra. Then V(A) is
minimal, i.e., it forms an atom.
Second, we will discuss the case when Ais innite and lower-bounded. Moreover,
the lower-bound has to be nearly term denable. An element b A is called nearly
term denable if there is an n-ary term t(x
1
, . . . , x
n
) such that t(a
1
, . . . , a
n
) = b holds
unless a
1
= = a
n
= 1.
LEMMA 5.1.2. Let Abe a strictly simple FL-algebra with a bottom element nearly
term denable by an n-ary term t. Then V(A) is a minimal variety. Moreover, if A
t
is
342 Rostislav Hor ck
a strictly simple FL-algebra with a bottom element nearly term denable by the same
term t, then V(A) V(A
t
) if and only if Aand A
t
are isomorphic.
Proof. Let Dbe a subdirectly irreducible member of V(A). By J onssons Lemma there
is an ultrapower B = A
I
/U, a nontrivial subalgebra C of B and a homomorphism
f : C D such that f(C) = D. In order to show that V(A) is minimal, we will show
that D contains A as a subalgebra. Clearly B contains an isomorphic copy of A (it is
a subalgebra of all congruence classes containing constant functions). We identify the
elements from this copy with the original elements from A. Thus B. Observe that
Asatises the following rst-order sentence expressing the fact that t(x
1
, . . . , x
n
) =
if at least one of x
1
, . . . , x
n
is not 1:
(x
1
, . . . , x
n
)(y) ((x
1
,= 1 or or x
n
,= 1) implies t(x
1
, . . . , x
n
) y) . (22)
Since universal sentences are preserved under taking subalgebras and ultrapowers, C
has to satisfy (22) as well. Consequently, C because = t(a
1
, . . . , a
n
) for any
a
1
, . . . , a
n
C at least one of them different from 1 (we can choose such elements
because C is nontrivial). Moreover, A is a subalgebra of C because it is generated by
(recall that A is strictly simple, so it is generated by any element different from 1).
Finally, we claim that the restriction of f on A has to be an isomorphism. Indeed, if not
then f() = 1 because Ais simple (it is a well-known fact from universal algebra that
each homomorphic image of a simple algebra is either isomorphic to this algebra or it is
trivial). Since is the bottom element of B and C as well, we get f[C] = 1 which
contradicts D being subdirectly irreducible. Consequently, D contains an isomorphic
copy of A. Hence V(A) is an atom.
To see the second part of the lemma, assume that V(A) V(A
t
). Since A is
subdirectly irreducible, A
t
is embeddable into A by the previous argument. As both
algebras A, A
t
are generated by , the embedding has to be onto.
Finally, in order to construct algebras generating an atom in a subvariety lattice, we
will need a method extending an RL
ei
-chain by a new neutral element for the multiplica-
tion. Let A = A, , , , , 1
A
be an RL
ei
-chain with a coatoma = max(A1
A
).
We will extend the 1-free reduct of A by adding a new neutral element 1
A

in order to
obtain an RL
e
-chain A

= A

, , , , , 1
A

, where A

= A 1
A

. The new
lattice order , is the extension of the original order letting a 1
A

1
A
. Thus 1
A
becomes a top element of A

. Let x A 1
A

and y A 1
A
. The operations
are extended as follows:
1
A

x = x = x 1
A

, 1
A

x = x, y 1
A

= 1
A
, 1
A
1
A

= a.
LEMMA 5.1.3. Let A be an RL
ei
-chain with a coatom a. Then A

is an RL
e
-chain.
Moreover, if Ais contractive then A

is contractive as well.
Proof. It is easy to check that A

, , 1
A

forms a commutative monoid. We will check

that A

satises the residuation property, i.e., xy z iff y x z. Since A is a

subuniverse of the 1-free reduct of A

, we have x y z iff y x z for all x, y, z

Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 343
= 0
2
= 1
2
2

= 1
2
r
= 0
2
r
2
r
= 0
2

e = 1
2

e = 1
2

r
= 0
2

r
Figure 8. Some strictly simple FL-chains.
A. Let x, y, z A1
A

and suppose that at least one of them equals 1

A

. If x = 1
A

then we have 1
A

y z iff y z = 1
A

z. Thus assume x ,= 1
A

. If z ,= 1
A

(i.e.,
y = 1
A

) then x z iff 1
A
x z. Consequently, x = x1
A

z iff 1
A

x z.
Now assume z = 1
A

. If x = 1
A
then 1
A
y 1
A

iff y a = 1
A
1
A

.
If x a then x y x 1
A

. Thus x y 1
A

iff y 1
A
= x 1
A

. Summing up,
A

is an RL
e
-chain. The moreover part follows easily since we have x x
2
for x A
and 1
A

= (1
A

)
2
.
Figure 8 shows several examples of strictly simple FL-chains which we will need
later. The rst one is the two element Boolean algebra which forms an FL
cw
-chain 2.
The second one is its corresponding RL
ci
-chain 2
r
. The last algebra 2

r
is obtained
from 2
r
by means of Lemma 5.1.3. Thus 2

r
is an RL
ec
-chain (recall that contractivity
together with integrality implies commutativity). Finally, the third algebra 2

is an
FL
ec
-chain whose 0-free reduct is 2

r
and 0 is interpreted as . All of them are clearly
simple because every nontrivial lter F (i.e., F ,= 1) has to contain . Thus all of
them have only two lters. It is also easy to check that they have no proper nontrivial
subalgebras. Consequently, each of these chains generates an atom in the subvariety
lattice (F'

) by Lemma 5.1.1.
5.2 Finitely many atoms
We start with the easier part proving that some of the subvariety lattices contain only
nitely many atoms. Note that if an FL-algebra Ais lower bounded (i.e., it has a bottom
element ), then it is bounded since is a top element usually denoted . Moreover
the set , is closed under multiplication, left and right divisions, and lattice opera-
tions. Thus it forms a subalgebra of the 0, 1-free reduct of A. Consequently, if A is
an FL
w
-algebra (i.e., 1 = and 0 = ), then , = 0, 1 forms a subalgebra of
A. Moreover, it is easy to check that this subalgebra is in fact the two element Boolean
algebra 2. Thus we have the following theorem (note that 2 is also commutative and
contractive).
THEOREM 5.2.1. Let w S e, c, w. Then (F'

S
) contains a single atom,
namely the variety V(2) generated by the two element Boolean algebra 2.
344 Rostislav Hor ck
F'

F'

ew
1
F'

i
F'

ei
2
0
F'

ec
2
0
F'

e
F'

c
F'

eo
2
0
F'

eco
2
F'

co
2
0
F'

o
F'

w
1
F'

ci
= F'

eci
2
F'

cw
= F'

ecw
1
Figure 9. The lattice of considered semilinear varieties of FL
S
-algebras and the numbers
of atoms in their corresponding subvariety lattices.
1'

c
2

0
1'

ec
2
1'

1'

ci
= 1'

eci
1
1'

i
? 1'

e
2

0
1'

ei
2
Figure 10. The lattice of considered semilinear varieties of RL
S
-algebras and the num-
bers of atoms in their corresponding subvariety lattices.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 345
The above theorem covers all cases where w S. Thus in the following text we
have to omit either the condition that 1 is a top element or that 0 is a bottom element. If
we remove the second condition, we obtain the following theorem.
THEOREM 5.2.2. The subvariety lattice (F'

ci
) contains two atoms, namely the va-
rieties V(2) and V(2
r
).
Proof. Let Abe a nontrivial algebra from F'

ci
. Since every algebra in F'

ci
is integral,
1 is the top element. Moreover, xy = x y because A is contractive. If 0 < 1 then
0, 1 forms a subalgebra of A isomorphic to 2. Thus assume 1 = 0. Let a < 1 be a
strictly negative element from A (see Lemma 3.0.7). Then a, 1 forms a subalgebra of
Aisomorphic to 2
r
.
If we remove integrality from F'

cw
, the number of atoms also increases as in the
previous theorem.
THEOREM 5.2.3. The subvariety lattice (F'

eco
) contains two atoms, namely vari-
eties V(2) and V(2

).
Proof. Let A be a subdirectly irreducible algebra from F'

eco
(i.e., A is a nontrivial
chain). Then 0 is the bottom element. Consider the FL
eco
-chain B = A/F, where F
is the maximal lter not containing 0 (i.e., the union of all lters not containing 0; it is
a lter because A is a chain). The algebra B is clearly simple. Recall that xy = x y
for negative elements x, y B because B is contractive. We claim that 0 is a subcover
of 1. Indeed, if there would be an element b B such that 0 < b < 1 then the lter
F(b) = b generated by b is nontrivial and proper (see Theorem 3.5.2). Thus B would
not be simple. Further B has a top element = 0 0. If = 1 then 0, 1 forms a
subalgebra of B isomorphic to 2. If > 1 then we claim that 0, 1, is a subuniverse
of B isomorphic to 2

. The set 0, 1, is clearly closed under multiplication. Recall

that 0, is closed under . Furthermore we have 1 x = x and 0 1 = . Thus
the only remaining case is 1. The value 1 has to be strictly less than 1. Since
0 is the subcover of 1, we must have 1 = 0. It is straightforward to check that the
subuniverse 0, 1, forms an algebra isomorphic to 2

.
Now we turn our attention to RL

-algebras. The situation for RL

ci
-algebras is again
easy. Let Abe a nontrivial algebra from (1'

ci
) and a A a strictly negative element
(see Lemma 3.0.7). Since the multiplication coincides with , the set a, 1 forms a
subalgebra of Aisomorphic to 2
r
. Thus we obtain the following theorem.
THEOREM 5.2.4. The subvariety lattice (1'

ci
) contains a single atom, namely the
variety V(2
r
).
We have seen that the subvariety lattice of contractive integral RL-chains contains
only a single atom. Even without integrality the situation is relatively easy if we have
commutativity.
THEOREM5.2.5. The subvariety lattice (1'

ec
) contains two atoms, namely varieties
V(2
r
) and V(2

r
).
346 Rostislav Hor ck
Proof. Let Abe a subdirectly irreducible member of (1'

ec
). Recall that xy = x y
for x, y 1 because A is contractive. Thus every negative element a A generates
a lter F(a) = a (see Theorem 3.5.2). Since A is subdirectly irreducible, there has
to be minimum nontrivial lter F(a) for a strictly negative a A (see Theorems 2.1.4
and 3.5.3). Then a is a subcover of 1. Set T = a 1. We claim that a, 1, T is a
subalgebra of A. If T = 1 then a, 1 is clearly closed under multiplication. It is also
closed under because 1 x = x for every x A and
1 a a a 1 = T = 1.
Thus a, 1 forms a subalgebra isomorphic to 2
r
.
Thus assume that T > 1. We check that a, 1, T is closed under multiplication.
Clearly, 1x = x = x1 for x a, 1, T. Further, a
2
= a. To see that aT = Ta
a, 1, T, note that
a aT = a(a 1) 1.
Since a is a subcover of 1, we get Ta = aT a, 1. Moreover, Ta ,= 1. Indeed,
if Ta = 1 then a = (Ta)a = Ta
2
= Ta = 1. Thus Ta = aT = a. Consequently,
aT
2
= (aT)T = a 1. Thus T T
2
a 1 = T. Hence a, 1, T is closed
under multiplication. Finally, we have to check that a, 1, T is closed under . Let
x a, 1, T. Clearly, 1 x = x. Next we show that a x a, 1, T. We have
a 1 = T by the denition. Then a T = a (a 1) = a
2
1 = a 1 = T.
Also aT = a, hence
T a a a 1 = T.
Finally, we show that T x a, 1, T. We have T 1 = (a 1) 1 a. On
the other hand, T 1 < 1 otherwise 1 T(T 1) T. Thus T 1 = a because
a is the subcover of 1. Moreover, a T a because Ta = a. On the other hand,
T a < 1 otherwise a T(T a) T. Thus T a = a. Finally, to see that
T T a, 1, T, note that T
2
= T and aT = a. Consequently,
T T T = T (a 1) = aT 1 = a 1 = T.
Thus a, 1, T forms a subalgebra of A. It is easy to see that it is isomorphic to 2

r
.
If we replace contraction by integrality, the number of atoms still remains nite.
Recall that Z = Z, +, , , , 0 (where x y = y x) denotes the additive -group
of integers viewed as an RL
e
-chain. One can apply to Z the conucleus (x) = x 0
(see Section 3.7) in order to obtain an RL
ei
-chain Z

= [Z] = Z

, +,

, , , 0.
Note that x

y = (y x) 0. Before we prove that the subvariety lattice (1'

ei
)
contains only two atoms, we will prove that Z

generates an atom.
LEMMA 5.2.6. The variety V(Z

) is an atom in (1'
ei
).
Proof. Assume that V(A) V(Z

) is a nontrivial subvariety of V(Z

). Thus A
is a nontrivial RL
ei
-algebra. First, observe that Z

satises the identity x xy = y.

Indeed, we have x

(x+y) = (x+yx)0 = y. Thus V(A) has to satisfy this iden-

tity as well. Let a be a strictly negative element in A (see Lemma 3.0.7). Consider the
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 347
submonoid B of A generated by a, i.e., B = a
k
[ k N. We claim that a
k+1
< a
k
for all k N. Indeed, if a
k+1
= a
k
for some k N, then
1 = a
k
a
k
= a
k
a
k+1
= a.
Further, we claim that B forms a subalgebra of A. It is clearly closed under ,
because B is a chain. It remains to check that B is closed under . Let k, n N. If
k n then a
n
a
k
. Thus a
n
a
k
= 1 B. Suppose that k > n. Then a
n
a
k
= a
n
a
n
a
kn
= a
kn
. Thus B forms a subalgebra which is obviously isomorphic
to Z

. Consequently, V(Z

) V(A), i.e., V(Z

) is an atom.
THEOREM 5.2.7. The subvariety lattice (1'

ei
) contains two atoms, namely varieties
V(2) and V(Z

).
Proof. Let Abe a subdirectly irreducible RL

ei
-algebra, i.e., Ais nontrivial and linearly
ordered. Thus A contains is a strictly negative element a < 1 as A is integral. We
will show that V(A) contains either 2 or Z

. Let N
+
denote the set of strictly positive
natural numbers and consider the non-increasing sequence a
k

kN
+. If a
k+1
= a
k
for
some k N
+
, then a
k
is idempotent. Indeed, we have
a
2k
= a
k+1
a
k1
= a
k
a
k1
= = a
k
.
Then it is easy to see that a
k
, 1 forms a subalgebra of A isomorphic to 2
r
. Thus in
this case we have 2
r
V(A).
Now assume that a
k+1
< a
k
holds for all k N
+
, i.e., a
k

kN
+ is a strictly
decreasing sequence. We will show that Z

ISHP
U
(A) V(A). Consider a
non-principal ultralter U on N and the corresponding ultrapower B = A
N
/U. Set

1 = 1, 1 . . ./U, a = a, a, . . ./U, and

b = a
k

kN
+/U. Let be the congruence
on B corresponding to the lter F(a) generated by a (cf. Theorem 3.5.3). In particular,
we have

1/ = a/ because

1 a = a F(a) and a

1 =

1 F(a). Further,
note that

b , F(a). Indeed, if

b F(a) = a
n
[ n N, then for a xed n N
and a set J in the ultralter U we have a
k
a
n
for all k J. However, a
k

kN
+
is strictly decreasing. Thus J has to be nite. Consequently, U has to be principal (a
contradiction).
We claim that the subalgebra of B/ generated by

b/ is isomorphic to Z

. First, it
is not trivial because

b , F(a). Second, we have to check that

b
m
/

b
n
/ =

b
nm
/
for n > m. Note that for all k N
+
we have
a
(nm)k
a
mk
a
nk
< a
(nm)k1
a a
(nm)k
.
Thus

b
nm
/

b
m
/

b
n
/ a/

b
nm
/ =

1/

b
nm
/ =

b
nm
/.
Unfortunately, the proof of Theorem 5.2.7 does not work without commutativity
(i.e., for the variety 1'

i
) because the lter F(a) is closed also under conjugates. Then
it is not clear how to prove that

b , F(a).
PROBLEM 5.2.8. What is the cardinality of atoms in (1'

i
)?
348 Rostislav Hor ck
5.3 Continuum many atoms
Now we focus on varieties whose subvariety lattice has continuum many atoms.
We start with the variety of RL

e
-algebras and construct 2

0
algebras generating atoms
in (1'

e
). Easy modications of these algebras give also 2

0
atoms in (F'

eo
) and
(F'

ei
).
Let A = A, +, , , , 0, 0 be the totally ordered Abelian -group (viewed
as an RL
e
-chain) given by the lexicographic product of two copies of Z, i.e., A =
Z
2
ordered lexicographically, + is dened component-wise and x, y u, v =
u x, v y. Recall that there are 2

0
innite subsets of 2 + Z

= 2 + z [
z Z

, where Z

denotes the set of non-positive integers. We construct for each in-

nite subset S 2+Z

an algebra A
S
from the algebra Aby means of a conucleus
S
and a nucleus . We dene the conucleus
S
by its image as follows:

S
[A] = 0, 0, 1, 0, 1, 11, z A [ z Sx, y A [ x 2.
To show that
S
[A] denes a conucleus
S
, we have to prove that
S
[A] is the image of
an interior operator which forms a submonoid (see Lemma 3.7.1). The set
S
[A] clearly
forms a submonoid of A since 1, x + 1, y = 2, x + y
S
[A]. Further we
have to check that
S
[A] is an interior system of A, i.e., if maxx
S
[A] [ x y
exists for every y A. The existence is obvious for y
S
[A]. Suppose that y ,
S
[A].
Then y has the rst component greater than or equal to 1. If y = u, v for u > 0,
then the maximum is 0, 0. If u = 0 then the maximum is 1, 0. If u = 1 then the
maximum exists since the set 1, S = 1, z A [ z S is dually well-ordered
and innite (thus there is a lower bound of y in 1, S). Hence
S
[A] is the image
of a conucleus and
S
[A] =
S
[A], +,

S
, , , 0, 0 forms an RL
ei
-chain, where
x, y

S
u, v =
S
(u x, v y).
Now consider the -retraction B = [
S
[A]] where (x, y) = x, y 3, 1.
The algebra B = [
S
[A]], +

S
, , , 0, 0 is again an RL
ei
-chain as follows
from Lemma 3.7.4, where x, y +

u, v = (x + u, y + v). The structure of B is

depicted in Figure 11. We dene the algebra A
S
as the subalgebra of B generated by
a = 1, 0. The algebra A
S
is an RL
ei
-chain. It is even a 4-potent RL
ei
-chain (i.e., it
satises x
5
= x
4
) because a
4
= (4, 0) = 3, 1 is the bottom element.
We will prove that each A
S
contains some important elements from B. This will
help us to prove that A
R
and A
S
are not isomorphic for R ,= S. First, note that there is
a term r(x) = x
2
x
4
such that its value in A
S
for x = a is
r
A
S
(a) = 2, 0

S
3, 1 =
S
(1, 1) = 1, 1.
Thus the element 1, 1 belongs to A
S
. Since the set S is dually well ordered, we
can index its elements by natural numbers, i.e., S = c
0
> c
1
> c
2
> .
LEMMA 5.3.1. For every A
S
and every n N there is a term s
n
(x) such that
s
A
S
n
(a) = 1, c
n
.
Thus 1, c
n
A
S
for all n N.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 349
0, 0 = 1
B
1, 0 = a
1, 1
1, S
2, 1
2, 0
2, 1
3, 1
3, 0
3, 1
Figure 11. The structure of the algebra B.
Proof. By induction on n. Let s
0
(x) be the term x r(x)
2
. Then
s
0
(a) = 1, 0

S
2, 2 =
S
(1, 2) = 1, c
0
,
since 1, c
0
is the subcover of 1, 1. Now assume that there is a term s
n
(x) such
that s
A
S
n
(a) = 1, c
n
. Let s
n+1
(x) = x r(x) s
n
(x). Then
s
A
S
n+1
(a) = 1, 0

S
2, c
n
1 =
S
(1, c
n
1) = 1, c
n+1
.
LEMMA 5.3.2. For every A
S
and every n N there is a term t
n
(x) such that
t
A
S
n
(a) = 2, n.
Thus 2, n A
S
for all n N.
Proof. By induction on n. For n = 0 it is clear since 2, 0 = a+

a = a+a. Assume
that there is t
n
(x) such that t
A
S
n
(a) = 2, n. Consider the term t
n+1
(x) = r(x)
x t
n
(x). Then
t
A
S
n+1
(a) = 1, 1

S
3, n =
S
(2, n + 1) = 2, n + 1.
The algebra A
S
is a simple RL

ei
-algebra because every nontrivial lter F contains
a and a
4
is the bottom element, i.e., F = A
S
. However, A
S
is not strictly simple since
350 Rostislav Hor ck
3, 1, 0, 0 forms a subalgebra isomorphic to 2
r
. In order to obtain a strictly
simple algebra, we will use the construction from Lemma 5.1.3 and extend A
S
by a
new neutral element.
LEMMA 5.3.3. The algebra A

S
is a strictly simple RL

e
-algebra with a nearly term
denable bottom element.
Proof. We will show that any element x ,= 1
A

S
generates A

S
. First, we can make the
top element 0, 0 = x x. Second, we can produce a = 1, 0 = 0, 0 1
A

S
since 1, 0 is the coatom of A
S
. Since A
S
is generated by a, we are done. Moreover
the bottom element a
4
is nearly term denable by the term x
4
(x 1)
4
.
In order to invoke Lemma 5.1.2, we have to prove that A
R
and A
S
are not isomor-
phic for different sets R, S.
LEMMA 5.3.4. Let R, S 2 + Z

such that R ,= S. Then A

S
is not isomorphic
to A
R
. The same is true also for A

S
and A

R
.
Proof. Let us enumerate the elements of R, S as follows: R = d
0
> d
1
> d
2
>
and S = c
0
> c
1
> c
2
> . Suppose that f : A
S
A
R
is an isomorphism. Since
f is order-preserving, f must be the identity when restricted to the set
0, 0, 1, 0, 1, 1, 3, 0, 3, 1.
By Lemma 5.3.1 we have 1, c
n
A
S
, 1, d
n
A
R
and f(1, c
n
) = f(s
A
S
n
(a))
= s
A
R
n
(f(a)) = s
A
R
n
(a) = 1, d
n
for all n N. Assume that k is the least natural
number such that c
k
,= d
k
. Without any loss of generality suppose that c
k
> d
k
.
Lemma 5.3.2 has two consequences. First, 2, c
k
A
S
, A
R
. Second, we have
f(2, n) = f(t
A
S
n
(a)) = t
A
R
n
(f(a)) = t
A
R
n
(a) = 2, n for all n N. Thus we get
a
3
= f(a
3
) = f(3, 0) = f(1, c
k
+2, c
k
) = f(1, c
k
+

2, c
k
) =
1, d
k
+

2, c
k
= 3, d
k
c
k
3, 1 = 3, 1 = a
4
,
which is a contradiction since a
3
,= a
4
. It is straightforward to check that the above
argument works also for extended algebras A

S
and A

R
.
Now using Lemma 5.1.2, we know that each A

S
generates an atom in (1'

e
).
Moreover, distinct subsets R, S of 2 + Z

generate different atoms because A

R
and
A

S
are not isomorphic. Thus we obtain the following theorem.
THEOREM 5.3.5. There are 2

0
atoms in (1'

e
).
The above-mentioned RL
ei
-chains A
S
can be also used to construct atoms in the
subvariety lattices (F'

eo
) and (F'

ei
). First, A

S
can be viewed as an FL
eo
-chain
if we interpret the constant 0 as a
4
. Then the premises of Lemma 5.1.2 remain still
satised. Thus FL
eo
-chains A

S
generate 2

0
atoms in (F'

eo
). Second, observe that
we can make A
S
into an FL
ei
-chain by interpreting the constant 0 as a = 1, 0.
Consequently, the FL
ei
-chain A
S
becomes strictly simple because it is generated by a.
Moreover, the bottom element of A
S
is term-denable by the term 0
4
. Consequently,
using Lemma 5.1.2 again, we obtain the next theorem.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 351
THEOREM 5.3.6. There are 2

0
atoms in (F'

ei
) and (F'

eo
).
We have seen that the subvariety lattices of FL

ei
-algebras and FL

eo
-algebras have
continuum many atoms. The same can be proved also for the subvariety lattice of FL

ec
-
algebras. Let A = Z
2
, , , , , , 0, 0 be the RL
e
-chain arising on the
lexicographic product of two copies of the additive -group of integers Z extended by
bottom and top elements , (see Proposition 3.3.1). Thus we have x = , x =
for x ,= , and a
1
, b
1
a
2
, b
2
= a
1
+ a
2
, b
1
+ b
2
. The division is computed
by formulas x = , x = for x ,= , x = , x = for
x ,= , and a
1
, b
1
a
2
, b
2
= a
2
a
1
, b
2
b
1
. We dene a conucleus on A
by its image as follows:
[A] = , , 0, 0 a, b [ a > 0, b < 0.
To prove that [A] is the image of a conucleus, we have to check that [A] is a sub-
monoid and the image of an interior operator (see Lemma 3.7.1). It is easy to see that
[A] is closed under multiplication, i.e., it forms a submonoid. Further, we have to check
that for every y Z
2
, the maximum m = maxx [A] [ x y exists.
For y [A] it is obvious. Thus suppose that y , [A]. If y < 0, 0 then the max-
imum m = because it is the only strictly negative element in [A]. If y > 0, 0
then y = a, b for a 0 and b 0. Consequently, m = 0, 0 if a = 0 and
m = a, 1 if a > 0. Thus [A] is the image of a conucleus and the -contraction
[A] = [A], ,

, , , 0, 0 is an RL
e
-chain.
Let N
+
denote the set of strictly positive natural numbers. Given a subset S N
+
such that 1 S, we dene a nucleus
S
as follows:

S
(x) = x for x , , 0, 0,

S
(a, b) =
_
a, 1 if a , S or b = 1,
a, 2 otherwise.
Observe that the image of
S
is the following set:

S
[[A]] = , , 0, 0 k, 1 [ k N
+
k, 2 [ k S.
The image of
S
is well-ordered and contains . Thus
S
is a closure operator. We have
to check that
S
(x)
S
(y)
S
(xy). If x or y equals , or 0, 0 then the inequality
obviously holds. Thus assume that x, y , , , 0, 0. Then x = a
1
, b
1
and y =
a
2
, b
2
for some a
1
, a
2
1 and b
1
, b
2
1. We have
S
(x)
S
(y) a
1
+ a
2
, 2
and
S
(xy) a
1
+ a
2
, 2. Thus
S
(x)
S
(y)
S
(xy) holds.
Using the conucleus and nucleus
S
, we dene an FL
e
-chain
A
S
=
S
[[A]] =
S
[[A]],

S
,

, , , 1, 2, 0, 0,
where x

S
y =
s
(x y) and 0
A
S
= 1, 2. The structure of A
S
is depicted in
Figure 12. Note that A
S
is in fact an FL
ec
-algebra because x x
2
holds for every x as
the only strictly negative element is .
In order to invoke Lemma 5.1.2, we have to show that A
S
is strictly simple with a
nearly term denable bottom element and A
S
is not isomorphic to A
R
for S ,= R. We
start with a lemma showing properties of 0
A
S
.
352 Rostislav Hor ck

0, 0 = 1
A
S
1, 2 = 0
A
S
1, 1
2, 1
3, 2
3, 1
4, 1
5, 2
5, 1

S, 2 N
+
, 1
Figure 12. The structure of the algebra A
S
for S = 1, 3, 5, . . ..
LEMMA 5.3.7. Let S N
+
such that 1 S. Then A
S
satises for all k N
+
the
following:
1. (0
A
S
)
k
=
_
k, 2 if k S,
k, 1 otherwise.
2. 0
A
S

(0
A
S
)
k+1
= k, 1.
Proof. We will prove the rst claim by induction on k. For k = 1 the statement holds
since 1 S and 0
A
S
= 1, 2. Assume the validity for k, i.e., (0
A
S
)
k
= k, b for
some b 1, 2. We have
(0
A
S
)
k+1
= 0
A
S

S
(0
A
S
)
k
=
S
(1, 2 k, b) =
S
(k +1, 2 +b) = k + 1, b
t

for some b
t
< 0. Since 2 + b 2, we get b
t
= 2 if k + 1 S and b
t
= 1
otherwise.
Now we prove the second claim. By the rst claim we have (0
A
S
)
k+1
= k + 1, b
for some b 1, 2. Since b + 2 0, we obtain
0
A
S

(0
A
S
)
k+1
= (1, 2 k + 1, b) = k, b + 2 = k, 1.
LEMMA 5.3.8. The algebra A
S
is strictly simple with a nearly term denable bottom
element.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 353
Proof. The simplicity follows from the fact that each nontrivial lter F in A
S
contains
, i.e., F = A
S
. Further we claim that A
S
is generated by 0
A
S
= 1, 2. Let B be
the subalgebra of A
S
generated by 0
A
S
. First,
0
A
S

1
A
S
= (1, 2 0, 0) = (1, 2) = .
Thus B. Second, = B. Third, k, 2 B for k S by
Lemma 5.3.7. Finally, k, 1 B for any k N
+
by the same lemma. Thus B =
A
S
, i.e., A
S
is generated by 0
A
S
. Consequently, A
S
is strictly simple. The bottom
element is nearly term denable by the term x (x 1) because a, b

0, 0 =
(a, b) = for a, b > 0, 0.
It follows from Lemma 5.1.2 that each A
S
generates an atom in (F'

ec
). Thus it
sufces to prove that A
R
and A
S
are not isomorphic for R ,= S.
LEMMA 5.3.9. Let R, S N
+
such that R ,= S and 1 RS. Then A
R
and A
S
are
not isomorphic.
Proof. Suppose that f : A
R
A
S
is an isomorphism. Take the least element k where
R and S differ. Without any loss of generality assume k R and k , S. Then
(0
A
R
)
k
= k, 2, (0
A
S
)
k
= k, 1. Moreover, it follows from Lemma 5.3.7 that
0
A
R

(0
A
R
)
k+1
= k, 1,
0
A
S

(0
A
S
)
k+1
= k, 1.
Consequently, we obtain
f(k, 1) = f(0
A
R

(0
A
R
)
k+1
) =
0
A
S

(0
A
S
)
k+1
= (0
A
S
)
k
= f((0
A
R
)
k
) = f(k, 2).
Thus f is not one-to-one, so cannot be an isomorphism.
THEOREM 5.3.10. There are 2

0
atoms in (F'

ec
).
To complete Figures 9 and 10, we have to discuss the cardinality of atoms in the sub-
variety lattices for RL

c
-algebras and FL

co
-algebras. Continuum many atoms in (1'

c
)
were constructed in [16]. In fact, [16] constructs continuum many idempotent RL
c
-
chains (i.e., RL
c
-chains satisfying x = x
2
) generating atoms in (1'

c
). As shown
in [18] the construction can be further modied in order to prove that also (F'

co
)
contains continuum many atoms. The proofs of all these results are rather involved, so
we will state here the next theorem without a proof.
THEOREM 5.3.11. There are 2

0
atoms in (1'

c
) and (F'

co
).
354 Rostislav Hor ck
6 Historical remarks and further reading
Fuzzy logics as formal logical systems started to be investigated in 1990s. The
main contribution in this direction is H ajeks monograph [23]. This book introduces
syntactical calculi for fuzzy logics as well as their corresponding algebraic semantics.
Later it turned out that fuzzy logics can be viewed as a part of much broader class of so-
called substructural logics, i.e., logics lacking some of the structural rules of contraction,
exchange, left and right weakening. Hence we present fuzzy logics (i.e., in our sense
semilinear substructural logics) inside the hierarchy of substructural logics. For details
on substructural logics see the recent book [18] or [34, 35].
Our presentation of FL-algebras which form the equivalent algebraic semantics for
the base substructural logic FL, mainly follows [18]. It differs slightly in the section on
the Dedekind-MacNeille completion of an FL-algebra. Namely, in the construction of
the completion of an FL-algebra Awe use a downset monoid T(A) (i.e., an FL-algebra
on the set of all downsets) whereas [18] replaces the downset monoid with a powerset
monoid T(A) (i.e., an FL-algebra on the set of all subsets). We choose our approach
because T(A) is unnecessarily big in comparison with T(A). Consequently, the map
x x is an embedding of the -free reduct of A into T(A) (see Proposition 3.8.3)
which even preserves if A is linearly ordered. On the other hand, an analogous map
from A to T(A) given by x x preserves only the monoid structure of A. This
modication also allows us to obtain the structural characterization for FL-chains (see
Theorem 3.8.11). Concerning the construction of the downset monoid T(A), it is good
to mention that it can be further generalized as was known in category theory already in
1970s. Namely, the requirement that A has to be a (pointed) pomonoid can be weak-
ened in order to obtain an FL-algebra by the above construction; it sufces that A is a
poset endowed with a promonoidal structure (see [12]). Further, it is interesting to note
that the construction of the Dedekind MacNeille completion can be put into a more gen-
eral framework of residuated frames [17]. Using this framework it was described which
identities are preserved by the Dedekind-MacNeille completion (see [7, 8, 39]).
The section on completeness properties with respect to distinguished semantics con-
tains results scattered in various papers. In the literature the completeness property with
respect to the class of respective chains on [0, 1] is often called standard completeness.
The proofs of standard completeness presented here are completely new providing hope-
fully a better insight into matters. The original proofs are usually based on a method
invented in [29] proving the strong completeness property of F'

ew
with respect to the
class of respective chains on [0, 1]. This method was further modied by other authors.
In particular, [15] proves the standard completeness for "nF'

ew
and [30] for F'

w
. An
analogous result for F'

cw
goes back to 1950s (see [13]). On the other hand, the fact
that F'

does not enjoy the standard completeness was proved in [41]. Here we provide
a shorter proof together with new results that the same holds also for F'

S
if S c, o
and for "nF'

S
, Cy"nF'

S
if S c. Finally, as we mentioned already inside this
chapter, the standard completeness for F'

e
was proved in [32] but the used method is
proof-theoretical. As far as we know there is no algebraic proof of this fact. The nite
strong completeness property of a variety with respect to the class of all nite algebras
is known in the literature under the name nite embeddability property (FEP). Note that
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 355
for semilinear varieties the FEP is equivalent to the nite strong completeness property
with respect to the class of all nite chains, i.e., FSFC. The proofs of the FEP are
usually based on a method invented in [2], where it is proved that F'
ew
has the FEP. A
semilinear modication for F'

ew
and "nF'

ew
was given in [9]. Further modications
can be found in [3, 18, 3740]. In this chapter we provide a different presentation based
on conuclei and nuclei which works for semilinear varieties. For other varieties a related
approach using residuated frames can be used (see [17, 20]). The quasi-identity (21)
used in the proof of Theorem 4.2.7, showing that several non-integral semilinear vari-
eties do not possess the FEP, was introduced in [2]. An overview of many known results
on completeness properties of fuzzy logics can be found in [10].
We should point out that there are other results in the literature on (semilinear) va-
rieties of FL-algebras not mentioned in this chapter. Let us recall a few of them. Likely
the most studied variety of (semilinear) FL-algebras not covered in this chapter is the
variety of n-potent FL-algebras, i.e., FL-algebras satisfying x
n+1
= x
n
. In particular,
for n = 2 we obtain the class of idempotent FL-algebras. Note that idempotent FL
i
-
algebras coincide with FL
ci
-algebras. The constructions for proving various complete-
ness properties usually preserve the n-potency identity. Thus one can obtain analogous
completeness properties we presented here also for n-potent FL-algebras. An overview
for integral semilinear varieties can be found in [10]. The standard completeness for
the variety of idempotent FL

e
-algebras was proved in [32]. Other well-known identities
studied in the realm of FL-algebras are the identities xxy = y = yx/x expressing the
fact the monoid reduct is cancellative. These identities are not as easily preserved as the
n-potency. Hence one has to use a different method to prove a completeness property.
The nite strong completeness property for the variety of cancellative RL

ei
-algebras
with respect to the class of chains on (0, 1] was proved in [26]. In fact, [26] contains a
slightly different result from which the above-mentioned result easily follows. Namely,
[26] proves the nite strong completeness property with respect to the class of chains on
[0, 1] for a subvariety of F'

ew
axiomatized by x((x xy) y) = 1. This identity
expresses in an FL
ew
-chain Athat A0 forms a cancellative monoid. Nontrivial can-
cellative FL-algebras are necessarily innite so the varieties of cancellative FL-algebras
cannot posses the FEP. Hence it is not clear whether their (quasi)-equational theories
are decidable or not. In [27] it is proved that the variety of cancellative RL

ei
-algebras en-
joys the nite strong completeness property with respect to a countable class of chains.
Consequently, it is shown that the quasi-equational theory of this variety is decidable.
Last but not least are the identities x[x(xy)] = xy = [(xy)/x]x. They are known
under the name divisibility expressing that x y implies x = yz and x = z
t
y for some
z, z
t
. The divisible RL-algebras are known as GBL-algebras (see [22]). Moreover, the
divisible FL

ew
-algebras are known as BL-algebras. The variety of BL-algebras enjoys
the nite strong completeness property with respect to the class of chains on [0, 1]. All
details on BL-algebras can be found in Chapter V.
Finally, we will comment on the section about the subvariety lattice of semilinear
FL-algebras. A good overview on the structure of subvariety lattice of FL-algebras can
be found in [18]. The useful criterion presented in Lemma 5.1.2 comes from [16]. The
easy results on the subvariety lattices with nitely many atoms were known (see [16,
18]). The result showing that there are only two atoms in (1'

ei
) comes from [25].
356 Rostislav Hor ck
This paper also shows howto construct continuummany atoms in (1'

e
), (F'

ei
) and
(F'

eo
). The algebras generating atoms in the subvariety lattice of FL

ec
-algebras were
presented in [18]. Here we present how these atoms can be constructed from an -group
by means of a conucleus and a nucleus. Continuum many atoms were also constructed
for the subvariety lattice of RL

c
-algebras in [16]. A modication for FL

co
-algebras is
presented in [18]. Concerning other results not mentioned in this chapter, it is clear that
more is known about subvariety lattices for smaller semilinear varieties of FL-algebras.
For instance it is known that the subvariety lattice of idempotent RL

e
-algebras is count-
able (see [33]). A lot is known about the subvariety lattice of BL-algebras. These results
are mentioned in Chapter VI on MV-algebras and Chapter V on BL-algebras. There are
also results on almost minimal (semilinear) varieties of FL-algebras, i.e., varieties which
cover atoms in the subvariety lattice. For details see [18, 28, 31].
Acknowledgments
The author wishes to thank Clint van Alten for valuable comments and remarks
helping himto improve this chapter. The author has been supported by the grant ICC/08/E018
of the Czech Science Foundation (part of ESF Eurocores project LogiCCC FP006 Lo-
MoReVI) and the Institutional Research Plan AV0Z10300504.
BIBLIOGRAPHY
[1] Willem J. Blok and Don L. Pigozzi. Algebraizable Logics, volume 396 of Memoirs of the American
Mathematical Society. American Mathematical Society, Providence, RI, 1989. Freely downloadable
from http://orion.math.iastate.edu/dpigozzi/.
[2] Willem J. Blok and Clint J. van Alten. The nite embeddability property for residuated lattices, pocrims
and BCK-algebras. Algebra Universalis, 48(3):253271, 2002.
[3] Willem J. Blok and Clint J. van Alten. On the nite embeddability property for residuated ordered
groupoids. Transactions of the American Mathematical Society, 357(10):41414157, 2005.
[4] Kevin Blount and Constantine Tsinakis. The structure of residuated lattices. International Journal of
Algebra and Computation, 13(4):437461, 2003.
[5] T. S. Blyth and M. F. Janowitz. Residuation Theory, volume 102 of International Series of Monographs
in Pure and Applied Mathematics. Pergamon Press, 1972.
[6] Stanley Burris and H.P. Sankappanavar. A Course in Universal Algebra, volume 78 of Graduate Texts
in Mathematics. Springer-Verlag, 1981.
[7] Agata Ciabattoni, Nikolaos Galatos, and Kazushige Terui. Algebraic proof theory for substructural
logics: Cut-elimination and completions. Submitted.
[8] Agata Ciabattoni, Nikolaos Galatos, and Kazushige Terui. MacNeille completions of FL-algebras. To
appear in Algebra Universalis.
[9] Agata Ciabattoni, George Metcalfe, and Franco Montagna. Algebraic and proof-theoretic characteriza-
tions of truth stressers for MTL and its extensions. Fuzzy Sets and Systems, 161(3):369389, 2010.
[10] Petr Cintula, Francesc Esteva, Joan Gispert, Llus Godo, Franco Montagna, and Carles Noguera. Distin-
guished algebraic semantics for t-norm based fuzzy logics: Methods and algebraic equivalencies. Annals
of Pure and Applied Logic, 160(1):5381, 2009.
[11] Brian A. Davey and Hillary A. Priestley. Introduction to Lattices and Order. Cambridge University
Press, Cambridge, 1990.
[12] Brian J. Day. Construction of Biclosed Categories. PhD thesis, The University of Sydney, 1970.
[13] Michael Dummett. A propositional calculus with denumerable matrix. Journal of Symbolic Logic,
24(2):97106, 1959.
[14] M. Ern e, J. Koslowski, A. Melton, and G. Strecker. A primer on galois connections. In S. Andima et al.,
editor, Papers on General Topology and its Applications. 7th Summer Conf. Wisconsin., volume 704 of
Annals New York Acad. Sci., pages 103125, New York, 1994.
Chapter IV: Algebraic Semantics: Semilinear FL-Algebras 357
[15] Francesc Esteva, Joan Gispert, Llus Godo, and Franco Montagna. On the standard and rational com-
pleteness of some axiomatic extensions of the monoidal t-norm logic. Studia Logica, 71(2):199226,
2002.
[16] Nikolaos Galatos. Minimal varieties of residuated lattices. Algebra Universalis, 52(2):215239, 2005.
[17] Nikolaos Galatos and Peter Jipsen. Residuated frames with applications to decidability. To appear in the
Transactions of the AMS.
[18] Nikolaos Galatos, Peter Jipsen, Tomasz Kowalski, and Hiroakira Ono. Residuated Lattices: An Alge-
braic Glimpse at Substructural Logics, volume 151 of Studies in Logic and the Foundations of Mathe-
matics. Elsevier, Amsterdam, 2007.
[19] Nikolaos Galatos and Hiroakira Ono. Algebraization, parametrized local deduction theorem and inter-
polation for substructural logics over FL. Studia Logica, 83(13):279308, 2006.
[20] Nikolaos Galatos and Hiroakira Ono. Cut elimination and strong separation for substructural logics: An
algebraic approach. Annals of Pure and Applied Logic, 161(9):10971133, 2010.
[21] Nikolaos Galatos and James G. Raftery. Adding involution to residuated structures. Studia Logica,
77(2):181207, 2004.
[22] Nikolaos Galatos and Constantine Tsinakis. Generalized MV-algebras. Journal of Algebra, 283(1):254
291, 2005.
[23] Petr H ajek. Metamathematics of Fuzzy Logic, volume 4 of Trends in Logic. Kluwer, Dordrecht, 1998.
[24] Graham Higman. Ordering by divisibility in abstract algebras. Proceedings of the London Mathematical
Society, 2:326336, 1952.
[25] Rostislav Hor ck. Minimal varieties of representable commutative residuated lattices. To appear in
Studia Logica.
[26] Rostislav Hor ck. Standard completeness theorem for MTL. Archive for Mathematical Logic,
44(4):413424, 2005.
[27] Rostislav Hor ck. Decidability of cancellative extension of monoidal t-norm based logic. Logic Journal
of the Interest Group of Pure and Applied Logic, 14(6):827843, 2006.
[28] Rostislav Hor ck. Cancellative residuated lattices arising on 2-generated submonoids of natural numbers.
Algebra Universalis, 63(23):261274, 2010.
[29] S andor Jenei and Franco Montagna. A proof of standard completeness for Esteva and Godos logic
MTL. Studia Logica, 70(2):183192, 2002.
[30] S andor Jenei and Franco Montagna. A proof of standard completeness for non-commutative monoidal
t-norm logic. Neural Network World, 13(5):481489, 2003.
[31] Y. Katoh, Tomasz Kowalski, and M. Ueda. Almost minimal varieties related to fuzzy logics. Reports on
Mathematical Logic, 41:173194, 2006.
[32] George Metcalfe and Franco Montagna. Substructural fuzzy logics. Journal of Symbolic Logic,
72(3):834864, 2007.
[33] Jeffrey S. Olson. The subvariety lattice for representable idempotent commutative residuated lattices.
To appear in Algebra Universalis.
[34] Francesco Paoli. Substructural Logics: A Primer, volume 13 of Trends in Logic. Kluwer, Dordrecht,
2002.
[35] Greg Restall. An Introduction to Substructural Logics. Routledge, New York, 2000.
[36] Joseph G. Rosenstein. Linear Orderings. Academic Press, New York, 1982.
[37] Clint J. van Alten. The nite model property for knotted extensions of propositional linear logic. Journal
of Symbolic Logic, 70(1):8498, 2005.
[38] Clint J. van Alten. Completion and nite embeddability property for residuated ordered algebras. Alge-
bra Universalis, 62(4):419451, 2009.
[39] Clint J. van Alten. Preservation theorems for MTL-chains. Logic Journal of the Interest Group of Pure
and Applied Logic, 19(3):490511, 2011.
[40] Clint J. van Alten and James G. Raftery. Rule separation and embedding theorems for logics without
weakening. Studia Logica, 76(2):241274, 2004.
[41] Sanmin Wang and Bin Zhao. HpsUL is not the logic of pseudo-uninorms and their residua. Logic
Journal of the Interest Group of Pure and Applied Logic, 17(4):413419, 2009.
ROSTISLAV HOR CIK
Institute of Computer Science
Academy of Sciences of the Czech Republic
Pod Vod arenskou v e z 2
358 Rostislav Hor ck
182 07 Prague 8, Czech Republic
Email: horcik@cs.cas.cz
Chapter V
H ajeks BL Logic and BL-Algebras
MANUELA BUSANICHE AND FRANCO MONTAGNA
1 Introduction
In his famous book [58], H ajek considers the problem of nding a basic fuzzy logic
which is a common fragment of the most important fuzzy logics, namely ukasiewicz,
G odel and product logics. With that purpose, he introduced basic logic (BL), and he
proposed it for the role of basic fuzzy logic. The formulas of BL are built from proposi-
tional variables and from the constant by means of two binary connectives & and .
BL is axiomatized as follows:
(A1) ( ) (( ) ( ))
(A2) ( & )
(A3) ( & ) ( & )
(A4) ( & ( )) ( & ( ))
(A5) ( ( )) (( & ) )
(A6) (( & ) ) ( ( ))
(A7) (( ) ) ((( ) ) )
(A8)
The only rule of BL is modus ponens:
(MP)

H ajeks proposal was greatly supported by [27], where it was shown that BL is the
logic of all continuous t-norms (roughly speaking, a t-norm is a weakly increasing com-
mutative monoid operation on [0, 1] with 1 as neutral element) and their residuals. In
this chapter, we will discuss the equivalent algebraic semantics, in the sense of Blok
and Pigozzi, of BL, that is, the class of BL-algebras. Many algebraic properties of
BL-algebras correspond to logical properties of BL. For instance, the lattice of subvari-
eties of the variety of BL-algebras corresponds to the lattice of axiomatic extensions of
BL, a class Kof algebras generates the whole variety of BL algebras if and only if BL is
360 Manuela Busaniche and Franco Montagna
complete w.r.t. K, etc. Hence, when treating BL-algebras we will implicitly obtain prop-
erties of BL logic. Among other things, a source of interest for this algebraic semantics
is the lack of a reasonable proof theory for BL (all the proof systems introduced so far,
see e.g. [16] are based on semantics and look like model-checking algorithms rather than
genuine sequent calculi). Hence, BL-algebras are the main tool for reasoning inside BL.
Although BL-algebras constitute a rather recent topic, the research on this sub-
ject has produced many signicant results, and nowadays we can say that we know a
fair amount about them. This chapter does not cover all the state of the art (a signif-
icant topic which is not treated here is the subject of free BL-algebras, which will be
discussed in Chapter IX), but we will present some aspects of BL-algebras which in
our opinion are very relevant. The next section deals with some general facts about
BL-algebras. In particular, we will see that BL-algebras form a variety of algebras that
can be presented in many ways: as the variety generated by all continuous t-norm alge-
bras, as the variety of basic (i.e., prelinear) bounded hoops, and as the variety of commu-
tative, integral, bounded, divisible and representable (i.e., prelinear) residuated lattices.
In that section, we also discuss some special subvarieties of the variety of BL-algebras,
namely MV-algebras, product algebras and G odel algebras.
Then, in Section 3, we will study some representation theorems for BL-algebras.
The idea is to decompose BL-algebras into special and better known algebras (in a sim-
ilar way as natural numbers are decomposed as product of special natural numbers, the
prime numbers). First of all, every BL-algebra decomposes into a subdirect product of
totally ordered BL-algebras (also called BL-chains). Hence, we can concentrate on the
decomposition of BL-chains. For algebras arising from a continuous t-norm, there is an
important decomposition theorem in the literature: the Mostert and Schields theorem.
This Theorem states that any continuous t-norm algebra decomposes as an ordinal sum
of ukasiewicz and product t-norm-algebras, called the ukasiewicz (product respec-
tively) components of the t-norm. For general BL-chains, a similar result holds, but
with two basic differences: (1) it is convenient to use a slightly different notion of ordi-
nal sum, and (2) the components need not have a minimum element, and hence product
components are replaced by cancellative totally ordered hoops (roughly, product chains
deprived of 0).
But it is also possible to decompose a BL-algebra in one step: this idea leads to poset
product decomposition, a sort of generalization of ordinal sums and of direct product. It
turns out that BL-algebras are precisely the subdirect poset products of linearly ordered
MV-algebras and product algebras, see [69].
In Section 4 we treat different classes of BL-algebras that generate the whole vari-
ety. First of all, we prove that the variety of BL-algebras is generated as a quasivariety
by the class of all nite BL-chains and by the class of all BL-algebras arising from con-
tinuous t-norms. Then, we characterize the class of all BL-chains which generate the
whole variety of BL-algebras. We use these results to describe the signicant classes of
BL-chains with respect to which BL logic is respectively nitely strongly complete and
strongly complete, and hence we completely clarify the connections between BL logic
and BL-algebras. In Section 5 we investigate some important subvarieties of the vari-
ety of BL-algebras. We study the varieties of BL-algebras generated by ordinal sums of
nitely many Wajsberg hoops, and we prove that for every n > 1 there is a continuum of
Chapter V: H ajeks BL Logic and BL-Algebras 361
varieties of BL-algebras generated by ordinal sums of n Wajsberg hoops. The results of
this section are taken from [2]. Finally, we give a description of varieties generated by a
single BL-algebra arising from a continuous t-norm, see [45] and we give the equational
characterization of some particular subvarieties generated by a single BL-chain.
In Section 6 we discuss the amalgamation property in varieties of BL-algebras. This
property is closely related to the deductive interpolation, see [50]. We will prove that the
class of all BL-algebras has the amalgamation property, and hence BL has the deductive
interpolation property. The same is true for the most important logics extending BL,
like G odel logic G, ukasiewicz logic , product logic and SBL. However, there is a
continuum of subvarieties of BL which do not admit amalgamation, and hence there is a
continuum of extensions of BL which do not have deductive interpolation. The situation
is even worse for Craig interpolation: only a nite number of extensions of BL have
Craig interpolation.
In Section 7 we deal with completions of BL-algebras. A completion of a BL-
algebra A is a complete residuated lattice in which A embeds. There are special kinds
of completions in the literature. Among them, the MacNeille completion, in which the
existing suprema and inma are preserved, and the canonical completion, which has a
compactness property which will be formulated later. Unfortunately, in general neither
the MacNeille completion nor the canonical completion of a BL-algebra need to be a
BL-algebra. We will prove a very general negative result. Let us say that a BL-algebra
is n-potent iff it satises the equation x
n+1
= x
n
, and that a variety of BL-algebras V is
closed under completion iff every algebra in Vembeds into a complete algebra still in V.
Then, a variety of BL-algebras is closed under completions iff it is a variety of n-potent
BL-algebras. After characterizing subvarieties of BL-algebras that admit completion we
study canonical extensions. There are two possible candidates for the canonical exten-
sion of an algebra A, namely the canonical extension A

and the dual canonical exten-

sion A

(see Subsection 7.4). A class of algebras is called -canonical or -canonical

if it is closed under canonical or dual canonical extensions respectively. We analyze
-canonicity for each subvariety of BL-algebras and show that the only subvarieties of
BL-algebras that enjoy -canonicity are those that are nitely generated. On the other
hand, a subvariety of BL-algebras enjoys -canonicity if and only if it is a subvariety
of n-potent BL-algebras. Therefore a BL-algebra admits a completion if and only if it
admits a dual canonical extension.
In Section 8 we collect additional material on BL-algebras and related algebras, and
we suggest some further reading. To nish the chapter, in Section 9 we present some
historical remarks.
To have a better understanding of the present chapter, we recommend to read Chap-
ters II and IV of the present handbook.
2 Denitions and basic properties of BL-algebras
BL-algebras may be presented in three different ways: as the variety generated by
the t-norm-algebras, as bounded basic hoops or as divisible, representable, bounded
(commutative and integral) residuated lattices. We will briey discuss each of these
equivalent presentations.
362 Manuela Busaniche and Franco Montagna
2.1 Continuous t-norms
A t-norm is a binary operation from [0, 1]
2
into [0, 1] satisfying the following
conditions:
1. is commutative and associative,
2. is non-decreasing in both arguments, i.e, for all x, y, z [0, 1]
x y implies x z y z and z x z y,
3. 1 x = x and 0 x = 0 for all x [0, 1].
A continuous t-norm is a t-norm which is continuous as a map from[0, 1]
2
into [0, 1]
in the usual sense. For each continuous t-norm a residuum can be dened (see [58])
satisfying
x z y iff x z y.
The following are the most important examples of continuous t-norms and their
corresponding residua:
1. ukasiewicz t-norm, x

y = max0, x +y 1 and its implication, x

y =
min1, 1 x + y,
2. G odel t-norm, x
G
y = minx, y and its implication,
x
G
y =
_
y if x > y,
1 if x y.
3. Product t-norm, x

y = xy (ordinary product) and Goguen implication,

x

y =
_
y/x if x > y,
1 if x y.
The algebra [0, 1],

, max, min, 0, 1, is called the standard MV-algebra and

is denoted by [0, 1]

. The algebra [0, 1],

G
,
G
, max, min, 0, 1 is called the stan-
dard G odel algebra and is denoted by [0, 1]
G
. Finally, the algebra [0, 1],

,
max, min, 0, 1 is called the standard product algebra and is denoted by [0, 1]

.
From these t-norms, it is possible to construct other t-norms, by means of the so
called ordinal sum construction. To begin with, consider a closed interval [a, b] of the
real line, and take a continuous increasing bijection h from[0, 1] onto [a, b] (for instance,
h(x) = a + (b a)x). Then, every continuous t-norm induces a continuous monoid
operation
[a,b]
on [a, b] dened by x
[a,b]
y = h(h
1
(x) h
1
(y)). Now take any
family of pairwise disjoint open intervals (a
i
, b
i
): i I, with 0 a
i
< b
i
1.
For i I, let
i
be either the ukasiewicz t-norm or the product t-norm. For each
x, y [0, 1] dene:
x y =
_
x
i
[a
i
,b
i
]
y if x, y (a
i
, b
i
)
minx, y otherwise.
Chapter V: H ajeks BL Logic and BL-Algebras 363
It can be shown that the above dened operation is a continuous t-norm, which is called
the ordinal sum of the t-norms
i
[ i I with respect to the intervals [a
i
, b
i
] [ i I.
A famous theorem by Mostert and Schields [84] says that every continuous t-norm is the
ordinal sum of a family of ukasiewicz and of product t-norms. These t-norms are called
the ukasiewicz (product respectively) components of the t-norm . In many textbooks,
a maximal interval [a, b] entirely consisting of idempotents is called a G odel component.
With this terminology, the Mostert and Schields theorem can be rephrased as follows:
THEOREM 2.1.1. Every continuous t-norm can be decomposed into the ordinal sum of
ukasiewicz, product and G odel t-norms.
In [58], P. H ajek introduced, for each xed continuous t-norm , a propositional
calculus L() whose truth values are in the real segment [0, 1], is taken for the truth
function of the (strong) conjunction and the residuum

of becomes the truth func-

tion of the implication. After formulating logical axioms for BL, he proved that each
provable formula in BL is a tautology in each L() (soundness of BL). To prove the
completeness of the logic he started an algebraization of BL. This is the point at which
BL-algebras are introduced.
The completeness problem of the logic BL with respect to the continuous t-norm
semantics can be reformulated in algebraic terms. In order to do so, we need some
preliminary concepts.
Let be a continuous t-norm and let for x, y [0, 1], x

y = supz [0, 1] [
z x y. Then, the algebra [0, 1], ,

, max, min, 0, 1 is called a continuous

t-norm-algebra.
In a continuous t-norm algebra, the following conditions hold:
(1) [0, 1], , 1 is a commutative monoid.
(2) [0, 1], , (where x y = maxx, y and x y = minx, y) is a bounded
lattice with minimum 0 and maximum 1.
(3) (,

) form an adjoint pair, that is, x y z iff x y

z.
(4) x y = x (x

y).
(5) x

y y

x = 1.
DEFINITION 2.1.2. A BL-algebra is an algebra A, , , , , 0, 1 satisfying condi-
tions (1) . . . (5) above.
Recall that a class of algebras is equational if and only if it is a variety, i.e., it is
closed under subalgebras, homomorphic images and direct products. As we shall see
later, BL-algebras form a variety. The following question arises:
Problem: is it true that conditions (1) . . . (5) constitute a complete axiomati-
zation of the variety generated by the class of all t-norm algebras? In other
words, is it true that if an identity in the language , , , , 0, 1 does
not follow from (1) . . . (5), then it fails in some t-norm-algebra?
364 Manuela Busaniche and Franco Montagna
This is the algebraic equivalent of the so called standard completeness problem for
the logic BL. The problem was formulated by Petr H ajek in [57] and then was answered
afrmatively by Cignoli, Esteva, Godo and Torrens in [27]. In view of this result, we are
allowed to dene the class of BL-algebras as follows:
DEFINITION 2.1.3. The class of BL-algebras is the variety generated by the class of
all continuous t-norm algebras.
2.2 Hoops and BL-algebras
A hoop is an algebra A = A, , , 1 of type 2, 2, 0, such that A, , 1 is a
commutative monoid and for all x, y, z A:
1. x x = 1,
2. x (x y) = y (y x),
3. x (y z) = (x y) z.
Hoops were introduced in an unpublished manuscript by B uchi and Owens and
they were deeply investigated in [1, 2, 12, 13, 46]. Some basic properties of a hoop
A = A, , , 1 are:
1. A, , 1 is a naturally ordered residuated commutative monoid, where the order is
dened by x y iff x y = 1 and the residuation is
x y z iff x y z.
2. The partial order on a hoop is a semilattice order, with x y = x (x y).
3. For any x, y, z A the following hold:
(a) 1 x = x,
(b) x 1 = 1, i.e., 1 is the largest element in the order,
(c) x y (z x) (z y),
(d) x y x,
(e) x (x y) y,
(f) x (y z) = y (x z),
(g) x y (y z) (x z),
(h) x y implies y z x z and z x z y,
(i) x y implies x z y z,
(j) x y x.
EXAMPLES 2.2.1.
(1) Let C
n
be the algebra with domain a
0
= 1, a, ..., a
n
and with operations and
dened by a
h
a
k
= a
minh+k,n]
and a
h
a
k
= a
maxkh,0]
. Then, C
n
is
a nite hoop.
Chapter V: H ajeks BL Logic and BL-Algebras 365
(2) Let C

be the algebra with domain a

n
[ n N and with operations and
dened by a
h
a
k
= a
h+k
and a
h
a
k
= a
maxkh,0]
. Then C

is a
cancellative hoop, i.e., a hoop in which a x = a y implies x = y.
(3) Let Int be the interior operator in a topological space (X, T). Then the set of all
open elements of the space form a hoop with respect to = and dened by
Y Z = Int((X Y ) Z).
(4) Let be a continuous t-norm and

its residuum. Then [0, 1], ,

, 1 is a
hoop.
(5) The algebra (0, 1]

= (0, 1],

, 1, where

is ordinary product and

is dened by x

y = min
y
x
, 1 is a totally ordered hoop without minimum.
(6) Let C = 0 < < c
n+1
< c
n
< c
0
be a chain, and consider the poset P
consisting of C plus two incomparable elements a, b with 0 < a, b < c
n
for every
n. Let x y = infx, y and x y = 1 if x y and x y = y otherwise.
Then, P, , , c
0
is a hoop which is not closed under join.
A basic hoop is a hoop that satises the equation:
(((x y) z) ((y x) z)) z = 1
In every basic hoop Aan operation can be dened by
x y = ((x y) y) ((y x) x),
thus L(A) = A, , , 1 is a lattice with greatest element 1. In addition, every basic
hoop Asatises the equation:
(x y) (y x) = 1.
REMARK 2.2.2. Every totally ordered hoop is basic, but the converse does not hold.
Indeed, basic hoops form an equational class and hence they are closed under direct
products, but direct products of non-trivial hoops are not totally ordered.
We are going to introduce an alternative denition of BL-algebra (the equivalence
between the next denition and Denition 2.1.2 will be shown later).
DEFINITION 2.2.3. A BL-algebra is a bounded basic hoop, that is, it is an algebra
A = A, , , 0, 1 of type 2, 2, 0, 0 such that A, , , 1 is a basic hoop and 0 is the
lower bound of L(A).
According to this denition, a set B A is the universe of a subalgebra of a BL-
algebra A iff 1, 0 B and B is closed under and . Furthermore, if C A is a set
closed under and such that 1 C, then C = C, , , 1 is a basic hoop. For any
integer k, a BL-term in variables x
1
, x
2
, . . . x
k
is a string over the set S
k
= , , 0, 1,
x
1
, x
2
, . . . , x
k
, (, ) that arises froma nite number of application of the following rules:
1. 0, 1, x
1
, x
2
, . . . , x
k
are BL-terms,
2. if
1
and
2
are BL-terms, then (
1

2
) and (
1

2
) are BL-terms.
366 Manuela Busaniche and Franco Montagna
On each BL-algebra, the unary operation (negation) is dened by the equation:
x = x 0.
The BL-algebra Awith only one element, that is 0 = 1, is called the trivial BL-algebra.
The varieties of BL-algebras and basic hoops will be denoted by B' and BH, respec-
tively. It is known (see [63]) that both varieties are congruence distributive and congru-
ence permutable.
Let Abe a basic hoop. As previously mentioned, we denote by the (partial) order
dened on A by the lattice L(A), i.e. for a, b A, a b iff a = a b iff b = a b.
Interestingly, we have x y iff there is a z such that yz = x. Hence, this order is called
the natural order of A. When this natural order is total (i.e., for each a, b A, a b or
b a), Ais called a totally ordered hoop or a BL-chain in case Ais a BL-algebra.
2.3 BL-algebras as residuated lattices
DEFINITION 2.3.1 ([50, 90]). A commutative and integral residuated lattice (in the
sequel, just residuated lattice) is a system L, , , , , 1 where:
(a) L, , 1 is a commutative monoid.
(b) L, , is a lattice with top element 1.
(c) is a binary operation such that the residuation property holds:
x y z iff x y z.
Residuated lattices constitute the algebraic counterpart of substructural logics. In
[50], a more general denition of residuated lattice is given (the operation need not be
commutative and 1 need not be the top element), with the purpose to nd an algebraic
counterpart of non-commutative logics or of logics without weakening. In this more
general presentation, two implications are needed for the left and for the right residual.
Despite this, the denition above is sufciently general for our purposes.
The only non-equational dening property of residuated lattices is (c). However, in
the presence of commutativity residuation may be expressed by the equations (see [62])
x (y z) = (x y) (x z)
x (y z) = (x y) (x z)
(x (x y)) y = y
(x (x y)) y = y
Hence, residuated lattices constitute a variety.
DEFINITION 2.3.2. A residuated lattice is divisible iff it satises the identity x y =
x (x y), prelinear if it satises the identity (x y) (y x) = 1, cancellative
if it satises the identity x (x y) = y and bounded if it has a minimum element
m with respect to the order induced by the lattice operations, and its signature has an
additional constant 0 which is interpreted as m.
Chapter V: H ajeks BL Logic and BL-Algebras 367
The equation xy = x (x y) is equivalent to the following condition: if x y,
then there is a z (thought of as the division of x by y, whence the name divisible) such
that y z = x. The equation (x y) (y x) = 1 is satised in a residuated
lattice L iff L embeds into a subdirect product of linearly ordered residuated lattices,
whence the name prelinear. Finally, bounded residuated lattices correspond to logics
(with weakening and) with a propositional constant for falsum.
EXAMPLES 2.3.3. The following are examples of residuated lattices:
1. A Heyting algebra is a bounded residuated lattice in which and coincide.
Heyting algebras constitute the algebraic counterpart of intuitionistic logic. It is
easy to see that Heyting algebras are divisible and bounded residuated lattices, but
they are neither cancellative nor prelinear.
2. A lattice ordered Abelian group (-group for short) is an algebra G, ,
1
, , , 1
where G, ,
1
, 1 is an Abelian group, G, , is a lattice and the equation
x (y z) = (x y) (x z) holds.
The negative cone of an Abelian -group G = G, ,
1
, , , 1 is the algebra
G

= G

, , , , , 1 where G

= x G [ x 1, , , are the restric-

tions to G

of the analogous operations of G and is the binary operation de-

ned by x y = (x
1
y) 1.
The negative cone of an Abelian -group is a divisible and prelinear residuated
lattice, but it is not bounded.
3. Let A = [0, 1], , , max, min, 0, 1 where:
[0, 1] is the unit real interval,
x y =
_
min x, y if x + y > 1
0 otherwise
x y =
_
max 1 x, y if x > y
1 otherwise
Then A is a bounded and prelinear residuated lattice, but is neither cancellative
nor divisible. To check the failure of divisibility, let x =
1
3
and y =
1
4
. Then
x (x y) = 0, but min x, y =
1
4
.
Denition 2.1.2 can be rephrased as follows:
DEFINITION 2.3.4. A BL-algebra is a bounded divisible prelinear residuated lattice.
2.4 BL-chains and completeness
We have three different presentations of BL-algebras, namely:
the algebras in the variety generated by continuous t-norms (Denition 2.1.2),
bounded, divisible and prelinear residuated lattices (Denition 2.3.4),
bounded basic hoops (Denition 2.2.3).
368 Manuela Busaniche and Franco Montagna
We are going to prove that these three denitions of BL-algebras are equivalent.
Since the equivalence between the rst two denitions is trivial, it is sufcient to prove:
THEOREM 2.4.1.
(1) The , , 0, 1 reduct of a divisible, prelinear and bounded residuated lattice is
a bounded basic hoop.
(2) Let A be a bounded basic hoop. Let us dene x y = x (x y) and x y =
((x y) y) ((y x) x). Then, and are lattice operations which
make Aa bounded, divisible and prelinear residuated lattice.
Proof. (1) The only non-trivial property to verify is the inequality
(x y) z ((y x) z) z.
By [58, Lemma 2.3.16], it is sufcient to verify it in linearly ordered BL-algebras. Thus,
we can suppose, without loss of generality, that either x y or y x. If x y, the
above identity reduces to z ((y x) z) z, which is obviously true, and if
y x, then the equation reduces to (x y) z z z, which is also trivially true.
(2) It is sufcient to prove the claim for subdirectly irreducible bounded basic
hoops, which are linearly ordered. Suppose, without loss of generality, x y. Then,
minx, y = x = x 1 = x (x y) and maxx, y = y = 1 y = (x y) y.
Moreover, by the residuation property, y (y x) x and y (y x) x.
Hence maxx, y = y = (x y) y (y x) x. Therefore the above de-
ned operations and are the join and the meet operation, and make A a prelinear
and divisible residuated lattice. Modulo an extension by denitions, this denition is
equivalent to the denition of BL-algebras as bounded basic hoops. Indeed, bounded
basic hoops are subdirect products of totally ordered hoops. Moreover, in a totally or-
dered basic hoop, the meet and the join are denable by x y = x (z y) and
x y = ((x y) y) ((y x) x), and hence these denitions extend to all
bounded basic hoops. It follows that a bounded basic hoop is a divisible, prelinear and
bounded residuated lattice, that is, a BL-algebra.
Nowthat we have proved the equivalent denitions, we recall the subdirect represen-
tation theorem for BL-algebras, which is crucial for our study (see [58, Lemma 2.3.16],
[1, Theorem 2.8], and Chapter II of the present handbook).
THEOREM 2.4.2. Each BL-algebra is a subdirect product of BL-chains.
Therefore the lattice L(A) is distributive for each BL-algebra A.
Now is the appropriate moment to describe the connections between the logic BL
and the variety of BL-algebras. Recall that a valuation in a BL-algebra A is a homo-
morphism from the algebra of BL-formulas into A. Given a class K of BL-algebras, a
formula and a set of formulas , we say that:
is valid in K (and we write [=
K
) if for every A K and every valuation v in
A, we have v() = 1.
is semantic consequence of in K (and we write [=
K
) if for every A K
and every valuation v in A, we have that, if v() = 1 for all , then v()= 1.
Chapter V: H ajeks BL Logic and BL-Algebras 369
BL is complete with respect to K iff for every formula , we have that [=
K
iff

BL
, that is, is a theorem of BL.
BL is strongly complete (nitely strongly complete respectively) with respect to K
iff for every set of formulas (for every nite set of formulas respectively) and
for every formula , we have that [=
K
iff
BL
.
Then H ajek [58] proved the following:
THEOREM 2.4.3. BL is strongly complete with respect to both the class of all BL-
algebras and the class of all BL-chains.
Because the class of all continuous t-norm BL-algebras generates the whole variety
of BL-algebras, we obtain: Add some forward reference to where it is proved here in
this chapter; do not use more than two lines!
THEOREM 2.4.4. BL is complete with respect to the class of all continuous t-norm
BL-algebras.
Note that BL is not strongly complete with respect to this class. Consider the set
= p
n+1
(p
n
& p
n
) [ n N p p
n
[ n N, and let = p (p
0
& p).
It is readily seen that if N denotes the class of all continuous t-norm BL-algebras, then
[=
TN
, but -
BL
.
2.5 Important subvarieties of BL-algebras.
Some subvarieties of B' have been studied for their own importance, since they are
the algebraic counterpart of some well known logics. MV-algebras, for instance, the
algebras of ukasiewicz innite-valued logic, form the subvariety of B' characterized
by the equation:
x = x
(see [58]). MV-algebras are usually presented in terms of the two connectives and
, where x y = (x) y, instead of and , but the two presentations are term-
equivalent. For references about these algebras see [26]. The variety of MV-algebras is
denoted by MV and a totally ordered MV-algebra is called an MV-chain.
A G odel algebra (or prelinear Heyting algebra) H = H, , , 0, 1 is a Heyting
algebra (or relatively pseudocomplemented bounded distributive lattice, see [9]) which
satises the equation:
(x y) (y x) = 1.
These algebras are the algebraic counterpart of a superintuitionistic logic, called G odel
Logic or Dummett Logic and denoted by G, which is axiomatized over intuitionistic
logic by the axiom ( ) ( ). Observe that any G odel algebra H satises
the equations:
x y = x (x y),
x y = ((x y) y) ((y x) x).
Thus a G odel algebra H is a BL-algebra in which = , i.e., it satises the equation
x y = x y.
370 Manuela Busaniche and Franco Montagna
REMARK 2.5.1. It is easy to see that a BL-algebra is a G odel algebra if and only if it
satises the equation x x = x.
A product algebra is a BL-algebra that satises the following two equations:
(z ((x z) (y z))) (x y) = 1,
x x = 0.
Product algebras correspond to product fuzzy logic, denoted by , see [28, 58].
REMARK 2.5.2. Product algebras can be axiomatized over BL-algebras by the single
equation
(x) (((x y) x) y) = 1.
Of course, each of the standard t-norm based algebras introduced in Section 2.1 be-
longs to one of these varieties. More precisely, the standard MV-algebra [0, 1]

belongs
to MV, the standard G odel algebra [0, 1]
G
belongs to the subvariety of G odel algebras
and the standard product algebra [0, 1]

is a product algebra.
An SBL-algebra (pseudocomplemented BL-algebra) is a BL-algebra satisfying the
equation
x x = 0.
This equation can be interpreted in several ways: it says that negation is strict (i.e., in an
SBL-chain, the negation of any element is either 0 or 1). But it can also be interpreted
as a weak form of contraction, as it can be written as x x x, and hence as
x
2
0 x 0. In logical terms, if & implies a contradiction, then itself implies
a contradiction. Hence, SBL-algebras can be regarded as the algebraic counterpart of
the logic SBL dened as BL plus (( & ) ) ( ).
Finally, adding the equation
x x = 1
to the variety of SBL-algebras one gets the variety of Boolean algebras. For each BL-
algebra A, the complemented elements of L(A) form a subalgebra B(A) of A which
is a Boolean algebra. Elements of B(A) are called Boolean elements of A.
2.6 Implicative lters
An implicative lter of a BL-algebra A is a subset F A satisfying the following
conditions:
1. 1 F,
2. If x F and x y F, then y F.
An implicative lter is called proper provided F ,= A. Note that every implicative
lter F of a BL-algebra Ais the universe of a subhoop of A. An implicative lter F of
a BL-algebra A is called maximal iff it is proper and no proper implicative lter of A
strictly contains F.
Chapter V: H ajeks BL Logic and BL-Algebras 371
REMARK 2.6.1. If X is a subset of a BL-algebra A, the implicative lter generated by
X, that will be denoted by X, is the set
x A [ x
1
X...x
n
Xsuch that x
1
... x
n
x,
where repetitions are allowed, i.e., we may have x
i
= x
j
for some i ,= j.
Implicative lters characterize congruences in BL-algebras. If F is an implicative
lter of a BL-algebra A, then the binary relation
F
on A dened by:
x
F
y iff x y F and y x F
is a congruence of A (see [58, Lemma 2.3.14]). Moreover, F = x A [ x
F
1.
Conversely, if is a congruence relation on A, then the set F = x A [ x 1
is an implicative lter, and x y iff x y 1 and y x 1. Therefore, the
correspondence
F
F
is an order isomorphism from the set of implicative lters of A onto the set of congru-
ences of A, both ordered by inclusion.
Given a BL-algebra A and a lter F of A, we will denote the quotient set A/
F
by A/F. Since
F
is a congruence, dening on the set A/F the operations
(x/F) (y/F) = (x y)/F
and
(x/F) (y/F) = (x y)/F,
the system A/F, , , 0/F, 1/F becomes a BL-algebra called the quotient algebra of
Aby the implicative lter F. Moreover, the correspondence
x x/F
denes an homomorphism h
F
from Aonto the quotient algebra A/F.
An implicative lter F of a BL-algebra A well be called Boolean provided that
A/F is a Boolean algebra and it will be called an MV-lter in case A/F is an MV-
algebra.
2.7 Regular and dense elements
In the paper [31] there are some important results relating ukasiewicz innite-
valued logic and the logic BL. Recall that MV-algebras form the proper subvariety of
BL-algebras characterized by the equation:
x = x.
For details about MV-algebras the reader can see Chapter VI of the present handbook
and [26]. Given a BL-algebra Awe can consider the set
MV(A) = x A [ x = x.
The elements of MV(A) will be called regular elements of A. Since in any BL-algebra
the following equations are satised:
372 Manuela Busaniche and Franco Montagna
1. 0 = 0,
2. x = x,
3. (x y) = x y,
4. (x y) = x y,
the system MV (A) = MV(A), , , 0, 1 is subalgebra of A which is an MV-
algebra and the mapping
x x
is a retract from the BL-algebra Aonto MV (A).
Let A be a BL-algebra. Since denes a retraction from A into MV (A), for
each valuation v from the set of formulas of BL into A, the map v
t
dened as v
t
() =
v() is a valuation from the set of formulas of ukasiewicz innite-valued logic into
the MV-algebra MV (A). With this in mind, a result relating provability in BL with
provability in ukasiewicz innite-valued logic, that resembles the Glivenko theorem
relating classical and intuitionistic logic can be stated as follows (see [31, Theorem2.1]):
THEOREM 2.7.1. Let be a propositional formula in the language of BL and let
be the formula . Then:
1.
BL
if and only if is derivable in ukasiewicz innite-valued logic.
2.
BL
if and only if is derivable in ukasiewicz innite-valued logic.
Proof. If
BL
, then is a theorem of and hence is a theorem of . Con-
versely, if -
BL
, from Theorem 2.4.3, there is BL-chain B and a valuation v into B
such that v() < 1. But is a homomorphism from B into MV (B), and hence
the map v
t
: v() = v() is a valuation into MV (B). It follows that
v
t
() = v() < 1. Completeness of implies that is not derivable in ukasiewicz
innite-valued logic. The second statement follows from the rst, since is equiv-
alent to .
For a BL-algebra Awe dene
D(A) = x A [ x = 0.
The elements in D(A) will be called dense elements of A. It is easy to check that
D(A) = D(A), , , 1 is a proper implicative lter of A, and the congruence deter-
mined by D(A) is given by
x
D(A)
y iff x = y.
Therefore
A/D(A)

= MV (A).
Given an element x in a BL-algebra Ait is always the case x x. Then we have
x = x x = x (x x),
i.e., every element in a BL-algebra is the product of a regular element and a dense
element. This factorization need not be unique.
Chapter V: H ajeks BL Logic and BL-Algebras 373
3 Decomposition theorems for BL-algebras
In this section we will show that BL-algebras can be decomposed into special BL-
algebras (or into special subreducts of BL-algebras). We will examine rst the case of
totally ordered BL-algebras, and then the general case.
3.1 Ordinal sums
FromTheorem2.4.2 we can deduce that BL-chains play a key role in the structure of
BL-algebras. One way of characterizing BL-chains is studying the number and form of
some of their subhoops. To describe how the operations of a BL-chain Abehave among
some of its proper subhoops we use the notion of ordinal sum introduced by B uchi and
Owens in an unpublished manuscript and recalled in [46]. It is worth remarking that this
notion does not coincide with the notion of ordinal sum given in [27].
Let I, be a totally ordered set. For each i I, let A
i
= A
i
,
i
,
i
, 1 be a
hoop such that for every i ,= j, A
i
A
j
= 1. Then we can dene the ordinal sum as
the hoop

iI
A
i
=
iI
A
i
, , , 1 where the operations , are given by:
x y =
_

_
x
i
y if x, y A
i
,
x if x A
i
1, y A
j
and i < j,
y if y A
i
1, x A
j
and i < j.
x y =
_

_
1 if x A
i
1, y A
j
and i < j,
x
i
y if x, y A
i
,
y if y A
i
, x A
j
and i < j.
If in addition I has a minimumi
0
and A
i
0
is a bounded hoop, then

iI
A
i
denotes
the bounded hoop whose operations and are dened as before, and whose bottom
element is the minimum of A
i
0
. Note that, in general

iI
A
i
need not be a BL-
algebra, not even if the hoops A
i
are basic hoops and A
i
0
is a BL-algebra. However,
this is the case if in addition every A
i
is totally ordered.
In the sequel, we denote by A
1
A
2
A
n
the ordinal sum of n hoops
A
1
, . . . , A
n
, where the order of the index set is 1 < 2 < < n.
REMARK 3.1.1. In every residuated lattice, the monoid operation distributes over arbi-
trary existing suprema, that is
y

X =

(y X) =

y x [ x X.
The next example shows that even in a complete BL-chain, product does not distribute
over innite inma. Let A = L
3
C, where L
3
is the three-element MV-chain and C
is the negative cone of Z. Let a be the coatom of L
3
; then a
_
C ,=
_
(a C).
3.2 Decomposition of BL-chains into irreducible hoops
The results of this section can be found in [2, 20]. A Wajsberg hoop is a hoop that
satises the equation:
(x y) y = (y x) x.
374 Manuela Busaniche and Franco Montagna
A Wajsberg algebra is a bounded Wajsberg hoop. It is well-known that Wajsberg alge-
bras are term-equivalent to MV-algebras (see e.g. [26, 48]).
For brevity, we shall use the terminology o-hoop for a totally ordered hoop. An o-
hoop is irreducible if it cannot be written as the ordinal sum of two non-trivial o-hoops.
LEMMA 3.2.1. Let Abe any o-hoop and let a ,= 1 be an element in A. Dene
F
a
= x A 1 [ a x = x.
Then F
a
is downwards closed, and F
a
1 is the domain of a subhoop F
a
of A.
Proof. Observe that any x F
a
satises x < a and a x = x. In details, x = x a =
a(a x) = ax. Let x F
a
and assume y x. Since it is always the case y a y,
it is enough to see that (a y) y = 1. First note that a y a x = x. Then
(a y) y = (x (a y)) y = (x (x (a y))) y
= (x (a y)) (x y) = ((a x) (a y)) (x y)
= ((a (a x)) y) (x y) = ((x a) y) (x y)
= (x y) (x y) = 1.
This proves that F
a
is downwards closed. It follows that F
a
(hence F
a
1) is
closed under . Finally, F
a
1 is closed under , because if x, y F
a
1 then
a (x y) = (a x) y = x y.
Under the hypothesis and notation of the previous lemma we get:
LEMMA 3.2.2. Let a, b A be such that a < b < 1. Then
(i) F
a
F
b
.
(ii) If F
a
,= F
b
, then a F
b
.
(iii) If F
a
,= F
b
, then a b = mina, b.
Proof. For (i), take x F
a
. Hence a x = x and so x b x a x = x,
thus x F
b
. To obtain (ii), take y F
b
F
a
. Therefore b y = y and y < a y.
This last inequality implies that (a y) y ,= 1. Since a (a y) y and F
b
is
downwards closed it is enough to see that (a y) y F
b
. Clearly
b ((a y) y) = (a y) (b y) = (a y) y
thus (a y) y F
b
as desired. Lastly, for (iii), by (ii) we get a F
b
. Thus
b a = b (b a) = b a, and since A is a totally ordered, a b = mina, b.
THEOREM 3.2.3. For an o-hoop (BL-chain) Athe following are equivalent:
(i) Ais irreducible,
(ii) For all a, b A, b a = a implies b = 1 or a = 1,
Chapter V: H ajeks BL Logic and BL-Algebras 375
(iii) Ais a Wajsberg o-hoop (Wajsberg chain).
Proof. (i) (ii). Assume that Ais irreducible and let a, b A be such that b a = a.
If b ,= 1, Lemma 3.2.1 implies that F
b
is a subhoop of A. Consider S = A F
b
. S
is the universe of a subhoop S of A. In details, S is upwards closed, for any x S,
x y x yields that S is closed under and if x, y S it is easy to see that x y S.
Moreover, A = F
b
S. The irreducibility of A together with the fact that a F
b
,
imply a = 1.
(ii) (iii) We reproduce the proof given in [13]. Without loss of generality assume
that a < b with the intend of proving that (a b) b = (b a) a. Due to
our assumption we have (a b) b = 1 b = b. Therefore we only need to
prove that (b a) a = b, or equivalently ((b a) a) b = 1. Since
(b a) = ((b a) a) a, we obtain
(((b a) a) b) (b a)
= (((b a) a) b) (((b a) a) a)
= ((((b a) a) b) ((b a) a)) a
= (b (b ((b a) a))) a
= (b ((b a) a))) (b a)
= 1 (b a) = b a.
The assumptions (ii) and a < b yield ((b a) a) b = 1 as desired.
(iii) (i). Assume that A = B C for some non-trivial subhoops B and C of
A. Take b B 1 and c C 1. We get (c b) b = b b = 1 and
(b c) c = 1 c = c, thus Ais not a Wajsberg hoop.
A tower of irreducible o-hoops is a family = C
i
[ i I indexed by a totally
ordered set I, with rst element 0 such that:
1. C
i
= C
i
,
i
,
i
, 1 is an irreducible o-hoop,
2. C
i
C
j
= 1 for each i ,= j,
3. C
0
is a bounded o-hoop.
It is easy to see that for each tower = C
i
[ i I of irreducible o-hoops,
A

iI
C
i
is a BL-chain. We shall demonstrate the following theorem that gives
the unique decomposition of each BL-chain into an ordinal sum of irreducible hoops:
THEOREM 3.2.4. Each BL-chain A is isomorphic to an algebra of the form A

for
some tower of irreducible o-hoops.
376 Manuela Busaniche and Franco Montagna
Proof. We have already noticed that if an algebra is of the form A

for some tower of

irreducible o-hoops , then the algebra is a BL-chain.
We only need to prove that each BL-chain Ahas this form. As in Lemma 3.2.1, for
each a A, a ,= 1, let F
a
= x A 1 [ a x = x and let F
1
= 1. Dene the
relation on Aby:
a b iff x A, a x = x b x = x iff F
a
= F
b
.
Clearly is an equivalence relation. Lets check that for each equivalence class C,
the structure C
t
= C 1, , , 1 is a Wajsberg o-hoop.
(1) C
t
is totally ordered, because the order of Ais inherited.
(2) C
t
is closed under . In detail, let a, b C (that is, a b) with the intent of
proving that a b a. The fact that a b a and (ii) in Lemma 3.2.2 yield
F
ab
F
a
. For the opposite inclusion, let x F
a
= F
b
. Then
a b x = b (a x) = b x = x,
and x F
ab
. Moreover since x 1 = x for every x C and 1 1 = 1, we obtain
that C
t
is closed under .
(3) C
t
is closed under . Let a, b C. We intend to see that b a C
t
.
1. If b a, then b a = 1 C
t
.
2. If a < b, since a a = 1 ,= a, then a / F
a
= F
b
. Therefore a < b a
and b (b a) = b a = a ,= b. Hence b (b a) ,= minb, b a. By
(iii) in Lemma 3.2.2, b a and b are equivalent, thus b a C
t
.
On the other hand, if a C, then 1 a = a, a 1 = 1. Since 1 1 = 1, we
conclude that C
t
is closed under .
At this point we deduce from claims (1), (2) and (3) that for each equivalence
class C, C
t
is an o-hoop.
(4) C
t
is irreducible. Suppose conversely that C
t

= A
1
A
2
, for some non-trivial
hoops A
1
and A
2
. Let x A
1
1 and y A
2
1. Then x, y C and
F
x
= F
y
. From the denition of ordinal sum we have that y x = x < 1.
Therefore x F
y
, or equivalently x F
x
. But this can not happen because
x x = 1 ,= x. Then C
t
is irreducible.
(5) C is a convex set. Assume that a, b C and a < u < b. From (i) in Lemma 3.2.2,
we conclude that F
a
F
u
F
b
. But since a b, we have that F
a
= F
u
= F
b
.
Let I be the set of equivalence classes C. Since each equivalence class is a convex
set, the order of A induces an order on I, dened for C, D I, by C _ D iff either
C = D or for all x C and for all y D one has that x y. Thus I is a totally ordered
set. We shall denote by C
0
the equivalence class that contains the bottom element of A,
and by C
1
the class that contains the element 1. Therefore we have that:
Chapter V: H ajeks BL Logic and BL-Algebras 377
1. For each C I, C
t
is an irreducible o-hoop.
2. For each C ,= D I, C
t
D
t
= 1, for equivalence classes are pairwise
disjoint.
3. C
t
0
has a least element, because Ais a BL-chain.
Therefore = C
t
[ C I is a tower of irreducible o-hoops. Notice that if a C
and b D with C _ D and C ,= D, by (iii) and (ii) in Lemma 3.2.2, we conclude that
b a = a and a b = a. Then it is easy to verify that A

CI
C
t

= A.
It is clear that C
t
1
= 1. The previous theorem offers a constructive method for
decomposing BL-chains into non-trivial irreducible o-hoops.
THEOREM 3.2.5. Each non-trivial BL-chain admits a unique decomposition into non-
trivial irreducible o-hoops.
Proof. Suppose that A =

iI
C
i
=

jJ
D
j
, where C
i
and D
j
are non-trivial
irreducible o-hoops for each i I and j J, I and J are totally ordered sets. For each
i I and j J, one of the following cases occurs:
1. C
i
D
j
= 1 or,
2. there exists a C
i
D
j
such that a < 1.
We only need to see that if the second case happens, then C
i
= D
j
. Suppose on the
contrary, that C
i
,= D
j
, and let a < 1 be in C
i
D
j
. Without loss of generality
we can assume that there exists b C
i
D
j
and a < b. Observe that b / D
j
, thus
b < 1. Since b A, there exists k J such that b D
k
and clearly j < k. From
the denition of ordinal sum b a = a. But, since a, b C
i
and C
i
is irreducible,
Theorem 3.2.3 asserts that b = 1 or a = 1. The contradiction arises from the hypothesis
that C
i
,= D
j
.
Summarizing the previous results we obtain:
THEOREM 3.2.6. Each non-trivial BL-chain admits a unique (up to isomorphism) de-
composition as ordinal sum of non-trivial Wajsberg o-hoops.
We have noticed that bounded Wajsberg chains (irreducible bounded o-hoops) co-
incide with MV-chains. A characterization of bounded and unbounded o-hoops is given
in [13, Section 1]. A hoop is cancellative if its basic monoid is cancellative. Cancellative
hoops form a variety characterized by the equation
y = x (y x).
Cancellative o-hoops coincide with unbounded Wajsberg o-hoops (see [13]).
The most general example of an unbounded Wajsberg hoop is the negative cone of
an -group. In fact, the following result can be deduced from [4] (see also [13, 28]).
THEOREM 3.2.7. The following are equivalent conditions for a basic hoop A:
378 Manuela Busaniche and Franco Montagna
1. Ais a cancellative hoop,
2. there is an -group Gsuch that A

= G

,
3. A is in the variety of basic hoops generated by Z

, where Z denotes the additive

group of integers with the usual order.
Therefore we conclude:
THEOREM 3.2.8. If A is an irreducible o-hoop then either A is the hoop reduct of an
MV-chain (in case the set A is bounded), or A is isomorphic to G

for some totally

ordered Abelian group G.
COROLLARY 3.2.9. Each BL-chain is an ordinal sum of a family of hoops each of
which is an MV-chain, the hoop reduct of an MV-chain or it is of the form G

for a
totally ordered Abelian group G.
Let us consider a non-trivial BL-algebra A. We have proved that there exists a
unique tower of non-trivial irreducible o-hoops = C
i
[ i I such that A =

iI
C
i
. If i
0
denotes the rst element in I, then C
i
0
is a bounded Wajsberg o-hoop,
i.e, C
i
0
is an MV-chain. Besides, B =

iI\i
0
]
C
i
is an implicative lter of A.
Recall that a BL-algebra A is said to be simple provided it is non-trivial and the only
proper implicative lter of Ais the singleton 1. Therefore:
LEMMA 3.2.10. A BL-chain Ais simple iff it is simple MV-chain.
For n 2, we dene:
L
n
=
_
0
n 1
,
1
n 1
,
2
n 1
, . . . ,
n 1
n 1
_
.
The set L
n
equipped with the operations xy = max0, x+y1, x y = min1, 1
x + y, denes a nite MV-chain which shall be denoted by L
n
. Any nite MV-chain
is isomorphic to a chain of the form L
n
for some n 2 (see [26, Corollary 3.5.4])
Since nite Wajsberg o-hoops are reducts of nite MV-chains, with an abuse of
notation we shall refer to both algebras by the same symbol, and we will deduce their
structure from the context. Therefore L
n
will denote the MV-chain L
n
, , , 0, 1 as
well as the o-hoop L
n
, , , 1. We shall understand L
n
L
m
as the ordinal sum of
the MV-chain L
n
and the o-hoop L
m
.
Clearly, any nite o-hoop is bounded. Hence, the results of Theorem 3.2.8 and
Corollary 3.2.9 imply that a nite BL-chain can be uniquely decomposed into an ordinal
sum a nite MV-chain and nitely many reducts of nite MV-chains. These observa-
tions lead us to conclude:
THEOREM 3.2.11. Each nite BL-chain C is isomorphic to a chain of the form
k

i=0
L
r
i
for an integer k 0 and where each r
i
2 is an integer for i = 0, 1, . . . , k.
Chapter V: H ajeks BL Logic and BL-Algebras 379
REMARKS 3.2.12. The next results are stated without proof. The reader is invited to
verify them.
(1) A BL-algebra is said to be n-potent if it satises the equation x
n+1
= x
n
. Then,
every n-potent BL-chain is the ordinal sum of a tower of MV-algebras of the form
L
k
with k n + 1.
(2) Of course, the ordinal sum decomposition theorem works for G odel and for prod-
uct chains as well. While G odel chains are the ordinal sums of copies of the two
element MV-algebras which have a minimum component, product chains are the
ordinal sums of the two- elements MV-chain and a cancellative o-hoop.
(3) The Mostert and Schields theorem can be derived as a particular case from the
Aglian o-Montagna decomposition theorem.
(4) Any n-generated BL-chain is the ordinal sum of at most n + 1 Wajsberg compo-
nents, and any n-generated o-hoop is the ordinal sum of at most n components.
(5) A BL-algebra is subdirectly irreducible if and only if it is an ordinal sum of a
tower of Wajsberg chains such that the tower has a last element which is a subdi-
rectly irreducible Wajsberg hoop.
Hint for (3): (a) decompose the continuous t-norm algebra A as an ordinal sum of
Wajsberg hoops; (b) prove that if W is a Wajsberg component, then W 1 has a
supremum and an inmum, which are both idempotent elements; (c) prove that if W is
a cancellative component, then W inf(W) is the domain of a subhoop of A which
is a product algebra with minimum inf(W). Then the intervals for the decomposition
are precisely the intervals [inf(W), sup(W 1)], where W is a Wajsberg component
with more than two elements.
Hint for (4). In an o-hoop, one can prove that if a
i
U
i
, i = 1, ..., n, then for every
term t(x
1
, ..., x
n
), t(a
1
, ..., a
n
)

n
i=1
U
i
. (Induction on t). For BL-chains we must
also consider 0, which belongs to the rst component.
Hint for (5). A subdirectly irreducible BL-algebra A is totally ordered, and hence
it decomposes as an ordinal sum of a tower of Wajsberg hoops. Let F be the minimum
lter. Then there is at most one Wajsberg component W of Asuch that W F _ 1,
and this component must be the last one. Finally, F must be the minimum lter of W.
3.3 Decomposition into regular and dense elements
We present now an alternative way of decomposing BL-chains that can be useful for
some purposes. Given a BL-algebra Arecall that the sets of dense elements and regular
elements are given respectively by
D(A) = x A [ x = 0 and MV(A) = x A [ x = x.
THEOREM 3.3.1. For each BL-chain A, we have
A

= MV (A) D(A).
380 Manuela Busaniche and Franco Montagna
Proof. From Theorem 3.2.4 we know that there exists a tower = C
i
[ i I of
irreducible o-hoops, such that A =

iI
C
i
. Let 0 be the least element of I and let
B =

iI\0]
C
i
. Clearly B is an o-hoop and A = C
0
B. Therefore we only need
to prove that MV (A) = C
0
and that D(A) = B.
1. MV (A) = C
0
. Clearly C
0
MV(A), because C
0
is an MV-chain. Suppose
now that there exists x MV(A)C
0
. Thus x ,= 1. Hence x B, and x = 1.
But since x MV(A), x = x. The contradiction x = 1 results from the
hypothesis MV(A) C
0
,= , hence we may conclude that MV (A) = C
0
.
2. D(A) = B. If x B, then from the denition of ordinal sum x = 0 and so
x D(A). Hence B D(A). Assume that x D(A) B. Obviously x ,= 1.
Since x D(A), we have that x = 1. On the other hand, since x C
0
1,
we obtain that x = x, because C
0
is an MV-chain. The contradiction x = 1
make us conclude that B = D(A).
REMARK 3.3.2. Given a BL-algebra A, if MV (A)

= L
n
for some n 2, then
A

= MV (A) D(A).
3.4 Poset products
In this section, we propose a generalization of the ordinal sum construction which
will allow us to represent all BL-algebras (hence, not only the BL-chains) by means of
a poset-indexed family of MV-algebras and of product algebras. The results are taken
from [68, 69]. Observe that poset products are called poset sums in [68].
DEFINITION 3.4.1. Let P = P, be a poset and let A
p
[ p P be a collection of
commutative, integral and bounded residuated lattices. Up to isomorphism we can (and
we will) assume that all A
p
share the same neutral element 1 and the same minimum
element 0. The poset product

pP
A
p
is the algebra Adened as follows:
(1) The domain of A is the set of all maps h

pP
A
p
such that for all p P if
h(p) ,= 1, then for all q > p, h(q) = 0.
(2) The monoid operation and the lattice operations are dened pointwise.
(3) The residual is as follows:
(h g)(p) =
_
h(p)
p
g(p) if for all q < p h(q)
q
g(q)
0 otherwise
where the subscript
p
denotes realization of operations and of order in A
p
.
REMARKS 3.4.2.
(1) It is easy to verify that

pP
A
p
is closed under the above dened operations.
The proof is left to the reader.
(2) The condition: if h(p) ,= 1 and q > p, then h(q) = 0 is equivalent to the condi-
tion: if h(p) ,= 0 and q < p, then h(q) = 1.
Chapter V: H ajeks BL Logic and BL-Algebras 381
(3) If for all p P, A
p
is a subalgebra of B
p
, then

pP
A
p
is a subalgebra of

pP
B
p
.
In the sequel, we will often omit subscripts when there is no danger of confusion.
EXAMPLES 3.4.3.
1. If P is an antichain, that is, if p q iff p = q, then

pP
A
p
is just the direct
product

pP
A
p
.
2. Suppose that P is a nite and total order. Let for p P, A
t
p
be an isomorphic
copy of A
p
such that for p ,= q P, A
t
p
A
t
q
= 1. Then,

pP
A
t
p
is
isomorphic to

pP
A
p
via the isomorphism dened as follows:
(i) if a = 0, then (a) is the constantly zero function on P;
(ii) if a ,= 0, then let p P be such that a A
t
p
. Then, (a) is the function on
P dened as follows:
(iia) if q = p, then (a)(q) is the isomorphic copy of a in A
p
;
(iib) if q < p, then (a)(q) = 1;
(iic) if p < q, then (a)(q) = 0.
3. If P is a totally ordered poset with minimum, then the above dened map is
still an embedding of

pP
A
t
p
into

pP
A
p
, but in general it is not surjective:
suppose e.g. that every A
p
is the two element residuated lattice and that P is the
rational interval [0, 1] Q. Then,

pP
A
t
p
is countable, whereas

pP
A
p
is
uncountable: for every real [0, 1], the map h

dened for every q [0, 1] Q,

by h

(q) = 1 if q and h

(q) = 0 otherwise, is in

pP
A
p
, and if ,= ,
then h

,= h

.
4. If P is a poset and for all p P, A
p
is the two element MV-algebra, then

pP
A
p
is an idempotent commutative, integral and bounded residuated lattice,
and hence it is a Heyting algebra. In [69] it is shown that every Heyting algebra is
a subalgebra of a poset product of the form shown above.
5. Di Nolas and Lettieris representation theorem for nite BL-algebras [39] can be
expressed in the following form: every nite BL-algebra is the poset product of
a nite family of nite MV-chains, with respect to a poset which is a forest (i.e.,
the downset of every element is a chain). This result has been extended in [68] to
GBL-algebras (a generalization of BL-algebras).
PROPOSITION 3.4.4 ([68]). (a) The poset product

pP
A
p
of a collection of com-
mutative, integral, and bounded residuated lattices is a commutative, integral, and
bounded residuated lattice.
(b) If in addition all A
p
are divisible, then

pP
A
p
is in turn divisible.
382 Manuela Busaniche and Franco Montagna
Proof. That

pP
A
p
is a commutative, integral and bounded lattice ordered monoid
is clear. We still have to verify the residuation property. That for every q P and
h, k

pP
A
p
we have (h (h k))(q) k(q) is trivial when (h k)(q) = 0 and
follows from the residuation property in A
q
if (h k)(q) = h(q)
q
k(q).
It is left to verify that if f(q)
q
h(q) k(q) for every q P, then for all r P,
f(r) (h k)(r). Thus, suppose f(q)
q
h(q) k(q) for every q P. Let r P
be arbitrary. If for all q < r, h(q) k(q), then (h k)(r) = h(r)
r
k(r), and the
claim follows from the residuation property in A
r
.
Suppose now that there is some q < r such that h(q) > k(q). We claim that
f(r) = 0, and hence, f(r) (h k)(r), as desired. Suppose f(r) > 0. Then, for all
q < r, f(q) = 1. But then there would be a q < r such that h(q) f(q) = h(q) > k(q),
a contradiction. This concludes the proof of Claim (a).
(b) We have to prove that for all p P and h, k

pP
A
p
, h(p) (h k)(p) =
h(p) k(p). If for all q < p, h(q) k(q), then the claim follows from the divisibility of
A
p
. If there is a q < p such that h(q) > k(q), then we must have k(p) = 0, otherwise
k(q) = 1 h(q). Hence, 0 = (h k)(p) = h(p) (h k)(p) = k(p) = h(p) k(p),
as desired.
3.5 BL-algebras as subdirect poset products
In general, prelinearity is not preserved under ordinal sums and a fortiori it is not
preserved under poset products. We know that the ordinal sum of a family of BL-chains
is a BL-chain. However, prelinearity may fail to be preserved under poset products even
if the factors are BL-chains, as is shown by the next example.
EXAMPLE 3.5.1. Consider the three element poset P = a, b, c with b < a, c < a,
b , c and c , b. Let A
a
= A
b
= A
c
be the two element residuated lattice. Then

pP
A
p
is not a BL-algebra, even though the factors are BL-chains. Let h(a) =
h(b) = 0 h(c) = 1, and let k(a) = k(c) = 0, k(b) = 1. Then h, k

pP
A
p
, but
(h k)(a) = (k h)(a) = 0, as there are x, y < a such that h(x) > k(x) and
k(y) > h(y). Hence, the prelinearity equation (x y) (y x) = 1 fails and the
algebra is not a BL-algebra.
DEFINITION 3.5.2. A subalgebra Aof a poset product

pP
A
p
is said to be a sub-
direct poset product of the family A
p
[ p P with respect to the poset P if the
projections
p
, dened, for every p P and h

pP
A
p
, by
p
(h) = h(p), are
surjective.
DEFINITION 3.5.3. A poset P = P, is said to be a forest if for all p P, the set
q P [ q p is totally ordered.
The following results are slight modications of the ones in [69]
THEOREM 3.5.4. (1) Suppose that P is a forest and that for all p P, A
p
is a BL-
chain. Then

pP
A
p
is a BL-algebra.
(2) Every BL-algebra is a subdirect poset product of a family of MV-chains and
product chains.
Chapter V: H ajeks BL Logic and BL-Algebras 383
Proof. (1) We know that

pP
A
p
is a divisible, commutative, integral and bounded
residuated lattice (Proposition 3.4.4). Hence, we only need to prove that for all p P
and h, k

pP
A
p
, either (h k)(p) = 1 or (k h)(p) = 1. If for all q < p,
h(q) = k(q), then (h k)(p) = h(p)
p
k(p) and (k h)(p) = k(p)
p
h(p),
and the claim follows from the fact that A
p
is totally ordered. Suppose e.g. that there is
q < p such that h(q) > k(q). Then, we claim that: (a) k(p) h(p) and (b) for all r < p,
k(r) h(r). These conditions together imply that (k h)(p) = k(p)
p
h(p) = 1,
and the claim follows.
Hence, we only need to verify (a) and (b). As regards to (a), if h(p) < k(p), then for
all q < p, 1 = k(q) h(q), against our assumption. As regards to (b), if there is r < p
such that h(r) < k(r), then, since P is a forest, we have either q < r or r < q. But
if q < r, then k(r) > 0 implies k(q) = 1 h(q), a contradiction. A similar argument
shows that we can exclude q < r, and we have obtained a contradiction.
(2) Let us decompose Aas a subdirect product of totally ordered BL-algebras A
i
[
i I. Next let us decompose each A
i
as an ordinal sum A
i
=

jJ
i
W
i,j
, where
each W
i,j
is a totally ordered MV-algebra or a totally ordered cancellative hoop. Let
W

i,j
= W
i,j
if W
i,j
is a totally ordered MV-algebra, and W

i,j
= L
2
W
i,j
otherwise. Note that W

i,j
is either an MV-chain or a product chain. Now let P =
(i, j) [ i I, j J
i
. Dene a partial order _ on P by (i, j) _ (i
t
, j
t
) iff i = i
t
and
j j
t
. Clearly P = P, _ is a forest.
We associate to each a Athe function h
a
on P, dened by
h
a
(i, j) =
_
_
_
1 if a
i
W
i,h
for some h > j
a
i
if a
i
W
i,j
0 if a
i
W
i,h
1 for some h < j
It is readily seen that the map : a h
a
is an embedding of A into

(j,i)P
W

j,i
,
and that the projections
i,j
are surjective.
As a corollary, we obtain:
THEOREM 3.5.5. A commutative, integral and bounded residuated lattice is a BL-
algebra iff it is a subdirect poset product of a family of MV-chains and product chains
with respect to a poset which is a forest.
Several classes of BL-algebras, arising from many-valued logic, have an easy char-
acterization in terms of poset products. We collect all of themin the next theorem, whose
easy proof is left to the reader.
THEOREM 3.5.6. A BL-algebra is:
(a) An MV-algebra iff it is a subdirect poset product

pP
A
p
such that: (a1) each
A
p
is a totally ordered MV-algebra, (a2) P = P, is an antichain, that is,
p q iff p = q.
(b) A product algebra iff it is a subdirect poset product

pP
A
p
such that each A
p
is a totally ordered product algebra and P = P, is an antichain.
384 Manuela Busaniche and Franco Montagna
(c) A G odel algebra iff it is a subdirect poset product

pP
A
p
such that each A
p
is a two element MV-algebra and P = P, is a forest.
(d) An n-potent BL-algebra (i.e., a BL-algebra satisfying the equation x
n+1
= x
n
),
iff it is a subdirect poset product

pP
A
p
where P is a forest and every A
p
is
a nite MV-chain with n + 1 elements at most.
(e) An SBL-algebra iff it is a subdirect poset product

pP
A
p
such that each A
p
is
either an MV-algebra or a product algebra, P = P, is a forest, and if p P
is minimal, then A
p
is a product algebra (possibly with two elements).
(f) A Boolean algebra if it is a subdirect poset product

pP
A
p
, where P is an
antichain and every A
p
is the two-element MV-algebra L
2
.
REMARKS 3.5.7.
(1) Recalling that every cancellative o-hoop is a subhoop of (the reduct of) a (per-
fect) MV-chain, we can easily obtain that every BL-algebra embeds into a poset
product of MV-chains. Indeed, consider the proof of Theorem 3.5.4, (b). Replace
every cancellative o-hoop by an MV-chain in which it embeds, and get the desired
result.
(2) Although every innite product chain embeds into the poset product of two MV-
chains, the embedding cannot be subdirect, that is, the projection functions cannot
be surjective. The easy proof is left to the reader.
(3) By (1), if an equation holds in all poset products of MV-algebras, then it holds in
all BL-algebras. The converse is not true, consider e.g. the prelinearity equation.
3.6 BL-algebras as MV-product algebras with a conucleus
DEFINITION 3.6.1. A MV-product algebra (MVP-algebra for short) is a BL-algebra
which is isomorphic to a subdirect product of MV-algebras and of product algebras.
MV-product algebras constitute the variety of BL-algebras generated by all MV-
algebras plus all product algebras. This variety is axiomatized over the equations of
BL-algebras by the equation x (((y x) x) (x y)) = 1.
DEFINITION 3.6.2. A conucleus on a commutative, integral and bounded residuated
lattice Ais an operator on Asuch that:
(1) If x y, then (x) (y).
(2) (x) x.
(3) ((x)) = (x).
(4) (1) = 1.
(5) ((x) (y)) = (x) (y).
Chapter V: H ajeks BL Logic and BL-Algebras 385
Conditions (1), (2), (3) and (4) say that is an interior operator, and condition (5)
says that (A), the image of Aunder , is a submonoid of A.
Note that (A) is a subquantale of A, that is, (A) is a subposet and a submonoid
of A such that for all a A, sup m (A) [ m a exists and is in (A). This
immediately implies that (A) is closed under (nite or innite) joins existing in A.
Thus, (A) is closed under nite meets, but meets in A and in (A) may differ. In
general, the meet x

y of x and y in (A) is given by (x y). Moreover, (A) is

also residuated: given x, y (A), the greatest element z (A) such that z x y
is (x y). Thus, given a conucleus on a residuated lattice A, one obtains a sub-
quantale (A) which is also a residuated lattice with respect to the monoid operation
and the join of Aand with respect to the meet

and the residual

dened by
x

y = (x y) x

y = (x y).
Conversely, a subquantale B of A induces a conucleus on A dened by (x) =
supz B [ z x.
Conuclei are very important in algebraic logic: as a classical example, Heyting
algebras are categorically equivalent to Boolean algebras with a conucleus.
Now consider a BL-algebra B which is a subdirect poset product of MV-algebras
and of product algebras, B

pP
A
p
, where P is a forest. Let A =

pP
A
p
.
Clearly, Ais an MVP-algebra. On A, dene
(a)
p
=
_
a
p
if for all q < p, a
q
= 1
0 otherwise
Then, it is almost immediate to verify that is a conucleus on A. For instance, let
us verify condition (5) in the denition of conucleus: let x, y (A). Note that there
is q < p such that (x y)
q
< 1 iff either there is q < p such that x
q
< 1 or there is
q < p such that y
q
< 1. If one of these conditions holds, then (x y)
p
= (x y)
p
= 0.
Otherwise, (x y)
p
= (x y)
p
= x
p

p
y
p
.
Moreover the denition of poset product immediately gives (A) =

pP
A
p
.
LEMMA 3.6.3. The above dened conucleus is a monoid homomorphism and a meet
homomorphism.
Proof. Let x, y A. There is q < p such that (xy)
q
< 1 iff either there is q < p such
that x
q
< 1 or there is q < p such that y
q
< 1. If one of these conditions holds, then
(x y)
p
= (x) (y)
p
= 0. Otherwise, (x y)
p
= (x)
p

p
(y)
p
= x
p

p
y
p
.
Hence, preserves meet. The same argument shows that preserves .
The operator also has an additional property corresponding to prelinearity, namely,
((x) (y)) ((y) (x)) = 1. (lin

)
Indeed, the prelinearity equation for (A) =

pP
A
p
with replaced by

gives exactly the equation (lin

). Therefore:
386 Manuela Busaniche and Franco Montagna
THEOREM 3.6.4. Every BL-algebra embeds into an algebra of the form (A) where
A is an MVP-algebra and is a conucleus on A which is also a meet and a monoid
homomorphism and in addition satises (lin

).
The converse also holds:
THEOREM 3.6.5. Let be a conucleus on an MVP-algebra A which is a meet and a
monoid homomorphism and which in addition satises condition (lin

). Then, (A) is
a BL-algebra.
Proof. Property (lin

) says that the prelinearity condition (x

y) (y

x) holds
in (A), and hence it implies prelinearity. Since 0 (0) 0, we have that 0 (A),
and hence Ais bounded. It remains to verify divisibility, i.e. the equation x (x y)
= (xy) for x, y (A). Since x = (x) and is a monoid homomorphism, we get
x (x y) = (x) (x y) = (x (x y)) = (x y).
4 Classes generating the variety of BL-algebras
We are ready to investigate some signicant classes of BL-algebras which generate
the whole variety of BL-algebras. Once this aim is achieved, we present completeness
theorems.
4.1 General facts about ordinal sums
Following standard notation, if K is a class of algebras of the same signature, then
S(K), H(K), I(K) and P
U
(K) denote the classes of subalgebras, homomorphic images,
isomorphic images and ultraproducts of K respectively. V(K) is the variety generated
by K. If O is an operator on classes of algebras such that for any class K of algebras
of the same signature one has O(K) V(K), and K
1
, . . . K
n
are classes of hoops,
then

n
i=0
O(K
i
) will denote the class of all hoops of the form

n
i=0
A
i
, where for
i = 0, . . . , n, A
i
O(K
i
). In particular, if for i = 1, . . . , n, K
i
consists of a single
hoop, B
i
, then

n
i=0
O(B
i
) will denote the class

n
i=0
A
i
[ A
i
O(B
i
).
PROPOSITION 4.1.1. Let (A
i
)
iI
be a family of hoops. Then
S(

iI
A
i
) =

iI
S(A
i
),
i.e., the subhoops of

iI
A
i
are exactly the hoops of the form

iI
B
i
, where B
i
is a (possibly trivial) subhoop of A
i
. Moreover if

iI
A
i
is a bounded hoop (and
so I has a minimum and A
0
is bounded), then the basic subalgebras of

iI
A
i
are
exactly the subalgebras of the form

iI
B
i
, where B
0
is a bounded hoop which is a
subalgebra of A
0
and B
i
is a subhoop of A
i
.
It is easy to see that the ordinal sum of nitely many hoops is associative; hence we
feel free to write A
0
A
n
without parentheses. In this case, something more can
be said.
Chapter V: H ajeks BL Logic and BL-Algebras 387
PROPOSITION 4.1.2. Let A
i
[ i I be an ordered family of totally ordered hoops
and let A =

iI
A
i
. Then, B is a homomorphic image of A iff either there is
an initial segment J of I such that B =

jJ
A
j
or there is an i I such that
B

j<i
A
j
H(A
i
). A similar result holds if I has a minimum i
0
and A
i
0
is
bounded (hence Ais a BL-algebra).
Proof. Let F be the unique lter on A such that B = A/F. Then, F has either the
form

jI\J
A
j
for some initial segment J, and then B =

jJ
A
j
, or there is a
unique i I such that F intersects A
i
1 but A
i
, F, and then B =

j<i
A
j

(A
i
/(F A
i
)).
PROPOSITION 4.1.3. Let A
0
j
, . . . , A
n
j
, j J be hoops. Then P
U
(A
0
j
A
n
j
[
j J) P
U
(A
0
j
[ j J) P
U
(A
n
j
[ j J).
Proof. Let A =

jJ
(A
0
j
A
n
j
) and let U be any ultralter on J. Clearly A/U
P
U
(A
0
j
A
n
j
[ j J) and any member of P
U
(A
0
j
A
n
j
[ j J)
is of the form A/U for some ultralter U on J. For b A, let b/U be the class of
b modulo U. For m = 0, . . . , n 1, let X(b, m) = j J [ b
j
A
m
j
1 and
let X(b, n) = j J [ b
j
A
n
j
. Since the X(b, m) form a partition of J there is
exactly one m
b
such that X(b, m
b
) U. Of course m
b
really depends on b/U, in that if
b/U = c/U, then m
b
= m
c
.
Now for m = 0, . . . , n, let Y
m
= b/U [ b A and m
b
= m; it is easy to check
that Y
m
1 is the universe of a subhoop B
m
of A/U. If we take b/U Y
m
and
c/U Y
h
with m < h, then X(b, m) X(c, h) U and for j X(b, m) X(c, h)
one has b
m
j
c
m
j
= c
m
j
and c
m
j
b
m
j
= b
m
j
. It follows that b/U c/U = b/U and
c/U b/U = b/U and hence A/U = B
0
B
n
.
It remains to show that B
m
is isomorphic with (

jJ
A
m
j
)/U. If b/U Y
m
1,
then j J [ b
j
A
m
j
U. If c is dened by
c
j
=
_
b
j
if b
j
A
m
j
1 otherwise
then c

jJ
A
m
j
and c/U = b/U. Let now c
m
/U be the equivalence class of c
modulo U in

jJ
A
m
j
/U; one can easily verify that the map c
m
/U b/U = c/U
is an isomorphism from

jJ
A
m
j
/U to B
m
, as desired.
COROLLARY 4.1.4. If A
0
, . . . , A
n
are totally ordered hoops, then
P
U
(A
0
A
n
) B
0
B
n
[ B
i
P
U
(A
i
).
LEMMA4.1.5. If A
1
, . . . , A
n
are totally ordered Wajsberg hoops and A
0
is a Wajsberg
algebra, then
ISP
u
(
n

i=0
A
i
) = I(
n

i=0
SP
u
(A
i
)).
388 Manuela Busaniche and Franco Montagna
Proof. The left-to-right inclusion is a straight consequence of Proposition 4.1.1 and
Corollary 4.1.4; for the other we induct on n. If n = 0 there is nothing to prove. Let
then n > 0 and let B
i
ISP
u
(A
i
) for i = 0, . . . , n. If C = A
0
A
n1
, then by
inductive hypothesis B
1
B
n1
ISP
u
(C). Let C
I
/U be an ultrapower of C
in which B
0
B
n1
embeds. Then it is readily seen that B
0
B
n1
B
n
embeds into (C B
n
)
I
/U. On the other hand, if A
J
n
/V is an ultrapower of A
n
in
which B
n
embeds, by the same token C B
n
embeds into (C A
n
)
J
/V . Thus
B
0
B
n
ISP
u
(SP
u
(A
0
A
n
)) ISP
u
(
n

i=0
A
i
).
4.2 Generation by nite members
It is well-known that any algebra is isomorphic to an ultraproduct of its nitely
generated subalgebras. Moreover, by Birkhoff subdirect representation theorem, every
algebra is isomorphic to a subdirect product of subdirectly irreducible algebras. Since a
subdirectly irreducible BL-algebra is linearly ordered, it follows that every BL-algebra
is isomorphic to a subdirect product of an ultraproduct of nitely generated BL-chains.
In other words, the variety of BL-algebras is generated as a quasivariety by the class of
all nitely generated BL-chains. We are going to strengthen the above result, showing
that the variety of BL-algebras is generated as a quasivariety by its nite linearly ordered
members. First of all, we can prove:
LEMMA 4.2.1. Every nitely generated BL-chain is the ordinal sum of nitely many
Wajsberg hoops, and every Wajsberg component contains at least one generator.
Proof. By induction on the termt(x
1
, ..., x
n
), it is easy to prove that given a BL-chain A
with rst component W
0
and given elements a
1
, . . . , a
n
Awhich belong to Wajsberg
components W
1
,...,W
k
, each in the order given, we have that t(a
1
, ..., a
n
) W
0

W
1
... W
k
.
DEFINITION 4.2.2. An algebra A in a class K has the Finite Embeddability Property
(FEP) with respect to K if for any nite partial subalgebra A
t
of Athere exists a nite
algebra B K such that A
t
is embeddable in B. A class K has the FEP if each of its
members has the FEP with respect to K.
In [12] it is shown that, if a variety V has the FEP then V is generated as a quasiva-
riety by its nite members, and that if the class of subdirectly irreducible members of V
has the FEP, so does V. Finally, the class of linearly ordered Wajsberg hoops (algebras)
has the FEP (see the proof of [12, Theorem 3.9]). This fact is crucial to the following:
THEOREM 4.2.3. The class of BL-chains has the FEP. Hence the variety of BL-
algebras is generated as a quasivariety by its nite linearly ordered members.
Proof. Let C be a linearly ordered BL-algebra, let C
t
be a nite partial subalgebra of
C, and let Dbe the subalgebra generated by C
t
. Then by Lemma 4.2.1, Dis the ordinal
sumof nitely many Wajsberg hoops, each containing at least one generator. We proceed
by induction on the cardinality n of C
t
. If n = 1, then trivially D consists of a single
Chapter V: H ajeks BL Logic and BL-Algebras 389
Wajsberg component, and the claim follows from [12, Theorem 3.9]. If n > 1, then
take the last Wajsberg component W of D. Then D can be written as D = E W,
where E is a (possibly trivial) BL-chain. Moreover by [12] again, W has the FEP,
and C
t
W embeds into a nite linearly ordered Wajsberg hoop S. Finally, again by
Lemma 4.2.1, W contains at least one generator, therefore the cardinality of E C
t
is
less than n, and by the induction hypothesis E C
t
embeds into a nite BL-chain B.
It follows that C
t
embeds into the nite BL-chain B S, and the claim is proved.
REMARK 4.2.4. Although the variety of BL-algebras is generated as a quasivariety
by its nite members, no nite BL-algebra generates the whole variety of BL-algebras.
Indeed, the equation

x
i
x
j
[ i, j = 1, ..., n + 1, i ,= j = 1
holds in every BL-algebra with n elements, but it does not hold in all BL-algebras.
A well-known result on logical systems due to Harrop [61], applied to quasivarieties
(see [12, Corollary 2.6]) and combined with FEP yields the following:
COROLLARY 4.2.5. The quasi-equational theory of BL-algebras is decidable. (In
fact, since the language of BL-algebras has nite type, the universal theory of BL-
algebras is decidable). Thus the consequence relation in BL is decidable.
4.3 Generation by t-norm BL-algebras
As said before, in [27] it is shown that the class of BL-algebras is generated as a
variety by the class of all continuous t-norm BL-algebras, and hence BL is complete
with respect to such class of BL-algebras. Here below, we will prove a slightly stronger
statement.
THEOREM 4.3.1. The variety of BL-algebras is generated as a quasivariety by the
class of all t-norm algebras which are ordinal sums of nitely many copies of the uka-
siewicz t-norm algebra.
Proof. The variety of BL-algebras is generated as a quasivariety by the class of all nite
BL-chains. Clearly a nite BL-chain has nitely many Wajsberg components and all of
them are (reducts of) nite Wajsberg algebras. Now any nite Wajsberg algebra embeds
into the ukasiewicz t-norm algebra [0, 1]

, therefore any nite BL-chain B embeds

into the ordinal sum of nitely many copies of [0, 1]

. Let n be the number of such

copies, and let [0, 1]

i
be an isomorphic copy of [0, 1]

in [
i
n
,
i+1
n
]. Then B embeds
into

n1
i=0
[0, 1]

i
, which is easily shown to be a t-norm BL-algebra.
REMARKS 4.3.2.
(1) Any nite ordinal sum of ukasiewicz t-norms is a continuous t-norm. Indeed, the
corresponding algebra is the ordinal sum of nitely many bounded hoops, hence
it is a bounded hoop [12]. Thus it satises the condition x (x y) = x y,
which (in [0, 1]) is known to be equivalent to the continuity of .
390 Manuela Busaniche and Franco Montagna
(2) By Theorem 4.3.1, the (nite) consequence relation in BL is complete with re-
spect to continuous t-norms and their residuals, i.e. for every nite set of BL-
formulas and for every BL-formula , one has
BL
iff for every BL-algebra
B = [0, 1], , , , , 0, 1, where is a continuous t-norm and is its resid-
ual, and for every evaluation e in B such that e() = 1 for all , one has
e() = 1.
(3) No ordinal sum of product t-norms generates the whole variety of BL-algebras.
To verify this, consider the equation
(x x
2
) ((x x
2
) x) = 1.
This equation holds in any ordinal sum of product chains. Indeed, if x is an
idempotent, then x x
2
= 1, otherwise, x belongs to a cancellative component,
and hence (x x
2
) = x and (x x
2
) x = 1. However, the equation does
not hold in the standard MV-algebra on [0, 1]. To see this, take x =
1
3
. Then,
x x
2
=
2
3
< 1, and (x x
2
) x =
2
3

1
3
=
2
3
< 1.
4.4 Generation by single BL-chains
An easy consequence of Theorem 4.3.1 is that the variety of BL-algebras is gener-
ated by any BL-chain that is the ordinal sumof innitely many copies of the ukasiewicz
t-norm algebras. One may wonder if there are other signicant examples of BL-chains
generating the whole variety of BL-algebras. A characterization of such BL-chains,
which will be called generic in the sequel, is given in [2]. This characterization result is
sketched in the present section.
Our strategy is as follows: rst we nd some properties which are not shared by
all BL-chains and which can be expressed by means of equations. Then we show that
every BL-chain which does not satisfy any of the above properties is generic. In order
to express these properties we start from a denition.
DEFINITION 4.4.1. Given a BL-chain A, a Wajsberg component W of Aand a natu-
ral number n > 1, we say that
1
n
is in W iff W is the reduct of a Wajsberg algebra and
there is an a W such that
W
a = a
W
. . .
W
a
. .
n1 times
, where
W
and
W
denote the
negation (x
W
0
W
) and the ukasiewicz sum (
W
x
W
y) in W, thought of as a
Wajsberg algebra.
Note that
1
n
is in W iff the nite Wajsberg chain L
n+1
with n+1 elements embeds
into W. The properties we are trying to express are:
(i) There is a number n > 1 such that
1
n
is not in the rst Wajsberg component.
(ii) There are natural numbers n > 1 and m such that
1
n
is in at most m Wajsberg
components.
Our plan is to show that all properties (i) and (ii) can be expressed by means of identities.
We start from some notation: Let A =

iI
A
i
be a BL-chain, where I has a
minimum i
0
, every A
i
is a Wajsberg o-hoop and A
i
0
is bounded. For all a, b A we
Chapter V: H ajeks BL Logic and BL-Algebras 391
shall write a b if either a ,= 1 and b = 1 or a < b and a, b are not in the same
Wajsberg component. We shall write a b if a , b and b , a, i.e. either a = b = 1 or
both a, b are not 1 and belong to the same Wajsberg component.
For any BL-algebra A and a, b A we dene a b = (a a b) b; moreover na
is dened inductively by the following clauses
0a = 0;
(n +1)a = na a.
It is easy to see that if Ais a Wajsberg algebra, then the operator so dened coincides
with the ukasiewicz sum on A, and that na = a a
. .
n times
. The following lemma can
be proved by straightforward computation, using the denition of ordinal sum and some
elementary properties of Wajsberg hoops.
LEMMA 4.4.2. Let A, I, A
i
, i
0
, etc. be as dened above. Then for all a, b A the
following conditions hold:
1. (a b) b = (b a) a if and only if a, b A
i
for some i I;
2. (a b) b = 1 if and only if either b = 1 or b a;
3. if b , a, then (a b) b = a b;
4. if A
i
is a bounded Wajsberg hoop and a, b A
i
, then a b is the ukasiewicz
sum of a and b in A
i
;
5. if a b, then a b = b;
6. if b a or a, b A
i
and A
i
is a cancellative hoop, then a b = 1;
7. a A
i
0
if and only if a = a if and only if either a = 1 or a ,= 0.
Now we prove that property (i) above can be expressed by means of an identity.
LEMMA 4.4.3 ([2]). Consider for k > 1 the identity
((y x) x) ((k 1)x x) x y (
k
)
(where x y is an abbreviation for (x y) (y x)). Let Abe a BL-chain and let

iI
A
i
be its decomposition into Wajsberg o-hoops (so that I has a minimum i
0
and
A
i
0
is a Wajsberg algebra). If [I[ > 1 then A [= (
k
) if and only if
1
k
is not in A
i
0
.
Proof. Assume rst that
1
k
is not in A
i
0
. Let a, b A and set c = (k 1)a a. If
either a = 1 or b = 1, then (
k
) holds trivially for x = a and y = b, since a b = 1.
If b a or b a, again (
k
) holds, since (b a) a = a b. Assume then
that a b 1; then (b a) a = 1 and a b = b < 1, thus we have to show
that c b. If a / A
i
0
, then a = a 0 = 0 by denition of ordinal sum, and
c = (k 1)a a = (k 1)a a = 0 b. If a A
i
0
, then since
1
k
is not in
A
i
0
, c A
i
0
1. Moreover, since a b, b / A
i
0
, and c < b.
392 Manuela Busaniche and Franco Montagna
Conversely suppose that
1
k
is in A
i
0
. Let a A
i
0
be such that (k 1)a = a and
let b , A
i
0
(such a b exists because [I[ > 1); then (b a) a = (k 1)a a = 1,
but a b = b ,= 1. Thus (
k
) does not hold in A.
We now come to property (ii). For any m, k N with m 2 and k 1, consider
the inequality
k1

j=0
((x
j+1
x
j
) x
j
) ((m1)x
j
(x
j
x
3
j
))
k

j=0
x
j
. (
m,k1
)
LEMMA 4.4.4. Let A be a BL-chain, let

iI
A
i
be its decomposition as an ordinal
sum of Wajsberg o-hoops, and let m 2 be given. The following are equivalent.
1. There is a k > 0 such that (
m,k1
) holds in A.
2. The set i I [
1
m
is in A
i
has cardinality at most k.
A proof of Lemma 4.4.4 can be found in [2]. Since it is rather long we do not
reproduce it here. However, just to give the reader an idea of the techniques involved,
we give a proof of an easier statement saying that the fact that a BL-chain has only
nitely many Wajsberg components can be expressed by means of an identity.
LEMMA 4.4.5 ([2]). Let A =

iI
A
i
be a BL-chain , where every A
i
is a linearly
ordered Wajsberg hoop, and consider for any n > 0 the equation
n1

i=0
((x
i+1
x
i
) x
i
)
n

i=0
x
i
. (
n
)
Then A [= (
n
) if an only if [I[ n.
Proof. Suppose rst that [I[ n. Then for any a
0
, . . . , a
n
A, either a
i
= 1 for some
i n or else there is a k < n such that a
k
, a
k+1
. In the rst case the right hand side of
(
n
) is 1, hence the equation holds. In the second case (a
k+1
a
k
) a
k
= a
k
a
k+1
and thus
n1

i=0
((a
i+1
a
i
) a
i
) a
k
a
k+1

n

i=0
a
i
and again (
n
) holds in A.
Conversely suppose that [I[ > n. Then we can pick a
0
, . . . , a
n
A such that
a
0
a
1
a
n
1.
Hence
_
n1
i=0
((a
i+1
a
i
) a
i
) = 1 and
_
n
i=0
a
i
= a
n
,= 1. Thus (
n
) does not hold
in A.
Combining Lemmas 4.4.3, 4.4.4 and Theorem 4.2.3 we obtain:
THEOREM 4.4.6. Let A =

iI
A
i
be a BL-chain, where I has a minimum i
0
, each
A
i
is a Wajsberg o-hoop, and A
i
0
is bounded. Then Ais generic if and only if:
Chapter V: H ajeks BL Logic and BL-Algebras 393
1. for every n > 1,
1
n
is in A
i
0
,
2. for every n the set i I [
1
n
is in A
i
is innite.
Proof. The conditions are necessary because of Lemma 4.4.3 and Lemma 4.4.4. For the
sufciency, by Theorem 4.2.3, it is certainly enough to show that any nite BL-chain
(thus of the form L
n
0
L
n
k
) embeds into A. Since
1
n
0
1
is in A
i
0
, then L
n
0
embeds into A
i
0
. If m = (n
1
1) . . . (n
k
1), then each L
n
j
j = 1, . . . , k, is a
subalgebra of L
m+1
. By assumption there are innitely many i I such that L
m+1
embeds into A
i
, thus it is possible to nd 0 < i
1
< < i
k
I such that L
n
j
embeds
into A
i
j
for j = 1, . . . , k. Hence L
n
0
L
n
k
embeds into Aand Ais generic.
REMARK 4.4.7. Since the variety of BL-algebras is generated as a quasivariety by
the class of nite BL-chains, and since every nite BL-chain embeds into any BL-chain
satisfying the conditions of Theorem 4.4.6, it follows that any such BL-chain generates
the whole variety as a quasivariety. In logical terms, BL is complete with respect to any
of such chains also as regards to the nite consequence relation.
EXAMPLES 4.4.8.
(i) Let [0, 1]
Q
denote the Wajsberg subalgebra of [0, 1]

consisting of all rational ele-

ments of [0, 1]. Then [0, 1]
Q

i=2
L
i
clearly satises Theorem 4.4.6, therefore
it is generic. Note that none of its Wajsberg components (except the rst one)
generates the whole variety of Wajsberg algebras. To the contrary, the algebra
B = [0, 1]

nN
L
2
n
+1
is not generic, because L
4
does not embed in any
component of B except the rst one.
(ii) Let p be a prime and let U
p
be the subalgebra of [0, 1]
Q
whose base set is m/n [
n > 1, 0 m n, p does not divide n. Then U
p
generates the whole
variety of Wajsberg algebras, but the ordinal sum of innitely many copies of U
p
does not satisfy Theorem 4.4.6 (because no Wajsberg component contains
1
p
), and
hence it is not generic.
(iii) Although BL is a common fragment of ukasiewicz logic, product logic and
G odel logic, it is not just the intersection of these three logics. Indeed, the equa-
tion (x y) y (x y (y y
2
)) holds in all MV-algebras, product
algebras and G odel algebras, but not in any BL-algebra (consider e.g. the ordinal
sum of two copies of [0, 1]

).
4.5 Generating the variety of SBL-algebras
As we dened in Section 2.5, an SBL-algebra is a BL-algebra satisfying the identity
xx = 0. The logical counterpart of this identity is a kind of weak contraction, namely
( &). This principle, being a weak form of a structural rule, has a logical
interest, which gives a motivation for the study of SBL-algebras. From an algebraic
point of view, SBL-algebras are characterized as follows:
LEMMA 4.5.1. (i) A BL-chain Ais an SBL-algebra iff Ahas the formA = L
2
B
for some (possibly trivial) BL-chain B.
394 Manuela Busaniche and Franco Montagna
(ii) Hence SBL-algebras are precisely BL-algebras which are isomorphic to subdi-
rect products of BL-chains as shown in (i).
Proof. We only prove (i), as (ii) immediately follows from (i) and from the fact that
every BL-algebra is isomorphic to a subdirect product of BL-chains. Suppose that A =
L
2
B. Let a A. If a is in the rst Wajsberg component of A, then a 0, 1, and
a a = 0. If a is not in the rst component, then a = 0 and a a = 0.
Conversely suppose that the rst Wajsberg component of A has an element b / 0, 1,
thus b ,= 0 and b ,= 0. Since Ais linearly ordered, b b = minb, b , = 0.
We investigate the problems of generating the variety of SBL-algebras by nite
SBL-algebras, by t-norm SBL-algebras and by single SBL-chains.
As regards the generation by nite SBL-algebras, the argument used in the proof of
Theorem 4.2.3 yields:
THEOREM 4.5.2. The class of linearly ordered SBL-algebras has the FEP. Hence
the variety of SBL-algebras is generated as a quasivariety by its nite linearly ordered
members. These are precisely the ordinal sums of the form L
2
L
k
1
. . . L
k
n
.
As regards to the generation by t-norm SBL-algebras we have:
THEOREM 4.5.3. The variety of SBL-algebras is generated as a quasivariety by both
the class of t-norm algebras of the form [0, 1]

[0, 1]

. . . [0, 1]

and the class of

t-norm algebras of the form [0, 1]
G
[0, 1]

. . . [0, 1]

.
Proof. Any algebra of the form shown above is a SBL-algebra. Moreover, any algebra
of the form L
2
L
k
1
. . . L
k
n
embeds into the ordinal sum of the form [0, 1]

[0, 1]

. . . [0, 1]

, or of the form [0, 1]

G
[0, 1]

. . . [0, 1]

. Thus the claim

follows from Theorem 4.5.2.
We now turn to the problem of generating the variety of SBL-algebras by single
SBL-chains.
THEOREM 4.5.4. Let A =

iI
U
i
be a SBL-chain, where I is an ordered set with
minimum i
0
, the U
i
are linearly ordered Wajsberg hoops, and U
i
0
= L
2
. The following
are equivalent:
(i) Agenerates the whole variety of SBL-algebras.
(ii) For every n > 1, the set i I [
1
n
is in W
i
is innite.
Proof. If (ii) does not hold, then the identity (
n,k1
) holds in Afor k sufciently large
(this is shown as in Lemma 4.4.4, see [2]). On the other hand (
n,k1
) is not a valid
identity of SBL-algebras, because it fails e.g. in L
2
[0, 1]

, where [0, 1]

is the
ordinal sum of copies of [0, 1]

. Thus Adoes not generate the whole variety of SBL-

algebras. On the other hand, if (i) holds, then every nite SBL-algebra of the form
L
2
L
k
1
. . . L
k
n
embeds into A (the proof is as in Theorem 4.4.6). Hence by
Theorem 4.5.2, Agenerates the variety of SBL-algebras as a quasivariety.
Chapter V: H ajeks BL Logic and BL-Algebras 395
4.6 Generating classes and completeness theorems
Let us consider four signicant semantics for BL, namely the class of t-norm BL-
algebras (t-norm semantics), the classes of nite BL-algebras (nite semantics), the
class of BL-chains whose lattice reduct is the rational interval [0, 1] (rational semantics)
and the class of BL-chains over an ultrapower of [0, 1] (hyperreal semantics).
THEOREM 4.6.1 ([32]). Given an axiomatic extension L of MTL, its corresponding
variety V and a subclass K of V, we have the following:
1. L is complete with respect to K iff K generates V as a variety.
2. L is nitely strongly complete with respect to K iff K generates V as a quasivari-
ety.
3. L is strongly complete with respect to K iff every nite or countable chain in V
embeds into some algebra in K.
We have already seen that the class F"N of all nite BL-algebras, as well as the
class F"NC of nite BL-chains and the class N of all continuous t-norm BL-algebras
generate the class B' of all BL-algebras. It follows:
THEOREM 4.6.2. BL is nitely strongly complete with respect to any of F"N, F"NC
and N.
We have already seen that BL is not strongly complete with respect to N. Since no
innite BL-chain embeds into a nite BL-algebra, by Theorem 4.6.1 we have that BL
is not even strongly complete with respect to F"N or to F"NC. Hence, we may wonder
if there are signicant classes of BL-algebras such that BL is strongly complete with
respect to them. We will present two such classes:
1. The class 1A of all BL-chains on the rational interval [0, 1].
2. The class N

of all ultraproducts of continuous t-norm BL-algebras (these are

clearly BL-algebras having an ultrapower of [0, 1] as lattice reduct).
We will treat 1Arst. First of all, note that the lattice reduct of a countable and densely
ordered BL-chain is order-isomorphic to the rational interval [0, 1], and that any order
isomorphism induces on such interval the structure of a BL-chain in 1A. Hence, by
Theorem 4.6.1, in order to prove that BL is strongly complete with respect to 1A, it
sufces to prove that every nite or countable BL-chain Aembeds into a countable and
densely ordered BL-chain B. The construction of the desired chain B is obtained as
follows:
Decompose Aas an ordinal sum of Wajsberg hoops.
Recall that every ordered group G embeds into a minimum divisible ordered
group G
d
, called the divisible hull of G. Divisible means that for all g G
d
and
for every positive integer n, there is an element g
n
G
d
such that g = g
n
+...+g
n
(n times). Moreover if Gis nite or countable, then G
d
is countable.
396 Manuela Busaniche and Franco Montagna
Hence, every Wajsberg component W of Ais either a negative cone of an ordered
group G, and then it embeds into the negative cone W
d
of G
d
, or it is (the reduct
of) an MV-algebra, which is in turn an interval of an ordered group G, and then
it embeds into an interval W
d
of G
d
. In any case, every Wajsberg component W
of Aembeds into a densely ordered countable Wajsberg hoop.
Now replace in the decomposition of Aevery Wajsberg component W with W
d
.
We obtain a countable densely ordered BL-algebra in which Aembeds.
For N

it is easier to use a different argument. Let be any set of formulas and be a

formula such that -
BL
. Take a sequence of nite sets

0

1

n
. . .
such that

nN

n
= . Note that if , there is a natural number n such that for
m n,
m
. Since for all n,
n
-
BL
and
n
is nite, there are a BL-algebra
A
n
N and a valuation v
n
in A
n
such that v
n
() = 1 for all and v
n
() ,= 1.
Now let U be an ultralter on N containing all conite sets. Let
A = (

nN
A
n
)/U and v() = v
n
() [ n N/U.
Note that A N

and that v is a valuation into A. Moreover, for every , the

set n N [ v
n
() = 1 is in U, whereas the set n N [ v
n
() = 1 is in not U (it
is the empty set). Hence, by the ultraproduct theorem, v() = 1 for every and
v() < 1. It follows that ,[=
TN
. Summing-up, we have proved:
THEOREM 4.6.3. BL is strongly complete with respect to any of 1A or N

.
5 Varieties of BL-algebras
As it is well known, to each logic L which is an axiomatic extension of BL there
corresponds a subvariety of BL-algebras formed by those algebras in which all formulas
from L are true. On the other hand, to each variety V of BL-algebras one can assigned
a logic being an axiomatic extension of BL formed by all formulas which are true in
every member of V. The assignments are inverse of each other and establish a dual
lattice isomorphism between the lattice of axiomatic extensions of BL and the lattice of
all subvarieties of BL-algebras.
Unlike the lattice of varieties of MV-algebras, of G odel algebras or of product al-
gebras, the lattice of varieties of BL-algebras is very complicated. In this section we
investigate different subvarieties of BL-algebras. We prove that there are uncountably
many varieties of BL-algebras which are generated by a single chain. Already the variety
generated by the ordinal sums of two MV-algebras has uncountably many subvarieties.
Finally, we investigate the varieties of BL-algebras which are generated by a single
t-norm algebra. Surprisingly, it turns out that there are only countably many of them.
THEOREM 5.0.1. Le0t V
n
denote the variety axiomatized by
n+1
(i.e. the variety
generated by all ordinal sums of at most n+1 linearly ordered Wajsberg hoops, the rst
one bounded). Then:
Chapter V: H ajeks BL Logic and BL-Algebras 397
(1) the lattice of subvarieties of V
n
is countable if and only if n = 0;
(2) for every n 0 there are uncountably many subvarieties of V
n+1
which are not
subvarieties of V
n
.
In particular there are uncountably many subvarieties of the variety of BL-algebras.
Proof. Note that V
0
is just the variety of Wajsberg algebras, whose lattice of subvarieties
is countable and recursively presentable. For n > 0 and for any set X of prime numbers
let V
X,n
be the variety axiomatized by
n+1
and (
p,0
) [ p X. Clearly V
X,n
is a
subvariety of V
n
. Let now W
X
be the subalgebra of the MV-algebra on [0, 1] whose
universe is the set m/n [ m, nN, n>0, mn, p does not divide n, for any p X
and let for m > 0, W
X,m
= W
X
W
X
, m + 1 times. It is easily seen that
W
X,m
V
Y,n
if and only if m n and Y X. Hence if X ,= Y , then V
X,n
,= V
Y,n
which shows that, for n 0, V
n+1
has uncountably many subvarieties of the form
V
X,n+1
. Moreover none of these varieties can be a subvariety of V
n
.
THEOREM 5.0.2. For every n > 1 there are uncountably many varieties V of BL-
algebras such that:
(1) V contains all BL-chains which are ordinal sums of n Wajsberg components.
(2) Every chain in V is the ordinal sum of n + 1 Wajsberg components.
Proof. We will only sketch the proof (for a full proof see [2]). Consider for every set X
of prime numbers, the variety W
X
generated by all BL-chains which are either ordinal
sums of n Wajsberg components or ordinal sums of n + 1 Wajsberg components
U
1
, ..., U
n+1
such that for all p X, there is a j n +1 such that L
p
does not embed
into U
j
. Then, W
X
is axiomatized by
n+1
plus (
(p,n)
), and hence the chains in W
X
either have n components or have n + 1 components and for every prime p X, L
p
does not embed into some U
i
. Clearly, there is a continuum of such varieties and each
of them satises the required conditions.
5.1 Varieties generated by t-norm algebras
By Theorem 4.4.6, there is a t-norm, namely, the ordinal sum of innitely many
ukasiewicz t-norms, whose associated logic is BL.
We have seen that there is a continuum of varieties of BL-algebras, and hence there
is a continuum of schematic extensions of BL. Indeed, to every set X of prime numbers,
we can associate a variety V
X
whose chains either have just one Wajsberg component,
or are ordinal sums such that the rst component does not contain the rational
1
p
for
all p X. Clearly, these varieties are not generated by a t-norm algebra, because if the
rst Wajsberg component of a t-norm is innite, then it is isomorphic to the ukasiewicz
t-norm, and hence it has all rational numbers in [0, 1].
Hence, a natural question is: what is the cardinality of the class of all varieties of
BL-algebras generated by a continuous t-norm? And is it possible to nd an explicit
description of a generating t-norm algebra for all such varieties?
A naive guess would be that there are uncountably many varieties of t-norm alge-
bras. This guess is supported by the result proved next.
398 Manuela Busaniche and Franco Montagna
THEOREM 5.1.1. The class of all isomorphism types of continuous t-norms has the
cardinality of the continuum.
Sketch of the proof. Every continuous t-norm is uniquely determined by its restriction
to the rationals in [0, 1]. Since the cardinality of the set of functions from (Q [0, 1])
2
into [0, 1] is 2

0
, there are at most 2

0
continuous t-norms. Moreover for any set X
of natural numbers and for every natural number i, let
i
be the ukasiewicz t-norm if
i X and the product t-norm otherwise. Then, let
X
be the ordinal sum of the t-norms
described above. Then if X ,= Y ,
X
and
Y
are not isomorphic, and there are 2

0
isomorphism types of continuous t-norms.
Somewhat surprisingly, Zuzana Hanikov a [60] proved that every t-norm variety is
co-NP complete, and hence decidable (see Chapter X of this handbook). This imme-
diately implies that only countably many varieties of BL-algebras are generated by a
continuous t-norm algebra.
The goal of the present subsection is the study and the equational characterization
of all subvarieties of B' generated by single BL-chains on [0, 1] induced by continuous
t-norm. The main result of this subsection is the following: there is a countable class K
of continuous t-norms (whose members will be called canonical t-norms), such that:
1. For each continuous t-norm there is a canonical t-norm which generates the same
subvariety.
2. Any two distinct canonical t-norms generate different varieties.
3. Hence, the set of subvarieties of BL generated by t-norm BL-chains, is countable,
and a set of generating t-norms can be explicitly described.
Recall that given a continuous t-norm on [0, 1], the symbol [0, 1]

denotes the BL-chain

[0, 1], ,

, max, min, 0, 1, where

is the residual of .
A BL chain A is called regular if it is the ordinal sum of Wajsberg hoops of the
form [0, 1]

, (0, 1]

and L
2
, and Ahas a rst component (which is either [0, 1]

or L
2
).
The class of regular BL-chains will be called 1EG.
The class [0, 1]

[ a continuous t-norm is denoted by N. Members of N are also

known as t-norm BL-algebras. Clearly N 1EG, but the opposite inclusion does not
hold. For instance, the three element G odel algebra is in 1EG but not in N.
The subclass of 1EG consisting of those BL-chains which are nite ordinal sums of
Wajsberg hoops will be denoted by Fin.
DEFINITION 5.1.2. Let A 1EG. Then Fin(A) denotes the set of all BL-algebras
B =

n
i=0
W
i
1EG such that:
- For i = 0, . . . , n, W
i
is a Wajsberg hoop.
- There are components A
0
< . . . < A
n
of A such that A
0
is the rst component
of A, and for every i, if W
i
is isomorphic to [0, 1]

, then A
i
is isomorphic to
[0, 1]

, if W
i
is isomorphic to (0, 1]

then A
i
is isomorphic either to (0, 1]

or
to [0, 1]

, and if W
i
is isomorphic to L
2
then A
i
is isomorphic either to L
2
or
to [0, 1]

.
Chapter V: H ajeks BL Logic and BL-Algebras 399
In the sequel if is any continuous t-norm, we write Fin() for Fin([0, 1]

) and V()
for V([0, 1]

).
The next lemma is straightforward:
LEMMA 5.1.3. Let A 1EG. Then:
(i) Every nitely generated subalgebra of A is a subalgebra of at least one algebra
in Fin(A).
(ii) Every algebra in Fin(A) is in ISP
u
(A).
(iii) ISP
u
(A) = ISP
u
(Fin(A)).
Proof. (i) Let a
0
, . . . , a
n
be generators of a subalgebra B of A (without loss of gener-
ality, possibly adding 0 to the generators, we may assume that a
0
is in the rst compo-
nent), and let W
0
, . . . , W
k
be nitely many components of Asuch that a
0
, . . . , a
n

W
0
. . . W
k
. Without loss of generality we may assume that the components are
enumerated in increasing order, i.e., if i < j then W
i
< W
j
. Thus W
0
is the rst
component of A. Now by induction on the complexity we can see that for every term
t(x
1
, . . . , x
n
) one has:
t(a
1
, . . . , a
n
) W
0
. . . W
k
.
Since B is generated by a
0
, . . . , a
n
, B W
0
. . . W
k
, and W
0
. . . W
k

Fin(A). This proves (i).
(ii) We have seen that (Lemma 4.1.5) that if U
0
is a Wajsberg algebra and U
1
, . . . , U
k
are Wajsberg hoops, then
ISP
u
(U
0
. . . U
n
) = ISP
u
(U
0
) . . . ISP
u
(U
n
).
Now let W
0
. . . W
n
Fin(A), and let A
0
, . . . , A
n
be components of A as in
Denition 5.1.2. Recall that (0, 1]

ISP
u
([0, 1]

), because any cancellative linearly

ordered hoop is a subhoop of a linearly ordered Wajsberg algebra, and every linearly
ordered Wajsberg algebra is embeddable in an ultrapower of [0, 1]

, see [26]. Moreover,

L
2
is a subalgebra of [0, 1]

, hence it is in ISP
u
([0, 1]

). It follows that for every i,

W
i
ISP
u
(A
i
), therefore
W
0
. . . W
n
ISP
u
(A
0
) . . . ISP
u
(A
n
) = ISP
u
(A
0
. . . A
n
).
Since A
0
. . . A
n
S(A), the claim follows.
(iii) Fin(A) ISP
u
(A) by (ii), therefore ISP
u
(Fin(A)) ISP
u
(A). On the other
hand if FG(A) denotes the class of nitely generated subalgebras of A, then, by (i),
FG(A) S(Fin(A)), and A ISP
u
(FG(A)). Thus A ISP
u
(Fin(A)), and
ISP
u
(A) ISP
u
(Fin(A)).
400 Manuela Busaniche and Franco Montagna
One might wonder whether the operator Fin( ) commutes with the ordinal sum
operation. And one can easily notice that this is not true, i.e. in general Fin(A B)
is not equal to Fin(A) Fin(B), we only have Fin(A) Fin(B) Fin(A B),
because rst components of algebras belonging to Fin(B) can only be L
2
or [0, 1]

, but
not (0, 1]

. However, we can get the equality if we consider a weaker notion of Fin( )

by dropping this last condition.
In the sequel, we shall denote by 1EG
t
the class of all linearly ordered hoops which
are ordinal sums (possibly without rst component) of components isomorphic to [0, 1]

,
or to (0, 1]

or to L
2
.
DEFINITION 5.1.4. Let A 1EG
t
. Then Fin
t
(A) denotes the set of all nite ordinal
sums of Wajsberg hoops W
0
, . . . , W
n
such that the following conditions hold:
- Each W
i
is isomorphic either to L
2
, or to (0, 1]

or to [0, 1]

.
- There are components A
0
< . . . < A
n
of A such that for every i, if W
i
is iso-
morphic to [0, 1]

, then A
i
is isomorphic to [0, 1]

, if W
i
is isomorphic to (0, 1]

then A
i
is isomorphic either to (0, 1]

or to [0, 1]

, and if W
i
is isomorphic to
L
2
then A
i
is isomorphic either to L
2
or to [0, 1]

.
The next decomposition properties are easily proved from the denitions of Fin
and Fin
t
.
LEMMA 5.1.5. Let A, B 1EG
t
.
1. Fin
t
(AB) = Fin
t
(A) Fin
t
(B).
2. Moreover, if A 1EG then Fin(AB) = Fin(A) Fin
t
(B).
5.2 Varieties generated by regular BL-chains
In this section we characterize the relation of inclusion between varieties of the form
V(), V(), when and range over continuous t-norms, in terms of the inclusion of
the sets Fin() and Fin(). In fact we will prove a more general result: if A, B 1EG,
then V(A) V(B) iff Fin(A) Fin(B). We start from the following:
LEMMA 5.2.1. Let A 1EG. Then V(A) = V(Fin(A)).
Proof. V(A) = V(ISP
u
(A)) = V(ISP
u
(Fin(A))) = V(Fin(A)), the second
equality holds by Lemma 5.1.3 (iii)).
COROLLARY 5.2.2. Let A, B 1EG. Then V(A) V(B) iff A V(B) iff A
V(Fin(B)) iff Fin(A) V(Fin(B)). In particular, if and are continuous t-norms
then V() V() iff [0, 1]

V() iff [0, 1]

V(Fin()) iff Fin() V(Fin()).

EXAMPLES 5.2.3.
1. Fin([0, 1]
G
) consists of all the nite ordinal sums of components of the form L
2
.
Hence, given A N with innitely many components, we have that Fin(A)
Fin([0, 1]
G
), and hence V(A) V([0, 1]
G
).
Chapter V: H ajeks BL Logic and BL-Algebras 401
2. Let Abe the ordinal sum of innitely many components of the form[0, 1]

. Then,
Fin(A) = Fin, and hence Agenerates the variety of BL-algebras.
3. If A is the ordinal sum of nitely many ukasiewicz and product components,
then Fin(A) is nite, and hence every chain in the variety generated by A must
have a nite number of components.
4. The product algebra on [0, 1]

belongs to the variety generated by the ordinal sum

of two copies of the MV-algebra [0, 1]

. The reader is invited to verify this.

DEFINITION 5.2.4. We introduce the following terms:
e
[0,1]

(x) = (x x
2
) ((x x
3
) x
2
),
e
(0,1]

(x) = x x
2
,
e
L
2
(x) = (x x
3
) x
2
.
LEMMA 5.2.5. We have:
1. the equation e
[0,1]

(x) = 1 is valid in L
2
and in cancellative hoops and it is not
valid in any MV-chain with more than two elements.
2. the equation e
(0,1]

(x) = 1 is valid in L
2
and it is not valid in any MV-chain
with more than two elements nor in any non-trivial cancellative hoop.
3. the equation e
L
2
(x) = 1 is valid in any cancellative hoop and it is not valid in L
2
nor in any MV-chain.
Proof. The proof is easy using the observation that in any MV-chain with more than
two elements there always exists an element 0 < x < 1 such that (x x
3
) x
2
< 1.
For instance, let y such that 0 < y < 1. Then, x = y y meets our requirements.
DEFINITION 5.2.6. Let A =

iI
A
i
and B =

iI
B
i
be two BL-chains such
that, for each i I, A
i
and B
i
are Wajsberg hoops, where I has a minimum i
0
and
both A
i
0
and B
i
0
are bounded. We stress that we are assuming that the index set I is
the same for Aand B. Then we say that Aand B are of the same type if for each i I,
both A
i
and B
i
are either MV-chains with more than two elements, or cancellative
hoops, or isomorphic to L
2
.
DEFINITION 5.2.7. Let A =

n
i=0
A
i
be a BL-chain, and for each i = 0 . . . n let
e
A
i
be e
[0,1]

if A
i
is an MV-algebra with more than two elements, be e
(0,1]

if A
i
is
a non-trivial cancellative hoop and be e
L
2
if A
i
is L
2
. Then we dene the following
equation
[(
n1

i=0
((x
i+1
x
i
) x
i
) ) (x
0
x
0
) (
n

i=0
x
i
)] (
n

i=0
e
A
i
(x
i
)) = 1 (e
A
)
402 Manuela Busaniche and Franco Montagna
Notice that, by construction, the equation (e
A
) is not valid in A. Namely, one can
always nd a sequence of values 0 x
0
< x
1
< < x
n
< 1 such that for each
i = 0, . . . , n, x
i
A
i
, and thus (
_
n1
i=0
((x
i+1
x
i
) x
i
) ) (x
0
x
0
) = 1,
but
_
n
i=0
x
i
= x
n
< 1 and e
A
i
(x
i
) < 1 for all i = 0, . . . , n.
LEMMA 5.2.8. (i) Let D 1EG, and let A Fin. Then the equation (e
A
) is
valid in all B Fin(D) iff A , Fin(D).
(ii) Let Band Dtwo BL-chains of the same type, and let A Fin. Then the equation
(e
A
) is valid in B iff it is valid in D.
Proof. (i) Let A =

n
i=0
A
i
. First of all observe that the claim trivially holds if the
number of components of D is nite and smaller than n + 1. Let us assume that D has
n + 1 components or more.
As for one direction, assume A , Fin(D) and take B =

k
i=0
B
i
Fin(D).
If k < n then, again, (e
A
) holds in B, so assume n k. We want to show that
the equation (e
A
) holds in B. Actually we have only to check the equation for all
(n + 1)-tuples x
0
< x
1
< . . . < x
n
< 1 where x
j
B
i
j
(j = 0, . . . , n) with
i
0
= 0 < i
1
< < i
n
, since otherwise the evaluation of the rst disjunct of the
equation (e
A
) is already 1. Since A , Fin(D), there exists 0 j n such that one
the following conditions holds:
(a) A
j
is [0, 1]

and B
i
j
is either (0, 1]

or L
2
, or
(b) A
j
is (0, 1]

and B
i
j
is L
2
, or
(c) A
j
is L
2
and B
i
j
is (0, 1]

If we are in case (a) then e

A
j
(x
j
) = e
[0,1]

(x
j
) = 1 for all x
j
B
i
j
, if we are in
case (b) then e
A
j
(x
j
) = e
(0,1]

(x
j
) = 1 for all x
j
B
i
j
, and if we are in case (c)
e
A
j
(x
j
) = e
L
2
(x
j
) = 1 for all x
j
B
i
j
. Hence, in any case, e
A
j
(x
j
) = 1 and thus the
equation (e
A
) is satised by the tuple x
0
, . . . , x
n
. Therefore (e
A
) is valid in B.
Reciprocally, it sufces to recall that the equation (e
A
) is not valid in A(see Lemma
5.2.5), therefore Acannot belong to Fin(D).
(ii) If B and D are of the same type, then the cardinalities of the sets of Wajsberg
components of both algebras must be the same. If B and D have less components
than A, then the equation (e
A
) trivially holds in both B and D. Otherwise, the claim
follows from Lemma 5.2.5, since (i) the fact that expression e
[0,1]

(x) (resp. e
(0,1]

(x)
and e
L
2
(x)) gets value 1 for all x in a component W only depends on whether W is
L
2
or cancellative or an MV-algebra with cardinality > 2, and (ii) the relevant values of
the variables involved x
0
, x
1
, . . . , x
n
are only those taken in different components, with
x
0
in the rst component, otherwise the equation trivially holds.
Note that, by Lemma 5.2.8, it follows that if is a continuous t-norm and A Fin,
then (e
A
) holds in all elements of Fin() iff A / Fin().
COROLLARY 5.2.9. Let D 1EG. Then, Fin ISP
u
(Fin(D)) = Fin(D). In
particular, if is a continuous t-norm and A Fin ISP
u
(Fin()), then A Fin().
Chapter V: H ajeks BL Logic and BL-Algebras 403
Proof. Suppose A Fin ISP
u
(Fin(D)) Fin(D). Then the equation (e
A
) is valid
in any algebra of Fin(D), hence it is also valid in any algebra of ISP
u
(Fin(D)), hence
it will also be valid in A, which is a contradiction with the fact that (e
A
) is not valid
in A.
THEOREM 5.2.10. Let D, E 1EG. Then V(D) V(E) iff Fin(D) Fin(E).
Thus, in particular, if and are two continuous t-norms, then V() V() iff
Fin() Fin().
Proof. If Fin(D) Fin(E) then V(D) = V(Fin(D)) V(Fin(E)) = V(E).
For the inverse implication, we recall that any variety V of BL-algebras is congruence-
distributive, hence by J onssons Lemma (see e.g. [18]), if K is any class such that
V(K) = V, then every subdirectly irreducible member of V is in HSP
u
(K). It follows
that every subdirectly irreducible member of V(E) is in HSP
u
(Fin(E)). Now [0, 1]

,
(0, 1]

and L
2
are simple (hence subdirectly irreducible) Wajsberg hoops. Moreover
by [13], if W is a subdirectly irreducible Wajsberg hoop and H is any hoop, then the
ordinal sum H W is subdirectly irreducible. It follows that if V(D) V(E), then
Fin(D) HSP
u
(Fin(E)).
Now let A =

k
j=0
W
j
Fin(D) be arbitrary (as usual, W
j
are the Wajsberg
components of A). Then there is B =

iI
U
i
ISP
u
(Fin(E)) (where as usual U
i
are the Wajsberg components of B) and a lter F of B such that A

= B/F. Without
loss of generality we may suppose F ,= 1 and F ,= B. By Proposition 4.1.2, there
are two possibilities:
(a) There is an i I such that A

=

ji
U
j
. Then, A Fin ISP
u
(Fin(E)),
and A Fin(E) by Corollary 5.2.9.
(b) There is i I such that and a proper non-trivial lter F
i
on U
i
such that A

=

j<i
U
j
(U
i
/F
i
). Then we have i = k, U
j

= W
j
for j < i, and W
i
=
U
i
/F
i
. Since W
i
is a quotient of U
i
, it follows that if W
i
is [0, 1]

then U
i
is
an innite linearly ordered Wajsberg chain, if W
i
= (0, 1]

then U
i
is a linearly
ordered cancellative hoop, and if W
i
= L
2
then U
i
is a bounded linearly ordered
Wajsberg hoop.
If W
i
= L
2
, then W
i
is a subalgebra U
i
and thus A ISP
u
(Fin(E)) and,
by Corollary 5.2.9, A Fin(E). Otherwise, either W
i
is [0, 1]

and U
i
is an
innite MVchain, or W
i
is (0, 1]

and U
i
is a cancellative hoop, so in any case
W
i
and U
i
are of the same type and so Aand B are.
Now suppose A , Fin(E). Then by (i) of Lemma 5.2.8, (e
A
) is valid in Fin(E).
Hence, it is valid in ISP
u
(Fin(E)), and hence also in B. But by (ii) of Lemma
5.2.8, (e
A
) must be also valid in A, which is a contradiction.
COROLLARY 5.2.11. Let A, B 1EG. Then A V(B) iff A ISP
u
(B). Thus if
Asatises all equations valid in B, it also satises all universal formulas valid in B.
Proof. By Theorem 5.2.10, if A V(B) then Fin(A) Fin(B), and therfore
ISP
u
(Fin(A)) ISP
u
(Fin(B)). By Lemma 5.1.3, it follows that ISP
u
(A)
ISP
u
(B). The other implication is trivial.
404 Manuela Busaniche and Franco Montagna
5.3 Canonical representation of regular BL-chains
DEFINITION 5.3.1. In the sequel, (0, 1]

denotes the ordinal sum of copies of

(0, 1]

, [0, 1]

denotes the ordinal sum of copies of [0, 1]

and [0, 1]

denotes the
ordinal sum of copies of [0, 1]

.
DEFINITION 5.3.2. Let A =

iI
A
i
1EG
t
, where for i I, A
i
(0, 1]

, L
2
.
We say that A has innitely many alternations of (0, 1]

and L
2
if for every n N
there are i
0
< i
1
< . . . < i
n
I such that for j = 0, . . . , n 1, if A
i
j
= (0, 1]

then
A
i
j+1
= L
2
, and if A
i
j
= L
2
, then A
i
j+1
= (0, 1]

.
LEMMA 5.3.3. Let A 1EG
t
with no [0, 1]

component. Then:
(i) If Ahas innitely many L
2
components and no (0, 1]

component,
then Fin
t
(A) = Fin
t
([0, 1]
G
).
(ii) If Ahas innitely many (0, 1]

components and no L
2
component,
then Fin
t
(A) = Fin
t
((0, 1]

).
(iii) If Ahas innitely many alternations of (0, 1]

and L
2
,
then Fin
t
(A) = Fin
t
([0, 1]

).
Furthermore, if A 1EG, then in cases (i) and (iii), Fin
t
(A) = Fin(A).
Proof. It sufces to check that Fin
t
(A) consists respectively of all nite ordinal sums
of L
2
components in case (i), of all nite ordinal sums of (0, 1]

components in case
(ii), and of all nite ordinal sums of L
2
and (0, 1]

components in case (iii).

LEMMA 5.3.4. Let A 1EG
t
with no [0, 1]

component, and let B 1EG and

D 1EG
t
be arbitrary. Then:
(i) If Ahas innitely many L
2
components and no (0, 1]

component,
then V(B AD) = V(B [0, 1]
G
D).
(ii) If Ahas innitely many (0, 1]

components and no L
2
component,
then V(B AD) = V(B (0, 1]

D).
(iii) If Ahas innitely many alternations of (0, 1]

and L
2
,
then V(B AD) = V(B [0, 1]

D).
Moreover, if A 1EG, then in case (i) V(AD) = V([0, 1]
G
D), and in case (iii)
V(AD) = V([0, 1]

D).
Proof. The equalities easily follow from Lemma 5.3.3, because by Theorem 3.9 it suf-
ces to show, for instance in case (i), that Fin(BAD) = Fin(B[0, 1]
G
D),
and by Lemma 5.1.5 this holds since Fin
t
(A) = Fin
t
([0, 1]
G
). The remaining cases
are treated analogously.
DEFINITION 5.3.5. A regular BL-algebra H is said to be canonical iff either H =
[0, 1]

, or H = L
2
[0, 1]

, or H is a nite ordinal sum of components of the form

[0, 1]

, L
2
, [0, 1]
G
, (0, 1]

, (0, 1]

and [0, 1]

, where
Chapter V: H ajeks BL Logic and BL-Algebras 405
(i) each component [0, 1]
G
is not preceded and not followed by L
2
or by another
[0, 1]
G
(ii) each component (0, 1]

is not preceded and not followed by (0, 1]

or by another
(0, 1]

(iii) each component [0, 1]

is not preceded and not followed by L

2
, [0, 1]
G
, (0, 1]

,
(0, 1]

or by another [0, 1]

.
Due to Lemma 5.3.4, canonical regular algebras which are indeed isomorphic to
t-norm BL-algebras, i.e ordinal sums of [0, 1]
G
, [0, 1]

and [0, 1]

components, can be
identied by the following simpler denition.
DEFINITION 5.3.6. A BL-chain A is said to be a canonical t-norm algebra iff either
A isomorphic to [0, 1]

, or A is isomorphic to [0, 1]

[0, 1]

, or A is isomorphic
to a nite ordinal sum of components of the form [0, 1]

, [0, 1]

, [0, 1]
G
and [0, 1]

,
where each component [0, 1]
G
is not preceded and not followed by another [0, 1]
G
, and
each component [0, 1]

is not preceded and not followed by [0, 1]

G
, or by [0, 1]

or by
another [0, 1]

.
REMARK 5.3.7. Notice that this latter denition introduces a small difference with
respect to the one for regular algebras: L
2
[0, 1]

is not a t-norm algebra, so it

is replaced by [0, 1]

[0, 1]

, that generates the same variety, the variety of SBL-

algebras.
As a matter of example, L
2
[0, 1]

[0, 1]
G
(0, 1]

is not a canonical regular

algebra, [0, 1]

(0, 1]

[0, 1]

[0, 1]

is canonical regular algebra but not a

canonical t-norm algebra, and nally, [0, 1]
G
[0, 1]

[0, 1]

is a canonical t-norm
algebra.
THEOREM 5.3.8. (i) For every regular BL-algebra H there is a canonical regular
BL-algebra K such that V(H) = V(K).
(ii) For every continuous t-norm there is a canonical continuous t-norm such
that V() = V().
Proof. We consider the case of regular algebras and we distinguish several cases:
1. If H has innitely many [0, 1]

components and the rst component is [0, 1]

,
then V(H) = V([0, 1]

) is the variety of BL-algebras [1].

2. If H has innitely many [0, 1]

components and [0, 1]

is not the rst component,

then V(H) = V(L
2
[0, 1]

) is the variety of SBL-algebras.

3. If H has nitely many [0, 1]

components, say [0, 1]

1
, . . . , [0, 1]

n
, then H is of
the form
H = A
1
[0, 1]

1
A
2
A
n
[0, 1]

n
A
n+1
.
If A
i
is not a nite (or empty) ordinal sum of components isomorphic to L
2
or
(0, 1]

, then it must contain innitely many L

2
or innitely many (0, 1]

compo-
nents (or both) and no [0, 1]

component. We distinguish the following cases:

406 Manuela Busaniche and Franco Montagna
(a) A
i
has nitely many occurrences of (0, 1]

and innitely many occurrences

of L
2
. Then we replace any maximal ordinal sum of innitely many consec-
utive L
2
components in A
i
by [0, 1]
G
.
(b) A
i
has nitely many occurrences of L
2
and innitely many occurrences
of (0, 1]

. Then we replace any maximal ordinal sum of innitely many

consecutive (0, 1]

components in A
i
by (0, 1]

.
(c) A
i
has innitely many occurrences of (0, 1]

and innitely many occur-

rences of L
2
, but only nitely many alternations of (0, 1]

and L
2
. Then we
replace any maximal ordinal sum of innitely many consecutive L
2
compo-
nents in A
i
by [0, 1]
G
, and any maximal ordinal sum of consecutive (0, 1]

components in A
i
by (0, 1]

.
(d) A
i
has innitely many alternations of (0, 1]

and L
2
. Then we replace A
i
by [0, 1]

.
After performing operations (a), (b), (c) and (d) for all A
i
, we obtain a canonical
regular BL-algebra Awhich, by Lemma 5.3.4, generates the same variety as H.
As for the case of A = [0, 1]

being a continuous t-norm algebra, the above rst case is

the same, the second one reads
2
t
. If Ahas innitely many [0, 1]

components and [0, 1]

is not the rst component,

then V(A) = V([0, 1]

[0, 1]

) = V([0, 1]
G
[0, 1]

) is the variety of
SBL-algebras.
and the third one can be simplied as follows:
3
t
. If Ahas nitely many [0, 1]

components, let [a
1
, b
1
], . . . , [a
n
, b
n
] be such [0, 1]

components enumerated in increasing order. Let A

i
be any of the components
[0, a
1
], [b
1
, a
2
], . . . , [b
n1
, a
n
] and [b
n
, 1]. If A
i
is not a nite (or empty) or-
dinal sum of components isomorphic to [0, 1]

or to [0, 1]
G
(after having glued
together all contiguous [0, 1]
G
components), then it must contain innitely many
[0, 1]

components and no [0, 1]

component. In this case we replace A

i
by (an
isomorphic copy of) [0, 1]

.
After we have performed this operation for all A
i
, we obtain a regular t-norm
algebra B = [0, 1]

. By an iterated application of Lemma 5.3.4, we see that

V() = V().
EXAMPLES 5.3.9.
1. The BL-algebra H = L
2
[0, 1]

[0, 1]
G
(0, 1]

is not canonical, but there

is a canonical BL-chain which generates the same variety, namely K = [0, 1]

.
Notice that in this case, K is indeed a t-norm BL-algebra, while H is not.
2. Suppose that two continuous t-norms and have a different but nite number
of ukasiewicz components. Then, V() ,= V(). Indeed, assuming that has
n ukasiewicz components and that has m < n ukasiewicz components, the
ordinal sum of n ukasiewicz components is in Fin() but not in Fin().
Chapter V: H ajeks BL Logic and BL-Algebras 407
3. Suppose that two continuous t-norms and have a nite number of ukasiewicz
components and a different, but nite, number of components [0, 1]

. Then,
V() ,= V().
4. Take a nite ordinal sum of components of the form [0, 1]
G
, [0, 1]

, [0, 1]

,
[0, 1]

and [0, 1]

. Consider the reductions:

[0, 1]
G
W W, where W is either [0, 1]
G
or [0, 1]

.
W[0, 1]
G
W, where W is either [0, 1]
G
or [0, 1]

.
[0, 1]

W [0, 1]

, where W is either [0, 1]

or [0, 1]

.
W[0, 1]

[0, 1]

, where W is either [0, 1]

or [0, 1]

.
W[0, 1]

[0, 1]

, where is any nite sequence of words of the form

above and W is either [0, 1]

or [0, 1]

.
W[0, 1]

[0, 1]

[0, 1]

, where is any nite sequence as above and

W is either [0, 1]
G
or [0, 1]

or [0, 1]

.
[0, 1]

[0, 1]

, where is as above.
Then, a nite sequence as shown above represents a canonical t-norm iff no sub-
word of it can be reduced.
COROLLARY 5.3.10. Let H =

n
i=0
H
i
and K =

m
i=0
K
j
be two canonical
regular BL-chains, where H
i
, K
i
[0, 1]
G
, [0, 1]

, [0, 1]

, [0, 1]

, [0, 1]

. Then
V(H) = V(K) if and only if n = m and H
i
= K
i
for each i = 1, . . . , n.
Proof. Assume V(H) = V(K). We can restrict ourselves to the case of both Hand K
having a nite number of [0, 1]

components, otherwise either H = K = [0, 1]

or
H = K = L
2
[0, 1]

. We shall prove several claims.

Claim 1: H and K have the same number of [0, 1]

components and the same rst

Wajsberg components.
Proof of Claim 1. We prove that otherwise Fin(H) ,= Fin(K). Let l
H
and l
K
the
number of [0, 1]

components in H and K respectively. If l

H
> l
K
, then any ordinal
sum of Fin(H) with l
H
components [0, 1]

will not belong to Fin(K). The case of

l
H
< l
K
is analogous. Finally, consider the case l
H
= l
K
but with H and K starting
with different Wajsberg components. Without loss of generality, assume H starts with
L
2
and K not. Then L
2
[0, 1]

l
H
. . . [0, 1]

will belong to Fin(H) and not to

Fin(K).
Fromnowon, let us denote by [0, 1]

1
, . . . , [0, 1]

k
the k components [0, 1]

appear-
ing in both the canonical decompositions of Hand K. Then, let U and V be non-empty
ordinal sums with no [0, 1]

component such that H = . . .[0, 1]

i
U[0, 1]

i+1
. . .
and K = . . . [0, 1]

i
V [0, 1]

i+1
. . ..
Claim 2: Fin
t
(U) = Fin
t
(V ).
408 Manuela Busaniche and Franco Montagna
Proof of Claim 2. If Fin
t
(U) ,= Fin
t
(V ), let e.g. A be such A Fin
t
(V ) and A /
Fin
t
(U). Then B = [0, 1]

1
. . . [0, 1]

i
A[0, 1]

i+1
. . . [0, 1]

k
will belong
to Fin(K) but not to Fin(H), hence it would be Fin(K) ,= Fin(H).
Hence we can assume Fin
t
(U) = Fin
t
(V ). Now if both U and V have innitely-
many alternations L
2
/(0, 1]

, then U = V = [0, 1]

.
Claim 3: If U has nitely-many alternations L
2
/(0, 1]

(resp. (0, 1]

/L
2
) then the
same is true of V , and viceversa. Moreover, if the rst alternation in U is of type
L
2
/(0, 1]

(resp. (0, 1]

/L
2
), so it is in V .
Proof of Claim 3. If V has more alternations L
2
/(0, 1]

than U, we can always con-

sider a sufciently large ordinal sum A = L
2
(0, 1]

. . . L
2
(0, 1]

such that
B = [0, 1]

1
. . . [0, 1]

i
A[0, 1]

i+1
. . . [0, 1]

k
will belong to Fin(K) but
not to Fin(H). Here we have assumed that H and K start with a [0, 1]

component,
otherwise, if they start with a L
2
component, one has to add such component at the be-
ginning of B. The case of alternations (0, 1]

/L
2
is dealt with in the same way. Finally,
if U and V have the same number of alternations (both L
2
/(0, 1]

and (0, 1]

/L
2
), say
s, but, e.g., U starts with L
2
and V with (0, 1]

, then (

i=1,s
(L
2
(0, 1]

)) L
2
would belong to Fin
t
(U) but not to Fin
t
(V ), contradicting Claim 2.
Hence we can restrict ourselves to the case of both U and V having the same num-
ber (and type) of alternations. Without loss of generality, assume the rst alternation
in both is of the type L
2
/(0, 1]

. If t is the number of such alternations, then we can

decompose U and V in terms of Wajsberg components in the following form:
U =

i=1,t
(D
1
i
L
2
(0, 1]

E
1
i
),
V =

i=1,t
(D
2
i
L
2
(0, 1]

E
2
i
)
where the D
j
i
s are ordinal sums of copies of L
2
s and the E
j
i
s are ordinal sums of
copies of (0, 1]

s, possibly empty.
Claim 4: for each i, D
1
i
= D
2
i
and E
1
i
= E
2
i
.
Proof of Claim 4. it is easy to see that if for some i either D
1
i
,= D
2
i
or E
1
i
,= E
2
i
, then
Fin
t
(D
1
i
) ,= Fin
t
(D
2
i
) or Fin
t
(E
1
i
) ,= Fin
t
(E
2
i
) respectively. Assume w.l.g i = 1,
D
1
i
,= D
2
i
and let A Fin
t
(D
1
i
) and A / Fin
t
(D
2
i
). Then we easily have that
A(

i=1,t
(L
2
(0, 1]

))
belongs to Fin(U) and not to Fin(V ), contradiction.
Finally, from Claims 1, 2, 3 and 4, it follows that V(H) = V(K) only if they have
exactly the same components in their canonical decomposition, i.e. only if H = K.
Chapter V: H ajeks BL Logic and BL-Algebras 409
Another interesting corollary of the above theorem is the following.
COROLLARY 5.3.11. The set of varieties generated by regular BL-algebras is count-
able. In particular, the set of the ones generated by t-norm BL-algebras is so.
REMARK5.3.12. For continuous t-normBL-algebras, Corollary 5.3.11 can also be de-
rived from [60] where it is shown that every variety generated by a t-normBL-algebra is
co-NP complete, using the fact that there are only countably many co-NP sets. However,
here we have given an explicit description of the above varieties. In [45], the authors
also present an algorithm for axiomatizing all the varieties mentioned above, which is
not reproduced here.
5.4 Axiomatizing the variety generated by a nite BL-chain
For each r 2, the variety of MV-algebras (Wajsberg algebras) generated by L
r
is axiomatized by an equation in one variable only, which can be written in the form
t
r
(x) = 1 (see [38]). Now let t

r
(x) be the term obtained by replacing in t
r
(x) the
constant 0 by x
r
. Observe that if x L
r
and x < 1, then x
r
= 0. Thus t

r
(x) = t
r
(x) =
1. In case x = 1, then trivially t

r
(1) = 1. Clearly t

r
(x) = 1 is an equation in the
language of hoops, and the reader can verify that a Wajsberg hoop satises t

r
(x) = 1 if
and only if it is in the variety of Wajsberg hoops generated by L
r
.
Based on these axiomatizations of varieties of Wajsberg algebras and Wajsberg
hoops, we will now axiomatize varieties generated by a nite BL-chain. To achieve
such aim, we consider the nite BL-chain
C =
k

j=1
L
n
j
,
with each n
j
2. For each j = 2, . . . , k we let t

n
j
(x) = 1 be the equation in the
language of Wajsberg hoops previously dened, but for j = 1 we let t

n
1
(x) = t
n
1
(x).
The rst easy to prove result is:
LEMMA 5.4.1. Let D =

k
j=1
L
s
j
, where for each j = 1, . . . , k, s
j
is a natural
number greater or equal than 2. The following are equivalent:
1. L
s
1
is a subalgebra of L
n
1
and for j = 2, . . . , k , L
s
j
is a subhoop of L
n
j
2. For j = 1, 2, . . . , k, if x L
s
j
, then t

n
j
(x) = 1 holds in D.
Given a BL-chain C, we dene for x, y C, x y iff x < y and either y = 1
or x and y do not belong to the same Wajsberg component. Further, we dene x y =
(x y) y. Then, we know that if y x, then x y = 1, otherwise, x y = x y.
LEMMA 5.4.2. Let Dbe any BL-chain. Then, Dis a subalgebra of C iff the following
condition holds:
(1) For all x
1
, . . . , x
m
D, if x
1
x
m
< 1, then m k, and if in
addition x
1
= x
1
, then there are 1 = j
1
< j
2
< < j
m
k such that for
i = 1, 2, . . . , m, t

n
j
i
(x
i
) = 1.
410 Manuela Busaniche and Franco Montagna
Proof. Dis a subalgebra of C iff D =

h
r=1
L
s
r
where h k, L
s
1
is a subalgebra of
L
n
1
and there are 1 < j
2
< < j
h
k such that for r = 2, . . . , h, L
s
r
is a subhoop
of L
n
j
r
.
Assume that D =

h
r=1
L
s
r
is a subalgebra of C. Let x
1
, . . . , x
m
D be such
that x
1
x
m
< 1. Then m h k, as the number h of components of D is
k. Moreover, suppose x
1
= x
1
. Then, x
1
L
s
1
1. Since L
s
1
is a subalgebra
of L
n
1
, one must have t

n
1
(x
1
) = t
n
1
(x
1
) = 1. Now take 2 i m. Then there are
r
i
2 and j
i
2 such that x
i
L
s
r
i
and L
s
r
i
is a subhoop of L
n
j
i
. Thus t

n
j
i
(x
i
) = 1.
Then there are 1 < j
2
< < j
m
h such that for i = 2, . . . , m, t

n
j
i
(x
i
) = 1.
Conversely, suppose that D satises condition (1). Then D has h k Wajsberg
components, call them U
1
, . . . , U
h
. Take, for i = 1, . . . , h, x
i
U
i
1. Then,
x
1
x
h
< 1 and x
1
= x
1
. Hence, by (1), t

n
1
(x
1
) = 1, and hence U
1
is a
subalgebra of L
n
1
. Moreover, there is a sequence 1 = j
1
< j
2
< < j
h
k such
that for i = 2, . . . , h, t

n
j
i
(x
i
) = 1, and hence U
i
is a subhoop of L
n
j
i
. It follows that
D is a subalgebra of C.
Our next task is to express condition (1) by means of equations. For m = 2, . . . , k,
let S
m
be the set of nite sequences (j
1
, . . . , j
m
) of length m such that 1 j
1
< <
j
m
k, and let for s S
m
, t

s
=
_
m
i=1
t

n
j
i
(x
i
). For m = 2, . . . , k, we dene the
inequality
m1

i=1
(x
i+1
x
i
)
_

sS
m
t

s
_

i=1
x
i
. (
m
)
Consider the following equations:
k

i=1
(x
i+1
x
i
)
k+1

i=1
x
i
= 1. (1)
t

n
1
(x) = 1. (2)
THEOREM 5.4.3. Let D be a BL-chain. Then D satises (1), (2), and (
m
) for m =
2, . . . , k iff D is a subalgebra of C. Hence, the variety generated by C is axiomatized
by (1), (2), and (
m
), m = 2, . . . , k.
Proof. (1) is satised in all BL-chains which are ordinal sums of k Wajsberg com-
ponents at most, and (2) holds in a BL-chain when its rst Wajsberg component is a
subalgebra of L
n
1
. Hence (1) and (2) hold in a BL-chain D precisely when D is the
ordinal sum of m k Wajsberg components and the rst is a subalgebra of L
n
1
.
Now let us consider the equations (
m
), m = 2, . . . , k. We claim that a BL-chain
D satisfying (1) and (2) is a subalgebra of C iff it satises (
m
), for m = 2, . . . , k.
Suppose D is a subalgebra of C. Then D =

h
r=1
L
s
r
where h k, L
s
1
is a
subalgebra of L
n
1
and there are 1 < j
2
< < j
h
k such that for r = 2, . . . , h, L
s
r
is a subhoop of L
n
j
r
. We can assume that h 2. Let x
1
, . . . , x
m
D be given. We only
care for the case that x
1
x
m
< 1, because otherwise either
_
m1
i=1
(x
i+1
x
i
)
=
_
m1
i=1
x
i
or
_
m
i=1
x
i
= 1, and (
m
) trivially holds. But if x
1
x
m
< 1, then
Chapter V: H ajeks BL Logic and BL-Algebras 411
for i = 1, . . . , m there are 1 r
1
< < r
m
k ,1 j
1
< . . . j
m
k such that
x
i
L
s
r
i
1, L
s
r
i
is a subalgebra (or a subhoop) of L
n
j
i
, and j
1
= 1 iff r
1
= 1.
Then t

n
j
i
(x
i
) = 1. Taking s = (j
1
, . . . , j
m
) S
m
the equation t

s
= 1 holds and (
m
)
is satised.
Now suppose that Dis not a subalgebra of C. Since Dsatises (1) and (2), it is the
ordinal sum of m k components U
1
, . . . , U
m
such that U
1
is a subalgebra of L
n
1
,
and there is no sequence 1 = j
1
< j
2
< < j
m
such that for i = 2, . . . , m, U
i
is a
subhoop of L
n
j
i
. Nowtake, for i = 1, . . . , m, x
i
U
i
1. Then,
_
m1
i=1
(x
i+1
x
i
) =
1,
_
m
i=1
x
i
< 1, and for all s S
m
, t

s
< 1. It follows that (
m
) is not satised.
REMARK 5.4.4. Observe that if C =

k
j=1
L
n
j
is a nite BL-chain, since each of
its Wajsberg components is a nite MV-chain (or its hoop reduct), each of them is a
k-potent algebra for some k. Then there is n = maxn
j
[ j = 1, . . . k, such that C
is an (n 1)-potent BL-chain. Thus the subvariety generated by C also satises the
equation x
n
= x
n1
.
EXAMPLES 5.4.5.
1. Consider the MV-chain L
n
for some n 2. Let 1 x = x and (k + 1) x =
(x (k x)) for each k 1. Then the variety of MV-algebras generated by
L
n
can be axiomatized as a subvariety of MV-algebras by the following set of
equations:
x
n
x
n1
= 1 (3)
if n 4, for every p = 2, . . . n 2 that does not divide n 1:
(p x
p1
)
n
n x
p
= 1 (4)
Of course the conjunction of these equations yields a single equation, that we
previously denoted by t
n
(x) = 1. Observe that the equation t

n
(x) = 1 is the one
obtained from t
n
(x) = 1 replacing every occurrence of 0 by x
n
, so e.g. x is
replaced by x x
n
. According to Theorem 5.4.3, the subvariety of BL-algebras
generated by L
n
is axiomatized by:
(x)
n
(x)
n1
= 1 (5)
if n 4, for every p = 2, . . . n 2 that does not divide n 1:
(p (x)
p1
)
n
n (x)
p
= 1 (6)
(x
2
x
1
) x
1
x
2
= 1. (7)
2. Consider the BL-chain C = L
n
L
m
, where n, m are natural numbers greater
or equal than 2. Recall that x C is in the second Wajsberg component iff x is
a dense element, i.e., it satises x x = x. If x C is in the rst Wajsberg
component, then x x = 1. Thus the equation t

m
(x x) = 1 holds
in C. Conversely, suppose B = U
1
U
2
is a BL-chain that has only two nite
412 Manuela Busaniche and Franco Montagna
Wajsberg components and it satises t

m
(x x) = 1. Then, since for each x
in U
2
we have x = x x, we know from Lemma 5.4.1 that U
2
is a subhoop
of L
m
. Then the subvariety of BL-algebras generated by C can be axiomatized
following Theorem 5.4.3 or simply by
t
n
(x) = 1 (8)
t

m
(x x) = 1 (9)
2

i=1
(x
i+1
x
i
)
3

i=1
x
i
= 1. (10)
Notice that also for the specic case that C =

k
j=1
L
n
j
and there is m 2
such that for each j = 2, . . . k, n
j
= m, then axiomatization of the subvariety of
BL-algebras generated by C can be made simpler by:
t
n
1
(x) = 1 (11)
t

m
(x x) = 1 (12)
k

i=1
(x
i+1
x
i
)
k+1

i=1
x
i
= 1. (13)
Indeed, it is easy to see that C satises the previous set of axioms. On the other
hand, if D is a chain satisfying these three axioms, the last equation implies that
D is a chain with at most k Wajsberg components. The rst equation yields that
the rst component of Dhas to be a subalgebra of L
n
1
. Lastly, if U is a Wajsberg
component of Dthat is not the rst, then every element x U is a dense element,
thus x = x x. Since D satises the second equation, U is a chain in the
subvariety of Wajsberg hoops generated by L
m
. Thus U is a subhoop of L
m
and
D is a subalgebra of C.
5.5 Axiomatizing the variety generated by an arbitrary nite BL-algebra
Given any nite BL-algebra A, there are nite BL-chains A
1
, . . . , A
n
such that
A is a subdirect product of A
1
, . . . , A
n
. Now let t
i
(x
i
) = 1, where x
i
denotes a
nite string of variables, be an equation axiomatizing the variety generated by A
i
, i =
1, . . . , n. (Such an equation exists, as we have seen that the variety generated by a nite
BL-chain is axiomatizable by nitely many equations). We can assume, without loss of
generality, that if i ,= j, then the strings x
i
and x
j
have no variable in common. Then,
we claim that
_
n
i=1
t
i
(x
i
) = 1 axiomatizes the variety generated by A. Indeed, clearly
_
n
i=1
t
i
(x
i
) = 1 holds in all A
i
and hence it holds in all subalgebras of the product of
the A
i
, and hence it holds in A. Conversely, we have to prove that if a BL-algebra B
satises
_
n
i=1
t
i
(x
i
) = 1, then it belongs to the variety generated by A. Since B is
a subdirect product of BL-chains, it sufces to prove that each subdirectly irreducible
factor C of B belongs to the variety generated by A. We claim that for some i, C
satises t
i
(x
i
) = 1. Suppose not. Then, for i = 1, . . . , n we may nd a
i
C such that
t
i
(a
i
) < 1, and since C is a chain, we conclude that
_
n
i=1
t
i
(a
i
) < 1, a contradiction.
Therefore:
Chapter V: H ajeks BL Logic and BL-Algebras 413
THEOREM 5.5.1. Suppose that A is a nite BL-algebra and that A
1
, . . . , A
n
are its
subdirectly irreducible factors. Let for i = 1, . . . , n, t
i
(x
i
) = 1 be an equation which
axiomatizes the variety generated by A
i
. Assume that for each i ,= j the strings x
i
and
x
j
have no variable in common. Then, the variety generated by A is axiomatized by
_
n
i=1
t
i
(x
i
) = 1.
6 Amalgamation, interpolation and Beths property in BL
6.1 General facts about interpolation and Beths property
Let L be any logic. For every formula of L, let Var() be the set of variables occurring
in . We say that L has the deductive interpolation property if whenever
L
, there
is a formula , called an interpolant between and , such that
L
,
L

and Var() Var() Var(). The intuitive meaning of this property is that if
is derivable from , then the information from which is relevant to derive can be
expressed in the language (i.e. variables) common to and .
There is another form of interpolation, namely, the Craig interpolation, that is: a
logic L has the Craig interpolation if for any two formulas and , if
L
, then
there is a formula such that
L
,
L
and Var() Var() Var().
The rst form of interpolation can be generalized to the case where is replaced by
any set of formulas. Moreover, both forms of interpolation have two uniform versions,
that is:
A logic L has uniform right deductive interpolation if for every formula of L
and for every X Var(), there is a formula such that Var() = X,
L

and for every formula such that Var() Var() X, if
L
, then

L
.
A logic L has uniform left deductive interpolation if for every formula of L and
for every X Var(), there is a formula such that Var() = X,
L
and
for every formula such that Var() Var() X, if
L
, then
L
.
A logic L has uniform right Craig interpolation if for every formula of L and
for every X Var(), there is a formula such that Var() = X,
L

and for every formula such that Var() Var() X, if
L
, then

L
.
A logic L has uniform left Craig interpolation if for every formula of L and
for every X Var(), there is a formula such that Var() = X,
L

and for every formula such that Var() Var() X, if
L
, then

L
.
The formula is said to be a right (respectively left) uniform interpolant of with
respect to the set X of variables. The intuition is that the right interpolant is a for-
mula which contains all the information contained in which can be expressed in the
language restricted to the variables in X, and the left interpolant is a formula which
contains the minimum information in the above language which is sufcient to derive .
414 Manuela Busaniche and Franco Montagna
In the presence of weakening and of contraction, deductive interpolation and Craig
interpolation are equivalent, because
L
is equivalent to
L
. We will see
that this equivalence does not hold in general for fuzzy logics, because of the lack of
contraction.
A property related to interpolation is Beths denability property. This property can
also be formulated in (at least) two ways. For convenience, throughout this section we
write for ( ) ( ), instead of ( ) & ( ). Then we
introduce the following denition:
DEFINITION 6.1.1. We say that a formula (a, p) (p a propositional variable, a a
string of propositional variables) implicitly denes p in L iff (a, p), (a, q)
L
p q.
We say that (a, p) explicitly denes p in L if there is a formula (a) in the variables
a only such that (a, p)
L
p (a).
We say that a logic L has the deductive Beth property if whenever (a, p) implicitly
denes p in L, then (a, p) explicitly denes p in L.
As said before, Beths property has an alternative formulation in terms of implica-
tion. The denition of the implicative Beths property is obtained from the denition
of the deductive Beths property replacing the consequence relation by the implication.
More precisely:
DEFINITION 6.1.2. We say that L has the implicative Beth property iff whenever
L
((a, p) & (a, q)) (p q), then there is a formula (a) such that
L
(a, p)
(p (a)).
For the logics we are interested in, the implicative Beth property follows easily from
Craig interpolation. In detail if
L
((a, p)&(a, q)) (p q), then
L
((a, p)&p)
((a, q)) q). By Craigs interpolation property, there is a formula (a) in the
variables a such that:
L
((a, p) &p) (a) and
L
(a) ((a, q) q). The rst
condition implies
L
(a, p) (p (a)), and the second one, with q replaced by
p, implies
L
(a, p) ((a) p). To the contrary, we will see that in many-valued
logic deductive interpolation does not imply the deductive Beth property.
We remark that deductive interpolation and deductive Beths property may be gen-
eralized e.g. considering a (possibly innite) set of assumptions instead of a single as-
sumption, see [50].
In general, there are two kinds of proofs of interpolation and related properties:
syntactic proofs, usually by means of cut-elimination, and semantic proofs. Typically,
syntactic proofs are more informative, because they provide for an explicit construction
of an interpolant. However, in the case of BL, due to the lack of a good proof-theory, we
are forced to use semantic means. Interpolation and related properties are strictly related
to various kinds of amalgamation properties. We will investigate two of them, namely
amalgamation and strong amalgamation. In [80], it is shown that for superintuinistic
logics both deductive and implicative interpolation are equivalent to both amalgama-
tion and strong amalgamation. Moreover, Beths property is weaker than interpolation
in these logics. However, superintuitionistic logics have the deduction theorem and
many-valued logics do not in general, therefore the latter logics behave quite differently.
The situation of modal logics, investigated in full details in [81], is closer to that of
Chapter V: H ajeks BL Logic and BL-Algebras 415
many-valued logics, because several modal logics do not have the deduction theorem.
In [81], see also [49], deep relations are shown between interpolation, Beth property and
amalgamation properties in modal logics and in their corresponding varieties. For in-
stance, deductive interpolation is equivalent to amalgamation, implicative interpolation
is equivalent to superamalgamation, and strong amalgamation is equivalent to amalga-
mation plus Beths property.
6.2 Amalgamation
DEFINITION 6.2.1. Let K and K
t
be classes of algebras of the same type. Then a
V -formation in K is a tuple A, B, C, i, j such that A, B, C K and i and j are
embeddings from Ainto B and into C respectively.
Let A, B, C, i, j be a V -formation in K, let D K
t
and h, k be embeddings
of B and of C respectively into D. We say that D, h, k is an amalgam in K
t
of
A, B, C, i, j if the diagram
D

h

k
B C

i

j
A
commutes. We say that K has the amalgamation property (AP for short) if every
V -formation in K has an amalgam in K.
In the sequel, if A, Band C are algebras of the same type, we will write A = BC
to mean that Ais a subalgebra of both B and C and the domain of Ais the intersection
of the domains of B and C. The following lemma is almost obvious.
LEMMA 6.2.2. Let K be a class of algebras of the same type closed under isomorphic
images. Suppose that for every A, B and C in K such that A = B C, there are an
algebra D K and embeddings h and k of B and C respectively into D such that the
restrictions of h and k to Acoincide. Then K has the AP.
Notation. (a) With the same notation as in Lemma 6.2.2, we say that D, h, k is an
amalgam of A, B, C.
(b) In the sequel, BH, B', WH, W= MV, CH, P, G, and 'AG denote the classes
of basic hoops, of BL-algebras, of Wajsberg hoops, of Wajsberg algebras (identied
with the class of MV-algebras), of cancellative hoops, of product algebras, of G odel
algebras and of lattice ordered Abelian groups respectively.
(c) Given a class K of lattice ordered algebras, we denote by K
lin
the class of
linearly ordered members of K.
We recall that a class K of algebras has the Congruence Extension Property (CEP,
for short) iff whenever A, B Kand Ais a subalgebra of B, then for every congruence
of Athere is a congruence
t
of B such that =
t
A
2
.
We also recall that a prime lter of a hoop A is a lter F of A such that for all
a, b A, if a b F, then either a F, or b F. It is readily seen that the quotient
of a basic hoop modulo a prime lter is totally ordered.
Let us say that a variety K of hoops, possibly with additional operators and con-
stants, has the Prime Filter Extension Property (PFEPfor short) iff for every A, B K,
416 Manuela Busaniche and Franco Montagna
if Ais a subalgebra of B, then for every prime lter F of A, there is a prime lter G of
B such that F = G A.
LEMMA 6.2.3. Let K be a variety or a quasivariety of BL-algebras such that:
(i) K has the PFEP.
(ii) For any A K, for any lter F of A and for any a A such that a / F, there
is a prime lter P of Asuch that a / P and F P.
(iii) Every V -formation in K
lin
has an amalgam in K.
Then K has the AP.
Proof. Let A, B, C in K be such that A = B C. For any b B 1, take a prime
lter P
b
of B such that b / P
b
. Let M
b
= A P
b
. Then M
b
is a prime lter of A, and
by the PFEP there is a prime lter Q
b
of C such that Q
b
A = M
b
. Let B
b
= B/P
b
,
A
b
= A/M
b
, and C
b
= C/Q
b
. Then A
b
, B
b
and C
b
are in K
lin
, as M
b
, P
b
and Q
b
are prime lters.
Similarly if c C 1 there is a prime lter Q
c
of C such that c / Q
c
. Moreover,
M
c
= Q
c
A is a prime lter of A, and we can extend it to a prime lter P
c
of B such
that P
c
A = M
c
. Let A
c
= A/M
c
, B
c
= B/P
c
, C
c
= C/Q
c
, and note that A
c
, B
c
and C
c
are in K
lin
.
Now for d (B C) 1, A
d
embeds into both B
d
and C
d
via the embeddings
i
d
and j
d
dened by i
d
(x/M
d
) = x/P
d
and j
d
(x/M
d
) = x/Q
d
(it is readily seen that
i
d
and j
d
are well-dened). After replacing A
d
, B
d
and C
d
with isomorphic copies,
we can suppose that A
d
= B
d
C
d
. By (iii), there is an amalgam D
d
, h
d
, k
d
of
A
d
, B
d
, C
d
such that D
d
K.
Moreover, for b B 1, b/P
b
,= 1, and for c C 1, c/Q
c
,= 1. Thus the
mappings f, g dened by f(x) = x/P
d
[ d (B C) 1 and g(y) = y/Q
d
[
d (B C) 1 are embeddings of B and C respectively into

d(BC)\1]
B
d
and

d(BC)\1]
C
d
respectively. Moreover, f and g coincide on all the elements
of A.
Finally,

d(BC)\1]
B
d
and

d(BC)\1]
C
d
embed into

d(BC)\1]
D
d
via the embeddings h and k dened by h(x)
d
= h
d
(x/P
d
), and k(x)
d
= k
d
(x/Q
d
)
respectively. Let D =

d(BC)\1]
D
d
. Then, D K, and D, h f, k g is an
amalgam of A, B, C.
It is well-known that any variety of (bounded or unbounded) hoops has the CEP:
any congruence of a hoop A is associated to a lter F of A, and if A is a subalgebra
of B and G is the lter of B generated by F, then G A = x A [ a
1
, . . . , a
n
F
such that a
1
. . . a
n
x = F. Thus if
t
is the congruence of B associated to G, one
has =
t
A
2
.
LEMMA 6.2.4. Any variety V of basic hoops or of BL-algebras satises conditions (i)
and (ii) of Lemma 6.2.3.
Chapter V: H ajeks BL Logic and BL-Algebras 417
Proof. (i) Let A, B in V be such that Ais a subalgebra of B, and let F be a prime lter
of A. The set K of all lters G of B such that G A = F is non-empty (since V has
the CEP), and it is closed under unions of chains. By Zorns Lemma, K has a maximal
element P. We claim that P is prime. Let for X A, Fg(X) denote the lter generated
by X. Suppose by contradiction that for some a, b B we have a b P, a / P and
b / P. Then by the maximality of P there are c, d A F such that c Fg(P a),
and d Fg(P b). Thus c d Fg(P a) Fg(P b).
Now Fg(P a) Fg(P b) = Fg(P a b). Indeed, we have a b
Fg(P a) Fg(P b), therefore Fg(P ab Fg(P a) Fg(P b).
On the other hand, if x Fg(P a) Fg(P b), then there are p, q P and
m, n N such that p a
n
x and q b
m
x. If k = maxm, n and r = p q, then
r (a b)
k
= (r a
k
) (r b
k
) x. So x Fg(P a b), and the other inclusion
is proved. Summing-up, c d Fg(P a b), and since a b P, it follows that
c d P A = F. Since F is prime, either c F or d Fa contradiction.
(ii) That condition (ii) holds in any BL-algebra is proved in [58], and the proof for
basic hoops is quite similar. The idea is the following: given an implicative lter F and
an element a / F, using Zorns Lemma one obtains an implicative lter M which is
maximal among all congruence lters which contain F and do not have a as an element.
Then along the lines of the proof of (i) we can prove that the maximality condition
mentioned above implies that M is prime.
We proceed with the proof that if V is any of W, CH, WH, and P, then V has the
AP. By Lemmas 6.2.3 and 6.2.4, it is sufcient to prove that V
lin
has the AP. We recall
the following result, see [3, 88].
LEMMA 6.2.5. The classes 'AG and 'AG
lin
have the AP.
COROLLARY 6.2.6. The classes W
lin
, CH
lin
, WH
lin
, and P
lin
have the AP. Hence
(Lemmas 6.2.3 and 6.2.4), the varieties W, CH, WH, and P have the AP.
Proof. The result for W
lin
(and more generally for W) is due to Mundici [86], but since
a similar argument will be used also for other varieties, we briey sketch a proof. We
will use some known results. Recall that a strong unit of an -group G is an element
u G such that for all g G there is a positive integer n such that g nu. Given an
-group Gwith a strong unit u, (G, u) denotes the MV-algebra with domain the closed
interval [0, u] of Gand with operations and dened by x y = (x +y u) 0 and
x y = (ux+y)u. Given an MV-algebra A, there is a (unique up to isomorphism)
-group with strong unit (G, u) denoted by
1
(A), such that (G, u) = A, see [85].
We also use another construction due to [37]. The co-radical of an MV-algebra
is the intersection of its maximal non-degenerate lters, and an MV-algebra A is said
to be perfect if for all a A, either a or a is an element of the co-radical of A.
Given a perfect MV-algebra A, its co-radical is the domain of a cancellative subhoop of
A, denoted by (A). Conversely, given a cancellative hoop H, there is a (unique up to
isomorphism) perfect MV-algebra, denoted by
1
(H) whose co-radical is H. Finally,
given an -group G, we denote by (G) its negative cone, which is a cancellative hoop,
418 Manuela Busaniche and Franco Montagna
and conversely, given a cancellative hoop H, we denote by
1
(H) the (unique up to
isomorphism) -group whose negative cone is H.
All these constructions may be expressed in a more general way as equivalences of
categories, but in our treatment we do not need such level of generality.
Now let A, B, C be linearly ordered Wajsberg algebras as in Lemma 6.2.2. Let
(G, u) =
1
(A), (H, v) =
1
(B) and (K, w) =
1
(C). Note that G, H and K
are linearly ordered. Moreover modulo isomorphism we can suppose that u = v = w
and that G = H K. By Lemma 6.2.5 there are a lattice-ordered Abelian group F
and embeddings h and k from H and K respectively into F such that h and k coincide
on the domain of G. Thus in particular h(u) = k(u). After replacing F by its convex
lattice ordered subgroup generated by h(u), we may suppose that h(u) is a strong unit
of F. Then, denoting by (h) ((k) respectively) the restriction of h (k respectively)
to (G, u) ((H, v) respectively), we have that (F, h(u)), (h), (k) is the desired
amalgam, and the claim is proved.
For CH
lin
we can use a similar argument: given cancellative o-hoops A, B, C as
in Lemma 6.2.2, we construct an amalgam D, h, k in LAG of
1
(A),
1
(B),

1
(C) and then we take (D), h
t
, k
t
, where h
t
and h
t
are the restrictions of h and
k to B and to C respectively.
We now discuss WH
lin
. Let A, B, C, be linearly ordered Wajsberg hoops satisfy-
ing the conditions of Lemma 6.2.2. If Ais the reduct of a Wajsberg algebra, then B and
C must also be reducts of Wajsberg algebras (as the image of 0 must be an idempotent
element different from 1), and the claim follows from the AP for W
lin
. Now suppose
that A is a cancellative hoop. If B and C are also cancellative hoops, then the claim
is obvious because CH
lin
has the AP. If one of B and C (or both) is the reduct of a
Wajsberg algebra, let A
t
=
1
(A), and let B
t
= B if B is the reduct of a Wajsberg
algebra, and B
t
=
1
(B) otherwise. Similarly, let C
t
= C if C is the reduct of a
Wajsberg algebra, and C
t
=
1
(C) otherwise. Then A, B and C are (isomorphic
to) subalgebras of (the hoop reducts of) A
t
, B
t
and C
t
respectively. Moreover, up to
isomorphism we may assume that the conditions of Lemma 6.2.2 are satised by A
t
,
B
t
and C
t
. For instance, if B is a Wajsberg algebra, then A is included in the radical
of B, therefore A
t
is a subalgebra of B. Since W
lin
has the AP, there is an amalgam
D, h, k of A
t
, B
t
, C
t
such that D W
lin
. Let h
t
and k
t
be the restrictions of h and
k respectively to B and to C. Since A is a subalgebra of A
t
, D, h
t
, k
t
is the desired
amalgam, and the claim follows.
Finally, we come to P
lin
. It follows from [28] that a linearly ordered product algebra
is the ordinal sum of the two-element hoop L
2
and a cancellative hoop. Thus the claim
follows from the AP for CH
lin
.
THEOREM 6.2.7. B' and BH have the AP.
Proof. We prove the claim for B', because the proof for BH is similar. Once again,
it is sufcient to prove that B'
lin
has the AP. Let A, B and C be BL-chains such
that A = B C. We can write B and C as B =

iI
W
i
, C =

jJ
U
j
, where
W
i
and U
j
are linearly ordered Wajsberg hoops, I and J are linearly ordered sets with
minimum i
0
and j
0
respectively, and W
i
0
and U
j
0
are bounded. Let
I
and
J
denote
Chapter V: H ajeks BL Logic and BL-Algebras 419
the orders of I and of J respectively, and let I(A) = i I [ A W
i
,= 1, and
let J(A) = j J [ A U
j
,= 1. Without loss of generality, we can suppose
that I J = I(A) = J(A) = M. Thus i
0
= j
0
M (as 0 (W
i
0
A) 1
and 0 (U
j
0
A) 1). Now we can represent A, B and C as ordinal sums with
reference to the same index set Y = I J, possibly using trivial summands (ie., hoops
consisting of 1 only). Dene, for a, b Y , a b iff either a, b I and a
I
b, or
a, b J and a
J
b, or a J, b I and there is m M such that a
J
m
I
b, or
a I, b J and there is no m M such that b
J
m
I
a. Moreover let for a Y ,
W
t
a
= W
a
if a I, and let W
t
a
be a trivial hoop otherwise. Similarly, let U
t
a
= U
a
if a J, and let U
t
a
be a trivial hoop otherwise. Finally let V
t
a
= V
a
if a M, and let
V
t
a
be a trivial hoop otherwise. Then we have A =

aY
V
t
a
, B =

aY
W
t
a
and
C =

aY
U
t
a
. Moreover for all a Y , V
t
a
= W
t
a
U
t
a
. Since WH
lin
has the AP,
for a Y , there is an amalgamH
a
, h
a
, k
a
of V
t
a
, W
t
a
, U
t
a
such that H
a
WH
lin
.
Now let H =

aY
H
a
, and let h: B
h
H and k: C
k
H be dened as follows:
h(x) =
_
h
a
(x) if x W
t
a
1
1 if x = 1
k(x) =
_
k
a
(x) if x U
t
a
1
1 if x = 1
It is readily seen that H, h, k is an amalgam of A, B, C.
REMARK 6.2.8. The variety of SBL-algebras has the AP too. To see this, it is sufcient
to prove the claim for V -formations consisting of linearly ordered SBL-algebras. These
are precisely the ordinal sums of the two-element BL-algebra and a BL-chain, and
the proof that such V -formations have an amalgam in the class of SBL-chains is quite
similar to the proof of Theorem 6.2.7.
Theorems 6.2.7 and 6.2.6 tell us that the most interesting varieties of BL-algebras
have the AP. There are however many subvarieties of B' without the AP. For instance,
in [38] it is shown that a variety of MV-algebras (hence a variety of Wajsberg algebras)
has the AP iff it is generated by a single linearly ordered algebra. Thus, for instance the
variety generated by the Chang algebra (
1
(Z)) and by the three-element Wajsberg
algebra does not have the AP. Moreover:
THEOREM 6.2.9. There are uncountably many varieties of BL-algebras which are all
generated by a single BL-chain and do not have the AP.
Proof. Recall that for n N, L
n
is the (unique up to isomorphism) linearly ordered
Wajsberg algebra with n elements. Let X be a set of prime numbers, and let V
X
be the
variety generated by all Wajsberg algebras plus all ordinal sums of the form W
0
W
1
,
where W
1
is any non-trivial Wajsberg chain, and W
0
is a linearly ordered Wajsberg
algebra such that for all p X, L
p+1
does not embed into W
0
. In Theorem 5.0.1
we have seen that if X and Y are different sets of prime numbers, then V
X
,= V
Y
.
Thus there are uncountably many varieties of this form. Moreover, each variety V
X
is generated by the single linearly ordered algebra W
X
[0, 1]

, where W
X
is the
subalgebra of [0, 1]
X
consisting of 0 and 1 plus all rationals of the form
n
m
, where
0 n < m and for every prime p X, p does not divide m. Thus in order to prove
420 Manuela Busaniche and Franco Montagna
the claim it is sufcient to show that if X ,= , then V
X
does not have the AP. Now let
A = L
2
, B = [0, 1]

and C = L
2
[0, 1]

. Clearly A, B and C are in V

X
, and A
is a subalgebra of both B and C. Moreover, if both B and C embed into D, then D
must have a subalgebra isomorphic to [0, 1]

[0, 1]

. Thus in order to prove that V

X
does not have the AP, it is sufcient to show that [0, 1]

[0, 1]

/ V
X
. But from the
proof of Theorem 5.0.1, it follows that V
X lin
consists of all MV-chains plus all ordinal
sum of the form AW, where Ais an MV-chain such that, for p X, L
p+1
does not
embed into A, and W is any Wajsberg o-hoop. Hence, [0, 1]

[0, 1]

/ V
X
, and V
X
does not have the AP.
The above varieties of BL-algebras without the AP are constructed ad hoc. One
may wonder if there are more interesting varieties of BL-algebras without the AP. The
answer is positive. First of all, for n > 1, let G
n
denote the variety generated by the
n-element linearly ordered G odel algebra. As a consequence of a very general result
about AP in superintuitionistic logics, see [49, Theorem 6.39] or [80], we obtain:
THEOREM 6.2.10. For n > 3, G
n
does not have the AP.
We now introduce other examples of varieties of BL-algebras without the AP.
In [27], the authors prove that, even though the logic BL is a common generalization
of ukasiewicz logic , G odel logic G and product logic , the variety of BL -algebras
is not just the join of the corresponding varieties W, G and P. There the authors axiom-
atize the joins of Wand G, of Wand P, of G and P, and of W, G and P. Let us denote
these varieties by WG, WP, GP, and WGP respectively.
THEOREM 6.2.11. Let V be any variety such that V WGP and V contains either
WG or WP. Then V does not have the AP. In particular, none of WG, WP and WGP
has the AP.
Proof. Let V be as in the claim of Theorem 6.2.11. Let A be the two-element BL-
algebra and B = [0, 1]

. Moreover, if V WG, then let C = [0, 1]

G
, otherwise let
C = [0, 1]

. Suppose by contradiction that there is an amalgamD, h, k of A, B, C

with D V. Let b and c be the images of
1
2
under h and k respectively. Then b = b,
whereas c < 1, and c = 0. Now let us decompose D as a subdirect product of an
indexed family of subdirectly irreducible BL-algebras D
i
[ i I. As shown in [27],
every subdirectly irreducible member of WGP is either a Wajsberg algebra or a G odel
algebra or a product algebra. Now take an index i such that c
i
< 1. Since b
i
= b
i
, D
i
needs to be a Wajsberg algebra, because in a product algebra or in a G odel algebra has
no xed point. But then in D
i
we cannot have c
i
< 1 and c
i
= 0, and a contradiction
has been reached.
Another rather natural collection of varieties of BL-algebras without the amalgama-
tion property is the following: for every positive natural number n, let V
n
denote the
variety generated by all ordinal sums of at n Wajsberg hoops at most. We have seen in
Section 3 that V
n
is axiomatized by (the dening equations of B' plus) the equation
n1

i=0
((x
i+1
x
i
) x
i
)
n

i=0
x
i
. (
n
)
Chapter V: H ajeks BL Logic and BL-Algebras 421
Then, we can prove:
THEOREM 6.2.12. For all n > 2, the variety V
n
does not have the AP.
Proof. Let n[0, 1]

be the ordinal sum of n copies of [0, 1]

, denoted by [0, 1]
,1
, . . . ,
[0, 1]
,n
. Let h
i,j
be the unique isomorphism between [0, 1]
,i
and [0, 1]
,j
. Consider
the V -formation (n 1)[0, 1]

, n[0, 1]

, n[0, 1]

, i, j, where i(1) = j(1) = 1, and if

x < 1, then i(x) = x and j(x) = x if x [0, 1]
,i
with i < n1, and j(x) = h
n1,n
(x)
if x [0, 1]
,n1
. Then, the reader is invited to verify that if D, h, k is an amalgam
of (n 1)[0, 1]

, n[0, 1]

, n[0, 1]

, i, j, then D must contain (n + 1)[0, 1]

, which
is not in V
n
.
Interestingly, there are varieties of BL-algebras which are not generated by a single
linearly ordered BL-algebra and have the AP. One example is given by GP. Since every
linearly ordered element of GP is either a G odel algebra or a product algebra, GP is not
generated by a single linearly ordered algebra. Moreover:
THEOREM 6.2.13. GP has the AP.
Proof. By the remark just following Lemma 6.2.4, it is sufcient to prove that given A,
B and C in GP
lin
such that A = B C, there is an amalgam D, h, k of A, B, C
such that D GP (possibly D / GP
lin
). Now any algebra in GP
lin
is either a G odel
algebra or a product algebra. If B and C are either both product algebras or both G odel
algebras, the claim follows from the fact that both G and P have the AP. If, say B is a
G odel algebra and C is a product algebra, then Amust be the two-element BL-algebra.
Indeed, in a G odel algebra every element is idempotent, whereas in a linearly ordered
product algebra the only idempotents are 0 and 1. Dene:
D = B C, h(x) =
_
(x, 1) if x > 0
(x, 0) otherwise
k(y) =
_
(1, y) if y > 0
(0, y) otherwise
It is readily seen that D, h, k is the desired amalgam, and that D GP.
REMARKS 6.2.14.
(1) Let V
1
and V
2
be varieties of SBL-algebras such that every chain in V
1
V
2
is isomorphic to the two-element BL-chain. Then, the join of V
1
and V
2
has the
AP. (The argument is similar to the one used in the proof of Theorem 6.2.13).
(2) We say that a variety V has the joint embeddability property if for any two mem-
bers A, B of V there is an algebra C V in which A and B embed. It is easy
to verify that if a variety has APand it has a model which embeds into every al-
gebra of the variety, then the variety has the joint embeddability property. Hence,
W, WH, CH, G, P, B' and BH have the joint embeddability property.
422 Manuela Busaniche and Franco Montagna
6.3 Deductive interpolation and Beths property, and strong amalgamation
In [87], see also [50, Corollary 5.41], various connections between amalgamation
and interpolation have been established.
THEOREM 6.3.1 ([50]). If L is an algebraizable logic whose equivalent algebraic se-
mantics is a variety V of commutative (and possibly integral and bounded) residuated
lattices, then V has the APiff L has the deductive interpolation property.
Now let G
n
denote the n-valued G odel logic, and let G, G, and L(X), X
any set of prime numbers, denote the logics whose equivalent algebraic semantics is
GP, WG, WP and V
X
, respectively, (see the proof of Theorem 6.2.9 for the denition
of V
X
). Then:
THEOREM 6.3.2. (1) The logics G, , , G, BL, SBL, G
2
and G
3
have deductive
interpolation.
(2) The logics G
n
(with n > 3), G, , and L(X) (X ,= ) do not have deductive
interpolation.
EXAMPLES 6.3.3.
1. Although the formulas p (p q) and r (r q) do not have any Craig
interpolant, they do have a deductive interpolant, that is, q.
2. The reader may want to verify directly that G
4
has no interpolation. Let p
q = (p q) ((q p) p). An easy computation shows that for every n,
in the G odel chain G
n
with n elements one has p q = 1 if either p < q or
p = q = 1, and p q = p q otherwise. Now in G
4
we have p q, q r [=
(r s) (r s). Finally, let 0 = a < b < c < d = 1 be the elements of G
4
,
and, by way of contradiction, let (r) be a deductive interpolant of p q, q r
and (r s) (r s). If p = 0, q = b and r = c, then (r) must be 1.
Moreover, by induction on (r), we can prove that if (r) = 1 for r = c, then
(r) = 1 for r = b. Now let r = b and s = c. Then, (r) = 1, r s = 1, but
r s = c < 1, a contradiction.
3. Since G
3
has the interpolation property, the formulas in the previous example have
an interpolant in G
3
. The reader is invited to nd one.
We now investigate the deductive Beth property. First of all, every superintuitionistic
logic has Beths property, see [76] or [49, Theorem 4.38]. (In the above mentioned
theorem, Beths property is formulated in terms of implication, but in G odel logics the
implicative formulation is equivalent to the deductive one, because the deduction theo-
rem holds in any extension of G.) Therefore:
THEOREM 6.3.4. Any schematic extension of G odel logic G has both the deductive
and the implicative Beth property.
However:
Chapter V: H ajeks BL Logic and BL-Algebras 423
THEOREM 6.3.5. If L is an axiomatic extension of BL which is not an extension of G,
then L does not have the deductive Beth property. Thus, an axiomatic extension of BL
has the deductive Beth property iff it extends G.
Proof. Let V be the variety associated to L. Then V is a variety of BL-algebras which
is not contained in G, and therefore there is a linearly ordered algebra A V which is
the ordinal sum of an indexed family of Wajsberg components, one of which, W say,
has more than two elements (otherwise every element of V would be a G odel algebra).
Taking the subalgebra of Agenerated by the rst Wajsberg component of Aand by W,
we can assume that either A = W or A = H W for some bounded Wajsberg hoop
H. Hence, either W V or L
2
W V. If W is cancellative, then A = H W,
and V contains an innite product chain (and hence it contains all product algebras).
The same conclusion holds if A = H W and W is a perfect Wajsberg chain. If
A = W where W is a perfect and innite Wajsberg chain, then V contains Changs
algebra (Z
lex
Z, 1, 0), where Z is the ordered group of integers and
lex
denotes
the lexicographic product. Otherwise, since W has more than two elements, it contains
a nite Wajsberg chain with more than two elements, and V contains either an element
of the form L
2
L
n
, or an element of the form L
n
, with n > 2.
Hence, we have to distinguish the following cases:
(a) V contains a BL-chain of the form L
2
L
n
or of the form L
n
, for some n > 2.
(b) V contains [0, 1]

.
(c) V contains Changs algebra.
Now consider, for every n, the formula
n
(a, p) = p (p
n2
a). We rst prove
that for every n > 2,
n
(a, p),
n
(a, q)
L
p q. Since any variety of BL -algebras
is generated as a quasivariety by its linearly ordered members ([1, 58]), it is sufcient to
check that for every linearly ordered algebra B Vand for every evaluation i in Bsuch
that i(
n
(a, p)) = i(
n
(a, q)) = 1, one has i(p) = i(q). Suppose i(
n
(a, p)) = 1. If
i(a) = 1, then i(p) = 1. Now assume that i(a) ,= 1; then i(p) > i(a) (otherwise
i(p) = i(p
n2
a) = 1, and i(a) = 1). Moreover, i(a) and i(p) are in the same
Wajsberg component, otherwise we would obtain i(p
n2
a) = i(a), and nally
i(p) = i(a) = 1. Consider the function F(i(p)) = i(p
n2
a) as a function of
i(p) (for i(a) < 1 xed), when i(p) ranges over the Wajsberg component Z which
i(a) belongs to and i(p) i(a). If i(p) = i(a), then F(i(p)) > i(p), and F(i(p)) is
decreasing in [a, sup(Z)). Thus there can be at most one value i(p) in Z (and at most
one value of i(p) in A) such that F(i(p)) = i(p). Hence, i(p) is uniquely determined
from i(a). Therefore, if i(
n
(a, p)) = i(
n
(a, q)) = 1, then i(p) = i(q).
Now we prove that in case (a) there is no formula (a) such that
n
(a, p)
L
p
(a). If L
2
L
n
V, then let e(a) = min(L
n
), and if L
n
V, then let e(a) = 0.
Note that e(
n
(a, p)) = 1 iff e(p) is the unique coatom of L
n
. Then, for any formula
(a) in the variable a only, e((a)) is in the subalgebra of L
n
(of L
2
L
n
respectively)
generated by e(a), which in the rst case consists of 0 and 1, and in the second vase
consists of e(a), 0 and 1. Thus e(p) ,= e((a)).
424 Manuela Busaniche and Franco Montagna
Nowwe prove that in cases (b) and (c) there is no formula (a) such that
3
(a, p)
L
p (a). In case (b), let e(a) =
1
2
. Then, e(
3
(a, p)) = 1 iff e(p) =
_
1
2
. Moreover
for any formula (a) in the variable a only, e((a)) is in the subalgebra of [0, 1]

gener-
ated by e(a), which consists of 0, 1 and of all elements of the form (
1
2
)
n
, n N. Hence,
e(p) =
_
1
2
,= e((a)).
Finally, in case (c), let e(a) = 1, 2. Then, e(
3
(a, p)) = 1, 0 iff e(p) =
1, 1, and for every formula (a), e((a)) belongs to the subalgebra of (Z
lex
Z,
1, 0) generated by e(a), which in turn consists of all elements of the form 1, 2n
or 0, 2n with n N. Hence, there cannot be a formula (a) with e((a)) = e(p) =
1, 1.
Beths property is also related to strong amalgamation, a strengthening of the AP
which is dened below. In the rest of this section we will obtain a characterization of
the varieties of BL-algebras having the strong amalgamation property.
DEFINITION 6.3.6. A class K of algebras has the Strong Amalgamation Property
(SAP for short) iff whenever A, B and C are in K and i and j are embeddings from A
into Band fromAinto C respectively, there is an amalgamD, h, k of A, B, C, i, j
such that D K, and in addition h(B) k(C) = (h i)(A) = (k j)(A). In this case,
D, h, k is said to be a strong amalgam of A, B, C, i, j.
Then in [72, Theorem 50], it is shown that if a variety V of commutative (possi-
bly integral, possibly bounded) residuated lattices is the equivalent algebraic seman-
tics of an algebraizable logic L and if V has the SAP, then L has the deductive Beth
property. Moreover, in [49, Theorem 6.1], (see also [80]), it is shown that a variety of
pseudoboolean algebras has the SAP iff it has the AP. Now in [80] (see also [49, The-
orem 6.39]), a characterization is given of the (nitely many!) varieties of Heyting
algebras with the SAP. An immediate consequence of this characterization is the fol-
lowing:
THEOREM 6.3.7. A variety V of G odel-algebras has the SAP iff it is either the trivial
variety, or G, or G
2
or G
3
.
From Theorems 6.3.7 and 6.3.5, we deduce:
THEOREM 6.3.8. A variety V of BL-algebras has the SAP iff it is either the trivial
variety, or G, or G
2
or G
3
.
6.4 Craig interpolation and implicative Beth property
Due to the presence of contraction and weakening, in G odel Logic Craig interpola-
tion is equivalent to deductive interpolation, therefore G has interpolation of both kinds.
As shown in [8], Galso has uniform interpolation. However, very few many-valued log-
ics have Craig interpolation: as shown in [74], the only extensions of ukasiewicz logic
with Craig interpolation are classical logic and the inconsistent logic. Moreover, in [8]
it is shown that, with the exception of G odel logic, no t-norm logic, i.e., no many-valued
logic which is complete with respect to a BL-algebra over [0, 1], has Craig interpolation.
A slight modication of the proof of this result gives:
Chapter V: H ajeks BL Logic and BL-Algebras 425
THEOREM 6.4.1. The only consistent schematic extensions of BL having Craigs in-
terpolation are G, G
3
, and classical logic.
Proof. By [49, Theorem 6.42], the only consistent extensions of G having Craig inter-
polation are G, G
3
and classical logic. Thus it is sufcient to prove that if a schematic
extension L of BL is not an extension of G, then it does not have Craig interpolation. To
this purpose, let (p, q) = (p (p q)), and (q, r) = r (r q). As noted in [8],

BL
(p, q) (q, r). In order to conclude the proof, it is sufcient to prove that if
the equivalent algebraic semantics of a schematic extension L of BL is a variety V of
BL-algebras which is not included in G, then (p, q) and (q, r) have no interpolant in
L. First note that under the above assumptions, V contains a BL-chain Awhich has the
form

iI
W
i
, where each W
i
is a linearly ordered Wajsberg hoop, and (at least) one
of them, W say, has more than two elements. Distinguish the following cases:
(a) W is a Wajsberg algebra. Let e be an evaluation such that e(q) = min(W)
and e(p) = e(r) W e(q), 1 (this is possible because W has more than two
elements). Then e(q) < e((p, q)) e((q, r)) < 1. Let (q) be a formula in the
variable q only. Then e((q)) belongs to the subalgebra of A generated by 0, 1 and
e(q). Since e(q) is an idempotent, such a subalgebra only consists of 0, 1 and e(q). If
e((q)) = 1, then e((q)) > e((q, r)); if either e((q)) = 0 or e((q)) = e(q), then
e((q)) < e((p, q)). So (q) cannot be an interpolant of (p, q) and (q, r).
(b) W is a cancellative hoop. Let a W1, and let e be any evaluation such that
e(p) = e(r) = a, e(q) = a
2
. Then e((p, q)) = e((q, r)) = a. Now let (q) be any
formula in the variable q only. Then, e((q)) belongs to the subalgebra of A generated
by 0, 1 and e(q) = a
2
. Such subalgebra consists of 0, 1 and all elements of the form
a
2n
, n N, and a is not an element of that subalgebra. Thus (q) is not an interpolant
of (p, q) and (q, r).
The negative result contained in Theorems 6.3.5 and 6.4.1 extend to the implicative
Beth property:
THEOREM 6.4.2. Let L be any schematic extension of BL. Then L has the implicative
Beth property iff L extends G.
Proof. Since every superintuitionistic logic has the implicative Beth property [76], it is
sufcient to prove that if L is a schematic extension of BL whose equivalent variety V is
not contained in G, then L does not have the implicative Beth property. If V , G, then
there is a linearly ordered algebra A V such that one of the Wajsberg components W
of A has more than two elements. As in the proof of Theorem 6.3.5 we may assume
that W is either a perfect Wajsberg algebra or a cancellative hoop, or a nite Wajsberg
algebra. In the rst two cases, let (p, q) = q (q p); in the third case let
(p, q) = q (q
n1
p), where n + 1 > 2 is the cardinality of W. As in the proof
of Theorem 6.3.5, we may nd an algebra A V and an evaluation e in A such that
e((p, q)) = 1 and e(q (p)) ,= 1 for every formula (p) in the variable p only. Thus
there is no formula (p) in the variable p only, such that (p, q)
L
q (p). In order
to conclude that L does not have the implicative Beth property, it is sufcient to prove
that
BL
((p, q) & (p, r)) (q r).
426 Manuela Busaniche and Franco Montagna
Let us write (p, q) as q (q
i
p), where i can be either 1 or n1. By symmetry
reasons, it is sufcient to prove that

BL
((p, q) & (p, r)) (q r),
which by the residuation property amounts to prove that

BL
((p, q) & (p, r) & q) r.
Let (p, q, r) = (p, q)&(p, r)&q. We prove that for any linearly ordered BL-algebra
A and for any evaluation e in A, one has e((p, q, r)) e(r). This clearly implies the
claim.
Since e((p, q, r)) e(q), if e(q) e(r), the claim is trivial. If e(r) < e(q), then
e(q
i
p) e(r
i
p), and since e((p, q) & q) e((q
i
p), we get:
e((p, q, r)) e((q
i
p) & (p, r)) e((r
i
p) & (r (r
i
p))) e(r),
and the claim is proved.
EXAMPLES 6.4.3.
1. Let (p, q) = q (q p). From the proof of Theorem 6.4.2, we know that
(p, q) implicitly denes q in BL. Since G has both the deductive Beth property
and the implicative Beth property, (p, q) must explicitly dene q. Now it is
readily seen that
G
(p, q) q, and hence letting (q) = we have (p, q)
G
q (p). However, does not work if we reason in an extension of BL which
does not extend G.
2. We have seen that the formulas p (p q) and r (r q) do not have a Craig
interpolant in a logic L which extends BL but does not extend G. However, q is a
Craig interpolant of the above formulas in G.
3. Let L be an extension of BL. Then, the formulas p p and q q have an
interpolant in L iff L extends SBL. Indeed, if L extends SBL, then is a Craig
interpolant. Otherwise, an interpolant should not contain variables, and hence it
should be provably equivalent either to or to . But if L does not extend SBL,
then -
L
(q q) and -
L
(p p) , and hence the above formulas do
not have a Craig interpolant.
7 Completions of BL-algebras
A lattice C is a completion of a lattice L if C is a complete lattice and L is a
sublattice of C. By a complete BL-algebra we understand a BL-algebra Awhose lattice
reduct L(A) is complete. A subvariety V of BL-algebras is said to admit completions
if for every algebra A V there is a complete BL-algebra B V and an embedding
: A B. We usually identify Awith (A).
Every lattice admits a completion. One of the main problems, when dealing with
algebras with a lattice reduct, is how to extend the extra operations of the algebra to the
Chapter V: H ajeks BL Logic and BL-Algebras 427
extended lattice. Of course the goal is to nd a completion such that the extension of
an algebra remains in the variety of the original one. Sometimes there are more than
one way to extend the extra operations, and choosing the right one requires a careful
investigation.
Some completeness theorems of propositional and predicate logics are obtained
from results of completions. The two most studied methods for completing lattice-
based algebras are canonical extensions and MacNeille (or Dedekind-MacNeille) com-
pletions. The MacNeille completions provide completeness of various predicate logics
with respect to their algebraic semantics, and canonical extensions are the keys for many
completeness results in non-classical propositional logics.
Canonical extensions are based on the Stone representation for Boolean algebras:
the canonical extension of a Boolean algebra A is the embedding of A into the power
set of its Stone space. This construction, introduced by J onsson and Tarski in [70, 71],
was then generalized for distributive lattices, lattices and posets with different internal
operations in [5254] and [41]. On the other hand, MacNeille completions are a gen-
eralization of Dedekinds construction of the reals by cuts of rationals (see [34]). The
MacNeille completion of a poset P is the minimum complete lattice that extends P.
While canonical extensions have the advantage of being atomic, MacNeille completions
have the advantage that existing joins and meets are preserved.
In the present section we will study completions of BL-algebras. Our rst goal is to
investigate which subvarieties of BL-algebras admit completions. Secondly we analyze
canonical extensions of BL-algebras.
7.1 Negative results about completions
In [75] completions of some subvarieties of BL-algebras are studied within the
framework of GBL-algebras (see 8.2.1 for denition). The authors exhibit some condi-
tions under which a GBL-algebra can not be completed.
The proofs presented next are based on the general method developed in [75] to
decide if a GBL-algebra can be completed.
We recall the denition of Changs MV-algebra C = (Z
lex
Z, 1, 0). For
simplicity, one can consider the universe of C to be the set C = a
n

nN
b
n

nN
with a total order given by 0 = b
0
< b
1
< b
2
. . . < . . . < a
2
< a
1
< a
0
= 1, i.e.:
_

_
b
n
< a
m
for all n, m N,
b
n
< b
m
if n < m,
a
n
< a
m
if n > m.
C is the algebra C, , , 0, 1, where a
0
= 1, b
0
= 0 and the product is given by
b
n
b
m
= 0, a
n
a
m
= a
n+m
, a
n
b
m
=
_
b
mn
if m n,
0 otherwise.
Observe that is uniquely determined as
b
n
a
m
= 1, a
n
b
m
= b
n+m
428 Manuela Busaniche and Franco Montagna
a
n
a
m
=
_
a
mn
if n m,
1 otherwise
b
n
b
m
=
_
a
nm
if n m,
1 otherwise.
We shall show that any complete extension of this algebra does not satisfy the divis-
ibility condition x (x y) = y (y x) = x y.
THEOREM 7.1.1. The BL-algebra C does not admit a completion within the variety
of BL-algebras.
Proof. Assume (absurdum hypothesis) that there is a complete BL-algebra D and an
embedding : C D. As usual, we identify (C) with C. Let A = a
n

nN
and
B = b
n

nN
. Then a =
_
A and b =
_
B are elements of D C. Observe that
_
B b
1
=
_
b
n
b
1
[ n N = a. Since b
n
a b
n
a
n
b
0
< b
1
for all n N
we have
b (b b
1
) =

B (

B b
1
) =

B a =

b
n
a [ n N < b
1
.
Then the divisibility condition is not satised since b (b b
1
) ,= b
1
= b b
1
.
COROLLARY 7.1.2. Let V be a variety of BL-algebras such that C V. Then V does
not admit completions.
Komoris classication of subvarieties of MV-algebras shows that any subvariety
of MV-algebras that is not nitely generated contains Changs MV-algebra C (see [26,
Theorem 8.4.4].) This fact together with the previous corollary yield:
THEOREM 7.1.3. Let S be a variety of MV-algebras that is not nitely generated and
let V B' be such that S V. Then V does not admit completions.
In [75], the authors also exhibit a product algebra P such that any complete exten-
sion of P does not satisfy divisibility. To make the chapter self-contained we sketch
the idea of the construction of this product chain presented in [75]. Let (C) be the
co-radical of Changs MV-algebra C. Since C is a perfect MV-algebra, we know that
(C) is a cancellative hoop. Thus P = L
2
(C) is a BL-chain, which is a product
chain. If we denote by P = 0, a
n

nN
the universe of P, then it is easy to see that
P is a subdirectly irreducible product algebra. But unlike C, the lattice reduct of P is a
complete lattice. However, for a nonprincipal ultralter U, the ultrapower R = P
I
/U
is not embeddable in any complete BL-algebra. To prove this fact we have to check
some properties of the BL-chain R. They are:
1. There exists a R such that a < 1 and for all x R, if a x < 1 then a = x.
If we denote by A = a
n

nN
, then the hoop A is isomorphic to the cancellative
hoop (C), with isomorphism given by a
n
a
n
. We identify A with a
n

nN
.
2. For every 0 < x R 1 there is x
t
R such that for all y R, if x y < x
t
then x = y. We call x
t
the successor of x in R. Observe that for each k N, a
k
is the k-th predecessor of 1.
3. For every x, y R, if x > 0 and y is the k-th successor of x, then y x = a
k
.
Chapter V: H ajeks BL Logic and BL-Algebras 429
4. For every x, y R, if x > 0 and y is the k-th successor of x, then a
k+1
y < x.
5. There exist c, b in R A such that 0 c < b and b is the successor of c.
The reader can check that most of these properties are satised by the original chain P,
and since they are rst order expressible they are inherited by the ultrapower R. Now
take b = b
1
and let b
k
be the k-th successor of c (the (k 1)-th successor of b
1
).
Consider the totally ordered set B = b
k

kN
, that has no supremum in R. Observe
that if e A B, then e is the j-th successor of c and the k-th predecessor of 1. Thus
c R A is the (k + j)-th predecessor of 1. The contradiction implies A B = .
Suppose that Rcan be embedded into a complete BL-chain S by an embedding ,
and as usual, identify (R) with R. Let e =
_
B and d =
_
A. Clearly e and d are
elements of S, since S is complete. As in the case of Changs MV-algebra completion
we have:
_
B b
1
=
_
b
n
b
1
[ n N = d. Since for each b
k
there is a
k+1
such
that b
k
a
k+1
< b
1
we have b
k
d b
k
a
k+1
c < b
1
for all k N. Therefore
e (e b
1
) =

B (

B b
1
) =

B d =

b
k
d [ k N < b
1
.
Then the divisibility condition is not satised since e (e b
1
) ,= b
1
= e b
1
.
The existence of this product algebra, together with the fact that the only non-trivial
subvariety of product algebras is the variety of Boolean algebras yield:
THEOREM 7.1.4. Every variety of BL-algebras V that contains the variety of product
algebras does not admit completions.
7.2 k-subvarieties and classication
Given a natural number k 1, a subvariety V of BL-algebras is called a k-sub-
variety if all algebras in V are k-potent algebras, i.e., if they satisfy the equation:
x
k
= x
k+1
. (14)
Recall that an MV-chain Ais k-potent if and only if A

= L
n
for some k +1 n.
We say that a BL-chain Ais a k-chain if it is k-potent, i.e., if
A

iI
L
r
i
where I is a lower bounded totally ordered set and 2 r
i
k + 1 for each i I.
REMARK 7.2.1. Observe that every nite BL-chain is a k-chain for some k and if V is
a k-subvariety of BL-algebras for some k 1, then every BL-chain in V is a k-chain.
It is also easy to prove that every nitely generated subalgebra of a k-chain is nite.
We will prove the following classication Theorem:
THEOREM 7.2.2. Let V be a subvariety of BL-algebras. Then one and only one of the
following happens:
1. There is k 1 such that V is a k-subvariety.
430 Manuela Busaniche and Franco Montagna
2. V contains the variety of product algebras or Changs MV-algebra C is in V.
Proof. Assume that V is a k-subvariety. Since Changs MV-algebra C is not k-potent
for any k, then C / V. The standard product algebra [0, 1]

is not k-potent for any k.

Since this algebra generates the variety of product algebras, V cant contain the variety
of product algebras.
Assume that V is not a k-subvariety for any k 1. We distinguish two cases:
1. For every k, V does not satisfy the equation (x)
k+1
= (x)
k
. Observe that
for any element x in a BL-chain B, the element x is in the rst Wajsberg
component of B. Then for every k, there is a BL-chain B
k
V such that the
rst Wajsberg component W
k
of B
k
is not a k-potent MV-chain. Obviously
each W
k
V. If there is k such that W
k
V is an innite MV-chain, then
Changs MV-algebra C is in V, since C is in every non-nitely generated subva-
riety of MV-algebras (see [26, Theorem 8.4.4]). Otherwise, W
k
, k N is an
innite set of non-isomorphic nite MV-chains in V. By [26, Proposition 8.1.2],
the variety generated by this set is the variety of all MV-algebras. Thus Changs
MV-algebra C is in V.
2. There is n 1 such that the equation (x)
n+1
= (x)
n
holds in V. In this
case, we follow a proof given in [10]. Our assumption implies that if B is a BL-
chain in V, its rst Wajsberg component is isomorphic to L
s
for some s n +1.
Since V is not a k-subvariety, for every natural number k there is a BL-chain B
k
and an element b
k
B
k
such that b
k+1
k
< b
k
k
and b
k
is not in the rst component
of B
k
. Now take an ultraproduct B of all B
k
modulo a non-principal ultralter.
From the ultraproduct theorem, B is a BL-chain and there is an element b B
such that b
k+1
< b
k
for every k. Clearly, for all k, b
k
is in the same Wajsberg
component W of B. Consider the subalgebra A of B generated by b. The set
b
k
[ k N1 Ais a subreduct of B isomorphic to the negative cone of Z.
Hence Ais an innite product chain. Since every innite product chain generates
the variety of product algebras, this variety is contained in V.
7.3 Completions in k-subvarieties
A subset L of a lattice C is said to be join-dense in C if every element of C is the
join in C of elements of L. Dually one can dene meet-dense. A completion C of a
lattice L is called a MacNeille completion of L if L is both, join- and meet-dense in C.
Every bounded lattice L has a MacNeille completion L, which is unique up to
isomorphisms. The idea of the construction is the following: for any subset A of L, let
A
u
be the set of upper bounds of A and A
l
the set of lower bounds of A. A cut is a pair
A, B of subsets of L such that A = B
l
and B = A
u
. If we order the set of cuts by
A, B A
t
, B
t
iff A A
t
iff B
t
B,
then we obtain a complete lattice L. The map
x x, x = a L [ a x, a L [ a x
Chapter V: H ajeks BL Logic and BL-Algebras 431
from L into L is a lattice embedding and the image of L is both meet and join-dense
in L. For details see [34].
An important result that will help us deal with completions of BL-algebras is the
next one:
LEMMA 7.3.1 ([89]). Direct products interact well with MacNeille completions, i.e., if
L
i
[ i I is a family of lattices, then

iI
L
i

=

iI
L
i
.
Let I be a lower bounded totally ordered set and let 0 be its bottom element. It
is easy to see that the MacNeille completion I of I is a lower bounded totally ordered
set and 0 is its lower bound. Moreover, if i I I and i
1
, i
2
I are such that
i
1
< i < i
2
, then the join and meet density of I in I yields elements j
1
, j
2
I such that
i
1
< j
1
< i < j
2
< i
2
. Therefore we obtain the following result.
LEMMA 7.3.2. Let I be a lower bounded totally ordered set and let I be its MacNeille
completion. Let i
1
< i
2
< . . . < i
n
be a nite subchain of I. Then there exists a nite
subchain j
1
< j
2
< . . . < j
n
of I such that if i
k
I then j
k
= i
k
.
We will be concerned with k-chains of the formA =

iI
L
r
i
. Consider the index
set I of Aand let I be its MacNeille completion. For each i I, let
s
i
=
_
r
i
if i I,
2 if i I I.
We dene the BL-chain
C(A) =

iI
L
s
i
.
Observe that C(A) is also a k-chain and Ais a subalgebra of C(A).
THEOREM 7.3.3. The MacNeille completion of the lattice reduct of a k-chain Ais the
lattice reduct of C(A).
Proof. We denote by A and C(A) the corresponding lattice reducts of A and C(A).
Using the notation previously stated,
A =

iI
L
r
i
and C(A) =

iI
L
s
i
.
Let : C(A) 1 I be given by (a) = i if a L
s
i
1. Clearly is a well-
dened function that preserves the order of C(A), i.e., if a, b C(A) 1 and a b,
then (a) (b). Moreover, if (a) < (b) for some a, b C(A) 1, then a < b.
Using this function we check that C(A) is a complete lattice and that A is join-dense
and meet-dense in C(A).
432 Manuela Busaniche and Franco Montagna
C(A) is a complete lattice: Let B C(A); we can assume that 1 / B. Consider the set
(B) = i I [ i = (a) for some a B
and j =
_
I
(B), that exists because I is a complete chain.
If j (B), then B (L
s
j
1) ,= . Take b to be the maximum element in
B (L
s
j
1). Since (b) (a) for every a B, we conclude that b a for
every a B. Thus b =
_
C(A)
B. On the other hand, if j / (B), we claim that
0
j
=
_
C(A)
B, where 0
j
is the least element of L
s
j
. Clearly 0
j
> a for every a B,
because (a) < j. If there is b C(A) such that b a for all a B, then (b) j.
Thus b 0
j
. Similarly one can prove that
_
C(A)
B exists in C(A).
A is join-dense in C(A): Take c C(A) A. Then c = 0
j
for some j I I. Since I
is the MacNeille completion of I, there is a subset J of I such that j =
_
I
J. Consider
the set B A given by
B = 0
i
[ i J.
We will see that 0
j
=
_
C(A)
B. Indeed, 0
j
0
i
for each i J. Take b C(A) such
that b 0
i
for each i J. Then (b) i for each i J, thus (b) j. Then b 0
j
and we are done.
The proof that A is meet-dense in C(A) is dual to the previous one.
We have proved that every k-chain admits a completion within the variety of BL-
algebras, and that the completion is also a k-chain. We intend to see that every algebra
in a k-subvariety admits a completion, and that the completion is in the k-subvariety.
We present some results that will lead us to that conclusion.
THEOREM 7.3.4. Let Abe a k-chain for some k N. Then every nite subalgebra of
C(A) is isomorphic to a nite subalgebra of A.
Proof. According to Proposition 4.1.1, if B is a nite subalgebra of C(A) there is a
nite subset J = 0 = j
0
< j
1
< . . . < j
n
I, such that
B =

jJ
L
t
j
and for each j J, L
t
j
is a subalgebra (subhoop) of L
s
j
. Notice that if j I I then
the algebra L
s
j
is the chain L
2
. Then we can assert
L
t
j
=
_
is a subalgebra of L
s
j
if j I,
L
2
if j I I.
From Lemma 7.3.2, we get a nite subset K = 0 = k
0
< k
1
< . . . < k
n
I
such that if j
w
I then k
w
= j
w
. For each k K take
L
u
k
=
_
L
t
j
if k = j for some j J
L
2
if k ,= j for any j J.
Chapter V: H ajeks BL Logic and BL-Algebras 433
Since L
2
is a subalgebra of L
m
for every integer m 2 we get that
B
t
=

kK
L
u
k
is a subalgebra of A. Clearly B

= B
t
, thus B is isomorphic to a subalgebra of A.
THEOREM 7.3.5. For any k-chain A, the BL-chain C(A) is in the subvariety of BL-
algebras generated by A.
Proof. Assume that is an equation in the language of BL-algebras which is not satis-
ed by C(A). Since C(A) is a locally nite algebra, we can assert that is not satised
by a nite subalgebra B of C(A). B is isomorphic to a subalgebra of A, hence is not
satised by A. We conclude that every equation satised by Ais satised by C(A) and
the result of the theorem follows.
We have developed the tools to prove the main theorem of the section.
THEOREM 7.3.6. Let V be a k-subvariety. Then V admits completions.
Proof. For each chain B in V, consider the injective morphism
B
: B C(B) that
exists, since B is a k-chain. The BL-algebra A V is a subdirect product of chains.
Then there is an injective morphism
: A

iI
B
i
where for each i I, B
i
is a k-chain in V. The morphism:

iI
B
i

iI
C(B
i
)
given by (a)
i
=
B
i
(a
i
), is injective. Therefore, the composition is an injective
morphism fromAinto

C(B
i
). Lemma 7.3.1 together with Theorem 7.3.5, yield that
the algebra

C(B
i
) is a complete BL-algebra in V. Thus every algebra in V admits a
completion.
REMARK 7.3.7. Observe that with the proof of the previous theorem one can not es-
tablish a relation among the lattice reduct of the completion of a BL-algebra Aand the
MacNeille completion of the lattice reduct of A.
THEOREM 7.3.8. A subvariety V of BL-algebras admits completions if and only if
there is k 1 such that V is a k-subvariety.
Proof. We have proved in Theorem 7.3.6 that every k-subvariety admits completions.
We have also seen that no subvariety of BL-algebras containing the variety of product
algebras admits completions and that no subvariety of BL-algebras containing Changs
MV-algebra C admits completions. Then the result follows from Theorem 7.2.2.
434 Manuela Busaniche and Franco Montagna
7.4 Canonical extensions of BL-algebras
Canonical extensions of subvarieties of BL-algebras are studied in [21, 23]. We will
summarize the most important results of those papers in the present section.
Let L
t
be a completion of a lattice L. An element of L
t
is called closed if it is the
meet in L
t
of elements of L. Dually, if an element of L
t
is the join of elements of L, it is
called open. We denote by K(L
t
) and O(L
t
) the sets of closed and open elements of L
t
,
respectively.
The completion L
t
is called dense if each element of L
t
is both: the join of the closed
elements below it and the meet of the open elements above it. L
t
is called compact if
given D, U L, such that
_
L
D
_
L
U, then there exist nite sets D
t
D and
U
t
U such that
_
L
D
t

_
L
U
t
. In such a case, L = K(L
t
) O(L
t
). If a
completion L
t
of L is dense and compact it is called a canonical extension of L.
Every lattice L has a canonical extension L

which is unique up to isomorphism.

The construction of the canonical extension of a lattice is based on the notion of Galois
connection of lattices (see [11]). Recall that a pair of maps f : L L
t
and g : L
t
L
between lattices L and L
t
is a Galois connection if f and g are order inverting and
if x f(g(x)) and y g(f(y)) for all x L
t
, y L. We sketch the idea of the
construction of the canonical extension following [52]: let L be a bounded lattice. Let
T be the set of all lters of L and let J be the set of all ideals of L. The maps from
the power set of T into the power set of J and from the power set of J into the power
set of T given by
(X) = I [ I F ,= for each F X
(Y ) = F [ I F ,= for each I Y
form a Galois connection. Let G(, ) be the set of elements of TT that are in the
image of and G(, ) the set of elements of TJ that are in the image of . Then
G(, ) and G(, ) are complete lattices and the restrictions of to G(, ) and of
to G(, ) are mutually inverse dual order isomorphisms. Let : L G(, ) be
given by
(x) = F T [ x F.
Then is a lattice embedding. It is left to verify that G(, ) is a compact and dense
complete extension of L.
REMARK 7.4.1. The MacNeille completion of a lattice can be constructed using ideas
similar to the previous construction, but instead of considering the sets of all lters and
all ideals in the lattice, one restricts to the sets of principal lters and principal ideals
in the lattice.
The MacNeille completion of a totally ordered set does not necessary coincide with
its canonical extension. For example, consider Changs MV-algebra C with universe
C = a
n

nN
b
n

nN
and the total order given by 0 = b
0
< b
1
< b
2
. . . <
. . . < a
2
< a
1
< a
0
= 1. Then the MacNeille completion of the lattice reduct of C
is the lattice C
t
that arises by adding an extra element a =
_
a
n

nN
=
_
b
n

nN
to C, because this is the minimum complete lattice that extends C. Since compact-
ness is not satised by this extension C
t
, to obtain the canonical extension C

of C
Chapter V: H ajeks BL Logic and BL-Algebras 435
one needs to add two different elements a =
_
a
n

nN
and b =
_
b
n

nN
to C.
Another example is given by the real interval [0, 1] with its usual order. Since it is
a bounded complete lattice its MacNeille completion coincides with the original lat-
tice, but the canonical extension is the complete lattice L whose universe is the set
[0, 1] 1, 0, 1 (0, 1), (1, 1) with the lexicographic order in the product. More
precisely, the embedding f : [0, 1] L given by f(x) = (x, 0) has an image order-
isomorphic to [0, 1], and the canonical extensions of this image is L (see [23] for a
detailed proof).
Though the canonical extension and the MacNeille completion may not coincide,
an analogous of Lemma 7.3.2 holds for canonical completions.
LEMMA 7.4.2. Let I be a totally ordered set and let I

be its canonical completion.

For each nite subchain i
1
< i
2
< . . . < i
n
of I

, there is a nite subchain j

1
< j
2
<
. . . < j
n
of I such that if i
k
I, then j
k
= i
k
.
Proof. If i
k
I

I, then
i
k
=

c K(I

) [ c i
k
.
Since I

is a totally ordered set, there exists c K(I

) such that i
k1
< c i
k
<
i
k+1
. Similarly, there exists d O(I

) with i
k1
< c i
k
d < i
k+1
. Since
O(I

) K(I

) = I, it cannot be the case c = d. Then c < d and, the denitions of

closed elements and open elements together with the fact that I

is totally ordered yield

j (c, d) such that j I. Take j
k
= j. If i
k
I take j
k
= i
k
. Now the sequence
j
1
< j
2
< . . . < j
n
of I satises the desired result.
Let X and Y be algebras with lattice reducts X and Y respectively. We show a
method to extend an order preserving map f : X Y , to the canonical extensions X

and Y

. In fact, we consider two extensions f

, f

: X

Y

, that can be computed
according to the next denition (see [41, Lemma 3.4]).
DEFINITION 7.4.3. For every order preserving map f : X Y we dene:
1. f

(c) =
_
f (x) [ c x X for every c K (X

).
2. f

(a) =
_
f

(c) [ a c K (X

).
3. f

(o) =
_
f (x) [ o x X for every o O(X

).
4. f

(a) =
_
f

(c) [ a o O(X

).
In case f : X Y is an order reversing map, the maps f

and f

are dened by
the following procedure:
Consider the function g : X
d
Y , where X
d
denote the lattice whose order is
obtained by reversing the order of X and g(x) = f(x) for every x X.
Compute g

, g

: (X
d
)

Y according to Denition 7.4.3, using the fact that

K
__
X
d
_

_
= O(X

) and O
__
X
d
_

_
= K (X

) .
436 Manuela Busaniche and Franco Montagna
Since
_
X
d
_

= (X

)
d
, let f

, f

: X

Y be such that f

(x) = g

(x) and
f

(x) = g

(x) for each x X

.
Let f :

n
i=1
X
i
Y be a map that preserves the order in some coordinates and
reverses it in others. The extensions f

, f

: (

n
i=1
X
i
)

Y

can be computed
following the previous procedure coordinatewise and recalling that
(X Y )

= X

Y

,
K ((X Y )

) = K (X

) K (Y

) ,
O((X Y )

) = O(X

) O(Y

) .
Let A = A, , , 1 be a hoop and the natural order of A. Recall that pre-
serves the order in both coordinates while reverses the order of the rst coordinate
and preserves it in the second one. According to the previous denition there are two
candidates to extend the algebra A to the complete lattice A

. The canonical exten-

sion A

= A

, 1 and the dual canonical extension A

= A

, 1
with the extension of the operations computed as previously dened. As an example, if
a, b K(A

) then
a

b =

c d [ a c, b d and c, d A ,
and if a O(A

) and b K(A

) we obtain
a

b =

c d [ c a, b d and c, d A .
A subvariety V of BL-algebras is called -canonical (or simply canonical) when-
ever A

V for every A V. Analogously, V is called -canonical (or dual canoni-

cal) whenever A

V for every A V.
There are many results, positive and negative, about canonicity of classes of alge-
bras. Two of the most important are the following theorem and its corollary.
THEOREM 7.4.4 ([54]). If a class Kof similar algebras with distributive lattice reducts
is closed under ultraproducts and -canonical (-canonical) extensions then the variety
generated by K is -canonical (-canonical).
COROLLARY 7.4.5. If V is a nitely generated variety of algebras with a distributive
lattice reduct, V is -canonical and -canonical.
We will investigate and canonicity in subvarieties of BL-algebras. As we have
observed in the previous section if a subvariety V of BL-algebras contains the vari-
ety of product algebras or it contains Changs MV-algebra C, then V does not admits
completions. In particular, V is not closed under or canonical extensions. This is
summarized in the next theorem:
THEOREM 7.4.6. Let V be a subvariety of BL-algebras that contains the variety of
product algebras or it contains Changs MV-algebra C. Then V is neither -canonical
nor -canonical.
Chapter V: H ajeks BL Logic and BL-Algebras 437
Observe, that due to Theorem 7.2.2, it only remains to study canonical extensions
of k-subvarieties.
7.5 Canonical extensions of k-subvarieties
In [21] the canonical extension of the underlying poset of an ordinal sum of Wa-
jsberg hoops is studied. We will explain how this extension can be computed when
the hoops involved are totally ordered. Let A be a BL-chain and let

iI
W
i
be
its decomposition into non-trivial Wajsberg hoops. Consider the totally ordered set
A =

iI
W
i
. The canonical extension A

is the totally ordered set obtained by the

following procedure:
1. Compute the canonical extension I

of the totally ordered set I.

2. For each i I, compute the canonical extension W

i
of the poset W
i
.
3. For each i I

I, take an element 0
i
that is not in

iI
W

i
.
4. For each i I

, let R
i
= W

i
1 if i I and R
i
= 0
i
otherwise.
5. Compute the ordered sum R =
iI
R
i
; that is a totally ordered set where x y
if x R
i
, y R
j
and i < j or x, y R
i
and x
i
y.
6. Add a greatest element 1 to the ordered set R.
Then the canonical extension of the poset

iI
W
i
is the totally ordered set A

=
R

1.
The careful reader might have noticed that, when each W

i
is a hoop, then A

is
the lattice reduct of

iI

S
i
, where
S
i
=
_
W

i
if i I,
L
2
if i I

I.
We will use Denition 7.4.3 to compute the extensions of the operations of the BL-
chain

iI
W
i
to the poset A

and we will observe that while A

coincides with

iI

S
i
, A

might not even be a hoop. To achieve such aim we consider the set
M = 0, 1 with the usual order and the operations and given by:
0 1
0 0 0
1 0 1
0 1
0 0 1
1 0 1
The system M = 0, 1, , , 1 is not a hoop. Observing that L

n
= L

n
= L
n
we are ready to prove the next theorem.
THEOREM 7.5.1. Let I be a lower bounded totally ordered set. Let A =

I
L
s
i
be a
BL-chain, with s
i
2 for each i I. Then
1.
_
iI
L
s
i
_

iI

D
i
, where D
i
=
_
L
s
i
if i I,
L
2
if i / I,
438 Manuela Busaniche and Franco Montagna
2.
_
iI
L
s
i
_

iI

C
i
, where C
i
=
_
L
s
i
if i I,
M if i / I.
REMARK 7.5.2. Though the operation of ordinal sum was dened only for hoops, one
can extend the denition for any algebra of the same type as hoops. That is what is to
be understood in the statement of the previous theorem when summing the algebra M.
Proof. We denote C =

iI

C
i
and D =

iI

D
i
. Their corresponding universes
(as well as their lattice reducts) are denoted by C and D respectively. As usual, A

is the
canonical extension of the poset A. We dene
D
: D I

by
D
(x) = i iff x D
i
.
Analogously we dene
C
: C I

.
Though if I

,= I the structure C is not a hoop, it possesses a natural order given

by the poset sum (

C
i
1) 1. From the comments before this theorem we
know that the lattices A

, C and D are isomorphic.

Without danger of confusion, for each j I

I, let 0
j
be the unique element in
C
j
1 as well as the unique element in D
j
1.
The set of closed and open elements of A

are:
K(A

) = 0
j
[ j K(I

) I A,
O(A

) = 0
j
[ j O(I

) I A.
To prove 1., we have to see that the canonical extensions

and

of the operations
of A, that are dened according to Denition 7.4.3, coincide with the operations
D
and

D
in D, given by the denition of ordinal sum. Clearly if a, b A, then a
D
b = a

b
and a
D
b = a

b. We check rst that

=
D
in the rest of the elements of A

.
From the fact that
D
and

are commutative, we only consider the following cases:

Case 1: a, b O(A

). This being the case, the possibilities are

1. a = b = 0
j
for some j O(I

) I. Then
0
j

0
j
=
_
D
c d [ c, d 0
j
and c, d A
=

D
c [ c 0
j
and c A = 0
j
= 0
j

D
0
j
.
2. a D
j
1, b D
k
1 with j < k. Since a, b O(A

), j, k O(I

). Thus
there exists l I such that j < l k and
a

b =
_
D
c d [ c a, d b and c, d A
=
_
D
c d [ c a, d b,
D
(d) l and c, d A
=
_
D
c [ c a, c A = a = a
D
b.
3. b = 1. Easily we obtain a

b = a = a
D
b.
Case 2: a, b / O(A

) O(A

). One of the following cases occurs:

Chapter V: H ajeks BL Logic and BL-Algebras 439
1. a = b = 0
j
for some j / I. Hence
0
j

0
j
=
_
D
c

d [ 0
j
c, d and c, d O(A

)
=
_
D
c [ 0
j
c and c O(A

) = 0
j
= 0
j

D
0
j
.
2. a D
j
1, b D
k
1 and j < k. Let l O(I

) be such that j l < k.

Then a

b =

D
c

d [ a c and b d, c, d O(A

) and
D
(c) l
=

D
c
D
d [ a c < b d, c, d O(A

) and
D
(c) l
=

D
c [ a c, c O(A

) = a = a
D
b.
The implication of Ais order reversing in the rst coordinate and order preserving in
the second coordinate. Therefore, by Lemma 7.4.3 and the remarks below it, to compute

, we have to consider rst a, b O((A

d
A)

) = K(A

) O(A

).
Case 1: a K(A

) and b O(A

). Note that in this case, if a, b D

j
1 for some
j I

, then j K (I

) O(I

) = I. Therefore, we have the following possibilities:

1. a D
j
1, b D
k
1 with j < k. Since a K(A

) and b O(A

),
j K (I

) and k O(I

). Thus there exist l, m I such that j l < m k

and
a

b =
_
D
c d [ a c, d b and c, d A
=
_
D
c d [ a c, d b,
D
(c) l,
D
(d) m and c, d A
=

D
c d [ a c < d b and c, d A
= 1 = a
D
b.
2. a D
j
1, b D
k
1 with j > k. Thus b a and
a

b =
_
D
c d [ a c, d b and c, d A
=
_
D
c d [, d b a c and c, d A
=
_
D
d [ d b, b A = b = a
D
b.
3. a = 1 or b = 1. It is easy to see that a

b = b = a
D
b.
Case 2: a, b / K(A

) O(A

). Then one of the following cases occurs:

1. a = b = 0
j
for some j I

I.
0
j

0
j
=
_
D
c

d [ c 0
j
d, c K(A

) and d O(A

)
=
_
D
c
D
d [ c 0
j
d, c K(A

) and d O(A

)
= 1 = 0
j

D
0
j
.
440 Manuela Busaniche and Franco Montagna
2. a D
j
1, b D
k
1 and j < k. Since a b,
a

b =
_
D
c

d [ c a, b d, c K(A

) and d O(A

)
=
_
D
c
D
d [ c a b d, c K(A

) and d O(A

)
= 1 = a
D
b.
3. a D
j
1, b D
k
1 and j > k. By the density of I

there exist l O(I

)
and m K (I

) such that k l < m j. Then

a

b =
_
D
c

d [ c a, b d, c K(A

) and d O( A

)
=
_
D
c
D
d [ b d < c a,
D
(c) m,
D
(d) l
=
_
D
d [ b d, d O(A

) and
D
(d) l = b = a
D
b.
To prove 2. notice that the sum C involves algebras of the form M when the index
is in I

I, whereas in (1) it involves L

2
. Observe that the operations
M
and
L
2
coincide, while
M
and
L
2
are different only in the pair 0, 0.
Therefore for (2), we are just going to verify that 0
j

0
j
= 0
j

C
0
j
, when
0
j
C
j
1 with j / I. This being the case j / K(I

) O(I

), thus 0
j
/
K(A

) O(A

). Using the fact that - and -canonical extensions coincide on closed

and open elements we obtain
0
j

0
j
=
_
C
b

c [ c 0
j
b , b O(A

) and c K(A

)
=
_
C
b

c [ c 0
j
b, b O(A

), c K(A

) and c < b
=
_
C
b
D
c [ c 0
j
b, b O(A

), c K(A

) and c < b
=
_
C
c [ c 0
j
, c K(A

) = 0
j
= 0
j

C
0
j
.
COROLLARY 7.5.3. The variety of G odel algebras is not -canonical.
Proof. Take a lower bounded totally ordered set I that is not complete. The above
theorem asserts that the -canonical extension of the G odel algebra

I
L
2
is not a
BL-chain. Thus the variety of G odel algebras is not closed under -canonicity.
THEOREM 7.5.4. Let Vbe a k-subvariety of BL-algebras that is not nitely generated.
Then V contains the variety of G odel algebras.
Proof. Since V is a k-subvariety every chain in V is a k-chain. Let S be a set of chains
that generates V. The hypothesis that V is not nitely generated implies that one of the
following cases must occurs:
Case 1: S contains an innite set of non-isomorphic nite BL-chains. Denote by
B
j

=

iI
j
L
n(j,i)
(j J)
these elements of S, where J is an innite set of indices, and for each j J, I
j
is a
nite totally ordered set. Let [I
j
[ be the cardinal of the set I
j
. Since V is a k-subvariety,
it can not be the case that the set [I
j
[ [ j J is bounded, because that will contradict
Chapter V: H ajeks BL Logic and BL-Algebras 441
the fact that the set B
j
[ j J is an innite set of non-isomorphic BL-chains. Then
the set [I
j
[ [ j J is unbounded. For each j J, let A
j
=

I
j
L
2
. Clearly for
each j J, the algebra A
j
is a nite G odel algebra isomorphic to a subalgebra of B
j
.
We conclude that
A
j
[ j J V
is an innite set of non-isomorphic nite G odel chains and it is well known that this
implies V contains the variety of G odel algebras.
Case 2: S contains an innite BL-chain B. Then B is of the form

I
L
s
i
, with I an
innite set and 2 s
i
k + 1. Then the algebra H =

I
L
2
, is a subalgebra of
B which is an innite G odel algebra. Every innite G odel chain generates the whole
variety of G odel algebras, thus V contains the variety of G odel algebras.
Note that any nitely generated subvariety of BL-algebras is a k-subvariety for some
k N.
THEOREM7.5.5. A subvariety of BL-algebras is -canonical iff it is nitely generated.
Proof. One of the implications is immediate from Corollary 7.4.5. For the other one,
assume that V is not nitely generated. From the classication theorem 7.2.2, we get
that V is a non-nitely generated k-subvariety or it either contains the variety of prod-
uct algebras or Changs MV-algebra. If the rst case happens the result follows from
Corollary 7.5.3 and Theorem 7.5.4. If the second case occurs, the result follows from
Theorem 7.4.6.
We now turn our attention to -canonicity. As we did for -canonicity, due to
Theorem 7.4.6 and Theorem 7.2.2 it is left to study -canonicity only in k-subvarieties.
LEMMA 7.5.6. For each k-chain B the dual canonical extension B

is in the subvari-
ety of BL-algebras generated by B.
Proof. Following the arguments of Theorem 7.3.5, we just need to see that every nite
subalgebra of B

is isomorphic to a nite subalgebra of B. Consider the decomposition

of B into non-trivial Wajsberg chains, i.e.,
B =

iI
L
s
i
for some bounded totally ordered set I, with lower bound 0, and for each i I, 2
s
i
k + 1. By Theorem 7.5.1, the dual canonical extension of B is the k-chain
B

iI

L
t
i
,
where for each i I, t
i
= s
i
and for each i I

I, t
i
= 2. Let Abe a nite subalgebra
of B

. Then there is a nite set J = 0 = j

0
< j
1
< . . . < j
n
I

such that
A =

j
i
J
L
r
j
i
.
442 Manuela Busaniche and Franco Montagna
Observe that if j
i
I

I, then r
j
i
= 2, while if j
i
I, then L
r
j
i
is a subalgebra
(subhoop) of L
t
j
i
. Using Lemma 7.4.2 we can obtain a chain K = 0 = k
0
< k
1
<
. . . < k
n
I, such that if j
i
I, then k
i
= j
i
. Following the arguments in the proof
of Theorem 7.3.4, we can assert that the chain A
t
=

k
i
K
L
u
k
i
where u
k
i
= r
j
i
if
k
i
= j
i
and u
k
i
= 2 otherwise, is a subalgebra of B and A

= A
t
. Thus every nite
subalgebra of B

is isomorphic to a nite subalgebra of B and the proof is concluded.

THEOREM 7.5.7. Let V be a subvariety of BL-algebras. Then V is -canonical if and
only if V is a k-subvariety for some k N.
Proof. Assume that V is a k-subvariety for some k N. Let S be the set of all chains
in V. Clearly V is generated by S. Since V is a k-subvariety, each element of S is a
k-chain. By Lemma 7.5.6, S is closed under -canonical extensions. An ultraproduct
of BL-chains in V is a BL-chain in V, hence S is closed under ultraproducts. Now an
application of Theorem 7.4.4 proves that V is -canonical.
Now assume that V is not a k-subvariety. Due to Theorem 7.2.2 and Theorem 7.4.6
we conclude that V is not -canonical.
8 Further investigations of BL-algebras
8.1 Introduction
In this section we briey report on some additional results on BL-algebras. First
we consider generalizations of BL-algebras. Such generalizations can be of two types:
(a) investigation of varieties which arise by weakening the language of BL-algebras (ba-
sic hoops, BL-BCK-algebras) or the axioms of BL-algebras (psBL-algebras, bounded
GBL-algebras) or both (hoops, HBCK-algebras, GBL-algebras); (b) investigation of
BL-algebras with an additional structure (e.g., BL-algebras with constants). This last
investigation will not be treated here because it is studied in Chapter VIII of the present
handbook.
Moreover, two additional specic topics will be discussed, the rst one essentially
logical, namely, interpolation, Beths property and uniform interpolation in extensions
of BL, possibly with additional operators, and the second one purely algebraic, that is,
sheaf representation of BL-algebras by Boolean products or by weak Boolean products.
8.2 Generalizations of the notion of BL-algebra
As said before, BL-algebras are divisible, commutative, integral, bounded and rep-
resentable residuated lattices. One may wonder what we obtain if we remove any of the
above conditions from the denition of BL-algebra. If we remove divisibility, we get
MTL-algebras, which are already treated in Chapter VI of this Handbook, and hence
will not be discussed here. If we remove all the other conditions, we get GBL-algebras,
see [51], that is, divisible residuated lattices.
But BL-algebras may also be regarded as bounded and representable hoops. Al-
though equivalent to the previous one, this alternative denition introduces a difference
when we investigate fragments. Indeed, if we remove representability and boundedness
Chapter V: H ajeks BL Logic and BL-Algebras 443
from this second denition, we obtain the notion of hoop, which is different from the
notion of commutative and integral GBL-algebra, in that in a hoop the join operation,
which is a primitive operation in GBL-algebras, is not denable in general.
After a general overview of GBL-algebras, we will briey discuss some special
kind of integral and bounded GBL-algebras, namely, the pseudo-BL-algebras (psBL-
algebras in the sequel). The main difference between BL-algebras and psBL-algebras
is that the latter are not commutative and not even representable, although they satisfy
some non-commutative version of the axiom of representability.
Then we will concentrate on those generalizations of BL-algebras in which divisi-
bility, commutativity and integrality is preserved.
8.2.1 GBL-algebras
AGBL-algebra is a divisible residuated lattice. Recall that a (possibly non-integral,
possibly non-commutative) residuated lattice is an algebra A, , , , , /, e such that
A, , e is a monoid, A, , is a lattice and the following residuation property holds:
x y z iff y xz iff x z/y.
A residuated lattice is said to be divisible, or a GBL-algebra, iff, whenever x y,
there are z, u such that zy = yu = x or equivalently if the identities ((x y)/y)y =
y(y(x y)) = x y are satised.
If we add any of commutativity, integrality, representability or boundedness to the
denition of GBL-algebras, we obtain the class of commutative, of integral, of repre-
sentable and of bounded GBL-algebras respectively.
EXAMPLES 8.2.1. Examples of GBL-algebras:
1. Lattice ordered groups (or -groups for short). These are algebras of the form

G, , , ,
1
, e
_
where G, , e is a monoid, G, , , is a lattice and the iden-
tities x(yz) = xyxz and (yz)x = yxzx hold. An -group becomes
a GBL-algebra with respect to the operations xy = x
1
y and y/x = y x
1
.
2. 0-free subreducts of Heyting algebras (also called Brouwerian algebras). These
are GBL-algebras in which x y = x y. These algebras are commutative
and integral residuated lattices, and hence the residuals and / coincide and are
denoted by . In particular, a Heyting algebra is a bounded GBL-algebra in
which the identity x y = x y holds.
3. GMV-algebras. These are GBL-algebras in which the identities y/((xy)e)) =
((y/x) e)y = x y holds. Note that -groups are GMV-algebras and MV-
algebras are precisely the commutative, integral and bounded GMV-algebras.
Here are some interesting facts about GBL-algebras and GMV-algebras:
1. Every GBL-algebra is isomorphic to a direct product of an -group and an integral
GBL-algebra [51].
444 Manuela Busaniche and Franco Montagna
2. Every integral and bounded GMV-algebra (also called psMV-algebra) is repre-
sentable as an interval [e, u], with e u, of an -group, see [42]. Moreover every
integral GMV-algebra is isomorphic to the image of a negative cone of an -group
with respect to a nucleus [51].
3. The lattice reduct of a GBL-algebra is distributive [51].
4. Every totally ordered integral GBL-algebra is the ordinal sum of a tower of totally
ordered integral GMV-algebras, which are either intervals of -groups or negative
cones of -groups [43].
5. Let n be a positive integer. A GBL-algebra is said to be n-potent if it satises
the identity x
n+1
= x
n
. Then, for every positive integer n, we have that every
n-potent GBL-algebra is commutative and integral [68]. In particular, every nite
GBL-algebra is commutative and integral [67].
6. Finite GBL-algebras are precisely the poset sums of nite BL-chains with respect
to a nite poset [68].
A pseudo BL-algebra (psBL-algebra for short) is an integral and bounded GBL-
algebra which satises the identities
xy yx = y/x x/y = e (ps)
Pseudo BL-algebras have been studied by several authors, see e.g. [36, 43, 47, 65, 77].
For a more detailed bibliography, the reader is invited to consult [64], where one can
also nd several examples of psBL-algebras.
Somewhat surprisingly, although the identity (ps) is reminiscent of the axiom of
prelinearity, it does not imply representability. For a counterexample, one can consult
e.g. [67]. Representability is axiomatized in [77] by the following axioms:
xy ((z (yx))/z) = y/x (z((x/y) z)) = 1.
Hence, representable psBL-algebras constitute a variety, which is a proper subvari-
ety of the variety of psBL-algebras.
Dvure censkijs ordinal sum representation for totally ordered and integral GBL-
algebras holds for psBL-algebras as well, with the only difference that in the case of
psBL-algebras the tower should have a rst component which is a psMV-algebra.
One may wonder whether representable psBL-algebras correspond to the continu-
ous pseudo t-norm (a pseudo t-norm is dened in analogy to a t-norm, with the only
difference that commutativity is removed). Unfortunately the answer is negative: every
continuous pseudo t-norm on [0, 1] is commutative, see [47].
8.2.2 Hoops and commutative and integral GBL-algebras
As said before, BL-algebras may be equivalently regarded as representable and
bounded hoops or as basic (i.e., representable), commutative, integral and bounded
GBL-algebras. This is true because meet is denable in any hoop, and in any basic
hoop also join is denable. But if we remove the representability condition, join is
Chapter V: H ajeks BL Logic and BL-Algebras 445
no longer denable, and hence bounded hoops and bounded commutative and integral
GBL-algebras are different classes.
For hoops, a weaker form of ordinal sum decomposition has been proved in [13],
namely:
THEOREM 8.2.2. Every subdirectly irreducible hoop is the ordinal sum of a proper
subhoop of it and of a non-trivial Wajsberg hoop.
This property is weaker than the Aglian` o-Montagna decomposition theorem, but
it is more general, because it also holds for non-representable hoops. Moreover, it is
very powerful. For instance, it allows to prove the nite embeddability property (a
strengthening of the nite model property) for hoops, and hence the decidability of the
quasiequational theory of hoops.
The Blok-Ferreirim theorem has been extended to commutative and integral GBL-
algebras in [68]: every subdirectly irreducible commutative and integral GBL-algebra is
the ordinal sum of a proper subalgebra and of a subdirectly irreducible commutative and
integral GMV-chain. For the sake of precision, this result has been proved for a slightly
wider class of GBL-algebras. In any case, it implies the nite embeddability property
for the class of commutative and integral GBL-algebras. In particular, this result implies
the well-known fact that the quasiequational theory of Heyting algebras is decidable.
Finally, in [69], the Blok-Ferreirim decomposition theorem has been used in the
proof of the following theorem:
THEOREM 8.2.3. Every commutative and integral GBL-algebra embeds into the poset
sum of a family of MV-chains.
8.3 Implicative subreducts of BL-algebras and of their generalizations
BCK-algebras were introduced by K. Is eki as algebraic models of Merediths BCK-
logic, a non-classical propositional logic containing implication as the only connective.
A BCK-algebra is an algebra A = A, , 1 of type (2, 0) satisfying the following
identities and quasi-identity:
1. x x = 1
2. x (y x) = 1
3. (x y) ((y z) (x z)) = 1
4. x (y z) = y (x z)
5. if x y = 1 and y x = 1 then x = y.
This presentation of BCK-algebras is the dual of the original one (see [66]). The
class BCKof all BCK-algebras is a quasivariety which is not a variety (see [91]). Some
of the most important results on BCK-algebras that have helped develop the theory of
residuated lattices can be found in [14, 15, 30, 55, 56].
A bounded BCK-algebra is an algebra A = A, , 0, 1 of type (2, 0, 0) such that
the reduct A, , 1 is a BCK-algebra and 0 is a constant that satises the equation
0 x = 1.
446 Manuela Busaniche and Franco Montagna
The class of bounded BCK-algebras, that will be denoted by BBCK is a quasivari-
ety which is not a variety.
Given a BCK-algebra A we can dene at most one binary operation satisfying
the equation
(x y) z = x (y z).
If this operation exists, it is associative, commutative and satises x 1 = 1 for
all x A. Then in the expanded algebra A
t
= A, , , 1 the residuation property
x y z iff x y z holds. In fact, A
t
is a partially ordered commutative residuated
integral monoid; i.e., a pocrim. If the operation is dened in a bounded BCK-algebra,
then we get a bounded pocrim.
Reciprocally, given a pocrim A, the implication reduct of Ais a BCK-algebra.
Since hoops are particular cases of pocrims, the , 1 subreducts of hoops form
a subvariety of BCK, denoted by HBCK that can be dened as the subquasivariety of
BCK that satises the equation:
(x y) (x z) = (y x) (y z).
The elements of HBCK are called HBCK-algebras. For further results about pocrims
and HBCK-algebras see e.g. [15].
The implicative subreducts of BL-algebras are called BL-BCK-algebras. These
algebras are precisely the HBCK-algebras satisfying the prelinearity equation
((x y) z) (((y x) z) z) = 1.
Indeed, it is readily seen that every hoop BCK-algebra satises the above equation if
and only if it decomposes as a subdirect product of totally ordered hoop BCK-algebras.
8.4 Further results about BL-algebras
8.4.1 Interpolation in extensions of BL
Although BL, ukasiewicz logic and product logic satisfy the deductive interpo-
lation property, none of them has the Craig interpolation property. Indeed, let :
(p (p r)) and : (q (q r)). Then, the formula is provable in any ex-
tension of BL, but if L is any extension of BL which does not extend G, then has
no interpolant in L. To the contrary, G odel logic has the uniform interpolation property,
see [8]. As noted in [49], sometimes the lack of interpolation in a logic L is due to the
weakness of the language of L. Then, the problem is to nd a conservative extension of
L in a richer language in which Craig interpolation holds.
In [8], Baaz and Veith solve the problem for ukasiewicz logic and for product
logic. Indeed, they prove that if is extended by a division operator d
n
for every integer
n > 1, together with the axioms nd
n
() and (d
n
() (n 1)d
n
()), then the
resulting logic has uniform interpolation. Moreover if a conservative extension of has
Craig interpolation, then the division operators must be denable in it.
For product logic one has a similar result. In this case, one needs n-root operators
instead of division operators.
A partial result for BL has been given in [83]: indeed, there is an axiom
n
which
is satised in a BL-chain iff it has at most n Wajsberg components. Let BL(n) be BL
Chapter V: H ajeks BL Logic and BL-Algebras 447
plus
n
. Then, although BL
n
is stronger than BL, it proves the same formulas in less
than n variables, and in [83] it is shown that for every n there is a conservative extension
of BL
n
which has uniform interpolation.
8.4.2 Sheaf representation of BL-algebras
Sheaf representations of algebras are very useful, because they reduce the study
of algebras to the study of the stalks, which are usually better known structures. For
background on sheaf representations of universal algebras see e.g. [33, 79].
In [78] a denition of BL-sheaf space (or sheaf space of BL-algebras) is presented,
following the general line of sheaf spaces of universal algebra. Let A be a non-trivial
BL-algebra and let Spec(A) be the set of prime implicative lters of A. For each P
Spec(A) we denote by O(P) the proper lter of A given by a A [ a x =
for some x A P. It is proved that the family of lters O(P) [ P Spec(A)
determines a sheaf representation of A.
Particular cases of sheaf representations are the Boolean product representation and
the weak Boolean product representation. For denitions see [18, 19].
In the case of BL-algebras these representations can be useful when the Boolean
base space is a known space and the stalks of the representations are BL-algebras with
a simpler structure, such as BL-chains. Unfortunately this is not always the case.
Concretely, for each BL-algebra A consider the set Spec(B(A)) to be the set of
ultralters of the Boolean skeleton of A. For each U Spec(B(A)), we denote by U
the implicative lter of Agenerated by U. Then the correspondence
a (a/U)
USpec(B(A))
gives a representation of Aas a weak Boolean product of the family
A/U [ U Spec(B(A))
over the Boolean space Spec(B(A)). The important matter about this representation of
BL-algebras as weak Boolean product is that the stalks of the representation are directly
indecomposable BL-algebras. Though in general there is nothing we can say about
directly indecomposable BL-algebras, there are some subvarieties of BL-algebras whose
directly indecomposable members are easy to manage structures. For instance, if for
each U Spec(B(A)) the lter U is a prime implicative lter of A, then each stalk
A/U is a BL-chain. Below we present two examples of the use of the weak Boolean
product representation:
EXAMPLES 8.4.1.
Free algebras in subvarieties generated by a BLn-chain. A BLn-chain is a BL-chain
whose rst Wajsberg component is the n-elements MV-chain. Consider a BLn-chain A,
and let C be the totally ordered hoop such that A

= L
n
C. Denote by Vthe subvariety
of BL-algebras generated by A and by F
V
(X) the free BL-algebra in V over the free
generating set X. In [22] it is proved that the Boolean skeleton B(F
V
(X)) is the free
Boolean algebra over the partially ordered set Y = (x)
n
i
(x) [ x X, i = 1, . . . n
where for each i = 1, . . . , n,
n
i
(x) are unary operators, known as n-valued Moisil
operators, dened over X. The following theorem was proved as [22, Theorem 3.9]:
448 Manuela Busaniche and Franco Montagna
THEOREM 8.4.2. The free BL-algebra F
V
(X) can be represented as a weak Boolean
product of the family
F
V
(X)/U [ U Spec B(F
V
(X))
where B(F
V
(X)) is the free Boolean algebra over the poset Y =
n
i
(x) [ x X,
i = 1, . . . , n 1. Moreover, for each U Spec B(F
V
(X)) there exists m 2 such
that m1 divides n 1 and
F
V
(X)/U = L
m
F
W
(Z)
where Z = x/U [ x X, x U and W is the variety of basic hoops
generated by C.
Free algebras in subvarieties with a Boolean retract. In [29] free BL-algebras in sub-
varieties of BL-algebras with a Boolean retract are described. These are subvarieties
such that for each algebra Ain the subvariety there exists a retraction
A
, from Aonto
B(A), with the identity as associated section. Every subvariety of SBL-algebras and
subvarieties of bipartite MV-algebras are subvarieties with a Boolean retract. If V is a
subvariety of BL-algebras with a Boolean retract, and F
V
(X) is the free algebra in V
over a set X of free generators, then Spec(B(F
V
(X))) is the Cantor space over 2
]X]
.
Using the weak Boolean product representation of BL-algebras, it is proved that if V
is a subvariety of BL-algebras with a Boolean retract, then the free algebra F
V
(X) is
isomorphic to a weak Boolean product of a family
(F
V
r (S)) [ S 2
X

over the Cantor space 2

]X]
, where F
V
r (S) is a free algebra in the subvariety V
r
of basic
hoops generated by the radical algebras of V and is map from V
r
into V.
9 Historical remarks and further reading
The history of BL and of BL-algebras is rather recent, and hence this section will be
rather short. BL logic and BL-algebras were introduced by Petr H ajek [57, 58]. Since
BL is not isolated from the other fuzzy logics, we may wonder what are the forerunners
of this logical system. For sure, the work on t-norms [73, 84]), as well as the work on the
three main fuzzy logics, namely ukasiewicz [24, 26], G odel [40], and product [28, 59]
played an important role in the birth of BL. Indeed, the motivations for BL were both
nding a common generalization of the three main fuzzy logics and nding the logic of
all continuous t-norms.
While Zadeh [93] was more interested in fuzzy sets and their applications rather
than in the logical principles, H ajek payed attention not only to the semantics or to the
applications, but also to the syntactic aspects. Now ukasiewicz, G odel, and product
logics are based on residuation, and hence they belong to the realm of substructural log-
ics (see [50]). H ajek was aware of this fact. He explicitly mentioned that BL-algebras
are residuated lattices. The paper [84] showed that the logic of any continuous t-norm
can be reduced to the three main fuzzy logics. Indeed, every continuous t-norm decom-
poses as an ordinal sum of G odel, ukasiewicz and product components. This might
Chapter V: H ajeks BL Logic and BL-Algebras 449
have suggested that the problem of nding a common fragment of the main fuzzy logics
and the problem of nding the logic of continuous t-norms were related.
Besides the prelinearity axiom, the main contribution of H ajek to the axiomatization
of BL was nding an axiom which formalizes the continuity of the t-norm. The axiom is
( & ( )) ( & ( )),
or, in algebraic terms,
(x (x y)) = (y (y x)).
Interestingly, this axiom was already present in the denition of hoops (see [13, 17]),
though its role there was not the formalization of continuity, but the formalization of
divisibility: if a b, there is c such that a = b c. Hence, in a sense, [17] was a fore-
runner of BL, but as far as we know, H ajek was not aware of this monograph when he
presented BL-algebras. Thinking of divisibility as a continuity condition was a source
of errors: it is not true that in a BL-chain product is continuous with respect to the order
topology. In particular, the monoidal operation in a BL-algebra does not distribute over
innitary meets, and the principle (x)((x) &) ((x(x)) &) is not valid in all
BL-chains, although it is valid in all BL-chains on [0, 1].
After formalizing BL logic and BL-algebras, H ajek started a deep investigation.
His rst result was that BL is complete with respect to the class of BL-chains, see [58].
But the challenge was to prove that BL is the logic of all continuous t-norms and their
residuals. In [57], under some supplementary conditions, H ajek proved that this result is
true, i.e., he proved that BL extended by two axioms is the logic of continuous t-norms.
In [27] it was proved that the additional axioms introduced by H ajek are redundant.
In other words, BL, as axiomatized by H ajek, is the logic of all continuous t-norms and
their residuals. This is perhaps the rst important result about BL, though [57] already
contains the main ideas for this result.
The paper [57] also contains some very useful ideas for an extension of the Mostert
and Shields theorem to BL-chains. Indeed, H ajek tries to decompose BL-chains into
ordinal sums of sum-indecomposable components, i.e., of components which are no
longer decomposable as ordinal sums. What is still missing in [57] is a characterization
of sum-indecomposable components as Wajsberg hoops.
The paper [13] is a fundamental step towards an ordinal sum decomposition of BL-
chains: it proves that any subdirectly irreducible hoop is the ordinal sum of a proper
subhoop and a subdirectly irreducible Wajsberg hoop.
Putting ideas from [57] and from [13] together, Aglian o and Montagna proved in [2]
a generalization of the Mostert and Schields theorem: any BL-chain can be decomposed
as an ordinal sum of Wajsberg hoops. A proof of this result without the use of the axiom
of choice was given in [20]. In the same way as the Mostert and Shields decomposition
result is fundamental in the theory of continuous t-norm, the decomposition result of [2]
is a very powerful tool in the study of BL-algebras. For instance, it provides an alter-
native proof of the completeness theorem in [27], as well as an alternative proof of the
co-NP completeness of BL given in [7]. Finally, it constitutes an important tool for the
study of subvarieties of the variety of BL-algebras. Already in [27] it was pointed out
that the join of the varieties of MV-algebras, of product algebras and of G odel algebras
450 Manuela Busaniche and Franco Montagna
does not exhaust the whole variety of BL-algebras. The lattice of subvarieties generated
by a single component was described in [35]. However, most varieties of BL-algebras
are not generated by a single component. In particular, in [2] it is shown that there is a
continuum of varieties of BL-algebras.
The logic BL was mainly studied by semantic means, through its algebraic coun-
terpart, the class of BL-algebras. Although several fuzzy logics do have analytic proof
systems, see for instance [5, 6, 25, 82], in many cases proof systems are based on seman-
tics. Moreover, the proof theory of BL is still unsatisfactory. For instance, the paper [16]
has not attracted much attention from the proof theorists of Fuzzy Logic. Hence, several
syntactic properties which are commonly treated syntactically, in the case of BL have
been proved by semantic means.
An important example of proof-theoretical result which has been investigated by
semantic means is the interpolation property. The rst important results on interpolation
are contained in [8]. There it is proved that the only continuous t-norm logic having
Craigs interpolation is G odel logic. Moreover, although neither ukasiewicz nor prod-
uct logic have Craig interpolation, it is possible to add operators on them in such a way
that the extended logics do have interpolation. In the case of ukasiewicz logic, it is
sufcient to add division operators and for the case of product logic it is sufcient to
add n-root operations. In view of the results of [87], where it is shown that amalgama-
tion implies Robinson property, the failure of interpolation caused a misunderstanding
for a while, because the most important extensions of BL do have the amalgamation
property. The misunderstanding was due to the fact that in many fuzzy logics Craig
interpolation and deductive interpolation are not equivalent. This problem constituted
the main motivation for [83], where a rather complete treatment of interpolation in BL
is given.
So far, we have described the birth of BL and its developments. To make our history
of BL complete, we also have to say a few words about its children. The idea of nding
more and more general fuzzy logics, which is one of the sources of inspiration of BL, has
led to other new and interesting logics, namely: Esteva and Godos logic MTL, which is
obtained by relaxing the divisibility condition, or alternatively, by replacing continuous
t-norms by left-continuous t-norms, see [44], Uninorm logic UL, see [82, 92], obtained
by relaxing both divisibility and integrality, and psBL, a non-commutative version of
BL, see [36] for an algebraic treatment.
Acknowledgments
We would like to thank Roberto Cignoli, Petr H ajek and Daniele Mundici who
highly contributed to motivate our interest in fuzzy logic. We are also grateful to Pe-
ter Jipsen and Antoni Torrens for their important suggestions to improve this chapter.
Finally, the rst author would like to express her sincere thanks to Leonardo Cabrer, for
his valuable comments on Section 7.
BIBLIOGRAPHY
[1] Paolo Aglian` o, Isabel Maria Andr e Ferreirim, and Franco Montagna. Basic hoops: An algebraic study
of continuous t-norms. Studia Logica, 87(1):7398, 2007.
Chapter V: H ajeks BL Logic and BL-Algebras 451
[2] Paolo Aglian` o and Franco Montagna. Varieties of BL-algebras I: General properties. Journal of Pure
and Applied Algebra, 181(23):105129, 2003.
[3] Norman L. Alling. On order divisible groups. Transactions of the American Mathematical Society,
94(3):498514, 1960.
[4] K. Amer. Equationally complete classes of commutative monoids with monus. Algebra Universalis,
18(1):129131, 1984.
[5] Matthias Baaz, Agata Ciabattoni, and Christian Georg Ferm uller. Hypersequent calculi for G odel logics:
A survey. Journal of Logic and Computation, 13(6):835861, 2003.
[6] Matthias Baaz, Agata Ciabattoni, and Franco Montagna. Analytic calculi for monoidal t-norm based
logic. Fundamenta Informaticae, 59(4):315332, 2004.
[7] Matthias Baaz, Petr H ajek, Franco Montagna, and Helmut Veith. Complexity of t-tautologies. Annals of
Pure and Applied Logic, 113(13):311, 2002.
[8] Matthias Baaz and Helmut Veith. Interpolation in fuzzy logic. Archive for Mathematical Logic,
38(7):461489, 1999.
[9] Raymond Balbes and Philip Dwinger. Distributive Lattices, volume 1. University of Missoury Press,
Columbia, 1974.
[10] Matteo Bianchi and Franco Montagna. Supersound many-valued logics and Dedekind-MacNeile com-
pletions. Archive for Mathematical Logic, 48(8):719736, 2009.
[11] Garrett Birkhoff. Lattice theory, volume 25 of American Mathematical Society ColloquiumPublications.
American Mathematical Society, Providence, RI, third edition edition, 1967.
[12] Willem J. Blok and Isabel Maria Andr e Ferreirim. Hoops and their implicational reducts. Algebraic
Methods in Logic and Computer Sciences, Banach Center Publications, 28:219230, 1993.
[13] Willem J. Blok and Isabel Maria Andr e Ferreirim. On the structure of hoops. Algebra Universalis,
43(23):233257, 2000.
[14] Willem J. Blok and James G. Raftery. On the quasivariety of BCK-algebras and its subvarieties. Algebra
Universalis, 33:6890, 1995.
[15] Willem J. Blok and James G. Raftery. Varieties of commutative residuated integral pomonoids and their
residuation subreducts. Journal of Algebra, 190(2):280328, 1997.
[16] Simone Bova and Franco Montagna. Proof search in H ajeks Basic Logic. ACM Transactions on Com-
putational Logic, 9(3), 2008.
[17] J.R. B uchi and Owens T.M. Complemented monoids and hoops. Unpublished manuscript.
[18] Stanley Burris and H.P. Sankappanavar. A Course in Universal Algebra, volume 78 of Graduate Texts
in Mathematics. Springer-Verlag, 1981.
[19] Stanley Burris and Heinrich Werner. Sheaf constructions and their elementary properties. Transactions
of the American Mathematical Society, 248(2):269309, 1979.
[20] Manuela Busaniche. Decomposition of BL-chains. Algebra Universalis, 54(4):519525, 2004.
[21] Manuela Busaniche and Leonardo M. Cabrer. Canonicity in subvarieties of BL-algebras. Algebra Uni-
versalis, 62(4):375397, 2010.
[22] Manuela Busaniche and Roberto Cignoli. Free algebras in varieties of BL-algebras generated by BL
n
-
chains. Journal of the Australian Mathematical Society, 80(3):419439, 2006.
[23] Leonardo M. Cabrer. Non-canonicity of BL-algebras. Reports on Mathematical Logic, 44:107125,
2008.
[24] Chen Chung Chang. Algebraic analysis of many-valued logics. Transactions of the American Mathe-
matical Society, 88(2):467490, 1958.
[25] Agata Ciabattoni, Christian G. Ferm uller, and George Metcalfe. Uniform rules and dialogue games for
fuzzy logics. In Proceedings of LPAR 2004, volume 3452 of Lecture Notes in Articial Intelligence,
pages 496510. Springer, 2005.
[26] Roberto Cignoli, Itala M.L. DOttaviano, and Daniele Mundici. Algebraic Foundations of Many-Valued
Reasoning, volume 7 of Trends in Logic. Kluwer, Dordrecht, 1999.
[27] Roberto Cignoli, Francesc Esteva, Llus Godo, and Antoni Torrens. Basic fuzzy logic is the logic of
continuous t-norms and their residua. Soft Computing, 4(2):106112, 2000.
[28] Roberto Cignoli and Antoni Torrens. An algebraic analysis of product logic. Multiple-Valued Logic,
5(1):4565, 2000.
[29] Roberto Cignoli and Antoni Torrens. Free algebras in varieties of BL-algebras with a Boolean retract.
Algebra Universalis, 48(1):5579, 2002.
[30] Roberto Cignoli and Antoni Torrens. Glivenko like theorems in natural expansions of BCK-logic. Math-
ematical Logic Quarterly, 50(2):111125, 2003.
[31] Roberto Cignoli and Antoni Torrens. H ajek basic fuzzy logic and ukasiewicz innite-valued logic.
Archive for Mathematical Logic, 42(4):361370, 2003.
452 Manuela Busaniche and Franco Montagna
[32] Petr Cintula, Francesc Esteva, Joan Gispert, Llus Godo, Franco Montagna, and Carles Noguera. Distin-
guished algebraic semantics for t-norm based fuzzy logics: Methods and algebraic equivalencies. Annals
of Pure and Applied Logic, 160(1):5381, 2009.
[33] Brian A. Davey. Sheaf spaces and sheaves of universal algebras. Mathematische Zeitschrift, pages
275290, 1973.
[34] Brian A. Davey and Hillary A. Priestley. Introduction to Lattices and Order. Cambridge University
Press, Cambridge, 1990.
[35] Antonio Di Nola, Francesc Esteva, Pere Garcia, Llus Godo, and Salvatore Sessa. Subvarieties of BL-
algebras generated by single-component chains. Archive for Mathematical Logic, 41(7):673685, 2002.
[36] Antonio Di Nola, George Georgescu, and Afrodita Iorgulescu. Pseudo-BL-algebras, part III. Multiple-
Valued Logic, 8(56):673714, 717750, 2002.
[37] Antonio Di Nola and Ada Lettieri. Perfect MV-algebras are categorically equivalent to Abelian -groups.
Studia Logica, 53(3):417432, 1994.
[38] Antonio Di Nola and Ada Lettieri. One chain generated varieties of MV-algebras. Journal of Algebra,
25(2):667679, 2000.
[39] Antonio Di Nola and Ada Lettieri. Finite BL-algebras. Discrete Mathematics, 269(13):93112, 2003.
[40] Michael Dummett. A propositional calculus with denumerable matrix. Journal of Symbolic Logic,
24(2):97106, 1959.
[41] J. Michael Dunn, Mai Gehrke, and Alessandra Palmigiano. Canonical extensions and relational com-
pletness of some substructural logics. Journal of Symbolic Logic, 70(3):713740, 2005.
[42] Anatolij Dvure censkij. Pseudo MV-algebras are intervals in -groups. Journal of the Australian Mathe-
matical Society, 72(3):427445, 2002.
[43] Anatolij Dvure censkij. Agliano-Montagna type decomposition of linear pseudo hoops and its applica-
tions. Journal of Pure and Applied Algebra, 221(3):851861, 2007.
[44] Francesc Esteva and Llus Godo. Monoidal t-norm based logic: Towards a logic for left-continuous
t-norms. Fuzzy Sets and Systems, 124(3):271288, 2001.
[45] Francesc Esteva, Llus Godo, and Franco Montagna. Equational characterization of subvarieties of BL-
algebras generated by t-norm algebras. Studia Logica, 76(2):161200, 2004.
[46] Isabel Maria Andr e Ferreirim. On Varieties and Quasivarieties of Hoops and Their Reducts. PhD thesis,
University of Illinois at Chicago, Chicago, 1992.
[47] Paul Flondor, George Georgescu, and Afrodita Iorgulescu. Pseudo t-norms and pseudo-BL-algebras.
Soft Computing, 5(5):355371, 2001.
[48] Josep Maria Font, Antonio J. Rodrguez Salas, and Antoni Torrens. Wajsberg algebras. Stochastica,
8(1):531, 1984.
[49] Dov M. Gabbay and Larisa L. Maksimova. Interpolation and Denability: Modal and Intuitionistic
Logics, volume 46 of Oxford Logic Guides. Clarendon Press, Oxford, 2005.
[50] Nikolaos Galatos, Peter Jipsen, Tomasz Kowalski, and Hiroakira Ono. Residuated Lattices: An Alge-
braic Glimpse at Substructural Logics, volume 151 of Studies in Logic and the Foundations of Mathe-
matics. Elsevier, Amsterdam, 2007.
[51] Nikolaos Galatos and Constantine Tsinakis. Generalized MV-algebras. Journal of Algebra, 283(1):254
291, 2005.
[52] Mai Gehrke and John Harding. Bounded lattice expansions. Journal of Algebra, 238(1):345371, 2001.
[53] Mai Gehrke and Bjarni J onsson. Bounded distributive lattices with operators. Mathematica Japonica,
40(2):207215, 1994.
[54] Mai Gehrke and Bjarni J onsson. Bounded distributive lattice expansions. Mathematica Scandinavica,
94:1345, 2004.
[55] Joan Gispert and Antoni Torrens. Bounded BCK-algebras and their generated variety. Mathematical
Logic Quarterly, 53(2):206213, 2007.
[56] Joan Gispert and Antoni Torrens. Boolean representation of bounded BCK-algebras. Soft Computing,
12(10):941954, 2008.
[57] Petr H ajek. Basic fuzzy logic and BL-algebras. Soft Computing, 2(3):124128, 1998.
[58] Petr H ajek. Metamathematics of Fuzzy Logic, volume 4 of Trends in Logic. Kluwer, Dordrecht, 1998.
[59] Petr H ajek, Llus Godo, and Francesc Esteva. A complete many-valued logic with product conjunction.
Archive for Mathematical Logic, 35(3):191208, 1996.
[60] Zuzana Hanikov a. A note on the complexity of propositional tautologies of individual t-algebras. Neural
Network World, 12(5):453460, 2002. Special issue on SOFSEM 2002.
[61] Ronald Harrop. On the existence of nite models and decision procedures for propositional calculi.
Proccedings of the Cambridge Philosophical Society, 54:113, 1958.
Chapter V: H ajeks BL Logic and BL-Algebras 453
[62] James B. Hart, Lori Rafter, and Constantine Tsinakis. The structure of commutative residuated lattices.
International Journal of Algebra and Computation, 12:509524, 2002.
[63] Pawe Idziak. Lattice operations in BCK-algebras. Mathematica Japonica, 29(6):839846, 1982.
[64] Afrodita Iorgulescu. Algebras of Logic as BCK algebras. ASE Editor, Bucharest, 2008.
[65] Afrodita Iorgulescu. Classes of examples of pseudo-MV-algebras, pseudo-BL-algebras and divisible
bounded non-commutative residuated lattices. Soft Computing, 14(4):313327, 2010.
[66] Kiyoshi Is eki and Shotaro Tanaka. An intoduction of the theory of BCK-algebras. Mathematica Japon-
ica, 23(1):126, 1978.
[67] Peter Jipsen and Franco Montagna. On the structure of generalized BL-algebras. Algebra Universalis,
55(23):226237, 2006.
[68] Peter Jipsen and Franco Montagna. The Blok-Ferreirim for normal GBL-algebras and its applications.
Algebra Universalis, 60(4):381404, 2009.
[69] Peter Jipsen and Franco Montagna. Embedding theorems for classes of GBL-algebras,. Journal of Pure
and Applied Algebra, 214(9):15591575, 2010.
[70] Bjarni J onsson and Alfred Tarski. Boolean algebras with operators. American Journal of Mathematics,
73:891939, 1951.
[71] Bjarni J onsson and Alfred Tarski. Boolean algebras with operators II. American Journal of Mathematics,
74:127162, 1952.
[72] Hitoshi Kihara and Hiroakira Ono. Interpolation properties, Beth denability properties and amalgama-
tion properties for substructural logics. Journal of Logic and Computation, 20(4):823875, 2010.
[73] Erich Peter Klement, Radko Mesiar, and Endre Pap. Triangular Norms, volume 8 of Trends in Logic.
Kluwer, Dordrecht, 2000.
[74] Yuichi Komori. Super-ukasiewicz propositional logics. Nagoya Mathematical Journal, 84:119133,
1981.
[75] Tomasz Kowalski and Tadeusz Litak. Completions of GBL-algebras: Negative results. Algebra Univer-
salis, 58(4):373384, 2008.
[76] Georg Kreisel. Explicit denability in intuitionistic logic. Journal of Symbolic Logic, 25(4):389390,
1960.
[77] Jan K uhr. Pseudo-BL algebras and PRl-monoids. Mathematica Bohemica, 128:199208, 2003.
[78] Laurentiu Leustean. Sheaf representations of BL-algebras. Soft Computing, 9(12):897909, 2005.
[79] Saunders MacLane and Ieke Moerdijk. Sheaves in Geometry and Logic. Springer, New York, 1992.
[80] Larisa L. Maksimova. Craigs theorem in superintuitionistic logics and amalgamable varieties of pseu-
doboolean algebras. Algebra i Logika, 16(6):643681, 1977.
[81] Larisa L. Maksimova. Modal logics and varieties of modal algebras: the Beth property, interpolation
and amalgamation. Algebra i Logika, 31(2):145166, 1992.
[82] George Metcalfe and Franco Montagna. Substructural fuzzy logics. Journal of Symbolic Logic,
72(3):834864, 2007.
[83] Franco Montagna. Interpolation and Beths property in propositional many-valued logics: A semantic
investigation. Annals of Pure and Applied Logic, 141(12):148179, 2006.
[84] Paul S. Mostert and Allen L. Shields. On the structure of semigroups on a compact manifold with
boundary. The Annals of Mathematics, Second Series, 65:117143, 1957.
[85] Daniele Mundici. Interpretations of AFC

-algebras in ukasiewicz sentential calculus. Journal of

Functional Analysis, 65:1563, 1986.
[86] Daniele Mundici. Free products in the category of Abelian -groups with strong unit. Journal of Algebra,
113(1):89109, 1988.
[87] Hiroakira Ono. Interpolation and the Robinson property for logics not closed under the Boolean opera-
tions. Algebra Universalis, 23(2):111122, 1986.
[88] Keith R. Pierce. Amalgamations of lattice ordered groups. Transactions of the American Mathematical
Society, 172:249260, 1972.
[89] Mark Theunissen and Yde Venema. MacNeille completions of lattice expansions. Algebra Universalis,
57(2):143193, 2007.
[90] Morgan Ward and Robert P. Dilworth. Residuated lattices. Transactions of the American Mathematical
Society, 45:335354, 1939.
[91] Andrzej Wro nski. BCK-algebras do not form a variety. Mathematica Japonica, 28(23):211213, 1983.
[92] Ronald R. Yager and Alexander Rybalov. Uninorm aggregation operators. Fuzzy Sets and Systems,
80(1):111120, 1996.
[93] Lot A. Zadeh. Fuzzy sets. Information and Control, 8(3):338353, 1965.
454 Manuela Busaniche and Franco Montagna
MANUELA BUSANICHE
Instituto de Matem atica Aplicada del Litoral
FIQ, CONICET-UNL
Guemes 3450
S3000GLN-Santa Fe, Argentina
Email: mbusaniche@santafe-conicet.gov.ar
FRANCO MONTAGNA
Department of Mathematics and Computer Science,
Pian dei Mantellini 44,
53100 Siena, Italy
Email: montagna@unisi.it
INDEX
-form, 146
3-sequent, 242
algebra
FL
e
-algebra, 222
Lindenbaum-Tarski, 192
Almost-Implicational Deduction Theo-
rem, 139
amalgam, 415
Amalgamation Property, 415
AP, 415
axiom, 107, 214
schema, 108, 214
axiomatic extension, 289
axiomatic system, 107
nitary, 107
basic substructural logic, 289
basis, 294
BCK
algebra, 445
bounded, 445
Beth property
deductive, 414
implicative, 414
BL
algebra, 363, 365
canonical, 405
complete, 426
congruence, 371
dense element, 372
lter, 370
n-potent, 379
quotient, 371
regular element, 371
standard, 303
axiomatic system, 359
chain, 366
FEP, 388
standard completeness, 364, 369
term, 365
BL-BCK-algebra, 446
Boolean algebra, 302, 311
Boolean element, 370
canonical model, 194
CEP, 415
CHL
hypersequent calculus, 263
standard semantics, 263
classical logic
sequent calculus, 219
closed set, 116
(nitely) -irreducible, 118
maximal w.r.t. an element, 118
saturated, 118
closure
operator, 116
nitary, 116
system, 116
base, 117
inductive, 116
completeness, 171, 322
nite strong, 171, 322
standard, 248, 354
strong, 171, 322
completion, 316, 427
Dedekind-MacNeille, 316, 319, 427
regular, 316
complexity of a formula, 214
congruence
logical, 114
congruence distributive, 293
Congruence Extension Property, 415
conjugate
iterated, 140, 306
left, 140, 306
right, 140, 306
456 INDEX
consecution, 106
consequence relation, 287
equational, 121, 291
predicate, 181
semantical, 110
conuclear contraction, 313
conucleus, 311, 384
core rule set, 224
cut elimination, 217, 236, 266
cut rule, 217
decidability, 242
deduction terms, 139
basic, 137
Dense Extension Property, 167
density
axioms, 273
elimination, 243, 267
rule, 243
Density Property, 167, 248
DEP, 167
derivation, 214
disjunction, 144
downset, 294
principal, 294
downset monoid, 316
DP, 167
dual well partial order, 333
dwpo, 333
embedding, 119
partial, 293
subdirect, 292
equation, 121
equational class, 291
evaluation, 110
expansion, 109
axiomatic, 109
conservative, 109
extension, 109
conservative, 241, 267
FEP, 323, 354, 388
lter, 109, 306
dense, 167
generated, 111
implicative, 370
generated, 371
maximal, 370
proper, 370
linear, 155
logical, 111
prime, 151, 415
lter-distributivity, 111
nitary companion, 109
Finite Embeddability Property, 388
nite embeddability property, 323, 354
nite tree, 214
rst-order quantiers, 265
FL
algebra, 299
commutative, 300
contractive, 300
cyclic, 301
integral, 300
involutive, 301
semilinear, 310
square-increasing, 300
axiomatic system, 136
chain, 299
Proof by Cases Property, 143
FL
c
algebra, 300
FL
e
algebra, 222, 300
axiomatic system, 136
Proof by Cases Property, 143
sequent calculus, 221
FL
ew
axiomatic system, 135
Proof by Cases Property, 143
FL
i
algebra, 300
FL
o
algebra, 300
forest, 382
formula
directed set, 172
predicate, 179
prenex, 267
INDEX 457
propositional, 106
fragment, 241
G odel logic
algebra, 311, 369
axiomatic system, 369
decomposition rules, 250
hypersequent calculus, 229, 250
sequent calculus, 252
sequent of relations calculus, 253
Galois connection, 295
GBL-algebra, 443
Giless dialogue game, 258
Glivenko theorem, 372
Goguen implication, 362
good sequence, 333
group
lattice ordered, 302
groupoid, 297
lattice ordered, 297
partially ordered, 297
residuated, 298
Herbrand theorem, 268
approximate, 271
Herbrand universe, 269
Heyting algebra, 302, 367
Higmans lemma, 333
homomorphism
matrix, 119
strict, 119
hoop, 364
basic, 365
cancellative, 365
totally ordered, 366
Wajsberg, 373
hypersequent, 222
atomic, 223
double p-marked, 245
formula translation, 231
marked, 237
p-regular, 245
single-conclusion, 223
hypersequent rule
local, 245
hypersequent version, 223
IMTL
algebra, 311
hypersequent calculus, 228
IMTL
n
hypersequent calculus, 230
inference, 107, 214
inference rule, 107, 214
interpolation, 274
Craig interpolation, 413
uniform left, 413
uniform right, 413
deductive interpolation property, 413
uniformleft deductive interpolation,
413
uniform right deductive interpola-
tion, 413
intuitionistic logic, 289
involutive pair, 303
IPC, 159
implicational fragment, 146
sequent calculus, 221
isomorphism, 119
IUL
algebra, 311
hypersequent calculus, 228
IUML
hypersequent calculus, 229
J onssons lemma, 293
join-dense set, 312
L, 184
L
w
, 201
L
m
, 184
L

, 170
L

, 160
L

, 152
ukasiewicz logic
witnessing axioms, 201
-group, 302
-monoid, 297
residuated, 299
language
458 INDEX
predicate, 179
propositional, 106
lattice, 214
division, 303
associative, 303
involutive, 303
proof system, 215
residuated, 301
lattice ordered Abelian group, 367
Leibniz
congruence, 114
operator, 115
LEP, 156
Lindenbaum
matrix, 109
Linear Extension Property, 156
logic
(almost) MP-based, 137
Abelian logic, 263
algebraically implicative, 122
algebraizable, 304
CHL, 263
classical, 218, 290
cross ratio logic, 263
disjunctional, 144
disjunctive, 144
lter-distributive, 111
nitary, 108
FL, 134
FL
e
, 135, 221
FL
ew
, 135
G odel-Dummett, 229, 249, 369
IMTL, 228
IMTL
n
, 230
inconsistent, 107
intuitionistic, 221, 289
IPC, 159
IUL, 228
IUML, 229
lattice-disjunctive, 144
ukasiewicz, 190, 254
MAILL, 214, 221
MALL, 221
modal fuzzy, 276
MTL, 228
MTL
n
, 230
NM, 230
p-disjunctional, 144
preSkolem, 198
product, 261, 370
propositional, 107
Rasiowa-implicative, 125
regularly implicative, 125
RM
ND
, 229
SBL, 370
semilinear, 155, 291, 307
SL, 126
SL
c
, 130
SL
e
, 130
SL
i
, 130
SL
o
, 130
SMTL, 230
substructural, 132, 221, 289
basic, 135, 289, 290
UL, 228
IUML, 229
weakly implicative, 113
WNM, 230
ukasiewicz logic
rst-order, 190, 270
game semantics, 258
hypersequent calculus, 254
hypersequent calculus for rst-order,
272
one variable fragment of rst-order,
271
sequent calculus, 257
standard semantics, 254
MAILL, 214
axiomatic system, 214
sequent calculus, 221
MALL
sequent calculus, 221
matching axioms and rules, 233
matrix
dense linear, 167
logical, 109
reduced, 114
mid-hypersequent theorem, 268
INDEX 459
mirror image, 128
modal fuzzy logics, 276
modal G odel logics, 277
modus ponens, 113
monoid
lattice ordered, 297
partially ordered, 297
monolith, 292
MTL
algebra, 311
standard, 303
hypersequent calculus, 228
implicational fragment, 241
MTL
Z
semilinearity, 166
MTL
n
hypersequent calculus, 230
multiset, 219
MV-algebra, 372
Chang, 427
standard, 362
MVP-algebra, 384
nearly term denable element, 341
negation, 300
negative cone, 313
NM
hypersequent calculus, 230
nuclear retraction, 313
nucleus, 311
o-hoop, 374
irreducible, 374
tower, 375
operator
closure, 294
interior, 294
ordinal sum, 362
p-disjunction, 144
p-protodisjunction, 143
parameter, 143
partial embeddability, 174
partial subalgebra, 293
partially embeddable, 294
PCP, 141, 144
Peirces law, 219
PEP, 151
PFEP, 415
polarity, 295
pomonoid, 297
residuated, 299
pomonoid pointed, 316
poset product, 380
poset relation, 296
nuclear, 318
poset sum, 380
predicate
consequence relation, 181
evaluation, 180
formula, 179
closed, 180
language, 179
model, 181
canonical, 194
exhaustive, 183
witnessed, 200
sentence, 180
structure, 180
safe, 181
term, 179
closed, 180
substitutable, 180
theory, 180
-Henkin, 197
-Henkin, 192
Henkin, 197
linear, 192
restricted Henkin, 195
variable, 180
bound, 180
free, 180
Prime Extension Property, 151
Prime Filter Extension Propert, 415
product
direct, 119
reduced, 119
subdirect, 120, 292
product logic
algebra, 370
460 INDEX
axiomatic system, 370
hypersequent calculus , 261
standard semantics, 261
proof, 108
in algebra, 111
Proof by Cases Property, 141, 144
FL, 143
FL
e
, 143
FL
ew
, 143
strong, 147
transfer theorem, 148
proof system, 214
analytic, 217
cut-free, 217
proper subalgebra, 341
propositional
consecution, 106
formula, 106
language, 106
logic, 107
theory, 107
variable, 106
protodisjunction, 143
psBL
algebra, 444
psMTL
r
algebra, 311
quantied G odel logic, 274
quantier
elimination, 274
propositional, 273
rules, 265
quantiers
shifting law, 265
quasi-equational class, 291
quasivariety, 292
relational hypersequent rules, 264
residual of a map, 295
residuated binary map, 298
residuated lattice, 366
bounded, 366
commutative, 222
divisible, 366
prelinear, 366
residuated map, 295
residuation property, 298
RL-algebra, 301
RM
ND
hypersequent calculus, 229
rule, 107, 214
admissible, 214
almost substitutive, 238
communication, 226
contraction, 219, 229, 289
external, 223
density, 243
derivable, 214
exchange, 289
invertible, 224
left weakening, 289
mingle, 229
mix, 228
nullary, 228
n-contraction, 230
quantier rules, 265
right weakening, 289
split, 227
structural, 219, 226
substitutive, 237
weakening, 219, 228
external, 223
SAP, 424
SBL
algebra, 370
axiomatic system, 370
semantics
rational, 177
real, 177
semigroup, 297
lattice ordered
residuated, 299
lattice ordered, 297
partially ordered, 297
semihoop
prelinear, 311
Semilinearity Property, 156
transfer theorem, 156
INDEX 461
sequent, 219, 288
atomic, 219
atomic implicational, 250
of relations, 253
single-conclusion, 219
sequent of relations, 279
shifting modality axioms, 277
single-conclusion version, 221, 223
Skolem form, 270
Skolemization, 268, 271
SL
axiomatic system, 135
SL
c
axiomatic system, 130
SL
e
axiomatic system, 130
SL
i
axiomatic system, 130
SL
o
axiomatic system, 130
SLP, 156
MTL
hypersequent calculus, 230
sPCP, 147
strictly simple algebra, 341
Strong Amalgamation Property, 424
subcover, 324
subdirect product, 292
subdirect poset product, 382
subdirect representation, 120
subdirectly irreducible, 292
subformula property, 217, 226
submatrix, 119
subquantale, 385
substitution, 106
substructural logic, 289
sudirectly irreducible, 120
nitely, 120
t-norm, 302, 362
ukasiewicz, 362
continuous, 362
G odel, 362
product, 362
theory
deductively directed set of theories,
195
-Henkin, 197
-Henkin, 192
Henkin, 197
predicate, 180
propositional, 107
restricted Henkin, 195
transfer theorem, 117
Proof by Cases Property, 148
Semilinearity Property, 156
truth stressers, 276
UL
algebra, 311
hypersequent calculus, 228
ultraproduct, 119
IUML
hypersequent calculus, 229
uniform interpolant, 413
upset, 294
principal, 294
variety, 292
semilinear, 307
V -formation, 415
weak implication, 113
weakly implicative logic, 113
Whitmans condition, 218
witnessed
extension, 201
model, 200
witnessing axioms, 200
WNM
hypersequent calculus, 230