Você está na página 1de 563

A Model Theoretic Oriented Approach to

Partial Algebras
Peter Burmeister
1
December 8, 2002
1
This book was published 1986 by the Akademie-Verlag Berlin, Volume 32 in
the series Mathematical Research; Lector was Dr. Reinhard Hoppner. It was
printed in the German Democratic Republic. It has been transferred into L
A
T
E
X
by Ulrich Thiemann. In connection with this translation also the notation for
the direction used for the composition of homomorphisms has been changed (the
rst morphism is written to the right of the second one, etc.). So far most of the
diagrams are still missing. Also an index is still missing. The errors and misprints
in the original version have not yet been corrected. The bibliography has been
extended about 1990 and not yet been updated moreover, it may still contain
many titles, which treat partial operations (which was the keyword for the search
in the Zentralblatt), but which are not really concerned with partial algebras in
the sense of this book. In order to get a better idea of the material in the book,
the article: Tools for a Theory of Partial Algebras, which has been published in:
General Algebra and Applications (Eds.: K.Denecke and H.-J.Vogel), Research
and Exposition in Mathematics, Vol. 20, Heldermann Verlag Berlin, 1993, pp.
1232, has been added to a revised introduction.
December 8, 2002
2 Preface
Preface
This book is the rst of two parts about partial algebras. The second
part [Re84] (Structural Induction on Partial Algebras) has been written by
H.REICHEL and provides several applications to computer science. More-
over, the theory presented there is more specially tailored to the interests of
computer science and contains some deeper treatment of some special partial
algebras, so called hierarchical partial algebras.
In comparison to the second part the general theory of partial algebras
in this rst one is developed to a broad extent (as far as it already exists),
although not all features which have already been studied can be presented
within the scope of this book (see section 22 for some more details). Thus
choices have been made according to the interests of the author and of some
colleagues with whome he has a close collaboration. In this connection espe-
cially H.Andreka, I.Nemeti and I.Sain as well as B.Wojdylo and A.Pasztor
should be mentioned whose encouragement to a great extent nally caused
the author to write this book.
Actually the idea of a book on universal algebra which also treats partial
algebras as fundamental structures had already been pursued at the end of
the sixties by the now deceased teacher of the author, J urgen Schmidt, who,
however, never nished this project. One reason might have been that in
the early seventies there seemed to be almost no larger interest or possibility
of applications for partial algebras. This, however, has changed during the
last years, when the necessity of a more rigorous theoretical foundation of
computer science and the growing engagement in abstract (algebraic) data
types as they are e.g. used in programming languages like PASCAL or
ADA caused among computer scientists also a growing interest in universal
algebra. As a next step some of them also turned their interest towards
partial algebras, since many operations occurring in computer science are
only partial: Operations and the evaluation of functions or expressions used
or dened in programs sometimes are not everywhere dened either because
of the delimitations given by the used computer or because of the denition
of the expressions. Moreover, the abstract theory of algorithms shares many
features with the one of partial algebras (see [Ti81]), and this led H.THIELE
in [Th66] to develop a logic for partial algebras which is basically the one we
use here. Further applications or hints to corresponding literature (see
e.g. [AN79f] as such a source) can also be found in the literature mentioned
in subsection 22.2.
December 8, 2002
Preface 3
A quite old source of interest in the study of partial operations is the one
concerning partial recursive functions (see e.g. [Kl52], [Mal74] and [Hoe72],
also [Ku77]). Another one may come from quantum mechanics, e.g. quantum
logic, in theoretical physics, where the fact that not all phenomena connected
with the objects of interest can be observed and described at the same time
gives rise to partial operations (see [KoSp68]).
Despite a possible growing interest in features of partial algebras only
few people have investigated or really used these structures so far, since
most possible users had great reservations against these structures whose
treatment seemed to be so much more complicated than the one of total
algebras and for which they did not see much of a theory around. As a
matter of fact almost every concept from the theory of total algebras often
splits into more than three corresponding concepts for partial algebras, and
for a long time it seemed hard to distinguish some rules behind it but this
has changed. On the other hand one has either the possibility of completing a
partial algebra in a more or less satisfactory way, or one can consider partial
algebras as special relational systems whose theory provides some tools in
order to handle such structures but not always the most adequate ones.
Nevertheless there has been developed in particular during the last
decade a theory of partial algebras which may really be a starting point
for the application of these structures. It is based on category and model
theoretical methods and provides some insight into the reasons for the fact
that there really is such a wealth of concepts around for partial algebras
one reason for this richness being the fact that the theory of relational
systems can totally be embedded into the theory of partial algebras. But the
theory of partial algebras is much more appropriate to the task of dealing
with relations which in fact are the graphs of functions than the theory of
relational systems is. Thus we think it to be worth-while to write a book
like this one which introduces into the theory of partial algebras and shows
up their special features in comparison to total algebras. One further reason
is the fact that out of the theory of partial algebras one may get a better
understanding of many features of so-called many-sorted or heterogeneous
algebras which are those structures mostly used in computer science. One
may say that this is too large a detour for the understanding of heterogeneous
algebras, but as long as one has not yet decided on how to treat errors, even
by using many-sorted algebras one cannot avoid partiality as the common
example of the abstract data type STACKS (see the introduction and later
sections) shows. And partial algebras include all such structures, especially
December 8, 2002
4 Preface
when one uses the approach via (what we shall call) partial S-algebras which
is only sort of a formal but very useful further generalization, all of whose
nal applications cannot yet be foreseen.
The present book has another feature which makes it dierent from al-
most any other presentation of universal algebra so far, namely the fact that
also innitary operations are allowed. Although this requires at some places
some more technical details, we hope that this fact does not burden the under-
standability too much, in particular since we can avoid almost everywhere the
treatment of ordinal numbers (following J.Schmidt). Nevertheless on some
occasions one has to deal with cardinality restrictions, where in the nitary
case the occurring bounds usually can be replaced by the cardinality #
0
of
the set of all natural numbers. One reason for allowing innite arities is the
possible application of innitary structures in connection with so-called
0
-
posets, which are used in some part of computer science (cf. D.LEHMANN,
A.PASZTOR [LePa80] and A.PASZTOR [Pa83], [Pa84]). Innitary opera-
tions also arise in possible applications in topology (see G.A.EDGAR [Ed73]),
or in the theory of ordered sets.
The objection might be raised that in this book we do not provide enough
really important applications for the theory presented. There might be some
point to this, since this is the more theory oriented part of a two-volume
project, and for some more special part of the theory of partial algebras the
second volume of this project (written by H.REICHEL) provides some really
important examples; however, the special theory presented there cannot cover
all their features. Therefore the general theory provided here shows how one
can proceed in situations not covered by the more special theory. Moreover
we hope that the existence of a theory for partial algebras like this one will
encourage possible users to produce new applications. Although we do not
know what a possible user will really need, we hope that in this book he will
nd enough information in order to develop by himself the special part of the
theory necessary for him. And thus we touch a further intention of this book
together with its second part , namely to encourage other mathematicians
or computer scientists to use these two books as a platform from which they
can start a further development of the theory.
Such a rst treatment and representation of a very general theory like
this one on partial algebras will probably not be all around as smooth as one
would like it to be, and we beg the reader in advance to pardon this. The rst
chapter of this book has to a great extent already been the contents of lectures
of J.Schmidts at Bonn (Germany) in 1965/66 (see [Sch66b]) and Athens
December 8, 2002
Preface 5
(Georgia, USA) in 1967/68 (see [Sch68]), as well as of the authors at Bonn
in 1971 and at Darmstadt for several times, while most of the material in the
subsequent chapters has not yet been presented in this form and framework
before; and it contains at some places research results not yet presented
anywhere else. Only a rough overview of the guiding ideas has been given in
[B82] (see also [AN83a] for a considerable portion of the logic and category
oriented aspects), where especially the model theoretical aspects of the theory
have been presented (without proofs). However the whole presentation is
new as far as it is based on the study of what we call partial S-algebras.
Therefore we might not have found the best way of representation right at
the beginning. On the other hand we wanted this book to appear as quickly
as possible because of the other reasons mentioned above. If not every one
who contributed to the theory of partial algebras presented here has received
due credit, we apologize for that: it is not due to bad will but only to lack
of space (e.g. as far as the schools of J.Slomi nski and H.-J.Hoehnke are
concerned), to some forgetfulness or even to ignorance on part of the author.
Let us also mention that we do not treat relations separately, since they can
easily be encoded as partial operations without any loss of algebraic features;
however, observe that in this context the class of all relational systems of some
given type is not an equational but only an implicational subvariety of the
class of all partial algebras of the corresponding type.
Finally I want to express my gratitude to all those who have supported
this project. In the rst place I have to mention my family for its patience
and understanding, and in particular my wife Hildegund who also did a lot
of proof-reading and lled in most of the special symbols. Moreover I thank
Hans-J urgen Hoehnke, Hajnal Andreka, Istvan Nemeti, Ildiko Sain, Ana
Pasztor and Manuel Siegmund-Schultze for valuable remarks and discussions,
and also together with Peter Winkler, Peter Alles and Wolfgang Thumser
for proof-reading all or parts of the book. I should not forget to thank
all those who helped me getting started on text processing on a personal
computer. I also want to thank the Akademie-Verlag for publishing this
book, e.g. Dr. R. Hoppner for his patience and his slight pressure.
Darmstadt, July 1985 Peter Burmeister
December 8, 2002
6 Preface
December 8, 2002
Contents
0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
0.1 Historical remarks . . . . . . . . . . . . . . . . . . . . 19
0.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 23
0.3 Similarity types and partial algebras . . . . . . . . . . 33
0.4 Terms, existence equations, elementary implications . . 36
0.5 Preservation and reection of formulas homomor-
phisms and some of their properties . . . . . . . . . . . 42
0.6 Modeling many-sorted (partial) algebras . . . . . . . . 45
0.7 Encoding elementary implications by epimorphisms . . 51
0.8 Factorization systems . . . . . . . . . . . . . . . . . . . 53
0.9 A Meta Birkho Theorem . . . . . . . . . . . . . . . . 56
I Basic denitions and properties 61
1 Homogeneous and heterogeneous partial algebras . . . . . . . 62
1.1 Partial algebras . . . . . . . . . . . . . . . . . . . . . . 63
1.2 Heterogeneous partial algebras . . . . . . . . . . . . . . 67
1.3 Cardinal and ordinal dimension . . . . . . . . . . . . . 72
2 Homomorphisms etc. . . . . . . . . . . . . . . . . . . . . . . . 74
2.1 Homomorphisms . . . . . . . . . . . . . . . . . . . . . 75
2.2 Isomorphisms . . . . . . . . . . . . . . . . . . . . . . . 77
2.3 Endomorphisms and automorphisms . . . . . . . . . . 80
2.4 Full and closed homomorphisms . . . . . . . . . . . . . 81
2.5 Congruence relations . . . . . . . . . . . . . . . . . . . 85
2.6 Factor algebras . . . . . . . . . . . . . . . . . . . . . . 91
2.7 The Homomorphism Theorem . . . . . . . . . . . . . . 93
2.8 The congruence lattice of a full homomorphic image . . 98
2.9 The analogue of the Second Isomorphism Theorem of
E.Noether . . . . . . . . . . . . . . . . . . . . . . . . 99
7
8 CONTENTS
2.10 Mono- and epimorphisms . . . . . . . . . . . . . . . . . 101
2.11 Special observations concerning partial S-algebras . . . 102
3 Substructures, closed subsets, generation . . . . . . . . . . . . 106
3.1 Substructures of partial algebras . . . . . . . . . . . . . 106
3.2 Comparison of substructures . . . . . . . . . . . . . . . 112
3.3 The global process of generation . . . . . . . . . . . . . 114
3.4 Local generation (generation from below) . . . . . 115
3.5 Algebraic Induction . . . . . . . . . . . . . . . . . . . . 118
3.6 Generation, Epimorphisms . . . . . . . . . . . . . . . . 119
3.7 The analogue of E.Noethers First Isomomorphism The-
orem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4 Products and Limits . . . . . . . . . . . . . . . . . . . . . . . 125
4.1 (Direct) Products (homogeneous case) . . . . . . . . . 125
4.2 Pullbacks and products of partial S-algebras . . . . . . 131
4.3 (Global) Coproducts and multiple pushouts . . . . . 135
4.4 Direct limits (= directed colimits) . . . . . . . . . . . . 140
4.5 Reduced products and ultraproducts . . . . . . . . . . 147
4.6 Some properties of the operators H, .., o, .. and T, .. . . 150
5 Universal solutions, Freeness, Terms . . . . . . . . . . . . . . . 159
5.1 The general concepts of universal solution and freeness 160
5.2 Partial Peano algebras . . . . . . . . . . . . . . . . . . 165
5.3 Free completions . . . . . . . . . . . . . . . . . . . . . 170
5.4 The Generalized Recursion Theorem . . . . . . . . . . 173
5.5 The algebraic quasi-ordering on partial algebras . . . . 174
5.6 Tree representation of terms . . . . . . . . . . . . . . . 177
5.7 (Global) Term operations and polynomials . . . . . . . 179
5.8 The Epimorphism Theorem . . . . . . . . . . . . . . . 184
5.9 The initial closed congruence relations of a free com-
pletion . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5.10 On the cardinalities of M-generated partial algebras . . 188
5.11 On the existence of K-universal (K-) solutions . . . . . 190
6 The equational theory of partial algebras . . . . . . . . . . . . 195
6.1 Existence-equations . . . . . . . . . . . . . . . . . . . . 196
6.2 Description of the model classes (E-varieties) . . . . . . 201
6.3 More facts about freeness, fully invariant congruences . 207
6.4 Characterization of E-equational theories . . . . . . . . 211
December 8, 2002
CONTENTS 9
II A Basis for a two-valued model theory for partial algebras 225
7 A rst order language for (innitary) partial (S-) algebras . 227
7.1 The syntax . . . . . . . . . . . . . . . . . . . . . . . . 227
7.2 The semantics . . . . . . . . . . . . . . . . . . . . . . . 233
7.3 Coincidence and Substitution Theorem . . . . . . . . . 235
8 Elementary implications, universal Horn formulas . . . . . . . 241
8.1 Some basic kinds of universal Horn formulas . . . . . . 241
8.2 Algebraic encoding of universal Horn formulas . . . . . 247
8.3 Preservation properties of algebraic operators rst
visit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
8.4 Diagonal-ll-in property and projectivity motivation 253
9 Preservation (reection) of formulas by mappings . . . . . . . 256
9.1 Preservation and reection . . . . . . . . . . . . . . . . 256
9.2 Examples for the reection of formulas by homomor-
phisms . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
10 Factorization systems of partial algebras . . . . . . . . . . . . 271
10.1 The Galois connection induced by the unique-diagonal-
ll-in property . . . . . . . . . . . . . . . . . . . . . . . 271
10.2 Factorization systems . . . . . . . . . . . . . . . . . . . 275
10.3 Examples of factorization systems for partial algebras . 286
III Birkho type results for elementary implications 297
11 Direct limits revisited . . . . . . . . . . . . . . . . . . . . . . . 298
11.1 Morphisms between directed systems . . . . . . . . . . 298
11.2 A device to measure smallness . . . . . . . . . . . . 301
11.3 Direct limits and the operator . . . . . . . . . . . 303
11.4 Sections, retractions and direct limits . . . . . . . . . . 306
11.5 Some further concepts connected with direct limits . . 307
11.6 Proof of Theorem 8.3.4 . . . . . . . . . . . . . . . . . . 308
12 A Meta Birkho Type Theorem . . . . . . . . . . . . . . . . . 312
12.1 HoT
r,k
-cones . . . . . . . . . . . . . . . . . . . . . . . 312
12.2 A Meta Birkho Theorem . . . . . . . . . . . . . . . . 314
12.3 Compatible classes H and o in Alg(S) . . . . . . . . 324
13 Construction of examples for the Meta Birkho Theorem . . . 330
13.1 k-dual algebroidalness of
op
(S) for some classes o of
Alg(S)-morphisms . . . . . . . . . . . . . . . . . . . . . 330
13.2 H-projective partial S-algebras . . . . . . . . . . . . . 333
13.3 Examples in Alg(S) for the Meta Birkho Theorem . . 337
December 8, 2002
10 CONTENTS
13.4 Special examples . . . . . . . . . . . . . . . . . . . . . 342
14 Closed classes of epimorphisms (elementary implications) . . . 348
14.1 Preparatory results . . . . . . . . . . . . . . . . . . . . 348
14.2 The general theorem . . . . . . . . . . . . . . . . . . . 350
14.3 Closed sets of E-, ECE- and QE-equations and of H
f
o
c
T
r,k
-
implications . . . . . . . . . . . . . . . . . . . . . . . . 353
14.4 Closed sets of ECH- and Horn-formulas . . . . . . . . . 362
IVSpecial Topics 365
15 Algebraic operations . . . . . . . . . . . . . . . . . . . . . . . 366
15.1 Superposition . . . . . . . . . . . . . . . . . . . . . . . 366
15.2 Algebraic partial operations of the rst and second kind372
15.3 Algebraic partial operations of the third kind . . . . . 378
15.4 Algebraic total operations . . . . . . . . . . . . . . . . 384
15.5 More about algebraic partial operations and terms; al-
gebraic partial functions . . . . . . . . . . . . . . . . . 388
15.6 Algebraic partial operations and homomorphisms . . . 390
16 Semilattices of (weak) completions . . . . . . . . . . . . . . . 392
16.1 The semilattice of (weak) minimal completions . . . . . 393
16.2 The complete lattice of normal minimal completions . . 397
17 Independence, free partial algebras and primitive classes . . . 401
17.1 Some properties of independence and independence classes401
17.2 Relatively free completions of partial algebras . . . . . 403
17.3 Relatively free completions of relatively free partial al-
gebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
17.4 Characterization of heterogeneous relatively free par-
tial algebras . . . . . . . . . . . . . . . . . . . . . . . . 415
17.5 Generators and relations . . . . . . . . . . . . . . . . . 418
17.6 Initial objects and representation by canonical terms . 420
18 Dependence of operations on arguments and element basis
property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
18.1 General concepts . . . . . . . . . . . . . . . . . . . . . 427
18.2 Dependence on variables and algebraic independence . 431
18.3 A counterexample for the innitary homogeneous case 437
19 Congruence relations revisited . . . . . . . . . . . . . . . . . . 439
19.1 Closed congruence relations of nitary partial algebras 439
19.2 P-reecting homomorphisms and P-closed congruence
relations . . . . . . . . . . . . . . . . . . . . . . . . . . 445
December 8, 2002
CONTENTS 11
20 Subdirect representations of partial algebras . . . . . . . . . . 450
20.1 The general concept and remarks . . . . . . . . . . . . 450
20.2 Classes of full and surjective homomorphisms arbi-
trary monomorphisms . . . . . . . . . . . . . . . . . . 452
20.3 The cases of full respectively closed monomorphisms . 457
20.4 Subdirectly irreducible mono-unary partial algebras . . 462
21 More facts about E-varieties . . . . . . . . . . . . . . . . . . . 466
21.1 Initial homomorphic extensions . . . . . . . . . . . . . 466
21.2 On the lattice of initial closed congruence relations
homogeneous case . . . . . . . . . . . . . . . . . . . . . 472
21.3 Relativized algebraic partial operations homogeneous
case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
481
22 Survey on some additional concepts and results . . . . . . . . 482
22.1 Other concepts of homomorphisms . . . . . . . . . . . 482
22.2 Further topics and hints to applications . . . . . . . . . 485
23 Some denitions and facts from set theory . . . . . . . . . . . 488
23.1 Some remarks on notations . . . . . . . . . . . . . . . 488
23.2 Some remarks on an axiomatic approach . . . . . . . . 493
23.3 Ordinals . . . . . . . . . . . . . . . . . . . . . . . . . . 497
23.4 Cardinals . . . . . . . . . . . . . . . . . . . . . . . . . 501
24 Ordered sets, algebraic lattices and Galois connections . . . . 505
24.1 Partially ordered sets and lattices . . . . . . . . . . . . 505
24.2 k-directed posets, k-inductivity, k-algebraic lattices . . . 507
24.3 Closure systems and closure operators . . . . . . . . . 510
24.4 Galois connections . . . . . . . . . . . . . . . . . . . . 513
December 8, 2002
12 CONTENTS
December 8, 2002
List of Tables
0.1 Compatible pairs (H, o) (see Proposition 0.9.4) . . . . . . . . 59
0.2 Premises and conclusions for some special operators . . . . . . 60
1.1 Ordinal and cardinal dimensions and rank . . . . . . . . . . . 73
8.1 Preservation (via validity) of formulas by algebraic operators . 251
9.1 Some properties of homomorphisms denable by reection of
formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
10.1 Factorization systems in TAlg(S) . . . . . . . . . . . . . . . . 294
12.1 Compatibility of the operators H and o . . . . . . . . . . . . 328
13.1 Pairs of operators H, o . . . . . . . . . . . . . . . . . . . . . . 338
13.2 Shape of the implications (see 13.3.3) . . . . . . . . . . . . . . 340
20.1 Positive cases of (c, /)-subdirect . . . . . . . . . . . . . . . . 462
13
14 LIST OF TABLES
December 8, 2002
List of Figures
0.1 Example of a partial algebra . . . . . . . . . . . . . . . . . . . 34
0.2 Layer model of generation from below . . . . . . . . . . . . . . 35
0.3 Example of a (partial) interpretation . . . . . . . . . . . . . . 41
0.4 e

is the epimorphism encoding the elementary implication . . 52


0.5 ftools3.2=FKext; e

is an A-extendable epimorphism, . . . . 52
0.6 ftools3.3=Frebf; Reection by m of the el. implication . . . . 53
0.7 ftools4.1; P is H-projective . . . . . . . . . . . . . . . . . . . . 56
1.1 Example of the specication of a partial algebra by a drawing 68
2.1 Compatibility property of a many-sorted homomorphism . . . 76
2.2 Bijective endomorphism, but no isomorphism . . . . . . . . 79
2.3 Examples of properties of homomorphisms . . . . . . . . . . . 83
2.4 f and g are full, but gf is not full. . . . . . . . . . . . . . . . 84
2.5 The equivalence generated by two congruences is in general no
congruence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.6 A four element lattice . . . . . . . . . . . . . . . . . . . . . . . 93
2.7 Diagram completion for surjective mappings . . . . . . . . . . 94
2.8 Dual homomorphism theorem for full and injective f . . . . . 95
2.9 Two diagrams corresponding to Theorem 2.7.4 . . . . . . . . . 96
2.10 f
(2)
[
1
] is a congruence on B . . . . . . . . . . . . . . . . . . 99
2.11 Eect of a (closed) quotient homomorphism on the (closed)
congruences . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2.12 An analogue of the Second Isomorphism Theorem . . . . . . . 100
3.1 f3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.2 Layer model of generation from below (nitary case) . . . . . 117
3.3 A non-closed homomorphism preserving closed subsets . . . . 121
3.4 An analogue of the First Isomorphism Theorem . . . . . . . . 123
3.5 Necessity of the assumptions of closedness . . . . . . . . . . . 124
4.1 The induced product morphism . . . . . . . . . . . . . . . . . 125
15
16 LIST OF FIGURES
4.2 A closed graph in the product need not represent a homomor-
phism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.3 About the maximal (closed) homomorphic extension . . . . . 130
4.4 h: The induced multiple pullback morphism . . . . . . . . . . 132
4.5 g: The induced coproduct morphism . . . . . . . . . . . . . . 135
4.6 Example of a coproduct with a non-injective canonical injec-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.7 The induced morphism for direct limits . . . . . . . . . . . . . 141
4.8 Iteration of reduced products (using direct limits) . . . . . . . 154
5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.3 This should be Figure 5.3 . . . . . . . . . . . . . . . . . . . . 178
5.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.6 t
A
(w) = a . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
5.7 t
A
(w

) does not exist . . . . . . . . . . . . . . . . . . . . . . . 182


5.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
6.1 f6.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
6.2 f6.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.3 f6.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.4 f6.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.5 f6.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
6.6 f6.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
6.7 f6.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
6.8 f6.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
8.1 f8.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.2 f8.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.3 f8.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
8.4 f8.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.5 f8.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
8.6 f8.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
9.1 f9.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
10.1 The diagonal-ll-in property and the operators and
op
. . . 272
10.2 ad (ii) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
10.3 ad (v)/(vi) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
10.4 ad (vii)/(viii) . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
10.5 ad (ix) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
December 8, 2002
LIST OF FIGURES 17
10.6 ad (x) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
10.7 f10.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
10.8 f10.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
10.9 f10.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
10.10f10.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
10.11f10.11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
10.12f10.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
10.13f10.13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
10.14f10.14 (old f 10.1) . . . . . . . . . . . . . . . . . . . . . . . . . 283
10.15f10.15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
10.16f10.16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
10.17f10.17 (old f10.2) . . . . . . . . . . . . . . . . . . . . . . . . . 290
10.18f10.18 (old f10.3) . . . . . . . . . . . . . . . . . . . . . . . . . 290
10.19f10.19 (old f10.4) . . . . . . . . . . . . . . . . . . . . . . . . . 291
10.20f10.20 (old f10.5) (see 10.3.12) . . . . . . . . . . . . . . . . . . 295
11.1 f11.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
11.2 f11.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
11.3 f11.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
12.1 f12.1 (see 12.1.2.(i)) . . . . . . . . . . . . . . . . . . . . . . . . 313
12.2 f12.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
12.3 f12.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
12.4 f12.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
15.1 f15.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
15.2 f15.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
15.3 f15.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
17.1 f17.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
17.2 f17.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
17.3 f17.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
18.1 f18.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
18.2 f18.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
19.1 f19.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
19.2 f19.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
19.3 f19.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
20.1 f20.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
21.1 f21.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
December 8, 2002
18 LIST OF FIGURES
22.1 f22.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
22.2 f22.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
24.1 f24.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
24.2 f24.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
December 8, 2002
Introduction 19
0 Introduction
0.1 Historical remarks and motivation
Algebra has a long history which started with the investigation of rules for
very special calculations with natural numbers more than three thousand
years ago. Later (positive) rational numbers and calculations with them
were added, irrational numbers were discovered about two thousand years
ago, zero and negative numbers were added some centuries ago, and only
about two centuries ago it was accepted that complex numbers form a suit-
able extension of real numbers. During all those centuries algebra was the
mathematical theory which mainly dealt with the solvability of equations in
special ranges of numbers and with the description of such solutions. And the
extension of the world of numbers at each stage was caused by the desire
to add solutions of the equations under consideration mostly linear equa-
tions or polynomial eqations of low degree to the existing and accepted
domain (i.e. to extend or complete a given special partial algebra). More-
over, one wanted to transfer the rules of calculations as far as possible also to
the extended domain. Finally with the introduction and wide acceptance of
complex numbers a stage was reached, where every linear or polynomial equa-
tion was theoretically solvable within this domain, and other questions
arose, in particular the one of which polynomial equations could be solved
by expressions using n-th roots and rational operations only. This led
among others to the introduction and investigation of abstract groups,
elds and rings, independently from a special concept of numbers. Thus
quite dierent problems than before intrigued algebraists: construction, in-
vestigation and description in particular classication of axiomatically
dened algebraic structures. While, at the beginning, this mainly concerned
structures close to groups, rings, elds, modules and vector spaces, in the
second half of the 19th century the development of mathematical logic and
its algebraization e.g. by G. Boole, C.S. Peirce and E. Schr oder led
to the investigation of quite new structures like Boolean algebras or lattices,
and methods nowadays contributed to universal algebra were already used.
In 1889 A.N. Whitehead wrote a book
1
A treatise on universal algebra,
where the idea of universal algebra was founded, yet no signicant results in
this direction provided. Only some thirty years later abstraction in algebra
1
See [?].
December 8, 2002
20 Introduction
went so far as to investigate sets with (mostly nitary) operations in general
(up to isomorphism) and to get relevant results , and this is sometimes
quoted as the hour of birth of universal algebra, which is usually closely
related to the paper [Bi35] of Garrett Birkhoff. In this paper the general
denition of an abstract algebra already allows partial operations
2
, while
he speaks of uniform algebras, if each operation is dened exactly on all
sequences (of the carrier set) of some xed length. The results and examples
in [Bi35], however, only concern uniform algebras (i.e. what we shall call total
algebras).
At the beginning universal algebra was pushed forward at least to a
large extent just in order to nd a general background for the dierent al-
gebraic structures arising within mathematics itself, partly supported by the
hope that such a theory could be useful in connection with problems within
special algebraic theories. But it has turned out that some of these hopes
could not be fullled; universal algebra rather provides problems for classical
algebra than that it solves problems for classical algebra. Yet it proves to
be useful in particular when new algebraic structures arise, since it provides
all the known results common to all algebraic structures (in general or un-
der certain restrictions like being equationally or implicationally dened).
Thus, during the last two decades, it has proved to be quite useful in par-
ticular for some aspects of theoretical computer science, where computers,
programming languages, programs or abstract data types can be interpreted
as, be described by or be better understood when considered as algebraic
structures.
Until about 1964 universal algebras have with few exceptions been
considered with total and nitary operations only, while then almost simul-
taneously J. Slomi nski in Poland and J. Schmidt in Germany started to
allow both innitary and partial operations in their investigations. Among
their motivations may have been besides trying to reach utmost generality
in their concepts and results the following ones:
In 1953 T. Evans, proved in [Ev53] the equivalence of the solvability
of the word problem for varieties of total universal algebras and of
the decidability of the embeddability problem w.r.t.
3
the variety under
consideration for nite partial algebras satisfying all those formulas of
2
The denition in [Bi35] of the proper domain of an operator of an abstract algebra
even allows to include sequences of dierent lengths, what we shall not do in what follows.
3
We abbreviate with respect to by w.r.t..
December 8, 2002
0 Introduction 21
some special kind (which can be interpreted w.r.t. special semantics as
some kind of equations
4
), which are valid in the given variety.
S.C. Kleene introduced in 1952 in his book Introduction to Meta-
mathematics
5
three kinds of equalities for partial recursive func-
tions, which later reappeared in Slomi nskis papers, although it seems
that he did not know Kleenes denitions. It can be said in gen-
eral that the study of partial recursive functions motivated quite a few
mathematicians, like H.-J. Hoehnke and his school, to investigate
partial algebras.
In 1963 G. Gr atzer and E.T. Schmidt used partial algebras as
starting point for their constructions in order to prove that every non-
trivial algebraic lattice is isomorphic to the congruence lattice of some
total universal algebra. In this connection they also had to prove
some elementary results about partial algebras. Here a general method
showed up to construct total algebras with prescribed properties by
starting with an easily constructed partial algebra and to complete it by
using some universal constructions. This is e.g. a well-known method in
group theory to describe groups by generators and relations (meaning
in the terminology of universal algebra of partial algebras to construct
by the generators and the terms in the relations a partial algebra, to
factorize it by the congruence relation generated by the relations, and
then to nd its universal solution within the given variety of groups).
In the early sixties category theory was developing quite rapidly, and
(small) categories are just a very special instance of partial algebras.
Another motivation for an investigation of sets with partial operations
stems from algorithm theory. In particular in 1966 H. Thiele developed in
connection with algorithm theory a logic for partial algebras
6
, which is the
same that was (re)developed (without knowing about Thieles approach)
during the following ten years by H. Hoeft, P. Burmeister, and by
H. Andr eka, I. N emeti and I. Sain. It is basically the one, which we use
in this book.
4
Cf. Evans equality in section 8.
5
See [Kl52].
6
See [Th66].
December 8, 2002
22 Introduction
The investigations of J. Slomi nski and of J. Schmidt have then been
continued and extended during the late sixties and early seventies (and in
part also later) by their students B. Wojdylo, H. H oft and P. Burmeis-
ter. During that period also R. Kerkhoff and V. Poythress worked on
partial algebras, e.g. in connection with the so-called weak equality. And in
1968 S. Kochen and E.P. Specker applied in [KoSp68] partial algebras
and this weak equality in a description of some phenomena in quantum me-
chanics. In the early seventies H. Reichel and H. Kaphengst, motivated
by the applicability of their results in computer science, worked on partial
algebras. A similar motivation led H. Andr eka, I. N emeti, I. Sain and
A. Pasztor to investigate partial algebras. However, they started their
research via category theory in order to achieve utmost generality, i.e. al-
though the intended examples for their results are mainly partial algebras,
these results apply to all categories satisfying some general conditions
7
Starting with the soft ware crisis in computer science in the seventies
theoretical computer scientists got a growing interest in universal algebra,
however mainly in many-sorted algebras
8
(which somehow try to hide par-
tiality, or at least to make the still remaining exceptional cases more easily
handable). And in connection with this development interest in many-sorted
partial algebras was growing, too. We refer to the list of references for the
names to be mentioned in this connection. However, we want to end these
historical remarks by mentioning W. Craigs paper [Cg89], part I, in which
he indicated that the kind of implications, which we call existentially condi-
tioned existence equations or briey ECE-equations are (almost
9
) equivalent
to what he calls Kleene-equalities, if one only adds in the one-sorted
case to the given similarity type a binary operation symbol, say , as a
logical constant, which is always to be interpreted as a total rst projection.
This result somehow rounds up the scenery, since it proves, what we always
have been feeling, that ECE-equations are the appropriate substitute in the
case of partial algebras for equations in the case of total algebras (namely,
Kleene-equality between two terms t and t

means the equality of the graphs


7
See e.g. [AN83a], [AN79a] and [NSa82], the latter two papers have given rise to a Meta
Birkho Type Theorem, a version of which we have included in section 12 of [B86]), and
which will also be treated in volume 2 of this book.
8
See e.g. the papers of the so-called ADN-gpoup, e.g. [GoTcWa78], which is to be
considered as one of the mile stones in this development.
9
For Craig and most other logicians and mathematicians this is indeed a complete
equivalence, since they only consider non-empty models.
December 8, 2002
0 Introduction 23
of the term operations induced by t and t

, respectively, in each model of the


Kleene-equality t
K
= t

).
0.2 Examples
Now this is a point, where partial algebraic structures might come into the
picture, since often the operations occurring in computer science are not
dened on the whole set of elements involved. One has tried to overcome
this fact in several ways. Let us consider some examples in order to illustrate
this:
Example 0.2.1 A Mealy-type automaton is usually algebraically described
as a systemA = (S, I, O; , ) which consists of three non-empty sets, namely I
the set of all input elements , O the set of all output elements and S
the set of all possible (internal) states of the automaton. Furthermore
one has two functions: : S I S, the so-called transition function,
and : S I O, the so-called output function (an input i ( I) which
is applied to A while in a state s ( S) causes A to switch into the state
(s, i) ( S) and yields an output (s, i) ( O)).
This is a prototype of a so-called many-sorted algebra. Considered as a
partial algebra it would correspond to (S

I

O;

)

designating
disjoint union , where and

as well as and

in each case have the


same graph, but

as well as

are now binary partial operations on the


carrier set S

I

O.
Thus one method is to subdivide the carrier set of an algebraic struc-
ture into subsets, such that one can specify for each argument of any given
operation, from which subset it may be taken. This has led to the study of
so-called many-sorted or heterogeneous algebras. But even this may not help
directly in every case, as the following example will show.
Example 0.2.2 A data type Stack of Integers is usually described
except for error values as the so-called initial object in the class of all
two-sorted (partial) algebras A = (S, I; 0, Succ, Pred, Push, Top, Pop, E),
such that I and S are non-empty sets containing 0 and E (the empty stack!)
as distinguished elements, respectively (i.e. 0 and E are so-called nullary
constants), and the operations behave as follows
10
:
10
e
= is used throughout this book as the syntactical symbol for equality between terms.
This is done for reasons to be explained later in section 6
December 8, 2002
24 Introduction
Succ, Pred : I I, satisfying (1) Succ(Pred(i))
e
= i
and (2) Pred(Succ(i))
e
= i for every i I,
Push : S I S
Pop : S S, satisfying (3) Pop(Push(s, i))
e
= s,
Top : S I, satisfying (4) Top(Push(s, i))
e
= i (s S, i I).
The intuitive meaning of these operations is the following: If I is really the set
of integers, then Succ(i) := i +1 and Pred(i) := i 1, moreover Push(s, i)
puts integer i as a new entry on top of the stack s, while Top(s) reads the
item at the top of the stack s, if there is any such item, and is undened,
if s = E, i.e. when s is the empty stack; similarly Pop(s) throws away the
item at the top of the stack s if s is non-empty (s ,= E) and is undened if
s is empty (s = E). 0 and E are so-called nullary constants which in this
connection are required to be dened, what we shall formally express later
by the axioms 0
e
= 0 and E
e
= E.
At this point we should already say something about the use of equations
as axioms in connection with partial algebras, which will more formally be
introduced in Chapter I, section 6 (see also section 7): There has long been
a great uncertainty among algebraists on how to generalize the concept of an
identity from total algebras to partial ones; the basic concept which we shall
use here means that for all possible replacements of the variables occurring
on any side of an identity both sides of the identity have to exist (i.e. the
corresponding expressions (terms) have to be evaluable in the partial algebra
under consideration), and the values have to be equal. Because of these
hidden existence statements in this kind of semantics we also call equations
for partial algebras existence equations (more or less for didactical reasons)
and stress this by writing
e
= instead of = this corresponds to the habit
of logicians of using a dierent equality sign in the object language and in
the meta-language, respectively. We would like to stress that we shall use the
usual equality symbol in the meta-language in the same way as existence-
equality of the object language: For the given expressions both sides have
to exist and have to take the same value! With this starting point at
hand almost all known interpretations of identities for partial algebras can
be expressed by validity of usually relatively simple open rst order formulas
(in general elementary implications or conjunctions of them). Actually this
starting point is the key for a successful introduction of a powerful two-valued
model theory for partial algebras, as we shall formally do in section 7.
December 8, 2002
0 Introduction 25
Returning to Example 0.2.2 we realize that so far we have specied a
really partial two-sorted algebra. In order to make it a total algebra we have
to introduce for each sort at least one error-value in order to be able to
assign a value to Pop(E) and Top(E), respectively. If, however, we would
like to keep the above axioms valid also in the completion, we will end up
with an innite set of dierent error-values, although this is not what one
usually wants in the concrete case, showing (or knowing in advance) that the
error-values are not original, but only some necessity for the nal concrete
implementation. Here the theory of partial algebras allows a wide range of
possibilities of completion, and in connection with a specic implementation
one has to choose the most appropriate one.
Nevertheless (total) many-sorted algebras may help in so far as one
often introduces a new sort Boolean which applies to exactly two elements
True and False, and every partial (many-sorted) operation is replaced by a
function which maps its graph to True and everything else which would be
possible because of correct sorts to False. This method imposes strong
and sometimes not wanted restrictions on the admissible homomorphisms
i.e. in connection with the comparison of structures. On the other hand this
method can also be applied in connection with relations, where it just yields
the characteristic function of the relation under consideration (again with the
indicated heavy restrictions on homomorphisms). For a detailed discussion
of such completions in computer science we refer to [GoTcWa78] and [Go78]
from the ADJ-group, where such methods are also applied in connection with
so-called error algebras.
We think that it might usually be easier rst to handle partiality in its
own right and then to propose adequate completions, for instance to use
in Example 0.2.2 above a so-called heterogeneous normal one-point comple-
tion (i.e. wherever necessary, to a set of a specic sort exactly one new point
is added and all undened items of operations with output in this sort are
mapped onto this new element, the so-called error-element of the sort un-
der consideration). Other completions are conceivable, too (see Chapter IV).
Thus we present in this book an introduction into the theory of partial al-
gebras, which as we hope may serve as a basis for further investigations
using partial algebras or for further research on partial algebras.
There is one well-known example of homogeneous or heterogeneous
partial algebras, where usually nobody thinks of completing them: small
categories. Let us briey sketch the homogeneous description the two-
sorted one is presented in detail in section 1 in 1.2.1.(iii) (see also 8.1.1 for
December 8, 2002
26 Introduction
further items):
Example 0.2.3 As carrier set of a description of a given small category C
as a (homogeneous) partial algebra one chooses the set M := Mor C of all
C-morphisms (the reader may e.g. think of the set of all linear mappings
between nite powers of a given eld Q considered as vector spaces over Q
or even more concretely of the set of all nite matrices with components
in Q). Fundamental operation is a (binary) partial multiplication (the
composition of morphisms), which is dened for two elements f, g of M (i.e.
for two morphisms) iff (if and only if) the codomain (target) of f is equal to
the domain of g since in this approach objects are not included, they are
replaced by their identity morphisms. In order to have a possibility within
the partial algebraic structure to distinguish somewhere the domain and the
codomain of an (abstract) morphism one also introduces two unary total op-
erations here denoted by Dom

and Cod

( in order of distinguishing
these operations from the two-sorted ones which have objects as values),
which map a morphism f onto the identity morphism Dom

(f) respectively
Cod

(f) of the domain respectively of the codomain of f. Thus small cat-


egories can be considered as partial algebras (M; , Dom

, Cod

) with one
binary and two unary (partial) operations such that one has for arbitrary f,
g, h in M (observe that in a composition of morphisms the morphism on the
right is applied rst):
(i) Dom

and Cod

are unary total operations, is a binary partial oper-


ation, and g f is dened iff Cod

(f) = Dom

(g).
(ii) The multiplications f Dom

(f) and Cod

(f) f are always dened and


each has value f.
(iii) Dom

(g f) = Dom

(f) and Cod

(g f) = Cod

(g), whenever g f is
dened.
(iv) If g f and h g are dened, then h (g f) and (h g) f are dened,
and they are equal: h (g f) = (h g) f (weak associativity).
The class Cat

of all partial algebras (M; , Dom

, Cod

) satisfying (i) through (iv)


above is then identical with the class of all abstract small categories described
by their set of morphisms only.
We shall see that this class of all small categories is not axiomatizable by
equations (since a weak homomorphic image of a category need not be a
December 8, 2002
0 Introduction 27
category), however by elementary implications of a very simple structure
namely by such elementary implications which behave like equations when-
ever the model under consideration is a total algebra (see 8.1.1 and the con-
cept of ECE-equations i.e. of existentially conditioned existence-equations
in 8.1.2).
Category theory has developed its own special concepts, but it should be
observed that e.g. functors between (small) categories are nothing else but
the usual homomorphisms between the corresponding partial algebras.
Let us add a further example of partial algebras:
Example 0.2.4 In theoretical computer science so-called -complete posets
play an important role. These are partially ordered sets (posets) (P; ) in
which each countable ascending chain (i.e. -chain) has a supremum: i.e.
if p
0
, p
1
, p
2
, . . . , p
n
, . . . (n ) are elements of P such that p
n
p
n+1
for
each n , then the supremum sup( p
n
[ n ) of this -chain exists.
The morphisms are the mappings preserving order and the partial supremum
sup. D.Lehmann and A.Pasztor have observed in [LePa80] that -
posets can be considered as innitary partial algebras:
The partial order relation is presented by a binary partial operation,
say , such that x y is dened iff x y; and if x y, then x y := x
(say). Moreover, one has an -ary partial operation, say Sup, such that
Sup( x
n
[ n ) is dened iff x
n
x
n+1
for all n , and if Sup( x
n
[ n )
is dened, then its value is the supremum of the argument chain.
The axioms for -posets are basically of the same nature (besides con-
taining innite conjunctions) as those for small categories (namely ECE-
equations, see 13.4.4).
Thus all that is said in this book on classes denable by ECE-equations
is applicable to the classes of (small) categories respectively -posets, and no
new structure theory is needed. More material on varieties of -posets and
-continuous algebras (i.e. -posets which have - and Sup-compatible
operations, see 13.4.5) can be found in [Pa83] and [Pa84].
Example 0.2.4 already suggests that the theory of partial algebras as
presented in this book includes the theory of (partial) algebras with relations,
since each n-ary relation, say , on some set A considered as set of n-
tuples with elements from A (n > 0) can be encoded by a partial n-ary
operation, say

, on A, whose domain is , and which maps each n-tuple


December 8, 2002
28 Introduction
from onto, say, its rst component. Again we obtain a variety dened
by ECE-equations.
We do not claim that it be always better to treat relational systems as
partial algebras; we only want to stress that although we do not treat
relations separately most of the material contained in this book can also
be applied without any additional eort to partial algebras with relations.
However, we should also mention that a reader who is mainly interested in
the algebraic theory and model theory of elds another standard example
for partial algebras, since the formation of a multiplicative inverse is a partial
operation might be disappointed by this book, since our model theoretic
oriented results in Chapters II and III apply at most to Horn varieties (i.e.
classes axiomatizable by universal Horn formulas), and the class of all elds is
not denable by universal Horn formulas. But the model theoretic language
which we use for partial algebras at least allows us to axiomatize the class
of all elds by universal formulas the existential quantiers usually needed
are hidden in our concept of (existence-) equations (see sections 6, 7 and 14,
and in particular 13.4.11).
After these introductory examples which may already provide a slight
idea of some intentions of this book, let us say a few words in general about
the main concepts a larger survey of the basic ideas and concepts guiding
us here has been presented in the survey article [B82] :
We try to present a theory of partial algebras which ts neatly between
the theories of relational systems (and two-valued model theory) and of total
algebras, respectively. This has been the main guideline in particular for
the choice of the basic concept of a homomorphism: it should in each case
coincide with the corresponding concepts of the two theories. Then most
of the properties usually encountered with homomorphisms between partial
algebras have a model theoretic description or interpretation (see Chapter II).
Modest tools from category theory are also applied, when we thought that
it would help a better understanding, and in order to relate the results pre-
sented here to more category oriented papers in theoretical computer science:
we hope that our in parts parallel use of category theoretical and algebraic
language and the translations between these two languages will help to ap-
ply universal algebra, in particular partial algebras, as well as the abstract
results from category theory in concrete situations. Because of limitations
in space, time, etc., this book cannot cover all the aspects of the theory of
partial algebras, but we hope that it can serve as a good introduction.
December 8, 2002
0 Introduction 29
Since each chapter is preceded by a brief survey of the contents in its
sections, we only mention here the main intentions of each chapter:
Chapter I provides basic algebraic (and related category theoretical) con-
cepts with respect to partial algebras, such as are usually contained in an
introductory text on universal algebra, and ends with the E-equational the-
ory.
In Chapter II an introduction to a two-valued (extended) rst order log-
ical language (extended in order to include the innitary and many-sorted
cases) is presented and an investigation of its applications to a structuring
of the properties of homomorphisms is added.
In Chapter III we prove a meta Birkho-type theorem for elementary
implications and universal Horn formulas (due to H.Andr eka, I.N emeti
and I.Sain) and study to a broad extent its impact on the theory of partial
algebras.
In Chapter IV several topics are presented, most of them concerning
concepts related to (relative) freeness. It should be observed that, as long as
we move in front of a category theoretical background, many-sortedness
does not cause much trouble, but when we dive in Chapter IV into intrinsic
more set theoretic properties of (constructions of special kinds of) partial
algebras, then one encounters a lot of complications in part explicable by the
fact that for many-sorted partial algebras existence-equations are on a more
elementary level, where the dierent sorts are expressed by unary relations
and axioms for them actually elementary implications and not equations
(of this level). This may explain besides further instances given in the text
, why in our opinion basic concepts from model theory have to be used
on a very early stage in order to understand the behaviour of partial algebras
and of concepts related to them.
In an Appendix basic denitions and facts from set theory and order
theory are collected as an introduction to the main text. Moreover, some
hints on further aspects of the theory (or of other theories) of partial algebras
are given (section 22).
December 8, 2002
30 Introduction
Introduction
Partial algebras rather than total ones can be considered as basic mathemati-
cal structures implemented on computers, although one mostly species only
total ones. Many-sorted algebras are basically partial algebras, too. This
note is meant to indicate some of the basic tools for an adequate presenta-
tion of the theory of partial algebras in such a way that e.g. specications
of (many-sorted) partial algebras as abstract data types can easily be per-
formed, as will be shown by some examples. For a more detailed survey we
refer to [?]
11
.
In this note we want to discuss the model theoretic tools rather than the
universal algebraic ones. Namely, in a theory of partial algebras of some
given similarity type one can also model all relational structures of the same
type. One has also to be able to speak about the existence of interpretations
of terms. Therefore one needs a powerful language in order to describe all
features of the theory most eciently. In particular model theoretic concepts
are needed much more and much earlier than in the universal algebraic
theory of total algebras.
This need is taken care of in our approach by the fact that we base it on
existence equations (briey referred to as E-equations t
e
= t

for terms t and


t

). Their semantics form the basis of the semantics of a rst order language,
which has the same syntax as the one (with equality) for total algebras. In
particular, in this connection a special kind of elementary implications
namely what we call existentially conditioned existence equations (abbrevi-
ated as ECE-equations:
_
n
i=1
t
i
e
= t
i
t
e
= t

) , gets new meaning


and importance. ECE-equations behave like ordinary equations, when the
models are restricted to total algebras. They also allow to dene within
the accompanying rst order language all those kinds of equalities, which
are sometimes used in connection with partial algebras, too, like weak or
strong equality
12
. Among others relational structures modeled as partial
algebras then form an axiomatic class dened by special ECE-equations. In
addition, homomorphisms as basic structure preserving mappings between
partial algebras as well as half of their most important properties can be de-
ned as mappings preserving or reecting E-equations of special forms. The
11
The book [B86], in which almost all of the proofs omitted here and in [?] can be found,
is at the moment out of print, however a new version is in preparation (cf. the reference
[B9?])
12
See the end of subsection 0.4, [B86] or [?].
December 8, 2002
0 Introduction 31
missing half of properties can then be dened via the concept of factor-
ization systems taken from elementary category theory and stemming from
a Galois connection modeling the reection of elementary implications.
Because of all this it proves to be quite useful for a good description of the
arising concepts to have besides the terminology and constructions from uni-
versal algebra (homomorphisms, generalized recursion theorem, epimorphism
theorem
13
, free partial algebras
14
) also such from logic (existence equations
and elementary implications), model theory (preservation and reection of
formulas by mappings), and also in connection with model theoretic con-
cepts from (elementary) category theory (factorization systems)
15
.
As an application of the concepts and tools presented here we give a (sim-
plied) formulation of a Meta Birkho Theorem from Andr eka, N emeti
and Sain together with some of its applications.
As already mentioned at the beginning, much of the motivation for the
consideration of partial algebras and their implicational logic stems from the
possibility of their applications in computer science, some of which we list
below:
A computer can usually only represent a nite (weak) relative sub-
structure of the structure under consideration, and this is usually only
a partial algebra.
Many-sorted algebras are in general implemented as partial algebras
(as can be seen from the error message type mismatch, if an input
sequence does not t to the type restrictions).
Many-sorted (partial) algebras can be modeled quite easily by homo-
geneous partial algebras and a homomorphism into the sort algebra
(or specication algebra).
The description of many-sorted (partial) algebras as ordered pairs con-
sisting of a homogeneous partial algebra and a (closed) homomorphism
into the partial specication algebra allows without problems to gen-
eralize to overloading.
13
See [Sch70].
14
See [B70]. Moreover, most of the universal algebraic concepts needed for a presentation
without proofs are collected in Denition 0.3.1, i.e. we do not present here the universal
algebraic concepts and theorems just mentioned in more detail.
15
We do not use much from category theory. Those who want to learn a little bit more
about it, are referred e.g. to [HS73] or to the more recent [AdHS90].
December 8, 2002
32 Introduction
In order to keep the presentation as simple as possible, we restrict our
considerations mainly to the one-sorted case of nitary partial algebras to-
gether with a discussion of how to treat many-sorted (partial) algebras within
this framework as indicated above.
December 8, 2002
0 Introduction 33
0.3 Similarity types and partial algebras
Let be any set, the elements of which will be called operation symbols.
In addition, let : N
0
be a mapping from into the set of natural
numbers including zero; for , () will be interpreted as the arity of
the operation symbol .
or more precisely the pair (, ) will be called a similarity type
or briey a type (in computer science it is usually called a signature). If not
explicitly stated dierently, in what follows we shall always assume that we
are given an arbitrary but xed similarity type (, ), and that all partial
algebras under consideration are of this (same) type.
Let A be any set, and let n be any natural number including zero. An n-
ary partial operation on A is a function : dom A where dom
designates the domain of such that dom A
n
:= (a
1
, . . . , a
n
) [
a
1
, . . . , a
n
A; i.e. is a partial function out of A
n
into A, and we some-
times denote this fact by writing : A
n
A. Observe that, for n = 0,
A
0
= contains just the empty sequence, and therefore a partial nullary
operation (a partial nullary constant) on A is either empty or distinguishes
exactly one element of A and it is then usually identied with this element.
If dom = A
n
, then is called a total operation on A.
Let (, ) be any similarity type. Then a partial algebra A of type is an
ordered pair (A, (
A
())

), where A is any set, called the carrier set of A,


and, for each ,
A
() is a ()-ary partial operation on A providing the
interpretation of the partial operation symbol with respect to A.
A
() is
called the fundamental operation of A corresponding to the operation symbol
. Usually
A
() will be abbreviated
16
by
A
. A partial algebra A is called
total (a total algebra), if all fundamental operations
A
are total ( ).
For binary operations we shall often use e.g. in examples inx notation,
i.e. we write a
A
b instead of
A
(a, b).
Observe that occasionally we write in connection with some partial al-
gebra A that
A
(a
1
, . . . , a
()
)) = a, what means:
A
(a
1
, . . . , a
()
) exists
and has the value a.
In Figure 0.1 we give an example of a partial algebra. Here, we have
chosen := (2, 1, 1, 0, ), with := (
1
,
2
,
3
,
4
), the indices corresponding
16
This means that the so-called forgetful functor, which maps each partial algebra to its
carrier set, i.e. which forgets the partial algebraic structure, is indicated by forgetting the
understroke. This yields a notation for the partial algebra A as (A, (
A
)

), which seems
to be recursive or to contain a self-reference, but which is just meant to be suggestive.
December 8, 2002
34 Introduction
A :
a

A
2

A
4
= c
A
3

A
1

A
1

A
2
b
e
d
1
2
`
_

`
`
`
`
`
`
`
`
`
`
`

`
`
`
`

Figure 0.1: Example of a partial algebra


to the place in the type-sequence.
A := a, b, c, d, e,
graph
A
1
:= ((a, a), b), ((e, b), d),
graph
A
2
:= (c, a), (a, e),
graph
A
3
:= (c, e),

A
4
:= c.
Of fundamental importance is the principle of generation, which we present
together with further basic universal algebraic concepts, which will be needed
later, in the following denition:
Denition 0.3.1 Let A be any partial algebra of type , and let M be any
subset of its carrier set A.
M is called a closed subset of the partial algebra A, if for all
and for every sequence (a
1
, a
2
, . . . , a
()
) M
()
dom
A
one has

A
(a
1
, a
2
, . . . , a
()
) M.
A relative subalgebra M of the partial algebra A with carrier set M is
a partial algebra on M such that M carries as structure the (total)
restrictions to M of the fundamental operations of A, i.e. (M, (
M
)

)
with

M
:=
A
(M
()
M)
(for the graphs of the partial mappings) for each .
December 8, 2002
0 Introduction 35
,
,
,
,
,
,

a
1
a
2
a
r
a
()1
a
()
a =
A
(a
1
, . . . , a
r
, . . . , a
()
)

>
>
>

_
_

B
1
A
M M = B
0
A
M
B
2
A
M
B
k
A
M
B
k+1
A
M
A
Figure 0.2: Layer model of generation from below
A relative subalgebra on a closed subset M of the partial algebra A is
called the subalgebra of A with carrier set M.
If we assume, that M is not necessarily closed, then we denote by
c
A
M the smallest closed subset of A containing M, i.e. c
A
M is the
intersection of all closed subsets of A containing the set M, and it is
then easily seen that c
A
M is itself a closed subset of A.
The subalgebra of A on c
A
M is denoted by C
A
M, and it is called the
subalgebra of A generated by the subset M, and M is called its gener-
ating subset. See Figure 0.2 to get an idea of the process of generation.
Observe that on the direct product B :=
iI
A
i
of carrier sets of partial
algebras A
i
, for i I, the induced product structure (
B
)

is
dened componentwise in case that it exists in all components, i.e.
dom
B
:= (a
1
, a
2
, . . . , a
()
) [ (a
1
(i), a
2
(i), . . . , a
()
(i)) dom
A
i
for each i I ,
December 8, 2002
36 Introduction
and if (a
1
, a
2
, . . . , a
()
) dom
B
, then

B
(a
1
, a
2
, . . . , a
()
) := (
A
i
(a
1
(i), a
2
(i), . . . a
()
(i)) [ i I ).
B is then called the direct product of the family (A
i
)
iI
, and it is denoted
by

iI
A
i
or, in the case that I = 1, 2, by A
1
A
2
. If A
i
= A for
each i I, then one writes B =: A
I
.
A congruence relation on the partial algebra A is an equivalence relation
on A A, which is at the same time a closed subset of AA.
If A A is a binary relation on A, then Con
A
designates the
smallest congruence relation of A containing the relation .
Let be a congruence relation on the partial algebra A. Then we
dene for each a A as usual the congruence class of a as [a]

:=
b A [ (a, b) , and the quotient set A/ := [a]

[ a
A. And nat

: A A/, a [a]

(a A) then designates the


corresponding quotient mapping. Finally A/ can be provided with the
weakest partial algebraic structure on A/ which exists, since is
a congruence relation on A and which yields the partial algebra A/ :=
(A/, (
A/
)

) such that nat

becomes a homomorphism.
Observe that, for a given partial algebra A, the operators c
A
and Con
A
are closure operators on the sets A and A A, respectively.
0.4 Terms, existence equations, elementary implica-
tions
Denition 0.4.1 (of terms and of the term algebra) Assume X to be
any set, which we want to be disjoint from . Let (X )
+
be the set of
all nite sequences of length 1 of elements from X written as words
l
1
l
2
. . . l
n
with l
1
, l
2
, . . . , l
n
X , where n designates the length of such a
word. Now we dene terms of type with variables in X as follows:
(T1) x is a term for each variable x X.
(T2) If is any operation symbol, and if t
1
, t
2
, . . . , t
()
are terms, then
t
1
t
2
. . . t
()
is a term.
December 8, 2002
0 Introduction 37
(T3) Only words formed according to rules (T1) and (T2) are terms.
The set of all terms of type with variables in X is denoted by T(X, ).
The set T(X, ) of all such terms can now easily be transformed into
a total algebra T(X, ), i.e. into a total algebra (T(X, ), (
T(X,)
)

), as
follows:
For operation symbols and for terms t
1
, t
2
, . . . , t
()
T(X, ) set

T(X,)
(t
1
, t
2
, . . . , t
()
) := t
1
t
2
. . . t
()
.
T(X, ) is called the term algebra of type with variables in X.
Denition 0.4.2 Let X be a set of variables.
(i) A pair (t, t

) of terms t and t

of T(X, ) is called existence equation


(briey E-equation) and written as t
e
= t

(or even as t
e
=
X
t

, if it
has to be distinguished from t
e
=

, when t, t

T(, ), and only


assignments
17
starting from the empty set are allowed). Moreover,
we write Eeq
X
for the set of all E-equations with variables in the set X.
(ii) For any partial algebra A of type a mapping v : X A will be called
an assignment (of X in A). In connection with each such assignment v
of the set X of variables in a partial algebra A one has an interpretation
of terms, which we denote by v

, and which is dened recursively as


follows:
For each variable x X one has v

(x) := v(x).
If is any operation symbol, and if t
1
, t
2
, . . . , t
()
are terms,
for which v

(t
i
) exists, say v

(t
i
) = a
i
, (1 i ()), and
such that
A
(a
1
, a
2
, . . . , a
()
) also exists, then v

(t
1
, t
2
, . . . , t
()
)
exists, too, and takes the value
A
(a
1
, a
2
, . . . , a
()
).
The interpretation v

(t) of some term t T(X, ) exists iff it is


dened by one of the two cases above. In Figure 0.3 we give an
idea of the computation of such an interpretation for a similarity
type with one unary operation symbol.
Let, in addition, A be a partial algebra of type , v : X A a
assignment, and t
e
= t

an E-equation in Eeq
X
.
17
See below.
December 8, 2002
38 Introduction
(iii) We say that A satises the E-equation t
e
= t

w.r.t.
18
the assignment v
(in symbols: A [= t
e
= t

[v]), iff (t, t

) ker v

, i.e. iff
the interpretation v

(t) of t by v exists, and


the interpretation v

(t

) of t

by v exists, and
these interpretations are equal: v

(t) = v

(t

).
(iv) We say that the E-equation t
e
= t

holds (is valid) in the partial al-


gebra A (in symbols: A [= t
e
= t

), iff A satises t
e
= t

for every
assignment v : X A, i.e.:
A [= t
e
= t

iff A [= t
e
= t

[v] for every assignment v : X A.


(v) If t and t

are identical terms, then t


e
= t will be called a term existence
statement (briey TE-statement), since one has for v : X A:
A [= t
e
= t[v] iff the interpretation of t w.r.t. v exists.
The last part of the above denition shows that in the semantics of E-
equations the diagonal of T(X, ) T(X, ) gains importance, while for
total algebras A the statement A [= t
e
= t is always trivially true. Thus
this gives additional expressive power to E-equations in connection with par-
tial algebras, and this will become even more obvious in connection with
elementary implications.
Based on existence equations between terms with variables in some
set X as atomic formulas one may now build all the well-known formulas
of a rst order language /(X, ) with equality. And the semantics of existence
equations are extended to arbitrary rst order formulas in the usual way.
We shall restrict in this note our considerations to elementary implica-
tions, which are formulas of the form

iI
t
i
e
= t

jJ
t

j
e
= t

,
which belong to /(X, ), if the two sets I and J are nite.
For such elementary implications of arbitrary length the semantics are
extended in a natural way:
Let A be any partial algebra of type , and v : X A any assign-
ment. Then we say that A satises an elementary implication
_
iI
t
i
e
= t

i

18
w.r.t. is to be read as with respect to.
December 8, 2002
0 Introduction 39
_
jJ
t

j
e
= t

w.r.t. the assignment v (in symbols: A [= (


_
iI
t
i
e
= t

i

_
jJ
t

j
e
= t

)[v]), iff
whenever A [= t
i
e
= t

i
[v] for all i I, then A [= t

j
e
= t

[v] for all j J.


Moreover, such an elementary implication holds in A, iff it is satised in A
w.r.t. every assignment v : X A.
Particular elementary implications are
ECE-equations (i.e. existentially conditioned existence equations)

iI
t
i
e
= t
i
t
e
= t

,
QE-equations (i.e. quasi existence equations)

iI
t
i
e
= t

i
t
e
= t

,
where the conclusion consists of one existence equation only.
Observe that in connection with total algebras Aan ECE-equation
_
iI
t
i
e
=
t
i
t
e
= t

is satised or holds in A, respectively, iff t


e
= t

is satised re-
spectively holds in A, i.e. in connection with total algebras the semantics of
ECE-equations reduce to those of E-equations.
Special elementary implications occur in connection with some further
equational concepts, which are also frequently used as axioms for the de-
scription of (classes of) partial algebras.
Weak equations t
w
= t

(for t, t

T(X, )) are in our approach special


ECE-equations:
t
w
= t

:= (t
e
= t t

e
= t

t
e
= t

).
Strong equations (or Kleene equations) t
s
= t

(for t, t

T(X, )) are
conjunctions of special ECE-equations:
t
s
= t

:= ((t
e
= t t
e
= t

) (t

e
= t

t
e
= t

)).
December 8, 2002
40 Introduction
In particular P. Kosiuczenko has recently used a combination of E-equations
and weak equations in order to characterize axiomatic classes of partial al-
gebras, in which each partial algebra has a permutable respectively dis-
tributive lattice of closed congruence relations
19
.
Classes denable by E-, ECE- or QE-equations are called E-, ECE- or
QE-varieties, respectively. As an example for ECE-varieties let us list the
axioms for the class of all small categories considered as homogeneous partial
algebras
20
of type
(, ) := (Dom

, Cod

, , (Dom

, 1), (Cod

, 1), ( , 2))
satisfying the axioms
1. xDom

x
e
= x, (this implies Dom

x
e
= Dom

x, i.e. Dom

has to be
total),
2. Cod

xx
e
= x, (this implies Cod

x
e
= Cod

x, i.e. Cod

has to be
total),
3. yx
e
= yx Cod

x
e
= Dom

y Cod

yx
e
= Cod

y Dom

yx
e
=
Dom

x,
4. yx
e
= yx zy
e
= zy z yx
e
= zyx.
One might have expected in addition the QE-equation
5. Dom

y
e
= Cod

x yx
e
= yx,
but this one can be proved from the other axioms, showing that the class of
all small categories is really an ECE-variety and not only a QE-variety
21
.
19
See [Kos90]; closed congruence relations are introduced in Denition 3.6.
20
In this case the carrier set of a small category is the set of morphisms, while one
usually also considers small categories as two-sorted partial algebras, where one has in
addition a set of objects, which are represented in the homogeneous case by the identity
morphisms, i.e. by the values of the operations Dom

and Cod

see below. For


categories in general the classes of objects and morphisms need no longer be proper sets.
Cf. e.g. [AdHS90] for more details.
21
Note that 3. and 5. imply that small categories form what H.Reichel calls in [Re84]
a hep-variety.
December 8, 2002
0 Introduction 41

x
x
x
x
x
x
y
y
y
y
y
y
a
1
a
3
a
5
a
7
a
8
a
2
a
4
a
6

.
.
.
.
.
.
.
.
.
.
.
.
.
.
`
`

`
`
`
`
`
`
`
`
`
`
`
`
`
`

.
.
.
.
.
.
T(x, y, ) dom v

A
v

v
v
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
end
of
dom v

Figure 0.3: Example of a (partial) interpretation


December 8, 2002
42 Introduction
0.5 Preservation and reection of formulas homo-
morphisms and some of their properties
One application of our formulas is the classication of many important prop-
erties of mappings between partial algebras by reection and preservation of
formulas.
Denition 0.5.1 Let A, B be partial algebras of type , F /(X, ) a
formula w.r.t. a set X of variables, and let f : A B be any mapping. We
say that
(i) f preserves the formula F, iff for every assignment v : X A one has
that A [= F[v] implies that B [= F[f v];
(ii) f reects the formula F, iff for every assignment v : X A one has
that B [= F[f v] implies that A [= F[v].
This denition can also be applied to innitary elementary implications.
With this well known model theoretic concept we can now easily intoduce
our notion of homomorphism as a basic model theoretic one, which is closely
related to existence equations:
Denition 0.5.2 Let A, B be partial algebras of type , and let f : A B
be any mapping. We say that f is a homomorphism from A into B (in
symbols: f : A B), iff f preserves all the existence equations
x
1
. . . x
()
e
= y,
for all and for all x
1
, . . . , x
()
, y X (where one may assume that
these variables are pairwise distinct).
This approach means that a mapping f : A B is a homomorphism
from A into B, iff for every fundamental operation symbol and for
every sequence (a
1
, a
2
, . . . , a
()
) A
()
the existence of
A
(a
1
, a
2
, . . . , a
()
)
implies the existence of
B
(f(a
1
), f(a
2
), . . . , f(a
()
)) and that one then has
the usual compatibility condition
f(
A
(a
1
, a
2
, . . . , a
()
)) =
B
(f(a
1
), f(a
2
), . . . , f(a
()
)).
Moreover, one has the
December 8, 2002
0 Introduction 43
Proposition 0.5.3 Let A, B be partial algebras of type , and let f : A B
be any mapping. Then f is a homomorphism f : A B from A into B, iff
f preserves all existence equations in /(X, ) for all sets X of variables.
The relationship between preservation and reection of formulas is very
close, since one has
Lemma 0.5.4 Let A, B be partial algebras of type , f : A B any map-
ping and F any formula of /() or an arbitrary elementary implication.
Then
f preserves F iff f reects F.
Let us denote by Alg() the category with the class of all partial algebras
of type as class of objects and with the class of all homomorphisms between
partial algebras of type as class of morphisms.
Denition 0.5.5 In the following table some properties of homomorphisms
are listed together with the sets of formulas, the reection of which charac-
terizes them. Dierent variables are assumed to be distinct. If an operation
symbol occurs, then the reection of this kind of formulas for all
is meant. If some kind of TE-statement t
e
= t or E-equation t
e
= t

is
mentioned, this means reection of all such axioms. Observe that injective
and initial is equivalent to injective and full (where full is dened as
inducing the structure on the image set).
Property of homomorphisms Kind of reected formulas
injective x
e
= y
closed x
1
. . . x
()
e
= x
1
. . . x
()
closed t
e
= t
initial x
1
. . . x
()
e
= y
injective and closed x
e
= y, x
1
. . . x
()
e
= x
1
. . . x
()
injective and closed t
e
= t

injective and initial x


e
= y, x
1
. . . x
()
e
= z
Properties of homomorphisms dened by reection of formulas
It may attract attention that such important properties like epimorphic,
surjective, full and surjective (quotient), etc. do not occur in this table. In
December 8, 2002
44 Introduction
order to get their characterization we need the concept of a factorization
system, which is discussed in a later section.
It should be observed that classes of mappings dened by the preservation
and/or reection of formulas are always closed with respect to composition
and contain all isomorphisms (and hence all identity homomorphisms). Thus
any class of partial algebras as class of objects together with a class of map-
pings between them dened by the reection of formulas taken as class of
morphisms always forms a category. For more details see e.g. [B86] or [?].
Having now the concepts of homomorphisms and closed homomorphisms
around, it is useful to introduce the following additional concepts:
Denition 0.5.6 and Remarks Let f : A B be any homomorphism.
(i) One denes the kernel ker f of the homomorphism f as the set of all
pairs of elements of A, which are identied by f:
ker f := (a, a

) [ a, a

A and f(a) = f(a

) .
(ii) Observe that the kernel of f is a congruence relation on A.
(iii) If f is a closed homomorphism, then ker f is what we shall call a closed
congruence relation on the partial algebra A, i.e. equivalently it
is a congruence relation on A, which satises in addition:
For every operation symbol and for any sequences (a
1
, a
2
, . . . , a
()
),
(b
1
, b
2
, . . . , b
()
) A
()
satisfying (a
i
, b
i
) for 1 i () one
has
(a
1
, a
2
, . . . , a
()
) dom
A
implies (b
1
, b
2
, . . . , b
()
) dom
A
.
(iv) The set of all closed congruence relations of a partial algebra A always
forms a principal ideal (w.r.t. set theoretic inclusion) of the set of all
congruence relations of A, i.e.
There always exists a largest closed congruence relation of A.
And if a congruence relation of A is contained in a closed congru-
ence relation of A ( ), then is also a closed congruence
relation of the partial algebra A.
December 8, 2002
0 Introduction 45
(v) If the largest closed congruence relation of the partial algebra A is equal
to its largest congruence relation A A, then one has for each partial
operation
A
of A that
A
is either empty or total (and vice versa).
Having all these concepts, one can prove the following
Proposition 0.5.7 Let X be any set of variables, A any partial algebra and
v : X A any assignment, and let domv

be the relative subalgebra of the


term algebra T(X, ), which has the set of all terms interpreted in A by v as
its domain. Then one has
(i) v

is a closed homomorphism v

: domv

A, and domv

is gener-
ated by the set X of variables under consideration.
(ii) v

is the largest homomorphic extension of the assignment v to a


relative subalgebra of T(X, ) which is generated by X.
(iii) For any existence equation t
e
= t

with terms t, t

T(X, ) one has


A [= t
e
= t

[v], iff (t, t

) ker v

.
0.6 Modeling many-sorted (partial) algebras
Computer scientists need rather many-sorted than only one-sorted (partial)
algebras; and we have now the concepts available, which allow us to indi-
cate how partial and total many-sorted algebras can be treated within the
category of one-sorted partial algebras of the corresponding similarity type.
From this it will not be dicult to conclude that the basic category theoret-
ical constructions work for heterogeneous partial algebras quite analogously
as for the homogeneous ones. However, we also want to point out some dif-
ferences, e.g. concerning the model theoretic properties. More about this can
be found in [B86].
Let us rst recall that a many-sorted similarity type or signature :=
(S,

) is usually specied by
a set, say S, the elements of which are called sorts,
a set, say , of operation symbols,
December 8, 2002
46 Introduction
a mapping

from into the set S

S consisting of all pairs

() =
(s

1
. . . s

()
, s

) (for ), of which the rst component is a nite


empty or non-empty word w = s

1
. . . s

()
of elements from S, the
length () of which is just the arity of the operation symbol , while
for 1 i () the i-th letter s

i
indicates that the i-th argument of
each realization of has always to be of sort s

i
; the second conponent,
s

, of

() designates the sort of the value of any realization of ,


whenever this value exists.
A many-sorted (partial) algebra A of signature is then dened as an
ordered pair ((A
s
)
sS
, (
A
)

), where (A
s
)
sS
is a family of sets, A
s
being
called the phylum or carrier set of sort s, and if

() = (s

1
. . . s

()
, s

),
then
A
is a (possibly partial) mapping from (or out of) A
s
1
. . . A
s
()

into A
s
. A homomorphism, say h : A B, from a many-sorted (par-
tial) algebra A into a many-sorted (partial) algebra B is then usually de-
ned as being a sequence (h
s
: A
s
B
s
)
sS
of mappings h
s
between corre-
sponding phyla, such that
A
(a
1
, . . . , a
()
) = a (a
i
A
s

i
, a A
s
) implies

B
(h
s
1
(a
1
), . . . , h
s
()
(a
()
)) = h
s
(a), or in the total case just
h
s
(
A
(a
1
, . . . , a
()
)) =
B
(h
s
1
(a
1
), . . . , h
s
()
(a
()
)).
These denitions suggest that:
Dierent phyla of a many-sorted partial algebra of signature may
without loss of generality be assumed to be disjoint. (Inclusions
should be specied by appropriate unary operations.)
The specication of the signature can be considered as the description
of a particular homogeneous partial algebraic structure (
S
)

of the
homogeneous similarity type = (())

on the set S of sorts, where


for each ,
S
is dened only on the sequence (s

1
, . . . , s

()
), and
for this it takes the value s

. We shall call this partial algebra S the sort


algebra (or specication algebra) for the signature under consideration.
One can replace now the family (A
s
)
sS
of phyla of a many-sorted
partial algebra by its disjoint union, say A

:=

A
s
, in which case
each many-sorted (partial) operation
A
becomes a partial operation
on A

, which we shall designate again by


A
.
December 8, 2002
0 Introduction 47
The original partition of A into phyla can be replaced by a mapping,
say v
A
: A

S such that v
A
(a) = s iff a A
s
(a A

, s S). v
A

will be called the canonical sort mapping for A

.
One then has the following
Proposition 0.6.1 With the concepts and notations introduced above for
any many-sorted partial algebra A the canonical sort mapping v
A
is always
a homomorphism
22
v
A
: (

_
sS
A
s
, (
A
)

) S;
and this homomorphism is closed iff A is a total many-sorted algebra.
Conversely, if A

is any partial algebra and v


A
: A

S is a ho-
momorphism, then A

can be considered as a many-sorted partial algebra


((
1
(s))
sS
, (
A
)

) which is total iff v


A
is closed.
Thus, we can consider a many-sorted partial algebra of signature and
sort algebra S as a pair (A

, v
A
: A

S), where A

is a partial algebra of
similarity type = (())

23
, and v
A
is a homomorphism (which is closed
iff the many-sorted algebra is total).
A homomorphism h : A B between many-sorted partial algebras then
corresponds to a homomorphism h

: A

such that v
B
h

= v
A
,
where the graph of h

is the disjoint union of the graphs of the mappings


h
s
: A
s
B
s
(s S); and vice versa: if h

: A

is a homomorphism
between partial algebras of type provided with homomorphisms v
A
: A


S and v
B
: B

S such that v
B
h

= v
A
, then one has that (h

[
As
: A
s

B
s
)
sS
is a homomorphism between the corresponding many-sorted partial
algebras A and B. If A and B are total, then h

is closed.
In category theory, for any category C and any object S the category
C S, the objects of which are pairs (C, v : C S), where C is an
object and v is a morphism of C, and where morphisms h : (C, v) (D, w)
are morphisms h : C D of C, which satisfy w h = v, is called a comma
category.
Theorem 0.6.2 The above denitions establish an equivalence between the
category Alg() of all partial many-sorted algebras of signature = (S, ,

)
22
Remember that we assume in particular the phyla to be pairwise disjoint!
23
Here () is the length of the word in the rst component of

().
December 8, 2002
48 Introduction
and the comma category Alg() S of the category Alg() of all partial
algebras of the corresponding homogeneous type with respect to the sort
algebra S.
Example 0.6.3 Let us consider the signature for stack automata of integers
(in the way computer scientists usually represent it):
sorts : Stack, Integer
operations : 0 : Integer
Succ, Pred : Integer Integer
Empty : Stack
Push : Stack Integer Stack
Pop : Stack Stack
Top : Stack Integer.
This means that we have two nullary, four unary and one binary (possibly
partial) operations, e.g. Push : Stack Integer Stack means that
push is a binary operation with rst argument of sort Stack, second
argument of sort Integer and value of sort Stack.
Let us add that the axioms for the specication of the initial data type
24
Stack of Integers can then be formulated with s as variable of sort
Stack and x as variable of sort Integer as
25
. Observe that the rst three
of the following axioms just specify the initial data type of Integers. The
concepts of terms and of satisfaction are generalized to many-sorted
partial algebras in the obvious way. Observe, that in E-equations t
e
= t

the terms t and t

have always to be of the same (output) sort. If one


wants to use the full specicational power of the rst order language based
on E-equations, each many-sorted E-equation (or more generally each rst
order formula) should be accompanied by a set of variables, which includes
the free variables of the formula! We often omit it, when it is identical with
the set of free variables of the formula.
24
An initial data type w.r.t. some specication is an object of the specied category,
which allows exactly one homomorphism into each other object of the category under
consideration. An initial object of a category is determined up to (unique) isomorphism.
Therefore, in some approaches to the specication of a data type in computer science
see e.g. [GoTcWa78] , the initial object of an axiomatic class is if it exists
considered as the data type specied by a given signature and a given set of axioms.
25
We use here a notation for the terms, which employs brackets, since the names for the
operation symbols consist of more than one letter
December 8, 2002
0 Introduction 49
0
e
= 0,
Pred(Succ(x))
e
= x,
Succ(Pred(x))
e
= x,
Empty
e
= Empty,
Pop(Push(s, x))
e
= s,
Top(Push(s, x))
e
= x.
The initial data type specied in this way is then a partial algebra F,
where 0
F
, Succ
F
, Pred
F
, Empty
F
and Push
F
are total operations, while
Pop
F
and Top
F
are partial, however both of them are everywhere dened
except on Empty
F
, since every other element of F will be a value of Push
F
.
Example 0.6.4 In the above example the integers under consideration have
been specied as an innite total algebra. However, on a computer only a
nite part, say an interval Z
lk
:= [l, k] of it (with l, k N, the set of all
positive integers) can be realized
26
. In analogy to the above specication we
choose the similarity type
(, ) = (0, Succ, Pred, (0, 0), (Succ, 1), (Pred, 1)).
Z
lk
can then be specied as the initial object F, say, of the class K
lk
dened
by the axioms
27
:
Succ
k
(0)
e
= Succ
k
(0),
Pred
l
(0)
e
= Pred
l
(0),
Succ(x)
e
= Succ(x) Pred(Succ(x))
e
= x,
Pred(x)
e
= Pred(x) Succ(Pred(x))
e
= x.
If one wanted to specify in this connection only by E-equations, this would
also be possible, but then the number of axioms would depend on the numbers
k and l, what would be less convenient. Thus this example shows among
others the usefulness of the employment of ECE-equations as axioms.
26
We do not treat here the fact, that the implementable set of stacks will also be nite.
We leave this to the reader. We have chosen this example, since e.g. in Borlands
Turbo Pascal and many other compilers integers are implemented in such a way that
succ(32767) = 32768, and we think this to be quite annoying. One intention of this note
is it to show that the theory of partial algebras provides enough and very ne tools for
the specication of the real data types and not only of the ideal ones.
27
Succ
n
(x) stands as abbreviation for Succ(. . . Succ(
. .
n times x)...)
December 8, 2002
50 Introduction
One advantage of our comma category approach to (a theoretical treat-
ment of) many-sorted (partial) algebras may become obvious from the ob-
servation that it easily also allows to handle overloading: one just has to
drop the requirement that in the sort algebra each partial operation is to
be dened on exactly one sequence. Everything which has been observed so
far does not depend on this requirement. Because of the fact that one uses
variables of dierent sorts (or the corresponding equivalent representation
with a sort mapping) in order to formulate axioms, dierent instances of the
same operation can satisfy dierent axioms.
Example 0.6.5 Let us consider integers, sequences of integers of length n
(for some xed natural number n) and nn square matrices of integers with
addition and multiplication as operations:
sorts : Int, Seq
n
, Mat
n
operations : + : Int Int Int
Seq
n
Seq
n
Seq
n
Mat
n
Mat
n
Mat
n
: Int Int Int
Int Seq
n
Seq
n
Mat
n
Seq
n
Seq
n
Mat
n
Mat
n
Mat
n
Int Mat
n
Mat
n
Among the axioms one may formulate
28
(x, y : Int; xy
e
= yx)
(x, y, z : Int; xyz
e
= x yz)
(x, y : Int, z : Seq
n
; xyz
e
= x yz)
etc., while there is no commutativity law for matrix multiplication.
28
Observe that this time we include the specication of the variables and their sorts in
front of the ; and before the formula really starts (cf. footnote 24, where we observed
that in connection with many-sorted (partial) algebras this should be the rule).
December 8, 2002
0 Introduction 51
0.7 Encoding elementary implications by epimorphisms
Elementary implications play an important role as axioms in computer sci-
ence, since they dene classes, in which initial objects always exist. Since
we have already seen e.g. in the case of partial algebras that one has
dierent kinds of varieties, one would like to have kind of meta theorems
e.g. for the description of such kinds of varieties in a more or less
uniform way. Category theory is a useful tool in order to achieve such a
goal. And we even do not have to go too deep into category theory. With
the tools to be provided in order to be able to formulate and prove a Meta
Birkho Theorem for dierent kinds of elementary implications and even
more general kinds of formulas we can also get a possibility to characterize
those properties of homomorphisms, which we could not so far describe by
reection of formulas (namely dierent kinds of epimorphisms).
In category theory an epimorphism e : A B is a morphism, which
has the property that for any two further morphisms f, g : B C the fact
that f e = g e always implies f = g. For partial algebras a homomorphism
e : A B is an epimorphism in the category Alg(), iff e is dense, i.e. iff the
image set e(A) generates B.
There is now a close connection between epimorphisms and elementary
implications, as has been observed independently by B.Banaschewski and
H.Herrlich in [BaH76], H.Andr eka and I.N emeti in [AN83] and R.John
in [J77]. This connection will be the basis for the application of category the-
ory in the theory of partial algebras.
Denition 0.7.1 Let
:= (

iI
t
i
e
= t

jJ
t

j
e
= t

).
be an elementary implication of arbitrary length. Then we assign to a
homomorphism e

: P

as follows: Set
29
P
0
:= var t
i
, t

i
[ i I , C
0
:= P
0
t

j
, t

[ j J .
Moreover, let P
0
and C
0
be the relative subalgebras of T(X, ) with
carrier sets P
0
and C
0
, respectively. Let

P
0
:= Con
P
0
(t
i
, t

i
) [ i I , P

:= P
0
/
P
0
,

C
0
:= Con
C
0
(t
i
, t

i
) [ i I (t

j
, t

) [ j J ), C

:= C
0
/
C
0
.
29
Observe that, for an elementary implication , var designates the set of all variables
occurring in , and that T designates the set of all subterms of terms occurring in a
set T of terms.
December 8, 2002
52 Introduction
gures disabled
Figure 0.4: e

is the epimorphism encoding the elementary implication .


Figure 0.5: ftools3.2=FKext; e

is an A-extendable epimorphism,
A is e

-injective.
Then e

: P

is the encoding homomorphism induced by the inclusion


mapping from P
0
into C
0
(see Figure 0.4)
30
Lemma 0.7.2 and Denition
(i) For a given elementary implication the encoding homomorphism e

:
P

is an epimorphism. Moreover, holds in a partial algebra A,


iff for every homomorphism f : P

A there exists a (unique) ho-


momorphism g : C

A such that g e

= f. One says in such


a situation that e

is an A-extendable epimorphism, and that A is


injective w.r.t. e

31
.
(ii) Every epimorphism e : P C encodes a possibly innitary
implication; namely, if X is a generating subset of P and : X
P the inclusion mapping, R a generating subset of ker

, and S a
generating subset of ker (e )

, then

e
:= (

(t,t

)R
t
e
= t

(s,s

)S
s
e
= s

)
is an elementary implication encoded by e. If the sets R and S can be
chosen to be nite, then
e
/(var , ).
30
In what follows, Epi, Mono, Iso and Hom designate the classes of all epi-
morphisms, monomorphisms, isomorphisms and homomorphisms, respectively, between
partial algebras of type .
31
Cf. Figure 0.5 and Denition 0.7.3.
December 8, 2002
0 Introduction 53
gures disabled
Figure 0.6: ftools3.3=Frebf; Reection by m of the el. implication
encoded by the epi. e

Denition 0.7.3 If c is a class of epimorphisms in the category Alg(), then


we dene
Inj c := A Alg() [ each e c is A-extendable ,
to be the class of all models of the set of all elementary implications induced
by epimorphisms from c (cf. Figure 0.5 and Lemma 0.7.2.(i)).
Observe that existence equations t
e
= t

are special kinds of elemen-


tary implications, where the premise is empty; however for the encoding
epimorphism
e
: P

one has P

:= var , which is a discrete partial


algebra. P

only allows a homomorphism into the empty partial algebra, if


it is empty itself; else trivially holds in , and Inj e

is also true.
0.8 Factorization systems
The relation Dip between arbitrary homomorphisms as dened below
is motivated by the category theoretical interpretation of reection of an
elementary implication encoded by the epimorphism e

: P


by a homomorphism m. It is easily realized that Figure 0.6 represents this
situation. Observe that d is unique and also satises m d = q, since e

is an
epimorphism.
Denition 0.8.1 A pair (e, m) of homomorphisms e : P C, m : A B
has the unique-diagonal-ll-in-property, i.e. it satises Dip(e, m), iff for all
homomorphisms p : P A and q : C B satisfying q e = m p there
exists a unique homomorphism d : C A such that d e = p and m d = q.
Observe that here e need not to be an epimorphism.
As usual such a relation induces two operators which are called here
and
op
of a Galois correspondance applicable in this connection to
classes of homomorphisms:
December 8, 2002
54 Introduction
Denition 0.8.2 Let c, / Hom be two classes of homomorphisms. Then
one denes
32
(c) := m Hom [ Dip(e, m) for all e c ,

op
(/) := e Hom [ Dip(e, m) for all m /.
Of special interest for us are the situations where the related classes (c)
and
op
(c) satisfy some additional properties:
Denition 0.8.3 Let c, / Hom be arbitrary classes of homomorphisms
of partial algebras of type . The pair (c, /) is said to form a factorization
system in Alg(), if the following conditions are satised:
(FS1) / c = Hom,
(FS2) / / = /, c c = c,
(FS3) Iso / c,
(FS4) c / Dip.
The motivation for the name factorization system stems from the fol-
lowing equivalent description:
Lemma 0.8.4 c, / Hom form a factorization system, iff (FS1), (FS2)
and (FS3) are satised together with
(FS4) The factorization in (FS1) is unique up to isomorphism, i.e. if f =
m e = m

with e, e

c, m, m

/, then there is a unique


isomorphism i from Cod(e) onto Cod(e

),
33
such that e

= i e and
m = m

i.
Factorization systems are abundant in Alg(), since we have the
Theorem 0.8.5 Let c Epi and / Mono in Alg(), then
(
op
(c), (c)) as well as (
op
(/),
op
(/))
are factorization systems in Alg().
32
For more details about factorization systems see G.E.Strecker [S72], [AdHS90],
Chapter IV, or [B86], section 10.
33
Cod(e) designates here the target algebra of the homomorphism e.
December 8, 2002
0 Introduction 55
Thus in particular classes of homomorphisms, which are dened by the re-
ection of some set of E-equations or elementary implications, always are the
right hand partners of factorization systems. Special examples are described
in
Proposition 0.8.6 In the category Alg() we have among others the non-
trivial factorization systems (c, /) shown in the table below, where e : P
C in c and m : A B in /.
c is the class of all / is the class of all
homomorphisms e which are homomorphisms m which are
TAlg()-extendable epimorphisms closed
epimorphisms closed and injective
full and surjective (= quotients) injective
surjective initial and injective
(=full and injective)
surjective, and initial
c C

C
(C
()
) implies #e
1
(c) = 1
Some interesting factorization systems
Proposition 0.8.6 shows us that the most interesting properties of homo-
morphisms, which have shown up so far, are either denable by the reection
of existence equations or are their partners in a factorization system (repre-
senting all the reected epimorphisms). The only exception here from this
rule are full homomorphisms.
In particular, the fact that the classes Ext of all TAlg()-extendable epi-
morphisms and /
c
of all closed homomorphisms form a factorization sys-
tem shows that our extensions of assignments to closed homomorphisms on
X-generated relative subalgebras of the term algebra T(X, ) are denable
within the category Alg() without using partial mappings between partial
algebras: for a homomorphism f : A B the pair (id
A
domf

, f

) is just
its (Ext, /
c
)-factorization (up to isomorphism).
December 8, 2002
56 Introduction
Figure 0.7: ftools4.1; P is H-projective
0.9 A Meta Birkho Theorem
Now we have almost all the tools available, which are needed for the for-
mulation (and the proof) of a (still quite restricted version of a) result by
H.Andr eka, I.N emeti and I.Sain (see [AN82] and [NSa82]) which yields
many Birkho-type theorems for partial algebras w.r.t. very dierent kinds
of implications.
In what follows we shall use the letters Op and o both for classes of
homomorphisms and for special operators induced by them; H for H-
homomorphic images and o for o-subobjects, i.e. for K Alg() we
dene
HK := B Alg() [ there are A K and f : A B in H,
o K := A Alg() [ there are B K and f : A B in o .
Recall that one has the following concept dual to injectivity:
Denition 0.9.1 Let H be a class of homomorphisms. A partial algebra P
is called H-projective, iff for every homomorphism h : A B from H and
for every homomorphism f : P B there is a homomorphism g : P A
such that h g = f (see Figure 0.7).
As the last preparation for the following theorem we have to specify dif-
ferent kinds of epimorphisms (with respect to their application).
Denition 0.9.2 Let e : P C be an epimorphism, then we say that
(i) e is an H-epimorphism (for H-images), iff P is H-projective;
(ii) e is an o-epimorphism (for o-subobjects), iff e
op
(o);
December 8, 2002
0 Introduction 57
(iii) e is always a T-epimorphism (for products);
(iv) e is a T
r
-epimorphism (for reduced products
34
), iff P is totally nite
35
.
From the results of H.Andr eka, I.N emeti and I.Sain in [AN82] and [NSa82]
one can extract the following
Theorem 0.9.3 Meta Birkho Theorem Let (O, /) be a category of
partial algebras with O Alg() and the class / Hom of homomor-
phisms
36
such that
(O, /) has products and direct limits,
every A O is the direct limit (in (O, /)) of totally nite partial
algebras belonging to O.
Moreover, let H, o / be classes of morphisms such that:
(1) Each A O is the H-image of an H-projective P O.
(2) Every H-projective object P O is the direct limit of totally nite
H-projective partial algebras from O.
(3) (
op
(o), o) is a factorization system in (O, /).
(4) If g f H and f
op
(o), then g H.
34
In what follows we assume that the reader is familiar with the concepts of reduced
products (w.r.t. some lter), of ultraproducts (i.e. reduced products w.r.t. some ultralter),
and of direct limits, which are rather directed colimits cf. also [B86] or [?].
35
A partial algebra A is called totally nite, iff the set A as well as the disjoint union
of all graphs of fundamental operations of A are nite. This is a special case of strong
smallness (see subsection 11.2) for the category of all partial algebras of some nitary
type with all homomorphisms as morphisms. In the category of all total algebras of
nitary type with all homomorphisms as morphisms a total algebra A is called strongly
small (or Neutotally nite), i it is the quotient of a nitely generated term algebra with
respect to a nitely generated congruence relation. In an arbitrary subcategory Alg()()
the description may be quite dierent and depends on the factorizability through an
arbitrary directed system for homomorphisms into its direct limit object, where these
homomorphisms start from such a (strongly small) partial algebra.
36
A reader, who is familiar with the concept of strong smallness should understand
totally nite in what follows as strongly small; else one should take as O either the
class of all partial algebras or the class of all total algebras of some nitary type , and
/ to be the class of all homomorphisms between objects of O.
December 8, 2002
58 Introduction
(5) From each A O there starts up to isomorphism only a set of
op
(o)-
morphisms.
(6) o = e : P C [ e
op
(o) and P and C are totally nite .
Let K O be any subclass of partial algebras and let T be one of the operators
T or T
r
. Then
Inj c
HSF
K = HoT K,
where c
HSF
designates the class of all HoT-epimorphisms, which hold in K.
For applications of this theorem we only consider as (O, /) the category
Alg() of all partial algebras of type and of all homomorphisms between
them. However, some ECE-varieties might also do. Moreover, let in this
category
1 be the class of all isomorphisms,
o
w
be the class of all injective homomorphisms,
H
w
be the class of all surjective homomorphisms,
/
c
be the class of all closed homomorphisms,
/
i
be the class of all initial homomorphisms,
H
f
be the class of all full and surjective homomorphisms,
H
i
:= /
i
H
w
be the class of all initial and surjective homomorphisms,
H
b
:= o
w
H
w
be the class of all bijective homomorphisms,
H
c
:= /
c
H
w
be the class of all closed and surjective homomorphisms,
o
i
:= /
i
o
w
be the class of all initial and injective homomorphisms,
o
c
:= /
c
o
w
be the class of all closed and injective homomorphisms.
Then we get the
Proposition 0.9.4 In Table 0.1 it is indicated by + in a row for a class
of homomorphisms chosen for H and in the column for a class of homo-
morphisms chosen for o, when it is known that this pair (H, o) satises the
assumptions of Theorem 0.9.3. A missing entry means that the corresponding
pair has not yet been investigated.
This yields already 52 dierent Birkho type theorems, since we have 2
possibilities for the operator T in Theorem 0.9.3. However, not all resulting
closure operators are distinct, since e.g. for classes closed w.r.t. H
w
closure
w.r.t. the operators T and T
r
is equivalent.
December 8, 2002
0 Introduction 59
H o o
i
o
w
/
i
o
c
/
c
1 + + + + +
H
c
+ + + + +
H
f
+ + + + +
H
w
+ + + + +
H
i
+ + +
H
b
+ + +
Table 0.1: Compatible pairs (H, o) (see Proposition 0.9.4)
Some descriptions of premises and conclusions derived from the condi-
tions (1) through (6) from Theorem 0.9.3 on the implications under consid-
eration are collected for special operators in Table 0.2 (equality = here
really means that the terms have to be identical and X is the set of variables
under consideration).
It should be observed that one consequence of Theorem 0.9.3 is the fact
that the quasi-primitive classes K, i.e. classes which satisfy K = 1o
c
T K, are
exactly the classes denable by elementary implications with no restrictions
on the lengths of premise or conclusion. Moreover, observe that
E-equations correspond to the operator H
w
o
c
T(= H
w
o
c
T
r
),
nite ECE-equations correspond to the operator H
c
o
c
T
r
, and
nite QE-equations correspond to the operator 1o
c
T
r
.
For the description of closed sets of E-equations, ECE-equations and QE-
equations, respectively, see [?], Theorem 3.4, or, if proofs are of interest,
[B86], section 14
37
, where even more general results e.g. a Meta-Theorem
for closed sets of implications
38
by R.John
39
can be found.
37
See e.g. Theorem 14.3.3.
38
See [B86], Theorem 14.2.2
39
See [J77].
December 8, 2002
60 Introduction
Restrictions on the premise
_
iI
t
i
e
= t

i
in case of H as:
1 no restrictions
H
w
t
i
= t

i
is a variable (i I)
H
f
t
i
= t

i
=
i
x
1i
. . . x
(
i
)i
(i I,
i
),
and for (k, i) ,= (k

, i

) the variables x
ki
and x
k

i
are distinct
H
c
t
i
= t

i
, arbitrary term (i I)
Restrictions on the conclusion t
e
= t

with respect to the premise


_
iI
t
i
e
= t

i
in case of o as:
o
w
t, t

t
i
, t

i
[ i I X
o
i
t arbitrary term, t

t
i
, t

i
[ i I X
o
c
t, t

arbitrary terms
/
i
t arbitrary term, t

t
i
, t

i
[ i I X,
and not both of t, t

are variables
/
c
t = t

arbitrary term
Table 0.2: Premises and conclusions for some special operators
December 8, 2002
Chapter I
Basic denitions and properties
In this rst chapter we present a great part of the universal algebraic prop-
erties of partial algebras. Starting with basic denitions and concepts and
their main properties we develop the theory of partial algebras up to their
equational theory. As already indicated in the introduction, heterogeneous
(also called many-sorted) partial algebras are included in the considerations,
since they are the basic structures considered e.g. in the initial algebra ap-
proach to abstract data types of the so-called ADJ-group
1
. However, our
treatment of heterogeneous partial algebras is based on the investigations
of homogeneous (i.e. one-sorted) partial algebras, since we think that this
point of view reveals a lot of special features of heterogeneous algebras. In
a category theoretical language one would say to our approach that the cat-
egory of all heterogeneous partial algebras of some given similarity type is
considered as a comma category Alg() S of the corresponding category
of homogeneous partial algebras, where is the corresponding homogeneous
type, and S is a homogeneous partial algebra encoding the heterogeneous
similarity type, the so-called sort algebra
2
.
In section 1 homogeneous and heterogeneous partial algebras are dened
in the usual way, and some concepts connected with cardinals and ordinals
related to the similarity type are discussed.
In section 2 we introduce homomorphisms as structure preserving map-
pings from the source- into the target-object, study their relationship with
congruence relations, and prove the Homomorphism Theorem and its conse-
quences (e.g. Isomorphism Theorems). Besides injectivity, surjectivity and
1
See e.g. [GoTcWa78].
2
Cf. subsection ??.
61
62 Chapter I Basic denitions and properties
bijectivity, also closedness and fullness are recognized as important proper-
ties of homomorphisms between partial algebras. As a by-product we get the
description of heterogeneous partial and total algebras indicated above.
In section 3 closed subsets and the process of generation of partial algebras
by sets are investigated also in connection with the properties of homo-
morphisms mentioned above. One very important concept in this section
is the one of algebraic induction in Computer Science called structural
induction.
Section 4 is dedicated to the construction of new partial algebras from
given ones, e.g. direct products, coproducts, reduced products and direct
limits are described.
Section 5 treats the fundamental concept of free (partial) algebras in its
generalized form of universal solutions. E.g. absolutely free completions,
terms and term operations are introduced and investigated, and a general
Recursion Theorem as well as a homomorphism theorem for epimorphisms
and an existence theorem for universal solutions are presented. Moreover, up-
per bounds for the cardinalities of partial algebras depending on the cardinal
dimension and the cardinality of a generating subset are given. This section
lays the basis for all concrete con structions in connection with equational
and implicational theories.
Finally, section 6 contains the introduction of the concept of E-equations
and the Birkho-type E-equational theory for partial algebras. E-equations
will then form the basis for a (in the nitary case rst order) language al-
lowing the introduction of more general model theoretic concepts and inves-
tigations in the following two chapters.
1 The concept of homogeneous and hetero-
geneous partial algebras
In this section the concepts of homogeneous and heterogeneous partial al-
gebras of a given similarity type, as already intuitively introduced in the
introduction, are made precise. For set theoretical prerequisites see section
?? in the appendix.
December 8, 2002
1 Homogeneous and heterogeneous partial algebras 63
1.1 Partial algebras
Since they form the fundament of (partial) operations and algebraic struc-
tures, we rst consider
Denition 1.1.1 (of sequences) Let A and K be any sets
3
. A K-ary
sequence (sequence of type or arity K) (with elements) in A is any mapping
from the set K into the set A. We shall use dierent notations for sequences
as suitable , e.g. a : K A, (a
k
)
kK
, (a
k
[ k K), (ak [ k K)
or (a(k) [ k K); if K is some natural number, say n, then we also use
the notation (a
0
, . . . , a
n1
) or sometimes (a
1
, . . . , a
n
), since we shall usually
identify the set A
n
of all mappings from 0, . . . , n 1 into A with the set
A A . . . A (n factors) of all n-tuples of elements of A.
In general we denote by A
K
the set of all K-ary sequences in A.
#A always designates the cardinality of set A
4
Remark 1.1.2 Notice that A
1
or A
{k}
for any one-element set is usually
identied with the set A itself, if no confusion can arise, and that A
0
= A
Lleer
consists of exactly one element, namely the empty sequence : A
0
= .
Especially we have
5
(i) A
K
= iff A = and K ,= .
(ii) #A
K
1 iff #A 1 or K = .
(iii) #A
K
= 1 iff #A = 1 or K = .
Denition 1.1.3 (of partial operations) and Remarks As before let
A and K be any sets. A partial operation of arity K on the set A (a K-
ary partial operation on A) is any partial mapping out of A
K
into A, in
symbols: : dom A, where dom A
K
is its domain
6
.
3
Often K is assumed to be an ordinal number, but in general this assumption will not
be necessary
4
See subsection 23.4 in the Appendix.
5
iff means if and only if.
6
We allow arbitrary sets as arities, since once one gets acquainted to this fact, and
except for the case of superposition in section 15 it is not necessary to use always
ordinal numbers, if one has to include innitary operations. However, those readers, who
like and know ordinal numbers, can always think of the arity K or () below to be
an ordinal. Then it is quite convenient to think of it as the set of all smaller ordinal
numbers. In this case one has K + 1 = K K. In the nitary case, it means that
K = 0, 1, . . . , K 1. Thus it should not cause trouble, when we say that K be a set.
More about ordinal numbers can be found in the appendix 23.3. If one does not want
at all to think of innitary operations, then one should think of an arity K to be just a
natural number (i.e. a nite ordinal).
December 8, 2002
64 Chapter I Basic denitions and properties
If, however, dom = A
K
, then is called a (total, full or complete) K-
ary operation on A, while in the case of dom = is called the empty or
discrete K-ary partial operation on A.
Observe that for any given arity K we usually have exactly one discrete
partial operation of type K on any set A, while we usually have quite a lot
of total K-ary operations on A.
If K = is an empty arity, then any K-ary partial operation is called a
nullary partial operation on A, and if its domain is non-empty, then it is often
called a (nullary) constant in A. Notice that the total nullary operations
on a set A are in one-to-one correspondence with the elements of A, and they
are usually identied with their unique value.
We would also like to mention that every K-ary partial operation on a
set A can be considered in a natural way as a (K K)-ary relation on A,
i.e. as a subset of A
K{K}
, namely by considering
graph := (a, (a) [ a dom,
the graph of , having a distinguished index, namely K
7
. On the other
hand, every K-ary relation R on a set A for a non-empty set K, i.e. every
subset R A
K
, can be considered as a K-ary partial operation on A by
distinguishing some element, say k
0
in K, and dening
R
: R A by

R
(a) := a(k
0
) iff a R,
i.e.
R
is a partial K-ary projection of A onto the k
0
-th component.
If the reader follows all the subsequent denitions and statements care-
fully, he may easily realize that everything that is said about partial opera-
tions and partial algebras equally applies to relations and relational systems,
only that something might either become trivial (as for instance everything
concerning generation
8
), while other items will become a little bit more
complicated (as for instance some descriptions by rst order axioms will
need some additional axioms).
While mappings f from a set A into a set B are usually noted as f : A
B, the fact that a partial mapping is a K-ary partial operation on a set A
and not necessarily everywhere dened on A
K
will sometimes be indicated by
the notation : A
K
A, i.e. will always indicate a partial mapping.
7
Notice, that in the general context we shall usually apply functions from the left, for
binary concrete operations inx notation is often used, too.
8
Cf. section 3
December 8, 2002
1 Homogeneous and heterogeneous partial algebras 65
Denition 1.1.4 (of partial algebras). Throughout this book will des-
ignate an arbitrary, but xed set, which we shall call the set of operation
symbols; the elements of , i.e. the operation symbols themselves will usu-
ally be denoted by
9
, , ,. . . or
1
,
2
, . . .
10
.
Moreover, will designate a sequence of empty or non-empty sets (often
of natural or even arbitrary ordinal numbers considered as the set of smaller
numbers) indexed by the elements of , i.e. := (())

(sometimes more
briey := (K

). This sequence or mapping will be called the


similarity type (or briey: type) under consideration
11
. When all sets ()
( ) are nite, then the type is said to be a nitary type, otherwise we
speak of an innitary type.
Let A be any empty or non-empty set. A (homogeneous) par-
tial algebra A of (similarity) type with carrier set A is an ordered pair
(A, (
A
)

)(=: A)
12
), such that, for each ,
A
is a ()-ary partial
operation on A:
A
: A
()
A.
A
is called the fundamental partial
operation of A corresponding to the operation symbol ( ).
If each fundamental operation of a partial algebra A is total, then A
is called a total (or complete (universal)) algebra since we shall always
deal with universal or partial universal algebras, we shall usually omit the
adjective universal, since no confusion with algebras, i.e. hypercomplex
systems in the classical sense, is to be feared. If, on the contrary, each
fundamental partial operation is discrete (empty), then A will be called the
discrete partial algebra of type with carrier set A, and we shall often identify
9
Namely will usually be used as denotation for some initial ordinal number.
10
See 1.1.5.(ii) below.
11
The similarity type should be dened more precisely as the ordered pair (, ), but
since uniquely determines as its index set or domain, we usually suppress a reference
to it.
12
This denition of a partial algebra A of type seems to contain a cycle, but it just
wants to be suggestive. If one wants to be more precise, one should dene a partial
algebra A as an ordered pair (A, (
A
())

), where A is carrier set of A, and


A
:
PO(C),
A
() PO
()
(C) for , provides the interpretation of the ()-ary partial
operation symbol as a ()-ary partial operation on the carrier set C with respect to
A.
A
() is called the fundamental operation of A corresponding to the operation symbol
. In general but not always we shall use the letter A as the name for the carrier
set of the partial algebra A (i.e. the so-called forgetful functor, which maps each partial
algebra to its carrier set, i.e. which forgets the partial algebraic structure, is indicated by
forgetting the understroke). Moreover,
A
() will be abbreviated by
A
. This yields
a notation for the partial algebra A as (A, (
A
)

), which seems to be recursive or to


contain a self-reference, but which is just meant to be suggestive.
December 8, 2002
66 Chapter I Basic denitions and properties
it with the carrier set A itself sometimes, in order to stress discreteness,
we shall also write A
d
for the discrete partial algebra of type on A.
Given a similarity type , Alg() will designate the class (species) of all
partial algebras of similarity type , while TAlg() will designate its subclass
of all total algebras of type . Partial algebras of the same similarity type
will be called similar; if they are of nitary type, they are called nitary,
otherwise innitary.
Remarks 1.1.5 (i): If not explicitly stated dierently, partial algebras con-
sidered at the same time e.g. in connection with homomorphisms (cf.
section 2) or classes of partial algebras are always assumed to be similar.
(ii): What we have introduced above is the notation which we shall use
in denitions and general statements. However, when we consider special
examples, where only nitely many operations are given, then if, say =

1
,
2
, . . . ,
n
, we shall often consider rather as ordered n-tuple :=
(
1
,
2
, . . . ,
n
) than as set, the similarity type as = (K
1
, K
2
, . . . , K
n
),
and any partial algebra A of type as (A;
A
1
,
A
2
, . . . ,
A
n
) in agreement
with the common notation. In very special cases, such as groups, rings, and
derived structures, we shall also omit as usual the upper index A. E.g. we
denote (multiplicative) groups as total algebras (G; ,
1
, e) of type (2, 0, 1),
or elds as partial algebras (F; +, 0, , , 1,
1
) of type (2, 0, 1, 2, 0, 1), where
all operations apart from
1
are total, and where only 0 does not belong
to dom
1
.
It should be observed, too, that the above specication of a partial algebra
with nitely many operations depends on the order of the listing of the
fundamental operations
13
, e.g. (A; +, ) and (A; , +) have to be considered
as dierent (where + and are dierent operations, but in both cases
the same symbol designates the same operation).
When we deal with binary operations, we shall often use the inx no-
tation, writing the operator between the two arguments, especially when
dealing with classical algebraic structures like semigroups, groups, rings,
elds, categories, etc.. The dot for multiplication will often be omitted.
Presentation of Examples 1.1.6 Usually we shall describe a concrete par-
tial algebra by the specication of its carrier set and by the specication of
the graphs of its fundamental partial operations. But sometimes we shall
13
However, there are concepts ignoring such facts, if the arities allow it, like the one of
weak homomorphisms, but we do not treat them in this book.
December 8, 2002
1 Homogeneous and heterogeneous partial algebras 67
rather sketch nitary partial algebraic structures graphically, indicating the
elements of the carrier set by small circles. And the application of an oper-
ation to a sequence is indicated by either an arrow, when the operation is
unary (i.e. when the arity is a one-element set!), or by lines usually la-
beled with the position in the sequence, when this matters starting at the
arguments and leading to a symbol (small rectangle or circle dierent from
the one indicating the elements) representing the partial operation under
consideration. From this symbol there starts an arrow leading to the value
of the operation with respect to the given sequence. Dierent unary partial
operations will usually be distinguished by a dierent form of the arrows:
continuous line, broken line or dotted line usually we shall not need more
than three dierent unary partial operations in order to illustrate some fea-
ture; non-empty fundamental constants will be distinguished by writing the
corresponding operation at the circle representing the value. Thus such ex-
amples may look like the one in Figure 1.1, where, say := (2, 1, 1, 0, ),
with := (
1
,
2
,
3
,
4
), the indices corresponding to the place in the type-
sequence. A partial algebra A of this type may then be described as follows
(observe that we identify one-element sequences with the elements of the
carrier set!):
A := a, b, c, d, e,
graph
A
1
:= ((a, a), b), ((b, e), d),
graph
A
2
:= (c, a), (a, e),
graph
A
3
:= (c, e),

A
4
:= c.
1.2 Heterogeneous partial algebras
Remarks and Examples 1.2.1 . (i): Mainly in Computer Science, but
to some extent also in Mathematics there arise algebraic structures which
can be considered as partial algebras, where the domains of the partial fun-
damental operations can more easily be dened by subdividing the carrier
set, say A, into a family, say (A
s
)
sS
, of usually but not necessarily dis-
joint subsets
14
such that all arguments in a xed position of a fundamental
14
Even when the sets are not disjoint, they are primarily treated, as if they were disjoint.
Inclusions should be specied by introducing the corresponding inclusion mapping as fun-
December 8, 2002
68 Chapter I Basic denitions and properties
A :
a

A
2

A
4
= c
A
3

A
1

A
1

A
2
b
e
d
1
2
`
_

`
`
`
`
`
`
`
`
`
`
`

`
`
`
`

Figure 1.1: Example of the specication of a partial algebra by a drawing


operation belong to exactly one of these subsets, and all values of a specic
operation have to belong to a well-dened subset. Then either this subdivi-
sion and specication allows a full description of the domains of the partial
operation, or hopefully only a few exceptional cases are left (as the
application of POP or TOP to the empty stack L in example ?? in the intro-
duction (STACK OF INTEGERS)). For such exceptional cases a special
error handling will then become necessary (and will mostly be easier as in
a homogeneous treatment), if e.g. the algebraic specication of an abstract
data type is concerned.
(ii): Probably the most well-known example from Mathematics al-
though usually described a little bit dierently are vector spaces V , where
the carrier set is divided into two disjoint sets, namely the sets V
s
of scalars
and V
v
of vectors, respectively, and V
s
carries the structure of a eld, while
V
v
carries the structure of an abelian group, and there is a multiplication,
say : V
s
V
v
V
v
satisfying the well-known axioms for the multiplication
of vectors with scalars.
(iii): Another example are small categories C, where C is also subdivided
into two disjoint subsets, usually denoted by Ob (C) and Mor(C), the subsets
of objects and morphisms, respectively cf. the dierent, but equivalent def-
damental operations or what one calls hidden functions with the appropriate axioms
specifying their injectivity. Non-empty intersections should be introduced as additional
sets with the corresponding inclusion operations, etc..
December 8, 2002
1 Homogeneous and heterogeneous partial algebras 69
inition in the introduction, where we only used one sort of elements, namely
the one of morphisms. Small categories can then be dened as partial alge-
bras of type (2, 1, 1, 1), corresponding to the set , Dom, Cod, 1 in
the same order as in the type , such that
: Mor (C) Mor (C) Mor (C) is a partial operation, and
Dom, Cod : Mor (C) Ob(C) as well as
1 : Ob (C) Mor (C) are total operations.
designates the composition of morphisms, while Dom and Cod
assign to each morphism its domain (start object) and its codomain (target
object), and 1 assigns to each object its canonical isomorphism (if the cate-
gory is realized by sets with mappings, this is the identity automorphism, and
then id is used instead of 1), such that the usual axioms for categories
are satised
15
for all m, m

, m

Mor (C) and for all A Ob (C):


(C1) m

m is dened iff Cod m = Dom m

.
(C2) If m

m and m

are dened, then (m

) m and m

(m

m)
are dened and equal. (Associativity of composition)
(C3) If m

m is dened, then Dom(m

m) = Dom(m) and Cod(m

m) =
Cod(m

).
(C4) 1
A
the value of 1 on A satises:
(a) Dom(1
A
) = Cod(1
A
) = A;
(b) if 1
A
m is dened, then 1
A
m = m;
(c) if m 1
A
is dened, then m 1
A
= m.
Notice that abstract categories in general i.e. not necessarily small ones
are dened similarly, with the only dierence that Ob (C) and Mor (C) are
allowed to be proper classes, and the additional restriction, that for any two
objects A and B the class
Mor (A, B) := m Mor (C) [ Dom(m) = A and Cod(m) = B
is a proper set.
15
Cf. HERRLICH & STRECKER [HS73] or ADAMEK, HERRLICH & STRECKER
[AdHS90].
December 8, 2002
70 Chapter I Basic denitions and properties
(iv): The two examples mentioned above and those of the introduction
suggest that it might be helpful in some cases to have an appropriate lan-
guage to handle such situations, although they can also be treated within the
framework of partial algebras in general. Although we shall present in sec-
tion 2 a dierent approach to heterogeneous or many-sorted partial algebras,
which relates them more closely to the homogeneous case, we briey present
in this subsection the straightforward generalization to the partial case of
their usual denition (cf. P.I.HIGGINS [?], H.LUGOWSKI [Lu76], the ADJ-
group, e.g. [GoTcWa78], H.A.KLAEREN [Kla83] , H.EHRIG & B.MAHR
[ErMh85] or H.REICHEL [Re84] or [Re87]). The reader who has diculties
in understanding the following denition of heterogeneous partial algebras
should skip this subsection and take the characterization in subsection ?? as
a denition.
Denition 1.2.2 (of heterogeneous partial algebras). Throughout this
book S will designate an arbitrary but xed non-empty set, which will be
called the set of sorts, and the elements of S will be denoted by s, s

, s, . . .
or s
1
, s
2
, . . ., and will be called sorts.
As in 1.1.4, designates the set of fundamental operation symbols, and
= (())

is the family of arities under consideration. A heterogeneous


similarity type or signature is dened as a mapping
:
_
S
()
S [
assigning to each fundamental operation symbol an ordered pair
() :=

:= ((s

[ ()), s

) S
()
S.
A heterogeneous (or many-sorted) partial algebra A of signature is an
ordered pair
((A
s
)
sS
, (
A
)

)
consisting of a family A := (A
s
)
sS
of the phyla or carrier sets of the dif-
ferent sorts and a family (
A
)

of the (partial) many-sorted fundamental


operations

A
:
()
A
s

A
s

(for each ); i.e. we have in particular


() dom
A

()
A
s

.
December 8, 2002
1 Homogeneous and heterogeneous partial algebras 71
We shall call the components of the sequence (s

)
()
the input sequence
of the fundamental operation symbol , and its components are called the
input sorts, while s

is called the output sort of .


When we have equality in () for all , then A is called a (total)
heterogeneous or many-sorted algebra
16
. All other concepts are dened as in
the homogeneous case.
By Alg() we denote the class and later the corresponding category
of all heterogeneous partial algebras of signature . Moreover, by TAlg()
we denote the subclass and later the full subcategory of all total het-
erogeneous algebras of signature .
Remarks 1.2.3 (i): One could replace by a collection of two mappings,
say
Input :
_

S
()
and Output : S,
such that for each one has
Input() S
()
and () = (Input(), Output()).
Notice that Input() is itself a ()-ary sequence. A heterogeneous (or
many-sorted) similarity type or signature is then a pair (, ) or a triplet
(, Input, Output) notice that, as in the homogeneous case, is im-
plicitely xed by this denition. We shall sometimes refer to these two map-
pings Input and Output, when we want to single out the components of
().
(ii): The above denition 1.2.2 diers slightly from the usual one, since
we allow also innitary heterogeneous (partial) operations and therefore we
cannot work just with the monoid S

of nite empty or non-empty


words over S for the input sequences, where is then usually just a mapping
from into S

S.
Thus, let us look at the above denition for the nitary case in more
detail, and let us assume that all arities are natural numbers, e.g. let us
consider one total ()-ary heterogeneous operation, say
A
, on (A
s
)
sS
.
Then Input() =: s

0
s

1
. . . s

()1
S

and Output() =: s

such that

A
is a mapping from A
s

0
A
s

1
. . . A
s

()1
into A
s
. If is a partial
operation of the above type, its domain has to be a subset of the above
cartesian product, while all the other data remain the same.
16
In some literature it is called a -algebra.
December 8, 2002
72 Chapter I Basic denitions and properties
(iii): We shall discuss some more examples of heterogeneous partial al-
gebras in section 2, thus we do not give further ones at this place.
(iv): As far as the presentation of examples is concerned, we apply the
same methods as discussed in 1.1.6. The dierent phyla are usually sur-
rounded by a closed curve, and sometimes the whole structure is surrounded
by an additional curve, especially, when several structures are considered at
the same time. Often a sketch of this kind will substitute the formal set
theoretical description.
1.3 Cardinal and ordinal dimension
In what follows we shall often need some cardinal and ordinal numbers con-
nected with a given similarity type . In the heterogeneous case the cardi-
nality #S of the set of sorts plays an additional role. We recall from the
Appendix that the successor cardinal of a cardinal number n is denoted by
n
+
. Although we shall rst need them in section 3, we introduce them al-
ready here in order to compare them, and and to avoid these denitions to
be too much scattered around.
Denition 1.3.1 (of ordinal dimension) The ordinal dimension := ()
of a given similarity type is the least regular ordinal number
17
such that
# > #() for each .
The ordinal dimension will mainly be needed in connection with the
local description of generation (from below) of subalgebras and other sub-
structures
18
.
Denition 1.3.2 (of rank and cardinal dimension) The cardinal dimen-
sion d := d() of a given similarity type is the least additively inaccessible
(i.e. regular) cardinal number
19
c such that c > #() for each .
Often we shall need the innite cardinal dimension d

:= max
0
, d,
and in connection with heterogeneous partial algebras we sometimes also
need the heterogeneous innite cardinal dimension d
&
:= maxd

, (#S)
+
,
where S is the corresponding set of sorts. Another useful cardinal derived
from the similarity type is the rank r := r() := supc [ c < d().
17
Cf. section 23 in the Appendix.
18
Cf. section 3
19
Cf. the Appendix.
December 8, 2002
1 Homogeneous and heterogeneous partial algebras 73
Properties of (s := sup#() [ ) d d

r
= . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 0 0
0
0
,= s = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1
0
0
s = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0
2
0
1
1 < s
0
, all () nite . . . . . . . . . . .
0

0

0

0
s =
0
, at least one () innite . . . .
1

1

1

0
s =

>
0
additively accessible . . . .
+1

+1

+1

s =

>
0
additively inaccessible,
s = #() for some
+1

+1

+1

s =

>
0
additively inaccessible,
s > #() for some

Table 1.1: Ordinal and cardinal dimensions and rank


Remarks and Examples 1.3.3 (i): The cardinal dimension will mainly
be used in connection with inductivity properties of subalgebra lattices and
direct limits, especially as an upper bound for other cardinalities, while the
rank is usually more precise, when a lower bound for cardinalities is needed.
(ii): In Table 1.1 we compare the values of the above cardinal and ordinal
numbers. As an additional tool we use s := sup#() [ . As a rule
one can state that d = # except for the case when d = 2, i.e. when all
oparation symbols are unary at least one unary operation exists or
nullary, since then =
0
, as for all nitary types, where not all operations
are nullary. We do not prove the statements included there, since they easily
follow from the denitions above and the properties of cardinal and ordinal
numbers.
December 8, 2002
74 Chapter I Basic denitions and properties
2 Homomorphisms, congruence relations and
factor algebras
Next to the concept of partial algebras the one of homomorphisms between
similar partial algebras, i.e. of structure preserving mappings between them,
is of fundamental importance. They are used for the comparison of (parts
of) algebraic structures and say, via quotient homomorphisms to carry
structure from one set to another one.
There have been introduced and used dierent concepts of morphisms
between partial algebras so far somewhere in the literature. We think that
among them the one of a homomorphism which we shall introduce here is in
some way the most natural one (see also subsection 22.1 for a more detailed
discussion). This opinion is based in part on its close relationship to the
corresponding concepts for total universal algebras as well as for relational
systems. Also it has close connections to model theory, as we have already
indicated in the introduction, and which will become still more obvious espe-
cially in Chapter II. The basic concept of homomorphisms is in some respect
relatively weak and allows quite many additional properties, which will be
discussed in more detail in sections 9 and 10. In this section we restrict our
consideration of additional properties to closedness
20
and fullness in subsec-
tion 2.4, which are besides (and in particular fullness in connection with)
injectivity, surjectivity and epiness among the most important ones.
At some occasions
21
we shall be forced to deal with homomorphisms start-
ing only from relative subalgebras of special (total!) algebras and which
cannot be extended to all of the domain algebra. Thus one might think
that a concept of morphisms based on partial mappings would be more ade-
quate. However, in all those cases, where such partial mappings become nec-
essary, the concept of the factorization system
22
of (TAlg()-extendable epi-
morphisms, closed homomorphisms) will show that these morphisms (seem-
ingly based on partial mappings) are nicely tting into the category of partial
algebras with those homomorphisms dened in this section.
Moreover, this concept of homomorphism allows in the easiest way a
relationship to the well known concept of congruence relation, which is of
20
By some other authors it is called strongness, see e.g. Gratzer [G68] or Wirsing et
al..
21
See e.g. section 5.
22
See section 10.
December 8, 2002
2 Homomorphisms etc. 75
such basic importance in universal algebra with respect to total algebras
and it will also become of basic importance for partial algebras, although
we shall need at some occasions for quotients and epimorphic images some
additional aspects
23
.
Although we shall treat for quite a while explicitely only homogeneous
partial algebras, and will present at the end of this section in subsection
?? another way how to introduce heterogeneous partial algebras which
allows a treatment more closely related to homogeneous partial algebras as
the original one , we already introduce in the following subsection the
concept of homomorphisms between heterogeneous partial algebras. This
will give a better understanding of what we introduce in subsection 2.1. The
reader should extend by himself as an easy exercise all the results of this
section to the heterogeneous case
24
.
2.1 Homomorphisms
Denition 2.1.1 (of homomorphisms)
(i) (Homogeneous case): Let A and B be two (homogeneous) partial
algebras of the same similarity type = (())

, and let f : A B be any


mapping from A into B. f or more precisely the triplet (A, graph (f), B),
although we shall use this notation very rarely is a homomorphism from
A into B and we shall indicate this fact by writing f : A B iff
(*) for all , for all a A
()
and for all b A the following holds:
if
A
(a) = b, then
B
(fa) = f(b), where f a := (fa

[ ()),
and where the implication is to be read in the following way: Whenever

A
(a) exists and has value b, then
B
(f a) exists and has value f(b)
this value is also often written as fb. We shall indicate thae fact that f is a
homomorphisms from A into B by writing f : A B. By Hom(Alg())
we denote the class of all homomorphisms between partial algebras of type
.
(ii) (heterogeneous case): Let A and B be heterogeneous partial al-
gebras of a heterogeneous signature = ((s

[ ()), s

. Then a
homomorphism f : A B between heterogeneous partial algebras is a family
(f
s
: A
s
B
s
)
sS
of mappings such that for each , for each a dom
A
and for all b A
s
one has
23
See section 5.
24
See also the remarks in 2.1.2.(iii).
December 8, 2002
76 Chapter I Basic denitions and properties
gures disabled
Figure 2.1: Compatibility property of a many-sorted homomorphism
(**) if
A
(a) = b, then
B
(fa) = f
s
(b), where fa := (f
s

[ ().
For

= ((s
1
, . . . , s
n
), s) this situation is illustrated in the commutative
diagram of Figure 2.1.
Remarks 2.1.2 (i): Notice that in the case of total algebras the implication
2.1.1.(*) reduces to the well-known condition f(
A
a) =
B
(fa), and that
in the case of partial algebras there is some sort of completeness require-
ment to the structure of B with respect to the structure of A, but not vice
versa. This is also expressed in the equivalent description of (*) for the graphs
of the fundamental operations: (*) For all one has: f
(){()}
[graph
A
] graph
B
.
Notice that f
(){()}
is the mapping induced by f by componen-
twise application. The simple proof of the equivalence between (*) and (*)
is left to the reader as an exercise; the clue is the equality (fa,
B
(fa)) =
(fa, f(
A
a)).
(ii): Notice that all homomorphisms between total algebras in the usual
sense are examples for Denition 2.1.1.(i). In addition each identity mapping
from a discrete partial algebra on some set A onto any partial algebra on A
is a homomorphism in the sense of 2.1.1.(i). More concrete examples can be
found e.g. in the gures of the following subsections.
(iii): As far as heterogeneous partial algebras are concerned we only want
to point out at this place that homomorphisms between them treat dierent
phyla as if they were disjoint. This will provide us with another denition
which is from a category theoretical point of view equivalent with the usual
one of heterogeneous partial algebras in subsection ??. Until then we shall
only deal with homogeneous partial algebras in order that the fundamental
concepts will not be too much burdened with additional technicalities, which
at the end will prove to be unnecessary. However, the reader may
consider it as an exercise to extend the concepts and results of the following
subsections to heterogeneous partial algebras, before we do this at the end
of section 2 in connection with the alternative representation. It is easy
to realize that linear mappings between vector spaces and more generally
all homomomorphisms between similar total heterogeneous algebras are
December 8, 2002
2 Homomorphisms etc. 77
homomorphisms in the sense of 2.1.1.(ii), and that 2.1.2.(ii) also applies to
the heterogeneous case.
Proposition 2.1.3 Let A, B, C Alg().
(i) The identical mapping id : A A is always a homomorphism from A
into itself: id
A
: A A (sometimes also denoted by 1
A
).
(ii) If f : A B and g : B C are homomorphisms, then their com-
position g f (mostly also written as gf), where (g f)(a) := g(f(a))
(a A), is also a homomorphism g f : A CU.
(iii) If f : A B is any mapping, then f is always a homomorphism from
the discrete partial algebra A
d
into B, i.e. f : A
d
B also written
as f : A B.
Proof . Use condition 2.1.1.(*).
Remark 2.1.4 Properties 2.1.3.(i) and 2.1.3.(ii) together with the well-known
properties of mappings in general tell us that (for every similarity type ) the
class Alg() as class of objects together with the class Hom (:= Hom Alg())
of all homomorphisms between its objects forms a (concrete) category
25
, which
we denote by Alg().
2.2 Isomorphisms
2.2.1 In any category C a C-morphism f : A B is an isomorphism iff
there exists a morphism g : B A such that g f = 1
A
and f g = 1
B
; g is
then called an inverse of f and is uniquely determined by f. Two C-objects
A and B are called isomorphic in symbols: A

= B , iff there exists an
isomorphism, say i : A B, between them. Iso (C) designates the class of
all isomorphisms in C.
Because of 2.1.4 we can adopt the same denition and notation for iso-
morphisms between partial algebras, and we get the following criterion:
Proposition 2.2.2 Let A, B Alg(), and let f : A B be any homomor-
phism. Then the following statements are equivalent:
25
Cf. [HS73] or [?] for basic denitions from category theory.
December 8, 2002
78 Chapter I Basic denitions and properties
(i) f is an isomorphism between A and B.
(ii) f is bijective, and the inverse mapping f
1
: B A is also a homo-
morphism f
1
: B A.
(iii) f is bijective, and for all : f
(){()}
[graph
A
] = graph
B
.
(iv) f is bijective, and for all and for all a A
()
the following
holds:
a dom
A
iff fa dom
B
(i.e. f
(())
[dom
A
] = dom
B
).
Proof . One has to check the denitions of homomorphism and isomor-
phism above, and to observe the following two facts: gf = id
A
and fg = id
B
iff g = f
1
(whence (i)(ii)), and the equivalence between 2.1.1.(*) and
2.1.2.(*), which establishes the other equivalences.
Remarks 2.2.3 (i): From the equivalence of (i) and (iv) in 2.2.2 it easily
follows in the case when both A and B are total algebras, that every bi-
jective homomorphism from A onto B is an isomorphism, while this it not
true for arbitrary partial algebras. E.g. id
A
: A A is always a bijective
homomorphism, but it is an isomorphism iff A is discrete, too.
(ii): But even a bijective homomorphism from some partial algebra onto
(a so-called endomorphism
26
itself need not be an isomorphism, as the follow-
ing example shows (as an exercise one should decide, whether it is necessary
to choose the carrier set to be innite!):
Let N
0
be the set of all natural numbers including 0, and let 2N
0
and
2N
0
+1 designate the subsets of all even (respectively odd) numbers. Consider
the partial algebra (N
0
; ) of type (1), where graph := (n, n) [ n 2N.
Moreover, let f : N
0
N
0
be given by
f(n) :=
_
_
_
n + 2 , if n 2N
0
,
0 , if n = 1 ,
n 2 , else .
This situation is illustrated in Figure 2.2 the larger circles indicate the
operation , which also shows that f is bijective, but no isomorphism,
since f
1
(0) , dom.
The statement in 2.2.3.(i) about total algebras is strengthened in
26
See subsection 2.3 for more details.
December 8, 2002
2 Homomorphisms etc. 79

`
_
`
_
`
_
`
_
`
_
`
_
_
`

_
`

(N
0
; )
(N
0
; )
f
0 2 3 4 5 6


>
>
>
>
>
>
>
>.
>
>
>
>
>
>
>
>.
>
>
>
>.
Figure 2.2: Bijective endomorphism, but no isomorphism
Proposition 2.2.4 Let A be any non-empty partial algebra of type . Then
the following statements are equivalent:
(i) A is total.
(ii) Every bijective homomorphism f : A B from A onto any similar
partial algebra B is an isomorphism
27
.
Proof Let A be total, and let f : A B be any bijective homomorphism.
Then 2.2.2.(iv) easily implies that B is total, and that f
1
is also a homo-
morphism, thus f is an isomorphism.
Conversely, assume that A is not total. Since A ,= , we may choose some
xed element c in A and dene for every a total operation

on A by

(a) :=
_

A
a , if adom
A
,
c , else,
for any a A
()
.
A

:= (A, (

) is called an inner one-point completion of A, and obvi-


ously id
A
: A AU

is a bijective homomorphism, which is not an isomor-


phism.
Proposition 2.2.5 Let A, B and C be similar partial algebras and f : A
B, g : B CU be any homomorphisms.
(i) id
A
: A A is always an isomorphism.
(ii) If f is an isomorphism, then f
1
: B A its inverse is an
isomorphism, too.
(iii) If f and g are isomorphisms, then gf : A C is an isomorphism, too.
27
And B is then a total algebra, too.
December 8, 2002
80 Chapter I Basic denitions and properties
Corollary 2.2.6 For every similarity type the isomorphism relation

=
on Alg() is an an equivalence relation, i.e. it is reexive, symmetric and
transitive. The equivalence classes of

= are proper classes, and they are


called isomorphism types.
Remark 2.2.7 It should be noted that initial algebraic specications of ab-
stract data types only yield the specied (partial) algebras up to isomorphism.
Two isomorphic algebraic structures cannot be distinguished algebraically.
Nevertheless an appropriate set theoretic or algebraic model
28
may of-
ten help to get a better feeling for the structure than this might be achieved
by other models.
2.3 Endomorphisms and automorphisms
Denition 2.3.1 Let A be any partial algebra. The homomorphisms from
A into itself are called endomorphisms of A, and those endomorphisms of A,
which are at the same time isomorphisms, are called automorphisms of A.
By End (A) and Aut A we denote the sets of all endomorphisms (respectively
automorphisms) of A.
Proposition 2.3.2 For every partial algebra A one has:
(i) End A := (End A); , id
A
) is a monoid,
(ii) Aut A := (Aut A; , id
A
,
1
) is a group,
where is the composition of homomorphisms.
Proof This is an immediate consequence of previous statements and the
fact that the composition of mappings and therefore of homomorphisms
is associative, whenever it is dened (and here it is always dened).
Remark 2.3.3 It should be noticed that in general as the example in
2.2.3.(ii) shows a bijective endomorphism need not be an automorphism;
but it is a simple set theoretical argument to show that every bijective endo-
morphism of a partial algebra with a nite carrier set is an automorphism.
From 2.2.4 one may easily infer that every bijective endomorphism of a total
algebra is an automorphism.
28
These may be obtained e.g. by using well known algebraic constructions or parts of
well known structures.
December 8, 2002
2 Homomorphisms etc. 81
Sometimes it might be useful to know the following result:
Proposition 2.3.4 (CAYLEY, BIRKHOFF) Let M := (M; , e) be any
monoid, and dene for every m M the mappings
m
: M M, x mx,
and
m
: M M, x x m (x M).
(i) (End (M, (
m
)
mM
); , id
M
) = (
m
[ m M; ,
e
)

= M,
i.e. every monoid is isomorphic to the endomorphism monoid of some
unary total algebra.
(ii) If M is a group, then M is isomorphic to the automorphism group of
(M, (
m
)
mM
) in the way described in (i).
Remarks 2.3.5 (i): Notice that in the above result the similarity type de-
pends on #M. In [PuTr80] by A.PULTR and V.TRNKOVA much stronger
results of similar nature can be found.
(ii): The results in 2.3.4 show that every semigroup, monoid or group
can be considered up to isomorphism as a semigroup, monoid or group of
mappings on a suitable set. The algebraic structure is used to select the
appropriate mappings.
2.4 Full and closed homomorphisms
2.4.1 The denition of a homomorphism between partial algebras as given
in 2.1.1 is a relatively weak one, since it contains among others a state-
ment only about the (minimal) richness of the structure in the codomain
depending on the structure on its domain, but not vice versa. This fact
and more precisely some model theoretic aspects of 2.1.1 (e.g. reection
of formulas) which we shall discuss in detail in Chapter II allows to in-
troduce many more interesting properties of homomorphism between partial
algebras than can be done in the case of total algebras. We shall discuss this
in more detail in later chapters. In this introductory chapter we restrict our
considerations to the following two properties, which are both trivially sat-
ised by all homomorphisms between total algebras. It should be observed
that according to denition 2.1.1.(ii) what follows also applies without
changes to heterogeneous partial algebras.
Denition 2.4.2 Let f : A B be any homomorphism between partial
algebras of the same type = (())

.
December 8, 2002
82 Chapter I Basic denitions and properties
(i) f is said to be a full homomorphism from A into B iff one has:
For every , and for every a A
()
:
(fh) If fa dom
B
and
B
(fa) f[A],
then there exists a

dom
A
such that fa = fa

.
(ii) f is said to be a closed homomorphism from A into B iff one has:
(ch)
For every and for every a A
()
one has:
If fa dom
B
, then a dom
A
.
Remarks 2.4.3 (i): Notice that in G.GR

ATZERs book [G68] those homo-


morphisms, which we call closed, are called strong, and this terminology
is also used in many papers on partial algebras. But in papers of J.SCHMIDT
and of the author this property has already been called closed, while strong
was used for what GR

ATZER and also we here in this book call full. Confu-


sion is even greater, since strong has been used in papers of J.S LOMI

NSKI
and B.WOJDY LO in a third meaning
29
. This is the main reason that we try
to avoid strong in connection with homomorphisms throughout this book.
(ii): Intuitively spoken a homomorphism is em full, iff the restriction to the
exact image set of the structure in the codomain is the one induced by
the structure of the preimage via the homomorphism
30
. Similarly, a homo-
morphism is closed, iff the domains of the operations in the preimage are
exactly the inverse images of the domains of the corresponding operations in
the codomain.
Proposition 2.4.4 Let f : A B be any homomorphism between similar
partial algebras.
(i) f is full, iff for every one has:
f
((){()})
[graph
A
] = graph
B
(f[A])
(){()}
.
(ii) f is closed, iff for every one has: (f
(())
)
1
= dom
A
.
(iii) If f is closed, then it is full, but not vice versa.
29
See [BWo82] for further details.
30
See the Homomorphism Theorem below and the concept of relative subalgebras in
section 3.
December 8, 2002
2 Homomorphisms etc. 83

_
`

_
`

_
`

_
`

_
`

_
`

_
`

f
1
f
2
f
3
f
4
neither full
nor closed
full, but
not closed
closed (and
therefore full)
bijective, but
no isomorphism

`
`
`
`
`
`
` ` ` `
>
>
>
>
>
>
>
>

Figure 2.3: Examples of properties of homomorphisms


Proof (i) and (ii) follow immediately from the properties (fh) and (ch),
respectively, and it is easily seen that (ch) implies (fh), whence we have the
rst statement in (iii). Finally observe that the homomorphism f
2
indicated
in Figure 2.3 is full, but not closed, while f
1
is not even full.
Proposition 2.4.5 (i) The composition of closed homomorphisms is closed.
(ii) The composition of full homomorphisms is not necessarily full.
(iii) Let f : A B and g : B C be homomorphisms, then one has:
(a) If f and g are full, and if g is injective or if f is surjective, then
gf is full.
(b) If gf is full and g is injective, then f is full.
(c) If gf is full, and if f is surjective, then g is full.
(d) If gf is full and surjective, then g is full and surjective.
(e) If gf is closed, then f is closed.
(f ) If gf is closed, and if f is surjective, then g is closed.
Proof In what follows let f : A B and g : B C be arbitrary homo-
morphisms.
December 8, 2002
84 Chapter I Basic denitions and properties

_
`

_
`

_
`

f
g

`
`
`

Figure 2.4: f and g are full, but gf is not full.


Ad (i): Let f and g be closed, and assume gfadom
C
for some a A
()
and some . Then fa dom
B
, since g is closed, therefore a dom
A
,
since f is closed.
Ad (ii): Consider Figure 2.4.
Ad (iii): The simple proofs are all quite similar and will, moreover, follow
from more general results in 9.1.7 and 9.2.16.(iv) in connection with reection
and weak reection of formulas by mappings. Therefore we leave them at
this point as exercises.
In connection with the properties of homomorphisms discussed above we
also get new criteria for isomorphisms, when we apply Proposition 2.4.4 to
the results of Proposition 2.2.2:
Proposition 2.4.6 Let f : A B be a homomorphism between similar
partial algebras. Then the following statements are equivalent:
(i) f is an isomorphism.
(ii) f is bijective and closed.
(iii) f is bijective and full.
Proposition 2.4.7 Any homomorphism starting from a total algebra is closed.
This is an immediate consequence of the Denition 2.4.2.(ii) of closedness.
December 8, 2002
2 Homomorphisms etc. 85
2.5 Congruence relations
2.5.1 In group theory the surjective homomorphisms starting from a given
group G are characterized (up to unique isomorphism) by the normal sub-
groups of G. For rings and Boolean algebras this is done by so-called ideals,
but in each case these are special subobjects, while already in the cases of
semigroups, monoids or lattices one cannot nd subobjects which do this
job. But it is also well known in the cases of groups, rings or Boolean alge-
bras that the distinguished subobjects mentioned above induce equivalence
relations on the carrier sets such that these relations are compatible with
the fundamental operations. This fact can be generalized to universal and
partial algebras.
Denition 2.5.2 Let A be any partial algebra of some type (())

. A
relation A A is called a congruence relation (or briey a congruence)
on A, iff it satises (Cong 1) and (Cong 2) below:
(Cong 1) is an equivalence relation on A.
(Cong 2) For every and for any two sequences a, b dom
A
one has:
If (a
k
, b
k
) for each k (), then (
A
(a),
A
(b)) .
will be called a closed congruence relation (or a closed congruence) on A,
iff satises in addition:
(Cong 3c) For every and for any sequences a, b A
()
one has:
If a dom
A
, and (a
k
, b
k
) for all k (), then b dom
A
.
By Cong A we denote the set of all congruence relations on A, while Cong
c
A
designates the set of all closed congruence relations on A. Moreover, we
introduce:
Cong A := (Cong A; ), and Cong
c
A := (Cong
c
A; ), the corresponding
partially ordered sets with respect to set theoretical inclusion.
The most immediate consequences of this denition are collected in the
Corollary 2.5.3 For every partial algebra A and for every homomorphism
f : A B one has:
(i) The identity relation :=
A
:= (a, a) [ a A and the universal
relation :=
A
:= AA are always congruence relations on A. is
always closed, while is closed iff every
A
is either total or discrete
(some may be total while others may be discrete!).
December 8, 2002
86 Chapter I Basic denitions and properties
(ii) The relation
f
:= (a, a

) [ a, a

A, f(a) = f(a

) is always a
congruence relation on A. It is called the congruence relation induced
by f or the kernel of f, and it is usually denoted by ker f.
(iii) If f is closed, then ker f is closed, too.
Proof Ad (i): The rst statements in (i) are obvious from the denition.
Assume that each fundamental operation of A is either total or discrete, then
obviously (Cong 3c) is satised by . Now assume that is closed, and let
a, b A
()
such that a dom
A
; then a and b are componentwise equivalent
with respect to , and therefore b dom
A
by (Cong 3c) showing that
A
is total whenever it is not discrete and is closed.
Ad(ii): It is well known that
f
is always an equivalence relation. Let
a, a

dom
A
such that f(a) = f(a

), then
f(
A
(a)) =
B
(fa) =
B
(fa

) = f(
A
(a

)),
showing that also (Cong 2) is satised.
Ad (iii): If, nally, f is a closed homomorphism, then a dom
A
, and
fb = fa ( dom
B
), then also b dom
A
by the closedness of f.
Recalling that d designates the cardinal dimension, and that d

= max
0
, d
we get:
Proposition and Notation 2.5.4 Let A be any partial algebra. Then:
(i) The set theoretical intersection of arbitrarily many congruence relations
on A is always a congruence on A, thus Cong A is a closure system
on AA; and the corresponding closure operator
31
is denoted by Con
A
,
i.e. for every binary relation M on A one has
Con
A
=

[ Cong A, and .
Con
A
is called the congruence (relation) of A generated by .
(ii) The closure system Cong A is d

-inductive, and Con


A
is a d

-algebraic
closure operator. Hence, for each A A one has:
Con
A
=
_
Con
A
[ and # < d

.
31
See section 24 in the Appendix.
December 8, 2002
2 Homomorphisms etc. 87
(iii) Cong
c
A forms an initial segment of Cong A (i.e. whenever ,
Cong A, , and is closed, then is closed). Moreover, Cong
c
A
is d

-inductive.
Proof It is routine to check that the intersection of any family of congru-
ences of a partial algebra A is itself a congruence, where the intersection of
an empty family is dened to be
A
(see subsection 2.11 for the intersection
of the empty family of congruences of heterogeneous partial algebras). Thus
(i) easily follows from the general facts about closure systems and closure
relations (cf. section 24).
Ad (ii): Let B Cong A be any d

-directed system of congruences of A; we


have to show that

:=

B is again a congruence on A: Obviously,

is reexive and symmetric, and it is also transitive, since (a, b), (b, c)

imply (a, b) and (b, c)

for some ,

B. By the d

-directedness
of B there is in B containing both and

; therefore (a, c)

.
Now assume a, b dom
A
and (a
k
, b
k
)

for all k (); thus for


each k () we have (a
k
, b
k
)
k
B. Since #() < d

, and since
B is d

-directed, there is

B containing all
k
, and hence all (a
k
, b
k
).
Since

is a congruence, we get (
A
(a),
A
(b))

, showing that

is a congruence relation on A. If, in addition, all congruences of B


are closed, then

is closed, too: Namely, assume that a, b A


()
are
componentwise

-equivalent as above, and that a dom


A
; then we can
nd as above a

B such that (a
k
, b
k
)

for every k ()
(notice that B is non-empty, since it is d

-directed, and the empty subset


has an upper bound in B). Since

is closed by assumption, we may infer


that b dom
A
, i.e.

satises (Cong 3c). Finally let us observe


that Con
A
[ , # < d

is d

-directed, since d

is additively
inaccessible, and therefore the union of less than d

-many sets of cardinality


less than d

again has cardinality less than d

. Thus the union of the above


set of d

-small generated congruences is again a congruence contained in


Con
A
, but containing ; therefore equality has to hold.
Ad (iii): That Cong
c
A is d

-inductive has already been shown in connection


with (ii); and it is easy to realize that any congruence contained in a closed
congruence satises (Cong 3c) and is therefore closed, too.
In the nitary case we can say more in connection with closed congru-
ences:
December 8, 2002
88 Chapter I Basic denitions and properties
Lemma 2.5.5 If the similarity type under consideration is nitary, then the
supremum of any two closed congruences on A is again closed, moreover it
is the equivalence theoretical supremum of the two congruences under consid-
eration, i.e.:
For any two closed congruences ,

Cong
c
A one has that their equiv-
alence theoretical and their congruence theoretical supremum are identical,
i.e.
()
Con
A
(

) =
_
(

)
n
[ n <
0
= sup
e
(,

) =: sup
e

Cong
c
A,
where

:= (r, s) [ there exists t A such that (r, t)

and (t, s)
is the relational product of and

, (

)
0
:=
A
and (

)
n+1
:=
(

)
n
(

).
Moreover, there exists a largest closed congruence
c,A
on A, namely the
set theoretical union of all closed congruences on A is again a closed congru-
ence on A.
2.5.6 Notice that the assumptions of Lemma 2.5.5 are in particular satised
for every nitary total algebra A and all congruences on A.
Observe, too, that, for every set A and for every binary relation on
A,

n<
0

n
is the transitive closure of , and that, for any equivalence
relations and

on A one has

, and for arbitrary relations


one has to use rather

n<
0

than

n<
0

to get the transitive


closure.
Proof of 2.5.5: W.l.o.g. let all () be natural numbers, say k

, let
:=

)
n
[ n <
0
be the equivalence theoretical supremum of

, and let := Con


A
(

) be its supremum within the congruence


lattice. Then by transitivity is contained in every equivalence relation
containing both and

, and therefore in particular in . If we can show


that is itself a congruence, then = : Obviously, is an equivalence
relation, since so are and

. Thus there remains to show (Cong 2) and


(Cong 3c): By the denition of there exist for any (a, b) elements z
0
:=
a, z
1
, . . . , z
2n1
, z
2n
:= b, such that for 0 i < n one has: (z
2i
, z
2i+1
)
and (z
2i+1
, z
2i+2
)

. We shall call such a sequence a

-chain of
length n between a and b (such a sequence can always be assumed to have
December 8, 2002
2 Homomorphisms etc. 89
2n + 1 elements, since we may repeat elements because of reexivity!). Let
us now consider any operation symbol , and let us set for its arity
k := k

( = ()). Moreover, let a = (a


1
, . . . , a
k
), b = (b
1
, . . . , b
k
) A
k
,
such that a dom
A
. We claim that b dom
A
and (
A
(a),
A
(b)) :
First let us observe that we may choose the

-chains between a
i
and b
i
to be all of the same length n := maxn
1
, . . . , n
k
, where n
i
is the original
length of such a chain (1 i k), by repeating elements if necessary.
Thus, let z
i
0
= a
i
, z
i
1
, . . . , z
i
2n1
, z
i
2n
= b
i
be the corresponding

-
chains, and let z
r
:= (z
1
r
, . . . , z
k
r
) (0 r 2n). By the closedness of
and

and by (Cong 2) it is not dicult to prove by induction on r that


z
r
dom
A
(since z
0
= a dom
A
), and that (
A
(z
2i
),
A
(z
2i+1
)
and (
A
(z
2i+1
),
A
(z
2i+2
)

; thus (
A
(a),
A
(b)) , since z
0
= a and
z
2n
= b; moreover, b dom
A
. As a consequence of what we have
just shown we get that Cong
c
A is (
0
-) directed, since it is non-empty and
contains with any two congruences also their supremum. By inductivity its
union is a closed congruence; this was to be shown.
Remarks and Examples 2.5.7 (i): Closedness of both congruences in-
volved in Lemma 2.5.5.() is necessary. For instance one might consider the
ve-element partial algebra of Figure 2.5, where the closed curves indicate
the congruence classes of a congruence
1
, while the broken ones indicate the
classes of a congruence
2
. Then (d, e) belongs to the congruence relation
generated by
1

2
, but not to the equivalence relation generated by it.
(ii): Notice that in the innitary case there exists in general no greatest
closed congruence relation. As an example consider the partial algebra (N; )
(N designating the set of all natural numbers) of similarity type (N) such that
dom := a := (a
0
, . . . , a
n
, ..) [ innitely many elements of N
occur innitely often in a ;
(a) := 0,
whenever a dom.
Let
n
:= (k, l) [ 1 k, l n for all n N.
Then each
n
is a closed congruence relation on (N; ), since from a
sequence belonging to dom only nitely many values can be equivalent
modulo
n
to dierent elements in the corresponding component of a second
sequence, which therefore has to have innitely many elements of N occurring
innitely often, while
n
trivially satises (Cong 1) and (Cong 2). But the
union (= supremum) of all the
n
is , which is not closed (cf. 2.5.3.(i)).
(iii): As an application of the above result one may consider a partial
Mealy-automaton (I

S

O; , ) =: A of type (2, 2), where : I S S
December 8, 2002
90 Chapter I Basic denitions and properties

_
`

_
`

d
a
b
c
e
1
:

2
:
` `
Figure 2.5: The equivalence generated by two congruences is in general no
congruence
and : I S O. Then a reduction of A is a closed homomorphism
f : A B onto a similar automaton B such that the restriction of f to I O
is the identity mapping (the input sets (I) (respectively the output sets (O))
of A and B are assumed to be the same, and only the set of states S =: S
A
and
S
B
may dier). Then it is easy to check that there exists a smallest reduct
of A corresponding to the largest closed congruence relation of A, which has
only one-element congruence classes for the elements of I O. Such a largest
closed congruence relation exists because of the following lemma:
Lemma 2.5.8 . Let A be a nitary partial algebra and M any subset of A.
Let B be a chain or directed system of (closed) congruence relations which
have only one-element congruence classes for each element of M (i.e. none
of them identies two dierent elements of M). Then

B is a (closed)
congruence of A with the same property with respect to M. In particular
there always exists (in the nitary case!) a maximal (closed) congruence
on A not identifying element of M with any element of A. And in the case
of closed congruences there is even a largest one with these properties.
If we require above that the congruences in B do not identify any two
dierent elements from M, then

B also has this property, and therefore


(still only in the nitary case) there always exists a maximal (respectively
maximal and closed) congruence with these properties.
Proof As far as closed congruences are concerned, this is an immediate
consequence of Lemma 2.5.5() and of the inductivity of Cong
c
A, and in
December 8, 2002
2 Homomorphisms etc. 91
general case of the inductivity of Cong A. And in both cases one has to
observe that, for any (
0
-) directed system B of congruence relations, its
union

B cannot identify any element of M with any element of A (or with


any other element of M), if no member of B does so.
2.6 Factor algebras
2.6.1 In Corollary 2.5.3 we have already realized that the equivalence re-
lation
f
induced by a homomorphism f : A B is always a congruence
relation on A. One aim of the following construction is to show the converse:
that every congruence relation on A is induced by a (full) homomorphism
(and every closed congruence relation by a closed homomorphism), which
(in each case) can also be assumed to be surjective). This will just be the
natural projection nat

from A onto its factor algebra A/.


Moreover, the semantics of abstract data types (i.e. so-called initial al-
gebras
32
) are always factor algebras of the corresponding term algebra
33

generated by the empty set of variables with respect to (fully invariant


34
)
congruence relations generated by the axioms of the corresponding speci-
cations. And this yields a strong motivation for the investigation of factor
algebras in general also in computer science.
Denition 2.6.2 and Construction of Factor Algebras. Let A be any
partial algebra of type (())

, and let be any congruence on A. By [x]

we denote the congruence class of x A with respect to , and we dene as


usual:
A/ := [x]

[ x A, the set of all congruence classes of elements of A,


and nat

: A A/, x [x]

, the natural projection from A onto A/.


In order to get an algebraic structure on A/, the quotient set of A with
respect to , we dene for each :
dom
A/
:= ( [x
k
]

[ k () ) [ ( x
k
[ k () ) A
()
, and there is
( y
k
[ k () ) dom
A
such that
(x
k
, y
k
) for each k () ;
and if ([x
k
]

[ k ()) dom
A/
, and if (y
k
[ k ()) dom
A
32
See 17.6.1.
33
Cf. 5.2.3.
34
See 6.3.5.
December 8, 2002
92 Chapter I Basic denitions and properties
satises (x
k
, y
k
) for each k (), then we dene

A/
([x
k
]

[ k ()) := [
A
(y
k
[ k ())]

.
Notice that this denition is independent from the choice of the sequence
y. Namely, let z := (z
k
[ k ()) dom
A
be a sequence, which is also
componentwise -equivalent to x. Then by the transitivity of we
have (y
k
, z
k
) for each k (), and by (Cong 2) (
A
(y),
A
(z))
, i.e. the two representative sequences lead to the same congruence class.
Thus for each
A/
is a well-dened ()-ary partial operation
on A/ such that nat

becomes a full homomorphism nat

: A A/ :=
(A/, (
A/
)

). A/ is called the factor algebra or quotient algebra of A


with respect to the congruence notice that we shall omit here the
adjective partial even though A/ will usually also be partial when A is
partial.
We reformulate the above observations in
Proposition 2.6.3 Let A be any partial algebra, and let be any congruence
on A. Then the quotient algebra A/ as dened in 2.6.2 is a partial algebra
on A/ such that the natural projection nat

becomes a full and surjective


homomorphism nat

: A A/, and is the congruence relation on A


induced by nat

: = ker nat

.
We shall see later that the structure on A/ is the weakest one on A/,
which allows nat

to be a homomorphism.
Moreover, is closed iff nat

is closed.
Proof From the construction in 2.6.2 most of the statements are obvi-
ous, one might even say that one makes the structure of A/ in such a
way that nat

becomes a full and surjective homomorphism. By recalling


2.5.2.(Cong 3c) and the denition of closed homomorphisms in 2.4.2.(ii) also
the last statement becomes obvious in connection with 2.6.2.
Corollary 2.6.4 The quotient algebra of any partial algebra A is always of
the same similarity type as A. And if a fundamental operation, say
A
, of A
is total (discrete), then
A/
is total (discrete) for each quotient algebra A/
of A, too.
December 8, 2002
2 Homomorphisms etc. 93

a
b
0
1

Figure 2.6: A four element lattice


Remark 2.6.5 It should be observed that the homomorphisms between par-
tial algebras as well as the congruence relations of any given partial algebra
depend very much on the chosen similarity type. Because of the equiva-
lence between full and surjective homomorphisms and congruence relations
which will be made even more evident in the next subsection it suf-
ces to consider this dependence for congruence relations. As an example
we consider the four element lattice (cf. its Hasse diagram in Figure 2.6)
both as a lattice (M
2
; , ) of type (2, 2) and as a meet semilattice (M
2
; ).
Let the equivalence relation on M
2
have the equivalence classes 1 and
o, a, b; then it is easy to check that is a congruence relation of the meet
semilattice but not of the lattice.
That every semigroup homomorphism between groups is already a group
homomorphism, and that every semigroup congruence of a group is already
a group congruence are important exceptions which cannot be generalized.
e.g. a non-surjective semigroup homomorphism between two semigroups of
type (2) which both have a (left as well as right) neutral element need not
carry the neutral element of the source semigroup onto the one of the target
semigroup, i.e. it need not be a monoid homomorphism.
2.7 The Homomorphism Theorem
The core of diagram completion theorems, among which the Homomorphism
Theorem is one of the most important ones, is the following set theoretical
result.
Lemma 2.7.1 (Diagram Completion Lemma) Let f : A B and g :
A C be arbitrary mappings such that f is surjective, and let
f
and
g
be
the equivalence relations on A induced by f and g, respectively. Then there
exists a mapping h : B C such that g = h f, iff
f

g
(see Figure 2.7).
December 8, 2002
94 Chapter I Basic denitions and properties
gures disabled
Figure 2.7: Diagram completion for surjective mappings
Moreover, if h exists, then:
(a) h is uniquely determined by f and g.
(b) h is injective, iff
f
=
g
.
(c) h is surjective, iff g is surjective.
Proof Let us rst assume that h : B C exists and satises g = h f,
and let (a, b)
f
; then g(a) = h f(a) = h f(b) = g(b), i.e. (a, b)
g
.
Conversely, assume
f

g
; then in order to achieve g = h f we
have to dene:
() h(b) := g(a), whenever f(a) = b, for all b B.
Since f is surjective, h assigns to each b B at least one value. Let us now
assume that f(a) = f(a

) = b; then (a, a

)
f
, and because of
f

g
we have (a, a

)
g
, i.e. g(a) = g(a

). Therefore h(b) is uniquely determined


by ().
For the rest of the proof let us assume that
f

g
is satised, i.e. that
h exists. (a) is then an immediate consequence of the surjectivity of f, and
for the same reason both directions of (c) hold. In order to prove (b) let us
assume
f
=
g
. Then h(b) = h(b

), b = f(a) and b

= f(a

) imply g(a) =
g(a

), and therefore b = b

. Conversely, let h be injective and (a, a

)
g
;
then h f(a) = g(a) = g(a

) = h f(a

), and f(a) = f(a

) because of the
injectivity of h; hence (a, a

)
f
.
Corollary 2.7.2 Let, in addition to the assumptions of the above Lemma,
A, B and C be similar partial algebras, and let f and g be homomorphisms
between them such that f is full (besides being surjective). Then h is a
homomorphism, whenever it exists.
Moreover, if h exists, then:
(d) h is a full homomorphism, iff g is a full homomorphism.
December 8, 2002
2 Homomorphisms etc. 95
gures disabled
Figure 2.8: Dual homomorphism theorem for full and injective f
(e) h is an isomorphism, iff g is a surjective and full homomorphism, and

f
=
g
.
(f ) If g is closed, then h is closed, too; if f is closed, then h is closed iff g
is closed.
Proof Let the assumptions of the Corollary be satised. Let b dom
B
;
since f is surjective and full, there is a sequence a dom
A
such that
f a = b. Thus h b = h f a = g a. Therefore h b dom
C
, since g is a
homomorphism. Therefore:
h(
B
(b)) = h(
B
(f a)) = (h f)(
A
(a)) = g(
A
(a)) =
C
(g a) =
C
(h fa) =
C
(b),
showing that h is a homomorphism, too.
Now, (d) is a consequence of 2.4.5.(iii).(a) and (c) and the surjectivity of f;
(e) follows from 2.4.6.(iii) using 2.7.1 above. And (f) is a consequence of
2.4.5.(iii).(f) and 2.4.5.(i).
Remarks 2.7.3 (i): The results above show that the role of normal sub-
groups in group theory or ideals in the theory of rings or Boolean lattices is
played in general by congruence relations. Thus in universal algebra
the kernel of a homomorphism f between similar partial algebras is indeed
the induced congruence relation
f
; therefore it really deserves this notation
as kernel:
f
= ker f.
(ii): It is usually less known since almost trivial that 2.7.1 and 2.7.2
have dual statements:
(a) Let f : A B and g : C B be mappings such that f is injective.
Then there exists a (unique!) mapping h : C A such that f h = g,
iff g[C] f[A] (see Figure 2.8). h is surjective, iff g[C] = f[A], and
it is injective, iff g is injective.
(b) If A, B and C are partial algebras such that f and g above are homo-
morphisms f being full and injective , then h is a homomorphism
whenever it exists. Moreover, h is full iff g is full; if g is closed, then h is
December 8, 2002
96 Chapter I Basic denitions and properties
gures disabled
Figure 2.9: Two diagrams corresponding to Theorem 2.7.4
closed; if f is closed, then h is closed iff g is closed; h is an isomorphism
iff g is injective and full, and f[A] = g[C].
Proof The easy proofs are left as exercises.
The results of 2.6 and 2.7 so far combine to the
Theorem 2.7.4 (First Homomorphism Theorem) Let g : A C be
any homomorphism between similar partial algebras A and C, and set :=
ker g.
(i) Then there exists a unique injective homomorphism i : A/ C such
that g = i nat

.
(ii) i (from (i)) is an isomorphism iff g is surjective and full.
(iii) If j h is any factorization of g, where h : A D is a full and surjective
homomorphism and j : D C is an injective homomorphism (and
g = j h), then there exists a unique isomorphism k : A/ D such
that h = k nat

and j k = i.
This can also be expressed by saying that the diagrams in Figure 2.9 are in
all parts commutative.
Remark 2.7.5 With respect to what we shall discuss in Chapter II, sec-
tion 10, the statement in 2.7.4.(iii) means that the class of all full and surjec-
tive homomorphisms as the left factor and the class of all injective
homomorphisms as the right factor form a factorization system in
the category of all partial algebras and their homomorphisms, since it can
easily be checked that both classes are closed with respect to composition
and contain all isomorphisms.
Denition 2.7.6 Let A and B be similar partial algebras, then we introduce
the following notation:
December 8, 2002
2 Homomorphisms etc. 97
B is called a (weak) homomorphic image of A, iff there exists a surjec-
tive homomorphism from A onto B.
B is called a full (or strong) homomorphic image of A, iff there exists
a full and surjective homomorphism from A onto B.
B is called a closed homomorphic image of A iff there exists a closed
and surjective homomorphism from A onto B.
Remarks 2.7.7 (i): The First Homomorphism Theorem tells us that the
total information about a full homomorphic image B of some partial algebra
A with respect to a full and surjective homomorphism, say f : A B is
contained in the ordered pair (A, ker f), i.e. within the structure of A. Hence
there is up to unique isomorphism only a set of representatives of the classes of
equivalent full homomorphic images of A (equivalence meaning that they
belong to the same congruence relation). But it should also be observed that
two full homomorphic images of A may be isomorphic without the existence
of an isomorphism carrying one of the full and surjective homomorphisms into
the other one: Take e.g. the void type (i.e. sets), and consider for a set A
with at least two elements the two projections pr
1
: AA A ((a, b) a)
and pr
2
: A A A ((a, b) b), then we have two full homomorphic
images of AA, which are not equivalent in the above sense, but which are
isomorphic (via id
A
; we could have done the same, say, for a total structure
on A and for the product A A (see section 4)).
(ii): It should be observed that in the case of total algebras the three
kinds of homomorphic images dened above coincide; therefore in this case
each homomorphic image of a total algebra is totally described within A by
its kernel. We shall discuss later how to describe other kinds of homomorphic
images of partial algebras in close connection with A and the corresponding
homomorphism. From 2.7.2 we may already infer that the closed homomor-
phic images of A are (up to isomorphism) in one-to-one correspondence with
the closed congruence relations of A.
(iii): It should have become clear so far from the above results, why
we had to require the surjective homomorphism f in 2.7.2 to be full, since
only then we can guarantee that an existing diagram completion of the set
theoretical case is always a homomorphism.
According to the work of Bourbaki
35
this fact is described as: Each full
and surjective homomorphism is a nal homomorphism, and the quotient
35
See e.g. [Bou61], Chapitre I, ]2.
December 8, 2002
98 Chapter I Basic denitions and properties
structure on a factor algebra is a nal structure. Similarly, 2.7.3.(ii).(b)
tells us that each full and injective homomorphism is according to Bour-
baki an initial homomorphism, and the relative substructure on a sub-
set of A is an initial structure
36
.
2.8 The congruence lattice of a full homomorphic im-
age
First we state a simple but useful fact:
Lemma 2.8.1 Let f : A B and g : B C be homomorphisms; then
ker f ker g f = (f
(2)
)
1
[ker g].
Proposition 2.8.2 Let f : A B be a homomorphism,
1
Cong A,

2
Cong B.
(i) If f is full and surjective, and ker f
1
, then f
(2)
[
1
] Cong B.
(ii) (f
(2)
)
1
[
2
] Cong A and ker f (f
(2)
)
1
[
2
].
(iii) If f is closed and surjective, and if
1
and
2
are closed congruence
relations, respectively, and if ker f
1
, then f
(2)
[
1
] and (f
(2)
)
1
[
2
]
are closed congruence relations, too.
Proof Ad (i): From the Diagram Completion Lemma 2.7.1 and its corollary
we conclude that there exists a homomorphism g : B A/
1
such that
g f = nat

1
. Obviously, f
(2)
[
1
] ker g, but because of the surjectivity of f
also the other inclusion holds, showing that f
(2)
[
1
] = ker g is a congruence
on B (see Figure 2.10).
Ad (ii): Notice that ker f = (f
(2)
)
1
[
B
] (f
(2)
)
1
[
2
] = ker (nat

2
f),
where the last equality follows from 2.8.1 and shows that (f
(2)
)
1
[
2
]
Cong A.
Ad (iii): If f is closed and surjective, and if
1
is closed and contains
ker f, then the homomorphism g in the proof of (i) is closed according to
36
See 3.1.2 and also 9.2.1.
December 8, 2002
2 Homomorphisms etc. 99
gures disabled
Figure 2.10: f
(2)
[
1
] is a congruence on B
if
1
ker f.
2.4.5.(iii).(f); therefore f
(2)
[
1
] is a closed congruence according to (i) and
2.5.3.(iii). Also from 2.5.3 we conclude that (f
(2)
)
1
[
2
] is closed, since
nat

2
f is closed
37
.
Corollary 2.8.3 Let f : A B be a full and surjective homomorphism.
(i) f
(2)
induces an isomorphism between the interval ([ker f,
A
]; ) of
Cong A and all of Cong B ( f
(2)
[]).
(ii) If, in addition, f is closed, then f
(2)
induces, too, an isomorphism
between the principal end ([ker f, ]; ) of Cong
c
A and all of Cong
c
B
on the other hand (see Figure 2.11).
Proof It remains only to show that the assignment f
(2)
[] is injective,
but this follows easily from the assumption that ker f .
Remark 2.8.4 We would like to observe that in the case, when f above is
closed, but not necessarily surjective, then the preimage of a closed congru-
ence is also always closed.
2.9 The analogue of the Second Isomorphism Theorem
of E.Noether
Notation 2.9.1 If A is any partial algebra, and and

are congruences
on A such that

, then we denote by

/ the congruence nat


(2)

]
of the factor algebra A/, i.e. the image of

in A/ (see 2.8.2.(i)).
Theorem 2.9.2 Let A be any partial algebra, and let ,

Cong A such
that

. Then one has


A/


= (A/)/(

/),
where the isomorphism is given by the assignment [a]

[[a]

/
(a A).
37
Cf. 2.7.2.(f) and 2.4.5.(i)).
December 8, 2002
100 Chapter I Basic denitions and properties
,
,
,
,
,
_

_
`

_
`

A
ker f
f
(2)
Cong A

B
Cong B
,
,
,
,
,
_

_
`

_
`

A
ker f
f
(2)
Cong
c
A
Cong A

B
Cong
c
B
Cong B
f is closed
. . . . . . . . . . . .

. . . . . .

Figure 2.11: Eect of a (closed) quotient homomorphism on the (closed)


congruences
gures disabled
Figure 2.12: An analogue of the Second Isomorphism Theorem
Proof Consider the diagram in Figure 2.12 with

= ker nat

= ker (nat

/
nat

),
where nat

, nat

and nat

/
and therefore also nat

/
nat

are full
and surjective homomorphisms. Thus the required isomorphism exists ac-
cording to 2.7.2.(e).
2.9.3 The Second Isomorphism Theorem of group theory
38
is an immediate
consequence of the above theorem, since there is a one-to-one correspondence
between the normal subgroups and the congruence relations of a group.
The analogue of the First Isomorphism Theorem of group theory can only
be formulated in the next section, when substructures are around.
Observe that in 2.9.2 for a closed congruence

the congruences and

/ are closed, too.


38
Cf. e.g. van der Waerden [vdW66], p. 150.
December 8, 2002
2 Homomorphisms etc. 101
2.10 Mono- and epimorphisms
2.10.1 In classical algebra and abstracted from it in category theory
two kinds of morphisms play an important role: the monomorphisms and
their dual, the epimorphisms. We have seen that in the theory of partial
algebras there are additional properties which are, as we shall see in
Chapter II, of a more model theoretical nature, since they are connected
with the preservation or reection of special formulas , but mono- and
epimorphisms are sort of a basis for those additional properties, too.
Denition 2.10.2 Let C be any category. A morphism f : A B between
any two C-objects A and B is called a monomorphism, iff for any two mor-
phisms g, h : C A one has that f g = f h always implies g = h. Dually, f
is called an epimorphism, iff for any two morphisms k, l : B D k f = l f
always implies k = l. Mono C and Epi C designate the corresponding classes of
all monomorphisms and of all epimorphisms in the category C, respectively.
Since we have not yet discussed the process of generation, we can not yet
characterize epimorphisms in the category of partial algebras of some given
type, but it is already easy to realize that each surjective homomorphism is
an epimorphism. And we have:
Proposition 2.10.3 In the category Alg() of all partial algebras of simi-
larity type with the homomorphisms as morphisms, the monomorphisms
are exactly the injective homomorphisms. Moreover we have since this
holds in arbitrary categories that the composition of monomorphisms is
a monomorphism, and the composition of epimorphisms is an epimorphism,
each isomorphism is a mono- as well as an epimorphism, but in contrast
to species of total algebras and to the classes of hierarchical equoids
39
, in
general a bimorphism, i.e. a monomorphism which is at the same time an
epimorphism, need not be an isomorphism.
Proof That every injective homomorphism is a monomorphism follows al-
ready from the fact that every injective mapping has the cancellation prop-
erty (for all other mappings) required for monomorphisms. Conversely, let
f : A B be a homomorphism which is not injective, and let a, b A be
dierent elements such that f(a) = f(b) (such elements then exist!). Let
39
Cf. H. Reichel [Re85] or [Re86].
December 8, 2002
102 Chapter I Basic denitions and properties
C := (c; ) be a discrete one-element partial algebra of type , and let
g, h : C A , g(c) := a, h(c) := b. Obviously, g and h are homomor-
phisms from C into A satisfying f g = f h, but g ,= h. Therefore f is not
a monomorphism.
Let A be a total algebra on a non-empty carrier set, and of a non-trivial
similarity type, then id
A
: A A is a bimorphism which is not an isomor-
phism.
Remarks 2.10.4 (i): The above argument can be repeated in every full
subcategory of Alg(), which has at least one free object with at least one
generator (cf. sections 5, 6 and 17).
(ii): Notice that in the book [Pi68] of R.S.Pierce epimorphisms are
dened as full and surjective homomorphisms described in our terminol-
ogy.
2.11 Special observations concerning partial S-algebras
In the previous subsection we have dened (heterogeneous) partial S-algebras
and have shown that this concept is equivalent to the one of heterogeneous
partial algebras. Now we extend the concepts introduced so far for homoge-
neous partial algebras to the case of partial S-algebras (in this subsection we
still use the reminder ):
Denition 2.11.1 Let S be any heterogeneous sort algebra throughout
the rest of this book S will have this meaning, even if we do not stress it
explicitly and let A be any partial S-algebra.
(i): An S-homomorphism f : A B is full or closed, respectively, iff
the ordinary homomorphism f : A B is full, respectively closed.
(ii): A relation AA is an S-congruence of A, iff it is the kernel of
an S-homomorphism starting from A, and it is called a closed S-congruence
of A, iff it is the kernel of a closed S-homomorphism starting from A.
(iii): Let A := (A, v
A
) and B := (B, v
B
) be any S-sets. Then we dene
A B :=

( A
s
B
s
[ s S ), which later on will also be denoted by AB,
since in the case of S-sets this is the correct product (see section 4). Let
us also mention that in what follows we shall understand by A
B
the set
A
B
of all S-mappings f : B A, i.e. satisfying v
A
f = v
B
.
Moreover, one has to set, for each , v
()
() := s

in S ( ()).
December 8, 2002
2 Homomorphisms etc. 103
(iv): By Cong A and Cong
c
A we denote from now on the sets of all
S-congruences respectively closed S-congruences of A. The corresponding
ordered sets are also denoted as in the homogeneous case; only in special
cases we use Cong A or Cong
c
A.
(v): End A and Aut A, the endomorphism monoid and the automor-
phism group of A, respectively, are dened as in the homogeneous case and
will also be denoted as End A, respectively Aut A. By HomAlg(S) we
denote the class of all S-homomorphisms.
Proposition 2.11.2 Let A be any partial S-algebra. Then
(i) Cong A is the set of all congruence relations of A the underlying
homogeneous partial algebra , which are contained in ker v
A
= AA
, i.e. ker v
A
(=:
A
from now on!) is the largest element of the complete
lattice Cong A := (Cong A; ).
(ii) (Cong
c
A; ) is the initial segment of all those closed congruences of A
contained in ker v
A
. If the type is nitary, then there also exists a
largest element
c
(A) in Cong
c
A, namely
c
(A) =
c
(A) ker v
A
,
where
c
(A) is the largest closed congruence relation of A.
Proof Ad (i): Let A and B be any partial S-algebras, f : A B
any full and surjective homomorphism, and := ker f. Since f is an S-
homomorphism, iff v
B
f = v
A
, we have Cong A, iff ker v
A
, as can
be easily derived from the Diagram Completion Lemma 2.7.1.
Ad (ii): This is an immediate consequence of (i) and the denition of closed
S-homomorphisms.
Observations 2.11.3 The above result and the observations in ?? show
that all the preceding results on homomorphisms and congruences carry over
to S-homomorphisms, also as far as full and closed (S-) homomorphisms
and congruences are concerned, as long as one observes the restriction given
by ker v
A
in connection with congruences. For instance we always have to
remember that now
A
means

A
:= ker v
A
= A A =

_
sS
A
s
A
s
.
As an example we reformulate Corollary 2.8.3.(i) as
December 8, 2002
104 Chapter I Basic denitions and properties
Proposition 2.11.4 Let f : A B be a full and surjective S-homomorphism.
Then f
(2)
induces an isomorphism between the interval ([ker f,
A
; ) of
Cong A and all of Cong B ( f
(2)
[]).
Also for closed S-congruences the corresponding result is true. Finally
let us observe:
Proposition and Notation 2.11.5 Let A be a partial S-algebra and
AA any relation. Then there exists a smallest S- congruence, say

, of A
containing , iff
A
. And if

exists, then it is given by

:= Con
A

and will be denoted by Con
A
. Each relation
A
will be called an S-
relation on A.
2.11.6 Let us repeat that in what follows partial algebra A will always
mean partial S-algebra A for some arbitrary but xed heterogeneous sort
algebra S of the (homogeneous) type under consideration, even when S
will not be explicitely mentioned (or only at the beginning of a section),
except when we stress that it is homogeneous (then S is the one-element
total algebra). Moreover, mappings and homomorphisms will always have to
be compatible with the sort mappings.
It should be observed, too, that we allow empty carrier sets of homo-
geneous and empty phyla of heterogeneous partial algebras, therefore the
sort homomorphisms v
A
need not be surjective. In particular, in the het-
erogeneous case we would otherwise get into trouble with the existence of
subalgebras generated by subsets which do not contain elements from each
phylum.
Example 2.11.7 In order to get a better feeling for all these concepts,
let us consider, for the class (category) Cat of all small categories C
(i.e. Ob C and Mor C are always sets) as class of heterogeneous partial alge-
bras the concept of homomorphisms, which here are usually called functors:
We have in this case
40
S := mor, ob, graph
S
:= ((mor, mor), mor),
graph Dom
S
:= ((mor), ob) =: graph Cod
S
, graph 1
S
:= ((ob), mor). For
every category C Cat we then have Ob C := v
1
C
[ob], Mor C := v
C
[mor].
And if C and C

are any two small categories, then a functor f : C C

, i.e.
an S-homomorphism from C into C

, is any mapping f : C C

such that
40
Cf. 1.2.1.(iii).
December 8, 2002
2 Homomorphisms etc. 105
(a) f
ob
[Ob C] Ob C

and f
mor
[Mor C Mor C

;
(b) f
mor
(m

C
m) = f
mor
(m

)
C

f
mod
(m) for all composable m, m

Mor C;
(c) f
ob
(Dom
C
(m)) = Dom
C

(f
mor
(m)) and f
ob
(Cod
C
(m)) = Cod
C

(f
mor
(m))
(m Mor C);
(d) f
mor
(1
C
A
) = 1
C

f
ob
(A)
(A Ob C).
Except for (a) these are the usual compatibility conditions for homomor-
phisms known from the homogeneous case.
December 8, 2002
106 Chapter I Basic denitions and properties
3 Substructures, subalgebras, closed subsets,
generation
One of the most important features of algebraic structures is the fact that
by means of the process of generation usually very large subsets of a given
structure can be described by specifying a comparatively small subset as a
generating subset, whenever the large subset under consideration can be
described as a so-called closed subset. This fact may also serve as the most
important motivation for dealing with partial algebras rather than describing
them as relational systems, where generation is not so easily dealt with as in
an algebraic context (cf. also [Re84] or [Re87] for further motivation and ex-
amples in this direction from computer science, where this is one of the most
important reasons for the application of universal algebra at all). Of special
importance in connection with the concept of generation is the principle of
algebraic induction which is called in Computer Science structural induction.
It should also be observed that most data types considered in Computer
Science are in reality represented by innite (total or partial) algebras,
while every computer can represent only a nite part of them, which will
then be algebraically spoken a relative subalgebra or even only a weak
relative subalgebra of the original one, motivating also the introduction of
these concepts in this section (while this fact also motivates the consideration
of partial algebras in general).
As already mentioned at the end of the previous section, we tacitly assume
in what follows some xed type and a heterogeneous sort algebra S of
type (cf. ??). Moreover, all partial algebras are of type , and each partial
algebra A is provided with a xed sort homomorphism v
A
: A S such that
all homomorphisms between such partial S-algebras are compatible with the
corresponding sort homomorphisms. If we do not need it for some special
reason we usually omit the reminder referring to the existence of this
sort homomorphism.
3.1 Substructures of partial algebras
3.1.1 (Category theoretical concept of subobjects) In Category The-
ory a subobject of an object A in a category C is in most generality an ordered
pair (B, f), where f : B A is any monomorphism in C very often addi-
tional properties are requested.
December 8, 2002
3 Substructures, closed subsets, generation 107
For our category Alg(S) of all partial S-algebras of some similarity type
together with their homomorphisms the simplest concept would mean that
a subobject of a partial algebra A is a pair (B, i), where i : B A is an
injective homomorphism (cf. 2.10.3).
If (B, i) is any subobject of A in this sense, then we may consider A

:=
i[B] and graph
A

:= i
(()+1)
graph
B
; therefore i

: B A

(i = id
A
i

)
becomes an isomorphism, and (A

, id
A
) is another subobject, which we shall
call the weak (relative) subalgebra of A induced by (B, i).
Two subobjects (B, i) and (C, j) of A are called equivalent, whenever
there is an isomorphism k : B C such that j k = i. Thus (B, i) and
(A

, id
A
) above are equivalent, and (A

, id
A
) is kind of a natural represen-
tative of the equivalence class of (B, i). These observations motivate the
denition below.
But it should also be observed that this concept of subobject of par-
tial algebras is too weak to describe generation. Therefore we strengthen
the concept of subobject below by requiring additional properties for the
monomorphisms under consideration, and the most appropriate one will be
closedness. We shall almost always consider only the natural representa-
tive introduced above.
Denition 3.1.2 Let A and B be any similar partial algebras such that
B A. Then B is said to be a
weak (relative) subalgebra of A iff id
B
: B A is an injective homo-
morphism
relative subalgebra of A iff id
B
: B A is a full and injective homo-
morphism
(closed) subalgebra of A iff id
B
: B A is a closed and injective
homomorphism
C A is said to be a closed subset of A, iff for every fundamental operation

A
and for every sequence a dom
A
C
()
one has
A
a C.
We denote by
Csubs A the set of all closed subsets of A
Wsuba A the set of all weak subalgebras of A
Rsuba A the set of all relative subalgebras of A
December 8, 2002
108 Chapter I Basic denitions and properties
Suba A the set of all (closed) subalgebras of A.
Proposition 3.1.3 Let A, B be similar partial algebras, B A. Then:
(i) B is a weak subalgebra of A iff for all :
graph
B
graph
A
(ii) B is a relative subalgebra of A iff for all :
graph
B
= graph
A
(B
()
B)
(iii) B is a subalgebra of A iff for all :
graph
B
= graph
A
(B
()
A)
i.e. iff B is a relative subalgebra of A, and B is a closed subset of A.
Proof (i) is obvious, (ii) follows from 3.1.2 and 2.4.4.(i), while (iii) can
easily be derived from 3.1.2, 2.4.2.(ii) and 2.4.4.(ii).
Remarks 3.1.4 (i) It should be observed that each subalgebra is a rela-
tive subalgebra, and that each relative subalgebra is especially a weak
subalgebra. Moreover, A is always a subalgebra of itself; and if A has no
(non-empty) nullary constants, then the empty subset is always closed,
too.
(ii) The importance of weak and relative subalgebras lies in their role as
small building blocks, as we shall see e.g. in section 4 in connection
with direct limits and also in later sections; and we have already men-
tioned that they may be used to approximate an innite algebra which
cannot be completely represented, say, in a computer (e.g. the sets of
all integers, respectively of all rationals or reals are by far too large, and
only special relative subalgebras of the corresponding algebras can be
represented, and it depends on the hardware as well as on the software,
how large this part is for a given computer and compiler or interpreter
of a programming language).
In connection with the concept of generation they are of less impor-
tance, since there is a (weak) relative subalgebra of A on every subset
of A.
December 8, 2002
3 Substructures, closed subsets, generation 109
The situation is dierent for subalgebras, which only exist on closed
subsets. Thus we will study in this section mainly the properties of
subalgebras and closed subsets.
(iii) It should be noticed that as with homomorphisms the concept
of a closed subset and hence of a subalgebra heavily depends on the
similarity type chosen. E.g. if one denes groups as is usually done
in group theory by only one binary operation, then the subalgebras
in the sense of 3.1.2 are all subsemigroups of the given group, and
one has to add at least the formation of the inverse as a fundamental
operation in order to get all subgroups including the empty subalgebra
as subalgebras. Only when we consider groups as universal algebras
of type (2,1,0) (or in some equivalent way), we get coincidence for
groups between our concept of subalgebras and the usual concept of
subgroups. One reason for this conceptual dierence is the fact that
the universal algebraic concept of subalgebras uses as little of model
theory as possible (i.e. tries to keep the axioms as simple as possible),
while usually in group theory the axioms are logically relatively complex
(containing existential quantiers explicitly).
The concept of a subalgebra, as usually used in universal algebra, fully
describes substructures for classes which are described by equations
but not by more complex axioms, and it is well known that one can
describe groups by equations, when one chooses (2,1,0) as similarity
type.
(iv) When considering any weak subalgebra B of some partial S-algebra
A (or just when writing B A) the sort mapping v
B
is always the
restriction of v
A
to B, and thus it is a homomorphism from B into S.
Thus it is usually not necessary to refer to S in connection with sub-
structures of partial S-algebras, anyway.
Proposition 3.1.5 Let A be any partial algebra of similarity type , and let

:=

s
[ s S

c
a
[ a A and

:= ( 2 [ s S )( 0 [ a
A) ( indicating concatenation of sequences), where each constant c
a
has
output sort v
A
(a).
Then Cong A is exactly the set of all closed subsets of the following partial
algebra B of similarity type

, where:
B := A A (i.e. A

(!)), c
a
B
:= (a, a) for every a A,
B
s
is a binary
December 8, 2002
110 Chapter I Basic denitions and properties
total heterogeneous operation dened on ((a, b), (c, d)) [ a, b, c, d A
s
=
v
1
A
(s) ,

B
s
((a, b), (c, d)) :=
_
(a, d), if b = c,
(b, a), if b ,= c;
and for all one has
dom
B
:= ((a
k
, b
k
) [ k ()) [ ( a
k
[ k () ), ( b
k
[ k () ) dom
A

and if ( (a
k
, b
k
) [ k () ) dom
B
then

B
( (a
k
, b
k
) [ k () ) := (
A
( a
k
[ k () ), (
A
( b
k
[ k () )).
v
B
is the mapping induced by v
A
.
Proof The denition of B guarantees that each closed subset of B is a
binary relation on A, the constants c
a
B
yield reexivity, while the
B
s
take
care of symmetry and transitivity of the closed subsets of B on each phylum
of A. Finally the denition of the partial operations
B
enforces compatibility
of the equivalence relations with the fundamental operations of A.
3.1.6 As a consequence of Proposition 3.1.5 everything which we shall say
about generation of closed subsets or subalgebras of some partial algebra
A will also apply to the generation of congruence relations in connection
with the algebraic structure of B dened above on A A. Notice, how-
ever, that for the structure of B above the cardinal dimension is always
d

= maxd,
0
, where d is the cardinal dimension of , and its ordinal di-
mension is max,
0
, being the ordinal dimension of (cf. 1.3.1). More-
over, 3.1.5 yields examples of closed subsets, and as we have already observed
in 3.1.4.(iii) our concept of subalgebra yields the correct subobjects of total
algebras, whenever one considers a class of algebras dened by equations
(identities). This can e.g. be done for semigroups, rings, groups, lattices,
boolean algebras, vector spaces, hypercomplex systems, etc..
Proposition 3.1.7 For every closed homomorphism into a total algebra the
domain is a total algebra, too.
Corollary 3.1.8 Every subalgebra of a total algebra is again total.
December 8, 2002
3 Substructures, closed subsets, generation 111
gure uncompleted
Figure 3.1: f3.1
Proposition 3.1.9 Let f : A B be any homomorphism, and let C and D
be subalgebras of A and B, respectively. Then:
(i) f
1
[D] is a closed subset of A, too: f
1
[D] Csubs A.
(ii) If f is closed, then f[C] is a closed subset of B, but this need not be
the case for arbitrary homomorphisms, even f[A] need not be a closed
subset of B in general.
Proof (i) With the conditions of the Proposition assume a (f
1
[D])
()

dom
A
; then f(a) D
()
dom
B
, since since f is a homomorphism,
and
B
(f(a)) = f(
A
(a)) D, since D is closed. Thus
A
(a) f
1
[D],
showing that f
1
[D] is closed in A.
(ii) Let f be closed, and b (f[C])
()
dom
B
. Then, since f as well as C
are closed, there exists a dom
C
such that f(a) = b and
A
(a) C.
Thus f(
A
(a)) =
B
(f(a)) =
B
(b) f[C]. Hence f[C] is closed in
B. In order to show that arbitrary homomorphisms do not preserve
closed subsets, we present Figure 3.1 with partial algebras of type (1).
Corollary 3.1.10 Let f : A B be any homomorphism and C a total
subalgebra of A. Then f[C] is the carrier set of a total subalgebra of B.
Proof Examine the proof of 3.1.9.(ii) for the case when C is total.
Proposition 3.1.11 Let B be a closed subset of A, and C any subset of A
(i) Then C Csubs B iff C B and C Csubs A, in particular the con-
cepts of closed subset and subalgebra are transitive and hereditary.
(ii) If f : A D is a closed homomorphism, then f[
B
: B D is closed,
too, where B is the subalgebra of A with carrier set B.
(iii) If Cong
c
A, then
B
:= (B B) Cong
c
B.
Proof All the three statements easily follow from the denitions of the
related concepts.
December 8, 2002
112 Chapter I Basic denitions and properties
3.2 Comparison of substructures
It is sometimes necessary to compare partial algebraic structures having the
same carrier set. For that purpose we dene
Denition 3.2.1 Let A be any set, O

:= (

and O

:= (

partial
algebraic (S-) structures of type on A. We say that O

(or (A, O

)) is
weaker or poorer than O

(or (A, O

)), and that O

is richer than O

, iff
id
A
: (A, O

) (A, O

) is an (S-) homomorphism.
Remarks 3.2.2 (i) Let A Alg(S) be any partial algebra, B A any
subset. Then the relative subalgebra B of A with carrier set B carries
the richest (S-) structure O

on B such that id
B
: (B, O

) A is a
homomorphism; and it has the property that for every homomorphism
f : C A, which satises f[C] B, the mapping f : C B is also a
homomorphism f : C B (cf. 2.7.3.(ii)).
Moreover, recall, that a mapping consists of a domain, graph and
codomain, and that in f : C B f usually represents the graph,
while at other occasions f stands for the whole triple).
The relative substructure of A on B is the richest weak relative sub-
structure of A on B, the discrete substructure on B is the weakest or
poorest one.
(ii) Let O
i
:= (
i
)

(i I) be partial algebraic structures on a set A.


Then
O

:=

O
i
[ i I := (


i
[ i I [ )
(see 23.1.2) is a partial algebraic structure on A dened by
graph

:=

graph
i
[ i I , ( ).
O

is the richest structure on A such that for each i I, id


A
: (A, O

)
(A, O
i
) is a homomorphism; it is called the intersection or inmum of
the given family of partial algebraic structures on A.
If, for each , graph

:=

graph
i
[ i I is still the graph of
a partial operation, then O

:= (

( O
i
[ i I ) (see 23.1.2)
is the poorest partial algebraic structure on A such that, for each i
I, id
A
: (A, O
i
) (A, O

) is a homomorphism. It is called whenever it


December 8, 2002
3 Substructures, closed subsets, generation 113
exists the supremum of the given family of partial algebraic structures
on A. Therefore the partial algebraic structures on any set A form a
conditionally complete lower semilattice (see 24.1).
(iii) Let B
i
:= (B
i
, (
B
i
)

)(i I) be any family of (weak) relative subal-


gebras of some partial algebra A. Then
B :=

( B
i
[ i I ) := ((

B
i
[ i I , (

)
)
, v
B
)
with
graph

:=

graph
B
i
[ i I
is the largest and richest weak relative subalgebra C of A such that
for each i I id
C
: C B
i
is a homomorphism. We call B the
intersection of the (weak) relative subalgebras B
i
(i I). If all B
i
are
relative subalgebras of A, then so is B. The sort homomorphism of B
is the restriction to B of the given sort homomorphism for A.
Because of its importance we formulate separately
Proposition 3.2.3 Let A be any partial algebra, and let ( B
i
[ i I )
(Suba A)
I
be any family of subalgebras of A. Then
B :=

B
i
[ i I Suba A.
In particular the intersection of an arbitrary family of closed subsets of A is
again a closed subset of A (and if I = , the empty set, then B := A).
Hence Csubs A is a closure system on A which does not depend on v
A
,
and we denote by c
A
the corresponding closure operator (cf. the Appendix,
section 24).
Proof Let b B
()
dom
A
. Then b B
()
i
dom
B
i
for each i I,
and
A
(b) B
i
[ i I = B, i.e. B is closed.
December 8, 2002
114 Chapter I Basic denitions and properties
3.3 The global process of generation
3.3.1 The statement of Proposition 3.2.3 and the results listed in the Ap-
pendix 24 now tell us that for every subset M A of any partial algebra A its
closure with respect to c
A
is given by c
A
M =

B [ B Csubs A, M B.
We call c
A
M the closed subset of A generated by M, and C
A
M designates
the subalgebra of A generated by M.
If c
A
M = A, then M is called a generating subset or a set of generators
of the partial (S-)algebra A, and A is said to be generated by M.
It should be observed that this general global process of generation does
not depend on the fact that we might consider a partial S-algebra; only
the partial algebraic structure of A is responsible for the size and structure
of C
A
M.
Proposition 3.3.2 Let A be any partial algebra of similarity type with
cardinal dimension d, and let c := max2, d. Then
(i) Csubs A is a c-inductive closure system on A.
(ii) (Csubs A, ) is a c-compactly generated (and hence a c-algebraic) lat-
tice. Its c-compact elements are exactly those closed subsets of A which
have a generating subset of cardinality less than c.
In particular, one has for every subset M A
() c
A
M =
_
c
A
N [ N M, #N < c .
Proof (ii) is an immediate consequence of (i) and the general results on
closure systems indicated in the Appendix, section 24. In order to prove
(i) let B Csubs A be a c-directed system of closed subsets of A, and let
B :=

B. We have to show that B is closed in A:


Therefore let and b = ( b
k
[ k () ) B
()
dom
A
. Since
B =

B there exists for each k () a set B


k
B such that b
k
B
k
(notice that B ,= , since it is c-directed). c-directedness of B now implies
the existence of a set B

B such that B
k
B

for each k () (observe


() < c); since B

is closed,
A
(b) B

B, showing that B is closed.


December 8, 2002
3 Substructures, closed subsets, generation 115
3.4 Local generation (generation from below)
3.4.1 So far generation of a closed subset has been a global process, i.e.
we had to know the whole closure system of all closed subsets (or because
of c-inductivity c := max2,
0
at least the c-compact closed sets) in
order to determine whether a given subset of a partial algebra A is closed or
not, and to compute its closure in the latter case (cf. 3.3.2.()).
In the case of partial algebras the decision about closedness or the con-
struction of the closure can be accomplished by using the fundamental par-
tial operations. And this can be done by an iterative local process without
any knowledge about the closure system at all. This is achieved in analogy
to the fact that the linear subspace generated by some set M in a vector
space V over some eld K is the set of all nite K-linear combinations of
elements of M. A similar process of generation from below is well known
for semigroups, groups and other algebraic structures. One generalization
of this procedure is described in the following denition. The constructions
under consideration only use the partial algebraic structure and nothing in
connection with a possibly existing sort homomorphism.
Denition 3.4.2 Let A be any partial algebra of some similarity type =
(())

, and let M A be any subset. We dene operators D


A
and B
A
:
(i)
DM := D
A
M := M
_

A
[M
()
] [ ,
and
B M := B
A
M := D
A
M M.
(ii) Moreover we dene recursively for any ordinal and any limit ordinal
,= 0
D
0
M := D
0
A
M := M
D
+1
M := D
+1
A
M := D
A
D

A
M
D

M := D

A
M := D
A

A
M [
B
0
M := B
0
A
M := M
B
+1
M := B
+1
A
M := D
+1
A
M D

A
M
B

M := B

A
M := D

A
M

A
M [
= D

A
M

A
M [
December 8, 2002
116 Chapter I Basic denitions and properties
In analogy to topology B
A
is called the Baire- or Borel-operator on A, and
B

A
M is called the -th Baire- or Borel-class of M. D
A
is called the step-
operator for (or: on) A. It is obvious that, for ,= , B

A
M B

A
M = ;
therefore we say for a B

A
M that a has step-number or rank with respect
to M, in symbols:
A
(a, M) := .
In contrast to the one-step closure operator c
A
( i.e. c
A
c
A
M = c
A
M),
the operator D
A
is called a many-step closure operator on A; the reason for
this will become obvious from the following results.
Lemma 3.4.3 Let A be any partial algebra, M A. Then c
A
M = M iff D
A
M =
M; i.e. M is closed with respect to the operator c
A
iff M is closed with respect
to the operator D
A
.
This observation shows that D
A
might be the correct candidate for local
generation.
Theorem 3.4.4 For every partial algebra A of similarity type , which has
the ordinal dimension (cf. 1.3.1), and for every set M A one has:
(i) c
A
M = D

A
M
(ii) If > 1 then we already have c
A
M =

A
M [ < .
3.4.5 We shall see below that Theorem 3.4.4 yields the optimal general re-
sult, but it should also be noticed that in concrete cases the iteration process
may stop much earlier; e.g. if V is a linear space of nite dimension m ,= 0
over the eld of real numbers, and if M is a basis of V , then one has for
:= log
2
m| + 1 that V = D

V
M (r| is here the least integer greater or
equal to r); and this bound is sharp; therefore we cannot get below 3.4.4.(ii),
if M is innite.
Proof (of Theorem 3.4.4) If = 0, and = ; thus c
A
M = M = D
0
A
M.
If = 1, then c
A
M = Mc
A
= D
1
A
M, and obviously c
A
M ,= M = D
0
A
M
, if c
A
, M.
Now assume > 1. Because of Lemma 3.4.3 it suces to show that D
A

A
M [
< =

A
M [ < (notice that just by the denition of D
A
one has N D
A
N c
A
N for every subset N of A, and inductively:
D

A
N c
A
N for every ordinal and for every subset N of A). There-
fore assume b := ( b
k
[ k () ) (

A
M [ < )
()
dom
A
for
December 8, 2002
3 Substructures, closed subsets, generation 117
,
,
,
,
,
,

a
1
a
2
a
r
a
()1
a
()
a =
A
(a
1
, . . . , a
r
, . . . , a
()
)

>
>
>

_
_

B
1
A
M M = B
0
A
M
B
2
A
M
B
k
A
M
B
k+1
A
M
A
Figure 3.2: Layer model of generation from below (nitary case)
some . Since is regular and () # there exists an < such
that b
k
[ k ()

A
M [ < , and therefore
A
(b) D

A
M

A
M [ < . This argument shows that c
A
M

A
M [ < .
3.4.6 We would like to observe that obviously for each subset M of a partial
algebra A one has c
A
M =

A
M [ < , and that this is a disjoint
union. This is the justication of the rank or step-number dened in 3.4.2.
We illustrate the situation concerning the Baire-classes in Figure 3.2. In
particular one has: If an element a of c
A
M has step-number , then there
are and a dom
A
such that each component of a belongs to c
A
M,
has a step-number less than , and the supremum of all step-numbers of
components of a is either 1, if is a successor ordinal, or it is , if is
a limit ordinal. Moreover it should be noticed that B

A
M = iff

A
M [
< is a closed subset of A.
December 8, 2002
118 Chapter I Basic denitions and properties
3.5 The principle of algebraic induction (i.e. of struc-
tural induction)
3.5.1 (Algebraic or Structural Induction) The process of generation
from below along Baire-classes as described in the previous section may sug-
gest to prove theorems for partial algebras by transnite induction on step
numbers, when a generating subset is given. In fact this is often done. But
in most such cases it is more elegant to apply the method of algebraic in-
duction, which is mostly called structural induction in Computer Science (in
mathematical logic and in connection with absolutely free structures this
method usually appears in the form of induction on the structure of terms
or formulas):
Let A be any partial algebra of some similarity type = (())

, and
let M, N A be any subsets (where N is usually dened to be the set of
all elements of A having some given property usually described by a rst
order formula (cf. Chapter II)):
In order to prove c
A
M N (or that a given property holds for all
elements of c
A
M) it is sucient to prove:
(i) x N for every x M (i.e. M N) basis of the induction.
(ii) If for some and some b dom
A
(c
A
M)
()
one has
b N
()
(induction hypothesis), then
A
(b) N (induction step).
Notice that one version of (ii) says that N is itself a closed subset of A
containing M (cf. (i)), therefore it has to contain c
A
M, since c
A
is a closure
operator.
If = (1), A = (N; Succ), M = 0, where Succ(n) := n + 1 is the
successor function for natural numbers, then the above principle of algebraic
induction is nothing else than the principle of complete induction for the set
of natural numbers, which thus turns out to be a special case of algebraic
induction.
Applications of the principle of algebraic induction are as in the ex-
amples below sometimes not so evident; better examples can be found in
section 5.
Proposition 3.5.2 For each subset M of a partial algebra A one has
() c
A
M = M
_

A
[(c
A
M)
()
] [
December 8, 2002
3 Substructures, closed subsets, generation 119
i.e. each element of the closure of a set M is either an element of M itself
or it is the value under an application of a fundamental partial operation to
an argument sequence within its closure.
Proof It is obvious that the right-hand side of () is contained in the left-
hand side, and since it contains M and is easily seen to be closed, the other
inclusion also holds.
Proposition 3.5.3 Let f, g : A B be two homomorphisms between simi-
lar partial algebras; then the equalizer a A [ f(a) = g(a) is a closed
subset of A.
Corollary 3.5.4 (1) Let f, g : A B be two homomorphisms between sim-
ilar partial algebras which coincide on some generating subset M of A : f[
M
=
g[
M
; then f and g are identical: f = g.
Corollary 3.5.5 (2) Let A and B be similar partial algebras, M a gener-
ating subset of A, f
0
: M B any mapping. Then there is at most one
homomorphism f : A B extending f
0
, i.e. satisfying f[
M
= f
0
. In par-
ticular, every homomorphism is uniquely determined by its values on (i.e. by
its restriction to) a generating subset of its domain.
Corollary 3.5.6 (3) Let f : A A be any endomorphism; then the set
x A [ f(x) = x of all xed points of f is a closed subset of A.
3.6 Some further results concerning generation, epi-
morphisms
We are now able to characterize epimorphisms (cf. 2.10.2):
Proposition 3.6.1 Let f : A B be any homomorphism; then the follow-
ing statements are equivalent:
(i) f is an epimorphism.
(ii) c
B
f[A] = B, i.e. the image of f is dense in B.
Proof (ii) (i) Let g, h : B C such that g f = h f. Then g and h
coincide on the generating set f[A] of B, and therefore they are equal
according to 3.5.4.
December 8, 2002
120 Chapter I Basic denitions and properties
(i) (ii) Assume that f[A] does not generate B; then we have to show
that f is not an epimorphism. We achieve this by constructing a partial
algebra C and two dierent homomorphisms g and h from B into C
such that g f = h f:
Let D := c
B
f[A], E := B D and E

a copy of E disjoint from B,


and i : E

E a bijection, which we extend by the identity on D to a


bijection i

: D

E. Then we dene: C := B E

;
v
C
(c) :=
_
v
B
(c), if c B,
(v
B
i)(c), if c E

.
For each we set
dom
C
:= c C
()
[ c dom
B
or i

(c) dom
B
,
and if c dom
C
, then either
C
(c) :=
B
(c) or
C
(c) := i

1
(
B
(i

(c))).
Hence h := i

1
becomes a closed and injective homomorphism from B
into C; and setting g := id
B
, we get indeed g f = h f, but g ,= h.
The following observation is closely connected to Proposition 3.1.9.
Proposition 3.6.2 Let f : A B be any homomorphism between similar
partial algebras, and let M A be any subset. Then:
(i) f[c
A
M] c
B
f[M], and in general equality is false.
(ii) If f is a closed homomorphism (i.e. in particular, if A is total), then
f[c
A
M] = c
B
f[M].
(iii) If f is an epimorphism, and g : B C is any homomorphism, then
g[B] c
C
(g f[A]), and if g is closed, then equality holds.
Proof Ad (i) By Proposition 3.1.9.(i) f
1
(c
B
f[M]) is a closed subset of A
which obviously contains M and therefore also c
A
M. Applying f again
yields (i).
That usually equality does not hold in (i) is easily seen from the example
in the proof of Proposition 3.1.9.
Ad (ii) If f is closed, then 3.1.9.(ii) implies that f[c
A
M] is a closed subset
of B which contains f[M] and therefore c
B
f[M]. The statement now
follows from part (i).
Ad (iii) 3.6.1 and (i) imply g[B] = g[c
B
f[A]] c
C
g[f[A]] = c
C
(g f[A]),
and if g is closed, then (ii) implies equality.
December 8, 2002
3 Substructures, closed subsets, generation 121
,
,
,
,
,
,

dPfeil?

` `

>
>
>
>
> >.
>
>
>
>
> >.


rPfeil?

:
:
0 0
1 1
2 2
Figure 3.3: A non-closed homomorphism preserving closed subsets
Remark 3.6.3 Closed homomorphisms have received their name mainly be-
cause they preserve closed subsets. But it should be noticed that contrary to
closed continuous mappings in topology there are non-closed homomorphisms
between partial algebras which also preserve closed subsets; thus there is no
full analogy: Consider the type (1,1) for operation symbols and

, and let
A = B = 0, 1, 2, graph
A
:= graph
B
:= ((0), 1), ((1), 2), ((2), 0), graph

A
:=
and graph

B
:= ((1), 0) as shown in Figure 3.3 (f is the identity
on 0, 1, 2).
But we do not know any syntactical description of homomorphisms pre-
serving closed subsets, and what we call according to J urgen Schmidt
closed homomorphisms seems to be the smoothest approximation. The
following concept of, say, semiclosed homomorphisms would also do in this
connection, as can easily be seen:
For every , a A
()
satisfying f(a) dom
B
there exists a


(c
A
(a[()]))
()
dom
A
such that f(a) = f(a

).
It should also be observed that especially the case of partial S-algebras
shows us that closed homomorphisms are exactly those homomorphisms
modeling totalness, and in particular all homomorphisms between total
S-algebras are closed.
Lemma 3.6.4 Let A and B be possibly dierent similar partial algebras on
the same carrier set A, and let graph
A
graph
B
for all , i.e. A is
a weak subalgebra of B. Then:
(i) Csubs B Csubs A.
(ii) For each M A one has c
A
M c
B
M.
Proof Each subset of A closed with respect to the structure of B is obvi-
ously also closed with respect to the structure of A; this yields (i). Now, if
December 8, 2002
122 Chapter I Basic denitions and properties
M A , then c
A
M =

C Csubs A [ M C C Csubs B [ M
C = c
B
M.
Corollary 3.6.5 (1) If A is any weak subalgebra of B, then one has for any
M A that c
A
M c
B
M.
Proof For C := (B, (
A
)

) the assumptions of 3.6.4 are satised for C


and B, and moreover one has c
A
M = c
C
M for each M A, hence the
statement of 3.6.5.
Corollary 3.6.6 (2) Let A be a weak subalgebra of B and B a weak sub-
algebra of C, and assume c
B
A = B, c
C
B = C. Then A generates C, i.e.
c
C
A = C.
Proof c
C
A = c
C
c
C
A c
C
c
B
A = c
C
B = C.
Corollary 3.6.7 (3) Let f : A B be any epimorphism and M a generat-
ing subset of A, then f[M] is a generating subset of B.
3.7 The analogue of E.Noethers First Isomomorphism
Theorem
Denition 3.7.1 Let A be any partial algebra, Cong A any congruence
relation on A, and B any subset of A. Then the -saturation [B]

of B is
the subset [B]

:= a A [ (a, b) for some b B. Observe that the


corresponding sort mapping v
[B]

: [B]

S is the restriction to [B]

of the
sort homomorphism v
A
: A S. By [B]

we denote the relative subalgebra


of A with carrier set [B]

. Moreover, we let
B
designate the restriction of
congruence to the relative subalgebra B of A, i.e.
B
:= (B B).
Obviously
B
Cong B.
Proposition 3.7.2 Let B be a closed subset of a partial algebra Aand a
closed congruence of A. Then [B]

is also a closed subset of A.


Proof Let nat

: A A/ be the natural projection. Then obviously


[B]

= nat

1
[nat

[B]]. Since nat

is closed (cf. 2.6.3), nat

[B] is closed in
A/ (cf. 3.1.9.(ii)); therefore [B]

is closed in A according to 3.1.9.(i).


December 8, 2002
3 Substructures, closed subsets, generation 123
gures disabled
Figure 3.4: An analogue of the First Isomorphism Theorem
Theorem 3.7.3 Let A be any partial algebra, B a relative subalgebra of A,
and Cong A a congruence on A. Then there exists a bijective homomor-
phism
f : B/
B
B

/
[B]

dened by f : [b]

B
[b]

for every b B
(cf. Figure 3.4 and notice that [b]

designates the congruence class of b with


respect to ).
If B is a closed subset of A, and if is a closed congruence on A, then
f is an isomorphism.
Proof Consider the diagram in Figure 3.4. There exists an induced ho-
momorphism f : B/
B
[B]

/
[B]

according to the Diagram Comple-


tion Lemma 2.7.1. f is injective, since ker nat

B
= ker (nat

[B]

id
B
). As
nat

[B]

id
B
is also surjective, f is even bijective.
Therefore there only remains to show that f is full, if B and are closed
(cf. 2.4.6.(iii)):
For brevity set B

:= [B]

, and

:=
[B]

=
B
, and assume ( [b
k
]

[ k
() ) dom
B

, where each b
k
B.
Then there exists a sequence a := ( a
k
[ k () ) dom
B

such that
nat

a = ( [b
k
]

[ k () ). By 3.7.2 B

is closed in A, and by 3.1.11.(iii)

is a closed congruence on B

. Thus b := ( b
k
[ k () ) dom
B

. Since
id
B
: B B

is closed, we get b dom


B
. Hence ( [b
k
]

B
[ k () )
dom
B/
B
, showing that f is full (closed).
Corollary 3.7.4 Let B be any subalgebra and a closed congruence of a
partial algebra A such that [B]

= A. Then A/ is isomorphic to B/
B
:
A/

= B/
B
.
Remark 3.7.5 We leave it as an exercise to show that both assumptions,
namely the closedness of B as well as the closedness of , are necessary
assumptions in 3.7.3 in order to get an isomorphism; i.e. if we omit only one
of them, then we get counterexamples. The partial algebra in Figure 3.5 may
be a good candidate.
December 8, 2002
124 Chapter I Basic denitions and properties
,
,
,
`
_

a
b
c
A
Suggestions:
B := a, b and := Con
A
(a, c) or
B := a, c and := Con
A
(a, b).
Figure 3.5: Necessity of the assumptions of closedness
December 8, 2002
4 Products and Limits 125
gures disabled
Figure 4.1: The induced product morphism
4 Direct and reduced products, coproducts,
direct limits and related constructions
In this section we discuss some basic constructions of new partial algebras
from families of given ones. So far we started with one partial algebra and
ended with a smaller one, whereas we shall now get large structures from
a family of possibly relatively small ones. Products, coproducts and direct
limits also play an important role in category theory, while reduced products
will be needed in connection with model theoretic investigations.
We recall that we shall use the conventions introduced at the end of sec-
tion 2 concerning partial S-algebras encoding heterogeneous partial algebras.
4.1 (Direct) Products (homogeneous case)
Denition 4.1.1 Let C be any category and ( A
i
[ i I ) Ob C)
I
for
some set I be any family of C-objects.
A product of the family ( A
i
[ i I ) is a pair (A, ( p
i
[ i I )) such
that A Ob C, and for each i I p
i
is a C-morphism p
i
: A A
i
, and
whenever there is any C-object B Ob C with a source (g
i
: B A
i
)
iI
i.e. a family of C-morphisms starting from the same object then there
exists a unique C-morphism g : B A the product morphism induced
by the family ( g
i
[ i I ) (see 4.1.4 below) satisfying g
i
= p
i
g for each
i I, i.e. making each diagram in Figure 4.1 commutative. If I = , then
the above denition means that for every B Ob C there exists exactly one
C-morphism from B into A i.e. A is a so-called right zero or terminal object
of the category C under consideration.
An immediate consequence of category theory is the following
Proposition 4.1.2 If in a category C a product (A, ( p
i
[ i I )) of some
family ( A
i
[ i I ) in C exists, then it is unique up to unique isomorphism.
Proof Let (B, ( q
i
[ i I )) be another product of the given family, then we
have two unique product morphisms: p : A B , induced by ( p
i
[ i I ),
December 8, 2002
126 Chapter I Basic denitions and properties
and q : B A induced by ( q
i
[ i I ), which satisfy for each i I
p
i
q p = q
i
p = p
i
= p
i
1
A
and q
i
p q = p
i
q = q
i
= q
i
1
B
, and the
uniqueness of the induced C- morphisms implies q p = 1
A
and p q = 1
B
,
showing that p an q are isomorphisms.
In the situation described in the above proof we shall say that the product
objects A and B are canonically isomorphic, since the isomorphisms p and
q (inverse to each other) are uniquely determined by the context (here by
the fact that both (A, ( p
i
[ i I )) and (B, ( q
i
[ i I )) are products of the
same family ( A
i
[ i I ) of C-objects).
In what follows we at rst restrict our considerations to the homogeneous
case, and only later on we shall also get a description of products in the
heterogeneous case (of partial S-algebras).
4.1.3 (Construction of direct products of families of homogenous
partial algebras) Let ( A
i
[ i I ) (Alg())
I
be any family of similar
homogeneous partial algebras, where I is any set.
(a) Let A := a : I

A
i
[ i I [ a(i) A
i
for each i I be the set
of all choice functions from I into the family ( A
i
[ i I ) of all carrier
sets, i.e. the so called cartesian (or direct) product of all carrier sets:
A =:

( A
i
[ i I ).
(b) For each and for every sequence a = ( a
k
[ k () ) A
()
dene a dom
A
iff a(i) := (a
k
(i) [ k ()) dom
A
i
for every
i I, and if a dom
A
, then set
A
(a) := (
A
i
(a(i)) [ i I) i.e. the
partial operations
A
are dened componentwise.
(c) For each i I the mapping pr
i
:= pr
A
i
: A A
i
, a a(i)(a A) is
called the i-th projection.
If I = is the empty set, then we assume A to be the one-element total
algebra.
If I contains exactly one element, say, i, then we (can) identify A in a
natural way with A
i
(see the following Theorem).
A, or more concisely (A, ( pr
i
[ i I )), is called the direct product of the
family ( A
i
[ i I ) of partial algebras, and we introduce the following
notation:
A =:

( A
i
[ i I ) (sometimes also A =: ( A
i
[ i I )),
December 8, 2002
4 Products and Limits 127
and if I is nite and non-empty, say I = 1, 2, . . . , n, then we write
A =: A
1
A
2
. . . A
n
, and the elements of A are identied with the
ordered sequences (a
1
, a
2
, . . . , a
n
), where a
i
A
i
for each i I.
If A
i
= B for each i I , then we write A =: B
I
and in the nite case,
where I = 1, . . . , n, we write A =: B
n
.
(d) If ( f
i
: A
i
B
i
[ i I ) is any family of homomorphisms, then
the unique homomorphism (see below) f : A :=

( A
i
[ i I )

( B
i
[ i I ) induced by the family ( f
i
pr
A
i
[ i I ) is denoted by

( f
i
[ i I ).
Theorem 4.1.4 Let ( A
i
[ i I ) (Alg())
I
be any family of similar partial
algebras and (A, ( pr
i
[ i I )) its direct product.
(i) For each i I the i-th projection is a homomorphism pr
i
: A A
i
, and
(
A
)

is in fact the richest structure (initial structure in the sense


of Bourbaki) on A such that all projections are homomorphisms.
(ii) Let ( g
i
: B A
i
[ i I ) be a source of homomorphisms. Then there is
a unique homomorphism g : B A dened by g(b) := (g
i
(b) [ i I)
for each b B such that for each i I one has g
i
= pr
i
g, i.e. for
each i I the diagram in Figure 4.1 with pr
i
instead of p
i
is
commutative. We then also use the notation g := g
i
[ i I ).
(iii) (A, ( pr
i
[ i I )) is a product in the category theoretical sense of the
family ( A
i
[ i I ).
Proof That each projection is a homomorphism ((i)) is obvious from the
denition of the structure of A in 4.1.3, which also shows that the homomor-
phism condition is violated as soon as the structure of A is enriched at some
place. Moreover, (iii) easily follows from 4.1.1, once we have shown (ii).
Ad (ii): If g : B A exists such that pr
i
g = g
i
for each i I, then
g b = ( g
i
b [ i I ) for each b B according to the denition of projections,
hence there is at most one such g. And if we actually dene g like that, it
is a mapping having the required commutativity properties. But it is also a
homomorphism:
For let b dom
B
, then g
i
(b) dom
A
i
for each i I, therefore g(b)
dom
A
by 4.1.3 and the denition of g. Moreover,
B
(g(b)) = (
A
i
(g
i
(b)) [
i I) =
A
(g(b)).
December 8, 2002
128 Chapter I Basic denitions and properties
Corollary 4.1.5 Let f, g : B

( A
i
[ i I ) be any two homomorphisms
such that pr
i
f = pr
i
g for all i I , then f = g.
Remarks 4.1.6 (i) Notice that according to and with the notation of
4.1.4.(ii) g already exists for every family of mappings, and that it
is a homomorphism just when all g
i
are homomorphisms.
(ii) If I ,= , then

( A
i
[ i I ) ,= iff , for each i I, A
i
,= because of
the Axiom of Choice which actually is equivalent to this fact.
(iii) Notice that, for A :=

( A
i
[ i I ),dom
A
= iff dom
A
i
= for at
least one i I.
(iv) Observe that in Proposition 3.1.5 we have actually enriched the struc-
ture of AA by a sucient amount of nullary constants and an addi-
tional binary operation for each s S in order to guarantee reexivity,
symmetry and transitivity of the closed subsets.
Proposition 4.1.7 Let ( f
i
: B A
i
[ i I ) be a family of homomorphisms
(I any set) and f : B

( A
i
[ i I ) the induced product homomorphism.
Then
ker f =

( ker f
i
[ i I ).
Proof One has (b, b

) ker f iff f(b) = f(b

), iff (f
i
(b) [ i I) = (f
i
(b

) [
i I), iff (b, b

( ker f
i
[ i I ).
Denition 4.1.8 Let ( f
i
: B A
i
[ i I ) be a source of mappings. It is
said to separate B (to be a separating family of B), iff for any two dierent
elements b and b

of B there exists an i I such that f


i
(b) ,= f
i
(b

).
Proposition 4.1.9 Let ( f
i
: B A
i
[ i I ) be a source of homomorphisms.
Then the following statements are equivalent:
(i) ( f
i
: B A
i
[ i I ) separates B.
(ii) The induced homomorphism f : B

( A
i
[ i I ) is injective.
(iii)

( ker f
i
[ i I ) =
B
.
December 8, 2002
4 Products and Limits 129
,
,
,
,
_
`

_
`

f
`

Figure 4.2: A closed graph in the product need not represent a homomor-
phism
We add some further results concerning direct products and homomor-
phisms, which will be of great importance in section 5.
Proposition 4.1.10 Let A, B Alg(), and let f : A B be any mapping;
then:
(i) If f is a homomorphism (f : A B), then graph f is a closed subset
of A B.
(ii) If B is total, and if graph f is a closed subset of A B, then f is a
homomorphism f : A B.
(iii) In general, if graph f is a closed subset of AB, then f need not be a
homomorphism from A into B.
Proof (i) is obvious; and if B is total, then the implication concerning
the existence of the values of operations becomes trivial, while closedness of
graph f guarantees the compatibility condition for homomorphisms. Finally
consider Figure 4.2 in connection with (iii).
Lemma 4.1.11 Let f : A B be a homomorphism, g any mapping out of
A into B such that graph g graph f, and let domg be the relative subalgebra
of A on the domain of g.
(i) g : domg B is a homomorphism.
(ii) graph g is a closed subset of A B iff domg is a closed subset of A.
Proof (i) is obvious. Thus assume graph g to be closed in A B, and
let a dom
A
(domg)
()
. Then g(a) = f(a) dom
B
, and therefore
graph g being closed (
A
(a),
B
(g(a))) graph g, i.e.
A
(a) domg,
showing that domg is closed in A. The converse is proved in a similar
way.
December 8, 2002
130 Chapter I Basic denitions and properties
gures disabled
Figure 4.3: About the maximal (closed) homomorphic extension
Corollary 4.1.12 Let f : A B be a homomorphism and M A any
subset of A. Then the following statements are equivalent:
(i) c
A
M = A.
(ii) c
AB
graph f[
M
= graph f.
The following technical lemma will become quite useful later on (see 5.4.1
and 21.1):
Lemma 4.1.13 Let A, B, C, D Alg(), C a weak subalgebra of A such that
C generates A, D a relative subalgebra of B. Moreover, let f : C D be a
homomorphism which allows a homomorphic extension

f : A B, and let
graph f
#
:= c
AD
graph f. Then one has:
(i) graph f
#
is the graph of a homomorphism f
#
: domf
#
D (domf
#
being the relative subalgebra of A with carrier set domf
#
), and domf
#
is generated by C.
(ii) If E is a C-generated relative subalgebra of A, and if g : E D is a
homomorphic extension of f (i.e. g[
C
= f), then graph g graph f
#
.
(iii) If f
#
is closed, and if g : E D is any closed homomorphic extension
of f (within A) such that E is a C-generated relative subalgebra of A,
then g = f
#
. If A is total, then f
#
is a closed homomorphism from
domf
#
into D. The assumptions of the lemma are sketched in the
commutative diagrams of Figure 4.3.
Proof Ad (i) Since

f : A B is a homomorphism, graph

f is a closed
subset of A B such that graph f graph

f. Therefore
graph f
#
:= c
AD
graph f c
AB
graph f graph

f.
Hence f
#
: domf
#
B is a homomorphism according to Lemma 4.1.11;
but since f
#
[domf
#
] D, we may consider f
#
as a homomorphism
into D. Since graph f
#
= c
AD
graph f
#
[
C
, Corollary 4.1.12 implies
that C generates domf
#
.
December 8, 2002
4 Products and Limits 131
Ad (ii) Now, let g : E D as in (ii) of the Lemma, then Corollary 4.1.12
implies
graph g = c
ED
graph f c
AD
graph f = graph f
#
.
Ad (iii) Let f
#
: domf
#
D be a closed homomorphism, and let g : E
D be another closed homomorphic extension of f such that c
E
C = E.
Then because of (ii) E domf
#
, and graph g graph f
#
. The
inclusion domf
#
E is now proved by algebraic induction:
One has C E; and if b domf such that b =
domf
#
(b), where
b E
()
, then one has f
#
(b) = g(b) dom
D
. Hence b dom
E
,
since g is closed. Therefore b =
E
(b) =
A
(b) E. This shows
E = domf
#
and g = f
#
.
Finally, let A be total, and let a (domf
#
)
()
such that f
#
(a)
dom
D
. Since a dom
A
A being total one has
(
A
(a),
D
(f
#
(a))) graph f
#
= c
AD
graph f.
Thus a dom
domf
#
, i.e. f
#
is a closed homomorphism.
4.2 Pullbacks and products of partial S-algebras
4.2.1 Before we can characterize the products in the category Alg(S) of all
partial S-algebras, we have to describe so-called pullbacks of homogeneous
partial algebras. In this connection we also need the concept of a sink of
homomorphisms or mappings, i.e. of such a family, say, ( f
i
: A
i
B [ i I )
of homomorphisms which all have the same target or codomain (here B ).
Denition 4.2.2 Let C be any category, I be any non-empty set, and let
f := ( f
i
: A
i
B [ i I ) be any sink of C-morphisms. A multiple pullback
of f is a source g := ( g
i
: P A
i
[ i I ) such that
(i) f
i
g
i
= f
j
g
j
for all i, j I, and
(ii) for every source h := ( h
i
: Q A
i
[ i I ), which satises f
i
h
i
=
f
j
h
j
for all i, j I there exists a unique C-morphism h : Q P
such that h
i
= g
i
h for all i I, i.e. the diagram in Figure 4.4 is
commutative.
December 8, 2002
132 Chapter I Basic denitions and properties
gures disabled
Figure 4.4: h: The induced multiple pullback morphism
If I has exactly two elements, then one is speaking of a pullback rather than
of a multiple pullback.
Corollary 4.2.3 Let I be any non-empty set, f = ( f
i
: A
i
B [ i I )
any sink of C-morphisms, and g = ( g
i
: P A
i
[ i I ) as well as g

=
( g

i
: P

A
i
[ i I ) be any two pullbacks of f. Then there exists a unique
isomorphism p : P

P such that g

i
= g
i
p for all i I, i.e. a pullback
whenever it exists is uniquely determined up to unique isomorphism.
The proof is similar to the one of the uniqueness of products (cf. 4.1.2),
therefore we omit the details.
Theorem 4.2.4 Let I be any non-empty set, and let f = ( f
i
: A
i
B [
i I ) be any sink of homomorphisms. Then there exists a multiple pullback
of f. Moreover:
(i) If (A, ( pr
i
: A A
i
[ i I )) = (

( A
i
[ i I ), ( pr
i
[ i I )) is the
direct product of the family ( A
i
[ i I ) of corresponding homogeneous
partial algebras, and if P is the relative subalgebra of A with carrier set
P := a A [ f
i
(pr
i
(a)) = f
j
(pr
j
(a)) for all i, j I ,
then P is a subalgebra of A, and, for p
i
:= pr
i
[
P
(i I), (P, ( p
i
: P
A
i
[ i I )) is a pullback of f (in the heterogeneous case we dene
v
P
:= v
B
f
i
p
i
for any i I.)
(ii) If i

I, and f
i
is surjective, full and surjective, injective or closed for
all i I i

, then p
i
has the same property.
Proof Ad (i) Let h := ( h
i
: Q A
i
[ i I ) be any source such that
f
i
h
i
= f
j
h
j
for all i, j I. Then there exists a homomorphism
h : Q A, where A :=

( A
i
[ i I ), such that pr
i
h = h
i
for all
i I. First we show that h[Q] P:
December 8, 2002
4 Products and Limits 133
Thus consider q Q and q := h(q) A; for all i, j I we have
(f
i
h
i
)(q) = (f
j
h
j
)(q), i.e. f
i
(pr
i
(q)) = (f
i
pr
i
)(h(q)) = (f
i
h
i
)(q) =
(f
j
h
j
)(q) = (f
j
pr
j
)(h(q)) = f
j
(pr
j
(q)), i.e. q P. Moreover, the
denition of v
P
does not depend by denition of P on the special
i chosen from I and is a homomorphism into S such that e.g. v
P
h =
v
B
f
i
pr
i
h = v
B
(f
i
h
i
) = v
Q
, since f
i
and h
i
are assumed to be S-
homomorphisms (i I). From the denition of P it now easily follows
that (P, ( p
i
[ i I )) is indeed a pullback of f.
We still have to show that P is a closed subset of A: In order to
realize this let us observe that each P
ij
:= a A [ f
i
(pr
i
(a)) =
f
j
(pr
j
(a)) (i, j I) is a closed subset of A according to 3.5.3, and
therefore P =

P
ij
[ i, j I is also a closed subset of A.
Ad (ii) Now assume i

I (observe that I ,= ), and let all f


i
be surjective
(injective, full or closed) for all i I i

=: I

.
If A
i
= , then also A and P are empty, and p
i
is an isomorphism
having all these properties, and also for I = i

p
i
is an isomorphism.
Thus we may assume #I 2 and A
i
,= . If, nevertheless, A = , then
A
i
= , for some i I

; but in the case of surjectivity this would imply


B = contradicting A
i
,= and the fact that f
i
is a mapping from A
i

into B the cases of injectivity or of being full and surjective are no


problem, either. And if all f
j
(j I

) are closed, then B and hence A


i

cannot have any existing nullary constants, and the empty mapping p
i

is closed, too. Thus we may additionally assume that A and hence all
A
i
are non-empty:
(a) Let all f
i
(i I

) be surjective, and consider a

A
i
, f
i
(a

) =:
b

B. Then we can nd (by the Axiom of Choice) a sequence a A


such that a(i

) = a

and f
i
(a(i)) = b

for each i I, i.e. a P, and


p
i
(a) = a

, showing that p
i
is surjective.
(b) Let all f
i
(i I

) be full and surjective, and let a P


()
such
that p
i
(a) dom
A
i

. Then (f
i
p
i
)(a) dom
B
, and because of
the surjectivity and fullness of all f
i
(i I

) we can nd sequences
b
(i)
dom
A
i
for each i I

such that f
i
(b
(i)
) = (f
i
p
i
)(a), and all
these sequences compose together with p
i
(a) to a sequence b dom
A
satisfying p
i
(b) = p
i
(a), and which actually belongs to P, showing that
p
i
is fullsurjectivity already follows from (a).
The cases of injectivity respectively of closedness can be treated in a sim-
ilar way and the details are left here as exercises. Moreover, these results
December 8, 2002
134 Chapter I Basic denitions and properties
follow in a more general concept from 10.2.5 and 10.1.2.(viii) together
with 10.3.4.(i) for closedness, respectively with 10.2.8.(iv) and 2.10.3 for
injectivity.
The preceding result also contains the construction of products within the
category Alg(S), since the product of ( A
i
[ i I ) in Alg(S) is by the def-
inition of this category nothing else than the pullback of the family ( v
A
i
:
A
i
S [ i I ) within the category Alg() of homogeneous partial algebras.
Hence we have the
Corollary 4.2.5 In the category Alg(S) products exist for arbitrary families
of partial S-algebras indexed by a set, and for non-empty index sets they
can be computed within Alg() as the multiple pullback of the corresponding
family of sort homomorphisms. For the empty index set the corresponding
product in Alg(S) is given by (S, ), where v
S
:= id
S
. The corresponding
product object will also be denoted by A :=

( A
i
[ i I ), if no confusion
may arise, else we shall write A :=

S
( A
i
[ i I ). Also arbitrary multiple
pullbacks exist in Alg(S), and they can be constructed by just computing the
corresponding multiple pullback within Alg() according to 4.2.4.
Remarks 4.2.6 (i) Theorem 4.2.4 and its proof stand as an example how
one can compute arbitrary so-called limits in the categories Alg()
and Alg(S) as substructures of the corresponding direct product of the
objects involved. For the general concept of limits in category theory
we refer to books on category theory like [HS73] or [Ml71].
(ii) All the results in 4.1 on products and the relationship between homo-
morphisms and their graphs considered as substructures of the direct
product between domain and codomain now carry over to partial S-
algebras without any signicant changes one only has to realize that
the corresponding subobjects are always already contained in the multi-
ple pullback object with respect to the sink of the sort homomorphisms,
i.e. in the product object of the partial S-algebras under consideration.
Proposition 4.2.7 Let (f
i
: A
i
B
i
) be a family of homomorphisms, A :=

( A
i
[ i I ) and B :=

( B
i
[ i I ). If all f
i
are, respectively, surjective,
injective, full or closed, then f :=

( f
i
[ i I ) has the corresponding
property.
December 8, 2002
4 Products and Limits 135
gures disabled
Figure 4.5: g: The induced coproduct morphism
Proof The statement concerning injectivity or surjectivity (using the Ax-
iom of Choice) is easily realized. Let f
i
be full for all i I, let
, a A
()
with f(a) =: ( ( b
ik
[ i I ) [ k () ) dom
B
and
b := ( b
i
[ i I ) :=
B
(f(a)) f[A], and therefore a(i) dom
B
i
for
each i I. Then, since each f
i
is full, there is a
i
=: ( a
ik
[ k K ) dom
A
i
such that f
i
(a
i
) = f
i
(a(i)) and b
i
=
B
i
(f
i
(a
i
)) = f
i
(
A
i
(a
i
)). Thus, for
a

:= ( ( a
ik
[ i I ) [ k () ), we get a

dom
A
and f(a

) = f(a) = b.
In the case of closedness the proof runs in a similar way (exercise!).
4.3 (Global) Coproducts and multiple pushouts
In 4.1 we have seen that the concept of direct products is actually a category
theoretical one. Now, in category theory one gets for every notion a dual one
by reversing the arrows. This leads to the following
Denition 4.3.1 Let C be any category, and ( A
i
[ i I ) any family of
C-objects I any set. A pair (( j
i
: A
i
A [ i I ), A) is called a coproduct
(or free sum) of this family, if, for each i I, j
i
is a C-morphism such that
for any sink g := ( g
i
: A
i
B [ i I ) there exists a unique C-morphism
g : A B satisfying, for each i I, g
i
= g j
i
, i.e. all the diagrams like
Figure 4.5 commute.
If I = , then this denition means that for every B Ob C there exists
exactly one C-morphism from A into B, i.e. A is a so-called left-zero or initial
object of the category under consideration.
In the same way as for products one proves
Proposition 4.3.2 If in a category C a coproduct (( j
i
[ i I ), A) of some
family ( A
i
[ i I ) in C exists, then it is unique up to unique isomorphism.
Remarks 4.3.3 (i) The expression free sum is used in algebra espe-
cially, when all j
i
are injective, but we do not require this in general.
Despite this fact these morphisms are often called the injections or
December 8, 2002
136 Chapter I Basic denitions and properties
embedding morphisms corresponding to a coproduct construction.
The investigation whether these embeddings are really injective is
then usually rather involved and its answer for a given class of total or
partial algebras is an important structural insight.
(ii) There is a basic dierence in the construction of products and coprod-
ucts with respect to interesting subcategories of (partial) algebras of
any given type: While in all subclasses K of Alg(S) (respectively the
corresponding full subcategory) of interest for us, in which products
exist (like primitive and quasi-primitive classes (see 4.6.3), E-varieties,
ECE-varieties, QE- varieties, etc. (see Denition 8.1.2)), the construc-
tion of the products does not depend on the special subclass but is the
same as in the global case, coproducts with respect to K are in general
special epimorphic images of the global coproduct (in the terminol-
ogy of section 5: the K-universal K-solution of the global coproduct).
This fact, too, makes the theory of coproducts usually much more com-
plicated as the one of products, and within this book we shall not need
coproducts very often explicitely. However, it should be observed that
in connection with the important concept of K-free K-algebras with ba-
sis M, denoted by F(M, K), is nothing else but the coproduct within
K of the family (F(m, K))
mM
of (pairwise isomorphic) K-free K-
algebras with one free generator. And this coproduct is total, if K is,
say, a quasi-primitive class of total S-algebras, while the global coprod-
uct of this family would be in a case without nullary constants
just their disjoint union (see 4.3.4 below), i.e. really partial, if there is
at least one at least binary fundamental operation around.
(iii) Coproducts are special cases of colimits, and in this book we shall
rather use a more special sort of colimits, namely directed colimits,
also called direct limits (this notation has historical reasons), which we
shall introduce in the next subsection.
4.3.4 (Construction of Coproducts of Partial Algebras) Let ( A
i
[
i I ) be any set-indexed family of partial algebras of similarity type =
(())

. Let

:= [ () ,= 0 ,
0
:=

, and

:= (())

.
For each i I we call A

i
:= ((A
i
, (
A
i
)

), v
A
i
) the

-reduct of A
i
.
(a) We rst dene the coproduct of the

-reducts of the A
i
:
A

:= ((A

:=
_
i A
i
[ i I , (
A

), v
A
),
December 8, 2002
4 Products and Limits 137
where v
A
(i, a) := v
A
i
(a) for all (i, a) A

, and for all

:
a = ( (i
k
, a
k
) [ k () ) dom
A

iff i
k
= i
k
=: i for all k, k

(),
and ( a
k
[ k () ) dom
A
i
, andifa dom
A

corresponding to
the index i I, then
A

(a) :=
A
i
((i, ( a
k
[ k () ))) we shall
indicate this by sloppily writing graph
A

:=

( igraph
A
i
[ i I )
for every

.
Finally we dene, for each i I, j

i
: A
i
A

, a (i, a) (a A
i
).
(b) We now take care of the appropriate identication of the existing
nullary constants and set

:= ((i,
A
i
), (i

,
A

i
)) [ i, i

I,
0
;
A
i
,
A

i
exist;
Moreover, let := Con
A

be the congruence relation on A

generated
by

. Finally, we set A := A

/, v
A
:= v
A

/
, j
i
:= nat

j
i
: A
i
A,

A
:=
A

/
for all

and
A
:= j
i

A
i
for all
0
for which

A
i
exists for any i I.
If I = , then we dene A := , the empty partial algebra.
Proposition 4.3.5 With the assumptions and notation of 4.3.4 (( j
i
[ i
I ), A) is a coproduct in all of Alg(S) of the family ( A
i
[ i I ).
A is usually denoted by

( A
i
[ i I ) or by

( A
i
[ i I )).
Notice that there is no dierence between the coproduct of ( A
i
[ i I )
within Alg() and of the one of ( A
i
[ i I ) in Alg(S) except for the sort
homomorphism.
Proof If I = , then the statements are obvious. Further, we can observe
that A and all the mappings j
i
(i I) are well dened, that each j
i
is a
homomorphism with respect to the similarity type

preserving constants
whenever they exist according to the denition of

and ; therefore they


are homomorphisms j
i
: A
i
A. Now, let ( g
i
: A
i
B [ i I ) be any
family of homomorphisms. Then there is obviously a unique

- homomor-
phism g

: A

(where B

is the

-reduct of B ) mapping (i, a) onto


g
i
(a), such that g
i
= g

i
for every i I. But whenever
A
i
and
A

i
exist
for some
0
, then g
i

A
i
= g
i

A

i
, and ((i,
A
i
), (i

,
A

i
)) ker g

,
showing that

ker g

, i.e. ker g

. Therefore there exists a unique


December 8, 2002
138 Chapter I Basic denitions and properties

,
,
,

,
, `
_
lPfeil?

,
,
,
, `
_
lPfeil?

,
,
_

+ + =
A
1
A
2
A
3
A
1
+A
2
+A
3
`
` `
`
`
`
`
`

-
`
`
`
`
`
`

-
Figure 4.6: Example of a coproduct with a non-injective canonical injection
homomorphism g : A B such that g

= g nat

(according to the Dia-


gram Completion Lemma 2.7.1). Finally, for every i I one has g
i
= g

i
=
g nat

i
= g j
i
. The uniqueness of g is obvious from the above construc-
tion, from which it also follows, that there is no dierence in the objects
and mappings, when we consider partial algebras in the given family with or
without their sort homomorphisms.
Notation 4.3.6 For nite index sets we usually write A
1
+A
2
and A
1
+A
2
+
. . . +A
n
instead of

( A
i
[ i 1, 2 ) respectively

( A
i
[ i 1, 2, . . . , n ).
Examples 4.3.7 (i) Let := (1, 0, 0), := , , , and consider Fig-
ure 4.6, where the dotted lines indicate the non-trivial congruence
classes of on the disjoint union A

of carrier sets (cf. and

in 4.3.4). The canonical injections should be obvious, but notice that


the one starting from A
1
is not injective.
(ii) In the category of all abelian groups the carrier set of

( A
i
[ i I ) is
given by a

( A
i
[ i I ) [ a(i) ,= 0
i
for only nitely many i I ,
j
i
(a) := a
(a,i)
such that a
(a,i)
(i) = a, and a
(a,i)
(k) = 0
k
, otherwise.
Notice that a
(a,i)
[ a A
i
, i I is up to a natural bijection equal
to

(A
i
i)/Con ((0
i
, i), (0
j
, j)) [ i, j I . In particular one has
A +B = A B.
(iii) But notice, that the above construction in (ii) does not yield the co-
product in the case of groups in general, where the carrier set of G+H
is usually described as the set of all words in elements of G and H
December 8, 2002
4 Products and Limits 139
which are assumed to be disjoint such that no two elements of the
same group are allowed to stand together (the neutral elements are not
allowed at all, and the new neutral element is represented by the empty
sequence). Multiplication of such sequences is the usual concatenation
followed by a reduction process, reducing the new sequence recursively
by replacing two neighbouring elements from the same group by their
product in this group, or by deleting both of them, if this product
becomes a neutral element.
Before we conclude this subsection we briey discuss multiple pushouts, the
concept dual to multiple pullbacks:
Denition 4.3.8 Let C be any category, (A, ( f
i
: A A
i
[ i I ) a source
of C-morphisms. A sink (( g
i
: A
i
P [ i I ), P) of C-morphisms
sometimes only the target object P of this family is called a multiple
pushout of the given sink iff
(a) g
i
f
i
= g
j
f
j
, and
(b) if (( h
i
: A
i
B [ i I ), B) is another sink of C-morphisms satisfying
h
i
f
i
= h
k
f
k
for all i, k I, then there is a unique C-morphism
h : P B such that h g
i
= h
i
for all i I. One gets for all i, j I
a commutative diagram dual to the one in Figure 4.4.
As for products (cf. 4.1.2) and pullbacks one gets the
Corollary 4.3.9 If a multiple pushout exists, then it is unique up to unique
isomorphism.
The concrete construction of multiple pushouts is sketched in the following
Proposition 4.3.10 Let (C, ( f
i
: C A
i
[ i I )) be a source of homomor-
phisms between similar partial algebras, (( j
i
: A
i
A [ i I ), A := ( A
i
[ i
I )) its coproduct,

:= (f
i
(c), f
k
(c)) [ c C,i, k I , and := Con
A

the congruence generated by

. Then a multiple pushout of the given family


is obtained in
(( h
i
:= nat

j
i
: A
i
A/ [ i I ), A/).
December 8, 2002
140 Chapter I Basic denitions and properties
4.3.11 Observe that in the second part of the proof of 3.6.1 we have con-
structed a pushout ((g, h), C) for the source (A, (f, f)) (for f : A B with
c
B
f[A] ,= B).
4.4 Direct limits (= directed colimits)
4.4.1 In the last subsection we have briey discussed coproducts and multi-
ple pushouts, which form special instances of the category theoretical concept
of colimits. The most important kind of colimit for our purpose is the one
of a (k-) directed colimit (for historical reasons also called (k-) direct limit)
(for some cardinal k.) In this subsection we only discuss the fundamentals of
this concept and return to it in Chapter III, when we have additional tools
around, since k-direct limits are a very important concept in connection with
model theoretic results. Nevertheless we think it useful to have this concept
already in this introductory chapter. For the concept of k-directed partially
ordered sets see section 24.
Denition 4.4.2 Let C be any category and k
0
any innite cardinal.
(i) A directed system is any family /
I
:= ( A
i
, f
il
: A
i
A
l
[ i, l I,i l ),
where I = (I; ) is a directed poset, A
i
Ob C for each i I, and each
f
il
is a C-morphism (i l), such that
(DS1) f
ii
= 1
A
i
for every i I.
(DS2) If i l m in I,then f
lm
f
il
= f
im
.
If (I; ) is in addition k-directed, then /
I
will be called a k-directed
system.
(ii) Let /
I
be a directed system in C; then f := (( f
i
: A
i
A [ i I ), A)
is called its directed colimit or its direct limit (other notations are:
colimiting cocone or inductive limit) and this is indicated by writing
f =:
lim

A
I
(often only A =
lim

A
I
) iff (DL1) through (DL3) below
hold:
(DL1) A Ob C, f
i
Mor (C) for each i I.
(DL2) If i l in I, then f
l
f
il
= f
i
(i.e. f is compatible with /
I
).
December 8, 2002
4 Products and Limits 141
gures disabled
Figure 4.7: The induced morphism for direct limits
(DL3) If g := (( g
i
: A
i
B [ i I ), B) is any system compatible with
/
I
, (i.e. g
l
f
il
= g
i
for all i l in I), then there exists exactly
one C-morphism g : A B such that g f
i
= g
i
for every i in
I (i.e. one has commutativity of the diagram in Figure 4.7); g is
then called the direct limit of g or the colimit morphism induced
by g.
If I is k-directed, then f above is also called a k-direct limit (k-directed
colimit). Often only the object A above is denoted as
lim

A
I
.
As before one easily proves:
Proposition 4.4.3 Every (k-) direct limit is unique up to unique isomor-
phism.
Proposition 4.4.4 (Existence of direct limits for partial algebras) Let
be any similarity type of cardinal dimension d and d

:= maxd,
0
, let
I := (I; ) be any directed poset, and let /
I
:= ( A
i
, f
il
[ i l in I ) be a
directed system of partial algebras in Alg(S). Then one has:
(i)
lim

A
I
=: (( f
i
[ i I ), A) exists and is given by (( nat

j
i
[ i
I ),

( A
i
[ i I )/R), where := Con

( A
i
|iI )

(see 4.3.4.f ), and

:= (j
i
(a), j
l
(b)) [ i, l I, a A
i
, b A
l
, and there is m I : i, l
m and f
im
(a) = f
lm
(b).
(ii) If I is d

-directed, then in (i) =

, and we may even dene A :=


(

i A
i
[ i I )/

, where

:= ((i, a), (l, b)) [ i, l I,


a A
i
, b A
l
, and there is m I : i, l m and f
im
(a) = f
lm
(b) ,
f
i
: a [(i, a)]

(i I, a A
i
), v
A
: [(i, a)]

v
A
i
(a), and for
we have dom
A
:= ([(i
k
, a
k
)]

[ k ()) [ there is l I
with ( f
i
k
l
[ k () ) dom
A
l
, and if ([(i
k
, a
k
)

[ k ())
dom
A
with respect to some l I, then
A
([(i
k
, a
k
)]

[ k ()) :=
[
A
l
((l, (f
i
k
(a
k
) [ k ()))]

.
December 8, 2002
142 Chapter I Basic denitions and properties
Proof Since in what follows we shall always assume /
I
to be d

-directed,
we shall only prove (ii) and leave the details of a proof of (i) as an exercise
to the reader as well as the proof of

= R stated in (ii).
Ad (ii) Thus we assume that I is d

-directed. We use the notation of 4.3.4,


especially the one concerning

,
0
and A

, and we claim:
(a)

is a congruence on A

containing the relation

from 4.3.4 identify-


ing all pairs arising from corresponding existing fundamental constants
(therefore A is a well-dened partial S-algebra):
(DS1) implies reexivity of

, symmetry is already contained in the


denition of

, while transitivity of

follows from directedness of I


: If ((i, a), (i

, b)), ((i

, b), (i

, c))

, then there are l

i, i

and
l

, i

such that f
il
(a) = f
i

l
(b) and f
i

l
(b) = f
i

l
(c). Since I
is directed, there is m l

, l

in I, thus f
im
(a) = (f
l

m
f
il
)(a) =
(f
l

m
f
i

l
)(b) = f
i

m
(b) = (f
l

m
f
i

l
)(b) = (f
l

m
f
i

l
(c) = f
i

m
(c).
Therefore

is an equivalence relation. If, for


0
and for some
i, i

in I,
A
i
and
A
i

exist, then there is l i, i

in I, such that
f
il

A
i
= f
i

l

A
i

, i.e. ((i,
A
i
), (i

,
A
i

))

.
It remains to show that

is compatible with all (non-nullary!) oper-


ations of A

:
Let

, and ((i, a
k
), (l, b
k
))

(k ()), and ( a
k
[ k () )
dom
A
i
, ( b
k
[ k () ) dom
A
l
. By assumption there is for each
k () some m
k
i, l in I such that f
im
k
(a
k
) = f
lm
k
(b
k
). Since
I is d

-directed and () < d

, there is some m m
k
(k ())
in I. Thus we get (f
im

A
i
)( a
k
[ k () ) =
A
m
(f
im
(a
k
) [ k
()) =
A
m
(f
lm
(b
k
) [ k ()) = (f
lm

A
l
)( b
k
[ k () ), i.e.
((i,
A
i
( a
k
[ k () )), (l,
A
l
( b
k
[ k () )))

.
(b) The structure dened in 4.4.4 on A is indeed the quotient structure of A

with respect to

as far as non-nullary operations are concerned: This


is an immediate consequence of the denition of A and of what we have
already shown for (a).
(c) (( f
i
[ i I ), A) is a d

-direct limit of /
I
:
That A is indeed a partial S-algebra follows immediately from (a), and
that all f
i
are homomorphisms satisfying (DL2) is a consequence of
the denition of

and the proof of (a). In order to show (DL3)


assume that (( g
i
[ i I ), B) is a sink compatible with /
I
. Then
December 8, 2002
4 Products and Limits 143
there is g

: A

(in the notation of 4.3.4 and of 4.3.5) such that


g

i
= g
i
(i I). Because of the Diagram Completion Lemma 2.7.1 it
remains to show that

ker g

:
Thus assume ((i, a), (l, b))

Then there is m i, l in I such that


f
im
(a) = f
lm
(b), therefore g

((i, a)) = (g

i
)(a) = g
i
(a) = (g
m
f
im
)(a) =
(g
m
f
lm
)(b) = g
l
(b) = (g

l
)(b) = g

((l, b)), i.e. ((i, a), (l, b)) ker g

.
Notice that everything is compatible with the sort homomorphisms.
Remark 4.4.5 The following example shows that in 4.4.4.(ii) d

-directedness
is really needed:
Let A
i
:= (0 N
0
) (1 N
0
) (0, ), (1, ), i N, > i for all
i N, where N
i
:= l N
0
[ l i = i, i + 1, i + 2, . . ., and
f
ij
:
_
_
_
(0, n) (0, n), n 0,
(1, n) (0, n), if i n < j,
(1, n) (1, n), if n j
i.e. successive identication;
:= (N), dom
A
i
:= a
(i)
o
, a
(i)
1
,
where
a
(i)
0
(n) :=(0, n) for all n N,
a
(i)
1
(n) :=(0, n) for 0 n < i,
a
(i)
1
(n) :=(1, n) for n i;

A
i
(a
(i)
0
):=(0, ),

A
i
(a
(i)
1
):=(1, ).
A

=[(0, (0, n))]

, [(0, (0, ))]

, [(0, (1, ))]

[ n N
o
,
A =[(0, (0, n))]

, [(0, (0, ))]

[ n N
o
,
where the elements [(0, (0, ))]

and [(0, (1, ))]

are dierent in A

.
Thus

is no congruence relation on A

, and the index set (N; ) is not

1
-directed.
The following observations are sometimes quite useful:
Proposition 4.4.6 Let /
I
:= ( A
i
, f
il
: A
i
A
l
[ i l I ) with direct
limit f := (( f
i
[ i I ), A) for a d

-directed set I, let g := (( g


i
[ i I ), B)
be any sink compatible with /
I
, and let g : A B be the direct limit of g.
Then we have:
December 8, 2002
144 Chapter I Basic denitions and properties
(i) If all f
il
are isomorphisms, injective, closed, surjective, epimorphisms
or full, then the f
i
have the same property (i I).
(ii) If all g
i
are isomorphisms, injective, closed, surjective, epimorphisms
or full, then g has the same property.
Proof Since the other statements will be proved in a more general con-
text in section 11 see 11.1.5, 11.1.6, 11.3.2 and 11.3.3 using concepts
from model theory and about factorization systems, we only prove now the
statement about fullness, leaving the others here as exercises.
Ad (i) Let all f
il
be full, let a := ( a
k
[ k () ) A
()
i
such that f
i
(a)
dom
A
. Then there are l I and b := ( b
k
[ k () ) dom
A
l
such that f
l
(b) = f
i
(a). Then directedness and d

-directedness of I
imply as in the proof of (a) for 4.4.4 the existence of m m
k

i, l (k ()) such that f
im
(a
k
) = f
lm
(b
k
). But (f
im
(a) =)f
lm
(b)
dom
A
m
and fullness of f imply the existence of a

dom
A
i
such that
f
im
(a

) = f
im
(a), and also f
i
(a

) = (f
m
f
im
)(a

) = (f
i
f
im
)(a) = f
i
(a),
i.e. f
i
is full.
Ad (ii) Let all g
i
be full, and consider c A
()
such that g(c) dom
B
.
Then there exist i I and a A
()
i
such that c = f
i
(a), i.e. g(c) =
(g f
i
)(a) = g
i
(a). Since g
i
is full, there is a

dom
A
i
such that
g
i
(a) = g
i
(a

). Hence g(c) = (g f
i
)(a) = g
i
(a) = g
i
(a

) = (g f
i
)(a

),
where f
i
(a

) dom
A
showing that g is full.
We now present some important examples of the representation of partial
algebras by small (weak) substructures:
Proposition 4.4.7 Let be a similarity type of cardinal dimension d, let
k d

be any regular innite cardinal, and let M be some generating subset


of some partial algebra A. Then consider the following sets:
I
0
I
0
C
A
N) [ N M, #N < k ;
I
1
I
1
B [ B is a relative subalgebra of A, #B < k
and B = c
B
(B M) ;
I
2
I
2
B [ B is a weak subalgebra of A, #B < k
and B = c
B
(B M) ;
I
3
I
3
B [ B Wsuba A, #B < k, B = c
B
(B M),
and #(

dom
B
[ ) < k .
December 8, 2002
4 Products and Limits 145
In each case dene B B

iff B B

and graph
B
graph
B

for all
; and for B B

let f
BB
:= id
BB
: B B

be the identity mapping


on B.
Then (I
n
; ) is a k-directed set, and /
n
(A, M) := ( B, f
BB
[ B
B

in I
n
) is a k-directed system with k-direct limit (( id
BA
[ B I
n
), A)
for 0 n 3.
Proof Since k is assumed to be a regular innite cardinal ( d

), k-directedness
of each poset (I
n
; ) (0 n 3) is easily realized. Notice too, that for
0 k < m 3 each object from I
m
is contained in some object from I
k
and
each object from I
k
contains one from I
m
, while all these objects are weak
relative subalgebras of A. We therefore only prove the statement for I
3
, the
other proofs run similar (exercise!) or will follow from later more general
observations (cf. section 11):
(DL1) and (DL2) are obvious. In order to prove (DL3) let g := ( g
B
: B
C [ B I
3
) be a sink compatible with /
3
(A, M). If we can show that
graph g :=

graph g
B
[ B I
3
is the graph of a homomorphism from
A into C, then the proof is nished, since the required commutativities are
then obvious. Obviously, g is a function, namely if a domg, a B,
a B

for some B, B

I
3
, then g(a) := g
B
(a) = g
B
(a) because of the
compatibility condition. Moreover, M domg is obvious, since k 2. We
prove A domg and the homomorphism condition for g at the same time
by algebraic induction:
If a A


A
[A
()
] [ , then a M domg, and no homomor-
phism condition has to be proved. Thus assume a =

A
(b), b dom

A
,
and that for each k ()

there is B
k
I
3
such that b
k
B
k
. Then dene
B := ((a

B
k
[ k ()

, (
B
)

), v
A
[
B
), where
graph
B
:=
_
b, a)

graph

B
k
[ k ()

, if =

graph
B
k
[ k () , if

.
Then it is obvious k d

being regular that #B < k, #(dom


B
) < k,
and B = c
B
(B M) (cf. 3.6.6), i.e. B I
3
. Since g
B
: B C is a
homomorphism, we get a domg as well as g(a) = g(
A
(b)) = g
B
(
B
(b)) =

C
(g
B
(b)) =
C
(g(b)).
Corollary 4.4.8 (1) If A = c
A
M Alg(S), then A =

B [ B I
n
,
0 n 3 (the union being taken for each component of the structure,
cf. 3.2.2).
December 8, 2002
146 Chapter I Basic denitions and properties
A useful tool is the following result, which follows from the proof of 4.4.7
for I
0
:
Corollary 4.4.9 (2) (Tacking Lemma) Let A and B be similar partial
algebras cardinal dimension d, M a generating subset of A. Then the
following statements are equivalent for every mapping f : M B:
(i) For every subset N of M of cardinality less than d

the restriction f[
N
of f to N has a homomorphic extension f
N
: c
A
N B.
(ii) f has a homomorphic extension f

: A B.
Remarks 4.4.10 (i) Observe that in 4.4.7 and 4.4.8 the assumption k
d

is necessary for the general statement, since for a regular cardinal


k < d

it is possible e.g. in a Peano algebra, which we shall discuss in


section 5 that we need k elements of M A in order to generate
some element of A.
(ii) The examples in 4.4.7 as well as the following one in 4.4.11 concern-
ing congruence relations and quotients are of special model theoretic
interest, as we shall see in Chapter III. Especially for partial alge-
bras, a combination of these two principles: stepwise generation of the
elements and the structure on one side and stepwise enlarging of the
congruence relations, form an important tool in connection with the
generation of algebraic structures. In the theory of total algebras one
usually already starts with a total free (cf. section 5) algebra and one
only has to generate an appropriate congruence, while for partial alge-
bras one additionally has to construct the elements and the admissible
structure.
(iii) If is a nitary similarity type, and if A is any M-generated (partial)
algebra of type , then the statement of 4.4.7 in connection with I
3
says that each element a A can be reached in A from a nite subset
of M using only a nite portion of the structure of A. Nevertheless
the minimal cardinality of this portion may be larger than hardware
and software of a computer allow to represent. Thus the size of a
computer imposes additional restrictions in such a way that usually
only a very restricted part of A can really be represented and dealt
with in a computer, and dierent kinds of computers respectively of
December 8, 2002
4 Products and Limits 147
software can represent dierent (weak) substructures of A. If A is
innite, then in all cases the representable part of A will always be a
partial algebra. However, in general it will be larger than what can be
represented from A at the same time, since one can consider it as the
union of all simultaneously representable parts of A (if the software
is designed that way).
Proposition 4.4.11 Let A be any partial algebra, a congruence on A
generated by some (admissible) relation

, k d

a regular cardinal, and


I
4
:= Con
A

, #

< k , ordered by set theoretical inclusion.


If

in I
4
, then dene p

: A/

A/

to be the homomor-
phism induced according to the Diagram Completion Lemma 2.7.1.
Then /
4
(A,

) := ( A/

, p

in I
4
) is a k-directed system
with direct limit (( nat
/
: A/

A/ [

I
4
), A/).
If

= then we may choose an arbitrary k


0
.
Proof (DL1) and (DL2) are obvious. Let ( f

: A/

C [

I
4
)
be a sink of homomorphisms compatible with /
4
(A,

). It is then easily
seen (cf. 2.5.4.(ii)) that

ker (f

nat

) [

I
4
, and 2.7.1 yields
a unique homomorphism f : A/ C satisfying f nat

= f

for each

I
4
.
4.5 Reduced products and ultraproducts
4.5.1 Closely connected with the construction of direct products (and di-
rect limits) is the one of reduced products. These play an important role
in model theory, especially in the form of ultraproducts. In this book we
shall mainly be concerned with reduced products, e.g. in Chapter III. For
more information about reduced products and ultraproducts in particu-
lar from a category theoretical point of view the reader should consult
[AMakMkiN79], [HiN81], [Hi81], [Sa83] and [AN80].
4.5.2 (Filters and Ultralters) Let I be any set. We recall that a lter
on I is any system T Subs I of subsets of I such that the following axioms
hold:
(F1) T , = (or, equivalently in view of (F3): I T).
(F2) F, F

T implies F F

T.
December 8, 2002
148 Chapter I Basic denitions and properties
(F3) F T and F F

I imply F

T.
If, in addition, one has
(F4) / T (or equivalently in view of (F3): T , = Subs I),
then we call T a proper lter on I. And if, moreover, T satises
(F5U) For all M I either M T or I M T,
then T is called an ultralter on I; i.e. an ultralter is always a proper lter
which is obviously maximal with respect to this property, and it is well known
that every maximal proper lter on I is indeed an ultralter on I.
Obviously, for every M I, M := N I [ M N is a lter
on I which is proper if M ,= , and which is an ultralter if M = i is a
one-element set. M is called the principal lter of I generated by M.
Other important examples of lters are the so-called generalized Frechet
lters on I: Let
0
c #I be any cardinal number, then T
c
:= F I [
#(I F) < c is a non-principal proper lter for every innite set I nite
sets possess only principal lters. Although no non-principal ultralters are
explicitly known, it follows from Zorns Lemma that every proper lter is
contained in some ultralter.
Let k be any cardinal and T a lter on I. T is called k-complete, if in
addition to (F2)
(F2,k) For every subset / T, where #/ < k, one has

/ T, i.e.
the intersection of less than k elements of T also belongs to T.
Obviously every lter is k-complete for 3 k
0
. If
0
c #I, and
if c is additively inaccessible, then T
c
is c-complete, since

I D
k
[ k
K = I

D
k
[ k K. The intersection of k-complete lters on I is
again a k-complete lter on I.
With respect to the fact that direct limits are available, the easiest way to
dene reduced products is the following one, which we shall usually use later
on:
Denition 4.5.3 Let I be any set, ( A
i
[ i I ) any family of similar partial
algebras, and let T be any k-complete lter on I. We rst introduce on T
the dual ordering with respect to inclusion, i.e. for F, F

T set: F
F

iff F F

. Thus we set T := (T; ) = (T; ). For each F T set


December 8, 2002
4 Products and Limits 149
A
F
:=

( A
i
[ i F ); and for F G T dene p
FG
: A
F
A
G
by
A
F
a a[
G
, i.e. p
FG
( a
i
[ i F ) := ( a
i
[ i G).
We call p
FG
the restriction mapping with respect to F and G. Obviously,
p
FG
is a homomorphism from A
F
into A
G
induced by the family ( pr
i
:
A
F
A
i
[ i G) of projections. Hence it is easy to realize that /
F
:=
( A
F
, p
FG
[ F G T ) is a k-directed system (exercise!), and its k-direct
limit (( p
F
[ F T ),
lim

A
F
) is called a (category theoretical) k-reduced
product of ( A
i
[ i I ) with respect to the lter T.
In particular, we introduce for the resulting object the notation

k
( A
i
[
i I )/T :=
lim

A
F
(=:

( A
i
[ i I )/T if k d

, since then the structure


is independent of how much k is larger than d

, as the construction of direct


limits in 4.4.4 shows).
If T is an ultralter, then

k
( A
i
[ i I )/T is called an ultraproduct.
Remarks 4.5.4 (i) In model theory one usually denes reduced products
in another way, which we want to sketch briey:
Let ( A
i
[ i I ) be any family of partial algebras which have non-
empty phyla for all s S if any phylum is empty for some i I,
then the corresponding phylum of

( A
i
[ i I ) will be empty , T a
d

-complete lter on I. On the cartesian product A := ( A


i
[ i I )
we dene, for a, b A, I
ab
:= i I [ a(i) = b(i) (the equalizer of a
and b), and moreover:
()
F
:= (a, b) A A [ I
ab
T .
The assumptions on T easily allow us to show that
F
is a congruence
on A. But the structure of the reduced product A/T, whose carrier
set is A/
F
, is for partial algebras usually richer than the quotient
structure (for total algebras we may dene A/T := A/
F
): Let us
denote the elements of A/T := A/
F
by a/T, and sequences of them
by a/T.
One denes for and for any sequence a/T (A/T)
()
:
a/T dom
A/F
iff there is some corresponding sequence a

A
()
such that a
/
T = a

/T and i I [ a

(i) dom
A
i
T; and if
a/T dom
A/F
with corresponding sequence a

, then
A/F
(a/T) :=
( b
i
[ i I )/T, where b
i
:=
A
i
(a

(i)), if a

(i) dom
A
i
, and any
element of appropriate sort from A
i
is taken for b
i
, otherwise.
December 8, 2002
150 Chapter I Basic denitions and properties
(ii) With our assumptions in (i) of non-empty phyla the two denitions
of reduced products are equivalent, while with empty phyla allowed in
(i), the denitions may be inequivalent, since A
I
may have an empty
phylum, while this is non-empty in A
F
for some F T (see [ABN80],
p. 345).
Because of such features we prefer the category theoretical denition.
(iii) With the notation from (i) and 4.5.3 one has:
If T := M is a principal lter, then A/T

=

( A
i
[ i M ).
If i I [ A
i
is total T, then A/T is total, even when some A
i
are
discrete.
Another example, where A is discrete, but A/T is total, is given as
follows:
Let := N, the set of natural numbers, := (1)
nN
, I := N, A
n
:= Q,
the set of rational numbers (n N), and
m
A
n
is nowhere dened, if
m = n, and it is multiplication by m, otherwise. T will be the Frechet
lter of all subsets of I with nite complement.
(iv) The relationship between B :=

S
( A
i
[ i I )/T (heterogeneous case)
and C :=

( A
i
[ i I )/T (homogeneous case) is described as follows:
B is always a subalgebra of C .
Namely because of 4.2.4.(i), for each F T, the carrier set of

S
( A
i
[
i F ) is a closed subset of

( A
i
[ i F ), and in a generalization of
4.4.6.(ii) we shall see in section 11 that this implies that B is a closed
subset of C.
4.6 Some properties of the operators H, .., o, .. and T, ..
4.6.1 In the previous subsections we have introduced several constructions
of new partial algebraic structures from (families of) given ones. These con-
structions lead to algebraic closure operators on the class of all partial (S-)
algebras of some given type, which especially gain importance in connection
with implicational model theory in Chapters II and III, where we shall also
need their behaviour with respect to other operators. Some of their proper-
ties will be investigated in this subsection.
Denition 4.6.2 Let K Alg(S) be any subclass. We dene
December 8, 2002
4 Products and Limits 151
HK := H
w
K := B Alg(S) [ there are A K and a surjective homo-
morphism f : A B = class of all weak homomorphic images of
K-algebras;
H
f
K := B Alg(S) [ there are A K and a full and surjective homo-
morphism f : A B = class of all full homomorphic images of
K-algebras;
H
c
K := B Alg(S) [ there are A K and a closed and surjective
homomorphism f : A B = class of all closed homomorphic
imagesof K-algebras;
1 K := B Alg(S) [ there are A K and an isomorphism f : A B
= class of all isomorphic copies of K-algebras;
o K := o
c
K := B Alg(S) [ there are A K and a closed monomor-
phism f : B A = class of all closed subalgebras (up to isomor-
phism) of K-algebras;
o
f
K := o
r
K := B Alg(S) [ there are A K and a full monomorphism
f : B A= class of all relative subalgebras (up to isomorphism)
of K-algebras;
o
w
K := B Alg(S) [ there are A K and a monomorphism f : B A
= class of all weak subalgebras (up to isomorphism) of K-algebras;
T K := B Alg(S) [ there are a set I and a family ( A
i
[ i I ) K
I
such that B

=

( A
i
[ i I ) = class of all products of (set indexed)
families of K-algebras;
T
r,c
K := B Alg(S) [ there are a set I, a c-complete lter T on I and
a family ( A
i
[ i I ) K
I
such that B

=

( A
i
[ i I )/T =
class of all c-reduced products (up to isomorphism) of (set indexed)
families of K-algebras; here c is any innite cardinal;
T
r
K := T
r,d

K;
T
u
K := B Alg(S) [ there are a set I, a d

-complete ultralter T on I
and a family ( A
i
[ i I ) K
I
such that B

=

( A
i
[ i I )/T
= class of all ultraproducts (up to isomorphism and for d

-complete
ultralters) of (set indexed) families of K-algebras;
December 8, 2002
152 Chapter I Basic denitions and properties
o K := K = K and the empty algebra;
e K := K 1 o = K and all products for the empty index set.
Operators T
+
and T
r
+
are dened like the corresponding operators T and T
r
,
except that only non-empty index sets and proper lters are allowed.
Let O be the semigroup generated by the above operators applica-
tion from the left , i.e. of all nite sequences
1

2
. . .
n
for some natural
number n 1, where each
i
(1 i n) is one of the above operators.
For ,

O we dene

iff for every K Alg(S) one has K

K;
and we set =

, if

and

.
For convenience we add to the generating set of operators of O the oper-
ator i : i K := K for all K Alg(S), i.e. the identity operator.
For later use we dene
Denition 4.6.3 Let K Alg(S). K is called
(i) primitive, iff HoT K = K, and
(ii) quasi-primitive, iff oT K = K.
Lemma 4.6.4 Let O, then one has on Alg(S) (if T
u
is concerned,
has to be nitary):
(i) i = i = , and K K

Alg(S) implies K K

.
(ii) i ; and 1, O, 1 imply 1 and 1 .
(iii) 1 = 1.
(iv) 1 = , if H, H
f
, H
c
, o, o
r
, o
w
, 1, T, T
r
, T
u
, T
+
, T
r
+
or if
O is a sequence which somewhere contains one of these operators.
(v) = i, H, H
f
, H
c
, o, o
r
, o
w
, 1, T, T
r
, T
u
, T
+
, T
r
+
, o, e.
(vi) T = T
+
e = eT
+
and T
r
= T
r
+
e = eT
r
+
.
December 8, 2002
4 Products and Limits 153
(vii) 1 H
c
H
f
H,
1 o o
r
o
w
,
1 T
u
T
r
,
1 T,
T
+
T T
r
,
T
+
T
r
+
T
r
.
(viii) Let H, H
f
, H
c
, 1 o, o
r
, o
w
, T, T
r
, T
u
, T
+
, T
r
+
,
then 1 1, , and 1 1.
(ix) With the notation from (viii) we have that each of the operators ,
1, , 1, , 1 and 1 is a closure operator on Alg(S).
(x) When we only allow non-empty phyla, or only sort algebras S where
S has less than d

elements, or if we require that all lters are d


&
-
complete, for d
&
:= maxd

, #S, then we have for any regular car-


dinal c d
&
: T
r,d
&
+
HT
+
, and P
r,d
&
HT.
(xi) If c d

(respectively c d
&
(cf.(x)) is some regular cardinal, then
every statement above about T
r
is also true for T
r,c
.
Proof (Sketch) The statements (i), (ii), (iii), (iv), (vi) and (vii) are easily
realized from the denitions. In what follows K is any subclass of Alg(S).
Ad (v) is obvious. If i, H, H
f
, H
c
, o, o
r
, o
w
, 1, o, e,
is easily seen.
Let us argue for products rst: Let for each i I (A
i
:=

( B
ij
[ j
J
i
), ( p
ij
[ j J
i
)) be products, and let (A :=

( A
i
[ i I ), ( p
i
[ i
I )) be their product. We claim that (A, ( p
i
p
ij
[ i I, j J
i
)) is a
product of ( B
ij
[ i I, j J
i
):
Let ( f
ij
: C B
ij
[ i I, j J
i
) be a source of homomorphisms.
Then we have for each i I a unique f
i
: C A
i
satisfying p
ij
f
i
= f
ij
(j J
i
). These f
i
induce a unique homomorphism f : C A satisfying
p
i
f = f
i
for each i I; thus p
ij
p
i
f = p
ij
f
i
= f
ij
for each i I and
j J
i
. If I ,= and some J
i
,= , then also

( i J
i
[ i I ) ,= .
In the case of reduced products let with the above notation T
i
be a d

-complete lter on J
i
(i I), and T be a d

-complete lter on I.
Then it is not too dicult, but combined with some computation, to
see that ( :=

( i G
i
[ i I ) [ G
i
J
i
(i I) and i I [ G
i

December 8, 2002
154 Chapter I Basic denitions and properties
gures disabled
Figure 4.8: Iteration of reduced products (using direct limits)
T
i
T is a d

-complete lter on J :=

( i J
i
[ i I ), and that

(B
ij
[ (i, j) J)/(

=

( B
ij
[ j J
i
)/T
i
[ i I)/T, where one
can use in the case of non-empty phyla that (b
ij
[ (i, j) J)/( (( b
ij
[
j J
i
)/T
i
[ i I)/T, and (b
ij
[ (i, j) J)
G
(b

ij
[ (i, j) J) iff i [
I
( b
ij
|jJ
i
)( b
ij
|jJ
i
)
T
i
T, iff (( b
ij
[ j J
i
)/T
i
[ i I)/T = (( b

ij
[
j J
i
)/T
i
[ i I)/T (for notation see 4.5.4); for the structure the
argumentation is similar.
For the general case we sketch the main ideas using the argumenta-
tion via direct limits and some results, which will only be proved in
section 11:
First we introduce some further notation (recall that we start with the
notation introduced above): For M J = (i, j) [ i I and j J
i
,
for i I and for H J
i
dene
M
i
:= j J
i
[ (i, j) M ,
F
M
:=i I [ M
i
T
i
,
B
iH
:=

( B
ij
[ j H ) (projections p
iHj
: B
iH
B
ij
for j H),
B
G
:=

(B
ij
[ (i, j) G), for G J,
and for H H

T
i
let f
iHH
: B
iH
B
iH
be the homomorphism
induced by the family ( pr
iHj
: B
iH
B
ij
[ j H

) of projections and
for G G

in J let g
GG
: B
G
B
G
be the homomorphism induced by
the family (pr
Gij
: B
G
B
ij
[ (i, j) G

) of projections. Moreover, let


lim

( B
iH
,f
iHH
|HH

F
i
)
=: (( f
iH
[ H T
i
), B
i
),
lim

( B
G
,g
GG
|GG

G )
=: (( g
G
[ G ( ), B).
And let D
F
:=

( B
i
[ i F ) for F I (with projections pr
Fi
:
D
F
B
i
for i F), and d
FF
: D
F
D
F
be induced by the family
( pr
Fi
[ i F

),
lim

( D
F
,d
FF
|FF

F )
=: (( d
F
[ F T ), D). Dene for
G ( (cf. Figure 4.8)
q
G
:= f
iG
i
g
G{i}G
i
[ i F
G
) : B
G
D
F
G
,
December 8, 2002
4 Products and Limits 155
i.e. we get pr
F
G
i
q
G
= f
iG
i
g
G{i}G
i
for each i F
G
. If G G

in (,
then the denitions above (in prticular that the g
GG
are projections
pr
GG
between direct products) show that q
G
g
GG
= d
F
G
F
G

q
G
, i.e.
the family (m, ( q
G
[ G ( )) with m : ((; ) (T; ) (G F
G
)
is a morphism from the directed system ( B
G
, g
GG
[ G G

( )
into the directed system ( D
H
, d
HH
[ H H

T ) (see 11.1.1), and


according to 11.1.4.(i) this morphism induces a homomorphism q : B
D between the direct limit objects, such that, for all G (, q g
G
=
d
F
G
q
G
(see the in all parts commutative diagram in Figure 4.8, where
we assume G

G in ( and G

i
T
i
). We have to show that q is an
isomorphism:
(a) (Surjectivity) Let a
lim

( D
F
,d
FF
|FF

F )
. Then there are
F T and a D
F
such that d
f
(a) = a, where a =: ( a [ i F ) D
F
.
Thus, for each i F there are F
i
T
i
and b
i
=: ( b
ij
[ j F
i
) B
iF
i
,
such that a
i
= f
iF
i
(b
i
). Dene G :=

( i F
i
[ i F ), which
obviously belongs to (, and b := (b
ij
[ (i, j) G); then it is easy to
realize that a = q
G
(b) and a = (d
F
G
q
G
)(b) = (q g
G
)(b), showing that
q is surjective.
(b) (Injectivity) Let b, b

B such that q(b) = q(b

). Then there
are G, G

( (and because of directness of ( we may assume G = G

)
and b, b

B
G
with b = (b
ij
[ (i, j) G), b

= (b

ij
[ (i, j) G), such
that g
G
(b) = b, g
G
(b

) = b

, i.e. (d
G
q
G
)(b) = (q g
G
)(b) = (q g
G
)(b

) =
(d
G
q
G
)(b

). Then there is F T, F F
G
such that (d
F
G
F
q
G
)(b) =
(d
F
G
F
q
G
)(b

) =: c =: ( c
i
[ i F ), i.e. for G

:=

i G
i
[ I
F (yielding F = F
G
), we have c = (q
G
g
GG
)(b) = (q
G
g
GG
)(b

) (see
again Figure 4.8). For each F we have therefore
c
i
= pr
Fi
(c) = (pr
Fi
d
F
G
F
q
G
)(b)
= (f
iG

i
g
G

{i}G

i
g
GG
)(b)
= (f
iG

i
g
G

{i}G

i
g
GG
)(b

)
Observe that f
iG

i
belongs to the colimiting cocone of
lim

( B
iH
,f
iHH
|HH

in F
i
)
.
Thus there is, for each i F, an H
i
G

i
= G
i
in T
i
such that
(f
iG

iHi
g
G

{i}G

i
g
GG
)(b) = (f
iG

iHi
g
G

{i}G

i
g
GG
)(b

), i.e. g
G{i}H
i
(b) =
g
G{i}H
i
(b

), yielding b
ij
= b

ij
for all (i, j)

iH
i
[ i F =: G
#
.
Then G
#
(, since f T and H
i
T
i
for all i F, and g
GG
#(b) =
g
GG
#(b

),i.e. g
G
(b) = g
G
#(g
GG
#(b) = g
G
#(g
GG
#(b

) = g
G
(b

), showing
that q is injective.
December 8, 2002
156 Chapter I Basic denitions and properties
(c) (Fullness) Let b B
()
such that q(b) dom
D
. Thus there
are F T and a dom
D
F
such that d
F
(a) = q(b). Then, for each
i F, a(i) dom
B
i
, i.e. there are H
i
T
i
(for each i F) and c
i
=:
( ( c
ijk
[ j H
i
) [ k () ) dom
B
iHi
such that a(i) = f
iHi
(c
i
). Set
G :=

i H
i
[ i F and c := ( ( c
ijk
[ (i, j) G) [ k () )
B
G
()
. Then c dom
B
G
and q
G
(c) = aGU, whence g
G
(c) dom
B
and (d
F
q
G
)(c) = d
F
(a) = q(b) = (q g
G
)(c) show that q is full. (a), (b)
and (c) imply that q is an isomorphism.
Assume now that the similarity type is nitary, that T and all T
i
(i I) are ultralters, and let ( be the lter dened above. We claim
that ( is also an ultralter: Thus assume M J, M / (; then F :=
i I [ M
i
:= j J
i
[ (i, j) M T
i
/ T. Let

M := J M,

F := I F. Then

F T, since T is an ultralter, and for each i

F
M
i
/ T
i
. Therefore

M
i
:= J
i
M
i
= j J
i
[ (i, j)

M T
i
,
since T
i
is an ultralter. Thus

M ( , showing that ( is an ultralter
(observe that ( cannot be equal to Subs J).
In the remaining cases the argumentation is similar.
Ad (viii) VY YV Let A be a 1-substructure of a -homomorphic image
B of C K, i.e. we have f : C B in (and w.l.o.g. A B).
Let (D, (f

: D A, g : D C)) be the pullback of ((id


AB
, f), B).
According to 4.2.4.(ii) f

inherits all the important properties from f,


respectively g those from id
AB
showing that A 1 K; hence 1 1.
In order to show 1 ,= 1 choose to be the type of commutative
rings with unit: := (2, 1, 0, 2, 0) (S the one-element total algebra, i.e.
homogeneous case), let D := Z be the ring of integers and A := Z/3Z
its quotient (f

:= nat

3Z
), and consider D as a subring of the ring Q
of rational numbers. Obviously A 1 Q, but A / 1 Q, since
Q only allows homomorphisms onto itself or into the one-element ring,
and A is not a substructure of any of them.
WY YW Let A be a -product of -homomorphic images of K-
algebras, say, for ( C
i
[ i I ) K
I
we have f
i
: C
i
B
i
(i I)
and a d

-complete lter T on I (which is just I in the case of prod-


ucts and an ultralter in the (nitary) case of ultraproducts) such that
A

=

( B
i
[ i I )/T =: B with an isomorphism g : A B. If
C :=

( C
i
[ i I )/T, then let, for each F T, f
F
: C
F
B
F
be the
direct product of the family ( f
i
[ i F ). Then (id
F
, ( f
F
[ F T ) is a
conal morphism between directed systems in the sense of 11.1.1, which
December 8, 2002
4 Products and Limits 157
induces, according to 11.1.4.(i), a homomorphism f : C B. As stated
in 4.2.7, each f
F
inherits properties like surjectivity, fullness or closed-
ness from the f
i
; and in 11.1.6, respectively in 11.3.3 we show that f
inherits these properties from the f
F
. And this argument is also suitable
to prove 1 1. In general inequality also holds in the two state-
ments under consideration; for more information see H.H oft [Hoe72]
or D.Pigozzi [Pig66].
Ad(ix) Now, for two operators A and Z with the properties (i),(ii),(iv),(v)
and ZA AZ, AZ is a closure operator : monotony and extensity
follow from (ii) and (i), while AZAZ = A(ZA)Z AAZZ = AZ
shows idempotency.
Ad(x) Let ( A
i
[ i I ) K
I
be any family of K-algebras and T an at least d

-
complete lter on I. When all phyla are non-empty, then C :=

( A
i
[
i I )/T is obviously a (weak) homomorphic image of A :=

( A
i
[ i
I ) (via p
I
, see 4.5.3). If empty phyla are allowed, but #S > d

, then for
every s S for which C
s
,= , there exists an F
s
T such that A
is
,=
for all i F
s
. Set F
s
:= I for all other s S. Then d

-directedness
of T implies F :=

( F
s
[ s S ) T, and C is then a homomorphic
image of A
F
:=

( A
i
[ i F ). However, if T is a d

-complete lter,
then again the set F constructed above belongs to T.
That the assumptions are necessary is shown by the following example,
where we consider only sets and let I =: S be any innite set, A :=
((0 S) (1 S), v
A
), where v
A
(j, s) := s (j 0, 1, and s S),
and A
i
:= A (0, i), (1, i) with the restriction of v
A
as sort mapping
v
i
(i I), i.e. all phyla contain zero or two elements and exactly one
phylum is empty. Let T be the Frechet lter on I of all subsets of I,
whose complement has cardinality less than I. Then all A
F
(F T)
in fact all products with non-empty index set of any family ( B
j
[
j J ) A
i
[ i I
J
have at least one empty phylum, while in

( A
i
[ i I )/T each phylum is innite. Now let be nitary,
and consider A T
r
K, i.e. we have a set I, a lter T on I and a
family ( A
i
[ i I ) K
I
such that A

=

( A
i
[ i I )/T. If I = ,
then A is isomorphic to the so-called empty product and hence belongs
to oTT
u
K. Therefore let I be non-empty. If T is not proper, then the
result is the same as if I were empty; therefore T can also be assumed to
be proper. Now let u(T) := ( [ ( T is ultralter on I . Then the
Axiom of Choice tells us that u(T) ,= and T =

u(T). We claim that


December 8, 2002
158 Chapter I Basic denitions and properties
A is isomorphic to a subalgebra of

( A
i
[ i I )/( [ ( u(T)):
Let /
F
:= ( A
F
, p
FF
[ F F

T ) be the direct system belonging


to A, and let /
G
:= ( A
G
, q
GG
[ G G

( ) be the direct system for


any ( u(T) with direct limit (( q
G
[ G ( ), A(()). Since /
F
is a
subsystem of /
G
the subfamily ( q
F
[ F T ) induces a homomorphism
q
G
: A A((). Hence we get a product homomorphism q : A

(A(() [ ( u(T)) =: B. It remains to show that q is injective and


closed: Observe that for any ( u(T) and for any F T the system
([
F
:= G ( [ G F is always an ultralter on F.
(a) (Injectivity) Let a, a

A, a ,= a

, and let a, a

A
F
for some
F T such that p
F
(a) = a and p
F
(a

) = a

. Then the equalizer


F
aa
(see 4.5.4.(i)) does not belong to T[
F
, i.e. (I F) F
aa
/ T and
therefore there is some ( u(T) such that F
aa
does not belong to ([
F
,
whence (I F) F
aa
/ (, implying G := F F
aa
(, since ( is
an ultralter. This shows that q
G
(q
FG
(a)) ,= q
G
(q
FG
(a

)) in A((), and
therefore q(a) ,= q(a

), i.e. q is injective.
(b) (Closedness) Let a A
()
such that q(a) dom
B
; assume
a =: p
F
(a) for some F T. Then pr
G
(q(a)) dom
A(G)
for each (
u(T), i.e. there is F G(() ( such that q
FG(G)
(a) domvarphi
A
G(G)
.
Let H := i F [ a(i) dom
A
i
; then H G(() for each ( u(T),
i.e. H

u(T) = T, and therefore p


FH
(a) domvarphi
A
H
, implying
a = p
H
(p
FH
(a)) dom
A
. This shows that q is closed.
Ad(xi) The proof of this statement is left as an exercise for the reader. How-
ever, observe that I is a c-complete lter on I for every cardinal c,
and the assumption c > d

guarantees that all the statements con-


cerning the structure remain true. This ends the sketch of the proof of
Lemma 4.6.4.
December 8, 2002
5 Universal solutions, Freeness, Terms 159
5 Universal solutions, free partial algebras,
Peano algebras, terms and termoperations
The concepts of (relatively) free algebras in particular of absolutely free
algebras in TAlg(S), i.e. of term algebras (see section 5.2) and more gener-
ally the one of a K-universal solution of a partial algebra A for some class K
of (partial) algebras play a fundamental role in universal algebra and also in
the theory of partial algebras. They provide a tool to characterize (partial)
algebraic structures up to isomorphism by specifying K usually by rst
order axioms and a possibly easily describable partial algebra A
(e.g. in group theory the construction of K-universal solutions corresponds
to the characterization of a group by (axioms and) by generators and re-
lations, see 5.1.2 for an example). These concepts are of basic importance
in Computer Science, too, since they include the concept of initial algebras
(see 5.11.12 and section 17.6), i.e. of the K-free K-algebra K-freely gener-
ated by the empty set (usually for an (E-) equational (or implicational)
class K). Namely, initial algebras are those algebraic structures which arise
as the semantics of algebraic specication in connection with one branch of
the specication of abstract data types (the other one using so-called nal al-
gebras). Much motivation for this initial algebra approach to abstract data
types can be found in [GoTcWa78] written by the so-called ADJ-group.
For later purposes it is useful to start with the most general concept of
K-universal solutions (connected with the category theoretical concept of a
left adjoint functor for the embedding functor from K into Alg(S) which
will be avoided in this context). And starting from these general concepts
we mainly discuss in this section the case when K = TAlg(S) is the class of
all total S-algebras. The results presented here then form the basis of the
treatment of more general situations and of a model theoretic approach to
partial algebras in later chapters.
In connection with the most important structures of this section the
term algebras (Peano algebras, word algebras) and more generally the abso-
lutely free completions three dierent approaches are given: the category
theoretical one homomorphism extension property for all total algebras
, an axiomatic one characterizing the internal structure in an axiomatic
way , and an explicit set theoretical construction. In so far this section
is also important as a prototype for modelling the semantics of an algebraic
specication of an abstract data type. As before we usually consider partial
December 8, 2002
160 Chapter I Basic denitions and properties
gures disabled
Figure 5.1:
gures disabled
Figure 5.2:
S-algebras without mentioning it, except when we explicitly restrict consid-
erations to the homogeneous case.
5.1 The general concepts of universal solution and free-
ness
Denition 5.1.1 Let C be any category, K Ob C any class of C-objects
and f : A B a C-morphism.
(a) f is called K-extendable, iff for every C K and for every C-morphism
g : A C there exists a C-morphism h : B C such that h f =
g (i.e. the diagram in Figure 5.1 is commutative). (cf. Herrlich,
Strecker [HS73], Denition 37.8; in [NSa82], [AN83] and other pa-
pers of Andr eka, N emeti and Sain f is said to be valid in K and K
is said to be injective w.r.t. f)
(b) f is called K-universal, if it is a K-extendable epimorphism, and if
for every K-extendable epimorphism f

: A B

there exists a C-
morphism l : B

B such that l f

= f (cf. Figure 5.2). Observe that


l is then an epimorphism, too.
(c) If f : A B is a K-universal epimorphism, then (f, B) is called a
K-universal solution of A, and if, in addition, B K, then (f, B) is
called a K-universal K-solution of A.
Examples 5.1.2 Let us present already at this point some simple examples:
(i) For every A Alg(S) and for every K Alg(S) the identity mapping
id
A
: A A is K-extendable.
December 8, 2002
5 Universal solutions, Freeness, Terms 161
(ii) Let F(M, Gr) be the free group with free generating set M, then
id
M
: M F(M, Gr) is a Gr-universal homomorphism (where M
is considered as a discrete partial algebra, and (id
M
, F(M, Gr)) is a
Gr-universal Gr-solution of M. If Z designates the set of all integers,
then (Z; +, 0, ) is the free (abelian) group, freely generated by the
empty set, i.e. the initial object in Gr.
(iii) When e designates the neutral element of F(a, b, Gr), let (a, b; a
4
e
=
e, b
2
e
= e, b a
e
= a
3
b) be the specication of a group by genera-
tors and relations. Then this corresponds to a partial algebra A :=
(e, a, b, a
2
, a
3
, c;
A
, e
A
,
1A
) where graph
A
:= ((a, a), a
2
), ((a, a
2
), a
3
), ((a, a
3
), e), ((b, b), e), ((b, a), c), ((a
3
, b), c),
e
A
= e and
1A
is the empty operation.
The Gr-universal Gr-solution of A is then the dyhedral group D
8
, i.e.
the group characterized by the above generators and relations.
(iv) Let R be the class of all rings with unit and R R be any ring,
x / R, R[x] the polynomial ring with the variable x and coecients
in R. Then id
R{x}
: R x R[x] is the R-universal R-solution of
R x = R +x
d
.
(v) Let A be any set and A A be any relation. If is a reexive
quasi-order on A, then the relation
:= (a, b) A A [ (a, b) and (b, a)
is an equivalence relation on A, and the quotient structure (A/ ; /
) is a poset which is the P-universal P-solution of (A; ) in the sub-
category P of all posets within the category Q of all quasi-orders with
quasi-order preserving mappings, where
/ := ([a]

, [b]

) [ (a, b)
= ([a]

, [b]

) [ a, b A, (a

, b

) for some a

[a]

, b

[b]

.
Proposition 5.1.3 Let C be any category, K any class of C-objects and A
Ob C.
(i) Any K-universal C-morphism for A is unique up to unique isomor-
phism, i.e. if f : A B and f

: A B

both are K-universal epimor-


phisms, then there exists a unique isomorphism i : B B

such that
i f = f

.
December 8, 2002
162 Chapter I Basic denitions and properties
(ii) If f : A B is a K-extendable C-epimorphism and B K, then f is
K-universal, and (f, B) is a K-universal K-solution of A.
(iii) The composition of K-extendable C-morphisms is K-extendable and each
isomorphism is K-extendable.
Proof Ad (i) Let f : A B and f

: A B

be K-universal epimor-
phisms. Then there exist according to 5.1.1.(b) morphisms g : B B

and h : B

B such that g f = f

, h f

= f, 1
B
f = f = h f

=
h g f. Since f is an epimorphism, this implies 1
B
= h g. Similarly
we can show g h = 1
B
; therefore g and h are isomorphisms inverse to
each other. Uniqueness follows, since both f and f

are epimorphisms.
Ad (ii) Since B K, 5.1.1.(b) and (c) are easily veried.
Ad (iii) These statements are immediately realized.
Denition and Notation 5.1.4 (i) The result of 5.1.3.(i) allows to in-
troduce a special notation for K-universal epimorphisms, whenever they
exist: We write r
A,K
: A F(A, K), if this is a K-universal solution
of A Alg(S) for some K Alg(S). If A is discrete, we usually write
F(A, K) instead of F(A, K), and if F(A, K) K, then we call this par-
tial algebra the K-free K-algebra, K-freely generated by A (even when
r
A,K
is not injective), and this expression is sometimes also used, when
A is not discrete (instead of K-universal K-solution of A).
(ii) Let M A Alg(S): If id
M
: M C
A
M is K-extendable, i.e. if every
mapping f : M B K extends to a homomorphism

f : C
A
M B,
then M is called a K-independent (or K-free) subset of A; if K = B,
then we write B-independent rather than B-independent. If
M generates A and is A-independent, then M is called a basis of A,
and A itself is called a (relatively) free partial algebra. Moreover we
introduce the independence class of M with respect to A as
ind
A
M := B Alg(S) [ M is B-independent in A.
Similarly, if E is any class of homomorphisms, we introduce the injec-
tivity class
Inj E := A Alg(S) [ each e E is A-extendable .
December 8, 2002
5 Universal solutions, Freeness, Terms 163
Remarks 5.1.5 (i) Although the notions which we have introduced above
belong more or less to the same concept, they have developed somewhat
independently and do not show this interconnection directly. Notice,
for instance, that ind
A
M = Inj id
M
: M C
A
M .
(ii) Notice that all direct products on empty index sets (i.e. isomorphic
to (S, id
S
)) belong to Inj E and therefore also to ind
A
M for every class
E of homomorphisms. More generally we have the result in 5.1.6 below.
(iii) For those who know more about category theory let us observe that for
special subclasses K of Alg(S) forming as full subcategories so-called
epireective subcategories in particular the classes K which are
closed with respect to direct products and subalgebras form such epire-
ective subcategories the assignment A F(A, K) (A Alg(S))
yields a functor which is left adjoint to the inclusion functor from K
into Alg(S).
(iv) Above we have introduced freeness via the homomorphism extension
property; in 17.4 we add further characterizations of freeness including
unique representability, as it is well known from, say, the theory of K-
vector spaces (homogeneously specied as (V ; +, 0, , (k)
kK
)), where
each basis can be characterized as a maximal subset having the exten-
sion property for linear mappings, and as a set for which each element
of the vector space has a unique representation as a linear combination
with elements from this set.
Similar characterizations are known for the bases of groups and semi-
groups. For the characterization of bases using the unique representabil-
ity property we still need the concepts of (global) term operations
(see 5.7) or of algebraic operations (see section 15). See also 6.3 for
further information about free partial algebras.
Proposition 5.1.6 Let K Alg(S) be any subclass.
(i) Every K-extendable homomorphism is also T K-extendable, and if it is
an epimorphism, then it is even oT K-extendable.
(ii) Every K-independent subset M of some partial algebra A is also HoT K-
independent, and if M ,= , then it is even HoTo K independent.
M is B-independent for all B Alg(S) satisfying v
M
[M] , v
B
[B].
December 8, 2002
164 Chapter I Basic denitions and properties
(iii) Every injectivity class for epimorphisms is quasi-primitive, and every
independence class is primitive.
Proof (iii) is an immediate consequence of (i) and (ii), respectively.
Ad (i) Let e : B C be K-extendable, let I be any set, ( A
i
[ i I )
K
I
, (A, ( pr
i
[ i I )) a product of this family and f : B A any
homomorphism. Since A
i
K Inj e, there exist g
i
: C A
i
such
that g
i
e = pr
i
f (i I). Then ( g
i
[ i I ) induces a homomorphism
g : C A satisfying g
i
= pr
i
g (i I). Hence pr
i
f = g
i
e = pr
i
g e
(i I), i.e. f = g e, showing that A Inj e. If e is an epimorphism,
D K, A a subalgebra of D, and f : B A any homomorphism,
then id
A
f : B D implies that there exists g : C D such that
id
A
f = g e. By 3.6.2.(iii) we have g[C] A, and therefore already
g : C A, satisfying g e = f (see also 2.7.3).
Ad (ii) Because of (i) and 5.1.5.(i) we only have to show that a K-independent
set is also HK-independent: Let D K, h : D E any surjective
homomorphism, and f : M E any mapping (M a K-independent
subset of A). Since h is surjective, the Axiom of Choice yields a mapping
g

: M D such that h g

= f. Since M is D-independent in A, we
get a homomorphism g

: C
A
M D such that g

id
M
= g

: Then
h g

: C
A
M E is a homomorphism satisfying h g

id
M
= h g

= f,
i.e. extending f. If M ,= , then there does not exist a mapping into
the empty partial algebra ; therefore M is trivially -independent in
any partial algebra containing M, and a similar argument yields the
more general statement.
The last argument even yields 5.1.7, while 5.1.8 follows from 5.1.6.(i):
Corollary 5.1.7 (1) Let e : B C be any homomorphism and D Alg(S)
such that v
B
[B] v
D
[D] ,= , then D Inj e.
Corollary 5.1.8 (2) Every K-universal epimorphism is also oT K-universal.
And if (r
A,K
, F(A, K)) is a K-universal solution of A, then it is also an oT K-
universal solution of A.
December 8, 2002
5 Universal solutions, Freeness, Terms 165
Remark 5.1.9 For each class K Alg(S) and each partial S-algebra A the
K-universal solution F(A, K) is in some sense the best approximation of A
with respect to K (comparable to the approximation of an n-dimensional real
vector v by an element in the closure of a convex subset of the n-dimensional
real space R
n
): In 5.11.7 we shall show that we always have F(A, K) oT K.
Moreover, for B oT K we always have B

= F(B, K) (i.e. (r
B,K
, F(B, K))

=
(id
B
, B)), and therefore oT K = F(A, K) [ A Alg(S) , i.e. oT K is the
smallest class containing (K and) all K-universal solutions in a category
theoretical language oT K is the epireective hull of K (see 10.2.17). The close
relationship between quasi-primitive classes oT K and injectivity classes will
become obvious in 10.2.14.
In the following two subsections we shall investigate this approximation
problem for the class K := TAlg(S) of all total S-algebras, which can be
closely related to the problem of error handling in Computer Science, as
we shall observe in more detail in a later remark (see 5.3.2.(ii)). In subsec-
tion 5.11 we shall treat the general case.
5.2 Partial Peano algebras
5.2.1 So far we have only dealt with very abstract concepts, and except for
the hints in 5.1.9 it is not yet quite clear, in what instances the morphisms and
objects dealt with in 5.1 will really exist (see 5.11 for the general existence
statements and constructions). As a starting point for almost all further
constructions serve the so-called term algebras with variable set X, which
are in the terminology of 5.1 just the TAlg(S)-free TAlg(S)-algebras
F(X, TAlg(S)) with TAlg(S)-basis X. They are used for the formulation of
equational axioms (see section 6) as well as for the semantics via so-called
canonical term representations (see subsection 17.6). In order to show the
existence of term algebras even for the innitary case and to get a
better knowledge of their internal structure, we rst introduce the concept
of partial Peano algebras.
Denition 5.2.2 Let B Alg(S) be any partial algebra of similarity type
= (())

, and let X B be any subset. Then B is called a partial


Peano algebra on X (with Peano basis X), iff the following generalized Peano
axioms are valid in B:
(P1) For every and for every a dom
B
:
B
(a) / X
(i.e. X

(
B
[dom
B
] [ ) = ).
December 8, 2002
166 Chapter I Basic denitions and properties
(P2) For any ,

and for any a dom


B
, a

dom

B
one has:

B
(a) =

B
(a

) implies =

and a = a

.
(P3) (Axiom of Induction) c
B
X = B.
5.2.3 Notice that (P2) says that each partial operation is injective and that
the images of dierent partial operations are disjoint, while (P3) means that
X generates B; and because of (P1) X is obviously the smallest possible
generating subset of B.
If a partial Peano algebra is total (as an S-algebra), then it is called a
Peano algebra (or for reasons to be discussed later a word algebra, or an (S)
term algebra on X).
Examples 5.2.4 (i) A discrete partial algebra A
d
is always a partial Peano
algebra on A.
(ii) Let

: N N, n

:= n + 1 be the successor function on the set N of


natural numbers, then (N;

) is a Peano algebra of type (1) on 0, and


in this case (P1), (P2) and (P3) are nothing else than the usual Peano
axioms for the natural numbers.
(iii) The intuitive construction of partial Peano algebras of a given type runs
as follows (the heterogeneous case is similar, but observe the correct
phyla): Let X be any set; for each and for (some) b X
()
choose a new element (b) not contained in X and dierent from

(b

)
if ,=

or b ,= b

. Then do the same thing with the collection


of new and old elements as long as you wish and can , observing
that to sequences (and corresponding operation symbols) which have
already been considered in some step there will not be assigned a new
element in a later step. Then it is easy to realize that at each step this
procedure yields a partial Peano algebra on X.
More examples are obtained with the following Lemma, whose proof is a
nice example for an algebraic (i.e. structural) induction.
Lemma 5.2.5 Let B be any partial Peano algebra on X B, A any weak
subalgebra of B, and let Y := A

(
A
[dom
A
] [ ). Then A is a
partial Peano algebra on Y .
December 8, 2002
5 Universal solutions, Freeness, Terms 167
Proof (P1) is satised because of the denition of Y ; and (P2) holds in A,
because it is true in B. Thus it remains to prove (P3), and this is done by
proving by algebraic induction for x B the statement
H(x) := x A implies x c
A
Y :
Assume x X; if x A, then x Y by (P1) for B and the denition of Y .
Thus, assume x B, x =
B
(b) for some and b dom
B
, and let
the statement H(b(k)) be true for every k (). Moreover, let us assume
that x A (otherwise H(x) is already true). Then either x Y yielding
x c
A
Y , or there exist

and b

dom

A
such that x =

A
(b

) (by
the denition of Y ). Since graph

A
graph

B
, and since B satises (P2),
we get

= and b

= b. Thus b(k) A for each k (). Therefore, by


the induction hypothesis one has b(k) c
A
Y (k ()). But this implies
x =
A
(b) c
A
Y , since c
A
Y is closed in A. This proves A c
A
Y .
The following results will positively solve the question of existence of total
(S-) Peano algebras in a constructive way.
Theorem 5.2.6 For every type , every sort-algebra S of type and every
S-set X there exists a total (S-) Peano algebra P(X, S) with Peano basis X.
(The proof contains a construction).
Proof (a) We rst consider the homogeneous case, using the construction
of R.Kerkhoff in [Ke65]: If X = and if contains no nullary
constants, then set P(, S) := ; this is obviously a Peano algebra over
X. Else dene
() Z := (0 X) (1 )
_
( () [ ),
the disjointness of the images of X and will later become important
and let
() Z

:= (z
1
, z
2
, . . . , z
n
) [ n N, n 1, z
1
, z
2
, . . . , z
n
Z ,
be the set of all non-empty nite sequences of elements of Z. Moreover,
set G := Subs Z

, the set of all subsets of Z

, and introduce on G a total


algebraic structure as follows (for and ( M
k
[ k () ) G
()
):
()

G
( M
k
[ k () ) :=
((1, )) (k, z
1
, . . . , z
n
) [ k (), (z
1
, . . . , z
n
) M
k

December 8, 2002
168 Chapter I Basic denitions and properties
notice that
G
= ((1, )), if () = 0. We have a natural injection
j : X G by x ((0, x)), and dene W as the subalgebra of G
generated by j[X]. Observe that there is an algebra W

containing X
and isomorphic to W over a homomorphic extension of j (see 23.2.8
in the Appendix), and it is easy to realize, that W

is a Peano algebra
over X once we have shown that W is a Peano algebra over j[X]:
(P3) is satised by the denition of W.
Ad (P1) If
W
( M
k
[ k () ) j[X], then
W
( M
k
[ k () ) =
((0, x)) for some x X. By ()
W
( M
k
[ k () ) contains
exactly one sequence of length 1, namely ((1, )), which is dierent
from ((0, x)). Thus
W
( M
k
[ k () ) / j[X], a contradiction.
Ad (P2) Assume that
W
( M
k
[ k () ) =

W
( N
l
[ l ()

).
Since each element of W contains only one sequence of length 1, and
since this is ((1, )) respectively ((1,

)) above, we may conclude


=

and () = ()

. Now, let k () and (z


1
, . . . , z
n
) M
k
,
then (k, z
1
, . . . , z
n
)

W
( N
l
[ l ()

), i.e. (z
1
, . . . , z
n
) N
k
showing that M
k
N
k
(k ()). Symmetry of the situation now
yields equality everywhere, which was to be shown.
(b) We now consider the heterogeneous situation and present a more in-
tuitive description of a total (S) Peano algebra connected with the
above construction: Since each fundamental operation is dened in S
on at most and usually exactly one sequence, it is associated with
at most (exactly) one output sort. For given X = (X, v
X
: X S)
we have already associated an element of S with each element of X. If
w W

X, then w =
W

( w
k
[ k () ) and we set v
W
(w) := s

.
Thus we get v
W
: W

S, and we redene the structure on W

as dom

:= a (W

)
()
[ v
W
(a(k)) = s
k
for k () and

:=
W

[
dom
.
Let Y := (W

, (

); then Y is a weak relative subalgebra of W

,
and (C
Y
X, v
W
[
C
Y
X
) is a partial Peano algebra and a total S-algebra
with Peano basis X.
Remarks 5.2.7 (i) In the nitary case the proof is usually given by con-
sidering the set W of all empty or non-empty words, i.e. nite se-
quences from elements in X

(disjoint union!) written as words:
w = (z
1
, . . . , z
n
) =: z
1
. . . z
n
(z
1
, . . . , z
n
X

), such that the struc-
ture is dened by
W
(w
1
, . . . , w
()
) := w
1
. . . w
()
, concatenation of
December 8, 2002
5 Universal solutions, Freeness, Terms 169
words ( , (w
1
, . . . , w
()
W
()
).
The desired Peano algebra is then obtained as P := C
W
X (the one-
element word x is identied with the corresponding element x from X
and vice versa). In the heterogeneous case one then chooses the appro-
priate weak subalgebra as in the proof (part (b)) above.
(ii) Another proof of the existence of total (S)Peano algebras will follow
from the Generalized Recursion Theorem in subsection 5.4.
One of the most important features of Peano algebras is the following
Theorem 5.2.8 (Recursion Theorem) Let P be any partial Peano alge-
bra on X, C TAlg(S) any total algebra and f : X C any mapping. Then
there exists a unique homomorphic extension f

: P C of f, i.e. f

[
X
= f.
Proof Let f

:= c
PC
graph f. When we have shown that f

= graph f

is
the graph of a mapping f

, then 4.1.10.(ii) (which is easily extended to partial


S-algebras) implies that f

will be a homomorphism (between partial S-


algebras), and from 4.1.12 and (P3) we may infer that domf

= P. Therefore
we prove by algebraic induction on p P that H(p) := There exists exactly
one c C such that (p, c) f

is true:
We rst assume p X: Then (p, f(p)) f

; if also (p, c) f

and c ,= f(p),
then this can only happen, when there are and (p
k
, c
k
) f

(k ())
such that
P
( p
k
[ k () ) = p X, and this is impossible by (P1). Hence
we have proved the basis of induction.
Assume now p P X, p =
P
( p
k
[ k () ), and let H(p
k
) be true for
each k (). Moreover, let, for each k (), c
k
be the unique element
from C such that (p
k
, c
k
) f

. Since by (P2) p can be represented


as a value of a fundamental operation in exactly one way, and since each
p
k
satises H, p satises H, too, i.e. c :=
C
( c
k
[ k () ) is the unique
element of C satisfying (p, c) f

. Hence the principle of algebraic induction


yields that f

is the graph of a mapping.


Consequences of this result, e.g. the fact that P(X, S) is canonically iso-
morphic to F(X, TAlg(S)), will be discussed later in a more general context.
December 8, 2002
170 Chapter I Basic denitions and properties
5.3 Free completions
In this subsection we describe and discuss TAlg(S)-universal TAlg(S)-solutions
of partial algebras A. The corresponding object F(A, TAlg(S)) for A will be
shown later to be canonically isomorphic to the so-called (absolutely) free
completions of A (see 5.3.8).
Denition 5.3.1 Let A Alg(S) be any partial algebra. A total algebra
B TAlg(S) is called an (absolutely) free completion of A iff the following
axioms hold:
(FC0) A B and for each : graph
A
graph
B
.
(FC1) For each and for each b B
()
one has:

B
(b) A implies
B
(b) =
A
(b)
(thus in particular b dom
A
( A
()
)).
(FC2) For all ,

and for all b B


()
, b

B
()

one has:

B
(b) =

B
(b

) / A implies =

and b = b

.
(FC3) c
B
A = B (i.e. the completion B of A is minimal).
Remarks 5.3.2 (i) The axioms (FC0) through (FC3) are closely related
to the generalized Peano axioms in 5.2.2: (FC0) says that A is a weak
subalgebra of B, while (FC1) implies that it is even a special
relative subalgebra of B relative subalgebras satisfying (FC1) will
be called normal relative subalgebras according to J.Schmidt [Sch70].
(FC1) means that no value of a fundamental operation which lies in A
can come from the outside (cf. (P1)). (FC2) says that outside of A
(P2) is satised, while (FC3) the axiom of induction corresponds
to (P3). Altogether we get that outside of A a free completion
like B behaves like a partial Peano algebra.
(ii) In connection with the algebraic specication of abstract data types ab-
solutely free completions can be considered as an important tool for the
handling of error values (see [GoTcWa78]). Let a partial algebra A
be the semantics of an algebraic specication, and let B be a free com-
pletion of A. Then, whenever something is not dened in A, B contains
December 8, 2002
5 Universal solutions, Freeness, Terms 171
a special value (error value) for this situation at least if the argu-
ment sequence has the appropriate length or size ; and whenever
a fundamental operation uses an error value, then again there will be a
special value (message) for this situation. Thus an error situation can
be traced down to its exact origins (the latest normal arguments) in
each conceivable error situation which can successively be described
by applications of fundamental operations to sequences of the correct
form however with arguments which were not allowed. It should
also be observed that under the above restrictions the homoge-
neous free completion (even of a heterogeneous partial algebra A) will
be useful in order to handle the cases of type mismatch. As already
observed earlier, those errors stemming from argument sequences of
incorrect length will not have their corresponding values in the free
completion. However, except for that fact, each total error algebra
of A generated by A and containing A as a relative subalgebra, can be
considered as a (special) homomorphic image of a free completion of A
(special, since the restriction to A of the natural homomorphism has
to be (full and) injective).
Corollary 5.3.3 B TAlg(S) is a total (S) Peano algebra, iff B is the
free completion of a discrete partial algebra.
The existence of free completions also follows from the one of Peano al-
gebras:
Theorem 5.3.4 For every partial algebra A there exists a free completion B.
Proof We shall actually indicate two methods of proof, both providing
eective constructions relative to the one of Peano algebras.
(i) Let P := P(A, S) be a total Peano algebra on A. Dene a relation

on P as follows: (p, q)

, iff there is b dom


A
such that p =
A
(b)
and q =
P
(b). Let := Con
P

be the congruence of P generated


by

. Then P/ satises (FC0) through (FC3) with respect to its


relative subalgebra on nat
[A]
, which is isomorphic to A over nat

[
A
.
More details and an intrinsic description of can be found in [G68a].
(ii) Let W be a total Peano algebra on A, whose existence follows from 5.2.6,
and dene for and w W
()
:

#
(w) :=
_

A
(w), if w dom
A
,

W
(w), else.
December 8, 2002
172 Chapter I Basic denitions and properties
Let W
#
:= ((W, (
#
)

), v
W
) and B := C
W
#A. Then (FC0), (FC1),
(FC2) and (FC3) follow directly from the denition of B and the fact
that W satises the Peano axioms (P1), (P2) and (P3).
We are now able to present a complete characterization of partial Peano al-
gebras:
Corollary 5.3.5 A is a partial Peano algebra, iff each (and it suces to
say: some) free completion of A is a total Peano algebra.
Proof It has been proved in 5.2.5 that weak relative subalgebras of Peano al-
gebras are always partial Peano algebras. On the other hand it is easy to
check that any free completion of a partial Peano algebra satises the axioms
(P1), (P2) and (P3).
Remark 5.3.6 The rst version of the proof of 5.3.4 yields the more familiar
construction, while the second one can be interpreted in the way that from
each congruence class in the rst construction a canonical representative is
chosen. Thus we get by (ii) sort of a representation by canonical terms in
the sense of [GoTcWa78] (cf. also section 17.6), which sometimes gives easier
and better insight into the structure of the free completion than the factor
algebra.
Theorem 5.3.7 (Recursion Theorem for Free Completions) Let A be
any partial algebra, B a free completion of A, and C any total algebra sim-
ilar to A. Then there exists for every homomorphism f : A C a unique
homomorphism g : B C extending f (i.e. g[
A
= f).
Proof Since B is a free completion of A, axioms (FC2) and FC3) tell us that
B

:= ((B, (

), v
B
) is a partial Peano algebra on A, when we dene

to
be the restriction of
B
to B
()
dom
A
. Hence the Recursion Theorem 5.2.8
yields a homomorphic extension f

: B

C extending f. Since f

[
A
= f is
a homomorphism from A into C by assumption, f

is compatible with all of


the structure of B, i.e. f

: B C is a homomorphism.
December 8, 2002
5 Universal solutions, Freeness, Terms 173
Corollary 5.3.8 Let A Alg(S) be any partial algebra. Then A possesses
up to unique isomorphism exactly one free completion, say B, and (id
A
, B)
is the TAlg(S)-universal TAlg(S)-solution of A, i.e. B

= F(A, TAlg(S))
over id
A
.
Proof From 5.3.7 we infer that id
A
: A B is a TAlg(S)-extendable
epimorphism for every free completion B of A; therefore it is a TAlg(S)-
universal TAlg(S)-solution according to 5.1.3.(ii), while 5.1.3.(i) yields the
uniqueness of the free completion up to unique isomorphism.
5.4 The Generalized Recursion Theorem
In connection with the recursion theorems proved so far we get a very useful
tool, which will become the basis of a model theoretic approach to the theory
of partial algebras.
Theorem 5.4.1 (Generalized Recursion Theorem) Let f : A B be
any homomorphism between similar partial algebras A and B. Then there
exists a closed homomorphic extension (f

)
B
:= f : domf

B of f
such that domf

is an A-generated relative subalgebra of the free completion


F(A, TAlg(S)) of A; moreover, we have:
(i) graph f

= c
F(A,TAlg(S))B
graph f.
(ii) f

is the largest homomorphic extension of f to an A-generated relative


subalgebra (what we shall call an A-initial extension) of F(A, TAlg(S)).
(iii) Let

f : F(A, TAlg(S)) F(B, TAlg(S)) be the homomorphic extension
of f which exists according to 5.3.7. Then domf

=

f
1
[B] and f

f[
domf

.
Proof According to the Recursion Theorem 5.3.7 there exists a homomor-
phic extension

f : F(A, TAlg(S)) F(B, TAlg(S)). Moreover, A and B
are relative subalgebras of F(A, TAlg(S)) and F(B, TAlg(S)), respectively.
Thus according to Lemma 4.1.13 graph f

:= c
F(A,TAlg(S))B
graph f is the
graph of a homomorphic extension of f to an A-generated relative subalgebra
domf

of FU(A, TAlg(S)); and since F(A, TAlg(S)) is total, f

: domf


B is closed. And 4.1.13.(iii) implies that f

is the largest homomorphic


extension of f to an A-generated relative subalgebra of F(A, TAlg(S)) with
respect to B. Thus it remains only to show (iii):
December 8, 2002
174 Chapter I Basic denitions and properties
Obviously domf



f
1
[B]. Therefore because of (ii) we only have to
show that A generates

f
1
[B]. This is done by algebraic induction on q
F(A, TAlg(S)) with respect to the property H(q) := q

f
1
[B] implies
q c

f
1
[B]
A:
H(q) is true for all q A. Let q

f
1
[B], q / A. Then q =
F(A,TAlg(S))
( c
k
[
k () ). Let H(c
k
) be true for all k (). Since

f(q) B, we
have

f(q) =
F(B,TAlg(S))
(

f(c
k
) [ k () ) =
B
(

f(c
k
) [ k () ), as
F(B, TAlg(S)) satises (FC1). Therefore

f(c
k
) B, i.e. c
k


f
1
[B], for
each k (). Because of H(c
k
) we have c
k
c
f
1
[B]
A, and therefore also
q =
F(A,TAlg(S))
( c
k
[ k () ) c
f
1
[B]
A.
Remark and Notation 5.4.2 The Generalized Recursion Theorem yields
the foundation for a model theory of partial algebras, as we shall see especially
in Chapter II. Therefore the closed initial homomorphic extension of 5.4.1
will occur quite often. Hence it seems natural that it gets a short notation.
We shall preserve the notation f

in connection with homomorphisms f :


A B only for this closed A-initial homomorphic extension of f described
in Theorem 5.4.1 and if it might not be clear what target algebra is used,
we shall write (f

)
B
to indicate that B is the target the starting algebra
should always be clear, else we write (f

)
A,B
. If f is composed, f = h g,
then we shall write f

= (h g)

for the largest closed initial extension.


Remark 5.4.3 Theorem 5.4.1 provides us with another possibility for de-
scribing F(M, TAlg(S)) with respect to F(M, TAlg()) and the set M =
(M, v
M
) of variables, namely:
F(M, TAlg(S)) = domv

M
.
The straightforward proof is left to the reader.
5.5 The algebraic quasi-ordering on partial algebras
Denition 5.5.1 Let A Alg(S) be any partial algebra. Then one can
dene on A relations
A
and _
A
as follows:

A
:= (a, a

) A A [ there are , k
0
() and b dom
A
such that a

=
A
(b) and a = b(k
0
) .
December 8, 2002
5 Universal solutions, Freeness, Terms 175
If a
A
a

, then we say that a is an immediate predecessor of a

. The relation
_
A
is the reexive and transitive closure of
A
. Thus, by denition, _
A
is
a quasi-order on A which is called the algebraic quasi-order on A and if it
is a partial order, then it is called the algebraic partial order on A induced
by A (on A). (Note that the relationship of A to S does not have any
direct inuence on this quasi-order, it only restricts the structure of A.)
If M A, we denote by M the initial segment in A with respect to the
above quasi-order, i.e.
M := x A [ x _
A
m for some m M ;
in particular we denote m also by m.
M designates the relative subalgebra of A with carrier set M. M is
often called the set of all predecessors of elements of M in A.
If A = F(M, TAlg(S)) is a Peano algebra, then the elements of A are also
called terms, and for any term t A the elements of t are called subterms
of t.
Proposition 5.5.2 Let A be a partial Peano algebra on some set X. Then
one has in connection with the rank function
A
(, X):
a
A
b implies
A
(a, X) <
A
(b, X) for all a, b A
(notice that a
A
b iff a _
A
b and a ,= b, as usual). Therefore _
A
is then a
partial order on A.
Proof Let A be a partial Peano algebra on some set X, and let a
A
b for
some a, b A. Then b =
A
(a) and a = a(k) for some , a dom
A
and k (). Hence, by (P1) b / X, i.e.
A
(b, X) > 0. Then (P3),
Theorem 3.4.4 and observation 3.4.6 imply that there exist

and
c dom

A
such that
A
(c(l), X) <
A
(b, X) for every l ()

, and that
b =

A
(c). (P2) then immediately implies =

and a = c, i.e. a = c(k)


satises
A
(a, X) <
A
(b, X). A transitivity argument shows the general
implication, which immediately implies that a
A
b and b
A
a cannot hold
at the same time. Hence
A
is anti-symmetric.
Corollary 5.5.3 (1) In a partial Peano algebra B _
B
only allows nite
strictly descending chains.
Proof Notice that the step numbers are ordinals, which have this property.
December 8, 2002
176 Chapter I Basic denitions and properties
Corollary 5.5.4 (2) Let A be any partial algebra, B a relative subalge-
bra of F(A, TAlg(S)), and A B. Then A generates B, iff B = B
(w.r.t. _
F(A,TAlg(S))
).
Similarly, if B is any relative subalgebra of a partial Peano algebra P with
Peano basis A, and A B, then c
B
A = B, IFF B = B.
Proof Since F(A, TAlg(S)) becomes a partial Peano algebra, when we for-
get about the structure of A in it, we have only to prove the second part:
Assume that there is an element c B B, then there is b B which has
predecessors in P which do not belong to B. Because of 5.5.3 we may even
assume that b has at least one immediate predecessor of this kind. Thus b
would belong to every generating subset of B (cf. 5.2.5), but it cannot belong
to A (whose elements do not have immediate predecessors in P at all), i.e.
A cannot generate B.
Conversely, if B = B, then one can prove by algebraic induction (exercise!)
that A generates B (cf. the proof of 5.2.5).
Lemma 5.5.5 (i) Let B be any partial Peano algebra on M, t B. t is
a relative subalgebra of B generated by M
t
:= t M; and for any a
t t / c
D
(M D), where D := t a. The set M
t
:= t M is often
denoted by var t, the set of variables occuring in t.
(ii) Let w : M A be any mapping, and t F(M, TAlg(S)). Then
t domw

iff t domw

; moreover, w

(t) only depends on w[


Mt
.
Proof Ad (i) The rst part easily follows from 5.5.4 and the fact that al-
ready t F(M
t
, TAlg(S)), while the second one can be proved by
algebraic induction on t using the fact that for t =
F(M,TAlg(S))
( t
k
[ k
() ) one has t = t

t
k
[ k () .
Ad (ii) From 5.5.4 respectively from 5.5.5.(i) we infer that t is the smallest
relative subalgebra of F(M, TAlg(S)) which is generated by a subset
of M (namely M
t
) and contains t. Since w

is the largest M-initial


homomorphic extension of w (cf. 5.4.1 and 5.5.4), the rst statement
follows, while the second one follows from the fact that M
t
(is the small-
est subset of M which) generates t.
December 8, 2002
5 Universal solutions, Freeness, Terms 177
Remark 5.5.6 It should be observed that, for M N and t F(M, TAlg(S)),
t can also be considered as an element of F(N, TAlg(S)), but while there may
be many mappings w : M A for which t domw

, there may be no map-


ping from N into A with this property, namely when there is no mapping at
all from N into A; and this happens iff v
A
[A] , v
N
[N]. (In the homogeneous
case it can only happen that M but not N allows a mapping into A when
M and A both are empty, but then there is no element from F(M, TAlg()),
which can be mapped into A.) Thus it will be necessary at least for the
heterogeneous case to x to each term t F(M, TAlg(S)) the Peano basis
with respect to which it is considered, i.e. (M : t) will be the correct nota-
tion for a heterogeneous term (notice that in Part II (by H.Reichel [Re87])
terms are written as (v : t) rather than (M : t), where v means our v
M
).
5.6 Tree representation of terms
5.6.1 In this subsection we treat a representation of terms (i.e. elements of
a Peano algebra) by rooted trees, which has many applications in Computer
Science and is a useful tool. For a better understanding the reader should be
familiar with the basic terminology of graph theory (cf. e.g. F.Harary [Hy72],
Chapters 4 and 16). Since terms belong to the most fundamental algebraic
and logical concepts, it is important to get a good understanding of
them.
5.6.2 We shall consider rooted (vertex-) labeled directed trees G := ((V, E, r), l),
where V is the set of points or vertices, E the set of directed lines or directed
edges, which are ordered pairs of elements of V , r V is the root and
l : V X (X any set, the set of labels) is a labeling of the vertices with
elements of X such that (V, E) is a directed (out-) tree with root r (i.e. all
edges are directed away from r, and even when we forget about directions
each vertex can be reached from r in exactly one (nite) path). Two la-
beled rooted trees ((V, E, r), l) and ((V

, E

, r

), l

) are isomorphic, iff there is


a bijection f : V V

such that f(r) = r

, (v, v

) E iff (f(v), f(v

)) E

and l

f = l. If X is any set and c any cardinal, then we denote by Tr(X, c)


the set of all isomorphism types (e.g. a set containing exactly one element
from each isomorphism class) of rooted trees labeled with elements from X,
and where the vertex set has cardinality less than c.
5.6.3 Let = (())

be any similarity type of innite cardinal dimension


d

, and let M be any set, which we assume w.l.o.g. to be disjoint from .


December 8, 2002
178 Chapter I Basic denitions and properties
Figure 5.3: This should be Figure 5.3
Then we supply the set Tr(M , d

) with an algebraic structure as follows


(we only treat the homogeneous case, in the heterogeneous case one has to
observe the sort mapping of M and the output sorts of the fundamental
operations as in part (b) of the proof of Theorem 5.2.6):
Let , and
t := (((V
k
, E
k
, r
k
), l
k
: V
k
M ) [ k ()) (Tr(M , d

))
()
be any ()-sequence of rooted trees. Then

Tr(M,d

)
(t) := ((V, E, r), l),
where
V := (, )

( V
k
k [ k () )
E := ((, ), (r
k
, k)) [ k ()

( ((a, k), (b, k)) [ (a, b) E


k
[ k () )
r := (, )
l : V X : (, ) , (a, k) l
k
(a), (a V
k
, k ()).
An example for := (2, 1, 0), := ,

, M := m, m

, m

can be seen
in Figure 5.3, where we have drawn the trees as usual in graph theory and
labeled the vertices directly with their labels.
Notice that we have already chosen such trees as arguments which shall
later on represent terms (i.e. elements of F(M, TAlg()); moreover, it should
be observed that in the diagrams we have attached only the labels to the
vertices, not the names of the vertices, and that by reading such a tree one
starts at the left hand leaf (end vertex, vertex with out-degree 0) which also
is the left hand leaf of the rst argument, and then proceedes successively to
the right (see 5.6.4).
5.6.4 Let

M := G
m
Tr(M, d

) [ m M , where G
m
:= ((m, , m), l
m
:=
id
{m}
). Then it is not too dicult to realize that j : M

M, m G
m
(m M) induces an isomorphism

j : F(M, TAlg()) C
Tr(M,d

)

M =:
Trees(M, ).
December 8, 2002
5 Universal solutions, Freeness, Terms 179
For instance the trees in Figure 5.3 have the following preimages (repre-
sented as words in (X )

, i.e. by inverse Polish parenthesis-free nota-


tion):

m, mm

, and

mmm

.
This might give an idea of how to read the trees as elements of Trees(M, ) (

=
F(M, TAlg())). Via the isomorphism above one may now also get the
trees corresponding to terms in F(M, TAlg(S)), since this is actually an M-
generated relative subalgebra of F(M, TAlg()) (see e.g. 5.7.3 below for a
further example).
5.6.5 (i) Notice that one can easily prove, say by algebraic induction, that
for every tree G Trees(M, ) every path is nite, i.e. every maximal
directed path starting from the root ends with a leaf, and the leaves
are labeled with elements from M
0
(
0
again being the set of all
fundamental nullary operation symbols), while all the other vertices
are labeled by elements from
0
.
(ii) It is also easily realized that for a nitary similarity type the value

Trees(M,)
(G,

M) of the rank function on Trees(M, ) for G Trees(M, )
computes either as the length of the longest path in G from the root to
a leaf, if for no such longest path the leaf is labeled by a fundamental
constant, or one has to add 1.
(iii) Let G = ((V, E, r), l) Trees(M, ) be any rooted tree (identied with
the corresponding isomorphism class), v V any vertex of G.
Let G
v
:= ((V
v
, E
v
, v), l[
Vv
) be given by
and
V
v
:= w V [ there is a directed path in G from v to w,
E
v
:= (a, b) E [ a, b V
v
.
Then G
v
is also a rooted tree belonging to Trees(M, ), and with the
isomorphism

j from 5.6.4 one obviously has for any t F(M, TAlg(S))
that

j[t] = G
v
[ v V (G) , where G =

j[t].
5.7 (Global) Term operations and polynomials
One very important consequence of the Generalized Recursion Theorem 5.4.1
is the fact that it allows us to dene term operations and polynomials also
for partial algebras (observe 5.5.6).
December 8, 2002
180 Chapter I Basic denitions and properties
Denition 5.7.1 Let M be any set, A Alg(S) any partial algebra, and
t F(M, TAlg(S)) any term. Then t denes on A a partial M-ary term
operation (M : t)
A
as follows:
dom(M : t)
A
:= w A
M
[ t domw

,
and
w dom(M : t)
A
implies (M : t)
A
(w) := w

(t).
(M : t)
A
is called the (global) term operation on A corresponding to the
term t or more precisely to (M : t). In this connection M is called the
set of variables under consideration, while the mappings w : M A are
sometimes called valuations of M in A or assignments of values in A to the
variables in M. Moreover, t (or (M : t)) is said to be evaluable in A with
respect to the assignment w : M A iff t domw

(iff w dom(M : t)
A
).
For every t

t, (M : t

) is called an M-subterm of t (or (M : t)).


Remarks 5.7.2 (i) Notice that in the homogeneous case it is not neces-
sary to mention the set of variables in connection with a term, while
we have already mentioned in 5.5.6, why this is necessary in the het-
erogeneous case.
(ii) The tree model is probably the most intuitive representation of a term.
In this connection a valuation in A corresponds to the process that
each label of a leaf of such a tree is replaced by an admissible element
of A, and each other label, which then is an operation symbol, say ,
is replaced by the corresponding fundamental operation
A
of A. The
operations are successively evaluated as long as possible, proceeding
from the leaves to the root. Whenever we reach a vertex such that the
corresponding operation is not evaluable on the values of its successors
or when it is a leaf, if there is no admissible element for it in A ,
then the whole term is not evaluable. Hence a term (M : t) is evalu-
able, iff each M-subterm of (M : t) is evaluable. Notice that the term
operations are obtained by repeated superposition of fundamental
operations, starting with the so-called projection operations (cf. sec-
tion 15 for more details), since for each variable m M the mapping
w w(m) = w

(m) = w(M : m)
A
(w A
M
) is a total M-ary oper-
ation on A, called the m-th M-ary projection and denoted by (e
M
m
)
A
,
i.e. (e
M
m
)
A
: A
M
A, (e
M
m
)
A
(w) := w

(m) in most cases one writes


December 8, 2002
5 Universal solutions, Freeness, Terms 181
gures disabled
Figure 5.4:
gures disabled
Figure 5.5:
e
M
m
instead of (e
M
m
)
A
, when the context yields the corresponding partial
algebra A. Notice that (e
M
m
)
A
is independent from any structure on A,
but it depends on v
A
and v
M
, since for v
M
[M] , v
A
[A] it is empty.
Example 5.7.3 In order to get a better feeling for the evaluation of terms
we consider a type = (2, 1, 0) for = (,

) and M := m, m

, m

as
in 5.6.3, and we write terms as words, as in 5.6.4, and rewrite it with brackets
for better structuring; moreover we represent it as a tree (cf. Figure 5.4).
Let the term t (rather (M : t), but we are in a homogeneous case) be
given as follows:
t := mm


= (((m, (m

, (m

)))

).
For the evaluations of t we consider the partial algebra A sketched in Fig-
ure 5.5, where we assume
A
to be commutative. We evaluate t e.g. for w
and w

, where
graph w := (m, b), (m

, a), (m

, c) and graph w

:= (m, a), (m

, b), (m

, c)
by replacing the labels of the tree successively by its evaluations, starting
with the leaves (cf. Figures 5.6 and 5.7).
In connection with this example it is relatively easy to determine domt
A
,
and we suggest it as an exercise, giving only some hints:
Starting with the root one realizes that here the value exists iff the out-
put of its successor labeled with

is b (since always
A
= c). Observing the
structure of Aone thus realizes that domt
A
= (b, a, a), (b, a, c), (b, c, a), (b, c, c),
and t
A
(w) = a for every w domt
A
.
The following results are of great importance in the sequel:
December 8, 2002
182 Chapter I Basic denitions and properties
gures disabled
Figure 5.6: t
A
(w) = a
gures disabled
Figure 5.7: t
A
(w

) does not exist


Lemma 5.7.4 Let M be any set, f : A B any homomorphism, and
w : M A any valuation. Then domw

dom(f w)

equality holding
when f is closed , and ker w

ker (f w)

.
Proof Obviously f w

is a homomorphic extension of f w to an M-
generated relative subalgebra of F(M, TAlg(S)). Since (f w)

is the largest
such extension, we get domw

dom(f w)

, and if f is closed, then f w

is already a closed such extension, and equality has to hold (cf. 5.4.1.(ii)).
Since f w

= (f w)

[
domw

, ker w

ker (f w

) ker (f w)

.
Corollary 5.7.5 Each homomorphism is compatible with each induced term
operation, i.e.:
If f : A B, t F(M, TAlg(S)), w : M A and a A, then
(M : t)
A
(w) = a implies (M : t)
B
(f w) = f(a)
i.e. if (M : t)
A
(w) exists , then so does (M : t)
B
(f w) and one has f((M :
t)
A
(w)) = (M : t)
B
(f w).
Proof With the notation of the Corollary and the assumption that (M :
t)
A
(w) exists one has t domw

dom(f w)

; therefore (f w)

(t) exists
and satises because of t domw

= dom(f w

) : (M : t)
B
(f w) =
(f w)

(t) = (f w

)(t) = f((M : t)
A
(w)) (= f(a)).
Remark 5.7.6 Corollary 5.7.5 may form the basis for the possibility of ex-
tending rational equivalence to partial algebras, but we do not go into details
in this book, since as far as we know this concept has not yet been inves-
tigated for partial algebras.
Proposition 5.7.7 Let A be any partial algebra of cardinal dimension d and
let M be a generating subset of A. Then:
December 8, 2002
5 Universal solutions, Freeness, Terms 183
(i) For every a A there exists a term t F(M, TAlg(S)) such that
a = (M : t)
A
(id
M
).
(ii) For every a A there exists a subset N of M satisfying #N <
max2, d with a term t F(N, TAlg(S)) such that a = (N : t)
A
(id
N
).
Proof Let i := id
M
: M A, then i

is surjective, since M generates A


(see 5.8.1 below). Thus i
1
(a) ,= ; let t i
1
(a), then i

(t) = a =
(M : t)
A
(i). Since A = c
A
M =

( c
A
N [ N M, #N < max2, d )
(cf. 3.3.2.()), the second statement follows in a similar way.
Closely connected with the concept of global term operations is the one
of global polynomials with coecients in some partial algebra A and some
set X of variables:
Denition and Remark 5.7.8 Let A Alg(S) be any partial algebra and X
any set (of variables), and we may assume w.l.o.g. that A X = . Then
A
X
:= A + X
d
is a partial algebra, which is the coproduct in Alg(S) of A
and the discrete partial algebra on X. F(A
X
, TAlg(S)) =: F(AX, TAlg(S))
is called the algebra of global partial polynomials with coecients in A, and
each p F(A
X
, TAlg(S)) is called a (global partial) polynomial on A, since
by the Generalized Recursion Theorem 5.4.1 we have:
Let f : A B be any given homomorphism (e.g. B = A and f = id
A
),
then we may dene for p F(A
X
, TAlg(S)) and for w : X B: domp
B,f
:=
w B
X
[ p dom(f w)

f w meaning the mapping having as


graph the disjoint union of the graphs of f and w , and if w domp
B,f
,
then p
B,f
(w) := (f w)

(p).
If B is a total homogeneous algebra, then this concept generalizes the one
of polynomials in rings.
Because of the close relationship between global polynomials and global
term operations as far as the technicalities are concerned , we do not
present here more details on polynomials (but cf. section15).
They have been discussed more thoroughly in the thesis of M.Siegmund-
Schultze [Si84]. When dealing with them one always has to recall the
Generalized Recursion Theorem 5.4.1 and its main consequences.
December 8, 2002
184 Chapter I Basic denitions and properties
5.8 The Epimorphism Theorem
Another very important tool provided by the Generalized Recursion Theorem
will be discussed in this subsection. It is an analogue to the Homomorphism
Theorem which deals with quotient mappings, while this new one concerns
epimorphisms. As a preparation we present the following characterization of
epimorphisms:
Lemma 5.8.1 The homomorphism f : A B is an epimorphism iff its
greatest A-initial extension f

in F(A, TAlg(S)) is surjective.


Proof Let f : A B be an epimorphism, i.e. f[c
B
A] = B. Then f

:
domf

=: C B being closed satises f

[C] = f

[c
C
A] = f

[c
B
A] =
f[c
B
A] = B (cf. 3.6.2.(ii)), i.e. f

is surjective.
If, conversely, f

: C B is surjective, then the same calculation as above


yields that f is dense, hence an epimorphism.
The following result is due to J.Schmidt [Sch70].
Lemma 5.8.2 (Diagram Completion for Epimorphisms) Let f : A
B be an epimorphism, g : A C any homomorphism.
(a) Then the following statements are equivalent (see Figure 5.8 of the
proof ):
(i) There exists a unique homomorphism h : B C such that h f =
g.
(ii) ker f

ker g

(within F(A, TAlg(S))).


(b) If h : B C satisfying h f = g exists, then:
(iii) h is an epimorphism, iff g is an epimorphism.
(iv) h is surjective, iff g

[
domf

is surjective (hence especially, when g


is surjective).
(v) h is injective (i.e. a monomorphism), iff ker f

= ker g

(domf

)
2
.
(vi) h is closed, iff domf

= domg

.
(vii) h is closed and injective, iff ker f

= ker g

.
(viii) h is an isomorphism, iff g is an epimorphism and ker f

= ker g

.
December 8, 2002
5 Universal solutions, Freeness, Terms 185
gures disabled
Figure 5.8:
(ix) h is full, iff g

[
domf

is full.
Proof Ad (a) Let us consider the diagram in Figure 5.8, and let us assume
that h : B C exists such that h f = g. Then the diagram is
commutative in all fully drawn parts (i
1
, i
2
being identity injections).
Since h exists, we have g = h f = h f

i
1
, i.e. h f

is an A-initial
extension of g. Since g

is the largest such extension, we have D :=


domf

E := domg

, therefore the embedding i := id


D
: D
E exists and satises i
2
= i i
1
. Hence g

i = g

[
D
= h f

, and
therefore ker g

ker (g

[
D
) = ker (h f

) ker f

. Conversely, if
ker f

ker g

, then domf

domg

, i.e. i above exists, i : D E.


f

is full (even closed) and surjective (cf. 5.8.1), and we can apply 2.7.2
getting a unique homomorphism h : B C such that h f

= g

i.
Then h f = h f

i
1
= g

i i
1
= g

i
2
= g. Thus the equivalence
in (a) is proved.
Ad (b) Let h : B C exist such that h f = g. (iii) follows, since f is an
epimorphism: v h f = w h f implies v h = w h, and hence v = w
, if h is an epimorphism; if v h = w h, then v h f = w h f; thus
v = w, if g = h f is an epimorphism.
It is not too dicult to derive facts (iv) through (ix) from the Diagram
Completion Lemma 2.7.1 and its Corollary 2.7.2, since f

is closed
(hence full) and surjective, occasionally using 5.8.1; therefore the proofs
are left as exercises. Finally, (ix) is a consequence of 2.4.5.(iii).(a)
and (c), since f

is full and surjective.


As a corollary we get what J.Schmidt has called the General Homo-
morphism Theorem, and what we call the
Theorem 5.8.3 (Epimorphism Theorem) Let f : A B and g : A
C be any two epimorphisms. Then the following statements are equivalent:
(i) ker f

= ker g

.
(ii) There exists an isomorphism h : B C such that h f = g.
December 8, 2002
186 Chapter I Basic denitions and properties
The following result (cf. [BJPa78]) will become important e.g. in sec-
tion 10:
Corollary 5.8.4 Let f : A B be any homomorphism, E := domf

, and
e : A E the natural injection.
(i) Then f is the composition of the TAlg(S)-extendable epimorphism e
followed by the closed homomorphism f

.
(ii) Assume that f = g e

, where e

: A C is any TAlg(S)-extendable
epimorphism and g : C B a closed homomorphism. Then there
exists an isomorphism j : C E such that j e

= e and f

j = g.
Proof Ad (i) Since every homomorphism f

: A D into some total


algebra D allows a homomorphic extension

f

: F(A, TAlg(S)) D
such that f

:=

f

[
E
satises g

e = f

, e is TAlg(S)-extendable, and
since A generates E, e is an epimorphism. Hence (i) is proved, since f

is indeed closed.
Ad (ii) Assume f = g e

as described in the assumption of (ii). If we


can show that ker e

= ker (e

, then 5.8.3 implies the existence of


an isomorphism j : C E such that j e

= e, and then g e

=
f

e = f

j e

; therefore g = f

j, since e

is an epimorphism.
Now, ker e

=
E
, and f

as well as g (e

are closed A-initial


extensions of f; thus dom(e

= dome

= E. Since e

is TAlg(S)-
extendable, and since e can be considered as a homomorphism from A
into F(E, TAlg(S)), we get a homomorphism j

: C F(E, TAlg(S))
such that j

= e, and therefore j

(e

= e, showing that (e

has
to be injective, too, i.e. indeed ker e

= ker (e

.
Remarks 5.8.5 (i) In [Sch70] ker f is used for what we denote by ker f

.
In [BWo82] we have called ker f

in an even more general context


the J. Schmidt-kernel (briey: S-kernel) of the homomorphism f.
(ii) Theorem 5.8.3 characterizes the epimorphisms starting from a given
partial algebra A up to unique isomorphism by closed congruences
on A-initial relative subalgebras of F(A, TAlg(S)); thus we have up
to isomorphism only a set of epimorphisms starting from A and there-
fore only a set of epimorphic images (up to isomorphism). In the lan-
guage of category theory this means that the category Alg(S) with
December 8, 2002
5 Universal solutions, Freeness, Terms 187
all homomorphisms as morphisms is co-(well-powered) with respect to
epimorphisms.
(iii) In section 10 we will realize that 5.8.4 just describes a factorization
system on the category of all partial S-algebras (see also [BJPa78]).
5.9 The initial closed congruence relations of a free
completion
For a still better understanding of Theorem 5.8.3 let us realize that 5.8.1
implies the
Proposition 5.9.1 Let A be any partial algebra, C an Agenerated relative
subalgebra of F(A, TAlg(S)), a closed congruence on C, nat

: C C/
the natural projection, and nat

[
A
: A C/ its restriction to A. Then
nat

[
A
is an epimorphism.
Denition 5.9.2 Let B be any partial algebra, A B any subset. Then we
denote by Rsuba
A
B the set of all A-generated relative subalgebras of B (to-
gether with the restriction of the sort homomorphism v
B
); and by Cong
c,A
B
we denote the set of all closed congruence relations on A-generated relative
subalgebras of B:
Cong
c,A
B :=

( Cong
c
C [ C Rsuba
A
B),
Cong
c,A
B := (Cong
c,A
B; ).
Notice that, for ,

Cong
c,A
B,

always implies dom dom

,
where dom := x B [ (x, y) for some y B.
The elements of Rsuba
A
B are sometimes also called A-initial segments
of B, while the elements of Cong
c,A
B are called A-initial closed congruences
of B.
The motivation for this denotation is taken from 5.5.4.
Proposition 5.9.3 Let A Alg(S) , then:
(i) Rsuba
A
F(A, TAlg(S)) is a closure system on F(A, TAlg(S)) with small-
est element A (cf. 3.2.2.(iii)).
(ii) Cong
c,A
F(A, TAlg(S)) is a closure system on F(A, TAlg(S))F(A, TAlg(S))
with smallest element
A
and largest element
F(A,TAlg(S))
.
December 8, 2002
188 Chapter I Basic denitions and properties
(iii) For a family Cong
c,A
F(A, TAlg(S)), dom

( dom [
C ).
(iv) Cong
c,A
F(A, TAlg(S)) is d

-inductive.
Proof Because of 5.5.4, and since the intersection of initial segments of
a quasi-ordered set is again an initial segment, (i) is easily realized (cf. also
3.2.2.(iii)). (iii) is obvious, too. In order to prove (ii) we have to show that, for
Cong
c,A
F(A, TAlg(S)), :=

is a closed congruence on dom =: B:


Thus assume a, b B
()
, a dom
B
, and (a(k), b(k)) for each k
(). Since each

is closed, we get b dom


dom

), therefore
b dom
B
, and (
B
(a),
B
(b)) = . (iv) is proved in a similar way
observing the proof of 2.5.4.(ii).
Corollary 5.9.4 Let ( f
i
: A B
i
[ i I ) be a family of homomorphisms
such that f

i
: domf

i
B
i
are their closed initial extensions, and let
f : A

( B
i
[ i I ) be the induced homomorphism. Then domf

( domf

i
[ i I ) and ker f

( ker f

i
[ i I ).
Proof Let C :=

( domf

i
[ i I ) and g
i
:= f

i
[
C
: C B
i
. Then there
exists the induced homomorphism g : C

( B
i
[ i I ) =: B. We claim
that g = f

:
C is obviously A-generated, and by 4.1.7 ker g = ( ker f

i
C
2
[ i I ) =
(

( ker f

i
[ i I )) C
2
= ( ker f

i
[ i I ) (cf. the denition of C) is a
closed congruence. In order to show g = f

we still have to show that g is


closed:
Let c C
()
such that g(c) dom
B
. Then g
i
(c) dom
B
i
(i I), i.e.
c

( dom
domf

i
[ i I ) = dom
C
.
5.10 On the cardinalities of M-generated partial alge-
bras
From the results obtained so far we can also infer some observations concern-
ing the cardinalities of partial algebras.
Lemma 5.10.1 Let M be any set. Then
#F(M, TAlg(S)) 2
#M+#+
0
+

( ()|)
= 2
sup( #M,#,
0
,( ()|))
.
December 8, 2002
5 Universal solutions, Freeness, Terms 189
Proof In the proof of Theorem 5.2.6 we have realized F(M, TAlg(S)) as a
subset of the power set Subs Z

of the set Z

of all nite sequences with ele-


ments in M

() [ =: Z. Now #Z #+#M+

( () [
) =: s

. Let s := s

+
0
. Then it is well known from cardinal arith-
metic that #(Z
n
) s
n
= s for every natural number n, since s is innite.
Hence #(Z

) =

(#( Z
n
[ n N)) (
0
) s = s. Thus #F(M, TAlg(S))
#Subs Z

= 2
#(Z

)
2
s
. There also follows from cardinal arithmetic that
#+#M +
0
+

( () [ ) = sup( #M, #,
0
, () [ ), since
both cardinals are innite and

( () [ )
0
+# sup( () [
) =
0
+ # + sup( () [ ).
Proposition 5.10.2 Let A be any partial algebra of similarity type , and
let M be a generating subset of A. Then #A #F(M, TAlg(S)).
Proof Because of the Generalized Recursion Theorem 5.4.1 i := id
M
: M
A extends to a closed homomorphism i

: domi

A. Since M generates A,
i

is surjective, and therefore #A #(domi

) #F(M, TAlg(S)).
Corollary 5.10.3 If A Alg(S) and M generates A, then we have for the
rank r of (cf. 1.1):
#A 2
max{#M,#,
0
,r}
.
These results show that an M-generated partial algebra cannot become
too large with respect to #M and the data of its similarity type. However,
if one has only nitary operations, then one has an even better cardinality
bound:
Proposition 5.10.4 Let be a nitary similarity type, and let M be any
set.
If = or if has only nullary constants, then #F(M, TAlg()) = #M +
#.
If M = = [ () = 0 , then F(M, TAlg()) = , else #F(M, TAlg()) =
max#M, #,
0
.
Moreover, one always has #F(M, TAlg(S)) #F(M, TAlg()).
Proof In the rst case we have F(M, TAlg()) = M

. If M is empty,
and if there are no nullary constants, then there is nothing to start with
December 8, 2002
190 Chapter I Basic denitions and properties
for a generation. Since a heterogeneous Peano algebra is always a relative
subalgebra of the corresponding homogeneous Peano algebra, also the last
statement is obvious. In the remaining case there is at least one at least
unary operation, and M

[ () = 0 ,= ; thus the set of natural
numbers with the successor function as fundamental operation shows that
F(M, TAlg()) will be at least countably innite. Moreover, #F(M, TAlg())
will always be at least as large as #M and #. In order to get an upper
bound we use the construction of F(M, TAlg()) sketched in 5.2.7.(i): It
yields
#F(M, TAlg()) #
_
(M

)
n
[ n
0
max#M, #,
0
.
5.11 On the existence of K-universal (K-) solutions
We now have enough tools at hand to provide the construction of a K-
universal solution for a given partial algebra A with respect to a class K
of partial S-algebras of the same type. The following procedures are essen-
tial for although a little more general than the eective construction
of the semantics of algebraic specications of abstract data types, although
we present here rather the theoretical background than algorithms. We shall
see this in more detail at the end of the next section.
The results of the previous subsections and the Diagram Completion
Lemma will immediately motivate the following
Denition 5.11.1 Let A Alg(S), K Alg(S) be given.
(i) Rsuba
A
K := domf

[ f : A B for some B K
designates the set of all A-initial segments of F(A, TAlg(S)) induced
by homomorphisms from A into some K-algebra.
Cong
c,A
K := ker f

[ f : A B for some B K
designates the set of all A-initial congruences induced by some homo-
morphism from A into some K-algebra.
(ii) char
A
K :=

Cong
c,A
K is called the A-characteristic of K.
As an immediate consequence of 5.9.3 we get
Lemma 5.11.2 Let A Alg(S), K Alg(S). Then

Rsuba
A
K = domchar
A
K,
and char
A
K is a closed congruence relation on

Rsuba
A
K.
December 8, 2002
5 Universal solutions, Freeness, Terms 191
Remarks 5.11.3 (i) The concept of an A-characteristic of some class K
is a generalization of the concept of an M-characteristic of a partial al-
gebra or a class of partial algebras M any set , as it was introduced
by J.Schmidt in [Sch62] and [Sch64] as a generalization of the char-
acteristic of a ring R with unit, which is with respect to 5.11.1
the Z-characteristic of R, (Z := (Z; +, 0, , , 1) designating the ring of
integers). One should only observe that in ring and eld theory, when
(0, p) generates the corresponding congruence of Z, then the natural
number p stands for this congruence, since it generates the correspond-
ing ideal.
(ii) Having a closer look to Cong
c,A
K, Rsuba
A
K and char
A
K we should
have written more precisely Cong
c,A
(F(A, TAlg(S)); K) etc., taking
pattern from Denition 5.9.2 resp. Proposition 5.9.3.
(iii) domchar
A
K is the set of all those elements of F(A, TAlg(S)) which are
interpreted by every homomorphism from A into some K-algebra.
Hence, if A is discrete, domchar
A
K is the set of exactly those terms
t F(A, TAlg(S)) for which (A : t)
B
is a total term operation in
every K-algebra B. Therefore it could be considered as a completeness
degree for K. The meaning of char
A
K will become more obvious in the
next section about the equational theory of partial algebras.
(iv) Notice that in the case, when Rsuba
A
K and hence Cong
c,A
K is
empty, i.e. when there is no homomorphism from A into any K-algebra,
then

Rsuba
A
K = F(A, TAlg(S)) and char
A
K =
F(A,TAlg(S))
, since
these are the largest elements of the corresponding complete lattices
(ordered by inclusion).
Theorem 5.11.4 Let A Alg(S), K Alg(S). Then the restriction r to A
of the natural homomorphism
nat
char
A
K
:

Rsuba
A
K

Rsuba
A
K/char
A
K =: F
(i.e. r : A F), is a K-universal epimorphism.
Proof 5.9.1 implies that r is an epimorphism. From the denition of r
and the Diagram Completion Lemma 5.8.2 for epimorphisms it immediately
December 8, 2002
192 Chapter I Basic denitions and properties
follows that r : A F is K-extendable. There remains the extremality
property to be shown:
Let r

: A F

be any other K-extendable epimorphism, and let C :=


dom(r

, := ker (r

. Then 5.8.2 implies that ker f

for every
f : A B (B K), i.e. char
A
K, therefore another application of 5.8.2
yields the existence of a homomorphism h : F

F such that h r

= r.
This shows that r is K-universal.
For the following we need a technical lemma.
Lemma 5.11.5 Let A Alg(S), K Alg(S). Then there exists a subset
K

K such that for every B K and for every homomorphism f : A B


there exist B

and f

: A B

such that domf

= dom(f

and
ker f

= ker (f

, i.e. such that


()
(domf

)/ker f

[ f : A B, B K = (domf

)/ker f

[ f : A B

, B

.
Proof Let M := (C, ) [ C Rsuba
A
F(A, TAlg(S)), Cong
c
C, and
let M K be the following relation: ((C, ), B) , iff (C, ) M,
B K, and there exists a homomorphism f : A B such that domf

= C
and ker f

= .
By Proposition II.7.8 in A.Levy [Lev79] (cf. also section 23) there exists
a subrelation

such that, for each (C, ) M, [(C, )]

:= B
K [ ((C, ), B)

is a set and [(C, )]

,= iff [(C, )]

,= for each
element of M.
Since M is a set, we get that
K

:= B K [ B [(C, )]

for some (C, ) M =


_
[(C, )]

[ (C, ) M
is a set. Because of the denitions of M, and

the set K

has the required


properties.
From 5.11.5.() we immediately get the
Corollary 5.11.6 Let A Alg(S), K Alg(S), and let K

K be any subset
such that 5.11.5.() above is satised. Then one has
char
A
K =

ker f

[ f : A B, B K
=

ker f

[ f : A B, B K

= char
A
K

December 8, 2002
5 Universal solutions, Freeness, Terms 193
and
domchar
A
K =

Rsuba
A
K = domchar
A
K

.
Proposition 5.11.7 With the assumptions and the notation of Theorem 5.11.4
we have F oT K.
Proof From Lemma 5.11.5 we know that there is a subset K

of K such that
() of 5.11.5 is satised. Let P := domchar
A
K and := char
A
K. Then 5.11.6
yields P =

Rsuba
A
K =

Rsuba
A
K

and = char
A
K

; therefore F =
P/. Consider the index set (!) I := (f, B) [ B K

, f : A B, and
dene D :=

( B [ (f, B) I ). By the denition of char


A
K

and P each
f : A B ((f, B) I) extends to a homomorphism

f
(f,B)
: P B (since
char
A
K

ker f

). Thus we get an induced homomorphism u : P D such


that pr
(f,B)
u =

f
(f,B)
for each (f, B) I. Obviously ker u =

ker

f
(f,B)
[
(f, B) I = char
A
K

= , i.e. by 5.9.4 and its proof u is closed. The


First Homomorphism Theorem 2.7.4 implies that u induces an isomorphism
i : P/ = F C
D
u[P]. Hence F is isomorphic to a subalgebra of a product
of K-objects, i.e. F oT K.
Observing the denitions of K-universal solutions and quasi-primitive
classes we immediately get
Theorem 5.11.8 (Existence Theorem for K-universal K-solutions) Let
K Alg(S) be any quasi-primitive class of partial algebras. Then there ex-
ists for every partial algebra A Alg(S) its K-universal K-solution, and one
realization of it is
(r
A,K
, F(A, K)) := (nat
char
A
K
[
A
,

Rsuba
A
K/char
A
K).
Reviewing the preceding arguments one easily realizes that for 5.11.8 the
closedness of K with respect to products on the empty index set is only
needed in the case, when there does not exist a homomorphism from A into
any K-algebra. Thus we have the
Corollary 5.11.9 (1) Let K Alg(S) such that oT
+
K = K. Then a K-
universal K-solution exists for every partial algebra A Alg(S) which allows
at least one homomorphism into some K-algebra.
December 8, 2002
194 Chapter I Basic denitions and properties
5.11.10 The observation 5.11.9 will become important in connection with
classes dened by universal Horn formulas, which we shall consider in Chap-
ters II and III.
Corollary 5.11.11 (2) Let K = oT
+
K Alg(S), and let M be any set.
If A K [ v
M
[M] v
A
[A] , = , then the K-universal solution of M
exists. And if for each s S satisfying #M
s
2 there is A K such that
v
M
[M] v
A
[A] and #A
s
2, then the K-free K-algebra K-freely generated
by M exists with an injective embedding r
M,K
.
Remarks 5.11.12 (i) In the case of homogeneous partial algebras the
assumptions in 5.11.11 just mean K ,= respectively if #M 2,
then there is A K with #A 2.
(ii) If F(, K) exists, then it is called the initial K-algebra. Obviously
F(, K) exists, whenever oT
+
K = K ,= . Initial K-algebras play
an important role in connection with the algebraic specication of
abstract data types in computer science (see 17.6, and for more de-
tails and examples e.g. H.Reichel [Re87], respectively [GoTcWa78]
by the so-called ADJ-group or H.Ehrig & B.Mahr [ErMh85] or
H.A.Klaeren [Kla83] for the total case).
(iii) See also Remark 5.1.9, whose statements have now been proved, or can
easily be derived from what we have shown in this section.
December 8, 2002
6 The equational theory of partial algebras 195
6 The equational theory of partial algebras
Although we shall introduce in Chapter II a general model theory for partial
algebras including an equational theory, we develop the equational theory of
partial algebras based on existence equations here separately because
of its importance in Universal Algebra and Computer Science. Equations are
the basic axioms treated in Universal Algebra, respectively used in the alge-
braic specication of abstract data types. Also one might say that (almost)
all of the development of Algebra is connected with the theory of solvabil-
ity of (special kinds of) equations. It should not be concealed that quite a
few approaches to an equational theory for partial algebras are conceivable
(see e.g. S.C.Kleene [Kl52] and J.Slominski [Sl68] for the rst treatments
of several approaches to an equational theory for partial algebras, while in
R.John [J75] and [J78] one might possibly nd the most comprehensive dis-
cussion of dierent (axiomatic) approaches to equational theories for partial
algebras; in 8.1.3 and 22.2.1 we shall discuss several of these proposals in
some more detail). In this book we mainly deal with what we shall call
existence equations for several reasons:
(i) As far as we can judge it, they have a theory which is most parallel to
the equational theory of total algebras what is not too astonishing,
since for an existence equation (M : t
e
= t

) valid in a partial algebra A


the global term operations (M : t)
A
and (M : t

)
A
both have to be
total.
(ii) It is quite easy to take them as a basis for a general rst order logical
language with equality for partial algebras whose semantics is quite
similar to the usual one for total algebras (cf. [B82]). At least in the
homogeneous case the formulas are the same as those for total alge-
bras, only in the semantics one gets dierences in particular, since the
existence equations (M : t
e
= t), which are trivially valid in all total
algebras, become of special importance for partial algebras (they take
in our approach the role of new atomic formulas like t with the
semantics that (the evaluation of) the term t has to exist.
(iii) It should not be denied that the author has a special interest in them,
since he ran upon them in connection with the investigation of free
partial algebras (see [B70]) and the description of primitive classes of
December 8, 2002
196 Chapter I Basic denitions and properties
partial algebras (see [B71]). These papers quoted were amongthe start-
ing points of the investigations of H.Andr eka, I.N emeti and I.Sain
(see e.g. [AN83], [AN79a], [AN82] and [ANSa82] as well as [ABN80]),
who extended the theory using category theoretical tools and general-
izing these beginnings enormously always having the applications to
partial algebras in mind; the presentation of the applications of their
results to partial algebras in particular of their Meta-Birkho The-
orem will be the main purpose of Chapter III.
(iv) Most of the other approaches to equational theories referred to above
for partial algebras can be expressed in the language based on ex-
istence equations as special instances of what we shall call in 8.1.2
existentially conditioned existence equations (ECE-equations) i.e.
of special implications , or as two-element sets of them. And the
diculties in the desription of their theories if one has any at all
can in most cases easily be explained in connection with these repre-
sentations.
6.1 Existence-equations
Denition 6.1.1 (i) Let = (())

be any xed similarity type of


cardinal dimension d, (d

= maxd,
0
), let S be any sort algebra
(specifying a heterogeneous similarity type as indicated in ??), and
let X be any xed set such that each phylum of X has (at least) car-
dinality d

for most applications max2, d would suce. Moreover,


let d
&
:= maxd, (#S)
+
, where (#S)
+
is the least cardinal greater
than #S.
X will be called the (global) set of variables under consideration. More-
over, let T := F(X, TAlg(S)) be the Peano algebra on X, and in this
context we shall call T the global term algebra, its elements are called
(global) terms (with variables in X). In the homogeneous case, T would
be sucient, but since we also want to include the heterogeneous case
in the form of partial S-algebras with possibly empty phyla, we also
have to consider the set T
mathrmloc
:=

( M F(M, TAlg(S)) [ M
X,#M < d
&
) of all local terms with variables in X, which are writ-
ten as (M : t) (for t F(M, TAlg(S))). If A Alg(S) is any par-
tial algebra and M X any subset of variables, then every map-
ping w : M A will be called a (local) valuation or assignment for
December 8, 2002
6 The equational theory of partial algebras 197
M in A, and w

: domw

A its closed M-initial extension in


F(M, TAlg(S)) (or F(X, TAlg(S))) will be called the corresponding
evaluation (or corresponding interpretation) of terms.
(ii) The elements (M, t, t

) MF(M, TAlg(S))
2
which satisfy v
F(M,TAlg(S))
(t) =
v
F(M,TAlg(S))
(t

) will be called existence-equations (briey: E-equations),


and will formally be written as (M : t
e
= t

) here we do not impose


any cardinality restriction on M!
For each M X, (#M < d
&
) we introduce the sets
Eeq
M
:= (M : t
e
= t

) [ t, t

F(M, TAlg(S)) and


v
F(M,TAlg(S))
(t) = v
F(M,TAlg(S))
(t

)
of all E-equations on M, and
Eeq

X
:=
_
( Eeq
M
[ M X, #M < d
&
)
of all small E-equations with variables in X (i.e. connected with a small
set (#M < d
&
) of variables).
We say that an E-equation (M : t
e
= t

) is satised in a partial algebra


A Alg(S) under a valuation v : M A and we denote this by
A [= (M : t
e
= t

)[v] , iff (t, t

) ker v

(i.e. i both evaluations


v(M : t)
A
and v(M : t

)
A
exist in A and are equal: v(M : t)
A
= v(M :
t

)
A
(or, equivalently, v

[t] = v

[t

])).
If A [= (M : t
e
= t

)[v], then v is called a solution of the E-equation


(M : t
e
= t

).
We say that an E-equation (M : t
e
= t

) is valid in A and denote this


by A [= (M : t
e
= t

) , iff A satises (M : t
e
= t

) for every assignment


v : M A, i.e. A [= (M : t
e
= t

)[v] holds for every v A


M
.
(iii) Let K Alg(S) be any class of partial algebras, let G Eeq

X
be any
set of E-equations, and let M X, (#M < d
&
); then we dene:
Eeq
M
K := (M : t
e
= t

) Eeq
M
[ A [= (M : t
e
= t

) for every A K,
the set of all E-equations on M, which are valid in K, i.e. in all K-
algebras;
Eeq K := (M : t
e
= t

) Eeq

X
[ A [= (M : t
e
= t

) for every A K,
December 8, 2002
198 Chapter I Basic denitions and properties
the set of all E-equations with (a small set of) variables in X, which
are valid in K;
Mod G := A Alg(S) [ A [= (M : t
e
= t

) for every (M : t
e
= t

) G.
For A Mod G we also write A [= G (G is valid in A).
Mod G is called the class of all models of the system G of E-equations,
and Eeq K is called the (E-) equational theory of the class K.
Remarks 6.1.2 (and further concepts)
(i) Notice that A [= (M : t
e
= t

)[v] means that (t, t

) ker v

, i.e. the
values of (M : t)
A
and (M : t

)
A
exist with respect to v and are equal,
i.e. A [= (M : t
e
= t

)[v] includes three statements:


(a) t domv

,
(b) t

domv

,
(c) v[t] = v[t

].
Thus the statement A [= (M : t
e
= t

)[v] is false means that at least


one of the statements (a), (b) or (c) is false.
(ii) If t and t

are identical terms, then A [= (M : t


e
= t)[v] still says
that t domv

, i.e. that v(M : t)


A
exists. This is the reason, why
we speak of existence-equations rather than of equations: Namely
the diagonal-expressions (M : t
e
= t), which are trivially satised
and valid in all total algebras, get a real meaning for partial algebras.
Because of this semantic property we call E-equations of the form (M :
t
e
= t) term-existence-expressions (briey: TE-expressions).
From results which we shall derive in later chapters we shall realize
that exactly the fact that TE-expressions get a real (non-trivial) mean-
ing for partial algebras while being of no real importance for total
algebras causes the wealth of concepts, e.g. of signicant properties
of substructures and homomorphism, for partial algebras in compari-
son to total algebras. With a more extended rst order language and
the corresponding model theory at hand we shall see in Chapters II
and III that these concepts and properties can be organized quite well
by model theoretic concepts especially using TE-expressions (and their
negations).
December 8, 2002
6 The equational theory of partial algebras 199
(iii) It should also be observed that validity of an E-equation (M : t
e
= t

)
in a partial algebra A does not only mean that (M : t)
A
and (M : t

)
A
take the same values, whenever they are dened, but also that they
both are total M-ary term operations on A, i.e. A [= (M : t
e
= t

) iff
dom(M : t)
A
= dom(M : t

)
A
= A
M
and (M : t)
A
= (M : t

)
A
.
When we have to describe really partial term operations, we shall have
to use elementary implications between conjunctions of E-equations, as
we shall introduce them in the next chapter.
(iv) It will become clear in the next subsections (cf. 6.4.9.(iii)), why we
consider d
&
rather than d

as the cardinality restriction for the subsets


of variables. It would suce to consider sets M in X such that #M
s
<
d

for each s S. However, this would make some assumptions still


more complicated, and we keep to the coarser assumption: #M < d
&
.
6.1.3 (On the Galois-connection related to validity of E-equations)
Validity of E-equations establishes a Galois-connection between Eeq

X
(re-
spectively Eeq
M
for any M X) and Alg(S) (cf. subsection 24.4).
Thus the operators Eeq, Mod, Mod Eeq and Eeq Mod (or Eeq always re-
placed by Eeq
M
and Eeq

X
by Eeq
M
notice that in the combinations one
rst has to apply the right-hand operator and only afterwards the left-hand
operator) have the following properties for G G

Eeq

X
(Eeq
M
) and
K K

Alg(S):
(1) Eeq K

Eeq K and Mod G

Mod G
(2) K Mod Eeq K and G Eeq Mod G
(3) Mod Eeq K Mod Eeq K

and Eeq Mod G Eeq Mod G

(4) Eeq Mod Eeq K = Eeq K and Mod Eeq Mod G = Mod G
Mod Eeq Mod Eeq K = Mod Eeq K and Eeq Mod Eeq Mod G = Eeq Mod G
(5) Eeq
M
Mod Eeq K = Eeq
M
Mod Eeq
M
K = Eeq
M
K.
The easy straightforward proofs, which also follow from the general principles
of Galois-connections (cf. subsection 24.4) are left as exercises to the reader.
In connection with (5) it has to be observed that because of the fact that
we also treat heterogeneous partial algebras with possibly empty phyla an
E-equation cannot be separated from the set of variables with which it has
been connected. Notice that in the homogeneous case this would not have
December 8, 2002
200 Chapter I Basic denitions and properties
been necessary. Namely, every mapping from some non-empty subset of X
into some homogeneous partial algebra can be extended to all of X. One
only has to treat the empty subset separately, since we allow empty models.
Thus Eeq Mod is a closure operator on Eeq

X
, and Mod Eeq behaves like
a closure operator on Alg(S) observe that Alg(S) is not a set. It is the
main aim of this section to characterize the closed subsets of Eeq

X
, which
are called (E-) equational theories, and the closed subclasses of Alg(S), which
are called (E-) equationally denable classes or E-varieties (here we cannot
drop the prex E-, since we shall consider later also other variety concepts).
The next lemma is an immediate consequence of the above denitions
and section 5.
Lemma 6.1.4 Let A Alg(S), K Alg(S), M X (#M < d
&
) and
G Eeq

X
. Then
(i) Eeq
M
A =

ker v

[ v A
M
= char
M
A;
Eeq A =

Eeq
M
A [ M X =

char
M
A [ M X, #M < d
&
.
(ii) Eeq
M
K =

Eeq
M
A [ A K = char
M
K;
Eeq K =

Eeq
M
K [ M X =

char
M
K [ M X, #M < d
&
.
In the homogeneous case it is sucient to consider and X instead of all
(small) subsets of X.
(iii) A Mod G, iff G Eeq A, i.e. iff G

ker v

[ v A
M
[
M X, #M < d
&
, or in other words iff , for each M X with
#M < d
&
, G Eeq
M
char
M
A.
Remark 6.1.5 Since we have introduced into the notation of terms and E-
equations the reference to the corresponding set of variables in order to be
able to check, whether or not it allows a mapping into a partial algebra A
under consideration, i.e. whether v
M
[M] v
A
[A], it would suce to replace
M by v
M
[M]in these notations. Then, for A Alg(S), K Alg(S), one
could replace the second statements in 6.1.5.(i) and (ii), respectively, by
(i) Eeq A =

char

{ Xs|SS

}
A [ S

S ;
(ii) Eeq K =

char

{ Xs|SS

}
K [ S

S .
December 8, 2002
6 The equational theory of partial algebras 201
Thus one has to compute the equations for fewer sets, what is particularly
useful, if S is a very small set, although the results obtained will become
somewhat coarser in other respects. In what follows we keep with the old
notation, however the reader should always be aware of the possibilities al-
lowed by (i) and (ii) in the alternative approach. In particular most of the
remarks concerning the homogeneous case are based on this observation.
6.2 Description of the model classes (E-varieties)
Proposition 6.2.1 Let G Eeq

X
, K := Mod G. Then:
(i) HK o K T
r,d
&
K K.
(ii) (domEeq
M
K)/Eeq
M
K

= F(M, K) K.
Proof Ad (i):
(a) Let A K, f : A B a surjective homomorphism and v : M B
an assignment for some M X, (#M < d
&
). Since f is surjective,
there exists using the Axiom of Choice an assignment w : M A
such that v = f w. Now, by 6.1.4.(iii) and 6.1.3.(1) A Mod G implies
G Eeq
M
char
M
Mod G char
M
A ker (f w

) ker v

, and
therefore G Eeq B easily follows. Thus HK K.
(b) Let A K, B a subalgebra of A, M X (#M < d
&
) and v : M B.
Then ker v

does not depend on the fact that we have chosen B instead


of A as the target. Thus G Eeq
M
Eeq
M
A Eeq
M
B, i.e.
B K.
(c) Let ( A
i
[ i I ) be a family of K-algebras, T a d
&
-complete lter
on I and (( p
F
[ F T ), A) the direct limit of the d
&
-directed system
( A
F
:=

( A
i
[ i F ), p
FF
: A
F
A
F
[ F F

T ) (i.e. a
reduced product), and let v : M A be any assignment for M X,
#M < d
&
. Since each element of A has a preimage in some A
F
, since
#M < d
&
, and since T is d
&
-complete, there are F T and using
the Axiom of Choice a mapping w : M A
F
such that p
F
w = v.
Since G Eeq
M


char
M
A
i
[ i F

ker (pr
i
w)

[ i
F = ker w

, we also get G Eeq


M
ker (p
F
w

) ker v

, showing
that A Mod G.
(ii) follows immediately from 5.11.4, 5.11.7 and (i): F(M, K) oT K
Mod K.
December 8, 2002
202 Chapter I Basic denitions and properties
Because of HMod ( Mod G for any set G of E-equations we have in
particular
Corollary 6.2.2 Let G Eeq

X
. Then Mod G contains (S, id
S
) the empty
product , and for every A also each inner extension of A, i.e. each B
Alg(S) for which B = A and id
A
: A B is a homomorphism.
Many of the technical results needed in the proof of the following main
results are collected in
Lemma 6.2.3 Let N and M be any sets and K Alg(S). Moreover, let
p := nat
char
M
K
: domchar
M
K domchar
M
K/char
M
K = F(M, K)
(cf. 5.11.8). Then:
(i) Let P and Q be partial Peano algebras on M and N, respectively, and let
j : N M be an injective mapping, which extends to a homomorphism

j : Q P. Then

j is injective.
(ii) If N M, then char
N
K char
M
K (c
F(M,TAlg(S))
N)
2
, and if
() v
M
[m] v
domchar
N
K
[domchar
N
K],
then we have equality.
(iii) If j : N M is an injective mapping, then it induces a homomorphism

j : F(N, K) F(M, K); and if (ii).() is satised, then



j is closed and
injective.
(iv) For every set Y and terms t, t

F(Y, TAlg(S)) with v


Y
[t] = v
Y
[t

]
there are a set M X, an injective mapping j : M Y and
terms q, q

F(M, TAlg(S)) such that j has a homomorphic extension

j : F(M, TAlg(S)) F(Y, TAlg(S)) satisfying #M < d


&
,

j(q) = t,

j(q

) = t

, and such that for every A Alg(S) and every v : Y A we


have A [= (Y : t
e
= t

)[v] iff A [= (M : q
e
= q

)[v j]. Moreover


() A [= (Y : t
e
= t

) iff A [= (M : q
e
= q

).
and M can be chosen such that

j is closed and injective.
December 8, 2002
6 The equational theory of partial algebras 203
(v) If for some s S there is A K such that the phylum A
s
of A has at
least two elements, then the restriction p[
Ms
is injective.
(vi) Let X and Y both be sets (of variables) satisfying that each phylum
has cardinality at least d

, and let for the moment , for each K


Alg(S), Eeq

X
K and Eeq

Y
K designate the sets of all E-equations valid
in K with variables in X, respectively Y . Then the operators Mod Eeq

X
and Mod Eeq

Y
are identical.
Proof If K = , then char
M
K = F(M, TAlg(S))
2
, and all statements be-
come trivial. Hence we can assume in what follows that K ,= .
Ad (i) This can either be proved by algebraic induction, or one may argue
as follows: One can nd a set N

N such that j extends to a bijection


j

: N

M, which then extends to an isomorphism



j

: F(Q

(N

N), TAlg(S)) F(P, TAlg(S)) between total Peano algebras and


is an extension of

j, which therefore has to be injective.
Ad (ii) Since every (S-) mapping from M into A K can be restricted to N,
but not necessarily every mapping from N into A can be extended
to one from M into A, we easily get the inclusion stated in (ii), and
realize that equality will not be true in general. Now assume (), and
let (t, t

) char
M
K (c
F(M,TAlg(S))
N)
2
, and let v : N A (A K)
be any assignment for N. Condition () now says possibly by an
application of the Axiom of Choice, if v
M
[M N] is innite that
we can get an extension w : M A of v, whose closed homomorphic
extension w

: M A obviously coincides with v

on (domw

)
F(N, TAlg(S)) = domv

: It is obvious that domv

is contained in the
intersection, and from 5.5.5 we may infer that also the other inclusion
holds. And therefore two terms t and t

from F(N, TAlg(S)) can be


identied by w

iff they are identied by v

. The denition of char


N
K
together with the rst argument then yields the desired equality.
Ad (iii) This is an immediate consequence of (ii), (i) and 5.8.2.
Ad (iv) Let (Y : t
e
= t

) be any E-equation with respect to an arbitrary set Y .


Let var
Y
(t, t

) := Y (t t

) be the set of all elements of Y really


needed (cf. 5.5.5) to generate t and t

. Then #var (t, t

) < d

. From
each nonempty phylum we pick an element y
s
Y
s
(s S) and dene
Y (t, t

) := var
Y
(t, t

) y
s
[ s S, Y
s
,= . Obviously #Y (t, t

) < d
&
and we have an injective mapping j

: Y (t, t

) X. Dene M :=
December 8, 2002
204 Chapter I Basic denitions and properties
j

[Y (t, t

)] and j := j

1
: M Y . Obviously j

has a homomorphic
extension

j

: c
F(Y,TAlg(S))
Y (t, t

) c
F(X,TAlg(S))
M, which is obviously
an isomorphism (since it is easy to realize that c
F(Y,TAlg(S))
Y (t, t

)

=
F(Y (t, t

), TAlg(S)) and c
F(X,TAlg(S))
M

= F(M, TAlg(S)), and j

is bi-
jective!). Its inverse then induces a closed and injective homomorphism

j : F(M, TAlg(S)) F(Y, TAlg(S)). Dene q :=



j

(t) and q

:=

j

(t

).
Then the construction of Y (t, t

) and the fact that



j

is an isomorphism
imply the rest of the statement, in particular ().
Ad (v) With the assumptions of (v) we have for any two dierent elements m
and m

of the phylum M
s
a mapping w
m,m
: M A
m,m
K such that
w
m,m
(m) ,= w
m,m
(m

). Hence the induced mapping, say w, from M


into the product of all these K-algebras separates all these pairs, i.e.
it is injective on M
s
. Since there is a homomorphism g from F(M, K)
into this product such that w = g r
M,K
, we infer that r
M,K
[
Ms
= p[
Ms
is
injective, and it is easy to realize that p is its closed M-initial extension.
Ad (vi) Assume A Mod Eeq

X
KMod Eeq

Y
K. Then we have (M : t
e
= t

)
Eeq

Y
K, which is not valid in A. But this is easily seen to contradict (iv)
and the fact that A Mod Eeq

X
K.
Remark 6.2.4 6.2.3.(vi) justies that we may x for a set of variables for
E-equations any set X which only satises that each phylum has at least
cardinality d

, as we have done at the beginning.


We are now able to prove the generalization to partial (S-)algebras of
G.Birkhoffs rst main theorem on equational theories (cf. [Bi35]).
Theorem 6.2.5 (First Birkho Theorem) Let K Alg(S) be any class
of partial algebras and let X be any set of variables such that each phylum
of X has cardinality at least d

. Then:
(a) Mod Eeq K = HoT K.
(b) The following statements are equivalent:
(i) K = Mod Eeq K.
(ii) K = HoT K.
December 8, 2002
6 The equational theory of partial algebras 205
Proof (b) is an immediate consequence of (a), and HoT K Mod Eeq K
follows from 6.2.1 and 4.6.4.(viii).
Thus assume A Mod Eeq K. From 5.11.7 and 5.11.8 we know that F(A, K)
the oT K-free oT K-algebra with basis r
A,K
[A] exists and belongs to oT K
Mod Eeq K. Hence there is to show that i := id
A
: A A induces a
necessarily surjective homomorphism

i : F(A, K) A. And with respect
to 5.11.8 and 5.8.2.(a) we only have to show that char
A
K ker i

.
Thus assume (A : t
e
= t

) char
A
K. Then we have, according to 6.2.3.(iv),
M X, q, q

c
F(M,TAlg(S))
M and an injective mapping j : M A such
that #M < d
&
and C [= (M : q
e
= q

) iff C [= (A : t
e
= t

) for all C Alg(S).


Since (A : t
e
= t

) char
A
K iff D [= (A : t
e
= t

) for all D K, this


implies (M : q
e
= q

) Eeq K, and therefore A [= (M : q


e
= q

), respectively
A [= (A : t
e
= t

); thus (t, t

) ker i

, which was to be shown.


Remarks 6.2.6 (i) The preceding result shows that the E-varieties are
exactly the primitive classes. In the homogeneous case one often
takes assignments only for the whole set X. Then every E-variety
always contains the empty partial algebra since this then allows no
assignment and hence each E-equation is trivially valid in it or one
also has to forbid the empty algebra. If we proceed in the same way
for many-sorted (partial) algebras, we would get a whole class of non-
isomorphic trivial models contained in each E-variety, namely all
those with at least one empty phylum or one has to allow only partial
S-algebras with surjective sort homomorphism, i.e. forbid empty phyla,
which causes trouble in connection with substructures and especially
with the concept of generation. This shows that with each concept one
runs into trouble, and one just has to make a decision, as we have done
in this book.
(ii) It should be noticed that 6.2.3.(ii).() is in general really necessary for
the statements in 6.2.3.(ii) and (iii), i.e. the homomorphism

j from 6.2.3.(iii)
need in general be neither injective nor closed, nor full. In order to see
this, let us consider the following example:
:= (0, 0, 0, 1, 1, 2) := (,

, , , ) K := A, B,
where S, A and B are sketched in Figure 6.1. It is easy to realize that
(assuming b

X) A = F(, K) and B = F(b

, K), since there is


December 8, 2002
206 Chapter I Basic denitions and properties
gure uncompleted
Figure 6.1: f6.1
no assignment from b

into A, and the only possible homomorphism


f : A B is neither injective nor full nor is its image a subalgebra
of B. In the homogeneous case such an eect can only be obtained,
when K contains the empty algebra and there is a nullary fundamental
constant which is dened in each non-empty K-algebra. Better fea-
tures can be obtained, when one considers hierarchical equoids, as this
is done in Part II [Re87], but they do not form an E-variety.
(iii) Notice that stack automata (which we have considered in the introduc-
tion) with natural numbers as input form an E-variety, where we have
the following specication (in the notation of Part II [Re87]):
Stack(Nat) is denition
sorts States, Nat
oprn 0 Nat
Succ(Nat) Nat
Pred(Nat) Nat
States
Top(States) Nat
Pop(States) States
Push(States, Nat) States
axioms x : States, y : Nat
( : 0
e
= 0 )
( y : Pred(Succ(y))
e
= y )
( :
e
= )
(x, y : Top(Push(x, y))
e
= y )
(x, y : Pop(Push(x, y))
e
= x)
end Stack(Nat).
Let K be the class of all stack automata with natural numbers as input,
i.e. all partial algebras of the above type satisfying the above axioms.
As the stack automaton with natural numbers as input elements one
then declares F(, K), the so-called initial algebra of K. It is really
December 8, 2002
6 The equational theory of partial algebras 207
partial, since Pop(), Top() and Pred(0) are not dened. For
further information about this example see 6.4.11.(iii)).
(iv) Notice that the class of all small categories (cf. Example 3 of the in-
troduction or 1.2.1.(iii)) is not an E-variety, since it is not closed with
respect to homomorphic images (cf. 8.1.1). We shall see in 8.1.1 that
they form an ECE-variety (where an ECE-equation is an elementary
implication with a conjunction of TE-expressions in the premise).
6.3 More facts about freeness, fully invariant congru-
ences
6.3.1 In 5.1.4 and 5.11.11 we have already mentioned K-free algebras, and
in 6.2.5 they played an important role in the proof, showing that in an
E-variety K each K-algebra A is a homomorphic image of the K-free K-
algebra F(A, K) which then exists, as one can derive from the proof of 6.2.5
and it is then characterized in 5.11.8.
From 5.1.6 and 6.2.5 we may derive that every independence class is
an E-variety. If we consider only homogeneous partial algebras and do not
allow the empty partial algebra, then Remark 6.2.6 easily yields that also the
converse is true this is also the case, when E-equations and assignments
are always dened with respect to the whole set X of variables.
In this subsection we want to learn more on free partial algebras, since
the results e.g. in 6.1.4 and 5.11.4 show that the K-free K-algebras encode
the E-equations valid in K. Hence we shall use them for the characterization
of closed sets of E-equations.
Proposition 6.3.2 Let K Alg(S) and let M be any set. Then domchar
M
K
is freely and K-freely generated by M.
Proof Because of 5.9.3 M generates P := domchar
M
K. If v : M A
K is any assignment, then there is v : F(M, K) A such that, for p :=
nat
char
M
K
, v p extends v (observe 6.2.3.(v)), i.e. M is K-independent in P.
Now assume v : M P to be any mapping (see Figure 6.2), and let v

:
domv

P be its closed M-initial homomorphic extension. Then p v

:
domv

F(M, K) K is an M-initial closed extension of p v, and


since P 1oT K ind
P
M by the denition of P and the closedness of p
: P dom(p v)

= dom(p v

) = domv

; therefore v

[
P
: P P is a
homomorphic extension of v to P, showing that M is independent in P.
December 8, 2002
208 Chapter I Basic denitions and properties
gure uncompleted
Figure 6.2: f6.2
gure uncompleted
Figure 6.3: f6.3
A closely related result is the following one (see also 17.4.1):
Proposition 6.3.3 Let F Alg(S) be a free partial algebra with basis M,
and let P := domid
M

be the corresponding relative subalgebra of F(M, TAlg(S))


with respect to id
M
: M F.
Then M freely generates P, and for every t P, (M : t)
F
is a total M-
ary term operation on F, and only elements of P induce total M-ary term
operations on F.
Proof For the proof that M is a basis of P let f : M P be any mapping,
(see Figure 6.3) and let f

: domf

P be its closed M-initial extension.


Moreover, let p := (id
M
)

: P F. Since M freely generates F, p f has


a homomorphic extension g : F F (g[
M
= p f). Now p f

is a closed
M-initial homomorphic extension of p f, and g p is just an M-initial one;
therefore P domf

, i.e. f

P
: P P is a homomorphic extension of f
with respect to P.
Now, let t P be arbitrary, and v : M F any assignment (see Figure 6.4.
Then one has a homomorphic extension v : F F of v within F, and
v p : P F is an M-initial homomorphic extension of v as a mapping out
of P into F. Thus t domv

, i.e. v dom(M : t)
F
.
If, on the other hand, dom(M : t)
F
= F
M
for some t F(M, TAlg(S)), then
id
M
dom(M : t)
F
, i.e. t domid
M

= domp = P.
gure uncompleted
Figure 6.4: f6.4
December 8, 2002
6 The equational theory of partial algebras 209
6.3.3, 2.7.2 and 5.1.4.(ii) then yield:
Corollary 6.3.4 With the notation and the assumptions of 6.3.3 we have
ker id
M

= char
M
F = char
M
ind
F
M.
Denition 6.3.5 A congruence on a partial algebra A is called fully in-
variant iff , for every endomorphism f : A A, f
(2)
[] .
By Cong
fi
A (respectively Cong
c,fi
A) we denote the set of all fully invari-
ant (closed) congruence relations on A.
Proposition 6.3.6 For every A Alg(S) Cong
fi
A is a closure system
on A A.
Proof Obviously A A =
A
is fully invariant. Closedness with respect
to arbitrary meets can be shown using that the image of a meet is always
contained in the meet of the images.
Lemma 6.3.7 Let A Alg(S) and Cong
fi
A and s S such that
(A
s
A
s
) ,= A
s
A
s
. If M A freely generates A, then (M
s
M
s
) =

Ms
.
Proof If (m, n) (M
s
M
s
), m ,= n, then for every a, b A
s
the
mapping f : M A dened by f(m) := a, f(l) := b for all l M
s
m,
and f(l) := l otherwise, has a homomorphic extension

f : A A satisfying
(

f(m),

f(n)) = (a, b) , contradicting the assumptions on .
Proposition 6.3.8 Let f : A B be a homomorphism.
(i) If f is full and surjective, if ker f is fully invariant, and if M is a
basis of A, then f[M] is a basis of B. And if for some s S, #B
s
>
1, then f[
Ms
is injective. Moreover, M is B-free in A, and for each
homomorphism g : A B we have ker f ker g.
(ii) If B has a basis N, and if f maps some generating subset M of A
into N such that f[
Ms
is injective for every s S for which #B
s
> 1,
then ker f is fully invariant.
December 8, 2002
210 Chapter I Basic denitions and properties
gure uncompleted
Figure 6.5: f6.5
Proof Ad (i) Let us rst observe that f[M] generates B, since B = f[A] =
f[c
A
M] c
B
f[M]. If all phyla of B have at most one element, then
(i) is trivial. Hence assume that #B
s
> 1 for some s S; thus 6.3.7
implies that f[
Ms
is injective.
Let g : f[M] B be any mapping (cf. Figure 6.5). Then using the
Axiom of Choice and the surjectivity of f we can nd a mapping
h : M A such that f h = g f[
M
. Since M is a basis of A, h has
a homomorphic extension

h : A A. Since ker f is fully invariant,
we have

h
(2)
[ker f] ker f, and ker f ker (f

h) easily follows. Since


f is full and surjective, the Diagram Completion Lemma 2.7.1 implies
the existence of a homomorphism g : B B such that g f = f

h.
Thus g f[
M
= f

h[
M
= f h = g f[
M
, i.e. g[
f[M]
= g, showing that g
extends g. This shows that f[M] freely generates B.
Now, assume that g : M B is a mapping from M into B (cf. Fig-
ure 6.6). Since f[
Ms
is injective for every s S for which B has
a non-trivial phylum, there is h : f[M] B such that g = h f
M
.
Since f[M] is a basis of B, h extends to an endomorphism

h of B, and

h f is the homomorphic extension of g we have looked for. Moreover


ker f ker (

h f), implying the last statement.


Ad (ii) Let the assumptions of (ii) be satised, let h be any endomorphism
of A (cf. Figure 6.7) and assume (a, b) ker f. We have to show
that (f h)(a) = (f h)(b). Since f
Ms
is injective, whenever #B
s
>
1, we can nd a mapping g : f[M] B such that g f
M
= f h
M
(g : f(m) f(h(m)), m M). Next we extend g to a mapping g

:
N B, which then can be extended to a homomorphism g : B B,
i.e. g
f[M]
= g. Hence g f
M
= g f
M
= f h
M
, implying g f = f h, and
(f h)(a) = ( g f)(a) = ( g f)(b) = (f h)(b), what we wanted to show.
Corollary 6.3.9 (1) For any subclass K Alg(S) and for any set M
char
M
K = ker (r
M,K
)

,
and this is a closed and fully invariant congruence relation on domchar
M
K.
December 8, 2002
6 The equational theory of partial algebras 211
gure uncompleted
Figure 6.6: f6.6
gure uncompleted
Figure 6.7: f6.7
Proof Closedness of char
M
K and char
M
K = ker (r
M,K
)

are already known


(cf. 5.11.2 and 6.3.4).
Since F(M, K) is Mod Eeq K-freely generated by r
M,K
[M], it is also freely gen-
erated by r
M,K
[M]. If, for some s S, #F(M, K)
s
> 1, then 6.2.3.(v) implies
that r
M,K
[
Ms
is injective. Thus 6.3.8.(ii) implies full invariance of char
M
K.
6.3.2 and 6.3.4.(ii) together with 6.3.9 imply
Corollary 6.3.10 (2) For each subclass K Alg(S) and for each set M,
Eeq
M
K is a closed and fully invariant congruence relation on a relative sub-
algebra, say P, of F(M, TAlg(S)) such that M freely generates P.
6.4 Characterization of E-equational theories
6.4.1 We now have the tools to describe the closed sets of E-equations. The
description is a little bit more involved than it would be in the homogeneous
case.
In the following results we compare sets G
M
and G
N
of E-equations on
dierent sets M and N such that N is a subset of M; G
N
G
M
will
mean that (t, t

) [ (N : t
e
= t

) G
N
(t, t

) [ (M : t
e
= t

) G
M
;
domG
M
:= t, t

[ (M : t
e
= t

) G.
Theorem 6.4.2 (Second Birkho Theorem) Let X be any set of vari-
ables such that each phylum has cardinality not less than d

. Let G Eeq

X
and for each M X (#M < d
&
) let G
M
:= G F(M, TAlg(S))
2
. Then the
following statements are equivalent:
(i) Eeq Mod G = G.
December 8, 2002
212 Chapter I Basic denitions and properties
(ii) For each M X (with #M < d
&
), G
M
is a closed and fully invariant
M-initial congruence on a relative subalgebra domG
M
of F(M, TAlg(S)),
which is freely generated by M.
Moreover, for each N M the inclusion mapping from N into M
induces a homomorphism i
NM
: domG
N
domG
M
such that (a)
and (b) below are true:
(a) i
(2)
NM
[G
N
] G
M
(cf. 6.4.1);
(b) If v
M
[M] v
F(N,TAlg(S))
[domG
N
], then i
(2)
NM
[G
N
] = G
M
(c
F(M,TAlg(S))
N)
2
.
(iii) G = Eeq F(M, Mod G) [ M X,#M < d
&
.
(iv) G = Eeq F(M, Mod G) [ M :=

X
s
[ s S

[ S

S .
(v) G =

(M : t
e
= t

) [ (t, t

) ker (r
M,Mod G
)

[ M X, #M <
d
&
.
Moreover, G = Eeq Mod G always implies G = Eeq (domG
M
)/G
M
[ M X, #M < d
&
.
Proof Since Eeq
M
K = char
M
K (in the sense of 6.4.1), (ii) easily follows
from (i) by 6.2.3.(ii) and (iii) in connection with 6.3.9. Moreover, it is easily
seen in connection with 6.1.3 that (iii) implies (i); and the equivalence of
(iii), (iv) and (v) immediately follows from 5.11.4, 6.1.4, 6.1.5, 6.2.1, 6.3.4
and 6.3.9 with respect to 6.3.4 it does not matter, whether or not r
M,Mod G
is injective. And the last observation follows from (ii), 5.11.4, 6.2.1.(ii) and
the proof of 6.2.1.(i) for the operator H. Thus it remains only to show that
(ii) implies (iii):
Let K := F(M, Mod G) [ M X, #M < d
&
; then K Mod G implies
Eeq K Eeq Mod G G (cf. 6.1.3.(ii)). 6.2.1.(ii) implies that for every class
K

Alg(S) one has


Eeq K

ker (r
M,K
)

[ M X, #M < d
&

= Eeq F(M, K

) [ M X, #M < d
&
.
Set, for each M X with #M < d
&
, F
M
:= (domG
M
)/G
M
. Then (ii)
implies that F
M
is freely generated by M

:= nat
G
M
M, that G =

G
M
[
M X, #M < d
&
, and that always G
M
= ???(ker [M, F
M
])

. Thus, if
we can show, that, for each M, F
M
Mod G and that M

(Mod G)-freely
generates F
M
, then F
M

= F(M, Mod G), and (ii) (iii) has been proved.


December 8, 2002
6 The equational theory of partial algebras 213
gure uncompleted
Figure 6.8: f6.8
Asume f : N F
M
for some N X, #N < d
&
. We have to show that
N

ker f

(see Figure 6.8): Let N

:= N M, graph f

:= graph f graph id
M\N
.
Then f

: N

F
M
extends f to N

, and since N

M and v
N
[N

]
v
F(M,TAlg(S))
[domG], and since therefore (ii).(b) can be used, i
MN
induces an
injective and closed homomorphism i
#
MN
: F
M
F
N
, i.e. w.l.o.g. F
M
can
be considered as a subalgebra of F
N
and also domG
M
as a subalgebra
of domG
N
. Therefore ker (f

= ker (i
MN
(f

) ker nat
G
M
= G
M
,
where the last inclusion follows from 6.3.8.(i). Since (ii).(b) also applies to
the subset N of N

, we get in the same way G


N
ker f

, showing that
F
M
Mod G, since N and f were arbitrarily chosen. Then the denition
of F
M
shows that nat
G
M
[
M
: M F
M
is (Mod G)-universal i.e. F
M

=
F(M, Mod G), completing the proof.
Remarks 6.4.3 (i) In the homogeneous case, and when we forget about
the empty partial algebra, then we can always only consider E-equations
of the form (X : t
e
= t

) X the global set of variables , and


we get Mod G = Mod Eeq F(X, Mod G), moreover, for every class
K Alg(), Mod Eeq K = Mod Eeq
X
K = Mod Eeq F(X, K), and
Eeq
X
K is a closed and fully invariant congruence relation on the rel-
ative subalgebra dom(Eeq
X
K) which is freely generated by X. Hence
we have a strong analogy to the original Second Birkho Theorem for
total algebras, where Eeq K is always a fully invariant congruence re-
lation on F(X, TAlg()), and every fully invariant congruence on
F(X, TAlg()) is just Eeq F(X, TAlg())/. In the heterogeneous
case the situation becomes a little bit more complicated. Namely we
can restrict our considerations to the global E-equations with respect
to X only when we forget about all models which have at least one
empty phylum (or we have to consider these partial algebras as trivial
models contained in each model class of an E-equational theory in
this restricted sense). Then again the E-equational theories would be
in one-to-one correspondence to the closed and fully invariant congru-
ences on relative subalgebras of F(X, TAlg(S)) freely generated by X;
December 8, 2002
214 Chapter I Basic denitions and properties
while in the general situation which we have considered above, we need
a set K with usually more than one partial algebra in order to get for
a given set G of E-equations that Eeq Mod G = Eeq K.
(ii) Observe that the fully invariant congruence relations of a partial alge-
bra Aare exactly the congruences of the partial algebra ((A, (
A
)

( e [
e End A)), v
A
), i.e. where we enrich the structure by all endomor-
phisms of A. Since P
M
:= dom(Eeq
M
K) is always freely gener-
ated by M, we get for any t P and for any endomorphism e of P
that e(t) = (M : t)
P
(e[
M
), and the condition of full invariance for
Cong
fi
P is equivalent with the fact that for any (t, t

) and
for every v : M P (which by assumption induces an endomorphism
v : P P) we have ((M : t)
P
(v), (M : t

)
P
(v)) , i.e. is closed
with respect to substitution as this fact is usually described, since
(M : t)
P
(v) is obtained from the term t by replacing each variable x
occurring in t by v(x).
We realize that this fact of closedness with respect to substitution,
which is well known for equational theories of total algebras, also ex-
tends to E-equational theories of partial algebras. We shall see below,
how one can exploit this in order to construct E-equational theories
from a given set of E-equations by a set of derivation rules. In compar-
ison with the total case the situation for homogeneous partial algebras
is more complicated, as one has to construct dom(Eeq Mod G) besides
Eeq Mod G (with respect to a set X of variables). In the heterogeneous
case one then also has to observe the interaction of E-equations on
dierent sets of variables.
The following lemma will also be useful in later sections.
Lemma 6.4.4 Let G Eeq

X
be any set of E-equations, and let
G := (M : t

) [ there is (M : t
e
= t

) G such that t

(t t

)
be the set of all M-subterms of M-terms in G.
Let Y be any set, let
Q := ((F(Y, TAlg(S)), ((M : t)
F(Y,TAlg(S))
)
(M:t)G
), v
F(Y,TAlg(S))
),
P := c
Q
Y , and let P be the corresponding relative subalgebra of F(Y, TAlg(S)).
Then P is freely generated by Y .
December 8, 2002
6 The equational theory of partial algebras 215
Proof Let v : Y P be any mapping; we have to show that P domv

,
(where v

= (v

)
P
), and we do this by algebraic induction on x P:
If x Y , then x domv

. Assume x P Y ; then there are (M :


t) G and a P
M
such that x = (M : t)
F(Y,TAlg(S))
(a). Assuming as
induction hypothesis that a (domv

)
M
, we get for the global homomorphic
extension v : F(Y, TAlg(S)) F(Y, TAlg(S)) that (M : t)
F(Y,TAlg(S))
( v(a)) =
v((M : t)
F(Y,TAlg(S))
(a)) (cf. 5.7.5), but by assumption and the denition of P,
(M : t)
F(Y,TAlg(S))
( v(a)) as well as (M : t)
F(Y,TAlg(S))
(a) both belong to P;
therefore x = (M : t)
F(Y,TAlg(S))
(a) P domv

.
Remark and Example 6.4.5 The procedure sketched in Lemma 6.4.4 may
be used to generate the elements of T
Y
(G) := domEeq
Y
Mod G, but unfor-
tunately the smallest fully invariant congruence relation G

generated by G
in P from Lemma 6.4.4 in general need not be closed, as the reader may nd
out by considering the following (homogeneous) example:
:= (2, 2), := (,

),
G := (xx, x), (xxxx

, xxxx

).
Here (x : xx) and (x : xxxx

) are the only total unary term


operations directly derivable fromG, however, as the calculation in 6.4.11.(iv)
(3
rd
example) of E-equations derivable from G shows, (x : xx

) will also
be a total term operation in domEeq
{x}
Mod G, and this is just based on
rule (7) in 6.4.8 below corresponding to the fact that Eeq
{x}
Mod G is a closed
initial congruence relation on its domain.
Nevertheless we have the following
Proposition 6.4.6 With the same assumptions and notation as in Lemma 6.4.4
we get: Let

G be the smallest fully invariant congruence relation on P P be-
ing dened as in 6.4.4 generated by G, i.e. containing

Y,G
:= ((M : t)
P
(v), (M : t

)
P
(v)) [ (M : t
e
= t

) G and v : M P .
Moreover, let G

:= Eeq
Y
Mod G and P

:= domG

.
Then there is a natural isomorphism j : P/

G P

/G

= F(Y, Mod G)
induced by

G[y] G

[y] (y Y ).
December 8, 2002
216 Chapter I Basic denitions and properties
Proof Obviously

G G

; therefore the homomorphism j exists according


to the Diagram Completion Lemma 2.7.1 and 2.7.2. From 6.4.4 and 6.3.8.(i)
we realize that P/

G is freely generated by nat
G
[Y ]. Applying 6.3.9 to K :=
P/

G and using 4.3.8 we infer that G

:= ker (nat
G
)

is a closed and fully


invariant congruence on an Y -initial relative subalgebra freely generated
by Y of F(Y, TAlg(S)), which satises
Y,G
G

G G

. Moreover,
the denition of P and

G imply that P/

G Mod G, and therefore j has an
inverse, i.e. it is an isomorphism.
Remark and Example 6.4.7 (i) Thus, if one only wants to construct
F(Y, Mod G) and is not interested in Eeq
Y
Mod G, then 6.4.6 suces. If
one really has to construct Eeq
Y
Mod G from G, one may apply to G the
complete system of rules of derivation of E-equations following below
in 6.4.8. We only formulate them for the case of E-equations with
d
&
-small sets of variables, but it is easy to see that one can also get
the following results allowing larger subsets of the underlying set X of
variables. One should also observe, that the rules below usually are of
only theoretical interest, if the signature allows innitary operations,
or if we have innitely many sorts. Nevertheless the following results
hold even for the general case.
(ii) When we apply 6.4.4 and 6.4.6 to the example in 6.4.5, then it is easy
to realize that a model F

= F(x, Mod G) can be obtained as follows:
F := n [ n <
0
, where x corresponds to 0.

F
(z) := z and

F
(z) := z + 1 for all z F.
Such a model can then often be used for an easier computation of
consequences of given axioms.
6.4.8 (Rules of Derivation for E-equations) Let X be any set of vari-
ables such that each phylum of X has cardinality at least d

. Let G Eeq

X
be any set of E-equations. An E-equation (M : t
e
= t

) Eeq

X
is said to be
derivable from G in symbols: G (M : t
e
= t

) , iff there is a d
&
-small
subset G

of G (i.e. #G

< d
&
) such that (M : t
e
= t

) is derivable from G

in a sequence of length less than d


&
of applications of the following rules
of derivation of E-equations. Notice that letters x stand for elements of X,
t, t

, t

, t
i
, t
k
, etc. stand for terms (i.e. roughly speaking, for elements of
F(X, TAlg(S))), M, N, M
i
, etc. stand for d
&
-small subsets of X, while H,
December 8, 2002
6 The equational theory of partial algebras 217
H
i
, etc. stand for d
&
-small subsets of G, and for any index set I we assume
that #I < d
&
. Above a line we write the list of premises of a rule, while
below a line there stands the conclusion if a rule does not have any line,
then we have an empty set of premises. The rules should be read in the
following way: If one knows about the derivability of all the sequences above
a line, then one may conclude that the expression to the right of below
the line is derivable from the left hand side.
(E0)
H
i
(M
i
: t
i
e
= t

i
) (i I)
H (M
i
: t
i
e
= t

i
) [ i I (M : t
e
= t

)
H

H
i
[ i I (M : t
e
= t

) Transitivity of .
(E1)
(M : x
e
= x) (x M X).
(E2)
H (M : t
e
= t

), whenever (M : t
e
= t

) H.
(E3)
H (M : t
e
= t

)
H (M : t

e
= t)
(E4)
H
1
(M : t
e
= t

)
H
2
(M : t

e
= t

)
H
1
H
2
(M : t
e
= t

)
(E5)
H (M : t
e
= t

)
H (M : t
e
= t)
(E6)
H (M : (t
k
[ k ())
e
= (t
k
[ k ()))
H (M : t
k
e
= t
k
) ( , k ())
December 8, 2002
218 Chapter I Basic denitions and properties
(E7) Let and k () then
H
k
(M : t
k
e
= t

k
)
H (M : (t
k
[ k ())
e
= (t
k
[ k ()))
H

H
k
(M : (t
k
[ k ())
e
= (t

k
[ k ()))
(E8) Let f : M F(M, TAlg(S)), x t
x
(x M) be any assignment
observe v
X
(x) = v

X
(t
x
) (x M) , and let

f : F(M, TAlg(S))
F(M, TAlg(S)) be its homomorphic extension i.e. in each t F(M, TAlg(S))
each x M is substituted by t
x
; then
H (M : t
e
= t

)
H
x
(M : t
x
e
= t
x
) (x M)
H

H
x
[ x M (M :

f(t)
e
=

f(t

))
(E9)
H (M : t
e
= t

)
H (M N : t
e
= t

)
(E10) Let M N X, t
i
F(M, TAlg(S)) (i I) and t, t

F(N, TAlg(S))
such that N (t t

) M and v
X
[N] v

X
[M

t
i
[ i I ];
then
H (N : t
e
= t

)
H
i
(M : t
i
e
= t
i
) (i I)
H

H
i
[ i I (M : t
e
= t

)
Remarks 6.4.9 (i) Rules (E8), (E9) and (E10) together can be replaced
by the following rule
(E8*) Let f : M F(N, TAlg(S)), x t
x
(x M) be any as-
signment observe v
X
(x) = v

X
(t
x
) (x M) , and let

f : F(M, TAlg(S)) F(N, TAlg(S)) be its homomorphic ex-


tension i.e. each x M in each t F(M, TAlg(S)) is substi-
tuted by t
x
; then
December 8, 2002
6 The equational theory of partial algebras 219
H (M : t
e
= t

)
H
x
(N : t
x
e
= t
x
) (x M)
H

H
x
[ x M (N :

f(t)
e
=

f(t

))
The argumentation in the derivation of (E8*) from the other rules and
from our knowledge about homomorphisms between term algebras fol-
lows the idea of the proof of the Second Birkho Theorem, the other
direction is even simpler; we leave the details as an exercise.
(ii) In the case of total (S-) algebras the above system of rules could be
simplied: (E0), (E2), (E3), (E4) and (E9) would remain unchanged,
(E5) and (E6) would become superuous, and we could simplify the
other rules to:
(E1t)
(M : t
e
= t) t F(M, TAlg(S)), M X
(E7t)
H
k
(M : t
k
e
= t

k
) (k ())

H
k
[ k () (M : (t
k
[ k ())
e
= (t

k
[ k ()))
(E8t) Let f : M F(M, TAlg(S)) be any assignment, and let

f : F(M, TAlg(S)) F(M, TAlg(S)) be its homomorphic ex-


tension, i.e. each x M in each t F(M, TAlg(S)) is substi-
tuted by f(x), then
H (M : t
e
= t

)
H (M :

f(t)
e
=

f(t

))
(E10t) Let M N X and t, t

F(M, TAlg(S)) such that v


X
[N]
v
X
[F(M, TAlg(S))], then
H (N : t
e
= t

)
H (M : t
e
= t

)
These are except for some dierences in the formulation exactly the
rules presented in J.A.Goguen & J.Meseguer [GoMe81].
December 8, 2002
220 Chapter I Basic denitions and properties
(iii) Let us present a very simple example showing that we really have to
take care of the cardinality #S of the set S of sorts in connection with
the cardinalities of our sets of variables:
Let S := N be the set of natural numbers, :=
n
,

n
[ n N,
and := (0)

a type only containing nullary constants such that

n
and

n
are of sort n, and let K := A
n
[ n N, where A
n
:=

m
,

m
[ m N n , v
n
:= v
An
being the restriction of v
F(,TAlg(S))
to A
n
, and all possible constants are dened in A
n
:
A
n
m
:=
m
and

m
A
n
:=

m
. If we consider only E-equations with a nite set of vari-
ables, then K would satisfy no non-trivial E-equation, and we would
have Mod Eeq
nite
K = TAlg(S), while in HoT K those partial alge-
bras, all of whose phyla are non-empty, are isomorphic to S, contradict-
ing the First Birkho Theorem (and we actually have the non-trivial
E-equations (N :
m
e
=

m
) (m N)).
But then it is easy to realize in connection with rules (E8) and (E10)
that we sometimes also need innite premises of the rules in order to
be able to apply the rules for substitution and restriction of variables
properly.
(iv) Let us briey comment upon the connection between 6.4.2.(ii) and the
rules in 6.4.8, where we assume that all the sets H and H
k
with any
index k are assumed to be equal to an arbitrary but xed set G of
E-equations with variables in X, where G

be the set of all equations


derivable from G according to the above rules, T(G) := (M : t) [ M
X, #M < d
&
, t F(M, TAlg(S)), G (M : t
e
= t) , and where M is
always an arbitrary d
&
-small subset of X:
(E0) just means that, if we add to G E-equations already derived
from G and get a set G
#
, then all that is derivable from G
#
is al-
ready derivable from G. Hence we can add to G E-equations already
derived from G in order to derive new ones from G.
(E1) says that with respect to some set M of variables at least
the TE-statements (M : x
e
= x) are always derivable without extra
premises, i.e. (M : x) T(G). Moreover it says that G is reexive on
the set of variables.
(E2) says that all E-equations in G are always derivable from G, i.e.
G G

.
December 8, 2002
6 The equational theory of partial algebras 221
(E3) says that G

M
is always a symmetric, (E4) that it is a transitive
and (E5) that it is a reexive relation on its domain, which is T(G)
M
:=
(M : t) [ (M : t) T(G) , i.e. that it is an equivalence relation on
its domain, and T(G) = domG

M
.
(E6) corresponds to the fact that domG

M
is always an M-initial seg-
ment of F(M, TAlg(S)).
(E7) means that G

M
is a closed congruence relation on domG

M
.
(E8) corresponds to the fact that G

M
is fully invariant on its domain,
and that this is freely generated by M.
(E9) guarantees that 6.4.2.(ii).(b) is satised by G

and (E10) corre-


sponds to 6.4.2.(ii).(a).
Theorem 6.4.10 (Soundness and Completeness) The set of rules pre-
sented in 6.4.8 for the derivation of E-equations is correct (sound) and com-
plete, i.e. if G Eeq

X
is any set of E-equations, and if G

is the set of all


E-equations derivable from G according to 6.4.8, then G

= Eeq Mod G.
Proof Soundness just means (a) G

Eeq Mod G =:

G, while completeness
means (b) Eeq Mod G G

.
The discussion in 6.4.9.(iv) of the rules of derivation in 6.4.8 shows that they
do nothing else and nothing more but to enforce properties on G

,
which have have to hold for G

according to 6.4.2.(ii), thus they are sound,


and (a) is proved.
On the other hand, the fact that G

has according to 6.4.9.(iv) all the


properties required in 6.4.2 for G

, and that G

is the smallest set with these


properties, and contains G, also proves (b).
Remarks 6.4.11 (i) In subsection 14.3 we shall return to this topic in a
more general context.
(ii) The above rules now allow a stepwise procedure to construct F(M, Mod G)
for some given set G of E-equations, but for lack of space we do not go
into more details here.
(iii) If one wants to present an explicit description of a free partial alge-
bra, it is often possible to nd term operations which can be used to
December 8, 2002
222 Chapter I Basic denitions and properties
generate the carrier set as a partial Peano algebra such that the fun-
damental partial operations can then be dened by using the given
E-equations. See e.g. 6.4.7.(ii), where (x : xx

) is such a term oper-


ation generating a partial Peano algebra, which yields the underlying
set of F(x, Mod G)). Also in [B68], Satz 2.2, such a construction has
been eectively used to characterize the free algebras of some special
variety.
In 6.2.6.(iii) we have given the E-equational axioms for Stack(Nat),
the stack automata of integers. When we want to characterize F :=
F(, Stack(Nat)) in the way mentioned above, we rst use the con-
stant 0 and the operation Succ to generate the underlying set of
integers, since the rst axiom says that 0
F
has to be dened, while
the second axiom in 6.2.6.(iii) requires that Succ
F
has to be total
(cf. 6.4.8.(E6) and 6.1.2.(iii)). Pred
F
is then dened by again us-
ing the (second) axiom (y : Succ(Pred(y))
e
= y), however it only
requires the existence of Pred
F
(z), if z is a value of Succ
F
, and there-
fore Pred
F
(0
F
) remains undened. The axiom ( :
e
= ) says that
the empty stack
F
has to exist, and the last two axioms imply that
Push
F
has to be total (as a subterm of a term in an axiom. Thus we can
use , the terms Succ
n
(0) (for n <
0
) where Succ
0
(0) := 0 and
Succ
n+1
(0) := Succ
n
(Succ(0)) and what we can generate from
these terms in the Peano algebra by only using Push in order to gen-
erate the elements of sort States in F. Push
F
is then dened canon-
ically; and for the denition of Top
F
and Pop
F
we can use the axioms
(x, y : Push(Top(x, y))
e
= y) and (x, y : Push(Pop(x, y))
e
= x),
which show that Top
F
(
F
) and Pop
F
(
F
) will not be dened, since
the axioms only require that these operations are dened on values
of Push, and in the Peano algebra is no such value and therefore
neither in F.
(iv) For the example in 6.4.5 we derive some E-equations. Let us recall:
:= (2, 2), := (,

),
G := (xx, x), (xxxx

, xxxx

).
Behind the derivation we always mention the line and the rules which
December 8, 2002
6 The equational theory of partial algebras 223
we have used:
1. G xx
e
= x (E2)
2. G xxxx

e
= xxxx

(E2)
3. G xxxx

e
= xx

(E7) for 1., 1. and 2.


4. G xx
e
= xx (E5) for 1.
5. G xxxx
e
= xx (E8) for 1., 4. (xx substituted for x in 1.)
6. G xxxx
e
= x (E4) for 1., 5.
7. G xx

e
= xxxx

(E3) for 3.
8. G xx

e
= xx

(E5) for 7.
9. G xx

xx

e
= xx

(E8) for 1. and 8.


etc..
December 8, 2002
224 Chapter I Basic denitions and properties
December 8, 2002
Chapter II
A Basis for a two-valued model
theory for partial algebras
In Chapter I basic universal algebraic properties of many-sorted partial al-
gebras have been investigated, ending with a description of an (existence-)
equational theory. Moreover several dierent properties for homomorphisms
and in consequence of subobjects have been presented. However, it is
well known (see e.g. [TcWaWr78]) that often algebraic specication of ab-
stract data types needs innitely many equations, while it may be possible
to specify in an equivalent way by nitely many quasi-equations (elementary
implications). And in connection with partial algebras elementary implica-
tions (and even negations of TE-expressions, as the axiomatization of elds
in 13.4.1 shows) are even more necessary than in the total case for the speci-
cation of a class of partial algebras (in which the initial object may be meant
to represent the intended abstract data type).
Elementary implications and their category theoretical equivalent of
epimorphisms are also a useful tool to describe and handle most of the
important properties of homomorphisms (together with the concept of fac-
torization systems). Thus it seems worth-while to have at hand a more
comprehensive model theoretic language and model theoretic and cate-
gory theoretic concepts than only that of existence equations (respectively
only those of mono-, epi- and isomorphisms). Moreover, we think it useful
to have a language at hand which also allows a treatment of innitary op-
erations and innitary conjunctions (and quantications) of formulas, since
these more general concepts will e.g. allow an easy description of -complete
posets and (partial) algebras (see 13.4.4 and 13.4.5) and of properties of the
225
226 Chapter II A Basis for a two-valued model theory for partial algebras
class of all these structures.
This chapter and the following one will therefore be dedicated to develop
the fundaments of the theory of universal Horn formulas based to a large
extent on category theoretical methods, as they have been developed in
part independently by B.Banaschewski and H.Herrlich (cf. [BaH76])
for the total (homogeneous) case and by R.John (see [J77]) as well as by
H.Andr eka and I.N emeti (in part supported by I.Sain; see e.g. [AN83],
[AN82], and [NSa82] , which will become of special importance in Chap-
ter III).
In this chapter we mainly present some basic concepts which will allow a
better description and understanding of many features and concepts occur-
ring in connection of partial algebras:
In section 7 we present a comprehensive formal language together
with a semantics and some fundamental theorems for a two-valued model
theory for (possibly innitary) partial algebras, which is a dialect of the
usual rst order language, i.e. when all arities and #S are nite. We do not
treat relations separately, since we do not want to burden the presentation
still more, and since relations can also be considered as partial operations
(partial projections) as indicated in 1.1.3.
In section 8 elementary implications and universal Horn formulas are con-
sidered: After having distinguished several dierent kinds of universal Horn
formulas some of them only arising in connection with partial algebras, like
TE-expressions, NTE-expressions, ECE-equations, etc. we encode them
algebraically and category theoretically, and investigate the preservation of
universal Horn formulas by algebraic operators, which can be considered as
a basic tool for the description of properties of algebraic operators.
In section 9 preservation and reection of (universal Horn) formulas by
mappings is studied in some detail and several examples are studied.
In section 10 the category theoretical concept of factorization systems is
studied, in particular the close relationship of the diagonal-ll-in property
with the reection of positive universal Horn formulas by mappings. It is
shown by the presentation of several examples that each category of all partial
S-algebras of some given signature allows quite a lot of relevant and non-
trivial factorization systems.
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 227
7 A rst order language for (innitary) par-
tial (S-) algebras
As in Chapter I we always tacitly assume that every partial algebra (or
set), say A, is endowed with a sort homomorphism v
A
: A S into a
heterogeneous sort algebra S (see ??), and that all mappings are compat-
ible with these sort mappings. We present here a tentative formulation of
a two-valued logic for (possibly innitary) partial algebras, whose presenta-
tion is strongly oriented to compatibility with the category theoretical for-
mulation in [AN79a]. The nitary homogeneous case has intensively been
discussed in [B82] (the basic ideas for such a logic are already contained
in H.Thiele [Th66]), while the nitary heterogeneous case is investigated in
a recent paper of H.Kaphengst [Ka82]. We have been informed that our
logic here has been intensively studied by Zs.Markusz in [Mr83a] for total
heterogeneous algebras.
7.1 The syntax
We start with the denition of a formal possibly innitary language for
partial heterogeneous algebras. It is designed to help formulate properties
of partial S-algebras and of homomorphisms between them. Although later
on we shall restrict considerations to universal formulas and in particular
to universal Horn formulas we also include a proposal for a treatment
of quantiers. Since the language will be used in this book for description
rather than for algorithmic calculations, innite arities and formulas will not
bother too much.
Denition 7.1.1 Let = (())

be any similarity type, d its cardinal


dimension, let S be any heterogeneous sort algebra of type , let be any
ordinal, # = c, and q and v be any cardinals, v regular, such that, for
d
&
= maxd,
0
, (#S)
+
, d
&
, q, c c; and let X be any set such that for
each sort s S each phylum X
s
has cardinality v. We are going to dene a
language L := L
c,q,d,v
(, S, (X, v
X
)) =: L
c,q
(S, X) as follows:
(i) The alphabet of L is given by
(a) The logical symbols, i.e.
December 8, 2002
228 Chapter II A Basis for a two-valued model theory for partial algebras
- the elements x of the set X of variables; each variable x has
a sort v
X
(x);
- all subsets of X of cardinality less than v;
- sentential connectives , ,
_

(for , , and
_

see 7.1.2.(v));
- quantiers , ,

;
- the identity symbol
e
=;
- brackets ( , ) , colon : (vertical line [ and curly brackets , );
(b) non-logical constants
- constants for partial operations ( ), which shall also in-
clude individual constants as nullary operation symbols (see 7.1.2.(iii));
(- the element symbol).
(ii) Terms of L are of the form (M : t) for elements t F(M, TAlg(S)),
where M X with #M < v.
(iii) Atomic formulas of L are all E-equations (M : t
e
= t

), where t, t


F(M, TAlg(S)) with v

M
(t) = v

M
(t

) and M X, #M < v (one


might be tempted to write (M : t)
e
= (M : t

), but we shall always


use a notation like the above one).
(iv) Formulas of L are dened recursively as follows:
- each atomic formula of L is a formula of L;
- if (M : ) is a formula of L, then (M : ) is a formula of L;
- if (M : ) and (M

) are formulas of L, then (MM

: (

))
is a formula of L;
- if T is a non-empty set of formulas of L, #T < c, then (M : )
is a formula of L, where M :=

N [ (N : ) T , and
:=
_

[ (N : ) T (note that #M < v);


- if M X, #M < v, x M and Y M with #Y < q, and if
(M : ) is a formula of L, then (M x : x ) and (M Y :

Y ) are formulas of L;
- if (M : ) is a formula of L, and if N X such that #N < v,
then (M N : ) is a formula of L.
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 229
Remarks 7.1.2 (i) Actually a formula as dened above consists of two
parts: the set of variables it is related to, and the second compo-
nent, which we shall call the body of the formula (or basic formula).
We have to refer explicitly to M because of the heterogeneous sit-
uation allowing empty phyla (see Markusz [Mr83] and [Mr83a] or
Matthiessen [Mt76] for more details).
(ii) The last possibility of the formation of formulas has been included in
order to show that variables which have been taken away in connection
with quantication can nevertheless occur in the set of variables of a
formula, where the body has been quantied.
(iii) We could treat individual constants separately, as it is often done, but
we try to get along without separate constants, to avoid further details;
therefore we shall have to include axioms of the form ( : c
a
e
= c
a
), when
c
a
is a constant which has to be interpreted see the denition of an
interpretation below.
(iv) Notice that the logical symbols , , ,
_

, x, and x

Y can be
considered as operations on L, and except for
_

they have a xed


arity and satisfy the Peano axioms, while we have simplied the use
of
_

, which is often (usually in formal presentations) replaced by a


family (
_

[ 0 < < ), where is the smallest ordinal of cardinality c,


and
_

has arity . For we wanted to avoid the inclusion of too much


arithmetic of ordinals (a more rigorous presentation of such languages
can be found e.g. in [K64]; for a treatment of innite arities of non-
logical constants see [HnMoT85]). We do not want to deal here with
more rigorous logical problems (cf. [ANSa82]), but we just want to have
a language at hand, which is able to express some features encountered
in connection with partial algebras. It should be observed that, already
with our approach, one can make proofs or denitions by algebraic
induction or recursion.
(v) In order to have not too many basic logical operators we shall take
in what follows as basic operators only (negation),
_

(c-restricted
conjunction) and

(q-restricted generalization) and dene the others


used above and further usual operators as follows, whenever the right
December 8, 2002
230 Chapter II A Basis for a two-valued model theory for partial algebras
hand sides are dened:
(

) :=
_

(binary conjunction)
(

) :=
_

=
_

(binary disjuntion)
_

T :=
_

[ T (c-restricted disjunction)
(

) :=
_

= (

) (implication)
(

) := ((

) (

)) (equivalence)
x :=

x (generalisation of one variable)


x :=

x (existential quantication with one variable)

Y :=

Y (q-restricted existential quantication).


Often we shall include and

into our basic considerations, es-


pecially when we start considering semantics. We denote the extended
language in the same way as the original one.
(vi) Actually L is a disjoint family L = ( L
M
[ M X, #M < v ), where
each L
M
is the set of all L-formulas whose set of variables is M. Only
quantication (and adding new variables) change the set of variables.
(vii) In later applications the above parameters will often be chosen as fol-
lows: c = q = v = d
&
.
Denition 7.1.3 For every formula of L := L
c,q,d,v
(, S, (X, v
X
)) we dene
the sets var , fvar , and subf of variables, free variables, subterms and
subformulas of , respectively, recursively on the structure of (we omit one
pair of brackets, if there would be two of them):
(i) If = (M : t
e
= t

), i.e. is atomic, then


:= (M : t

) [ t

t t

,
fvar := var := X t

[ (M : t) ,
subf := .
(ii) If = (M : ), then
:= (M : ), fvar := fvar (M : ),
var := var (M : ), subf := subf (M : ).
(iii) If = (M :
_

T), #T < c, T a set of basic L-formulas, then


:=

(M : ) [ T , fvar :=

fvar (M : ) [ T ,
var :=

var (M : ) [ T , subf :=

subf (M : ) [ T .
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 231
(iv) If = (M :

Y ), Y X, #Y < q, then
:= (M Y : ), fvar := fvar (M Y : ) Y,
var := var (M Y : ), subf := subf (M Y : ).
If = (M : ), then we dene
:= t [ (M : t) (M : ) , fvar := fvar (M : ),
var := var (M : ), subf := [ (N : ) subf (M : ) .
A variable x X is called a bound variable of a formula (M : ), if there
is a subformula (N :

Y ) subf (M : ) such that x Y ; and by


bound (M : ) we denote the set of all bound variables in (M : ).
The following results are easily proved either by observing that t
F(M, TAlg(S)) implies fvar (M : t) M, or by induction on the structure
of a formula:
Lemma 7.1.4 For every formula = (M : ) one has:
(i) fvar var = X ; and fvar = var holds iff for every subfor-
mula subf , which is of the form = (N :

), there holds
Y var (N Y :

) = .
(ii) fvar (M : ) M, but not necessarily var (M : ) M.
(iii) #M, #var() < v and #X = v.
Since substitution in formulas of variables by terms may cause conicts
between free and bound variables (see 7.1.6 below), we rst consider the
situation using admissible replacements of variables by terms , where
such a conict is avoided:
Denition 7.1.5 (i) Let (N : ) L, let M X, #M < v, and let f :
N F(M, TAlg(S)) be any (S-) mapping (!) and

f : F(N, TAlg(S))
F(M, TAlg(S)) its homomorphic extension. We say that f is admis-
sible for (N : ), if for every subformula of (N : ) of the form
(N

Y ) we have Y

f(x) [ x fvar (N

Y ) = and
fvar (N

Y ) bound (N

Y ) = .
December 8, 2002
232 Chapter II A Basis for a two-valued model theory for partial algebras
(ii) Let f : N F(M, TAlg(S)) be an admissible mapping for (N : ),
then we say that we get (M : f()) by (simultaneous uncritical) substi-
tution from (N : ) with respect to f, if f() is the following formula:
(a) If = t
e
= t

, then f() :=

f(t)
e
=

f(t

);
(b) if = , then f() := f();
(c) if =
_

T, then f() :=
_

f() [ T ;
(d) if =

Y , then f() :=

Y (f
Y
(), where f
Y
: N Y
F(M Y, TAlg(S)), dened by f
Y
(x) := f(x), if x N Y , and
f
Y
(x) := x, if x Y .
We write Subst((N : ), f, (M : f())) in order to indicate this situa-
tion.
Remarks 7.1.6 (i) Notice that f in 7.1.5 is always admissible for (N : ),
when does not contain any quantier i.e. no basic subformula of
the form

Y ; such a formula will be called quantier free.


(ii) Moreover observe that f
Y
in 7.1.5.(d) is admissible for (N Y : ), if
f is admissible for (N : ), since the concept of subformula is transi-
tive. Admissibility just means that the corresponding substitution does
not cause any collision between free and bound variables occurring
in the new formula. A variable x X is bound in a formula (N : ),
if has a subformula

Y such that x Y fvar .


(iii) One could dene a generalized substitution for (N : ), which is always
applicable, by rst changing bound variables to avoid collisions in
connection with f; but we do not go into details here, since later on we
shall mainly deal with formulas without quantiers, or only with such
of the form

Y , where is quantier free.


(iv) It should be observed that the relation to be a subformula of some
formula is a quasi-order and indeed even a partial order on L (the proof
is similar to the one for terms; but observe that
_

only depends on
the set of its arguments and not on a special sequence formed by them;
therefore the original Peano algebra has already been factored with
respect to a fully invariant and closed congruence.)
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 233
(v) If we only deal with homogeneous partial algebras, then we can simplify
considerations by only considering formulas (X : ) for the whole set
of variables and allowing the empty partial algebra as a trivial model
for every theory, i.e. we can then restrict considerations to the basic
formulas.
7.2 The semantics
Based on the semantics already introduced in section 6 for E-equations we
now briey sketch the fundaments of semantics for all of L = L
c,q
(S, X).
Denition 7.2.1 Let A Alg(S), and let (M : ) be an L-formula.
(i) Any mapping w : M A (compatible with the sort mappings!) is
called an assignment or valuation for (M : ) with values in A.
(ii) Let w : M A be an assignment with values in A. The corresponding
interpretation I
w
(or mostly only denoted by w

) is then dened as
follows:
- I
w
(x) := w

(x) = w(x) for every variable x in M, and I


w
(x) is
undened for x X M.
- I
w
() :=
A
for every fundamental operation symbol .
- If (N : t) is any term, then (N : t) is interpreted by I
w
iff M = N
and t domw

; and if this is the case, then I


w
(M : t) := w

(t) =
(M : t)
A
(w) according to section 5.
We could extend the interpretation in an obvious way to formulas, but
the following denition of satisfaction of a formula in A with respect to an
interpretation is understandable without this.
Denition 7.2.2 Let (M : ) be an L-formula, and let A Alg(S).
(i) Let w : M A be any valuation with values in A; then we dene re-
cursively, what it means that A satises (M : ) under the valuation w
(i.e. that (M : ) is satised in A for w) in symbols: A [= (M : )[w]:
(a) If = t
e
= t

, then A [= (M : t
e
= t

)[w] iff (t, t

) ker w

(cf. 6.1.1.(ii)).
December 8, 2002
234 Chapter II A Basis for a two-valued model theory for partial algebras
(b) If = , then A [= (M : )[w] iff it is not true that A [= (M :
)[w].
(c) If =
_

T, #T < c, T a set of basic L-formulas, then A [= (M :


_

T)[w] iff for every T there holds A [= (M : )[w].


(d) If =

Y , #Y < q, then A [= (M :

Y )[w] iff for every


mapping w

: M Y A satisfying w

[
M\Y
= w[
M\Y
one has
A [= (M Y : )[w

].
(ii) (M : ) is valid in A in symbols: A [= (M : ) , iff one has
A [= (M : )[w] for every assignment w : M A.
7.2.3 Thus we have A [= (M : ), whenever there is no assignment from M
into A, i.e when A has empty phyla, and at least one of the corresponding
ones of M is non-empty. In the proofs below we shall usually omit the trivial
case based on this observation.
Lemma 7.2.4 Let (M : ) L, A Alg(S), Y M, #Y < q.
(i) A [= (M : ) iff A [= (M Y :

Y ); hence, if #M < q, then


A [= (M : ) iff A [= ( :

M).
(ii) If w : M A is any assignment for (M :

Y ), and if N := M Y ,
then A [= (M :

Y )[w] iff there exists an assignment w

: N A
such that w

[
N\Y
= w[
N\Y
and A [= (N : )[w

].
(iii) If = (t
e
= t

) for some t, t

F(M, TAlg(S)), then we have for any


w : M A: A [= (M : (t
e
= t

))[w] iff t / domw

or t

/ domw

or
w

(t) ,= w

(t

); in particular, A [= (M : (t
e
= t))[w] iff t / domw

(i.e. iff t is not interpreted by I


w
).
Proof (i), (ii) and (iii) are straightforward consequences of the preceding
denitions, and we only prove (ii):
We have for any assignment w : M A that A [= (M :

Y )[w] iff
it is not true that A [= (M :

Y )[w], iff it is not true that for all


assignments w

: N = M Y A satisfying w

[
N\Y
= w[
N\Y
it is the case
that A [= (N : )[w

], iff there exists an assignment w

: N A such that
w

[
N\Y
= w[
N\Y
and A [= (N : )[w

].
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 235
Remarks 7.2.5 (i) A broad discussion of the properties of existence equa-
tions and the motivation of this kind of logic in general is contained
in the survey article [B82] of Burmeister; therefore because of lack
of space we do not go here into further details. Let us only recall
that we have hidden existential quantiers in the E-equations: Let us
consider every ()-ary operation
A
of output sort s of some par-
tial algebra (A, v
A
) as a (() + 1)-ary relation, say

:= graph
A
,
on (A, v
A
), and set A

:= ((A, (

), v
A
); then we have for any as-
signment w : M A that A [= (M : ( x
k
[ k () )
e
= ( x
k
[
k () )[w] iff for some variable y v
1
X
(s) x
k
[ k () ,
A

[= (M :

(( x
k
[ k () ), y))[w]. Thus the satisfaction of
the TE-expression for A is equivalent to the satisfaction of an existen-
tial formula with respect to A

. As sketched in [B82], 2.18, this can be


extended to E-equations between arbitrary terms, where we get quite
complex existential expressions.
(ii) Let us further observe that A [= ( : x x
e
= x) is equivalent with
the fact that, for s := v
X
(x), A
s
,= , i.e. the phylum of A of the
sort s of x is non-empty. Thus our language L is not only strong
enough to talk about evaluability and non- evaluability of terms, but
also about emptiness and non-emptiness of phyla of models: Namely
A [= ( : x x
e
= x) iff there exists an (S-) assignment w : x A such
that A [= (x : x
e
= x)[w]; but since this formula is trivially satised by
every assignment, there only remains as non-trivial consequence that w
exists, and this is equivalent to the fact that A
s
is non-empty.
7.3 Coincidence and Substitution Theorem
7.3.1 For lack of space we do not present here many details about the lan-
guage which we have just introduced. For the nitary homogeneous case it
has been discussed to a great extent including a completeness theorem
in [B82], while for the nitary heterogeneous case a similar discussion, which
even contains proofs, is presented in [Ka82] by H.Kaphengst for a language
almost identical with the above one.
In order to provide a better feeling for the properties of such a language
for heterogeneous partial algebras, we only briey discuss the Coincidence
Theorem and the Theorem of Substitution, which are of great importance,
but dier in our case quite a bit from the usual ones, since we now have
December 8, 2002
236 Chapter II A Basis for a two-valued model theory for partial algebras
to observe evaluability of terms and compatibility with the sort mappings
(cf. also section 6 for the properties of E-equations). A reader who is not so
familiar with logic is referred e.g. to H.Hermes [He63].
Theorem 7.3.2 (Coincidence Theorem) Let A Alg(S) be any partial
algebra, let (M : ) and (N : ) be any L-formulas belonging to the same
basic formula . If v : M A and w : N A are any two (S-) assignments
such that v[
fvar
= w[
fvar
, then
A = (M : )[v] iff A [= (N : )[w].
Proof We use induction on the structure of for all possible M, N, v, w:
For atomic formulas, i.e. for E-equations the statement follows from 5.5.5.(ii).
We omit the cases of negation and c-restricted conjunction, which are
quite simple, and assume that =

Y , #Y < q in X, and further that


A [= (M :

Y )[v], i.e. that for all v

: MY A satisfying v

[
M\Y
= v[
M\Y
we have A [= (M Y : )[v

]. By symmetry we now only have to show that


A [= (N :

Y )[w] for (every) w : N A with w[


fvar
= v[
fvar
:
First case: There is no assignment v

: M Y A (compatible with
the sort mappings!), which means that there is some y Y , say of sort
s S, such that A
s
= namely this cannot occur for m M, since v
maps M in a compatible way into A. But then there also is no (compatible)
w

: N Y A, and trivially A [= (N :

Y )[w].
Second case: There is some compatible v

: M Y A; then there also


is some compatible w

: N Y A, and in both cases it can then be


arranged that the restrictions to M Y (respectively N Y ) coincide with
the corresponding restrictions of v respectively w. Now, given such a w

,
i.e. w

[
N\Y
= w[
N\Y
, we consider v

: M Y A, where v

[
Y \fvar (

Y )
:=
w

[
Y \fvar (

Y )
, and the restriction of v

to the rest of M is the same as the


corresponding restriction of v. Then by the construction of v

and the
assumptions on v and w it is obvious that v

[
M\Y
= v[
M\Y
as well as
v

[
fvar
= w

[
fvar
. The rst equality implies A [= (M Y : )[v

], while the
second one implies together with the induction hypothesis for v

and M Y
on one side and w

and N Y on the other side that A [= (N Y : )[w

].
Since w

was arbitrary (except for the necessary restrictions), we get A [=


(N :

Y )[w].
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 237
Remark 7.3.3 It should be stressed once more that we cannot derive in
the heterogeneous case from the Coincidence Theorem that A [= (M : )
implies A [= (fvar : ), as it can be done in the homogeneous case! But
observing that the proof of 7.3.2 heavily uses the simultaneous existence of
assignments, we have at least the following result which can be strengthened,
as the corresponding conditions for E-equations suggest.
Corollary 7.3.4 If A Alg(S), and if (M : ) and (N : ) are L-formulas
for the same basic formula such that v
X
[M] v
X
[N], then
A [= (M : ) implies A [= (N : ).
Theorem 7.3.5 (Substitution) Let A be any partial algebra, and let and
(N : ) be an L-formula. Moreover, let w : M A be any assignment and
f : N domw

a mapping (both compatible with the sorts!), such that f is


admissible for (N : ). Then
A [= (N : )[w

f] iff A [= (M : f())[w].
In particular, if M = N, then we get the usual form for some m M:
A [= (M : )[w
(w

f)(m)
m
] iff A [= (M : ()
f(m)
m
)[w],
where, for each n M,
w
(w

f)(m)
m
(n) :=
_
w(n), if m ,= n,
(w

f)(m), if m = n;
and we get ()
f(m)
m
from by replacing each free occurrence of m in
by f(m).
Proof The proof is by induction on the structure of :
(a) Let = (t
e
= t

). Then
A [= (N : t
e
= t

)[w

f] iff
(t, t

) ker (w

f)

, iff (t, t

) ker (w

domw

), iff
(f

(t), f

(t

)) ker w

, iff (

f(t),

f(t

)) ker w

, iff
A [= (M :

f(t)
e
=

f(t

))[w], iff A [= (M : f())[w].


Since the cases, when = and =
_

T, are straightforward
using 7.1.5 and the induction hypothesis , we only consider the case
December 8, 2002
238 Chapter II A Basis for a two-valued model theory for partial algebras
(b) Let =

Y . Then let M

:= fvar and f

:= f[
M
; therefore
M

Y = , and the Coincidence Theorem yields A [= (N : )[w

f] iff
A [= (M

: )[w

]. Since it is not too dicult to prove (say also by


induction on the structure of formulas using 7.1.5 and the Coincidence
Theorem) that
() f[
fvar
= f

[
fvar
implies f() = f

(),
it is sucient to prove
() A [= (M

: )[w

] iff A [= (M : f

())[w].
We therefore introduce a mapping f

:= f

Y
: Y M

Y domw

by
setting f

(x) := f(x) = f

(x), if x M

= fvar , and else f

(x) := x
(cf. 7.1.5.(ii).(d)). As in 7.1.6.(ii) we conclude that f

is admissible for
(M

Y : ). Let us rst assume that A [= (M

: )[w

], i.e. A [=
(M

Y : )[u] for all u : M

Y A satisfying u[
M
= w

. Let
v : M Y A be such that v[
M\Y
= w[
M\Y
. Observe that for all
x M

(v

)(x) = (v

)(x) = ((w[
M\Y

f)(x) = (w

)(x);
thus A [= (M

Y : )[v

]. The induction hypothesis then yields


A [= (M Y : f

())[v]. Hence A [= (M : f

Y ))[w].
Conversely, let us assume that A [= (M : f

Y ))[w], and let u :


M

Y A satisfy u[
M
= w

. Then dene v : M Y A as
follows: v(y) := u(y)(= (v

)(y)) for all y Y , and v(x) := w(x)


for all x (M Y ). Hence we get (v

)(x) = v

(x) = u(x) for all


x Y , and (v

)(x) = (w

)(x) = u(x), if x M

(observe that
Y f

[M

] = ), i.e. v

= u. Thus observing () for f

and f

as
well as the Coincidence Theorem A [= (M Y : f

())[v] implies by
induction hypothesis A [= (M

Y : )[v

], i.e. A [= (M

Y : )[u].
The choice of u then implies A [= (M

Y )[w f

], completing the
proof.
Corollary 7.3.6 (Theorem of Substitution for Validity) Let A Alg(S),
M, N X v-small subsets, (N : ) an L-formula and f : N domEeq
M
A
a mapping admissible for (N : ). Then
A [= (N : ) implies A [= (M : f()).
Proof Let w : M A be any assignment. Then w

f is also an assign-
ment, since domEeq
M
A domw

. Since f is admissible for (N : ),


A [= (N : ) implies A [= (N : )[w

f], and therefore by 7.3.5


A [= (M : f()[w]. Hence A [= (M : f()).
December 8, 2002
7 A rst order language for (innitary) partial (S-) algebras 239
Denition 7.3.7 Let K Alg(S) be any class of partial algebras and T
L
c,q
(S, X) =: L any set of L-formulas. We then dene
K [= T iff A [= for every A K and for every T
(for one-element sets the brackets are omitted). With this notation we get:
Th K := L [ K [= and Mod T := A Alg(S) [ A [= T .
Th K is called the theory of K, and Mod T is called the model class of T. K is
said to be axiomatizable (by T) iff K = Mod T for some suitable set T L.
Remarks 7.3.8 (i) Since [= is a binary relation between Alg(S) and L,
Th and Mod are just the closure operators corresponding to the related
Galois connection, whence they satisfy the same properties as observed
for Eeq and Mod (cf. 6.1.3).
(ii) As we have already seen in section 6 it is sometimes useful not to con-
sider the whole language, but only some special part of it, for instance
all E-equations and later on all elementary implications, and then the
operator Th will be replaced by the corresponding operator, while
Mod is always the same operator.
(iii) In the rest of this book we shall mainly be concerned with so-called
open (or universal) formulas. Within the scope of this book it is not
possible to continue the discussion of the general language which we
shall use here mainly as a background for the description of phenomena
and not as a mathematical theory to be studied in its own right. We
would only like to mention some further facts:
If S is nite and if the similarity type is nitary, then it is not dicult
to convince oneself in connection with the results of model theory for
relational systems by replacing partial operations by their graphs
(cf. 7.2.5.(i)) and the phyla by disjoint unary relations, whose domains
cover the whole set, and by requiring by appropriate axioms that the i-
th component of some relation has to be of the appropriate sort that
axiomatizable classes of such partial (S-) algebras are exactly those
classes which are closed with respect to the formation of ultraproducts,
ultraroots and isomorphic copies A Alg(S) is an ultraroot of some
K-algebra, iff some ultrapower of A belongs to K (i.e. we have the
December 8, 2002
240 Chapter II A Basis for a two-valued model theory for partial algebras
Theorem of Keisler and Shelah). We have the strong feeling that this
is no longer true, when the set S of sorts is innite.
Moreover still assuming S to be nite and to be nitary , the above
encoding also shows that the Compactness Theorem, Los Lemma (=
Ultraproduct Theorem), Skolems Theorems etc. also hold for nitary
partial algebras. Thus our language L is just a suitable shorthand so
that we do not have to carry along troublesome axioms. Moreover,
it should be easily realized that it would be almost impossible to get
our results in sections 5 and 6 on terms, if we had to represent them
by existential formulas as indicated in 7.2.5.
(iv) The scope of the above language was mainly motivated by compatibility
with Part II by H.Reichel [Re84] or [Re87] and with the category
theoretical encoding of formulas by Andr eka, N emeti and Sain (as
quoted at the beginning of this chapter), which we shall heavily use in
the case of open formulas in this and the next chapter.
December 8, 2002
8 Elementary implications, universal Horn formulas 241
8 Elementary implications, universal Horn for-
mulas
8.1 Some basic kinds of universal Horn formulas
8.1.1 Usually the equational theory of algebraic structures is still consid-
ered as part of universal algebra, while the theory of elementary implications
is often already counted as belonging to model theory. In connection with
partial algebras the E-equational theory is not of the same importance as
equational theory is for total algebras, but elementary implications, i.e. for-
mulas of the form (M : (

E t
e
= t

)), where E is any set of E-equations,


gain fundamental importance, e.g. in connection with the description of the
domains of partial algebras (see e.g. Reichel [Re84] or [Re87]).
For instance the class Cat

of all small categories considered as homoge-


neous partial algebras (M; , Dom

, Cod

) the domain M being the set of all


morphisms of similarity type (2, 1, 1) can be axiomatized by the following
axioms (universal quantiers and some parenthesis are omitted):
(C1*) (x y
e
= x y y z
e
= y z) x (y z)
e
= (x y) z,
(C2a*) x y
e
= x y Dom

(x y)
e
= Dom

(x),
(C2b*) x y
e
= x y Cod

(x y)
e
= Cod

(y),
(C3*) x y
e
= x y Cod

(x)
e
= Dom

(y),
(C4a*) Dom

(x) x
e
= x,
(C4b*) x Cod

(x)
e
= x.
If we consider (small) categories as two-sorted partial algebras getting
the class Cat of type = (2, 1, 1, 1), where = ( , Dom, Cod, 1),
and S is given by S := ob, mor, graph (
S
) := ((mor, mor), mor),
graph Dom
S
:= ((mor), ob) =: graph Cod
S
and graph 1
S
:= ((ob), mor),
where the argument of 1 is written as usual as an index, then we can formu-
late the axioms as (fvar : ), where is one of the following formulas (m,
m

and m

being of sort mor and o being of sort ob):


(C1) (m m

e
= m m

e
= m

) m (m

)
e
=
(m m

) m

;
December 8, 2002
242 Chapter II A Basis for a two-valued model theory for partial algebras
(C2a) m m

e
= m m

Dom(m m

)
e
= Dom(m),
(C2b) m m

e
= m m

Cod(m m

)
e
= Cod(m

),
(C3) m m

e
= m m

Cod(m)
e
= Dom(m

),
(C4a) 1
o
m
e
= 1
o
m 1
o
m
e
= m,
(C4b) m 1
o
e
= m 1
o
m 1
o
e
= m,
(C5a) Dom(1
o
)
e
= o,
(C5b) Cod(1
o
)
e
= o,
(C6a) 1
Dom(m)
m
e
= m,
(C6b) m 1
Cod(m)
e
= m.
Notice that in both cases the elementary implications which occur as ax-
ioms only contain TE-expressions in the premise (cf. 6.1.2.(ii)), that TE-
expressions are trivially satised in all total algebras, and in this case all
the above axioms behave like E-equations. In each case the denability
condition
(m, m

: Dom(m

)
e
= Cod(m) m m

e
= m m

)
respectively Dom

(y)
e
= Cod

(x) x y
e
= x y is a consequence of the
other axioms. (we omit below the reference to the sets of variables and argue
semantically): Assume for morphisms a and b that Dom(b) = Cod(a). Hence
we can infer from (C6a) and (C6b) that 1
Dom(b)
a = a and b 1
Dom(b)
= b.
And by (C1) (b 1
Dom(b)
) a = b (1
Dom(b)
a) thus again by (C6a) and (C6b),
b a = b a, meaning that a and b are composable. It is also easy to see that
the homomorphic image of a small category need not be a small category:
Let the category C have exactly four objects, say o
1
, o
2
, o
3
and o
4
and besides
the images of Dom
C
and Cod
C
only two further morphisms m : o
1
o
2
and
m

: o
4
o
3
(using the usual notation for morphisms, indicating e.g. that
Dom
C
(m) = o
1
and Cod
C
(m) = o
2
); when we identify o
2
with o
4
and take
the quotient with respect to the induced congruence, say , then we get a
partial algebra, which is not a (small) category, since [m] and [m

] would
be composable in the quotient, if this were a category, and they are not.
December 8, 2002
8 Elementary implications, universal Horn formulas 243
Since we know that E-varieties are closed with respect to homomorphic
images, we may conclude that neither Cat

nor Cat are axiomatizable by


E-equations.
Moreover, almost all the examples in [Re84] and [Re87] by H.Reichel
are not denable by E-equations, but only by elementary implications.
In general it will not even be sucient to use elementary implications of the
form (M :

F t
e
= t

), where F is a set of TE-expressions (we shall call


such axioms ECE-equations).
Hopefully having motivated the reader a little bit by these observations we
now make the above concepts precise by introducing those kinds of open for-
mulas which will be most important for further considerations the reader
should always keep in mind that we usually deal with partial S-algebras
without explicitely mentioning this.
Denition 8.1.2 Let (M : ) be an L-formula (cf. section 7); then (M : )
(possibly withsome cardinality restrictions as in section 7) is called
TE-expression (i.e. term existence expression), iff = (t
e
= t) for some
t F(M, TAlg(S));
E-equation (i.e. existence equation), iff = (t
e
= t

) for some t, t


F(M, TAlg(S));
ECE-equation (i.e. existentially conditioned existence equation), iff =
(
_

F t
e
= t

), where F is some set of TE-expressions with variables


in M and t, t

F(M, TAlg(S));
QE-equation (i.e. quasi-existence equation), iff = (
_

F t
e
= t

),
where F is a set of E-equations and t
e
= t

a single E-equation with


variables in M;
elementary implication, if = (
_

F
_

), where F and F

are
sets of E-equations with variables in M;
NTE-expression (i.e. negation of a TE-expression), iff = (t
e
= t) for
some term t with variables in M;
NE-equation (i.e. negation of an E-equation), iff = (t
e
= t

) for some
terms t, t

with variables in M;
December 8, 2002
244 Chapter II A Basis for a two-valued model theory for partial algebras
ECH-formula (i.e. existentially conditioned universal Horn formula),
iff either =
_

F with F a set of TE-expressions, should one


exchange and
_

here and on similar occasions? or is an


ECE-equation, both with variables in M;
SECH-formula (i.e. special existentially conditioned universal Horn for-
mula), iff either = (t
e
= t) or is an ECE-equation, each with
variables in M;
H-formula (i.e. universal Horn formula), iff either =
_

F, where
F is a set of E-equations, or is a QE-equation, each with variables
in M.
If K Alg(S) is any class of partial algebras, then TEK, Eeq K, ECEeq K,
QEeq K, Impl K, NTEK, NEeq K, ECHK, SECHK and Horn K (cf. ECHEeq K,
SHEeq K, respectively Heq K in [B82]), respectively, are the corresponding
sets of L-formulas valid in K when the operators carry an index M, then
we want to indicate that we only want to consider such axioms with M as
corresponding set of variables ; the model classes of a set of the correspond-
ing axioms are called TE-, E-, ECE-, QE-, Impl-, NTE-, NE-, ECH-, SECH-,
and Horn- varieties, respectively, while TE, Eeq, ECEeq, QEeq, Impl, NTE,
NEeq, ECH, SECH and Horn designate the corresponding sets of formulas
of L, respectively.
Remarks 8.1.3 (i) It should be observed that the hep-varieties of H.Reichel
in Part II are very special QE-varieties.
(ii) For a long time it has been quite confusing among those dealing with
partial algebras that validity of identities could be dened in so dif-
ferent ways, and that all these denitions coincided in the case of
total algebras with the usual denition of validity of equations (cf. the
papers of Slomi nski, Kerkhoff, H oft, Poythress, John, Ed-
wards, Kochen & Specker in the references). Therefore it should
be noticed that all ECE-equations and all conjunctions of them with
the same conclusion reduce to mere equations in the case of total alge-
bras, since then all TE-expressions are trivially satised. In each case
of a notion of validity (except for the one called strong validity by
Slomi nski [Sl68], where satisfaction refers to the congruence generated
by ker w

(w : M A) in the completion F(M, TAlg(S)) of domw

)
December 8, 2002
8 Elementary implications, universal Horn formulas 245
one actually deals with validity with respect to a set of ECE-equations
(or special conjunctions of them) with special prescribed form of the
premise. Let us briey discuss the most frequently used concepts (only
for the homogeneous case, whence we omit the reference set of vari-
ables):
(a) Weak validity of t
e
= t

: If t and t

both are interpreted, then


their interpretations are equal. This is just validity of the
ECE-equation (t
e
= t t

e
= t

) t
e
= t

) (weak equation:
t
w
= t

).
Weak validity can already be found in S.C.Kleenes book [Kl52].
It has been studied among others e.g. by R.Kerkhoff in [Ke70],
by V.S.Poythress in [Po73], by H.H oft in [Hoef70] and [Hoef73],
(somewhat generalized by P.Burmeister in [B71]); and S.Kochen
and E.P.Specker have used it in [KoSp68] for aspects of quan-
tum logic. Recently L.Rudak has formulated and proved in [Ru83]
a completeness theorem for weak equality, thus lling a gap in the
syntactical treatment left by all the previous papers. A seman-
tical description of weak varieties by closedness with respect to
algebraic operators has not yet been given without some refer-
ence to an operator based on syntactical properties since one
is dealing with a very special kind of ECE-equations, this is not
too surprising. In his doctoral thesis [?] L.Rudak has got a more
satisfactory description using some additional restriction for the
weak variety under consideration.
(b) Validity (in Slomi nski [Sl68]: in recent literature of computer
science this is often called strong validity and the correspond-
ing equality strong equality, while W.Craig in [Cg89] calls
it Kleene equality, since it occurs already in [Kl52]) of t
e
= t

: If either t or t

is interpreted, then both terms are equally


interpreted. This is described by validity of the formula
(t
e
= t t
e
= t

) (t

e
= t

t
e
= t

) or of the two-element
set (t
e
= t t
e
= t

), (t

e
= t

t
e
= t

) of ECE-equations (we
shall call it strong equation and denote it by t
s
= t

). In [Cg89]
and [ACgN87] completeness theorems are given for this equational
theory, respectively for some dialects of it. [Cg89] also contains
many many results for a better algorithmic treatment of the syn-
tax of Kleene equality (strong equality).
December 8, 2002
246 Chapter II A Basis for a two-valued model theory for partial algebras
(c) Evans validity of t
e
= t

(cf. T.Evans [Ev51]): If t is inter-


preted, and if all strict subterms of t

are interpreted, then t

is
interpreted with the same value as t, and the same holds for t
and t

interchanged. This is the description of the following


conjunction of ECE-equations:
(t
e
= t
_

q
e
= q [ q (t

) t
e
= t

)
(t

e
= t

bigwedge

p
e
= p [ p (t t) t
e
= t

)
(Evans equation t
E
= t

)
Notice that A [= (t
s
= t

) iff t
A
= t

A
, i.e. strong equations are valid
in A, iff the corresponding induced term operations are equal. And
validity of Evans equations is connected with the embeddability as
relative subalgebras (one has to take all elements of A as fundamental
constants into the type getting, say, A
A
; if B
A
is any partial algebra
similarly extended with the constants for A, then one has to consider
all Evans equations valid in B
A
which contain no free variable; if all
these are also valid in A
A
, then A is fully embeddable into B). Weak
equations are e.g. closely related to quantum logic (cf. Kochen and
Specker [KoSp68]).
(iii) It can be said that R.Kerkhoff [Ke70] and V.Poythress [Po73]
deal in their papers with specic SECH-formulas, namely with NTE-
expressions and weak equations, since they consider weak equa-
tions, where the terms can only be taken from a given subset of the
term algebra, while all other terms are never allowed to be interpreted.
Quite a spectrum of concepts of validity for equations can be found to-
gether with general completeness results and a general kind of algebraic de-
scription of the corresponding varieties in the doctoral thesis of R.John [J75]
(see also [J77]). In the terminology of this book we would say that John
there axiomatizes the relationship between the premise and the conclusion
of the kind of ECE-equations corresponding to an equational concept.
The following denition is of great importance for model theoretic inves-
tigations:
Denition 8.1.4 A set F of L-formulas is preserved by some algebraic oper-
ator, say A (see 4.6 for examples of operators) iff , for every class K Alg(S)
December 8, 2002
8 Elementary implications, universal Horn formulas 247
and for every subset F

F, K Mod F

implies A K Mod F

, i.e. if A is
obtained from K-algebras by using A (say products, homomorphic images,
etc.) and if K [= F

, then A [= F

.
8.1.5 Notice that the First Birkho Theorem (cf. also 6.2.1) for E-equations
especially tells us that E-equations are preserved by homomorphic images,
subalgebras and direct as well as d
&
-reduced products.
8.2 Algebraic encoding of universal Horn formulas
8.2.1 Before we further discuss preservation of formulas by algebraic oper-
ators, we present an algebraic description of those formulas which will be
of interest for us. Among others B.Banaschewski and H.Herrlich used
this method in [BaH76] for total algebras, while, independently, R.John
in [J77] and H.Andr eka and I.N emeti in [AN83], [AN78] and [AN79a] in-
dependently applied it to partial algebras (and in [AN78] and [AN79a] even
to arbitrary categories and for arbitrary formulas, see also Hien & Sain
[HiSa83]). We shall deal in this book only with Horn formulas.
Denition 8.2.2 Let (M : ) be any elementary implication, = (
_

F
_

G), where F, G Eeq


M
(cf. 6.1.1). Then we associate with (M : ) four
partial algebras T
p
(M : ) and T
c
(M : ), which are M-initial relative
subalgebras of F(M, TAlg(S)), and Pr(M : ) and Co(M : ), which are
special quotients of them and an epimorphism between the two latter
ones (observe the sort homomorphisms, which are always added in a natural
way; moreover M
d
designates the discrete partial S-algebra on M cf. 1.1.4
and + designates the (binary) coproduct)
T
p
(M : ) :=M
d
+(
_

F), the partial algebra of the premise terms in ,


Pr(M : ) :=T
p
(M : )/Con
T
p
(M:)
F, the premise algebra for (M : ),
T
c
(M : ) :=M
d
+, the partial algebra of all subterms of terms in ,
Co(M : ) :=T
c
(M : )/Con
T
c
(M:)
(F G), the conclusion algebra for (M : );
i.e. we have the commutative diagram in Figure 8.1, where nat
p
(M : ) and
nat
c
(M : ) are the corresponding quotient homomorphisms.
The induced homomorphism e(M : ) : Pr(M : ) Co(M : ) then ex-
ists according to the Diagram Completion Lemma and the fact that ker (nat
p
(M : ))
ker (nat
c
(M : ) id) by construction. e(M : ) is called the encoding epi-
morphism for (M : ).
December 8, 2002
248 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 8.1: f8.1
gure uncompleted
Figure 8.2: f8.2
Proposition 8.2.3 Let (M : ) = (M :
_

F
_

G) be any elementary
implication with encoding epimorphism e(M : ) : Pr(M : ) Co(M : ),
let A Alg(S), and let w : M A be any assignment.
(i) e(M : ) is in fact an epimorphism.
(ii) A [= (M :
_

F)[w] iff there exists a homomorphism f


w
: Pr(M : )
A satisfying f
w
nat
p
(M : ) id
M
= w.
(iii) A [= (M : )[w], iff : whenever w induces a homomorphism f
w
:
Pr(M : ) A satisfying f
w
nat
p
(M : ) id
M
= w, then there exists
a homomorphism g
w
: Co(M : ) A such that g
w
e(M : ) = f
w
(cf. Figure 8.2).
(iv) If A Alg(S), then A [= (M : ) iff A Inj e(M : ) i.e. iff e(M :
) is A-extendable (cf. 5.1.4). In the homogeneous case we may
replace M by fvar , if fvar is non-empty, or else by any one-element
set, if M is non-empty.
Proof Since M generates both the domain and the codomain of e(M : )
which is induced over id
M
, (i) is obvious. (iv) is a consequence of (iii),
and the last remark follows from the Coincidence Theorem. Let us rst
prove (ii):
Assume that f
w
exists, satisfying f
w
nat
p
(M : ) id
M
= w; then T
p
(M :
) = M(
_

F) domw

and F ker w

, since F ker (nat


p
(M : ))
ker (f
w
nat
p
(M : )) ker w

, i.e. A [= (
_

F)[w].
Conversely, if A [= (
_

F)[w], i.e. if F ker w

, then the Diagram Comple-


tion Theorem 5.8.2 for epimorphisms implies the existence of f
w
.
December 8, 2002
8 Elementary implications, universal Horn formulas 249
Ad (iii): If A [= (M :
_

F)[w] is not true, then f


w
does not exist, but (M : )
is trivially satised by w in A, and conversely. Hence the case of A [= (M :
_

F)[w], i.e. of the existence of f


w
, is indeed the only interesting one. Then
A [= (M : )[w] iff (F G) ker w

, and by (ii) this is equivalent to the ex-


istence of g
w
: Co(M : ) A satisfying g
w
nat
c
(M : ) id
T
p
(M:)
id
M
=
w = f
w
nat
p
(M : ) id
M
. Commutativity of the upper right hand square in
Figure 8.2 yields f
w
nat
p
(M : ) id
M
= g
w
nat
c
(M : ) id
T
p
(M:)
id
M
=
g
w
e(M : ) nat
p
(M : ) id
M
. Since nat
p
(M : ) id
M
is an epimorphism,
this is equivalent to f
w
= g
w
e(M : ).
Corollary 8.2.4 (1) If F is any set of E-equations with variables in M,
and if T
p
:= M
d
+ (
_

F), then e
F
:= nat
Con
T
p
F
id
M
: M T
p
/Con
T
p
F
encodes (M :
_

F).
Corollary 8.2.5 (2) Each epimorphism e : P C represents an elemen-
tary implication (M :
_

F
_

G) (for some suitable cardinality c re-


stricting the lengths of the conjunctions), namely if e is given, then let M
be some generating subset of P, let i : M P be the identity as a map-
ping out of F(M, TAlg(S)), and let F and F G be some generating subsets
of the closed congruence relations ker i

and ker (e i)

, respectively, such
that, for T
p
:= M
d
+ F and T
c
:= M
d
+ (F G): P

= T
p
/Con
T
p
F and
C

= T
c
/Con
T
c
(F G). If a cardinality restriction by c is given, then e
represents a c-restricted implication iff M, F and G can be chosen such that
#M, #F, #G < c.
Since (M :
_

F) can be considered as an implication with an always


false conclusion:
Corollary 8.2.6 (3) If F Eeq
M
, A Alg(S) and w : M A, then A [=
(M :
_

F)[w] iff there exists in the diagram corresponding to Figure 8.2


for :=
_

F no homomorphism f
w
: Pr(M :
_

F) A (see 8.2.4 for


Pr(M :
_

F)) satisfying f
w
nat
p
(M :
_

F) id
M
= w; and A [= (M :

F) iff Pr(M :
_

F) does not allow any homomorphism into A.


Remark 8.2.7 It should be observed that one has for every A Alg(S), for
every w : M A, for every basic L-formula , and for every non-empty set
F of basic L-formulas that
A [= (M :
_

F)[w] iff
A [= (M :
_

[ F )[w], iff
A [= (M : )[w] for every F.
December 8, 2002
250 Chapter II A Basis for a two-valued model theory for partial algebras
Hence, if (M :
_

F) is just an elementary implication, then this can


always be replaced by a set of QE-equations, i.e. the language of all QE-
equations is as powerful as the one of all elementary implications, although
the latter is much more comprehensive. If we had allowed an empty set F
of E-equations in the conclusion, then we would get a formula which would
be always true (cf. 8.2.3), and we would still have the same implicational
model theory as before.
8.3 Preservation properties of algebraic operators
rst visit
8.3.1 In Denition 4.6.2 we have presented a list of operators on Alg(S)
which are of interest for us, in Denition 8.1.2 we have given a list of special
kinds (sets) of formulas, and in both cases we have a partial order related to
set theoretical inclusion. We now want to investigate a little bit more closely
the relationship via preservation between these two lists, and this order will
be useful, as e.g. the following observation shows:
Lemma 8.3.2 Let Q and Q

be two algebraic operators on Alg(S) such that


Q Q

, and let F and F

be two sets of L-formulas such that F F

. Then:
(i) If Q

preserves F

(cf. 8.1.4), then Q preserves F.


(ii) If Q does not preserve F, then Q

does not preserve F

.
Proof Assume that Q

preserves F

, and let K Alg(S) such that K [= F

for any subset F

of F. Since F

is also a subset of F

, we get QK Q

K
Mod F

by our assumptions. (ii) is just the contrapositive of (i).


Remark 8.3.3 From 4.6.4 and 8.1.2 we get the Hasse diagrams for the corre-
sponding partial orders as shown in Figure 8.3. For some of the sets or classes
of formulas the formulas are restricted in size by some cardinality c d
&
as
in the denition of L, and this is indicated by an upper index. For reduced
products the upper index k ( d

) indicates that the lters under considera-


tion are at least k-complete. Notice that lters containing only the index set
are complete for all cardinals. Assume k k

. Observe, too, that E-equations


can be considered as special ECE-equations with an empty premise.
December 8, 2002
8 Elementary implications, universal Horn formulas 251
gure uncompleted
Figure 8.3: f8.3
TE Eeq ECEeq
c
QEeq
c
Impl
c
NTE NEeq ECH
c
SECH
c
Horn
c
H + +
H
f
+ +
H
c
+ + + + + +
o
w
+ +
o
r
+ +
o + + + + + + + + + +
T + + + + +
T
+
+ + + + + + + + + +
T
r,k

T
r,k
+

e + + + + +
1 + + + + + + + + + +
(o)
(T
u
) + + + + + + + + + +
Table 8.1: Preservation (via validity) of formulas by algebraic operators
(for explanations see Theorem 8.3.4)
Theorem 8.3.4 Let the sets of formulas dened in 8.1.2 be taken in L
c,q
(S, X) =:
L, and consider the operators from 4.6.2. Then Table 8.1 presents the preser-
vation properties of these operators with respect to the indicated sets of for-
mulas ( + means preservation, means no preservation, means
preservation only in the case of d
&
, c k, means d

k). In the case


of ultraproducts #S as well as all arities and formulas have to be nite. In
connection with L, o preserves no formulas of the indicated kinds, while in
the case, when the assignments always concern the whole set X of variables,
then o preserves all Horn formulas.
Proof At this place we only consider the case of products, the rest of the
December 8, 2002
252 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 8.4: f8.4
proof of 8.3.4 will be postponed until subsection 11.6, since it will follow more
easily from later results.
Preservation of all elementary implications: Let e : P C be the encod-
ing epimorphism of an elementary implication, and let (A :=

( A
i
[ i
I ), ( pr
i
[ i I )) be a product such that in each A
i
the implication encoded
by e is valid, and let p : P A be any homomorphism. If I = , then A

= S
satises all E-equations and therefore also all elementary implications; but
note that it does not satisfy any negation of a TE-expression, since it is to-
tal, and therefore we get all the negative entries in Table 8.1. Thus assume
I ,= ; then, for each i I, pr
i
p induces (by assumption) a homomorphism
q
i
: C A
i
satisfying q
i
e = pr
i
p. ( q
i
[ i I ) then induces a homomor-
phism q : C A satisfying pr
i
q = q
i
for each i I; therefore p = q e, by
the properties of the projections of a product (see also Figure 8.4) showing
that e is valid in A. If I ,= and if (M : ) is an NTE-formula encoded by
the partial algebra C

, such that in all A


i
(M : ) is valid, then the exis-
tence of a homomorphism from C

into A would contradict this assumption


pr
i
f would show that A
i
would satisfy (M : ) thus (M : ) is also
valid in A (cf. 8.2.6). Because of 8.3.2 the entries in Table 8.1 for products
and for the operator e are proved.
Also without proofs at this place we mention here the following results
whose proofs (see subsection 13.3, e.g. 13.3.4) will follow from the main
theorem of section 12.
Theorem 8.3.5 Let K Alg(S) be any class of partial algebras, and let /
c
be an additional algebraic operator, dened by
/
c
K := A Alg(S) [ there exist B K and a
closed and surjective homomorphism f : A B,
i.e. the class of all closed preimages of K-algebras. Then we have for
k := maxd
&
, c:
(i) Mod TEK = H/
c
T K;
December 8, 2002
8 Elementary implications, universal Horn formulas 253
(ii) Mod Eeq K = HoT K;
(iii) Mod ECEeq
k
K = H
c
oT
r,k
K;
(iv) Mod QEeq
k
K = oT
r,k
K;
(v) Mod ECH
k
K = H
c
oT
r,k
+
K;
(vi) Mod SECH
k
K =
_
H
c
oT
r,k
+
K, if NTEK = ,
H
c
oT
r,k
+
K, else;
only, if each assignment interprets all of X;
(vii) Mod Horn
c
K = oT
r,k
+
K.
8.3.6 We shall see in later sections, e.g. in sections 12 and 13, that especially
within the framework of ECE-equations even ner distinctions can be made
than in the above result.
8.4 Diagonal-ll-in property and projectivity moti-
vation
Observations 8.4.1 Let Q, T be classes of homomorphisms between partial
(S-) algebras, and at the same time, for any class K Alg(S), let
H
Q
(K) := B Alg(S) [ there exists f : A B, f Q, for some A K
be the class of all Q-images of K-algebras, and
o
T
(K) := A Alg(S) [ there exists f : A B, f T , for some B K
be the class of all T -preimages of K-algebras.
Moreover, let (M : ) be any elementary implication encoded by the epimor-
phism e : P C, and let f : A B be any homomorphism.
(i) Let f be a T -homomorphism, and assume B Inj e (cf. 5.1.4). We
want to know, when A Inj e is satised, i.e. when (M : ) respec-
tively e is retractable against T -homomorphism (hereditary with
respect to T -subobjects ), since such a situation occurred in connec-
tion with E-equations and o instead of T , and since we will often meet
similar situations below. Therefore, let p : P A be any homo-
morphism. Then B [= (M : )[f p] (since B Inj e) implies the
December 8, 2002
254 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 8.5: f8.5
gure uncompleted
Figure 8.6: f8.6
existence of a homomorphism q : C B such that q e = f p, i.e. we
get a commutative square as the fully drawn one in Figure 8.5.
What we need is a homomorphism d : C A satisfying d e = p. Since
e is an epimorphism, we may conclude in the case of the existence of
such a d from q e = f p = f d e that q = f d, and that d is
unique. Hence, with our assumptions, A Inj e, iff each such square
as in Figure 8.5 has a unique diagonal-ll-in d making both triangles
commutative.
(ii) Let us now assume f Q and A Inj e (i.e. A [= (M : )), and this
time we ask when we can conclude that also B [= (M : ). Thus assume
r : P B. If we want to apply A Inj e, we need that for each
such r : P B there exists a (not necessarily unique) homomorphism
p := p
r
: P A such that f p = r. Namely under such circumstances
we get a homomorphism d : C A satisfying d e = p; and this implies
f d e = f p = r, i.e. f d is then the homomorphism we have looked
for in connection with r. Again we get a diagram similar to the one in
Figure 8.5 see Figure 8.6. But this time we need a special property
for A, namely that it be e-projective (see below).
These two observations motivate to consider the following two denitions
(see e.g. [N82] or [NSa82]; [NSa82] contains several historical notes on the
subject):
Denition 8.4.2 We say that any two homomorphisms e : P C and f :
A B have the unique diagonal-ll-in-property (in symbols: Dip(e, f))
1
iff
1
For our purposes here it would not matter, whether or not the diagonal-ll-in is
required to be unique right from the beginning, since the implications are encoded by
epimorphisms and hence uniqueness follows in all these applications. Yet since we want
December 8, 2002
8 Elementary implications, universal Horn formulas 255
for every p : P A and q : C B satisfying q e = f p (i.e. making the
fully drawn part in Figure 8.5 commutative) there exists a unique homomor-
phism d : C A such that d e = p and f d = q.
As is well known (cf. section 24), any such relation induces a Galois con-
nection (Galois correspondence), here between HomAlg(S) and HomAlg(S),
with two closure operators, which in this connection are often called and

op
, i.e. for any / HomAlg(S) we have:
(/) := b HomAlg(S) [ Dip(a, b) for all a /,

op
(/) := a HomAlg(S) [ Dip(a, b) for all b /.
Denition 8.4.3 Let P Alg(S), and f : A B. P is called f-projective,
iff for every homomorphism r : P B there exists a homomorphism p :
P A satisfying f p = r (see Figure 8.6.(a)). (The same notion is also
used for this property in arbitrary categories.)
If / HomAlg(S), and if P is f-projective for each f /, then we
say that P is /-projective. By Pj / we denote the class of all /-projective
partial (S-) algebras.
8.4.4 Returning to our observations in 8.4.1 we realize that in connection
with the preservation of elementary implications by Q-homomorphic im-
ages (case (ii)), we will have to study the class Pj Q of all Q-projective
objects.
And in connection with the rst observation we can now say that all
formulas of some given set F of elementary implications are preserved by the
operator T (see 8.4.1), iff for the set, say encEp F, of encoding epimorphisms
of all F-formulas, i.e. encEp F := e(M : ) : Pr(M : ) Co(M : ) [
(M : ) F , we have T (encEp F) this time for the class T .
We shall prove in section 10 that for every class E of epimorphisms
in Alg(S) the pair (of classes) (
op
((E)),
op
(E)) is a so-called factor-
ization system; and our observations show that factorization systems form
the appropriate background for an (algebraic) treatment of the preservation
of implications by any kind of subobjects (see also the next section for an-
other motivation; and see the Remark in [AN79], p. 148). And in section 12
we shall see how this concept relates to the meta-theorem of Andr eka,
N emeti and Sain about the axiomatizability of classes of partial algebras
by elementary implications.
to apply later the theory of factorization systems where uniqueness is required in such
situations, we already require it here.
December 8, 2002
256 Chapter II A Basis for a two-valued model theory for partial algebras
9 Preservation (reection) of formulas by map-
pings
The universal algebraic theory of total algebras usually gets along with ho-
momorphisms, where injectivity, surjectivity or being an epimorphism are
the only additional relevant properties. We have already seen that we need
for partial algebras in addition at least closedness and fullness. But this is
only the top of an iceberg, as we shall see with more examples at the end
of this section. Many users of partial algebras complain as much about the
multitude of properties for homomorphisms between partial algebras (usu-
ally calling them dierent concepts of homomorphisms) as the multitude
of notions of validity of equations (see 8.1.3). As in the latter case the
main reason for the richness of properties of homomorphisms between par-
tial algebras is the newly gained importance of term existence expressions
(respectively of their negations), which play only a trivial role in the case
of total algebras. And concepts which are usually considered part of model
theory have to be used in order to organize this wealth of properties, as we
want to show in this section. We could already realize in section 8 and we
shall get new evidence for it here and in the next section that the category
theoretical concept of a factorization system has a model theoretic aspect in
connection with implications, and it is a very useful tool also in connection
with the concepts described in this section. One should know in advance
that the concepts of preservation (respectively reection) of a formula by a
mapping (homomorphism) are not actually entirely dierent ones, but are
like two sides of the same coin.
9.1 Preservation and reection
Denition 9.1.1 Let A, B Alg(S) be arbitrary partial algebras, (M : )
any L-formula and f : A B any mapping (not necessarily a homomor-
phism, but always compatible with the sort mappings).
(i) We say that f preserves (M : ) with respect to A and B, iff for every
assignment w : M A we have:
A [= (M : )[w] implies B [= (M : )[f w].
(ii) We say that f reects (M : ) with respect to A and B, iff for every
December 8, 2002
9 Preservation (reection) of formulas by mappings 257
assignment w : M A we have:
B [= (M : )[f w] implies A [= (M : )[w].
If we assume which we shall usually do that f is a homomorphism, then
we shall usually omit the addendum with respect to A and B, since then
the structures of A and B are already involved. If F is any set of formulas,
then f preserves (reects) F iff f preserves (reects) each formula in F. In
a similar way we say that a set F of formulas is preserved (reected) by a
class / of mappings, if this is done by each mapping.
Remarks 9.1.2 (i) It is immediate from Denition 2.1.1 that homomor-
phisms between A and B are exactly those mappings from A into B
which preserve all formulas (M

: ( x
k
[ k () )
e
= y) ( ),
where M

:= x
k
, y [ k () , y / x
k
[ k () , and where
k, k

() and k ,= k

imply x
k
,= x
k
(in this case we shall say
that the sequence x := ( x
k
[ k () ) is injective or without repeti-
tions). It is then easy to realize that homomorphisms also preserve
all atomic formulas (E-equations); namely, if (M : t
e
= t

) is given, if
f : A B is any homomorphism, and if w : M A is any assign-
ment such that A [= (M : t
e
= t

)[w], then w

(t) = w

(t

), and this
implies (f w

)(t) = (f w

)(t

) (cf. 5.7.5), i.e. B [= (M : t


e
= t

)[f w].
Moreover, homomorphisms trivially preserve all formulas of the form
(M : (x
e
= x)) (x M X), since we always have A [= (M : x
e
= x);
therefore (M : (x
e
= x)) is never satised, neither in A, nor in B.
While in the case of total algebras we can extend this observation to all
NTE-expressions (M : (t
e
= t)), we shall see below that preservation
of all NTE-expressions by f : A B is equivalent to closedness of f.
As we have mentioned before, the individuality of TE-expressions
and their negations in connection with partial algebras is the clue for
the basic dierence between total and partial algebras and the man-
ifold of partial-algebra-concepts, into which almost any total-algebra-
concept splits.
(ii) For the above observation, and for what follows about closed homo-
morphisms, it is important that we are able to restrict assignments to
the set of free variables, as long as the empty partial algebra or hetero-
geneous partial algebras with empty phyla are allowed. Namely if we
December 8, 2002
258 Chapter II A Basis for a two-valued model theory for partial algebras
had only allowed assignments for all of X into A, and if A had some
empty phylum, then preservation or reection of a formula would mean
nothing in connection with A for any formula, since there would be no
assignments into A at all.
In order to show that there is not much dierence between the concepts
of preservation and reection of formulas by mappings, we state:
Proposition 9.1.3 Let A and B be any similar partial algebras, f : A B
any mapping, and (M : ) any formula. Then with respect to A and B
f preserves (M : ) iff f reects (M : ).
Proof Let w : M A. If f preserves (M : ), and if B [= (M : )[f w],
then A [= (M : )[w] cannot hold, i.e. f must reect (M : ). The converse
argument is similar.
9.1.4 We shall in general use either reection or preservation in such a way
that we may always deal with the more simple kind of formulas. Moreover, we
shall usually assume in what follows that the mappings under consideration
preserve all E-equations, i.e. are homomorphisms.
Examples 9.1.5 (i) An (S-) mapping f : A B is injective iff , for
every a, b A (of the same sort), a ,= b implies f(a) ,= f(b), i.e. iff
f preserves (x, y : (x
e
= y)) (f reects (x, y : x
e
= y)), for any
distinct variables x, y X of the same sort. If we only consider those
formulas (x, y : x
e
= y), where v
X
(x) = v
X
(y) S

S for a proper
subset, then reection of these formulas means injectivity of f[
v
1
A
[S

]
.
(ii) Recall that a homomorphism f : A B is closed iff , for every
and for every a A
()
: f(a) dom
B
implies a dom
A
, i.e. iff
f reects all formulas (x[()] : (x)
e
= (x)) ( , x X
()
preserving sorts); it suces to require that f reects such a formula
for each for at least one such sequence without repetitions.
We shall see below that this will be equivalent to the fact that f reects
all TE-expressions.
Further examples will follow later. But observe that full homomorphisms
cannot be dened by reection of formulas, since the original sequence (as-
signment) might be changed.
December 8, 2002
9 Preservation (reection) of formulas by mappings 259
9.1.6 The following lemma is only formulated with respect to reection,
but because of 9.1.3 one gets the corresponding results for preservation by
interchanging reects and preserves at each occurrence, and if a special
kind of formula is referred to, then it has to be replaced by its negation.
Lemma 9.1.7 Let A, B and C be any similar partial algebras, f : A B
and g : B C any mappings, and = (M : ),
i
= (M
i
:
i
) (i I) any
formulas (#I c, see 7.1). Then:
(i) If both f and g reect , then so does g f.
(ii) If g f reects , and if g preserves , then f reects .
(iii) If g f reects , and if f is surjective and preserves , then g re-
ects .
(iv) If f reects
i
for each i I, then f reects ( M
i
[ i I :
_


i
[
i I ) as well as ( M
i
[ i I :
_


i
[ i I ).
(v) If 0 < #Y < q, Y X, and if f reects (M : ), then f reects
(M Y :

Y ), and if f is surjective, then it also reects (M Y :

Y ).
(vi) If f reects (M : ), and if M N X, then f reects (N : ); and
if fvar M

M, and if there exists an assignment, say v, from M


into A, then f also reects (M

: ).
(vii) Let N X, let q : M F(N, TAlg(S)) be any mapping and let f
reect (M : t
e
= t

) (N : q(m)
e
= q(m)) [ m M Then f reects
(N : q(t
e
= t

)) (cf. 7.1.5).
Proof Since the proofs of (i), (ii) and (iii) are relatively straightforward,
we leave them as exercises observe in (iii) that surjectivity of f allows an
assignment into B to be factorized by one into A via f.
Ad(vi) We only prove the second case (the proof of the rst part should be
obvious): Let v : M A be given; let w : M

A be any assign-
ment, and assume B [= (M

: )[f w]. Then B [= (M : )[f w

]
where w

[
M
:= w, and w

[
M\M
:= v[
M\M
, because of the Coincidence
Theorem. Since f reects (M : ), we have A [= (M : )[w

], thus
A [= (M

: )[w] again by the Coincidence Theorem.


December 8, 2002
260 Chapter II A Basis for a two-valued model theory for partial algebras
(iv) is now an immediate consequence of (vi) and the denition of A [= (M :
_


i
[ i I ) (cf. 7.2.2).
Ad (v) We only prove the case of the existential quantier, the easier case
of the universal quantier is left as an exercise:
Let f be surjective, and let w : MY A be any assignment such that
B [= (MY :

Y )[f w]. Then there exists an assignment u : M B


such that u[
M\Y
= f w and B [= (M : )[u]. Since f is surjective,
there exists an assignment w

: M A such that u[
Y
= (f w

)[
Y
and
w

[
M\Y
= w; hence u = f w

. Thus A [= (M : )[w

], since f reects
(M : ); hence A [= (M Y :

Y )[w].
Ad (vii) Let v : N A be an assignment such that B [= (N : q(t
e
=
t

))[f v], and let t


m
:= q(m) (m M). 6.4.8.(E6) and the fact
that B [= (N : t( t
m
[ m M )
e
= t

(t
m
[ m M))[f v] i.e.
(f v)

(t) = (f v)

(t

) imply B [= (N : t
m
e
= t
m
) [ m
M [f v]. Moreover, the assumption that f reects (N : t
m
e
= t
m
)
yields A [= (N : t
m
e
= t
m
)[v] (m M). Hence we may dene u : M A
by u(m) := v

(t
m
) = (v

q)(m) (m M). Therefore (f v)

(t) =
(M : t)
B
( (N : t
m
)
B
(f v) [ m M ) = (M : t)
B
( (M : m)
B
(f u) [
m M ) = (M : t)
B
(f u); similarly (f u)

(t

) = (M : t

)
B
(f u) =
(f v)

(t

) = (f v)

(t) = (M : t)
B
(f u). Thus B [= (M : t
e
=
t

)[f u]. Since f reects (M : t


e
= t

), we get A [= (M : t
e
= t

)[u], i.e.
t
A
(u) = t

A
(u), and therefore (M : t)
A
(v

q) = (M : t

)
A
(v

q); and
this shows that A [= (M : q(t
e
= t

))[v].
Corollary 9.1.8 (1) Let f : A B be any homomorphism and (M : )
any formula.
(i) If A [= (M : ) and if f is surjective and preserves (M : ), then
B [= (M : ).
(ii) If B [= (M : ) and if f reects (M : ), then A [= (M : ).
Corollary 9.1.9 (2) Let f : A B and g : B C be any homomorphisms
and (M : t
e
= t

) any E-equation.
(i) If f and g both reect (M : t
e
= t

), then so does g f.
(ii) If g f reects (M : t
e
= t

), then so does f.
December 8, 2002
9 Preservation (reection) of formulas by mappings 261
(iii) If f is surjective and g f reects (M : t
e
= t

), then g reects (M : t
e
=
t

).
Proof Observe that homomorphisms preserve all E-equations.
Proposition 9.1.10 Closed homomorphisms reect all TE-expressions.
Proof (By algebraic induction) Let f : A B be closed. Then f trivially
reects all (M : x
e
= x) (x M X). Thus assume t := ( t
k
[ k () )
F(M, TAlg(S)) and that f reects (M : t
k
e
= t
k
) (k ()). By assumption
f reects all (fvar (x) : (x)
e
= (x)), where x := ( x
k
[ k () ) is any
injective sequence of variables (of correct sorts!) since f is closed. Consider
the mapping q : x
k
t
k
(k ()) from fvar (x) into F(M, TAlg(S)), and
let w : M A be an assignment, for which B [= (M : t
e
= t)[f w]. Then
B [= (M : t
k
e
= t
k
) [ k () [f w] Now 9.1.7.(vii) with fvar (x) for M
and M for N implies that f also reects (M : t
e
= t), since (t
e
= t) =
q((x)
e
= (x)).
Of fundamental importance for later applications is the following
Theorem 9.1.11 Let f : A B be any homomorphism and = (M :
_

t
i
e
= t

i
[ i I
_

p
j
e
= p

j
[ j J ) be any elementary implication
with encoding epimorphism e : P C. Then f reects , iff the pair (e, f)
has the diagonal-ll-in property: Dip(e, f) (cf. 8.4.2).
Proof The above statement follows with no great eort from 8.4.1.(i) and 8.2.3.
9.2 Examples for the reection of formulas by homo-
morphisms
Denition 9.2.1 Let f : A B be any homomorphism. f is called initial
(an initial homomorphism) iff , for every mapping g : C A starting from
some partial algebra C, g is a homomorphism from C into A iff f g is a
homomorphism. In the dual situation, i.e. if, for every D Alg(S) and for
every mapping h : B D, h is a homomorphism from B into D iff h f is a
homomorphism, then f is called nal (a nal homomorphism).
December 8, 2002
262 Chapter II A Basis for a two-valued model theory for partial algebras
Remark 9.2.2 It is easy to realize that every full and injective homomor-
phism is initial, and that every full and surjective homomorphism is nal
but the converses are not true.
Proposition 9.2.3 For any homomorphism f : A B the following state-
ments are equivalent:
(i) f is initial.
(ii) For every a


A
[A
()
] [ one has #(f
1
[f(a)]) = 1,
and if a A
()
such that f(a) dom
B
and
B
(f(a)) f[A], then
a dom
A
.
(iii) f reects all formulas of the form (y x[()] : (x)
e
= y) (where
we may assume that x is injective respecting the sorts! , and y /
x[()]).
Proof Let us rst observe that in Denition 9.2.1 of initiality the implica-
tion g is a homomorphism implies that f g is a homomorphism is trivial;
thus we have only to prove the other implication.
(i) implies (ii) (by contraposition) Let a


A
[A
()
] [ , say
a =

A
(b), and let f(c) = f(a) for some c ,= a. Dene C := c
b[()

], graph (

C
) := (b, c), graph
C
:= , else. It is easy to see
that id
C
: C A is no homomorphism, while f id
C
is one. Hence f is
not initial. Similarly, if b A
()
dom
A
, but
B
(f(b)) = f(c), then
dene C as above, and we get the same result.
(ii) implies (iii) Let (ii) be satised, let (N : (x)
e
= y) be given as in (iii)
N := y x[()] , and let w : N A be an assignment such that
B [= (N : (x)
e
= y)[f w], i.e.
B
((f w)(x)) = (f w)(y). Then the as-
sumptions of (ii) are seen to be satised, and (ii) implies w(x) dom
A
and f
1
[(f w)(y)] = w(y); therefore necessarily
A
(w(x)) = w(y),
i.e. A [= (N : (x)
e
= y)[w], i.e. f reects (N : (x)
e
= y).
(iii) implies (i) Assume (iii), and let g : C A be any mapping such that
f g is a homomorphism f g : C B. Then f g preserves all formulas
(N : (x)
e
= y) and f reects them; hence the dual statement (see 9.1.6)
of Lemma 9.1.7.(ii) implies together with Remark 9.1.2.(i) that g is
a homomorphism, i.e. f is initial.
December 8, 2002
9 Preservation (reection) of formulas by mappings 263
gure uncompleted
Figure 9.1: f9.1
Remarks 9.2.4 (i) In analogy to the denition of initial homomorphisms
one may call homomorphisms f : A B relatively closed, when they
satisfy for every and for every a A
()
: f(a) dom
B
and
B
(f(a)) f[A] imply a dom
A
, since f

: A C is closed,
where graph f

:= graph f and C is the relative subalgebra of B with


carrier set f[A]. It is easily realized that relatively closed homomor-
phisms cannot be described by reection of formulas, since composi-
tion of relatively closed homomorphisms need not be relatively closed
(cf. 9.1.7.(i)), as the example in Figure 9.1 shows (for a homogeneous
example of type := (1)), where f and g are relatively closed g is
even closed , but g f is not relatively closed. Notice that all initial
homomorphisms are relatively closed.
(ii) Proposition 9.2.3 shows that a homomorphism is injective and full iff
it is injective and initial. Thus the injective and full homomorphisms
can be dened via reection of special formulas, namely (x, y : x
e
= y)
and (y x[()] : (x)
e
= y). But full homomorphisms alone cannot
be dened by reection of formulas, since this class is not closed with
respect to composition. The class of full and surjective homomorphisms
is closed with respect to composition, but we do not know of any set
of formulas which could dene this class via reection. However, in
the next section we shall see that this latter class is connected with a
class dened by the reection of some formulas through a factorization
system.
Denition 9.2.5 Let P be any set of terms (together with their set of vari-
ables), and let F
P
:= (M
t
: t
e
= t) [ X t M
t
F(M, TAlg(S)), (M
t
:
t) P be any set of TE-expressions. We say that a homomorphism
f : A B is P-reecting iff f reects all formulas from F
P
.
Remarks 9.2.6 (i) If in 9.2.5 P = or P (M : x) [ x M X ,
then P-reecting homomorphisms are just homomorphisms, but if P =
T
v
loc
:=

M F(M, TAlg(S)) [ M X, #M < v v being taken


from 7.1.1 , or at least if (x) [ , x an injective sequence in X
()

December 8, 2002
264 Chapter II A Basis for a two-valued model theory for partial algebras
P, then all P-reecting homomorphisms are closed (therefore we have
sometimes already called them P-closed). Thus the concept of P-
reecting homomorphisms is a common generalization of the concepts
of homomorphisms and of closed homomorphisms.
(ii) If a homomorphism reects (M : ( t
k
[ k () )
e
= ( t
k
[ k () )),
then it need not reect (M : t
k
e
= t
k
) for any k (); for instance
consider := (1), := , A := a, B := b, b

, graph
A
:= ,
graph
B
:= ((b), b

), and graph f := (a, b). Then f trivially reects


(x : ((x))
e
= ((x))), but it does not reect (x : (x)
e
= (x)).
Thus, if we want P to include all subterms of terms occurring in P
(where the sets of variables are not changed!), then we say that we
consider P-reecting homomorphisms. This distinction will gain some
importance in later chapters, since P-reecting homomorphisms have
the nicer theory.
In a similar way as P-reecting homomorphisms generalize closed homo-
morphisms and homomorphisms, initial and relatively normal homomor-
phisms can be generalized as follows:
Denition 9.2.7 Let P be any set of terms,
G
P
:= (M y : t
e
= y) [ (M : t) P, y X fvar t, v

X
(t) = v

X
(y) .
Then a homomorphism f : A B is called P-initial (or P-normal) iff f
reects all formulas from G
P
. If P = T
v
loc
(see 9.2.6.(i)), then we call a
P-normal homomorphism relatively normal.
Remark 9.2.8 It is easy to see from 9.2.3 that initial homomorphisms are
P-initial (= P-normal) ones, when we choose
P := (yx[()] : (x)
e
= y) [ y X x[()], v

X
((x)) = v
X
(y),
x X
()
respecting the sort mappings .
Recalling the concept of normal relative subalgebras from 5.3.2, where
we have met them in connection with free completions, we get the following
motivation for the notation relatively normal:
Proposition 9.2.9 Let A be a relative subalgebra of B such that A gener-
ates B. Then B is a normal extension of A iff id
A
: A B is a relatively
normal homomorphism.
December 8, 2002
9 Preservation (reection) of formulas by mappings 265
Proof First assume that B is an A-generated normal extension of A, and
let = (M y : t
e
= y) be any of the formulas considered in 9.2.7. If
t X, then is reected by id
A
, since id
A
is injective. Thus let us assume
that t = ( t
k
[ k () ), and also that (M z
k
: t
k
e
= z
k
) [ k ()
is reected by id
A
, where we assume the z
k
to be mutually distinct and not
contained in M y. Let w : M y A be any assignment such
that B [= (M y : t
e
= y)[id
A
w]. Then this is only possible, when
(M : t)
B
(w[
M
) A, i.e.
B
( (M : t
k
)
B
[ k () )(w[
M
) A. Thus
normality of A in B implies (M : t
k
)
B
(w[
M
) A (k ()), and
B
( (M :
t
k
)
B
[ k () )(w[
M
) = (M : t)
A
(w[
M
) = w(y), i.e. id
A
reects .
Conversely, let id
A
be relatively normal and c
B
A = B. Let b dom
B
,

B
(b) =: a A. By assumption we can nd sets N A, M X,
#M, #N < v, a bijection w : M N and terms t
k
F(M, TAlg(S)) such
that b(k) = (M : t
k
)
B
(w) (k ()). Hence we get for t := ( t
k
[ k () )
and w

: M y A with w

[
M
:= w, w

(y) := a (where y X M,
v
X
(y) = v

X
(t)) that B [= (M y : t
e
= y)[id
A
w

], and therefore
A [= (M y : t
e
= y)[w

], i.e. (M : t)
A
(w) = w

(y) = a =
A
( (M :
t
k
)
A
(w) [ k () ). Thus (M : t
k
)
A
(w) = (M : t
k
)
B
(w) = b
k
A for all
k () and (M : t)
B
(b) = (M : t)
A
(b), i.e. B is a normal extension of A.
9.2.10 (i) It is still an open problem for us, whether one can describe nor-
mality of an extension B of A in general i.e. when B is not necessarily
generated by A by some model theoretic concept similar to reection
or preservation of formulas. We conjecture that this is not the case.
(ii) It should be observed that a relatively normal homomorphism is always
injective. We shall call a homomorphism almost normal, if it reects
all formulas (M y : t
e
= y), where t F(M, TAlg(S)) M is any
term which is not a variable, y / M, and y M X, #M < v.
Thus it dierence to relatively normal homomorphisms is the fact that
it is not necessarily injective.
Let us now recall and generalize from Model Theory the following deni-
tions:
Denition 9.2.11 A mapping f : A B is an elementary embedding of A
into B iff , for every L-formula (M : ) and for every assignment w : M A,
one has A [= (M : )[w] iff B [= (M : )[f w].
December 8, 2002
266 Chapter II A Basis for a two-valued model theory for partial algebras
f is an L-pure embedding of A into B if the above is true whenever is of
the form =

Y (
_

F), where F is a set of E-equations, #F < c, #Y < q.


Hence we have because of 9.1.3:
Proposition 9.2.12 Let A, B Alg(S), and let f : A B be any mapping.
Then the following statements are equivalent:
(i) f is an elementary embedding of A into B.
(ii) f preserves every L-formula (M : ) with respect to A and B.
(iii) f reects every L-formula (M : ) with respect to A and B.
If we have a nitary language, then it is well known that these are equivalent
to
(iv) For every formula (x : ) (i.e. #fvar 1) f reects ( : x) with
respect to A and B.
Similarly one has
Proposition 9.2.13 Let A, B Alg(S), and let f : A B be any mapping.
Then the following statements are equivalent:
(i) f is an L-pure embedding of A into B.
(ii) f reects every L-formula of the form (M :

Y F), F Eeq
M
, #F <
c, M, Y X, #M < v, #Y < q.
(Every L-pure embedding is injective!)
9.2.14 There exists an approach by Z.Pawlak (cf. [BtRaSk75]) to describe
computers basically as partial algebras with one unary partial operation,
i.e. as (A;
A
). The morphisms used in this connection are called machine
morphisms and are dened as homomorphisms f : A B such that a /
dom
A
with a [A] implies f(a) / dom
B
, i.e. f has to preserve the
formulas (x, y : (x)
e
= y ((y)
e
= (y))) or, equivalently, f has to
reect the formulas (x, y : (x)
e
= y (y)
e
= (y)).
December 8, 2002
9 Preservation (reection) of formulas by mappings 267
Some properties of homomorphisms denable by reection of formulas (see 9.2.15):
No. Property of the homomorphism Type of the reected formula
1. injective (monomorphism) x
e
= y (x, y X, v
X
(x) = v
X
(y))
2. closed (x)
e
= (x) ( , x X
()
injective,
v
X
x = v
()
)
closed t
e
= t (t arbitrary term)
3. injective and closed x
e
= y, (x)
e
= (x) (see 1. and 2.)
4. initial (x)
e
= y (see 2., y / x[()])
5. injective and initial (inj. & full) x
e
= y, (x)
e
= y (see 1. and 4.)
6. relatively normal t
e
= y (t arbitrary term, y / fvar t)
7. P-reecting (M : t
e
= t) (for all (M : t) P; P T
loc
)
8. P-initial = P-normal (M y : t
e
= y) (for all (M : t) P, y / fvar t)
9. Pawlaks machine morphism (x)
e
= y (y)
e
= (y)
10. elementary embedding (M : ) (for all L-formulas)
11. L-pure embedding (M :

Y (
_

F)) (F Eeq

X
, #F < c, #Y < q)
Table 9.1:
December 8, 2002
268 Chapter II A Basis for a two-valued model theory for partial algebras
Table 9.2.15 We collect in Table 9.1 the examples obtained so far about
properties of homomorphisms denable by reection of formulas: If the set
of variables is just the set of free variables of the body of the formula, this
set is omitted. Arbitrary combinations of these properties are possible.
Recall that f : A B is a homomorphism, iff f preserves all E-equations
(atomic formulas), iff f preserves all formulas (x[()] y : (x)
e
= y),
where x X
()
is an injective sequence compatible with the sort mappings,
y X x[()], v
X
(y) = v

X
((x)).
The properties have been discussed at the following places: 1: 9.1.5.(i),
2: 9.1.5.(ii), 3: 9.1.10, 4: 9.2.3, 6: 9.2.7, 7: 9.2.5, 8: 9.2.7, 9: 9.2.14, 10: 9.2.12,
11: 9.2.13.
Observations 9.2.16 (concerning additional concepts)
(i) We have already observed that fullness cannot be described by reec-
tion of formulas, nor do we know sets of formulas describing surjectivity,
nality and related properties by their reection. For some properties
we shall nd tools related to reection in the next section in connection
with factorization systems, but fullness will then only be describable
by some process of relativization. Nevertheless one can nd con-
cepts describable in a model theoretic way to characterize the properties
mentioned above:
We say that a mapping f : A B weakly reects a formula (M :
) with respect to A and B iff , for every assignment w : M A,
B [= (M : )[f w] implies the existence of an assignment w

: M A
satisfying f w

= f w and A [= (M : )[w

].
And we say that f retracts (M : ) with respect to A and B, iff for
every assignment u : M B (!), for which B [= (M : )[u], there exists
an assignment u

: M A such that u = f u

and A [= (M : )[u

].
(ii) It is now easy to realize that a homomorphism f : A B is full, iff
f weakly reects all formulas of the form (y x[()] : (x)
e
= y)
( ), where y X x[()], v
X
(y) = v

X
((x)), and x is an
injective ()-sequence in X respecting the sort mappings.
Thus we may already conclude that properties of homomorphisms de-
ned by weak reection are in general not inherited by the composition
of such homomorphisms.
December 8, 2002
9 Preservation (reection) of formulas by mappings 269
(iii) The simplest example for retraction is the class of surjective homomor-
phisms, since obviously a mapping f : A B is surjective iff it retracts
(x : x
e
= x) for all x X. Similarly one easily sees that f : A B
is a full and surjective homomorphism iff f retracts all formulas of the
form (x : x
e
= x) and (x[()] : (x)
e
= (x)), x X and x an
injective ()-sequence respecting the sort mappings. In addition
nal homomorphisms can be characterized by retraction, since (as we
mention without proof) for f : A B the following statements are
equivalent:
(a) f is nal.
(b) f is full, and for all and b B
()
, b dom
B
implies

B
(b) b[()] f[A].
(c) f retracts all formulas of the form (x[()] : (x)
e
= (x)), where
and x is any injective, sort respecting ()-sequence in X.
Hence a homomorphism is full and surjective, iff it is nal and sur-
jective.
(iv) Finally let us show that the properties of full homomorphisms formu-
lated in 2.4.5.(iii) follow from properties of the concept of weak reec-
tion, which always tacitly refer to the given structures:
Let A, B, C Alg(S), f : A B and g : B C be mappings
(compatible with the sort mappings), and let = (M : ) be some
L-formula. Then we have
(a) If f is surjective and weakly reects , then f retracts .
(b) If f and g both weakly reect , and if f is surjective or g is
injective, then g f weakly reects .
(c) If g f weakly reects , and if g is injective and preserves , then
f weakly reects .
(d) If g f weakly reects , and if f is surjective and preserves ,
then g weakly reects .
Proof (a) follows directly from the denitions in (i) if f is surjective.
Ad (b) Let f and g both weakly reect , and let w : M A be any
assignment such that C [= (M : )[g f w]. If g is injective, then
December 8, 2002
270 Chapter II A Basis for a two-valued model theory for partial algebras
B [= (M : )[f w], since g weakly reects , and there is no assignment
u : M B dierent from f w satisfying g f w = g u. Since f weakly
reects , there now exists v : M A satisfying f v = f w (and
g f v = g f w) and A [= (M : )[v]. If g is not injective, but f
is surjective, then we rst have u : M B satisfying g u = g f w
and B [= (M : )[u], and since f is surjective and weakly reects , we
have v : M A such that f v = u = f w whence g f v = g f w
and v can be chosen such that A [= (M : )[v].
Ad (c) Let w : M A be such that B [= (M : )[f w]. Hence g
preserving C [= (M : )[g f w]. Since g f weakly reects , we
have w

: M A such that g f w

= g f w and A [= (M : )[w

].
Since g is injective, we may conclude f w = f w

, showing that f
weakly reects .
Ad (d) Let u : M B satisfy C [= (M : )[g u]. Since f is surjective,
there is w : M A with f w = u. Since g f weakly reects , we
have w

: M A satisfying g f w

= g f w and A [= (M : )[w

].
Since f preserves , we get B [= (M : )[f w

], showing that g weakly


reects .
December 8, 2002
10 Factorization systems of partial algebras 271
10 Factorization systems of partial algebras
As we have already indicated in section 8.4 and in 9.1.11, the diagonal-ll-
in property (Dip) (cf. 8.4.2) has a fundamental importance in connection
with the reection of elementary implications by homomorphisms and with
algebraic operators connected with elementary implications. Since elemen-
tary implications are encoded by epimorphisms, the diagonal-ll-in is even
unique in all cases considered so far, and this allows factorization systems to
enter the scene, as we shall see below. Actually we shall not meet examples,
where the diagonal-ll-in will not be unique; therefore we shall treat in this
presentation only the unique case, although the more general case in many
instances yields the same results.
We start this section in a more general category theoretical context, which
will show that the concepts treated here can also be applied in greater gen-
erality, say in nice subclasses of Alg(S), especially in so-called full (epire-
ective) subcategories.
In the category Alg(S) with homomorphisms as morphisms a full epire-
ective subcategory is in particular given by any E-, ECE- or QE-variety 1
together with all homomorphisms between 1-algebras.
Actually the concept of factorization systems was to our knowledge rst
considered in more detail in general topology, since, owing to lack of more
than one interesting factorization system, this concept was not then of much
interest in the study of total algebras.
For partial algebras the situation is dierent because of the wealth of
interesting factorization systems, as we shall see below. The general the-
ory of partial algebras much resembles the theory of (Hausdor) point set
topology, while total algebras correspond to compact topological spaces (e.g.
the free completion F(A, TAlg(S)) of A is comparable to the Stone-

Cech-
compactication of a Hausdor space).
10.1 The Galois connection induced by the unique-
diagonal-ll-in property
10.1.1 By replacing in Denition 8.4.2 Alg(S) by an arbitrary category
C = (Ob C, Mor C; , Dom, Cod, 1) and the word homomorphism by C-
morphism and HomAlg(S) by Mor C the concept (unique) diagonal-
ll-in property and the operators and
op
can be generalized to arbitrary
categories (see also Figure 10.1, which indicates for arbitrary classes H, o
December 8, 2002
272 Chapter II A Basis for a two-valued model theory for partial algebras
gures disabled
Figure 10.1: The diagonal-ll-in property and the operators and
op
Mor C, how and in which direction the operators and
op
work).
And in 10.1.2 below the basic properties of the operators and
op
are
formulated and proved (see G.E.Strecker [S72]). In 10.3.12 and and the
related Table 10.1 the reader can nd a survey of classes of homomorphisms
between partial S-algebras, for which we will show below that they are classes
of the form (H) respectively
op
(o) (even as partners in a factorization
system (cf. 10.2.1 and 10.2.5 below)).
Theorem 10.1.2 Let H Mor C be any class of morphisms of some given
category C (say Ob C = Alg(S), Mor C = HomAlg(S). Then:
(i) (H) contains all isomorphisms from C.
(ii) (H) (H) := s s

[ s, s

(H) (H), i.e. (H) is closed with


respect to composition.
(iii) (H)H Iso C, i.e. (H) and H have only isomorphisms in common;
and if Iso C H, then (H) H = Iso C.
(iv) If H

H, then (H) (H

), H
op
(H), and
op
(H) =
(H).
(v) If g f (H), and if g is a monomorphism (i.e. if g Mono C), then
f (H), i.e. (H) is left cancellable with respect to monomor-
phisms.
(vi) If H Epi C is a class of epimorphisms of C, then (H) is left can-
cellable: g f (H) implies f (H).
(vii) (H) is closed with respect to pullbacks, i.e. if (P; p, q) is a pullback
of (f, g; C) (in particular f p = g q, cf. 4.2.2) with f (H), then
q (H).
(viii) (H) is closed with respect to multiple pullbacks, i.e. if (P; ( p
i
[ i I ))
is a multiple pullback of (( f
i
[ i I ); C) (cf. 4.2.2), and if for some
i

I f
i
[ i I i

(H), then p
i
(H) (#I 2, I a set)
in particular: if f
i
[ i I (H), then p
i
[ i I (H).
December 8, 2002
10 Factorization systems of partial algebras 273
(ix) (H) is closed with respect to products, i.e. if ( f
i
: A
i
B
i
[ i I ) is
a family of C-morphisms (where I is any set), if f
i
[ i I (H),
and if f :

( A
i
[ i I )

( B
i
[ i I ) is in the case when these
products exist the induced product morphism, then f (H).
(x) If H Epi C, then (H) is closed with respect to induced product mor-
phisms with respect to non-empty index sets; i.e. if ( f
i
: B A
i
[
i I ) is a source of C-morphisms (I a non-empty set) such that
f
i
[ i I (H), if the product (

( A
i
[ i I ); ( p
i
[ i I ))
exists and f : B

( A
i
[ i I ) is the induced product morphism,
then f (H).
Dually one has for o Mor C:
(i)
op

op
(o) contains all isomorphisms from C.
(ii)
op

op
(o)
op
(o)
op
(o).
(iii)
op

op
(o) o Iso C; and if Iso C o, then
op
(o) o = Iso C.
(iv)
op
If o

o, then
op
(o)
op
(S

), o
op
(o), and
op

op
(o) =

op
(o).
(v)
op
If g f
op
(o), and if f is an epimorphism, then g
op
(o), i.e.

op
(o) is right cancellable with respect to epimorphisms.
(vi)
op
If o Mono C is a class of monomorphisms, then
op
(o) is right
cancellable: g f
op
(o) implies g
op
(o).
(vii)
op

op
(o) is closed with respect to pushouts (cf. and dualize (vii)).
(viii)
op

op
(o) is closed with respect to multiple pushouts. (cf. and dual-
ize (viii)).
(ix)
op

op
(o) is closed with respect to coproducts. (cf. and dualize (ix)).
(x)
op
If o Mono C, then
op
(o) is closed with respect to induced co-
product morphisms in connection to non-empty index sets. (cf. and
dualize (x)).
December 8, 2002
274 Chapter II A Basis for a two-valued model theory for partial algebras
Because of the importance of these concepts and results in later sections
(since we cannot assume the reader to be much familiar with category theoret-
ical argumentation, and since we do not know any place containing detailed
proofs of these statements) we sketch the proofs of these facts in connec-
tion with the operator , the dual statements then just follow by reversing
the arrows. The reader is advised to draw in each case the corresponding
diagram for which we do not have enough space (cf. Figure 8.4).
Proof Ad (i) Since isomorphisms are invertible, the statement is trivial.
Ad (ii) Let s, s

(H) such that s

s exists. Let h H, p, q Mor C


such that q h = s

s p (see Figure 10.2, then we rst have a unique


d

Mor C such that d

h = s p and s

= q, since s

(H); and
next we have a unique d Mor C such that d h = p and s d = d

;
therefore s

s d = s

= q, showing that we have Dip(h, s

s), i.e.
s

s (H).
Ad (iii) f (H) H implies Dip(f, f), and choosing in Figure 8.4 p :=
1 and q := 1 (the identities for the corresponding objects), then the
diagonal-ll-in d satises d f = 1 and f d = 1, i.e. f Iso C. If
Iso C H, then, by (i), H (H) = Iso C.
Ad (iv) These are just properties of the closure operators related to a Galois
connection.
Ad (v)/(vi) Assume h H, g f (H), and let morphisms p, q be given
(see Figure 10.3 such that f p = q h. Since then also g f p = g q h,
we have a unique d satisfying d h = p and g f d = g q. If g is
a monomorphism, then the last equation implies f d = q; if h is an
epimorphism, then q h = f p = f d h implies q = f d, i.e. Dip(h, f)
in both cases.
Ad (vii)/(viii) Let (P; ( p
i
[ i I )) be a multiple pullback of (( f
i
[ i
I ); C), and assume for some i

I that f
i
[ i I i

(H)
(#I 2, I a set); moreover compare Figure 10.4, let h H, and
let p and q be morphisms, such that q h = p
i
p. Then we have for
each j I i

that f
i
q h = f
i
p
i
p = f
j
p
j
p. Thus (f
j
(H))
there is a unique morphism d
j
satisfying d
j
h = p
j
p and f
j
d
j
= f
i
q.
Since (P; ( p
i
[ i I )) is a multiple pullback, there is a unique d such
that p
j
d = d
j
and p
i
d = q (j I i

). Since also d
j
h = p
j
p =
p
j
d h (j I i

) and q h = p
i
p = p
i
d h, the morphisms p
December 8, 2002
10 Factorization systems of partial algebras 275
gures disabled
Figure 10.2: ad (ii)
gures disabled
Figure 10.3: ad (v)/(vi)
and d h can be considered as the unique morphism induced by the
family ( d
j
h [ j I ), where d
i
:= q. Thus d h = p, and we have
Dip(h, p
i
).
Ad (ix) The argumentation is almost identical with the one for (viii); there-
fore we leave it as an exercise (however, cf. Figure 10.5).
Ad (x) Assume H Epi C, and let ( f
i
: B A
i
[ i I ) be a source of
(H)-morphisms, where I is any set, let the product (A :=

( A
i
[
i I ); ( p
i
[ i I )) exist, and let f : B A be the induced product
morphism. Moreover, let h H, an let p and q be morphisms (cf.
Figure 10.6, such that q h = f p. Then p
i
q h = p
i
f p = f
i
p for
each i I. Since each f
i
belongs to (H), we get for each i I a unique
morphism d
i
such that d
i
h = p and f
i
d
i
= p
i
q = p
i
f d
i
. Since h
is an epimorphism, the rst family of equations implies d
i
= d
j
=: d for
all i, j I, and since (A; ( p
i
[ i I )) is a product, the last equation of
the second family implies q = f d; thus: Dip(h, f).
Before we discuss examples in Alg(S), we rst investigate the relationship
of Dip to factorization systems.
10.2 Factorization systems
A survey about examples in Alg(S) discussed in the following subsections
can be found in Table 10.1 (see 10.3.12). However, before we consider them
gures disabled
Figure 10.4: ad (vii)/(viii)
December 8, 2002
276 Chapter II A Basis for a two-valued model theory for partial algebras
gures disabled
Figure 10.5: ad (ix)
gures disabled
Figure 10.6: ad (x)
in detail, we rst investigate the relationship of the relation Dip to fac-
torization systems (see e.g. 10.2.2 and 10.2.5). Much of the material in this
section is taken from [NSa82].
Denition 10.2.1 Let C be any category, and let H, o Mor C be any
classes of morphisms. Then one says that (H, o) is a factorization system
or that C is a bicategory with respect to H and o , iff the following axioms
hold:
(FS1) o H = Mor C, i.e. each f Mor C has an (H, o)-factorization
(h, s) Ho such that s h = f.
(FS2) Iso C H o.
(FS3) H H H and o o o.
(FS4) The (H, o)-factorization of any C-morphism is unique up to unique
isomorphism, i.e. whenever there are h, h

H and s, s

o such that
s h = s

, then there exists a unique isomorphism d Iso C such


that d h = h

and s

d = s.
If (H, o) is a factorization system, then we call H its left factor and o its
right factor.
Proposition 10.2.2 Let (H, o) be a factorization system in the category C.
Then:
(i) H o = Iso C.
(ii) H H = H and o o = o.
December 8, 2002
10 Factorization systems of partial algebras 277
gure uncompleted
Figure 10.7: f10.7
(iii) If p, p

: A B are C-morphisms such that p h = p

h and s p = s p

for some h H and s o, then p = p

.
(iv) Ho Dip.
(v) If h

h, h H, then h

H; if s

s, s

o, then s o.
Proof Ad (i) Let f : A B, f H o. Since Iso C H o and
1
A
, 1
B
Iso C, (1
A
, f) and (f, 1
B
) are (H, o)-factorizations of f; thus
there exists by (FS4) a unique isomorphism i : B A such that i f =
1
A
and f i = 1
B
, i.e. f is an isomorphism. Combined with (FS2) we
get Iso C = H o.
Ad (ii) Because of (FS1) and (FS2), (FS3) implies H Iso C H H H
H, and the same for o; therefore (ii) follows.
Ad (iii) Let the assumptions of (iii) be satised, and let (h

, s

) and (h

, s

) be
(H, o)-factorizations of p and p

, respectively. By (FS3) h

h, h

h H
and s s

, s s

o, and the assumptions imply s

(h

h) = s

(h

h),
(s s

) h

= (s s

) h

and (s s

) (h

h) = (s s

) (h

h); hence (FS4)


implies the existence of unique isomorphisms d, d

and d

(see Fig-
ure 10.7) such that
(a) d (h

h) = h

h and s

d = s

;
(b) d

= h

and (s s

) d

= s s

;
(c) d

(h

h) = h

h and (s s

) d

= s s

.
From (a) and (b) we get
(d) d (h

h) = h

h and (s s

) d = s s

(e) d

(h

h) = h

h and (s s

) d

= s s

.
The uniqueness of the isomorphisms in each case yields that (d) and (c)
imply d

= d, while (e) and (c) imply d

= d

. Hence (a) and (b) imply


s

d = s

and d h

= h

, thus p = s

= (s

d) (d
1
h

) = s

=
p

.
December 8, 2002
278 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 10.8: f10.8
gure uncompleted
Figure 10.9: f10.9
Ad (iv) Assume h H, s o, let p, q Mor C such that q h = s p, and
let (h

, s

) and (h

, s

) be (H, o)-factorizations of p respectively q (see


Figure 10.8). Since now s

(h

h) = (s s

) h

, (FS3) and (FS4) imply


the existence of a unique isomorphism i satisfying i (h

h) = h

and
(s s

) i = s

. Dene d := s

i h

; then d h = s

h i h

h = s

= p
and s d = s s

i h

= s

= q, i.e. d is a diagonal-ll-in for h and s.


If we have another diagonal-ll-in d

satisfying d

h = p and s d

= q,
then (iii) implies d = d

, showing the uniqueness, i.e. Dip(h, s).


Ad (v) (v) follows from (iii) (cf. 10.1.2.(v), (v)
op
) and is left as an exercise
(cf. Figure 10.9).
Remark 10.2.3 10.2.2.(iii) shows that even without requiring H Epi C
or o Mono C which would immediately imply the uniqueness of the
diagonal-ll-in in 10.2.2.(iv) this uniqueness can be proved, when (H, o) is
a factorization system. Moreover, the pairs (h, s) H o share properties
of epimorphisms and monomorphisms. In this book we shall have in most
cases H Epi Alg(S), but the factorization system (nal homomorphisms,
bijective homomorphisms) (see 10.3.11.(iv)) shows that this is not necessary,
while the factorization system (TAlg(S)-extendable epimorphisms,closed ho-
momorphisms) (see 10.3.4.(i)) shows that the right factor need not consist of
monomorphisms. In [Pa79], Beispiel 1.12 and Beispiel 1.13, A.Pasztor has
given examples of factorization systems in categories C, where neither the
left factor is contained in Epi C, nor is the right factor contained in Mono C.
From I.N emeti, I.Sain [NSa82] we take the following
Proposition 10.2.4 Let C be any category, H, o Mor C. Then the follow-
ing statements are equivalent:
December 8, 2002
10 Factorization systems of partial algebras 279
gure uncompleted
Figure 10.10: f10.10
gure uncompleted
Figure 10.11: f10.11
(i) (H, o) is a factorization system in C.
(ii) H and o satisfy (FS1), (FS5) and (FS6), where
(FS5) Iso C H H and o Iso C o.
(FS6) H o Dip (i.e. each pair (h, s) H o has the unique
diagonal-ll-in property: Dip(h, s)).
Proof Let (H, o) be a factorization system, then (FS2) and (FS3) obviously
imply (FS5), and (FS6) has been proved in 10.2.2.(iv).
Conversely, let us assume that (FS1), (FS5) and (FS6) are satised. First
assume h

, h

H, and let (h, s) be an (H, o)-factorization of h

; let
p := h and q := 1
Cod(h

)
(see Figure 10.10. Applying (FS6) twice yields
diagonal-ll-ins d

and d

=: d such that rst d

= p(= h) and s d

=
q h

= h

, while in the second case d h

= d

and s d = q = 1
Cod(h

)
. From
the last equation it easily follows that s is an epimorphism. Now (d s) h =
d (h

) = h = 1
Dom(s)
h, and s (d s) = (s d) s = s = s 1
Dom(s)
. Since
d s and 1
Dom(s)
can be considered as diagonal-ll-ins of the same diagram
(see Figure 10.11), the uniqueness in (FS6) implies d s = 1
Dom(s)
. Altogether
we have shown that s is an isomorphism inverse to d; therefore (FS5) implies
h

Iso C H H. The argument for o o o is similar.


In order to prove (FS2) let (h, s) be an (H, o)-factorization of i Iso C,
i.e. s h = i; thus (i
1
s) h = 1
Dom(h)
and s (h i
1
) = 1
Cod(s)
(see Fig-
ure 10.12). Then one easily realizes (h i
1
s) h = h = 1
Cod(h)
h, and
gure uncompleted
Figure 10.12: f10.12
December 8, 2002
280 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 10.13: f10.13
s (h i
1
s) = s 1
Cod(h)
; as above, (FS6) implies h i
1
s = 1
Cod(h)
=
1
Dom(s)
. This shows that h and s are isomorphisms; therefore i = s h
Iso C H o Iso C H o (by (FS5)), hence (FS2).
In order to prove (FS4) let f Mor C, and let (h, s) and (h

, s

) be two
(H, o)-factorizations of f (see Figure 10.13): f = s h = s

. Dip(h, s

)
implies for (p, q) := (h

, s) the existence of a unique d satisfying d h = h

and s

d = s, while a similar choice for Dip(h

, s) yields a morphisms d

such that d

= h and s d

= s

; thus h = d

(d h) = 1
Cod(h)
h and
(s d

) d = s = s 1
Cod(h)
, and (FS6) implies d

d = 1
Cod(h)
, and in a similar
way one can show that d d

= 1
Cod(h

)
. Hence d and d

are isomorphisms
uniquely determined because of (FS6), and (FS4) is satised.
That the ties between the unique diagonal-ll-in-property Dip and fac-
torization systems are even closer is shown by the following result (cf. [NSa82]):
Proposition 10.2.5 If (H, o) is a factorization system in a category C, then
(H) = o and
op
(o) = H (uniqueness of the diagonal-ll-in need not even
be required here!).
Proof Assume f (H), let f = s h, (h, s) H o. Because of
Dip(h, f) there exists d such that d h = 1
Dom(f)
and f d = s. Hence
(h d) h = h (d h) = h 1
Dom(f)
= h = 1
Cod(h)
h, and s (h d) = (s h) d =
f d = s = s 1
Cod(h)
. 10.2.2.(iii) then implies h d = 1
Cod(h)
, showing that
h is an isomorphism, i.e. f = s h o Iso C o because of (FS5). Hence
(H) o. (FS6) implies o (H); therefore equality holds.
op
(o) = H
is proved in a similar (dual) way.
From H.Herrlich and G.E.Strecker [HS73], exercise 34K, we take
the following
Denition 10.2.6 Let C be any category.
(i) Every monomorphismm Mono C(Epi C) is called a strong monomor-
phism of C, every epimorphism e Epi C
op
(Mono C) is called a strong
epimorphism of C.
December 8, 2002
10 Factorization systems of partial algebras 281
(ii) C is called well-powered, iff for every object A Ob C there exists a
set, say Sub(A) (B, m) [ m : B A, m Mono C such that for
every C-monomorphism n : C A there exist (B, m) Sub(A) and
an isomorphism i : C B such that m i = n. C is called co-(well-
powered), if C has the analogous (dual) property for epimorphisms.
Example 10.2.7 For every similarity type and for every (heterogeneous
sort algebra) S Alg() the category (Alg(S) HomAlg(S); , Dom, Cod, 1)
is well-powered and co-(well-powered). Namely for every A Alg(S)
(B, id
BA
: B A) [ B is a weak subalgebra of A
is such a representative set for monomorphisms into A; and
(nat

[
A
: A C/, C/) [ is a closed, A-initial congruence
in F(A, TAlg(S)) with domain C
is such a representative system for epimorphisms (cf. 5.8).
As far as the strong epimorphisms and the strong monomorphisms in Alg(S)
are concerned, we have the following
Proposition 10.2.8 In Alg(S) we have:
(i) (Epi Alg(S)) = m [ m is a closed and injective homomorphism,
i.e. here the strong monomorphisms are exactly the closed and injective
homomorphisms.
(ii)
op
(Mono Alg(S)) = e [ e is a full and surjective homomorphism,
i.e. the strong epimorphisms are exactly the full and surjective homo-
morphims.
(iii) (Epi Alg(S), (Epi Alg(S))) is a factorization system in Alg(S).
(iv) (
op
(Mono Alg(S)), Mono Alg(S)) is a factorization system in Alg(S).
(v) In TAlg(S) the full subcategory of Alg(S) of all total S-algebras the
two factorization systems corresponding to (iii) and (iv) coincide.
Proof Because of 10.2.5 it is sucient for (i) to (iv) to prove
December 8, 2002
282 Chapter II A Basis for a two-valued model theory for partial algebras
(a) (Epi Alg(S), all closed and injective homomorphisms) is a factorization
system of Alg(S);
(b) (all full and surjective homomorphisms, Mono Alg(S)) is a factorization
system of Alg(S).
Because of Theorem 10.1.2 and the general properties of epimorphisms and
monomorphisms in both cases only (FS1) and (FS4) have to be proved, but in
connection with (a) one easily realizes that every homomorphism f : A B
can be factored uniquely up to isomorphism into f = i f

, where f

: A
c
B
f[A], graph f

= graph f, and i is the closed inclusion homomorphism from


the f[A]-generated subalgebra of B into B. Concerning (b), (FS1) and (FS4)
follow from the First Homomorphism Theorem 2.7.4.
Since in TAlg(S) all TAlg(S)-homomorphisms are closed, all TAlg(S)-epimorphisms
are full and surjective and all TAlg(S)-monomorphisms are closed and injec-
tive, we have (v).
10.2.9 Proposition 10.2.8 may give an idea, how useful the following gen-
eral result may be for the case of partial algebras, while in the case of total
algebras at least as long as the left factor is required to consist of some
class of epimorphisms and the right factor has to consist of some class of
monomorphisms only one relevant factorization system exists. Notice
that, trivially, (Mor C, Iso C) and (Iso C, Mor C) are always factorization sys-
tems.
Theorem 10.2.10 Let C be any co-(well-powered) category, which has mul-
tiple pushouts (i.e. for every set I and for every source (A; ( f
i
[ i I )) a
multiple pushout exists). And let (Epi C, all strong C-monomorphisms) be a
factorization system in C, then for any class c Epi C, (
op
(c), (c)) is a
factorization system of C, and
op
(c) Epi C.
A dual statement holds also for monomorphisms.
Proof Assume c Epi C, let f : A B be any C-morphism, and let
( A
e
i
D
i
m
i
B [ i I ) be a representative system for all factorizations
of f: e
i

op
(c), m
i
e
i
= f for each i I, and if A
e
D
m
B is
a factorization of f (m e = f, e
op
(c)), then there exist i I and
an isomorphism d : D
i
D such that d e
i
= e and m d = m
i
. Since
c Epi C, we have (c) (Epi C), i.e.
op
(c)
op
(Epi C) = Epi C by
December 8, 2002
10 Factorization systems of partial algebras 283
gure uncompleted
Figure 10.14: f10.14 (old f 10.1)
assumption; therefore
op
(c) consists only of C-epimorphisms. Thus, and
since C is assumed to be co-(well-powered), we may assume that I is a set.
Let (( j
i
[ i I ); C) be a multiple pushout of (A; ( e
i
[ i I )). Notice that
I ,= , since f 1
A
is a factorization of f with 1
A

op
(c) (cf. (FS2)). Let
m : C B be the C-morphism induced with respect to the family ( m
i
:
D
i
B [ i I ) by the properties of a multiple pushout. Hence m j
i
= m
i
(i I). Since I ,= , and since it can always be achieved that #I 2 (e.g.
by counting a factorization twice), we may conclude from 10.1.2.(viii)
op
that
j
i

op
(c) for each i I, i.e. e := j
i
e
i

op
(c). Hence there exist i

I
and an isomorphism d : D
i
C such that d e
i
= e and m d = m
i
.
Without loss of generality we may assume that C = D
i
and d = 1
C
.
We now want to show that m (c) =
op
(c): Let v
op
(c) and
p, q Mor C such that q v = m p (see Figure 10.14), let (r, s; P) be a
pushout of (X := Dom(v); p, v), and let u : P B be the morphism induced
by (m, q; B); thus u r = m, and u s = q. Because of 10.1.2.(vii)
op
: r

op
(c), hence r e
op
(c), and therefore we may assume as above that
there is i

I such that P = D
i
, r e = e
i
and u = m
i
. In particular we
have j
i
: P C such that j
i
r e = e; therefore j
i
r = 1
C
, since e is an
epimorphism.
Dene d := j
i
s : Y C, where Y := Cod(v). Then we have d v =
(j
i
s) v = j
i
r p = p, and m d v = u r p = m p = q v. Since
v is an epimorphism, this implies m d = q and also the uniqueness of d,
i.e. d is a unique diagonal-ll-in, and we have Dip(v, m); therefore m
(c). This shows that (
op
(c), (c)) satises (FS1). (FS5) follows from
Theorem 10.1.2, since obviously Iso C
op
(c) (c) and by denition

op
(c) (c) Dip, i.e. we have (FS6). Hence (
op
(c), (c)) is indeed
a factorization system.
Remark 10.2.11 Since Alg(S) is co-(well-powered) and has multiple pushouts,
and since (Epi Alg(S), all strong monomorphisms) is a factorization system
in Alg(S), all the classes of homomorphisms in particular those in Table 9.1
of subsection 9.2 dened by reection of some set of implications form the
right factor of some factorization system (cf. 8.4.1.(i) and 8.4.4).
December 8, 2002
284 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 10.15: f10.15
10.2.12 In order that we do not become too much involved here with more
special category theoretical concepts we investigate the following specializa-
tion of Theorem 10.2.10 only for the category of partial S-algebras any sub-
class closed with respect to subalgebras and direct products would do, too.
Let us recall now from Denition 5.1.1 the concept of K-extendable homo-
morphisms, and let us dene for K Alg(S):
Ext K := e Epi Alg(S) [ e is K-extendable .
Remarks 10.2.13 (i) 5.1.6 implies: Ext K = Ext oT K for every K
Alg(S).
(ii) Let K Alg(S), K = oT K, let A Inj Ext K, and let (r
A,K
, F(A, K))
be its K-universal solution. Since then r
A,K
Ext K, A Inj Ext K
implies that there exists d : F(A, K) A such that d r
A,K
= id
A
.
Since r
A,K
is an epimorphism, we additionally have from r
A,K
d r
A,K
=
r
A,K
id
A
= id
F(A,K)
r
A,K
that r
A,K
d = id
F(A,K)
, showing that r
A,K
is an
isomorphism; thus A K. Thus every quasiprimitive class is uniquely
determined as Ext K for some K Alg(S). In particular:
Proposition 10.2.14 For every subclass K Alg(S) : Inj Ext K = oT K.
Theorem 10.2.15 For every subclass K Alg(S) one has:
(i) If A oT K and f : A B, then f (Ext K).
(ii) (Ext K, (Ext K)) is a factorization system.
Proof Ad (i) Let e Ext K, e : P C, A oT K, f : A B, and let
p, q be given such that f p = q e (see Figure 10.15. Since A oT K
and e Ext K, there exists d : C A such that d e = p. Since e is
an epimorphism, d is unique and f d e = f p = q e implies f d = q,
showing that Dip(e, f) and f (Ext K).
December 8, 2002
10 Factorization systems of partial algebras 285
gure uncompleted
Figure 10.16: f10.16
Ad (ii) Since Alg(S) satises the assumptions of Theorem 10.2.10, we only
have to show that
op
(Ext K) Ext K. Therefore let e
op
(Ext K),
e : P C, let A K and let f : P A be any homomorphism. Con-
sider the K-universal solutions (r
P,K
, F(P, K)) =: (r
P
, F
P
) and (r
C,K
, F(C, K)) =:
(r
C
, F
C
). Then f induces a homomorphism

f : F
P
A such that
f =

f r
P
(see Figure 10.16, and r
C
e induces a homomorphism q :
F
P
F
C
satisfying q r
P
= r
C
e. Because of (i), q (Ext K); thus
there exists a unique diagonal-ll-in d : C F
P
satisfying d e = r
P
;
hence

f d e =

f r
P
= f, showing that e Ext K.
Following H.Herrlich [H72], we dene
Denition 10.2.16 Let K Alg(S). Then the elements of (Ext K) are
called K-perfect homomorphisms.
Remark 10.2.17 We would like to recall from category theory the following
concepts:
Let / Ob C for some category C. Then / is called a reective subclass
of C, iff
(a) for every B Ob C there exist an object F(B, /) / and a morphism
r
B,A
: B F(B, /) such that for every A /and for every f : B A
there exists a unique

f : F(B, /) A satisfying f =

f r
B,A
; and
(b) for every f : B C in Mor C there exists a morphism F(f, /) :
F(B, /) F(C, /) such that r
C,A
f = F(f, A) r
B,A
, F(1
C
, /) =
1
F(C,A)
and F(g f, /) = F(g, /) F(f, /).
This means that F(, /) is a functor from C into the full subcategory of C
with object class / and all C-morphisms between /-objects. In category
theoretical terminology F(, /) is called a left adjoint functor for the inclusion
functor from / into C.
A reective subcategory / is called epireective, if, for each B /,
r
B,A
is an epimorphism of C; it is called monoreective, if each r
B,A
is a
December 8, 2002
286 Chapter II A Basis for a two-valued model theory for partial algebras
monomorphism; and H-reective where (H, (H)) is a factorization system
in C , if each r
B,A
belongs to H. In particular, for K Alg(S) the class
Inj Ext K, i.e. each quasiprimitive subclass, yields always an epireective full
subcategory of Alg(S).
It is proved in category theory that each monoreective full subcate-
gory of C is epireective (cf. [HS73], Proposition 36.3). The epimorphisms
of C which are also monomorphisms are called bimorphisms. In TAlg(S)
all bimorphisms are isomorphisms, since closed and injective epimorphisms
are isomorphisms. Thus in TAlg(S) the only abstract monoreective sub-
class (abstract meaning to be closed with respect to isomorphic copies)
in Ob TAlg(S) is TAlg(S) itself, while in Alg(S) there are usually abundantly
many monoreective subcategories, namely all quasiprimitive subclasses /
containing TAlg(S), since then, for every B Alg(S), F(B, TAlg(S)) /,
and r
B,A
has to be a left factor of id
B
: B F(B, TAlg(S)); therefore r
B,A
is
injective. And such classes can be obtained e.g. as Inj c
P
, where P T
loc
v is
any set of terms and c is the set of all encoding epimorphisms for (M : t
e
= t)
for (M : t) P (cf. 8.2.4 and 10.3.5).
10.3 Examples of factorization systems for partial al-
gebras
10.3.1 In what follows we want to describe some further in our opinion
important factorization systems in the category Alg(S) of partial algebras.
In order to get a formal description of factorization systems connected with
P-reecting homomorphisms for any set P of terms we have to introduce
some generalization of the concept of a term and a term operation. As far
as we know, this concept has been rst introduced by A.Obtulowicz in
connection with Lawvere style investigations of the category Alg() of all
partial algebras of some given type . We shall consider this concept later in
section 15 in more detail; here we only need the denition and some special
items.
Denition 10.3.2 (i) Let M X, #M < d
&
, and let tD F(M, TAlg(S)).
Then (M : t [ D) is called a conditioned term (briey: c-term; more
precisely we should say conditioned S-term) with set M of variables.
(ii) If A Alg(S) and (M : t [ D) is a conditioned term, then the induced
December 8, 2002
10 Factorization systems of partial algebras 287
c-term operation (M : t [ D)
A
is dened by:
dom(M : t [ D)
A
:=

dom(M : p)
A
[ p t D
= a A
M
[ (M : p)
A
(a) exists for every p t D,
and if a dom(M : t [ D)
A
, then (M : t [ D)
A
(a) := (M : t)
A
(a).
(iii) Let P T
loc
(see 6.1.1) be any set of terms, then we recall that P :=
(M : t) [ t p for some (M : p) P . If A Alg(S), then let us
dene
A, P) := (A, ( (M : t [ p)
A
[ t p and (M : p) P )).
If necessary, A, P) is connected with S by v
A
, since v
A
is compati-
ble with S-term operations and therefore also with conditioned S-term
operations.
Examples 10.3.3 (i) Let (I

Z

O; , ) be a partial Mealy automa-
ton, i.e. for S := i, z, o we have graph
S
= ((i, z), z), graph
S
=
((i, z), o), where I is the set of input elements, Z the set of states,
and O the set of output elements, and i, z, o are the corresponding
sorts in S. Then, sometimes, in connection with the process of re-
duction only terms having their value in O may be of interest, like
(i

, i

, z

: (i

, (i

, z

))) (i

, i

are of sort i, z

is of sort z). In this


connection a term like (i

, i

, z

: (i

, z

)) might only be of inter-


est, when it occurs as a subterm with value in O, i.e. e.g. in the form
( i

, i

, z

: (i

, z

) [ (i

, (i

, z

)) ).
(ii) Let P be the smallest homogeneous partial Peano algebra on a, b, c
of type := (2, 2), in which

yyxx
e
=

yyxx is valid, and


in which
P
(a, b),
P
(a, c) and
P
(b, c) exist, and let Q be the rel-
ative subalgebra of P generated by a, b, c by the conditioned term
operations (x, y : xy [

xyxy)
P
and (x, y :

xyxy)
P
.
Then Q is the smallest partial Peano algebra on a, b, c, in which

xxxx
e
=

xxxx is valid; in particular


P
(a, b),
P
(a, c) and

P
(a, b) do not belong to Q, since none of these elements occurs in a
sequence in dom

P
.
As a further preparation for the description of the factorization system
connected with P-reecting homomorphisms let us observe that we have as
December 8, 2002
288 Chapter II A Basis for a two-valued model theory for partial algebras
an immediate consequence of Corollary 5.8.4 of the Epimorphism Theorem
(see also [BJPa78]):
Proposition 10.3.4 Let /
c
be the class of all closed homomorphisms in Alg(S).
Then
(i) (Ext TAlg(S), /
c
) is a factorization system, i.e. the closed homomor-
phisms are exactly the TAlg(S)-perfect homomorphisms.
(ii) An epimorphism e : A B is TAlg(S)-extendable (i.e. belongs to
Ext TAlg(S)), iff e = i e

, where i is an isomorphism, and e

is a full
and injective epimorphism from A into an A-generated relative subal-
gebra of F(A, TAlg(S)) such that graph e

=
A
.
Proposition 10.3.5 Let P T
loc
(see 6.1.1) be any set of terms, and let
/
P
HomAlg(S) designate the class of all P-reecting homomorphisms
of Alg(S); then:
(i) (
op
(/
P
), /
P
) is a factorization system in Alg(S).
(ii) A homomorphism e : A B belongs to
op
(/
P
) iff
()
e is a full and injective epimorphism (i.e. full bimorphism)
such that there exists a full and injective homomorphism j
B
:
B F(A, TAlg(S)) satisfying j
B
e = id
A
; and if C is the
relative subalgebra of F(A, TAlg(S)) on j
B
[B], then c
C,P
A =
C.
(iii)
op
(/
P
) Ext Mod (M : t
e
= t) [ (M : t) P , and if P = P, then
equality holds.
Proof Ad (i) For each (M : t) P consider e(M : t) : M
d
t M,
the encoding epimorphism of (M : t
e
= t) (cf. 8.2.2 and 8.2.4). Then
/
P
= ( e(M : t) [ (M : t) P ), implying that
op
(/
P
)
Epi Alg(S) and that (
op
(/
P
), /
P
) is a factorization system because
of Theorem 10.2.10.
For the proofs of the other two items we only sketch the ideas and leave
details as exercises.
December 8, 2002
10 Factorization systems of partial algebras 289
Ad (ii) Knowing already that we deal with the factors of a factorization sys-
tem, we have to see how we can describe in a way unique up to isomor-
phism a decomposition of any given homomorphism, say f : A C, in
an epimorphism e : A B followed by a P-reecting homomorphism,
say g : B C.
Having already 10.3.4 at hand, and knowing that /
P
contains /
c
, we
get
op
(/
P
) Ext TAlg(S). Since the classes involved are closed with
respect to isomorphisms, we may assume that B is an A-generated rela-
tive subalgebra of F(A, TAlg(S)), and graph e =
A
. In order to make g
P-reecting we have to take to A as many elements of F(A, TAlg(S))
as are required by the reection of TE-expressions induced by P and
allowed by C. This idea leads to the formal description in (ii).
Ad (iii) For any set P of terms let F
P
:= (M : t
e
= t) [ (M : t) P .
Then it is obvious (cf. section 6) that Mod F
P
= Mod F
P
. It is also
easy to realize that e(M : t) [ (M : t) P Ext Mod F
P
;
therefore
op
(/
P
)
op
(/
P
) Ext Mod F
P
. Hence it remains to
show (cf. 10.2.15) that each (Mod F
P
)-extendable epimorphism, say
e : A B where we can assume B to be an A-generated relative
subalgebra of F(A, TAlg(S)) belongs to
op
(M
P
): But in order that
e be (Mod F
P
)-extendable, id
A
: A F(A, Mod F
P
) observe that
Mod F
P
is monoreective, since it contains TAlg(S) has to be extend-
able to B, thus B has to be isomorphic (over id
A
) to a relative subalgebra
of F(A, Mod(F
P
)). On the other hand it is not too dicult to realize
because of P being closed with respect to subterms that B is gener-
ated by A with respect to the term operations induced by the elements
of P, showing in connection with (ii) that e belongs to
op
(/
P
).
10.3.6 Proposition 10.3.5 yields an explicit description of a usually relatively
large set of factorization systems, and because of 9.2.6.(ii) not all of them
are connected with quasivarieties, namely usually not in the case, when P is
not closed with respect to subterms. Two similarly large families will follow
below, but rst we consider one single very important factorization system,
which has been studied e.g. in the thesis of A.Pasztor [Pa79] with respect
to subcategories of Alg(S).
Proposition 10.3.7 Let H
w
be the class of all surjective and o
i
the class
of all injective and initial homomorphisms. Then (H
w
, o
i
) is a factorization
system in Alg(S).
December 8, 2002
290 Chapter II A Basis for a two-valued model theory for partial algebras
gure uncompleted
Figure 10.17: f10.17 (old f10.2)
gure uncompleted
Figure 10.18: f10.18 (old f10.3)
Proof This is easy to realize, when one observes that every homomorphism
f : A B can be decomposed into a surjective homomorphism f

: A B


graph f

:= graph f, B

being the relative subalgebra of B with carrier set f[A]


followed by the full and injective (= initial and injective) embedding of B

into B, and that this factorization is unique up to unique isomorphism.


Lemma 10.3.8 Let (c, /) and (H, o) be factorization systems of Alg(S)
such that the following properties are satised:
(a) / consists of all homomorphisms, which reect some given set F of
formulas.
(b) Each h in H is surjective and preserves all formulas from F.
(c) Either c consists of epimorphisms, or o consists of monomorphisms.
Then ((H /) c, o /) = (H c, o /) is a factorization system
in Alg(S).
Proof We rst prove that the conditions of the Lemma imply
op
(/o) =
(H /) c:
For this purpose assume k
op
(/o), let (e, m) be its (c, /)-factorization,
and let (h, s) be an (H, o)-factorization of m(see Figure 10.17). Lemma 9.1.7.(iii)
implies that s /, because s h belongs to /, h is surjective and pre-
serves F. Since c Epi Alg(S) or o Mono Alg(S), either 10.1.2.(vi)
or 10.1.2.(v) imply together with s h / that h /.
Thus h H/, s o/: Moreover, since k = s h e
op
(o/), and
since either h e Epi Alg(S) or o/ Mono Alg(S) (see (c)), 10.1.2.(v)
op
or
10.1.2.(vi)
op
imply s
op
(o/). Hence the axioms of factorization systems
December 8, 2002
10 Factorization systems of partial algebras 291
gure uncompleted
Figure 10.19: f10.19 (old f10.4)
easily imply that s is an isomorphism, and s h H/, i.e. k = (s h) e
(H /) c.
Moreover, (H/) c
op
(o /): To prove this consider a commutative
diagram like the one in Figure 10.18 (without the broken lines). Then there
exists a unique d
1
such that d
1
e = p and m d
1
= q h, since m /. Since
also m o and h H (!), there exists a unique d
2
such that d
2
h = d
1
and m d
2
= q. Now d
2
h e = d
1
e = p showing that d
2
is the desired
diagonal-ll-in (uniqueness easily follows from the assumptions).
Altogether we get (H /) c =
op
(o /) (= H c, since we have seen
that we did not need in the last argument that h had to belong to /).
We still have to show that
op
(o /) o / (the other inclusion
being obvious): Let l
op
(o /); then we decompose l as shown in
Figure 10.19 and argue in a similar way as for k getting that h e ((H
/) c) (H /) c Iso Alg(S), showing that s h e o /.
Corollary 10.3.9 Let P be any set of terms for Alg(S), and let us dene
o
w
:= Mono Alg(S), H
f
:= class of all full and surjective Alg(S)-homomorphisms.
Then
((H
f
/
P
)
op
(/
P
), /
P
o
w
) and ((H
f
/
P
)
op
(/
P
), /
P
o
i
)
are factorization systems in Alg(S).
Proof The assumptions of Lemma 10.3.8 are satised for (
op
(/
P
), /
P
)
and (H
f
, o
w
) (respectively (H
w
, o
i
)).
Remark 10.3.10 When we remember that in Lemma 10.3.8 we have (H
/) c = H c, and when we dene also for later use:
H
P
:= /
P
H
w
,
H
fP
:= /
P
H
f
,
o
P
:= /
P
o
w
,
o
iP
:= /
P
o
i
,
December 8, 2002
292 Chapter II A Basis for a two-valued model theory for partial algebras
then we can formulate the result of 10.3.8 and 10.3.9 also in the shorter
form: For every set P of terms (
op
(/
P
), /
P
), H
fP

op
(/
P
), o
P
), and
H
P

op
(/
P
), o
iP
), are factorization systems.
10.3.11 (Some further examples of factorization systems) We briey
present some further examples of factorization systems in Alg(S), which
might be of interest, but which will not be much used later in this book.
In this connection we recall that because of Theorem 10.2.10 all classes of
homomorphisms which are dened by the reection of a given set of elemen-
tary implications are right factors of factorization systems, where the left
factor consists of a class of epimorphisms. In what follows we want to give a
brief description of some such left factors.
(i) Let /
i
be the class of all initial homomorphisms in Alg(S) (see 9.2),
let f : A B be a homomorphism, and let (s, j) be a (H
w
, o
i
)-
factorization of f (see 10.3.7). If then (e

, i

) is a (
op
(/
i
), /
i
)-
factorization of s, then (e

, j i

) is a (
op
(/
i
), /
i
)-factorization of f,
and we can restrict our considerations to the case, when f is surjective:
Let (id
A
, f

) be a (Ext TAlg(S), /
c
)-factorization of f (cf. 10.3.4), ob-
serve that every initial homomorphism is relatively closed, and every
surjective and initial homomorphism is closed. Assume that (e, i) is the
(
op
(/
i
), /
i
)-factorization we look for, where e : A C. Since for
surjective f also i is surjective and being initial also closed, C has
to be a closed homomorphic image of domf

(see 5.8.2.(vi)) in such a


way that for a


C
[C
()
] [ one has #(i
1
(i a) = 1. Since
i is closed, we get i[


C
[C
()
] [ =


B
[B
()
] [ .
Hence let := (a, b) [ a, b domf

, f

(a) = f

(b)


B
[B
()
] [
. Then is a closed congruence on domf

, and it is not
too dicult to see that C

= (domf

)/. These observations show


perhaps together with some additional calculations that
op
(/
i
)
consists of all surjective (!) epimorphisms e : A C of Alg(S) satisfy-
ing that c C


C
[C
()
] [ implies #(e
1
(c)) = 1.
(ii) Relatively normal homomorphisms are dened to reect all formulas
(fvar t y : t
e
= y), where t is any term and y / fvar t any variable of
correct sort; in particular (x, y : x
e
= y), for x ,= y any variables of the
same sort, is such a formula; thus relatively normal homomorphisms are
injective; moreover, they are initial. Let /
n
be the class of all relatively
December 8, 2002
10 Factorization systems of partial algebras 293
normal Alg(S)-homomorphisms. We claim that
op
(/
n
) consists of
all epimorphisms e : A C such that for every c C there exist a
term (M : t), a sequence c dom(M : t)
C
and an m / such that
c(m) = c and (M : t)
C
(c) e[A] equivalently: e
op
(/
n
) iff
e : A C) is an epimorphism such that (e[A]) = C with respect to
the algebraic quasiorder on C.
(iii) We have not yet seen the necessity to work out a detailed description of
the left factors of P-initial (= P-normal) homomorphisms respectively
injective and P-normal homomorphisms.
(iv) Final homomorphisms form the left factor of a factorization system
in Alg(S) such that this left factor does not only consist of epimor-
phisms. It is not dicult to realize that the corresponding right factor
consists of the class of all bijective homomorphisms in Alg(S).
10.3.12 We conclude this Chapter by presenting in Table 10.1 a survey on
some factorization systems in Alg(S) described earlier in this section (the
rst eight numbers correspond to those in Table 9.1) and in Figure 10.20 the
closure system on HomAlg(S) generated by the most interesting classes of
homomorphisms considered so far (or which will be considered in sections
12 and 13). In addition to the notation already introduced, let us use the
following one: c
w
designates the class of all epimorphisms, 1 the class of
all isomorphisms, / (with some index) stands for the class of all homo-
morphisms with the additional property indicated in the index, similarly, H
(with some index) designates a class of surjective and o (with some index)
a class of injective homomorphisms with the additional properties shown in
the index. The additional properties are abbreviated by: f for full, i
for initial, c for closed, f
i
for nal, b for bijective, and P
for P-reecting it should be observed that P is a parameter, and that
in connection with Figure 10.20 we have separating examples for the cases
when the similarity type concerns at least two at least unary operation sym-
bols, say and

, such that for some sequence x of variables (x) belongs


to P, while for all admissible sequences t of terms,

(t) does not belong


to P (for more details see [BWo82a]; the nal version of Figure 10.20 is
due to R.Wille (see also [Wl84] for the general principles of drawing such
concept lattices)).
Observe that already c
w
, /
c
, /
f
, /
i
, /
f
i
, /
P
, H
w
and o
w
generate
this closure system.
December 8, 2002
294 Chapter II A Basis for a two-valued model theory for partial algebras
No Left factor: Class of all . . . Right factor: Class of all . . .
1 full and surj. homs (H
f
) inj. homs (o
w
)
2 TAlg(S)-ext. homs (A B, B

= over id
A
to some A-initial segment of F(A, TAlg(S)))
closed homs (/
c
)
3 epis (c
w
) closed monos (o
c
)
4 surj. homs f : A B, where
a B [B
()
] [ implies #(f
1
(a)) = 1
initial homs (/
i
)
5 surj. homs (H
w
) initial monos (o
i
= o
f
)
6 epis f : A B, where for b B there exist
(M : t), b dom(M : t)
B
, m M: b = b(m) and
(M : t)
B
(b) f[A]
relatively normal homs (/
n
)
7 epis f : AB, where B is isom. over id
A
to a
P-generated A-initial segment of F(A, TAlg(S))
P-re. homs (/
P
)
7 (Mod F
P
)-ext. epis P-re. homs (/
P
)
8 not yet described in detail (problem) P-initial (= P-normal) homs
9 e H
f

op
(/
P
) P-re. monos (o
P
)
10 e H
w

op
(/
P
) P-re. initial homs (/
iP
)
11 nal homs (/
f
i
) bijective homs (H
b
)
Table 10.1: Factorization systems in TAlg(S)
(see 10.3.12)
December 8, 2002
10 Factorization systems of partial algebras 295
gure uncompleted
Figure 10.20: f10.20 (old f10.5) (see 10.3.12)
December 8, 2002
296 Chapter II A Basis for a two-valued model theory for partial algebras
December 8, 2002
Chapter III
Birkho type results for
elementary implications
As can already be realized from the examples in Part II (and also from those
in Chapter II of this part), the role of equations as most important axioms
in the total case is taken over by elementary implications in the case of par-
tial algebras, in particular by ECE-equations. Now H.Andr eka, I.N emeti
and I.Sain have shown in [AN83], [AN82] and [NSa82] that category theo-
retical methods help to provide a meta-theorem which yields Birkho type
theorems for a wealth of special kinds of elementary implications. This meta-
theorem will be the core of this chapter and is presented and proved in
section 12, while the rest of this chapter is then dedicated to some of its
applications within the category Alg(S) of partial algebras.
In section 11 we rst revisit direct limits and introduce some concepts
which shall be needed in connection with the main result. In particular
morphisms between directed systems are introduced and investigated, and it
is investigated how properties of homomorphisms are inherited by the col-
imiting cocone (respectively, the induced morphisms) between direct limits.
A very important feature of direct limits is the fact that they provide the
possibility of measuring smallness. At the end we prove Theorem 8.3.4.
In section 12 we formulate and prove the Meta-Birkho Theorem for
elementary implications, and a rst attempt is made to test some concepts in
the theory of partial algebras as to whether they could be used in connection
with this theorem.
In section 13 we provide as many examples as possible within the cate-
gory Alg(S) for the main theorem of section 12, i.e. of characterizations of
297
298 Chapter III Birkho type results for elementary implications
classes of partial algebras axiomatizable by special elementary implications.
Finally, in section 14 we present descriptions of closed classes of some
special kinds of elementary implications.
11 Direct limits revisited
In this section we continue the study of direct limits started in subsection 4.4,
and we especially investigate how they behave in connection with model the-
oretic concepts introduced in the meantime. We shall use again the notation
introduced in subsection 4.4, and the reader is advised to reread this subsec-
tion.
11.1 Morphisms between directed systems
Denition 11.1.1 Let C be any category, I := (I; ) and J := (J; ) be
directed sets, and A
I
:= ( A
i
, f
il
: A
i
A
l
[ i, l I, i l ) and B
J
:=
( B
j
, g
jl
: B
j
B
l
[ j, l J, j l ) be two directed systems in C. Then a
pair (m, ( m
i
[ i I )) is called a morphism between directed systems, here
from A
I
into B
J
iff :
(D1) m : (I; ) (J; ) is a monotone mapping, i.e. i i

implies
m(i) m(i

).
(D2) For every i I, m
i
: A
i
B
m(i)
is a C-morphism.
(D3) For any i, i

I with i i

one has m
i
f
ii
= g
m(i)m(i

)
m
i
(i.e. a
compatibility condition between the two systems).
Such a morphism is called conal, if m[I] is conal in J, i.e. if for every j J
there is i I such that j m(i).
If I = J and m = id
I
, then we call (id
I
, ( m
i
[ i I )) a special morphism.
If I J,
I
=
J
(II) and m
i
= id
A
i
for each i I, then (id
I
, ( id
A
i
[
i I )) is called an inclusion morphism between the given directed systems.
11.1.2 Notice that for a directed system A
I
(with direct limit (( f
i
[ i
I ), A)) and a sink ((m
i
: A
i
B)
iI
, B), this family can be considered as a
special morphism from A
I
into B
I
:= (B
i
:= B, g
ij
:= id
B
[ i j I). This
observation will often tacitly be used.
December 8, 2002
11 Direct limits revisited 299
Lemma 11.1.3 The class of all morphisms (conal morphisms, respectively
special morphisms) between directed systems (for any xed poset I, if neces-
sary) is closed with respect to composition and therefore forms, together with
the corresponding directed families, a category. In particular, for directed
systems A
I
, B
J
and C
K
and morphisms m

= (m, ( m
i
[ i I )) : A
I
B
J
and n

= (n, ( n
j
[ j J )) : B
J
C
K
we dene their composition as
n

:= (n m, ( n
m(i)
m
i
[ i I )) : A
I
C
K
.
Proof (D1) and (D2) are easily veried. For (D3) (draw a diagram!) we
compute:
n
m(i

)
m
i
f
ii
= n
m(i

)
g
m(i)m(i

)
m
i
= h
nm(i)nm(i

)
n
m(i)
m
i
.
If m = n = id
I
and I = J = K, then n m = id
I
. If m

and n

are conal,
and if we consider k K, then there are j J, i I such that k n(j),
j m(i), i.e. k n(j) nm(i), showing that n

is conal.
Proposition 11.1.4 Let A
I
, B
J
and C
K
be directed systems in Alg(S), and
let m

: A
I
B
J
and n

: B
J
C
K
be morphisms between them. Moreover,
let (( f
i
[ i I ), A), (( g
j
[ j J ), B) and (( h
k
[ k K ), C) be direct limits
of A
I
, B
J
and C
K
, respectively. Then:
(i) m

, n

and n

induce unique homomorphisms m : A B, n : B


C and p : A C, respectively, such that m f
i
= g
m(i)
m
i
(i I),
n g
j
= h
n(j)
n
j
(j J), and p f
i
= h(i)
nm(i)
n
m(i)
m
i
(i I),
respectively. Moreover: n m = p.
(ii) If m

is conal, and if all m


i
(i I) are isomorphisms, then m is
an isomorphism. If, in addition, I J, m
i
= id
A
i
and f
ij
= g
ij
(i j I), then one may choose (( f
i
[ i I ), A) := (( g
i
[ i I ), B).
(iii) Let o HomAlg(S) such that (
op
(o),
op
(o) is a factorization sys-
tem, let m

be conal, and assume m


i
[ i I
op
(o), then
m
op
(o) (it is sucient that, for some i
0
I, m
i
[ i
0
i I

op
(o) ).
(iv) Let o HomAlg(S) such that (
op
(o),
op
(o) is a factorization sys-
tem, let f
ii
[ i i

I
op
(o), then f
i
[ i I
op
(o).
December 8, 2002
300 Chapter III Birkho type results for elementary implications
Proof Ad (i) For each i I we have g
m(i)
m
i
: A
i
B such that, for
i j I, one has g
m(j)
m
j
f
ij
= g
m(j)
g
m(i)m(j)
m
i
= g
m(i)
m
i
;
from (DL3) we therefore infer the existence of a unique m : A B such
that, for each i I, one has m f
i
= g
m(j)
m
i
. In the same way we get
n and p with respect to n

and n

. Since p f
i
= h
nm(i)
n
m(i)
m
i
=
n g
m(i)
m
i
= n m f
i
for each i I the uniqueness of the induced
morphism in (DL3) implies p = n m.
Ad (ii) Let m

be conal, let each m


i
be an isomorphism and let m : A B
be the induced homomorphism according to (i). We observe that, for
each j J, we have some i I satisfying m(i) j; therefore q
j
:=
f
i
m
1
i
g
jm(i)
: B
j
A is a homomorphism which is independent from
the choice of i: if we have i

I such that m(i

) j, then there is l I,
i, i

l, and we get
q
j
= f
i
m
1
i
g
jm(i)
= f
l
f
il
m
1
i
g
jm(i)
= f
l
m
1
l
g
m(i)m(l)
g
jm(i)
= f
l
m
1
l
g
jm(l)
= . . . = f
i
m
1
i

g
jm(i

)
.
In the same way we get, for j j

in J and for l I satisfying j

m(l):
q
j
g
jj
= f
l
m
1
l
g
j

m(l)
g
jj
= f
l
m
1
l
g
jm(l)
= q
j
. Thus ( q
j
[ j
J ) induces a homomorphism q : B A. Now it is not dicult to
conclude (because of the uniqueness of the induced homomorphisms)
that q m = id
A
and m q = id
B
(since m
1
i
m
i
= id
A
i
and m q g
j
=
m f
i
m
1
i
g
jm(i)
= g
m(i)
m
i
m
1
i
g
jm(i)
= g
j
id
B
j
= id
B
g
j
).
Ad (iii) Assume m
i
[ i I
op
(o), let s S, s : U V , a, b
HomAlg(S) such that s a = b m. Then (draw a diagram!) for each i
I: b g
m(i)
m
i
= b m f
i
= s a f
i
; because of Dip(m
i
, s) there exists
a unique d
m(i)
: B
m(i)
U such that d
m(i)
m
i
= a f
i
and s d
m(i)
=
b g
m(i)
. Dening, for arbitrary j J, d
j
: B
j
U by d
j
:= d
m(i)
g
jm(i)
for some i I, it is not dicult to realize that (( d
j
[ j J ), U) is
compatible with B
J
, therefore inducing a unique homomorphism d :
B U such that for each j J one has d g
j
= d
j
. Since for each
i I we have a f
i
= d g
m(i)
m
i
= d m f
i
, the uniqueness property
in (DL3) implies a = d m. Similarly we get, for each j J, b g
j
=
s d
j
= s d g
j
, and therefore, because of (DL3) and (ii), b = s d
showing that m
op
(o).
Ad (iv) Assume now that f
ii
[ i i

I
op
(o), let i I, s o,
s : U V , and a, b HomAlg(S) such that b f
i
= s a. Observe
that I
i
:= i

I [ i i

is conal in I. For each i

I
i
we have
December 8, 2002
11 Direct limits revisited 301
s a = b f
i
= b f
i
f
ii
and f
ii

op
(o). Therefore there exists a
unique d
i
: A
i
U such that d
i
f
ii
= a and s d
i
= b f
i
. Let
i

i; then d
i
f
i

i
f
ii
= a = d
i
f
ii
, and s d
i
= b f
i
=
b f
i
f
i

i
= s d
i
f
i

i
. 10.2.2.(iii) then implies d
i
f
i

i
= d
i
, and
we therefore get a unique d : A U such that d f
i
= d
i
for each
i

I
i
. Since d
i
= a, this yields d f
i
= d
i
. For each i

I
i
we have, in
addition, s d f
i
= s d
i
= b f
i
. Therefore the uniqueness in (DL3)
implies s d = b, showing that f
i

op
(o).
From 11.1.4.(iv) and 10.3 (see Table 10.1) we infer the
Corollary 11.1.5 Let (A
i
, ( f
ij
[ i j I )) be a directed system with direct
limit (( f
i
[ i I ), A). Assume that each f
ij
is surjective, full and surjective,
respectively an epimorphism (or has any other property described on the left
hand side of Table 10.1). Then each f
i
(i I) has the same property.
Similarly 11.1.4.(iii) implies together with 10.3
Corollary 11.1.6 Let (m, ( m
i
[ i I )) : A
I
B
J
be a conal morphism
between directed systems, and let m be the induced homomorphism between
the corresponding direct limits. Then one has: If each m
i
(i I) is surjective,
full and surjective, respectively an epimorphism (or has any other property
described on the left hand side of Table 10.1), then m has the same property.
11.2 A device to measure smallness
Results similar to those of Proposition 11.1.4 and its corollaries can also be
obtained with respect to the operator , but not in the same general-
ity. As a preparation we need the following denition from Banaschewski
and Herrlich (see [BaH76], Andr eka and N emeti [AN83] and other pa-
pers):
Denition 11.2.1 Let B Ob C for some category C, and let k be any
innite regular cardinal. Then B is called k-small iff for every k-directed
system A
I
= (A
i
, f
ij
[ i j I ) in C with k-direct limit
()
(( f
i
[ i I ), A), and for every C-morphism g : B A there
exist some i I and h : B A
i
such that f
i
h = g.
December 8, 2002
302 Chapter III Birkho type results for elementary implications
B is called strongly k-small iff one has in addition to ():
()
For any k-directed system A
I
with k-direct limit as in () and
for any two C-morphisms h, h

: B A
i
satisfying f
i
h =
f
i
h

there exists j i in I such that already f


ij
h = f
ij
h

.
Proposition 11.2.2 Let k d

and B Alg(S). Then the following state-


ments are equivalent:
(i) B is k-small.
(ii) B is strongly k-small.
(iii) #B < k and #(( dom
B
[ )) < k.
Proof (iii) (ii) Let (iii) be satised, let A
I
be a k-directed system with
direct limit (( f
i
[ i I ), A). Let g : B A be any homomorphism.
Because of the d

-directedness of I (d

k) we know by 4.4.4 that there


exists, for every b B, some i(b) I such that g(b) = f
i(b)
(a

b
) for
some a

b
A
i(b)
. By the k-directedness of I we get some j in I such
that having chosen i(b) and a

b
for each b B we have i(b) j and
g(b) = (f
j
f
i(b)j
)(a

b
). Dene h

: B A
j
by h

(b) := f
i(b)j
(a

b
).
Let b dom
B
, then g(b) dom
A
, and 4.4.4.(ii) implies the existence
of i(b) I such that there is a
b
dom
A
i(b)
satisfying f
i(b)
a
b
= g b.
Since I is k-directed for an innite regular cardinal k there exists, because
of the assumptions about B, some i I such that j, i(b) i for each
and for each b dom
B
. Then it is easy to realize that h :=
f
ji
h

: B A
i
is a homomorphism which satises f
i
h = g. This
shows that B is k-small.
Now assume that p, q : B A
i
are two homomorphisms for some
i I such that f
i
p = f
i
q. Then a similar argument as above shows
the existence of some j i in I such that (f
ij
p)(b) = (f
ij
q)(b) for all
b B, i.e. f
ij
p = f
ij
q, showing that B is even strongly k-small.
(ii) (i) is obvious.
(i) (iii) (by contraposition): Assume that B violates (iii). Then (by 4.4.7)
B =
lim

A
I
3
(B,B)
where each C in I
3
(B, B) satises (iii), and B can-
not be isomorphic to any such C. By choosing g := id
B
: B
lim

A
I
3
(B,B)
= B we realize that there cannot exist any homomorphism
h : B C for any C I
3
(B, B) such that id
C
h = id
B
= g.
December 8, 2002
11 Direct limits revisited 303
Remarks 11.2.3 (i) Observe that in 11.2.2 the arguments proving (iii)
(ii) (i) remain true in general for arbitrary innite regular car-
dinals, and therefore the objects described in 11.2.2.(iii) are always
strongly k-small in Alg(S). But the argument which we have used in
proving (i) (iii) fails in general (see 4.4.5). We did not check,
whether 11.2.2 remains true also for
0
k d

, since we shall not


need this later. Thus we leave it as a problem to the reader.
(ii) In the case of total (heterogeneous) algebras the classes of k-small and
strongly k-small algebras do not coincide. Without proof we state that
for k d

an algebra A TAlg(S) is
(a) strongly k-small in TAlg(S), iff #A k; and it is
(b) k-small, iff A is isomorphic to the quotient T(M, TAlg(S))/ of
a k-small generated Peano algebra (#M < k) with respect to a
k-small generated congruence .
(iii) From 11.2.2 it easily follows that full homomorphic images of (strongly)
k-small partial algebras (k d

) are also (strongly) k-small.


(iv) For historical remarks also for the following concept see [NSa82].
Denition 11.2.4 A category C is called strongly k-algebroidal (see [BaH76]),
iff each C-object is the direct limit of strongly k-small objects.
11.2.5 4.4.7 and 11.2.3 together show that Alg(S) is strongly k-algebroidal
for every regular cardinal k d

.
11.3 Direct limits and the operator
Now we are able to prove an analogue to Proposition 11.1.4.(iii) and (iv):
Proposition 11.3.1 Let A
I
= (A
i
, f
ik
[ i k I ) and B
J
= (B
j
, g
jl
[
j l J ) be k-directed systems (k d

, regular), and let (( f


i
[ i I ), A)
respectively (( g
j
[ j J ), B) be their k-direct limits. Moreover, let m

=
(m, ( m
i
[ i I )) : A
I
B
J
be a conal morphism between directed systems.
Finally assume c HomAlg(S) such that domain and codomain of each c-
morphism are strongly k-small, and that (
op
(c), (c)) is a factorization
system. Then we obtain:
December 8, 2002
304 Chapter III Birkho type results for elementary implications
(i) If each m
i
belongs to (c) (for all i i
0
in I for some i
0
), then so
does the colimit morphism m : A B.
(ii) If f
ik
[ i k I (c), then f
i
[ i I (c).
Proof Ad (i) Let e : X Y be any c-morphism and a, b HomAlg(S)
such that m a = b e. Since X and Y are strongly k-small, there exist
i I, j J, a

: X A
i
, b

: Y B
j
such that a = f
i
a

and b = g
j
b

.
Since m[I] is conal in J, since I and J are directed, and since X and Y
are strongly k-small, there exists k I such that m(i), j m(k) and
m
k
f
ik
a

= g
jm(k)
b

. m
k
(c) implies the existence of a unique
d
k
: Y A
k
such that f
ik
a

= d
k
e and m
k
d
k
= g
jm(k)
b

. Therefore,
a = f
k
f
ik
a

= f
k
d
k
e and m f
k
d
k
= g
m(k)
m
k
d
k
= b, showing
that d := f
k
d
k
is the diagonal-ll-in we have looked for. 10.2.2.(iii)
and 10.2.5 then imply the uniqueness of d.
Ad (ii) Let i I be arbitrary, consider e c as in the proof of (i) and
a, b HomAlg(S) such that f
i
a = b e. The facts that Y is (strongly)
k-small and A
I
is k-directed imply the existence of k I (w.l.o.g. k i)
and b

: Y A
k
such that f
k
b

= b. Since X is strongly k-small


there exists j k such that f
ij
a = f
kj
b

e. W.l.o.g. k = j, i.e.
f
ik
a = b

e. Since f
ik
(c), there exists a unique d : Y A
i
such
that d e = a and f
ik
d = b

. Then f
i
d = f
k
f
ik
d = f
k
b

= b. Hence
Dip(e, f
i
). As in (i) we can infer the uniqueness of d.
From section 9 (cf. Table 9.1) and section 10 (see Table 10.1) we can infer
the
Corollary 11.3.2 (1) Let (A
i
, f
ij
[ i j I ) be a d

-directed system
in Alg(S), and let (( f
i
[ i I ), A) be its direct limit. If all f
ij
are injective,
closed, initial, relatively normal, P-reecting, respectively P-initial (for some
set P of terms), then each f
i
has the same property.
Corollary 11.3.3 (2) Let (m, ( m
i
[ i I )) : A
I
B
J
be a conal mor-
phism between k-directed systems, and let m :
lim

A
I

lim

B
J
be the
induced homomorphism. If each m
i
is injective, closed, initial, relatively
normal, P-reecting, respectively P-initial (for some set P of terms), then
m has the same property.
December 8, 2002
11 Direct limits revisited 305
11.3.4 The results of 11.1.4 and 11.3.1 can also be applied in connection
with properties, which are not necessarily directly related to a factorization
system, but only so-called relativized properties like fullness or relative
closedness:
f : A B is full, iff for every (H
f
, o
w
)-factorization (s, m) of f (i.e.
f = m s) the monomorphism m is initial;
f : A B is relatively closed, iff for every (H
w
, o
i
)-factorization (s, m)
of f the epimorphism s is closed;
f : A B is relatively P-reecting, iff for every (H
w
, o
i
)-factorization
(s, m) of f the epimorphism s is P-reecting.
Thus the following results will show that all these properties are preserved in
the sense of 11.3.2 as well as in the sense of 11.3.3 by k-direct limits (k d

)
of k-directed systems.
Lemma 11.3.5 With the same general assumptions and notation on A
I
, B
J
and m

as in 11.3.1 let (H, o) := (


op
(c), (c)) be any factorization system
as in 11.3.1. Moreover, for each i I let (h
i
, s
i
) be an (H, o)-factorization
of m
i
, h
i
: A
i
E
i
, s
i
: E
i
B
i
. Then there exists for i j I a
homomorphism e
ij
: E
i
E
j
such that, for i j k I, one has e
ij
h
i
=
h
j
f
ij
, g
m(i)m(j)
s
i
= s
j
e
ij
, e
ii
= id
E
i
and e
jk
e
ij
= e
ik
, i.e. A
I
:= (E
i
, e
ij
[
i j I ) is a k-directed system, and h

:= (id
I
, ( h
i
[ i I )) : A
I
A
I
and s

:= (m, ( s
i
[ i I )) : A
I
B
J
are conal morphisms between k-
directed systems, and the induced homomorphisms

h :
lim

A
I

lim

A
I
and s :
lim

A
I

lim

B
J
belong to H, respectively to o the assumptions
concerning the existence of c are only needed for s S.
Proof Since for i j I we have g
m(i)m(j)
s
i
h
i
= g
m(i)m(j)
m
i
=
m
j
f
ij
= s
j
h
j
f
ij
, where h
i
H and s
j
o, there exists a unique
diagonal-ll-in e
ij
: E
i
E
j
such that the usual commutativity relations
hold. The uniqueness of diagonal-ll-ins also easily implies e
jk
e
ij
= e
ik
for i j k I. The special assumptions on (H, o) imply together
with 11.1.4.(iii) respectively 11.3.1.(i) that

h H and s o.
December 8, 2002
306 Chapter III Birkho type results for elementary implications
11.3.6 Since (H
w
, o
i
) and (H
f
, o
w
) satisfy the assumptions on (H, o) in
Lemma 11.3.5, it is now easy to realize that k-direct limits (k d

) of k-
directed systems preserve fullness, relative closedness or P-reection, since
they preserve surjectivity (together with fullness), injectivity, initialness,
closedness and P-reection.
11.4 Sections, retractions and direct limits
The following two properties of morphisms are often useful in proving that
some morphism is in fact an isomorphism:
Denition 11.4.1 Let C be any category, A, B Ob C, s : A B , r :
B A C-morphisms. s is called a section iff there is r

: B A such that
r

s = 1
A
, and r is called a retraction iff there is s

: A B such that
r s

= 1
A
.
If r is a retraction or s a section, then A is called a retract of B.
11.4.2 It should be observed that every section is a monomorphism, and that
every retraction is an epimorphism. Moreover, in 11.4.1 r

is a retraction and
s

is a section, too.
The simple category theoretical proof of the following statement is left as
an exercise (see [HS73], 6.7 and 6.15):
Proposition 11.4.3 Let C be any category, f Mor C. Then the following
statements are equivalent:
(i) f is an isomorphism.
(ii) f is a section and an epimorphism of C.
(iii) f is a retraction and a monomorphism of C.
Proposition 11.4.4 In Alg(S) every section is injective and closed, and
every retraction is full and surjective.
Proof Let s : A B be a section with corresponding retraction r : B A.
Since r s = id
A
is closed and bijective, s is closed and injective, and r is
surjective. Since the restriction of r to s[A] is still closed and surjective, it is
easy to realize that r is full.
December 8, 2002
11 Direct limits revisited 307
Proposition 11.4.5 Let m

: A
I
A
I
and n

: A
I
A
I
be special mor-
phisms between directed systems, and let m :
lim

A
I

lim

A
I
respectively
n :
lim

A
I

lim

A
I
be the induced colimit homomorphisms. Assume that
each m
i
is a section and n
i
its corresponding retraction. Then m is a section
and n its corresponding retraction.
Proof Since n
i
m
i
= id
A
i
, we get n m = id by 11.1.4.(i) and (ii).
11.5 Some further concepts connected with direct lim-
its
In this subsection we describe some category theoretical concepts and their
partial algebraic interpretations in Alg(S), which will be needed for the meta
theorem of H.Andr eka, I.N emeti and I.Sain on implicationally denable
classes, which we shall present in the next section. In 11.2.5 we have already
realized that Alg(S) is a strongly d

-algebroidal category.
Denition 11.5.1 Let C be any category, k any innite cardinal. A class o
of C-morphisms is called k-inductive iff any k-direct limit of o-subobjects
is again an o-subobject. I.e.: o is k-inductive iff for any k-directed system
(A
i
, f
ij
[ i j I ) with direct limit (( f
i
[ i I ), A) and for every
compatible sink (( m
i
: A
i
B [ i I ), B) with m
i
[ i I o one has
m o for the induced C-morphism m : A B.
Remarks and Examples 11.5.2 Propositions 11.1.4.(iii) and 11.3.1.(i) tell
us that in Alg(S) we have: For every factorization system (c, /) of Alg(S),
c is k-inductive for every innite regular cardinal k. Moreover, / is k-
inductive for k d

, if / = (c

), where c

c, and for each c

-morphism,
say e : X Y , X and Y are strongly k-small.
In particular, the classes of all epimorphisms, all surjective homomor-
phisms, respectively of all full and surjective homomorphisms are k-inductive
for every innite regular cardinal k; and the classes of all injective homo-
morphisms, of all closed, closed and injective, initial, P-reecting, injective
and initial, injective and P-reecting respectively relatively normal homo-
morphisms are d

-inductive and k-inductive for every regular cardinal k d

.
Denition 11.5.3 Let C be any category and k any innite regular cardinal.
A class c of C-morphisms is called k-dual algebroidal (cf. [AN82], p.16f) iff
December 8, 2002
308 Chapter III Birkho type results for elementary implications
gure uncompleted
Figure 11.1: f11.1
every c-morphism e : A B for which A is strongly k-small is a direct
limit of c-morphisms f
ij
for which Dom(f
ij
) as well as Cod(f
ij
) are strongly
k-small. More precisely: c is k-dual algebroidal iff , for every c-morphism
e : A B with strongly k-small domain A, there exists a directed system
(A
i
, f
ij
[ i j I) with direct limit (( f
i
[ i I ), B) such that there is i

I
satisfying e = f
i
, and for i j I the domain as well as the codomain of
f
ij
are strongly k-small, and f
ij
c.
Remark 11.5.4 Notice that the above concept is a generalization of the fact
that every (c-) congruence relation is a directed union of nitely (or d

-small)
generated (c-) congruence relations (cf. 4.4.11). If c consists of all full and
surjective homomorphisms, then 4.4.11, 11.2.2 and 11.2.3.(iii) show that c
is d

-dual algebroidal. And in combination with 4.4.7 we may conclude that


many classes of epimorphisms in Alg(S) are d

-dual algebroidal, as we shall


see in section 13.1.
11.6 Proof of Theorem 8.3.4
The reader should recall the contents of subsection 8.3 and observe, that
the statements of Theorem 8.3.4 concerning the operators T, T
+
and e have
already been proved in 8.3.
11.6.1 The operators H (= H
w
) and H
f
: That (full) homomorphic images
preserve c-equations, has been shown in 6.2.1.(i). That ECE-equations are
in general not preserved is shown by the example sketched in Figure 11.1 for
= (1) and F = (x : x
e
= x x
e
= x).
The homomorphism of Figure 11.1 shows that negations of TE-expressions
here (x : (x
e
= x)) are not preserved by (full) homomorphic im-
ages.
11.6.2 That there is some sort of ECE-equations which is preserved by full
homomorphic images will be discussed in section 13 (e.g. in 13.3).
December 8, 2002
11 Direct limits revisited 309
gure uncompleted
Figure 11.2: f11.2
gure uncompleted
Figure 11.3: f11.3
11.6.3 The operator H
c
: By denition closed homomorphisms preserve all
negations of TE-expressions (since they reect all TE-expressions). Therefore
also arbitrary disjunctions of such formulas are preserved (cf. 9.1.7.(iv)).
Thus, if f : A B is a closed and surjective homomorphism, if A [= (M :
_

(t
i
e
= t
i
) [ i I ), and if w : M B is any assignment, then the
surjectivity of f implies that we can choose w

: M A (always compatible
with the sort mappings!) such that w = f w

. Since f preserves our formula


and A [= (M :
_

(t
i
e
= t
i
) [ i I )[w

], we get B [= (M :
_

(t
i
e
= t
i
) [
i I )[f w

] (w = f w

).
Moreover, we show below that surjective and closed homomorphisms
preserve ECE-equations, and therefore the same argument as above shows
that H
c
preserves ECE-equations: Assume that A [= (M :
_

t
i
e
= t
i
[
i I t
e
= t

), and let w : M B be an assignment, such that


B [= (M :
_

t
i
e
= t
i
[ i I )[w]. As above we have w

: M A
such that f w

= w. Since f reects (M :
_

t
i
e
= t
i
[ i I ), we get
A [= (M :
_

t
i
e
= t
i
[ i I )[w

] and therefore A [= (M : t
e
= t

)[w

]. Thus
B [= (M : t
e
= t

)[f w

], since f preserves E-equations.


Finally we have to show that closed homomorphic images preserve neither
negations of E-equations nor QE-equations. For this purpose consider, for
:= (1), the formulas F := (x : (x
e
= x)) and G := (x : x
e
=
x x
e
= x), and consider Figure 11.2, where A := (N; ) is the set of
natural numbers with successor operation . Then F and G are valid in A,
but not in B.
11.6.4 The operators o
w
and o
r
: Let f : A B be an injective homo-
morphism. Since f preserves atomic formulas, it reects negations of them.
Thus, if in B the negation of an E-equation is valid, then this is also valid
in A. Now consider F := (x : x
e
= x), ( := (1)), then Figure 11.3
shows that (weak) relative subalgebras do not preserve TE-expressions.
December 8, 2002
310 Chapter III Birkho type results for elementary implications
11.6.5 The operators o (= o
c
) and 1: Since closed monomorphisms re-
ect all elementary implications (i.e. epimorphisms) and therefore also their
disjunctions, 9.1.8.(ii) implies that subalgebras preserve all elementary im-
plications. Moreover, they reect negations of E-equations and therefore
also their disjunctions. Thus 9.1.8.(ii) implies that subalgebras preserve all
universal Horn formulas. The same trivially holds for isomorphic copies.
In preparation for the next step we prove the following
Lemma 11.6.6 Let F := (M :
_

t
k
e
= t

k
[ k K
_

t
l
e
=

t

l
[
l L) be any elementary implication with #K, #L < k for some innite
regular cardinal k d
&
, and let e : P C be its encoding epimorphism.
Let A
I
:= (A
i
, f
ij
[ i j I ) be a k-directed system with k-direct limit
(( f
i
[ i I ), A), and assume that A
i
[= F for all i I. Then A [= F.
Proof Observe that the assumptions on F imply that P and C are strongly
k-small, that each f
ij
trivially preserves and reects F, and that each f
i
reects F, because F is valid in A
i
. Now, assume that we have g : P A.
Since P is strongly k-small, there exist i I and g

: P A
i
such that
g = f
i
g

. Since F is valid in A
i
, there exists h

: C A
i
such that
g

= h

e. Thus f
i
h

e = f
i
g

= g. This shows that A [= F.


11.6.7 The operators T
r,k
, T
r,k
+
, T
u
and
lim

k
: Since the sort algebra S
always belongs to T
r,k
K for any class K, T
r,k
cannot preserve negations of
TE-expressions. If k d
&
, c (c being the cardinal connected with the
language L, see 7.1.1) is an innite regular cardinal, then Lemma 11.6.6 and
the denition of k-reduced products in 4.5.3 show that the operators T
r,k
, T
r,k
+
and
lim

k
preserve elementary implications (restricted by c). Thus it remains
to show that T
r,k
+
preserves formulas of the kind F := (M : (
_

t
k
e
= t

k
[
k K)) (#K < c):
Let P be the relative subalgebra of F(M, TAlg(S)) with carrier set MF.
Then we recall that validity of F in A means that there is no homomorphism
from P into A. Let ( A
i
[ i I ) be a non-empty family of partial algebras
such that F is valid in each A
i
, and let ( be a proper k-complete lter on I.
Since P is strongly k-small (even strongly c-small), there cannot exist any
homomorphism from P into

k
( A
i
[ i I )/(, since this would imply the
existence of a homomorphism from P into some A
i
. Hence in this k-reduced
December 8, 2002
11 Direct limits revisited 311
product F is also valid. Since ultralters are proper lters, the index set
has to be non-empty; thus T
u
is contained in the operator T
r,
0
+
, and like this
operator it preserves all kinds of nite formulas under consideration.
Remarks 11.6.8 (i) The statement about ultraproducts also follows from
the fact that the Ultraproduct Theorem or Los Lemma holds in all cat-
egories (cf. [AN78]):
A nitary and nite formula F holds in an ultraproduct A :=

0
( A
i
[
i I )/( iff i I [ A
i
[= F (.
(ii) We would also like to mention without proof that formulas of the kind
(M :
_

t
k
e
= t

k
[ k K
_

t
l
e
=

t

l
[ l L) for an at least
two-element set L are only preserved by the operators o
c
, 1, e and T
u
.
December 8, 2002
312 Chapter III Birkho type results for elementary implications
12 A Meta Birkho Type Theorem
12.1 HoT
r,k
-cones
12.1.1 In a series of papers (see [AN83], [AN82] or [NSa82]) H.Andr eka,
I.N emeti and I.Sain have presented category theoretical axiomatizabil-
ity results which yield Birkho type results of the form Mod Cone
HSF
K =
HoT K, where His an operator of H-images, o an operator of o-preimages
for classes H and o of morphisms, respectively, and T is an operator of some
special kind of reduced products (for some class F of lters). We do not
present here the most general result from [AN82], but one which is relatively
close to it. And we shall present it in a category theoretical form, since in
this way it can relatively easily be applied to nice subclasses of Alg(S) like
E-, ECE- or QE-varieties, for example. Finally we prove an even weaker
version. One should observe that almost all the classes of objects appearing
in PART II are QE-varieties.
In 8.4 and also somehow in 11.6 we have tried to motivate, how one
can describe axioms which are preserved by such operators H, o and T,
respectively. We do not claim that those axioms are the only ones which
do the job, but the Main Theorem in subsection 12.2 will show that under
certain conditions these axioms are sucient. Thus we ask the reader to
recall the concepts of H-projective objects, of Pj H from 8.4.3, and also of
the operator
op
from 8.4.2 or from section 10. The following denition now
presents an encoding of formulas which, as we shall show to some extent
in 12.2, are then preserved by the operators they have been connected with
(see e.g. Table 8.1).
Denition 12.1.2 (i) Let C be any category, let H and o be classes of
C-morphisms, let k be any innite regular cardinal or (i.e. no
cardinal bound at all), and let I be any set:
A family (D, ( f
i
[ i I )) (f
i
: D Cod(f
i
), i I) is called a cone; let
Cone designate the class of all set indexed cones in C; and in particular,
let us dene:
December 8, 2002
12 A Meta Birkho Type Theorem 313
gure uncompleted
Figure 12.1: f12.1 (see 12.1.2.(i))
Cone
H
:= (D, ( f
i
[ i I )) [ D Pj H, (H-cones);
Cone
S
:= (D, ( f
i
[ i I )) [ f
i

op
(o) for all i I , (o-cones);
Cone
P
r,k := (D, ( f
i
[ i I )) [ #I = 1, D is strongly k-small , (T
r,k
-cones);
Cone
P
r,k
+
:= (D, ( f
i
[ i I )) [ #I 1, D is strongly k-small , (T
r,k
+
-cones);
Cone
e
:= (D, ( f
i
[ i I )) [ I ,= , (e-cones);
Cone
Pu
:= (D, ( f
i
: D D
i
[ i I )) [ D and D
i
are strongly
0
-small, I nite ,
(T
u
-cones).
If we designate strongly k-small objects in diagrams by
k
and o-
morphisms by
S
, we can sketch the situation in Figure 12.1. (If
nothing is said about I, then the set I may be empty; for I = the
cone starting with D is written as (D, ).)
(ii) Let H, o, T, T
r,k
, T
r,k
+
, e, T
u
, and let X be the string formed from
the operator symbols in (usually written in the order HoTT
r,k
T
r,k
+
eT
u
),
then:
Cone
X
:= Cone
Q
[ Q .
(iii) Let c := (D, ( f
i
[ i I )) Cone, and A Ob C. We say that c holds
(is valid) in A and that A is injective for c (in symbols: A [= c), iff for
every C-morphism q : D A there exist some i I and f : Cod(f
i
)
A such that q = f f
i
(see Figure 12.2); (D, ) holds in A iff there is no
morphism from D into A.
If K Ob C and c Cone Cone, then we dene
K [= c iff A [= c for all A K;
A [= Cone iff A [= c for all c Cone;
K [= Cone iff A [= Cone for all A K;
and
Inj Cone := A Ob C [ A [= Cone ;
Cone
X
K := c Cone
X
[ K [= c .
December 8, 2002
314 Chapter III Birkho type results for elementary implications
gure uncompleted
Figure 12.2: f12.2
Remark 12.1.3 Inj and Cone
X
are for each string X corresponding
to a subset of H, o, T, T
r,k
, T
r,k
+
, e, T
u
for any given innite regular car-
dinal k (including ) the operators of a Galois correspondence which
denes a language (Cone
X
, Ob C, [=) in the sense of abstract model theory
(see [NSa82]). In the case that we consider the category Alg(S) of par-
tial algebras, each cone, say (D, ( d
i
[ i I )), represents an implication
(M :
_

t
l
e
= t

l
[ l L
_

t
ij
e
=

t

ij
[ j J
i
[ i I ),
where in the case that I is empty we replace the conclusion by a statement
which is always false, and the whole formula then becomes equivalent to
(M : (
_

t
l
e
= t

l
[ l L)). In the proof and the applications of the main
theorem we shall restrict our considerations to the cases where I is an at
most one-element set, but the general theory goes further.
Observe, too, that for k = we get T
r,
= T, T
r,
+
= T
+
.
12.2 A Meta Birkho Theorem
Denition 12.2.1 Let C be any category and H Mor C. We recall from 8.4.3
that a C-object A is called H-projective, iff for every H-morphism, say
h : B C, and for every f : A C there is a g : A B such that
h g = f. When Pj H designates the class of all H-projective C-objects, then
we say that C has enough H-projectives, iff for every C Ob C there are
P Pj H and h : P C in H, i.e. every C-object is the H-image of some
H-projective object.
12.2.2 Observe that all partial Peano algebras are H
c
-projective for the
class H
c
of all closed and surjective homomorphisms, and that each partial
algebra is a closed and surjective homomorphic image of some suitable partial
Peano algebra.
The main theorem is hidden in the following
Lemma 12.2.3 Let k be an innite regular cardinal or (i.e. there may be
no bound on the cardinalities), and let C be a strongly k-algebroidal category
December 8, 2002
12 A Meta Birkho Type Theorem 315
(see 11.2.4) in which direct products and direct limits of k-directed systems ex-
ist. Moreover, let H and o be classes of C-morphisms, for which we consider
the following properties:
(1) C has enough H-projectives (see 12.2.1).
(2) Every H-projective P Pj H is the k-direct limit of strongly k-small
H-projective objects.
(3) (
op
(o), o) is a factorization system such that
op
(o) Epi C.
(4) If g f H and f
op
(o), then g H, i.e. H is right cancellable
with respect to
op
(o)-morphisms.
(5) C is
op
(o)-co-(well-powered) (i.e. from each C-object there start up to
isomorphism only a set of
op
(o)-objects).
(6) o = (
op
(o)
k
k
), where
op
(o)
k
k
:= e
op
(o) [ Dom(e) and Cod(e) are strongly k-small .
Let K Ob C be any class of C-objects, and recall that HK designates the
class of all H-images of K-objects, while o K designates the class of all o-
preimages of K-objects. Then we have the following statements:
(i) Without any assumptions on H and o we have for any string Z of
operators from H, o, T
r,k
+
, T
r,k
, e, T
u

Z(K) Inj Cone


Z
K.
(ii) (1), (3), (4) and (5) imply
Inj Cone
HS
K = Ho K, and
Inj Cone
HSe
K = Hoe K.
(iii) (1) through (7) imply
Inj Cone
HSP
r,k
+
K = HoT
r,k
+
K, and
Inj Cone
HSP
r,k K = HoT
r,k
K.
(iv) If k = , then (1), (3), (4) and (5) imply
Inj Cone
HSP
+
K = HoT
+
K, and
Inj Cone
HSP
K = HoT K.
December 8, 2002
316 Chapter III Birkho type results for elementary implications
(v) (1) through (7) imply
Inj Cone
HSPu
K = HoT
u
K, and
Inj Cone
HSPue
K = HoT
u
e K.
12.2.4 The rest of this subsection will be dedicated to the proof of this
Lemma, or at least to the proofs of (i), (iii) and (iv) in it, and also in
connection with (i) we restrict our considerations to the operators HoT,
where T is either T
r,k
+
or T
r,k
. The statements in (ii) and (v) will not be used
later, and therefore we refer the interested reader to their proofs in [NSa82]
and [AN82]. But rst we collect the statements (by slightly weakening them)
in the following
Corollary 12.2.5 (Meta Birkho Theorem) With the notation and all
assumptions from 12.2.3 one has for any operator T i, T
r,k
, T
r,k
+
, e, T
u
, eT
u

(cf. 4.6.2),
Inj Cone
HSF
K = HoT K.
12.2.6 Before we go into the details of the proof of 12.2.3, we already briey
comment on them:
(i) Part (i) of the proof (see 12.2.7) will be closely connected with the
proof of Theorem 8.3.4 (see Table 8.1 and also subsection 11.6).
(ii) Let us turn to the proof of the other inclusion in 12.2.3.(iii) which
will be the main part of the following proof, since we do not prove all
the stated facts (compare Figure 12.3): While in the proof of the
original Birkho Theorem (for equations) one works with the Peano
algebra freely generated by the carrier set of the algebra A under con-
sideration i.e. which satises all specied formulas valid in K , one
has to replace the Peano algebra in the general case by an H-projective
object, say Q, together with an H-morphism from Q to the object A
(H takes the place of the class of all quotient mappings). Inductiv-
ity of the Peano algebra is replaced by the assumption that the H-
projective object Q is the direct limit of a k-directed system consist-
ing of strongly k-small H-projective objects Q
j
in order to get the
premises of the implications to be considered down to the desired size.
December 8, 2002
12 A Meta Birkho Type Theorem 317
For each such premise the possible conclusions are collected from
the (
op
(o), o)-factorizations of morphisms from the Q
j
into K. Since
we want to form products of the possible conclusions and necessary
K-objects, we need the
op
(o)-co-(well-poweredness). With the sets of
necessary conclusions respectively necessary K-objects two directed
systems are formed, the latter yielding a k-reduced product, say K, of
K-objects, the other one yielding a direct limit, say R, which is an o-
subobject of K and allows an H-morphism into A. Verication of this
last statement is especially involved and needs an extra Lemma 12.2.14,
where we need that
op
(o)-morphisms starting with strongly k-small
objects can be approximated as direct limits by k-directed systems of

op
(o)-morphisms between strongly k-small objects i.e. admissible
implications (needing strong k-algebroidalness of C and k-dual alge-
broidalness of
op
(o)).
12.2.7 (Proof of 12.2.3.(i)) Since we restrict the proof to operators HoT
with T T
r,k
+
, T
r,k
, the corresponding cones c Cone
HSF
have the form
c = (D, d : D D

) or c = (D

, ) (see 12.1.2.(i), the latter is only admissible,


if T = T
r,k
+
), where D and D

are strongly k-small H-projective objects and


d
op
(o). Now, assume K Ob C, A HoT K and c Cone
HSF
. Thus
there are an H-morphism h : B A, and an o-morphism s : B

k
( A
i
[
i I )/f, where, for some set I, A
i
[ i I K, and f is a k-complete
lter on I (observe that -completeness of f means that f is a principal lter
on I, and T
r,
= T, T
r,
+
= T
+
); the directed system connected with the
above reduced product is denoted as in 4.5 by ( A
F
, p
FG
[ F G f ), where
A
F
:=

( A
i
[ i F ) and p
FG
is the projection from A
F
onto A
G
.
(a) f contains the empty set: Then

k
( A
i
[ i I )/f =: T is just the
terminal element in C, i.e. each C-object allows exactly one morphism
into T. We then have only cones of the form c = (D, d : D D

).
Assume f : D A; since D is H-projective and h H, there is
f

: D B such that h f

= f. Because of the nature of T there


is a (unique) q : D

T such that q d = s f

. Since s o and
d
op
(o), there is a unique d

: D

B such that d

d = f

(and
s d

= q). Thus h d

d = h f

= f, showing that A [= c, since f was


arbitrarily chosen; hence A Inj Cone
HSF
K.
(b) f is a proper lter, I ,= : Let either c = (D, d : D D

) or c

= (D, )
be in Cone
HSF
K. In either case assume f : D A, and as in (a) we
December 8, 2002
318 Chapter III Birkho type results for elementary implications
get f

: D B satisfying h f

= f. Since D is strongly k-small and


since f is k-complete, there are G f and f
G
: D A
G
such that
p
G
f
G
= s f

. If we have started with c

, then p
Gi
f
G
: D A
i
for some i G contradicts A
i
K and c

Cone
HSF
K. Thus in
this case there cannot be any morphism from D to A, i.e. A [= c

.
If we consider c, then c holds in every A
i
(i G), and therefore
also in A
G
, and we get g : D

A
G
such that g d = f
G
. Now,
s f

= p
G
f
G
= p
G
g d, s o and d
op
(o). Thus there exists
d

: D

B such that d

d = f

(and s d

= p
G
g). As in (a) this
completes the proof.
12.2.8 (Sketch of the proof of 12.2.3.(iii) and (iv)) From what we
have said earlier, (iv) is a special case of (iii) for k = . Assume that
we have already proved HoT
r,k
+
K = Inj Cone
HSP
r,k
+
K. Then HoT
r,k
K =
HoT
r,k
+
(K T) = Inj Cone
HSP
r,k
+
(K T), where T designates the terminal
object of C, representable as a product with empty index set. Since all
cones (D, d : D D

) hold in T, while no cone (D, ) does, we have


Cone
HSP
r,k
+
(K T) = Cone
HSP
r,k K; therefore (iii) (and (iv)) follow in the
case of T = T
r,k
, too.
Thus because of (i) it remains only to show Inj Cone
HSP
r,k
+
K HoT
r,k
+
K.
This proof follows to a great extent the one in [ABN80] for the statement
given here in 8.3.5.(iv), and which is based on a proof of A.I.Malcev
in [Mal73] for quasivarieties (compare also the proof of the First Birkho
Theorem 6.2.5) in [AN82] a proof for a more general result can be found.
We have to show that A Inj Cone
HSP
r,k
+
K implies A HoT
r,k
+
K:
Thus assume A Inj Cone
HSP
r,k
+
K. Since we have assumed in (1) that
C has enough H-projectives, there exist Q Pj H and an H-morphism h :
Q A. Because of assumption (2) we may choose a k-directed system,
say Q
J
:= (Q
j
, q
jj
[ j j

J ) with direct limit (( q


j
[ j J ), Q)
(i.e. representing Q) such that each Q
j
is a strongly k-small H-projective
object; and in order to make some notation simpler, we assume that Q
j
,= Q
j
,
whenever j ,= j

in J.
Then the main line of the rest of the proof is as follows:
(a) For each j J we choose a set K
j
K and a set c
j

op
(o)
such that each c
j
-morphism starts at Q
j
and in some way c
j
and K
j
represent all morphisms from Q
j
into some K-object with respect to

op
(o):
December 8, 2002
12 A Meta Birkho Type Theorem 319
If K K, f : Q
j
K and if (e, s) is a (
op
(o), o)-factorization
of f, then there are K

K
j
and f

: Q
j
K

such that f

has a
(
op
(o), o)-factorization (e

, s

) satisfying
(a1) e

E
j
, and
(a2) There is an isomorphism i : Cod(e

) Cod(e) such that i e

= e.
Claim (a) The choices in (a) above are possible, and c
j
,= and K
j
,= ,
in particular there exists a morphism from Q
j
into some K-object for
each j J. (See 12.2.10.)
(b) Now, we dene the set K
A
:=

K
j
[ j J which is rele-
vant for the morphisms starting from A into some K-object (in the
equational case this was the set of all representatives of subalgebras
of K-algebras which were generated by homomorphic images of A,
cf. Rsuba
A
K in 5.11). Dene I :=

Mor (Q
j
, K) [ j J, and K K
A
,
which is also a set. For each j J we dene I
j
:=

Mor (Q
j
, K) [
j j

J, K K
A
.
Claim (b1) I
j
[ j J is the base of a k-complete proper lter
f := L I [ there is j J with I
j
L on I. (See 12.2.11.)
For what follows compare Figure 12.3 on page 321: For each j J dene
K
j
:=

( Cod(f) [ f I
j
), and let f
j
: Q
j
K
j
be the (product) morphism
induced by the source ( f q
jj
[ j j

J, f Mor (Q
j
, K), K K
A
).
Moreover, let
D
j
:=

( Cod(e
f
) [ f Mor (Q
j
, K), j j

J, K K
A
, e
f
c
j
,
(e
f
, s
f
) is a xed (
op
(o), o)-factorization of f q
jj
).
Let e

j
: Q
j
D
j
be the (product) morphism induced by the source ( e
f
[
f I
j
), and let s

j
:=

( s
f
[ f I
j
) : D
j
K
j
be the product in the
sense of 4.1.3.(d) of the family of morphisms s
f
o occurring in the denition
of D
j
. Since o =
op
(o) is closed with respect to products (see 10.1.2.(ix)),
we get s

j
o (j J).
Let (e
j
, s

j
) be a (
op
(o), o)-factorization of e

j
, such that R
j
:= Cod(e
j
) =
Dom(s

j
). Then s
j
:= s

j
s

j
is an o-morphism, and (e
j
, s
j
) is a (
op
(o), o)-
factorization of f
j
. Since I
l
I
j
for j l J, we have a projection
morphism (or restriction morphism) k
jl
: K
j
K
l
induced by the family
( p
K
j
f
: K
j
Cod(f) [ f I
l
) of projections.
December 8, 2002
320 Chapter III Birkho type results for elementary implications
Claim (b2) k
jj
= 1
K
j
and k
lm
k
jl
= k
jm
for j l m J,
i.e. K
J
:= ( K
j
, k
jl
[ j l J ) is a k-directed system conal in
K
(f;)
:= ( K
F
, k
FG
[ F G f) let (( k
j
[ j J ), K) be its
direct limit, which is obviously isomorphic to the reduced product

( Cod(f) [ f I )/f (cf. 11.1.4.(ii)), i.e. K T


r,k
+
K. (See 12.2.12.)
Claim (b3) k
jl
f
j
= f
l
q
jl
for every j l J, i.e. f

:= (id
J
, ( f
j
[ j
J )) is a conal morphism between the k-directed systems Q
J
and K
(f,)
let

f : Q K be the induced colimit morphism. (See 12.2.13.)
(c) Claim (b3) implies for j l J, when we decompose f
j
and f
l
:
(k
jl
s
j
) e
j
= s
l
(e
l
q
jl
). Observing that e
j

op
(o) and s
l
o, we
get from Lemma 11.3.5 for j l J a unique morphism r
jl
: R
j
R
l
(recall R
j
= Cod(e
l
)) such that r
jl
e
j
= e
l
q
jl
and s
l
r
jl
= k
jl
s
j
. And
in particular ( R
j
, r
jl
[ j l J ) =: R
J
is a k-directed system with
colimit, say, (( r
j
[ j J ), R) and e

:= (id
J
, ( e
j
[ j J )) : Q
J
R
J
as well as s

:= (id
J
, ( s
j
[ j J )) : R
J
K
J
are special morphisms
between k-directed systems with induced morphisms, say, e : Q R
and s : R K, respectively. Obviously s e =

f.
Claim (c1) e
op
(o) and s o, in particular R oT
r,k
+
K. (See 12.2.15.)
(d) Because of Claim (c1) it remains to prove that there exists an H-
morphism from R to A. Since for each j J the morphism e
j
belongs
to
op
(o) and starts at a strongly k-small H-projective object; thus
(Q
j
, e
j
) is an HoT
r,k
+
-cone.
Claim (d1) For each j J we have e
j
Cone
HSP
r,k
+
K. (See 12.2.16.)
Thus, because of (3), e
j
is a K-extendable epimorphism. Since A
Inj Cone
HSP
r,k
+
K, we get for each j J a unique g
j
: R
j
A satisfying
h q
j
= g
j
e
j
.
Claim (d2) ( g
j
[ j J ) is compatible with R
J
, i.e. for each j l J
we have g
l
r
jl
= g
j
. (See 12.2.17.)
Hence we obtain a unique morphism g : R A such that g r
j
= g
j
for each j J.
Claim (d3) g e = h. (See 12.2.18.)
December 8, 2002
12 A Meta Birkho Type Theorem 321
gure uncompleted
Figure 12.3: f12.3
Since e
op
(o) and h H, assumption (4) together with Claim (d3)
yields g H, nally showing that A HoT
r,k
+
K.
This completes the sketch of the proof of 12.2.3.(iii) and (iv) observe
that, for k = , J above is a one-element set.
12.2.9 We now prove the claims stated in 12.2.8. At some instances we shall
refer to the fact that we have proved results before for the category Alg(S),
which are actually true in general for arbitrary categories or at least for
categories satisfying the assumptions of 12.2.3. We shall not stress the cor-
responding references more explicitly.
12.2.10 (Proof of Claim (a)) Since by 12.2.3.(5) C is
op
(o)-co-(well-
powered), there exists for each j J a set c

j

op
(o) such that each

op
(o)-morphism starting from Q
j
is represented up to isomorphism in c

j
.
Let
c
j
:= e c

j
[ there are s o and K K such that s e : Q
j
K,
i.e. c
j
consists of all c

j
-morphisms which are the rst factor of a (
op
(o), o)-
factorization of a morphism from Q
j
into some K-object. For each e c
j
choose some K
e
K such that there is s o satisfying s e : Q
j
K
e
. Dene
K
j
:= K
e
[ e c
j
. Obviously K
j
,= , if c
j
,= , and the latter is true,
if there is at least one morphism from Q
j
into some K-object: Assume that
there is no such morphism; then (Q
j
, ) Cone
HSP
r,k
+
K, but this contradicts
the facts that h q
j
: Q
j
A and A Inj Cone
HSP
r,k
+
K. Thus we have for
each j J at least one morphism from Q
j
into some K-object, c
j
as well as
K
j
are non-empty, and they are sets by construction. In particular, if there
is some A Inj Cone
HSP
r,k
+
K, then K has to be non-empty.
12.2.11 (Proof of Claim (b1)) Consider J

J, #J

< k, then k-
directedness of J implies the existence of an upper bound of J

, say j. Thus
I
j
I
j
(j

). Since I
j
,= by Claim (a), this easily implies that f is a
proper k-complete lter.
December 8, 2002
322 Chapter III Birkho type results for elementary implications
gure uncompleted
Figure 12.4: f12.4
12.2.12 (Proof of Claim (b2)) This is a simple consequence of the def-
inition of k
jl
as a restriction morphism between products and the fact that
for j l m j

and f Mor (Q
j
, K

) for some K

K
A
we have
p
Km
f
k
lm
k
jl
= p
K
j
f
k
jl
= p
K
j
f
= p
Km
f
k
jm
.
12.2.13 (Proof of Claim (b3)) Again by the denition of K
j
and f
j
in 12.2.8.(b), we have, for j l J and f I
l
, i.e. for f Mor (Q
j
, K

) for
some j

l and K

K
A
, that p
K
l
f
f
l
q
jl
= f q
lj
q
jl
= f q
jj
= p
K
j
f
f
j
=
p
K
l
f
k
jl
f
j
, and the statement follows.
For the proof of Claim (c1) we need an auxiliary lemma (cf. [NSa82],
Lemma 11(b)):
Lemma 12.2.14 If C is a strongly k-algebroidal category k an innite reg-
ular cardinal , o a class of C-morphisms satisfying conditions (3), (5), (6)
and (7) from 12.2.3, then o = (
op
(o)
k
k
), where
op
(o)
k
k
:= e
op
(o) [
Dom(e) and Cod(e) are strongly k-small .
Proof Since (
op
(o), o) is a factorization system, 10.2.5 implies o =
op
(o).
Therefore it is sucient (cf. 10.1.2.(iv)) to prove
op
(o)
k
k
)
op
(o):
Thus, assume that s (
op
(o)
k
k
), let h
op
(o), h : A B, s : C D,
and let p, q Mor C such that q h = s p. We have to show the existence of
a unique diagonal-ll-in d : B C, d h = p, s d = q.
Since C is strongly k-algebroidal, A is the direct limit, say (( h
i
[ i I ), A) of
a k-directed system, say A
I
:= (A
i
, h
ij
[ i j I ). For each i I let (t
i
, r
i
)
t
i
: A
i
E
i
in
op
(o), r
i
: E
i
B in o be a (
op
(o), o)-factorization of
h h
i
: r
i
t
i
= h h
i
= r
j
(t
j
h
ij
). t
i

op
(o) and r
j
o imply the existence
of a unique diagonal-ll-in e
ij
: E
i
E
j
, and the uniqueness of diagonal-ll-
ins (or the fact that t
i
is an epimorphism) easily implies e
jl
e
ij
= e
il
for each
i j l I (cf. 11.3.5). Hence E
I
:= (E
i
, e
ij
[ i j I ) is a k-directed
system with colimiting cocone, say (( e
i
[ i I ), E).
The denition of the e
ij
immediately implies that t

:= (id
I
, ( t
i
[ i I )) :
A
I
E
I
and r

:= (id
I
, ( r
i
[ i I )) : E
I
(B, 1
B
[ i j I ) =: B
I
December 8, 2002
12 A Meta Birkho Type Theorem 323
are morphisms between k-directed systems. Let

t : A E and r : E B
be the induced colimit morphisms (consider the commutative diagram in
Figure 12.4). Obviously r

t = h; moreover, r o, since o is k-inductive,


and

t
op
(o) by 11.1.4.(iii). In addition, since h = r

t is a
op
(o)-
morphism, too, the uniqueness of such factorizations implies that r is an
isomorphism. It is therefore now sucient to construct a diagonal-ll-in,
say d, starting from E instead of B (since then p = d

t = (d r
1
) h and
q = q r r
1
= s (d r
1]
).
In order to obtain d, we use that
op
(o) is k-dual algebroidal:
For each t
i
: A
i
E
i
we have a k-directed system E
(i)
I
i
:= (E
(i)
k
, t
(i)
kl
[ k l
I
i
) with colimiting cocone ((t
(i)
k
[ k I
i
), E
i
) such that each E
(i)
k
is strongly
k-small, t
i
= t
(i)
o
for some o I
i
, and t
(i)
kl

op
(o) for any k l I
i
and
i I. Observe that, for o k I
i
: (q r
i
t
(i)
k
) t
(i)
ok
= q h h
i
= (s p) h
i
and s
op
(o)
k
k
). Thus there exists for each i I and each k I
i
a
d
(i)
k
: E
(i)
k
C such that d
(i)
k
t
(i)
ok
= p h
i
and s d
(i)
k
= q r
i
t
(i)
k
= q r e
i
t
(i)
k
.
Since each t
(i)
ok
is an epimorphism, it is easy to realize that d
(i)
l
t
(i)
kl
= d
(i)
k
for k l I
i
, i I. Therefore there exists a unique colimit morphism
d
i
: E
i
C such that d
i
t
(i)
k
= d
(i)
k
(k I
i
), and in particular d
i
t
i
=
d
i
t
(i)
o
= d
i
t
(i)
k
t
(i)
ok
= d
(i)
k
t
(i)
ok
= p h
i
.
Now, for i j I the facts that t
i
is an epimorphism and d
j
e
ij
t
i
=
d
j
t
j
h
ij
= p h
j
h
ij
= p h
i
= d
i
t
i
imply that d
j
e
ij
= d
i
, i.e. the family
( d
i
[ i I ) is compatible with A
I
and therefore induces a morphism d : E
C satisfying d e
i
= d
i
(i I). The uniqueness of induced colimit morphisms
and the facts that ( d
i
t
i
: A
i
C [ i I ) is a family compatible with A
I
and that d
i
t
i
= p h
i
= d e
i
t
i
= d

t h
i
implies d

t = p. Since

t is an
epimorphism, s d

t = s p = q h = q r

t implies s d = q r, i.e. we have


got the morphism d which we have looked for.
12.2.15 (Proof of Claim (c1)) The fact that e
op
(o) follows from 11.1.4.(iii),
while s o is now an immediate consequence of Lemma 12.2.14 and Propo-
sition 11.3.1.(i). From Claim (b2) then R oT
r,k
+
K follows.
12.2.16 (Proof of Claim (d1)) Assume g : Q
j
K for some K
K. Then by the assumptions on c
j
and K
j
there exists a (
op
(o), o)-
factorization, say (e
g
, s
g
), of g such that e
g
c
j
, and there are K


K
j
and f : Q
j
K

such that f has a (


op
(o), o)-factorization (e
g
, s

)
December 8, 2002
324 Chapter III Birkho type results for elementary implications
(see (a)); in particular e
g
= e
f
and f I
j
. Thus we have (see Figure 12.3))
g = s
g
e
g
= s
g
e
f
= s
g
p
D
j
f
s

j
e
j
, showing that g factors through e
j
,
i.e. e
j
Cone
HSP
r,k
+
K.
12.2.17 (Proof of Claim (d2)) Since each e
j
is an epimorphism, this
follows from (g
l
r
jl
) e
j
r = g
l
e
l
q
jl
= h q
l
q
jl
= h q
j
= g
j
e
j
(see Fig-
ure 12.3).
12.2.18 (Proof of Claim (d3)) Since g
j
e
j
= h q
j
for each j J, the
uniqueness of induced colimit morphisms implies, as often before, that indeed
g e = h.
12.2.19 (Conclusion of the proof of 12.2.3(iii) and (iv)) Now all claims
of 12.2.8 have been proved, and we get Inj Cone
HSP
r,k
+
K HoT
r,k
+
K, which is
what we had to show.
As already mentioned the missing parts of the proof of 12.2.3 can be
found in [NSa82] and [AN82] (in the latter paper the statement is even more
general than in 12.2.3) or the reader may construct proofs by himself along
the line of the above one.
Remark 12.2.20 We have just presented the approach of Andr eka, N emeti
and Sain to handle the abundance of axiomatizability results for Horn for-
mulas. One should observe that R.John in [J77] for the general case and
in [J75] for special kinds of ECE-equations has got similar results by com-
bining the operators o and T to one, called /, and /is based on a class M
of special cocones (sinks). His approach appears perhaps to be more ele-
gant, but we have the feeling that the one in part presented here follows more
closely the lines of existing proofs (cf. the proof of Birkhoff in [Bi35] for
equational classes of universal algebras or the one of Malcev in [Mal73]
for quasivarieties), and the special axiomatizability results can be read more
easily from it. For further historical remarks see the paper [NSa82].
12.3 Compatible classes H and o in Alg(S)
Denition and Remark 12.3.1 The rest of this section as well as the
next section are now dedicated to the investigation of examples of the Meta
Birkho Theorem in the category Alg(S). In particular we shall study fam-
ilies of classes H and o of homomorphisms satisfying the conditions (1)
December 8, 2002
12 A Meta Birkho Type Theorem 325
through (7) of Lemma 12.2.3. That Alg(S) is strongly k-algebroidal for each
innite regular cardinal k d

has already been shown in 11.2.5. More-


over, products as well as direct limits of (k-) directed systems exist in Alg(S).
Thus Alg(S) satises the general assumptions about C of 12.2.3. We also
know already (see Table 10.1) that Alg(S) has in general innite families of
factorization systems.
In this subsection we want to investigate possible combinations of classes
H and o of homomorphisms (respectively of the corresponding operators)
satisfying the compatibility-condition 12.2.3.(4), and we shall call such
pairs of classes compatible. Observe that 12.2.3.(4) is the only condition
in 12.2.3 which connects H and o.
As far as H is concerned, we shall always assume that it be a class of
surjective homomorphisms because of condition 12.2.3.(1) that there should
always be enough H-projectives , although surjectivity is not necessary for
the existence of H-projectives in general: with respect to the class T
i
of all
nal homomorphisms (see 9.2.1 and 9.2.16.(iii)) it is not too dicult to show
by using 9.2.16.(iii).(c) the following:
()
A partial Peano algebra P on some set X is T
i
-projective iff
T = X B
1
X
P
such that for each x X there are exactly
one , exactly one a dom
P
and exactly one k ()
such that x = a(k).
(i.e. the domains of the operations
P
consist of disjoint injective sequences
also disjoint for dierent operations , which cover X). Thus such a P can
only be discrete when it is empty, and therefore no non-empty discrete par-
tial algebra can be the nal homomorphic image of an T
i
-projective partial
algebra.
Lemma 12.3.2 For any strongly k-algebroidal category C and any class o
of C-morphisms satisfying
op
(o) Epi C the class H := Iso C of all isomor-
phisms in C is compatible with o and also satises 12.2.3.(1) and (2).
Proof It is obvious that every C-object is Iso C-projective, and strong k-al-
gebroidalness of C implies 12.2.3.(2). Let h = g f Iso C, f
op
(o)
Epi C, then we have h
1
g f = 1, and therefore f is a section as well as an
epimorphism, i.e. an isomorphism (see 11.4.3). Thus g = h f
1
Iso C.
Corollary 12.3.3 Let the assumptions of Lemma 12.2.3 be satised for the
category C, for H, o Mor C, and for some innite regular cardinal k (
means that k =
0
). [2] Then one has for every class K Ob C:
December 8, 2002
326 Chapter III Birkho type results for elementary implications
(i) Inj Cone
S
K = o K;
(ii) Inj Cone
Se
K = oe K;
(iii) Inj Cone
SPu
K = oT
u
K;
(iv) Inj Cone
SPue
K = oT
u
e K;
(v) Inj Cone
SP
r,k
+
K = oT
r,k
+
K;
(vi) Inj Cone
SP
r,k K = oT
r,k
K;
(vii) Inj Cone
HS
K = Ho K;
(viii) Inj Cone
HSe
K = Hoe K;
(ix) Inj Cone
HSPu
K = HoT
u
K;
(x) Inj Cone
HSPue
K = HoT
u
e K;
(xi) Inj Cone
HSP
r,k
+
K = HoT
r,k
+
K;
(xii) Inj Cone
HSP
r,k K = HoT
r,k
K;
(xiii) Inj Cone
SP
+
K = oT
+
K;
(xiv) Inj Cone
SP
K = oT K;
(xv) Inj Cone
HSP
+
K = HoT
+
K;
(xvi) Inj Cone
HSP
K = HoT K.
Proof In the cases (i) through (vi), (xiii) and (xiv) we have chosen H =
Iso C (see 12.3.2), in the cases (iii), (iv), (ix) and (x) the assumptions of 12.2.3
have to be satised for k =
0
, and in the cases (xiii) to (xvi) one has to
choose k = ; but then, by assumption, the conditions of 12.2.3 are satised.
12.3.4 In 12.3.3 we have listed separately all possibilities for the operator T
in 12.2.5, and we have distinguished between arbitrary admissible opera-
tors H satisfying 12.2.3.(i) and (ii), and for which we still have to investigate
compatibility with admissible operators o and the case H = 1 (isomor-
phisms), where every admissible o i.e. such o satisfying 12.2.3.(3), (5),
(6) and (7) can be chosen.
Lemma 12.3.5 Let o HomAlg(S), and let H =: H
F
H
w
be the class of
all surjective homomorphisms reecting some given set F of formulas, such
that F is preserved by
op
(o), and assume that g f H
F
and f
op
(o)
together imply surjectivity of f. Then H
F
is compatible with o, i.e. H
F
is
right cancellable with respect to
op
(o).
Proof Since surjectivity of g f implies surjectivity of g, this is a conse-
quence of 9.1.7.(ii).
December 8, 2002
12 A Meta Birkho Type Theorem 327
Since by 10.3.7 and 10.1.2.(iv)
op
(o) H
w
iff o contains the class
of all full and injective homomorphisms, and since for closed h = g f, where
f is an epimorphism, f is closed and surjective, we get:
Corollary 12.3.6 Let o HomAlg(S), and let H H
w
be the class of all
surjective homomorphisms reecting some given set F of E-equations.
(i) Then H is right cancellable with respect to
op
(o), whenever o =

op
(o) contains the class o
i
of all full and injective homomorphisms.
(ii) If, in addition, each h in H is closed, then H is left cancellable with
respect to
op
(o), whenever
op
(o) Epi Alg(S).
Proposition 12.3.7 If H is the class either of all full and surjective or of
all surjective homomorphisms in Alg(S), then H is right cancellable for any
class o of homomorphisms.
Proof If g f is surjective, then so is g; and if, in addition, g f is full, then
g is full see 2.4.5.(iii).(d).
Combining 12.3.7 with 12.3.6 (or with 12.3.5) we get the
Corollary 12.3.8 If o HomAlg(S), and if H is the class of all full and
surjective homomorphisms reecting some given set F of E-equations, then
H is compatible with o, whenever
op
(o) contains the class of all full and
injective homomorphisms.
Lemma 12.3.9 Let H be the class of all surjective respectively of all full
and surjective homomorphisms reecting some set F of E-equations; and
let o be the class (a): of all, (b): of all injective, (c): of all initial, respec-
tively (d): of all injective and initial homomorphisms in each case reecting
some set F

of E-equations. Then F

F implies that H and o are compat-


ible.
Proof We already know that under the conditions of the Lemma g f H
implies that g is (full and) surjective. Thus we only have to show that g
reects F: Lemma 10.3.8 tells us that for o = / ^ with / being the
class of all F

-reecting homomorphisms and ^ being one of the classes


in 12.3.9.(a) through (d), and therefore satisfying
op
(^) H
w
, one has

op
(o) =
op
(/) (
op
(^) /). Therefore f can be represented as f =
December 8, 2002
328 Chapter III Birkho type results for elementary implications
H o o
i
o
w
/
i
1 o
c
/
c
/
P
o
P
/
iP
o
iP
/
n
/
P

i
o
P

i
1 2 2 2 2 2 2 2 2 2 2 2 2 2
H
c
6 6 6 6 6 6 6 6 6 6 6 6 6
H
f
8 8 8 7 7 7 7 7 7 7 7 7 7
H
w
6 6 6 7 7 7 7 7 7 7 7 7 7
H
P
6 6 6 9* 9* 9* 9*
H
fP
8 8 8 9* 9* 9* 9*
H
i
6 6 6
H
Pi
6 6 6 9* 9*
H
b
6 6 6
Table 12.1: Compatibility of the operators H and o
(see 12.3.10)
v u, where u
op
(/) and v
op
(^) / H
w
. Since g v preserves
all E-equations, and since g v u reects F ( F

), 9.1.7.(ii) implies that u


reects F and therefore F

. Thus u /
op
(/) Iso Alg(S). Hence
f = v u is surjective and preserves all E-equations, and 9.1.7.(iii) implies
that g reects F, showing that g H.
Corollary 12.3.10 For some special instances of the operators H and o the
information about compatibility derived from the results in 12.3.2 through 12.3.9
is collected in Table 12.1. No entry at some place means only that this case
has not yet been investigated, while an entry m ( m ) says that com-
patibility of the corresponding pair (H, o) follows from 12.3.m (under the
condition P

P, where P and P

are any sets of terms). As abbreviations


we use the notation corresponding to the one in 4.6.2 and in 10.3.12, and in
addition to the one in 10.3.12 we use Pi for P-initialness (for any set P
of terms).
Remark 12.3.11 The compatibilities listed in Table 12.1 do not mean that
we actually already have examples for Lemma 12.2.3, since all the other
conditions of this Lemma have not yet been checked; this will be done
in detail in the next section. We only know from Table 10.1 that for the
classes o listed in Table 12.1 (
op
(o), o) is a factorization system satisfying

op
(o) Epi Alg(S), and that, for k d

, o from Table 12.1 is k-inductive


December 8, 2002
12 A Meta Birkho Type Theorem 329
(see 11.3.1). Thus we have assumptions 12.2.3.(3) and (7). Moreover, co-
(well-poweredness) of Alg(S) with respect to epimorphisms (see the Epimor-
phism Theorem 5.8.3) and the fact that
op
(o) Epi Alg(S) imply that
Alg(S) is
op
(o)-co-(well- powered); therefore also assumption12.2.3.(5) is
satised in Table 12.1. Since for k d

all the basic assumptions about the


category C in 12.2.3 are satised in Alg(S), it remains only to investigate
conditions 12.2.3.(1), (2) and (6) for the pairings of Table 12.1. This will
be done in section 13.
It should be observed that still other combinations of properties of ho-
momorphisms could be included in Table 12.1, which therefore only presents
some arbitrary choice. But even now we have already 33 pairs with one and
10 pairs with two parameters which usually may run through innitely many
distinct values.
December 8, 2002
330 Chapter III Birkho type results for elementary implications
13 Construction of examples for the Meta
Birkho Theorem
In this section we continue the discussion of examples for the Meta Birkho
Theorem 12.2.5, which we have already begun in 12.3. Our starting point
is Table 12.1 of compatible operators; we shall deal only with the operators
already considered in this table, and we shall also use its abbreviations.
Because of Remark 12.3.11 it remains only to investigate conditions (1), (2)
and (6) of Lemma 12.2.3.
Actually we shall not discuss all the operators considered in Table 12.1,
but only those about which we already know a bit more (e.g.
op
(/
P

i
)
has not yet been investigated); the investigation of the others and possibly
further operators is left as an exercise or problem. For some other operators
we only state the results and leave the details of the proofs as exercises.
13.1 k-dual algebroidalness of
op
(S) for some classes o
of Alg(S)-morphisms
An immediate consequence of 11.2.2 is
Lemma 13.1.1 Let f : A B be a full and surjective homomorphism, and
let A be strongly k-small for some regular innite cardinal k d

. Then B is
strongly k-small.
Corollary 13.1.2 The classes H
f
(=
op
(o
w
)) and H
c
are k-dual algebroidal
(see 11.5.3) for every regular cardinal k d

.
Lemma 13.1.3 The classes c
w
= Epi Alg(S) (=
op
(o
c
)) and H
w
(=
op
(o
r
))
are k-dual algebroidal for each regular innite cardinal k d

.
Proof Let e : A B be any epimorphism, A
s
C
m
B its (H
f
, o
w
)-
factorization, where we may assume that C is a weak subalgebra of B. Then
B = c
B
C. Let A be strongly k-small, then so is C because of 13.1.1. From
4.4.7 and 11.2.2 we know that B is the k-direct limit (union!) of all weak
subalgebras D of B, which are strongly k-small and generated by some subset
of C (i.e. D I
3
(B, C)). But then B is also the k-direct limit of the system
of all D

C (see 3.2.2 for

), where D I
3
(B, C) with the inclusion
homomorphisms; and each D

C is strongly k-small. It is easy to check


December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 331
that for D D

I
3
(B, C) the inclusion mapping id
CD
: D

C D

C
as well as the homomorphism f
AD
:= f : A D

C are epimorphisms.
Thus, adding A as a least element to I
3
(B, C) and the f
AD
as correspond-
ing homomorphisms, we get f
A
= f : A B, and therefore the desired
approximation of f. The argument for H
w
is similar.
Lemma 13.1.4 Let
op
(o) Epi Alg(S) for some factorization system (
op
(o), o),
and let k d

be a regular cardinal. Assume that we have for e : A B


in
op
(o) with A strongly k-small:
Let C be the weak subalgebra of B with carrier set e[A] and the struc-
ture induced from A by e; and if D is any C-generated strongly k-small weak
subalgebra of B containing C, then let there be a strongly k-small weak sub-
algebra D

of B containing D such that e : A D

belongs to
op
(o).
Then
op
(o) is k-dual algebroidal.
Proof Let e : A B be in
op
(o). With the assumptions of 13.1.4 set
I
5
(B, C) := D [ D is a strongly k-small C-generated weak subalgebra of B
and e : A D is in
op
(o).

Then the assumptions of 13.1.4 say that I


5
(B, C) is conal in I
3
(B, C); thus
B is the direct limit of I
5
(B, C). Adding to I
5
(B, C) the inclusion mappings,
A and the epimorphisms from A into D I
5
(B, C) with the same graph as e,
then 10.1.2.(v)
op
implies that all morphisms of this k-directed system belong
to
op
(o) showing that
op
(o) is k-dual algebroidal.
In view of 10.3.8 and 10.3.9 the following Lemma will be useful:
Lemma 13.1.5 Let P be any set of TE-expressions, and let k d

be any
regular cardinal. Then
op
(o
P
) and
op
(o
iP
)are k-dual algebroidal.
Proof (Sketch) We only prove the statement for o
iP
, the proof of the
other one is similar. Combining the results in 10.3.5 and 10.3.9 we may say:
()
g : C D belongs to
op
(o
iP
) iff G := domg

and g

have
the following properties: If G
P
:= c
G,P
C, then each element
of G G
P
is identied by ker g

with some element of G


P
.
Now consider f : A B in
op
(o
iP
) = H
w

op
(/
P
) =: c; let F := domf

,
F
P
:= c
F,P
A. Let J be the set of all pairs (D, ) such that D is a strongly
December 8, 2002
332 Chapter III Birkho type results for elementary implications
k-small A-generated relative subalgebra of F, Cong D, ker f

,
# < k and each element of D D
P
is -equivalent with some element
of D
P
.
For elements of J dene (D, ) (D

) iff D is a relative subalgebra of D

and

. Then it is easy to realize that (J; ) is a k-directed poset.


If (D, ) (D

), then there are unique homomorphisms f

: D/
D

satisfying f

nat

= nat

[
D
, and f

: D B satisfying f

nat

=
f

[
D
. Moreover, D/ and D

are strongly k-small. From () we may in-


fer that each f

and each f

belong to c and that ( D/, f

[ (D, )
(D

) in J ) is a k-directed system with k-direct limit (( f

[ (D, )
J ), B), (A,
A
) J and f

A
= f. This shows that
op
(o
iP
) is k-dual alge-
broidal.
Combining these results and using the notation from 12.3.10 we get the
Proposition 13.1.6 Let k d

be a regular cardinal. The classes


op
(o)
are k-dual algebroidal, where o is one of the classes 1, o
c
, o
r
, o
w
, o
P
, o
iP
,
/
c
, /
i
, /
P
, /
n
.
Proof Since the treatment of I is trivial because of k-algebroidalness of Alg(S),
it remains to treat o
P
, o
iP
, /
c
, /
i
, /
P
, /
n
. The statements for /
c
and /
P
can easily be derived from 13.1.4 and the descriptions of their
left factors in the corresponding factorization system. Since
op
(o
P
) =

op
(o
w
)
op
(/
P
) and
op
(o
iP
) =
op
(o
i
)
op
(/
P
), their k-dual alge-
broidalness follows from 13.1.5. The proofs for /
i
and /
n
are left as
exercises.
13.1.7 Remarks and Problems
(i) The details of the proofs of the cases, where o in 13.1.6 is /
i
or /
n
have not been given, since they are quite technical using the character-
izations of
op
(/
i
) respectively of
op
(/
n
), and we do not yet know
how important they will be. The cases of /
iP
, /
P

i
and o
P

i
have
not yet been investigated by us. The positive solution of one of the
following two problems would help a lot.
(ii) All the results in 13.1.6 would be trivial consequences if the following
conjecture (or even a stronger one) could be proved:
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 333
Conjecture: Let F L
d

,d
be any set of formulas for Alg(S), let
/
F
be the class of all homomorphisms reecting all formulas
from F. Then
op
(/
F
) is k-dual algebroidal for every regular
cardinal k d

.
We have no idea whether this conjecture is true. But it should be
observed that
op
(/
F
) is generated by a set c
F
of epimorphisms
between strongly k-small objects as
op
(/
F
) =
op
((c
F
)).
(iii) The solution of the following problem would probably be of great help
for the above one:
Problem: Let C be any (nice) category and c Epi C be any
class of epimorphisms. Is there a complete set of rules inducing a
concept of derivability of morphisms from a given class such that
a C-morphism e is derivable from c iff e
op
((c)) ?
Perhaps such results are known in category theory, but we have not yet come
upon them, and some experts whom we asked did not know about them,
either.
13.2 H-projective partial S-algebras
The results of this subsection are motivated by conditions 12.2.3.(1) and (2)
and the wish to apply this meta theorem in as many instances as possible.
We recall that we have already agreed in 12.3 to choose the class H of 12.2.3
as a class of surjective homomorphisms. Again we use the notation of 12.3.10.
Remark 13.2.1 Condition 12.2.3.(2) that every H-projective object of Alg(S)
has to be the direct limit of a k-directed system of strongly k-small H-
projective objects is obviously satised for k d

as long as we know that


each weak subalgebra generated by a subset of a xed generating set of an
H-projective object is H-projective, too. 12.2.3.(2) then follows from 11.2.2
and 4.4.7. Thus our main objective is to characterize H-projective objects
for dierent classes H.
Lemma 13.2.2 Pj H
w
is the class of all discrete partial algebras.
Proof That each discrete partial algebra is projective with respect to sur-
jective homomorphisms is a consequence of the Axiom of Choice. If A is not
discrete, then id
A
: A A cannot be lifted against id
A
: A
d
A in H
w
.
December 8, 2002
334 Chapter III Birkho type results for elementary implications
Corollary 13.2.3 H
w
satises conditions 12.2.3.(1) and (2).
Lemma 13.2.4 (i) Pj H
f
consists of all partial algebras P satisfying
() P is a partial Peano algebra on some Peano basis, say X.
() For all ,

, for all a dom


P
and b dom
P
, and for all
k (), l (

):
If ,=

, or if =

, but a ,= b, or if k ,= l, then a(k) ,= b(l),


and we always have a(k), b(l) X.
I.e. P is a partial Peano algebra on some set X such that P = X
B
1
P
X = D
1
P
X, and each element of X occurs at most once in any
sequence belonging to some of the fundamental partial operations (thus
these argument sequences are injective and mutually disjoint).
(ii) H
f
satises conditions 12.2.3.(1) and 2) for every regular cardinal k
d

.
Proof Let f : A B be any full and surjective homomorphism, let P be a
partial Peano algebra on X satisfying () and (), and let p : P B be any
homomorphism. Let x X. If there are , a dom
P
and k ()
such that x = a(k), then , a and k are uniquely determined by x. We have
p(a) dom
B
, and since f is full and surjective, there is c dom
A
such
that f(c) = p(a), and we choose one such c and dene q(a(l)) := c(l) for
each l () (in particular the q-value of x determines the q-value of any
other argument of a). If x is not in the domain of any such sequence, then
we dene q(x) arbitrarily in f
1
(p(x)). For each element a =
P
(a) of P we
dene (recursively) q(a) :=
A
(q(a)). Obviously q : P A satises f q = p.
It is easily seen that any partial algebra D can be represented as a quotient
of a partial Peano algebra P satisfying () and (): Choose for some o / :
X := (o A)

( a () [ and a dom
A
);
P X :=

( dom
A
[ );
graph
P
:= (( (, a, k) [ k () ), (, a)) [ a dom
A
, ;
p : P A : l(o, a) a,
(, a, k) a(k),
(, a)
A
(a).
It is easy to realize that P satises () and () and that p : P A is a full
and surjective homomorphism).
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 335
If Q is any partial algebra which does not satisfy () and () for any gen-
erating subset, then there is a congruence on a partial Peano algebra P
which satises () and () such that P/ is isomorphic to Q. Then id
P/
cannot be lifted against nat

, i.e. Q cannot belong to Pj H


f
.
Remark 13.2.5 It is easy to see that all partial algebras are Iso Alg(S)-
projective, and that 12.2.3.(1) and (2) are satised for any regular cardinal
k d

in connection with the operator 1.


Remark 13.2.6 Observe that for the class H
b
of all bijective homomor-
phisms of Alg(S) we have Pj H
b
= Pj H
w
(since we allow the Axiom of Choice),
and therefore we know from 13.2.3 that conditions 12.2.3.(1) and (2) are sat-
ised for H
b
and regular cardinals k d

.
But this means that switching between H
b
and H
w
in the cases of o as
o
r
, o
w
or /
i
where 12.2.3 is applicable to H
b
as well as to H
w
does not
change the corresponding sets of formulas. Therefore we get from Table 12.1
that
H
w
o = H
b
o,
H
w
oT
r,k
= H
b
oT
r,k
,
H
w
oT
r,k
+
= H
b
oT
r,k
+
,
where o is one of the operators o
r
, o
w
or /
i
.
However, it should also be observed that H
w
is additionally compati-
ble with other operators o, while we did not investigate their compatibility
with H
b
.
Let us add some more general results
Lemma 13.2.7 Let H and H

be classes of surjective homomorphisms; then


Pj (H H

) Pj H Pj H

A +B [ A Pj H, B Pj H

.
Proof Obviously Pj (H H

) Pj H Pj H

. Let (j
1
, j
2
; A
1
+A
2
=: B) be
a coproduct such that A
1
Pj H and A
2
Pj H

, and let f : E F be
any (H H

)-morphism, g : B F. By assumption there are h


i
: A
i
E
such that f h
i
= g j
i
(i 1, 2). Since (j
1
, j
2
; B) is a coproduct, there
is h : B E such that h j
i
= h
i
. Therefore g j
i
= f h
i
= f h j
i
(i 1, 2). Thus f h = g.
Lemma 13.2.8 Let P
1
and P
2
be partial Peano algebras, and let (j
1
, j
2
; P
1
+
P
2
) be their coproduct. Then j
1
and j
2
are injective.
December 8, 2002
336 Chapter III Birkho type results for elementary implications
Lemma 13.2.9 (i) Let H be a class of surjective homomorphisms and
(
op
(o), o) be a factorization system. Moreover, let e : Q P be
in
op
(o) with H-projective Q. Then P is c-projective for c := Ho.
(ii) If, in addition, Alg(S) has enough H-projectives, if H is the left factor
of a factorization system (it suces that H is right cancellable with
respect to epimorphisms), and if
op
(o) Epi Alg(S), then Alg(S) has
enough c-projective objects.
(iii) With the assumptions of (i) and (ii) we can also state that each c-
projective object P is isomorphic to a subalgebra of a
op
(o)-homomorphic
image of an H-projective object.
Proof Take the assumptions and notation of (i), let f : C D, f c,
and let g : P D. Since Q is H-projective, there is h : Q C such
that f h = g e. Now the assumptions on e and f imply the existence of a
diagonal-ll-in d : P C satisfying in particular f d = g, showing that P
is c-projective.
Now let the assumptions on H and o of (ii) also be satised, and let A
Alg(S). By assumption there exist Q Pj H and an H-morphism q : Q A.
Let q = s g be a (
op
(o), o)-factorization of q, g : Q P. By the rst part
P Pj (H o). Since g is an epimorphism, and since H is right cancellable
with respect to epimorphisms, we have s H ( o), i.e. A is an (H o)-
homomorphic image of an (H o)-projective object.
Ad (iii) Let P Pj c; since H has enough H-projectives, there are Q
Pj H and an H-morphism f : Q P. Let f = s t be a (
op
(o), o)-
factorization of f, t : Q Z. Because of (i) Z is c-projective. Since H is
right cancellable with respect to epimorphisms, s belongs to E, and therefore
id
P
: P P can be lifted against s to a homomorphism u : P Z satisfying
s u = id
P
. Thus u is a section and therefore closed (see 11.4.4), i.e. P is
isomorphic to a subalgebra of Z.
Proposition 13.2.10 Let H be one of the classes H
w
or H
f
, let P be any
set of terms, and let o be the class of all P-reecting homomorphisms. Then
c := Ho satises 12.2.3.(1) and (2). In particular each c-projective object
can be obtained in the way described in 13.2.9. More precisely using 10.3.2
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 337
and 13.2.4 we get
Pj H
P
= Q [ Q is a partial Peano algebra on some set X and Q = c
Q,P
X
Pj H
fP
= Q [ Q is a partial Peano algebra on some set X, and Q contains some
X-generated relative subalgebra B Pj H
f
such that Q = c
Q,P
B
Proof Let c be either H
P
or H
fP
and H the corresponding class H
w
re-
spectively H
f
. Using 10.3.5 the property in the description of Pj c applies
at least to those objects which are
op
(/
P
)-homomorphic images of H-
projective objects. By 13.2.9.(iii) each c-projective object is a subalgebra of
such a partial algebra, and it is easily seen (take the preimage of u[P] with
respect to t in the proof of 13.2.9.(iii)) that it also has to have the correspond-
ing property. Moreover, the fact that H
w
and H
f
have property 12.2.3.(2)
and the above description of Pj c imply that c has property 12.2.3.(2).
Corollary 13.2.11 Pj H
c
is the class of all partial Peano algebras, and H
c
satises conditions 12.2.3.(1) and (2).
Remark 13.2.12 Since H
i
H
c
and H
Pi
H
P
, we have Pj H
c
Pj H
i
,
and Pj H
P
Pj H
Pi
for any set P of terms. In particular the category theo-
retical denition of initialness implies that Pj H
i
= Alg(S), and therefore H
i
satises conditions 12.2.3.(1) and (2). Moreover, Pj H
i
= Pj 1, and therefore
we have for the closure operators connected with 12.2.3 that H
i
oT = 1oT,
where o o
i
, o
w
, /
i
and T e, T
r,
,T
r,k
+
, T
u
, T
u
e (see Table 12.1).
We leave a discussion of H
Pi
and similar operators as a problem to
interested readers.
13.3 Examples in Alg(S) for the Meta Birkho Theo-
rem
We now combine the results from 12.3, 13.1 and 13.2 and get the following
Theorem 13.3.1 Let k d

be a regular cardinal, and let P and P

be any
sets of terms. Table 13.1 shows, which operators H and o are now known
to satisfy the conditions of Lemma 12.2.3 (the Meta Birkho Theorem). An
entry means a positive answer, whenever P

P, while + means a
positive answer in general; we use the notation from Table 12.1. Missing
entries only mean that these pairs have not yet been investigated.
December 8, 2002
338 Chapter III Birkho type results for elementary implications
H o o
i
o
w
/
i
o
c
/
c
/
P
o
P
o
iP
/
n
1 + + + + + + + + +
H
c
+ + + + + + + + +
H
f
+ + + + + + + + +
H
w
+ + + + + + + + +
H
P
+ + +
H
fP
+ + +
H
i
+ + +
H
b
+ + +
Table 13.1: Pairs of operators H, o
satisfying the assumptions of the Meta Birkho Theorem (see 13.3.1)
Remarks 13.3.2 (i) We are quite certain that for all the positive entries
in Table 12.1 the corresponding operators H and o also satisfy con-
ditions 12.2.3.(1) and (2) and (6), and therefore lead to examples of
Theorem 12.2.5 in Alg(S).
It is quite likely that the proofs of the facts that the assumptions
of 12.2.3 are satised for a given pair (H, o) above could in most in-
stances be derived from some more general principles or results which
are not known to us at this moment. Such general principles and results
could make Theorem 12.2.5 and corresponding results of H.Andr eka
and I.N emeti even more valuable. The preceding investigations and
results are only meant to give a feeling for its possibilities and for some
methods to treat the corresponding problems.
(ii) We did not yet look for counterexamples for the cases where there is
no entry in Table 13.1. Thus it might be a good exercise to look for
positive or negative answers in those cases.
(iii) Table 13.1 contains 54 positive entries but since the H-operators
1 and H
i
, respectively H
w
and H
b
yield the same axioms in connec-
tion with their positive entries for o-operators, we actually have to
subtract 6 possibilities. 6 of these entries yield families with two (al-
most independent) parameters, while 18 of them yield families with
one parameter, and if the type is rich enough, these families are in-
nite. Omitting the row corresponding to 1 and observing 12.3.3 we get
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 339
from Table 13.1 624 dierent model theoretic closure operators, 336 of
which correspond to innite families. This illustrates how sensitive an
instrument Theorem 12.2.5 provides for partial algebras.
(iv) However this richness of the theory makes it dicult for the user to
nd the appropriate axioms and operators for his applications. We do
not want to hide that the situation may be even more complex, since
there are model theoretic approaches to parts of implicational partial
algebra, which do not t into the picture drawn so far: In computer
science (see Wirsing et al.) so called strong equality is often used
see 8.1.3.(b). In connection with other investigations, so called weak
equality is used see 8.1.3.(a) , e.g. in logics of quantum mechan-
ics (cf. S.Kochen, E.P.Specker [KoSp68]). For these concepts the
above instruments are not sucient. Here R.John (see [J75], [J77],
[J78]) has provided another approach for the description of model the-
oretic closure operators. But the application of his approach to partial
algebras looks less elegant than the presented one, and we do not have
the space to discuss it here in more detail.
13.3.3 Shape of the Implications The Meta Birkho Theorem does not
only tell us which model theoretic closure operators are around, in connec-
tion with 12.1.2 it also tells us the shape of the corresponding elementary
implications, once we know Pj H and
op
(o). The starting point is 12.1.2,
and we only consider HoT
r,k
-cones (k d

a regular cardinal), i.e. epimor-


phisms e : A B, where A Pj H is strongly k-small and e
op
(o),
which have all the essential features of other kinds of axioms. For the
corresponding elementary implication we can restrict our considerations to
one-element conclusions, i.e. we may only consider implications of the form
(M :
_

t
i
e
= t

i
[ i I t
e
= t

), #M, #I < k
_

t
i
e
= t

i
[ i I
corresponds to the positive diagram of A above, while t
e
= t

is taken from
the positive diagram of B.
The characterizations of Pj H and
op
(o), which we have got earlier for
some operators H and o then yield the restrictions collected in Table 13.2.(a)
and (b) (observe that we do not know nice conditions for /
n
; = means
formal identity of terms).
1
1
The original part (b) of Table 13.2, i.e.
December 8, 2002
340 Chapter III Birkho type results for elementary implications
(a) Restrictions on the premise
_

t
i
e
= t

i
[ i I in case of H are:
1 no restrictions
H
w
t
i
= t

i
and it is a variable from M (i I)
(actually the implication is an E-equation);
H
f
t
i
= t

i
and it has the form
i
( x
ki
[ k K

i
) (i I,
i
)
such that (k, i) ,= (k

, i

) implies x
ki
,= x
k

i
;
H
c
t
i
= t

i
is arbitrary (i I);
H
P
(t
i
= t

i
) c
F(M,TAlg(S)),P
M, i.e. each t
i
is generated from M
by iterative application of conditioned terms derived from P;
H
fP
I = J

K, and for j J the conditions for H
f
hold, while,
for k K, (t
k
= t

k
) c
F(M,TAlg(S)),P
( t
l
[ l J M).
(b) Restrictions on the conclusion t
e
= t

with respect to
_

t
i
e
= t

i
[ i I
as premise: The epimorphism induced by the implication must be for
o
w
a full and surjective homomorphism;
o
i
an arbitrary surjective homomorphism;
o
c
an arbitrary epimorphism;
/
i
an arbitrary hom., injective on the variables not identied to terms;
/
c
an epimorphism, where nothing must be additionally identied;
Table 13.2: Shape of the implications (see 13.3.3)
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 341
13.3.4 Proof of Theorem 8.3.5 All statements except for 8.3.5.(vi) follow
from 12.2.3 and Table 13.2. It remains to show 8.3.5.(vi), where we have to
assume that we use semantics in which each assignment interprets all of X,
i.e. we can omit the reference to the set of formulas:
If NTEK = , i.e. if there is no term t such that K [= (t
e
= t), then
Mod SECH
k
K is an ECE-variety and therefore equal to H
c
o
c
T
r,k
K. There-
fore, assume that NTEK ,= , say (q
e
= q) NTEK. We claim that
in Mod SECH
k
K =: K

also all formulas


_

t
i
e
= t
i
[ i I are valid
which are valid in K, i.e. that K

= Mod ECH
k
K = H
c
o
c
T kK: Namely
K [=
_

t
i
e
= t
i
[ i I iff K [=
_

t
i
e
= t
i
[ i I q
e
= q: va-
lidity of the implication and the fact that q
e
= q is never satised imply that
the premise is also never satised; the opposite direction is trivial.
13.3.5 Remark and Counterexample The restriction in 8.3.5.(vi) is in
particular satised, if we restrict considerations to the homogeneous case and
forbid the empty model.
The restriction is necessary, since we can consider a type := (1, 1, 1, 1)
with unary operation symbols ,
1
,
2
and
3
such that for S := s, s

, s

:
graph
S
:= ((s), s), graph
S
i
:= ((s

), s

), i 1, 2, 3.
Let x be a variable of sort s, y a variable of sort s

, a an element
of sort s, b an element of sort s

and c an element of sort s

, and con-
sider for all (i, j, k) with i, j, k = 1, 2, 3 the partial algebras A
ij
:=
(b) Restrictions on the conclusion t
e
= t

with respect to
_

t
i
e
= t

i
[ i I as premise:
o
w
t, t

t
i
, t

i
[ i I M;
o
i
t arbitrary, t

t
i
, t

i
[ i I M;
o
c
t, t

arbitrary;
/
i
t arbitrary, t

t
i
, t

i
[ i I M
and not both of t, t

are variables;
/
c
t = t

arbitrary;
o
P
t, t

c
F(X,TAlg(S)),P
( t
i
, t

i
[ i I M);
/
P
t = t

c
F(X,TAlg(S)),P
( t
i
, t

i
[ i I M);
o
iP
t arbitrary, t

c
F(X,TAlg(S)),P
( t
i
, t

i
[ i I M).
was wrong or at least too special. For the moment we have replaced it by more descriptive
characterizations, and in some cases this will be the best thing achievable. This error has
been copied into the corresponding tables of [B93a] and [B93b].
December 8, 2002
342 Chapter III Birkho type results for elementary implications
(a, b, c;
A
ij
,
A
ij
1
,
A
ij
2
,
A
ij
3
) such that graph
A
ij
k
:= graph
A
ij
:= and
graph
A
ij
i
:= graph
A
ij
j
:= ((b), c). Finally, set K := A
ij
[ i ,= j in 1, 2, 3 .
Then in K the axioms (x : ((x)
e
= (x))) and (y : (
1
(y)
e
=

1
(y)
2
(y)
e
=
2
(y)
3
(y)
e
=
3
(y))) are valid, but from the corre-
sponding elementary implication (x, y :
1
(y)
e
=
1
(y)
2
(y)
e
=
2
(y)

3
(y)
e
=
3
(y) (x)
e
= (x)) we can only derive (x, y : (
1
(y)
e
=

1
(y)
2
(y)
e
=
2
(y)
3
(y)
e
=
3
(y)). Since we do not see any other
negation of an E-equation valid in K and not needing a variable of sort s,
Mod SECH
k
K contains as model a partial algebra B on B := b, c, where

B
i
is total for i 1, 2, 3; however, B is not contained in H
c
o
c
T
r,k
+
K, while
(x : ((x)
e
= (x))) excludes the product with empty index set, of which
B is a subalgebra.
We do not have the feeling that in the general heterogeneous case a result
similar to 8.3.5.(vi) can be proved.
Remark 13.3.6 Because of its category theoretical nature the Meta Birkho
Theorem also applies to other categories, e.g. to E-varieties in Alg().
13.4 Special examples
13.4.1 Total (S-) algebras The class TAlg(S) of all total (S) algebras is
a TE-variety of Alg(S) dened by the axioms
(T

)
(var x : x
e
= x), v
X
(x) = dom
S
, x without repetitions, for all .
Thus TAlg(S) is closed among others with respect to the operators H
w
, /
c
and T
r,d

.
13.4.2 Relational systems As already mentioned earlier, relational sys-
tems of type may be considered as partial algebras of type , since with
each K

-ary relational symbol we can associate a K

-ary operation
symbol

and choose some xed k

in the nitary case this will


usually be the rst component , and we then replace the relational sys-
tem M := (M, (
M
)

) by the partial algebra A(M) := (M, (


A(M)

)
satisfying
()
a dom
A(M)

iff a
M
; and if a
M
, then
A

(M)(a) := a(k

).
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 343
The homomorphisms between relational systems are then the same as the
homomorphisms between the corresponding partial algebras. It is easily seen
that the class of all partial algebras corresponding in this way to relational
systems of a given type forms an ECE-variety in Alg(S) dened by the set
(RelS

) [ of axioms, where for :


(RelS

)
(var x :

(x)
e
=

(x)

(x)
e
= x(k

)), x X
K

injective.
The shape of the axioms shows that they are H
f
o
w
T
r,d

-axioms, i.e. that


H
f
o
w
T
r,d

= On the other hand the special form of the partial op-


erations in implies that is not H
w
-closed, since H
w
does not impose
restrictions on the inner extensions of the factor algebras.
13.4.3 Partially ordered sets (posets) According to 13.4.2 posets can
be considered as partial algebras of type := (2), and they form an ECE-
variety dened by the following axioms (we use the inx notation xy for
(xy)):
(Pos1) xy
e
= xy xy
e
= x, (relation)
(Pos2) xx
e
= xx, (reexivity)
(Pos3) xy
e
= xy yx
e
= yx x
e
= y, (antisymmetry)
(Pos4) xy
e
= xy yz
e
= yz xz
e
= xz, (transitivity).
(Pos3) and (Pos4) are no H
f
o
w
T
r,d

-axioms, since both premises violate


the form for H
f
, and the conclusion of (Pos4) is not of the form for o
i
,
but only of the form for o
c
. Observe, however, that in (see 13.4.2) the
concepts of subalgebra and relative subalgebra coincide. The behaviour
of posets shows that the class Pos of all posets is indeed closed with respect
to the operators H
c
, o
c
(here equal to o
r
), T
r,
0
and T
u
, while it is neither
closed with respect to full homomorphic images nor with respect to weak
subalgebras.
13.4.4 -continuous posets (the class Pos

) It is shown in [LePa80]
that the class Pos of all -continuous posets ( =
0
) which are an impor-
tant tool in D.Scotts xed point approach to the description of programs
in computer science is an innitary homogeneous ECE-variety: := (2, ),
December 8, 2002
344 Chapter III Birkho type results for elementary implications
multiplication being the binary, the -ary operation; axioms:
Pos

i := (Posi) for 1 i 4;
(Pos

5)
_

x
i
x
i+1
e
= x
i
x
i+1
[ i ( x
i
[ i )
e
= ( x
i
[ i ),
(Pos

6) ( x
i
[ i )
e
= ( x
i
[ i )
_

x
i
x
i+1
e
= x
i
x
i+1
[ i .
If there has to be a smallest element, usually denoted by (and called
bottom), then the following axioms have to be added:
(Pos

7)
e
= ,
(Pos

8) x
e
= x.
Because of (Pos

5) and (Pos

7) these axioms only allow H


c
, o
c
and T
r,#
1
as compatible operators, and we do not see how to do better.
13.4.5 Partially ordered and -continuously ordered algebras Let
(, , ()

) be any family of operation symbols corresponding to a simi-


larity type = (2, 0, (K

) such that (Pos1), (Pos2), (Pos3), (Pos4),


(Pos

7), (Pos

8) as well as for each

the axioms (T

) (see 13.4.1)
and (OA

) are valid, where for x := ( x


k
[ k K

), y := ( y
k
[ k K

)
(OA

x
k
y
k
e
= x
k
y
k
[ k K

(x) (y)
e
= (x) (y),
meaning that has to be monotone. Once more we deal with a homogeneous
ECE-variety.
Assume that we have an additional -ary partial operation satisfying
(Pos

5), (Pos

6) and that we have for each

and for each sequence


x := ( x
i
[ i ) with x
i
= ( x
ik
[ k K

) (i ) the axiom
(CA

)
_

( x
ik
[ i )
e
= ( x
ik
[ i ) [ k K

( ( x
i
) [ i )
e
= ( ( x
ik
[ i ) [ k K

).
Thus we get the ECE-variety of -continuous algebras. This could also be
done for S-algebras and partial S-algebras, although in the latter case the
axioms (OA

) and (CA

) become a little bit more complicated, since some


applicabilities of have to be assumed in the premises. Thus everything
that has been said so far or will be said later on ECE-varieties is applicable
to -continuously ordered algebras. In particular free objects exist, but as
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 345
we shall see in sections 16 the partial operations and are trivial on
them, i.e. only dened on constant sequences. But since also all universal
solutions exist, this can be overcome by starting with a structure in which
the partial order , and therefore the supremum , are less trivial. The class
of -continuous algebras plays a fundamental role in the semantics oriented
parts of computer science. For more details see e.g. the papers of the ADJ-
group (see 22.2.3) or of A.Pasztor [Pa83] and [Pa84] and consult [AN79f]
with its references.
13.4.6 hep-varieties The preceding examples already provide an idea
about the relevance of the results of sections 12 and 13. Observe that the
hep-varieties considered in Part II are QE-varieties, since their basic idea is
the following one: for each partial operation its domain is the set of all
solutions of some set G

of E-equations, i.e.
() (M

: (x)
e
= (x)

t
i
(x)
e
= t

i
(x) [ i I

).
In order to avoid loops it is additionally required that there exists a mapping
h :
0
(hierarchy mapping) such that each operation in h
1
(0) has to be
total, and if h() = n, then in G

there occur only fundamental operation


symbols

satisfying h(

) < n. This hierarchy condition implies that each


bimorphism is an isomorphism, and therefore the objects and morphisms of
hep-varieties can be considered to be still relatively close to total algebras
and their homomorphisms.
As it can be easily seen from the axioms for small categories and their
consequences (see 8.1.1), the class Cat

of all small categories is such a hep-


variety which is even an ECE-variety (for more details see Part II).
13.4.7 Commutative monoids with partial inverse Next, let us con-
sider the class CMpI of commutative monoids with partial inverse commu-
tativity is only chosen for simplicity of further observations. I.e. we consider
a type = (2, 0, 1) for = (, e,
1
) we use the usual inx notation for the
multiplication even omitting satisfying the following axioms:
(CMpI1) x(yz)
e
= (xy)z,
(CMpI2) xy
e
= yx,
(CMpI3) xe
e
= x,
(CMpI4) x
1
e
= x
1
xx
1
e
= e.
December 8, 2002
346 Chapter III Birkho type results for elementary implications
Observe that each of (CMpI1) and (CMpI2) implies totality of multiplication,
(CMpI3) implies totality of the nullary constant e, while formation of the
inverse may remain partial. We deal here with H
f
o
c
T
r,
0
-axioms. That we
cannot dene this class only by E-equations is easily seen.
13.4.8 If we add in 13.4.7 the axiom of the uniqueness of the inverse
(CMpI5) xy
e
= e x
1
e
= y,
then this is only a QE-equation; but the class dened by the axioms (CMpI1)
through (CMpI5) is easily seen to be closed with respect to closed homomor-
phic images. Thus it has to be an ECE-variety, and (CMpI5) may be replaced
by the two ECE-equations
(CMpI5

) e
1
e
= e and
(CMpI5

) (xy)
1
e
= (xy)
1
x
1
e
= x
1
(the proof is left as an exercise).
13.4.9 A proper QE-variety which is not an ECE-variety is given by the
class CSgI of all commutative semigroups with generalized inverse, where
= (2, 1), operations (,
1
), and where we have besides (CMpI1), (CMpI2)
and (CMpI3) the axioms
(CSgI4) x
1
e
= x
1
(xx
1
)(xx
1
)
e
= xx
1
and
(CSgI5) xy
e
= z zz
e
= z xz
e
= x yz
e
= y x
1
e
= y.
I.e. we do not have a central unit, but we may have local units such
as z in (CSgI5) and each existing generalized inverse refers to one such
local unit.
It is easy to check that CSgI is not closed with respect to closed homo-
morphic images (see [B82], Example 5.1), i.e. (CSgI5) cannot be replaced by
ECE-equations.
13.4.10 Weak or relative embeddability into total algebras When
we consider Table 13.2, then we also get a theoretical answer to the question,
when a partial algebra is weakly or fully embeddable into some total algebra
of some given variety 1 of total algebras. Namely, we have
1o
w
T
r,d

1 = Mod Cone
ISwP
r,d

1 = 1o
w
1, and
1o
i
T
r,d

1 = Mod Cone
IS
i
P
r,d

1 = 1o
i
1,
December 8, 2002
13 Construction of examples for the Meta Birkho Theorem 347
since each E-variety is closed with respect to T
r,d

, i.e. to reduced products.


Therefore, in order to belong to 1o
w
1 a partial algebra A has to satisfy
all axioms of the form (M :
_

t
i
e
= t

i
[ i I (t
e
= t

)), t, t

t
i
, t

i
[
i I X, #I < d

, which are valid in 1. If A shall belong to 1o


i
, it
has to satisfy all axioms of the form (M :
_

t
i
e
= t

i
[ i I (t
e
= t

)),
t

t
i
, t

i
[ i I X, t arbitrary, (where #I < d

) which are valid in 1.


It is a great problem to investigate for which varieties 1 there are re-
cursive or even nite subsets of these axioms which only have to be tested
for embeddability or even to decide whether there are any such varieties
besides TAlg(S).
We shall return to questions of embeddability in Chapter IV, but we do
not have enough space to discuss results of T.Evans [Ev51], I.Fleischer [Fl81]
or I.N emeti [N75].
13.4.11 Fields The well known fact that the class F of all elds is not
closed with respect to direct products (also for non-empty index sets) shows
that F is not denable by universal Horn formulas, although the fact that
E-equations hide some existential quantiers allows us to dene elds by a
collection of universal formulas. In order to do this, let us consider the type
= (2, 0, 1, 2, 0, 1) for = (+, 0, , , 1,
1
); then we have the axioms:
(F1) (x +y) +z
e
= x + (y +z),
(F2) (x y) z
e
= x (y z),
(F3) x +y
e
= y +x,
(F4) x y
e
= y x,
(F5) x (y +z)
e
= x y +x z,
(F6) x + 0
e
= x,
(F7) x 1
e
= x,
(F8) x + (x)
e
= 0,
(F9) (0
e
= 1),
(F10) (0
1
e
= 0
1
),
(F11) (x
e
= 0) x x
1
e
= 1.
Axiom (F11) is the only one which is not a universal Horn formula, but it
cannot be replaced by universal Horn formulas as observed above.
13.4.12 Further similar examples are elaborated in [N75].
December 8, 2002
348 Chapter III Birkho type results for elementary implications
14 Closed classes of epimorphisms (elemen-
tary implications)
14.1 Preparatory results
14.1.1 In the sense of abstract model theory we have investigated in sections
12 and 13 dierent (implicational) abstract languages, and we have investi-
gated the corresponding semantical closure operators Inj Cone
HSF
(see 12.1.2.(iii)).
In this section we are concerned with a characterization of the syntactical
closure operators Cone
HSF
Inj for some types of abstract languages discussed
earlier, namely for the case T T, T
+
, T
r,k
, T
r,k
+
, and some H, o as related
in Table 13.1. The closure operators for other cases of H, o and T are left as
open problems (or exercises, since we cannot exclude that the answer might
already exist somewhere).
14.1.2 (General assumptions and notation) Let C be any category,
H, o Mor C, k an innite regular cardinal or , such that C is strongly k-
algebroidal, possesses direct products and direct limits of k-directed systems,
and let H and o satisfy assumptions 12.2.3.(1) through (7). Recall that Pj H
designates the class of all H-projective C-objects, and Pj
k
H designates the
class of all strongly k-small H-projective objects.
c
H
S
:= e
op
(o) [ Dom(e) Pj H, and
c
H,k
S
:= e
op
(o) [ Dom(e) Pj
k
H
designate the classes of all
op
(o)-morphisms with H-projective (strongly
k-small) domain. These are HoT-morphisms.
For P Ob C, K Ob C we introduce Mor(P, K) :=

Mor(P, K) [
K K. For Ob C = Alg(S) and e c
H
S
Epi Alg(S) let (N : P(e)
C(e)) be the corresponding implication according to 8.2.2 such that e is
the encoding epimorphism of this implication, i.e. if e : A B, then N
is a generating subset of A, P(e) and C(e) are sets of E-equations with
variables in N such that P(e)/Cong
P(e)
P(e)

= A over id
N
(i.e. id
N
induces the isomorphism), and (P(e)C(e))/Cong
(P(e)C(e))
P(e) C(e)

=
B over e id
N
.
We shall always assume that T T, T
+
, T
r,k
, T
r,k
+
.
Lemma 14.1.3 With the notation and assumptions of 14.1.2 let C be a cat-
egory, let K Ob C satisfy K = HoT K, and let P, P

Ob C be arbitrary.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 349
(i) If Mor(P, K) = , then Mor(P, P

) ,= implies Mor(P

, K) = .
(ii) If Mor(P, K) ,= , then P has a K-universal K-solution (r
P,K
, F(P, K)) =:
(r
P
, F
P
) such that r
P

op
(o) and F
P
K, and if q : P

F
P
, then
also Mor(P

, K) ,= and there exists a unique morphism f


q
: F
P
F
P
such that q = f
q
r
P
.
14.1.4 In the language of category theory 14.1.3.(ii) means: if Mor(P, K) ,=
for each P Ob C, then K is
op
(o)-reective in C and (P F
P
, (e : P


P) (f
r
P
e
: F
P
F
P
)) is the description of the left adjoint functor with
respect to the inclusion functor into C of the full subcategory of C with object
class K.
Proof (of 14.1.3) (i) is obvious. Ad (ii): Since C is
op
(o)-co-(well-
powered), there is a set c(P, K)
op
(o) such that for each e c(P, K)
one has:
(1) Dom(e) = P.
(2) There are f Mor(P, K) and g o such that g e = f.
(3) If f Mor(P, K), and if (h

, s

) is a (
op
(o), o)-factorization of f, then
there exists an isomorphism i such that i h

c(P, K).
Thus c(P, K) is a representative system for all
op
(o)-morphisms leaving P
and occurring in a (
op
(o), o)-factorization of some f Mor(P, K). For
each e c(P, K) let Cod(e) =: C
e
. Then C
e
K by (2) and o K K. Let
B :=

( C
e
[ e c(P, K) ), then B K, since K is T
+
-closed and c(P, K) ,= .
Let e

: P B be the product morphism induced by c(P, K), and let (r


P
, s
P
)
be its (
op
(o), o)-factorization. Then F
P
:= Cod(r
P
) K, and it is routine
to realize that (r
P
, F
P
) is the K-universal solution of P. (cf. subsections 5.11
and 12.2). The rest of the statement then follows immediately from the fact
that also (r
P
, F
P
) is the K-universal solution of P

.
Theorem 14.1.5 Let the same assumptions be satised as in 14.1.2 and 14.1.3
except that we assume T = T
r,k
+
, and let
N
k
(K) := P [ P Pj
k
H, Mor(P, K) = ,
c
k
(K) := r
P
[ (r
P
, F
P
) is a K-universal solution of P Pj
k
H with Mor(P, K) ,= .
Then obviously c
k
(K) c
H,k
S
and moreover:
December 8, 2002
350 Chapter III Birkho type results for elementary implications
(i) K = Inj (N
k
(K) c
k
(K)),
(ii) Cone
HSF
K = N
k
(K) e c
H,k
S
[ Dom(e) N
k
(K) e c
H,k
S
[
m e = r
Dom(e)
for some m Mor C, Mor((, Dom)(e), K) ,= .
14.1.6 Observe that we get T =T
+
for k = , and if T = T
r,k
, then we
have to omit in 14.1.5 the consideration of N
k
. Notice that above we have
identied P with (P, ) and e with (Dom(e), e), and we shall do so, too, below.
Proof (of 14.1.5) From 12.2.3 we know using our assumptions that
K = Inj Cone
HSF
K, and N
k
(K)c
k
(K) Cone
HSF
K by denition of these sets
observe that each r
P
c
k
(K) is K-extendable. Therefore Inj Cone
HSF
K
Inj (N
k
(K) c
k
(K)).
Ad (ii) Let e c
H,k
S
, Dom(e) =: P, such that m e = r
P
for some m Mor C,
and let f : P K, K K; then P Pj
k
H, and there is g : F
P
K
such that f = g r
P
= g m e, showing that e Cone
HSF
K.
Conversely, assume that c Cone
HSF
K. Then either c is a cone consist-
ing of a single object, say c = (P, ), P Pj
k
H such that Mor(P, K) = ,
i.e. c N
k
(K), or c = (e : P C) is a
op
(o)-morphism with P Pj
k
H.
In the former case c is of the desired kind. In the latter case we
have either F
P
K for the K-universal solution (r
P
, F
P
) of P, and
c Cone
HSF
implies the existence of m : C F
P
such that m c = r
P
,
or Mor(P, K) = and there are no further restrictions on e (in each case
c c
H,k
S
).
Ad (i) Assume A Inj (N
k
(K) c
k
(K)). Then, for each P N
k
(K), we have
Mor(P, K) = . Now assume that (e : P D) Cone
HSF
K and
f : P A. If Mor(P, K) ,= , then because of (ii) there is an m
such that m e = r
P
. Since r
P
c
k
(K), there is g : F
P
A such that
f = g r
P
= (g m) e; otherwise e N
k
(K), contradicting the existence
of f. This shows that A Inj Cone
HSF
K = K and implies equality in (i).
14.2 The general theorem
14.2.1 The results of 14.1 provide an idea of what closed sets of HoT-cones
have to look like (cf. also R. John [J77], Proposition 2). The following the-
orem presents an abstract description of the corresponding closure operator.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 351
As in 14.1 we restrict our considerations to the case where negations of con-
junctions of equations are allowed, i.e. T = T
r,k
+
. The other cases for T can
be obtained by setting k = , respectively by neglecting the cones corre-
sponding to negations, in particular with the notation of 14.2.2 below
we then have P
2
= . Observe that sometimes we identify a cone with the
corresponding object or mapping without mentioning it explicitly.
Theorem 14.2.2 Let C be a category, H, o Mor C, k a regular innite car-
dinal and T = T
r,k
+
such that the assumptions of Lemma 12.2.3 are satised;
moreover, let us use the notation of 14.1.2 and 14.1.3. Let E Cone
HSF
,
dene P
2
:= P Pj
k
H [ (P, ) E, P
1
:= Pj
k
H P
2
, and for P P
1
,
e, e

c
H,k
S
, e : P D and e

: P D

, for which there is q : D D

satisfying q e = e

, set q =: q
P, e

e
and e
P
e

(observe that q is unique, if


it exists, since e is an epimorphism).
(a) The following statements are equivalent:
(i) E = Cone
HSF
Inj E.
(ii) The following conditions are satised:
(1) For each P P
1
the class e [ e : P Cod(e) in E is
non-empty, quasiordered by
P
, and has a largest element,
say r
P
: P F
P
.
(2) For each P P
1
, P

P
2
one has Mor(P

, F
P
) = .
(3) If P, P

Pj
k
H, then
() P P
2
and Mor(P, P

) ,= imply P

P
2
,
() P P
1
and Mor(P

, P) ,= imply P

P
1
.
(4) If P, P

P
1
, p Mor(P, F
P
), then there is a unique d
p
:
F
P
F
P
such that p = d
p
r
P
.
(5) r
P
[ P P
1

op
(o).
(6) E = P
2
e [ e : P D in c
H,k
S
, P P
1
, m e = r
P
for
some m Mor(D, F
P
) e [ e : P D in c
H,k
S
, P P
2
.
(b) If (ii) above is satised, then Inj E = Inj (P
2
r
P
[ P P
1
) = HoT T
P
[ T P

.
Proof Ad (a) (i) implies (ii) is an easily veried consequence of 14.1.3
and 14.1.5 in particular (1) and (6) follow from 14.1.3.(ii) and 14.1.5.(ii),
where here (r
P
, F
P
) = (r
P,Inj E
, F(P, Inj E) is the Inj E-universal solution
of P.
December 8, 2002
352 Chapter III Birkho type results for elementary implications
(ii) implies (i): Dene K
0
:= F
P
[ P P
1
, and for use in (b): E

:=
(P, ) [ P P
2
(P, r
P
: P F
P
) [ P P
1
(E

is usually identied
with P
2
r
P
[ P P
1
.) And let E be dened as in (ii).(6), where we
now have to think of the corresponding cones. It is obvious that indeed
E Cone
HSF
and E

E. We want to show that E = Cone


HSF
K
0
, since
then Cone
HSF
K
0
= Cone
HSF
Inj E, proving (a): E Cone
HSF
K
0
: Let
c E. If c = (P, ) or c = (P, e : P Cod(e)) with P P
2
, then
c Cone
HSF
K
0
by (ii).(2). Therefore assume that c = (P, e : P D)
with P P
1
, e
op
(o) and some m
e
satisfying m
e
e = r
P
. With (4)
we see that for f : P F
P
(P

P
1
) there is d
f
: F
P
F
P
such that
f = d
f
r
P
= (d
f
m
e
) e, showing that c Cone
HSF
K
0
.
Cone
HSF
K
0
E

: Let c Cone
HSF
K
0
, then we have to distinguish
two cases:
Case 1: c = (P, ), P Pj
k
H: Then Mor(P, K
0
) = , i.e. P has to belong
to P
2
, because for P P
1
we have r
P
: P F
P
K
0
.
Case 2: c = (P, e : P D): Since c Cone
HSF
K
0
, we have P Pj
k
H
and e
op
(o), i.e. e c
H,k
S
. If P P
2
, then c c by (6). Thus
assume P P
1
. Then c Cone
HSF
K
0
implies that for r
P
: P F
P
there has to exist g : D F
P
such that r
P
= g e, showing with (6)
that c E.
Ad (b) Let E be closed. From the denition of E

above, we have Inj E


Inj E

, but the description of E in (ii).(6) shows that Inj E

Inj E. E


E Cone
HSF
and E = Cone
HSF
K
0
show in connection with 12.2.3 that
Inj E

= Inj E = HoT K

.
Remark 14.2.3 The above Theorem 14.2.2 shows what closed sets of im-
plications have to look like, when we translate them into epimorphisms.
Families of morphisms satisfying condition 14.2.2.(ii).(4) have been called
strongly natural families in R. John [J77], where it has been observed
that in the case when P
1
Pj
op
(o) we only have to require natural fam-
ilies ( r
P
: P F
P
[ P P
1
), i.e. we only need that for every p : P P

(P, P

P
1
) there exists a unique d
P
: F
P
F
P
such that d
p
r
P
= r
P
p.
We repeat that in the case when T allows products with empty index
sets, then P
1
= Pj
k
H.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 353
14.3 Closed sets of E-, ECE- and QE-equations and of
H
f
o
c
T
r,k
-implications
14.3.1 Theorem 14.2.2 now allows us to describe without great additional
eort closed sets of E-, ECE
k
- and QE
k
-equations as well as closed sets of
H
f
o
c
T
r,k
-implications (k d

, regular). In 14.4 we shall treat the case when


we also allow negations of conjunctions of E-equations. The descriptions of
the kinds of axioms listed above are easier than other ones, since we do not
have any restrictions on the scope of the conclusions (having o = o
c
), while
this is dierent for other specications of the operator o. Within the scope
of this book we shall not treat the other cases which as far as we know
have not yet been treated elsewhere in the literature. Theorem 14.2.2 and
the following results provide an idea of how such descriptions in more general
cases could be approached, but we leave it as an open problem to go into the
details there.
Notation 14.3.2 In order to translate Theorem 14.2.2 for the categories
Alg(S) and TAlg(S) and for some special kinds of QE-equations into a set
theoretical and algebraical language we use the following representation of
elementary implications ??? = (M :
_

t
i
e
= t

i
[ i I t
e
= t

): ??? =
(P, C), where P := (t
i
, t

i
) [ i I (x, x) [ x M , C := (t, t

).
For a given regular cardinal k and any set M let Subs
k
M := N M [
#N < k be the set of all k-small subsets of M.
Let as before be X a set of variables, v
X
: X S such that
#(v
1
X
(s)) = k d
&
(s S), let, for M X, T
M
:= F(M, TAlg(S))
be the set of all terms with variables in M,
T
:= (t, t) [ t T
X
,

M
:= (y, y) [ y M , Eeq
X
:= (t, t

) [ t, t

T
X
, v

X
(t) = v

X
(t

)
be the set theoretical equivalent of the set of all E-equations with variables
in X; let
f
:= (y, y), ((x), (x)) [ x X
K
injective, sort respecting,
y X, be the set of all TE-statements with respect to the 0-th
and rst Baire classes of F(X, TAlg(S)) with injective argument sequences,
and Subs

k

f
:= K Subs
k

f
[ ((x), (x)), (

(y),

(y)) K, ,


imply x[K

] y[K

] = be the set of all k-small subsets of


f
, which
describe H
f
-projective objects.
Let Prem be a variable for one of the sets Subs
k
Eeq
X
, Subs
k

T
, Subs
k

X
or Subs

k

f
, i.e. a variable for the set of admissible premises concerning the
cases of QE-, ECE- and E-equations, respectively of H
f
o
c
T
r,k
-implications.
Thus, for a given value of Prem the set of elementary implications under
December 8, 2002
354 Chapter III Birkho type results for elementary implications
consideration is described as
Imp
Prem
:= (, (t, t

)) [ = (t
i
, t

i
) [ i I Prem, var (t
e
= t

) var .
If Q Imp
Prem
and Prem with set M of variables, then Q() :=
(t, t

) [ (, (t, t

)) Q, t, t

X implies t, t

var is the set of all


conclusions of Q-implications with premise . Observe that in general Q()
should contain (x, x) for each x X satisfying v
X
(x) v

X
[], but we
omit new variables here with respect to later applications.
Closed sets of implications of premise type Prem are then described as
follows:
Theorem 14.3.3 With the notation introduced in 14.3.2 let Q Imp
Prem
be any set of elementary implications of length less than k and with premises
restricted by Prem. Then the following statements are equivalent:
(i) Q = Imp
Prem
Mod Q.
(ii) Q has the following properties (I1) through (I4) for any ,

Prem
with M := X respectively M

:= X

:
(I1) Q() is an M-generated relative subalgebra of F(M, TAlg(S)).
(I2) Q() is a closed congruence relation on Q().
(I3) Q().
(I4) For every homomorphism f : Q(

) satisfying (f(x))(f)[Q]
Q(

) there exists a homomorphic extension f

: Q()
Q(

) which satises f
(2)

[Q()] Q(

).
Corollary 14.3.4 (1) With the assumptions and notation of Theorem 14.3.3
let Q Imp
Prem
be a closed set, i.e. Q = Imp
Prem
Mod Q.
(i) If Prem with set M of variables, then Q() = ker f

[ f :
A Mod Q, ker f

.
(ii) Let ,

Prem for sets M respectively M

of variables such that


M M

, v

X
[] = v

X
[

] and
X
=

X
, i.e. and

rep-
resent the same conjunction of E-equations, but for dierent reference
sets of variables. Then there exists a homomorphism g :

Q()
satisfying g[
M
= id
M
, moreover Q() = c
Q(

)
M and Q() =
Q(

) (Q())
2
.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 355
Remarks 14.3.5 (i) (I1) through (I3) just say that, for each Prem
with set M of variables, Q() is a closed congruence containing
on an M-generated relative subalgebra of F(M, TAlg(S)). (I4) is
the crucial axiom corresponding to full invariance in connection with
E-equations. Thus this result is an immediate generalization of the
Second Birkho Theorem for equations (see 6.4.2). (I4) encodes strong
naturalness (see 14.2.2.(ii).(4)) of the family ( r
P
[ P P
1
) in 14.2.2.
(ii) For the dierent meanings of the choice of Prem see 14.3.2.
For E-equations (I3) becomes trivial, while in the case of E- and ECE-
equations as well as of H
f
o
c
T
r,k
-implications (I4) reduces to
(I4) For every homomorphism f : Q(

) there exists a homo-


morphic extension f

: Q() Q(

) satisfying f
(2)

[Q()]
Q(

).
(iii) In the case of non-empty phyla one can restrict considerations in 14.3.3.(ii)
to X-generated relative subalgebras of F(X, TAlg(S)) (see [ABN80],
Theorem 3.4.5 for the homogeneous case): In this case we can assume
that, for every Prem, Q() contains (x, x) for every x X, and if
we include all these elements, then all phyla of Q() are non-empty
and Q() allows assignments from X (respectively assignments start-
ing from subsets of X can be extended to all of X without violating
e.g. the assumptions in (I4)).
(iv) In the case of total algebras references to relative subalgebras have to
be replaced by such to subalgebras. And in the case of non-empty phyla
we can always refer to the total term algebra on X. See in particular
the following Corollary (2).
Corollary 14.3.6 (2) With the notation introduced in 14.3.2 let us consider
the case of total algebras: Let Q Imp
Prem
be any set of elementary implica-
tions of length less than k, let Mod
t
Q designate the class of all total algebras
in which Q is valid; and let Mod
t
e
Q be the class of all models of Q which are
total and have non-empty phyla.
(a) The following two statements are equivalent:
(i)
t
Q = QEeq Mod
t
Q.
December 8, 2002
356 Chapter III Birkho type results for elementary implications
(ii)
t
Q has the following properties for ,

Prem with sets M re-


spectively M

of variables:
(I1)
t
Q() is a congruence on F(M, TAlg(S)) containing .
(I2)
t
For every homomorphism f : F(M, TAlg(S)) F(M

, TAlg(S))
f
(2)
[] Q(

) implies f
(2)
[Q()] Q(

).
(b) The following two statements are equivalent:
(i)
t
e
Q = QEeq Mod
t
e
Q.
(ii)
t
e
Q has the following properties for ,

Prem with sets X of


variables:
(I1)
t
e
Q() is a congruence on F(X, TAlg(S)) containing .
(I2)
t
e
For every endomorphism f : F(X, TAlg(S)) F(X, TAlg(S))
f
(2)
[] Q(

) implies f
(2)
[(Q)] Q(

).
Proof (of 14.3.3) Theorem 14.3.3 is just a translation to partial algebras
of Theorem 14.2.2, and we only recall a few items:
Let (r
P
: P F
P
, F
P
) be the Mod Q-universal Mod Q-solution of the premise
algebra P := ()/Cong

, and let nat

: P be the corresponding
natural homomorphism. Then the Diagram Completion Lemmas 2.7.1, 2.7.2
and 5.8.2 tell us that Q() = ker (r
P
nat

is a closed congruence, and that


especially because of 14.2.2.(ii).(4) all the other properties are satised
and conversely imply the corresponding properties in 14.2.2.(ii).
Proof (of 14.3.4)
Ad (i) From the denitions and results in 5.11 we see that, with respect
to M := X and P := ()/Cong

, Q() is the preimage of


char
P
Mod Q, with respect to the natural homomorphism fromF(M, TAlg(S))
onto F(P, TAlg(S)) induced by the natural mapping from M into P, and
(i) is just the algebraic translation of this fact.
Ad (ii) In view of (i) this can now be proved in a similar way as 6.2.3.(ii).
We leave the details as an exercise to the reader.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 357
Remark 14.3.7 In [Hoe82] H.-J. Hoehnke has shown for the total ho-
mogeneous nitary case that for Q = QEeq Mod
t
e
Q the set Q() [
Subs
#
0
F(X, TAlg())
2
is a fully invariant algebraic closure system (see 14.3.8
for the denition) on Cong F(X, TAlg()), X being a countably innite set of
variables. Moreover, he shows that Q() [ Subs
#
0
F(X, TAlg())
2
[
Q Imp
Prem
, Q = QEeq Mod
t
e
Q Prem = Subs
k
Eeq
X
is the set of all
fully invariant algebraic closure systems on Cong F(X, TAlg()). At the end
he conjectures that a similar result is true for partial algebras. We show
below that he is right.
Following H.-J. Hoehnke in [Hoe82] we dene fully invariant closure
systems as follows in 14.3.8. Our denition becomes a bit more complicated
than his, since besides partiality we also have to take care of the dierent
reference sets of variables. We only give this denition here for the closure
system of M-initial closed congruences (M Y ) of a total Peano algebra
F(Y, TAlg(S)) on some set Y usually our standard set X of variables.
Denition 14.3.8 Let Y be any (S-) set.
(i) In analogy to Denition 5.9.2 we dene
Rsuba
Y
F(Y, TAlg(S)) :=
_
Rsuba
M
F(M, TAlg(S)) [ M Y
to be the set of all M-generated relative subalgebras of F(Y, TAlg(S))
for some subset M of Y . Similarly we dene
Cong
c,Y
F(Y, TAlg(S)) :=
_
Cong
c
A [ A Rsuba
Y
F(Y, TAlg(S))
to be the set of all M-initial closed congruences of F(Y, TAlg(S)) for
subsets M of Y , thus obtaining a k-algebraic closure system k d
&
,
regular (with respect to set theoretical inclusion) on Subs F(Y, TAlg(S))
2
.
(ii) A closure system H on Subs F(Y, TAlg(S))
2
is called fully invariant iff
for any closed set C H and for any closed and surjective homomor-
phism f : A C for some A Rsuba
Y
F(Y, TAlg(S) C being as
usual the initial relative subalgebra of F(Y, TAlg(S)) generated by the
domain of C one has for B := (f
(2)
)
1
[C] that B = A and B H.
Observation 14.3.9 Let H be a closure system on a closure system ( on a
complete lattice L; then H is a closure system on L itself, since the meet on
( and L is the same.
December 8, 2002
358 Chapter III Birkho type results for elementary implications
The next lemma characterizes some kind of fully invariant closure sys-
tems, which we shall need later:
Lemma 14.3.10 Let H be a closure system on Cong
c,Y
F(Y, TAlg(S) for
some (S-)set Y with closure operator c
H
such that, for all D F(Y, TAlg(S))
2
,
c
H
D is generated by Y D. Then the following statements are equivalent:
(i) H is fully invariant.
(ii) For every H H, for every M Y and for every homomorphism
g : F(M, TAlg(S)) F(Y H, TAlg(S)) one has (g
(2)
)
1
[H] H.
(iii) For every D Subs F(Y, TAlg(S))
2
, for every H H and for every
homomorphism f : D H which satises f
(2)
[D] H there exists
a homomorphic extension f

: c
H
D H such that f

(2)
[c
H
D] H.
14.3.11 (i) Observe that the condition on the closure system H in 14.3.10
that c
H
D is generated by Y D for all D F(Y, TAlg(S))
2
is
equivalent to the fact that for each subset M Y the set F(M, TAlg(S))
2
is closed in particular, since to each subset W of F(Y, TAlg(S)) the
set Y W of variables occuring in W is uniquely determined.
(ii) The fact that we deal in 14.3.10 with a closure system on Cong
c,Y
F(Y, TAlg(S))
trivially implies that in 14.3.10.(ii) f
1
[D] = (f
(2)
)
1
)[D]; therefore
this does not have to be required there separately.
Proof (of 14.3.10) Observe that by 2.8.2.(iii) the preimage of a closed
congruence with respect to a closed and surjective homomorphism is closed.
(i) (iii): Let Hbe a fully invariant closure system on Cong
c,Y
F(Y, TAlg(S)),
and let the assumptions of (iii) be satised. Let

f : F(D X, TAlg(S))
F(N, TAlg(S)), where N := f[D] X, be the homomorphic extension
of f.

f is closed and surjective, and H

:= H F(N, TAlg(S))
2
is closed
(see 14.3.11.(i)).
Thus D (

f
(2)
)
1
[H

] = (

f
(2)
)
1
[H] =: D

, and D

belongs to H. Therefore
monotony of c
H
implies c
H
D D

, and f

:=

f[
C
H
D
is a homomorphic
extension of f: f

: c
H
D H; moreover f

(2)
[c
H
D H.
(iii) (ii): Let H, g be given as in (ii), let D := (g
(2)
)
1
[H]. Then D,
H and f := g[
D
: D H satisfy the assumptions of (iii). However, by
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 359
the denition of D and f there has to hold: f

= f and c
H
D = D, when we
apply (iii).
(ii) (i): Let C H, A Rsuba
Y
F(Y, TAlg(S)), and let f : A C be
a closed and surjective homomorphism. Let M be the Peano basis of A, N the
one of C, and let

f : F(M, TAlg(S)) F(N, TAlg(S)) be the homomorphic
extension of f. Dening D := (

f
(2)
)
1
[C] as in (iii) (ii), closedness and
surjectivity of f and 4.1.13.(iii) imply that D = A and D H.
Theorem 14.3.12 Let k d

be a regular cardinal, X the canonical set of


variables (i.e. #(v
1
X
(s)) = k for each s S). There is a one-to-one cor-
respondence between the closed sets Q of k-small QE-equations and the fully
invariant k-algebraic closure systems H (respectively closure operators c
H
)
on Cong
c,X
F(X, TAlg(S)), which satisfy that c
H
H is generated by XH
for all H F(X, TAlg(S))
2
.
This correspondence is given by the assignments
Q c
Q
such that, for D F(X, TAlg(S))
2
, c
Q
D =

Q(D

) [ D

Subs
k
D.
H Q
H
such that Q
H
:= (D, (t, t

)) [ D Subs
k
F(X, TAlg(S))
2
, (t, t

) c
H
D.
Proof First assume that Q = QEeq Mod Q. Since k d

, 14.3.3 implies
that, for each D F(X, TAlg(S))
2
, c
Q
D is a closed congruence on a relative
subalgebra c
Q
D of F(X, TAlg(S)) generated by X c
Q
D because of the
k-inductivity of these concepts. Let D F(X, TAlg(S))
2
be arbitrary. It
is obvious that always D c
Q
D, and that D D

implies c
Q
D c
Q
D

(see (I3) and (I4)). By denition c


Q
is k-algebraic. In order to show that
c
Q
c
Q
D c
Q
D assume E c
Q
D, #E < k. For each E there is
E

D such that #E

< k and c
Q
E

. Set E

:=

[ E .
Since k is regular, #E

< k, and E c
Q
E

. Thus c
Q
E c
Q
E

c
Q
D
showing that c
Q
is idempotent. Finally assume D, D

F(X, TAlg(S))
2
, f :
D c
Q
D

such that f
(2)
[D] c
Q
D

. We know by 14.3.3.(ii).(I4) that for


each D

D with #D

< k we have f
(2)
[c
Q
D

] c
Q
D

. Thus f
(2)
[c
Q
D] =
f
(2)
[

c
Q
D

[ D

Subs
k
D] =

f
(2)
[c
Q
D

] [ D

Subs
k
D c
Q
D

showing that c
Q
is the closure operator of a fully invariant closure system,
since the assumption of 14.3.10 is trivially satised.
Conversely, let Hbe a fully invariant closure system on Cong
c,X
F(X, TAlg(S)),
which satises that c
H
D is generated by XD for each D F(X, TAlg(S))
2
.
Then it is easily seen that Q
H
satises properties 14.3.3.(ii).(I1) through (I4).
Thus Q is a closed set of implications.
December 8, 2002
360 Chapter III Birkho type results for elementary implications
The following result might be of some interest:
Proposition 14.3.13 With the notation of Theorem 14.3.12 let Q be a
closed set of k-small QE-equations. Then one has:
For all M X, for all M-generated relative subalgebras A of F(X, TAlg(S))
and for all closed congruences on A: A/ Mod Q iff H
Q
.
In particular Q() := c
Q
= ker f

[ f : A B Mod Q,
ker f

.
Proof It easily follows from 14.2.2.(a) and the Epimorphism Theorem that,
for k-small M and H
Q
, A/ belongs to Mod Q, where A = ; namely
for = Q(D) we have that (nat

[
D
, A/) is the Mod Q-universal solution
of D. In the general case we get A/ as the k-direct limit of such Mod Q-
objects, and therefore it also belongs to Mod Q (see 11.6.7).
The converse follows from similar considerations as those in connection with 14.3.4,
when we start with the relation of F(A X, TAlg(S)) and obtain that
Q() := c
Q
=

ker f

[ f : A B Mod Q, ker f

.
Remarks 14.3.14 (i) We have no idea whether there is a description of
a similar nature as in 14.3.12 for closed sets of ECE-equations. Here
the closed sets which one starts with are generated by k-small subsets
of
T
.
(ii) For more practical applications one often looks for complete sets of rules
of derivability, as we have already presented them for E-equations in
section 6. Such a list which can easily be derived from Theorem 14.3.3
has already been presented in [AN83a] for the kinds of axioms under
consideration (see also [B82], but observe that in both papers the axiom
corresponding to 14.3.15.(ii).(IR8) is incorrectly formulated, since the
essential assumption (M

q
m
e
= q
m
) [ m M is missing).
In what follows we use the implicational notation rather than the set
theoretical one.
Proposition 14.3.15 With the notation and assumptions of 14.3.2 let Q
Imp
Prem
i.e. we treat simultaneously E-, ECE- and QE-equations as well
as H
f
o
c
T
r,k
-implications. Then the following statements are equivalent:
(i) Q is closed, i.e. Q = Imp
Prem
Mod Q.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 361
(ii) Q is closed with respect to the following derivation rules, where the
premises and

are in each case taken from Prem, with sets M, M


X of variables, respectively, and #I < k:
(IR1) (M : x
e
= x) for all x M.
(IR2) (M : t
i
e
= t

i
), i I, when =
_

t
i
e
= t

i
[ i I .
(IR3) (M : t
e
= t

) (M : t
e
= t).
(IR4) (M : t
e
= t

) (M : t

e
= t).
(IR5) (M : t
e
= t

), (M : t

e
= t

) (M : t
e
= t

).
(IR6) (M : ( t
k
[ k () )
e
= ( t
k
[ k () ))
(M : t
k
e
= t
k
), k (), .
(IR7) (M : t
k
e
= t

k
) [ k () , (M : ( t
k
[ k () )
e
= ( t
k
[ k () )
(M : ( t
k
[ k () )
e
= ( t

k
[ k () )), .
(IR8) Let =
_

t
i
e
= t

i
[ i I :
(M : t
e
= t

), (M

q
m
e
= q
m
) [ m M ,
(M

t
i
( q
m
[ m M )
e
= t

i
( q
m
[ m M ) [ i I
(M

t( q
m
[ m M )
e
= t

( q
m
[ m M )),
t, t

, t
i
, t

i
, q
m
arbitrary terms (i I, m M)
on the corresponding sets of variables.
Proof Assume Prem with set M of variables, let Q Prem, and dene
D(Q; (M : )) := (t, t

) T
2
X
[ (M : t
e
= t

) is derivable from Q by
the rules in 14.3.14 .
Then (IR1), (IR3), (IR4), (IR5), (IR6) and (IR7) are equivalent to the fact
that D(Q; (M : )) is a closed congruence on its domain which is an M-
generated relative subalgebra of F(X, TAlg(S)) (for the latter see (IR1),
(IR3) and (IR6)). (IR2) takes care of the fact that (t
i
, t

i
) [ i I
D(Q; (M : )), when =
_

t
i
e
= t

i
[ i I . (IR8) is equivalent to (I4)
in 14.3.3, it is a general substitution rule. Thus the equivalence of (i) and (ii)
is established by Theorem 14.3.3.
Remark 14.3.16 Observe that the above rules can be simplied in the case
of total algebras (or of non-empty phyla). We leave a formulation of the
corresponding simplications to the reader as an exercise.
December 8, 2002
362 Chapter III Birkho type results for elementary implications
14.4 Closed sets of ECH- and Horn-formulas
14.4.1 In this subsection we want to discuss briey the cases of closed
sets of ECH-formulas respectively of universal Horn formulas, i.e. we now
allow negations of the form (M :
_

t
i
e
= t
i
[ i I ) respectively
(M :
_

t
i
e
= t

i
[ i I ) (#I < k, k d
&
, regular) in addition to ECE-
equations or QE- equations, respectively. As before, we let Prem be the set of
admissible premises, i.e. in this case Prem = Subs
k

T
or Prem = Subs
k
T
2
X
(see 14.3.2), where T
X
:= F(X, TAlg(S)) for the usual set X of variables.
Moreover, the conjunctions of the elements of elements of Prem now yield
those formulas which can be negated. As before let Q designate a closed
set of Ho
c
T
r,k
-axioms for H H
c
, 1 , and let Neg Q

be a set of premises
in Prem corresponding to a closed set of negations of conjunctions of the el-
ements of admissible premises such that Q

= Q (X :
_

[
) [ Neg Q

is a closed set of ECH- respectively of Horn-formulas.


Then we call Q the positive part of Q

, and we have
Lemma 14.4.2 With the notation and assumptions of 14.4.1 and ,


Prem we have:
(i) Neg Q

implies that Q(

) =
F(var

,TAlg(S))
, whenever there is a
homomorphism (necessarily compatible with the sort homomorphisms!)
from /Cong

into Q(

)/Q(

) (in particular for =

).
(ii) If Neg Q

, v

X
[] v

X
[

] and Q(

) =
F(var

,TAlg(S))
, then

Neg Q

.
(iii) If Neg Q

and

, then

Neg Q

.
(iv) The positive part Q of a closed set Q

satises 14.3.3.(ii).
Proof When we recall that Pj H
c
consists of all partial Peano algebras
and that Pj 1 consists of all partial algebras, then all these statements are
immediate consequences of Theorem 14.2.2 and the fact that for a closed
set Q

of ECH- respectively of Horn-formulas with positive part Q we have:


Neg Q

implies (, (t, t

)) Qfor every admissible (t, t

) F(var , TAlg(S))
2
(since can not be satised in any model of Q

). And therefore, if the as-


sumptions of (ii) are satised, but

/ Neg Q

, then F(var

, TAlg(S))/F(var

, TAlg(S))
would be a model of Q

allowing a homomorphism from /Cong

into
it, thus contradicting Neg Q

.
December 8, 2002
14 Closed classes of epimorphisms (elementary implications) 363
Remark 14.4.3 Observe that Neg Q

above is in general not uniquely de-


termined by the positive part Q of Q

, since for some sets of phyla there need


not exist a negation of a conjunction of (T)E-equations depending only on
variables of these sorts, but there may be a premise depending only on these
sorts such that all corresponding models only have one-element phyla. How-
ever, we always have Neg Q

Prem [ Q() =
F(var ,TAlg(S))
=: Q

.
And in the homogeneous case we only have the possibilities Neg Q

=
or Neg Q

= Q

.
In the general case 14.4.2.(ii) shows that every set V v

X
[] [
Q

=: Subs
S
Q

which is an upper end with respect to set theoretical inclu-


sion within (Subs
S
Q

; ) is in one-to-one correspondence with a possible set


Neg Q

:= Q

[ v

X
[] V . This is easily seen from 14.2.2 and
the above remarks.
We would also like to mention that the cases of closed sets of ECH- or
Horn-formulas (and even other cases) can be treated like the corresponding
ones for closed sets of ECE- or QE-equations, if we add to
T
respectively
T
X
T
X
a new element False (not depending on any variables or sorts)
such that
A [= (M :

t
i
e
= t

i
[ i I False) iff A [= (M :

t
i
e
= t

i
[ i I
Moreover, False implies Q() =
F(X,TAlg(S))
.
Thus we have proved the following
Proposition 14.4.4 Let k d
&
be a regular cardinal. Then a set Q

=
QNeg Q

of encoded ECH
k
- respectively Horn
k
- formulas is closed iff Q is a
closed set of ECE
k
- respectively QE
k
-equations and Neg Q

Prem such that


there is an upper end V of (Subs
S
Q

; ) (see 14.4.3) such that


Neg Q

= Q

[ v

X
[] V .
Remark 14.4.5 The reader might have realized in particular while dealing
with the last sections that a lot of sometimes relatively easy work can still
be done in connection with special realizations of applications of Theorems
12.2.5 and 14.2.2. But we hope that the results discussed so far form a good
basis, and that everybody who has to deal with similar problems not treated
December 8, 2002
364 Chapter III Birkho type results for elementary implications
here has at least found some help in this chapter for an easier treatment of
them.
We did not yet touch a related topic, which could be of some interest,
namely the axiomatization by so called notions of validity of equations in
partial algebras (see 13.3.2.(iii)). We do not have the space here and refer
e.g. to R.John [J75].
December 8, 2002
Chapter IV
Special Topics
In this chapter we mainly treat or extend some fundamental concepts which
we either touched so far not at all or only very briey. Moreover, in most
sections of this chapter we now investigate the intrinsic set theoretical respec-
tively algebraical structure of some concepts rather than treating them from a
category theoretical or model theoretical point of view which we mostly used
in the previous chapters. In this connection the fact that we also include
heterogeneous partial algebras into our considerations (but not the fact that
we treat them in the form of partial S-algebras) gains much more importance
than in the previous sections and causes on several occasions complications
which do not occur in the homogeneous case.
In section 15 we treat two concepts of superposition of operations and
derived concepts of algebraic partial operations of a given partial alge-
bra. At some instances there is a close connection to the concept of term
operations introduced in section 5.
In section 16 we briey touch some general concepts connected with
possible (minimal) completions of partial algebras, but we do not go into
details concerning the existence of special kinds of completions within some
kind of variety.
In section 17 some additional properties of independence and the rela-
tionship between independence classes and E-varieties are considered. The
existence with respect to special properties of relatively free comple-
tions within an E-variety, say K, for the members of K and in particular for
the K-freely generated K-algebras is investigated and (relatively) free partial
algebras are characterized. Finally, the so-called representation of initial
objects of (QE-) varieties by canonical terms is investigated.
365
366 Chapter IV Special Topics
In section 18 we return to the investigation of algebraic (respectively
term) operations from the point of view of dependence on variables. More-
over, the existence of a minimal generating subset (element basis) of a free
generating subset for any given element of a free partial algebra is investi-
gated.
In section 19 we characterize the process of generation of closed con-
gruences of nitary partial algebras, and we characterize the congruences
belonging to P-reecting surjective homomorphisms for some set P of terms
closed with respect to the formation of subterms.
In section 20 several concepts of subdirect representability of partial
algebras are investigated, the corresponding subdirectly irreducible objects
are characterized, and representation theorems are proved.
In section 21 we show, how some of the tools used in particular for model
theoretic investigations within the whole species of partial algebras can be
relativized to E-varieties satisfying some special conditions which are to a
great extent satised in the homogeneous case.
15 Algebraic operations
15.1 Superposition
15.1.1 In subsection 5.7 we have already briey treated term operations and
polynomial operations on a given partial algebra. In this section we return to
this concept from a dierent starting point, namely from the one of algebraic
partial operations, which are derived directly from the fundamental partial
operations and total projection operations of a given partial S-algebra.
There are several ways to dene partial operations composed from the
fundamental operations (and total projections) of a given partial algebra.
While some of them yield the same results, others result in sets of partial
operations which dier from each other. In each case one includes some
collection of identity operations as starting points, namely either the total
identity operation of each phylum, or a collection of total projections. A
comprehensive study of superposition of partial functions is contained in
[Hoe72] by H.-J.Hoehnke (see also A.I.Malcev [Mal74], &2).
Observe that we always more or less tacitly consider partial S-algebras
and S-sets if not stated otherwise for some sort algebra S, i.e. each
set, say M, is connected with a sort mapping v
M
: M S which is to
December 8, 2002
15 Algebraic operations 367
be a homomorphism from M into S, whenever M carries some given partial
algebraic structure.
15.1.2 Denition of basic concepts Let A, M be any S-sets, i.e. they
are connected with sort mappings v
A
: A S and v
M
: M S.
(i) An M-ary partial operation with output sort s on A is some mapping
f : D A
s
(for some s S, A
s
= v
1
A
(s), where D is a subset of
the set A
M
of all sort-respecting mappings from M into A, i.e. of all
mappings q : M A satisfying v
A
q = v
M
.
Observe that for some partial operation we always assume that all its
values are of the same sort. Moreover, observe that an M-ary partial
operation, say m, does not only consist of its graph, but it is actually
a triplet (A
M
, graph m, A
s
) consisting of the global domain which
contains the actual domain (i.e. in the above case: D), its graph and
its codomain. Thus partial operations with empty actual domain can
well be distinguished by their arities and the sort of their target sets.
(ii) By PO
M
(A; S) we denote the set of all M-ary partial operations on A,
i.e.
PO
M
(A; S) :=

A
D
s
[ s S, D A
M
, and by
O
M
(A; S) :=

A
A
M
s
[ s S
its subset of all total M-ary (sort compatible) operations. If S has only
one element, then we also simply write PO
M
(A), respectively O
M
(A).
(iii) For each m M we denote by e
M
m
O
M
(A; S) the m-th total M-
ary projection e
M
m
: A
M
A
s
(s := v
A
(m)): e
M
m
(a) := a(m) for each
a A
M
.
E
M
(A; S) := e
M
m
[ m M designates the set of all M-ary total
projections of A.
(iv) Let A be any partial S-algebra of type with cardinal dimension d, and
let X be any (S-) set, such that each phylum of X has cardinality d

.
Then we dene for d
&
= maxd

, (#S)
+
:
PO(A; X; S) :=

PO
M
(A; S) [ M X, #M < d
&
,
O(A; X; S) :=

O
M
(A; S) [ M X, #M < d
&
,
E(A; X; S) :=

E
M
(A; S) [ M X, #M < d
&
,
December 8, 2002
368 Chapter IV Special Topics
the sets of all partial, respectively total projection operations on A with
d
&
-small arity in X.
On some occasions we shall need ordinals as arities. Therefore, let
(d
&
) be the least ordinal such that # = d
&
. Then we dene
PO(A; ; S) := PO(A; (d
&
); S) :=

PO
,v
(A; S) [ < (d
&
), v : S ,
O(A; ; S) := O(A; (d
&
); S) :=

O
,v
(A; S) [ < (d
&
), v : S ,
E(A; ; S) := E(A; (d
&
); S) :=

E
,v
(A; S) [ < (d
&
), v : S .
Fundamental for the denition of algebraic structure on PO(A; ; S) and PO(A; X; S)
are the following two denitions of superposition (for a better understanding
see the examples in 15.1.5):
Denition 15.1.3 (i) Superposition of the rst kind Let h PO
M
(A; S),
let g
m
PO
Lm
(A; S) have output sort v
M
(m) for each m M, and
let N :=

M
L
m
:=

m L
m
[ m M , or, if M and L
m
(m M) are each ordinals, then let :=

M
L
m
:=

( L
m
[ m M )
in the sense of ordinal arithmetic (see section 23) be the disjoint
union respectively the ordinal sum of the arities, and let v
N
be
the disjoint union of the sort mappings of the sets L
m
(m M) (re-
spectively of those connected with the given ordinals). Then we dene
a partial operation h g := h ( g
m
[ m M ) PO
N
(A; S) (or
h g := h ( g
m
[ m M ) PO
,v
(A; ; S) in the ordinal case) with the
same output sort, say s, as h as follows:
For every a A
N
and some a A
s
we have
[h( g
m
[ m M )](a) = a iff h( g
m
( a(m, l) [ l L
m
) [ m M ) = a;
i.e. h g is dened on an N-ary (sort-respecting!) sequence a and has
value a, iff for each m M g
m
is dened on the sequence ( a(m, l) [
l L
m
) and takes a value, say b
m
:= g
m
( a(m, l) [ l L
m
), and h is
dened on ( b
m
[ m M ) and takes the value a : h( b
m
[ m M ) = a.
In the case of ordinal superposition of the rst kind we have for
M = and L

( < )
[h( g

[ < )](a) = a iff h( g

(a() [

[ < ) <

[ ) ) = a.
December 8, 2002
15 Algebraic operations 369
(ii) Superposition of the second kind Let h PO
M
(A; S), g
m

PO
L
(A; S) of output sort v
M
(m) for each m M; then we dene a
partial algebraic operation h ( g
m
[ m M ) PO
L
(A; S) as follows:
For every a A
L
and for every a A one has
[h ( g
m
[ m M )](a) = a iff h( g
m
(a) [ m M ) = a,
i.e. iff for every m M there exists some b
m
A such that g
m
(a) = b
m
and h( b
m
[ m M ) = a.
15.1.4 Remarks and further Denitions
(i) Observe that we can state equalities as in 15.1.3 iff both sides of the
equation are dened at the same time and equal, i.e. if we use in the
metalanguage the same conventions for = as we use in the object
language for
e
=.
(ii) For superposition of the rst kind the ordinal version is often more
convenient than the set theoretical one; namely, if we have a xed
set, say X, from which we choose the arities under consideration as
subsets, then it is usually not guaranteed that for L
m
X (m M)
and M X we will also have

M
L
m
X. On the other hand one
is usually interested in having for some regular bound on the arities
a representative system for all partial operations on a set. In this
connection the switch to ordinals can provide such a possibility: One
may choose a well-ordering for the given set X; then this well-ordering
will also induce a well-ordering on each of its subsets, and this can
be used for the superposition usually it is of no great importance
what specic set we have used as arity: we may dene two partial
operations, say f PO
M
(A; S) and g PO
N
(A; S) to be equivalent,
iff there is a sort respecting bijection, say i : N M, such that for
each a A
M
and a A one has f(a) = a iff g(a(i)) = a (i.e. i
and id
A
induce a sort preserving bijection between graph f and graph g:
(a, a) graph f iff (a(i), a) graph g).
(iii) If A is any partial algebra, then the ordinal superposition of the rst
kind induces a total algebraic structure, say (

, of type on
PO(A; ; S) (respectively on PO(A; X; S), when we well-order X and
relate to each subset of X its induced well-order type) in the following
way, when we well-order each K

( ), say by an ordinal

:
December 8, 2002
370 Chapter IV Special Topics
Let f

PO(A; ; S) be a

-ary partial operation on A for each <

;
then dene the

[ <

)-ary partial operation

( f

[ <

) :=
A
( f

[ <

).
Similarly we get from the superposition of the second kind a partial
algebraic structure, say ( )

, on PO(A; X; S), dened as follows:


Let ( f
k
[ k K

) PO
M
(A; S)
K
for some d
&
-small subset M of X
be sort compatible with K

, then:
( f
k
[ k K

) :=
A
( f
k
[ k K

)
observe the remark on empty mappings in 15.1.3.(i).
The corresponding total and partial algebras are denoted by PO(A; ; S)
(=: PO(A; (d
&
); S)) respectively by PO(A; X; S).
One should observe that, for each M X (with #M < d
&
) and for
each non-nullary ,
PO
M
(A;S)
:= |PO
M
(A) (| meaning that
the domain is restricted to (PO
M
(A; S))
K
, and the target is restricted
to PO
M
(A; S)), is a total operation on PO
M
(A; S). However, if K

= ,
then
PO
M
(A;S)

M
(
M
here denoting the empty sequence of M-ary
partial operations) is either a total M-ary operation with value
A
, if
this constant is dened in A, otherwise it is the empty M-ary operation
with the same output sort as .
Thus PO
M
(A; S) := (PO
M
(A; S), (
PO
M
(A;S)
)

) is a total algebra of
type , while the corresponding restriction of (

to PO
,v
(A; S)
( < (d
&
), v : S) is usually partial, in most cases even empty.
15.1.5 Remarks and Examples
(i) Observe that by the above denition PO
M
(A; S) is not just a relative
subalgebra of PO(A; X; S) (this will only be the case if we change the
denition for the nullary constants, the non-nullary partial operations
are just the restrictions of the corresponding global ones).
(ii) For the following examples let us assume that each arity is nite, say
n := 0, 1, . . . , n 1, and for simplicity let us consider the homoge-
neous case, where we do not mention S in the notations:
December 8, 2002
15 Algebraic operations 371
gure uncompleted
Figure 15.1: f15.1
gure uncompleted
Figure 15.2: f15.2
Each h PO
n
(A) can then be considered as a black box with n
entrances and one outlet (see Figure 15.1) such that for an ad-
missible sequence a = (a
0
, . . . , a
n1
) i.e. a domh, each component
given on the corresponding entrance yields h(a) at the outlet; else
there is no reaction at the outlet.
In the case of superposition of the rst kind, say for h (g
0
, g
1
, g
2
),
where, say, g
0
PO
3
(A), g
1
PO
2
(A) and g
2
PO
4
(A), we connect
the corresponding black boxes as shown in Figure 15.2, i.e. all the
entries of the operations g
0
, g
1
respectively g
2
are independent. It is
easily seen in this connection that it does not matter whether or not
the arity of h is empty.
In the case of superposition of the second kind the corresponding pic-
ture for the black boxes looks as in Figure 15.3, when we consider
h PO
3
(A) and g
0
, g
1
, g
2
PO
2
(A); i.e. rst the variables are copied
as often as needed (here 3 times) into the corresponding entrances of the
partial operations, and the outlets of the g
i
are led into the appropriate
entrances of h.
Obviously these are two rather extremal possibilities for the combi-
nation of partial operations. Using the unary total identity opera-
tion e
1
0
= id
A
(for which in the diagrams the corresponding box
is usually omitted) we can write e.g. h(g(x
0
, x
1
, x
2
), x
3
, x
4
) as [h
(g, e
1
0
, e
1
0
)](x
0
, . . . , x
4
), but we have diculties in representing e.g. h(g(x
0
, x
1
, x
0
), x
2
, x
1
)
by superposition of the rst kind.
gure uncompleted
Figure 15.3: f15.3
December 8, 2002
372 Chapter IV Special Topics
In both cases superposition of the second kind is applicable and yields
[h(g(e
5
0
, e
5
1
, e
5
2
), e
5
3
, e
5
4
)](x
0
, x
1
, x
2
, x
3
, x
4
), respectively [h(g(e
3
0
, e
3
1
, e
3
0
), e
3
2
, e
3
1
)](x
0
, x
1
, x
2
)
(the reader should draw the corresponding black box schemes).
This shows that the concept of superposition of the second kind is more
exible when we include for each arity the total projection operations
among the basic tools which usually consist of the fundamental opera-
tions on a partial algebra A.
(iii) It easily follows from Denition 15.1.3 that for any set M one has for
the corresponding total projections:
() If h PO
M
(A; S), then h = h ( e
M
m
[ m M ), but
() If g
m
PO
M
(A; S) (m M), then e
M
m
( g
n
[ n M ) = g
m
[
( domgn|nM )
.
Thus the total projections behave left neutral, but they are right
neutral only in a restricted sense.
We have already observed similar features in connection with condi-
tioned term operations which we had to introduce in 10.3 (cf. 10.3.2)
in order to be able to describe Mod F
P
-extendable epimorphisms. We
shall see below that there is indeed a close connection.
(iv) Observe that the total projection e
M
m
behaves on a partial algebra A in
the same way as the term operation (M : m)
A
, when M is a subset of
the set of variables. Since PO
M
(A; S) is a total algebra, the mapping
e : M E
M
(A; S), m e
M
m
, has a homomorphic extension eval
M
:
F(M, TAlg(S)) PO
M
(A; S) (if necessary, we write eval
A
M
).
15.2 Algebraic partial operations of the rst and sec-
ond kind
Lemma 15.2.1 Let A be any partial algebra, M a d
&
-small set of variables.
Then the homomorphism eval
M
: F(M, TAlg(S)) PO
M
(A; S), which has
been dened in 15.1.5.(iv), satises for each t F(M, TAlg(S)) : eval
M
(t) =
(M : t)
A
; i.e. eval
M
provides a link for the evaluation of terms as term
functions on a more abstract level.
The straightforward proof by algebraic induction starting with 15.1.5.(iv)
is left as an exercise to the reader.
December 8, 2002
15 Algebraic operations 373
This observation gives rise to the following denition, where we also treat
substitution of the rst kind.
Denition 15.2.2 Let A be any partial (S-) algebra.
(i) The elements of the subalgebra of PO(A; (d
&
); S) (with respect to su-
perposition of the rst kind) generated by the set id
As
[ s S of
total identity operations for the phyla of A are called algebraic (par-
tial) operations of the rst kind on A, and the whole set is denoted by
APO1(A; (d
&
); S) (= c
PO(A;;S)
id
As
[ s S ).
(ii) Let M be any set. Then
eval
M
[F(M, TAlg(S))] = c
PO
M
(A;S)
E
M
(A; S) =: APO
M
(A; S)
is called the set of all M-ary algebraic (partial) operations (of the second
kind) on A the additions in brackets are often suppressed.
For any set X with #X
s
d

for each s S we denote by APO2(A; X; S) :=

( APO
M
(A) [ M X, #M < d
&
) the set of all algebraic partial
operations of the second kind on A with arities in X (restricted in car-
dinality by d
&
).
By APO1(A; ; S), APO2(A; X; S) respectively APO
M
(A; S) we denote
the relative subalgebras of the corresponding partial algebras of oper-
ations.
Remark 15.2.3 Observe that APO2(A; X; S) is generated as a subalgebra
of PO(A; X; S) by G :=

( E
M
(A; S) C
M
(A; S) [ M X, #M < d
&
),
where C
M
(A; S) := c
M,A

O
M
(A; S) [ , K

= ,
A
is total , such
that c
M,A

(a) =
A
for each a A
M
.
Observe that C

(A; S) could be omitted from G, since it will be pro-


vided by the algebraic structure of PO(A; X; S). Within each PO
M
(A; S)
the subset APO
M
(A; S) is generated by the M-ary total projections only,
but for the global set APO(A; X; S) we also need the extended funda-
mental constants (i.e. extended to each M X, #M < d
&
) in order to get
everything that is induced by terms with d
&
-small subsets of X for variables.
The relationship between algebraic operations of the rst and of the sec-
ond kind will follow from the subsequent results.
December 8, 2002
374 Chapter IV Special Topics
Lemma 15.2.4 Superposition of the second kind is associative. I.e. let
A Alg(S), f PO
M
(A; S), g
m
PO
N
(A; S) for each m M and h
n

PO
L
(A; S) for each n N. Then [f ( g
m
[ m M )] ( h
n
[ n N ) =
f (g
m
( h
n
[ n N ) [ m M ).
Proof The statement is easily derived from Denition 15.1.3.
Denition 15.2.5 Let A, M and N be any sets, p : M N any mapping,
and let h PO
M
(A; S). Then
h
p,N
:= h ( e
N
p(m)
[ m M )
is an N-ary operation on A obtained from h by a transformation of variables
along p.
Proposition 15.2.6 Let A Alg(S), M, N sets, p : M N a mapping.
Then p induces a homomorphism p

: PO
M
(A; S) PO
N
(A; S), h h
p,N
(h PO
M
(A; S)), whose restriction to APO
M
(A; S) is a homomorphism
p

: APO
M
(A; S) APO
N
(A; S) between the corresponding total algebras of
algebraic operations on A.
Proof From 15.2.5 it immediately follows that p

: PO
M
(A; S) PO
N
(A; S)
is a mapping. Let , h
k
[ k K

PO
M
(A; S), then: p

(
PO
M
(A;S)
( h
k
[
k K

)) = p

(
A
( h
k
[ k K

)) = [
A
( h
k
[ k K

)] ( e
N
p(m)
[ m
M ) =
A
( h
k
( e
N
p(m)
[ m M ) [ k K

) (using 15.2.4) =
A
(p

h
k
[ k
K

) =
PO
N
(A;S)
(p

h
k
[ k K

).
Thus p

: PO
M
(A; S) PO
N
(A; S) is a homomorphism.
Now assume that h = e
M
m
; then h
p,N
= e
M
m
( e
N
p(m

)
[ m

M ), and this implies


that (e
M
m
)
p,N
= p

(e
M
m
) = e
N
p(m)
. The compatibility of homomorphisms with
the algebraic closure operator now implies p

[APO
M
(A; S)] APO
N
(A; S),
and the last statement follows.
Proposition 15.2.7 For every partial algebra A each algebraic partial oper-
ation on A of the rst kind is equivalent to some algebraic partial operation
of the second kind, but in general not vice versa.
December 8, 2002
15 Algebraic operations 375
Proof (by algebraic induction) For simplicity we assume that X and each
K

are ordinals see 15.1.4.(ii): Obviously id


As
= e
1
0
APO2(A; X; S),
when the only element 0 of 1 is mapped to s (s S). Therefore assume
that h =
A
( h
k
[ k K

), and that each h


k
is equivalent to

h
k

APO
L
k
,v
k
(A; S) (k K

). Let N := ( L
k
[ k K

) with the induced sort


mapping v
N
. Because of 15.2.6 g :=
A
(

h
k
( e
N

( L

[ < k ) <

( L

[ k ) ) [ k K

) APO
N,v
N
(A; S); and it is easy to check that
for each a A
N
and b A:

A
[( h
k
[ k K

)(a) = b iff g(a) = b, i.e. again


A
( h
k
[ k K

) is
equivalent to some algebraic partial operation of the second kind.
In order to see that the converse implication does not hold, one may con-
sider a binary fundamental operation symbol, say ( = (2)), and A :=
F(x, y, TAlg()) homogeneous case! , h :=
A
(e
1
0
, e
1
0
) APO
1
(A),
i.e. h(x) :=
A
(x, x) (x A). But when we consider the algebraic (partial)
operations of the rst kind on A, then the only unary one is id
A
, the only
binary one is
A
, etc.; however we do not get h, which is unary, but not the
identity.
The following consequence corresponds to 15.2.4 and yields another gen-
eralization of the associativity of the composition of unary functions (observe
that in the case of arbitrary sets one has to include a bijection between the
arities):
Lemma 15.2.8 Let M, N
m
, L
n
(m M, n

( N
m
[ m M )) be
arbitrary ordinals. Let h PO
M
(A; S), g
m
PO
Mm
(A; S) (m M), f
n

PO
Ln
(A; S) (n

( N
m
[ m M ). Then [h ( g
m
[ m M )]( f
n
[ n

( N
m
[ m M ) = h ( g
m
(f
n
[

( N
l
[ l < m) n <

( N
l
[ l m) [
m M ).
The connection between algebraic partial operations of the rst and sec-
ond kind is described by the following two results:
Lemma 15.2.9 Let A be any partial algebra, M any set. Then every ele-
ment of APO
M
(A; S) can be derived from some element of APO1(A; ; S) by
transformation of variables.
Proof (by algebraic induction)
December 8, 2002
376 Chapter IV Special Topics
(a) We assume a xed well-ordering for each K

, . For each m M,
e
M
m
= id
As
(e
M
m
) where s := v
M
(m) p : 1 M, 0 m.
(b) Let
A
be a nullary fundamental operation of A, which is dened in A.
Then the empty mapping p : M yields the total M-ary constant
operation
A
p,M
with value
A
of APO
M
(A; S).
(c) Assume h APO
M
(A; S), h =
A
( h
k
[ k K

), such that for


each h
k
there exists g
k
APO1(A; ; S) of arity
k
(and sort mapping v
k
)
with p
k
:
k
M such that h
k
= p

k
(g
k
). We then dene for :=

(
k
[ k K

) a mapping p : M by p() := p
k
(

), where

k
:=

[ < k ) =:
k

<
k+1
, i.e.

is uniquely determined
by ordinal arithmetic. And we obtain for every a A
M
and b A
that h(a) = [
A
( h
k
[ k K

)](a) =
A
( (g
k
)(a)
p
k
,M
[ k K

) =
(
A
( g
k
[ k K

)(a))
p,M
(the reader should ll in missing details as
an exercise).
Proposition 15.2.10 Let X = (d
&
) and each K

( ) be ordinals
together with their sort mappings such that each phylum of X has cardinality
at least d

. Then APO2(A; X; S) is the closure of APO1(A; ; S) with respect


to transformation of variables.
Remark 15.2.11 In this connection one has to distinguish between the fact
that X has to be well-ordered in order that each subset carries a well-ordering
and that we may have for a given family of subsets a unique well-ordering
on the sum of their order types. Since the sort mapping of this sum is given
by the disjoint union of the sort mappings of the summands, we cannot just
take such a sum as the initial segment of X, but we have to take the subset
with the smallest elements of X having the corresponding well-ordering type
and the same sort mapping as the ordinal obtained as sum. This can always
be obtained by the properties of well-orderings, and we shall not make this
more explicit here.
Proof (of 15.2.10) From 15.2.6 we infer that APO2(A; X; S) is always
closed with respect to transformation of variables, and from 15.2.7 we then
obtain that all partial operations with d
&
-small arity in X obtained from ele-
ments of APO1(A; ; S) by transformation of variables belong to APO2(A; X; S).
The converse inclusion follows from 15.2.9.
December 8, 2002
15 Algebraic operations 377
15.2.12 In the nitary case one can say more about the basic principles
(i.e. operations on PO(A; (d
&
); S)), by which one can generate the alge-
braic partial operations of the second kind from those of the rst kind see
e.g. W.M.Gluschkow, G.J.Zeitlin, J.L.Justschenko [GlZJu80], p. 73f,
and also H.-J.Hoehnke [Hoe72].
Lemma 15.2.13 APO1(A; ; S) and APO2(A; X; S) are closed with respect
to superposition of the rst kind, where only algebraic partial operations of
the rst kind are allowed on the outer part: Let h APO1(A; ; S) with
arity, say , and let ( g

[ < ) be a sequence of algebraic partial operations


of A of the rst kind (respectively of the second kind). Then h( g
x
i [ < )
is an algebraic partial operation of the rst (second) kind. And only if A is
total, APO2(A; X; S) = AO2(A; X; S) is closed with respect to superposition
of the second kind with respect to arbitrary elements of AO2(A; X; S).
Proof (by algebraic induction on APO1(A; ; S)) If h = id
As
, then the
statement is trivial. Let h =
A
( f
k
[ k K

) for some and


f
k
[ k K

APO1(A; X; S) of arity (assuming that X and all arities


are ordinals or at least well-ordered; moreover, we only treat the second kind,
the other case is similar). Assume that, for each k K

, h
k
:= f
k
( g

[ < k ) <

[ k ) ) APO2(A; X; S), and let


k
be the
corresponding arity; then the mappings p
k
:
k
,

[ < k ) +
( <
k
, k K

) yield algebraic partial operations (h


k
)
p
k
,
, and we get
h ( g

[ < ) =
A
( h
k
[ k K

) =
A
( (h
k
)
p
k
,
[ k K

), showing
that we obtain indeed an algebraic partial operation of the second kind.
The statement about total algebras follows from the fact that in this case the
projection operations are also right neutral 15.1.5.(iii).() does no longer
yield any restriction on the domain which is now all of A
M
.
Remarks 15.2.14 (i) It is easily seen that APO1(A; ; S) is not closed
with respect to superposition of the second kind: Let A be the free
semigroup (type (2)) on an at least two-element free generating set M.
Then we get for the binary operation : (x, y) ((x, y), (x, y)) =
[ (, )](x, y) is an algebraic operation of the second, but not of the
rst kind.
(ii) APO2(A; X; S) is not closed with respect to superposition of the second
kind: Let := (2, 1, 0), := (, h, n), and let F := F(x, y, K) be
December 8, 2002
378 Chapter IV Special Topics
the K-free partial algebra on x, y in the E-variety K dened by the
E-equations (xy)
e
= (xy) and h(n)
e
= n, i.e. h
F
has a one-element
domain, while n
F
and
F
are total. Consider g := e
2
0
(e
2
0
, h
F
e
2
1
);
then domg = F n
F
. We claim that g is not an algebraic partial
operation of the second kind: g is a binary partial projection onto
the rst component. We scan the rst Baire classes of APO
2
(F) with
respect to E
2
(F):
0
th
class: e
2
0
, e
2
1
are total projections.
1
st
class: h
F
e
2
0
, h
F
e
2
1
, n
F

2
(
2
the empty sequence of binary
algebraic partial operations on F),
F
(e
2
0
, e
2
1
) =
F
,
F
(e
2
1
, e
2
0
),

F
(e
2
0
, e
2
0
),
F
(e
2
1
, e
2
1
).
2
nd
class: h
F
(
F
(e
2
0
, e
2
1
)), h
F
(
F
(e
2
1
, e
2
0
)), h
F
(
F
(e
2
0
, e
2
0
)),
h
F
(
F
(e
2
1
, e
2
1
)), and operations ending with
F
at their right
end.
g does not coincide with any of the above operations, and it can be seen
that the Baire classes of higher rank cannot contain g either. Therefore,
although the projection operations are algebraic, it is in general not
allowed within the scope of algebraic partial operations of the second
kind to substitute algebraic partial operations for their variables.
(iii) Let us just mention without proof that for A Alg(S), h APO
1
(A; S)
and g APO(A; X; S) one always has h g APO(A; X; S).
15.3 Algebraic partial operations of the third kind
15.3.1 In the theory of homogeneous total algebras so-called clones of
(total!) operations play an important role, i.e. for a given set A and a set X of
variables one considers subsets C of OM(A; X) which are closed with respect
to superposition (of the second kind): if h
l
[ l L C
M
:= h C [
h has arity M and f C
L
, then f ( h
l
[ l L) C. And it is well known
that for any total algebra A the set APO2(A; X) of all algebraic operations
of A is a clone (see 15.2.13), and if a clone for A contains all M-ary projections
for all d

-small subsets of X, then it is the clone of all algebraic operations


(of the second kind) on a suitable total (homogeneous) algebraic structure
on A.
December 8, 2002
15 Algebraic operations 379
The observations at the end of subsection 15.2 show that the same is
not true for partial algebras. We have already realized in other situations
that the term operations (which are by denition the same as the algebraic
operations of the second kind) are not sucient to describe everything in
connection with partial algebras, and we had to introduce conditioned term
operations. We shall see that these conditioned term operations which
also arise by category theoretical investigations, see A.Obtulowicz [O81]
form clones, namely those generated by the algebraic partial operations of
the second kind with respect to superposition (of the second kind).
In order to get a better feeling for this third kind of algebraic partial
operations we start with a seemingly weaker denition and show that we
indeed get clones.
Denition 15.3.2 Let A be any partial algebra and M any set. Then we de-
ne the set APO3
M
(A; S) of all M-ary algebraic partial operations of the third
kind to be the subset of PO
M
(A; S) which is generated by APO
M
(A; S) with
respect to superposition of the second kind, where we allow at the top only
algebraic partial operations of the second kind, i.e. we take APO2(A; X; S)
(for the usual set X of variables) as set of fundamental operations of a par-
tial algebra A

(and dene S

in a similar way) and dene APO3


M
(A; S) as
APO2
M
(A

; S

).
APO3(A; X; S) :=

( APO3
M
(A; S) [ M X, #M < d
&
) designates
the set of all algebraic partial operations of the third kind with d
&
-small sets
of variables in X.
By APO3
M
(A; S) respectively APO3(A; X; S) we denote the relative sub-
algebras of PO
M
(A; S) respectively PO(A; X; S).
Since all fundamental operations
A
are algebraic partial operations of
the second kind (
A
=
A
( e
K
k
[ k K

)), we get
Lemma 15.3.3 For each A Alg(S) and for each set M, APO3
M
(A; S) is
a (total) subalgebra of (the total algebra) PO
M
(A; S).
The following characterization of algebraic partial operations of the third
kind provides more insight into their nature and relates them to conditioned
terms:
Proposition 15.3.4 Let A be any partial algebra, M any set. Then h
APO3
M
(A; S) iff there exist a d
&
-small set L, an l

L and a sequence
f := ( f
l
[ l L) (APO
M
(A; S))
L
such that h = e
L
l
f.
December 8, 2002
380 Chapter IV Special Topics
Proof (by twofold algebraic induction) If already h APO
M
(A; S), then
we may choose L = l

= 1, f = (h) : h = e
1
0
(h). Thus assume that
h = h

( g
n
[ n N ), where h

APO
N
(A; S) (N X, d
&
-small), and that
for each g
n
we already have a representation g
n
=: e
Ln
l

n
( g
nl
[ l L
n
) of the
desired kind. If h

= e
N
n
, then we dene L :=

( n L
n
[ n N ) and get
h = e
N
n
(e
Ln
l

( g
nl
[ l L
n
) [ n N ) = e
L
(n

,l

n
)
( g
nl
[ (n, l) L), i.e. h has
the desired representation.
Next assume that h

=
A
( f
k
[ k K

) and that f
k
( g
n
[ n N ) =:
e
L
k
l

k
( f
kl
[ l L
k
) =: f

k
is of the desired form for each k K

. Then we
have for each a A
M
and for each b A that h(a) = b iff a

( domf
kl
[
l L
k
, k K

), b := ( f

k
(a) [ k K

) dom
A
and b =
A
(b), iff
a

( domf
kl
[ l L
k
, k K

), b

:= ( f

kl

k
(a) [ k K

) dom
A
and
b =
A
(b); however, h

:=
A
( f
kl

k
[ k K

) APO
M
(A; S) because
of all the intermediate assumptions. Therefore we may dene for some new
index l

: L := l

( kL
k
[ k K

) and f
l
:= h

, f
(k,l)
:= f
kl
(k K

,
l L
k
). Then h = e
L
l
( f
l
[ l L) is of the desired form.
Conversely, if h is of the above form, then it follows from the denition
of APO3(A; X; S) that h belongs to this set.
Remark 15.3.5 15.3.4 shows that the graph of each algebraic partial op-
eration of the third kind is a subset of the graph of an algebraic partial
operation of the second kind (i.e. term operation), while its domain is just
the intersection of the domains of algebraic operations of the second kind
(i.e. term operations). Since each algebraic partial operation of the second
kind is a term operation, we get for algebraic partial operations of the third
kind the following representation:
Let h APO3
M
(A; S), i.e. h = e
L
l
( f
l
[ l L); since each f
l
is of
the second kind, we have f
l
= (M : t
l
)
A
for some t
l
F(M, TAlg(S)), and
therefore h = ( M : t
l
[ t
l
[ l L )
A
is a conditioned term operation.
Conversely it is easily seen that each conditioned term operation, where the
condition is d
&
-small, yields an algebraic partial operation of the third kind.
Thus we get:
Corollary 15.3.6 For each partial algebra A and the usual set X of variables
the sets APO3(A; X; S) of all algebraic partial operations of the third kind
and the set of all conditioned term operations on A, for which the restricting
set of terms is d
&
-small, are identical.
December 8, 2002
15 Algebraic operations 381
Proposition 15.3.7 For each partial algebra A and each set M the sets
APO3
M
(A; S) and APO3(A; X; S) are closed with respect to superposition of
the second kind.
Proof Using 15.3.4, let g = e
N
n
( f
n
[ n N ) APO3
L
(A; S), and,
for each l L, g
l
= e
I
l
i

l
( f
li
[ i I
l
) APO3
M
(A; S). Then for each
a A
M
and b A we have (g ( g
l
[ l L)(a) = b iff a

( domf
l

i
, f
n

dom( f
li

l
[ l L) [ l

L, i I
l
, n N ) and f
n
( f
li

l
(a) [ l L) = b;
and an argument similar to the one at the end of the proof of 15.3.4 shows
that g ( g
l
[ l L)APO
M
(A; S).
Remark 15.3.8 This shows that indeed the algebraic partial operations of
the third kind (and this means: the d
&
-small conditioned term operations)
form a clone of partial operations. We cannot go into more details here,
however we refer in particular to the quoted papers of A.Obtulowicz.
Remark 15.3.9 We have seen that the set of algebraic partial operations of
the second kind is not closed with respect to superposition; and if we want
to apply arbitrary superposition (of the second kind), we are led to alge-
braic partial operations of the third kind which can be considered as special
domain restricted algebraic partial operations of the second kind. How-
ever, when we remember that each algebraic partial operation of the second
kind stems from some term, we may avoid this extension of APO2(A; X; S)
to APO3(A; X; S) in connection with superposition: if we have to superpose
algebraic partial operations of the second kind as term operations! , we
superpose the associated terms and then consider the resulting term oper-
ation on A the graph of the resulting algebraic partial operation (of the
second kind) then contains the graph of the algebraic partial operation of
the third kind which we would get by the usual superposition. One should
only observe that this requires the choice of a specic term for each involved
algebraic partial operation, but as subsection 15.5 will show dierent
choices will lead to the same results.
We mention this eect, since it has an analogue in some programming lan-
guages which allow to pass the names of functions as parameters to another
function or subroutine (procedure). We have detected this possibility e.g. in
some implementations of the programming language Pascal, for instance
in Pascal MT+ (from Digital Research for personal computers). The
short programs TestVal and TestParm below show this eect:
December 8, 2002
382 Chapter IV Special Topics
PROGRAM TestVal (INPUT,OUTPUT);
VAR Argument : REAL;
FUNCTION Square (Arg1 : REAL) : REAL;
BEGIN (* Square *)
Square := Arg1 * Arg1
END; (* Square *)
FUNCTION Inverse (Arg2 : REAL) : REAL;
BEGIN (* Inverse *)
Inverse := 1 / Arg2
END; (* Inverse *)
FUNCTION Test1 (V1, V2 : REAL) : REAL;
BEGIN (* Test1 *)
Test1 := V1
END; (* Test1 *)
BEGIN (* TestVal *)
WRITELN(Enter any real number: );
READLN(Argument);
WRITELN;
WRITELN(TestVal(,Argument,) =,Test1(Square(Argument),
Inverse(Argument)))
END. (* TestVal *)
PROGRAM TestPar(INPUT,OUTPUT);
VAR Argument:REAL;
FUNCTION Square(Arg1:REAL):REAL;
BEGIN (* Square *)
Square:= Arg1 * Arg1;
END; (* Square *)
FUNCTION Inverse(Arg2:REAL):REAL;
December 8, 2002
15 Algebraic operations 383
BEGIN (* Inverse *)
Inverse:= 1 / Arg2;
END; (* Inverse *)
FUNCTION Test2(FUNCTION F1(X:REAL):REAL;
FUNCTION F2(U:REAL):REAL):REAL;
BEGIN (* Test2 *)
Test2:= F1(Argument);
END; (* Test2 *)
BEGIN (* TestPar *)
WRITELN(Enter any real number: );
READLN(Argument);
WRITELN;
WRITELN(TestPar(,Argument,) = ,Test2(Square,Inverse));
END. (* TestPar *)
TestVal as well as TestPar are unary algebraic partial operations on
R := (REALS; Square, Inverse), where REALS is the set of real numbers
accepted by the Pacsal compiler under consideration. The algebraic partial
operation of the third kind under consideration is e
2
0
(Square, Inverse), i.e.
substitution into a projection.
It is evaluated by TestVal (i.e. by Test1) in the usual way of super-
position of the second kind. Thus, entering 0 (zero) in TestVal yields an
error message, since division by zero is detected in evaluating Inverse(0).
TestPar corresponds to the formal evaluation using the terms, therefore
TestPar yields for the argument 0 the value Square(0), i.e. 0. For all argu-
ments dierent from 0 TestVal and TestPar yield the same value.
Usually the implementations of the two kinds of passage of variables as
discussed above have so-called side-eects which we do not discuss here.
We would like to mention that similar eects as those described in connec-
tion with TestPar occur in programs, which dene some function by using
IF. . . THEN. . . ELSE. . . in connection with the values of partial functions.
December 8, 2002
384 Chapter IV Special Topics
15.4 Algebraic total operations
Before we study in some more detail the algebras of algebraic total M-ary
operations of a given partial algebra A let us recall from 5.1.4 that a subset M
of a partial algebra A is called B-independent (= B-independent), if each
mapping from M into B can be extended to a homomorphism from C
A
M
into B. Closely related to results in subsection 6.2 is the following one which
we shall need below:
Lemma 15.4.1 Let A, B be similar partial algebras, M a B-independent
subset of A and N M. If one of the following three conditions is satised,
then N is B-independent in A, too:
(i) There exists an (S-) mapping from M into B.
(ii) There is no (S-) mapping from N into B.
(iii) v
A
[N] = v
A
[M].
Proof If (ii) is satised, then the statement is trivial. If (i) or (iii) is
satised, and if f : N B is any mapping, then f can be extended to a
mapping g : M B, g
N
= f. If g : C
A
M B extends g, then its restriction
to C
A
N is the desired homomorphic extension of f.
15.4.2 It is easy to realize that the conditions in Lemma 15.4.1 are necessary
as long as empty phyla are allowed; in particular, in the homogeneous case
the conditions of 15.4.1 can only be violated when N = B = , but M ,= .
Another lemma which we shall need below is the following one:
Lemma 15.4.3 Let A be any partial algebra, B a (relatively) normal (see
5.3.2 and 9.2.7) relative subalgebra of A, and let M B be a generating
subset of A. Then M also generates B.
Proof Let C := c
B
M, then C B. We prove B C by algebraic
induction on A:
Obviously M B C. Assume that b =
A
(a) B, and that a(k) B
implies a(k) C for each k K

. Since B is relatively normal in A, and since


b B, we have a B
K
; thus a C
K
, and nally b =
A
(a) =
B
(a) C
by the denition of C.
December 8, 2002
15 Algebraic operations 385
We now return to operations on A:
Lemma 15.4.4 For any partial algebra A and for any set M the relative
subalgebra of PO
M
(A; S) of all total M-ary operations on A, i.e. O
M
(A; S),
carries the induced structure of the power D := A
(A
M
)
; and it is a normal
relative subalgebra of PO
M
(A; S).
Proof Let , f := ( f
k
[ k K

) D
K
, g D; then
D
(f) = g iff
for every a A
M
: (
D
(f))(a) = g(a), iff
A
( f
k
(a) [ k K

) = g(a), i.e. iff


g =
A
f. If g
k
[ k K

PO
M
(A; S), then dom[
A
( g
k
[ k K

)]

( domg
k
[ k K

). i.e. if at least one g


k
is not total, then
A
( g
k
[ k K

)
cannot be total. This shows that O
M
(A; S) is normal in PO
M
(A; S).
Denition 15.4.5 For every set M and for every partial algebra A we de-
note by
F
M
(A; S) := APO
M
(A; S) O
M
(A; S)
the set of all algebraic total M-ary operations on A (sometimes also called
in the literature the set of full M-ary algebraic operations), and F
M
(A; S)
designates the relative subalgebra of APO
M
(A; S) on F
M
(A; S).
Proposition 15.4.6 Let A be any partial algebra, M any set. Then F
M
(A; S)
is a normal relative subalgebra of APO
M
(A; S) freely and A-freely generated
by E
M
(A; S). Moreover, F
M
(A; S) is a subalgebra of O
M
(A; S).
Proof Obviously E
M
(A; S) F
M
(A; S). That F
M
(A; S) is normal in APO
M
(A; S)
is shown in the same way as the corresponding result in 15.4.4. Now, 15.4.3
implies that E
M
(A; S) generates F
M
(A; S), and from 15.4.4 and 15.4.5 we in-
fer that F
M
(A; S) has to be a closed subset of O
M
(A; S). Thus F
M
(A; S)
H
w
o
c
T /, and therefore 5.1.6.(ii) implies that it suces to show that
E
M
(A; S) is A-independent in F
M
(A; S):
Let f : E
M
(A; S) A be any mapping; then let f

: M A, m f(e
M
m
)
(m M), consider the projection homomorphism pr
f
: A
(A
M
)
A, and let

f := pr
f
[
F
M
(A;S)
. We claim that

f[
E
M
(A;S)
= f:
If g A
(A
M
)
then pr
f
(g) = g(f

), and pr
f
(e
M
m
) = e
M
m
(f

) = f

(m) = f(e
M
m
).
December 8, 2002
386 Chapter IV Special Topics
Remark 15.4.7 Let A be any partial algebra, and let M be any set which
does not allow any mapping into A this is equivalent to the fact that
v
M
[M] , v
A
[A]. Then all M-ary projections on A and therefore also all
algebraic partial or total operations will have empty domain and can only be
distinguished by their output sort. In particular all elements of APO
M
(A; S)
are (many-sorted) total. All phyla of E
M
(A; S) and of APO
M
(A; S) have at
most one element.
It should also be observed that for those s S for which A has a one-
element phylum and M
s
is non-empty (but no further restriction on M), the
corresponding phylum of E
M
(A; S) (respectively, of F
M
(A; S)) also consists
of one element only, since we only have constant mappings onto this phylum
of A.
In both situations the mapping m e
M
m
(m M) is injective only in the
case when each phylum of M has at most one element.
Proposition 15.4.8 For every partial algebra A and for any sets M and N
one has:
(i) F
M
(A; S) o
c
T /, F
M
(A; S)

= F(M, o
c
T /).
(ii) If one of the following conditions is satised, then F
M
(F
N
(A; S); S)

=
F
M
(A; S):
(a) M N and v
M
[M] = v
N
[N].
(b) M N and v
N
[N] v
A
[A].
(c) H
w
o
c
T F
N
(/; o) = H
w
o
c
T /.
Proof The proof of (i) is contained in the one of 15.4.6. Therefore it remains
only to prove (ii):
Let us assume (a) or (b): If there is no mapping from M into A, then there is
also no mapping from N into A, and the statement easily follows from 15.4.7.
Thus let us assume that we have at least one mapping from N into A in
case (b) this is always guaranteed , and then we also have at least one
mapping from M into A as well as into F
N
(A; S).
First let us observe that E
M
(F
N
(A; S); S) is H
w
o
c
T F
N
(/; o)-independent
in F
M
(F
N
(A; S); S). Therefore the mapping f : E
M
(F
N
(A; S); S) F
N
(A; S),
e
M
m
(F
N
(A; S)) e
N
m
(A) (m M) observe that we have to distinguish the
December 8, 2002
15 Algebraic operations 387
projections on dierent sets and that our assumptions (by 15.4.7) indeed yield
a mapping extends to a homomorphism

f : F
M
(F
N
(A; S); S) F
N
(A; S).
Since H
w
o
c
T F
N
(/; o) H
w
o
c
T /, the mapping g : E
N
(A; S)
E
M
(F
N
(A; S); S), dened by g(e
N
n
(A)) := e
M
n
(F
N
(A; S); S), if n M, and
g(e
N
n
(A)) := e
M
m
(F
N
(A; S); S) for some xed m

M, if n / M, extends
to a homomorphism g : F
N
(A; S) F
M
(F
N
(A; S); S). Obviously g f is
the identity mapping on E
M
(F
N
(A; S); S), and therefore g

f is the identical
isomorphism of F
M
(F
N
(A; S); S). This shows (see 11.4) that

f is a section
and therefore closed and injective. Since we have a mapping from N
and therefore from E
N
(A; S) into A, the assumptions of 15.4.1 are satis-
ed for e
N
m
(A; S) [ m M E
N
(A; S), i.e. e
N
m
(A; S) [ m M is A-
independent in F
N
(A; S), and therefore F
M
(F
N
(A; S); S)

= C
F
N
(A;S)
e
N
m
(A; S) [
m M

= F(M, o
c
T /)

= F
M
(A; S).
Finally assume that (c) is true: Then obviously ind
F
M
(F
N
(A;S);S)
E
M
(F
N
(A; S); S)
H
w
o
c
T /. Since F
M
(F
N
(A; S); S) belongs to the E-variety H
w
o
c
T /,
we again get F
M
(F
N
(A; S); S)

= F(M, H
w
o
c
T /)

= F
M
(A; S).
The following result will be useful in the next subsection:
Lemma 15.4.9 Let A be any partial algebra, M, N any sets.
(i) If g := ( g
n
[ n N ) (F
M
(A; S))
N
and f APO
N
(A; S), then
f g APO
M
(A; S).
(ii) Each endomorphism of F
M
(A; S) is extendable to an endomorphism
of APO
M
(A; S).
Proof Ad (i) We prove (i) by algebraic induction on f: If f = e
N
n
, then
f g = g
n
(since totality of all the g
n
does not impose any domain
restriction). Thus, assume that f =
A
( f
k
[ k K

) and that f
k
g
APO
N
(A; S) for all k K

. Then it is obvious that also fg =


A
( f
k
g [
k K

) belongs to APO
M
(A; S).
Ad (ii) We dene for f End F
M
(A; S) and h APO
M
(A; S):

f(h) :=
h ( f(e
M
m
) [ m M ). By (i) it is easily realized that

f is indeed a
mapping

f : APO
M
(A; S) APO
M
(A; S). We leave the details of a
proof that

f is an endomorphism of APO
M
(A; S) as an exercise.
December 8, 2002
388 Chapter IV Special Topics
For further results on free partial algebras, e.g. on F
M
(A; S), see sec-
tion 17.
15.5 More about algebraic partial operations and terms;
algebraic partial functions
15.5.1 We already know by Denition 15.2.2 that each M-ary algebraic
partial operation, say h, of the second kind of some partial algebra A is the
homomorphic value with respect to eval
M
: F(M, TAlg(S)) PO
M
(A; S)
induced by mapping m M onto the m-th M-ary projection e
M
m
(A) of
some term t
h
with variables in M: h = (M : t
h
)
A
. And in 15.3.7 we have seen
that each M-ary algebraic partial operation of the third kind corresponds in
a similar way to a conditioned term operation. Now, in general, eval
M
is
not injective, and here we briey want to investigate its kernel it should be
noticed that eval
M
heavily depends on the chosen partial algebra A, although
we do not always indicate it in the notation.
Proposition 15.5.2 Let A be any partial algebra, M a set, t, t

F(M, TAlg(S)).
Then (M : t)
A
= (M : t

)
A
(i.e. (t, t

) ker eval
M
) iff A [= ((M : t
e
= t t

e
=
t

t
e
= t

) iff A [= ((M : (t
e
= t t
e
= t

) (t

e
= t

t
e
= t

)).
15.5.3 Observe that the axioms in 15.5.2 correspond to the concept of
strong validity of equations in partial algebras as mentioned in 8.1.3.(ii).(b).
Proof If (M : t)
A
= (M : t

)
A
, then, for all a A
M
and b A, (M :
t)
A
(a) = b iff (M : t

)
A
(a) = b, but this just means that in A the above
formula is valid. Conversely it is as easily seen that for t, t

F(M, TAlg(S))
validity in A of the above formula implies (M : t)
A
= (M : t

)
A
.
15.5.4 Problems We do not know which congruences of F(M, TAlg(S))
can occur as kernels of canonical homomorphisms eval
A
M
onto algebras of
algebraic partial operations. The interested reader will nd some information
in this direction in J.Slominski [Sl68] in connection with the concept of
validity. Our Lemma 15.4.9 provides an idea that not every congruence
on F(M, TAlg(S)) can occur as kernel of some eval
A
M
, since every mapping
f : M M induces an endomorphism

f : F(M, TAlg(S)) F(M, TAlg(S))
which has to satisfy f
(2)
[] , but it need not be fully invariant either
(cf. subsection 6.3).
December 8, 2002
15 Algebraic operations 389
We also do not know whether it has already been investigated which other
properties the set of all kernels of evaluation morphisms eval
A
M
for xed M
and arbitrary A has. Is it closed with respect to intersections? Is it d

-
algebraic? etc.
Some of the results in the next subsection will also provide some more
information about these questions.
Proposition 15.5.5 Let A be any partial algebra, M A, a c
A
M. Then
there exists h APO
M
(A; S) such that a = h(id
M
) i.e. every a c
A
M
is the value of a suitable M-ary algebraic partial operation applied to the
identity sequence of M.
Proof i := id
M
: M M extends to a closed homomorphism i

:
domi

A (out of F(M, TAlg(S))) onto c


A
M. Therefore there is t
domi

such that i

(t) = (M : t)
A
(i) = (eval
A
M
(t))(id
M
), eval
A
M
(t) APO
M
(A; S).
In the same way one proves
Corollary 15.5.6 1 For each sequence a A
M
and for each element b of
the closed subset c
A
(a[M]) generated by the image of a there is an M-ary
algebraic partial operation h APO
M
(A; S) with value b at a: b = h(a).
Corollary 15.5.7 2 Each mapping f : E
M
(A; S) A has a closed ho-
momorphic extension out of APO
M
(A; S): f

: domf

A. In particular,
if g : M A is dened by g m := f(e
M
m
), then domf

= eval
A
M
(domg

),
and domf

is an E
M
(A; S)-generated relative subalgebra of APO
M
(A; S).
Moreover we have, for each h APO
M
(A; S): h domf

iff g domh.
Proof We have g dom(M : t)
A
iff t domg

, iff g dom(eval
A
M
(t)).
From these equivalences all the statements are easily derived.
15.5.8 In the notation of Poythress [Po73] 15.5.6 says that E
M
(A; S) is p-
free for A in the notation of Kerkhoff [Ke70] E
M
(A; S) is called conform
free for A in APO
M
(A; S).
Finally we would like to mention another concept closely related to the
one of algebraic partial operations: the one of algebraic partial functions or
polynomial functions:
December 8, 2002
390 Chapter IV Special Topics
Denition 15.5.9 Let M be any set, A any partial algebra. An M-ary par-
tial operation p PO
M
(A; S) is called an M-ary algebraic (partial) function
(also M-ary polynomial function) on A, if there exist a set N containing M
and an algebraic partial operation h, such that p is obtained from h by re-
placing each argument belonging to some element of N M by some xed
element of A, i.e. iff there exists a mapping b : (N M) A such that for
each a A
M
and for ab denoting the sequence whose graph is obtained as
the disjoint union of the graphs of a and b p(a) := h(ab).
15.5.10 The set of all algebraic partial functions on a partial algebra A
derivable from algebraic partial operations with d
&
-small arities in X can
be considered as the closure of the set of all total projections with d
&
-small
arities in X with respect to the structure on PO(A
A
; X; S
S
), where A
A
:=
(A, (
A
)

(
a
)
aA
), and each
a
is a (non-empty) nullary constant with
value a.
This shows how all the results on algebraic partial operations can be
extended to algebraic partial functions. This concept is useful in particu-
lar in connection with the investigation of congruence relations (see subsec-
tion 19.1). They also arise in connection with the evaluation of global partial
polynomials in A, as can easily be realized.
15.6 Algebraic partial operations and homomorphisms
15.6.1 From 5.7.5 we know that for each M-ary term (M : t) and each
homomorphism f : A B one has that (M : t)
A
(a) = b implies (M :
t)
B
(f(a)) = f(b) for each a A
M
and b A, i.e. that homomorphisms are
compatible with term operations. Thus, if (M : t)
A
= h APO
M
(A; S),
(M : t)
B
= h

APO
M
(B; S), then h(a) = b implies h

(f(a)) = f(b) for


each a A
M
, b A. But this does not mean that f induces in general
a homomorphism, say f
#
, from APO
M
(A; S) into APO
M
(B; S), since this
would be equivalent with the fact:
()
ker eval
A
M
ker eval
B
M
, i.e. for any terms t, t

F(M, TAlg(S)):
(M : t)
A
= (M : t

)
A
(M : t)
B
= (M : t

)
B
.
But in general this cannot be expected when f is not surjective and not full.
However, even when f is full and surjective this need not be the case as the
homomorphism f in Figure 11.1 in 11.6.1 shows: f is full and surjective, but
APO
1
(A) = e
1
0
,
A
, , APO
1
(B) = e
1
0
,
B
,
B

B
, , and f cannot induce
December 8, 2002
15 Algebraic operations 391
a homomorphism from APO
1
(A) into APO
1
(B), since the identity would go
onto the identity, and f
#
would have to be surjective as a homomorphism
between total algebras.
However we have the following
Proposition 15.6.2 Let f : A B be a closed and surjective homomor-
phism. Then, for each set M, f induces a homomorphism f
#
: APO
M
(A; S)
APO
M
(B; S) such that, for every t F(M, TAlg(S)), (M : t)
B
= f
#
((M :
t)
A
).
Proof We prove 15.6.1.(): Assume t, t

F(M, TAlg(S)) such that (M :


t)
A
= (M : t

)
A
. Let b dom(M : t)
B
be arbitrary. Since f is surjective and
closed, there is a dom(M : t)
A
such that f(a) = b (Axiom of Choice). (M :
t)
A
= (M : t

)
A
now implies that (M : t)
A
(a) = (M : t

)
A
(a), and therefore
(M : t)
B
(b) = (M : t)
B
(f(a)) = f((M : t)
A
(a)) = f((M : t

)
A
(a)) = (M :
t

)
B
(f(a)) = (M : t

)
B
(b).
Symmetry of the assumptions implies (M : t)
B
= (M : t

)
B
. Thus 15.6.1.()
is satised, and the Diagram Completion Lemma implies the existence of
f
#
: APO
M
(A; S) APO
M
(B; S) having the required properties.
In a similar way one proves a corresponding result for algebraic total
operations, which needs weaker conditions on f, since we do not have to
bother with denedness:
Proposition 15.6.3 Let M be any set and f : A B a surjective homo-
morphism. Then f induces a homomorphism f
&
: F
M
(A; S) F
M
(B; S)
such that, for every t (eval
A
M
)
1
[F
M
(A; S)], (M : t)
B
= f
&
((M : t)
A
.
15.6.4 Notice that surjectivity of f is necessary in 15.6.2 as well as in 15.6.3,
even in the case of total algebras, since algebraic partial operations which
coincide on f[A] may dier outside of this subset of B (example?).
December 8, 2002
392 Chapter IV Special Topics
16 Semilattices of (weak) completions
Although we have claimed in the introduction that we want to treat partial
algebras in their own right, it cannot be denied that the topic of completions
of partial algebras or questions about their weak, full or even normal embed-
dability into a total algebra of some prespecied kind are very important, in
particular in computer science, but also in mathematics in general. If, for
instance, in computer science a program is specied which uses partial func-
tions i.e. functions which are not everywhere dened or which are badly
dened e.g. because of the possibility of ending in a loop which they cannot
leave then anyway the program should stop after some time in one way
or another, either yielding a so-called error value or by some implemented
or enforced time limit (meaning that the program is interrupted from the
outside after some time). Each such ending of the program could be con-
nected with some new value, thus completing the operation dened by the
program, if this is done for all possibilities. In this way the implementation
yields a total operation, although it was originally partial with respect to its
basic meaning. A neat treatment of the program should distinguish these
dierent kinds of values. These dierent kinds should not be mistaken for
dierent sorts as occurring in many-sorted algebras - these dierent sorts
just hide some kind of partiality, and a good compiler will tell by an error
message, when one has used at some place a wrong sort (often called data
type, e.g. in the programming language Pascal) of variables, thus com-
pleting this many-sorted operation to a one-sorted one. In connection with
a many-sorted partial operation (given by some program, or often specied
in connection with some data type) also the error-values will have a specic
sort (if there are no error-messages in connection with the use of wrong
sorts (data types)). For the treatment of error-values in connection with the
specication of abstract data types by means of many-sorted (total) alge-
bras see the papers of the ADJ-group concerning this topic, in particular
[GoTcWa78] or [Go78].
In pure mathematics we refer to the famous paper of T. Evans [Ev51]
relating in connection with a given variety, say K, of total algebras the solv-
ability of the (full) embeddability problem for a special class of partial al-
gebras to the solution of the word problem for K. We do not want to go
into details about this here. Another area where the completion of partial
algebras plays an important role is the question of whether the universal so-
lution of a so-called partial lattice within the variety of all lattices is nite
December 8, 2002
16 Semilattices of (weak) completions 393
or innite, and related topics in lattice theory. Another instance is given by
the constructions of direct sums (coproducts, often called free products)
within given equational classes of total algebras. Here we want to mention
e.g. lattice theory (see e.g. Gr atzer [G78]). Also the questions in strong
amalgamation theory are actually of the kind concerning whether special
partial algebras constructed from the amalgams can be fully embedded into
total algebras of the given variety (but we cannot treat this within the scope
of this book, either). This shows that also within the theory of total algebras
the completion of partial algebras plays an important role.
In section 17 we shall also mainly study completions, but this time free
completions within E-varieties.
16.1 The semilattice of (weak) minimal completions
The material of this subsection is basically contained in [BSch67].
Denition 16.1.1 (i) Let A be any partial algebra; a total (S-) algebra B
is called a weak completion of A, if A is a weak subalgebra of B, a
completion, if A is a relative subalgebra; and a normal completion, if
A is normal in B (see 5.3.2). If A generates its (weak) completion B,
then B is called a minimal (weak or normal) completion of A.
(ii) Let B and C be minimal weak completions of A such that id
A
induces
a (necessarily surjective) homomorphism from B onto C; then we say
that B is greater than C and denote this by
B ~
A
C or B ~ C,
assuming that the reference to A is obvious.
Remarks 16.1.2 (i) Notice that we want this denition to be understood
in such a way that the . . . completion of any partial S-algebra is again
an S-algebra and therefore in general not homogeneous. But what we
shall say in the following will also apply to homogeneous completions
of partial S-algebras, when we just forget about S.
(ii) A minimal weak completion B of A is normal iff id
A
: A B is
relatively normal in the sense of section 9 (see 9.2.7).
December 8, 2002
394 Chapter IV Special Topics
(iii) It is easy to realize that ~ is a quasi-order on the class of all minimal
weak completions of any given partial algebra A, and if there holds
B ~ C as well as C ~ B, then obviously B and C are isomorphic
over id
A
, and we say that they are equivalent and denote this by B C.
It is well known that any quasi-order factored with respect to the in-
duced equivalence relation yields a partial order. This also holds for
classes, when one chooses a representative system, and such a repre-
sentative system is easily provided in our case:
Let A := F(A, TAlg(S)) be the free completion of A. Then we know
from section 5 that A is a normal minimal completion of A, and that
every homomorphism from A into some total algebra has a homomor-
phic extension to A. Thus every minimal weak completion, say B,
of A corresponds to a unique congruence, say
B
on A.
As an immediate consequence of the Homomorphism Theorem we get the
Proposition 16.1.3 For any two weak completions B and C of a partial
algebra A one has:
(i) B ~ C iff
B

C
.
(ii) B C iff
B
=
C
.
Notice that the quasi-order between weak minimal completions of A is
just dual to the inclusion between the corresponding congruences.
Theorem 16.1.4 Let be an arbitrary congruence on A. Then:
(i) =
B
for some weak minimal completion B of A iff (A A) =
id
A
.
(ii) =
B
for some minimal completion B of A iff
(A (A
_

A
[A
K
] [ )) = id
A
.
16.1.5 16.1.4.(ii) can be interpreted as follows: If g(a) = g(x), a A, and
x A


A
[A
K
] [ = D
A
A (where a and x are both of the same
sort), then x = a, where g is a homomorphism with kernel .
December 8, 2002
16 Semilattices of (weak) completions 395
Proof (of Theorem 16.1.4)
Ad (i) B is a minimal weak completion of A iff it contains A as a generating
weak subalgebra, i.e. iff the induced restriction to A of the corresponding
homomorphism, say g, from A onto B is injective, and this is equivalent
to
g
(A A) = id
A
.
Ad (ii) First, consider =
B
, where B is a minimal completion of A.
Assume a A, x A


A
[A
K
] [ , and g(a) = g(x), where
g := g
B
is the induced surjection from A onto B. If x A, then a = x
by (i), since B is also a weak minimal completion of A. If x =
A
(a),
where a(k) A for all k K

, then we have
a = g(a) = g(x) =
B
(g(a)) =
B
(a),
hence a =
A
(a), since A is a relative subalgebra of B. Therefore
a =
A
(a) = x, since A is a relative subalgebra of A.
Conversely, let the congruence satisfy the condition of 16.1.4.(ii).
Then there is a total algebra B

with a homomorphism g

from A
onto B

, which induces (as its kernel). By hypothesis, the restric-


tion h of g

to A is injective, and we can construct a total algebra B


which contains A at least as a weak relative subalgebra, and which is
isomorphic to B

, say via h

, induced by h : h

h = id
A
(cf. Appendix,
subsection 23.2).
Then id
A
is a homomorphism from A into B, and we claim that A is
even a relative subalgebra of B: In fact, assume that a =
B
(a) for
some a A
K
and some , and consider x :=
A
(a) A. Then
(h

)(a) = a =
B
(a) =
B
((h

)(a)) = (h

)(x).
Since ker h

= , this implies by hypothesis a = x or


A
(a) = a A.
By (FC1) (cf. 5.3.1) we obtain a =
A
(a). Thus B is a completion of A,
and even a minimal one, since A generates A and the homomorphism
h

is surjective, carrying A onto A. Moreover, = ker h

=
B
,
completing the proof.
Remarks 16.1.6 (i) Theorem 16.1.4 provides a comparatively concrete
survey on all weak minimal respectively on all minimal completions of
a given partial algebra A. Notice that the set of all inner completions
is characterized with respect to 16.1.4.(i) by the additional property
December 8, 2002
396 Chapter IV Special Topics
that A has a non-empty (and hence a one-element) intersection with
each congruence class of the corresponding congruence relation of A.
(This is obvious, since the corresponding surjective homomorphism is
onto A.)
(ii) By replacing each (weak) minimal completion B of A by its corre-
sponding factor structure A/
B
we do not get a weak completion but
a structure which is canonically isomorphic to it. Thus a representa-
tive system for all weak minimal completions of A forms a set, which
bijectively corresponds to the set of all congruence relations on A
satisfying (AA) = id
A
. And this set also contains a representative
from each isomorphism class of minimal completions of A.
Moreover, let (B
j
)
jJ
be any non-empty family of representatives of iso-
morphism classes of weak minimal completions of A. Then this family
has a supremum in the poset introduced above, namely the isomor-
phism class of minimal completions corresponding to :=


B
j
[
j J , which obviously also satises the condition in 16.1.4.(i). No-
tice that the corresponding weak completion is canonically isomorphic
to the subalgebra generated by the image of A in the direct product of
the family ( B
j
[ j J ).
(iii) Instead of dealing with a poset of isomorphism types of weak minimal
completions of A we can switch, because of the preceding results, to the
poset of congruence relations of A satisfying the condition in 16.1.4.(i)
or in 16.1.4.(ii) in the case of minimal completions. Since the inter-
section of a non-empty family of such congruence relations still satises
these conditions, they form in each case a conditionally complete meet-
semilattice, which has in general in each case no greatest element
but many maximal elements: In the case of weak minimal completions
each inner completion obviously corresponds to a maximal congruence
relation, i.e. inner completions are minimal elements in the set of iso-
morphism types of weak minimal completions. In the case of minimal
completions (with full embeddings) one has to add to A at least one
new element to each phylum which belongs to the output sort of a re-
ally partial operation in order to get a full embedding of A into such a
minimal completion, and all sequences in A
K
dom
A
( ) have
then to be mapped onto the corresponding new element. Let B be
the corresponding new carrier set. There is now no condition for the
December 8, 2002
16 Semilattices of (weak) completions 397
values of fundamental operations on B for sequences which contain at
least one of the new elements except that they have to belong to the
appropriate phylum, and it is easily realized that any choice will then
yield in the end a minimal element in the poset of isomorphism types
of minimal completions of A, since the corresponding congruence rela-
tion cannot be extended without violating condition 16.1.4.(ii) (proof
as an exercise!). Thus in general we also have in this subsemilattice of
minimal completions more than one minimal element (see [BSch67] for
an example that a minimal element of this subsemilattice of minimal
completions of even a homogeneous partial algebra A can have more
than one additional element).
We thus get the following
Proposition 16.1.7 The representative weak minimal completions respec-
tively minimal completions of a partial algebra A constitute in each case a
conditionally complete join semilattice with greatest element A. In general
this is not a lattice.
16.2 The complete lattice of normal minimal comple-
tions
16.2.1 (The (heterogeneous) normal one-point completion) The sit-
uation becomes much nicer, when we consider normal minimal completions
(see 5.3.2). Let us rst observe that in this case the ordered set of repre-
sentatives of minimal normal completions of a given partial algebra, say A,
has (up to isomorphism) exactly one minimal element, namely the so-called
(heterogeneous) normal one-point completion, which we shall denote by A
o
:
For each s S we add a new element, say
s
, to A
s
:= v
1
A
(s) (such that

s
/ A
s
and v
A
(
s
) := s for each s S). Moreover we dene for each sort
respecting and for each a (A

)
K
:

(a) :=
_

A
(a), if a dom
A
,

s
, else, where s is the output sort of .
A
o
is then the subalgebra of A

generated by A (i.e. new elements are added


to some phylum only when they are absolutely necessary for the completion).
December 8, 2002
398 Chapter IV Special Topics
It is easy to realize that A
o
is in fact a normal minimal completion of A,
and that it corresponds to the congruence relation
:=
A

:= id
A

_
((A)
s
A
s
)
2
[ s S
(the normality of A within A guarantees that is in fact a congruence
relation). Any of the two descriptions of A
o
above shows that A
o
is minimal
(a homomorphic image over id
A
) with respect to all other normal minimal
completions of A.
Next we consider a criterion similar to those in 16.1.4:
Theorem 16.2.2 A minimal completion B of A is normal iff the corre-
sponding congruence relation
B
of A satises

B
(A A) = id
A
.
16.2.3 Since id
A
is the restriction of g : A B to A, this condition
(strengthening the one of Theorem 16.1.4) means that g(x) A implies
(and is implied by) x A, for all x A (and implies g(x) = x). This
observation is helpful for the
Proof (of Theorem 16.2.2) Let B be a normal minimal completion of A.
We prove the implication of 16.2.3 by algebraic induction on x: The inductive
beginning for x A is trivial. Induction hypothesis: g(x
k
) A implies
x
k
A, for all k K

, for some fundamental operation symbol , and for


some family x := ( x
k
[ k K

) of elements x
k
A. We have to prove
our implication for x :=
A
(x). Thus assume g(x) A, i.e.
B
( g(x
k
) [ k
K

) A. Since B is a normal completion of A, this implies g(x


k
) A.
Thus even
B
(g(x)) =
A
(g(x)), and by induction hypothesis: x
k
A, i.e.
g(x
k
) = x
k
for all k K

. Thus
A
(x) is dened, and since A is a relative
subalgebra of A, x =
A
(x) =
A
(x) A.
Conversely, let a minimal completion B of A fulll the condition of Theorem
16.2.2, and assume
B
(b) A (b = ( b
k
[ k K

)). By the Axiom of Choice


there are elements x
k
A such that b
k
= g(x
k
), for all k K

, and we have
g(
A
( x
k
[ k K

)) =
B
(b) A. By assumption
A
( x
k
[ k K

) A;
by (FC1) (normality of A in A), we get b
k
= g(x
k
) = x
k
A, for all k K

,
completing the proof.
December 8, 2002
16 Semilattices of (weak) completions 399
16.2.4 Since the relation
B
A A above is reexive and symmetric,
the condition of Theorem 16.2.2 is equivalent with

B
id
A
((A A) (A A));
or,as
id
A
((A A) (A A)) =
A
o
,
our condition in Theorem 16.2.2 means that
B

A
o
. This again makes
evident (as we already noted above) that the (heterogeneous) normal one-
point completion A
o
of A is a normal extension of A and minimal with respect
to all other such completions.
Moreover we have the
Corollary 16.2.5 (1) A minimal completion B of A is normal iff B ~ A
o
.
It is even more important that now, when we have a minimal element,
we have the following counterpart to 16.1.7:
Corollary 16.2.6 (2) The representative normal minimal completions of
any partial algebra A constitute a complete lattice with greatest element A
and least element A
o
.
Remarks 16.2.7 (i) The heterogeneous one-point completion plays an
important role in computer science, when the algebraic specication of
an abstract data type yields a partial algebra which has to be (heteroge-
neously) completed in a most simple way. Here the heterogeneous one-
point completion is the easiest way and produces the minimal possible
number of so called error values, as the new elements are then called.
In any way one usually wants a normal minimal completion of the data
type specied. A detailed discussion of the completion of heteroge-
neous partial algebras by error-values can be found in [GoTcWa78]
and in [Go78] (notice that Proposition 7 in [GoTcWa78] is more or less
vacuous, since it only applies in the case when the algebra, which is to
be extended, is already total). It should however be mentioned that
the concept of error algebras as introduced in [Go78] by Goguen is
not so easily subsumed by the concept of a free partial algebra in our
sense, which is then normally completed by some error values, since
one of the main conditions in [Go78] is that a homomorphism between
error algebras has to carry error elements into error elements.
December 8, 2002
400 Chapter IV Special Topics
(ii) We would like to mention as a sort of natural completion of a partial
algebra the one using an algebra which is often called the global of a
given partial algebra (this construction is also used for total algebras):
Recall that Subs A designates the disjoint union of all power sets of
phyla of A; then we equip this set with a total (heterogeneous) structure
as follows:
For every and for every sequence M := ( M
k
[ k K

)
(Subs A)
K
we set:

Subs A
(M) :=
A
( m
k
[ k K

) [ m
k
M
k
for each k K

.
It is easy to realize that the set of all one-element subsets in Subs A
forms a relative subalgebra of Subs A which is canonically isomorphic
to A, and the subalgebra generated by this set is canonically isomorphic
to the normal heterogeneous one- point completion A
o
of A.
Denition 16.2.8 Let (M : t
e
= t

) be any E-equation. It is called regular,


if var t = var t

, i.e. if both terms contain the same variables.


Proposition 16.2.9 Let G be any set of regular E-equations, and A
Mod G be any model of G. Then the heterogeneous one-point completion A
o
of A is also a model of G.
Proof If A is total, then A
o
= A, and the statement is trivial. Let e :=
(M : t
e
= t

) be any E-equation in G, and let u : M A


o
be any assignment.
Let N := var t = var t

. Then we have to distinguish two cases:


(i) u[N] A: Then e is already satised in A with respect to u.
(ii) u[N] , A: Then there is at least one new element contained in u[N].
Therefore the normality of the completion implies that both (M : t)
A
o
(u)
and (M : t

)
A
o
(u) are new elements not contained in A (proof by alge-
braic induction), and they have to be the same new element (since t
and t

are of the same sort). Thus e is also satised in A


o
with respect
to u in this case.
December 8, 2002
17 Independence, free partial algebras and primitive classes 401
17 Independence, free partial algebras and
primitive classes
In earlier sections (e.g. in sections 5, 6 and 15) independence and partial
algebras freely generated by some set or even by some other partial al-
gebra have been discussed to some extent as well as their relationship to
primitive classes. In this section we want to continue this discussion and
add further results. At the beginning we want to discuss some inheritance
properties of independence, followed by an investigation of the free comple-
tions within primitive classes as well as by a characterization of free partial
algebras freely generated by some set a similar characterization for free
extensions of partial algebras within some kind of variety (in particular of
universal solutions) is missing in general and might be an interesting problem
(however see 17.6.4). Free partial algebras or more generally maximal free
extensions of partial algebras play an important role in computer science
in connection with the initial algebraic specication of abstract data types.
Thus this will also be discussed in Part II in some more specic context. As
in most sections before we treat arbitrary partial S-algebras for some given
nonempty homogeneous partial algebra S of the given similarity type without
usually mentioning it explicitly.
17.1 Some properties of independence and indepen-
dence classes
17.1.1 We recall from 5.1.4 that a subset M of some partial algebra A is B-
independent in A for some partial algebra B, iff every assignment w : M B
has a homomorphic extension w

: E := c
A
M B satisfying v
B
w

= v
E
.
Recall, too, that ind
A
M designates the class of all partial algebras D for
which M is D-independent in A, and that it is called the independence
class of M in A. In addition we know (cf. 5.1.6) that every independence
class is primitive, i.e. it is closed with respect to (direct) products, (closed)
subalgebras and (weak) homomorphic images.
Lemma 17.1.2 Let M be any subset of some partial algebra A, and let B
be any partial algebra. Then the following statements are equivalent:
(i) M is B-independent in A.
December 8, 2002
402 Chapter IV Special Topics
(ii) Every subset N of M of cardinality less than d
&
, and which satises
v
A
[N] = v
A
[M], is B-independent in A.
If v
A
[M] v
B
[B], then the condition v
A
[N] = v
A
[M] is not necessary
in (ii). This applies in particular, when B = A, i.e. when we deal with a
so-called algebraically independent subset M of A.
Proof The implication from (i) to (ii) is trivial, since the assumptions of (ii)
imply that every assignment for N into B can be extended to an assignment
of M into B. The converse is an immediate consequence of 4.4.9.
If v
A
[M] v
B
[B], then there exists at least one assignment from M into B,
and even in the case that v
A
[N] = v
A
[M] is violated, any assignment from N
into B can be extended to one starting from M; hence the equivalence of (i)
and (ii) can also be proved.
Remark 17.1.3 Notice that in general the condition v
A
[N] = v
A
[M] is
essential even when we are in the homogeneous case, since otherwise it might
happen that there does not exist any assignment from M into B but that
it exists for its subset N in the homogeneous case this might only happen
when B and N are void and M is nonvoid.
Let us recall from ?? that in the homogeneous case the rank r of
a given similarity type is dened as r := sup( n [ n < d ). Moreover recall
from sections 5, 6, and 15 that, for every quasi-primitive class K, F(M, K)
designates the free partial K-algebra, K-freely generated by r
M,K
[M], where
r
M,K
: M F(M, K) is the K-universal homomorphism for M, and that
in general r
M,K
is injective iff there exists at least one injective assignment
w : M B from M into some K-algebra B.
Proposition 17.1.4 Let K be any primitive class of partial algebras. Then
the following statements are true:
(i) In the case of homogeneous partial algebras it is true for every set M
of cardinality #M r + 1 that ind
F(M,K)
M = K , or even more
precisely K = ind
F(M,K)
M ind
F(,K)
.
(ii) In the case of arbitrary partial S-algebras let M be such that, for ev-
ery s S, #v
1
M
(s) r + 1, and for every subset Z S of S
let M
Z
:= v
1
M
[Z] be the preimage of Z under v
M
. Then we have
K =

ind
F(M
Z
,K)
M
Z
[ Z S .
December 8, 2002
17 Independence, free partial algebras and primitive classes 403
Proof Since M is always nonempty by assumption, the empty partial alge-
bra always trivially belongs to the independence class of M, and whether or
not the empty partial algebra belongs to K is decided by ind
F(,K)
, which
always contains K, but which contains iff F(, K) is itself empty. This ob-
servation easily shows that (i) is an immediate consequence of (ii), which as
a matter of fact is a consequence of Theorem 6.4.2.(iii) in combination with
the fact that the K-universal homomorphism r
M,K
: M F(M, K) represents
the set of all E-equations with variables in M which are valid in K. Since
each E-equation needs from each phylum of M less than d + 1 and hence at
most r +1 elements, the statement can easily be derived from 6.4.2.(iii). The
details are left as an exercise.
17.2 Relatively free completions of partial algebras
Remark 17.2.1 In this subsection we continue the study of completions of
partial algebras started in section 16; but this time we restrict considerations
to completions within a given E-variety. Although we are especially inter-
ested in the relation between K-freely generated K-algebras for some given
E-variety K and their free completions within K, part of this can be dis-
cussed in a more general context, and we shall do this in this subsection. In
the next one we shall then consider the other problem, and in a later sub-
section we shall even realize that there is in some respect no great dierence
between the free completion if it exists of a K-free partial algebra and the
one of an arbitrary partial algebra.
Actually, we shall see that the situation is relatively simple in the homo-
geneous case, which has already been represented in [B73], but in the het-
erogeneous case, things become relatively complicated, since, in some way,
E-varieties of heterogeneous partial algebras are from the homogeneous
model theoretic point of view already sort of implicational classes (one
introduces a fundamental relation for each sort, and the heterogeneous E-
equations (Y : t
e
= t

) can then be considered as an implication with a


conjunction in the premise saying that for each non-empty sort of Y there
exists an element satisfying the corresponding relation, and with t
e
= t

as
conclusion). Actually we shall observe already for E-varieties of heteroge-
neous partial algebras phenomena which appear in the homogeneous case at
the earliest in connection with ECE-equations. But since for ECE-equations
the description of cases of full embeddability into the relatively free com-
pletion is even more complicated, we do not treat this situation within the
December 8, 2002
404 Chapter IV Special Topics
scope of this book. Nevertheless the results of this subsection can provide a
slight idea of what can happen in this more general case.
As a preparation of our results we need some notation:
Denition 17.2.2 (i) Let K be any class of partial algebras, then we
designate by K
t
the class of all total K-algebras.
(ii) Let A be any partial algebra; then we designate by

A
the disjoint
union of the set of all complements of domains of fundamental oper-
ations of A (i.e. the set of all undened sequences of A observe
many-sortedness!):
_
A
:=
_
(A
()
dom
A
) [ .
(iii) Let A and B be similar partial algebras. Then B is called an inner
extension of A, iff A = B and A is a weak subalgebra of B. If, in
particular, B is total, then B is called an inner completion of A.
It is easy to realize for the case, when B is an inner extension of A, that
there exists a mapping z : domz A such that domz

A
, and for every
one has
graph
B
= graph
A
(a, z(, a)) [ (, a) domz .
In connection with this mapping we also call B a z-extension of A (and we
denote B by A
z
), and if B is total, then we speak of a z-completion.
As a preparation for our embeddability results we prove
Lemma 17.2.3 Let K be any class of similar partial algebras closed with
respect to subalgebras and homomorphic images, let A be any partial algebra
of the same type, and let r
A,K
be the K-universal epimorphism for A (which
exists according to subsection 5.11), then we have
(i) If A allows an injective homomorphism into some K-algebra, then r
A,K
is injective, too.
(ii) If A allows a full and injective homomorphism into some K-algebra,
then r
A,K
is full and injective, too.
Proof Since injectivity as well as the property to be full and injective (i.e.
initial and injective), can be described by the reection of formulas, which are
preserved by homomorphisms (cf. section 9, e.g. 9.2.15), this is an immediate
consequence of Lemma 9.1.7.(ii).
December 8, 2002
17 Independence, free partial algebras and primitive classes 405
Proposition 17.2.4 Let K be any E-variety, let A be any K-algebra, and
consider the following conditions:
() A is empty.
() To A and F(A, K
t
) there belong the same subsets of S of sorts for which
these partial algebras have nonempty phyla, i.e. v
A
[A] = v
F(A,Kt)
[F(A, K
t
)].
() K has a set, say (, of dening E-equations such that for any (Y : t
e
=
t

) ( there exists an assignment from Y into A.


() For those elements of S for which A has a non-empty phylum, F(A, K
t
)
has an at least two-element phylum.
Then we have the following statements:
(i) K
t
is again an E-variety, each inner extension of A belongs to K, and
if () holds, then A allows inner completions, and these belong to K
t
.
(ii) () holds iff v
A
[A] is a closed subset of S.
(iii) If A satises one of the conditions (), (), or (), then r
A,Kt
is
injective.
(iv) If A satises one of the conditions (), (() and ()) or (() and
()), then r
A,Kt
is injective and full.
(v) If K is denable by so-called regular E-equations (cf. 16.2.8), then
F(A, K
t
) is always a normal minimal completion of A, i.e. r
A,Kt
is
injective and relatively normal (and in particular full).
Proof Ad (i) Since for a given E-variety K homomorphic images, subalge-
bras and direct products of total K-algebras are again total, we imme-
diately get the rst part of the statement. If := (Y : t
e
= t

) Eeq K,
and if B is an inner extension of A, then Y allows an assignment into B
iff it allows one into A, and since is valid in A it is valid in B.
If () is satised, then all fundamental operations whose input sorts
are non-empty in A have also output sorts which are non-empty in A,
and therefore A allows inner completions, and they belong to K and
therefore to K
t
.
December 8, 2002
406 Chapter IV Special Topics
Ad (ii) Since v
F(A,Kt)
is closed, v
F(A,Kt)
[F(A, K
t
)] is always a closed subset
of S, and the statement easily follows.
Ad (iii) If A is empty, then the statement is trivial. Thus assume that
A is non-empty. If () is satised, then the statement is implied by
(i) and 17.2.3.(i). Finally assume that () holds, but that () is
violated: Then we construct a total K-algebra B containing A as a weak
subalgebra:
Let
A
:= [ the elements of graph
S
are contained in v
A
[A] ,
and let B := A

b
s
[ s Sv
A
[A] such that v
B
(b) :=
_
v
A
(b) if b A,
s if b = b
s
.
Dene B to contain A as relative subalgebra and having no additional
structure (
B
:=
A
for every ). Then () implies that B and
each inner completion of B belong to K. Therefore the statement follows
from 17.2.3.(i).
Ad (iv) If A = , then the statement is trivial. Therefore we may assume
that A ,= . Let us abbreviate F := F(A, K
t
). Consider g : F(F, K)
F to be the homomorphism induced by id
F
, and let (

be the
quotient structure on F with respect to g joined with the structure of A
(observe that A can be considered as a subset of F, because of the
injectivity of r
A,Kt
which follows from (i) or (iii) and our assumptions).
Obviously F

:= (F, (

) belongs to K as an inner extension of


some K-algebra. Let A
s
and F
s
be the phyla of sort s of A and F,
respectively, and let S
A
and S
F
be the subsets of S of sorts, for which
the corresponding phyla in A and F, respectively, are non-empty. For
each s S
F
we now choose a
s
, b
s
F
s
such that a
s
,= b
s
whenever
s S
A
, and we dene on F the following two structures:

a
(c) :=
_

(c), if c dom

,
a
s
, when s is the output sort of
and
b
(c) :=
_

(c), if c dom

,
b
s
, when s is the output sort of
for every and for every c F
()
. Dene F
c
:= (F, (
c
)

)
for c a, b, let G := F
a
F
b
, and let i : A G be dened by
i(a) := (a, a) for each a A. Then it is easy to realize from the above
construction of G that i is a full and injective homomorphism, and that
G is a K
t
-algebra. This proves (iv).
December 8, 2002
17 Independence, free partial algebras and primitive classes 407
Ad (v) This statement is a consequence of 16.2.9, since A

will belong to K
t
.
Why we cannot state much more in the heterogeneous case is shown in
the following
Example 17.2.5 Since in the many-sorted case E-equations are also pro-
vided with a set of variables, their inuence also depends on the evaluability
of this set, and the situation of completions within E-varieties can become
rather complicated for heterogeneous partial algebras. Consider:
:= (,

), := (1, 1, 1), S := s, s

,
S
,

S
: s s,

S
: s s

Moreover, let x, y be distinct variables of sort s, x

, y

be distinct variables of
sort s

, respectively, and let


g :=(x : (x)
e
= (x)),
g

:=(x, x

: (x)
e
=

(x)),
g

:=(x, y, x

: x
e
= y),
g

:=(x, x

, y

: x

e
= y

),
g

:=(x :

(x)
e
=

(x)).
Let A be the following partial S-algebra:
A
s
:=
0
= 0, 1, 2, 3, . . . , , graph
A
:= (i, i + 1) [ i
0
,
A
s
:= , graph
A
:= graph
A
:= .
Then it is easy to realize that A belongs to each of the following E-varieties:
K
1
:=Mod g
K
2
:=Mod g, g

K
3
:=Mod g, g

(omit

from the type!)


K
4
:=Mod g, g

, g

K
5
:=Mod g, g

, g

(omit

from the type!)


K
6
:=Mod g

K
7
:=Mod g, g

, g

since even A = F(0, K


i
) for 1 i 5 (0 having sort s), as only g has some
inuence on A, but the other three E-equations cannot be evaluated in A.
December 8, 2002
408 Chapter IV Special Topics
On the other hand the phylum corresponding to s

has to be non-empty in
each F(A, (K
i
)
t
), and therefore now all the E-equations under consideration
are really evaluable, yielding the following properties of r
A,(K
i
)t
:
r
A,(K
1
)t
is full and injective,
r
A,(K
2
)t
is injective, but not full,
r
A,(K
3
)t
and r
A,(K
5
)t
are full, but not injective, (

being omitted!)
r
A,(K
4
)t
is neither full, nor injective.
We have also considered K
5
, since in this case each phylum of each total
K
5
-algebra has at most one element, i.e. (K
5
)
t
is a trivial E-variety, although
K
5
is non-trivial.
Finally, let B := F(0, K
6
) (= 0
d
)), and C := F(0, K
7
) (0 having
sort s). Then C is a partial Peano algebra in which
C
and

C
are total and

C
is empty, and C also belongs to K
6
, B is a relative subalgebra of C as
well as of F(B, (K
6
)
t
), and F(B, (K
6
)
t
) and F(C, (K
6
)
t
) are isomorphic (the
isomorphism being induced by id
B
), but r
C,(K
6
)t
is not injective (in this case
it is full, but a similar example, only with dierent E-varieties, would show
that it need not even be full under the circumstances above).
For ECE-varieties K, where K
t
is also always an E-variety (!), one gets
similar eects even in the homogeneous case by taking as axioms besides g
ECE-equations with

(x)
e
=

(x) as premise and one of the E-equations g

or g

(or g

) as conclusion.
This example shows that we need in the heterogeneous case relatively
strong restrictions on the E-varieties K (or, equivalently, on their dening E-
equations) in order to get injective and/or full homomorphisms r
A,K
. 17.2.4
is the best we can do at the moment in general. But in the homogeneous
case we have the
Corollary 17.2.6 Let K be any E-variety of homogeneous partial algebras,
and let A be any K-algebra. Then r
A,Kt
is always injective, and it will only
happen that r
A,Kt
is not full, when both A and F(A, K
t
) contain exactly one
element and A is not total.
Remark 17.2.7 Notice that the last part of the example above shows that
for E-varieties K of heterogeneous partial algebras a K-algebra need not have
a completion within K, not even a weak completion in the sense of section 16
on completions in general. In the homogeneous case there exists at least a
December 8, 2002
17 Independence, free partial algebras and primitive classes 409
weak completion, and if F(A, K
t
) has at least two elements, then A also has
a completion within K. As we have indicated in 17.2.1, heterogeneous E-
varieties already have features which one observes in the homogeneous case
only for ECE-varieties. That the results above are not so far from the study
of K-free partial K-algebras, K-freely generated by some set, follows from the
results in the subsection on initial partial algebras below. In this subsection
here we now draw the special conclusions from the above results for K-free
partial K-algebras on sets. We restrict the results to criteria for r
F(M,K),Kt
to
be full and injective the case of injectivity alone is handled similarly.
Corollary 17.2.8 Let K be any E-variety, and let M be any set such that
F(M, K) satises one of the following conditions:
(+) If a phylum of F(M, K
t
) is non-empty, then it contains at least two
elements, and the corresponding phylum of F(M, K) is non-empty, too.
(++) K = ind
F(M,K)
M, and F(M, K
t
) has an at least two-element phylum for
each s v
F(M,K)
[F(M, K)].
Then r
F(M,K),Kt
is full and injective.
Proof Notice that (+) implies () and () of 17.2.4; thus the statement
follows from 17.2.4.
If K is the independence class of M with respect to F(M, K), then it is easily
seen that 17.2.4.() has to be satised (cf. section 6), while the second part
of (++) is just 17.2.4.(); hence r
F(M,K),Kt
is full and injective.
17.3 Relatively free completions of relatively free par-
tial algebras
In order to consider the structure of F(M, K
t
) more intensively we present
Denition 17.3.1. For the homogeneous case the following material can be
found in [B73].
Denition 17.3.1 (i) Let S be any sort-algebra, and let S

be any subset
of S. Then the relative structure of S on S

denes a similarity type


December 8, 2002
410 Chapter IV Special Topics
of many-sorted partial algebras, which is called the S

-reduction of the
given one. It satises
(S

) := [ , and each input and output sort of belongs to S

;
(S

) := (())
(S

)
;

:=
S
for each (S

).
If A is any partial algebra, and if we forget about all phyla of A of sorts
not in S

and about all fundamental operations of A not in (S

), then
we call this partial algebra the S

-reduct of A and denote it by A[


S
.
Notice that this can also be dened for arbitrary weak subalgebras S

of S, only (S

) and (S

) have then to be modied into the correspond-


ing (S

) and (S

), where now only those operations are considered,


which are non-empty in S

.
If K is any class of partial S-algebras and S

any weak subalgebra of S,


then we denote by K(S

t) the class of all partial K-algebras for which


each fundamental operation corresponding to some (S

) is total,
and we call it the class of all S

-total K-algebras. If S

is the relative
subalgebra of S on S

, then we also use the notation K(S

t) instead
of K(S

t).
(ii) Let B := A
z
be a z-extension of a given partial algebra A(cf. 17.2.2.(iii)).
Then we dene on B a mixed step operator G
A,z
as follows:
G
A,z
M := c
A
D
A
z
M
for all M A = B. Moreover, we dene as usual
G
0
A,z
M := M, and G

A,z
M := G
A,z
(
_
G

A,z
M [ < )
for all ordinal numbers > 0.
Let = () be the ordinal dimension of the given similarity type
(see 1.3.1), then
Lemma 17.3.2 (i) Let A Alg(S) and let z : domz

A
A be any
mapping such that the sort of z(, A) is the output sort of ((, a)

A
). Then we have for all M A:
c
A
z
M = G

A,z
M,
December 8, 2002
17 Independence, free partial algebras and primitive classes 411
and in the case of > 1 we even get
c
A
z
M =
_
G

A,z
M [ < .
(ii) Let K Alg(S) be any E-variety, A K and z as in (i). Then
we have for every ordinal number > 0 and for every subset M
of A that (G

A,z
M, (
A
)

) and (G

A,z
M, (
A
z
)

) are relative subal-


gebras of A, respectively of A
z
, and they are both K-algebras. Moreover
(G

A,z
M, (
A
z
)

) is generated by M.
Proof (i) can easily be proved in the same way as for the step operator D
(cf. subsection 3.4).
(ii) is also easily checked, since, for each > 0, G

:= G

A,z
M is a closed
subset of the K-algebra A (by the denition of G
A,z
); hence the relative
subalgebras (G

A,z
M, (
A
)

) and (G

A,z
M, (
A
z
)

) of A and A
z
, respec-
tively, are K-algebras. That M generates (G

A,z
M, (
A
z
)

) can be proved
by transnite induction.
The remaining part of this subsection is basically dedicated to the proof
and to the discussion of consequences of the following
Theorem 17.3.3 Let K be any primitive class of similar partial algebras,
such that K contains a K-algebra with at least two elements in each phylum,
and let M be any set such that v
F(M,K)
[F(M, K)] is a closed subset of S (or
K = ind
F(M,K)
M).
Then there exist a set N containing M and a bijective mapping z :

F(N,K)
(N M) satisfying z(, a) / a[()] for all (, a)

F(N,K)

such that the identity mapping id
M
induces an isomorphism between F(M, K
t
)
and F(N, K)
z
.
Before we prove this result, we want rst to discuss its immediate conse-
quences, and, moreover, to prove some useful preparatory facts. In particular
we immediately get a result similar to the inner axiomatic characterization
of absolutely free completions of partial algebras.
Corollary 17.3.4 Let K be a primitive class of partial algebras which con-
tains a K-algebra with at least two elements in each phylum for each sort s
in S. Let A K
t
be a total K-algebra and M any subset of A such that ei-
ther v
F(M,K)
[F(M, K)] is a closed subset of S or K = ind
F(M,K)
M. Then id
M
induces an isomorphism between F(M, K
t
) and A, iff A has the following
properties:
December 8, 2002
412 Chapter IV Special Topics
(RFC 1) M generates A.
(RFC 2) There exists a weak subalgebra B of A with carrier set B = A
such that the assignment z :

B
A, dened by z(, a) :=
A
(a)
for every (, a) domz, and the set N := M z[

B
] fulll the
following conditions:
(a) M z[

B
] = ,
(b) z is injective, and z(, a) / a(()) for all (, a)

B
,
(c) id
N
induces an isomorphism between B and F(N, K).
Assumptions and Notation 17.3.5 For the remainder of this subsection
we shall use the following assumptions and abbreviations:
K Alg(S) shall designate an E-variety which contains a K-algebra with
at least two elements in each phylum, and as before K
t
designates the subclass
of all total K-algebras. We assume that M N ,= are any two sets and
z : domz

F(N,K)
(N M) shall be an injective mapping satisfying
z(, a) / a(()) (the sort of z(, a) being the output sort of ) for all
(, a) domz. Finally assume that either v
F(M,K)
[F(M, K)] is a closed subset
of S or that K = ind
F(M,K)
M, and in the latter case additionally F(, K)

=
c
F(Q,K)
for each non-empty set Q.
Moreover, we introduce the abbreviations
F := F(N, K), G := G
F,z
, (

) := (
F
)

, (
z
) := (
F
z
)

,
and we shall also denote by (

) and (
z
) the structures of relative sub-
algebras of F and F
z
, respectively (for single operations we just omit the
brackets).
Lemma 17.3.6 In addition to the assumptions in 17.3.5 let A be any M-
generated relative subalgebra of F
z
. Then
(i) A = c
F
(N A).
(ii) F(M, K)

= C
F
M (A, (

))

= F(N A, K).
(iii) The restriction mapping z

:= z[
z
1
(NA)
is a bijection from domz

onto
(N A) M, and domz

(A,(

))
.
December 8, 2002
17 Independence, free partial algebras and primitive classes 413
(iv) N

:= N G

M = N D
F
z

M [ < = M z[domz

((

M [ < )
()
dom

) [ )], ( > 0)
(v) G

M = c
F
N

, ( > 0).
(vi) C
F
z
M (and each G

M ( > 0)) is K
t
-freely generated by M.
(vii) If additionally domz =

F
, then id
M
induces an isomorphism between
C
F
M and F(M, K
t
).
Proof Ad (i) Since A K, id
NA
can be extended to a homomorphism,
say g : c
F
(N A) =: B A, such that g = id
B
observe that,
because of the fact that M generates A and because of the assumptions
on F(M, K) or K, we may apply 17.1.2 in order to realize that B is
isomorphic to F(N A, K). Hence we have M (N A) B A.
Since z

[domz

] (N A) B (for z

as dened in (iii)), B
z

is a
subalgebra of A containing M observe that B cannot contain any
sequence of

A
, which z maps outside of B , therefore B = c
A
M = A,
and we even get equality between A and B.
Ad (ii) Again the assumptions on K and M imply the rst equality and the
inclusion, while the last equality has already been shown in connection
with the proof of (i).
Ad (iii) This is an immediate consequence of the proof of (i).
Ad (iv) These statements are an immediate consequence of the denitions
of G
A,z
and (
z
), and of the fact that N c
F
Q = Q for each subset Q
of N because of the A-freeness of N within F (cf. (v)).
Ad (v) This follows from (i) and 17.3.2.(ii).
Ad (vi) This statement is proved by transnite induction: Assume D K
t
,
and let g : M D be any mapping. Since M is K-free in F(M, K), g has
a homomorphic extension g

: F(M, K) D, whose restriction to M


is a homomorphism g
0
: (M, (
z
)) D extending g. Now let > 0
be any ordinal number such that for every ordinal number < there
exists a homomorphism g

: (G

M, (
z
)) D satisfying g

[
M
= g.
Since (G

M, (
z
)) is generated by M, there holds, for any two ordinal
numbers and which satisfy < < , that graph g

graph g

.
Hence g

with graph g

:=

graph g

[ < maps A

:= (

M [
< , (
z
)) homomorphically into D. The other statements of the
December 8, 2002
414 Chapter IV Special Topics
Lemma imply that G

M = c
F
N

, where N

is dened as in (iv) and


satises N

[ < z[

(A,(

))
domz].
Let g

be the restriction of g

to NA

, and let g
0
: N

D be dened
as follows:
graph g
0
:= graph g

(z(, a),
D
(g

(a))) [ (, a)
_
(A,(

))
domz .
Because of the assumptions on z, g
0
is well dened. Therefore we can as-
sume g

to be the homomorphic extension of g


0
to C
F
N

= (G

M, (
z
))
(into D). Let be the ordinal dimension of the given similarity type,
then g

:= g

: c
F
z
M D is the homomorphic extension of g, whose
existence has been claimed.
Ad (vii) This follows from (vi) and the uniqueness up to isomorphism of the
K
t
-universal K
t
-solution.
The preceding results now make it relatively easy to prove Theorem 17.3.3:
17.3.7 (Proof of Theorem 17.3.3) We rst show that there exists a set N

containing M such that each phylum of N

M has a cardinality not less than


the cardinality of

F(N

,K)
, since this will then allow an injective mapping
z

F(N

,K)
N

. Thus take n := 2
#M+#S+#+#+#
0
, where is the
ordinal dimension of the given type. Consider any set N

which contains M
and whose phyla, say N

s
:= v
1
N

(s), for each s S satisfy #(N

s
M
s
) =
n. Then N

is a set of the kind we have looked for (cf. the arguments in


subsection 5.10): For each we have #(F(N

, K)
()
) (n #S)
#()

n.
This suces to allow an injective mapping z

as described above, since it


now easily follows that also #

F(N

,K)
n.
Dene A := C
F(N

,K)
z M, N := N

A and B := F(N, K) considered as a


weak subalgebra of A; then (RFC 1) and (RFC 2) (cf. 17.3.4) are easily seen
to be satised (z is dened to be the resriction of z

to

A
).
Remark 17.3.8 We can get another kind of proof of Theorem 17.3.3, since
(RFC 2)(b) can be achieved recursively by using a direct limit construction
and observing facts in the proof of 17.3.6.(vi). Thus we get a version where
we do not need the above cardinality argument. But notice that then we have
also to nd sets disjoint from the given ones, whose elements may become the
distinct values for the operations undened in the previous step (see 23.2).
December 8, 2002
17 Independence, free partial algebras and primitive classes 415
17.4 Characterization of heterogeneous relatively free
partial algebras
As before we assume throughout this section that S is a heterogeneous sort
algebra. The results in the previous sections and subsections now enable
us to formulate and prove the following Characterization Theorem of Free
Partial Algebras for the nontrivial case, and we shall show in the subsequent
remarks that the theorem remains true - except for the equivalence between
(v) and the other items - even for the cases when A has one-element phyla.
Theorem 17.4.1 (Characterization Theorem of Free Partial Algebras)
Let A be any partial S-algebra such that each nonempty phylum of A has at
least two elements, and let M be any generating subset of A. Then the fol-
lowing statements are equivalent:
(i) A is a free partial algebra with M as a basis (i.e. M is A-independent
in A).
(ii) A = F(M, K) for some nontrivial primitive class K of partial S-algebras.
(iii) A is isomorphic to the partial algebra F
M
(A; S) of all total algebraic op-
erations of type M on A. This isomorphism is induced by the mapping
e
M
m
m (m M) from E
M
(A; S) onto M.
(iv) For every element a A there exists exactly one algebraic partial op-
eration g
a
APO
M
(A; S) such that a = g
a
(id
M
). And for every
g APO
M
(A; S) id
M
domg implies that g is total, i.e. g F
M
(A; S).
(v) There exists a free total algebra B with basis M such that A is a
relative subalgebra of B, and such that for every algebraic operation
h F
M
(B; S) of B the following is true: If h(id
M
) A, then the re-
striction h[
M
A
: A
M
B is an element of F
M
(A), i.e. an algebraic total
operation on A.
(vi) Let F := F(M, TAlg(S)) be the (S) total Peano (or term-) algebra
on M of type , and let i be the identity mapping id
M
on M consid-
ered as a mapping out of F into A. Then i : F domi

A is
a closed, fully invariant and surjective homomorphism, and domi

is
freely generated by M.
December 8, 2002
416 Chapter IV Special Topics
Proof The equivalence of (i) and (ii) is trivial, since we can choose K :=
ind
A
M, and the equivalence of (i) and (iii) easily follows from 15.4.6 and 15.4.8.
The equivalence of (i) (or (ii)) with (vi) has been proved in section 6 (cf. e.g.
6.3.9).
(i) (iv) We already know that e
M
m
m induces an isomorphism, say f :
F
M
(A; S) A such that f(h) = h(id
M
) for every h F
M
(A; S). Thus
there corresponds to every element of A exactly one from F
M
(A; S).
Now assume h APO
M
(A; S), such that id
M
domh, and let a A
M
be arbitrary. Since M is a basis of A, a extends to an endomor-
phism, say a

of A. From 5.7.5 and 15.2.1 we may infer that al-


gebraic partial operations commute with homomorphisms; therefore
a

(h(id
M
)) = h(a

id
M
) = h(a). This shows that a domh, and
that h is therefore total on A.
(iv) (i) By 15.4.6 e
M
m
m induces a homomorphism, say f : F
M
(A; S)
A. The assumptions of (iv) imply that f is injective. From 15.5.7 we
know that f has a closed homomorphic extension, say f

, and from (iv)


there follows that f

is injective and that domf = domf

, i.e. f = f

showing that f is closed and therefore also surjective, hence an isomor-


phism. Thus A is freely generated by M.
(i) (v) Because of the assumptions on A we may apply 17.2.8 in con-
nection with K := ind
A
M, and A is fully embeddable over M into
F(M, K
t
) =: B. Without loss of generality we may even assume that
A is a relative subalgebra of B. Consider h F
M
(B; S) such that
h(id
M
) A, and let a A
M
be any assignment which is obvi-
ously also an assignment from M into B. Since M freely generates
A and B, a can be extended to endomorphisms, say a

and a

, of A
and B, respectively, where a

is the restriction of a

to A. Therefore
A (a

h)(id
M
) = (a

h)(id
M
) = h(a

id
M
) = h(a). This shows that
the restriction of h to A is a total operation on A. By algebraic in-
duction it can also be shown that the restriction of h to A is even an
algebraic operation on A; therefore it belongs to F
M
(A; S).
(v) (i) Assume that A is a relative subalgebra of the free total algebra B
with basis M, such that (v) holds. Let f : M A be any assignment.
Then f has a homomorphic extension f

: B B. Since B is isomorphic
to F
M
(B, S), there exists for every element b of B exactly one algebraic
operation h
b
F
M
(B; S) such that b = h
b
(id
M
). Now assume that
December 8, 2002
17 Independence, free partial algebras and primitive classes 417
gure uncompleted
Figure 17.1: f17.1
b belongs to A; then we have f

b = (f

h)(id
M
) = h
b
(f

id
M
) =
h
b
(f). Since A b = h
b
(id
M
) and f A
M
we can infer with (v) that
f

b = h
b
(f) A. This shows that the restriction f

[
A
of f

is an
endomorphism of A; therefore M is free in A. Since M generates A by
general assumption, M is a basis of A.
Remarks 17.4.2 (i) Before we discuss the nontriviality assumptions of 17.4.1
let us observe that 17.4.1.(iv) and (v) generalize E.Marczewskis con-
cept of freeness of total algebras (cf. [Mar58]), and that freeness with
respect to sets is closely related to algebraic total operations, as we
have already seen in section 6.
(ii) From our observations in section 15 and on other occasions it is easy
to realize that items (i), (ii), (iii) and (vi) of 17.4.1 are equivalent for
every partial algebra A generated by M. Since M is assumed to be a
subset of A, the proof of the equivalence between (i) and (iv) carries
over to the arbitrary case.
Only in connection with (v) do we need the assumption of non-triviality
of A: As a simple example consider the case when we have only two
sorts, say s and s

, and two unary operations, say : s s

and

: s

, and A
s
:= M
s
:= a
1
, a
2
, and A
s
:= M
s
:= a, such
that
A
maps a
1
and a
2
on a and

A
is empty (see Figure 17.1).
Then it is easy to see that A is freely generated by M, but because of
a M each completion B of A which has M as basis can only have
a one-element phylum of sort s

and therefore A cannot be a relative


subalgebra of B. Because of this (v) cannot be equivalent to (i).
Since a is generated by a
1
or a
2
, M is not independent in a clo-
sure theoretical sense, while in the homogeneous case an algebraically
independent set M is also closure independent: m / c
A
(M m)
for each m M.
Notice that we cannot nd an example similar to the above one, when
S has only one element: Let A := a; if M = , then a has to be a
fundamental constant of A, and if A is partial, then it is easily seen
December 8, 2002
418 Chapter IV Special Topics
that F := F(A, TAlg()) is freely generated by M and contains A as a
relative subalgebra; if M = a and A is partial, then dene on F a
new structure (

as follows:

(a) :=
_
b, if a is a constant sequence with value b, and graph
A
,= ,
F(a), else.
Then C
(F,(

)
M is freely generated by M and contains A as a relative
subalgebra.
Since then also the rest of the proof of the equivalence between (i) and (v)
in 17.4.1 remains true, we can state the
Corollary 17.4.3 In the homogeneous case 17.4.1.(i) through 17.4.1.(vi)
are equivalent for each partial algebra A with respect to some generating sub-
set M of A.
17.5 Generators and relations
17.5.1 In particular in group theory it is customary to specify groups by
presenting a set, say M, of generators and a set, say G, of equations among
terms in these generators usually called relations and to say that the
group under consideration is the free group generated by the set M satisfying
the relations in G. This procedure can easily be extended to arbitrary
varieties of total algebras, but it seems to be not so well known that it
also applies to partial algebras. We shall treat for lack of space in this
subsection only the mere existence and the general principles. More detailed
information in connection with hep-varieties can be obtained in Part II.
The main dierence between the partial and the total case lies in the fact
that in the total case one need not care about which terms are dened and
which are not; but in the partial case this is an essential item. The concept of
E-equations takes care of this necessity and provides a suitable tool to handle
denedness, while the principle of freeness takes care of undenedness in
the specication.
Let us make this more precise by an example:
We rst want to specify for := (2) ( := ) within the whole
species Alg() the free partial algebra, say F, generated by M := a, b, c and
satisfying the set G := ((a, a), (b, b)), ((c, c), (c, c)), (((a, b), c), (a, c))
December 8, 2002
17 Independence, free partial algebras and primitive classes 419
gure uncompleted
Figure 17.2: f17.2
of relations - actually these relations are formulated in a language for the ex-
tended type

:= (2, 0, 0, 0), extended by three constants for a, b and c (which


we have also written as a, b and c, as is often done in this connection):
First a partial Peano algebra, say P, is constructed as a relative subalge-
bra of F(M, TAlg(S)) consisting of all terms and subterms contained in G, i.e.
P := a, b, c, (a, a), (b, b), (c, c), (a, b), (a, c), ((a, b), c).
P describes the scope of the partial algebra to be specied; and in order
to take care of the relations we only have to factor P with respect to the
congruence relation generated by G in P, i.e. F := P/Con
P
G. Thus F is
described by Figure 17.2. If it is desired that for instance (x, x) should
be a total algebraic operation, then one has to specify the free partial al-
gebra presented by M and G in the E-variety K characterized by the set
Q := ((x, x), (x, x)) of E-equations, i.e. one has specied F

:= F(F, K)
(which would be described in connection with Figure 17.2 by letting start
from each element of the upper row an innite path, each arrow representing
the formation of
F

(d, d), where d is the corresponding argument).


What has been sketched above is made more precise in 17.5.2 and 17.5.3:
Denition 17.5.2 Let M be any set, let G F(M, TAlg(S))F(M, TAlg(S))
be a set of E-equations, let Q be any set of E-equations with respect to any
set X of variables, and let K := Mod Eeq Q be the E-variety dened by Q.
Then we say that a partial algebra A together with a mapping r
A
: M A
is a free K-algebra presented by M and G, if A is a K-algebra generated
by r
A
[M] such that G ker r

A
, and for every mapping g from M into
any K-algebra B which saties G ker g

, there exists a homomorphism


g : A B satisfying g = g r
A
(see Figure 17.3).
Proposition 17.5.3 Let M be any set, let G be a set of E-equations with
respect to F(M, TAlg(S)), let Q be any set of QE-equations with respect to
some set X of variables, and let K := Mod QEeq Q be the QE-variety dened
December 8, 2002
420 Chapter IV Special Topics
gure uncompleted
Figure 17.3: f17.3
by Q. Then the free partial K-algebra presented by M and G exists and is
determined up to unique isomorphism, and it is denoted by F((M, G), Q) (or
by F((M, G), K), if K is the QE-variety dened by Q), and the canonical
mapping from M into F((M, G), Q) is denoted by r
M,G
.
In particular, F((M, G), Q) can be obtained as follows:
Let P := (M G) be the relative subalgebra of F(M, TAlg(S)) on the
set M joined with the set of all subterms of terms occurring in G; then
(r
M,G
, F((M, G), Q)) is the K-universal K-solution of P/Con
P
G.
Proof The description of F((M, G), Q) at the end of 17.5.3 together with 17.5.2
easily veries the statement in 17.5.3.
17.5.4 Notice that we have already used similar constructions in section 6
and Chapter II, when we translated validity of sets of E- or QE-equations
into injectivity with respect to special epimorphisms.
17.6 Initial objects and representation by canonical
terms
17.6.1 In connection with the algebraic specication of abstract data types
in computer science so-called initial objects within varieties specied by equa-
tions or elementary implications play a fundamental role. We recall that in
category theory an object A of a category C is called an initial object of C iff
for every C-object B there exists exactly one C-morphism from A into B.
Roughly speaking the initial algebra approach to algebraic specication
of abstract data types then consists in nding a set of axioms such that
the algebraic structure under consideration becomes the initial object in the
subcategory of (partial) algebras in which the axioms are valid.
Lemma 17.6.2 shows the connection between initial objects and free (par-
tial) algebras:
Lemma 17.6.2 Let K be any non-empty Horn-variety (cf. 8.1.2). Then one
has:
December 8, 2002
17 Independence, free partial algebras and primitive classes 421
(i) For every partial algebra A which allows at least one homomorphism
into some K-algebra, the K-universal solution F(A, K) belongs to K.
(ii) K has an initial object, and this is naturally isomorphic with F(, K),
the K-free K-algebra K-freely generated by the empty set.
Proof Ad (i) K is assumed to be a universal Horn-variety, therefore the
statement follows from 5.11.9 and 8.3.5.(vii).
Ad (ii) F := F(, K) exists in K, since the empty set allows a mapping
(homomorphism) into every partial algebra, and K is assumed to be
non-empty. The same reason and the K-freeness of F allows at least
one homomorphism from F into each K-algebra. And since a homomor-
phism is uniquely determined by its restriction to a generating subset,
and since the empty set allows exactly one mapping into each set, this
homomorphism is unique, therefore F is an initial object in K. It is well
known in category theory (exercise!) that there can be up to isomor-
phism at most one initial object in any category. Thus (ii) is proved.
Since each E-, ECE-, or QE-variety is always a non-empty Horn-variety,
we have the
Corollary 17.6.3 Each E-, ECE- or QE-variety K of partial algebras has
an initial object, and this is represented by F(, K).
The limitations of the above results and of the initial algebra approach
to abstract data types in computer science are shown by the following result:
Proposition 17.6.4 Let A be any partial algebra of type (with respect to
some sort algebra S), and let M be a generating subset of A. Moreover,
let
M
be the similarity type obtained from by adding a nullary operation
symbol, say c
m
, to the set of operation symbols for each m M such
that c
m
is of the same sort as m (the set of operation symbols thus obtained
will be designated by
M
). We then convert S into a partial
M
-algebra S
M
by dening c
S
M
m
:= v
A
(m), and similarly we convert A into A
M
by dening
c
A
M
m
:= m (in each case for each m M). Then A
M
is an initial K-algebra
for some suitable E-variety K of partial S
M
-algebras.
December 8, 2002
422 Chapter IV Special Topics
Notice that the above procedure is well known in model theory as the
extension of the underlying language by a suitable set of constants corre-
sponding to some subset M of A.
Proof Since M generates A, there exists for each a A a term, say (M : t
a
)
on M (considering M also as a set of variables) such that a = (M : t
a
)
A
(id
M
).
Let i : M c
m
[ m M , i(m) := c
m
(m M) be a bijective mapping.
Dene for F
M
:= F(, TAlg(S
M
)).
G := ( : (M : t)
F
M
(i)
e
= (M : t

)
F
M
(i)) [ t, t

F(M, TAlg(S)) and (M :


t)
A
= (M : t

)
A
(id
M
) (notice that equality in the meta-language is used
with the same semantics as existence equality in the object language).
Then it is easy to check (exercise!) that A
M
is an initial object in the E-
variety K := Mod
S
M
Eeq
S
M
G.
Remarks 17.6.5 (i) The above result shows that theoretically each par-
tial algebra can be algebraically specied as an abstract data type in
connection with the initial algebra approach. We have only forgotten
to mention that this specication has to be an explicit, i.e. a recursive
one. Namely, the real problem only starts when one tries to nd a
recursive algebraic specication. We cannot go into more details here
on this problem, we only wanted to show the limitations for the praxis
of a too general formulation of the initial algebra approach.
(ii) A similar eect is met in connection with the following method of rep-
resenting partial algebras, which also has a non-constructive general
solution. Nevertheless it is a very useful tool like the initial algebra
approach , whenever there is an explicit way to eectively determine
such a representation as we want to discuss.
The following denition and result are slight generalizations of Deni-
tion 10 and Theorem 8, respectively, in [GoTcWa78].
Denition 17.6.6 A partial algebra A is said to be a canonical term algebra
iff A is a subset of F(M, TAlg(S)) for some term algebra F(M, TAlg(S)) on
a suitable set M of variables, such that
(i) A is closed with respect to the formation of subterms, and
(ii) the relative subalgebra on A of F(M, TAlg(S)) is a weak subalgebra
of A.
December 8, 2002
17 Independence, free partial algebras and primitive classes 423
This can also be formulated by saying that a weak subalgebra of A with
carrier set A is an initial relative subalgebra of some suitable term algebra.
While in [GoTcWa78] this concept is restricted to total heterogeneous alge-
bras generated by the empty set, we allow here a more general situation!
Theorem 17.6.7 Each partial algebra A is isomorphic to some canonical
term algebra.
Proof Let M be any generating subset of A, and let f := (id
M
)

: domid
M


A be the closed initial extension of id
M
with respect to the term algebra
F(M, TAlg(S)). Let 1 := I [ I is an initial segment of domf, and the
restriction of f to I is injective .
It is easy to see that the union of any chain of elements of 1 ordered by
set theoretical inclusion also belongs to 1; therefore 1 is inductive and has
maximal elements according to Zorns Lemma (notice that M belongs to 1,
therefore 1 is non-empty). Let C 1 be such a maximal element; we
then claim that f[
C
is surjective and therefore bijective (namely we show by
algebraic induction that each element of domf is congruent modulo ker f to
some element of C, and since f is surjective, this also implies surjectivity of
f[
C
):
First we show that each element of B
domf
1
C
is congruent to some element
of C: Assume that t := ( t
k
[ k () ) B
domf
1
C
. If f(t) / f[C],
then f[
C{t}
would be injective, and C t would also be an initial segment
of F(M, TAlg(S)), contradicting the maximality of C. Now assume that for
each ordinal satisfying 0 < each element of B

domf
C is congruent
modulo ker f to some element of C, and assume t := ( t
k
[ k () )
B

domf
C, then there is by induction hypothesis for each k () some term
t

k
C such that (t
k
, t

k
) ker f. Since ker f is a closed congruence relation
and ( t
k
[ k () ) dom
domf
, we get t

:= ( t

k
[ k () ) B
domf
1
C
and (t, t

) ker f. By assumption there is t

C such that (t

, t

) ker f;
therefore (t, t

) ker f, which is what we wanted to show.


f[
C
is then a bijective homomorphism from the relative subalgebra of F(M, TAlg(S))
on C onto A, and the structure on C can be enriched in such a way that
f[
C
becomes an isomorphism from C onto A, while C is a canonical term
algebra.
December 8, 2002
424 Chapter IV Special Topics
Remarks and Examples 17.6.8 (i) Following more the idea of proof
in [GoTcWa78] one can proceed in a dierent way in order to prove 17.6.7:
Again let M be any generating subset of A, and let f := (id
M
) :
domid
M

A be the closed initial extension of id


M
with respect to
the term algebra F(M, TAlg(S)). For brevity we set B := domf and
denote the congruence class of t B with respect to ker f by [t]. Let
= () be the ordinal dimension of the type , and for each t B
let (t) designate the step number of t with respect to M (cf. subsec-
tion 3.4). This time we proceed by transnite induction on the step
number of terms to produce a family ( C
n
[ n ) such that:
(1) If t C
n
, then (t) n.
(2) If t B such that [t] has a representative t

with (t

) n, then
there is a unique representative of [t] in C
n
.
[(3) If
B
( t
k
[ k () ) C
n
, then t
k
C
n
for each k ().
One can now ll in the details of the proof (exercise!) in such a way
that one can infer that the set C of canonical terms which one ends
up with in the above proof can even be chosen in such a way that
for each (t, t

) ker f with t C the step number of t with respect


to M is always less or equal to the one of t

(exercise!), i.e. that the


representative in C is always of minimal complexity.
(ii) The above result is sort of a normal form theorem for arbitrary par-
tial algebras with only the drawback that neither C nor the equiva-
lence of any element of domf to its representative in C is in general
computable. Nevertheless the above theorem can be a useful tool in
proving freeness (or correct specication) of some partial algebra under
consideration, whenever a canonical term algebra representing it can
be explicitly constructed. Actually such a construction was a rst item
for the author showing the usefulness of applying partial algebras (cf.
the proof of Satz 2.2 in [B68]).
(iii) Several applications of the above principle of representation of partial
algebras can be found in [GoTcWa78]. Another example is given by
representing a free semigroup freely generated by a set, say a, b, c, d,
by introducing additional constants a, b, c, d to the given type = (2)
and considering in F(, TAlg(2, 0, 0, 0, 0)) the subset generated from
December 8, 2002
17 Independence, free partial algebras and primitive classes 425
the empty set by the nullary operations corresponding to the intro-
duced four constants and the unary operations corresponding to the
terms x a, x b, x c and x d ( being the operation symbol
for the binary operation). The resulting subset of the Peano algebra
F(, TAlg(2, 0, 0, 0, 0)) may be endowed in a natural way with a binary
multiplication (composition of the corresponding term operations) in
order to yield a canonical term algebra for the free semigroup freely
generated by a, b, c, d. In this example we have applied another prin-
ciple often used in connection with the algebraic specication of ab-
stract data types in computer science, namely the principle of hidden
functions (here the hidden functions are the four auxiliary nullary con-
stants) , which is even more useful in connection with problems of nite
axiomatization (cf. [Go78]).
(iv) In order to show that for the study of initial objects in computer science
the aspects taken from universal algebra yield only the trivial part of
the problem, while the main diculties lie in questions of computabil-
ity, we present the following lemma, whose last statement shows us
that each partial algebra can be described as the initial object of a
suitable E-variety just by introducing enough additional constants and
formulating a suitable set of E-equations.
Lemma 17.6.9 Let K be any E-variety of partial S-algebras, A K be any
K-algebra, and
A
be the extension of the underlying similarity type by
nullary fundamental constants, say c
a
, for each element a of A according
to 17.6.4. Let K
A
be the E-variety of all partial S
A
-algebras dened by the
set Eeq K D
+
(A) of E-equations, where
D
+
(A) := ( : t
e
= t

) [ t, t

F(, TAlg(S
A
)), and A [= ( : t
e
= t

) ( : c
a
e
= c
a
) [ a A
is equivalent to the model theoretic positive diagram of A (cf. Chang, Keisler [ChKs73]).
Then one has:
(i) There is a natural bijection
A
between K
A
and the class
Hom

(A, K) := (f, B) [ f : A B and B K,


where for any C K
A
we dene
A
(C) := (f
C
, (C, (
C
)

) =: C
A
)
(C
A
is called the -reduct of C; f
C
is induced by a c
C
a
(a A)).
December 8, 2002
426 Chapter IV Special Topics
(ii) This bijection
A
satises (r
A,Kt
, F(A, K
t
)) =
A
(F(A
A
, (K
A
)
t
)), i.e.
the -reduct of the (K
A
)
t
-universal (K
A
)
t
-solution of A
A
is naturally
isomorphic to the K
t
-universal K
t
-solution of A.
(iii) A
A
is an initial object of K
A
.
Proof Ad (i) Assume C K
A
, and as indicated let C
A
designate the -
reduct of C. Since C = D
+
(A), it is easy to realize that f
C
: A
A
C,
a c
C
a
(a A) is a homomorphism, which remains a homomorphism
for the reducts, therefore we can dene
A
(C) := (f
C
, C
A
) as indicated.
Since f
C
is the only possible homomorphism from A
A
into C,
A
is a
mapping, and it is obviously injective; that it is also surjective holds,
since f : A B implies that B
A
:= (B, (
B
)

) (c
B
A
a
:= f(a))
aA
)
K
A
, and
A
(B
A
) = (f, B).
Ad (ii) This is an immediate consequence of (i), since C
A
(K
A
)
t
iff C
A

K
t
.
Ad (iii) This now follows from 17.6.4
December 8, 2002
18 Dependence of operations on arguments and element basis property 427
18 Dependence of operations on arguments
and element basis property
18.1 General concepts
Closely related to the concept of a basis is at least in the homogeneous
case the one of an element basis, which concept is connected with the
one of dependence of operations on their arguments. In this subsection we
try a generalization to many-sorted partial algebras of material contained in
J.Schmidt [Sch64a].
Let us note that as before we denote for every set N and for every sort s
(of S) by N
s
the subset of N of all elements of sort s: N
s
:= v
1
N
(s).
Denition 18.1.1 Let A be any partial algebra, M any subset and x any
element of A. A subset M

of M is called an element basis of x in A with


respect to M, iff the following two conditions hold:
(EB 1) x c
A
M

;
(EB 2) for every subset M

of M one has x c
A
M

iff M

.
If an element basis of some element x of A exists with respect to M, then it
is uniquely determined and we shall denote it by xM.
We say that the subset M of A has the element basis property in A, iff
each element x of c
A
M has an element basis in A with respect to M.
Remarks 18.1.2 (i) Let A be any partial algebra; let M be any subset
of A and x A such that the element basis xM of x in A for M exists.
Then xM is obviously uniquely determined by the other data and is
given by
xM =

[ M

M and x c
A
M

.
(ii) So far nothing at all has been said about the existence of element bases
in general. It is obvious (cf. section 5) that the Peano basis of any
partial Peano algebra (especially of any term algebra) has the element
basis property. And if a subset M of a partial algebra A has the element
basis property, then so has each subset of M.
We shall see below that it is not easy in general at least in the really
many-sorted or in the innitary case to give criteria for when the
December 8, 2002
428 Chapter IV Special Topics
element basis exists. In the nitary homogeneous case we can show
that except for some restrictions on the values of constant functions all
algebraically independent sets have the element basis property.
(iii) One instance when the concept of an element basis may become im-
portant, is given in connection with relativized logic, i.e. when one
tries to formulate equations or more general formulas already within
some special theory, as is e.g. often done in group theory, where one
formulates equations or so-called relations for groups within a
given free group rather than in the abstract term algebra thus being
able to use associativity and other features in order to simplify the
presentation of the axioms.
Before we discuss in more detail the existence of element bases, let us
introduce the concepts of dependence of a partial operation on an argument
and of constant algebraic operations together with some facts about them.
Denition 18.1.3 (i) Let M be any set and L any subset of M. Moreover
let A be any partial algebra and h an M-ary partial operation on A. We
say that h does not depend on (is independent of) L iff for all elements c
of A and for all M-ary sequences a, b A
M
satisfying a[
(M\L)
= b[
(M\L)
one has h(a) = c iff h(b) = c.
Thus h does depend on L iff there are two M-ary sequences a and b
of A which coincide on all but the elements of L such that either h is
dened on one of them but not dened on the other, or h is dened on
both of them but h(a) ,= h(b), i.e. the values are dierent.
In case that L = l we also say that h does not depend (or depends)
on its l-th argument or simply that h does not depend (or depends,
respectively) on (the variable) l.
(ii) Let us dene a constant partial operation to be such a partial operation
which does not depend on any of its arguments.
(iii) Finally let C(A) designate the set of all values of constant algebraic
operations of A.
18.1.4 It is easy to realize that the above denition automatically implies
that a constant operation is either total or empty. Namely when for one
argument there is no element in the partial algebra for which the operation
December 8, 2002
18 Dependence of operations on arguments and element basis property 429
gure uncompleted
Figure 18.1: f18.1
takes a value, then it obviously has an empty domain, and otherwise it has
to be total.
Lemma 18.1.5 For any partial algebra A, C(A) is a closed subset of A; and
if A is homogeneous and if there is at least one non-empty nullary fundamen-
tal constant, then C(A) = c
A
.
Proof If C(A) is empty, then there are no constant algebraic operations
in A (in particular if there are no nullary constants in the type under con-
sideration), and is closed in A. If C(A) is non-empty, then assume ,
a dom
A
C(A)
K
and b :=
A
(a). For each k K

there then exist some


constant, say N
k
-ary, algebraic operation h
k
and some N
k
-ary sequence c
k
such that a(k) = h
k
(c
k
). Obviously h :=
A
( h
k
[ k K

) (superposition
of the rst kind!) is a constant algebraic operation of A with b among its
values, i.e. b C(A).
Let nally A be homogeneous and c
A
be non-empty, i.e. it contains at least
one element which then may be used to generate within c
A
all elements
of C(A). The converse inclusion is trivial.
In the case of many-sorted partial algebras A with at least two sorts C(A)
may contain several other closed subsets even when c
A
is non-empty. The
following example may provide an idea of what the situation might then look
like:
Example 18.1.6 Let us assume that S := s, s

, := (,

), :=
(0, 1, 2), and
S
:= s,
S
: s

,
S
: (s, s

) s. The partial
algebra A which we want to consider is then described in Figure 18.1. It
is then easy to realize that C(A) = a, b, c, that c
A
= a, and that b
is another closed subset of A properly contained in C(A) and even disjoint
from c
A
. We have added two additional elements in A in order to make

A
dependent on each of its arguments.
It should be easy to realize that by using more sorts the lattice of
closed subsets of A contained in C(A) can be made arbitrarily large.
December 8, 2002
430 Chapter IV Special Topics
The situation is once more dierent when we only allow closed subsets
with non-empty phyla, since then it may happen that C(A) is not of this
kind. We do not discuss this further here.
Observation 18.1.7 Since we are just discussing C(A), we would like to
mention that by introducing also D(A) as the intersection of all non-empty
closed subsets of A J.Schmidt has shown in [Sch64a] (Corollary to The-
orem 4) that there are exactly the following four possible cases for homoge-
neous partial algebras A:
I. c
A
= C(A) = D(A);
II. = c
A
C(A) = D(A);
III. = c
A
= C(A) D(A);
IV. = c
A
= C(A) = D(A).
That even for partial homogeneous algebras these are the only possibilities
([Sch64a] only deals with total algebras) can easily be derived from Lemma
18.1.5. In order to show that they are all possible, let us quote the following
examples from [Sch64a]:
Example I. Remember that c
A
,= iff A has at least one non-empty
nullary constant, and this can easily be realized. The equalities then follow
from 18.1.5.
Example II. Consider a group A as an algebra with multiplication and
inversion as the only fundamental operations, both non-nullary. Then x
x x
1
represents a constant algebraic operation with the unit element e as
constant value. And since e is a closed subset, C(A) = e. Neverthe-
less c
A
= , since one has forgotten to introduce the nullary operation
corresponding to the unit element e as a fundamental operation.
Example III. Let A consist of two dierent elements a, b; let
A
be the only
non-trivial permutation of A considered as the only fundamental operation.
Then there is no constant algebraic operation at all, and therefore C(A) = ,
whereas D(A) = A, since A is the only non-empty closed subset of A.
Example IV. In order to realize IV we can choose any algebra with at least
two elements such that any one-element subset is a subalgebra (e.g. let A be
a lattice, an ane space, etc.).
December 8, 2002
18 Dependence of operations on arguments and element basis property 431
18.2 Dependence on variables and algebraic indepen-
dence
A rst connection between the concepts of dependence on variables and the
element basis property on one side and algebraic independence on the other
is indicated by the following
Lemma 18.2.1 Let M be some algebraically independent subset of some par-
tial algebra A, let M

, M

be subsets of M such that for, some s S, M

s
and
M

s
are disjoint, and assume that g APO
M

(A; S) and h APO


M

(A; S)
are algebraic partial operations such that g id
M
= h id
M
. Then g (and
similarly h) does not depend on any argument of sort s. In particular, if M

and M

are disjoint, then g is constant.


Proof Clearly we assume that M

s
is non-empty. Let c := g id
M
= h id
M
;
since M

and M

are independent, this implies that g and h are total algebraic


operations on A, i.e. we do not have to worry about the domains. Assume
that a, b A
M

are any sequences which dier at most on arguments from M

s
,
and let f, f

: M A be any mappings such that f[


M
= a, f

[
M
= b, and
f[
M
= f

[
M
(this is possible, since M

s
M

s
= ). Since M is indepen-
dent in A, f and f

have homomorphic extensions, say f

, f

: C
A
M A,
respectively. As f

and f

permute with algebraic partial operations, we get


g(a) = g (f

id
M
) = f

(g id
M
) = f

(h id
M
) = h (f

id
M
)
= h (f

id
M
) = f

(h id
M
) = f

(g id
M
) = g (f

id
M
) = g(b),
showing that in fact g takes the same value on a and b.
Lemma 18.2.2 Let M be some algebraically independent subset of some
partial algebra A with algebraic closure operator c := c
A
. Let M

, M

be
subsets of M. Then
c M

c M

_
C(A), if M

= ,
= c (M

), if M

s
M

s
,= for all s S.
For the case when for some phyla the intersections are empty while they are
nonempty for others, see the remarks below.
Proof As we have shown earlier each element of, say, c M

is the value of an
M

-ary algebraic partial operation applied to id


M
. Since M

is independent
December 8, 2002
432 Chapter IV Special Topics
in A as a subset of the independent set M, each such algebraic operation is
total on A. If M

and M

are disjoint, then each element of c M

c M

is the value of a total algebraic constant operation and therefore it belongs


to C(A) (cf. 18.2.1).
Now assume that M

s
M

s
,= for each sort s. Then there is a homomor-
phism f from CM into A such that f(a) M

for all a M

and
f(b) = b for all b M

. A similar calculation as in the proof of 18.2.1


yields for every c c M

c M

that c = g(id
M
) = h(id
M
) for appropriate
algebraic partial operations g and h, i.e.
c = h(id
M
) = h(f id
M
) = f h(id
M
) = f g(id
M
) = g(f id
M
) c (M

).
Thus c M

c M

c (M

); since the converse inequality is trivial,


even equality holds.
Corollary 18.2.3 Let A be any partial algebra such that c
A
= C(A), and
let M be an independent subset of A such that all elements of M are of the
same sort. Then
c
A
M

c
A
M

= c
A
(M

)
for all subsets M

, M

M.
Notice that the assumption on M is in particular satised when A is
homogeneous.
Remarks and Examples 18.2.4 (i) The hypothesis c
A
= C(A) in 18.2.3
means that the sets of values of constant algebraic operations of a cer-
tain arity L and of arity , respectively, are the same; or as c
A
= C(A)
is equivalent with C(A) c
A
, we may simply say that any value of a
constant algebraic operation is the value of a nullary algebraic opera-
tion. As a matter of fact, this condition is at least in the homogeneous
case a very weak one; for in the only case when it does not hold orig-
inally for homogeneous partial algebras (see case 18.1.7.II), one may
slightly alter the algebraic structure of A by introducing one single new
fundamental operation, namely the nullary operation corresponding to
some certain element of C(A), thus obtaining a partial algebra, say
A

, with constants, whose closure operator, say c

, fullls condition
c

= C(A

) = C(A).
In the inhomogeneous case c
A
= C(A) can also be obtained by adding
forgotten nullary constants to the type, but in general more than one
might be necessary.
December 8, 2002
18 Dependence of operations on arguments and element basis property 433
gure uncompleted
Figure 18.2: f18.2
(ii) While in the homogeneous case only the values of constant algebraic
total operations may cause some trouble in connection with the element
basis property for independent subsets, the situation is much more
complicated in the inhomogeneous case. Let us illustrate the situation
in the many-sorted case by an example:
S := s, s

, := , := (2), graph
S
:= ((s, s

), s

). The
(total) S-algebra A is indicated in Figure 18.2. Let M := a, b, a

,
M

:= a, a

, M

:= b, a

. It is easy to see that M is independent,


that M generates A, that M

s
M

s
= and M

s
M

s
= a

, = , and
that c M

c M

(= a

, b

, c

) is neither contained in C(A) (= ), nor


is it equal to c (M

) (= a

).
Notice that
A
does not depend on its 0-th argument but on its 1
st
one.
(iii) This simple example may give an idea how complicated the situation
can become in the many-sorted case. We leave a more detailed discus-
sion as a problem if there is any need for it. Let us only nally observe
that in the example of A above the element b

has no element basis


with respect to M.
In connection with the following results one should observe the dierent
assumptions concerning arities (nitary or innitary) and number of
sorts (homogeneous or arbitrary).
Proposition 18.2.5 Let A be any partial algebra such that for c := c
A

C(A) = c , let h be an algebraic partial operation of type L on A such that
either all elements of L are of the same sort or each non-empty phylum of L
has at least two elements, and let a domh be a sequence without repetitions
(i.e. injective) of type L in A. Let the range M

of a be minimal for the


generation of x := h(a), i.e. x / c M

for all proper subsets M

of M

. Then
h depends on each of its variables.
Proof Let us assume the existence of an index l
0
L such that h does
not depend on l
0
(i.e. on its l
0
-th argument). In case L = l
0
h would
December 8, 2002
434 Chapter IV Special Topics
be constant; therefore x C(A), i.e. x c by hypothesis, contradicting
the minimality of M

= a(l
0
) , = . In case L ,= l
0
we might dene a
transformation of L onto K := (L l
0
) which carries l
0
into l
1
,= l
0
,
leaving all other elements of L xed (this is possible, since the phylum of L
containing l
0
has to contain by our assumptions at least one further element).
As the sequences a and ( a((l)) [ l L) only dier on the argument l = l
0
,
and as h does not depend on its l
0
-th argument, we obtain
x = h(a) = h(( a((l)) [ l L));
therefore x c M

, where M

:= a((l)) [ l L. But since a is a sequence


without repetitions, we get a(l
0
) / M

, thus M

is a proper subset of M

,
again contradicting the minimality of M

.
Remark 18.2.6 Our example in 18.2.4.(ii) shows that restrictive assump-
tions on M are necessary in the above proposition. Namely in this example
we can generate c from M only by using
A
which is neither constant nor
does it depend on its 0-th argument, nor can it be transformed into an alge-
braic operation depending on every argument, since its arity does not allow
a transformation (over S !) into a strictly smaller arity.
Proposition 18.2.7 Let A be a partial algebra with nitary fundamental
partial operations such that C(A) = c again we set c := c
A
, and let
M be a subset of A such that all elements of M are of the same sort. Then
each element x c M can be represented in the form x = h(a), where h is
an algebraic partial operation of some (nite!) type L depending on each of
its arguments, and a is an L-ary sequence without repetition and arguments
in M.
We may choose L = M

, where M

is a nite subset of M, and a(l) = l


for all l L = M

, i.e. a = id
M
.
Proof Since the fundamental operations are nitary, c is an (#
0
-) alge-
braic closure operator; therefore there exists for each x c M a nite subset
F M such that x c F. Thus (by Tarskis niteness criterion) there exists
a minimal subset, say M

F, such that x c M

, and as a consequence x
can be represented in the form x = h(id
M
), where h is an M

-ary algebraic
operation. Since id
M
obviously is a sequence without repetitions and all of
its arguments are of the same sort, Proposition 18.2.5 implies that h depends
on each of its variables.
December 8, 2002
18 Dependence of operations on arguments and element basis property 435
We shall show in 18.3.2 that this result cannot be generalized to partial
algebras with innitary operations.
While the last results did not need independence, this concept becomes
relevant in the following one.
Proposition 18.2.8 Let A be a partial algebra such that C(A) = c c :=
c
A
, let h be an algebraic partial operation of type L on A which depends on
each of its variables, let a domh be a sequence without repetitions of type L
in A, and let its range, say M

, be a subset of the independent subset M of A.


Then M

is a minimal subset of M for the generation of x := h(a), and if all


elements of M are of the same sort i.e. in particular if A is homogeneous
, then it even is the minimum subset with this property, i.e. M

for
all M

M such that x c M

.
Proof First, let us assume the existence of some proper subset M

such that x c M

. Then there would be an index l


0
L such that a(l
0
) /
M

. Therefore x c a(l) [ l K where K := L l


0
, hence x = g( a(k) [
k K ) where g is a partial algebraic operation of type K on A. Let i be the
identity transformation from K into L, then (cf. 15.2.5) h(a) = x = g( a(k) [
k K ) = g( a(i(k)) [ k K ) = g
i,L
(a).
Since g
i,L
is an algebraic partial operation as well as h, and since a is a se-
quence without repetitions in the independent set M, we obtain h = g
i,L
(cf.
Theorem 17.4.1 in the total case this is Marczewskis criterion of indepen-
dence in [Mar58]). But then h does not depend on its l
0
-th argument; for let
c and d be sequences of type L in A which dier only on l
0
, and observe that
h is total M being independent , then
h(c) = g
i,L
(c) = g( c(k) [ k K ) = g( d(k) [ k K ) = g
i,L
(d) = h(d).
This contradicts our hypothesis on h. Thus, as x c M

, M

is a minimal
subset M

M such that x c M

.
Now assume that all elements of M are of the same sort. Then M

even is
the minimum of all these subsets M

; for A satises C(A) = c . Therefore


18.2.3 implies that x c M

c M

= c (M

). Thus M

= M

,
i.e. M

.
December 8, 2002
436 Chapter IV Special Topics
Theorem 18.2.9 Let A be a partial algebra such that C(A) = c c := c
A
and let M be an independent subset of A such that all elements of M are
of the same sort. Then for an element x c M the following properties are
equivalent:
(i) The element basis for x in A with respect to M exists.
(ii) x can be represented in the form x = h(id
M
), where h is an M

-ary
algebraic (even total) operation on A depending on each of its variables,
and M

M.
Moreover, when one of these statements is true, then the above representation
of x is unique and M

is the element basis for x in A with respect to M


(unique irredundant coordinate representation).
Proof (i) (ii): Let M

be the element basis for x in A with respect


to M. Because of x c M

, x can be represented in the form x =


h(id
M
), where h is an algebraic (total!) operation of type M

in A.
But sequence id
M
being without repetitions, its range being a minimal
set M

such that x c M

, and all its elements being of the same sort,


h depends on each of its variables according to Proposition 18.2.5.
(ii) (i): The algebraic operation h depending on each of its variables,
sequence id
M
being without repetions, its range M

being a subset
of the independent set M, all of whose elements are of the same sort,
M

is the element basis of x in A with respect to M according to


Proposition 18.2.8. This proves also the uniqueness of M

and h in this
representation of x (M

itself being independent!).


We now turn to the element basis property; as we have already seen in
18.2.4.(ii) the element basis of an element need not always exist even with
respect to independent subsets (see also 18.3.2 below). For purely closure
theoretic results concerning the element basis and the element basis property
see [Sch64a]. Here we state as a consequence of the preceding results:
Theorem 18.2.10 Let A be a partial algebra with nitary fundamental par-
tial operations such that C(A) = c
A
, and let M be an independent subset
of A such that all elements of M are of the same sort. Then M has the
element basis property in A; moreover, the element bases for x in A with
respect to M are nite for all elements of c
A
M.
December 8, 2002
18 Dependence of operations on arguments and element basis property 437
18.3 A counterexample for the innitary homogeneous
case
Let us now consider a counterexample for the innitary case (see [Sch64a]):
I.e. we want to show that even in a total homogeneous algebra A without
constants, with innitary fundamental operations, an independent subset M
of A need not have the element basis property. As a preparation we prove
Proposition 18.3.1 For a homogeneous partial algebra A let h be a non-
constant algebraic partial operation of type L which does not depend on any
of its arguments, let a be an L-ary sequence without repetitions in A, and
let its range M be an independent subset of A. Then the element x := h(a)
(which belongs to c
A
M) does not possess an element basis in A with respect
to M.
Proof Let M

M be an arbitrary subset such that x c


A
M

. Let
K be the set of all l L such that a(l) M

; then there is an algebraic


operation g of type K on A such that x = g( a(k) [ k K ). Let i be the
identity transformation from K into L and b := ( a(k) [ k K ); then we get
(cf. the proof of 18.2.8)
h(a) = x = g(b) = g( a(i(k)) [ k K ) = g
i,L
(a).
Since a is a sequence without repetitions in the independent set M, and since
h and g
i,L
are algebraic partial operations which coincide on a, we obtain
g
i,L
= h, and therefore h( c
l
[ l L) = g( c
l
[ l K ) for each sequence
( c
l
[ l L) A
L
. Therefore, since h is non-constant and does not depend
on any of its arguments, the same is true with g. One concludes that K is
innite (since otherwise g could not be non-constant without depending on
any of its arguments), and in particular, K contains two distinct elements,
say k
0
and k
1
. We dene a transformation p from K onto K k
0
which
carries k
0
into k
1
, leaving all other elements of K xed. As in the proof
of 18.2.5 we obtain x = g( a(p(k)) [ k K ). Therefore x c
A
M

, where
M

:= a(p(k)) [ k K = M

a(k
0
) is a proper subset of M

: M

is therefore not the element basis of x in A with respect to M (in fact, k


0
has been an arbitrary element of K, thus the intersection of all such M

is
empty!).
December 8, 2002
438 Chapter IV Special Topics
Examples 18.3.2 (i) While every nitary partial operation not depend-
ing on any of its arguments is constant, this need not be the case for
an innitary partial operation, as any example of a (not too trivial)
Frechet-Urysohn limit space shows (for instance take the limit space
connected with the real or the rational line): We may dene a partial
operation h of type
0
by h( x
n
[ n
0
) := lim
n
x
n
, if this limit
exists, otherwise undened.
It is easy to realize that h does not depend on any of its arguments,
nor is it constant. If one chooses one xed element, say a, of the carrier
set and denes a as value of h whenever h is originally undened, one
even gets a total operation with these properties.
(ii) An example of innitary partial algebras in which independent sets
need not have the element basis property can now be constructed as
follows:
Let K be the class of all (total) algebras A with precisely one fundamen-
tal operation of type
0
which does not depend on any of its variables.
It is easily shown that K is an E-variety: For each argument n one has
to take two
0
-sequences a
n
and b
n
(of variables) without repetitions
which dier exactly in the n-th argument, and one has to consider the
E-equation h(a
n
)
e
= h(b
n
). The set of all these E-equations denes K.
Example (i) above shows that K is non-trivial in the sense that it con-
tains at least one algebra of more than one element. Therefore the
K-free K-algebra F := F(
0
, K) exists (with an innite basis). Obvi-
ously the only fundamental operation
F
is non-constant and does not
depend on any of its arguments (since the examples constructed in (i)
belong to K in particular the one with the rational line as carrier
set; therefore we may construct a surjective homomorphism from F
onto this algebra , showing that
F
is non-constant); hence by 18.3.1
no element of F possesses an element basis with respect to
0
in F,
although it is even true that C(F) = c
F
, i.e.
0
does not have the
element basis property in F.
December 8, 2002
19 Congruence relations revisited 439
19 Congruence relations revisited
19.1 Closed congruence relations of nitary partial al-
gebras
19.1.1 (Introductory Remarks and Denitions) In section 2 we have
already studied congruence relations of partial algebras in general and closed
congruence relations in particular, and we have seen that not every pair of
elements (of the same sort) in a given partial algebra will belong to a closed
congruence relation, but that in the nitary case there exists a largest closed
congruence relation containing all those pairs, and that any congruence rela-
tion contained in a closed one is also closed. Hence a syntactical description
of this largest closed congruence relation is equivalent to a syntactical de-
scription of those pairs generating a closed congruence relation.
Moreover, we have already seen that the lattice of all closed congruence
relations behaves much more like the congruence lattices of total algebras
than the lattice of all congruence relations does. We shall give some more
evidence for this observation in this subsection.
First of all let us recall that the set of all M-ary algebraic partial func-
tions of a given partial algebra A is just the set APO
M
(A
A
; S
A
) of all M-
ary algebraic operations of the partial algebra A
A
, where all elements of A
are additional fundamental constants. In this subsection we shall need in
particular the sets APF
s
(A; S) := APO
{s}
(A
A
; S
A
) of all unary partial al-
gebraic functions of input sort s for each s S. Finally, we consider
APF(A; S) :=

APF
s
(A; S) [ s S , the set of all unary algebraic partial
functions of A. Let us observe in this connection that one can obtain any
algebraic partial function of A by substituting elements of A of the appro-
priate sort for some set (possibly empty) of variables in an algebraic partial
operation of A.
Finally, for any given nitary partial algebra A, let
c
:=
c
(A) designate
the largest closed congruence relation of A.
Lemma 19.1.2 Let A be any partial algebra. Then an equivalence relation
on A is a (closed) congruence relation of A iff it is a (closed) congruence
relation of A
A
.
Proof This statement is obvious, since every equivalence relation is com-
patible with constant operations.
December 8, 2002
440 Chapter IV Special Topics
Of interest for us are the following consequences:
Corollary 19.1.3 Each congruence relation of a given partial algebra is
compatible with all partial algebraic functions.
Proposition 19.1.4 Let A be a nitary partial algebra, and let be any
equivalence relation on A. Then the following statements are equivalent:
(i) is a closed congruence relation of A.
(ii) is compatible with all unary algebraic partial functions of A, and if
f is any unary algebraic algebraic function on A, and (a, b) any
element of , then f is dened on a iff f is dened on b.
Proof It is an immediate consequence of 19.1.3 that (i) implies (ii). Thus
let us assume (ii): Let a, b A
()
such that for each l () one has
(a(l), b(l)) , and a dom
A
. For simplicity we assume that () =
1, . . . , k is the set of the rst k natural numbers (,= 0), and that c(i) = c
i
for c a, b and 1 i k. Let x
i
be a variable of the suitable sort
and dene for 1 i k p
i
(x
i
) :=
A
(b
1
, . . . , b
i1
, x
i
, a
i+1
, . . . , a
k
), whenever
this is dened and makes sense. Thus we obtain k unary algebraic partial
functions on A. We then have a
1
domp
1
, and it is easy to realize because
of the assumptions on using the inx notation for relations that all
of the following elements are dened and satisfy

A
(a) = p
1
(a
1
) p
1
(b
1
) = p
2
(a
2
) . . . p
k1
(b
k1
) = p
k
(a
k
) p
k
(b
k
) =
A
(b).
Because of the transitivity of this shows that b dom
A
and (
A
(a),
A
(b))
, i.e. that is a closed congruence relation on A.
This suggests the following
Theorem 19.1.5 Let A be any nitary partial algebra, and let a, b A be
any two elements of the same sort, say s. Then the following statements are
equivalent:
(i) (a, b)
c
.
(ii) The so-called principal congruence relation
(a,b)
:= Con
A
(a, b) is
closed.
December 8, 2002
19 Congruence relations revisited 441
(iii) Each unary algebraic partial function of A is dened on a iff it is dened
on b.
(iv) For M := s each unary term (M : t) F(M, TAlg(S
A
)) satises
a dom(M : t)
A
A
iff b dom(M : t)
A
A
.
Proof The equivalence of (i) and (ii) is obvious from what has already been
said in 19.1.1. Because of 19.1.4 (iii) is an immediate consequence of (i). Thus
it remains only to show that (iii) implies (i):
For this purpose we dene
0
:= (a, b) [ a and b are elements of A of the
same sort, and each algebraic partial function is dened on a iff it is dened
on b .
From what we have shown so far, we can infer that
c

0
. The Theorem
will be proved, when we have shown that
0
is a closed congruence relation
on A (since
c
is the greatest one):
It is obvious from the denition that
0
is reexive and symmetric; and
transitivity of
0
is not less obvious. Hence
0
is an equivalence relation.
Because of Proposition 19.1.4 it is therefore sucient to show that
0
is
compatible with all unary algebraic partial functions of A: Therefore we
assume (a, b)
0
and p APF
s
(A; S), where s is the sort of a and b,
moreover consider q APF
s
(A; S), where s

is the output sort of p. From


subsections 15.2 and 15.5 we know that the composition q p is then a unary
algebraic partial function, too. Thus we have by the assumptions on
0
:
p(a) domq iff a domq p iff b domq p iff p(b) domq,
showing that also (p(a), p(b))
0
, completing the proof that (iii) implies (i).
The equivalence between (iii) and (iv) is an immediate consequence of the
close relationship between algebraic partial operations and partial term op-
erations (cf. subsection 15.5).
The important role that is played by unary algebraic partial functions in
connection with closed congruence relations is also stressed by the following
result which is the analogue for closed congruence relations to the corre-
sponding result for total algebras (cf. Gr atzer [G68], 10, Theorem 3).
December 8, 2002
442 Chapter IV Special Topics
Theorem 19.1.6 Let A be any nitary partial algebra, and (a, b)
c
(A).
Then for any two elements x and y of A of the same sort, say s, the following
statements are equivalent:
(i) (x, y)
(a,b)
.
(ii) There exist a natural number n, a sequence x = z
0
, z
1
, . . . , z
n
= y of
elements of A of the same sort s, and a sequence p
0
, . . . , p
n1
of unary
partial algebraic functions of A such that p
i
(a), p
i
(b) = z
i
, z
i+1
, for
i = 0, 1, . . . , n 1.
See Figure 19.1 in 19.1.8.(i) below for an illustration.
Proof Because of 19.1.3, (i) is trivially implied by (ii), since a (closed) con-
gruence relation is transitive and compatible with all unary partial algebraic
functions. In order to show that (i) implies (ii) it is sucient to show that
the relation
:= (x, y) A
2
[ x and y satisfy (ii)
is a congruence relation on A, namely closedness will then follow from the
facts that (a, b) is contained in
c
(A) and that (p(a), p(b)) belongs to
(a,b)
for each unary algebraic partial function p of A, i.e. all of is contained
in
(a,b)
and therefore =
(a,b)
.
The formulation of the denition of immediately implies that is sym-
metric and transitive (one just reverses or composes the sequences under
consideration). Reexivity of follows from the fact that each element c
of A induces a total unary algebraic function for each input sort with con-
stant value c. Hence is an equivalence relation.
Because of 19.1.4 and our observations so far it is now sucient to show the
substitution property (i.e. compatibility with the algebraic structure) only
for unary algebraic partial functions:
Thus let p be any unary algebraic partial function of A, and assume (u, v)
, u domp. By the denition of there exists a sequence u = z
0
, . . . , z
n
=
v, with the associated sequence of unary partial algebraic functions p
0
, . . . , p
n1
,
such that z
i
, z
i+1
= p
i
(a), p
i
(b), 0 i n 1.
Notice that, for each i, z
i
domp, since u domp, and (z
i
, z
i+1
)
c
(0 i n 1) apply 19.1.4. We have to show that (p(u), p(v)) . But
this is immediately realized because of the sequences p(u) = p(z
0
), . . . , p(z
n
) =
p(v) and p p
0
, . . . , p p
n1
, which then satisfy p(z
i
), p(z
i+1
) = (p p
i
)(a), (p p
i
)(b),
0 i n 1.
December 8, 2002
19 Congruence relations revisited 443
gure uncompleted
Figure 19.1: f19.1
Since the supremum of closed congruence relations in a nitary partial
algebra is the equivalence theoretical one, we immediately get the
Corollary 19.1.7 Let A be any nitary partial algebra, and
c
. Then
the following statements are equivalent for any c, d A (of the same sort):
(i) (c, d) Con
A
.
(ii) There exist a natural number n, a sequence c = z
0
, z
1
, . . . , z
n
= d
of elements of A of the same sort s, pairs of elements (a
i
, b
i
) ,
and unary algebraic partial functions p
i
of A, 1 i n, such that
p
i
(a
i
), p
i
(b
i
) = z
i1
, z
i
, for i = 1, . . . , n.
Remarks 19.1.8 (i) The description of a principal closed congruence re-
lation
(a,b)
given by 19.1.6 can be visualized by Figure 19.1.
(ii) The above results show the important role that is played by unary alge-
braic partial functions in connection with closed congruence relations,
and the equivalences in the results above also imply that this will no
longer be true when we drop the assumption of closedness. Actually
there cannot be given in general an upper bound below the cardinal
dimension for the cardinality of the arities of algebraic partial functions
needed to describe compatibility of arbitrary congruence relations:
As a simple example consider the following homogeneous partial alge-
bra A of type := (n)
n
0
\{0}
, i.e. we have =
n
[ n
0
0 , and
A :=

a
n
1
, . . . , a
n
n
, b
n
1
, . . . , b
n
n
, c
n
, d
n
[ n
0
0 ; nally we assume
graph
n
:= ((a
n
1
, . . . , a
n
n
), c
n
), ((b
n
1
, . . . , b
n
n
), d
n
), for each n
0
0.
When all the elements listed for A are distinct, then it is easy to see
that there are equivalence relations on A, for which one needs algebraic
partial functions of arbitrary high nite arity in order to show that they
are congruence relations on A. For instance consider the equivalence
relation on A generated by
:= (a
n
j
, b
n
j
), (c
n
, d
n
) [ 1 j n, and n
0
0 even .
December 8, 2002
444 Chapter IV Special Topics
(iii) A characterization of
c
(A) dierent from the one in 19.1.5 has been
given by I.Kupka in his habilitation thesis (see [Ku77]), where the
main line of the ideas and representations in this connection are correct,
but the proofs and Lemma 2.3.(21) there contain some incorrect parts;
a correct version of this description is contained in [Si82] (see also the
appendix of [Si84]), while a third description has been communicated
by Kupka to the author, but we do not know whether or where it will
be published. Since the version in [Ku77] needs additional replacement
operators and is somewhat more involved, we do not give here any more
details, since we hope that Theorem 19.1.5 will serve most purposes.
From section 2 we already know that the congruence lattice of any ni-
tary partial algebra is algebraic, and that the closed congruence relations
form a principal ideal in it. We conclude this subsection with a general
representation theorem in this connection, which has been discovered inde-
pendently by J.Berman (cf. [Bm71]) and the author (cf. [B71a]) for the case
of homogeneous partial algebras. As a preparation we need the following
Lemma 19.1.9 Let A be some nitary total algebra, let be any congruence
relation of A, and dene B := (A, (
A
)

(id
C
)
CA/
). Then one has
(i) Cong A = Cong B, and
(ii) Cong
c
B is the principal ideal of Cong A generated by .
Proof Since unary partial operations which behave like the identity on
their domain in the corresponding phylum, do not inuence compatibility
of an equivalence relation with the algebraic structure, the rst statement is
obvious. But the second statement is not less trivial: Since all the operations
of A are total, closedness of a congruence relation of B only depends on the
additional operations of B. Let

be any congruence relation of B. If

, and C A/, then for any (a, b)

one has a domid


C
iff b domid
C
, showing that

is closed; but in case that

, there
is (a, b)

. Let a C A/; then, by assumption, b / C, i.e.


a domid
C
, but b / domid
C
, showing that

is not closed.
Since each algebraic lattice is isomorphic to the congruence lattice of some
nitary total (homogeneous) algebra (see [G68] for references and a proof,
and the second edition of [G68] for even more references on more simple
proofs and additional results), we immediately get from 19.1.9 the
December 8, 2002
19 Congruence relations revisited 445
Theorem 19.1.10 Let L be any algebraic lattice, and let I be any principal
ideal of L. Then there exists a nitary homogeneous partial algebra, say A,
such that L is isomorphic to the congruence lattice of A, and the restriction
to I of this isomorphism is a bijection from I onto the set of all closed
congruence relations of A.
19.2 P-reecting homomorphisms and P-closed con-
gruence relations
19.2.1 In subsection 9.2 we have introduced the notion of a P-reecting
homomorphism as a concept which combines those of ordinary and closed
homomorphisms in a common generalization. We have the feeling that such
a concept might become of interest in connection with some questions in
computer science, e.g. when one tries to dene a concept of reduction for
automata which are given as partial automata with states, input and output
sorts, and where for reduction only such partial term operations or functions
are of interest, which have as a resulting sort an output sort one might
even restrict the consideration to such algebraic partial functions, whose
variables only take states as values. As we shall see in the next section, P-
reecting homomorphisms are also important in connection with subdirect
representations in E-varieties.
We recall that for a given set P of terms homomorphisms are called P-
reecting iff they reect all terms in P here reection of a term, say
of (M : t), means reection of the TE-statement (M : t
e
= t). Thus the
properties of homomorphisms reecting a given set of formulas as described
in subsection 9.1 especially apply in this case.
In this subsection we want to continue the study of P-reecting homomor-
phisms especially by considering the congruence relations which they induce.
For this purpose we introduce the
Denition 19.2.2 Let P be any set of terms for some type , and let be
any congruence relation of some partial algebra A of the same type. Then
is called a P-closed congruence relation of A, iff the natural projection
nat

: A A/ is a P-reecting homomorphism.
Cong
P
A designates the set of all P-closed congruence relations of A.
December 8, 2002
446 Chapter IV Special Topics
Proposition 19.2.3 Cong
P
A is always an initial segment of Cong A, the
congruence lattice of A. And any full, P-reecting and surjective homo-
morphism g : A B induces a bijection between the set of all P-closed
congruence relations of A which contain ker g (on one side) and Cong
P
B
(on the other side).
Proof Let be any P-closed congruence relation of A, and assume

.
Then there is a homomorphism g : A/

A/ such that nat

= g nat

(cf. the Homomorphism Theorem). Since g is a homomorphism and therefore


preserves all terms, Lemma 9.1.7.(ii) implies that nat

is P-reecting. The
surjectivity of nat

implies that g is P-reecting, too; therefore ker g = /

is P-closed. The fact that the composition of P-reecting homomorphism is


again P-reecting then implies the rest of the above proposition.
Proposition 19.2.4 Let A be any partial algebra of cardinal dimension d.
Then Cong
P
A is d

-inductive for d

:= max#
0
, d

.
Proof Let R Cong
P
Abe any d

-directed subset. Then L := (A/, nat


,
:
A/ A/

R
is a d

-directed system of P-reecting homomorphisms.


Let f
0
: A
lim

L be the induced homomorphism from A = A/, then


subsections 4.4 and 11.3 imply that f
0
is full and P-reecting and it is easily
seen that ker f
0
=

[ R, i.e. ker f
0
is P-closed.
Remark 19.2.5 An internal characterization of P-closed congruence rela-
tions is not as simple as one might think at the rst glance, where one would
hope that the following concept might work:
Let P be any set of terms, A a partial algebra of the corresponding type,
and let be a congruence relation of A. Let us call weakly P-closed iff , for
every (M : t) P, and for every a, b A
M
, which satisfy a dom(M : t)
A
and (a(m), b(m)) for each m M, one has b dom(M : t)
A
.
Then it is easy to realize that every P-closed congruence relation is weakly
P-closed, but the converse is not true, as the following simple example shows:
Let := (1, 1), := ,

, A := x, x, y,

, and A the relative subalge-


bra on A of the homogeneous Peano algebra with Peano basis x, y. Finally
let := Con
A
x, y and P := z, z, z

(for simplicity we omit the


one-element arities!). Then it is easy to realize that is weakly P-closed,
but not P-closed, since (

)
A/
is dened on x

, but (

)
A
is not dened
on x.
December 8, 2002
19 Congruence relations revisited 447
gure uncompleted
Figure 19.2: f19.2
gure uncompleted
Figure 19.3: f19.3
This example and especially the corresponding Figure 19.2 show, what
an internal characterization of P-closed congruence relations has to look
like. Roughly speaking this can be described as follows: If one represents
each term in P by its corresponding tree, then one has to split each node
into two new ones, thus disconnecting the tree (now a forest) at this place,
labeling the new leaf by a new variable of the corresponding sort, and leav-
ing the old label at the new root. We show this by a more complex ex-
ample: S := s, s

, s

, graph
S
:= ((s), s

), graph
S
:= ((s

, s

), s),
graph
S
:= ((s

, s, s

), s

), and where we want to consider for variables x


of sort s, y of sort s

, and z of sort s

the term (x, x

, y, y

, z, z

, z

:
xx

yy

) =: (M, t). Figure 19.3 then shows rst the tree


representation of (M : t) on the left, and on the right the corresponding
-split forest whose interpretation (including the -relationships indicated
by ellipses a root represents a whole element) leads in the factor to the
existence of the corresponding term operations (see 19.2.6 below):
We do not present here an explicit formulation for the general situation,
since we have the feeling that this is better illustrated in connection with the
above examples rather than by a clumsy formal description: is (M : t)-
closed iff : whenever the above variables are evaluated by two sequences of
elements of A, say (a, a

, b, b

, c, c

, c

) and (a

, a

, b

, b

, c

, c

, c

) such that
these sequences are componentwise -congruent, and if all the elements of
the -split forest are interpreted with respect to the rst sequence, then they
are interpreted with respect to the second sequence, and for all these latter
interpretations the root is -congruent with the root of the interpretation
connected with the rst sequence.
In the case that P is closed with respect to subterms, we can give more
easily a more formal description of P-closed congruence relations. (Recall
that we have already shown in 9.2.6 that the set of terms reected by a ho-
momorphism need not be closed with respect to the formation of subterms!)
December 8, 2002
448 Chapter IV Special Topics
Proposition 19.2.6 Let P be a set of terms which is closed with respect
to the formation of subterms, and let A be any partial algebra of the type
under consideration. Then for any congruence relation of A the following
statements are equivalent:
(i) is P-closed.
(ii) If (M : p) = (M : ( p
k
[ k K

)) P is any term, if a dom


A
is any sequence, and if w : M A is any assignment such that w
domp
A
k
and (a(k), p
A
(w)) for every k K

, then ( p
A
k
(w) [ k
K

) dom
A
.
Proof From what has already been observed in 19.2.5 it is easily seen that
(i) implies (ii) (since P is closed with respect to subterms!). The converse
implication is proved in showing by algebraic induction that (ii) implies (),
where:
()
If (M : p) P and nat

w domp
A/
, then w domp
A
(for every w A
M
).
If p is a variable, then (M : p)
A
and (M : p)
A/
are at the same time either
empty or total, and () is trivially satised. Now assume that p is given
as in (ii), and assume nat

w dom(M : p)
A/
for some w A
M
. Then
nat

w dom(M : q)
A/
for every subterm (M : q) of (M : p), and in
particular nat

w dom(M : p
k
)
A/
for every k K

. By the induction
hypothesis we therefore have w domp
A
k
for every k K

. Moreover, since
nat

is full, there exists some a dom


A
such that nat

a = ( nat

(p
A
k
(w)) [
k K

). Because of (ii) we now get ( p


A
k
(w) [ k K

) dom
A
, and this
implies ().
Since () has now been proved for every (M : p) P, is shown to be
P-closed.
Proposition 19.2.7 Let K be any E-variety, let D := Eeq K be the set of all
terms with d
&
-small sets of variables in X inducing total term operations on
each K-algebra, and let P D be any set of terms. Then each homomorphism
between any two K-algebras is P-reecting, and in particular we have for any
K-algebra A that Cong
P
A = Cong A, i.e. each congruence relation of A is
P-closed.
Proof This is an immediate consequence of the fact that the TE-statements
corresponding to the terms in D are already valid in each K-algebra.
December 8, 2002
19 Congruence relations revisited 449
19.2.8 (Concluding Remarks)
(i) From the results in section 9 quite a few conclusions can be drawn
for P-reecting homomorphisms. For this purpose let us dene for
a given set H of homomorphisms TermH to be the set of all terms
(with arities less than d
&
contained in a suitable set X of variables)
which are reected by all homomorphisms within H, and for a given
set P of terms let Mor P be the set of all homomorphisms reecting
all terms in P. It is easily seen that Term and Mor are the operators
of a Galois connection, and that TermMor and MorTerm are closure
operators on their corresponding underlying classes. As we have already
observed TermMor P is usually not closed with respect to the formation
of subterms, but it trivially contains all terms (M : x) for any variable
x M of any given sort, and for any subset M of the set X of variables
under consideration (notice that one starts with assignments taking
their values in the domain of a homomorphism!). Moreover, the results
in section 9 show that TermMor P is always closed with respect to
substitution, and if it is closed with respect to subterms, then for any
xed subset M X the terms (M : p) TermMor P form the carrier
set of a relative subalgebra of F(M, TAlg(S)) freely generated by M.
(ii) In the homogeneous case let X be a set of cardinality d

:= maxd, #
0
,
and consider all terms as elements of F(X, TAlg()). Then P

:=
TermMor P for some subset P F(M, TAlg(S)) which is closed with
respect to subterms , iff P

is the carrier set of a free relative subal-


gebra of F(M, TAlg(S)) freely generated by X, and then at the same
time we get that Mor P

= Mor P is the set of all ind


P

X-perfect homo-
morphisms in Alg(). A similar result can be formulated in the het-
erogeneous case with respect to the TE-variety dened by TermMor P,
as long as P is closed with respect to subterms.
December 8, 2002
450 Chapter IV Special Topics
20 Subdirect representations of partial alge-
bras
20.1 The general concept and remarks
20.1.1 In the theory of total algebras the concept of subdirect representation
of an algebra has been introduced in order to have a description of algebras
by smaller and hopefully simpler building blocks. Despite the fact that one
has the theorem of G.Birkhoff that every algebra is a subdirect product of
subdirectly irreducible algebras, the theory of subdirect representations has
rather more theoretical than practical importance in most instances, since
for only very few varieties one denitely knows all the subdirectly irreducible
members. On the other hand the fact that a variety possesses up to isomor-
phism only a set of subdirectly irreducible members, i.e. that it is residually
small, usually allows one to draw quite a few other conclusions.
In this section we want to generalize the concept of subdirect represen-
tation to partial algebras. But there are many dierent properties of ho-
momorphisms which could play a role (and actually will do); e.g. there is
a concept of subdirect representation of partial algebras contained in the
book of R.S.Pierce [Pi68], another one in an exercise of G.Gr atzers
book [G68] (cf. exercise 96 to Chapter 3); both of them will be considered
later.
In order to get a more general background we introduce in this subsection
what might be a general concept of subdirect representation. As we have been
told, such a concept has already been investigated by I.Sain in [Sa74] (in
Hungarian).
Denition 20.1.2 Let C be any category with products, and let c and /
be classes of epimorphisms respectively monomorphisms of C which are each
closed with respect to composition and both contain all isomorphisms of C.
A family ( e
i
: A B
i
[ i I ), I ,= , of C-morphisms is called an (c, /)-
subdirect representation of an object A of C, and A is called an (c, /)-
subdirect product of the family (B
i
)
iI
, if each morphism e
i
belongs to c
(i I), and if moreover the induced C-morphism m : A

( B
i
[ i I )
from A into the product of the family ( B
i
[ i I ) belongs to /. A is
called (c, /)-subdirectly irreducible, if each (c, /)-subdirect representation
of A contains an isomorphism (i.e. if there is an i I such that e
i
is an
isomorphism).
December 8, 2002
20 Subdirect representations of partial algebras 451
Remarks 20.1.3 (i) The above concept may seem to be too general, but
when one considers existing or so far proposed concepts of subdirect
representations in connection with relational systems or partial alge-
bras, then it seems to be the most narrow one covering all of them.
In a so-called concrete category, i.e. when there is a full and faithful
functor into the category Set of all sets with ordinary mappings as
morphisms as it is the case with Alg(S) or TAlg(S) , then it is quite
customary to restrict the class c of epimorphisms to be contained in
the class of all morphisms whose corresponding value in Set, i.e. whose
underlying mapping, is surjective; and we shall do so, too, in what
follows. But since we know that each epimorphism starting from some
partial algebra A can be characterized by an initial closed congruence
relation of F(A, TAlg(S)), it seems by no means necessary to make this
restriction in order to get satisfactory results (observe that in [BSi85]
also epimorphisms in general are treated, and cf. [Wl69]).
(ii) In the case of total algebras the most powerful tool in connection with
subdirect representations and the decision for subdirect irreducibil-
ity are congruence relations and in the case of relational systems
H.E.Pickett shows in [Pc64] that here equivalence relations also play
a fundamental role. Our results and observations below will follow the
usual pattern as far as possible.
(iii) For partial algebras the discussion of (c, /)-subdirect representations
will usually be quite simple, as long as /is chosen to be the class of all
monomorphisms. If then c is a class of epimorphisms reecting some
given set, say F, of formulas, then by the results of subsection 9.1
/ can without loss of generality be also restricted to the class of all
monomorphisms reecting F, since the embedding morphism will in-
herit this reection property. The situation will become more involved,
when / is supposed to satisfy additional requirements e.g. when the
monomorphisms have to be full or P-reecting, and P-reection is not
required for the elements of c.
(iv) As in the case of total algebras the proof of the existence of enough
(c, /)-subdirectly irreducible objects in order to have the result
that each partial algebra has an (c, /)-subdirect representation with
(c, /)-subdirectly irreducible objects needs the Axiom of Choice (as
a matter of fact, this existence is even equivalent with the Axiom of
December 8, 2002
452 Chapter IV Special Topics
Choice, as has been shown by G.Gr atzer in [G85]). Therefore we
shall have to restrict our considerations to nitary similarity types.
20.2 Classes of full and surjective homomorphisms
arbitrary monomorphisms
Lemma 20.2.1 Let be a nitary similarity type, let / be the class of all
injective homomorphisms in Alg(S), and let c be a composition closed class
of full and surjective homomorphisms in Alg(S) such that for each partial
algebra A the class of all c-morphisms starting from A is represented up to
isomorphism by an inductive set Cong
E
A of congruence relations of A, and
such that c is right cancellable with respect to c-morphisms (i.e. g f c
and f c imply g c). Then one has for A Alg(S):
(i) Let e : A B be any c-morphism, then
(a) e induces an isomorphism, say e

, (with respect to set theoretical


inclusion) between Cong
E
A [ ker e and Cong
E
B.
(b) B is (c, /)-subdirectly irreducible, iff
(SI)

Cong
E
A [ ker e , = ker e, whenever the index set
for the intersection on the left side is non-empty;
ker e will then (when (SI) is satised) be called a completely c-meet
irreducible congruence relation of A.
(ii) A is (c, /)-subdirectly irreducible iff
A
satises (SI), i.e. iff
A
is
completely c-meet irreducible.
(iii) Let a, b A be elements of the same sort, a ,= b. Then there exists
a maximal c-congruence relation, say
E
(a, b), in Cong
E
A such that
(a, b) /
E
(a, b); and
E
(a, b) is completely c-meet irreducible.
20.2.2 Notice that the congruence relation
E
(a, b), which we consider in
20.2.1.(iii), is in general not uniquely determined by (a, b), but that the rest
of the statement of 20.2.1.(iii) applies to each such congruence relation. And
if we use later on the notation =
E
(a, b), then this means that could
serve as
E
(a, b).
December 8, 2002
20 Subdirect representations of partial algebras 453
Proof Ad (i): Let e : A B be any c-morphism (thus B is isomorphic to
A/ker e). In order to prove (a) let us rst consider Cong
E
A such
that ker e . Since c is assumed to be right cancellable with respect
to c-morphisms, nat
/ker e
also belongs to c, i.e. e

() := /ker e
Cong
E
B. This shows that e

is the restriction of the isomorphism be-


tween the corresponding intervals of congruence relations as induced by
the Homomorphism Theorem, and therefore e

preserves set theoretical


inclusion and is injective. Surjectivity follows from the fact that c is as-
sumed to be closed with respect to composition; thus, for

Cong
E
B,
we get that nat

e belongs to c, and its kernel is the preimage of

with respect to e

.
Now let us prove (b): To begin with, let ker e be completely c-meet
irreducible, and let ( e
i
: B C
i
[ i I ) be an (c, /)-subdirect
representation of B. Let C designate the direct product of the family
(C
i
)
iI
, and let m : B C be the product homomorphism induced by
the family (e
i
)
iI
. By assumption
B
= ker m =

ker e
i
[ i I . It
is then easy to realize that e
1
(
B
) = ker e =

ker e
i
e [ i I ,
and because of (SI) this can only happen when ker e = ker e
i
e for
some i

I. This, however, implies ker e


i
=
B
, and e
i
has to be an
isomorphism, as it is in addition full and surjective. Thus B is (c, /)-
subdirectly irreducible.
Next, let us assume that ker e does not satisfy (SI). It is then easy to
see that ( nat
/ker e
: B A/ [ Cong
E
A, and ker e ) yields an
(c, /)-subdirect representation of B, which contains no isomorphism
(cf. (a)). Thus the equivalence of (i)(b) has been proved.
Ad (ii): (ii) is just a corollary of (i).
Ad (iii): Let
E
(a, b) be the set of all elements of Cong
E
A avoiding (i.e. not
containing) (a, b). Then
A

E
(a, b), since c contains all isomor-
phisms. Moreover, the union of any non-empty chain (or directed set)
in
E
(a, b) also avoids (a, b) and belongs to Cong
E
A, since Cong
E
A is
assumed to be inductive. Thus
E
(a, b) is itself inductive and therefore
contains maximal elements, one of which will be denoted by
E
(a, b).
It now easily follows that
E
(a, b) satises (SI), since each element of
Cong
E
A, which properly contains
E
(a, b) also contains (a, b), and the
intersection of all such congruence relations (if there are any) will also
contain (a, b) and has therefore to be dierent from
E
(a, b).
December 8, 2002
454 Chapter IV Special Topics
As an immediate consequence we get the following generalization to par-
tial algebras of G.Birkhoffs theorem on subdirect representations:
Theorem 20.2.3 Let / be the class of all injective homomorphisms of a
species Alg(S) with nitary operations, and let c be a composition closed class
of full and surjective homomorphisms such that c contains all isomorphisms,
and is right cancellable with respect to c-morphisms, and, for each A
Alg(S), Cong
E
A is inductive.
Then every Alg(S)-algebra A has an (c, /)-subdirect representation of
(c, /)-subdirectly irreducible partial algebras.
Proof According to Lemma 20.2.1.(iii) there exists for any partial alge-
bra A and for any distinct elements a, b A of the same sort a maximal
completely c-meet irreducible congruence relation
E
(a, b) avoiding (a, b).
If
E
(a, b) =
A
, then A is itself (c, /)-subdirectly irreducible. Other-
wise we get


E
(a, b) [ (a, b) A
2
, a ,= b, v
A
(a) = v
A
(b) =
A
, and
( nat

E
(a,b)
: A A/
E
(a, b) [ (a, b) A
2
, a ,= b, v
A
(a) = v
A
(b) ) is an
(c, /)-subdirect representation of A of (c, /)-subdirectly irreducible par-
tial algebras.
Corollary 20.2.4 (1) Theorem 20.2.3 applies in particular in the case of
nitary partial algebras, when / is the class of all injective homomorphisms:
/ := o
w
, and c is the class of all full and surjective
(i) homomorphisms: c := H
f
; or
(ii) P-reecting homomorphisms (for any set P of terms): c := H
fP
(in
this case we may even take / := o
wP
, the class of all P-reecting and
injective homomorphisms); or
(iii) closed homomorphisms: c := H
c
(in this case we may even choose
/ := o
c
, the class of all closed and injective homomorphisms).
Proof For each of the three cases above we already know that c is closed
with respect to composition and contains all isomorphisms; moreover, if g f
and f belong to c, then 2.4.5.(iii).(c) implies that g is full, and that g is
surjective, is trivial; since f preserves all TE-statements and is surjective,
9.1.7.(iii) implies that g reects the same TE-statements, which are reected
by g f: Thus c is also right cancellable with respect to c-morphisms, and
therefore satises the assumptions of Theorem 20.2.3. That Cong
E
A is in-
ductive for each A has been proved earlier for each of the three cases. The
statements about / follow from 20.1.3.(iii).
December 8, 2002
20 Subdirect representations of partial algebras 455
Thus we have in the case of partial algebras the following three Birkho-
Theorems for subdirect representability:
Corollary 20.2.5 (2) Let Alg(S) be a nitary species of partial algebras.
Then we have for each set P of terms:
(i) Each Alg(S)-algebra is the (H
f
, o
w
)-subdirect product of (H
f
, o
w
)-subdirectly
irreducible Alg(S)-algebras.
(ii) Each Alg(S)-algebra is the (H
fP
, o
wP
)-subdirect product of (H
fP
, o
wP
)-
subdirectly irreducible Alg(S)-algebras.
(iii) Each Alg(S)-algebra is the (H
c
, o
c
)-subdirect product of (H
c
, o
c
)-subdirectly
irreducible Alg(S)-algebras.
Remark 20.2.6 It is our opinion that the three concepts of (c, /)-subdirect
representation presented above are among the most important ones for par-
tial algebras. The result on (H
f
, o
w
)-subdirect products is already contained
in the book of R.S.Pierce on abstract algebras (cf. [Pi68], Chapter 2). The
importance of the results in 20.2.4.(ii) and (iii) will follow from the follow-
ing results about their relationships to E- and ECE-varieties, respectively.
It is easily seen that for any E-variety K and for the case that c is the
class of all full and surjective homomorphisms all the (H
f
, o
w
)-subdirectly
irreducible partial algebras needed for an (H
f
, o
w
)-subdirect representation
of some K-algebra belong to K. But it is also easy to realize that not each
(H
f
, o
w
)-subdirect product of (H
f
, o
w
)-subdirectly irreducible K-algebras be-
longs to K. This is a point where P-closedness is needed for suitable sets P of
terms, as we shall see below, while closedness is an adequate concept for ECE-
varieties in general. These concepts have also been studied and applied in
the thesis of M.Siegmund-Schultze ([Si84]), where they have been needed
in order to transfer the results of W.Taylor on residually small varieties
and equational compactness of total algebras to partial algebras.
Denition 20.2.7 Let K be any class of partial algebras, and let c and /
be suitable classes of epi- and monomorphisms. Then we dene:
(i) si
(E,M)
K is the class of all (c, /)-subdirectly irreducible K-algebras.
(ii) Ps
(E,M)
K is the class of all (c, /)-subdirect products with respect to
families of c-morphisms into K-algebras.
December 8, 2002
456 Chapter IV Special Topics
Lemma 20.2.8 Let K be any E-variety, let D := deg(K) := Eeq K be the
completeness degree of K, i.e. the set of all (admissible) terms which induce
total term operations in each K-algebra, let P D be a set of terms, and
let m : A B be an injective and P-reecting homomorphism into some
K-algebra B. Then A also belongs to K.
Proof Let (M : t
e
= t

) be any E-equation valid in K, and let w : M A be


any assignment. Since by assumption (M : t) and (M : t

) belong to P, since
both terms are interpreted in B with respect to m w, and since m reects
these terms, both of them are interpreted in A with respect to w; moreover
they are equally interpreted, since this is the case in B, and m is injective.
Thus the E-equation is also valid in A, i.e. A belongs to K.
Theorem 20.2.9 Let K be any E-variety of nitary partial algebras. Then
we have for any set P of terms which contains D = Eeq K:
K = Ps
(H
fP
,S
wP
)
(si
(H
fP
,S
wP
)
K),
i.e. for D P any E-variety K is just the class of all (H
fP
, o
wP
)-subdirect
products of (H
fP
, o
wP
)-subdirectly irreducible K-algebras.
Proof It is obvious that K is contained in Ps
(H
fP
,S
wP
)
(si
(H
fP
,S
wP
)
K) (be-
cause of 20.2.5.(ii), and since an E-variety is closed with respect to arbitrary
homomorphic images and therefore also with respect to P-reecting homo-
morphic images). Now 20.2.8 shows that each domain of a P-reecting injec-
tive homomorphism into some K-algebra belongs to K, if D P, therefore
also the other inclusion follows, since K is closed with respect to products of
K-algebras.
Theorem 20.2.10 Let K be any ECE-variety of nitary partial (S-) alge-
bras. Then
K = Ps
(Hc,Sc)
(si
(Hc,Sc)
K),
i.e. each ECE-variety K is the class of all (H
c
, o
c
)-subdirect products of
(H
c
, o
c
)-subdirectly irreducible K-algebras.
Proof Since K is closed with respect to closed homomorphic images, all
the (H
c
, o
c
)-subdirectly irreducible partial algebras needed for an (H
c
, o
c
)-
subdirect representation of a K-algebra belong to K. And since K is closed
with respect to subalgebras (and therefore with respect to the domains of
closed and injective homomorphisms into K-algebras), all (H
c
, o
c
)-subdirect
products of families of K-algebras belong to K. From these observations the
equality stated above follows.
December 8, 2002
20 Subdirect representations of partial algebras 457
Remark and Problem 20.2.11 In considering the last two theorems one
can realize that in connection with subdirect representations ECE-varieties
have the smoothest representation theorem when one uses (H
c
, o
c
)-subdirect
representations. As we have already mentioned above the results of M.Siegmund-
Schultze in [Si84] show that this is really a very suitable concept and is
closely connected with a concept of equational compactness, where one not
only allows E-equations, but also negations of TE-statements. This observa-
tion also backs our opinion that ECE-varieties are an adequate generalization
to partial algebras of the concept of variety for total algebras. But we do
not want to deny that the concept of hep-varieties (which are in general
very special QE-varieties) as introduced by H.Reichel (cf. Part II) is some-
times more useful, and still has to a great extent a nice theory. We did not
yet consider the question of what might be the most convenient concept of
(c, /)-subdirect representation in connection with hep-varieties if there
is any at all , and we leave this investigation as an open problem for the
reader.
20.3 The cases of full respectively closed monomor-
phisms
20.3.1 The cases which we want to discuss in this subsection are analogues
to a subdirect representation theorem for relational systems of H.E.Pickett
(cf. [Pc64]), where we now have to assume that / is the class of all full (or
even closed) and injective homomorphisms, while c is a class satisfying the
same assumptions as in 20.2.1 or 20.2.3.
Since the monomorphisms are now supposed to be full (or even closed),
we have to separate each sequence which is not contained in the graph (do-
main) of a fundamental operation, say
A
, from the graph (domain) of this
operation by a suitable c-congruence relation. As a preparation we dene:
Denition 20.3.2 Let A be any set, E an equivalence relation on A, and
a K-ary relation on A. Then we introduce
E := b A
K
[ there is a such that (a(k), b(k)) E for each k K;
i.e. E consists of all K-sequences of A which are E-equivalent with
some sequence in .
December 8, 2002
458 Chapter IV Special Topics
Quite useful in the sequel will also be the following
Lemma 20.3.3 Let c and c

be classes of epimorphisms and / and /

be
classes of monomorphisms closed with respect to composition and each con-
taining all isomorphisms. Moreover assume that c

c and /

/. Then
each (c, /)-subdirectly irreducible object will also be (c

, /

)-subdirectly ir-
reducible.
Proof Each (c

, /

)-subdirect representation of an object A is also an


(c, /)-subdirect representation of A, by assumption, and if A is (c, /)-
subdirectly irreducible, then it contains an isomorphism; thus A is also
(c

, /

)-subdirectly irreducible.
Lemma 20.3.4 Let S be a nitary partial sort algebra, let / be the class
of all full (respectively closed) and injective Alg(S)-homomorphisms, and let
the class c of epimorphisms have the same properties as in 20.2.1 or 20.2.3.
Then a family ( e
i
: A B
i
[ i I ) of c-morphisms is an (c, /)-subdirect
representation of A iff (SR) and (SRf ) (in the closed case (SR) and (SRc))
below are satised, where
(SR)

ker e
i
[ i I =
A
,
(SRf )

(graph
A
) (ker e
i
) [ i I = graph
A
for each .
(SRc)

(dom
A
) (ker e
i
) [ i I = dom
A
for each .
Proof We only treat the case of full and injective homomorphisms; the
other case is similar. It is clear that (SR) is eqivalent to the fact that the
induced homomorphism e : A

( B
i
[ i I ) is injective. Furthermore it
is easy to see that the set described in (SRf) is equal to

(e
K
i
e
i
)
1
[graph
B
i
] [ i I ,
observing that all the e
i
are full. And when (SR) is satised, then the equality
of the above intersection with graph
A
( ) is equivalent with the fact
that e is full.
December 8, 2002
20 Subdirect representations of partial algebras 459
Lemma 20.3.5 Let E and M satisfy the same assumptions as in Lemma 20.3.4,
let A be any partial algebra, and let Cong
E
A be arbitrary.
(i) Assume that for some we have (a, a) (A
K
A) graph
A
(respectively a A
K
dom
A
). Then there exist maximal congru-
ence relations
E,f,a,a
(respectively
E,c,a
) in Cong
E
A with respect to the
property (a, a) / (graph
A
)
E,f,a,a
(respectively a / (dom
A
)
E,c,a
).
(ii) A/ is (c, o
f
)-subdirectly irreducible iff either =
E
(a, b) for some
(a, b) A
2
, or =
E,f,a,a
for some (a, a) (A
K
A) graph
A
, for
some a A
s
, and for some .
(iii) A/ is (c, o
c
)-subdirectly irreducible iff either =
E
(a, b) for some
(a, b) A
2
, or =
E,c,a
for some a A
K
dom
A
, and for some
.
Proof We only prove the existence of
E,f,a,a
and (ii); the other case is
similar.
Ad (i): As in the proof of 20.2.1.(iii) we have to show that R
E,f,a,a
:=
Cong
E
A [ (a, a) / (graph
A
) is inductive:
Obviously, R
E,f,a,a
is non-empty, since it contains
A
. Let D be any
directed subset of R
E,f,a,a
, and let

be its set theoretical union. Since


Cong
E
Ais inductive,

is an c-congruence. If we had (a, a) (graph


A
)

, then there would exist (b, b) graph


A
and
k
D for each k K

and for k := s

such that (a(k), b(k))


k
(k K

), and (a, b)
s
.
Since D is directed, there would be

D containing all
k
and satis-
fying all the equivalences contrary to the assumption that D R
E,f,a,a
.
Thus R
E,f,a,a
is inductive, and therefore it contains maximal elements,
one of which we may denote by
E,f,a,a
.
Ad (ii): First let us show that in both cases for in (ii) A/ is (c, o
f
)-
subdirectly irreducible: If =
E
(a, b), then this follows from 20.2.1.(iii)
and 20.3.3, since o
f
o
w
. Now assume that =
E,f,a,a
. We already
know from 20.2.1.(i) that we have a bijection e

from the set of all


c-congruence relations of A which contain , onto Cong
E
A/. By as-
sumption on we have (nat

a, nat

(a)) / graph
A/
, but for any

Cong
E
A which properly contains , we have (nat

a, nat

(a))
graph
A/

. Therefore ( nat

/
: A/ A/

Cong
E
A)
cannot yield an (c, o
f
)-subdirect representation of A/, or expressed
December 8, 2002
460 Chapter IV Special Topics
in another way : Each (c, o
f
)-subdirect representation of A/ has to
contain an isomorphism, showing that A/ is (c, o
f
)-subdirectly irre-
ducible.
Conversely, let A/ be (c, o
f
)-subdirectly irreducible, and consider
( nat

/
: A/ A/

Cong
E
A). This cannot be an
(c, o
f
)-subdirect representation of A/. Let C be the direct product of
the A/

occurring in this family of c-morphisms, and let m : A/ C


be the induced product homomorphism. Then if (SR) is violated
=
E
(a, b) for a suitable pair (a, b) A
2
, a ,= b, v
A
(a) = v
A
(b), for
which m(nat

(a)) = m(nat

(b)); or if (SRf) is violated there exists


such that

(graph
A
)

Cong
E
A , = graph
A
.
Hence we have (a, a) (A
K
A) graph
A
such that nat

does
not identify (a, a) with any element of graph
A
, but each nat

(for

Cong
E
A) does. This shows that =
E,f,a,a
.
Corollary 20.3.6 Let A be any nitary partial algebra, and let c be a com-
position closed class of full and surjective homomorphisms, right cancellable
with respect to c-morphisms and containing all isomorphisms, and for which
Cong
E
A is inductive. Then:
(i) A is (c, o
w
)-subdirectly irreducible, iff there is (a, b) A
2
such that
each non-trivial c-congruence relation of A contains (a, b).
(ii) A is (c, o
f
)-subdirectly irreducible, iff either there is (a, b) A
2
such
that each non-trivial c-congruence relation of A contains (a, b), or there
is (a, a) (A
K
A) graph
A
for some such that for every
non-trivial c-congruence relation one has (a, a) (graph
A
) .
(iii) A is (c, o
c
)-subdirectly irreducible, iff either there is (a, b) A
2
such
that each non-trivial c-congruence relation of A contains (a, b), or there
is a A
K
dom
A
for some such that for every non-trivial
c-congruence relation it is true that a (dom
A
) .
Proof (i), (ii) and (iii) follow directly from 20.3.5 and its proof.
Theorem 20.3.7 Let S be a nitary partial sort algebra, and let c be a com-
position closed class of full and surjective homomorphisms, right cancellable
with respect to c-morphisms and containing all isomorphisms, and for which
Cong
E
A is inductive for every partial algebra A. Then one has for every
partial algebra A:
December 8, 2002
20 Subdirect representations of partial algebras 461
(i) A has an (c, o
f
)-subdirect representation of c-morphisms into (c, o
f
)-
subdirectly irreducible partial algebras.
(ii) A has an (c, o
c
)-subdirect representation of c-morphisms into (c, o
c
)-
subdirectly irreducible partial algebras.
Proof The proof is similar to the one of Theorem 20.2.3, except that we
now have to consider families of
E
(a, b) and
E,f,a,a
(respectively
E,c,a
).
Therefore we omit the details.
With the notation from 20.2.7 we then get the
Corollary 20.3.8 Let K be an E-variety of nitary partial algebras. Then
K = Ps
(H
f
,Sc)
(si
(H
f
,Sc)
K).
The argument is similar to the one in 20.2.10.
Remarks 20.3.9 (i) Corollary 20.3.8 shows that the concept of (H
f
, o
c
)-
subdirect representation is very suitable for E-varieties, but it is simi-
larly adequate for all H
f
-closed ECE-varieties. Actually the results of
this subsection provide suitable concepts of (c, /)-subdirect represen-
tation for quite a lot of special concepts of ECE-varieties as discussed
in Chapter III, especially in section 13. But the explicit treatment
of the concepts of (c, /)-subdirect representation as discussed in this
subsection is much more dicult than for the cases treated in subsec-
tion 20.2.
(ii) Notice that for total algebras all the concepts of subdirect representa-
tion considered so far coincide with the usual one for total algebras.
(iii) We have restricted our consideration here to the case when c is con-
tained in the class of all full and surjective homomorphisms of Alg(S)
in order to provide an idea of possibilities of subdirect representations
of partial algebras and of their treatment. In [BSi85] one can also nd
a discussion of the cases when c is the class of all surjective homomor-
phisms (and / is the class of all full and injective homomorphisms)
which was introduced by G.Gr atzer in exercise 96 to Chapter 3
of [G68] , and of the case when c is the class of all epimorphisms
(and / is the class of all closed and injective homomorphisms) this
December 8, 2002
462 Chapter IV Special Topics
c / Quotation
full and surjective hom.s injective hom.s 20.2.5
full and surjective hom.s full and injective hom.s 20.3.7
full and surjective hom.s closed and injective hom.s 20.3.7
P-reecting full and surj. hom.s P-reecting and inj. hom.s 20.2.5
closed and surjective hom.s closed and injective hom.s 20.2.5
surjective hom.s full and injective hom.s *5.8
epimorphisms closed and injective hom.s *6.4
Table 20.1: Positive cases of (c, /)-subdirect
representation theorems; see 20.3.10
case was rst introduced by R.Wille in [Wl69]. In both cases the
congruences on A are no longer sucient to characterize the (c, /)-
subdirectly irreducible components of A, but one has to consider the
A-initial closed congruence relations of F(A, TAlg(S)) and for surjec-
tive homomorphisms only those for which each congruence class has a
non-empty intersection with A.
Table 20.3.10 As a resume of the results in this section combined with
those in [BSi85] we get Table 20.1 for concrete positive cases of Birkho
type results on the existence of (c, /)-subdirect representations with respect
to (c, /)-subdirectly irreducible partial algebras (the numbers refer to the
theorem where the corresponding result is proved, * meaning that this is
a quotation from [BSi85]).
It is not dicult to add further examples, which are easily derivable
with the previous methods and statements, but we think that the examples
presented here will do for most cases.
20.4 Subdirectly irreducible mono-unary partial alge-
bras
Remarks 20.4.1 (i) In order to get a better feeling for the dierent con-
cepts of subdirect irreducibility, we present these concepts in more de-
tail in the most simple case where we have only one unary partial
operation. This is also the only similarity type, where in the total case
all subdirectly irreducible algebras of this type are known (cf. [JoSe67]
December 8, 2002
20 Subdirect representations of partial algebras 463
and [We70]). In the total case all subdirectly irreducible algebras are
known only for very few varieties. The characterizations of all (H
f
, o
w
)-
respectively of all (H
c
, o
c
)-subdirectly irreducible mono-unary partial
algebras was rst given in [Si82], while all the examples presented here
can be found in [BSi85]. Since all subdirect products of total alge-
bras only use total subdirect factors, all subdirectly irreducible total
(mono-unary) algebras will also be subdirectly irreducible for all other
concepts.
(ii) Since for mono-unary partial algebras and for every non-trivial set P
of terms which is closed with respect to subterms every P-reecting
homomorphism is already closed, we will not dicuss P-reection here;
the interested reader may investigate for himself cases of P-reection
where P is not closed with respect to subterms.
(iii) Observe that mono-unary partial algebras can be considered as directed
graphs in the sense of graph theory without multiple edges and with
outdegree 1, namely by taking graph
A
as the corresponding rela-
tion between the elements of A, which then will also be called vertices,
while the elements of graph
A
will be called (directed) edges. Thus to
a mono-unary partial algebra will be attributed some graph theoretical
property, whenever the corresponding graph has this property; espe-
cially such a mono-unary partial algebra will be called connected, when-
ever the corresponding graph is connected (as an undirected graph, i.e.
by forgetting directions of the edges), and a maximal connected part
of a given graph will be called a (connected) component: Notice that
every graph is a disjoint union of its connected components. We shall
also represent a mono-unary partial algebra by its corresponding graph,
whenever necessary.
Notation 20.4.2 Let us give names to some special mono-unary partial
algebras:
(i) In what follows, + will always denote disjoint union of graphs, respec-
tively of the corresponding mono-unary partial algebras. Moreover, n
and m will designate natural numbers (
0
) or (meaning inn-
ity in the sense that a path of the graph has no nite beginning or
end point).
December 8, 2002
464 Chapter IV Special Topics
gure uncompleted
Figure 20.1: f20.1
(ii) For every n
0
and m, m


0
, m, m

/ 0, we consider the
mono-unary partial algebras described by their graphs in Figure 20.1
(notice that C
1
= I
0
).
In the proof of the next theorem one needs the following lemma and its
corollary taken from [Si82].
Lemma 20.4.3 Let A be a mono-unary partial algebra with partial part P
(union of all connected components which are really partial, i.e. contain at
least one partial element) and total part T; then one has the isomorphism
Cong
c
A = Cong
c
P Cong T,
where the right side is considered as the usual direct product of lattices.
Proof (Sketch) The assignment
Cong
c
P Cong T Cong
c
A, (,

is a lattice isomorphism. We omit the somewhat long computation, whose


details are available in the preprint [Si82] (or see the appendix of [Si84]).
From this lemma there follows that the atoms in Cong
c
A are exactly the
closed congruence relations
T
and
P

with atoms in Cong


c
P
and

in Cong T, respectively, and we have the


Corollary 20.4.4 A mono-unary partial algebra A = P

T is (H
c
, o
c
)-
subdirectly irreducible, iff one of the two parts is (H
c
, o
c
)-subdirectly irre-
ducible, and the other one is closed-simple.
Theorem 20.4.5 Up to isomorphism all (c, /)-subdirectly irreducible mono-
unary partial algebras are given in the cases below by the following lists:
(i) Total case:
(A) I
n
, C
n
, C
n
+I
0
, and I
0
+I
0
,
where n

= p
k
, p prime, k
0
(k ,= 0), n
0
.
December 8, 2002
20 Subdirect representations of partial algebras 465
(ii) c := H
f
, / := o
w
: The list (A) and
(B) I

0
, I

1
, I
0
+I

0
, and I

0
+I

0
.
(iii) c := H
c
, / := o
c
: The list (A) and
(C)
I

n
, I

;m,m
, I

;m,m
+I
0
, I

n
+I

n
, I

n
+I
n
,
I

n
+C
p
k, I

n
+I

n
+I
0
, I

n
+C
p
k +I
0
,
where n, n


0
, p, k, n


0
, p prime, and m, m

(
0
0)
. (Notice that I

n
+I
0
+I
0
is contained in the last case.)
(iv) c := H
f
, / := o
c
: The lists (A), (B) and
(D) I

n
, and I

n
+I
0
,
where n
0
.
(v) c := H
f
, / := o
f
: The lists (A), (B) and
(E) I

2
.
For the proof we refer to [BSi85] or leave it as an exercise.
Remark 20.4.6 We would like to mention a paper of W.Bartol, Z.Ras
and A.Skowron [BtRaSk75], where they study mono-unary partial alge-
bras (called iterative systems) used as models of digital computers of the
von Neumann type. For a mono-unary partial algebra A they explicitly
describe the greatest element in Cong
c
A concluding that Cong
c
A is a com-
plete lattice, and they give a necessary and sucient condition for Cong
c
A
to be a distributive lattice or a Boolean algebra.
Moreover, they show that every mono-unary partial algebra is the (H
c
, o
c
)-
subdirect product of (H
c
, o
c
)-subdirectly irreducible mono-unary partial al-
gebras, but they do not use this representation in the subsequent discussion
of a stored program computer being a mono-unary partial algebra with
additional structure imposed on the carrier set and the fundamental partial
operation.
December 8, 2002
466 Chapter IV Special Topics
21 More facts about E-varieties
Throughout this book there are already scattered many facts about E-varieties.
In this section we want to present some additional results especially about
this kind of variety, and in particular such, which will show us that at
least in the homogeneous case E-varieties and the free (partial and total)
algebras behave almost in the same way as the corresponding whole species
and its absolutely free (partial and total) algebras. In particular one can
formulate results about equational and implicational logic in them (almost)
in the same way, as this has been done earlier within a whole species.
Main restrictions in addition to homogeneity will be the assumptions that
the bases of the free partial and total algebras under consideration have the
element basis property, and that the variety has a partial algebra with at least
two elements; we have seen in section 17 that this is not always satised. We
conjecture that the element basis property is not necessary for the results
below, but we have not succeeded in managing without this assumption.
Nevertheless the basic generalization of the Recursion Theorem holds
without such an assumption, and we shall prove this rst. To some extent
we shall also treat arbitrary heterogeneous partial algebras, but the main
conclusions will then be restricted to homogeneous ones. Most of the results
presented here are generalizations of results rst presented in [B71].
21.1 Initial homomorphic extensions
Remark 21.1.1 In 4.1.13 we have proved a quite general homomorphism ex-
tension lemma; however, for this section it is not yet general enough. Namely
in 4.1.13 we have assumed in the notation of that lemma that D is a rel-
ative subalgebra of B, while in this section we need a similar result for the
case when D need only be a weak subalgebra of B. Since we want to apply
such a result in the case when B = F(D, K
t
) for some E-variety K, the results
in section 17 show that in the heterogeneous case D need not be a relative
subalgebra of B even weak subalgebra is not guaranteed, while in the
homogeneous case this is always satised (see 17.2.6).
Lemma 21.1.2 Let A, A

, B and B

be similar partial algebras, such that A


is a weak subalgebra of A

, and B is a weak subalgebra of B

. Moreover, let
f : A B be any homomorphism which possesses a homomorphic extension
f

: A

. Then the following statements are true:


December 8, 2002
21 More facts about E-varieties 467
(i) f

with graph f

:= c
A

B
graph f is a homomorphism f

: E B,
and A generates E, where E := domf

, and E carries the largest weak


substructure of A

which still allows f

to be a homomorphism. If,
in particular, B is a relative subalgebra of B

, then E is a relative
subalgebra of A

.
(ii) If C is any weak subalgebra of A

generated by A, and if g : C B is
any homomorphic extension of f, then graph g graph f

, and C is a
weak subalgebra of E.
(iii) If A is total, then f

: E B (from (i)) is a closed homomorphism.


And if g : D B is any other closed homomorphic extension of f
starting from an A-generated weak subalgebra D of A

, then g = f

(including D = E).
Proof Ad (i): Because of 4.1.11.(i) f

is a homomorphism from domf


the relative subalgebra of A

with carrier domf

into the relative subalgebra


of B

with carrier B, since obviously graph f

graph f

. From this we
already get the last statement in (i). For each we dene
graph
E
:= (a, a) graph
A

[ a dom
domf

and f

(a) dom
B
.
It is easy to realize that this is indeed the largest weak substructure of A

on E allowing f

to be a homomorphism from E into B. It can now easily


be shown (e.g. by algebraic induction or by induction on the step numbers)
that also graph f

= c
EB
graph f, and 4.1.12 implies that A generates E.
(ii) and (iii) are proved in a similar way as in 4.1.13.
Remark 21.1.3 The results above generalize Theorem 2 (Hutters Theo-
rem) in J.Schmidt [Sch70], where it is required that A

be a normal exten-
sion of A (only the homogeneous case is considered). But the next theorem
below will show that normality is not necessary for the existence of a largest
and then closed initial homomorphic extension, say f

, within a relatively
free completion F(A, K
t
) of a given homomorphism f : A B within a prim-
itive class K. But if we want the domain of f

to be a relative subalgebra of
F(A, K
t
), then B has in general to be a relative subalgebra of F(B, K
t
), as
follows from 21.1.2.(i) if there is a sort s S for which each K-algebra has
an at most one-element phylum, then this can usually easily be violated.
In this context the main application of Lemma 21.1.2 is the following
December 8, 2002
468 Chapter IV Special Topics
Theorem 21.1.4 Let K be any E-variety, and K
t
its subclass of all total K-
algebras. Let A and B be K-algebras such that A and B are weak subalgebras
of F(A, K
t
) and F(B, K
t
), respectively, and let f : A B be any homo-
morphism. Then there exists a largest initial homomorphic extension, say
f

: E B, of f from an A-generated weak relative subalgebra E of F(A, K


t
)
into B E having as much structure as possible. In particular we have
(i) graph f

= c
F(A,Kt)B
graph f.
(ii) f

: E B is a closed homomorphism.
(iii) f

is the right factor of a (K


t
-extendable homomorphisms, K
t
-perfect
homomorphisms)-factorization of f (see 10.2.16), and the embedding
of A into E is the corresponding left factor.
(iv) E is a K-algebra (generated by A).
(v) If B is a relative subalgebra of F(B, K
t
), then E is a relative subalgebra
of F(B, K
t
).
(vi) If we are in the case of homogeneous partial algebras, then A and B are
always at least weak subalgebras of F(A, K
t
) and F(B, K
t
), respectively,
and if K contains an at least two-element partial algebra, then E is
always a relative subalgebra of F(A, K
t
) possibly apart from the case
that both F(A, K
t
) and F(c
B
f[A], K
t
) contain only one element.
Proof Since f has a homomorphic extension f

: F(A, K
t
) F(B, K
t
)
f can also be considered as a homomorphism from A into F(B, K
t
)
Lemma 21.1.2 is applicable, where A

:= F(A, K
t
) and B

:= F(B, K
t
). Thus
the main statement of 21.1.4 as well as (i), (ii) and (v) immediately follow.
Since E is a homomorphic image with respect to the rst projection applied
to the subalgebra on graph f

of the direct product of the K-algebras F(A, K


t
)
and B, (iv) is also true.
Ad (iii): Let i : A E be the canonical embedding; we rst show that
i is a K
t
-extendable epimorphism: That i is an epimorphism follows from
the fact that A generates E; and since each homomorphism from A into
some K
t
-algebra even extends to all of F(A, K
t
), it especially extends to E,
showing that i is K
t
-extendable. It remains to show that f

is K
t
-perfect:
Consider the commutative diagram in Figure 21.1, where h is assumed to be
December 8, 2002
21 More facts about E-varieties 469
gure uncompleted
Figure 21.1: f21.1
K
t
-extendable, and where we have already added the bottom rectangle for
the K
t
-universal solutions.
Since f

belongs to (Ext K
t
) (i.e. is K
t
-perfect), there is a homomorphism
d

: V F(A, K
t
) such that d

h = i
1
p and f

= i
2
q. Since f

is just
the restriction of f

to E, and since f

is closed, we only have to show that


d

[V ] is contained in E (the rest of the proof then being straightforward), and


we prove this by algebraic induction on the elements of V , using that h[U]
generates V , and that p[U] is contained in E: Thus d

h u = i
1
p u = p u
(u U) provides the initial condition; assume that a dom
V
, and that
d

maps all elements of a into E; then q(a) dom


B
, and therefore d

(a)
dom
E
(by the denition of f

and E). Thus


E
(d

(a)) = d

(
V
(a)) belongs
to E, completing the proof.
Ad (vi): If F(B, K
t
) contains at least two elements, or if B is empty, then B is
a relative subalgebra of F(B, K
t
), and the statement follows from (v). Thus
it remains only the case that without loss of generality B and F(B, K
t
)
both have exactly one element; but if A and F(A, K
t
) are both empty, then
the statement is trivial, and if F(A, K
t
) has only one element, we are in the
possibly exceptional case; thus we may assume that F(A, K
t
) has at least
two elements. Since this remaining situation probably will not occur that
often in practice, we only briey sketch the proof: One may use the fact that
each operation in B is either total or empty, and one will realize that the
relative subalgebra of F(A, K
t
) generated by A with respect to the non-empty
operations of B is identical with E.
Because of its importance we introduce the
Denition 21.1.5 Let K be any E-variety, and let f : A B be any homo-
morphism between K-algebras. Then the largest closed initial homomorphic
extension, if it exists according to Theorem 21.1.4, will always be designated
by f

: domf

B; except for the case when K


t
is the class TAlg(S) of all
total algebras, when we shall mostly use the old notation f

instead of f

.
If we know that domf

is indeed a relative subalgebra of F(A, K


t
), then we
shall use as before the notation domf

for the domain of f

, otherwise the
December 8, 2002
470 Chapter IV Special Topics
fact that we only underline d of dom shall indicate that it might only
be a weak subalgebra of F(A, K
t
).
Since each mapping f : M B into some homogeneous K-algebra B has
a homomorphic extension to F(M, K), the above theorem immediately yields
the
Corollary 21.1.6 With the same assumptions as in Theorem 21.1.4, and
the additional one that we are in the homogeneous case, each mapping f :
M B into some K-algebra B has a largest closed initial homomorphic
extension f

: domf

B, to which all the facts in Theorem 21.1.4 apply,


and for which there always holds that F(M, K) is at least a weak subalgebra
of domf

.
Corollary 21.1.7 In every non-trivial E-variety K of homogeneous partial
algebras each K-algebra, say A, is a closed homomorphic image of a weak K-
subalgebra of a suitable K
t
-freely generated K
t
-algebra. More precisely: If M is
any generating subset of A, then (id
M
)

: dom(id
M
)

A is a closed and sur-


jective homomorphism from an M-generated weak K-subalgebra of F(A, K
t
)
onto A and if F(M, K
t
) has at least two elements, it is even a relative sub-
algebra and the pair (dom(id
M
)

, ker (id
M
)

) is a presentation of A within
F(A, K
t
) which contains all the information about A.
Remarks 21.1.8 (i) In some papers on partial algebras, like in [Ke70] by
R.Kerkhoff and in [Po73] by V.S.Poythress a concept of struc-
ture compatible mappings is used which is based on partial mappings.
In [Ke70] these mappings are called Konformismen, and in [Po73]
p-morphisms, and if we use them, we shall call them conformisms: Let
A and B be any partial algebras; a conformism between A and B is a
closed homomorphism from a weak subalgebra of A into B (notice that
in this concept the structure in the preimage is in some way stronger
than the one in the image, while this is just the other way round for
the homomorphisms which we use in this presentation). Then in the
terminology of the authors quoted above we have just proved that
for every E-variety K of homogeneous partial algebras and for every
set M the K
t
-free K
t
-algebra F(M, K
t
) K
t
-freely generated by M is also
conformism-freely generated by M with respect to K in the sense of
Corollary 21.1.7.
December 8, 2002
21 More facts about E-varieties 471
(ii) The results above also show that to some extent the K
t
-universal solu-
tions F(A, K
t
) of K-algebras behave at least in the homogeneous case
in a similar way as we have already observed it with respect to the
general TAlg(S)-universal solutions F(A, TAlg(S)) i.e. the absolutely
free completions in the whole species do.
A connection between the results above and closed initial homomorphic
extensions within absolutely free completions is given by the following
Proposition 21.1.9 Let K be any E-variety, and f : A B any homomor-
phism between K-algebras. Let f

: domf

B and f

: domf

B be
the largest closed initial extensions with respect to F(A, TAlg(S)) respectively
F(A, K
t
) assuming that f

exists. Moreover, let p

: F(A, TAlg(S))
F(A, K
t
) be the homomorphic extension of id
A
, and let p be its restriction
to domf

.
Then p induces a homomorphism p
f
: domf domf

, and this is a
closed and surjective homomorphism satisfying f

= f

p
f
.
Proof Let i := id
A
: A domf

, and let i

be its closed initial extension


within F(A, TAlg(S)), then f

is a closed initial extension of f, and be-


cause of its uniqueness we get f

= f

, and i

= p
f
immediately follows.
The preceding observations also lead to a generalization of the Epimor-
phism Theorem (see 5.8.2 and 5.8.3).
Theorem 21.1.10 (Epimorphism Theorem for Homogeneous E-Varieties)
Let K be any E-variety, and let f : A B and g : A C be any homomor-
phisms between K-algebras, where f is an epimorphism, and let f

and g

be
the closed and initial homomorphic extensions of f and g within F(A, K
t
), re-
spectively. Then there exists a homomorphism h : B C such that h f = g,
iff ker f

ker g

and domf

is a weak subalgebra of domg

in the case that


the domains of f

and g

are relative subalgebras of F(A, K


t
), the relation
between the kernels is sucient.
Proof While the necessity of the condition is trivial, its suciency follows
from an application of the homomorphism theorem, using the fact that a
closed epimorphism is also full and surjective. If the domains of the ini-
tial extensions are relative subalgebras, the relation between the domains is
implied by the one between the kernels.
December 8, 2002
472 Chapter IV Special Topics
Remark 21.1.11 The results presented in this subsection show that despite
the fact that many thinks work for E-varieties as they do in the whole species,
facts become nevertheless to some extent more complicated; and this becomes
even worse if we consider ECE-varieties, where in general we do not even
know about the existence of relatively free completions, which then have to
be replaced by direct systems of relatively free extensions. Mainly because
of lack of space we do not go here into more details, but on the other hand
we do not yet know about any detailed investigation in this direction. Some
hints on how to treat ECE- and QE-varieties can be found in the presentation
of [ABN80].
21.2 On the lattice of initial closed congruence rela-
tions homogeneous case
Remark and Notation 21.2.1 (i) In the last subsection we have seen
that at least in the homogeneous case the information about all M-
generated K-algebras in an E-variety K is contained in the pairs (E, ),
where E is any M-generated weak K-subalgebra of F(M, K
t
), and
is a closed congruence relation on E (cf. 21.1.6 and 21.1.7) and the
case, when E is not already a relative subalgebra can occur only when
F(M, K
t
) has only one element. Because of the diculties which we
have encountered in sections 17 and 18 and hence also in 21.1 for the
general heterogeneous case, we restrict our considerations in this sub-
section to the homogeneous case.
(ii) Let us introduce the following notation: Let K be any (abstract) class
of partial S-algebras, let A be any K-algebra, and let M be any subset
of A. Then we denote by Wsubalg(M, A, K) and Rsubalg(M, A, K) the
sets of all M-generated weak respectively relative subalgebras of A be-
longing to K. Similarly, we write Cong
c,w
(M, A, K) respectively Cong
c,r
(M, A, K)
for the set of all pairs (E, ), where is a closed congruence relation
on E, and E is any element of Wsubalg(M, A, K) or Rsubalg(M, A, K),
respectively, i.e. they are each the union of the sets of closed congruence
relations on an M-generated weak respectively relative K-subalgebra
(say E) of A, labeled by their domain (i.e. by E) we shall call them
M-initial closed K-congruence relations. In the case of Cong
c,r
(M, A, K)
we shall usually omit mentioning the domain, since then it is uniquely
determined by the congruence relation.
December 8, 2002
21 More facts about E-varieties 473
Lemma 21.2.2 Let K be any E-variety, let A be any K-algebra, and M
any generating subset of A. Then Cong
c,w
(M, A, K) and Cong
c,r
(M, A, K) are
complete lattices with respect to set theoretical inclusion in each component.
The inmum of a family B of elements of one of these sets is given as fol-
lows: inf B = (D,

), where D is the subalgebra generated by M in

E [
(E, ) B, and

is the intersection of DD with

[ (E, ) B.
Proof Since (A,
A
) is the largest element in both sets of M-initial closed
congruence relations, the Lemma is proved once that we have shown the
statement about the inma. Since id
D
has a homomorphic extension (id
D
)
E
:
F(D, K) E for each (E, ) B, and since the graph of all these mappings
is always the same, we have a surjective homomorphism (id
D
)

: F(D, K)
D, showing that D is indeed a K-algebra a similar argument shows that
the intersection of weak or relative K-subalgebras of any K-algebra is again a
weak or relative K-subalgebra. If we can show that

is a closed congruence
relation on D, then (D,

) is indeed the inmum of B: Thus assume that


a, b D
K
, that a dom
D
, and that the sequences a and b are compo-
nentwise congruent modulo

; then the same assumptions apply for every


(E, ) instead of (D,

). But is a closed congruence relation on E, there-


fore a and b both belong to the domain of
E
, and (
E
(a),
E
(b)) . But
this easily implies that
D
is dened on b, and that (
D
(a),
D
(b))

,
showing that

is indeed a closed congruence relation on D.


21.2.3 Although we do not have counterexamples, we cannot prove in gen-
eral that inf B described above is always already the componentwise set the-
oretical intersection of the elements of B. Nevertheless the results obtained
so far show that in E-varieties one can formulate at least in the homoge-
neous case a local model theory in a similar way as we have done for
the whole species in Chapter II. But if we could restrict considerations at
least to Cong
c,r
(M, F(M, K
t
), K) for any given set M, this would make the
corresponding results much smoother. Thus we dedicate the rest of this sub-
section to the study of cases when this special situation arises. We shall
heavily need the element basis property in what follows, although we have
no idea whether this property is really necessary. If there will really be prac-
tical interest in this area of investigations, then there still remains a lot to
be done, and at this place we only want to oer an idea of possible methods.
Theorem 21.2.4 Let M and N be any sets, M N, let K be an E-variety,
and let z : U
F(N,K)
N be an injective mapping such that the z-completion
December 8, 2002
474 Chapter IV Special Topics
A := F(N, K)
z
is isomorphic over id
M
to F(M, K
t
), the (total) K
t
-free
K
t
-algebra (cf. subsection 17.3). Moreover, let each x F(N, K)C(F(N, K))
have an element basis in N, and assume either M ,= or that c
F(N,K)
=
C(F(N, K)) (cf. subsection 18.1).
Then Rsubalg(M, F(M, K
t
), K) is a closure system on F(M, K
t
).
Proof Since M generates the whole algebra, the empty intersection of ele-
ments of Rsubalg(M, F(M, K
t
), K), namely A, again belongs to Rsubalg(M, F(M, K
t
), K).
Now let B := (A
j
)
jJ
be any non-empty family of elements of Rsubalg(M, F(M, K
t
), K).
Because of 17.3.6.(i) we have for each j J that A
j
= c
F(N,K)
(N A
j
). We
abbreviate N
j
:= N A
j
(j J) and B :=

A
j
[ j J . Then B is a
relative subalgebra of A, and because of our assumptions it is easy to realize
that C(F(N, K)) c
F(N,K)
M B. By possibly introducing new nullary
fundamental operations with value x for each x C(F(N, K)) we change the
subalgebra lattice of F(N, K) in such a way that C(F(M, K)) becomes the
smallest closed subset. Then N has the element basis property, and there-
fore we get for the closure operator c

with respect to the possibly extended


structure of A that
B =

A
j
[ j J =

N
j
[ j J = c

( N
j
[ j J ) :
We only have to prove the last equality, and the inclusion of the right hand
side in the left hand side is trivial; in order to prove the other inclusion
let x be any element of

N
j
[ j J . Since J is non-empty, and
since x has an element basis xN in N, we have xN N
j
for each j
J, and therefore xN

N
j
[ j J , showing that x also belongs to
c

N
j
[ j J ). (Because of C(F(N, K)) B this equality would also
have been obtained without changing the similarity type.) Thus we get for
N

:=

N
j
[ j J that B = c
F(N,K)
N

. We prove indirectly that M


generates B, i.e. we assume that B

:= c
B
M ,= B. Since B belongs to K
as an intersection of relative subalgebras of A belonging to K, we also get
B

K, where B

is the relative subalgebra of A on B

. Again 17.3.6.(i)
implies that for N

:= N B

we get B

= c
F(N,K)
N

. Hence the assumption


B ,= B

implies N

,= N

. With again designating the ordinal dimension


under consideration and G := G
F(N,K),z
as in 17.3 we recall from 17.3.6 that
A = c
A
M = G (

M [ < ); therefore we have for each a A an


ordinal (a) which is the least ordinal such that a G

M [
< . By assumption there is a N

, and we can choose such an


December 8, 2002
21 More facts about E-varieties 475
a N

, for which

:= (a) is minimal. It is then easy to realize that


we also have

< (b) for each b BB

(apply the fact that each b BB

has an element basis bN in N, and that 17.3.6.(v) implies that for b G

M
we also have bN G

M). Because of M BB

we have

> 0. Moreover
we have
() B
_
G

M [ <

= B

_
G

M [ <

.
For the chosen a N

satisfying (a) =

there exists exactly one


sequence a (

M [ <

)
K
for some suitable such that
(, a) domz and a = z(, a) =
A
(a). Since a B =

A
j
[ j J
we get a

A
K
j
[ j J = B
K
. Because of () this implies a B
K
,
and since B

is a closed subset of B, this contradicts the assumption that


a BB

. Hence B = B

, showing that Rsubalg(M, F(M, K


t
), K) is a closure
system on A = F(M, K
t
).
Corollary 21.2.5 Let K be any E-variety such that for each non-empty
set N each x F(N, K) C(F(N, K)) has an element basis in N. Then
Rsubalg(M, F(M, K
t
), K) is a closure system on F(M, K
t
) for each non-empty
set M.
Combining Corollary 21.2.5 with Theorem 18.2.10 we easily get, when d
designates the cardinal dimension of the given similarity type, the
Corollary 21.2.6 Let K be any non-trivial E-variety of homogeneous partial
algebras of cardinal dimension d #
0
. Then Rsubalg(M, F(M, K
t
), K) is a
closure system on F(M, K
t
) for each non-empty set M and if M is empty,
then we have additionally to assume that c
F(N,K)
= C(F(N, K)) for some
non-empty set N.
Because of the fact that it is the main consequence of Theorem 21.2.4 we
designate the next corollary as
Theorem 21.2.7 Let K be any non-trivial class of homogeneous partial al-
gebras such that for each non-empty set N each x F(N, K) C(F(N, K))
has an element basis in N. Then Cong
c,r
(M, F(M, K
t
), K) is a closure sys-
tem on F(M, K
t
) F(M, K
t
) for each non-empty set M forgetting about
the rst component, which is always induced by the second one. And if M
is empty, then we have additionally to assume that c
F(N,K)
= C(F(N, K))
for some non-empty set N.
December 8, 2002
476 Chapter IV Special Topics
Remarks 21.2.8 (i) Observe that 21.2.7 has a corollary similar to 21.2.6.
Moreover, similar results are also true in the inhomogeneous case under
additional assumptions derived from earlier observations; nevertheless
the formulations are much more complicated, and we do not give them
here.
(ii) Observe that for every non-trivial E-variety of nitary homogeneous
partial algebras the isomorphism classes (over M ,= ) of M-generated
K-algebras represented as triples (M, f, A), where f : M A, and
f[M] is a generating subset of A form a complete lattice, induced
by a quasi-ordering, where (M, f, A) ~ (M, f

, A

) iff there exists a


homomorphism g : A A

such that g f = f

; this corresponds to
the relation ker f

ker f

(cf. Theorem 21.1.10).


21.3 Relativized algebraic partial operations homo-
geneous case
Remarks and Denition 21.3.1 In this subsection we want to present
some considerations concerning the formulation of relativized algebraic
partial operations with respect to an E-variety. We recall that it is quite
natural to formulate e.g. polynomials for rings and groups and other equa-
tionally dened total algebras within the special variety rather than for the
whole species. One might be tempted to do the same for partial algebras.
In this subsection we shall only briey treat the homogeneous case just
as an example to show some of the eects arising for partial algebras:
Let K be any non-trivial E-variety of homogeneous similar partial alge-
bras, let M be any non-empty set, and A be any non-empty K-algebra
we assume additionally that each non-empty subalgebra, say B, of F(M, K
t
)
has at least two elements, or B has no proper non-empty weak subalgebra
in K. Then the results in 21.1 imply that each assignment f : M A has
a largest closed homomorphic extension f

: domf

A, where domf

is
an M-generated relative subalgebra of F(M, K
t
). As in section 5 we can
introduce for each t F(M, K
t
) an M-ary partial operation, say t
A
, on A
by dening domt
A
:= a A
M
[ t doma

, and t
A
(a) := a

(t) for each


a domt
A
.
We then call t
A
a K-algebraic M-ary partial operation on Aor a relativized
(to K) algebraic partial operation on A.
December 8, 2002
21 More facts about E-varieties 477
Finally we introduce APO
M
K
(A) := t
A
[ t F(M, K
t
) , the set of all
K-algebraic M-ary partial operations on A.
The remainder of this subsection is mainly devoted to the investigation of
the relationship between K-algebraic partial operations on A and algebraic
ones as investigated in section 15.
Lemma 21.3.2 Let K be a non-trivial E-variety of homogeneous partial al-
gebras, let A K such that each non-empty subalgebra of F(A, K
t
) has at least
two elements or has no proper non-empty weak subalgebra in K, let M be any
non-empty set, p := id
M
, and let p

: F(M, TAlg()) F(M, K


t
) be its ho-
momorphic extension. Let A be an arbitrary K-algebra and e : M E
M
(A),
dened by e(m) := e
M
m
for all m M (cf. section 15), the canonical map-
ping which maps M onto the set of all M-ary total projections of A
M
. Let
e

: dome

F
M
(A) and e

: dome

F
M
(A) be the largest closed ho-
momorphic extensions of e within F(M, TAlg()) and F(M, K
t
), respectively;
and for f : M A let p
f
: domf

domf

be the (closed and surjective


cf. 21.1) restriction of p

to domf

.
Then we have the following statements:
(i) For each q F(M, K
t
): graph q
A
=

graph t
A
[ t p
1
(q) , and
q
A
is a partial operation on A.
(ii) ker f

= (p
f
1
p
f
1
)[ker f

].
(iii) t
A
[ t dome

= q
A
[ q dome

= F
M
(A).
(iv) ker e

is a fully invariant congruence relation on dome

.
(v) dome

is freely and A-freely generated by M.


Proof Observing the notation introduced in the lemma we get:
Ad (i): From 21.1.9 we know that for each f : M A we have f

=
f

p
f
, and p
f
: domf

domf

is closed and surjective. Now, let


g FU(M, K
t
), and let f domg
A
, i.e. g domf

. Then there is
t domp
f
such that g = p
f
(t) , thus g
A
(f) = t
A
(f). Hence we have the
inclusion in (i) from the left in the right hand side. Conversely, assume
that f domt
A
for some t p
f
1
(g). Then we have t domf

,
and therefore p
f
(t) = g domf

, therefore the other inclusion easily


December 8, 2002
478 Chapter IV Special Topics
follows. Let us observe, in addition, that for t, u p
1
(g) which satisfy
f domt
A
domu
A
, we have t
A
(f) = f

(t) = f

(p
f
(t)) = f

(g) =
f

(p
f
(u)) = f

(u) = u
A
(f).
(ii) is easily realized.
Ad (iii): We already know from section 15 that the rst and the third set are
equal. p
e
maps dome

onto dome

, and (i) implies that all elements


of dome

also imply total operations on A, which have to be identical


with the ones induced by the corresponding elements of dome. Since p
e
is surjective, equality follows.
(iv) and (v) are proved in the same way as this has been done for ker e

and dome

in section 6.
Example 21.3.3 In order to get a better idea of 21.3.2.(i) we consider the
E-variety R

of all homogeneous partial algebras A := (A; +, 0, , , 1, h)


of type := (2, 0, 1, 2, 0, 1) such that (A; +, 0, , , 1) is an associative and
commutative ring with multiplicative unit 1, and h is an arbitrary additional
unary partial operation satisfying h(1) 1
e
= 1. Let
s := ((x +y) + (h(x) + ((h(x)))) +h(z)) and
t := ((x +y) + (h(y) + ((h(y)))) +h(z))
be two elements of F(x, y, z, TAlg(2, 0, 1, 2, 0, 1)), and let
Z

:= (Z; +, 0, , , 1, h
Z

)
be the ring of all integers with an additional unary partial operation satisfying
graph h
Z

:= (1, 1). Then we have


doms
Z

= ((1, b), 1) [ b Z

and
domt
Z

= ((a, 1), 1) [ a Z

,
while s and t have the same image in F(x, y, z, R

t
), namely g := x + y +
h(z), satisfying
domg
Z

= ((a, b), 1) [ a, b Z

.
Remarks and Problems 21.3.4 (i) As long as the elements of F(M, K
t
)
possess a normal form within F(M, TAlg(S)), each K-algebraic partial
operation of A K will also be algebraic, but it is conceivable that
this is not the general situation, although we have not yet computed a
suitable example.
December 8, 2002
21 More facts about E-varieties 479
(ii) We also conjecture, but have not yet checked in detail, that, for each
(non-empty) set M, for each non-trivial E-variety of (homogeneous)
partial algebras and for each K-algebra A, APO
M
K
(A) is the K
t
-universal
K
t
-solution of APO
M
(A).
(iii) As it has already been hinted upon earlier, we have also the possibility
of relativizing arbitrary X-ary partial polynomial functions for given
K-algebras A by considering F(A

X, K) and F(A

X, K
t
), where
at least in the homogeneous case for non-trivial A the former can be
considered as a relative subalgebra of the latter, and each assignment
f : X A also has a largest closed (A X)-initial closed homomorphic
extension with respect to the identity mapping on A within F(A

X, K
t
).
(iv) In those cases where each M-generated K-algebra (for some E-variety K)
is representable by a closed congruence relation on an M-generated
relative subalgebra of F(M, K
t
), the closed M-initial extensions of as-
signments f : M A into K-algebras A provide a tool for formulating
a basic model theoretic language within the primitive class. But it
should be observed that it is in general coarser than the global one,
e.g. it might not be possible to distinguish two globally dened ECE-
subvarieties on this local level, meaning that at least one of them might
not be locally denable as an ECE-variety.
(v) Most of the trouble which arises for E-varieties of heterogeneous par-
tial algebras as discussed earlier could be avoided by a more re-
strictive denition of the concept of E-equations in the heterogeneous
case, e.g. by only allowing as E-equations (M : t
e
= t

) such expressions
which satisfy v
M
[M] = v
var {t,t

}
[var t, t

]. But this would cause other


mostly syntactical diculties, and we repeat that we have chosen the
concept presented so far because of compatibility with the approach
of the ADJ-group in computer science, the concept treated in Part II
by H.Reichel, and the category theoretical approach of H.Andr eka,
I.N emeti and I.Sain.
December 8, 2002
480 Chapter IV Special Topics
December 8, 2002
Appendix
In section 22 we briey mention some concepts and results which could not
be treated in detail in this book, and we give some further hints to literature
concerning applications of the theory presented so far. In section 23 some
concepts and results from set theory are presented without proof which
are often used. In section 24 some concepts and results from the theories
of partial orders, lattices, closure systems and Galois correspondences are
collected, which are often used throughout this book.
481
482 Appendix
22 Survey on some additional concepts and
results
22.1 Other concepts of homomorphisms
22.1.1 This book is based on a concept of homomorphisms which corre-
sponds to those in general used for relational systems or total algebras, re-
spectively; we make this more precise for the homogeneous case for any given
similarity type = (())

:
The category TAlg() of all total algebras of type is a full and epire-
ective subcategory of Alg(); and in the way described in the introduction,
in 1.1.3 and in 13.4.2 even the category RelS() of all relational systems of
type i.e. the class of all relational systems (M, (
M

), where, for each


,
M

is a subset of M
()
(i.e.
M

M
()
is a ()-ary relation
on M) and with homomorphisms f : (M, (
M

) (N, (
N

) which
are mappings from M into N such that, for each , f
()
[
M

]
N


can be considered as a full and epireective subcategory of Alg(). Moreover,
Alg() can be considered in a natural way as a full subcategory of RelS(

),
with

:= (() ())

, where each ()-ary partial operation


A
is
replaced by the corresponding (() ())-ary relation graph
A
.
However there are other concepts around, and we want to discuss some
of them and to show that they can be modeled without great diculties
within the presented theory.
Denition 22.1.2 Let A, B Alg(S), h : A B any partial mapping
out of A into B:
(i) h is called a quomorphism (short for quasi-homomorphism) out of A
into B iff one has for every and for every a A
()
:
(QH)
If a dom
A
(domh)
()
and
A
(a) domh,
then (h a) dom
B
and h(
A
(a)) =
B
(h a).
(ii) h is called a conformism (see [Ke70], or a p-morphism in [Po73]) out
of A into B iff one has for every and for every a A
()
:
(C)
If a (domh)
()
, and if (h a) dom
B
,
then a dom
A
and h(
A
(a)) =
B
(h a).
December 8, 2002
22 Survey on some additional concepts and results 483
(iii) h is called a partial algebraic mapping out of A into B iff one has for
every and for every a A
()
:
(PAM)
If a dom
A
(domh)
()
,
A
(a) domh, and (h a) dom
B
,
then h(
A
(a)) =
B
(h a).
Remarks 22.1.3 (i) The above concepts have already been used some-
where in the literature. Note that quomorphisms are dierent from
the partial homomorphisms in H.-J.Hoehnkes [Hoe72] and also
dierent from the morphisms considered in [Hoe76]. He also studied the
category Par of all sets with partial mappings (see also B.Wojdylo [Wo72]).
(ii) Notice that partial mappings h out of A into B which satisfy (QH)
and (C) are sometimes called conformic partial homomorphisms (see
[Hoe72], p.10). The interpretation homomorphisms v : F(M, TAlg(S))
A for an assignment v : M A are always such conformic partial ho-
momorphisms, but we have seen in Chapter II that they can also be
expressed in the category Alg(S) as the right factor of a (TAlg(S)-
extendable,TAlg(S)-perfect)-factorization of v : M
d
A.
(iii) It might be useful to observe that, for a partial mapping h : A B,
we can characterize it being one of the above morphisms out of A
into B in the following way, where the right hand side always has to
be satised for all :
quomorphism iff graph (h
A
[
(domh)
() ) graph (
B
h
()
);
conformism iff graph (h
A
[
(domh)
() ) graph (
B
h
()
);
conformic quomorphism iff graph (h
A
[
(domh)
() ) = graph (
B
h
()
);
(conformic partial homomorphism)
partial algebraic mapping iff graph (h
A
) (dom(
B
h
()
) B)
graph (
B
h
()
).
(iv) Using a composition of two morphisms, where the target of the rst one
is discrete, one easily realizes that the smallest category containing the
class Alg(S) and either (all homomorphisms and all conformisms) or (all
quomorphisms and all conformisms) or all partial algebraic mappings,
respectively, contains in each case all partial mappings between the
carrier sets of the partial algebras involved, i.e. the morphisms have in
December 8, 2002
484 Appendix
each of these cases nothing to do with the algebraic structure of the
objects.
On the other hand all homomorphisms and all conformic partial ho-
momorphisms are quomorphisms, while the above remark shows that
homomorphisms and conformisms cannot be combined in one category
which seems to be relevant for partial algebras. Since among con-
formisms in particular the conformic partial homomorphisms seem to
be of greater interest, one might think that the category containing
them together with our homomorphisms i.e. the category with quo-
morphisms as morphisms might be the most suitable one for partial
algebras.
Because of the following reasons we did not take quomorphisms as the
basic concept for morphisms:
(a) The reasons already mentioned in 22.1.1: The category (Alg(S),homomorphisms)
as well as the category (TAlg(S),homomorphisms) are neither full
nor epireective subcategories of a category (Alg(S),quomorphisms),
and all the category theoretical concepts considered in Chapter I
would dier very much in the new category from the descriptions
given in this book.
(b) However, full and epireective subcategories inherit a lot of prop-
erties from the original one, while an arbitrary subcategory of a
given category may have only little connection with the given one.
Thus we think it natural that the two categories (Alg(S),homomorphisms)
and (Alg(S),quomorphisms) should be studied separately, and within
the scope of this book we could only study one of these categories
in detail.
(c) In the category (Alg(S),quomorphisms) the description of the ba-
sic concepts becomes much more complicated and therefore appli-
cation of such a theory even when it is developed seems rel-
atively unlikely. Moreover, there are (category theoretical) tools
to handle also quomorphisms, conformisms, etc. within the frame-
work of the category (Alg(S),homomorphisms) by considering pairs
of homomorphisms with special properties (see 22.1.4 below).
22.1.4 (Special categories of pairs) The concepts of morphisms described
above can be modeled within the category Alg(S) (again we silently assume
December 8, 2002
22 Survey on some additional concepts and results 485
gure uncompleted
Figure 22.1: f22.1
gure uncompleted
Figure 22.2: f22.2
that the class of morphisms is the class of all homomorphisms) by replacing
a partial mapping, say h : A B, by a suitable pair, say (i, h

) of two
mappings: i : A

:= domh A (i := id
A
) and h

: A

B (graph h

:=
graph h), see Figure 22.1, and in the case of the above morphisms we only
have to choose special properties for each component in the following way:
h is a quomorphism iff i is a full monomorphism and
h

is an arbitrary homomorphism;
h is a conformism iff i is a monomorphism and
h

is a closed homomorphism;
h is a conformic quomorphism iff i is a full monomorphism and
h

is a closed homomorphism;
h is a partial algebraic mapping iff i is a monomorphism and
h

is a homomorphism.
Composition of two partial mappings is then in this translation constructed
in each case by a suitable pullback as indicated in Figure 22.2, where (A

; p, q)
is the pullback of (h

, i
g
; B).
22.1.5 We leave the discussion at this place, but we hope that the reader
might have gotten the idea, that the concepts presented in this book are not
as special as he might have thought.
22.2 Further topics and hints to applications
22.2.1 (ECE-varieties)
(i) As already mentioned at several other places, there exists another
approach to special ECE-varieties mainly started by J.Slominski
December 8, 2002
486 Appendix
in [Sl68], in some instances further developed by H.H oft in [Hoef73]
and treated in most generality by R.John in [J75]. We mean the no-
tions of validity of equations in partial algebras (see 8.1.3). Some
contributions with respect to weak validity (i.e. weak equations)
can also be found in R.Kerkhoff [Ke70], P.Burmeister [B71],
V.S.Poythress [Po73], and L.Rudak [Ru83], where nally the prob-
lem of a description of closed sets of weak equations has been solved.
Applications of this concept can be found in G.A.Edgar [Ed73] (to
topology) and in S.Kochen and E.P.Specker [KoSp68] (to the logic
of quantum mechanics). In a diplom thesis [Oe81] at the Technische
Hochschule Darmstadt G.Oehme has combined the results of [Ke70]
and [Po73] with our concept of ECE-varieties, and he proves that an
1o
c
T
+
-closed class K of (homogeneous) partial algebras possesses for
each set M a conform-free K-algebra iff for every ECE-equation
t
e
= t

which is valid in K the weak equation t


e
= t t

e
= t

t
e
= t

is also valid in K.
(ii) In his doctoral thesis [Si84] M.Siegmund-Schultze has studied the
concept of residual smallness for E- and ECE-varieties. Especially for
ECE-varieties he gets very satisfactory results in particular in con-
nection with eqational compactness by allowing also negations of TE-
statements together with E-equations.
22.2.2 (Category theoretical approaches) Besides the approach to a
theory of partial algebras presented in this book there are other ones making
more use of category theory by using a Lawvere style approach, where partial
algebras are functors from some special kind of small categories into the
category of all sets with partial mappings. Since space and time limitations
do not allow us to go into more details here and to give it due credit, we would
just like to mention in this connection the school around H.-J.Hoehnke with
papers (among others) from Hoehnke [Hoe76], [Hoe82] (see also [Hoe84]),
J.Schreckenberger [Sre80] and H.-J.Vogel [V79] and [V84]. Another
approach in this direction is made by A.Obtulowicz in [O81] and [O82];
his results might be quite useful for the study of ECE-varieties, since his
concept of equations covers our concept of ECE-equations.
22.2.3 (Applications in computer science) We conclude our discussion
of the theory of partial algebras by hints mainly due to H.Andr eka,
December 8, 2002
22 Survey on some additional concepts and results 487
I.N emeti and I.Sain to references concerning applications in computer
science of the results presented in this book:
In a series of papers [AN79f] H.Andr eka and I.N emeti present a sur-
vey of applications of universal algebra in computer science, which also indi-
cates at many places main directions of applications of partial algebras see
also [AN83].
In [Mr83] and [Mr83a] Zs.Markusz thoroughly investigates the problems
and diculties of dening validity for a logic as presented here in section 7
for many-sorted partial algebras (including empty sorts).
The papers [Pa83] and [Pa84] of A.Pasztor apply the theory of partial
algebras to obtain results about the concrete categories used in the approach
by D.S.Scott and the so-called ADJ-group (S.L.Bloom, J.A.Goguen,
J.Meseguer, J.W.Thatcher, E.G.Wagner, J.B.Wright) to compu-
tational semantics. E.g. she proves that the structures used there are simply
ECE-varieties of partial algebras, thus easily obtaining important results
as consequences of the theory presented here. See also the other papers
of A.Pasztor.
The papers [T84] and [T84a] of A.Tarlecki contain important applica-
tions of to computer science the results presented here in sections 12 and 8.2.
In the papers [GuL82], [GuL82a], and in particular in [GuL80] by R.Guitart
and C.Lair mainly the results presented in sections 12 and 8.2 are combined
with heavily functorial category theoretical approaches.
Further applications of partial algebras mainly in computer science
can be found among others in the following papers: I.Guessarian and
J.Meseguer [GeMe84], I.G.Rosenberg [R83], B.Mahr and J.A.Makowsky
[MhMk84] and [MhMk84a], J.A.Makowsky [Mk84], E.Knuth and L.Ronyai [KnRo83],
and M.Broy and M.Wirsing [BrWi81] and [BrWi81a] (see also [BgBrTuWi81]).
December 8, 2002
488 Appendix
23 Some denitions and facts from set theory
In addition to subsection 1.1 and the notation scattered throughout the book
we want to collect here some remarks on the basic set theoretical notation
as well as on the set theoretical background, together with some facts on
cardinals and ordinals. For more information about set theory we refer to
J.Schmidt [Sch66a], D.Monk [Mo76], A.Levy [Lev79], and with respect
to cardinals and ordinals in particular to H.Bachmann [Bc67].
From an axiomatic point of view we use the Godel-Bernays-von Neumann
approach. See also Remark 24.4.10.(ii).
23.1 Some remarks on notations
23.1.1 (Some notational conventions) From a naive point of view we
distinguish proper classes, sets and objects in the following way:
Classes of algebraic structures or sets of subsets (or substructures) of
some other set (or structure) are denoted by capital Gothic letters, as in
K Alg() (to be read as the class K is a subclass of the class Alg() of
all partial algebras of similarity type ). Observe that a set is always a class
which is an element of some other class as possibly specied by the axioms
of axiomatic set theory.
Sets sometimes also proper classes are usually denoted by capital
Latin letters, and their elements are denoted by small Latin letters from the
beginning or the end of the alphabet. Exceptions are made when these ele-
ments are sequences (see 1.1.1; then the small Latin letters are underscored)
or sequences of sequences (see e.g. 4.1.3; then we use small Gothic letters for
sequences and underscore them if they designate a sequence of sequences).
always designates the empty set, i.e. the set with no elements.
Moreover, closure systems are denoted by capital calligraphic letters, and
the same holds for categories not further specied (they usually are under-
scored because of their partial algebraic structure).
Capital letters are underscored when we want to express that we deal
with an algebraic or relational structure on the corresponding carrier set.
Ordinals are denoted by small Greek letters, cardinals by bold-faced cap-
ital letters.
23.1.2 (Some set theoretical notation)
December 8, 2002
23 Some denitions and facts from set theory 489
(i) designates the element relation. Negation: / : a / A, if a is
not an element of A.
(ii) () designates (proper) set theoretical inclusion, i.e. A B
means that each element of the class A also belongs to the class B (A
and B may be equal), while A B means A B and A ,= B (i.e.
not A = B). Negation of A B: A , B. ( is sometimes
more explicitely expressed by ,.)
(iii) For some given property P, x [ P(x) designates the class of all
sets x having property P. We sometimes write x A [ H(x) , if
P(x) means x A and H(x) for some class A and some property H.
(iv) Let A and B be any classes; then:
A B := a [ a A and a / B designates the dierence class of A
without B.
A B := a [ a A and a B designates the intersection or meet
of A and B.
A B := a [ a A or a B designates the union or join of A
and B.
A

B := (0A) (1B) is called the disjoint union of A and B.

A := b [ b a for each a A =:

a [ a A =:

( a [ a A)
designates the meet or intersection of the class A (observe that for each
class its members are again sets).

A := b [ there exists a A such that b a =: b [ b a for some a


A =:

a [ a A =:

( a [ a A) designates the union or join


of A.
Similarly, we dene for indexed families ( A
t
[ t T ) of sets or classes:

( A
t
[ t T ) := a [ a A
t
for each t T ,

( A
t
[ t T ) := a [ a A
t
for some t T ,
and as in the case of two sets we dene the disjoint union of the family
( A
t
[ t T ) as

( A
t
[ t T ) :=

( t A
t
[ t T )
(similarly

A :=

( a a [ a A) for every class A).


December 8, 2002
490 Appendix
(v) In connection with the inmum or supremum of partial algebraic (weak)
substructures we use

and

, respectively, i.e.:

( A
t
[ t T ) means (

( A
t
[ t T ), (

(
A
t
[ t T ))

), and
graph

(
A
t
[ t T ) :=

( graph
A
t
[ t T ).
Similarly,

( A
t
[ t T ) means (

( A
t
[ t T ), (

(
A
t
[ t
T ))

), where graph

(
A
t
[ t T ) :=

( graph
A
t
[ t T ), and

( A
t
[ t T ) is only dened, when for each we again get a
partial operation.
(vi) For any two sets a and b the ordered pair (a, b) is a set such that for
any two additional sets a

and b

:
() (a, b) = (a

, b

) iff a = a

and b = b

.
Usually an ordered pair is realized by the so-called Kuratowski-pair
(a, b) := a, a, b. But since we also use this symbolism, when
a and b may be proper classes (see e.g. the notation in connection
with factorization systems), we want to mention that there exists a set
theoretical realization of an ordered pair for possibly proper classes
(then being again a class) having the above property (): Such a pair
is described in detail in [Sch66a], pp 96 as
() (a, b) := x [ x a Subs y [ y b
(following an idea of Quine), where Subs A := M [ M is a set and
M A designates for any class A the class of all subsets of A, the
so-called power-class of A. () yields a set iff both a and b are sets.
(vii) The direct or cartesian product of two classes A and B is dened to
be the class of all ordered pairs with rst component in A and second
component in B:
A B := (a, b) [ a A and b B.
If A
1
, A
2
, . . . , A
n
are classes, then we dene the n-fold cartesian product
A
1
A
2
. . . A
n
:= (. . . (A
1
A
2
) . . .) A
n
,
but usually we do not distinguish this product from the set of all n-ary
choice sequences (see 1.1), since there is a canonical bijection between
December 8, 2002
23 Some denitions and facts from set theory 491
these two sets; and in the case of possibly innitely many factors, i.e.
for

( A
i
[ i I ), we only use the representation of its elements by
choice functions (I being any set; see also 4.1.3):

( A
i
[ i I ) := a [ a : I
_
( A
i
[ i I ) and a(i) A
i
for every i I .
23.1.3 (Relations)
(i) A binary relation between two classes A and B is simply a subclass
A B of ordered pairs with rst component in A and second
component in B. We denote by
dom := a A [ there is b B such that (a, b) , the domain
of , by
rng := b B [ there is a A such that (a, b) , the range of ,
and by
eld :=
dom rng , the eld of (this is used in particular if A = B).
If A = B, we call a binary relation on A.
(ii) Moreover, if A B, then we dene its converse
1
by

1
:= (b, a) [ (a, b) B A. Thus rng =
dom
1
.
(iii) If AB is a binary relation, A

A and B

B, then we dene
[A

] := b B [ there is a A

such that (a, b) = rng ( (A

B)),
similarly we dene
1
[B

] := dom( (A B

)).
(iv) If T is any set, and if ( A
t
[ t T ) is a family of sets (or even of classes),
then a T-ary relation on this family is a subclass

( A
t
[ t T ).
For some special kinds of relations see 24.1.
23.1.4 (Functions and (partial) mappings)
(i) In set theory a function f is just a binary relation on V the class of
all sets which is unique in its second component: (a, b), (a, b

) f
implies b = b

;
December 8, 2002
492 Appendix
(ii) In this book a function f as above will usually be called the graph
of a (partial) mapping (i.e. f = graph g) for some (partial) mapping
g : A B, since we consider mappings as the morphisms in the cat-
egory Set of all sets, and similarly partial mappings as the morphisms
in a category Par (again with V as class of objects):
We dene a partial mapping g out of a class A into a class B in-
formally as an assignment assigning to each element a of A at most
one element of B denoted by g(a), if it exists, and formally as a
triplet (A, graph g, B) (sometimes written as (A, f, B), or in symbols
as g : A B), where A is called the start object and B is called
the target object of g, and graph g := (a, g(a)) [ a A and g(a)
exists A B is a function in the sense of (i). The fact that
(a, g(a)) graph g is also often indicated by writing a g(a) (think-
ing of the informal denition of g as an assignment).
In analogy with 23.1.3.(i) we dene the (exact) domain of a partial
mapping g as
domg := a [ a A and there is b B such that (a, b) graph g .
In general this is not the category theoretical domain, which is Dom(g)
(:= A).
If domg = A, then we write g : A B and by calling g a mapping
from A into B.
Instead of g : A B we also sometimes write g : A domg
B or g : domg A B.
If, in particular, A designates a set of variables, we call a mapping
w : A B from A into some set B an assignment.
(iii) Let A be a class and B a set. By A
B
we denote the class of all mappings
from B into A: A
B
:= g [ g : B A.
(iv) A partial mapping g : A B is called
injective, if a ,= a

implies g(a) ,= g(a

) for every a, a

domg;
surjective (onto), if for each b B there is some a domg such that
b = g(a);
bijective, if g is an injective and surjective mapping from A onto B.
December 8, 2002
23 Some denitions and facts from set theory 493
(v) Observe that we always apply mappings from the left. As a consequence
we then also denote the composition of partial mappings f : A B
and g : B C by g f (: A C). In some situations we write gf
rather than g f.
In particular, id
A
designates the identity mapping from some class A
onto itself or sometimes into some class B containing A as a subclass
satisfying id
A
(a) := a for each a A. It usually has to be taken
from the context, whether or not A = B (at some occasions we shall
be more explicit by writing id
AB
(: A B)). Observe that we apply
operators (e.g. the ones dened in section 4) from the left, too, just
like mappings operator meaning here that the argument as well as
the value may be a proper class, and we should extend our set theory
by conglomerates (see 23.2.1)).
(vi) Observe that, for the composition g f of two partial mappings, we
have dom(g f) = a domf [ f(a) domg , and a dom(g f)
implies (g f)(a) := g(f(a)).
(vii) Let f : A B be a partial mapping, set := graph f, and let
A

A, B

B; then we dene as in 23.1.3: f[A

] :=
A
, the
image of A

with respect to f, and f


1
[B

] :=
1
B

, the preimage of B

with respect to f. If f is bijective, then the inverse f


1
is a mapping
f
1
: B A dened as the triplet (B, (graph f)
1
, A), i.e. f
1
(b) = a
iff f(a) = b.
(viii) If f : A B, f

: A

are mappings, then we dene (f f

) : A
A

BB

by (f f

)(a, a

) := (f(a), f

(a

)) for each (a, a

) AA

.
If A = A

, B = B

and f = f

, then we write f
(2)
instead of f f.
Similarly, if f : A B is a mapping, and if T is any set, then we
denote by f
(T)
: A
T
B
T
the mapping induced by f and satisfying
A
T
a := ( a
t
[ t T ) ( f(a
t
) [ t T ) =: f
(T)
(a) B
T
(i.e. it is
dened componentwise by f).
23.2 Some remarks on an axiomatic approach
23.2.1 On most occasions set theoretical concepts may be used in a naive
sense. But at some places some reference is made to set theoretical axioms
December 8, 2002
494 Appendix
(e.g. to the Axiom of Choice), and the Axiom of Foundation is tacitly as-
sumed in some set theoretical constructions. For an axiomatic approach to
set theory the reader may think of the well-known Zermelo-Fraenkel axioms
including the Axiom of Choice; however because of our frequent treatment of
classes the Neumann-Bernays approach seems more appropriate as presented
in part in [Sch66a].
On some occasions we shall even apply operators to proper subclasses
of classes, and the reader may think this embedded in an approach using
conglomerates as a next level, where a conglomerate is a collection of classes
(including sets which are just classes which are elements of some other class),
but proper conglomerates cannot be elements of other conglomerates (see
also the appendix of [HS73]); conglomerates behave otherwise like classes,
but we shall use them only in a very naive way in the text without explicitly
mentioning conglomerates there.
23.2.2 (Axioms of set theory)
Let St be the property of a class to be a set, i.e. to be an element of
another class:
St(a) : M(a M),
and let x, y, z, a, b, M, N be variables for classes (lower case letters usually
indicating that they will actually range over sets because of the context).
Then we have the following axioms (the set theoretical notation being taken
from 23.1, and the logical one from section 7):
Axiom (scheme) of Comprehension: Mx(x M St(x)
H(x))
for every rst order property H in a language for classes (i.e. having as
fundamental non-logical symbol ), where M is not free in H.
Meaning: For every property H there is a class M containing exactly the
sets with property H.
Axiom of Extensionality: y(y M y N) M = N
Meaning: Two classes are equal iff they contain the same elements.
Axiom of the Null-Set: St()
Meaning: The empty class is a set (follows from other axioms!).
December 8, 2002
23 Some denitions and facts from set theory 495
Axiom of Subsets: St(a) b a St(b)
Meaning: Each subclass of a set is again a set.
Axiom of the Binary Union: St(a) St(b) St(a b)
Meaning: The union of two sets is again a set.
Axiom of the One-Element Set: St(a) St(a)
Meaning: For every set a the class containing exactly a as only element is
again a set.
Axiom of Union: St(a) St(

a)
Meaning: For every set its union is again a set.
Inverse of the Axiom of Union: St(

a) St(a)
Meaning: If the union of a class a is a set, then a is already a set; this axiom
implies in particular the Power Set Axiom: St(a) St(Subs a).
Axiom of Replacement: Function(a) St(doma) St(rng a)
Meaning: If a is a function (in the sense of 23.1.4.(i)), and if its domain is a
set, then its range is a set.
Axiom of Innity: St()
Meaning: There exists an innite set; as a special item for an innite set the
set of all natural numbers is chosen, where a set theoretical denition
of runs as follows: :=

x [ x y(y x y y x) .
Axiom of Foundation (Regularity Axiom)
M(x(x M) y(y M y M = ))
Meaning: Every non-empty class M contains an element which is minimal
with respect to the -relation; in particular there is no class which possesses
an innite descending -chain, is an irreexive partial order on the class
of all sets, and thus no class can be an element of itself.
Axiom of Choice
M(z(z M z ,= ) a(Function(a) doma = M z(z M
a(z) z)))
Meaning: Each class of non-empty sets possesses a choice function.
December 8, 2002
496 Appendix
Two well known consequences of the Axiom of Choice are the following
ones:
23.2.3 (Lemma of Kuratowski-Zorn (Zorns Lemma)) Let (H; ) be
a non-empty inductive subset of some set M. Then H has a maximal element
(see 24.1).
23.2.4 (Theorem of Well-Ordering) On every set there exists a well-
ordering.
We also want to draw some frequently needed consequences from the Ax-
iom of Foundation (see Levy [Lev79], II, 5, 6, and 7, or Monk [Mo76], &15):
Proposition 23.2.5 Let V designate the class of all sets and Ord the class
of all ordinals (see 23.3). Then there exists a rank function rnk : V Ord
such that for every ordinal the class V() := A [ A V and rnk(A) <
is a set, and for any sets a and A, a A implies rnk(a) < rnk(A).
In what follows we assume a xed rank function (which can explicitely
be specied).
Proposition 23.2.6 Let A be any set, and let A V be any relation.
Then there exists a relation

such that

is a set,
dom

=
dom, and rng

is a set. In particular we may choose

:= (a, b) [
rnk(b) =

rnk(c) [ (a, c) having the required properties.


From the following result we often use in particular its corollary (similar
results can be found in [Sch66a] (end of &18) without using the Axiom of
Foundation).
Proposition 23.2.7 Let A be any set. Then B := x, A [ x A is a set
disjoint from A with a canonical bijection i : B A, x, A x (x A).
(Since rnk(x, A) > rnk(A) > rnk(x), we have A B = .)
December 8, 2002
23 Some denitions and facts from set theory 497
Corollary 23.2.8 Let A and B be any sets, and let f : A B be any
injective mapping from A into B. Then there exists a set C with a bijection
g : C B such that A C and the restriction g[
A
of g to A is equal to f.
Proof Consider C

:= x, A B [ x B f[A] , dene i : C

B,
x, A B x (x B f[A]), C := A C

and nally g : C B for each


c C by g(c) :=
_
f(c), if c A
i(c), else
Obviously, g is a bijective mapping from C onto B, whose restriction to A
equals f.
23.2.9 If, in 23.2.8, A and B are partial algebras, and if f : A B is
a full and injective homomorphism, then the construction in 23.2.8 can be
extended to provide C with a partial algebraic structure such that g becomes
an isomorphism g : C B. In this case A becomes a relative subalgebra
of C (and if f is closed, then A is even a subalgebra of C).
23.2.10 Another consequence of the Axiom of Foundation is the fact that for
every set M the set MM contains besides the elements of M exactly one
element namely M which is not an element of M. This fact is frequently
used in this book.
23.3 Ordinals
Denition 23.3.1 A class A is called -transitive iff for all classes x: x
A x A (i.e. iff , for all x, y : y x A y A). A class A is
an ordinal, iff A is an -transitive set, and each element of A is -transitive.
(From a more naive point of view an ordinal is a distinguished element in
a class of all well-ordered sets isomorphic to some given well-ordered set, as
will be made more explicit below.) Ord := x [ x is an ordinal designates
the class of all ordinals.
Proposition and Denition 23.3.2 As an immediate consequence we have
that Ord: =: 0 is the -smallest ordinal. Moreover, x Ord implies
x x Ord; x x =: x + 1.
December 8, 2002
498 Appendix
Remark 23.3.3 This approach leads to a model of ordinals which contains
the natural numbers in the following way (see also 23.2.2):
0 := , 1 := , 2 := , = 0, 1, 3 := 0, 1, 2, . . . , n :=
0, 1, . . . , n 1, n + 1 := n n,. . . . Finally, :=
0
:=

n [ n is a
natural number , and this is again an ordinal as well as +1, +2, . . .; this
can be extended without ending. is called the set of natural numbers.
Proposition 23.3.4 Ord is totally (irreexively) ordered by the -relation:
x, y Ord implies x = y or x y or y x. If A Ord and A ,= , then

A A is an ordinal; indeed

A is the smallest element of A with respect


to the -relation. Dene for x, y Ord: x y iff x = y or x y; then
(Ord; ) and (A; ) are well-ordered classes for every ordinal A. 0 is the
smallest element of Ord and of every ordinal A ,= .
Denition and Notation 23.3.5 In what follows we shall use lower case
Greek letters , , , . . . to denote ordinals. is called a successor ordinal, if
there is an ordinal such that = +1, otherwise, if ,= 0, then is called
a limit ordinal (we shall often use the letter to designate limit ordinals).
Proposition 23.3.6 For ordinals and we have:
(i) iff ; and implies + 1 or + 1 = .
(ii) If C , then

C or

C = ;

( + 1) = .
(iii) =

(iff is a successor cardinal) or =

(iff = 0
or is a limit ordinal).
(iv) = Ord [ < .
(v) There is no ordinal such that < < + 1.
(vi) For any set A Ord,

A Ord is the supremum of A in (Ord; ).


(vii) < iff +1 < +1 iff +1 < ; +1 = +1 iff = .
23.3.7 (Principles of transnite induction) Let A Ord (or = Ord)
and B be any class.
(1) If ( A)(( < B) B), then A B.
(2) If 0 B and ( A)( B + 1 B) and ( a)((( is a
limit ordinal) ( < B)) B), then A B.
December 8, 2002
23 Some denitions and facts from set theory 499
Denition 23.3.8 A class is called well-founded iff is a relation and
for every non-empty class A eld there is an x A such that A y [
(y, x) = .
23.3.9 Because of the Axiom of Foundation is a well-founded relation on
the class V of all sets, and so is its restriction to the class Ord or to any
ordinal.
23.3.10 (General Recursion Principle) Let be a well-founded relation
such that, for all x eld , the class y [ (y, x) is a set, and let
F : (eld ) V V be a mapping. Then there is a unique mapping
G : eld V such that, for all x eld : G(x) = F(x, G[
{ y|(y,x)}
) =
F(x, (y, G(y)) [ (y, x) ).
This applies in particular to the -relation on Ord, where it is sometimes
formulated as follows:
Let A ,= be an ordinal or A = Ord, let B be any class, a B any set,
and let F : B B, G : C B be mappings, where C := f [ f B for
some A. Then there is a unique function H : A B such that
(1) H(0) = a;
(2) H( + 1) = F(H()) for every with + 1 A;
(3) H() = G(H[

) for every limit ordinal A.


Denition 23.3.11 (i) The addition of ordinals + : Ord Ord Ord is
the unique mapping dened recursively (with , , Ord) by:
(1) + 0 = ;
(2) + ( + 1) = ( +) + 1;
(3) + =

( + [ < ), if =

,= 0.
(ii) Let Ord and (

[ < ) Ord

, then we dene the sum

[
< ) of the sequence (

[ < ) recursively (on ) as follows:


(4)

[ < 0 ) = 0 if = 0;
(5)

[ < + 1 ) = (

[ < )) +

, if + 1 ;
(6)

[ < ) =

[ < ) [ < ), if 0 ,= =

.
December 8, 2002
500 Appendix
23.3.12 Observe that there is a unique order isomorphism between the sum
(

[ < ); ) of a sequence of ordinals and the poset ( (, ) [ <

, < ; ), where (, ) (, ) iff < or ( = and (<

)).
This observation together with 23.3.13 below show that for every <

[ < ) there exist unique < and <

such that =

[
< ) +.
Proposition 23.3.13 +( +) = ( +) +. < iff there is > 0
such that + = . If < , then + +. 0 + = + 0 = ;
+ . If < , then + < + .
0
iff 1 + = . In
general + ,= + (1 + = ,= + 1).
Denition 23.3.14 (i) Let P = (P; ) be a poset, and let H P. H
is called conal in P (a conal subset of P), iff for all p P there is
h H such that p h (iff H = P).
(ii) Let , Ord be any ordinals. is called conal with iff there exists
a strictly increasing order preserving mapping f : (; ) (; ) such
that f[] is conal in (; ) (cf. 11.1.1).
(iii) An ordinal is called regular iff is not conal with any ordinal < .
In every other case is called singular.
Proposition 23.3.15 If is conal with , then . If = +1, then
is conal with 1. 0, 1, are the rst three regular ordinals; in particular
0 and 1 are the only regular ordinals which are no limit ordinals. is a
regular ordinal, iff for every < and for every mapping f : there is
< such that f[] .
In connection with the ordinal dimension we mainly need the
Corollary 23.3.16 Let be a regular ordinal, and let M be a set allowing
a bijective mapping, say g : M, from some < onto M. Then there
exists, for every mapping f : M , an ordinal < such that f[M] ,
i.e. f[M] is bounded in .
December 8, 2002
23 Some denitions and facts from set theory 501
23.4 Cardinals
23.4.1 While ordinals generalize the types of well-orderings from nite to
innite sets, cardinals generalize the process of determining (by counting)
the number of elements of sets from nite to innite sets (i.e. of comparing
magnitudes by numbers).
Denition 23.4.2 A set A is called equipotent with a set B iff there is a
bijective mapping from A onto B. A set A is called a cardinal number or
simply a cardinal iff A Ord, and A is not equipotent with any A.
Card := [ is a cardinal designates the class of all cardinals. Small
gothic letters like m, n, p, q, c, d, etc. are used in this book as variables or
names for cardinals (mainly since cardinal arithmetic is quite dierent from
ordinal arithmetic).
Proposition 23.4.3 The relation of equipotency is an equivalence relation
on the class V of all sets. Every equivalence class except that of the empty set
is a proper class; and every equivalence class contains exactly one cardinal
(this is a consequence of the Axiom of Choice).
Denition 23.4.4 For any set A let #A designate the unique cardinal m
equipotent with A (m =: #A). #A is called the power or cardinality of A
(sometimes also, in analogy to the nite case, #A is called the number of
elements of A).
Lemma 23.4.5 Card is a system of unique representatives for the system
of equipotency classes (this is no class!), and if A is any member of any such
equipotency class, then #A is the unique cardinal representing this class.
If A B, then #A #B.
Proposition 23.4.6 Let be any ordinal, and A be any set. Then # ;
and if and A are equipotent, then #A . is a cardinal iff
# = . If , then #( + 1) = #.
Proposition 23.4.7 For any sets A and B the following statements are
equivalent:
(i) #A #B.
(ii) There exists an injective mapping from A into B.
(iii) A = , or there is a surjective mapping from B onto A.
December 8, 2002
502 Appendix
Corollary 23.4.8 (Cantor-Bernstein-Theorem) If A and B are sets such
that each is equipotent with a subset of the other set, then A and B are equipo-
tent. (This can be proved without using the Axiom of Choice).
Proposition 23.4.9 Each natural number is a cardinal: Card.
Card; is the smallest innite cardinal. Every set has a smaller cardinal
than its power set: #A < #Subs A. For every Ord there is m Card
such that < m. If G Card is a set, then

G is a cardinal.
Denition 23.4.10 (i) For any ordinal let
+
designate the smallest
cardinal greater than (observe 23.4.9.(iv) and Card Ord).
(ii) The mapping # : Ord Card (usually denoted by ) is dened
recursively (the argument being written as an index!): #
0
:= ;
#

+
1
:= (#

))
+
; #

:=

( #

[ < ), if =

,= is a
limit ordinal.
Proposition 23.4.11 # is a strictly increasing mapping which maps Ord
onto Card [ . Thus, in particular, Card is a proper class.
Lemma 23.4.12 Let m, n be cardinals. Then: m < n
+
iff m n, and
m < n iff m
+
n. m
+
#Subs m (equality for each innite m would
be the statement of the Generalized Continuum Hypothesis, which we do not
need in this book). If is an equivalence relation on some set A, then
#(A/) #A. m
+
= [ # m.
Denition 23.4.13 Let A be any set. A is called: nite iff #A < #
0
,
innite iff #A #
0
, denumerable iff #A = #
0
, countable iff #A #
0
,
uncountable iff #A > #
0
. The ordinals representing innite cardinals
are called initial ordinals; and the initial ordinal representing n is denoted
by (n).
Proposition 23.4.14 If A is a nite set and B A is a proper subset, then
#B < #A; and a set A is innite iff A is equipotent with a proper subset
of A.
December 8, 2002
23 Some denitions and facts from set theory 503
Denition 23.4.15 (Cardinal arithmetic) Let m, n Card.
(i) Then m + n := #(m

n) is called the sum of these two cardinals
and mn := #(m n) is called their product. And if T is any set and
( m
t
[ t T ) Card
T
, then

( m
t
[ t T ) := #(

( m
t
[ t T )) is
called the sum and

( m
t
[ t T ) := #(( m
t
[ t T )) is called the
product of this family of cardinals.
(ii) m
n
:= #(m
n
) is called the n-th power of m (the exponentiation of m
with n) we hope that the same notation for the power of cardinals
in cardinal arithmetic and of their power just as sets will cause no
confusion.
23.4.16 As an immediate consequence of the denitions in 23.4.15 addition
and multiplication of (arbitrarily many) cardinals are commutative and
associative (i.e. they do not depend on any bijective change of the index
set, nor do they depend on any subdivision of this index set and rst adding
(respectively multiplying) with respect to the subsets and then adding (mul-
tiplying) the results).
Proposition 23.4.17 Let m and n be cardinals, and let at least one of
them be innite; then: m + n = maxm, n; and if m ,= 0 ,= n, then
mn = maxm, n.
Proposition 23.4.18 Let ( m
t
[ t T ), ( n
t
[ t T ) Card
T
.
(i) If m
t
= 0 for some t

T, then

( m
t
[ t T ) = 0.
(ii) If there is some t

T such that m
t
is innite, or if t T [ m
t
,= 0
is innite, then

( m
t
[ t T ) = sup(# t T [ m
t
,= 0 , m
t
[ t T).
(iii) If m
t
n
t
for each t T, then

( m
t
[ t T )

( n
t
[ t T ) and

( m
t
[ t T )

( n
t
[ t T ).
(iv) If m
t
n
t
and n
t
2 for each t T, then

( m
t
[ t T )

( n
t
[
t T ).
December 8, 2002
504 Appendix
Denition 23.4.19 A cardinal m is called additively inaccessible or regular
iff it is either 0, 1, 2 or it is ininite and its corresponding initial ordinal is a
regular ordinal; else m is called additively accessible or singular.
Proposition 23.4.20 is a regular cardinal. If is a regular ordinal,
then #(= ) is a regular cardinal. A cardinal m ,= 2 is regular iff its cor-
responding initial ordinal is a regular ordinal (but observe that 2 is regular
as a cardinal, but singular as an ordinal). For each cardinal m its suc-
cessor cardinal m
+
is regular. #

is additively accessible (i.e. a singular


cardinal).
December 8, 2002
24 Ordered sets, algebraic lattices and Galois connections 505
24 Ordered sets, algebraic lattices and Galois
connections
24.1 Partially ordered sets and lattices
24.1.1 ((Partial) Orders) Let A be any set or class. Recall that a binary
relation on A (i.e. A A) is
reexive iff (x, x) for all x A,
irreexive iff (x, x) / for all x A,
symmetric iff (y, x) , whenever (x, y) ,
antisymmetric iff (x, y) and (y, x) implies x = y,
transitive iff (x, y) and (y, z) implies (x, z) .
A reexive, symmetric and transitive binary relation on A is called an
equivalence relation on A; (if A is a set) each equivalence relation on A
induces a partition [a]

[ a A of A, where [a]

:= b A [ (a, b)
is the equivalence class of a with respect to and two equivalence classes are
either equal or disjoint; conversely, if Z := A
i
[ i I is a partition of A,
then

Z
:= (a, b) [ a, b A, a and b belong to the same class of Z
is an equivalence relation on A.
A reexive and transitive relation on A is called a quasi-order on A, and
if is in addition antisymmetric, then it is called a partial order (sometimes
also: partial ordering or simply order) on A, and (A; ) is called a partially
ordered set or briey a poset sometimes irreexivity rather than reexivity
is required, and in this book we shall usually indicate a reexive partial order
by a symbol like , _, (set theoretic inclusion) or similar symbols,
and an irreexive partial order by a symbol like <, , , etc., and we
shall write a b (inx notation) rather than (a, b) .
If is a partial order on A which additionally satises
for all x, y A one has (x, y) or (y, x)
(or x = y in the case of irreexivity),
then is called a linear or total order on A, and (A; ) is called a linearly
or totally ordered set or simply a chain.
If is a (quasi-) order on A, then the relation dened by a b
iff b a (for all a, b A) is called the dual (quasi-) order on A.
December 8, 2002
506 Appendix
Lemma 24.1.2 Let A be any set, A A.
(i) If is a reexive quasi-order on A, then the relation := (a, b)
AA [ (a, b) and (b, a) is an equivalence relation on A, and
the quotient structure (A/; /) is a poset (the universal solution of
(A; ) in the subcategory of all posets with (quasi-) order preserving
mappings), where
/ := ([a]

, [b]

) [ (a, b)
= ([a]

, [b]

) [ a, b A, (a

, b

) for some a

[a]

, b

[b]

.
(ii) If (A; ) is a poset (respectively a chain), and if M A, then (M;
(M M)) is a poset (respectively a chain).
24.1.3 (i) Let (A; ) be a poset, M A any subset.
u A is called an upper (respectively lower) bound of M, if m u
(respectively u m) for every m M.
s A is called a supremum or least upper bound (respectively inmum
or greatest lower bound) of M (and denoted by supM (infM) iff m s
(s m) for all m M, and if m s

(s

m) for all m M and


some s

A, then s s

(respectively s

s) if it exists.
Notice that in a poset an inmum respectively supremum of some sub-
set M is uniquely determined by M, whenever it exists (this need not be
true in a quasi-order), and it is denoted by infM (respectively supM).
If supM (infM) exists and belongs to M for some M A, then supM
is called the largest element or maximum maxM of M (respectively the
smallest element or minimum minM of M). If sup exists in A, then it
is the smallest element of A, if inf exists, then it is the largest element
of A.
An element a A is called maximal (minimal) if, for all b A, a b
implies a = b (b a implies a = b).
If, for two not necessarily distinct elements a, b A, supa, b
(infa, b) exists, then it is often denoted by ab, sup(a, b) (respectively
ab, inf(a, b)). We always have sup(a, a) = a = inf(a, a), and if it exists,
then sup(a, b) = sup(b, a), inf(a, b) = inf(b, a).
December 8, 2002
24 Ordered sets, algebraic lattices and Galois connections 507
(ii) Let (P; ) be a poset. If each non-empty subset of P has a smallest
element, then is called a well-ordering of P, and (P; ) is called a
well-ordered set, well-ordered by . Observe that every well-ordered set
is in particular a chain. More information about well-ordered sets can
be obtained in 23.3 in connection with ordinals.
24.1.4 A poset A := (A; ) in which for any two elements a, b of A sup(a, b)
exists (inf(a, b) exists) is called an upper or join- (lower or meet-) semilat-
tice; and if A is both an upper and a lower semilattice, then it is called a
lattice. If, for every M A, supM (infM) exists, then A is called a complete
upper semilattice or complete sup-semilattice (complete lower semilattice or
complete inf-semilattice), and if A is both a complete upper and complete
lower semilattice, then it is called a complete lattice.
If for each subset M A one has: if M has an upper bound, then M
also has a supremum, and if M has a lower bound, then M has an inmum,
then A is called conditionally complete; in particular a conditionally complete
upper semilattice is an upper semilattice which is conditionally complete.
24.1.5 It is well-known that every total algebra (A; ) with a binary idem-
potent (i.e. a a = a), commutative (i.e. a b = b a) and associative (i.e.
(ab)c = a(bc)) operation is an upper semilattice, when one denes,
for all a, b A: a b iff a b = b (or dually a lower semilattice by dening:
a b iff a b = a), in which sup(a, b) = a b (inf(a, b) = a b in the dual
case).
Conversely, for every, say, upper semilattice (A; ) the partial algebra
(A; sup(., .)) is a total algebra of type (2), in which sup(., .) is idempotent,
commutative and associative and satises a b iff sup(a, b) = b.
Similarly, (A; ) is a lattice, iff (A; , ) is a total algebra of type (2, 2)
such that both and are idempotent, commutative and associative, and
we have the absorption laws a(ab) = a and a(ab) = a; and moreover:
a b iff a b = b (iff a b = a).
24.2 k-directed posets, k-inductivity, k-algebraic lattices
Denition 24.2.1 Let k be any cardinal and A := (A; ) be any poset. A
is called (upwards k-directed observe that we shall not use downwards
k-directed iff every subset M of A with cardinality less than k (#M < k)
has an upper bound in A.
December 8, 2002
508 Appendix
A is simply called directed iff every nite subset has an upper bound, i.e.
iff it is #
0
-directed.
24.2.2 (Special cases) Let A be any poset.
A is 0-directed does not mean anything for A.
A is 1-directed iff the empty subset has an upper bound, i.e. iff A ,= .
A is 2-directed iff A is 1-directed (iff A ,= ).
A is 3-directed iff A is k-directed for any cardinal with 3 k #
0
,
iff every nite subset of A has an upper bound in A
(iff A ,= and any two elements of A have an upper
bound in A, iff A is directed).
24.2.3 If k

k and A is a k-directed poset, then A is k

-directed.
If k

is an additively accessible cardinal, if k is the least additively inac-


cessible cardinal greater than k

, and if A is k

-directed, then A is k-directed.


Denition 24.2.4 Let k be any cardinal, and let (L; ) = L be a complete
lattice.
(i) A subset M of L is called k-inductive iff for any k-directed subset T of M
the supremum sup
L
T of T formed in L belongs to M: sup
L
T M.
(ii) a L is called k-compact iff one has for every k-directed subset D of L:
a sup
L
D iff there is d D such that a d;
(iff , for every T L, a sup
L
T implies a sup
L
T

for some T

L,
#T

< k).
(iii) L is called k-compactly generated or k-algebraic iff every l L is the
supremum of a set of k-compact elements of L, i.e. in particular
l = sup
L
c L [ c l and c is k-compact in L.
(iv) If k = #
0
, then a k-algebraic lattice L is simply called algebraic, and
k-compact elements are called compact.
December 8, 2002
24 Ordered sets, algebraic lattices and Galois connections 509
gure uncompleted
Figure 24.1: f24.1
gure uncompleted
Figure 24.2: f24.2
Lemma 24.2.5 Let k k

be any cardinals and L be any complete lattice.


Then:
(i) The fact that a L is k-compact implies that a is k

-compact.
(ii) The fact that M L is k-inductive implies that M is k

-inductive.
(Since every k

-directed set is k-directed.)


24.2.6 (Special cases) Let (L; ) = L be a complete lattice, M L,
a L:
(i) k = 0: M is 0-inductive iff , for every T M, sup
L
T M, i.e. iff (M; )
is a complete sup
L
-subsemilattice of L in this case M := (M; ) is
also a complete lattice, since 0
L
:= sup
L
M, but inf
M
need not
coincide with inf
L
, since for T M, inf
M
T = sup
L
m M [ m is a
lower bound of T , and this set of lower bounds of T within M may
have a supremum t which is strictly smaller than that of the set of all
lower bounds of T within L.
inf
L
L is not 0-compact. The 0-compact elements of L are sometimes
called the completely supremum-prime elements of L (see (ii)). If L
is the four-element lattice given in Figure 24.1 by its Hasse-diagram
(there is a line upward from a to b if a < b and if there is no other
element between a and b), then the two atoms (the elements joined
by a line with the smallest element) are 0-compact; however the lattice
in Figure 24.2 does not have any 0-compact elements.
(ii) k = 1 or k = 2: M is k-inductive iff , for every non-empty subset T
of M, sup
L
T belongs to M: T M and T ,= implies sup
L
T M.
Hence M need not be a complete sublattice. a L is k-compact iff
a = inf
L
L or a is completely supremum prime: if a sup
L
D for some
non-empty set D, then a d for some d D.
December 8, 2002
510 Appendix
(iii) 3 k l < #
0
: Then, because of 24.2.1, M is k-inductive iff it is
l-inductive, and a L is k-compact iff a is l-compact. (For k > 0,
k-directedness of M implies M ,= .)
Lemma 24.2.7 Let L be a complete lattice, k 2 a regular cardinal. Then
the supremum of less than k k-compact elements of L is k-compact; hence the
subset L
k
of L of all k-compact elements of L is k-directed.
24.3 Closure systems and closure operators
Although we keep the following survey quite general, we do not claim that
all assumptions are presented in greatest generality.
Denition 24.3.1 Let L = (L; ) be a complete lattice (in most applica-
tions L will be the complete lattice (Subs E; ) of all subsets of some set E
with set theoretical inclusion as partial order).
(i) A subset H L is called a closure system in L iff for every subset
B H also its inmum with respect to L belongs to H:
(CS) B H implies inf
L
B H
(if L = (Subs E; ), then this means

B H).
(ii) A mapping c : L L is called a closure operator in L iff c has the
following properties for all a, b L:
(C1) a b implies c a c b (c is monotone);
(C2) a c a (c is extensive);
(C3) c c a = c a (c is idempotent).
(iii) If c is a closure operator in L, and if b = c a, then we say that a
generates b with respect to c (a c-generates b).
24.3.2 (i) In particular we have, if L = (Subs E; ), that H is a closure
system on E and c is a closure operator on E, respectively, iff H
Subs E, c : Subs E Subs E, and for B H and M, N E:
(CS)

B H
(C1) M N implies c M c N;
(C2) M c M;
(C3) c c M = c M.
December 8, 2002
24 Ordered sets, algebraic lattices and Galois connections 511
(ii) In 24.3.1 we have introduced the more general concepts of a closure
system and a closure operator in a complete lattice, since in applications
we often take into consideration the algebraic structure on a set and
consider complete lattices with closure systems and closure operators
on Wsuba A with the inclusion of partial algebraic structures for any
given partial A (see subsections 3.1, 3.2 and 23.1).
Lemma 24.3.3 Let L = (L; ) be a complete lattice, and let c be a closure
operator on L. Then one has for every family ( a
t
[ t T ) L
T
:
c inf
L
( c a
t
[ t T ) = inf
L
( c a
t
[ t T ) and c sup
L
( c a
t
[ t T ) = c sup
L
( a
t
[ t T ).
Theorem 24.3.4 Let L be any complete lattice.
(i) Let H be a closure system in L, and dene c
H
: L L by c
H
a :=
inf
L
( h H [ a h); then c
H
is a closure operator in L (called the
closure operator corresponding to the closure system H), and one has
H = a L [ c
H
a = a .
(ii) Let c : L L be a closure operator in L. Then the set H
C
:= c
L
=
c a [ a L is a closure system in L (called the closure system
corresponding to the closure operator c), and one has for each a L:
c a = inf
L
( b H
C
[ a b ).
(iii) With the assumptions and notation from above one has H
C
H
= H and
c
H
C
= c; i.e. there is a one-to-one correspondence between the closure
operators and the closure systems in L described by (i) and (ii).
Theorem 24.3.5 Let k 2 be any (regular) cardinal, let L = (L; ) be a
complete k-algebraic lattice, let c be a closure operator in L, and let H be its
corresponding closure system. Then the following statements are equivalent:
(i) H is k-inductive in L.
(ii) For every k-directed system B L one has c sup
L
B = sup
L
( c b [ b
B).
(iii) For every a L one has c a = sup
L
( c b [ b a, b is k-compact in L).
Observe that (i) and (ii) are equivalent without L being k-algebraic (or k being
regular).
December 8, 2002
512 Appendix
Lemma 24.3.6 For every cardinal k 2 and for every set E, (Subs E; )
is a k-algebraic complete lattice. And if, in addition, k is regular, then the
k-compact elements of (Subs E; ) are exactly the subsets of E of cardinality
less than k.
Lemma 24.3.7 For each closure system H in a complete lattice (L; ) the
subposet H := (H; ) is itself a complete lattice such that inf
H
= inf
L
, in
particular 1
L
:= inf
L
belongs to H.
Moreover, for each B H, sup
H
= c
H
sup
L
B; and if B H is k-directed,
then (because of 24.3.5.(ii)): sup
H
B = sup
L
B.
Proposition 24.3.8 Let k #
0
be a regular cardinal, let L be a k-algebraic
complete lattice, and let H be a closure system in L with closure operator c.
(i) If H is k-inductive, then H := (H; ) is a k-compactly generated (i.e.
k-algebraic) complete lattice.
(ii) Let H
k
designate the set of all k-compact elements of H; then H
k
=
h H [ there is a k-compact l L such that c l = h, i.e. the k-
compact elements of H are exactly the closures of k-compact elements
of L.
Moreover, we have the following connection between abstract k-algebraic
lattices and subalgebra lattices of (total) algebras:
Theorem 24.3.9 Let L = (L; ) be a complete lattice, and let k 2 be a
regular cardinal. Then L is k-algebraic iff there is a partial (or even a total)
algebra A with cardinal dimension k such that L

= ( c
A
M [ M A; ).
Proof (Sketch) (ii) implies (i) because of 3.3.2. (i) implies (ii): Let
A := L
k
be the set of all k-compact elements of L including 0
L
:= inf
L
L, and
let c
A
0
:= 0
L
be the nullary constant of A.
For each a L
k
dene a unary (total) operation
A
a
mapping b L
k
onto a,
if a b, and onto 0
L
, otherwise. For each cardinal c with 2 c < k dene
a total c-ary operation
A
c
by
A
c
( c
i
[ i c ) := sup
L
( c
i
[ i c ); then 24.2.7
implies that the value is k-compact and therefore belongs to A = L
k
. Then
L

= (Csubs A; ).
December 8, 2002
24 Ordered sets, algebraic lattices and Galois connections 513
Since every complete lattice L is (#L)
+
-algebraic, we get the
Corollary 24.3.10 Each complete lattice is isomorphic to the lattice of all
closed subsets of a suitable (possibly innitary) total algebra.
In a similar way one shows:
Proposition 24.3.11 Let k 2 be a regular cardinal, and let H be a k-
inductive closure system on some set A (i.e. in (Subs A; )). Then there is
a total algebraic structure (
A
)

of cardinal dimension k on A such that


(H; ) = (Csubs A; ).
Namely, let f : Subs A A be a choice function (f(M) M for
each non-empty M A). For each set M A with #M < k, and for each
a c
H
M dene for M = :
A
,a
:= a, and else:

A
M,a
( a
m
[ m M ) :=
_
a , if a
m
= m (m M)
a
f(M)
, else.
24.4 Galois connections
24.4.1 The concept of Galois connections (sometimes also called Galois cor-
respondences) originates from Galois theory in connection with the solution
of polynomial equations with coecients in some eld, but similar facts can
be observed quite often in mathematics, in particular in connection with log-
ics, and in category theory in connection with pairs of adjoint functors and
with factorization systems. We restrict this survey to posets and in particular
to power sets. Up to minor changes we take this survey from [Sch68a].
Denition 24.4.2 Let P := (P; ) and Q := (Q; ) be two posets, and let
: P Q, : Q P be two mappings such that for each p P and for
each q Q one has:
q (p) iff p (q);
then the pair (, ) is called a Galois connection (Galois correspondence)
between P and Q. If, in particular, P = (Subs A; ) and Q = (Subs B; )
(A and B sets), then (, ) is called a Galois connection on A and B.
December 8, 2002
514 Appendix
Proposition 24.4.3 Let ( : P Q, : Q P) be a Galois connection
between two posets (P; ) and (Q; ), and p, p
1
, p
2
P, q, q
1
, q
2
Q. Then:
(G1) p
1
p
2
implies (p
2
) (p
1
) and q
1
q
2
implies (q
2
) (q
1
)
(i.e. and are antimonotone mappings).
(G2) p ((p)) and q ((q)).
(G3) (((p))) = (p) and (((q))) = (q).
We now restrict considerations to Galois connections on some sets, say A
and B.
Corollary 24.4.4 (1) Let A and B be any sets and let : Subs A Subs B
and : Subs B Subs A be two mappings. Then (, ) is a Galois connection
on A and B iff and are antimonotone with respect to set theoretical
inclusion, and and are extensive and idempotent (see 24.3.1).
Corollary 24.4.5 (2) With the assumptions of 24.4.4 the operators
and are closure operators on A and B, respectively.
The closed sets of the corresponding closure systems, say / on A and 1
on B, are sometimes called left-closed and right-closed, respectively.
Corollary 24.4.6 (3) The mapping := [
L
: / 1is an anti-isomorphism
(i.e. a bijection which is antimonotone in both directions, here with respect
to set theoretical inclusion) with inverse := [
R
.
In particular the Galois connection (, ) on A and B is uniquely deter-
mined by the quadruple ( , , , ), since = and = .
Moreover, each quadruple (c,

c, , ) determines a Galois connection on


A and B iff c and

c are closure operators on A and B, respectively, and
is an anti-isomorphism from the closure system H
C
on A onto the closure
system H
C
on B with inverse .
Theorem 24.4.7 A mapping : Subs A Subs B is one part of a Galois
connection on the two sets A and B iff (

( M
t
[ t T )) =

( (M
t
) [ t
T ) for all families ( M
t
[ t T ) of subsets of A (in particular () = B).
December 8, 2002
24 Ordered sets, algebraic lattices and Galois connections 515
Of great importance for the theory of Galois connections on two sets is
the following
Theorem 24.4.8 Let AB be a binary relation between any two sets
A and B.
(i) Dene for every M A and N B:
(M) := b B [ (a, b) for all a M ,
(N) := a A [ (a, b) for all b N .
Then (, ) is a Galois connection on A and B.
(ii) Conversely, if (, ) is a Galois connection on A and B, then
:= (a, b) A B [ a (b) = (a, b) A B [ b (a)
is a binary relation on A B inducing (, ) according to (i).
Corollary 24.4.9 The Galois connections (, ) on given sets A and B are
in one-to-one correspondence to the binary relations A B.
Remarks 24.4.10 (i) Thus the characterization of a certain binary rela-
tion AB is essentially the task to determine the left-closed and
the right-closed sets of the corresponding Galois connection, in par-
ticular to describe explicitly the corresponding closure operators and
closure systems.
(ii) If A and B are proper classes (or at least one of them is), then the
above concepts and observations can be transferred to this situation
with the necessary care with respect to the set theoretic concepts. In
particular one now gets for a given relation A B operators
and dened as in 24.4.8.(i) such that their compositions and
behave like closure operators on A and B, respectively. (One may
introduce as an additional level conglomerates (see 23.2.1)).
(iii) Let us observe, too, that Galois correspondences have become quite
popular now and have many applications in dierent sciences in
the form of concept lattices (see R.Wille [Wl82] and [Wl84]).
December 8, 2002
516 Appendix
December 8, 2002
Bibliography
[AdHS90] J. Adamek, H. Herrlich, and G. Strecker. Abstract and Con-
crete Categories The Joy of Cat. John Wiley & Sons Inc.,
1990.
[ABN80] H. Andreka, P. Burmeister, and I. Nemeti. Quasi-equational
logic of partial algebras. Bull. Sect. Logic, Pol. Acad. Sci.,
9:193199, 1980.
[ABN81] H. Andreka, P. Burmeister, and I. Nemeti. Quasivarieties of
partial algebras - A unifying approach towards a two-valued
model theory for partial algebras. Studia Sci. Math. Hungar.,
16:325372, 1981.
[ACgN87] H. Andreka, W. Craig, and I. Nemeti. A system of logic for
partial functions under existence-dependent Kleene equality.
Manuscript, 1987.
[AMakMkiN79] H. Andreka, E. Makai, L. Marki, and I. Nemeti. Reduced
products in categories. In: Contributions to General Al-
gebra, pages 2545, (Proc. Conf. Klagenfurt 1978), 1979.
Verlag Johannes Heyn.
[AN78] H. Andreka and I. Nemeti. Los-lemma holds in every cate-
gory. Studia Scientiarum Math. Hung., 13:361376, 1978.
[AN79f] H. Andreka and I. Nemeti. Applications of universal algebra,
model theory and categories in computer science (Survey and
bibliography). Parts I and II in CL&CL (Comp. Linguist.
Comp. Lang.), Vol. 13, 1979, pp. 152282 and Vol. 14, 1980,
517
518 BIBLIOGRAPHY
pp. 4365. Part III in Fundamentals of Computation The-
ory 81 Lect. Notes Comp. Sci., Vol. 117, 1981, pp. 281290.
Part IV: Preprint Math. Inst. Hung. Acad. Sci. 1984, 1979.
[AN79] H. Andreka and I. Nemeti. Formulas and ultraproducts in
categories. Beitrage zur Algebra und Geometrie, 8:133151,
1979.
[AN79a] H. Andreka and I. Nemeti. Injectivity in categories to repre-
sent all rst order formulas I. Demonstratio Math., 12:717
732, 1979.
[AN79b] H. Andreka and I. Nemeti. On the more general notion of
factorization. Preprint of Math. Inst. Hungar. Acad. Sci. 26,
1979.
[AN79c] H. Andreka and I. Nemeti. Weak products in universal alge-
bra and model theory. Preprint of Math. Inst. Hungar. Acad.
Sci. 39, 1979.
[AN80] H. Andreka and I. Nemeti. Direct limits and ltered col-
imits are strongly equivalent in all categories. In: Algebra
and its Applications, volume 9 of Banach Center Publica-
tions, pages 7588. PWNPolish Scientic Publishers, War-
saw, 1980.
[AN82] H. Andreka and I. Nemeti. A general axiomatizability the-
orem formulated in terms of cone-injective subcategories.
In: Universal Algebra, volume 29 of Colloq. Math. Soc. J.
Bolyai, pages 1335, (Proc. Coll. Esztergom 1977), 1982.
North-Holland Publ. Co. Amsterdam.
[AN83a] H. Andreka and I. Nemeti. Generalization of the concept of
variety and quasivariety to partial algebras through category
theory. Dissertationes Mathematicae (Rozprawy Matematy-
czne), Warszawa, 204, 1983. preprint 1976.
[AN83] H. Andreka and I. Nemeti. Importance of universal alge-
bra for computer science. In: P. Burmeister, B. Ganter,
C. Herrmann, K. Keimel, W. Poguntke, and R. Wille, ed-
itors, Universal Algebra and its links with logic, algebra,
December 8, 2002
BIBLIOGRAPHY 519
combinatorics and computer science, pages 204215, 25. Ar-
beitstagung Allgemeine Algebra, Darmstadt, 1983. Helder-
mann Verlag, Berlin 1984.
[ANSa77] H. Andreka, I. Nemeti, and I. Sain. Cone injectivity and
some Birkho-type theorems in categories. Math. Inst.
Hung. Acad. Sci., Preprint 19/77, 1977.
[ANSa82] H. Andreka, I. Nemeti, and I. Sain. A complete logic for rea-
soning about programs via non-standard model theory. The-
oretical Computer Sci., 17:193212, 259278, 1982.
[APa84] H. Andreka and A. Pasztor. Relative epis need not be sur-
jective. Preprint, to appear in Algebra Universalis, 1984.
[ASa81] H. Andreka and I. Sain. Connections between algebraic logic
and initial algebra semantics of CF languages I. In: Math-
ematical Logic in Computer Science, volume 26 of Colloq.
Math. Soc. J. Bolyai, pages 2583, (Proc. Salg otarj an 1978),
1981. North-Holland Publ. Co., Amsterdam.
[AtKw84] J.P. Antoine and W. Karwowski. Partial -algebras of
Hilbert-space operators. In: Operator algebras, ideals,
and their applications in theoretical physics, volume 67 of
Teubner-Texte Math., pages 2939, (Proc. Int. Conf. Leipzig
1983), 1984.
[AtKw85] J.P. Antoine and W. Karwowski. Partial -algebras of closed
linear operators in Hilbert space. Publ. Res. Inst. Math. Sci.,
21:205236, 1985.
[Am86] M. Armbrust. An equivalence-theoretic equivalent of the ax-
iom of choice. Z. Math. Logik Grundl. Math., 32:95, 1986.
[Arn83] Y.N. Arnaudov. Structural sketches for 3-categories. God.
Vissh. Uchebn. Zaved., Prilozhna Mat., 19(2):8794, 1983.
(Bulgarian, English summary (1984)).
[Bc67] H. Bachmann. Transnite Zahlen. Springer-Verlag, Berlin,
2nd edition, 1967.
December 8, 2002
520 BIBLIOGRAPHY
[Bh85a] A. Bahamonde. Partially additive monoids. Cah. Topologie
Geom. Dier. Categoriques, 26:221244, 1985.
[Bh85] A. Bahamonde. Tensor product of partially-additive
monoids. Semigroup Forum, 32:3153, 1985.
[BaH76] B. Banaschewski and H. Herrlich. Subcategories dened by
implications. Houston Journ. Math., 2:149171, 1976.
[BanKt86] W. Bandler and L.J. Kohout. On new types of homomor-
phisms and congruences for partial algebraic structures and
n-ary relations. Intern. J. Gen. Systems, 12(2):149157,
1986.
[Bar] C.M.de Barros. Quelques structures algebriques denies par
des lois de composition partielles et associatives. C. R. Acad.
Sci. Paris Ser. A-B, pages 256, A163A166.
[Bt71] W. Bartol. On the existence of machine homomorphisms I.
J. Bull. Acad. Polon. Sci., Ser. Sci. Math. Astron. Phys.,
19:856869, 1971.
[Bt72] W. Bartol. On the existence of machine homomorphisms II.
J. Bull. Acad. Polon. Sci., Ser. Sci. Math. Astron. Phys.,
20:773777, 1972.
[Bt74] W. Bartol. Algebraic complexity of machines. Bull. Acad.
Polon. Sci., Ser. Sci. Math. Astron. Phys., 22:851856, 1974.
[Bt74a] W. Bartol. Programy dynamiczne oblicze n (Dynamic pro-
grams of computations). PWNPolish Scientic Publishers,
Warsaw, 1974.
[Bt86] W. Bartol. Weak subalgebra lattices. Manuscript, submitted
to Comment. Math. Univ. Carol., 1986.
[Bt87] W. Bartol. Weak subalgebra lattices of mono-unary partial
algebras. Manuscript, submitted to Comment. Math. Univ.
Carol., 1987.
December 8, 2002
BIBLIOGRAPHY 521
[Bt89] W. Bartol. Termal algebras an approach to inference on
incomplete data. In: K. Hatkowska and B. Stawski, editors,
Univ. and Appl. Algebra, pages 112, (Proc. 5th Univ. Alg.
Symp. May 1988, Turawa, Poland), 1989. World Scientic
Publ. Co., Singapur.
[BtNiRu85] W. Bartol, D. Niwi nski, and L. Rudak. Completion varieties.
Coll. Math., 50:1318, 1985. Preprint Math. Inst. Univ.
of Warsaw.
[BtNiRu86] W. Bartol, D. Niwinski, and L. Rudak. Extensiones de al-
gebras parciales (Extensions of partial algebras). Sociedad
Matematica Mexicana, Aportationes Matematicas, 1:6567,
1986. (Spanish).
[BtNiRu87] W. Bartol, D. Niwinski, and L. Rudak. Semiisomorphism vs.
isomorphism of inner extensions of a partial algebra, 1987.
[BtNiRu87a] W. Bartol, D. Niwinski, and L. Rudak. The semilattice of
inner extensions of a partial algebra, 1987.
[BtRaSk75] W. Bartol, Z. Ras, and A. Skowron. Theory of Computing
Systems. Banach Center Publications (Warszaw), 2:101165,
1975.
[BuPf66] V. Baumann and J. Pfanzagl. The closure operator in par-
tial algebras with distributive operations. Applications to set
algebra, measure theory and linear spaces. Math. Zeitschr.,
92:416424, 1966.
[BsGsStb86] G. Baumslag, D. Gildenhuys, and R. Strebel. Algorith-
mically insoluble problems about nitely presented solvable
groups, Lie and associative algebras. J. Pure Appl. Algebra,
39:5394, 1986.
[BbKh84] U.D. Bekbaev and D. Khadzhiev. First extension of partially
dened operators. Dokl. Akad. Nauk UzSSR, 4:57, 1984.
[BbKh84a] U.D. Bekbaev and D. Khadzhiev. Second extension of par-
tially dened operators. Dokl. Akad. Nauk UzSSR, 5:69,
1984.
December 8, 2002
522 BIBLIOGRAPHY
[Bel86] E.G. Belaga. Fast Computation of linear nite-dimensional
operators over arbitrary rings. In: Algebraic algorithms and
error-correcting codes, volume 229 of Lect. Notes Comp.
Sci., pages 238246, (Proc. 3rd Int. Conf. Grenoble 1985),
1986.
[Bel87] E.G. Belaga. Constructive universal algebra: an introduc-
tion. Theor. Comp.Sci., 51:229238, 1987.
[BdHu85] A. Bendell and S. Humble. A reliability model with states of
partial operation. Nav. Res. Logist. Q., 32:509535, 1985.
[BeKaRe81] K. Benecke, H. Kaphengst, and H. Reichel. Partial algebras
dened by generators and relations. Manuscript, pages 219
232, submitted to Algebra Universalis, 1981.
[BeRe81] K. Benecke and H. Reichel. Equational partiality. Algebra
Universalis, 16, 1983. pages 219232 in Manuscript 1981.
[Bn86] A. Bergmann. Finalstrukturen in ZFC im Hinblick auf par-
tielle Algebren. Math. Schr., 173:147148, 1986. Disserta-
tion, Bonn.
[BgBrTuWi81] J.A. Bergstra, M. Broy, J.V. Tucker, and M. Wirsing. On the
power of algebraic specications. In: Mathematical Founda-
tions of Computer Science 81, volume 118 of Lect. Notes
Comp. Sci., pages 193204, (Proc. 10th Conf. Strbske Pleso
1981), 1981. Springer Verlag, Berlin.
[Bm71] J. Berman. Strong congruence lattices of nite partial alge-
bras. Algebra Universalis, 1:133135, 1971.
[BoMuMi] A. Bertoni, G. Mauri, and P.A. Miglioli. Model theoretic
aspects of abstract data specication. Manuscript, U. Milano.
[BhDa93] S. K. Bhakat and Ph. Das. Fuzzy subalgebras of a universal
algebra. MR 93m:08009, 1993.
[Bi35] G. Birkho. On the structure of abstract algebras. Proc.
Cambridge Philos. Soc., 31:433454, 1935.
December 8, 2002
BIBLIOGRAPHY 523
[BiLi70] G. Birkho and J.D. Lipson. Heterogeneous algebras. J.
Combinatorial Th., 8:115133, 1970.
[BloEk85] S.L. Bloom and Z. Esik. Axiomatising schemes and their
behaviours. J. Comput. Syst. Sci., 31:375393, 1985.
[Boh76] K.A. Bohan.

Uber eine analytische Darstellung gewisser
nichtlinearer Funktionale, die auf rellen Zahlenfolgen
gegeben sind. Funkc. Anal., 6:2534, 1976. Russian.
[Boe87] F. Borner. A comparison of two denitions of partial
clone. Manuscript, 1987.
[Bou61] N. Bourbaki. Toplogie generale (Elements de Mathematique,
Livre III). Hermann, 1961.
[Br86] M. Broy. Partial interpretations of higher order algebraic
types. In: Math. foundations of comp. sci., volume 233 of
Lect. Notes Comp. Sci., pages 2943, (Proc. 12th Symp.
Bratislava 1986), 1986.
[BrWi81] M. Broy and M. Wirsing. On the algebraic extensions of
abstract data types. volume 107 of Lect. Notes Comp. Sci.,
pages 244251, (Proc. Peniscola, Spain, April 1981), 1981.
Springer-Verlag, Berlin.
[BrWi81a] M. Broy and M. Wirsing, editors. An analysis of semantic
models for algebraic specications, (Proc. of Internat. Sum-
mer School M unchen), 1981.
[BrWi83] M. Broy and M. Wirsing. Algebraic denition of a functional
programming language and its semantic models, RAIRO.
Inf. Theor., 17:137161, 1983.
[B68] P. Burmeister.

Uber die Machtigkeiten und Un-
abhangigkeitsgrade der Basen freier Algebren I. Fund.
Math., LXII:165189, 1968.
[B70] P. Burmeister. Free partial algebras. J. reine und angewandte
Math., 241:7586, 1970.
December 8, 2002
524 BIBLIOGRAPHY
[B70a] P. Burmeister.

Uber die Machtigkeiten und Unabhangig-
keitsgrade der Basen freier Algebren II. Fund. Math.,
LXVII:165189, 1970.
[B71a] P. Burmeister. On problem 12 in G.Gr atzers book Uni-
versal Algebra. Notices Amer. Math. Soc., 18:401, 1971.
Abstract.
[B71] P. Burmeister. Primitive Klassen partieller Algebren. Ha-
bilitationsschrift, Bonn, 1971.
[B73] P. Burmeister. An embedding theorem for partial algebras
and the free completion of a free partial algebra within a
primitive class. Algebra Universalis, 3:271279, 1973.
[B81] P. Burmeister. Quasi-equational logic for partial algebras.
In: Fundamentals of computation theory, volume 117 of
Lect. Notes Comp. Sci., pages 7180, (Proc. int. FCT-Conf.
Szeged 1981), 1981.
[B82] P. Burmeister. Partial algebras survey of a unifying ap-
proach towards a two-valued model theory for partial alge-
bras. Algebra Universalis, 15:306358, 1982.
[B86] P. Burmeister. A model theoretic oriented approach to par-
tial algebras. Introduction to theory and application of par-
tial algebras. Part I, volume 32 of Mathematical Research.
Akademie-Verlag, Berlin, 1986.
[B88] P. Burmeister. Closed sets of universal Horn formulas for
many-sorted (partial) algebras. In: Mathematical Problems
in Computation Theory, volume 21 of Banach Center Pub-
lications, pages 129143. PWNPolish Scientic Publishers,
Warsaw, 1988.
[B9?] P. Burmeister. Partial Algebras Including a Treatment
of Many-Sorted and Innitary Algebras. Akademie-Verlag,
Berlin, 199? in preparation.
December 8, 2002
BIBLIOGRAPHY 525
[B93a] P. Burmeister. Partial Algebras An Introductory Survey.
In: Sabidussi, editor, Algebras and Orders, (Proc. of the
Summer School at Montreal), 1992. Kluver Publ. Co.
[B93b] P. Burmeister. Tools for a theory of partial algebras. In:
K.Denecke and H.-J.Vogel, editors, General Algebra and
Applications, volume 20, Research and Exposition in Math-
ematics, 1993. Heldermann Verlag Berlin.
[BJPa78] P. Burmeister, R. John, and A. Pasztor. On closed
morphisms in the category of partial algebras. In:
H. Kautschitsch, W.B. M uller, and W. Nobauer), editors,
Contributions to General Algebra, pages 6976, (Proc. Conf.
Klagenfurt 1978), 1979. Verlag Joh. Heyn.
[BRe82] P. Burmeister and H. Reichel. Stand und Entwicklungsten-
denzen der Theorie partieller Algebren. Mitt. Math. Ges.
DDR, 34:4974, 1982. Preprint Nr. 198, Techn. Hochschule
Darmstadt, (English).
[BSch66] P. Burmeister and J. Schmidt.

Uber die Dimension einer
partiellen Algebra mit endlichen oder unendlichen Operatio-
nen II. Zeitschr. Math. Logik und Grundlagen der Mathe-
matik, 12:311315, 1966.
[BSch67] P. Burmeister and J. Schmidt. On the completion of partial
algebras. Coll. Math., 17:235245, 1967.
[BSi85] P. Burmeister and M. Siegmund-Schultze. Subdirect rep-
resentations of partial algebras. In: G. Eigenthaler, H.K.
Kaiser, W.B. M uller, and W. Nobauer, editors, Contri-
butions to General Algebra 3, pages 7598, (Proc. of
the Vienna Conference, June 21-24, 1984), 1985. Verlag
B.G.Teubner, Stuttgart. Preprint Nr. 867, TH Darmstadt,
1984.
[BWo82] P. Burmeister and B. Wojdylo. A category theoretical back-
ground for homomorphism theorems. Preprint Nr. 699, TH
Darmstadt, 1982. To appear in Coll. Math.
December 8, 2002
526 BIBLIOGRAPHY
[BWo] P. Burmeister and B. Wojdylo. The meaning of category
theoretical notions in some categories of partial algebras.
Preprint, Techn. Hochschule Darmstadt (to appear).
[BWo82a] P. Burmeister and B. Wojdylo. Properties of homomor-
phisms and quomorphisms between partial algebras. In:
Contributions to General Algebra 5, pages 6990, (Proc.
Salzburg Conf. 1986), 1987. Verlag Holder-Pichler-Tempski,
Wien, Verlag B.G.Teubner, Stuttgart. Preprint 1982.
[CrvMad91] R. Madarasz C. Crvenkovic. On aproblem of partial algebras.
MR 91c:08007, 1991.
[Ca78] V.E. Cazanescu. Categories dalgebres universelles I. Math.
Rev. Anal. Numer. Theor. Approx., Math., 20:520, 1978.
French.
[ChKs73] C.C. Chang and H.J. Keisler. Model Theory. North-Holland
Publ. Co., 1973. 2nd edition 1977.
[Cr84] C.C. Christian. Eine Bemerkung zur fast quasigeordneten
Ein-Element-Kompaktizierung von Algebren.

Osterr. Akad.
Wiss., Math.Naturwiss. Kl. 193, Sitzungsber. Abt. II, pages
443454, 1984.
[Ci81] J.M. Cibulskis. An algorithm for the construction of free
lattices over non-free generators. Bull. Calcutta Math. Soc.,
73:1726, 1981.
[ClKr81] D.M. Clark and P.H. Krauss. Topological quasivarieties.
Preprint, Math. Dept. State Univ. of New York, New York,
USA, 1981.
[C65] P.M. Cohn. Universal Algebra, 1965. 2nd rev. ed. D.Reidel
Publ. Co., Dordrecht, 1981.
[Co77] L. Coppey. Stabilite de decompositions dans les
precategories. Esquisses Math., 28, 1977. French.
[Cg89] W. Craig. Near-equational and Equational Systems of Logic
for Partial functions. The J. of Symb. Logic, 54:759827,
11811215, 1989.
December 8, 2002
BIBLIOGRAPHY 527
[CuDk85] R.J. Cunningham and A.J.J. Dick. Rewrite systems on a
lattice of types. Acta Inf., 22:149169, 1985.
[Cp72] G. Cupona. A class of partial algebras. (Macedonian). Fac.
Sci. natur. Univ. Kiril et Metodij, Skopje Annuaire, Math.
Phys. Chim., 22:537, 1972.
[Cz73a] J. Czelakowski. Another approach to partial Boolean alge-
bras. Pol. Acad. Sci. Inst. Philos. Sociol., Bull. Sect. Logic,
2:175177, 1973.
[Cz73b] J. Czelakowski. On embedding of partial Boolean algebras
into Boolean algebras. Pol. Acad. Sci. Inst. Philos. Sociol.,
Bull. Sect. Logic, 2:178181, 1973.
[Cz73] J. Czelakowski. Some remarks on transitive partial Boolean
algebras. Pol. Acad. Sci. Inst. Philos. Sociol., Bull. Sect.
Logic, 2:166174, 1973.
[Cz73c] J. Czelakowski. A local property of embeddings for partial
Boolean algebras. Pol. Acad. Sci. Inst. Philos. Sociol., Bull.
Sect. Logic, 2:182184, 1973.
[Cz74] J. Czelakowski. Logics based on partial Boolean -algebras I.
Studia Log., 33:371396, 1974.
[Cz75] J. Czelakowski. Logics based on partial Boolean -
algebras II. Studia Log., 34:166174, 1975.
[Cz78] J. Czelakowski. On extending of partial Booleanalgebras to
partial -algebras. Colloq. Math., 40:1321, 1978.
[Cz79] J. Czelakowski. Partial Boolean algebras in a broader sense.
Stud. Logica, 38:116, 1979.
[Cz81a] J. Czelakowski. Partial Boolean algebras in a broader sense
and Boolean embeddings. Colloq. Math., 45:171180, 1981.
[Cz81] J. Czelakowski. Partial Boolean algebras in a broader sense
as a semantics for quantum logic. Rep. Math. Logic, 11:49
56, 1981.
December 8, 2002
528 BIBLIOGRAPHY
[Da78] B.A. Davey. Topological duality for prevarieties of universal
algebras, Studies in foundations and combinatorics. Adv.
Math., Suppl. Stud., 1:6199, 1978.
[DaW83] B.A. Davey and H. Werner. Dualities and equivalences for
varieties of algebras. In: Contr. to lattice theory, volume 33
of Colloq. Math. Soc. J. Bolyai, pages 101275, (Szeged
1980), 1983. Preprint, Math. Schr. Kassel 17/80, 1980.
[DeHkRo82] J. Demetrovics, L. Hannak, and L. Ronyai. On the free
spectra of maximal clones. C. R. Math. Acad. Sci., Soc. R.
Can., 4:363366, 1982.
[Den92] K. Denecke. Strong regular varietties of partial algebras. MR
92e:08007, 1992.
[D] K.H. Diener. On induction and recursion in universal alge-
bra.
[D66] K.H. Diener. Order in absolutely free and related algebras.
Coll. Math., 14:6372, 1966.
[Di70] M.B. Dihtjar. Homomorphisms and congruences of partial
-operatives. Studies in Alg. 2, Izdat. Saratov Univ, pages
322, 1970. Russian.
[Di71] M.B. Dihtjar. Congruences of partial -operatives associ-
ated with ordered sets. Ordered sets and lattices 1, Izdat.
Saratov Univ, pages 310, 1971. Russian.
[Di73] M.B. Dihtjar. Direct product of partial -operatives. Stud-
ies in Alg. 3, Izdat. Saratov Univ., pages 913, 79, 1973.
Russian.
[Di74] M.B. Dihtjar. Almost onto homomorphisms of partial -
operatives. Issled. Algebr., 4:1826, 1974. Russian.
[Di77] M.B. Dihtjar. Epimorphisms and monomorphisms of the
category of partial -operatives. Ordered sets and lattices 4,
Izdat. Saratov Univ, pages 2427, 133, 1977. Russian.
December 8, 2002
BIBLIOGRAPHY 529
[DrKp83] A. Drapal and T. Kepka. Group modications of some par-
tial groupoids. Ann. Discrete Math., 18:319332, 1983.
[Du91] J. Duda. Malcev-type theorems for partial congruences. MR
91b:08003, 1991.
[Ea84] D. Easdown. Biordered sets are biordered subsets of idem-
potents of semigroups. J. Aust. Math. Soc., Ser. A 37, pages
258268, 1984.
[Ea84b] D. Easdown. Biordered sets of eventually regular semigroups.
Proc. London Math. Soc., III. Ser. 49, pages 483503, 1984.
[Ea84a] D. Easdown. A new proof that regular biordered sets
come from regular semigroups. Proc. R. Soc. Edinburgh,
Sect. A96, pages 109116, 1984.
[Ea85] D. Easdown. Biordered sets come from semigroups. J. Al-
gebra, 96:581591, 1985.
[EaHl84] D. Easdown and T.E. Hall. Reconstructing some idempotent-
generated semigroups from their biordered sets. Semigroup
Forum, 29:207216, 1984.
[E69] H.D. Ebbinghaus.

Uber eine Pradikatenlogik mit partiell
denierten Pradikaten und Funktionen. Arch. math. Logik,
12:3953, 1969.
[Ed73] G.A. Edgar. The class of topological spaces is equationally
denable. Algebra Universalis, 3:139146, 1973.
[Eh82] C. Ehresmann. Oevres compl`etes et commentees. Parte IV-
1. Esquisses et completions. Supplement des Cahiers Top.
Geom. Di., 1982.
[ErMh85] H. Ehrig and B. Mahr. Fundamentals of Algebraic Speci-
cation 1. Springer, 1985.
[Ei81] G. Eigenthaler.

Uber Einbettungsfragen bei frei erzeugten Al-
gebren. Sitzungsbericht Abt. II,

Osterr. Akad. Wiss., Math.-
Naturwiss., Kl. 190, 1981.
December 8, 2002
530 BIBLIOGRAPHY
[Ei83] G. Eigenthaler. Einige Bemerkungen uber Clones und inter-
polierbare Funktionen auf universellen Algebren. Beitr. Alg.
Geom., 15:121127, 1983.
[Ei83a] G. Eigenthaler. On the lattice of polynomial algebras over an
algebra and its generalizations. In: Contr. to lattice theory,
volume 33 of Colloq. Math. Soc. J. Bolyai, pages 361372,
(Szeged 1980), 1983.
[Ek80] Z. Esik. Identities in iterative and rational algebraic theories.
Comput. Linguist & Comput. Lang., 14:183207, 1980.
[Ev51] T. Evans. The word problem for abstract algebras. Journ.
London Math. Soc., 26:6471, 1951.
[Ev53] T. Evans. Embeddability and the word problem. Journ. Lon-
don Math. Soc., 28:7680, 1953.
[EvLr77] T. Evans and C.C. Lindner. Finite embedding theorems for
partial designs and algebras. Semin. Math. Super. 56 (1971),
Les presses de lUniv. Montreal, 1977.
[Es83] A.E. Evseev. On the algorithmic solvability of the problem of
exterior extendability of partial groupoids. In: Algebraic op-
erations and orderings, Interuniv. Collect. Sci. Works, pages
2631, Leningrad, 1983. Russian.
[Es84] A.E. Evseev. Survey of investigations of partial groupoids.
In: Properties of semigroups, Interuniv. Collect. Sci. Works,
pages 3976, Leningrad, 1984. Russian.
[Fe75] S. Feferman. Non-extensional type-free theories of partial op-
erations and classications I. In: Isilc Proof Theor. Symp.,
dedic. Kurt Schutte, volume 500 of Lect. Notes Math., pages
73118, (Proc. int. Summer Inst. Logic Colloq. Kiel 1974),
1975.
[Fl75] I. Fleischer. On extending congruences from partial algebras.
Fund. Math., 88:1116, 1975.
December 8, 2002
BIBLIOGRAPHY 531
[Fl80] I. Fleischer. Extending a partial equivalence to a congruence
and relative embeddings in universal algebras. Fund. Math.,
106:1317, 1980.
[Fl81] I. Fleischer. Equational classes of partial algebras.
Manuscript, 1981.
[FlR79] I. Fleischer and I. Rosenberg. The Galois connection between
partial functions and relations. Pac. Journ. Math., 79:9397,
1979.
[FRd80] N. Francez and M. Rodeh. A distributed abstract data
type implemented by a probabilistic communication scheme.
Preprint, Technion Comp. Sci. Dept. Haifa, 1980.
[Fd77] J. Fried. Simulations of Pawlak machines as fuzzy mor-
phisms of partial algebras. Comment. Math. Univ. Carolina,
18:343350, 1977.
[Fr74] M. Froda-Schechter. Sur les homomorphismes des struc-
tures relationelles I. Studia Univ. Babes-Bolyai, Ser. Math.-
Mech., 19:2025, 1974.
[Fr78] M. Froda-Schechter. Algebres partielles et structures re-
lationelles. Studia Univ. Babes-Bolyai, Ser. Math.-Mech.,
23:7275, 1978. French.
[FuKgStr84] M. Funk, O.H. Kegel, and K. Strambach. On group univer-
sality and homogeneity. In: Universal algebra and its links
with logic, algebra, combinatorics and computer science, vol-
ume 4 of Res. Expo. Math., pages 173182, (Proc. 25. Ar-
beitstagung Allgemeine Algebra, Darmstadt 1983), 1984.
[Gi78] D.W.H. Gillam. A concrete representation theorem for in-
tervals of multirelations. Z. Math. Logik Grundl. Math.,
24:463466, 1978.
[GlZJu80] W.M. Gluschkow, G.J. Zeitlin, and J.L. Justschenko. Alge-
bra, Sprachen, Programmierung. Akademie-Verlag, Berlin,
1980.
December 8, 2002
532 BIBLIOGRAPHY
[Go78] J.A. Goguen. Abstract errors for abstract data types. In: For-
mal Description of Programming Concepts, pages 491525,
(Proc. Ifip Working Conference, St.Andrews 1977), 1978.
[GoMe81] J.A. Goguen and J. Meseguer. Completeness of many-sorted
equational logic. Manuscript; appeared in Sigplan Notices,
1981.
[GoTcWa78] J.A. Goguen, J.W. Thatcher, and E.G. Wagner. An ini-
tial algebra approach to the specication, correctness, and
implementation of abstract data types. IBM Research Re-
port RC-6487, 1976, and: Current Trends in Programming
Methodology, IV: Data Structuring (R.Yeh, ed.), Prentice
Hall,1978, pages 80144.
[Gon87] S.S. Goncharov. Axiomatizable classes with strong homo-
morphisms. Studia Log., 46:113120, 1987.
[G68a] G. Gratzer. On polynomial algebras and free algebras. Can.
J. Math., 20:575581, 1968.
[G68] G. Gratzer. Universal Algebra. D.van Nostrand Co, 1968.
2nd ed., Springer-Verlag, 1979.
[G78] G. Gratzer. General Lattice Theory. Akademie-Verlag,
Berlin, 1978.
[G85] G. Gratzer. Axiom of Choice is equivalent with Birkhos
subdirect irreducible representation. Preprint, Univ. of Man-
itoba, Canada, 1985.
[GSdt63] G. Gratzer and E.T. Schmidt. Characterization of congru-
ence lattices of abstract algebras. Acta Sci. Math. 24, Szeged,
pages 3459, 1963.
[GWe67] G. Gratzer and G.H. Wenzel. On the concept of congruence
relation in partial algebras. Math. Scand., 20:257280, 1967.
[Gr72a] G. Grimeisen. Extensions of topological partial algebras to
topological algebras. In: Theory of sets and topology, pages
199207. Veb Deutscher Verlag der Wiss. Berlin, 1972.
December 8, 2002
BIBLIOGRAPHY 533
[Gr72] G. Grimeisen. On the saturation of a topolgical partial alge-
bra with respect to a congruence relation. In: General Topol-
ogy and its relations to modern Analysis and Algebra III,
pages 151153, (Proc. third Topol. Symp. Prag 1971), 1972.
[Gd72] S.P. Gudder. Partial algebraic structures associated with or-
thomodular posets. Pacic J. Math., 41:717730, 1972.
[GeMe84] I. Guessarian and J. Meseguer. Axiomatization of
if..then..else revisited. Laboratoire Inf. Theor. et Pro-
grammation, Preprint 84-18, April 1984.
[Gu81] R. Guitart. The theory of sketches (revisited). Journees Fais-
ceaux et Logique (19th P.S.S.L.), U. Paris-Nord & S.M.F.,
Paris, 23 et 24 mai 1981.
[Gu82] R. Guitart. Quest-ce que la logique dans une categorie?
Cahiers de Top. Geom. Di., 23:115148, 1982.
[GuL80] R. Guitart and C. Lair. Calcul syntaxique des mod`eles et
calcul des formules internes. Diagrammes, 4, Dec 1980.
[GuL82] R. Guitart and C. Lair. Existence de diagrammes localement
libres II. Diagrammes, 7, 1982.
[GuL82b] R. Guitart and C. Lair. La continuite pour representer les
formules. Diagrammes, 7, 1982.
[GuL82a] R. Guitart and C. Lair. Limites et colimites pour presenter
les formules. Diagrammes, 7, 1982.
[Hd88b] L. Haddad. Completenes theory for partial nite algebras.
University of Waterloo, Manuscript, 1988.
[Hd88a] L. Haddad. Maximal partial clones determined by quasi-
diagonal relations. University of Waterloo, Manuscript,
1988.
[Hd88] L. Haddad. Maximal partial clones determined by the are-
exive relations. University of Waterloo, Manuscript, 1988.
December 8, 2002
534 BIBLIOGRAPHY
[Hd88c] L. Haddad. Partial Sheer operations. University of Water-
loo, Manuscript, 1988.
[HdR86] L. Haddad and I. Rosenberg. Les clones partiels maximaux
determines par des relations areexives. Universite de Mon-
treal, Preprint No. 86-15, 1986.
[HdR87] L. Haddad and I.G. Rosenberg. Crit`ere general de
completude pour les algebres partielles nies (General com-
pleteness criterion for nite partial algebras). C. R. Acad.
Sci. Paris, Ser. I 304, 17:507509, 1987. French.
[HdR93] L. Haddad and I.G. Rosenberg. Completeness theorem for
nite partial algebras. MR 93j:08005, 1993.
[HdRSw86a] L. Haddad, I. Rosenberg, and D. Schweigert. A maximal
partial clone and a Slupeckitype criterion. Universite de
Montreal, Preprint No. 86-17, 1986.
[HmHr82] J. Hagemann and Chr. Herrmann. Arithmetical locally equa-
tional classes and representation of partial functions. In:
Universal algebra, volume 29 of Colloq. Math. Soc. J. Bolyai,
pages 345360, (Proc. Colloq. Esztergom 1977), 1982.
[HalJak93] E. Haluskova and D. Jakubikova-Studenovska. Partial mo-
nounary algebras with common closed quasi-endomorphisms.
Czech. Math. J., 43(118):225240, 1993.
[Hy72] F. Harary. Graph Theory. Addison Wesley Publ. Comp,
1972.
[Ha76] R. Hartwig. Ein Ableitungsbegri f ur Pseudogleichungen in
uber partiellen Algebren interpretierten Kalk ulen. Preprint,
Leipzig, 1976.
[Ha77] R. Hartwig. Pseudogleichungen in formalen Kalk ulen uber
partiellen Algebren. Vortr age zu Grundlagen der Informatik,
Heft 27/77, TU Dresden, pages 1322, 1977.
[Ha84] R. Hartwig. An algebraic approach to the syntax and seman-
tics of languages with subscripted variables. Period. Math.
Hung., 15:6171, 1984.
December 8, 2002
BIBLIOGRAPHY 535
[HnMoT85] L. Henkin, D. Monk, and A. Tarski. Cylindic Algebras
Part II. North-Holland Publ. Co, 1985.
[He63] H. Hermes. Einf uhrung in die mathematische Logik. Teubner
Verlag, Stuttgart, 1963. English translation: Introduction
to Mathematical Logic. Unitext Springer-Verlag, 1973.
[H72] H. Herrlich. Perfect subcategories and factorizations. In:
Topics in Topology, volume 8 of Coll. Math. Soc. Janos
Bolyai, pages 387403, (Keszthely, Hungary), 1972.
[HS73] H. Herrlich and G.E. Strecker. Category Theory an In-
troduction. Allyn and Bacon, 1973. 2nd ed.: Heldermann-
Verlag.
[Hi81] B.H. Hien. Formulas and ultraproducts in categories revis-
ited. Preprint of Math. Inst. Hungar. Acad. Sci., No. 56,
1981.
[Hi83] B.H. Hien. On a problem of algebraic model theory. Bull. of
Section in Logic, Warszaw-Lodz, 11:103108, 1983.
[Hi84] B.H. Hien. In which categories are rst order axiomatiz-
able hulls characterizable by ultraproducts? Periodica Math.
Hungar., 15, 1984.
[HiN81] B.H. Hien and I. Nemeti. Problems with the category the-
oretic notion(s) of ultraproducts. Bull. of Section of Logic,
Wroclaw, 10(4):122127, 1981.
[Hi82] B.H. Hien and I. Sain. Elementary classes in the injective
subcategories approach to model theory. Preprint of Math.
Inst. Hungar. Acad. Sci., No. 15, 1982.
[HiSa83] B.H. Hien and I. Sain. In which categories are rst order
axiomatizable hulls characterizable by ultraproducts? Cahier
Top. Geo. Di. XXIV-2, pages 215222, 1983.
[Hg63] P.J. Higgins. Algebras with a scheme of operators. Math.
Nachr., 27:115132, 1963.
December 8, 2002
536 BIBLIOGRAPHY
[Hoe63] H.J. Hoehnke. Charakterisierung der Kongruenzklassen-
Teilsysteme in binaren partiellen Algebren. Math. Nachr.,
25:5964, 1963.
[Hoe72] H.J. Hoehnke. Superposition partieller Funktionen. Studien
zur Algebra und ihre Anwendungen, pages 726. Akademie
Verlag, Berlin, 1972. ed.: Hoehnke.
[Hoe77a] H.J. Hoehnke. Bemerkungen zum Kalk ulbegri f ur partielle
Algebren im Sinne von R.Hartwig. Manuscript, 1977.
[Hoe77] H.J. Hoehnke. On partial recursive denitions and programs.
In: Fund. Comp

Theory, volume 56 of Lect. Notes Comp.
Sci., pages 262274, (Proc. int. Conf. Poznan-Kornik 1977),
1977.
[Hoe82] H.J. Hoehnke. Fully invariant algebraic closure systems
of congruences and quasivarieties of algebras. Manuscript,
1982.
[Hoe76] H.J. Hoehnke. On partial algebras. In: Universal algebra,
volume 29 of Colloq. Math. Soc. J. Bolyai, pages 373412,
(Proc. Coll. Esztergom 1977), 1982. North-Holland Publ.
Co., Amsterdam. Preprint Akad. d. Wiss. d. DDR, Berlin,
1976.
[Hoe84a] H.J. Hoehnke. On Yoneda-Schreckenbergers embedding of
a class of monoidal categories. Submitted to Beitrage zur
Algebra und Geometrie, 1984.
[Hoe84] H.J. Hoehnke. On the Lemma of Yoneda-Schreckenberger in
certain monoidal categories. Preprint, 1984.
[Hoe86] H.J. Hoehnke. On quasivarieties of partial algebras. Their
generation and their subquasivarieties I. Beitr. Alg. Geom.,
21:145160, 1986.
[Hoe86a] H.J. Hoehnke. On quasivarieties of partial algebras. Their
generation and their subquasivarieties II. Beitr. Alg. Geom.,
22:103114, 1986.
December 8, 2002
BIBLIOGRAPHY 537
[Hoe87] H.J. Hoehnke. Monoidal structure of Malcev clones, their
theories and completions. Seminarber. Humboldt-Univ.
Berlin, Sekr. Math., 93:5769, 1987.
[Hoef70] H. Hoft. Equations in partial algebras. Dissertation, Univ.
of Houston, Texas, 1970.
[Hoef72] H. Hoft. Operators on classes of partial algebras. Algebra
Universalis, 2:118127, 1972.
[Hoef73] H. Hoft. Weak and strong equations in partial algebras. Al-
gebra Universalis, 3:203215, 1973.
[Hoef73a] H. Hoft. A characterization of strong homomorphisms. Coll.
Math., 28:189193, 1973.
[Hoef74] H. Hoft. On the semilattice of extensions of a partial algebra.
Coll. Math., 30:193201, 1974.
[Hoeft75] M. Hoft. The order-sum in classes of partially ordered alge-
bras. Journ. Reine und Angew. Math., 278/279:7892, 1975.
[I66] A.A. Iskander. Partial universal algebras with preassigned
lattices of subalgebras and correspondences. Mat. Sborn.
(N.S.), 70(112):438456, 1966. Russian.
[I71] A.A. Iskander. Subalgebrassystems of powers of partial uni-
versal algebras. Pac. Journ. Math., 38:457463, 1971.
[I73] A.A. Iskander, editor. On subalgebras of partial universal
algebras, (Proc. Conf. Univ. Houston), 1973.
[Is86] F.A. Ismajlov. Superassociative algebras of partial ho-
momorphisms of graphs. Doklady Akad. Nauk Az. SSR,
42(8):36, 1986. Russian.
[IsMf86] F.A. Ismajlov and L.G. Mustafaev. Abstract characteristics
of superassociative algebras of homomorphisms. Izv. Akad.
Nauk Az. SSR, Ser. Fiz.-Tekh. Mat. Nauk, 1:37, 1986. Rus-
sian.
December 8, 2002
538 BIBLIOGRAPHY
[In81] L. Ivanov. P-recursiveness in iterative combinatory spaces.
Serdica, 7:281297, 1981. Russian.
[Jak81] D. Jakubikova-Studenovska. On congruence lattices of nite
partial unary algebras. Comment. Math. Univ. Carolina,
22:357364, 1981.
[Jak82] D. Jakubikov a-Studenovsk a. Partial monounary algebras
with common congruence relations. Czech. Math. Journ.,
32(107):307326, 1982.
[Jak85] D. Jakubikov a-Studenovska. Endomorphisms and connected
components of partial monounary algebras. Czech. Math. J.,
35(110):467490, 1985.
[Jak86] D. Jakubikov a-Studenovsk a. Endomorphisms of partial
mono-unary algebras. Czech. Math. Journ., 36(111):376
392, 1986.
[Jak88a] D. Jakubikova-Studenovska. Convex subsets of partial mo-
nounary algebras. Czech. Math. Journ., 38(113):655672,
1988.
[Jak88] D. Jakubikova-Studenovska. On completions of partial mo-
nounary algebras. Czech. Math. Journ., 38(113):256268,
1988.
[Jak89] D. Jakubikov a-Studenovsk a. On the lattice of convex subsets
of a partial monounary algebra. Manuscript, 1989.
[Jak90b] D. Jakubikova-Studenovska. Convex subsets of partial mo-
nounary algebras. MR 90e:08002, 1990.
[Jak90a] D. Jakubikov a-Studenovsk a. On the lattice of convex subsets
of a partial monounary algebra. MR 90h:08004, 1990.
[Jak92] D. Jakubikova-Studenovska. Intervals in partial monounary
algebras. MR 92k:08012, 1992.
[Ja83] T.M.V. Janssen. Foundations and applications of Montague
grammar. Mathematisch Centrum, Amsterdam, 1983.
December 8, 2002
BIBLIOGRAPHY 539
[Ja86] T.M.V. Janssen. Foundations and applications of Montague
grammar. CWI Tracts 28, Mathematisch Centrum, Amster-
dam, 1986. Part 2: Applications to natural language.
[Jz82] G. Jarzembski. Partial monadic algebras. Bull. Acad. Sci.
Polon., XXX(910):407413, 1982.
[Jz83] G. Jarzembski. Spectral algebraic theories and their algebras.
In: Category Theory, volume 1920 of Seminarberichte Fer-
nuniversitat Hagen, pages 225243, (Proc. Conf. Oberwol-
fach 1983), 1983.
[Jz84] G. Jarzembski. Partially monadic functors. Journal of Pure
and Appl. Alg., 32:315326, 1984.
[Jz88] G. Jarzembski. Weak varieties of partial algebras. Alg.
Univ., 25:247262, 1988.
[Jz92] G. Jarzembski. Programs in partial algebras - a categorical
approach. MR 92g:68005, 1992.
[JzLk84] G. Jarzembski and M. Lema nczyk. Birkho variety theorem
for monadic algebras over epireective subcategories of -
models and its connection with a problem of the existence of
nonsurjective epimorphisms. Demonstr. Math., 17:939954,
1984.
[JeSl91] J. Jezek and V. Slavik. Free lattices over join-trivial partial
lattices. MR 91a:06007, 1991.
[J77] R. John. A note on implicational subcategories. In:
B. Csakany and J. Schmidt, editors, Contributions to Uni-
versal Algebra, volume 17 of Colloq. Math. Soc. J. Bolyai,
pages 213222, (Proc. Coll. Szeged 1975), 1977. North-
Holland Publ. Co., Amsterdam.
[J75] R. John. G ultigkeitsbegrie f ur Gleichungen in partiellen
Algebren. Math. Zeitschrift, 159:2535, 1978. Dissertation
TH Darmstadt 1975.
December 8, 2002
540 BIBLIOGRAPHY
[J78] R. John. G ultigkeitsbegrie f ur Gleichungen in partiellen
Algebren. Math. Zeitschrift, 159:2535, 1978.
[Jn71] J.S. Johnson. Congruences and strong congruences of partial
algebras. Manuscript, 1971.
[JoSe67] J. Johnson and R.L. Seifert. A survey of multi-unary alge-
bras. Report, Univ. of California, Berkeley, 1967.
[Js72] B. Jonsson. Topics in universal algebra. Lect. Notes in
Math., 250, 1972. Springer-Verlag.
[Kaa83] K. Kaarli. Compatible function extension property. Algebra
Universalis, 17:200207, 1983.
[Kai84] H.K. Kaiser. Interpolation in universal algebra. In: Univer-
sal algebra and its links with logic, algebra, combinatorics
and computer science, volume 4 of Res. Expo. Math., pages
2940, (Proc. 25. Arbeitstag. Allg. Alg., Darmstadt 1983),
1984.
[KmAr84] S. Kamin and M. Archer. Partial implementations of ab-
stract data types: A dissenting view of errors. In: Semantics
of data types, volume 173 of Lect. Notes Comp. Science,
pages 317336, (Proc. int. Symp. Sophia-Antipolis/France
1984), 1984.
[Ka81] H. Kaphengst. What is computable for abstract data types?
In: Fundamentals of Computation Theory, volume 117 of
Lect. Notes Comp. Sci., pages 173181, (Proc. Conf. Szeged
1981), 1981. Springer-Verlag, Berlin.
[Ka82] H. Kaphengst. Zum Aufbau einer mehrsortigen Logik.
Manuscript, Dresden, 1982.
[KaRe77] H. Kaphengst and H. Reichel. Initial algebra semantics for
non-context-free languages. In: Fundamentals of Comp. Sci.,
volume 56 of Lect. Notes Comp. Sci., pages 120126, (Proc.
int. Conf. Poznan-Kornik 1977), 1977. Springer-Verlag.
December 8, 2002
BIBLIOGRAPHY 541
[K64] C.R. Karp. Languages with expressions of innite length.
North-Holland Publ. Co., 1964.
[Ky] D. Kelly. Complete rules of inference for universal sentences.
Studia Sci. Math. Hungar.
[Kel80] G.M. Kelly. A unied treatment of transnite constructions
for free algebras, free monoids, colimits, associated sheaves
etc. Seminarber. Fachber. Math., Fernuniv. 6, pages 582,
1980.
[KelPu78] G.M. Kelly and A. Pultr. On algebraic recognition of direct-
product decompositions. J. Pure Appl. Algebra, 12:207224,
1978.
[Ke65] R. Kerkho. Eine Konstruktion absolut freier Algebren.
Math. Annalen, 158:109112, 1965.
[Ke70] R. Kerkho. Gleichungsdenierbare Klassen partieller Alge-
bren. Math. Annalen, 185:112133, 1970.
[Kla83] H.A. Klaeren. Algebraische Spezikation. Springer Verlag,
1983.
[Kl52] S.C. Kleene. Introduction to Metamathematics. North-
Holland Publ. Co, 1952. 5th reprint, 1967.
[KnRo83] E. Knuth and L. Ronyai. Closed convex reference schemes.
In: A junction between computer science, cylindric and par-
tial algebras, (Proc. IFIP Conf. Hungary), 1983. Preprint,
MTA-SZTAKI, March, 1982. II/32.
[KoSp68] S. Kochen and E.P. Specker. The problem of hidden variables
in Quantum Mechanics. J. Math. Mech., 17:5987, 1968.
[KogSo93] S.R. Kogalovskii and V.V. Soldatova. Congruence lattices
of partial algebras II (Russian). MR 93h:08005, 1993.
[Kol82] J. Kollar. The category of idempotent 2-unary algebras con-
taining a given subalgebra. In: Universal algebra, volume 29
of Colloq. Math. Soc. J. Bolyai, pages 483493, (Proc. Col-
loq. Esztergom 1977), 1982.
December 8, 2002
542 BIBLIOGRAPHY
[Kop76a] O. Kope cek. Construction of all machine homomorphisms.
Bull. Acad. Sci. Polon., Ser. Sci. Math. Astron. Phys.,
24:655658, 1976.
[Kop76] O. Kope cek. Homomorphisms of partial unary algebras.
Czech. Math. Journ., 26(101):108127, 1976.
[Kop77] O. Kope cek. The category of connected partial unary alge-
bras. Czech. Math. Journ., 27(102):415423, 1977.
[Kop78b] O. Kope cek. Existence of monomorphisms of partial unary
algebras. Czech. Math. Journ., 28(103):426473, 1978.
[Kop78] O. Kope cek. Homomorphisms of machines I. Arch. Math.
(Brno), 1:4550, 1978.
[Kop78a] O. Kope cek. Homomorphisms of machines II. Arch. Math.
(Brno), 2:99108, 1978.
[Kop79] O. Kope cek. The category of connected partial and complete
unary algebras. Bull. Acad. Pol. Sci., Ser. Sci. Math., 27:337
344, 1979.
[Kop83] O. Kope cek. A note on some cardinal functions on unary
algebras. In: Contr. to general alg. 2, pages 221227, (Proc.
Conf. Klagenfurt 1982), 1983.
[Kor81] W. Korczynski. Some properties of partially monoidal cate-
gories I. Pr. Nauk., Wysza Sk. Pedagog. Czestochowie, Ser.
Mat.-Przyr., 4:2134, 1981. Polish.
[Kc86] I. Korec. L-continuous partial functions. In: Universal alge-
bra, volume 43 of Colloq. Math. Soc. J. Bolyai, pages 217
240, (Colloq. Szeged 1983), 1986.
[Kos90] P. Kosiuczenko. Malcev type conditions for partial algebras.
Manuscript, Politechnika Warszawska, 1990.
[Kb83] V. Koubek. Full embeddability into categories of generalized
algebras. Algebra Universalis, 17:120, 1983.
December 8, 2002
BIBLIOGRAPHY 543
[Kb84] V. Koubek. Subalgebra lattices, simplicity and rigidity. Acta
Sci. Math., 47:7183, 1984.
[KbRt79] V. Koubek and J. Reiterman. Categorical constructions of
free algebras, colimits and completions of partial algebras. J.
Pure Appl. Algebra, 14:195231, 1979.
[Kra83] P. Kratochvil. Sequential convergence as a Partial operation.
In: General topology and its relations to modern analysis
and algebra V, volume 3 of Sigma Ser. Pure Math., pages
442447, (Proc. 5th Topol. Symp. Prag 1981), 1983.
[Ki86] H.J. Kreowski. Partial algebra ow from algebraic spezica-
tions. Unviversitat Bremen, Manuscript, 1986.
[KiSm84] H.J. Kreowski and H.W. Schmidt. Some algebraic concepts
of the specication language Segras and their initial se-
mantics. GMD-Stud., 93, 1984.
[Kk69] V.T. Kulik. On the lattice of strong congruence relations
of a partial universal algebra I. All-Union Symposium on
Semigroup Theory, Sverdlovsk, pages 4142, 1969. Russian.
[Kk70] V.T. Kulik. On greatest strong congruence relations of par-
tial universal algebras. Studies on Algebra 2, Izdat. Saratov
Univ, pages 4046, 1970. Russian.
[Kk71] V.T. Kulik. The compact elements of the lattice of all strong
congruence relations of a partial universal algebra. Ordered
sets and lattices 1, Izdat. Saratov Univ, pages 4348, 1971.
Russian.
[Ku77] I. Kupka. Partial algebras for representing semantics of in-
formation processing. Habilitationsschrift, Hamburg, 1977.
Preprint am Fachbereich Informatik Hamburg, 1980.
[KurWo86] A. Kurpiel and B. Wojdylo. Description of partial algebras
by segments. In: Univ. Alg., volume 43 of Colloq. Math. Soc.
J. Bolyai, pages 283292, (Proc. Coll. Szeged 1983), 1986.
December 8, 2002
544 BIBLIOGRAPHY
[Kue86] K.D. K ursten. The completion of the maximal Op*-algebra
on a Frechet domain. Publ. Res. Inst. Math. Sci., 22:151
175, 1986.
[La69] W.A. Lampe. On related structures of a universal algebra.
PhD thesis, Pennsylvania State University, 1969.
[La72] W.A. Lampe. The independence of certain related struc-
tures of a universal algebra I. Partial algebras with useless
operations and other lemmas. Algebra Universalis, 2:99112,
1972.
[La72a] W.A. Lampe. The independence of certain related structures
of a universal algebra II. Algebra Universalis, 2:270283,
1972.
[La72b] W.A. Lampe. The independence of certain related structures
of a universal algebra III. Algebra Universalis, 2:286295,
1972.
[La72c] W.A. Lampe. The independence of certain related structures
of a universal algebra IV. Algebra Universalis, 2:296302,
1972.
[LaMy80] W.A. Lampe and D. Myers. Elementary properties of free
extensions. Algebra Universalis, 11:269284, 1980.
[LePa80] D. Lehmann and A. Pasztor. Epis need not be dens. Theo-
retical Computer Science, 17:151162, 1982. Preprint Techn.
Dept. of Comp. Sci., Haifa, 1980; On a conjecture of
Meseguer.
[Ld75] V.B. Lender. Replica equivalences on lattices. Mat. Zap.,
9(3):6072, 1975. Russian.
[Lev79] A. Levy. Basic Set Theory. Springer-Verlag, 1979.
[Ly78] J. Levy-Bruhl. Demi groupe large modulaire. Appli-
cations (Large modular semigroups. Applications). In:
Sistemi binari e loro applicazioni, pages 6577, (Conv.
Taormina/Messina), 1978. French.
December 8, 2002
BIBLIOGRAPHY 545
[Lr77] C.C. Lindner. Intersection preserving nite embedding the-
orems for partial quasigroups.

Aquationes math., 15:7990,
1977.
[Lr77a] C.C. Lindner. Totally symmetric and semi-symmetric quasi-
groups have the intersection preserving nite embeddability
property. Periodica mathematica Hung., 8:3339, 1977.
[Lin83] R. Linn. Kompositionsvollstandigkeit von Algebren mit
(m, n)-stelligen Operationen. Mitt. Math. Semin. Gieen,
160, 1983.
[Lp82a] J. Lipinska. Connection of a generalized inverse semigroup
with other algebraic systems with one binary partial opera-
tion. Zesz. Nauk. Politechn. Slask. 686, Mat. Fiz., 39:305
310, 1982. Polish.
[Lp82] J. Lipinska. A generalized inverse semigroup and its prop-
erties. Zesz. Nauk. Politechn. Slask. 686, Mat. Fiz., 39:287
295, 1982. Polish.
[Liu85] S. Liu. A class of partial algebraic systems. J. Beijing Norm.
Univ., Nat. Sci., 1:1724, 1985. Chinese, English summary.
[Lj76] E.S. Ljapin. Abstract characterization of partial groupoids of
words with synonyms. In: Algebraic Theory of Semigroups,
volume 20 of Colloq. Math. soc. J. Bolyai, pages 341356,
(Proc. Coll. Szeged 1976), 1979.
[Lj79] E.S. Ljapin. The weak associativity of partial operations.
In: Semigroup varieties and semigroups of endomorphisms,
pages 95112, 1979. Russian.
[Lj81] E.S. Ljapin. Partielle Operationen in der Theorie der Halb-
gruppen. In: Semigroups, volume 855 of Lect. Notes Math.,
pages 3348, (Proc. Conf. Oberwolfach 1978), 1981.
[Lj82a] E.S. Ljapin. On interior extension of partial actions to com-
plete associative actions. Sov. Math., 26(7):4955, 1982.
translation from: Izv. Vyssh. Uchebn. Zaved., Mat. 1982,
No.7 (242),1982, pages 4044.
December 8, 2002
546 BIBLIOGRAPHY
[Lj82] E.S. Ljapin. An abstract characterization of a class of partial
transformation groupoids. Sov. Math., 26(6):3039, 1982.
translation from: Izv. Vyssh. Uchebn. Zaved., Mat, 1982,
No.6 (241), 1981, pages 3036.
[Lu76] H. Lugowski. Grundz uge der Universellen Algebra. Teubner,
Leipzig, 1976.
[Ml71] S. MacLane. Categories for the Working Mathematician.
Springer-Verlag, 1971.
[Md85] J.J. Madden. l-groups of piecewise linear functions. In: Or-
dered algebraic structures, volume 99 of Lect. Notes Pure
Appl. Math., pages 117124, (Proc. AMS Meet. Cincinnati
1982), 1985.
[MhMk84a] B. Mahr and J.A. Makowski. Characterizing specication
languages which admit initial semantics. Theor. Computer
Science, 31:4959, 1984.
[MhMk84] B. Mahr and J.A. Makowski. An axiomatic approach to se-
mantics of specication languages. In: A.B. Cremers and
H.P. Kriegel, editors, Theor. Computer Science, volume 145
of Lect. Notes Comp. Sci., pages 211219, (6th GI Confer-
ence, Dortmund 1983), 1984. Springer-Verlag, Berlin.
[Mk84] J.A. Makowski. Model theoretic issues in theoretical com-
puter science. Part I. In: G. Lolli, G. Longo, and A. Marcja,
editors, Relational data bases and abstract data types, Logic
Coll. 82, pages 303345. North-Holland Publ. Co, 1984.
[Mal71] A.I. Malcev. Quasi-primitive classes of abstract algebras.
The Metamath. of Alg. Syst., pages 2731. North-Holland
Publ. Co, 1971.
[Mal73] A.I. Malcev. Algebraic Systems. Springer-Verlag, 1973.
[Mal74] A.I. Malcev. Algorithmen und rekursive Funktionen. Vieweg,
1974.
December 8, 2002
BIBLIOGRAPHY 547
[Ma78] S.S. Marchenkov. A method for constructing maximal sub-
algebras of algebras of general recursive functions. Alge-
bra Logic, 17:383392, 1979. translation from: Algebra
Logika 17, pages 581595, 1978.
[Mar58] E. Marczewski. A general scheme of independence in mathe-
matics. Bull. de lAcad. Polonaise des Sci., Ser. Math. Astr.
et Phys., 6:731736, 1958.
[Mek84] I. Marek. Consequence operations dened by partial matri-
ces. Rep. Math. Logic, 17:4755, 1984.
[Mr83a] Zs. Markusz, editor. Dierent validity concepts in many-
sorted logic, (Proc. Conf. Gyor, Hungary), 1983. to appear.
[Mr83] Zs. Markusz. On rst-order many-sorted logic. Mta Sztaki
Tanulmanyok, 151:185, 1983.
[Mw71] W. Markwald. Pradikatenlogik mit partiell denierten Funk-
tionen. Arch. Math. Logik, 14:1023, 1971.
[Mt76] G. Matthiessen. Theorie der heterogenen Algebren. Univer-
sitat Bremen, MathematikArbeitspapiere Nr. 3 (Disserta-
tion), 1976.
[Mn80] J. Meakin. Constructing biordered sets. In: Semigroups,
pages 6784, (Proc. Conf. Clayton/Australia 1979), 1980.
[Mer83] J. Mervartova. On some properties of genomorphisms of
C-algebras. Arch. Math., Brno, 19:99107, 1983.
[Mb78] I. Mikenberg. From total to partial algebras. In: Math. Logic,
volume 39 of Lect. Notes on Pure and Appl. Math., pages
203223, (Proc. First Brasil Conf. State Univ. Campinas
1977), 1978. Dekker, New York.
[Mb80] I. Mikenberg. A closure for partial algebras. In: Mathemat-
ical logic in Latin America, volume 99 of Studies on Logic
and Found. of Math., pages 238253, (Proc. Symp. Santiago
1978), 1980. North-Holland Publ. Co., Amsterdam.
December 8, 2002
548 BIBLIOGRAPHY
[Mik84] A.B. Mikhailov. Semigroups of matrices with inserted mul-
tiplication over a semigroup. In: Properties of semigroups,
pages 104121, Leningrad, 1984. Russian.
[Moe81] A. Mobus. Relational-Algebren. Dissertation,
Mathematisch-Naturwissenschaftliche Fakult at der Univer-
sitat D usseldorf, 1981.
[Moe83] A. Mobus. Alexandrov compactication of relational alge-
bras. Arch. Math., 40:526537, 1983.
[Mo76] J.D. Monk. Mathematical Logic. Springer-Verlag, 1976.
[Ms51] A. Mostowski. On the rules of proof in the pure functional
calculus of the rst order. J. Symbolic Logic, 16:107111,
1951.
[MzR80] J.C. Muzio and I.G. Rosenberg. Large classes of functionally
complete operations I. In: Multiple-valued logic, pages 94
101, (Proc. 10th int. Symp. Evanston), 1980.
[MzR82] J.C. Muzio and I.G. Rosenberg. Large classes of functionally
complete groupoids I.

Aquationes Math., 25:274288, 1982.
[Na79] K.S.S. Nambooripad. Structure of regular semigroups. In:
Regular Semigroups, pages 101117, (Proc. Symp. Northern
III. Univ. DeKalb), 1979.
[N76] I. Nemeti. Axiomatisability of classes of partial algebras via
category theory. Dissertation, Budapest, 1976.
[N76a] I. Nemeti. On a property of the category of partial algebras.
Comp. Linguistic & Comp. Lang., 11:510, 1976.
[N75] I. Nemeti. Partial algebra theory via categories. Disserta-
tion, Hungar. Acad. Sci., 1975 (in Hungarian); appeared as
Partial Algebras (in Hungarian) as booklet by Szamalk,
1976.
[N78] I. Nemeti. From hereditary classes to varieties in abstract
model theory and partial algebra. Beitrage zur Algebra und
Geometrie, 7:6978, 1978.
December 8, 2002
BIBLIOGRAPHY 549
[N81] I. Nemeti. Connections between cylindric algebras and initial
algebra semantics of CF languages. II. In: Mathematical
Logic in computer Science, volume 26 of Colloq. Math. Soc.
J. Bolyai, pages 561605, (Proc. Salg otarj an 1978), 1981.
North-Holland Publ. Co., Amsterdam.
[N82] I. Nemeti. On notions of factorization systems and their
applications to cone- injective subcategories. Periodica Math.
Hungar., 13(3):229235, 1982.
[NSa82] I. Nemeti and I. Sain. Cone-implicational subcategories and
some Birkho-type theorems. In: Universal Algebra, vol-
ume 29 of Colloq. Math. Soc. J. Bolyai, pages 535578,
(Proc. Coll. Esztergom 1977), 1982. North-Holland Publ.
Co., Amsterdam.
[Nn83] A.A. Nikitin. On freely generated projective planes. Algebra
Logic, 22:4547, 1983. translation from: Algebra Logika 22,
No.1, 1983, pages 6178.
[Nk87] T. Nipkow. Are homomorphisms sucient for behavioural
implementations of deterministic and nondeterministic data
types? In: Stacs 87, Theoretical aspects of computer sci-
ence, volume 247 of Lect. Notes Comput. Sci., pages 260
271, (Proc. 4th annual Symp. Passau), 1987.
[No82a] M. Novotny. Characterization of the category of co nnected
machines. Found Control Eng., 7:169179, 1982.
[No82] M. Novotny. Contextual grammars vs. context-free algebras.
Czech. Math. J., 32(107):529547, 1982.
[O81] A. Obyulowicz. Theories classifying partial algebras. sub-
mitted to: Proceedings Conf. on Category Theory, Gum-
mersbach, 1981.
[O82] A. Obyulowicz. The logic of categories of partial func-
tions and its applications. Diss. Math., 241:164165, 1986.
Preprint 1982.
December 8, 2002
550 BIBLIOGRAPHY
[Oe81] G. Oehme. Konformfreie Algebren und ihr Zusammenhang
mit speziellen ECE-Variet aten. Diplom thesis, Fachbereich
Math., TH Darmstadt, 1981.
[Pal91] M. Palasinski. Remark on subdirect representation theorem
for ECE-varieties of partial algebras. Studia Scient. Math.
Hungar., 26:471473, 1991.
[Ps71] A. Pasini. Sul reticolo di congruenzi forti in algebre parziali
nitarie. Boll. Un. Mat. Ital., 4:630634, 1971.
[Pa74] A. Pasztor. Generalizing identites to partial algebras through
category theory. Master thesis, Eotv os Lorand Univ., Bu-
dapest, 1974.
[Pa77] A. Pasztor. A category-theoretical characterization of surjec-
tive homomorphisms of partial algebras. Studia Sci. Math.
Hung., 12:251256, 1977.
[Pa79] A. Pasztor. Faktorisierungssysteme in der Kategorie
der partiellen Algebren Kennzeichnung von (Homo)-
Morphismen. Dissertation, TH Darmstadt, 1979, appeared
in: Hochschulsamml. Naturw.: Math. Bd. 1, Hochschulver-
lag, Freiburg, 1979.
[Pa82] A. Pasztor. Ordered algebras are monoreective in the cat-
egory of ordered partial algebras. In: Trees in Algebra
and Programming, pages 227253, (Proc. 7th Coll. Lille,
France), March 46 1982.
[Pa83] A. Pasztor. Chain-continuous algebras - a variety of partial
algebras. Ann. Soc. Math. Pol. Ser.

IV, Fundam. Inform.,


6:275288, 1983.
[Pa83b] A. Pasztor. Epis of some categories of Z-continuous partial
algebras. Acta Cybernet., 6:111123, 1983.
[Pa83a] A. Pasztor. The epis of the category of ordered algebras and
Z-continuous homomorphisms. Cahiers Top. Geom. Di.,
24:203214, 1983.
December 8, 2002
BIBLIOGRAPHY 551
[Pa86] A. Pasztor. Continuity of operations does change the epis.
In: Algebra, combinatorics and logic in computer science,
volume 42 of Colloq. Math. Soc. J.Bolyai, pages 645648,
(Colloq. Gyor, 1983, Vol.2), 1986.
[Pa84] A. Pasztor. On the variety concept for -continuous alge-
bras. Applications of a general approach. In: Mathematical
foundations of programming semantics, volume 239 of Lect.
Notes Comp. Sci., pages 215248, (Proc. Conf. Kansas State
Univ., Manhatten, April 1112, 1985), 1986. Dept. Math.,
Carnegie Mellon Univ., Pittsburgh, USA, Research report
84-2, 1984.
[Pa91] A. Pasztor. A sup-preserving completion of ordered partial
algebras. MR 91k:68138, 1991.
[PaSta86] A. Pasztor and R. Statman. Scott induction and closure
under omega-sups. Theor. Comp. Sci., 43:251263, 1986.
[Pe83] U. Petermann. On algorithmic logic with partial operations.
In: Logics of programs and their applications, volume 148 of
Lect. Notes Comp. Sci., pages 213223, (Proc. Symp. Poz-
nan 1980), 1983.
[Ph84] M. Petrich. Inverse semigroups. Pure and Applied Mathe-
matics. J.Wiley & Sons, New York, 1984.
[Pf67] J. Pfanzagl. Homomorphisms for distributive operations in
partial algebras. Applications to linear operators and mea-
sure theory. Math. Zeitschr., 99:270278, 1967.
[Pc64] H.E. Pickett. Subdirect representations of relational systems.
Fund. Math., 56:223240, 1964.
[Pi68] R.S. Pierce. An Introduction to Abstract Algebras. Holt,
Rinehart and Winston, 1968.
[Pig66] D. Pigozzi. On some operations on classes of algebras. No-
tices Amer. Math. Soc., 13:829, 1966.
December 8, 2002
552 BIBLIOGRAPHY
[PidSic85] D. Pigozzi and J. Sichler. Homomorphisms of partial and of
complete Steiner triple systems and quasigroups. In: Univer-
sal algebra and lattice theory, volume 1149 of Lect. Notes
Math., pages 224237, (Proc. Conf. Charleston 1984), 1985.
[Pz84] G. Pilz. Algebra. Ein Reisef uhrer durch die schonsten Ge-
biete. Schriftenreihe f ur Mathematik, 5. Linz: Univer-
sitatsverlag Rudolf Trauner, III, 1984.
[Pl85] G. Plotkin, editor. Denotational Semantics with Partial
Functions, Lecture Notes Logic, (Meeting, Stanford Univ.,
July 8-19), 1985.
[Po73] V.S. Poythress. Partial morphisms on partial algebras. Al-
gebra Universalis, 3:182202, 1973.
[Pr84] S. Premchand. Independence of axioms for biordered sets.
Semigroup Forum, 28:249263, 1984.
[PuTr80] A. Pultr and V. Trnkova. Combinatorial, Algebraic and
Topological Representations of Groups, Semigroups and Cat-
egories. Academia, Praha, 1980.
[Qu84] R.W. Quackenbush. Equationally complete discriminator
varieties of groupoids. Proc. Am. Math. Soc., 90:203206,
1984.
[Ra79] J. Rachunek. Quasi-orders of algebras. Cas. Pestovani Mat.,
104:327337, 1979.
[Rc81] J. Rachunek. Reexive and antisymmetric relations and
their systems. Czech

Math. J., 31(106):521530, 1981.
[Ra70] Z. Ras. On the algebraic properties of computing machines.
Bull. Acad. Polon. Sci., Ser. Sci. Math. Astron. Phys.,
18:613618, 1970.
[Ra72] Z. Ras. Congruence relations in computing machines. Bull.
Acad. Polon. Sci., Ser. Sci. Math. Astron. Phys., 20:313317,
1972.
December 8, 2002
BIBLIOGRAPHY 553
[Red87] E. Redi. A characteristic of polycategories of quasi-group
functions. Tartu Riikl.

Ul. Toimetised, 764:98108, 1987.
Russian.
[Re78a] H. Reichel, editor. Limit-colimit doctrines in computer sci-
ence, Banach Center Publications (to appear), (Banach Cen-
ter Semester, Warszawa), 1978.
[Re79] H. Reichel. Theorie der

Aquoide. Dissertation B, Humboldt-
Universitat, Berlin, 1979.
[Re78b] H. Reichel, editor. Algebraic specications of abstract data
types, volume 9 of Banach Center Publications (to appear),
(Banach Center Semester, Warszawa 1978), 1980.
[Re81] H. Reichel. Homomorphism theorem for equationally partial
algebras. Comment. Math. Univ. Carol., 24:115, 1983.
[Re84a] H. Reichel. Partial algebras a sound basis for structural
induction. In: Universal algebra and its links with logic,
algebra, combinatorics and computer science, number 4 in
Res. Expo. Math., pages 230240, (Proc. 25. Arbeitstagung
Allgemeine Algebra, Darmstadt 1983), 1984.
[Re84] H. Reichel. Structural Induction on Partial Algebras. Part
II of Introduction to Theory and Applications of Partial
Algebras, Part I: B86, Mathematical Research, Vol. 18,
Akademie-Verlag, Berlin, 1984.
[Re85] H. Reichel. Behavioural validity of conditional equations in
abstract data types. In: Contr. to general algebra 3, pages
301324, (Proc. Conf. Vienna 1984), 1985.
[Re86] H. Reichel. Initial restrictions of behaviour. In: Formal
models in programming, pages 285295, (Proc. IFIP TC2
Working Conf. Wien 1985), 1986.
[Re87] H. Reichel. Initial computability, algebraic specications and
partial algebras. The Intern. Ser. Monogr. Comp. Sci., 2,
1987. Oxford Sci. Publ., Clarendon Press.
December 8, 2002
554 BIBLIOGRAPHY
[Ri82] H. Riedel. Category theory and universal algebra. Preprint,
Univ. of Waterloo Canada, 1982.
[Rob91] A. Robinson. Equational logic of partial functions under
Kleene equality: a complete and an incomplete set of rules.
MR 91c(or e):03031, 1991.
[Rg88] P. Rogowska. Total congruencesin unary partial algebras.
Master Thesis, Warsaw, 1988. Polish.
[R81] I.G. Rosenberg. Large classes of functionally complete
groupoids II. In: Multiple-valued logic, pages 259262,
(Proc. 11th int. Symp. Oklahoma City & Norman), 1981.
[R83] I.G. Rosenberg. Galois theory for partial algebras. In: Uni-
versal algebra and lattice theory, volume 1004 of Lect. Notes
Math., pages 257272, Proc. 4th int. Conf. Puebla 1982),
1983. Preprint 1983.
[R88] I.G. Rosenberg. Partial algebras and clones via one-point
extensions. Universite de Montreal, Rapports du recherche
du Dep. de Math. Stat., Nr. 88-16, 1988. to appear in: Contr.
to General Algebra 6.
[R91] I.G. Rosenberg. Partial algebras and clones via one-point
extension. MR 91j:08006, 1991.
[Rz66] V.V. Rozen. Partial idempotent operatives associated with
ordered sets. Izvest. VUZ, Mathem., 4, 1966. Russian.
[Rz69] V.V. Rozen. Partial -operatives and their ideals. Volzhskiy
Mat. Sb., 7, 1969.
[Rz71] V.V. Rozen. The partial operations min and inf in ordered
sets. Izdat Saratov Univ., Ordered sets and lattices 1, pages
7485, 1971. translation from: Uporjadoc. Mnozestv. Re-
setki 1, 1971, pages 7485.
[Rz73] V.V. Rozen. Partial operations in ordered sets. Izdat. Sara-
tov Univ, 1973. Russian.
December 8, 2002
BIBLIOGRAPHY 555
[Ru80] L. Rudak. The lattice of varieties of 1-algebras. Manuscript,
1980.
[Ru80a] L. Rudak. The lattice of weak varieties. Manuscript, 1980.
[Ru83] L. Rudak. A completeness theorem for weak equational logic.
Algebra Universalis, 16:331337, 1983.
[Sa74] I. Sain. Category theoretical investigations in order to gen-
eralize identities and quasi-identities. Master Thesis, Eotv os
Lorand Univ., Budapest, 1974. Hungarian.
[Sa77] I. Sain. On classes of algebraic systems closed with respect to
quotients. In: Algebra and its applications, volume 9 of Ba-
nach Center Publ., pages 127131, (Banach Center Semester
Warsaw 1978), 1980. PNW - Polish Scientic Publishers,
Warszawa. Preprint Math. Inst. Hung. Acad. Sci, 1977.
[Sa83] I. Sain. Applicability of category theoretic notions of ultra-
products. Mathematical Lapok, 31(1-3):143167, 1983. Hun-
garian.
[SbIgHo82] T. Sakabe, Y. Inagaki, and N. Honda. Rules of inference
for equations in many-sorted partial algebras. In: Theory of
formal languages and automata theory, volume VI of Res.
Inst. Math. Sci., pages 247255, (Proc. Symp. Kyoto 1982,
Kyoto Univ.), 1982. Japanese.
[Sj86] V.N. Salij. Quasi-Boolean lattices and associations. In: Uni-
versal algebra, volume 43 of Colloq. Math. Soc. J. Bolyai,
pages 429454, (Proc. Colloq. Szeged 1983), 1986.
[SauSto82] N. Sauer and M.G. Stone. A Galois correspondence between
algebras and endomorphisms. In: Universal algebra, vol-
ume 29 of Colloq. Math. Soc. J. Bolyai, pages 691695,
(Proc. Colloq. Esztergom 1977), 1982.
[Sche76] B. Schepull.

Uber Quasivariet aten von partiellen Algebren.
Dissertation, Akad. d. Wiss. d. DDR, Berlin, 1976.
December 8, 2002
556 BIBLIOGRAPHY
[Sch62] J. Schmidt. Die Charakteristik einer Allgemeinen Algebra I.
Arch. Math., XIII:457470, 1962.
[Sch64] J. Schmidt. Die Charakteristik einer Allgemeinen Algebra II.
Arch. Math., XV:286301, 1964.
[Sch64a] J. Schmidt. Some properties of algebraically independent sets
in algebras with innitary operations. Fund. Math., LV:123
137, 1964.
[Sch65] J. Schmidt.

Uber die Dimension einer partiellen Algebra mit
endlichen oder unendlichen Operationen. Zeitschr. Mathem.
Logik und Grundl. Math., 11:227239, 1965.
[Sch66a] J. Schmidt. Mengenlehre I (Einf uhrung in die axiomatische
Mengenlehre). Bibliogr. Institut, HTB, 56/56a, 1966.
[Sch66] J. Schmidt. A general existence theorem on partial algebras
and its special cases. Coll. Math., 14:7387, 1966.
[Sch68a] J. Schmidt. Direct sums of partial algebras and nal alge-
braic structures. Can. Journ. Math., 20:872887, 1968.
[Sch70] J. Schmidt. A homomorphism theorem for partial algebras.
Coll. Math., 21:521, 1970.
[Sm84] H.W. Schmidt. Towards a net-theoretic notion of type based
on predicate-transition nets. Arbeitspap. GMD, 117, 1984.
[Sch66b] J. Schmidt. Allgemeine Algebra. Lecture notes at the U. of
Bonn, 1965/66.
[Sch68] J. Schmidt. Seminar on Universal Algebra. Lecture notes
at the U. of Athens, Ga., 1967/68.
[Sre80] J. Schreckenberger.

Uber die Einbettung von dht-
symmetrischen Kategorien in die Kategorie der partiellen
Abbildungen zwischen Mengen. Preprint der Akad. d. Wiss.
d. DDR, ZAMM, 1980.
[Sr77] P. Schreiber. Grundlagen der Mathematik. VEB Deutscher
Verlag der Wissenschaften, Berlin, 1977.
December 8, 2002
BIBLIOGRAPHY 557
[Sc77] D. Scott. Identity and existence in intuitionistic logic. In:
M.P.Fourman, editor, Applications of Sheaves, volume 753
of Lect. Notes Math., pages 660696, (Durham Proceedings
1977), 1979. Springer-Verlag.
[Sz78] K. Seitz. Application of categorical and quasi-categorical
methods for investigation of systems in chemical engineer-
ing. Dep. Math. Karl-Marx-Univ. Econ. Budapest, 6:1623,
1978.
[SzBal79] K. Seitz and J. Balazs. Investigation of generating algo-
rithms of the partial group P(H$H, $), with special emphasis
on applications in chemical engineering. Notes on algebraic
systems II, Dep. Math. Karl-Marx-Univ. Econ. Budapest,
5:1023, 1979.
[SzBlBal76] K. Seitz, T. Blickle, and J. Balazs. The study of special
equations over the partial algebraic system S(F$$F$, $, $).
Dep. Math. Karl-Marx-Univ. Econ. Budapest, 5, 1976.
[Sy84] V.A. Shlyk. Subadditive characterization of the facets of
polyhedral sets related to partial algebra. Dokl. Akad. Nauk
Az. SSR, 28:980983, 1984. Russian.
[Si82] M. Siegmund-Schultze. Subdirectly irreducible and equation-
ally compact mono-unary partial algebras. Preprint Nr. 695,
TH Darmstadt, 1982.
[Si84] M. Siegmund-Schultze. Residual kleine Variet aten partieller
Algebren. Dissertation, Fachbereich Mathematik, TH Darm-
stadt, 1984.
[Sl59] J. Slomi nski. The theory of abstract algebras with inni-
tary operations. Dissertationes Mathematicae (Rozprawy
Matematyczne), Warszawa, 18, 1959.
[Sl60a] J. Slomi nski. On the common embedding of abstract quasi-
algebras into equationally denable classes of abstract alge-
bras. Bull. Acad. Polon. Sci., Ser. Sci. Math. Astron. Phys.,
8:277282, 1960.
December 8, 2002
558 BIBLIOGRAPHY
[Sl60] J. Slomi nski. On the embedding of abstract quasi-algebras
into equationally denable classes of abstract algebras. Bull.
Acad. Polon. Sci., Ser. Sci. Math. Astron. Phys., 8:1117,
1960.
[Sl62] J. Slomi nski. On the solving of systems of equations over
quasi-algebras and algebras. Bull. Acad. Polon. Sci., Ser.
Sci. Math. Astron. Phys., 10:627635, 1962.
[Sl64] J. Slomi nski. A theory of extensions of quasi-algebras to
algebras. Dissertationes Mathematicae (Rozprawy Matem-
atyczne), Warszawa, 40, 1964.
[Sl66] J. Slomi nski. A theory of p-homomorphisms. Coll. Math.,
14:135162, 1966.
[Sl68] J. Slomi nski. Peano-algebras and quasi-algebras. Dis-
sertationes Mathematicae (Rozprawy Matematyczne),
Warszawa, 62, 1968.
[Sl70] J. Slomi nski. On mappings between quasi-algebras. Coll.
Math., 21:521, 1970.
[Sl86] J. Slomi nski. On the characterization of inside presentations
and inside quotients of partial algebras. Demonstr. Math.,
19:977991, 1986.
[So83] F.N. Sokhatskij. On positional algebras. Mat. Issled.,
71:104117, 1983. Russian.
[So87] F.N. Sokhatskij. Positional algebras. The Belousov algebras.
Mat. Issled., 95:101120, 1987.
[St73] V. Strassen. Berechnungen in partiellen Algebren endlichen
Typs. Computing, 11:181196, 1973.
[S72] G.E. Strecker. Epireection operators vs perfect morphisms
and closed classes of epimorphisms. Bull. Austral. Math.
Soc., 7:359366, 1972.
December 8, 2002
BIBLIOGRAPHY 559
[Sv82] J. Svrcek. A contribution to the theory of decompositions of
partial algebras. Acta Univ. Palacki. Olomuc., Fac. Rerum
Nat. 69, Math. 20, pages 2325, 1982.
[Szi84] J. Szigeti. A note on reective subcategories dened by par-
tial algebras. Comment. Math. Univ. Carolina, 25:319323,
1984.
[Tm82] D. Tamari. A graphic theory of associativity and wordchain
patterns. In: Combinatorial theory, volume 969 of Lect.
Notes Math., pages 302320, (Proc. Conf. SchloRauis-
chholzhausen 1982), 1982.
[T84a] A. Tarlecki. Note: Quasi-varieties in abstract algebraic in-
stitutions. Preprint Univ. of Edinbourgh, 1984.
[T84] A. Tarlecki. On the existence of free models in abstract al-
gebraic institutions. Preprint, Univ. of Edinbourgh, 1984.
[T85] A. Tarlecki. On the existence of free models in abstract al-
gebraic institutions. Theor. Comput. Sci., 37:269304, 1985.
[TcWaWr78] J.W. Thatcher, E.G. Wagner, and J.B. Wright, editors. Data
type specication: Parametrization and the power of speci-
cation techniques, (Proc. of the 10th annual Symposium,
San Diego, Cal.), 1978.
[Th66] H. Thiele. Wissenschaftstheoretische Untersuchungen in
algorithmischen Sprachen I (Theorie der Graphschemata-
Kalk ule). VEB Deutscher Verlag der Wissenschaften, Berlin,
1966.
[Tin78] G. Tinhofer. On the simultaneous isomorphism of spe-
cial relations (isomorphism of automata). In: Datenstruk-
turen, Graphen, Algorithmen, pages 205213, (3. Fachta-
gung graphenth. Konz. Inf. Linz 1977), 1978.
[Ti81] J. Tiuryn. A survey of the logic of eective denitions.
Preprint MIT/LCS/TR-246, MIT Laboratoy for Computer
Science, Cambridge, Massachusetts, 1981.
December 8, 2002
560 BIBLIOGRAPHY
[To88] V.V. Topentcharov. Structures algebriques n-aires
generalisant les categories. Alg. Univ., 25:147173, 1988.
[ToAn82] V.V. Topentcharov and Y.N. Arnaudov. Two algebraic va-
rieties related with Cat. In: N-ary structures, pages 2934,
(Proc. Symp. Skopje), 1982.
[Tk83] V.S. Trokhimenko. Connectivity relation on ordered Menger
algebras. Ukr Math. J., 35:388393, 1983. translation from:
Ukr. Mat. Zh. 35, No 4, pages 461466, 1983.
[Tk84] V.S. Trokhimenko. Theory of restrictive Menger algebras.
Ukr. Math. J., 36:7680, 1984. translation from: Ukr. Mat.
Zh. 36, No 1, 1984, pages 8287.
[UStj78] J. Usan and Z. Stojakovic. Orthogonal systems of partial
operations. Zb. Rad. Prir.-Mat. Fak., Univ. Novom Sadu,
8:4751, 1978.
[Va65] V.V. Vagner. The theory of relations and the algebra of
partial mappings. Theory of Semigroups and Appl. I, Izdat.
Saratov Univ, pages 3178, 1965. Russian.
[Va79a] V.V. Vagner. On the theory of generalized rings, modules
and linear algebras. Sov. Math., 23(3):716, 1979. trans-
lation from: Izv. Vyssh. Uchebn. Zaved., Mat, 1979, No 3
(202), pages 1227.
[Va79] V.V. Vagner. Restrictive sums of external operatives and
their natural extension. Sov. Math., 23(2):414, 1979. trans-
lation from: Izv. Vyssh. Uchebn. Zaved., Mat, 1979, No 2
(201), pages 724.
[Va79b] V.V. Vagner. The algebra of binary relations and its applica-
tions to dierential geometry. Di. Geom., 4:15131, 1979.
Russian.
[VhPd85] I.I. Valutseh and N.I. Prodan. On the extension of partial
algebras to total algebras of a quasi-variety. Mat. Issled.,
83:3943, 1985. Russian.
December 8, 2002
BIBLIOGRAPHY 561
[vdW66] B.L. Van der Waerden. Algebra, Erster Teil. Springer-
Verlag, 1966.
[Vs79] S. Vasilach. Direct limits in the categories of non-nitary
heterogeneous algebras. Rend. Circ. Mat. Palermo II, Ser. 28,
pages 337350, 1979.
[Vd86] V.V. Vdovin. On homomorphisms of projective planes I.
Sib. Mat. Zh. 27 No. 1 (155), pages 3544, 1986. Russian.
[Ve82a] Y.P. Velinov. Combinatorial arrows in a polycategory. In:
N-ary structures, pages 5763, (Proc. Symp. Skopje), 1982.
[Ve82] Y.P. Velinov. Polycategories and recursiveness. In: N-ary
structures, pages 4756, (Proc. Symp. Skopje), 1982.
[V79] H.J. Vogel. On Birkho-algebras in dht-symmetric cate-
gories. In: Finite Algebras and Multiple-Valued Logic, vol-
ume 28 of Coll. Math. Soc. J.Bolyai, pages 759779, (Proc.
Szeged, Hungary 1979), 1981. North- Holland Publ. Co.
[V82] H.J. Vogel. Eine Beschreibung von Verkn upfungen f ur par-
tielle Funktionen. Rostocker Math. Kolloq., 20:2132, 1982.
[V84] H.J. Vogel. Eine kategorientheoretische Sprache zur
Beschreibung von Birkho-Algebren. Report R-Math-06/84,
Akad. d. Wiss. d. DDR, 1984.
[We70a] G.H. Wenzel. Extensions of congruence relations on inni-
tary partial algebras, A problem of G.Gr atzer. Fund. Math.,
67:163169, 1970.
[We70] G.H. Wenzel. Subdirect irreducibility and equational com-
pactness in unary algebras A; f). Arch. Math., 21:256264,
1970.
[W78] H. Werner. Einf uhrung in die allgemeine Algebra. B.I.
Hochschultaschenb ucher Band 120, 1978.
[Wh69] T.P. Whaley. Endomorphisms of partial algebras. Notices
AMS, 16:817818, 1969. Abstract.
December 8, 2002
562 BIBLIOGRAPHY
[Wl69] R. Wille. Subdirekte Produkte und konjunkte Summen. Jour-
nal f ur Math., 239/240:333338, 1969.
[Wl82] R. Wille. Restructuring Lattice Theory: An Approach Based
on Hierarchies of Concepts. Ordered Sets, pages 445470,
1982. editor: I.Rival; D.Reidel Publ. Co., Dordrecht.
[Wl84] R. Wille. Line diagrams of hierarchical concept systems.
Internat. Classication, 11:7786, 1984.
[Wk82] J. Winkowski. An algebraic description of system be-
haviours. Pr. Inst. Podstaw Inf. Pol. Akad. Nauk, 471:315
340, 1982. also in: Theor. Comp. Sci. 21.
[Ws72] M. Wischnewski. Partielle Algebren in Initialkategorien.
Math. Z., 127:8391, 1972.
[Wc88] D. Wojciechowska. Hofts equality in unary partial algebras.
Masterthesis, Warsaw, 1988. Polish.
[Wo70] B. Wojdylo. On some problems of J.Slomi nski concerning
equations in quasialgebras. Coll. Math., 21:14, 1970.
[Wo72] B. Wojdylo. Categories of quasi-algebras. N.Copernicus Uni-
versity, Torun Preprint No. 2, 1972.
[Wo73] B. Wojdylo. Remarks on lattices of congruences relations of
quasi-algebras. Coll. Math., 27:187191, 1973.
[Wo75] B. Wojdylo. On equationally denable classes of quasi-
algebras. Collection of papers on the Theory of Ordered
Sets and General Algebra, Acta Fac. Rerum Natur. Univ.
Comenian Math., Special No, 1975, pages 5557, 1975.
[Wo77] B. Wojdylo. Many sorted algebras and their application
in computer science. Instytut Matem. UMK, Preprint 15,
Torun, 1977.
[Wo79] B. Wojdy lo. Programming language from algebraic point of
view. In: Contributions to General Algebra, pages 405421,
(Proc. Conf. Klagenfurt, Mai 1978), 1979. Verlag Joh.Heyn
Klagenfurt.
December 8, 2002
BIBLIOGRAPHY 563
[Za68] A. Zajtz. Algebraic objects. Zeszty nauk. Uniw. Jagiellon-
ski 167, Prace mat., 12:6779, 1968.
[Ze79] I.

Zembery. A note on categories of partial algebras. Math.
Slovaca, 29:157158, 1979.
[Zl92] P. Zlatos. On surjective kernels of partial algebras. Acta
Math. Univ. Comenian. (N.S.), 61:8590, 1992.
December 8, 2002

Você também pode gostar