Galois Connections and Applications

Galois Connections and Applications
Mathematics and Its Applications
Managi ng Editor:
M. HAZEWINKEL
Centre for Mathematics and Computer Science, Amsterdam, The Netherlands
Volume 565
Galois Connections and
Applications
Edited by
K. Denecke
University of Potsdam,
Potsdam, Germany
M. Erne
University of Hannover,
Hannover, Germany
and
S. L. Wismath
University of Lethbridge,
Lethbridge, Canada
Springer Science+Business Media, B.V.

A C. J.P. Catalogue record for this book is available from the Library of Congress.
ISBN 978-90-481-6540-7 ISBN 978-1-4020-1898-5 (eBook)

DOI 10.1007/978-1-4020-1898-5
Printed on acid-free paper
All Rights Reserved

© 2004 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 2004.
Softcover reprint of the hardcover 1st edition 2004
No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microtilming, recording
or otherwise, without written permission from the Publisher, with the exception
of any material supplied specitically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work
Contents
Preface . Vll
M. Erne
Adjunctions and Galois Connections: Origins, History and Devel-
opment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
G. Janelidze
Categorical Galois Theory: Revision and Some Recent Develop-
ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
M. Erne
The Polarity between Approximation and Distribution . . . . . 173
K. Denecke, S. l. Wismath
Galois Connections and Complete Sublattices . . . . . . . . . . 211
R. Poschel
Galois Connections for Operations and Relations . . . . . . . . 231
K. Kaarli
Galois Connections and Polynomial Completeness . . . . . . . . 259
K. Gtazek, St. Niwczyk

Q-Independence and Weak Automorphisms . . . . . . . . . . . 277
A. Szendrei
A Survey of Clones Closed Under Conjugation . . . . . . . . . . 297
P. Burmeister
Galois Connections for Partial Algebras . . . . . . . . . . . . . 345
K. Denecke, S. l. Wismath
Complexity of Terms and the Galois Connection Id-Mod . . . . 371
Vl Contents
J. lambek
Iterated Galois Connections in Arithmetic and Linguistics . . . 389
I. Chajda, R. Hala~
Deductive Systems and Galois Connections . . . . . . . . . . . 399
J. Slapal
A Galois Correspondence for Digital Topology . . . . . . . . . . 413
w. Gahler
Galois Connections in Category Theory, Topology and Logic . . 425
R. Wille
Dyadic Mathematics - Abstractions from Logical Thought . . . 453
Index . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. . 499
Preface
Galois connections provide the order- or structure-preserving passage

between two worlds of our imagination - and thus are inherent in hu-
man thinking wherever logical or mathematical reasoning about cer-
tain hierarchical structures is involved. Order-theoretically, a Galois
connection is given simply by two opposite order-inverting (or order-
preserving) maps whose composition yields two closure operations (or
one closure and one kernel operation in the order-preserving case).
Thus, the "hierarchies" in the two opposite worlds are reversed or
transported when passing to the other world, and going forth and back
becomes a stationary process when iterated. The advantage of such an
"adjoint situation" is that information about objects and relationships
in one of the two worlds may be used to gain new information about
the other world, and vice versa. In classical Galois theory, for instance,
properties of permutation groups are used to study field extensions.
Or, in algebraic geometry, a good knowledge of polynomial rings gives
insight into the structure of curves, surfaces and other algebraic vari-
eties, and conversely. Moreover, restriction to the "Galois-closed" or
"Galois-open" objects (the fixed points of the composite maps) leads
to a precise "duality between two maximal subworlds".
Though the name "Galois connections" refers to the classical example
of Galois theory, where subfields of a field and subgroups of the au-
tomorphism group are related by a dual lattice isomorphism, a richer
supply of examples comes from certain relations of annihilation or
orthogonality, well-known from various branches of algebra, geome-
try and topology- and in fact, this was the starting point for Garrett
Birkhoff when he invented in the thirties of the past century his general
"polarities" as Galois connections between power sets, arising from re-
lations between the ground sets. The more general order-theoretical
type of Galois connections (alias Galois connexions or Galois corre-
spondences) as defined above is due to Oystein Ore (about 1940) .
Vlll Preface
One or two decades later, the even more general notion of adjoint
functors was discovered by workers in category theory, and various
very useful types of categorical Galois correspondences were subse-
quently invented.
Concerning the position of Galois connections within mathematics, it
has to be admitted that deep theorems rarely are immediate conse-
quences of the general theory of Galois connections, but usually require
some extra tools and ideas stemming from the specific theory under
consideration. But the general framework, supporting the intuition
and suggesting the appropriate concepts, is established by the discov-
ery of the "right" underlying Galois connection, followed by a good
characterization of the "Galois-closed" (or "Galois-open") elements or
sets. And there is no doubt that many proofs become shorter, more
elegant and more transparent in the language of Galois connections.
The aim of this book is to present not only the main ideas of gen-
eral Galois theory, of concepts related to Galois connections, and of
their appearance in various classical and modern areas of mathemat-
ics, but also to show how they can be applied in other disciplines. The
book consist of 15 contributions, written by authors who are special-
ists in diverse fields of mathematics and who use Galois connections
in their work. Although the majority of the contributions are dealing
with themes of classical and universal algebra, the reader will also find
many logical, order-theoretical, topological and categorical aspects -
and, last but not least , several practical applications, sometimes even
in extra-mathematical fields like linguistics, digital representation of
graphics, or data analysis. In fact, the methods of modern formal
concept analysis, based on polarities arising from certain relations be-
tween objects and attributes, have applications in quite diverse non-
mathematical areas of sciences and arts. Readers interested in that
theory, its mathematical background and its applications, are referred
to the monograph Formal Concept Analysis - Mathematical Founda-
tions by B. Ganter and R. Wille (Springer-Verlag, 1999).
In the first contribution, which is also aimed at "novices" in the the-

ory of Galois connections, M. Erne presents the basic ideas and con-
cepts: order-theoretical Galois connections in their original contravari-
ant form with anti tone (order-inverting) maps and in their covariant
form with isotone (order-preserving) maps, also referred to as adjunc-
Preface IX
tions or "mixed" Galois connections. The classical examples come

from polarities and axialities, i.e. Galois connections and adjunctions,
respectively, between power sets. A journey through three centuries
of mathematics illustrates the historical development of the main con-
cepts in that area. After a discussion of diverse examples in classical
Galois theory and other parts of algebra, the pioneer achievements
of the twentieth century are sketched, in particular, Birkhoff's idea
of polarities and Ore's Galois connexions. The journey ends with a
survey of residuation theory, where binary operations having adjoint
translations are the central tool. More recent developments, in par-
ticular, the categorical types of Galois correspondences and the even
more general adjoint functors had to be omitted in this historical in-
troduction, due to space limitations. Readers interested in these topics
but not familiar with categorical thinking may consult, for example,
the book by Adamek , Herrlich and Strecker: Abstract and Concrete
Categories (J. Wiley & Sons, 1990).
In Dyadic Mathematics - Abstractions from Logical Thought, R. Wille
sets out the philosophical, logical and mathematical ideas furnishing
the foundations of formal concept analysis, and hence of concrete ap-
plications of Galois connections. He demonstrates how human logical
reasoning is based on concepts and how these may be formalized math-
ematically by forming contexts and their concept lattices. As the title
indicates, the emphasis is on dyadic conceptions and results in order
and lattice theory, contextual logic, algebra, and geometry. This sur-
vey article is of interest not only for people looking for applications
of Galois connections, but also for those who want to understand the
logical and philosophical background.
Three other contributions are devoted to some order- and lattice-

theoretical aspects of Galois connections and their applications in
other fields.
In "Galois Connections and Complete Sublattices " by K. Denecke and
S. L. Wismath, three different methods are described to characterize
complete sublattices of a complete lattice, by using a conjugate pair
of additive closure operators, by special closure or kernel operators
defined on the given complete lattice, and by so-called Galois-closed
subrelations. Basic tools here are contexts and concept lattices, known
from formal concept analysis.
X Preface
The frequently observed phenomenon that approximation properties

of certain relations on ordered sets are closely related or even equiva-
lent to certain (infinite) distributive laws for the given ordered struc-
tures or some of their completions, is studied in "The Polarity between
Approximation and Distribution" by M. Erne. It turns out that the
right framework for those investigations is a certain polarity coming
from an elementary relation defined in terms of a fixed closure opera-
tion on a frame (alias locale, a pointfree abstraction of topology). Via
that polarity, the Galois-closed approximating subsets are in one-to-
one correspondence with the sublocales containing all closed elements,
and the Galois-closed strongly approximating subsets correspond to
the completely distributive sublocales. These results encompass many
specific representation theorems in order theory and related fields such
as domain theory and pointfree topology.
Adjoint pairs of maps on the integers or natural numbers are studied
by J. Lambek in "Iterated Galois Connections in Arithmetic and Lin-
guistics". He points out that such adjoint pairs simply correspond to
infinite subsets with an infinite complement, and that for maps that
are unbounded on both sides, all iterated left and right adjoints ex-
ist. In the second part, some categorical generalizations lead to the
notion of pregroups, which have interesting applications in a modern
treatment of linguistic problems. In particular, they provide a new
algebraic approach to the structure of sentences in natural languages.
Often it is necessary to define Galois connections for proper classes

rather than sets. In order to circumvent set-theoretical difficulties in
that extended setting, one usually assumes the existence of a "com-
mon universe" . One of the most important Galois connections of the
class-theoretical type comes from the relation of satisfaction between
algebras and equations, or more generally, between partial algebras
and formulas. (The dual relation is the relation of validity.) This ap-
proach provides fundamental connections between algebraic structures
and the equations (or formulas) they satisfy. The resulting Birkhoff-
Tarski isomorphism theorems then establish dualities between varieties
and equational theories, or between more general classes like quasivari-
eties and implicational theories, etc. Four papers here deal with these
and some other polarities arising in universal algebra.
Preface XI
The survey paper "Galois Connections for Partial Algebras" by P.

Burmeister deals with such polarities in the rather general context of
many-sorted partial algebras, which are of still increasing importance
not only in universal algebra but also in computer science. Since in
that general realm there are several kinds of equations, one also has
various corresponding relations of validity. Another polarity discussed
in that paper comes from the relation between maps and formulas
of a certain signature, saying that the map reflects the formula (or
equivalently, preserves the negated formula) . Here, the resulting po-
larity yields several interesting factorization systems for morphisms,
known from category theory and universal algebra. In the last part of
Burmeister's paper, the power of attribute exploration via formal con-
cept analysis is demonstrated for the discovery of interdependencies
between properties of homomorphisms.
A standard algebraic approach to fragments of implicational or intu-

itionistic logic are Hilbert algebras and their deductive systems (con-
taining the constant "true" and closed under modus ponens) . In De-
ductive Systems and Galois Connections, I. Chajda and R. Halas in-
vestigate a general algebra A with a constant 0 and its so-called b-
deductive systems, where b is a binary term function of the algebra.
These are closed under certain inference rules and form a closure sys-
tem that is isomorphic, via a suitable axiality, to the system of all con-
gruences 8 satisfying a8c iff b(a, c)80 and b(c, a)80 (b-regular con-
gruences for short). The basic tools for applications to congruence the-
ory are so-called Fichtner terms of a variety, i.e. binary terms such that
every congruence of any algebra in the variety is b-regular (it suffices
to ensure that property for the identity relations). The b-deductive
systems of any algebra in a variety having a Fichtner term turn out to
be precisely the congruence kernels. In the second part, the polarity
induced by the relation R = {(b, D): Dis a b-deductive system of A}
between binary terms and subsets of an algebra is investigated, and
its Galois-closed sets are characterized for certain cases.
In " Complexity of Terms and the Galois Connection Id - Mod", K.

Denecke and S. L. Wismath study the influence of term complexity
on the Galois connection between identities and algebras. Calling a
variety k-normal when both sides of any of its nontrivial equations
have complexity at least k, they extend several central results on nor-
Xll Preface
mal varieties to the k-th level and illustrate the limitations of such
extensions.
In the survey 11 Q-Independence and Weak Automorphisms as Galois
Connections", K. Glazek and S. Niwczyk present two further kinds of
Galois connections appearing in universal algebra. These are related
to general notions of independence and of weak automorphisms. In
the first case, the basic relation for the required polarity is defined
between subsets and partial maps of a fixed algebra: a subset X is
related to a partial map p if the latter extends to a homomorphism
on the subalgebra generated by X whenever X is the domain of p. In
the second case, the basic relation is defined between permutations of
a fixed set and clones over that set, and it holds iff the permutation
preserves the clone (in a natural way).
Another fundamental type of Galois connection for universal algebra

is a central theme in several papers. This is the Galois connection
(Pol, I nv) arising from the invariance relation "!preserves r/' between
operations of an algebra A and n-ary relations on A. In that case, one
says that f is a polymorphism of (!, or that (! is invariant under f.
In Galois Connections for Operations and Relations, R. Poschel gives
a systematic overview of the manifold Galois connections obtained by
restriction or extension of the connection (Pol , I nv) to specified or
generalized ranges, respectively. The main theorems yield intrinsic
characterizations of the Galois-closed sets and the corresponding clo-
sure operators on both sides. The general (not necessarily finite) case
requires certain local closure operators providing interpolation, while
on the relational side, one has to consider so-called relational clones,
for example, various kinds of Krasner algebras or clones, corresponding
(via Galois connection) to classes of endomorphisms and (weak) auto-
morphisms (different from those considered by Gtazek and Niwczyk).
Moreover, the (Pol, Inv) approach provides nice solutions of simulta-
neous concrete characterization problems for subalgebra lattices, con-
gruence lattices and automorphism groups. On the other hand , there
are useful extensions of (Pol , Inv) to partial functions , multifunctions
and heterogeneous functions (cf. the article by P. Burmeister).
In Galois Connections and Polynomial Completeness, K. Kaarli ex-
plores different generalizations and variations of primality and polyno-
Preface Xlll
mial completeness (meaning that all functions, or at least all relation-

preserving functions, are polynomials). His approach makes system-
atic use of the Galois connection (Pol , Inv ) and demonstrates how
and why local interpolation plays such a crucial role in the infinite
case. Thus, for example, a local term function is one that can be in-
terpolated by a term function on every finite subset, and an algebra is
called locally primal if every function (in several variables) is a local
term function. Variations of that concept are (local) congruence pri-
mality, (local) endo- or automorphism primality etc. (cf. t he survey
article by R. Poschel). Special attention is paid to varieties having a
majority or near-unanimity term, to order affine complete lattices, to
algebras with distributive lattice reduct, and to endoprimal algebras.
A. Szendrei in her contribution A Survey of Clones Closed Under Con-
jugation studies important complete sublattices of the lattice of all
clones of operations on a finite set, using a Galois connection that can
be derived from (Pol, Inv).
The last three papers are devoted to themes where categorical methods
help to generalize and to improve classical results of topology and
Galois theory.
W. Gahler shows in his survey Galois Connections in Category The-
ory, Topology and Logic how partially ordered monads (a categorical
notion very close to adjoint situations) provide the right framework
for a very general development of abstract categorical topology and
convergence theory. The basic idea here is to suitably generalize the
"classical" filter monad, which has proved basic for many questions
of convergence theory in the past. But, as recent developments have
shown, there is also a large overlap of such topological theories with
non-classical logics and computer sciences (see, for example, t he book
by St. Vickers: Topology via Logic (Cambridge Tracts in Th. Comp.
Sc. 5, Cambridge University Press, 1989). In the second part of this
paper, various types of fuzzy filters in locales and quantales are stud-
ied more thoroughly, and their interdependencies are illustrated by
several diagrams.
J. Slapal's note A Galois Correspondence for Digital Topology de-
scribes an adjoint situation between the category of (generalized) clo-
sure spaces in the sense of Cech (where only extensivity and preser-
XIV Preface
vation of inclusion and of emptiness is required) and the category of

relational systems of a given (possibly infinite) arity. This extends the
well-known binary case, leading to a concrete functorial isomorphism
between partially ordered sets and Alexandroff (discrete) spaces. How-
ever, the general case is much more involved. For the objects of the
resulting corefiective subcategory of closure spaces, connectedness co-
incides with a certain kind of "discrete path-connectedness". There-
fore , the study of such objects is helpful in digital topology, where
these notions of connectedness play a crucial role.
In the final paper, G. Janelidze gives a survey of Recent Developments
in Categorical Galois Theory. This rather advanced paper requires
some familiarity with categorical methods and concepts. For read-
ers who are not experts in that area, some preparatory literature is
recommended, for example the book Galois Th eories by F. Borceux
and G. Janelidze (Cambridge Studies in Adv. Math. 72, Cambridge
University Press, 2001), where the development from classical Galois
theory to the modern categorical versions is described more compre-
hensively. In the present article, one subsection deals with categori-
cal versions of A. R. Magid's Galois theory for rings, and another is
concerned with central extensions of universal algebras. But, a bit
surprisingly, the theory begins with some purely topological consider-
ations on categorical characterizations of connectedness and compo-
nents. A Galois structure is then defined as a triple consisting of an
adjoint pair of functors and two subclasses of objects in the involved
categories, called fibrations, having suitable closure properties with re-
spect to certain categorical constructions. Many examples are given,
but most of them also require some categorical background. The fun-
damental theorem of Galois theory (in a categorical form) is based
on Grothendieck's formulation saying that the opposite category of
subextensions of the field extension E I K is equivalent to the category
of transitive Gal (E I K)-sets. The chosen approach confirms that cat-
egorical language provides many improvements and simplifications of
classical results: besides various related phenomena, it explains why
the Galois theory of fields and the theory of covering spaces are just
identical, if interpreted the right way.
The plan for a book on Galois connections originated with the partici-
pants of a conference on that subject held at the University of Potsdam
Preface XV
in 2001. The editors hope that this volume, though certainly not com-
prehensive enough to cover all the important themes in this area, will
help to make an easy but beautiful concept more popular. Thus, if
the value of Galois connections as a unifying concept in quite diverse
branches of mathematics, but also its usefulness as a practical tool
both in mathematics and other areas, become more evident by the
contributions in this book, our main intention will have been fulfilled.
***
Klaus Denecke Marcel Erne Shelly L Wismath

Potsdam Hannover Lethbridge
October 2003
Adjunctions and Galois Connections:
Origins, History and Development
M. Erne
Abstract
Galois connections build the abstract framework not only for classical and
modern Galois theory involving groups, fields and rings, but also for many
other algebraic, topological, order-theoretical, categorical and logical theo-
nes.
We sketch the development of Galois connections, both in their covariant
form (adjunct ions) and in the contravariant form (polarities) through the
last three centuries and illustrate their importance by many examples. The
main steps in the development are:
the theory of polynomial equations (Lagrange, Galois),
the modern Galois theory (Dedekind, Artin),
the origins of lattice theory (Dedekind, Schroder),
the polarities and lattice-theoretical aspects (Birkhoff),
the order-theoretical Galois connections (Ore),
the logical calculus (Boole, Peirce, Schroder),
and the residuation theory (Krull, Ward, Dilworth).
Besides sporadic occurrences of adjunctions and Galois connections in im-
portant mathematical theorems, we discuss diverse contributions to a sys-
tematical theory of adjunction and residuation, and we touch on various
applications to topology, logic, universal algebra and formal concept anal-
ysis.
Mathematics Subject Classification: 01-02, 06A15, 06D15, 12F10.
Key words: Adjoint map, Axiality, Closure operator, Complete lattice, Ga-
lois connection, Polynomial equation, Field, Group, (Partially} Ordered set,
Polarity.
K. Denecke et al. (eds. ), Galois Connections and Applications, 1-138.

© 2004 Kluwer Academic Publishers.
2 M. Erne
Contents
1 The Idea of Adjunction and Galois Connections
1.1 Adjunctions between Ordered Sets

1.2 Closure and Kernel Operations
1.3 Galois Connections
1.4 Polarities and Axialities
1.5 Formal Concept Analysis
2 The Roots of Galois Connections
2.1 Prelude on Symmetric Functions

2.2 Joseph-Louis Lagrange: Permutation Groups, Partially
Symmetric Functions and Interpolation
2.3 Evariste Galois' Memoir on the Solvability of Equations
2.4 Richard Dedekind's Lectures on Algebra
2.5 Dedekind's Order- and Lattice-Theoretical Concepts
2.6 Ernst Schroder's Algorithms and Calculi
2. 7 David Hilbert's Nullstellensatz
3 Galois Connections in the Twentieth Century
3.1 From Dedekind to Noether, Artin and van der Waerden:

The Modern Galois Theory
3.2 Garrett Birkhoff's Polarities
3.3 Oystein Ore's Galois Connexions
3.4 Adjunctions as Galois Connections of Mixed Type
4 Residuation Theory
4.1 From Boole to Schroder: The Logical Calculus

4.2 Dedekind's Invention of Lattice-Ordered Groups
4.3 Wolfgang Krull's Ideal Domains and Quantales
4.4 Morgan Ward's and R. P. Dilworth's Residuated Lattices
4.5 T. S. Blyth and M. F. Janowitz: Residuation Theory
4.6 Heyting Algebras, Locales and Pointfree Topology
Adjunctions and Galois Connections 3
1 The Idea of Adjunctions and Galois Connections

Mais je n'ai pas le temps, et mes idees ne sont pas encore
bien developpees sur ce terrain, qui est immense.
Evariste Galois, in the letter to his
friend Auguste Chevalier, written in
1832, on the night before the duel
A Galois Adjunction
A Galois Connection
(Duel Adjunction)
4 M . Erne
Galois connections are the main theme of this book. So, readers not
familiar with that topic might wish to know: where do they come
from, where are they used , and who were the prominent pioneers of the
theory? I hope I will be able to answer at least in part these questions.
Certainly it cannot be the intention of a historical introduction to
cover all relevant instances of Galois connections, and one or the other
expert will miss his own favorite theme. What I have tried to do is to
collect together a series of situations where Galois connections enter
(or may be introduced into) the picture of an interesting classical
theory or problem. A comprehensive historical discussion of the roots
of order and lattice theory can be found in the book by Mehrtens
[92]. For historically oriented treatments of classical Galois theory,
two comprehensive sources are Edwards [39] and Tignol [138].
In the 19th century, the concept of Galois theory was not yet born,
though we shall have occasion to encounter it implicitly in the work of
the prominent pioneers of Galois theory: Lagrange, Abel , Galois and
Dedekind.
In the 20th century, milestones on the way to the modern view of Ga-
lois connections are Birkhoff's "polarities", Ore's "Galois connexions",
and Schmidt's "Galois correspondences of mixed type", nowadays bet-
ter known as "adjunctions" . Particular mention should be made of the
residuation theory initiated by Krull, Dilworth and Ward , continued
by Blyth and Janowitz, and leading to the modern theory of "quan-
tales". We also briefly touch on the practice-oriented interpretation
of relations and polarities as "contexts" in the sense of formal concept
analysis (see R. Wille's contribution "Dyadic Mathematics - Abstrac-
tions from Logical Thought" in this volume). However, reasons of
limited space forbid us to have a closer look at the immense field of
categorical Galois connections and functorial adjunctions.
A number of examples from order theory, universal algebra and topol-
ogy, showing how the tool of adjoint maps may facilitate or clarify con-
siderably mathematical problems, will accompany our journey through
the world of Galois connections. The examples chosen are sometimes
rather elementary, but nevertheless some of them have caused break-
throughs not only in the development of mathematical theories, but
also in their applications.
1.1 Adjunctions Between Ordered Sets

In their most general sense, adjunctions and/or Galois connections
make it possible to relate two "worlds" of (more or less mathemati-
cal) objects with each other in order to gain information about one
world by passing to the other, perhaps better known world (cf. [15]).
Slightly more precisely, the passage should be made so that certain
order structures are preserved or inverted. In a nutshell, two typical
"microcosmic worlds" could be the left and the right hand side of an
equality or inequality. The usual process of simplifying or evaluating
such an (in)equality is to shift some "junk" from one side to the other
and then to transform the lightened side suitably; then, if necessary,
to shift parts back from the other side, and so forth.
An old example for such a procedure is the following problem from
Abu Kamil's Algebra (about 880 A.D.):
Under what conditions is it possible to solve the equation
-x 2 + 2bx + c = D ?
(Here D stands for a square). The answer given by Abu Kamil is:
If c < 0 then - c must be less than b2 , and b2 + c must be a sum of
two squares. It is unlikely that Abu Kamil already was aware of the
following subtle difference: while the second condition is necessary
and sufficient for the existence of a solutions within the rationals, the
first one is necessary and sufficient for the existence of a solution in the
reals. The argument here is a typical case of shifting parts between the
two sides of an inequality, adopting the fact that in the reals, squares
are just positive numbers:
-x 2 + 2bx + c > 0 ¢:? c > x2 - 2bx ¢:? c + b2 > x 2 - 2bx + b2 ¢:?
c+b2 >(x-b) 2 :::::} c +b2 > 0 ¢:? -c<b2 .

Notice that the implication arrow may be inverted if x is close enough
to b.
Of course, the treatment of similar inequalities, requiring the possi-
bility of adding or subtracting values on both sides, or of multiplying
them simultaneously by positive numbers etc. , relies on the properties
of an ordered (semi)group or field.
It is not difficult to make precise what is generally needed in order to
shift parts of an (in)equality from one side to the other: the adequate
6 M. Erne
description of such situations is by means of adjoint pairs of maps
P~Q
T
between (quasi-)ordered sets, subject to the defining condition
U=r) P::; Q(q) {::} A(p) ::; q,
which turns out to be equivalent to the requirement that A and 12 be
isotone (i.e. order-preserving) maps satisfying the inequalities
p :S Q(A(p)) and A(g(q)) :Sq.
As usual, we mean by a quasi-ordered set one with a reflexive and
transitive relation, called the order relation. If the relations involved
in an adjunction are partial orders (i.e. antisymmetric), then the left
or lower adjoint A is uniquely determined by its right or upper adjoint
Q, and conversely. Notice that rotating the above little diagram by
180° leads to an adjoint pair (g, A) between the dual posets Qd and
pd_ In the introductory example, A could stand for the addition of
a constant, and Q for the subtraction of the same constant. More
generally, in any (additively written) partially ordered group G with
isotone translations
Ta : X f--+ X +a
the "subtractions"
aa : y f--+ y- a
are both left and right adjoint to Ta , being inverse to Ta· Generally, a
poset with a binary operation * is said to be residuated iff each of the
translations x f--+ x *a and x f--+ a* x has a right adjoint (see Chapter
4).
Notice that one always may take, as a special instance of a partial
order, the equality relation =. In that specific case, an adjoint pair of
maps is just a pair of mutually inverse bijections, and the usual treat-
ment of equations by shifting parts of one side to the other appears
as a special case of the general handling of inequalities.
A similar, perhaps somewhat unexpected type of adjunction arises
when the first quasi-order is an equivalence relation and the second is
the identity relation. For example, the "world of congruences modulo
a prime number p" is translated into the "world of the Galois field
Fp = 'lljp'll" (consisting of the residue classes modulo p) , by means of
the quotient map
A : 'll---t Fp, n f---t n + p'll .
Any left inverse (! : Fp ---t 7l of A, picking one representative from each
residue class, is both left and right adjoint to A, since
A(n) = q -R n Q(q) mod p.
Thus, all statements, problems and solutions in Gauss' theory of con-
gruences modulo a prime number [61] may be translated into the the-
ory of Galois fields Fp , and vice versa.
An obvious question in the present context is this: which maps be-
tween ordered sets do have a left or a right adjoint? Although the
answer is very simple, it seems not to have been stated explicitly be-
fore the middle of the 20th century (see e.g. Pickert [114], Schmidt
[123]). Calling sets of the form
(p] = {x E p: X~ p}
principal (order) ideals and sets of the form
[p) = {X E p : p ~ X}
principal dual ideals or principal filters, one has the following
Characterization Theorem for Adjoint Maps
(1) A map A : P ---t Q between quasi-ordered sets has a right adjoint
if and only if it is residuated, i.e. inverse images of principal ideals
under A are again principal ideals. If P is a poset, the right adjoint
(! : Q ---t P is then given by
Q(q) = max{p E P: A(p) ~ q}.

(2) Dually, a map (! : Q ---t P has a left adjoint if and only if it is
residual, i.e. inverse images of principal dual ideals under (! are again
principal dual ideals. If Q is a poset, the left adjoint A : P ---t Q is
given by
A(p) = min{q E Q: p ~ Q(q)}.
(3) Any residuated map between posets preserves all existing joins, and
any residual map all existing meets.
8 M. Erne
(4) Conversely, any join (meet) preserving map between complete lat-
tices is residuated (residual).
Indeed, if a map A : P---+ Q between complete lattices preserves joins
then
(! : Q ---+ P, q H V{p E P : A(p) ::; q}
is the right adjoint of A; and dually, if(! : Q---+ P preserves meets then
A: P---+ Q, pH (\{q E Q: p::; Q(q)}

is the left adjoint of(!.
1.2 Closure and Kernel Operations

Adjunctions are intimately related to the notion of closure and ker-
nel operations. Set-theoretical closure and kernel operators are well-
known from algebra, topology, geometry and many other mathemati-
cal disciplines. The easiest order-theoretical abstraction is the follow-
ing: a closure operation of a partially ordered (or quasi-ordered) set
P is a map ry : P ---+ P such t hat
x::; ry(y) {::} ry(x) ::; ry(y).
In other words, closure operations are in one-to-one correspondence
(via surjective corestriction) to left adjoints of inclusion maps between
ordered sets. K ernel operations are defined dually; hence, their surjec-
tive corestrictions are just the right adjoints of inclusion maps. Calling
a subset C of a poset P a closure range if for each x E P there is a least
x E C with x ::; x, one immediately verifies that 'Yc : P ---+ P , x H x
is a closure operation, and that the assignment C H 1c yields a dual
isomorphism between the poset of all closure ranges in P (ordered by
inclusion) and the poset of all closure operations of P (ordered point-
wise). The closure ranges in a complete lattice are just the meet-closed
subsets. Analogous statements hold for kernel operations and kernel
ranges.
On the other hand, an arbitrary adjunction (A, (!) always gives rise
to a closure operation (!OA = A(! and a kernel operation A o (! = (!A ,
because of the inequalit ies p ::; Q(A(p)) , A(Q(q)) ::; q and the resulting
equations
)q?A = -\, g-\ g = g.

(In order to be flexible in due course, we use two notations, ,\ g and
go,\, for the same composition of maps, given by
pA(! = (-\g)(p) = go-\(p) = g(-\(p))
and meaning that one applies first,\ and then g.)
Denoting the ranges of,\ and g by P,\ and Qg, respectively, one has the
following fundamental result, in the covariant form with isotone maps
probably formulated first by J. Schmidt [123], while the contravariant
form with antitone maps is already to be found in Ore's pioneering
paper [38]:
Adjunctions and Closure-Kernel Isomorphisms

An arbitrary adjunction (-\ , R) between posets P and Q induces mu-
tually inverse isomorphisms ,\ : Qg----* P,\ and 1! : P-\----* Qg . Further-
more, the fixpoint set of the closure operation 'Y = go,\ is the closure
range Qg, and the fixpoint set of the kernel operation K, = ,\o g is the
kernel range P-\. The original adjunction may be reconstructed from
these data by -\(p) = ~o'Y(P) and g(q) = (!oK,(q). Under this construc-
tion, the adjoint pairs of maps between P and Q are in one-to-one
correspondence with isomorphisms between closure ranges of P and
kernel ranges of Q.
(2
p Q
,\
'Y =go,\ c K,=Aog

(2
Qe
,\
Adjoint pairs compose in the obvious manner: if(-\ , g) is an adjunction

between P and Q, and (O", T) is an adjunction between Q and R, then
( ,\ O", T g) = ( O" o ,\, goT) is one between P and R. The essence of the
previously established correspondence may be formulated as a

Homomorphism Theorem for Adjunctions
Every adjunction between posets has a decomposition, unique up to
10 M. Erne
isomorphism, into a "closure adjunction" (constituted by a closure op-

eration and an inclusion map), a "perfect adjunction" (constituted by
two isomorphisms), and a "kernel adjunction" (constituted by an in-
clusion map and a kernel operation). In particular, every adjunction
factorizes into an "epi-adjunction" (with surjective left adjoint) and a
"mono-adjunction" (with injective left adjoint).
1.3 Galois Connections

What are Galois connections? Many authors use that term as a syn-
onym for what we called adjunctions. But we prefer the convention
to distinguish between the "covariant" case of adjunctions and the
"contravariant" case of Galois connections. Thus, inverting the order
in the second ordered set Q, one arrives at the classical definition of
Galois connections (alias Galois correspondences), where the defining
equivalence becomes "symmetric" :
p :S Q(q) {::} q :S A(p).
A well-known equivalent definition of Galois connections, due to Ore
(see [38] and Section 3.3) requires that A and Q be antitone (i.e. order-
inverting) maps satisfying
p :S Q(A(p)) and q :S A(Q(q)).
By definition, both compositions of the partners of a Galois connec-
tion are closure operations, and their ranges are dually isomorphic.
Moreover, by the above correspondence between adjoint pairs and
hull-kernel isomorphisms,
every pair of mutually inverse dual isomorphisms between

closure ranges is induced by exactly one Galois connection
between the original ordered sets.
While adjoint maps are isotone (order-preserving), the partners of a

Galois connections are always anti tone (order-reversing). Certainly,
an advantage of adjunctions, compared with Galois connections, is
that they compose in a natural way. However, there is also a (less
natural) way of composing Galois connections (see Picado [113]). The
fixpoints on both sides of a Galois connection are said to be Galois
closed (while in an adjunction (A, Q), the fixed points of AO(! are said to
be Galois open, because >..oe is a kernel operation). The fundamental

problem of Galois connections is to determine or characterize (in some
practical way) the Galois-closed objects. In concrete situations, this
is often a difficult task (as we know from classical Galois theory). A
rather easy example is
The Galois Connection Between Subsets and Isotone Self-
maps
For any subset Y of a complete lattice L, the map
).. (Y) = Ay : L--+ L , X f---1 1\ {y E y :X ::::: y}
is a closure operation.
On the other hand, for any isotone self-map 1 of L, the set
e('y) = {x E L: 1(x) :S: x}
is meet-closed, hence a closure range. The pair ().. , f2) is a Galois
connection between the power set of L and the pointwise ordered com-
plete lattice of all isotone self-maps of L. The Galois-closed subsets
are precisely the closure ranges, and the Galois-closed self-maps are
the closure operations. Thus, the above Galois connection induces the
known dual isomorphism between the complete lattice of closure oper-
ations and that of closure ranges.
Indeed, an isotone map 1 : L--+ L is Galois closed iff >..(e('y)) :S: /,
which means that 1\ {a E L: x :S: a and 1 (a) :S: a} :S: 1(x) for all x E
L. Clearly, this implies x :S: 1(x) , and if x :S: I (Y) then y :S: a and
1(a) :S: a entail 1(x) :S: 1('y(y)) :S: 1('y(a)) :S: 1(a) :S: a, whence
1(x) :S: 1\{a E L: y :S: a and 1(a) :S: a} :S: 1(y).
Thus, 1 is a closure operation. Conversely, if the latter is assumed
then for x E L, we have x :S: 1(x ) and lb(x)) :S: 1(x) , hence
1\{a E L: x :S: a and 1(a) :S: a} :S: 1(x) .
That the Galois-closed subsets are just the meet-closed ones is easily
checked. Specifically, the Galois-closed subsets of power sets (under
the Galois connection between subsets of the power sets and inclu-
sion preserving maps) are just the closure systems, i.e. those systems
which are closed under arbitrary intersections; and the Galois-closed
inclusion-preserving maps are precisely the closure operators.
Many authors speak of a perfect adjunction or a perfect Galois connec-
12 M. Erne
tion if both partners are surjective (respectively, all involved elements

are Galois closed). Helpful tools in the determination of perfect ad-
junctions or Galois connections are often certain scales; most gener-
ally, these are merely functions J.L on a poset P so that p < p' implies
J.L(P) =f=. J.L(p'). In concrete cases, such a scale may be a cardinality,
dimension or degree function, etc. Clearly, every injection is a scale,
but not conversely.
Epi-Adjunctions and Right Perfect Galois Connections
For an adjunction or Galois connection (A, g) between quasi-ordered
sets P and Q, the following conditions are equivalent:
(a) .Xo g = idQ.
(b) A o g is injective.
(c) .Xog is surjective.
(d) g is injective.
(e) A is surjective.
(f) There is a map J.L : P ----+ N such that f-LOg is a scale.
(g) There is a scale fl: Q----+ N such that fl = flo .Xog. IfQ is partially
oTdered, (g) may be replaced with
(h) There is an isotone scale fl : Q----+ N such that fl S flo .Xo g.
The implications (a)==>-(b)==>-(d)==>-(h)==>-(g)==>-(f) and (a)==>-(c)==>-(e)
are obvious. For (f)==>-(a), note that .Xog(q) sq and J.LO&JOA o(l(q) =
JlP&~(q), hence .Xog(q) = q. Similarly, (e) implies (a) because AO(lOA =A.
Combining that result with its dual (exchanging left and right), one
obtains the
Theorem on Perfect Galois Connections
For an adjunction or Galois connection (.X , (2) between quasi-ordered
sets P and Q, the following conditions are equivalent:
(a) A and g aTe mutually inverse bijections.
(b) Ao g and (lOA are injective.
(c) Ao g and go.X are surjective.
(d) A and (2 aTe injective.
(e) A and g are suTjective. (f) There are maps J.L: P----+ N , fl : Q ----+ N
such that flo A and f-LOg aTe scales.
(g) There aTe scales J.L :P----+ N, fl :Q ----+ N with J.L = flo.X, fl = J.LO&J.
If P and Q aTe paTtially ordeTed, (g) may be replaced with
(h) There are isotone scales p,: P-+ N, jl: Q--+ N with
and jl ::; p,o (].
For (a) ::::;. (g) and (h) , take

p,:P-+PxQ, pH(p, ..\(p)) and jl:Q-+PxQ, qH(Q(q),q) .
Note that two maps ,.\ and (} form a perfect adjunction iff(} is both
a left and a right adjoint of ,\ , which means that they are mutually
inverse isomorphisms. Thus, a perfect Galois connection consists of
two mutually inverse dual isomorphisms.
1.4 Polarities and Axialities

A fundamental fact concerning Galois connections, pointed out in the
early sources by Birkhoff [3] , Everett [54] and Ore [38], is that all
polarities in the sense of Birkhoff, that is, all Galois connections be-
tween power sets, may be constructed in a unique way from relations
between the underlying sets. The partners of the induced Galois con-
nection associate with any subset of the one set the collection of all
elements of the other that are in relation to each element of the former
subset. We shall focus on that topic in Sections 3.2 and 3.3, but for
easy reference, we sum up already now the basic connections between
relations, polarities und dualities of closure systems.
Relations and Polarities
Every relation R between two sets A and B gives rise to a Galois
connection (R---t , +--R) between the power sets of A and B , by sending
X~ A to the set g -+ (X) = XR = {y E B: xRy for all x EX},
Y ~ B to the set +--R(Y) = YR = {x E A: xRy for all y E Y}.

The ranges of these two maps are dually isomorphic closure systems.
Conversely, any Galois connection between power sets and, conse-
quently, any dual isomorphism 1> between two closure systems comes
14 M. Erne
from a unique relation R between the underlying sets in that manner,

where xR y means that y belongs to the image of the closure of { x}
under .
For example, this construction provides a natural framework for the
Fundamental Theorem of Galois Theory, where the relation in question
is the "fix relation" between automorphisms and elements of a field:
CJRy <=> CJ(y) = y.
One map of the induced Galois connection sends each subset of the
automorphism group to the subfield of all elements fixed by these au-
tomorphisms; and in the opposite direction, the other map sends each
subset to the group of all automorphisms fixing the elements of that
set. Each Galois-closed subset of the automorphism group is then
a subgroup, and each Galois-closed subset of the field is a subfield.
The important (and more difficult) part of the Fundamental Theo-
rem of Galois Theory states that in the case of a Galois extension,
this restriction yields a dual isomorphism between the closure system
of all intermediate fields of the extension and the closure system of
all subgroups of the corresponding automorphism group (the 11Galois
group ") . However, that is not the way Galois himself attacked the
solvability of polynomial equations. We shall come back to that clas-
sical theme when we discuss Galois' famous memoir [56, 57] in Section
2.3.
Of course, there is also a covariant counterpart of polarities, obtained

by composing the polarities wit h dual isomorphisms via set comple-
mentation. Although such axialities, i.e. adjunctions between power
sets, are quite natural, t hey seem to be less common t han polari-
ties. Let us summarize the main facts about such adjunctions be-
tween power sets and the inducing relations (see the Primer on Galois
Connections [15] for details) .
Relations and Axialities
Every relation R between two sets A and B gives rise to an adjunction
(-+R, R +- ) between the power sets of A and B , by sending X ~ A to
the set -+R(X) = {y E B : x Ry for some x E X }, Y ~ B to the set
R +- (Y) = { x E A : xRy implies y E Y} .
Hence, the range of -+R is a kernel system, the range of R+- is a
closure system, and the two systems are isomorphic via restriction
and corestriction of these maps.
Conversely, any adjunction between power sets and, consequently, any
isomorphism 1> between a closure system and a kernel system comes
from a unique relation R between the underlying sets in that manner,
where xRy means that y belongs to the image of the closure of {x}
under 1>.
Polarities and axialities are related via the complement operator c by
the equations
Notice that if F is a function rather than a general relation, then

the above sets -+F(X) and F+-(Y) denote the usual image and preim-
age sets, respectively. Thus, in particular, any function F : A -1- B
gives rise to an adjoint pair of "lifted maps" -+F and p +- between
the corresponding power sets. As usual, we write F(X) for -+F(X) if
no confusion is likely to arise, but we avoid the ambiguous notation
F- 1 (Y) for F+-(Y). The adjointness condition is here the well-known
equivalence
F(X) ~ Y {::} X ~ F +- (Y).
For more examples not mentioned in this survey, see [15].
The aforementioned correspondence between relations and axialities
may be represented in the form of a further perfect adjunction:
Isomorphism Theorem for Axialities
Assigning to each relation R the map -+R yields an isomorphism be-
tween the power set of Ax A (i.e. the set of all relations on A) and the
complete lattice of all residuated maps on the power set of A. More-
over, this is also a semigroup isomorphism, sending the composition
product of relations to the composition product of mappings.
In order to demonstrate how polarities and axialities may simplify ar-

guments and computations, let us conclude this section with a typical
class of examples arising frequently in elementary linear algebra.
Let D be a division ring (skew field) and D~ the space of mxn-matrices
over D. Note that by the associative and distributive laws for matrix
operations, D;:_: is a ring (with unit), D~ is a left D;;_:-module, and
16 M. Erne
D;;" is a right D~-module. Denote by .C(D~) the closure system of all

left D~-submodules of D~, and by R(D;;") the closure system of all
right D~-submodules of D;;" .
We now consider, for any three natural numbers k, m , n , the orthogo-
nality relation j_=kj_m between A= D'k and B = D;;" given by
x j_ y {::} xy = 0.
Again by the associative and distributive laws, it is clear that for
any subset X of A, the Galois-closed set X j_ belongs to R(D-:;), and
on the other hand , for any subset Y of B, the Galois-closed set Y_1_
belongs to .C(D'k)· The point is now that, conversely, every submodule
S E .C(D'k) (and every T E R(D:;")) arises in that fashion. For the
proof, one has to find for any x E A \ S a matrix y E B such that
sy = 0 for all s E S but x y =/::- 0. Since S is a left D~-module, it is easy
to see that the rows of all matrices inS form a subspace U of the row
space Df, and that S consists precisely of those matrices whose rows
belong to U. Thus, x tj_ S means that at least one row x j of x is not
in U. Therefore, we find a basis {b 1 , ... , bn} of Df so that b1 = Xj and
{b2 , .. . br} is a basis for U. Solving the system of linear equations
'L.~= l biYi = (1, 0, ... , 0) ,
we get a column fj E D~ with Xjfj = b1 fj = 1 and bifj = 0 for i > 1,
hence ufj = 0 for all u E U. Consequently, the matrix y = (fj , ... , fj) E
D-:; has the required properties xy =/::- 0 and sy = 0 for all s E S.
Thus, we have established the
Dual Isomorphism Theorem for Matrix Submodules
For any division ring D and natural numbers k, m, n , the polarity
associated with the orthogonality relation between D''k and n-:; induces
a dual isomorphism between the left submodule lattice .C(D~) and the
right submodule lattice R( n-:;). In particular, for fixed n , each of the
lattices .C(D~) is isomorphic to th e row subspace lattice .C(D~) , and
each of the lattices R(D-:;) is isomorphic to the column subspace lattice
R(D~).
Notice that , specifically, .C(D~) is the lattice of all left ideals in the
matrix ring D~ , and R( D~) is that of all right ideals.
In the present context, not only polarities, but also certain axialities
are of particular use. Consider the "column relation" C between D~
and n; given by xCy iff y belongs to the column space of x, I.e.

xb = y for some b E n;.
Isomorphism Theorem for Ideals of Matrix Rings
The axiality associated with the column relation between D~ and n;
induces an isomorphism between the left ideal lattice £(D~) and the left
subspace lattice £( n;). Moreover, every left ideal of D~ is principal.
However, it should be emphasized that not all Galois-open subsets of
n; with respect to that axiality are subspaces, but the Galois-open
sets are the unions of subspaces, hence form an Alexandroff topology
(closed under arbitrary unions and intersections). The isomorphism
theorem is an immediate consequence of the previous remarks concern-
ing submodules, i.e. ideals of D~, and the associated column spaces.
Summing up the previous definitions and facts, we have four orthog-
onality relations
l_=ll_l~nrxn; with orthogonal spaces X _i and Y_1
K = nl_ 1 ~ D~ x n; with right kernel spaces X K
L = 11_ n ~ D1 x D~ with left kernel spaces YL
N = nl_n ~ D~ x D~ with annihilator ideals XN and YJV .
Furthermore, we have the column relation
c ~ D~ X n; with column spaces --+c(X)
and the row relation
R ~ nr XD~ with row spaces--+ Rd(Y)
where Rd denotes the dual of the relation R. These six relations are
connected via the relation product
RS = {(x, z ) : x Ry and yS z for some y}
as indicated in the following commutative diagram:
R c
K L
Dn_ _ _ __ _ Dn
n N n
18 M . Erne
Thus, NC = K, RN = L , LC =..l , RK =..l , and so RNC =..l.

Lifting the relations to the associated polarities, respectively, axialities
and suitable restriction yields the following commutative diagram of
isomorphisms (perpendicular) and dual isomorphisms (horizontal or
diagonal):
.C(Dn } + - - -- --
.C(D~ ...-------R(D~)
These isomorphisms are the starting point in John von Neumann's in-
genious coordinatization theory for complete, complemented modular
lattices [145] .
1.5 Formal Concept Analysis

This modern, practice-oriented and mathematically founded theory
has been initiated and propagated by Rudolf Wille and his coworkers
since the seventies of the last century. Comprehensive sources about
that theme are the books [58] and [59], the latter containing also a
long list of references to further research on the subject.
Formal concept analysis combines the mathematical ideas of polarities
and lattice theory with philosophical concepts (in particular, that of
intent and extent), with methods of ordering and classification, and
with diverse algorithmic tools for applications, not only in mathe-
matical but also in various extra-mathematical areas, from technical
engineerings, biology, medicine, history and social sciences up to fine
arts and music. Some of the prominent aims and topics formulated by
Wille in [153] are:
• Hierarchical classification of objects

• Minimal systems of implications between sets of attributes
• Combination of individual concept systems
• Reduction to basic concepts

• Conceptual scaling and measuring
• St ructures for the representation of knowledge data
But Wille also views formal concept analysis as an important tool for
the purpose of "restructuring lattice theory". In the introduction to
his inspiring article [153], he writes:
. . . we go back to the origin of the lattice concept in nineteenth-

century attempts to formalize logic, where the study of con-
cepts played a central role (cf Schroder {131}) . Tradi-
tional philosophy considers a concept to be determined by
its "extent" [extension} and its "intent" {intension, com-
prehension}: the extent consists of all objects (or entities)
belonging to the concept while the intent is the multitude
of all attributes (or properties) valid for all those objects.
Let us explain briefly the mathematical framework for t hese ideas

and the fundamental notions of formal concept analysis. For more
background, see [59] and [6, 48, 49].
Mathematically, a (formal} context is a triple :OC = ( J, M, I), consisting
of a set J (of "subjects" or "objects"), a set M (of "marks " or "at-
tributes") and an ("incidence ") relation I ~ J x M. The concepts are
the pairs (X, Y) with X = Y1 ~ J (the "extent") and Y = X 1 ~ M
(the "intent"). Ordered by "specialization", i.e. inclusion of the first
component or, equivalently, by dual inclusion of the second compo-
nent, they form a complete lattice, the so-called concept lattice BK
Obviously, this lattice is isomorphic to the closure system of all ex-
tents (via projection onto the first component) and dually isomorphic
to that of all intents (via projection onto the second component ). The
resulting dual isomorphism between these two closure systems is just
that induced by the polarity associated with the relation I. If the
maps
r : J-t B:OC, j t-+ ( {jf1 , {jf) and p,: M -7 B:OC, m t-+ ( {m}I, {m}/)
are injective, one speaks of a purified context. Specific purified contexts
are the basic contexts or L-contexts, where J is a join-dense subset of
a complete latt ice L (i.e., every element of L is a join of elements
20 M . Erne
in J), M is a meet-dense subset of L (defined dually) , and I is the

relation between J and M induced by the order relation of L . As we
saw before, a context is purified iff it is isomorphic to a basic context.
Moreover, every context K has a canonical "purification" lKQ , obtained
by passing from j to 'Y(j) , from m to J-t(m) , and from I to the relation
induced by the order of BK. By definition, lKo is a basic context such
that K and lKo have isomorphic concept lattices. Indeed, the concept
lattice of any basic context is isomorphic to the original lattice L , by
sending x E L to the concept
(J n t x, M n t x) = ( {j E J : j ~ x} , {m E M : x ~ m}) .
More precisely, one has the

Fundamental Theorem of Formal Concept Analysis
An arbitrary complete lattice L is isomorphic to the concept lattice of a
context K = ( J, M, I) iff there is a join-dense embedding ;y : J -t L and
a meet-dense embedding ji : M -t L such that jIm {:} i (j) ~ ji( m)
in L.
This theorem extends the well-known characterization of the Dedekind-
MacNeille completion or completion by cuts of an arbitrary poset as
the complete lattice (unique up to isomorphism) in which the poset is
join- and meet-densely embedded (see [3, 28, 91] and Section 3.2) .
The interplay between contexts and complete lattices is very fruitful in
both directions: for a given context, a careful inspection of the concept
lattice provides insight into hierarchies, implications and independen-
cies that wouldn't be obvious from large context tables. On the other
hand, at least every finite lattice and, more generally, every doubly
based lattice, that is, every complete lattice L whose V-irreducible el-
ements form a join-dense subset J and whose /\-irreducible elements
form a meet-dense subset M , is coded entirely by its basic context
(J, M, I). The latter may be regarded as a "logarithm " or a "base " of
the whole lattice, usually having much smaller size than L . As pointed
out in [49] and in [52], the passage from a context to its concept lattice
and, in the reverse direction, from a complete lattice to the context
of irreducible elements, provides a functorial adjunction between the
category of complete lattices with complete homomorphisms and the
category of contexts with so-called conceptual morphisms. Restriction
to doubly based lattices on the one hand and to purified contexts on
the other hand even yields a categorical equivalence.

A useful generalization are the so-called multi-valued contexts, where
the incidence relation is substituted by a map ~ : J x M--+ N. Of
course, this includes the usual notion of contexts, by taking the two-
element set {0, 1} for Nand replacing relations by their characteristic
functions. But the multi-valued approach is much more flexible and
has manifold modern applications, for example to fuzzy concept analy-
sis and fuzzy logic [20, 21, 140], relational data bases [23], information
systems [111, 112], and Chu spaces [115] .
Some hundreds of further references to themes involving adjunctions
and Galois connections may be found via the MathSciNet using the
relevant key words.
2 The Roots of Galois Connections

It is almost impossible to trace back the idea of adjunctions and Galois
connections to its historical origins. Actually, as a mathematical con-
cept they probably did not emerge before t he 20th century. However,
they are already involved, more or less implicitly, in certain construc-
tions and computations of earlier centuries. Of course, the examples
selected below are far from being representative for all such occur-
rences in ancient mathematical contexts. Nevertheless, they show how
certain considerations are made easier or at least more transparent by
the language of adjunctions or Galois connections.
We shall start with the theory of symmetric functions, initiated by
Girard, Newton, Waring and Vandermonde. Then we touch upon La-
grange's ingenious contributions to the solution of algebraic equations
and, of course, we discuss Galois' pioneering work on the same theme.
Among the historical background literature on the theory of polyno-
mial equations, we recommend the recent book by J.-P. Tignol [138],
which combines historical thoroughness with mathematical accuracy.
Then we pass to the field- and ring-theoretical ideas of Dedekind de-
veloped in his early Lectures on Algebra [26), and we have a look at
Hilbert's "Nullstellensatz" from the perspective of Galois connections.
Other interesting aspects are Schroder's contributions to universal al-
gebra in t he light of Galois connections and Dedekind's early inves-
22 M . Erne
tigations on closure operators, lattices ( "Dualgruppen" ) and lattice-

ordered groups (to be discussed later in Section 4.2).
2.1 Prelude on Symmetric Functions

An equation consists of several terms, some known and some
unknown, some of which are together equal to the rest; or
rather, all of which together are equal to nothing; for this is
often the best form to consider.
Translated from: Rene Descartes, La Geometrie

Historically, the starting point for Galois theory, and hence for the
appearance of Galois connections, were investigations concerning the
solvability of polynomial equations. For Galois and Lagrange, a crucial
tool in that undertaking were so-called symmetric {rational) functions
and, in particular, symmetric polynomials. A polynomial or rational
function in several variables is symmetric iff it remains fixed under
the exchange of any two variables, and consequently, under arbitrary
permutations of all variables. A rational function that is invariant
only under a certain subset T of permutations may be called partially
symmetric {with respect toT). On the other hand , for a given rational
function f , those permutations of the variables that leave f unchanged
form a group, the isotropy group or stabilizer!(!) .
By its very definition, the notion of (partially) symmetric functions
comes from an elementary Galois connection: the "fix relation"
B -1 ! {::} B(f) = !
gives rise to a a polarity in Birkhoff's sense (cf. Sections 1.4 and
3.2) between S(X), the full permutation group of X = {x 1 , . . . , xn}
and K(X) , the field of rational functions over a ground field K. This
polarity associates with any set T ~ S(X) of permutations the subfield
T 4 of all rational functions in K(X) fixed by each (} E T, and in
the opposite direction with any set F of rational functions the group
F-j = nUU) : f E F} of all permutations fixing each f E F. In
particular, S(X) 4 is the field Ks(X) of all symmetric functions.
The following easy result , mentioned in Dedekind's Lectures on Al-
gebra [26] (see Section 2.4) , may be regarded as a forerunner of the
Fundamental Theorem of Galois Theory.
Theorem on Partially Symmetric Functions

The assignment f HI(!) is a surjection from K(X) (and even from
the polynomial ring K[X]) onto the subgroup lattice of S(X). Hence,
the Galois connection associated with the fix relation induces a dual
isomorphism between the subgroup lattice of the symmetric group S(X)
and the closure system form ed by the "fields of partially symmetric
functions", i.e. those subfields of K(X) that consist of all rational
functions fixed by the members of a subgroup of S(X) .
Indeed, given a subgroup G of S(X), one easily finds polynomials
f E K[X] that are fixed precisely by the members of G, for example
f(xl, ... , Xn) = fleEG L:~=l B(xk )k.
Now to the history of symmetric functions. The elementary symmetric
functions in the variables x 1 , ... , Xn,
Sk = Bk,n = L:l~}l < ... <jk~n fl~=l Xjm = L:Y~X, #Y=k f]y EY Y'
are, up to signs, the coefficients of the polynomial
(*) TI7=(x- Xj ) =
1 L: ~=o Bk,n(-x)n-k,
a result probably due to Albert Girard (1629) in its full generality
(see [64]), while specific cases for small degrees were already known to
Frangois Viete (Francesco Vieta; see [143]).
Historians attribute the first major results on symmetric polynomials
to Isaac Newton (see, for example, [39]). About 1666, Newton listed in
his papers a whole sequence of formulas relating symmetric functions
of the roots of a cubic polynomial with its coefficients. In these notes,
one also finds the case n = 8 of the famous formula
L:f=l(-1 )j aj,n Bk- j,n + ksk ,n = 0
connecting t he power sums
aj,n L...-i=l x.;,J.
= ""n
with the elementary symmetric functions Sk,n· This formula occurs, in
a slightly different form and without proof, in Newton's Arithmetica
Universalis from 1707. It enables one to compute recursively t he
power sums from the elementary symmetric functions, and vice versa,
by the earlier mentioned process of shifting parts of the equation from
one side to the other. At least for n :::; 4, t he formula was already
24 M. Erne
known to Girard [64].

As a consequence of Girard's equation (*), expressing the coefficients
of a polynomial as symmetric functions of the roots, one has t he follow-
ing formulas providing the passage between the elementary symmetric
functions in n and in n-1 variables, respectively:
(+) Sk,n = Sk,n-1 + Sk-1,n-1Xn,
(-) sk,n-1 = 2:1=oSk-j,n( -xn)j.

Although Newton wrote down only a list of specific formulas involving
symmetric polynomials, it is assumed today that he was aware of the
general
Fundamental Theorem on Symmetric Polynomials
Every symmetric polynomial may be expressed uniquely as a polyno-
mial in the elementary symmetric fun ctions.
Mathematicians of the 18th century used that theorem without any
inhibition. Lagrange qualifies it as "self-evident" but mentions that
a "rigorous demonstration depends on the principles" developed in
his memoir on the solvability of algebraic equations [85]. Construc-
tive proofs (by recursion) of the existence part appeared at the same
time in Waring's Meditationes Algebraicae [152] and in Vandermonde's
M emoire sur la resolution des equations [141], while the uniqueness
part seems to have been neglected before Gauss' demonstration in
[60]. The first complete and rigorous proof of both parts and a correct
argument for the extension to rational functions is probably due to
Dedekind [26] (see Section 2.4). Today, we find many different proofs
in the text books, some of them relying on parts of the more general
Galois theory. But the historical order was just the converse, starting
with Girard's formula and culminating in Lagrange's and Galois' the-
ory of polynomial equations, based on the Fundamental Theorem on
Symmetric Polynomials.
For the sake of convenience and easy reference, we add here two slightly
stronger order-theoretical results in terms of rings; the first one ex-
tends Waring's method of difference exponents [152], and t he second
imitates Dedekind's reduction method from n to n - 1 variables [26].
Of course, the ring-theoretical terminology is borrowed from the math-
ematics of the 20th century, but it could be replaced easily with (less
definite) phrases in the language of Waring, Lagrange and Galois.
Let R be an integral domain , that is, a commutative ring with 1 and

no zero divisors. We denote by R[X] the ring of all polynomials over
R in a set X of variables; if X= {x} , we put, as usual , R[x] = R[X].
The symbol Rs[X] will denote the subring consisting of all symmetric
polynomials (fixed by all permutations of X). Mathematicians are
familiar with the representation of polynomials as linear combinations
of monomials. But what is a monomial? The easiest, though rather
abstract definition uses the additive monoid w[x] of all functions cp
from X tow = {0, 1, 2, ... }with finite support {x E X : cp(x) -1- 0}
and identifies monomials as functions from w[x] into R having value 1
for one argument m E w [X] and value 0 for all others. Denoting that
monomial by xm, the product of two monomials is then given by "the
exponential law" xmxm' = xm+m' 0
Starting with 1 instead of 0, we mean by wn the set of all finite se-

quences (m 1 , . .. , mn) of natural numbers, and by w~ the subset of
all decreasing sequences among them. These sets are ordered lexico-
graphically in the usual way (from left to right). That lexicographical
ordering is a well-ordering, so it may be used for induction proofs as
in the case of natural numbers. The "reverse sum operator"
I; : Wn --7 w~, (m1 , ..., mn) H (m1 + ... + mn , ... , ffin-1 + mn, mn)
and the "reverse difference operator"
.6. : W~ --7 wn, (m1 , ... , mn) H (m1- m2 , ..., mn- l- mn , mn)
induce mutually inverse bijections between wn and w~. (However, :E
and .6. become isomorphisms, hence form a perfect adjunction only if
wn and w~ are ordered anti-lexicographically from right to left) . For a
finite set
one has a canonical isomorphism between the additive monoid wn

and the multiplicative monoid of all monomials over Xn , sending
m = (ml , ... , mn) to xm = n
xr;:k Now , the (multi-)degree MU) of a
0
non-zero polynomial f E R[Xn] is defined to be the lexicographically

largest "exponent" m of monomials occurring in f (supplemented by
the setting M(O) = -oo). Thus, one may write fin the form
J = >-.(!) xMU) + r
where >-.(!) is the "leading coefficient", xMU) is the "leading mono-
26 M. Erne
mial", and r is the ((remainder polynomial" having smaller degree

than f. A polynomial with leading coefficient 1 is said to be monic.
(Note that our definition requires monomials to be monic.)
There is a well-defined substitution homomorphism
(J : R[Xn]-+ Rs[XnJ, f(xl, ... , Xn) r-+ f(sl,n, ... , Sn,n)·
Now, the Fundamental Theorem on Symmetric Polynomials amounts
to a
Perfect Adjunction Between Polynomials and Symmetric Poly-
nomials
The substitution homomorphism CJ induces an algebraic and order-
theoretical isomorphism between the rings R[Xn] and Rs[XnJ, ordered
by divisibility.
The inverse isomorphism 6 : Rs[Xn]-+ R[Xn] may be defined recur-
sively by 6(0) = 0 and
6(s) = ,\(s) xMs)+ 6(8) otherwise, where
Jt(s) = 6(~t(s)), 8 = s- ,\(s)CJ(xMs)) E Rs[Xn], ~t(8) < ~t(s).
Furthermore, It and {t satisfy the degree equations
~tUg)= ~tU) + ~t(g) and Jt(st) = Jt(s) + Jt(t).
In particular, they induce isotone scales on the partially ordered sets
of all monic polynomials and of all symmetric monic polynomials, re-
spectively.
Proof. The first of the latter two equations is an easy consequence of
the (distributive) product definition for polynomials, and the second
follows from the first by the fact that 6 preserves sums. Note that
the degree ~t( s) of a symmetric polynomial s is in fact decreasing, by
definition of the lexicographical order. Hence, ft and Jt are scales pre-
serving the divisibility relation. Iterated application of the equation
~tUg) = ~tU) + ft(g) yields
ft(CJ(xm))= ~t(f1sj,nmi) ='Lmjft(Sj,n) = ~(m) .
As ~ is inverse to 6, we conclude that
ft(CJ(xMsl)) = ~(Jt(s)) = ~(6(~t(s))) = ~t(s).
From this, we infer the inequality
~t(8) = ~t(s- ,\(s)CJ(xMsl)) < ~t(s),
which makes the recursive definition of owell-founded.

Now, it is easy to prove inductively the equation a(o(s)) = s for all
symmetric polynomials s E Rs[X], starting with a(o(O)) = 0 :
a(o(s)) = .\(s)a(xfi(s)) + a(o(s-.\(s)a(x fi( s)))) = s.
As :E is injective, so is J.-l o a on monomials. Hence, for any non-zero
polynomial f = L am xm, in the linear combination
a (f) = L ama(xm)
any two summands have distinct leading monomials. The largest of
them is also the leading monomial of a(!) , whence a(!) is not zero.
This proves injectivity of a and shows that ois inverse to a. Since the
latter is an algebraic isomorphism, so is the former, and a(!) divides
a(h) iff f divides h. D
Though J.-l(a(xm)) = :E(m) entails ji(a(xm)) = ~(J.-l(a( xm)) = m =

J.-l(xm), the diagram below unfortunately does not commute!
0
R[Xn] Rs[Xn]
~~/>
wn U { -oo}
For example, if t = x 12 x 2 + x 1x 22 and s = a(t) then J.-l(s) = x 13 xl ,

hence
We now formulate the second extension of the Fundament al Theorem

on Symmetric Polynomials.
Characterization Theorem for Rings of Symmetric Polyno-
mials
A sequence of rings Rn satisfies the conditions
(0) Ro = R, (1) Rn ~ R s [Xn], (2) Rn-1 ~ Rn[xn], and
(3) fn( xn) = 0 for some monic polynomials fn E Rn[y] of degree n
if and only if Rn = Rs[Xn] for all n.
We prove the "only if" part by induction (the "if" part is easier).
By (0), we have Ro = R = Rs [0] = R s [Xo] . Assume Rn-1= Rs[Xn- 1]
28 M. Erne
and consider any g E Rs[Xn]· Expanding g according to powers of Xn

yields coefficients that are symmetric in x 1 , .. . , Xn_1 . Thus,
(4) Rs[Xn] ~ Rs[Xn-1Hxn],
and by the induction hypothesis, g belongs to Rn-dxn ]· Now, we infer
from (2) that there is a polynomial hE Rn[y] of minimal degree such
that h(xn) =g. This degree must be less than n, because otherwise
subtraction of a suitable multiple of the polynomial fn in (3) would
reduce the degree. But by symmetry, the polynomial h-g E Rs [Xn][Y]
has the distinct roots x 1 , ... , Xn, and it would have the same degree as
h (note t hat g is here a constant) unless h- g is the zero polynomial.
Thus, g = h belongs toRn . D
The Fundamental Theorem on Symmetric Polynomials now follows

immediately by taking for Rn the ring R[s 1,n, ... , Sn,n] generated by
the elementary symmetric functions: (0) and (1) hold by definition,
the recursion (-) ensures (2), and the polynomial fn = n~= 1 (x- Xk)
satisfies (3).
It is not difficult (cf. Section 2.4) to derive from the Fundamental
Theorem on Symmetric Polynomials the
Fundamental Theorem on Symmetric Functions
A rational function in n variables over a field is symmetric if and
only if it has a (unique) representation as a rational function in the
elem entary symmetric functions.
Another consequence of the Characterization Theorem is
Corollary 1. Any subset Y of the set X = Xn of variables satisfies

the equation
Rs[X\ Y] [Y] = Rs[X] [Y].
For the proof, observe first that (2) for Rn = Rs[Xn] together with
(4) gives
Rs[Xn-1] [xn] = Rs[Xn] [xn]·
The general formula is then obtained by induction.
Finally, substituting for the indeterminates t he (pairwise distinct)
roots of a polynomial, one arrives at the useful
Corollary 2. If a polynomial f E R[x] has the set Z of roots, then

Rs [Z \ Y] [Y] = R[Y] for all Y ~ Z, in particular

Rs[Z] =R and Rs[Z\ {z}][z] = R[z] for all z E Z.
The last two formulas say that any symmetric polynomial (or rational)
expression in the roots of a given polynomial must already belong to
the ground ring (or field , respectively) , and that a polynomial (or
rational function) that is symmetric in all but one root is actually
a polynomial (or rational function) in that root . This is the form
in which Galois and his predecessor Lagrange used the Fundamental
Theorem on Symmetric Functions.
2.2 Joseph-Louis Lagrange: Permutation Groups,

Partially Symmetric Functions and Interpolation
Voila, si je ne m e trompe, les vrais principes de la resolution
des equations et ['analyse la plus propre a y conduire; tout
se reduit, comme on voit, a une espece de calcul de combi-
naisons, par lequel on trouve a priori les resultats auxquels on
do it s 'attendre.
J. - L. Lagrange in Article 109 of his memoir
on the algebraic solution of equations (1771)
About a century after Newton, it was Lagrange who developed a rather

comprehensive theory of partially symmetric functions. Lagrange's
immense mathematical work [85] is one of the most impressive cre-
ations of human kind. His contributions to the theory of polynomial
equations [86] and its basic group-theoretical aspects alone made him
a pioneer in the fundaments of Galois theory. Though not using ex-
plicitly the abstract notion of a group, Lagrange systematically dealt
with
• permutation groups of the roots of a polynomial equation and

• the change of rational functions under permutations of the vari-
ables.
These are the essential tools not only in Lagrange's own considera-
tions on the solvability of polynomial equations, but also in the later
expansion of the theory by Abel, Galois and Dedekind.
30 M. Erne
In order to make Lagrange's results more easily accessible, we antici-

pate a few additional ideas and notations due to Dedekind (see Section
2.4) and feel free to include some modern terminology. Thus, given a
field K, we denote, as usual, by K(A) the field obtained by adjoining
the elements of A, and we write K(a 1 , •.• ,am) for K( {a 1 , . . . ,am}). In
particular, as before, K(X) or K(x 1 , ... , xm) designates the field of all
rational functions in the variables x 1 , ... , Xm over K (Lagrange, Ga-
lois and Dedekind preferred the letter m instead of n). Similarly, if
S = { s 1,m, ... , Sm,m} is the set of elementary symmetric functions in
m variables, then the Fundamental Theorem on Symmetric Functions
may be expressed by the identity Ks(X) = K(S).
Denote by B(f) or f(} (!) the rational function obtained from f E
K(X) by applying the permutation(} to the variables. Then, in mod-
ern language,
I(!) = {0 E S(X) : O(f) = f}
is the stabilizer off with respect to the symmetric group
S(X) ~ S( {1, ... , m}) =Sm.
In his famous memoir Reflexions sur la resolution algebrique des equat-
ions [86], Lagrange implicitly uses the quasi-order
f :S g {::> I(g) ~I(!)
and calls two rational functions similar if they are equivalent with re-
spect to that quasi-order, meaning that they have the same stabilizer.
One of the similarity classes is that of all symmetric functions.
Another useful ingredient in Lagrange's theory of polynomial equa-
tions is what nowadays is termed
Lagrange's Subgroup Theorem
The order (size) of any subgroup of a finite group G divides the order
of G.
But what Lagrange really stated in Articles 97-99 of his memoir [86]
is, rephrased in modern terms, the following more concrete
Theorem on Minimum Polynomials for Rational Functions
For f E K(X), the minimum polynomial (the least monic polynomial
with root f ) over K (S) is given by
mj(X) = ngEfSm (x- g)
where fSm is the "orbit" {fO: () E S(X)} = {f(xa(l) , ... ,Xa(m)): CJ E

Sm}·
The order of the stabilizer I(!) is m!jd, where d is the degree of m 1
and m! is the order of Sm ~ S(X). Moreover,
mJ(x)m!/d = TioES(x)(x- fO).
Lagrange did not carry through a general proof but considered a few
particular cases, among them two-, three- and four-element subgroups
of S(X) and the case I(!) = S(X) , which amounts to the Fundamen-
tal Theorem on Symmetric Functions. Conversely, that theorem is the
base for a rigorous proof of the general statement. How the general
theorem extends to algebraic elements instead of rational functions,
and how it provides a perfect adjunction between subgroups and min-
imum polynomials will be discussed in Sections 2.3 and 2.4, dealing
with the later progrees made by Galois and Dedekind.
The essence of Article 104 in Lagrange's memoir [86] is the following
remarkable statement (in English translation):
If t and y are any two [rational} functions ... fin the roots
of a polynomial equation} and if these functions are such
that every permutation of the roots which changes y also
changes t, one can, generally speaking, express y rationally
in terms oft and ... [the coefficients of the equation].
In modern form, that statement reads as follows:

Theorem on Stabilizer Groups of Rational Functions
For any two functions t, y E K(X), the group I(t) is contained
in the group I (y) if and only if there is some polynomial q over K (S)
such that y = q(t). Hence
I(t) ~ I(y) {::} y :S t {::} K(S) (y) ~ K(S) (t).
Thus, there is a perfect adjunction between the similarity classes of

K(X) and the simple field extensions of K(S) contained in K(X).
The "if" part is obvious. For the "only if" part, Lagrange first assumes
(in Article 100 of the memoir) that t and y are similar functions in
x 1 , . .. , Xm and concludes that if t 1 , . .. , tn are the various "forms" ob-
tained by permutation of the variables in t , the same permutations
32 M. Erne
yield the different forms y 1 , ... , Yn of y. Then he remarks that the

sums
M1 = 2:7= 1 tJ Yi (j = 0, ..., n-1)
are symmetric functions, hence rationally expressible by the elemen-
tary symmetric functions (the coefficients of the general polynomial
equation of degree m). It is now clear that, solving that system of
linear equations for Yi (given the values ti), one obtains each Yi as
a rational function of all the values t 1 , ... , tn. But Lagrange's crucial
observation is that Yi is already expressible rationally in terms of the
coefficients M 1 and the corresponding single ti. For that, he carries
through rather long and complicated calculations, which, admittedly,
lead to more general statements, also involving equations with mul-
tiple roots. A rather easy proof (as proposed by Edwards [39]) uses
Cramer's rule. But still easier and more elegant is Dedekind's idea to
use another fundamental tool invented by Lagrange, namely his inter-
polation polynomials. How that approach works will be explained in
Section 2.4.
Combining the above result with the Theorem on Partially Symmetric
Polynomials, one already obtains the major part of the Fundamental
Theorem of Galois Theory for the "generic case" of variaLle roots:
namely, a dual order embedding of the subgroup lattice of the sym-
metric group S(X) in the closure system of intermediate fields between
K(S) and K(X). In order to see that this Galois connection is per-
fect on both sides (hence an isomorphism) , one only has to assure
that each intermediate field occurs in that way as a simple extension
- and that remaining statement is just a special case of the Theorem
on Primitive Elements, due to Abel [1] and Galois [56] (see the next
section).
Fundamental Theorem of Galois Theory for Rational Func-
tions
The subgroup lattice of the symmetric group S(X) is dually isomorphic
to the lattice of fields between K(S) = K s (X) and K(X).
The main problem in extending the above result to the general situ-
ation of "numerical" roots r 1 , ... , r m of a polynomial equation instead
of variables is the definition of the group I (f) in that case (where f is
now an element of K(r 1 , ... , rm) rather than a rational function), and
the solution of that problem was one of Galois' major achievements.
Adjunctions or Galois connections sometimes arise in a somewhat un-

expected way, for example in group theory. In what follows, we shall
have occasion to see how Lagrange's Subgroup Theorem together with
a suitable adjunction between divisors and subgroups (also called "di-
visors" by Galois and Dedekind) leads to a basic theorem on cyclic
groups. As demonstrated by Gauss, Abel and Galois, such groups are
of central importance for the solvability of certain equations and for
the theory of finite fields. In particular, the reasoning below yields the
existence of primitive roots modulo prime numbers, a central tool in
number theory developed by Gauss.
Given a finite group G of order n whose neutral element is denoted
by 1, let L(G) designate the closure system (hence complete lattice)
of all subgroups of G, and D(n) the lattice of all divisors of n. By
Lagrange's Subgroup Theorem, one has a map
>..a: L(G) ---t D(n), H f-7 #H,
assigning to each subgroup its order (cardinality). If G is commuta-
tive, one also has a map in the opposite direction,
Qa: D(n) ---tL(G), dr-7 {g E G : gd = 1},
sending each divisor d of n to a subgroup (!) of G.
For h E G, let G(h) denote the cyclic subgroup generated by h. If
G is finite then the order o(h) of G(h) divides the order of G, and in
particular, h o(h) = 1.
Characterization Theorem for Cyclic Groups

Given a finite group G of order n, the following statements are equiv-
alent:
(a) G is cyclic.
(b) G is isomorphic to the additive group Zn of integers modulo n.
(c) Aa and (}a are mutually inverse isomorphisms.
(d) Aa and (}a form an adjoint pair.
(e) For all divisors d of n, the equation gd = 1 has at most d solutions.
Proof. The equivalence (a) ¢::? (b) and the implications (a)::::} (c)::::} (d)
=}(e) are straightforward: if G is cyclic of order n, say G = G(h),
34 M . Erne
then for each d E D(n), the unique subgroup of G having order dis
G(hnfd).
(e):::}(a). Ford= o(h) , we have G(h)r:;_pc(d), and# pc(d)s_d forces
pc(d)=G(h). Now, put ca(d) = #{hEG: o(h)=d}. From (a){:} (b)
ford instead of n we infer that pc(d) =G(h) ::::::= 'lld ::::::= Zn(n/d). Thus
ca(d) = c12 a(d)(d) = Czn(n/d)(d) s_ Czn (d) ,
and so there is an injection i from G into 'lln with o(h) = o(i(h)).
Since both groups have the same order, i must be onto, and as 'lln
does have an element of order n , so does G . D
The implication (e):::} (a) applies to any finite subgroup G of the mul-
tiplicative group of a field K , because the polynomial xn - 1 has at
most n roots in K. This gives a short proof of the famous
Theorem on Multiplicative Subgroups of Fields
All finite subgroups of the multiplicative group of a field K are cyclic.
In particular, if K is finite then K \ { 0} is a cyclic group under mul-
tiplication.
2.3 Evariste Galois' Memoir on the Solvability of

Equations
Get Ouvrage n'ayant pas ete compris, les propositions qu 'il
renferme ayant revoquees en doute, j'ai du me contenter de
donner, sous forme synthetique, les principes generaux et une
seule application de ma theorie. Je supplie mes juges de lire
du moins avec attention ce peu de pages.
Galois' preface to the second version of his memoir,
crossed out by himself but added later by his friend
Auguste Chevalier in the posthumous publication.
Perhaps the most famous and controversial paper in the history of alge-
bra is Galois' Memoire sur les conditions de resolubilites des equations
par radicaux, dated January 1831 but not published until1846 by Liou-
ville in the Journal de Mathematiques (see [56], or [57] for a later com-
mented edition, or Edwards [39] for an English translation). Hundreds
of elaborations, supplements, comments, refinements, and historical
remarks have been written by later authors about that memoir (see,
for example, [139] and the bibliography therein). Almost everybody

dealing with Galois connections cites the Fundamental Theorem of
Galois Theory as the "classical prototype" of such connections. How-
ever, that theorem, stated in its modern form as a dual isomorphism
between the subfield lattice of a Galois extension and the subgroup
lattice of the corresponding Galois group, was formulated explicitly
quite some time after Galois. Surprisingly, when looking at the vaste
literature on Galois theory, one misses an investigation of the role of
Galois connections in Galois' original papers. At least implicitly, a
few adjunctions and Galois connections already emerge in Galois' pi-
oneering memoir - though, admittedly, one will not find any explicit
order-theoretical terms in Galois' formulations. However, the natural
order of fields and groups via inclusion is inherent in the whole theory,
and so are two other natural orders, namely the divisibility relation be-
tween numbers and that between polynomials. As mentioned earlier,
in other writings Galois even used the word "divisor" for subgroups, a
convention adopted by Dedekind in his Lectures on Algebra [26] (see
the next section) .
Galois' memoir starts with a series of "principles" and "definitions".
Instead of giving explicit rigorous definitions of fields and groups, Ga-
lois describes how "rational" or "rationally known quantities" are ad-
joined stepwise on the way towards the solution of a polynomial equa-
tion, and he indicates how "groups of permutat ions" have to be formed
and investigated for that purpose.
Given a finite set Z = {a, b, c, ... } of m elements ("letters"), Galois
means by a "permutation" (of the letters) any m-tuple whose entries
are the distinct elements of Z - or, in modern terminology, any bijec-
tion between m = {1, ... , m} and Z; while a substitution in the sense of
Galois (and Dedekind) is a bijection (} : Z -+ Z (hence a permutation
of Z in the usual sense also adopted by Lagrange). Of course, it is easy
to establish a one-to-one correspondence between these two concepts:
denoting by P(Z) the set of all "permutations" 7r : m-+ Z and fixing
one such 1r, one obtains bijections
l 1r : S ( Z) -+ P (Z), (} r---+ 1r (} = (} o 1r,
r 7r : S (m) -+ P ( Z), a r---+ a 1r = 7r o a ,
which become group isomorphisms if the product of two permutations
36 M. Erne
'1/J, ~ is defined as 'lj;7r- 1 ~ (whence

1r is the neutral element) . Thus,
Sm = S(m), S(Z) and P(Z) are pairwise isomorphic groups in our
modern sense.
Galois gives a double description of his "groups" : first , as collections
of "permutations" with the following two properties:
(i) each of the permutations arises from one of them by a [bijective]
substitution of the letters,
(ii) the set of substitutions involved is always the same, no matter
which permutation one starts from .
Alternatively, Galois takes those substitutions as the constituents of
his groups and concludes that such collections of substitutions must
be closed under composition (which by the finiteness of the underlying
set suffices to assure the existence of inverse elements).
The passage between the two group concepts may be formulated in
the framework of a perfect Galois connection. In order to avoid ambi-
guities, let us speak in the present context of a grouping block when we
mean a set of permutations satisfying the above conditions (i) and (ii) ,
while subgroups of the symmetric group S(Z) will be referred to as
substitution groups. Hence, for any grouping block G and 1r E G , the
set Sa = 1r- 1 G is a substitution group not depending on the chosen
1r, and is referred to as the substitution group of G . Furthermore, let
us call two grouping blocks G , H equivalent if there is a a E Sm with
aG =H. By a grouping, we mean a partition of P(Z) into equivalent
grouping blocks. This is the central tool in Galois' theory. Now, a
sequence of straightforward verifications yields:
The Galois Correspondence Between Subgroups and Group-
ings
(1) For a fixed permutation 1r, the map T H 1rT is a bijection between
the subgroup lattice of S(Z) and the set of grouping blocks containing
7f.
(2) Th e map T H { 1r T : 1r E P ( Z)} is a dual isomorphism between

the subgroup lattice of S(Z) and the collection of all groupings, ordered
by r-efinement.
(3) In the opposite dir-ection, th e map G H Sa is a surjection from
the set of grouping blocks onto the subgr-oup lattice of S(Z) , and two
grouping blocks are equivalent iff they have the same substitution group .
A slightly different point of view would be to consider, instead of

groupings, the associated equivalence relations 2 on P(Z): these
are characterized by the property that 1r 2 '1/J entails a1r 2 a'l/J for all
a E Sm. Such equivalence relations might be called left congruences on
P(Z) . By the above remarks, they are in one-to-one correspondence
with the subgroups of S(Z). If, moreover, 1rO 2 'lj;O for all substitu-
tions e E S(Z) whenever 1r 2 '1/J, then we could speak of a congruence.
It is then easy to check that the congruences precisely correspond to
the normal subgroups of S ( Z). On the other hand, the congruences
are associated with those groupings which have the property that one
passes from one block of the partition to another by one single sub-
stitution. And, as Galois discovered, that is just the condition that
reflects the splitting of a polynomial into factors under the adjunction
of all roots of an auxiliary equation. Or, in modern terms: under the
Galois connection between fields and groups, normal field extensions
correspond to normal subgroups.
By the above reasoning, there exists a perfect adjunction between
left congruences on P(Z) and subgroups of S(Z). And in fact, that
adjunction is induced by an axiality (see Section 1.4) between the
power sets of P(Z) x P(Z) and of S(Z), which in turn comes from a
very natural function between these sets, namely
F: P(Z) x P(Z) --t S(Z), (1r, '1/J) f--t 1r- 1 'lj;.
The axiality c--+F , F+---) sends any left congruence 2 ~ P(Z) X P(Z)
to the subgroup {1r- 1 'lj; : 1r 2 '1/J}, and in the reverse direction , any
subgroup T of S(Z) to the left congruence {(1r, '1/J): 1r- 1 'lj; E T} - which
happens to be a congruence just in case Twas a normal subgroup.
After having stated his "principles", Galois formulates his famous four
basic lemmas on which he founds his theory of the solvability of poly-
nomial equations. We summarize them in slightly modernized termi-
nology.
Galois' Four Basic Lemmas

I. Any algebraic element a over a field K is the root of a ((minimum
polynomial" ma,K that divides any other polynomial vanishing at a .
II. Given a finite subset Z and an infinite subset A of a field, there is
a linear form f with coefficients in A such that all values v obtained
38 M. Erne
by substituting elements of Z in that form are distinct.

III. Any such value is a primitive element of K ( Z): K ( Z) = K (v).
IV. Ifv and v' are conjugate (i.e. have the same minimum polynomial)
then g( v) = 0 {::} g( v') = 0 for any rational function g.
Notice that Lemma II may be shown by an elementary induction on
the size of Z , without using any higher algebraic tools. After Galois,
any rational function f with the property stated in Lemma II (ex-
cept linearity) or its value v obtained by substituting the roots of the
proposed equation is referred to as a Galois resolvent for the given
equation. Alternatively, some authors mean by a Galois resolvent the
minimum polynomial for v.
Then Galois proceeds to his famous Proposition I, building the basis
of Galois theory, which he formulates as follows:
Theorem
Let an equation be given whose roots are a, b, c, . . . There will always
be a group of permutations of the letters a, b, c, ... which will have the
following property:
1. that each function invariant under the substitutions of this group
will be known rationally;
2. conversely, that every function of the roots which can be determined
rationally will be invariant under these substitutions.
In a footnote, Galois explains what he means by "invariant under
these substitutions":
We call here invariant not only a function whose form is
invariant under mutual substitutions of the roots, but also
those whose numerical value is unchanged by these substi-
tutions.
Galois' "demonstration" consists of a few typical laconic statements.

Referring to Lemmas III and IV, he considers all conjugates v =
v 0 , . .. , Vn- l of some resolvent v. Then, choosing polynomials r.p = r.p 0 , ... , i.pm- 1
such that r.p 0 ( v) , ... , r.pm-1 ( v) are the roots of the original equation, he
forms the collection of all tuples (r.pi(vj) : i < m) (j < n) and con-
cludes:
I say that this group has the stated properties.
That the so-defined tuples actually are permutations of the roots is an

easy consequence of Lemma IV, but the second "group" property is not
so obvious. Most of the proofs given by later authors use the notion of
field automorphisms - undoubtedly an elegant tool of modern Galois
theory, but one that is not contained in Galois' approach. Again, a
rigorous proof was supplemented by Dedekind (see Section 2.4).
Galois emphasizes the fact that adjoining new quantities may "reduce"
the equation inasmuch as the polynomial in question may decompose
into smaller factors after the adjunction. Of course, in that context,
the word "adjunction" is to be understood in the field-theoretical and
not in the order-theoretical sense. Nevertheless, there is also an order-
theoretical adjunction behind these remarks of Galois, which may be
described as follows (using the later notion of field extensions).
The intermediate fields L between a field K and an extension E form
a closure system :F(E: K), hence a complete lattice with respect to
inclusion. If E is a simple algebraic extension K(a), we shall write
:F(a, K) instead of :F(E : K). For each divisor h of the minimum
polynomial f = ma,K in E[x], the field Aa,K(h) := K(h) obtained
by adjoining the coefficients of h to K is a member of :F(a, K) (note
that formally h is just the sequence of its coefficients). Let M (a, K)
designate the set of all monic divisors hoff in E [x] that are irreducible
over K (h) and vanish at a - in other words, the minimum polynomials
of a over extensions of K contained in E. Then we have a map Aa,K
from M(a, K) to :F(a, K). In the opposite direction, associate with
any L E :F(a, K) the minimum polynomial f2a,K(L) = m a,L of a over
L, to obtain a map f2a,K from :F(a, K) to M(a , K). We write g qhand
call g a strong divisor of h if g divides h in K(g)[x]. Strong divisibility
is easily seen to be a partial order on monic polynomials.
The Perfect Galois Connection for Minimum Polynomials
For any algebraic element a over K, the pair (Aa,K, f2 a,K) is a perfect
Galois connection between M (a , K) , the set of minimum polynomials
of a dividing ma,K, ordered by strong divisibility, and :F (a , K) , the
interval of all fi elds between K and K (a), ordered by inclusion. Hence,
M (a, K) and :F( a, K) are dually isomorphic finite lattices. The degree
functions
p: M(a,K) ----+wand {L: :F(a,K) ----+w , L H [K(a):L]
40 M. Erne
are scales with j), = J.L o (} and J.L = j), o A.

Proof. That A and (} form a Galois connection between the dually
ordered sets M (a, K) and :F( a, K) follows from Lemma I: for any
h E M(a, K) and L E :F(a, K), we have the equivalences
.\(h)= K(h) ~ L {:} hE L[x] and h(a) = 0 {:} f!(L) = ma,L ~ h .
Indeed, g = ma,L divides h in L[x] iff g divides h in K(g)[x], because
both are monic polynomials (cf. Gauss' Lemma) .
In order to see that A and (} are mutually inverse dual isomorphisms,
it suffices (by the Theorem on Perfect Galois Connections, Section
1.3) to prove the statements about J.L and jj,. These two functions are
related by
J.L(f!(L)) = J.L(ma,L) = [L(a): L] = [K(a): L] = jj,(L).
If L ~ L' and jj,(L) = jj,(L') then L = L', because the vector spaces L
and L' have the same dimension [L :K] = jj,(K)/P,(L) = jj,(K)/P,(L') =
[L': K], so that not only J.L but also j), = J.LO (} is a scale.
As any hE M(a, K) is monic, irreducible over K(h) and has the root
a, it must coincide with the minimum polynomial f!(.\(h)) = ma,K(h)·
Thus, f!OA is the identity map. It follows that J.L =fLO f!OA = jj,o.\ .
Since :F( a, K) is a lattice, so is M (a , K), and since the latter is finite ,
so is :F(a, K). D
From the fact that j), inverts the divisibility relation (by the degree
formula for fields) and (} is anti tone, it follows that J.L = j), o A is an
isotone scale. Thus, the degree of a strong divisor of a polynomial h
in M (a, K) is not only smaller than but a divisor of the degree of h
(cf. Lagrange's Theorem on Minimum Polynomials!)
An immediate consequence of the above theorem is the "only if" part
in the
Theorem on Primitive Elements
A field extension E : K is simple and algebraic (i .e. E = K (a) for
an algebraic element a over K) if and only if the interval :F(E : K)
is finite. For characteristic 0, these two conditions are equivalent to
finiteness of [E: K] .
For the "if" part (cf. [90, 98]), one first observes that the degree
[E : K] must be finite (otherwise, one would obtain an infinite chain
of intermediate fields). Then, one distinguishes the following cases:
if K is finite then so is E, being a finite-dimensional vector space

over K. Hence, any generating element of the cyclic group E* (see
the Theorem on Multiplicative Subgroups of a Field) may serve as a
"primitive element" generating E. If K is infinite then one proceeds
by induction and assumes in the induction step that E = K(b, c)
for algebraic elements b, c over K. If F(E : K) is finite, there exist
different elements k,k' E K such that K(b+kc) = K(b+k'c). Hence,
for a = b + kc, it follows that t he elements c = (a-b-k' c)(k-k') - 1
and b = a-kc belong to K(a), i.e. E = K(a). Third, if m c,K has no
multiple roots (which is for sure if K has characteristic 0), a different
argument works: consider k E K and assume mc,K(c') = 0 = mb,K(b')
for some c' f- c and b' = b + k (c- c') . Since there are only finitely
many such k = (b'-b)(c-c') - I, we may choose a k E K so that for
a = b + kc, the polynomials mc,K and mb,K(b + k(c - x)) E K(a)[.r]
have the greatest common divisor x- c in K(a)[x]. Again, we get
c E K(a) and b =a- kc E K(a), i.e. E = K(a).
2.4 Richard Dedekind's Lectures on Algebra

In einer Vorlesung als Privatdocent in Gottingen behandel-
ten Sie die damals noch ganz unzugiingliche Theorie von Ga-
lois. Aber bei allem verstiindnisvollen Eingehen auf fremde Art
wujJten Sie doch stets Ihr Selbst zu behaupten, dem Uberkomme-
nen Ihr personliches Gepriige aufzudrucken.
Laudatio of the Royal Academy at

Berlin on Richard Dedekind (1902)
The Lectures on Algebra alluded to in the quoted Laudat io on the

occasion of the 50th anniversary of Dedekind's Ph. D. t hesis proba-
bly contain the first elaborated version of Galois' theory, including a
systematic elementary theory of (abstract!) finite groups. Unfortu-
nately, the unpublished manuscript of these lectures, written some-
where between 1858 and 1868, had been lost for more than a hundred
years until its rediscovery by Purkert [116] and t he commented edition
by W. Scharlau in [121] on the occasion of t he 150th anniversary of
Dedekind's birthday (1981).
Content and style of the rediscovered manuscript demonstrate the
power of an ingenious mathematician whose striking ideas and de-
42 M. Erne
ductive accuracy put him far ahead of most , if not all of his contem-
poraries. Indeed, if some antiquated idioms and long-winded verbal
formulations in the text would be replaced by modern terms and for-
malisms, an almost up-to-date exposition of group and field theory
and other parts of algebra would arise. Unlike Lagrange, Dedekind
always endeavored to giving concise and elegant proofs, and on the
other hand, unlike Galois, he did not omit important arguments or
constructions. The reader of the present survey will already have
remarked that when one tries to make certain statements of those
forerunners precise, one has to refer to Dedekind's work again and
again.
In the preparatory Chapter I on groups, Dedekind deals not only
with permutations, but also with (finite) abstract groups, normal sub-
groups, and even with factor groups. Following Galois, he speaks of
substitutions rather than permutations and defines their product ac-
cording to the (rather modern) convention that the substitution ()()'
first applies () and then ()'. Then he proves the associative law and the
cancellation laws and remarks that these two "Fundamental Laws"
are the base for all further conclusions, so that they may be used for
an abstract definition of finite groups, occurring also in many other
mathematical contexts. (Dedekind is well avvare of the fact that the
finiteness hypothesis is essential for the desired conclusions.)
As a first highlight, he derives Lagrange's Subgroup Theorem and
some of its prominent consequences from the partition
G = H() + HO' + HO" + ...
of a group G into cosets of a subgroup H , mentioned already in Ga-
lois' letter to his friend Auguste Chevalier [57). Since all cosets have
the same cardinality asH, their number (the index of the subgroup)
multiplied by the order of H must give the order of the entire group
G. But Dedekind goes further , analyzes decompositions into two-sided
cosets () H ()' and shows, for example, that the intersection of all conju-
gates e- 1 H () is a normal subgroup ( "eigentlicher Divisor"). Perhaps
his most modern achievement in the area of group theory is the ob-
servation that the cosets of a normal subgroup again form a group
under the "complex product" - a conclusion that requires a degree of
abstraction entirely alien to other mathematicians of his time.
In Chapter II, Dedekind develops in a concise and rigorous way the
main pieces of Lagrange's theory of partially symmetric functions and

their behavior under substitutions. In particular, he proves the funda-
mental fact that any representative system for the cosets of the stabi-
lizer I(t) of a rational function t produces all different "forms" oft ob-
tainable by substitutions of the variables. After a careful proof of the
existence and uniqueness of the representation of symmetric polyno-
mials by means of elementary symmetric functions, he derives the Fun-
damental Theorem on Symmetric Functions, observing that if f and
h are polynomials with no common divisors so that f / h is symmetric,
then both f and h must be symmetric. But Dedekind does not leave it
at that but points out that this conclusion requires (among other tools)
Gauss' Lemma, and that one may circumvent that tool by the follow-
ing argument: in the rational function f /h = (! fi ... fk)/(hfi ... fk) ,
where the polynomials j 1 are the different results obtained from f by
permuting the variables in f , the numerator is symmetric by defini-
tion, and so is the denominator hfi ... fk = f fi ... fk · h/ f.
As mentioned earlier, the symmetric group S(X) of all bijections of
X= {x 1 , ... ,xm} acts on the polynomial ring R[X] = R[x 1 , ... ,xm] by
e(f)(xl, ... , Xm) = JB(xl , ... , Xm) = f(B( x!), ... , e(xm)) for() E S(X) .
In his proof of Lagrange's Theorem on the Stabilizers of Rational

Functions, Dedekind uses the explicit representation of
Lagrange's Interpolation Polynomials
For any m distinct elements t 1 , ... , tm and arbitrary (not necessarily
distinct) elements y 1 , .. . , Ym of a field K , there is a unique polynomial
q E K[x] with q(t1) = y1 for all i = 1, ..., m, namely
( ) "'m
q x =
YjP( x )
L.Jj = l (x -tj )P'(tj )'
h
w ere
p(x) = n ::l(x- ti) and p'(x ) = L:;l :~~~.
This applies to the following situation. Choose a set of representat ives

01 =id, ... ,Om for the "cosets" GO of G = I(t) in S(X) , so that S(X) is
the disjoint union of the cosets GB1 (j = 1, ... , m) . Then the elements
44 M . Erne
are pairwise distinct (because t (}i = t (}j implies G (}i = G (}j , hence
i = j).
On account of the inclusion I(t) ~ I(y) , the elements Yj = (}j(y) =
y(}j E K(X) satisfy the implication 8(ti) = lj :::;. 8(yi) = Yj for each
() E S(X). Indeed,
t()i()=t()j {::} ()i(J(Jj- 1 E I(t) :::;. ()i()()j- 1 E I(y) {::} y()i() = y()j·
Now, the interpolation polynomial q E K(X)[x] with q(tj) = Yj has
the required property q(t) = q(t 1 ) = y 1 = y. From the construction
it is rather plausible that q is symmetric, but an exact proof requires
some care. First, one observes that p(x) = f1~ 1 (x - ti) is symmetric,
because application of any () E S(X) merely permutes the elements ti
(cf. Lagrange's Theorem on Minimum Polynomials). Hence, ()(p) = p
and ()(p') = p'. As ()(ti) = lj implies ()(yi) = yj , one obtains 8(q)(x) =
""'m O(w)O(p)(x) _
L.Jj = l (x-O(f:i))p'(O(fj)) -
( )
q X .
0
In Chapter III, Dedekind supplements most of the missing proofs in

Galois' memoir, but he also makes clear that a great part of Galois
theory, even in the general case of "numerical roots" , may be developed
along the lines of Lagrange's theory. For a better understanding, let us
restate Galois' Fundamental Theorem more formally, using Dedekind's
concepts and terminology.
Fundamental Theorem on Fixpoint Fields
Let f (x) = 0 be a polynomial equation over K with no multiple roots.
Let Z = {r 1 , . . . , rm} be the set of these roots, s a Galois resolvent for
the equation (whence K(s) = K(Z) ), and s 0 = s, ... , sn-l its conju-
gates. Then:
(1) There are rational functions f i E K( x b ... , Xm) with ri = fi(s) .
(2) The maps ()j : ri r-t f i(sj) (j < n) form a subgroup G(f) of S(Z).
(3) An element of K(Z) is invariant under each ()j iff it belongs to K.
(4) The maps ()j extend to K-automorphisms of K(Z).
(5) The K-automorphism group of K(Z) is isomorphic to G(f).
(6) The group G(f) does not depend on the choice of the resolvent s.
(7) Th e groups G(f) and G(ms,K) are isomorphic.
All of these statements are proven in Dedekind's Lectures (however,
without explicit mention of the notions of fields and automorphisms,

and only for the case of characteristic zero). For a careful elaboration
of some details, see Edwards [39] and Tignol [138] . The group G(f) =
G(f, K) is referred to as the Galois group of the equation f(x) = 0
over K.
Dedekind's view of the situation is quite modern , emphasizing the
advantage of group actions on a set: his crucial observation in that
context is that the Galois group of an equation f(x) = 0 operates on
the corresponding "splitting field" K(Z). One has to assure that the
definition of the action of the Galois group on K(Z) = K(r 1 , ... , rm),
given by
u(} = g(r10 , ... , rmO) if u = g(r1, ... , rm)
does not depend on the choice of the rational function g (which by no
means is uniquely determined by u). To that aim, represent each ri in
the form fi(s) (with fi E K(x)) as a rational expression of the resol-
vent s. Now, if g(r 1, ... , rm) = h(r 1, ... , rm) then the rational function
(g- h)(ft(x) , ... , fr(x)) vanishes at s and consequently also at each
conjugate sO , where (}runs through the Galois group. In other words,
g(r10, ... , rmO) = h(r10, ... , rmO).
Using consequently the machinery of the Galois group action, Dedekind
extends Lagrange's Theorem on Stabilizers to the present setting of
"numerical roots". Denoting again by I(t) the stabilizer oft in the
Galois group G(f) , that is, the subgroup of all (} E G(f) with t(} = t,
one obtains by the same proof as in the case of rational functions:
Dedekind 's Theorem on Stabilizers
If s is a Galois resolvent for an equation f(x) = 0, then for any two
elements t, y of the field extension K ( s) one has the equivalence
I(t) ~ I(y) {::} K(s)(y) ~ K(s)(t).
We are now in a position to prove another variant of the Fundamental
Theorem of Galois Theory that might be rather close to Galois' own
ideas, namely an
Isomorphism Between Subgroups and Minimum Polynomials
Lets be a Galois resolvent for the polynomial equation f(x) = 0 over
K. Then the subgroup lattice g (!, K) of the Galois group G = G (!, K)
is isomorphic to the lattice M(s, K) of all minimum polynomials for
46 M. Erne
s. The following maps are mutually inverse isomorphisms, hence form

a perfect adjunction:
A: CJ(J, K)--+ M(s, K) , H H PH(x) = TioEH(x- sO) ,
(!: M(s, K)--+ Q(J, K), pH Hp ={BEG: p(sB) = 0} .
Proof. First, we must verify that both mappings are well-defined. In

order to see that for each polynomial p E M(s , K) the set HP is a
subgroup of G, consider substitutions B, (}' E G (and their extensions
to K-automorphisms) with p(sB) = p(sB') = 0. If we can prove
the equation p(sBB') = 0 then Hp is closed under composition and
therefore (by finiteness) a subgroup of G. There are rational functions
t and t' in K(x) with sO= t(s) , sO'= t'(s) and sOB'= t(t'(s)) . Hence
we have pot(s) = p(sB) = 0, and as pot belongs to K(p)(x) , it follows
that p = ms,K(p) divides the numerator of pot; now, p(sB') = 0 implies
p(sBB') = pot(sB') = 0, as desired.
For any subgroup H of G , it is clear that PH is a monic divisor of
PG = ms,K with PH(s) = 0, since H contains the identical substitution.
For each(}' E H, we compute
PHB'(x) = TioEH(x- sOB') = Tio EH(x- sO) = PH(x).
Now, consider any hE K(pH )[x] with h(s) = 0. As(}' fixes all elements
of K(pH), it follows that h(sB') = h(s)B' = 0 for each(}' E H , whence
PH divides h. This shows that PH is the minimum polynomial of s
over K(PH ), i.e. PH E M(s , K).
That the two mappings are mutually inverse is easily checked: on the
one hand, we have
PHp(x) = Tio EG:p(sO)=o(x- sO)= p(x)
for p E M(s, K); and on the other hand , for HE CJ(J, K) we get
HPH = {B' E G: Tio EH (sB'- sO)= 0} =H.
Finally, the equivalence
H ~ Hp {::} \f(} E H (p(sB) = 0) {::} PH qp
shows that we have a perfect adjunction. 0
One of the various interesting consequences of the last theorem is

Dedekind's observation that the degree of the minimum polynomial is
always equal to the order of the corresponding group.

Composing the previously derived adjunctions and Galois connections
between groups, minimum polynomials and fields, one obtains the
Fundamental Theorem of Galois Theory in the form of a perfect Ga-
lois connection between the subfields of a field extension and the sub-
groups of the corresponding automorphism group, as established by
Emil Artin(see [5] and Section 3.1). Let us connect the various Fun-
damental Theorems of Galois Theory in a diagram.
subgroups of the GALOIS minimum

Galois group 9(!, K)-- - -- - M(s, K) polynomials
DEDEKIND LAGRANGE
subgroups of the Q(s , K) ______ F(s , K) field extensions

automorphism group ARTIN contained in K (s)
2.5 Dedekind's Order- and Lattice-Theoretical

Concepts
Es steht alles schon bei Dedekind.
Emmy Noether
It is evident from our introductory remarks about the idea of adjunc-

tions and Galois connections that the fundamental ingredients in that
theory are
(1) order relations,
(2) isotone (order-preserving) maps,
(3) closure operations.
Therefore, it is certainly of interest to search for the first appearance
of these notions in the literature.
48 M. Erne
As Scharlau communicates in the memorial issue [121], he discov-

ered in the posthumous works of Dedekind an unpublished prelim-
inary draft (written probably quite some time before 1860) of the
Lectures on Algebra, which contains the following remarkable lattice-
theoretically inspired sentence (translated):
General Theorem Suppose S1 , S 2 etc. ar-e fields rce-
biete"} containing the coefficients of the equation f(x) and
let G 1 , G 2 etc. denote the cor-r-esponding gr-oups of the equa-
tion. Then their- gr-eatest common divisor- [ G 1 n G 2 } is the
gr-oup with r-espect to the system sls2 [the least common
multiple of S 1 and S2}.
The first structural investigation involving (at least implicitly) the
above three order-theoretical notions is probably Dedekind's famous
treatise on the axiomatics of natural numbers [29] "Was sind und
was sollen die Zahlen ?" (1887). Here Dedekind introduces "systems"
(something like naive sets) and mappings, and derives in a rigor-
ous way many elementary properties from a few definitions. Thus,
Dedekind's method may be regarded as the first successful attempt to
develop a partly axiomatic, partly "naive" set theory for the founda-
tions of mathematics.
Though Dedekind defines set inclusion by membership of elements,
he does not introduce an extra symbol for the latter, but uses the
same symbol -< both for membership and inclusion. Aside from that
little inconsistency, his development is so careful and precise that a
few minor changes of derived notions or facts into axioms would pro-
vide an axiomatic theor-y of or-der-ed sets, regarding -< merely as an
order relation rather than as the concrete containment relation. In-
deed, Dedekind's initial conclusions (which had to be more verbal than
formal, because his "systems" are not defined axiomatically) are:
A-< A (r-eflexivity),
A -< B -< A ::::} A = B ( antisymmetr-y),
A -< B -< C ::::} A -< C (tr-ansitivity).
Next, he derives the defining rules for the union, alias supr-emum
( 11Composed system") and inter-section, alias infimum ( Community"),
11
which in modern notation read as follows:

A -< B for all A in A {::} UA -< B,
B -< A for all A in A {:} B -< nA.

Moreover, he observes that if for each A E A there is a B E B with
A ~ B then UA ~ UB, and dually for intersections. Of course, in
the abstract order-theoretical setting, one has to interpret the union
and intersection symbols as joins and meets.
Then Dedekind explains what he means by a mapping, its image, the
restriction to subsets, etc. For a given mapping <p he denotes, as usual
still today, by rp (a) the image of an element a, and by cp( A) the image of
set A, consisting of all cp(a) with a E A. In order to avoid ambiguities,
we employ a slightly different symbol and write rp(A) for the latter. At
a few points, the abstraction of Dedekind's theory from set inclusion
to arbitrary order relations requires some care, because not all results
may be transferred without additional hypotheses. The most impor-
tant of these hypotheses is the preservation of unions, which is clear
for the "lifted" maps between power sets, carrying subsets of the do-
main to their images, while for maps <p between arbitrary ordered sets,
one would merely require that they be isotone (i.e. order preserving) :
A -< B ::::? rp(A) -< cp(B),

but not always that <p should preserve arbitrary joins, which for maps
between complete lattices is equivalent to saying that <p is a left ad-
joint.
After having established the associative law both for the composition
of mappings and for the binary union operation (but not for binary in-
tersection), Dedekind considers one-to-one mapppings, which he calls
"iihnlich" (resembling) or "deutlich" (distinctive) , and derives the ba-
sic properties of such mappings and their inverses.
The crucial notions for his set-theoretical introduction of the natural
numbers are then the "self-mappings" (mappings from a set into itself)
and what he calls "Ketten" (chains) , namely, systems K with cp(K) -<
K. In the general setting, this nomenclature is not very suggestive, but
of course, what Dedekind had in mind are chains of natural numbers,
closed under the successor function. Perhaps, he also thought of the
descending chain
... rpn(K) -< ... -< rp(K) -< K.
Let us distill the order-theoretical essence: given a partial order -< and
50 M. Erne
an isotone self-map rp on a complete lattice C, call an element K of C

(rp-) closed if rp( K) -< K; as Dedekind proves, the system Ccp of all such
elements is closed under arbitrary meets, hence a closure system in the
set-theoretical case (cf. Section 1.2). Dedekind's dual conclusion that
this system will also be closed under unions depends on the hypothesis
that the power set map rp comes from a map on the ground set and
therefore preserves unions. Order-theoretically, it is easy to show that
if a map rp preserves arbitrary, finite or directed joins etc., then the set
of all rp-closed elements is closed under formation of the corresponding
joins.
Dedekind denotes the least member K of Ccp that satisfies A -< K by A 0
or rp 0 (A). Thus, he introduces here (without using that terminology)
the closure A 0 of A and the closure operator rp 0 associated with Ccp (or
induced by rp). Dedekind's crucial discovery is that a closure operator
rp 0 enjoys and is determined by the following properties:
A -< rp 0 (A) E Ccp and A-< K <=? rp 0 (A) -< K forK E Ccp,
which say that the corestricted closure map rp 0 : C---+ Ccp is left adjoint
to the inclusion map '1/J : Ccp Y C. Dedekind also establishes two of the
three characteristic properties of closure operators, viz. extensivity
and isotonicity, but not explicitly the third, idempotency. Instead,
he derives some related formulas and the shortest of all conditions
characterizing closure operators:
B-< 'Po(A) <=? <po(B) -< rpo(A) .
It is a historical curiosity that the latter characterization was re-
invented several times by various authors until the seventies of the
20th century.
Fundamental for the classical Galois theory are finite permutations, in

particular the fact that an injective or surjective self-map of a finite
set is already bijective. Apparently, that coincidence was assumed
tacitly by Galois and his contemporaries; as a consequence, through-
out the whole 19th century, the bijectivity of "substitutions" was not
postulated explicitly.
Again, it was an ingenious mental act of Dedekind to realize that a
proof of such seemingly "obvious" facts requires an axiomatic defi-
nition of finiteness, not only an intuitive comprehension of the word
"finite". And even bolder was his decision to declare finiteness of a set
S by the postulate that there be no injective self-map from S onto a
proper subset:
Ein System S heiflt unendlich, wenn es einem seiner echten

Teile iihnlich ist; im entgegengesetzten Falle heiflt S ein
endliches System.
In other words, Dedekind characterizes finiteness of S by the property

that all injective self-maps of S are surjective! But he also observes
that, in order to derive theorems about finite sets via induction, one
often needs another definition , namely the existence of a bijection onto
an initial segment of the natural numbers. Dedekind proves the equiv-
alence of the two definitions, making use of a weak choice principle,
later called the Axiom of (Countable) Dependent Choices. Today, we
know that in set theories without choice principles, the second def-
inition of finiteness is strictly stronger than the first one, and also
that the existence of an infinite set is not provable on the base of
the other elementary set-theoretical axioms. Curiously, Dedekind's
"proof" of the existence of an infinite set strongly contrasts with the
formal proofs for all other statements he formulates before and after.
At this point, Dedekind uses philosophical rather than mathematical
phrases and claims that (translated)
the world S of thoughts, that is, all entities that may be
the object of reasoning, is infinite; because if s is such an
element of S then the thought that s can be the object of
my reasoning, is again an element of S ...
and so on. There exist some hints in the literature that in later years,
Dedekind himself was not satisfied with that sort of argumentation.
An interesting question in that context is this: can finiteness of a set
S also be defined equivalently by the postulate that every surjective
self-map of S be injective?
That the answer is in the affirmative is well-known, but probably not
the following simple proof using a suitable adjunction. It requires,
however, the hypothesis that the power set PS of a finite set is still
finite (which is easily verified by induction if the second definition
of finiteness is accepted). The argument is then as follows . A map
52 M. Erne
rp : S---+ S is obviously injective or surjective, respectively, if and only

if so is the power set map
-+rp: PS---+ PS, X H {yES I :3x EX (rp(x) = y)}.
But the latter is left adjoint to the inverse image map
rp+-: PS ---tPS, Y H {xES I :3y E Y (rp (x) = y)}.
And we know that injectivity of one partner of the adjunction is equiv-
alent to surjectivity of the other. This shows that the two statements
• every injective self-map of a finite set is surjective

• every surjective self-map of a finite set is injective
are actually equivalent .
Having discussed a few implicit occurrences of adjunctions and Galois

connections in the early work of Dedekind, we should now mention
briefly his pioneering achievements in lattice theory - which certainly
is an indispensible tool for the general theory of Galois connections.
The two papers to be cited here are those from 1897 and 1900, respec-
tively,
Uber Zerlegung von Zahlen durch ihre groflten gemeinsamen Teiler
{On decompositions of numbers by their greatest common divisors} [30]
and
Uber die von drei M oduln erzeugte Dualgruppe
{On the lattice generated by three modules) (31].
Much of the material in these two articles has been rediscovered and
presented by Birkhoff in his fundamental monograph Lattice Theory
[3]. We find it opportune to confirm here Noether's "headline" quo-
tation by pointing out that , if not "everything", at least a lot of nice
results (and their proofs) of lattice theory can already be found in the
two papers by Dedekind.
The theme in the title of the first paper is a rather specific number-
theoretical one; probably, it did not receive much resonance among
algebraists. But that paper contains the two basic concepts of formal
lattice theory, namely Dedekind's "Dualgruppen" and what is nowa-
days called a lattice-ordered group (see Section 4.2). Dedekind accu-
rately mentions Schroder's work (see the next section) and appreciates
his priority in some of the relevant discoveries; but he also makes clear
that his own emphasis is on the applications to ideal and number the-
ory rather than logic and set theory ( "Systemlehre "), and that his
method provides considerable simplification of Schroder's results and
proofs. Indeed, Dedekind's definitions and deductions are of striking
elegance and precision, though he rarely used order-theoretical argu-
ments and completely avoided representations by diagrams (perhaps
a "forbidden tool" at that time).
We adapt Dedekind's terminology to the modern common use of lattice-
theoretical symbols and write V instead of+ for the join and 1\ instead
of - for the meet. Dedekind's "Fundamental Laws" are (1) the two
commutative laws, (2) the two associative laws, and (3) the absorption
laws
(3') a V (a 1\ b) = a and (3") a 1\ (a V b) = a.
Thus, Dedekind's "Dualgruppen" are just what later have been bap-
tized "lattices" by Birkhoff (Ore called them "structures"). At the
end of the 19th century, the later meaning of "groups" was not yet
standard, and some mathematicians used that term for every (or at
least every finite) semigroup. At that time, the name "Dualgruppen"
was chosen quite reasonably, pointing to the two commutative semi-
group structures and their duality (in the sense that every true lattice
identity entails its dual, obtained by exchanging the two fundamental
operations). Unlike many other authors, Dedekind does not include
(as stated by Birkhoff) the idempotency laws
(4') a V a = a and (4") a 1\ a = a
in the list of axioms, but he derives them as a first step from the
absorption laws in the shortest possible way. Then he emphasizes
that the distributive laws
(5') (a 1\ b) V (a 1\ c) =a 1\ (b V c)
(5") (a V b) 1\ (a V c) = a V (b 1\ c)
(characterizing his "Dualgruppen vom I dealtypus ") imply each other
and the modular laws (characterizing his aDualgruppen vom Modulty-
pus")
(6') (a 1\ b) V (a 1\ c) = a 1\ (b V (a 1\ c))
(6") (a V b) 1\ (a V c)= a V (b 1\ (a V c)).
54 M. Erne
Moreover, he shows that either of these two laws is equivalent to the

self-dual identity
(M) (a V (b !\c))!\ (b V c)= (a!\ (b V c)) V (b !\c) .
In order to avoid misinterpretations, it should be mentioned that
Dedekind has systems of fractional ideals (of a number field) in mind
when he speaks of the "Idealtypus", while the lattice of all ideals of
a ring in general fails to satisfy the distributive laws, though being
always modular. Next, Dedekind describes, by means of 5 x 5 Cayley
tables, the smallest non-modular lattice (now known as the pentagon)
and the smallest modular but non-distributive lattice (the diamond) ,
which he concedes "took him some effort". And Dedekind also was
the first to show (in the second paper [31]) that modular lattices are
characterized by the exclusion of pentagon sublattices; the analogous
and deeper characterization of distributive lattices by exclusion of pen-
tagon and diamond is due to Birkhoff [3] . But the equivalence of the
distributive laws to the self-dual "median law "
(a V b)!\ (b V c)!\ (b V c) = (a!\ b) V (b !\c) V (b !\c)
and its proof are already contained in Dedekind's second paper, too.
The central subject of his investigation is the free modular lattice
with three generators, which Dedekind describes by a Cayley table
covering two pages. Since this 28-element lattice occurs in many later
investigations of other authors (for example, in the paper [150] by
Ward and Dilworth to be discussed in Section 4.3) , we sketch here a
diagram whose points carry the original notation used by Dedekind.
a c
That Dedekind rarely used order-theoretical arguments does not mean

he would not have known the order characterization of lattices. In the
first paper, he gave a purely set-theoretical description of lattices in
terms of their principal (dual) ideals, hence implicitly in terms of an
order relation he called "divisibility ", denoted by > and defined by
a > b {::} a V b = b {::} a 1\ b = a.
(Dedekind wrote > instead of ~ because he had the interpretation
of divisibility of ideals in mind, which is dual to set inclusion.) In
the second paper, he explicitly stated the three properties character-
izing partial orders (reflexivity, antisymmetry and transitivity) , and
observed that the operation V is the "greatest common divisor " with
respect to divisibility (the join with respect to the dual relation) , while
1\ is the "least common multiple" (the meet for the dual relation).
Since a map between complete lattices preserves joins iff it is left

adjoint, it is evident that (finite or infinite) distributive laws are closely
related to adjunctions for unary operations. But also modularity can
be described in terms of certain adjunctions, as we shall see below from
a fundamental characterization of modularity, again due to Dedekind.
Two elements a, b of a lattice L form a modular· pair, written aM b, if
(x 1\ b) V (a 1\ b) = ( ( x 1\ b) V a) 1\ b for all x E L ,
or equivalently, x ~ b implies x V (a 1\ b) = (x V a) 1\ b; if the dual
condition holds, one writes aM*b (cf. [3]). The interval [a, b] consists
of all elements x with a ~ x ~ b. Now a straightforward verification
yields the so-called
Transposition Principle for Intervals
For any two elements a, b of a lattice L , the maps
'1/Ja : [a 1\ b, bJ ---+ [a, a V bJ, x H x V a and
'Pb : [a, a V b] ---+ [a 1\ b, bJ, y H y 1\ b
form an adjoint pair-. Furth ermore,
aMb {::} '1/Ja is injective {::} 'Pb is surjective,
aM*b {::} 'Pb is injective {::} '1/Ja is sur-jective .
Hence, L is modular-iff th e adjunctions ('1/Ja, 'Pb) are always (left , right)
perfect, in other words, '1/Ja and 'Pb are always mutually inverse isomor-
phisms.
56 M . Erne
For finite modular lattices, Dedekind concludes that all maximal chains
have the same length, and that the degree ( "Stufe ") or height func-
tion a, associating with any element a the length of a maximal chain
between the least element and a, satisfies the identity
a( a)+ a(b) =a( a A b)+ a( a V b).
Of course, if a and bare related and a( a) = a(b) then a= b, so that a
is in fact a scale in the sense of Section 1.3. Conversely, the existence
of a function with these two properties forces modularity (cf. Birkhoff
[3] and von Neumann [145]). Indeed, in a pentagon with atoms a, b
and a coatom c above a,
a( a)+ a( b) = a(a A b)+ a( a V b) = a(cA b)+ a(c V b) =a(c)+ a( b),
hence a(a) = a(c), while a < c.
A further adjunction, not contained explicitly in Dedekind's work but
closely related to the above one, concerns
Distributivity and Product Representation of Lattices
For any two elements a, b of a bounded lattice L, the maps
A: (0, a] X (0, b]---+ L, (u, v) r---t u Vv and Q: L ---+[0, a] X (0, b], X r---t
(aAx,bAx)
form an adjoint pair. Furthermore, (-X, Q) is a perfect adjunction iff
(a, b) is a bi-modular, complementary and distributive pair, i.e.
aMbMa, aAb=O and x=(aAx)v(bAx) forall xEL.
Every direct product representation L c:= £ 1 x L 2 is, up to isomorphism,
of the above form. In particular, any complementary distributive pair
of elements in a modular lattice gives rise to a direct product repre-
sentation, and conversely.
This result is of importance, for example, in the famous decomposition
and coordinatization theory for complemented modular lattices due to
Garrett Birkhoff and John von Neumann (see [3] and [145] for details) .
2.6 Ernst Schroder's Algorithms and Calculi

Schroder's work may be seen as the ultimate fu~filment of the
endeavour of logicians over half a century to evolve an adequate
treatment of the formal side of logic.
G. T. Kneebone
Ernst Schroder probably was not only the inventor of the first semi-
formal system of lattice theory (perhaps even before Dedekind), but
also the first mathematician to investigate the interplay between logic
and (a restricted form of) universal algebra, involving certain Galois
connections. Although he did not use that terminology nor the precise
mathematical definition of such connections, they occur implicitly in
at least three problem areas discussed extensively by Schroder in his
monographs, articles and lectures [127] - [131]:
(1) The role of the distributive law and its (non-)provability in his
''logical calculus", the first axiomatic approach to lattice theory.
(2) The foundation of (parts of) universal algebra and its links to
lattice theory, exemplified by the equational theory of quasi-
groups.
(3) The idea of extents and intents of concepts, leading a century

later to the formal concept analysis initiated by R. Wille [59,
153].
In honor of Schroder's pioneering work in the area between mathe-

matics, logic and philosophy, a modern institution founded by Wille,
dealing both with philosophical and mathematical aspects of formal
concept analysis, its applications and related theories, was baptized
Ernst-Schroder-Zentrurn fur· Begriffiiche Wissensverarbeitung.
Schroder's early investigations of certain systems of axioms and pos-
tulates building the framework for the later theory of lattices will
be discussed more thoroughly in Section 4.1. May it suffice for the
moment to point out that, in his own later work, Schroder carefully
distinguished between the "identical calculus" (or "calculus of iden-
tities"), the earliest manifestation of a systematic theory of Boolean
logic and algebra, and the "logical calculus", which, in the terminology
58 M . Erne
of the 20th century, is lattice theory combined with certain semantic

interpretations by classes, statements, conclusions, etc.
Schroder was quite proud of having found a counterexample that dis-
proves the general validity of the distributive law in the "logical cal-
culus". For his debate with Peirce about that point, see Section 4.1.)
The counterexample alluded to, though outdated soon by Dedekind's
much simpler construction (see the previous section), opened a pro-
found insight into the axiomatics of lattice theory. But it is also of
interest with regard to the development of Galois connections, because
it stemmed from Schroder's theory of "algorithms", perhaps the first
historical instance of an equational theory and its models, and there-
fore of a class-theoretical Galois connection induced by the relation of
satisfaction. He considers a special type of "functional equations" for a
multiplicatively written operation, together with two other operations
": " and "/" so that
ab =c {::} c = a : b {::} b = a/ c.
Thus, he deals with nowadays so-called quasigroups. Observe that the
above rules just define residuated groupoids (see Chapter 4) with the
identity relation as order. Schroder deduces the "principles"
b = a(b: a)= (bja)a = aj(a: b)= a : (a/b)= (ba)ja = (ab) : a
and considers the finite set U of all identities derivable from these prin-
ciples and containing exactly three - possibly equal - variables on both
sides. By a "group of formulas" or an "algorithm" he means "a system
of formulas in U that does not imply any formula in U (by virtue of
the principles) that it does not already contain. " In other words, using
modern terminology, an "algorithm" is a deductively closed system or
an equational theory of quasigroups. Schroder points to the difference
between the [closure] system of all algorithms as a model for his "log-
ical calculus" and the system of all subsets of U (which is a model for
Boolean algebra, his "identical calculus" ).
By virtue of the Galois connection induced by the relation of satis-
faction, M f= E, meaning that M is a model for E (or satisfies the
equation E), the model classes for sets of equations always form a clo-
sure system and consequently a (complete) lattice. This is essentially
what Schroder has in mind when he considers his "groups of formulas"
as the "members of a logical calculus", while the notion of "closure"
seems not to occur explicitly in that work. Again, it becomes evident

that the word "group" had a very vague meaning at the end of the
19th century. Later, Schroder wrote in his Lectures on the Algebra of
Logic [131, p.630] (free English translation):
Given a system of things, called "elements", and a process

deriving new elements from the given ones, we may add
those new elements to the system ... {and after iteration
of that process} the given elements together with the de-
rived ones form a system, which we may call the "group"
deter-mined by the former elements.
Thus, Schroder's "groups" essentially are systems closed under certain

processes, hence closure systems or, in the more specific realm of alge-
braic operations, systems of subalgebras. The fundamental idea that
extending the system of algorithms, axioms or rules decreases the sys-
tem of "groups" and vice versa, inherent in Schroder's work, opened
the way to the class-theoretical Galois connections between equational
theories and varieties of modern universal algebra.
Whereas the logical product of "algorithms" (or "groups" ) is simply
their intersection, the logical sum is not the union but the collection
of all formulas derivable from those in their union by virtue of the
'principles'. In other words, the sum is the Galois closure of the union
with respect to satisfaction.
Schroder disproves the distributive law by what he calls "exemplifica-
tion". The counterexample he gives (together with a Venn diagram)
is a proper inclusion
AB + CB c (A+C)B
where A and C are the algorithms of associative and commutative
quasigroups, respectively, and B is a certain algorithm consisting of
consequences of the identity ab = a : b = afb. Admittedly, as Hunt-
ington said, that example is rather complicated (e.g. compared with
that given by Dedekind) , but for Schroder it was both natural and
substantial in view of his elaborated theory of quasigroups. In a later
edition, he cited three further counterexamples due to other authors,
two of which are of interest from the perspective of later developments
of lattice theory and its connections with geometry.
60 M . Erne
One of these examples, due to Voigt [144], may be regarded as a typical

"context" (as considered in formal concept analysis, see Section 1.5) ,
hence as an instance of a Galois connection. The "symmetry relation"
between quadrangles and orthogonal maps rise to a non-distributive
(and not even modular) "concept lattice", consisting of seven concepts:
square
D
rectrglf/ ~bus
trapezoid [ ~llelogt.m I kite
r\,"'~jJ
t>
quadrangle
However, Voigt and Schroder did not consider that whole lattice, but
argued that the symmetry group of the square, which is the concep-
tual synthesis of the rectangle and the rhombus, contains not only the
axial symmetries of these two "constituents" , but further symmetries
(obtained by rotation). Hence, the symmetry groups of the square, the
rectangle, the rhombus and the parallelogram together with the group
of rotations form a non-distributive (but modular) lattice. This "di-
amond" lattice may be regarded as a sublattice of the closure system
of all subgroups of the dihedral group D 4 of order 4 x 2 = 8 (which are
represented by the incidence-preserving permutations of the vertices of
a square). Notice that the cyclic group generated by the permutation
(1234) cannot be represented as a symmetry group of quadrangles,
because a quadrangle fixed by that permutation must alredy be a
square and has therefore the full symmetry group D 4 . But the whole
subgroup lattice contains a lot of further 5-element non-distributive
sublattices, among them also non-modular ones.
The other example is due to Korselt [79] . The elements are the points,
lines and planes of 3-dimensional euclidean space, together with the
empty set and the whole space - in other words, the affine subspaces of
IR3 . Since these objects are closed under arbitrary intersections, they
form a complete lattice, ordered by containment. This closure system
consists of all Galois-closed sets with respect to the incidence relation
between points and planes. Of course, the join of two affine subspaces
is the least affine subspace containing both, and this is almost never
their union. A plane, two parallel lines contained in it, a point on
one of the lines, and the empty space form a pentagon sublattice, so
that the whole lattice cannot even be modular (though being upper
semimodular). Later, Korselt modified his example to a projective
geometry [80, 81]. Thus, he made a first step towards the lattice-
theoretical development of (projective) geometry.
Let us conclude this section with a quotation (in free English trans-
lation) from Schroder's "Operationskreis". It concerns the "concept
of concept", fundamental for the philosophical foundation of formal
concept analysis (see Section 1.5):
The linguistic expression of a class usually is a common

name and gives rise to the foundation of a concept, in
which the essential attributes shared by all individuals be-
longing to the class are thought of being comprised. In
contrast to those attributes, the so-called "intent" of the
62 M. Erne
mentioned concept, the class itself represents its "extent ",

so that computing with class symbols actually means to deal
with the concepts represented by their extent.
Since the passage from extents to intents turns the whole concept
lattice "up-side down", this approach to lattice theory makes the du-
ality principle particularly evident. And it is the merit of Schroder
to have pointed out that duality, which was missing in Boole's earlier
development (Section 4.1).
2. 7 David Hilbert's Nullstellensatz

Sie haben da ein unermeflliches Feld mathematischer Forschung
erschlossen, welches als 'Mathematik der Axiome' bezeichnet
werden konnte und weit iiber das Gebiet der Geometrie hin-
ausreicht.
Hurwitz in a letter to Hilbert, 1903
Hilbert's axiomatization of geometry [74] is closely related , via inci-

dence relations, with certain Galois connections that also played a role
in the later lattice-theoretical developments of geometry by Menger
[94, 95, 96], von Neumann [145], Birkhoff [12, 14] and others. We
shall address here a related fundamental discovery due to Hilbert that
is based on a very natural Galois connection, namely his famous "Null-
stellensatz".
One of the most frequently occurring types of relations leading to
interesting Galois connections is the "zer-o relation" or "vanishing re-
lation"
f _L z {::} f( z ) = 0
between functions f from a fixed structured set S into a structured set
T having a "zero element" 0, and elements z of S. For example, one
may consider a topological space S and the ring C (S) of all continuous
functions from S into the reals. The interplay between topological
and algebraic properties in that context proved extremely fruitful , as
documented, for example, in the book Rings of Continuous Functions
by Gillman and Jerison [63]. Let us mention only one aspect of Galois
connections in that context. The partners of the Galois connection
(, '11) associated with the above zero relation l_ send any subset Y
of the space S to the ideal (Y) = Y1_ of all continuous functions
vanishing on Y, and in the opposite direction to each set F of functions
in C(S) its zero set
'li(F) = p.l = {z E S: f( z ) = 0 for all f E F}.
By continuity of the maps, each zero set is closed; but when are all
closed subsets of the space S Galois closed, i.e. zero sets? Necessary
and sufficient for that coincidence is that for any closed set Y and any
point z outside Y, there be a continuous real function vanishing on Y
but not at z. In other words, the space must be completely regular.
Thus, the aspect of Galois connections makes it evident why com-
pletely regular spaces play a central role in the theory of continuous
real functions.
Algebraic versions are obtained by considering endomorphism rings or
polynomial rings instead of rings of continuous functions . A rather ele-
mentary but useful tool of linear and geometric algebra is the following
result, obtained from the Dual Isomorphism Theorem for Submodules
(see Section 1.4) by passing from endomorphism rings to matrix rings
via the selection of bases:
Galois Connection Between Ideals and Subspaces
Let A denote the endomorphism ring of a finitely generated vector
space B. The polarity associated with the vanishing relation between A
and B induces a dual isomorphism between the lattice of all (principal)
left ideals of A and the lattice of all subspaces of B.
Now, let us turn to the central theme of this section and consider an
algebraic closure C of a field K. For the ring K[x] of polynomials
in one variable over K, the situation is easy to describe. For each
subset Y of C, the polynomials having all elements of Y as roots form
a (principal) ideal. Hence, there is a unique monic fy E K[x] dividing
precisely those polynomials which vanish on Y. If Y is finite then fy
is the product of the distinct minimum polynomials of elements in Y;
otherwise, fy is the zero polynomial. Denoting by Z 1 the zero set of
any f E K[x], we have the equivalences
fy divides f {::} f(y) = 0 for all y E Y {::} Y ~ ZJ ,
showing that the maps
64 M. Erne
 : P(C)---+ K[x], Y H fy and \ll: K[x]-+ P(C), f H Zf

form an adjoint pair, where K[x] is ordered by divisibility and the
power set P (C) by inclusion. A subset of C is closed under conjugation
iff it contains with c E C all other roots of the minimal polynomial of c
over K. The above adjunction then induces an isomorphism between
the closure system of all subsets of C that are either finite and closed
under conjugation or equal to C, and the lattice of all monic squarefree
polynomials over K.
Focussing (like Galois) on one particular equation f(x) = 0, where
f is separable (has no multiple roots) , one obtains an isomorphism
between the closure system of conjugation closed subsets of z 1 and
the finite lattice of all monic divisors h of f that are irreducible over
K(h).
More interesting and of fundamental importance for algebraic geom-
etry is the case of a polynomial ring Rn = K[x 1 , ... , xn] in several
variables. Given a subset F of Rn , denote by V(F) its common zero
set, i.e.
V(F) = F j_ = {c = (c1, ...,en) E en: f(c) = 0 for all f E F}.
In algebraic geometry, such zero sets are referred to as (algebraic) K-
varieties. The radical of an ideal I in a commutative ring R with unit
is the set
Vi= {fER: fm E I for some m}.
A radical ideal is an ideal I that coincides with its own radical, which
is tantamount to saying that P E I implies f E I .
There is a beautiful Galois connection relating these two notions and so
building an important bridge between algebra and geometry. First of
all, note that V(F) coincides with V(I), where I is the ideal generated
by F. The equations V(0) = en and V(F) u V(G) = V(FG) show
that the closure system of all K-varieties is closed under finite unions ,
hence topological. The corresponding system of open sets is called the
Zariski topology on en .
Isomorphism Theorem for Algebraic Varieties
Let C be an extension of the field K and Rn = K[x 1 , ... , Xn] the poly-
nomial ring in n indeterminates over K . The following statements
are equivalent:
(a) For each n, the Galois connection associated with the zero rela-
tion between Rn and eninduces a dual isomorphism between the
closure system of K -varieties in en
and that of radical ideals of
Rn .
(b) For each n and F ~ Rn, the zero set V(F) is nonempty if (and
only if) F generates a proper ideal.
(c) For each n , the Galois closure of any ideal I of Rn is its radical
Vi.
(d) For each n and all ideals I , J ~ Rn, V(I) ~ V(J) implies VJ ~
Vi.
(e) For each n and F, G ~ Rn
with V(F) ~ V(G) there is an r such
that each product of r factors in G belongs to the ideal generated
by F.
Proof. (a) :::} (b). Let I= (F) denote the ideal generated by F . If
V(F) = V(I) = V( Vi) = 0 then by the dual isomorphism, Vi= Rn ,
i.e. 1 E I = fln .
(b) :::} (c). Let I be any ideal of Rn. If fm E I then clearly f van-
ishes on V(I), which means that f belongs to the Galo~s closure of I.
Conversely, if the latter is the case, consider the ideal I generated by
I[x] U {1 - xf} in the polynomial ring R = Rn[x]. If 1 rf. i then by
(b) , there would exist a (y , z) E en X e ~ en-t-1 with g(y , z) = 0 for all
g E i , in particular g(y) = 0 for all g E I and 1- z f (y) = 0, contradict-
ing the hypothesis that f belongs to the Galois closure of I (and there-
fore f(y) = 0) . Hence, 1 E i, say 1 = g+ (1- xf)h for some g E I[x]
and h E Rn[x]. Iff is not the zero polynomial, it is invertible in the
quotient field of fln. Substituting 1/ f for x, one obtains 1 = g(1/ f) ,
and for the degree m of g, it follows that fm E I. (This idea is due to
Rabinovich [117]) .
(c):::} (e). Since V( G) = V( (G)) and (G) is finitely generated by
Hilbert 's Basis Theorem (see [72], [90] or [142]), one may assume that
G is finite, say #G = k. Now, V(F) ~ V(G) implies G ~ y'(F) by (c).
Hence, by finiteness of G, there is an exponent s such that g 8 E (F)
for all g E G, and for r = sk it follows that g1 •.• gr E (F) whenever
g1 , .. . , 9r E G (because one factor must occur at least s times).
The implications (e):::} (d) and (d):::} (a) are straightforward. D
66 M. Erne
The radical ideals are precisely the intersections of prime ideals, and
the latter correspond to the (union-)irreducible K -varieties via the
dual isomorphism induced by the zero relation.
That the situation described above frequently occurs in practice (for
example, when Cis the field of complex numbers) is assured by
Hilbert's Nullstellensatz (Zero Set Theorem)
The equivalent conditions in the previous theorem are fulfilled if C is
algebraically closed.
Various rather short proofs of that theorem are to be found in the
modern standard literature on algebra and algebraic geometry (see,
for example, [84] or [90]) . Historically, it has to be noted that in his
famous 1893 paper "Uber die vollen Invariantensysteme" [73], Hilbert
formulated the Nullstellensatz in a slighly different form, namely as
condition (e) (which often is lost in modern presentations) for ho-
mogeneous polynomials and the complex number field C. Actually,
Hilbert's primary purpose was the application to the theory of invari-
ants. Some of his concepts come close to the idea of Galois closure,
when he considers for a given set of (homogeneous) polynomials the
collection of all polynomials vanishing on the common zero set of the
given ones. However, the notion of Galois connections does not arise
explicitly in that context.
For finite fields, the situation becomes particularly simple.
Theorem on the Polynomial Completeness of Finite Fields
For a field K, the following conditions are equivalent:
(a) K is finite (a Galois field).
(b) For any a E Kn ther·e is a polynomial fa E K[x 1 , ... , Xn ] with
fa(a) = 1 and fa(b) = 0 forb E Kn \{a}.
(c) Any function g : Kn --+ K is a polynomial function.
(d) Each subset of Kn is the zero set of a polynomial.
Hence, for any .finite field K , the Galois connection associated with the
zero r·elation between K[x 1 , .. . , Xn ] and Kn is right perfect. In other
words, there is a dual isomoTphism between the poweT set of Kn and
the closure system of all radical ideals of K[x 1 , . .. , xn], under which
the singletons correspond to the prime ideals (by associating with any
point a E Kn the ideal of all polynomials vanishing at a).
For (a) ::::} (b), use the modified interpolation polynomials

fa(xl, ... , Xn) = TI7=l TicEK\{aj} ~~=~ for a= (a1, ...,an) E Kn .
For (b)::::} (c), take the polynomial f = LaEKn g( a) fa to obtain f (b) =
g(b) for all bE Kn. The implications (c)::::}(d)::::}(a) are clear.
A modern algebraic variant of the correspondence between varieties

and radical ideals should be mentioned at the end of this section,
because it involves another natural Galois connection. Let R be an
arbitrary commutative ring with 1 and S(R) its spectrum, the set of all
prime ideals, equipped with the hull-kernel topology (also sometimes
referred to as the Zariski topology) . The closed sets for this topology
are the collections
ci>(B) = { p E S(R) IB ~ P} (B ~ R)
and they are in fact the Galois-closed subsets of S(R) with respect to
the membership relation E between elements and prime ideals of R .
On the other hand, the Galois-closed subsets of Rare the intersections
of prime ideals, and these are known to be precisely the radical ideals.
(For the latter conclusion, one needs Krull's Separation Lemma, see
Section 4.2). Thus, we have the
Isomorphism Theorem for Radical Ideals
The polarity associated with the membership relation between elements
and prime ideals of a commutative ring R gives rise to an isomorphism
between the closure system of radical ideals and the hull-kernel topology
on the spectrum. Under that isomorphism, the prime ideals correspond
to the irreducible closed sets.
The similarity between the last theorem and that on algebraic varieties
is not casual, but there is an explicit link via continuous functions . For
any element z of en' the polynomials in Rn = K[xl , ..., Xn] vanishing
at z form a prime ideal Pz, and the function z H Pz is continuous
as a map from en with the Zariski topology to S(Rn) with the hull-
kernel topology. Indeed, in modern terminology, the latter space is the
sobrification or sober reflection of the former. Here, a space is sober
iff each irreducible closed set is the closure of a unique point (see e.g.
Johnstone [78]). Any topology is lattice-isomorphic to that of the
sobrification. In the present situation, both topologies are isomorphic
to the system of all radical ideals in Rn.
68 M . Erne
3 Galois Connections in the Twentieth Century
Die "abstrakte", "formate" oder· "axiomatische" Richtung, der

die Algebra ihren erneuten A ufschwung verdankt, hat vor all em
in der Gruppentheorie, der Korpertheorie, ... und der Idealthe-
orie . . . zu einer· Reihe von neuartigen Begriffsbildungen, zur
Einsicht in neue Zusammenhiinge und zu weitreichenden Re-
sultaten gefiihrt.
From the introduction to

van der Waerden's Algebra
3.1 From Dedekind to Noether, Artin and van der

Waerden:
The Modern Galois Theory
Richard Dedekind was one of the first mathematicians to present the
Galois theory in its modern (field-theoretic) form. As we saw, great
parts of that theory are already contained in his unpublished Lectures
on Algebra [26]; others are to be found in his celebrated Supplement
XI to Dirichlet's Lectures on Number Theory [27]. A concise and well
readable survey on the state of Galois theory at the beginning of a
new century is Dedekind's 1901 paper [32]:
Uber die Permutationen des Korpers aller· algebraischen Zahlen (On
the per·mutations of the field of all algebraic numbers)
which opened the door to the algebra of the twentieth century, includ-
ing already first investigations on infinite and normal field extensions.
Ernst Steinitz [136], Emmy Noether [102], Emil Artin [5] and B. L.
van der Waerden [142] are prominent representatives of the modern
algebra in the first half of the twentieth century. They worked in the
tradition of Dedekind and brought Galois Theory to a new culmination
point.
Below, we collect together those ingredients of the Fundamental The-
orem of Galois Theory that can be found in Dedekind's later work.
Then we shall point out how Galois connections simplify the consid-
erations. We also include some of the ingenious ideas contributed by
Art in.
By a "permutation" of a subfield K of the field C of complex num-

bers, Dedekind means in [32] a monomorphism 1r from K into C (that
is, an isomorphism onto a subfield of C). Dedekind uses here the
modern notation an for the image of a under 1r . Of particular inter-
est from the modern order-theoretical perspective is his idea to order
not only the fields by the subfield ("divisor") relation, but also the
monomorphisms, calling a monomorphism <p : K ----t C a "divisor" of
a monomorphism 'ljJ : L ----t C (and 'ljJ a "multiple" of <p) if (P) of
all elements on which these monomorphisms coincide, and that the
common restriction of all 1r E P to K is then the greatest common
divisor (the so-called "remainder") of P. Next, Dedekind cites two
important theorems from his Supplement [31].
Theorem on Permutations of Finite Field Extensions
If the field M is a finite extension of the field K and 1r : K ----t C is any
monomorphism then the set M1r of all multiples of 1r with domain M
satisfies K = if>( M1r) , 1r is the remainder of M 1r, and the degree [M: K]
is its cardinality. In particular, if H is the set of all monomorphisms
from M into C fixing K pointwise then K in turn is the fixpoint set
of H, and
#H=[M:K].
Thus, in modern t erminology, the Galois connection between sets of
monomorphisms from M into C and fields K having M as a finite
extension, induced by the fix relation ( G)
is the fi eld of its common fixpoints, then M is a finite extension of
K whose degree is the order of G, and G is the automorphism group
f(K) of M over K:
[M:K] = #G and G = f(K) = f(if>(G)).
The second equation follows from the first by the general theorem
in Sect ion 1.3. In the booklet Galoissche Th eorie [5], Art in gives a
nice proof of the last theorem, using basic tools of linear algebra and
70 M . Erne
another theorem due Dedekind, namely the

Character Independence Theorem
Any set of homomorphisms from a semigroup into the multiplicative
group of a field is linearly independent.
Artin's proof of the Theorem on Automorphism Groups shows that
the equation [M: K] = #G includes the infinite case, saying that the
group G is infinite if and only if so is the extension M : K. The proof of
the Character Independence Theorem is a quick but tricky induction
which we omit here .
Consider a 1 , ... ,an E G and b1 , . . . , bm, c1 , . . . , Cn E M such that the bj
are linearly independent but satisfy 2::::~ 1 ai(bj)ci = 0 for j = 1, ..., m.
Then, for any a = 2::::7= 1 aj bj E M with aj E K,
2::::7= 1 ai(a)ci = 2::::7= 1 aj 2::::7= 1 ai (bj)ci = 0.
Thus, if the bj generate M , all coefficients ci have to be zero, by the
Independence Theorem. This shows that n cannot be greater than m ,
hence #G :::;; [M : K], including the fact that if G is infinite then so is
[M:K].
We may now assume that G = {a 1 , ... , an}· If n < m , there is a
non-zero vector (d 1 , .. . , dm) E Mm with
2::::7= 1 ai - 1 (bj) dj = 0 for all i = 1, ... , n.
Again by the Independence Theorem, the trace a = 2::::~ 1 ai does
not vanish everywhere, so one may assume a(dk) i= 0 for some k.
Application of ai and summation over i yields the identity
0 = 2::::~ 1 2::::7= 1 bjai(dj) = 2::::7= 1 bj a(dj).
But the values a(dj) are fixed by the automorphisms ai, hence ele-
ments of K, which however contradicts the linear independence of the
bj. Therefore, the dimension (degree) of M over K is exactly n.
An extension M: K where K is the fixpoint field of some finite set P
of automorphisms of M is referred to as a (finite) Galois extension.
The subsequent equivalent characterizations of such extensions can be
found in any modern textbook on classical algebra (see, for example,
[5, 90, 98, 142]).
Characterization Theorem for Galois Extensions
For a finite field extension M: K, the following are equivalent:
(a) M: K is a Galois extension.

(b) [M: K] = #f(K).
(c) M : K is normal and separable.
(d) M is the splitting field of a separable polynomial over K.
In Dedekind's studies, separability plays no role since he confines his
considerations to subfields of C. However, he is aware of the difficul-
ties arising in the theory of infinite extensions; here, he shows (by a
countability argument) that every monomorphism on a subfield of a
field A of algebraic numbers has a "multiple" (that is, a monomorphic
extension) with domain A. Next, he explains why the above theorems
comprise the essence of Galois theory. Starting with a finite group G
of automorphisms of the field M and the fixpoint field K =( G) , he
formulates the crucial one-to-one correspondence between the inter-
mediate fields of M : K and the subgroups of G. Using the modern
language of Galois connections, one may rephrase his explanations as
follows (see [5] or [98]):
Fundamental Theorem of Galois Theory
For any Galois extension M: K, the polarity of the fix relation induces
mutually inverse isomorphisms and r between the closure system
g = Q(M: K) of all subgroups of the Galois group G = f(K) and the
closure system F = F(M: K) of all subfields of M containing K.
Let us see what is needed for the proof. We have the following four
maps:
the inclusion inverting map
: Q-+ F, H M {a EM: VaEH (a(a) =a)} ,
its (dual) Galois adjoint
f: F-+ Q, L M {a E G: VaEL (a(a) =a)} ,
the isotone ordinal scale into N = w \ {0} , ordered by divisibility,
~g : g -+ N, H M #H,
and the antitone degree scale

~;::F-+N, LM[M:L].
By the equivalence (a) <=? (g) in the Theorem on Perfect Galois Con-
nections (Section 1.3), the only two details that have to be checked
are the identities ~;:o = ~g and ~go r = ~;: . But these are just
the above equations
72 M. Erne
[M:(H)] = #H and #f(L) = [M :L] .

The first equation was established in Artin's proof of the First Theo-
rem on Finite Automorphism Groups. One obstacle arises in the ver-
ification of the second equation: one must know that whenever M: K
is a Galois extension then so is M : L , for every L E F(M : K). Of
course, one may solve that problem by invoking the Characterization
Theorem and observing that normality and separability are inherited
by the intermediate extensions M : L . But the proof of those facts is
not easy and requires a lot of extra definitions and tools. If one wishes
to circumvent that difficulty, one may alternatively use the following
result due to Artin:
Second Theorem on Finite Automorphism Groups
Let G be a finite group of automorphisms of a field M . Then each
intermediate field L between ( G) and M satisfies the equations
[G :f(L)] = [L: (G)], [M: L] = #f(L) ,
and M: L is a Galois extension.
For the proof, put K =( G) and H = f(L) . The first step is to prove
the inequality
[G:H]::; [L:K].
To that aim, Artin considers the set S = {aiL : a E G} and observes
that two automorphisms a , T E G coincide on L iff H a = H T. Thus,
the cardinality of S is the index [G: H], and the same argument as
in the proof of the First Theorem on Finite Automorphism Groups
yields the inequality #S ::; [L : K]. Now, using the degree and index
formulas, one obtains
[M:L] = [M:K]/[L :K] = #G/[L:K]::; #G/[G :H] =#H.
Thus we have
1::1;: ::; 1::1 9 o r ,
which together with the second degree formula #H = [M : (H)], i.e.
!:1g = !:1;:o 

provides all hypotheses needed in order to apply t he general theorem
on (Right) Perfect Galois Connections.
3.2 Garrett Birkhoff's Polarities

As mentioned earlier, many Galois connections (among them the clas-
sical one between subfields and subgroups) arise from a relation R
between a set A and a set B of functions whose domain includes A ,
so that R is defined in terms of a certain equation involving members
of A and B.
Apparently, Garrett Birkhoff was the first to point out the fundamen-
tal construction of Galois connections, termed polarities, from arbi-
trary relations between two sets. The initial publication containing a
systematic treatment of these ideas is the 1940 edition of Birkhoff's pi-
oneering Lattice Theory [3], a textbook that today is still a rich source
of inspiring concepts, facts and methods in the theory of lattices and
ordered sets.
Let us recall Birkhoff's construction, now well-known and applied
in hundreds of mathematical and extra-mathematical problems, and
sketched briefly in Section 1.4. It starts with a relation between (the
elements of) two sets A and B , that is, with a subset R of the carte-
sian product Ax B. Naturally associated with such a relation are two
canonical maps, <]) = R-+ and \If= +-R, between the power sets:
<]) : P(A) -t P(B), X H XR = {y E B : xRy for all x EX} ,
\If: P(B) -t P(A), Y H Yn = {x E A : xRy for all y E Y}.
While Birkhoff writes X* for xn and yt for Yn , Ganter and Wille
[59, 153] prefer to use the same symbolic notation X' andY' for these
sets. Since often several relations have to be considered simultane-
ously, we have chosen a notation that indicates the dependence of
the relation in question. Furthermore, as the relations need not be
symmetric, we distinguish between different positions.
As Birkhoff remarks next , the maps <]) and \II invert inclusion, and
their composites are extensive, that is,
X ~ X*t = XRn and Y ~ yt* = YnR·
Because of the immediate equations
\If o <]) o \II = \II and <]) o \II o <]) = <]) ,
it is now clear that the composite maps
\If o <]) : P(A) -t P(A) , X H X*t and
74 M. Erne
<P 0 '11 : P(B)--+ P(B) , y H yt*

are closure operators, and that <P and '11 induce mutually inverse dual
isomorphisms between the associated closure systems
Cq, = {<P(X) =X*: X~ A}= {Y ~ B : Y = yt*} ,
which is the range of <P as well as of <Po '11 , and
Cw = {w(Y) = yt: Y ~ B} ={X~ A : X= xt*} ,
which is the range of '11 and of '11 o <P. In particular, for X ~ Cw , one
has
(V X)*= n{x*: X EX} , (nX)* = V{X* : X EX} ,
and analogously for subsets of Cq, .
Birkhoff mentions the following four "classical" examples:
• A is a group of automorphisms of a field and o:Ry means a(y) = y
(leading to the Galois connection between subgroups and subfields) ,
• A is a set of functions on the space IRn , and <pRy means c.p(y) = 0
(leading to connections between varieties and Zariski topologies etc.),
• A= B is a ring, and xRy means xy = 0
(leading to the connection between left and right annihilator ideals),
• A = B is a partially (or quasi-)ordered set with order relation R
(leading to the completion via Dedekind-MacNeille cuts) .
The term "polarity" is borrowed from the following examples of sym-
metric relations (where A coincides with B and R with the dual rela-
tion) : if A = B is a finite-dimensional euclidean space with scalar
product ( , ) then, for the orthogonality relation x ..l y given by
(x, y) = 0, the map <P = '11 associates with any subset X its orthog-
onal space X _l_, and it induces a dual automorphism on the lattice of
all subspaces. In the case of (infinite-dimensional) Hilbert spaces, one
still obtains a dual automorphism , but only on the lattice of all closed
subspaces.
Two further examples mentioned in [3) are the symmetric relation Son
a group given by xSy {::} x y = yx, for which <P = '11 carries each sub-
group to its centralizer, and the symmetric inequality relation , which
leads to the dual automorphism of power sets via complementation.
As observed by Everett [54) and mentioned already in Section 1.4, not
only does every relation between two sets (which may be identical) give
rise to a Galois connection between the corresponding power sets, but
also conversely, every Galois connection between power sets arises in
that fashion from a unique relation: indeed, if : P(A) --t P(B) and
W : P(B) --t P(A) are two maps satisfying
X ~ w(Y) {:} Y ~ (X) (X ~ A, Y ~ B)
then there is a unique relation R ~ Ax B with = R-+ and W = +--R,
VlZ.
xRy {:} x E w({y}) {:} y E ({x}).

Of course, this is the only possible way to define R so that = R-+,
and if one defines R by that equivalence then one actually obtains
R-+(X) = {y: xRy for all x EX} =
{y E B: X~ w({y})} = (X),
and analogously +--R (Y) = W(Y) .
Many interesting isomorphism theorems of order theory, algebra and
topology later turned out to be induced by a rather simple polarity
- and indeed, the general theory says that every dual isomorphism
between two closure systems (respectively, every isomorphism between
a closure system and a kernel system) arises in that fashion from a
relation.
The example of an order relation and the associated polarity deserves
particular attention, also from the historical point of view. As ob-
served by MacNeille [91] in 1937, a few years before the publication
of Birkhoff's Lattice Theor-y, Dedekind's famous construction of the
reals from the rationals by so-called cuts [27], published first in 1872,
extends to arbitrary posets, providing a universal completion. There
are at least two alternative ways of defining cuts in arbitrary posets:
in MacNeille's original article, cuts are pairs (X, Y) with X= Y< (the
set of all lower bounds of Y) andY = x :s: (the set of all upper bounds
of X). This approach follows the idea of Dedekind and was generalized
later in formal concept analysis, where the notion of cuts was extended
to that of concepts (see Section 1.5). The obvious up-down duality
of that "two-sided" definition has certain conceptual advantages but
is technically less convenient than the second approach, preferred by
Birkhoff [3], where the first components Y:s; are regarded as (lower-}
76 M. Erne
cuts; Birkhoff also speaks of "closed ideals ", reminding of the fact
that they are closed sets with respect to the involved Galois connec-
tion . By definition , they are just the intersections of principal ideals.
It is evident (and follows from the general remarks on polarities) that
the lower cuts form a closure system and therefore a complete lattice,
dually isomorphic to the closure system of upper cuts (intersections of
principal filters) . As Birkhoff points out, a poset is complete if and
only if every cut is already a principal ideal, while in general, any
poset P is embedded in the closure system N P of all cuts by sending
each element to the principal ideal generated by that element:
TJ : p -+ N p' X H -!-X = {p E p : p ::; X} .
This completion by cuts, alias Dedekind-MacNeille completion or nor-
mal completion, may be characterized abstractly by various universal
properties (see [2, 10, 15, 40, 10, 45, 46, 124] for details) .
Characterizations of the Completion by Cuts
For an embedding c of a poset P in a complete lattice C , the following
conditions are equivalent:
(a) There is an isomorphism t : NP-+ C with c =to 'rf ·
(b) c( P) is join- and meet-dense in C.
(c) c preserves all existing joins, and c (P) is join-dense in C.
(d) For every order embedding c1 of Pin a complete lattice C' , there
is an order embedding t : C -+ C' with c' = t o c.
(e) For every join-dense embedding c' of P in a complete lattice C' ,

there is a unique join-dense embedding t : C -+ C' with c' = to c.
(f) For every cut continuous map c' from P into a complete lattice
C' , there is a unique join-preserving map t : C -+ C' with c 1 =
t 0 c.
(g) For every cut stable map c' from P into a complete lattice C',
there is a unique join- and meet-preserving map t : C -+ C' with
c' = t 0 c.
A map <p : P -+ Q between posets is (lower) cut continuous if inverse

images of cuts are cuts (i.e. closed ideals). Every residuated map is cut
continuous, every cut continuous map preserves joins, and on complete

lattices, all three types of maps coincide. Furthermore, <p : P ----+ Q is
cut stable (see [45]) iff
cp(X:S):::; = cp(X:S:::; ) and cp(X:::;):S = cp(X:::;:S) for all X ~ P.
Any such map is lower and upper cut continuous, but not conversely.
Precisely speaking, Dedekind's completion of the rationals gives a con-
ditionally complete lattice (in which all nonempty bounded subsets
have join and meet); in the general setting of posets, that construc-
tion may be imitated by taking the nonempty proper cuts only. Here,
Birkhoff proves the following general result:
Conditional Completion of Bidirected Posets
If P is an up- and down-directed poset (all finite subsets have upper
and lower bounds) then the nonempty proper cuts form "the'' condi-
tional completion, that is, a conditionally complete lattice in which P
is join- and meet-densely embedded (which forces uniqueness of the
conditional completion up to isomorphism).
Birkhoff's contributions to the lattice-theoretical development of ge-

ometry are immense and cannot be reported here (see [14] for one of
his own late reviews of the research in that field). Since (synthetic)
geometry is based on the idea of incidence relations, it is plain that
Galois connections and closure systems or operations play a crucial
role in that context. However, in order to avoid misunderstandings, it
should be pointed out that for geometries of arbitrary dimension, it
is not the incidence between points and lines (objects of dimension 1)
that creates the desired closure systems of "flats ", but that between
points and hyperplanes (objects of codimension 1). Indeed, calling
copoints the completely meet-irreducible members of an arbitrary clo-
sure system (e.g. the hyperplanes in the case of subspace systems) ,
one has the fundamental
Polarity Between Points and Copoints
Let C be any algebraic closure system on a set A (that is, directed
unions of closed sets are closed). Then the polarity associated with the
incidence relation between points and copoints has C as the system of
Galois-closed subsets of A , and the principal filt ers of closed sets are
the Galois-closed subsets of C.
78 M. Erne
Of course, this result requires the Axiom of Choice in form of a max-

imal principle, guaranteeing enough copoints in order to generate all
closed sets via intersection. Abstractly, this is the famous Birkhoff-
Frink Theorem, saying that in an algebraic lattice, every element is a
meet of completely meet-irreducibles. Many important results of uni-
versal algebra rely on that theorem (see e.g. [3] and [25]) .
3.3 Oystein Ore's Galois Connexions

Oystein Ore was well acquainted with Dedekind's style and methods,
and consequently interested in many questions combining order and
lattice theory with algebraic structures like groups and rings. Let us
mention here only two typical examples of that fruitful combination:
Ore's beautiful theorem characterizing the locally cyclic groups by
distributivity of their subgroup lattice [104] (for a short proof, see
[126, p.12]) , and Ore's famous theorem on the exchange of elements
in irredundant decompositions for elements in modular lattices (see
[103], and [3, 25] for improvements), generalizing the Steinitz exchange
theorem for bases of vector spaces.
Probably the first source where the abstract order-theoretical notion
of Galois connections emerged explicitly is Ore's 1944 paper Galois
Connexions [38]. As he remarks in the introduction, various causes
urged him to postpone a book on the subject, and it remains unclear
whether that book has ever been published in some form. Birkhoff
included the general notion of Galois connections in the second edition
of his Lattice Theory.
Ore himself changed between the terms "Galois connexions" , "Galois
connections" , and "Galois correspondences", the latter two expres-
sions being standard (and often synonymous) until today. Though
Ore's style is sometimes less precise than Dedekind's and several of
Ore's notational idiosyncracies did not survive ( "connexion" for "con-
nection", "structure" for "lattice", "dual topology" for "closure sys-
tem" , "correspondence" for "map" etc.), his brilliant structural ideas,
often based on some very simple but effective order-theoretical con-
cepts, founded a very fruitful theory of (dual) adjunctions, extended
later to the categorical concept of adjoint functors, monads and dual-
ities, developed by the next generation of mathematicians (see [3]) .
Like some of his mathematical ancestors and contemporaries, Ore used

the set-theoretical symbol C (or, not always consistently, the symbol
~) for the order relation; consequently, he interchanged the role of set-
theoretical and lattice-theoretical symbols, as they are used in modern
texts. In order to avoid ambiguities, we translate Ore's symbols into
the nowadays common ones (:::; for the order, V for the binary join,
and 1\ for the binary meet).
As Ore explains in the introduction, his paper arose from a series of
colloquium lectures on the theory of mathematical relations, given in
1941 at the Summer Meeting of the American Mathematical Society
at the University of Chicago. The mathematical part begins with a
brief review of the order-theoretical concept of closure operations (Ore
calls them "closure relations") . As mentioned in Section 1.2, such
closure operations x 1--t x are characterized by the three properties
of idempotency (x = x ), extensivity (x :::; x) and isotonicity (x :::;
y =? x:::; y). Ore cites his own papers [107, 108] and Ward [149] but
not the work of Dedekind, who had observed (in a slightly different
setting) that the three defining conditions for a closure operation may
be combined to a single one, namely x :::; y {::} x :::; y (see Section
2.5).
Ore defines isomorphisms between partially ordered sets P and Q in
the usual way as bijections cp : P ----t Q such that
x:::; y {::} cp(x) :::; cp(y),
and dual isomorphisms between P and Q as isomorphisms between P
and the dual of Q. Among many possible alternative characterizations
of such dual isomorphisms, the following two often occur in Ore's
paper:
p= x V y {::} cp(p) = cp(x) 1\ cp(y)
and dually,
p= x l\y {::} cp(p)=cp(x)Vcp(y)
where p = x V y means that p is the supremum (join, least upper
bound) of x andy, and dually, p = x 1\ y means that pis the infimum
(meet, greatest lower bound) of x and y. Note that the above equiv-
alent definitions of dual isomorphisms work not only in lattices, but
in arbitrary posets. Two posets are (dually) isomorphic if there exists
80 M. Erne
a (dual) isomorphism between them, and a (dual) automorphism is a

(dual) isomorphism between a poset and itself; finally, a self-inverse
dual automorphism is called an involution.
As a first theorem, Ore states the one-to-one correspondence between
closure operations and meet-closed subsets of complete lattices, ob-
tained by associating with any closure operation its range (=fixed
point set), and in the opposite direction with any meet-closed sub-
set P 1 the map that sends to each element the least member of P 1
above it. A suitable extension of that correspondence to arbitrary
partially ordered sets was discussed later by Robert Baer [7]; in that
general context, meet-closed subsets have to be replaced by "partial
ordinals" in Baer's terminology (closure ranges in the sense of Section
1. 2). Recall that we have reserved the term closure systems for the
set-theoretical version, i.e. for systems closed under arbitrary intersec-
tions, while by a closure range in a poset P we mean a subset P 1 such
that for each p E P there is a least p 1 E P 1 dominating p. Ore con-
firms that the closure operations of a complete lattice form a complete
lattice that is dually isomorphic, under the above correspondence, to
the closure system of all closure ranges.
After the preparatory remarks on closure operations, he introduces
his fundamental "Galois connexions" or "Galois correspondences" as
pairs of antitone (i.e. order-reversing) maps <p : P---+ Q and 'ljJ : Q ---+ P
whose composites are extensive, that is,
p ::; 'ljJ o <p (p) for p E P and q ::; <p o 'ljJ (q) for q E Q.
Ore does not mention the equivalent characterization by the condition
p::; '1/J(q) {::} q ::; <p(p) ,

but immediately notes the identities 'lj;o <p o 'ljJ = 'ljJ and <p o '1/J o <p = <p,
from which he deduces that <p o 'ljJ and 'ljJ o <p are closure operations
whose ranges are dually isomorphic. Everybody using Galois connec-
tions is aware of these basic and useful facts. The converse statement
however, also due to Ore, seems to be less known: namely, that given
two complete lattices P and Q with meet-closed subsets P 1 and Q1 ,
respectively, any dual isomorphism between P 1 and Q 1 is induced by
a unique Galois connection between P and Q. The partners of that
connection are obtained simply by composing the involved closure op-
erations with the appropriate dual isomorphisms. Again, that result

extends to arbitrary posets, when meet-closed subsets are replaced
with closure ranges (see Section 1.2). More precisely, ordering the
Galois connections between posets P and Q "pointwise" by
(cp,~)::; (cp',~') {::} Vp EP (cp(p)::; cp'(p)) {::} Vq EQ ( ~ (q)::;
~'(q)) ,
one obtains a poset P ® Q ~ Q ® P with the following property:

Theorem on the Ranges of Galois Connections
Via restriction to the ranges, the poset P ® Q of all Galois connections
between P and Q is isomorphic to the poset of all dual isomorphisms
between closure ranges of P and of Q, respectively. If P and Q are
complete lattices then so is P ® Q.
Ore defines "one-sided perfectness" of a Galois connection by requir-
ing that one of the partners be onto, and he observes that this is
equivalent to the postulate that the other partner be one-to-one. His
definition of "perfect Galois connections" is the usual one, requiring
injectivity (hence also surjectivity) of both partners. Thus, perfect
Galois connections consist of two mutually inverse isomorphisms (see
our introductory Section 1.3).
In the third section of his paper on Galois connections, Ore indicates
an ''interpretation of the theory in terms of continuous maps", but he
does not say precisely what he means by an "interpretation" . Also,
his claim that
any Galois connection between [complete lattices} P and

Q corresponds to a mapping of a space with a structure
of closed sets isomorphic to P onto a dense subspace of a
space whose structure of closed sets is isomorphic to the
dual structure Q* of Q
fails, no matter how the word "corresponds" is interpreted: for the

unique Galois connection between a one-element and a two-element
chain, there is no mapping between spaces with the claimed proper-
ties. The whole section about continuous mappings is a bit vague and
becomes more transparent in terms of adjunctions rather than Galois
connections. Indeed, it is possible to describe exactly the passage be-
tween adjoint maps and continuous maps (see e.g. [41] - [12]). Here,
82 M. Erne
as often, it proves advantageous to invoke the machinery of functo-

rial equivalences and adjunctions, the modern categorical extension
of order-theoretical adjunctions (see e.g. [3]). For the moment , the
following hint may suffice: every complete lattice may be regarded as
a closure space, by taking the principal ideals as closed sets; in that
setting, the residuated , i.e. left adjoint maps are precisely the contin-
uous ones, and for an arbitrary closure space S with closure system
C, the map
TJS : S--+ C, X H {X}
is continuous. Moreover, one has the following connection between
Continuous Maps and Adjoint Maps
Let C and C' be closure systems on sets S and S', respectively. Then
an ar·bitrary map rp : S --+ S' is continuous if and only if there is a
(unique) left adjoint, i.e. join-preserving map <P : C--+ C' with
<Po TJs = TJs' o rp.
Hence, the category of complete lattices and residuated maps is reflec-
tive in the category of closure spaces, with refl ection maps TJS·
Of course, the "lifted" map <P sends any closed set to the closure of its
image, and the right adjoint of <P sends any closed set t o its inverse
image. In other words, this adjunction is induced by the axiality
associated with the continuous map cp.
Let us mention only a few ideas in that area due to Ore. One of his
aphoristic statements is that
every complete structure {lattice} P can be considered to be

the structure of closed sets for a suitable topological space.
In order to understand that sentence, one must know that Ore means
by a topological space a set S equipped with a closure operator ( "clo-
sure relation" ) making the empty set closed, while finite unions of
closed sets need not be closed for him. Of course, such a "topological
representation" of a complet e lattice P is easily achieved, by endowing
the underlying set Po= P\ {0} (where 0 = 1\ P = V 0 is the least el-
ement) with the closure operator X H ] 0, V X], so t hat the "deleted"
principal ideals ] 0, p] = {x E P : 0 < x ::; p} form a closure system
isomorphic toP, and the empty set is closed. Now, Ore claims that
any substructure P1 of P with respect to crosscut [i.e. any

meet-closed subset] corresponds to a mapping of this space
onto a space whose structure of closed sets is isomorphic
to P1 .
Obviously, the mapping he has in mind is the closure operation as-
sociated with P 1 , restricted to P0 . This is in fact a continuous map
(between closure spaces), since the preimage of the deleted principal
ideal generated by p 1 in P 1 is the deleted principal ideal generated by
p 1 in P. For some more topological constructions involving adjunc-
tions, see the end of Section 3.4.
By a Galois connection within a complete structure, Ore means a Ga-
lois connection between a complete lattice P and P itself. By the
Theorem on Ranges of Galois Connections, those Galois connections
form a complete lattice isomorphic (via restriction) to the complete
lattice of all dual isomorphisms between pairs of meet-closed subsets of
P. Under that isomorphism, the dual automorphisms of meet-closed
subsets correspond to the Galois connections within P whose partners
have the same range. In particular, the involutions of meet-closed
subsets correspond to those Galois connections whose partners are
identical.
Deviating from Birkhoff's terminology, but in accordance with other
authors, Ore means by a polarity or orthogonality (alias orthocomple-
mentation) an involution under which the image is always complemen-
tary to its preimage. (Ore speaks of "the" complement, but we know
from the theory of vector spaces that complements are not necessarily
unique, even if such an involution exists; also, instead of Ore's nota-
tion a* for the orthocomplement, conflicting with his notation R* for
the dual relation and P* for the dual poset, most authors use the sug-
gestive symbol al_.) Then, Ore says that "because of the importance of
the applications such polarities are of particular interest" but gives no
such examples (in this paper). Of course, interesting applications are
to be found in the theory of Boolean lattices and, more generally, of
orthomodular lattices, in particular in the theory of Hilbert spaces and
their closed subspace lattices [3, 16, 145]. As observed in [34], rather
mild hypotheses force an antitone self-map <.p of a bounded lattice L
to be an orthogonality. Indeed, if
<.p(x) ~ x::::} x = 1 and <.p(x) = 1::::} x = 0 for all x E L
84 M. Erne
and cp is antitone then cp(x) must be a complement of x:

j := x V cp(x) 2: x =?- cp(j) ::::; cp(x) ::::; j =?- j = 1,
m := xl\cp(x) ::::; cp(x) ::::; cp(m) =?- cp(m) = mVcp(m) = 1 =?- m = 0.
If, moreover, cp is self-inverse then it is an orthogonality.
Next, Ore briefly considers Galois connections between a complete lat-
tice P and its dual, thereby anticipating the concept of adjunctions
(Galois connections of mixed type in the sense of Benado [11] and
Schmidt [123]). Ore observes (without proof) that any complete ho-
momorphism a : P---+ P of a complete lattice P gives rise to a Galois
connection (cp, 'ljJ) between P and its dual with
cp(x) =max{ y: a(x) = a(y)} and '1/J(x) = min{y: a(x) = a(y)}.
In Section 5, 0re reports and comments on his own and Birknoff's
results concerning the passage between relations and Galois connec-
tions. His claim that the theory of Galois connections be "expressible
in terms of relations" remains somewhat unclear. What he probably
means (but what he does not state explicitly) is the following
Representation of Galois Connections by Relations
Associating with an arbitrary Galois connection (cp, 'ljJ) between posets
P and Q the relation R s:;; P x Q with
xRy {::} x::::; '1/J (y) {::} y::::; cp(x),
one obtains an isomorphism between the poset P 0 Q of all Galois
connections and the poset of all G-ideals in P x Q, i.e. relations R
such that for each x E P, the set {x }R = {y E Q: xRy } is a principal
ideal of Q, and for each y E Q, the set {Y}R = {x E P: xRy } is a
principal ideal of P.
The following diagram (where T/P and TJQ denote the principal ideal
embeddings) commutes:
TJp
R+--
p (P)~===R==-->===:::::; p (Q)
Hence, every Galois connection is induced, up to isomorphism, by a

polarity.
Later authors called P 0 Q the tensor product of P and Q, though it
shares the desirable properties of usual tensor products only if P and
Q are complete lattices. For example, the "associative law"
P 0 (Q 0 R) ~ (P 0 Q) 0 R
valid for complete lattices, fails for arbitrary posets. The structure
of such tensor products has been investigated intensively by various
authors (see, for example, [89, 99, 133]). Remarkable is that the tensor
product of lattices need not be a lattice [77], and the tensor product
of distributive complete lattices need not be distributive, while the
tensor product of completely distributive lattices is always completely
distributive (see Raney [118] and Shmuely [134]).
As shown in [43], most of the deficiencies occurring in the theory of
tensor products for posets are remedied by replacing G-ideals with the
more general "multicuts" T ~ P x Q, defined by the condition
x xY ~ T =? x::::<- x y::::<- ~ T.
This definition also opens the possibility of defining tensor products
with an arbitrary number of factors (see [43] for details). A differ-
ent tensor product for complete lattices plays an important role in
formal concept analysis [59] and lattice-theoretical topology [49, 53].
For completely distributive lattices (but not for arbitrary distributive
complete lattices) it coincides with that defined by Galois connections.
Interestingly, Ore does not consider (binary) relations as the funda-
mental notion and functions as a derived special kind of relations, but
conversely, he defines relations as functions R : B---+ P(A) from a set
into a power set and then declares the expression xRy (not yRx !) by
x E R(y). Of course, both approaches lead to the same conclusions.
Next, Ore states and proves Everett's theorem saying that every Galois
connection between power sets comes from a unique relation between
the underlying sets [54]. In slightly extended and supplemented form ,
that theorem may be rephrased as follows:
Galois Connections Between Power Sets
For any two sets A and B , the following three collections form mutu-
ally isomorphic complete lattices:
86 M. Erne
(a) the power set P(A x B),

(b) the set of all Galois connections between P(A) and P(B) ,
(c) the set of all closure dualities, i.e. dual isomorphisms between
pairs of closure systems on A and B, respectively.
As mentioned earlier, an isomorphism between (a) and (b) is given by
associating with every relation R ~ A x B the pair (R--+, +-R), and
an isomorphism between (b) and (c) is obtained by restriction to the
ranges.
Since Ore means by a "dual topology" merely a closure system contain-
ing the empty set as a member, he has to insert minor modifications
in order to describe dual pairs of "topologies" by means of relations,
namely by those relations which have the property that no y E B
satisfies A= R(y), and no x E A satisfies Y = R*(x). By reasons of
analogy, we call a relation with that property unbounded, because in
the case of order relations, it just excludes the existence of greatest
and least elements. Let us summarize some of Ore's results about
specific relations and Galois connections.
Correspondence Between Relations and Closure Dualities
Under the bijection between relations and Galois connections,
(1) the complements of bijections correspond to perfect Galois con-
nections between power sets,
(2) the unbounded relations correspond to dual isomorphisms be-
tween closure systems with closed empty set,
(3) the symmetric relations correspond to the involutions of closure
systems,
(4) the symmetric anti-reflexive relations (graphs) correspond to or-
thogonalities of closure systems.
Ore points out (without formulating explicitly the isomorphism) that

(4) "gives an actual method for the construction of all complete struc-
tures which have a polarity".
Concerning (2) , it should be mentioned that if one wants to charac-
terize dualities between topological closure systems in the usual sense
(closed under finite unions), one must restrict the representing rela-
tions by the additional condition that all sets of the form R(bt) U R(b2 )
and R*(ar) U R*(a 2 ) be closed, which explicitly means that if neither

cRb 1 nor cRb 2 then there is a b with xRb whenever xRb 1 or xRb 2 ,
but not cRb, and dually. R.-E. Hoffmann [75] has observed that a
topology (in the usual sense) has a "dual" , i.e. , is dually (lattice) iso-
morphic to another topology, iff it has a base of strongly connected
open sets, where a nonempty subset U is strongly connected (Levine
[88]) iff for all open V, W, the inclusion U ~ V U W entails U ~ V or
U ~ W. Thus, under the isomorphism between relations and closure
dualities, the above type of relations are in one-to-one correspondence
with dual isomorphisms between topologies (having a base of strongly
connected sets).
In the final section, Ore touches on a concept he calls Galois theor·y for
relations. Today, one is familiar with Galois theories for any sort of au-
tomorphisms. In the specific setting of a relation Ron a setS, an auto-
morphism of (S, R) is a bijection a : S----+ S with xRy {::} a(x)R a(y),
thus generalizing the notion of automorphisms for order relations. As
in all similar situations, the automorphisms for R form a subgroup
G R = G(S, R) of the full permutation group on S. Again, the fix
relation induces a Galois connection ( , w) , where (X) is the set of
all elements fixed by each member of X ~ G R , and w(Y) is the set of
all automorphisms fixing each member of Y ~ S.
Ore focusses on the specific case of an equivalence relation E and the
associated partition PE. Here, G E is a direct product of groups G E,,,
consisting of automorphisms of the union SE,, of all blocks of size 11,
(((complete monomial groups" in Ore's terminology [105, 106]). Then
he describes the Galois-closed subsets of S.
The paper concludes with some remarks on general Galois theory,
starting with some group G of permutations of a setS and making use
of the polarity associated with the fix relation. Ore first considers the
case where G acts transitively on S. Here, the Galois-closed subgroups
of G are the intersections of conjugates of one fixed stabilizer. The
general case is then settled by partitioning the set S into its orbits
under the action of G.
Later generations have developed a broad spectrum of instances of
polarities or more general Galois connections for groups, rings , alge-
bras, topological spaces, and many other structures. But Ore's basic
concept of Galois connections remained unchanged until today.
88 M . Erne
3.4 Adjunctions as Galois Connections of Mixed

Type
During the forties and fifties , many pieces of the Galois connection
puzzle have been collected together by various authors, not all of which
can be cited here. But some of these contributions deserve explicit
mention.
Everett's early study from 1944 (see [54]) has already been discussed
in connection with Birkhoff's and Ore's work.
In 1948, G. Nobeling started one of the first systematic attempts to
develop a "pointfree" topology by means of lattice theory, initiated in
the thirties by Menger [97] and Ore [38], and culminating in the 1954
monograph Grundlagen der Analytischen Topologie [101]. His earlier
note [100] on that subject is one of the first sources where adjoint pairs
of maps (rather than Galois connections) between complete lattices
occur explicitly: Nobeling considers "regular homomorphisms", i.e.
maps r.p : P---+ Q preserving arbitrary joins, and their "inversions",
i.e. right or upper adjoints '¢ = r.pA, which preserve arbitrary meets
and are given by
r.pA(q) = V{p E P: r.p(p) ::S: q} ·

The original map is then determined as the left or lower adjoint '¢ v of
'¢ by the dual equation
'1/J v (p) = (\ { q E Q : '1/J(q) 2: p} .
As mentioned in the first chapter, the "covariant" counterpart of Ore's
system of axioms for Galois connections, obtained by reversing one of
the partial orders (but not both!) leads to the definition of adjoint
pairs of maps between arbitrary posets. Another early explicit occur-
rence of such a system is found in Benado's 1949 paper [11]:
p:::; p' :::} .\(p) :::; .\(p') ,

q :::; q' :::} (} (q) :::; (} (q') '
P:::; f2(.\(p)), .\(f2(q)):::; q.
Benado speaks in his "intercalation theory" for functions of "con-

nexions monotone d'espece mixte " ( "monotone connections of mixed
type}}).
Also of relevance for the development of an abstract adjunction theory

is Pickert's note Bemerkungen iibeT Galois- Verbindungen [114] from
1952. Being an expert in geometry and its foundations, Pickert was
interested, like Birkhoff, Menger and others, in the lattice-theoretical
background of geometrical constructions. Maybe his note [114] is the
first one where it is stated explicity that a Galois connection between
partially ordered sets is already determined by each of its partners,
and that a map between complete lattices is the partner of a Galois
connection if and only if it transforms joins into meets:
cp(V X) = (\ cp(X).
(We know that the "only if" part holds for arbitrary posets and all
existing joins.) Moreover, Pickert shows that a map
cp: p--+ Q, X H x*
between arbitrary posets is the partner of a Galois connection iff
it is anti tone, i.e. x 1 :::; x 2 ==> x 1 * 2: x 2 *, and each of the sets
{x E P : x* 2: y} with y E Q has a join yt with image yt* 2: y .
If that is the case then yt is even the greatest element of the set
{ x E P : x* 2: y} and the image of y under the second map of
the Galois connection. After those general remarks, Pickert turns to
two specific situations, one motivated by projective spaces, and the
other dealing with the fundamental possibilities of defining topologi-
cal spaces. Similarly as in Ore's considerations on continuous maps,
these examples lead to adjunctions (hence to axialities) rather than
Galois connections (polarities) between power sets. Pickert makes use
of the fact, obtained immediately by inverting the inclusion order of
one power set, that the left (respectively, right) adjoint maps between
power sets are just those which preserve arbitrary unions (respectively,
intersections).
Pickert's geometrically inspired construction begins with two sets M,
I and the set M I of all functions from I into M (here, I is regarded as
an index set and not an incidence relation, so confusion with the intent
MI of formal concept analysis is unlikely). In the specific case where
M is the set of all points in a projective space, I = {1, 2, 3}, and C ~
MI is the set of all collinear triples, the lines are just the maximal sets
X ~ M with XI ~ C. Generally, setting Y* = {X~ M : X 1 ~ Y}
90 M. Erne
for any subset Y of M 1 , one obtains a map Y ~ Y* from P(M 1 ) into

the "double power set" PPM that preserves arbitrary intersections.
Hence, it is one partner of a Galois connection between the dual of
P(M 1 ) and PPM or, otherwise stated, the right adjoint of a map
in the opposite direction, sending any system X ~ PM to the union
xt = U{X1 : X EX}. Apparently, this adjunction is an axiality and
comes from the relation R between PM and M 1 defined by
XRy ¢:} y E X 1.
The second topic concerns the definition of topological spaces by cer-

tain neighborhood systems on a set M. These are families Y = (Ya :
a E M) of subsets of PM satisfying the following axioms:
(Nl) Y E Ya :::? aE Y
(N2) Y E Ya and Y ~ Z ~ M :::? Z E Ya
(N3) y E Ya :::? ::JX E Ya(Y E n{Yx : X EX})
and they are topological neighborhood systems if in addition

(N4) Y,Z E Ya :::? YnZ E Ya ·
Pickert shows that the known one-to-one correspondence between such
topological neighborhood systems and topologies (due to Hausdorff)
comes from a suitable adjunction:
Theorem on Neighborhood Systems and Topologies
For any set M, the map
.X: PPM -+(PPM)M , X~ X*= ({Y: :3XEX (a EX~Y)}:
a EM)
preserves arbitrary unions and is therefore a left adjoint. The Galois-
closed subsets of PM are precisely the kernel systems on M, and the
Galois-open members of (PP M)M are precisely the neighborhood sys-
tems. Under the induced isomorphism, the topological neighborhood
systems correspond to the topologies on M.
We omit the proof details but add the remark that the above ad-
junction may be regarded as an axiality: just replace any family
Y = (Ya : a E M) by the relation {(Y, a) : Y E Ya , a E M}. This
yields an isomorphism between (PP M)M and P(P M x M) . Now, the
relation N ~ PM x (PM x M) defined by
X N (Y, a) {::} a EX~ Y

gives rise to an axiality between PPM and P(P M x M) with
-+N(X) = {(Y, a) E PM x M : ::JX E X(a EX~ Y)} and
N+-(Y) ={X E PM: a EX~ Y::::} (Y, a) E Y}.
In case X is a topology, the first equation just amounts to the usual
definition of neighborhoods, while the second leads to the definition
of open sets as neighborhoods of their points, interpreting (Y, a) E Y
as "Y is a neighborhood of a" .
A rather systematical investigation of Galois connections ( "Galois-
Korrespondenzen vom Hiillentypus ") and of adjunctions ( "Galois-Kor-
respondenzen vom gemischten Typus ") was carried through by Jiirgen
Schmidt in his 1953 paper Beitriige zur Filtertheorie II [123]. This pa-
per also contains applications to various set-theoretical constructions
involving filters, closure systems and topologies, and is partly based on
his Ph. D. dissertation dated 1952 [122]. Schmidt had collected care-
fully the material of other authors on that theme and cited, among
others, Benado [11], Birkhoff [3], Everett [54], Pickert [114], Nobeling
[100], Ore [38], and Riguet [119]. Schmidt formulates the whole the-
ory in the set-theoretical language, with inclusion as order etc., but
he emphasizes that all conclusions remain valid in complete lattices,
and many of them even in arbitrary partially ordered sets. Schmidt's
style is sometimes quite narrative (and a bit long-winded) , and the
formal parts are often supplemented by philosophical and historical
comments. Let us try to report some of his main ideas and results
in condensed form. After an introduction to the theory of relations
(pointing out similarities and differences to Ore's theory) , Schmidt re-
calls some of the basic facts about Galois connections due to Ore and
Pickert (in particular the fact that one partner is determined by the
other) and adds the succinct description by the equivalence
p:::; 'lj;(q) {::} q :::; cp(p).
Schmidt extends Ore's characterization of Galois connections by means
of dual isomorphisms between closure ranges to the setting of arbitrary
ordered sets, as described in Section 1.3. On account of t hat result,
Schmidt chooses the term "Galois correspondences of closure type ".
For him (and others), the main task in the investigation of concretely
92 M. Erne
given Galois connections is a "good" intrinsic characterization of the

Galois-closed elements or sets, that is, of the two ranges associated
with the closure operations obtained by composition of the partners
of the connection.
Next, he turns to the (covariant) adjunctions ("Galois correspondences
of mixed type"), for which he notes, besides Benado's system of axioms
[11], his own characterization
which today is the usual one. Moreover, he characterizes the left

(respectively, right) adjoint maps between complete lattices as those
which preserve arbitrary joins (meets). For the partners of an ad-
junction, he uses the names "Untere Finalkonjugierte" and "Obere
Initialkonjugierte". Apparently, these names allude to the special sit-
uation of continuous maps, which induce an adjoint pair of maps be-
tween the lattices of closed sets: the lower adjoint sends any closed
set to the closure of the image, and the upper adjoint takes preimages
of closed sets (see the previous section).
Schmidt recalls the one-to-one correspondence between relations and
polarities, due to Birkhoff, Ore and Everett. But he does not leave
it at that but discusses systematically the covariant version nowadays
called axialities (see Section 1.4 and [15]). Both the covariant and the
contravariant form of adjunctions remain a guiding theme for Schmidt
during a long period of his work (for an example of more categorical
flavor, see [125]).
The second part of the paper [123] focusses on the theme indicated
in its title, namely applications of Galois connections to the theory of
filters (and topological spaces). It appears adequate to mention here
a few more examples from topology, partly contained in Schmidt's
paper, partly due to other authors (the true origins cannot always be
localized; see also [15]).
(1) Given a set X, define a relation R between A=P(X) and B=XxX
by UR(y,x) {::} x E U implies y E U.
The Galois-closed subsets of A with respect to the associated polarity
are precisely the Alexandroff topologies on X, i.e. those kernel systems
which are at the same time closure systems. On the other hand, the
Galois-closed subsets of B are precisely the quasi-orders on X. Thus,

one obtains in a straightforward way the known dual isomorphism
between the lattice of Alexandroff topologies and that of quasi-orders
on X (see [4]) .
(2) Let F(X) denote the set of all filters on X, and put A = P(X),
B = F(X) x X. Similarly as before, consider the relation R ~ Ax B
with
UR(:F, x ) <=? x E U implies U E :F .
This time, the Galois-closed subsets of A are just all topologies, while
the Galois-closed subsets of B are the topological convergence rela-
tions. This leads to the well-known equivalent description of topologi-
cal spaces by open sets or by convergence (cf. [22, 114, 135]). Example
(1) is obtained by restricting the codomain of the present relation to
principal ultrafilters i: = {Y ~X : x E Y}.
(3) For the symmetric relation U 6 V <=? U n V =f:. 0 on a power set
P(X), the associated polarity induces an involutorial dual automor-
phism of the lattice S(X) of all stacks (subsets U of P(X) such that
U E U and U ~ V ~ X imply V E U) . Under that involution, the
lattice of filters is dually isomorphic to to the lattice of grills [22] (i.e.
complements of set ideals), and the fixed points of that dual isomor-
phism are the ultrafilters.
(4) Now, take the complementary relation U j_ V <=? U n V = 0 on a
power set P(X), or more generally, on a topology T ~ P(X). In that
situation, the Galois-closed subsets of T turn out to be t he principal
ideals {U E T: U ~ V} generated by regular open sets V (i.e. sets
that coincide with the interior of their closure). Thus, the lattice of
Galois-closed sets is here isomorphic to the complete Boolean lattice
of all regular open sets (cf. Glivenko [65], Birkhoff [3] and Frink [55]).
(5) Finally, Schmidt constructs, via a suitable axiality, an isomorphism
between the filter lattice and a certain topology on the set of all ultra-
filters. But this construction is a special instance of the isomorphism
between any spatial fram e L , that is, a complete lattice in which every
element is the meet of a subset of the set Sp(L) of all meet-primes,
and its hull-kernel topology or spectral topology, consisting of all sets
a = {p E Sp (L) : a i p} (a E L) .
94 M . Erne
Clearly, the isomorphism sends a to ii, and its inverse sends any open
set U to 1\ (L \ U). If L happens to be a closure system on a set A (as
in the case of the filter lattice F(X)), then this perfect adjunction is
induced by the axiality corresponding to the relation
R = {(x, P) E Ax Sp(L): x tj_ P}
(cf. the analogous construction of spectral topologies in Section 2.8).
4 Residuation Theory
Recall that a map between ordered sets is residuated if inverse images

of principal ideals are again principal ideals, and that residuated maps
are just the same thing as left adjoint maps, in other words, maps that
have a right adjoint. A residuated groupoid is a set S equipped with a
binary operation A : S x S --t S and a partial order (sometimes only
a quasi-order) such that each of the left translations
xA: S--+ S, y ~ A(x,y)
and each of the right translations
Ay : s --t s' X ~ A(X' y)
is residuated. More generally, a partially ordered (general) algebra is
said to be residuated if all translations, obtained by fixing one coor-
dinate in the basic operations, are residuated.
In order to prevent misunderstandings, it should be pointed out that a
binary operation A : P x P --t P on a poset is almost never residuated
(when P x Pis endowed with the componentwise order). For example,
the binary meet operation 1\ : L x L --t L of a lattice L with more
than one element does not even preserve binary joins, hence cannot
be residuated: indeed, choosing a=/= b in L, one obtains
1\((a,b)V(b,a)) = 1\((aVb,bVa)) = aVb =/= al\b = (al\b)V(li'Ja) =
I\( a, b) V 1\(b, a).
Algebraic residuation theory originates from at least two sources:
1. The investigation of the "laws of thought", due to Boole and com-
pleted by Schroder. Boole's vague observation that the fundamental
operations of joining and intersecting should be somehow "invertible"
was made precise by Schroder who invented the concepts of "minimal
difference" and "maximal quotient", which are close to the residuals

for translations coming from the lattice operations of join and meet.
The first section of this chapter is devoted to these ideas.
2. The investigation of relationships between ideals of rings and frac-
tional ideals of fields, as developed by Dedekind in his famous Sup-
plement XI to Dirichlet's Lectures on Number Theory [27]. It would
require too much space to survey Dedekind's ideal theory here. In-
stead, we shall concentrate, in the second section of this chapter, on
his basic (and more general) ideas concerning partially ordered groups.
In the third section, we shall briefly review Krull's pioneer work on
abstract ideal theory. In the fourth section, we sketch the theory of
residuated lattices, due to Ward and Dilwort h. The fifth section con-
tains a short survey of more recent aspects of residuation theory as
presented in the monograph by Blyth and Janowitz [16]. Specializa-
tion to locales, i.e. quantales whose multiplication is the binary meet,
leads us to the current vital theme of pointfree (but not pointless)
topology.
4.1 From Boole to Schroder: The Logical Calculus

To 1ae odJTo '&11a i:nraexcw TE K,a~ f.1TJ i:nraexnv
f!xMJI/aTOV
,.., ' '
TW O!VTW K,(Y~ K,O!TO!
J~'
TO O!VTO
v '
...
AvT7J biJ 1fa<7wv t<7T~ (3E(3mwTaT7J Twv l!xexwv ... It is
impossible that the same quality should both belong and not
belong to the same thing . . . This is the most certain of all
principles ... Aristotle, Metaphysica, III.3.
Quoted by George Boole in The Laws of Thought (1854)
George Boole is known as the pioneer in the development of formal

logic and its links to algebraic operations. Though the mathematical
theory of Boolean algebras carries his name, he did not invent exactly
that algebraic theory (see below), and a great part of his work is of a
more philosophical than a purely mathematical nature. In his essay
from 1847, "The Mathematical Analysis of Logic" [17], and later in
the famous monograph from 1854, "An Investigation of the Laws of
Thought" [19], Boole developed a broad spect rum of ideas with great
influence on later investigations of the mathematical foundations of
96 M. Erne
logical thinking, and so of lattice theory. It is not the place here

to report Boole's reflections in adequate length, but a few aspects
pointing to certain rudiments of adjunctions should be mentioned in
our historical expedition. For a comprehensive discussion of Boole's
theories of logic and probability, we recommend the monograph by T.
Hail perin [69].
Boole's style is a mixture of syntactic conclusions and semantic heuris-

tics, and sometimes of a priori postulates with a posteriori justifi-
cations of certain rules and methods by a wealth of examples. Let
us try to abstract the main mathematical postulates and deductions
from his work (without going into details of semantic interpretation).
Boole introduces two operations (connecting certain objects of think-
ing): the "combination" of x andy, written simply as juxtaposition xy
and describing the "class of things to which the names or descriptions
represented by x and y are simultaneously applicable"; and the "ag-
gregate" or "conjunction" of x andy, written additively as x + y and
describing the class of all things that belong to x or toy (respectively,
have one of the two properties described by x and y). Two aspects of
Boole's reasoning should be pointed out: first , for him, such "sums"
are meaningful or "interpretable" only if the summands are disjoint ,
i.e., no object has the properties x and y in common. Thus, one is
tempted to say that, in contrast to the total operation of "multipli-
cation" xy, Boole's "addition" x + y is only a partial operation. This
already shows that a formalized theory comprehending Boole's basic
ideas is not Boolean algebra in our sense (where the addition is a to-
tal operation, too) . From a modern perspective, roughly speaking one
might say Boole considered that fragment of Boolean algebra in which
the join operation of Boolean lattices coincides with the addition of the
associated Boolean rings, introduced and studied much later by M. H.
Stone [137]. But probably this is not what Boole had in mind: he freely
carried through calculations involving terms containing sums that vio-
late the disjointness postulate, calling them "not interpretable" , until
he arrived at a term that is "logically interpretable". This is the most
controversial part of his theory, appreciated by some commentators as
"highly ingenious", by others as "dark and mysterious" .
The second, somewhat surprising aspect is that Boole paraphrases the

addition both by the word "and" (applied to substantives, as in his
example "men and woman"), but also by the exclusive "or" (applied
to adjectives, as in 11beings that are men or women"). At first glance,
this simultaneous use of the words "and, or" seems to conflict with
our interpretation of the meet (intersection) described by the particle
11
and", while join (union) is described by the particle 11or". The in-
tuition that 11and" is in some sense dual to "or" was made precise by
Schroder, who distinguished between extent and intent of a concept
(see Sections 1.5 and 2.6); here, the conjunction of attributes (form-
ing the intent) corresponds to the intersection of the classes of objects
having these attributes (the extent). As indicated in Section 1.5, this
is a central aspect of formal concept analysis [59].
For the basic two operations, Boole explicitly postulates the law of
commutativity, while the law of associativity is not mentioned at all
(perhaps because juxtaposition and conjunction of more than two
members usually are written or phrased without parentheses) . As
a fundamental law of logic, Boole emphasizes the law of idempotency
xx = x, written as x 2 = x (and referred to as the "Law of Duality" by
Boole himself ; however, that terminology is not very suggestive and
was criticized by Schroder [129]). Thus, the Boolean multiplication
(juxtaposition) obeys the identities valid in what nowadays is called
a semilattice. For BooZe's addition, however, the dual law x + x = x
plays no role, in view of his hypothesis that the classes occurring in an
"interpretable" sum have to be disjoint. The next fundamental axiom
is the distributive law
z(x + y) = zx + zy,
which Boole motivates by examples. Then he introduces a third (par-
tial) operation he denotes by the subtraction symbol "-" . From the
point of adjunctions, this operation is of particular interest. Boole's
own comment:
But the very idea of an operation affecting some positive

change seems to suggest to us the idea of an opposite or
negative operation, having the effect of undoing what the
former one has done.
Accordingly, Boole gives the interpretation "except" to the subtraction

symbol "-". Shortly after, he states that from x = y + z one may
deduce the equation x - z = y: again the idea of shifting parts of
98 M. Erne
an equation from one side to the other! One should remark that the
equivalence
( ~) X - Z = y {::} X = y+Z
holds only under the assumption that "z is subsumed under x ", which
may be described by any of the equivalent identities
z = zx, z - zx = 0, z(l - x) = 0,
using Boole's symbols 0 for "Nothing" and 1 for "the Universe", sub-
ject to the rules
Ox = 0 and lx = x.
But Boole prefers to choose an "indetermined class variable" v and
expresses the above situation by the equation
z = vx.
Dealing with equalities exclusively, Boole does not mention the so
defined order of subsumption, all the less the symbolic notation z :S
x. Nevertheless, the equivalence (~) may be regarded as a special
instance of an adjunction, viz. for the equality relation as order. A bit
problematic is the domain of definition for the operations involved in
that adjunction. If one wishes to avoid partial operations, one has to
consider, for fixed z, the mutually inverse isomorphisms
X H X - Z and y H y+Z
between the intervals
[z, l] = {x: z :S x} and [0,1 - z] = {y: y :S 1 - z}.
Of course, this is not only a (perfect) adjunction with respect to the
identity relation = , but also with respect to the order relation :S.
An alternate (but not Boole's) point of view is to regard y + z as the
join and x - z as the relative complement x z' in a Boolean algebra B.
Then the equivalence
x-z:Sy {::} x:Sy+z
shows that we have an adjoint pair of maps x H x- z and y H y + z,
defined on the whole of B, whose surjective restrictions yield the above
mutually inverse isomorphisms. A third (and perhaps the most con-
vincing) alternative would be to take for + the addition and for - the
subtraction in a Boolean ring, in which case addition and subtraction
of a fixed element again are mutually inverse isomorphisms, this time

defined on the whole ring.
Boole claims that one must have not only the distributive law for the
addition but also that for the subtraction,
z(x- y) = zx- zy .
And it is not difficult to observe that the equivalence (~) together
with distributivity of the multiplication over addition formally (not
only intuitively) entails distributivity over subtraction: under the as-
sumption that y is subsumed under x, the equivalence (~) yields
x = (x- y) +y and then z (x - y) + zy = z (( x - y) + y) = zx ,
hence
z (x - y) = zx - zy.
A central point in Boole's reasoning is the fundamental law of contra-

diction,
x(1- x) = 0,
for which he quotes the sentence by Aristotle at the beginning of this
section, and which he deduces from the law of idempotency, written
in the form x - x 2 = 0 or x1 - x x = 0. The law of contradiction is
then the starting point for Boole's famous construction of the normal
("developed") form of any Boolean expression, involving an arbitrary
number of variables, composed by multiplication, addition, subtrac-
tion and division. Explicitly, Boole notes the result for functions in
one, two or three variables, in particular:
+ f(1 , O) x (1-y
f(x, y) = f(1, 1) x y
+ f(O, 1)(1- x)y + f(0,0)(1- x)(1-y).
In the subsequent chapters, Boole makes extensive use of the Normal
Form Theorem in his methods of reduction and solution of logical
equations and the elimination of variables. Although these methods
doubtless are ingenious and have kept their importance until today,
a rigorous demonstration of their validity is often missing in Boole's
original work.
Apparently, Ernst Schroder was the first to succeed in putting the
whole theory, now called Boolean algebra, on an axiomatic base allow-
100 M. Erne
ing for exact proofs and conclusions. In the preface to his early 1877
treatise "Der Operationskreis des Logikkalkuls" [129] he writes (free
English translation):
The Boolean theory still suffers from certain incomplete-

ness deficiencies. Among the lacks in BooZe's otherwise
most admirable and attractive method I have remarked,
I would like to mention that BooZe includes an entirely
alien element into his investigations . . . the whole ballast
of algebraic numbers. Invoking the latter indeed forces one
to give up the interpretability of the intermediate steps of
calculation, because one must compute with symbols like
2, -1, ~, ~, which do not admit any logical interpretation.
While it remains a hopeless undertaking to understand the
meaning of each executed single operation, one is led, in a
surprising but mentally not satisfactory way, to the desired
and, indeed, in its own fashion correct result.
In fact , Boole also considered rational functions containing quotients,

which, superficially, do not admit any interpretation by logical vari-
ables. For example, in order to solve the equation
x = yz
for z , he wrote down the developed form
z = ~ = xy + ~x(1- y) + 0(1- x)y + §(1- x)(1- y)
claiming that the quotients § and ~ would have a "very important
logical interpretation", namely:
... The symbol § indicates that a perfectly indefinite portion

of the class, i.e. some, none or all of its members have to
be taken. Any other symbol [than 0, 1 or §] as coefficient
indicates that the constituent to which it is prefixed must
be equated to 0. It follows hence that if the solution of a
problem, obtained by development, be of the form
w =A+ OB + §C + ~D ,
that solution may be resolved into the following equations,
vzz.
w = A + vC, D = 0.
In the above example, this leads to the correct solution

z=xy+v(1-x)(1-y) , x(1-y)=O
with an indeterminate variable v. Of course, the simpler equation
z = x + v(1- y) together with x(1- y) = 0 would work as well.
In a metaphor, one might say that Boole virtuously juggled with the
balls of logic, throwing them from one hand to the other but not caring
much about what happened with them in the air, as long as he was
able to have the desired result in hand at the end. On the contrary,
Schroder carefully observed the balls during their flight through the
a1r.
In his "Operationskreis " [129] Schroder presents a semi-formal system
for Boolean logic, consisting of the "commutation laws" and "asso-
ciation laws" for the class operations of "sum " and "product " (inter-
pretable either as logical or as set-theoretical operations) , the laws of
idem potency, one "distribution law" (the "obvious one" known from
ring operations), and the rules aa' = 0 , a+ a' = 1 for the existence
of complements a' (Schroder writes a 1 instead of a' and calls it the
negation of a). Concerning the four rules
Oa = 0, 1 +a= 1
1a =a, 0 +a= a
Schroder carefully notes that the first two of them follow from the as-
sociation laws, the idempotency laws and the complement laws, while
he believes that the other two cannot be derived from the axioms,
though being mutually equivalent, on account of the equations
1a = a1 = a (a + a') = aa + aa' = a + 0 = 0 + a .
And he is right: taking for addition and multiplication the same semi-
lattice operation, one gets 0 = Oa = 0 + a = a , which is impossible if
the semilattice has more than one element (but this argument is not
Schroder's) .
Besides many further equations valid in Boolean logic, he derives the
absorption laws
a+ (ab) =a and a( a+ b) =a
102 M. Erne
from the distribution law and the rules for the universal bound 1:
a(a +b) = aa + ab =a+ ab = al + ab = a(l +b) = al = a;
and now he concludes that the dual distribution law must hold, too:
(a + b) (a + e) = aa + ab + ae + be = a + ae + be = a + be.
After having proved and applied de Morgan's laws
(a V b)' = a' 1\ b' and (a 1\ b)' = a' V b'
he states in conclusion that (in contrast to Boole's original system)
the whole calculus obeys the rule of dualization: every true formula
remains valid when sum and product are exchanged at all occurrences
in the formula.
For the historical origins of residuation theory, the fourth section of
t he "Operationskreis" is of major relevance. Schroder points out that
in the logical calculus with distribution and complements, the dual
equations
x +b= a and xb = a
if solvable, do possess certain extremal solut ions. Namely, x + b = a
is solvable if and only if a'b = 0, in which case the largest solut ion
is x = a, while the smallest solution is the "minimal difference" x =
a - b = ab'. Dually, the equation xb = a is solvable iff ab' = 0, and
then the smallest solution is x = a, while the largest solution is here
the "maximal quotient" x = ~ =a + b'.
Schroder derives a whole series of rules for these two "principal values
of subtraction and division"; but he does not proceed explicitly to
the final idea of residuation theory, namely to search for t he extremal
solutions of the inequalities x + b 2: a (which could be written as an
equation, e.g. (x + b)a = a) and xb S a (equivalently, x b +a = a)
which always exist and are given by x = ab', respectively, x = a + b'.
In his later work, Schroder singled out the absorption laws as "partic-
ularly characteristic for the logical calculus" and included them into
the list of axioms but dropped the distribution and complement laws
from the axiomatic framework of his "logical calculus". He was familiar
with the early (and fragmentary) set- and lattice-theoretical ideas of
the Grassmann brothers [66, 67], who already used t he symbols U and
n and emphasized their dual behavior, in particular the equivalence
of the equations
an b =a and aU b = b,
both expressing that a be "included" in or "subsumed" under b. In
his extensive Lectures on the Algebra of Logic [131], Schroder often
returned to that su bsum ption relation as a basic concept for logic (or,
interpreted as set inclusion, for the theory of classes). Curiously, as
a forerunner of the inclusion symbol ~ ' he chose the Euro symbol =€
for subsumption. Comparing the two pioneers of lattice theory with
each other, one might say that Dedekind preferred the operational ap-
proach, while Schroder later emphasized the relational (hence order-
theoretical) tools. His basic principles for the subsumption are reflex-
ivity and transitivity; instead of demanding the law of antisymmetry,
he defines equality by the equivalence
a = b : {::} a =€ b and b =€ a
(later a common use in set theory). The existence of a least element 0
and a greatest element 1 are postulated verbally (rather than by the
formulas 0 =€ a and a =€ 1) . The definitions of (binary) join and meet
are again divided into a formal part:
if c =€ a and c =€ b then c =€ ab ,
if a =€ c and b =€ c then a + b =€ c,
and a verbal part, describing ab and a + b in the usual naive set-
theoretical sense of intersection and union by means of the word "Ge-
biet" (domain, field, area) . The "logical calculus" defined in that
semi-axiomatic way apparently is what later would become the order-
theoretical definition of (bounded) lattices. In order to obtain the full
"identical calculus" (of Boolean algebra), Schroder adds a weak form
of the "distribution law ":
if be = 0 then a (b + c) =€ ab + ac
and the defining condition for complements
aa' =€ 0 and 1 =€ a+ a'
together with the postulate that for each a at least one such a' exists,
"obtainable as remainder by omitting a from 1".
Schroder knew and appreciated Peirce's attempts to formalize logi-
cal thinking [110]. A purified list of order-theoretical postulates for
104 M. Erne
lattices (as formal systems of logic) was given later by Huntington

[76], however only in connection with additional postulates defining
Boolean algebras. One of t hese axioms is t he following "syllogistic
principle" used by Peirce [110] :
if a =€ b' is false then there is an x =/:- 0 with x =€ a and x =€ b,
which together with the postulate bb' =€ 0 (and the meet properties of
ab) is easily seen to be equivalent to
ab = 0 {:::? a =€ b' {:::? ab' = a ,
the later characterization of pseudocomplements in the sense of Glivenko
[65] and Frink [55]. This is now a very interesting axiom, both in view
of the historical controversy about Schroder's distributivity postulate
(provable or not?) and the early occurrence of an adjunction. In fact,
the equivalences
ab =€ c" {:::? ba =€ c" {:::? bac' = 0 {:::? b =€ ( ac')'
show that under the assumption c" = c (which follows from uniqueness
of complements and was derived by Schroder from the distribution
law), each of the unary meet operations
Aa: b H ab
is left adjoint to the "relative pseudocomplementation map"
r2a : c H (ac')'.
And by the general fact that left adjoint maps preserve joins, this
yields the distribution law
a( b + c) = ab + ac
and moreover its infinitary variant for all existing joins. The same clue
may be used to give a very short proof of the famous Glivenko-Frink
Theorem on the
Booleanization of Pseudocompletemented Semilattices
The skeleton
S' = {a' : a E S} = {a E S : a" = a}
of a pseudocomplemented semilattice S is always a Boolean lattice
(with a' as complement of a, ab as m eet and and (a'b')' as join of
a and b).
(Compare the concise argument above with t he rather complicated
proof of the distributive law in [68].)

Using the aforementioned "syllogistic principle", Peirce had deduced
the distributive law; later, corresponding with Schroder, he admitted
that the conclusion would be false, because of Schroder's counterexam-
ple (see Section 2.6). But later still, Peirce added a marginal remark
saying his proof was not an error at all. Again half a century later,
in a footnote of [3], Birkhoff stated that "Schroder corrected the er-
roneous impression of Peirce that all lattices were distributive". But,
as Mehrtens [92] has pointed out, Peirce did not deal with any axiom
system of general lattices but based his conclusion on the above "syllo-
gism". Nevertheless, Schroder certainly was right when he claimed to
have shown that the distributive law does not follow from the "logical
calculus", that is, from the axioms of lattice theory.
Let us summarize in a nutshell the role of adjunctions and Galois
connections in the early lattice-theoretically flavored logical investiga-
tions:
Logical Rules Derived from Adjunctions and Galois Connec-
tions
(1) Any partner '1/J : x H x' of a Galois connection on a lattice (e.g.
any pseudocomplementation) converts joins to m eets, hence satisfies
de Morgan's law
(X V y) 1 = X1 1\ y' ,
while the dual law does not follow in general.
(2) Any left adjoint (residuated} unary operation y Hay obtained from
some binary operation (x, y) H xy by fixing one argument preserves
joins, hence satisfies the distributive law
a(x V y) = ax V ay.
4.2 Dedekind's Invention of Lattice-Ordered Groups
Although the name "lattice-ordered groups" or "l-groups " was not used
by Dedekind , he invented that important modern concept and discov-
ered its basic properties. Of t he many interest ing aspects of the later
extensively elaborated theory of l-groups (see e.g. Birkhoff [3] for a
small excerpt), we shall touch on t hose only which occur in Dedekind's
106 M . Erne
work and provide the origin of the theory of residuated lattices. Al-
ready in the first (1897) paper on his "Dualgruppen" (see [30] and
Section 2.4), Dedekind defines abelian groups exactly the same way
we do, but writes them always multiplicatively. Moreover, he consid-
ers such groups endowed with an extra semilattice operation, denoted
by + , and postulates the distributive law
a (b + c) = ab + ac.
In order to give a succinct version of Dedekind's basic theorem on
such structures, let us (historically) anticipate the notion of a partially
ordered group, that is, a group G with a partial order such that left
and right translations are order isomorphisms, hence left and right
adjoint to the inverse translations:
ax ~ y {::} x ~ a- 1 y and xa ~ y {::} x ~ ya- 1 .
As in Section 2.5, we write V instead of+ in order to avoid confusion
with other possible additions. Dedekind's crucial observation is that
a 1\ b = ab / (a V b)
defines a second operation such that (G , V, 1\) becomes a lattice ( "Du-
algruppe"). His proof is entirely algebraic, but of course, one may
also invoke the fact that inversion is a dual automorphism. Hence,
x H ax- 1 b is a dual automorphism, too, and consequently a( a V b) - 1 b
must be the meet of aa- 1 b = b and ab- 1 b = a. Thus, in the non-
commutative case, the above equation has to be replaced with
a 1\ b =a( a V b) - 1 b
(see Birkhoff [3]). Now , the non-commutative version of Dedekind's
theorem and an extension to infinite joins and meets reads as follows:
Charcterization Theorem for Lattice-Ordered Groups
For any partially ordered group G, the following statements and their
duals are all equivalent:
(a) The join a V e exists for the neutral element e and each a E G.
(b) G is a semilattice with (a V b)c = ac V be.
(c) G is a semilattice such that c= 1\B implies
av c= 1\{av b: b EB}.
(d) G is a distributive lattice (with the underlying order).
Proof. (a) :::} (b). Being order isomorphisms, the translations pre-
serve joins. Thus, ((ab- 1 ) Ve)b = abb- 1 V eb = aV b exists, and (av
b)c = ac V be.
(b) :::} (c). If c = 1\ B and x ~a V b for all bE B, then c ~ b entails
a V b =a V b V c and c ~(a V c) 1\ b =(a V c)(a V b) - 1 b ~(a V c)x - 1 b,
hence x(a V c)- 1 c ~ b for all bE B. Consequently x(a V c) - 1 c ~ c, and
finally x ~a V c. This proves the equation aV c=f\{aV b: b EB}.
(c) :::} (d) :::} (a) and the self-duality of the statements are clear, be-
cause inversion is a dual automorphism. D
The implication (b) :::} (d) for commutative groups and the follow-
ing aesthetic proof is due to Dedekind. Iterated application of the
lattice identities and the commutative and distributive laws for the
multiplication yields
(a V b V c) ( ab V be V ca) = (a V b) (b V c) ( c V a),
an identity Dedekind uses at some places in his general ideal theory.
Now,
a 1\ (b V c) = a(b V c) = a(ab V be V ca) = __!l;Q_ ac _
aVbVc (avb)(avc) aVb v ave -
(a 1\ b) V (a 1\ c).
In view of the preservation of all existing joins and meets by the unary
lattice operations x H a V x and x H a 1\ x, one might guess that they
could possess adjoints (as the multiplicative translations do). But
that is not the case, by lack of universal bounds for any non-trivial
lattice-ordered group. Again, it was already Dedekind who observed
that such a group cannot possess any elements of finite order (see E.
Noether's comment in the 1931 edition of Dedekind's Collected Works
[31]).
4.3 Wolfgang Krull's Ideal Domains, and Quantales

In the twenties of the past century, Wolfgang Krull created with in-
genuity an abstract framework establishing the formal background for
the ideal theory developed before by Dedekind, Noether and Artin, but
also for many other mathematical structures (like locales or frames)
that became of interest only decades after Krull's pioneering work.
108 M. Erne
The commutative case is treated by Krull in his 1924 paper [82] , and
the non-commutative case in his 1928 paper [83]. We shall discuss only
a few basic definitions and facts from the first two chapters of the lat-
ter, which extends much of the material contained in the former. The
later chapters, being of minor relevance for the idea of residuation ,
will not be commented on here. Krull 's principal idea is the abstrac-
tion from concrete ring- and ideal-theoretical situations to a structure
whose "individuals" are not the elements of a ring, but the "ideals"
(in a generalized sense) themselves. So he writes in the introduction:
Wir legen nicht etwa einen nichtkommutativen Ring zu-

grunde, aus dessen Elementen die in Betracht kommenden
!deale gebildet werden; wir gehen vielmehr von einem "Ide-
albereich" aus, dessen Individuen die !deale selbst darstellen,
die fiir uns - unabhiingig von ihrer sonstigen Eigenart -
allein durch die fiir sie giiltigen Verkniipfungsregeln (Teil-
barkeitsverhiiltnisse, Multiplikation) charakterisiert sind. Das
Vorbild fiir eine derartige Behandlung der Idealtheorie findet
sich bei Dedekind.
Thus, Krull shows that it is possible to develop the ring-theoretical

notions of prime (ideal), radical, isolated component etc. in his very
general abstract setting of "ideal domains ". Though he does not point
out explicitly the role of adjunctions in his theory, the tool of quotients
(residuals) is fundamental and indispensible for a great part of his
constructions and conclusions.
Krull axiomatically introduces an ideal domain as a set with a partial
order (called divisibility relation) such that any two elements a, b have
a least common multiple (lcm) a n b and a greatest common divisor
(gcd) a+ b, and moreover, even arbitrary (not necessarily finite) sub-
sets have a gcd. So far, this just defines a complete lattice in the sense
of Birkhoff. Furthermore, Krull postulates an associative (but not
necessarily commutative) extra multiplication that distributes from
both sides over arbitrary gcd 's and has the "two-sided ideal property",
meaning that the least common multiple always divides the product
of two elements. Of course, this nomenclature is still in the tradition
of arithmetic and algebra. But the divisibility relation of concrete ide-
als is dual to set inclusion. Hence, in modern t erminology, one would
refer to the dual notions and speak of a complete (two-sided) residu-

ated semigmup if one has a complete lattice with an extra associative
multiplication that distributes over arbitrary joins (denoted by the
symbol V) and is dominated by the binary meet operation:
a·VB=V(a·B), VB·a =V(B·a), and a·b~al\b.
A short modern name for structures satisfying the previous two dis-
tributive laws is quantale, and quantales with the third property a·b ~
a/\ b are said to be two-sided (see Rosenthal [120] for a comprehen-
sive discussion of quantales and their applications). From the his-
torical point of view, it is remarkable that Krull was one of the first
mathematicians dealing with arbitrary joins (or meets) in an algebraic
context, but he resisted introducing a symbolic notation for such an
infinite operation, writing the left distributive law in the following
form:
a· (b1 + b2 + ... ) = a· b1 + a· b2 + ...
Apparently, the "finitary" standpoint in algebraic contexts was quite
strong at that time and also in the subsequent decades (cf. the various
editions of van der Waerden's Algebra [142]).
The crucial definition (with regard to residuation theory) is now that
of left quotients b- 1a and right quotients ab- 1 by
b- 1a=V{cibc~a} and ab- 1 =V{clcb~a}.
Krull notes at once that these suprema (gcd's in his terminology)
are in fact maxima (least elements with respect to his dual order of
divisibility), which could also be expressed by the typical adjunction
equivalences
ab ~ c {:::} a ~ cb- 1 {:::} b ~ a- 1c.
Altough the use of the exponent symbol - l is perhaps a bit dangerous,
because actual inverse elements rarely will exist, the notation is quite
convenient and rather stable, on account of the rules
b- 1(c 1a) = (cb)- 1a, (ac 1)b- 1 = a(bc) - 1, (b- 1a)c- 1 = b- 1(ac- 1) ,
so that one could even omit some parentheses and write b- 1ac 1 for
the last term. Krull notes without proof some further elementary rules
like
b- 1(1\A) = l\{b- 1a I a E A},
110 M. Erne
a(V B) - 1 = (\{ab - 1 I bE B},

referring to Dedekind's Supplement XI [27] for the case of modules,
and to his own paper [82] for the general commutative case.
In the second chapter, Krull gives an elegant proof of a first "prime
ideal theorem". He defines prime ideals (or simply primes) p by the
usual condition that ab ::; p implies a ::; p or b ::; p but does not require
explicitly that a prime should be distinct from the top element. The
primes p with p ~ a are referred to as the primes of a. (Recall that
our ~ is denoted by ::; in Krull's paper and is regarded as a divisibility
relation; accordingly, he speaks of a "highest prime (ideal) of a if he
means a minimal prime p ~ a in our sense) . The theorem in question
provides the first step towards the derivation of prime decompositions
from suitable chain conditions.
Theorem on Minimal Primes
Suppose Q is an ideal domain (two-sided quantale) satisfying the fol-
lowing quotient chain condition: there is no properly descending se-
quence of elements a1 such that a1+1 is always a left or right quotient
of a1 . Then, for each a E Q, there is a finite antichain of primes
P1, ... , Pn ~ a such that (Pt·. ·PnY ::; a for suitably large exponents r.
These primes are precisely the minimal primes of a .
Proof. As usual, an antichain is a set of pairwise incomparable ele-
ments; thus, p1 ::; Pk is excluded for j =J. k. Krull's proof is similar
to the common argument leading to prime decompositions for natural
numbers or elements of distributive lattices with chain conditions, but
a bit more subtle. It suffices to show that if an element a fails to
have the required property then so does a smaller quotient of a (this
would then produce a properly descending chain of quotients). Un-
der the contrapositive assumption on a, this element cannot be prime,
so there exist elements b, c with be ::; a but b 1;. a and c 1;. a. The
quotients a 1 = ac- 1 and a2 = a 1 - 1a satisfy a 1 a2 ::; a but a < a 1 and
a < a2 . (Indeed, ac ::; a implies a ::; ac 1 = a 1 , while the assumption
a 1 ::; a together with be ::; a would entail b ::; ac- 1 = a 1 ::; a; and
similarly, a 1 a ::; a implies a ::; a 1 - 1a = a2 , while a2 ::; a together with
a 1 c::; a would entail c::; a 1 - 1a = a2 ::; a.) If either a1 (j = 1, 2) would
have the required property (Pj,l···Pj,nJrj ::; a1 with suitable pairwise
incomparable primes Pj,k, then the minimal elements in the union of
these two antichains would form a finite antichain {p1 , ... , pn} with
(p 1 .•. pny 1 +r2 ~ a 1 a 2 ~ a, in contrast to the assumption on a. Hence,

either a 1 or a 2 fails to have the property in question. That a finite
antichain {p1 , ... , Pn} of primes Pi 2: a with (p 1 ... pnY ~ a must contain
all minimal primes p of a is clear, because (p 1 .. ·Pn Y ~ p entails Pi ~ p
for some j and then Pi = p by minimality of p. On the other hand,
each Pi must be a minimal prime of a, because a ~ p ~ Pi implies, as
before, Pk ~ p ~Pi for some k, so that p =Pi= Pk· D
Under the quotient chain condition, any sequence built by elements of

the form ab-n = a(bn)- 1 becomes stationary; its finally constant value
is denoted by ib(a) and called a right isolated component (ideal) of a.
The following equations for the functions ib are easily verified:
and ib preserves finite meets. Analogously for left isolated components.

Another useful result, stated as Theorem 2, is the following:
br ~ c implies ib 2: ic.
Indeed, induction gives ab- n 2: a en, and for n sufficiently large,
ab-n = ib(a), acn = ic(a). Applying this to the above t heorem,
Krull obtains the
Theorem on Isolated Components
If c is a product of all minimal primes of b (in some order) then ib = ic.
Adding the hypothesis that there is no ascending chain of primes,
Krull now develops a purely abstract prime (ideal) theory, covering
much of the material known from ring theory.
From the perspective of adjunctions, it might be of interest that Krull's
famous Separation Lemma (also in the abstract version for quantales)
is equivalent to the Boolean Prime Ideal Theorem, as shown in [9] -
and that the main ingredient in the proof is again a certain residuated
map. In the classical ring-theoretical setting, the Separation Lemma
says that
every ideal disjoint from a non-empty multiplicative subset M is
contained in a prime ideal disjoint from M.
For the general quantalic version (which covers the above and many
other separation lemmas, see [51]), one needs the notion of open m-
filters; these are subsets S of a quantale Q such that
112 M . Erne
(1) a, bE Sand ab ~ c imply c E S (m-filter pr-operty) , and

(2) every directed set with join in S already intersects S (openness).
Separation Lemma for Quantales
Any element a outside an open m-filter S in a two-sided quantale Q
has a prime outside S .
Of course, this is easily shown with the help of Zorn's Lemma - but
the point is that it can be derived from (and is therefore equivalent
to) the
Boolean Prime Ideal Theorem
Every non-trivial Boolean algebra contains a prime ideal.
The main tool in the proof is the map s : Q ---+ Q defined by
s (a) = V{x E Q I V y E Q (x V y E S ::::} a V y E S)} .
One shows that sis a closure operation on Q whose image Q 8 = s(Q)
is a locale (that is, a quantale whose multiplication is the binary meet).
Hence the corestriction s : Q---+ Qs is left adjoint to the inclusion map
from Qs into Q. Moreover, the top element e of Q is compact in
Q 8 , and its preimage s<-({e}) is the original open m-filter S. Now,
the Prime Element Theorem, known to be equivalent to the Boolean
Prime Ideal Theorem [8], yields a prime in Q 8 distinct from e, which
is then also a prime for a in Q outside S.
4.4 Morgan Ward's and R. P. Dilworth's Residuated

Lattices
A few years after Krull's discoveries, independently Morgan Ward and
his student Robert P. Dilworth dealt with the same theme. Their work
may be regarded as the first systematic investigation of residuated op-
erations. They became acquainted with Krull's work only after they
had finished a great part of their investigations. The overlap is not
broad, but of course, residuation plays a common central role in both
theories. While Krull focussed on notions of primeness and decompo-
sition theorems, Dilworth and Ward were more interested in general
questions of existence and uniqueness of residuated multiplications
and in the consequences thereof. In a retrospective, Dilworth writes:
Mor-gan Ward ... was also impressed with the usefulness of

residuation in the study of ideals in a ring. When I began
my graduate work under his direction he suggested that I
first make an axiomatic study of residuation over a lattice
. . . On the basis of this work, he and I became convinced
that much of the theory of ideals of a commutative ring
could be carried out for lattices with a multiplication (resid-
uated lattices). W e were able to carry through much of the
decomposition theory but were unable to get the deeper re-
sults of ideal theory. The main obstacle was obtaining a
suitable notion of 'principal element' in a multiplicative
lattice. Our attempts were only partially successful. It was
many years later {1962) that I discovered the right general-
ization and began the deeper theory in my paper 'Abstract
Commutative Ideal Theory'.
From a whole series of early papers on the subject (see (35, 36, 146,
147, 148, 150, 151]) we pick here the one that provided the most signif-
icant progress at its time (and covered the main results from its fore-
runners), viz. the paper entitled Residuated Lattices (150] from 1938.
The main purpose in t hat article is, similarly as in Krull's first paper,
an abstract foundation of the commutative ideal theory init iated by
Dedekind and pushed forward considerably by the impressive work
of Emmy Noether. Generalizing Dedekind's lattice-ordered groups,
Ward and Dilworth consider an arbitrary (not necessarily complete)
lattice L endowed with an extra commutative monoid operation, called
multiplication and operating distributively over the join. Furthermore,
they assume that the greatest element i should coincide with the neu-
tral element for the multiplication. Thus, such a multiplication must
satisfy the rules
a·z = a,
a·(b ·c) = a·(c·b) ,
a· (b V c) = (a· b) V (a· c),
and conversely, these identities (together with a :::; i) imply that one
has a partially ordered commutative monoid with
a/\ b 2:: a·b 2:: (a/\ b)·(av b) .
In particular,
114 M. Erne
aV b = i implies a 1\ b = a·b and aV (b·c) = aV c.

On the other hand, they consider so-called residuations, i.e. binary
operations : satisfying
a: b = i {:::} a ~ b,
(a:b):c = (a:c):b,
(a/\ b) :c =(a: c) 1\ (b:c).
In the tradition of divisor and ideal theory, Ward and Dilworth used
here the notations [a, bJ ("greatest common divisor") for the meet and
(a, b) ("least common multiple") for the join. Furthermore, they wrote
a ::) b instead of ~ for the order relation defined by
a~ b {:::} b ~a {:::} a 1\ b = b {:::} a V b =a.
Only later, they passed to the now common symbols 1\ for the meet, V
for the join, and ~ for the order relation. Also, they did not postulate
the existence of a greatest element explicitly, but most of their con-
clusions work only under that additional assumption. Instead, they
included the following rules in the definition of a residuation:
a ~ b implies a:c ~ b:c and c:b ~ c:a ,
c:(aVb) = (c:a)/\(c:b).
But these rules are easy consequences of the others: first, one derives
from the first two postulates for residuations the useful equivalence
c:a ~ b {:::} c:b ~ a
which say that ('1/Jc, '1/Jc) with '1/Jc(a) = c : a is a Galois connection.

This proves the equation c : (a V b) = (c: a) 1\ (c: b) and, a fortiori,
the implication a ~ b =} c : b ~ c : a. Similarly, the implication
a ~ b =} a: c ~ b: c follows from the equation (a/\ b) : c = (a : c) 1\ ( b: c).
Other consequences of the above postulates are the identities
a:a = i, i:b =i , (al\b):b = a:b=a:(avb),
and the inequalities
a: b ~ a, a: (a: b) ~ a V b,
(a V b): c ~ (a:c) V (b:c) , c : (a 1\ b) ~ (c :a) V (c:b).
Verifying a sequence of technical details, Ward and Dilworth prove the
following remarkable
Correspondence Between Multiplications and Residuations

(1) Every complete lattice L with a multiplication · such that
a·VB=V{a·blbEB} forall B~L
admits a residuation given by c: a 2 b {::} a · b ::::; c.

The same holds if L is merely a lattice but the sets {x I a· x ::::; c}
satisfy the ascending chain condition.
(2) Every complete lattice L with a residuation : such that
1\C:a=/\{c:a I cEC} for all C~L
admits a multiplication given by a · b ::::; c {::} c: a 2 b.

The same holds if L is merely a lattice but the sets {y I y: a 2 b}
satisfy the descending chain condition.
Moreover, the above two constructions are inverse to each other.
These facts and something more can be shown very easily with the
help of adjunctions. Call a commutative partially ordered monoid
(M, ·) residuated if each of the translations Aa : x t---7 a·x is residuated
(in other words, if each of the sets {x I a· x ::::; c} with a, c E M
has a greatest element). Similarly, we mean by a commuting residual
groupoid a partially ordered set M with an operation : such that the
maps f2a : y t---7 y: a are residual (i.e. , the sets {y I y : a 2 b} with
a, b E M have least elements) and commute with each other. The
top element i is said to be discriminating if it determines the order by
a: b = i {::} a 2 b. The operation : is then automatically a residuation
in the sense of Dilworth and Ward. (However, the converse conclusion
is problematic for infinite lattices; therefore, the chosen nomenclature
of "residuation" might be a bit misleading.) The above correspondence
may be now be rephrased as a
Theorem on Residuated Monoids
Sending each residuated monoid (M, ·) to the ordered groupoid (M, :)
with c : a = max{ xI a· x ::::; c}, one obtains a one-to-one correspon-
dence between commutative residuated monoids having a neutral top
element and commuting residual groupoids having a discriminating top
element. The bijection in the opposite direction is obtained by sending
(M, :) to the partially ordered monoid (M, ·) with a ·b = min{y I y: a 2
b}.
116 M. Erne
In fact, a commutative residuated monoid may be regarded as a set

M together with a family of commuting left adjoint maps Aa (a E
M) containing the identity map; on the other hand, a commuting
residual groupoid may be regarded as a set M together with a family of
commuting right adjoint maps r!a (a E M). And the passage between
left and right adjoints,
a· X= Aa(x) 'S Y {::} X 'S r!a(Y) = y:a
gives the desired bijection, because
Aa(i) = a·i =a'S b {::} i 'S r!a(b) = a:b {::} i =a: b.
Note that the "left translations" x f----+ c: x invert the order and trans-
form all existing joins into meets:
a=VB =? c:a=f\{c:blbEB}.
Indeed,
x'Sc:a {::} a'Sc:x {::} VbEB (b'Sc:x) {::} x'S/\(c:B).
The non-commutative case is more complicated (see for example Dil-
worth [36], Blyth and Janowitz [16]).
After having established the correspondence between "multiplications"
and "residuations", Ward and Dilworth give some necessary and some
sufficient conditions for the existence of such pairs linked by adjunc-
tion. They claim the existence of a unique divisor-free element (later
called coatom) would be a sufficient condition. That this is not true in
general is easily seen by looking at the componentwise ordered com-
plete lattice
{(x, 0): 0 'S x < 1} U {(0, 1) , (1, 1)}
with least element 0 = (0, 0) , greatest element i = (1, 1) and unique
coatom c = (0, 1): for a = (a , 0) and b = ((3, 0) with 0 < (3 < a < 1,
one obtains
(a· b) V (a·c) 'S (a 1\ b) V (a 1\ c) =a 1\ b = b < a= a· i =a· (b V c),
no matter how one tries to define the multiplication. Apparently,
Ward and Dilworth mean a sligthly stronger hypothesis, namely that
the unique coatom dominates all elements distinct from the top. In
that case, it is easy to define a residuated multiplication by a· i =
a, i · b = b and a· b = 0 otherwise. This type of examples shows
that no non-trivial lattice identity is necessary for the existence of a

residuation - a remarkable discovery at a time where experts believed
that a reasonable residuation theory might be possible in modular or
similar lattices only (as suggested by the classical case of ideal and
submodule lattices). On the other hand, Ward and Dilworth observed
that certain weakenings of the distributive law
(a 1\ b) V c = (a V c) 1\ (b V c)
are necessary for the existence of a residuation. From the inequality
ab = a· b ~ a 1\ b and the distributive laws, valid in any residuated
lattice, one derives the distributive inequality
(a V c) (b V c) = ab V ac V cb V cc ~ ab V c,
which in turn entails the semidistributive law
a V c = b V c = i ::::} ab V c = i ::::} (a 1\ b) V c = i ,
It follows that each coatom c has to be prime and, a fortiori, /\-prime ,
that is,
a 1\ b ~ c ::::} ab ~ c ::::} a ~ c or b ~ c.
Motivated by the background from ideal theory, Ward and Dilworth
call two elements a, b of a residuated lattice L coprime if aV b = i. As
remarked earlier, this implies ab = a 1\ b and, moreover,
ab V c =(a/\ b) V c = (aV c) 1\ (b V c)= (aV c)(b V c) for all c E L ,
since (av c) V (b V c) = (av b) V c = i entails
(a V c) 1\ ( b V c) = (a V c) (b V c) ~ ab V c ~ (a V c) 1\ ( b V c) .
Now, it is not hard to prove:
Local Distributive Laws for Residuated Lattices
Let A be a finite set of pairwise coprime elements in a residuated lattice
L. Then
1\ A V c = 1\ (A V c) ( = 1\ {a V c I a E A}) for all c E L.
If A does not contain the top element i then the map
cp : 2A ~ p A -+ B = {I\ X I X ~ A}' X H 1\ X
is an isomorphism, and consequently B is a Boolean lattice. In partic-
ular, if A is a finite set of coat oms then the elements above 1\ A form
a Boolean lattice isomorphic to 2A.
118 M . Erne
Proof. The identity 1\ A V c = 1\ (A V c) is obtained from the proven

case A = {a, b} by induction. While Ward and Dilworth establish
distributivity and complementation of B by explicit calculations, we
prefer an easier argument using the Galois connection (cp, 'ljJ ) with
'ljJ : B -+ P A, b r+ {a E A I b :::; a}
(note that just by definition, b:::; cp(X) = 1\X {::} X ~ '1/J (b)) . As cp
is onto, it suffices to prove the inclusion '1/Jo cp(X) ~ X for X ~ A. But
c E A \X and c E '1/Jo cp(X), i.e. 1\X:::; c, would imply c = 1\X V c =
1\ (XV c) = i ~ A. If A consists of coatoms then each c ::::: 1\ A satisfies
c = 1\A V c = 1\(A V c)= 1\{a E A I c:::; a} E B. 0
Years later, Dilworth demonstrated that if each interval of a lattice

has the above property and satisfies suitable chain conditions then
each element has a unique irredundant join-decomposition into join-
irreducible elements (actually, Dilworth proved the dual). In the six-
ties, Crawley and Dilworth were able to weaken the finiteness condi-
tions considerably (see [24, 38]) , as presented in the fascinating book
Algebraic Theory of Lattices [25].
That the local distributivity at the top is not sufficient for the exis-
tence of a residuation can be checked by a look at the free modular
lattice with three generators (see the diagram in Section 2.5). Using
Dedekind's notation, Ward and Dilworth argue as follows. If : was a
residuation then
a0 : d' = (a' 1\ a"') : d' = a' : d' 1\ a"' : d' = a' : d' 1\ i = a' : d' ::::: a'
and analogously b0 : d' ::::: b' , c0 : d' ::::: c'. Now, from h := a0 : d' =
b0 : d' = c0 : d' it would follow that h ::::: a' V b' V c' = d"" = i ,
contradicting a0 £_ d'. But, since it is not clear how to confirm the
claimed equality for h, I propose the following alt ernative. If there
would exist a (residuated) multiplication · then
a'· d' =a'· (b 0 V c0 ) = (a' ·b0 ) V (a' ·c0 )
:::; (a' 1\ b0 ) V (a' 1\ co) :::; d1 ;
similarly b' · d' :::; d1 and c' · d' :::; d 1 , whence
d' = (a' V b' V c')·d' = (a'·d') V (b'·d') V (c'·d'):::; d1 ,
a contradiction.
It should be mentioned that the local distributive law for residuated
lattices cannot be extended to infinite subsets A, as a straightforward
inspection of the cofinite topology on an infinite set shows. Although
this is a complete distributive lattice with ascending chain condition in
which every element is a meet of coatoms, it is certainly not Boolean,
and the infinite distributive law a V 1\ X = 1\(a V X) fails.
While even a finite distributive lattice may possess several residu-
ations, Ward and Dilworth showed that a complemented residuated
lattice must already be Boolean, and the multiplication must be the
meet operation. Since their computations are rather complicated, we
add here a short proof for a slightly stronger result , providing one of
dozens of possible axiomatizations for Boolean algebras (cf. Hunting-
ton [76]).
Characterization Theorem for Boolean Algebras
If a join-semilattice S with least element 0 and greatest element i car-
ries a commutative binary operation · and a unary operation ' such
that
a·i =a, a·a' = 0, a V a'= i , and a· (b V c) = (a· b) V (a· c)
then a· b is the binary meet, hence (S, V , ·, 0, i,') is a Boolean algebra.
Indeed, a V a' V b = i implies
a A b::; a A (a' V b) =a· (a' V b) =a· a' V a·b = a·b.
The next chapter of [150] is devoted to the study of so-called Noether

lattices. These are lattices satisfying the ascending chain condition
and admitting a residuation so that every meet-irreducible element p
is primary, i.e. p 2: a· b implies p 2: a or p 2: b8 for some s. Of
course, this class includes all ideal lattices of Noetherian rings, whence
the name. Ward and Dilworth impressively demonstrate that many
ring-theoretical results, in particular the fundamental decomposition
theorems, remain valid in the general realm of Noether lattices. A
crucial role is played by elements they call principal; such elements c
are characterized by the property that for each a ::; c, there is a b with
a = b·c. In the residuated case, this may be expressed by the equation
a = ( c: a) ·c. It will be convenient to call an element p power-principal
if all powers pk are principal. A central result in the lattice-theoretical
120 M . Erne
abstraction of ideal theory is now the

Theorem on Modular Noether Lattices
Let L be a modular residuated lattice satisfying the ascending chain
condition. Then the following statements are equivalent:
(a) L is a N oether· lattice.
(b) Each element has an (ir-redundant) finite primary decomposition.
(c) For all a, b E L there is a number k such that a/\ bk ::::; ab.
Sufficient for these properties is that
(d) Each element is a join of power-principal elements.
Proof. (a)::::} (b). By the ascending chain condition, each element is the
meet of a finite number of meet-irreducibles, and these are primary.
(b)::::} (c). Represent ab as a meet of primary elements p 1 , .. . , Pn. Then
ab ::::; Pj implies a 1\ bki ::::; Pj for suitable exponents ky, and for their
sum k, it follows that a/\ bk ::::; p 1 1\ ... 1\ Pn = ab.
(c) :::}(a). Let m be a meet-irreducible element and consider a, b E L
with ab ::::; m but a 1: m. Then we have d = a V m > m, and we
find an s such that d 1\ b 8 ::::; db= ab V mb::::; m < d. By modularity,
m = ( m V b 8 ) 1\ d, and as m is irreducible, m = m V b 8 ~ b 8 , proving
primarity of m.
(d)::::}(a): For any power-principal element p::::; b, the ascending chain
condition yields a k such that dp: pk = dp: pk+l_ Now, there is a q
with qpk = d 1\ pk, whence
qpk+l = ( d 1\ pk)p ::::; dp ::::} q ::::; dp: pk+l = dp: pk ::::} d 1\ pk = qpk
::::; dp.
Replacing b with pin (c)::::}(a) , one concludes that m ~ pk. Since
b is a finite join of such power-principal elements Ph the sum s of
the corresponding exponents kj with m ~ Pj kj then has the desired
property m ~ b8 • D
Clearly, every ring ideal is a join of principal ideals, and these are
closed under the formation of powers and even of products (if the ring
is commutative). Thus, the above theorem actually applies to classical
ring theory. Modifying the notion of principal elements, Dilworth later
found still more powerful results (see his own comment and [37]) .
In the last part, Ward and Dilworth deal with conditions under which
a residuated lattice is distributive. They show that the postulates
(a: b) V (b: a) = i
a: (b 1\ c) = (a : b) V (a: c)
(b V c) : a= (b: a) V (c : a)
are mutually equivalent and entail distributivity. Another sufficient
condition for distributivity is provided by the following
Distributive Law for Principal Elements
In a residuated lattice, the unary meet operations x H a 1\ x preserve
all principal joins: if c = V B is a principal element then a 1\ c =
V(ai\B).
Indeed, if d ~ a 1\ b for all b E B then
d:c = 1\{d:b I bE B} ~ 1\{(al\b) :b I bE B} = 1\{a:b I bE B}
= a:c,
hence d ~ (a:c)c =a/\ c.

An immediate consequence is the following
Corollary A complete lattice admits a residuated multiplication such
that all elements are principal iff it is a locale (that is, a complete
lattice satisfying the distributive law a 1\ VB = V(a 1\ B) for all
subsets B).
Of course, the prominent examples of residuated lattices in which all
elements are principal are the ideal lattices of principal ideal domains.
But also in every lattice-ordered group, all elements are trivially prin-
cipal. This yields an alternative proof of Dedekind's theorem on the
distributivity of lattice-ordered groups. But recall that such a group
never can be a complete lattice, lacking a greatest and a least element.
4.5 T. S. Blyth and M. F. Janowitz: Residuation

Theory
The first comprehensive monograph about residuation theory was writ-
ten by T. Blyth and M.F. Janowitz in the early seventies [16], after a
period of
122 M. Erne
approximately 20 years where the general theory of residu-

ated mappings seems to have lain dormant
as the authors write in the introduction. According to the mam

themes, the material is divided into three chapters:
1. Foundations
2. Coordinatizing Baer Semigroups
3. Residuated Algebraic Structures
Besides some basic and some more advanced facts on orders and lat-
tices, the first chapter contains, of course, a section about adjoint
pairs of mappings. The emphasis is here on descriptions of various
properties for mappings between ordered sets by means of principal
ideals. For example, so-called quasi-residuated mappings A are char-
acterized by the condition that inverse images of principal ideals be
order ideals (i.e. nonempty down-sets) , or equivalently, by isotonicity
together with the existence of a map [} in the reverse direction so that
s
Ao [} id. The characterizing equivalence for closure operations, here
termed closure mappings, is written as an equation:
1'+-[y) = 1'+-[!'(y)),
where /'+- [x) denotes the preimage of the principal dual ideal gener-
ated by x. Clearly, kernel maps are characterized by the dual equa-
tion. After these "forerunners", residuated mappings A are introduced
as mappings such that inverse images of principal ideals are again
such, and are characterized by the existence of a map[} in the reverse
direction with AO[l S id and (lOA ~ id. Thus, the map [} is the uniquely
determined right adjoint or residual of A : P --+ Q and denoted by A+.
The map
'D(A) =A+: 'D(Q) -+'D(P)
between the dually ordered sets 'D(Q) = Qd and 'D(P) = pd is then
again residuated, with
'Do'D(A) = A and 'D(Ao cp) = 'D(cp) o 'D(A).
As a consequence of the definitions, the closure mappings are pre-
cis ely the composites of a residuated map with its right adjoint, and
the kernel mappings are precisely the composites in the opposite or-
der. Of basic importance for structural investigations is the fact that
the composition of residuated maps yields again residuated maps, so

that the ordered sets together with the residuated maps form a self-
dual category under the self-inverse endofunctor D . A further useful
property of the duality functor D is that it preserves the pointwise
order between residuated maps. Thus,
A ~ p, {::} D(A) ~ D(p,) ,
or explicitly (observing that D (P) and D(Q) carry the dual order),
Vp E P(A(p) ~ p,(p)) {::} Vq E Q(p,+(q) ~ A+(q)).
Immediate from these facts is the following table, providing the trans-
lation of properties for residuated maps to their adjoints, and vice
versa.
residuated map A residual map (! = A+ identity or
(left adjoint) (right adjoint) inequality
injective surjective (!OA = id
surjective injective AO(! = id
isomorphism isomorphism (!OA = id, AO (! = id
extensive contractive A 2: id, (! ~ id
contractive extensive A ~ id, (! 2: id
idempotent idempotent Au A= A, (!0(! = (!
closure map kernel map A= (!OA, (! = AO(!
kernel map closure map A= AO(! , (! = (!OA
By the above remarks, the residuated maps on a fixed ordered set P

form a (partially) ordered semigroup Res(P) under composition and
pointwise order, and the functor D induces an isomorphism between
Res( F) and Res(Pd).
As indicated in the title, the second chapter deals with a special class
of partially ordered semigroups, the so-called Baer semigroups. They
are the basic tool for the coordinatization of arbitrary bounded lattices
L, i.e., the isomorphic representation of L as the lattice of certain left
or right principal ideals of suitable semigroups or rings. This theory is
closely related to but easier than John von Neumann's famous theory
of continuous geometries [145] (which implicit ly contains some further
interesting examples of Galois connections), but of course also with
Reinhold Baer's work on ring theory and projective geometry (see e.g.
[6]).
124 M. Erne
A left {right) ideal of a semigroup Sis a subset I with S I ~ I (respec-

tively, IS ~ I), and a (two-sided) ideal is both a left and a right ideal.
Since the left (right, two-sided) ideals of a semigroup are closed under
arbitrary unions and intersections, they form an Alexandroff topol-
ogy. Subsets of the form xS = {xy : y E S} and Sy = {xy : x E S}
are referred to as principal left or right ideals, respectively. Denote
by .CI(S) and RI(S) the closure systems of all left and right ideals,
respectively, and by P.CI(S) and PRI(S) the systems of all principal
left and right ideals, respectively.
If K is a left ideal and M is an arbitrary subset of S then the left
residual, quotient or K -annihilator of M is the left ideal
LK(M) ={xES: xM ~ K} (where xM = {xm: mE M}),
and if K is a right ideal then the right residual, quotient or K -anni-
hilator of M is the right ideal
RK(M) ={xES: Mx ~ K} (where Mx = {mx: mE M}) .
Thus, if K is a fixed ideal, the equivalence
I ~ L K ( J) {::} I J ~ K {::} J ~ RK (I)
for IE .CI(S) and J E RI(S) gives rise to an obvious Galois connec-
tion between .CI(S) and RI(S) , but in general not between P.CI(S)
and PRI(S), nor to an isomorphism between .CI(S) and RI(S) . The
relation inducing the polarity in question is given by xRKY {::} xy E
K.
Now, Blyth and Janowitz introduce Baer semigroups as pairs (S, K)
consisting of a semigroup Sand a nonempty ideal K such that the left
and right K -annihilator ideals L K ( x) and RK (x) of singletons {x} are
principal and generated by idempotent elements. It is then easy to see
that S must have a two-sided identity element 1 and that K is of the
form kS = Sk for a central idempotent k (i .e. kk = k and kx = xk
for all x E S). Thus, a Baer semigroup may also be defined as a pair
(S, k) with a central idempotent k such that for each x E S, there
are idempotent elements kx , xk with Lk(x) = S kx and Rk(x) = xkS
(where Lk stands for LK and Rk for RK ). It is then also a custom to
say that S is a Baer semigroup with focus k . Often, k will be a zero
element 0, satisfying xO = Ox = 0. In that case, it will suffice to speak
of the Baer semigroup S. A Baer ring is defined by the postulate
that its multiplicative semigroup be a Baer semigroup with focus 0.

Of course, any monoid with 0 and 1 having no proper zero divisors
is a Baer semigroup (with focus 0). Here, all left or right annihilator
ideals consist of the zero element only. Other important examples of
Baer semigroups are:
(1) Any multiplicative subsemigroup Sofa Baer ring containing all
idempotents; in particular, the multiplicative semigroup of any
regular ring in the sense ofvon Neumann [145]. In such rings, for
each x there is a y with xyx = x; equivalently, each principal left
or right ideal is an annihilator; here, one may take kx = 1 - xy
and xk = 1 - yx).
Examples: all rings with identity having no proper zero divisors
(in particular all division rings), all Boolean rings, all matrix
rings over division rings, and all rings of bounded operators of
Hilbert spaces.
(2) The semigroup R = P(X x X) of all relations on a set X,
with the product RS = {(x,z): 3y (xRySz)} and the empty
relation as fo cus. Here, for S E R , an easy computation gives
R0(S) = S 0 R with 5 0 = {(x, x ) : x E X\ SX} = S 0 S 0 , and
analogously for L 0 ( S).
(3) All pseudocomplemented semilattices. They are just those Baer

semigroups with focus 0 in which all elements are central idem-
potents. Similary, any element k of a Brouwerian semilattice
S (see the next section) gives rise to a Baer semigroup (S, k).
In that situation, Lk(x) = Rk(x) = xk S, where xk is the rela-
tive pseudocomplement of x with respect to k, i.e. t he greatest
element y with x 1\ y :::; k.
(4) The semigroup S = Res(L) of all residuated self-maps of a boun-

ded lattice L, with the constant zero map as focus. Here, the
construction of idempotent generators for annihilator ideals is a
bit more involved. It is easy to check that for any x E L, the
map Bx : L--+ L, defined by Bx(Y) = y for y ::S x and Bx(Y) = x
otherwise, is an idempotent element of S. Now, given arbitrary
(), <p E S, one observes t hat <p E R 0 (0) , i.e. Oo<p = 0, is equivalent
to <p(l) :::; e+(o) = : X. But <p(1) :::; X entails <p = Bxo<p (because
<p(y) :::; x for ally E L) ; and conversely, if <p = Bxo <p then <p(x) =
126 M. Erne
Ox(<p(l)) ::; Bx(l) = x. From this it follows that Ro(O) = Oxo S.

A similar construction gives L 0 (0) = So '1/Jy with y = 0(1) and
~;y(x) = 0 if x ::; y, '1/Jy(x) = x V y otherwise.
Theorem on Ideal Lattices of Baer Semigroups

For any Baer semigroup (S, k),
.ck = Lk(S) = {Lk(x): XES} and nk = Rk(S) = {Rk(x): XES}
are dually isomorphic bounded lattices. The Galois connection (Lk, Rk)
induces mutually inverse dual isomorphisms between these lattices.
Proof. By definition of Baer semigroups, .Ck and Rk are the ranges of
the Galois connection (Lk, Rk) between .CPT(S) and RPT(S). Clearly,
kS = Sk = Lk(l) = Rk(l) is the least and S = Lk(k) = Rk(k) is the
greatest element of .Ck and of Rk. The only detail that remains to
be shown is that the intersection of two members eS, fS E nk (with
idempotent e, f) is again in Rk· This and the corresponding state-
ment for Lk together with the dual isomorphism between those posets
then give the desired lattice properties. First, note that eS, fS E Rk
imply Rk(ke) = Rk(Lk(eS)) = eS and Rk(k!) = fS. Next, put
h =kef and g = fhk. Idempotency off yields hg = hhk E kS and
g = hkg = fg = gg. Now, if y E gS then y = gy = fhky E fS, hence
key = hhky E kS, and so y E Rk(ke) n fS = eS n fS. Conversely,
y E Rk(ke) n fS implies key E kS andy= fy, hence hy =key E kS
and so y = fy E hkS, i.e. y E fhkS = gS. Thus, we have
gS = eSnfS = Rk(ke)nRk(kJ) and therefore gS = Rk(Lk(gS)) E Rk·
0
Generalizing various notions of coordinatizing objects in geometry and

algebra, Blyth and Janowitz say that a Baer semigroup (S, k) coordi-
natizes a poset E in case E is isomorphic to the right k-annihilator
poset Rk(S), hence dually isomorphic to Lk(S). Then they prove the
powerful
Coordinatization Theorem for Bounded Lattices
Any bounded lattice L is coordinatized by the Baer semigroup S
Res( L) with focus 0, i.e. L is isomorphic to R 0 ( S) and dually isomor-
phic to L 0 ( S). Hence, the following three statements on a poset E are
equivalent:
(a) E is a bounded lattice.
(b) Res(E) is a Baer semigroup with focus 0.

(c) E can be coordinatized by a Baer semigroup.
Proof. We have already seen that S = Res(L) is a Baer semigroup,
and that any member of R 0 (S) is of the form R0 (0) = rp o S where
rp is an idempotent element of S. For any such rp, it easily follows
that rp(1) = e+(o). Therefore, one obtains a well-defined map A :
R 0 (S) ---t L by setting A(rpo S) = rp(1) = e+(o). Indeed, each of the
implications
'lj;oS ~ rpoS::::} 'ljJ = rpo'ljJ::::} 'l/;(1) = rp('l/;(1))
::::} 7/J(1) ~ rp(1) = e+(o) ::::} '1/J E Ro(O)
may be inverted because the last statement implies the first. This
shows that A is an order embedding. In order to prove surjectivity,
pick any x E L and put O(y) = 0 if y ~ x and O(y) = 1 otherwise.
Then(} E S, and one finds an idempotent rp with R 0 (8) = rpoS. Thus,
A(rpo S) = rp (1) = e+(o) =X. 0
The major part of the second chapter is devoted to a series of more
difficult theorems characterizing modular complemented lattices, or-
thomodular lattices and Boolean lattices by means of various specific
classes of coordinatizing Bacr semig;roups. This provides a nice frame-
work for John von Neumann's coordinatization of continuous geome-
tries [145]. These are join- and meet-continuous, modular, comple-
mented and irreducible complete lattices and may be represented as
certain ideal lattices of regular rings. Von Neumann's proof of that
deep result is one of the longest and most difficult proofs in lattice
theory and its relationship to classical algebra.
The third chapter of the book by Blyth and Janowitz deals with diverse
kinds of residuated algebraic structures having a binary operation, i.e.
residuated groupoids. Among them are Querre semigroups, Dubreil-
Jacotin semigroups and various types of semigroups admitting gen-
eralizations of the Glivenko-Frink theorem about the Booleanization
of pseudocomplemented semilattices. Useful tools in that theory are
diverse types of congruences like Molinaro congr-uences, zigzag equiva-
lences etc. which ensure that the image of a given residuated groupoid
or semigroup under a (residuated) homomorphism becomes a partially
ordered (quasi) group, a Boolean latt ice or a similar structure having
a specific kind of residuation.
128 M. Erne
4.6 Heyting Algebras, Locales and Pointfree

Topology
Long overdue now is a remark concerning residuated meet operations
and their relationship to non-classical logics, modelled by Brouwerian
(semi)lattices or Heyting algebras. The reader familiar with intuition-
istic logic will be aware of the strong similarity between the system of
axioms for residuations and systems of axioms for implicationallogic.
For example, the rule
a : b = i iff a :::: b
resembles the "deduction theorem"
a --7 b = T iff a f- b
providing the basic connection between formal implication, deduction
of formulas, and metalinguistic conclusion. However, in order to make
the translation more consistent, it is better to translate a : b into
b --7 a and b ~ a into b f- a if the multiplication is replaced with
1\ and interpreted as conjunction ( "and"). Then the other rules for
residuations obtain the following form:
c --7 (b --7 a) = b --7 (c --7 a) ,
c --7 a!\ b = (c --7 a) 1\ (c--+ b) ,
aV b--+ c =(a--+ c) 1\ (b--+ c),
all well-known from intuitionistic logic and other formalized languages
of logic. The identity
a: (b·c) = (a:b) :c
turns into the rules of importation and exportation:
cl\ b --7 a= c --7 (b--+ a),
while the equivalence
cl\ b f- a iff c f- b --7 a
may be viewed as a "semiformal" variant of these rules. Of course,
these coincidences are not casual: indeed, the algebraic counterpart
of intui tionistic logic is the notion of H eyting algebra [71]. This is a
bounded lattice with residuated meet operation, or, what amounts to
the same thing, with a residuated multiplication making all elements
idempotent (then the inequality a/\ b = (a/\ b) 2 ~ ab ~ a/\ b shows

that ab must coincide with the meet a/\ b).
By the basic fact that in complete lattices, the left adjoint maps are
precisely the join-preserving ones, it is clear that the complete Heyting
algebras are just the frames or locales, characterized by the distributive
law
a 1\ VB = V {a 1\ b: bE B}.
There exists an immense literature on that topic, of which we only
mention Johnstone's excellent book Stone Spaces [78]. Perhaps some-
what unexpectedly, in the last few decades logic and topology amalga-
mated to a quite powerful synthesis, profiting from both theories (and,
of course, from lattice theory and the theory of adjoint maps). The
theories that arose from that fruitful interplay are pointfree topology
and the more recent formal topology.
Let us conclude our journey through the world of adjunctions and Ga-
lois connections by the remark that meanwhile there exist extremely
powerful categorical extensions of the modest order-theoretical notion
of adjunctions, like categorical Galois connections and adjoint functors
(see, for example, the early paper by Lawvere [87] on adjunctions, the
nice article by Herrlich and Husek [70] on four types of categorical Ga-
lois connections of increasing generality, and the inexhaustible source
Abstract and Concrete Categories by Adamek, Herrlich and Strecker
[3]). Another very modern theory in that context is the theory of
Chu spaces [115], which combines categorical methods with the idea
of multivalued contexts, known from formal concept analysis. For
quite recent material on the role of (classical and categorical) Galois
connections in pointfree topology and its links to classical topology,
we refer to the paper General Stone Duality (52] and to the article The
polarity between approximation and distribution in this volume.
The slogan is: Adjoint functors arise everywhere.
S. Mac Lane in
Categories for the Working Mathematician
130 M. Erne
References
[1] Abel, N.-H.: (Euvres Completes. (L. Sylow and S. Lie, eds.) , Gr0endahl
& S0n, Christiania, 1881.
[2] Abian, A. : On definitions of cuts and completion of partially ordered
sets. Z. Math. Logik Grundl. der Math. 14 (1968), 299-309.
[3] Adamek, J., Herrlich, H. , Strecker, G.: Abstract and Concrete Cate-
gories. John Wiley& Sons, Inc., New York, 1990.
[4] Alexandroff, P.: Diskrete Raume. Mat. Sb. (N.S.) 2 (1937) , 501-518.
[5] Artin, E.: Galoissche Theorie . Teubner, Leipzig. 1959.
[6] Baer, R.: Linear Algebra and Projective Geometry. Academic Press, New
York, 1952.
[7] Baer, R.: On closure operators. Arch. Math. 10 (1959) , 261- 266.
[8] Banaschewski, B.: Prime elements from prime ideals. Order 2 (1985) ,
211- 213.
[9] Banaschewski, B., Erne, M.: On Krull's separation lemma. Order 10
(1993), 253- 260.
[10] Banaschewski, B., Bruns, G.: Categorical characterization of the Mac-
Neille completion. Arch. Math. (Basel) 43 (1967) , 369- 377.
[11] Benado, M.: Nouveaux theoremes de decomposition et d'intercalation a
la normalite a. C. R. Acad. Sci., Paris 228 (1949) , 529-531.
[12] Birkhoff, G.: Combinatorial relations in projective geometries. Ann.
Math. 36 (1935), 743-748.
[13] Birkhoff, G.: Lattice Theory, Amer. Math. Soc. Coll. Publ. 25, Provi-
dence, R. I.; 1st ed. 1940, 2nd ed . 1948, 3d ed. 1973.
[14] Birkhoff, G.: Ordered sets in geometry. In: I. Rival (ed.) , Ordered Sets. ,
Proceedings NATO ASI Ser. C, Math. and Phys., No. 83 (1982), D.
Reidel Publ. Co. , Dordrecht- Boston- London, 1981, 407- 443.
[15] Bishop, A.: A universal mapping characterization of the completion by
cuts. Algebra Universalis 8 (1978) , 349- 353.
[16] Blyth, T., Janowitz, M.: Residuation Theory. Pergamon Press, Oxford,
1972.
[17] Boole, G.: A general method in analysis. Phil. Trans. of the Royal Soc.
London 134 (1844), 225- 282.
[18] Boole, G.: The Mathematical Analysis of Logic. Cambridge, 1847. Repr.
Oxford, 1951.
[19] Boole, G.: An Investigation of The Laws of Thought, Dover Publ., Lon-
don, 1854. Repr. New York, 1958.
[20] Bvelohlavek, R.: Fuzzy Galois connections. Math. Logic Q. 45 (1999) ,
497- 504.
[21] Bvelohlavek, R.: Fuzzy Galois connections and fuzzy concept lattices:
from binary relations to conceptual structures. Discovering the world
with fuzzy logic, 462-494, Stud. Fuzziness Soft Comput. 57, Physica,
Heidelberg, 2000.
[22] Choquet, G.: Convergences. Ann. Univ. Grenoble Sect. Sci. Math. Phys.
(N.S.) 23 (1948) , 57-112.
[23] Codd, E.F.: A relational model for large shared data banks. Comm.
ACM 13.6 (1970), 377- 387.
[24] Crawley, P.: Decomposition theory for nonsemimodular lattices. Trans .
Amer. Math. Soc. 99 (1961) , 246- 254.
[25] Crawley, P., Dilworth, R.P.: Algebraic Theory of Lattices. Prentice-Hall
Inc., Englewood Cliffs, New Jersey, 1973.
[26] Dedekind, R.: Vorlesungen iiber Algebra. Manuscript (1858-1868).
Printed in [121]
[27] Dedekind, R.: Uber die Theorie der ganzen algebraischen Zahlen. Sup-
plement XI to: Dirichlet, P.G.L.: Vorlesungen iiber Zahlentheorie. pt
ed. Braunschweig, 1863; 2nd ed. 1871, 3d ed. 1879, 4th ed. 1894.
[28] Dedekind, R.: Stetigkeit und Irrationale Zahlen. Vieweg, Braunschweig,
1st ed. 1872, lOth ed. 1969.
[29] Dedekind, R.: Was sind und was sollen die Zahlen? Vieweg, Braun-
schweig, 1st ed. 1887, lOth ed. 1969.
[30] Dedekind, R.: Uber Zerlegung von Zahlen durch ihre gr613ten gemein-
samen Teiler. Festschrift der Technischen Hochschule zu Braunschweig
bei Gelegenheit der 69. Versammlung Deutscher Naturforscher und
Arzte, 1-40 (1897). Reprint with concluding comment by E. Noether in:
R. Fricke (ed.), Richard Dedekind, Gesammelte Mathematische Werke,
2nd vol., Vieweg, Braunschweig, 1931 , pp. 103- 147.
[31] Dedekind, R.: Uber die von drei Moduln erzeugte Dualgruppe. Math.
Annalen 53 (1900) , 371-403. Reprint with concluding comment by E.
Noether in: R. Fricke (ed.) , Richard Dedekind, Gesammelte Mathema-
tische Werke, 2nd vol., Vieweg, Braunschweig, 1931, pp. 236- 271.
[32] Dedekind, R.: Uber die Permutationen des Korpers aller algebraischen
Zahlen. Festschrift zur Feier des hundertfiinfzigjiihrigen Bestehens der
Koniglichen Gesellschaft der Wissenschaften zu Gottingen. Abh. der
Math.-phys. Klasse (1901) , 1-17.
[33] Deiters, K., and Erne, M: Negations and contrapositions of complete
lattices. Discrete Math. 181 (1998) , 91-111.
[34] Diercks, V., Erne, M., Reinhold , J.: Complements in lattices of varieties
and equational theories. Algebra Universalis 31 (1994) , 506- 515.
[35] Dilworth, R.P.: Abstract residuation over lattices. Bull. Amer. Math .
Soc. 44 (1938), 262- 268.
132 M. Erne
[36] Dilworth, R.P.: Non-commutative residuated lattices. Trans . Amer.

Math. Soc. 46 (1939), 426- 444.
[37] Dilworth, R.P.: Abstract commutative ideal theory. Pacific J. Math 12
(1962), 481- 498.
[38] Dilworth, R.P., Crawley, P.: Decomposition theory for lattices without
chain conditions. Trans. Amer. Math. Soc. 96 (1960), 1- 22.
[39] Edwards, H.M.: Galois Theory. Springer-Verlag, New York - Berlin -
Heidelberg, 1984.
[40] Erne, M.: Einfiihrung in die Ordnungstheorie. B.l. Wissenschaftsverlag,
Mannheim - Heidelberg, 1982.
[41] Erne, M.: Lattice representations for categories of closure spaces, in:
Bentley, H. L. et al. (eds.), Categorical Topology, Proc. Conf. Toledo,
Ohio 1983, Heldermann Verlag, Berlin, 1984, pp. 197-222.
[42] Erne, M.: Order extensions as adjoint functors . Quaest. Math. 9 (1986) ,
149-206.
[43] Erne, M.: Tensor products for bounded posets revisited. Order 7 (1990) ,
295- 314.
[44] Erne, M.: The ABC of order and topology, in: Herrlich, H. and Porst,
H.-E. (eds.), Category Theory at Work, Heldermann Verlag, Berlin, 1991,
57-83.
[45] Erne, M.: The Dedekind-MacNeille completion as a reflector. Order 8
(1991), 159-173.
[46] Erne, M.: Algebraic ordered sets and their generalizations, in: Rosen-
berg, I. and Sabidussi, G. (eds.), Algebras and Orders, Proc. Montreal
1992, Kluwer, Amsterdam, 1994.
[47] Erne, M., Koslowski, J., Melton, A., Strecker, G.: A primer on Galois
connections. In: S. Andima et. al. (eds.): Papers on General Topology
and Applications. Annals of the New York Academy of Sciences 704,
1993, pp. 103- 125.
[48] Erne, M.: Distributive laws for concept lattices. Algebra Universalis 30
(1993)' 538-580.
[49] Erne, M.: Tensor products of contexts and complete lattices. Algebra
Universalis 31 (1994) , 36-65.
[50] Erne, M.: Z-continuous posets and their topological manifestation. Appl.
Cat. Structures 7 (1999) , 31- 70.
[51] Erne, M.: Prime ideal theory for general algebras. Appl. Cat. Structures
8 (2000), 115- 144.
[52] Erne, M.: General Stone duality. In: M. M. Clementino et al. (eds.),
Proc. IV Iberoamerican Conf. on Topology, Coimbra 2001. Topology and
its Appl., (to appear)
[53] Erne, M., Reinhold, J .: Ordered one-point-compactifications, stably con-

tinuous frames and tensors. Quaestiones Mathematicae 22 (1999) , 63- 81.
[54] Everett, C.J.: Closure operators and Galois theory in lattices. Trans.
Amer. Math. Soc. 55 (1944) , 514- 525.
[55] Frink, 0.: Pseudo-complements in lattices. Duke Math. J. 29 (1962) ,
505- 514.
[56] Galois, E.: Memoire sur les conditions de resolubilites des equations
par radicaux (1831). Published by J. Liouville in the Journal de
Mathematiques 1846.
[57] Galois, E.: CEuvres Mathematiques. With an introduction by E. Picard.
Gauthier-Villars, Paris 1897.
[58] Ganter, B., Wille, R., Wolff, K. E. (eds.) : Beitriige zur Begriffsanalyse.
B.I. Wissenschaftsverlag, Mannheim- Wien - Zurich, 1987.
[59] Ganter, B., Wille, R. : Formal Concept Analysis - Mathematical Founda-
tion. Springer-Verlag, Berlin - Heidelberg - New York, 1999
[60] Gauss, C. F.: Demonstratio nova altera theorematis omnem functionem
algebraicam rationalem integram unius variabilis in factores primi vel
secundi gradus resolvi posse. Comm. soc. regiae scient. Gottingensis re-
centiores 3 (1816). Reprinted in: C. F. Gauss , Werke Bd. III, Georg
Olms, Hildesheim, 1981 , pp. 31-56.
[61] Gauss, C. F.: Disquisitiones Arithmeticae. Apud Gerh. Fleischer Iun.
Lipsia, 1801. Reprinted in: C. F . Gauss, Werke Bd. I, Herausg. Konig.
Ges. Wiss. Gottingen, 1870.
[62] Gierz, G., Hofmann, K.H. , Keimel , K ., Lawson, J.D., Mislove, M., and
Scott, D.S.: A Compendium of Continuous Lattices, Springer-Verlag,
Berlin - Heidelberg - New York, 1980.
[63] Gillman, L., Jerison, M.: Rings of Continuous Functions. Van Nostrand ,
Princeton. Reprint: Graduate Texts in Math. 43, Springer-Verlag, Berlin
- Heidelberg - New York, 1976.
[64] Girard, A.: Invention Nouvelle en l'Algebre (1629) , reimpression par D.
Bierens De Haan, Mure Freres, Leiden, 1884.
[65] Glivenko, V.: Sur quelques points de Ia logique de M. Brouwer. Bull.
Acad. Sci. Belgique 15 (1929), 183- 188.
[66] Grassmann, H.: Gesammelte Mathematische und Physikalische Werke .
ed. F. Engel. 3 Vol. , Leipzig, 1894- 1911.
[67] Grassmann, R.: Die Formenlehre oder Mathematik. Stettin , 1872.
Reprint Hildesheim, 1966.
[68] Gratzer, G.: General Lattice Theory. Bikhiiuser, Basel, 1978.
[69] Hailperin, T.: Boote 's Logic and Probability. Studies in Logic 85, North-
Holland, Amsterdam, 1976.
134 M. Erne
[70] Herrlich, H., Husek, M.: Galois connections categorically. J. Pure Appl.
Algebra 68 (1990), 165- 180.
[71] Heyting, A.: Die formalen Regeln der intuitionistischen Logik. Sitzungs-
ber. PreujJ. Akademie der Wiss., Phys.-Math. Klasse (1930) , 42-56.
[72] Hilbert, D.: Uber die Theorie der algebraischen Formen. Math. Annalen
36 (1890), 473- 534.
[73] Hilbert, D.: Uber die vollen Invariantensysteme. Math. Annalen 42
(1893), 313- 373.
[74] Hilbert, D.: Grundlagen der Geometrie. 7th ed., Teubner, Stuttgart,
1930.
[75] Hoffmann, R.-E.: Topological spaces admitting a "dual", in: Cate-
gorical Topology, Lecture Notes in Math. 719, Springer-Verlag, Berlin-
Heidelberg-New York, 1979, pp. 157- 166.
[76] Huntington, E. V.: Sets of independent postulates for the algebra of logic.
Trans. Amer. Math. Soc. 5 (1904), 288-309.
[77] Johnson, Jr., A.: A lattice whose residuated maps do not form a lattice.
J. Nat. Sci. and Math. 9 (2) (1969), 283-284.
[78] Johnstone, P. T.: Stone Spaces, Cambridge University Press, Cambridge,
1982.
[79] Korselt , A.R. : Bemerkungen zur Algebra der Logik. Math. Ann. 44
(1894), 156-157.
[80] Korselt, A.R.: Uber die Grundlagen der Geometrie. Jahresber. DMV 12
(1903), 402- 407.
[81] Korselt, A.R.: Uber die Logik der Geometrie. Jahresber. DMV17 (1908) ,
98- 124.
[82] Krull, W.: Axiomatische Begriindung der allgemeinen Idealtheorie.
Sitzungsberichte der physikalisch-medizinischen Sozietiit zu Erlangen 56
(1924), 47-63.
[83] Krull , W.: Zur Theorie der zweiseitigen !deale in nichtkommutativen
Bereichen. Math. Z. 28 (1928), 481- 503.
[84] Kunz, E.: Introduction to Commutative Algebra and Algebraic Geometry.
Birkhauser, Basel, 1985.
[85] Lagrange, J.-L.: CEvres Complf~ tes. Ed. J.A. Serret; G. Darboux. 14
volumes, Gauthier-Villars, Paris, 1867- 1892.
[86] Lagrange, J.-L.: Reflexions sur la resolution algebrique des equations.
Nouveaux Memoires de l'Academie Royale des Sciences et Belles-Lettres
de Berlin, 1770-1771. See also [85], vol. 3, pp. 205-421.
[87] Lawvere, F.W.: Adjointness in foundations. Dialectica 23 (1969), 281-
296.
[88] Levine, N.: Strongly connected sets in topology, Amer. Math. Monthly
72 (1965)' 1098- 1101.
[89] Lisa, J.: Cardinal sums and direct products in Galois connections. Com-
ment. Math. Univ. Carolinae 14 (1973) , 325- 338.
[90] Lorenz, F.: Einfuhrung in die Algebra. B.l. Wissenschaftsverlag, Mann-
heim, 1992.
[91] MacNeille, H.M.: Partially ordered sets. Transactions Amer. Math. Soc.
42 (1937), 90-96.
[92] Mehrtens, H.: Die Entstehung der Verbandstheorie. Gerstenberg, Hildes-
heim, 1979.
[93] Melton, A., Schmidt, D. A., Strecker, G. E.: Galois connections and com-
puter science applications. In: D. Pitt et al. (eds.) : Category Theory
and Computer Programming (Guilford 1985), Lecture Notes in Computer
Science 240 (1986) , 299- 312.
[94] Menger, K.: Bemerkungen zu Grundlagenfragen IV. Jahresber. DMV 37
(1928), 309- 325.
[95] Menger, K.: New foundations of projective and affine geometry. Ann.
Math . 37 (1936) , 456- 482.
[96] Menger, K.: Non-euclidean geometry of joining and intersecting. Bull
Amer. Math. Soc. 44 (1938) , 821-824.
[97] Menger, K.: Topology without points. The Rice Institute Pamphlet 27
(1940), 80- 107.
[98] Meyberg, K: Algebra, Teil 2. C. Hanser Verlag, Miinchen - Wien, 1976.
[99] Nelson, E.: Galois connections as left adjoint maps. Comm. Math. Univ.
Carolinae 17 (1976) , 523- 541.
[100] Nobeling, G.: Topologie der Vereine und Verbande. Arch. Math. (Basel)
1 (1948) ,154- 159.
[101] Nobeling, G.: Grundlagen der Analytischen Topologie. Springer-Verlag,
Heidelberg, 1954.
[102] Noether, E.: Abstrakter Aufbau der Idealtheorie. Math. Annalen 96
(1926)' 36-61.
[103] Ore, 0.: On the foundations of abstract algebra, II, Annals of Math. 37
(1936), 265- 292.
[104] Ore, 0.: Structures and group theory II. Duke Math. J. 4 (1938) , 247-
269.
[105] Ore, 0.: Theory of equivalence relations. Duke Math. J. 9 (1942) , 573-
627.
[106] Ore, 0.: Theory of monomial groups. Trans . A mer. Mat. Soc. 51 (1942) ,
15-64.
[107] Ore, 0.: Combinations of closure relations. Ann. of Math. (2) 44 (1943),
514- 533.
[108] Ore, 0.: Some studies on closure relations. Duke Math. J. 10 (1943) ,
573- 627.
136 M. Erne
[109) Ore, 0.: Galois connexions. Trans. Amer. Math. Soc. 55 (1944), 493-
513.
[110] Peirce, C. S.: Collected papers, Vol. III. Exact Logic, Vol. IV. The Sim-
plest Mathematics. 2nd printing, Cambridge, Mass., 1960.
[111) Pawlak, Z.: Rough concept analysis. Bull. Polish Acad. Sci.: Technical
Sciences 33 (1985), 495- 498.
[112) Pawlak, Z.: Rough sets. Kluwer, Dordrecht- Boston 1991.
[113) Picado, J.: The quantale of Galois connections. Preprint, University of
Coimbra, 2001.
[114) Pickert, G.: Bemerkungen iiber Galois-Verbindungen.
Ar·ch. Math. (Basel) 3 (1952) , 285- 289.
[115) Pratt, V.: Chu spaces. Notes for the School on Category Theory and
Applications, University of Coimbra, 1999.
[116) Purkert, W.: Ein Manuskript Dedekinds iiber Galois-Theorie. NTM-
Schriftenr. Geschichte Naturw. Technik Med. 13, Heft 2 (1977), 1- 16.
[117] Rabinovich, E.: Ein Beweis des Hilbertschen Nullstellensatzes. Math.

Annalen 102 (1929) , 33.
[118] Raney, G.: Tight Galois connections and complete distributivity. Trans .
Amer. Math. Soc. 97 (1960), 418-426.
[119] Riguet, J.: Relations binaires, fermetures , correspondances de Galois.
Bull. Soc. math. France 76 (1948), 114- 155.
[120) Rosenthal, K. I.: Quantales and Their Applications. Longman Scientific
& Technical, Harlow, Essex, 1990.
[121) Scharlau, W. (ed.): Richard Dedekind 1831-1981, Festschrift 1981.
Vieweg, Braunschweig, 1982.
[122) Schmidt, J.: Die Grundla9en der Theorie der halbgeordneten Mengen.
Ph. D. Diss., Berlin, 1952.
[123) Schmidt, J.: Beitrage zur Filtertheorie, II. Math. Nachr. 10 (1953), 197-
232.
[124] Schmidt, J.: Zur Kennzeichnung der Dedekind-MacNeilleschen Hiille
einer geordneten Menge. Arch. Math. (Basel) 7 (1956) , 241-249.
[125] Schmidt, J.: Every join-completion is the solution of a universal problem.
J. Austral. Math. Soc. 17 (1974), 406- 413.
[126] Schmidt, R.: Subgroup Lattices of Groups. Expositions in Math. 14. De
Gruyter, Berlin - New York, 1994.
[127] Schroder, E.: Lehrbuch der Arithmetik und Algebra fur Lehrer und
Studierende. Vol. I: Die sieben algebraischen Operationen. Leipzig, 1873.
[128] Schroder, E.: Uber die formalen Elemente der absoluten Algebra. Stutt-
gart, 1874.
[129] Schroder, E.: Der Operationskr·eis des Logikkalkuls. Leipzig, 1877; repr.
Darmstadt, 1966.
[130] Schroder, E.: Uber Algorithmen und Calculn. Archiv der Math. und
Physik (2nd series) 5, (1887), 225-278.
[131] Schroder, E.: Vorlesungen uber die Algebra der Logik. (Exakte Logik). 3
Vol., Leipzig, 1890-1905; repr. New York, 1966.
[132] Schroder, E.: Algebra und Logik der Relative. Erste Abteilung. Leipzig,
1895.
[133] Shmuely, Z.: The structure of Galois connections. Pacific J. Math. 54
(1974), 209- 225.
[134] Shmuely, Z.: The tensor product of distributive lattices. Algebra Univer-
salis 9 (1979), 281- 296.
[135] Sonner, H.: Die Polaritiit zwischen topologischen Riiumen und
Limesriiumen. Arch. Math. (Basel} 4 (1953) , 461- 469.
[136] Steinitz, E.: Algebraische Theorie der Korper. Journal fur Mathematik
137 (1910), 167- 309.
[137] Stone, M.H.: The theory of representation for Boolean algebras. Trans .
Amer. Math. Soc. 40 (1936), 37- 111.
[138] Tignol, J.-P.: Galois' Theory of Algebraic Equations. World Scientific,
Singapore 2001.
[139] Toti Rigatelli, L.: Evariste Galois (translated from the Italian by J.
Denton). Birkhiiuser, Basel, 1996.
[140] Umbreit, S.: Formale Begriffsanalyse mit unscharfen Begriffen. Ph.
Diss., University of Halle, 1994.
[141] Vandermonde, A. T .: Memoire sur Ia resolution des equations. Histoire
de l'Acad. Royale des Sciences (avec les memoires de Math. f3 de Phys.
pour la meme annee, tires des registres de cette Acad.} (1771), 365-416.
[142] Van der Waerden, B.: Algebra, Erster und Zweiter Teil. pt_7th ed.,
Springer-Verlag, Berlin - Heidelberg - New York, 1936-1966.
[143] Vieta, F.: The Analytic Art. Translated by T. R. Witmer, Kent State
Univ. Press, Kent, Ohio, 1983.
[144] Voigt, A.: Was is Logik? Vierteljahresschrift fur wissenschaftliche
Philosphie 16 (1892) , 189-332.
[145] Von Neumann, J.: Continuous Geometry. Reprint of Lectures from
1935/6. Princeton, 1960.
[146] Ward, M.: Residuation in structures over which a multiplication is de-
fined. Duke Math. J. 3 (1937) , 627- 636.
[147] Ward, M.: Some arithmetical applications of residuation. A mer. J. Math.
39 (1937), 921- 926.
[148] Ward, M.: Structure residuation. Ann. Math. 39 (1938) , 558- 568.
138 M . Erne
[149] Ward, M.: The closure operators of a lattice. Ann. of Math. (2) 43
(1942)' 191- 196.
[150] Ward, M., Dilworth, R.P.: Residuated lattices. Trans. Amer. Math. Soc.
45 (1939), 335- 354.
[151] Ward, M., Dilworth, R.P.: Evaluations over residuated structures. Ann.
Math. 40 (1939), 328- 338.
[152] Waring, E.: Meditationes Algebraicae. Translated by D. Weeks, Amer.
Math. Soc., Providence, RI, 1991.
[153] Wille, R.: Restructuring lattice theory: An approach based on hierar-
chies of concepts. In: I. Rival (ed.) , Ordered Sets (Banff 1981}, NATO
ASI Ser. C, Math. and Phys. , No. 83 (1982) . D. Reidel Pub!. Co. , Dor-
drecht- Boston- London, 1981, 445- 470.
Author's address:
Marcel Erne
Department of Mathematics
University of Hannover
D-30167 Hannover
Germany
e-mail: erne@math.uni-hannover.de
Categorical Galois Theory: Revision
and Some Recent Developments
G. Janelidze
Abstract
This expository paper presents a short review of categorical Galois theory,

with special attention to the connection with A. R. Magid's Galois the-
ory of commutative rings and most recent developments in the theory of
generalized central extensions. In the last section some open questions are
proposed.
AMS Mathematics Subject Classification: 18A40, 06A15.
Key words: Categorical Galois theory, Galois theory of commutative rings.
1 Introduction
The so-called fundamental theorem of classical Galois theory (GFT

for short) describes the lattice of sub extensions of a finite Galois field
extension E I K , and - according to its long formulation presented in
many modern textbooks in algebra - also says a few words about
the morphisms and especially automorphisms of those subextensions.
However all this information is contained in one sentence, which is
a corollary of a very special case of Grothendieck's formulation of
GFT: it says that the opposite category of subextensions of E I K is
equivalent to the category of transitive Gal (E I K) - sets. Categorical
language also brings many other improvements and simplifications,
and in particular allows algebraic geometers to explain that the Galois
theory of fields and the theory of covering spaces are not just analogous
but identical; in particular the fundamental group (of a scheme) is
139
K. Denecke et at. (eds.). Galois Connections and Applications, 139-171.
140 G. Janelidze
nothing but the Galois group of the universal covering, i.e. the largest
Galois group.
Grothendieck's GFT has been extended to an elementary-topos-theo-
retic context in various ways by M. Barr, Diaconescu, J. Kennison , G.
Wright, A. Joyal and M. Tierney, I. Noerdijk, M. Bunge and others,
and there are many other non-topos-theoretical extensions, such as the
one proposed by Y. Diers (19]; A. R. Magid's theorem [55, Theorem
IV.31] seems to be especially important - since it is the final step in
the extension of (separable) Galois theory from fields to commutative
rings. Magid's theory is indeed beyond the topos-theoreticallevel since
it involves profinite groupoid actions on profinite topological spaces,
and those do not form a topos.
A purely categorical approach to Magid 's theory was first briefly de-
scribed in [27] and then in [28], where it was already transformed into
categorical Galois theory, with a new example of central extensions of
groups - which is very far from Grothendieck's Galois theory. These
results were reported to Saunders Mac Lane, and presented in his talk
[54] - and the next paper (30] was written according to his suggestion,
with an additional page written by him for the introduction. Various
further developments are briefly described in the book (4] , whose main
purpose is to show a precise chain of generalizations from the classical
level of finite field extensions to the purely categorical level.
This expository paper presents a short review of categorical Galois
theory (with special attention to the connection with A. R. Magid 's
Galois theory of commutative rings and most recent developments in
the theory of generalized central extensions) , where however I was try-
ing to exclude the material that would copy the corresponding parts
of [4], for example the categorical approach to covering spaces de-
scribed in Chapter 6 of [4] ; some " minimal" definitions from [4] are
replaced by more refined ones, especially useful for formulations of
open questions and for a few observations that did not exist when the
manuscript of [4] was submitted for publication.
The paper is organized as follows: Section 1 shows how the topological
notion of connectedness becomes categorical, and transforms it into
what is called Galois structure, which is an abstract adjunction be-
tween categories with specified classes of morphisms called fibrations.
Several examples of Galois structures are given. The term "fibration"
Recent Developments in Categorical Galois Theory 141
is used for the first time; in fact it is suggested by the example studied
in [8](see Example 2.12).
Since the role of Galois group( oid)s in general is played by internal pre-
categories, the internal precategories and their actions are described
in Section 3; a sufficient condition on a precategory action to become
a category/ groupoid action is also given.
The categorical GFT in Section 4 is formulated essentially as in [32],
and then some related notions are defined and some examples are
described in Section 5.
The links and applications to Galois theory of commutative rings are
briefly described in Section 6, and the theory of generalized central
extensions is considered in Section 7.
Several open questions are listed in Section 8.
I would like to thank Klaus Denecke and the other organizers of the In-
ternational Conference on Galois connections in Potsdam for inviting
me; this paper is an extended version of the talk I gave there.
2 Connectedness and Galois structures
Usually connectedness is considered as a topological notion, but as

the following simple and well known theorem shows, there are several
ways to define it in a purely categorical language:
Theorem 2.1. The following conditions on a topological space A ar·e

equivalent:
(a) A is connected;
(b) A is not the initial object in the categor-y Top of topological spaces,
and if A~ B+C in Top, then either- B or- C is canonically isomor-phic
to A;
(c) A is not the initial object in Top, and if A~ B + C, then either-
B or- C is the initial object in Top;
(d) any mor-phism fr-om A to a copr-oduct in Top factor-s thr-ough one
of the copr-oduct injections;
(e) the functor hom(A ,-) : Top ----7 Sets preserves coproducts.
Moreover, replacing Top with a general category C with coproducts,
we still have the implications (e)-=*(d}, (d}-=*(b), (c)-=*(b), and if the
coproducts in C ar-e universal and disjoint (in A. Gr-othendieck's ter-
minology - that is, C is infinitar-y extensive in modern terminology),
142 G. Janelidze
then the converse implications also hold. In particular the conditions

(b) - (e) are equivalent whenever C is locally connected in the sense
of 2.2(c) below.
Definition 2.2. Let C be a category with coproducts and finite limits.

Then
(a) an object A inC is said to be connected if the functor hom( A , -) :
C ---+ Sets preserves coproducts;
(b) the category C is said to be connected if its terminal object is
connected;
(c) the category Cis said to be locally connected if every object inC
is a coproduct of connected objects.
For the readers familiar with topos theory, we point out that C is not
required to be a topos; however the first example in the following list
in fact recalls the topos-theoretic motivation for 2.2(c).
Example 2.3. (a) Let A be a topological space. The categories

(Top ._!.. A) of bundles over A and Shv(A) of sheaves (of sets) over
A are connected if and only if so is A ; the category Shv(A) is locally
connected if and only if A is a locally connected topological space.
(b) If a category C is locally connected, then so is (C ._!.. A) , for any
object A in C; the category (C ._!..A) is then connected if and only if
so is A.
(c) The category Cat of all categories is connected and locally con-
nected; an object A in Cat is connected if and only if for every two ob-
jects a and bin A, there exists a zigzag of morphisms a ---++--- ... ---++--- b
between a and b. The category (Cat._!.. A) is always locally connected,
and is connected if and only if so is A (by (b)). This easily implies
the same properties for the functor category Sets A. In particular the
following structures form categories that are connected and locally
connected: M-sets (where M is a monoid) , presheaves (of sets) over a
fixed topological space, graphs, simplicial or cubical sets, etc.
A locally connected category C can always be identified, up to a cate-
gory equivalence, with the category Fam(Conn(C)) of families of its
connected objects. Under this identification
A;:::::; LA in C corresponds to A= (A)iE inFam(Conn)(C); (2.1)

1
iEl
we will write I(A) instead of I in order to consider it as a functor
I :C ----t Sets, (2.2)

which in fact comes from the 2-functor Fam : Cat ----t Cat as
I= Fam(!conn(c)) : Fam(Conn(C)) ----t Fam(l) =Sets, (2.3)

where !conn(C) : Conn(C) ----t 1 is the unique functor from Conn(C)
to the terminal object in Cat.
If in addition C is connected, then the functor (2) has a right adjoint
H : Sets ----t C, (2.4)

defined by H(X) =X· 1, where X· 1 is the coproduct of " X copies"
of a fixed terminal object in C. And again, the adjunction bfCr:;Sets
is nothing but the image of the adjunctionConn(C)r:;l under the 2-
functor Fam. Moreover, the functor H itself has a right adjoint,
namely hom(1,-).
Let us now assume that C is equipped with a fixed adjoint pair of func-
tors (2.2) and (2.4) , which were determined so far only up to an isomor-
phism of course. This suggests to say that each object A in C has its
!(A)-indexed family (A)iEJ(A) of connected components, in which each
Ai is defined as the pullback
H({i})
t
HI(A)
(2.5)
where 'TJA is the unit of the adjunction, and the right hand vertical
arrow is the image of the inclusion map {i} ----t I(A) under the functor
H.
Our next step is to make the notion of connectedness much more ab-
stract by replacing Sets with an arbitrary category X. The structure
which we obtain provides an appropriate level of generality for categor-
ical Galois theory and therefore it is called Galois structure, alt hough
the definition we are giving here is slightly different from the original
144 G. Janelidze
one given in [32]. Its absolute version (see Definition 2.4( e)) is closely
related to various other structures that occur in general connected-
ness/ radical/ torsion theories (see [57] and references there) .
Definition 2.4. A Galois structure r on a category C with finite

limits consists of an adjunction
(I , H, rJ , E) : C---+ X (2.6)
and two classes F and of morphisms in C and X respectively, whose
elements are called fibrations ; the following conditions on fibrations
are required:
(a) all pullbacks along fibrations exist, and the classes of fibrations are
pullback stable;
(b) the classes of fibrations are closed under composition and contain
all isomorphisms;
(c) the functors I and H preserve fibrations .
Such a Galois structure is said to be finitely complete, absolute, ad-
missible, or closed, if it satisfies the following conditions respectively:
(d) the categories C and X have all finite limits (usually there are all
small limits, but we will never use them) ;
(e) all morphisms in C and in X are fi brations;
(f) for every object C in C and every fibration <p : X ---+ I (C) in
X, the composite of the canonical morphisms I(C x HI(C ) H(X)) ---+
IH(X)---+ X is an isomorphism;
(g) for every object A inC, the morphism 'TJA : A---+ H I(A) is a fibra-
tion, and for every object X in X , the morphism Ex : I H(X) ---+X is
an isomorphism.
Let us consider/ recall some examples:
Example 2.5. The following data determine a Galois structure sat-

isfying all additional conditions (i.e. it is finitely complete, absolute,
admissible, closed) of Definition 2.4: C any connected and locally
connected category, X = Sets, I and H the functors (2.2) and (2.4)
respectively, and F and the classes of all morphisms in C and X
respectively.
Example 2.6. The category C of locally connected topological spaces

has an admissible Galois structure defined as in Example 2.5 but with
F being the class of local homeomorphisms.
Example 2. 7. Any adjunction (I, H, TJ, c) : C ---+X becomes a Galois

structure if we take F and to be the classes of isomorphisms in C
and X respectively. It is always admissible, and closed if and only if
it is a category equivalence.
Example 2.8. Any pair (C, F) consisting of a category C and a
class F of morphisms inC satisfying the conditions 2.4(a) and 2.4(b) ,
determines an identity Galois structure; it consists of the identity ad-
junction (1, 1, 1, 1) : C ---+ C with F = , and it is always admissible
and closed.
Example 2.9. There is an obvious finite version of example 2.5. It
involves finitary local connectedness instead of local connectedness and
Finite Sets instead of Sets. That is, the category C is required
to have finite coproducts, and every object in C must be a finite
coproduct of connected objects. In particular C could be the opposite
category of finite-dimensional commutative (unitary) algebras over a
field.
Example 2.10. There is also a profinite version of Example 2.5, with

X = Profinite Spaces = Stone Spaces. It is in fact described in
[10]. Let us restrict ourselves here to the following two special cases:
(a) Cis the opposite category of commutative (unitary) rings. In this
case the functors I and H have the following description: I(A) is the
Boolean spectrum (=Pierce spectrum) of the ring A, i.e. it is defined
either as the Stone space (=the space of maximal ideals) of the Boolean
algebra of idempotents in A, or as the space of connected components
of the Zariski spectrum of A; H(X) is the ring of continuous maps
from the space X to the ring of integers equipped with the discrete
topology.
(b) C is the category of compact Hausdorff spaces, I carries spaces
to the spaces of their connected components, and H is the inclusion
functor.
In both cases all the additional conditions of Definition 2.4 hold. In
the case (a) the admissibility was used in [27] (see also [28]) in order
to show that A.R. Magid's Galois theory [55] extends to general cate-
gories. The case (b) was used in [11] to give a categorical description
of the monotone-light factorization of continuous maps of compact
spaces.
146 G. Janelidze
Example 2.11. Let C be a variety of universal algebras, X a subva-

riety in C, (I, H, TJ, c) : C ---+ X the reflection - inclusion adjunction ,
and F and <P the classes of regular epimorphisms (=surjections) in
C and X respectively. This Galois structure is finitely complete and
closed. Moreover, as shown in [34] (in fact in a more general context) ,
it is admissible whenever C is congruence modular.
Example 2.12. Let C be the category of simplicial sets, X the cat-

egory of (small) groupoids, (I , H , TJ, c) : C ---+ X the fundamental
groupoid - nerve adjunction , and F and <P the classes of Kan fibra-
tions in C and X respectively. The admissibility of this Galois struc-
ture is used in [8], and the same can be done for many other Quillen
homotopy structures.
Most of these and some other examples actually occur in various Galois
theories mentioned in [4, Appendix A4].
We will use the notation of Definition 2.4 for a fixed Galois structure
r, and for each object C in C , write
(Ic, He, TJc, cc) : F(C)---+ <P(J(C)) (2.7)
for the induced adjunction between the category F( C) of fibrations
over C and the category <P(J(C)) of fibrations over J(C) .
Note that for an object (X, <p) in <P(J(C)) - the (X, <p) - component
of the counit cc is nothing but the composite involved in Definition
2.4(f). Therefore r is admissible if and only if this counit is an iso-
morphism. On the other hand the admissibility condition 2.4(f) well
agrees with the intuitive idea of connectedness, since applied to the
map {i} ---+ I(A) used in the pullbacks of the form (2.5) , it just says
that the connected components must themselves be connected.
3 Precategory actions
From now on we have to assume that the reader is familiar with some
basic notions of internal category theory, described for instance in [4] .
This should not create problems for "non-experts" since these notions
are as simple and natural as their well-known special cases, such as
• internal groups ( = group objects) used for example in topology
and analysis, where internal groups in the category of Hausdorff
spaces are called topological groups;
• internal equivalence relations used for example in Universal Al-

gebra, where they are called congruences.
Let C be an internal category in a category C with finite limits and

I : C ---+ X an arbitrary functor. If I preserves finite limits, then
we can say that the image J(C) is an internal category in X; if not ,
then still, there are natural ways to say that J(C) is a " generalized
internal category" and to define an internal !(C)-action in X. Trying
to use the simplest possible structure carried by J(C), one arrives at
the following
Definition 3.1. (a) An internal precategory P in a category X is a

diagram
p d
(3.1)
in X with de = 1 = ce,dp = cq,dm = dq, and em = cp.

(b) Let P be an internal precategory as above. An internal P-action in
X is a triple A= (A 0 , 1r, ~),in which 1r: A 0 ---+ P 0 and~: P 1 x(d,1r)Ao---+
A 0 are morphisms in X in such that all pullbacks along 1r exist (and so
P 1 x(d,1r)Ao and all pullbacks below are well-defined), and the diagrams
P1 x (d,1r) Ao Ao
proj1 j jw (3.2)
pl Po
c
148 G. Janelidze
<p.q>xl 1 X~ < e1r , 1 >
(3.3)
commute. The category of internal F-actions in X will be denoted by

XP.
The notation we are using may not seem very natural if one considers
an internal precategory as a truncated simplicial object; however it
follows the internal-categorical notation: indeed, when P is an internal
category,
• Pi(i = 0, 1, 2) are its object of objects, object of morphisms, and
object of composable pairs respectively,
• d and c the domain and codomain,
• p and q the pullback projections,
• e and m the identity and the composition.
Remark 3.2. If X = Sets, it is convenient to use the following "

mixture" of the categorical and simplicial terminology and notation
for a precategory P above:
• The system g -=::+ P0 = (Po, P 1 , d, c) is called the underlying graph
of P; accordingly we will write u : x --+ y and call u a morphism
from the object x to the object y in P whenever u is an element in
g with d(u) =a and c(u) =b. We will also write e(x) = lx, and call
(P0 , P 1 , d, c, e) the underlying reflexive graph of P .
• The elements of P 2 are called triangles in P , and displayed as
w
X z
whenever p(t) = u, q(t) = v and m(t) = w ; if P was a category, then

the triangles would be the same as composable pairs of morphisms, t
would be identified with (v , u) and w = vu.
The F-actions can be described as:
(a) the diagrams D : (P1 DP0 ) --+ Sets with D(lx) = lv(x) for every
object x and D(v)D(u) = D(w) for every triangle (2.4);
(b) the morphisms of precategories from P to Sets;
(c) the functors from L(P) to Sets, where Lis the left adjoint to the
inclusion functor from the category of categories to the category of
precategories; explicitly, L(P) is generated by the underlying reflexive
graph of P and the set of identities {m(t) = q(t)p(t)it E P2} (where
m(t) = q(t)p(t) is of course the same as w = vu in the situation (3.4)).
Since the definition of a precategory is purely equational with no limits
involved, it is indeed clear that any functor I : C --+ X carries the
internal precategories in C into internal precategories in X. Thus,
if C is an internal category in C, then I (C) is at least an internal
precategory in X. Moreover, the following fact is well known:
Proposition 3.3. In the notation above, I( C) is an internal category

whenever the canonical morphisms
I(Cl X(d,c) C1) --+ I(Cl) X(I(d),I(C)) I(CI) {3.5}
I(Cl X(d,c) cl X(d,c) Cl) --+ I(Ct) X(I(d),I(C)) I(CI) X(I(d),I(C)) I(Cl) (3.6}
are isomorphisms. Under this condition, if C is an internal groupoid,
then so is I(C).
The following special case will be important for our purposes:

Let p : E --+ B be a morphism in C , for which there exist the
pullbacks E x 8 E, E x 8 E x 8 E ~ (E x 8 E ) xE (E x 8 E), and
E x 8 E x 8 E x 8 E ~ (E x 8 E) xE (E x 8 E) xE (E x 8 E). Then
there is the internal equivalence relation Eq(p) on E corresponding to
150 G. Janelidze
p (= the kernel pair of p) , and this equivalence relation can be con-

sidered as an internal groupoid inC ; its " precategory part" displays as
E (3.7)
and Proposition 3.3 gives
Corollary 3.4. In the notation above, I(Eq(p)) is an internal groupoid

whenever the canonical mor·phisms
I((E XB E) XE (E XB E))-+ I(E XB E) XJ(E) (E XB E) (3.8)
I((E XB E) XE (E XB E) XE (E XB E)) -+ I(E XB E) XJ(E) (E XB

E) XJ(E) I(E XB E) (3.9)
are isomorphisms.
4 The fundamental theorem of Galois theory
In this section r = ((I , H , T] ,c): C-+ X , F ~ Mor(C) , .P ~ Mor(X))

denotes a fixed Galois structure in the sense of Definition 2.4, and
p: E-+ B a fibration in C , which we will usually write as (E, p) and
call a fibration over B.
Consider the following two diagrams:

JE
F(E) (I(E))
p!
j 1J(p)! (4.1)
F(B) (I(B))
JB
HE
F(E) (I(E))
p* r l(p)· (4.2)
1
F(B) (I(B))
HB
the horizontal arrows here are as in (2.7), and the vertical arrows
form the appropriate change-of base adjunctions. The diagram (4.1)
obviously commutes, and since the diagram (4.2) is obtained from
(4.1) by replacing all arrows with their right adjoints, it commutes
too, up to a canonical isomorphism.
Definition 4.1. A fibration (A, f) over B is said to be
(a) a trivial covering (of B) , if t he diagram
---HI(A)
!HI(!) (4.3)
------HI(B)
is a pullback, or, equivalently, if the morphism
(4.4)
is an isomorphism.
152 G. Janelidze
(b) split over (E,p), if p*(A, f) is a trivial covering of E.

The categories of trivial coverings of B and of fibrations over B split
over (E,p) will be denoted by TrivCov(B) and Spl(E,p), respectively.
In other words Spl (E, p) is defined as the pullback
Spl(E,p TrivCov(E)
inclusion j j inclusion (4.5)
F(B) F(E)
p*
If the ground Galois structure r is admissible (see Definition 2.4(f)),

then clearly a fibration (A, f) over B is a trivial covering if and only
if there exists a fibration (X, c.p) over I(B) with
(4.6)
that is, the functor H8 induces an equivalence
TrivCov(B) rv CJ!(I(B))
(4.7)
Therefore, under the admissibility condition, the pullback (4.5) can
be rewritten (up to an equivalence) as
Spl(E , p) <f!(I(E))
(4.8)
F(B) F(E)
p*
Since the diagram (4.2) commutes, the equivalence (4. 7) tells us that
the functor p* carries trivial coverings to trivial coverings, and that
TrivCov(B) is obviously contained in Spl(E , p) .
If we ask now, what information on F(B) can be obtained using the
category X, there is a trivial answer - namely the equivalence (4. 7)
that describes TrivCov(B) , but there is also a non-trivial one- namely
Theorem 4.3 below, which describes the larger category Spl(E , p)
whenever (E, p) is a monadic extension in the sense of
Definition 4.2. We will say that (E , p) is a monadic extension , or that

pis an effective descent morphism, if the functor p* : F(B) ---+ F(E)
is monadic.
Theorem 4.3. (The categorical form of the fundamental the-

orem of Galois theory). If the ground Galois structure r is admis-
sible and ( E, p) is a monadic extension of B , th en there is a canonical
category equivalence
Spl(E , p) rv XI( E q(p)) n <f!

(4 .9)
where XI(Eq(p)) n <f! is the full subcategory in X 1 ( E q(p)) with objects all
triples A = ( A 0 , 1r, ~), in which 1r : A 0 ---+ I (E) is a fibration.
The first full proof of this theorem is given only in [32], but as was
already mentioned, the special case proved in [27] (in detail in [28])
is sufficient in order to show that the fundamental theorem of Galois
theory of commutative rings, whose most general version is due to A.
R. Magid [55], extends to a purely categorical context. The same is
shown in [4] on a more elementary level but with some long calculation
omitted.
The (internal) precategory I(Eq(p)) should be called the Galois pre-
category of the extension (E , p) and denoted by Gal(E , p) . It is a
154 G. Janelidze
groupoid if the conditions of Corollary 3.4 are satisfied, and if so, then
it is a group if and only if E is connected, i.e. I(E) is a terminal object
in X.
5 Normal extensions, coverings, and related notions
Using the notation of the previous section and assuming again r to

be admissible, let us recall
Definition 5.1. A fibration (A, f) over B is said to be a covering, if

there exists a monadic extension (E, p) of B such that (A, f) is split
over (E , p); if in addition the morphism I(!) : I (A) ~ I (B) is an
isomorphism, then we will say that (A, f) is a connected covering. A
monadic extension (E, p) of B is said to be
(a) Cartesian-normal, if it satisfies the conditions of Corollary 3.4, i.e.
the morphisms (3.8) and (3.9) are isomorphisms;
(b) normal, if it is split over itself;
(c) a weakly universal covering, if it is a covering and every covering
of B is split over it;
(d) projective, if for every monadic (A, f) extension of B there exists
a morphism g: E ~A with fg = p;
(e) a projective covering, if it is a covering, and for every covering
(A, f) that is a monadic extension of B, there exists a morphism
g : E ~ A with f g = p;
(f) a universal covering, if it is a weakly universal and projective cov-
ering, and every endomorphism of it is an isomorphism.
Remark 5.2. The universal coverings in the sense of [4, Definition
6.6.5(ii)] (which was adopted especially for the theory of ordinary cov-
ering spaces) are not the same as those of Definition 5.l(f) here, but
they always are weakly universal coverings in the sense of Definition
5.1 (c); the converse needs an additional condition, such as closedness
under composition of the class of all covering morphisms.
Clearly every weakly universal covering is normal, and as shown in [30],
every normal extension is Cartesian-normal- and clearly the converse
Recent Developments m Categorical Galois Theory 155
implications are not true in general. It can also be proved that the
existence of a weakly universal covering implies that every projective
covering is weakly universal, and that every connected projective cov-
ering is universal. These and some other links (see also Section 6)
between the notions listed in Definition 5.1 in some sense are visible
already in the following simple cases:
r as in r as in ras in ras in
Example 2.5 Example Examp le 2.9 Example 2.11
2.8
with C = G-Sets , with C as in the last w ith C = Groups and
where G is a sentence there; X = Abelian Groups;
groupj B = 1 B the g round field B any group ;
Trivia l Trivial G - sets All B-algebras that are the e xtensions of B
fibrations
coverings i.e. sets equipped fin ite products that are obtained as
with the trivial of copies of B pullbacks of
action of G abelian exte nsions
of abe lianizations of B
Monadic nonempty Monadic Nondegen e rat ed All extensions
exte nsions G -sets extensions B -algebras of B
Coverings All G -sets All Separa ble B -algebras , Central extensions
fibrations
i.e. finite products of of B
separable field
extensions o f B
Connected Tra n sitive G-sets Isomorph. Separable field Central extensions
coverings with exac t ly one extensions of B (E, p) of B , s u ch t h at
(non empty ) orbit the hom omorp hi s m
HI(p,O ) : HI(E ,D)
--+ Ht (B, D) is an
isomorphis m
Cartesian N onempty G -sets, All monad. Nonempty finite products A ll exte n sions
normal in whi ch e very ex tensions of those field extensions of B
extensions stabilizer of B in wh ich the s ubfield
is a normal of separable elements is a
subgroup Galois extension of B
Normal As above As above Nonempty finite Ce ntral extensions
extensions products of of B
Galo is extensions of B
W eakly Nonempty As above Nonempty finite products Weakl y unive rsal
universal si m pie ( = free) of c opies of the separable central extensions
cove ring s G-sets closure of B , provided that i. e . weak ly ini t ia l
clos ure is a finite o bjects in t h e
extension of B; otherwise category of ce n t ra l
t here are no weak ly extensions o f B
universa l coverings of B
Proj ective As above Projective As a bove As above
monadic
cove rings extensions
Projective As above As a bove Non e mpty finite produc t s Free extensions , i.e.
monadic of c opi es of t h e a lgebraic the extensions ( E, p)
ex tensions closu re of B prov ided t h at of B in which E is a
clos ure is a finite free g roup
extension of B; otherwise
there are no projective
m onadic extension s of B
Universal S imple Projective The se p a r a ble closure of The universal central
monadic
coverings transitive e xte nsion s B, provided that clos ure is exte ns ion of B
a ll w hose
G-sets e ndom. a finite extensio n of B i prov ided B is perfect ,
a re isom. otherwise there a re no i.e. B =[B , B ]
uni versa l co ve rings of B
If (E,p) and (E',p') are fibrations over B and p' factors through p,
then clearly Spl(E,p) <::; Spl(E',p'). Therefore whenever (E,p) is a
156 G. Janelidze
projective monadic extension of B, the category Spl(E,p) coincides

with the category Cov(B) of all coverings of B. Therefore Theorem
4.3 yields:
Corollary 5.3. In the assumptions of Theorem 4. 3, if (E , p) zs pr-o-

jective, then there is a canonical category equivalence
Cov(B) rv XI(Eq(p)) n <]_) • (5.1}
If (E, p) is Cartesian-normal and has the property Spl (E, p) = C ov (B) ,

then the internal groupoid I ( Eq(p)) should be called a fundamental
groupoid of B, and if every morphism in X is a fibration , then it is
uniquely determined up to the Morita equivalence. However it is de-
sirable to make it unique up to the internal equivalence, which could
be achieved by requiring (any of) the additional conditions involved
m
Proposition 5.4. Let (E , p) and (E', p') be Cartesian-normal monadic

extensions of B. If any of the following two conditions holds, then the
Galois gr-oupo ids I (Eq(p)) and I ( Eq(p')) are equivalent as intemal
groupoids:
(a) The morphisms p and p' factor through each other; in particular
this is true if ( E, p) and ( E', p') are both projective;
(b) Spl(E,p) = Spl(E',p') , every morphism in X is a fibration, and
the Morita equivalence of intemal groupoids in X implies their intemal
equivalence; in particular this implication holds for· X = Sets.
6 Galois theory of commutative rings
The notion of a Galois extension of a commutative ring with a finite

Galois group was first introduced by M. Auslander and 0 . Goldmann
[1], and then several equivalent definitions were found and the funda-
mental theorem of Galois theory for these extensions was proved by S.
U. Chase, D. K. Harrison, and A. Rosenberg [15]. The separable clo-
sure of a connected commutative ring was constructed by G. J. Janusz
[51], and its abelian version by D. K. Harrision [25]. Galois groupoids
in connection with idempotents were introduced by 0. Villamayor and
D. Zelinsky, first in the finite and then in the infinite case - as the ti-
tles of [63] and [64] show. Not mentioning many further developments,
such as connections with A. Grothendieck's work [24] on etale cover-

ings (that work, where in some sense Galois t heory of commutative
rings was already discovered, was in fact done long before it was pub-
lished), Galois descent and cohomology, non-commutative separable
algebras and Chase-Sweedler's Hopf Galois extension, Harrison func-
tor and generalized Kummer theory, and separable polynomials, and
restricting our attention to Galois correspondences and related cate-
gory equivalences, we still have to mention three important books:
(i) S. I. Chase and M. E. Sweedler [16] consider the finite Galois exten-
sion of commutative rings as a special case of what they call (cofiat)
Galois objects in a category with finite products and coequalizers.
A general theory of Galois objects is developed in the first chapter,
written by S. U. Chase. Among other things it is shown t hat in the
opposite category of commutative rings every Galois object is cofiat;
however the authors do not make the important conclusion that they
could use torsors instead of the Galois objects (the reader should not
consider this as a critical remark: note that the book was written more
than 30 years ago!).
(ii) F . R. DeMeyer and E. Ingraham's book [18] is mostly " non-
commutative", but it has a chapter on Galois theory of connected
commutative rings.
(iii) A. R. Magid wrote several papers on the extension of Galois theory
of separable algebras from the connected case to the case of arbitrary
commutative rings, and completed that work in his book [55]. Magid's
Galois theory contains the Villamayor-Zelinsky case [64], and in fact
it is the final step in generalizing the separable Galois theory of com-
mutative rings. Magid's notion of a componentially locally strongly
separable algebra perfectly combines the following two possible ways of
involving infinity: on the one hand such an algebra may have infinitely
many idempotents, and on the other hand its Boolean localizations (in
the sense of Pierce spectrum - see [59]) may themselves be an infinite
extension of the corresponding Boolean localizations of the ground
nng.
Let us now briefly recall the main known definitions, constructions,
and results that are needed for or/and belong to Magid's theory; all
rings and algebras below are supposed to be commutative (and with
a unit, and the homomorphisms are supposed to preserve it) , and R
always denotes a ring.
158 G. Janelidze
6.1. (a) Let I demp( R) be the Boolean algebra of idempotents in R .

For each maximal ideal x in Idemp(R) the set Rx = {reir E R , e E
x} is an ideal in R and the quotient ring Rx = RIRx is called the
Boolean localization of R at x . The ring Rx is always connected,
i.e. I demp(Rx) = {0, 1}. If A is an R-module, or an R-algebra,
then Ax is still defined as Ax = AIAx ~ A ® R Rx. Given a ring
homomorphism f : R ---t S and a y in I demp( S) , we will write y n
R = {e E Idemp(R)if(e) E y} and , regarding S as an R-algebra,
also regard Sy as an Rynwalgebra; clearly the Sy is one of Boolean
localizations of SynR ·
(b) An R-algebra S is said to be:
(bt) separable, if it is projective regarded as an S ® R S- module,
(b 2 ) strongly separable, if it is separable and projective as an R-
module, and the canonical homomorphism R ---t Sis injective;
(b 3 ) locally strongly separable, if every finite subset of it is contained
in a strongly separable subalgebra;
(b 4 ) componentially locally strongly separable, if the canonical homo-
morphism R ---t S is injective, and Sx is a locally strongly separable
Rx-algebra for every x E Idemp(R) , or equivalently, Sy is a strongly
separable Rynwalgebra for every y E Idemp(S) ;
(b 5 ) the separable closure of R , if it is componentially locally strongly
separable, and has the following two properties inside of the category
of componentially locally strongly separable R-algebras: every object
has a morphism into it, and every morphism from it is an injective
homomorphism;
(b 6 ) normal, if it is componentially locally strongly separable, and for
every x E Idemp(R), every two R-algebra homomorphisms from S to
the separable closure of Rx have the same image.
(c) For a normal R-algebra S the objects of the Galois groupoid
Gal (S / R) are the maximal ideals of the Boolean algebra I demp( S) ,
and a morphism in Gal(S/ R) from x toy is an R-algebra (iso)morphism
from Sx to Sy (although there is also a good reason to make it a mor-
phism in the opposite direction). The morphisms in Gal(S/ R) can
also be described as maximal ideals in Idemp(S ® R S) , which helps
to make Gal(S/ R) an internal groupoid in the category of profinite
topological spaces. As follows from the results of [27] (the detailed
proofs are in [28]), an R-algebra S is normal if and only if it is normal
in the sense of Definition 5.1 (b) with respect to the Galois structure
described in Example 2.10(a) , and its profinite Galois groupoid is the

same as the I(Eq(p)) considered in the previous two sections, where
p : E ~ B is the canonical homomorphism from R to S regarded as
a morphism S ~ R in the opposite category of commutative rings.
(d) A.R. Magid [55] proves that every ring R admits a unique separable
closure R, and calls its Galois groupoid Gal(R/ R) the fundamental
groupoid of R. One of his important discoveries is that the Boolean
algebra Idemp(R) is larger than the Boolean algebra Idemp(R), and
in fact it is the completion of I demp(R). The Main Theorem IV.31 in
[55] asserts that there is a category equivalence
( Componentially Locally Strongly Separable R-algebras ) 0 P rv
(Profinite Spaces)Gal(R/ R)n (Epimorphisms of profinite spaces) ,
(6.1)
in the notation used in previous sections. The epimorphisms in the
category of profinite spaces are of course just (continuous) surjections,
but moreover, as explained in [55], the space of objects of Gal(R/ R)
is extremely disconnected (which is the same as to say that the corre-
sponding Boolean algebra I demp( R) is complete), which makes every
epimorphism into it a split epimorphism.
After this we list the main additional observation brought by categor-
ical Galois Theory:
6.2 (a) Let r = ((I,H, TJ ,c): C ~ X , F = Mor(C), = Mor(X))
be the Galois structure described in Example 2.10(a). A fibration
(A, f) over B is nothing but a (commutative unitary) B-algebra, i.e.
B and (A, f) play the same roles as Rand S in 6.1. As follows from
the results of [28], (A , f) splits over (E, p) in the sense of Definition
4.1 (b) if and only if for every x E I ( E ® B A) the Boolean localization
(E ® 8 A)x is isomorphic to ExnE· In fact this easily follows from the
results of R. S. Pierce [59] (recall that E ® 8 A is the same as the
pullback E x 8 A in C).
(b) The characterization (a) of split objects helps to prove that the
following conditions on (A, f) are equivalent:
(b 1 ) (A, f) is a covering in the sense of Definition 5.1 ;
(b 2 ) (A, f) is split over the separable closure of B ;
(b 3 ) for x E I(B), Ax is a locally strongly separable Ex-algebra.
(c) The crucial part of the proof of (b) is to show that every strongly
separable algebra is split over some monadic extension. This was done
in [28] using the finite Galois theory of [15] (and probably known to A.
160 G. Janelidze
Grothendieck before) ; however as shown later in [9], even that can be

done categorically. Note also that the categorical notion of separability
used in [9] in fact applies only to the " finite" situations of Example
2.8. The general notion of separability was introduced and studied in
[47], and of strong separability in [49].
(d) As follows from (b) , the Magid 's equivalence (6.1) is a part of a
larger equivalence
(R-algebras satisfying (b3) ) 0 P rv (Profinite Spaces )Gal(R/ R) (6.2)
provided by categorical Galois theory (see Theorem 3.3). Moreover, as
also shown in [28], there is a categorical modification, which directly
gives (5.1). Still, this brings the question for the existence of a class of
morphisms in C = (Commutative Rings )0 P , which declared as the
class of fibrations in C , would form a Galois structure with the class
of epimorphisms in X = Profinite Spaces- and then makes (5.1) a
special case of Theorem 4.3. The answer is affirmative, and there are
several possibilities including faithfully flat extensions (see [50]) .
(e) The separable closure in the sense of A. R. Magid [55] is a non-
connected universal covering in the sense of Definition 5.1 (see 6.1(d)) ,
and no other example like that among the known concrete Galois
theories has been described so far.
(f) In the last (fifth) chapter of [55] the equivalence (5.1) is used to
describe the Galois correspondence for commutative rings, which ex-
tends the classical formulation of the fundamental theorem of Galois
theory. This result has been improved using the categorical approach
in [12].
(g) As shown in [35] the coverings of commutative rings satisfy the
categorical conditions that make them a part of a factorization system,
generalizing the (purely inseparable, separable) factorization for the
finite field extensions - which however satisfy much stronger conditions
of [11] . The problem of describing " the other part" is far from being
solved, but some related results are obtained in [47] and [48].
7 Central extensions of universal algebras
Let r = ((I,H,rJ,c): C-+ X , F ~ Mor(C) , 1> ~ Mor(X)) be the

Galois structure described in Example 2.11. As we already mentioned,
in this case the coverings become central extensions in the sense of [34],
and if C and X were the varieties of groups and of abelian groups,
respectively, then a fibration (i.e. an extension) (A , f) would be a

covering if and only if it is a central extension in the usual sense, i.e.
f is a surjective homomorphism whose kernel is contained in the cen-
tre of A. As shown in [34], the same is true if C was any variety of
D-groups, X a subvariety in C, and the " usual" central extensions
are defined as in A.S.-T. Lue's paper [53] (earlier a special case was
studied by A. Frohlich [20], and after Lue these central extension were
also studied by J. Furtado-Coelho [21]). On the other hand, as sug-
gested by Commutator theory (see also 8.3 below) developed in uni-
versal algebra, an extension (A, f) should have been called central if
[Ax A, A x 8 A] = 1A, i.e. if the commutator of the largest congruence
on A with the kernel-congruence of f is trivial (the universal alge-
braists would write [1 , a] = 0, where a denotes the kernel-congruence
of f; they would also say that 1 and a centralize each other) - such
extensions were called algebraically central in [36] and [37], in distinc-
tion from the central extensions introduced in [34], which hence were
called categorically central. Although neither Frohlich and his follow-
ers, nor universal algebraists seemed to be interested to compare their
notions of central extension with each other, or with the one intro-
duced by S.A. Huq [26], the comparison with the categorically central
extensions has been made in [36] and [37], and then by D. Bourn and
M. Gran [6] and M. Gran alone in [22] and [23] . The results in fact
came rapidly one after another as shown in the following:
Theorem 7 .1. List of Comparison
Since the algebraically central extensions are defined using a com-

mutator, and the most general case where all known definitions of
commutator are known to be equivalent is the case of congruence
modular varieties, we restrict ourselves to that case, i.e. we assume C
to be congruence modular. Another reason for this restriction is the
admissibility mentioned in Example 2.11. The algebraically central
extensions should coincide with the categorically central ones if we
take X to be the subvariety of abelian algebras in C, i.e. the algebras
A with the property [A x A, A x 8 A] = 1A· The fact that every cate-
gorically central extension is algebraically central, was proved in [36],
and the converse was proved
(a) for varieties of n-groups in [36].
(b) then for Mal'tsev (=congruence permutable) varieties in [37],
162 G. Janelidze
(c) and finally for congruence modular varieties in [23].

Since the categorically central extensions were actually studied in [34]
in the context of Barr exact categories (which is more general than the
context of varieties), and since there is a categorical notion of com-
mutator introduced by M. C. Pedicchio [58] for Barr exact Mal'tsev
categories (which were introduced by A. Carboni, J. Lambek and M.
C. Pedicchio [14]; see also A. Carboni, G. M. Kelly and M. C. Pedic-
chio [13]), the two notions of central extensions can also be considered
in any Barr exact Mal'tsev category. The fact that they coincide was
proved
(d) for semi-abelian categories (in the sense of [39]); in fact mainly the
Bourn protomodularity [5] was used) in [6],
(e) and then for Barr exact Mal'tsev categories in [22].
The situation with Huq's notion of centrality is more complicated; it

agrees with the others in " nice" semi-abelian varieties, such as the
varieties of groups or of rings, but not in general (see [7]).
In the notation above, let us now assume for simplicity that C is
a Mal'tsev variety, X any subvariety in C, B an object in C, and
consider the question of existence of a universal covering of B.
Lemma 7.2. Every fibration in C is a Cartesian-normal (monadic}
extension, and every covering is a normal extension.
Proof: The pullbacks involved in Corollary 3.4 are pullbacks of split
epimorphisms, and therefore the functor I preserves them by [22,
Proposition 2.1] - which proves the first assertion; the second one is
proved in [34] (in fact in the more general case of Goursat varieties).
Thus, if (E,p) is a fibration over B with free (or, more generally,
projective) E then it yields a good way to define the fundamental
groupoid of B, namely as I(Eq(p))- and it is determined uniquely up
to an internal equivalence by Proposition 5.4. Let us hence fix such
an (E,p), and write
d
----+
I(Eq(p)) = IT(B) = (II(B)2 ~ IT(B)I ~ II(B)o) (7.1)
~ ~
for the fundamental groupoid. The equivalence (4.1) then becomes
Cov(B) rv xrr(s) n ~ - (7.2)

Considering groupoid actions as a special case of monoid actions in

the appropriate monoidal category of spans, we can speak of II(B)
acting on itself via the multiplication, and write that object of xrr(B)
as (II(B)I, c, m) . Moreover, since according to the construction of
IT(B), the morphism cis obviously a fibration, (II(B)I, c, m) belongs
to XII( B) n and therefore has the corresponding object (E~, p~) in
Cov(B).
Lemma 7 .3. The fibration (E~, p~) above is a projective weakly uni-
versal covering.
Proof: Since every fibration inC is a monadic extension, so is (E~,p~).

The projectivity follows then from the equivalence (7.2) and the fact
that (II(B)I , c, m) is a free II( B) - action. Furthermore, the projec-
tivity implies weak universality by the second assertion of Lemma 7.2
(or using the results of [34] directly).
After that as an easy exercise with internal groupoid actions using the
fact that the diagram
I(E xs E) righ~row I(E) ---+ I(B)

(7.3)
is a coequalizer diagram, we obtain
Theorem 7 .4. The following conditions on an object B in C are

equivalent:
(a) B admits a connected universal covering;
(b) the object I (B) is projective in X ;
(c) the fundamental groupoid IT(B) is internally equivalent to an in-
t ernal groupoid in which every morphism is an automorphism.
Remark 7.5. (a) The covering (E~,p~) considered above is nothing

but the reflection of (E, p) into the category of coverings in the sense
of [35], and, if X was the category of abelian objects in C, then due
to the connection with the algebraically central extensions, it could
be described as
E~ = [E x E, E x 8 E] with the induced p~ - see [37].
(b) If Chad a zero object (which is equivalent t o the requirement that
every algebra has a unique one-element subalgebra), and I(B) was the
trivial algebra, then the universal covering of B would be the initial
164 G. Janelidze
object in the category Cov(B) , and that is what is usually called the
universal central extension. In the case of !1-groups the existence of
universal central extensions is a part of what A. Frohlich's school calls
theory of Baer invariants; we already mentioned the papers [20), [53),
[21), but much deeper links are to be established.
8 Some open questions
8.1 The classification theorem for covering spaces of a " good" space is
shown in detail in [4] to be a special case of what is presented here as
Corollary 5.3 of Theorem 4.3. Does it mean that the notion of Galois
structure is rich enough to be used for developing " abstract homo-
topy theory" = " higher dimensional Galois theory" to be applicable
to other examples of categorical Galois theory? If so, this approach
should be in the same relationship with D. Quillen's approach [60), as
the general adjunctions with the adjunctions formed by the functors
(2.2) and (2.4). One possible way of developing higher dimensional
Galois theory is indicated by a very special example of so-called dou-
ble central extensions of groups [31], further extended in [33].
8.2 A further generalization of Galois theory to so-called variable cate-
gories is proposed in [41] and then an intermediate case studied in [44].
After that A. Joyal- M. Tierney's theorem on presentations of geomet-
ric morphisms of toposes [52] and the Tannaka duality become parts
of (generalized) categorical Galois theory. However it seems that they
should even be special cases of the " straightforward" two-dimensional
version of the very special situation studied in [11] and in [38].
8.3 The universal-algebraic notion of commutator (of two congruences)
was invented by J.D.H. Smith [62] for Mal 'tsev varieties, and then
generalized in various directions by various authors, although all defi-
nitions agree in the congruence modular case; some relevant references
are given in [40), where the commutators are described via internal cat-
egorical structures ( "pseudogrupoids" - rather than " pregroupoids"
used in [58] in [58]). As explained in Section 6 above the notion of
commutator theory perfectly agrees with the categorical notion of cen-
tral extension, at least in the congruence nodular case. However this
comparison makes sense only in a very special case of X being the
category of abelian objects inC (see 7.1). Is it possible to define the
notion of commutator with respect to any subvariety/ " good" reflec-
tive subcategory X? This should probably be related with the theory

of Bear invariants (see the references in 7.5 (b)).
8.4 The existence of a universal covering proved in Section 6 for the
Galois structures in Mal'tsev varieties, and in fact proved in [55] for
the Galois theory of commutative rings, is also known in many cases,
and the proof is especially simple if X = Sets. But what is the most
appropriate general context for such a proof?
8.5 Is there a convenient admissible Galois structure, large enough to
make Magid's equivalence (5.1) and the similar Grothendieck's equiv-
alence for etale coverings of schemes from [24] subequivalences of a
larger equivalence? A good answer to this question will probably in-
volve an infinite version of M. Barr's Galois theory of commutative
rings ([2],[3]) in a topos. It seems that some progress in this direction
has already been made by B. Mesablishvili.
8.6 Unlike the case of commutative rings, where Magid's Galois theory
is obviously "the right one" , it is not at all clear what is the right level
of generality for the theory of covering spaces. For instance, is it possi-
ble to put together naturally (like in 8.5 above?) the locally connected
and the compact case? Recall that these two cases are described in
[4] (for the second one see also [11]) in the categorical language. Note
also that a generalized theory of covering spaces may need further in-
vestigations in topological descent theory, and especially of the etale
descent, and we recall:
(a) The complete description of effective descent morphisms (= monad-
ic extensions) of topological spaces, was obtained by J. Reiterman
and W. Tholen [61] in terms of ultrafilter convergence structure. This
result is mentioned in [45], which together with [46] and [50] provides
a general exposition of descent theory.
(b) Examining the simple case of finite topological spaces (=finite
preorders), a link with old Grothendieck - Giraud type results was
established [42], and the effective etale descent morphisms of finite
topological spaces were characterized [43]. There is a hope - thanks
also to the recent work of M. M. Clementino, D. Hofmann, and W.
Tholen - to generalize these results back to infinite spaces.
8. 7 Answering a question of S. MacLane, the relationship between
the notion of Picard - Vessiot extension and the categorical Galois
theory was established in [29], but no further investigations towards
categorical Galois theory of differential rings/fields have been made,
166 G. Janelidze
and no connection with P. Deligne's work [17] has been established. An

observation about that is made by A. R. Magid in [56]. For instance
I do not know the answer to E. Kolchin 's question: can the strongly
normal extensions of differential fields (in the sense of E. Kolchin) be
described categorically?
References
[1] M. Auslander and 0. Goldman, The Brauer group of a commu-

tative ring, Trans. AMS 97, 1960, 367-409.
[2] M. Barr, Abstract Galois theory, Journal of Pure and Applied

Algebra 19, 1980, 21-42.
[3] M. Barr, Abstract Galois theory II, Journal of Pure and Applied
Algebra 25, 1982, 227-247.
[4] F. Borceux and G. Janelidze, Galois Theories, Cambridge Stud-

ies in Advanced Mathematics 72, Cambridge University Press,
2001.
[5] D. Bourn, Normalization equivalence, kernel equivalence and

affine categories, Lecture Notes in Math. 1488, Springer, 1991 ,
43-62.
[6] D. Bourn and M. Gran, Central extensions in semi-abelian cat-

egories, Journal of Pure and Applied Algebra, 175, 2002, 31-44.
[7] D. Bourn, M. Gran, G. Janelidze and M. C. Pedicchio, On two

commutators in semi-abelian categories, in preparation.
[8] R. Brown and G . Janelidze, Galois theory of second order cover-

ing maps of simplicial sets, Journal of Pure and Applied Algebra
135, 1999, 23-31.
[9] A. Carboni and G. Janelidze, Decidable (=separable) objects

and morphisms in !-extensive categories, Journal of Pure and
Applied Algebra 110, 1996, 219-240.
[10] A. Carboni and G. Janelidze, Boolean Galois theories, Georgian

Mathematical Journal 9, 4, 2002, 645- 658.
[11] A. Carboni, G. Janelidze, G. M. Kelly and R. Pare, On local-

ization and stabilization of factorization system, Applied Cate-
gorical Structures 5, 1997, 1-58.
[12] A. Carboni, G. Janelidze and A. R. Magid, A note on Galois

correspondence for commutative rings, Journal of Algebra 183,
1996, 266-272.
[13] A. Carboni, G. M. Kelly and M. C. Pedicchio, Some remarks on

Mal'tsev and Goursat categories, applied Categorical Structures
1, 1993, 385-421.
[14] A. Carboni, J. Lambek and M. C. Pedicchio, Diagram chasing

in Mal'tsev categories, Journal of Pure and Applied Algebra 69,
1990, 271-284.
[15] S. U. Chase, D. K. Harrison and A. Rosenberg, Galois theory

and cohomology of commutative rings, Mem. AMS 52, 1965,
15-33.
[16] S. U. Chase and M. E. Sweedler, Hopf algebras and Galois the-

ory, Lecture Notes in Math. 97, Springer 1969.
[17] P. Deligne, Categories tannakiennes, "Grothendieck Festschrift"

2, Birkhauser 1990, 111-195.
[18] F. R. DeMeyer and E. Ingraham, Separable algebras over a com-

mutative ring, Lecture Notes in Math. 181, Springer 1971.
[19] Y. Diers, Categories of Boolean sheaves of simple algebras, Lec-

ture Notes in Math. 1187, Springer 1986.
[20] A. Frohlich, Baer-invariants of algebras, Trans. AMS 109, 1963,

221-244.
[21] J. Furtado-Coelho, Varieties of 0-groups and associated func-

tors, Ph. D. Thesis, University of London , 1972.
[22] M. Gran, Algebraically central and categorically central exten-

sions, Coimbra University Math. Preprint 01-02 , 2001.
168 G. Janelidze
[23] M. Gran, Commutators and central extension in universal al-

gebra, Journal of Pure and Applied Algebra 174, 2002, 249 -
261.
[24] A. Grothendieck, Revetements etale et groupe fundamental ,

SGA 1, expose V, Lecture Notes in Math . 224, Springer 1971.
[25] D. K. Harrison, Abelian extensions to commutative rings, Mem.

AMS 52, 1965, 1-14.
[26] S.A. Huq, Commutator, nilpotency and solvability in categories,

Quart. Journal Math. Oxford (2) 19, 1968, 363-389.
[27] G. Janelidze, Magid's theorem in categories, Bull. Georgian

Acad. Sci. 114, 3, 1984, 497-500 (in Russian).
[28] G. Janelidze, The fundamental theorem of Galois theory, Math.

USSR Sbornik 64 (2) , 1989, 359-384.
[29] G. Janelidze, Galois theory in categories: the new example of

differential fields, Proc. Conf. Categorical Topology in Prague
1988, World Scientific 1989, 369-380.
[30] G. Janelidze, Pure Galois theory in categories, Journal of Alge-

bra 132, 1990, 270-286.
[31] G. Janelidze, What is a double central extension? (the question

was asked by Ronald Brown) , Cahiers de Topologie et Geometrie
Differentielle Categorique XXXII-3, 1991, 191-202.
[32] G. Janelidze, Precategories and Galois theory, Lecture Notes in

Math. 1488, Springer, 1991, 157-173.
[33] G. Janelidze, Higher Dimensional Central Extensions: A Cate-

gorical Approach to Homology Theory of Groups, Talk at the
International Meeting in Category Theory, Halifax (Canada) ,
1995.
[34] G. Janelidze and G. M. Kelly, Galois theory and a general notion

of a central extension, Journal of Pure and Applied Algebra 97,
1994, 135-161.
[35] G. Janelidze and G . M. Kelly, The reflectiveness of covering

morphisms in algebra and geometry, Theory and Applications
of Categories 3, 1997, 132-159.
[36] G. Janelidze and G. M. Kelly, Central extensions in universal

algebra: a unification of three notions, Algebra Universalis 44,
2000, 123-128.
[37] G. Janelidze and G. M. Kelly, Central extensions in Mal'tsev

varieties, Theory and Applicat ion of Categories 7, 10, 219-226.
(38] G. Janelidze, L. Marki and W. Tholen, Locally semisimple cov-

erings, Journal of Pure and Applied Algebra 128, 1998, 281-289.
[39] G. Janelidze, L. Marki and W . Tholen, Semi-abelian categories,

Journal of Pure and Applied Algebra, 168, 2002, 367- 386.
[40] G. Janelidze and M. C. Pedicchio, Pseudogroupoids and com-

mutators, Theory and Applications of Categories 8, 15, 2001,
408- 456.
[41] G. Janelidze, D. Schumacher and R. H. Street, Galois theory

in variable categories, Applied Categorical Structures 1, 1993,
103-110.
[42] G. Janelidze and M. Sobral, Finite preorders and topological

descent I, Journal of Pure and Applied Algebra, 175 (1 - 3),
2002, 187 - 205.
[43] G. Janelidze and M. Sobral, Finite preorders and topological

descent II: Etale descent, Journal of Pure and Applied Algebra,
174, 2002, 303 - 309 .
[44] G. Janelidze and R. H. Street, Galois theory in symmetric

monoidal categories, Journal of Algebra 220, 1999, 174-187.
[45] G. Janelidze and W . Tholen, Facets of Descent I, Applied Cat-

egorical Structures 2, 1994, 245-281.
[46] G. Janelidze and W. Tholen, Facets of Descent II, Applied Cat-

egorical Structures 5, 1997, 229-248.
170 G. Janelidze
[47] G. Janelidze and W. Tholen, Functorial factorization , well-

pointedness and separability, Journal of Pure and Applied Al-
gebra 142, 1999, 99-130.
[48] G. Janelidze and W. Tholen, Extended Galois theory and dis-

sonant morphisms, Journal of Pure and Applied Algebra 143,
1999, 231-253.
[49] G. Janelidze and W. Tholen, Strongly separable morphisms in

general categories, in preparation.
[50] G. Janelidze and W. Tholen, Facets of Descent III, in prepara-

tion.
[51] G. J. Janusz, Separable algebras over commutative rings, Trans.

AMS 122, 1966, 461-479.
[52] A. Joyal and M. Tierney, An extension of the Galois theory of

Grothendieck, Mem. AMS 309, 1984.
[53] A.S.-T. Lue, Bear-invariants and extensions relative to a variety,

Proc. Cambridge Philos. Soc. 63, 1967.
[54] S. Mac Lane, Galois theory in categories (work of G. Janelidze),

Talk on International Meeting in Category Theory, Lovain-la-
Neuve (Belgium), 1987.
[55] A. R. Magid, The separable Galois theory of commutative rings,

Marcel Dekker, 1974.
[56] A. R. Magid, Review on " Selected works of Ellis Kolchin"

Bulletin AMS 37, 3, 2000, 337-342.
[57] L. Marki, R. Mlitz and R. Wiegandt, A general Kurosh-Amitsur

radical theory, Communications in Algebra 16, 1988, 249-305.
[58] M. C. Pedicchio, A categorical approach to commutator theory,

Journal of Algebra 177, 1995, 647-657.
[59] R. S. Pierce, Modules over commutative regular rings, Mem.

AMS 70, 1967.
[60] D. Quillen, Homotopical algebra, Lecture Notes in Math. 43,

Springer, 1967.
[61] J. Reiterman and W. Tholen, Effective descent maps of topo-

logical spaces, Topology and its Applications 57, 1994, 53-69.
[62] J . D. H. Smith, Mal'tsev varieties, Lecture Notes in Math. 554,

Springer, 1976.
[63] 0. Villamayor and D. Zelinsky, Galois theory for rings with

finitely many idempotents, Nagoya Math. Journal 27, 1966, 721-
731.
[64] 0. Villamayor and D. Zelinsky, Galois theory with infinitely

many idempotents, Nagoya Math. Journal 35, 1969, 83-98.
Author's address
A. Razmadze Mathematical Institute
Georgian Academy of Sciences
1 M. Alexidze Street
Tbilisi 0193
e-mail:george_janelidze@hotmail.com
The Polarity between Approximation
and Distribution
M. Erne
Abstract
We provide a general framework for the sporadic observation that certain

distributive laws for standard completions are equivalent to certain approx-
imation properties of the underlying ordered sets, as known from the theory
of continuous posets (domains).
Every closure operation /j on a locale (frame) L with join-dense range
X gives rise to a polarity (Galois connection) induced by the relation
(} = { (y, z) : z :::; y =? {jz :::; y }. Using that polarity, we show that the
sublocales of L containing X are precisely the systems Z 6 of all Z-8-closed
elements y (satisfying /jz :::; y whenever z E Z and z :::; y) for so-called
locally approximating subsets of L, i. e. subsets Z ;;;:? X such that for all
x E X and z E Z with x :::; 8z , there is a w E Z with w :::; z and x = 8w.
Thus, the complete lattice of all sublocales containing X is dually isomor-
phic to the complete lattice of all Galois-closed locally approximating sets.
Similarly, we show that in the case of a completely distributive lattice L,
those meet-closed subsets which are completely distributive and contain
X are precisely the systems of Z-8-closed elements for so-called strongly
approximating subsets Z.
These two polarity theorems have manifold applications in order theory
and related fields, for example in domain theory and pointfree topology.
Mathematics Subject Classification: 06A15, 06B35, 06D10, 06D22.
Key words: Approximation, Interpolation, Closure, Distributive, Continu-
ous poset, Frame, Galois connection, Locale, Nucleus, Sublocale.
173
K. Denecke eta/. (eds. ), Galois Connections and Applications, 173-210.
174 M. Erne
0 Introduction
Approximation, interpolation and distribution are central notions of

applied mathematics. In order to prevent misunderstandings a pri-
ori, it should be pointed out that in the context we are concerned
with, the use of these terms is rather general and perhaps a bit vague
from the perspective of applied mathematics. However, theoretical
computer science and other mathematically founded disciplines use
ordered structures as modelling hierarchies of information, represen-
tation, computation etc. In that context, certain auxiliary relations -<
play a basic role, and it is a custom to call them approximating if each
object is the join (supremum, least upper bound) or "limit" of allele-
ments preceding it with respect to-< (see, e.g., [18]). It is known from
lattice theory, domain theory and related fields that such an approx-
imation property is often equivalent to certain (infinite) distributive
laws for the underlying structure or one of its completions.
Notational Remark. Given an arbitrary relation -<, we shall use the
abbreviations
x-< = {y: x-< y}, X-<= {y: :Jx E X(x-< y)},
-< z = {y : y -< z}, -< Z = {y : :Jz E Z (y -< z)}.
If the relation -< is idempotent, i. e. transitive and interpolating (the
latter meaning that x -< z implies x -< y -< z for some y) then, by
an old result due to Raney [23], the complete lattice of all -<-saturated
sets U = U-< is completely distributive, i.e., arbitrary joins distribute
over arbitrary meets. A frequently observed phenomenon is that "ap-
proximation implies interpolation". (We apologize for the "abuse" of
terms that often have a specified meaning in applied mathematics.)
To record two typical examples, let us recall that a complete lattice L
is completely distributive iff L is supercontinous, that is, the relation <J
defined by
x <J y ~ y :::; V W implies x :::; w for some w E W
(where, as usual, V denotes the join) has the approximation property
and, as a consequence, the interpolation property (see [23]). Equiv-
alently, for each element x of L, there is a least downset (lower set,
decreasing set, order ideal) whose join dominates x. Similarly, arbi-
trary meets distribute over directed joins iff the way-below relation«,
The Polarity between Approximation and Distribution 175
given by
x « y {:::} y :::; VW implies x :::; VF for some F ~w W
(where F ~w W means that F is a finite subset of W) , has the ap-
proximation and consequently the interpolation property [18]. In other
words, L is a continuous lattice in the sense of Scott [24]. Perhaps even
more interesting is the fact that a dcpo or domain (a poset in which all
directed subsets have joins) is continuous ( that is, for each element
there is a least directed downset whose join dominates that element)
iff the Scott topology (equivalently, the system of Scott-closed sets)
is completely distributive (see [18, 21]) . The Scott-open sets are then
precisely those of the form U«.
A similar characterization holds for lattices with completely distribu-
tive ideal lattices. Indeed, a much more general theory has been
developed for so-called Z-distributive and Z-continuous lattices and
posets (see, for example, [4, 5, 7, 12, 22, 25]) . Here, instead of arbi-
trary, directed or finite subsets, one considers a general system Z (in
[5, 7, 8, 14] denoted by M) of subsets of a poset P. The Z-join ideals,
i. e., those downsets that contain with any member of Z its join (pro-
vided it exists) form a closure system z v. On the other hand , passing
from Z to the system Z l\ of all Z-downsets
.}Z = {y: y:::; z for some z E Z}
where Z is a member of Z or a singleton {x}, one may assume that
Z is a standard extension, i. e. a collection of downsets including at
least all principal ideals
.}y = {x: x :S: y} (yEP) .
Moreover, there is no essential loss of generality in assuming t hat all
members of Z have joins - in other words, that the underlying poset
is Z-complete. Then, it turns out that z v is completely distributive
whenever Z is strongly approximating, meaning that the Z-below ideals
JJ-zy = n{z E ZA: y :S:V Z}
belong to ZA and, moreover, the Z-below relation, defined by
X <<z Y {:::} X E JJ-zy ,
has both the approximation and the interpolation property. Under
rather mild hypotheses, the converse implication is true as well. (Con-
176 M. Erne
sult the aforementioned references, in particular [5] and [12], for de-
tails.)
Often, one has to work with ordered structures possessing weaker local
approximation properties. For example, the ideal lattice of a join-
semilattice S is known to be distributive (and even a frame) iff for
all finite subsets Z and all x E S with x :=::; V Z, there is a finite
W ~-!-Z with x = VW (cf. [8, 19]). Or, the Scott topology of a dcpo
is a dual frame iff the corresponding condition is fulfilled for directed
instead of finite subsets (for complete lattices, this is equivalent to
meet-continuity; cf. [14, 18]). More generally, it was shown in [8] that
the Z-join ideals of a Z-complete poset P form a frame whenever Z
is locally approximating, that is,
ZEZ and x::;VZ imply x=VW for some WEZ with W~ -t-Z.
Again, under suitable weak assumptions on Z (for example, that Z 11
be closed under binary intersections), the converse also holds true.
A recent result established in [14] is that all T 0 closure systems that
are frames arise as Z-join ideal systems from locally approximating
standard extensions. Similarly, we shall demonstrate in the present
note that all completely distributive T0 closure systems come from
strongly approximating standard extensions in that manner.
An appropriate and elegant concept for these investigations is that of
Galois connections and, in particular, polarities in the sense of Birkhoff
[3]. For surveys on that topic, see [15] and the article Adjunctions and
Galois Connections: Origins, History and Development in this volume.
Recall that polarities are pairs ( <P, \J!) of anti tone (inclusion inverting)
maps between power sets such that the two possible compositions are
extensive. Any such Galois connection between power sets comes from
a unique relation (! between the underlying sets A and B such that
<P(Y)=Y 12 ={zEB:y(!z forall yEY} (Y~A) ,
\J!(Z) = Z 12 ={yEA: y (! z for all z E Z} (Z ~B).
In the present situation, the polarity results from a relation canonically
associated with a given closure operation. Often, this operation is the
cut operator ~ of a poset P = (X, :=::;), assigning to each subset Y of
X the intersection of all principal ideals containing Y. The crucial
relation is then defined by
Y Q Z {::} Z ~ Y implies tlZ ~ Y.

One partner of the induced polarity, restricted to the completely dis-
tributive lattice .C of all downsets of P , maps any collection Z of
downsets to the closure system
zt:>. = {Y E .C: Y l? Z for all Z E Z} = {Y E .C: tl2 Y ~ Y} ,
where
tl2 Y = U{tlZ: Z E Z , Z ~ Y}.
The members of zt:>. are called Z-tl-ideals. T hey coincide with the
Z-join ideals provided P was Z-complete (see [10, 11, 12]) . The other
partner of the polarity sends any system Y of downsets to the system
Yl? = {Z E .C: Yl?Z for allY E Y} = {Z E .C: tlZ ~ fyZ},
where
fyZ = n{Y E Y: Z ~ Y}.
A standard extension Z is then Galois closed with respect to the above
polarity iff Z = (zt:>. )1?, t hat is, Z E .C and tlZ ~ f z.c.Z imply Z E
Z.
Our main result in that context will be that under the above polarity,
those standard completions which are frames bijectively correspond to
the Galois-closed, locally approximating standard extensions, and the
completely distributive standard completions are in one-to-one cor-
respondence with the Galois-closed, strongly approximating standard
extensions.
With regard to topological aspects, the local and global approximation
properties we shall discuss might also be referred to as local or global
density properties, respectively.
1 The Polarity Between Sublocales and Locally

Approximating Sets
We are now going to study such polarities in a more general "frame-
work", namely that of arbitrary frames (locales) instead of lattices
of downsets. A fram e or locale is a complete lattice satisfying the
distributive law
178 M. Erne
x 1\ VY = V(x 1\ Y) (x E L, Y ~ L)
where x 1\ Y is a shorthand notation for {x 1\ y : y E Y} .
From the viewpoint of Galois connections, it is worthwile to note that
a complete lattice L is a frame iff for each x E L , the unary meet
operation z f-----7 x 1\ z has a right adjoint , sending y to the relative
pseudocomplement or residual
x....-.ty = y: x = V{ z E L :x 1\ z ~ y}.
A subframe of L is closed under arbitrary joins and finite meets,
whereas a sub locale Y of L is closed under arbitrary meets (/\-closed)
and under residuation, the latter meaning that y: x lies in Y for all x E
L and y E Y. For more background about frames , see, for example,
[20].
Henceforth, we are fixing
- a frame L with order relation ~ '
- a closure operation b on L , and
- a subset X of the range L 8 = b[L] that is join-dense in L.
Thus, each element of L is a join of elements from X. For later use,
we note the known fact that a join-dense subset X of L is /\-closed
in L (that is, for each subset of X , the meet formed in L belongs to
X) if and only if X is a complete lattice with respect to the induced
order. It is an experience in order and category theory that often
the reduction to join-dense subsets simplifies constructions and proofs.
This will be a guiding principle in the subsequent considerations. Most
of the definitions will involve the closure operation b, and some of
them also the join-dense subset X, although that dependence will not
be mentioned explicitly.
It will be convenient to write bx for b(x) etc. Recall the most succinct
characterization of closure operations by the equivalence
x ~ by {::} bx ~ by.
Every subset Y of L gives rise to a closure operation IV defined by
/YX = 1\ {y E Y : X ~ y },
and Y coincides with £ 1Y iffY is /\-closed in L. More generally, for
every preclosure operation, that is, for every map f3 : L --t L satisfy-
ing
x :::; y =? f3x :::; f3y and x :::; f3x,

the set
£!3 = {x E L: f3x = x}
of all fixpoints is /\-closed , and the corresponding closure operation
rLf3 may be constructed by transfinite iteration of /3, setting
/3 = /3, /3"+ = /3 o {3" , and /3>-x = V{f3"x: K,:::; ..\}

1 1
for limit ordinals ,.\. By set-theoretical restrictions, there is a first

ordinal /'1, such that {3" = /3"+ 1 , and this map will be denoted by {-J.
It is the smallest closure operation that dominates f3 (pointwise) and
coincides, therefore, with the closure operation of £ !3 = £ 73. In the
case of (pre )closure operations on power sets, we shall speak, as usual,
of (pre)closure operators, and /\-closed subsets of power sets P A are
referred to as closure systems on A . (Often, that terminology is also
used in the general context of complete lattices.)
Given a fixed subset Z of L with X ~ Z, we frequently shall work
with the preclosure operation
52 : L---+ L, y 1----'t V{5z : z E Z, z :S y }.
Its fixpoints, referred to as Z-5-closed elements, form a /\-closed subset
Z"={yEL:zEZandz:Sy imply 5z :Sy} ,

and the corresponding closure operation is 52 = / zJ . Without partic-
ular emphasis, we shall make use of the inequalities
52 y:::; (5 2 ry:::; 52 y:::; 5y for all y E L
and the equalities
52 z = (5 2 rz = 52 z = 5z for all z E Z.
A fundamental ingredient for our investigations will be the relation
Q={(y, z )ELxL:z:Sy =? 5z :Sy}.
To capture the dependence on 5, the letter{! may be read as an "upside
down 5". The induced polarity consists of the two maps
Yl----'t Y{J = {z E L : Vy E Y (yQz)} = {z E L: 5z :S /YZ } and
Zl----'tZ 8 = {y E L: Vz E Z (yQ z )} = {y E L: 52 y:::; y}.
Observe that the other inequality y :::; 52 y holds by join-density of Z 2
180 M. Erne
X, and that the inequality r YZ :::; 6z holds whenever L 0 is contained

in Y. In that situation, one has
ye = {z E L: 6z = r YZ} .
In particular,
£0 =£e.
One of the two possible compositions gives the Galois closure
zoe= {z E L: 6z = 6Z z } ,
and Z is Galois closed if Z = zoe, which means that 6z = 6Z z implies
z E Z. By the usual rules for Galois connections,
Y C
-
zo {::} Z C- ye l
each set of the form ye is Galois closed, and if Z is Galois closed then
y = zo ::::} z= ye .
Notice that zo is always /\-closed and contains L 0, whereas the ex-
ample below shows that Galois-closed subsets need not even be 1\-
subsemilattices (but see Lemmas 6 and 12).
Example 1. Let Dn denote the lattice of all divisors of the natural

number n, and consider t he downset frame L of D 36 . (In the present
context, a downset contains with any number all divisors of it). The
principal ideals of D 36 (in other words, the divisor lattices Dn for
n I 36) together with the set {1 , 2, 3} form a closure system Y of
downsets, hence a /\-closed subset of L. Denote the corresponding
closure operator by 1 and the restriction of the cut operator of D 36 to
L by 6. Then, with respect to the polarity associated with 6, we have
a Galois-closed set
z = ye = {z E L : rZ = 6z } .
The downsets a = {1, 2, 3, 4} and b = {1 , 2, 3, 9} belong to Z since
1a = 6a = D 12 and r b = 6b = D 18 , but an b is not in Z because
r(a n b) = {1, 2, 3} # D 6 = 6(a n b) . Hence, Z is Galois-closed but
not n-closed.
However, as in every Galois connection, the Galois-closed sets arising

from any closure operation 6 via the relation 12 do form a closure
system.
Henceforth, we are fixing a set Z with X ~ Z ~ L. First, we list

some necessary and sufficient conditions ensuring that 8 preserves fi-
nite meets of elements from Z (cf. [8) for the case of set-theoretical
closure operators). Denote by )..Z the /\-subsemilattice of L generated
by z.
Lemma 1. The following statements are equivalent:
(a) y E L , z E Z and y :::; 8z imply y :::; 8(y !\ z ).
(b) xEX,zEZ andx:s;8z imply x=8(x !\z).
(c) y !\ 8z1!\ ... !\ 8zn:::; 8(y !\ z1 !\ ... !\ Zn) for yEL and ziEZ .
(d) 8 preserves finite meets of elements from Z.
(e) 8 induces a !\-homomorphism from :>..Z to L 15 •
Proof. (a)==> (b) is clear by the inclusion X ~ L 15 , which means x = 8x
for all x E X.
(b)=>(c) is shown by induction. For n = 1, use join-density of X
and the hypothesis that for x E X, the inequality x :::; y !\ bz implies
x = 8(x !\ z ) :::; 8(y !\ z) .
Assume the claim has been verified for n . Then, for Zn+l E Z, the
previous argument and the induction hypothesis give
y !\ 8z1 !\ ... !\ 8zn !\ 8zn+l :::; 8(y !\ 8z1 !\ ... !\ 8zn !\ Zn+l)
:::; bb(y !\ Z1 !\ ... !\ Zn !\ Zn+l) = 8(y !\ Z1 !\ ... !\ Zn+l)·
The implications (c)==> (d)==> (e)==> (a) are straightforward. 0
By a pseudonucleus on L, we mean a preclosure operation {3 : L -----+ L

such that
x !\ y :::; z = {Jz implies x !\ {Jy :::; z,
which is tantamount to the inequality
x!\{3y :::;/3(x!\y) ,
because /3(x !\ y) is the smallest z with x !\ y :::; z = {Jz . A prenucleus
(cf. [1]) is a preclosure operation {3 satisfying the slightly stronger
condition
x!\{Jy :::; f3(x!\y),
and a nucleus is a closure operation that preserves binary (hence fini-
tary) meets. In the implication chain
182 M . Erne
nucleus
-0-
extensive /\-homomorphism
-0-
prenucleus
-0-
pseudonucleus
-0-
preclosure operation
none of the arrows may be inverted in general. However, an idem-

potent pseudonucleus j3 is already a nucleus, since j3 = 7J entails
x 1\ j3y::::; j3(x 1\ y) and then, replacing x with j3x,
j3x 1\ j3y::::; j3(j3x 1\ y) ::::; j3j3(x 1\ y) = j3(x 1\ y) ::::; j3x 1\ j3y.
Next, we improve two known elementary facts, namely: (1) that for
any prenucleus {3, the associated closure operator 7J is a nucleus [1],
and (2) that the nuclei bijectively correspond to the sublocales [20].
Lemma 2. The following conditions on a preclosure operation j3 :

L ---+ L are equivalent:
(a) j3 is a pseudonucleus.
(b) 7J is a nucleus.
(c) £!3 = £73 = {z E L: j3z = z} is a sublocale of L.
In particular, the dual isomorphism r 1----7 L '"Y between the complete
lattice of all closure operations on L and that of all 1\ -closed subsets
of L induces a one-to-one correspondence between nuclei and sublocales
of L.
Proof. (a)=?(c). If z = j3z E £!3 then for y = z:x, one obtains

x 1\ y::::; z, hence x 1\ j3y::::; z, i.e. j3y::::; z:x = y E £ !3 .
(c)=?(a). Ifxl\y::Sz=j3z theny::Sz:xEL!3,i.e. j3(z:x):s;z:x
and therefore x 1\ j3y::::; x 1\ j3(z:x)::::; z .
The equivalence (b) {::} (c) is now obtained for 7J instead of j3. o
If £!3 contains the join-dense subset X , one can say more:
Lemma 3. Let {3 : L ------t L be any preclosure operation whose set of

fixpoints, Y = £f3, contains the join-dense subset X, and let r be the
corresponding closure operation (thus r = 7J and Y = L 7 ) . Then the
following statements are all equivalent:
(a) {3 is a pseudonucleus.
(b) r is a nucleus.
(c) r induces a fram e homomorphism from L onto £f3.
(d) r preserves finite m eets.
(e) £f3 is a sub locale of L.
(f) £f3 is a frame (with respect to the induced order).
(g) x E X and y E L imply X 1\ rY = r( x 1\ Y) .
(h) x EX, y E L and x::; rY imply x = r(x 1\ y).

(i) x EX andy E Limply x 1\ {3y::; r(x 1\ y).
(j) x EX, y E L and X::; {3y imply x = r(x 1\ y) .
Proof. In the implication circuit

(a)=? (b) =? (c)=? (d) =? (e)=? (f)=? (g)=? (h) =? (i) =? (j) =?(a),
only the following three implications remain to be verified:
(f) =? (g). Joins in the complete lattice Y are given by Vy W
r(V W). Let x E X and y E L. The inclusion X ~ Y and the
join-density of X together with the frame property of Land Y yield
r(x 1\ y) = r(V{ x 1\ x': x' E X,x'::; y}) =
Vy { x 1\ x' : x' E X, x' ::; y} = x 1\ Vy { x' E X : x' ::; y} =
x 1\ r(V{ x' EX: x' ::; y}) = x 1\ rY ·
(h) =? (i). For x' EX with x' ::; x 1\ {3y, it follows that x' ::; r Y, hence
x' = r (x' 1\ y) ::; r (x 1\ y). Now, join-density of X yields X 1\ {3y ::;
r(xl\y) .
(j) =?(a). Analogously. o
Our primary goal is to find handy conditions on r5 under which t he set
Z 8 of all Z-6-closed elements becomes a frame . As we shall see, it is
sufficient (but not necessary) to require that
184 M. Erne
for all x E X and z E Z with x :S: 8z,

there is a w E Z with w :S: z and x = 8w,
in which case we say Z is locally approximating; on the other hand,
the same condition for w E L instead turns out to be necessary (but
not sufficient) for Z 8 to become a frame. Therefore, we also need
certain intermediate properties and call Z locally {3-approximating for
any {3 : L ---+ L if
for all x E X and z E Z with x :S: 8z,
there is a w E L with w :S: z and x = {3w.
Apparently, that condition may be condensed to the following:
.1: E X, z E Z and x :S: 8z imply x = {3(x 1\ z),
which in turn, under the hypothesis {3 :S: 8, is equivalent to
y E L, z E Z andy :S: 8z imply y :S: {3(y 1\ z),
by join-density of X ~ L 8 . Thus, the local {3-approximation prop-
erty may be defined without reference to the selected set X. Observe
also that Z is locally 8-approximating iff it satisfies the equivalent
conditions in Lemma 1.
Finally, call Z locally /\-approximating if
x E X, z E Z and x :S: 8z imply x = 8(x 1\ z) and x 1\ z E Z
and note the following implications:
locally /\-approximating
-ll-
locally approximating
-ll-
locally 8z -approximating
-ll-
locally 8Z -approximating
-ll-
locally 8-approximating
None of these implications may be inverted, not even in the rather

specific "standard situation" known from domain theory, where L is
the frame of all downsets of a dcpo, 8 is the cut operator, X is the set
of all principal ideals, and Z is the standard extension of all directed
downsets (see the introduction and Section 4, and for counterexamples,

[13]).
However, if Z is closed under binary meets, the five properties in the
above implication diagram are equivalent. (Indeed, x E X ~ Z , w :::; z
and x = 6w entail w :::; x 1\ z :::; x = 6x , hence x = 6(x 1\ z) and
X 1\ z E Z.)
Corollary 1. Z is locally 6-approximating iff the 1\-subsemilattice

generated by Z is locally (1\-) approximating.
Therefore, it is often possible to assume that Z be closed under finite

meets (/\-closed for short) , which considerably simplifies the theory
(see, for example, Corollary 2) .
Lemma 3, applied to f3 = 5z and '"'/ = 5z , yields the desired necessary
and sufficient criteria for the frame property of Z 8 .
Proposition 1. The following statements are equivalent:

(a) 5z is a pseudonucleus.
(c) 6Z is a nucleus.
(d) 6Z induces a frame homomorphism from L onto Z 8 .
(e) Z 8 is a sublocale of L.
(f) Z 8 is a frame (with the induced order) .
(e) Z is locally 6Z -approximating.
Each of these conditions is fulfilled if Z is locally approximating.
Lemma 4. A map f3 : L ----+ L with f3 :::; 6 satisfies the inequality

x 1\ 6zy:::; f3(x 1\ y) for all x,y E L
iff Z is locally /3-approximating.
Proof. If the stated inequality holds then x E X , z E Z and x :::; 6z

imply x = x 1\ 6z = x 1\ 5z z :::; f3(x 1\ z) :::; f3x :::; 6x = x.
Conversely, suppose Z is locally /3-approximating. Since X is join-
dense in L, we get for arbitrary x, y E L:
x 1\ 6zy 1\ V{6z : z E Z , z :::; y} = V{x 1\ 6z : z E Z , z :::; y}
=x
= V{x' EX: :Jz E Z(x':::; xl\6z,z:::; y)}
186 M. Erne
::; V{x' EX: ::Jz E Z (x' = (J(x'l\ z)::; x, z::; y)}

::; (J(x 1\ y). o
The special case (3 = c5z amounts to
Proposition 2. c5z is a prenucleus iff Z is locally c5z -approximating.
In case X is the least join-dense subset, consisting of all V-irreducible

elements (a situation to be discussed more thoroughly in Section 4),
we have another convenient coincidence:
Lemma 5. If X consists of V-irreducible elements then every locally

c5z -approximating subset Z is already locally approximating.
This holds because for V-irreducible x E X ~ Z,

x = c5zw = V { c5v : v E Z, v::; w} implies
x = c5v for some v E Z with v::; w.
As a consequence of the previous facts, we note:
Proposition 3. c5z is a nucleus (and Z 8 is the associated sublocale)

iff Z is locally c5z -approximating and c5z [L J is contained in Z 8 . If X
consists of V-irreducible elements then these two conditions are also
equivalent to the local approximation property for Z combined with
idempotency of c5z.
Corollary 2. If Z is /\-closed then the following conditions are equiv-

alent:
(a) Z 8 is a fram e (and a sublocale of L).
(b) c5z is a nucleus.
(c) c5z is a prenucleus.
(d) c5 induces a /\-homomorphism from Z into L 8 .
(e) Z is locally (/\-)approximating.
An example in [13] shows that if Z fails to be /\-closed, it may be

locally approximating but not locally /\-approximating.
Recall that Z is Galois-closed iff c5z = c5z z implies z E Z (and that
the converse implication is always true). For Galois-closed sets, the
situation is simpler, on account of
Lemma 6. Every Galois-closed locally oz -approximating subset Z is

/\-closed, hence locally 1\-appmximating.
Pmof. For y, z E Z, we compute, using Proposition 1:

O(y 1\ z)::::; 6y 1\ 6z = oZy 1\ oZz = 6Z(y 1\ z ) ,
Hence, y 1\ z is a member of the Galois-closed set Z. Clearly, the top
element of L belongs to any Galois-closed subset. 0
We now show that every sublocale containing Lli arises in the form
zli. Simultaneously, we generalize the fact that Lli is a sublocale iff L
is locally appmximating (see Lemma 3) .
Proposition 4. A subset Y of L is a sublocale containing Lli iff Z =

ye is locally ( /\-) appmximating and satisfies Z li = Y.
Pmof. Z = Yl1 = {z E L : 6z = r Y Z} is Galois closed. Suppose Y

is a sublocale with Lli ~ Y and consider elements x E X , z E Z with
x ::::; 6z . As 6z = rYZ, Lemma 3 yields
x = rY(x 1\ z) ::::; 6(x 1\ z ) ::::; 6x =x,
and sox= rY(x 1\ z) = 6(x 1\ z ), whence x 1\ z E ye = Z. This already
shows that Z is locally /\-approximating. The inclusion Y ~ yeli = zli
is clear. For the converse inclusion, consider y E zli and x E X with
X ::::; r¥Y· Setting z = X 1\ y , we obtain (using Lemma 3 as before)
X= rYZ = 6z, whence z E ye = Z , and z ::::; y yields X ::::; oZy = y (by
the choice of y E zli) . Again by join-density of X , we conclude that
r¥Y::::; y E Y.
On the other hand, Proposition 1 tells us that for all locally approx-
imating subsets Z, the set Y = zli is a sublocale (with Lli ~ Y).
0
Corollary 3. Z is Galois closed and locally (/\-)approximating iffY=

zli is a sublocale of L with Z = ye.
Let us call an element z E L o-nuclear if for all x E X , x ::::; oz

implies x = 6(x/\z). Alternatively, we may describe 6-nuclear elements
without reference to the join-dense subset X , by postulating that y ::::;
6z implies y ::::; 6(y 1\ z ) , or even shorter, that y 1\ 6z ::::; o(y 1\ z ) for
o
ally E L. Clearly, each fixpoint of is 6-nuclear, and by Lemma 3, o
is a nucleus iff each element of L is o-nuclear. Furthermore, we note
188 M . Erne
that a subset is locally 6-approximating just iff it consists of 6-nuclear

elements, and conclude:
Lemma 7. A Galois-closed subset of L is locally approximating iff

it consists of 6-nuclear elements. Hence, the Galois-closed locally ap-
proximating subsets of L form a closure system whose greatest member
consists of all 6-nuclear elements.
Proof. Only the last claim requires a verification. The set N of all
6-nuclear elements is /\-closed by Lemma 1. It is a sublocale by Propo-
sition 1, and by Proposition 4, its closure N 8 e is a Galois-dosed locally
approximating subset containing N . By maximality, N must coincide
with its Galois closure. o
Summarizing the previous results, we arrive at
Theorem 1. Let L be a fmme with a specified closure opemtion 6

whose mnge L 8 is join-dense in L. Then the sublocales containing L 8
are precisely the sets of all Z -6-closed elements for locally approximat-
ing sets Z.
Furthermore, the polarity associated with 6 induces a dual isomor-
phism between the closure system of all sublocales of L containing L 8
and the complete lattice of all Galois-closed locally approximating sub-
sets of L.
Corollary 4. The sublocales of L containing L 8 are precisely the sets

Z 8 where Z runs through all 1\-subsemilattices of L consisting of 6-
nuclear elements. Moreover, the 6-nuclear elements form the largest
sublocale Y of L so that 6 restricted to Y becomes a nucleus.
2 Completely Distributive Lattices and

Approximating Subsets
We now turn to the "global" analogues of the previous results, with
the same hypotheses as at the beginning of Section 1, but this time
focussing on maps and distributive laws that involve infinite meets.
Although , at first glance, one might conjecture that the main steps in
the local theory developed in Section 1 have their strict analogues for
the global situation (and this will be justified in a few instances) , there
are some situations where the global analogues fail at least partially
or are more complicated. For example, Lemma 3 has no exact coun-
terpart for maps that preserve arbitrary meets. To be more explicit: it
is not true in general that if a prenucleus (3 preserves arbitrary meets,
the associated nucleus 73 would also preserve arbitrary meets.
Example 2. Consider the complete chain

L = { +n : n E w +1= w U { w}}
and the prenucleus (3 with (Jw = w, (3( -w) = -w and (Jn = n + 1
otherwise. Apparently, (3 preserves arbitrary meets. But this no longer
holds for the nucleus'"'! = 73, because '"'! ( -w) = -w , whereas '"'(X = w for
all other x E L. In the complete lattice of all nuclei on L , we observe
that '"'( is the join of the nuclei '"'fn with '"'inn = n + 1 and '"'fnX = x
otherwise, where n runs through the integers, and each '"'in preserves
arbitrary meets!
As in the situation described in the previous section, we consider a

join-dense subset X of a complete lattice Land a subset Z of L con-
taining X. Inspired by the theory of Z-continuous posets and their
Z-below ideals (cf. [4, 12, 22, 25, 26] and Section 4) , we define the
Z -core of y E L (with respect to the given closure operation <5) by
Ozy = 1\{z E Z: y :S oz },
and the Z -below part of y by
f2zY = V{ozx: x E X ,x <J y}.
Recall that x <J y means that the element x belongs to each downset
whose join dominates y, and note that Ozx coincides with {}zX when-
ever xis a V-prime element of L (i.e. x <J x EX) . The least /\-closed
subset containing Z is given by
AZ = {I\ W : W ~ Z}.
From the definition of Oz , one concludes that Oz = OAz and that
Ozy:::; OzOY :::; y for ally E AZ. If o is the identity map then Oz is the
closure operation '"Yz associated with Z. Also, by definition , we have
{}zY :::; Ozy and {}zY :::; y for all y E L. Moreover, since x <J V W is
equivalent to x <J w for some w E W, we see that {}z preserves joins.
In all, we have:
190 M . Erne
Lemma 8. Qz is a dual preclosure operation that preserves arbitrary

joins. It has therefore a right adjoint that is a preclosure operation
preserving arbitrary meets, in particular a nucleus.
In Proposition 6, we shall find necessary and sufficient conditions un-

der which the right adjoint of {!z is just Oz.
Let us begin our investigation with the global counterpart of Lemma
1.
Lemma 9. The following statements are equivalent:

(a) y ~ OOzY for each y E L.
(b) x = OOzx for each x EX .

(c) y 1\ 1\ o[W] ~ o (y 1\ 1\ W) for all y E L and W ~ Z .
(d) o preserves arbitrary meets of elements fmm Z .
(e) o induces a 1\ -homomorphism between AZ and L 8 .
(f) The restriction o : AZ---+ L 8 has a left adjoint (oz).
Pmoj. The implication circle (a)=:;.(f)=}(e)=:;.(c):::;.(d):::;.(b):::;.(a) is
easily established: for (a)=:;.(f) , recall that OzOy ~ y for y E AZ. For
(b)=:;.(a), use join-density of X, and for (e):::;.(c), apply the corre-
sponding implication in Lemma 1 to z1 = 1\ W (n = 1) and AZ instead
of Z, using the equation oz1 =(\o[W]. o
Next, let us note that the following two conditions on a subset X of a
complete lattice L are equivalent:
• L is completely distributive and X is join-dense in L
• X is a stmng base, i.e. y = V{xEX: x <J y} for all yEL.
Henceforth, we shall assume that
L is a completely distributive lattice with a strong base X ~ Z.
Consequently, the relation <J is interpolating. As for frames , there
is a nice characterization of complete distributivity in terms of Galois
connections: a complete lattice Lis completely distributive iff the join
map V from the downset lattice onto L has a left adjoint (viz. the
"well-below" map y f-----t <Jy) . Clearly, in every complete lattice, the
join map has a right adjoint, the principal ideal map y f-----t .!,. y = {x E
L:x~y} .
As mentioned earlier, Lemma 3 can be transferred to the global level

only in part.
Lemma 10. For a closure operation 'Y on L with join-dense range
Y = L 1 , the following statements are equivalent:
(a) 'Y -l [ -1.- y] is a principal ideal for each y E L .
(b) 'Y has a left adjoint.
(c) 'Y induces a complete homomorphism onto Y.
(d) 'Y preserves arbitrary meets.
(e) Y is a completely distributive sublocale of L.
(f) Y is a completely distributive lattice.
Hence, a join-dense subset of L is a completely distributive complete
lattice iff it is a sublocale whose nucleus pr·eserves arbitrary meets.
Proof. The equivalences (a){:}(b){:}(c){:}(d) are well-known (cf. [15]).
For (c)=} (e) , apply Lemma 3 and the fact that complete homomor-
phisms preserve complete distributivity. Clearly, (e) implies (f).
(f)=}(d). Let W ~Land put Yw = {y E Y: y ::; w} for wE W.
Then each Yw is a downset in Y. By join-density of Y , we get "'(W =
'Y(V Yw) = V y Yw, and by complete distributivity,
1\ {"'(W : w E W} = 1\ y { Vy Yw : w E W} =
vyn{Yw: wE W} = Vy{yEY: y:::::: 1\ W} =
'Y (V{Y E Y: y::; 1\ W}) = 'Y (/\ W). o
Without the density assumption , condition (f) is strictly weaker than
the other statements (a) - (e).
Now, we formulate the global counterparts of the local approximation
properties introduced in Section 1. Given a map (3 : L ----7 L with
(3 ::; <5, we say Z is (globally or uniformly) (3-approximating if
x = f38z x for all x E X ,
(globally or uniformly) approximating if
x = 88zx and 8zx E Z for all x E X,
and strongly ((3-) approximating if, in addition , 8z is idempotent (hence
a closure operation). Note that corresponding local and global approx-
imation properties differ only by the position of quantifiers:
192 M . Erne
Z is locally approximating iff

VxEX\fzEZ3wEZ(x = 6w and (x::;6 z =? w:Sz)) ,
Z is (globally) approximating iff
\fxEX3wEZ\fzEZ(x = 6w and (x:S6z =? w:Sz)) ,
and similarly for the ,8-approximation properties. In particular, every
(,8-) approximating set is locally (,8-) approximating. Notice that in
the second description of approximating sets, w must be the Z-core
of x. But even a strongly approximating set need not be locally /\-
approximating (the condition x/\z E Z for z E Z may fail; see Example
4 at the end). However, Z is approximating iff so is AZ, and clearly
x 1\ z is in AZ for x E X and z E AZ. Hence, in many instances of the
global approximation theory, there is no essential loss of generality in
assuming that Z be /\-closed.
By definition, Z is 6-approximating iff it satisfies the equivalent con-
ditions in Lemma 9. In particular, the whole lattice L is (strongly)
approximating iff 6 preserves arbitrary meets, i.e. LJ is completely
distributive (cf. Lemma 10.)
Similarly as in the local case, we have the following implications be-
tween the various global approximation properties:
strongly approximating
-0-
( globally) approximating
-ll-
6z -approximating
-ll-
6Z -approximating
-0-
6 -approximating
If Z is 1\-closed then 6z x belongs to Z for all x E X, so that the last

four approximation properties are equivalent. From Lemma 9 and the
equation 6z = 6Az, we immediately infer:
Corollary 5. Z is ,8 -approximating iff so is the 1\ -closed subset AZ

gen erated by Z . In par·ticular, the following three statements are equiv-
alent:
(a) 6 preserves arbitrary meets of elements from Z .

(b) Z is 6 -approximating.
(c) AZ is (6-) approximating.
Lemma 9, applied to the closure operation 6Z , amounts to
Proposition 5. Z 6 is completely distributive (and a sublocale of L)

iff 6Z has a left adjoint, that is, 6Z preserves arbitrary meets (and
induces therefore a complete J!!!momorphism onto Z 6 ). Each of these
conditions implies that Z is 6Z -approximating.
Likewise, the global counterpart of Proposition 2 involves certain ad-

joint maps.
Proposition 6. The following conditions are equivalent:

(a) f}z is left adjoint to 6z .
(b) 6z preserves arbitrary meets.
(c) Z is 6z -approximating.
Proof. (a)==>- (b). Right adjoints preserve meets.

(b)==>-(c). By the inclusion X~ L 6 , we get for x EX:
6z6zx:::; 6x = x:::; f\{6z: z E Z,x::; 6z} =
/\{6zz: z E Z,x::; 6z} = 6z6zx.
(c)==>-(a). Suppose y:::; 6zw = V{6z: z E Z , z:::; w}. For x EX
with x <l y, we find a z E Z such that z :::; w and x :::; 6z . Hence
6zx :::; z :::; w, and f2zY = V{6zx : x E X, x <l y} :::; w. (For this part,
the hypothesis (c) is not needed.)
Conversely, assume f2zY :::; w , i. e. 6zx :::; w for all x E X with
x <l y. Any such x satisfies x = 6z6zx :::; 6zw. By the strong base
property of X, it follows that y = V{x E X : x <l y} :::; 6zw. In
all, we have established the required adjunction equivalence f2zY <
w 9 y :::; 6zw. o
In analogy to Lemma 5, we have
Lemma 11. If X consists of V-ir-reducible elements then every 6z-

approximating subset is already approximating.
194 M . Erne
Proof. If xis V-irreducible then x = 6z6zx implies x = 6z for some

z E Z with z ::; 6zx. It follows that 6zx = z E Z and x = 66zx . o
The conditions in Proposition 6 imply that Z is 6Z-approximating,
but what about the converse? We have not settled that problem in
general, but results in [12] show that under reasonable circumstances
(in domain theory, say) , the properties of 6z- and 6Z-approximation
are equivalent.
Combining the previous results with the fact that one partner of an
adjunction is idempotent iff so is the other, we obtain:
Proposition 7. The following statements are equivalent:
(a) zb is the range of 6z and completely distributive.
(b) 6z is a closure operation preserving arbitrary meets.
(c) 6z induces a complete homomorphism from L onto zb.
(d) (! z is idempotent and left adjoint to 6z .
(e) Z is strongly 6z -approximating.
Each of these conditions is fulfilled when
(f) Z is strongly approximating,
and if X consists of V-irreducible elements, (f) is also a consequence
of the other conditions.
From Corollary 5 and Propositions 6 and 7, we infer:

Corollary 6. Let Z be a A-closed subset of L containing Lb, and
consider the following conditions:
(a) zb is completely distributive.
(b) zb is a sub locale whose nucleus preserves arbitrary meets.
(c) 6Z induces a complete homomorphism from L onto zb.
(d) 6z preserves arbitrary meets.
(e) 6 induces a A-homomorphism from Z onto Lb .
(f) Z is approximating.
In general, (a){:}(b){:}(c):=>(d){:}(e){:}(f). If 6z is idempotent, all
six conditions are equivalent to the strong approximation property of
z.
Thus, for /\-closed Z, the situation becomes considerably simpler, but

remains still more complicated than in Corollary 2. Example 2.7 in [5]
shows that even for /\-closed approximating subsets Z, the /\-closed
subsets Z 6 need not be completely distributive, although we know
from Proposition 3 that Z 6 must be a sublocale, hence a frame in its
own right. Another example in [5] shows that complete distributivity
of Z 6 is not sufficient for idempotency of bz .
The global counterpart of Lemma 6 is more involved, too. What we
need is a handy criterion ensuring idempotency of bz or {!z, respec-
tively. The following definition will be helpful: Z is (12z- )invariant if
12z[Z] ~ Z.
Lemma 12. (1) If Z is '"'( -approximating for some closure operation
'Y on L with '"'( :S b then '"'( = '"'( f2z·
(2) Every invariant b-approximating subset Z of L is strongly approx-
imating.
(3) Every Galois-closed (bZ-)approximating subset Z of Lis /\ -closed,
invariant and strongly approximating.
Proof. (1) 'YY = 'Y(V{x EX: x<J y}) =
'Y(V{'"'fbzx: x E X ,x <J y}) =
'Y(V{bzx: x E X ,x <J y}) = 'YI2zY·
(2) x E X ~ Z implies z := {!z{!zX E Z , by invariance of Z. Using
(1) for b instead of'"'(, we get bz = b {!zX = bx = x . The definition of
bz yields bzx :S z :S {!zbzx :S bzx, and it follows that bzx = {!zbzx =
z E Z and bbzx = x. Thus, Z is approximating. Furthermore,
(}zY = V{Qzbzx : x <J Y} :S f2z(}zY :S QzY ,
establishing idem potency of (}z and so, by Proposition 7, of bz .
(3) Put 'Y = bZ. Galois-closedness of Z means that z E Z is equivalent
to '"'(Z = bz. If Z is '"'(-approximating then, by (1),
'"'f(}zZ :S bQzZ :S bz = '"'(Z = '"'f (}zZ ,
whence (}zZ E Z. Thus, Z is invariant and (2) applies. We con-
clude that 'Y coincides with bz and preserves arbitrary meets (again
by Proposition 7) . In particular, for W ~ Z , we obtain
1\ 'Y[W] = '"'!(/\ W) :S b(/\ W) :S 1\ b[W] = 1\ 'Y [W]
196 M. Erne
and therefore 1\ W E Z. o
After these remarks about invariant and Galois-closed subsets, we are
now in a position to establish the global version of Proposition 4.
Proposition 8. A subset Y of L contains L 8 and is completely dis-
tributive (hence a sublocale) iff Z = ye is strongly approximating and
satisfies Y = Z 8 . In particular, L 8 is completely distributive iff L is
strongly approximating.
Proof. If Y is of the form Z 8 for a strongly approximating set Z

then, by Proposition 7, Y is a completely distributive sublocale with
y::) L8.
Now assume that, conversely, Y contains L 8 and is completely dis-
tributive, hence a sublocale with nucleus I · We have already re-
marked that Z = ye is Galois closed. For x E X ~ Y n Z , we
get 16zx ::=:; 66zx ::=:; Jx = x and
x ::=:; l\{6z: z E Z,x ::=:; Jz} = 1\{lz : z E Z ,x ::=:; Jz} = 16zx ,
because 1 preserves meets (Lemma 10). Hence, each x E X satisfies
16zx = 66zx = x and 6zx E Z. This together with Lemma 12(3)
shows that Z is strongly approximating. The equation Y = Z 8 follows
from Lemma 10 and Proposition 4. 0
Corollary 7. Z is Galois closed and strongly approximating iffY =

Z 8 is a completely distributive sublocale of L with Z = ye.
There is no exact counterpart of Lemma 7 for approximating sets, be-

cause neither the completely distributive sublocales nor the approxi-
mating Galois-closed subsets of L always do form a complete lattice.
Still worse, Example 3 will demonstrate that in general, there is no
least completely distributive sublocale containing L 8 and no greatest
approximating Galois-closed subset of L.
Summarizing the facts established in Propositions 7, 8 and Corollary
7, we may now formulate the global version of Theorem 1.
Theorem 2. Let L be a completely distributive lattice with join-dense
subset X, and let 6 be a closure operation on L with X ~ L 8 . Then
a subset Y of L contains L 8 and is a (/\-closed) completely distribu-
tive lattice iff there e.'Eists a strongly approximating subset Z of L with
Y = Z 6 . Moreover, the assignment Z f-----+ Z 6 yields a dual order iso-

morphism between the system of all (strongly} approximating Galois-
closed subsets of L and the system of all completely distributive /\-
closed subsets (sublocales) of L containing L 6 .
We conclude this section with the remark that many of the previous
results and their local counterparts admit a common generalization
to the situation of meets formed by fewer than K elements, where K
is a regular infinite cardinal. For example, call a subset Z of L K-
approximating if for each x E X and each subset W of Z with fewer
than K elements, x :::; /\ 6[W] implies x = 6z for some z E Z with
z :::; /\ w, and call a complete lattice K-distributive if vn
y = /\ V[Y]
for all systems Y of less than K downsets. Then, among other facts
generalizing those obtained before, one may show that if a /\-closed
subset Y of a K-distributive complete lattice with closure operation 6 ;::::
/Y is K-distributive, too, then the set Z = Yt? is K-approximating and
satisfies Y = Z 6 ; and the converse holds at least if 6z is idempotent.
3 The Z-below Relation and the Interpolation

Property
As before, L denotes a completely distributive lattice with strong base

X, and Z a subset of L containing X. Following a central idea in the
theory of Z-continuous posets (see the introduction and Section 4) ,
one is tempted to introduce a Z-below relation on L by X<lzY ¢:} x:::;
b"zy or by x <J zY ¢:} x :::; {]zY , but neither of these attempts does
lead to the desired results. It turns out that the "right" definition of
the Z -below relation is the following:
X<lzY ¢:} X<l(]zY
which, by the approximation and interpolation properties of the rela-
tion <J is tantamount to
x <lzY ¢:} x <J b"zx':::; x' <J y for some x' EX.
On the other hand, <lz determines {]z by the equation
f2zY = V{u E L: U<Jzy} .
The join-preservation property of {}z together with the equivalence
198 M. Erne
x <1 VW {::} x <1 w for some w E W

entails the rule
x <1 z VW {::} x <1 z w for some w E W.
By an auxiliary relation (cf. [18]) on a poset P = (X,::::;) , we mean a

relation -< such that
u ::::; v -< w ::::; y ::::} u-< y ::::} u::::; y.
In particular, any auxiliary relation is transitive (but we do not re-
quire, as in [18], that the sets {x : x -< y} be directed). Thus, an
auxiliary relation is idempotent iff it has the interpolation property.
It is easy to see that the relations <lz are always auxiliary and, in
particular, transitive. As we have seen in Section 3, it is important
to have good information about conditions under which the maps c5z
and {!z are idempotent.
Lemma 13. The relation <lz is idempotent (interpolating) iff the map
{!z is idempotent (h ence a dual closure operation). Sufficient for idem-
potency is the equation
c5zx = {!zbzx for all x E X,
and this is also necessary if the elements of X are V-prime.
Proof. Recall that r!zY = V{u: U<lzY} (y E L). Hence, if<lz IS
idempotent then
r!zr!zY = V{u: U<lz V{v: v <lzy}} =
V{u: U<lzV<lzY for some v} = V{u: U<lzY} = r!zY·
Conversely, suppose {!z is idempotent. Then, as {!z preserves joins,
u <1 zY {::} u <1 r!zY = V{ezv: v <1 r!zY}
{::} u <1 {!zV for some v <1 {!zY {::} u <1 zV <1 z Y for some v .
By definition of {!z, the equation c5zx = {!zbzx for all .T E X entails
r!zY = V{ez c5zx : x EX, x <1 Y} ::::; r!zr!z Y ::::; r!zY,

hence idempotency of r!z· On the other hand, if each X E X is v-
prime, i.e. x <1 x, then c5zx = {!zX, and consequently, idem potency of
{!z entails bzx = {!zbzx. o
Invoking Proposition 7 once more, we obtain the remarkable coinci-

dence of three idempotency conditions.
Corollary 8. Suppose Z is approximating. If one of the relations,

respectively, functions <lz, {!z and 6z is idempotent then so are the
others.
This together with Lemma 12 (2) confirms the introductory remark

that often "interpolation follows from approximation".
We prove two technical lemmas that simplify arguments in order to

ensure idempotency.
Lemma 14. Let n be a positive integer and 'Y a closure operation on

L with 'Y :::; 6. Then the following statements are equivalent:
(a) x = "(6zx for all x EX; that is, Z is "(-approximating.
(bn) x = "((2zn6zx for all x EX .
(en) X= "((}znX for all x EX.
Proof. (a)::::} (bn)· Put Xn = {2zn6zx. Then x = "(Xo , and the equations
"fXn+l = "( (}zXn = 'Y (V {6zx' : x' EX, x' <l Xn}) =
'Y(V {"(6zx' : x' E X, x' <l Xn}) =
'Y(V {x' E X : x' <l Xn}) = "(Xn

show (via induction) that x = "fXn for all nEw.
(bn)::::} (en). Since all occurring maps are isotone and {}z is contractive,
we obtain x = "((2zn6zx :S "( {!znx :S "(X :S 6x = x .
(en) :::}(a). x = "f(}znX :S "( (}zX :S "( 6zx :S "(X :S 6x = x. o
Lemma 15. Suppose Z is 6-approximating. Then, for any integer

n ~ 1 and any element x E X, the following are equivalent:
(a) {2z6zx E Z. (a') 6zx = {2z6zx E Z.
(bn) {2zn6zx E Z. (b~) 6zx = {2zn6zx E Z.
(en) {}zn+lx E Z. (c~) 6zx = {!zn+lx E Z.
Proof. (a)::::} (a'). For z := {2z6zx E Z, we have x = 6z by Lemma 14.

Hence, 6zx :S z and z :S 6zx.
(bn)::::}(b~) and (cn)::::}(c~) are shown analogously (using Lemma 14).
200 M. Erne
(a')::::}(c~). By induction, 6zx = f2z6zx yields

6zx = f2zn+16zx :::; f2zn+lx:::; f2z2 X :::; f2zOzx = 6zx.
(c~)::::}(b~). OzX = f2zn f2zX:::; f2zn6zx:::; 6zx.
(b~)::::}(a'). 6zx = f2zn6zx:::; f2z6zx:::; 6zx. o

Applying the last lemma to the case n = 1 and invoking Lemma 13
leads to
Corollary 9. Z is 6-approximating and satisfies f2zf2z E Z for all

x E X iff Z is approximating and f2z6zx = Ozx for all x E X. These
two conditions are equivalent to the strong approximation property if
X consists of V-primes.
Now, combining the results in Section 3 with those in Section 2, we

finally arrive at
Theorem 3. Consider the subsequent conditions:

(a1) Z is 6-approximating and invariant.
( a2) Z is o -approximating and f2 z f2 z x E Z for all x E X.
(b1) Z is approximating with interpolating Z -below relation.
(b2) Z is approximating and f2zf2zX = f2zX for all x EX.
(c1) Z is strongly approximating.
(c2) Z is approximating.
(d1) Z is strongly 6z -approximating.
(d2) Z is Oz -approximating.
(e1) zo is a completely distributive sub locale with nucleus 6z.
(e2) Oz preserves arbitrary meets.
(fl) 6Z preserves arbitrary meets.
(f2) Z is 6Z -approximating.
(g1) 6Z has a left adjoint.
(g2) 6 preserves meets of subsets of Z.
(h1) zo is completely distributive (and a sublocale) .
(h2) Z is o-approximating.
In general, one has the following diagram of implications:
(a1) =} (b1) {::} (cl) =} (d1) {::} (e1) =} (fl) {::} (g1) {::} (h1)
-U- 11 -0- -U- -0- -0- -U- -0-
(a2) {::} (b2) =} (c2) =} (d2) {::} (e2) =} (f2) =} (g2) {::} (h2)
If each x EX is V-prime then the conditions (a2), (b1) , (b2) , (cl) ,

(d1) and (e1) are all equivalent, and so are (c2) , (d2) and (e2) .
If Z is (\-closed, (c2) , (d2) , (e2) , (f2) , (g2) and (h2) are equivalent.
If Z is invariant, all 16 conditions are equivalent.
4 Standard Extensions and Completions

Many effective applications of the previous results are obtained in the
following "standard situation", alluded to already in earlier sections.
Let P = (X,:::;) be an arbitrary poset, .C the completely distribu-
tive frame of all downsets, .6. the cut operator, and X the set of all
principal ideals. Hence, X is contained in the range of .6., called the
M acNeille completion, normal completion or completion by cuts of P
and denoted by N. The condition X ~ Z ~ .C means that Z is a
standard extension. By definition , zb. is the system of all Z-.6.-ideals
(i.e. downsets Y such that .6-Z ~ Y for all Z E Z with Z ~ Y; see the
introduction and [10, 11, 12]). Thus, zb. is always a standard comple-
tion, that is, a closure system of downsets that contains all principal
ideals (cf. [16]). Note that every T0 closure system is the standard
completion of a unique poset. (Take the specialization order given by
x :::; y iff x belongs to the closure of y.) The largest standard comple-
tion is the Alexandroff completion .C, often also denoted by A. It is
not only completely distributive but even superalgebraic: its V-prime
( = supercompact) elements are the principal ideals and form the least
join-dense subset X.
Conversely, if L is a complete lattice whose V-prime elements form a
join-dense subset X then this is the smallest one, and L is isomorphic
to the Alexandroff completion .C of the subposet P = (X, :::;), where
:::; is the order induced from L. In fact , the join map V : .C--+ L is an
isomorphism sending X to X and N to AX. The closure operation
/x associated with X is, up to isomorphism, the cut operator .6. of P ,
202 M. Erne
via the equation lx o V = Vo ~.

A standard extension Z is locally ~-approximating iff for all Z E
Z and x EX n ~Z there is some W ~ Z with x = V W, locally
approximating iff, in addition, W may be chosen from Z, and locally
/\-approximating iff, moreover, one may take W = .}x n Z. Similarly,
Z is ~-approximating iff for all x E X there is a W such that x = V W
and W ~ Z for all Z E Z with x E ~z, and approximating iff, in
addition, W may be taken from Z. In the latter case, W must be the
(modified) Z-below ideal
jJ.z X= n{z E Z: X E ~Z}
which is the Z-core and also the Z-below part of the principal ideal
generated by x . Recall that a strongly approximating standard exten-
sion is one such that ~z is idempotent, i. e., the sets
~zy = U{~Z: ZEZ, Z~ .,j..Y}
are always Z-~-ideals. Thus, a standard extension Z is (strongly)

approximating iff the underlying poset is (strongly) Z -precontinuous in
the sense of [12]. Also, the Z-below relation between downsets extends
the original Z-below relation between elements as defined in [12]. The
crucial role of the Z-below maps in the theory of Z-continuous posets
was already pointed out in [5].
Since X consists of (all) V-prime elements of .C, every locally ~z-ap
proximating standard extension Z is already locally approximating,
and every ~z -approximating standard extension Z is already approx-
imating. This observation makes some of the "standard" applications
easier to handle than the general case discussed in Sections 1-3.
Propositions 3 and 7 have the following consequence for standard ex-
tensions:
Proposition 9. Let Z be a standard extension with idempotent op-

erator ~z. Then Z is locally or strongly approximating iff Z/c;,, the
closure system of Z - ~ - ideals, is a frame or completely distributive,
respectively. In both cases, Z/c;, is a sublocale of the Alexandroff com-
pletion and ~z is the corresponding nucleus.
A downset Y belongs to the Galois closure of Z iff the Z-~-ideal

generated by it is already a cut. Hence, Z is Galois closed iff ~zy =
~y implies Y E Z. The Polarity Theorems 1 and 2 amount to the
following "standard" version (cf. [14] for the first part and [12] for the
second):
Theorem 4. Associating with each locally approximating standard ex-
tension Z of a poset P the collection zt:. of all Z-tl-ideals, one obtains
precisely all standard completions that are frames . Similarly, the com-
pletely distributive standard completions of P are precisely the systems
of Z-tl-ideals coming from strongly approximating standard extensions
Z, In both cases, restriction to Galois-closed extensions makes the re-
spective correspondences bijective.
A slight modification of the tl-ideal construction is often more adapted

to concrete applications, because it works with joins instead of cuts
but does not require any completeness properties of the involved poset.
In the most general situation of a closure operation oon a frame L , the
trick is to restrict the associated polarity relation f2 to the cartesian
product LX J where J = o- 1 [X] . Let us speak of a ]-closed set z if
it is contained in J and for all y E J , it follows from o2 y E X that
y E Z. It is easy to check that a subset has that property iff it is
Galois closed with respect to the restricted relation f2 n(L x J) (i. e.
Z 6 {} n J = Z), and that Proposition 4 and Theorem 1 remain valid for
such ]-closed sets. More precisely:
Proposition 10. Let X be a join-dense subset of a locale L with
X ~ J = o- 1 [X], and let Y, Z be arbitrary subsets of L. Then Y is a
sublocale of L with L 6 ~ Y and Z = Y{} n J iff Z is a J -closed, locally
(A-) approximating subset with Y = zo.
Specifically, for a given poset P , we denote by .:J the standard ex-
tension consisting of all downsets possessing a join (so that in fact
.:J = tl- 1 [X] for the system X of principal ideals) . Then, by defini-
tion, a standard extension Z ~ .:J is (strongly) approximating iff the
underlying poset is (strongly) Z-continuous in the sense of [4], [22]
and [25], and Z is .:J-closed iff each member of Z has a join and Z
contains all lower sets Y for which the Z-join ideal generated by Y
is principal. Thus, we obtain for the closure system z v of all Z-join
ideals the following variant of Theorem 4:
Corollary 10. For every poset there is a dual isomorphism, sending
Z to zv, between the complete lattice of all .:J -closed, locally approx-
imating standard extensions and the complete lattice of all standard
204 M. Erne
completions that are frames. Under this isomorphism, the :J -closed,

strongly approximating standard extensions correspond to the com-
pletely distributive standard completions. The inverse isomorphism
maps standard completions to the collection of all downsets whose clo-
sure is a principal ideal.
Corollary 11. The completely distributive T 0 closure systems are pre-

cisely the Z-join ideal systems of strongly Z-continuous posets.
The definitions of Z-cores and Z-below parts have manifold realiza-

tions in algebra, topology, order theory and computer science. Let us
mention a few examples.
(1) Consider an element x in a Z-inductive poset, as introduced by
Wright, Wagner and Thatcher [26) (see also [11]). The downset gen-
erated by the core o-2 x (consisting of all Z-compact elements below
x) is just our Z-core of the corresponding principal ideal (and gener-
alizes, therefore, the aforementioned construction of Z-below ideals).
Z-inductivity requires that o- 2 x E Z and x = V o- 2 x . Hence, for ev-
ery Z-inductive poset, the standard extension Z A of all Z-downsets
is approximating.
(2) In [6), the (neighborhood} core of a point x in a topological space
(X, T) was defined as the intersection of all neighborhoods containing
x. This is nothing but the 7-core of {x} in the power set P X if we take
as closure operator the identity map on P X. (If we take the closure
operator of the space instead, we get the intersection of all closed
neighborhoods.) It was observed in [6) that complete distributivity of
the topology Tis equivalent to local supercompactness, requiring that
each point has a neighborhood base consisting of cores, and to some
other conditions generalized in Lemma 10 of the present paper.
(3) Even in the seemingly rather special case Z = L, where OLY =
1\ { z E L : y ~ c5 z}, there exist interesting applications. Consider a
closure system Y on a set X and let r be the corresponding closure
operator. Then £ = {U W : W ~ Y} is the Alexandroff topology of
all downsets with respect to the associated specialization quasiorder ~
given by x ~ y {:::} x E r {y}. The £-core of Y E £ is easily computed
by the formula
(Y = n{Z E £ :Y ~ r Z}
where the interior relation ( is defined by

x(y ? y E int(tx) ? y tf_ f{ z EX : x-£ z}.
It was shown in [16] that assigning to each closure operator its in-
terior relation, one obtains a one-to-one correspondence between all
closure operators with a given specialization order :::; that preserve
intersections of downsets, and the approximating, interpolating, aux-
iliary relations for :::; ; moreover, that the corresponding lattices of
open (respectively, closed) sets are precisely the completely distribu-
tive ones (where A is closed iff A = r A and U is open, i.e. X \ U is
closed, iff U = U() . Using those results from [16], we obtain:
Theorem 5. For any poset P = (X,:::;) , the following objects form
four mutually isomorphic posets:
(1) The completely distributive standard completions of P.
(2) The closure operators having specialization order:::; and preserv-
ing arbitrary meets of downsets.
(3) The approximating and interpolating auxiliar·y relations on P .
(4) The approximating Galois-closed standard extensions, ordered by
dual inclusion.
It may happen that these posets are neither isomorphic nor dually
isomorphic to the corresponding posets of standard completions etc.
for the dual of P. For example, the chain w of natural numbers has
exactly two completely distributive standard completions, whereas its
dual has only one. The example below shows that descending chains in
the above four isomorphic systems need not possess any lower bounds
(so that completeness is violated).
Example 3. The "deleted real unit square"
Q = ]0, 1]2 U { (0, 0)} = {(x1, x 2) : 0 <xi :::; 1 or xi = x2 = 0},
ordered componentwise by the usual :::;, is a frame but not a dual
frame, hence not completely distributive. The relations (n on Q de-
fined by
(xi,x2)(n(YI,Y2) ? XI :::; Yh X2:::; Y2 and XIX2:::; ~
are easily seen to be auxiliary, approximating and idempotent, but
their intersection is { (0, 0)} x Q, which certainly cannot contain any
approximating relation (though being auxiliary and idempotent).
206 M. Erne
Now, let us have a look at a few specific standard extensions (see

[9, 10, 11, 12] for more background) .
(1) The smallest standard extension X , consisting of all principal ide-
als (X-below ideals), is always strongly approximating and opposite to
the largest standard completion £ = X b. , which is always completely
distributive.
(2) On the other hand, for the largest standard completion£ and the
smallest standard completion N =£b. , we have (cf. [8]) :
Corollary 12. The Alexandroff completion£ of a poset is locally

(/\-)approximating iff the M acNeille completion N is a frame, and £
is (strongly) approximating iff N is completely distributive.
(3) Consider the standard extension V of all directed downsets of a

dcpo P . Here, the V-core, respectively, V-below part of .,t. x is the
way-below ideal of x. Invariance of V postulates that for any directed
downset D, the set ~ D be again directed. But this is certainly the
case in any continuous poset, because ~ D is a directed union of
directed downsets, namely the way-below ideals of the elements in D .
Now, the results in Section 3 confirm that the way-below relation of a
continuous poset has the interpolation property [18]. As remarked in
the introduction, V is (strongly) approximating, i.e. P is a continuous
poset, iff the Scott completion vv is completely distributive.
Similar arguments apply to any union complete subset system Z [26] ,
showing that for such systems the Z-below relation of a Z-continuous
poset is idempotent and the system of Z-join ideals is completely
distributive (see [4, 5, 12, 22]) .
For a more comprehensive discussion of local approximation properties
of the standard extension V , we refer to [13] and [14] . Here we only
mention the fact that the local approximation property of V entails
the frame property for the Scott completion, but not conversely!
(4) For the standard extension of all finitely generated downsets,
F= {.,t..F: F ~wX} ,
the associated preclosure operator f:j.F is idempotent, hence a closure
operator. It assigns to each subset Y of X the Frink ideal generated
by Y,
6;:y = U{6F: F ~wY}.

Thus, the corresponding closure system
Ft:. = {Y ~ X : 6F ~ Y for all F ~w Y}
consists of all ideals in the sense of Frink (17]. In that context, the
results of Section 1 reduce to known characterizations of posets with
distributive ideal lattices (8]:
Corollary 13. For any poset P = (X, ::;) , the following conditions
are equivalent:
(a) The lattice Ft:. of all Frink ideals is distributive (a frame) .
(b) Ft:. is a sub locale of£, the frame of all downsets.
(c) The cut operator 6 preserves finite intersections in Ft:..
(d) The ideal operator 6F preserves finite intersections in .C.
(e) 6F induces a (pre)nucleus on .C.
(f) For F ~wX and x E 6F, there is an E ~w.J-.F with x = VE .
(g) F is locally approximating.
A V-semilattice P with these properties is called distributive (see,
for example, [19]). In (8], an example of a V-semilattice has been
given whose ideal lattice fails to be distributive, although the sys-
tem of all downsets (not only the finitely generated ones) is locally
6-approximating. Of course, that failure cannot happen with 1\-
semilattices, because then the finitely generated downsets form a 1\-
semilattice, too, and Corollary 1 applies. For the "global" situation,
we get:
Corollary 14. The lattice Ft:. of all Frink ideals of a poset P is

completely distributive iff F is (strongly) approximating, that is, for
each x E P there is a least F E F with x E 6F. In case P is a V-
semilattice, the latter condition holds iff each element is a finite join
of V -prime elements, in other words, P is a free semilattice over a
poset.
These free semilattices are, up to isomorphism, just the standard ex-

tensions F of their subposets of V-prime elements, and conversely, the
standard extension of all finitely generated downsets of an arbitrary
208 M . Erne
poset is "the" free V-semilattice over that poset. Our final exam-
ple shows that the standard extension :F of a V-semilattice may be
(strongly) approximating without being locally /\-approximating.
Example 4. The "double real square"
C = [0, 1] X [0, 1] X { 0, 1}
is completely distributive. For any number r with 0 < r < 1, the
subset
S = C \ (({r} X [0, 1] X {0}) U ([0, 1] X {r} X {0}))
is a V-subsemilattice whose elements are joins of finitely many V-prime
elements. For x = (1, 1, 0) , y = (r, 1, 1) , z = (1 , r, 1) and F ={.{y,z},
we obtain x ~ V F, but {.x n F is not finitely generated.
References
[1] Banaschewski, B., Another look at the localic Tychonoff theorem,
Comm . Math. Univ. Carolinae 29 (1988) , 647-656.
[2] Banaschewski, B. and Hoffmann, R.-E. (eds.) , Continuous Lattices.

Bremen 1979, Lecture Notes in Math. 871 , Springer-Verlag, Berlin,
Heidelberg, New York, 1981.
[3] Birkhoff, G., Lattice Theory. Amer. Math. Soc. Coll. Publ. 25 , 3rd ed.,
Providence, R.I., 1973
[4] Bandelt, H.-J. and Erne, M., The category of Z-continuous posets, J.
Pure Appl. Algebra 30 (1983) , 219-226.
[5] Bandelt, H.-J. and Erne, M. , Representations and embeddings of M-

distributive lattices, Houston J. Math. 10 (1984), 315-324.
[6] Erne, M., Scott convergence and Scott topologies on partially ordered
sets II, in: [2], pp. 61-96.
[7] Erne, M., Homomorphisms of M-distributive and M-generated

posets, Preprint No. 125, lnstitut fur Mathematik, Universitat Han-
nover, 1981.
[8] Erne, M., Distributivgesetze und die Dedekind'sche Schnittver-

vollstandigung, Abh. Braunschweig. Wiss. Ges. 33 (1982) , 117-145.
[9] Erne, M., Adjunctions and standard constructions for partially ordered
sets, in: Eigenthaler, G. et al. (eds.) , Contributions to General Algebra
2, Proc. Klagenfurt Conf. 1982, Holder-Pichler-Tempsky, Wien, 1983,
pp. 77-106.
[10] Erne, M., Order extensions as adjoint functors , Quaestiones Math. 9

(1986), 146-204.
[11] Erne, M., Algebraic ordered sets and their generalizations, in: Rosen-
berg, I. and Sabidussi, G. {eds.), Algebras and Orders, Proc. Montreal
1992, Kluwer, Amsterdam, 1994.
[12] Erne, M., Z-continuous posets and their topological manifestation,
Appl. Cat. Structures 7 (1999) , 31-70.
[13] Erne, M., Continuity properties and Scott frames. Preprint , University
of Hannover, 2000
[14] Erne, M., Pultr, A., Sichler, E. , Closure frames and web spaces.
Preprint, Hannover-Prag, 2000
[15] Erne, M., Koslowski, J. , Melton, A., Strecker, G. , A primer on Galois

connections. In: S. Andima et al. (eds.), Papers on General Topology
and its Applications. 7th Summer Conf. Wisconsin. Annals New York
Acad. Sci. 704, New York, 1993, 103-125.
[16] Erne, M. and Wilke, G., Standard completions for quasiordered sets,
Semigroup Forum 27 (1983) , 351-376.
[17] Frink, 0., Ideals in partially ordered sets. Amer. Math. Monthly 61
(1954), 223-234.
[18] Gierz, G., Hofmann, K. H ., Keimel, K ., Lawson, J. D. , Mislove, M. ,

Scott, D. S., A Compendium of Continuous Lattices. Springer-Verlag,
Berlin- Heidelberg - New York, 1980
[19] Gratzer, G., General Lattice Theory, Birkhauser Verlag, Basel, 1978
[20] Johnstone, P. T. , Stone Spaces. Cambridge Studies in Advanced

Math., No. 3, Cambridge University Press, Cambridge, 1982
[21] Lawson, J. D., The duality of continuous posets, Houston J. Math. 5

(1979), 357-386.
[22] Novak, D., Generalization of continuous posets, Trans. Amer. Math.

Soc. 272 (1982), 645-667.
210 M. Erne
[23] Raney, G. N., A subdirect-union representation for completely dis-

tributive complete lattices, Proc. Amer. Math. Soc. 4 (1953), 518-522.
[24] Scott, D. S., Continuous lattices. In: Toposes, Algebraic Geometry and
Logic, Lecture Notes in Math. 274, Springer-Verlag, Berlin - Heidel-
berg - New York, 1972.
[25] Venugopalan, G. , Z-continuous posets, Houston J. Math. 12 (1986),

275-294.
[26] Wright, J. B., Wagner, E. G., and Thatcher, J. W., A uniform ap-
proach to inductive posets and inductive closure, Theor. Comp. Sci. 7
(1978), 57-77.
Author's address:
Marcel Erne
University of Hannover
D-30167 Hannover
Germany
e-mail: erne@math.uni-hannover.de
Galois Connections and Complete
Sublattices
K. Denecke, 5. L. Wismath*
Abstract
Any Galois connection (between two sets) has associated with it two clo-
sure operators, and hence two complete lattices (dually isomorphic to each
other). Set-theoretical Galois connections are induced by relations, and
it is interesting to know which subrelations of an inducing relation will in
turn induce complete sublattices of the two lattices. Complete sublattices
of the Galois-lattices may also be obtained using conjugate pairs of closure
operators. It is also well known that the set of fixed points of a closure (or
a kernel) operator defined on a complete lattice forms a complete lattice,
and in this case too we can look for properties which ensure that this lattice
is a complete sublattice. This paper surveys the results known about such
subrelations, conjugate pairs of closure operators, and closure and kernel
operators. We also give an application of the closure and kernel opera-
tor method, which generalizes several well-known examples from universal
algebra using the Galois connection (Id,Mod) .
AMS Classification: 06B23, 06A15.
Key Words: Galois connection, Closure operator, Kernel operator, Com-
plete sublattice.
1 Introduction and Notation

A Galois connection is a pair of mappings with certain properties, be-
tween two sets of objects, usually of different kinds. Such a connection
can provide a useful tool for studying properties of one kind of object,
based on the properties of the other (usually more well-known) kind
*Research of the second author supported by NSERC of Canada.
211
K. Denecke et al. (eds.), Galois Connections and Applications, 211-229.
212 K . Denecke, S. L. Wismath
of objects. In the classical Galois theory, for instance, properties of

permutation groups are used to study field extensions. In universal
algebra, the Galois connection (I d, Mod) between classes of algebras
and sets of identities allows us to study algebraic structures either
algebraically, in terms of the usual operators H , S and P , or equa-
tionally, in terms of the identities satisfied by a given algebra or class
of algebras. The Galois connection (Pol , Inv) between sets of oper-
ations and sets of relations, on a fixed base set, is also important in
universal algebra, and combines the structure-theoretic point of view
with the logical approach.
For any Galois connection between sets A and B , the corresponding
collections of closed sets form complete lattices, which are dually iso-
morphic to each other. We remark that the propositions formulated
here for sets are also valid for proper classes. For the two examples
mentioned above, we have the lattices of all varieties and of all equa-
tional theories of a fixed type, and the lattices of all clones and of all
relational clones on a fixed base set. Since such lattices are usually
very large and complex, it is useful to find ways to study pieces of
them, particularly their complete sublattices. In this paper we give
a survey of results on three such methods for constructing complete
sublattices of the lattice of closed sets obtained from a Galois connec-
tion: conjugate pairs of closure operators, Galois-closed subrelations,
and closure and kernel operators with particular properties. We ap-
ply these results to produce a general method for finding complete
sublattices, and illustrate it with some examples for the connection
(I d, Mod) . This method encompasses several well-known examples
involving normal and regular identities, and also provides a frame-
work for the introduction of some new examples.
In the remainder of this section we fix the main definitions and nota-
tion to be used throughout. We let A and B be fixed non-empty sets.
A Galois connection between A and B is a pair (t , f1) of mappings
between the power sets P(A) and P(B) ,
t: P(A) ----t P(B) and f1: P(B) ----t P(A) ,
such that for all T, T' ~ A and all S , S' ~ B the following conditions
are satisfied:
(i) T ~ T' =? t(T) 2 t(T') , and S ~ S' =? f.l(S) 2 f.l(S');

Galois Connections and Complete Sublattices 213
(ii) T ~ w(T), and S ~ ip(S).

That is, the two mappings in a Galois connection reverse inclusions,
(map larger sets to smaller images), and applying the composite of the
two mappings to a given set produces a set which contains the original
one. It is well known that for any Galois connection (i, p) between
A and B, the composite mappings w and if-1 are closure operators
on A and B respectively, meaning that they are extensive, monotonic
and idempotent mappings. The sets closed under pi are precisely the
sets of the form p(S) , for some S ~ B; the sets closed under if-1 are
precisely the sets of the form i(T), for some T ~ A.
In general, for any closure operator r defined on a set A the set 1i'Y of
allr-closed subsets of A forms a complete lattice. In this lattice, the
meet operation, also the greatest lower bound or infimum with respect
to the partial order of set inclusion, is the operation of intersection.
The join operation however is not usually just the union: we have
vs = n{H E 1i'Y , H ;;;:> uB}

for every B ~ 1i'Y .
In the special case of a Galois connection between sets A and B, we
obtain two closure operators pi and if-1, and thus two complete lattices,
1iJ.LL and 1iLJ.L' of closed sets. These two lattices are dually isomorphic
to each other.
A relation between the sets A and B is simply a subset of Ax B . Any
relation R between A and B induces a Galois connection, as follows.
We can define t he mappings
i : P(A) --t P(B) , p: P(B) --t P(A) ,
by
i(T) : {s E E I VtET ((t,s) E R)},
p(S) : {tEA I VsES ((t, s)ER)} .
Then it is easy to verify that the pair (i, f.1) forms a Galois connec-
tion between A and B, called the Galois connection induced by R.
Conversely, any Galois connection between A and B determines a re-
lation R ~ A x B. Thus there is a one-to-one correspondence between
relations from A to B and Galois connections between A and B.
214 K. Denecke, S. L. Wismath
For example, the Galois connection (I d, Mod) between classes of alge-

bras and sets of identities corresponds to the relation R of satisfaction:
for any algebra A and any identity u ~ v, the pair (A, u ~ v) is in
R iff the algebra A satisfies the identity u ~ v. Similarly, the Galois
connection (Pol, Inv) is based on the relation of preservation (of a
relation by an operation).
2 Conjugate Pairs of Closure Operators
Given a relation R from A to B, and the corresponding Galois con-

nection (i, tJ) between A and B, we obtain the complete lattices 1-lJJ.~
and 1-l~p, of closed subsets of A and B respectively. Our goal in this
section is to construct, from the given relation R and the induced Ga-
lois connection, a new relation and Galois connection, which produce
complete sublattices of 1-lJJ.~ and 1-l~w The tool we use is the concept
of a conjugate pair of closure operators.
Definition 2.1. A closure operator 1 defined on a set A is said to be

additive if for all T ~ A, 1 (T) = U 1 (a). (Note that we write 1 (a)
a ET
for 1( {a})).
We can show easily that when 1 is an additive closure operator, the

least upper bound operation on the lattice 1-l'Y agrees with U B (see
[Rei] or [DG]). We always have U B ~ 1 (U B) because of the ex-
tensivity of I · Conversely, if a E U B then a E B E B for some
set B E B, and since B E 1-£7 we have 1 (a) ~ U B and 1 (U B) =
U 1(a) ~ U U B = U B. This means that U B is 1-closed and
aEUB aEUB
VB = U B. In other words, when 1 is an additive closure operator
on A, the corresponding closure system forms a complete sublattice of
the lattice (P(A); 1\ = n,v = U) of all subsets of A.
(That the additive closure operators are in one-to-one correspondence
with union-closed closure systems (Alexandroff topologies) is well-
known. In the thirties of the past century Alexandroff dicovered
a one-to-one correspondence between additive closure operators and
quasi-orders.)
Definition 2.2. Let 1 1 be a closure operator defined on the set A and

let 1 2 be a closure operator defined on the set B. Let R ~ A x B
be a relation between A and B . Then 'Yl and ')'2 are called conjugate
with respect to R , and we say that the pair ('Yl , ')'2 ) is a conjugate pair
with respect toR, if for all tEA and all s E B, 'Y1 (t) x {s} ~Riff
{t} x 'Y2 (s) ~ R.
This property of conjugacy of two closure operators is defined in terms
of individual elements. When the two operators are also additive, we
can extend this to sets of elements. Thus when ('Yl, 'Y2 ) is a pair of
additive closure operators, 'Yl on A and ')'2 on B , and they are conju-
gate with respect to a relation R ~ Ax B , then for all T ~ A and all
S ~ B we have T x 'Y2 (S) ~ R if and only if 'Y1 (T) x S ~ R .
The two closure operators w and Lf..L , obtained from a Galois connec-
tion induced by a relation R , are always conjugate with respect toR.
But f..LL and Lf..L need not be additive in general.
Definition 2.3. {{DWB}) Let 'Y := {1'1 , 'Y2 ) be a conjugate pair of ad-
ditive closure operators, with respect to a relation R ~ A x B . Let Ry
be the following relation between A and B:
Ry := {(t ,s) E Ax B I 'Y1 (t) x {s} ~ R} .

We now have two relations and Galois connections between A and B.
We have the original relation R, with induced Galois connection (L, J.L)
between A and B , and corresponding lattices of closed sets. We also
have the new relation R7 and its induced Galois connection, which we
shall denote by (L7 , J.L 7 ) . The following theorem gives some properties
relating the two Galois connections.
Theorem 2.4. ([DWB}) Let 'Y = {1'1 , 'Y2 ) be a conjugate pair of addi-
tive closure operators with respect to R ~ A x B . Then for all T ~ A
and S ~ B , the following properties hold:
{i) J.L7 (T) = J..Lb1 (T)) ,
{ii) f-L'Y (T) ~ J.L(T),
{iii) 'Y2(J..L'Y (T)) = J.L7 (T) ,
{iv) ')'1(L(J..L7 (T))) = L(J.L7 (T)) ,

and dually,
{i ') ~,(S) = ~(r2(S)),
{ii 'J ~,(s) ~ ~(s),
{iii') 'Yl(~ 1 (S)) = ~,(S),
{iv ') 'Y2 (M(~ 1 (S))) = M(~ 1 (S)),
( v I) ~, (J-l, (T)) = ~(J-l(rl (T))) .
The next theorem is the "Main Theorem for Conjugate Pairs of Clo-
sure Operators", from [Rei]. It shows that when we consider sets
which are closed under the original Galois connection from R, there
are four equivalent conditions for such sets to also be closed under the
new connection from R1 .
Theorem 2.5. {Main Theorem for Conjugate Pairs of Additive Clo-

sure Operators) Let R be a relation between sets A and B, with cor-
responding Galois connection (J-l, ~). Let 'Y = (11 , ')'2 ) be a conjugate
pair of additive closure operators with respect to the relation R. Then
for all sets T ~ A with ~(J-l(T)) = T the following propositions {i) -
{iv) are equivalent; and dually, for all sets S ~ B with J-l(~(S)) = S ,
propositions (i') - (iv) are equivalent:
{i) T = ~,(M 1 (T)),
{ii) ')'1 (T) = T,
{iii) J-l(T) = J-l,(T),
{iv) 'Y2(J-l(T)) = J-l(T); and dually,
{i ') S = J-l,(~ 1 (S)),
{ii ') 'Y2 (S) = S,
{iii') ~(S) = ~,(S),
{iv ') 'Yl(~(S)) = ~(s).

Proof: We prove the equivalence of (i) ,(ii) ,(iii) and (iv); the equiva-
lence of the four dual statements can be proved dually.
(i) ::::} (ii) We always have T ~ --y1 (T) , since --y1 is a closure opera-
tor. Since iJ..t is a closure operator we also have --y1 (T) ~ iJ..t( --y1 (T)) =
i-y(J..t-y(T)) = T, by 2.4 (v') and (i).
(ii) ::::} (iii) We have ~-t(T) = ~-tb1 (T)) = /-t-y (T) by (ii) and 2.4 (i) .
(iii) ::::} (iv) We have --y2 (~-t(T)) = --y2 (~-t-y (T)) = /-t-y (T), using 2.4 (iv)
and (iii).
(iv) ::::} (i) Since the i-y /-t-y-closed sets are exactly the sets of the form
i-y(S), we have to find a set S ~ B with T = i-y (S). But we have
i-y(~-t(T)) = i('"Y2 (~-t(T))) = t(~-t(T)) = T , by 2.4 (i') and our assumption
that T is ij..t-closed. •
Before we use this Main Theorem to produce our complete sublattices,

we need the following additional properties.
Theorem 2.6. Let R be a relation between sets A and B , with Galois

connection (J..t, i) . Let --y = (--y1 , --y2 ) be a conjugate pair of additive
closure operators with respect to R . Then for all sets T ~ A and
S ~ B, the following properties hold:
(i) 1'1 (T) ~ i(J..t(T)) {:} t(~-t(T)) = L-y (M-y (T));
(ii) --y1 (T) ~ t(~-t(T)) {:} 1'1 ( i(J..t(T))) = i(J..t(T)) ;
(i ') --y2 (S) ~ ~-t(t(S)) {:} J..t(i(S)) = /-t-y (t-y (S));
(ii ') --y2(S) ~ ~-t(t(S)) {:} 1'2(~-t(t(S))) = ~-t(t(S)) .
Proof: We prove only (i') and (ii') ; the others are dual.
(i') Suppose that --y2( S) ~ ~-t( t( S)). Since J..tL is a closure operator
we have J..t(t(S)) = J..t(t(J..t(t(S)))) 2 J..t(i('"Y2(S))) = J..t -y (t-y (S)) , by our
assumption and by 2.4 (v). Also S ~ --y2(S) , and hence we have
J..t(i(S)) ~ J..t(i('"Y2(S))) = /-t-y (i-y (S)) , again by 2.4 (v). For the converse
we have 1'2(5) ~ J1(~,(!'2 (S))) = /1-y (~,-y (S)) = J1(~,(S)) , using the exten-
sivity of f1L,, 2.4 (v) and our assumption.
(ii') Let ')'2(5) ~ (!1(~,(5)). Then S ~ ')'2(S) implies that ')'2 (J1(~,(S)))
~ ')'2 (!1(~,(1'2 (5)))). We also have ')'2 (J1(~, (!'2(S)))) = /'2(J1(~,-y (S))) by
Theorem 2.4 (i'), and ')'2 (J1(~,(!'2 (S)))) = J1(~,-y (S)) by 2.4 (iv') . In ad-
dition, J1(~,-y(5)) = J1(~,(!'2 (S))) ~ J1(L,(J1(~,(S)))) = J1(~,(S)). Altogether
we obtain J'2 (J1(~,(5))) ~ 11(~,(5)). The opposite inclusion is always
true, since ')'2 is a closure operator. Conversely, 5 ~ J1(~,(5)) implies
')'2 (5) ~ ')' 2 (!1(~,(5))) = 11(~,(5)) , by the extensivity of f1L, , the mono-
tonicity of ')'2 and our assumption. •
Now we have the machinery to produce our complete sublattices. From

the original relation R and Galois connection (J1 , ~,) we have two (du-
ally isomorphic) complete lattices of closed sets, the lattices 1iJ.LL and
1iw We also have two complete lattices of closed sets from the new
Galois connection (~,"1 , /1-y ) induced by R-y . Our result is that each new
complete lattice is in fact a complete sublattice of the corresponding
original complete lattice.
Theorem 2. 7. Let R be a relation from A to B , with induced Galois

connection (J1, ~,). Let I' = (1'1 , ')'2 ) be a conjugate pair of additive
closure operators with respect to R . Then the lattice 1iJ.L-y L-y of sets
closed under J11 L-y is a complete sublattice of the lattice 1iJ.Lu and dually
the lattice 1iL-y J.L-y is a complete sublattice of the lattice 1iw
The proof can be found in [Rei] .

Conjugate pairs of additive closure operators are an important way to
construct complete sublattices of given closure lattices. We may also
consider the set of all such conjugate pairs, and define on it an order
relation, as follows: for a = (1'1 , I'D and {3 = (1'2 , I'~) we set
a:::; {3 :{:} (VT ~ A)(V5 ~ B)[!'1 (T) ~ !'2 (T) and 1'~(5) ~ I'~(S)].
When a :::; {3, it can be shown that the lattice 1iJ.Lf3 Lf3 is a sublattice of
1iJ.La La , and dually that 1iLf3 Jl.f3 is a sublattice of 1iLa J.La. The following
additional properties may also be verified:
(i) J'I(~,(J1(T))) = ~,(J1(T)) <¢:::::::} T = ~,(J1 (!' 1 (T))) , and

The theory of conjugate pairs of additive closure operators has been

generalized in [Arw] to the situation of conjugate pairs of extensive,
additive operators.
3 Galois-Closed Subrelations
In the previous section we used a pair of additive closure operators,
which is conjugate with respect to the relation R, to determine a new
relation R1 , which in turn induces a new Galois connection and closure
operators. We showed that the sets closed under these new operators
form complete sublattices of the original lattices of closed sets. Now
we want to focus on the relations such as R, which determine com-
plete sublattices of our original lattices. In general, we can look at
subrelations R' ~ R, and consider the complete lattices of closed sets
obtained from the Galois connection induced by R'. When are these
lattices complete sublattices of the lattices obtained via R? In this
section we describe a property of subrelations R' of R which is suf-
ficient to guarantee that the new complete lattices obtained from R'
will be complete sublattices of the original lattices, and show that any
complete sublattices of our original lattices arise in this way. The
property we need is called the Galois-closed subrelation property.
Definition 3.1. Let R and R' be relations between sets A and B,
and let (p,, i) and (p,' , i1 ) be the Galois connections between A and B
induced by R and R', respectively. The relation R' is called a Galois-
closed subrelation of R if:
1) R' ~ R, and
2) for all T ~ A and all S ~ B , we have (p,'(T) = S and i'(S) =

T =? p,(T) = S and i(S) = T ).
The following equivalent characterizations of Galois-closed subrela-
tions were given in [GW], [ADP] and [Den].
Proposition 3.2. Let R' ~ R be relations between sets A and B .

Then the following are equivalent:
(i) R' is a Galois-closed subrelation of R;
(ii) For any T ~A, if t'p/(T) = T then p,(T) = p,'(T) , and for any
S ~ B, if p,'t'(S) = S then t(S) = t'(S);
(iii) For all T ~ A and for all S ~ B the equations t' p,' (T) = L p,' (T)
and p,' t' ( S) = p, t' ( S) are satisfied.
The proof is straightforward and will be skipped here.

It is also straightforward to show that if "( := (11 , "(2 ) is a pair of ad-
ditive closure operators which are conjugate with respect to a relation
R ~ A x B, then the relation Ry of Definition 2.3 is a Galois-closed
subrelation of R .
The significance of Galois-closed subrelations lies in the fact that there

is a one-to-one correspondence between Galois-closed subrelations of
a relation R ~ A x B and complete sublattices of the corresponding
lattices 1-lqt and 1-l 1u of closed sets. The remainder of this section is
devoted to a scetch of the proof of this claim. Specifically, we will
show that any Galois-closed subrelation R' of the relation R yields
a lattice of closed subsets of A which is a complete sublattice of the
corresponding lattice 1-l~tL for R , and conversely that any complete
sublattice of the lattice 1-l~tL occurs as the lattice of closed sets induced
from some Galois-closed subrelation of R. Dual results of course hold
for the set B.
Theorem 3.3. ([GW}, [Den}, and [ADP}) Let R ~ Ax B be a rela-

tion between sets A and B, with induced Galois connection (p,, t). Let
1-l~ tL be the corresponding lattice of closed subsets of A.
(i) If R' ~ A x B is a Galois-closed subrelation of R, then the class

UR' := 1-l~'~t' is a complete sublattice of 1-l~w
(ii) If U is a complete sublattice of 1-l~IL ' then the relation
Ru := U {T x p,(T) I T E U}
is a Galois-closed subr·elation of R.
(iii) For any Galois-closed subrelation R' of R and any complete sub-
lattice U of 1lq. , we have
URu = U and Run,= R'o
Scetch of the Proof: (i) We begin by verifying that any subset of

A which is closed under the operator ~, 1.1/ is also closed under ~p, , so
that the lattice 1{~, J1.' is at least a subset of 1lq.~,o Let T E 1{~, Jl/, so that
~' p,' (T) = T 0 By Proposition 302, parts (ii) and (iii), we have
Therefore, 1{~, Jl/ ~ 1{ ~ w

Next we have to show that this subset is in fact a sublatticeo This
means showing that for any family {Tj I j E J} of sets in 1{~, J1.' , both
the sets 1\ {Tj I j E J} and V {Ti I j E J} are in 1{~, JL' 0
1-l,l' 1-l, I'

We know from part (i) that the collection {Tj I j E J} is also a family
of sets in 1{~ w Since
it is enough to show that
By repeated use of Lemma 1.1 and Proposition 302, we check that
~'p,'~p,(u Tj) = ~p,(u Tj)o

jEJ jEJ
This shows that ~ p,( U Tj ) is also a fixed point under ~' p,' , so that it
jEJ
t oo is an element of 1{~, JL' 0
(ii) Now let U be any complete sublattice of 1l~w We will prove that
the relation Ru is a Galois-closed subrelation of R o First, for each
non-empty T E U we have p,(T) = {s E B I Vt E T, (t , s) E R} and
then T x p,(T) ~ Ro Therefore Ru ~ R o To show that the second
condition of the definition of a Galois-closed subrelation is met, we
let (p,' , ~') be the Galois connection between sets A and B induced by
Ru, and assume that tl(T) = S and t'(S) = T for some T <;;;; A and
S <;;;; B. Our goal is to prove that
p,(T)=S and t(S)=T. (*)

For each t E T, we define
Dt = n {T' E u It E T' and s <;;;; p,(T')} .

The following facts can be verified:
(a) p,'( {t}) = p,'(Dt)·
(b) t'p,'({t}) = Dt.

(c) T= U Dt.
tET
(d) p,(T) = p,'(T) =Sand t(S) = t'(S) = T , and (*) holds.

From the fact that for any set T E U , if p,(T) = S , then t(S) = t'(S) ,
we have t( S) = t' ( S) = t' p,' (T) = T. This shows that (*) holds, com-
pleting the proof of (ii) that Ru is a Galois-closed subrelation of R.
(iii) Now we must show that for any complete sublattice U of1lqj,, and
any Galois-closed subrelation R' of R , we have URu = U and Run, = R'.
We know that URu := 1£L, Jl/ , the lattice of subsets of A closed under
the closure operator t' p,' induced from the relation Ru. This means
that T E URu iff t' p,' (T) = T. First let T E URu, and let S be the
set p,'(T). Then we have t'(S) = T , and since Ru is a Galois-closed
subrelation of R we conclude that
p,(T) = S and t(S) = T .
If T = 0 then T E U , and forT -1- 0 we can show that T = U Dt (see

tET
(c) above). But now T = tp,(T) = tp,(U Dt) = U{Dt It E T} E U .
t ET
This shows one direction, that URu <;;;; U.
For the opposite inclusion, let T E U. Then using the fact that U is a
sublattice of 1lLM' we have
t' p,' (T) = t' p,(T) = t p,(T) = T.

This shows that T E 1£,,~ 11,, which is equal to URu. We now have the
required equality URu = U .
Now let R' be a Galois-closed subrelation of R, and set
UR' ·- HL'f.L' = {T ~A I /p/(T) = T} , and

RuR, ·- U{T X p,(T) IT E UR' }.
We will show that RuR' = R'.
First, if (t,s) E R' then s E p,'({t}). SettingS:= p,'({t}), we have
s E Sand i'(S) = i1 p,'({t}). Now taking T := i1 p,'({t}), we have
i'p,'(T) = T, soT E UR' and p,'(T) = Sand i'(S) = T. Therefore
p,(T) = S and i(S) = T.
Since t E i'p,'({t}) = T and s E S = p,(T) , we get (t,s) E T x p,(T)
and T E UR'· Hence (t, s) E RuR, and we have shown that R' ~ RuR,.
To show the opposite inclusion, letT E UR', and letS= p,(T). Then
from Facts 1 and 2 we have
p,' (T) = p,(T) = S and i1 ( S) = i( S) = i p,(T) = T.

Therefore T x p,(T) ~ R' , and RuR' ~ R'. Altogether, we have RuR, =
R'. This completes the proof of part (iii) , and of our theorem. •
4 Closure and Kernel Operators
It is well known that there is a 1-1-correspondence between closure

operators 'P : .C --+ .C on a complete lattice .C and closure systems S
on .C (that is, subsets of .C closed under arbitrary meets). This cor-
respondence is given by the following maps. For any closure operator
'P : .C --+ .C, we get the closure system
S := Fix( 'P) := H rp = {T E .C I 'P(T) = T}
of all fixed points of (/); and for any closure system S we have the
closure operator 'P defined by
t:
'P(T) = 'Ps(T) := 1\{T' E S IT' 2 T} forT E .C.
Moreover, for any closure system S and any closure operator cp, we
have 1icps = S and cp1i<p = cp.
There is also a dual 1-1-correspondence between kernel operators and
kernel systems on a complete lattice £. A kernel operator is a mapping
which is intensive, monotonic and idempotent, and a kernel system on
£ is a family S of subsets of£ which is closed under arbitrary joins.
Then for kernel operators 'ljJ : £ ---* £ on £ and kernel systems S on
£,we set
S ·- Fix( 'ljJ) := {T E £ I '¢ (T) = T} ;

[_
'l/J(T) ·- 'l/Js(T) := V{T' E SIT'::::; T} forTE£.

We have Fix('l/Js) = Sand '¢Fix('I/J)= '¢, for any kernel systemS and
any kernel operator 'ljJ on £.
A result of A. Tarski ([Tar]) shows that for any closure operator cp on

a complete lattice£, the closure system (fixed-point set) 1icp is always
a complete lattice with respect to ::::;. However, it is not necessarily
a sublattice of £. Thus we look for some additional condition under
which a complete sublattice is obtained. The answer (and its dual for
kernel operators) is given in the following well-known theorem ([Tar]).
Theorem 4.1. Let £ be a complete lattice.
(i) If cp is a closure operator on £ which satisfies

[_ [_
cp(v{rj u E J}) = v{cp(Tj) u E J} (*)
for every index set J, then the set of all fixed points under cp,
1icp = {T E £ I cp (T) = T}
is a complete sublattice of£ and cp(£) = 1icp ·
(ii) Conversely, if 1i is a complete sublattice of£, then the function

cp1i which is defined by
[_
cpH(T) := / \ {T' E 1i IT::::; T'}

is a closure operator on .C with cp1l(.C) = 1t, and <{)1£ satisfies the con-
dition (*). Moreover, 1tcp1i = 1t and <{)1£'{' = cp.
(iii) If '1/J is a kernel operator on .C which satisfies

£ £
'1/J(/\{Tj I j E J}) = f\{ '1/J (Tj) I j E J} (**)
for every index set J, then the set of all fixed points under '1/J,
1{1/J = {T E .c I '1/J (T) = T}

is a complete sublattice of .C and '1/J (.C) = 1t1/J.
(iv) Conversely, if 1t is a complete sublattice of .C then the function

'l/J1i which is defined by
'l/J1l(T) := v
£
{T' E 1t IT' :s; T}
is a kernel operator on .C with 'l/J1i (.C) = 1t1/J, and 'ljJ satisfies the con-
dition (**). Moreover, 1t1/J1i = 1t and 'l/J1i..p = '1/J.
5 An Application
We saw in the previous section that any join-preserving closure opera-
tor on a complete lattice induces a complete sublattice of the lattice. In
this section we describe a technique to produce such a join-preserving
closure operator. Our technique is a relatively simple one, but when
applied to the Galois connection (I d, Mod) leads to some interesting
and new examples of complete sublattices of the lattice of all varieties.
As before, we assume that ( L, J.L) is a Galois connection between two
sets A and B. Let E ~ B be a subset of B which is closed under
the closure operator LJ.L, so that LJ.L(E) = E. We define the following
functions.
Definition 5.1. Let cp~ : P(B) -+ P(B) be defined by S H cp~(S) =

S n E, for every S ~ B.
Let cp~ : P(A) -+ P(A) be defined by T H cp~ (T) = J.L( cp~(L(T))) , for
every T ~A.
Then the following proposition is easy to prove.
Proposition 5.2. (i) The operator cp~ is a kernel operator on B which

preserves arbitrary unions and non-empty intersections, so that for
any indexed family (Sj)j EJ of subsets of B, we have
cp~(n{Si IJ E J}) n { 'P~ ( sj) I j E J} and

cp~(u{Sj I j E J}) u {cp~(Sj) I j E J}.
(ii) The operator cp~ is a closure operator on the lattice 1-lll-£ of the
subsets of A closed under the closure operator JU, and it preserves
non-empty joins.
As a consequence of this Proposition and Theorem 4.1 , we have the

following result.
Theorem 5.3. Let E ~ B be a subset of B closed under the closure

operator ifl, and let cp~ : P(A) --+ P(A) be the closure operator de-
fined by cp~(T) = Jl(cp~(i(T))), where cp~(i(T)) = i(T) n E for every
T E 1-l/1-£" Then the set of all cp~-closed subsets of A forms a complete
sub lattice of the lattice 1-l!l-£. •
The next proposition lists some additional properties of the operator

cp~ which we shall need.
Proposition 5.4. (i) i(Jl(cp~(S))) ~ cp~(i(Jl(S))), and if Sis a subset

of B for which i(fl(S)) = S, then also i(Jl(cp~(S))) = cp~(i(Jl(S))).
(ii) For all sets S and S' ~ B, we have i(Jl(cp~(ifl(S) U iJl(S')))) ~
cp~(ifl(S uS')).
Now let T be a subset of A which is closed under the operator Jli , so
that T E 1-£11-£· It is clear that the collection of all subsets ofT which
are closed under Jli forms a sublattice £(T) of 1-£11-£· We will denote by
£(cp~(T)) the lattice of all subsets of cp~(T) which are closed under
Jli. Note that the set cp~(T) is itself such a closed set. Then we get
the following result.
Theorem 5.5. LetT E 1-l!l-£ be a set such that cp~ (T) =1- T. Assume
that the operator cp~ satisfies the inclusion
(*)cp~(i~-t(SuS')) ~ if-t(cp~(i~-t(S)Ui~-t(S')))),forS =i(U1 ) andS' =

i(U2) with U1 and U2 E .C(T). Then the mapping 1/JA : .C(T) -t .C(T)
defined by U H cp~(U) for all U ~ T is a lattice homomorphism.
Proof: It follows from Proposition 5.2 that the map 1/JA preserves
joins, and we have to prove now that it also preserves meets. We have
1/JA (UI 1\ U2) = cp~ (U1 1\ U2) = ~-t(cp~(i(UI 1\ U2))) = f-t(cp~(if-t(i(UI) U
i(U2)))) = f-t(if-t(cp~(if-t(i(UI))) U cp~(if-t(i(U2 ))) ) , by (*)and Proposi-
tion 5.4(ii). Since i(cp~(Ui) = cp~(i(Ui)), fori= 1, 2, we see that
cp~(U1/\ U2) = ~-t(if-t(VJ~(if-t(i(UI))) U cp~(if-t(i(U2)))))
= f-t(if-t(cp~(i(UI) u cp~(i(U2 )))
= f-t(if-t(i(cp~(UI)) U i(VJ~ (U2))))
= cp~(UI) 1\ cp~(U2). •
As our source of examples we now consider the following situation.

Let T be any type of algebras, with W7 (X) the set of all terms of
type T on a countably infinite alphabet X. We take A to be the class
Alg(r) of all algebras of type T, and B to be the set W7 (X) 2 of all
equations of type T. We use the maps i = Id and 1-t = Mod, where
for any K ~ Alg(r) and any~~ W 7 (X)2
I d(K) = { u ~ v E W 7 (X)2 I A satisfies u ~ v for all A E K},
Mod(~)= {A E Alg(r) I A satisfies u ~ v for all u ~ v E ~}.
It is well known that these two maps form a Galois connection between
Alg(r) and W7 (X) 2, induced by the relation of satisfaction between
algebras and identities. The corresponding complete lattices are the
lattice .C(r ) of all varieties of type r, and the lattice £(r) of all equa-
tional theories of type T.
Now we take E = E(r) to be any property of identities which is an
equational theory, that is, any element E E £(r). The map cp~ with
~ H ~ n E is then a kernel operator which preserves meets. By Theo-
rem 5.3, the mapping cp~ = f-tip~ i = M odip~ldis a closure operator on
the lattice £(T), and preserves joins. Then the collection of cp~-closed
varieties forms a complete sublattice of .C(r) . This corresponds to the
Galois connection induced by the Galois-closed subrelation
Re = {(A, s ~ t) E Alg(r ) x E I A satisfies s ~ t}.
Corollary 5.6. Let E(r) be any property of identities which forms an
equational theory. Then the class of all ip~ - closed varieties of type T
forms a complete sublattice of the lattice ..C( T) of all varieties of type

T.
Two well-known examples of this corollary arise from taking E = E(T)

to be either the set N (T) of all normal identities of type T , or the set
Reg(T) of all regular identities of type T. Other kinds of identities
which have been used in this way are the externally compatible or
0-compatible identities. These various kinds of identities have been
studied by E. Graczy:riska, I. I. Mel 'nik and J . Ptonka (see [Gra], [Mel]
and [Pto]).
in [DW] we will introduce a countably infinite family of new such
choices for E, based on the idea of complexity of terms and identities,
and apply our theory to this new setting. This will allow us to general-
ize the notions of normal varieties and normalization of a non-normal
variety, to k-normalizations, for each natural number k.
References
[Arw] Arworn, S. Groupoids of Hypersubstitutions and G-solid vari-

eties, Shaker-Verlag, Aachen 2000.
[ADP] Arworn, S, K. Denecke and R. Poschel, Closure Operators on

Complete Lattices, to appear in Proc. International Conference on
Ordered Algebraic Structures, Nanjing, 1998.
[DWB] Denecke, K. Wismath, S. L., Universal Algebra and Applicatu-

ions in Th eoretical Computer Science, Chapman and Hall/CRC, Boca
Raton 2002.
[Den] Denecke, K., Clones closed with respect to closure operators,

Multi. Val. Logic, Vol. 4 (1999), 229 - 247.
[DW] Denecke, K. and S. L. Wismath, Complexity of Terms and the

Galois-Connection Id-Mod, this volume pp. 91-107.
[DG] Dikranjan, D. and E. Giuli, Closure Operators I, Topology Appl.

27 (1987), 129- 143.
[GW] Ganter, B. and R. Wille, Formate Begriffsanalyse, Springer 1996.
[Gra] Graczynska, E., On normal and regular identities and hyper-iden-

tities, Proc. Universal Algebra and Applied Algebra (Turawa 1988),
World Scientific Publishing, Teaneck, N.J. 1989, 107-135.
[Mel] Mel'nik, I. 1., Nilpotent shifts of varieties (in Russian), Mat. Za-
metki 14, No. 5, 1973. English translation: Math. Notes 14, 1973,
962-966.
[Pro] Plonka, J., On Varieties of Algebras Defined by Identities of Some

Special Forms, Houston Journal of Mathematics, Vol. 14, no. 2, 1988,
253- 263.
[Rei] Reichel, M., Bi-Homomorphismen und Hyperidentitiiten, Disser-

tation, Universitat Potsdam 1994.
[Tar] Tarski, A., A lattice theoretical fix point theorem and its applica-
tion, Pacific. J. Math. 5 (1955), 285- 310.
Authors' addresses:
Klaus Denecke
University of Potsdam
Institute of Mathematics
Am Neuen Palais
14415 Potsdam
Germany
e-mail: kdenecke@rz.uni-potsdam.de
Shelly Wismath
Dept.of Mathematics and CS
University of Lethbridge
Lethbridge, Ab.
Canada T1K-3M4
e-mail: wismaths@cs.uleth.ca
Galois Connections for Operations
and Relations
R. Poschel
Abstract
This paper reports on various Galois connections between operations and
relations. Several specifications and generalizations are discussed.
AMS Mathematics Subject Classification: 08A55, 03B10.
Key words: Operations, Relations, Clones, Relational algebras.
Introduction
Operational systems and relational systems are basic structures in

algebra and logic. Here we focus on one fundamental connection be-
tween operations and relations which is based on the invariance rela-
tion "an operation f preserves a relation r/'. This induces a Galois
connection 1 (Pol- lnv , see 1.3, 2.2) between sets of operations and
sets of relations. Many well-known algebraic concepts fit into this
framework (automorphisms, subalgebras, congruences, see e.g. proof
of 3.5).
There are numerous results on various Galois connections between
operations and relations characterizing the Galois closed sets. In this
paper we undertake the attempt to collect and systematize these re-
sults and to give a kind of survey. Therefore almost no proofs are
given and besides new formulations there are no new results. More-
over, we were not able to give a complete survey. We mainly report on
1 In [McKMT87, p. 147] this Galois connection is emphasized as the most basic
Galois connection in algebra. In [Wi103] it is mentioned as a general framwork for
the development of dyadic algebra.
231
232 R. Poschel
results which can be found in the following (not so easily accessible)

papers: [Pos80a] (some of the results of this report can also be found
in [Pos79]), [Pos84], [BorOO] and [BorPS].
For an orientation to which Galois connections are discussed the reader
may at first consult the Table 1 in Section 3 and Table 2 in Section 4.
1 Basic notions and notations

1.1. Operations and Relations. For operations f : An ----t A and
relations [!~Am (m, n E w+ = {1 , 2, 3, ... }) on a fixed base set A we
introduce the following notation:
Op<nl(A) ·- {j I J : An ----t A} (n-ary operations)

Op(A) ·- U~= l Op(nl(A) (finitary
operations)
Tr(A) ·- Op< 1 l(A) (transformations,
i.e. unary
permutations)
Sym(A) ·- {f E Tr(A) I f bijective} (permutations)
Rel(m) (A) ·- {QI[!~Am} (m-ary relations)
Rel(A) ·- U:=1 Rel(ml(A) (finitary relations)
~(A) ·- Rel< l(A)
1 (power set)
Eq(A) ·- {BE Rel( )(A) I e equivalence} (equivalence
2
relations)
The (A) may be omitted if it is clear from the context. 0-ary functions
or relations are not considered (mainly for technical reasons).
An m-tuple r E Am sometimes will be regarded as a mapping r :
m ----t A with m := {1, ... , m }, and its components are given by r =
(r(1), ... , r(m)).
1.2. lnvariance. Operations and relations can be connected by the

invariance property: A relation [! E Rel(m) (A) is invariant for an
operation f E Op(n)(A) (we also say, f preserves [!,or f is a polymor-
phism of [!, notation [Wil03] : f t> [!) if f[[! , . .. , Q] ~ [!, i.e. , if for all
r 1, ... , rn E [!we have j[r1 , . . . , rn] E [!. Here them-tuple f[rl , .. . , rn]
is defined component-wise by
f[rl, ... , rn](i) := f(rl(i) , ... , rn(i)) (i Em) .

Operations and Relations 233
Remark: In the universal algebra setting this means that f2 is (the

carrier of) a subalgebra of the algebra (A; f)m (m-th direct power);
from the relational point of view f is a (relational) homomorphism
from the relational system (A; f2)n = (An; f2n) into (A; f2).
For permutations f E Sym(A) we define: f strongly preserves f2 if
f[f2] = {! , i.e. , if the mapping f2 ----t f2 : r H f[r] is bijective.
1.3. The operators Pol- Inv. For F ~ Op(A) , G ~ Sym(A) and
Q ~ Rel( A) we define
PolAQ ·- {f E Op(A) I every f2 E Q is invariant for f} ,

Pol~) Q ·- Op(nl(A) n PolA Q, (n EN)
EndAQ ·- Tr(A) n PolA Q (endomorphisms),
wAutA Q · - Sym(A) n PolA Q (weak automorphisms) ,
AutAQ ·- {f E wAutA Q I f - 1 E wAutA Q}
(au tomorphisms),
InvAF ·- {f2 E Rel(A) I f2 is invariant for every f E F} ,
Inv(m) F ·- Rel(ml(A) n Inv F (mE N) ,
A
SubAF ·- Inv~) F (subalgebras) ,
ConAF ·- Eq n Inv A F (congruence relations) ,
slnvA G ·- {f2 E Rel(A) I
f2 is strongly invariant for every g E G} .
Note that the notation Pol stands for polymorphisms, not for polyno-
mials! The subscript A may be omitted if A is clear from the context.
The notion of weak automorphism applies here to relational systems
and differs from the weak automorphisms for algebras introduced by
A. Goetz ([Goe66]) and investigated in many papers of K. Glazek
(e.g. [Gla97]).
1.4. Clones. Recall, a clone is a set of operations (on a fixed base

set A) which is closed with respect to composition and contains all
projections ef E Op(n) given by ef(x 1 , . . . , Xn) =Xi· The composition
![91 . . . , 9n] off E Op(n) and 91 , . . . , 9n E Op(m) is the m-ary function
defined by
234 R. Poschel
The least clone containing F ~ Op(A) (i.e. , the clone generated by F)

will be denoted by (F) or (F)op(A)· Likewise, (H)Tr(A) and (G)sym(A)
denote the submonoid of Tr(A) generated by H ~ Tr(A) and the
subgroup of Sym(A) generated by G ~ Sym(A), respect ively.
1.5. Relational algebras (clones). We define the following (set-

theoretical) operations:
(1) ~A (nullary operation: to contain the diagonal (or equality)

relation ~A := {(a, a) I a E A}),
(2) n (intersection of relations of the same arity),

(3) x (product for m-ary 12 and s-ary a:
12 X a :={(al, ... , am, b1 , ... , bs) E Am+s I (a1, ... , am) E (!,
(b1, ... , b8 ) E a}) ,
(4) pr1 (projection onto a subset I of coordinates: For m-ary (! and

I = { i1, . . . , it} with 1 :S i 1 < i 2 · · · < it :S m we define
pr1 (Q) := {(aip . .. , ait) l3aj (j E {1, ... , m} \I):

(a1, ... , am) E Q} ),
(5) 1r a (permutation of coordinates: For m-ary (! and a permutation

a of { 1, . . . , m} let
1fa(r!) := {(aa(l), ... , aa(m)) I (at, ... , am) E Q} ),
(6) U (union of relations of the same arity),
(7) ---, (complementation: • (! := Am\ (! for m-ary Q)

A set Q ~ Rel(A) of relations is called a relational algebra, weak
Krasner algebra or Krasner algebra, respectively, if Q is closed with
respect to (1)-(5) , (1)-(6) or (1)-(7), respectively. These closures will
be denoted by [Q]RA, [Q]wKA and [Q]KA, resp. For finite sets A these
algebras are also called clones (relational clone, (weak) Krasner clone) .
For infinite A see 2.4.
Note t hat relational algebras contain relations of arbitrary (finite) ar-
ity and are much more general than Tarski 's relation algebras ( [Tar41]).
1.6. Remark. Relational algebras can be characterized also via clo-

sure with respect to first order formulas . To each first order formula
rp(R1, ... , Rq; XI, ... , Xm) with free variables xi, ... , Xm and mi-ary re-
lation (predicate) symbols~ (i E {1 , ... , q}) one can assign an oper-
ation (called logical operation)
according to
Hereby, for atomic formulas e.g. ~(x , y) we interpret f= {Ji(a, b) as

(a, b) E fli (for elements a, bE A). E.g. , the formula
rp(RI, R2; xi, x2) := 3z : RI (xi , z) 1\ R2(z, x2)

(with binary relation symbols) induces the operation relational prod-
uct:
for binary relations Q2, Q2 E Rel (2) (A) .

Let (3, /\, ... ) denote the set of all first order formulas that contain
only the indicated quantifier 3, the indicated connectives /\, . . . and
relation symbols and variables.
Let Lop A (3, /\, ... ) := { Fcp I rp E (3, /\ , .. . )} denote the corre-
sponding logical operations. Then we have (cf. [PosK79, 2.1.3]) for
Q <;;;; Rel(A):
Q relational algebra {:} Q closed w.r. t. Lop A (3, /\, =) ,

Q weak Krasner algebra {:} Q closed w .r. t. Lop A (3 , /\ , V, =) ,
Q Krasner algebra {:} Q closed w.r.t. LopA(3, /\ , V, •, =).
2 Galois connections and Galois closures
2.1. Galois connections. We use here the canonical definition of a

Galois connection between sets A and B , namely as a pair of mappings
236 R. Poschel
cp : ~(A) ---+ ~(B), 1/J : ~(B) ---+ ~(A) induced by a binary relation
I~ Ax B via
cp(X) := {bE B IVa E X : (a , b) E I},

'lj;(Y) := {a E A I Vb E Y : (a , b) E I}
for X ~ A, Y ~ B. It is well-known that such pairs (cp, 1/J ) of map-

pings can be characterized by the following properties: Both cp and 1/J
are anti-monotone (i.e. inclusion-reversing) and their compositions cp'lj;
and 'lj;cp are closure operators (i.e. , extensive, monotone, idempotent).
Therefore elements X E ~(A) andY E ~(B) satisfying X= 1/J (cp(X))
and Y = cp('lj;(Y)) are called Galois closed w.r.t . (cp, 1/J ). The sets of
Galois closed elements are
{7j;(Y) I Y ~ B} ~~(A) and {cp(X) I X~ A}~ ~(B) ,
and they form dually isomorphic complete lattices (w.r.t. inclusion) .
2.2. With the above definitions in 2.1 , the (strong) invariance relation
[> ( cf. 1.2) gives rise to the following Galois connections:
Pol- Inv between Rel(A) and Op(A) ,

End- Inv between Rel(A) and Tr(A) ,
wAut- lnv between Rel(A) and Sym(A) ,
Aut- slnv between Rel(A) and Sym(A).
Whenever a Galois connection is given there arises the question of

how to characterize the corresponding closure operators and the Galois
closed sets. This is well-known for these listed Galois connections (cf.
e.g. [PosK79]) and we summarize the results for finite A in the next
theorem (for notation see 1.4, 1.5 and 1.6).
Theorem 2.3. Let A be finit e and let F C Op(A) , H C Tr(A) ,

G ~ Sym(A), Q ~ Rel(A). Then
(F)op(A) = Pol lnv F , [Q]RA = lnv Pol Q ,

(H)n(A) = End lnv F , [Q]wKA = lnv End Q ,
(G)sym(A) =Aut Inv G , [Q]KA = lnv Aut Q.
In order to include results for infinite sets A we still need some more
notions and notations.
2.4. Relational clones and local closures. For b= (c1 , . .. , em) E
Em and a mapping f E A B let
For a family (l>i)iEI of relations Qi E Rel(mi) (A) (I being an arbitrary

index set) we introduce the general superposition with respect to given
bE Em, bi E Emi (i E I) for some set E as follows:
The strong superposition w.r.t. ii E Am, iii E Ami is given by
sS(a,ai) (l>i)iEI := {f[ii] If E Tr(A) surjective and Vi E I : ![iii] E l>i}·

Let Q ~ Rel(A). The least set of relations which contains Q and the
trivial (unary) relations 0, A , and is closed w.r.t. general superposition,
will be denoted by [Q](S)·
Note that arbitrary intersections are a particular case of general su-
perposition: niEI l>i = \E,bi) (l>i)iEl forb= bi E Am (mi = m , i E I) .
Further we introduce local closure operators which are in fact a kind of
interpolation operation (and can also be characterized topologically) .
Let F ~ Op(A), G ~ Sym(A) , Q ~ Rel(A) , mE N+. Then
m-LocF := {f E Op(n)(A) I 'iiit , ... , iin E Am3g E F:

f[iit , · · ·, iin] = g[iit, · · · , iln]} ,
LocF := n00
m =l
m-LocF ,
Loc0 G := {f E Sym(A) I Vn E N+va E An3g E G : f[ii) = g[ii]} ,

m-LOCQ := {Q E Rel(A) I Vr1 , ... ,rm E Q3a E Q:
{r1, ... , rm} ~a~ Q} ,
LOCF := n00
m=l
m-LOCF.
Let [Q]Rc := LOC[Q)(s) and [Q)wKc := 1-LOC[Q)(S)· We call Q ~

Rel(A) a relational clone (weak Krasner clone, resp.) if Q = [Q]Rc
238 R. Poschel
(Q = [Q]wKc, resp.). It should be noted that the definition of rela-

tional clone here differs from that in other papers (e.g. [Pos80a]) where
sets Q with Q = [Q](s) were also called relational clones.
One easily checks that 1-LOC Q is the closure of Q with respect to
arbitrary unions.
Further, let [Q]Kc be the closure of Q with respect to [.]wKc (weak
Krasner clone), complementation (-,) (cf. 1.5) and strong superposi-
tion (symbolically we express this by (KC) = (WKC)&(-,)&(sS)). Sets
Q with Q = [Q]Kc are called Kmsner clones.
A weak Krasner clone will be called a Pre-Kmsner clone if it is closed
with respect to strong superposition and contains the inequaltity re-
lation v = VA := { (x, y) E A 2 I x =J y} . The corresponding
closure is denoted by [Q]PKC (thus we have symbolically (PKC) =
(WKC)&(v)&(sS)).
Remark: In [BorOO] it is shown that (Pre-)Krasner clones are just the
(Pre-)Krasner algebras (cf. 1.5) which are closed w.r.t . strong super-
position and arbitrary (i.e. also infinite) intersections and unions. If
A is countable then one can even drop the closure w.r.t. strong super-
position.
With these notions we can state the following theorem which ex-
tends 2.3 to infinite sets A.
Theorem 2.5. Let F ~ Op(A) , H c Tr(A) , G c Sym(A), Q c

Rel(A). Then
Loc(F)op(A) =Pol Inv F, [Q]Rc = Inv Pol Q,

Loc(H)Tr(A) = End Inv F, [Q]wKc = Inv End Q ,
Loc 0 (G)Tr(A) = wAut Inv G, [Q]PKC = Inv wAut Q,
Loc 0 (G)sym(A) = AutsinvG, [Q]Kc = sinv Aut Q.
Remark: We have sinv G = Inv G whenever G s; Sym(A) is a permu-

tation group. Therefore e.g. sinv Aut Q = Inv Aut Q .
In the next sections we report on further Galois connections which
appear in connection with operations and relations. There are several
(mutually connected) possibilities for specialization and generalization
of those Galois connections considered in the previous section, e.g.:
(A) Restricting the operations and/or relations under consideration

but keeping the invariance relation " (! is invariant for f".
(B) Generalizing the operations (e.g. to partial operations or multi-

operations) and/ or relations with "canonically" modified invari-
ance relation.
(C) Considering "natural" closure operators on operations and/or

relations and trying to characterize the closed sets as Galois
closures of a suitable Galois connection.
3 Galois connections with restricted operations and

relations
We start with the approach (A) which was already described in [Pos80a,
15.1].
3.1. The Galois connections EPol-Rinv. Let E ~ Op(A) and

R ~ Rel( A). Then EPol - RJnv defined by
EPolQ :=En PolQ for Q ~ R,

Rlnv F := Rnlnv F for F ~ E
is a Galois connection between R and E. For motivation, E and R

may be regarded as those sets of operations and relations which are
of particular interest or especially important in some context.
For arbitrary E and R , there is no general proceedure for how to
characterize the Galois closed sets of operations or relations. However,
if we assume that E is a locally closed clone and R is a relational clone
then (by 2.5) there exist F0 ~ Op(A) and Q 0 ~ Rel(A) such that
E = Pol Q 0 and R = Inv F0 .
Then we have the following characterization for the Galois closures.
Theorem 3.2. Let E , R , F0 , Q 0 be as above. Then
EPol Rlnv F =En Loc(Fo u F)op(A) for F ~ Op(A) '

Rlnv EPol Q = R n LOC[Qo U Q](s) = R n [Qo U Q]Rc for Q ~ Rel(A).
240 R. Poschel
For finite A the local operators can be omitted and [.]Rc can be replaced
by [.]RA· ForE = Op(A) orR = Rel(A) one can choose R 0 = 0 or
Fo = 0.
Proof. The result immediately follows from Theorem 2.5 and E n
PolQ = PolQ 0 n PolQ = Pol(Q 0 u Q) , Rn Inv F = Inv F 0 n Inv F =
lnv(Fo U F). 0
Remark 3.3. Let G := Pol R = Pollnv F0 (then R = Inv G). Every

pair (E, R) as above gives rise to the induced closure operator
cls(E,G)(H) := EPol Rlnv H
on the principle ideal.,j..E =.,l..cAE := {HI HE LA and H ~ E} in the

lattice LA of all clones in Op(A). In case of finite A we have
cls(E,G)(H) = E 1\ (GV H) (see Fig. 1)
(A, V denoting meet and join in the lattice LA) according to Theo-
rem 3.2 (where the local operators can be omitted since A in finite) ;
for infinite A one has to restrict to the lattice of locally closed clones.
Op(A)
GVH
•
'.: ·.
/ •G
Fig. 1: The closure cls(E,a)(H) = E 1\ (G V H) in .,j..E

Let E = AAk and let G be the clone of all term operations of an algebra
A = (A; (fi)iEI) from a class JC of algebras. The cls(E,a)-closed sets
H ~ AAk (monoids in case k = 1) can be equipped with the algebraic
structure inherited from A. They were investigated in [JakMP03] as
so-called operation JC-algebras (transformation JC-algebras for k = 1)
and generalize the concept of near-rings.
In particular cases the Galois closures may be characterized more
intrinsically. Note, e.g. , that the Galois connections End- Inv and
wAut- Inv coincide with EPol- Inv forE= Tr(A) and E = Sym(A),
respectively (and trivial R = Rel(A)).
As a further example we mention the following characterization theo-
rem where the arity of operations or relations will be restricted (E =
Op(m)(A) orR= Rel(m)(A)).
Theorem 3.4. Let mEN+) F ~ Op(A)) Q ~ Rel(A) . Then
Pol Inv(m) F = m-Loc(F)op(A) ,

Inv Pol(m) Q = m-LOC[Q]Rc = m-LOC[Q](s).
Many other specializations of the Galois connection Pol - Inv were

investigated, in particular, special Galois closures were characterized
for various purposes. We collect some of these specializations in Ta-
ble 1 below where we briefly mention which Galois closures have been
characterized and give some references. The first two columns show
how to choose the sets E and R in order to apply Theorem 3.2 (how-
ever, although our approach via relational clones can be applied to
every particular case there are often other characterizations which are
more practicable; some details can be found e.g. in [Pos80a] and in
the references of Table 1).
In order to read Table 1 correctly we mention that sometimes the
reader sees a set F of operations where a set of relations should be
expected. Here we use the convention that an n-ary function f : An --?
A also may be treated as (n + 1)-ary relation, namely the graph of f
which we denote by
Thus, e.g., the so-called bicentralizer Pol Pol F ofF means Pol(Pol F•)•
(with p• := {r I f E F} for a set F of operations) . Note that e.g.
242 R. Poschel
an automorphism f of an algebra A= (A ; F) may be characterized in

different ways:
f E Aut A {:} f E Aut F• {:} r E Inv F .

Once the Galois closures are characterized the way is open for applica-
tions. As an example we t ake the so-called concrete characterization
of related structures like the subalgebra lattice Sub A , the automor-
phism group Aut A or the congruence lattice Con A of an algebra
A= (A; F). The problem may be posed as follows:
The concrete simultaneous characterization problem : Given a set L ~
s,p(A) of subsets of A , a set C ~ Eq(A) of equivalence relations on
A and a set G ~ Sym(A) of permutations on A. Does there exist
an algebra A such that G is the automorphism group of A , L is the
subalgebra lattice of A , and C is the congruence lattice of A ?
Note that we do require exact equality (e.g. C = ConA) and not
equality up to isomorphism (the latter is the abstract characteriza-
tion problem which sometimes is easier to answer and gives satisfac-
tory globally valid results; here are some references for the abstract
characterization problem: [Bir46], [BirF48], [GraS63], [Sch63] , [Sch64],
[GraL67], [J6n74], [SchT79]).
The answer to the concrete characterization problem in principle is
already provided with Theorem 3.2, but we shall make it more explicit:
Theorem 3.5 (cf. [Pos80a, (7.2)]). For a given permutation group
G ~ Sym(A) and given L ~ s,p(A) , C ~ Eq(A), there exists an algebra
A= (A; F) such that G =Aut A, L =Sub A and C =Con A if and
only if
c• = Sym(At n [Q]Rc
L = s,p(A) n [Q]Rc
C = Eq(A) n [Q]Rc
where Q := c• U L U C. The algebra A can be chosen with at most

m-ary operations if and only if [Q]Rc is replaced by m-LOC[Q]RC·
Proof. Note that f E Aut A {:} r

E Sym(A)• n Inv F , U E Sub A
{:} U E Inv(l) F = Sub F and (} E Con A {:} (} E Eq(A) n Inv F.
Thus the theorem immediately follows from Theorem 3.2. (For more
details we refer to [Pos80a]). 0
In Table 1 we also mention several further results for the concrete

characterization problem.
Tab. 1: Several closures of the Galois connections EPol-Rinv

(E ~ Op(A) , R ~ Rel(A))
E R Galois closure References

Op(A) Rel(A) Pollnv F [BodKKR69a] ,
[BodKKR69b] ,
[Gei68], [BakP75],
[Rom76], [Rom77a],
[Rom77b],[PosK79],
cf. 2.3, 2.5
InvPolQ [Gei68] ,
[BodKKR69a],
[BodKKR69b],
[Sza78], [PosK79] ,
[Pos79], [Pos80a],
cf. 2.3, 2.5
Generalization to infinitary relations or operations
Pollnv 00 F [Ros72], [KraP76] ,
[Poi81]
Inv 00 PolQ [Ros79]
Pol 00 Inv 00 F} [Kra76b], [KraP76],
lnv 00 Pol 00 Q [Poi81]
arity restrictions
Op(A) RellmJ(A) PollnvlmJ F [Gei68], [BakP75] ,
[Pos80a]
lnv(m) PolQ [Ros78]
Op(A) Rel(l)(A) PolSubF [Sch82, Thm. 1.6] ,
[Pos80a]
SubPolQ see below
Op(m)(A) Rel(A) lnv Pol(m) Q [Sza78] , [Pos80a]
to be continued at next page
244 R. Poschel
Table 1, continued from previous page

Tr(A) Rel(A) InvEndQ} [Kra38], [Kra50],
Endlnv F [Kra76a],
[Kra86], [Gou68],
[BodKKR69a],
[BodKKR69b],
[Pos80a] , [BorOO]
Sym(A) Rel(A) lnv AutQ } [Kra38],
(slnv Aut Q) [BodKKR69a] ,
wAutlnv F [BodKKR69b],
Autslnv F [Gou72a], [Pos80a],
[BorOO]
Sym(A) Rel(m)(A) Autlnv(m) F [Wie69]
restriction to (graphs of) operations only
Op(A) Op(A)• PolPolF [Sza78, Thm. 13] ,
[Faj77] , [Dan77](for
IAI = 3) , (also Kuz-
necov, cf. [Val76])
Op(A) Tr(A)• PolEndF [SauS82], ([Rei82]
implicit operations)
EndPolQ see below
concrete characterization problems
concrete characterization of Aut A
Op(A) Sym(A)• AutPolQ [Jon68] (cf. [Jon72,
(2.4.3)]) , [Kra50],
[ArmS64] , [Sza75],
[Bre76]
concrete characterization of AutA
for algebras A with at most m-ary
operations
Op(ml(A) Sym(A)• AutPol(m) Q [Plo68], [Gou72a],
[Jon72, (2.4.1)]
concrete characterization of Sub A
Op(A) s,p(A) SubPolQ [BirF48] (cf.
[J6n72 , (3.6.4)]) ,
[Gou68], [Gou72b]
to be continued at n ext page

for unary algebras:
[J6n72, (3.6.7)],
[JohS67]
concrete characterization of Con A
Op(A) Eq(A) Con Pol Q [Arm70](partial
solution), [J6n72,
(4.4.1)], [QuaW71] ,
[Wer74], [Dra74]
Pol Con F for p-rings (A; F)
[Isk72]
concrete character·ization of End A
Op(A) Tr(A)• EndPolQ [Lam68], [Gra168],
[SauS77a], [Sto69],
[Sto75], [Jez72],
[Sza78, Thm. 15]
concrete characterization of
Op(A) Sym(A)• u '+J(A) Aut A & Sub A [Sto72] , [Gou72b],
cf. 3.5
Sym(A)• U Eq(A) Aut A & Con A [Wer74](conjecture)
cf. [Pos80b], (for
simple A [Sch64]),
cf. 3.5
Op(A)• u Sub(A) End A &Sub A [SauS77b] (cf.
[J6n74])
Op(A)•usub(A) Aut A & SubA & ConA
UEq(A)
[Sza78], [Pos80a] ,
cf. 3.5
Generalization to partial functions and multi/unctions see 4.1
4 Further Galois connections
Now let us consider further Galois connections from the points of view
(B) and (C) mentioned at the end of Section 2.
Concerning (B) we mention here some generalizations of the opera-
tions under consideration while the other side of the Galois connection
246 R. Poschel
- the relations Rel(A) - remains unchanged. The invariance relation

which induces the Galois connection is a more or less straightforward
generalization and we refer to the references.
4.1. Galois connections with generalized operations.
• partial functions (i.e. mappings f : B -----t A with B C An):
[Ros83] ([FleR78]), [HadR92] (cf. also [Fre66]) ,
[Rom81], [BorHP91], [SusK94],
• multifunctions (i.e. mappings f : An-----t \f.\(A)): [BorOl], [DreP],
• heterogeneous functions (i.e. mappings f : A 1 x Ain -----t Aio for

a family (Ai)iEI of base sets, i 0 , i 1 , ... , in E I): [Pos73] .
Another generalization can be found in [PosROO] where cofunctions
(i.e. mappings f : A -----t {1, .. . , n} x A) and corelations (i.e. subsets
(! ~ { 1, .. . , n} A) are considered instead of functions and relations.
A Galois connection which generalizes many features of functions,
cofunctions, partial functions and multifunctions was investigated in
[Ro:BOO].
Finally we demonstrate (C) point of view.
4.2. Closures of relational systems as Galois closures.
In Table 2 below we show how several closures of systems of rela-
tions can be characterized by Galois connections. These closures can
be classified by logical operations (cf. 1.6) of particular sets of first
order formulas (cases ( 1)- (8)) or other operators on relations (cases
(9)-(15)) .
However the investigations were not really motivated by this classifica-
tion. In cases (1)-(8) the algebraic structure of the relational systems
under consideration was of most important interest and gave such a
hierarchy of relational systems. In cases (9)-(15) also the closed sets
of operations were the starting point of interest.
We mainly report on results from [BorPS](cases (1)-(8)) and [Pos84]
(cases (9)-(14)) and refer to these papers for further details. Some of
the cases are already discussed in this paper and were mentioned in
Table 1; however they fit into the scheme used here and therefore they
are listed once more.
Table 2 is organized as follows. The first column shows the closure by
indicating under which operations the relational system is to be closed.

The next column gives a notation for the closure (if it was int roduced
somewhere). The last two columns indicate the Galois connection and
briefly describe the Galois closed sets (the last column hereby shows
what is the operational counterpart for the Galois connection under
consideration). For simplicity we assume that the base set A is finite
for the cases (1)-(8) . The generalization to arbitrary base sets A for
(1),(2),(3) is given in (9),(14),(15). In cases (6)-(9) one has essentially
to consider n-complete relational systems and locally closed opera-
tional system in order to get the results also for infinite A (for details
see [BorPS]).
We shall provide here explicitly only those notions in Table 2 which
concern the relational systems (for the operational part see the above
mentioned papers).
A Boolean system , briefly BS, is a set Q ~ Rel(A) of relations which
is closed with respect to the Boolean operations (union U, intersection
n, complement ---,, cf. also 1.5) , contains as constants the sets 0 and
Am (m E J:~+) and is closed with respect to the following operation
W s: Rel(n)(A)---+ RelCml(A) for all mappings s : rr---+ m (n,m EN+ ):
Ws(Q) := { (al , . .. , am) E Am I (as(l), ... , as(n)) E Q} for (} E Rel(n) (A).
A Boolean system with identity, briefly BSI, is a Boolean system which

contains the diagonal relation ~A (cf. 1.5).
A Boolean system with projections, briefly BSP, is a Boolean system
which is additionally closed under the projections pr, (cf. 1.5).
Tab. 2: Relational closures as Galois closures
[
Galois connection
Nota- rela- dosed opera-
Closure closed
tion tiona} system tiona} system
I for fimte base set A:
slnv - Aut (cf. 2.3 and Tab. 1)
(1) Lop A (:3, /\, V, •, =) [Q]KA Krasner alge- group of per-
bra (cf. 1.5, mutations
1.6)
to be continued on next page
248 R. Peschel
Galois connection
Closure closed rela- closed opera-
tional system tional system
Inv- End (cf. 2.3 and Tab. 1)
(2) LopA(3, 1\, V, =) [Q]wKA weak Kras- monoid of
ner algebra unary func-
(cf. 1.5, 1.6) tions
Inv- Pol (cf. 2.3 and Tab. 1)
(3) LopA(3, 1\, =) [Q]RA relational alge- clone of fini-
bra (cf. 1.5, tary functions
1.6)
Inv- pPol
(4) LopA(/\,=) weak system down-closed
with identity clone of fini-
tary partial
functions
Inv- mPol
(5) Lop A(/\) weak system of down-closed
relations clone of
finitary multi-
functions
slnv- sEnd (sbmEnd, resp.)
(6) Special mo-
noid of unary
functions (cf.
LopA(::l,/\,V,---,) [Q]BSP BSP (cf. 4.2) [BorPS, 7.9])
(6') (down-closed
involuted mo-
noid of bitotal
multifunc-
tions, resp.)
Galois connection
slnv - spmEnd
(7) LopA(A,V,--, , =) [Q)BSI BSI (cf. 4.2) down-closed
involuted
monoid of
pp-multifunc-
tions (partial
permutations)
slnv -smEnd
(8) LopA(A,V,•) [Q]ss BS (cf. 4.2) down-closed
involuted mo-
noid of unary
multifunctions
I for arbitrary base set A:
slnv- Aut (cf. 2.5 and Tab. 1)
(9) (KC) = [Q]Kc Krasner clone locally closed
(WKC)&( --,)&(sS) (cf. 2.4) group of per-
mutations
Inv- wAut (cf. 2.5 and Tab. 1)
(10) (PKC) = [Q)PKC Pre-Krasner locally closed
(WKC)&(v)&(sS) clone (cf. 2.4) monoid of per-
mutations
lnv- sur-End
(11) (WKC)&(sS) InvH for sets locally closed
H of surjec- monoid of sur-
tive unary jective unary
functions functions
Inv - inb-End
(12) (WKC)&( --,) InvH for lo- locally closed
cally invertible locally invert-
monoids H ible monoid
of unary
functions
250 R. Poschel
Galois connection
Inv - inj-End
(13) (WKC)&(v) InvH for locally closed
monoids H monoid of
of injective injective unary
functions functions
Inv - End (cf. 2.5 and Tab. 1)
(14) (WKC) = [Q]wKc weak Krasner locally closed
(RC)&(1-LOC) clone (cf. 2.4) monoid of
unary func-
tions
Inv - Pol (cf. 2.5 and Tab. 1)
(15) (RC) = [Q] Rc relational locally closed
(S)&(LOC) clone (cf. 2.4) clone of fini-
tary functions
There are many interdependences between the operations on relations

mentioned in this paper. E.g., with intersections and the operations
Ws (cf. 4.2) one can express the operations (3) and (5) from 1.5. Of-
ten it depends on the cardinality of the base set (finite, countable, un-
countable) which operations on relations are needed for the characteri-
zation of the Galois closed sets (detailed investigations can be found in
[BorOO]) and how they are interrelated (e.g. (KC) <* (RC)&( •) holds
for finite and countable but not for uncountable A, [DroKMPOl]) .
Finally we mention that there is an interesting interpretation of the

operations of relational algebras (and Krasner algebras) as operations
of the Peircean algebraic logic (PAL) proposed also by R.W. Burch
in [Bur91] and modelled mathematically in power context families ,
see [Pol02] and [ArnOl], for more details we refer to [HerP03]. PAL is
closely related to the existential graphs that Peirce developed in the
late 1890s, and to the conceptual graphs (cf. e.g. [Sow92]).
References
[Arm70] M. Armbrust. On set-theoretic characterization of congru-

ence lattices. Z. Math. Annalen, 16: 417- 419, 1970.
[ArmS64] M. Armbrust and J . Schmidt. Zum Cayleyschen Darstel-

lungssatz. Math. Annalen, 154: 70-73, 1964.
[Arn01] M. Arnold. Einfiihrung in die kontextuelle Relationenlogik.

Diplomarbeit, Technische Universitiit Darmstadt, 2001.
[BakP75] K.A. Baker and A.F . Pixley. Polynomial interpolation and

the chinese remainder theorem for algebraic systems. Math.
Z., 143: 165-174, 1975.
[Bir46] G. Birkhoff. Sobre los grupos de automorfismos. Un. Math.

Argentine, 11: 155-157, 1946.
[BirF48] G. Birkhoff and 0. Frink. Representations of lattices by sets.

Trans. Amer. Math. Soc. , 64: 299- 316, 1948.
[BodKKR69a] V.G. Bodnarcuk, L.A. Kaluznin, N.N. Kotov, and B.A.

Romov . Galois theory for Post algebras I. Kibernetika (Kiev),
(3): 1- 10, 1969. (Russian) .
[BodKKR69b] V.G. Bodnarcuk, L.A. Kaluznin, N.N. Kotov, and B.A.

Romov. Galois theory for Post algebras II. Kibernetika (Kiev) ,
(5): 1-9, 1969. (Russian).
[BorOO] F. Borner. Krasneralgebren. Habilitationsschrift (UniversiUit

Potsdam), Logos Verlag, 2000.
[Bor01] F . Borner. Total multifunctions and relations. In I. Cha-

jda, M. Droste, G. Eigenthaler, W.B. Muller, and R. Poschel,
editors, Contributions to General Algebra, volume 13. Verlag
Johannes Heyn, Klagenfurt , 2001. (Proceedings of the 60th
Workshop on General Algebra, Dresden, June 2000) .
[BorHP91] F. Borner, L. Haddad, and R. Poschel. Minimal partial

clones. Bull. Austr. Math. Soc., 44: 405-415, 1991.
252 R. Poschel
[BorPS] F. Borner, R. Poschel, and V. Sushchansky. Boolean sys-

tems of relations and Galois connections. Acta Sci.Math.
(Szeged)[XXX]. (submitted).
[Bre76] D.A. Bredhin. Inverse semigroups of local automorphisms

of universal algebras. Sibirsk. Mat. Z. , 17: 499- 507, 1976.
(Russian).
[Bur91] Robert W. Burch. A Pearcean Reduction Thesis: the founda-

tions of topological logic. Texas Tech University Press, 1991.
[Dan77] A.F. Danil'cenko. On parametrical expressibility offunctions

of he 3-valued logic. Algebra i logika, 16: 397- 416, 1977.
[Dra74] H. Draskovicova. On a representation of lattices by congru-

ence relations. Mat cas., 24: 69- 75, 1974.
[DreP] Th. Drescher and R. Poschel. Multiclones and relations.

Multiple- Valued Logic[XXX]. (to appear).
[DroKMP01] M. Droste, D. Kuske, R. McKenzie, and R. Poschel.

Complementary closed relational clones are not always Kras-
ner clones.
Algebra Universalis, 45: 155- 160, 2001.
[Faj77] S. Fajtlowicz. Duality for algebras. In Contributions to univer-

sal algebra, volume 17 of Colloq. Math. Soc. J. Bolyai. North-
Holland, 1977.
[FleR78] I. Fleischer and I.G. Rosenberg. The Galois connection be-

tween partial operations and relations. Pac. J. Math., 79:
93- 97, 1978.
[Fre66] R.V. Freivald. Functional completeness for not everywhere

defined functions of the algebra of logic. Diskretnyi Analiz, 8:
55- 68, 1966. (Russian) .
[Gei68] D. Geiger. Closed systems offunctions and predicates. Pacific

J. Math ., 27: 95- 100, 1968.
[G~a97] K. G~azek. A cert ain Galois connect ion and weak automor-
phisms. Acta Univ. Patacki. Olomouc, Fac. rer. nat. , Mathe-
matica, 36: 15- 26, 1997.
[Goe66] A. Goetz. On weak isomorphisms and weak homomorphisms

of abstract algebras. Colloq. Math. , 14: 163- 167, 1966.
[Gou68] M. Gould. Multiplicity type and subalgebra structure in uni-

versal algebras. Pacific J. Math ., 26: 469- 485, 1968.
[Gou72a] M. Gould. Automorhism groups of algebras of finite type.

Canadian J. Math, 24: 1065- 1069, 1972.
[Gou72b] M. Gould. Automorphism and subalgebra structure in al-

gebras of finite type . Algebra universalis, 2: 369- 374, 1972.
[Gra,L67] G. Gratzer and W.A. Lampe. On subalgebra lattices of

universal algebras. J. Alg. , 7: 263-270, 1967.
[GraL68] G. Gratzer and W.A. Lampe. Representation of some trans-

formation semigroups as endomorphism semigroup universal
algebras II. Notices A.M.S. , 15: 625, 1968.
[GraS63] G. Gratzer and E.T. Schmidt. Characterizations of congru-

ence lattices of abstract algebras. Acta Sci. Math . (Szeged},
24: 34-59, 1963.
[HadR92] L. Haddad and I. Rosenberg. Completeness theory for finite

partial algebras. Algebra Universalis, 29: 378-401 , 1992.
[HerP03] J. Herreth Correia and R. Poschel. The power of PAL.

Manuscript in preparation, 2003.
[Isk72] A. Iskander. Algebraic functions on p-rings. Colloq. Math.,

25: 37- 41, 1972.
[JakMP03] D. Jakubikova-Studenovska, D. Masulovic, and

R. Poschel. Completeness criteria and invariants for
operation and transformation algebras. B eitriige zur Algebra
und Geometrie (Contributions to Algebra and Geometry) ,
2003. (to appear).
254 R. Poschel
[Jez72] J. Jezek. Algebraicity of endomorphisms of some relational

structures. Acta Univ. Carolinae- Math. et Phys. , 13: 43- 52,
1972.
[JohS67] J. Johnson and R.L. Seifert. A survey of multi-unary alge-

bras. Mimeographed seminar notes, U.C.Berkeley, 1967.
[Jon68] B. Jonsson. Algebraic structures with prescribed automor-

phism groups. Call Math., 19: 1- 4, 1968.
[Jon72] B. Jonsson . Topics in universal algebra, volume 250 of Lectur·e

notes in Mathematics. Springer-Verlag, Berlin - Heidelberg -
New York, 1972.
[Jon74] B. Jonsson. Some recent trends in general algebra. Proc.

Symp. pur·e Math. , 25: 1- 19, 1974. (Proc. Tarski Symp.,
Berkeley 1971).
[Kra38] M. Krasner. Une generalisation de la notion de corps. J.

Math. pure et appl. , 17: 367- 385, 1938.
[Kra50] M. Krasner. Generalisation abstraite de la theorie de Galois.

Algebre et Theorie des Nombres, 24: 163-168, 1950. (Coli.
Intern. Centre nat. Rech. Sci.) .
[Kra76a] M. Krasner. Endotheorie de Galois abstraite et son theoreme

d'homomorphie. C.R. Acad. Sci. Paris, Ser. A , 282: 683- 686,
1976.
[Kra76b] M. Krasner. Polytheorie de Galois abstraite dans le cas infini

general. Ann. Sci. Univ. Clearmont, Ser. Math., 13: 87- 91,
1976.
[Kra86] M. Krasner. Abstract Galois theory and endotheorie. Acta

Sci. Math ., 50: 253- 286, 1986.
[KraP76] M. Krasner and B. Poizat. Theorie des Galois pour les

algebres de Post infinitaires. Preprint, 1976. (for an extended
version see [Poi81]).
[Lam68] W.A. Lampe. Representations of some transformation semi-

groups as endomorphism semigroups of universal algebras I.
Notices A.M.S., 15: 625, 1968.
[McKMT87] R.N. McKenzie, G.F. McNulty, and W .F. Taylor. Al-

gebras, Lattices, Varieties , volume I. Wadsworth & Brooks,
Monterey, 1987.
[Plo68] E. Plonka. A problem of B. Jonsson concerning automor-
phisms of a general algebra. Call. Math ., 19: 5- 8, 1968.
[Poi81] B. Poizat. Theorie de Galois pour les algebres de Post infini-
taries. Z. Math . Logik Grundl. Math ., 27: 31- 44, 1981.
[Pol02] S. Pollandt. Relation graphs: a structure for representing re-
lations in contextual logic of relations. In U. Priss, D. Corbett,
and G. Angelova, editors, Conceptual Structures: Integration
and Interfaces. Proc. 1Oth Int. Conf. on Conceptual Struc-
tures, volume 2393 of LNAI, pages 34- 47. Springer, 2002.
[Pos73] R. Poschel. Postsche Algebren von Funktionen iiber einer
Familie endlicher Mengen. Z. Math. Logik Grundlagen Math. ,
19: 37- 74, 1973.
[Pos79] R. Poschel. Concrete representation of algebraic structures

and a general Galois theory. In H. Kautschitsch, W .B. Miiller,
and W. Nobauer, editors, Contributions to General Algebra,
pages 249- 272. Verlag J . Heyn, Klagenfurt , 1979. (Proc. Kla-
genfurt Conf. 1978).
[Pos80a] R. Poschel. A general Galois theory for operations and rela-
tions and concrete characterization of related algebraic struc-
tures. Report R-01/80, Zentralinstitut fiir Mathematik und
Mechanik, Akademie der Wissenschaften der DDR, Berlin,
1980. (with German and Russian summaries), 101 pp.
[Pos80b] R. Poschel. On a conjecture of H. Werner. Algebra Univer-
salis, 10: 133- 134, 1980.
[Pos84] R. Poschel. Closure properties for relational systems with

given endomorphism structure. Beitriige zur Algebra und Ge-
ometrie, 18: 153-166, 1984.
256 R. Poschel
[PosK79] R. Poschel and L.A. Kaluznin. Funktionen- und R elatione-

nalgebren. Deutscher Verlag der Wissenschaften, Berlin, 1979.
Birkhauser Verlag Basel, Math. Reihe Bd. 67, 1979.
[PosROO] R. Poschel and M. Rof3iger. A general Galois theory for
cofunctions and corelations. Algebra Universalis, 43: 331-
345, 2000.
[QuaW71] R. Quackenbush and B. Wolk. Strong representation of
congruence lattices. Algebra Universalis, 1: 165-166, 1971.
[Rei82] J. Reitermann. The Birkhoff theorem for finite algebras. Al-
gebra Universalis, 14: 1- 10, 1982.
[Rom76] B.A. Romov. Local characterizations of Post algebras I.

Kibernetika, (5): 38- 45, 1976. (Russian).
[Rom77a] B.A. Romov. Local characterizations of Post algebras II.
Kibernetika, (1): 12-20, 1977. (Russian).
[Rom77b] B.A. Romov. Galois connection between iterative Post al-

gebras and relations on an infinite set. Kibernetika, (3): 62- 64,
1977. (Russian).
[Rom81] B.A. Romov. The algebras of partial functions and their
invariants. Kibernetika, (2): 1-11 , 1981. (Russian) .
[Ros72] I. Rosenberg. A classification of universal algebras by infini-
tary relations. Algebra Universalis, 1: 350-354, 1972.
[Ros74] I. Rosenberg. Une correspondance de Galois entre les algebres
universelles et des relations dans le meme universe. C.R.
Acad.Sci., Paris Ser. A, 279: 581-582, 1974. (Engl. sum-
mary: C.R. Acad.Sci., Paris Ser. A 280(1975), 615-616).
[Ros78] I. Rosenberg. The subalgebra system of direct powers. Algebra
Universalis, 8: 221- 227, 1978.
[Ros79] I.G. Rosenberg. On a Galois connection between algebras

and relations and its applications. In H. Kautschitsch, W.B.
Muller, and W. Nobauer, editors, Contributions to General
Algebra, pages 273-289. Verlag J. Heyn, Klagenfurt, 1979.
(Proc. Klagenfurt Conf. 1978).
[Ros83] I.G. Rosenberg. The Galois theory for partial clones. In R.S.
Freese and O.C. Garcia, editors, Universal Algebra and Lattice
Theory, volume 1004 of Lecture Notes in Math ., pages 257-
272. Springer-Verlag, 1983. (Proceedings Puebla 1982) .
[RoBOO] M. Rofiiger. A unified general Galois theory. Multiple Valued

Logic, 5(3): 239- 258, 2000.
[SauS77a] N. Sauer and M.G. Stone. The algebraic closure of a semi-

group of functions . Algebra Universalis, 7: 219- 233, 1977.
[SauS77b] N. Sauer and M.G . Stone. Endomorphism and subalgebra

structure; a concrete representation . Acta Sci. Math. (Szeged},
38: 397- 402, 1977.
[SauS82] N. Sauer and M.G. Stone. A Galois correspondence between

algebras and endomorphisms. Call. Math. Soc. J. Bolyai, 29:
691- 695, 1982. (Proceedings of the Esztergom Coli., 1977) .
[Sch63] E.T. Schmidt. Universale Algebren mit gegebenen Auto-

morphismengruppen und Unteralgebrenverbanden. Acta Sci.
Math ., 24: 251- 254, 1963.
[Sch64] E.T. Schmidt. Universale Algebren mit gegebenen Automor-

phismengruppen und Kongruenzverbanden. Acta Math . Acad.
Sci. Hung., 15: 37-45, 1964.
[Sch82] J. Schmidt. Clones and semiclones of operations. Call. Math.

Soc. J. Bolyai, 29: 705- 723, 1982. (Proceedings of the Es-
ztergom Coli. , 1977) .
[SchT79] B.M. Schein and V.S. Trohimenko. Algebras of multiplace

functions. Semigroup forum , 17: 1- 64, 1979.
[Sow92] J.F. Sowa. Conceptual graphs summary. In T .E. Nagle, J.A.

Nagle, L.L. Gerholz , and P.W. Eklund, editors, Conceptual
structures: Current Research and Practice, pages 3- 52. Ellis
Horwood, 1992.
[Sto69] M.G. Stone. On certain endomorphism semigroups in univer-

sal algebras. Notices A.M.S. , 16: 203, 1969.
258 R. Poschel
[Sto72] M.G. Stone. Subalgebra and automorphism structure in uni-

versal algebras; a concrete characterization. Acta Sci. Math.
(Szeged} , 33: 45- 48, 1972.
[Sto75] M.G. Stone. On endomorphism structure for algebras over a

fixed set. Call. Math ., 33: 41- 45, 1975.
[SusK94] V.I. Sushchansky and N.Y. Kormysheva. Galois connection
between coherence algebras and inverse semigroups of partial
permutations over a finite set. Dopovidi AN Ukraine, (12):
7- 10, 1994. (Ukrainian).
[Sza75] L. Szabo. Characterization of related semigroups of universal
algebras. Acta Sci. Math . (Szeged} , 37: 143- 147, 1975.
[Sza78] L. Szabo. Concrete representation of related structures of
universal algebras. Acta Sci. Math. {Szeged}, 40: 175- 184,
1978.
[Tar41] A. Tarski. On the calculus of relations. J. Symbolic Logic, 6:

73- 89, 1941.
[Val76] 1.1. Valuce. Mappings, Algebraic aspects of theory. lzd. Stiinca,

Kishinev , 1976.
[Wer74] H. Werner. Which partition lattices are congruence lattices?

Call. Math. Soc. J. Bolyai {Lattice Theory) , 14: 433- 453,
1974.
[Wie69] H. Wielandt. Permutation groups through invariant relations

and invariant functions. Lecture notes, 1969.
[Wi103] R. Wille. Dyadic Mathmatics - Abstraction from Logial

Thought. Manuscript, 2003.
Author's address:
Reinhard Poschel
Institut fur Algebra
TU Dresden
D - 01062 Dresden
GERMANY
e-mail: poeschel@math.tu-dresden.de
Galois Connections and Polynomial
Completeness
K. Kaarli*
Abstract
The aim of this paper is to emphasize the importance of the Galois theo-
retic approach in the study of polynomial completeness problems. We show
that reasoning in terms of Galois connections can be used for describing
(local) polynomial functions , for classifying and introducing new types of
polynomial completeness properties, etc. Special attention is paid to al-
gebras with a majority term, in which case the local term functions of an
algebra A are determined by the subuniverses of A 2 .
AMS Math ematics Subject Classification: 08A40, 06A15.
Key words: Galois correspondence, Polynomial completeness, Primality,
Order functional completeness, Endoprimality.
1 Introduction
In the earlier literature polynomially complete was a synonym of func-

tionally complete. Here we use the notation and terminology of the
monograph [10] where polynomial completeness is a property of alge-
bras (or classes of algebras) whose characteristic feature is richness in
polynomial functions, in some sense. So there are many polynomial
completenesses, and functional completeness is just one of them. Re-
call that a finite algebra A is called functionally complete if all finitary
functions defined on its universe A are polynomials. However, there
*Supported by grant no. 4353 from the Estonian Science Foundation

259
260 K. Kaarli
are many other completenesses, and a typical one is affine complete-

ness: an algebra A is affine complete if its polynomial functions are
exactly the congruence preserving functions . Since the term functions
are special polynomial functions, we consider the property of primal-
ity and its variations also as special kinds of polynomial completeness.
Recall that a finite algebra A is called primal if all finitary functions
defined on its universe A are term functions. One of several varia-
tions of primality is endoprimality: an algebra A is endoprimal if its
term functions are exactly the finitary functions defined on A which
permute with all endomorphisms of A .
The examples of polynomial completenesses given so far are strictly
universal algebraic because they make sense for any class of (universal)
algebras. If we restrict ourselves to a certain special class of algebras
then it may happen that all polynomial (term) functions of algebras
of that class preserve some specific relation which can be defined only
for algebras of that class. In this situation it is reasonable to consider
such algebras for which the polynomial (term) functions are exactly
the functions on their universe which preserve that specific relation. A
typical example is lattices and the order relation, and the correspond-
ing polynomial completeness property is order functional complete-
ness, also known as order polynomial completeness. The combination
of order functional completeness and affine completeness is order affine
completeness: a lattice is order affine complete if its polynomial func-
tions are exactly the functions on their universe which preserve the
congruences and the order relation.
It is also possible (and useful) to consider instead of polynomial or
term functions local polynomial or local term functions, respectively.
For example an algebra (not necessarily finite) is locally functionally
complete if all functions defined on its universe are local polynomials,
that is, can be interpolated by a polynomial at any finite subset of
their domain. It will turn out that the approach based on Galois con-
nections is especially useful in the case of local versions of polynomial
completenesses.
2 The basic Galois connection

Let A be an arbitrary nonempty set. We denote by F(A) the set of
all finitary functions on A and by R(A) the set of all finitary relations
Galois Connections and Polynomial Completeness 261
on A. We also use symbols Fn(A) and Rn(A) for denoting the sets
of all n-ary functions and relations, respectively, on A. Thus, every
f E Fn(A) is a function from An to A and every r E Rn(A) is a
subset of An. Note that we do not require that a function f E Fn(A)
be essentially n-ary. IfF is any subset of F(A) then its n-ary part is
Fn = F n Fn(A) . The n-ary part of R ~ R(A) is defined similarly.
Our Galois connection is generated by the fundamental binary relation
"preserves" between F(A) and R(A) which we describe now. Let
f E Fn(A) and r E Rm(A). We say that f preserves r if for every
m x n matrix (aij) with all its rows in r , we have:
(!(an , ... , aln), . . . , f(aml , ... , amn)) E r.
Instead of saying "f preserves r" one often says: "r is !-invariant" or
"f respects r" or "r supports f" or (especially if r is an equivalence
relation) "f is r-compatible".
The context (F(A), R(A) , preserves) generates the Galois connection
between the ordered sets (p(F(A)) , ~)and (p(R(A)) , ~) in the usual
way (say, as in formal concept analysis) . This connection is established
by the mappings:
cp: p(R(A)) --7 p(F(A)),

R c--+ R'P = {f E F(A) If preserves all r E R}
and
{} : p(F(A)) --7 p(R(A)),

F c--+ pe = {r E R(A) Iris preserved by all f E F}.
This basic Galois connection can be extended to partial functions but

also restricted to functions and/or relations of restricted arities, to
equivalence relations, etc.
The central question related to any Galois connection is to describe the
Galois-closed members of the two given ordered sets. In the present
situation this means to describe: a) the sets of relations having the
form pe where F ~ F( A) and b) the sets of functions having the form
R'P where R ~ R(A) . The answers to these questions are well known ,
especially in the case of a finite set A. In order to formulate them,
we introduce the notion of relational clone, also known as relation al-
gebr-a or coclone. In fact there is no generally accepted definition of
262 K. Kaarli
the relational clone. Different authors use different basic operations.

Perhaps the most economical definition is one due to R. Poschel [13]
(see also R. Poschel's contribution "Galois Connections for Operations
and Relations" in this volume): the relational clone on A is a subset
R s;;; R(A) which contains all diagonals and is closed with respect to
the so-called general superposition operation. The latter actually cov-
ers many different simpler operations which are obtained by choosing
appropriate functions the general superposition operation depends on.
We prefer to give a definition containing more but simpler operations
rather than a single complicated operation.
Definition 2.1. A subset R of R(A) is said to be a relational clone,

if it contains the binary diagonal 6.t
= {(a , a) I a E A} of A and is
closed with respect to the following operations:
1. renumbering of coordinates, that is, assigning torE Rn(A) the

relation ra = {(xa(l) , . . . , Xa(n)) I (x1 , ... , Xn) E r} where O" is a
permutation on the set {1 , .. . , n} ;
2. adding a coordinate, that is, assigning to r E Rn(A) the relation

r X A E Rn+l(A) ;
3. projection, that is, assigning to r E Rn+l the relation
4. intersection, that is, assigning tor, s E Rn(A) their intersection

r n s E Rn(A).
It easily follows from the definitions that every relational clone on A

contains the diagonals 6.~ = { (a , ... , a) E An I a E A} for every n ,
and their direct products. The set of all such "generalized" diagonals
is the smallest relational clone on A. We denote it by V(A) . It is not
difficult to see that the binary part of every relational clone is closed
with respect to the usual relation product.
Theorem 2.2. (see [14], [16]) Let A be a finite set. Th en

1. a subset F s;;; F(A) is Galois-closed iff it is a clone on A , that
is, F contains all projection maps and is closed with respect to
superposition;
2. a subset R ~ R(A) is Galois-closed iff it is a relational clone on

A.
In case of an infinite set A it is impossible to characterize Galois-closed
sets of functions and relations in purely algebraic terms; it is necessary
to impose additional conditions of a topological nature. Given a subset
F ~ F(A) we define its closure F as follows: a function f E Fn(A)
belongs to F iff it can be interpolated by members of F on every
finite subset of An. The clone F ~ F(A) is called a local clone on
A if it coincides with F . We also have local relational clones; all we
need to define them is to add to the definition of relational clone the
requirement of closure under arbitrary intersections.
Theorem 2.3. ([13]) Let A be an arbitrary set. A subset F ~ F(A)

is Galois-closed iff it is a local clone on A. A subset R ~ R(A) zs
Galois-closed iff it is a local relational clone on A.
3 Primality and its generalizations
Let A = (A; F) be an algebra with universe A and set of fundamental

operations F . Then F{! is the collection of all subuniverses of An ,
n = 1, 2, ... , that is,
It is well known that a function on A is a local term function of A iff

it preserves all subuniverses of all finite direct powers of A . Hence,
F(!cp = LocTerm A
is the clone of all local term functions of A. An algebra A is called

locally primal if LocTerm A= F(A) , that is, if all functions on A are
local term functions of A. Since F(A){! = V(A) , we conclude that an
algebra A is locally primal iff its finite direct powers have only trivial
subuniverses. Here we call a subuniverse R of An trivial if it is a direct
product of diagonals .6.~ , m :::; n. Since the congruence relations and
the graphs of endomorphisms of A are subuniverses of A 2 , the locally
primal algebras must be simple and have as their only endomorphism
the identity map. Consequently, if we wish to generalize the notion of
264 K. Kaarli
local primality to nonsimple algebras or to algebras having nontrivial

endomorphisms, we must require that the local term functions are
exactly the functions which are compatible with all congruences or
the functions which permute with all endomorphisms. Using our basic
Galois connection the corresponding notions (and some other related
ones) can be defined as follows.
Definition 3.1. An algebra A is called:
• locally congruence primal if (Con A)'P = LocTerm A;
• locally endomorphism primal if (EndA)'P = LocTermA;

• locally automorphism primal if (Aut A) 'P = LocTerm A;
• locally quasiprimal if (Partlso A) 'P = LocTerm A.
The corresponding non-local notions are defined by replacing the op-
erator LocTerm with Term in the above definition. The reader may
wonder why our preference is for the local versions of primalities. The
answer is the following. In the literature the primal algebras are usu-
ally finite by definition. Then the primality of A = (A; F) is equiva-
lent to pe = 'D(A). In the case of an infinite A the latter property,
however, is equivalent to local primality, not to primality. So, in view
of our basic Galois connection between finitary relations and finitary
functions local primality is a natural generalization of primality to
infinite algebras.
In the older literature subalgebra primal and congruence primal alge-
bras were called semiprimal and hemiprimal, respectively. The term
endoprimal is usually used instead of endomorphism primal. The sym-
bol Partlso A denotes the local clone of partial isomorphisms of A ,
that is, the collection of all isomorphisms between subalgebras of A.
Quasiprimality has turned out to be the most successful generalization
of primality. Note that we have not presented here the complete list
of primalities one can find in the literature. It should be mentioned
that some of them, for example paraprimality, cannot be defined via
our basic Galois connection.
The usual way to show that an algebra A is, say, locally congruence
primal is to take a function f E (Con A) 'P and to prove that f is a
local term function . An alternative way to do this is suggested by the
fundamental theorem on Galois connections.
Theorem 3.2. An algebra A is locally congruence pr-imal iff Con A

generates the local r-elational clone U{Sub An In = 1, 2, . . . }. Similar·
statements hold for the other- local primalities listed in Definition 3.1.
Pr-oof We give the proof only for congruence primality. If A is locally
congruence primal then (Con A) 'P = LocTerm A, implying that
(ConA)'Pe = U{SubAn In= 1, 2, ... }.
This proves the necessity of our condition. The sufficiency follows by
applying the operator r.p to the equality (1) .
We shall denote by A + the algebra obtained from A by adding to
F all elements of A as nullary operations on A. Formally we write
A + = (A; F U A). Now the subuniverses of (A +)n are precisely the
subuniverses of An which contain the diagonal b.;';. These are called
diagonal subuniverses of An. Hence we may write
(F U A)l! = U{Subdiag An In= 1, 2, ... }
= U{Sub (A +t In= 1, 2, ... }.

The (local) term functions of A+ are the (local) polynomial functions
of A. Thus
(F U A)ecp = LocTermA + = LocPolA. (1)
An algebra A is called locally functionally complete if LocPol A
F(A), that is, if all functions on A are local polynomials of A. It
follows from (1) that A is locally functionally complete iff A + is locally
primal iff (F u A)e = V(A) .
Since polynomial functions do not preserve all subuniverses and do
not permute with all (partial) endomorphisms, there are not so many
generalizations of local functional completeness. In fact there is only
one such generalization which makes sense for all varieties of algebras.
It is based on the fact that congruences are preserved by polynomial
functions.
Definition 3.3. An algebra A is called locally affine complete if Con A
= LocPolA.
Analogously with Theorem 3.2 we have the following theorem.
Theorem 3.4. An algebra A is locally affine complete iff Con A gen-
erates the local relational clone U{Sub (A+)n In= 1, 2, ... }.
266 K . Kaarli
4 Varieties with near-unanimity term

Theorem 2.2 shows, in particular, that the notion of relational clone
is an adequate tool for the formalization of the collection of all sub-
universes of finite powers of a finite algebra. Due to Theorem 2.3, for
arbitrary algebras a similar role is played by local relational clones. It
is natural to ask if it is possible to formalize the collection of subuni-
verses of An, for a given n . This would correspond to the restriction
of the basic Galois connection to the set of n-ary relations. An answer
can be derived from general results of [13]. This answer, however,
seems to be too complicated, in general. Here we present the solution
recently obtained by J. W. Snow for the important special case where
the clone of functions contains a near-unanimity function.
Recall that a function f E Fn(A), n ~ 3, is called a near-unanimity
function if the equality f(a 1 , ... , an) =a holds whenever all but per-
haps only one of the elements a 1 , ... , an E A are equal to a. Let V be
a variety of algebras of type T. We say that a term t of type T is a
near-unanimity term for V if it induces a near-unanimity function on
every A E V.
A ternary near-unanimity term is called a majority term. If m is a
majority term for a variety V then the identities
m(x, x, y) ~ m(x, y, x) ~ m(y, x, x) ~ x
hold in V. Every variety with lattice reduct has a majority term

(x 1\ y) V (y 1\ z) V (z 1\ x).
Varieties with near-unanimity term have many interesting features
(see [1] and [10]). For us the most important is the following.
Theorem 4.1. Let V be a variety with an (n+ 1)-ary near-unanimity
term and let A E V. Then (SubAn)'P = LocTermA.
Hence, in the case of a variety with an (n + 1)-ary near-unanimity
term, if we want to check whether a given function on an algebra
A is a local term function, it suffices to test it only on subuniverses
of An. This made it even more important to try to characterize in
more or less feasible terms of An the systems of subsets of An which
can appear as systems of subuniverses of some algebra A = (A ; F)
where F contains an (n + 1)-ary near-unanimity function. For the
case n = 2 and A finite this problem was posed by C. Bergman [2].
Since the collection of sub universes of A 2 is closed with respect to the

operations of taking converse v' intersection n and relation product o,
C. Bergman asked: is it true that for every R ~ R 2 (A) containing
llt and A 2 , and closed with respect to the operations v, n and o,
there exists a set F ~ F(A) containing a majority function and such
that F(! n R 2 (A) = R? The answer to a more general question was
given by J. W. Snow in his Ph.D. thesis [16]. In order to formulate his
result we define some operations on the set Rn(A). First, for every
mapping a : {1, ... , n} ---+ {1 , ... , n} let Pa : An ---+ An be defined by
the formula:
Pa(al, ... , an)= (aa(l) , . .. , aa(n)).
Then both Pa and pa- l (taking of full preimage) can be considered
as unary operations on Rn(A) . Second, we define an n-ary opera-
tion Cn which generalizes the binary relation product operation: given
r 1 , ... , rn E Rn(A) the relation cn(r 1 , . . • , rn) is the set of all n-tuples
(a 1 , ... , an) E An for which there exists an+l E A such that
(a1, ... , an-l , an+I) E r1 ,

(a1, ... , an- 2, an, an+I) E r2 , ... , (a2 , . .. , an , an+I) Ern.
Theorem 4.2. ([16], Theorem 2.10) Let A be a finite set. Assume that
a subset R ~ Rn(A) contains tl~ and An, and is closed with respect to
the operations n, Cn, Pa and pa- l for every mapping a. If R ep contains
an (n + 1)-ary near-unanimity function , then R ep(! n Rn(A) = R.
In case n = 2 Theorem 4.2 answers Bergman's question. For any set A

we consider the algebra R2 (A) = (R2 (A); 0, 1 ,~ n, o) where the nullary
operation symbols 0 and 1 are interpreted as llt and A2 , respectively.
It is easy to see that in case n = 2 all operations Pa and pa-l can be
expressed via the operations of R2 (A). Consequently Theorem 4.2 has
the following corollary.
Corollary 4.3. Suppose that A is a finite set andRE R 2 (A). If Rep

contains a majority function , then R ep(! n R 2(A) is the subuniverse of
R2 (A) generated by R.
In the sequel we indicate some applications of the latter result. It
seems very likely that these results of Snow have infinite analogs. We
conjecture that Corollary 4.3 remains valid in the case of an arbitrary
268 K. Kaarli
set A if we only require that R is closed with respect to arbitrary

intersections.
For later use we introduce the following notation. For an algebra A,
the subalgebra of R2 (A) consisting of all diagonal subuniverses of A 2
will be denoted by 5 2 (A).
5 Order affine complete lattices
A lattice L is called order functionally complete if {.::;}'P = PolL and

order affine complete if (Con L U {::;} )'P = PolL. One gets the local
version of this notion by replacing the operator Pol with LocPol in the
above definition. The study of order functionally complete lattices was
initiated by D. Schweigert [15]. Later finite order functionally com-
plete and order affine complete lattices were characterized by R. Wille
([17], [18]), and M. Kindermann showed that tolerances and the order
relation of a finite lattice determine its polynomial functions and that
a finite lattice L is order functionally complete iff TolL = { .6.~, L 2 }
([11]). Recall that a tolerance of an algebra A is a reflexive (in other
words, diagonal) and symmetric subuniverse of A 2 .
Kindermann's original proof implicitly uses a Galois theoretic ap-
proach. He shows that every diagonal subuniverse of L 2 , where L
is a finite lattice, can be obtained from tolerances and the order rela-
tions ::; and ~ of L using two binary operations: n and +, the latter
denoting formation of the sublattice generated by two given sublat-
tices. In view of Snow's result (Corollary 4.3), the algebra S2 (L) is
generated by tolerances and the order relation ::; of L, for any finite
lattice L.
In [10] we extended the basic results about finite order functionally
complete and order affine complete lattices to the infinite case. In
particular, we proved the following analogs of Kindermann's results.
It is possible that the results were known earlier but we are not aware
of the corresponding references.
Theorem 5.1. ([10], Lemma 5.3.18) For an arbitrary lattice L,
(TolL U {::;} )'P = LocPol L .
Theorem 5.2. ([10], Theorem 5.3.40) A lattice L is locally order func-

tionally complete iff .6.~ and L 2 are its only tolerances.
Obviously the sufficiency part of Theorem 5.2 follows from Theo-

rem 5.1. Indeed, if L has only trivial tolerances, then
LocPolL =(TolL U {:::;})cp = {~} cp,
saying that the local polynomial functions of L are precisely its order
preserving functions.
In the finite case the necessity part of Theorem 5.2 can be easily
derived from Theorem 5.1 and Corollary 4.3. Suppose that L is a
finite order functionally complete lattice. Then {~} cp = PolL which
implies
Now Corollary 4.3 implies that the order relation ~ generates the
algebra S2 (L). However, it is easy to see that the subalgebra of R2 (L)
generated by ~ has the universe {~ , ;::::: , .6. 2 , L 2 }. In particular, L has
no nontrivial tolerance relation.
A similar way of reasoning works also in the case of order affine com-
pleteness defined by the equality (Con LU{ ~} )'P = PolL . If we assume
that L is finite and apply to this equality the mapping (! then Corol-
lary 4.3 implies that L is order affine complete iff the algebra 5 2 (L)
is generated by the congruences and the order relation of L. On the
other hand it follows from Theorem 5.1 that the algebra S2 (L) is gen-
erated by the tolerances and the order relation of L. Consequently we
have the following result.
Theorem 5.3. A finite lattice L is order affine complete iff all toler-
ances of L are contained in the subalgebra of S2 (L) generated by the
congruences and th e order relation of L .
This is quite close to R. Wille's [18] characterization of finite order

affine complete lattices. Wille's original result is given in terms of
decreasing V-endomorphisms but it can be easily translated into the
language of tolerances. Given two tolerance relations S and T of a
lattice L, we introduce their symmetrized relation product
S *T = { (a, c) I (3b E L) (a 1\ c, b) E S , (b, a V c) E T} .

Then Wille's theorem can be stated as follows .
270 K. Kaarli
Theorem 5.4. A finite lattice L is order affine complete iff every

tolerance of L can be represented as a meet of symmetrized relation
products of congruences of L.
An infinite version of this theorem is presented in [10].
6 Local polynomial functions of algebras with

distributive lattice reduct
G. Gratzer ([7]) proved that all Boolean algebras are affine complete
and then characterized affine complete bounded distributive lattices
([8]). He proved, in particular, that every bounded distributive lattice
L is order affine complete. This result was generalized to arbitrary
distributive lattices by D. Dorninger and D. Eigenthaler ([6]). In the
language of our Galois connection their result reads as follows.
Theorem 6.1. Let L be a distributive lattice. Then
(Con L U {:S} )"' = LocPol L. (2)
In other words, every distributive lattice is locally order affine com-

plete.
This result is based on the fact that every distributive lattice L is

a subdirect power of the 2-element lattice D 2 and every congruence
preserving function f E Fn (L) can be represented via its "coordinate
functions": f = (fi)iEI where every fi is an n-ary function on the i'th
subdirect factor Li ':::::' D 2 of L. Clearly f is order preserving iff so are
all the k Now a polynomial p which interpolates f at {aI, ... , am} ~
Ln can be found by constructing, in a uniform way, the polynomials
Pi E PolD 2 which interpolate fi at {a}, ... , ar} ~ {0, 1}n. Note
that the existence of the interpolating polynomials fi follows from
Theorem 5.2. The important point is that these polynomials can be
constructed in a uniform way.
Theorem 6.1 stimulated the search for similar characterizations of local
polynomial functions for other varieties which are close to the variety
of distributive lattices. We describe here the result of this search in
the case of Kleene algebras. The work was done jointly by M. Haviar,
M. PlosCica, and the author of the present paper (see [9] and also [10]).
Recall that a Kleene algebra is a bounded distributive lattice equipped

with an additional unary operation 1 , so that the following identities
are satisfied:
(xl\y) 1 ~X 1 Vy', (xvy)'~x'l\y 1 , 0 1 ~1, 1 1 ~0 ,
x" ~ x, (x 1\ x 1 ) V y V y 1 ~ y V y 1 •
It is well known that the variety of Kleene algebras is generated by
the 3-element Kleene chain K 3 . (Obviously there is only one way to
define a unary operation 1 on the 3-element lattice so that the result is a
Kleene algebra.) Therefore every Kleene algebra is a subdirect product
of copies of K 3 and its 2-element subalgebra K 2 with universe {0, 1},
which in fact is the 2-element Boolean algebra B 2 . Trying to find
an analog of Theorem 6.1, we first studied the diagonal subuniverses
of K 3 2 . It turned out that the algebra S2 (K 3 ) is generated by the
so-called uncertainty order
~ = { (0, 0), (a , a), (1, 1) , (0, a) , (1 , a)},
where a is the third element of K 3 . Hence we have the equality

{~} 'P =Pol K 3 . As the next step, we showed that using the subdirect
decomposition, there is a natural way to extend the uncertainty order
to an arbitrary Kleene algebra. As in the case of distributive lattices,
it was possible to construct the interpolating polynomials for subdirect
factors in a uniform way. The final result was the following.
Theorem 6.2. Let K be a Kleene algebra. Then
(Con L U {~} )'P = LocPol K. (3)
Using this description of local polynomial functions it was not difficult

to characterize locally affine complete Kleene algebras. The necessary
and sufficient condition we obtained is given in lattice theoretic terms
but its actual meaning is that all congruence preserving functions must
also preserve the uncertainty order (formally, that ~ E (Con L) 'P!?).
Thus the formula (3) gives a hint for how to find a characterization of
locally affine complete Kleene algebras.
Kleene algebras form a small subvariety of the variety of Ockham al-
gebras [3] , which can be defined as bounded distributive lattices with
272 K . Kaarli
an additional unary operation ' satisfying the identities of Kleene alge-

bras, except the last two. In [10) we found an analog of Theorems 6.1
and 6.2 for Stone algebras. The latter form a subvariety of the variety
of Ockham algebras defined by the additional identities
x 1\ x' ~ 0, x' V x" ~ 1.
It turned out that in the case of Stone algebras the role of ::; in (2)
and ~ in (3) is played by the relation
::;a= {(x,y) lx' = y' andx::; y}
which we called the Glivenko order. Recently my doctoral student

V. Kuchmei observed that there is a common definition of the un-
certainty order of Kleene algebras and the Glivenko order of Stone
algebras. Moreover, he found analogs of Theorem 6.2 for some other
varieties of Ockham algebras, for example for de Morgan algebras.
However, he also noticed that a similar characterization of local poly-
nomial functions cannot be extended to all Ockham algebras. These
results are not published yet.
7 Endoprimal algebras
An algebra A is called endoprimal if (End A) 'P = Term A . L. Marki

and R. Poschel ([12]) proved that a distributive lattice is endoprimal iff
it is not relatively complemented. B. Davey ([4]) observed that endo-
primality is closely related to the theory of natural dualities. Namely,
he noticed that every finite algebra which is dualisable by its endomor-
phism monoid, is endoprimal, too. Using the machinery of the theory
of natural dualities, he and J. G. Pitkethly ([5]) produced a number of
endoprimal algebras in various classes of universal algebras. Recently,
in joint work with H. Priestley on endoprimality of Kleene algebras,
we discovered an alternative direct approach for describing endoprimal
algebras. The results are not published yet, so we only briefly outline
the main ideas of our approach. It is based on the following theorem
whose proof is not very complicated.
Theorem 7.1. Assume that the algebra A is subdirect in n iEI Ai and

that there exists an algebra M such that the following conditions hold:
• Ai is a subalgebra of M for each i E I;
• A has a subalgebra S isomorphic toM;
• there exists p E I such that Av = M .
Then, given any function f E (End A )h, there exists a function g E

Fn(M) such that f = (Ji)iEI where every fi is the restriction of g to
Ai and the equality
holds for every homomorphism a : Ai --1 Aj.
An immediate corollary of Theorem 7.1 is that if the three conditions

are satisfied and the algebra M is endoprimal, then so is A. In par-
ticular, every Boolean algebra B is endoprimal because B contains a
2-element subalgebra (unless it is trivial) and can be represented as
a subdirect product of 2-element algebras. However, the 2-element
Boolean algebra B 2 is primal and then obviously endoprimal.
In general, the importance of Theorem 7.1 is that it reduces the ques-
tion of endoprimality of A to a question about functions on a smaller
algebra M. It is especially useful if we have a majority term and M
is finite, because then every function on M can be effectively tested
with respect to being a term function .
The following theorem is an almost direct consequence of Theorem 7 .1.
Theorem 7.2. Assume that M , A, and Ai, i E I, satisfy the con-

ditions of Theorem 7.1. Also suppose that M is finite and generates
a variety which has a majority term. Then A is endoprimal if the
algebra 52 (M) is generated by the set
Here Hom(Ai, Aj) denotes the set of all graphs of homomorphisms

from Ai to Aj and Aij is the canonical projection of A in A x Aj.
Theorem 7.2 shows, in effect, that if the assumptions of the theorem
are satisfied then A is endoprimal if it has enough homomorphic im-
ages. For example, it easily implies the sufficiency part of Marki's
and Poschel's theorem about endoprimality of distributive lattices.
Indeed, if a distributive lattice L is not relatively complemented then
274 K. Kaarli
it has a 3-element homomorphic image. This allows us to embed L

subdirectly into D 2 1 so that one of the 2-fold projections Lij is a 3-
element lattice. Clearly the 3-element sub universe of D 2 2 generates
the algebra S2 (D2).
One can use similar arguments for finding sufficient conditions for
endoprimality in the case of other algebras with majority term, for
example in the case of Stone and Kleene algebras. It is possible to
show, using Corollary 4.3 and adding one more natural assumption
on the subdirect decomposition, that in the finite case the condition
of Theorem 7.2 is necessary, too. However, so far we have not found
a good application of this result because all known proofs of non-
endoprimality are based on explicit construction of a function f E
(End A)'~' \Term A which remains valid in case of infinite algebras.
For example, if L is a relatively complemented distributive lattice
then we get a suitable function f by defining f(x, y, z) = u iff u is the
complement of y in the interval [x 1\ y, y V z].
Acknowledgment I wish to thank Reinhard Peschel who helped me

to find necessary references and kindly let me use his unpublished
manuscript.
References
[1] Baker, K. A., Pixley, A. F., Polynomial interpolation and the

Chinese remainder theorem for algebraic systems, Math. Z. 143
(1975), 165- 174.
[2] Bergman, C., Categorical equivalence of algebras with a majority

term, Algebra Universalis 40 (1998), 149- 175.
[3] Blyth, T., Varlet, J. C. , Ockham Algebras, Oxford Univ. Press,

1994.
[4] Davey, B. A., Dualisability in general and endodualisability in

par·ticular, Logic and Algebra. Lecture Notes in Pure and Applied
Mathematics 180, Marcel Dekker, New York, 1996, 437-455.
[5] Davey, B. A., Pitkethly, J. G. , Endoprimal algebras, Algebra Uni-

versalis 38 (1997), 266- 288.
[6] Dorninger, D., Eigenthaler, D. , On compatible and order pre-

serving functions on lattices, Banach Center Publ. 9, Semester
1978, Universal Algebra and Applications, 1982, 97- 104.
[7] Gratzer, G., On Boolean functions (Notes on lattice theory II) ,

Rev. Math. Pures Appl. (Bucarest) 7 (1962), 693- 697.
[8] Gratzer, G., Boolean functions on distributive lattices, Acta

Math. Acad. Sci. Hung. 15 (1964) , 195-201.
[9] Haviar, M., Kaarli , K., PlosCica, M. , Affine completeness of

Kleene algebras, Algebra Universalis 37 (1997) , 477- 490.
[10] Kaarli, K. , Pixley, A. , Polynomial completeness in algebraic sys-

tems. Chapman & Hall/CRC, Boca Raton - London - New York
- Washington, D.C. , 2001.
[11] Kindermann , M. , Uber die Aquivalenz von Ordnungspolynomvoll-

stiindigkeit und Toleranzeinfachheit endlicher Verbiinde, Contri-
butions to General Algebra (Proc. Klagenfurt Conf. 1978) , Verlag
J. Heyn, Klagenfurt, 1979, 145- 149.
[12] Marki, L., Poschel, R., Endoprimal distributive lattices, Algebra

Universalis 30 (1993) , 272- 274.
[13] Poschel, R. , Concrete representation of algebraic structures and

general Galois theory, Contributions to general algebra. Proc.
Klagenfurt Conf., May 25- 28, 1978. Verlag Joh. Heyn, Klagen-
furt, 1979, 249- 272.
[14] Poschel, R. , Kaluzhnin , L. A., Functionen- und Relationenalge-

bren, VEB Deutscher Verlag der Wissenschaften, Berlin 1979.
[15] Schweigert, D., Uber endliche, ordnungspolynomvollstiindige Ver-

biinde, Monatsh. Math. 78 (1974) , 68- 76.
[16] Snow, J. W., R elations on Algebras and Vari eties as Categories,

Ph. D. Dissertation, Vanderbilt Univ. , Nashville, Tennessee.
[17] Wille, R. , Eine Characterisierung endlicher, ordnungspolynom-

vollstiindiger Verbiinde, Arch. Math. (Basel) 28 (1977) , 557- 560.
276 K. Kaarli
[18] Wille, R., Uber endliche, ordnungsaffinvollstiindige Verbiinde,

Math. Z. 155 (1977) , 103-107.
Author's address:
Institute of Pure Mathematics
University of Tartu
50090 Tartu
Estonia
e-mail: george_janelidze@hotmail.com
Q-lndependence and Weak
Automorphisms
K. Glazek, 5. Niwczyk
Abstract
The aim of this survey is a presentation of two kinds of Galois connections,
which appear in Universal Algebra but are not well known. They are related
to the notions of general independence and of weak automorphism. There
are still interesting open problems from these areas.
AMS Mathematics Subject Classification: 06A15, 08B20, 08A40, 08A35,
08A30, 03B50, 20B 25, 20B21.
Key words: General Galois connection, Formal context, Marczewski inde-
pendence, Independences with respect to a family of mappings, Weak endo-
morphism, Weak automorphism, G-clone.
1 Introduction
Classical Galois theory for fields establishes the correspondence be-

tween extensions of fields and field automorphisms. G . Birkhoff, 0.
Ore, J. Riguet (see [1], [38], [43]) and others have investigated suitable
general correspondences between some ordered sets. Such correspon-
dences are extensively used in formal concept analysis (see [7]) .
Moreover, there are several Galois connections (Galois corresponden-
ces) used in Universal Algebra. For instance, the Galois correspon-
dence between identities and varieties (of the same type) is well-
known (see [1], 3rd ed., Chapter VI) . A more general Galois correspon-
dence between hyperidentities and hypervarieties (of the same type)
has recently been studied by many authors (see [4]) . On the other
hand, the Galois connection Inv-Pol, a connection between functions
(operations) and relations, was discovered by M. Krasner and A.V.
277
278 K. Gtazek, St. Niwczyk
Kuzniecov, and intensively explored by many researchers, e.g., by L.A.

Kaloujnine (Kaluznin) , R. Poschel, I. Rosenberg and others (see [2],
[24], [28], [42], [41], [44], [45], [46], [53]) .
In this paper, we want to describe two other Galois correspondences
which are less well known, namely the Galois connections related to
the notions of independence (with respect to a set of mappings) and
of weak automorphism.
Some interesting problems from these areas are still open. Among
others, it would be interesting to know a description of the Galois-
closed sets with respect to the Galois connection related to the notion
of weak automorphism. This problem is still open and is equivalent
to the concrete characterization problem of weak automorphisms (see
Section 4). In the last section we also propose a general setting for
the connections studied here.
Investigations of the general notion of independence and of weak ho-

momorphisms of general algebras were initiated by E. Marczewski, his
students and co-workers.
2 Preliminaries
Let A be nonempty set and ((}n) (A) , JR(n) (A) the set of all n-ary
operations and the set of all n-ary relations of A , respectively (for
00
any positive integer n) . Let us put O(A) = U ((}n) (A) and JR(A)
n=l
00
U JR(n) (A). The set of all projections will be denoted by JE(A). A

n=l
subset <C(A) ~ O(A) is called a clone (in the sense of P. Hall) over
the set A if it contains the set of all projections and is closed under
compositions of operations (see [3]) . If IF ~ O(A), then (IF) denotes
the clone generated by IF (i.e. the smallest clone containing the set IF) .
The set of all clones over the set A will be denoted by £(A); this set
forms a lattice, under inclusion. For a given algebra 21 = (A ; IF) the set
of term operations of this algebra is the clone (IF) , which is denoted by
'I(21). Two algebras 21 = (A ; IF) and~ = (A ; <G) are term equivalent if
they have the same sets of term operations, i.e. (IF) = (<G). Similarly,
for a subset B ~ A, we denote by (B)'l! (or just (B)) the subalgebra
of the algebra 21 = (A ;IF) generated by B.
Q-lndependence and Weak Automorphisms 279
The first Galois connection we shall describe uses the concept of in-
dependence. Let 2l = (A ; JF) be an algebra with a carrier A and a
set lF of fundamental operations. A nonempty set I ~ A is called
M -independent (or independent in the sense of Marczewski; see [35],
[36] and [9]) if for any finite system a1, ... , an E I of different elements
and for any pair j, g E T(n) (2l) of n-ary term operations the equality
j(a1, ... ,an) = g(a 1, ... ,an) implies f =gin A. We also assume (unlike
the assumption in [35]) that the empty set is M-independent. This
notion is a special case of the notion of independence with respect to
a set of mappings, to be considered in Section 3.
Now we introduce the necessary background for the second Galois con-
nection we shall consider, using weak automorphisms. For an algebra
2l = (A; JF) letT be a mapping from A into A. Consider the following
two conditions:
(a) For every f E 'I'(2l) (say: n-ary) there exists g E 'I'(2l) such that
(*) (\fa1 , ... , an E A) (T (! (a1, ... , an)) = g (T (a1), ... , T (an))).
One can verify that if T is " onto" , then g is uniquely determined (see
[19]).
(b) For every g E 'I'(2l) (say: n-ary) there exists f E 'I'(2l) such that
the equality (*) holds.
One can observe that if T is " 1-1" , then f is uniquely determined.

The mapping Tis said to be a semi-weak endomor-phism (see [11], [31])
of the algebra 2l if condition (a) holds. A semi-weak endomorphism
is a special case of an T-mor-phism in the sense of A. Goetz [22], and
is related to the notion of anamor-phism in the sense of W. S. Hatcher
and S Whitney [27].
If the mapping T satisfies condition (b), then T is a special case of
an l-mor-phism (see [22]) and is related to the notion of catamor-phism
(see [27]). Moreover, if T fulfills both conditions (a) and (b), then Tis
said to be a weak endomor-phism of the algebra 2l (see [19]).
Every permutation (bijection) CJ of the set A induces a bijection a :
f H g of O(A), where the operation g is defined by f and CJ in the
following way:
280 K . Gtazek, St. Niwczyk
We denote by SA the set of all permutations (bijections) of A . If, for

the permutation a E SA and the algebra 2t =(A; JF), the equality
a('I'(2t)) = 'I'(2t)
holds, then a is said to be a weak automorphism of the algebra 2t.
It is worth adding, that - in the definition of a weak automorphism
- it is not enough to assume the inclusion a('I'(2t)) ~ 'I'(2t) (in this
case a is a semi-weak endomorphism with injective ol We denote by
Aut (2t) and W Aut (2t) the groups of all automorphisms and all weak
automorphisms of the algebra 2t, respectively. One can verify that
Aut (2t) is a normal subgroup of the group W Aut (2t) (see [48], [39]).
3 The Galois connection related to Q-independence

In all further considerations of this section, we fix an algebra 2t =
(A; JF).
We will denote by M(A) (or by M for short) the set of all mappings
p : T ----+ A from all nonempty subsets T ~ A to the carrier A , i.e.
M(A) = {p: pEAT, T ~A}, and by H(A) (or just H) the set of all
mappings p : T ----+ A (for T ~ A), which possess an extension to a
homomorphism p: (T) ----+A (PIT= p) .
E. Marczewski ([36], p. 137 and [37]) observed that several notions
of independence weaker than M -independence fall under a common
scheme, namely one can express them in the language of mappings
and their extensions to homomorphisms. He proposed the following
definition:
Definition 3.1. A nonempty set T ~ A is called independent with

respect to the set Q ~ M(A) , or Q-independent for short, if Q nAT ~
H(A).
The system of all Q- independent sets will be denoted by I nd21. (Q) ( or

I nd( Q) for the fixed algebra 2t). A subset which is not Q-independent
is called Q-dependent.
A number of versions of independence can be viewed as examples
of Q-independence. If we put Q = M , we obtain M-independence,
i.e. independence in the sense of Marczewski. Taking Q = S =
{ p Ip E (T)T, T ~ A} gives the S-independence introduced by J.
Schmidt in [47] . A similar claim holds for So-independence, where

S0 = {pIpE Tr, T ~A} , introduced by S. Swierczkowski in [49].
Moreover two other notions of independence were defined and in-
vestigated in [9] . These are A1-independence, defined by putting
Q = A1 = {fir I f E T(l) , T ~A}, and R-independence, obtained
by taking Q = R, where R ~ M is the subset of all injective map-
pmgs.
A mapping p E AT, where T ~ A, is called diminishing if for every
j, g E T(l) and for each a E T the equality f(a) = g(a) implies
f(p(a)) = g(p(a)). If Q = G is the set of all diminishing mappings,
then Q-independence becomes the G-independence introduced by G.
Gratzer in [25] (and modified in [37]).
It is worthwhile to remark that Q-independence satisfies the following
important properties, needed for a Galois connection:
(1) ('I/Q1 , Q2 ~ M) [Ql <;;;; Q2:::? Ind(Q2) ~ Ind(QI)],
(2) ('1/Q ~ M) [Ind(M) ~ Ind(Q) ~ Ind(H) = 2AJ ,
(3) for an arbitrary family Qi ~ M ,i E I

n Ind(Qi) =
iE/
Ind(U Qi),
iE/
U Ind(Qi)
iE/
~ Ind(n Qi)·
iEI
These properties suggest that we can consider a special Galois con-

nection
M :2 Q f---t Ind(Q) ~ 2A and 2A :2 J f---t Q 1 ~ 111[, where Q 1 is the
maximal (with respect to inclusion) set of mappings with Ind(Q 1 ) =
J. We develop this Galois connection as follows.
For an algebra 2{ we will construct a formal context in the sense of R.
Wille (see [52] and [7]), in which the objects of t he context are subsets
of the carrier of 2{ and the attributes are mappings (see [20]) .
Let 2A be the power set of A . We define a relation e ~ 2A x M in the
following way:
For a nonempty subset X E 2A and a mapping p EM one declares
We also assume that
('1/p E M) [0 e p].
Now, the dually adjoint operators forming our Galois connection are
defined as follows:
M 2Q f-t I( Q) = {X ~ A I (Vp E Q) [X [J p]} '
2A 2 J f-t Q(J) = {p EM I (VX E J) [X [J p]}.
Let us define Galois closure operators by the equalities:
v(J) = i(Q(J)),
6(Q) = Q(i(Q)).
In this way, we have constructed the formal context J = (2A , M, LJ) of

subsets and mappings. We will call J a context of Q-independences.
For Q ~ M, we have Ind(Q) = I(Q).
One can consider, for an arbitrary set Q ~ M, a maximal (with respect
to inclusion) set Q such that Ind(Q) = Ind(Q) (see (9]). It was proved
that:
(4) for a family J of subsets of the carrier A of 2l there exists a
set Q of mappings such that
J = Ind(Q) iff Ind(M) ~ J.
It is known (also from (9]) that the set D(2l) of all maximal sets
of mappings with set-theoretic join and meet operations, and with
complementation defined by
(a) Q' = (M \ Q) U H ,
is a complete atomic Boolean algebra.
In a context J these families (Q and J) are determined by the closure
operators 6 and v.
Q=6(Q) ¢:} Q=Q ,

J = V(J) {:} Ind(M) ~ J.
In further considerations, a set Q = 6 (Q) of mappings is said to be

closed instead of maximal. Similarly, J = V (J) is called a closed
system of subsets of A.
Both closure operators V and 6 are additive, meaning that
v(UAi) = U v(Ai), where Ai ~ 2A,

i EI iEl
~(uQi) = U.-6.(Qi), where Qi ~ M.

'iEI iEI
For an algebra 2t, a set X ~ A is called C-independent if, for any

a E X, a tf. (X \ {a}) . The system of all C -independent sets of the
algebra 2t will be denoted by C-I nd2r (or C- I nd if 2t is fixed). It is
easy to see that Ind(M) ~ C-Ind. Therefore, by (4), there exist a set
Q ~ M of mappings such that C-Ind = Ind(Q) (see [9]).
Similarly to the definition (a) above for the complementation of dosed
sets of mappings, we define a complementation on dosed systems of
(independent) subsets of A as follows:
(/3) J' := (P \ J) U V(0) = (P \ J) U Ind(M) ,
for any dosed set J ~ 2A = P.
If a subset T is not M-independent and is Q-independent for some
Q ~ M , then Tis (~(Q))'-dependent.
There is a simple relationship between complementations in lattices
of closed sets of subsets of A and dosed sets of mappings, namely:
(5) for any system Q ~ M , we have I((~(Q))') = (I(Q))'.
We know (see [9]) that:
(6) ifthe set I(Q) is hereditary 1 , then the set i ((~(Q))') is of finite
character2 .
From statement (5), one can easily conclude that the converse of im-
plication (6) is not true.
The lattice J(2t) of all v-closed sets, with set-theoretic join and meet
operations and complementation defined by (/3), is dually isomorphic
to D(2t).
For any subset T ~ A, T tf. Ind(M), we have:
(7) the set ATuH is an atom of,Q(2t) and {T}Uind(M) is an atom

of J(2t),
1 A system J of sets is hereditary if from X E J follows that every subset Y of
X belongs to J.
2 A system J of sets is of finite character if a set T belongs to J whenever any
finite subset F ofT is in J.

(8) the sets (M \AT) U Hand P \ {T} are co-atoms of the lattices
J(Qt) and Q(Qt) respectively.
Let us recall another notion of independence, the so-called t-indepen-

dence, introduced by J . Plonka (see (40] and [23]). For an algebra
Qt = (A; JF) a set X ~ A is called t-independent if for any finite system
of different elements a 1 , ... , an E X and for any n-ary term operation
f E 'I(n) (Qt) we have
We denote by t-Ind<J. (or t-Ind if Qt is fixed) the system of all t-

independent sets of the algebra m.
It is easy to show (see [23]) that:
(9) Ind(M) ~ t-Ind ~ C-Ind;
(10) X is t-independent iff f (a 1 , . .. ,an ) tJ. X for any non-trivial term

operation f E 'I(n) (Qt) and different elements a 1 , . . . , an E X.
One can construct (in a rather artificial way, see [9]) a set Q ~ M of
mappings such that
(!) t-Ind<J. = Ind<J.(Q) ,
for any algebra Qt. The following problem (see [10], Problem 7.5) seems
to be interesting and still open:
Problem Does there exist a uniform method for constructing a set Q

(valid for all algebras Qt) such that equality (1) holds?
4 The Galois connection related to weak

automorphisms
In this section we want to describe another Galois connection, this

time a connection between bijections of a set A and clones over A.
Let <J E SA and lffi E .C, where .C = .C( A) is the set of all clones over A
(see Preliminaries) . Then we define a relation 8 between elements of
SA and .C(A) as follows:
where a is a mapping induced by a (or in other words a is a weak

automorphism of the algebra (A , lB\) ; see Preliminaries). We will say
that a preserves the clone lB\.
The relation 8 determines a Galois correspondence (see (13], (17], (18])
between subsets G ~ SA and families F of clones over A, given by the
two mappings:
G f-t F(G) = {lB\ E £(A) I (Va E G)[a e lB\]}'
Note some simple properties of this mappings (see (17], (18]) , namely:
(i) G({JE}) = G({0°)}) = G({O}) =SA, where 0°) is a set of all

constant operations of the set A ,
(ii) lE, (()) E F(G),
(iii) G({lB\}) = WAut((A,lB\)) ,
(iv) if G ~ SA (i.e. G is a subgroup of SA) and lB\ E F(G), then

G ~ WAut((A,lB\)) ,
(v) if lB\i E F(G) (for i E J), then / U lB\i) E F(G) and more
\tEl
generally F( G) is a sublattice of the lattice of all clones over A,
(vi) F(G) = F((G)) = n F({a}) , where (G) is a subgroup of SA

u EG
generated by the set G ~ SA,
(vii) G(F) = n G({lB\}) ~ sA.

IBEF
Define the operators V : 25 A ---+ 25 A and 6. : 2.c ---+ 2.c in the following
way:
v(G) = G(F(G)),
6.(F) = F(G(F)).
Of course, \7 and 6 are closure operators.

In other terminology (see (24]) , if G:::; SA, then clones from F(G) are
called G-clones and F(G) is the lattice of all G-clones. It follows from
general facts on Galois-closed subrelations t hat are discussed in the
paper [5] of this volume that the lattice of all G-clones is a complete
sublattice of the lattice of all clones.
From property (vii), we know that v(G) is a subgroup of SA for every
G ~ SA. The question of whether the converse implication is true is
very natural. So, sometimes ago we formulated the following ques-
tions:3
Question A. Is it true t hat 'i7 (G) = G iff G :::; SA ?
This problem is equivalent to the next two problems:
Question B (concrete characterization problem). Is it true

that for every group G:::; SA there exists a clone lllS E £(A) such that
W Aut((A, lllS)) = G?
Question C (see [24]). Are the lattices of G 1-clones and G 2-clones

different for different groups G 1 , G 2 :::; SA ?
If the answer will be affirmative, then the lattice of lattices of G-clones

would be dual to the lattice of subgroups of SA . But recently we were
informed by A. Szendrei that this is not the case. 4
Example. Let 0 =I= B C A and let lF be the set of all mappings

fulfilling the following condition:
j(x1, ... , Xn) E B iff Xi E B for some 1 :::; i :::; n.
The set 11' = lF U lE is a clone and 'f E F( G) iff B is invariant under all
permutations from G:::; SA. Then one can observe that if the subgroup
G is maximal with respect to inclusion, having the set B as a union
oforbits, then v(G) =G.
Let E be the identity permutation on set A . Using this example for
A = {0, 1, 2} we can easily show that {c, (01)} , {c, (12)} and {c, (02)}
3 4 see "Added in proof" at the end of t his pa per.
are 'V-closed. We know that there are infinitely many elements in

F(A 3 ) and there are 48 elements in F(S3 ) (see [24]) which shows us
that A 3 is 'V-closed. Later on we will observe that {c} is V-closed.
For solving this problem, another Galois connection is helpful, namely
the connection between all finitary operations of the set A and all
relations defined on A (see [2] and [42]). Now, we recall some notions
which we need in order to describe this Galois connection.
Let f E (())(m) (A) and {! E JR(n) (A) . We say that the operation f pre-
serves the relation {! whenever
[(Vj E {l, ... , m}) ((xlj,··· , Xnj) E f!):::}
((!(xu , ... , XIm) , J (x21, ... , X2m), ... , f (xnl, ... , Xnm)) E f!)] .
For lF ~ (()) (A) and § ~ lR (A) let us define two operators
Pol(§) := { f E (()) (A) If preserves each {! E §},
Inv (JF) := {f! E JR(A) I each f ElF preserves f!}.
These operators Pol and Inv form the Galois connection mentioned
above (see [2], [42] and [46]); and for each lF ~ (())(A) , Pol (Inv (JF)) =
Lac ( (lF)) where
Loc(lF) = {f E (())(n) (A) I n 2: 1, ('If finiteB ~ An)(3 g E lF)[fls =
9ls]}
is the so-called local closure of the set IF (see [46]). If the set A is finite
then the Galois -closed sets of operations are exactly the clones.
For a permutation a E SA and a relation{! E JR(n) (A) let
a(f!) := {(a(ai) , ... ,a(an)) I (a l, ···,an) E f!}.
a
It is known ([24]) that for§ ~ lR (A), (Pol(§)) = Pol (a(§)) . If the
permutation a maps the set §onto itself, then so does a. If the set of
relations§ is Galois-closed, then for the clone C = Pol (S) , (C) = C a

iffa(S) = s.
Now, let us return to the example above. In order to show that the sub-
group {c} ::; S 3 is Galois-closed (\7 -closed with respect to the Galois
connection related to weak automorphisms) , one can easily show that
the only permutation which maps the clone Pol ( ~ ~ ~ ) (where
1r
1 2 ) ts
= ( 01 2 . a b"mary re1atwn
. m. wh"tch e1ements are wntten
. as
1
columns) onto itself, is the identity E. Indeed , let f be an operation of
one variable which preserves the relation 1r , and let f (0) = 1. Then we
must have f (1) = 2 and f (2) = 1. If a E S3, then let g =a(!) , i.e.
g (x ) =a(! (a- 1 (x ))). If a= (01) , then g (0) = 2 and g (1) = 0, and
thus g does not preserve the relation 1r. Analogously we can show that
the only permutation a E S3 , for which g preserves 1r , is the identity.
The Galois connection considered in this section, between permuta-

tions and clones, can be extended to a Galois connection between all
unary mappings from A to A (i.e. elements of the set (())( l) (A)) and
clones, by redefining the relation e in any of the three following ways.
Let TA = AA ( = ((J)(l) (A) ), T E TA and B E .C(A). We can define e
as
(c) T e B {::} T fulfills the condition (a) for the algebra (A; B) (see
Preliminaries) ;
(d) T e B {::} T fulfills the condition (b) for the algebra (A ; B) ;
(e) T e B {::} T is a weak endomorphism of the algebra (A ; B) .
Now, if H ~ TA and F ~ .C (A) , then taking each of definitions (c) ,

(d) , (e) for the relatione and defining operators ii, J as follows :
H H F(H) = {BE .C(A) I (Vr E H) [reB]} '

F HH(F) = { T ETA I ('l;lllli EF) [reB]} '
one gets three different Galois connections.

If we consider definition (d) and H ~ TA is a subsemigroup of t he
semigroup TA, t hen F(H) is the set of all H -clones (see [24]).
5 Some compatibilities and Galois connections
In this section we will present a construction which generalizes the

notions of weak automorphisms and preservation of a relation by an
operation. Let A be a fixed non-empty set. For every e E JR(n) (A),
i.e. e ~ An, we can define an n-ary relation {f on (())(m) (A), i.e. {f ~
(({])(m) (A))n as follows:
(JI, ... , fn) E {f iff
((fi (xn, · · ·, XIm), f2 (x21 1 • • ·, X2m), · · ·, fn (xn1 1 • • · , Xnm)) E g)],
fi (xu,.:... , XIm) )
I.e. ( E [!.
fn (xn1 , · · ·, Xnm)
E[! E[!
This definition encompasses a number of special cases. An operation

f satisfies (!, ... ,f) E {f exactly when f preserves the relation e (see
Section 4). In this way, for any set of relations § ~ JR( A) and for any
set of operations lF ~ ({]) (A) , we have
Pol(§)= { f E ({])(A) I (V {! E §) [(!, ... ,f) E Q]} and

Inv (JF) = { {! E JR(A) I (V f E JF) [(!, ... ,f) E Q]} .
Let a E SA (or more generally, a E A A). Define Qu as follows:
(x, y) E Qu ¢:? a(x) = y.
Then
(! ,g) E Q;, iffa(f(xl, ... ,xn)) =g(a(xi), ... ,a(xn)),
which means that (f) =g. a
Such a situation can be generalized in the following way.
A pair of clones A, JBl E £(A) are [)-COmpatible (where f2 is a binary

relation on A) if for every f E A there exists g E JBl such that (!,g) E
{j and for every g E JBl there exists f E A such that (!, g) E {j. If
A = JBl, then the pair A, JBl is f2u-compatible (for a E AA) iff a is a
weak endomorphism of 2( = (A, A) (or a weak automorphism , if a is
a permutation of set A).
Now, for i ::::: 1, we define the relation 8i between elements of JR(A)
and £ (A) as follows:
(} 8i C iff ((fi,. · · , fi- 1, fi, fi+1, · · ·, fn) E Q& f1, · · ·, fi- 1, f i+1 · · ·, fn
E C :::} fiE C) &((1:/fi E C)(3JI, ... , fi- 1, f i+I ... , fn E C)) [(JI, ... ,
fi-1, fi, fi+I, · · ·, fn) E Q),
where f2 E JR(A) is a relation of arity ::::: i and C E £(A). Thus the
Galois connection related to weak automorphisms from Section 4 can
be generalized in the following way:
§ r-+ F(§) = {C E £(A) 1 ei en,

(vf2 E §) [f2
;: r-+ ii(F) = { f2 E JR.( A) (vc E F) [f2 ei en,
1
where § ~ JR(A) , F ~ £( A) and i ::::: 1.

In particular, if§ = { f2u I a E G} for some group G :S SA of permu-
tations and i = 2, then F(§) is the lattice of all G-clones.
Added in proof:
The authors obtained an information from Agnes Szendrei that the
answer to question A (and for two its equivalent forms, namely ques-
tions B and C) is NO. T he answer is finally due to S. Marchenkov
([34]) and it is based on the papers [29], [30], [32], [33] (see also [51]
in this volume).
The authors would like to thank A. Szendrei for this important infor-
mation.
References
[1] G. Birkhoff, Lattice Th eory (2nd ed.). Arner. Math. Soc., New
York 1948 (cf. also the 1st ed. from 1940, and the 3rd ed., Provi-
dence, RI, 1967).
[2] V.G. Bodnarcuk, L.A. Kaluznin, V.N. Kotov, and B.A. Romov,
Galois theory for Post algebras. I, II. (Russian. English summary),

Kibernetika (Kiev) 1969, no. 3, 1-10; ibid. 1969, no. 5, 1-9.
[3] P.M. Cohn, Universal Algebra (2nd ed.) . D. Reidel Publ. Co.,
Dordrecht 1981.
[4] K. Denecke and S.L. Wismath, Hyperidentities and Clones. Gor-

don & Breach Sci. Publ. , London 2000.
[5] K. Denecke, S.L. Wismath, Galois connenctions and complete

sublattices. Galois Connections and Applications (Editors: K. De-
necke, M. Erne, S. L. Wismath) , Kluwer, 2003.
[6] S. Fajtlowicz, Properties of the family of independent subsets

of a general algebra. Colloq. Math. 14 (1966), 225-231.
[7] B. Ganter and R. Wille, Formale Begriffsanalyse. Mathematis-

che Grundlagen. Springer-Verlag, Berlin 1996 (English transl.:
Formal Concept Analysis. Mathematical Foundations. Springer-
Verlag. Berlin 1999).
[8] K. Glazek, Weak automorphisms of integral domains. Colloq.

Math. 22 (1970), 41-49.
[9] K. Glazek, Independence with respect to family of mappings in

abstract algebras. Dissert. Math. 81 (1971) .
[10] K. Glazek, Some old and new problems in the independence the-
ory. Colloq. Math. 42 (1979), 127-189.
[11] K. Glazek, Weak homomorphisms of general algebra and related

topics. Math. Seminar Notes (Kobe) 8 (1980), 1-36.
[12] K. Glazek, On weak automorphisms of finite fields . Colloq. Math .

Soc. J. Bolyai, vol. 28 (Finite Algebras and Multiple-Valued
Logic), North-Holland, Amsterdam 1981 , 275-300.
[13] K. Glazek, Weak automorphisms and some Galois connection.

Abstracts of Short Communications, International Congress of
Mathematicians, Kyoto (Japan) 1990, 11.
[14] K. Glazek, Semigroup of weak endomorphisms of universal alge-

bras. Proceedings of the International Symposium on the Semi-
group Theory (Kyoto 1990) , Matsue (Japan) 1990, 85-102.
[15] K. Glazek, On weak automorphisms of some finite algebras. Con-

temp. Math. 131 (1992), 99-110.
[16] K. Glazek, Morphisms of gener·al algebras without fixed fundamen-

tal operations. General Algebra and Applications, Heldermann-
Verlag, Berlin 1993, 89-112.
[17] K. Glazek, Algebras of Algebraic Operations and Morphisms

of Algebraic Systems. Acta Univ. Wratislaviensis, no. 1602, Wyd.
Uniw. Wrodawskiego, Wrod aw 1994, (Polish).
[18] K. Glazek, A certain Galois connection and weak automorphisms.

Acta Univ. Palack. Olomouc. Fac. Rerum Natur. Math. 36
(1997), 15-26.
[19] K. Glazek and J. Michalski, Weak homomorphisms of general

algebra. Comment. Math. 19 (1977) , 211-228.
[20] K. Glazek and S. Niwczyk A new perspective on Q-independences.

General Algebra and Applications, Shaker Verlag, Aachen 2000,
61-71.
[21] A. Goetz, On weak isomorphisms and weak homomorphisms of

abstract algebras. Colloq. Math. 14 (1966), 163-167.
[22] A. Goetz, A generalization of the direct product of universal al-

gebras. Colloq. Math. 22 (1971), 167-176.
[23] K . Golema-Hartman , Exchange property and t-independence. Col-

loq. Math. 36 (1976) , 181-186.
[24] V.V. Gorlov and R. Poschel, Clones closed with respect to permu-
tation groups or transformation semigroups. Beitriige zur Algebra
und Geometrie 39 (1998), 128-204.
[25] G. Gratzer, A new notion of independence in univer·sal algebra.

Colloq. Math. 17 (1967) , 225-234.
[26] G. Gratzer, Universal Algebra (2nd ed.). Springer-Verlag, New

York 1979.
[27] W. S. Hatcher and S. Whitney, Absolute Algebra. Teubner,

Leipzig 1978.
[28] N. van Hoa (Khoa) , On the structure of self-dual closed classes

of three-valued logic P 3 ,(Russian) Diskret. Mat. 4 {1992) , no. 4,
82-95.
[29] N. Van Hoa (Khoa) , Families of closed classes of k-valued logic

that are preserved by all automorphisms (Russian), Discret. Mat.
5, (1993) , no. 4, 87-108.
[30] N. van Hoa, Description of self- closed classes preserved by inner

automorphisms (Russian) , Dokl. Akad. Nauk Belarusi 38 (1994) ,
no. 3, 16-19.
[31] M. Kolibiar, Weak homomorphisms in some classes of algebras.

Studia Sci. Math. Hungarica 19 {1984) , 413-420.
[32] S. S. Marchenkov, Basic relations for the S - classification of

functions of multi-valued logic, Diskret. Mat. 8 {1996) , no. 2, 99-
128.
[33] S. S. Marchenkov, A -classification of functions of many-valued

logics (Russian), Diskret. Mat. 9 {1997) , no. 3, 125-152.
[34] S. S. Marchenkov, S-classification of functions of many-valued

logics (Russian) , Dokl. Akad. Nauk 366 (1999) , no. 4, 455-457.
[35] E. Marczewski, Independence and homomorphisms in abstract al-

gebras. Fund. Math. 50 (1961) , 45-64.
[36] E. Marczewski, Independence in abstract algebras. Results and

problems. Colloq. Math. 14 (1966), 169-188.
[37] E. Marczewski, Independence with respect to a family of map-

pings. Colloq. Math. 20 (1968) , 11-17.
[38] 0. Ore, Galois connexions. Trans. Amer. Math. Soc. 55 (1944) ,

493-513.
[39) E. Plonka, Note on weak automorphisms of algebras having a ba-

sis. Colloq. Math. 24 (1971), 7-10.
[40) J. Plonka and W. Poguntke, T-independence in distributive lat-

tices. Colloq. Math. 36 (1976) , 171-175.
[41) R. Poschel, A general Galois theory for operations and relations

and concrete characterization of related structures. Report R-
01/80, Akademie der Wissenschaften der DDR, Institut fiir Math-
ematik, Berlin 1980.
[42) R. Poschel and L.A. Kaluznin , Funktionen- und Relationenalge-

bren. VEB Deutscher Verlag der Wissenschaften, Berlin 1979.
Birkhauser Verlag, Basel and Stuttgart 1979.
[43) J . Riguet , Relations binaires, fermetures , correspondences de Ga-

lois. Bull. Soc. Math. France 76 (1948) , 114-155.
[44) I.G. Rosenberg, La structure des functions de plusieurs variables

sur un ensemble fini. C. R. Acad. Sci. Paris, Ser. A-B , 260 (1965) ,
3817-3819.
[45) I.G. Rosenberg, Uber die funktionelle Vollstiindigkeit in den

mehrwertigen Logiken. Rozpravy Ceskoslovenske Akad. Ved., Ser.
Mat. Nat. Sci., v. 80 (1970) , no. 4, 3-39.
[46) M. RoBiger, A unified characterization of clones. Contributions

to General Algebra 12 (2000) , 355-369.
[47) J. Schmidt, Eine algebraische Aquivalente zum Auswahlaxiom.

Fund. Math. 50 (1962) , 485-496.
[48) J. R. Senft, On weak automorphisms of universal algebras. Dis-

sert. Math. 74 (1970) , 1-35.
[49) S. Swierczkowski, Topologies in free algebras. Proc. London Math.

Soc. (3) 14 (1964) , 566-576.
[50) S. Swierczkowski, A sufficient condition for independence. Colloq.

Math. 9 (1962) , 39-42.
[51] A. Szendrei, A Survey of Clones Closed Under Conjugation, Ga-

lois Connections and Applications (Editors: K. Denecke, M. Erne,
S. L. Wismath),Kluwer, 2003.
[52) R. Wille, Reconstructing lattice theory: an approach based on hi-

erarchies of concepts. Ordered Sets, D. Reidel Publ. Co., Dor-
drecht 1982, 445-470.
[53) S.V. Yablonski!, G.P. Gavrilov, and V.B. Kudryavtsev , Functions

of the algebra of logic and Post classes (in Russian) . Izd. "Nauka",
Moscow 1966.
Authors' addresses:
K. Glazek, St. Niwczyk ,
Institute of Mathematics,
Technical University of Zielona G6ra,
ul. Pog6rna 50 ,
65-246 Zielona G6ra, Poland,
e-mail: k.glazek@im.pz.zgora.pl
s.ni wczy k@im. pz.zgora.pl
A Survey of Clones Closed Under
Conjugation *
/
A. Szendrei
Abstract
There is a Galois connection between the lattice of clones on a set A and

the group of permutations on A that is determined by the relation that
a permutation conjugates a clone onto itself. The Galois-closed sets on
the clone side are the lattices La of all clones that are closed under con-
jugation by all members of some permutation group G. In this paper we
discuss the coarse structure of the lattice La when A is finite and G is
a 2-homogeneous permutation group, describe La completely for the case
when G is the group of all permutations, and discuss for which groups G
the lattice L a is finite.
AMS Mathematics Subject Classification: Primary OBA05, Secondary 08A40,
08A35, 03B50
Key words: Clone, Conjugation, Weak automorphism.
1 Introduction
Many important symmetries of algebras are not automorphisms. For

example, inversion T G ----+ G, g H g- 1 is an antiautomorphism of a
group, and this is a type of symmetry that is not an automorphism
(unless G is abelian). Transpose T Mn(IR) ----+ Mn(IR), A H AT is
an antiautomorphism of a matrix ring, which is a symmetry that is
not an automorphism (unless n = 1). Complementation '"'/: B ----+ B,
*This material is based upon work supported by the Hungarian National Foun-
dation for Scientific Research (OTKA) grants no. T 026243, T 034175, T 37877.
297
K. Denecke et at. (eds.), Galois Connections and Applications, 297-343.
298 A. Szendrei
b H b in a Boolean algebra is a symmetry that is not an auto-

morphism. If K is a field and a is an automorphism of K , then
the coordinatewise action of a defines a symmetry 'Y : Kn -t Kn ,
(k 1 , ... , kn) H (a(ki) , . .. , a(kn)) of then-dimensional K-vector space
Kn that is not an automorphism (unless a= id) .
The kind of symmetry exhibited by these examples is the focus of this
paper. For the general concept, let A be a set and 'Y be a permutation
of A. Iff: An -t A is an operation on A, we define the conjugate of
f by "f to be
The conjugate of a set F of operations by 'Y is 1F = {'! : f E F} ,

and the conjugate of an algebra A= (A; JI, h, ...) by 'Y is the algebra
' A = (A; 'fi ,'h , ...). It is easy to see that 'Y is an automorphism
of A if and only if ' f i = fi for all fundamental operations fi of A.
The common feature of the symmetries mentioned in the preceding
paragraph is that they are permutations 'Y of the underlying set of an
algebra A = (A; JI, h , .. .) such that A and 1A have the same term
operations. Such a permutation 'Y is called a weak automorphism of
A . In other words, "( is a weak automorphism of A if and only if the
clone <r of term operations of A satisfies the equality ' <f = <r.
For any fixed permutation 'Y on a set A , conjugation by 'Y yields an
automorphism <r H ' <f of the lattice of all clones on A . The con-
dition ' <f = <r defines a Galois connection between clones on A and
permutations on A as follows: we assign to every set of clones the
collection of all permutations that fix every clone in the set, and to
every set of permutations the collection of all clones that are fixed by
every permutation in the set. 1 It is easy to see that the Galois-closed
sets of clones are complete sublattices of the lattice of all clones, while
the Galois-closed sets of permutations are subgroups of the symmetric
group on A. For a permutation group G , the members of the Galois-
closed set { <r : 1 <r = <r for all 'Y E G} of clones corresponding to G
will be called G-closed clones. It is easy to see from the definitions
that the family of G-closed clones includes all clones <r that contain
G, and also all clones <r for which G is a group of automorphisms of
the algebra (A; <r) .
1 This Galois connection is discussed in detail in the paper [8] in this volume.
Clones Closed Under Conjugation 299
Our aim in this paper is to give an overview of some results on the

lattice of G-closed clones on a finite set when G is a large permuta-
tion group. In Section 3 we discuss Szabo's theorem [38] which can be
used to determine the coarse structure of the lattice of G-closed clones,
including all coatoms of the lattice, when G is a 2-homogeneous per-
mutation group (a property slightly weaker than 2-transitivity). In
Section 4 we present a complete description of all SA-closed clones
where SA is the symmetric group on A. This description is a modified
version of results obtained earlier by Hoa [15] and Marchenkov [28]. In
Section 5 we briefly discuss a recent result [20] which lists all groups
G such that the family of G-closed clones that contain all constants
is finite. For each major theorem mentioned above we outline a proof
which, in some cases, differs essentially from the proofs published ear-
lier. Our goal is to emphasize the advantages of combining different
methods: using operational and relational arguments simultaneously,
which is facilitated by the Galois connection between operations and
relations 2 , and applying localization and globalization from tame con-
gruence theory in combination with these arguments.
Finiteness of the base set is crucial in these results and methods. For
the case when the base set A is infinite, Goldstern and Shelah [10]
obtained interesting results on a special class of SA-closed clones. They
studied the lattice of all clones that contain the full transformation
monoid on A, and found that the structure of this lattice depends
heavily on partition (Ramsey) properties of the cardinality IAI of the
base set. Specifically, if IAI is a weakly compact cardinal, then this
lattice has exactly 2 coatoms, while if IAI is, say, the successor of an
uncountable regular cardinal, then this lattice has 221 A 1 coatoms.
2 Preliminaries and examples
For any set A, a clone on A is a collection of finitary operations on

A that contains the projection operations and is closed under compo-
sition. For any set F of operations on A, [F] will denote the clone
generated by F. In other words, [F] is the collection of all term op-
erations of the algebra A = (A; F), which is called the clone of A ,
and is also denoted by Clo(A). If A = (A; F) is an algebra, then A c
2 For more details about this Galois connection, see [34] in this volume.
300 A. Szendrei
will denote the algebra that arises from A by adding all constants as
fundamental operations. The clone of A c , that is, the clone generated
by F and all constants, is called the clone of polynomial operations of
A , and is denoted by Pol(A). For n ~ 1, Clon(A) and Poln(A) will
denote the set of n-ary term operations, and the set of n-ary polyno-
mial operations of A , respectively. A finite algebra A is called primal
if Clo( A) is the clone of all operations on A , and functionally com-
plete if A c is primal. We will call A a projection algebra if Clo( A) is
the clone of projections. Two algebras are said to be term equivalent
if they have the same clones, and they are said to be polynomially
equivalent if they have the same clones of polynomial operations.
An operation f on A is idempotent if it satisfies the identity f(x, ... , x)
= x . An algebra A is idempotent if all fundamental operations (and
hence all term operations) of A are idempotent. A clone <t is idempo-
tent if all operations in <tare idempotent. A ternary operation f is a
Mal'tsev operation if it satisfies the identities f(x , y , y) = f(y , y, x) =
x; f is a minority operation if it satisfies the identities f(x , y , y) =
f(y , x , y) = f(y, y , x ) = x ; and f is a majority operation if it satisfies
the identities f(x , y , y) = f(y , x , y) = f(y , y , x ) = y.
We will use the notation 'TA. , CA , SA , and AA for the full transformation
monoid, its subsemigroup of all constants, the symmetric group, and
the alternating group on A , respectively. For IAI = 4, VA will denote
the Klein group on A.
Let --y E SA. For an n-ary operation f on A the conjugate of f by --y is
the operation
The conjugate of a set F of operations by --y is defined as 'YF = Pf :

f E F} . It is straightforward to check that 7 <t is a clone for every
clone <t and permutation --y, and the mapping <t 1---7 7 <t defines an
inner automorphism of the lattice of clones on A for every --y E SA.
If the clone of an algebra A is fixed by this automorphism , that is,
if 7 Clo(A) = Clo(A) holds, then --y is called a weak automorphism
of A (see [8] or [9]) . The weak automorphisms of A form a group,
which is called the weak automorphism group of A , and is denoted by
WAut(A). It follows immediately from the definitions that for any
algebra A, the automorphisms of A as well as the permutations g
with g , g- 1 E Clo 1 (A) are weak automorphisms of A. In fact , it is
clone of all operations
clone of
[A, v]
[0, 1]
clone of projections
Fig. 1: SA-closed clones on A= {0, 1}
easy to check that the automorphism group Aut(A) of A is a normal

subgroup of the weak automorphism group WAut(A), and so is the
largest subgroup {g E SA : g,g- 1 E Clo 1 (A)} of Clo 1 (A). If A is
finite, then the latter group is SAn Clol(A).
For a group G ~ SA and a set F of operations on A we will say that
F is G-closed if 1F = F for all 1 E G. In particular, it follows that
a clone It on A is G-closed if and only if it is fixed by every inner
automorphism of the clone lattice that is induced by a permutation
from G, or equivalently, It is G-closed if and only if G is a subgroup
of the weak automorphism group of the algebra (A; It) . Clearly, if G
is the one-element group, then every clone is G-closed.
2.1 Example. On a 2-element set A = {0, 1} the only nontrivial

group is G =SA, and the SA-closed clones can be easily determined ,
using Post's description [36] of all clones. Figure 1 shows the lattice of
all these clones. In the diagram 0, 1, A, V, - denote the Boolean algebra
operations, p(x, y, z) = x+y+ z is the unique minority operation, and
m(x, y, z ) = (x A y) V (x A z) V (y A z ) is the unique majority operation
onA={0,1}.
From now on we will assume that A is a finite set with at least three
elements, and A is the underlying set of all algebras and clones consid-
302 A. Szendrei
ered. A relation 12 ~ Am is called a compatible relation of an algebra

A= (A; F) if 12 is (the underlying set) of a subalgebra of the m-th di-
rect power Am of A. Examples of compatible binary relations include
congruences and the graphs of automorphisms. (The graph of an n-ary
partial function h is defined to be the (n + 1)-ary relation consisting
of all tuples (a 1 , ... , an, h(a 1 , ... , an)) such that (a 1 , .. . , an) is in the
domain of h.) If 12 is a compatible relation of the algebra (A; f), we
will also say that f preserves (2.
It is well known (see e.g. [34, 35]) that for finite A, clones on A can
be characterized as the Galois-closed sets of operations in the Galois
connection between operations on A and relations on A that is defined
by the condition "f preserves 12". This Galois connection assigns to
every set of operations the collection of all relations that are preserved
by every operation in the set, and to every set of relations the collection
of all operations that preserve every relation in the set. It is easy to
see that Galois-closed sets of operations are clones. The fact that all
clones are in fact Galois-closed provided the base set is finite, can be
expressed as follows.
2.2 Claim. [34, 35] On a finite base set A, an operation f belongs
to a clone Q.: if and only if f preserves all compatible relations of the
algebra (A; Q.:).
For any fixed permutation group G s;;; SA one can modify this Galois
connection so that the Galois-closed sets of operations are exactly the
G-closed clones (cf. e.g. [11]). For any 1 E SA and any relation (} s;;; Am
let
7 12 = 1(12) = { (r(x1), ·. ·, I(Xm)) : (x1, ... , Xm) E 12}.
Since 7f preserves r2 for all 1 E G if and only if f preserves 7 12 for

all 1 E G, these two conditions define the same Galois connection,
and it follows that the Galois-closed sets of operations in this Galois
connection are exactly the G-closed clones. In other words:
2.3 Claim. [11] On a finite base set A, either one of the following
conditions characterizes G-closed clones Q.::
(1) 7f E Q.: for every f E Q.: and 1 E G;
(2) 7 12 is a compatible relation of the algebra A = (A; Q.:) for every

compatible relation 12 of A and for every 1 E G.
The same argument that proves that clones on a finite set form an
atomic and dually atomic, algebraic and dually algebraic lattice, also
proves that G-closed clones form an atomic and dually atomic, alge-
braic and dually algebraic lattice for all G. In fact, the lattice of G-
closed clones is a complete sublattice of the lattice of H-closed clones
whenever H ~ G ~ SA, and in all these lattices the smallest element
is the clone of projections, and the largest element is the clone of all
operations. 3 One cannot expect that the lattice of G-closed clones is
much simpler than the lattice of all clones, unless the permutation
group G is fairly large.
Next we summarize the notions and basic facts about permutation
groups that we will need in this survey. A permutation group G ~ SA
is called primitive if the unary algebra (A; G) is simple, and IGI > 1
when IAI = 2. For 1 < k < IAI , G is said to be k-transitive if for
any two lists b1 , ... , bk and c1 , ... , ck of pairwise distinct elements of
A there exists "( E G such that ci = "f(bi) for all i; G is said to
be k-homogeneous if for any two k-element subsets B, C of A there
exists 'Y E G such that C = 1(B). A !-transitive (or equivalently,
!-homogeneous) group is briefly called transitive. Clearly, primitive
permutation groups are transitive, k-transitive permutation groups
are k-homogeneous, and k-homogeneous permutation groups are also
(IAI - k)-homogeneous. It is also easy to see that 2-homogeneous
permutation groups are primitive.
The two most well known infinite families of 2-transitive groups are
the affine linear groups and the projective linear groups. Let K A be
a finite vector space with underlying set A. The automorphism group
of K A is the general linear group G L(K A) , and its subgroup consisting
of all automorphisms with determinant 1 is the special linear group
SL(KA). The affine general linear group AGL(KA) is the subgroup of
SA generated by GL(KA) and by the group of translations TR(KA) =
{x+a: a E A}; AGL(KA) is a 2-transitive group on A for every vector
space KA. We note that the permutation group AGL(KA) determines
3 This is a consequence of general facts on Galois-closed subrelations that are

discussed in the paper [7] in this volume, because the following is true for arbitrary
permutation groups G, H <;;; SA: if H <;;; G , then the relation between operations
and relations on A that is defined by the condition " "~f preserves f2 for all "'( E G"
is a Galois-closed subrelation of the relation defined by the condition " "~f preserves
f2 for all "'( E H " .
304 A. Szendrei
the vector space KA up to the choice of 0; equivalently, AGL(KA)

uniquely determines the affine space KAid (see Example 2.11). The
projective general linear group PGL(K A) is the quotient of GL(K A) by
its center consisting of all scalar multiplications in K A, and PSL(K A)
is its subgroup that arises from SL (K A) in a similar way; both act 2-
transitively on the set of one-dimensional subspaces of K A (the points
of a projective geometry). If we identify the base set A of the vector
space KA with Kd (d = dim KA), then each automorphism of the
field K yields a weak automorphism of K A by acting coordinatewise
on A = Kn (see the Introduction), and this weak automorphism has
a natural permutation action on the set of one-dimensional subspaces
of KA as well. The permutation group generated by AGL(KA), resp.
PGL(KA), and the corresponding permutations induced by the field
automorphisms is denoted by AfL(KA), resp. PfL(KA). We will omit
the subscript K if K is a prime field, because then the vector space
K A is term equivalent to its underlying elementary abelian group A.
Clearly, we have AfL(A) = AGL(A) and PfL(A) = PGL(A) in this
case. Moreover, ArL(KA) ~ AGL(A) for every vector space KA with
underlying abelian group A. In statements where the vector space
KA is relevant up to isomorphism only and dim(KA) = d, /K/ = q,
we write GL(d, q), AGL(d, q), PGL(d, q), etc., in place of GL(KA) ,
AGL(KA), PGL(KA), etc.
The 2-homogeneous permutation groups (on a finite set) are com-
pletely classified: a description of 2-homogeneous groups that are not
2-transitive can be found e.g. in [18], Chapter XII, Section 6; the clas-
sification of 2-transitive groups, which is based on the classification of
finite simple groups, is summarized e.g. in [4]. There are two types of
2-homogeneous groups:
• Affine 2-homogeneous groups: these are 2-homogeneous groups

G such that TR(A) ~ G ~ AGL(A) for some elementary abelian
p-group A (p prime); TR(A) is the unique minimal normal sub-
group of G.
• Almost simple 2-transitive groups: these are 2-transitive permu-

tation groups G ~ SA such that G has a unique minimal normal
subgroup N, N is a nonabelian simple group, and - with one
exception when N is primitive but not 2-transitive (/A / = 28)
- N is 2-transitive on A.
After these preparations we will look at some examples of clones that

are G-closed for fairly large permutation groups G.
2.4 Example. Let A = (A; T) be a unary algebra. Since we are

interested in the clone of A, we may assume without loss of generality
that Tis a submonoid of 'TA_, and hence T = Clo 1 (A). Clearly, Clo(A)
is G-closed if and only if the group P = TnSA of permutations in T as
well as the su bsemigroup U = T \SA of nonsurjective transformations
in T are G-closed. It is easy to see that P is G-closed if and only
if P is a normal subgroup of a group H ~ SA with G ~ H (H can
be chosen to be WAut(A)). G-closed semigroups U of nonsurjective
transformations are not so easy to characterize. However, it is easy
to verify that if G is 2-homogeneous and U is a (nonempty) G-closed
subsemigroup of 'TA_ \SA, then CA ~ U.
These observations apply to the monoid T = Q:(l) of unary operations
of any G-closed clone Q:. Thus we get the following
Corollary 2.5. IfQ: is a G-closed clone and G is 2-homogeneous, then
one of the following conditions holds:
• Q:(l) = { id}, that is, Q: is a clone of idempotent operations;
• Q: contains all constants;
• Q:(l) is a transitive permutation group; in fact, Q:(l) is a nontriv-

ial normal subgroup of the 2-homogeneous group WAut( (A; Q:))
containing G.
In the special case G = SA there is a transparent characterization

for the G-closed transformation monoids. To describe these monoids,
define the type of an equivalence relation e on A to be the increasing
sequence K, = ( k1 , k2 , ... , kr) of positive integers that lists the sizes
of the e-classes (hence k1 + k2 + · · · + kr = IAI). A transformation
f E 'TA_ is said to have kernel type K, if K, is the type of ker(f). There
is a natural partial ordering of kernel types induced by the inclusion
ordering of equivalence relations: if K, and A are kernel types, then
K, < A exactly when there are equivalence relations e, and e>. of types
K, and A respectively for which e, ~ e>..
2.6 Theorem. [21] Let A be a finite set, and let U be a subsemigroup

of 'TA_ \ SA. The following conditions are equivalent:
306 A. Szendrei
(i) U is SA -closed;
(ii) there is a filter F of kernel types on A such that the members of
U are exactly the transformations whose kernel types belong to
F.
Combining this theorem with some observations made in Example 2.4

we get the following corollary.
Corollary 2. 7. For a finite set A, a submonoid T of 74 is SA -closed

if and only if T = P U U, where P is one of the permutation groups
{id}, VA (IAI = 4) , AA, SA, and either U = 0 or U is a subsemigroup
of 74 \ SA that satisfies condition (ii) of Th eorem 2. 6.
2.8 Example. For 2 < m < IAI let 9\n denote the clone consisting
of the projections and all operations whose range has size at most m.
An important subclone of 9\2 is the following clone 113: it consists of
the projections and all operations of the form
(n 2: 1) (2.1)
where <p 1 , ... , 'Pn are mappings A ---7 {0, 1}, +is addition modulo 2,
and cp is any mapping {0, 1} ---7 A. If IAI is even, then the projections
and all operations of the form (2.1) where the kernel of each 'Pi (i =
1, ... , n) has two even size blocks form a proper subclone in 113; this
clone will be denoted by 113*. It is easy to check that all the clones
9\n, 113, and 113* are SA-closed.
It is well known [3, 22, 37) that the clones containing 74 form a chain
whose members are exactly the clones [74), 113 U [74), and 9\n U [74)
(m = 2, ... , IAI). The coatom 9\IAI-I U [74) of this chain is called
the Slupecki clone. The description of all clones containing 74 was ex-
tended in [12] to all clones containing SA. Since all clones that contain
SA are SA-closed, these results will be covered by the description of
SA-closed clones discussed in Section 4.
2.9 Example. Another rich class of SA-closed clones is the class con-
sisting of the clones Clo(A) where A is a homogeneous algebra, that
is, A has S A as its automorphism group. These clones were described
in [6] and (23). The description was extended in [25) to the case when
the automorphism group of A is AA (IAI > 3); interestingly, these
clones are also SA-closed.
Examples of homogeneous algebras include the algebras (A; t), (A; d) ,

(A; s), and (A; £i) (i 3, ... , IAI) whose operations are defined as
follows:
z if x = y, X if X= y ,
t(x, y, z) = { x d(x, y , z) = { z
itx=/=y, if X =/= y,
s(x, y,z) = G if X= y,
if X= z,
otherwise,
and
if x 1 , .. • , xi are pairwise distinct ,

otherwise.
The operation t is called the ternary discriminator, and d is called

the dual discriminator. Note that on the 2-element set A = {0, 1} , s
coincides with the operation x+y+z and d coincides with the majority
operation m from Example 2.1.
2.10 Example. Let A be a quasiprimal algebra, that is, a finite

algebra such that t is a term operation of A. Equivalently, A is
quasiprimal if and only if every operation on A that preserves the
isomorphisms between subalgebras of A is a term operation of A .
Isomorphisms between subalgebras are briefly called internal isomor-
phisms. Suppose that WAut(A) is 2-homogeneous and A is not idem-
potent. If Clo(A) contains all constants, then A has no proper sub-
algebras and no nontrivial automorphisms, therefore A is primal and
WAut(A) = SA· In view of Corollary 2.5 the only remaining possi-
bility is that Clo 1 (A) is a transitive normal subgroup of WAut(A).
The transitivity of Clo 1 (A) implies that A has no proper subalgebras,
therefore the only internal isomorphisms of A are its automorphisms.
The automorphism group Aut(A) of A cannot be trivial, because
A is not primal. Therefore Aut(A) is also a transitive normal sub-
group of WAut(A). Hence the classification of 2-homogeneous groups
shows that, for some elementary abelian p-group (p prime) , we have
Clo 1 (A) = TR(A) = Aut(A) and WAut(A) ~ AGL(A). It follows
that A is term equivalent to (A; t , TR(A)), because both algebras are
308 A. Szendrei
quasiprimal and they have the same internal isomorphisms. Thus

every permutation from AGL(A) is a weak automorphism of A, so
WAut(A) = AGL(A) holds in this case.
2.11 Example. An algebra A is called affine if for some ring R
there exists a module RA on the universe A of A such that A is
polynomially equivalent to RA. It can be proved that in this case
A has the same idempotent term operations as the affine module
RAid= (A;x-y+z,{rx+(1-r)y: r E R}). In particular, x -y+ z
is a term operation of A. Now we will look at two important classes
of affine algebras, and determine the weak automorphism groups of
these algebras. One class consists of those finite affine algebras that
are polynomially equivalent to some vector space K A , and the other
consists of all finite simple affine algebras. If A is a finite simple affine
algebra, then A is polynomially equivalent to a finite simple module;
up to term equivalence, a finite simple module is of the form RA where
R = End(KA) and KA is a vector space.
Thus, let us assume that A is polynomially equivalent to a module
RA where R = K orR= End(KA) for a finite vector space KA· Since
x - y + z is the only Mal'tsev operation in t he clone of A , it must
be conjugated to itself by every 1 E WAut(A). Thus WAut(A) ~
Aut( (A; x - y + z)) =Aut( Aid)= AGL(A). In the case when R = K,
the set of binary idempotent term operations of A is {gk : k E K}
where gk(x , y) = kx + (1- k)y; this set is closed under conjugation
by all1 E WAut(A), therefore we get that WAut(A) ~ AfL(KA). In
the case when R = End(K A), those ternary compatible relations of
A that are graphs of binary idempotent operations on A are exactly
the graphs of the operations gk (k E K); by Claim 2.3 this set of
compatible relations is closed under conjugation by all 1 E WAut(A),
hence we conclude again that WAut(A) ~ ArL(KA).
Now suppose that the weak automorphism group of A is 2-homogen-
eous. If A is idempotent, then it follows from our general remarks
about affine algebras that A is term equivalent to RAid. If Clo(A) con-
tains all constants, then by definition A is term equivalent to RAe. Ac-
cording to Corollary 2.5, there is only one more possibility, namely that
Clo 1 (A) is a transitive permutation group. Since ux + a E Clo 1 (A)
implies r(ux +a)+ (1 - r) x = (1 - r(1 - u) )x + ra E Clo 1 (A) for
all r E R and some of these operations are not permutations unless
u = 1, we get that Clo 1 (A) ~ TR(A). As Clo 1 (A) is transitive, equal-
ity holds. Hence A is term equivalent to the algebra RAtr that we

get from RAid by adding all translations from TR(A) as basic opera-
tions. It is straightforward to check that whether A is term equivalent
to RAid or RAe or RAtr, every permutation from AfL(KA) is a weak
automorphism. Thus WAut(A) = AfL(KA).
2.12 Example. The m-th matrix power of an algebra B, denoted
B[ml, is the algebra on the underlying set Bm whose operations are
the following: (i) all term operations of the m-th direct power of B,
(ii) the diagonal operation ~ defined by
~((xli)~l' · · ., (xmi)~l) = (xn, ... ,Xmm),
and (iii) for each permutation 1r E Sm, the unary operation U1r: Bm ~
Bm that permutes the coordinates of Bm according to 1r.
If B is a nontrivial unary algebra, then every term operation of B[m]
depends on at most m variables, and the m-ary idempotent term op-
erations of B[m] that depend on all variables are exactly the operations
that arise from ~ by permuting variables. Therefore this set of opera-
tions is closed under conjugation by all permutations 1 E WAut(B[ml).
This implies that WAut(B[ml) ~ S 8 Wr Sm, the wreath product of S 8
and Sm (with the product action), which is the subgroup of S 8 m gener-
ated by all permutations 1 1 x · · · x rm (ri E S8 ) acting coordinatewise,
and all u7r (1r E Sm) that permute the coordinates. It is not hard to
show that the permutation group S 8 Wr Sm is not 2-homogeneous on
Bm if IBI 2: 2 and m 2: 2. Thus, for a nontrivial unary algebra B ,
WAut(B[ml) is not 2-homogeneous unless m = 1.
3 G-closed clones when G is 2-homogeneous

In [38] L. Szabo classified all finite algebras A such that A expanded
with its weak automorphisms is simple. This class includes every alge-
bra whose weak automorphism group is primitive. The classification
is more transparent when the weak automorphism group is assumed
to be 2-transitive (Theorem 5.6 in [38]). In this section we discuss
this result and sketch its proof which nicely combines methods from
clone theory and tame congruence theory. The extraction of a direct
proof for the 2-transitive case reveals that for the classification in The-
orem 5.6 of [38] it suffices to assume the slightly weaker condition that
the weak automorphism group is 2-homogeneous.
310 A. Szendrei
A remarkable consequence of this classification is an explicit descrip-

tion of all maximal G-closed clones when G is 2-homogeneous. In
particular, we will see that for all 2-homogeneous permutation groups
G that are not contained in an affine group AG L( d, p), there are ex-
actly two maximal G-closed clones: the clone of idempotent operations
and the Slupecki clone. The maximal G-closed clones were determined
earlier by Hoa [16] and Marchenkov [26] under much stronger transi-
tivity assumptions on G; their assumptions imply k-homogeneity for
every k (1 < k < IAI); cf. Theorem 5.2.
3.1 Theorem. [38] Let A be a finite algebra with at least three ele-
ments. If the weak automorphism group of A is 2-homogeneous, then
one of the following conditions holds for A:
(0) A is a primal algebra;
(I) A is a simple idempotent algebra that is not affine;
(U) A is an essentially unar·y algebra that is term equivalent to (A; M)

or (A ; M)c where M is a normal subgroup of a 2-homogeneous
permutation group on A;
(S) A is a simple algebra such that Clo(A) is contained in the Slupecki

clone, and contains a unary operation that is neither constant
nor a permutation;
(A) A is an affine algebra; namely A is term equivalent to one of the

algebras RAid, RAtr, or RAe (see Example 2.11) wher·e A is an
elementary abelian group on A and for a subfield K of End(A),
either R = K orR= End(KA);
(Q) A is a quasiprimal algebra that is term equivalent to the algebra

(A; t , TR(A)) for some elementary abelian group A.
This theorem allows us to determine the coarse structure of the lattice

of G-closed clones for any 2-homogeneous permutation group G . For
a given 2-homogeneous permutation group G , the lattice of G-closed
clones consists of all clones (!: = Clo(A) where A is one of the algebras
listed in the theorem such that G ~ WAut(A). Figure 2 displays this
lattice, and indicates the types of the clones according to Theorem 3.1.
In Figure 2 clones of types (A) and (Q) are denoted by circles, and all
clone of all operations (0)
clone of projections
Fig. 2: G-closed clones on A (IAI 2': 3) for 2-homogeneous G
other clones by bullets. This distinction is made to emphasize that G-

closed clones of types (A) and (Q) exist if and only if G ~ AGL(d,p)
for some prime p. This follows from Examples 2.10 and 2.11, because
if A is as in (A), then WAut(A) ~ AGL(A), and equality holds for at
least one A in (A), while if A is as in (Q), then WAut(A) = AGL(A).
Therefore, if G is not contained in any affine linear group AGL(d,p)
(p prime), then the lattice of G-closed clones looks much simpler than
Figure 2: all circles and adjacent lines can be deleted. In particular,
we see that there are exactly two maximal G-closed clones: the clone
of idempotent operations and the S~upecki clone.
Next assume that G is a subgroup of an affine linear group AGL(d,p)
and contains the translation subgroup TR(d,p) of AGL(d,p). Then,
up to the choice of 0, there is a unique elementary abelian p-group A
such that TR(A) ~ G ~ AGL(A). Hence there is a unique G-closed
clone of type (Q), and a unique family of G-closed clones of type (A) ,
which form three disjoint intervals
[Clo(Aict), Clo(EAict)], [Clo(Atr), Clo(EAtr)], [Clo(Ac), Clo(EAc)]

(E = End(A)) (3.1)
as depicted in Figure 2, and each interval is isomorphic to the lattice
of all subrings R of End(A) such that either R = K orR= End(K A)
for a subfield K of End(A) with G ~ AfL(KA). So, in this case,
there are exactly four maximal G-closed clones. Finally, it remains to
312 A. Szendrei
consider the case when G is a 2-homogeneous permutation group such

that G ~ AGL(d,p) for some d and some prime p, but the normal
subgroup G n TR(d, p) of G is trivial. I am indebted toP. P. PaJfy for
pointing out that in this case G cannot be of affine type, and if G an
almost simple 2-transitive permutation group, then the classification
of these groups can be used to show that the only possibility for G and
AGL(d,p) is the following: IAI = 8, G = PSL(2, 7), and d = 3, p = 2.
It is not hard to check that if we are given a 2-homogeneous group
G = PSL(2, 7) on an 8-element set A, then there are two essentially
different elementary 2-groups A on A such that G ~ AGL(A). Hence
there are two G-closed clones of type (Q), and two disjoint families
of G-closed clones of type (A). It follows, in particular, that there are
exactly six maximal G-closed clones.
Sketch of proof of Theorem 3.1. Let A be a finite algebra such that

IAI ;:::: 3 and G = WAut(A) is 2-homogeneous. Assume first that A
is not simple, and let B( u, v) (u =I= v) be a minimal congruence of
A. Since G = WAut(A) is 2-homogeneous, every nontrivial principal
congruence of A is of the form () (1( u), 1(v)) = '( B( u, v)) for some
1 E G, and is therefore a minimal congruence. Suppose A has a unary
polynomial operation f E Poh (A) which is neither a permutation nor
a constant. Let B be a nonsingleton kernel class of f, b E B a fixed
element, and let C = A\ B. Then for any b' E B \ {b} and c E C
we have B(f(b), f(c)) = B(f(b'), f(c)) ~ B(b, c) n B(b', c) and B(b, b') ~
B(b, c) V B(b', c); hence the minimality of the principal congruences
implies first that B(b, c) = B(b', c) and then that B(b, c) = B(b, b').
Since this equality holds for all b' E B \ {b} and c E C, we see that
the minimal congruence B( b, x) ( x E A\ { b}) is independent of x, and
is therefore the full relation. Thus A is simple, which contradicts our
assumption. This proves that every unary polynomial operation of A
is a permutation or a constant. Hence we can apply PaJfy 's Theorem:
3.2 Theorem. [33] If A is a finite algebra with at least three elements

such that every unary polynomial operation of A is a permutation or
a constant, then either A is essentially unary or A is polynomially
equivalent to a vector space.
Thus, A satisfies condition (U) (see Example 2.4) or condition (A)

with R = K (see Example 2.11).
From now on we will assume that A is simple. If A is an idempotent

algebra, then either (I) holds, or A is a simple idempotent affine al-
gebra. Applying again the facts established in Example 2.11 we get
that A is term equivalent to RAid with R = End(KA) as in (A) .
Assume now that A is simple and not idempotent. Let Q: = Clo(A). It
follows from Corollary 2.5 that either Q:(l) is a t ransitive permutation
group, or Q:(l) contains all constants. In either case, A has no proper
subalgebras, therefore we can apply the following strengthening of
Rosenberg's Primal Algebra Characterization Theorem:
3.3 Theorem. [40, 41] Let A be a finite simple algebra with no proper
subalgebras. Then one of the following conditions holds for A:
(a) A is quasiprimal;
(b) A is affine;
(c) A is isomorphic to an algebra term equivalent to U (m] for some
2-element unary algebra U and some integer m 2: 1;
(d) A has a compatible k-regular relation (k 2: 3);
(e) A has a k -ary compatible central relation (k 2: 2);

(f) A has a compatible bounded partial order.
Moreover, if all fundamental operations of A are surjective, then (a),
(b) , or the following condition holds for A :
(c)' A is isomorphic to an algebra term equivalent to (U; f) [m] for
some finite set U (lUI 2: 2), some m 2: 1, and for some permu-
tation group r on u which acts primitively on u.
For the precise definition of a k-regular relation the reader is referred
to [40, 41] or [35]. To be able to follow the rest of the proof it suffices to
know that every k-regular relation is a k-ary totally reflexive, totally
symmetric relation (}which is distinct from the full relation Ak; here
totally reflexive means that f2 contains all n-tuples whose coordinates
are not pairwise distinct, and totally symmetric means that (} is invari-
ant under permuting coordinates. A k-ary central relation is a totally
reflexive, totally symmetric relation (} #- A k such that {c} x A n - l <:;;; f2
for at least one c E A.
314 A. Szendrei
Now we apply Theorem 3.3. In the case when Q:(l) is a transitive

permutation group, every term operation of A is surjective. Therefore
condition (a) , (b) , or (c)' holds for A. In fact , (c)' can hold only if m =
1, because otherwise WAut(A) is not 2-homogeneous by Example 2.12.
If (c)' holds with m = 1, then A is essentially unary, so A satisfies
condition (U) by Example 2.4. If (b) holds, then by Example 2.11 , A
must be term equivalent to RAtr with R = End(KA) as in (A). If (a)
holds, then by Example 2.10, A must satisfy condition (Q) .
It remains to consider the case when A is simple and Q:(l) contains
all constants. The latter condition means that A is term equivalent
to A c. This is a perfect setting for applying tame congruence theory.
In fact , we need only a very small portion of the theory which we
summanze now.
For a finite algebra A and for e E Pol 1 (A) we call e an idempotent
unary polynomial of A if e2 = e, and we call e a minimal idempotent
polynomial of A if it is an idempotent unary polynomial which is not
constant, and has minimal range (with respect to inclusion) among the
nonconstant idempotent unary polynomials of A. 4 The range e(A) of
a minimal idempotent polynomial of A is called a minimal set; if A is
simple, then it can also be called a trace. For any idempotent unary
polynomial e of A we define the induced algebra on the range N = e(A)
of e as follows:
AIN = (N; {epi N : p E Pol( A)}).
If e, e' are idempotent unary polynomials of A , we say that the sets

N = e(A) and N' = e'(A) are polynomially isomorphic if there exist
unary polynomials g, h E Pol 1 (A) such that g maps N onto N' , h maps
N' onto N , and hgi N = idN , ghiN' = idN'; g , hare called polynomial
isomorphisms.
3.4 Theorem. [17] Let A be a finite simple algebra.
(1) For every minimal idempotent polynomial e of A , the induced

algebra AIN on N = e(A) is polynomially equivalent to one of
the following algebras:
4 Notice that in this context the word 'idempotent' is applied to unary oper-
ations only, and has a different meaning than in the definition of an idempotent
operation, as introduced at the beginning of Section 2.
1. a simple unary algebra (N; r) where r ~ SN is a group:

2. a one-dimensional vector space on N;
3. a 2-element Boolean algebra (INI = 2);
4. a 2-element lattice (INI = 2);
5. a 2-element semilattice (INI = 2).
(2) Any two traces of A are polynomially isomorphic.

It is easy to see that iffor two idempotent unary polynomials e, e' of A
the sets N = e(A) and N' = e'(A) are polynomially isomorphic, then
the induced algebras AIN and AIN' have isomorphic clones. Therefore
the type 1-5 of A IN ( N = e( A)) in Theorem 3.4 is independent of
the choice of the minimal idempotent polynomial e of A; this number
is called the type of the finite simple algebra A.
Now let us return to the proof of Theorem 3.1 in the case when A is
simple and Q:(l) contains all constants. We can apply Theorem 3.3 to
A. We will look at cases (a)-(f) separately. If A is quasiprimal, then
A must be primal as Q:(l) contains all constants; hence (0) holds for
A. If A is affine, then <!: = Clo(A) = Pol(A) is the polynomial clone
of a finite simple module; therefore, by Example 2.11, A must be term
equivalent to RAe with R = End(KA) as in (A). Case (c) cannot hold
for A; otherwise our assumption that IAI ~ 3 would imply m > 1,
but then by Example 2.12 the weak automorphism group of U[m] is
not 2-homogeneous. Hence we are left with the cases when one of
conditions (d), (e), or (f) holds for A. Note that in these cases A
cannot be affine or quasiprimal.
Next we show that A cannot have a binary compatible central rela-
tion or a nontrivial compatible partial order. Suppose first that (} is
a binary compatible central relation of A, and select c E A so that
{ c} x A ~ (}. Let e = e2 E Q:(l) be a minimal idempotent polynomial
of A, and let N = e(A). Then e(f2) = f21N is a compatible relation
of the induced algebra AIN such that e(f2) is reflexive and symmet-
ric. Moreover, since { c} x A ~ (}, an application of e shows that
{ e(c)} x e(A) ~ e(f2). Thus e(f2) is either the full relation or it is
a compatible central relation of AIN· However, none of the algebras
listed in Theorem 3.4 (1) admit compatible central relations. There-
fore e(f2) is the full relation for every minimal idempotent polynomial
e of A. Since the weak automorphism group G of A is 2-homogeneous
316 A. Szendrei
and every conjugate 1e ('-y E G) of a minimal idempotent polynomial

e of A is again a minimal idempotent polynomial of A , we conclude
that f2 contains every pair from A. This contradicts our assumption
that f2 is central, and hence proves that A does not have a compatible
binary central relation.
Suppose now that A has a nontrivial compatible partial order, and let
f2 be a minimal such order. As before, for every minimal idempotent
polynomial e of A, e(f2) = f21 N is a compatible relation of the induced
algebra AIN· Clearly, e(f2) = f21N is a partial order on N. Assume
e(f2) = f21 N is trivial (i.e. , the equality relation). Then f2 ~ ker(e);
hence the symmetric transitive closure of f2 is a nontrivial congruence
of A , which is impossible, as A is simple. Thus e(f2) is a nontriv-
ial compatible partial order of AIN· Among the algebras listed in
Theorem 3.4 (1) only some of the 2-element algebras admit nontrivial
compatible partial orders. Hence N is a 2-element set, and the two
elements of N are comparable with respect to f2· Since the weak au-
tomorphism group G of A is 2-homogeneous and every conjugate 'e
(r E G) of a minimal idempotent polynomial e of A is again a mini-
mal idempotent polynomial of A , we conclude that any two elements
of A are comparable with respect to fl. Thus (A ; f2) is a chain. As f2
was chosen to be a minimal nontrivial compatible partial order of A ,
every ' f2 (r E G) is a minimal nontrivial compatible partial order of
A. This implies that ' f2 = f2 or ' f2 = fl- 1 holds for all r E G. Since
(A ; f2) is a chain with more than two elements, G cannot be transitive.
This contradiction proves that A has no nontrivial compatible partial
order.
Thus A has a compatible k-ary regular or central relation f2 with
k ~ 3. Let a = n('fl : r E G). Clearly, a is a k-ary totally reflexive,
totally symmetric compatible relation of A. Moreover, by construc-
tion , 1a = a for all r E G. Let A* denote the algebra with underlying
set A whose operations are the surjective term operations of A and
the permutations from G. The properties of a ensure that a is a
compatible relation of A*. Moreover, by definition, the fundamental
operations of A are surjective, and because of G ~ Clo(A*), the alge-
bra A* is simple and has no proper subalgebras. Therefore conditions
(a) , (b) , or (c)' from Theorem 3.3 hold for A* in place of A . Since A*
has a totally reflexive, totally symmetric compatible relation of arity
~ 3, A* is neither quasi primal nor affine. Thus (c)' must hold for
A*. Since the permutations from G are weak automorphisms of A*,

it follows from Example 2.12 that m = 1, that is, the algebra A* is
essentially unary. The definition of A* shows now that every surjec-
tive term operation of A is essentially unary, that is, the clone of A
is contained in the Slupecki clone. Since all constants are term oper-
ations of A and A is simple, we get that (S) holds for A unless every
unary term operation of A is constant or a permutation. In the latter
case Theorem 3.2, combined with the fact that A is not affine, yields
that A satisfies condition (U) (cf. also Example 2.4). This completes
the proof of Theorem 3.1. 0
Remark 3.5. (1) Every algebra A that satisfies the assumptions of

Theorem 3.1 and condition (I) is functionally complete, and hence is
simple of type 3. To verify that A is functionally complete, observe
that every weak automorphism of A is a weak automorphism of the
algebra Ac as well; now, if we apply Theorem 3.1 to the algebra Ac,
we see that none of the conditions can hold for A c except (0), hence
A c is primal.
(2) Every algebra A that satisfies the assumptions of Theorem 3.1
and condition (S) is of type 1, 2, or 3. This follows from the fact that
every simple algebra of type 4 or 5 has a connected compatible partial
order (see Theorem 5.26 in [17]); however, in the course of proving
Theorem 3.1 we saw that if an algebra A satisfies the assumptions
of the theorem and the additional conditions that A is simple and
CA ~ Clo(A), then A has no nontrivial compatible partial order.
4 SA -closed clones
In this section we give a complete description of the SA-closed clones
on a finite set A. The case IAI = 2 was discussed in Example 2.1,
so we will assume throughout this section that A has at least three
elements. The description of SA-closed clones for IAI = 3 was first
published in [13]. As regards the case IAI ~ 4, in [14] Hoa found the
maximal SA-closed clones and described all SA-closed clones that are
contained in the Slupecki clone; in [15] he announced the completion
of the description of all SA-closed clones. Later Marchenkov [27, 28]
used a purely relational approach to describe all SA-closed clones that
are not contained in the Slupecki clone. The description we present
318 A. Szendrei
here is similar to Hoa's and Marchenkov's. However, the proof we

will outline differs essentially from the earlier proofs: we will combine
arguments with operations and relations, and will make essential use
of some ideas and results of tame congruence theory.
Clearly, SA is a 2-homogeneous group. Therefore we can start our
search for all SA-closed clones by applying Theorem 3.1 and the sub-
sequent discussion on the lattice of G-closed clones to the case G = SA.
It follows from Corollary 2.7 that the two disjoint intervals of clones
of type (U) are 3- or 4-element chains corresponding to the chain of
normal subgroups of SA. For IAI :2: 5 the group SA is not an affine
general linear group, therefore there exist no SA-closed clones of type
(A) or (Q). For IAI = 4 we have SA = AGL(2, 2) , and up to the choice
of 0, there is a unique 2-group A such that SA = AGL(A). In this
case the intervals in (3) are 3-element chains. Therefore there are nine
SA-closed clones of type (A) and one of type (Q) . For IAI = 3 we have
SA = AGL(1, 3), and up to the choice of 0, there is a unique group A
such that SA = AGL(A). In this case each interval in (3) is a single-
ton. Therefore there are three SA -closed clones of type (A) and one of
type (Q). In either case, the following two classes of SA-closed clones
It remain to be described:
(S) It is the clone of a simple algebra (A ; It) such that It is contained

in the Slupecki clone, and contains at least one transformation
that is neither constant nor a permutation; and
(I) It is an idempotent clone such that (A; It) is a simple algebra,

but not affine.
4.1 Theorem. [14] Let A be a finite set with IAI :2: 3. The SA-closed
clones It on A that satisfy condition (S) are the following:
(1) [T] where T is an SA-closed transformation monoid on A (see

Corollary 2. 7) such that CA ~ T C£_ SA U CA;
(2) 113 U [T] where T is as in (1);
(2)* 113* U [T] if IAI is even, where T is as in (1) and the kernel type
of each nonsurjective member ofT consists of even numbers;
(3) ~ U [T] where 2 < m < IAI and Tis as in (1).

The representation of each SA-closed clone of type (S) described in

Theorem 4.1 will be unique if we require T to be the unary part of
the clone, or equivalently, if we require
• in case (2) that T contains all transformations with range of size

at most 2;
• in case (2) * that T contains all transformations with range of

size at most 2 and with kernel type consisting of even numbers;
and
• in case (3) that T contains all transformations with range of size

at most m.
It is easy to see that for two SA-closed clones of type (S) which are
given in this unique form as XU (T] and X' U [T'] (X = 0, Q5 , Q5* , or
9\m) we have XU (T] ~X' U [T'] if and only if X~ X' and T ~ T'.
Sketch of proof of Theorem 4.1. ([20]) It is easy to see that all clones
listed in the theorem are SA-closed and satisfy condition (S). Con-
versely, let ~ be an SA-closed clone satisfying condition (S) , and let
A = (A;~). Since ct is SA-closed, so is the transformation monoid
T = ct(l). Therefore Corollary 2. 7 applies, and shows in particular
that T contains all constants. Hence ct = Pol(A) and T = Pol 1 (A).
If ct is an essentially unary clone, then there is nothing more to prove
to see that (1) holds.
From now on we will assume that ct is not essentially unary. Let r
denote the maximum of the cardinalities of ranges of operations p E ct
that depend on more than one variable. Since ct is contained in the
SJupecki clone and has an operation that depends on more than one
variable we have 2 < r < IAI. We know from condition (S) that A is
simple. Corollary 2. 7 implies that for some positive integers k1 and k2 ,
T = Poh (A) contains all transformations of kernel type ( k1 , k 2 ). Thus
the minimal idempotent polynomials of A have 2-element ranges, and
every 2-element subset of A is a trace. Remark 3.5 implies that A is
of type 1, 2, or 3.
Now let us fix a trace N of A , and following [19] let us call a set
of the form M = f(N, N, ... , N) with f E Pol(A) = ct a multitrace
of A. Furthermore, let m denote the size of the largest multitrace
in A . It is easy to see that m < r. Indeed, if m = 2 then this is
320 A. Szendrei
clear since r ;::: 2. If m > 2 and M = f(N, N , . . . , N) is a multitrace

with IMI = m, then f must depend on more than one variable, so
m = lf(N, N, . . . , N)l < lf(A , A, ... , A) I< r.
We will now show that m = r and every r-element subset of A is
a multitrace. First we argue that every m-element subset of A is
a multitrace. Let M = f(N , . . . , N) (! E <r) be a multitrace of A
such that m = IMI, and let 1r E SA · Then nM = nf(nN, ... , nN).
Since nN has size 2 it is a trace, so there is a polynomial isomorphism
g: N ---+ nN. The polynomial f'(x 1 , . .. , Xn) = nf(g(xi) , ... , g(xn))
witnesses the fact that nM = f'(N, ... , N) is a multitrace. This shows
how to express any subset of size mas a multi trace. It remains to prove
that mf-r. Our argument is based on Yablonskil's Lemma:
4.2 Lemma. {42} Let A be a finite set and let f (x 1 , ... , Xn) be an oper-
ation on A that depends on more than one variable. If If (A, A , ... , A) I
> m for some m > 1, then there exist m-element subsets M 1 , ... , Mn ~
A for which lf(MI , M2, ... , Mn)l > m.
If m < r , then A has a polynomial f such that f depends on more than
one variable, and lf(A , ... , A) I= r > m. By Lemma 4.2 there are m-
element subsets M1 , .. . , Mn such that lf(M1 , M2, . . . , Mn)l > m. But
all m-element subsets are multitraces, so each Mi equals fi(N , ... , N)
for some fi E <r. Hence
f(MI , M2 , ... , Mn) = f(JI(N , ... , N), ... , fn(N, ... , N))
is a multitrace of size larger than m , which is impossible. This proves

our claim that m = r and every r-element subset of A is a multitrace.
Now we are in a position to use the structure theorem for multitraces
of types 1, 2 , and 3 .
4.3 Theorem. [19] If N is a trace of the simple algebra A of type
1,2, or 3 , and M = f(N , ... , N) is a multitrace, then
(1) .71.1 = e (A) for some idempotent unary polynomial e E Pob (A),
and
(2) the induced algebra AIM is term equivalent to
(i) a matrix power (A IN ) (k] if A is of type 1 or 2;

(ii) a primal algebra if A is of type 3.
Suppose first that A is of type 1 or 2. By Theorem 4.3 the size of

every multitrace of A is a power of INI = 2. On the other hand , it
is easy to see that if M = f(N, ... , N) is a multitrace of A, then
so is every subset M' of M. Indeed, Corollary 2.7 and part (1) of
Theorem 4.3 yield that T = Poh (A) contains a unary polynomial h
with h(M) = M', and hence M' = hf(N, ... , N) is a multitrace. Thus
A cannot have a multitrace of size > 2. Hence m = 2. This means
that every operation f E Q:: that depends on more than one variable
has 2-element range, and hence the range is a trace. If A is of type 1,
then such an f cannot exist by [17], Theorem 5.6, Claim 3.
Hence A is of type 2. The arguments in the preceding paragraph
imply that Q:: = (Q:: n !Jh) u [T] . We will prove that Q:: n 9\2 equals s.B
or s.B*, and hence Q:: is one of the clones described in (2) or (2)*. Let
N = { 0, 1}, and let x + y denote a polynomial of A whose restriction
to N is the vector space addition on N. Since traces are polynomially
isomorphic, every operation f' E Q:: n 9\2 with range in a trace N'
has the form pf for an operation f E Q:: n 9\2 with range in N and
a polynomial isomorphism p: N --+ N' , p E T. Furthermore, in the
fourth paragraph of the proof of Theorem 13.5 of [17] it is shown that
any polynomial operation of a type 2 simple algebra A that has range
in a trace N is constructible from unary polynomials of A and from
x+y; in fact, any polynomial is a sum of unary polynomials. Therefore
we have to determine the unary operations in Q:: n 9\2 whose range is
contained inN.
A unary operation f: A --+ N = {0, 1} from Q:: n 9\2 can be thought
of as a characteristic function on A which may be identified with its
support U1 ={a E A: f(a) = 1}. The family S of subsets of A that
are supports of unary operations f: A--+ N = {0, 1} from Q:: n 9\2 has
the following properties:
(i) S contains 0, A , and at least one nonempty proper subset of A .
(ii) S is closed under symmetric difference, EB.
(iii) If U is in S, then every subset V ~ A with lVI = lUI is also in

s.
To verify these properties one can use the following facts: the con-
stant polynomials into N and a minimal idempotent polynomial e
322 A. Szendrei
with e(A) = N belong to Q:; the support off+ g is Uf EB U9 ; and Q: is

SA-closed.
It follows easily from these properties that S contains a set of size 2;
moreover,
• if S contains a 1-element set, then S is the set of all subsets of
A, while
• if S contains a 2-element set but no 1-element set, then IAI is

even and S is the set of all subsets of A of even cardinality.
Thus, in the first case, Q: n 9t2 contains every unary operation f: A ~
N, and the description of Q: given above yields that Q: n 9t2 = ~ '
and hence Q: is one of the clones in (2). In t he second case, a unary
operation f: A ~ N belongs to Q: n 9t2 if and only iff has kernel type
(k 1 , k2 ) where both k1 and k2 are even, whence we get that Q: n 9t2 =
~ * . It follows also from Theorem 2.6 t hat t he kernel type of each
nonsurjective member of Q:(l) consists of even numbers, therefore Q: is
one of the clones in (2)*.
Finally, we prove that if A is of type 3, t hen Q: is as described in (3).
We est ablished earlier that every operation from Q: that depends on
more than one variable has range in a multitrace of size m. Thus
Q: S: 9'-tm U [T]. What remains to show is that 9'-tm S: Q:, which is the
assertion that any operation f: An ~A whose range is of size at most
m is in Q:. Let M be a multitrace of A of size m such that M contains
the range of f, and let N' S: M be a trace of A. Choose polynomials
PI, ... ,Pk E Pol 1 (A) with range inN' which separate the points of A .
(The existence of these polynomials is guaranteed by Theorem 2.8( 4)
of [17].) View p = (p 1 , ... , Pk): A ~ Mk as a polynomial injection of
A into Mk. Since f: An ~ M, we can try to find h E Polkn(AIM)
that allows us to factor f as
An fin (Mk)n ~ M .
The existence of such a factorization depends on the ability to interpo-

late the partial operation f · (pn) - 1 : (Mk) n ~ M by a total operation
h: (Mk)n ~ M that is a polynomial of A IM· We can do t his since
AI M is primal (see Theorem 4.3). Thus, f agrees wit h some polyno-
mial operation of A of the form h · pn. This concludes t he proof of
Theorem 4.1. D
Now we turn to the description of SA-closed clones <!: that satisfy

condition (I). These clones are most conveniently described by data
that involve compatible relations as well as information on how the
operations from <!: restrict to certain subalgebras. Accordingly, the
proof that is outlined here uses arguments with operations as well as
arguments with relations. This combination of methods seems more
effective than restricting to a purely operational or purely relational
approach.
We will call a binary relation on A a cross if it has the form
fl:[b, B; c, C] = ( {b} X C) u (B X { c}) (bE B, c E C)
for some B, C ~ A with IBI, ICI ~ 2. An r x s cross is a cross

fl:[b, B; c, C] such that IBI = r and ICI = s.
For an algebra A, Iso(A) will denote the family of all internal isomor-
phisms of A (see Example 2.10 for the definition). It is easy to see
that a subset B of A is a subalgebra of A if and only if ida E Iso(A).
For convenience we will allow the empty bijection id0 as an internal
isomorphism, and hence the empty set 0 as a subalgebra of A. We
will say that Iso(A) is m-complete if all bijections between subsets
of A of size < m are internal isomorphisms of A (in particular, all
subsets of A of size < m are subalgebras of A). Clearly, Iso(A) is
0-complete for every algebra A, and Iso(A) is 1-complete if and only
if A is idempotent.
Now let<!: be an idempotent SA-closed clone on a finite set A (IAI ~ 3),
and let A= (A;<!:). It is easy to see that in this case
(4.4) IT= Iso(A) is a family of bijections between subsets of A (con-
sidered as binary relations) such that
(0) IT contains id0, idA, and all bijections between 1-element

subsets of A,
(1) IT is closed under intersection,
(2) IT is closed under multiplication and inversion, and
(3) IT is closed under conjugation by permutations from SA.
Applying properties (1) and (3) to ida E IT (B ~A) one can see that
(4.5) if A has a proper subalgebra B of size IBI = k, then every subset
of A of size < k supports a subalgebra of A .
324 A. Szendrei
Moreover,
(4.6) if B, C are k-element subalgebras of A , then conjugation by any

permutation 1 E SA such that r (B) = C yields an isomorphism
Clo(B) ---+ Clo(C) ; in particular, it follows that Clo(B) is Ss-
closed.
We will use the following two parameters associated to A: subA is the

maximum size of a proper subalgebra of A; croA is the largest number
r such that A has a compatible r x 2 cross, if A has such a compatible
cross, and croA = 1 otherwise. Clearly, we have 1 < subA < IAI.
It is also easy to see that croA = IAI or croA < subA . The next
theorem shows that if It = Clo(A) satisfies condition (I), then it is
determined by the following data: the internal isomorphisms of A ,
the parameter croA, and the clone of a 2-element subalgebra of A
(if any). In particular, it follows that there are only finitely many
SA-closed clones that satisfy condition (I).
4.7 Theorem. [20] (cf. [15, 28]) Let A be a finite set with IAI ;::: 3,
and let A = (A; It) where It is an SA -closed clone on A that satisfies
condition (I). Then one of the following conditions holds for A:
(Q) A is quasiprimal;
(P) subA ;::: 2, the 2-element subalgebras of A are term equivalent to

2-element affine spaces, and an operation f is a term operation
of A if and only iff preserves the internal isomorphisms of A
and !Is E Clo(B) for all 2-element subalgebras B of A;
(D) r = croA ;::: 2, Iso(A) is 2-complete if r > 2, and an operation f

is a term operation of A if and only if f preserves the internal
isomorphisms of A and
all r x 2 crosses if r > 2,

all reflexive 2 x 2 crosses if r = 2.
(E) r = croA ;::: 2, Iso(A) is r-complete if r > 2, and an operation f

is a ter-m oper-ation of A if and only if f preserves the internal
isomor-phisms of A and !Is is a pr-ojection for all r-element
subalgebms B of A.
This classification shows, in particular, that if A satisfies the assump-

tions of Theorem 4.7, then A has the dual discriminator (see Exam-
ple 2.9) as a term operation if and only if A is of type (Q) or (D), and
A is paraprimal (see [5]) if and only if A is of type (Q) or (P).
Before outlining the proof of Theorem 4. 7 we will show how this the-
orem can be transformed into an explicit description of all SA-closed
clones satisfying condition (I). First we have to determine all possible
sets Iso(A) of internal isomorphisms, and then list for each such set ][
all SA-closed clones ([:that satisfy conditions (I) and][= Iso( (A;<!::)) .
The next lemma explicitly describes all families ][ of bijections that
satisfy conditions (4.4) (0)- (3) above. The description uses the follow-
ing notation. Bijm(A) denotes the collection of all bijections between
m-element subsets of A (0 < m < IAI). For even IAI, BijfAI/ 2(A) de-
notes the set of all bijections B --+ C between IAI/2-element subsets
B, C of A such that either B = C or B, C are complements of each
other. Finally, for a nontrivial normal subgroup N of SA, Restrm(N)
denotes the set of bijections between m-element subsets of A that are
restrictions of members of N. Clearly, Restr 2(VA) = Bij~(A) if IAI = 4,
Restrm(AA) = Bijm(A) if 1 < m < IAI-2, and Restrm(SA) = Bijm(A)
if 1 < m < IA I - 1.
4.8 Lemma. Let A be a finite set with IAI 2 3 and let IT be a family
of bijections between subsets of A such that IT satisfies conditions (4.4)
(0)-(3). Then IT is one of the following families:
(i) ]k,l = U(Bij 8 (A) : 0 < S < k) U {idE : B ~ A, IBI < l} U {idA}
for some integers k, l with 1 < k < l < IAI;
(ii) (a) ]3,! U U(VB : B ~ A, IBI = 4) for some l with 3 < l < IAI,
(b) ]k,luU(AB: B ~A, IBI = k+1) for some k,l with 2 < k <
l < IAI,
(c) ]k,l U U(SB : B ~A, IBI = k + 1) for some k, l with 1 < k <
l < IAI;
(iii) ]IAI/2- l,l U BijfAI/ 2(A) for some l with IAI/2 < l < IAI, if IAI is
even;
326 A. Szendrei
(b) .JT2,3 U Restr3(VA), if IA I = 4,
(c) .JT1,3 U Bij~ (A) U Rest r3(VA), if IAI = 4;
(v) (a) .JTk,k UN for some k :::: 1 with IA I - 3 < k < IA I and f or
some nontrivial normal subgroup N of SA such that k :::: IAI- 2
if N =SA,
(b) .JTIAI- 2,IAI-1 U RestriAI-1(AA) UN for a nontrivial normal sub-

group N of SA with N ~ AA,
(c) .JT2,3 U Rest r3(VA) U VA , if IAI = 4,

(d) .JT1,3 U Bij~(A) U Restr 3(VA) U VA, if IAI = 4,
(e) .JT1,2 U Bij~(A) U VA , if IAI = 4.
These families of bijections b etween subsets of A are naturally ordered
by inclusion, and form a lattice. Figures 3- 5 show these lattices for
IAI :::: 5, IAI = 4, and IAI = 3, respectively. In the diagrams the trian-
gular array of large bullets represents the families.JTk,t (1 < k < l < IAI)
from (i) ; in particular, the top bullet corresponds to .JTIAI- 1,1AI-1, the
bottom rightmost bullet to .JT 1,1, and the bottom leftmost bullet (at the
right angle of the triangle) to .JT 1 ,1AI-l· T he small bullets and t he circles
represent the families listed in (ii) and (iii) , respectively (type (iii) ex-
ists only if IAI is even); for each pair k, l with k < l these form a chain
between .JTk,t and .JTk+l,t · The families in (iv) are denoted by diamonds
in Figures 3- 5. Finally, the families in (v) , which are distinguished
from those in (i)-(iv) by the property that they contain nonidentity
permutations from S A, are denoted in Figures 3-5 by squares.
The main ingredient of the proof of Lemma 4.8 is to show t hat if I
is a family of bijections between m-element subsets of A for a fixed
integer 1 < m < IAI such t h at I satisfies conditions (4.4)(2)-(3) and
I contains a bijection between two distinct m-element subsets of A,
then the following holds for I:
• Ifm < IAI-1 then either I= Bijm(A) , or IAI is even, m = IA I/2,
and I = Bij[AI/ 2(A). In both cases S 8 ~ I for all m-element
subsets B of A.
• If m = IAI - 1, then there exists a nontrivial normal subgroup

N in SA such t hat I = RestriAI- 1 (N).
:::~
...
.. .. ..
Fig. 3: Lattice of possible families of internal isomorphisms, IAI 2: 5

328 A.. Szendrei
Fig. 4: Lattice of possible families of internal isomorphisms, IAI = 4

Fig. 5: Lattice of possible families of internal isomorphisms, IAI = 3
The next theorem will show that for every family TI of bijections from
Lemma 4.8 there exists an SA-closed clone Q: that satisfies conditions
(I) and TI = Iso( (A; <r)). An explicit list of all such clones is also given
in the theorem. We will use the following notation. If TI = .Jf k,l U ... is
one of the families listed in Lemma 4.8, then we define D(TI) to be the
set of all operations on A that preserve all members of TI (as binary
relations). Clearly, D(TI) is the clone of an idempotent quasiprimal
algebra. Furthermore, let
1' 2 (TI) = {J E D(TI) : JIB E Clo( (B; 1\, v)) for all 2-element
subsets B of A}
= {J E D(TI) : J preserves all reflexive crosses on
2-element subsets of A},
1'r(TI) = {J E D(TI) : J preserves all r x 2 crosses} (3 < r < IAI),
~(TI) = {J E D(TI) : JIB E Clo( (B; x + y + z)) for all
2-element subsets B of A},
<Er(TI) = {J E D(TI): JIB is a projection for all r-element
subsets B of A} (2 < r < IAI).
The two descriptions of 1' 2 (TI) are equivalent, because a reflexive 2 x 2

cross is a partial order on a 2-element subset of A. Note that if k 2 2,
then there is a third description as well, namely
1' 2 (TI) = {J E D(TI): J preserves all 2 x 2 crosses},

330 A. Szendrei
because every 2 x 2 cross is a relational product of a reflexive 2 x 2

cross and a bijection from Bij 2 (A) ~ ll . This means that in the case
when k 2': 2, ~ 2 (ll) can be defined by a condition analogous to the
condition defining ~r(ll) for r 2': 3.
4.9 Theorem. [15, 28, 20] Let A be a finite set with IAI 2': 3, and
let ll = ]k,l U ... be a family of bijections from Lemma 4 .8. The SA-
closed clones <t that satisfy condition (I) and have the property that
ll = Iso( (A; <!:)) are the following.
(1) If k 2': 2, then there are exactly k + l + 1 such clones <t, namely
O(ll), ~r(ll) (r = 2, . .. , l orr = IAI), ~(ll), and <Es(ll) (s =
2, ... ' k).
(2) If l > k = 1, and Ss ~ ll for all 2-element subsets B of A , then

<t is one of the four clones .Q(ll), ~ 2 (ll), ~(ll) , or <E 2 (ll).
(3) If l > k = 1,and ll n Ss = {ids} for all2-element subsets B of

A, then <t equals .Q(ll) or ~ 2 (ll).
(4) Finally, if k = l = 1, then <t must be the clone .Q(ll) .

The clones listed above are pairwise distinct.
In Figures 3-5 the four cases for ll that are distinguished in Theo-
rem 4.9 are indicated as follows . The families ll with k 2': 2 are those
that appear in the triangular region. The families ll with l > k = 1
are inside the rectangular region, and the two subclasses distinguished
by the property "Ss ~ ll for all 2-element subsets B of A" and by the
complementary property "ll n Ss ={ids} for all 2-element subsets B
of A" are separated by a dashed line. The remaining families ll are
those with parameters k = l = 1. Figure 6 shows the clones listed in
Theorem 4.9 for four typical families ll = llj (j = 1, 2, 3, 4). For each j
the family ll = llj represents the case when ll satisfies the assumptions
of part (j) of Theorem 4.9, and the corresponding clones are ordered
by inclusion.
The fact that every SA-closed clone satisfying condition (I) is one
of the clones listed in Theorem 4.9 is immediate from Theorem 4.7.
What is new in Theorem 4.9 is the statement that the clones X(ll)
(X= 0, ~ ' ~n or <Es) appearing in the theorem are pairwise distinct,
and that for every such clone we have ll = Iso( (A; X(ll))). Both of
Fig. 6:
332 A. Szendrei
these statements can be proved by finding enough operations in the

given clones that witness the required properties; for example, the
operations in Example 2.9 can be used to show that :t(JI) =1- :t'(JI) if
:t =1- :t'. In this way we can also prove that for two clones :t(JI) and
:t' (JI') appearing in the theorem we have :t(li) ~ :t' (JI') if and only if
li 2 JI'- whence :t'(JI) ~ :t'(ll') where :t'(JI), too, is one of the clones
appearing in the theorem - and :t(ll) ~ :t'(JI) holds by Figure 6.
This completes the description of SA-closed clones. From the discus-
sions at the beginning of Section 4, after Theorem 4.1, and in the
preceding paragraph we also know the comparability relation between
SA-closed clones that are of the same type (U) , (A) , (Q), (S), or (I).
The comparabilities between SA-closed clones of different types are
easy to determine from the descriptions of the clones. Putting all this
information together, and using Figure 2 as a 'skeleton' one can get a
full description of the lattice of SA-closed clones for each finite set A
with IAI ~ 3.
Outline of proof of Theorem 4-7. ([20]) It is well known and easy to

check that the collection of all compatible relations of an algebra is
closed under the following constructions: intersection, direct product,
permutation of coordinates, and projection onto some of the coordi-
nates. If A is an idempotent algebra, then the family of compatible
relations of A is also closed under substituting an element a E A into
a relation (! = (!(x 1 , . . . , xn) ~ An to yield an (n - 1)-ary relation
(!(XI , ·· ·, xi- l , a, xi ,···, xn- d = {(xi , . . . , Xn- 1): (x l , . . . ,Xi- l,a , xi ,
... , Xn - l) E (!}. Furthermore, if Clo(A) is an SA-closed clone, then
the collection of compatible relations of A is closed under conjugation
by all permutations from SA (cf. Claim 2.3). Therefore, if A satis-
fies the assumptions of Theorem 4. 7, then the collection of compatible
relations of A is closed under all these constructions. We use this
fact throughout the proof when we construct new compatible rela-
tions from given ones. For a relation(!~ An and I~ {1 , .. . , n} , pr1 (!
will denote the projection of(! onto its coordinates in J. The size of(!
is the number max(lpri r!l : 1 < i < n).
We will call a relation reduced if no permutation of coordinates trans-
forms it into a direct product of relations of smaller arity, and no
projection onto two coordinates is a bijection between two sets. In
particular, a binary compatible relation of an algebra is reduced if
and only if it is neither a direct product of two subalgebras, nor an

internal isomorphism. It follows easily from Claim 2.2 that
(4.10) an operation belongs to the clone of A if and only if it pre-

serves the internal isomorphisms of A and the reduced com-
patible relations of A.
Therefore, to prove Theorem 4.7, we have to understand t he reduced

compatible relations of the algebras A = (A; <!:) where <!: is an SA-
closed clone satisfying condition (I). The following theorem helps re-
duce most of our considerations to binary relations.
4.11 Theorem. [39] Let A be a finite idempotent algebra, let n ~ 2,
and let CJ be an n-ary reduced compatible relation of A . For 1 < i < n
let B i denote the subalgebra of A whose universe is pri CJ. Then one
of the following conditions holds:
(i) The algebra A has a reduced binary compatible relation {! of the

same szze as a;
(ii) B 1 , . .. , B n are isomorphic simple affine subalgebras of A.
Now we outline the main steps of the proof of Theorem 4. 7, describing

the key ideas for each step. Let A = (A;<!:) where <!:is an SA-closed
clone on a finite set A (IAI ~ 3) and <!:satisfies condition (I). Combin-
ing property (4.5) with the assumption that A is not affine, and then
Theorem 4.11 with statements (4.6) and (4.10) we get the following.
(4.12) If B is an affine subalgebra of A, then lEI < 2; hence,

(4.13) if A has no binary reduced compatible relation, then (Q) or
(P) holds.
Therefore, from now on, we assume that A has a binary reduced com-
patible relation. Using the existence of such a relation {! and the
constructions described at the beginning of the proof along with some
of the properties of A established earlier, one can prove the following
preparatory claims on subalgebras, compatible crosses, and internal
isomorphisms of A.
(4.14) subA ~ 2.
(4.15) If B <:;;; A is not simple, then B is a projection algebra.
334 A. Szendrei
(4.16) croA 2: 2.
(4.17) If r;;[b, B; c, C] with B = {b, b'} and C = {c, c'} is a 2 x 2 com-
patible cross of A, then the bijection i: B ---+ C, b f-----t c', b' f-----t c
belongs to Iso( A).
(4.18) Every 2-element subset of A is a subalgebra, and either every
2-element subalgebra B of A is term equivalent to the unique
lattice on B, or every 2-element subalgebra B of A is term
equivalent to the unique majority algebra on B, or every 2-
element subalgebra B of A is a projection algebra (cf. Fig-
ure 1).
(4.19) If croA 2: 3, then Iso(A) is 2-complete.
(4.20) If Iso(A) is 2-complete and A has a compatible r x 2 cross,
then all r x 2 crosses are compatible relations of A.
For the proof of (4.14) it suffices to look at the relations o( a, x), o( x, a)

(a E A) and on 'o (r E SA)· To show (4.15) one can apply Theo-
rem 3.1 and claim (4.12) to the subalgebra B whose clone is 5 8 -closed
by (4.6). (4.16) follows by arguing that either on (B X C) is a 2 X 2
cross for some 2-element subsets B, C of A, or one of the relations
o o o- 1 , o- 1 o o is a nontrivial equivalence relation on a subset of A
and hence (4.15) applies. Thus A has a compatible 2 x 2 cross in both
cases. The proof of (4.17) requires slightly different arguments for the
cases B = C, IBnCI = 1, and BnC = 0. To prove (4.18) we consider
a compatible 2 x 2 cross r;; = r;;[b, B; c, C] of A, which exists by (4.16)
and (4.5). With the internal isomorphism i provided by (4.17) we get
that r;; o i - 1 is a reflexive compatible cross of the 2-element subalge-
bra B of A. This fact, combined with (4.5), (4.6), and Example 2.1,
implies (4.18). (4.19) can be derived from (4.17) by observing that if
r;;[b, B; c, C] is a compatible r x 2 cross of A, then the 2 x 2 crosses of
the form on (B' x C) with bE B' ~ B , IE' I = 2, yield enough bijec-
tions between 2-element subsets to conclude by properties (4.4)(2) - (3)
that R = Iso(A) is 2-complete. Finally, (4.20) is a special case of the
following more general fact:
(4.21) If Iso(A) is 2-complete and Tis a binary compatible relation

of A such that pr2 T is a 2-element set, then
(rx <5)T = {('r (x), 6(y)): (x, y) E r}

is a compatible relation of A for all 1 , c5 E SA.
Indeed, (lx<l~ = "~rot- for some bijection t- between 2-element subsets

of A . The next two claims are the core of t he proof of T heorem 4.7;
they are crucial in proving that t he family Iso(A), the parameter croA,
and the 2-element subalgebras of A completely determine the reduced
compatible relations of A.
(4.22) If {! is a reduced binary compatible relation of A then the size

of{! is at most croA; moreover, if{! is not a cross, then croA 2: 3
and the 2-element subalgebras of A are projection algebras.
(4.23) If r = croA 2: 3 and the 2-element subalgebras of A are pro-
jection algebras, then the r-element subalgebras of A are also
projection algebras and Iso(A) is r -complete.
For the proof of (4.22) let B = pr 1 {!, C = pr2 {!, and r = croA. We can
assume without loss of generality that IBI 2: ICI. For any compatible
r x 2 cross K;[d, D; v, V] (dE C, V = { v, v'}) of A, {! o K;[d, D ; v, V] is
a compatible relation of A of the form
A = (Ex {v}) U (E' x {v'}) with E :2 E ' -1- 0

where E = U(Q(x, d') : d' E C n D) and E' = Q(x, d). Were IBI >
r, we could select K;[d, D; v, V] (using conjugation) so that lEI > r
and 1 < IE'I < lEI . Therefore the same argument as in the proof
of (4.19) would imply that Iso(A) is 2-complete. Hence, by (4.21),
an intersection of some relations of the form (lxid)A with r (E) = E
(r E SA) would be a compatible lEI x 2 cross of A. This contradiction
proves the statement on the size of (!. If {! is not a cross then, after
replacing {! with Q- 1 if necessary, we can select a compatible cross
K;[d, D; v, V] (using conjugation) so that d E C , D :2 C , and for the
compatible relation A we have 2 < IE' I < lEI (< lE I) . The size of A
is lEI 2: 3, therefore r = croA 2: lEI 2: 3 by the first part of (4.22).
Thus (4.19)- (4.20) imply that every r x 2 cross is a compatible relation
of A. Hence for any r x 2 cross K;[e, E";v' , V] with e E E \ E' and
E'U {e} s:;; E" , r = AnK;[e, E" ; v', V] is a compatible relation of A such
that T o r - 1 is a nontrivial equivalence relation on the set E ' U { e }.
Thus, by (4.15) and (4.5)-(4.6), all subalgebras of A of size IE'I + 1
are projection algebras.
336 A. Szendrei
To establish (4.23) let us fix a 2-element subalgebra C of A . It suffices

to prove that for every term operation f = f(x 1 , .. . , xn) of A such
that f is projection onto the first variable inC, f is projection onto the
first variable in every r-element subalgebra B of A ; thus the r-element
subalgebras of A are projection algebras, and any bijection between
them or between smaller subalgebras is an internal isomorphism of A .
Let C = { c, c'} and let b1 , .. . , bn be arbitrary elements of B. Since
~[bb B; c, C] is a compatible relation of A by (4.19)- (4.20), and since
f(c', c, . .. , c) = c' , it follows that f(bl , b2, ... , bn) = b1 .
Now we can complete the proof of Theorem 4.7 through the following
three claims:
(4.24) If croA > 2 and the 2-element subalgebras of A are not projec-
tion algebras, then (D) holds.
(4.25) If croA > 2 and the 2-element subalgebras of A are projection
algebras, then (E) holds.
(4.26) (D) or (E) holds also if croA = 2.
Let r = croA. In all three claims (4.24)- (4.26), the statements in

(D) or (E) , respectively, about croA and Iso(A) , and the necessity of
the conditions for f to be a term operation of A follow from (4.16) ,
(4.19) , (4.23), and (4.18) , (4.20), (4.23). It remains to check in each
case that every operation f satisfying the conditions described in (D)
or (E), respectively, is a term operation of A. First assume that r 2:: 3
and that the 2-element subalgebras of A are not projection algebras.
Then A has a term operation g that is a majority operation on some
2-element subalgebra of A. By (4.19) Iso(A) is 2-complete, therefore
g is a majority operation throughout A. (4.22) shows that every
reduced binary compatible relation of A is a cross. Combining this
with statements (4.19)- (4.20) and the fact that every r 1 x r 2 cross is a
relational product of an r 1 x 2 and a 2 x r 2 cross one can see that every
operation f that satisfies the conditions described in (D) preserves all
reduced binary compatible relation of A. The following theorem of
Baker and Pixley, modified in the spirit of claim (4.10) , shows that f
is a term operation of A:
4.27 Theorem. [1] If a finite algebra A has a majority term oper-
ation, then an operation on A is a term operation of A if and only
if it preserves the internal isomorphisms of A and the binary reduced
compatible relations of A.
If r ~ 3 and the 2-element subalgebras of A are projection algebras,
then Theorem 4.11 and (4.22) yield that every reduced compatible
relation of A has size < r. All such relations are preserved by every
operation f that satisfies the conditions described in (E). Thus it fol-
lows immediately from (4.10) that f is a term operation of A. Finally,
if r = 2, then we know from Theorem 4.11 and (4.22) that every re-
duced compatible relation a of A has size < 2. It is not hard to prove
by induction on the arity n of a, and using (4.17), that for any indices
1 < i < j < n there exist internal isomorphisms '"ij: pri a --t prj a.
Therefore a arises from a compatible relation a' of the 2-element sub-
algebra B of A on B = pr1 a by applying the internal isomorphisms
t 12 , ... , t 1n in the 2nd, ... , n-th coordinates. Thus every operation f
that satisfies the condit ions described in (D) or (E), respectively, for
r = 2 preserves a. This holds for every reduced compatible relation
a, so we get from (4.10) that f is a term operation of A. D
In conclusion we note that the description of SA-closed clones was

extended by Marchenkov to AA-closed clones. For JAI = 3 there are
2No AA-closed clones (see [24]) , but for JAI ~ 4 there are only finitely
many (see [29, 30, 31, 32]), and their description is very sili1ilar to
the description of the SA-closed clones. In fact, it turns out that if
JAI is not divisible by 4, then every AA-closed clone is in fact S A-
closed. If JAI is divisible by 4, then there exist AA-closed clones ~
satisfying condition (I) that are not SA-closed, because there exist
AA-closed families of internal isomorphisms that are not SA-closed.
Such families of internal isomorphisms exist for the following reason:
if JAI is divisible by 4, then BijfA I/ 2 (A) can be partitioned into two
subsets of equal size so that both families are AA-closed and satisfy
(4.4)(2), but they are not SA-closed.
5 Finiteness of the lattice of G-closed clones

The results discussed in the preceding section show that the lattice of
G-closed clones on a finite universe A is finite if G = SA or if JAI ~ 4
and G = AA· One wonders: Are there any other permutation groups
G for which the lattice of G-closed clones is finite? At the present time
the answer to this question is not known. However, the main theorem
338 A. Szendrei
of this section shows that there are very few possible candidates for
such a G. This theorem determines all groups G for which the number
of G-closed clones that contain all constants is finite.
5.1 Theorem. [20] For a permutation group G ~ SA where A is

finite, IAI 2: 3, the following conditions are equivalent:
(i) The lattice of G-closed clones that contain all constants is finite.
(ii) G satisfies the following combinatorial condition:
(*)for every integerk (2 < k < IAI), for every (k+1)-element

subset B of A , and for every k-element subset C of B,
there exists a k-element subset C' of B such that C' # C
and C' = '!'(C) for some ')' E G.
(iii) G is one of the following groups: SA, AA, AGL(1 , 5) (IAI = 5),
PSL(2, 5) (lA I = 6), PGL(2, 5) (IAI = 6), PGL(2, 7) (IAI = 8),
PGL(2, 8) (IAI = 9), PfL(2, 8) (IAI = 9) .
Outline of proof. To prove the implication (i)::::}(ii) we assume that
(ii) fails, and construct infinitely many G-closed clones that contain
all constants. To witness the failure of (ii) we fix an integer k (2 <
k < lA I), a k-element subset C of A , and a (k + 1)-element subset B =
Cu{O} of A so that '!'(C) =J- C' for every 1' E G and for every k-element
subset C' of B that contains 0. For notational simplicity let us assume
that C = {1, 2, ... , k }. Using these sets we construct an infinite se-
quence f2n (n = 2, 3, .... ) of relations and an infinite sequence fn (n =
3, 4, ... ) of operations on A as follows : f2n is the (n + k- 1)-ary relation
consisting of all tuples (x 1 , ... ,xn, y1 , ... , yk-l) E An+k-l such that
l{xl, ... ,Xn,Yl,···,Yk- dl < k+1 and if{xl , . . . ,Xn, Yl , ···,Yk- d =
'!'(C) for some ')' E G then I{x 1 , .. . , Xn} I 2: 2; f n is the n-ary opera-
tion on A such that fn(1 , ... , 1, 0, 1, ... , 1) = 1 (with 0 occurring in
the j-th position) for every j (1 < j < n) , fn(c , . .. , c) = c for all
c = 2, ... , k, and fn(x 1 , ... ,xn) = 0 for all remaining arguments. It is
not hard to check that
• every f2n is reflexive and satisfies 'Yf2n = f2n for all')' E G; therefore
the clone lt(f2n) consisting of all operations that preserve the
relation f2n is G-closed and contains all constants;
• Q.:(t?n- d 2 Q.:(gn) for all n 2: 3;

• the inclusions are proper, as fn E Q.:(t?n-d \ Q.:(gn) for all n 2: 3.
The proof of the implication (ii)::::}(iii) is purely group theoretical. It
is clear that condition (ii) holds for G if G is k-homogeneous for every
k (1 < k < JAJ) . Therefore the implication (ii)::::}(iii) generalizes the
following classical theorem:
5.2 Theorem. [2] The permutation groups on a finite set A (JAI 2: 3)
that are k-homogeneous for every k (1 < k < JAI) are the following:
SA, AA, AGL(1, 5) (JAI = 5), PGL(2, 5) (JAI = 6), PGL(2, 8) (JAI =
9), and PfL(2, 8) (JAJ = 9).
If G satisfies the weaker assumption (*), then condition (*) for k = 2
immediately implies that G must be 2-homogeneous. It is not hard to
argue that condition (*) fails for the group AG L( d, p) unless pd < 5.
This implies that every affine 2-homogeneous group G that satisfies
(*) is listed in (iii). For almost simple 2-transitive groups G we use
the classification theorem to prove that (*) fails unless G is listed in
(iii)
0
Finally, to establish the implication (iii)::::}(i) we extend the method of

proof of Theorem 4.1 presented in the previous section from the case
G =SA to all groups G listed in (iii). D
References
[1] K. A. Baker, A. F. Pixley, Polynomial interpolation and the

Chinese remainder theorem for algebraic systems, Math. Z. 143
(1975), 165- 174.
[2] R. A. Beaumont, R. P. Peterson, Set transitive permutation

groups, Canad. J. Math. 7 (1955) , 35- 42.
[3] G. A. Burle, Classes of k-valued logic which contain all functions

of a single variable (Russian), Diskret. Analiz (Novosibirsk) 10
(1967), 3- 7.
[4] P. J. Cameron, Permutation Groups, London Mathematical So-

ciety Student Texts 45, Cambridge University Press, Cambridge,
1999.
340 A. Szendrei
[5] D. M. Clark, P. H. Krauss, Para primal algebras, Algebra Uni-

versalis 6 (1976), no. 2, 165- 192.
[6] B. Csakany, T. Gavalcova, Finite homogeneous algebras. I, Acta

Sci. Math. (Szeged) 42 (1980), 57-65.
[7] K. Denecke, S. L. Wismath, Galois connenctions and complete

sublattices, Galois Connections and Applications (Editors: K. De-
necke, M. Erne, S. L. Wismath), Kluwer, 2003.
[8] K. G!azek, St. Niwczyk, Q-Independence and weak automor-

phisms, Galois Connections and Applications (Editors: K. De-
necke, M. Erne, S. L. Wismath), Kluwer, 2003.
[9] A. Goetz, On weak isomorphisms and weak homomorphisms of

abstract algebras, Colloq. Math. 14 (1966), 163- 167.
(10] M. Goldstern, S. Shelah, Clones on regular cardinals, Fund. Math.

173 (2002), no. 1, 1- 20.
[11] V. V. Gorlov, R. Poschel, Clones closed with respect to permu-

tation groups or transformation semigroups, Beitdige Algebra
Geom. 39 (1998), 181-204.
[12] L. Haddad, I. G. Rosenberg, Finite clones containing all permu-

tations, Canad. J. Math. 46 (1994) , 951- 970.
[13] Nguen Van Khoa [Nguen Van Hoa], On the structure of self-dual
closed classes of three-valued logic P 3 (Russian), Diskret. Mat. 4
(1992), no. 4, 82- 95.
[14] Nguen Van Khoa, Families of closed classes of k-valued logic that
are pr·eserved by all automorphisms (Russian), Diskret. Mat. 5
(1993) , no. 4, 87- 108.
[15] Nguen Van Khoa, Description of closed classes that are preserved
by all inner automorphisms of k-valued logic (Russian) , Dokl.
Akad. Nauk Belarusi 38 (1994), no. 3, 16- 19, 122.
[16] Nguen Van Khoa, On closed classes of k- valued logic that are self-
dual with respect to transitive groups (Russian), Diskret. Mat.
8 (1996), no. 1, 129--159; translation in: Discrete Math. Appl.
6 (1996), no. 2, 179-206; announcement of results in: On the

structure of closed classes of k-valued logic that are self-dual with
respect to transitive groups (Russian), Dokl. Akad. Nauk Belarusi
38 (1994), no. 6, 17- 20, 121 (1995).
[17] D. Hobby, R. McKenzie, The Structure of Finite Algebras, Con-

temporary Mathematics 76, American Mathematical Society,
Providence, RI, 1988.
[18] B. Huppert, N. Blackburn, Finite Groups III, Grundlehren

der Mathematischen Wissenschaften, vol. 243, Springer-Verlag,
Berlin-New York , 1982.
[19] K. A. Kearnes, E. W. Kiss, M. Valeriote, Minimal sets and vari-

eties, Trans. Amer. Math. Soc. 350 (1998) , no. 1, 1- 41.
[20] K. A. Kearnes, A. Szendrei, Clones closed under conjugation, in

preparation.
[21] I. Levi, R. B. McFadden , Sn-normal semigroups, Proc. Edinburgh

Math. Soc. (2) 37 (1994), no. 3, 471- 476.
[22] A. I. Mal'tsev, A strengthening of the theorems of Slupecki and

Yablonski£ (Russian) , Algebra i Logika 6 (1967) , 61- 75.
[23] S. S. Marchenkov , Homogeneous algebras (Russian) , Problemy

Kibernet. 39 (1982), 85- 106; announcement of results in: On
homogeneous algebras (Russian) , Dokl. Akad. Nauk SSSR 256
(1981), no. 4, 787- 790.
[24] S. S. Marchenkov, Closed classes of self-dual functions of many-

valued logic. II (Russian) , Problemy Kibernet. 40 (1983), 261-
266.
[25] S. S. Marchenkov, Classification of algebras with an alternating

group of automorphisms (Russian), Mathematical problems in cy-
bernetics, No. 2, 100- 122, " Nauka", Moscow, 1989; announce-
ment of results in: On the classification of algebras whose auto-
morphism group is the alternating group (Russian) , Dokl. Akad.
Nauk SSSR 265 (1982), no . 3, 533- 536.
342 A. Szendrei
[26] S. S. Marchenkov, G-precomplete classes of many-valued logic

(Russian), Diskretn. Anal. Issled. Oper. 3 (1996), no. 3, 47- 70,
111.
[27] S. S. Marchenkov, Basic relations for the S-classification of func-

tions of multivalued logic (Russian), Diskret. Mat. 8 (1996), no.
1, 99- 128; translation in: Discrete Math. Appl. 6 (1996) , no. 2,
149- 178.
[28] S. S. Marchenkov, S-classification of functions of many-valued

logic (Russian) , Diskret. Mat. 9 (1997) , no. 3, 125- 152; trans-
lation in: Discrete Math. Appl. 7 (1997), no. 4, 353- 381; an-
nouncement of results in: S -classification of idempotent algebras
with finite support (Russian), Dokl. Akad. Nauk 348 (1996), no.
5, 587- 589.
[29] S. S. Marchenkov, A-classification of finite injective functions

(Russian), Diskretn. Anal. Issled. Oper. Ser. 1 4 (1997), no. 2,
15- 42, 115.
[30] S. S. Marchenkov, A-closed classes of idempotent functions of

many-valued logic that are defined by two-place relations (Rus-
sian), Diskretn. Anal. Issled. Oper. Ser. 1 5 (1998), no. 1, 32- 59,
104.
[31] S. S. Marchenkov, A-closed classes of many-valued logic that con-

tain constants (Russian), Diskret. Mat. 10 (1998) , no. 3, 10-26;
translation in: Discrete Math. Appl. 8 (1998), no. 4, 357-374.
[32] S. S. Marchenkov, A-classification of idempotent functions of

many-valued logic (Russian), Diskretn. Anal. Issled. Oper. Ser.
1 6 (1999), no. 1, 19--43, 97; announcement of results of [31], [32]
in: Dokl. Akad. Nauk 366 (1999), no. 4, 455- 457.
[33] P. P. Pa1fy, Unary polynomials in algebras I, Algebra Universalis

18 (1984) , 262- 273.
[34] R. Poschel, Galois connections for operations and relations, Ga-

lois Connections and Applications (Editors: K. Denecke, M. Erne,
S. L. Wismath), Kluwer, 2003.
[35] R. Poschel, L. A. Kaluznin, Funktionen- und Relationenalge-

bren, VEB Deutscher Verlag der Wissenschaften, Berlin, 1979;
Birkhauser Verlag, Basel and Stuttgart, 1979.
[36] E. L. Post, The Two- Valued Iterative Systems of Mathematical

Logic, Ann. Math. Studies 5, Princeton University Press, Prince-
ton, N.J., 1941.
[37] J. Slupecki, Completeness criterion for systems of many-valued

propositional logics, C. R. Seances Soc. Sci. Lettres Varsovie Cl.
III 32 (1939), 102- 109. English translation: Studia Logica 30
(1972), 153- 157.
[38] L. Szabo, Algebras that are simple with weak automorphisms,

Algebra Universalis 42 (1999), 205- 233.
[39] A. Szendrei, Clones in Universal Algebra, Seminaire de

Mathematiques Superieures, vol. 99, Les Presses de l'Universite
de Montreal, Montreal, 1986.
[40] A. Szendrei, Simple surjective algebras having no proper subalge-

bras, J . Austral. Math. Soc. Ser. A 48 (1990) , 434- 454.
[41] A. Szendrei, Strongly Abelian minimal varieties, Acta Sci. Math.

59 (1994), 25- 42.
[42] S. V. Yablonski}, Functional constructions in a k-valued logic

(Russian), Trudy Mat. Inst . Steklov. 51 (1958), 5- 142.
Author's address:
Agnes Szendrei
Bolyai Institute
University of Szeged
Aradi vertanuk t ere 1
H-6720 Szeged, Hungary
email: A. Szendrei @math.u-szeged. hu
Galois Connections for Partial
Algebras
P. Burmeister
Abstract
In connection with partial algebras one has many more relevant polarities
(i.e. Galois connections induced by binary relations) than in the case of
total algebras. On one side there are many different subsets of the set of
first order formulas, which one wants to use as a concept of identity in
some special context, and where one is interested in the closure operators
induced by restricting the validity of first order formulas to this special sub-
set. On the other hand the polarity induced by the reflection of formulas
by mappings allows us to keep track many interesting properties of homo-
morphisms between partial algebras, while others can be related to these
via factorization systems - which can be considered as special pairs of cor-
responding closed classes (in Formal Concept Analysis one would call such
pairs "formal concepts") of the polarity induced by the (unique) diagonal-
fill-in property on the class of all homomorphisms. - Moreover, having
an interesting set of properties of homomorphisms, the relation "a homo-
morphism has a property" can be used to apply the method of attribute
exploration from Formal Concept Analysis in order to elaborate a basis for
all implications among these properties and on the other hand a small but
"complete" set of counterexamples against all non-valid implications.
In this note we want to describe some' of these polarities or correspond-
ing pairs of interest in them, and we shall present them in the context of
many-sorted partial algebras, since this context seems to be less known.
Moreover, we want to give an example of an attribute exploration as men-
tioned above.
AMS Mathematics Subject Classification: 08A55, 03B10.
Key words: Partial algebra, Polarity, Formal context.
345
346 P. Burmeister
1 Introduction
When we speak in this survey article about a Galois connection, we

mean a polarity (t, +) in the sense of [Bir67] , i.e. a Galois connection
induced by a binary relation I between two sets G and M: 1
Let I ~ G x M, and for A ~ G and for B ~ M define
At := { rn E M I (g , rn) E I for all g E A } , (1.1)
and
B.J. := { g E G I (g, rn) E I for all rn E B } . (1.2)
Then (t, +) forms a polarity. In Formal Concept Analysis (FCA for
short) the triplet (G, M , I) is called a (formal) context- the funda-
mental structure of FCA - , and a pair (A, B) with A ~ G , B ~ M
and with At = B and B+ = A is called a (formal) concept, A is
called its extent and B its intent, and we shall adopt here this way of
speaking.
There are in Universal Algebra and therefore also in the theory of par-
tial algebras important but more or less "trivial Galois connections"
arising in connection with closure systems like those of all closed sub-
sets or of all congruence relations, since every closure system <!: on some
set A can be considered to be induced by the relation RE ~ A x <!:,
where, for a E A and H E <!:, one has:
(a, H) E RE if and only if aE H.

However, there are two main sources for applications of Galois connec-
tions occurring to us immediately, when we think of - many-sorted
-- partial algebras of a given type or signature I:= (S, D, T, rJ , a): 2
Model theoretic polarities in connection with identities:
The relation f= of validity of a first order formula in a (many-sorted)
partial algebra for some given signature I: can be restricted to subsets
:F ~ .c~ (Y) of special interest within the first order language .CdY)
with equality which will here be interpreted as existence equality 3 -
1Since they could even be proper classes, we shall assume the existence of set
theoretical universes, and that all will take place in such a universe.
2 For more detailed definitions of some of the basic notions concerning many-
sorted partial algebras see the next section.

3 See [B86], [B92] or [B93] or below.
Galois Connections for Partial Algebras 347
with some countably infinite S-set Y of variables. In each case it gives

rise to a Galois connection, and in the case of partial algebras there
are many interesting sets F ~ .Cy:, (Y), in particular all (existence)
equations (X; t ~ t') (X~ Y finite , and t , t' any terms using variables
only from X), all weak equations (X ; t ~ t') := (X; t ~ t 1\ t' e t' ::::}
t e t'), strong equations (X ; t :b t') : (X ; (t ~ t ::::} t ~ t') 1\ (t' ~
t' ::::} t e t')), all ECE-equations 4 (X; /\~ 1 ti ~ ti ::::} t ~ t') , or all
quasi-existence equations (X ; /\~=I ti ~ t~ ::::} t ~ t') . But one can
also consider all so-called regular strong equations, which are strong
equations, where both t and t' are definitely using the same variables
in their inductive construction. We shall discuss some of the closure
operators connected with such sets.
The classification of properties of homomorphisms:
The classification and investigation of homomorphisms between par-
tial algebras yield another range of applications of some special polar-
ities:
Defining properties through the reflection of formulas:
A homomorphism from a partial algebra A into a partial algebra 1ffi is
just an S-mapping between the carrier sets, which preserves 5 all exis-
tence equations. However, in general it does not reflect any existence
equation - except for ({x};x ~ x) for variablEs x -,even not any of
the form (X; t ~ t) for some proper term t - having the meaning that
the interpretation of the term t exists - , which is always reflected in
the case of total algebras, and this is one source for many properties
of homomorphisms, which are of interest in the case of partial alge-
bras. When we denote by <J the relation of reflection of a formula by
a mapping, this means that such a property can be related to a formal
concept of the formal context of the polarity induced by the relation
<J.
Defining "epimorphic properties" as extents of factorization
systems:
"Reflection of formulas" allows us to describe properties of mappings
between partial algebras like homomorphisms, injective homomorph-
isms, closed homomorphisms, initial homomorphisms, etc. , but it does
not yet allow us to characterize surjectivity, epimorphy and a wide
4 Short for existentially conditioned existence equations.
5 See the next section. Observe that preserving a formula is equivalent to re-
flecting its negation.
348 P. Burmeister
range of other "epimorphic" properties. For a characterization of such

properties one can use the polarity induced by the existence of the
unique diagonal-fill-in- denoted in this note by 0 - , which is are-
lation between the class HomE of all homomorphisms between partial
algebras of a given signature and itself. Some special formal concepts
of the formal context (HomE , HomE , 0) are called factorization sys-
tems. And if the intent of such a factorization system (i.e. its right
hand component) corresponds to a class of homomorphisms defined
by the reflection of some set of formulas , its extent (i.e. its "left hand
partner") will usually be a class of homomorphisms corresponding to
one of the "missing epimorphic properties".
Investigation of interdependencies of properties of homomor-
phisms using "attribute exploration":
Having many interesting properties of homomorphisms around, some
of which are already combinations of other ones, one is also interested
in all possible combinations of them. Here methods from FCA can
be very useful. The main tool in FCA applicable for the investiga-
tion of the interdependence of the properties of homomorphisms (or
of their combinations) is the so-called attribute exploration. We shall
briefly explain in this note the method of attribute exploration, which
is based on the additional polarity induced by the relation of satis-
faction of an attribute implication by a homomorphism, and we shall
present an example.
In order to avoid set theoretical difficulties we shall assume - as
already mentioned - that all our considerations take place in a set
theoretical universe which is itself a set. This will indeed allow us in
particular to consider factorization systems as formal concepts of the
corresponding polarity, as we already indicated above.
We present our observations for many-sorted partial algebras, since
this context seems to be less known.
2 Some basic definitions
2.1 Fundamentals of the theory of partial algebras

A signature 6 ~ = (S, n, T , ,.,, a) consists of
We only present here the most fundamental concepts of the theory of partial
6
algebras needed in this note; for more details cf. [B86J , [B93J or in the internet
- a non-empty set S, the elements of which are interpreted as sorts,

- a set D of operation symbols,
- an arity function T : D ---+ N0 , which assigns to each operation
symbol wED a non-negative integer T(w), the arity of w;
- a mapping TJ : D ---+ S* := u~=O sn
assigning to each operation
symbol w E D a sequence TJ(w) =: (s 1 , ... S7 (w)) (of length T(w)) of
input sorts,
- and a mapping a : D ---+ S assigning to each operation symbol w E D
its output sort a(w). 7
A partial algebra A:= (A , (wA)wEn) of signature ~ is then an ordered
pair consisting of a so-called S-set A := (As)sES as its carrier set,
where, for s E S, As is called the carrier or phylum of sort s of
A; and, for each w E D, wA : Ary(w) 2 domwA ---+ Aa(w) - with
A7J(w) := Asl X . .. X Asr(w) for TJ(w) =: (sl , ... S'T(w)) - is a partial
operation on A, the fundam ental operation of type (TJ(w), a(w)) of A
corresponding to the operation symbol w. The fundamental operation
wA is called total, iff dom wA = A7J(w), and A is called a total algebra, iff
each fundamental operation of A is total. If TJ(w) is the empty word,
then wA is either empty or total, and then it just fixes an element from
Aa(w), which we call the fundamental constant of A induced by w.
By PAig~ we denote the class of all partial algebras of signature ~'
and by TAig~ we denote its subclass of all total algebras of signature
~-
In the following let Y = (Ys)sES be an S-set , where the elements of Ys
are called variables of sort s (for s E S). When we speak of a global
S -set, say Y, of variables we shall always assume that each phylum
Ys is (at least) countably infinite, and that the phyla are mutually
disjoint and disjoint from D. For any S-set U = (Us)s ES we denote
by '-lJ(U) the set of all S-subsets V = (Vs)sES of U (i.e. where one has
Vs ~ Us for each s E S). V will be called a finite S-subset of U, if the
disjoint union over all phyla of V is finite. By '.P!in (U) we designate
the set of all finite S-subsets of U. In the following, X will always
denote a finite S-subset of the set Y of variables under consideration.
Terms - of some sort s - are defined in the usual recursive way.
[BOO].
One often considers the pair (ry(w), u(w)) as the value of one single function
7
~ then mostly also denoted by 17, and one omits t he arity function T, which is
implicit in our ry, but we think that the above notation is more convenient.
350 P. Burmeister
By TI;(X)s we designate the set of all terms with output sort s E S

and with variables in X E s,pfin(Y) . For t E TI;(X)s we denote by
var(t) ~ X the set of variables "really occurring in t" (because of the
recursive definition oft) . Moreover, by 'JI'E(X) we designate the total
term algebra on TI;(X) := (TE(X)s)sES , where w1I'E(X)(t1, ... , t 7 (w)) :=
wh ... t 7 (w) (as a word in (D U UsES Xs)*).
It is well-known, that one has to include for the definition of iden-
tities in a first order language for a many-sorted signature in some
way a reference to the variables under consideration (one actually
needs the set of sorts to which the referenced variables belong) , when-
ever one does not exclude empty phyla, but such an exclusion would
usually exclude too many structures. For simplicity we use such
a reference for all formulas, i.e. we define our first order language
LI;(Y) := U{ .CI;(X) I X E s,pfin(Y)} as follows:
For X E s,pfin(Y), s E Sand t, t' E TI;(X)s , (X ; t e t') is an atomic
formula, which we call an existence equation (E-equation for short) .8
The special case (X; t ~ t) gets meaning in the case of partial alge-
bras and is called a term existence statement (abbr. TE-statement).
Arbitrary first order formulas are then defined recursively - almost
- as usual:
• each atomic formula of .CE(Y) is a formula of .CE(Y);
• if (X; ) is a formula of .CE(Y), then (X; •) is a formula of

.CE(Y) (negation) ;
• if (X; ) and (X'; ') are formulas of .CE(Y), then
(XU X'; ( 1\ ')) (conjunction),
(X U X'; ( V ')) (disjunction),
(XU X'; (::::} ')) (implication) , and
(XU X'; ({::} ')) (equivalence) are formulas of .CI;(Y) ;
• if (X; ) is a formula of LI;(Y), and if x E Xs for some s E S ,

then 9 (X\ {x};(Vx)) and (X\ {x};(:3x)) are formulas of
.CE(Y).
8 One could avoid the extra reference to variables, if (X ; t ~ t') were replaced
by 1\xEX x ~ x =? t ~ t' (cf. Definition 4 below)- since one only has to reference
some extra variables in order to get full expressive power of the language.
9 Observe that we abbreviate by X\ {x}, for x E X 8 , the S-set Z , for which
Zs = Xs \ {x} and Zs' = Xs' for s' E S \ {s}.

• if (X; ) is a formula of L.:dY) , and if X' E ~fin(Y) , then (XU

X'; ) is a formula of L.:dY).
Formulas of special interest in connection with identities for partial
algebras are besides E-equations (as usual we omit some brackets,
whenever possible):
(a) existentially conditioned existence equations (ECE-equations for
short)
(X; /\~ 1 ti _::_ ti =* t ~ t') ,
(b) quasi-existence equations ( QE-equations for short)
(X; A:~ l ti _::_ t~ =* t _::_ t') '
(c) or (conjunctions of) ECE-equations of some very special kind like
(ci) weak equations, i.e. ECE-equations of the form
(X; t ~ t') :- (X ; t ~ t 1\ t' ~ t' =* t _::_ t') ,
(c2 ) strong equations or KLEENE-equations
(X; t ::b t') : (X ; (t ~ t =* t e t') 1\ (t' e t' =* t ~ t')) ,
(c 3 ) regular· strong equations (X; t s t') , where - because of their
inductive construction - t and t' have the same set of variables, and
this is equal to X, i.e. where var(t) = var(t') =X.
(d) As a generalization of QE-equations we speak of elementary im-
plications, when we allow arbitrarily long (even infinite) premises and
conclusions in an extended infinitary language:
( Z; 1\iEI ti e t~ =* 1\jEJ Pi ~ P~ ) ,
where Z may be an arbitrarily large S-set of variables (with mutually
disjoint phyla all disjoint from D) .
In the following we shall always consider a fixed signature I: and a
fixed S-set Y of variables, countably infinite (in every phylum) , and
disjoint with D; and all partial algebras are assumed to be of this
signature (if not stated otherwise) . For partial algebras A or lH\ etc. ,
A and B etc. shall always designate their carrier sets, respectively.
The basis for the semantics is (partial) interpretations :
Let A be a partial algebra, X E ~fin (Y) , and v : X ----+ A any S-
mapping (i.e. any S-indexed family (vs : Xs ----+ As)sES of mappings) ,
called an X -valuation. 10 Then the (partial) interpr·etation induced by
v, denoted by v, is the mapping out of TdX) into A with smallest
domain dom v ~ TdX) such that
10 Since A may have empty phyla , one should not use only "global" valuations,
since there might exist none, while there may be lots of "local" valuations.
352 P. Burmeister
• V8 (y) = V 8 (y) for ally E Xs and s E S .
• For w E n with (7J(w) , a(w)) =: ((sl , . .. 'Br(w)) , s), and forti E

Tr;(X)s; (1 ~ i ~ T(w)) one has: lfiis;(ti) = : ai is already defined
for 1 ~ i ~ T(w) , and if wA(a 1, .. . , ar(w)) =: a is defined in A,
then ii 8 (wt1 ... tr(w)) is defined with value a.
We say that an E-equation (X; t ~ t') (of sorts) is satisfied in A with

respect to the valuation v - in symbols: A f= (X ; t ~ t')[v] - , iff t E
dom V8 and t' E dom V8 and ii 8 (t) = ii 8 (t')Y We say that (X; t ~ t')
is valid in A - in symbols: A f= (X; t ~ t') - ,iff A f= (X ; t ~ t')[v]
for all valuations v : X ---+ A. As usual, satisfaction and validity are
carried over recursively to arbitrary formulas of .CdY) .
Let A and Ill\ be partial algebras, let (X; ci>) E .Cr; (Y) be any formula,
and let f :A---+ B be any S-mapping. We say that f reflects (X; ci>)
- in symbols: f <J (X; ci>) - , iff, for all X-valuations v : X ---+ A,
Ill\ f= (X; ci>)[f o v] implies A f= (X; ci>)[v]. Conversely, we say that f
preserves (X;ci>), iff for all X-valuations v: X---+ A, A f= (X;ci>)[v]
implies Ill\ f= (X; ci>)[f o v]; however, this is equivalent to f <J (X; -,ci>),
and therefore the relation <J is sufficient.
Let A be any partial algebra and B any S-subset of A . Then B is said
to be a closed subset of A, iff, for every w E n, and for every sequence
Q E domwA n B 71 (w) one has wA(g_) E Bu(w). If B
is a closed subset of A, then (B, (w AIB 11 (w))wEn) will be called the
subalgebra of A with carrier B , and it will be denoted by Ill\. By CAM
we shall designate the smallest closed subset of A containing M <;; A
- and by (;_AM the corresponding subalgebra. - Observe, that in
cases where a subset of the carrier of some partial algebra is defined
by some operator, then underlining the operator means formation of
the relative subalgebra 12 on the defined subset (as in the case (;_AM).
A subset D of the carrier of A is called an initial segment of A, iff, for
every (a1, ... , ar(w)) E domwA, the fact that wA(a1, ... , ar(w)) E Du(w)
implies ai E D 11 (w)(i) for 1 ~ i ~ T(w). By .!-AM we shall designate
11 Observe that, fort = t' , A f= (X; t ~ t)[v] still has the nontrivial meaning that
"t E dom v8 " , i.e. that t is interpreted w.r.t. v (or, in other words, v interprets t).
12 Note that, for an arbitrary subset B of A one defines the rela-
tive subalgebra lB := B of A with carrier B to be the partial algebra
(B, (wAn (B''I(w) X B,.(w)))wErJ
the smallest initial segment of A containing M S: A - and by ±_}1

the corresponding relative subalgebra. Observe that, for v : X ---+ A,
dom vis always an initial segment of 'Il\:::(X).
In the rest of this subsection let (~ )iEI be any fixed set-indexed family
of partial algebras of signature ~:
The direct pmduct lB := niEI ~ has as carrier set the set theoretical
cartesian S-product B := X Ai := (X Ais)sES of the carriers. And,
iEI iEI
for w E n, one has dom wiE := { ( ( ai)iEI ' ... ' (a~(w) )iEI) E BTJ(w) I
(ai, ... , a~(w)) E domw~, for each i E I}.
For ((ai)iEI , ... , (a~(w))iEI) E domw~E, one defines
wlB((ai)iEJ, ... , (a~(w))iEI) := (w~(ai, ... , a~(w)))iEI·
In the case of many-sorted (partial) algebras the reduced product ]!)) :=
(f1iEI~)/F of the family (~)iEI w.r.t. a filter F on the index set
I is defined as follows: Let D 0 := U{ X Ai I J E F}. Moreover,
iEJ
for a := (ai)iEh, b := (bi)iEh E D 0 (for J1, J2 E F) ,13 define
In := Jl' and In,b := { i E In n Ib I ai = bi }. Moreover, define
on D 0 an equivalence relation ();: := { (a, b) E (Do,s);ES I In,b E F} .
Then the quotient S-set D := D 0 /()F is the carrier of ]!)), the ele-
ments of which will be denoted by a/ F (for a E D 0 some arbitrar-
ily chosen representative). And, for w E n, define dom WIIJi := { Q I
Q = (adF , ... , nr(w)/F) E DTJ(w) and Ig,_ := (n~~{ Ink) n {i E I I
(ai, ... , a~(w)) E domw~} E F}. And for (ai/F, ... , a7 (w)/F) E
domwiDi define wiDi(ai/F, . .. , aT(w)/F) := (w~ (at, ... , a~(w) )) iEI,jF.
As a further construction we shall need the mixed product of a family
(~)iEI, as it was recently introduced by Grzegorz Binczak: 14
Define P0 := U{ X Ai I J S: I}. We shall define In for a E P0 as
iEJ
above. For i E I, set dom H i := {a E P0 I i E In}, where H i : P0 2
dom Jri---+ ~ is a "generalized projection" with 1ri(a) := ai, whenever
a = (aj )jEla and i E In.
A partial algebra Miscalled a mixed pmduct of the family (~)i E I , iff
there is a subset M 0 S: P0 and a partial algebraic structure (wMo )wEn
satisfying:
13 Observe that all elements in any such a sequence a have to be of the same
sort, and a and b should be of the same sort, too.
14 See [BiOI].
354 P. Burmeister
(1) For every w E n, (a1, ... , a7 (w)) E M'(}(w) , and a E Mo ,a(w) one
has:
If wMo (a 1 , ... , aT(w)) =a, then
(a) Ia ~ Ia 1 n . . . n Iar<wl,
(b) 1ri(a) = wA., (1ri(a1), ... , 7ri(a7 (w))) for every i E Ia.
(2) Let e be the congruence relation on Mo generated by the set

({(a, b) E (Mos) 2 I Ia ,b = Ian Ib} )sES . Then e is a closed
congruence relation (i.e. the natural projection nate : Mo --+
Mo/0 induced by (}is a closed homomorphism as defined in the
table at the end of subsection 3.1) , and M = Mo/B, the usual
quotient algebra.
2.2 Fundamentals of Formal Concept Analysis

The basic structures of Formal Concept Analysis 15 (FCA for short) are
formal contexts lK := (G, M , I) , where G and Mare arbitrary sets, the
elements of which are called objects and attributes, respectively, and
where I ~ G x M is any binary relation. The polarity (t, .J.) defined
by (1.1) and (1.2) in the introduction plays a central role in FCA. The
pairs (A, B) with A~ G and B ~ M satisfying At= Band B+ =A
are called formal concepts. If (A, B) is a formal concept, then A is
called its extent and B its intent. By ~(IK) we designate the set of all
formal concepts of the formal context IK. Formal concepts are ordered
by set theoretical inclusion of the extents:
The ordered set (~(IK) , ::::;) always forms a complete lattice. One
has two mappings 1-L : M --7 ~(IK) - with J-L(m) := ( { m}.!. , {m }H)
(mE M)- and 1: G--+ ~(IK)- with ! (g):= ({g}t+, {g}t) (g E G)
- assigning to the attributes and objects their "generated formal con-
cepts". In line diagrams of concept lattices the name of the attribute
m is usually written a little above the circle representing the formal
concept J-L(m), and the name of the object g is usually written a little
below the circle representing the formal concept ! (g) (cf. Figure 2).
15 Cf. [GW99].
Let 1K := (G, M, I) be a formal context. For P, C ~ M we call P -t C

an attribute implication. And we say that the attribute implication
P -t C holds in 1K - and denote this by 1K FFCA P -t C - , iff, for
every object g E G, P ~ {g F implies C ~ {g F. For more details on
FCA see [GW99].
3 Properties of homomorphisms
3.1 About the polarity induced by the relation <J

Since special homomorphic images are needed for the description of
classes of partial algebras defined by some kinds of identities, we first
consider the relation <J of reflection of formulas by mappings, since
this allows us to define homomorphisms and a lot of their properties.
Let Mapr; designate the class of all S-mappings between the carri-
ers of partial algebras of signature :E (within our universe), i.e. the
class of all triplets (A, f, lffi) with A, lffi E PAigr; and f : A -t B an
arbitrary S-mapping between the carrier sets, and let us consider
the formal context .IK<J := (Mapr;, .CI:(Y), <J). And let (t<J, -l-<l) be
the polarity corresponding to this formal context (cf. (1.1) and (1.2)
of the introduction). Moreover, let Homr; designate the clas..-; of all
structure preserving mappings, i.e. homomorphisms f : A -t lffi be-
tween partial algebras. This means that for such an f one has, for
all w E n and for all sequences Q := (al, ... 'Q,T(w)) E A7J(w)' that Q E
domwA implies f o Q := (!1J(w)(l)(al), ... , f1J(w)(T(w))(aT(w))) E domwJH
and !cr(w)(w A(g_)) = wJH(f o g).
Theorem 1 LetS) ~ Mapr;. Then th e following statements are equiv-

alent:16
(i) S) = Homr;.
{ii) S) = {(X; •i ~ t') I X~ ~fin(Y), t, i' E TL:(X) }+<l.
(iii) S) = { ( { xl, ... 'XT(w), y }; •WXl ... XT(w) e y) I X-t E y1)(w)(i) (1 :S;
i :S; T(w)) , y E Yu(w), and wE r2 }+<l.
16 Observe that we write {x 1 , .. . ,Xr(w),;t/} as abbreviation for S-sets (Xs)sES
with Xs = { z I (z = Xi and rJ(w)(i) = s and 1 ~ i ~ T(w)) or (z = ;t1 and
a(w) = s)} (for s E S).
356 P. Burmeister
{iv) SJ = { ( { X1 , ... , XT(w) 1 Y }; •WX1 ... X 7 (w) ~ y) I X i E Y11 (w)(i) (1 :::;

i :::; T( w)) , y E Yu(w), all Xi are mutually distinct and distinct
from y, and wE S1 }-!<1.
This means that the usual homomorphisms between partial algebras

are exactly those mappings between partial algebras which reflect all
negations of E-equations and therefore they are exactly those map-
pings which preserve all E-equations.
As already mentioned above, a "usual" homomorphism with a proper
partial algebra as start object in general does not reflect TE-statements,
while homomorphisms between total algebras trivially reflect all TE-
statements. Therefore it should not be astonishing that a great part of
the wealth of interesting properties of homomorphisms between par-
tial algebras can be described in a model theoretic way. In particular,
many of them like injectivity, closedness, initialness 17 and their com-
binations (like "full and injective", what is equivalent to "initial and
injective") can be defined by the reflection of special E-equations, i.e.
the class of all homomorphisms having such a property is the extent
of a formal concept of the formal context ~, where the intent is gen-
erated by one of the sets of negations of E-equations mentioned in
Theorem 1, and in addition by the kinds of E-equations indicated in
Table 1 below, in which we omit the reference to the set of variables,
since it is in each case the set of all variables (of appropriate sort)
occurring in any of the terms involved . 18
17 In the sense of Bourbaki in [Bou5 7]: A homomorphism f : A ---+ B (A, Ia E

PAigE) is initial, iff, for all homomorphisms g : Ia ---+ <C, f is a homomorphism
from A into Ia iff for every <C E PAigE and for every mapping g : C ---+ A, g is a
homomorphism from <C into A iff f o g is a homomorphism from <C into A By
lnitiaiE we designate the class of all initial homomorphisms for PAigE .
18 Observe that a homomorphism f : A---+ Ia is full, if it induces on f(A) the full
relative structure of Ia. Morover, lnitiaiE consists of all those homomorphisms f :

A---+ Ia for which the preimage of every element from UwE 0 w3 (domwllll ) contains
at most one element (c.f. e.g. [B86], Proposition 9.2.3) .
notation class of all kind of additionally

reflected formulas
Monor; injective homomorphisms X .£ y (X, y E Ys, s E S)
Closedr: closed homomorphisms t ~ t (t E TL:(X) 8 , s E S,
X E ~fin(Y))
Closedr; closed homomorphisms w(g) .£ w(g) (w E S1, Q
E Y 77 (w))
Monor; closed closed injective horns X ey, t ~ t (... )
'
Monor; full full injective horns X e y, W (g) ~ Z,
'
x, y E Ys, s E S
Q E Y 77 (w), z E Ya(w),
z ~ var(g),
wEn)
lnitialr; initial homomorphisms w(g) y, (g E Y7J(w),
~
Y E Ya·(w), y ~ var(g),
w E n)
Table 1
3.2 .. Epimorphic properties" of homomorphisms and

factorization systems
Most of the "interesting" properties of homomorphisms not character-
izable by reflection of some set of formulas can be described in con-
nection with extents of formal concepts called factori zation systems
(w.r.t. some polarity) in category theory: 19
Let us recall that a homomorphism f : A --+ lB is an epimorphism, iff
CIBf(A) = B (i.e. iff f(A) generates JB). An epimorphism e : A --+ lB
is TAig r; -extendable, iff for all homomorphisms f : A --+ <C with
<C E TAigr; , there exists a unique homomorphism g : lB --+ <C such
that g o e = f. A surjective homomorphism f : A --+ lB is full (i.e. a
quotient homomorphism), iff f "induces the structure on JB". Define
the formal context~ := (Homr;, Homr;, 0 ), where, fore, mE Homr; ,
one says that (e, m) satisfies the unique diagonal-fill-in property ~
here denoted by 0 ~ iff, for any p, q E Hom r;, mop = q o e implies
19 In [AdHS90] they are now called factorization structures.
358 P. Burmeister
the existence of a unique d E Hom~ such that do e = p and mod = q .

A factorization system (£, M) is then any formal concept of~ such
that in addition to being a formal concept w.r.t. 0 one has:
-lso~ ~ EnM,
- Eo£~£,
- MoM~ M, and
- M o£ =Hom~.
Theorem 2 Let ~ .CE(Z) be any set of elementary implications,

where Z is any global S -set of variables, and let M ·- +<l . Then
(M-1-0 , M) is always a factorization system. 20
Thus, in particular, when M is a class of homomorphisms defined via

the reflection of some set of QE-equations, then M+o is its partner in
a factorization system and consists of a class of epimorphisms (which
encode the elementary implications under consideration) - and this
is the reason why we call it here an "epi-factor". The proofs of the
statements in Table 2 can mainly be found in [B86], subsections 10.2
and 10.3. 21
In addition, observe that the "epi-factor" corresponding to Initial~
consists of all those surjective homomorphisms f : A -+ lB for which
the preimage of every element from B \ U wEO wlB(domwlB) contains
exactly one element.
°
2 Cf. e.g. [B86], Remark 10.2.11 -observe that there and in other books and
papers the operators have a different notation than we have used in this note in
order to have a homogeneous notation. Very often one writes A(£) instead of [ t0 ,
and A0 P(M) instead of M.!-0 .
21 Observe that, what one often ·- and we here, too - calls a "mono-factor",
need not consist only of monomorphisms. As an example take the class Closed~ of
all closed homomorphisms. However, the factorization systems considered origi-
nally usually consisted of a class of epimorphisms as extent and a class of monomor-
phisms as intent. - Observe, too, that we have in the theory of partial algebras an
interesting factorization system (all final homomorphisms, all bijective homomor-
phisms ) , where the final homomorphisms between partial algebras - which form
the dual concept to initial homomorphisms in the sense of Bourbaki [Bou57] -
are exactly those homomorphisms, which fully induce the structure on the image
algebra, but they need not be surjective and therefore not epimorphic. Moreover,
the bijective homomorphisms are not defined by t he reflection of formulas.
"epi-factor" "mono-factor"
(class of all full and surjective homomorphisms, Menor; )
(class of all TAigr;-extendable epimorphisms, Closedr: )
(Epir;= class of all epimorphisms, Monor:,closed )
(class of all surjective homomorphisms, Monor:,full )
(class of all final homomorphisms, cl. of all
biject. horns).
Table 2
4 Polarities derived from the relation f=

The relation f= of validity of a first order formula in a (many-sorted)
partial algebra for some given signature :E of the fundamental opera-
tions under consideration can be restricted to subsets :F ~ .Cr:(Y) of
special interest within the first order language .Cr:(Y). As mentioned
earlier, :F is usually chosen to be - for arbitrary X E SJJfin (Y), and
terms t, t' E Tr:(X)s and ti, t~ E Tr:(X)si (i E {1, ... , n}) -
- the set of all E-equations,
- the set of all ECE-equations,
- the set of all QE-equations,
- or the set of all (special conjunctions of) ECE-equations of some
special kind like
- the set of all weak equations,
- the set of all strong equations or KLEENE-equations ,
- the set of all regular strong equations,
- or various other similar concepts of special equalities.
With each such set of special formulas one has the problem of de-
scribing the closed sets/classes of the induced Galois connection on
the syntactical and on the semantical side, respectively, i.e. to find
so-called Birkhoff-type theorems and Birkhoff-Tarski-type theorems.
ForE-, ECE- and QE-equations this has been no great problem, and
the results can be found e.g. in [B86] or [B93]. However, in the cases
of weak and strong equations the problems have been much harder,
and only recently G. Binczak has solved the "semantic problem" for
weak equations in a satisfactory way (see [BiOl]) by inventing a new
operator Pm (which he calls the formation of mixed products as defined
360 P. Burmeister
Form I corresponding semantic operator I

E-equations 'HwSc P = 'HwSc Pr
ECE-equations HcScPr
QE-equations IScPr
weak equations IPm
Table 3
in this note at the end of subsection 2.1).

In [Hoe73], H.Hoft has characterized closed sets of weak equations as
what he calls weakly invariant relations. We do not give the details
here. Moreover, the problems for strong equations are still unsolved ,
while Bozena and Bogdan Staruch have solved in [StSt94] the prob-
lems for regular strong equations. 22 Again we refer here to the litera-
ture. On the other hand William Craig has observed in [Cr89] (for the
homogeneous case) that for the extension of the language by a "log-
ical" binary operation symbol, which is always interpreted as a total
binary first projection, KLEENE-equations in this extended language
and ECE-equations (in the original or extended language) have the
same expressive power (when the empty algebra is excluded) - and a
similar observation can be made in the case of heterogeneous partial
algebras (see [B95]).
Theorem 3 For many-sorted par·tial algebras one has the following

semantical operators shown in Table 3 for the descr·iption of the closure
Mod Form(~) of classes~ of partial algebras w.r·.t. some sets For·m of
special QE-equations (with the involved operators defined below):23
Here the operators are defined as follows, for any class ~ ~ PAigi;:
• Hw(~) := { lB\ E PAigi; I there exists a surjective homomorphism

A --+ lB\ for some A E ~ } ;
• He(~) := { lB\ E PAigi; I there exists a closed and surjective

22 Regularity of (X; t ~ t') means that - according to the recursive construction
of terms - t and t' contain the same variables.
23 If Sis infinite, then the equality 1lwScP = 1lwScPr no longer holds , and one
then has to take 1lwScP,. as semantic operator for E-equations, if one wants to
keep the language finitary (see [B95]).
homomorphism A ----+ lR for some A E R} ;
• I(Jt) := { lR E PAigE IJR is isomorphic to some Jt-algebra};
• Sc(R) := { lR E PAigE I lR is a (closed) subalgebra of some R-

algebra};
• P(.st) := { lR E PAigE I there exist a set I and a family (.1\)iEI of

R-algebras SUCh that JR = rriEJ .1\ is the direct product of this
family};
• Pr(R) := { lR E PAigE I there exist a set I, a filter :F on I and

a family (.I\ )iEI of Jt-algebras such that lR = (f1iEI .1\) / F is a
reduced product of this family} ;
• Pm (Jt) := { lR E PAigE I there exist a set I and a family (.I\ )iEI

of Jt-algebras such that lR is a mixed product of this family}.
The descriptions of the closed sets of formulas under consideration is

a little more involved and not discussed here in all cases.
We only want to give a description of closed sets of E- ECE- and QE-
equations. Here a set theoretical representation is useful: 24
Definition 4 Let
n
t- := (X; f\ ti ~ t~ =} t ~ t')
i= l
be any QE-equation. Then t- may be set theoretically represented by

an ordered pair ({(x,x) I x EX} U {(ti,tD 11 SiS n},(t,t')) E
~fin(TE(X) 2 ) x TE(X)2. If t- is an ECE-equation, then the correspond-
ing pair belongs to ~fin({(t,t) It E TE(X)}) x TE(X) 2 , and if~ is an
E- equation, then the corresponding pair belongs to ~fin ({ (x, x) I x E
24 For E-equations in the homogeneous case (excluding the empty algebra) the
simplest description of closed sets is by saying that they are closed and fully
invariant congruence relations on relative subalgebras lF of 'li' E (Y), such that lF is
freely generated by Y. The following generalizes this for the case when the empty
partial algebra is allowed, too , and to heterogeneous partial algebras (with empty
phyla allowed).
362 P. Burmeister
X}) x TdX) 2 . In each case we can represent X as var(L,). Since ev-

ery t E TdX) can be considered as an elem ent of T~(Y) {because of
X ~ Y and the recursive definition of terms), we can define
PremE := ~fin(({(y,y) I Y E Ys})sEs),
PremEcE := ~fin(({(t , t) It E TdY)s})sES) , and

PremQE := ~fin(T~(Y) x TdY)).
For P ~ TdY) 2 we define
var(P) := U (var(t) U var(t')).

(t,t')EP
And we obtain
Eeqy = U {P} x T~(var(P)) 2 ,

PEPrems
ECEeqy = U {P} x T~(var(P)) 2 ,

PEPremscE
QEeqy = U {P} x Tdvar(P)) 2 ,
PEPremQE
for the sets of all set theoretical encodings of E-, ECE- or QE-equations
with variables in Y, respectively. Now, for
Prem E {PremE, Prem Ec E, PremQE},
we consider in the following
Q ~ U ({P} x T~ (var(P)) 2 )
PEPrem
to be any set of set theoretically encoded elementary implications of
the corresponding type. For P E Prem we define
Q(P) := {(t , t') I (P, (t , t')) E Q}.
For any class R of partial algebras defin e

lmpPrem(R) := {(P, (t, t')) I P E Prem , t, t' E Tdvar(P)), R f= (var(P);
1\(p,p')EP p e p' ::::} t _:_ t')}

and set -!-E to be the relative subalgebra of 'IT' = '.II'L:(Y) consisting of
all subterms of terms occurring in E s;;;; TI'. (Y) 2 , and let supp E :=
Uct,t')EE {t, t'}) (i.e. the support of E) be the set of all terms occur-
ring as at least one component of a pair in E, and, moreover, let supp E
designate the relative subalgebra of'.II'I'-(Y) with carrier suppE. - -
With the above notation one has the following description of closed
sets of elementary implications of one of the three kinds of Prem:
Theorem 5 25
Let Prem E {Preme, Premece, Premqe}, and let Q s;;;; UPEPrem({P} x
TI'.(var(P)f) be any set representing elementary implications connected
with Prem.
(a) Then the following statements are equivalent:
(i) Q = lmpPrem(Mod(Q)).
(ii) Q has the following properties (11) through (14) for any
P, P' E Prem:
(11) suppQ(P) is a var(P)-generated relative subalgebra of
'.II'I'.(var(P)) - in particular one has suppQ(P) =
-!- Q(P).
(12) Q(P) is a closed congruence relation on supp Q(P).
(13) P s;;;; Q(P).
(14) For every homomorphism f : -!-P ~ supp Q(P') satis-
fying(! x f)(P) s;;;; Q(P'), there exists a homomorphic
extension fPP' : supp Q(P) ~ supp Q(P'), which satis-
fies (JPP' X fPP')(Q(P)) s;;;; Q(P').
25 The proof of this theorem for the homogeneous case can be found first -
formulated for QE-equations - in [ABN81] (and in another form in [AN83]). Later
it appeared in [B86] and, without proof, in [B93]. Yet in all three cases (Il)
contained an error, since we there refer to "'Q(P) rather than to suppQ(P), and
±Q(P) is trivially generated by var(P), i.e.-then (Il) does not contain any non-
trivial condition. We think that in this set theoretical form, and formulated for
heterogeneous partial algebras the theorem is formulated here for the first time.
364 P. Burmeister
(b) If Q = lmpPrem(Mod(Q)), and P E Prem, then Q(P)

= n{ker f~ If: ±P---+ A., A. E Mod(Q) and p ~ ker !~}.
Again factorization systems come into the picture in connection with

the Meta Birkhoff Theorem of Hajnal Andn~ka, Istvan Nemeti and
Ildiko Sain (see [AN82] and [NSa82]) characterizing closed model class-
es of universal Horn formulas in a very general category theoretical
way. Namely, the class S of "admissible subobjects" there has to
correspond to the "mono-factor" of a factorization system. And the
class of epimorphisms used for the "admissible epimorphisms" has in
some way to be compatible with S-!-0 (for more details cf. e.g. [B92] or
[B86]) .
5 u Attribute exploration" uses further polarities
Attribute exploration is a method fromFCA , where the user or expert

fixes a very large context 1U = (GI[], Ml[], I u) of interest- with some
finite set Mu of attributes - as so-called universe, and where a pro-
gram (like "Conlmp" 26 ) asks the expert in a systematic way, whether
some attribute implications computed by the program hold ia 1U. The
aim of the procedure is to get a list I of attribute implications holding
in 1U, from which all other attribute implications holding in 1U can be
derived. And at the same time one wants to produce a subcontext
lK = (G, Mu, I) of 1U, which contains for each attribute implication
not holding in 1U a counterexample. - These data then allow one to
compute the concept lattice of 1U up to isomorphism.
We present an example for a homogeneous mono-unary signature (i.e.
S = { s} ), n = { w }, T : w t----+ 1 (rt and a are obvious), 27 where the set
Gu equals HomE. The list of attributes is shown in Table 4 together
with their abbreviations used at different occasions in order that e.g.
the implications do not become too long. 28
26 Cf. [BOOa].
27 The result holds for all homogenoeus signatures with at least one at least
unary operation, but it might look different, if we have e.g. a signature with only
one unary operation mapping elements of one sort t o elements of a different sort,
since then one cannot produce examples like Hom6 and Hom7 below.
28 Moreover, the program "Conlmp" accepts only names with at most 9 charac-
ters.
The set Mu; of attributes in implications in the

full name context
injective injective lllJ
full&surjecti ve full&surj f&sur
initial&injective init&inj ini&inj
surjective surjectiv sur
closed closed cl
TAigE-extendable TA-extend TA-ext
closed&injective clos&inj cl&inj
epimorphic epimorph ep1
initial initial llll
A0 P(initial)= (lnitiaiE)-1-0 LOinitial LOini
Table 4: The attributes of the formal context FactSys
The algorithm may start with an empty list of objects or a list of

objects entered in advance, and one can also enter some implications
as so-called background implications in advance.
In connection with the algorithm of attribute exploration the follow-
ing list of so-called Duquenne-Guigues-implications is produced:
1. { LOinitial} =? { surjectiv, epimorph}

2. { clos&inj} =? {injective, init&inj , closed, initial}
3. { TA-extend} =? {injective, init&inj , epimorph, initial}
4. {closed, epimorph} =? { full&surj , surj ectiv}
5. { surjectiv} =? { epimorph}
6. { surjectiv, epimorph, initial} =? { full&surj , closed}
7. { init&inj } =? {injective, initial}
8. { full&surj } =} { surjectiv, epimorph }
9. { full&surj, surjectiv, closed, epimorph, initial, LOinitial} =? Mu;
10. { injective, initial } =? { init&inj }
11. {injective, closed} =? { init&inj , clos&inj, initial}
12. {injective, surjectiv, epimorph} =? { LOinitial}
13. {injective, full&surj , surjectiv, epimorph, LOinitial} =? Mu;
Moreover, the following "complete" list of counterexamples has been

366 P. Burmeister
produced as collected in a formal context "FactSys" shown in Table 5.

The object names correspond to the homomorphisms shown in Fig-
ure 1.
lllJ f& ini& sur cl TA- cl& ep1 llll LO

sur lllJ ext lllJ llll
Homl X X X X X
Hom2 X X X X X
Hom3 X X X X X
Hom4 X X X X
Hom5 X X X X X
Hom6 X X X X
Hom7 X X X X
Table 5: The formal context FactSys
Homl Hom2 Hom3
Hom4
Hom5 Hom6 Hom7
Fig. 1: Sketches of the homomorphisms of the context FactSys
In Figure 2 we finally show the line diagram of the resulting concept

lattice (113(FactSys), ::;).
Fig. 2: The concept lattice (113(FactSys), ::;)
The relationships to polarities (induced by the relation of satisfaction

of an attribute implication by the universe and by the formal sub-
contexts at every intermediate step) of the methods involved in the
algorithm of attribute exploration have been indicated already in the
introduction and in subsection 2.2, and we cannot here go into more
detail.
References
[AdHS90] J. Adamek, H. Herrlich, G. E. Strecker. Abstract and

Concrete Categories - The Joy of Cats. John Wiley &
Sons, Inc., 1990.
[ABN81] H. Andreka, P. Burmeister, I. Nemeti. Quasivarieties

of partial algebras - A unifying approach towards a
two-valued model theory for partial algebras. Studia Sci.
Math. Hungar. 16, 1981, pp. 325- 372.
[AN82] H. Andreka, I. Nemeti. A general axiomatizability theo-

rem formulated in terms of cone-injective subcategories.
368 P. Burmeister
In: Universal Algebra (Proc. Coil. Esztergom 1977) , Col-

loq. Math. Soc. J. Bolyai, Vol. 29, North-Holland Publ.
Co., Amsterdam, 1982, pp. 13- 35.
[AN83] H. Andreka, I. Nemeti. Generalization of the concept of

var-iety and quasivar-iety to par-tial algebras thmugh cat-
egor-y theor-y. Dissertationes Mathematicae (Rozprawy
Mat.) No. 204, Warszawa, 1983.
[Bi01] G. Binczak. A character-ization theor-em for- weak var-i-

eties. Algebra univers. 45, 2001 , pp. 53- 62.
[Bir67] G. Birkhoff. Lattice Theor-y. American Mathematical So-

ciety, 1967.
[Bou57] N. Bourbaki. Theor-ie des ensembles. Chapitr-e 4, Str-uc-

tur-es. Hermann, Paris, 1957.
[B82] P. Burmeister. Parlial algebras - sur-vey of a unifying

appr-oach towar-ds a two-valued model theor-y for- par-tial
algebras. Algebra Universalis 15, 1982, pp. 306- 358.
[B86] P. Burmeister. A Model Theor-etic Or-iented Appr-oach to

Parlial Algebras. Intr-oduction to Theor-y and Application
of Par-tial Algebras - Par-t I. Mathematical Research
Vol. 32, Akademie-Verlag, Berlin, 1986.
[BWo87] P. Burmeister, B. Wojdylo. Pr-oper-ties of homomor--

phisms and quomor-phisms between par-tial algebras. Con-
tributions to General Algebra 5 (Proc. Salzburg Conf.,
1986). Verlag Holder-Pichler-Tempski, Wien, Verlag
B.G.Teubner, Stuttgart, 1987, pp. 69- 90.
[B92] P. Burmeister. Tools for- a Theor-y of Par-tial Algebras.

General Algebra and Applications (Eds.: K.Denecke and
H.-J.Vogel) , Research and Exposition in Mathematics,
Vol. 20, Heldermann Verlag Berlin, 1993, pp. 12- 32.
[B93] P. Burmeister. Par-tial Algebras ~ An Intr-oductory

Survey. In: Algebras and Orders (Proceedings of the
NATO Advanced Study Institute and Seminaire de
mathematiques supeneures on algebras and orders,

Montreal, Canada, July 29 - August 9, 1991; Eds.:
I.G.Rosenberg and G. Sabidussi) , NATO ASI Series C,
Vol. 389, Kluver Academic Publ., Dordrecht, London,
1993, pp. 1- 70.
[B95] P. Burmeister. On the equivalence of ECE- and gener-

alized KLEENE-equations for many-sorted partial alge-
bras. Contributions to General Algebra 9 (Proceedings
on the conference on Universal Algebra and its Applica-
tions held at Linz in June 1994). Verlag Holder-Pichler-
Tempsky, Wien 1995- Verlag B.G. Teubner, Stuttgart,
pp. 91- 106.
[BOO] P. Burmeister. Lecture Notes on Universal Algebra -

Many Sorted Partial Algebras. Fragment of lectures
hold in the summerterms of 1994, 1996, 1998 and 2000
at the Darmstadt University of Technology.
http:/ /www.mathematik.tu-
darmstadt.de/ rvburmeister1
Choose "Lecture Notes on Partial Algebras (Frag-
ment)".
[BOO a] P.Burmeister. Conimp - Ein Programm zur Formalen

Begri.ffsanalyse.
G. Stumme, R. Wille (Eds.) Begri.ffiiche Wissensverar-
beitung: Methoden und Anwendungen, Springer-Verlag,
2000, pp. 25- 56
A largely extended English translation: Formal Concept
Analysis with Conimp: Introduction to the Basic Fea-
tures (Preprint No. 2284 of the Darmstadt University
of Technology)
can be accessed on the WWW-server:
http:/ /www .mathematik.tu-darmstadt.de/agsjag1/
Software /DOS-Programme /Welcome_en. html
[Cr89] W. Craig. Near-equational and equational systems of

logic for partial functions. I and II. The J. of Symb.
Logic 54, 1989, pp. 759- 827 and pp. 1181-1215.
370 P. Burmeister
[GW99] B. Ganter, R. Wille. Formal Concept Analysis - Mathe-

matical Foundations. Springer, 1999.
[HS73] H. Herrlich, G. E. Strecker. Category Theory- An Intro-

duction. Allyn and Bacon, 1973 (2nd ed. : Heldermann-
Verlag).
[Hoe73] H. Hoft. Weak and strong equations in partial algebras.

Algebra univers. 3, 1973, pp. 203- 215.
[NSa82] I. Nemeti, I. Sain. Cone-implicational subcategories and

some Birkhoff-type theorems. Universal Algebra (Proc.
Coli. Esztergom 1977) , Colloq. Math. Soc. J. Bolyai,
Vol. 29, North-Holland Publ. Co. , Amsterdam, 1982, pp.
535- 578.
[StSt94] B. Staruch, B. Staruch. Strong regular varieties of partial

algebras. Algebra univers. 31 , 1994, pp. 157-176.
Author's address:
Peter Burmeister
FB4, AG1, Darmstadt University of Technology
SchloBgartenstraBe 7
D-64289 Darmstadt, Germany
email: burmeister@mathematik. tu-darmstadt.de
Complexity of Terms and the Galois
Connection ld-Mod
K. Denecke, 5. L. Wismath*
Abstract
A non-trivial identity s ~ t is normal when neither of the terms s and t is
a variable. Using several common measures of the complexity of a term,
this requires exactly that both s and t have complexity ::::: 1. We generalize
this definition to any integer k ::::: 1, by saying that a non-trivial identity
s ~ t is k-normal when both s and t have complexity ::::: k. A variety will
be called k-normal when all its non-trivial identities are k-normal. Using
results from the theory of Galois connections and complete sublattices, we
show that the collection of all k-normal varieties of a fixed type forms a
complete sublattice of the lattice of all varieties of the type. We also gen-
eralize to the k-normal case the results of Graczynska ([Gra]) and Melnik
([Mel]) describing normal varieties and the mapping taking any variety to
the least normal variety containing it.
AMS Classification: 08B05, 08A55.
Key words: Complexity, Normalization, Galois connection, Complete sub-
lattice.
1 Introduction
An identity s ~ t of type T is said to be normal if either both of s and

t are equal to a variable x or neither of them is a variable. A variety
is called normal if all of its identities are normal. The properties of
normal varieties, and of the mapping which takes any variety to the
least normal variety to contain it, have been studied by Mel'nik ([Mel])
*Research of the second author supported by NSERC of Canada.
371
and Graczynska ([Gra]). In this paper we introduce a generalization

of the concept of a normal variety, called a k-normal variety, where
k is any natural number. Using a measure of the complexity of a
term, which assigns to any term an integer 2:: 0 and for which a term
has complexity zero iff it is a variable, we can characterize non-trivial
normal identities s : : : : t as those in which both s and t have complexity
2:: 1. We will say that an identity s : : : : t is k-normal, for k 2:: 0, if both
s and t have complexity 2:: k , and call a variety k-normal if all its
non-trivial identities are k-normal. We show that the collection of all
k-normal varieties of a fixed type forms a complete lattice, and study
the mapping Nk which takes any variety to the least k-normal variety
containing it. We show that most but not all of the theorems from
[Mel] and [Gra] generalize to the k-normal case. In particular, for any
variety V which is not normal, the mapping Nk is shown to be an
injective lattice homomorphism on the lattice of subvarieties of V.
We introduce our k-normal varieties as an example of a more general
construction, based on Galois connections. It was shown in [DW] that
given a Galois connection between two sets A and B, any Galois-closed
subset of B may be used to define a kernel operator on the power set
P(B) and then a closure operator on P(A). Moreover, the collection of
sets closed under this closure operator forms a complete sublattice of
the lattice of all Galois-closed subsets of A. In this section we describe
these results for the particular setting in which we will use them ,
with the Galois connection (Id, Mod) between classes of algebras and
sets of identities. Using as our starting point the set of identities of
complexity 2:: k, we use this construction to produce the lattices of
k-closed, or k-normal, varieties. In Sections 2 and 3 we generalize the
results of Graczynska and Mel'nik to the k-normal case.
Let T = (ni)i EI be any type of algebras, with operation symbols fi
of arity ni. Let X= {x 1 ,x 2 ,x 3 , ... } be a countably infinite set of
variables. We denote by W7 (X) the set of all terms of type T on the
alphabet X. As our base sets for a Galois connection, we consider
the sets A= Alg(T) of all algebras of type T, and B = W7 (X)2 of all
equations of type T. We use the maps I d: P(A) ---t P(B) and Mod:
P(B) ---t P(A), where for any K ~ Alg(T) and any E ~ W7 (X) 2
Id(K) = {u:::::::: v E WAX) 2 1 A satisfies u:::::::: v for all A E K},

Mod(E) ={A E Alg(T) I A satisfies u:::::::: v for all u:::::::: vEE}.
Complexity of Terms and the Galois Connection ld-Mod 373
It is well known that these two maps form a Galois connection between
Alg(T) and W7 (X) 2 , induced by the relation of satisfaction between
algebras and identities. The corresponding complete lattices are the
lattice £(T) of all varieties of type T, and the lattice E(T) of all equa-
tional theories of type T.
Now we take P(T) to be any property of identities which is an equa-
tional theory, that is, any element P(T) E E(T). It follows from [DW]
that the map <p~ with .Ef--t .En P(T) is a kernel operator which pre-
serves both unions and intersections, and the mapping <pj, = J.L<p~&
= M od<p~I d is a closure operator on the lattice £(T), and preserves
joins. Moreover the collection of <pj,-closed varieties forms a complete
sublattice of£(T). This corresponds to the Galois connection induced
by the Galois-closed subrelation
Rp = {(A, s ~ t) E Alg(T) x P(T) I A satisfies s ~ t}.
Theorem 1.1. Let P(T) be any property of identities which forms an

equational theory. Then the class of all <pj,-closed varieties of type T
forms a complete sublattice of the lattice £( T) of all varieties of type
T.
Several well-known examples of this corollary have been studied by a

number of authors. Taking P( T) to be the set of all normal identities
of type T , or the set of all regular identities of type T, gives the lat-
tices of normal and regular varieties of type T, respectively (see [Gra],
[Plo]). Externally compatible identities and 0-compatible identities
have also been studied (see [Chr]). Our goal is to describe a family of
new examples, based on the idea of complexity of a term. Since nor-
mal identities may be characterized by complexity, our examples will
include normal identities, but also generalize t he concept of a normal
identity or variety to that of k-normal, for any natural number k 2:: 1.
There are several obvious ways to measure the complexity of a term of
type T. Rather than giving the inductive definitions here, we will de-
scribe these measures more graphically, using the fact that for finitary
type any term may be viewed as a tree, with a tree diagram. The most
commonly used measure of complexity is the depth of a term, which
is defined as the length of the longest path from the root to a node in
the tree diagram. We can also count the total number of occurrences
of operation symbols in a term, or the total number of occurrences of
variables. (For technical reasons, we will modify the latter measure to

count instead one less than the total number of variables.) For each
of these, we have a function c : W7 (X) --+ N0 , which assigns to each
term t a number c(t), and having the property that c(t) = 0 iff t is
a variable. We will call such a function c a complexity function for
terms.
Now suppose we have such a complexity function. An identity s ~ t
of type T is said to be normal if either s = t or both c( s) and c( t) are
~ 1, and a variety V is called normal if all its identities are normal.
To extend this idea, we define, for any k ~ 0,
Pk(T) = {s ~ t E WT(X) 2 1 c(s) ,c(t) ~ k}U{s ~ t E W7 (X) 2 1 s = t} .
To apply Theorem 1.1 , we need the set Pk (T) to be an equational
theory. For the three complexity measures mentioned above, the
depth , variable-symbol count (decreased by 1) and the operation sym-
bol count, this is easily seen to be the case; the set Pk (T) is closed under
the five deduction rules of equational logic. Applying our construction
above, we define two maps, Nt on sets of identities and N;( on sets
of algebras of type T:
Nt : L:--+ L: n Pk(T),
N;(: V--+ Mod(IdV n Pk(T)) .
Then we have the following result.
Theorem 1.2. The map N;( : .C(T) --+ .C(T) defined by N;((V)
Mod(IdV n Pk(T)) is a join-preserving closure operator on .C(T), and
the class of all N ;(-closed varieties forms a complete sub lattice of the
lattice of all varieties of type T.
We will denote the lattice of all N;(-closed varieties of type T by .Ck(T) .

Fork= 0, the set Pk(T) is just the set of all identities of type T, and N;(
is the identity mapping on .C(T) . In this case the sublattice we obtain
is all of .C(T). Fork= 1, we obtain the set of all normal identities of
type T, and the corresponding lattice .C 1(T) of all normal varieties. In
analogy with this normal case, we shall refer to the Nk-closed varieties,
for k ~ 1, as k-normal varieties.
Since for any k > 0 we have Pk (T) -;2 Pk+ 1 ( T) , for any variety V
we have N;((V) ~ N;(+ 1 (V) . Moreover we get a chain of complete
sublattices
Similarly, for each k :2: 0 the complete sublattice £k(T) determines a

Galois-closed subrelation of the basic relation R of satisfaction between
algebras and identities (see [ADP]):
Rk = u{v X Id(V) Iv E £k(T)}
= {(V,s ~ t) Is~ t E IdV and Nf(V) = V}.
This gives us a chain of Galois-closed subrelations,
It is easy to show that the elements in these two chains, of sublattices

and of Galois-closed subrelations, are all distinct.
Example 1.3. The medial variety of type (2) is defined by one identity
f(j(x, y), f(z, w)) ~ f(j(x, z), f(y, w)). Using the depth function as
our complexity measure, this identity has complexity of 2 on each side.
This means that V = Nt(V) = Nt(V) C Nf(V). The type (2) variety
V = SL of semilattices has Nt(V) = V =1- Nt(V) =1- Nt(V) .
2 k-Normalizations
The lattice of all normal varieties of a fixed type T , and the mapping
Nt taking any variety V to the least normal variety Nt(V) to contain
V, have been studied by Mel'nik ([Mel]) and Graczynska ([Gra]) . In
this section we show that some of their results can be extended to our
k-normal varieties and mappings N;(, fork :2: 2.
Lemma 2.1. Let k :2: 0. For any variety V, the variety N;((V) is the
smallest k-normal variety to contain V.
Proof: Since N;( is a closure operator, we have V ~ N;((V) and

N;((V) is k-normal. If W is any k-normal variety containing V, then
V ~ W implies N;((V) ~ N;((W) = W. Thus N;((V) is the least
such variety. •
For this reason we refer to the variety N:(v) as the k-normalization

of V. In the case that k = 1 we shall continue to speak of normal
identities and varieties. Now we want to consider the mapping N: on
the lattice £(T) of all varieties of type T. We know from Theorem 1.2
that N: is always a join-homomorphism on £(T). Graczynska proved
in [Gra] that for k = 1, this mapping is a lattice homomorphism, and
when restricted to the lattice £(V) of subvarieties of a non-normal
variety V, it is also injective. The basic tool in her proofs is the fact
that when a variety V satisfies a non-normal identity of the form x ~ t,
where x is a variable and t is a term with c(t) ~ 1, we can use this
identity to "inflate" other non-normal identities in V to normal ones.
In fact, if our term t has complexity k ~ 1, we can inflate identities to
complexity ~ k, which means that we can extend Graczynska's proofs
to the k-normal case, given the right conditions. We begin with some
useful preliminary results.
Lemma 2.2. Let k,p ~ 0. Th en N:N: = N:\ for r = max{k,p}.
Proof: By definition we have for any variety V

N:N:(V) = ModNf(IdModNff(IdV))
= ModNf(N/!(IdV)), since Nff(IdV) is an equational
theory
= Mod(Nf(T) n Nff(T) n IdV)
= Mod(N/?(IdV))
= N ;4(V). •
Lemma 2.3. Let V be any variety.
(i) If V is N:-closed for some k ~ 1, then V is N:-closed for all
1 :::::; p :::::; k.
(ii) If V is not normal, then V is not N: -closed for any k ~ 1.
Proof: (i) We have V ~ Nf(V) ~ Nt(V) ~ · · · ~ N:(v) , so if V =

N:(v) then V = N:(v) for all 1 :::::; p:::::; k.
(ii) This follows directly from (i). •
We will denote by Con(~) the set of consequences, under the usual

deduction rules (1) - (5), of any set ~ of identities. Since our class
Pk (T) of identities is an equational theory, it follows that for any set

I: of identities we have Con(N{(L:)) ~ N{(Con(L:)). In particular,
when I: =I dV is an equational theory defined by a variety V, we have
Con(N{(IdV)) = N{(Con(IdV)).
We now fix a particular non-normal variety V , and consider the be-
haviour of the functions Nk when restricted to the subvariety lat-
tice .C(V). Since V is non-normal, it satisfies a non-normal identity
x : : : : : t, where t is some term with complexity p 2: 1. By identifica-
tion of variables if necessary, we see that V also satisfies an identity
x::::::::: t(x, ... , x) in which x is the only variable to occur. We will refer
to the complexity p of the term t as a non-normality index of V , and to
the identity x::::::::: t(x, ... , x) as a non-normality witness for V of level
p. Notice that by replacing each x in t(x, ... , x) by the whole term
t(x, ... , x), and iterating this process, we can produce non-normality
witnesses for V of arbitrarily large level k 2: p . Throughout the next
several proofs, we will use the following notation. For any term u, we
let u* be the result of replacing each variable x occurring in u by the
term t(x, ... , x). It is clear that u::::::::: u* is always a consequence of the
non-normality witness identity x : : : : : t(x, ... , x), and is an identity of
V. In addition, any such term u* has complexity 2: p.
Lemma 2.4. Let V be a non-normal variety, with a non-normality
witness identity e of the form x::::::::: t(x, . .. , x) of level p. Then I dV =
Con(N{(IdV) U {e}), for any k 2: 1.
Proof: We will first show that IdV = Con(Nff(IdV)U{e}), where pis
the complexity of the term t from the identity e. Since Nff (I dV) U{ e}
~ IdV, we have Con(Nff(IdV) U {e}) ~ IdV, and we have to prove
the opposite inclusion. Let u::::::::: v be any identity of V. Then V also
satisfies u* : : : : : v*, and this identity is in N {(I dV). Since u : : : : : u* and
v : : : : : v* are consequences of e, we see that u : : : : : v is a consequence of the
identities u : : : : : u*, u* : : : : : v* and v* : : : : : v , all of which are consequences
of N{(IdV) U {e}. This shows that IdV ~ Con(N{(IdV) U {e}).
Now let 1::; k::; p . Then N{(IdV) ~ N{(IdV), so we have
IdV = Con(N:(IdV) u {e}) ~ Con(N{(IdV) u {e}) ~ IdV,

giving us the equality Con(N{(IdV)U{e} = IdV , fork::; p. The final
claim follows from the fact that , as noted above, V has non-normality
witnesses of arbitrarily large level p. •
This Lemma will be used frequently in our subsequent proofs. It tells

us that all identities of V , including those of complexity ::; k, can
be deduced from complexity 2:: k identities of V plus one non-'normal
identity of V. But we point out here that the Lemma is not true for
sets of identities which are not equational theories. That is, we might
ask whether if we take a set of identities () which is a basis for I dV ,
meaning that Con(B) = IdV, we still have IdV = Con(Nf(B) U {e}).
The answer is no, since we have to be able to use all the consequences
of(), not just those in Con(Nf(B)). This fact will be illustrated in
Example 3.7.
The same inflation technique is used for the next proof, which gener-
alizes the key Lemma in Graczynska's proof.
Lemma 2.5. Let I: and r be two sets of identities. Let e be an identity

of the form x ~ t(x, .. . , x) for some term t of complexity p 2:: 1, such
that e E I: U f . Then for any 1 ::; k ::; p,
Con(Nf!(Con(L::)) u N{(Con(f))) = N{(Con(L:: U f)).
Proof: We always have Con(Nf(Con(L::))uNf(Con(f))) ~ Nf(Con

(I: U r)) (see [DW]), so we must prove the opposite inclusion. Sup-
pose that u ~ v is an identity which is a consequence of I: U r and for
which both u and v have complexity 2:: k. Then there is a sequence of
identities e1 , .. . , em, using only identities from L::Uf and the five rules
of deduction, which ends in u ~ v and which is a proof of u ~ v. Now
we consider the "inflated" sequence ei, ... , e~_ 1 , u* ~ v*. If ei was an
identity from I:, then ei is in Con(L::), and similarly for r . Adding to
this inflated sequence the identities u ~ u*, v ~ v* and u ~ v, we get
a sequence which is a proof of the identity u ~ v, in which each term
used has a complexity 2:: k. This shows that u ~ v is a consequence
of Nf(Con(L::)) U Nf(Con(r)). •
We shall need the result of this Lemma in a particular setting: when

I: = I dU1 and r = I dU2 , for two subvarieties U1 and U2 of a non-
normal variety V, for which we have a non-normality witness e of
level p. Since such non-normality witnesses then exist for arbitrarily
large levels p as well, we have the following result.
Corollary 2.6. Let V be a non-normal variety, and let U1 and U2 be

any subvarieties ofV . Then for any k ~ 1, Nf(Con(IdU1 UldU2)) =
Con(Nf(Con(IdU1 ) U Nf(Con(IdU2))). •
Now we are ready to prove our main theorem. Recall that for any
variety V, we denote by £(V) the subvariety lattice of V.
Theorem 2.7. Let V be a non-normal variety. Then for every k ~ 1,

the function Nf : £(V) ---+ £(Nk(V)), defined by Nf : U H Nf(U) ,
is an injective lattice homomorphism.
Proof: We know from Theorem 1.2 that the map Nf always preserves
joins. We now show that on subvarieties of V it also preserves meets.
We have
Nf(Ul 1\ U2) ModNf(Id(Ul 1\ U2))
ModNf(Con(IdU1 u IdU2))
ModCon[Nf(Con(IdUI)) U Nf(Con(IdU2))],
by Corollary 2.6
ModCon[Nf(IdUI) U Nf(IdU2)],
since I dUi is an equational theory
ModCon(IdNf(U1 ) U IdNf(U2)) ,
since IdNf(Ui) = Nf(IdUi)
Modld(Nf(UI) 1\ Nf(U2))
Nf(UI) 1\ Nf(U2)·
Thus Nf is a lattice homomorphism on £(V). For the injectivity,
let U1 and U2 be any subvarieties of V, and let e be a non-normality
witness for V. Then
Nf(UI) Nf(U2) =? ModNf(IdUI) = ModNf(IdU2)

=? Nf(IdUI) = Nf(IdU2) ,
since N f (I dUi) is an equational theory
=? Nf(IdUI) U {e} = Nf(IdU2) U {e}
=? Con(Nf(IdUI) U {e}) = Con(Nf(IdU2) U {e})
=? IdU1 = IdU2, by Lemma 2.4
=} ul = U2. •
We denote by T R the trivial variety T R = Mod( x ~ y) of arbitrary
type T.
Corollary 2.8. Let k 2': 1, and let V be a non-normal variety. The

mapping Nf is a lattice-embedding of the lattice £(V) into the interval
lattice [Nk(TR), Nk(V)].
It was essential, in all our proofs, that we had a non-normal identity e

from I dV which we could use to inflate any identities to a complexity
2': k. In particular, we could not have carried out these proofs for
a non-normal variety V. In fact, the maps Nf are not in general
injective, as the following example shows.
Example 2.9. Let V be the type (2) variety Mad( {f(x, f(y, z)) ~
f(f(x, y),
z ), x ~ f(f(x, x), f(f(x, x), f(x, x))), f(x , y) ~ f(y, x)} ), and let the
complexity function c be the depth function. Since V satisfies identities
of varying depths, we have V C Nf(V) C Nf(V) C Nf(V), and in
particular V is not normal. However, by Lemma . { 2 the map N f has
Thus although Nf is injective on the lattice of subvarieties of V, it is

not injective on the lattice £( Nf(V)) of subvarieties of Nf(V) .
This example also illustrates another feature of the k-normalizations.

When V is a non-normal variety and e is any non-normal identity of
V, we have seen in Lemma 2.4 that the normal identities of V along
with e are enough to produce all the identities of V. An important
consequence of this is that the (usual) normalization of V covers V
in the lattice £(r): any variety W which has V ~ W c Nf(V)
must satisfy all the normal identities of V plus at least one addi-
tional non-normal identity of V, and hence must equal V. But this
argument does not extend to the k-normalization case for k > 1. In
the previous example, adding the one additional non-normal identity
x ~ f(f( x,x ),f(f(x,x),f(x,x))) to the set N.f (IdV) gives all the
identities of V, but Nf(V) does not cover V. We can however prove
the following special case, analogous to Proposition 1 of [Gra].
Lemma 2.10. Let V be a non-normal variety, with non-normality

witness e of the form x ~ t(x, ... , x) for some term t of complexity
p 2': 1. Let W be a variety such that W ~ N f (V). If W is non-normal,
then W ~ V.
Proof: We will show that W must also satisfy the identity e. Then us-
ing Lemma 2.4 and the fact that W ~ N;((V) means that IdN;((V) ~
IdW, we have
IdV = Con(IdN;((V) u {e}) ~ Con(IdWU {e}) ~ IdW,
and hence W C V.
To show that W satisfies e, we will show that it can be deduced from
the set N{(IdV) of identities along with some non-normal identity f
of W. If e = f this is obvious, so we assume that e =/::- f. We can write
f as x ~ p(x, ... , x) for some term p with complexity r ~ 1, and by
inflating if necessary we may assume that r > k . Now consider the
sequence
x ~ p(x, ... , x) ~ p(t(x , ... , x), ... , t(x , . . . , x)) ~ t(x , ... x).
The first identity is just f itself; the second one is a consequence of
e E I dV and has complexity on each side ~ r > k , and the third
identity is a consequence of f . This shows that we can deduce e, or
x ~ t(x, ... , x), using only identities from Nf(I dV) U {f}. These
identities are all in IdW, and hence W satisfies e. •
This Lemma has an important consequence in the case k = 1. For

any variety W ~ Nf(V) , either W is normal, or W ~ V. When W is
normal we have Nf(TR) ~ W = Nf(W) ~ Nf(V); otherwise when
W is not normal we have T R ~ W ~ V. Graczynska used this fact
to show that the lattice of subvarieties of Nf(V), for a non-normal
variety V, consists of two copies of the lattice .C(V): the non-normal
subvarieties of V, and for each of these its normalization which is a
normal subvariety of Nf(V).
We ask then whether this result holds for k > 1: is it true that if
W ~ N;((V), then either W is k-normal or W ~ N;(_1 (V)? If this
were the case, we could characterize the subvariety lattice .C(N;( (V))
as consisting of k copies of the lattice .C(V). However, the following
example shows that this is not true.
Example 2.11. We take our type to be (2), with one binary op-
eration symbol f. Let V be the variety V = M od(f (x, f (y , z)) ~
f(f(x, y), z), x ~ f(x , x)) . This is the variety of idempotent semi-
groups, or bands. Taking complexity to be the depth, this variety is
382 K. Denecke, 5. L. Wismath
not normal, and hence not k-normal for any k ~ 1. We will fix k = 2,
and consider the varieties V C Nt(V) C Nf(V). Since both terms in
the associative law f(x , f(y , z )) ~ f(f(x, y) , z ) have depth 2, the va-
riety Nf(V) is still a variety of semigroups. We will follow the usual
convention of writing semigroup terms as words (with the binary op-
eration denoted by juxtaposition and brackets omitted}.
Now let W be the subvariety of Nf(V) defined by the set of identities
N .f (I dV) plus the one additional identity xi ~ xf. This additional
identity is a normal identity of V, and the term xi has depth 1, so
we have Nt(V) ~ W C Nf(V), and W is a normal variety. But it
t
is also true that N (V) is properly contained in W , since the normal
identity x 1 x 2 ~ xix 2 holds in Nt(V) but cannot be deduced from the
identities of W.
This example shows that it is possible to have a non-normal variety V
and a normal variety W with W c Nt"(V) but W neither k-normal nor
a subvariety of Nt_1 (V). Again this highlights the special properties
of non-normal identities. For any variety W with W ~ Nt(V), either
all identities of W are normal, or adding any one non-normal identity
to the set of all normal ones gives all the identities of V. At higher
levels, this no longer happens: our example shows that we can add
one non-2-normal identity to the 2-normal identities of V , without
obtaining all such non-2-normal identities.
As we remarked above, our proofs only apply in the situation of a non-
normal variety V, where we can use a non-normal identity to "inflate"
other identities. Another consequence of this is that we cannot carry
out any of our work at a higher level, that is for varieties V which are
normal but not k-normal for some k ~ 2. A non-k-normal identity
s ~ t can only be used to inflate terms in which an instance of s or t
occurs, not arbitrary terms. Consider as an example a type (2) normal
variety which satisfies an identity of the form f(x , x ) ~ f(x , f(x , x )).
There is no way to use this to inflate the term on the left of the identity
f(x, y) ~ f(f(x, y), f( x, y)). In this respect normality is a very strong
property.
3 Structural Theorems
Our proofs so far have focussed on identities satisfied in V and Nt(V),
to generalize the results of Graczyri.ska. A different approach was taken
by Mel'nik in [Mel], where a structural characterization of the algebras

in Nf(V) was given. In this section we generalize this characteriza-
tion, and consider Mel'nik's results on finding an equational descrip-
tion of Nf(V) in terms of a basis of identities for V. Throughout this
section we assume a fixed type T of algebras and identities.
Definition 3.1. Let k ~ 1. For any algebra B of type T, we define
nk(B) = {b E B I b = tB (bl, ... 'bm) for some m -ary term t with

c(t) ~ k, and some b1 , ... , bm E B}.
We set 0. 0 (B) =B.
Lemma 3.2. For any algebra B, we have
B = flo(B) ~ flt(B) ~ fl2(B) ~ 03(B) ~ ... ,
and each flk+l (B) is a subalgebra of flk (B) and of B.
Proof: Let f be an operation symbol of type T, of arity n. Let

bt, ... , bn E flk(B). Then there are terms t1, ... , tn and elements bij E
B such that bi = tf (xi 1 , ... , XimJ, for 1 ~ i ~ n, and with each term
ti of complexity~ k. Then fB(bt,···bn) = fB(tf(xu, ... ,xlm1 ), ••• ,
t~(Xnt 1 ••• , XnmJ) = wB(bu, ... , bnmn), where w is the term fB(tt, ... ,
tn) and has complexity ~ k. This shows that fB(b 1 , •.. , bn) is in
flk(B), and hence that flk(B) is a subalgebra of B. •
Definition 3.3. Let k ~ 1. Let A and B be algebras of type T. We

call B a k-normal extension of A if
(i) A is isomorphic to nk (B), and
(ii) flk(B) is a retract of B , that is, there is a surjective homomor-
phism r.p: B-+ fl(B) which is the identity mapping on flk(B).
Fork= 1, Definitions 3.1 and 3.3 give exactly Mel'nik's definitions of

a normal extension of an algebra. Now we want to generalize Mel'nik's
theorem that for V a non-normal variety, an algebra is in Nf(V) iff
it is a 1-normal extension of an algebra in V.
Theorem 3.4. Let V be a non-normal variety, and let k ~ 1. Then

an algebra B is in N,t(V) iff B is a k-normal extension of an algebra
in V.
Proof: =?: Let BE N;((V). Let x ~ u(x, . . . , x) be a non-normality

witness for V for which the complexity of u is 2: k. We define a map
cpu on B by cpu(x) = u(x, ... , x). We will show that D.k(B) is in V , and
that the mapping cpu : B ---+ D.k(B) is a homomorphism which is the
identity mapping on D.k(B). Our proof is based is on the key fact that
B, and hence also its subalgebra D.k(B), are in N;((V), and so satisfy
all the identities s ~ t of V for which both s and t have complexity
~ k. In particular, they satisfy all consequences of x ~ u(x , ... , x)
which have complexity 2: k .
To show that Ok(V) is in V , we will show that D.k(B) satisfies all the
identities of V. We have just remarked that it satisfies all such identi-
ties of complexity 2: k , and by Lemma 2.4 it will suffice to show that
Ok(B) also satisfies the non-normal identity x ~ u(x, ... , x). That is,
we need to show that for any bE Ok(B) , we have b = u(b, ... , b). But
bE Ok(B) means that there is a term t of some arity m and complexity
2: k, and some elements b1, . .. ,bm E B such that b =t 8 (b 1, ... , bm) ·
Then
is a consequence of x ~ u(x , ... , x) in which both sides have complex-

ity 2: k. Hence(*) holds in Band b = u(b, .. . , b) .
Next we show that the map cpu is a homomorphism on B. Let f be any
operation symbol of type T , of arity n . We must show that for any
X1, ... , Xn in B, we have cpu(J(xl , ... , Xn)) = f(cpu(xi) , . .. , cpu(xn)).
This is equivalent to u(J(x 1, . . . , Xn) , . . . , J(x 1, .. . , Xn)) = J(u(x 1, .. . ,
xi), ... , u(xn, ... , Xn)). But this is a consequence of x ~ u(x , ... , x)
with complexity on both sides 2: k , and hence does hold in B .
To finish this direction of the proof we must show that cpu is the iden-
tity on Ok(B). For any b = t 8 (b1 , ... , bm) in Ok(B) , we have cpu(b) =
u(b, . . . , b)= u(t 8 (b1, . . . , bm) , ... , t 8 (b1 , ... , bm)) = t 8 (b1, ... , bm) = b,
using the fact that u(t(x1 , ... , Xm), ... , t(x1 , ... , Xm)) ~ t(x1 , . .. , Xm)
is a consequence of x ~ u(x, ... , x) with complexity on both sides 2: k.
{:::: Let B be a k-normal extension of some algebra A in V. Then A
is isomorphic to Ok(B) , which is then also in V , and there is a ho-
momorphism cp on B which is the identity on D.k(B). To show that
B is in Nk (V), we must show that B satisfies any identity s ~ t of
V in which both s and t have complexity 2: k. This condition means
that for any b1, ... , bm in B , we have t 8 (b1 , . .. , bm) and s 8 (b1 , ... , bm)
both in rlk(B), and since rlk(B) is in V it satisfies s ~ t. Therefore

s 8 (b 1 , ... , bm) = t 8 (b 1, .. . , bm) holds, and B satisfies s ~ t. •
It follows immediately from Lemma 3.2 that if B is a k-normal ex-

tension of an algebra A isomorphic to rlk(B) in V , then B is also a
p-normal extension for any p 2: k .
Now we want to use this characterization theorem, as Mel 'nik did for
k = 1, to describe the identities of Nt(V) in terms of the identities
for V. We deal first with a special case. Let k 2: 1, and let u be
a term with depth 2: k . Let V be the non-normal variety M od(x ~
u(x, ... , x)).
For each operation symbol j , of arity n , from our type, let E 1 be the
set of identities
j(u(x1, ... , x1), x2, ... , Xn) ~ f(xl, u(x2 , ... , x2), X3, ... , Xn) · · ·
~ f(xl, ... , Xn - 1, u(xn, ... , Xn)) ~ u(f(x1, ... , Xn) , ... , f(xl , ... , Xn)).
Let r be the set of all identities of the form u(t(xl , ... ' Xm), ... ' t(xt ,
... , Xm)) ~ t(x 1 , ... , Xm) , where tis an m-ary term of complexity 2: k.
Denoting by F the set of all operation symbols of type T , we set E =
U Et Ur.
/EF
Lemma 3.5. Let k 2: 1, and let V = Mod(x ~ u(x, ... , x)) with c(u)
2: k. Let E be the set of identities constructed above. Then Nt(V) =
ModE.
Proof: All the identities in E are consequences of the identity x ~

u(x, ... , x) in V, and have complexity 2: k on both sides, so we have
E ~ Nf(IdV) and hence Nt(V) ~ModE. For the opposite inclusion,
let B be any algebra in ModE. We want to show that B is a k-
normal extension of rlk(B) E V, which by Theorem 3.4 will give us
B E Nt(V) as required.
The identities in r guarantee that rlk(B) satisfies X ~ u(x, ... 'x) ,
and so rlk(B) is in V. We define a mapping <fJu : B -t B by <fJu(x) =
u(x, .. . , x). It remains to show that <fJu is a homomorphism on B
and is the identity mapping on rlk(B). For any operation symbol
j, of arity n, and any b1, .. . , bn in B , we have <fJu(f(b1, . .. , bn)) =
f(r.pu(bi), ... , <fJu(bn)) iff u(f(b1 , ... , bn) , ... , f(bl , ... , bn)) = f(u(bl,
... , bl), ... , u(bn, ... , bn)). But this is a consequence of the identities
in E f, and does hold in B. This makes <fJu a homomorphism on B .
The identities in r then make 'Pu the identity mapping on rlk(B), since
any element of this set has the form t 8 ( b1 , ... , bm) for some term t and
elements b1, ... , bm. •
Notice that for k = 1, the set r of identities can be replaced by a

simpler set of identities, those of the form u(J(x1, ... , Xn ), ... , f(x 1,
... , Xn)) ~ f(xl, ... , Xn) for fan operation symbol of type T of arity n ;
it is clear that these identities yield as consequences all the identities
of r. We can produce similar simplifications for k = 2.
Next we consider the more general case of any non-normal variety
V. We consider a variety having an equational basis which can be
divided in a k-normal part I: and in non-k-normal identities of the
form x ~ Uj (x, · · · , x) for terms Uj of type T. For each term we form
u
the sets I;~: ' :E~j = I;/; and ruj ' I.;Uj = :E~j u fUj . Then we have
i E/
Theorem 3.6. Let k 2: 1 and let V = Mod{L:U U{x ~ uj(x, · · · , x)}}

jEJ
be a non-k -normal variety where v (Uj) 2: k and where I: consists only
of k-normal equations.
Then Nt(V) = M od{:E U U I.;ui}.
jEJ
Proof. Since N{;(IdV) U D.wT(X) is an equational theory and since

IdNt(V) = IdModN{;(IdV) we have to show that N{; (IdV)Ub.wT(X)
= E(:E U U I.;ui), i.e. N{;(E(L: U U {x ~ Uj(x, · · · ,x)})) U D.wT(X) =
jEJ jEJ
E(:E U U I.;ui), where E denotes the operator which assigns to each
jEJ
set of equations its closure under application of the five rules of deriva-
tions for equational theories. The generalization of Lemma 3.5 gives
E(N{;(E(L:)) U U N{; (E({ x ~ uj(x, · · · ,x)})) = N{;(E(L: U U {x ~
jEJ jEJ
uj(x, · · · , x)} )) U D.wT(X) · Since I: consists only of k-normal equa-
tions, the set of all consequences of I:, i.e. E(:E) consists also only
of normal equations. Therefore N{;(E(L:)) U D.wT(x) = E(:E). Then
by previous results we have N{;(E({x ~ uj(x, · · · ,x)})) = I.;ui,j E J
and therefore U N{;(E({x ~ uj(x,··· ,x)})) = U I.;ui . This gives
jEJ jEJ
E(E(:E) U U I.;ui) = N{; (E(L: U U { x ~ uj(x, · · · , x )}) U D.wT(x). The
jEJ jEJ
sets I.;ui are equational theories since N{; (E({ x ~ uj(x , · · · ,x)}) U
~w.,.(x) are equational theories. Therefore E(~ui ) = ~ui and this gives
E(E(~) U U E(~ui)) = Nff(E(~ U U{ x ~ uj(x, · · · , x)}) U ~W.,. (X) ·
j EJ j EJ
The left hand side can be replaced by E(~ U U ~ui) U ~w.,.(x) , and
j EJ
this finishes the proof. •
The following example shows that without the assumption of Theorem
3.6 we will not get a basis for the k-normalization.
Example 3.7. We use type (2) and complexity equal to depth, and
since our varieties will all satisfy associativity we write terms as words.
Using k = 1 and the non-normal identity x ~ x 2 , the set ~ above
consists of the 4 identities
since as remarked above we may omit r .

Let() = {x(yz) ~ (xy) z , x ~ x 2 ,x ~ xy} . This is a basis for the
variety LZ of left-zero semigroups. But Nf (()) contains only the as-
sociative law, so our potential basis ()* = ~UNf (()) contains only the
associative law and the four identities from ~ - But all thes e identities
are regular (have the same set of variables occurring on each side} and
hence we cannot deduce the non-regular identity x y ~ x z from them .
This identity does hold in LZ and in Nt(LZ), so we conclude that()*
is not a basis for Nt(LZ).
References
[ADP] Arworn, S., K. Denecke and R. Poschel, Closure Operators on
Complete Lattices, Ordered Algebraic Structures: Nanjing; Algebra,
Logic and Applications Series Volume 16, Gordon and Breach Science
Publishers, 2001 , 1 - 22.
[Chr] Chromik, W ., Externally Compatible Identities of Algebras,

Demonstratio Mathematica 23, 1990, 345 - 355.
[DW] Denecke, K. and S. L. Wismath, Galois Connections and Com-

plete Sublattices, this volume pp. 49-73, 2001.
[Gra] Graczynska, E., On normal and regular identities and hyperiden-

tities, Proc. Universal Algebra and Applied Algebra (Turawa 1988),
World Scientific Publishing, Singapore, New Jersey, London, Hong
Kong 1989, 107-135.
[Mel] Mel'nik, I. I., Nilpotent shifts of varieties (in Russian) , Mat. Za-
metki 14, No. 5, 1973. English translation: Math. Notes 14, 1973,
962-966.
[Plo] Plonka, J. On Varieties of Algebras Defined by Identities of Some

Special Forms, Houston Journal of Mathematics, Vol 14, no. 2, 1988,
253- 263.
Authors' addresses:
Klaus Denecke
University of Potsdam
Institute of Mathematics
Am Neuen Palais
14415 Potsdam
Germany
e-mail: kdenecke@rz. uni-potsdam.de
Shelly Wismath
Dept.of Mathematics and CS
University of Lethbridge
Lethbridge, Ab.
Canada T1K-3M3
e-mail: wismaths@cs.uleth.ca
Iterated Galois Connections in
Arithmetic and Linguistics
J. Lambek
Abstract
Galois connections may be viewed as pairs of adjoint functors, specialized
from categories to partially ordered sets. We study situations that permit
iterations of such adjoints. While their occurrence in elementary number
theory is a curiosity, they play a crucial role in a new algebraic approach
to sentence structure in natural languages.
AMS classification: 18A40, 06A15,03B65.
Key words: Adjoint functors , Iteration of adjoint functors, Sentence struc-
ture in natural languages.
1. Adjunction and complementation. We begin by looking at
two well-known functions in number theory:

p(n) = the nth prime (for good measure, we put p(O) = 0),
1r(n) = the number of primes ~ n.
Inspection of the following table leads to a curious observation
first made in [1]:
n p(n) p(n) +n K(n) 1r(n) +n+1

0 0 0 0 1
1 2 3 0 2
2 3 5 1 4
3 5 8 2 6
4 7 11 2 7
5 11 16 3 9
389
390 J. Lambek
We note that the sets
{p(n) + nln EN} , {1r(n) + n +linEN}
are complementary subsets of N. The proof, though tricky, is quite

easy. It has nothing whatever to do with properties of prime num-
bers and depends only on the fact that p and 1r constitute a Galois
correspondence:
p(x) ::; y {::} x ::; 1r(y) .
We borrow the terminology of category theory and call p the left ad-
joint of 1r.
In general, let f, g : N --7 N be order preserving functions such that
f( x ) ::; y {::} x ::; g(y) ,
then
f(x) +x::; x+y {::} x+y::; g(y) +y
{::} g(y)+y+l>x+y.
It follows immediately that the sets
(i) F = {f( x ) + xix EN} , G + 1 = {g(y) + y + lly EN}
have no elements in common. Moreover, consider the range
0 ::; x ::; n , 0 ::; y = n - x ::; n .
Then F and G + 1 together have exactly n + 1 elements between 0

and n, hence all the elements between 0 and n . Thus F and G + 1 are
complementary sets.
We leave as an exercise the converse observation: if F and G are
infinite subsets of N so that F and G + 1 are complementary (hence
0 E F), then F and G have the form (i) with f left adjoint to g.
2. Iterated adjoints.
When does an order preserving function g : N --7 N have a left adjoint?

More generally, let
f : (X,::;) --7 (Y, ::;) , g : (Y, ::;) --7 (X , ::;)

Iterated Galois Connections in Arithmetic and Linguistics 391
be a Galois connection, that is
(ii) f(x) ::; y {::} x::; g(y).
This equivalence may also be written:
(iii) f(x) = n{y E Ylx::; g(y)}.
Given g order preserving, we can find its left adjoint f = l if and

only if
(a) g preserves infs,
(b) the inf (iii) exists.
This is a special case of Freyd's Adjoint Functor Theorem in category
theory.
Return now to the special case X = Y = N. Then
f(x) = min{y E Nix ::; g(y)}.
In this case, (a) holds trivially and (b) holds if and only if {y E Nix ::;
g(y)} =/=- 0 for each x in N, that is, provided g is unbounded.
We note that then f = l will also be unbounded, since f(x) ::; b
would imply x::; g(b). Therefore, ge also has a left adjoint gee, and so
on.
In summary, if g : N----+ N has a left adjoint gf-, then it also has iterated
left adjoints gfe, ge.ef etc.
One way to construct these is to look at the corresponding subsets of
N:
The functions
correspond to the sets
where Fe denotes the complement of F.

What about the right adjoints? (ii) may also be written thus:
(iv) g(y) = U{x E Xlf(x)::; y}.
Given j, its right adjoint g = r exists if and only if

392 J. Lambek
(a') f preserves sups,

(b') the sup (iv) exists.
In the special case X = Y = N, U = max, hence
(a') holds if and only if f(O) = 0 (the maximum of the empty set
being 0),
(b') holds if and only if f is unbounded.
Thus the order preserving function f : N -t N has a right adjoint _r
if and only if f(O) = 0 and f is unbounded.
What about iterated right adjoints?
correspond to the sets
F, Fe - 1, (Fe - 1Y - 1, · · · .
Now
_r(o) = 0 iff 0 E Fe- 1, i.e. 1 E Fe, i.e. 1 (j_ F,
_rr(O) = 0 iff 0 E (Fe- 1Y- 1, i.e. 1 E (Fe- 1)e,
i.e. 1 (j_ Fe- 1, i.e. 2 (j_ Fe, i.e. 2 E F.
Continuing in this way, we see that iterated right adjoints exists if

and only ifF = {f(x) + xix E N} is the set of even numbers, i.e.
f(x) + x = 2x, i.e., f is the identity function.
For example, iff = p, pr = 1r and prr = 1rr , where 1rr (y) = p(y + 1) -1.
But 1rr(O) = 1 =/= 0, so prrr = 1rrr does not exist.
However, let .f : Z -t Z be order preserving and unbounded on both
sides. Then f is left adjoint to g if and only if
F = {f(x) + x jx E Z}, G + 1 = {g(y) + y + 1jy E Z}

are complementary subsets of Z (see [2]). In this case all iterated
left and right adjoints exist, as was pointed out in [3]. Note that,
in general, f · and _r are distinct. For example, if .f(x) = 2x, then
_r(x) = [x/2] and .f£(x) = [(x + 1)/2].
3. Adjunction in 2-categories.
Adjoints are usually defined in the 2-category of all small categories,

where
0- cells small categories,
1 -cells functors ,
2- cells natural transformations.
The usual definition carries over to any 2-category: let f : A ---t B
and g : B ---t A be 1-cells, then f is left adjoint to g if and only if
there exist 2-cells
E : fg ---t 1s
such that
(v)
As perhaps the simplest example of a 2-category, let us look at any

partially ordered monoid, where
0 - cells just one,

1 -cells = elements,
2 - cells : f ::=; g .
(Note that there is at most one 2-cell between two 1-cells and we write
f ::=; g for f ---t g.) Here elements f and g form an adjoint pair if and
only if
1 :::; gf, fg:::; 1.
The equations (v), like other equations between 2-cells, are automat-
ically satisfied.
As an example consider the partially ordered monoid of all order pre-
serving functions N ---t N under composition. Then f is left adjoint to
g provided
f(x) :S y {::} x :S g(y).
4. Pregroups.
We shall introduce a couple of definitions.

394 J. Lambek
A left pr-egroup is a partially ordered monoid in which every element

has a left adjoint. Of course, every partially ordered group is a left
pregroup in which fe = f- 1 . More interesting is the partially ordered
monoid of all unbounded order preserving functions N --+ N.
A right pr-egroup is a partially ordered monoid in which every element
has a right adjoint. Finally, a pregroup is both a left and a right
pregroup. Again, every partially ordered group is a pregroup, but
so is the partially ordered monoid of all order preserving functions
Z--+ Z which are unbounded on both sides. In this example, jl-:/::- r
in general.
Recent applications to linguistics make use of the free pregroup gener-
ated by a partially ordered set of so-called basic types. Given a basic
type a, one forms simple types
Compound types, or just types, are strings of simple types, say
the ai being simple types. In particular, when n = 0 one obtains the

empty string 1.
The types form a monoid under concatenation which is partially or-
dered due to the order of basic types and the following contractions
and expansions
Both contractions and expansions are needed to prove that the com-
pound types form a pregroup with adjoints
Without loss of generality, one may assume that all contractions pre-
cede all expansions.
It follows that to show A ::; /3, where f3 is a simple type, no expansions
are needed. In particular, to show that a string of words of compound
type A is a sentence of type s in a natural language it suffices to check

that A ::; s by repeated contractions.
5. Linguistic applications.
We shall discuss briefly how free pregroups may help to investigate

certain aspects of three European languages:
(1) Chomskyan traces in English [3,4],
(2) word order in German [5],
(3) clitic pronouns in French [6] .
It so happens that the same list of basic types will do for the fragments
of the languages discussed here.
1fj = _j-th person pronoun (_j = 1, 2, 3),

s1 = declarative sentence in present tense,
o= direct object,
P2 = past participle,
q1 = yes-or-no question in present tense,
q= yes-or-no question, q1 ::; q,
w = wh-question,
'l = infinitive of intransitive verb.
English.
I see her
7fl (7rrsld') 0 ::; sl
L...____j L..__j
I have seen her

1r1 ( 1rr s1p~)(p2o£)o
L...____j '-----.1 '---..J
::; s1
have I seen her?

(qlp~1rf)1f1 (p2o£) o ::; ql
I L.__jI L..__j
whom have I seen - ?

(wouq£)(qlp~7rf)1fl(P2o£) ::; w
I L--J I L.__j II
Note that q£q1 ::; q£q ::; 1. The dash here represents a Chomskyan
trace, which is put in for comparison only. In writing whom rather
396 J. Lambek
than who, I am following the late Inspector Morse.
she is seen -
1r3 ( 1r3r s1 o££P2f.) (P2o£) < s1
'----...J I L.__j I
she has been seen -

1r3 (1r3slp~)(p2oap~) (p2o£)
'----...J L.__j I L.__j I
German.
du siehst ihn
1r2 (1r2s1o£) o ::::; s1
'----...J L...._j
siehst du ihn?
(q1o£7r~)(1r2 o) ::::; ql
I L...____j I
ich habe ihn gesehn

1r1 (1rrs1p~ ) o (orp2) ::::; s1
'----...J I L_____j I
ich kann ihn sehen

1r1 (1rrs1i£) o (ori) ::::; s1
'----...J I L_____j I
er kann gesehn werden

7f3 r(0 r P2 ) (P20
·£)
( 1f3S(l r rr z·)
'----...J I I l___j I I
kann er gesehn werden ?
1r3 (or P2 ) ( p 2r orr z·) <
f)
( qlz·£1r3 _ ql
I '----...J I l___j I I
French.
j e veux dormir
7fl (7fr sl i~') i ::::; sl
L..______j L___j
j e veux voir Jean

7fl (7fr sl i~') (io£) 0 ::::; sl
'----...J t_j L...._j
je veux le voir
1f 1
L_____j
( nr sl ie) (ioeeie) (ioe)
L___j I L__j I
::::; sl
Note that the three languages considered here all require the simple
type oee, but German also requires orr. So far, I have not come across
a language that requires oeee or orrr.
References
1. J. Lambek and L. Moser, Inverse and complementary sequences

of natural numbers, Amer. Math. Monthly 61 (1954) , 454-458.
2. J. Lambek, Some Galois connections in elementary number the-

ory, J. Number Theory 47 (1994) , 371-377.
3 ........ , Type grammar revisited , in: A. Lecomte et al. (eds) ,

Logical aspects of computational linguistics, Springer LNAI 1582
(1999) , 1-27.
4 ........ , Pregroups: a new algebraic approach to sentence structure,

in: C. Martin-Vide et al. (eds) , Recent topics in mathemati-
cal and computational linguistics, Editura Academici Romane,
Bucharest 2000.
5........ ,Type grammar meets German word order, Theoretical Lin-

guistics 26 (2000), 19-30.
6. D. Bargelli and J. Lambek, An algebraic approach to French

sentence structure, in: Groote et al. (eds.) , Logical aspects of
computational linguistics, Springer LNAI 12099 (2003) , 62 - 78.
Author's address:
J. Lambek, McGill University,
Montreal, Canada
Deductive Systems and Galois
Connections
I. Chajda, R. Halas
Abstract
The concept of a b-deductive system will be introduced. An interconnec-

tion between b-deductive systems and congruence kernels will be shown.
We define a Galois connection between sets of binary term functions and
systems of subsets of a given algebra and study closed sets with respect
to the induced closure operators and the algebraic properties of systems of
subsets related to a given term function.
AMS Mathematics Subject Classification: 08A30, 06A15,08B05.
Key words: b-deductive system, Congruence kernel, Fichtner term, Hilbert
term, Galois connection.
1 Introduction
The concept of a deductive system was introduced by A. Diego [2]

for so-called Hilbert algebras which can be considered as an algebraic
counterpart of intuitionistic logic. More precisely, a Hilbert algebra is
an algebra 1i = (H; ·, 1) of type (2,0) satisfying the following axioms:
(H1) x · (y·x) = 1
(H2) (x · (y · z)) · ((x · y) · (x · z)) = 1
(H3) X· y = 1 andy· X= 1 imply X= y.
Hilbert algebras can be considered as a fragment of any propositional

logic containing only the logical connective implication and the con-
stant 1 which is considered as the truth-value "true". As usual, instead
399
K. Denecke et at. (eds.), Galois Connections and Applications, 399411.
400 I. Chajda, R. Halas
of the symbol "*", generally used for implication, we will here (for
the sake of brevity) use " ·" .
It was shown by A. Diego [2] that the operations of a Hilbert algebra
satisfy the following properties:
(1) X· X= 1, 1 = 1, 1 ·X=
X· X
(2) x · (y · z) = y · (x · z)
(3) x · (y · z) = (x · y) · (x · z).
Moreover, on the base set of a Hilbert algebra a binary relation :::; can
be introduced by setting
x ::::; y if and only if x · y = 1
and, due to (H3), (1) , (2) and (3) , it is a partial order on H with the
greatest element 1.
Now, deductive systems on a Hilbert algebra 1l = (H; ·, 1) are subsets
containing the constant 1 and closed under the inference rule Modus
Ponens; formally, a subset D ~ His a deductive system of 1l if
1 ED and
a E D and a · c E D give c E D.
This concept was generalized from Hilbert algebras to weakly regular
algebras by the first author in [1]. In what follows we get another
definition which is a generalization of the previous one for Hilbert
algebras but not so restrictive as that in [1] for investigations of Galois
connections.
2 b-deductive systems
From now on let A = (A , F) be an algebra and let T 2 (A) denote
the set of all binary term functions of A . We introduce the following
concept:
Definition 1. Let A= (A , F) and b E T 2 (A). By a b-translation is
meant every unary func tion
where f E F is n-ary, i E {1 , ... , n} and a 1 , . . . , an E A (i.e. every

translation) and every unary function
r(x) = b(x, a) or r( x) = b(a, x)
Deductive Systems and Galois Connections 401
for every a E A.
We are ready to introduce our main concept:
Definition 2. Let A = (A , F) be an algebra with a constant 0 and
b E T 2 (A). A subset D <:;;; A is called a b-deductive system of A
whenever
(i) 0 ED
(ii) a E D and b(a, c) E D , b(c, a) E D imply c E D
(iii) a E D implies b(O , a) E D and b(a, 0) E D
(iv) if b(a, c) ED and b(c, a) ED then also b(T(a) , T(c)) ED for every
b-translation T .
Remark 1. If 1l = (H; ·, 1) is a Hilbert algebra and its constant
1 is considered as 0 in the previous definition, then the concept of
b-deductive system for b(x, y) = x · y coincides with the already men-
tioned one. It is an easy exercise to verify that the condition (iii) is
trivially satisfied with respect to (1) and, due to (2) and (3) , also (iv)
is evident. It is not too complicated to show that also (ii) is satisfied.
At first , we will study the interconnection between b-deductive systems
and congruence kernels. For this, we introduce the following concept:
Definition 3. Let A= (A, F) be an algebra with 0 and b E T 2 (A) ,

let D <:;;; A. A binary relation e D on A defined by the setting
(a,c) E 8D if and only ifb(a,c) ED and b(c,a) ED (*)
will be called b-induced by D.

Theorem 1. Let A= (A, F) be an algebra with 0 and bE T 2 (A) , let
D <:;;;A. If the relation eD, b-induced by D, is a congruence on A and
D = [O]ev, then D is a b-deductive system of A.
Proof. Suppose 8D E Con A and D = [O]ev· Clearly D satisfies (i).
We prove (ii): let a E D, b(a, c) E D and b(c, a) ED. Since D = [O]ev ,
we have (a, 0) E 8 D and, by (*), also (a, c) E 8 D thus also (c, 0) E 8 D
proving c E [O]ev =D.
For (iii) , suppose a E D. Then (a, 0) E 8D and, by ( *), we conclude
b(O, a) E D and b(a, 0) E D. Now we prove (iv): let b(a, c) E D
and b(c, a) E D . Then (a, c) E 8D and, due to 8D E Con A , also
(T (a) , T (c)) E 8 D for every b- translation T. Applying (*) once more,
we have b(T(a), T(c)) ED. D
Theorem 2. Let A = (A , F) be an algebra with 0 and assume that

bE T 2 (A) satisfies b(x, x) = 0. If D is a b-deductive system of A then
the relation 8D, b-induced by D , is a congruence on A and D = (O]ev·
Proof. By using the identity b(x, x) = 0 and (i) of Definition 2, 8D

is clearly reflexive. Symmetry of e D is evident. We prove transitiv-
ity: let (a,c) E 8D and (c, d) E 8D. By(*) we have b(a, c) ED,
b(c, a) E D and b(c, d) E D , b(d, c) E D . Consider the b-translation
T(x) = b(d, x). Due to (iv) , b(c, a) E D and b(a, c) E D imply
b(T(c), T(a)) E D and b(T(a) , T(c)) E D , i.e. b(b(d, c) , b(d, a)) E D
and b(b(d, a), b(d, c)) ED. Applying (ii) we conclude b(d, a) ED.
If we use the b-translation T(x) = b(x , d) , we obtain analogously
b(a,d) ED. Hence,(*) implies (a, d) E eD, i.e., eD is also transitive.
Applying (iv) we easily derive that eD has also the substitution prop-
erty with respect to every translation and because it is reflexive and
transitive, also with respect to every f E F , i.e., 8D is a congruence
on A.
If a E D then, by (iii) , b(O, a) E D and b(a, 0) E D thus (a, 0) E 8D
giving a E [O]ev · Conversely, if a E [O]ev then (a, O) E 8D and, by
(*),also b(O,a) ED and b(a, O) ED. By (i) we have 0 ED and ,
applying (ii) we obtain a ED. Thus D = [O]ev · D
Remark 2. If Dis a b-deductive system of A and b(x, x ) = 0 holds in

A then 8D is the greatest congruence on A with the kernel D. Indeed,
suppose E Con A and [O].p = D. If (a, c) E then
(b(a, c) , 0) = (b(a , c) , b(a, a)) E and
(b(c, a) , 0) = (b(c, a) , b(a, a)) E 

thus b(a, c) ED and b(c, a) ED giving (a, c) E 8D , i.e., ~ 8D.
3 Fichtner terms
An algebra A = (A, F) with 0 is said to be weakly regular , if each

congruence on A is determined by its kernel , i.e. if e, E Con A
then
[O]e = [O].p implies 8 =.
A variety V with 0 is weakly regular if every A E V has this property.

Weakly regular varieties were characterized by K. Fichtner [3] as
follows:
A variety V with 0 is weakly regular if and only if there exist binary
terms b1 , ... , bn (n ~ 1) such that
b1 (x,y) = · · · = bn(x,y) = 0 if and only if x = y. (**)
One can immediately check that the variety of all Hilbert algebras
is weakly regular (where 1 is considered as our constant) due to the
condition (H3) . Moreover, we need only one term and n = 2 in the
Fichtner condition (**)since one can take b1 (x, y) = x·y and b2 (x, y) =
b1 (x, y) = y · x (It was shown by A. Diego [2] that the class of all
Hilbert algebras is a variety). This motivates us to introduce the
following concept:
Definition 4. A binary term b of a variety V with 0 is called a
Fichtner term if
b(x, y) = b(y, x) = 0 if and only if x = y
holds in V.
Remark 3. Hence, every variety V having a Fichtner term is weakly
regular.
Let b be a binary term of a variety V and A E V. The term function
on A induced by b will be denoted by the symbol bA. Analogously for
oA.
Lemma 1. Let b be a Fichtner term of a variety V, let A E V and
8 E ConA. Th en
Proof. Suppose (a, c) E 8 for 8 E Con A. Then
(bA(c, a), oA) = (bA(c, a), bA(a, a)) E 8 and
(bA(a, c), OA) = (bA(a, c), bA(a, a)) E 8

proving bA(a, c) E [oA]e and bA(c, a) E [OA]0 .
Conversely, let bA(a, c) E [oA]e and bA(c, a) E [oA]e for some e E

ConA. In the quotient algebra A/8 we have
Since b is a Fichtner term , it yields [a]e = [c]e whence (a, c) E 8. D
We are ready to prove

Theorem 3. Let b be a Fichtner term of a variety V with 0. Let
A = (A, F) E V and D ~ A. Then D is a congruence kernel (i.e.
D = [OA]e for some 8 E Con A) if and only if D is a bA-deductive
system of A.
Proof. Let b be a Fichtner term of V, let A = (A, F) , 8 E Con A
and D = [OA] 8 . Then D satisfies (i) . We prove (ii) : if a E D and
bA(a, c) E D , bA(c, a) E D then (a , oA) E D and, by Lemma 1, also
(c, a) E 8 thus (c, OA) E 8 giving c E [OA]e =D.
For (iii) , let a E D. Then (a , OA) E 8 and , by Lemma 1, we have
bA(OA , a) ED and bA(a, OA) ED.
For (iv), let bA(a, c) E D and bA(c, a) E D and letT be a b-translation.
By Lemma 1, (a , c) E e thus also (T(a) , T(c)) E e and, once more by
Lemma 1, bA(T(a) , T(c)) ED.
Thus D is a bA-deductive system of A. The converse follows from
Theorem 2. D
Examples. (1) As already mentioned, b(x , y) = x · y is a Fichtner

term of the variety of Hilbert algebras (where 1 is the constant under
consideration).
(2) For BCK-algebras, one can take b(x , y) = x·y and easily show that
for every variety of those algebras, b is its Fichtner term. Analogously
it can be shown that it is a Fichtner term for the variety of implication
algebras.
(3) For the variety of relatively pseudocomplemented semilattices (the
so called Brouwer semilattices where 1 is the constant), b( x , y) = x * y
(the relative pseudocomplement of x with respect toy) is the Fichtner
term.
4 Galois connections
Let A= (A, F) be an algebra with 0. We can define a binary relation

R ~ T 2 (A) x Exp A (where Exp A denotes the set of all subsets of A)
as follows:
(b, D) E R if and only if D is a b-deductive system of A.
This relation R defines a Galois connection (g, a) as follows:

for a subset 1i ~ T 2 (A) we take
g(1i) = {D ~A; Dis a b-deductive system of A for each bE 1i};
if 1i is singleton, say 1i = {b}, we will write simply g(b) instead of

g({b}).
For .C ~ Exp A we take
a(.C) ={bE T 2 (A) ; each DE .Cis a b-deductive system of A}.

As one knows well, a g and ga are closure operators. We are interested
in closed sets under these operators. Forb E T 2 (A) denote the subset
g(b) of Exp A by the symbol DedA(b). Of course, g(b) = DedA (b) is
the set of all b-deductive systems of A.
The proof of the following theorem is straightforward and hence omit-
ted:
Theorem 4. Let .C ~ Exp A for an algebra A = (A , F) with 0 and
let bE T 2 (A). Then
(a) g(b) = DedA(b) is a complete lattice with respect to set inclusion
with the greatest element A where the meet operation coincides with
set intersection;
(b) Q(a(.C)) = n{DedA(b) ; bE a(.C)}.
Hence, Theorem 4 shows us the structure of closed sets under the
operator Qa . In what follows , we are interested in the second closure
operator at least in particular cases.
Theorem 5. Let b be a Fichtner term of a variety V . Let b0 be a
binary ter·m ofV such that b0 (x,x) = 0. Then b~ E a(Q(bA)) for each
A E V if and only if b0 is also a Fichtner term of V .
Proof. Let b be a Fichtner term of a variety V with 0. By Remark
3, V is weakly regular. Let A = (A , F) be an arbitrary algebra of
406 I. Chajda , R. Halas
V . By Theorem 3, {OA} E DedA (bA), i.e. {OA} E Q(bA) . Since

b~ E a-(Q(bA)), {OA} is also a b~-deductive system of A.
Suppose now b~(a, c) = oA and b~(c, a) = oA. Then b~(a , c) E {OA}
and b~(c, a) E {OA} and , by Theorem 2, (a, c) E 8v for D = {OA}.
Since V is weakly regular and the congruences 8 D and wA (the identity
relation on A) have the same kernel , then 8v = WA and hence (a, c) E
WA, 1.e. a= c.
The converse is trivial thus the equality b0 (x , y) bo(Y , x) = 0 iff
x = y holds in V and b0 is a Fichtner term of V. D
Corollary 1. Let b be a Fichtner term of a variety V and A E V.

Then Q(bA)::: ConA.
Proof. By Theorem 4, Q(bA) = DedA (bA) and , due to Theorem 3, it
is the lattice of all congruence kernels of A. Since A is weakly regular,
this lattice is clearly isomorphic to Con A . D
Remark 4. Of course, if b(x , y) is a Fichtner term of a variety V then

b0 (x, y) = b(y, x) is also a Fichtner term of V.
Example. If V is a variety of non-commutative groups then for every
natural number n the term bn(x, y) = x-n · y- 1 · xn+l is a Fichtner
term of V.
One could mention that the least element of DedA (bA) is not described
in Theorem 4 (a). However, if {OA} is that of DedA (bA) then it has
an interesting consequence in weakly regular varieties:
Theorem 6. Let V be a weakly regular variety and b(x, y) be a binary
term of V satisfying b(x, x) = 0. If {OA} is a bA-deductive system of
every A E V then b is a Fichtner term of V.
Proof. Let A = (A, F) E V and let {OA} E DedA (bA) for the binary
term function bA induced by the term b(x, y) of V satisfying b(x , x) =
0. Suppose bA(a, c) = bA(c, a) = oA. Then bA(a, c) E {oA} and
bA(c, a) E {OA} thus (a, c) E 8v forD= {OA} by Theorem 2. Since
A is weakly regular and 8v and WA have the class {OA} in common,
we have 8v = w and hence a = c. Together with b(x, x) = 0 this
yields
b(x , y) = b(y, x) = 0 iff x = y
thus b is a Fichtner system of V . D
For not necessarily weakly regular varieties we get the following result:
Theorem 7. Let V be a variety with 0 and b(x , y) be a binary term of

V satisfying b(x, x) = 0. Then the following conditions are equivalent:
(i) {OA} is a bA-deductive system on each A E V;
(ii) For each A E v and 8 E Con(A), [oA]e is a bA-deductive system
of
A;
(iii) the bA-induced relation 8{oA} is a congruence with kernel {OA}
for each A E V.
Proof. (ii)=?(i) is trivial and (i)=?(iii) by Theorem 2. It remains to
prove (iii)=?(ii). Let A = (A , F) E V. Since {OA} is the kernel of
8{oA} E ConA, we have by(*) of Definition 3
Suppose now 8 E ConA and D = [OA] 8 . Of course, oA ED. We need

to verify the conditions (ii) , (iii) and (iv) of Definition 2.
Suppose a E D and bA(a, c), bA(c, a) E D = [OA] 8 . Then we have
[a]e = [OA]e and
bAfe ([oA ]e, [c)e) = bAfe([c]e, [oA)e) = [oA)e
thus, applying (* * *) in A/8, we obtain [c]8 = [OA]e whence c E

[OA]e = D proving (ii) of Definition 2.
If a E D = [oA]e then [a] 8 = [OA]e thus, by b(x, x) = 0 we infer
bAfe([a]e, [oA]e) = bAfe( [oA]e, [a]e) = bAfe([oA]e, [oA)e) = [OA)e
thus bA(a, oA), bA(oA, a) E [OA]e = D proving (iii) of Definition 2.

Now, let bA(a, c), bA(c, a) E [OA]e = D. By the assumption, the
bA/ 8 -induced by {[oA]e} relation on A/8, i.e. 8 {[oA]e}' is a congru-
ence on A/8 thus bAI8 ([a] 8 , [c) 8 ) = bAI 8 ([c] 8 , [a) 8 ) = [OA]e whence
([a)e, [c)e) E 8{ [0A]e}' i.e. also
(T([a)e) , T([c)e)) E 8 {[oA)e}
for every b-translation T. One can easily infer [bAI8 (T(a) , T(c))]e =
bAI8 (T([a] 8 ), T([c) 8 )) = [OA]e and hence bA(T(a), T(c)) E [OA]e =D.
0
In the remaining part of the paper, we are interested in lattice proper-
ties of Q(b) = DedA (b). For this, we can ask for some more restrictive
408 I. Chajda , R. Hala?5
properties of the Fichtner term b of V . In accordance with the prop-

erty (1) of Hilbert algebras, we can introduce the concept of Hilbert
term as follows:
Definition 5. Let V be a variety with 0 and b be its Fichtner term.
We call b a Hilbert term whenever it satisfies
b(x , 0) = 0 and b(O , x) = x.

Examples. For Hilbert algebras, implication algebras and BCK-
algebras the term b(x , y) = x · y is the Hilbert term (where the con-
stant is denoted by 1 in the first two cases). For relative pseudocom-
plemented semilattices, b(x , y) = y * x is the Hilbert term.
Theorem 8. Let b be a Hilbert term of a variety V and A E V. Then
DedA (bA) is infinitely V -distributive, i.e.
holds forD, D 1 E DedA (bA) and an arbitrary index set r.

Proof. Since V has a Hilbert term b, V is weakly regular. By Corollary
1, there exists an isomorphism h between DedA (bA) and Con A given
by h(D) = 8D, i.e.
8D 1 vD2 = 8D 1 V 8D 2
(where Von the right hand side is in ConA).
Suppose D, D"~ E DedA (bA) for "'( E f and a E D n V{D'Y; "'( E f} .
By Theorem 3, there are 8 , 8"~ E Con A such that D = [OA]e , D 1 =
[OA]e')' for"'( E f. Hence, (OA , a) E 8 and, due to the compactness of
con A, there exist "'/1' ... ' "'/k E r such that (oA ' a) E e'Yl v . . . v e 'Yk.
Hence, there exist Co, Ct, ... 'Ck E A with Co = oA , Ck =a and
Then (bA(cj, a), oA) = (bA(cj , a), bA(cj , oA)) E e and

(bA(cj+l, a), OA) = (bA(cj+l , a), bA(cj+l , OA)) E 8
thus also (bA(cj, a), bA(cj+l , a)) E 8 for j = 0, ... , k- 1.
We infer
a= bA(oA, a)= bA(c0 , a)(e n 8"JbA(c1, a)(e n 8"2 )bA(c2, a)···

bA(ck,a) = bA(a,a) = OA proving (a,OA) E (8n81'1 )V· · ·V(8n81'k).
With respect to the aforementioned isomorphism, we obtain
The converse inclusion is satisfied trivially thus DedA (bA) is infinitely

V-distributive. D
Corollary 2. If a variety V with 0 has a Hilbert ter-m then V is

congruence distr·ibutive.
Corollary 3. Let V be a var-iety with 0 and A E V. If V has a Hilbert
term b then DedA (bA) is pseudocomplemented.
Proof. Let b be a Hilbert term of V, let A E V and D E DedA (bA) .
Denote by
P = {C E DedA (bA); C n D = {OA} }.
Of course P =/- 0 since { oA} E P by Theorem 3. Set D* = V P. Then
D n D* = D n VP = V{D n C; C E P} = {oA}
by Theorem 7 thus D* is the pseudocomplement of D. D
We give a description of D* for D E DedA (bA).

Lemma 2. Let b be a Hilbert term of a variety V, let A= (A, F) E V
and DE DedA (bA).
(a) ifd ED and a E A then bA(a,d) ED
(b) if d 1 , d 2 ED and a E A then bA(d2 , bA(bA(d 1 , a), a)) ED.
Proof. By Theorem 3 we have 8D E Con A for the bA-induced by D
relation e D. Then
since d E D = [OA]e 0 and b is a Hilbert term. Thus bA(a, d) E D

proving (a).
Further, let d1, d2 E D and a E A. Then (d1, OA) E 8D and (d2, OA) E
8D and
thus also
(bA(d2, bA(bA(d1 , a) , a)) , bA(OA , bA(bA(OA , a), a))) E 8v

proving (b). D
Lemma 3. Let b be a Hilbert term of a variety V and A E V , C, D E

DedA (bA). Then C n D = {OA} if and only if bA(bA(d, c), c) = oA for
each c E C and d E D.
c
Proof. Suppose n D = {OA} , the least element of DedA (bA). By
(a) of Lemma 2 we have bA(bA(d, c), c) E C and, by (b) ,
bA(bA(d, c) , c)= bA(oA , bA(bA(d, c) , c)) ED
since oA ED and dE D . Thus bA(bA(d, c), c) E CnD = {OA} proving

bA(bA(d, c) , c) = OA.
Conversely, if x E C n D and bA(bA(d, c), c) = oA for each c E C and
d E D then it is satisfied also for c = x = d and we obtain
X= bA(OA , x) = bA(bA(x , x), x) = OA.
Hence C n D = {OA} . D
The following description of pseudocomplements in DedA (bA) for a

Hilbert term b is an immediate consequence of Lemma 3:
Theorem 9. Let b be a Hilbert term of a variety V and A E V and
DE DedA (bA). Then
References
[1] Chajda 1.: Congruence kernels in weakly regular varieties, South-

east Asian Bull. of Math. , 24(2000) , 15-18.
[2] Diego A.: Surles algebres de Hilbert , Collection de Logique Math.

Ser. A (Ed.Hermann, Paris), 21(1967) , 177- 189.
[3] Fichtner K.: Eine Bemerkung iiber Mannigfaltigkeiten universeller

Algebren mit Idealen, Monatsschr. d. Deutschen Akademie der
Wissenschaften, (Berlin), 12(1970), 21-45.
Author's addresses:
Department of Algebra and Geometry
Palacky University Olomouc
Tomkova 40
779 00 Olomouc, Czech Republic
e-mails: chajda@risc. upol.cz
hals@risc. upol.cz
A Galois Correspondence for Digital
Topology
v
J. 5/apa/
Abstract
We investigate a Galois correspondence between the category of closure
spaces (where closures are considered to be only grounded, extensive and
monotone) and the category of relational systems of a given arity (where
arities are considered to be ordinals) . We show that objects of the obtained
coreflective subcategory of the category of closure spaces are suitable for
applications to digital topology because their connectedness is a certain
type of path connectedness.
AMS Mathematics Subject Classification: 18B30, 18A40, 06A15.
Key words: Closure spaces, Relational systems, Digital topology.
While correspondences between closure operations and binary rela-

tions have been studied by many authors (see e.g. [2], [7], [8]) , the
same can not be said about correspondences between closure opera-
tions and relations of arities greater than two. The latter correspon-
dences will be investigated in the present note which is a continuation
of the author's previous papers [9] and [12] on the topic. We will deal
with the Galois correspondence in the categorical sense discovered in
[12]. This Galois correspondence between closure spaces and a-ary re-
lations (a > 1 an ordinal) will be be discussed with respect to possible
applications to digital topology.
We follow the categorical terminology of [1]. Throughout the paper,
all categories are considered to be constructs, i.e. , concrete categories
1The author acknowledges partial support from Grant Agency of the Czech
Republic, grant no. 201/00/1466.
413
K. Denecke et at. (eds.), Galois Connections and Applications, 413-424.
414 J. Slapal
(over Set) of structured sets and structure-compatible maps. For any

object A of a construct, we denote by IAI the underlying set of A.
As usual, we do not distinguish notationally between morphisms of
constructs and their underlying maps.
Recall that a Galois correspondence (in the covariant sense) between
given constructs A and B is a pair (F, G) of concrete (covariant)
functors F : A ---7 B , G : B ---7 A which forms an adjoint situation
(TJ, c) : F -1 G with concrete (i.e. , identity-carried) natural transfor-
mations 77 : ida ---+ Go F and E : FoG ---7 idA. F and G are then
called the Galois co-adjoint and Galois adjoint, respectively. For our
purposes we will use the following, equivalent definition of a Galois
correspondence between constructs (see [1]):
Definition 1. Let A , B be constructs and G: A---+ B, F: B---+ A

be concrete functors. Then the pair (F, G) is called a Galois correspon-
dence between A and B provided that the following two conditions
are equivalent whenever A E A , B E B are objects and f : IBI ---7 IAI
is a map:
(a) f : IF(B) I ---7 IAI is an A-morphism ,
(!3) f : lEI ---7 IG(A) I is a B-morphism.
By a closure operation u on a set X we mean a map u: exp X ---+ exp

X fulfilling u0 = 0, A~ X=} A~ uA , and A~ B ~X=} uA ~ uB.
Such closure operations were studied by E. Cech in [3] (who called
them topologies). A pair (X, u), where X is a set and u is a closure
operation on X, is called a closure space. Given a pair u , v of closure
operations on a set X , we put u ::::; v if uA ~ vA for each A ~ X .
Clearly, ::::; is a partial order on the set of all closure operations on
X. A closure operation u on a set X is called additive or idempotent
if A, B ~ X =} u(A u B) = uA u uB or A ~ X =} uuA = uA
respectively. A closure operation u on a set X which is both, additive
and idempotent is called a K uratowski closure operation or briefly a
topology and the pair (X , u) is called a topological space. If a topology
u on a set X is also completely additive (i.e. , u uj EJ Aj = ujEJ uAj
for any system {Aj; j E J} of subsets of X), then u is called an
Alexandroff topology and (X , u) is called an Alexandroff topological
space.
Many concepts known for topological spaces can be extended to clo-
A Galois Correspondence for Digital Topology 415
sure spaces in a natural way. Given a closure space (X,u), a subset

A ~ X is called closed if uA = A, and it is called open if X - A is
closed. If (X, u) is a closure space and x E X a point, then a subset
A ~X is said to be a neighbourhood of x if x ¢:. u(X- A). A complete
system of neighbourhoods of x is any system T(x) of neighbourhoods
of x such that for each neighbourhood A of x there exists B E T(x)
such that B ~ A. A closure space (X, u) is said to be a subspace of
a closure space (Y, v) if uA = vAn X for each subset A ~ X. We
will speak briefly about a subspace X of (Y, v). A closure space (X, u)
is said to be connected if 0 and X are the only subsets of X which
are both closed and open. A subset X ~ Y is connected in a closure
space (Y, v) if the subspace X of (Y, v) is connected. A maximal (with
respect to the set inclusion) connected subset of a closure space is
called a component of this space. All the basic properties of connected
sets and components in topological spaces (see e.g. [5]) are preserved
also in closure spaces. Given closure spaces (X,u) and (Y,v), a map
cp : X --+ Y is said to be a continuous map of (X, u) into (Y, v) if
f(uA) ~ vf(A) for each subset A~ X.
Throughout the paper, ordinals are identified with the sets of their
predecessors and cardinals with initial ordinals (so that the Axiom
of Choice is assumed). From now on, a will denote an ordinal with
a>l.
Given a set X, we denote by xa the set of all maps of a into X, i.e.,
the set of all sequences (xi I i < a) of type a consisting of elements of
the set X. If a sequence (xil i < a) is viewed as a set, we mean the
set {xi; i <a}. Any subset R ~ xa is called an a-ary relation on X
and the pair (X, R) is called an a-ary relational system. Given a-ary
relational systems (X, R) and (Y, S), a map cp : X --+ Y is called a
homomorphism of (X, R) into (Y, S) if the implication (xi! i < a) E
R:::} (cp(xi)l i <a) E Sis valid.
We denote by Clothe category of closure spaces with continuous maps
as morphisms, and by Rela the category of a-ary relational systems
with homomorphisms as morphisms.
For any object (X, u) E Clo we put Ga(X, u) = (X, ga(u)) where
ga(u) = {(xili <a) E xa; Xj E u{xi; i < j} for each j, 0 < j <a},
and for any object (X, R) E Rela we put Fa(X, R) = (X, fa(R)) where
fa(R)A = AU {x E X; there exist (xili < a) E Rand io, 0 < io <
416 J. ~lapal
a, such that x = Xio and xi E A for all i < i 0 }.

Then Fa : Rela ---+ Clo and Ga : Clo ---+ Rela are concrete functors
and we have:
Theorem 1. (Fa, Ga) is a Galois correspondence between Clo and

Rela.
Proof. Let (X, u) E Clo and (Y, R) E Rela be objects. Let f :
(Y, fa(R)) ---+ (X, u) be a continuous map and let (Yil i < a) E R.
Then Yio E fa(R){yi; i < i 0 } for each 'io , 0 < i 0 < a. Consequent ly,
f(:lfio) E u{f(Yi) ; i < io} for each io, 0 < io < o:. This yields (f(yi)I i <
a) E 9a(u). Hence, f: (Y, R) ---+ (X, 9a(u)) is a homomorphism.
Conversely, let f : (Y, R) ---+ (X, 9a(u)) be a homomorphism and let
A s:;: Y and y E fa(R)A. If y E A , then f(y) E f(A) s:;: uf(A). Let
y ¢: A. Then there exist (Yil i < o:) E R and i 0 , 0 < i 0 < o:, such
that y = Yio and Yi E A for all i < io. Thus (f(yi)l i < a) E 9a(u)
and, therefore, f(y) E u{f(Yi); i < i 0 } s:;: uf(A). We have shown that
f: (Y, f a(R)) ---+ (X, u) is a continuous map. D
Remark 1. Of the result ing properties of the Galois correspondence

from Theorem 1 let us mention the following ones:
(1) Fa o Ga 'S i dc lo and 'idRela 'S Gao Fa where 'S is induced by the
corresponding fibre preorder (i.e., given a set X, f a(9a(u) ) < 1L for
each closure operation u on X, and R s:;: 9a (fa (R)) for each o:-ary
relation R on X) .
(2) Fa(Rela) is coreflective in Clo.
(3) Ga(Clo) is reflective in Rela .
(4) If (X, u) E Clo, then (X, u ) E Fa(Rela) if and only if (X, u) =
Fa o Ga (X, u).
(5) If (X , R) E Rela , t hen (X, R) E Ga (Clo) if and only if (X, R) =
Gao Fa(X, R).
(6) The restrictions of Ga and Fa to Fa(Rela) and Ga (Clo) respec-
tively are concrete isomorphisms between Fa (Rela) and G a (C lo) that
are inverse to each other.
(8) Ga preserves initial sources (because Clo is a topological category).
(7) G 2 ( Clo) is the full subcategory of Rel2 whose objects are precisely
the reflexive binary relational systems, and F2 (Rel2 ) is t he full sub-
category of Clo whose objects are precisely t he closure spaces with
completely additive closure operations. (F2 , G 2 ) is the obvious Galois

correspondence between Clo and Rel2 given by:
G2(X,u) = (X,Ru) where xRuy {:} y E u{x},
F2(X, R) =(X, uR) where uRA =AU {y EX; 3x E A: xRy }.
Ru is a preorder provided that u is a topology, and R;;, 1 is then called
the specialization preorder of u. Conversely, if R is a preorder, then
uR is an Alexandroff topology (and vice versa).
Of course, closure operations f a(R) need not be additive. But, on the

other hand, we have:
Proposition 1. Let R be an a-ary relation on a set X. Then the

union of any system of closed subsets of (X, fa(R)) is a closed subset
of (X, fa(R)).
Proof. Let {Aj; j E J} be a system of closed subsets of (X, fa(R))
and let X E !a(R) uj EJ Aj. Then there are (xi I i < a) E R and
io, 0 < i 0 < a, such
that X = Xio and Xi E uj EJ Aj for all i < io. Especially, we have
.To E uj EJ Aj and so there exists Jo E J such that Xo E Ajo· Sup-
pose that {xi; i < a} is not a subset of Aj0 • Then there exists the
smallest ordinal i 1 < a such t hat Xi 1 tj_ Aj0 • Consequently, 0 < i 1
and Xi E Aj0 for all i < i1. Thus, we have Xi 1 E fa(R)Aj 0 =
Aj0 , which is a contradiction. Therefore, {xi; i < a} ~ Ajo and,
hence, X E Ajo = !a(R)Ajo ~ ujEJ !a(R)Aj· We have shown that
fa(R) U jEJ Ai ~ U j EJ fa(R)Aj. As the converse inclusion is obvious,
the proof is complete. D
In [12], the following result is proved:

Proposition 2. For any ordinal a > 1 there holds:
(I) Ga(Clo) is the full subcategory of Rela whose objects are pre-
cisely the a-ary relational systems (X, R) satisfying:
1o (xili < a) E R whenever there exists x E X with Xi = x

for all i < a,
2° if (xili < a) E xa has the property that for any io, 0 <
io < a, there exist (YjiJ < a) E R and Jo , 0 < )o < a,
such that xio = Y]o and {yj; j < Jo} ~ { xi ; i < io} , t hen
(xili < a) E R.
418 J. ~lapal
(II) Fa(Rela) is the full subcategory of Clo whose objects are pre-
cisely the closure spaces (X, u) having the following property:
(#) if A ~ X and x E uA, then there exist i 0 , 0 < i 0 < a,

and (xi li < i 0 ) E Aio such that Xj E u{ xi; i < j} for each
j, 0 < j < i 0 , and x E u{xi;i < i 0 }.
Remark 2. (a) Relational systems fulfilling the condition 2° from

Proposition 1 are studied in [10] and [11]. Note that for binary re-
lational systems 2° is trivially satisfied. A ternary relational system
(X, R) fulfills 2° if and only if each of the following six conditions
implies (x , y, z) E R:
(1) (x, y, t) E R, (x, z, u) E R,
(2) (x, y, t) E R, (y, y, z) E R,
(3) (x, y, t) E R, (y, z, u) E R,
(4) (x, x, y) E R, (x, z, u) E R,
(5) (x, y, t) E R, (y, x, z) E R,
(6) (x, x, y) E R, (y , z, u) E R.
(b) From Proposition 2 it immediately follows that any object (X, u) E
Fa(Rela), a> 1 an ordinal, satisfies the condition uA = U{ uB; B ~
A, card B < a} whenever A ~ X. Closure spaces (X, u) satisfying
this condition for a= 2 are called quasi-discrete in [4] . It is evident
that they are just the closure spaces with completely additive closures.
Thus, a topological space (X, u) fulfills the previous condition for a=
2 if and only if (X, u) is an Alexandroff topological space.
Of course, given a closure space (X, u), the closure operation u is

uniquely determined by systems of all neighbourhoods of points of
(X, u): for any subset Y ~X and any point x E X we have x E uY
if and only if Y n A f. 0 for every A E T(x) where T(x) denotes the
system of all neighbourhoods of x. But the same is valid also when
T(x) denotes a complete system of neighbourhoods of x. Therefore,
for any point x E X, it is useful to know a complete system of neigh-
bourhoods of x, especially one which is smaller than the system of all
neighbourhoods of x. It is obvious that, given an a-ary relation R on
a set X and a point x E X, a subset A ~ X is a neighbourhood of x
in (X, fa(R)) if and only if x E A and, whenever (xi I i < a) E R is
a sequence with x = Xio for some i 0 , 0 < i 0 < a, there exists i 1 < i 0
such that Xi 1 E A.
Complete systems of neighbourhoods of points of (X, fa(R)) can be
determined by using the following statement (and Remark 2(b)):
Proposition 3. Let (X, u) be a closure space such that uA = U{ uB;

B ~ A, cardB < a} whenever A~ X. For any x E X put T(x) =
{A ~ X for each subset Y ~ X fulfilling cardY < a and x E uY
there exists a non-empty subset Ay ~ Y such that A= U{Ayl Y ~
X, card Y < a, x E uY}} . Then T(x) is a complete syst em of
neighbourhoods of x in (X, u).
Proof. Clearly, a subset A ~ X is a neighbourhood of a point x
in (X, u) if and only if from x E uY it follows that Y n A "# 0 for
each subset Y ~ X. But the same is valid if we consider only the
subsets Y ~ X with cardY < a. Therefore, T(x) is a system of
neighbourhoods of x in (X, u). Let B ~ X be a neighbourhood of x
in (X, u). For any subset Y ~ X with cardY < a and x E uY put
Ay = Y n Band Z = U{Ay ; Y ~ X,cardY <a, x E uY}. Then
clearly Z ~ B and Z E T( x ). D
From Proposition 3 it follows that in any closure space (X, u) with

the property that uA = UxEA u{ x} whenever A ~ X, any point x E
X has a complete system of neighbourhoods consisting of just one
neighbourhood A (which must then be a least one with respect to the
set inclusion) given by A = {y E X; x E u{y }}. Especially, this
is valid for any object (X, u) E F2 (Rel 2 ) (for Alexandroff topological
spaces this fact is well known -see [6]).
Denote by Clau the full subcategory of Cla given by the closure spaces
whose closure is idempotent. Further, denote by TapA the full subcat-
egory of Clau whose objects are precisely the Alexandroff topological
spaces. TapA is a full subcategory of F2 (Rel2 ) and in [12] the following
statement is proved:
Proposition 4. Fa(Rela) n Clau =TapA.
Examples. 1. Let X = {a, b, c} and define u : exp X ----+ exp X as fol-

lows: u0 = 0, u{a} = {a , b}, u{b} = b, u{c} = c, u{a, b} = u{a,c} =
uX = X, u{b,c} = {b,c}. Then (X,u) E F3 (Rel 3 ) (but (X,u) ~
420 J. Slapal
F 2 (Rel 2 )). Clearly, the ternary relation R on X with F 3 (X, R) =

(X, u), i.e., with h(R) = u, is given by R = {(a , a, a), (b , b, b), (c, c, c),
(a,b,c), (a, a, b), (a,b,a), (a , b,b)}.
2. Let w denote the least infinite ordinal and let (w + 1, u) be the
closure space given by u(/) = 0, uA = w whenever A~ wand 0 <card
A < w, uA = w + 1 otherwise. Then (w + 1, u) E Fw+l(Relw+l) (but
(w + 1, u) tf. Fw(Relw)). Note that the space (w + 1, u) is additive, i.e. ,
u(A U B) = uA U uB whenever A, B ~ w + 1.
Digital topology is a branch of theoretical computer science which has

arisen for the study of geometric and topological properties of digital
images. The basic problem of digital topology is to provide the digital
plane, i.e., the set Z x Z (where Z denotes the set of integers), with a
topological structure suitable for the study of topological properties of
digital images (i.e., subsets of Z x Z) , especially their connectedness.
Among various topologies on Z x Z there is known only one which is
satisfactory for the needs of digital topology, the so-called Khalimsky
topology. We will show that there are other topological structures suit-
able for digital topology which can be found among closure operations
on Z x Z.
The Khalimsky topology coincides with the closure operation h(R)
where R is the binary relation on Z x Z given as follows:
R = {((x 1 ,x 2 ), (y 1 ,y2 )) E (ZxZ)x(ZxZ) ; whenever i E {1 , 2}, Xi=
Yi or xi is both, odd and fulfilling Ixi - Yi I = 1}.
It can easily be seen that R is a partial order on Z x Z . The relation
S = (R U R - 1) - ~ (where R- 1 denotes the inverse relation toR and
~ denotes the equality relation on Z x Z) is then called the adjacency
relation. Given points z 1 , z2 E Z x Z , z 1 is said to be adjacent to z2
provided that (z 1 , z2 ) E S . A subset A ~ Z x Z is connected in the
Khalimsky topology if and only if for any two points x, y E A there
exists a finite sequence z0, z1 , ... , zk of points of A such that x = z0, y =
Zk and (zj_ 1 , Zj) E S for all j = 1, 2, ... , k. Thus, in the Khalimsky
topology, connectedness is a certain type of path connectedness. This
is one of the facts that make the Khalimsky topology suitable for
solving problems of digital topology. We will show that the same is
valid also for the closure operations !a(R) provided that a is finite
(and R is an a-ary relation on Z x Z).
Lemma 1. Let R be an a-ary relation on a set X and (xi I i < a) E R

an element. Then the set {xi; i 1 and suppose that
{xi; i < i 1 } is connected for each i 1 , 0 < i 1 < i 0 . As {xi; i < i 1 } ~
{xi; i::::; i 1 } ~ !a(R){xi; i < i 1 }, the set {xi; i::::; i 1 } is connected for
each i1, 0 < i1 < io. Since no<il<io{xi; i::::; i1} #- 0, Uo<il <io {xi; i::::;
i 1 } is connected. But Uo 1.
Definition 2. Let R be an n-ary relation on a set X and let p be a

natural number with 1 < p < n. A sequence (Yi I i < p) of points of
X is called a connected element in (X, fn(R)) if there is a sequence
(xil i < n) E R such that Yi =xi for all i < p or Yi = Xp-1-i for all
i < p.
Thus, by Lemma 1, each connected element is a connected set.
Definition 3. Let R be an n-ary relation on a set X . A finite

nonempty sequence C = (xi I i < m) of points of X is called a path in
(X, fn(R)) if there is a finite increasing sequence (jkl k < p) of natural
numbers with j 0 = 0 and jp- 1 = m- 1 such that jk - jk_1 < n and
(xi I jk- 1 < j < jk) is a connected element in (X, f n(R)) for each k,
0 < k < p. Given points y, z E X, the path C = (xi I i < m) is said to
connect y and z provided that y = x 0 and z = Xm_ 1 .
Thus, one-member sequences are also considered to be paths. Clearly,

each path is a connected set and each connected element is a path. If
(Xi I i < m) is a path, then also its inverse, i.e. , the sequence (Yi I i < m)
where Yi = X m - 1 - i for all i < m, is a path. Further, if (xil i < m),
(Yil i < p) are paths such that Xm - 1 = y0 , then also their union, i.e.,
the sequence (zi l i < m + p - 1) where zi = xi for all i < n and
Zi = Yi- m+l for all i with m < i < p, is a path.
Theorem 2. Let R be an n-ary relation on a set X and A ~ X be

a subset. Then A is connected in (X, fn(R) ) if and only if any two
points of A can be joined by a path in (X, fn(R)) contained in A.
422 J. ~lapal
Proof. If A = 0, then the statement is trivial. Let A =J. 0. In

(X, fn(R)), if any two points of A can be connected by a path, then
A is clearly connected . Conversely, let A be
connected and suppose that there are points x, y E A which can not
be connected by a path contained in A. Let B be the set of all points
of A which can be connected with x by a path contained in A. Let
z E fn(R)B n A be a point and assume that z rJ. B. Then there
are (xil i < n) E R and i 0 , 0 < i 0 < n , such that z = Xio and
{xi; i < i 0 } ~ B. Thus, x and x 0 can be connected by a path
contained in A, and also x 0 and z can be connected by a path - the
connected element (xi I i < i 0 ) - contained in A. It follows that x and
z can be connected by a path contained in A , which is a contradiction.
Therefore z E B, i.e. , fn(R)B n A = B . Consequently, B is closed
in the subspace A of (X, fn(R)) . Further, let z E fn(R)(A- B) n A
be a point and assume that z E B. Then z rJ. A - B , thus there
are (xil i < n) E R and i 0 , 0 < i 0 < n , such that z = Xio and
{xi; i < i 0 } ~ A- B. As we have assumed that z E B , x can
be connected with z by a path contained in A . Further, z can be
connected with x 0 by a path- the connected element (xio- il i < i 0 ) -
contained in A. Consequently, x and x 0 can be connected by a path
contained in A, which is a contradiction with x 0 rJ. B. Thus z rJ. B ,
i.e., fn(R)(A- B) n A = A- B . It follows that A- B is closed in
the subspace A of (X, fn(R)). Hence A is the union of the nonempty
disjoint sets B and A- B closed in the subspace A of (X, fn(R)) .
But this is a contradiction because A is connected. Therefore any two
points of A can be connected by a path contained in A . D
We have shown that, for any n-ary relation Ron Z x Z , connectedness

in the closure operation fn(R) is a certain type of path connectedness.
Therefore, the closure operations fn(R) could be useful for applications
in digital topology.
Problem. When solving problems of digital topology it is often de-

sirable that the used topological structure on Z x Z should fulfill
an analogy of the Jordan curve theorem known for the real plane.
(Recall that the Jordan curve theorem states that any simple closed
curve separates the real plane into precisely two components.) For
the Khalimsky topology, the following analogy of the Jordan curve
theorem was proved in [6) : If Cis a finite connected subset of Z x Z
with at least four points such that, for each point z E C, there are
exactly two points in C adjacent to z, then C separates Z x Z into
precisely two components. Now, there arises a problem to formulate
and prove an analogy of the Jordan curve theorem for the closure op-
erations fn(R). This problem seems to be quite difficult in its general
setting but, with respect to applications to digital topology, it would
be sufficient to solve it by finding a convenient n-ary relation R on
Z x Z for each natural number n > 1. For n = 2 the problem has
been solved because, in consequence of [6], such a convenient binary
relation R on Z x Z is that one mentioned above for which h(R) is
the Khalimsky topology.
References
[1] J. Adamek, H. Herrlich and G.E. Strecker, Abstract and concrete

categories, Wiley-lntersci. Publ., New York, 1990.
[2] T.J. Ahlborn and T.N. Bhargava, On topological spaces associated

with digraphs, Acta Math. Sci. Hung. 19 (1968), 47-52.
[3] E. Cech, Topologi cal spaces, in: Topological Papers of Eduard

Cech, Academia, Prague, 1968, 436-472.
[4] E. Cech, Topological Spaces, revised by Z. Frolfk and M. Katetov,

Academia, Prague, 1966.
[5] R. Engelking, General Topology, Panstwowe Wydawnictwo

Naukowe, Warszawa, 1977.
[6] E.D. Khalimsky, R. Kopperman and P.R. Meyer, Computer

graphics and connected topologies on finite ordered sets, Topol-
ogy Appl. 36 (1990), 1-17.
[7] F. Lorrain, Notes on topological spaces with minimum neighbor-

hoods, Amer. Math. Monthly 76 (1969), 616-627.
[8] J. Slapal, On closure operations induced by binary relations, Rev.

Roum. Math. Pures Appl. 33 (1988), 623-630.
[9] J. Slapal, Relations and topologies, Czech. Math. J. 43 (1993) ,

141-150.
424 J. Slapal
[10] J . Slapal, On strong regularity of relations, Math. Bohemica 119

(1994) , 151-155.
[11] J. Slapal, On categories of strongly regular relational systems,

Tatra Mount. Math. Publ. 5 (1995) , 101-105.
[12] J. Slapal, A Galois correspondence between closure spaces and

relational systems, Quaest. Math. 21 , (1998) , 187- 193.
Author's address:
Technical University of Brno
616 69 Brno, Czech Republic
e-mail: slapal@um.fme.vutbr.cz
Galois Connections in Category
Theory, Topology and Logic
W. Gahler
Abstract
The notion of a Galois connection is important in different branches of
mathematics. It even is used for defining basic notions in several theories.
In this paper the role of Galois connections is demonstrated in reference to
known results and moreover in presenting new ones.
AMS Mathematics Subject Classification: 06D35,18B35,03E12.
Key words: Galois connection, Non-classical logic, MV-algebra, Partially
ordered monad, Fuzzy filter.
1. Different kinds of Galois connections

Standard kind. G. Birkhoff ([1]) introduced a notion of Galois con-
nection, related to power sets. Later on 0. Ore ([15]) and J. Schmidt
([18]) studied a generalization of this notion, which can be considered
as the standard kind of a Galois connection. Its "covariant form" is
defined as follows: A pair of isotone mappings f : (X, :S) -----t (Y, :S)
and g : (Y, :S) -----t (X, :S) with (X, :S) and (Y, :S) partial ordered sets, is
a Galois connection provided the equivalence f( x ) :S y -¢:::::::? x :S g(y)
holds for all x E X and y E Y . (!,g) is a Galois connection if and
only if for each y E Y the supremum U{x EX I j(x) :S y} exists and
coincides with g(y). Therefore, g will be called the sup-inverse of f.
Categorical notions of Galois connections. The following kind
of Galois connections, pointed out by H. Herrlich and M. Husek ([10]),
is more general. In this case instead of isotone mappings f : (X, :S
) -----t (Y, :S) and g : (Y, :S) -----t (X, :S) two concrete functors <p : A -----t B
and 'ljJ : B -----t A are given, where A and B are concrete categories over
425
426 W. Gahler
a category X. The pair (r.p, 1/J ) is called a concrete Galois connection

provided the following equivalence holds
r.pA ~ B E MorB ~ A ~ 'lj;B E MorA
for A E ObA, B E ObB and f E MorX. We obtain the classical

kind if we take (X,~) and (Y, ~) as the concrete categories A and
B, respectively, with X and Y the sets of objects and the partial
orderings of X and Y as the related sets of morphisms and interpreting
X trivially as a category having only one object and only the identity
morphism.
A further kind of Galois connection, also discussed in [10], is that
of an adjoint situation, which consists of a pair (<p, 1/J ) of covariant
functors <p : A ---+ B and 1/J : B ---+ A and two natural transformations
'T] : idA ---+ 1/J o <p and c : <p o 1/J ---+ id 8 (the unit and counit of the adjoint
situation) such that idcp = Ecp o <p'TJ and id'l/1 = 'lj;E: o 'Tl'l/1 ·
2. Galois connections and partially ordered monads

There are interesting applications of the standard kind of Galois con-
nections in general topology. Two examples related to partially or-
dered monads will be presented in this section.
Partially ordered monads. In the general theory of convergence
structures the notion of a partially ordered monad is basic. It is defined
by means of the notion of an almost complete semilattice. By an
almost complete semilattice we mean a poset (X , ~) in which all non-
empty suprema exist. Infima of subsets, even of pairs {x, y } , need not
exist.
Let acSLAT denote the category of almost complete semilattices, where
the morphisms are the mappings between almost complete semilattices
which preserve non-empty suprema. By a partially ordered monad
(over SET) (cf. [3, 6]) we mean a quadruple <P = (r.p, ~' ry , p,) which
fulfills the following conditions:
 consists of two natural transformations 'T] = ('TJx )x EObSET and p, =
(p,x )x EObSET of mappings 'T]x : X ---+ r.pX and f.Lx : r.pr.pX ---+ r.pX
respectively, such that the triple (r.p, ry, p,) is a monad over SET. Finally
we assume:
Galois Connections in Category Theory, Topology and Logic 427
(MO) rpX is empty in case X is empty.
(Ml) For any set X and each pair of different elements x and y of X,
the infimum of rJx(x) and TJx(Y) does not exist.
(M2) For all mappings f, g: Y ------t rpX from f ~ g it follows J-Lx o rpf ~
J-Lx o rpg, where ~ is defined argumentwise with respect to the
partial ordering of rpX.
(M3) For each set X, J-Lx: (rprpX, ~) ------t (rpX , ~)preserves non-empty
suprema.
The partial orderings ~ of the sets rpX are considered as finer rela-
tions. For each set X, the elements of rpX are called rp- objects on X,
and the minimal elements of rpX also ultra objects.
For each non-empty subset A of rpX the supremum V M exists.
MEA
Hence, whenever for a subset A of rpX a lower bound in rpX exists,
the infimum A M of this subset exists.
MEA
For each non-empty set X we denote the supremum V TJx(x) by
xEX
TJx[X]. A rp-object M on X for which M ~ TJx[X] holds, is called
stratified.
A rp-object Mona set X for which M < TJx(x) holds for some x EX,
is called a microobject at x. Because of condition (Ml), x is uniquely
associated to M. Of course, microobjects are special stratified <p-
objects.
Proposition 1 ([6]) If ther·e is a microobject at an element x of X,

then there is a microobject at any element y of X.
Microobjects are in some sense properly finer than points. They may
exist or may not. If there are no microobjects, then condition (Ml)
can be formulated as follows:
(Ml') For each set X, TJx : X ------t rpX is an injection and all values
TJx(x) are ultra objects on X.
The classical case of a partially ordered monad is that of the partially

order-ed filter- monad (F, ~' 7], J-L), which is given as follows: F is the
428 W. Gahler
filter functor-, which assigns to each set X the set FX of all (proper)
filters on X. ::; indicates that the sets FX are equipped with the finer
relations of filters, that is , the inversion of the inclusion. ry and p, are
natural transformations consisting of all mappings 'r/x : X --+ FX and
fJ,x: FFX--+ FX respectively, where for each x EX, 'r!x (x) = {M ~
X I X E M} and for each filter .c on FX , Mx(.C) = un
AE.CMEA
M. In
this classical case microobjects do not exist.
In the general case of a partially ordered monad <]) = (<p, ::; , ry , p,) there
are the following examples of Galois connections.
Example 1. Let f : X --+ <pY and e : Y --+X be mappings such that
foe= ryy. Then the sup-inverse of P,x o<p f : (<pX ,::; )--+ (<pY ,::; )
exists and is denoted by
<p;;: (<pY, ::;) --+ (<pX, ::;).
Example 2. For each surjection f : X --+ Y , the sup-inverse of the
acSLAT-morphism <pf : (<pX , ::;) --+ (<pY , ::;) exists, denoted by <p- f :
(<pY, ::;) --+ (<pX, ::;).
Partially ordered submonads. Let<]) = (<p, ::; , ry , p,) be a partially
ordered monad. By a partially ordered submonad of <]) we mean a
partially ordered monad W = (<p' , ::; , ry' , p,') such that
(1) (<p', ::;) :SET--+ acSLAT is a subfunctor of (<p, ::;)

(in particular, for each set X and each non-empty subset A of
<p'X the suprema of A with respect to (<p'X , ::; ) and to (<pX, ::;)
coincide).
(2) (<p', ry' , p,') is a submonad of (<p , ry, p,).
An example always exists: the partially ordered submonad of <]) of

all stratified <p-objects. For the related subfunctor <p1 of <p we have
<p'X = {ME <pX I M::; 'r/x[X] }.
3. Some topological notions

Let <]) = (<p, ::;, 7], p,) be a partially ordered monad.
A <P-converyence str-uctur·e on a set X is a subset T of <pX x X such
that, writing M--+ x instead of (M , x) E T , we have
(Cl) 'r!x(x ) --+ x for all .T EX,

(C2) M ---+ x and N::; M imply N---+ x, and
(C3) M---+ x implies M V TJx(x) ---+ x.
A <P-convergence structure Tis called a <P-limit structure if instead of

(C3) the following condition is fulfilled:
(C3') M ---+ x and N---+ x imply M V N---+ x.
In the filter case the <P-convergence structures are the convergence

structures in sense of D. C. Kent, introduced first in [13], and the
<P-limit structures are the limit structure in the usual sense, proposed
first by H.-J. Kowalsky in [14].
A mapping p : X ---+ rpX is called a <P-pretopology on X, provided that
TJx (x) ::; p( x) holds for all x E X. p can be interpreted as the special
<P-limit structure T = { (M, x) I M ::; p(x) } . For each <P-pretopology
p : X ---+ rpX, nb : rpX ---+ rpX defined by
nb = J1x o rpp,
is called the neighbourhood operator of p.

A <P-pretopology p is called a <P-topology provided that nb o p = p or,
equivalently, nb o nb = nb holds, that is, nb is idempotent.
Let T be a <P-convergence structure on a set X and let t 1 and t 2 be
the first and second projection ofT, respectively.
rpX X rpX
Fig. 1. The projections t1 and t2 ofT

and related cp-morphisms
For the mapping e : X ---+ T, defined by e(x) = (TJx(x), x) for all

x E X, we have h o e = TJx. Because of Example 1 of a Galois
connection the sup-inverse rp~t 1 of J1xorpt 1 exists. Since t 2 is surjective,
because of Example 2 of a Galois connection moreover the sup-inverse
430 W. Gahler
cp- t 2 of cpt 2 exists. We introduce the closure operator cl : cpX ---+ cpX
ofT ([3]) by means of the sup-inverse cp;t 1 defining
cl is a hull operator, that is, M ::::; elM holds for all M E cpX . If
M = elM, then M is called closed. Since cl is a hull operator, we
have that the infimum of any set of of closed cp-objects on X is closed,
as far as this infimum exists.
As neighbourhood operator nb : cpX ---+ cpX of the -convergence
structure T we mean the neighbourhood operator of the associated -
pretopology p ofT, which is defined by p(x) = V M for all x E X.
M--tx
nb is also a hull operator and can be introduced by means of the
sup-inverse cp- t 2 ([6]) by
This shows that the notions of a closure operator and of a neighbour-

hood operator are in some sense dual. More exactly, here we can
speak of an associated pair of Galois connections. In general topol-
ogy both notions of closure operator and of neighbourhood operator
are used for different reasons. This differs from the classical case of
a topological space and even of a pretopological space in which each
of the notions of a closure operator and of a neighbourhood opera-
tor determines uniquely the other one. This simple situation in the
classical case depends on the fact that for the interior and the clo-
sure of subsets of a space we have X\ intM = cl(X \ M), where
intM = { x E X I M E p( x)}. Because of this reason, classical pre-
topological spaces are sometimes called closure spaces. That in general
such a one-to-one correspondence does not exist, will be demonstrated
at the end of this paper by examples of different -pretopologies which
have one and the same closure operator.
In the following let T be a -convergence structure on a set X and let
nb be its neighbourhood operator. A cp-object M on X is called open
provided that M = nbM.
Proposition 2 ([6]) The set 0 of all open cp-objects on X fulfills the

following condition
(0) 0 is closed with respect to all non-empty suprema and all infima,
as far as these infima exist.
Each if>-topology p can be characterized by the set 0 of all open <p-

objects on X by taking
p(x) = M (1)
MEO,ryx(x):SM
for all x EX.
Proposition 3 ([6]) Let 0 be a subset of <pX which fulfills the condi-

tion (0) and let p be the mapping defined by (1). Then the following
are equivalent:
(a) p is a if>-topology with 0 the set of all open <p-objects on X.
(b) For each M E <pX we have nbM = 1\ N.

NEO,M:SN
4. Basic structures in non-classical logics

There is a close relation between the general topology and the different
types of non-classical first order predicate logics, given by means of
the associated basic structures of these logics. Concerning this logics
we refer to [11].
Frame case. A first kind of basic structures is given by frames . A
frame is a non-degenerate complete lattice (L , ~) which fulfills the
complete distributivity law
i EI i EJ
Non-degenerate means that the minimal element 0 and the maximal

element 1 differ.
In a frame (L, ~), by means of the implications
a ----+ y = max { x E L I a 1\ x ~ y }
432 W . Gahler
for each a E La Galois connection is given, consisting of the mappings

fa : L ---+ L and ga : L---+ L , where fa(x) = a 1\ x and ga(Y) = a---+ y.
Clearly, fa(x) :S y +-+ x :S ga(Y) holds for all x , y E L.
Quantale case. A structure more general than that of a frame , is
that of a commutative quantale (L, :::; , *) (cf. Rosenthal's paper [17]) .
It is defined as a non-degenerate complete lattice which is equipped
with a commutative and associative binary operation * such that the
complete distributivity law
iEI iEI
holds. We always assume that the maximal element 1 of the underlying

lattice is the unit element with respect to *·
Analogously as in case of a frame , in a commutative quantale impli-
cations a---+ bare defined and Galois connections are introduced. The
only difference is that 1\ is replaced by *, that is, here we have
a ---+ y = max { x E L I a * x :S y }
and for the mappings fa : L ---+ L and ga : L ---+ L, given by fa (x) = a*X
and ga(Y) =a---+ y, we have fa(x) :S y +-+ x :S ga(Y) for all x , y E L.
Clearly, a frame is a commutative quantale with 1\ : L 2 ---+ L the
related binary operation.
Further examples of commutative quantales are given by means of the
continuous t-norms * : £ 2 ---+ L. They are defined as follows:
(T1) L is the unit interval [0, 1] equipped with the usual ordering.
(T2) * is a continuous commutative and associative binary operation

with unit element 1.
(T3) * is order preserving, that is, a :::; band c :::; dimply a* c :::; b *d.
Because of the continuity of*, for at-norm the complete distributivity

law with respect to * holds.
We note some special examples:
Godel t-norm a* b =min{ a, b}
Product t-nor·m a * b = a · b (product of reals)
Lukasiewicz t-norm a* b = max{O , a+ b- 1}
Case of Girard monoids. A commutative quantale which fulfills the

condition of double negation -.(-·a) =a, is a complete Girard monoid.
Note that the negation of a is given by •a = a --t 0.
For the Godel t-norm and the product t-norm we have
if a> 0
if a= 0.
Hence, for these t-norms the condition of double negation is not ful-
filled and they therefore do not define Girard monoids.
Case of MY-algebras. By a complete MV-algebra we mean a com-
plete Girard monoid (L , ::; , *) which is divisible, that is, for all a, b E L
from a ::; b it follows a = b * c for some c E L.
Examples of complete MY-algebras:
(1) The Lukasiewicz t-norm *defines a complete MY-algebra ([0, 1],::;
'* ).
(2) All complete Boolean algebras are complete MY-algebras. They
can be characterized as those complete MY-algebras (L, ::;, *)for which
* is idempotent, that is, a * a = a for all a E L. In this case, then *
equals 1\.
Related logics. In the following there are listed basic structures
together with related non-classical logics:
Frames intuitionistic logic
Commutative quantales monoidallogic
Complete Girard-monoids - linear logic
Complete MY-algebras Lukasiewicz logic
The following is well-known in first order logic:
Proposition 4 ( [11]) A formula is valid with respect to all non-

degenerate complete MV-algebras if and only if it is valid with respect
to ([0, 1], ::;, *) with* th e Lukasiewicz t-norm.
Proposition 5 ([16]) A formula is valid with respect to all non-

degenerate complete Boolean algebras if and only if it is valid with
respect to the Boolean algebra consisting only of 0 and 1.
434 W. Gahler
5. Fuzzy filters in the frame case

There are different kinds of fuzzy filters. The following deals with one
of these kinds.
If (L, :S::) is a complete lattice, then for each set X the mappings f E
Lx are called fuzzy subsets of X. For each a E L let a denote the
constant fuzzy subset of X with value a.
Let L = (L, :S::) be a non-degenerate frame. A fuzzy filter on X in the
frame case, also called an L-filter, is a mapping M : Lx --* L such
that
(Fl) M(O) = 0 and M(I) = 1,
(F2) M(f 1\ g)= M(f) 1\ M(g) for all f,g E Lx.
In the following, a definition of the related partially ordered fuzzy

filter-monad (FL, :::;;, ry, J-t) , also called the partially ordered L-filter mo-
nad, is given.
For each set X, FLX consists of all L-filters on X. The partial ordering
on each set FLX is given by
M :S:: N ~ M(f) :::=: N(f) for all f E Lx.
For each mapping f : X --* Y, the mapping FLf : FLX --* FLY
assigns to each M E FLX the fuzzy filter FLf(M) on Y given by
FLf(M)(g) = M(g of) for all g E Lx .
For each set X, each x E X and f E Lx let 'TJx(x )(f) = f(x) , and
for each £ E FLFLX and f E Lx let J-tx(.C)(f) = .C(ef) , where ef :
FLX--* L is the mapping M r-+ M(f).
We distinguish some types of fuzzy filters. An L-filter M on a set X
is called
bounded if M(a) :::;; a holds for all a E L and
homogeneous if M(a) =a holds for all a E L.
As a specialization of the general property of being stratified, here we
have that an L-filter on a set X is
stratified if and only if M(a) :::=: a holds for all a E L.
Each fuzzy filter coarser than a bounded fuzzy filter is also bounded.
Each fuzzy filter finer than a stratified fuzzy filter is also stratified.
Proposition 6 There are three partially ordered submonads of (FL , ::;

, TJ, f.L), defined by all bounded fuzzy filters , all homogeneous fuzzy filters
and all stratified fuzzy filters, respectively.
The following is related to suprema and infima of fuzzy filters .
Proposition 7 Let X be a set. For the supremum of a non-empty

subset A of FLX we have
(v
M EA
M)(f) = A M(f)
M EA
(2)
for all f E Lx . If the infimum of a subset A of FLX exists, then
(A M)(f) =
MEA h 1\ ···1\ fnS!
Ml , ... ,MnEA
for all f E LX.
6. Principal fuzzy filters in the frame case

Let X be a non-empty set. For each mapping f E Lx and each a E L
with a ::; Uf by
iff g ' t
[/, a[(g) = { ; if f ::; g =!= I ,
if g =I
for all g E Lx, a bounded fuzzy filter [f, a] is defined, called a principal
fuzzy filter on X (first defined in [4]).
Proposition 8 For each bounded fuzzy filter M in the frame case

the infimum M 1\ TJx[X] exists, called the homogenization of M.
Since for each f E Lx the principal fuzzy filter [f, Uf] is bounded,
its homogenization [ f , Uf) 1\ TJx[X] exists. For any f E Lx and all
g E Lx we have [f](g) = V (Uf 1\a) vng . Clearly, [OJ= TJx[X].
f /\a5:c g
436 W. Gahler
In the following some representations of fuzzy filters by means of prin-

cipal fuzzy filters are given.
Proposition 9 ([8]) Let X be a set of at least two elements. Then

faT each fuzzy filter M on X we have:
M is the infimum of all pr·incipal fuzzy filters [f, a] coaT"ser than M ,
that is, of all principal fuzzy filters [f, a] for which a ::; M (f) 1\ U.f
holds:
M 1\
a ~M(f) 1\ Uf
[f, a].
Moreover then faT each stmtified fuzzy filter M on X we have:

M is the infimum of all fuzzy filters [.f] which are coarser than M ,
that is, of all fuzzy filters [.f] for which U.f ::; M (f) holds:
M= 1\ [f].
Uf~M(f)
Proposition 10 ([8]) Let X be a set of only one elem ent, then we

have:
Each bounded fuzzy filter on X can be represented analogously as in
Pmposition 9 by means of pTincipal fuzzy filters.
Non of the stmtified, non-homogeneous fuzzy filter·s on X can be rep-
resented in this way.
For any non-empty set X we have [ 0] 1\ [.f, nf ].

/ ELX
7. Fuzzy filters in the quantale case

In the following let L = ( L , ::;, *) be a non-degenerate commutative
quantale. A fuzzy filt er on a set X in the quantale case, also called an
L-filter, is a mapping M : Lx -+ L such that
(Ql) M(O) = 0 and M(l ) = 1.

(Q2) f ::; g implies M(J) ::; M(g) for all f ,g E L x.
(Q3) M(J) * M(g) ::; M(J *g) for all f ,g E L x .

If* is the infimum 1\ in (L , ::;), then the notion of a fuzzy filter in both
the frame case and in the quantale case coincide. Hence, the quantale
case is more general than the frame case.
In the quantale case, the related partially ordered fuzzy filter monad
(.ri, ::;, ry, fJ), also called the L-filter monad, is defined in the same way
as in the frame case. In particular, the L-filters Tlx(x) : Lx ---* L are
the same mappings as in the frame case.
Moreover, the notions of bounded, homogeneous and stratified fuzzy
filter are defined in the same way as in the frame case. There exist the
partially ordered fuzzy filter submonads of (FL, ::;, ry, 11) defined by all
bounded, homogeneous and stratified L-filters, respectively.
For each set X the supremum of a non-empty subset A of FLX has
the same representation as in the frame case given in Proposition 7
by equation (2). It follows that for each non-empty set X , T]x [X] =
V Tfx(x) also is a fuzzy filter in the quantale case.
xEX
For the infima of fuzzy filters in the quantale case we have:
Proposition 11 Let X be a set. If the infimum of a subset A of FLX

exists, then
(A
MEA
M)(J) = v
h •··· • fn S:: f
M1 , ... ,MnEA
for all f E Lx.
The following is an easy consequence of Proposition 11.
Proposition 12 Let X be a non-empty set and A a linearly ordered

subset of (FLX, ::;). Then the infimum 1\ M exists and for each
MEA
f E Lx we have
(A
MEA
M)(J) v
MEA
M(J) .
A bounded fuzzy filter on a set X will be called a bounded ultra fuzzy

filter if there is no properly finer bounded fuzzy filter on X. Analo-
gously the notions of homogeneous ultra fuzzy filter and of stratified
ultra fuzzy filter are defined.
438 W. Gahler
It follows by means of Zorn's lemma from Proposition 12 that:
Proposition 13 Let X be a non-empty set. In the quantale case we

have: Each bounded fuzzy filter has a finer bounded ultra fuzzy filter.
Analogously, each homogeneous and each stratified fuzzy filter has a
finer homogeneous and a finer stratified ultra fuzzy filter, respectively.
Remark. Clearly, this proposition also holds in the frame case. In

this case because of Proposition 8 we even have that each bounded
ultra fuzzy filter is a homogeneous fuzzy filter.
8. Principal fuzzy filters in the quantale case

In the quantale case we introduce a notion of principal fu zzy filter in
defining the mappings [f, a] : Lx ~ L exactly as in the frame case.
However only some of these mappings appear as fuzzy filters in the
quantale case. We have the following :
Proposition 14 For each set X , each f E Lx and each a :::; Uf

the mapping [f, a] is a fuzzy filter in the quantale case if and only if
f :::; f * f or a * a = 0 holds.
Clearly in the frame case we especially have f = f 1\ f for all f E Lx.

The following counterexample is obtained as an immediate conse-
quence of Proposition 14.
Proposition 15 Let a and b be elements of L such that b * b < b <

a :::; a* a holds. Moreover, let X be a set of at least two elements and
x 0 an element of X. Then [f, a] with f E Lx defined by
if X= Xo
f(x) = { :
if X -::j::. Xo
for all x E X, is not a fuzzy filter in the quantale case.

Under some conditions, in the quantale case, homogeneous fuzzy fil-

ters [f] exist.
Proposition 16 For any f E Lx for which f :::; f * f or Uf *Uf = 0

holds and there is an x EX with f(x) = Uj, the fuzzy filter infimum
in the quantale case
[f] = [f,Uf] !\1Jx[X]

exists and for each g E Lx we have [ f ](g) V (uf *a) v ng.
f*a~g
9. Distinguished and adjoint fuzzy filters, the frame

case
Let L = (L, :::;) be a non-degenerate frame. An L-filter M on a set X
is said to be distinguished ([9]) provided that
M(f v g) = M(f) v M(g)

holds for all f, g E Lx.
In the filter case distinguished means that this filter F is prime, that
is for non-empty subsets M and N of X from M UN E Fit follows
M E F or N E F. Hence, in this case distinguished means that F is
an ultra filter.
In the general frame case the property of a fuzzy filter to be distin-
guished is more complex. As will be shown in the example in Section
12, there are distinguished fuzzy filters which are neither homoge-
neous ultra fuzzy filters nor stratified ultra fuzzy filters and on the
other hand there are homogeneous ultra fuzzy filters which fail to be
distinguished.
In the following the zero-condition (Z) of L is important. (Z) means
that
a > 0 and b > 0 imply a!\ b > 0
for all a, b E L. This condition is needed for proving some types of

compactifications (cf. [9]). Clearly this condition is fulfilled in case
440 W. Gahler
(L, ~) is a complete chain.
Proposition 17 ([9]) If L fulfills the condition (Z), then each strati-

fied ultra fuzzy filter is distinguished.
If L is a non-degenerate complete chain, then each distinguished ho-
mogeneous fuzzy filter is a homogeneous ultra fuzzy filter.
Remark. ([9]) Propositions 13 and 17 imply that under the condition

(Z) each fuzzy filter has a finer distinguished fuzzy filter (which may
not be homogeneous).
In the following let an antitone involution c : L --t L be fixed. Clearly,
for all a, bEL we have a V b = c(c(a) 1\ c(b)).
By means of c each distinguished fuzzy filter M on X can be assigned
a further fuzzy filter Me defined by
Me(!) = c(M(c o !))
for all f E Lx. Me is called the c-adjoint of M.
Proposition 18 The c-adjoint of each distinguished fuzzy filter M is

also a distinguished fuzzy filter and we have (Me)e = M .
If M and N are distinguished fuzzy filters, then M ~ N implies
Me 2: Ne.
From the example in Section 12 we will obtain, that there are distin-
guished fuzzy filters which differ from their c-adjoints and also some
which do not.
If M and N are distinguished fuzzy filters and M = Ne , hence
also Me = N, then {M , N} will be called a c-adjoint pair. In case
M = Me, the fuzzy filter M will be called c-selfadjoint.
Proposition 19 Let M be a distinguished fuzzy filter. If one of the

fuzzy filter-s M and Me is bounded the other- one is stratified and vice
ver-sa. M is homogeneous if and only if Me is homogeneous.
Clearly, if we restrict ourselves to the partially ordered submonad of

all bounded fuzzy filters or of the partially ordered submonad of all
stratified fuzzy filters, then for each non-homogeneous distinguished

fuzzy filter, its c-adjoint does not appear.
10. Distinguished and adjoint fuzzy filters, the

quantale case
In the following let L = (L , ~' *) be a non-degenerate commutative
quantale. If * differs from /\ , for defining the notion of distinguished
fuzzy filter in the quantale case always an antitone involution c : L -t
L is to be fixed.
If not specified we always assume in the following that such a mapping
c is given. Let E9 denote the binary operation on L defined for all
a, bEL by
a E9 b = c(c(a) * c(b)).
If L is a Girard monoid, then as c we can take the negation -, : L -t L
If L = [0, 1], cis given by c(a) = 1- a and *is a continuous t-norm,

then E9 is the t-conorm associated to *· Both notions oft-norm and t-
conorm, introduced by Schweitzer and Sklar ([19]) , play an important
role in fuzzy logic.
For a fixed antitone involution c of L, an L-filter on a set X is called
c-distinguished provided that
M(f E9 g) ~ M(f) E9 M(g)
holds for all f, g E Lx. Notice that in this definition only an inequality
appears, which is in accordance with the inequality in condition (Q3)
of the definition of a fuzzy filter in the quantale case.
Analogously as in the frame case, to each c-distinguished fuzzy filter
M can be associated a further fuzzy filter Me defined by M e(!) =
c(M(c of)) for all f E Lx. Me is called the c-adjoint of M. Propo-
sitions 18 and 19 hold analogously in the quantale case. The notions
c-adjoint pair and c-selfadjoint fuzzy filter are defined analogously as
in the frame case.
In the example which will be presented in Section 13 all c-distinguished
fuzzy filters M are c-selfadjoint.
442 W . Gahler
11. "Basic example"

In this section a special example of fuzzy filters in the frame case will
be considered. For different reasons, this example has been already
considered in [9] in the general case and at first , restricted to homo-
geneous fuzzy filters also in [2] . We are here especially interested in
the distinguished fuzzy filters in this example.
Let X = {0, 1} and L = {0, ~ ' 1} . There are 9 fuzzy subsets on X ,
and among them there are 6 fuzzy subsets which are non-constant.
They are presented in Fig. 2.
L J fs
0
Fig. 2. All nonconstant mappings of X= {0, 1} into L = {0, ~ ' 1}.
There are 25 fuzzy filters in this example. They are shown in Fig. 3,
in which finer filters are situated more downwards. In particular there
are
9 bounded, non-homogeneous fuzzy filters , indicated by small circles,
11 homogeneous fuzzy filt ers, indicated by boldfaced dots, and
5 stratified, non-homogeneous fuzzy filt ers, indicated by small dots.
Whereas in this example all the principal fuzzy filters are the bounded,
non- homogeneous fuzzy filters , their homogenizations as well as H1s =
[!I] A [fs], H16 = [!I] A [!6], Hs2 = [is] A [!2], H52 = [!6]/\ [!2] are
the homogeneous fuzzy filters .
Moreover, Mo = [h] A [!6], M1 = [!4]/\ [is], Mo1 = [fs] 1\ [!6]
and 5 0 = [!I] A [fs]A [f6], S1 = [!2] A [fs] A [!6] are the stratified,
non-homogeneous fuzzy filters.
Proposition 20 Th e following hold:

(1) M0 and M 1 are the only micro fuzzy filters and the only stratified
ultra fuzzy filters .
(2) TJx(O), 7Jx(1), H16 and Hs2 are the only homogeneous ultra fuzzy
filters.
(3) In this example [h, 1], [!4 , 1], 7Jx(O), 7Jx(1), M0 and M1 are the
distinguished fuzzy filters, indicated in Fig. 3 additionally by overlines.
(4) {M 0 , [h, 1]} and {M 1 , [!4 , 1]} are adjoint pairs.
(5) Both fuzzy filter-s 1Jx(O) and 7Jx(1) are selfadjoint.
(0, 0]
TJx(
=[h]
Fig. 3. Graph of all fuzzy filters on X = {0, 1}

with L = {0, ~' 1} in the frame case.
Notice the following:

7Jx(O) and 7Jx(1) fail to be stratified ultra fuzzy filters.
H16 and H52 are homogeneous ultra fuzzy filters which are not distin-
guished.
The distinguished fuzzy filters [h, 1] and [f4 , 1] are neither homoge-
neous ultra fuzzy filters nor stratified ultra fuzzy filters.
12. Analogous lukasiewicz case

As in the preceding section let X = {0, 1} and let L be the set {0, ~, 1}.
We equip L with the usual ordering and with the binary operation *
444 W . Gahler
defined by
1
0 1
* 2
0 0 0 0
1 1
2 0 0 2
1 0 21 1
By a fuzzy filter here we mean an L-filter with L the quantale (L, :::; , *).
There are 22 fuzzy filters in this example. They are shown in Figure
4, where the finer fuzzy filters are situated more downwards. In par-
ticular, there are
11 bounded, non-homogeneous fuzzy filters, indicated by small circles,
and
11 homogeneous fuzzy filters , indicated by bold faced dots.
Stratified fuzzy filters which are non-homogeneous, in particular micro-
fuzzy filters, do not exist.
From Proposition 14 it follows:
Proposition 21 The mappings [!5, 1], [!6, 1], [f5] and [f6] are ex-
cluded as fuzzy filters with respect to (L , :::; , *).
There are four bounded non-homogeneous fuzzy filters which are not
single principal fuzzy filters. They are:
834 [h, ~], 1\ [!4, ~] B36 = [/s, ~] 1\ [!6, ~],

854 [f5, ~] 1\ [f4, ~], B56 = [f5, ~] 1\ [!6, ~].
Notice that for the principle fuzzy filters [!,a] appearing in this ex-
ample in some cases only the condition f :::; f * f (e.g. for [h, 1]) and
in some other cases only the condition a* a = 0 (e.g. for [h , ~]) is
fulfilled.
[0,0]
'TJX (0)
=[h]
Fig. 4. Graph of all fuzzy filters on X = {0, 1}

with L = {0, ~' 1} in the Lukasiewicz case.
Among the principal fuz zy fil ters there are three which are homoge-
neous, that is [ 0] = [J,
~] , [!I] = [/I ,~] and [h] = [h , ~ ].
The further homogeneous fuzzy filters are [h] = T,lx(O), [!4 ] = T,lx(1)
and
H12 = [!1] 1\ [h] , H14 = [!I] 1\ [j4 , ~] , H32 = [h, ~] 1\ [h] ,

H3 = [h, ~] 1\ [0], H4 = [j4 , ~ ] 1\ [OJ , H34 = [h, ~] 1\ [!4, n
The property of a fuzzy filter to be distinguished here we understand
with respect to the associated binary operation EB of * defined by
means of the antitone involution c with c(a) = 1 - a , that is we have
EB 0 1
2 1
0 0 2 1
1 1 1 1
2 2
1 1 1 1
446 W. Gahler
Proposition 22 In this example the ultra fuzzy filters 1Jx(O), 7Jx(l)

and H12are the only distinguished fuzzy filters (indicated in Fig. 4 ad-
ditionally by overlines). They are selfadjoint.
13. Balanced fuzzy filters
Let L = (L, :::;) be a non-degenerate frame. An L-filter will be called

balanced ([9]) whenever a finer distinguished homogeneous fuzzy filter
exists.
All balanced fuzzy filters are bounded. In the "basic example" H16
and H52 are homogeneous fuzzy filters which are not balanced.
Proposition 23 ([9]) Let L be a non-degenerate chain. Then we have:
(1) All balanced fuzzy filters define a partially ordered submonad of

the partially ordered fuzzy filter monad.
(2) All homogeneous balanced fuzzy filters define a partially ordered

submonad of the partially ordered fuzzy filter monad.
14. Fuzzy convergence structures
In the following let L = (L, :::;, *) be a non-degenerate commutative

quantale and let = ( cp, :::; , 1], J.L) be the partially ordered L- filter
monad or one of its partially ordered submonads. The special cases
we will consider are listed in the following table. We describe these
cases only by noting the related fuzzy filters.
Fuzzy filters of <P

Case (A) allfuzzy filters
Case (S) allstratified fuzzy filters
Case (B) allbounded fuzzy filters
Case (H) allhomogeneous fuzzy filters
Case (B') allbalanced fuzzy filters with
(L, :S:) a non-degenerate complete chain
and*= 1\
Case (H') all homogeneous balanced fuzzy filters with
(L, :S:) a non-degenerate complete chain
and*= 1\.
By the interior operator of a fuzzy pretopology p : X ---+ <pX we mean
the mapping int : Lx ---+ Lx defined for all f E Lx and x E X by
(intf)(x) = p(x)(J). (3)
According to the following proposition, the interior operator of a fuzzy

pretopology can be characterized independently of this fuzzy pretopol-
ogy.
Proposition 24 The interior operator int of a fuzzy pretopology fulfills

the following conditions:
(1) intf :S: f holds for all f E Lx and inti= I.
(2) f :S: g implies intf :S: intg for all f, g E Lx.
(3) intf * intg :::; int(J *g) holds for all f , g E Lx .
On the other hand, each mapping int : Lx ---+ Lx which fulfills these
conditions, is the interior operator of the fuzzy pretopology defined by
equation (3). The interior operator int characterizes a fuzzy topology
if and only if additionally into int = int.
This proposition shows that fuzzy pretopologies and fuzzy topologies

can be characterized only by fuzzy sets. Clearly in this proposition
the partial ordering and the products of fuz zy subsets are defined
argumentwise.
448 W. Gihler
In the frame case conditions (2) and (3) are equivalent to the condition
(2)' intf 1\ intg = int(f 1\ g) for all j, g E Lx.
In a -pretopological space a mapping f is called open provided that

intf =f.
Proposition 25 The set Tof all open fuzzy sets f E Lx of a -

pretopology fulfills the condition
(0') T is closed with respect to all suprema and all finite products
!1 * · · · * fn of fuzzy sets.
Each -topology p can be characterized as a subset T of Lx which ful-
fills condition (0') by taking intf = V g for all f E Lx.
g"S_ f,gET
Remark. In the frame case, (0') is a well-known condition for char-

acterizing a -topology. Notice that (0') and condition (0) in Propo-
sition 3 look completely different.
Proposition 26 The set r of all open fuzzy sets of a fuzzy topology

equipped with the argumentwise defined partial ordering and the argu-
mentwise defined product *, is a quantale.
Remark. Notice that for each subset M of r the supremum of

M in the lattice (r, :::;) is the argumentwise defined supremum x H
V f (x) (x E X) and the infimum of M in (r , :::;) is the interior of the
! EM
argumentwise defined infimum x H 1\ f(x) (x E X).
! EM
16. Closure operator of a fuzzy convergence structure

Let L = (L, :=:;;, *) be a non-degenerate commutative quantale and let
 = (rp, :::;, ry, p,) be specified as in Section 15. Moreover, let T be a
-convergence structure on a set X and let as usual t 1 : T ----+ <pX be
the first projection ofT.
Proposition 27 The sup-inverse 'P;t 1 : (/)X ---+ (/JT of J.Lx o'Pt 1 : (/JT ---+
'PX is given by
v (4)
for all M E 'PX and h E LT, wheTe joT each f E Lx, ef : 'PX ---+ L is
defined by ef (N) = N (f).
Because of this proposition the closure operator cl 'PX ---+ (/)X is

given as follows.
Proposition 28 For· each M E (/)X and f E LX we have
(elM)(!) = v
N(h)•···•NCJn) ~ f (x)
for all N -tx
(5)
In the following we assign to each finite subset { JI, ... , fn} of Lx

as its fuzzy set closuTe a fuzzy set cl (!1, ... , fn) E LX defined for all
X EX by
cl (h, · · ·, fn)(x) = V(N(h) * ·· ·* N(.fn)) · (6)

N-tx
Some properties of this fuzzy set closure are the following.

(1) h * · · · * fn = 0 implies cl(JI,, ... , fn) = 0,
(2) h :::; 91, ... , fn:::; 9n implies cl(h, ... , fn) :::; cl(91, ... , 9n) and
(3) h * ... * fn :::; cl (!1, ... 'fn) and cl (!1, ... 'fn, 91, ... '9m) :::;
cl (h, .. . , fn) * cl (91, ... , 9m) holds for all JI, ... , fn, 91, ... , 9m E Lx.
By means ofthe notion offuzzy set closure, equation (4) can be written
as follows:
(ciM)(.f) v
cl (JI ,.. ..Jn) :S; J
M(h*·· · *fn)· (7)
450 W. Gahler
Because of the result on finer ultra <p-objects in Proposition 13, we

have:
Proposition 29 The <p-objects N appearing in equations (5) and (6)

can be assumed to be ultra <p-objects.
The frame case. In the following we assume that * equals 1\. In this
case we have more results. Some are even more simple. For instance,
for all JI, ... , fn E Lx it follows that cl (JI , ... , j~) = cl (fi/\ .. .1\fn).
Hence, in the frame case, we only need the notion of fuzzy set closure
for single fuzzy sets. Thus instead of (6) we only need
(clf)(x) = V N(.f). (8)

N -+ x
Proposition 30 In the frame case the equations (4) , (5) and (7) can
be reduced, respectively, to
<p~t1(M)(h) v M(g) ,
t1 ::; h
v M(g),
e9 o
(ciM)(.f) (9)
N(g ) ~ f( x)
for all JV -4x
(ciM)(.f) V M(g).
clg :S: f
Because of Proposition 29 in the equations (8) and (9) the <p-objects

N can be assumed to be ultra <p-objects.
Proposition 31 In the cases (A) and (S) under the assumption that
(L, ~) fulfills the condition (Z) and in the cases (B') and (H') (in which
(L, ~ ) even is a complete chain) we have cl M V ciN = cl (M V N)
for all M,N E <pX .
Hence, under the assumption on stated in Proposition 31 we have

that the supremum of finitely many closed <p-objects is closed.
Examples. In the "basic example" in the cases (A), (B) and (B') as
well as in the "analogous Lukasiewicz case" we have: p and q defined
by p(O) = [!I], q(O) = [h, ~] and p(l) = q(l) = 7Jx(l) are different
-pretopologies which have one and the same closure operator.
References
[1] G. Birkhoff, Lattice theory, Amer. Math. Soc. Colloq. Publ. 25 , 1940.
[2] P. Eklund, W. Gahler, Completions and compactifications by means

of monads, in: Fuzzy Logic, State of Art, Kluwer (1993) 39-56
[3] W. Gahler, Monadic topology- a new concept of generalized topology,

in: Recent Developments of General Topology and its Applications,
Akademie Verlag 1992, 136-149.
[4] W. Gahler, The general fuzzy filter approach to fuzzy topology, I,

Fuzzy Sets and Systems 76 (1995) 205-224.
[5] W. Gahler, The general fuzzy filter approach to fuzzy topology, II,
Fuzzy Sets and Systems 76 (1995) 225-246.
[6] W. Gahler, Monadic convergence structures, Seminarberichte aus dem

Fachbereich Mathematik, Fernuniversitat Hagen 67 (1999) 111-130.
[7] W. Gahler, General topology - the monadic case, examples, applica-

tions, Acta Math. Hungar. 88 (4) (2000) , 279-290.
[8] W. Gahler, Extension structures and completions in topology and

algebra, Seminarberichte aus dem Fachbereich Mathematik, Fernuni-
versitat Hagen (2001).
[9] W. Gahler, P. Eklund, Extension structures and compactifications,

Universitat Bremen, Mathematik-Arbeitspapiere Nr. 54 (2000) 181-
205.
452 W . Gahler
[10] H. Herrlich, M. Husek, Galois connections categorically, J. Pure Appl.

Algebra 68 (1990) 165-180.
[11] U. Hohle, Monoidal logic, in: Fuzzy Systems in Computer Science,

Vieweg Verlag 1994, 223-234.
[12] U. Hohle, Locales and L-topologies, Universitat Bremen, Mathematik-

Arbeitspapiere Nr. 48 (1997) 223-250.
[13] D. C. Kent, On convergence groups and convergence uniformities,

Fundamenta Math. 60 (1967) 213-222.
[14] H.-J. Kowalsky, Limesraume und Komplettierung, Math. Nachr. 12

(1954) 301-340.
[15] 0. Ore, Galois connections, Trans. Amer. Math. Soc. 55 (1944)

493-513.
[16] H. Rasiowa, R. Sikorski, The mathematics of metamathematics,

Paristwowe Wydawnictwo Naukowe 1963.
[17] H. I. Rosenthal, Quantales and their applications, Pitman Research

Notes in mathematics 234, Longman, Burnt Mill, Harlow, 1990.
[18] J. Schmidt, Beitrage zur Filtertheorie II, Math. Nachr. 10 (1953)

197-232.
[19] B. Schweitzer, A. Sklar, Probabilistic metric spaces, North-Holland,

1983.
Author's address:
Werner Gahler,
Scheibenbergstr. 37,
12685 Berlin
Dyadic Mathematics -
Abstractions from Logical Thought
R. Wille
Abstract
Mathematics finally aims at supporting thought and action of human be-
ings. For fulfilling this aim, mathematical abstractions of logical thought
are essential. Because human logical reasoning is based on concepts as the
basic units of thought, the dyadic mathematization of concepts performed
in Formal Concept Analysis is such an abstraction. The dyadic nature of
concepts is grasped through the notion of a formal context with its object-
attribute-relation and its corresponding Galois connection. It is outlined
how this dyadic foundation may lead to dyadic conceptions and results in
order and lattice theory, contextual logic, algebra, and geometry, and how
that supports the development of a human-oriented mathematics.
AMS Mathematics Subject Classification: 06A15, 06B23, 03B, 08A, 051D
Key words: Dyadic Mathematics, Formal Concept Analysis, Ordered sets,
Lattices, Galois connection, Contextual logic, General Algebra, Geometric
measurement
Contents
1. Introduction
2. Galois Connections in Human Thought
3. Dyadic Order Theory and Logic
4. Dyadic Algebra and Geometry
5. Outlook
1 Introduction
Dyadic Mathematics is understood as a human-oriented development of
mathematics based on the conviction that the aim of mathematics finally
453
K. Denecke eta!. (eds. ), Galois Connections and Applications, 453-498.
454 R. Wille
lies in the support of thought and action of human beings (cf. [Wi02a]).
For elaborating this view, we have to answer the basic question: How
can mathematics support human thinking? Following Peirce's philosoph-
ical pragmatism (see for instance [Pe92]), we obtain as a short answer that
mathematics is able to promote human reasoning. More explicitly, logical
thinking as expression of human reason grasps actual realities by the ba-
sic forms of thought: concepts, judgments, and conclusions; mathematical
thinking abstracts logical thinking for hypothetically developing a "cosmos
of forms" of potential realities. This close relationship between logical and
mathematical thinking allows mathematics to support humans in their rea-
soning about realities (cf. [WiOl b]).
This paper gives a survey about basic attempts to unfold the idea of dyadic
mathematics. Because of the lack of space, most explanations can only
sketch the different findings and will be far from being complete. But, for
all the different approaches, a unifying basis evolves out of the understand-
ing that human reasoning is based on concepts as the basic units of thought
(cf. [SeOl]). Since a concept is constituted by its extension and its inten-
sion, concepts are of dyadic nature and may even be viewed as the origin
of the notion of Galois connection. This is outlined in Section 2 in con-
nection with an introduction to Formal Concept Analysis, which has been
developed out of a mathematization of "context" and "concept". A general
dyadic method for mathematizing transformations from ideas to concepts is
presented and applied to mathematize linear continua order-theoretically.
Section 3 points out that orders in general yield basic patterns of thought
with which humans grasp realities. Prominent mathematizations of those
patterns are general ordinal and contraordinal scales which are discussed
in their central role for a dyadic order theory. These mathematical abstrac-
tions of logical thought are related to the dyadically developed Contextual
Logic for which a close connection between the mathematized concepts and
judgments is essential. Section 4 reports on basic theorems of a possible
dyadic algebra and on applications in elementary linear algebra. A possible
dyadic geometry is then discussed through geometric measurement with its
four mathematization levels from (1) empirical situations to (2) empirical
models, (3) synthetic models, and (4) numerical (algebraic) models. Fi-
nally, an outlook suggests, in particular, future work on an extension of
dyadic mathematics to triadic mathematics in the sense of Peirce's three
universal categories.
Dyadic Mathematics -Abstractions from Logical Thought 455
2 Galois Connections in Human Thought

In human thought, concepts are basic for mentally reconstructing realities,
their conditions and relationships (cf. [SeOl]). For this task concepts are
formed as the basic units of thought in dynamic processes within social and
cultural environments. A concept is constituted by its extension, compris-
ing all objects which belong to the concept, and its intension, including
all attributes (properties, meanings) which apply to all objects of the ex-
tension. Concepts can only live in relationships with many other concepts
where the subconcept-superconcept-relation plays a prominent role. Being a
subconcept of a superconcept means that the extension of the subconcept
is contained in the extension of the superconcept which is equivalent to the
relationship that the intension of the subconcept contains the intension of
the superconcept (cf. [Wa73] , p.201). That reciprocal relationship between
extensions and intensions of concepts is the origin of Galois connections in
human thought . This becomes clear by the mathematization of contexts
and concepts in Formal Concept Analysis. Therefore, we start our dis-
cussion with an elementary introduction to Formal Concept Analysis (cf.
[GW99]).
Definition 1. A formal context is defined as a set structure lK := ( G, M , I)

for which G and Mare sets while I is a binary relation between G and M ,
i.e. I <;;;; G x M; the elements of G and M are called objects (in Ger-
man: Gegenstande) and attributes (in German: Merkmale) , respectively,
and gim, i.e. (g, m) E I , is read: the object g has the attribute m. The
following derivation operators are defined for arbitrary X <;;;; GandY <;;;; M:
X {mE M I gim for all g EX} ,

y {g E G I gim for all mE Y}.
Obviously, the two derivation operators satisfy the following three condi-
tions:
( 1) z1 c- z2 ===} z11 :J
-
z2'1 (2) z c- ziT ' (3) ziii = z 1 ·'
this means that the two derivation operators form a Galois connection be-
tween the power sets of G and M ordered by set-inclusion. The general
definition of a Galois connection based on a binary relation was first pub-
lished by G. Birkhoff in [Bi40] where he even mentioned the interpretation
of the binary relation as object-attribute-relation (as G. Birkhoff has told
the author, he was inspired to his general definition by the presentation
456 R. Wille
of self-adjoint operators in Hilbert spaces in M. Stone's influential book

[St32]).
The rich experiences in applications have shown that users of Formal Con-
cept Analysis could easily and successfully activate the Galois connection of
the object-attribute-relation in their logical thinking if they were familiar
with the content of the formalized context. But, in most cases, the users
were not able to mathematically understand Galois connections and For-
mal Concept Analysis; therefore it is important to establish enough bridges
between mathematical and logical thinking. The basic notion of a formal
context can be logically understood best via data tables. The example
shown in Fig.2 (see [PrOO]) represents a formal context with 21 objects
(types of sleeping bags) evaluated by 8 attributes (the crosses describe the
binary relation). Such data tables can also support the logical understand-
ing of the following mathematical definition of formal concepts and of the
subconcept-superconcept relation:
Definition 2. A formal concept of a formal context IK := (G, M,I) is a

pair (A, B) with A~ G , B ~ M, A= BI, and B =AI ; A and Bare called
the extent and the intent of the formal concept (A, B) , respectively. The
subconcept-superconcept-relation is formalized by
The set of all formal concepts of IK together with the defined order relation
is denoted by IB (IK).
0
A general method of constructing formal concepts uses the derivation oper-
ators to obtain, for X ~ G and Y ~ M , the formal concepts (X I I , X I) and
(YI,YII). For an object g E G , its object concept 19 := ({g}II , {gV) is
the smallest concept in IB(IK) whose extent contains g and, for an attribute
mE M, its attribute concept J-Lm := ({mV, {m}II) is the greatest concept
in IB(IK) whose intent contains m.
Basic Theorem on Concept Lattices. [Wi82] Let IK := ( G, M,I) be a

formal context. Then IB (IK) is a complete lattice, called the concept lattice
of (G, M,I), for which infima and suprema can be described as follows:
f\ (At, Bt)
t ET t ET l ET
V(At, Bt)
lET l ET l ET
One Kilo Bag X X X X
Sund X X X
Kompakt Basic X X X X
Finmark Tour X X X
Interlight Lyx X X X
Kompakt X X X
Touch the Cloud X X X
Cat's Meow X X X X
Igloo Super X X X
Donna X X X
Tyin X X X
Travellers Dream X X X X
Yeti light X X X X
Climber X X X
Viking X X X X
Eiger X X X
Climber light X X X X
Cobra X X X X
Cobra Comfort X X X
Foxfire X X X X
Mont Blanc X X X X
Fig. 1: Formal context about sleeping bags (c - cheap, m- medium-

prized, e - expensive, d- downy feather, s- synthetic fibre , g-
good , a - acceptable, b - bad)
458 R. Wille
In general, a complete lattice L is isomorphic to ~ (OC) if and only if there

exist mappings :Y : G ----+ L and P, : M ----+ L such that :YG is V-dense
in L (i.e. L = {V X I X~ :YG}) , P,M is /\-dense in L (i.e. L ={!\X I
X ~ P,M}), and gl m ~ ;yg ::::; P,m for g E G and m E M; in particular,
L ~ ~(L, L, ::S) and furthermore: L ~ ~(J(L) , M(L) , ::::;) if the set J(V) of
all V-irreducible elements is V-dense in L and the set of all !\-irreducible
elements is !\ -dense in L.
Fig. 2: Concept lattice of the context in Fig.2
Concept lattices can be effectively represented by labelled line diagrams, as

for example shown in Fig.2. In such diagrams the name of each object g
is attached to its represented object concept 1g and the name of each at-
tribute m is attached to its represented attribute concept ~-tm; by the basic
theorem, this labelling allows to read the extents, the intents, and the un-
derlying formal context from the diagram. Non-mathematicians are taught
to understand the represented concept lattice logically as a hierarchical
network of concepts where each concept is represented by a little circle so
that its extension (intension) consists of all the object (attributes) whose
names can be reached by a descending (ascending) path from that circle.
In this way, users learn to activate diagrammatically the Galois connection
of the object-attribute-relation. By our experiences we got the impression
that the labelled diagrams "speak" to those users who are familiar with the
content represented in the formal context; quite often, after a short glance
at the diagram, users even recognize mistakes in the underlying data con-
text (cf. [WiOOb]). This direct support of logical thinking indicates that the
contextual and holistic nature of concepts in human mind are remarkably
preserved by our mathematization of formal contexts and concepts.
Since concepts are the basic units of human thought, the mathematization
of concepts by formal concepts based on Galois connections can be used to
make mathematically explicit further ideas and conceptions. This shall be
demonstrated by establishing a general method for mathematizing transfor-
mations from ideas to concepts. Those transformations shall be understood
in the sense of the structure-genetic psychology of Jean Piaget [Pi59] (see
also [SeOl]) .
The mathematization of this transformation starts from an ordered set

C := ( C, ~) of formally represented ideas. The process of concept formation
is modelled by filters of C on the object side and by ideals of C on the
attribute side. In the ordered set C , a filter is defined as a subset F of
C for which a E F and a ~ b always imply b E F and for a, c E F there
is always an element d with d ~ a, c; an ideal is dually defined 1 . This
approach leads to the formal context IK(C) := (J(C), J(C) , b.) for which
J(C) is the set of all non-empty filters F of C and J(C) is the set of all
non-empty ideals I of C , while F b.I : {::} F n I -1- 0. Thus, a filter as
"object" has an ideal as "attribute" if and only if the filter and the ideal
meet in at least one formally represented idea. Important filters are the
!-maximal filters Fin J(C) for which I is an arbitrary ideal in J(C) and F
is maximal with respect to the property F n I = 0; the F-maximal ideals
are dually defined. The set of all !-maximal filters is denoted by Jo( C) and
the set of all F- maximal ideals by J 0 (C). The following theorem informs
about the basic structural properties of the concept lattice of !K( C) (cf.
[Ur78] ,[SW86],[Ha92]) :
Theorem 1. For· an ordered set C := ( C, ~), the map
" : X f--t ( { F E J( C) IX E F} , {I E J( C) IX E I})

is an order embedding of C into the concept lattice of the related context
JK(C) with "(x 1\ y) = "(x) 1\ "(Y) resp. "(x V y) = "(x) V "(Y) if x 1\ y resp.
x V y exists in C. In s:B(IK(C)), J(s:B(IK(C))) = /Jo(C) is V-dense and
M(IB(IK(C))) = J.LJo(C) is /\ -dense, i.e. IB(IK(C)) 2:! IB(Jo(C),Jo(C) , b.) .
1 Filters and ideals represent processes of converging ideas which are dual to
each other.
460 R. Wille
Proof For x E C, ({FE ~(C) I x E F} , {IE J(C) I x E 1}) is a formal

concept of the context JK.(C) which can be concluded using [x) E {F E
~(C) I x E F} = {(x]}~ and (x] E {IE J(C) I x E I} = {[x)}~. By
the equivalences x ~ y ¢:? (x] ~ (y] ¢:? {(x]}~ 2 {(y]}~ ¢:? 1[x) ~ 'Y[y),
it follows that L is an order embedding of C into ~(JK.(C)) . That L is
1\- resp. V-preserving is an immediate consequence of the equivalences
x 1\ y E F ¢:? (x E F andy E F) resp. x V y E I ¢:? (x E I andy E 1) .
Now, let Fa be an !-maximal filter of C . Then 1Fo has the infimum !Fol\f.l.l
as its only lower neighbour in ~( JK.(C)) which yields 1Fo E J(~(JK.(C)))
and hence /~o(C) ~ J(~(JK.(C))). Conversely, let b be a V-irreducible
formal concept of ~(JK.(C)) . Then, by the Basic Theorem, b has to be an
object concept, i.e., there is an F E ~(C) with b = 1F. For the unique
lower neighbour c of b, by Zorn's Lemma, there exists a maximal formal
concept ~ with c ~ ~ and b <f ~- It follows that ~ is /\-irreducible in
~(JK.(C)) and therefore an attribute concept by the Basic Theorem, i.e.
there is an I E J( C) with ~ = f.J.l . Because of 1F 1\ f.J.l = c, F is !-maximal
and hence FE ~o(C). Thus, /~o(C) = J(~(JK.(C))) is proved. Dually, we
obtain M(~(JK.(C))) = f.J.Jo(C).
Finally, let F be an arbitrary filter of C. For each a E C \ F and its

principal ideal (a], by Zorn's Lemma, there exists an (a]-maximal filter
Fa with F ~ Fa . For all a E C \ F , it follows that Fa E ~o(C) and
F = naEC\F Fa which has ,F = VaEC\F !Fa as consequence. Therefore,
by the Basic Theorem, /~o(C) is V-dense in ~(JK.(C)). Dually, we obtain
that f.J.Jo(C) is /\-dense in ~( JK.(C)) . D
Theorem 1 yields a general method how to transform the ordered set

of ideas to a concept lattice in which each formal concept is the supre-
mum of V-irreducible concepts and an infimum of /\-irreducible concepts.
This method shall now be concretized in making mathematically explicit
the Aristotelean conception of the time continuum which he uses to solve
Zenon's paradox of the flying arrow (see [Ar95],[Wi03c]). First, we start
with a suitable order-theoretic mathematization of ideas of one-dimensional
continuum structures which is illustrated by the left diagram in Fig.3:
1 L(l)
Fig. 3: A linear continuum represented by an ordered set and its con-

cept lattice
Definition 3. A linear continuum structure is defined as an ordered set

C := ( C, :S) for which
(1) Cis a V-semilattice with greatest element 1 and without
lower bound,
(2) the A-irreducible elements of C form two disjoint dense chains
cl and c2 without lower and upper bounds
and with Cl V C2 = 1 for all Cl E C1 and C2 E C2 ,
(3) CI A c2 = d1 A d2 implies c1 = d1 and c2 = d2
for all c1, d1 E C1 and c2 , d2 E C2,
(4) there is an antiisomorphism c-j 1---7 c f- from cl onto c2 so that
c = {1}UC1UC2U{c-j 1\ d I c-j E cl and dE c2 with cf- < d}.
The elements of C are called (linear) continua and the pairs (c-1, c f- ) of
corresponding elements c-1 E C 1 and cf- E C 2 are called the cuts of C. ()
Example 1. The linear continuum-structure C JR := (CJR , ~)of the ordered

field JR. of real numbers consists of the set CJR of all open intervals ]r, s[:=
{x E JR. I r < x < s} with r E JR. U { -oo} and s E JR. U { +oo}. In this
continuum-structure the supremum is the convex hull of the set-theoretic
union and the largest element is]- oo, +oo[; the 1\-irreduzible elements form
the dense chains C1 := {]- oo, s[l s E JR.} and C2 := {]r, +oo[l r E JR.}; the
assumed antiisomorphism between cl and c2 is given by ]-oo, r[ t--7]r, +oo[.
The cuts of CJR are therefore the pairs (]- oo, r[, ]r, +oo[) (r E JR.).
For making Theorem 1 applicable to linear continuum-structures C, the

following lemma determines the !-maximal filters und F-maximal ideals of
those continuum-structures.
Lemma 1.
F1 := C1 U {1} and F2 := C2 U {1} are the "extreme " !-maximal filters,
462 R. Wille
and h := {x E C I x ~ c for some c E Ct} and I2 := {x E C I x ~

c for some c E C2} are the "extreme" F -maximal ideals of C. The cuts
(c-1, c f-) of C determine the other I -maximal filters of IK{ C) by
Fe-<:= {x E c I X~ c-1/\ d for some dE c2 with c f- < d} ,

Fe'r := {y E c I y ~ c f- 1\ d for some dE cl with c-1 < d} ,
and the other F -maximal ideals of C by
I(e-<) := {x E C I X< c-1} and I(e'r-] := {y E C Iy ~ c f- },
I(e-<] := {x E C Ix ~ c-1 } and I(e'r-) := {y E C I y < c r- }.

Proof Since C is the disjoint union of F1 and h as of F2 and h, F 1 is
I2-maximal, his F1-maximal and F2 ish-maximal, his F2-maximal. Be-
cause Cis the disjoint union of Fe-<, I(e-<), and I(e 'r], Fe-< is an I(e-< rmaximal
filter and an I(e'rrmaximal filter , while I(e-<) and I(e'r] are Fe-1-maximal ide-
als; analogously, Fe'r is an I(e-<rmaximal filter and an I(e'rrmaximal filter ,
while I(e""'] and I(e'r) are Fe'r-maximal ideals.
Now, let Fo be a filter and let Io be an ideal of C so that Fo is Io-maximal

and Io is Fa-maximal; furthermore, let a:= V Io. By Definition 3{4) , there
are elements c-1 E C1 U {1} and df- E C2 U {1} with a= c-1/\ dr-. If c-1 Fo tt
then Fo must be equal to Fe'ri analogously, if df- t!. Fo then Fo must be
equal to Fd-<· It remains the case that c-1,df- E Fo which is equivalent to
a E Fo. If Ion Fe-< = 0 then Io must be equal to I(e""' ) and hence Fo = Fe""'i
analogously, if Io n Fd'r = 0 then Io must be equal to I(d'r-) and hence
Fo = Fd'r-. It remains to consider that there are elements e, f E Io with
e E Fe-< and f E Fd'r-· By Definition 3{4), it follows e V f = c-1/\ df- =a
and hence a E Io which contradicts a E Fo. D
Theorem 2. In the concept lattice of the formal context IK{C) :=(~(C) ,

J(C), ~) of a linear continuum-structure C,
(1) i{l) = 1F1 V 1F2 is the largest element,
{2) /~o (C) is the set of atoms, which is the disjoint union of
the sets
A1 :={IF!} U {!Fe 'r I cf- E C2} and
A2 := {JF2} U {!Fe-< I c-1 E Cl} ,
{3) tDo(C) = {JF1 V 1F IF E A1} U {JF2 V 1F IF E A2} is
the set of the /\-irreducible elements, which is the disjoint
union of the convex chains [JF1 , i{l)[ and [JF2 , i{l)[,
(4) each cut (c-1,cf-) yields that 1F1 V rFc-< = t-d(c-<J and
1F2 V r Fc't- = f.lJ(c 't-J, p,I(c-<J is a lower neighbour of
p,I(c""'] V rFc't- and an upper neighbour of p,I(c-<),
p,I(c't-] is a lower neighbour of p,I(c't-] V r Fc.., and an upper
neighbour of p,I(c't-) ,
(5) x = d-1 1\ cf- yields that t(x) := ({FE J(C) I x E F},
{IE J( C) IX E I}) = r Fc't- v rFd-1 = p,I(d-1] (\ f.lJ(c't-]·
Proof (1): I E {F1}.6. n {F2}.6. implies I = (1] = C, wherefore t(1)

1F1 V 1F2 and t(1) is the largest element of ~(IK(C)).
(2): By Theorem 1, rJo(C) is V-dense in ~(IK(C)) and constists exactly
of the V-irreducible formal concepts of IK( C). Therefore, by Lemma 1, the
formal concepts in 1Jo(C) must be exactly the atoms of ~(IK(C)), which
divide into the sets A1 and A2.
(3): By Theorem 1, p,Jo(C) is /\-dense in ~(IK(C)) and consists exactly of
the /\-irreducible formal concepts of IK(C). By Lemma 1, each of the /\-
irreducible concepts lies in one of the two chains [rF1, t(1)[ and [rF2, t,(1)[.
Because of b = 1\{p,I E [rFl,t(1)[1 p,I ~ b} V /\{f.l,[TF2,t(1)[1 p,I ~ b},
[rF1, t(1)[U[rF2, t(1)[ is the set of all /\-irreducible formal concepts of IK( C).
(4): Since{F1}.6. n {Fc-< }.6. = {IE J(C) I Icc-<] ~ I}) for any cut (c-1,cf- ) it
follows that 1F1 VrFc-< = p,I(c-<Ji analogously, we obtain 1F2VrFc't- = p,I(c't-]·
Because of {I(c-<]}.6. = {FE J(C) I c-1 E F} and {I(c-<) }.6. = {F E: J(C) I
x E F for some x < c-1}, p,I(c-<] is an upper neighbour of p,I(c""' ) and, because
of I(c-<] n Fc't- = 0 and I(d-<J n Fc't- -=f. 0 for all d-1 > c-1 in C1, p,I(c't-J is a lower
neighbour of p,I(c-<] V rFc't-; analogously, we obtain the third assertion of
(4).
(5): Let X = d-1 (\ cf- in c. Since (x] = (d-1] n (c f-] , it follows that t(x ) =
p,I(d-<J 1\ p,I(c't-J· [x) = F (c't-J n F(d-< J leads to t(x) = r F(c't-J n rF(d-< ]· 0
Since the linear continuum-structure C is embeddable into the related con-

cept lattice ~(IK(C)) by Theorem 1, this concept lattice yields an extended
conceptual frame for the continuum-structure C which is made explicit in
some aspects by Theorem 2. Important for our view is that "point con-
cepts" are established by the atoms of~(IK(C)) which have every other for-
mal concept as supremum by (2). By (4), each cut of the linear continuum-
structure yields two point concepts delimiting the cut; the supremum of
those point concepts could be viewed as a point in the usual sense. (5) in-
dicates how point concepts may represent the boundaries of continua in the
sense of Aristotle for whom durations do not consist of time points, time
points are only boundaries of durations (hence motion takes place only in
464 R. Wille
durations, but not in time points). These hints shall suffice to demonstrate
the fruitfullness of concept-analytic methods and to support the Aristote-
lean conception of continua. Finally, the concept-analytic results shall be
even more concretized within the example of the real continuum-structure.
Example 2. The real continuum-structure CIR decribed in Example 1 is
embedded by the map L into the concept lattice of the formal context IK( C JR)
by Theorem 1, which can be illustrated by a linear ordered set (IR, :::;)
extending (JR., :::;). (IR, :::;) is defined by
i :=(JR. x {- 1,+1}) U{( - oo, +1),(+oo,-1)} and
(r, u) :::; (s, v) : {::} r < s or (r = s and u:::; v);
The linear ordered set (IR \ {(-oo, + 1), (+oo, -1)} , :::;) obviously arises out
of (JR., :::; ) by dividing each real number r into t he two elements (r, -1) <
(r, +1). i can be mapped bijectively onto the set of all atoms of ~(IK( C JR))
by the map a with a( - oo, +1) := 1F1, a(r, - 1) := !F}--oo,r[ • a(r, +1) :=
IF]r,+ oo[• and a( +oo, - 1) := 1F2 . To simplify we define - oo := 1F1,
r- := !F}--oo ,r[• r+ := IF]r,+oo[• and +oo := 1 F 2. By a(r, u) :::; a(s, v) :
{::} (r, u) :::; (s, v), the linear order of (IR, :::;) is transferred to the set of
atoms of ~(IK(C JR )) ; with respect to this order :::;, -oo is the smallest
atom, +oo is the largest atom, and r- < r+ < s - < s+ if r < s in JR.
The continua of the real continuum-structure QIR are represented in the
concept lattice by the formal concepts t(]r, s[). By Theorem 2(5), we have
t(]r, s[) = (r+) V (s- ), which indicates that the atoms a with r+ :::; a:::; s -
are exactly the atoms below t(]r, s [); thus, it is meaningful to say that the
point concepts r+ and s- are the boundaries of the continuum concept
t(]r, s [). The cuts of the real continuum-structure are represented in the
concept lattice by the pairs (r - , r+ ); in this conceptual frame, r - and r+
stand for the indecomposable partial subpoints of the decomposable real
point which is represented by the formal concept (r-) V (r+) .
3 Dyadic Order Theory and logic

Humans grasp realities first of all in relational forms where classificatory
and comparative concepts play a basic role (cf. [Ca66]). Abstracting those
logical forms leads to the mathematical theory of (quasi-) ordered sets. A
dyadic view of (quasi-)ordered sets arose naturally by applying Formal Con-
cept Analysis in data analysis; for those applications, (quasi-)ordered sets
form the basis for different types of ordinal scales which are used to derive
concept lattices for a great variety of data sets. For making this mathe-
matically explicit we have to generalize the notion of a formal context:
Definition 4. A (complete) many-valued context is a set structure (G , M ,

W, I) where G, M, and Ware sets and I~ G x M x W such that (g , m , v) E
I and (g,m,w) E I imply v = w; for (g , m , v) =I we often write m(g) = w
and understand m as a map from G to W . The elements of G , M , and
W are called objects, attributes, and attribute values, respectively. More
generally, an ordinal context is a set structure (G , M, (Wm, Sm)mEM, I) for
which Sm is an order on Wm for all mE M and (G, M , UmEM{m} x Wm,I)
is a many-valued context with (g , n , (m, w)) E I::::} n = m. 0
For reaching a conceptual understanding of many-valued contexts, the
method of conceptual scaling has been developed [GW89] . This method
asks for a purpose-oriented conceptual interpretation of the (potential)
attribute-values of each attribute m ; the answers are represented by for-
mal contexts §m, called conceptual scales, which give the attribute values
a dyadic understanding:
Definition 5. A scaled context of a many-valued context lK := (G, M ,

W, I) is a pair (IK; TimEM §m) where the §m := (Gm , Mm , Im) are concep-
tual scales with m( G) ~ Gm and TI is a "product operator" aggregat-
ing the conceptual scales to a formal context (X mEM Gm, N , J) . The
scaled context (IK; TimEM §m) yields the derived context (G, N , ]) with
g]n : {::} (m(g))mEMJn . The (mostly used) method of plain scaling em-
ploys the semiproduct Z as product operator where N := UmEM {m} x Mm
and (gm)m EMJ(m, n) : {::} m(gm) = n; the (less used) product scaling em-
ploys the direct product where N := X mEM Mm and (gm)mEM J(nm)mEM :
{::} 3m E M : m(gm) = nm. 0
In Fig.4 the data table is shown from which the formal context in Fig.2 was
derived by conceptual scaling. In the first step the logical understanding
of the data table had to be mathematically abstracted to a many-valued
context. Since, in this case, the purpose was to support customers in buy-
ing a sleeping bag, the abstraction had also to make a shift from the more
technical terminology to a more customer-friendly language for which the
attributes "cheap", "medium-priced", "expensive" , "downy feather", "syn-
thetic fibre", "good", "acceptable", and "bad" were chosen. In Fig.5 the
definition of those attributes are given in terms of formal logic which con-
stitute the desired many-valued context and leads in this way to the formal
context in Fig.2 by plain scaling with ordinal scales.
An ordinal scale is a formal context determined by an ordered set P :=

(P, '5,). Out of the great variety of different types of ordinal scales (see
466 R. Wille
producer pd temperature t weight w price p filling material f
One Kilo Bag Wolfskin 7° c 940 g 149,- Liteloft
Sund Kodiak 3° c 1880 g 139,- hollow fibre
Kompakt Basic Ajungilak 0° c 1280 g 249,- MTI Loft
Finmark Tour Finmark 0° c 1750 g 179,- hollow fibre
Interlight Lyx Caravan 0° c 1900 g 239,- Thermolite
Kompakt Ajungilak -3° c 1490 g 299,- MTI Loft
Touch the Cloud Wolfskin -3° c 1550 g 299,- Liteloft
Cat's Meow The Nort h Face -7° c 1450 g 339,- Polarguard
Igloo Super Ajungilak -7° c 2060 g 279,- Terraloft
Donna Ajungilak -7° c 1850 g 349,- MTI Loft
Tyin Ajungilak -15° c 2100g 399,- Ultraloft
Travellers Dream Yeti 3° c 970 g 379,- goose down
Yeti light Yeti 3° c 800 g 349 ,- goose down
Climber Finmark -3° c 1690 g 329,- dug down
Viking Warm peace -3° c 1200 g 369,- goose down
Eiger Yeti -3° c 1500 g 419,- goose down
Climber light Finmark -7° c 1380 g 349,- goose down
Co bra Ajungilak -7° c 1460 g 449,- dug down
Cobra Comfort Ajungilak -10° c 1820 g 549,- dug down
Foxfire The North Face -10° c 1390 g 669,- goose down
Mont Blanc Yeti -15° c 1800 g 549,- goose down
Fig. 4: Many-valued context about sleeping bags
[GW99], Section 1.4), in this paper we only concentrate on two types which
are basic for the mathematical theory of ordered sets: the general ordinal
scale ([])£ := (P, P, :S:) and the general contraordinal scale ([JPJ!. := (P, P, l).
.0_
cheap (P, :S. 250)
A
medium-priced (P, > . 2501\ :S. 400)
A
expens1ve - (P, >. 400)
.0_
downy feather (F, a. goose down)
.0_
synthetic fibre (F, -, a. goose down)
good
.0_
((T, >. 0 1\ :S. 7) 1\ (G, :S .1000)) v
((T, >. -7 A:S.O)I\(G,:S.1400)) v
((T, > . - 15 1\ :S. -7) 1\ (G, :S .1700)) v
(T, :S. -15) 1\ (G, :S. 2000))
.0_
acceptable ((T, > .01\ ::;.7)A(G,:S.1400)) V
( (T, > . - 7 1\ :S . 0) 1\ (G, :S . 1700)) v
((T, >. - 151\ :::;. -7) 1\ (G,:S.2000)) v
(T, :S. -15)
bad -A ((T, >.01\ :::;. 7) 1\ (G, > .1400)) v
( (T, > . - 7 1\ :S. 0) 1\ (G, >. 1700)) v
((T, > . -151\ :::;. -7) 1\ (G,>.2000))
Fig. 5: Definitions of customer-friendly attributes for the data context

in Fig.2
The following theorem on general ordinal scales yields an important dyadic

understanding of ordered sets:
Dedekind-MacNeille-Completion Theorem. For an ordered set (P, :S

), t,X := ((x], [x)) for x E P defines an order embedding " of (P, :S) into
'B(P, P, :S); moreover, "V X= V~,x resp. "A X= A~,x if the supremum
resp. infimum exists in (P, :S). If K, is an ar-bitrary order embedding of
(P, :S) into a complete lattice L, then there is always an or·der embedding A
of 'B(P, P, :S) into L with K, = A o "·
Biedermann's Supplement. [Bi98a] For an order embedding /'1, of (P, :S)

into a complete lattice L, the subset L"" := U{[V K,A, A K,B] I (A, B) E
'B(P, P, :S)} is a complete lattice with r-espect to the order induced by the
order· of L; mo·reover, the map 1r"" : L"" -t 'B(P, P, :S) with 1r"" x := (A, B)
468 R. Wille
for x E [V K;A, 1\ K;B] is a surjective complete lattice homomorphism, z. e.

Ll'i) kenr"' ~ '13 (P, P, ~) , with i = 1r"' o K;.
Both, Theorem and Supplement, show that the concept lattice of the gen-
eral ordinal scale ([Jl.E is (up to isomorphism) the smallest complete lattice
into which the ordered set P := (P, ~) has an order embedding. This in
particular underlines the importance of the Galois connection given by the
derivation operators of ([Jl.E:
X xt {yEP I x ~ y for all x EX},

y y+ {x E P I x ~ y for all y E Y}.
The formal context JK(P) := (J(P) , J(P) , .6.) defined in Section 2 can be
viewed as a modified version of the general ordinal scale ([Jl.E which, in the
finite case, is even isomorphic to ([Jl.E.
The dyadic view of mathematical theories does not only give closer links
to logical thinking, it may also substantially enrich their structure the-
ory. In the scope of mathematical order theory this may become clear
by considering, for each ordered set P, the corresponding standard context
IKo (P) := (Jo(P), Jo(P) , .6.) which can be understood as an analogue to the
spectrum of a commutative ring with 1. Since '13 (JK( P)) ~ '13 (!Ko (P)) by
Theorem 1, P can still be represented by its standard context. This has
been elaborated in the form of topological representation theorems for 0-1-
lattices P including the special cases of distributive and Boolean 0-1-lattices
[Ha92]; recently, this has been generalized to 0-1-lattices with operators by
M. Gehrke and J. Harding [GH01]. Here the usefulness of the standard
context shall only be demonstrated by results about the order dimension.
Definition 6. For an ordered set P := (P, ~ ) , the order dimension dim(P)

is defined as the smallest cardinal number a so that P has an order em-
bedding into the direct product of a-many chains. A Ferrers relation is a
relation F between sets G and M (i.e. F ~ G x M) satisfying (g, m) E F ,
(h, n) E F, (g, n) rtF=> (h, m) E F. For a formal context lK := (G, M, I) ,
the Ferrers dimension f dim(JK) is defined as the smallest cardinal number
(3 so that I is the intersection of {3-many Ferrers relations between G and
M.
Theorem 3. [Wi89] The Ferrers dimension of a formal context lK is equal

to the order dimension of the concept lattice of JK, z. e.
fdim( JK) = dim('13(JK)).

As a direct consequence of Theorem 3 and the Dedekind-MacNeille-Comple-

tion Theorem for ordered sets (P, ~) , we obtain the equation dim(P, ~) =
dim(P, P, ~). Since a finite lattice L is always isomorphic to the concept
lattice of its standard context (J(L) , M(L) , ~) by the Basic Theorem on
Concept Lattices, we also have the equation
dim(P) = fdim(J(SB(fflp
-
)) , M(SB(fflp
-
)) , ~)
for each finite ordered set P. Theorem 1 indicates that this equation can
be obtained even for certain infinite ordered sets:
Theorem 4. Let P be an ordered set with finite order dimension . Then
dim(P) = fdim('Ko(P)) = dim(SJ3(0Co(P))).

Proof We first prove that fdim('Ko(P)) = fdim(Op_). Let the incidence
relation ~ of the general ordinal scale Op_ be the intersection of finitely
many Ferrers relations R1, ... , R, on P x P and let F b.ii : {::} :3x E
F3y E I : xRiY for i = 1, . . . , n , F E ~o(P), and I E Jo(P) . Then each
b.i is a Ferrers relation between Jo(P) and Jo(P) (obviously satisfying
b. ~ b.i) because, forF1b.ih and F2b.ii2, there are x1 E F1 , Yl E h ,
X2 E F2, and Y2 E h with X1~Y1 and X2~Y2 so that X1~Y2 or X2~Y1
and hence F1b.ih or F2b.ih· Now, let (F, I) E b.1 n · · · n b.n. Then
there exist Xi E F and Yi E I with Xi~Yi for i = 1, . .. , n. Since F
is a filter and I is an ideal, there are x E F with x ::; Xi and y E I
with y 2:: Yi for i = 1, ... , n and therefore xRiY for i = 1, ... , n which
implies X ~ y and hence F b.I, i.e. fj. = fj.l n ... n b.n. This proves that
OCo (P) has the same Ferrers dimension as fflp. Finally, we conclude that
dim(P) = fdim(Op_) = fdim('Ko(P)) = dim(SB('Ko(P))). D
The order dimension results can be succesfully applied to the task of establi-
shing well readable line diagrams of concept lattices which are essential for
reaching valuable interpretations of empirical data represented by formal
contexts. In this paper such applications can only be explained through
an example, the data context (G, M,I) of which is presented in Fig.6. The
Ferrers dimension of that context is 3 which is indicated by the numbers 1,
2, and 3 filled into the empty cells to represent the complements of the three
corresponding Ferrers relations R1 , R2 , and R3 , respectively. The three cells
(2, c), (4, b), and (6, g) show that two Ferrers relations do not suffice. The
concept lattices Li := SB(G, M, Ri) (i = 1, 2, 3) which are chains of length 5,
3, and 2 are depicted in Fig.7. An order embedding of the concept lattice of
470 R. Wille
II a b c d e f g h
1 Leech X X 1 1 1 1 X 1 1
2 Bream X X 1 1 1 1 X X 1
3 Frog X X X 1 1 1 X X 1
4 Dog X 2 X 1 1 1 X X X
5 Spike- weed X X 3 X 3 X 3 3 3
6 Reed X X X X 3 X 3 3 3
7 Bean X 2 X X X 1 2 2 2
8 Maize X 2 X X 2 X 2 2 2
Fig. 6: Context of an educational film "Living Beings and Water" .

The attributes are: a: needs water to live, b: lives in water,
c: lives on land, d: needs chlorophyll to produce food , e: two
seed leaves, f: one seed leaf, g: can move around , h: has limbs,
i: suckles its offspring.
(G, M, I) into L1 X L2 x L3 is given by i(A, B) := 1\mEB(ttlm, P,2m, p,3m) for

(A, B) E P,( G, M, I). This order embedding determines the line diagram
of P,(G, M, I) in Fig.8.
a,b,g
1
h
2
c a,c,d,f
3 8
4
e
7 a,b,d,:!:
d,e g,h,i
7 4
f b e,g,h ,i
5,6,8 1,2,3,5,6 1,2,3,4,7,8
Fig. 7: Concept lattices of the three Ferrers relations indicated in Fig.6
Now, we consider the second basic type of ordinal scales: the general con-
traordinal scales !!Jfl'_ ·- (P, P, l) based on ordered sets P ·- (P, :::;) .
Fig. 8: Line diagram constructed by a minimal grid embedding
These scales yield the important Galois connection given by the derivation
operators of OE. :
X xt {y E P I x '1. y for all x E X},

y yt {x E P I x '1. y for all y E Y}.
Concatenating those operators leads to the closure operators -!. and t defined
by .j..X := {y E P I y ::; x for some x E X} and t X := {y E P I y 2::
x for some x EX}; a subset X of Pis called an order ideal if X= .j..X and
an order· filter if X= tX. The formal context JKcd(P) := (J(P) , J(P), ~cd)
with F .6. cd I : ¢::> I n F = 0 can be viewed as a modified version of the
general contraordinal scale orl_0
Birkhoff's Theorem on Completely Distributive Concept

Lattices. For an ordered set P ·- (P, ::;) , Lv(x) := ((x] , P \ (x]) and
472 R. Wille
~A(x) := (P \ [x), [x)) for x E P define order embeddings ~ v : P ----7

23(0Ct) and ~A : P ----7 23((()fj1) where ~v (P) = J(23(0Ct)) is V-dense
and --;A(P) = M(23(0Cj)) is 7\-dense. The extents resp-:intents of oct
are exactly the order Ideals resp. order filters of P, and the extent and
the intent of a formal concept of OCj_ are always complements of each
other. Finally, 23(0Ct) is always completely distributive and, conversely,
for each completely distributive lattice L, i(y) := (J(L) n (y] , J(L) \ (y])
and~ (y) := (M(L) \ [y) , M(L) n [y)) for y E L define lattice isomorphisms
i: L ----7 23(0(~(£),:::;))) and~: L ----7 23(0(~(£) ,:::;))).
Supplement. For an ordered set P := (P, ::;) , ~cd(Sl, 23) :=

(USl,U23) for (Qt,23) E 23(JKcd(P)) defines an isomorphism ~cd from
23 (JKcd (P)) onto 23(0Cj_).
Proof For (Sl, 23) E 23(JKcd(P)) we have InF = 0 for all IE Qt and FE 23,
hence U Qt n U 23 = 0. U mis, as a union of ideals, an order ideal and U 23
is, as a union of filters , an order filter. Furthermore, (U Sl) U (U 23) = P
because otherwise there would be an a E P with a rj_ U Sl and a rj_ U 23 ,
i.e. (a] rj_ Qt and (a]~cd B for all BE 23, which would contradict Qt = 23~cd.
Thus, (U m, U 23) is a formal concept of oct.
Conversely, for (A , B) E
23(0Ci_), m:={IE J(P) I I~ A} and 23 :=[FE ~(P) I F ~ B} form the
formal concept (Qt,23) of occd(P) with (USl, U23) = ~cd(Qt , 23) =(A, B).
This proves that ~cd is a bijection. Since ~cd and its inverse are obviously
order-preserving, ~ cd is an isomorphism from 23 (JKcd (P)) onto 23 (OCj_). D
By Birkhoff's Theorem, 23(0Cj_) is (up to isomorphsm) uniquely determined

by the ordered set (M(23(0Cj_), ::;) (~ P) and therefore uniquely determined
by the property that the attribute implications A ----7 B (:{:?- At ~ st) of
oct have a basis of attribute implications with one-element premise. This
explains - because everyday thinking predominately uses implications with
one-element premises -, why contraordinal scales and their modifications
are so important in data analysis. A frequently used type of modified
contraordinal scales are the so-called simply implicational standard scales
defined as follows (cf. [SWWOl J, [Wi03a]):
Definition 7. An ordered set with inconsistent subsets is defined as a set

structure M := (M, ::; , Qt) where (M, ::;) is an ordered set and Qt is a set
of non-empty subsets of M having no lower bound in (M, ::;) . In the finite
case we consider, for any p E M, the set Jp of all minimal order ideals D of
(M, ~) with p ED and AnD =f. 0 for all A E 2t; let JM := UpEM Jp. Then
the finite formal context ((J)M := (J M , M, ;i) is called the (object reduced)
simply implicational standard scale derived from M. 0
Theorem 5. The intents of ((J)M are exactly the order filters of (M , ~)
equal to M or not containing any A E 2t. The set J(((J)M) \ {M} of all
proper intents is therefore an order ideal I of the distributive lattice of all
order filters of (M, ~) which satisfy -!F = 0 for all order filters F not in I.
Conversely, each such order ideal may be represented by the proper intents
of a simply implicational standard scale.
Proof For D E JM we have the object intent D;6 = M \D. Thus, the
intents of ((J)M are just the sets M \ UtET Dt with Dt E J M ( t E T) , which
are obviously order filters of (M, ~) equal to M or not containing any
A E 2t. Now, let F be any such order filter unequal M. Then M \ F is
an order ideal of (M, ~) with An (M \F) =f. 0 for all A E 2t. For each
p tf_ F we choose a minimal order ideal Dp in M \ F with p E Dp and
A n Dp =f. 0 for all A E 2t; then F = M \ Up EM\ F Dp. Thus, the intents
of ((J)M are exactly the order filters of (M, ~) equal to M or not containing
any A E 2t. Clearly, J(((J)M) \ {M} is an order ideal of the lattice F(M, ~)
of all order filters of the finite ordered set (M, ~) . Conversely, let I be an
order ideal of :F(M, ;S) satisfying ..1-F = 0 for all order filters F not in I,
and let 2li := {FE F(M, ~) \ {0} I F tf_ I} . Then M := (M , ~ ' 2li) is an
ordered set with inconsistent subsets and I= J(((J)M) \ {M}. D
The frequent occurrence in everyday human thought of attribute implica-

tions with one-element premise combined with attribute inconsistencies is
witnessed by the dominantly used conceptual scales isomorphic to simply
implicational standard scales in the practice of Formal Concept Analysis.
Among the scale types listed in [GW99] on page 57, the following scales
are isomorphic to simply implicational standard scales: one-dimensional
nominal, ordinal and interordinal scales, biordinal and multiordinal scales,
grid scales, (multiple) dichotomic scales, contranominal and contraordinal
scales.
An advantage of simply implicational standard scales is that an appropriate

line diagram of their concept lattices may be algorithmically established:
Let L be such a lattice with lL as its greatest element , and let M(L)
be the set of its /\-irreducible elements. First we determine a (minimal)
474 R. Wille
1\
chain decomposition C1, ... , Cn of M(L) and extend each chain Ck to Ck:=
C k U { 1L} (k = 1, ... , n) . Then a 1--7 ( c1 (a) , . .. , Cn (a)) , where Ck (a) is the
1\
smallest upper bound of a in Ck (k = 1, ... , n) , yields an isomorphism L
1\ 1\
from L \ {OL} onto an order filter of C 1 x · · · x Cn- Now, t(L \ {OL}) can
be drawn as an upper part of a grid which represents the direct product
1\ 1\ 1\
C1 x · · · x Cn where the chains Ck are located on straight lines. Finally,
the node representing OL has to be added properly {cf. [LSW97]). Fig.9
shows a concept lattice drawn by the described method {cf. [GW99], p.23).
Fig. 9: Line diagram of the concept lattice of a simply implicational

scale which essentially determines a satisfactory line diagram
of the concept lattice of the context in Fig.6 (cf. [GW99), p.23)
The dyadic view of mathematical order theory has been extended to a

dyadic view of mathematical logic [Wi96]. To keep the tight connection to
human thought, the development of a dyadic logic aims at the contextual
mathematization of the traditional philosophical logic which is based on
"the theory of the three essential main functions of thinking - concepts, judg-
ments, and conclusions" ([Ka88], p.6). The dyadic-logic theory of concepts
takes over the mathematical approach of Formal Concept Analysis. Since
judgments are logically understood as asserting combinations of concepts,
the dyadic-logic theory of judgments is built on so-called concept graphs in
which formal concepts are relationally combined. The dyadic-logic theory
of conclusions treats entailments between those concept graphs. The three
theories are semantically based on formal contexts. This is the reason why
the mathematical logic determined by the described dyadic view is called
"Contextual Logic" (cf. [WiOOa]). In this paper, we shall only discuss the
dyadic nature of concepts graphs which arise from the contextual math-
ematization of Sowa's conceptual graphs [So84]:
Definition 8.
(1) A power context family is a sequence i. := (!Ko, JKI , lK2, . . . ) of formal
contexts~:= (Gk,Mk,h) with Gk c:;;; (Go)k fork= 1,2, .... The formal
concepts of~ with k = 1, 2, . . . are called relation concepts, because they
represent k-ary relations on the object set Go by their extents.
(2) A relational graph is a structure (V, E , z;) consisting of two sets V and
E and a mapping v: E-+ Uk=I ,2 , ... Vk; the elements of VandE are called
vertices and edges, respectively, and v( e) = (VI, .. . , vk) is read: VI, ... , Vk
are the adjacent vertices of the k-ary edge e (lei := k is the arity of e; the
arity of a vertex is defined to be 0). Let E (k) be t he set of all elements of
VUE of arity k (k = 0, 1, 2, .. . ).
(3) A concept graph of a power context family i. := (!Ko, JKI, lK2, .. . ) with
~ := (Gb Mk, Ik) for k = 0, 1, 2, ... is a structure Q5 := (V, E , v, /'\,'e) for
which
- (V, E, v) is a relational graph,
- /'\,:VUE-+ Uk= O,I,2 , ... 1B(~) is a mapping such that 1'\,(u) E IB(~)
for all u E E(k),
- e: V-+ s.}J(Go)\{0} is a mapping such that e(v) c:;;; Ext(/'\,(v)) for all
vE vand, furthermore, e(vl) X ... X e(vk) c:;;; Ext(/'\,( e)) for all e E E
with v( e) = (VI, ... , Vk); in general, E xt( c) denotes the extent of the
formal concept c.
It is convenient to consider the mapping e not only on vertices but also on

edges: For all e E E with z;(e) = (vi, ... ' vk), let e(e) := e(vl ) X . .. X e(vk)·
0
As an example of a (small) power context family [Wi03b], Empedocles'
doctrine of the four basic elements is formally represented in Fig.10 by
contexts IKo and IK2 ; these contexts are intensionally determined by the
two oppositions cold+-+warm and dry+-+moist (JKI might be considered to
be the empty context (0, 0, 0)) ; Fig.ll shows the concept lattices of the
formal contexts in Fig.10. Examples of concept graphs of the power context
family represented in Fig.lO are depicted in F ig.12: The two concept graphs
together inform that the element water, belonging to the concept moisture,
is allied with the element air, belonging to the concept heat, and opposite
to the element fire, belonging to the concept heat too.
476 R. Wille
D
2
1j
<J)
·en0
::&
I
Ill 0
(water,earth)
(water,air)
uwaier AIX IX (water,fire)
arth X X IX (earth,air)
r X XIX (earth,fire)
re XXIX (air,fire)
Fig. 10: A power context family (lKo, IK2 ) representing Empedocles'

doctrine of the four basic elements
Fig. 11: The concept lattices of the formal contexts in Figure 1
A concept graph lB := (V, E, v, ,.,, g) of a power context family i. informs

on the elementary level that the object g belongs to the extent Ext(,.,(u))
for all u E E(k), g E g(u) , and k = 0, 1,2, ... where each ,.,(u) is a formal
concept of the corresponding formal context ~ . This indicates the dyadic
nature of concept graphs on the elementary level.
But the elementary information units give rise to further information by
Imoisture: water H allied H heat: air I

Imoisture: water H opposite H heat: fire I
Fig. 12: Two concept graphs of the power context family in Fig.lO
background implications coded in the power context family. This can be

made mathematically explicit as follows (cf. [Wi02b]): Let k = 0, 1, 2, ... ;
for A, C ~ Gk, lKA; satisfies A -t C if Ah ~ Ch and, for 123, 1) ~ 123(1KA; ),
lKA; satisfies 123 -t 1) if 1\ 123 :::; 1\ 1). The formal implications A -t C
and 123 -t 1) give rise to a closure system C(lKA;) on §imp(lKA:) := {(g, b) E
Gk x 123(1KA;) I g E Ext(b)} consisting of all subsets Y of §imp(lKA;) which
have the following property:
(Pk) If Ax 123 ~ Y and if lKA; satifies A-t C and 123 -t 1) then

c X 1) ~ Y .
Now we are able to define the "conceptual content" of a concept graph of
a power context family as the full information given by the graph:
Definition 9. For a power context family IK, the k-ary conceptual content
Ck(IB) (also called the lKA; -conceptual content) of a concept graph IB :=
(V, E , v, K, e) of IK is defined as the closure of {(g, K(u)) I u E E(k ) and g E
e(u)} with respect to the closure system C(lKA;) (k = 0, 1, 2, ... ). Then the
disjoint union
C( IB) := Co( ~B) u CI (~B) U C2( ~B ) U ...
is called the (IK-) conceptual content of the concept graph IB.
The conceptual contents give rise to an information quasi-order ;S on the

set of all concept graphs of a power context family IK: A concept graph
IB1 := (VI, E1 , vi, lil , l?l) is said to be less informative (more general) than a
concept graph IB2 := (V2 , E2 , v2 , K2, l?2) (in symbols: IB1 ;S IB2) if Ck(IB1) ~
Ck(IB2) for k = 0, 1, 2, ... ; IB1 and IB2 are called equivalent (in symbols:
IBI "' IB2) if IBI ;S IB2 and IB2 ;S IBI (i.e., Ck(IBI) = Ck(IB2) for k =
0, 1, 2, ... ). The set of all equivalence classes of concept graphs of a power
context family IK together with the order induced by the quasi-order ;S is
an ordered set denoted by i\oc). The ordered set r(IK) is even a complete
lattice as shown in [Wi02b].
Since every complete lattice is isomorphic to a concept lattice by the Basic

Theorem on Concept Lattices, the question arises whether there is a nat-
urally described formal context so that its concept lattice is isomorphic to
r(i) or, more concretely, the conceptual contents of the concept graphs of
i are specified as extents of that context. This would allow to understand
formal judgments dyadically as formal concepts and formal conclusions as
inferences between formal concepts. An affirmative answer to the stated
question is sketched in the rest of this section.
478 R. Wille
Definition 10. For a context lK. := {G, M, I), let §imP(JK.) := { (g, b) E G x
SB(JK.) I g E Ext{b)} and let scon(SB(JK.)) be the set of all convex subsets 6
ofSB{lK.)\ { (M 1 , M)} for which 6U{ (M 1 , M)} is a complete sublattice of the
interval
[(M 1 , M) , V 6] of IB{JK.). A conceptual information context corresponding
to lK. is defined by
JK:inf (K) := (§imP(JK.), 6 con {IB (JK.) ), ~)
with (g , b)~S: ¢} ['yg , b] nS =!= 0. 0

Now, we can formulate the main result about the conceptual contents of
formal contexts lK. (which are understood as the conceptual contents of the
concept graphs of a power context family having at the most IKo := lK. as a
non-empty context):
Basic Theorem on JK.-Conceptual Contents. (Wi03b] For a formal

context lK. with 011 = 0, the extents of the corresponding conceptual infor-
mation context JK:inf (JK.) are exactly the conceptual contents of K
Fig.13 shows the concept lattice of the conceptual information context

JK:inf (!Ko) derived from the formal context IKo of our example. According
to the Basic Theorem on JK.-Conceptual Contents, the conceptual contents
of IKo are represented as the extents of the depicted concept lattice.
The Basic Theorem on JK.-Conceptual Contents can be extended to the
general case of {limited) power context families. Let lK := (!Ko, JK.1 , ... ,
lKn) be a power context family with IKA; := (Gk,Mk,h) (k = 0, 1, ... , n).
The conceptual information context corresponding to lK is defined as the
formal context
JK!nf {JK) ·- JK.inf {JKo) + JK.inf {JK.l) + ... + JK.inf (JKn)
·n ·n ·n ·n
·- (uk=Oak , uk=OMk, uk=Oh uu(j#)=Oaj x Mk)-
An extent U of JK:inf (K) is said to be rooted if ((91, ... , gk) , bk) E U always
implies (gj , To) E U for j = 1 ... , k and To:= {Go,G~0 ). Now we are able
to state the desired theorem:
Basic Theorem on JK-Conceptual Contents. (Wi03b]

For a (limited) power context family lK with 0hh = 0 for k = 0, 1, 2, ... ,
the conceptual contents of the concept graphs of lK are exactly the rooted
extents of the cor-responding conceptual information context
JK:inf {JK).
Fig. 13: Line diagram of the concept lattice of the conceptual infor-
mation context Kinf (IKo) having as extents the conceptual
contents of the formal context IKo in Fig.lO
4 Dyadic Algebra and Geometry
Modern algebra is mainly understood as the structure theory of sets with

operations. It assists humans with a language for operational descriptions
of forms of thought, in particular of concepts and concept systems (cf.
[Wi81],[Wi01a]). Therefore, it seems appropriate to start the development
of dyadic algebra on the basis of formal contexts and concepts so that the
Galois connection between formal objects and attributes plays a basic role.
In [MMT87] it is stated that "the most basic Galois connection in algebra is

the one which is associated to the binary relation of preservation between
operations and relations" (p.l47). A general framework for this type of
Galois connection is given for arbitrary sets A with their sets Op(A) of all
480 R. Wille
operations on A and Rel(A) of all relations on A by the formal contexts

oc0 R (A) := ( Op(A), Rel(A) , l>) with
f l> (} : {::} (!(au , .. . , aln) , ... , f(aml , ... , amn)) E (}

for all (ali , ... , ami) E (} (i = l , ... ,n).
The corresponding Galois connection is constituted by the derivation op-

erators of oc0 R(A) as follows (F ~ Op(A) and Q ~ Rel(A)):
F ~--t pt> := {Q E Rel(A) If l> Q for all f E F} =: InvAF ,

Q I--t Q<J := {f E Op(A) I f l> Q for all Q E Q} =: PolAQ .
It is common to consider Op(A) as an algebra with the projection and

composition operations and Rel(A) as an algebra with the diagonal, in-
tersection, product, projection, and permutation operations (cf. [Po03]) .
These algebras allow useful algebraic descriptions of the extents and the
intents of oc0 R(A) with finite base set A:
Theorem 6. (cf. [PK79]) Let A be a finite set. Then the extents of the
operation-relation-context oc0 R(A) are exactly the subalgebras of the opera-
tional algebra Op(A), which are called the operational clones on A, and the
intents of the operation-relation-context OC0 R (A) are exactly the subalgebras
of the relational algebra Rel(A) , which are called the relational clones on
A.
Since the term operations of every (finite) algebra A := (A, F) form an

operational clone on A , the "complete, in-depth presentation of basic uni-
versal algebra" in [MMT87] starts with the definition of (operational) clones
of term operations. The dyadic view would even suggest to complete such
an approach by activating the corresponding relational clones to support a
more holistic understanding via formal concepts viewed as clones with oper-
ational clones as extents and relational clones as intents. This would allow
a flexible use of the two opposite types of descriptions, firstly by generating
subsets w.r.t. algebraic operations and secondly by specifying subsets w.r.t.
derivation operators.
The two basic types of descriptions may become more apparent by consid-
ering the important algebraic notion of an equational class. Let <.!.:17 be the
class of all algebras of similarity type a and let E 17 (X) be the set of all a-
equations over the (countable) set X of variables. Then the dyadic nature
of the notion of an equational class can be explicated by the formal context
OC( <.!.:17 ; X) := ( <.!.:17 , E 17 (X) , f=) with A f= p ~ q : {::} the equation p ~ q is
valid in the algebra A. The corresponding Galois connection is given by

the derivation operators of JK(([a-;X) as follows (Qt ~ ([a- and R ~ Ea-(X)):
Q{ t---+ mF= ·- {p ~ q E Ea-(X) I A F p ~ q for all A Em}

-· Id(~Jt) ,
R t---+ Rf= ·- {A E ([a- I A != p ~ q for all p ~ q E R}
-· Mod(R).
Birkhoff's Theorem on Equational Classes. (cf. [llu98]) The extents

of the algebra-equation-context JK(C[a-; X) are exactly the subvarieties of ([a- ,
i. c. the subclasses closed under· the formation of direct products (P), subal-
gebras (S), and homomorphic images (H), and the intents are exactly the
equational subtheories of Ea-(X) , i.e. the subsets closed under the deduction
rules of r·efiexivity, symmetry, transitivity, substitution, and replacement.
The dyadic view suggests to holistically understand the equational classes

a.'3 formal concepts of the algebra-equation-context JK( ([a-; X) with the va-
rieties of type () as extents and the equational theories of type () over the
variable set X as intents. The two types of descriptions present varieties
first as compositions of the operators P, S, and H applied to generating
classes of algebras, and secondly as derivations of specifying sets of equa-
tions. The two types of descriptions present equational classes firstly as
schernas of deduction rules applied to generating sets of equations and sec-
ondly as derivations of specifying classes of algebras.
Another useful idea is to view a concept lattice as a representation of the two

mutually inverse antiisomorphisms between the closure system of the ex-
tents and the closure system of the intents where the extent-intent-pairs of
the formal concepts represent the preirnagc-image-pairs. For instance, the
concept lattice of the operation-relation-context (algebra-equation-context)
yield the mutually inverse antiisomorphism between the closure system
of the operational clones (varieties) and the closure system of the rela-
tional clones (equational theories). Using such a dyadic view, an inter-
esting discovery has been made even in elementary linear algebra by con-
sidering the formal contexts OCr (V) := (V, V*, ..lr) where V is a finite-
dimensional vector space over a field F , V * is the dual space of V , r E F
and v ..lr <p: ? rp(v) = r . In the case r = 0, the concept lattice of ~1 (V)
represents the well-known antiisomorphisrns between the subspace lattices
of V and V * . In the case r -1- 0, an analogous result could be proved for
affine subspaces (cf. [Wi02c]):
482 R. Wille
Theorem 7. [Wl91] The concept lattice of the vector-space-context lKr (V)

with r =I 0 represents mutually inverse antiisomorphisms between the lat-
tices consisting of the total space of V resp. V* and of all affine subspaces
not containing zero; in particular, if V = JR.n = V*, v ..lr w : {::} v · w = r,
and r > 0, the concept lattice of lKr (V) represent the well-known inversion
in the hypersphere of radius .Jr and center 0 in the euclidian space JR.n.
Proof Obviously, for r =I 0, the derivations c.p j_r of the linear forms j_r = n c.pl. r, the extents of lKr (V) are exactly the affine
subspaces of V not containing 0 or equal V. The dual statement for the
intents can be dually proved. Now, the first claim of the theorem follows.
In the case V = JR.n = V* , v ..lr w : {::} v · w = r , and r > 0, the elements
u E JR.n with u E uj_r form the hypershere of radius .Jr and center 0 because
of u E uj_r {::} u · u = r {::} lui = .Jr. The derivation uj_r which contain
u is the tangential hyperplane on the hypersphere in u since u , v E u j_r
implies u · (v- u) = u · v- u ·u = r - r = 0. Furthermore we know from the
first claim that U H Uj_r induces a (self-inverse) antiisomorphism between
the stated lattices of affine subspaces. Together we can conclude that the
antiisomorphism is the inversion in the hypersphere of radius .Jr and center
0. D
Further pairs of antiisomorphisms have been studied via suitably chosen for-
mal contexts for or·dered vector spaces [Wl99], finite abelian groups [Vo95],
[Gr99], modules [KV96], and algebraic geometry [Be99]. General investiga-
tions of bialgebraic contexts in which both context sets carry an algebraic
structure are presented in [Vo94] . The here cited research has been inspired
by "Ideas of Algebraic Concept Analysis" as outlined in [VW94] (see also
[Wi76], [WiOla]).
For further developments of dyadic algebra, the representation and gener-

ation of mappings and morphisms as formal concepts are interesting (cf.
[Xi93], [Xi94],[Ga03]). In particular, the logic of distributive systems as a
theory of information flow presented in (BS97] and the related categorial
theory of Chu spaces (cf. [Pr02]) can be viewed as such developments. Here
only an application in linear algebra shall be presented. For this, further
notions concerning many-valued contexts are discussed:
Definition 11. For a set W , a W-context is defined to be a (complete)
many-valued context (G, M, W , I) which is clarified, i.e., (Vm EM: m(g) =
m(h)) ::::} g =hand (Vg E G : m(g) = n(g)) ::::} m = n. In the case of W-

contexts, we usually write <g, m>= w instead of (g,m,w) E I. For awE
W, the corresponding constant object 9w is characterized by < 9w, m > = w
for all m E M and the corresponding constant attribute mw is characterized
by <g,mw>= w for all g E G. From a W-context lK := (G , M , W,I) each
wE W gives rise to a formal context lKw := (G,M,_iw) with g l_w m:
{::} <g,m>= w. For W-contexts (G1,M1 , W, h) and (G2 , M2 , W, h) , the
dually paired context (Gl X G2,M2 X MI,rv) is determined by (gl , g2) rv
(m2,m1): {::} <g1,m1 >=<g2 , m2>· 0
Lemma 2. Let (cp,cpd) E 113(GI x G2 , M2 x M1 , "'). Then cp : G1 ---+
G2 is a map if and only if cpd : M2 ---+ M1 is a map. If cp and cpd are
maps between finite-dimensional vector spaces and their corresponding dual
spaces, respectively, and if< , > denotes the natural bilinear forms between
these vector spaces and their duals, then cp and cpd are linear maps.
Proof Let cp: G1-+ G2 be a map, let (g1,g2) E cp, and let (m2 , mi),
(m2, m1) E cpd. It follows < 91, m1 >=< 92 , m2 >=< 91 , m1 > and hence
m1 = m1 because domcp = G1 and (G1 , M1 , W , h) is clarified. The dual
proof yields that cp : G 1 ---+ G2 is a map if cpd : M2 ---+ M 1 is a map. In the
case of vector spaces, let (gi,92),(hi , h2) E cp and let {m2 , mi) E cpd. Then
we obtain < g1, m1 >=< g2 , m2 > and < h1 , m1 >=< h2 , m2 > which imply
<gl +hi, ml >=<g2 +h2, m2 >. Since (cp, cpd) E 113(Gl X G2, M2 X Ml, rv) it
follows that (g1 + h1, 92 + h2) E cp, i.e. cp(gl + h1) = 92 + h2 = cp(gi) + cp(h1 ).
For a scalar r we obtain < rg1, m1 >= r < g1 , m1 >= r < 92 , m2 >=<
rg2, m2 > and hence (rg1, rg2) E cp, i.e. cp(rgi) = rg2 = rcp(gi). Thus, the
linearity of cp is proved. The linearity of cpd follows dually. 0
Lemma 3. Let (cp,cpd) E 113(GI x G2,M2 x M1,"') and let cp: G1---+ G2
be a map. If cp(GI)l_w -::/- 0 then mw E M1 and cp(GI)_Lw = (cpd) - 1(mw);
dually, if cpd(M2)1_w -::/- 0 then 9w E G2 and cpd(M2) 1_w = cp- 1(gw)·
Proof The existence of an attribute min cp(GI)_iw yields w =<cp(g), m>
=< g, cpd(m) > for all g E G which is equivalent to cpd(m) = mw E M1
since the underlying W -context is clarified. This equivalence is formally
expressed by the identity cp(GI)l_w = (cpd) - 1(mw)· The dual claim can be
dually proved. 0
Lemma 2 and 3 can be successfully applied to bilinear F-contexts (V, V* , F,

E) composed by a finite-dimensional vector space V over the field F , its
484 R. Wille
dual space V, and the ternary relationE with (v, <p, r) E E: {::} <p(v) = r.
As an example of such an application, a dyadic-algebraic proof is given for
the following well-known theorem:
Theorem 8. For each matrix over a field, its row rank zs equal to its
column rank.
Proof Let A be an m x n-matrix over the field F. We consider the F-

contexts (nF, pn, F, En) and (mF, pm, F, Em) where PF is the vector space
of all p-columns over F, FP is the vector space of all p-rows over F, and
(a,b,r) E Ep : {::} b ·a = r for p E {m, n}. The matrix A gives rise
to the linear maps <{JA and '1/JA defined by <fJA(a) := A· a (a E Fn)
and '1/JA(b) := b ·A (b E Fm). < <fJA(a) , b >= b · <fJA(a) = b · A· a =
'1/JA(b) ·a=< a,'l/JA(b) >indicates that '1/JA = <p1. By Lemma 3, we have
(Im<p)1° = (<pA(nF))l_ 0 = (<p1) - 1 (mo) = Ker<p1. Finally, it follows that
column-rank(A) = dim(Im<pA) = m-dim(Im<pA) l_ 0 = m-dimKer<p1 =
dimlm<p1 = row-rank(A). 0
That algebra serves humans with a language for operational descriptions

becomes particularly apparent in geometry. Since Graeco-Roman times,
algebra plays a significant role in geometric measurement which is aim-
ing at supporting human thought and action by making realities graphic,
intelligible, and workable (cf. [WW03]) . Geometric measurement is to a
large extent prototypical for mathematically representing realities which
is indicated by the comprehensive theory of representational measurement
presented in the three volumes "Foundations of measurement" [KLST71].
To make clear the significance of geometry for representational measure-

ment, U. Wille has proposed a four-level-approach of geometric measure-
ment in [W197] which is illustrated in Fig.l4. The first level identifies the
relevant empirical situations of reality. In a first step of mathematization
empirical situations are grasped by empirical models which are located on
a second level. In our dyadic approach, such models are represented as
ordinal contexts. The third level is the level of synthetic geometry with
geometric spaces as representing structures. The dyadic view of geometry
leads to formal contexts the extents of which are just the subspaces of the
representing geometric spaces. Such formal contexts are derived by concep-
tual scaling from the ordinal contexts of the second level. A representation
on the third level is often sufficient to obtain satisfying interpretations of the
empirical data, particularly by using line diagrams of concept lattices in the
analytic level numerical models
coordinatization
geometrical models
synthetic level
representation
empirical models
formal level
formalization
empirical ituations
of reality
reality level
Fig. 14: The four levels of geometric measurement
dyadic case. The transformation of the third level of geometric models to

the fourth level of numerical models (algebraic models) can be understood
as (dyadic) coordinatization. Algebraic descriptions of empirical relation-
ships are finally given on the fourth level. The direct representation of
empirical by numerical models which is proposed in [KLST71] is, of course,
an important task of measurement, but it misses the possibility of concep-
tual data interpretation via the concept lattices of the third level.
In the following , a four-level-approach of geometric measurement is sketched
which may finally lead to decriptions of empirical situations by ordered vec-
tor spaces. The first definition determines how the ordinal contexts of the
second level may be transformed to suitable "geometric" formal contexts
of the third level (cf. [WW03]):
Definition 12. Let ][( := (G , M, (Wm, S::m)mEM,I) be an ordinal con-
486 R. Wille
text. By interordinal scaling of lK, we derive the formal context ~o :=

(G,UmEM{m} X Wm X {~m,2::m} , J) withgJ(m , wm , ~m): <=> m(g) ~m
Wm and gJ(m, Wm, 2::m) : <=> m(g) 2::m Wmi ~ 0 shall be called an interor-
dinally derived context. For the attribute triples (m, w, ~m), (m, w, 2::m) E
UnEM{n} x Wn x {~n, 2::n} with m(g) = w for some g E G, we shortly
write (g]m for (m, w, ~m)J and (g)m for (m, w , 2::m)J. For each m E M,
the equivalence relation 8m on G is defined by its equivalence classes
[g]m := [g)m n (g]m(= {h E G I m(g) = m(h)}) (g E G); furthermore,
D. := nmEM 8m. lK and ~0 are object-clarified if D. is the identity on G.
0
Geometrically important ordinal contexts are the ordered bilinear contexts
JK(V) := (V, V*, (F, ~),E) for a finite-dimensional ordered vector space V
over a (partially) ordered field (F, ~) with (v, cp, r) E E : <=> cp(v) =
r ( cf. [Wl95]). As the interordinally derived formal context we obtain
~ 0 (V) := (V, V* x F x {~ , 2::} , J) with gJ(cp , r, ~) : <=> cp(g) ~rand
gJ(cp, r, 2::) : <=> cp(g) 2:: r. The extents (g]m and (g)m of ~ 0 (V) (g E G
and m E M) are exactly the closed halfspaces of V. This indicates that
the interordinally scaled contexts can be used as basis for a dyadic theory
of closed convex sets of ordered vector spaces (cf. [Wl99]).
Usually, empirical situations cannot be directly modelled within structures

like ordered bilinear contexts. Therefore, we approach the final representa-
tion in several steps where embeddings (as defined next) play an essential
role.
Definition 13. An object-value-embedding (-isomorphism) of a clarified or-

dinal context lK := (G,M,(Wm,~m)mEM,I) to a clarified ordinal context
1\ 1\ 1\ 1\ 1\
JK:= (G, M, (W m, ~m)mEM, I) is a family (w, (vm)mEM) of injective (bijec-
1\ 1\ 1\
tive) maps w: G -+G and Vm: Wm -+Wm with Vm ~m Wm <=> vm(vm) ~m
vm(wm) (mE M) such that (g, m, w) E I <=> (w(g), m,
1\
v(w)) EI. In general, a context embedding (isomorphism) of a formal con-
/\ 1\ 1\
text (G, M, I) to a formal context (G, M, I) is a pair (w , a) of injective (bi-
/\ 1\ 1\
jective) maps w: G -+G and a: M -+M such that gim <=> w(g) I a(m).
1\
A context embedding (isomorphism) (w , a) of ~o to lKio is called oTdi-
1\ 1\
nal if there exist order embeddings vi; : (Wm, ~m) ---+ (W m, ~ m) with
1\ 1\
a(m, Wm, ~m) = (m, vi;(wm), ~m) and a(m, Wm, 2::m) = (m, vi;(wm), 2::m)
formE M. 0
Lemma 4. (w, (vm)mEM) is an object-value-embedding (-isomorphism) of a

clarified ordinal context IK := ( G , M , (Wm , ~m)mE M , I) to a clarified ordinal
1\ 1\ 1\ 1\ 1\
context][(:= (G, M, (W m, -::;m)mEM, I) if and only if the derived pair (w , oP)
1\
of maps with av(m, Wm , -::;m) := (m, Vm(wm) , -::;m) and av(m , Wm, ?:m) :=
1\
(m, vm(wm), ?:m) is an ordinal context embedding (isomorphism) of the in-
/\
terordinally derived context ~o to the interordinally derived context IKio ·
Conversely, (w, a) is an ordinal context embedding (isomorphism) of ~o
1\
to IKio if and only if the family (w, (v~)mEM) is an object-value-embedding
1\
(-isomorphism) of JK: to K The notions "object-value-embedding " and "or-
dinal context embedding" are also used in the case that only the cardinalities
1\
of the attr·ibute sets of JK: and ][( are equal.
Representations of ordinal contexts resp. interordinally derived contexts in

ordered bilinear contexts (over the reals) resp. their interordinally derived
contexts require strong assumptions. Therefore, to cover a wider spectrum
of empirical situations, geometric measurement in weaker ordered algebraic
structures have been approached. The basic axioms of this approach are
given by the anti ordinal dependences between attributes mo, m1, ... , mn in
an ordinal context IK := (G , M , (Wm , ?:m)mEM,I) and between attribute
extents (g]mj and [g)mj in its interordinally derived context ~ 0 , respec-
tively (i = 0, 1, ... , n):
(Ai) Vg, h E G (Vj E {0, 1, ... , n} \ {i} : (g)mj <;;;; (h]mj)
===? (h]mi <;;;; (g]mi·
In addition, for the equivalence classes [g]ij := nnk = O [g]mk (i, j = 0, 1, ... ' n
i# #j
with i =/= j), we consider the following solvability conditions:
(Pij) Vg, hE G: [g]ij n [h]mi =/= 0.
Since the solvability conditions are essential for coordinatizing ordinal con-
texts, the following embedding theorem (cf. [WW93] , [WW96]) is important
for improving the synthetic geometrical models:
Theorem 9. (Embedding Theorem)
A clarified ordinal context JK: (interordinally derived context ~ 0 ) with the
attribute set M = {mo , ml ,···, mn} satisfying (Ai) (i = 0, 1, . .. , n) can
1\
always be embedded into a clarified ordinal context ][( (interordinally derived
1\ 1\ 1\
context ~ 0 ) with the same attribute set so that ][( (IKio) satisfies (Ai) and
(Pij)(i,j = 0, 1, ... , n with i =I= j).
488 R. Wille
As a by-product of the Embedding Theorem, we obtain that the solvability

conditions cannot be rejected by finite data in the class of interordinally
derived contexts with a set of n + 1 attributes satisfying the antiordinal
dependency axioms. Hence the (Pij) are only technical conditions which
can be added without destroying connections to the empirical models. The
resulting contexts can now be coordinatized by ordered n-loops defined as
follows (cf. [WW96]):
Definition 14. An OTdered n-loop is an ordered algebra L_ := (L , j, 0, s) for

which f is an order-preserving n-ary operation on (L, s) uniquely solvable
in each of its components always respecting the order and with 0 as neutral
element. For each ordered n-loop L_, there is a corresponding clarified
ordinal context IK(L_) := (Ln, {mo, m1 , ... , mn}, (L, Si)iE{l,... ,n}Jd with
:So:=2:, mo := J, S(=S and mi(ai, ... , an) := ai fori= 1, ... , n so that
h can be defined by ((a1, ... , an), mj , a) E h : {::} mj(al, ... , an) = a
for j = 0, 1, ... , n; furthermore, the corresponding interordinally derived
context
OC;o (L) := (Ln' u~=o{ mi} X {Si, 2:i}, h) is determined by

L X
(ai, ... ,an)h(mj,b,sj): {::} mj(a1, ... ,an) :Smj(ai , ... , an) sb,
(a1, ... ,an)h(mj,b, 2:j): {::} mj(al, ... , an) S mj(ai, ... ,an) 2: b.
0
Theorem 10. (Coordinatization Theorem) A clarified ordinal con-
I\
text ][( ( inteTOrdinally derived context OC; 0 ) with the attribute set M =
{ mo, m1, ... , mn} satisfying (Ai) and (Pij) (i, j = 0, 1, ... , n with i =I= j)
is isomorphic to the ordinal context IK(L_) (interordinally derived context
0C; 0 (L)) of a suitable ordered n-loop L_.
The Embedding Theorem and the Coordinatization Theorem together yield

a representation of empirical models of the second level (and the geometric
models of the third level) into algebraic models of the fourth level. Those
representations can be explicitly described which gives rise to a general rep-
resentation theorem clarifying the basic role of the antiordinal dependency
axioms (cf. [WW95]):
Theorem 11. (General Representation Theorem)[WW96] For a claT-

ified oTdinal context OC with the attribute set M = {mo , m1 , ... , mn} and its
interordinally deTived context OC; 0 , the following statements are equivalent:
1. There exists an object-value-embedding of OC to the ordinal context

IK(L_) of a suitable ordered n-loop L_.
Receptor Violet Blue Blue Blue-Green

430 458 485 498
1 147 153 89 57
2 153 154 110 75
3 145 152 125 100
4 99 101 122 140
5 46 85 103 127
6 73 78 85 121
7 14 2 46 52
8 44 65 77 73
9 87 59 58 52
10 60 27 23 24
11 0 0 40 39
Fig. 15: Colour absorption of 11 receptors in a goldfish retina
2. ]I( satisfies the antior-dinal dependency axioms (Ai) fori = 0, 1, ... , n.

3. There exists an or-dinal embedding of ~o to the interordinally derived
context ~ 0 (L_) of a suitable order-ed n-loop L_.
4. ~o satisfies the antiordinal dependency axioms (Ai)( i = 0, 1, ... , n).
The remaining question about the uniqueness of the embeddings of the

General Representation Theorem is answered in [WW95]. In [Wl96], an
extension of the General Representation Theorem is proved in the case of
n + l attributes m1, ... , mn, mn+l, ... , mn+l for which each attribute set
{m 1 , ... ,mn,mt} (t = 1, .. . ,1) satisfies then + 1 antiordinal dependency
axioms. The Coordinatization Theorem can be specialized to charaterize
those ordinal contexts (interordinally derived contexts) which have a rep-
resentation by ordered Abelian groups and by ordered vector spaces (over
the reals), respectively (see [Wl95],[WW96]).
Although the embeddings of ordinal contexts into bilinear contexts over the
reals are not finitely axiomatizable in first order logic [WlOO], the General
Representation Theorem can still be used for concrete data which shall fi-
nally be demonstrated by the example context presented in Fig.15. The
data context in this figure describes the amounts of absorption of four colour
stimuli by eleven receptors in a goldfish retina (cf. [SF68]). The data con-
text is viewed as an ordinal context IKcot whose integer values are ordered in
490 R. Wille
the natural way. For representing such an ordinal context in a real vector
space, by the General Representation Theorem, we have to determine the
antiordinal attribute dependencies of the ordinal context (cf. [WW96]). In
[GW86], it is shown that the ordinal dependencies of an ordinal context
lK. := ( G, M, (Wm, ~m)mE M , I) are exactly the attribute implications of the
formal context 0Ca := (G 2 , M,I0 ) with (g , h)I0 m : {::} m(g) ~m m(h).
Since an ordinal dependency m1 , ... , mn ~ m is defined by (mi (g) ~ mi
mi(h)for all i = 1, ... , n) =? m(g) ~m m(h) , for checking the conditions
(Ai), it helps to extent lK. to the ordinal context lK := (G , M(JMd , (Wn , ~n
)nEMl.JMd,llJJd) where Md := {md I mE M}, wmd := Wm , v ~md w:
{::} w ~m v, and Jd := {(g , md , w) I (g , m , w) E I}. Then we obtain
the dependencies described by the (Ai ) as implications of lKo which can
be read from the line diagram of the concept lattice of lKv . For our ex-
ample, this line diagram is depicted in Fig.l6 (cf. [WW96]). It shows
Fig. 16: Concept lattice of Ko
that the conditions (Ai) are satisfied by the ordinal contexts correspond-
ing to the attribute sets {Violet430, Blue458dual , Blue - Green498}
and {Violet430, Blue485dual , Blue- Green498}. Therefore, we have
two ordered-2-loop-representations which can even be simultanously repre-
sented in the Euclidean plane as shown in Fig.17 (cf. [Wi92]). An interesting
outcome is that the four colours are "ordered" in the figure according to
the colour circle.
Blue-Green 498
Fig. 17: Representation of the data of Fig.15 in the Euclidean plane
5 Outlook
As already pointed out in the introduction, the presented survey only gives
a first insight into what dyadic mathematics could be. This hopefully is
stimulating to elaborate further the sketched ideas concerning a human-
oriented mathematics. Of course, comprehensive efforts are desirable. In-
stead of discussing possibilities of further developments of dyadic mathe-
matics, it shall only be explained that there are even good reasons for ex-
tending dyadic mathematics to triadic mathematics. Inspired by Peirce's
triadic theory of signs understood as a basic theory of human thought, a
Triadic Concept Analysis has already been invented based on triadic con-
texts which code when an object has an attribute under a certain condition
(see [LW95],[Wi95],[WZOO]). This led to mathematical investigations of
triordered sets and trilattices [Bi98b] and even of triadic Galois connections
492 R. Wille
[Bi97]. It turned out that Triadic Concept Analysis is a useful basis for a
modal extension of Contextual Logic (see [Wi98],[SW03]) . But also a triadic
view of algebra and geometry becomes meaningful if the degrees of freedom
of mathematizing realities are mathematized themselves on a third level.
Already Klein's Erlanger Programm [K172] may be viewed as triadic in this
sense and even more the theory of invariance and meaningfullness for rep-
resentational measurement [KLST71]. All this indicates how inspiring and
fruitful further human-oriented developments of mathematics might be.
References
[Ar95] Aristoteles Werke in deutscher Ubersetzung. Bd. 11: Physikvor-
lesung. 5. Aufl. Akademie Verlag, Berlin 1995.
[BS97] J. Barwise, J . Seligman: Information flow : the logic of dis-

tributive systems. Cambridge University Press, Cambridge, UK,
1997.
[Be99] T. Becker: Formal Concept Analysis and Algebraic Geometry.

Dissertation, TU Darmstadt. Shaker, Aachen 1999.
[Bi97] K. Biedermann: Triadic Galois connections. In: K. Denecke,

0. Liiders (eds.): General algebra and applications in discrete
mathematics. Shaker, Aachen 1997, 23~33 .
[Bi98a] K. Biedermann: Completion of triordered sets and trilattices.

In: D. Dorninger, G. Eigenthaler, H. K. Kaiser, H. Kautschitsch,
W . More, W. B. Muller (eds.): Contributions to general algebra
10. Verlag Johannes Heyn, Klagenfurt 1998.
[Bi98b] K. Biedermann: A foundation of the theory of trilattices. Dis-

sertation, TU Darmstadt. Shaker, Aachen 1998.
[Bi40] G. Birkhoff: Lattice theory. First edition. AMS, Providence,

R.I., 1940.
[Bu98] S. N. Burris: Logic for mathematics and computer science. Pren-

tice Hall, Upper Saddle River, N. J ., 1998.
[Ca66] R. Carnap: Philosophical foundations of physics. Basic Books,

Inc., New York 1966.
[DK03] F. Dau, J. Klinger: From Formal Concept Analysis to Contex-

tual Logic. FB4-Preprint, TU Darmstadt 2003.
[Du95] Duden - Das grof3e Worterbuch der deutschen Sprache in 8

Banden. 2. Aufl. Dudenverlag, Mannheim 1993- 95, S.2145.
[Er82] M. Erne: Einfiihrung in die Ordnungstheorie. B.I.-

Wissenschaftsverlag, Mannheim 1982.
[Ga03] B. Ganter: Relational Galois connections. (to appear)
[GW86] B. Ganter, R. Wille: Implikationen und Abhangigkeiten zwis-

chen Merkmalen. In: P. 0. Degens, H.-J. Hermes, 0. Opitz
(Hrsg.): Die Klassifikation und ihr Umfeld. Indeks-Verlag,
Frankfurt 1986, 171-185.
[GW89] B. Ganter, R. Wille: Conceptual scaling. In: F. Roberts (ed.):

Applications of combinatorics and graph theory to the biological
and social sciences. Springer, New York 1989, 139- 167.
[GW99] B. Ganter, R. Wille: Formal Concept Analysis: mathematical

foundations. Springer, Heidelberg 1999.
[GHOl] M. Gehrke, J. Harding: Bounded lattice expansions. Journal of

Algebra 238 (2001) , 345- 371.
[Gr99] A. Grof3kopf: Conceptual structures of finite abelian groups.

Dissertation, TU Darmstadt. Shaker, Aachen 1999.
[Ha92] G. Hartung: A topological representation of lattices. Algebra

Universalis 29 (1992) , 273- 299.
[Ka88] I. Kant: Logic. Dover Edition, Mineola 1988.
[KV96] K. A. Kearnes, F. Vogt: Bialgebraic contexts from dualities.

Australian Math. Soc. (Series A) 60 (1996), 389- 404.
[Kl72] F. Klein: Vergleichende Betrachtungen iiber neuere ge-

ometrische Forschungen. Verlag von Andreas Deichert, Erlangen
1872.
[KLST71] D. Krantz, R. D. Luce, P. Suppes, A. Tversky: Foundations

of measurement. Vol. 1, 2, 3. Academic Press, San Diego 1971,
1989, 1990.
[LSW97] R. Langsdorf, M. Skorsky, R. Wille, A. Wolf: An approach to au-

tomated drawing of concept lattices. In: K. Denecke, 0. Liiders
(eds.): General Algebra and Applications in Discrete Mathe-
matics. Shaker Verlag, Aachen 1997, 125--136.
494 R. Wille
[LW95] F. Lehmann, R. Wille: A triadic approach to Formal Concept

Analysis. In: G. Ellis, R. Levinson,W . Rich, J. F. Sowa (eds.):
Conceptual structures: applications, implementation and the-
ory. LNAI 954. Springer, Heidelberg 1995, 32- 43.
[MMT87] R. N. McKenzie, G. F . McNulty, W . F. Taylor: Algebras, lat-

tices, varieties. Wadsworth & Brooks, Monterey, 1987.
[Pe92] Ch. S. Peirce: Reasoning and the logic of things. Edited by

K. L. Ketner; with an introduction by K. L. Ketner and H. Put-
nam. Havard University Press, Cambridge 1992.
[Pi59] J. Piaget: La formation du symbole chez !'enfant-imitation, jeu

et reve - Image et representation. Delachaux et Niestle S.A.,
Neuchatel 1959.
[Po03J R. Poschel: Galois connections for operations and relations. In:

K. Denecke, M. Erne, S. L. Wismath (eds.): Galois connections
and applications. Kluwer, Dordrecht (to appear)
[PK79] R. Poschel, L. A. Kaluinin: Funktionen- und Relationenalge-

bren. VEB Verlag der Wissenschaften, Berlin 1979.
[PW86] W. Poguntke, R. Wille: Zur Restrukturierung der mathema-

tischen Ordnungstheorie. In: A.-M. Kempf, F. Wille (Hrsg.):
Mathematische Modellierung. McGraw-Hill, Hamburg 1986,
283-293.
[Pr02] V. R. Pratt: Chu spaces as semantic bridge between linear logic

and mathematics. Special issue on linear logic, TCS (2002).
[PrOO] S. Prediger: Terminologische Merkmalslogik in der Formalen

Begriffsanalyse. In: G. Stumme, R. Wille (Hrsg.): Begriffiiche
Wissensverarbeitung: Methoden und Anwendungen. Springer,
Heidelberg 2000, 99- 124.
[SF68] H. Schiffmann, Ph. Falkenberg: The organization of stimuli and

sensory neurons. Physiology and Behavior 3 (1968) , 197-201.
[SW03] L. Schoolmann, R. Wille: Concept Graphs with Subdivision: A

Semantic Approach. In: A. de Moor, W . Lex, B. Ganter (eds.):
Conceptual structures for knowledge creation and communica-
tion. LNAI. Springer, Heidelberg 2003, 271- 281.
[SeOl] Th. B. Seiler: Begreifen und Verstehen. Ein Buch iiber Begriffe
und Bedeutungen. Verlag Allgemeine Wissenschaft, Miihltal
2001.
[So84] J. F. Sowa: Conceptual structures: information processing in

mind and machine. Adison-Wesley, Reading 1984.
[SW86] J. Stahl, R. Wille: Preconcepts and set representations of con-

texts. In: W. Gaul, M. Schader (eds.): Classification as a tool
of research. North-Holland, Amsterdam 1986, 431-438.
[St32] M. H. Stone: Linear transformations in Hilbert space and their

applications to analysis. AMS Publ. XV, NewYork 1932.
[SWW01] S. Strahringer, R. Wille, U. Wille: Mathematical support for

empirical theory building. In: H. S. Delugach, G. Stumme
(eds.): Conceptual structures: broadening the base. LNAI 2120.
Springer, Heidelberg 2001, 169- 186.
[Ur78] A. Urquhart: A topological representation theory for lattices.

Algebra Universalis 8 (1978), 45- 58.
[Vo94] F. Vogt: Bialgebraic contexts. Dissertation, TU Darmstadt.

Shaker, Aachen 1994.
[Vo95] F. Vogt: Subgroup lattices of finite Abelian groups. In:

K. A. Baker, R. Wille (eds.): Lattice theory and its applica-
tions. Heldermann Verlag, Lemgo 1995, 241- 259.
[VW94] F. Vogt, R. Wille: Ideas of Algebraic Concept Analysis. In: H.-

H. Bock, W. Lenski, M. M. Richter (eds.): Information systems
and data analysis. Springer, Heidelberg 1994, 193- 205.
[Wa73] H. Wagner: Begriff. In: H. Krings, H. M. Baumgartner,

C. Wild (eds.): Handbuch philosophischer Grundbegriffe. Kosel,
Miinchen 1973, 191-209.
[Wi76] R. Wille: Allgemeine Algebra zwischen Grundlagenforschung

und Anwendbarkeit. Der Mathematikunterricht 22 (1976), 40-
64.
[Wi81] R. Wille: Versuche der Restrukturierung von Mathematik am

Beispiel der Grundvorlesung "Lineare Algebra" . Beit6ige zum
Mathematikunterricht 1981. Schroedel, Hannover 1981, 102-
112.
496 R. Wille
[Wi82] R. Wille: Restructuring lattice theory: an approach based on

hierarchies of concepts. In: I. Rival {ed.): Ordered sets. Reidel,
Dordrecht-Boston 1982, 445~470.
[Wi89] R. Wille: Lattices in data analysis: how to draw them with a

computer. In: I. Rival {ed.): Algorithms and Order. Kluwer,
Dordrecht-Boston 1989, 33~58.
[Wi92] R. Wille: Concept lattices and conceptual knowledge systems.

Computers & Mathematics with Applications 23 {1992), 493~
515.
[Wi95] R. Wille: The basic theorem of Triadic Concept Analysis. Order

12 {1995), 149~ 158 .
[Wi96] R. Wille: Restructuring mathematical logic: an approach based

on Peirce's pragmatism. In: A. Ursini, P. Agliano {eds.): Logic
and algebra. Marcel Dekker, New York 1996, 267-281.
[Wi98] R. Wille: Triadic concept graphs. In: M. L. Mugnier, M. Chein

(eds.): Conceptual structures: theory, tools and applications.
LNAI 1453. Springer, Heidelberg 1998, 194-208.
[WiOOa] R. Wille: Contextual Logic summary. In: G. Stumme (ed.):

Working with conceptual structures: Contributions to ICCS
2000. Shaker-Verlag, Aachen 2000, 265~276.
[WiOOb] R. Wille: Begriffiiche Wissensverarbeitung: Theorie und Praxis.

Informatik Spektrum 23 {2000), 357~369; shorter version in:
B. Schmitz {Hrsg.): Thema Forschung: Information, Wissen,
Kompetenz. Heft 2/2000, TU Darmstadt, 128- 140.
[Wi01a] R. Wille: Restructuring general algebra: an approach based on

the idea of operational descriptions. Lecture at the 62nd Arbeit-
stagung Allgemeine Algebra, Universitat Linz 2001.
[Wi01b] R. Wille: Mensch und Mathematik: Logisches und mathema-

tisches Denken. In: K. Lengnink, S. Prediger, F. Siebel (Hrsg):
Mathematik und Mensch: Sichtweisen der Allgemeinen Mathe-
matik. Verlag Allgemeine Wissenschaft , Miihltal 2001, 139~ 158.
[Wi02a] R. Wille: Kommunikative Rationalitat, Logik und Mathematik.

Math. Semesterberichte 49 {2002), 167~183 .
[Wi02b] R. Wille: Existential concept graphs of power context families.

In: U. Priss, D. Corbett, G. Angelova (eds.): Conceptual struc-
tures: integration and interfaces. LNAI 2393. Springer, Heidel-
berg 2002, 382~395 .
[Wi02c] R. Wille: Applications of lattices: Formal Concept Analysis.

In: A. V. Mikhalev, G. F. Pilz (eds.): The concise handbook of
algebra. Kluwer, Dordrecht 2002, 446~450.
[Wi03a] R. Wille: Truncated distributive lattices: conceptual structures

of simply implicational scales. Order (submitted)
[Wi03b] R. Wille: Conceptual contents as information- basics for Con-

textual Judgment Logic. In: A. de Moor, W. Lex, B. Ganter
(eds.): Conceptual structures for knowledge creation and com-
munication. LNAI. Springer, Heidelberg 2003, 1 ~ 15.
[Wi03c] R. Wille: Sind unsere Vorstellungen von Raum und Zeit richtig?
oder: Besteht ein Kontinuum a us Punk ten? In: L. Hefendehl-
Hebeker, S. HuBmann (Hrsg.): Mathematikdidaktik: Zwischen
Fachorientierung und Empirie. Franzbecker Verlag, Hildesheim-
Berlin 2003, 266~279.
[WW93] R. Wille, U. Wille: On the controversy over Huntington's equa-
tions. When are such equations meaningful? Mathematical So-
cial Sciences 25 {1993) , 173~180 .
[WW95] R. Wille, U. Wille: Uniqueness of coordinatizations of ordinal
structures. Contributions to General Algebra 9 (1995) , 321 ~324.
[WW96] R. Wille, U. Wille: Coordinatization of ordinal structures. Order

13 (1996), 281 ~284.
[WW03] R. Wille, U. Wille: Restructuring general geometry: measure-
ment and visualization of spatial structures. Contribution to
General Algebra 14. Johannes Heyn Verlag, Klagenfurt 2003
(to appear)
[WZOO] R. Wille, M. Zickwolff: Grundlagen einer Triadischen Begriff-

sanalyse. In: G. Stumme und R. Wille (Hrsg.): Begriffiiche
Wissensverarbeitung: Methoden und Anwendungen. Springer,
Heidelberg 2000, 125-150.
[Wl91] U. Wille: Eine Axiomatisierung bilinearer Kontexte. Mitt.
Math. Sem. GieBen 200 (1991) , 71~112.
498 W . Gahler
[Wl95] U. Wille: Geometric representation of ordinal contexts. Disser-

tation, Univ. GieBen 1995. Shaker Verlag, Aachen 1996.
[Wl96] U. Wille: Representation of ordinal contexts by ordered n-

quasigroups. European Journal of Combinatorics 17, 317- 333.
[W197] U. Wille: The role of synthetic geometry in representational

measurement theory. Journal of Mathematical Psychology 41
(1997), 71- 78.
[W199] U. Wille: Characterization of ordered bilinear contexts. Journal

of Geometry 64 (1999) , 167-207.
[WlOO] U. Wille: Linear measurement models - axiomatizations and

axiomatizability. Journal of Mathematical Psychology 44 (2000) ,
617-650.
[Xi93] W. Xia: Morphismen als formale Begriffe: Darstellung und

Erzeugung. Dissertation, TU Darmstadt. Shaker, Aachen 1993.
[Xi94] W. Xia: Galois connections and formal concept analysis.

Demonstratio Mathematica XXVII (1994) , 751-767.
{bf Author's address:

Technische Universitiit Darmstadt
Fachbereich Mathematik
SchloBgartenstr. 7, D-64289 Darmstadt
wille@ma thema tik. tu-darmstad t. de
Index
BCK-algebra, 404 of a term, 371
b-deductive system, 399 concrete category, 413
congruence kernel , 401
adjunction, 389 conjugate pair of additive clo-
adjunctions, 5 sure operators, ix
Alexandroff topology, 414 conjugate pair of closure opera-
algebraic varieties, 67 tors, 214
attribute exploration, 364 connectedness, 140, 422
axiality, 13 context, 364
coproduct, 142
band, 381 covariant functor, 414
Brouwer semilattice, 404 covering, 154
category d-deductive system, x
connected, 142 de Morgan algebras, 272
locally connected, 142 deductive system, 399
central extensions, 160 degree of the minimum polyno-
clone, 233 mial, 47
clones, 287 depth
closed, 364 of a term, 375
closure operation, 8, 414 digital topology, x, 420
additive, 414 dualgruppe, 52
completely additive, 414
idempotent, 414 ECE-equation, 347
closure operator, 213 elementary symmetric functions,
additive, 214 24
closure space, 413 endoprimal algebras, 272
connected, 415
closure systems, 79 fibration , 140
cohomology, 157 Fichtner terms, 402
commutative ring, 157 finite limits, 142
commutator, 161 fixed-point set, 224
complementation, 389 free extension, 156
complete lattice, 79, 214 free pregroup, 394
complete sublattice, 214 full subcategory, 417
complexity functional completeness, ix
500
functionally complete, 260 left pregroup, 393

fundamental theorem of classi- local closure, 287
cal Galois theory, 140 locally automorphism primal, 264
locally endomorphism primal, 264
Galois closed, 10, 236 locally primal, 264
Galois connection, 10, 235 locally quasiprimal, 264
Galois field, 7
Galois resolvent for an equation, Magid's theory, 140
45 Mal'tsev variety, 162
Galois structure, 140 many-valued logic, x
Galois theory, 87 mindepth, 375
Galois-closed subrelation, ix, 219 minimum polynomial, 38
general independence, x modus ponens, 400
general superposition, 237
Goursat variety, 162 natural transformation, 414
near-unanimity term, 265
Hilbert algebra, 400 normal extension, 154
implication, 400 Normal Form of Boolean Func-
implication algebras, 404 tions, 99
independence, 284 normal identity, 371
initial, 364 normal variety, 371, 375
initial object, 142
Ockham algebras, 272
injective, 364
operational system, 231
integral domain, 27
order affine complete lattices, 268
internal precategories, 141
order of a group, 30
internal precategory, 14 7
interpolation polynomial, 44 partial algebra, 345
intuitionistic logic, x, 399 perfect adjunction, 11
invariance property, 232 permutation, 36
isotone map, 11 permutation groups, vii
k-normalization, 375 polarity, viii, 13, 346
kernel operation, 8 polynomial completeness, 259
kernel operator, 224 pre-Krasner clone, 238
Khalimsky topology, 420 primality, ix, 263
Kleene algebras, 272 primitive elements, 40
Krasner clone, 238 principal dual ideal, 7
principal ideal, 7
left adjoint, 6 pullbacks, 144
Index 501
Q-independence, 277
radical ideals, 67
regular strong equation, 347
relational algebra, 234, 262
relational clone, 234, 262
relational system, 231 , 415
relative automorphism, vii
relative pseudocomplement, 404
right adjoint, 6
right pregroup, 393
Ring of symmetric polynomials,
25
semi-abelian category, 162

Stone algebras, 272
Stone space, 145
strong equation, 347
strong superposition, 237
subcontext, 364
symmetric functions, 21
symmetric polynomials, 21
t-independence, 284
terminal object, 142
topological category, 416
topology, 414
topos, 142
ultrafilter, 93
variety
congruence modular, 146
weak automorphism, x, 277

weak Krasner clone, 238
weakly regular algebra, 400
weakly universal covering, 155

Galois Connections and Applications

Uploaded by

Copyright:

Available Formats

Galois Connections and Applications

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Galois Connections and Applications

Uploaded by

Copyright:

Available Formats

Galois Connections and Applications

Mathematics and Its Applications

Springer Science+Business Media, B.V.

ISBN 978-90-481-6540-7 ISBN 978-1-4020-1898-5 (eBook)

Printed on acid-free paper

All Rights Reserved

K. Gtazek, St. Niwczyk

Galois connections provide the order- or structure-preserving passage

In the first contribution, which is also aimed at "novices" in the the-

tions or "mixed" Galois connections. The classical examples come

Three other contributions are devoted to some order- and lattice-

The frequently observed phenomenon that approximation properties

Often it is necessary to define Galois connections for proper classes

The survey paper "Galois Connections for Partial Algebras" by P.

A standard algebraic approach to fragments of implicational or intu-

In " Complexity of Terms and the Galois Connection Id - Mod", K.

Another fundamental type of Galois connection for universal algebra

mial completeness (meaning that all functions, or at least all relation-

vation of inclusion and of emptiness is required) and the category of

Klaus Denecke Marcel Erne Shelly L Wismath

K. Denecke et al. (eds. ), Galois Connections and Applications, 1-138.

1.1 Adjunctions between Ordered Sets

2 The Roots of Galois Connections

2.1 Prelude on Symmetric Functions

3 Galois Connections in the Twentieth Century

3.1 From Dedekind to Noether, Artin and van der Waerden:

4.1 From Boole to Schroder: The Logical Calculus

1 The Idea of Adjunctions and Galois Connections

1.1 Adjunctions Between Ordered Sets

c+b2 >(x-b) 2 :::::} c +b2 > 0 ¢:? -c<b2 .

description of such situations is by means of adjoint pairs of maps

Q(q) = max{p E P: A(p) ~ q}.

A: P---+ Q, pH (\{q E Q: p::; Q(q)}

1.2 Closure and Kernel Operations

)q?A = -\, g-\ g = g.

Adjunctions and Closure-Kernel Isomorphisms

'Y =go,\ c K,=Aog

Adjoint pairs compose in the obvious manner: if(-\ , g) is an adjunction

previously established correspondence may be formulated as a

isomorphism, into a "closure adjunction" (constituted by a closure op-

1.3 Galois Connections

every pair of mutually inverse dual isomorphisms between

While adjoint maps are isotone (order-preserving), the partners of a

be Galois open, because >..oe is a kernel operation). The fundamental

tion if both partners are surjective (respectively, all involved elements

For (a) ::::;. (g) and (h) , take

1.4 Polarities and Axialities

Y ~ B to the set +--R(Y) = YR = {x E A: xRy for all y E Y}.

from a unique relation R between the underlying sets in that manner,

Of course, there is also a covariant counterpart of polarities, obtained

Notice that if F is a function rather than a general relation, then

In order to demonstrate how polarities and axialities may simplify ar-

D;;" is a right D~-module. Denote by .C(D~) the closure system of all

and n; given by xCy iff y belongs to the column space of x, I.e.

Thus, NC = K, RN = L , LC =..l , RK =..l , and so RNC =..l.

1.5 Formal Concept Analysis

• Hierarchical classification of objects

• Reduction to basic concepts

. . . we go back to the origin of the lattice concept in nineteenth-

Let us explain briefly the mathematical framework for t hese ideas