PDFlib PLOP: PDF Linearization, Optimization, Protection

Page inserted by evaluation version

www.pdflib.com – sales@pdflib.com
 A Primer on Galois Connections
M. ERNE, J. KOSLOWSKI, A. MELTON,
AND G . E. STRECKER
Institut fur Mathematik
Universitat Hannover
0-30167 Hannover I
Germany
and
Department of Mathematics and
Department of Computer and Infomation Sciences
Kansas State University
Manhattan, Kansas 66506

ABSTRACT. The rudiments of the theoly of Galois connections (or residuation theory, as it
is sometimes called) are provided, together with many examples and applications. Galois
connections occur in profusion and are well known to most mathematicians who deal with order
theory; they seem to be less known to topologists. However, because of their ubiquity and
simplicity, they (like equivalence relations) can be used as an effective research tool throughout
mathematics and related areas. If one recognizes that a Galois connection is involved in a
phenomenon that may be relatively complex, then many aspects of that phenomenon immedi-
ately become clear, and thus, the whole situation typically becomes much easier to understand.

INTRODUCTION

Mathematicians are familiar with the following situation: there are two “worlds”
and two transforming functions back and forth between these worlds. Moreover,
after an object has been transformed from one world to the other and then back to
the first world, a certain stability is reached in that further transformations produce
the same results. In particular, three transformations always yield the same result as
just one (no matter where one starts). Further, both worlds often carry natural
orders. When the transforming processes respect these orders, then we frequently
have situations that are both simple and rich and that can be handled in a very
elegant way. Their simplicity is reflected in the succinct definition (see Definition l),
and their richness is manifested in the wealth of results that follow from the
definition (see Propositions 2-6). These pleasant situations are called Galois connec-
tions aftcr Evariste Galoisa whose work initiated the study of the connection between

Mathematics Subject Classification. Primary 06A15, 06-01, 06A06; Secondary 54-01, 54B99,
54899,68F05.
Key words and phrases. Galois connection, closure operation, interior operation, polarity,
axiality.
‘Galois’ main results were published fourteen years after his early death (at the age of 21 in a
duel) by Liouville in his Journal de mathimatiquespures et applqukes (1846). For a translation of
Galois’ original notes Memoire sur les conditions de rksolubiliti des iquationspar radicaux, see the
text by Edwards [ 141.
103
104 ANNALS NEW YORK ACADEMY OF SCIENCES

the world of intermediate fields of a field extension E : F and the world of subgroups
of the group of automorphisms of E that fix the subfield F (cf. Examples 1 and 19).
Today this area is known as Galois theory.
Since the proofs of many results in this paper are either well known or easily
obtained, we do not include them. Galois connections were originally expressed in a
symmetric but contravariant form with transformations that reverse (rather than
preserve) order. Early references to this form are [8], [22], [44], and [45]. We use the
covariant form since it is more convenient, for example, compositions of Galois
connections are handled more easily; it allows for more natural categorical explana-
tions (e.g., by means of adjunctions); and it is more applicable to computer science
situations (where relative information preservation is important). For references to
the covariant form, see [7], [51], [9], [32], [25], [35], [15], [42], [31], and [19].

THE DEFINITION AND SOME OF ITS CONSEQUENCES

We formulate all our results in terms of partially ordered sets (orposets), that is,
sets equipped with a reflexive, transitive, and antisymmetric relation. Everything can
easily be generalized to preordered sets (i.e., one may drop the antisymmetry
requirement) and even to preordered classes. Applications of these generalizations
can be found in Examples 10, 22, and 23, and in [lo]. For a poset 9 = (P,s ) , the
order-theoretical dual (P, 2 )is denoted by 9OP.
=*
1: Consider posets 9 = (P, I ) and d = (Q, E ) . If P
DEFINITION -+ Q and
Q 2P are functions such that for allp E P and all q E Q

psa*(q) iff a * @ ) ~ q (1)

then the quadruple IT = (9, IT *, IT*, d ) is called a Galois connection.

We also write 9 5 d (or sometimes just ( I T * ,T*))for the whole Galois

connection. The functions T * and IT* are called the coudjoint part and the udjoint
part of T, respectively. We write P , for the T*-image of Q, and 9"for (PT,I ) .
Similarly, Q,, denotes the T * -image of P,and @,, stands for (Q,, E ) . The elements
of P" (respectively, Q,) are called vclosed (respectively, .r-open). We say that T is a
coreflectionb (respectively, reflection) if IT * (respectively, IT* ) is a one-to-one func-
tion, and an interior (respectively, closure) connection if IT* (respectively, I T * ) is an
inclusion. The pair IT is called an isomorphism if both IT and IT* are one-to-one, and
hence mutually inverse bijections (see Proposition 3(5) and (9) below). We call a
function between posets (co)adjoint iff it is the (co)adjoint part of a Galois connec-
tion.
2:
DEFINITION
(1) A self-mapfof a poset (P,I) is said to be increasing (respectively, decreasing)
i f p 5 f@) (respectively, f@) 5 p) for eachp E P. Aclosure (respectively,

hIn computer science coreflectionsare frequently called embedding-projectionpairs.

ERNE et al.: GALOIS CONNECTIONS 105

interior) operation on (P,5 )is an order-preserving, idempotent, and increas-

ing (respectively, decreasing) self-map.
( 2 ) Aclosuresystem of (P, 5 )is a subset Q ofP such that for eachp E P there is a
smallest q E Q withp I q, called the closure o f p and denoted byp-. The
order-theoretically dual notions are interior system and interior po of p. By a
closure system on a set X , we mean a closure system of the power set lattice
P(X) = ( P ( X ) ,2).
Closure connections 7~ bijectively correspond to closure operations: assign to ?r
the composite IT* IT, (cf. Proposition 3(4) and 6). Mapping each closure operation
0

on a poset to its image yields a bijective correspondence with closure systems of the
poset.

1: Let Q be a closure (interior) system of 9.

PROPOSITION

(1) A subset A of Q has an infimum (supremum) in &' = (Q, I )iff it has an

infimum (supremum) in 9 , and whenever either exists, they are equal.
( 2 ) If B G Q has a supremum (an infimum) h in ~ 9 ,b- (b")is the supremum
then
(infimum) of B in d .
(3) If 9 is a complete lattice, then the closure (interior) systems of 9 are those
subsets that are closed under infima (suprema). Hence, under the induced
order, closure (interior) systems of 9 are also complete lattices. They need
not, however, be sublattices.

Galois connections behave well under composition and order-inversion.

2: If .9 -%d and d
PROPOSITION 9 are Galois connections, so are
071 TOP
9 L 9 , where p 0 7~ = (p* and ( Q , ) = d u P - 9 0 p
O ? ~ * , I T * O ~ * ) , =
(P,z ) ,where TOP = (IT,, IT.,). In particular, the .?r-openelements of Q are precisely
the .rPP-closed ones.
3: For any Galois connection 9 5 d
PROPOSITION
(1) Both T * and T*preserve order.
( 2 ) .?r* and IT* are mutual quasi-inverses, that is,
IT,OIT*OIT~ =IT* and I T * O T ~ ~ I T * = T * .

(3)pEP"iffpisafixedpointof1~*01~,,andqEQ,iffqisafixedpointof
?r* o r * .
(4) .?r* ?* is a closure operation on 9 that induces a closure connection
0

9 5 B-; similarly, IT*," ?r* is an interior operation on d that induces an

interior connection 6'a A d . Hence PT is a closure system of 9 , and QT
is an interior system of B , so the results of Proposition 1 apply.
(5) 9 vand 19,are isomorphic posets; the restrictions of IT* and T* to these
posets yield an isomorphism 9 5B -; and ?r canonically factors as the
closure connection .A followed by this isomorphism fi followed by the interior
106 ANNALS NEW YORK ACADEMY OF SCIENCES

connection ir; namely

(6) The functions IT * and IT* uniquely determine each other; in fact
l ~r*(q)l) and r * ( q ) = S U P ( b E P I I T * ( P )
~ * ( P ) = i n f ( h E Q 5; 91).
This justifies calling IT +. the coudjoint (or lower adjoint) of r * ,and calling IT*
the adjoint (or upper adjoin?)of IT * .

(6) For eachp E Pp its .rr*-inverseimage has r * (p) as thc smallest element;
similarly, the T % -inverse image of each q E (2, has IT*(q) as the largest
clement.
(7) r * preserves joins (i.e., suprema), and IT* prescrves meets (i.e., infima).
(8) r is a reflection (i.e., P* is one-to-one) iff P * is surjective iff P * IT* = LdQ;
0 '

and I T is a coreflection (i.e., r * is One-to-one) iff IT* is surjective iff IT* 0 T * =

idp.
Now we see that Galois connections can be characterized by other simplc
conditions that sometimes are easicr to verify than the preceding definition. Recall
that t p = [x E P i p 5 x] is the principal filter generated by p E P, and .1q =
'h
[ y E Q ly 5 q ] is theprkcipal ideal generated by q E Q. We write P ( Y ) + P ( X ) for
h
the inverse image function of X + Y.

PROPOSITION 4:Let 9 = (P, < ) and d' = (Q, f Q and

c ) be posets, and let P +
Q A P be functions.

(1) The following are equivalent:

(a) ( 9 , f , g , ~7)isaGaloisconnection.
(b) f and g preserve order, g f is increasing, and f g is decreasing.
0 0

(c) fprcserves order, and for each q E Q the largcst element off'( .1q ) is
dq).
(d) g preserves order, and for eachp E P the smallest element of g'( t p ) is
f(P).
(2) fis coadjoint ifffisresiduated, that is, ifff' preserves principal ideals. If .Yis
complete, this condition is equivalent to thc preservation of arbitrary SU-
prema by f .
(2) g is adjoint iff g is residual, that is, iff g' preserves principal filters. If d is
completc, this condition is cquivalent to thc preservation of arbitrary infima
by g-
PROPOSITION 5 : By assigning to each coadjoint self-map of a poset 9 its adjoint,
onc obtains a dual isomorphism (reversing both the composition and the order)
between the monoid of all coadjoint self-maps on 9 and the monoid of all adjoint
self-maps on 9, both ordered pointwise. In particular, this isomorphism prescrves
ERNE et al.: GAullS CONNECTIONS 107

idempotency. Moreover, coadjoints that are increasing (decreasing) correspond to

adjoints that are decreasing (increasing).
Finally, let us have a look at the relativization of Galois connections.
PROPOSITION 6: Let P be a closure system of a poset Z = (X, )I with closure
connection Z 29 = (P,I), and let Q be an interior system of a poset =
(Y, E ) with interior connection d = (Q, L ) y. Then every Galois connec-
tion 2? y w i t h p * (x) E Q for eachx E Xand p* (y) E P for eachy E Y induces a
Galois connection 9 5d such that the following diagram commutes

In particular, p and IT induce the same isomorphism (and the lower trapezoid
collapses iff IT is already an isomorphism).
Conversely, under the preceding hypotheses every Galois connection 9 5 13

% 9 -
is induced by a uniquc Galois connection F
Ti
y , namely
d 2y .Hence the set of all Galois connections from 9 to d
the composite

with the pointwise order (on the coadjoints) is isomorphic to the poset of all Galois
connections from 2to that restrict to ones from 9 to @ .

EXAMPLES OF GALOIS CONNECTIONS

T * , T * , @ ) is a Galois connection.
In each example of this section T = (9,
Many natural examples where 9 and d are power sets are given in the fifth and
sixth sections.
EXAMPLE (cf. [44],[3]): Let E : F be a field extension, and 9 be the set of
intermediate fields between F and E ordered by inclusion. Let G = G ( E ,F ) be the

subgroups of G ordered by reverse inclusion. Define 9 @ and d -

group of all field automorphisms of E that fix F pointwisc, and let d be the set of all
T*
9 by
IT * ( L ) = (g E G Jgfixes L pointwise]
IT*(H)= (a E El every h E H fixes u ] .
This is the Galois connection that arises from Galois theory in its modern form, as
presented, for example, by Artin [3]. (Of course, neither modern algebraic notions
such as field or automorphism group nor order-theoretical aspects occur explicitly in
108 ANNALS NEW YORK ACADEMY OF SCIENCES

Galois' original work.) By known theorems of classical algebra, every finite subgroup
H of G is +P-closed, that is, 7~ * IT* ( H ) = H , and IT* ( H ) : F is a finite field extension
with dimension card ( H ) . Conversely, if E : F is a finite extension, then card ( G )is
bounded by dim ( E : F ) , and hence IT is a reflection. If in addition E : F is a Galois
extension, that is, if F is IT-closed, then the fundamental theorem of Galois theory
states that R is an isomorphism; in this case every intermediate field between F and E
is r-closed.
h
EXAMPLE 2 (cf. [57], [27]): For a functionX + Y between two sets, let 9 and @
be the power sets of X and Y, respectively, both ordered by inclusion. Define the
'h
direct imagefunction 9 % d and the inverse imagefunction d -+ 9 by

h,(U) = ( y E YI with h ( x ) = y ] , the direct image of U under h,

h'(V) = [x EX(X(3,,,withh(x) =y], theinverseimageof Vunderh.
Since h, preserves unions and 9 is a complete lattice, by Proposition 4(2) h, is
the coadjoint part IT * of some Galois connection IT. The corresponding adjoint T*
turns out to be h', which by Proposition 3(8) must preserve intersections. The
elements of 9* are called h-saturated. Notice that IT* also preserves unions and
hence is a coadjoint.
-
If h is only a partialfunction X Y, that is, h is defined on only a subset X ' C X ,
then h , still is coadjoint. Its adjoint IT* then maps V C Y to the disjoint union of
h'(V) a n d X - X ' . In particular, ifX' z X , then IT* does not preserve the empty join
(hence by Proposition 3(8) it cannot be coadjoint), and the inverse image function h-
does not preserve the empty meet (hence it is not adjoint).
h
EXAMPLE 3 (cf. [27]): Even for a partial function X - Y the inverse imagc
function h' preserves unions, and hence by Proposition 4(2) is coadjoint. Let 9 and
& be the power sets of Y andX, respectively, both ordered by inclusion. (Note that
the rolcs of X and Yare reversed as compared to Example 2.) Define 9 3 d' to be
h-. According to Proposition 3(6) for corresponding adjoint IT* must map U 5. X to
the union of all those subsets of Ywhose inverse image under h is contained in U. It
turns out that T * ( U ) = Y - h , (X - 17).
Notice that h , preserves unions but need not preserve intersections, while the
adjoint to h
' we just described preserves intersections but need not preserve unions.
EXAMPLE 4 (cf. [16]): Let h be a continuous map between topological spaces or,
more generally, between closure spaces ( X , 2')and (Y, y ) (Let (Y and y are
closure systems on X and Y, respectively, and h ' maps y into Z).CThe modified
direct imagc function 2?
% maps C E 9 to (h+(C))-. Let h: denote the
restricted inverse image function from $/ to 2. Then T = (h;,h') is a Galois
connection from (F, C) to ( v ,C). In particular, h; preserves joins (but not
k
necessarily unions), and h l preserves meets (i.e., intersections). Also, X -+ Y is
continuous from (X, ti") to (Y,Y) iff there exists a coadjoint (= joint-preserving)

'If 9-is a topology on X,and 2 consists of the :!-closed sets, then @, 3)is a closure space.
ERNE et ai.: GALOIS CONNECTIONS 109

K ?X
Z --i) y such that K 0 qx = q y 0 k, where the “completion maps” X +2 and
Y5 send each point to its closure. Clearly, K then must agree with k;. This
construction is fundamental for the representation of topological spaces by complete
lattices and vice versa, and also for certain aspects in the theory of hyperspaces (see,
e.g., [25] and [18]).
EXAMPLE 5: Let ,9be the set of all open sets and d that of all closed sets of a

closure operation ( )- gives rise to a closure connection P(X) -

topological space with underlying set X, both ordered by inclusion. The topological
Y
d , and the
topological interior operation ( )” induces an interior connection 9‘2P(X).
Their composite y L is a Galois connection 9 --%@with T * (U)= U - and r *
0

( V ) = Y o .The wclosed members of 9 are the regular open (/) sets U = U-”, and the
P-open members of d are the regular closed (!) sets V = P-. Thus ( )- and ( )”
induce an isomorphism between the lattice of regular open sets and that of regular
closed sets.
Composing 7 with the isomorphism d’ 9OP given by complementation
yields a Galois connection 9 -%9OP with T‘*(U)= T‘*(U)= U -’ = U lo. Hence
the lattice 9 = 9 of regular open sets is self-dual. In fact, these lattices are
always Boolean (see Example 18).
Notice that the operation of taking the complement of the topological closure
does not induce a Galois connection from P(X) to P(X).p (e.g., in the usual topology
on the reals {0]-’-’= 0 2 (0)).Nevertheless the Galois connection T‘can be used to
prove the standard result that there are at most 14 sets that result from the
application of complement and closure to any subset U G X (since U-’ is open,
U-l-1 = P‘* (U) is a fixed point of T’*‘0 T’*).
As can be seen from the following example, a similar phenomenon occurs with
the compactness operator introduced in [13].
EXAMPLE 6: Let 9 be the set of all topologies on Xordered by inclusion, and let
d be 9OP. Consider the operator K that maps any topology Y on X to the topology
obtained by using the 7-compact sets as a base for the closed sets. Even though K
preserves order from 9 to d and vice versa (because K reverses inclusion), for every
infinite set the pair (K, K) is not a Galois connection (while for finite X it trivially is
one). For example, if a is any ordinal, and ‘7 is thc Alexandroff topology of all upper
sets (cf. Example 9), then every subset of a is compact, and consequently, K*(Y) is the
cofinite topology on a. For infinite a,this topology is neither coarser nor finer than
9. However, in all cases (K, K) induces a Galois connection between the K-images of
@‘ and 9. This follows immediately from the fact that K preserves order and that
K(Y)G K~(F) for any topology Y (cf. [13, Theorems I1 (8) and (9)]). Thus
Proposition 3(2) gives an instant proof via Galois connections of theorem I1 (10) of
[13] that K~ = K ~ .
EXAMPLE 7: Let 9 be the set of all topologies on Y,and let d be the set of all
h
topologies on X , both ordered by inclusion. Given a function X -+ Y, define T * and
110 ANNALS NEW YORK ACADEMY OF SCIENCES

rr* by

(Y)
,rr* = { h - ( V ) l V € P} (9
E P)
= { V C Ylh'(V) E -7)
Tr*(~Y) (YE a.).
Observing that h is a continuous function from (X, Y) to (Y, 9) iff ?r* ( Y )G 7iff
Y C_ v* ( Y )we , see that T * and v* are in fact that coadjoint part and the adjoint
part of a Galois connection 9 5d . Hence v + (9) is the coarsest topology 9on
X that makes h continuous from (X, .F)to (Y,9) (called the initial topology with
respect to h and P), and a*(.F) is the finest topology 9 ' on Y that makes h
continuous from (X, 7) to (Y,9) (called thefinal topology with respect h and ,Y).
The v-closure of a topology .Y on Y is a new topology that consists of all those
sets B C_ Y with h'(B) = h'(V) for some V E 9. On the other hand, the interior
operation ,rr* 0 v* maps a topology 7on X to a new topology that consists of all
thoseUE Y w i t h U = h ' ( B ) f o r s o m e B L Y . )
A topology Y on X is T-open iff (each member of) .o/is h-saturated (see Example
2), and a topology 9on Y is v-closed iff it is h-discrete, that is, h, (X) and all subsets
of Y - h , ( X ) are Y-open. By Proposition 3(5) the lattice of h-saturated topologies
o n X is isomorphic to the lattice of h-discrete topologies on Y . If h is one-to-one, thcn
every topology on X is h-saturated, and if h is onto, then every topology on Y is
h-discrete.

EXAMPLE 8 (cf. [45]: Let B(X) and F(X) bc the sets of all bases (for topologies)
and of all filters on X, respectively. Both B(X) and F(X) are subsets of the double
power set ofX, where -9E B(X) iff U,V E 9 andx E U n Vimply the existence of
some W E A? withx E W C_ U n V, and 9 E F(X) iff X E 9 and .7 is closed under
supersets and binary intersections. (To make F(X) into a complete lattice, we includc
the power set of X as a filter.) Define a Galois connection T from 9 = (B(X), C) to
d = (F(X), C_)x,the set of all functionsx- F(X) with the pointwise order induced
by the inclusion on F(X), by
rr* (a)(x) = (VC_XI3,, withx E B C V ] for allx E X
v * ( u ) = {ULX(Vx,,,UEu(x)}.

The rr-closed bases are precisely the topologies on X , in fact, T*T* (9 ) is the.
topology generated by the base 9 . Thc rr-open filter functionsX--!!- F(X) are the
topological neighborhood functions, that is, for each V E u ( x ) there exists a subset
U C Vwithx E U and U E u ( y ) for everyy E U. Thus we obtain the classical bijective
correspondence between topologies and functions that assign neighborhood filters to
points.

EXAMPLE
9: For a topology 7on X define a preorder I X by
x s F y = x E U impliesy E Ufor all U E Y. (4)
This condition is equivalent to [x)- C_ { y ] - .The preorder I iis called the speciulizu-
tion order o n Xwith respect to Y .It is a partial order iff Y is a To-topology.
ERNE et al.: GALOIS CONNECTIONS 111

For a preordered set (X, I ) each U C X generates an u p p r sef .T. U =

u[t x Ix E U ) .Among the topologies with specialization I there is a finest one, the
Alexondroftopology d g that, consists of all upper sets U = U C X (cf. [I]), and a
coarsest one, the upper topology Ys,that is generated by the complements of the
principal ideals J x , for x E X (cf. [25]). The closed sets with respect to ..dS are
precisely the open sets with respect to .dr, that is, the lower sets 1 V = U( 1x Ix E V ) .
Let ~Ybe the set of all topologies onXordered by inclusion, and let d be the set
of all preorders on X ordered by reverse inclusion. One obtains a Galois connection
~Y5d' by setting
a * ( . Y ) = I : ~ and T*(I) = d<.

Herc T is a reflection, that is, every preorder is P-open, and P is an isomorphism iff X
is finite. The P-closed topologies are precisely the Alexandroff topologies; the
corresponding lattice is isomorphic to the lattice of all preorders on X ordered by
reverse inclusion.
Among the topologies 7 with specialization order I those with the property
that every I -directed set with a least upper bound in U E :7 is eventually in U are
called order-consistent;the finest such topology is the Scott topology 9 ''It consists of
.
all those U G X that satisfyx E U (* U n D # 0 for every I-directed D G X with a
least upper boundx (cf. [25] and [19]). (Note that " =" implies U = t U.) In fact, all
topologies between V, and Y' arc order-consistent. Hence we may restrict P * to
the subposet 9'of order-consistcnt topologies on X to obtain a new Galois
connection T ' with P ' * ( I ) = .Yg. Now the T'-closed elements of 9 ' are precisely
the Scott topologies (cf. [41]).
EXAMPLE 10: Let ,Y be the set of all equivalcncc relations on a set X, ordered by
inclusion, and let d be the class of all functions with domain X , ordered
f h
by ( X Y ) E ( X + Z ) iff there exists some function Y 4 2 that satisfies g o f = h.
--+

Define 7~ * and P* by setting

This is not a Galois connection as defined in Definition 1for two reasons:

0 d is a proper class rather than a set;
The order on @ is only a preorder, rather than a partial order, that is, it lacks
antisymmetry.
However, neither of these problems creates a serious roadblock. Galois connections
can be defined for both proper classes and preorders without any changes; however,
some of the consequences and notions in Propositions 3-6 require slight modifica-
tions. For example, rather than P * = P * P* P * in Proposition 3(2) we have
0 0

T* = P * inthesensethatT* c T *
O P * O P ~ L alsoane an element
O T * O T *

p E P is wclosed iff p I P * P * (p) I p , and an element q E Q is .rr-open iff

q c 7~ * T* ( q ) c q. And so forth.
112 ANNALS NEW YORK ACADEMY OF SCIENCES

In this example the r-open functions with domain X are precisely the surjective
ones, and all equivalence relations on X are wclosed (so T is a corcflection). This
makes precise the usual interplay between equivalence relations and surjective
functions.
We conclude our list of examples with two related constructions that establish
interesting links between order theory, topology, algebra, and computer science.
EXAMPLE 11(cf. [58], [17], [20]): Let Z denote an R-invariant extension, that is, a
function assigning to each poset 9 a certain collection X (9) of lower sets including
all principal ideals such that every residuated map 9 + f d is Z -continuous, that is,
V E Z (@) implies f + ( V ) E Z (9). Equivalently, this means that every residual
map d 5 9 is %quasi-closed, that is, U E Z ( 9 ) implies J.g,(U) E Z ( @).
Possible choices for Z are: principal ideals, arbitrary lower sets, finitely generated
lower sets, directed lower sets, lower cuts (see Example 25), ideals, Scott-closed sets
(see Example 9).
An element p of a 2-complete poset 9 (where each Z E Z ( 9 ) has a
join = supremum) is called Z -compact or Z -prime if it belongs to every 2 E X (.9)
whose join dominates p, and =Pis called Z -compactlygenerated if every element of
9 is a join of Z-compact elements. By a 2-lattice we mean a 27-compactly
generated complete lattice. Of fundamental importance for various topological
representation theories are the following two facts:
(A) The X-lattices are, up to isomorphism, precisely the closure systems that
are closed under Z -unions.
This generalizes the known facts that the compactly generated complete lattices are
represented by algebraic closure systems, the V-primely generated complete lattices
are represented by lattices of closed sets in topological spaces, etc.
(B) A coadjoint map between Z -compactly generated posets preserves 2-com-
pactness iff its adjoint preserves Z -joins.

By definition, a map 9 + f B preserves Z-compactness iff it maps 27-compact

g
elements of 9 to Z-compact elements of d , and a map d + 9 preserves
Z -joins iff g(V2) = VgZ , for all Z E Z ( d ).
By combining (A) with the hyperspace construction of Example 4, one obtains an
equivalence between the category of Z -sober closure spaces (i.e., Toclosure spaces
where the point closures are precisely the Z-compact closed sets) with continuous
maps, and the category of Z -lattices with maps preserving joins and Z -compact-
ness. By (B) these categories are dual to the category of Z-lattices and maps
preserving meets and Z -joins.
In case 3 selects the finitely generated lower scts, we arrive at the known duality
between sober spaces and (dual) spatial frames (see, e.g., [25]).On the other hand,
for the special selection of all directed lower sets (“ideals” in the sense of [25]), these
considerations reduce to known dualities between algebraic lattices and (compact)
semilattices, etc. For reasons of limited space, we refer the reader to the literature,
especially to Hofmann and Stralka ([32]); throughout that study on compact topologi-
ERNE et UL: GALOIS CONNECTIONS 113

cal semilatices, a consistent theme is the continued application of Galois connections

(see also Example 12).
Now suppose that two R-invariant extensions Z and Z ’ are given.
Let Z S Z ’ be the category whose objects are the 3-sober closure spaces with a
base of Z‘-compact open sets, and whose morphisms are maps such that inverse
images of 3 ‘-compact open sets are again Z ‘-compact and open (in particular,
such maps are continuous).
Let Z C Z ’ be the category whose objects are those Z -lattices whose duals are
2 ’-lattices, and whose morphisms are maps that preserve arbitrary joins, 3-corn-
pactness, and 3’-meets.
Because of (B) it is straightforward to show that order-dualizing the objects and
passing to adjoint morphisms yields a dual isomorphism between the categories Z C Z ‘
and Z ‘ C Z , and this is the root for many other dualities. Indeed, the aforemen-
tioned hyperspace construction leads to an equivalence between 3 S Z ’ and 3 C Z I,

as well as to one between Z ‘ S Z and Z ‘ C Z . By composing these with the duality

earlier, one arrives at a nice “symmetric” duality between the categories Z S Z ‘ and
Z’SZ. Now, special choices of Z and 2 ’and suitable restrictions provide a
multitude of “classical” dualities, encompassing the dualities between
0 Sober spaces and spatial frames [25], [35]
0 Sober spaces with compact-open base and distributive semilattices [28]
Boolean spaces (Stone spaces) and Boolean lattices [54], [35]
0 semilattices and algebraic lattices [24], [32]
Alexandroff spaces (posets) and completely distributive algebraic lattices [ 161
and many others, including Lawson’s duality for continuous, respectively, algebraic
posets (see [38] and Example 12 below). For details consult [20].
EXAMPLE 12 (cf. [43], (51, [54]): A vital theme of modern order theory is
continuousposets, and more generally, Z -continuousposets (where Z denotes, as in
the previous example, an R-invariant extension). Such posets may be characterized
most conveniently in terms of Galois connections: a poset 9 = {P, I ) is X -complete
iff the principal ideal embedding of 9 into Z (9 ) given byx rf J x has a coadjoint
(namely, the join map V from 2 (9) to 9). A Z -complete poset 9 is Z -continu-

x -
ous iff its join map V is also adjoint (to the Z -below map
Jrx = n(zE z (SP)~X I vz)).
This approach makes many considerations on Z -continuous posets short, elegant,
and transparent.
A “good” class of morphisms is certainly formed by coadjoint maps that preserve
the 2-below relation (i.e., that satisfyf,( JZx) C J Z f ( x ) ) : via Galois connections,
the corresponding category is dually isomorphic to the category of Z -continuous
posets and adjoint maps preserving Z -joins. Full subcategories are formed by the
Z -algebraic posets, that is, Z -continuous and Z -compactly generated posets (cf.
Example 11). Since an element p is 2-compact iff p E J z p , preservation of the
Z -below relation here is tantamount to preservation of Z -compactness.
If 2 selects all lower sets, that is Z (P,I ) is the dual Alexandroff topology d 2
(cf. Example 9), then the Z-continuous posets are just the completely distributive
114 ANNALS NEW YORK ACADEMY OF SCIENCES

lattices (cf. [48]).Hence a complete lattice .9 is completely distributive iff the join
map V from sd2 to 9 is a complete homomorphism.
An easy Galois argument (explained in greater detail in Example 31) leads to the

0
Ax: Y -
following result on so-called 3 join ideals, that is, fixed points of the operator
.1 (VZIZE Z ’ ( 9 ) , Z L .1 Y ) .
For any Z -continuous poset with idempotent Z -below relation, the closure
system of 2-join ideals is completely distributive, and Az is the corresponding
closure operator.
For many R-invariant extensions occurring in practice (e.g., for each of the selections
mentioned at the beginning of Example 11, and more generally, for each 2 such that
U jY E Z ( 9 )whenever E 3 (2 (9), C)), evey .Z -continuous posets has an
idempotcnt 5-below relation (see [5]). For example, if Z selects the directed lower
sets, then ‘‘-continuous”
2 mcans “continuous,” “3 -algebraic” means “algebraic,”
and the 2-join ideals are just the Scott-closed sets. By the previous remarks, they
form a completely distributive lattice for any continuous poset (cf. [38] and [25]).

RESIDUATED SEMIGROUPS

In this section we briefly discuss a general concept that accounts for a large class
of interesting Galois connections.
DEFINITION 3: A partially ordered semigroup (P, I , .), that is, a poset .Y =
(P, I )equippcd with an order-prcserving associative multiplication .9 x .y ~ 7 ,
is called residuated if all left translations ZF 9 and all right translations
9 3 9 are residuated, that is, have adjoints .Y 9 and 9 9 ,respec-
tively; this means
s Ir\t 0 r’s I t 0 r < tJs.
Notice that, in addition to thc Galois connections (r . -, r\-) and (- .s, - / s )
from 9 to 9 given by the left and right translations, for every f E P the pair
(- / t , t\-) is a Galois connection from .!P to 9Op.
By Proposition 4(2) a complete partially ordered semigroup (9, .) is residuated
iff the multiplication distributes over arbitrary suprema, that is, for all r E P and
SCP
r * VS = V ( r . s l s E S) and VS * r= V{s * r l sE S).
Nowadays, complete residuated semigroups are referred to as quantales. For a
comprehensive trcatise on this modcrn branch of order theory, see the text by
Rosenthal[50]. Aclassic source on residuation theory is the monograph by Blyth and
Janowitz [9].
EXAMPLE 13 (cf. [18]):Let (A, .)be a semigroup, and let Z be a closurc system on
A that makes the left and right translations continuous. The complete latticc (2, C)
becomes a quantale when equipped with the multiplication R 0S = (r * s i rE R,
ERNE et UL: GALOIS CONNECTIONS 11s

C, 0 )may be regarded as a “residuated completion” of

s E SJ-. The quantale (Y,
the “generalized semitopological semigroup” (-01, 2, *).By Example 4,every continu-
ous semigroup homomorphism lifts to a residuated semigroup homomorphism
between the associated completions. These facts provide a broad spectrum of useful
constructions for order-theoretical and topological completions. Among these, we
only mention two specific examples: the MacNeille completion of a residuated
semigroup, and the Vietoris hyperspace of a continuous semilattice. For more details
see [21] and [MI.
EXAMPLE 14: In ring theory, the closure systems of additive subgroups, of left
ideals, of right ideals, and of two-sided ideals of a given ringA each play a major role.
Endowed with the multiplication RS = (Z~,or,sl Jr,E R, s, E S , n E o),each of these
closure systems becomes a quantale. However, only in the case of two-sided ideals
are there “natural” expressions for both the left and right residual maps given by
R\T = {a € A lVrERraE T ) and T / S = (a € A IVSEsasE TI
Morc topologically flavored variants of these examples are the quantales of all Icft,
right, or two-sided closed ideals, respectively, of a C *-algebra.
EXAMPLE 15: By a partial semigroup d ,we mean a setA equipped with a partial
binary operation . such that whenever r . s and s * t are defined, then so are (r .s ) . t
ands . (r . t ) , and these products are equal. Any such .d induces a quantale (P(A), @)
via
R @ S = (r . S I T E R , s E S, and r - s is defined)
R\T=[aEAJV~,,ifr.aisdefined,thenr.uE TJ
if a .s is defined, then a .s E TI.
TIS = (u EA IVxrES

Important subquantales of (P(A), €3) are the Alexandroff topologies of all left, right,
or two-sided ideals ofA, respectively.
EXAMPLE 16: Let 8 be a set (often called an alphabet), and let Z* be the free
monoid of Swords with concatenation as the operation. Subsets of Z* are called
languages over 8. By the previous remarks the languages form a quantale, and
consequently every language L gives rise to three Galois connections (I,@-, L\-),
{-@L, --lL),and (-\L, LJ-).
EXAMPLE 17: Define a partial semigroup operation . onA X A by setting (a,b) .
( d , c ) = (a,c ) iff b = d. Then for relations R, S 2 A x A the product R €9 S is thc usual
composite rclation ((a,c ) I 3 b with (a, b) E R and (6,c ) E S).So (P(A x A ) , 8)is a
quantale. The next two sections deal explicitly with Galois connections induced by
rchtions.
EXAMPLE 18 (cf. [26], [8], [23], [49], [28], [37]): Letp be a futed element of a
meet-semilattice eY = (P, I). We say 9’is p-pseudocomplemented if every element
r E P has ap-pseudocomplement r p (also denoted by r * p , r + p , r\p,p/r,p : r, etc.)
that satisfies r A s I p iff s I rp. Since A is symmetric, T * (r) = T*(r) = rp defines a
116 ANNALS NEW YORK ACADEMY OF SCIENCES

Galois connection 9 5 9OP. Hence thep-skeleton

P“ = Pp = (rplr E P ] = (r E P l r = rpp)
is a closure system of 9; in particular, 9 p = (Pp, I) is a meet-subsemilattice of 9 ,
and if 9 is complete, so is 9 p . The meet-semilattice 9 is called Brouwerian or

-
equivalent to saying that each of the unary meet operations 9
(with adjoint 9
r* -
-
relativelypseudocomplemented iff it isp-pseudocomplemented for eachp E P. This is
rA- .
is residuated
9). Hence the Brouwerian semilattices are those residuated
semigroups whose multiplication is the binary meet. Brouwerian lattices with least
element are also known as Hqting algebras, and complete Heyting algebras are
known as frames or IocaEes. Since residuated maps preserve joins, every Brouwerian
lattice is distributive, and the locales are precisely those complete lattices that satisfy
the infinite distributive law
rA V S = V ( r A s I s E S).
Lattices of this type play a considerable role in disciplines as diverse as congruence
theory, (intuitionistic) logic, and topology. In fact, residuation yields a duality
between the category of frames (with maps preserving arbitrary joins and finite
meets) and the category of locales (with maps adjoint to frame morphisms), which in
turn contains (an isomorphic copy of) the category of topological spaces (cf. Example
11).As was pointed out by Isbell [34], Banaschewski and Pultr [4],Johnstone [36],
and others, this aspect is of fundamental importance for a general theory of ”spaces
without points.”
Moreover, Galois connections of the present type produce Boolean lattices in
abundance:
Anyp-pseudocomplemented meet-semilattice 9 has ap-skeleton 9 p that is a
Boolean lattice with least elementp and greatest elementp.
To verify this claim, three simple observations suffice: for r, s E Pp
(1) (rp A sp) p is the join in . 9 p ;
(2) f l is the complement of r in g a
p ;
(3) (r A s p ) p is the relative pseudocomplement in 9 p (in particular, 9 p is
distributive).
In casep is the least element I,the preceding notions reduce to pseudocomplement,
pseudocomplemented (semi-) lattice, and skeleton, respectively.
Since every topology 9 is a locale with pseudocomplements U L = U-I, one
immediately concludes that the regular open sets form a Boolean lattice, namely the
skeleton of 7(cf. Example 5). Compare this concise argument with, for example, the
treatment in [29].

POLARITIES

All Galois connections between power sets are induced by relations between the
underlying sets in particularly simple ways. For sets A and B we first study the
ERNE et ~i.: GALOIS CONNECTIONS 117

contravariant case of Galois connections from P(A) to P(B)”P, so-called polarities

between P(A) and P(B). (Recall, however, that many authors use the term “Galois
connection” exclusively in the contravariant sense.) Since the 7r-open subsets of such
a Galois connection 7r form a closure system on B, the term “+P-closed’’ is preferable
in this context.

7 (cf. [8],[44]):
PROPOSITION
R C
(1) Any relation R G A x B induces a Galois connection P(A) A P(B)”P. The
components of R+ = (R,, R + ) are defined by
+

R+(U) := (bEBIVa,,,(a,b)ER] forUCA

R+(V) := (a EAlVbE,(a,b)ER} forVCB.
(2) If for a polarity P(A) A P(B)”P the relation IT 1 C A x B is defined by
(a,b ) E Irrl iff b E 7rTT*( { a ] )(or, equivalently, a E rr* ( ( b ) ) )then
, n = (IT[ ++.
(3) Every relation R G A x B satisfies J R + + J= R. Hence every polarity rr
between P(A) and P ( B ) comes from a unique relation, namely R = I T 1, and
vice versa.
(4) R + +:=(R++)”p = (R+,R , ) is the polarity betweenP(B) andP(A) induced by
the opposite relation RoP = {(b,a ) E B x A I < a , b ) € R].
The previous considerations admit a natural generalization. Let P be a closure
system on A , and let Q be a closure system on B. By a polarity between P and Q we
mean a Galois connection between the complete lattices (P, C) and (Q, 2). Since
every complete lattice is isomorphic to the closure system of all principal ideals, and
is dually isomorphic to the closure system of all principal filters, essentially all Galois
connections between complete lattices may be regarded as certain polarities. The
polarities between P and Q form a complete lattice with respect to the pointwise
ordering, called the ternorproduct of P and Q (see, e.g., [52].By Propositions 6 and 7,
every polarity between P and Q is induced by a unique relation R L A x B. Hence the
interior system of all relations R C_ A X B with R+, (P(A))C Q and R (P(B))C ++

9 is dually isomorphic to the tensor product of P and Q .

EXAMPLE 19: The classical Galois connection of Example 1 is induced by the

relation R = ((a,g) E E X G Ig(a) = a ) .The R++-closed subrets of E are precisely the
7-closed intermcdiatefields of E : F , and the R++-closedsubsets of G are precisely
the +P-closed subgroups of G, cf. Proposition 6.

EXAMPLE 20 (cf. [15]): ForA = P ( X ) and B = X x X , the Alexandroff reflection

of Example 9 is induced by the relation R = {(U,( x , y ) ) € A x Blx E U ory E U). For
V C_ P ( X ) the specialization order R+ ( V )= Iyl‘ is defined in analogy to formula
(4).The adjoint R maps a relation S on X to the Alexandroff topology sdls of all right
+

S-closed sets, that is, U E R + ( S ) = d.7iff x E U and @ , y ) E S together imply y E U .

Thus R+R+ ( Y)is the Alexandroff topology generated by V ,while R+R+(S) is the
reflexive transitive closure of S. Hence the R++-closed subsets ofA are precisely the
Alexandroff topologies on X , and the R++-closed subsets of B are precisely the
preorders on X .
118 ANNALS NEW YORK ACADEMY OF SCIENCES

EXAMPLE
-
21 (cf. [53]):For the set F(X) of all filters on X consider the relation
R C P(X) x (X x F(X)) given by (U, (x, F ) ) E R x tZ U or U E F . The R++-closed
subsets ofA are preciscly the topologies on X . The R++-closedsubsets o f X X F(X)
are precisely the topological convergence relations, that is, the relations C L X x F(X)
for which there exists a topology Y o n X such that a filter F Y-converges to a pointx
iff (x, F ) E C . This polarity makes prccisc the interplay between topologies and filter
convergence.
The relation R of Example 20 can be viewed as a restriction of the current
relation R if points are identified with principal ultrafilters. By restricting from
arbitrary filters to principal filters (which can be identified with subsets of X ) , one
obtains a polarity whose closed subsets of X X P(X) are precisely the topological
adherence relations.
EXAMPLE 22 (cf. [46], [2]): Let R be the relation on the class of all topological
spaces defined by (X, Y) E R iff all continuous functions from X to Yare constant.
Although this is a relation on a proper class, we still obtain a polarity on the
collection of all subclasses of the class of topological spaces. Given a class Z of
topological spaces, the spaces in R+ (a)are called Z-connected, and the spaces in
R+R+ ( Z ) are called E-disconnected. For example, if 2 denotes a two-element
discrete space, then “{2}-connected”means “connected” in the usual sense, and
“{2]-disconnccted”means “totally disconnected.”
EXAMPLE 23: LetA be the class of all continuous functions, and let B be the class
of all topological spaces. Definc relations R and S between continuous functions and
topological spaces as follows: x
for every continuous function X
one) continuous function Y
( h
L
-
Y, Z ) E R (respectively, (X L- Y, Z ) E S)iff
Z there is at least one (respectively, at most
Z with h = g 0 f. Then R+ maps a class A‘ of
continuous functions to the class of so-called &-injective spaces. The class S+(&)
consists of the &-separated spaces in the sense of Pumpliin and Rohrl [47]. For
example, if A is the class of all dense embeddings, then S, ) is precisely the class
of Hausdorff spaces. On the other hand, S + ( B )is the class of all surjective continuous
functions. A dual construction for R yields A’-projective spaces. All these relations
have analogs that yield useful polarities in general categories. A generalization of the
orthogonality relation R n S is used extensively in [ll].

the polarity P(A) -

24 (cf. Example 15): If .d = (A, .) is a partial semigroup, then for T C A
EXAMPLE
(-\T,?’/ -)
P(A)op is generated by the relation

{ ( a ,b ) E A x A la . b
-
is defined and a b E 7’1.
EXAMPLE 25 (cf. [56]): A “restructured” view of polarities with applications in
extramathematical disciplines is pursued in the modern theory of Formal Concept
Analysis. Here one studies a set A of “objects,” a set B of “attributes,” and a relation
R C A x B . The triple (A, B , R ) is interprcted as a contat. Pairs of the form (U, EM
P (A) x P(B)with R + ( U )= V and R+(V)= U are called concepts; U is the extent and
Vis the intent of the concept (U, V). The concepts form a complete lattice, isomorphic
ERNE et al.: GALOIS CONNECTIONS 119

to the closure system of all extents (= R++-closedsubsets ofA ordered by inclusion),

and dually isomorphic to the closure system of all intents (= R++-closedsubsets of B
ordered by inclusion). A thorough investigation of the concept lattice reveals
implications and dependencies between the various concepts involved.
If R is a preorder 5 on a set P, then R+(U) is the set of all upper (!) bounds and
R + ( U )is the set of all lower (!) bounds for U C P. The concept lattice of the context
(P,P, 5 ) is known as the Dedekind-MacNeille completion or the completion by cuts for
the preordered set .op = (P, I )(cf. [40] and [S]). It is well-known that a complete
lattice 2 is a normal completion of 9 (that is, 9 admits a join- and meet-dense
embedding into 2')iff 9is isomorphic to the completion by cuts of 9 . Two further
examples of such normal completions for 9 are the systems of all R++-closedsets
(lower cuts), ordered by inclusion, and the system of all R++-closedsets (upper cuts),
ordered by reverse inclusion.
EXAMPLE 26: As before, let 9 = (P,5 )be a preordered set, but let R be the
complementary dual relation 2.It is easy to see that R+(U)is the complement in P of
the upper set t U generated by U, while R+(V) is the complement of the lowerset 4 V
(see Example 9). Thus the isomorphism between the Alexandroff topology of all
upper (= R++-open)sets and that of all lower (= R++-closed)sets via complementa-
tion is induced by the relation R. The associated closure operations preserve
arbitrary unions.
If R is the inequality relation z on a set A , then R++is the isomorphism from
P(A) to P(A)Op given by complementation. We denote its inverse by

P(A)OP -zP(A ).
EXAMPLE 2 7 Consider a meet-semilattice 9 = (P, I ) with least element I,and
the relation R = ( ( a ,b) E P X Pla A b = I].For any U C P the set U1 = R+(U)=
R + ( U )is a lower set. If 9 is pseudocomplemented (cf. Example 18), then ULis even
a lower cut, namely the set of lower bounds for all pseudocomplements of members
of U. Any two lower sets U and V satisfy U n V = {I) iff U C V L ,which explicitly
describes the pseudocomplementation in the frame of lower sets. Hence the closure
system S ( ~ 9 of) all R++-closed (= R++-open)sets is the skeleton of this frame (see
again Example 18), and as such it is Boolean. It is now easy to see that S ( 9 ) is
actually a normal completion of the skeleton 9 l . In particular, this shows that
Boolean lattices have Boolean completions by cuts.
EXAMPLE 28: Again, let 9 = (P,I)be a meet-semilattice with least element I,
but this time consider R = [(a,b) E P x P J a A b f I].Here the R++-closed
(= R++-open)sets are upper sets. If 9 is even a Boolean lattice, then every upper set
is R++-closed;the R++-closureof U C P is just the upper set t U. In this specific
situation by Proposition 3(5) the Alexandroff topology &'< is actually self-dual, as
well as dual to . d ~ .
In particular, this applies to any power set lattice P(X). Here R+ = R + is the
section operator considered by Choquet [12] and others (see, e.g., [Sl] and [30]) in
connection with convergence-related questions: R+ ( Z ) = V = {V C XI Wu, yI U
V f 01. Moreover, 2P#is the stack generated by "M, that is, the collection of all sets
Vwith U C V C Xfor some U E 2Y. The well-known isomorphism between the lattice
120 ANNALS NEW YORK ACADEMY OF SCIENCES

of filters on X and the lattice of gnlls on X is obtained by a suitable restriction and

corestriction of R,. The ultrafilters are precisely the fuced points of this dual
isomorphism. This puts the interplay between limits and cluster points into a nice
symmetric framework: a point x is a cluster point of a stack Z (in some topological
space or convergence space) iffx is a limit of the stack 2P and vice versa.
The section operator has also been used to give a symmetric formulation of the
notion of complete distributivity: a complete lattice (X, I ) is completely distributive
iff for every collection %of subsets of X the following identity holds:

A(VUIU€ z }= V { / \ V (E 2P].
Since replacing %with the stack *%+# does not change either side of the equation, we
immediately see that complete distributivity is a self-dual property (see also Example
12).

AXIALITIES

We now consider the covariant case of Galois connections between power sets
both ordered by inclusion; we call such Galois connections axialities. The strong
similarity between the following result and Proposition 7, despite the different orders
on the second power set, is quite surprising.
PROPOSITION
8 (cf. [27]):

(1) Any relation R C A x B induces a Galois connection P(A )

components of R3' = (R3,R') are defined by
---.
RjV
P(B). The

R,(U) := (b E B13,,(a, b) E R anda E U ) forULA

R v ( V ) := (u € A IVbEB(a,b) E R implies b E V ) for V C B.

(2) If for an axiality P(A ) P(B) the relation S A x B is defined by (a,b)

E I .rr I iff b E T * ( ( aI), then TT = I
(3) Every relation R C A x B satisfies R = IR,vI. Hence every axiality TT from
P(A ) to P(B) comes from a unique relation R L A X B , namely R = I TI I, and
vice versa.
(4) RV3 = (R3,R v ) := (R0P)3vis an axiality from P(B) to P(A).
As in the contravariant case, any Galois connection between a closure system on
A and an interior system on B is induced by a unique axiality, hence by a unique
relation R C A X B (cf. Propositions 6 and 8).
The following theorem provides a common generalization of several results
encountered so far. It is based on the phenomenon, first recognized by Lawvere [39],
that quantification is adjoint to substitution. Specifically, it shows that each relation
between two sets not only induces a whole family of polarities, but also two families
of axialities, from which the polarities arise as a derived concept. By substituting
singletons for A, C, and B in parts (I), (2) and (3) of the theorem, respectively, one
ERNE ei al.: GALOIS CONNECTIONS 121

obtains (after suitably renaming the sets) Proposition 8( l),(4), and Proposition 7(1).
Specializing Theorem 1to the caseA = B, we arrive at Example 17.
Recall that the composite relations R G A x B and S G B x C is a subset of
A X C, namely R €9 S = {(a,c) € A X C J3bE,(a, b ) E R and (b,c ) E S].
THEOREM 1: (The bi-category of sets, relations, and inclusions is biclosed with
respect to €9.)
(1) If S G B X C, then for each set A the function P(A X B)-
3'P(A X C) is
-as

-
coadjoint and its (uniquely determined) adjoint P(A x C) -+ P(A x B )
mapsTCA x C t o T U S := ( ( u, b)€ A x BI VCEc(a,c)E $ ; j b , c ) E S).
(2) If R C A X B, then for each set C the function P(B x C) P(A x C)

m a p s T L 4 x C t o R D T := { ( b , c )E B x C( VGa(u,b) E R
-DT
-
RD-
is coadjoint and its (uniquely determined) adjoint P(A x C) .--) P(B x C)
(a,c)E T ) .
(3) If T G A x C, then for each set B the functions P(A x B ) -+ Pop( B x C)
TQ-
and P(B x C)Op ----) P(A x B ) constitute a Galois connection
{- D T, T Q -).
To obtain a complete characterization via relations of all possible Galois connec-
tions between the power sets ofA and B, respectively, we use the complementation-
induced Galois connections KS and KB of Example 26 to define Galois connections
R- - and Rv3 by

P(A)OP - R ~.
PW P(A)OP - q3
P(5)OP

Notice that (Rv3)OP = (Rop)3v.Writing Rc for the complement (A x B ) - R, we also

have

(Rc)-- = K, o Rv3 = R 3 0 K~ and = K';

(Rc)++ 0 Rv3 = Rv3 0 K'; (5)
as well as

T 4 S = (7'"€9 SOP)c R DT
and = (RoPC3 TC)". (6)
9: A relation R C A x B satisfies:
PROPOSITION
(1) R 3 ( U ) C R v ( U ) (respectively, R v ( U ) C R3 (U)) for every U G A iff each
a E A is R-related to at most one (respectively, at least one) b E B; in this
case we call R right unique (respectively, lefi total);
122 ANNALS NEW YORK ACADEMY OF SCIENCES

(2) R3(V) C Rv(V) (rcspcctively, R v ( V ) C R J ( V ) )for every V C B iff each6 E B

is R-related to at most one (respectively, at least one) a E A; in this case we
call R lej? unique (respectively, right total).

h
EXAMPLE 29: When a partial function A B is identified with its graph, that is,
A

h C A x B, then the axiality hgv is prccisely the Galois connection of Example 2, and
the axiality hv3 is precisely the Galois connection of Example 3. From Examplc 2 it
can be seen that h3 C hv, that is, h 3 ( V ) C_ h v ( V ) for all V C B. This property in fact
characterizes partial functions: by Proposition 9, R C A x B is the graph of a partial
-
functionA B iffR3 C R'. Moreover, R is the graph of a function fromA to B iff
R 3 = R'.

PROPOSITION 10: Using the backwards composition of rclations S R = R @I S (as

0

in the case of functions), one has (S R)3v= SIV RIV. Hcncc thc axialities on P(A )
0 0

form a quantale that is isomorphic to the quantale (P(A x A ), 0) of all relations on A

(cf. Example 17).

EXAMPLE 30 (cf. Example 15): If .d= (A, .) is a partial semigroup and R, S G A ,

then both of (R, @I -, R\-) and (- 8 S, - / S ) are axialities from P(A) to itself. The
first is induced by the relation R ' C A x A given by (a, b) E R ' iff b = r .a for some
r E R, and the second is induced by the relation S ' C A x A given by (a,b ) E S ' iff
b = a C_ s for somes E S.

EXAMPLE 31: Suppose R is an idempotent relation on A = B, that is, R = R @I R.

This is tantamount to saying that R is transitive and has thc interpolationproperty: for
( a , c ) E R there exists some b with (a, b ) E R and (b,c ) E R. Then, by Propositions 5
and 10, the axiality RgVhas the property that both parts are idempotent functions.
Hence, restriction and corestriction to the Alexandroff topology JZ'R = [ U L A I
Rg(U) C U ) = [U C A I U C_ R v ( U ) } (cf. Example 20) yield a Galois connection
dR5 dRwith the property that T * is an interior operation and IT* is a closure

operation. Since .GT?R is closed under arbitrary unions and intersections, it is a

completely distributive lattice, as are the imagcs under T * and T * . Thus for any
idempotent relation R, the RgV-open(= rr-open)sets and the R3v-closed (= wclosed sets
form isomorphic completely distributive lattices (cf. [48] and [6]). Passing to the
opposite relation, we obtain two further completely distributive lattices that are
dually isomorphic to the first ones, via complementation: the R3'-opcn (closed) sets
are the complements of the Rv3-closed (open) sets.
These arguments apply, for example to the so-called way-below relation << of a
continuousposet. This relation is known to be idempotent, and the << 3v-open sets
are precisely the Scott-open scts (cf. 12.51, [19], and Example 12). Hence the Scott
topology of a continuous posct is not only completely distributive but also dually
isomorphic to the lattice of way-below sets, that is, eVg-openscts. See also [33].
We conclude our primer on Galois connections with the following observation.
Givcn two relations R L A x Band S C B x C, there are at least three natural ways to
produce a polarity from P(A ) to P(C) by using the polarities and axialities induced by
ERNE et ai.: GALOIS CONNECTIONS 123

R and S, as indicated by the following diagram:

Explicitly,S+* R3v comes from the relation (R @ Sc)c = RoP D S , and Sv3 R++from
0 0

the relation (Re63 S)' = R U Sop (use Proposition 10 and (6)). In general, all three
polarities may be different, being induced by different relations.

FIGURE 1. A regular tessellationby Galois connectionsarising from one relation.

124 ANNALS NEW YORK ACADEMY OF SCIENCES

REFERENCES

1. ALEXANDROFF, P. 1937. Diskrete Raume. Mat. Sb. 2: 501-518.

2. ARHANGELSKII, A. V. & R. WIEGAND.1975. Connectedness and disconnectedness in
topology. Topol. Appl. 5: 9-33.
3. ARTIN,E. 1959. Galoissche Theorie. Teubner. Leipzig.
4. BANASCHEWSKI, B. & A. PULTR.1989. Cauchy points of metric locales. Can. J. Math.
41: 830-854.
5. BANDELT, H.-J. & M. ERNE.1983. The category of 2-continuous posets. J. Pure Appl.
Algebra 3 0 21%226.
6. - . 1984. Representations and embeddings of A -distributive lattices. Houston J.
Math. 10 315-324.
7. BENADO,M. 1949. Nouveaux thkortmes de d6composition et d’intercalation ‘a la nor-
malit6 a.C. R. Acad. Sci. 228 529-531.
8. BIRKHOFF, G. 1940. Lattice Theory, 1st ed. American Mathematical Society Colloquium
Publication AMS, Providence, R.I. (3rd ed. 1967).
9. BLYTH,T. & M. JANOWITZ. 1972. Residuation Theory. Pergamon. Oxford.
10. CASTELLINI, G., J. KOSLOWSKI & G. E. STRECKER. 1993. Closure operators and polarities.
This issue.
11. CASTEUINI,G., J. KOSLOWSKI & G. E. STRECKER.1992. A factorization of the Pumpliin-
Rohrl connection. Topol. Appl. 44:69-76.
12. CHOQUET,G. 1948. Convergences. Ann. Univ. Grenoble Sect. Sci. Math. Phys. (N.S.)
23: 57-112.
13. DE GROOT,J., H. HERRLICH, G. E. STRECKER & E. WATTEL.1969. Compactness as an
operator. Cornpos. Math. 21: 349-375.
14. EDWARDS, H. M. 1984. Galois Theory. Springer-Verlag. Berlin/New York.
15. ERN&M. 1982. Einfiihrung in die Ordnungstheorie. B.-I. Wissenschaftsverlag. Mann-
heim.
16. -. 1984. Lattice representations for categories of closure spaces. In Categorical
Topology (Toledo, Ohio, Aug. 1983), H. L. Bentleyet al., Eds.: 197-222. No. 5 in Sigma
Series in Pure Mathematics. Heldermann Verlag. Berlin.
17. - . 1986. Order extensions as adjoint functors. Quaest. Math. 9 149-206.
18. - .1990. Residuated completions of generalized topological semigroups. In Proc. Int.
Symp. on the Semigroup Theory (Kyoto, Japan), pp. 63-83.
.
19. -1991. The ABC of order and topology. In Category Theory at Work, H. Herrlich
and H.-E. Porst, Eds.: 57-83. No. 18 in Research and Exposition in Mathematics.
Heldermann Verlag. Berlin.
20. - . 1993. Algebraic ordered sets and their generalizations. In Algebras and Orders,
Proceedings AS1 Montreal (Montreal, Canada, 1991), G. Sabidussi and I. Rosenberg,
Eds. Kluver. Dordrecht, the Netherlands. In press.
21. ERNE,M. & J. Z. REICHMAN.1987. Completions for partially ordered semigroups.
Semigroup Forum 34: 253-285.
22. EVERETT, C. J. 1944. Closure operators and Galois theory in lattices. Trans. Am. Math.
SOC.55: 514-525.
23. FRINK,0.1962. Pseudo-complements in semi-lattices. Duke Math. J. 2 9 505-514.
24. GEISSINGER, L. & G. GRAVES.1972. The category of complete algebraic lattices. J. Comb.
Theory Ser. A 13: 332-338.
25. GIERZ,G., K. H. HOFMANN, K. KEIMEL,J. D. LAWSON, M. MISLOVE & D. S. SCOTT.1980.
A Compendium of Continuous Lattices. Springer-Verlag. Berlin/New York.
26. GLIVENKO, V. 1929. Sur quelques points de la logique de M. Brouwer. Bull. Acad. Sci.
Belgique 15: 183-188.
27. GOLDBLA~, R. 1979. Topoi, Studies in Logic, Vol. 98, Sec. 15.4 453-457. North-Holland.
Amsterdam, the Netherlands.
28. GRATZER,G. 1978. General Lattice Theory. Birkhauser. Basel, Switzerland.
29. HALMOS, P. R. 1963. Boolean Algebras. Van Nostrand. Princeton, N.J.
30. HERRLICH, H. 1974. Topological structures. In Topological Structures (Amsterdam, Nov.
ERNE et nl.: GALOIS CONNECTIONS 125

1973), P. C. Baayen, Ed.: 59-122. No. 52 in Mathematical Centre Tracts, Wiskundig

Genootschap of the Netherlands. Mathematisch Centrum.
31. HERRLICH, H. & M. HU~EK. 1990. Galois connections categorically. J. Pure Appl. Algebra
68: 165-180.
32. HOFMANN, K. H. & A. STRALKA. 1976. The algebraic theory of compact Lawson semilat-
tices. Diss. Math. (Rozprawy Mat.) 137: 1-54.
33. HRBACEK, K. 1976. Continuous completions. Algebra Univers. 2 8 230-244.
34. ISBELL, J. 1972. Atomless parts of spaces. Math. Scand. 31: 5-32.
35. JOHNSTONE, P. 1982. Stone Spaces. No. 3 in Cambridge Studies in Advanced Mathematics.
Cambridge Univ. Press. Cambridge, England.
36. JOHNSTONE, P. T. 1991. The art of pointless thinking. In Category Theory at Work, H.
Herrlich and H.-E. Porst, Eds.: 85-107. No. 18 in Research and Exposition in Mathemat-
ics. Heldermann Verlag. Berlin.
37. KATRI~AK, T. 1981. A new proof of the Glivenko-Frink theorem. Bull SOC.R. Sci. Liege
5 0 171.
38. LAWSON, J. D. 1979. The duality of continuous posets. Houston J. Math. 5: 357-386.
39. LAWVERE, F. W. 1969, Adjointness in foundations. Dialectica 23: 281-296.
40. MACNEILLE, H. M. 1937. Partially ordered sets. Trans. Am. Math. SOC.42: 416-460.
41. MELTON,A. 1989. Topological spaces for cpo’s. In Categorical Methods in Computer
Science (Berlin, Germany, Sept. 1988), H. Ehrig et al., Eds.: 302-314. No. 393 in
Lecture Notes in Computer Science. Springer-Verlag. Berlin/New York.
42. MELTON,A., D. A. SCHMIDT & G. E. STRECKER. 1986. Galois connections and computer
science applications. In Category Theory and Computer Programming (Guilford,
England, Sept. 1985), D. Pitt et al., Eds.: 299-312. No. 240 in Lecture Notes in
Computer Science. Springer-Verlag. Berlin/New York.
43. NOVAK,D. 1982. Generalization of continuous posets. Trans. Am. Math. SOC.272 645-
667.
44. ORE,0.1944. Galois connexions. Trans. Am. Math. SOC.55: 493-513.
45. PICKERT,G. 1952. Bemerkungen iiber Galois-Verbindungen, Arch. Math. (Basel) 3: 285-
289.
46. PREUSS,G. 1971. Eine Galois-Korrespondenz in der Topologie. Monatsh. Math. 75: 447-
452.
47. PUMPLUN,D. & H. ROHRL. 1985. Separated totally convex spaces. Manuscr. Math.
5 0 145-183.
48. RANEY,G. N. 1953. A subdirect-union representation for completely distributive com-
plete lattices. Proc. Am. Math. SOC.4: 518-522.
49. RASIOWA, H. & R. SIKOKSKI. 1963. The Mathematics of Metamathematics. Panstwowe
Wydawnictwo Naukowe. Warszawa.
50. ROSENTHAL, K. I. 1990. Quantales and Their Applications. Longman Scientific & Techni-
cal. Harlow, Essex, England.
51. SCHMIDT, J. 1953. Beitrage zur Filtertheorie, 11. Math Nachr. 10 197-232.
52. SHMUELY, Z. 1974. The structure of Galois connections Pac. J. Math. 54: 209-225.
53. SONNER, H. 1953. Die Polaritat zwischen topologischen Raumen and Limesraumen. Arch.
Math. (Basel) 4 461-469.
54. STONE,H. M. 1936. The theory of representations for Boolean algebras. Trans. Am. Math.
SOC.40:37-111.
55. VENUGOPALAN, P. 1986. Z -continuous posets. Houston J. Math. 12: 275-294.
56. WILLE, R. 1982. Restructuring lattice theory: An approach based on hierarchies of
concepts. In Ordered Sets (Banff, Canada, Aug./Sept. 1981), I. Rival, Ed.: 445-470.
No. 83 in Series C Mathematical and Physical Sciences, NATO Advanced Study
Institute. Reidel. Dordrecht, the Netherlands.
57. WINSKEL, G. 1987. Relating two models of hardware. In Category Theory and Computer
Science, D. Pitt et al., Eds.: 98-113. No. 283 in Lecture Notes in Computer Science.
Springer-Verlag. Berlin/New York.
58. WRIGHT,J. B., E. G. WAGNER & J. B. THATCHER. 1978. A uniform approach to inductive
posets and inductive closure. Theor. Comput. Sci. 7: 57-77.