Techniques

A b o u t this B o o k
This book is intended as a course in modal logic for students who have
had prior contact with modal logic and wish to study it more deeply. It
presupposes training in mathematics or logic. Very little specific knowledge is
presupposed, most results which are needed are proved in this book. Know-
ledge of basic logic propositional logic, predicate logic as well as basic
mathematics will of course be very helpful. The book treats modal logic as a
theory, with several subtheories, such as completeness theory, correspondence
theory, duality theory and transfer theory. Thus, the emphasis is on the inner
structure of the theory and the connections between the subdisciplines and
not on coverage of results. Moreover, we do not proceed by discussing one
logic after the other; rather, we shall be interested in general properties of
logics and calculi and how they interact. One will therefore not find sections
devoted to special logics, such as G, K 4 or $4. We have compensated for
this by a special index of logics, by which it should be possible to collect all
major results on a specific system. Heavy use is made of algebraic techniques;
moreover, rather than starting with the intuitively simpler Kripke-frames we
begin with algebraic models. The reason is that in this way the ideas can
be developed in a more direct and coherent way. Furthermore, this book
is about modal logics with any number of modal operators. Although this
may occasionally lead to cumbersome notation, it was felt necessary not to
specialize on monomodal logics. For in many applications one operator is not
enough, and so modal logic can only be really useful for other sciences if it
provides substantial results about polymodal logics.
No book can treat a subject area exhaustively, and therefore a certain
selection had to be made. The reader will probably miss a discussion of
certain subjects such as modal predicate logic, provability logic, proof theory
of modal logic, admissibility of rules, polyadic operators, intuitionistic logic,
and arrow logic, to name the most important ones. The choice of material
included is guided by two principles: first, I prefer to write about what I
understand best; and second, about some subjects there already exist good
books (see [182], [43], [31], [157], [224]), and there is no need to add another
one (which might even not be as good as the existing ones).
vi About this book
I got acquainted with modal logic via Montague Semantics, but it was
the book [169] by WOLFCANC RAUTENBERG that really hooked me onto this
subject. It is a pity that this book did not get much attention. Until very
recently it was the only book which treated modal logic from a mathematical
point of view. (Meanwhile, however, the book [43] has appeared in print,
which is heartily recommended.) However, twenty years have passed from its
publication and many strong and important results have been found, and this
was the reason for writing this book.
My intellectual credits go not only to WOLFGANG RAUTENBERG but also
to SIEGFRIED BREITSPRECHER w whose early death saddened me greatly
for teaching me algebra, HELMUT S ALZMANN for his inspiring introduction to
geometry and linear algebra, and to WALTER FELSCHER for his introduction
to logic and exact mathematics. Furthermore, I wish to thank KIT FINE for
making an exception and taking me as his student in Edinburgh. He too
taught me logic in his rather distinct way. More than anyone in the last
years, FRANK WOLTER has been an inspiration and collaborator. Without
him, this book would not have been written. Thanks to CARSTEN GREFE for
his help both with some of the pictures as well as modal logic, and thanks also
to ANDREAS BULL and MARTIN MITTELMAIER. Thanks to SAM DORNER,
KIT FINE, CLEMENS HENDLER, CARSTEN IHLEMANN, TOMASZ KOWALSKI
and TIMOTHY SURENDONK for careful proofreading and RAJEEV GOR~ and
MISHA ZAKHARYASCHEV for their advice in many matters. The final draft
was carefully read by HANS MIELKE and BIRGIT NITZSCHE. Special thanks
go to ARMIN ECKER for his never ending moral support.
No endeavour can succeed if it is not blessed by love and understanding. I
am fortunate to have experienced both through my wife J OHANNA DOMOKOS,
my parents, my brother and my sister. This book is dedicated to all those to
whom it gives pleasure. May it bring in its own modest way a deeper
understanding of the human spirit.
Berlin, March 1999

Marcus Kracht
Overview
The book is structured as follows. There are ten chapters, which are
grouped into three parts. The first part contains the Chapters 1 - 3, the
second part the Chapters 4 - 7 and the third part the Chapters 8 - 10. The
first part contains roughly the equivalent of a four hour one semester course
in modal logic. Chapter 1 presents the basics of algebra and general propo-
sitional logic inasmuch as they are essential for understanding modal logic.
This chapter introduces the theory of consequence relations and matrix se-
mantics. From it we deduce the basic completeness results in modal logic.
The generality of the approach is justified by two facts. The first is that in
modal logic there are several consequence relations that are associated with a
given logic, so that acquaintance with the general theory of consequence rela-
tions is essential. Second, many results can be understood more readily in the
abstract setting. After the first chapter follow the Chapters 2 and 3, in which
we outline the basic terminology and techniques of modal logic, such as com-
pleteness, Kripke-frames, general frames, correspondence, canonical models,
filtration, decidability, tableaux, normal forms and modal consequence rela-
tions. One of the main novelties is the method of constructive reduction. It
serves a dual purpose. First of all, it is a totally constructive method, whence
the name. It allows to give proofs of the finite model property for a large va-
riety of logics without using infinite models. It is a little bit more complicated
than the filtration method, but in order to understand proofs by constructive
reduction one does not have to understand canonical models, which are rather
abstract structures. Another advantage is that interpolation for the standard
systems can be deduced immediately. New is also the systematic use of the
distinction between local and global consequence relations and the introduc-
tion of the compound modalities, which allows for rather concise statements
of the facts. The latter has largely been necessitated by the fact that we allow
the use of any number of modal operators. Also, the fixed point theorem for
G of DICK DE JONGH and GIOVANNI SAMBIN is proved. Here, we deduce it
from the so-called Beth-property, which in turn follows from interpolation.
This proof is originally due to CRAIC SMORYI~SKI [200].
The second part consists of chapters on duality theory, correspondence
theory, transfer theory and lattice theory, which are an absolute necessity for
vii
viii Overview
understanding higher modal logic. In Chapter 4 we develop duality theory

rather extensively, starting with universal algebra and Stone-representation.
Birkhoff's theorems are proved in full generality. This will establish two im-
portant facts. One is that the lattice of normal modal logics is dually isomor-
phic to the lattice of subvarieties of the variety of modal algebras. Secondly,
the characterization of modally definable classes of generalized frames in terms
of closure properties is readily derived. After that we give an overview of the
topological and categorial methods of GIOVANNI SAMBIN and VIRGINIA VAC-
CARO, developed in [186] and [187]. Furthermore, we study the connection
between properties of the underlying Kripke-frame and properties of the un-
derlying algebra in a descriptive frame. We will show, for example, that
subdirect irreducibility of an algebra and rootedness of dual descriptive frame
are independent properties. (This has first been shown in [185].) We conclude
this chapter with a discussion of the structure of canonical frames and some
algebraic characterizations of interpolation, summarizing the work of LARISA
MAKSIMOVA. An algebraic characterization of Halld~n-completeness using
coproducts is derived, which is slightly stronger than that of [153]. Chap-
ter 5 develops the theme of first-order correspondence using the theory of
internal descriptions, which was introduced in MARCUS KRACHT [121]. We
will prove not only the theorem by HENDRIK SAHLQVIST [183] but also give
a characterization of the elementary formulae which are definable by means
of Sahlqvist formulae. This is done using a two-sided calculus by means of
which correspondence statements can be systematically derived. Although
this calculus is at the beginning somewhat cumbersome, it allows to compute
elementary equivalents of Sahlqvist formulae with ease. Moreover, we will
show many new corollaries; in particular, we show that there is a smaller
class of modal formulae axiomatizing the Sahlqvist formulae. On the other
hand, we also show that the class of formulae described by VAN BENTHEM in
[10] which is larger than the class described by Sahlqvist does not axiomatize
a larger class of logics. Next we turn to the classic result by KIT FINE [65]
that a logic which is complete and elementary is canonical, but also the result
that a modally definable first-order condition is equivalent to a positive re-
stricted formula. This has been the result of a chain of theorems developed by
SOLOMON FEFERMAN, ROBERT GOLDBLATT and mainly J OHAN VAN BEN-
THEM, see [10]. In Chapter 6 we discuss transfer theory, a relatively new
topic, which has brought a lot of insights into modal logic. Its aim is to study
how complex logics with several operators can be reduced to logics with less
operators. The first method is that of a fusion. Given two modal logics,
their fusion is the least logic in the common language which contains both
logics as fragments. This construction has been studied by FRANK WOLTER
in [233], by KIT FINE and GERHARD SCHURZ [67], and by FRANK WOLTER
and MARCUS KRACHT in [132]. For many properties T it is shown that a
Overview ix
fusion has T iff both fragments have T. In the last section a rather different
theorem is proved. It states that there is an isomorphism from the lattice of
bimodal logics onto an interval of the lattice of monomodal logics such that
many properties are left invariant. This isomorphism is based on the simu-
lations defined by S. K. THOMASON in [208~ 210]. Some use of simulations
has been made in [127], but this theorem is new in this strong form. Only
the simulations of THOMASON have these strong properties. Extensive use
of these results is made in subsequent chapters. Many problems in modal
logic can be solved by constructing polymodal examples and then appealing
to this simulation theorem. Chapter 7 discusses the global structure of the
lattices of modal logics. This investigation has been initiated by WIM BLOK
and WOLFGANG RAUTENBERG, whose splitting theorem [170] has been a
great impulse in the research. We state it here in the general form of FRANK
WOLTER [234], who built on [120]. The latter generalized the splitting theo-
rem of [170] to non-weakly transitive logics and finitely presentable algebras.
[234] has shown this use to be inessential; we show in Section 7.5 that there
exist splitting algebras which are not finitely presentable. In the remaining
part of this chapter we apply the duality theory of upper continuous lattices,
which are also called frames or locales (see [110]) to modal logic. One result
is a characterization of those lattices of logics which admit an axiomatiza-
tion base. This question has been put and answered for K 4 by ALEXANDER
CHAGROV and MICHAEL ZAKHARYASCHEV [42]. The argument used here is
rather simple and straightforward. We prove a number of beautiful theorems
by WIM BLOK about the degree of incompleteness of logics. The way these
results are proved deserves attention. We do not make use of ultraproducts,
only of the splitting theorem. This is rather advantageous, since the structure
of ultraproducts of Kripke-frames is generally difficult to come to terms with.
Finally, the basic structure of the lattice of tense logics is outlined. This is
taken from [123].
The last part is a selection of issues from modal logic. Some topics
are developed in great depth. Chapter 8 explores the lattice of transitive
logics. It begins with the results of KIT FINE concerning the structure of
finitely generated transitive frames and the selection procedure of MICHAEL
ZAKHARYASCHEV, leading to the cofinal subframe logics and the canonical
formulae. The characterization of elementary subframe logics by KIT FINE
is developed. After that we turn to the study of logics of finite width. These
logics are complete with respect to noetherian frames so that the structure
theory of KIT FINE [66] can be extended to the whole frame. This is the
starting point for a rich theory of transitive logics of finite width. We will
present some novel results such as the decidability of all finitely axiomatizable
transitive logics of finite width and finite tightness and the result that there
exist 13 logics of finite width which bound finite model property in the lattice
x Overview
of extensions of $4. The first result is a substantial generalization of [247],

in which the same is shown for extensions of K4.3. In Chapter 9 we prove a
series of undecidability results about modal logics using two main ingredients.
The first is the simulation theorem of Chapter 6. And the other is the use
of the logics K . a l t n . The latter have been studied by KRISTER SEGERBERG
[197] and FABIO BELLISSIMA [5] and their polymodal fusions by CARSTEN
GREFE [91]. The latter has shown among other that while the lattice of
K . a l t l is countable, the lattice of the fusion of this logic with itself has 2 ~~
many coatoms. Moreover, the polymodal fusions of K . a l t l can be used to
code word problems as decidability problems of logics. Using this method,
a great variety of theorems on the undecidability of properties is obtained.
This method is different from the one used by LILIA CHAGROVA [44], and
ALEXANDER CHAGROV and MICHAEL ZAKHARYASCHEV [41]. Their proofs
establish undecidability for extensions of K4, but our proofs are essentially
simpler. The proofs that global finite model property (global decidability) are
undecidable even when the logic is known to have local finite model property
(is locally decidable), are new.
We conclude the third part with Chapter 10 on propositional dynamic
logic ( P D L ) . This will be a good illustration of why it is useful to have a the-
ory of arbitrarily many modal operators. Namely, we shall develop dynamic
logic as a special kind of polymodal logic, one that has an additional com-
ponent to specify modal operators. This viewpoint allows us to throw in the
whole machinery of polymodal logic and deduce many interesting new and old
results. In particular, we will show the finite model property of P D L , in the
version of ROHIT PARIKH and DEXTER KOZEN [118], of P D L with converse,
by DIMITER VAKARELOV [217], and of deterministic P D L by MORDECHAI
BEN-ARI, JOSEPH I. HALPERN and AMIR PNUELI, [7']. Again, constructive
reduction is used, and this gives an additional benefit with respect to interpo-
lation. We have not been able to determine whether P D L has interpolation,
but some preliminary results have been obtained. Moreover, for the logic of
finite computations we show that it fails to have interpolation and that it
does not have a fixed point theorem. Largely, we feel that an answer to the
question whether P D L has interpolation can be obtained by closely analysing
the combinatorics of regular languages.
CHAPTER 1
Algebra, Logic and Deduction
1.1. B a s i c F a c t s a n d S t r u c t u r e s
In this section we will briefly explain the n o t a t i o n as well as the basic

facts and structures which will more or less be presupposed t h r o u g h o u t this
book. We will assume t h a t the reader is familiar with t h e m or is at least
willing to grant their truth.
SETS, FUNCTIONS. We write {x : ~(x)} for the set of all objects satisfying
~. Given a set S, p ( S ) denotes the powerset of S, ~S the cardinality of S. For
functions we write f : A -4 B to say t h a t f is a function from A to B, and
f : x ~-4 y to say t h a t f maps (in particular) x to y. T h e image of x under f is
denoted by f(x). We write f : A ~ B if f is injective, t h a t is, if f(x) = f(y)
implies x = y for all x, y E A; and we write f : A --, B if f is surjective, t h a t
is, if for every y C B there is an x C A such t h a t y ---- f(x). For a set S C_ A,
f[S] : = { f ( x ) : x E S}. We put f - l ( y ) := { x : f(x) = y}. For a set T _C B,
f-l[T] : = { x : f(x) C T}. If f : A - 4 B and g : B -4 C then g o f : A -4 C is
defined by (g o f)(x) : = g(f(x)). The image of f : A -4 B, denoted by im[f],
is defined by im[f] : = f[A]. M N denotes the set of functions from N to M .
If C C_ A then f r C denotes the restriction of f to the set C.
CARDINAL AND ORDINAL NUMBERS. Finite ordinal n u m b e r s are con-
structed as follows. We start with the e m p t y set, which is denoted by 0. T h e
n u m b e r n is the set { 0 , 1 , . . . , n - 1}. 'i < n ' is s y n o n y m o u s with 'i E n'. In
general, an ordinal n u m b e r is the set of ordinal n u m b e r s smaller t h a n t h a t
number. So, in constructing ordinals, the next one is always the set of the
previously c o n s t r u c t e d ordinals. There are two types of ordinal n u m b e r s dis-
tinct from 0, successor ordinals and limit ordinals. A n ordinal A is a successor
ordinal if it is of the form ~ U {~}, and a limit ordinal if it is not 0 a n d not
a successor ordinal. Finite numbers are successor ordinals, with the excep-
tion of 0. Ordinal n u m b e r s are well-ordered by the inclusion relation E. A
well-order < on a set R is a linear ordering which is irreflexive and such t h a t
any n o n e m p t y subset S C_ R has a least element with respect to <. Ordinal
numbers can be characterized as sets well-ordered with respect to E such t h a t
every element of an ordinal ~ is an ordinal. A n y well-ordered set is isomorphic
to a pair @, G), ec an ordinal. '~ < A' is s y n o n y m o u s with '~ C A'. T h e Axiom
4 1. A l g e b r a , Logic and Deduction
of Choice is equivalent to t h e s t a t e m e n t t h a t every set can b e w e l l - o r d e r e d .

T h r o u g h o u t t h i s b o o k we will b e w o r k i n g w i t h t h e s t a n d a r d s e t - t h e o r y Z F C
(ZERMELO--FRAENKEL set t h e o r y plus t h e A x i o m of Choice). (See [220] a n d
[114, 115] for an i n t r o d u c t i o n to set theory.)
O n e can define t h e sum, p r o d u c t a n d e x p o n e n t i a t i o n for ordinals. T h e
s u m a n d p r o d u c t of o r d i n a l s is g e n e r a l l y n o t c o m m u t a t i v e . 1 ÷ ~v = ~v, b u t
w + 1 ~ w. A n d 2 . w = w, b u t w . 2 = : ~ + w ~ w. C a r d i n a l n u m b e r s
are s p e c i a l k i n d s of ordinals, n a m e l y those o r d i n a l s A for which no # <
can b e m a p p e d onto ~. F i n i t e o r d i n a l s are cardinals. ~ also is a c a r d i n a l
here always d e n o t e d by R0. C a r d i n a l a r i t h m e t i c is different f r o m o r d i n a l
a r i t h m e t i c , e x c e p t for finite n u m b e r s . If c~ or fl is infinite, however, we have
a + fl = a • fi = max{a, fl}. A set M has cardinality a if t h e r e is a b i j e c t i o n
f : a --+ M . T h e set 2 M of all functions from M to 2 has t h e s a m e c a r d i n a l i t y
as t h e p o w e r s e t p ( M ) . T h e r e f o r e we will o c c a s i o n a l l y i d e n t i f y ~ ( M ) w i t h 2 M.
If a is a c a r d i n a l , a + d e n o t e s t h e least c a r d i n a l g r e a t e r t h a n a . T h i s is
always defined. For a = n we have a + = n + 1. For a = R~, K a n o r d i n a l ,
R+ : = R~+I. We know t h a t always a < 2 ~. T h e Generalized Continuum
Hypothesis (GCH) is t h e c o n j e c t u r e t h a t c~+ = 2 ~. For a = Ro, t h e c a r d i n a l i t y
of t h e set of n a t u r a l n u m b e r s , this is t h e Continuum Hypothesis (CH). C H
( a n d also G C H ) is a c t u a l l y i n d e p e n d e n t of Z F C . W e will s t a t e our r e s u l t s so
t h a t t h e y are i n d e p e n d e n t of CH a n d G C H .
LATTICES AND ORDERINGS. A partial order on a set S is a r e l a t i o n _<
which is (1.) reflexive, t h a t is, x <_ x for all x E S , (2.) transitive, t h a t is,
x _ < y a n d y < z i m p l i e s x < z for a l l x , y , z C S, a n d (3.) antisymmetric,
w h i c h m e a n s t h a t for all x, y C S if x _< y a n d y _< x t h e n x = y. A chain is
a p a r t i a l o r d e r in which for a n y two x, y we have x _< y or y < x. (If e i t h e r
x _< y or y < x we say t h a t x a n d y are comparable.) L e t X C S. T h e n
:-- {y: e X ) ( y _ x)}

$x := {y:(3xeX)(y>_x)}
If X = {x} t h e n we w r i t e i x a n d i"x r a t h e r t h a n $ { x } a n d 1"{x}. A set of
t h e form S X ('~X) for some X is called a lower cone (upper cone).
A lattice is a t r i p l e ~ = (L, M, W) satisfying t h e following laws for all
x,y,z E L
xn(ynz) = (zny)nz xU(yUz) = (xUy)Uz
xMy = yOx xUy = yUx
XRX ~ X xUx ~ X
xn(yux) = x xu(ynx) = x
T h e s e laws are referred to as t h e laws of associativity, commutativity, idem-

potence a n d absorption. We call x [~ y t h e meet of x and y a n d x U y t h e join
of x and y. In a l a t t i c e we c a n define a p a r t i a l o r d e r i n g < b y s e t t i n g x _< y
if[xUy-- y. It t u r n s o u t t h a t x < y i f f x ~ y = x. F r o m t h e laws a b o v e
1.1. Basic Facts and Structures 5
follows t h a t ~ is a partial order, and x R y is the greatest lower bound (glb) of

x and y, and x U y the least upper bound (lub) of x and y. Conversely, if _< is
a partial ordering on V such t h a t the glb and the lub for any pair of elements
exists, t h e n (V, glb, lub} is a lattice.
T h e r e is a principle of duality in lattices which states t h a t any law valid
in all lattices is transformed into a valid law if R and U are exchanged. This
is due to the fact t h a t the laws postulated for lattices come in pairs, one the
dual of the other. Let £ -- (L,n,U>. We say t h a t £op = (L,U,[7> is the
dual lattice of £. T h e partial order obtained for £op is simply the converse
ordering. (Usually, since x < y means t h a t in the graphical representation x
is below y, £op is obtained from £ by p u t t i n g everything upside down.)
We say t h a t ± is a bottom element if ± < x for all x, and t h a t 7- is a
top element if x < T for all x. A structure (L, ± , T, R, U / whose reduct to
and U is a lattice with top element 7- and b o t t o m element ± is called a
bounded lattice. A n element x is an atom if for no y, ± < y < x, and a coatom
if for no y, x < y < T. Two lattices £ = (L,[7, U) and 9)I = (M,N,U> are
isomorphic iff the ordered sets <L, <) and <M, _<) are isomorphic. If a lattice
also has infinite meets and joins, t h a t is, if the glb as well as the lub of infinite
sets exists, t h e n £ is called complete. These infinitary operations are denoted
by [7 and L]. We write [-]yeYY or simply [-]Y for the meet of Y. Similarly
for the join. Notice t h a t complete lattices always have a b o t t o m and a top
element. It can be shown t h a t a lattice has infinitary glb's iff it has infinitary
lub's. A n element x of a lattice is join compact if from x < L] Y we m a y
conclude t h a t x < UY0 for a finite Y0 c Y. A lattice is algebraic if every
element is the least upper b o u n d of a set of join c o m p a c t elements. A lattice
is distributive if it satisfies the identities
xn(yuz) = (xny)u(xnz)
~U(ynz) = (~uy)n(xuz)
It can be shoWn t h a t either of the two equations implies the other.
A filter is a n o n e m p t y set F _C L such t h a t F = t F and such t h a t if
x,y E F then also x [ 7 y E F . A filter is principal if it is of the form t x
for s o m e x E L. A s u b s e t I C L i s an ideal i f I = $ I a n d for x , y E I a l s o
x ~ y C I. A n ideal I is principal if I = $ x for some x C L.
A boolean algebra is a structure ~ = (B, 0, 1 , - , N, U> such t h a t the re-
striction ~ I {N, U, 0, 1} := {B, 0, 1, A, U) is a b o u n d e d distributive lattice,
and - : B --+ B is a function satisfying for all x:
n w x
X
XA --X ---- 0
xU -x --- 1
as well as the so-called de Morgan Laws
- ( x u y) = ( - x ) n ( - y ) , - ( x n y) = ( - x ) u ( - y ) .
6 1. A l g e b r a , L o g i c a n d D e d u c t i o n
We also use the notation n , c x x or n x , and likewise for U, as for lattices.

This is in general only defined if X is finite. A general reference for the kind
of concepts introduced so far is [37], [52] and [90].
CLOSURE OPERATORS. Let S be a set. A m a p C : ~(S) -+ ~ ( S ) is
called a closure operator if it satisfies the following properties, referred to as
extensivity, monotonicity and idempotence.
(ext) X C C(X)
(mon) X C_ Y implies C(X) C_C(Y)
(ide) C(C(X)) = C(X)
A set of the form C(X) is called a closed set. Obviously, C(X) is the smallest
closed set containing X. Given a closure operator C, any intersection of closed
sets is closed again. Thus, the closure C(X) can also be defined via
C(Z) = N ( Y : Y 2 X , Y = C(Y)}
The closed sets form a lattice, with • being standard set intersection, and
X U Y = C(X U Y). A closure operator is called finitary if it satisfies
(fin) C(Z) = U ( C ( E ) : E C_ X, E finite )
If a closure operator is finitary, the lattice of closed sets is algebraic, and
conversely. The join compact elements coincide with the sets C ( E ) , E finite.
For general reference see [52].
2-VALUED LOGIC. We will assume familiarity with classical logic, and in
some sections predicate logic and some model theory is required as well. Ne-
vertheless, for reference, let us fix here what we mean by classical logic. First
of all, to avoid a terminological clash, we talk of 2-valued logic when we refer
to classical logic in the usual sense. In fact, 2-valued only refers to the fact
that we allow exactly two t r u t h values, denoted by 0 and 1. Furthermore, in
the languages of 2-valued logic we have primitive sentence letters functioning
as propositional variables, and various logical symbols with fixed meaning
throughout this book. These are the constants verum T, and falsum ± , the
negation 7, the conjunction A, disjunction V, implication --+ and biimplication
++. We m a y identify the set of t r u t h values with the set 2 (= {0, 1}, see above).
W h e n we speak of boolean logic we mean 2-valued logic for the language in
which only T, -- and A are basic, and all other symbols are defined in the
usual way. A valuation is a function from the variables into the set 2. The
t r u t h value of a complex proposition is calculated using the t r u t h tables of
the symbols, which are standard. We say ~ comes out true under a valuation,
if it receives the value 1. A formula T is a tautology if it is true under all
valuations. Formally, boolean logic is the logic of the boolean algebra 2,
which is the algebra based on the set 2 with the usual interpretation of the
symbols. Again, using the set interpretation of numbers, A will come out
as intersection, V as union and -7 as relative complement. (Notice namely,
1.1. Basic Facts and Structures 7
that 0 = ~ , 1 = {2~} and 2 = {~, {~}}.) A good reference for basic logical
concepts is [1991 or [84].
BINARY RELATIONS. T h e binary relations over a set M form a boolean
algebra with respect to intersection, complement relative to M x M , b o t t o m
element 2~ and top element VM : = M x M . A n o t h e r special c o n s t a n t is the
diagonal, A M : = {(x, x> : x E M } . Moreover, the following operations can be
defined. First, for two relations R, S C_ M x M we can define the composition
R o S : = {(x, z) : (3y E M ) ( x R y S z)}. From the composition we define the
n - f o l d p r o d u c t R n of a relation by R ° : = AM and R n+l := R o R n. T h e
transitive closure R + of R is the union U0<n<~ R~. T h e reflexive transitive
closure R* of R is Un<~ Rn, or equivalently, R + U AM. Finally, for each
relation R there is the converse R ~ : = {(y,x) : x R y } . For the converse we
have the following identities
(Rus)- = R-us~
(RoS) ~ = S-oR ~
(R*y =
G R o u P s AND SEMIGROUPS. Given a set G and a binary operation • on

G which is associative, (G, .) is called a semigroup. If 1 C G is such t h a t
1 - x = x . 1 = x for all x E G, then 1 is called a unit. Moreover, (G, 1,.)
is called a monoid. A particular example is provided by strings. Let A be a
set. T h e n A* denotes the set of finite strings over A. (For m a n y purposes one
m a y define strings as functions from a n a t u r a l n u m b e r to A; however, for us,
strings are basic objects. T h e y simply are sequences of symbols.) Strings are
denoted by vector arrow, e. g. aT, ~7, if it is necessary to distinguish t h e m from
simple symbols. Given aT and ~, ~ 7 or aT. ~Ydenotes the c o n c a t e n a t i o n of
strings of ~ and ~. T h e e m p t y string is denoted by e. (A*,¢, ~ ) is a monoid.
A string ~ is a prefix of ~ if there is a ff such t h a t ~7 = aT-~/, a~ is a postfix of
if there exists a z / s u c h t h a t ~ = f f ~ . ~ is a substring if there are ~7 and
g such t h a t j = g ~ - g . E v e r y string ~ has a length, d e n o t e d by Jail. It is
defined inductively as follows:
le[ := 0,
faT[ := 1, if£c A,
:= Iel+I l.
Suppose we have an additional operation -1 : G ---} G such t h a t the following

laws hold for all x, y E G:
X • X -1 ---- l,
X -1 "X ~ I .
T h e n the structure (G, 1, -1 , .} is called a group.

8 1. A l g e b r a , Logic and D e d u c t i o n
1.2. Propositional Languages

A propositional language L consists of three things: (1) a set var of
(propositional) variables, (2) a set F of (propositional) function symbols and
(3) a function f~: F --+ ~:. f2(f) is the arlty of f . f2 is called the signature
of L. The cardinality of vat is a m a t t e r of choice. Unless otherwise stated
we assume t h a t ~var -- ~o. Hence we have countably and infinitely m a n y
variables. Sometimes we consider the case of languages with finitely m a n y
variables; these are called w e a k l a n g u a g e s . Since the set var is usually
fixed, L is uniquely identified by f~ alone.
To take an example, let us consider the language ~3, consisting of the
function symbols A, V, ~, _1_ and T, where f~(A) = f~(V) = 2, f~(~) = 1 and
f~(T) = f~(±) = 0. So, A and V are binary function symbols, which is to say
that their arity is 2; ~ is a unary function symbol, in other words a function
symbol of arity 1 and - - finally - - T and ± are nullary function symbols:
their arity is zero. The propositional languages are a family of languages
which are syntactically impoverished. There is only one type of well-formed
expression, t h a t of a proposition. The symbol f can also be understood as a
function taking f2(f) m a n y propositions, returning a proposition (see below
the definition of the t e r m algebra). Predicate logic knows at least two types
of meaningful expressions: terms and formulas, x + y is a term, x + y = 2 is a
formula. Natural languages have even more categories to classify meaningful
expressions, for example noun phrases, verb phrases, adjectival phrases and
so on. Thus, from this point of view, propositional languages are the poorest
kind of language, those with a single type only. The freedom lies in the set
of basic functions. There are quite meaningful n - a r y functions for every n,
for example e x a c t l y o n e o f Ao,A1,... o r A ~ - I . In boolean logic such
functions are rarely studied because they can be produced by composing A
and -~; however, in other logics - - e.g. intuitionistic logic - - these definability
results no longer hold.
Let X be a set. An X - s t r i n g is a finite string consisting of symbols from
X U F. For a string ~ we write ~ C S if all members from ~ are in S. T h e
set o f f ~ - t e r m s o v e r X, Tm~(X), is defined to be the smallest set of strings
satisfying
1. For all x C X, x C Tmn(X).
2. For all f • F and tk • Tm~(Z), k < f~(f), also
f~Q.. ~tn(:)_l • Tmn(Z)

This way of writing a term will be called p r e f i x n o t a t i o n otherwise also
known as P o l i s h N o t a t i o n . The more conventional notation with brackets
and binary function symbols in between their arguments will be referred to as
infix n o t a t i o n . Typically, we will write ~ for a sequence of elements other-
wise denoted by x 0 , . . . , xk-1 for some k. Moreover, for an n - a r y function f
1.2. Propositional Languages 9
we will write / ( t o , . . . , t n - , ) instead of f ~ t o t ~ . . ~ t~-l. If we do not want

to highlight the arity of f we will write f ( ~ f o r / ( t o , t , , . . . , tn_,). When the
length of the sequence, n, is suppressed in the notation, in the c o n t e x t / ( ~ ) ,
the sequence is assumed to be of the required length, namely ~ ( f ) . Finally,
if f is a binary termsymbol we also write to f tl or (to f t,) (depending on
readability) rather t h a n f~to~t~. (This is the infix notation.) The set X is
taken in the context of propositional logic to be the set of v a r i a b l e s , some-
times also denoted by var. We write var(t) for the set of variables occurring
in t and sf(t) for the set of subterms or subformulae of t. We define t h e m
formally as follows, vat(t) :-- sf(t) N X, and
:= {xl}
sS(S(t0,... := {S(t0,...
Nullary functions are also referred to as c o n s t a n t s . The cardinality of the set
of constants is usually considered independently of the cardinality of functions.
The cardinality of T m a ( X ) is infinite except in some trivial cases.
PROPOSITION 1.2.1. T m a ( X ) is finite exactly when (i) X is empty, there
are no constants and any number oS n-ary Sunctions Sor n > O, or (ii) X is
finite and there exist finitely many constants and no n-ary Sanctions Sot any
n > O. Is the cardinality os T m a ( X ) is infinite it is equal to the maximum of
~o and the cardinalities oS the set oS variables, the set os constants and the
set oS Sunction symbols.
PROOF. Let C denote the set of constants and F the set of functions of
arity > 0. Let a := ~X + ~C, /3 := ~F and V := ~Tnk~(X). Obviously, if
a = 0, V = 0 as well. If/3 = 0 then V -- a. This is finite if a is. The theorem
holds in all these cases. Now assume that F ¢ ~ and X U C ¢ ~. Define the
sets Si, i < w, as follows.
So := X UC
:= u : c_ I c F}
Then
Tm.(X) : U S{
i<w
If F ¢ ~ and So ¢ ~, then S i ¢ O and also Si C Si+l. Hence V is at least
1%. If So and F are finite, so is Si for every i < w. Hence the theorem holds
in these cases. Finally, let a be infinite. Then a = max{~X, ~C}. It can be
shown t h a t ~Si+l = ~Si"/3 = max{~Si,~}. Hence for i > 0, ~Si = ~$1. So,
= ~ S 1 = max{a,3}, showing the theorem. []
Typically, we think of the function symbols f as standing for functions, taking
~ ( f ) m a n y inputs and yielding a value. This is codified in the notion of an
fl-algebra. (For basic concepts of universal algebra see [37] and [89].) An
ft-algebra is a pair 2[ -- (A, I}, where A is a set, and I a function assigning to
each f C F an f ~ ( f ) - a r y function from A to A. This definition is somewhat

cumbersome, and is replaced by the following, less rigorous definition. An
f ~ - a l g e b r a (or simply an a l g e b r a ) is a pair 92 = (A, ( f ~ : f e F ) ) , where A
is a set, the u n d e r l y i n g s e t o f 92, and for each f C F, f~l is an f ~ ( f ) - a r y
function on A, in symbols f ~ : A n(f) -+ A. One very important f~-algebra is
the algebra of ~ - t e r m s over a given set X. Let us denote by f not only the
function symbol f , but also the function T m ~ ( X ) fl(f) -~ T m ~ ( X ) : (tj : j <
fl(f)) ~-~ f(to, t l , . . . , tfl(f)-l). Now, since the set of f~-terms is closed under
(application of) the functions f for every f E F, the following is well-defined.
Z m u ( X ) := (Trafl(X),F).
This is called the a l g e b r a o f f ~ - t e r m s o v e r X or simply the t e r m a l g e -

b r a ( o v e r X). In addition to the termalgebras we have the trivial algebra
1 := ({0}, ( f l : f E F)), where f l ( 0 , . . . , 0) = 0. In addition to the primitive
functions f ~ , f C F, we can form complex functions by composition of func-
tions; in algebra, for example, + and • are primitive functions, and x . (y + z),
x • y + x • z are complex functions (and the terms are distinct even though
they represent identical functions from R 3 to R). A t e r m - f u n c t i o n of 92 is
now generally defined to be any primitive or complex function of 92. Given a
set A a c l o n e o f t e r m - f u n c t i o n s is a set C l of functions f : A "~ -+ A for
some m < w satisfying
1. For all n < w the projections p~ : A n --+ A of an n-sequence to the
/ - c o m p o n e n t are in C1.
2. For a l l n < w i f f : A k -+ A i s i n C l a n d g i : A '~--+ A a r e i n C 1 for
i < k, then the composition f i g 0 , . . - ,gk-1] : A n -+ A is also in C1. It
is defined by
fig0,..., gk-1](S) := f(g0(~),..., gk_l(s))

where 5 E A%
Let 92 be an ~-algebra. The clone generated by the functions f ~ is called
the c l o n e o f t e r m f u n c t i o n s of 9/, and is denoted by Clo(9/). CIon(9/)
denotes the set of n - d r y t e r m functions of 9/. A p o l y n o m i a l of the algebra
92 is a termfunction of the algebra 9/A, where 9/A denotes the algebra 92
expanded by constants ca with value a for each a • A. We denote by Pol~ (9/)
the set of all n - d r y polynomials of 9/, and by Pol(9/) the set of polynomials
of 91. The elements of Poll(9~) are called t r a n s l a t i o n s . The reader m a y
verify the following fact. Given a polynomial p(~) • Pol~(9/) there is a t e r m
function f ( ~ , ~ ) • Clon+m(9/) for some m • w and some b • A TM such t h a t
p(~) = f ( d , b) for all ~ • A n. This means that every polynomial results from
a t e r m function by supplying constants for some of the arguments.
An ~ - h o m o m o r p h i s m (or simply a h o m o m o r p h i s m ) from the algebra
91 = ( A , ( f ~ : f • F}} to the algebra ~ = ( B , ( f ~ : f • F}} is a m a p
h : A --+ B such t h a t for all f E F and all elements aj C A, j < ~ ( f ) ,
h ( f ~ ( a o , . . . , a n ( / ) - l ) ) = f ~ (h(ao),... , h ( a n ( f ) - l ) ) •
In case h : A -+ B is a h o m o m o r p h i s m we write h : 92 -+ ~ . To rephrase the
formal definition, h o m o m o r p h i s m s are m a p s which preserve the structure of
the source algebra; the source elements compose in the same way in the source
algebra as their images do in the target algebra. A h o m o m o r p h i s m h : 92 --+ 92
is called an e n d o m o r p h i s m of 92. A bijective e n d o m o r p h i s m is called an
automorphism of 91. We write End(92) for the set of e n d o m o r p h i s m s of
92 and Aut(21) for the set of a u t o m o r p h i s m s of 92. End(92) is closed under
composition, and so the endomorphisms form a semigroup with idA as unit.
Moreover, if h : A -+ A is an a u t o m o r p h i s m , so is h -1 : A --+ A.
THEOREM 1.2.2. Let t2 be a signature and 92 an ~-algebra. Put

~nb(92) := (End(92),idA,o),
92ut(92) := (Aut(92),idA, - 1 , o ) .
Then ~n~(92) is a semigroup and 92ut(92) is a group.
A m a p h : A --+ B induces an equivalence relation ker(h) on A via ix, y) E

ker(h) iff h(x) = h(y). We call ker(h) the k e r n e l of h. Given an equivalence
relation 0 , the sets [x]® : = {y : x O y} are called the c o s e t s of (9. For
a set D, we write [D]® : = UxcD[x]O. W i t h ® = ker(h) we call the cosets
also f i b r e s of h. If h : 92 -+ ~ is a h o m o m o r p h i s m , it induces a special
equivalence relation on A, called a congruence. A c o n g r u e n c e ( r e l a t i o n ) on
92 is a set O C_ A × A which is an equivalence relation and for all functions f
and sequences ~ = ( x 0 , . . . , x ~ ( f ) - l ) and ~ = (Y0,... , Y ~ ( / ) - I ) if xj Oyj for
all j < f l ( f ) t h e n also f ~ ( ~ ) O f ~ ( ~ ) . E a c h congruence relation on 92 defines
a so-called f a c t o r a l g e b r a 92/0 whose elements are the cosets [x]O of the
equivalence relation, and the operations [f~](9 act blockwise in the following
way.
::
It is the fact t h a t O is a congruence relation t h a t makes this definition in-

dependent of the choice of the representatives. T h e reader m a y check this
for known cases such as groups, lattices etc. T h e m a p x ~-~ [x]O is an f l -
h o m o m o r p h i s m from 92 onto 92/(9. We write 92 --~ 92/(9 to highlight the fact
that the m a p is surjective. The following is now straightforwardly proved.
PROPOSITION 1.2.3. Let h : 92 ~ ~ be a surjective homomorphism. Then

(9 : = ker(h) is a congruence, and ~ ~- 92/(9. A n isomorphism is given by the
map ~ : h(x) ~-+ Ix](9.
Since h is surjective, ~ is defined on all elements of ~ and is well-defined
by construction of [x](9.
Let E C_ A × A. The smallest congruence relation containing E is denoted

by O ( E ) . The m a p E F-+ O ( E ) is a closure operator with closed sets being
the congruences. The congruences on an algebra form a complete lattice, [-]
being intersection and I I the smallest congruence containing the m e m b e r s
of the union. We denote this lattice by Con(92). The b o t t o m element of
this lattice is the d i a g o n a l A = {{a,a} : a C A}, and the top element is
V = A x A. An Mgebra is s i m p l e if it has only these congruences and they
are distinct. The congruence O ( E ) can be computed explicitly. Let
E ~ := {{t(a),t(b)}: {a,b} C E , t E Po11(92)}.
This is the closure of E under translations. T h e n ®(E) is the smallest equiv-
alence relation containing E ~'. The rationMe behind this is the following
characterization of congruence relations.
PROPOSITION 1.2.4. Let 92 be an algebra. A binary relation on A is a

congruence on 92 iff it is an equivalence relation closed under translations.
PROOF. Clearly, a congruence relation must be an equivalence relation.
Conversely, let O be an equivalence relation and t a translation. T h e n t(x) =
f ( x , d) for a t e r m function f ( x , ~). Assume x O y. Then
t(x) = / ( x , d) o / ( v , = t(y).
Hence O is closed under translations. Conversely, let O be translation closed
and ~ = (xi : i < n}, ~ = (Yi : i < n} be n-long sequences such t h a t xi O y~
for all i < n. Assume t h a t f is a an n - a r y t e r m function. Then we have
o
0 f(yo,Yl,X2,...,Xn-1)
O f(Yo,...,Yn-1).
By transitivity, f ( ~ ) O f(Y0. []
PROPOSITION 1.2.5. Let Q)~, i C I, be congruences of a given algebra 92.

Then let q2 be the transitive closure of the equivalence relation ~Jiel 0 i . Then
is a congruence on 92 and identical to U i e l O i . Hence, (x,y} E [ J ~ i O i iff
there is a number n < w, elements zi, i < n + 1, and a sequence j(i), i < n,
of elements in I such that
X = Z0 0 j ( 0 ) Zl O j ( 1 ) Z2 - . . Z n - 1 0 j ( n - 1 ) Zn -~ y
PROOF. [-JiEI Oi is symmetric and reflexive; so is its transitive closure, q .

Therefore, we only need to verify that q is closed under translations. To see
that, let (x,y} E ~. Then there exist elements z~, i < n + 1, and a sequence
j(i), i < n, of elements of I such t h a t
X -~- Z 0 0 j ( 0 ) Zl O j ( 1 ) Z2 . . - Z n - 1 0 j ( n - - 1 ) Zn = y
Now let f E Poll (9/) be a translation. T h e n

f ( x ) = f(zo) Oj(0) f ( z l ) Oj(1) f(z2) . . . f(Zn-1) Oj(n-1) f ( z n ) = f ( y )
Hence ( f ( x ) , f(y)> e q'. []
We derive the following useful consequence. Let (a, b) E [_] (Oi : i C I ) .

T h e n for some finite set Io C_ I, (a, b) C U (6)i : i C Io). Moreover, for a set E
of equations, O ( E ) = [_J(O(E0): E0 C_ E, ~E0 < iq0).
PROPOSITION 1.2.6. Let 92 be an f~-algebra. Then Con(92) is algebraic.
The compact elements are of the form O(E), E a finite set of equations.
Moreover,
(9(E) = U < O ( E 0 ) : E0 C E , ~E0 < I%)
T e r m algebras have the important property that for any function v : X -+

A where A is the underlying set of 92 there is exactly one ~ - h o m o m o r p h i s m
: ~ m n ( X ) --+ 92 such that ~ I X = v. ~ can be defined inductively as follows.
1. For x E X we have ~(x) = v(x).

2. For every f e F and t e r m s tk (k < f~(f)):
~(f(t0,... , t~(f)-1)) = f ~ ( ~ ( t 0 ) , . . . , ~(t~(/)-l)) •
We note the following useful fact. If h : 92 -~ ~ is a h o m o m o r p h i s m then
h o v : X -~ B, and h o v = h o ~ . Given v : X -+ A and a t e r m t =
t ( x 0 , . . . , x n - 1 ) then ~(t) E A. Hence each t e r m t defines a t e r m - f u n c t i o n
t ~ : A n -+ A on 92 such t h a t t ~ ( a o , . . . , a n _ l ) = ~ ( t ( x o , . . . , x n - 1 ) ) where
v(xi) = ai, i < n. A m a p cr : X --+ Tree(X) is called a s u b s t i t u t i o n .
A substitution defines a unique h o m o m o r p h i s m ~ : T r e e ( X ) --+ Tmf~(X);
conversely, any h o m o m o r p h i s m of this type is d e t e r m i n e d by a substitution.
So, substitutions are simply e n d o m o r p h i s m s of the t e r m algebra. We usually
write t ~ for c~(t). It is also c u s t o m a r y to write t[u/x] or t[x ~+ u] to denote the
result of a p p l y i n g to t the substitution c~ : x ~-+ u, y F-~ y (for all y ~ x). Given
a set V of variables we denote by t[u~/x : x E V] the result of simultaneously
substituting us for x for all x E V.
E x e r c i s e 1. Let A be a set, and E C_ A × A. Show t h a t the least equivalence

relation over A containing E , e ( E ) , can be c o m p u t e d in three steps. Let
r ( E ) = E U A A denote the reflexive closure, s(E) = E U E v the s y m m e t r i c
closure, and t(E) = E + the transitive closure. T h e n e(E) = t(s(r(E))).
E x e r c i s e 2. (Continuing the previous exercise.) Show t h a t if E is reflexive,

that is, E = r ( E ) then so is E~; and t h a t if E is s y m m e t r i c , then so is E ~.
Hence show t h a t O ( E ) can be c o m p u t e d in four steps as follows. (i) close E
reflexively, (ii) close E symmetrically, (iii) close under translations and (iv)
close E transitively.
E x e r c i s e 3. Show T h e o r e m 1.2.2.
1.3. A l g e b r a i c Constructions
We have already encountered the notion of a h o m o m o r p h i s m of 12-algebras
and congruences. Here we will state and prove some e x t r e m e l y useful theo-
rems a b o u t these constructions and also introduce some (more) notation. If
h : 91 -+ ~ is surjective we write h : 91 --- ff~ and call ~ a h o m o m o r p h i c
i m a g e of 91. If h is injective we write h : 91 ~ ~ . F u r t h e r m o r e , if A C B and
the n a t u r a l inclusion h : A -+ B : x ~-+ x is a h o m o m o r p h i s m we say t h a t 91
is a s u b a l g e b r a of ~ and write 91 < if3. If h is b o t h surjective and injective,
it is called an i s o m o r p h i s m . 91 and ~ are i s o m o r p h i c if there exists an
i s o m o r p h i s m h : 91 -+ if3. We know t h a t each h o m o m o r p h i s m from 91 to
induces a congruence on 91 and t h a t each congruence on an algebra is asso-
ciated with a surjective h o m o m o r p h i s m . T h e r e is a o n e - t o - o n e - c o n n e c t i o n
between congruences and the n a t u r a l factor algebras, where we c o m p u t e with
blocks rather t h a n elements. Moreover, the following holds.
PROPOSITION 1.3.1. (1.) Let 91 be an algebra and 01 C 02. Then 02
induces a congruence on 91/01, denoted by 0 2 / 0 1 . Moreover,
(91/ol)/(o2/ol) 91/o2.
(2.) Let h : 91 - , ~ be surjective and let 0 be a congruence on f~. Then there
exists a congruence • on 91 such that 91/~ -~ ~ / 0 . (3.) Let 91 < ~ be a
subalgebra and 8 B be a congruence on ~ . Then OB r A :---- OB N A × A is a
congruence on 92.
PROOF. (1.) 0 2 / 0 1 can be defined as follows. [[x]01](O1/O2) := [x]O2.
This is independent of the choice of x as a representative of the class [x]O1,
since 02 includes ~T. T h e m a p [x]O1 ~-~ [x]O2 is a h o m o m o r p h i s m with kernel
O2/O1, as is i m m e d i a t e l y verified. (2.) P u t q~ := { ( a , b ) : h ( a ) 0 h(b)}. Let
h/rb: [x]~ F-+ [h(x)]O. This is well defined because it does not d e p e n d on the
choice of x as a representative of its class. By definition, h / ~ is injective; it is
also surjective. It is also not hard to see t h a t it is a h o m o m o r p h i s m . (3.) P u t
OA := O e I A. OA is clearly an equivalence relation. Now let t E P o l l ( ~ ) ,
a, b E A. T h e n t(a),t(b) E A, since 91 is a subalgebra. Hence if (a, b) E OA
we have (t(a), t(b)) E OA by the fact t h a t OB is a congruence and 91 is closed
under translations. []
THEOREM 1.3.2. The map • ~-~ ~ / 0 is an isomorphism from the interval
[0, V] in the lattice Con(91) onto the lattice Con(91/0).
PROOF. By the previous proposition this m a p is bijective. It is easy to see
t h a t it is an order isomorphism, thus it is a lattice i s o m o r p h i s m as well. []
1.3. Algebraic Constructions 15
Next we define the product of algebras. Let 92j, j C J, be a family of algebras.

Then the underlying set of the product is the cartesian product of the sets,
X j c j A j , which is the set of functions s with domain J and value s(j) C Aj.
Denote this set by P. The operations are defined pointwise, that is, if f is an
n - d r y function symbol, and s o , . . . , Sn-1 E P , then
f•(so,... , sn-1)(j) := f ~ ¢ ( s o ( j ) , . . . , S n - l ( j ) ) .
Then the p r o d u c t of the 92j is
n 9 2 J := (P' { / ~ : f C F))
jCJ
The projection maps Pi : P --+ Ai : s ~-~ s(i) are homomorphisms. If (gj,
j E J , are congruences on the 92j then there is a natural congruence X j c j (9j
on the product defined by s (Xjej ®j) t iff for all j E J we have s(j) (9j t(j).
For every family of maps hj : 92j --+ ~ j there exists a map
H hi: H 92J II
jeI jCI jCI
If all hj are injective (surjective) then so is n j e z - The kernel of n j e I hj is
exactly X j e i k e r ( h j ) . However, not every congruence on the product can be
obtained in this way. (The easiest example are sets, that is, algebras with no
functions. There a congruence is just an equivalence relation. An equivalence
on a product set is not necessarily decomposable.)
If J¢ is a class of f~-algebras for a fixed f~, then by H(~:) we denote the
class of all algebras q3 which are homomorphic images of algebras in 9£, we
denote by 5(9£) the class of subalgebras of algebras in 9C and by P(K) the
class of algebras isomorphic to products of algebras in K. A v a r i e t y is a class
closed under all three operators H, S and P.
THEOREM 1.3.3. Let f~ be a signature and :K a class of ft-algebras. The
smallest variety containing 9£ is the class HSP(:~).
PROOF. All operators are individually closure operators. Namely, we
have K C O(:X) for all 0 C {H, S, P}. For if 92 is in :X, then since the identity
1~ : 92 --+ 92 is an isomorphism, 92 E S(K) as well as 92 C H(:K). Moreover,
92 is a product of 92 (take a singleton index set). Secondly, if K C_ L then
O(K) _C O(L) for 0 E {S, H, P}, as is immediate from the definition. We
will leave it as an exercise to show that HH(iK) _C H(K), SS(K) _C S(JC) and
PP(iK) _C P(K). Furthermore, 5H(9£) C HS(K). For let ~ < ~ and ~3 ~ 92/0
for some congruence (9. We may assume that ~ = 92/(9. Then C is a set of
blocks of the form [x]O. Put D := {x: Ix](9 e C}. Then [D]O = C. Moreover,
by the fact that C is closed under the operations If](9 it follows that D is closed
under all operations of 92. Hence D < 92. Since we have ho I D : ~ --~ ~,
it follows that D C HS(K). Also, we have noted that PH(K) _C HP(9£).
Finally, PS(gC) C S P ( ~ ) since the product of subalgebras is a subalgebra of

the product of the ~ j as we have noted above. W i t h these c o m m u t a t i o n
laws we have H(HSP(X)) C_ HSP(K), S(HSP(K)) C_ HSSP(K) -- HSP(JC) and
P ( H S P ( ~ ) ) C_ HPSP(9() _C HSPP(K) ---- HSP(gC). This shows the theorem. []
Let A be a cardinal number. An algebra 91 is A - g e n e r a b l e or A - g e n e -
r a t e d if there exist elements a~, /~ < A, such t h a t the smallest subalgebra
containing all these elements is 92 itself. Likewise, since the smallest subalge-
bra containing these elements is the set of all elements obtainable by applying
the functions to these elements, 92 is A-generated iff there exists a surjective
h o m o m o r p h i s m E m ~ ( X ) --~ 92, where X is a set of cardinality A. An algebra
92 is called f r e e l y A - g e n e r a t e d in a class ~: if it there is a set X C A of
cardinality A such that for every m a p v : X -+ B there is exactly one homo-
morphism h : 92 -+ ~ such that h l X = v. We write ~ for h. We say also t h a t
92 is freely generated by X. In the class of all ~-algebras, the t e r m algebras
~ m ~ ( X ) are freely generated by X. (They are also called a b s o l u t e l y f r e e
~-algebras.)
PROPOSITION 1.3.4. Let • be a class of algebras, and let 92 and ~ be
freely A-generated in K. Then 92 ~ f~.
PROOF. Let X C A and Y C_ B be subsets of cardinality A such t h a t 92
is freely generated by X and ~ freely generated by Y. By assumption, there
exist m a p s p : X - + Y a n d q : Y - - + X such t h a t q o p = 1 z a n d p o q = 1y.
Then :~ : 92 -+ ~ and ~ : ~ --~ 92 are (uniquely) defined extensions of p and
q. Moreoever, p o q r Y = 1y, and so p o q = lm, since there is exactly one
h o m o m o r p h i s m extending 1y, and the identity is a homomorphism. Likewise
q o p = 1~ is proved. Hence 92 and ~ are isomorphic. []
The previous theorem has established the uniqueness of free algebras.
Their existence is not generally guaranteed. To have free algebras for all
cardinalities of generating sets is a nontrivial property. A central theorem of
universal algebra states that all nontrivial varieties have free algebras. T h e
proof looks difficult, but the argument is simple. Take an cardinal A and
consider all pairs (f, 92) where 92 E ~: and f : A --+ A. Let S be the set of such
pairs. It is used as an index set in the product
~:= 1-I 92
(f,~)~s
Let C be the subalgebra generated by the functions s~, with # E A, where
s , ( ( f , 92)) = f ( # ) . L e t ~ : ~ ~ ~ be the inclusion map. We claim t h a t ff is
freely generated by the s~. To t h a t end, let v : s~ ~-+ a~ be any m a p into an
algebra 92 C K. Then let g be defined by g(#) := v(s,). The pair c = (g, 92) is
in S. Hence there is a projection p~ : ~ --+ 92 such that p~(s,) = s~(c) = a~.
The composition p~ o L is a homomorphism from ¢ to 92. By ~ r v (X) we denote
1.3. Algebraic Constructions 17
the algebra freely generated by X in V; and for a cardinal number A we denote

by ~rv()~) the freely )~-generated algebra in V. Given a set X of generators,
and a variety ~?, there may be several terms denoting the same element in the
free algebra ~ c v ( X ) , since it is in general a homomorphic image of the t e r m
algebra E m n ( X ) . Nevertheless, we will not distinguish between a t e r m t(~)
and its equivalence class in ~ r v ( X ) .
THEOREM 1.3.5. Let ~ be a variety of 12-algebras for a given 12. For
every cardinal 7 there exists in V a freely 7-generated algebra. Every algebra
of V is the homomorphic image of a free algebra in V.
The reader m a y care to note that it may happen t h a t the variety is trivial,
containing up to isomorphism only the algebra 1. In that case even though
we start off with a set of larger cardinality, the free algebra will be isomorphic
to 1, so the generators turn out to be equal as elements of the algebra. If
we insist on the generators as being elements of the free algebra then the
previous theorem is false just in case we have a trivial variety. However,
under a different reading the theorem makes sense, namely if we take the
following definition of a free algebra. An algebra 9/is free over X if there is a
m a p i : X -+ A such t h a t for any m a p j : X --+ B there is a h o m o m o r p h i s m
h : 9/ --+ ~ for which j ( x ) = h o i ( x ) for all x C X. In the case of the
trivial variety, X can have any cardinality, and yet A = {0}. In the sequel we
always assume to work with nontrivial varieties, whenever this should make
a difference.
THEOREM 1.3.6. Let V be a variety, i : X ,-~ Y and p : Y --, X . Then
~: ~ r v ( Z ) --* ~ r v ( Y ) and ~ : ~ r v ( Y ) --- ~ r v ( X ) .
PROOF. If i : X --~ Y there exists a q such t h a t q o i = 1x. It follows
that q o i is the identity on ~ : v ( X ) . Since q o i = ~o~, ~ is surjective and ~ is
injective. Similarly it is shown that ~ is surjective. []
PROPOSITION 1.3.7. Let V be a variety and g/ an algebra of~?. Then there
exists a free algebra ~ and a homomorphism h : ~ --, 9/.
For a proof note t h a t there exists a surjection v : A --- A. Hence we also have
that ~ : ~ r v (A) --- 92.
E x e r c i s e 4. Show that either a variety contains up to isomorphism only the

trivial algebra 1 or it contains infinite algebras.
E x e r c i s e 5. Let L be a language with signature ~. Let 7 be the cardinality

of T m n ( ~ ) . Then show that any variety of ~ - a l g e b r a s which is nontrivial
contains an algebra of any infinite cardinality ~ > 7.
E x e r c i s e 6. Show with a specific example that the claim of the previous

exercise need not hold for finite 5.
E x e r c i s e 7. Show that for any class of algebras, HH(gC) C H(:E) as well as

SS(gC) C_ S ( ~ ) and PP(X) C_ P ( ~ ) . Hint. For the last claim take a collection
(Ij : j E J) of pairwise disjoint index sets, and assume that there is an algebra
P2i~ for every ij E I j , j E J. T h e n each ~3j := 1-IkEb ~k is a product, and
:= I-[jEJ ~3j is a general element of PP(:K). Now let K := [-JjEJ/J and put
:= HkEK Q~k" Show that ¢ ~ 3 .
1.4. G e n e r a l L o g i c
In our view, logic is the study of t r u t h and consequence. In logic we study
(among other things) whether a statement T follows from some set A of other
statements. We usually write A ~- T if this is the case. We interpret this as
follows: if all X E A are true then so is T. Of course, we must specify what we
mean by being true. However, already on these assumptions there are some
nontrivial things that can be said about the relation ~-. To write t h e m down,
we will - - in accordance with our notation in connection with modal logic - -
use lower case Greek letters for terms of propositional logic, since these t e r m s
are thought of as formulae. We also will henceforth not distinguish between
L as a set of function symbols and the terms of L, namely the set Tmn(var);
given this convention ~-C_ p ( L ) × L. Moreover, we write E ~- F if for all
E F, E ~- ~. It is also customary to use E; A for E U A and E; T instead of
E [J {~}. This notation saves brackets and is almost exclusively used instead
of the proper set notation.
(ext.) If ~ E E then E ~- ~.
(mon.) If E C_ A then E ~- ~ implies A ~- ~.
(trs.) If E ~- F and F ~- ~ then E [- ~.
(Observe t h a t (mon.) is derivable from (ext.) and (trs.).) For suppose t h a t
E E. T h e n if all X E E are true, then ~ is true as well. Thus (ext.) holds.
Furthermore, if E ~- ~ and E C A and if all X E A are true, then all t e r m s
of E are true and so ~ is true as well; this shows (mon.). The third rule is
proved thus. If E ~- F and A ~- ~ and all terms of E are true then all formulae
of F are true by the first assumption and so T is true by the second.
In addition, there are two other postulates that do not follow directly
from our intuitions about truth-preservation.
(sub.) If E ~- ~ and c~ is a substitution then E ~ ~- ~ .
(cmp.) E ~- ~ iff there exists a finite E0 C_ E such t h a t E0 ~- ~.
The postulate (sub.) reflects our understanding of the notion of a variable. A
variable is seen here as a name of an arbitrary (concrete) proposition and thus
we m a y plug in all concrete things over which the variables range. T h e n the
1.4. General Logic 19
relation E F- ~ says t h a t for any concrete instances of the occurring variables,

the concretization of E ~- ~ is valid. So, the rule p A q ~- q A p - - being valid
- should r e m a i n valid under all concretizations. For example, we should
-
have Aristotle was a philosopher and Socrates was a carpenter ~- Socrates was
a carpenter and Aristotle was a philosopher. Suppose now t h a t we have a
substitution c~. T h e n - - contrary to what one might think - - ~ is not a
concretization of ~; however, every concretization o f t ~ is a concretization of
~. One m a y therefore think of ~r as a sharpening, since given T it returns a
more specific proposition ~ . Thus, E ~ ~- ~ simply is a s t a t e m e n t over a
subset of concretizations of E ~- ~. Since it holds for all of t h e m E ~ ~- T~
holds as well. T h e algebraic a r g u m e n t is simple. Namely, a concretization is,
formally speaking, just a h o m o m o r p h i s m / 3 : T m a ( X ) --+ 92. A concretization
of ~ is just the value/3(~(~)) under a h o m o m o r p h i s m . B u t / 3 ( ~ a) --/3ocr(~),
and since /3 o cr : ! m ~ ( X ) -+ 92, it is a concretization as well. Hence any
concretization of ~ is in fact a concretization of ~.
T h e last p o s t u l a t e (crop.) is called c o m p a c t n e s s , a t e r m borrowed from
topology. A n o t h e r t e r m is f i n i t a r y . (crop.) is not at all justifiable from our
notions of truth. In fact, it fails e. g. in logics with infinitary operations. We
have {~i : i e ~} ~- A ( ~ i : i • w) even t h o u g h no finite set F C {~i : i • w}
exists such t h a t F ~- A(Ti : i • w}. One might expect t h a t if we only have
finitary operations, the t r u t h of any t e r m if entailed by E will already be
entailed by some finite subset of E. But this is known to be false as well (see
[232]). Indeed, the most plausible justification comes from the i n t e r p r e t a t i o n
via deduction. For suppose we read E ~- ~, as there is a proof of ~ from E.
Then, since such a p r o o f is a finite object, we can use only finitely m a n y t e r m s
of E in it.
DEFINITION 1.4.1. Let L be a propositional language and F-C p ( C L ) × L.

~- is called a f i n i t a r y c o n s e q u e n c e r e l a t i o n over L if it satisfies (ext.),
(mon.), (trs.), (sub.) and (cmp.). A pair (L, F-} where L is a propositional
language and ~- a finitary consequence relation over L is called a ( p r o p o s i -
t i o n a l ) logic.
Notice the following. In contrast to the s t a n d a r d literature, we require

b o t h (sub.) and (cmp.), since we want to deal almost exclusively with such
logics. If (cmp.) is not assumed, we explicitly say so. Hence, in sequel, by a
consequence relation we actually u n d e r s t a n d a finitary consequence relation.
Moreover, when there is no risk of confusion we will speak of the l o g i c ~-; this is
justified especially when it is clear over which language ~- is defined. A general
reference for consequence relations is [232]. A logic ~- is called i n c o n s i s t e n t
if F- = p ( T m a ( v a r ) ) × T m a ( v a r ) . Alternatively, ~- is inconsistent if ~- p for
some p. It follows t h a t in an inconsistent logic F ~- T for all F and T. Likewise,
a set F (a formula T) is c o n s i s t e n t if there is a formula X such t h a t F K X
20 1. A l g e b r a ~ L o g i c a n d D e d u c t i o n
(~ Y X). T h e terms ~ satisfying ~ ~- p are called the t a u t o l o g i e s of the logic.

We denote the set of tautologies of ~- by Taut(~-).
A r u l e is a pair p = (E, p) where E is a finite set. E is called the s e t o f
p r e m i s e s o f p and ~ the c o n c l u s i o n . If ~E = n, p is called an n - a r y rule.
Furthermore, p is a p r o p e r rule if n > 0, and an a x i o m otherwise, p is a
d e r i v e d r u l e of ~- if p E ~-.
DEFINITION 1.4.2. Let L be a language and R a set of rules over L . T h e n

by ~_n we denote the least consequence relation F- such that R C_~-. R is a
r u l e base of ~- if F- = F-R.
To characterize ~_R define an i n s t a n c e of p = (E, ~} to be any pair

(A, ¢ / such t h a t there is a substitution c~ with ~ = ¢ and E ~ = A. E v e r y
rule justifies all instances of itself, by (sub.). Suppose t h a t a set of rules R
is given. T h e n E [_R ~ ii~ there is a proof tree deriving ~ from E. A p r o o f
t r e e of ~ from E is a tree (T, <} (where the root is the largest element with
respect to < and < is completely intransitive) labelled with formulae f r o m L
in such a way t h a t (i) the root has label ~, (ii) the leaves have labels from E
and (iii) if {Yi : i < n} is the set of elements > x, Yi has label ~i and x has
label ¢ , t h e n ({~i : i < n}, ¢} is an instance of some p • R.
A n alternative characterization is via sequences. Given R we define now
an R - p r o o f f r o m E t o ~ to be any finite sequence (Ti : i < A/ such t h a t
~ = ~ and for any n _ A either (i) ~ • E or (ii) for some set A such t h a t
A C_ ( ~ , : p < ~}, ( A , ~ } is an instance of a rule in R. Now set E ~-R ~ iff
there exists an R - p r o o f from E to ~. T h e reader m a y verify t h a t if we have
a proof tree, there exists an e n u m e r a t i o n of the nodes of a tree such t h a t the
corresponding labels, if written down according to t h a t enumeration, form an
R - p r o o f ; and t h a t if we have an R - p r o o f of ~ from E, we can define a p r o o f
tree for p from E. To see the correctness of this characterization of ~_R, we
prove the following theorem.
THEOREM 1.4.3. Let R be a set of rules and put E IF- ~ i f f there exists an
R - p r o o f of ~ f r o m E. Then I~- = F-R.
PROOF. It is not hard to see t h a t 1~- C ~_R. To show t h a t the two are
equal it suffices to establish t h a t l~- as defined is a consequence relation. (ext.)
If T • E t h e n the sequence consisting of ~ alone is a R - p r o o f of T from E.
(mon.) If c~ is a R - p r o o f of T from E t h e n it is also an R - p r o o f of T f r o m
any A D E. (sub.) If (~ is an R - p r o o f of ~ from E and a is a s u b s t i t u t i o n
then c~~ is an R - p r o o f of ~ from E ~. (cmp.) If G is an R - p r o o f of T from E
t h e n let E0 be the set of terms occurring b o t h in E and c~. Since c~ is finite,
so is E0. Moreover, c~ is a R - p r o o f of ~ from E0. (trs.) Suppose t h a t E I~- F
and t h a t F I~- ~. By (cmp.) we can assume t h a t F is finite, so w i t h o u t toss of
generality let F = {7i : i < n}. Let 5 be an R - p r o o f of ~ from F and fii an
1.4. General Logic 21
R - p r o o f of ¢i from E, i < n. Now let
It is straightforward to check that ( i s a proof of ~ from E. []

Let R be a set of rules. We say that a rule p can be derived from R if
p E F R. Alternatively, p = (A, ~) then p E F R if there exists an R - p r o o f of
from A. If p is a rule, let F +p denote the least consequence relation containing
both F and p. For a consistent F put
E(F):={n: there is a n n - a r y r u l e p eF such that F + p p }
E(F) contains the arity of all rules p such that F +p is a proper, consistent
extension of F. We call E ( F ) the e x t e n d e r s e t of F. Obviously, a conse-
quence relation is maxima] among the consistent consequences relations iff its
extender set is empty.
Let E be a set of formulae. A rule (A, p) is called a d m i s s i b l e for E if
E is closed under the application of p. T h a t is, if for some substitution cr,
cr[A] C E, then cr(~) E E. p is a d m i s s i b l e for F if p is admissible for Taut(F).
Equivalently, p is admissible for F if for all substitutions or, ~ is a tautology,
that is, ~ FR ~ , whenever all members of A ~ are tautologies. F is called
s t r u c t u r a l l y c o m p l e t e if every admissible rule of F is derivable. F is called
P o s t - c o m p l e t e if 0 ¢ E(F).
PROPOSITION 1.4.4 (Tokarz). (1) A consequence relation F is structurally
complete iff E(F ) C_ {0}. (2) A consequence relation is maximally consistent
if] it is both structurally complete and Post-complete.
PROOF. (2) follows immediately from (1). So, we show only (1). Let
F be structurally complete. Let p be a rule such that F +p is a proper and
consistent extension of F. Since F is structurally complete, the tautologies are
closed under p. It follows that p must be a 0 - a r y rule. For the other direction
assume t h a t F is structurally incomplete. Then there exists some p which is
admissible but not derivable, p is not an axiom. Hence F +p properly extends
F. Since F is consistent, p is not a tautology of F. Since p is admissible, p is
not a tautology of F +p, and so the latter is also consistent. []
It will be proved later that 2-valued logics with 7 , -~ and /~ is b o t h

structurally complete and Post-complete. Now, given F, let
T(F) := {F': Taut(F') = Taut(F)}
Then T(F) is an interval with respect to set inclusion. Namely, the least
element is the least consequence containing all tautologies of F. The largest
is the consequence F +R, where R is the set of all rules admissible for F. As
we will see in the context of modal logic, the cardinality of T(F) can be very
large (up to 2 ~° for countable languages).
Another characterization of logics is via consequence operations or via

theories. This goes as follows. Let (L, F) be a logic. Write E e = {~ : E F ~}.
The map E ~-+ E e is an operation satisfying the following postulates.
(ext.) E C_ E ~-.
(mon.) E C A implies E ~- C A ~-.
(trs.) E H- C E ~-.
(sub.) a l e ~] C_ (eriE]) ~ for every substitution a.
(crop.) E e = U ( E ~ : E0 C_ E, E0 finite}.
Actually, given (cmp.), (mon.) can be derived. Thus the map E ~-~ E ~ is a
closure operator which is compact and satisfies (sub.). This correspondence is
exact. Whenever an operation Cn : ~(L) --+ p(L) on sets of L - t e r m s satisfies
these postulates, the relation E Fen ~ defined by E Fen ~ iff ~ C Ca(E)
is a consequence relation, that is to say, (L, ben} is a logic. Moreover, two
distinct consequence operations determine distinct consequence relations and
distinct consequence relations give rise to distinct consequence operations.
Finally, call any set of the form E ~ a t h e o r y of F. By (trs.), theories
are closed under consequence, that is, E e not only contains all consequences
of E but also all of its own consequences as well. We may therefore say that
the theories of F coincide with the d e d u c t i v e l y c l o s e d sets. Given F, the
following can be said about F-theories.
(top.) L is a F-theory.
(int.) If T~, i E I, are F-theories, so is ~ ( T i : i C I).
(sub.) If T is a F-theory, a a substitution then
cr-l[T] = { ~ : qo~ • T} is a F-theory.
(cmp.) If Ti, i C I, is an ascending chain of F-theories then
U(T~ : i • I) is a b-theory.
Again, the correspondence is exact. Any collection T of subsets of L satisfying
(top.), (int.), (sub.) and (cmp.) defines a consequence operation E ~+ E ~ by
E ~ = A ( T : T e T , T _D E) which satisfies (ext.), (mon.), (trs.), (sub.) and
(cmp.). The correspondence is biunique. Different consequence operations
yield different sets of theories and different collections of theories yield differ-
ent consequence operations. The theories of a logic form a lattice. This lattice
is algebraic if the consequence relation is finitary. The converse does not hold;
this has been shown by BURGHARD HERRMANN and FRANK WOLTER [103].
E x e r c i s e 8. Give a detailed proof of Theorem 1.4.3.
E x e r c i s e 9. Let R be a set of axioms or l - d r y rules. Show that A F R ~ iff

there exists a 6 ~ A such that 6 F R ~.
1.5. Completeness of Matrix Semantics 23
E x e r c i s e 10. Show that every consistent logic is contained in a P o s t -

complete logic. Hint. You need Zorn's L e m m a here. For readers unfamiliar
with it, we will prove later Tukey's Lemma, which will give rise to a very short
proof for finitary logics.
E x e r c i s e 11. Show that in 2-valued logic
Thus if ~ =- ¢ is defined by F ~ +4 ¢, then = is a congruence relation. W h a t

is the cardinality of a congruence class? Hint. We assume that we have 1%
m a n y propositional variables. Show that all congruence classes must have
equal cardinality.
E x e r c i s e 12. Show that there is no t e r m p in T, V, A such t h a t ~ F -~p F

in classical logic. Hint. Show first that one can assume var(~) = {p}.
E x e r c i s e 13. Show that if (L, F) is a logic, the m a p E ~-+ E e is a finitary

closure operator.
E x e r c i s e 14. Show that if E ~-+ E e is a finitary closure operator, the system

of deductively closed sets satisfies (top.), (int.), (sub.) and (amp.).
E x e r c i s e 15. Show that if the system of F-closed sets satisfies (top.), (int.),
(sub.) and (cmp.), then (L, F) is a logic.
1.5. C o m p l e t e n e s s of Matrix Semantics

Fundamental for the study of algebraic logic is the notion of a logical
matrix. While for algebraic purposes we need only an algebra in order to
compute terms, truth is extraneous to the notion of an algebra. Boolean
algebras as such are neutral with respect to the notion of truth, we must
stipulate those elements which we consider as true. The link between 1 and
true is conventionally laid. One must be aware, therefore, t h a t this is just a
convention. We might, for example, consider 0 rather t h a n 1 as true, and it
turns out, t h a t the logic of the algebra (2, A) where 0 is considered true is the
same as the logic of (2, V) where 1 is considered true, if A is translated as V.
DEFINITION 1.5.1. A n ~ - m a t r i w for a signature ~ is a pair ~ = (92, D)

where 92 is an ~-algebra and D C A a subset. A is called the set of t r u t h
values and D the set of d e s i g n a t e d t r u t h values. A n a s s i g n m e n t or a
v a l u a t i o n into 9~ is a map v from the set of variables into A. We say that
v makes ~ t r u e in if2 i f ~ ( ~ ) C D; otherwise we say, it makes ~ f a l s e .
With respect to a matrix ff)~ we can define a relation ~-~ by

A ~-~ T ~ for all assignments v : If ~[A] C D then ~(T) C D .
Given a class S of matrices (for the same signature) we define
THEOREM 1.5.2. Let ~ be a signature. For each class S of ~-matrices,

'--s is a (possibly nonfinitary) logic.
PROOF. We show this for a single matrix. The full theorem follows from
the fact that the intersection of logics is a logic again. Let if2 E S. (ext.)
I f ~ E E andS[El C_ D t h e n ~ ( ~ ) E D, by assumption. (mon.) Let E C A
and E ~-~ p. Assume ~[A] C D. Then ~[E] C_ D as well, and so ~(~) E D,
by assumption. (trs.) Let E ~-~ F and F ~-~ ~. Assume ~[E] C D. T h e n
we h a v e , I F ] C D and so ~(p) E D. (sub.) Assume E ~-~ p and let cr be
a substitution. Then ~ o a is a homomorphism into the algebra underlying
Kit a n d ~ o c r [ E ] = ~[E~]. Hence i f ~ [ E ~] C_ D we also h a v e ~ ( ~ ) E D, as
required. []
THEOREM 1.5.3 (W6jcicki). For each logic (L, ~-) there exists a class S
oS matrices such that ~- = ~-s.
PROOF. Given the language, let S consist of all (~m~(var),T) where T
is a theory of %. First we show that for each such matrix if2, ~- C_ kv~. To
that end, assume E ~- ~ and that ~[E] C_ T. Now ~ is in fact a substitution,
and T is deductively closed, and so ~(~) E T as well, as required. Now
assume E ~ ~. We have to find a single matrix if2 of this form such that
E tz~ ~. For example, if2 := (~ma(var), E~). Then with ~ the identity map,
~[E] = E C E ~. However,~(~) = ~ • E ~ by definition o f E k and the fact
that E ~ ~. []
We add the remark that if 9Y~ is a matrix for ~-, then the set of t r u t h
values must be closed under the rules. The previous theorem can be refined
somewhat. Let ~ = (~, D) be a logical matrix, and (9 a congruence on 92. (9
is called a m a t r i x c o n g r u e n c e if D is a union of (9-blocks, that is, if x E D
then Ix](9 C_ D, and likewise, i f x ~ D then [ x ] ( g N D = ~. Then we can
reduce the whole matrix by (9 and define ff2/(9 := (92/(9, D/(9).
LEMMA 1.5.4. Let if2 be a matrix and (9 be a matrix congruence of 9~.
Then ~-~ = F ~ / ~ .
PROOF. Let E ~-~ p. Let v : var --+ A/(9 be a valuation such that
~[E] C_ [D](9. Then let w : var -+ A be defined by taking w(p~) E v(pi).
(Recall that v(p~) is a union of (9-blocks.) By assumption, [w[E]](9 C_ [0](9,
since [D](9 is a union of blocks, and (9 is a congruence. Hence w(~) ~ D, by
which ~(F) = [~(~)](9 e [D](9, as required. Now assume E ~-~/o ~. Let w
1.5. Completeness of Matrix Semantics 25
be a valuation such that ~[E] C D. Then define v to be the composition of

with the natural surjection x ~-+ [x](9. Then v[E] C [D](9. By assumption,
v(qo) e [D](9, so t h a t v(qo) = [x](9 for some x e D. Consider ~(qo). We know
that v(~o) = [~(W)](9 = [x](9. Thus N(~) = y for some y (9 x. Since D consists
of entire (9-blocks, y E D, as required. []
Call a matrix r e d u c e d if the diagonal, that is the relation A = {(x,x) :

x E A}, is the only matrix congruence. It follows that we can sharpen the
Theorem 1.5.3 to the following
THEOREM 1.5.5. For each logic (L,F-) there exists a class S of reduced
matrices such that ~ = ~ s.
Let S be a class of R-matrices. S is called a u n i t a l s e m a n t i c s for ~- if
~- = ~-~ and for all (P2, D) C S we have ~D < 1. (See JANUSZ CZELAKOWSKI
[49, 50]. A unital semantics is often called algebraic. This, however, is
different from the notion of 'algebraic' discussed in W i g BLOK and DON
PICOZZI [29].) The following is a useful fact, which is not hard to verify.
PROPOSITION 1.5.6. Let ~- have a unital semantics. Then in F- the rules

p; q; ~(p) F- ~(q) are valid for all formulae ~.
Notice that when a logic over a language L is given and an algebra 92
with appropriate signature, the set of designated truth values must always be
a deductively closed set, otherwise the resulting matrix is not a matrix for the
logic. A theory is c o n s i s t e n t if it is not the entire language, and m a x i m a l l y
c o n s i s t e n t if it is maximal in the set of consistent theories. One can show
that each theory is contained in a maximally consistent theory. A direct proof
for modal logics will be given below. It is interesting to note t h a t for classical
logics the construction in the proof of Theorem 1.5.3 can be strengthened by
taking as matrices in S those containing only maximally consistent theories.
For if E )/to then E; ~ is consistent and so for some maximally consistent A
containing E we have ~ C A. Taking v to be the identity, ~[E] = E C A,
but ~(~) ~ A, otherwise A is not consistent. Furthermore, there is a special
matrix, ~au~ = (~m~(var), ~ ) . Recall that ~ e are simply the tautologies of
a logic.
THEOREM 1.5.7 (WSjcicki). ~- is structurally complete iff F- = ~-~aut.

Notes on this section. The concepts of logical consequence and logical
matrix are said to date back to the work of JAN LUKASIEWICZ and ALFRED
TARSKI [141]. Many results of this section are claimed to have been folklore
in the 1930ies. Theorem 1.5.3 is due to RYSZARD WOJCICKI [229]. The
converse of the implication in Proposition 1.5.6 also holds on condition that
the logic has tautologies. This is proved in [50], where it is attributed to
unpublished work by ROMAN SUSZKO. The notion of structural completeness
has been introduced by W. A. POGORZELSKI [162] who proved also that

classical logic is structurally complete. For general reference on consequence
relations and algebraic semantics see RYSZARD WOJCICKI [232].
E x e r c i s e 16. Prove Proposition 1.5.7.
E x e r c i s e 17. Characterize ~-~ where 0Jr = (9/,D), where D = Z or D = A.
1.6. Properties of Logics

Logics can have a number of highly desirable properties which one should
always establish first whenever possible. The first is decidability. A logic
(L,~-) is said to be d e c i d a b l e if for all finite E and all terms ~o we can
decide whether or not E ~- ~. In other words, we must have a procedure or an
algorithm that yields a (correct) answer to the problem 'E ~- ~'. (To be exact,
we should speak of the problem 'E ~- ~o?', but the question mark generally
will be omitted.) This comprises two things (i) the algorithm terminates, that
is, does not run forever and (ii) the answer given is correct. Predicate logic
is not decidable; the reason is that its expressive power is too strong. Many
propositional logics, on the other hand, are decidable. 2-valued logics are an
example. This follows from the next theorem, first shown by R. HARROP in
[100].
THEOREM 1.6.1 (Harrop). Suppose that ~ = (92, D} is a finite logical
matrix. Then (L, ~-~ } is decidable.
PROOF. Basically, the decision algorithm is that of building ' t r u t h tables'.
Given a finite E and a term ~, there exist only finitely many valuations
v : var(E) U var(~) --+ T. It is a finite problem to compute all the values ~(¢)
for ~b C E to check whether ~[E] C_C_D and then to see whether ~(~) e D as
well. []
This procedure is generally slower than tableaux-methods, however only
mildly so (see [51]). Tableaux-methods allow for a guided search for falsifying
assignments which in many cases (and certainly in many applications) reduces
the search space rather drastically. However, the truth-table m e t h o d is in
certain cases also rather efficient. There is namely a certain tradeoff between
the length of a formula and the number of variables it contains. The length
of a computation for a given formula T depends exponentially on the number
of variables (so this is indeed expensive), but only quadratically on the length
of ~. (See Section 1.8.) For if ~ has n variables, and 9Jr has k elements,
then k n assignments need to be checked. Given a particular assignment, the
truth value of ~ with respect to that assignment can be checked simply by
induction on the constitution of ~. If ~ contains ~ many symbols different
from variables, then ~ many steps need to be performed. Each step takes
1.6. P r o p e r t i e s of Logics 27
time p r o p o t i o n a l to the length of ~o. In total we get a b o u n d on the t i m e of

c.e.t l.kn
Secondly, we investigate the notion of implication. Of p a r t i c u l a r im-
p o r t a n c e in logic are the modus ponens and the deduction theorem. To ex-
plain them, assume t h a t we have a binary t e r m f u n c t i o n ~ (p,q), w r i t t e n
p -~ q. T h e rule of m o d u s p o n e n s for ~ - - ( m p - - . ) for short - - is the
rule ({p, p --~ q}, q). T h e r e are m a n y connectives which fulfill m o d u s ponens,
for e x a m p l e A and -~. We write (mp.) for (mp--+.). -~ is said to satisfy the
d e d u c t i o n t h e o r e m with respect to F- if for all E, ~, ¢
(t) ¢,
A logic (L, k> is said to a d m i t a d e d u c t i o n t h e o r e m if there exists a t e r m
p --~ q such t h a t -~ satisfies m o d u s ponens and (t) holds. Given the deduction
t h e o r e m it is possible to t r a n s f o r m any rule different from (mp.) into an axiom
preserving the consequence relation. (To be precise, we can also rewrite (mp.)
into an axiom, but we are not allowed to replace it by t h a t axiom, while with
any other rule this is possible in presence of the deduction t h e o r e m and (mp.).)
For example, the rule p; q ~- p A q can be t r a n s f o r m e d into F- p --~ (q --~ (pA q)).
Hence it is possible to replace the original rule calculus by a calculus where
m o d u s ponens is the only rule which is not an axiom. In such calculi, which
are called r a p - c a l c u l i or also H i l b e r t - s t y l e c a l c u l i for --% validity of the
deduction t h e o r e m is equivalent to the validity of certain rules.
THEOREM 1.6.2. A n mp-calculus for -% (L, F-), has a deduction theorem
for ~ iff ~ satisfies modus ponens and the following are axioms of F-:
(wk.) p (q p),
(fd.) (p ~ (q --~ r)) --- ((p --~ q) --~ (p --~ r ) ) .
R e m a r k . (wk.) is the axiom of weakening and (fd.) is known as Freges
Dreierschlufl, n a m e d after GOTTLOB FREGE. T h e r e is also a r a t h e r conve-
nient n o t a t i o n of formulae using dots; it is used to save brackets. We write
~. --~ .~ for (~) ~ (~). For example, (fd.) can now be w r i t t e n as
(p. --- .q --- r) --- (p --~ q. --- .p --- r).
Now we prove T h e o r e m 1.6.2.
PROOF. ( 3 ) Suppose b o t h m o d u s ponens and (t) hold for ---. Now since
~p ~- qo, also ~p; ¢ b ~ and (by (t)) also ~ U ¢ ~ ~, and (again by (t))
b ~ --. (~b --~ T). For (fd.) note t h a t the following sequence
proves ¢. -~ .¢ --~ X; ~ --~ ¢; ~ ? X. A p p l y (t) three times and (fd.) is proved.

( ~ ) By induction on the length of an R - p r o o f d of ¢ from E U {~} we show
t h a t E ~- ~ --~ ¢. Suppose the length of d is 1. T h e n ¢ E E U {~}. T h e r e are
two cases: (1) ¢ E E. T h e n observe t h a t (¢ -~ (~ ~ ¢ ) , ¢, p -~ ¢> is a p r o o f
of ~ --~ ¢ from E. (2) ¢ = ~. T h e n we have to show t h a t E ~- %p--~ ~. Now

observe t h a t the following is an instance of (fd.): (~. --~ .(¢ -~ ~) --~ ~) --~
(~ --~ ( ¢ -~ ~). -~ .~ --~ ~). B u t ~. --~ .(¢ -~ ~) -~ ~ and ~. --~ .¢ --~ ~ are
b o t h instances of (wk.) and by applying m o d u s ponens two times we prove
%v--~ ~. Now let d be of length > 1. T h e n we m a y assume t h a t ¢ is o b t a i n e d
by an application of m o d u s ponens from some formulae X and X --~ ¢- T h u s
the p r o o f looks as follows:
... ,x,... ,x -~ ¢,... ,¢,...
Now by induction hypothesis E ~- ~ -~ X and E ~- ~. --~ .(X --~ ¢ ) . T h e n , as

-~ (X --~ ¢ ) . ~ .(%v --~ X. -~ .~ --~ 9 ) is a theorem we get t h a t E ~- ~ -~ ¢
with two applications of m o d u s ponens. []
T h e significance of the deduction t h e o r e m lies a m o n g other in the fact

t h a t for a given set E there exists at most one consequence relation F- with
a deduction t h e o r e m such t h a t E is the set of tautologies of ~-. For assume
A ~- ~ for a set A C_ L. T h e n by compactness there exists a finite set A0 C A
such t h a t A0 ~- ~. Let A0 : = {5i : i < n}. P u t
d e d ( A o , ~ ) : = 50 -~ (51 - + . . . (5n-1 - " ~ ) . . . )

Then, by the deduction t h e o r e m for -~
A ~- ~ ~=~ ~ ~- ded(A, ~)
THEOREM 1.6.3. Let F- and F' be consequence relations with Taut(~-) =

Taut(~-~). Suppose that there exists a binary termfunction -~ such that ~- and
F-' satisfy a deduction theorem for -~. Then ~ = ~-'.
~- has i n t e r p o l a t i o n if whenever ~ ~- ~b there exists a formula X w i t h

v a r ( x ) C_ var(~) ~ v a r ( ¢ ) such t h a t b o t h ~ ~- ) / a n d X ~- ¢. I n t e r p o l a t i o n is a
rather strong property, and generally logics fail to have interpolation. T h e r e
is a rather simple t h e o r e m which allows to prove interpolation for logics based
on a finite matrix. Say t h a t ~- has a c o n j u n c t i o n if there is a t e r m p A q such
t h a t the following are derivable rules: ({p, q}, p A q} and b o t h ({p A q}, p} and
({p A q}, q}. In addition, if ~- = t - ~ for some logical m a t r i x we say t h a t t- has
all c o n s t a n t s if for each t E T there exists a nullary t e r m - f u n c t i o n t such
t h a t for all valuations v ~(t) = t. (Note t h a t since var(t) = ;~ the value of t
does not depend at all on v.) This rather complicated definition allows t h a t
we do not need to have a constant for each t r u t h - v a l u e ; it is e n o u g h if t h e y
are definable from the others. For example in classical logic we m a y have only
-V = 1 as a primitive and then 0_ = ~ T . Notice also the following. Say t h a t
an algebra is f u n c t i o n a l l y c o m p l e t e if every function A N --+ A is a t e r m -
function of ~; and say t h a t ff is p o l y n o m i a l l y c o m p l e t e if every f u n c t i o n
A n --+ A is a polynomial function. T h e n any functionally complete algebra
1.6. Properties of Logics 29
is polynomially complete; the converse need not hold. However, if ~ has all
constants, then it is functionally complete iff it is polynomially complete.
THEOREM 1.6.4. Suppose that 9Jr is a finite logical matrix. Suppose that
~-~ has a conjunction A and all constants; then F-~ has interpolation.
PROOF. Suppose that ~(iff, q-) F- ¢ ( ~ ~'), where ~ = (ri : i < n). Clearly,
var(¢) ~ var(~), iff n > 0. We show that if n 7~ 0 there exists a
¢1 = ¢1(¢,r0,... rn- )
such that ~ ~- ¢1 ~_ ¢. The claim is then proved by induction on n. We put
¢1(¢,r0,... ,rn-2):= A<¢(g, r0,... ,qn_2, St): t e T>

Now observe that T ~- ¢ implies
q-) e ¢(¢,r0,...,
for every t. Now we apply the rule for conjunction for each t C T, and obtain
~(Y,~A<¢(0",r0,...,r~-2,t):teT> (=¢1).
Furthermore, ¢1 ~_ ¢, that is,
A ( ¢ ( g , r o , . . . ,rn_:,_t)) ~ ¢ ( ¢ , r 0 , . . . , r ~ - l )
For if ~(¢1) is true then for any extension v + with r~-i C dom(v +) we have
v + ( ¢ ) E D. []
The following is an instructive example showing that we cannot have in-

terpolation without constants. Consider the logic of the 2-valued matrix in
the language C' = {A,--}, with 1 the distinguished element. This logic fails
to have interpolation. For it holds that p A -,p ~- q, but there is no interpolant
if p ¢ q. This is surprising because the algebra ( 2 , - , A ) is polynomially
complete. Hence, polynomial completeness is not enough, we must have func-
tional completeness. In other words, we must have all constants. Let us also
add that the theorem fails for intersections of logics with all constants and
conjunction.
A property closely related to interpolation is Hallddn-completeness. It
is named after SOREN HALLDI~N,who discussed it first in [93]. (See also
[191].) A logic is called H a l l d 4 n - c o m p l e t e if for all formulae T and ¢ with
var(~o) r) var(¢) = 0 we have that if ~ ~- ¢ and ~ is consistent then ~- ¢.
2-valued logics are Halld4n-complete. Namely, assume that T is consistent.
Let v : var(¢) --+ 2 be any valuation. Since ~ is consistent there exists a
u : var(T) --+ 2 such that g(~) = 1. Put w := u U v . Since u and v have
disjoint domains, this is well-defined. Then N(T) = 1, and so N(¢) = 1. So,
g(¢) = 1. This shows that ~- ¢. The following generalization is now evident.
THEOREM 1.6.5 (Los & Suszko). Let 97[ be a logical matrix. Then ~-~ is
Hallddn-complete.
Thus, failure of Halld@n-completeness can only arise in logics which are

not determined by a single matrix. In classical logic, the property of Halld~n-
completeness can be reformulated into a somewhat more familiar form. N a m e -
ly, the property says that for ~ and ¢ disjoint in variables, if p V ¢ is a tautology
then either ~ or ¢ is a tautology.
Finally we turn to structural completeness. Recall t h a t structural com-
pleteness means that all admissible rules are derivable. The converse is always
valid.
THEOREM 1.6.6. Suppose that F2 is a logical matrix and ~-~ has all con-
stants. Then it is structurally complete and Post-complete.
PROOF. Suppose that p = ( { ~ 0 , ~ 1 , . . . , ~ - 1 } , ¢ ) - Assume n > 0. We

show t h a t if p is not derivable it is not admissible. So, assume p ¢ ~-~.
Then there is a v : var -~ T such t h a t ~ ( ~ o ) , ~ ( ~ 1 ) , - . - , ~ ( ~ - 1 ) E D while
~(¢) ¢ D. Consider the substitution c~(p) := v(p), where v(p) is the constant
with value v(p). Then ~ , ~ , . . . , ~ - 1 are in D under all homomorphisms,
so they are theorems. But ¢~ is not in D for any homomorphism. So p is
not admissible in ~-~. If n = 0, the addition of ¢ as an axiom makes the
logic inconsistent. For ¢~ is constant, with value ~ D. Hence, ¢~ ~-~ p. So,
adding ¢~ yields an inconsistent logic. Therefore, adding ¢ makes the logic
inconsistent. []
Notes on this section. In [140] a consequence relation b- is cMled uniform

if for sets F and A and a single formula ~ such that var(A) N var(F; ~) =
we have: if F; A ~- ~ then F ~- ~. Obviously a uniform consequence relation
is Halld@n-complete. It is shown in t h a t p a p e r that a consequence relation
is uniform iff it is of the form ~-~ for a single logical matrix. It was noted
by RYSZARD WOJCICKI [230] t h a t for consequence relations t h a t need not
be finitary this works only on the assumption t h a t ~- is regular. Finitary
consequence relations are always regular.
E x e r c i s e 18. Let S be a finite set of finite matrices. Show t h a t ~-s is

decidable.
E x e r c i s e 19. Let ~ = (T, D) be a logical matrix. Show t h a t a connective

( = termfunction) -* satisfies the rule of modus ponens for F-~ iff whenever
a C D and a - ~ b E D then also b E D; in other words, this is the t r u t h
table for - ~ :
1.7. Boolean Logic 31
-- D D
D , D
Read this as follows. D stands for any element of D, D for any element of
T - D. * stands for an arbitrary element. How m a n y binary connectives of
classical logic satisfy (mp.)?
E x e r c i s e 20. Let 9Y~= (T, D) be a logical matrix. Show t h a t --- satisfies the
deduction theorem iff it has the t r u t h table below.
--- D D
D D D
D D D
Thus the above t r u t h table requires only that a ---~ b ¢ D if a C D but b !~ D.
E x e r c i s e 21. Let ff)~ = (T, D) be a logical matrix. Show t h a t A is a conjunc-

tion iff it has the following truth table.
A D D
D D D
D D D
1.7. Boolean Logic

The result of the previous sections will now be applied to the most funda-
mental logic, namely boolean logic. This chapter m a y be skipped by all those
readers who are acquainted with the theory of boolean algebras. The main
purpose is to repeat whatever will be essential knowledge for the rest of this
book. Before we begin, let us agree that we will use the t e r m boolean logic
to denote what otherwise may also be called classical logic. The reason for
not using the latter is a clash in terminology, because there are also classical
modal logics. To distinguish them from the traditional classical logic we call
the latter boolean logic.
We distinguish between 2-valued logic, which is a logic whose semantics
consists of matrices with at most 2 elements. A set of termfunctions is c o m -
p l e t e or a b a s i s if the smallest clone of functions containing this set is the
clone of all term-functions. Examples of bases are { 1 , - , ; q } , {--+, ±}, and
{$, 1}, where p $ q := -~(p n q). Our set of primitive symbols is {1, - , A}.
This set is a basis, as is well-known. Notice that if we need only a polyno-
mially complete set of basic functions, T is redundant. (However, notice that
by the theorems and exercises of the previous section, in the language of
and A 2-valued logic does not have interpolation.)
DEFINITION 1.7.1. A b o o l e a n a l g e b r a is an algebra ~ = ( B , 1, - , N / o f

the signature (0, 1, 2 / such that N is c o m m u t a t i v e , associative and idempotent,
with neutral element 1, and - is a map satisfying the following conditions
with 0 := - 1 and x U y := - ( - x N - y ) .
-- -- X -~- X
x N -x = 0
(x n y) U (x n - y ) = x
R e c a l l t h a t in Section 1.1 we defined a b o o l e a n Mgebra as a b o u n d e d

d i s t r i b u t i v e l a t t i c e w i t h a n e g a t i o n s a t i s f y i n g t h e de M o r g a n laws a n d t h e
i d e n t i t i e s above. We need to verify t h a t if we define 0, U as above, t h e n
( B , 0 , 1, N , U , - } satisfies t h e d e s c r i p t i o n of Section 1.1. T h e m o s t difficult
p a r t is t h e p r o o f of t h e d i s t r i b u t i v i t y law.
LEMMA 1.7.2. I n a boolean algebra, U is associative, c o m m u t a t i v e , idem-

potent, and 0 U x = x.
T h e p r o o f of this fact is left as a n exercise. P u t x < y i f f x N y = x . It

follows t h a t x = y iff x <__y a n d y ~ x. For if t h e l a t t e r holds t h e n x -- x N y =
y N x = y. T h e n x --+ y = - x U y. Moreover, x ++ y : = (x -+ y) n (y --+ x).
For t h e p r o o f of t h e n e x t t h e o r e m observe t h a t
xN0=0
For x n 0 = x n (~ n -~) = (x n x) n -x = ~ n -~ = o.
LEMMA 1.7.3. In a boolean algebra, the following holds.

1. x<_yiffxN-y=O.
2. x<yiffx-ey=l.
3. x=yiffx~y=l.
4. x < y iff - y < - x .
PROOF. (1.) I f x < y t h e n x N - y = ( x N y ) N ( - y ) = z N ( y N - y ) = x N 0 - - 0.
If on t h e o t h e r h a n d x N - y = 0 t h e n x = ( x N y ) U ( x N - y ) = (xNy)U0 = xny,
by t h e p r e v i o u s t h e o r e m . Hence x < y. (2.) x N - y = - ( x -+ y), so x _< y iff
xN-y=0iffx--+y=l. (3.) x = y i f f x < y a n d y < x i f f x - - + y = l a n d
y --+ x = 1 i f f x ++ y = 1. (4.) x < y i f f x N - y = 0 (by (1.)) i f f - y N - - x = 0
iff - y < - x ( a g a i n by (1.)). []
LEMMA 1.7.4. I n a boolean algebra, the following holds.

1. ~ -~ (y -~ z) = (~ n y) -~ z.
2. x N y < z ¢=~ x < y - - e z
The law (2.) is k n o w n as the l a w oS r e s i d u a t i o n .
PROOF. (10 T h i s is also easily p r o v e d from t h e o t h e r laws. F o r x --+ (y -+

z) = - ( ~ n - ( - ( y n - z ) ) ) = -(~n(yn-z)) = -((~ny)n-z) = ( x n y ) -~ z.
(2.) x N y _ ~ z i f f ( x n y ) - 4 z = l i f f x - 4 ( y - 4 z ) - - - - l ( b y (1.)) i f f x _ ~ y - 4 z
by (2.) of L e m m a 1.7.3. []
LEMMA 1.7.5. In a boolean algebra, x n (x -4 y) < y.
PROOF. F r o m - y n x _< x n - y we deduce - y _< x -4 (x N - y ) , by

residuation. So, - ( x -4 - ( x -4 y)) < y, which is x N (x -4 y) < y. []
PROPOSITION 1.7.6. In a boolean algebra, x n (y U z) = (x n y) U (x N z).
PROOF. We have x n y < x since x N y n x ----x n y, and likewise x n z < x.

Furthermore, x N y < y < yUz. Hence x n y <_ y U z , and likewise x n z <<yUz.
This shows one inequality. For the other, observe t h a t
z n (y u z) n - ( x n y) n - ( x n z) = x n ( - y -4 z) n (x -4 - y )
n(x -+ -z)
< (-y-4z) n-yn-z
< zn -z
= 0
So, by (1.) of L e m m a 1.7.3, x N (y U z) < (x n y) U (x n z). []
A n alternative characterization of a boolean algebra is the following.

(B, 0, 1, - , N, U) is a boolean algebra iff its reduct to (0, 1, N, U) is a b o u n d e d
distributive lattice, and - a function assigning to each x E A its complement.
Here, an element y is a c o m p l e m e n t of x if y N x ---- 0 and y U x -- 1. In
a distributive lattice, y is uniquely defined if it exists. For let Yl and Y2 be
complements of x. T h e n Yl (xUy2) ~Yl = (XNyl) U (Y2 N y l ) = Y2 n y l , and
=
so Yl _< Y2. B y the same argument, Y2 _< Yl, and so Yl -= Y2. In addition, if
y is the c o m p l e m e n t of x, x is the complement of y. T h e second definition is
easily seen to be equivalent (modulo the basic operations) to the one of Sec-
tion 1.1. Call y a c o m p l e m e n t o f x r e l a t i v e t o z i f x n y ----0 and x U y ----z.
The law (x N y) U (x n - y ) ----x is (in presence of the other laws) equivalent to
the requirement t h a t x N - y is the complement of x N y relative to x. Namely,
(xny.)n(xn-y)=xn(yn-y)=O.
Given boolean logic, w h a t are the deductively closed sets in f13? To answer
this note t h a t we have x = y i f f x < y and y _ x i f f x ++ y = 1. Now i f S
is deductively closed, it must contain all tautologies and be closed under the
rule (mp.). So i f x E S and x - 4 y E S t h e n y E S. We deduce first of all
t h a t S is not empty; namely, 1 E S. (Recall t h a t 1 is assumed to be the value
of T, which is in the language; if it is not, then at least we have x U - x E S
for an a r b i t r a r y x if S is not empty.) Furthermore, if x E S and x < y t h e n
x-+y= 1 E S and so y E S. Finally, i f x , y E S t h e n a l s o x N y E S s i n c e
x -4 (y -4 x N y) = 1 and by applying the previous rule twice we get the
desired result. T h e following definition summarizes this.
DEFINITION 1.7.7. A f i l t e r in a boolean algebra 92 is a subset F of A

satisfying the following.
(ill.) 1 e F .
(fi_<.) I f x E F and x <_ y then y C F.
(tiM.) I f x , y E F t h e n x F ~ y C F.
A filter is trivial or i m p r o p e r if it is the entire carrier set. A maximal
proper filter is called an ultrafilter.
PROPOSITION 1.7.8. A subset of a boolean algebra is deductively closed iff
it is a filter.
PROOF. T h e direction from left to right has been shown already. N o w
let F be a t i l t e r . Consider x E F a n d x - + y C F . T h e n x M (x --+ y) E F .
But x M ( x - + y) = x F l ( - x U y ) = O . U . x g l y = x M y . So, x M y C F. Now
x My < y and so y C F as well. []
We have characterized the deductively closed sets; these t u r n out to be also

the sets which are congruence classes of the top element. F u r t h e r m o r e , if F
is a tilter, then the relation x ~'~F Y defined by x <-~ y C F is a congruence.
First of all, it is reflexive, since x ~ x = 1 C F . Second, it is s y m m e t r i c
since x <-+ y = y ~ x. A n d it is transitive, for if x HE y and y ~'~F Z t h e n
x <-+ y C F and y ++ z E F and
(x y) n (y z) = (x n y. u . - x n - y ) n (y n z. u . - y n - z )
= (xRyNz. U.-xN-yN-z)
< (xnz. U.-xn-z)
:Xt-~Z
and so x ~ z E F by (tiM.) and (fi<.). Finally, it has to be checked t h a t it

respects the operations. Now, if x ++ y C F , we have - x ++ - y = x e+ y C F
as well. Secondly, if x ++ y E F and z ++ u E F , t h e n also (x M z) ++ (y M u) _<
(x ++ z)M(y ++ u) C F , as one can check. So, if we have a filter F , we also have
a congruence, denoted by OF- Moreover, F = [1]OF. T h e h o m o m o r p h i s m
with kernel O F is denoted by hF. Clearly, if F and G are filters and F C_ G
then O F C_ OG, and conversely. We conclude the following.
PROPOSITION 1.7.9. Let 92 be a boolean algebra. The map f : 0 ~-~ F e : =

{x : x O 1} is an isomorphism from the lattice of congruences of 92 onto the
lattice of filters of 92. Furthermore, if F is a filter, then OF defined by x OF y
iff x e~ y C F is the inverse of F under f .
T h e following are equivalent definitions of an ultrafilter.
PROPOSITION 1.7.10. Let ~ be a boolean algebra not isomorphic to 1. A

filter U is an ultrafilter iff either of (1.), (2.), (3.).
(1.) For all x: - x E U iff x ~ U.
(2.) U is proper and for all x,y: if x U y • U then x • U or y • U.

(3.) The algebra ~ / U is isomorphic to 2.
PROOF. First we show that U is an ultrafilter iff U satisfies (3.). Namely,

U is maximal iff Ou is maximal in Con(92) iff the interval [Ou, V] ~ 2 iff 9.1/U
is simple (by Proposition 1.3.2). A boolean algebra is simple iff it is isomorphic
to 2. For if there exists an element x # 0, 1 then the filter F = {y : y > x}
is distinct from the trivial filters. Now we show t h a t U is an ultrafilter iff U
satifies (1.). Suppose that U is an ultrafilter. Then, by (3.), 9.1/U ~ 2. So
for every x E A, [x]Ou = 0 or [z]®u = 1. Thus either [x]Ou = 1, which
means x C U, or [-x]Ou = 1, which means - x C U. Both cannot hold
simultaneously. Next suppose that (1.) holds. Then if U C V, there exists
an x E V - U. By (1.), however, - x E U and so - x C V. Hence 0 C V, and
V is not proper. So, U is an ultrafilter. Thirdly, we show t h a t U satisfies (2.)
iff U satisfies (1.). Suppose U satisfies (1.). Then U is proper. For 0 ¢ U,
since 1 E U (by (1.)). Now suppose that x ¢ U and y ¢ U. T h e n - x , - y E U
and so ( - x N - y ) • U. Therefore, by (1.), x (J y • V. Hence U satisfies (2.).
Conversely, assume that U satisfies (2.). Then since (1 - - ) x tJ - x • U we
conclude that x • U or - x • U. Not both can hold simultaneously, since U
is proper. Hence x ¢ U iff - x • U. So, U satisfies (1.). []
We know that boolean logic is complete with respect to matrices ( ~ , F ) ,

where ~ is a boolean algebra and F a deductively closed set. This is so
because in the t e r m algebra the relation = : = {(x,y) : x ++ y = 1} is a matrix
congruence. (See exercises of Section 1.4.) The deductively closed sets of a
boolean algebra are the filters. The congruence associated with F , ®F, is a
matrix congruence and so we m a y factor again by the congruence (~F. Hence
boolean logic has a unital semantics. Now, suppose we can show that every
nontrivial filter is contained in an ultrafilter. Then we m a y further deduce
that boolean logic is complete with respect to matrices 9Yt = ( ~ , F ) where F
is an ultrafilter. By the previous theorem, 9Yt/®F = (2, {1}). So the matrix
consisting of the algebra 2 with the distinguished element 1 characterizes
boolean logic completely. This is less spectacular if we look at the fact that
boolean logic is actually defined this way; rather it tells us t h a t the set of
equations t h a t we have chosen to spell out boolean logic is complete - - for
it reduces an adequate set of matrices to just the algebra 2 with D = {1},
exactly as desired.
The next result is of extreme importance in general logic. It deals with the
existence of filters and ultrafilters - - or, as is equivalent by Proposition 1.7.10,
- with the existence of homomorphic images of boolean algebras. Let us take
-
a subset E _C A and ask when it is possible to extend E to a proper filter on

A.
PROPOSITION 1.7.11. The least filter containing an arbitrary subset E is

the set (E) defined by
<E> = { x : > r] X, Z a finite subset o: E}
(E) is proper iff every finite subset of E has a nonzero intersection. In this
ease we say that E has the f i n i t e i n t e r s e c t i o n property.
PROOF. First of all, E C_ (E}. For x >_ N{x}. (E} is also a filter. For if
Z = ~ , then N Z = 1, so (ill.) is satisfied. Next let x C (E} and x < y. We
know t h a t there is a finite X such that x ~ N X. Then also y > ~ X and so
y E (E). This shows (fi_<.). Finally, if x, y E (E} then x >_ N X , y >_ N Y for
some finite X , Y C_ E. Then x n y > A ( X U Y ) ; and X U Y is finite. To see
t h a t (E) is the least filter, observe t h a t for every finite subset X we must put
N X e (E}. For either Z = Z and so ~ Z must be added to satisfy (ill.), or
Z = {x} for some x, and then ~ Z = x must be added to satisfy E C_ (E).
Or ~X > 1 and then N X must be added to satisfy (riD.). Finally, to satisfy
(fi<.) all elements of (E} must be taken as well. So no smaller set suffices.
It is then clear that if for no finite set X we have ~] X = 0, (E} is not the
full algebra, since 0 ¢ (E}. But if (] X = 0 for some finite X then by (fi<.),
y E (E> for all y. []
Now we show that every proper filter is contained in an ultrafilter. Equiv-
alently, every set with the finite intersection property can be embedded in an
ultrafilter. The proof will be of quite general character, by appealing to what
is known as Tukey's Lemma. (See [215]. This l e m m a is actually the same as
OSWALD TEICHMULLER'S Principle D, 3rd Version from [205].) To state the
lemma, consider a set S and a property T of subsets. T is said to be of f i n i t e
c h a r a c t e r if ? holds of a set X C S iff it holds of all finite subsets of X. If
T is of finite character, then if T has T and S C_ T then also S has T. For
every finite subset of S is a finite subset of T, and so if S fails to have T, this
is because some finite subset So fails to have ~?, which implies t h a t T fails to
have T, since So C T.
THEOREM 1.7.12 (Tukey's Lemma). Suppose • is a property of subsets
of S and that T is of finite character. Then every set X C S having • is
contained in a set X* which is maximal for ~.
PROOF. By the axiom of choice, S can be well-ordered by an ordinal A;
thus S = {s~ : a < A}. By induction over A we define X~ for a _< A. We
put X0 := X. If X~ is defined, and a + 1 ¢ A then put X~+I := X~ U {s~}
if this set has ~, and X~+I := X~ otherwise. For a limit ordinal a we put
X~ := LJ~<~ x ~ . Now let X* := XA. We claim that X* has the desired
properties. X* contains X. It has ~P; for X0 has • by assumption. In the
successor step this remains true by construction and in the limit step by the
finite character. For either the sequence (X~ : ~ < c~) is stationary from a
certain ordinal /3. Then X a = XZ, in which case Xa has [P by induction

hypothesis. Or the sequence is strictly ascending. Then any finite subset of
X~ is contained in a X~ for some/3 < a; and then X~ has [P by the finite
character of ~P. Finally, we must verify that X* is maximal. Suppose that
X* C_ Y and Y has T. Take an element y C Y. y = s-r for some 7 < A.
Consider the definition of Xv+l. We know that X-y+1 _C X* C_ Y. Since Y
has T, X* U {s?} has [? as well, since ~P is closed under subsets. Therefore
X-y+1 = X-y U {s?}, showing y G X*. Since y was arbitary, X* = Y. Hence
the set X* is maximal. []
COROLLARY 1.7.13. Every set with the finite intersection property is con-
tained in an ultrafilter. In particular, every proper filter is contained in an
ultrafilter.
PROOF. By Tukey's Lemma. The property • is taken to be generates a

proper filter. It is of finite character. For X generates a proper filter iff X
has the finite intersection property iff every finite subset of X has the finite
intersection property iff every finite subset generates a proper filter. []
THEOREM 1.7.14. Boolean logic in the language T, ~ and A is the logic

o/the matrix 2 := ({0, 1}, 1, - , N, {1}}. Boolean logic is structurally complete
and Post-complete. It has interpolation and is Hallddn-complete.
For a proof note that we have established that boolean logic is the logic of
the matrix 2. Furthermore, it has all constants, since T has value 1 and J_
has value 0. It follows that it has interpolation by Theorem 1.6.4 (since the
matrix is finite, the logic has conjunction and all constants), and that it is
structurally complete and Post-complete by Theorem 1.6.6. It is Halld~n-
complete by Theorem 1.6.5.
E x e r c i s e 22. Show that { - , N} is polynomially complete and t h a t {1, - , N}

is functionally complete.
E x e r c i s e 23. Let - be an operation satisfying - - x -- x. Let N be a

binary oeration, and put x U y := - ( - x N - y ) . Show t h a t N is associative
(commutative, idempotent) iff U is associative (commutative, idempotellt).
Hint. Show first that x N y = - ( - x U - y ) , so that only one direction needs
to be established in each case. (This shows L e m m a 1.7.2.)
E x e r c i s e 24. Show that (x N z) ++ (y N u) _< (x +~ z) N (y ++ u).
E x e r c i s e 25. Show that each proper subspace of a vector space is contained

in a hyperplane, i. e. a maximal proper subspace.
1.8. S o m e N o t e s o n C o m p u t a t i o n and Complexity

In this section we will briefly explain the basic notions of computability
and complexity. Although we shall prove only very few results on complexity
of modal logics we shall nevertheless mention a fair number of them. This
section provides the terminology and basic results so t h a t the reader can at
least understand the results and read the relevant literature. The reader is
referred to MICHAEL R. GAREY and DAVIDS. JOHNSON [73] for an introduc-
tion into complexity theory. For our purposes it is best to define computations
by means of string handling devices. The most natural one is of course the
Turing machine, but its definition is rather cumbersome. We shall therefore
work with a slightly easier model, which is a mixture between a Turing ma-
chine and a so-called S e m i - T h u e P r o c e s s . Let us now fix an alphabet A.
Members of A shall be written in small caps to distinguish t h e m from symbols
that merely abbreviate or denote symbols or strings from A.
DEFINITION 1.8.1. Let A be a finite set and V ~ A. A s t r i n g h a n d -
ling m a c h i n e over A is a finite set T of pairs (2, y~ such that ~ and ~ are
strings over A U ( V } in which V occurs exactly once. We call members of T
instructions.
The symbol V is used to denote the position of the read and write head
of the Turing machine.
DEFINITION 1.8.2. Let T be a string handling machine over A and ~ and
strings over A U {V}. Then ~ ~ T ~ if there is a pair (~,~} C T and
strings 31 and 32 such that ~ = 3 1 ~ v - J 2 and ~ = 3 1 ~ 3 2 . In that case
we say that ~ is 1 - s t e p c o m p u t a b l e f r o m 2. g is n - i - l - s t e p c o m p u t a b l e
f r o m 2, in symbols ~ ~ + 1 ~, if there is a Z such that 2 ~ F ~ T j . ~ is
c o m p u t a b l e f r o m 2 if ~ ~ ~ for some n E w. We write 2 ~ ~.
We also call a sequence (~i : i < n) a T - c o m p u t a t i o n if for every
j < n - 1 we have ~y ~ T 2j+l. g i s called a h a l t i n g s t r i n g f o r T if
there is no string t h a t is computable from g. A h a l t i n g c o m p u t a t i o n is a
c o m p u t a t i o n whose last member is a halting string.
DEFINITION 1.8.3. Let f : A* --+ A* be a function and T a string hand-
ling machine over some alphabet C = A U B , where B is disjoint from A. We
say that T c o m p u t e s f with respect to B, E, ~: C B if for every ~ C A* the
following holds:
1. V ~ E ~ f ( ~ ) ~ # is computable from V ~ B ~ X ~ =.
2. V ~ E ~ f ( ~ ) ~ is a halting string for T.
3. There is no halting string for T different from V ~ E ~ f ( ~ ) ~ which is
computable from V ~ B~ ~ ¢~
f is called n o n d e t e r m i n i s t i c a l l y c o m p u t a b l e if there exists a string han-
dling machine that computes f (with respect to some elements).
1.8. Some Notes on C o m p u t a t i o n and Complexity 39
The elements B and E are used to m a r k the begin and the end of the
computation and ~ to m a r k the end of the input string. Notice t h a t it is
not possible to define a string handling machine that computes f ( ~ ) from
simpliciter. For we would have no means to distinguish whether we just
started the computation or whether we just ended it; nor could we tell where
the input ended. Notice that the symbol B also marks the begin of the input
string, whence a separate marker is not needed for that purpose. In the
definition above we shall say that T computes f ( ~ ) from ~ in n s t e p s if
Note t h a t it is not required that all computations terminate; but if a com-

putation halts, it must halt in the just specified string. If f is only a partial
function we require that if f is undefined on ~ then there is no halting com-
putation starting with V ~ B ~ . This allows us to define the notion of a
computable function from A* to B*, which is namely also a partial function
from (AUB)* to (AUB)*. We shall present a simple example. Let A = {R, S}
and let f : ~ ~ ~ a . Here is a string handling machine that computes this
function:
T1 := {(VB#, V E a # ) ,
(VBR,BRVc>, <VBs,BsVc>,
(VCR,~Vc), (Yes,sVc>,
(vc#, VDR#>,
(RVD, VDR), (SVD, VDS),
<BVD, rE>}
The verification that this machine indeed computes f is left to the reader.
The above definition of computability is technical. Yet, ALONZO CHURCH
advanced the conjecture - - also known as C h u r c h ' s T h e s i s - - that any func-
tion that is computable in the intuitive sense is also computable in the tech-
nical sense. (See STEVEN C. KLEENE [116] for details.) The converse is of
course clear. Notice that if A consists of a single letter, say A, then the set of
natural numbers can be identified simply with the set of n o n e m p t y sequences
over A, where n is represented by the sequence consisting of n consecutive As.
(ALAN TURING used n + 1 As to code n, but this is unnecessary given the
present setting.) It can be shown that under this coding any recursive function
on the natural numbers is computable, and that any computable function is
recursive. There are other codings of the natural numbers t h a t work equally
well (and are less space consuming), for example p-adic representations.
The above definitions can be generalized by admitting computation not
on one string alone but on several strings at the same time. This model is
more general but it can be shown that it does not generate a larger class of
computable functions (modulo some coding). The benefit is that it is much
easier to see that a certain function is computable. A string handling machine
with k-strings is a finite set of k-tuples of pairs of strings over AU {V} in which
V occurs exactly once. Computations now run over k-tuples of strings. T h e

definitions are exactly parallel. A replacement is done on all strings at once,
unlike a multihead Turing machine, which is generally allowed to operate on
a single tape only at each step. We can now define the notion of a computable
k - a r y function from A* to A* in much the same way. Notice t h a t for T to
compute f we shall simply require that the initial configuration is the sequence
(V~B~i~ : i < k}, and that the the halting configuration is the sequence
(V~E~f(H)~#, <V~E~ : i < k - I}>. The fact that the starting state and
the end state appear on each tape is harmless. Given f : (A*) k -+ A*, we
shall define a unary function f~ : (A U {c})* -+ A*, where c ~ A, as follows.
If ~ = ~ 7 0 ~ c ~ 1 ~ c ~ . . . ~ k _ l ~ c ~ , then
:= : i < hi)
Otherwise, f v ( ~ ) :__ c. The following now holds.
THEOREM 1.8.4. f is computable on a string handling machine with k
strings iff f v is computable on a string handling machine with one string.
The proof of this theorem is not hard but rather long winded.
W i t h these definitions we can already introduce the basic measures of
complexity. Let f : A* -+ A* and h : w -+ w be functions. We say t h a t
T c o m p u t e s f in h - t i m e if for all ~, T computes f ( ~ ) from Z in at most
h(l~[) steps. We say that T computes f in h - s p a c e if there is a c o m p u t a t i o n
of V ~ E ~ f ( ~ ) ~ # from V ~ B ~ x ~ in which every member has length _< I~[.
Typically, one is not interested in the exact size of the complexity functions,
so one introduces more rough measures. We say that T computes f in O(h)-
t i m e if there is a constant c such t h a t T computes f is c. h - t i m e for almost all
~. Analogously with space. Now, before we can introduce the m a j o r classes
of complexity, we shall have to define the notion of a deterministic machine.
DEFINITION 1.8.5. Let T be a string handling machine. T is called de-
t e r m i n i s t i c if for every string ~ there exists at most one string ~ which
is 1-step computable from ~. A function f : A* -+ A* is d e t e r m i n i s t i c a l l y
computable if there is a deterministic string handling machine that computes
f with respect to some elements.
It is easy to see that the machine T1 defined above is deterministic.
Namely, a string handling machine T is deterministic if for any pair (~, y-) E T
there exists no other pair (g, g> E T such t h a t g is a substring of ~. This is
obviously satisfied. The following theorem shall be stated without proof.
THEOREM 1.8.6. A function f : A* --+ A* is nondeterministically com-
putable iff it is deterministically computable.
DEFINITION 1.8.7. P denotes the class of functions deterministically com-
putable in polynomial time, N P the class of functions nondeterministically
1.8. Some Notes on C o m p u t a t i o n and Complexity 41
computable in polynomial time. Similarly, E X P T I M E ( N E X P T I M E ) de-

notes the class of function deterministically (nondeterministically) computable
in exponential time. P S P A C E denotes the class of functions deterministi-
cally computable in polynomial space.
The reason why there is no class N P S P A C E is the following result from
[189].
THEOREM 1.8.8 (Savitch). If a function is computable by a nondetermin-
istic machine using polynomial space then it is also computable by a determin-
istic machine using polynomial space.
We have
P _C NP _C PSPACE C E X P T I M E _C N E X P T I M E
Most inclusions are by now obvious. Notice t h a t if a deterministic machine
runs in h - t i m e it can write strings of length at most O(h). When using a
machine with several strings the complexity does not change. It can be shown,
namely, that if T is a string handling machine with k strings computing f in
O(h) time, then there is a string handling machine T v computing f~) in O(h 2)
time. Moreover, T v can be chosen deterministic if T is. It is therefore clear
that the above complexity classes do not depend on the number of strings on
which we do the computation, a fact that is very useful.
Talk about computability often takes the form of problem solving. A
problem can be viewed as a subset S of A*. To be exact, the problem that is
associated with S is the question to decide, given a string ~ C A*, whether or
not ~ E S.
DEFINITION 1.8.9. A p r o b l e m is a function f : A* --+ {0, 1}. f is trivial
if f is constant. We say that a problem f is C if f E C, we say that it is C-
h a r d if for any g C C there exists a p C PSPACE such that g = f o p . Finally,
f is e - c o m p l e t e if it is both in C and C - h a r d .
A problem that is C-complete is in a sense the hardest problem in C.
For any other problem is reducible to it. Given a set S, we shall denote by
'x C S?' the problem of deciding membership in S, which is simply defined as
the function t h a t gives 1 if the answer is 'yes' and 0 if the answer is 'no'. (This
is also the characteristic function of S.) For amusement the reader may show
that any nontrivial problem t h a t is in P is also P-complete. In connection
with these definitions we have
DEFINITION 1.8.10. A subset of A* is decidable if its characteristic func-
tion is computable.
We shall spend the rest of this section illustrating these concepts with
some examples from logic. We have introduced languages in Section 1.2.
To adapt the definitions of that section to the present context we need to
make the alphabet finite. This is not entirely straightforward, since we have
infinitely m a n y variables. Therefore, let F be the set of function symbols,
and X := {Pi : i E w} our set of variables. Let us first assume t h a t F is
finite. Take symbols P, 0 and 1 not occurring in F or X. So, we shall replace
the variable Pi by the sequence P~Z, where Z E {0, 1)* is a binary string
representing the number i in the usual way. T h a t is to say, we put p(0) := 0
and #(1) := 1, and if ~ = x o ~ x l ~ . . , x,,_l then i = #(~), where
"(e) := Z
j<n
In the same way we can obviously also code a countable F by means of a single
symbol F followed by a binary string. In this way, the entire logical language
can be written using just the symbols F, P, 0 and 1. We will however refrain
from using this coding whenever possible. We shall note here t h a t the typical
measure of length of a formula is the number of symbols occurring in it. We
call this the s y m b o l c o u n t of the formula. However, the actual string t h a t
we write to denote a formula can be longer. Since a variable counts as one
symbol, the coding is not length preserving. Rather, a formula with n symbols
is represented by a string of length at most n log 2 n if we allow renaming of
variables. Given qo, IPl counts the length of a minimal representing string.
This additional factor by which I~1 is longer t h a n the symbol count is usually
(but not always!) negligeable. If the set of function symbols is infinite, there
is no a priori upper bound on the length of the string in comparison to the
symbol count! It is for these reasons t h a t complexity is always measured in
terms of the length of the string representing the formula. There is another
way to represent a formula, which we refer to as the packed representation. It
is described in the exercises below since it will only be relevant in Section 3.6.
We shall now go into the details of certain basic string properties and
manipulations. First of all, let ~2 : F -+ w be a signature and X = {Pi : i E w}.
We shall provide a procedure to decide whether or not a given string is a term.
Define the w e i g h t pa of a symbol as follows.
p~(p~) := -1
p.(f) := f (y) - 1
For a string ~ = x o ~ x l ~ . . . ~ x n - 1 we put
i<n
It is useful to observe the following
LEMMA 1.8.11. Let ~ be a string such that p~(~) < O. T h e n there exists
a prefix ff of Z such that pa(~) = p~(~) - 1.
1.8. Some Notes on Computation and Complexity 43
This lemma is proved by induction on the length of ~. For the sake of

precision define an o c c u r r e n c e of a string ~ in ~ to be a pair (zT,2) such that
~ 7 ~ is a prefix of ~. Two string occurrences (zT,~) and (g, z~ o v e r l a p if either
(a) ~7 is a prefix of g and ~7 is a proper prefix of ~7~ or (b) ~ is a prefix of ~7
and ~ is a proper prefix of ~ .
PROPOSITION 1.8.12. A string ~ is an ~ - t e r m iff it has the property (P).
(P). p~(~) = - 1 and for every prefix ~ of~: Pn(Y) _> 0.
In particular, no proper prefix of a term is a term, and no two distinct occur-
rences of subterms overlap.
PROOF. We begin by showing how the other claims follow from the first.
If ~ is a term and ~ a proper substring, then Pn(~) > - 1 , and so ~ is not
a term. Next, let £ = ~70~71~72~g3~74, and assume that ~71~72 as well as
~2~73 are terms and both ~71 ¢ e and ~73 ¢ ~. Then we have pa(~72~73) =
pa(~72) + pa(~Ta). Now, since ~72~73 has Property (P), pa(~72) > 0. Now,
likewise - 1 = pa(~71) + Pa(g2), whence pa(ff~) < 0. So, ~71~g2 does not have
(P), and hence is not a term, contrary to our assumption.
We now show the first claim. This is done by induction on the length
of £. Clearly, for strings of length 0 and 1 the claim is true. (Notice that
p~(~) = 0.) So, let t~] > 1. Assume that the claim is true for all strings of
length < [~]. Assume first that ~ is a term. Since £ has length at least 2, the
first symbol of ~ is some f E F of arity at least 1. So,
---- f ~ 0 ~ ... ~xf~(/)_~
By inductive hypothesis, pn(£i) = - 1 for all i < ~ ( f ) . Hence we have
pn(Z) = ~ ( f ) - 1 - ~ ( f ) -- - 1 . Now it is easy to see that for no prefix ~7
of ~ pn(~7) _> 0. Hence £ has Property (P). Conversely, assume that ~ has
Property (P). Take the first symbol of £. It is some f E F of arity at least
1. Let g b e such that £ = f ~ . Then pn(g) _> - 1 . If the arity of f is
1, then ~ is a term iff ~ is. Then ~ has (P) and so is a term, by inductive
hypothesis. Therefore Z is a term as well. Now suppose that ~ ( f ) > 1. Then
pn(~) -= - f ~ ( f ) . Using Lemma 1.8.11 it can be shown that ~ is the product
of strings ~7i, i < ~ ( f ) , which all have (e). By inductive hypothesis the ~7i are
terms. So is therefore ~. []
We shall note here the following corollary. If F is infinite, we code it as
described above by strings of the form F~£. ~ is defined on these strings. We
shall recode the signature as a function from binary strings to binary strings.
Namely, we let
~'(~) := ~ - ~ ( ~ ( F ~ ) )
We shall say that ~ is computable if ~ is.
PROPOSITION 1.8.13. Assume that L is a language with a computable
signature. Then the set of terms over F is decidable.
The proof is not difficult.

It is possible to convert a term in prefix notation into a string in the
typical bracket notation and conversely. To that end, we assume that the
signature consists of symbols of arity < 2. The procedure is as follows. Take
a string ~ in prefix notation. Start from the left end. Assume ~ = ~ f ~ F and
n := t/71. Then let 3 be the smallest prefix of Z such that f ~ 3 is a term and let
~' = 3~7. Then let ~ := ~ ( ~ f ~ 3 ~ ) ~ g . We call this the insertion of brackets
at the place n. The procedure is now simply the following. Start at the left
end of ~ and add brackets whenever the described situation occurs. Continue
as long as possible. Call the resulting string 5(2). Now let b(~) be given. Pick
a symbol f of arity 2 following a symbol (. Then it is contained in a sequence
of the form ( ~ f ~ 3 0 ~ 1 ~) where u0 and 31 are terms. (This needs to be
defined for sequences which contain brackets but is straightforward.) Replace
this sequence by ( ~ g 0 ~ f ~ 3 1 ~ ) . Continue whenever possible. Finally, some
brackets are dropped (the outermost brackets, brackets enclosing variables).
The resulting sequence is a term in usual bracket notation.
To close, let us describe a procedure that evaluates a term in a finite
algebra given a particular valuation. It is commonly assumed that this can
be done in time proportional to the length of the string. The procedure is as
follows: identify a minimal subterm and evaluate it. Repeat this as often as
necessary. This procedure if spelled out in detail is quadratic in the length
of the string since identifying a minimal subterms also takes time linear in
the length of a string. As this procedure is repeated as often as there are
subterms, we get in total an algorithm quadratic in the length. We shall
now describe the algorithm in detail. So, let 92 be a finite f/-algebra. For
each element a of the algebra we assume to have a primitive symbol, which
we denote by a. Let the term be given as a string, and the valuation as a
sequence of pairs ( P ~ , / 3 ( P ~ ) ) . It is not necessary to have all values, just all
values for variables occurring in ~. We shall describe a procedure that rewrites
successively, until a particular value is determined. We start by inserting
/3(P~2) in place of p~2. We treat the elements as variables, assigning them
the weight - 1 . Let (3, ~) be an occurrence of a substring, w h e r e / j has length
> 1. We call p~(3) its e m b e d d i n g n u m b e r . An occurrence of a substring
is with maximal embedding number is of the form f~ao~al~... ~ a a ( f ) - l ,
where all di are (symbols denoting) elements of 92. The procdure is therefore
as follows. Look for an occurrence of a nontrivial substring with maximal
embedding number and compute its value. Replace that string by its value.
Continue as long as possible. The procedure ends when ~ is a single symbol.
Now let a formula qo of boolean logic be given. We can solve the problem
whether p is a theorem nondeterministically by guessing a valuation of the
variables of ~ and then evaluating T. A valuation is linear in the length of T.
Hence, the problem whether a formula is a theorem of boolean logic is in NP.
1.8. Some Notes on Computation and Complexity 45
Alternatively, the problem whether a boolean formula is satisfiable is also in

NP. Moreover, the following holds:
THEOREM 1.8.14 (Cook). Satisfiability of a boolean expression is N P -
complete.
This result m a y appear paradoxical since we have just proved t h a t satis-
liability is computable nondeterministically in O ( n 2) time. So, how come it
is the hardest problem in NP? The answer is the following. If a problem S is
in NP it is polynomially reducible to the satsifiability problem; the reduction
function is itself a polynomial p and this polynomial can have any degree.
Hence the harder S the higher the degree of p.
It has been shown by L. J. STOCKMEYER and A. R. MEYER [203]
that the problem of satisfiability of quantified boolean formulae is PSPACE-
complete. (Here, quantifiers range over propositional variables.)
E x e r c i s e 26. We can represent the natural number n either as the sequence

# - l ( n ) over {0, 1} or as a sequence o f n ls. Show that the mappings convert-
ing one representation to the other are computable.
E x e r c i s e 27. Show that if f , g : A* -+ A* are computable then so is g o f .

Show that if f and g are in C for any of the above complexity classes, then so
is g o f .
E x e r c i s e 28. Let A be a fixed alphabet, and let c (for comma) ( and ) be

new symbols. W i t h the help of these symbols we can code a string handling
machine T by a string Tt. (This is not unique.) Now let C be the enriched
alphabet. Let f : C* x C* -+ C* be defined as follows. If g d A* and ~ = T t
for some string handling machine using the alphabet A then f ( ~ , if) is the
result that T computes on ~ if it halts, and otherwise f is undefined. Show
that f is computable. (This is analogous to the Universal Turing Machine.)
E x e r c i s e 29. Prove L e m m a 1.8.11.
Exercise 30. Here is another way to represent a formula. Let ~ be a formula,

say Ap0 V Pl ~P0, which in infix notation is just (P0 A (pl V ~P0)). The string
associated with it is AP0 V p l - , p 0 . Now enumerate the subformulae of this
formula. We shall use for the n t h subformula the code s ~ , where ~ is the
binary representation of n:
S0 : P0
$1 : P1
sl0 : ~P0
sll : VplmP0
$100 : APOVPI=P0
Now replace to the right of this list the immediate subformulae by their re-
spective code, starting with the smallest subformulae that are not variables.
(For example, the least line becomes s l 0 0 A s 0 s l l . ) Finally, write these lines
in one continuous string:
sOPOsIPiS10-~sOsll V s l s l 0 s 1 0 0 A s O s l l
Denote the resulting string by ~ ' . Give upper and lower bounds for t ~ 1 in
comparison with I~I. Show that given a sequence ~ one can compute in linear
time a formula ~ such that ~ = ~ . This representation can be used to code
a set A of formulae as follows. Each subformula of A that is itself in A is
denoted not by s ~ but by s ~ . How is IA*I related to card(sf[A])?
CHAPTER 2
Fundamentals of Modal Logic I
2.1. Syntax of Modal Logics

The languages of propositional modal logic used in this book contain a
set var = {Pi : i C 7} of variables, a set cns of propositional constants, the
boolean connectives T, 7, A and a set {[]i : i E ~} of modal operators. E]i~
is read box i phi. T is in cns. With each operator [~i we also have its so-
called d u a l o p e r a t o r 0i defined by 0 i ~ := - ~ [ 3 i ~ . In what is to follow, we
will assume that there are no basic constants except T; and t h a t there are
countably m a n y propositional variables. So, 7 = 1% unless stated otherwise.
The number of operators, denoted by s throughout, is free to be any cardinal
number except 0; usually, s is finite or countably infinite, though little hinges
on this. Theorems which require that s has specific values will be explicitly
marked. By [P~ we denote the language of x - m o d a l logic, with no constants
and R0 m a n y propositional variables. Also, [P~ denotes the set of terms also
called well-formed formulae of that language. If ~ = 1 we speak of (the
language of) monomodal logic, if s = 2 of (the language of) bimodal logic.
Finally, -~ and E]j are assumed to bind stronger t h a n all other operators, A
stronger than V, V stronger than --+ and ++. So, [~jp A q -+ p is the same as
( ( @ p ) A q) -~ p.
~P~ is the set of terms or formulae or propositions as defined earlier.
Metavariables for propositional variables are lower case R o m a n letters, such as
p, q, r, metavariables for formulae are lower case Greek letters such as p, X, ¢.
In addition, rather than using indices to distinguish modal operators we will
use symbols such as II, Fq, [], rq, [], [], and similar abbreviations for their
duals. We denote by II[P~II the cardinality of the set of terms over ~P~. With
the cardinality of the set of variables, of the set of constants and the set of
operators the cardinality of ~P~ is fixed. Namely, by Proposition 1.2.1 we have
II ll = max{ 0, var, cns,
Since we standardly assume to have at most Ro variables and constants, the

latter reduces generally to II[P~II = max{lqo, ~}.
47
48 2. F u n d a m e n t a l s of Modal Logic I
The m o d a l d e p t h or ( m o d a l ) d e g r e e of a formula is the m a x i m u m

number of nestings of modal operators. Formally, it is defined as follows.
@(p) := 0 p E var U cns

dp(~) := @(~)
dp(~ A ¢) := max{@(~), @(9)}
@([]j~) := 1+@(~)
The boolean connectives will behave classically. As for the modal oper-
ators, the interest in modal logic lies in the infinitely m a n y possibilities of
defining their behaviour. First of all, according to the theory outlined in
Chapter 1, a modal logic must be a relation ~- C ~o(~) × ~ satisfying (ext.),
(mon.), (cut.), (sub.) and (cmp.). Moreover, we generally assume t h a t the
boolean connectives behave as in boolean logic. There is a special set of con-
sequence relations - - by no means the only ones - - which have a deduction
theorem for --+. Such consequence relations are fully determined by their sets
of tautologies. Indeed, it is standard practice to identify modal logics with
their set of tautologies. We will stick to t h a t tradition; however, we will see
in Section 3.1 that for a given set of tautologies there exist other consequence
relations with useful properties. Thus we call a set A C_ T~ a m o d a l logic
if A contains all tautologies of boolean logic, is closed under substitution and
modus ponens, t h a t is, if ~ E A then ~ C A for a substitution ~, and if ~ C A
and ~ --~ ¢ C A then 9 E A. The relation ~-h is then defined via
(cmp.) A ~-A T i f f there is a finite set A0 C A such that ded(Ao, ~) C A .
Let a logic A be given. Fix an operator [] of the language for A. [] is called
c l a s s i c a l in A if the rule (cl[].) is admissible; if (mo[Z.) is admissible in A,
[] is called m o n o t o n e in A. Finally, if (mn.) is admissible, and if A contains
the axiom of box distribution, which is denoted by (bd--~.) [] is called n o r m a l
in A.
(olD.)
k~9
~ []~. ~ .[]9
(mo[].) k []~.-~
k ~ - ~ 9.G9 (ran.) []-----~
~
k~
(bd--+.) ~- 5 ( ~ ~, 9). --+ . ~ -~ D 9

A normal operator [] of A is monotone in A; a monotone operator of A is clas-
sical in A. A logic is c l a s s i c a l ( m o n o t o n e , n o r m a l ) if all its operators are
classical (monotone, normal). Two formulae ~, ¢ are said to be d e d u c t i v e l y
or (locally) e q u i v a l e n t in a logic A if ~ ++ ¢ E A. Classical logics have the
property that ¢1 and 92 are deductively equivalent in A, if 92 results from 91
by replacing in 91 an occurrence of a subformula by a deductively equivalent
formula.
2.1. S y n t a x of Modal Logics 49
PROPOSITION 2.1.1. Let A be a classical modal logic and 91 ~ 92 E A.

Let ¢1 be any formula and let ¢2 result from replacing an occurrence of 91 in
¢1 by 92. Then ¢l ++ ¢2 E A.
PROOF. Notice t h a t the following rules are admissible in addition to
(cl[].), by the axioms and rules of boolean logic.
(cl A .) ~- a l ++ a2 ~- 3i ++ 132 ~- ai ++ a2
~- a i A 13i ++ a2 A 32 (cl-~.) ~- - - a i ++ -~a2
By induction on the constitution of ~1 it is shown t h a t ~bl ++ ¢2 E A, s t a r t i n g
with the fact t h a t 91 +-} ~2 C A and p ++ p C A the claim follows. []
PROPOSITION 2.1.2. Let A be a classical modal logic. Then for any for-
mula ~ there exists a formula ¢ which is deductively equivalent to 9 and is
composed from variables and negated variables, ± and T using only A, V, []j
and (}j, j < ~.
For this t h e o r e m it makes no difference whether the symbols ± , V and (}j,
j < a, are new symbols or merely abbreviations. T h e d u a l of a formula is
defined as follows.
pd :_~ p
(T) d := ±
(_~)d := T
(gA¢)d := 9dVCd
(gv¢)d := 9dACd
([]~)d := (>j~d
(0~) d := []j9 ~
The dual is closely related to negation. Recall, namely, that in boolean algebra
negation is an isomorphism between the algebras (A, -, •, U 1 and (A, -, U, M I.
The following theorem -- whose proof is an exercise -- states that for axioms
there typically are two forms, one dual to the other.
PROPOSITION 2.1.3. Let A be a classical modal logic. Then 9 --+ ¢ E A
iff cd __+ pd E A.
In m o n o t o n e logics we can prove t h a t an alternative version of the (bd-+.)
postulate, (bdA.), is equivalent.
(bdA.) ~- [ ] ( 9 A ¢). ++ -[]9 A [-1¢
PROPOSITION 2.1.4. Let A be a classical modal logic, and [] be a modal

9perator of the language of A. (1.) (mn.) is admissible for an operator [] in
h if [~T C A. (2.) If A contains (bd-+.) and []T then [] is normal in A. (3.)
If [] is monotone in A, the postulates (bdA.) and (bd--+.) are interderivable
in A.
50 2. F u n d a m e n t a l s o f M o d a l L o g i c I
PROOF. (I.) By assumption, []T E A and therefore ~- [Z-V ++ 7. Now

assume F- ~. Then F- ~ ++ T and so, by (cl[Z].), ~- Dqo ++ [~T. Using this
equivalence we get ~- [2~ ++ -U, that is, ~- []~, as required. (2.) By (I.), (mn.)
is admissible. (3.) Assume (bd-4.) is in A. Then, since ~- ~ A ~. -4 .~ and
A ¢. -4 .¢ we have, by (mo[].), U A ¢) -4 and U A ¢) -4 []¢.
Now by boolean logic we get ~- [D(~A¢). -4 .[~oA[Z¢. Next, since it holds that
~- ~. -4 .¢ -4 (~ A ¢), with (mo[3.) we get k- []~. -4 .[](¢ -4 (T A ~b)). Using
(bd-4.) and some boolean equivalences we get k- []~. -4 . [ ~ -4 7](~ A ~b),
from which the remaining implication of (bdA.) follows. Now assume (bdA.)
is in A. Then ~- [](~ -4 ~) A []T. -4 .~(~, -4 ~. A .T). Furthermore, we get
~- D(T --+ ¢. A .~) -4 [ ~ , by (mo[:].). Hence ~- ~ ( ~ -4 ~)A []T. -4 .[]~, which
is equivalent to (bd-4.). []
The smallest classical modal logic will be denoted by E~, the smallest
monotone logic by M~ and the smallest normal modal logic by K~ (after
SAUL KRIPKE). The index n is dropped when ~ = 1, or whenever no confusion
arises. Notice that since these logics are determined by their theorems, it is
enough to axiomatize their theorems. This is different from an axiomatization
of the proper rules (see Section 3.9). Notice, namely, that when our interest is
only in axiomatizing the theorems, we can do this using the admissible rules
for deriving theorems. A classical logic (monotone logic) can be identified
with its set A of theorems, which is a set containing all boolean tautologies
and which is closed under modus ponens, substitution and (cl[].) or, for
monotone logics, (moZ].). A q u a s i - n o r m a l logic is a modal logic containing
K~. A quasi-normal logic is normal iff it is closed under (mn.). Notice
that in a normal logic, the rule (mo[3.) is in fact derivable. The smallest
normal n-modal logic containing a set X of formulae is denoted by K ~ ( X ) ,
K ~ . X or K~ @ X, depending on the taste of authors and the circumstances.
The smallest quasi-normal logic containing a set X is denoted by K~ + X.
Similarly, if A is a (quasi-)normal logic then the result of adding the axioms X
(quasi-)normally is denoted by A @ X (A + X). In particular, there is a list of
formulae that have acquired a name in the past, and modal logics are denoted
by a system whereby the axioms are listed by their names, separated mostly
by a dot. For example, there are the formulae D = 0 T , 4 = 0 0 p --+ 0p.
The logic K ( 0 T ) is denoted by K D or also K . D , the logic K ( 0 0 p -4 0p) is
denoted by K4, and so forth. We will return to special systems in Section 2.5.
Let us now turn to the calculi for deriving theorems in modal logic. Let
a logic be axiomatized over the system K~ by the set of formulae X, that is,
consider the logic now denoted by K~ @ X. For deducing theorems we have
the following calculus. (We write L- ~ for the fact ~ E K~ @ X. Also, ~-Bc
means that T is a substitution instance of a formula derivable in the calculus
of boolean logic.)
2.1. Syntax of Modal Logics 51
(bc.) R ~ if RBc ~ (bd-+.) k 5~(~ -+ ¢). --+ . D ~ --+ Die

(i < ,~)
(ax.) R~forall~EX (mp.)
@b.) ~R~~ (mn.) ~ (i < ~)
Thus, there are axioms (pc.), (ax.), (bd-+.) and the rules (mp.), (sb.) and
(mn.) (for all operators). Notice that the way in which we have stated the
rules they are actually so called rule schemata. The difference is that while in
an ordinary rule one uses propositional variables, here we use metavariables for
formulae. Thus, with an ordinary rule the use of a rule schema like R ~ / R [3i~
is justified from the rule ~- Po/ R [3ipo by uniform substitution of ~ for P0.
In view of the fact proved below that any proof can be rearranged in such a
way that the substitutions are placed at the beginning, we can eliminate the
substitutions since we have rule schemata.
Although the ordering of the application of the deductive rules (mp.),
(sb.) and (mn.) is quite arbitrary it is actually the case that any proof can
be reorganized in quite a regular way by rearranging the rules. Consider an
application of (mn.) after (mp.) as in the left hand side below. There is an
alternative proof of R N ¢ in which the order is reversed. This proof is shown
to the right.
F-¢
e De
~ ~ s ( ~ -+ ¢)
This shows that (mn.) can be moved above (mp.). However, this does not
yet show that all applications of (mn.) can be moved from below applications
of (mp.). A correct proof of this fact requires more than showing that the
derivations can be permuted. One also needs to show that the procedure of
swapping rule applications will eventually terminate after finitely many steps.
In this case this is not hard to prove. Observe that the depth in the proof
tree of the particular application of (mn.) decreases. Namely, if the depth of
R ~ is i and the depth of ~- p -+ ¢ is j then the depth of R ¢ is max{i,j} + 1
and so the depth of R [3¢ is max{i, j} + 2. In the second tree, the sequent
R [Np has depth i + 1 and the sequent R [E(~ --+ ¢) is j + 1. Both are smaller
than max{i,j} + 2. Let the deepest applications of (mn.) be of depth 5. By
starting with applications of depth 5 we produce applications of depth < 5,
52 2. ~ n d a m e n t a l s of Modal Logic I
so the instances of (mn.) of depth ~ can all be eliminated in favour of (twice

as many) applications of depth < & Now it is clear that the reduction will
terminate.
Next we look at substitution. The place of (sb.) in the derivation can be
changed quite arbitrarily. This is due to the fact that our rules are schematic,
they are operative not only on the special formulae for which they are written
down but for all substitution instances thereof. So, if we apply a certain rule,
deriving a formula ~ and apply (sb.) with the substitution or, then in fact we
could have derived ~ directly by applying (sb.) on all premises of the rule
and then using the rule.
Notice that (~ --+ ¢ ) " is the same as I- ~,o- _~. ¢~. Finally, note t h a t (mn.)
can be permuted with (sb.), that is, the two derivations below are equivalent.
R~ k~
R ~" R DjT
k D;f k (D3 )
Notice t h a t (D~) ~ = [B(~°), so the second derivation is legitimate. We have

now established that (sb.) can always be moved at the beginning of the proof.
Hence it can be eliminated altogether, as we have remarked earlier. The proof
that this commutation terminates is here simpler than in the first case.
We can now prove an important theorem about the relation between
quasi-normal closure and normal closure. To state it properly, we introduce
the following important bit of notation. Given a set A of formulae, put
N°A := A
NA := {[]i6:i<~,66 A}
Nk+IA := N(NkA)
N-<kA := Uj_<kNJA
N~A := Uk~, NkA
The notation [](k)A is also used for [~-<kA. In all these definitions, A m a y
also be replaced by a single formula. If, for example, n = 2 then
[]2~9 = {[~0[~0~, [~0[Z]l~, I--]l[-]0~ , E]IE]I~ }
[ ~ A is effectively the closure of A under all rules (mn.) for each operator.
2.1. Syntax of Modal Logics 53
THEOREM 2.1.5. Let A be a normal logic. Then A @ A -- A + Y~"~A.

Moreover, ¢ C A ® A iff ¢ can be derived from A U []"~A s by using modus
ponens alone, where A s is the closure of A under substitution.
PROOF. We can arrange it that a derivation uses (mn.) only at the begin-
ning of the proof. (mn.) is a unary rule, so it can under these circumstances
only have been applied iteratively to an axiom. Likewise, the use of substitu-
tion can be limited to the beginning of the proof. []
In case []kA is finite, we also write ~kA in place of A []kA and []<-kA in
place of A []-<kA- Here, for a finite set F of formulae, A F simply denotes
the formula A(7 : 7 E F). However, notice that the latter definition is not
unambigous. First, we need to fix an enumeration of F, say, F = {7i : i < n}.
Next, we let A F := Ai<n 7i. The latter is defined inductively by
Ai<0"yi := T
Ai<I 7i := 3'0
A~<27i := ~oA71
In the last line,n _> 2. Technically speaking, A F depends on the enumeration

of F. This will not matter as long as we deal with formulae up to deductive
equivalence. However, in syntactic manipulations (forexample, normal forms)
the differences need to be dealt with. Here we fix beforehand a well-order on
the set of formulae. This well-order defines a unique order on F.
Finally, we introduce the notion of a compound modality. Suppose that
is a formula such that ~var(p) = 1 (that is, ~ contains occurrences of a
single sentence letter) and which is built using A and []j, j , ~ . Then ~ is
called a c o m p o u n d m o d a l i t y . For example, [ZoE]lp A [11[]0p is a compound
modality, and so is []o(pA[~lp). We say that p is n o r m a l in A if from ¢ E A
w e may infer ~ ( ¢ ) E A, and if ~P(¢1 --~ ¢2)" --+ "~3(¢1) --+ ~P(¢2) ~ A. So, a
normal compound modality satisfies the same axioms and rules as a normal
operator, except that it is not necessarily a primitive symbol of the language.
PROPOSITION 2.1.6. Let T(p) be a compound modaIity. If all operators

occurring in ~ are normal, so is ~.
PROOF. We show by induction that the compound modalities admit

(mn.) and satisfy (bdA.). We leave it to the reader to verify that the op-
erators are classical. Let p(p) = ¢1(p) A ¢2(P). Assume that ¢1 and ¢2 are
normal. Then
~(pAq) I- ¢ l ( p A q ) A ¢ 2 ( p A q )
~- ¢1(P) A ¢1(q) A ¢2(P) A ¢2(q)
~- ¢1 (P) A ¢2 (P). A "¢1 (q) A ¢2 (q)
A
and conversely. Thus T satisfies (bdA.). Now assume ? X. Then F- %bl(X) and
~- %b2(X), by assumption. Thus ~- %bl(X) A %b2(X), that is, ~- T(X), as required.
Now let ~ = [3j¢. Then we have
A q) Dj¢(p A []j(¢(p) A siC(p) A Die(q),
and conversely. Furthermore, from ~- X we may conclude ~- ¢(X), by normality
of ¢, and ~- [3j¢(X), by normality of [Nj. []
It follows that a compound modality can be rewritten into the form

Ai<n¢i(P), where each ¢i(P) consists just of a string of modal operators
prefixed to the variable. We will use the following notation for such opera-
tors. If cr is a finite sequence in n write Z]~ for the operator prefix obtained
in the following way.
Moreover, for the empty sequence e, [Z~ wilt be the empty prefix. So, K]~ =
(the two are syntactically equal). Modulo deductive equivalence in K~ any
compound modality is of the form Ai<k [Z~P • Let s be a finite set of finite
sequences in ~. Then
K]SP := A [~'P
~'ES
The following theorem is easily proved.
PROPOSITION 2.1.7. Any compound modality is deductively equivalent to
a compound modality of the form []~p, where s a finite set of finite sequences
of indices.
To make life easy, we use [] as a variable for an arbitrary compound
modality. With Y a set we write []Y for {[]6 : 6 E Y}.
PROPOSITION 2.1.8. Let A be a normal modal logic, X a set. Then ~ C
A ® X iff there is a compound modaIity [] and a finite set Y C__X s, X ~ the
closure of X under substitution, such that N Y ~A ¢.
PROOF. Clearly, we know that if ¢ E A @ X then ¢ can be derived from
A and a finite subset Y C_ X ~ using modus ponens and (mn.). Then ¢ can be
derived from h and a finite set Z of formulae of the form []iXi, i < m, Xi E Y,
using only (mp.). Put [~p := Ai<,~ Nip. Then [~Xi ~-K~ ~i)~i" Thus the
compound modality [9 and the set {Xi : i < m} fulfill the requirements. []
Let 7R1 and []2 be two compound modalities. We write []1 <A []: if
[~2P -+ [~lP E A. It is not hard to see that this ordering is transitive and
reflexive. P u t [~1 ~A m2 if both ~1 <A []2 and []2 <_A []1. We define
(~1 0 ~]~2)(P) := [::~t([~2P)

2.2. Modal Algebras 55
PROPOSITION 2.1.9. The compound modalities, factored by the equiva-

lence ~h, form a semilattice with respect to tJ and a monoid with respect to
o and [~. Moreover, the following distribution law holds.
m l o (m2 u o u o
The proof of this theorem is straightforward.
E x e r c i s e 31. Show that in classical logic, the formula ded(~, ~) defined in

Section 1.6 is equivalent to ACE~ ¢" -4 "~"
E x e r c i s e 32. Show that negation is a classical modal operator but not

monotone. Show also t h a t [Di~ ++ ~0i-%o is a theorem in E~.
E x e r c i s e 33. Show that in a classical logic, (mo[2.) is interderivable with

the rule (moO.).
~- ~--+ ¢
(moO.) -+ 0¢
E x e r c i s e 34. Prove Propositions 2.1.2 and 2.1.3.
E x e r c i s e 35. The least modal logic is the logic which is just the closure of
the tautologies of boolean logic under substitution and m o d u s ponens. Show
that in the minimal modal logic p is a theorem iff there exists a substitution
cr such t h a t ~ -- ¢~ for some non-modal, boolean tautology ¢. Show further
that X F- ~o iff there exists a substitution c~ such that X = Y~ and ~ = ¢~ for
some n o n - m o d a l Y and ¢ such that Y t- ¢. Thus the minimal modal logic is
decidable. Hint. Show that in each proof subformulae of the form [Dj~ can
be replaced by a variable p~ for some i.
E x e r c i s e 36. Show that all compound modalities are classical if the basic
operators are classical.
2.2. M o d a l A l g e b r a s
From the general completeness theorems for logics we conclude t h a t for
modal logics of any sort there is a semantics based on t e r m algebras and
deductively closed sets. Moreover, consequence relations for modal logics of
any kind are determined by matrices of the form (~m(var), A), where A is
deductively closed, or, to be more general, by matrices (92, D) where 92 =
(A, 1, - , V~, ( ' i : i E n)) is an algebra of an appropriate signature and D a
deductively closed set. We focus here on algebras whose reduct to the boolean
operations is a boolean algebra. Call an e x p a n d e d b o o l e a n a l g e b r a an
algebra (A, 1, - , D, F), where (A, 1, - , D) is a boolean algebra and F a set of
functions. We are interested in the question when a modal logic is complete

with respect to matrices over expanded boolean algebras.
THEOREM 2.2.1. Let h be a classical modal logic. Then h is determined
by a class of reduced matrices over expanded boolean algebras.
PROOF. Let 2ira(vat) be the algebra of terms. Define an equivalence
relation (9 on the set of formulae by ® by ~ ® ¢ iff ~ ++ ¢ C A. By
Proposition 2.1.1, e is a congruence relation on Tm(var). Moreover, if 5
is a formula and A a deductively closed set then either [5]0 A A = ~ or
[5]® C_ A. Hence ® is admissible in any matrix 8X :-- (~m(var), A), where A
is a deductively closed set with respect to [-h. 8X/O is an expanded boolean
algebra. Now, by the results of Section 1.5, [-h is the intersection of all
b ~ , where 87~ -- (Tm(var), A), A deductively closed in hA. Hence it is the
intersection of b ~ / o . []
The converse of this implication is not generally valid. This theorem explains
the fundamental importance of classical logics in the general theory of m o d a l
logic. We can now proceed to stronger logics and translate the conditions
that this logic imposes into the language of expanded boolean algebras. For
example, the postulate of monotonicity is reflected in the following condition.
(mna.) If a <_ b then m~a <_ mib
In the general theory it has been customary to reserve the t e r m modal algebra
for expanded boolean algebras which correspond to normal modal logics.
DEFINITION 2.2.2. A (poly-) m o d a l algebra is an algebra
-- ( A , I , - , N , ( ' , : i < ~})
where the following holds.
1. (A, 1 , - , N) is a boolean algebra.
2. I~1 = 1 and , ~ ( x N y) = mix. n . l ~ y for all i < ~ and x , y E A.
With n given, P2 is called x - m o d a l .
Let P2 and ~ be two boolean algebras. A m a p h : A -~ B with h(1) = 1
and h(x N y) = h(x) N h(y) is called a h e m i m o r p h i s m . The name derives
from Greek hemi- (half) and morphg (shape), just like homomorphism is from
Greek homo- (same) and morph~. So, the name says that a h e m i m o r p h i s m
preserves only half the shape. By definition, then, a modal algebra is a boolean
algebra expanded by a set of endo-hemimorphisms. We will expand on this
theme in Section 4.5.
The abstract machinery of general logic can provide us now with the
canonical definition of a model. A model consists of a matrix and a valuation.
A m a t r i x is a pair consisting of an algebra and a deductively closed set. In
classical logics we have seen that the algebras can be reduced to expanded
boolean algebras and that we can choose maximally consistent sets, that is,
ultrafilters. The general completeness theorem says that if in a logic A we
have X JZh W then there is a model for A such that X holds in the model but
does not.
DEFINITION 2.2.3. A n algebraic m o d e l is a triple 9)I = (91,/3, U> where
91 is a modal algebra, /3 a map from the set of variables into A and U an
ultrafilter in 91. We say that ~ holds in 9)~, in symbols 9)~ ~ ~, if-~(~) E U.
We write <91,U> ~ ~ if for all/3, (91,/3, U> ~ ~ and we write 91 ~ ~ if for all
ultrafilters U and all valuations t3 we have <91,/3,U> ~ ~.
PROPOSITION 2.2.4. Let 91 be a modal algebra. Then 91 ~ ~ iff for every
/3, -g(~) = 1.
PROOF. Suppose that 91 ~ ~. Then (91,/3, U} ~ ~ for all valuations/3 and
ultrafilters U. Hence, given/3, ~ ( ~ ) C U for every ultrafi]ter. So, g ( ~ ) -- 1 for
all g. Now assume 91 ~ ~. Then there exists a valuation/3 and an ultrafilter
U such that g ( ~ ) ~ U. Hence for this/3, ~ ( ~ ) ~ 1. []
PROPOSITION 2.2.5. Let ffI~ = (91,/3, U} be an algebraic model. Then
5.) i#
(ii.) 97t ~ ~ A ¢ iff ff~;t ~ ~ and gX ~ ¢.
This notion of algebraic model was chosen to contrast it with the geometric
models based on Kripke-models. However, recall from Chapter 1.5 the notion
of a unital semantics. A class of matrices based on modal algebras is called a
unital semantics in the sense of that definition if it has at Inost one designated
element. Since the set of designated elements is deductively closed, and so is a
filter, in an algebra of the class of modal algebras there is always a designated
element and it is the unit element, 1. By the Proposition 1.5.6, for a modal
logic which has a unital semantics we must have p; q; ~(p) F- ~(q). Putting -V
for p we get q; ~(-7) ~- ~(q). In particular, for ~,(q) := [Zjq we deduce that
q ~- E]jq. Although this is not a rule of the standard consequence relation, we
will show in Chapter 3.1 that for each logic there exists a consequence relation,
denoted by I~-h, with the same tautologies, in which the rules ({p},[~jp>,
j < a, are derived rules. Since we are mostly interested in tautologies only, it
is justified to say that there is a unital semantics for modal logic.
For an algebra 91 we write Th 91 -- {~ : 91 ~ ~} and for a class :K of algebras
we put Th K :-- A(Th 91 : 91 C K>. Furthermore, for a set A of formulae we
write AIg(A) to denote the class of algebras such that 91 ~ A. The operators
Th and AIg are antitonic. T h a t means, if tK C /5 then Th L ~ Th K, and if
A C_EthenAlgA_~AlgE.
PROPOSITION 2.2.6. Let A be a set of ~-modal formulae, and :~ a class
of t~-modal algebras. Then the following holds.
I. A c_ T h ~ i f f A I g A _D ~ .
2. A C _ T h A I g A .
3. K C_ AIg Th ~C.
4. A l g a = A I g T h A I g A .
5. Th ~ = Th A I g T h K .
PROOF. (1.) A C_ Th ~ iff for all 92 E :K we have 92 ~ A iff for all 92 C 9~

we have 92 E Aid A iff ~: C_ Aid A. (2.) From Aid A C_ Aid A we deduce with
(1.) t h a t A C_ T h A I g A . (3.) From Thg( _C Thg~ and (1.), ~ C_ A I g T h ~ .
(4.) By ( 3 . ) , A l g a C_ AIg T h A I g A . By (2.), A _CThAIgA, and so AlgA _D
Aid Th Aid A. The two together yield the claim. (5.) Analogous to (4.). []
Hence, the maps :~ ~-+ AIgTh ~ and A ~-~ Th Aid A are closure operators.
The closed elements are of the form Aid A and Th :~, respectively. The next
theorem asserts t h a t the closed elements are varieties and normal modal logics,
as expected.
PROPOSITION 2.2.7. For all A , A l g a is a variety of ~-modal algebras.
For all ~ , Th K is a normal x-modal logic.
PROOF. For the first claim it will suffice to show that Aid {~} is a variety.
For in general, Aid A = N~e/x Aid {~}. It is left as an exercise to verify t h a t
the intersection of varieties is a variety. So, let ~ be given. We have to show
t h a t Aid {T} is closed under products, subalgebras and homomorphic images.
First, if Th 92i _D ~ then also Th l'-[iei 92i _D ~). For let ~ : = Hi@I~ti and
qi : ~ --" 92i be the projection onto the ith component. Let ~, be a valuation
on ~ . Then gi := qio'7 is a valuation on 92~, and we have g~(~) = 1. However,
g i ( ~ ) = q i o ~ ( ~ ) - Hence, for a l l i E I , q i o ~ ( ~ ) = 1. Thus 7(T) = 1. So,
~ T. This shows closure of Aid • under products. Next let i : q~ --~ 92 and
92 ~ T. Suppose V is a valuation into ~ . Then g := i o V is a valuation into
92. By assumption, g ( T ) = 1. However, g ( ~ ) = i o 7(~) = 1; hence V(T) --- 1,
since i is injective. Thus Aid {T} is closed under subalgebras. Finally, we
show that if h : 92 --~ ~ then Th ~ _D Th 91. Now suppose that "y is a valuation
on ~ . Take a valuation g such t h a t h(g(p)) = ~/(p). Then 1 = ~ ( T ) implies
1 = h ( g ( ~ ) ) = ~(~). Thus, Aid {T} is also closed under homomorphic images;
and so it is shown to be a variety. Now, take a class :K of modal algebras. We
want to show t h a t its theory is a modal logic. We leave it to the reader to
verify t h a t the intersection of modal logics is again a modal logic. Hence we
m a y specialize on the case :K = {92}. Clearly, since 92 is an expanded boolean
algebra, Th 92 contains all boolean tautologies. Moreover, by definition, it is a
classical logic, contains [:]j~- and satisfies (bdA.). Hence we have a monotonic
logic by the first fact, and we have (bd--+.) by the equivalence of the latter
with (bdA.) in monotonic logics. []
Let us now note that we have shown that each set of formulae gives rise to
a variety of algebras. They can be obtained rather directly by appeal to
Theorem 1.5.5. It says t h a t a logic is determined by its reduced matrices.

Now define an equivalence -------Aby ~a = h ¢ iff ~a ~ ¢ C A. If A is classical, this
is a congruence. Hence put
A(var) :=
The boolean reduct of ~rA(var) is a boolean algebra. So, the algebra is an
expanded boolean algebra.
LEMMA 2.2.8. Th ~rA(var) = A.
PROOF. Suppose ~b ¢ A. Then ¢ ~ h T and so because for the na-

tural m a p , : ¢ ~-+ ¢ / --=A we have u(T) ¢ 1. Therefore we also have
(~rA(var), {1}) ~ ~. Hence there is an ultrafilter U not containing ~, and
for t h a t ultrafilter (~rA(var),u, U) ~ ~a, as required. On the other hand, if
E A and h : ~£m~(var) --4 ~ h ( v a r ) is a homomorphism, then let cr be a
substitution defined by a(p) = Cp for some Cp such that h(p) = r,(¢p). Then
h(p) = u o ~r(p) and so h = r, o or. In particular h(p) = u o a(T). Since p e A
we also have a(~a) e A by closure under substitutions. Thus u(a(T)) = 1, by
definition. Consequently, h(T) = I for all h, showing that ~:A(var) ~ ~P. []
This last theorem is extremely important. It tells us not only t h a t each logic
has an adequate set of algebras, it also tells us the following.
THEOREM 2.2.9. The map AIg is a one-to-one map from normal n-modal
logics into the class of varieties of n-modal algebras.
PROOF. Clearly, we have shown that each set of formulae defines a variety
of n - m o d a l algebras, and each class of n - m o d a l algebras defines a normal n -
modal logic. Furthermore, for two logics A ~ (9 the varieties must be distinct,
because either A ~ (9 or (9 ~ A. In the first case ~rh(var) ¢ AIg(9 and in the
second case ~re)(var) ~ AlgA. []
This shows that algebraic semantics provides enough classes to distinguish

logics. We will see in Section 4.2 that distinct varieties give rise to distinct
logics, so that the correspondence is actually exact. Notice also the following.
The cardinality of the free algebra is at most l t ~ ] l . Moreover, for a coun-
termodel of T we only need to consider finitely generated subalgebras of that
algebra.
THEOREM 2.2.10. Let A be a n-modal logic with at most countably many

variables, n > O. If ~ ~ A then there exists an algebra 92 of cardinality
<<_max{l%, n} such that 92 ~ ~. In particular, if n is at most countable, so is
92.

FIGURE 2.1
o/\
E x e r c i s e 38. Let ~3 ~ 1-[i~ 921. Show t h a t Th ~3 = NieI Th 92i.
E x e r c i s e 39. Show t h a t the (possibly infinite) intersection of varieties is a va-
riety again. Likewise, show t h a t the (possibly infinite) intersection of n o r m a l
m o d a l logics is a n o r m a l m o d a l logic again. Hint. Use closure operators.
2.3. K r i p k e - F r a m e s and Frames

T h e most intuitive semantics for normal (and also q u a s i - n o r m a l ) logics
is based on relational structures, so-cMled frames. A frame is defined in two
stages. First, a K r i p k e - f r a m e for T~ is a pair ~ = (f, (<1i : i < n)) where f
is a set, called the s e t o f w o r l d s , and each <1i, i < n, is a b i n a r y relation.
<1i is called an a c c e s s i b i l i t y r e l a t i o n . More precisely, <1i is the accessibility
relation a s s o c i a t e d w i t h [~i- It is not required t h a t the set of worlds be
nonernpty. Frames can be pictured in m u c h the same way as directed graphs.
In fact, with only one accessibility relation, the two are one a n d the same
thing. T h e worlds are denoted by some symbol, say *, and the relation is just a
collection of arrows pointing from a node x to a node y just in case x <1y. Some
authors do not use arrows; instead t h e y have an implicit convention t h a t the
arrows point from left to right or from b o t t o m to top. (This is similar to the
conventions for drawing lattices.) Especially when there is only one relation,
several s h o r t h a n d notations are used. First, o s t a n d a r d t y denotes a reflexive or
self-accessible point, while • denotes an irreflexive point. Figure 2.1 illustrates
this. T h e r e are four points, one is irreflexive. A n o t h e r convention is to use
• for reflexive and x for irreflexive points. In the case of more t h a n one
relation, K r i p k e - f r a m e s are the same as edge-coloured directed multigraphs.
Technically, colours are realized as indices decorating the arrows. Notice
t h a t the shorthands for reflexive and irreflexive points are now insufficient,
so it is generally b e t t e r to use a subscripted turning arrow instead (O1 for
example). In t h a t case we will use • for worlds t h r o u g h o u t . A s u b s t i t u t e
policy for a p o l y m o d a l K r i p k e - f r a m e ~ = (f, (<1i : i < ~}} is to present it as
a set {[i : i < ~}, [i = (f, <1i), of m o n o m o d a l frames. This makes d r a w i n g
of the frames simpler, a l t h o u g h imagining such a frame is m o r e difficult as
2.3. K r i p k e - F r a m e s and Frames 61
FIGURE 2.2
o0 o,1
c o m p a r e d to the coloured graphs. Each particular pair (x, y} such t h a t

x <~j y is also called a j - t r a n s i t i o n or simply t r a n s i t i o n of the frame. We
also write x ~ y for the fact t h a t there is a j - t r a n s i t i o n from x to y. This
n o t a t i o n is generalized to sets of sequences s in the following way. Given a set
of sequences s the symbols x ~ y and x 2+ y are synonymous. For a sequence
% x <1~ y is defined by induction on the length. If T = e (the e m p t y string),
x ~ y iff x = y; if T = ~-1 J, J < a, then x <~ y iff there exists a z such t h a t
x ~ r l z and z ~ j y. Finally, for s = (Ti : i < n}, where each Ti is a sequence,
x<l s y i f f x < ~ i y f o r s o m e i < n .
Typically, the idea of these pictures is grasped r a t h e r quickly once the
reader has played with t h e m for a while. Readers who wish to have more
examples of K r i p k e - f r a m e s might take a m a p of the bus connections of some
area. T h e points are the bus stops, and each bus line defines its particular
accessibility relation between these stops. A p o i n t e d K r i p k e - f r a m e is a
pair ([, x} where x C f . A model based on [ consists of two more things, a
valuation and a specification of a special reference world. A v a l u a t i o n is a
function/~ : var -+ 2 f. A K r i p k e - m o d e l is a triple 9Y~ := ([,t3, x}, where
x E f and t3 is a valuation on [. For every proposition in [P~ we can now say
whether it is accepted or rejected by the model. This is defined formally as
follows.
(redO.) ([,/3, x) ~ p iff x E/3(p)
(md-~.) ([,/3, x} ~ ~ iff (f,/3, x) ~
(mdA.) ([,~3, x) ~ ~ A ¢ iff (f, t3, x) ~ qp and ([,13, x) ~ ¢
(md[]i.) ([,~3, x)~Di~ iff forallywithx<~iy ([,t3, y ) ~
Notice t h a t when defining the model condition for a formula W it is not ne-
cessary to assume t h a t fl is defined on all variables; all t h a t is needed is
t h a t it is defined on all variables of ~. F r o m a theoretical point of view it is
mostly preferrable to assume fl to be a total function. However, for practical
c o m p u t a t i o n and decidability proofs (see Section 2.6) we will prefer fl to be
a partial function, defined at least on the relevant variables. If fl is a partial
function we also say t h a t it is a p a r t i a l v a l u a t i o n .
Given a valuation/3, we can extend/3 to a m a p / 3 which assigns to each

formula the set of worlds in which it is accepted.
(hex.) ~ ( ~ ) = { x : (~,/3, x) ~ ~}.
It is not hard to see that ~(--~) = S - ~ ( ~ ) and that ~ ( ~ A ¢) = ~ ( ~ ) N ~ ( ¢ ) .
Therefore, ~ is a boolean homomorphism from the boolean algebra of modal
propositions into the powerset algebra of S. Moreover,
= {x: (vy)(x y. .(f,/3,y}

Thus define the following operation on subsets of f
(algO.) R~a :-= { x : (Vy)(z <1i y. =~ .y • a)}.
Then/3(Oj~) = Ij/3(~).
DEFINITION 2.3.1. A (polymodal) f r a m e is a pair ~ -= ([,F) where ~ =

(f, (<~i : i < ~)) is a Kripke-frame and F a set of subsets of f such that
(F, f , - , R, (I~ : i < ~}) is a polymodal algebra. Alternatively, F is a set
of subsets closed under boolean operations and the operations I i defined via
(alg[2.) on the basis of the relations <~i underlying ~. A p o i n t e d f r a m e is a
pair (5, x) where ~ is a frame, and x • f a world.
Standardly, frames in our sense are called generalized frames. For the
purpose of this book, however, we want to drop the qualifying phrase gener-
alized. This has several reasons. First, the nongeneralized counterparts are
called Kripke-frames, and so there will never be a risk of confusion. Sec-
ond, from the standpoint of Duality Theory it is generalized frames and not
Kripke-frames that we should expect as natural structures. (See C h a p t e r 4.)
And third, given that there exist numerous incomplete logics it is not a luxury
but simply a necessity to use generalized frames in place of Kripke-frames.
In a frame ([,F), a set a C__f is called i n t e r n a l or a s i t u a t i o n if a • F.
If a ~ F, a is e x t e r n a l . So, a frame combines two things in one structure: a
K r i p k e - f r a m e and an algebra. Due to the presence of the underlying relational
structure, it is unnecessary to specify the operations of t h a t algebra and so it
shows up in an impoverished form only as a set of sets. From an ideological
standpoint we may call frames also realizations of algebras. More on t h a t
in Section 4.6. A v a l u a t i o n into a frame is a valuation into the underlying
Kripke frame which assigns only internal sets as values. Since the set F is
closed under all relevant operations, the following is proved by induction on
the structure of the formula.
PROPOSITION 2.3.2. Let (~, F) be a frame and t3 : var--+ 2 / be a valuation

on ~. If/3(p) is internal for all p • v a r then -~(~) is internal for all ~.
Moreover, -~ is a homomorphism from ~ m ~ . ( v a r ) to (F, 1, - , A, (mj : j < n)).
2.3. Kripke-Frames and Frames 63
A ( g e o m e t r i c ) m o d e l based on the frame 3 is a triple (3,/3, x), where

/3 is a valuation and x a world. Notice that models based on frames as well
as Kripke-frames are called geometric because they use a world to evaluate
propositions. Recall that in the algebraic model we had ultrafilters; an alge-
braic model is a triple (92,/3, U), where U is an ultrafilter on 92. We will see
in Chapter 4 that every modal algebra is isomorphic to an algebra of inter-
nal sets over some Kripke-frame. Thus the principal difference between the
algebraic and the geometric approach is the use of ultrafilters as opposed to
worlds. In that respect an algebraic model is still more flexible. For if we have
a world x, we also have a corresponding ultrafilter, U~ := {a : x E a}. But
not every ultrafilter is of this form. There exist, however, classes of frames
where every model based on an ultrafilter has an equivalent model based on
a world. These frames are called descriptive. More on that in Section 4.6.
Now we come to the interaction between models and logics. Consider a
model D2 = (3,/3, x). Define Th(~)I) := {~o: ~ > ~o}. Then Th(ffJI) is closed
under modus ponens. Moreover, for each ~o we have either ~o E Th(ff)~) or
-~W C Th(~3~) but not both, by (md~.). So, the set Tb(~))I) is a theory (by
mp-closure) and a maximally consistent theory. Now, suppose we abstract
away from the valuation; that is, we take the pointed frame (3, x) and define
(tpf.) (3, x) ~ ~ <=> for all valuations/3 we have (3,/3, x) ~ ~o
Then the theory of the pointed frame Th (3, x) = {~o : (3, x) > ~o} is still
closed under modus ponens; however, we no longer have either W E Th (3, x)
or -~o E Th (3, x), even though ~oV ~ E Th (3, x). Simply take ~o := p, where
p is a variable. However, Th (3, x) is closed under substitution. Hence, it is a
quasi-normal logic, since it contains all classical tautologies and (bd-+.). In
a last step we abstract from the worlds and consider the theory of the frame
alone.
(tfr.) 3 ~ ~o ¢=> for all worlds x and all valuations/3 (3,/3, x) ~ ~o
Th 3 := {~o : 3 ~ ~o} is called the t h e o r y o f 3- This time we not only have
closure under (mp.) and (sb.) but also under (mn.). Suppose, namely, that
~o E Th 3. Then for all points and all valuations (3,/3, y) ~ ~. Take a valuation
/3 and a point x. Since for all y such that x ~ y we already have (3,/3, Y) > ~,
we now have (3,/3, x) ~ [~i~o. Hence, since both x a n d / 3 have been chosen
arbitrarily, [7~o E Th 3-
THEOREM 2.3.3. Let (3, x) be a pointed frame. Then Th (3, x) is a quasi-
normal logic. For all frames Th 3 is a normal logic.
Now, given a quasi-normal logic A and a pointed frame (3, x), we say
that (3, x) is a p o i n t e d f r a m e for A if Th (3, x) _D A; likewise, 3 is a f r a m e
for A or a A - f r a m e if Th 3 D A. For a given logic A we denote the class
of A-frames by Frrn(A) and the class of A-Kripke-frames by Krp(A). We
FIGURE 2.3
f@
0
~A
w =,
0~ 0,1
conclude this section with an important theorem concerning the generated

substructures. Consider a generalized frame (~, F) and an internal set g E F.
g is called o p e n if for all x E g and all y such that x <~j y for some j then
also y C g. So, g contains all successors of points contained in g. We put
g
<~j := <~j M (g × g) and g := (g, (<~g: j < ~)). 9 is a Kripke-frame. Finally,
G := {aMg : a E F} = {b C g : b e F}. G is closed under relative complements;
for if b C G then g - b = g N ( f - b) E F; G is also closed under intersection.
Furthermore, if b E G, then
• ~b = {x~g:(Vy)(x<ly ~yCb)}
= (x e g: (vy)(x <5 y y e b)}
= gNijb,
since g is successor closed. So the m a p b ~-~ g n b is in fact a h o m o m o r p h i s m
of the modal algebras. Thus ® := <~, G} is a frame; we say, it is a g e n e r a t e d
s u b f r a m e o f 5. We denote this by ~ < 5. In the picture below the box
encloses a generated subframe.
THEOREM 2.3.4. Let ® be a generated subframe o f ~ and x E g. Then
Th <®, x> = Th <~, x>. Moreover, Th ~ _D Th 5.
N o t e s on this section. The idea of a Kripke-frame is generally credited to
SAUL KRIPKE ([134]), though it can already be found in works of RUDOLF
CARNAF ([38]) and SwIG KANGER. Also, BJARNI JONSSON and ALFRED
TARSKI in [113] presented a fully fledged algebraic theory of algebras for
modal logic. In it they also show that certain logics have the property t h a t the
algebraic structures of that logic are closed under completion (see Section 4.6),
which they use to show that these logics are complete with respect to K r i p k e -
frames with certain properties. The notion of a general frame first appeared
with S. K. THOMASON in [206]. Before that it was customary to use the
notion of a model, which was just a Kripke-frame together with a valuation.
In the language of generalized frames, a model was equivalent to a generalized
frame in which the internal sets were exactly the definable sets.
2.4. Frame Constructions I 65
FIGURE 2.4. A contraction
o 1 o,1 " " 4

00 0,i 00
E x e r c i s e 40. Prove T h e o r e m 2.3.4.
E x e r c i s e 41. Let ~(p) be a c o m p o u n d modality. We say t h a t ~(p) is based

on <~, if for all models: (5,/3, x) ~ ~(p) iff for all y ~> x we have (~, Z, Y) ~ P-
For example, [Zj is based on <~j. Show t h a t if p is a c o m p o u n d m o d a l i t y
based on <~ and ¢ is a c o m p o u n d m o d a l i t y based on • t h e n ~ A ¢ is based
on < U • , and ~[¢/p] is based on <~ o • .
E x e r c i s e 42. (Continuing the previous exercise.) Show t h a t [p?] defined by

[p?]q :-- p A q is a n o r m a l m o d a l operator. O n which relation is this o p e r a t o r
based?
2.4. F r a m e C o n s t r u c t i o n s I
In this section we will introduce a n u m b e r of ways of creating frames from

frames; in particular, these are the subframes, p-morphic images and disjoint
unions or direct sums. These notions will first be introduced on K r i p k e - f r a m e s
and then lifted to (general) frames. T h e most i m p o r t a n t notion is t h a t of a
p-morphism. Let ~ : f --+ g be a map. 7r is a p - m o r p h i s m from ~ to 9, in
symbols 77 : f --+ 9, if the following holds.
(pro1.) Forx, yef: ifx<ljythenzT(x)<~j~(y)
(pm2.) ForxC f anduCg: if77(x)<~juthenthereisayE fsuchthat
u = 77(y) and x <lj y.
We refer to ( p m l . ) as the first condition and to (pm2.) as the second condi-

tion on p - m o r p h i s m s . If 77 is injective, we write 7r : f ~ 9- If in addition 7r is
the identity on g, 9 is a generated subframe of [. If 77 is surjective we call 77 a
c o n t r a c t i o n and say t h a t 9 is a p - m o r p h i c i m a g e or c o n t r a c t u m (plural:
contracta) of ~. We write 77 : ~ ~ 9- Figure 2.4 provides an example. T h e
Kripke-frame on the right is a c o n t r a c t u m of the K r i p k e - f r a m e on the left.
A n o t h e r example is (w, <). T h e one-element K r i p k e - f r a m e consisting of a
reflexive point is a c o n t r a c t u m of (w, <). Each contraction 77 : f --~ 9 induces
an equivalence relation ~ on f by x ~ . y iff ~(x) = ~r(y). T h e equivalence
relations induced by p - m o r p h i s m s can be characterized intrinsically as fol-

lows. A n e t on ~ is an equivalence relation ~ such t h a t if x ~ x' and x <~j y
t h e n there exists a y' ~ y such t h a t x' ~ j y'. This latter condition is called the
n e t c o n d i t i o n . Given a net ~ , define Ix] := { x ' : x ~ x ' } , and p u t [x] <~j [y]
iff there exist x' E Ix] and y' E[y] such t h a t x' <~j y'. We denote by ~/~ the
K r i p k e - f r a m e with worlds Ix], x E f , and relations as just defined. We leave
it to the reader to verify t h a t the m a p x ~-+ [x] : [ --+ [ / ~ is a contraction if
is a net, and t h a t if ~v is a contraction, then the equivalence ~ induced by 7r
is a net. Nets are just a suitable way to picture contractions.
PROPOSITION 2.4.1 (Net Extension I). Suppose that ~ ~-~ 9 and that ~ is
a net on ~. Define ~ C g x g by x ~ y i f f (i.) { x , y } C f and x ~ y or 5 i . )
{ x , y } ~ f and x = y. T h e n ~ is a net on g.
T h e p r o o f is easy and omitted. This t h e o r e m is of e x t r e m e practical

i m p o r t a n c e ; it says t h a t if c is a c o n t r a c t u m of a generated s u b f r a m e ~ of 9,
t h e n it is a generated s u b f r a m e of a suitably defined contraction i m a g e ~ of
9, see picture below. T h e m a p s denoted by dashed arrows are in some sense
unique (we will come to t h a t later), this is why we put an e x c l a m a t i o n m a r k .
l '
If/3 is a valuation, and ~- a p - m o r p h i s m , rr is called a d m i s s i b l e for/3 if
for every x E 9 and every set/3(p) either ~r-i(x) C_/3(p) or 7 r - i ( x ) C -/3(p).
In other words, the partition t h a t 7r induces on f must be finer t h a n t h e
p a r t i t i o n induced by the sets ~ ( ~ ) . In t h a t case we can say t h a t / 3 i n d u c e s
a valuation 7 on 9 by taking 3'(P) := {Tr(x) : x E /3(p)}. We say t h a t 7 is the
i m a g e of/3 under 7r. Moreover, we will also write /3 for the v a l u a t i o n 7 if
no confusion arises. It should be clear t h a t every valuation on 9 can be seen
as the image of a valuation on ~ under a contraction. Now take an a r b i t r a r y
p - m o r p h i s m 7r : f ~ 9- T h e n the following i m p o r t a n t t h e o r e m holds.
PROPOSITION 2.4.2. Let 7r : ~ ~ 9 and let 7r be admissible f o r / 3 . Then
f o r every x E f
PROOF. For variables this is true by construction; the steps for -7 and
A are easy. Now let ~ = 0 j ¢ . Assume ([,/3, x} ~ 0 j ¢ . T h e n there is a y
such t h a t x <~j y and ([,/3, y} ~ ~b. By ( p m l . ) , 7r(x) <~j 7r(y), and by induction
hypothesis (9,/3, Tr(y)} ~ ¢. This gives (9,/3, Tr(x)) b 0 j ¢ . Now a s s u m e t h a t
the latter holds. T h e n for some u with 7r(x) <~j u we have (9,/3, u} ~ ¢.
By (pro2.) there exists a ~ such t h a t x <~j ~ and 7r(~) = u. B y induction
hypothesis, (~,/3, ~} ~ ¢ and so (~,/3, x} ~ O j ¢ , as required. []
A remark on the proof. As is often the case, the induction is easier to

perform using 0j rather than [:]j. Although the latter is a primitive symbol
of the language, we allow ourselves for the purpose of proofs to take either as
primitive and the other as composite, whichever is best suited for the inductive
step.
Now let r~ : f --+ g be a p-morphism. Let im[7~] C_ g be the set of points
~(x) for x E f . This is the so-called i m a g e of 7r. It is a subframe of g.
Moreover, the following geometric analogue of the first Noether Isomorphism
Theorem holds.
PROPOSITION 2.4.3. Let 7e : ~ --+ g be a p - m o r p h i s m . Then there are

maps p : f --~ im[~] and ~ : im[r~] ,-~ g such that 7~ = ~ o p.
PROOF. P u t p(x) := ~(x) and ~(u) := u. First, we show t h a t im[Tc] is

a generated subframe. This shows that ~ is a p - m o r p h i s m . To that end let
u :-- ~(x) and u < ~ v . Then by (pm2.) there is a y such t h a t x < l i y and
~(y) = v. Hence, ~ : im[~] ~ g. Now, consider the m a p p. Take x, y with
x <~ y. Then, by (pml.) ~(x) <~ ~(y), whence p(x) <~ p(y). Finally, let
p(x) <i u. Then, by definition, ~(x) <~i u and by (pm2.) there is a y such that
x <~ y and p(y) = ~(y) = u. This shows p : f -~ im[~]. That p is surjective
follows immediately from the definition. []
Take now Kripke-frames, [i = (fi, ( <~} : j < ~) ), i E I. Let
@ s, : = U{i} ×
iEI iEI
be the disjoint union of the sets fl- (For simplicity, we standardly assume
that the sets fi are pairwise disjoint and then we put ~ i E I fi := [-JicI fi. In
general, this is not without complications, however.) Based on this set we can
define the frame ~ i E I ~i, called the d i r e c t s u m or d i s j o i n t u n i o n , via
<~ := {((i,x),(i,y)):ieI;x, yefi;x<lj y}

@ice fi := (@iCI fi, (<~?: j < @}
Intuitively, the direct sum consists of several components, and two points are
j - r e l a t e d iff they are from the same component and are j - r e l a t e d in that
component. The components are just placed next to each other, with no
interaction. The direct sum has the following property.
THEOREM 2.4.4. Let fi, i E I, be Kripke-frames. Then there exist am-

beddings ei : fi ~ @ i c I [i. Moreover, f o r every K r i p k e - f r a m e b with p -
morphisms ~i : ~i -+ ~, i E I, there exists a unique ~r : ~ i E I ~i --+ ~, which is
a p - m o r p h i s m such that ~ o ei = t~i for all i E I.
PROOF. It is easy to check that the identity embeddings el : fl -+ ~ i c s fi

are injective p-morphisms. Now assume that O is given and ~i : ~i --+ ~,
i E I. We have to define 7r. Take an element z E ( ~ i c z fi. T h e r e is an

i E I such t h a t z = {i,x} for some x E f i . P u t 7r(z) : = hi(x). 7r as defined
is a p - m o r p h i s m . For if z <1i z ~ for some z and z ~, then there is an i and
x , x ' e f i such t h a t z = (i,x} and z' = {i,x'}. By construction, x < l j x ' .
T h e n zr(z) = hi(x) <lj hi(x') = ~r(z'), by the fact t h a t hi satisfies ( p m l . ) .
Now assume ~r(z) "~i u, z = (i,x). T h e n 7r(z) -- hi(x) and by the fact t h a t
hi satisfies (pm2.) we have a x' such t h a t x <~2 x' and h i ( x ' ) = u. T h e n
7r({i,x')) = hi(x') = u. So, 7r is a p - m o r p h i s m ; moreover, by definition we
get (Tr o ei)(x) = ~r({i,x)) --- hi(x) for x e f/. Finally, let us see w h y lr is
unique. So let ~ be another m a p satisfying all requirements. Let x E fi-
T h e n ~({i,=}) = (~ o ei)(x) -- hi(x) = (~ o e d ( x ) = ~({i, x}). []
A valuation d on (~i~I fi uniquely determines a valuation fli on fi by

{i} x /3i(P) = 5(p) A ({i} × f/). Conversely, given a family /3i, i e I, of
valuations into fi there is a unique valuation t~ies/3/ on ( ~ i e ] ~i defined by
:= U{i} ×/3i(p)
iCI iEI
T h e following theorems are left as exercises.
PROPOSITION 2.4.5. Let 9, ~, fi, i 6 I, be Kripke-frames.
1. If g ~ f then Th(~:) C Th(g).
2. I f g ~ f then Th(g) C_Th(}:).
3. Th(@ic x fi) = ~iez Th(fi).
THEOREM 2.4.6. Let A be a normal polymodal logic. Then if f is a K r i p k e -
frame for A, so is any generated subframe and any p - m o r p h i c image. More-
over, any direct sum of Kripke-frames for A is again a Kripke-frame for A.
To generalize these notions to frames we need to consider w h a t h a p p e n s
to the internal sets. First, consider a subframe [ --+ g, and let ~ = {f, 1F} as
well as ® = (9, G}. For simplicity we assume t h a t f C g. T h e m a p a ~-> a A f
is a b o o l e a n h o m o m o r p h i s m from { G , g , - , @ to {2f, f , - , N}. Consider now
the fact t h a t any valuation/3 on g defines a valuation 7 on ~, namely, 7(P) : =
/3(p) N f . In order for theorems like P r o p o s i t i o n 2.4.5 to hold we need t h a t for
every v a l u a t i o n / 3 the corresponding valuation 7 is a valuation on ~. Hence,
we must require t h a t a ~-~ a N f is a h o m o m o r p h i s m from (G, g , - , 7~} onto
f, -, n>.
DEFINITION 2.4.7. A map 7r : f -+ g is an e m b e d d i n g of the frame
~" = (f, F} in the frame 05 = (9, G} if (0.) r: is injective, (1.) 7r(f) C C,, (2.)
7r : f --~ g is a p - m o r p h i s m and (3.) the map zc-1 : a ~ {x : 7r(x) e a}
is a surjective homomorphism from {G,g, - , N, {lli : i C h}} to the algebra
{F,f,-,A,{lll : i C h}}. If all that is the case we write p : ~ ~ 05. If
in addition f C g, and 7r the natural inclusion map we call ~ a g e n e r a t e d
s u b f ' r a m e of 05 and write 7r : ~ <_ 05 or simply ~ < 05.
T h e surjectivity of the m a p 7r-1 for embeddings is actually quite impor-

tant for generalizing the factorization theorem, P r o p o s i t i o n 2.4.3. It would
fail otherwise. Several remarks are in order. First, notice t h a t while the m a p
on the frames goes from ~ to ~, the corresponding m a p 7r-1 on the algebras
goes from (G, g, - , n, ( l i : i E •)) to (F, f, - , •, ( l i : i E t~)). Moreover,
for generated subframes, rather t h a n requiring t h a t the m a p is a well-defined
h o m o m o r p h i s m we require t h a t the m a p is onto, t h a t is, all sets of F are
restrictions of sets in G. A last point is the requirement t h a t the image of
f under ~r is internal. This is added for theory internal reasons, since if this
definition is generalized to subframes, this clause is needed. However, it can
be shown to be unnecessary for generated subframes.
Next we t u r n to contractions. Again, we s t u d y the m a p a ~ ~r-l[a]. This
is a boolean h o m o m o r p h i s m from 2g to 2/, and we need to make sure t h a t it
is also a h o m o m o r p h i s m of the modal algebras on the frames. Hence, consider
the set c : = 7 r - l [ l j a ] . It contains all y such t h a t for all u with 7r(y) < j u we
have u C a. On the other hand, the set d : = l j ( T r - l [ a ] ) is the set of all y
such t h a t for all z such t h a t y <~j z we have 7r(z) E a. If y C c and y <~j z then
re(y) • l j a . By ( p m l . ) 7r(y)<~jlr(z), hence 7r(z) • a, and so z • 1r-l[a]. T h u s
y • d. Now let y • d. To see t h a t y • c, we have to show t h a t 7r(y) • lja.
So, assume ~r(y) <~j u. By (pm2.) there is a z such t h a t 7r(z) = u and y < j z.
T h e n z • 7r-t[a], by assumption on y, and so u • a, as required. So, the m a p
a ~ lr - l [ a ] is indeed a h o m o m o r p h i s m of m o d a l algebras.
DEFINITION 2.4.8. Let ~ and C5 be frames and 7c : f --+ g a map. 7r is a

c o n t r a c t i o n f r o m ~ to 03 if (0.) ~r is surjeetive, (1.) 7r is a p - m o r p h i s m and
(2.) 7r - 1 i8 an injective h o m o m o r p h i s m f r o m ( G , g , - , A , ( l j : j C n}) into
(F, f, - , A, ( I j : j • ~)). / f all that is the case, we write 7r : ~ --* ¢5.
A nontrivial example is ft :-- (w, <, ©} with © the set of finite and cofinite
subsets of w. Take the K r i p k e - f r a m e In = ( { 0 , . . . , n - 1}, <~) where i <~ j iff
i < j o r i - - j = n - 1 . Next, l e t ~ n : = ( { 0 , 1 , . . . , n - 1 } , ~ l ) , w h e r e ~ = nxn.
It turns out t h a t In is a c o n t r a c t u m of ft, while ~tn is a c o n t r a c t u m only for
n = 1. However, ~, is a c o n t r a c t u m of (w, <), the underlying K r i p k e - f r a m e of
ft. (It follows t h a t also all [n are contracta of (w, <).) We can, finally, define
the notion of a p - m o r p h i s m for frames.
DEFINITION 2.4.9. Let ~ and ® be frames and 7c : f --+ g a map. TC is a

p - m o r p h i s m f r o m ~ to 6 , in symbols 7r : ~ --~ ¢~, if (1.) 7r : ~ --+ ~ is a p -
m o r p h i s m of the underlying K r i p k e - f r a m e s and (2.) 7r-~ is a h o m o m o r p h i s m
from the modal algebra ( G , g , - , A , ( ' j : j • t~)) to ( F , f , - , A , (mj : j • ~)).
In fact, given (1.) it is not necessary to require rc-1 to be a h o m o m o r -

phism. Rather, it is enough to require
(pro3.) If a • G t h e n Tr-l[a] • F.
PROPOSITION 2.4.10. A map h : f --+ g is a p-morphism from ~ to @ iff

it satisfies (pro1.), (pro2.) and (pro3.).
We refer to (pm3.) as the third p - m o r p h i s m condition. Notice t h a t
with the definitions given an e m b e d d i n g is an injective p - m o r p h i s m and a
contraction a surjective p - m o r p h i s m . A generalized frame is contractible to
a K r i p k e - f r a m e if the algebra of sets is finite. This is the content of the next
theorem.
THEOREM 2.4.11. Let ~ = ([, F) be a generalized frame and F finite. De-
fine an equivalence relation x ~ y on points by
x ~ y ¢ , (Va c F)(x • a. .y • a ) .
Put [x] := ( y : x ~ y}, [f] := {[x]: x • f } and [x] < j [y] iff there exist ~ • [x]
and ~ • [y] such that ~ <~j ~. Then x ~-~ [x] is a p-morphism from ~ onto the
# a m e (If], ( < j : j < 2Esl).
PROOF. We have to check the second clause of the p - m o r p h i s m condition.
(The first is satisfied by definition of <lj on [f].) Let [x] <lj u. T h e n there is a
y such t h a t u = [y]. We have to show t h a t there is a ~ such t h a t x <~j ~" and
~ y. To see this, let ay be the intersection of all elements in F containing
y (equivalently, let ay be the unique a t o m containing y). Since F is finite,
ay • F. T h e n ~ ~ y i f f ~ • ay i f f a ~ = ay. It is checked t h a t x <~jy iff
ax ~_ ~ja v. Now since [x] <~j [y] there are ~ • [x] and ~ • [y] such t h a t ~ <lj ~.
It follows t h a t a t _< a~. Since a t = a~ and a~ -- ay we conclude a~ <__Ojay,
from which x • ~jay. And so there is a yt ~ y such t h a t x <~j yq So, the m a p
is a p - m o r p h i s m . We have seen t h a t the classes p - l ( [ x ] ) are internal sets of
the form as defined earlier. Hence, each set has a preimage. []
DEFINITION 2.4.12. Let ~i, i • I, be frames. The d i s j o i n t u n i o n of the

~i, denoted by ~ i ~ I ~i, is defined as follows. Then underlying Kripke-frame
is ~ i e I ~" A set is i n t e r n a l if it is a union [.Jiei{i} × el, where ai • F~ for
each i • I.
PROPOSITION 2.4.13. Let ~i, i • I, be frames. There exist embeddings
ei : ~ ~-~ ~ E I ~ such that for all ® and embeddings di : ~i ~ ¢3 there exists
a p-morphism Tc : ~ I ~i -+ @ such that dl = 7c o ei for all i • I.
PROOF. Follow the proof of T h e o r e m 2.4.4. T h e construction is com-
pletely analogous. We only have to check t h a t the m a p ~r satisfies the third
condition on p - m o r p h i s m s . To this end consider an internal set a of @. Let
a~ := dT,~[a]. B y the fact t h a t d~ is a p - m o r p h i s m , this is an internal set o f ~ .
Now, ~r-l[a] = [ . J ~ { i } × d:(~[a]. By definition of ~ e ~ ~ ' this is an internal
set. []
2.5. Some I m p o r t a n t Modal Logics 71
A n e t on a generalized frame 3 is a net --~ on [ such t h a t for each a E F

the set [a]~ := { y : ( 3 x E a ) ( x ~ y)} is a m e m b e r of F. We write 3 / ~ for the
quotient, which is defined by
3 / ~ := (~/~, ({[x] : x E a } : a E F})
PROPOSITION 2.4.14 (Net Extension II). L e t 3 be a f r a m e , and 0~ < 3.
L e t ~ be a n e t on ®. L e t x ~ y iff (i.) x, y E g and x ~ y or h i . ) x, y E f - g
and x = y. T h e n ~ is a n e t o n 3 .
PROOF. In view of Proposition 2.4.1 we only have to check t h a t for each

a E F we have [a]~ E F. To see this, let b := a N g and c := a N ( f - g). Since
g E F we have b, c E F. T h e n [a]~ = [b]~ U c. By a s s u m p t i o n on ~ , [b]~ E F.
Hence the claim is proved. []
E x e r c i s e 43. Let ~ be a frame and f C g . Show t h a t the m a p a ~ + a N f

is a boolean h o m o m o r p h i s m from 2 g to 2 I. Hence, the condition (2.) of
Definition 2.4.7 is necessary only to ensure t h a t this m a p is a h o m o m o r p h i s m
of the m o d a l algebras, t h a t is, t h a t it c o m m u t e s with the m o d a l operators.
Show t h a t it is sufficient.
E x e r c i s e 44. Show t h a t if in Definition 2.4.7 we did not require 7r-1 to be

surjective, t h e n there were injective p - m o r p h i s m s which are not embeddings.
E x e r c i s e 45. Let 7r : 3 --~ ~5 be a p - m o r p h i s m and x E g. Show t h a t if there

is no chain of length k x = x0 <~j xl <~j x2 . . . <~j xk then the same holds for
~r(x) as well. Show t h a t if lr(x) Cj 7r(x) then ~r-l(Tr(x)) is a j - a n t i c h a i n , t h a t
is, for all y , z E lr-l(Tr(x)) y ~ j z.
E x e r c i s e 46. P r o v e Proposition 2.4.5 and T h e o r e m 2.4.6.
E x e r c i s e 47. F o r m u l a t e and prove Proposition 2.4.5 and T h e o r e m 2.4.6 for

(generalized) frames instead of K r i p k e - f r a m e s .
E x e r c i s e 48. Let [ = (f, (<~j : j < n)) be a K r i p k e - f r a m e and G a s u b g r o u p

of the group 92~t([) of a u t o m o r p h i s m s of ~. P u t
[x] := { y : there exists g E G : g ( x ) -- y } .
Also, put [x] <]j [y] if there exist ~ E [z] and ~ E[y] such t h a t ~ j ~ . Show t h a t
x ~+ Ix] is a p - m o r p h i s m . (An a u t o m o r p h i s m of f is a bijective p - m o r p h i s m
from f to [. T h e a u t o m o r p h i s m s of a structure generally form a group.)
2.5. Some Important Modal Logics

A m o n g the infinitely m a n y logics t h a t can be considered there are a num-
ber of logics t h a t are of f u n d a m e n t a l importance. T h e i r i m p o r t a n c e is not
only historical but has as we will see also intrinsic reasons. We begin with
logics of a single operator. Here is a list of axioms together with their stan-
dard names. (In some cases we have given alternate forms of the axioms. T h e
first is the one standardly known, the second a somewhat more user friendly
variant.)
Inc ±
4 (}0p -+ 0p
D 0T
B p -+ N 0 p
T p - + 0P
5 Op -+ @Op
altl (}p -+ [Ep
Op A Oq. -+ .O(p A q)
1 [E(}p -+ O[lp
2 ODp -+ [~Op
3 OP A Oq. --> .O(P A Oq) V O(q A Op) V O(P A q)
G []([]p -+ p) ~ []p
Op --+ O(p A =OP)
Grz [](D(p -+ []p) -+ p) -+ p
p -+ 0(p A S ( ~ p -+ []~p))
In addition, there are also logics with special names. For example, $4 is
K 4 . T , $5 is K 4 . B T or, equivalently, S4.B. (The dot has no meaning; it
is inserted for readability. Occasionally, several dots will be inserted.) The
letter S stems from a classification by C. I. LEWIS, who originally introduced
modal logic as a tool to analyse conditionals. He considered five systems,
called $1 to $5, among which only the last two - - namely $4 and $5 - -
were based on normal modal logics. D originally comes from deontic, since
this postulate was most prominent in deontic logic, the logic of obligations.
Nowadays, D is associated with definal, lit. meaning without end, because in
frames satisfying this postulate every world must see at least one world. G is
named after KURT G6DEL, because this axiom is related to the logic derived
by interpreting [] as provable in Peano Arithmetic (= PA); the logic K 4 . G is
called the p r o v a b i l i t y logic. Often G is also called G L , where L stands for
M. H. L6B, who contributed the actual axiomatization. In [201], ROBERT
SOLOVAYshowed that if [] is read as it is provable in PA that then the logic of
this modal operator is exactly K 4 . G . For the history of this logic see the en-
tertaining survey by GEORGE BOOLOS and GIOVANNI SAMBIN,[32]. Finally,
the axiom G r z is named after the Polish logician GRZEGORCZYK. Usually, the
logic S 4 . G r z is called GRZEGORCZYKs logic. Some authors use G for K 4 . G
and G r z for S 4 . G r z . The reason will be provided in the exercises; namely,
it turns out that K . G as well as K . G r z contain the axiom 4. We will use
2.5. Some Important Modal Logics 73
FIGURE 2.5. Important Extensions of K
Inc
Th~Th~-~
K.a rz
the same convention here, if no confusion arises. 1 is known as MCKINSEY'S

axiom, therefore also denoted by M. 2 is called the GEACH axiom.
For theoretical purposes, the following two infinite series of axioms are
important.
alt~ Ai<,~+I(~pi. --~ • V i < j < n + l O(Pi A pj)

trsm ~<_mp._+ .o<_m+lp
The first holds in a Kripke-frame iff any point has at most n successors, the
second holds in a Kripke-frame iff any point that can be reached at all can
be reached in m steps. There are not so many polymodal logics which have
acquired fame. However, the most useful logics occur as extensions of a logic
with several operators in which the logic of a single operator on its own be-
longs to one of the systems above, and the axioms specifying the interaction of
these operators are of a rather simple type. Therefore, the polymodal logics in
which there are no axioms mixing the operators, are an important basic case.
The following notation is used here. Given two monomodal logics, A1 and A2,
the symbol A1 ® A2 denotes the smallest normal bimodal logic in which the
first operator satisfies the axioms of A1 and the second operator the axioms
of A2. A1 ® A2 is called the f u s i o n or i n d e p e n d e n t j o i n of the two logics.
Similarly, the notation ~)i<~ Ai denotes the fusion of # many modal logics;
in general each A~ can be polymodal as well. In this logic the operator Vii
satisfies exactly the postulates of Ai. Given a logic A, the operator [] is called
m - t r a n s i t i v e if [-]<_mp ~ •<_m+lp E A. [] is w e a k l y t r a n s i t i v e if it is
m-transitive for some m. A polymodal logic A is called w e a k l y t r a n s i t i v e if
there is a compound modality [] such that for every compound modality []',
[]p -+ []'p C A. If ~ is finite, we call A m - t r a n s i t i v e if it contains the axiom
trsm below and w e a k l y t r a n s i t i v e if it is m-transitive for some m. This
notion of weak transitivity coincides with the one defined earlier for arbitrary
~, as can easily be demonstrated.
trsm [~-<'~p.--+ • C~<-'~+1 p
(Recall the definition of [] from Section 2.1.) If each basic operator is weakly
transitive, A is said to be w e a k l y o p e r a t o r t r a n s i t i v e . K 4 ® K 4 is oper-
ator transitive, but not transitive (as can be shown). A logic is of b o u n d e d
o p e r a t o r a l t e r n a t i v i t y if for each operator there is a d such that that op-
erator satisfies altd. A is of b o u n d e d a l t e r n a t i v i t y if it has finitely many
operators and is of bounded operator alternativity. A logic A is called c y c l i c
if for every compound modality [] there exists a compound modality [] such
that p --+ [ ] ~ p E A.
Furthermore, the postulates (}iP -+ ()jP are considered. On Kripke-
frames they force the relation <~i to be included in <~j. Also, the postulates
(}i(}jp ++ Oj(}iP say that any point reachable following first <~ and then <lj
is reachable following <~j and then <Si - - and vice versa. This is a kind of
Church-Rosser Property with respect to i and j. Sometimes, only one impli-
cation is considered. Quite interesting is the following construction. Consider
a (polymodal) logic A with operators [Z]i, i < a. Add a new operator • = [ ~ .
Then call • a m a s t e r m o d a l i t y if it satisfies the postulates of K 4 and the
interaction postulates (}iP --+ Op. If • satisfies $5 it is called a u n i v e r s a l
modality.
Important bimodal logics are tense logics. A t e n s e logic is a normal
extension of
K.t := K2 O {p --+ s~>p,p -+ E ~ p }

2.6. Decidability and Finite Model P r o p e r t y 75
Here, we have used [] for []0 and [] for [Z1. If A is a m o n o m o d a l logic, then A.t
is obtained by interpreting [] as []. There is also the possibility to interpret
[] as E. Tense logical axioms can be derived from m o n o m o d a l axioms by
choosing either interpretation for the operator. Thus, if ~ is an axiom and we
want to interpret the o p e r a t o r as [], we write ~a+, and if we want to interpret
the o p e r a t o r as [] then we write T - . So, we have logics like $4.t, Sh.t and
K 4 . t . D + . D - . T h e latter is actually the same as K 4 D . t . D - . A logic A with
2k operators is called c o n n e c t e d if there exists a p e r m u t a t i o n 7r : 2k -+ 2k
such t h a t ~2 = id and for each i < 2k, p -+ [~i(}~(i),P -+ []~(~)(}~ E A. A
connected logic is cyclic.
E x e r c i s e 49. Recall from Proposition 2.1.3 t h a t K ® T -+ ¢ = K @ ¢ d - + ~pd.

Write down the axioms you obtain for the axioms presented in this section if
you apply this operation.
E x e r c i s e 50. Show t h a t K . G _D K 4 . Proceed as follows. Suppose we

have an intransitive K r i p k e - f r a m e f and we want to show t h a t it is also not
a K r i p k e - f r a m e for K . G . T h e n there m u s t be points x ~ y ~ z such t h a t
x :~ z. Now put as a valuation ~(p) := {y, z}. T h e n x ~ (}p; ~ 0 ( p A ~(}p).
But this trick works only for Kripke-frames. Nevertheless, it gives a clue to
a solution which is completely syntactical, and therefore completely general.
Assume t h a t {(}(}p, -~(}p} is consistent in K . G . T h e n p u t ~ : = p V (}p. Now
{(}(}P,~(}P) ~-K.G (}~. Show t h a t this leads to a contradiction. T h e relation
with the K r i p k e - f r a m e is the following. A violation of transitivity can be
d o c u m e n t e d by taking V(P) : = {z}. Now we have ~(T) = { y , z } = ~(p),
the desired set d o c u m e n t i n g the failure of G. (The p r o o f t h a t transitivity is
deducible in K . G is a t t r i b u t e d to DICK DE JONGH and GIOVANNI SAMBIN
in [9].)
E x e r c i s e 51. Show as in the previous exercise t h a t K . G r z D K 4 . Hence, 4

is dispensable in the axiomatization.
E x e r c i s e 52. Show t h a t S 4 . G r z = K @ p --+ 0 ( p A [ ] ( ~ p --+ []-~p)). (See

[14].)
E x e r c i s e 53. Show t h a t K 4 ® K 4 is operator transitive b u t not weakly
transitive. Hint. Consider the frame 3 = {w, <~, • / with x <~ y iff y = x + 1
and x is even, x • y iff y = x + 1 and x is odd.
2.6. Decidability and Finite M o d e l P r o p e r t y

Recall t h a t a logic A is called decidable if one can decide for every finite
set A and a formula ~ whether or not A ~-h ~. It follows from the deduction
t h e o r e m t h a t a logic is decidable iff for every formula ~ we can decide w h e t h e r

or not ~ E A. In other words, A is decidable iff the problem ' ~ E A?' is
c o m p u t a b l e iff the set A is decidable. We shall also say t h a t a m o d a l logic A
is C - c o m p u t a b l e or in C, where e is a complexity class, if ' p E A?' is in C.
Likewise, C - h a r d n e s s and C - c o m p l e t e n e s s of h are defined. Now let A be
a finite set. A set M C A* is called r e c u r s i v e l y e n u m e r a b l e if it is either
e m p t y or there exists a c o m p u t a b l e function f : w -+ A* such t h a t the set
ranger = f[w] = M. (So, M # ~ is recursively enumerable if we can, so to
speak, make an infinite list of M.) A set is c o - r e c u r s i v e l y e n u m e r a b l e or
c o - r . e. if its complement is recursively enumerable. M is called r e c u r s i v e
if it is recursively and co-recursively enumerable.
PROPOSITION 2.6.1 (Post). Let A be finite and M C_ A*. The problem

'x E M ? ' is decidable if] M is both recursively enumerable and co-recursively
enumerable.
First, let h : w --+ A* be a c o m p u t a b l e bijection. For a p r o o f note t h a t
if 'x E M ? ' is decidable, define f as follows. If M is not empty, pick ~ E M .
T h e n put f ( n ) : = h(n) if h(n) E M, otherwise f ( n ) : = i . This function
enumerates M . So, M is recursively enumerable. Likewise we show t h a t it
is co-recursively enumerable. Now assume t h a t M is b o t h recursively and
co-recursively enumerable, ~ # M and A* # M , and let f and g e n u m e r a t e
M and A * - M . Define h as follows, h(~) : = 1 if there is an n E w such
t h a t f ( n ) = J; h(~) : = 0 if there is a n n G w such t h a t g(n) = ~. It is easy
to see t h a t this is a c o m p u t a b l e function. It follows from P r o p o s i t i o n 1.8.13
t h a t if n _< w then the set of well-formed formulae is decidable. A particular
consequence of Proposition 2.6.1 is t h a t a logic A is decidable iff A is recur-
sive. Unfortunately, this does not apply to the particular axiomatizations.
There are recursively axiomatized (even finitely axiomatized) logics which are
undecidable. However, if A is recursive then it is recursively axiomatizable.
Moreover we have the following.
PROPOSITION 2.6.2. (t~ < ~1.) Let A be a recursively enumerable set of

modal formulae. Then K~ • A is recursiveIy enumerable.
PROOF. We show how to make an infinite list of the tautologies. Classical
logic is decidable, hence the tautologies of P C are recursively enumerable.
T h e primitive instances of (bd--+.), which are of the form
[hj(p0 -+ Pl)- -+ .½jP0 --+ [2jpl

are enumerable, since ~ is. Fix an e n u m e r a t i o n g of A, an e n u m e r a t i o n h of
all classical tautologies and an enumeration g of the instances of (bd-+.). P u t
f ( 3 i ) : = g(i), f(3i + 1) : = h(i), and f(3i + 2) :-- ~(i). This gives an enumer-
ation of the axioms. We show how to enumerate the theorems of K ~ G A.
T h e problem is t h a t we have to calculate the consequences of A with respect
2.6. Decidability and Finite Model Property 77
to the rules, namely, modus ponens, the necessitation rule and the substitu-
tion rule. The reader is asked to think about the fact that it is enough to
use finitary substitutions rather than substitutions. A substitution c~ is called
f i n i t a r y if or(p) # p only for finitely m a n y p. The finitary substitutions can
be enumerated. We leave this to the reader. (Basically, it amounts to showing
that the finite sequences of natural numbers are enumerable. In effect, this
is what we will be showing here as well, though in disguise.) Thus, assume
that the substitutions are somehow enumerated. Now begin the enumeration
of the theorems simply as a list. The list is produced in cycles. The n t h cycle
consists of f ( n ) and all one-step consequences of theorems of the previous
cycles, but with substitution restricted to the first n substitutions and (mn.)
restricted to the first n boxes according to the enumeration. If the list has k
entries up to the n t h cycle, then there are k x n consequences with respect
to (ran.), at most k x k consequences with respect to (mp.) and k x n conse-
quences with respect to (sb.). So in each cycle the list is finite. Let us show
that this list contains all theorems. The proof is by induction on the length of
the derivation of ~ from A. C a s e 1. ¢ is a classical tautology or a m e m b e r of
A. Then for some i W = f(i), and so ¢ is in the ith cycle. C a s e 2. ~ = D j ¢ .
By inductive hypothesis, ~ occurs in the list, say in the kth cycle. Let Dj be
the j t h modality according to the enumeration L Then ~ occurs in the cycle
m a x { k , j } + 1. C a s e 3. T is the result of applying modus ponens to ¢ --+
and ¢. By induction hypothesis, the latter are in the list. Then ~ is the next
cycle. C a s e 4. ~ = ~rk(¢). Let ¢ be in the ruth cycle and ~ := m a x { k , m } .
Then ~ occurs in the e + 1st cycle. This concludes the proof. []
The last proof applies as well to all other logics, classical, monotone etc.
Typically, since logics are mostly given in such a way that one can deduce
that they are enumerable, it is mostly the enumerability of the nontheorems
which is problematic. Only very recently, VALENTIN GORANKO in [85], has
given some proof procedures for enumerating the nontheorems directly. (The
proof of their correctness has interestingly been given by means of semantic
arguments.) Let us now say that a logic A is r e c u r s l v e l y a x i o m a t i z a b l e
( f i n i t e l y a x i o m a t i z a b l e ) if there exists a recursively enumerable (finite) set
A such that A = K~ ® A. And let us say that A is s t r o n g l y r e c u r s i v e l y
a x i o m a t i z a b l e if a recursive set axiomatizing A can be given. It is a priori
possible that all modal logics are finitely axiomatizable; it may, namely, very
well be t h a t although a logic can be axiomatized by an infinite set of formulae,
a finite set would have been enough. We will show below t h a t this is false.
The following is a consequence of the compactness theorem.
PROPOSITION 2.6.3. Let A be finitely axiomatizable and A = K~ @ A .

Then there exists a finite A0 C_ A such that A = K~ @ A0.
PROOF. Let A = K~ ~Y A. Since A is finitely axiomatizable, there is

a finite set F such that A = K~ G F . Then there exists a proof of A F
from A and the classical tautologies. This proof is finite, so it uses only a
finite number of formulas in A. Let them be collected in A0. Then we have
A _D K,~ @ Ao _D K,~ @F = A, and so A = K,~ @ Ao. []
So if A is a set of axioms for A such that no finite subset axiomatizes

A, then A cannot be finitely axiomatized. Decidability is usually brought
into correspondence with the finite model property defined below (see next
section). However, rather than with finite model property it is primarily
connected with constructibility of models, at least if the language is countable,
which we will now assume. Recall that we have shown that any logic A is
complete with respect to some class of algebras; in particular A is the theory
of the algebra ~A(var). This being so it is nevertheless not at all clear that
we can always produce these algebras. In particular, if A is undecidable, then
even though we can enumerate all formulae, we are not able to construct
~rh(var) from the definition, since we have
~:A(var) = ~Ym(var)/:A
where ~ ~A ¢ iff ~ ~ ¢ E A. The problem is simply that we cannot even
decide whether or not a formula ~ is in the class of T. On the other hand,
suppose that A is decidable. Then choose an enumeration for of the formulae;
for simplicity for(O) := T. Furthermore, we assume to have an inverse r~aa,
yielding for each formula ~ a k E w such that for(k) : ~. We are going
to produce an enumeration 7 of the equivalence classes as follows. We start
with V(0) : : for(O) -- T. Then ~'(i + 1) :-- for(k) where k is the smallest
number such that for no k' < k, for(k') -~h for(k). In other words, we
choose a subsequence si of w such that for(si+l) is the first formula in the
enumeration that does not belong to one of the already established equivalence
classes. Then 7(i) := for(si). Since the logic is decidable, this is indeed an
(algorithmic) enumeration of the equivalence classes. Furthermore, for each
we can decide whether or not it belongs to a given class, and we can compute
the number #(~) of the class of ~. Namely, take k := r~a~, and enumerate
all Sor(k') for all k' < k. Then calculate the 7(i) up to (at most) k and see
which is the first i such that 7(i) ~ p E A. Now, the algebra ~ A ( v a r ) / --A
formed by calculating with numbers instead of formulae. For example, let m
and n be given. The conjunction is a function eonj : w × w -+ w defined by
conj(i,j) := #(7(i) A 7(J)). Likewise, all other functions of ~rA(var) can be
reproduced as functions over w. Instead, we could also use the representatives
7(i) as the underlying set of the algebra.
DEFINITION 2.6.4. (~ < ~1.) A modal algebra is called effective if its

underlying set is w, and the functions 1, - , N, and I j , j < ~, are computable.
2.6. Decidability and Finite Model P r o p e r t y 79
In general, an algebra is called e3~fective if its underlying set is w and all basic
term-functions are computable.
DEFINITION 2.6.5. Let ~ be a variety of t~-modal algebras. V is said to
have c o n s t r u c t i b l e free algebras if for any finite set of generators the con-
gruence --A, defined by ~ ~ h ¢ iff ~9 t-} ¢ C A, is a decidable subset of the set
of pairs of terms.
W h a t we have shown is that if A is decidable, its variety has constructible
free algebras, which are also effective algebras.
PROPOSITION 2.6.6. Suppose that ~1 is an effective algebra. Then T5 ~ is
co-recursively enumerable.
PROOF. E n u m e r a t e all partial valuations into P2, and enumerate the for-
mulae. Given a partial valuation and a formula with variables in the domain
of that valuation we can compute the value of the formula under the given
valuation, since the algebra is effective. It is therefore possible to enumerate
all pairs (~, a) where ~ is a formula and a a value of ~ under some valuation.
Consequently, choosing among this set only the pairs for which a ~ 1 we
obtain an enumeration of the nontheorems. []
This proof needs some explanations. If an algebra based on the set w is
not effective, there is a formula ~ and a valuation fl such t h a t ~ ( ~ ) cannot
be determined even when fl(p) is known for all relevant variables. For by
definition of effectiveness the primitive functions are not all computable. So,
let fi be a primitive function o f ~ which is not computable. Then T(p~ :----fi(p~
is a formula such that ~ ( ~ ) cannot be computed for any given ft.
COROLLARY 2.6.7. A logic A over a countable language is decidable iff A
is rccursively axiomatizable and A is the logic of an effective algebra.
(It is enough to have completeness with respect to a recursively enumer-
able class of effective algebras in addition to recursive axiomatizability. This
has been pointed out to me by MICHAEL ZAKHARYASCHEV.)Above we have
seen that it is actually enough to have the free algebra be effective, with no
assumptions on axiomatizability. Why is it then that the constructibility of
~rA(var) is enough to guarantee the decidability, while otherwise effectiveness
is apparently not enough? The reason for that is that the theory of the free
algebra is the theory of (~rA(var), v), where ~ is the natural valuation. If we
have completeness with respect to such a pair then theoremhood is easy to
decide. In fact, then the assumption that the underlying algebra is effective
is sufficient.
DEFINITION 2.6.8. A logic A has the f i n i t e m o d e l p r o p e r t y (f~mp) if
for all ~ • A there exists a finite frame ~ such that ~ ~ ~. A is tabular if
there is a finite Kripke-frame [ such that A -- Th ~.
FIGURE 2.6
n n-ln-2 1 0
C ~ ~ ... ~ =~
By T h e o r e m 2.4.11 we know t h a t A has the finite model p r o p e r t y iff for

each n o n - t h e o r e m T there exists a finite K r i p k e - f r a m e ~ for A with f ~ T.
THEOREM 2.6.9 (Harrop). (n < R0.) Suppose that A has the finite model
property. If A is finitely axiomatizable, A is decidable.
PROOF. Suppose t h a t A = K ~ ® A for a finite A. T h e n A is recursively
enumerable; we need to show t h a t it is co-recursively enumerable. Let us
first show t h a t it is possible to enumerate the frames for the logic A. To
see that, observe t h a t in order to decide for ~ whether or not ~ ~ A we
iust have to check whether or not ~ ~ A. Since A is finite, this can be
decided in finite time. Hence, since we can e n u m e r a t e all frames, we can also
enumerate the A-frames. Furthermore, we can enumerate all models (5,/3, x}
where fl assigns values only for finitely m a n y variables. For each m o d e l we can
enumerate easily all formulas which are false. Hence we have an e n u m e r a t i o n
n : w x w -+ wff which returns for a pair (i, j ) the jth formula refuted by
model n u m b e r i. Since w x w can be enumerated, say by p : w --4 w x w, we
can finally enumerate all nontheorems by n o p. []
The use of the finite axiomatizability is essential. For recursive axiomatiz-
ability this theorem is actually false, see ALASDAIR UQUHART [216] and also
[122].
Finally we give the proof t h a t there are u n c o u n t a b l y m a n y logics. We
work here in m o n o m o d a l logic, t h a t is, there is just one operator. Let us take
the following frames. ¢~ : = ( { 0 , 1 , 2 , . . . ,n}, q} with i q j iff (a) j = i - 1 or
(b) i = j = n. Consider the following formulae
~n := O~+IT A O(S%L A ~[D~-l/-)
LEMMA 2.6.10. c~ ~ ~ m ±fin # m . (c~,j) ~ ~ iff j = n.
PROOF. Consider (¢,~,j) ~ ~n. If j > n then 0 [ ] n L is not satisfied; if
j < n then 0 n + l T is not satisfied unless j is reflexive; b u t if it is, the f o r m u l a
0 ( [ ] h i A _<2,-1_5) is not satisfied. So we m u s t have j = n. Now assume
m > n. T h e n 0[]]~± is not satisfied at j = n. If, however, m < n t h e n
0(@'~1 A ~ [ ] ~ - l y ) is false at j = m. T h u s m = n = j, as required. []
This innocent example has a n u m b e r of consequences. Take any subset

M C_ w and let t ( M ) = K ® { ~ : n C M } . T h e n L: 2 ~ -+ ~ ( K ) is injective.
For if M ~ N then there is a m 6 M but m ¢ N (or the other way around).
2.6. Decidability and Finite Model Property 81
Then all axioms of ~(N) are satisfied on Cm, but not all axioms of L(M). We
conclude that Frm(t(M)) # Prm(t(N)). Thus the two logics are different.
THEOREM 2.6.11. There are 2 ~° ~-modal logics, for all ~ > O. Moreover,
there exist 2 ~° many non-recursively axiomatizable logics and there exist re-
cursively axiomatizable, undecidable logics.
PROOF. We have seen that the m a p c is injective, so there are at least 2 ~°

logics. However, our language has countably m a n y formulae, and a logic is a
set of formulae, so there are at most 2 ~°. Since there can be only countably
m a n y algorithms, there are 2 ~° non-recursively enumerable subsets of w and
so there are 2 ~° m a n y non-recursively axiomatizable logics. Finally, take a
recursively enumerable, but non-recursive set M (such sets exist). The logic
L(M) is recursively enumerable, by definition. But it cannot be decidable,
since that would mean that we can decide ' ~ j C ~(M)', or, equivalently,
'j E M ' . []
There exist also finitely axiomatizable undecidable logics. The first was
established by STEPHEN ISARD [107], basically through coding the action of a
machine in modal logic. Subsequently, m a n y alternative ideas have been used,
for example undecidable problems of group theory by VALENTIN SHEHTMAN
([198]), the tiling problem in EDITH SPAAN [202] and T h u e - p r o b l e m s (see
among others MARCUS KRACHT [127]). We will return to this subject in
Section 9.4.
E x e r c i s e 54. Show the following variant of Proposition 2.6.3. Let 0 be a

logic and A be finitely axiomatizable over 0 . Suppose 0 = A @ A . Then there
exists a finite set Ao C A such that A = 0 @ Ao.
E x e r c i s e 55. Give an example of a logic ® which is finitely axiomatizable

as a normal extension of K1 but not as a quasi-normal extension.
E x e r c i s e 56. Let :K be a recursively enumerable set of effective ~ - m o d a l

algebras. Show that Th X is co-recursively enumerable.
E x e r c i s e 57. Show that Theorem 2.6.9 holds also for infinite ~. Hint. A
finite axiom system uses only finitely m a n y basic operators.
E x e r c i s e 58. Let T be a theory of classical predicate logic, in any given

signature. Show that T is recursively axiomatizable iffit is strongly recursively
axiomatizable.
2.7. N o r m a l F o r m s
This chapter introduces a very basic method for proving t h a t the logic
K~ is decidable, using the fact that it has the finite model property. This
proof was first given by KIT FINE [64]. The finite model property is a m o n g
the best-studied properties of logics. We will show a fair number of strong
results on the finite model property later but shall be content in this section
to show only a single result, namely the finite model property of the base
logic K~. There are m a n y proofs of this fact but only very few proofs are
constructive and do not presuppose heavy theory. For example, the proof by
filtration - - which we will present later - - presupposes t h a t we can show the
existence of at least one model, from which we then obtain a finite model.
The basic method here is syntactic in nature. We will start by proving t h a t
formulae can be rewritten into a somewhat more user-friendly form.
DEFINITION 2.7.1. A formula ~ is called strictly s i m p l e o f degree 0

if it is of modal degree 0 and of the form T or Aj<n qJ, n > 0, where each qj
is either a variable or a negated variable; moreover, no conjunct may occur
twice. 9o is called s i m p l e o f degree 0 if it is J_ or a nonempty disjunction
of pairwise distinct strictly simple formulae of degree O. ~ is called s t r i c t l y
s i m p l e o f degree d ÷ l if it is of modal degree d 4- 1 and of the form
A A A A
j<P j<q
where # is strictly simple of degree O, no conjunct of ~ occurs twice, s : p --+ ~,

t : q --+ ~ are functions, Xj are simple of degree < d and all u;j are strictly
simple of degree ~ d. ~ is s i m p l e o f degree d4-1 if it is modal degree
d 4- 1 and a disjunction of pairwise distinct strictly simple formulae of degree
~ d4- 1. Moreover, a modal formula of degree 0 is called s t a n d a r d o f degree
0 if it is simple of degree O; a modal formula of degree d4-1 is called s t a n d a r d
o f degree d4-1 if it is a disjunction of strictly simple formulae in which the
functions s are injective and in which for a subformuIa of the form Oj@ ¢ is
standard of degree dp(¢ ).
It is important to get used to simple formulae, so the reader is asked to

prove some easy properties of them.
LEMMA 2.7.2. (i) Any subformula of a simple formula is simple. (ii) If

a formula is simple of degree d + 1 it is composed from variables (constants)
and their negations, and formulae []j¢, where ¢ is simple of degree d, using
only A, V and Oj.
PROPOSITION 2.7.3. Every formula ~ can effectively be tranformed into
a simple formula deductively equivalent to ~ in K~.
2.7. Normal Forms 83
PROOF. We prove in a rather detailed way how to obtain a form t h a t

contains negation only directly in front of variables. The m e t h o d to convert
into full simple form is similar. Take a formula ~o and apply the following
reductions from left to right as often as possible.
It is clear that some reduction will apply as long as some operator is in

the scope of -~. Moreover, each step is an equivalence. (This follows from
Proposition 2.1.1.) Thus all we have to show is that there is a terminating
reduction series. To see this, we have to monitor two parameters, namely e,
the length of a longest subformula of the form ~ ¢ , and k, the number of the
subformulae of length t~ of the form -1¢. As long as there is a subformula
7 ¢ of length g, we apply the reduction algorithm to t h a t subformula. The
result is always a formula in which k is decreased by 1. If k = 1, then in the
next step g decreases. Proceeding this way, we will eventually reach g < 2,
which means that the subformutae are of the form ~Pi, Pi a variable, or ~ci,
ci a constant. This shows how to throw negation in front of variables and
constants. To obtain simple form use the following reductions
~oA(¢Vx) ~-~ (qoA¢)V(~PAx)
05(¢vx) "~ 05¢v0jx
~OAtO -~ to
~pV~o "~
It is proved analogously that these reduction terminate and that after their
termination the formula is in the desired form. []
PROPOSITION 2.7.4. Every formula ~o can be effectively transformed into

a deductively equivalent standard formula.
PROOF. First, transform ~o into simple form. Let X be a strictly simple
subformula of minimal degree that is not standard. Then it contains two con-
juncts of the form []jcrl, []jcr2. ~oi is defined by eliminating t h a t occurrence of
[~jcr2 and replacing the occurrence of []jO" 1 by K]j(Crl A or2). ~Ol is deductively
equivalent to p. Iterate this as often as possible. Now repeat this construc-
tion for other nonstandard subformulae of minimal degree. Each time the
construction is performed it either reduces the number of nonstandard sub-
formulae of least degree, or it increases the minimal degree of a nonstandard
subformula. The procedure terminates and yields a standard formula. []
The next definition is crucial for the definition of the finite model. In
order to understand it, we explain first the notion of in conjunction with.
First, say t h a t an occurrence of ¢ in X is a c o n j u n c t of X if this occurrence

of ¢ in X is only in the scope of A. An occurrence of ¢1 is in c o n j u n c t i o n
w i t h an occurrence of ¢2 in the formula ~ if ¢1 and ¢2 are conjuncts of some
subformula containing both occurrences of ¢1 and ¢2.
DEFINITION 2.7.5. Let ~ be a standard formula. ~ is called e x p l i c i t if

for every strictly simple subformula # A Ai<p [q~(i)xi A Aj<q (}t(j)a~j and every
j < q there exists an i such that t(j) = s(i) and a disjunct ~ of Xi, such that
every conjunct of a is a conjunct of czj.
THEOREM 2.7.6. For each ~ there exists a standard and explicit ¢ such
that ~ ~-~ ¢ E K~.
PROOF. First, turn ~ into standard form. Call a subformula Iqj(V~ ¢i)
u n l e a s h e d if for every OjX it occurs in conjunction with, one ¢i is such t h a t
all conjuncts of ¢i are conjuncts of X- Let 5 be the largest number such t h a t
there is a subformula []j (Vi ¢i) of degree 5 which is not unleashed. Now take
the subformulae which are of degree 5 and not unleashed. Let [Bj(Vi ¢i) be
one of them. Suppose it occurs in conjunction with a subformula Ojr. T h e n
add Vi ¢i as a conjunct to r; perform this for formulae of degree 5 which are
not unleashed. Now distribute A over V, and then 0j over V. We will then end
up with formulae of the form 0 j ( ¢ i A r ) in place of 0 j r . ¢i is not of the form
X1 V X2, and standard. Thus the resulting subformulae are simple. Finally, to
convert the formula into standard form we only have to drive V outside, and
so all subformulae [ ] J ( V i ¢ i ) of degree 5 are now unleashed. Thus, we m a y
proceed to smaller subformulae. Since we never change the modal degree of
the modal formulae involved, this procedure ends. []
We work through a particular example to give the reader a feeling for

these definitions. Take the language ~ = 2, the operators being [] and [].
Then let ~ be the formula
p. A A A v
We can push negation inside in the second conjunct, distribute ~ over V in

the third, and kill double negation.
p. A v A v ( ep)
Next, we can distribute V and get
[p. A •=p v A
V [p. A .•(tE[]-~pV El~p). A .~£>p]
Now the formula is in standard form. However, it is not explicit. Namely, in
b o t h disjuncts, the second conjunct is of the form []¢ while it is in conjunction
2.7. N o r m a l Forms 85
with a formula of the form ~X. Thus, we must add ¢ as a conjunct to X.

Subsequently, we can distribute ~ and V.
[p. A .r~([][]~p v []~p) A ~ ( ~ [ ] ~ p A ~p)]
v [p. A . ~ ( ~ p v []~p) A <~([]a~p A ep)]

V [p.A.I3(tEB~pVE]~p) A , O ( t E ~ p A ~ p ) ]
We have unleashed formulae of degree 3, but there have a p p e a r e d new for-
mulae of lower d e p t h with must also be unleashed. After distribution etc. the
formula is in s t a n d a r d and explicit form.
v [p. A . ~ ( ~ p v a,~p) A ~(~epA ~p)]

V [p. A .E~(E]Iht-~pV tEOp) A O(E][5~p A 4)(pA I]~p))]
v [p. A .r~(aB~pv ~ p ) A ~(~:~p A e(pA ~p))]
Call ~ c l a s h - f r e e if there do not exist occurrences of subformulae of the form
p/ and ~Pi, for some i, in conjunction with each other.
LEMMA 2.7.7. Let ~ be standard and explicit and not of the form X1 VX2.
Suppose that it contains an occurrence of a formula of the form Pi A -'Pi A w
which is not in the scope of a box. Then ~ is inconsistent in K~.
PROOF. By assumption, p/ and ~p/ are not in the scope of V and [qj, for
any j < ~. Clearly, from the assumptions, ~ is c o m p o s e d from p~ A -~p~ A w
and other formulae using only A and 0j. T h e n ~ is inconsistent. []
There is an algorithm which converts a formula into a clash-free for-

mula. Moreover, there is an algorithm which in addition preserves simplicity,
explicitness and being standard. Namely, suppose t h a t Pi and -~Pi are in con-
junction with each other. T h e n remove from ~ all subformula occurrences in
conjunction with these occurrences, and replace Pi and ~Pi together by ± . If
necessary, remove L in disjunction with some formula. This converts p into
clash-free and s t a n d a r d form. It is somewhat c u m b e r s o m e b u t not difficult
to verify t h a t the resulting formula is also explicit if p has been explicit.
Given a standard, explicit and clash-free formula ~ we build a set of
models as follows. Let us assume ~ = Vi<~ p~, ~ strictly simple. T h e n for
each ~ / w e build a separate model; the collection of the models is the m o d e l -
set of p. We will see t h a t ~ is consistent iff the model set is n o n - e m p t y iff
n > 0. T h u s assume n = 1, t h a t is, p is now strictly simple. Take a node
~ , as the root, and for each subformula OjX not in the scope of a box take
a point x x. Then, as it is directly verified, X is strictly simple, standard,
explicit and clash-free. For two subformulae 0pC and (}qX p u t x¢ <~j x x iff
OjX is a conjunct of ¢. (In t h a t case, j -- q.) A valuation is defined as
follows. Let x x be a point constructed for the formula X. If p/ is a conjunct
of X then x x ~/3(p~), and if -~p/ is a conjunct of X then x x ¢/3(p~). In case
where neither Pi nor -~pi is a c o n j u n c t / 3 ( p i ) can be fixed arbitrarily. Since X

is clash-free, a model can always be defined. A model for ~ thus o b t a i n e d is
called a d i r e c t m o d e l .
LEMMA 2.7.8. Let ~o be strictly simple, standard, explicit, and clash-free.

Let (~,~, x~} be a direct model of ~. Let OjX not occur in the scope of a box.
Then if ¢ is a conjunct of X, Xx e ~ ( ¢ ) .
PROOF. By induction on the constitution of ¢ (and X)- For ¢ a variable,

this is true by construction. Also, the definition is sound, by assumption.
Now assume ¢ = wl A w2. T h e n if ¢ is a conjunct of X so are wl and w2. B y
induction hypothesis, x x e ~(a~l) and x x e ~(a~2). T h e n x x • ~(wl A w2), as
required. T h e case g) = ¢1 V ¢2 does not arise. Next assume ¢ = Ojw. T h e n
there exists a j-successor x~ of x x. By induction hypothesis x~ • ~(w), a n d
so x x • fl(¢). Finally, let ¢ = • j w . Let y be a j - s u c c e s s o r of x x. T h e n
y = x~ for some 7- such t h a t 0j~- is a c o n j u n c t of X. By explicitness, some
disjunct wi of w (or X itself) is a conjunct of ~-. By hypothesis, x,- • ~(co/).
Hence x x • ~ ( ¢ ) . []
THEOREM 2.7.9. K,~ has the finite model property.
PROOF. Start with ~o and convert it into s t a n d a r d and explicit form. Let
be a disjunction of ~oi. If ~oi contains a clash we have ~oi ~- _1_by L e m m a 2.7.7.
If ~i does not contain a clash then by L e m m a 2.7.8 there exists a finite m o d e l
for ~oi. Hence, either all ~i contain a clash, in which case ~o ~- _t_, or there is a
model for some ~i and hence for ~. []
R e l a t e d to s t a n d a r d formulae are the normal f o r m s which are also in use

in boolean logic. T h e difference with s t a n d a r d formulae is t h a t n o r m a l forms
give rise to u n a m b i g u o u s direct models on a given set of variables. One can
define n o r m a l forms with respect to s t a n d a r d formulae by the following fact.
In addition to being s t a n d a r d (i) a n o r m a l form is consistent, (ii) a n o r m a l
form is always reduced; it contains no occurrences of qo in conjunction with a
different occurrence of ~o and no occurrence of ~o in disjunction with a n o t h e r
occurrence of ~o, and (iii) a normal form is complete for given m o d a l d e p t h
5; if ~ is in normal form, X of d e p t h less t h a n 5 and y) A X is consistent,
then ~ ~- X. (i) is easy to achieve. We can start from a simple formula and
just t h r o w away all disjuncts containing a clash. (ii) is likewise easy to get.
Simply drop multiple ocurrences of the same formula. (iii) is possible only
on one condition, namely t h a t we work over a finite vocabulary. By finite
vocabulary we m e a n b o t h t h a t there are finitely m a n y m o d a l o p e r a t o r s and
t h a t there only finitely m a n y propositional variables and constants. T h u s , let
us assume t h a t we have a finite set J C_ ~ and a finite set P = {Pi : i < n}.
T h e n the n o r m a l f o r m s o f d e g r e e k over J and P are defined inductively
2.7. N o r m a l F o r m s 87
as follows.
xo := AiEc Pi A Ai~c -~pi C C_ n
nf(J,P,O) := {X°c:CCn}
xk+l
D(D := AicD(j) (}jX~ A Ai~D(j) -~(}jX~ D ( j ) C_ nf(J, P, k)
xk+i
:= X~ A ~\jog
^ k+l
XD(j) C C n,D(j) C nf(J,P,k)
C,~ -- --
k+l
n f ( J , P , k + l) := {Xc,-~ : C C n,-D : J--~ p ( n f ( J , P , k ) ) }
PROPOSITION 2.7.10. Let ~ be a modal formula of depth k based on the
variables of P and the operators of J. Then there is a set • C nf ( J, P, k)
such that
PROOF. By induction on k. For k = 0 this is the familiar disjunctive

n o r m a l form for boolean logic. Now let k > 0. T h e n p is a b o o l e a n combina-
tion of variables and formulae (}j¢ with dp(¢) < k. B y inductive hypothesis
each ¢ is equivalent to a disjunction of a set ArC of n o r m a l forms of degree
k - 1. If N¢ -- O, then (}j¢ is equivalent to (}j J_, hence to L. If ArC ¢
then (}j¢ ~ (}j V N¢ ~_ V((}j)~ : • E N¢). After this rewriting, bring ~ into
disjunctive n o r m a l form. Each conjunct of ~ is now of the form
# = ~A A (}jXck-1 A A _~(}jXck-1
CcG CCH
where u is n o n m o d a l and G N H = O, G and H sequences of subsets of

nf(J, P, k - 1). This is not necessarily in n o r m a l form, since we m a y have
G U H C nf(J, P, k - 1). But if there is a (}iX which has not yet been included
in G U H we e x p a n d # by the disjunction (}iX V -~(}jX, and we get
# ~_ # A (}iX. v .# A ~ ( } j X
In this way we can e x p a n d # so as to include all (}iX, X E nf(J, P, k - 1). T h e
same with u. R e p e a t this procedure as often as necessary. Finally, we reach
normal form. []
LEMMA 2.7.11. Any two distinct normal forms of nf(J, P, k) are jointly
inconsistent.
PROOF. By induction on k. If k = 0, then let C, C ' _C n two distinct
subsets. W i t h o u t loss of generality we m a y assume C - C I ~ O. T h e n
¥ O A x ~ , --= ± . For there is a n i E C s u c h t h a t i •C'. T h e n x ° ~-pi but
)~o, k- ~pi. Now let k > O. If X and X ~ are distinct forms, then either they have
distinct n o n m o d a l components, or there is a j C J such t h a t D(j) ~ D'(j).
[n the first case we already have seen t h a t there arises a contradiction. In
Lhe second case, assume without loss of generality, w C D ( j ) - D'(j). T h e n
~- (}jaJ but X ~ ~- -~(}jw, a contradiction. []
PROPOSITION 2.7.12. The formulae of degree k with variables from the

set P = {Pi : i < n} and modal operators from J form a boolean algebra. The
number of atoms is bounded by the number b(j, n, k) defined by
b(j,n,O) := 2n
b ( j , n , k + 1) := b(j,n,O). 2jb(j''~'k)
PROOF. We have seen that each formula is equivalent to a disjunction of
normal forms, so there are at most 2 b(j'~'k+l) formulae, where b(j,n, k + 1)
is the cardinality of nf(J, P, k + 1). Moreover, since the normal forms are
mutually inconsistent, they form the atoms of this algebra. The n u m b e r of
atoms is obtained by multiplying the choices for a normal form of degree 0
with the choices of j - l o n g sequences of sets of normal forms of degree k. T h e
latter is nothing but 2 b(j'n'k)3 = 2 j'b(j'n'k). []
Thus, each set of normal forms individually presents a representative for a
class of equivalent propositions. With a normal form we can associate a direct
model as before. However, this time there are no clashes, and the valuation/3
is uniquely defined. For if x~ is given, then ~ is equivalent to F Api V ~ A-'Pi,
so t h a t it must be equivalent to either of them, showing t h a t Pi or ~Pi must
be a conjunct of ~.
Notice t h a t there are nontrivial propositions with no variables as we have
seen earlier. Their number is bounded by the modal degree and the set of
occurring operators. In m a n y logics, however, there are up to equivalence
only two distinct constant propositions, A_ and T. We say t h a t such logics
have t r i v i a l c o n s t a n t s .
THEOREM 2.7.13. A modal logic A has trivial constants iff for every op-
erator V]j either D j ± C A or (}iT E A.
The proof is simple and is omitted. Notice that the postulate • j _ k says
that no world has a j-successor, while 0 i T says that every world has a j -
successor.
Normal forms are closely connected with a technique called unravelling
introduced in Section 3.3. The method of unravelling can be used to show t h a t
K~ is complete with respect to completely intransitive trees, by first showing
that it is complete and then using unravelling to get a totally intransitive tree
from a model. This is somewhat better than the proof via normal forms, which
established completeness with respect to acyclic frames only. Furthermore,
we can show now the following rather important fact. (Recall the definition
of [] from Section 2.1.)
THEOREM 2.7.14. Let t~ be finite and 92 an n-generated x-modal algebra.
If92 ~ []k±, k > O, then 92 is finite with at most 2 b(~'n'k-1) elements.
PROOF. Let a o , . . . , an-1 be the generators of 92. An arbitrary element
of 92 is of the form ~ [ a 0 , . . . , a n - i ] , where ~ is a formula in the variables
2.7. Normal Forms 89
P0,.-. ,P,~-I. We will show that for any formula T there exists a formula [~]k
of degree ~ k such that [ k ± A ~ is deductively equivalent to [k_l_ A [~]k in
K~. Since the number of such formulae is at most 2 b(~'n'k-1), we are done.
So, let T be of depth > k. We assume that negation is in front of variables,
double negations killed. Let [~]k denote the formula obtained as follows.
[p]0 := ± bp]0 := ±
[p]k+1 := p :=
:= A[¢]k := V[¢]k
[0W]k+ := 0j[ ]k [DW]k+ :=
Roughly speaking, [P]k is the result of replacing all occurrences of subformulas

embedded exactly k times by modal operators by ±. Then ~ is the result
of replacing some occurrences of ± by some formulae X we have [~]k ~- ~.
(Notice, namely, t h a t the occurrences are not embedded in any negation.) It
remains to be shown that [ k ± ; p ~_ [T]k. This is done by induction on k and T.
The case k = 0 is straightforward. Now let k > 0. Let ~ = CAw. By inductive
hypothesis, [~]k = [¢]k A [w]k. By inductive hypothesis [ k ± ; ¢ ~_ [elk and
[]k_l_; w ~- [w]k. Hence [ k ± ; CAw F- [¢]kA[w]k. Analogously the case ~ = e V w
is treated. Now assume p = [ ] j ¢ and k > 0. Then [ k - l j_; ¢ ~_ [¢]k-1. From
this we obtain []k_k; [:]j¢ ~- £]j[¢]k-1. Since [Dj¢]k = Dj[¢]k-1 we obtain the
claim. Analogously for ~ = Oj¢. []
E x e r c i s e 59. Prove L e m m a 2.7.2.
E x e r c i s e 60. Let ([,/3, x) be a model. Show that there exists exactly one
normal form X of degree n for given J and P such that ([, fl, x) ~ X. This
form will be denoted by X~ (x).
E x e r c i s e 61. With notation as in the previous exercise, let w ~,~ x if

Xn(w) = Xn(X). Let w ~ x i f w ~ n x for all n. Show that the m a p x ~-~ x~
is a p - m o r p h i s m , and compatible with the valuation.
E x e r c i s e 62. Call ([,/3) condensed if ~ is the identity. Let ([,~,w0) be

condensed and [ generated by w0. Show that there exists a e-localic m a p
from X~(Wo) into [. (See Section 3.3 for a definition.)
E x e r c i s e 63. Let X and X+ be normal forms. Call X+ an elaboration if it is of

depth at least t h a t of X, and if X + ~- X. Characterize this notion syntactically.
E x e r c i s e 64. Show that in K . a l t l every formula is equivalent to a formula

of the following altl-form. For degree 0, a formula is in a l t l - f o r m iff it is in
disjunctive normal form. A formula of degree d + 1 is in a l t l - n o r m a l form iff
it is of the form either # A [~± and d = 1 or # A 0 ~ where # is a a l t l - f o r m

degree 0 and qo an a l t l - f o r m of degree d.
2.8. T h e L i n d e n b a u m - T a r s k i Construction
An important question from the model theoretic point of view is the
problem whether for a given logic h there is a frame ~ such that A = Th(~).
Notice that in Section 2.2 we have shown that there exist Mgebras of this sort.
We will now show that there are also frames with this property. The solution
is due to LINDENBAUM and ALFRED TARSKI. The basic idea is that a world is
a maximally consistent set of formulae. Given a logic A this is a set W C wff
such that W is consistent but no proper superset is. The intuition behind this
terminology is that a world is an existing thing and everything we say about
it should therefore be either true or false. Clearly, one is more inclined to say
that there can be only one world, the world we are living in, so speaking of
collections of worlds can then only be metaphorical. There are ways to get
around this apparent problem. For now we are interested in the connection
between our logic and the worlds that might exist. The basic result of this
section is that if we define a certain natural frame from A over a given set
of propositional variables then the theory of that frame will be exactly A.
This shows that every modal logic is the theory of a single frame and so frame
semantics is as good as algebraic semantics. In Chapter 4 we will see that this
is no accident. The construction is a specialization of a general technique to
form geometric models from algebraic models. We proceed as follows. First,
we show that there are enough maximally consistent sets (or worlds). This
proof is completely analogous to that of Corollary 1.7.13. Second, we establish
the frame based on these worlds. And thirdly we show that the logic of this
frame is A.
Let us begin with the question of the existence of worlds. With respect
to a logic A a w o r l d is a maximally A-consistent set of formulas. The next
lemma asserts that for any consistent collection of facts there is a world in
which it is realized.
LEMMA 2.8.1. Every A-consistent set is contained in a maximally A -
consistent set.
The proof is immediate from ~ k e y ' s Lemma. A maximally consistent
set is also deductively closed, as can easily be shown. We note the following
properties, of which we will make tacit use later on. These are easy to prove
(cf. Section 1.7).
LEMMA 2.8.2. Let W be a deductively closed set of formulae.
1. W is consistent iff for no ~: ~ E W and ~ E W.
2. p A x E W iff~aEW andxCW.
3. I f ~ E W orxEW then~VxEW.
2.8. The Lindenbaum-Tarski Construction 91
4. W is maximally consistent iff it is consistent and ~ V X E W implies

E W orxC W, forallT andx.
5. W is maximally consistent iff for all ~: ~ E W exactly if ~ ~ W .
We now need to specify the relations in which these worlds stand to each
other. For t h a t we have a plausible definition. Given a modal operator [~j
(and its dual 0 j ) and two worlds W and X we want to say that X is j-possible
at W if the total collection of facts in X is j-possible at W, t h a t is, we would
like to say that
W~jX ~ (}JAXEW
However, X is infinite, so we cannot define things in this way. We have no
infinite conjunction, only a finitary one. So our best approximation is
W <~j X ¢=~ for all finite subsets Xo of X: (}j A Xo E W
This is actually how we define the accessibility relation. However, this defini-
tion can be phrased more elegantly. Notice, namely, t h a t if Xo is a finite set
then A Xo E X . For from A X0 ¢ x follows V ( ~ : ~ E Xo) E X , so for one
E X0 also ~ E X, which cannot be.
To introduce the definition in its final form let us agree on the abbreviation
OjS := { 0 j ~ : ~ E S}. Then we define
(ace.) W <j X ¢:~ (}iX C W
There is an alternative characterization as follows.
LEMMA 2.8.3. Let W and X be worlds. Then the following are equivalent.
1. For all ~ E X : Qj~ E W .
2. For all []j~ E W : ~ E X .
PROOF. Suppose that the first holds and assume ~, ¢ X . Then - ~ E X
and so 0 j - ~ E W. Since [-K 0 j ~ • ~ .--7[~]j~ we also have ~[Zjqp E W.
Thus [Zi~ ~ W, by consistency of W. So, the second holds. Now suppose
that the second holds and assume 0j~a ¢ W. Then we have ~0jqa E W, thus
[ ~ j ~ E W, which by assumption implies - ~ E X. Hence ~ !~ X . So, the
first holds. []
The construction also yields enough worlds in the following sense. If (~jT E W
then there is an X such that W <~j X and T E X. For consider the set
S := { T } U { x : 7~jX E W } . If it is consistent, there is a world X _D S
and we must have W <~j X by construction. So we have to show that S is
consistent. Suppose it is not. Then there is a finite set So _C S which is
inconsistent. Without loss of generality So = {T} U To. Now To F-A ~ , and
so []iT0 ~-h []j-~T. Since [3jT0 C_ W we must have D j - ~ E W, which is to say
-~(}j~ E W, and so by consistency of W, (~jT !~ W. This is contrary to our
assumptions, however. Thus, S is consistent and successor worlds containing

exists.
PROPOSITION 2.8.4. Whenever W is a world and (}j~ E W there exists

a world X such that W <~j X and ~ C X .
Finally, the internal sets must be specified. This is most straightforward.

Internal sets are those which are specifiable by a proposition. Define ~ by
We call ~ : p ~ ~ the n a t u r a l v a l u a t i o n .
LEMMA 2.8.5. Let ~ := ( X : ~ C X ) a n d , : p ~ ~. Then for all ~,

=
PROOF. By induction on T. Suppose T = ~X. Then ~X = {W : -'X E W}

and by induction hypothesis P(X) = X -- ( W : X C W}. Since the worlds are
maximally consistent, X C W is the same as ~X ¢ W, so ~X = - X = - ~ ( X ) ,
from which the claim follows. Now assume T = )il A X2. We have X1 AI'~2 =
~'~M~'~, as is easily checked. Therefore F(X1 AX2) = Y(X1)M~.~(X2) = ~'~ M~'~ --
X1 A''~2. Finally, assume that ~ = (}iX. We have W E (}iX iff (}iX E W iff
there is a j-successor X such that X C X (by Proposition 2.8.4) iff there is
a j-successor X such that X 6 ~ iff W E e j ~ . Hence P(~) = P((}jX) =
Cj~(X) = CjX = 0iX = ~, as required. []
Under the supposition that the valuation is the natural valuation the
internal sets are exactly those which are values of the formulae of our modal
language.
DEFINITION 2.8.6. Let A be a normal logic. Denote the set of worlds

by WA and let WA = ~ : qo E wff}. Define "~i by X <j Y iff for all
[3j~ E X , ~ C Y (= (acc.)). Then the canonical f r a m e for A is the
frame ~anh(var) = (WA, (<~ : i < ~),WA). The underlying Kripke-frame
is denoted by CCmh(var). The global canonical m o d e l for A is the pair
(Canh(var), , ) where ,(p) = ~ = { W : p 6 W } . A local canonical m o d e l is
a triple (CanA(var), ,, X ) , where X is a world.
W h a t we know now is that if a formula ~ is not in A, then there exists a
W such that ~ E W, since - ~ is consistent with A. Then (¢anA, u, W) ~ -~T.
But can we be sure that A is the logic of the frame? Is it possible that there
are countermodels for axioms of A? We will show that this is not the case.
The reason is the choice of the internal sets. Notice that the internal sets can
be ' n a m e d ' in the canonical model by a formula. Namely, for every a E WA
there exists a ~ such that a = ~ = ~(~). So, let/3 be an arbitrary valuation.
Then for each p there exists a formula Tp such that/3(p) = P(Tp). Then for
2.8. The Lindenbaum-Tarski Construction 93
an arbitrary formula ~,
~ ( ¢ ) = ~ ( ¢ [ ~ p / p : p e var(¢)])
Thus, if a model exists for ¢ based on the valuation t3, then for some sub-
stitution a, a model for ¢~ exists based on the valuation ~. Another way of
seeing this is using the Theorem 2.8.8 below. Suppose namely t h a t an axiom
of A is violated. Then, by our arguments, a substitution instance ~ (which
is also an axiom) is violated on the canonical model. This means, however,
that there is a world W such that --p~ E W. Now, since - ~ is inconsistent
with A this simply cannot be. Hence we have shown that no axiom can be
refuted on any model based on the canonical frame.
THEOREM 2.8.7 (Canonical Frames). Let A be a normal polymodal logic.
Then A - Th ~anh(var).
In writing ~ani(var) we have indicated that the canonical frame also
depends on the set of variables that we have available. In fact, the structure
of the canonical frame is up to isomorphism determined only by the cardinality
of the set of variables. (We will see that even the logic really depends on the
latter.) We have assumed that this set has countably m a n y elements. Thus
we also write Eani(~0).
A different proof of Theorem 2.8.7 consists in showing t h a t the algebra
of internal sets is isomorphic to ~rA(var). This follows from the fact that the
algebra of sets is the direct image of the m a p p ~-~ 7. We have verified above
that this m a p is a homomorphism and that ~ = ~ iff ~-h ~ /-~ ¢"
THEOREM 2.8.8. The algebra of internal sets of Earth(vat) is isomorphic

to ~rh(var). An isomorphism is given by the map ~ ~-+ 7.
Analogous techniques can be used for quasi-normal logics. Suppose that
is a quasi-normal logic extending A. Then any q2-consistent set is con-
tained in a maximally ~-consistent set. For this set there is a model of the
form (~anA(var), ~, W}. Again, since a quasi-normal logic is closed under
substitution, we get that (ganh(var), W) ~ ~9.
THEOREM 2.8.9 (Canonical Frames for Quasi-Normal Logics). Any qua-

si-normal logic k~ ~_ A can be obtained by
kv = D Th (EanK(var),W}
wES
for a set S C WK.
In a normal logic we have S = WA by closure under (ran.).
As we have noted earlier in connection with free algebras, the structure of
Earth(vat) only depends on the cardinality of the set vat. Let it be a. Then we
also write ganA (a) and call it the a - c a n o n i c a l f r a m e for A. As it will turn
out, the structure of a canonical frame depends nontrivially on the cardinality

of the set of variables. The question is whether the canonical frames for A for
different cardinalities of the variables are related by certain p-morphisms. We
will show that a function between two cardinal numbers induces a p-morphism
of the associated canonical frames. So, take two cardinals a and /3 and a
function f : a -+/3. f induces a homomorphism hy : ~:m(a) -+ ~m(/3) defined
by hf(pi) := PI(i). hf is uniquely determined by f . This has been proved in
Theorem 1.3.6. Moreover, the Theorem 2.8.11 below mirrors Theorem 1.3.6,
showing that the maps between the algebras have a correlate for the canonical
frames.
LEMMA 2.8.10. Let X be a world in the language over (Pl : i </3} and
f : ~ -+ j3. Then h]l[X] is a world in the language with variables {p~ : i < c~}.
PROOF. (1.) h)-l[x] is deductively closed. For let ~ ; p -+ ¢ E h]l[X].
Then h f ( ~ ) C X and h f ( ~ ) --+ h i ( C ) G X. Since X is deductively closed,
hf(~b) e Z . So, ¢ e h]l[Z]. (2.) h)-l[x] is consistent. For if ~ ; - ~ C h]l[X]
then h i ( p ) E X and -~hf(~) E X. But X is consistent. So, either h f ( F ) ¢ Z
or -~hf(~) ¢ Z . Hence either ~ ¢ h } l ( X ) or ~ ~ h~l[Z]. (3.) h}l[X] is
maximally consistent. For let ~ V ¢ e h]l[Z]. Then h f ( ~ ) V h f ( ¢ ) e Z . By
maximality of X, h f ( p ) E Z or h f ( ¢ ) E Z . Thus we have ~ E h]l[Z] or
¢ h71[X]. []
THEOREM 2.8.11. Let a and j3 be cardinal numbers and f : a -+ /3 a
map. Let X I denote the map X ~+ h]l[X]. Then Xy : ~anA(t3) --+ ~anA(a).
Moreover, if f is injeetive, X I is surjeetive and if f is surjeetive then X I is
injective.
PROOF. By the previous lemma, X ] is a map between the sets of worlds.
Now assume that Y and Z are worlds over {Pi : i < 13} and Y <lj Z. We claim
that X I ( Y ) <j X I ( Z ). For let []j~ E X t ( Y ) = h]l[Y]. Then [~jhi(~) G Y.
Hence hi(Q) G Z, since Y <~j Z. This shows ~ e h]l[Z] = X I ( Z ) . The first
condition for p-morphisms is proved. Now assume that Y is a world over
{Pi : i < Z} and U a world over {Pi : i < a}; and let X I ( Y ) <~j U. Then for
every ~ j ~ such that ~ j h i ( ~ ) E Y we have ~ C U. Let Y~ := {~ : Z]j~ G Y}.
The set hI[U]UY~ is consistent. For take finite sets A0 C_ hi[U ] and A1 C Y~.
We have h}~[A0] C_ V and h f l [ ~ j A ~ ] C_ h]~[Y]. By assumption on Y and U,
h]l[A1] C U. Therefore, the set h)-~[A0; A~] is A-consistent. Then A0; A~ is
A-consistent as well. So, every finite subset of h I [U] UY~ is A-consistent. This
set is therefore A-consistent and has a maximally consistent extension. Call
it V. Then for every [:]j~ ~ Y we have ~ G V, and so Y <~j V. Furthermore,
h ~ [ Y ] D h~[hi[V]] ~_ V. Since h]~[Y] is consistent and e maximally
consistent, h}~[V] = U, and that had to be shown. This proves the second
2.8. T h e L i n d e n b a u m - T a r s k i Construction 95
p - m o r p h i s m condition. Finally, let ~ be an internal set of ~anA(a). Then

Z ] - l [ ~ ] = { X f ( Y ) : ~ • Y} = { Z : h y ( ~ ) • Z} = hf(~-~"). Therefore, X ] l [ ~ ]
is an internal set of ~anA(~) showing t h a t X f is indeed a p - m o r p h i s m .
Now assume t h a t f : a --+/3 is injective and let U be a world over the
set {Pi : i < a}. T h e n hf[U] is a world over {pf(i) : i < a}. Hence it
is a consistent set over {Pl : i < /3}. Let V 2 hi[U] be a world. T h e n
h}-l[Y] _D hil[hl[U]]. Hence, X I ( V ) = h]x[Y] = U. So X I is onto. Now
assume t h a t f is surjective and let Y and Z be two different worlds over
(Pi : i < /3}. T h e n there exists a formula ~ such t h a t ~ • Y b u t ~ • Z.
There exists a formula ¢ over (Pi : i < a } such t h a t h f ( ¢ ) = ~. (Simply
choose a f u n c t i o n g : / 3 - - + asuchthat g o f ( i ) = i , for a l l i < a. T h e n put
¢ = hg(~).) It follows t h a t ¢ • X I ( Y ) but ¢ ¢ X f ( Z ) . []
An application of these theorems can be found in so-called weak canonical

frames. Suppose we restrict ourselves to a finite subset, say { P 0 , . . . ,Pn-1}.
Languages and logics based on finitely m a n y variables are called w e a k in [66].
If A is a logic then A [ n denotes the set of t h e o r e m s of A in the variables
{Pi : i < n}. It is clear t h a t A = U i c ~ A I n. So a logic is d e t e r m i n e d
already by its weak fragments. As it turns out, Eanhrn(n) = Earth(n). (This
is straightforward to verify.) We call a frame of the form E a n h ( n ) , n • w,
a w e a k c a n o n i c a l f r a m e . It follows from T h e o r e m 2.8.11 t h a t if n < m,
~anA(n) ~-~ ¢ a n A ( m ) and so Th ~anA(n) D Th ~anA(m). T h e situation is
therefore as follows.
A = Th EanA(~0) C . . . C_ Th CanA(2) C_ Th EaCh(l) C Th Earth(0)

Equality need not hold. F u r t h e r m o r e , the following is easy to establish.
THEOREM 2.8.12 (Weak Canonical Frames). Let A be a normal polymodal

logic. Then A = [~ne~ Th Earth(n).
T h e question arises what h a p p e n s if a logic is equal to the theory of one
of its weak frames.
DEFINITION 2.8.13. Let A be a modal logic. We call A n - c h a r a c t e r i z e d

/Y
A = Th EanA(n) .
(9 is n - a x i o m a t i z a b l e o v e r A if(9 = A@A for some set A with var[A] C_

(Pi : i < n}. (9 is n - a x i o m a t i z a b l e if (9 is n - a x i o m a t i z a b l e over K ~ . If (9 is
n - a x i o m a t i z a b l e over A it need not be n - a x i o m a t i z a b l e simpliciter, for A itself
m a y not be n - a x i o m a t i z a b l e . Clearly, if (9 is finitely axiomatizable, it is also
n - a x i o m a t i z a b l e for some n, but the converse need not hold. F u r t h e r m o r e ,
for every n one can find a logic which is n + 1 - a x i o m a t i z a b l e , b u t not n -
axiomatizable.
E x e r c i s e 65. Let 9Y¢ = (~,/3, x) be a local model. Show that Thg~t, the
theory of the point x in the model, is a maximally consistent set.
E x e r c i s e 66. Show that two worlds X, Y in the canonical frame are different
i f f t h e r e i s a f o r m u l a w such that ~ E X and ~ ~ Y. Show that X :~j Y
iff there is a formula ~ such that [~jT E X but ~ ~ Y. Show that for every
ultrafilter U in WA, ['1 U ~ ~. (The first property is called differentiatedness,
the second tightness and the third compactness. See Section 4.6.)
E x e r c i s e 67. Let A1,A2 be two normal logics. Show that A1 _C A2 iff

EanA2 (R0) is a generated subframe of ¢anA1 (R0)- Hint. Every A2-consistent
set is also Al-consistent.
E x e r c i s e 68. Show that ffanK(~0) contains 2 ~° many worlds, and 2 ~° worlds

without successor. Show that if A is a normal logic and T, U worlds such that
T <~iU then there are 2 ~°/-predecessors of U. What about the weak canonical
frames?
E x e r c i s e 69. Let A be a logic and ~ a finite Kripke-frame for A. Show that

(~, 29) is a generated subframe of ganA(~0).
E x e r c i s e 70. Show that Th (n, >) is 0-axiomatizable. Hint. The axiom [ ] n ±

is obviously not sufficient, but a good start. Now choose additional constant
axioms carefully so that only the intended frames remain as frames for the
logic.
2.9. T h e L a t t i c e s o f N o r m a l a n d Q u a s i - N o r m a l Logics
Recall that a lattice is complete if any set S has a least upper bound,
denoted by [_]~Esx or simply by [_JS, and a greatest lower bound, denoted by
[-]~ESX or simply by [']S. A lattice has greatest lower bounds iff it has least
upper bounds. Hence, a lattice is complete iff it has least upper bounds. We
can rephrase completeness by using limits of directed systems. Let ~ -= (I, ~)
be a partially ordered set. This set is d i r e c t e d if for i, j E I there exists a
k E I such that i, j _< k. An indexed family (x(i) : i E I) is called an u p w a r d
d i r e c t e d s y s t e m o v e r 2 if x(i) < x(j) whenever i ~ j. For example, let S
be a subset of L. Take I := S <s°, the set of all finite subsets of S, ordered by
inclusion. For a d E I put x(d) := [_]d. Let S + := (x(d) : d C S<S°). S + is
an upward directed system. For an upward directed system X = {xi : i E I}
we write lim~ X to denote the least upper bound. It is clear that we have
lim~ S + = [_]S. (With S given, the set I is uniquely defined, and may be
dropped.) Analogously a d o w n w a r d d i r e c t e d s y s t e m is defined. For a
downward directed system we write lim~ S for the intersection. If S is upward
2.9. The Lattices of Normal and Quasi-Normal Logics 97
(downward) directed, so is x [ S S = { x U y : y E S } and x ~ S = { x ~ y : y E S}.

A lattice is called u p p e r c o n t i n u o u s if intersection commutes with upward
limits, t h a t is, if x ~ lim S = lim (x ~ S) for all upward directed sets S. It is
called l o w e r c o n t i n u o u s if join commutes with downward limits, t h a t is, if
x U lim S = lim (x U S). A complete lattice is c o n t i n u o u s if it is b o t h upper
and lower continuous.
THEOREM 2.9.1. A complete, distributive lattice is upper continuous iff
it satisfies the law (jdi.) and lower continuous iff it satisfies the law (mdi.).
(jdi.) anLJB -- [.J(amB)
(mdi.) aU~B = [-](aUB)
PROOF. We show only the first claim, the second is dual. Let B be
a set and B + be the family of finite joins of elements of B. This is an
upward directed system. Then, by distributivity, the family a ~ B + defined
by a R B + := (a [~ U d : d E B <~° ) is identical to the family
(a R B) + := (UzEd a N x : d • B<~°) .
Thus we have the following identities.
anUB _- a~limB +
= lira a ~ B +
_- lim (a ~ B)+
= U(a B)
[]
DEFINITION 2.9.2. A locale is a complete, distributive and upper con-

tinuous lattice. A homomorphism from a locale 2. to a locale ~ is a map
h : L -+ M commuting with finite intersections and arbitrary joins.
We need to comment on this definition. Locales are also called frames in the
literature and what we have called a homomorphism of locales is actually a
homomorphism of frames (see [110]); a homomorphism between locales goes
in the opposite direction. (The open sets of a topological space form a locale;
the maps between them are continuous maps. More about this in Chapter 7.)
However, due to a clash in terminology we have departed from this convention.
Logics in general are ordered by set inclusion. Moreover, they form a
lattice with respect to this ordering. Normal modal logics are identified with
their tautologies, so A1 _< A2 is equivalent with the fact t h a t all tautologies of
A1 are tautologies of A2. It is possible to spell out exactly how to compute the
join and meet of two logics if their axiomatization is known. Moreover, we will
show that the lattice of normal logics is distributive and upper-continuous.
First of all, however, note that if Ai, i • I, is an indexed family of normal
logics, then the intersection f ~ c z Ai is a normal modal logic as well. The
reason is simply that if each Ai is individually contains the classical tautologies
and the (bd--+.) postulates and is closed under the rules (sub.), (mp.) and
(mn.), so is the intersection. (The reader might also recall t h a t logics are
defined as closed sets of a closure operator; an intersection of any n u m b e r
of closed sets is always closed.) The union, however, is generally not closed
under these operations. On the other hand, if the logics A~ are axiomatized as
K~ G X/, then the least logic containing all A / m u s t be K~ • U i c I x i . So, an
axiomatization of the union is quite easily obtained. We are interested in an
axiomatization of the meet as well. To this end, let us concentrate on the case
of a finite intersection, say of K~ @ X1 and K~ G X2. The next theorem tells
us how the intersection can be axiomatized. The notation ~o ~/~b is used to
denote the disjunction of T~ and ~b for some suitable renaming a of variables
such t h a t var(~ ~) N var(¢) = Z.
THEOREM 2.9.3. (i) Let A1 = K,~ 4- X1 and A2 = K,~ 4- X2 be quasi-
normal logics. Then A1 7~ A2 = K,~ 4- Y where
Y= {~v¢:~x~,¢ c x2}.
(ii) Let A1 = K G X1 and A2 = K ® X2 be two normal modal logics. Then
AINA2=K@Y where
Y = {KtT V [~b : ~o E Xl,~b E X2,[~ a compound modality} .
R e m a r k . The logic A1 N A2 defined above does not depend on the choice
of the renaming (r of variables in qo V ¢.
PROOF. The second claim follows from the first as follows. Assume t h a t
If A1 = K ~ G X i . Then A1 = K ~ + { ~ q p : qo E X1, [] compound}, and similarly
for A2. Then by (i), A1 N A2 = K~ ÷ Z for the set
Z := {~:~1(~ ~/ [~2~b : ~ C X l , ¢ E X2, ~1, [~2 compound}
However, the set Y as given in (ii) is actually sufficient. For notice t h a t if
~ ( p ) := [~l(p)Agq2(p) then in K~ we have [ ] a V ~ j 3 ~- 7qlaV[]2/3; and so every
m e m b e r of Z can be deduced from a m e m b e r of Y. Since the intersection is
a normal logic, we actually have Az N A2 = K~ ® Y. So let us prove the
first claim. Let ~0 Q ¢ E Y. Then A1 ~- qo Q ~ and A2 ~- qo V ¢ so t h a t
K~ + Y C_ K~ + X 1 as well as K~ + Y C_ K ~ + X 2 . For the converse
implication let X E K G X1 as well as X E K G X2. Then X can be derived
from substitution instances ~0~ of some T E X1 using modus ponens, and from
instances ¢ " of some ¢ E X2 using modus ponens. Now, in boolean logic,
a formula X follows both from a set V of formulae and from a set W iff it
follows from {a V 13 : a E V,/3 E W}. (Namely, we have A v -+ x as well
as A W ~ x and so A v v A w. -+ .x, from which the claim follows by the
distributivity laws.) So we can deduce X from formulae of the form ~ r V ¢v
with T E X1, ~ E X2 for some substitutions T, V. Generally, it is not possible
2.9. The Lattices of Normal and Quasi-Normal Logics 99
to find a single substitution p such that ~ r V ¢ " = (~ V ¢)P, since r and v

might disagree on a common variable. This is why we take instead of p the
formula ~ , which is disjoint in variables from ~. Then there is a substitution
p such that ~ r V ¢ " = ( ~ V ¢)P. []
We remark here t h a t the infinite meet of logics cannot be given a canonical

axiomatization in this way. This should be at least plausible from the con-
struction of the set Y above. This construction breaks down in the infinite
case. This is the reason why lattices of modal logics a r e not continuous, that
is, do not satisfy all infinite distributive laws. At the m o m e n t it is not proven
that the construction does not work. We will provide an explicit counterex-
ample in the exercises.
THEOREM 2.9.4. The set of quasi-normal x-modal logics operators is a

locale.
PROOF. Let (9 = K~ + X, Ai = K~ + Yi. Then
e n UiEIAi : (K,¢ + X ) Iq (K~ + UiEIYi)

= K,~+{~ 9¢ : ~ c X,¢ c U<,YJ
= K~ + U i e i { ~ ~/¢ : ~ E x , ¢ c Yi }
= U<~ K ~ + { ~ 9 ¢ : ~ c X,¢ c Y J
= [..J~eI ( K ~ + X ) M(K+Yi)
= [J~ez O ~ A ~
Moreover,
e u ( & n A2) = K,~ + X U {(p 9 '¢' : (p E Y1,'¢ E Y2}

= K,~+{~o~/¢:~ocXUYI,~EXUY2}
= ( e u A 1 ) n (e u & )
The step from the first to the second line needs justification. P u t A :=
X U { ~ ~/ ¢ : ~ C Y1,¢ E Y2}, and E := {~ Q ¢ : ~ C X U Y I , ¢ C X U Y 2 } .
Assume ~ C A. Then either p C X or p is of the form ~ Q X where ¢ E Y1 and
X E ]I2. Assume the first. Observe that ~ Q ~ C E. Clearly, ~ E K~ + ~ ~/~,
and hence ~ c K ~ + E . Now assume ~ ~ X . Then it is of the f o r m ¢ VX,
with ~ C Y1 and X C Y2. Then also ~p C E. This shows K~ + A C K~ + E. For
the converse inclusion, assume ~ C E. Then ~ = ~ 9 X where ~ C X U Y1 and
x E X U Y ~ . If e i t h e r ¢ E X o r x C X , then~CK~+Xandso~EK~+A.
However, if ~ C Y1 and X E Y2, then p C K~ + A, since it is in A modulo
renaming of some variables. []
THEOREM 2.9.5. The set of normal a-modal logics forms a locale. More-
over, the natural embedding into the locale of quasi-normal logics is continu-
ous, that is, it is a homomorphism with respect to the infinitary operations.
PROOF. Clearly~ since the infinite intersection of normal logics is a normal

logic, the embedding is faithful to arbitrary intersections. W h a t we have to
show is that the quasi-normal join of normal logics is also normal. To see this,
let 8 i , i E I, be logics and let 9 be deducible via (rap.) from X C_ [-Jic1 8 i .
Then we know that [3j~ is deducible from [~jX. By normality of the 8 i ,
E]jX C_ Ui~i 8i. []
DEFINITION 2.9.6. Let A be a normal modal logic and 8 a quasi-normal

logic. The locale of normal extensions of A is denoted by ~ A; the locale of
quasi-normal extensions of 8 is denoted by Q 8. We usually speak of the lat-
tice of (normal) extensions, rather than of the locale of (normal) extensions.
Some authors use NExt A instead of ~ A and Ext A for 0 A. In algebraic

terms, the underlying set of g 8 forms a principal filter in ~ K~. As before,
the lattice of normal extensions is a complete subtattice of 0 8 , the lattice of
quasi-normal extensions. Notice that when 8 is not normal then 8 ~ ~ 8 .
Nevertheless, ~ 8 is a principal filter in 0 8 induced by the normal closure of
8 , which is unique. We will rarely study lattices of quasi-normal extensions
and be concerned only with lattices of normal extensions. The whole lattice
K~ is extremely complex even for n = 1 as we shM1 see. However, for some
strong logics the extension lattices are completely known, such as the logics
K . a l t l , $5 and even $4.3. A large part of the study in modal logic has been
centered around classifying extensions of certain strong logics in order to gain
insight into the structure of the whole lattice g K1.
Let ~ = (I, _) be a linearly ordered set. A c h a i n o v e r ~ is an indexed
family over ~, that is, an order preserving map j : ~ --~ ~ K~. The chain j
is p r o p e r l y a s c e n d i n g if for x , y C I such that x < y we have j ( x ) C j(y).
The following is easy to establish.
PROPOSITION 2.9.7. Let j : ~ -+ g K be a properly ascending chain. Then

if ~ < ~o, ~ is at most countable. And if ~ > ~o, ~ has cardinality < t~.
Observe namely that a logic can be identified with its set of tautologies.
That set is either countably infinite (in case n < ~0) or of size n. The same
holds for properly descending chains.
Recall from Section 1.1 that an element x is join compact if for every
family Yi, i E I, such that x _< U ~ I Y i there exists a finite J C I such that
x ~ [_]i~jYi. A logic is join compact in the lattice of extensions of £ K~ only
if it is finitely axiomatizable. For let A = K~ G X, and X = {~i : i C I}.
Then A < [.]icxK~ G Ti. Hence by join compactness there is a finite J C I
such that A < [.JiejK~ ® Pi- So A = K~ ® {Ti : i E J}. Hence A is finitely
axiomatizable. Now let A = K~ • {~a~ : i < n}. Assume A ~ Uic10~. So,
for each i < n, Ti E [_JiciOi. A proof of ~i is finite and hence uses only
finitely many axioms. Consequently, there is a finite set J(i) C I such that
2.9. T h e Lattices of N o r m a l and Q u a s i - N o r m a l Logics 101
94 E Uj~j(i)8~. P u t J : = U~<~ y(i). T h e n A _< UjEJSi. So A is join

compact.
THEOREM 2.9.8. A logic is join compact in the lattice ~ K,~ iff it is finitely
axiomatizable. Every element in ~ K,~ is the join of compact elements; in other
words, ~ K,~ is algebraic.
Let us close with an i m p o r t a n t concept from lattice theory, t h a t of a
dimension. In distributive lattices (in fact in m o d u l a r lattices already) one
can show t h a t if x is an element such t h a t there exists a finite chain y :=
Yo < Yl < Y2 < ... < Y~ -- x of length n such t h a t there is no u such t h a t
Yi < u < Yi+l, then any other such chain is finite as well and has length n. n
is called the d i m e n s i o n of x over y and the c o d i m e n s i o n of y relative to x.
If the lattice has a b o t t o m element 2 , the d i m e n s i o n of x is defined to be
the dimension of x over i . If the lattice has a top element, the c o d i m e n s i o n
of x is the codimension of x relative to T. T h e following t h e o r e m is of great
theoretical i m p o r t a n c e and well worth remembering. It has been shown in
DAVID MAKINSON [146].
THEOREM 2.9.9 (Makinson). There exist exactly two logics of codimen-

sion 1 in the lattice ~K1, namely, the logic of the one-point reflexive frame
and the logic of the one-point irreflexive frame.
PROOF. Let us first see t h a t the logics A. = Th [ ] a n d Ao = Th [ ] are
of codimension 1. To see that, we show in t u r n t h a t Ao = K E3 D_l_ and
Ao = K @ p ++ [3p. Note first t h a t these axioms are surely contained in the
t h e o r y of the corresponding frames, so the axiomatization yields a logic which
is possibly weaker, in each of the two cases. For the inclusions '_C' observe
the following. In K @ K]]_ we have Fq9 ++ T, so any formula is equivalent to a
n o n m o d a l formula, and in K ® p ++ Kip we have [~9 ++ 9, so again any m o d a l
formula is equivalent to a n o n m o d a l formula, by a simple induction. T h u s
any axiom extending either logic can be written into a form 9, 9 Ilonmodal.
But if 9 does not hold in either logic, it does not hold in classical logic as
well. But there is no strengthening of classical logic which is consistent, by
T h e o r e m 1.7.14. So, K @ [ ] i is maximally consistent, and contained in A.,
so the two must be equal. Likewise, K E3 p ++ 7qp = Ao.
Let (9 ~ Ao = K G [ ] 2 , the logic of the o n e - p o i n t irreflexive frame. T h e n
® ~ K ® [D_L is inconsistent. Thus El± is inconsistent with (9 and so we have
0 T E (9, t h a t is, A _D K . D . T h e n any consistent formula w i t h o u t variables
is equivalent to either ± or T. Thus the m o d a l algebra 2 ° on two elements
0, 1 such t h a t II0 = 0 and II1 = 1 is a (9-algebra. B u t 2 ° ~ Kip ++ p, and so
T h 2 ° = Ao, so t h a t (9 C_ Ao, as required. []
Notes on this section. T h e locales of m o d a l logics are in general very

complex. C h a p t e r 7 is devoted to the s t u d y of these locales. T h e results
Of WIM BLOK a n d WOLFGANG RAUTENBERG have been groundbreaking in

this area. Lately, KRACHT [131] has investigated the automorphisms of some
locales of modal logics. These group measure the homogeneity of the locales.
Even though some results have been obtained (concerning, for example, the
locale of extensions of $4.3), for the standard systems even the size of the
groups is still unknown.
E x e r c i s e 71. Let K . t r s m be the logic defined by the axiom

trSm : = [~<_mp _+ [Tm+lp .
Show that
K = [-]m~K.trsm .
This provides an example of an infinite intersection of logics which cannot
be given a canonical axiomatization in terms of the axioms of the individual
logics.
E x e r c i s e 72. Show that if Ai, i E I, all have the finite model property, then
so does [-]ieiAi. Similarly for completeness.
E x e r c i s e 73. Call a logic A canonical if CanA V A. (See Section 3.2.) Show

that if A~, i C I, are canonical, so is l l i c i A i . Moreover, if A1,A2 are canon-
ical, so is A1 ~ A2. So, the canonical logics form a subtocale of the locale of
normal logics.
E x e r c i s e 74. Show that the recursively axiomatizable logics form a subtat-

tice with [7 and •. Show that they do not form a sublattice with [_J.
E x e r c i s e 75. Show that there is no order preserving injective map from the
ordered set of the real numbers into ~ K1.
E x e r c i s e 76. Show that the finitely axiomatizable logics are closed under
finite unions. Moreover, if ® is weakly transitive, the finitely axiomatizable
logics in ~ 0 are also closed under finite intersections.
E x e r c i s e 77. Show that a logic is Halld@n-complete iff it is not the intersec-

tion of two quasi-normal logics properly containing it.
* E x e r c i s e 78. Show that the infinite intersection of decidable logics need

not be decidable.
CHAPTER 3
F u n d a m e n t a l s of M o d a l Logic II
3.1. L o c a l a n d G l o b a l C o n s e q u e n c e R e l a t i o n s
With a modal logic A typically only the relation ~-A is considered as an
associate consequence relation. However, in many applications it is useful to
take a stronger one, which we will call the global consequence relation. It is
denoted by l]- h and defined as follows.
DEFINITION 3.1.1. Let A be a modal logic. Then A I~-A ~ iff ~ can be

derived from A and A using the rules (rap.) and (ran.): ({p}, [3jp) (j < t~).
We say that ~ f o l l o w s globally f r o m A in A if A I~h ~. IF-A is called the
global consequence relation o f A.
In the light of the definitions of Section 1.4, ([P~, ~-A) and (T~, I~-A) are
actually two different logics with identical sets of tautologies. However, since
it is customary to identify modal logics with their set of tautologies, we will
differentiate F-A and I~-A by using the qualifying adjectives local and global,
respectively. In ~-A the rule (mn.) is only admissible, whereas in IF-A it is
derivable.
The geometric intuition behind the notions of global versus local conse-
quence is as follows. Take a geometrical model 97t := (3,/3, x) and a formula
~. We say that ~ holds locally in 9~t if (3,/3, x) ~ ~ and that ~ holds glob-
ally if (3,/3} ~ T. Alternatively, we may distinguish between local and global
models. A local model is a triple (3,/3, x) with 3 a frame,/3 a valuation into
3 and x a world. A global model is a pair 91 := (3,/3)- A local extension
of 91 by x is the triple 91z := (3,/3, x). We say that a local model 9Yt is a
local A-model for T if 9Jr is a model based on a frame for A and T holds
locally in it; and we say that a pair 91 is a global A-model for ~ if every
local expansion of 91 is a local A-model for ~. By the deduction theorem for
~-A and Theorem 2.8.7 we have the following completeness result. A F-h Tiff
for every A-frame and every local model 9~t based on it, ~ is true in 9)~ if A
is true in 9)I. It will be shown below that an analogous completeness result
holds with respect to H-A. Fundamental for the completeness is the following
fact. (Recall that []~A was defined to be the closure of A under (mn.).)
103
104 3. F u n d a m e n t a l s o f M o d a l L o g i c I I
PROPOSITION 3.1.2 (Local-Global). For any given logic A, A I~-A ¢ iff

[]~A ~-A ¢ iff []A 0 ~-h ¢ for some compound modality [] and a finite set
AoCA.
PROOF. In Section 2.1 it was proved that any derivation of a formula ¢
can be transformed into a derivation of ¢ in which all applications of (mn.)
are done before (rap.). Hence, if A I~-A ~ then ~ is derivable from the closure
of A under (mn.) by means of (rap.) only. This shows the first equivalence.
The second holds by compactness and the fact that for any pair [], []t of
compound modalities there exists a compound modality []" such that • ~ p ~-A
rap; []tp. []
PROPOSITION 3.1.3. Let A be a modal logic. Then A I~-A ~ iff for every
global model ~ = (~,/3) such that ~ ~ A, we have fit ~ ~ if 9~ ~ A.
PROOF. Assume A If-A ~. Then []~A ~-A ~. Now let fit = (~,/3) be a
global A-model and assume 9~ ~ A. Take a local expansion fft~ := (~,/3, x).
Then ~ ~ ~]~A. Hence 9 ~ ~ T. Since this does not depend on the choice
of x, gt ~ ~,. Now assume that A K A ~. Then ~ A Y A T. Thus the set
~ A U (-~T} is consistent and is therefore contained in a maximally consistent
set W. Let ~ be the subframe of ~anA(var) generated by W, and let n be the
natural valuation, defined in Section 2.8. Then (5, ~) a A, but (~, ~) t~ ~, as
required. []
The previous theorem established the correctness of the notion of global con-
sequence. From now on the relation ~-A will also be called the l o c a l c o n -
s e q u e n c e r e l a t i o n if that qualification is necessary. Many notions that we
have defined so far now split into two counterparts, a local and a global one.
However, some care is needed due to the fact that many definitions take ad-
vantage of the fact that F-A admits a deduction theorem whereas t~-A generally
does not (see below). For example, by the definitions of Section 1.6 a logic
is called g l o b a l l y c o m p l e t e if for every finite set A of formulae and each
formula T if A KA ~ then there exists a Kripke-frame ~ for A and a valuation
/3 such that ([,/3) ~ A but (~,/3) ~ ~. (Instead of a finite set A we may just
take a single formula 5, e. g. A A.) If the frame can always be chosen finite
then we say that A has the g l o b a l f i n i t e m o d e l p r o p e r t y . Likewise, A is
g l o b a l l y d e c i d a b l e if for finite A the problem 'A I~-A ~' is decidable, that is,
we have an algorithm which for any given finite set A and formula T decides
(correctly) whether or not A I~-A ~. Similarly, for a given complexity class
we say that A is g l o b a l l y C - c o m p u t a b l e ( g l o b a l l y C - h a r d , g l o b a l l y
C - c o m p l e t e ) if the problem 'A I~-A ~?' is in ~ (is C-hard, is C-complete).
A is l o c a l l y d e c i d a b l e (locally c o m p l e t e etc.) if it is decidable (complete
etc.) simpliciter.
We have seen earlier that the semantics for the local consequence relation
leads to pairs (P~, F) where 9A is a modal algebra and F a filter. Since the set
3.1. Local and Global Consequence Relations 105
of designated elements must be closed under all rules, for I~-h we must now
also require the set of designated elements to be closed under the algebraic
counterpart of the rule (mn.).
DEFINITION 3.1.4. Let 91 be a modal algebra, and F C A a filter. F is
called o p e n if it satisfies (fiNN.): If a E F and j < ~ then also " j a C F .
LEMMA 3.1.5. Let 92 be an algebra and F be an open filter. Define 0 F by

a OF b iff a ++ b C F . T h e n OF is a congruence.
PROOF. In Section 1.7 we have shown that OF is a congruence with

respect to the boolean reduct. Hence, we only need to verify that if a @E b
then also mja b E mjb. So, suppose that a O F b. Then, by definition,
a ~-+ b E F. Thus, a --+ b E F, from which mj(a --~ b) C F, since F is open.
Hence, by (mp.) closure, n j a --~ mjb E F. Similarly, mjb --+ m j a C F is
shown, which together with mja --+ mjb gives mja ++ l i b E F. And that had
to be demonstrated. []
Usually, we write 92/F instead of 92/0F.
THEOREM 3.1.6. Let A be a modal logic. The global consequence relation

of A has a unital semantics.
PROOF. Let $ be the set of all pairs (92, F ) where (i.) 92 is a modal
algebra, (ii.) F an open filter in 92, (iii.) F-(~,F ) _D If-h and (iv.) (9/, F ) is
reduced. By the results of Section 1.5,
IF-A _-- ~ l- (91,F)

(91,F}ES
By Lemma 3.1.5 (92, F ) is reduced only when F = {1}. Thus, t[-h has a unital
semantics. []
As a useful consequence we note the following theorem.

LEMMA 3.1.7. Let A be a modal logic, ~1, ~2 and ¢ be modal formulae.
Then
~1 ~ ~2 IrA ¢(~1) ++ ¢(~2) .
The characterization of [P-A in terms of matrices fits into the geometrical
picture as follows. If 92 is a A-algebra and (92, {1}) a reduced matrix, then a
valuation/3 into that matrix makes ~ true just in case ~(~) = 1. If 92 is the
algebra of internal sets of a frame 5, then 1 is simply the full underlying set,
namely f . So, the corresponding geometrical model is nothing but the global
mode] (5,/3)-
Now we turn to the interconnection between local and global properties
of a logic.
106 3. F u n d a m e n t a l s of Modal Logic II
PROPOSITION 3.1.8. I f A is globally decidable (has the global finite model

property, is globally complete) then A is locally decidable (has the local finite
model property, is locally complete).
We will now prove that K~ has the global finite model property. The
proof is an interesting reduction to the local property. Notice t h a t K ~ has
the local finite model property, by Theorem 2.7.9.
LEMMA 3.1.9. Let k := ~(sf(~) t2 sf(~b)). Then we have
t~-K~ ¢ iff [] <-2~ ~ t- K~ ¢ .
PROOF. Surely, if []-<2~ ~-Z~ ~, then also ~ tFK~ ~. So, assume

[] -<2k~ ]~K~ ¢. Then there exists a finite model (f, t3, w0) ~ ~ -<2k~; ~ ¢ . More-
over, by T h e o r e m 3.3.8 we may assume that ~ is cycle-free, and t h a t between
any pair of points there exists at most one path. Let A := sf(~) U sf(¢) and
put S(y) = {X C A : (~,~,y) ~ X}. Let g be the set of all y in f such t h a t
along any p a t h from w0 to y there are no two distinct points v and w such
that S(v) = S(w). Then any p a t h from w0 to y E g has length < 2 k, because
there are at most 2 k subsets of A. Now define ~ j on g as follows, y ~ j z
iff(1.) y < ~ j z or (2.) for s o m e u C g w e h a v e y . ~ j u a n d S(z) = S(u). P u t
7(P) := fl(P) N g. We will now show that for every y C g and X C A
¢*
This is true for variables by construction. The steps for negation and con-
junction are clear. Now let X = 0iS. If (f,~, y) ~ 0j5 then for some z such
that y <]j z we have (f,f~,z) ~ 5. There are two cases. C a s e 1. z C g. T h e n
by induction hypothesis, (g, 7, z) ~ 5. From this we conclude (g, 7, Y) ~ 0jh,
s i n c e y ~ j z. C a s e 2. z C g . Then there is a u C g s u c h t h a t S(u) = S(z).
Therefore, by construction of 9, Y 4 j u. Furthermore, (~,/3,u) ~ 5 by defi-
nition of S ( - ) . So, ( g , % u ) ~ 5 by induction hypothesis. From this follows
(g,%Y) ~ 0jh, since y *lj u. This exhausts the two cases. Now suppose
(g, ~, y) ~ ~jh. Then (g, % z) ~ 5 for some z such t h a t y ~ j z. By induction
hypothesis, (~,~,z) ~ 5. Ify<~jz, then also (~,~,y) ~ ~jh. If, however, y ~ z,
then there is a u such t h a t y <~ u and S(u) = S(z). By definition of S ( - ) ,
([,f~,u) ~ 5, from which (~,Z,Y} ~ (}j5 as well. Now since from w0 there is
always a p a t h of length < 2 ~ to any point y G g, we have (I, ~, Y) ~ ~a for all
y G g, and so (g,f~,y) ~ ~ for all y. Consequently, (9,~,w0) ~ ~ a ; - ~ ¢ , as
required. []
THEOREM 3.1.10. K~ has the global finite model property.
Notice t h a t even if ~ was originally cycle-free, we might have inserted cycles

into g. This is in some cases unavoidable. For example, there is no finite
cycle-free model against 0 T I~-K p, but there are infinite cycle-free models
3.1. Local and Global Consequence Relations 107
as well as finite models with cycles. The bound for k in the proof can be
improved somewhat (see exercises). This theorem has a great number of
strong applications as we will see in the next section.
VALENTIN GORANKO and SOLOMON PASS'/have shown in [87] that the
global properties of a logic A correspond to the local properties of a logic
A m which arises from A by adding a so-called universal modality. (See Sec-
tion 2.5.) Recall that A m is defined by
A m := A ® S5({0jp--+ 0 ~ p : j < a})
Abbreviate [3~ by • . It is not hard to show that the canonical frames for A m
satisfy two properties. (1.) The relation • (:= <~) is an equivalence relation
on the set of points, (2.) For all j < ~, <~/ C_ 4. (It also follows from the
results of Section 3.2.) By completeness with respect to canonical frames, A N
is complete with respect to frames satisfying (1.) and (2.). Moreover, a rooted
generated subframe of a frame ~ satisfying (1.) and (2.) actually satisfies (1'.)
• = f × f . Thus, A I is complete with respect to frames satisfying (1'.) and
(2.). It is easy to construct such frames when given a frame ~ for A. Namely,
put • := f × f ( = : <1~),[I := (f,(<~j : j < n + l ) ) and~nu := (IN,F). Since
m~a -- ~ iff a ~ f , and • ~ f = f , IF is closed under m. Consequently, ~tt is
well-defined. Moreover, if ~ is a frame for A, ~tt is a frame for A m - - and
conversely.
Recall the definition of ~P~ from Section 2.1. Continuing our present
notation we write ~P~(•) for the language obtained from 2~ by adding 77~.
Let us agree to call a formula p l a i n if it does not contain any universal
modality. So, 9 E 2~(m) is plain iff 9 C ~P~. A d e g r e e 1 c o m b i n a t i o n of
plain formulae is a formula 9 E ~P~(•) such that no • occurs in the scope of
another occurrence of m.
PROPOSITION 3.1.11. Let A be a modal logic, A a set of plain formulae,
and 9 a plain formula. Then A I~-A 9 iff • A ~-Am 9- In particular,
~A 9 iff ~-A m 9 •
PaOOF. Suppose that A /]- A 9- Then []~°A ~-A 9 and so []~A ~-hm 9.
Now • A ~-Am ~ A , as can be shown easily. Hence • A ~-Am 9. Now assume
Uhat A ~zA 9- Then there exists a global model fit := (~,/3) such that fit ~ A
~nd fft~ 9- Thus for some local extension 9~, fits ~ 9. Then 9X := (~l,/3, x)
is a local A I - m o d e l such that 9X ~ • A ; ~9, as required. []
A sharper theorem can be established. Before we can prove it, however, we

aeed the following auxiliary theorem concerning simplifications of formulae.
LEMMA 3.1.12. Let A be a polymodal logic and A m be the extension by a
universal modality. Then any formula in 2 ~ ( • ) is deductively equivalent to a
degree 1 combination of plain formulae.
PROOF. Let ~ be given. We can assume T to be in normal form. The

following equivalences are theorems of K~ for [] -- [:]j, j < ~, or [] --- ~.
[]lp ~ l~.J_v mp
• (qv~p) ++ ~ q v , , p
EB(qVCp) ++ ~ q V C p
E~(qn ,.p) ++ []qAE~,.p
m(q n Cp) ++ []_Lv ([]q n Cp)
Analogous equivalences can be derived for the dual operator from these upper
six equivalences. The lemma now follows by induction on the degree of ~. []
THEOREM 3.1.13 (Goranko &=Passy). For the following properties ~3, A
has ~3 globally iff A II has ~3 locally: decidability, finite model property,
completeness.
PROOF. One direction follows from Proposition 3.1.11; namely, if A m has
locally, A has ~ globally. So we have to prove the converse direction. The
idea to the proof is to reduce a statement of the form 'bAa ¢ ' to a boolean
combinations of problems of the form 'X IkA T'. (Moreover, this reduction will
be effective, so it is enough for the proof of decidability of 'k h . ¢ ' if we show
the problems 'X IVA ~' to be decidable.) Now start with 'b A, ¢'. Transform
¢ into conjunctive normal form. This does not affect theoremhood of ¢.
So, without loss of generality ¢ can be asumed to be already in conjunctive
normal form. Moreover, we have seen in Lemma 3.1.12 that ¢ is deductively
equivalent to a degree 1 combination of plain formulae. So, we may as well
assume that it is already of this form. If ¢ is of the form ¢1 A ¢2 the problem
'bArn ¢' is equivalent to the conjunction of 'bArn ¢i' and 'hA.. ¢2'- Hence,
assume now that ¢ is not of that form. Then ¢ is of the form Vi<n .piv~av~-,
where all Pi, cr and ~- are plain. We claim that bAU Vi<n mpiV ¢(rv~- iff either
for some i < n ~r IF A Pi, or -~cr IF A ~-. To see this, assume that the left hand
side fails. Then there exists a local All-model (~,/3, x} ~ Ai<n ~mpi; m~a; -~-.
We may assume that ~ = ®ll for some A-frame (5. Then 93~ := ((5,/3) is a
global A-model and 9Jr ~ ~cr as well as ~ J~ pi for all i < n and 9Y¢ ia T, as
required. Now assume that the right hand side fails. Then there exist global
A-models ~ i = (~i,/3i} such that ~ i ~ ~ , ff)ll ~ pl for all i < n and a global
A-model 9~ = ((5, 7) such that 92 ~ a and ~ ~ ~-. In particular, 9 ~ ~ -~- for
some local extension 9~, of 92. P u t E) := ~i<,~ ~i G (5. Let 5 := ~i<,~/31 • V.
Then (~ll,5, x} ~ A i < , - ~ l p i ; I1-~;-~-. This concludes the proof in the case
of decidability. For completeness and finite model property, notice that in the
previous construction if ~ and (5 are (finite) Kripke-frames, so is ~ll. []
Notes on this section. The universal modality has enjoyed great popularity in
3.2. Completeness, Correspondence and Persistence 109
modal logic. It was observed in [86] that the universal modality has in con-
junction with nominals the same expressive power as the difference operator,
explored by MAARTEN DE RIJKE in [177]. It was proved subsequently that
A m shares few properties with A (apart from those which they must share by
virtue of the results of this section). In passing from A to A m finite model
property can get lost (FRANKWOLTER [235]), decidability (EDITHSPAAN
[202]) and even completeness, see Section 9.6. The number k in Lemma 3.1.9
cannot be significantly reduced. In MARCUS KRACHT [130] it is shown that
asymptotically k must be at least as large as 2 cv~, where n is the size of ~.
Moreover, EDITH SPAAN [202] has shown that K . a l t l is globally P S P A C E -
complete and that K~ is globally EXPTIME-complete. So, adding the uni-
versal modality can raise the complexity to any higher degree.
E x e r c i s e 80. Show the following improved bounds for global to local reduc-
tion. Define A := sf(~) U var(¢). Further, let # := max{dp(¢), 2~A}. Show
that
E x e r c i s e 81. Let A be a logic. Let r ( T , ¢ ) be such that ~ ]~-h ¢ iff

[]<r(v,¢)T ~-A ¢. Now let A be a logic which is locally decidable, but not
globally decidable. (Such logics exist, see Section 9.4.) Show that r ( ~ , ¢ ) is
not computable.
E x e r c i s e 82. (R. E. LADNER [137].) Let A C $5 be consistent. Show

that satisfiability of a formula is NP-complete. Hint. Clearly, the problem is
NP-hard. To show that it is in NP, show that any formula can be reduced to
a formula of depth 1.
E x e r c i s e 83. Show that a modal logic is locally tabular iff it is globally

tabular.
3.2. C o m p l e t e n e s s , C o r r e s p o n d e n c e a n d P e r s i s t e n c e
Clearly, Kripke-models are easier to handle than canonical frames. Mostly,
it is easier to reason in a Kripke-structure than to reason syntactically by shuf-
fling formulae. Moreover, canonical models are very difficult structures. All
the knowledge of a logic is coded in the canonical frame, so there is little
hope that we can use the canonical frame in any effective way. However, the
abstract existence of such a frame alone can provide us with many important
results. Consider the basic logic K~. We know that Kripke frames satisfy
all postulates of K~. Now, if ~ is not a theorem of K~, then we can base a
countermodel for ~ on the canonical frame, t h a t is we have a world X such
that
where •(p) := {X : p C X}. However, as the Kripke structure underlying

t h a t frame satisfies K~, we can actually forget the internal sets. Then, using
the same valuation we have
We have established now that K~ is complete by using the canonical frame

and 'forgetting' the internal sets. If this is possible, a logic is said to be
canonical or c-persistent.
DEFINITION 3.2.1. A logic A is called a - c a n o n i c a l if for every fl < a,

A C_ Th canA(fl). A logic A is c a n o n i c a l or c - p e r s i s t e n t if it is a-canonical
for every a, and it is called w e a k l y c a n o n i c a l if it is ~o-canonical.
The following is an easy consequence of the definition.
PROPOSITION 3.2.2. Let a, fl be cardinal numbers and a < ft. Let A be

f3-canonical. Then A is also a-canonical.
PROOF. By assumption, there exists an embedding f : a ~-~ ft. By

Theorem 2.8.11 this induces a contraction X f : ~anA(fl) --- ~anA(a). By
assumption, C~nA(fl)~ A. Then also COmA(a)~ A. []
DEFINITION 3.2.3. A logic A is called c o m p l e t e if it is the logic of its

Kripke-structures. A is called a - c o m p a c t if every consistent set based on
< c~ many variables has a model based on a Kripke-frame for A. A is called
s t r o n g l y c o m p a c t or simply c o m p a c t if A is a-compact for every a and A
is called w e a k l y c o m p a c t if it is Ro-compact.
Obviously, if a logic is compact, it is also weakly compact, and if it is weakly

compact it is complete. Neither of the converses hold; there are complete logics
which are not weakly compact and there are logics which are weakly compact
but not strongly compact. (This has been shown first in [66], who also defined
the notions of weak and strong compactness.) The reader m a y verify t h a t
logics axiomatized by constant axioms are 1-compact. Compactness is rather
closely connected with a different property of logics, called complexity (see
[82]).
DEFINITION 3.2.4. A logic A is a - c o m p l e x if every fl-generable algebra,

where fl < a, is isomorphic to a subalgebra of the algebra of all subsets of a
Kripke-frame for A.
a - c o m p l e x i t y is not directly equivalent with a - c o m p a c t n e s s if a is finite.

Rather, the right notion to choose here is global a-compactness. A logic A
is g l o b a l l y a - c o m p a c t if for every pair consisting of a set • and a formula
T, based together on 13-many variables,/3 < a, if • ~zA ~ then there exists a
Kripke-frame [ ~ A, and a valuation/3 such that ([,/3) ~ • but (~,/3) ~ ~. If
a logic is locally a + 1-compact then it is also globally a - c o m p a c t .
THEOREM 3.2.5 (Wolter). Let A be a t~-modal logic. A is globally a -

compact iff it is a-complex. Moreover, if A is globally a + 1-compact, it is
locally a-compact.
PROOF. Suppose A is a - c o m p l e x and take a set • and a formula T such
that • ~A ~. Assume that • and ~ are based on/3 m a n y variables,/3 < a.
Then there exists a model (92,7) ~ ~ such that (92, V) ~ P. By assumption,
there exists a K r i p k e - f r a m e ~ such that 92 is a subalgebra of the algebra ~ of
subsets of [. Let ~ : 92 ~ ~ be an embedding. Then 5 := L o "7 is a valuation
on ~. Now, for every ¢ ~ • we have ~ ( ¢ ) = 1, and so 5(¢) = 1 as well. Thus
([, 5} ~ ~. However, since ~(p) ~ 1, also ([, 5) ~ ~. For the converse assume
that A is globally a - c o m p a c t . Let 92 be a/3-generable algebra, where/3 < a.
Then choose a generating set X such that SX = / 3 . For each a ~ X take a
variable p~ and let V be the valuation defined by 7(P~) := a. ~ is surjective
by choice of X. Let U(92) be the collection of ultrafilters of 92. For every
U ~ U(92), ~ - I [ u ] := { ~ : ~(~) • U} is a maximally A-consistent set in the
variables pi, i </3. Now A is a - c o m p a c t and therefore for any ~-~[U] there is
a A - K r i p k e - m o d e l (gu, 5u, xu) ~ 5-1[U]. We assume that the gu are disjoint
and take g := ~ g u and 5(p) := ~ u h u ( p ) . It then follows t h a t g ~ A and
that the algebra ~ generated by the 7(P~), a • A, is a subalgebra of 91. We
show that in fact ~ ~ 92. Namely, this holds if
{ ~ : ~(~) = g} = { ~ , : 5(~) = 1 } .
In fact,
~(~) = 1 <=> (VEB)~(~) = 1
(VEB)(VU e c u
(vm)(vu []
(vm)(vu c
= g
Now for the second claim. Assume that • is a locally consistent set, based on
/3 m a n y variables,/3 < a. Then let q2 := {pc, -+ ~ : T C ~}. ~ ] ~ ; p ~ is based
on < a + 1 variables and is also consistent. For take a finite subset S; S is
without loss of generality of the form [](p~ --+ 4)0);p~, [] a compound modality
and a50 a finite subset of ~. By assumption, ~0 has a K r i p k e - m o d e l ([,/3, x),
since ~0 is locally consistent. (If a logic is globally a + 1-compact, it is also
complete for formulas with < a many variables.) Now put /3+(p~) :-- {x}.
Then
(~,fl+,x} ~ [](pa --+ ~0);P~ •
Hence S is consistent. Since S was arbitrarily chosen, [ ] ~ q ; p a is consistent.
Consequently, [ ] ~ q ]/A ~P~, from which • ~A ~P~. By a + 1-compactness,
there exists a Kripke-model (9,7) ~ k~ such that (9,7, Y) ~ P~ for some y.
Then (9, 7, Y) ~ ~, by construction of ¢2. []
COROLLARY 3.2.6 (Wolter). Let a be infinite, A a normal modal logic.

Then A is a-complex iff it is globally a-compact iff it is locally a-compact.
If a ~ /3 and A is a-canonical then A is not necessarily/3-canonical. It is
an open problem, for example, whether Rl-canonicity implies R2-canonicity.
It is possible to show that we cannot give any finite bound no such that a
logic is canonical if it is n0-compact. A simple but instructive example is
the case of logics extending K . D . The 1-canonical frame consists of a single
reflexive point (since there are only trivial constant propositions) and so every
extension is 1-canonical. But they are not all Rl-canonical. For example,
take the logic G r z . 3 . By a theorem of KIT FINE in [63] @rz.3 is weakly
compact. (An exercise with hints how to prove this theorem is provided in
Section 6.5.) We show that it is not Rl-compact, from which follows (with
the next theorem) that it is not R1-canonical.
Y : = {po} u {G(p -+ :i e v {D(p :i < j <
Each finite subset is satisfiable, taking a suitably large chain (n, ~}. For take
a finite subset X. Without loss of generality we assume that X is the subset
containing the formulas in which all and only the variables up to number no
occur. If i < no, then ifpi is true at a point x, it must have a successor y at
which Pi+l holds. Moreover, y :~ x. So let us take ~(no) :--- ({0, 1 , . . . , no}, _<}.
This is a G r z . 3 - f r a m e . (For let (~(n0),/3, k} ~ 0p. Let g be the largest
number such that g e /3(p). Then (~(n0),/3, g) ~ p;~(-~p --+ • ~ p ) , since for
any successor j > g if j ~ -~p then j > g and j ~ [~-~p, by choice of g.
Hence (~(no),/3, k) ~ (}(p A ~ ( ~ p --+ [-l-~p)), and that had to be shown.) If
we put/3(pl) := {i}, then (~(n0), /3, 0> ~ X. Now let f be a G r z . 3 - f r a m e and
(~,/3, x) ~ Y. Then pick x~ C/3(pi). Then for all i, xi <~ xi+~ (by linearity of
the frame) and x# ~ xi for all j > i. Thus we have a strictly ascending chain
of points. P u t ~/(p) := {x2k: k C w}. We have
([, % xo) Op; O(p A -+ O- p) ) .
Thus, ~ is not a frame for G r z . Contradiction.
In the exercise we have put a proof that G is not 2-compact. This has
been shown by WARREN GOLDFARB (see [32]) and also in [66]. However, G
is 1-compact.
PROPOSITION 3.2.7. Let A be a logic. IrA is a-canonical, it is a-compact.

In particular, if A is (weakly) canonical it is (weakly) compact.
The proof is easy, once we know that logics are generally complete with
respect to their canonical frame. The converse of this statement is false.
FRANK WOLTER [240] shows that the tense logic of the reals is compact
but not canonical. So, which logics are canonical? This question has been
answered for all standard systems. Generally, failure of canonicity is hard to
demonstrate, whereas the contrary is normally straightforward. K . D , K4,
Kh, K . a l t l , $4.3, $5 are all canonical, G and G r z are not.
EXAMPLE. We show that $4 is canonical. The proof for canonicity is in two
stages. First we show so called correspondence of the axioms with properties
of Kripke frames. Let f = (f, <~) be a Kripke frame. Then the following holds.
(cot.) [ ~ p --4 0p iff ~ is reflexive
(co4.) [ ~ 0 0 p -4 0p iff [ is transitive
First (cot.). Suppose that [ is reflexive, ~ is a valuation and x a world. Then
(f, 13, x) ~ p implies (~, 13, x) ~ 0p. Suppose now that [ is not reflexive, say
x ~ x. Then let fl(p) := {x}. Then we have ([,j3, x) ~ p;-~Op. Now (co4.).
Suppose [ is transitive, then it satisfies 4. For let ([,t3, x) ~ 00p. Then there
are y and z such that x < l y ~ z (f, fl, z) ~ p . By transitivity, x < ~ z and so
(f, fl, x) ~ 0p. Now assume ~ is not transitive. Then there are points x <~y <~z
such that x ~ z. Choose ~(p) := {z}. Then (f, fl, x) ~ 00p; ~0p.
So, if we can show that the Kripke structure underlying the canonical
frame is reflexive and transitive, we have proved that $4 is canonical. Assume
then that the canonical stucture is not reflexive. Then for some set X, X ¢~ X;
by definition, there must be a formula p such that ~ E X but 0 ~ ¢ X.
Hence ~ A ~ 0 ~ E X, and so the canonical frame does not satisfy the axiom
p --4 0p. Contradiction. Hence the structure is reflexive. Now assume it is
not transitive, that is, there are sets X <~Y <~Z such that X ~ Z. Then there
is a formula ~ such that ~ E Z but 0 p ¢ X. However, again by definition,
0 ~ E Y and so 0 0 p E X, showing that ~ 0 ~ A 0 0 ~ E X, in contradiction to
the assumption that our frame satisfies the axiom (4.).
LEMMA 3.2.8. Let s be a finite set of finite sequences over n. Let X and
Y be worlds such that for all ~ we have []s~ C X only i f ~ E Y . Then X <1sY
in the canonical frame for K~.
PROOF. First we show the claim for sequences. This in turn is shown
by induction on the length of the sequence. If ~ has length 0 the claim is
immediate. So let cr be a sequence and cr = j ~ - for some sequence ~- and
j<n. Let A0 := { ¢ : [ 3 j ¢ E X } , A 1 : = { 0 ~ ¢ : ¢ E Y } a n d A : = A o U A ~ .
Then E]jAo C X and 0iA1 _C X. We claim that A is consistent. To that
end, let A0 C_ A0 and A~ C_ A~ be finite sets. There exists a 5 E A1 such
that 5 ~-A 51 for all 5' E A1, as is easily shown. Now assume t h a t A0; Ai is
A-inconsistent. Then A0; 5 P A L . From this it follows that [3jA0 ~-A [ ] j ~ 5
and so [~jA0; (}i5 is inconsistent in A. However, [~jA; (}j5 C_ X and X is
A-consistent. Contradiction. Therefore, A is consistent and there exists a
world Z containing A. Then X <~j Z by definition of the canonical frame, and
Z <~ Y by induction hypothesis. Hence X < J Y. Now let s := {ai : i < n}
be a set of finite sequences over n and assume that X ~ Y. T h e n for every
i < n there exists a ~ such t h a t [Z]~'p C X but ~ ~ Y. P u t ~ := Vi<,, ~i.
Then [ : ] ~ E X but ~ ~ Y. []
THEOREM 3.2.9. Let s and t be finite sets o/finite sequences over n. Then
K~ ® [3Sp --~ [~tp is canonical. Moreover, the canonical frame satisfies <~t C
<:is.
PROOF. We prove the second claim first. Let X and Y be worlds such
that X <~t y . Assume that E * p E X. Then also [3t~ E X. Since X <1t Y we
have ~ C Y. Hence by the previous l e m m a X < ~ Y . So, <t C q~. To show t h a t
K~ D D~p --+ [3tp is canonical it is enough to show t h a t if [ is a K r i p k e - f r a m e
such t h a t qt C <~ then [ ~ [Z~p -+ Dtp. But this is straightforward. []
As an example, let [ = (f, q, 4) be a bimodal Kripke-structure. ~ satisfies
0p -+ Op iff q C ~ . [ satisfies 0Op --+ O(}p iff <~o • C_ • oq. Also, ~ satisfies
0Op ++ O0p iff q and • commute.
DEFINITION 3.2.10. Let ~ a modal formula and a be a first-order formula
in the language of predicate logic with <~ and equality. ~ c o r r e s p o n d s to a
iff a Kripke structure ~ satisfies ~ exactly if it satisfies a.
In light of this definition the following correspondences are valid.
AXIOM FIRST-ORDER PROPERTY DESCRIPTION
0T ( w ) ( 3 y ) ( x < y) definality
p -+ Op ( w ) ( x < x) reflexivity
p -~ []Op ( w y ) ( x < y -~ y < 5) symmetry
OOp ~ Op ( W y z ) ( ~ < y A y < z. -+ .~ < z) transitivity
O[3p -+ ½(}p ( W y z ) ( ~ w ) ( ~ < y A ~ < z. - + . convergence
yqwAzqw)
Op --+ [~p ( W y z ) ( x < y A x < z. -+ .y = z) partial functionality
(}p --+ [~Op (Vxyz)(z < y A x < z. -+ .y < ~) euclideanness
.3 (Vxyz)(x < y A x .q z. --+ . local connectedness
y<~zVy=zVzqy)
For polymodal logics similar correspondences can be established. I m p o r -

tant are the tense postulates p --+ D0(}lp and p -+ [:]100p- A frame satisfies
the first iff the relation <So is contained in the converse of q l , t h a t is, if x <10 y
implies y , ~ z. For a proof assume that x q0 Y ~1 x. Then put ~(p) := {x}.
T h e n (~,/3, y) ~ [:]l~p and so (~,fl, x) ~ p; 0 0 [ ] l ~ P . So the axiom is violated.

Assume then t h a t the frame satisfies the first order property. Pick a valuation
/3 and a point x. Assume ([,/3, x} ~ p. Take a y such t h a t x <~0 Y. T h e n y <]1 x
and so (~,/3, y) ~ Q p . Thus, as y was arbitrary, (~,/3, x) ~ [Z0(}~p, which had
to be shown.
THEOREM 3.2.11. A bimodal Kripke structure satisfies K . t iff <~o is the

converse of 41.
We conclude with a general theorem on logics of b o u n d e d alternative.
THEOREM 3.2.12 (Bellissima). Every logic of bounded alternative is ca-

nonical and hence compact and complete.
PROOF. Let A be of alternative a and ~ an axiom of A. We want to show
t h a t canA(R0) ~ ~. To see that, assume t h a t there is a point x and a valuation
/3 such t h a t (cnnA(l%),/3, x) ~ ~ . Let d be the m o d a l d e p t h of ~. T h e n there
are at most 1 + a + a 2 + . . • + O¢d = ctd+l--1
a-1
points reachable in at m o s t d steps
from x. Let the set of these points be X. We claim now t h a t for any set T C X
there is an internal set T ° in ~anA(~q0)SUCh t h a t T ° coincides with T on the d -
transit of x. T h u s if we put 7(P) :=/3(P)° then we have (¢anA(R0), V, x) ~ - ~ ,
and since the V(P) are internal, we now have (gnnh(R0), V, x) ~ ~, which had
to be shown. Now for the existence of these sets. Let X = {xo,... , Xn-1}.
For each pair i, j we have a set S(i, j) containing xi but not xj. (For if xi and
xj are different, t h e y represent different ultrafilters, and so there is a formula
such t h a t ~ C xl but 9~ ¢ xj. Now let S(i,j) : = ~ = { X : ~ C X } in the
canonical frame.) T h e n let S(i) = ~j#i S(i, j). S(i) contains xi but none of
the j. Thus, for any subset Y C_ X put y 0 = [.J{S(i) : i E Y}. y o is internal.
This concludes the proof. []
E x e r c i s e 84. Prove the remaining correspondence properties.
E x e r c i s e 85. Show t h a t a frame ® satisfies D iff the underlying Kripke

structure satisfies (Vx)(3y)(x ,~ y). Show t h a t K . D is canonical.
E x e r c i s e 86. Show t h a t S5 is canonical. Hint. Show t h a t the axiom B is

valid in a Kripke structure iff t h a t structure is symmetric. P r o c e e d as with
$4.
E x e r c i s e 87. Show t h a t the tense logic K . t is canonical and complete.
E x e r c i s e 88. Show t h a t a finite K r i p k e - f r a m e satisfies G if[ it is transitive

and irreflexive.
E x e r c i s e 89. Show t h a t a finite K r i p k e - f r a m e satisfies G r z iff it is reflexive,

transitive and a n t i s y m m e t r i c , t h a t is, if x <l y and y <~ x then x = y.
E x e r c i s e 90. Show t h a t if K ® ~ is canonical for all T E X , t h e n K (~ X is

canonical as well.
E x e r c i s e 91. Let A be complete. Define the Kripke-consequence F-kAfor A

as follows. • F-~ ~ iff for every A - K r i p k e - f r a m e f if ([,/3, x) ~ (I, t h e n also
(~,/3, x) ~ ~. Show t h a t ~-A C ~-k
A and t h a t equality holds iff A is c o m p a c t iff
k-~ is finitary (or compact).
* E x e r c i s e 92. Show t h a t if A is not complete, then it m a y well be t h a t t-k

is finitary. In t h a t case, however, ~_k must be strictly stronger t h a n k-A.
* E x e r c i s e 93. Show t h a t G is 1 - c o m p a c t b u t not 2 - c o m p a c t . Hint. To

show the first claim show t h a t one can only define boolean c o m b i n a t i o n s of
s t a t e m e n t s of the form [ ] ~ l A ~ D n - l L , stating t h a t there exists no upgoing
chain of points of length n + 1. Show t h e n t h a t any infinite set of such
formulae if jointly consistent has a model based on a K r i p k e - f r a m e . For the
second claim, consider the following formulae. P u t
~ := - ' 0 ( P A D i ± ) A 0 ( p A [~i+z±) .
Show t h a t the following set is consistent, but has no model based on a K r i p k e -

frame
{0 o} u i
(This is the solution of WARREN GOLDFARB as given in [32].)
* E x e r c i s e 94. Let us consider a fixed set {Pi : i < n} of sentence letters and
define for a set S C_ n the formulae Xs :-- A ~ s p i A Aids ~Pi" Consider the
following formulae
:= A <>(xs ,, D±). .[]([];o ;o) []po .

SC_n
Show t h a t the logic K1 (9 qon is n - 1-canonical but not n - c o m p a c t .
3.3. F r a m e Constructions II
This chapter is devoted to the question of making models as small as
possible or to produce models t h a t satisfy certain properties. Before we enter
this discussion, we introduce some very i m p o r t a n t terminology. First, let
= (f, (<1i : i < ~)) be a K r i p k e - f r a m e and g C_ f. Let g = (g, (<1~ : i < ~))
with <~ = <Si ~ g × g. T h e n we call g a s u b f r a m e o f [, in symbols g _ [. For
a general frame 5, ~ is a s u b f r a m e if g C F, g U__[, and G = {a N g : a C F}.
In general, for any subset p C_ f , {p • a : a E F} is called the t r a c e a l g e b r a
3.3. F r a m e Constructions II 117
of ~ in qh. T h e trace algebra is actually closed under c o m p l e m e n t and union,

as is verified. Moreover,
myb = {xEg: (VyE>g-3x ) ( y • b ) }
= {x • g (vy >5 x)(y • g. ~ .y • b)}
= gn"~(9 + b)
Hence, the s u b f r a m e based on an internal set is always well-defined. A valua-
tion/3 on [ defines a valuation 7 on fl in the n a t u r a l way, by 7(P) :=/3(P) VIg;
this valuation is often also denoted b y / 3 . Let us be given a K r i p k e - f r a m e ~:
and a subset S. We put sucj(S) = {y : (3x • S ) ( x <~y y)}. T h e m - w a v e
Wave~(S) and the m - t r a n s i t T r y ( S ) of S in f are defined as follows.
Wave~(S) := S
Waver (S) := U~<~ sucj(S)
Wave?+l(S) := Wave~(Wave?(S))
Tr?(S) := U,<~ Wavei(S)
Try(S) := U~ T~i(S)
Trf(S ) is called the t r a n s i t of S in ~. All these definitions just define subsets
of frames; but if ~ = (f, (<qj : j < nil is a p o l y f r a m e and h C f then we
can regard h n a t u r a l l y as a s u b f r a m e 0 = (h, (<l~ : j < n)) where <~ :=
<~j (3 h × h. Similarly, we write f f r ~ ( S ) for the s u b f r a m e based on the m -
transit of S. If there exists an element w such t h a t f = ~:r[({w}) (which
we will also denote by i r ~ ( w ) ) then [ is called r o o t e d and w the r o o t of [.
Rooted frames are sometimes also called o n e - g e n e r a t e d . However, we will
avoid this terminology. All definitions apply equally to generalized frames,
where the waves and transits are c o m p u t e d with reference to the underlying
K r i p k e - f r a m e s . P u t x ~ y if y C Trf(x) and x -< y if Try(y) C Try(x). P u t
Cy(x) = {y : x ~ y ~ x} and call it the c y c l e of x. A set M is a c y c l e if it
is of the form Cy(x) for some x. A frame is c y c l i c if its underlying set is a
cycle. T h e d e p t h of a point in a frame ~ is defined as follows.
dp(x) := {dp(y) : x -< y}

This is generally not a good definition, e. g. in the f r a m e (w, <>. Hence,
we require t h a t the definition is applied only to points for which there is no
infinite chain (x~ : i E w} such t h a t x0 = x and xi -< X~+l. In this case
the d e p t h is well-defined and yields an ordinal number. (This n u m b e r is not
necessarily finite.) In all other cases, x is said to have no depth.
DEFINITION 3.3.1. Let ~ be a Kripke-frame. A p r e c l u s t e r is a maximal

set P of points such that for all x , y C P and all j < n we have sucj(x) =
sucj(y). A c l u s t e r is a maximal connected set C of points such that for all
x , y C C and all j < n we have sucj(x) = sucy(y).
A cluster is a cycle in a p r e c l u s t e r , b u t n o t conversely. G i v e n a f r a m e

a n d a p r e c l u s t e r P the m a p collapsing P into a single p o i n t c a n b e t u r n e d
into a p - m o r p h i s m . T h i s can be d e r i v e d from a m o r e g e n e r a l t h e o r e m w h i c h
we will now present. Let ~ be a frame a n d ~ _ 3 be a s u b f r a m e . A s s u m e
t h a t for all x , y C g a n d all j < n we have s u c j ( x ) N ( f - g ) = sucj(y)n(f-g).
T h e n we call ® a l o c a l s u b f r a m e of 5.
PROPOSITION 3.3.2 (Net E x t e n s i o n I I I ) . Let 3 be a frame and ~5 C 3 be

a local subframe. Let ~ be a net on ¢b. P u t x ~ y if (i.) x , y E g and x ~ y
or (ii.) x, y E f - g and x = y. Then ~ is a net o n 3 .
PROOF. L e t x <]j y a n d x ~ x r. If x ~ g t h e n x = x r a n d t h e n t r i v i a l l y
x r <~jy. So, s u p p o s e t h a t x E g. T h e n x r E g a n d x ~ x r. A s s u m e y ¢ g.
B y a s s u m p t i o n on g, s u c j ( x ) N ( f - g) = s u c i ( x ' ) N ( f - g), so t h a t x ' <~j y.
A s s u m e now y C g. Since ~ is a n e t t h e r e exists a yr such t h a t y ~ y~ a n d
x ~<~jyr. T h i s shows t h a t ~ i s a n e t o n g . N e x t , let a C F. T h e n a = bUc
where b : = a N g a n d c : = a N ( f - g). Now [a]~ = [b]~ U [c]~ = [b]~ U c • F,
since [b]~ • ]F by the fact t h a t ~ is a net on 3- []
T h i s t h e o r e m allows several g e n e r a l i z a t i o n s , b u t in this form it is m o s t useful

( a n d s i m p l e ) . T h e r e is a n o t h e r v e r y useful o p e r a t i o n t h a t allows to r e p l a c e a
s u b f r a m e by a larger s u b f r a m e . I n this c o n s t r u c t i o n we say t h a t K is o b t a i n e d
by b l o w i n g u p 3 b y a p - m o r p h i s m .
THEOREM 3.3.3 (Blowing Up). Let ¢& E ~ and c : 2) --* ~5. A s s u m e that
h is disjoint with f (and so also with g). Define a new frame ~ as follows.
k := ( S - g ) u h
h
: • • S,y • <J c(y)}

: • • h,y • y}
IK :={aUb:a•F, bEH, ang=~}
If G = {c[a] : a • H} then Yt is a frame. The map d defined by d(x) : = x if
¢ h and := otherwise is a p-morphism of onto 3.
PROOF. F i r s t we verify t h a t d is a c o n t r a c t i o n of t h e K r i p k e - f r a m e s . To
see t h a t , a s s u m e x <~jk y. If { x , y } C k - h t h e n x <~ y a n d d(x) = x as well as
y = d(y). If { x , y } C h t h e n x <]hy. T h e n d(x) = c(x) a n d d(y) = c(y) a n d so
by a s s u m p t i o n on c, c(x) <~gjc(x). Hence d(x) <Jj d(y). N e x t , let x • k - h a n d
y • h. T h e n x <~jk y iff x <~jS c(y). Since x = d(x) a n d c(y) = d(y), t h e c l a i m
follows. Likewise for t h e l a s t case, t h a t x • h a n d y • k - h. T h i s shows t h e
first p - m o r p h i s m c o n d i t i o n . N e x t , let x • k, u • f a n d a s s u m e d(x) <Jj u. L e t
{x, u} C f - g. T h e n x = d(x) a n d u = d(u) a n d x <ljk u, as r e q u i r e d . N e x t
a s s u m e {x, u} C_ g. T h e n , since c is a p - m o r p h i s m , t h e r e exists a y • h such
3.3. Frame Constructions II 119
that x <~hy and c(y) = u. Then also d(y) -- u and x ~ y. Third, assume x E g
and u C f - g. Then d(u) = u and x <jk u by construction. The remaining
case is also straightforward. To see that ~ is a frame, we must show t h a t
K is closed under boolean operations and ¢~. The boolean operations are
straightforward. To show closure under ck, we take a set a E K. It is of
the form al [2 a2 where al E F and al C_ f - g and a2 C H. By assumption,
b2 :--- c[a2] • G and so also b2 • F. (10 ~ j ka l = ( ( k - h) • ¢ ~ a l ) U (h A ¢~al).
( k - h ) N C jka l = ( f - g ) N ¢~al • F. h A C kj a l = c-1 [g n ¢~al] • H. So this
case is settled. (2.) Cja2 k = ( ( k - h) n ¢]a2) U (h n ¢~a2). (k - h) n ¢jk a 2 =
c-X[(f - g) N ¢~b2] • F. h 5 ¢~a2 = c-l[g A ¢~b2]. This concludes the proof
that ~ is a frame. Finally, c -1 : F --+ K is clearly injective. This concludes
the proof of the theorem. []
COROLLARY 3.3.4 (Multiplication). Let ~ be a frame, and ~) KK_5. There
exists a natural p-morphism c : (~icI ¢3 --~ ~ : (x, i) ~-~ x. The result Sl of
blowing up ~ by c is a frame. We say that ~ has been obtained from ~ by
m u l t i p l y i n g ¢5 (~I t i m e s ) .
The depth of a modal formula corresponds quite directly to the bounded
transits.
PROPOSITION 3.3.5. Let ([,~,x) ~ p and k = dp(~). Then
PROOF. By induction on k = dp(~). If k = 0, then it is easy to check

that ({x}, ~, x) ~ ~ as well, by the fact that the evaluation clauses are only
(md~.), (rod A .) and so do not change the world at which we evaluate. Now
suppose that p = D i e , dp(¢) = k. Then
(~,Z,x)~[~j¢ ca forally~>jx ([,Z,y)~¢
Ca for all y E>j x (~r~(y),Z,y) ~ ¢
ca forallyE>jx (~rT+l(x),Z,y)~g)
[]
Having established that a consistent formula has a finite model we can

also give some bounds for the size of such a model. We can actually use the
bounds obtainable from the proof directly, but we seize the opportunity to
introduce the method of filtration. By itself it is a rather crude method for
proving finite model property, and m a n y people have found ingenious ways
of refining it so that it allows to give proofs for m a n y standard systems.
Here we only present the most basic variant. Suppose t h a t we have a model
(~, 13, x) ~ p. Assume t3 to be defined only for the variables of ~. Now let
X := sf(p). Let v ~ x w iff they satisfy the same formulae of X , t h a t is, for
all ¢ C X {~,/3, v) ~ ¢ ca (~, 13, w) ~ ~b. This is an equivalence relation. Let

[v] := { w : v ~ z w} and [f] = {[v] : v e f } . T h e r e are at m o s t 2 ~x distinct
classes, so ~[f] < 2 ~X. Now p u t [v]~j[w] iff there exists ~ e Iv] and @ e [w]
such t h a t ~ <lj ~. Next define 7 by ~(p) := {Iv] : { ~ , ~ , v ) ~ p}. This is
well-defined by the fact t h a t if v ~ p then p E var(~) and if v ~ x w t h e n also
w ~ p. This model is said to be o b t a i n e d by f i l t r a t i n g the original model.
Then
{{If], {-~j: J < ~ ) ) , 7 , Ix]) ~
Namely, by induction on ~ E X we show t h a t Iv] ~ ~ i f f v ~ ~. This is
straightforward for variables, and the only critical step is ¢ = (}iX. Here,
assume Iv] ~ (>iX- T h e n [w] ~ X for some w such t h a t [v]-~j [w]. T h e n there
exist ~ and ~ such t h a t ~ ~ x v, ~ ~ x w and ~ < j ~ . T h u s ~ ~ X by
construction and induction hypothesis and so ~ ~ (>iX, showing v ~ (}iX by
v ~ x v'. Conversely, if v ~ ()iX then for some j - s u c c e s s o r w we have w ~ X
and so [w] ~ X by induction hypothesis. B y construction, [v]-~j [w] a n d so
[v] ~ (}iX, as required.
THEOREM 3.3.6. Let ~ be consistent in K,:. Then there exists a model

with at most 2 k points, where k = ~sf(~).
N o r m a l forms are closely connected with a technique called unravelling.

There are more or less cautious variants of unravelling. T h e m o s t basic
m e t h o d is the following. Suppose we have a pointed K r i p k e - f r a m e (~,w0).
A p a t h of length r in {~, To) is a function ~r : r + 1 -+ f such t h a t 7r(0) = wo
and for a l l i < r + l w e haveTr(i)<jTr(i+l) for s o m e j < a. We say t h a t
7r(0) is the b e g i n p o i n t and 7r(r) the e n d p o i n t of 7r, denoted by ep(Tc). T h e
length of cr is denoted by g(Tr). We say t h a t 7r+ : m --+ f e x t e n d s 7r : n --+ f
if n ~ m and for all i < n, 7r(i) = 7r+(i). T h e u n r a v e l l i n g of degree r of
(f, wo) is denoted by u~(f, Wo), and defined as follows. It is a f r a m e based on
the set of p a t h s of length < r. T h e relation <~j is defined by 7r <~j 7r+ iff (i.)
f(~r +) = g(~r) + 1, (ii.) ~r+ extends ~r and (iii.) ep(~r) <~ ep(7c+). So, 7r <~j ~r+ if
7r+ is the p a t h t h a t goes just one step further t h a n ~r and to a point which is
j - a c c e s s i b l e from the end point of ~r. u~ thus defined has a unique generating
point, n a m e l y the p a t h 7r0 : 1 -+ f , sending 0 to w0; we denote it by {To).
T h e m a p e p ( - ) : ~r --+ ep(rr) sending each p a t h to its end point is not quite
a p - m o r p h i s m . However, with respect to points of restricted depth, it is as
good as a p - m o r p h i s m . We formalize this as follows. Say t h a t h : f --+ g is
n - l o c a l i c with respect to S C f if the following holds.
(gmf.) If x , y e Trn(S) and x <~j y then h(x) <~j h(y)
(grab.) If x e Tr~-~(S) and h(x) <~j u then
x <~j y for some y such t h a t h(y) = u
It is easy to verify t h a t the m a p ep : u~(~,w0) --+ f is r-locaiic with r e s p e c t to
3.3. F r a m e Constructions II 121
(w0). T h e next t h e o r e m then says t h a t whenever we can satisfy a formula

of d e p t h < r at wo then it can be satisfied in ur([,Wo) at the p a t h 7to.
PROPOSITION 3.3.7. Let ~ and fl be Kripke-frames. Suppose that h : f -+
g is n-localic with respect to S and that w C S. Let/3, 7 be valuations such
that 7(P) = h[/3(p)]. Then for all formulae ~ of degree < n
T h e p r o o f is an easy induction on ~ and is left to the reader. T h e frame

um is a s u b f r a m e of u~ if m _< n. It consists of the m - t r a n s i t of ~0 in u~. T h e
total unravelling u~ is the union of all un, n C w. This is well-defined.
THEOREM 3.3.8. Let (~,wo) be a pointed Kripke-frame. The map ep,
sending each path beginning at wo to its end point, is a p-morphism from
) onto (f, w0).
PROOF. T h e m a p is onto by definition of the transit. Now take 7r, zr+
such t h a t ~r <~i 7r+" T h e n ep(Tr) <~j ep(Tr+), by construction. Next assume
ep(Tr) <lj u. T h e n the p a t h 7r+ defined by extending 7r by just one m o r e point,
n a m e l y u, is a well-defined p a t h and we have lr <~j lr + by construction. []
T h e m e t h o d of unravelling can be used to show t h a t K ~ is complete with
respect to completely intransitive trees, by first showing t h a t it is complete
and then using unravelling to get a totally intransitive tree from a model.
This is s o m e w h a t b e t t e r t h a n the proof via n o r m a l forms, which established
completeness with respect to acyclic frames only. Finally, let us note t h a t if
= (~,F} is a frame and q : tt~(f, w0) --- ~ a total unravelling, then we can
define a s y s t e m U of sets by U := {q-l[a] : a E F}. By the fact t h a t we have
a p - m o r p h i s m , £[~(~, wo) := (u~(f, w0), U} has the same m o d a l theory as 5.
This fact is of some importance.
We r e m a r k here t h a t there is a more extreme variant of unravelling, which
is as follows. Define a t~-path to be a sequence lr = (wo, A0, Wl, A1, w 2 , . . . , wn)
such t h a t wi <~A, Wi+l for every i < n. We say t h a t 7r starts in wo and ends
in w,~. P u t 7r <~j 7r' ifTd = ( w 0 , A o , w l , A 1 , . . . ,w~,An,w~+l for some An <
and some Wn+l C f . T h e n we can form the frame ~;~(~, Wo) of all ~ - p a t h s in [
s t a r t i n g at w0. (The definition of ~ ( [ , w o ) is analogous. It is the s u b f r a m e of
~ - p a t h s of length < n.) T h e m a p ( sending 7r to the sequence (wi : i < n + 1}
is a p - m o r p h i s m onto u~([, wo). It follows t h a t there is a p - m o r p h i s m from
~ ( [ , w0) onto the transit of wo in ~, namely the m a p sending each ~ - p a t h to
ints end point. In ~ ( f , w0) for any pair x and y of points there is at m o s t
one relation <~j such t h a t x <~j y. This is not so in u~([,w0). For practical
purposes the difference between these constructions is only marginal.
E x e r c i s e 95. Show by m e a n s of filtration t h a t K ~ has the global finite m o d e l

property.
Exercise 96. Generalize the method of unravelling to unravellings generated

by sets rather than points. T h a t is, form the frame consisting of paths starting
in a given set S.
Exercise 97. A forest is a disjoint union of trees. Show that every K r i p k e -

frame is the p - m o r p h i c image of a forest of intransitive, irreflexive trees.
3.4. Weakly Transitive Logics I

The notion of a weakly transitive logic plays a pivotal role in m o d a l logic.
Many theorems hold only for weakly transitive logics. In this section we will
collect some elementary facts about them. In Section 4.3 we will return to
that subject m a t t e r again. First recall that a logic is weakly transitive iff
there exists a maximal modality with respect to _~A, where _(h is defined by
[] --<A []~ iff •~p -+ • p C A (see Section 2.1). Moreover, we have seen in
T h e o r e m 3.2.9 that logics axiomatized by axioms of the form []p -+ []~p are
canonical and that the frame property determined by them is elementary.
PROPOSITION 3.4.1. Let A be a weakly transitive logic with master modal-
ity []. Then [] = [~s for some finite set of paths s and A is complete with
respect to frames satisfying
y C Tr(x) ca x <~s y.
PROOF. The first claim has been shown in Section 2.1. For the second
claim we show that the canonical frame for A satisfies the property. Suppose
that Y E T r ( X ) in ~anh(var). Then there exists a sequence a of numbers
< a such t h a t X <o y . Now []Sp -+ []~p E A. Thus, by T h e o r e m 3.2.9,
<~ C_ <~, which means that X <~s Y, as desired. []
PROPOSITION 3.4.2. The weakly transitive n-modal logics form a filter in

the lattice ~ K~. This filter is not principal. For a given compound modality
[] there exists a least logic such that [] is maximal with respect to ~ i .
PROOF. The first claim is straightforward. To see t h a t there is no least
weakly transitive logic observe that K ~ has the finite model property. So, if
( K~ there exists a finite ~ such that ~ ~ T. For some n, ~ is n-transitive.
Hence, T ~ K~.trsn. This shows that
K~ = N K~.trs~
new
However, K~ is not weakly transitive since we standardly assume ~ > 0.

Finally, let [] be given. Then put
A := K~ ® {[]p -+ []'p : []~ compound}
This is the smallest logic in which [] is maximal with respect to _~i. []
3.4. Weakly Transitive Logics I 123
PROPOSITION 3.4.3. Let A be a logic. A is weakly transitive iff f o r every

A-algebra and every principal open filter F of 94, F is principal as a boolean
filter.
PROOF. Suppose that A is weakly transitive with master modality []. Let
F be a principal open filter, generated by the element a. We claim t h a t ~ a
is the smallest element of F. To see this, observe that the least open filter
containing a set E is the least boolean filter containing the set E D := {~re :
e C E , [ ] r compound}. This is the algebraic analogue of Proposition 3.1.2.
Since [] is the master modality, we have []e _< []re for all []r. Hence, the open
filter generated by E is the boolean filter generated by [ ] E := {[]e : e C E}. In
particular, if E = {a} this shows that the open filter is principal as a boolean
filter. Now assume t h a t A is not weakly transitive. Then let 94 := ~A({P}).
The open filter generated by (the equivalence class of) p in 9.1 is not principal
as a boolean filter; otherwise it has a smallest element. This element is of the
form []p for some compound modality []. Then, for any compound modality
m r, []p _< • r p , which is the same as []p --+ []rp C A. Hence A is weakly
transitive, contrary to our assumption. []
The following is proved in [30] using algebraic methods.
THEOREM 3.4.4 (Blok &: Pigozzi). Let A be a modal logic. IF-A admits a
deduction theorem iff A is weakly transitive.
PROOF. Suppose I~-A admits a deduction theorem. Then there exists a

term p --- q such that for all sets of formulae A and formulae p and ¢
($) A; w I~-A¢ *~ A/~-A w -+ ¢
Now since p --- q IF-Ap --- q we deduce that p - , q;p If-A q. By T h e o r e m 3.1.2
there exists a compound modality [] such that [](p - - q); ~ p ~-A q. By the
Deduction Theorem for the local consequence of A, ~ ( p --- q) ~-A []P -+ q.
Now let []r be an arbitrary compound modality. Replacing q by []~p we get
[](p _, [~rp) ~-A []P --+ []'P- Notice now that [](p --- []'p) is a theorem of
A; for since p IF-A []rp we immediately get I~-A p --- E]rp, using (1:). And so
[](p __, ~rp) is a theorem as well. Hence we have ~-A []P --+ []~P. This shows
that A is weakly transitive. For the converse, assume A is weakly transitive.
Then there exists a compound modality [] such that []p --+ []~p E A for all
compound modalities []r. P u t p --- q := []p -+ q. We claim that (:~) holds
with respect to - - . For assume A; p I~-A~b. By Theorem 3.1.2 there exists a
[]r such that []'A; []'~o ~-A ¢. Since []T b A []t~p, we also have []'A; []~o ~-A ¢
and so m~A kA []T -~ ¢. Hence A Ikh []~ --+ ¢, as desired. The other
direction of (~) is straightforward. []
We close with the following useful observation. Given t h a t A is m - t r a n s i t i v e

and that we have finitely m a n y operators, then []p := []<rap for some m is a
master modality (though clearly not the only one). Hence if t~ < R0 a weakly
transitive logic is m - t r a n s i t i v e for some m. There is an analogue of the next
theorem without the assumption ~ < R0, but we leave the generalization to
the reader.
THEOREM 3.4.5. (n < R0.) If a modal logic is m-transitive then every

extension of A can be axiomatized by formulae of modal depth <_ m ÷ 1.
PROOF. Suppose that h is m - t r a n s i t i v e and let p be a formula. Take a
fresh variable q¢ for each subformula and define w as follows.
A A ( q ¢ ~ ++ q¢ A qx: ¢ A X 6 sf(~))
A A(qs ¢ Cjq¢: csj¢ c #@))
(Obviously, the variables q¢ must be pairwise distinct and distinct from the
variables of 9.) Consider a model (3,/3, w0) ~ []<mwA~q~. Let 3 be rooted at
w0. Then it follows that (3,/3) ~ w, since A is m-transitive. By induction on ¢
it is shown using L e m m a 3.1.7 t h a t (3,/3) ~ q¢ ++ ¢- Hence (3,/3, w0) b ~ .
Conversely, assume (3,')',x) ~ ~T. P u t "~(q¢) := ~ ( ¢ ) . Then (3,')',x)
N<-mwAq¢. Thus A O ~ = AGN-<'~w --+ q~, and we have dp([]<-'~w --+ q~) =
m+l. []
E x e r c i s e 98. Show that a weakly transitive logic is globally decidable iff

it is locally decidable. Likewise for globally complete and global finite model
property.
E x e r c i s e 99. Formulate and prove a version of Theorem 3.4.5 t h a t does not

restrict n to be finite.
3.5. S u b f r a m e L o g i c s
In [66], KIT FINE introduced the notion of a subframe logic for logics
extending K 4 and proved that all subframe logics have the finite model pro-
perty. This will be shown again in Chapter 8.3. In WOLTER [244] this
concept was extended to general logics and it was shown t h a t there exist
subframe logics without the finite model property. Nevertheless, subframe
logics have been established as an important tool in modal logic. T h e notion
of a subframe logic is based on the concept of a subframe as defined previously.
The algebraic concept corresponding to it is the notion of a relativization.
DEFINITION 3.5.1. Let 9.1 = (A, 1 , - , N , ( . j : j < ~)} and b C A. Put

Ab := {c : c <_ b} and 92b := (Ab, b , - , A , ( ] l ~ : j < ~)) where - is the
relative complement and m~c := b 71 I j ( b -+ c). An algebra f8 is called a
r e l a t i v i z a t i o n of 92 if 98 = 925 for some b E A.
3.5. Subframe Logics 125
DEFINITION 3.5.2. A logic is called a subframe logic if its class of frames

is closed under taking subframes. Alternatively, a logic is a subframe logic if
its class of algebras is closed under relativizations.
THEOREM 3.5.3 (Wolter). (~ < ~0") Every subSrame logic of bounded op-
erator alternative has the finite model property.
PROOF. E v e r y logic of b o u n d e d alternative is complete by T h e o r e m 3.2.12.
Hence if p !Z A t h e n there is a K r i p k e - f r a m e f such t h a t (~, x} ~ ~ for some x
and [ ~ A. Let d be the m o d a l depth of ~. T h e n ~ ( x ) ~ p. ~:~(x) is finite,
and by the fact t h a t A is a subframe logic it is also a frame for A. []
In C h a p t e r 8.3 we will show the subframe t h e o r e m of [66]. T h e p r o o f is

s o m e w h a t long and tedious. However, there are some restricted variants which
can be proved with less effort. We present one here. It is illustrative in the
sense t h a t it d e m o n s t r a t e s t h a t the subframe p r o p e r t y is not straightforward
in case we fail to know a b o u t completeness.
THEOREM 3.5.4 (Fine). Every subframe logic extending G has the finite
model property.
PROOF. Observe t h a t the G - a x i o m states t h a t for every set a and x C a
there is a m a x i m a l successor, t h a t is, a point y such t h a t x <l y E a but no
successor of y is in a. Furthermore, by transitivity, if u has a successor in
a then it has a m a x i m a l successor in a as well. T h e m a x i m a l points of a
can be defined by a N • - a . Let ~ • A. T h e n for n : = ~var(~) we have
(~anA(n),/3, wo) ~ -,~ for some /3 and x. Now let M be the set of points x
such t h a t x ~ ¢ for some 0 ¢ C X U {Qp} but x ~ ~ ¢ . M is an internal set
and so defines a subframe 9Yr. There exists a w* E M such t h a t w* ~ ~. For
if w0 ¢ M then w0 ~ 0 ~ and so w0 ~ 0 ( ~ A [2]-~). Hence there exists a w*
such t h a t w0 <~w* and w* ~ ~; [ 3 ~ . From this follows w* E M . Let "7 be the
restriction of/3 to M . T h e n (gYt, "7, w*) ~ ~. This is proved by showing t h a t
forallxCXandxEM
This holds for X = P by definition of % The steps for A and ~ are straight-
forward. Now let X = 0 ¢ . From left to right is immediate. Now as-
sume t h a t (¢anA(n),/3, x) ~ 0¢. T h e n there exists a y such t h a t x <~ y
and (ffanA(n),/3, y) > ¢ ; E J ~ b . It follows t h a t y C M ; by induction hy-
pothesis, (ffJt, 3,,y) ~ ¢. By assumption on A, 92I ~ A. 92~ is not nec-
essarily finite. However, ffJt D [ ] m £ for some m. For let x C M ; put
~(x) : = {0X E X : (giTt,%x) ~ 0X}. If x <~ y then t?(y) C I?(x). Let
m : = ~{0X : 0X E X } . T h e n 9Yt ~ Din_l_. By T h e o r e m 2.7.14, the set of inter-
nal sets is finite. By T h e o r e m 2.4.11 the refinement m a p is a p - m o r p h i s m . It
has a finite image. Hence A has the finite model property. []
Now we show that the satisfiability of a formula ~ on a subframe of ~ can

be translated into the satisfiability of another formula, which can be derived
syntactically from p.
qSx := qAx
(~A¢)$X :---- ((£$x) A('C'$X)

([3~)$X :---- xASj( X-+ (~$X))
We call T $ X the l o c a l i z a t i o n o f T to X.
LEMMA 3.5.5. Let ~ be a frame, ® = ~g. Let/3 : var --+ f and 7 : var --+ g
such that 13(p) = g and 7(q) = -~(q A p) for all q • p. Then -~(~ $ p) = "~(~).
PROOF. For a variable q, x e ~(q $ p) iff x E ~(q A p) iff x E /3(q) n 13(p)
iff x E g and x E 7(q). Further, x E ~ ( ( - , ~ ) $ p) iff x E ~ ( p A ~ ( ~ $ p)) iff
x E g and x 9~7(~) i f f x E g and z E 7 ( - ~ ) . The step for conjunction is
straightforward. Now we turn to [:]j.
c Z ( ( ~ w ) $ p)
i~ • c Z(pA ~j(p --~ (~ $p)))
i~ x c g and x ~ " 9 ( P -~ (~ $ P))
iff xEgandforeveryyEgsuchthatyDjx:yE~(~$p)
iff x E g and for every y E g such that y E>j x : y E 7 ( ~ )
iff xcT(Dj~)
This ends the proof. []
Consequently, p --+ (T $ p) holds in ~ iff T holds in all subframes of 5- Now

define p s / := p __+ (~ $ p), where p is a variable not occurring in ~. (For our
purposes it will not m a t t e r which variable gets chosen.)
THEOREM 3.5.6. Let A = K~ G A be a logic. The smallest subframe logic
containing A, A s/ is axiomatizable by Kn • ASf. The largest subframe logic
contained in A is equal to h~f = K~ @ { T s / : p s i E A}.
If follows immediately that a logic A is a subframe logic iff for every ~ E A
also ~ / E A.
COROLLARY 3.5.7 (Wotter). Let S F K~ denote the set of n-modal sub-
frame logics. This set forms a complete lattice with the operations of ~ K~.
The natural embedding of S~: K~ into ~ K~ commutes with infinite meets and
joins.
PROOF. Let A~, i E I, be a set of subframe logics. Let ~ be a frame and ®
a subframe o f ~ . Then i f ~ ~ I_Jlh~, then ~ ~ Ai for all i E I. By assumption,
® ~ Ai for all i E I , and so ¢5 ~ [_JiAi. This shows that the infinite join is a
subframe logic. Next assume that T E ~±A~. Then for all i E I , ~ E Ai. By
FIGURE 3.1
w+l w 3 2 1 0
assumption on the Ai, ~s/ E Ai for all i C I. Hence p s / E [-]xAi. Hence, the
infinite meet is a subframe logic. []
In WOLTER [234] an infinite series of incomplete subframe logic has been

constructed. The simplest is the following logic, which is actually one of
the earliest examples of an incomplete logic, taken from VAN BENTHEM [9].
Moreover, in CRESSWELL [48] it is shown that this logic is decidable, thus
showing that there exist decidable, but incomplete logics. CRESSWELL uses
RABIN's Theorem, but the result follows easily from Corollary 2.6.7. Take the
frame Yt~+l to be (w + 2, <~, @} with
a<~i3 iff { 13<a and t3, a < w + 2

13_<a and 13<a=w
For the algebra O of sets we take the smallest algebra of sets on t h a t frame.
This is an infinite algebra. We will approach its structure in stages. First
of all, take the generated subframe of finite numbers and let F be the trace
algebra on t h a t set. We claim that Y is nothing but the algebra of finite and
cofinite sets. To show this, two things are required. We have to show that
it is closed under all operations, and that each set is definable by a constant
formula. The first is not so hard. The finite and cofinite sets are closed
under intersection and complement. Furthermore, if a C w is an arbitrary
subset, let n be the largest number such that [0, n] C a; n exists iff a ~ w.
Then . a = [0, n + 1], as is readily checked. Thus, ma is finite whenever
a ~ w. Moreover, mw = w, which is cofinite. This shows the closure under
the operations. Now, let us define the following constant formulae.
f(O) := []~
f(n + 1) := Of(n). A.i~-~Of(n)
It is straightforward to verify that f(n) can only be true at n, so that all
singleton sets are 0-definable, that is, definable by means of a constant for-
mula. The smallest boolean algebra containing them is the algebra of finite
and confinite sets. Lets go one step further and take the subframe generated
by w. Here it turns out that the trace of @ contains all finite sets which do
not contain w, and all infinite sets which do contain w. For the extension of
the formulae f(n) is still {n}, n < w. The set of these sets is closed under the
operations, as is easily verified. Finally, we let us consider O. Observe that
the extension of the formula f(w + 1) is exactly {w + 1} where

f(~ + 1) := ~ 0 o ± .
This means that the full algebra consists of all sets whose trace relative to
the frame generated by w is either finite and does not contain w, or is infinite
and contains w.
We claim that Th ~+i is a subframe logic. To that end take an internal
set g in that frame. If it is finite, it does not contain the point w and the
trace algebra is the powerset algebra. Therefore, the frame is isomorphic to a
generated subframe. If g is infinite, however, it contains w, and the subframe
is isomorphic to either ~+i or the subframe generated by w. We conclude
that Th fl~+1 is a subframe logic.
THEOREM 3.5.8 (Wolter). Th ~ + 1 is an incomplete subframe logic.
PROOF. To begin, G . 3 is a subframe logic; it has the finite model prop-
erty and is therefore complete. Now, take the logic A with the following
axioms.
[~(0p -+ 0(p A ~0p)), 0p A 0q. -+ .0(p A 0q) V 0(q A 0p) V 0(P A q) .
Take a frame ~ for A, and a point x. Then - - by choice of the axioms - -
for every successor y of x the subframe generated by y satisfies G . 3 . ~ + 1
satisfies the axioms of A, since the subframe generated by w satisfies G . 3 .
Now consider the logic A ® ~ with
:= 5q A ~5(q/~ O~q) A 5p. -+ . 5 0 p .
This formula implies in conjunction with the other axioms that when we have
a violation of the G - a x i o m at a point x then x must have a reflexive successor.
(This successor can of course be x.) ~ C Th 9/~+1. Now, any K r i p k e - f r a m e for
A ® ~ must be a frame for G . 3 . For if we have a K r i p k e - f r a m e ~ for A and x
a point then either the transit of x is a G . 3 - f r a m e or only the transit without
x is. In the latter case x has a reflexive successor; let it be y. T h e n since
(~,x} ~ [ ] ( 0 p - + (}(pA-~0p)) and x<ly we have (~,y} ~ 0 p - + 0 ( p A ~ 0 p ) . This
enforces that the transit of y in ~ is a frame for G. But this cannot be, since
then we must have y ~ y. Contradiction. The proof is now almost complete.
First, we have
A ® ~ c_ Th ~,+~ C_G.3.
Equality of the last two cannot hold, because the formula 0pA ~0(pA-~0p)) is
satisfiable in ~,o+i. All logics in between AO~ and G.3 have the same Kripke-
frames. Hence, any such logic if not equal to G.3 is incomplete. Th ~2~+i is
such a logic. []
Now, even if subframe logics may be incomplete, in case of their com-

pleteness we can show t h a t they are complete with respect to frames of size
R0 if n < Ri and n if l~l _< n. This may not seem such an improvement. How-
ever, it is a priori not clear t h a t complete logics are complete with respect to
countable models even in the case n = 1. Secondly, the proof method itself is
well-worth remembering. It will be used in many variations throughout this
book. For extensions of K 4 this theorem has first been proved in KIT FINE
[66].
THEOREM 3.5.9. (~ < R0.) Let A be a subframe logic and suppose that A
is complete with respect to Kripke-frames. Then A is complete with respect
to Kripke-frames of cardinality < ~o.
PROOF. Let 7 9 ¢ A. Then there exists a A-model <~,~,wo} ~ 9, ~ a
Kripke-frame. P u t So := {w0}; let s0 be the subframe of [ based on So, and
let 70(P) := /3(p) N So. Inductively we define sets Sn; given Sn, ~ is the
subframe based on Sn and %~(p) := ~(p) N Sn. The construction of the Sn
is as follows. Suppose that there exists a [ ] j ¢ C sf(9) and a x E S~ such
that <[,~,x> ~ ~[~i¢ but <s~,%,x} ~ D i e . Then let y := y(x,[~y¢) C f be a
point such that x <lj y and <[,/3, y} ~ 7 ¢ . Then let
S +l := sn u { y ( x , [ ] j ¢ ) : x • sn, • sf(9),
Finally, we put g := U~c~ S~, 6(p) :=/~(p) N g. g is finite if S,+1 = S , for
some n, else g is countably infinite. We claim that for every x • g and every
¢ • sf(9)
The proof is by induction on ¢. For variables this is immediate; for ¢ = ¢1 A¢2

and ¢ = -~¢1 this is also immediate. Now assume finally that ¢ = [Zi~-.
Suppose that <~t,(~, x> 1~ [~j~-. Let n be the smallest number such t h a t x • S , .
Then, by construction, there exists a point y • Sn+I such t h a t x <~j y and
<f,/3, y> ~ ~7-. By induction hypothesis, <~t,5, y> ~ ~v. Since y • g we have
<9, 5, x> ~ ~[~j~-. The converse direction is straightforward. []
We give an example to show that completeness is necessary for the proof

method to work properly. Take the logic Th 12~+1 defined above. Let ~(p) be
a set containing the point w. Then <ft~+l,~,w / ~ p. It is clear that $1 = So
in the construction. However, the frame based on a single reflexive point is
not a frame for the logic. This shows that in the incomplete case we cannot
get rid of the restriction that the subframe is based on an internal set. Since
internal sets need not be countable, the proof methods fails in this case as it
stands.
Let us define the K u z n e t s o v - l n d e x , Kz(A), of a complete logic A to be
the supremum of the cardinalities of minimal frames refuting nontheorems of
A.
Kz(A) := supv¢hinf{~f : ~ 9}
The Kuznetsov-index is finite if A is tabular, and > ~o otherwise. We have

shown that if A is a complete subframe logic, the Kuznetsov-index does not
exceed ~0. In general, following CHAGROV and ZAKHARYASCHEVthe com-
plexity of a logic is a function fA such that fA(n) is the s u p r e m u m of the
cardinalities of minimal models refuting nontheorems of length n. If A has the
finite model property and n is finite then fA(n) is finite for every n. Clearly,
the Kuznetsov-index is the supremum of all fA(n). It is shown in CHAGROV
and ZAKHARYASCHEV [43] that there exists a logic with K u z n e t s o v - I n d e x ~ ,
where "G is the so-called beth-function; roughly, ~ is the A-fold iteration of
~he exponentiation function. In KRACHT [129], for each countable ordinal A a
logic with Kuznetsov-Index ] ~ is constructed. Furthermore, it is shown t h a t
there exists a logic whose Kuznetsov-Index is the least strongly inaccessible
cardinal.
E x e r c i s e 100. Show that $5 is a subframe logic.
E x e r c i s e 101. Show that G . 3 is a subframe logic.
E x e r c i s e 102. Show that A @ ~ = Th f ~ + l and that they are immediately

below G . 3 , i. e. there is no logic (~ with A • ~ C 0 C G.3.
* E x e r c i s e 103. Show that there is a descending chain of R0 m a n y incomplete

subframe logics. Hint. Extend the construction of fl~+l above to frames fl~
based on ordinal numbers ~ < w x ~o. You have to p u t / 3 <~7 iff 13 = ~ x k + m
and(a.) 7=wxk+nandm<nor (b.) 7 = w x ( k + l ) . Let the algebra of
sets be the minimal algebra. Now show that all the logics of ~ are different.
E x e r c i s e 104. Let ~ ___ ~0. Show that if a subframe logic is complete with
respect to Kripke-frames, it is complete with respect to Kripke-frames of car-
dinality _< n.
E x e r c i s e 105. Assume that ~ < ~1- Show t h a t a subframe logic is globally

complete with respect to Kripke-frames if it is globally complete with respect
to Kripke-frames which are finite or countabty infinite. Show the same for
~l-compactness.
E x e r c i s e 106. Let ~ be a frame, and S = Try(S). Show t h a t a ~-+ a A S is

a h o m o m o r p h i s m of (F, 1 , - , n, ( ' j : j < n)} onto the trace algebra over S.
Remark. This shows that in contrast to arbitrary subframes, the generated
subframes need not be based on internal sets.
E x e r c i s e 107. Let n be countable and A a canonical n - m o d a l logic. Show

that the Kuznetsov-index of h is _< 2 ~°.
3.6. Constructive Reduction 131
3.6. C o n s t r u c t i v e R e d u c t i o n
In Section 3.1 we have proved the global finite model property for the basic
logic K~. We will now use this proof to obtain a number of other results on
(global) finite model property using a technique which we call constructive
reduction. This technique is syntactic. The standard situation is that certain
properties have been established for a logic A, for example K , and t h a t we
consider an extension A G A for some set of axioms A. It would be rather
unfortunate not to be able to use knowledge about A for A ® A. However, in
the overwhelming number of cases nothing can be inferred for A ® A from A.
On the other hand, many standard systems are an exception to this. Before we
investigate the formal background of this method, let us see some nontrivial
applications.
THEOREM 3.6.1. Let A have the global finite model property and let X be
a constant .formula. Then A G X has the global finite model property as well.
PROOF. We show that
From right to left is trivial. From left to right, take a proof of ¢ from
in A ® X. We know that we can move substitutions at the beginning of the
proof. Now X is constant, so we cannot derive anything but X from X using
substitutions. Hence the proof is a proof in A of ¢ from ~ together with X
using (mn.) and (mp.), as required. []
THEOREM 3.6.2. K~ has the global finite model property.
PROOF. Let A be a set of formulae. P u t
X4(A) : : {[~]X --} F][~X : E]X E 4 [ t ] } .
We show that
V) IbK4¢ ¢* ~;X4({~,¢})I~K ¢
From right to left is straightforward. For the direction from left to right,
assume ~; X4({~, ¢}) I~-z ¢ is not the case. Then there exists a finite model
(f,/3,x} such that (f,/3) ~ ~a;X4({~,¢}) but (f,/3, x} ~ ¢. f = (,f,<@ Let •
be the transitive closure of <~. We show that for all subformulae ) / o f ~ or
and all worlds y
(t) (f, •,/3, y) ' x ¢* (,f, 4,/3, y) ~ x
This then establishes (,f, •,/3) ~ p and (,f, •,/3, x} ~ =¢. (,f, • ) is transitive;
therefore (f, • ) ~ K 4 . We show (t) by induction on X. For variables there
is nothing to show. The steps for ~ and A are straightforward. Now let
X = [:]Xq Assume (,f, • , ~ , y ) J~ []X ~. Then there is a z such that y • z and
(.f, •,/3, z} ~ ~X'. By induction hypothesis, (,f, <~,/~, z) ~ ~X'. By definition
of • there is a chain y = Y0 <~Yl 4 . . . <l Yn = z. Now (,f, <~,/3, Yn-~) ~ -~[~X'. If
n - 1 > 0 t h e n (f, 4,/3, y~-2) b ~[312X'. Since [3 X' -+ [2[2 X' ff X 4 ( { ~ , ¢ } ) and
(f, <,fl, y~-2) ~ X 4 ( { ~ , ¢ } ) we must have (f, <~,/3, Y~-2) ~ 912X'- Iterating
this argument we get (f, q,/3, y} ~ 9[2)/'. So, (f, <,/3, y) N [2X'. Clearly, if
(f, <,/3, Y) J~ [~X' then (f, 4, fl, y) N [2X', since 4 C ~. []
A note on the reduction sets. Since (} is not a primitive symbol, some care
is needed in the formulation of the reduction sets. The following definition of
a reduction set for K 4 will not do.
Y4(A) := {00X -+ 0)~: 0X E sf[A]}
(Here, (} abbreviates 9[29.) Take for example the set A = {-~[Z][2p,[2p}. It is
K4-consistent. Yet, there is no subformula of a formula of A that matches
(}X for some X. Hence, Y4(A) = ~. But A is clearly K-consistent. So, this
definition of the reduction sets does not work. The reader may pause to reflect
on why the chosen reduction sets actually avoid this problem.
THEOREM 3.6.3. The basic tense logic K . t has the global finite model
property.
PROOF. Let
x , ( a ) := { ~ ( -~ [~09[--11X : [~1)~ ~ 8f[A]}
u {gx -+ [2 9 ox : [2ox d i a l } .
We show that
(t)
Proceed as in the previous proof. Let 93t = (f,/3, wo) be a local model where
= (f, 4o, <h) is a finite K2-Kripke-frame such that (f,/3) ~ V); Xt({~, f2})
and ([,fl, wo) ~ 9¢. Let 40 := 40 U q 5 and " 1 := 411._J 4 0 . Then the frame
(f, 40, "1) is a tense frame, for 4 o = (q0 U 4~') ~ = <1o U <h ----411. For all
X C sf(~) U sf(@) we have
(t) (/, x (/,<o,41,fl, v) x.
This is clear for variables; the steps for -~ and A are straightforward. Now let
X -- DoT. From left to right is clear. Now the direction from right to left.
Assume that (f, "o, "~l,fi, Y) t~ []o7. Then there is a w such that y ~o w and
(f, ao, al,/3, w} ~ 97. By induction hypothesis, (f, 4o, <~1,/3, w} ~ ~ r . If
y <SOW, we are done; for then (f, <o, 41,/3, y} t~ E]or (= X). Otherwise w 41 y.
Now, (f, 40, <:~1,/3,w) ~ [~l-n[~07, since (f, 40, 41,/3, w) ~ X t ( { ~ , ¢ } ) . Thus
(f, q0, q l , y } ~ -~D0r. So, (f, qo,4~,/3, y) )~ [20"5. The step X = [217- is
analogous. []
Informally, we say that a property ~ pushes up or can be pushed up from
A to A (3 A if we can prove that A ® A has ~ on the condition that A has
~. Properties that can be dealt with in this way are among other decidabil-
ity, finite model property, completeness and interpolation. The fundamental
property in this connection is decidability. Notice that given A, A, A and ~,

there is a set Y C []wAS such that
(Recall that A s denotes the closure of A under substitution.) We call Y a

local r e d u c t i o n s e t for A and ~. Moreover, if A is finite then Y can be
chosen finite. A l o c a l r e d u c t i o n f u n c t i o n for A @ A w i t h r e s p e c t t o A is
a function X : ~([P~) -+ ~v(iP~) such that (i) X ( A U {~}) is a local reduction
set for A and ~ and (ii) X ( A ) is finite whenever A is finite. We will write
X ( ~ ) rather t h a n X({T}). Note that there are two cases. (a.) ~ ¢ A ® A.
Then Y = ~ is a reduction set, (b.) T E A G A. Then there is a proof of
from A in []~A ~ using only modus ponens. Obviously, this proof uses only a
finite subset of []~A ~, and we take Y to be this subset. Similarly, there is a
finite set Y C A s such that
A r~-A®A ~ ¢=} A ; Y I~-A ~fl
Such a Y is called a g l o b a l r e d u c t i o n s e t for A and ~. A g l o b a l r e d u c t i o n
f u n c t i o n f o r A @ A w i t h r e s p e c t t o A is a function X : go(T~) -+ gv(iP.)
such that (i) X ( A U (~}) is a global reduction set for A and ~ and (ii) X ( A )
is finite whenever A is finite. We note the following properties of reduction
sets. The proof is left as an exercise.
PROPOSITION 3.6.4. (1.) There exists a reduction function X for A @ A

with respect to A such that (a.) var[X(A)] C_ var[A], (b.) X ( A ) = U ( X ( A 0 ) :
A0 C A, A0 finite} (2.) Let X , Y : ~ ( ~ ) -+ ~(~,~) be functions mapping
finite sets to finite sets such that X ( A ) C_ Y(A) for all A. Then if X is a
(global/local) reduction function for A • A with respect to A then so is Y.
As (lb.) shows, we may always assume that the reduction function is
determined by its values on finite sets. This means that we may actually
restrict our attention to functions from finite subsets of T~ to finite subsets of
T~. If A@A is decidable and A enumerable, a local reduction set can always be
constructed. For suppose that ~ is given. Then start enumerating the proofs
of A @ A in which (sub.) is applied before (mn.) and (mn.) before (mp.); in
parallel, enumerate the nontheorems of A ® A. If ~ is a theorem, it will occur
at end of a proof II. The reduction set will then consist in all formulae to
which only (mp.) is applied in H. (This is more t h a n necessary, but certainly
a sufficient set.) If 9~ is a nontheorem, then the e m p t y set is a reduction set
for T. Similarly, if A ® A is globally decidable, it allows the construction of
global reduction sets. Conversely, suppose we are able to produce for each
a local reduction set. Then decidability can be pushed up. To see the first,
assume A is decidable and let ~ be given. Construct X ( { ~ } ) . Since A is
decidable, we can decide X({~}) F A ~ , which by definition of the reduction
sets is nothing but kAeA T. Similarly for global decidability. For the purpose
of the next theorem, a computable function from p(9~) to p(~P~) is a function

f such t h a t (i.) f ( A ) = U ( f ( A 0 ) : Ao c A, A0 finite} and (ii.) there exists
an algorithm computing f ( A ) for any given finite A.
DEFINITION 3.6.5. A logic A ® A is said to be locally c o n s t r u c t i v e l y

reducible to A if there is a computable local reduction function for A G A and
A. A G A is said to be globally c o n s t r u c t i v e l y reducible to A if there is a
computable global reduction function.
THEOREM 3.6.6. Suppose A • A is globally (locally) constructively re-
ducible to A. Then A 0 A is globally (locally) decidable if A is.
In m a n y cases, it is possible to show that other properties can be pushed
up as well. For example, for transitivity we have established t h a t X4 : A ~-+
{[]X -+ [][3X : [IX 6 sf[A]} is a global reduction function, and t h a t fur-
thermore any frame satisfying both ~ locally and X4({~}) globally satisfies
also when the relation q is replaced by the transitive closure. It then is a
K 4 - f r a m e and a finite transitive model for ~. So if A is a m o n o m o d a l logic
whose frames are closed under passing from q to its transitive closure then
we can push up the global finite model property from A to A.4. We will show
here that m a n y of the standard systems mentioned in Section 2.5 have global
finite model property. The following are global reduction functions.
X4(LX) := {[]X -+ [][]X: []X 6 sfgX]}
Xr(a) :: {[]X -~ x : SX c , f i l l }
Xs(a) :: {~X-~ D~DX:~X • #[a]}
Xo(A) :: {~[]X-~ ~(xV ~x): SX • #[ZX]}
XGrz(LX) := { ~ X -~ ~D(X V < ~ ( X -~ D X ) ) : DX • #[ZX]}
The reader may check that the formulae are indeed axioms. The reduction
of A.4 to A has been proved for those logics whose class of frames is closed
under passing from <~ to the transitive closure. Now, for reflexivity we claim
that if the class of frames for A is closed under passing from q to its reflexive
closure, denoted by q°, then the above function achieves global reduction.
Namely, suppose that we have a A-frame [ and {~,/3} ~ XT({~b; ~o}). Let [" be
obtained by changing < to its reflexive closure. By definition, [" ~ A and so
~" ~ A.T. By induction on the set sf(~o) we show t h a t for all w in the transit
of x
The only critical step is X = D r . From left to right this follows from the
fact t h a t if x q y then also x <~" y. For the other direction, assume we have
{,',/3, w} > [2r. Then there is a v such that w q " v and (~',/3, v} > -~r. If
v ~ w, we are done for then also w q v. So assume the only choice for v is
v = w and that w ~ w. Then we have (~,/~, w} > [~r. But (~,/3, w} ~ [BT --+ r ,
by choice of the reduction function. Hence <f, ~, w> b 7-, and so (~', ~, w> ~ T,
a contradiction. So there always is a successor v ~ w, and it is safe to close
reflexively.
The proof for B is the same as for tense logic, so we will omit it here. We
can use this technique iteratively to show that a logic defined by a mixture
of reflexivity, transitivity or symmetry axioms has the finite model property.
However, since each particular pushing up has its preconditions, some care is
called for. The idea is always the following. Assume A has the global finite
model property; then constructively reduce A • ~ to A. This works if we can
be sure that the procedure that turns a frame ~ for K~ into a frame ~ for
K~Op also turns a frame for A into a frame for A(9~. The proof of the
following theorem illustrates this.
THEOREM 3.6.7. Let A be a finitely axiomatizable polymodal logic based

on postulates of reflexivity, transitivity and symmetry for its operators, in any
combination. Then A has the global finite model property.
PROOF. Let A = K,~®R®SGT, where R is a set of reflexivity postulates,

S a set of symmetry postulates and T a set of transitivity postulates. First
of all K~ has the global finite model property, so we need to consider finite
Kripke-frames only. We will start with K~ and add the postulates one by
one. First, we add all ~ E R. The map ~ ~-~ f~ is defined by taking the
reflexive closure of the relation <j for some j. Since each of the reflexivity
postulates concerns a different operator, it does not matter in which order we
add the axioms. In the end we obtain a frame ® satisfying all ~ E R. Next
we turn to S. The map [ ~+ [~, ~ C S, is now the map which turns <~ into its
symmetric closure. Since the symmetric closure of a reflexive relation is again
reflexive, [~ satisfies all postulates of R. Moreover, the symmetric closure
of one relation does not interfere with any other relation, so f~ satisfies all
symmetry postulates that ~ satisfies. Thus we can construct a frame for RO S.
Now we turn to T; the transitive closure of a reflexive relation is reflexive,
and the transitive closure of a symmetric relation is again symmetric. []
There remain the sets for G, G r z and altl. Now, both G and G r z are
transitive logics. (This is the content of some exercises in Section 2.5.) We will
now show that the functions above establish a reduction from G to K 4 and
a reduction from G r z to S4. The first of these has been shown by PHILIPPE
I3ALBIANI and ANDREAS HERZIG in [3].
THEOREM 3.6.8. G has the global finite model property.
PROOF. We establish that the reduction function is a reduction function

[rom the logic to K4, which has global finite model property by Theorem 3.6.2.
Moreover, K 4 is transitive, so we only need to consider reductions where the
antecedent is identical to T. Thus let f be a finite transitive frame and

0:,/3, Wo) ~ ~;~S<-~{~DX --~ < S ( x V ~ X ) : CSX • ~S(~)} •
Now, pick points from the frame as follows. P u t So := {Wo}. The sets Sn are
now defined inductively. Let x C Sn and [3X E sf(T) such that (f,/3, x} ]~ DX,
and no successor of x in S,~ - {w0} exists such that ([,/3, y} ~ ~X. Then, by
assumption on the reduction function, ([,/3, x} ~ (>(~X A [~X). Hence there
exists a ~ such that ~ ~ ~X; [3X. (Moreover, if x = wo, then ~ ~ Wo. For ~ is
irreflexive, and so wo q ~ implies wo ~ 5.) It follows that ~ is irrefiexive. P u t
S,~+1 := S~ U {~}. The selection ends after some steps, since f is finite. Call
the resulting set g. Let <3g := <3N (g x g) - {(Wo, Wo}}. Then put 0 := (g, qg)-
(Alternatively, we might simply take 9 to be the subframe consisting of Wo and
all irreftexive points from ~, with the transition w0 --+ wo being removed.) 0 is
transitive and irreflexive, hence it is a frame for G. P u t 7(P) :=/3(P) N g. We
now show that for every subformula ¢ of p and every point y E g, /0, 7, Y} ~
iff if,/3, Y} ~ ~- This holds for variables by construction, and the steps 7, A are
straightforward. Now let (9, 7, Y} ta [Z]X. Then also (~,/3, y} ta []X. Conversely,
suppose t h a t if,/3, Y} ~ E]X, for some ~X C sf(~). Then also (9,/3, Y} )L DX,
since a successor z for y has been chosen such that (f,/3, z} ~ X; Z]~X. By
induction hypothesis, (~,7, z} ~ X. Moreover, y <~g z. For if y ~ w0 this
holds by definition of qg. For y = w0 observe that either wo qS w0, and then
z ~ w0, since z ~ z. From this follows w0 qg z. Or else, w0 ~S w0, in which
case wo <~g z anyway. And so (0, 7, Y} ~ []X, as required. []
THEOREM 3.6.9. Grz has the global finite model property.

PROOF. As in the previous proof, this time reducing to S4. By Theo-
rem 3.6.7, S4 has the (global) finite model property. Let (~,/3, w0} a finite
S 4 - m o d e l such that
We select a subset g of f in the following way. We start with So := {w0}.

S~+1 is defined inductively as follows. If x E S~ and (~,/3, x} ~ -~[:]X, but no
y exists in Sn such that x < y and ([,/3, y} ~ -'X, then we choose a successor
of x as follows. By choice of the reduction function there is a successor
y of x such t h a t y ~ ~X;[~(X --+ DX). Therefore (i) the entire cluster C(y)
satisfies ~X, (ii) no point in a cluster succeeding C(y) and different from C(y)
satisfies X. T h e n S~+1 := S~ U {y}. This procedure comes to a halt after
finitely m a n y steps. The resulting set is called g, and the subframe based on
it ~. It is directly verified that g contains at most one point from each cluster.
(Moreover, the selection procedure produces a model whose depth is bounded
by the number of formulae in sf(~) of the form ~X as can easily be seen.)
So all clusters have size 1. 9 is reflexive and transitive, being a subframe of
[. So, 9 is a G r z - f r a m e . Let 7(P) :=/3(p) N g. It is shown as in the previous
proof t h a t for every subformula X of ~a and every x E g, (g, 7, x/ ~ X exactly

when (f,/3, x / ~ X- In particular, (~,7, w0) ~ ~. This concludes the proof. []
Obviously, to have local reduction functions is much stronger t h a n to have

global reduction functions. Yet, for practical purposes it is enough to com-
pute global reduction functions, since most standard systems are globally
decidable. Thus, the additional gain in establishing a local reduction, which
is possible in m a n y cases, is rather marginal. Moreover, as we will see, m a n y
properties can be pushed up even when we have global reduction functions
and no local functions. We shall end the section by a few remarks of complexi-
ty. We shall state here without proof that the size of the reduction sets of this
section is quadratic in the size of the initial set (if the size is simply the sum
of the lengths of the formulae contained in it). This is actually easy to verify.
However, if we change to the packed representation (see the exercises of Sec-
tion 1.8) then the increase is only linear. To verify this is left as an exercise.
It follows t h a t the logics discussed in this section are globally E X P T I M E ,
since K is. One has to take note here that the typical complexity measures
are established with respect to the length of the set, not with respect to the
length of the packed representation, which can in extreme cases be exponen-
tially smaller. Yet, they can typically be redone with respect to the length of
the packed representation. To verify that K is globally E X P T I M E even with
respect to the packed representation, tableaux can be used. Moreover, using
tableaux one can show that K 4 is globally PSPACE from which the same
follows for the systems extending K 4 . Unfortunately, reduction sets do not
allow to show that K 4 is in PSPACE.
E x e r c i s e 108. Show Proposition 3.6.4.
E x e r c i s e 109. Show that the symmetric closure of a transitive relation does

not need not be transitive again.
E x e r c i s e 110. In the next three exercises we will show a rather general

theorem on reduction of A.G and h . G r z to A for logics containing K 4 . Let
A _D K 4 have finite model property. Say that a A is a cofinal subframe logic
if it is closed under taking away from a finite frame any set of points which
is not final. Here a point x is final if for all y x ~ y implies y ~ x. Now let
(f,/3/ and ~ be given, and [ be finite. Say that x is u - m a x i m a l if for some
subformula X, X is satisfied at x and whenever y satisfies X and x <1 y, then
also y <~x. Show now that if (f,/3, x / ~ ~, and if we take the subframe g of all
~,a-maximal points, then (~,/3, y) ~ ~ for some y. Show t h a t every final point
of f is in ~. Thus, if A is a cofina] subframe logic, and f is a finite frame for
A, so is ft. Give the global reduction sets!
E x e r c i s e 111. (Continuing the previous exercise.) Now show t h a t the re-

duction function given above for G establishes that if A is a cofinal subframe
logic, then there exists a global reduction function for A.G to A.
E x e r c i s e 112. (Continuing the previous exercise.) Show t h a t the reduction

function for G r z establishes t h a t there is a reduction function any A . G r z to
A, provided that A is a subframe logic.
E x e r c i s e 113. Using the L e m m a 3.1.9, produce local reduction sets for the
logics for which global reduction sets have been given.
E x e r c i s e 114. Let X be any of the reduction functions of this section. Show

that there is a constant cx such t h a t for every set A: I X ( A ) •] < clA*],
where A $ is the packed representation of A.
3.7. I n t e r p o l a t i o n and B e t h T h e o r e m s
Recall from Section 1.6 the definition of interpolation. Interpolation is
defined with respect to the consequence relation. Since a modal logic admits
several consequence relations, we have several notions of interpolation, in
particular global and local interpolation.
DEFINITION 3.7.1. A modal logic A has local i n t e r p o l a t i o n if for every
pair ~ and ¢ of formulae with ~ F-A ¢ there is a X such that var(x ) C_
var(~) N var(¢) and ~ F-A X as well as X F-A ¢. A has global interpolation
if for every pair ~, ¢ of formulae with ~ l~-A ¢ there is a X such that var(x ) C
var(~) N vat(C) and ~ I~A X as well as X IF-A¢.
These definitions are taken from [151], though the terminology used here
is more systematic. Since we have a deduction theorem for local deducibility,
we can reformulate local interpolation in such a way that it depends only on
the set of theorems. A has the C r a i g I n t e r p o l a t i o n P r o p e r t y if whenever
--+ ¢ C A there exists a X which is based on the common variables of ~ and
¢ such that ~ --+ X; X --+ ¢ C A. A logic has the Craig Interpolation P r o p e r t y
iff it has local interpolation.
PROPOSITION 3.7.2. If A has local interpolation it also has global inter-
polation.
PROOF. Suppose that A has local interpolation. Let ~ I~-A ¢. T h e n for
some compound modality [] we have ~ ~-A ¢- Whence by local interpolation
there is a X with var(x) C_ var(~) N car(C) such that [ ~ ~-A X and X ~-A C-
Hence ~ I~-AX as well as X IF-A¢. []
The converse implication does not hold, as has been shown in [151]. Inter-
polation is closely connected with the so-called Beth property. It says, in
3.7. Interpolation and Beth Theorems 139
intuitive terms, that if we have defined p implicitly, then there also is an ex-
plicit definition of p. An explicit definition is a statement of the form X e+ p
where p ¢ var(x). An implicit definition is a formula ¢(p, q~, such that the
value of p in a model is uniquely defined by the values of the variables q. The
latter can be reformulated syntactically. In a logic f-, ~(p, q~ i m p l i c i t l y de-
fines p if ~(p, q0; ~(r, q0 ~- P e+ r. Given A, we may choose ~- to be either ~-h
or rF-h. This gives rise to the notions of local and g l o b a l implicit definitions.
DEFINITION 3.7.3. A is said to have the local B e t h P r o p e r t y if the
following holds. Suppose ~(p, q~) is a formula and
~(p, q~; ~(r, q~) F-A p ++ r.
Then there exists a formula X ( ~ not containing p as a variable such that
q3 p .
Analogously, the global B e t h p r o p e r t y is defined by replacing ~-A by I~-A.
The notion of definability was introduced by BETH in [16] under the name
Padoa's Method. The lack of the deduction theorem for the global consequence
makes the global Beth property somewhat more difficult to handle than the
local equivalent. For the local Beth property we can actually prove that it is
equivalent to the Craig Interpolation Property.
THEOREM 3.7.4 (Maksimova). Let A be a classical modal logic. Then A
has local interpolation iff it has the local Beth property.
PROOF. Suppose first that A has local interpolation. Assume that
defines q implicitly, that is,
(t) ~(p, ~ ; ~(r, ~ ?A P ++ r.
Then we also have ~(p, q~;p ~-A ~(r, q~) --+ r and thus by interpolation there is
a x(q~) such that
We claim that X(~ is the desired explicit definition, that is, that the following
holds.
q3 p ++ x(qg.
One implication holds by definition of the interpolant; for ~(p, q~;p ~-A X ( ~ .
For the other direction, observe that we have x(q") ~-A ~ ( r , ~ --+ r. Using
the deduction theorem we can derive ~(p,q-) ~-A x(q~) -+ r. Now replace r
by p, and the desired conclusion follows. Now for the converse, assume that
A has the local Beth Property. We will show that if ~(p, q3 ~-h ¢(r, q3 then
there is a x(q~ such that ~(p, q") ~-A x(q~ f-A ¢(r, q-). Let us call this the
1-interpolation property, since we can get rid of a single variable in p and a
single variable in ¢. The n-interpolation property is formulated similarly, but
with the difference that we can eliminate up to n variables in the premiss and
up to n in the conclusion. One can easily show that 1-interpolation property

implies n-interpolation for every n, and hence the local interpolation property.
We leave this to the reader. The hard part is to show 1-interpolation. Thus,
assume ~(p, q~ ?A ¢(r, ~. Define
5z(p,~ := (p -+ ~(;,@ A (¢(;,~ -+ ;).
Since we have 51(p,q~;51(r,q~ ~A P ++ r, we get a formula x1(q~ such that
We then have
x~(@ bA p v ¢(p,@.
Now define
52(;,~ := (~p -+ ~ ( p , @ A ( ¢ ( ; , ~ -+ ~p).
Again, it is checked that 52 is an implicit definition and so we can use Beth's
property again to get a x2(q-) with ~(p,q~ F-A p ++ x2(q~. After some boolean
rewriting
Next we define the following formula
~3(;, ~ := (p -+ ~ ( ~ p , ~ ) A (¢(p, q3 -+p).

We have ~ ( ~ p , ~ ~-h ~b(r,~ and so it is checked that 53(p,q-') also provides
an implicit definition of p. And so we get a third formula x3(q~), such that
53(p, ~ ~-A P ++ X3(~. From this we get
~o(~p,q~;p ~-A x3(q~, x3(q~ ~-A p V ¢(p,q-).
Substituting ~p for p in the first statement we get
Finally, define
54(p, ~ := (~p -~ ~(~p, q3) A (¢(p, q3 -~ ~p)
This defines p implicitly, and so we have a x4(q-') with 54(p, q~ ~-A P ~ x4(q-).
We get after rewriting
Substituting ~p for p in the first statement we get
~(P, q~;P ~-A -~x4(q~)•

The desired interpolant is
x(~ := (x~(~ A ~x~(@ v (-~x~@ A x~(~).
Namely, T(p,q~;p ~-h xl(q-) A ~x4(q-) and SO ~Y(p,q~); "Tp [-A ~x2(q~ A x4(q-), SO
that p(p, q3 kA X(~- Furthermore, X l ( ~ A-~x4(q") kA ¢(p, q~ and in addition
~x2(q-~ A x3(q~ [-A ¢(P, q-), from which x(q-) ~-A ¢(P, q~, as required. []
THEOREM 3.7.5. A classical modal logic with local interpolation also has
the global Beth-property.
PROOF. Assume that ~(p, q-); p(r, q~ Ikh p ++ r. Then for some compound
modality [] we have
q3; q3 FA p r.
This can now be rearranged to
[]~(P, q3;P FA []~(r,q3 -+ r.
We get an interpolant x(q~ and so we have []~(p,q~;p F-A x(q-), from which
[]9)(P,q-) ~-A P -+ x(q~) • So ~(p,q-) ~-A P -+ x(q~. And we moreover have
x(q-) b-h []T(r,q~ -+ r, from which we get []~(r,q~ ~-i x(q~ -+ r, and so
p(r, q~) [-A x(q~) -+ r. Replacing r by p we get the desired result. []
The picture obtained thus far is the following.
local interpolation . global interpolation
1
local Beth Property , global Beth Property
It can be shown that there exist logics without global interpolation while
having the global Beth Property and that there exist logics with global inter-
polation without the global Beth Property. An example of the first kind is
the logic G.3. See [150].
Recall from Section 1.6 the notion of Halld6n-completeness of a logic.
As with interpolation the notion of Halld6n-completeness of A splits into (at
least) two different concepts.
DEFINITION 3.7.6. Let A be a modal logic. A is locally H a l l d d n - c o m p l e t e
if whenever ~ ~-A ¢ and var(~) N var(¢) = ~ we have ~ F-A ± or [-A ¢. A is
globally t t a l l d d n - c o m p l e t e if whenever ~ IkA ¢ and var(~) N var(¢) =
we have ~ IkA 5_ or I~-A¢.
Global Halld6n-comp]eteness is called the P s e u d o R e l e v a n c e P r o p -
e r t y in [153]. In the literature, a logic A is called Hallddn-complete if for
and ¢ disjoint in variables, i f ~ V ¢ E A then ~ E A or ¢ C A. Clearly, this lat-
ter notion of Halld6n-completeness coincides with local Halld~n-completeness.
This follows from the deduction theorem, since ~ FA ¢ is equivalent to
>A ~ a -+ ~b. Local (global) Halld6n-completeness nearly follows from the cor-
responding interpolation property. Namely, if we have >A TV~b then ~ T >A ~b.
We then get a constant formula X such that -~a >A X and X >A ~P. On the con-
dition t h a t we can choose X to be either T or 3_ we get our desired conclusion.
For X = T yields ~-A ~b and X = I yields >A ~. Thus, if A has trivial constants
(see Section 2.6) then interpolation implies Halld6n-completeness. But this
is exactly the condition we need anyway to have Halld@n-completeness. For
notice t h a t always >A ~[Z]j_LV [ ] j ± . So if A is Halld6n-complete then we have
either >A ~ [ ~ j ± or >A [~j_l_. These are exactly the conditions under which A
has trivial constants.
PROPOSITION 3.7.7. A logic A is (locally/globally) Hallddn-complete only
if it has trivial constants. If A has trivial constants and has (local/global)
interpolation then it is (locally/globally) Hallddn-complete.
Finally, we will establish some criteria for interpolation. Assume t h a t we
have a logic A ® A and global reduction sets for A @ A with respect to A.
Let us say t h a t the reduction sets s p l i t if there exists a reduction function
X such t h a t (i.) for all sets A, var[X(A)] C var[A] and (ii.) X ( ~ --+ 9) =
x(~) u x ( ¢ ) .
THEOREM 3.7.8. Suppose that A ® A can be globally reduced to A with
splitting reduction sets. Then A • A has local (global) interpolation if A has
local (global) interpolation. Moreover, A • A is locally (globally) Hallddn-
complete if A is.
PROOF. Assume qa >AeA ~b. Then >Ae,* ~ --+ ~b and a f o r t i o r i I}-ne,*
~a -+ ~b. By global reduction we get X(9~ --+ ~b) I~-A ~ --+ ~b and so for some
compound modality []
[]x(~ -+ 9) >A ~ -+ 9.
This is the same as
[]x(~); []x(¢) >A ~ -+ ¢,
by the fact that the reduction sets split. We can rearrange this into
[ ] X ( ~ ) ; ~ >A [ ] X ( ¢ ) -~ ~.
By assumption on X , var[X(~o)] C_ var(~o) and var[X(~)] C_ var(¢). By local
interpolation for A we obtain a X in the common variables of ~o and ~b such
that
w; ~x(w) >A X >A []x(¢) ~ 9.
From this follows that ~o }-ae,* X F-AeA !b, by the fact t h a t the reduction sets
only contain instances of theorems. Pushing up global interpolation works
essentially in the same way. Now for Halld@n-completeness, assume t h a t
~o >Re,* ~b for ~o and ~b disjoint in variables. Then
~; ~ x ( ~ ) >A r e x ( t ) + 9.
The left hand side is disjoint in variables from the right hand side, and so
either the left hand side is inconsistent or the right hand side a theorem. In
the first case, W ~-AeA d_. In the second case t-AeA ¢, as required. The proof
for global Halld6n-completeness is analogous. []
In the next section we will prove that K has local interpolation. We conclude
from that the following theorem.
COROLLARY 3.7.9. The logics K.altl, K4, K . B , K . T , K . B T , S4, $5,

G and Grz have local interpolation.
This can be generalized to polymodal logics as well, namely to those which

have no interaction postulates for the operators, and whose logical fragments
for the individual operators is one of the above logics. This, however, will be
a consequence of a far more general result on so-called independent fusions to
be developed in Chapter 6. As an application we will prove a rather famous
theorem, the so-called Fixed Point-Theorem of SAMBIN and DE JONGH (see
[184]; for the history see also [32]). It is a theorem about the provability logic
G. We say that a formula ¢(q~) is a f i x e d p o i n t of ~(p, q~ with respect to p
in a logic A if
¢(3 3.
THEOREM 3.7.10 (Sambin, de Jongh). Let 9~(P,q~ be a formula such that

every occurrence of p is modalized. Then T(p, q~ has a fixed point for p in G.
PROOF. Consider a formula ~(p, q~) in which the sentence letter p occurs
only modalized, t h a t is, in the scope of an D. We show that for every finite
Kripke-frame and valuation fl on ~' there exists one and only one extension
•+ such that ([, fl+) ~ p ++ ~(p, q~. To see this take a finite K r i p k e - f r a m e f.
The accessibility relation is transitive and cycle-free, t h a t is, irrefiexive, fl+
will be defined by induction on the depth of a point, that is, the length of a
maximal ascending chain starting at that point. To start, consider a point
x without successors. Then, since p occurs only modalized, ~(p, q~) holds at
x iff 9~(±, q~ holds at x; so the value of p at x is well-defined and does not
depend on p. P u t x C fl+(p) iff z C fl+(9~(J-,q~). Now assume that the claim
has been established for points of depth d. Let x be of depth d ÷ 1. We have
to show t h a t x C fl+(p) iff x C fl+(9~(P,q~). We can regard ~ as a boolean
combination of formulae not containing p and formulae of the form [~X. Since
the value of p is already fixed on points of lesser depth, we know whether or
not x C fl+(DX); also, the value of formulae not containing p is fixed at x.
Hence there is a unique way to assign p at x. This completes the proof of the
existence and uniqueness of fl+.
The first consequence is that because the extension is unique on finite

frames, we have
p +~ ~(p, q3; r ++ ~(r, q-) I~G p ++ r .
Taking ~I(P, q~ := P ++ ~(p, q~ we now have a global implicit definition of p.
Since G has local interpolation it has the global B e t h - p r o p e r t y and so there
exists an explicit definition, that is, a ¢(q~ such that
From this we deduce that

p ~ ~(p, q~ I~G ¢(q~ ~ ~(¢(q3, q~.
(Simply replace p by ¢(qO).) We claim now that in fact
~G ¢ ( ~ ~ ~(¢(q3, q~.
To see this, take a finite flame f for G and a valuation of (. Then there exists
an extension •+ of/3 giving a value to p such that (f, ~3+) ~ p ++ ~(p, q-). In
~his model we have (~,f~+} ~ ~b(q~ ++ ~(¢(q~,q~). But then also ([,j3) ~ ¢(q-) ++
~(%b(q~, q-), as required. []
The proof via the Beth-property has first been given by SMORYI~SKI [200].
It is worthwile to note that many direct proofs of this theorem have been
given in the past, e. g. SAMBIN [184], [188], REIDHAAR--OLsON [175] and
GLEIT and GOLDFAaB [76]. The difficulty of these proofs varies. This proof
via the B e t h - p r o p e r t y reduces the problem to that of the interpolation of
G modulo the Theorem 3.7.5. The difficulty that most proofs face is that
the construction of the interpolant is not disentangled from the actual fixed
point property. For notice that the special property of G is that for formulae
~(p, @ in which p occurs only modalized, the fixed point equation p ++ ~(p, q~
can be globally satisfied in one and only one way for given (~. This is proved
semantically. The uniqueness of the solution then gives rise to an explicit
definition because of the Beth-property. T h a t means that the solution is
effabIe. The existence of a solution then allows to deduce that this explicit
definition is a fixed point of ~. Note that the previous proof did not depend on
G, only on some critical features of G. It trivially also holds for all extensions
of G. It has been observed by MAKSIMOVA that this can be used to show that
all extensions of G have the global B e t h - p r o p e r t y (see [149]). To show this,
two auxiliary facts must be established, which we shall give as exercises. Call
a formula y) @ b o x e d if every occurrence of a variable from ~ is in the scope
of some modal operator.
LEMMA 3.7.11. Let A be a logic containing G. Let qi, i < n, be distinct
variables and p a variable not contained in ~. For a set S C n define Xs by
XS := A l e s qi A AiEn-S ~qi. If ~(p, q~ is ~-boxed and
A ~- x s -~ ~(p,
then already A ~- ~(p, q~.

LEMMA 3.7.12. Let p(p,q~ be a formula. Then there exist g-boxed for-
mulae ¢1(P, q~, ¢2(P, q~), XI(P, q~) and X2(P, q~ such that
c ~ ~(p, qO ++ ((p v ¢1(p, q-))A (~pv ¢~(p, q-)))
c e ~(p,~ ++ ((p A x l ( p , ~ ) v (~p A x~(p,q-)))
THEOREM 3.7.13 (Maksimova). Let A be a logic containing G. Then A
has the global Beth property.
PROOF. Suppose that p(p, g) is a global implicit definition of p in A.
Then ~(p, q-); p(r, q~ I~-i p ++ r. Using Lemma 3.7.12 we get g-boxed formulae
XI(P, q-) and X2(P,q~ such that
(t) A~ v(p,~ ++ ((p A x~(p, q3) v (~p A X2(p, q3))
Since we also have
A ~ [ ] ~ ( p , q-) A s - < ~ ( r , ~ -+ (p ++ r)
we now get
A ~ (D~(p, q0 A ~ v ( r , ~ A p A X~(P, q-) A ~ r A x2(r, q-)) -+ (p -+ r)

This formula has the form (# A p A ~r) -+ (p --+ r), where # is g-boxed. This
is equivalent to 7 p ~/~p V r, or (-~p A r) -+ ~#. By use of Lemma 3.7.11 we
deduce that A ~- #, that is,
We substitute p for r and obtain
A ~ ~ ( p , q3 -+ (x~(p, q3 -~ ~x~(~,q0) -
Now from this and (~) it follows after some boolean manipulations
By the fixed point theorem for G there is a ¢(q~ such that
~ []-<'(p ++ x~(p, q0) -+ (p ++ ¢(q-)).

So we obtain
h e []<l~(p, g) _+ (p ++ ¢(q3),
which is nothing but
~(p, ~ IrA p ++ ¢(q~.
[]
E x e r c i s e 115. Show that if a logic has 1-interpolation then it has n -

interpolation for every n E w.
E x e r c i s e 116. (RAUTENBERG[171].) Let A have local interpolation and let

A be a set of constant formulae. Then A • A has local interpolation.
E x e r c i s e 117. As in the previous exercise, but with global interpolation.
E x e r c i s e 118. Show that the logic of the following frame is not Halld@n-
complete but has interpolation.
Exercise 119. Show that any quasi-normal logic determined by a single

rooted frame (5, x/ is Halld~n-complete.
E x e r c i s e 120. Show that if a logic has the local B e t h - p r o p e r t y it also has

the global Beth-property.
E x e r c i s e 121. (MAKSIMOVA[149].) Show L e m m a 3.7.11.
E x e r c i s e 122. (MAKSIMOVA[149].) Show L e m m a 3.7.12. Hint. You m a y

assume t h a t ~(p,p~ is is normal form. Now reduce the case where it is a
disjunction to the case where it is not.
3.8. Tableau Calculi a n d I n t e r p o l a t i o n

In this chapter we will introduce the notion of a tableau (plural tableaux),
a popular method for showing the decidability of logics. Since we have al-
ready established the decidability via the finite model property, we do not
need tableaux for this purpose. However, there are additional reasons for
studying tableau methods. One is the connection between tableau calculi and
interpolation. This connection has been explored by WOLFGANG RAUTEN-
BERG in [171]. However, we will show at the end that a proof via normal
forms is also possible. Tableaux are also an efficient technique for checking
the satisfiability of a formula in a logic. They will allow to give bounds on
the complexity of the satisfiability problem. Tableaux can be described as
a method of deriving a contradiction as fast and efficiently as possible. A
tableau can be seen as computing along branches in a model to see at which
point we reach an inconsistency. To help in understanding these remarks, let
us describe a tableau calculus for K~, which we denote here by CK. To keep
3.8. Tableau Calculi and Interpolation 147
the calculus short, we assume to have only the symbols -7, A and V]j. The
other symbols are treated as abbreviations. The calculus operates on sets of
formulae. As usual, X; ~o denotes X 0 {~} and X; Y denotes X t2 Y. Thus
X; to; to is the same as X; ~o. The rules are as follows.
x;~ (AE) X;~A¢
X;~ X;~o;¢
x; A ¢) []iX;
(VE) (OjE)
x; :IX; x;
X;Y
(w)
X
The last rules is known as w e a k e n i n g . We abbreviate by ( O E ) the set of
rules { ( O j E ) : j < n}. A O K - t a b l e a u for X is a rooted labelled tree, the
labels being sets of formulae, such that (i) the root has label X , (ii) if a node
x has label Y and a single daughter y then the label of y arises from the label
of x by applying one of ( ~ E ) , (AE), ([]E) or (w), (iii) if x has two daughters
y and z then the label of x is Y; ~(~o A ~b) and the labels of y and z are Y; ~
and Y; ~ ¢ , respectively. There are possibly several tableaux for a given X.
(VE) introduces a branching into the tableau, and it is the only rule that does
so. A branch of a tableau closes if it ends with p; ~p for some variable p.
The tableau closes if all branches close.
PROPOSITION 3.8.1. If X has a CK-tabIeau which closes then X is in-
consistent in K~.
PROOF. By induction on the length of the tableau. Clearly, at the leaves
we have sets of the form p; ~p, and they are all inconsistent. So, we have to
prove for each rule that if the lower sets are inconsistent, so is the upper set.
This is called the correctness of the rules. For example, if X is inconsistent,
then X; Y is inconsistent, so (w) is correct. The boolean rules are straightfor-
ward. For (DE), assume that X; ~ o is inconsistent, that is, X ~-K~ ~. Then
FljZ ~-K,, [~j~, that is, []jX; ~[]j~o is inconsistent. []
The calculus is also complete; that means, if no tableau closes, then X
is consistent. We prove this by showing that whenever there is no closing
tableau there is a model for X. Before we proceed, let us remark that one can
remove weakening from the tableau calculus. However, this is possible only if
([]jE) is replaced by the following rule.
(OhE,) X; oj: x[] := {x: c x}

X[]; ~ o
Removing weakening is desirable from a computational point of view, because
the rule of weakening introduces too many options in a search space. Given
a set X we can apply 2 ~x - 2 nontrivial weakenings but very few of t h e m

turn out to be sensible. Instead, if one intends to speed up the proof search
one has to implement a different calculus. Moreover, one can try to close a
branch as early as possible, e. g. when hitting a set containing ~ and ~ ,
where ~ can be any formula. One can also introduce priorities on the rules,
such as to prefer applying ( ~ E ) and (VE) before any other rule, and to delay
(VE) (or ( D E ' ) ) . To show that the tableau calculus is complete we shall have
to show that if no tableau for a set X closes then there is a model for X ,
which will be enough to show that X is consistent. To understand how such a
model can be found we imagine the tableau as computing possible valuations
at worlds in a model. We start somewhere and see whether we can fulfill
X. The rules ( ~ E ) , (AE) and (VE) are local rules. They allow us to derive
something about what has to be true at the given world. By using the rule
( D E ' ) , however, we go into another world and investigate the valuation there.
Thus, ( D E ' ) is not local; we call it the step rule. Once we have made a step
there is no returning back. This is why one should always apply local rules
first. In fact, we can also prove t h a t any closing tableau gives rise to another
closing tableau where ( D E ' ) has been applied only when nothing else was
possible. Call a set d o w n w a r d s a t u r a t e d if it is closed under ( ~ E ) , (VE)
and (AE). Alternatively, X is downward saturated if (i) for every -~-~ E X
also ~ 6 X , (ii) for every ~ A ¢ G X both ~o 6 X and ¢ 6 X , and (iii) for
every ~ ( p A ¢) 6 X either -,~ C X or ~ ¢ 6 X.
LEMMA 3.8.2. Suppose there is a closing tableau for X . Then there is a

closing tableau for X where (DE') is applied only to downward saturated sets.
PaOOF. Consider an application of ( D E ' ) where an application of (AE)

has been possible instead:
X; ~ A ¢; []jY; -~[2]jX
Y; ~X
Replace this derivation by
X; ~ A ¢; DjY; -~[3jX
X; ~; ¢; DjY; -'[]jX
Y; ~X
By assumption, there is a closing tableau for Y; X, so the latter tableau can

be brought to close as well. Similarly for ( ~ E ) . With (VE) we get
X; ~ V ¢; []jY; -~[3jX
Y; ~X
which we replace by
X;@ V ¢ ; D j Y ; ~ V ] j X
X; ~; DjY; 713jX X; @;DjY; -~[]jX
Y; x Y; 7x
Again, Y; X has a closing tableau by assumption, so the latter derivation has
a closing tableau. This process of swapping derivations yields less and less
offending instances, so it terminates in a derivation where ( D E ' ) is applied
only when all formulae are either variables, negated variables or of the form
7DjX, DjX. []
Call a GK-tableau g o o d if the rule (DE/) is applied to downward saturated
sets. Now we use good tableau-derivations to find our model. Assume t h a t
no closing tableau for X exists, hence also no closing good tableau. The
frame will be based on worlds wz for downward saturated Z. Let S x be
the set of all sets in any tableau for X which is on a branch that does not
close. By assumption, X E Sx. Within S x let Satz be the subcollection of
downward saturated sets in Sx. By assumption, for each Z E S x there exists
a saturated Z* E Satz containing X. Let Z, Y C Satx. Then put Y <lj Z iff
Z is a saturation of a set U obtained by applying ( [ ] j E I) to Y. This defines
the frame G a t z := (Satz, ( ~ j : j < n)). Furthermore, let Y E/3(p) iffp E Y.
By assumption, for no p we have both p E Y and 7p E Y and so the definition
is not contradictory. We will now show that if T E Y then (®atx,/3, Y) ~ T.
By definition of Satx we never have both ~ E Y and 7 T E Y. Now, let
= ¢1 A ¢ 2 . I f ¢ l A ¢ 2 E Y then b o t h ¢1,¢2 C Y and so by induction
hypothesis the claim follows. Next, let T = ~ 7 ¢ . If ~ E Y, then also ¢ E Y
and applying the induction hypothesis, the claim follows again. Finally, if
~(¢1 A ¢2) E Y then either 7¢1 E Y or 7~b2 E Y. In either case we conclude
(Gatx,/3, Y) ~ 9, using the induction hypothesis. Now assume ~ = 712]j¢ for
some ¢. By construction there is a downward saturated Z such t h a t Y <lj Z
and 7 ¢ E Z. By induction hypothesis, (Gatx,/3, Z) ~ 7 ¢ , showing t h a t
(Gatx,/3, Y) ~ ~. Now let T = [Zj¢ and take a j-successor Z of Y. This
successor is of the form U* where U* is the saturation of U, which in turn
results from Y by applying ([BE'). Thus (Gatx,/3, U*) ~ ¢, by induction
hypothesis. Hence (Gatx,/3, Y) ~ ~.
THEOREM 3.8.3. A set X of modal formulae is K~-consistent iff no K,~-
tableau for X closes. K,~-satisfiabiIity is in PSPACE.
PROOF. Only the last claim needs proof. It is enough to see that we can
do the tableau algorithm keeping track only of a single branch, backtracking
to a branching point whenever necessary. (This is not entirely straightforward
but can be achieved with a certain amount of bookkeeping.) So, we need to
show that a branch consumes only polynomial space. We see however that
each branch of the tableau is of depth _< dp(T) and that each node is of length
_< 2. []
We note that by a result LADNER [137], satisfiability is also P S P A C E -

complete. (In fact, LADNER proves in that paper that all logics in the interval
[K, S4] are PSPACE-hard.) With each tableau calculus we can associate a
dual tableau calculus. Notice that for formulae a dual has been defined by
reversing the roles of A and V as well as • j and 0j. In particular, if ~ is a
formula, ~ p is the same as ~d[~p/p]. The dual tableau calculus consists in
reversing the roles of A and V, Eli and (}j. Here a set X is read disjunctively,
and we attempt to show that X is a theorem. For example, the following is a
rule of the dual calculus
((}D) (}iX; ~(}j~

X; -,~
For suppose t-K V XV~(fl. Then T ~-K V X is valid, and so is (}j~ [-U (}j V X-
Hence -,(}jTV V (}iX is a theorem. It is the same to show that X has a closing
tableau as it is to show that ~ X has a closing dual tableau. Notice that the
dual of (VE) is (AE) and has no branching, while the dual of (AE) is (VE)
and does have branching. The dual of (w) is (w), and the dual of ( = E ) is
(=E) as is easily checked.
THEOaEM 3.8.4. ~ C K,~ iff there is a closed dual tableau for {~}.
Let us now see in what ways a tableau calculus helps in finding inter-
polants. Consider W ~-K ~, that is, ~; 7 ¢ is K-inconsistent. Then we need to
find a X in the common variables of ~ and ¢ such that ~; =;g is inconsistent
and X; = ¢ is inconsistent as well. The idea is now to start from a closing
tableau for W; 9 ¢ and define X by induction on the tableau, starting with the
leaves. To that end, let us introduce a marking or signing of the occurring
formulae in a tableau. In the tableau we introduce the marks a and c, where
a stands for antecedent and c for consequent. Each formula is thus signed,
whereby we signal whether it is a part of 9~ or a part of ¢ in the tableau. We
start with the initial set ~a; (=¢)c. The rules are then as follows. The tableau
operates on sets of the X; W, where X remains unchanged when passing from
above the line to below (except for ([]E)). In the marked rule, each formula
in the set X inherits its previous mark, while the newly occurring formulae
after applying the rule inherit their mark from ~, the formula on which the
rule operates. (There is a slight twist here. A set may contain a formula y)
that is marked ~a as well as ~c. In that case we will treat these two as distinct
formulae. Alternatively, we may allow a formula to carry a double mark, i. e.
in this case y)ac.) For example, the marked rule (VE) looks as follows.
X~;Y~;-~(~ A ¢) c
In (DE), a formula [~j¢ of ~ j X has the same mark as the corresponding ¢

in X below. This concludes the marking procedure.
Now suppose that we have a closing tableau. We will now define two
tableaux, one with formulae marked a and i and the other with formulae
marked as i and c. The first will be a tableau for 9; ~X and the second a dual
tableau for X; ~%b. Again, the same rules for marking apply, i stands here
for interpolant. The construction will be by unzipping the original tableau as
follows. Each time we have a set X in the tableau, we show t h a t it is of the
form X = Ya; Z c and that there is a formula Xi such t h a t Y~; (-~X)i has a
closing tableau and ~ZC; (~X) i has a closing dual tableau. This will yield the
desired conclusion; for if 9; ~X closes, we get 9 t- X, and if ¢; ~X dual closes,
we get ~- ~X V ¢, that is, X t- ¢. In the actual proof we will construct ~X
instead of X, which makes the proof easier to read. So, we will construct a X
such t h a t 9; X has a closing tableau, and 9 ¢ ; X a closing dual tableau.
A dual tableau for ~ X is isomorphic to a tableau for X d, the dual of
X. So there is an inherent duality in the construction, which we will use
extensively. The proof will be by induction. We begin with the leaves. There
are four possibilities how a branch can close, either as (aa) pa; (~p)~ or as
(ac) p~; (~p)C or as (ca) pC; (~p)~ or as (cc) pC; (-~p)L In case (aa) we choose
X := T, in case (ac) we chose X := ~P, in case (ca) we choose X := P, and
in case (cc) we put X := J_. In each case the claim is directly verified. Now
suppose that (w) has been applied.
X 1a ., X 2C,. ~ a . ,Y~c
(w)
x?;
By induction hypothesis, there is an interpolant X for X~; X~. It is easy to
a. C.
verify that X also is an interpolant for X1, X2, Y{a . , Y~.
C
Next, suppose (DE)
has been applied, say on the antecedent formula ~ D j g .
[~j xa ; []j yc; ( ~ h j g )a

X~; yc; (~9)a
By induction hypothesis we have a closing tableau for Xa; ~ga; Xi and a dual
closing tableau for Xi; (~y)c. Now consider the following rule applications.
D x°; ( 2jx)
_~yc; Xi
By assumption, the lower line has a closing tableau; hence the upper line has
a closing tableau as well. This shows that ([~jX) is an interpolant for the
premiss of (DE). Now suppose that the rule has operated on a consequent
formula, (~[~j¢)c. Then we have to put as the new interpolating formula the
dual formula, (}iX = -~E3j~X. The argumentation now is completely dual. We

have
~jxa; (~E];~x) ~
Xa; ~ X i 0 ~ x ~ ; (~D~¢)~; ((bx) ~
~X~; -~b¢; Xi
X ~; Xi
The lower lines have a closing tableau, and so do therefore the premisses. In
the case where (--E) and (AE) have applied, we choose for the new interpolant
the old one. It is easy to check that this satisfies the requirements. In the
first case, if ( ~ E ) has operated on an a - f o r m u l a we get
x°; ( ~ 9 ) ° ; x '
Xa; 9~; Xi
So, since Xa; 9a; Xi has a closing tableau, so has Xa; (-~-~9)a; X~. On the dual
side, X c has remained unchanged, so nothing is to be proved. Dually if the
rule ( ~ E ) has applied to a c-formula. In case of (AE) and an a - f o r m u l a we
get
x°; (9 A ¢)°;x ~
x a ; 9a; ~ba;X i
By induction hypothesis, the lower set has a closing tableau; thus there is one
for the upper set. Since X ~ did not change, there is nothing to prove for X c.
If (AE) has applied to a c-formula (9 A ~b), the case is exactly dual to an
application of (VE) to an a-formula. Now, finally, the rule (VE). Since we
introduce a split, there are now two interpolation formulae, X1 and X2, one
for each branch. Suppose first that an antecedent formula has been reduced.
Then we put X := X1 A X2.
x°; (9(9 A ¢))°; (x~ A x) *

x°; (~9)=; (x, A x=)~ x°; (~¢)°; (x, A x=)* --X~; (Xx A x2) ~
za; (mg)a ; xil; xi2 m
z°; (¢) a. i. i
, x l , x2
x°; (~9)°; xl xo; (=¢)a; x~
Both tableaux can be brought to close. The first by the fact t h a t we have cho-
sen Xi appropriately. The second by the fact that b o t h X~; ~XC and Xi2; ~ X c
close by induction hypothesis. This concludes the inspection of all rules.
THEOREM 3.8.5. K~ has local interpolation. Moreover, an interpolant

for 9 and #3 can be constructed from a closing tableau for 9; =~.
For special logics extending K1 there exist tableau calculi which allow to
construct interpolants. We will display the relevant rules below.
~ X ; -,[3~ X; []~
(4.) (t.)
X; []X; ~ X;~
[]X; ~ D ~ DX; -~[~
(g-) (grz.)
X; V1X; -.~; []~ x ; [3x; ~ ; [~(~ -+ ~ )
DX; --[3Y; -~[Z~
(altl.)
X;Y;~
It is an easy matter to verify that these rules are sound. Their completeness
is harder to verify directly, but follows easily with the help of the reduction
sets.
Now, finally, for the promised proof of interpolation for K~ that does
not make use of tableaux. In fact, it will show not only interpolation but a
stronger property of K~ called uniform interpolation.
DEFINITION 3.8.6. Let A be a modal logic. A has local u n i f o r m inter-

p o l a t i o n if (i.) given ~ and variables ~, there exists a formula X such that
var(x ) C_ ~ and for all formulae ¢ such that var(~) A var(¢) = ~ we have
~o [-h X ~-A •, and (ii.) given ¢ and variables ~, there exists a formula X such
that vat(x) c ¢ and/or all formulae such that var( ) • vat(e) = ¢ we have
~-h X I-h ¢.
The property (i.) alone is called u n i f o r m p r e i n t e r p o l a t i o n and the
property (ii.) alone u n i f o r m p o s t i n t e r p o l a t i o n . By definition, if a logic
has uniform preinterpolation the interpolant does not depend on the actual
shape of ¢ but only on the set of shared variables. Since ~ ~-A ¢ i f f 7 ¢ ~-A - ~
either of (i.) and (ii.) is sufficient for showing uniform interpolation. We will
now show that K~ has uniform preinterpolation. For simplicity we take the
case of a single operator, that is, we prove the statement for K1. But the
generalization is easy enough to make. The central idea is that when we have
~- ¢, and we have a variable p that occurs in T but not in ¢ we want to
simply erase the variable p in T and define a formula ~q-(P) (or simply pq-)
such that ~q-(P) ? ¢. Define pT(p) as follows. An occurrence of the variable
p is replaced by 5- if it is embedded in an even number negations, and by 5_
otherwise. Given this definition it is easily shown that ~ ~- ~q-. This is one
half of what we need; the other half is ~T ~_ ¢. Now here things can go wrong.
Take, for example, T := pA ~p and ¢ := _l_. Clearly, T ~- ¢. Given the current
definition, ~7- = T A 95_. This formula is equivalent to T. Hence T }z 5-. To
surround this problem, put ~ into standard and explicit form. We will show
in the next lemma that any model for pq- based on an intransitive tree is the
p-morphic image of a model for ~. Here, an intransitive tree is a frame (f, <~)
154 3. ~ n d a m e n t a l s of Modal Logic II
such that it contains no cycles with respect to < and in which x q z and y < z
implies x = y.
LEMMA 3.8.7. Suppose that ~ is standard, explicit and clash free. Then
for any finite model (~,/3, x} for ~ T(p) based on an intransitive tree there exists
a model (~, 7, xt) for ~ such that
i (g, q) is an intransitive tree and there exists a contraction c : ~ - , ~,
ii for all q # p,/3(q) = c[7(q)]
iii x = c(x').
PROOF. By induction on the modal depth of p. Assume that the depth
is 0. Then ~ -- Vi<m ~i, each ~i a conjunction of sentence letters or their
negations. Then ~T = V~<,~ v~. Suppose that (~,/3, w0} ~ qoT. We assume
that/3 is defined only on the variables of pT. Then there exists an i < m such
that (~,/3, x> ~ u [ . Now put 7(q) :=/3(P) if q 5£ p and 7(P) := {wo} if p is a
conjunct of ui and 7(P) := ~8, else. Since p is clash free, this is well-defined.
Then (I, 7, w0} ~ ul, and so (~, % Wo) b ~. Now let dp(~) > 0. Then ~, is a
disjunction of formulae 9oi of the form
~ i ---- # A A o¢; A/-1 X.
j<n
Assume further that (~,/3, wo) ~ ~T. /3 is defined only on the variables of (pT
Then for some i we have ([,/3, wo) ~ ~i. Furthermore,
= A A ocJ-/, T .
j<n
Put ~o := [, sO(p) := {w0} i f p is a conjunct of p, and f f ° ( p ) : = £~ otherwise.
For q 5£ p put 7°(q) :=/3(@. Then (~o, 7 0 wo) ~ #. Let the set of successsors
of Wo be suc(wo) = {x~ : a < A}. A is finite. Inductively, for each a < A we
perform the following operation. C a s e 1. (~a, 7 ~, xa) b @T for some j. Then
put J := {k : @k ++ ~bj C K~}. Let ~ be the transit of x~ in [~. By induction
hypothesis, there exists a model (bk,Sk,Yk} ~ ¢k for each k E J and a p -
morphism ek : Ok -~ g satisfying (ii) and (iii). Form f~+~ by blowing up the
frame ~ to (~kcJ Dk. (See Theorem 3.3.3.) There exists a p-morphism d a+l :
f~+l __~ is, obtained by extending Gk~J ek to f a + l . Moreover, 7a+l(p) :----
7~(p) U [[~ k ~ j ( d a+l'~-i
, [Sk(P)], and for q -¢ p, ~/a+X(q) :__ (d~+~)-~[7~(q)]"
C a s e 2. ([~,7~,x~} ~ %b~-for all j. Then at least ([~,7~,x~} ~ XT, and
we proceed as follows. We know that X is a disjunction of simple standard,
explicit and clash-free formulae ~-~. For some i we have (~a, 7 a, x~} ~ (~_)T.
Let ~ be the transit of x~ in I. By induction hypothesis there exists a ~j
and a p-morphism e : I) -* ~, a 50 and y satisfying (ii) and (iii) such that
(b,5 °,y} ~ % and so ([),50 ,y} ~ X. Now blow up g to b in [. This defines
[~+~. There exists a p-morphism d ~+~ : f~+~ --~ ~ . P u t 7'~+~(p) :-- 7~(p) U
(da+~)-l["/a(p)], and let 7a+~(q) := (da+~)-~[7~(q)] for q 5£ p.
It is clear that ( ~ , 7 ~, xz) ~ Cj iff ([a+l, ./a-t-l, Xf~) ~ ~)j for all /3 < O~,
and ([~,7~,x~) ~ (¢j)T iff ( ~ + I , V ~ + I , x z ) ~ (¢j)T for/3 > a. Furthermore,
([~+1,7~+1,w0) ~ #. Put 9 := [~- Then the composition c :-- d 1 o d 2 o . . . o d ~ :
t~ --~ ~. For q ¢ p, 7(q) := 7~(q), is the result of blowing up/3 by c. Moreover,
c(wo) = w0, and
(~, % To) ~ ~ ,
which had to be shown. []
THEOREM 3.8.8. K~ has uniform local interpolation.
PROOF. Suppose that ~(g,q-) ~z ¢(~,7r). Let To(/Y,q~ be a standard, ex-
plicit and clash-free formula deductively equivalent to T. Let/Y consist of
the variables Pi, i ~ n. Now for i < n, let Ti+l(iff,~ := ~i(:ff, q-)T(p,).
X := ~,~. Then the variables Pi do not occur in X- Moreover, ~ ~- X. Now
assume that X )z ¢. Then there exists a finite intransitive tree [o, a valua-
tion/30 and a wo such that (~0,/3o,W0) ~ X;~¢. By the previous lemma, if
([i,/3i, wo) ~ ~n-i; ~ and ~i is an intransitive tree then there exists a model
(~i+l,13i+l,Wo) ~ ~ n - i - 1 ; ~ such that ~i+~ is an intransitive tree. Hence, by
induction, we have (In,/3=, wo) ~ To; ~¢. This means that ~ ~z ¢, since ~o is
deductively equivalent with ~. This contradicts the fact that ~ f- ~b. []
It should be noted that the assumption that K . is complete with respect to
finite intransitive trees is rather essential for the proof method. The method
of reduction sets cannot be applied to show uniform interpolation for the stan-
dard systems. The notion of uniform interpolation has been introduced by
ANDREW PITTS [161]. It has been shown subsequently by SILVIO GHILARDI
and M. ZAWADOWKSI [74] and ALBERT VISSER [223] that K , G r z and G
have uniform interpolation, but that $4 lacks uniform interpolation. Further-
more, FRANK WOLTER [243] proves that uniform interpolation is preserved
under fusion. From the latter follows already that polymodal K has uniform
interpolation if only K has uniform interpolation.
Notes on this section. The notion of a downward saturated set first ap-
peared in HINTIKKA [104], who gave a tableau calculus for K 4 and other sys-
tems. Tableau calculi have attracted much interest in machine based theorem
proving. The literature is too large to be adequately summarized here. Suffice
to mention here the work by MELVIN FITTING [70], RAJEEV GORI~ [88] and
MARTIN AMERBAUER [1]. Furthermore, [224] contains an overview of proof
theory in modal logic. Tableau calculi are closely connected to Gentzen-
calculi. A Gentzen calculus operates on pairs of sets of formulae, (F, A),
written F ~- A. It is possible to reformulate a Gentzen calculus as a tableau
calculus. The method for proving interpolation employed above is quite sim-
ilar to the one introduced by S. MAEHARA in [144].
E x e r c i s e 123. Show that the rule (4.) is sound and complete for K4. Hint.
Use the reduction sets for K4.
E x e r c i s e 124. Show that the rule (t.) is sound and complete for K . T , and
that the rules (t.) and (4.) together are sound and complete for S4.
Exercise 125. Show that (g.) is sound and complete for G, and that (grz.)
is sound and complete for G r z . Hint. You have to use two reductions in
succession here, first one to K 4 and then one to K .
E x e r c i s e 126. Show that (alt>) is sound and complete for K . a l t l .
E x e r c i s e 127. Show that ~o }--K ~pT(p).
E x e r c i s e 128. (WOLTER [243].) Show that if AIg A is a locally finite variety

and A has interpolation, then A has uniform interpolation. It follows, for
example, that S5 has uniform interpolation. Hint. The notion of a locally
finite variety is defined in Section 4.8. You may work with the following char-
acterization: AlgA is locally finite iff CanA(n) is finite for every n C ~o.
E x e r c i s e 129. Show that G and G r z are in PSPACE. Hint. Show that

the length of a branch for in a tableau for ~ is bounded by the number of
subformulae of ~.
E x e r c i s e 130. Here is a tableau cMculus for determining whether or not

/k I[-K, ~ ~. Introduce a formula diacritic ~'. Now replace the rule ( D I E ) by
-~[]j~o; DjX; Y~
-~,; X; Y~
Furthermore, add the rule
X ; y t'
X ; Y~'; Y
Show that A I~-K~ ~P iff A$;~O has a closing tableau. Show then that the
problem ' ¢ I[-K~ ~0.9' iS in EXPTIME.
3.9. Modal Consequence Relations

In this section we wilt study the lattice of all consequence relations.
Thereby we will also elucidate the role of the consequence relations kA and
I~-A. We start by defining the notion of a modal consequence relation and give
some alternative characterizations.
DEFINITION 3.9.1. A m o d a l c o n s e q u e n c e r e l a t i o n is a consequence re-
lation over [P~ which contains at least the rule (rap.) and whose set of tautolo-
gies is a modal logic. If F- is a modal consequence relation and A := Taut0- )
3.9. Modal Consequence Relations 157
then F is a modal consequence relation f o r A. F is n o r m a l ( q u a s i -

n o r m a l , classical, m o n o t o n e ) if A is.
In sequel we will deal exclusively with normal consequence relations. No-
tice that if F C ~o(?~) × ~P~ is a consequence relation, then we m a y define
F ~ : = F N (~(?~) × ~P~) if A < ~;, and if A > ~ let F ~ denote the least
consequence relation over ?~ containing F. A special case is A = 0.
PROPOSITION 3.9.2. Let F be a consistent modal consequence relation.
Then its reduct to the language T, ~ and A is the consequence relation of
boolean logic.
PROOF. Taut(F °) contains all boolean tautologies. For we have
Taut(F °) = Taut(F) FI ?o.
The rule of modus ponens is contained in F °. Therefore, F ° contains the
consequence relation F2. If F is consistent, we have p Y q, so p y0 q, and so
F ° is consistent too. However, F2 is maximal, and so equal to F °. []
If F is a normal consequence relation, we denote by Q(F) the lattice of
extensions of F and by £(F) the lattice of normal extensions. We will show
below that this is a complete and algebraic lattice. Now let A be a modal
logic. Then define
CRel(A) := {F: Taut(F) = A}
As we have seen in Section 1.4 of Chapter 1, CRel(k) is an interval with respect
to inclusion. The smallest m e m b e r is in fact F , , as follows immediately from
the definition. The largest member will be denoted by F ~ ; it is structurally
complete. As is clear, for a given set of tautologies there exists exactly one
structurally complete logic with the given tautologies, and at most one logic
with a deduction theorem for --+. FX~ is the structurally complete logic with
respect to A and F , is the logic with a deduction theorem for -4.
PROPOSITION 3.9.3. Let A be a modal logic. Then
CRel(A) = {F : FA C F C_ F'~}
Moreover, FA is the unique member of CRel(A) having a deduction theorem
for --+ and F'~ the unique member which is structurally complete.
To see some more examples, consider the rule ({[~p},p). It is admissible
in K . For assume that qo := p~ is not a theorem. Then there exists a model
([,/3, x) ~ - ~ . Consider the frame 9 based on f U {z}, where x ~ f , and the
relation *I:= <~ U {(z, y) : y e f } . Take 7(P) := 13(p). Then (i~, 7, z) ~ -~K]qo.
We warn the reader here that even though for any modal consequence relation,
3 p F p is equivalent to p F 0p, the rule ({p}, 0p) is not admissible in K despite
Lhe admissibility of ({[]p},p). Take p := T. 0 T is not a theorem of K .
Similarly, the so-called Macintosh rule ({p -4 [Zp}, 0p --+ p) is not admissible
for K . Namely, put p := [33_. [2]/ -+ []F33_ is a theorem but <}[31 -+ [33_
is not. Notice also t h a t if a rule p is admissible in a logic O we m a y not
conclude that p is admissible in every extension of ®. A case in point is the
rule ({[Z]p},p), which is not admissible in K ® [3_L.
Recall the notation F R. This denotes the consequence relation generated
by the rules R. At present we m a y tacitly assume t h a t R contains (mp.).
Equivalently, F R is the least modal consequence relation containing R. Notice
that for every modal consequence relation ~- there exists an R with F- = ~_n
(for example the set of all finitary rules of F itself).
PROPOSITION 3.9.4. The set of modal consequence relations over ~,~ forms
an algebraic lattice. The compact elements are exactly the finitely axiomatiz-
able consequence relations. The lattice of quasi-normal consequence relations
is the sublattice of consequence relations containing FK..
PROOF. Clearly, the operation [7 is set intersection, and F1 II ~-2 is the
smallest consequence relation containing b o t h [-1 and F2. If F1 = F nl and
F2 = F n2 then F1 U ~-2 = t-RluR2- With this latter characterization it is easy
to define the infinite union. Namely, if ~-i = ~-R~ for i E I, put [_]z ~-i := Fs,
where S := Uz Ri. All rules are finitary by definition. Therefore, if a rule
is derivable in F s, then it is derivable already from a finite union of the Fi.
It follows t h a t a finitely axiomatizable consequence relation is compact, and
that a compact consequence relation is finitely axiomatizable. Moreover, the
lattice is algebraic, since F R = L.JpER ~-p" The last claim is a consequence
of the fact that F' is quasi-normal iff Taut(l-') is quasi-normal iff Taut(F')
contains K~. []
PROPOSITION 3.9.5. For each quasi-normal logic A and each quasi-normal
consequence relation ~-',
FA C I-' ~ AC Taut(l-')
Taut(-) commutes with infinite intersections, FA with infinite intersections
and joins.
PROPOSITION 3.9.6. ~ (FK~) is a complete sublattice of Q (t-K.).
PROPOSITION 3.9.7. In monomodal logic, F A is maximal iff A is a coatom.
PROOF. Clearly, if FA is maximal in ~(FK), A must be a coatom. To
show the converse, we need to show is t h a t for a maximal consistent normal
logic A, FA is structurally complete. (It will follow that CReI(A) has exactly
one element.) Now, A is Post-complete iff it contains either the formula IN-I-or
the formula p ++ [Zp. Assume t h a t FA can be expanded by a rule p = (A, T).
Then, by using the axioms p can be transformed into a rule p' = (A', ~')
in which the formulae are nonmodal. (Namely, any formula in a rule m a y
be exchanged by a deductively equivalent formula. Either [ZT E A and any
subformula D X may be replaced by T, or p e+ []p C A and then DX m a y be

replaced by X.) A nonmodal rule not derivable in ~-A is also not derivable in
its boolean fragment, ~_o. By the maximality of the latter, adding p' yields
the inconsistent logic. []
In polymodal logics matters are a bit more complicated. We will see t h a t

there exist in fact 2 ~° logics which are coatoms in ~ K2 without their conse-
quence relation being maximal. Moreover, we will see that even in monomodal
logics there exist 2 ~° maximal consequence relations, which are therefore not
of the form F-A (except for the two abovementioned consequence relations).
Notice that even though a consequence is maximal iff it is structurally com-
plete and Post-complete, Post-completeness is relative to the derivable rules.
Therefore, this does not mean that the tautologies are a maximally consistent
modal logic. We define the T - s p e c t r u m of A, TSp(A), to be the cardinality
of the set of consequence relations whose set of tautologies is A.
TSp(A) := card CRel(A)
To characterize the choices for the T - s p e c t r u m we will first deal with a seem-
ingly different question.
PROPOSITION 3.9.8. Let A be a normal logic. Then the following are
equivalent.
1. ~-h -----I~-h.
2. IkA admits a deduction theorem for -4.
3. A D _ K , ~ O { p - 4 [ ] i p : j < ~ }.
4. A is the logic of some set of Kripke-frames containing only one world.
PROOF. Clearly, if (1.) holds, then (2.) holds as well. Now let (2.) be the
case. Then since p l~-h Dip, by the deduction theorem, I~-A p -4 [[]jp, which
gives (3.). From (3.) we deduce (1.) as follows. Since p;p -4 []jp ~-A []jP, and
p -4 []jp E A, we get p ~-A [Zjp. Therefore, It-A=~-A. Finally we establish the
equivalence of (3.) and (4.). Assume (4.). Then clearly the formulae p -4 []jp
are axioms, since they hold on any one-point frame. Now assume t h a t (4.)
fails. Consider a rooted generated subframe ~ of Cana (vat) consisting of more
than one point. Such a frame must exist, by assumption. Let X be the root
and Y a j-successor. Then there exists a set ~ such t h a t X ~ ~ but Y ~ ~.
Now put fl(p) := ~. It follows that (~,fl, X ) ~ p;~Djp. Hence (3.) fails as
welt. []
Now let us return to the question of T - s p e c t r a . Clearly, if ~-A # t~-A then

the T - s p e c t r u m of A cannot be 1. We will show now t h a t the converse almost
holds.
PROPOSITION 3.9.9. Let A be a modal logic. Then the following are equi-
valent.
1. The T-spectrum of h is 1.
2. ~-h is structurally complete.
3. A is the logic of a single Kripke-frame containing a single world.
4. A is a fusion of monomodaI logics of the frames [ ] or [ ] .
PROOF. The equivalence between (1.) and (2.) is immediate. The equi-
valence of (3.) and (4.) is also not hard to show. If A is a fusion of logics
for one-point frames it contains for each operator either the axiom Djh- or
p ++ [:]jp. It means that the relation <~j is on all frames e m p t y or on all
frames the diagonal. Hence the generated subframes of the canonical frame
are one-point frames and they are all isomorphic. Finally, we show (2.) ~=~
(3.). Assume (3.). Then by the fact t h a t the ~-A is the logic of a single algebra
based on two elements, and has all constants, it is structurally complete. Now
let (3.) fail. There are basically two cases. If A is not the logic of one-point
frames, then ~-A is anyway not structurally complete by the previous theorem.
Otherwise, it is the intersection of logics determined by matrices of the form
(92, D), D an open filter, 9.1 the free algebra in R0 generators. (In fact, the
freely 0-generated algebra is enough.) 9.1 contains a constant c such t h a t
0 < c < 1. Namely, take two different one point frames. Then, say, []o is the
diagonal on one frame and e m p t y on the other. Then c := ¢01 is a constant
of the required form. The rule {{(}0-[-},P} is admissible but not derivable. []
The method of the last proof can be used in m a n y different ways.
LEMMA 3.9.10. Let A be a logic and X a constant formula such that nei-
ther X not -~X are inconsistent. Then the rule P[X] := {{X}, ± ) is admissible
for A but not derivable in F-A.
PROOF. Since X !~ A and var(x ) = ~, for no substitution or, X~ C A.
Hence the rule P[X] is admissible. If it is derivable in FA then ~-h X -+ ± , by
the deduction theorem. So ~X • A, which is not the case. So, P[X] is not
derivable. []
THEOREM 3.9.11. Let A be a logic such that ~r,(O) has infinitely many
elements. Then TSp(A) = 2 s°.
PROOF. Let A C FrA(0). Call A a b l o c k if 0, 1 ~ A and for every
a , b • A, i f a # b then a n b = 0. For each a • A there exists a constant
formula Xa whose value in ~rA(0) is a. So, for a subset S C_ A, put
F-sA:= ~-, U U aes FP[x°]
We claim that (1) FA S • T(FA), (2) if S ¢ T then ~-A s = F T, and (3) ~rA(0)
contains a block of size R0. A d (1). By the previous lemma, all the rules
P[X,] are admissible, by the requirement that a • A and A is a block. A d
(2). Suppose that a • S - T. Then let U be the closure under (mp.) of Xa in
A. We claim t h a t U is a theory of F T but not of FA
S. To be a theory of F~ for
some set R no more is required t h a n it be closed under (mp.) and consistent

if it is does not contain any Xa, a • R, or else be inconsistent (and contain
all formulae). Now since a ¢ T and Xa is consistent, the (mp.) closure does
not contain any Xb, b • T, since a N b = 0 (which means t h a t Xa ~-h -'Xb)"
Since a • S, X~ ~-A S ±- A d (3). We distinguish three cases. C a s e A . ~rh(0)
has infinitely m a n y atoms. T h e n the a t o m s form a block of size at least R0-
C a s e B . ~rA(0) has no atoms. T h e n there exists a sequence (ci : i • w} such
t h a t 0 < ci+l < ci < 1 for a l l i < w. P u t ai : = c i - c i + l . T h e n 0 < ai
since ci > ci+l and ai < 1 since ci < 1. Furthermore, let i < j. T h e n
a~ Naj = (c~ N - c i + l ) N (cj N - c j + l ) = cj N - c i + l _< ci+l n - c i + l = 0. So there
exists an infinite block. C a s e C. There exist finitely m a n y atoms. T h e n let
be the boolean algebra underlying ~rA(0). We claim t h a t ~ ------92 x ~ , where
92 is finite and ~ is atomless. (This is left as an exercise.) Now ~ contains
an infinite block, (bi: i • w}. P u t ai : = (1, b~}. T h e set A : = { a i : i • w} is a
block in 92 x ~ . []
COROLLARY 3.9.12. Let A be a monomodal logic and A C G.3. Then

TSp(A) = 2 ~°.
PROOF. G . 3 has infinitely m a n y distinct constants, n a m e l y [~"L, n C w.

This applies as well to A. []
Let us remain with G . 3 a little bit. Consider the consequence k~. a. We claim
t h a t it is maximal. To see this we need to show t h a t it is P o s t - c o m p l e t e . This
follows from a general fact t h a t we will establish here.
THEOREM 3.9.13. Let A be O-characterized. Then F'~ is maximal.
PROOF. Let IF ~ ~-~. T h e n Taut(IF) ~ A. Since A is 0-characterized,

there is a c o n s t a n t X such t h a t A C A • X C Taut(IF). Therefore, X ¢ A.
Two cases arise. C a s e 1. -'X ¢ A. T h e n the rule P[X] is admissible in A
and so derivable in A T. Therefore P[X] E lh, and so since IF X, also Ih ±. So,
IF is inconsistent. C a s e 2. ~X E A. T h e n Taut(IF) is inconsistent. So iF is
inconsistent as well. []
This t h e o r e m makes the search for m a x i m a l consequence relations quite

easy. Let us note in passing t h a t there are consequences relations ~-1 and F2
such t h a t
Taut(l-1 U I-2) :fi Taut(~-l) U Taut(F-2) .
Namely, let F I : = F~. a and F 2 : = FK,[3_L. T h e n Taut(k1 U F2) = K • l , but
G . 3 U K G E3_L = K O D / .
We will now investigate the cardinality of the set of c o a t o m s in £ (FK).
We will show t h a t there are exactly 2 ~°. Of course, showing t h a t there are
at least t h a t m a n y is enough; there cannot be more. In the light of the
FIGURE 3.2. ~M, M = { 1 , 3 , 4 , . . . }
4o 3° 1°
0 0
4"
L1
3" 2" 1" 0"
previous theorem, we are done if we can find 2 ~° distinct logics which are
0-characterized. Let M C_ w. P u t TM : : {n" : n C w} U {n ° : n E M}. P u t
(1.) x=m°,y=n ° andm>n
x<y ¢=> or (2.) x=m °,y=n ° andm>n
or (3.) x=m°,y=n°andm=n
Let ~M be the algebra of 0-definable sets. P u t ~TM := (TM, q, ~M)- We will
show now t h a t if M ~ N then Th ~M ~ Th ~Y. To see this, we show t h a t
every one-element set (n ° } in TM is definable by a formula x ( n ) t h a t depends
only on n, not on M. First, take the formula
8(n) := [2~+1± A ~ [ ~ ±
5(n) defines the set { n ' } . Now put
x(n) := 05(n) A ~5(n + 1) A = 0 5 ( n ÷ 1)
It is easily checked that x(n) defines {n°}. Hence, if n ¢ M , x(n) Th

So, -~x(n) 6 Th ~M if~ n 9~ M. This establishes t h a t if M # N , Th ~M
Th ~TN.
THEOREM 3.9.14. The lattice of normal monomodal consequence rela-
tions contains 2 ~° many coatoms.
Notes on this section. In contrast to the theory of modal logics, the theory
of modal consequence relations is not very well developed. Nevertheless, there
has been significant progress in the understanding of consequence relations,
notably through the work of VLADIMIR RYBAKOV. In a series of papers (see
[178], [179], [180], [181] as well as the book [182] and references therein)
he has investigated the question of axiomatizing b ~ by means of a rule basis.
A m a j o r result was the solution of a problem by HARVEY FRIEDMAN to
axiomatize the calculus of admissible rules for intuitionistic logic. In the
more philosophically oriented literature, certain special rules have been in the
focus of attention. Typically, the notion of a rule is somewhat more general.
It is a pair (A, F), where A and F are sets of formulae. It is a d m i s s i b l e for
a logic A, if for every substitution a such that A ~ _C A we have F ~ N A ~ ~.

Examples are the rule of margins
({p -+ Dp}, {p,
the Macintosh rule (which is of course also a rule in our sense), the (strong)
rules of disjunction
( { V Dpi}, {pi : i < n})
i<n
and the weak rules of disjunction
( { V []Pi: [] compound }, {Pi: i < n}>
i<n
The latter are nonfinitary rules. See for example work by BRIAN CHELLAS
and KRISTER SEGERBERG [46] and TIMOTHY WILLIAMSON [227], [226] and
[228].
E x e r c i s e 132. Let A D K~ @ {p --+ []jp : j < ~}. Show that A is the logic
of its 0-generated algebra. It follows that it has all constants. However, ~-h
is not necessarily structurally complete. Can you explain why?
E x e r c i s e 133. Let 91 be a boolean algebra with finitely many atoms. Show

that 91 -~ 83 × ~, where ~ is finite and ~ is atomless.
E x e r c i s e 134. Show that the rule(s) of disjunction are admissible for K.

Hint. Start with models refuting ~i := P~. Now build a model refuting
CHAPTER 4
Universal Algebra and Duality Theory
4.1. M o r e o n P r o d u c t s
In this chapter we will develop the algebraic theory of modal algebras,
taking advantage of some strong theorems of universal algebra. First, we know
from a theorem by GARRETH BIRKHOFF that equational classes correspond
o n e - t o - o n e to varieties, and that the lattice of modal logics is dual to the
lattice of varieties. Second, by using the representation theory of boolean al-
gebras by MARSHALL H. STONE we can derive m a n y useful results about gen-
eral frames, in particular deriving a theorem about modally definable classes
of Kripke-frames. Fuller expositions on universal algebra can be found in
[37], IS9].
We have to begin by talking more about products. A generalization of
the (direct) product of algebras is the so-called subdirect product. We call
92 a s u b d i r e c t p r o d u c t of the ~ i , i E I, if 92 is a subalgebra of the direct
product 1-[icI ~ i such that for each projection lri : 1-Lcx ~ i --" ~ i we have
7ri[A] = Bi. In other words, if 92 is projected onto any factor of the product,
we get the full factor rather than a proper subalgebra. Moreover, also every
algebra isomorphic to 92 will be called a subdirect product of the ~ i . An
alternative characterization of this latter, broader notion is the following. ¢
is a s u b d i r e c t p r o d u c t of the ~ , i E I, if there exists an embedding j : ~
1-LCz ~ i such that for every i C I, 7rioj : ff ~ ~ i . To see a nontrivial example
of a subdirect product, take the algebra 92 of the frame ~ in Figure 4.1. It is
a subdirect product of the algebra ~ of the frame ®; simply take the direct
product ~ × ~ , which is isomorphic to the algebra over ¢~ O ®. Now take
the subalgebra g generated by the encircled sets. It is isomorphic to the
algebra 92. (The sets define the frame 2).) An algebra is called s u b d i r e c t l y
i r r e d u c i b l e (s.i.) if for every subdirect product h : 92 --~ I-IicI ~ i we have
that 7ri o h : 92 --~ ~ i is an isomorphism for some i E I. There are some useful
theorems on subdirect products and subdirect irreducibility. Recall that there
is a smallest congruence on 92, A A = { ( a , a ) : a E A}, also denoted by A, and
a largest congruence VA = A × A, denoted by V when no confusion arises.
The congruences of an algebra form an algebraic lattice.
167
168 4. U n i v e r s a l Algebra and Duality Theory
FIGURE 4.1. A subdirect p r o d u c t
PROPOSITION 4.1.1. Let91 be a subdirect product of the algebras ~ i , i E I.

Let 7ri : [Ii~I ~3~ --+ ~3i be the projections. Then we have
A ker(Tri FA) = A A .
icI
Conversely, if AA = N~er Oi then 91 is a subdirect product of the algebras

91/0~, i C I. The embedding of 91 into r L e i 9 1 / o ~ is defined by the map
h(a) :: ([a]®~ : i E I } .
PROOF. An element of the direct p r o d u c t is a sequence a = (a(i) : i E I}.
For two such elements we deduce from (a, b} E Ni~i ker(T:i F A) t h a t (a, b) E
ker(~r~ I A) for all i C I, t h a t is, a(i) = b(i) for all i. Hence a -- b. This proves
the first claim. For the second observe first t h a t h is a h o m o m o r p h i s m . Now
h(a) = h(b) implies [a](gi -- [b]O~ for all i E I, t h a t is (a, b) E ®~ for all i.
Hence a : b, since N i e I ®i : AA. Now 7ri o h[A] -- A/(9~, because if [a]®i is
given, t h e n ~ o h(a) = ~ ( ( [ a ] e ~ : i • I ) ) = [a]e~. []
THEOREM 4.1.2. 91 is subdirectly irreducible iff there exists a congruence

(9 such that every congruence ~ A contains (9. Equivalently, 91 is subdirectly
irreducible iff A is N-irreducible in the lattice of congruences of 91.
PROOF. We show the second claim first. Let N i c I (9i = A for some O~
which are all different from A. T h e n consider the subdirect r e p r e s e n t a t i o n
h : 91 ~-~ 1-Lei(91/O~). T h e n none of the m a p s ~i o h is injective since their
kernel is exactly 01. So, 91 is subdirectly irreducible. Let on the other h a n d
Ni~i ®i ~£ A for all O~ ~£ A. T h e n in a subdirect representation h : 91
rI~ci ~ we have NicI ker(Tri o h) : A and thus (9i : A for some i • I,
showing 91 to be subdirectly irreducible. Now for the first claim. Let 91 be
subdirectly irreducible. T h e n let ® := N ¢ # A q. Since A is N - i r r e d u c i b l e ,
(9 ¢ A. Conversely, if there exists a congruence (9 # A such t h a t every
congruence # A contains (9, then (9 = N ~ # ~ (I), so A is N - i r r e d u c i b l e . []
In case t h a t 91 is subdirectly irreducible, the smallest congruence a b o v e

A is called the m o n o l i t h of 91. Now recall the n o t a t i o n (9(E) for the smallest
4.1. More on Products 169
congruence containing E. We say E g e n e r a t e s 6)(E). A congruence is called

p r i n c i p a l if there is a one-membered E generating it. If E = {(a,b)} we
write 6)(a, b) for 6)(E). It is easy to see that the monolith of a subdirectly
irreducible algebra is principal. The following is due to [17].
THEOREM 4.1.3 (Birkhoff). Every algebra is the subdirect product of sub-
directly irreducible algebras.
PROOF. For each pair a, b E A of elements consider the set ~(a, b) of con-
gruences such that (a, b} 9~ ~(a, b). Suppose we can show that each ~(a, b) has
a maximal element q;(a, b) if a # b. Then let 6)(a, b) be the least congruence
containing (a,b}. 6)(a, b) ~ ~(a,b), by definition. Consider now any congru-
ence 6)' ~ ~2(a, b). By the fact that q(a, b) is maximal for not being above
6)(a, b), we must have 6)' _~ ~(a, b) U 6)(a, b). Thus, 92/e2(a, b) is subdirectly
irreducible by Theorem 4.1.2. Moreover, we have A = Na#b ~(a, b), hence 92
is a subdirect product of the 92/~(a, b).
But now for the promised proof of the existence of q ( a , b). We consider
the following property ~Pof subsets of A × A. S has [P iff 6)(S) does not contain
(a, b/. All we have to show is that [P is of finite character. For then by Tukey's
Lemma, maximal sets exist. By Proposition 1.2.6,
6)(S) = U(6)(S0) : So C_ S, So finite },
and so the claim is obvious. []
We will close this discussion of decomposition by a criterion of decompos-
ability into a direct product.
DEFINITION 4.1.4. Let 92 be an algebra. 92 is called directly reducible if
there exist algebras ~3 and ~ such that ~B # 1 and ~C # 1 and 92 ~- ~3 × C. If
92 is not directly reducible it is called directly irreducible.
Let 92 be an algebra and 6), • E Con(92). 6) and q2 are said to p e r m u t e
if6) o ~ = ~ o 6 ) . In this case, 6) U ~ = 6 ) o ~ .
THEOREM 4.1.5. An algebra 92 is directly reducible iff there exist congru-
ences 6) and ~P, both different from XTA and AA, such that (i.) 6) [~ • = V A,
(ii.) 6) [~ • = AA, (iii.) 6) and • permute.
PROOF. Suppose 92 is directly reducible. Then there exist algebras ~3
and C, such that ~B > 1 and ~C > 1, and an isomorphism h : 92 -+ ~ × C.
We may actually assume that 92 -- ~3 × ft. Let Pl and P2 be the canonical
projections from ~3 × ~ onto ~3 and C. These are surjective homomorphisms.
Let 6 ) : = ker(pl) and • := ker(p2). Then 6) # VA since ~3 is not isomorphic
to 1. Also, 6) # AA since C is not isomorphic to 1. Likewise it holds that
q2 is different from VA and A A. Now (xo, xl}6)(yo,yl} iff x0 = Y0 and
(Xo, Xl}kO(yO,Yl) i f f x l = Yl. It is easy to see that 6) o k~ = ~ o 6) = VA,
showing (i.) and (iii.), and t h a t O N • = AA, showing (ii.). Now a s s u m e

conversely t h a t (9 and • are nontrivial congruences of 92 satisfying (i.), (ii.)
and (iii.). Let h e : 92 -~ 92/(9 and h~, : 92 --~ 9.1/k~ be the n a t u r a l m a p s
with kernel (9 and ~ , and let h := ( h e , he} : 92 -+ 92/(9 x 92/~ be the m a p
defined by h(a) := ([a]O, [a]qd}. This m a p is a well-defined h o m o m o r p h i s m .
We have to show t h a t h is bijective. (1.) h is injective. Let h(a) = h(b).
T h e n [a](9 = [b](9 and [a]~ = [b]~, hence [a]((9 M ~ ) = [b]((9 M ~ ) . By (ii.),
a = b. (2.) h is surjective. Let ([a](9, [b]kg} C A / ( 9 x A / ~ . Since (9 o ~ = VA,
by (i.) and (iii.), there exists a c such t h a t a ( g c q J b . T h e n [c](9 = [a](9 and
[clOd = [b]~ and so h(c) = ([a](9, [b]~}, as required. []
An algebra has p e r m u t i n g c o n g r u e n c e s if all nontrivial congruences
p e r m u t e pairwise. A class of algebras is c o n g r u e n c e p e r m u t a b l e if all its
m e m b e r s have p e r m u t i n g congruences. T h e following is due to A. I. MALCEV
[155].
THEOREM 4.1.6 (Malcev). Let V be a variety of algebras. ~2 is congruence
permutable iff there exists a term p(x, y, z) such that for all 92 C V and all
a, b E A
p~ (a, a, b) = b, p~ (a, b, b) = a .
PROOF. Assume t h a t there exists a t e r m p(x, y, z) with the p r o p e r t i e s
given above. T h e n let a and b be elements such t h a t a (9 o ~ b. T h e n there
exists a c such t h a t a (9 c • b. Hence
a = p ~ ( a , b , b ) k~p~(a,c, b) ( g p ~ ( c , e , b ) = b.
So a ¢2 o (9 b. Now assume t h a t all m e m b e r s of ~2 have p e r m u t i n g congruences.
T h e n in particular the algebra 92 := ~:rv ({x, y, z}) freely generated by x, y and
z has p e r m u t i n g congruences. Let (9 := (9((x,y}) and • := (9((y,z}). T h e n
(x, z) E ( 9 o ~ , and so (x, z) E ~o(9. Hence there exists an element u such t h a t
x q u(9 z. This element is of the form pV~(x, y, z) for some t e r n a r y t e r m f u n c -
tion p. We claim t h a t x = p ~ ( x , y , y ) and y = p ~ ( x , x , y ) . This is enough
to show the theorem. To t h a t end, consider the canonical h o m o m o r p h i s m
h e . Modulo an isomorphism, h e : ~ r v ( { x , y , z } ) ---, ~ r v ( { x , y } ) : x ~ x , y F-+
y , z ~+ y. P u t ~ := ~ v ( { x , y ) } . h . ( p ~ ( x , y , z ) ) = p ~ ( x , y , y ) = h~,(x) = x,
since h~,(u) = h , ( x ) . So, p ~ ( x , y , y ) = x holds in the algebra freely gen-
erated by x and y. To see t h a t the equation holds in any algebra, let
be any algebra in V, and let a , b E D be elements. Let j : ~ --~ ©
be the unique h o m o m o r p h i s m satisfying j ( x ) = a and j ( y ) = b. T h e n
p g ( a , b , b ) = j ( p ~ ( x , y , y ) ) = j ( x ) = a. Hence the first e q u a t i o n holds in
all algebras. Analogously for the second equation, using the congruence (9
instead. []
An algebra is said to be c o n g r u e n c e distributive if the lattice of con-
gruences is distributive.
4.1. More on Products 171
THEOREM 4.1.7. An algebra is congruence distributive if there is a ternary

termfunction re(x, y, z) such that for all a, b E A:
re(a, a, b) -- re(a, b, a) = re(b, a, a) = a.
PROOF. First of all, from lattice theory we get 0 R • C_ O R (~ U ~ ) and
O m ~ _C O ~ ( O U ~ ) , so that ( O R ~ ) U ( O R ~ ) _C OR(~U~2). For the converse
inclusion assume (a,b) C (9 ~ (~ U ~). Then a O b and there is a sequence ci,
i < n + 1, such that
a = Co ~ c l ~c2 ~ c a . . . ~ c , ~ _ l ~c,~ = b
We then have re(a, ci, b) ® re(a, ei, a) = a and re(a, ci, b) 0 re(b, ci, b) = b for
all i so that
re(a, ci, b) 0 a 0 b 0 m(a, ci+l, b)
and re(a, c~, b) • re(a, Ci+l, b) for even i. Similarly for odd i it is shown that
re(a, ei, b) ~g re(a, ci+l, b). Therefore re(a, el, b) ((9 ~ ~) rn(a, ci+l, b) for even i
and re(a, ci, b) ((9 • ~ ) re(a, ci+l, b) for odd i. Thus (a, b) E ((9 N ~) U ((9 N
• ). []
For example, let P2 be an algebra in which there are termfunctions ~, U
such that {A, n, U) is a lattice. Then 2[ is congruence distributive. Namely,
take
re(x, y, z) := (x u y) n (y u z) n (z u x).
This termfunction satisfies the above criterion. If in addition there is a
termfunction ~ such that (A, R, U, ~} is a boolean algebra, then 9~ has per-
muting congruences. For take
p(x,y,z):=(xnz)u(xny'nz')U(x'ny'nz).
If y -- z, this reduces to p(x, y, y) = (x R y) U (x R y') U (x' R y' ~ y) = (x ~ y) U
(x R y') = x; if x -- y this gives p(x, x, z) = (x R z) U (x R x'[q z') U (x'R x ' R z) =
(x R z) U (x' N z) ----z. It also follows that subalgebras, homomorphic images
and products of similar congruence distributive algebras are again congruence
distributive, if the same t e r m can be chosen in all algebras. This is the case
with modal algebras.
COROLLARY 4.1.8. The variety of modal algebras has permuting congru-

ences and is congruence distributive.
Congruence distributivity is important in connection with reduced pro-
ducts. Let I be an index set, I ] i e I ~ i be a product. Let F be a filter on 21
and let (gF be the set of all pairs (a, b) such that {i: a(i) = b(i)} E F. (gF is
a congruence. This is left as an exercise. Congruences of this form are called
filtral. We define the F - r e d u c e d product of the ~ i by
I-[ i := (1-[
F i6I
If F is an ultrafilter, we speak of an u l t r a p r o d u c t . Given a class iK of

algebras, Up(K) denotes the closure of :K under ultraproducts. Let us note
t h a t if F C_ G then there is a surjective h o m o m o r p h i s m I-IF ~3i -+ 1-[c ~ - T h e
following t h e o r e m is due to BJARNI J()NSSON [111], also known as Jdnsson's
Lemma.
THEOREM 4.1.9 (J6nsson). Let :K be a class of algebras and ~? = HSP(iK)
be a congruence distributive variety. If 91 C V is subdireetly irreducible then
P.1E H S U p ( ~ ) .
PROOF. Let h : ~3 --- 9.1 for ~ ~-+ 1-[~1 ~i- (This is not necessarily a
subdirect product.) T h e n put • :-- h - I [ A A ] ; this is a congruence. Moreover,
is R-irreducible in ~3. For if (91 M (92 = ~ we have
((91 (92)/¢ = (91/¢ n (92/¢ = zxA
in Con(92). This implies (91/~ = AA or (92/~ = AA, which is n o t h i n g b u t
(91 = ~ o r (92=¢5.
For subsets S of I let (gs denote the congruence induced on 1-Ii~I ~i by
the principal filter 1" S = {T : S C_ T C I}. F u r t h e r m o r e , let D be the set
of all S _C 1 s u c h t h a t (gs I B _C ~, t h a t is, • = ~ U ( ( g s I B). Choose
U to be a m a x i m a l filter contained in D. T h e n (gu = U((gs : s E u ) , and
so (gg I B C_ ~. All there is to be done is to show t h a t U is an ultrafilter
over I. For then the m a p q3 -+ 1-Iu~i has the kernel (gu I B. So this
induces an e m b e d d i n g 98/((9u I B) ,-~ l~u ¢.i. Since (gu [ B C ~, there is a
h o m o m o r p h i s m ~ / ( ( g v [ B) -+ 92.
Now observe t h a t if S, T C_ I then
®SuT [ B = ((gs [ B) M ((gT [ B).
Therefore we have for S, T E D
= • u I B) = (¢ u ((gs I B)) u ((gT I B)).
Since ¢ is m-irreducible in Con(~) we have either ¢ = ¢ U ((gs I B ) or
= • U (®T [ B). And so we conclude t h a t
If S U T C D then S C D or T E D.
Furthermore
If S E D and S C_ T then T C D.
F r o m these properties we can derive t h a t U is an ultrafitter. For if not, there
is a set S such t h a t neither S E U nor I - S E U. T h u s there are sets K , L E U
such t h a t S N K • D and ( I - S) M L ~ D. We show the existence of K ; the
existence of L is proved analogously. Suppose t h a t ~ C D. T h e n D = 2 z and
U is an ultrafilter by construction. Now assume ~ ¢ D. Consider the s y s t e m
of sets V := {T : T D_ S A K , K C U}. If all S A K E D, we have a s y s t e m of sets
which is a filter containing U and fully contained in D. In particular, S G V,
4.1. More on P r o d u c t s 173
contradicting the m a x i m a l i t y of U. So there is a K C U such t h a t S N K ~ D.

Likewise, we have an L E U such t h a t ( I - S) N L ~ D. P u t M : = K N L. U
is a filter, so M C U. Thus also M E D. Now M = ( S A M ) U ( ( I - S) N M )
but neither set is in D. This is a contradiction. []
By the fact t h a t m o d a l algebras are congruence distributive we can now

infer t h a t a subdirectly irreducible algebra in the variety generated by some
class of m o d a l algebras ~2 is an image of a subalgebra of an u l t r a p r o d u c t
from 9(. Let us return, however, to the criterion of subdirect irreducibility.
First of all, a congruence O on a boolean algebra 92 defines a h o m o m o r p h i s m
h e : 21 ~ 9.1/® with kernel ®. P u t F e : = h~)1(1). This is a filter, as is easily
checked. A n d we have a C F iff a ® 1. Moreover, suppose t h a t a O b. T h e n
he(a) = he(b) and so h e ( a ~ b) = 1, whence a ++ b C F . In other words, fil-
ters are the congruence classes of the top elements and t h e y are in o n e - t o - o n e
correspondence with congruences on the algebra. Let 92 be a boolean algebra.
We have seen in Section 1.7 t h a t the m a p f : O ~-~ F o = {a : a O 1} is a
o n e - t o - o n e m a p from the lattice of congruences of 9.1 onto the lattice of filters
on 92. Furthermore, if F is a filter, then OF defined by a OF b iff a ++ b C F
is the inverse under f . In m o d a l algebras, we have to take open filters rather
t h a n filters. Recall from Section 3.1 t h a t a filter is open if it is closed under
(tim.) If a C F then I j a E F
We have shown in L e m m a 3.1.5 t h a t open filters indeed correspond to ho-
momorphisms. A n y intersection of open filters is again an open filter, so the
open filters form a complete lattice, with the join F U G defined as the smallest
open filter containing b o t h F and G.
THEOREM 4.1.10. The map f : ® ~ Fe : = {a : a O 1} is an isomorphism

from the lattice Con(92) of a modal algebra 9.1 onto the lattice of open filters
on 92, with inverse F ~+ ® r : = {(a,b} : a ++ b E F } .
In particular, if C C A then the smallest open filter containing C, denoted

by {C}, can be obtained as follows.
(C> = { d : (3m)(~C0 C~in C)(d > • n Co)}.
[f c is finite, then put c : = n C. T h e above condition then reduces to the

:ondition t h a t d > [~c for some c o m p o u n d m o d a l i t y [~.
We now wish to characterize subdirectly irreducible m o d a l algebras. We
know t h a t 9./is subdirectly irreducible iff it has a m o n o l i t h O. O is principal,
say ® = ®(a, b). Now, in terms of filters this means t h a t there exists a unique
minimal open filter in the set of open filters ¢ {1} and this filter is generated
by a single element o (for example o = a ~ b). To say t h a t the filter generated
by o is minimal is to say t h a t every other filter generated by an element a 7~ 1
174 4. U n i v e r s a l A l g e b r a a n d D u a l i t y Theory
must contain o, which in turn means that there must be a compound modality
[] such that o ~ []a. Thus we obtain the following criterion of [170].
THEOREM 4.1.11 (Rautenberg). A modal algebra is subdirectly irreducible

i f / t h e r e exists an element o # 1 such that for every a 6 A - {1} there is a
compound modality [] such that o > []a. Such an o is called an o p r e m u m of
91.
For a K r i p k e - f r a m e [ the question of subdirect irreducibility of the algebra

ffYta([) of all subsets of f has a rather straightforward answer. We w a r n the
reader t h a t this t h e o r e m fails in general. This is discussed in Section 4.8.
THEOREM 4.1.12. For a K r i p k e - f r a m e [ the algebra 9~a([) of all subsets

of f is subdirectly irreducible iff ~ is rooted.
PROOF. Suppose t h a t ~ is rooted at x. T h e n take o := f - {x}. We claim

t h a t o is an opremum. So take a set a C_ f such t h a t a ~ f . T h e n there
exists a y C f such t h a t y ~ a . By assumption, there is a k E w such t h a t
a finite p a t h exists from x to y. T h e n x ~ []a for some c o m p o u n d m o d a l i t y
[9. T h u s • a C_ o. By T h e o r e m 4.1.11, 9~ta([) is subdirectly irreducible. N o w
assume t h a t there is no single point x which can generate I. T h e n there exist
two n o n e m p t y sets S and T such t h a t for X : = Tr(S) and Y : = T r ( T ) ,
f = X U Y , but X # f and Y # f . Suppose there exists an o p r e m u m , o.
Since m X D_ X , we must have o _D X . Similarly o D Y, and so o D X O Y : f ,
a contradiction. []
E x e r c i s e 135. Show t h a t the filtral congruence (gF defined above is indeed

a congruence.
E x e r c i s e 136. Show t h a t a boolean algebra is subdirectly irreducible iff it

is isomorphic to 2. Hence BA = HSP(2).
E x e r c i s e 137. Suppose t h a t @ = (G, 1, -1 , .) is a group and (9 a congruence

on G. Show t h a t the congruence class of the unit I must be a n o r m a l subgroup.
Furthermore, suppose t h a t [a](9 is given. T h e n [1](9 -- a - 1 - [a](9 = {a - 1 . b :
b C [a](9}. Thus, show t h a t there is a o n e - t o - o n e correspondence between
congruences on ® and normal subgroups.
E x e r c i s e 138. Let Z2 be the additive group of integers m o d u l o 2. T h a t is,

we have
0 1 I -I
0 0 1 0 0
1 1 0 1 1
Show that the latticeof congruences of the group ~ x ~2 is not distributive.
Hint. Use the previous exercise.
4.2. Varieties, Logics and Equationally Definable Classes 175
Exercise 139. A vector space 93 over a field F can be made into an alge-
bra by adding a unary function 0r for each r E iF. Its action is defined by
Or(v) -= r.v, where r . v is the usual scalar multiplication. (Why is this compli-
cated definition necessary?) Again, a congruence defines a normal subgroup.
Moreover, this subgroup must be closed under all Or. Show t h a t there is a
o n e - t o - o n e correspondence between congruences on 93 and subspaces.
Exercise 140. Continuing the previous exercise, show t h a t a vector space is

subdirectly irreducible iff it is one-dimensional.
4.2. Varieties, Logics and Equationally Definable Classes

Two of the most fundamental theorems of universal algebra are due to
BIRKHOFF of which the first states that a class of algebras is definable by
means of equations iff it is a variety, that is, closed under H, S and P. The
second gives an explicit characterization of all equations that hold in a variety
which is defined by some given set of equations. We will prove b o t h theorems
in their full generality and derive some important consequences for modal
logics. To start, let f~ be a signature, L a language for f~ and t(~), s(~)
two terms based on the variables xi, i < n. An f~-algebra 92 satisfies the
equation s(~) ~ t(£'), written 9_] b s(~) ~ t(~), if for all d C_ A, s a ( g ) = ta(d).
Furthermore, V ~ s(e) ~ t(e) if for all 92 E V, 9_] b s(e) ~ t(e). Often we
write s ~ t, dropping the variables.
PROPOSITION 4.2.1. The class of all algebras which satisfy an equation
s ~ t is closed under taking products, subalgebras and homomorphic images.
The proof of this theorem is routine and left as an exercise. We will prove
here that the converse is true as well. Let 3? be a variety, and and X a set of
variables. Then put
Eqx(V ) := {(s(e),t(e)) : 3? b s(e) ~ t(e),£ c x}
In case that X = {xi : i C aJ} put Eq(3?) := Eqx(3? ). Let E be a set of
equations over X. Define AIg(E) to be the class of algebras satisfying E.
By Proposition 4.2.1, AIg(E) is a variety. Moreover, for any class K of ft-
algebras we always have K C_ AIgEq(K). What we have to show is that for
a given variety 3? we have 3? = Alg Eq(3?). There is a way to restate this
using free algebras. Recall from Section 1.3 that an algebra ~rv(Y) is said
to be a freely Y-generated algebra if for every algebra 9.1 C 3? and every map
v : Y -+ A there is a homomorphism ~ : ~rv(Y) -+ 92 such that ~ I Y = h.
We have seen in Theorem 1.3.5 that a variety has free algebras for all sets
Y. Moreover, Alg Eq(3?) has free algebras because they can be obtained from
the term algebras. Namely, if ~m~(Y) is the algebra of Y-terms over the
language/5 with signature f/, define a congruence (9 by s(~) O t(~) iff for all
9[ C 3? and all g C_ A we have s~(5) = t~(d). Then E r n a ( Y ) / O is freely
176 4. U n i v e r s a l A l g e b r a a n d D u a l i t y T h e o r y
generated by Y in AlgEq(~?). For let v : Y --+ A be a map, 9ff an algebra over

A. T h e n there is a unique extension ~ : i r o n ( Y ) --+ 92. Let s(~) O t(aT). T h e n
~(s(~)) = s~(v(#.)) = t~(v(~)) = ~(t(~)), by definition of (9. Hence there
exists a unique homomorphism g : Emn(Y)/{9 --+ 92.
PROPOSITION 4.2.2. For any variety "V, ;~rv(X) is a subdirect product of

the ~ r v ( E ) , E a finite subset of X .
PROOF. Let E be a finite subset of X. We let 92(E) be the subalgebra

of ~ v ( X ) generated by the terms x~, xi • E. It is easy to see t h a t 92(E) is
isomorphic to ;~rv(E). Let ~E be a m a p ~E : X -+ E such that t~E I E = idE.
This m a p can be extended to a homomorphism g E : ~ r v ( Z ) -~ a r v ( E ) . ( T h a t
this m a p is onto follows from Theorem 1.3.6.) Let F be the collection of all
finite subsets of X. We have
N ker(-~E) = A .
EEF
For if ~ v ( X ) > s(a3) ~ t(aT), then let E consist of the variables in ~. Now
a r v ( E ) > s(2) ~ t(~) and thus -~E(S(~)) = s(2) ~ t(2) = gE(t(~)). (Here we
write s(2) also for the equivalence class of the t e r m s(aT) in the free algebras.)
By Proposition 4.1.1, ~ r v ( X ) is a subdirect product of the ; ~ v ( E ) . []
COROLLARY 4.2.3. For every variety V, ~2 = HSP(;~v(R0)).
Now, if we are able to show that E m a (X)/(9 is also the freely X - g e n e r a t e d

algebra of V, we are obviously done. For then the classes V and Alg Eq(V) con-
tain the same countably generated free algebras. Now recall the construction
of a free algebra from the algebras of V from Section 1.2.
PROPOSITION 4.2.4. For a class 9C, and a set X
• sP(JC).
PROOF. Let Q := {(s(aT),t(aT)): a~c_ X , K ~ s ~ t}. For each e < Q we
pick a witness algebra 92~. This means that there is a sequence g in 92~ such
that s 9~' (d) ~ t ~' (g). Thus we have a m a p v~ : X ~ A such t h a t v~ : x~ ~ a~,
i < n. This extends to a homomorphism ~ : 2~m~(X) --+ 92~. Now let
h : E m ~ ( X ) --+ 1-[~cQ 92~ be the canonical h o m o m o r p h i s m defined by all the
v~. We show now that ker(h) is exactly Eqx(~: ). For if e = s(aT) ~ t ( J )
holds in all algebras, it holds in their product as well, and so e E ker(h).
However, if it does not hold in all algebras of ~ , then h(s(~)) # h(t(~)), for
~e(S(~)) = s ~" (g) ¢ t ~ (g) = ~(t(aT)), as had to be shown. []
THEOREM 4.2.5 (Birkhoff). A class of ~-algebras is definable by means

of equations over a language L of signature ~ exactly if it is a variety.
In addition, an equational theory Eq(~) can actually be identified with a

particular kind of congruence on ~ma(X), the so-called fully invariant con-
gruence.
DEFINITION 4.2.6. A congruence (9 on an algebra 92 is called f u l l y i n -

v a r i a n t if it is compatible with all endomorphisms of 92. That is, (9 is fully
invariant if whenever h : 92 --+ 92 is an endomorphism and a 0 b then also
h(a) (9 h(b).
Our aim is to show that all modal logics are theories of certain classes of
algebras. In order to do this, we need to be explicit about what equations
can be derived from other equations. The following axiomatization is due to
BIRKHOFF. Unlike in propositional calculi, we derive equations from sets of
equations and not terms from sets of terms. We write F F-v t ~ s if t ~ s
can be derived in finitely many steps by applying one of the following rules in
addition to (ext.), (mon.) and (trs.) of Section 1.3.
(V1) k-vs~s
(V2) s~t~-vt~s
(V3) s~t;t~u~-vs~u
(V4) so ~ t 0 ; . . . ;sn-1 "~tn-1 }-, f ( S o , . . . , S n - 1 ) ~ f ( t o , . . . , t n - z )
s(x0,... ,xn-1) t(xo,...
s(u0,.., t(u0,...
(V1), (V2) and (V3) are axioms of pure equality. (V4) is known as the re-
p l a c e m e n t r u l e , (V5) as the s u b s t i t u t i o n rule. Notice that this calculus
also satisfies (sub.) and (cmp.), the latter by the fact that it is a calculus
defined by finitary rules, so it is finitary by the nature of a proof. In (V4),
f is any n - a r y term function. However, it can be shown that it is enough
to require the validity of (V4) only for the basic functions, fi, i C I. Now
take a set F of equations. Consider the congruence on ! r n a ( X ) induced by
F. The condition of reflexivity of (9 corresponds to (V1), the symmetry to
(V2) and the transitivity to (V3). Finally, (V4) corresponds to the compati-
bility with all functions. Hence, the calculus without substitution defines the
smallest congruence induced by F. Now a substitution is nothing but an en-
domorphism of the term algebra so (V5) enshrines the requirement that the
congruence derivable from F is fully invariant.
THEOREM 4.2.7. Let F be a set of equations over L of signature fL The
smallest fully invariant congruence on ~ m a ( X ) containing F is the set of all
s ~ t such that F ~-v s ~ t.
Let us call a set of the form Eq(:X) an e q u a t i o n a l t h e o r y . Then we have

COROLLARY 4.2.8 (Birkhoff). A set of equations is an equational theory
iff it is closed under the rules of ~-v.
PROOF. F i r s t of all, t h e rules (V1) - (Vh) are correct. T h a t is, given a n

a l g e b r a 92 a n d given a rule, if 92 satisfies every p r e m i s s of a rule t h e n 92 also
satisfies t h e conclusion. ( V 1 ) . For all v : X ~ A we have g ( s ) = g ( s ) for
a n y t e r m s. ( V 2 ) . A s s u m e t h a t for all v : X --* A, ~(s) = g ( t ) . T h e n for
all v : X -~ A a l s o , ( t ) = ~(s), showing 92 ~ t ~ s. ( V 3 ) . A s s u m e t h a t
92 ~ s ~ t ; t ~ u. Take a m a p v : X --~ A. T h e n ~(s) = ~(t) as well as
g ( t ) = ~(u), from which g ( s ) = ~(u). T h u s 92 ~ s ~ u. ( V 4 ) . A s s u m e t h a t
9.1 b si ~ ti for a l l i < n. Take v : X --+ A. T h e n ~ ( f ( s 0 , . . . , s n _ i ) ) =
f~(g(s0),... ,g(sn-1)) = f~(g(to),... ,g(t~-i)) = g(f(to,... , tn-i)). Hence
92 ~ f ( s o , . . . , s ~ - i ) ~ f ( t 0 , . . . , f n - i ) . ( V h ) . A s s u m e 92 ~ s ~ t. Define
a s u b s t i t u t i o n cr by cr : xi F-+ ui, i < n, or : xl ~-~ xi for i > n. T h e n
o a : X -+ A, a n d t h e h o m o m o r p h i s m e x t e n d i n g t h e m a p is j u s t v o c~,
since it coincides on X w i t h g o ~ r . N o w let v : X -+ A be given. T h e n
g(s(uo,..., un-i)) = g(~(s)) = go~(s) = go~(t) = g(t(u0,..., U~-l)). Thus
92
N o w let r b e any set of e q u a t i o n s a n d 8 its closure u n d e r (V1) t o (Vh).
L e t 92 : = ~ma(x)/r. T h e n if s ~ t ¢ r we have 92 ~ s ~ t. F o r j u s t t a k e
t h e c a n o n i c a l h o m o m o r p h i s m h o : E m a ( X ) --+ 92 w i t h kernel (9. Since (9 is
a congruence, this is w e l l - d e f i n e d a n d we have ho(s) ¢ ho(t), as r e q u i r e d .
N e x t we have to show t h a t 92 ~ F. To see t h a t t a k e a n y e q u a t i o n s ~ t E F
a n d v : X ~ A. Since A i s g e n e r a t e d b y t e r m s over X m o d u l o ®, we c a n
define a s u b s t i t u t i o n ~ such t h a t ~ = h o o ~. N a m e l y , p u t or(x) : = t(~),
w h e r e t(ff) E h;X(h(x)) is freely chosen. T h e n ho(cr(x)) = v(x), as r e q u i r e d .
Now ~ ( s ) ® ~ ( t ) , by closure u n d e r s u b s t i t u t i o n , so t h a t ~ ( s ) = ho(-~(s)) =
ho(-~(t)) = ~(t). T h u s 92 ~ F. []
THEOREM 4.2.9. There is a one-to-one correspondence between varieties

of f~-algebras and fully invariant congruences on the freely countably generated
algebra. Moreover, V1 C V2 iff for the corresponding congruences 01 D 02.
One h a l f of this t h e o r e m is a c t u a l l y T h e o r e m 2.2.9. T h e converse d i r e c t i o n

was a c t u a l l y much h a r d e r to prove b u t m a k e s t h e result all t h e m o r e useful.
C o n s i d e r now w h a t this m e a n s for m o d a l logic. (We will h e n c e f o r t h w r i t e
a g a i n p a n d q for variables i n s t e a d of x a n d y, as well as ~ a n d ~ for f o r m u l a e
i n s t e a d of s a n d t.) W e have seen earlier for p r o p o s i t i o n a l f o r m u l a e p a n d
t h a t 92 ~ ~ ~ ~ iff92 ~ p ~ ¢ ~ Y. T h e l a t t e r is n o t h i n g b u t 92 ~ ~ ~ ¢ , now
r e a d in t h e s t a n d a r d sense. T h u s t h e e q u a t i o n a l t h e o r y of a class K of m o d a l
a l g e b r a s c a n be i n t e r p r e t e d as t h e logical t h e o r y of a class. W e will show
t h a t if we have a class of m o d a l algebras, t h e n t h e set of e q u a t i o n s ~ ~ -2
in t h a t class is a m o d a l logic, a n d conversely. So t h e two a l t e r n a t i v e w a y s
to specify classes of a l g e b r a s - - n a m e l y v i a e q u a t i o n s a n d v i a logical a x i o m s
- - coincide. T h e e q u a t i o n a l t h e o r y of p o l y m o d a l a l g e b r a s is as follows. T h e
p r i m i t i v e f u n c t i o n s y m b o l s are here t a k e n to be T , 2 , 7, A, V a n d t h e m o d a l
operators rqj. The following equations must hold.
pAp ~ p pVp ~ p
pAq ~ qAp pVq ~ qVp
pA(qAr) ~ (pAq) Ar pV(qVr) ~ (pVq) Vr
pA(qVp) ~ p pV(qAp) ..~ p
pA(qVr) ~ (pAq) V(pAr) pV(qAr) ~ (pVq) A(pVr)
pAT ~ p pV± ~ p
~(pAq) ~ (~p) V(~q) ~(pVq) ~ (~p) A(=q)
pA(=p) ~ ± pV(~p) ~ T
~(~p) ~ p ~T ~ ±
l]j(pAq) ~ (~jp) A (~jq) I]]jT ~ T
We call this set of equations Mal. The first five rows specify that the algebras
are distributive lattices; then follow laws to the effect t h a t there is a top and
a b o t t o m element, and that there is a negation. Finally, there are two laws
concerning the box operators. We define 10 -+ X by (910) V X, 10 ++ X by
(10 A X) V ((910) A (=X)) and (>j10 by ~[]d=10.
PROPOSITION 4.2.10. (i) Mal; 10 ~ X ~-y 10 ++ X ~ T. (ii) Mal; 10 ~+ X ~

Tkv 10~X.
PROOF. The proof will be a somewhat reduced sketch. A full proof would
consume too much space and is not revealing. The reader is asked to fill in
the exact details. We perform the proof only to show how the calculus works
in practice. (i) Mal; 10 ~ X F v 10 V (~X) ~ X V (~X); 10 V (710) ~ X V (~10), by
applying (Vd). Furthermore, Mal F v 10 V (~10) ~ T; X V (~X) ~ T. Applying
(V2) and (V3) we get Mal;10 ~ X ~-v !oV (~X) ~ T ; x V (~10) ~ T. So,
Mal;10 ~ X F v (10V ( ~ x ) ) A ( x V ( ~ T ) ) ~ T A T , by (Vd). Now by (V5) we
have Mal t-v T A T ~ T; thus Mal;10 ~ X F-v (W V (~X)) A (X V (710)) ~ T.
Applying the distributivity law twice we arrive at
MaJ; 10 ~ X ~ v (10 A X) V (10 A ( ~ @ ) V ((=X) A X) V ((~X) A (910)) ~ V

We can replace 10A -~10 as well as (~X) A X by J_ and drop b o t h occurrences of
/ from the disjunction. Commutativity of A yields Mal; 10 ~ X ~ v (10 A X) V
((~9~)A(~x)) ~ T, the desired result, i. e. 10 ~ X ~ T. (ii) (We will now write
10 ~ X rather than Mal ~-v 10 ~ X.) Assume (10Ax)V ((~10)A (~X)) ~ T. Then
10A((9~Ax)V((~cp)A(~x))) ~ 10AT. Now 10AT ~ 10and so we have 10A((10Ax)V
10. Distributing 10 we get 10.
From there with associativity ((10 A ~) A X) V ((10 A (~T)) A (~X)) ~ 10 which
results in (10Ax) V ( L A (~X)) ~ X, by (Vd) with 10A10 ~ 10 and TA (~10) ~ ±.
This gets reduced to 10 A X ~ ~. Similarly, one can derive ~ A X ~ X. So,
Mal; 10 ~ ¢ ~. T ~-v 10 ~ X- []
Let F be a set of equations. We put Th(F) := {p : F ~ y ~ ~ T}. Now

let A be a modal logic. Then we define Eq(A) := {p ~ ¢ : ~ ~ ¢ C A}. T h e
proof of the next theorem is left as an exercise.
THEOREM 4.2.11. Let F be a set of equations, A be a set of formulae.

Then the following holds.
1. Th(F) is a normal modal logic.
2. Eq(A) is an equational theory of modal algebras.
3. Th Eq Th(r) = T h ( r ) .
4. Eq Th Eq(A) = Eq(A).
COROLLARY 4.2.12. There is a dual isomorphism between the lattice of

normal t~-modal logics and the lattice of varieties of normal modal algebras
with ~ operators.
This is a considerable strengthening of Proposition 2.2.7, L e m m a 2.2.8

and T h e o r e m 2.2.9. For now we do not only know that different logics have
different varieties associated with them, we also know that different varieties
have different varieties associated with them. The relation between equational
calculi and deductive calculi has been a topic of great interest in the study of
general logical calculi, see [29]. WIM BLOK and DON PIGOZZI have tried to
isolate the conditions under which a mutual translation is possible between
these two deductive formulations of a logic. It would take us too far afield to
discuss these developments, however.
Notes on this section. In a series of papers WOLFGANG RAUTENBERG
partly together with BURGHARD HERRMANN have studied the possibility to
export an axiomatization of a variety in the Birkhoff-calculus to a H i l b e r t -
style proof system for the logic determined by some unital semantics over
that variety, see [102], [174] and [173]. The equational rules present no
problem, likewise the rule of substitution. However, the rule of replacement
is not straightforward; it may lead to an infinite axiomatization. (See next
section on t h a t theme.) Therefore, the so-called finite replacement property
was defined. It guarantees t h a t adding a finite set of instances of the rule of
replacement will be sufficient for validity of all rule instances. It is a theorem
by ROBERT C. LYNDON [142] that the equational theory of any 2-element
algebra is finitely axiomatizable. In [102] it has been shown t h a t all varieties
of 2-element algebras have the finite replacement property. It follows t h a t
the logic of any 2-element matrix is finitely axiomatizable. For 3-element
algebras b o t h theorems are false.
E x e r c i s e 141. Prove Proposition 4.2.1
E x e r c i s e 142. Prove Theorem 4.2.9

4.3. Weakly Transitive Logics II 181
4.3. W e a k l y T r a n s i t i v e Logics II
In this section we will prove some connections between purely algebraic
notions and properties of logics. Again, weakly transitive logics play a funda-
mental role. The following definition is due to T. PRUCNAL and A. WROI~SKI
[166].
DEFINITION 4.3.1. Let ~- be a consequence relation. A set A(p,q) :=
{51 (p, q): i E I} is called a set of equivalential t e r m s f o r ~- if the following
holds
(e¢) A(p,p)
(eq2) A(p,q) F- A(q,p)
(eq3) A(p, q); A(q, r) [- A(p,r)
(eq4) Ui<.(/) A(pi,qi) ~- A(f(p~, f(q~)
(eq5) p;A(p,q) F- q
is called equivalential if it has a set of equivalential terms, and f i n i t e l y
equivalential if it has a finite set of equivalential terms. If A (p, q) ----{5(p, q)}
is a set of equivalential terms for F- then 5(p, q) is called an equivalential
t e r m for ~.
Let us investigate the notion of an equivalential logic for modal logics.
Clearly, for any modal logic A, ~-A is always equivalential; a set of equivalential
terms is the following.
A(p, q ) : = {m(p +4 q): [] a compound modality} .
Moreover, I~-Ais always finitely equivalential; p +4 q is an equivalential term for
I~-A. Thus, the only remaining question is whether f-A is finitely equivalential.
Note that if ~-A is finitely equivalential it also has an equivalential term. For
if A(p, q) = {5~(p, q) : i < n} is a finite set of equivalential terms for ~-A then
5(i9, q) := Ai<n 5i(p, q) is an equivalential term.
PROPOSITION 4.3.2. Let A be a modal logic and A(p, q) a set of equiva-
lential terms for A. Then the following holds.
1. A(p, q) ~-A P +4 q
2. p ++ q I~-A A(p, q)
3. A(p, q) [F-A A(p ++ q, T)
4. A(p ++ q, T) tf-A A(p, q)
PROOF. (1.) follows from (eq5) and (eq2) with the deduction theorem.
For (2.) note that p ++ q I~-A A(p,p) ++ A(p,q) (by Proposition 3.1.7). By
(eql) we get p ++ q lt-A A(p,q), as desired. To prove (3.) observe that
p ++ q lt-A A(p ++ q, T), again by Proposition 3.1.7. Since we have established
that A(p, q) I~-Ap ++ q, the third claim follows. For (4.) observe that, by (10,
A(p +4 q, T) I~-A p +4 q and that p ++ q I~-A A(p, q) (by (2.)). []
Now let A be a modal logic and E a set of formulae. Let E ~- := {¢ : E t-A ~b}
and E i~ := {¢ : E N-A ¢}. Consider the set A(E,-F) e where
A(~],T) :--- {~(~,T): ~ C A , ~ • ~]}.
This set is closed under (mp.). We show that it is closed under (mn.). Consider
first ¢ E N. We clearly have A ( N , T ) t-h A(tb, T) and so by (eq4) also
A(N, T) i-A A(E~¢, []T) for all compound modalities. Since m T is a theorem,
it can be substituted by T. By (eqh), A(EB¢, T) ~- FR¢. So, A(N, T) ~- Fq¢
for all N and ¢ E N. This shows that A ( N , T ) t-A N~N. From this we
get A(N, T)N-A N. Hence, A(E, T) F-A []~A(N, T), by (2) of the previous
theorem, since Z t~-h A(N, T). Now let A(N, T) ~-h ~ for some ~. Then
E]jA(E, T) [-A [Zj~. Since A(E, T) ~-h [3jA(E, T), we have succeeded to
show that [~jy) E A(E, T) u. Hence A(E, T) ~ _D 2 ~. The converse inclusion
is a consequence of (2) of Proposition 4.3.2.
PROPOSITION 4.3.3. Let A be a modal logic and A(p,q) a set of equiv-
alential terms. Then for any set
A(E, ~-)~ = Ei~ .
DEFINITION 4.3.4. A variety 17 has equationally definable p r i n c i p a l

c o n g r u e n c e s ( E D P C ) if there exists a number n E w and terms si(w, x, y, z),
t i ( w , x , y , z ) , i < n, such that for every 94 C 17 and a,b,c,d E A
(c,d) C ®(a,b) iff s~i (a,b,c,d) = t~(a,b,c,d) for all i < n .
PROPOSITION 4.3.5. A variety 3? of modal algebras has E D P C if] there
exists a term u(x, y) such that for all 92 E 3? and elements a, b C A, b is in the
open filter generated by a iff u ~ (a, b) = 1.
PROOF. Suppose that 3? has EDPC, and let si(w, x, y, z) and t~(w, x, y, z),
i < n, be terms defining principal congruences in 3?. Then put
:= A si(x, T , y , T) ++ ti(x, T , y , T) .
i<n
Then u~(a,b) = 1 ifffor a l l / < n we have s ~ ( a , T , b , T ) = t ~ ( a , T , b , T ) iff

(b, T} C O(a, T) iff b is in the open filter generated by a. Conversely, let u(x, y)
be a term such that for all algebras 92 E 3? b is in the open filter generated
by a iffu~(a,b) = 1. Then let n = 1, s o ( w , x , y , z ) := u(w e+ x , y ~ z) and
t o ( w , x , y , z ) := T. Let a,b,c,d E A, 92 C 3?. Then s~o (a,b,c,d) = tao (a,b,c,d)
iff u~(a +~ b,c ~ d) = 1 iff c ~ d is in the filter generated by a ~ b iff
(c, d) e e ( a , b). []
We say that u(x, y) d e f i n e s p r i n c i p a l o p e n filters in 3? if for all 92 C 3? and
a, b C A we have u ~ (a, b) = 1 iff b is in the open filter generated by a. V has
d e f i n a b l e p r i n c i p a l o p e n filters ( D P O F ) iff there exists a u(x, y) defining
open filters. By the previous theorem, a variety of modal algebras has E D P C
iff it has definable open filters. The following theorem has been obtained in
[28].
THEOREM 4.3.6 (Blok ~: Pigozzi & Khhler). For any normal modal logic
A the following are equivalent:
1. [-h is finitely equivalential.
2. AlgA has DPOF.
3. Aid A has EDPC.
4. I[- h admits a deduction theorem.
5. A is weakly transitive.
PROOF. We have shown earlier that (4.) ~=~ (5.) and we have shown in
Proposition 4.3.5 that (2.) ~=> (3.). We show that (1.) ~ (2.) ~ (4.) and
(5.) ~ (1.). Assume (1.). Then there exists an equivalential term 5(p, q) for
[-A. Now put u(p,q) := 5(p,T) --+ q. Let 92 E AlgA and a E A. Then by
Proposition 4.3.3 the set F := {b : b > 5~(a, T)} is the open filter generated
by a. So b C F iff u~(a,b) = 1. Hence u(p,q) defines principal open filters.
Hence (2.) is proved. Now assume (2.). Suppose that u(p, q) defines principal
open filters in Aid A. We claim that u(p, q) satisfies a deduction theorem for
rF-A. Namely, let A be a set of formulae, and let ~ and ¢ be formulae. Let 92
be a A-algebra, F an open filter in 92. We can actually assume that F = {1}.
Then A F-(~,F ) u(~, ¢) iff for every valuation/3 such that ~[A] C {1} we have
u ~ ( ~ ( T ) , ~ ( ¢ ) ) = 1 ifffor all valuations/3 such that ~[A] _ {1}, 3 ( ¢ ) is in the
open filter generated by ~(T) iff for every valuation/3 such that ~[A; ~] C_ {1}
we also have ~ ( ¢ ) = 1 iff A ; ~ F-(~,F ) ¢. Thus A ; ~ I~-h ¢ iff A i~-h U(~,¢).
This shows (4.). Finally, assume (5.). Let A be weakly transitive with master
modality [~. Then [~(p ++ q) is an equivalential term for F-A. []
DEFINITION 4.3.7. Let 9.1 be an algebra. A ternary termfunction t(x, y, z)

is called a t e r n a r y d i s c r i m i n a t o r t e r m f o r 91 if the following holds
t~(a,b,c ) = { c ifa=b
a otherwise
Let V be a variety. V is called a d i s c r i m i n a t o r v a r i e t y if there exists a class

:K of algebras such that ~ is the least variety containing • and there exists a
term t(x, y, z) which is a discriminator term for all 92 E K.
We remark here that except in trivial cases a discriminator for an alge-

bra 92 cannot be a discriminator for 92 × 92. This is why the definition of a
discriminator variety is somewhat roundabout.
PROPOSITION 4.3.8. Let 92 be an algebra, and t(x, y, z) a ternary discrim-

inator for 92. Then 92 is simple.
PROOF. Let ® # AA be a congruence. Then there exist a , b E A such

t h a t a # b and a ® b . Then
a = t~(a,b,c) ® t~(a,a,c) = c
Hence O = V A , and so 9.1 is simple. []
PROPOSITION 4.3.9. Let 9~ be a class of algebras, and assume that t(x, y, z)

is a discriminator term for all algebras of •. Then it is a discriminator term
for all members of HSOp9~.
PROOF. Let t(x, y, z) be a discriminator t e r m for 9.1; then it is obviously
a discriminator t e r m for every subalgebra of ~. It is also a discriminator t e r m
for every homomorphic image of 9.1, for the only images up to isomorphism are
9g and the trivial algebra. Finally, let ~ be an ultraproduct of ~ i , i E I, with
ultrafilter U over I. Then the congruence Ou is defined as in Section 4.1.
We write ~u instead of [~]O~. Let ~,b,~ C ~ i c z Ai. Assume t h a t the set
D defined by D := {i : ai = bi} is in U. Then a u = bu and moreover
D C_ {i : t~(a~,bi,ci) = ci}, whence t~(-Su,-bu,-~u) = -Su. Now assume
that D • U. Then ~u ~ bu, and D C {i : t ~ ( a i , bi,c~) : ai}. Thus
t ~ (~u, b~, ~ ) = ~ . []
In the remaining part of this section we shall be concerned with the re-
lationship between three properties of a variety of modal algebras: being a
discriminator variety, being semisimple and being weakly transitive and cyclic.
Semisimplicity is defined as follows.
DEFINITION 4.3.10. An algebra is called s e m i s i m p l e if it is a subdirect
product of simple algebras. A variety is called s e m i s i m p l e if it consists en-
tireIy of semisimple algebras.
PROPOSITION 4.3.11. Let V be a congruence distributive variety. Then if
V is a discriminator variety, ~ is semisimpIe.
PROOF. Suppose V is a discriminator variety. Then it is generated by a
class ~ of simple algebras. If ~ is subdirectly irreducible, it is by J6nsson's
Theorem contained in HSUp ~ . By Propositions 4.3.9 and 4.3.8, ~ is simple.
[]
It can be shown in general that a discriminator variety is congruence dis-

tributive, so that the previous theorem actually holds without assuming the
congruence distributivity of V. (See exercises below.) Varieties of modal al-
gebras have however already been shown to be congruence distributive, so we
do not need to work harder here.
We shall now prove that a semisimple variety of modal algebras is weakly
transitive on condition that it has only finitely many operators. Let V be a
semisimple variety of n - m o d a l algebras, ~ finite. We assume t h a t there is
a m o d a l i t y K]i such t h a t V ~ [~iP ~ P. This makes life a little bit easier.

Obviously, given this assumption V is weakly transitive iff it satisfies the
equation []kx = [ ] k + l x for some k E w. For a simple 92 C V we p u t
f~t:={ccA:c~land []-c=0}.
We call c d e n s e i n 92 if it is in t2~. For - [] - c is the closure of c. Denote by

Vs the class of simple algebras from ~?. Obviously, one of the following must
hold for our variety V:
(A) (Vn e c0)(~92 • ~ s ) ( ~ c • ~92)([]nc > 0 and []'~ (c -4 []c) ~ c),
(B) (3n • w)(g92 • Vs)(Vc • a ~ ) ( [ ] n c > 0 implies []n (c -4 []c) _< c).
We will refer to V as being of type A or of type B depending on which of the
above holds. In case B obtains, we shall denote the least n u m b e r such t h a t B
holds for V by N .
DEFINITION 4.3.12. Let 92 be a modal algebra, n < R0. Call c • A d e e p

i n 92 if for all m • w we have []<-m+lc < ~<--mc.
Since ~ c < c by our assumptions, c is deep if [ ] m + l c < ~ m c for all

m. Obviously, it is enough to require this to hold for almost all m. For if
[]m+lc = []mC for some m, then equality holds for almost all m. Now, using
the u l t r a p r o d u c t construction we can easily show the following:
LEMMA 4.3.13. Suppose that V is not weakly transitive. T h e n there exists

an algebra in V containing a deep element.
LEMMA 4.3.14. For every c E 92 such that []kC = 0 there is a dense b > c.
PROOF. To see this, ]et b : = - gain - c , where m is m a x i m a l with the

p r o p e r t y []m __ C > O. Then, [] - b = [ ] m + l c = O, and c < - gain - c = b, as
required. []
LEMMA 4.3.15. Let k • w. If for all 92 • Vs and all c • gl~ we have

[]kc = O, then V satisfies ~ k + l x = []kx.
PROOF. Let ~ • Vs. T h e n by the previous lemma, for any n o n - u n i t

element c of A there is a dense b such t h a t c _< b. It follows t h a t 0 = gakb >_
[]kc. H e n c e w e h a v e ~ k c = O f o r a l l c • A - { 1 } . Thus, 9 2 ~ k + 1 x = [ ] k x . []
LEMMA 4.3.16. If92 • VS, and V is semisimple of type B. I f c is deep in

92 then
zN((a c -4 -4 -4 < c
PROOF. Suppose t h a t c is deep. T h e n ~ k c > [~k+lc for all k. Now let

m be given. T h e n ~ m c -4 g]'~+lc belongs to f~9~, for [] - (~'~c -4 []m+lc) =
[]"~CA[] -- ~ m + l c = ~m+lcMY~ -- [~m+lc < g~m+lC~ -- []m+l C = 0. Moreover,
[~;~N([]rn __~ ~ m + l c ) ~ [~N []m+l C > []N+m+2c, since c is deep. Thus, since
is of type B,
-+ -+ < c.
[]
THEOREM 4.3.17 (Kowalski). If V is semisimple, then ~ is weakly tran-

sitive.
Suppose for contradiction t h a t V is semisimple and not weakly transitive.

W i t h o u t loss of generality we m a y assume t h a t it does not satisfy [ ] n + l x --
~ x , for any given n C w. Now, for any n, we take a simple algebra 92,~
falsifying [ ] n + I x = []~x. Then, by L e m m a 4.3.15 there is a cn C An such
t h a t []nCn > 0 and [] - cn = 0, t h a t is, c~ E ~ .
Now, let U be a nonprincipa] ultrafilter over w. P u t 92 : = I-Ine~ 92~, and
c : = (c~ : n C w)/U. T h e n for all n E w we get [ ~ c > [ ] n + l c > 0 and
[] - c = 0. So, c is b o t h deep an dense.
Obviously, 92 is a subdirect p r o d u c t of subdirectly irreducible algebras,
which, by our assumption, are also simple. We will derive a c o n t r a d i c t i o n
from this. More precisely, we will derive a contradiction from the a s s u m p t i o n
t h a t all subdirectly irreducible m e m b e r s of H(92) are simple.
Consider the congruence O = O(c, 1) on 92. By the choice of c, our (9 is
neither the diagonal nor the full congruence. As (9 is principal, there m u s t be
a congruence II covered by (9. W i t h the choice of H, our reasoning splits into
two cases, one for either of the two types.
Type A. If ~? is of type A, then, in 92 we have []n(c --+ [~c) ~ c for all

n C a;; hence (9(e -~ []c, 1) is strictly below ®. We choose a H -< O from the
interval I[0(c --+ []c, 1), 0]. Since the lattices of congruences are algebraic
there always is one.
Type B. If V is of type B, then we just choose any H -< ®. This case has
one feature t h a t deserves to be spelled out as a separate fact.
LEMMA 4.3.18. Let ~? be of type B, and m E w. If q~ C Con(92) is a

congruence satisfying []mc q~ ~m+l C then • > (9.
PROOF. By the construction, c is deep in 92. It follows from L e m m a 4.3.16

that
_< c .
Now, if (I) E Con(92) satisfies []'~c • []m+l C, then ~ N ( ( ~ C --+ []'~+IC) --+
[]([]me ---~ [~mq-lc)) (~ 1, which implies c (I) 1. So, • > (9. []
Now we return to the main argument. We shall develop it for b o t h types

together, splitting the reasoning only when necessary. Take the set of congru-
ences
F := {55 ¢ Con(A) : 55 > II and 55 2~ ® } -

Let 62 := LJF; by congruence distributivity, 62 E F. T h e n 92/62 is subdi-
rectly irreducible. Observe t h a t we cannot have F = {II}, for in such a case
92/62 = 92/II, and this would be subdirectly irreducible b u t n o n - s i m p l e , since
® • V. Thus, since 62 :~ 0 , we obtain t h a t (9 U 62 is full (which by congru-
ence p e r m u t a b i l i t y equals @ o 62). For otherwise 92/62 would be a n o n - s i m p l e
subdirectly irreducible algebra in V. Now, as 92/62 = (92/II)/62, and since
principal congruences remain principal in h o m o m o r p h i c images, we can shift
the whole a r g u m e n t to 92/II.
Let ~ := 92/II. As ® and 62 are above II in Con(92) we shall write @ and
62 in place of @ / I I and 62/II from Con(~). Thus, for instance, we will say
t h a t (9 (and not @/II) is an a t o m of Con(~). We shall also write c instead
of [c]n.
Now we have an algebra %; a principal congruence (9 E Con(~3) such t h a t
(9 >- 0; further, there is a congruence 62 - - which is n o n - p r i n c i p a l in general
- which is the largest congruence not containing (9, and (9 o 62 = 1. Hence,
-
there is an element d E B - {1} with (d, 1) E @ and (d, 0) E 62. T h e latter is

equivalent to ( - d , 1) E 62. It follows t h a t d > Nmc for some m E w. Moreover,
as a ~ / H C_ a ~ , we obtain t h a t c ( = [c]ri) E a ~ . Thus, c E a ~ - {1}.
T h e s t a t e m e n t ( - d , 1) E 62 can be further broken up as follows: first,
62 = U{55 c @(92): 0 < 55 _< ~ } ,
where @(92) stands for the set of all c o m p a c t congruences of 92 (which are
also the principal congruences of 92). Thus, each 55 above is of the form
a5 = (9(b, 1) for some b ¢ 1. Let C C B be the set of all such b E B. Passing
from congruences to open filters we obtain: - d E [UbEc[N~b:n E w)).
Secondly, C is a downward directed set (in fact, a filter, but we do not
need that). For take b 0 , . . . ,bk-1 E C, and let b = b0 n . . . N bk-1. Then,
(9(b, 1) = (9(b0, 1) V . . . V @(bk_~, 1), and all the congruences on the r i g h t -
hand side of the equation are below 62 by definition. Thus, so is (9(b, 1);
hence b E C, as needed. Moreover, all b C C s a t i s f y V n E co : Nnb ?: a;
otherwise some congruence below • would contain (9, which is impossible.
Thirdly, we have: - d C [UbEc[Nnb: n C co)) iff - d > do A . . . N dk-1,
with di E [[]~bi : n E w) (0 < i < k - 1). Since [] distributes over meet, this
gives
-de [[]n(bon...nb~_~):nCco),
and by the previous a r g u m e n t bo n -.. M bk-1 C C.
G a t h e r i n g all this together, we get - d C [N~b : n E co) for some b E B
such t h a t Vn E co : ~]~b ~ c. On the other hand, d C [~]~c : n E co). Therefore,
d >_ [~mc for some m, and - d > • k b , for some k. Thus, []'~c < d < - ~ k b;
in particular, []'~c <_ - • k b.
Consider q := - • k b ~ • m c = N]kb V Nmc. As q > • m c we have:
O(q, 1) < O ( • m c , 1) = {~(c, 1) = O .
Since 0 is an atom, O(q, 1) is either 0 or ®.

Let us first deal with the case (~(q, 1) = 0 = O(c, 1). We t h e n have:
[N~q < c, for some r. This yields: c > • ~ ( N k b V • ' ~ c ) > N~+kb. T h u s ,
c >_ • ~ + k b , which is a contradiction.
T h e remaining possibility is ®(q, 1) = 0. T h e n we have q = 1, and
t h a t means - • k b _< • ' ~ c . Together with the inequality from the previous
p a r a g r a p h , this implies - • k b = • ' ~ c . Two cases arise.
Type A. By the choice of H we have t h a t c -+ [~c H 1 in 91, a n d so

c rl [] c. Hence c = []c in ~S, and we get []kb = - c . Further, since c is dense,
we have [] - c = 0, from which we get []k+lb = O. However, (b, 1) E ~ , and
so 0 = []k+lb k~ 1. Hence ¢2 is full. Contradiction.
Type B. If either [] - []kb = [ ] ' ~ + l c < []'~c, or [] - N'~c = []k+lb <

[]~b, then the congruences (9(5, 1) and O(e, 1) have a n o n - t r i v i a l intersection;
namely, b o t h contain the pair ([]m+lc V ~k+lb, 1). This pair is n o t in A.
This c a n n o t happen, for then (9 [7 ~ > 0, which contradicts the definition of
q'. So, we obtain t h a t []'~+1c = []mc and [] -- []mc = - []m c in %. Hence,
[]'~+1c H []'~ c in 91, and since %? is of t y p e B, we can apply L e m m a 4.3.18
to get t h a t H >_ 6}. This, however, c a n n o t h a p p e n either, since H -~ O, by its
definition. This completes the proof of T h e o r e m 4.3.17.
Recall from Section 2.5 t h a t A is called cyclic if for every basic m o d a l
o p e r a t o r [3~ there exists a c o m p o u n d m o d a l i t y [] such t h a t p -+ [Di ~ p E A.
If A is weakly transitive with master m o d a l i t y [] this is equivalent to t h e
requirement t h a t [] satisfies $5.
LEMMA 4.3.19. (~ < i% 0 Let A be a modal logic. Then if A is weakly

transitive and cyclic, AIg A is a discriminator variety and semisimple.
PROOF. Assume t h a t A is weakly transitive with m a s t e r m o d M i t y [].
First we show t h a t AIg A is semisimple. Let 91 be an algebra and let a C A.
Assume t h a t a is open, t h a t is, a = []a. We claim t h a t - a is also open.
Namely, from a _< []a we conclude t h a t - [] - ( - a ) < - a . Therefore
-a < [](- []-(-a)) < [] - a _< - a .
Hence [] - a = - a . Thus - a is open. Now assume t h a t 91 is not simple. T h e n

there exists a proper open filter F such t h a t OF is not the monolith. It is easy
to see t h a t F can be assumed to be generated by a single element a. By weak
transitivity, we can also assume t h a t a is open, and t h a t F = {b : b >_ a}.
Then G := {b: b > - a } is also an open filter. Moreover, F A G = {1}. Hence

OG ~-](~F = AA. Hence 91 is directly decomposable. Thus, AIg A is semisimple.
Now put t ( x , y , z ) := m(x ~ y) A z. V .-~ [] (x ++ y) A x. We claim that
t(x, y, z) is a discriminator term. To that effect, take a subdirectly irreducible
algebra. It is simple, by since A l g a is semisimple. So, 0 is an opremum and
[]a = 0 iff a ¢ 1. Let a = b, and c be any element. Then tm(a,b,c) =
[](a~b) V/c.U.-m(a<-4b)Aa=c. Let n o w a C b . Thena++b¢ 1 and
[](a ++ b) = 0. Thus t a ( a , b , c ) = a. This shows that AlgA is a discriminator
variety. []
THEOREM 4.3.20 (Kowalski). (n < ~0.) Let A be a modal logic. Then the
following are equivalent:
1. A cyclic and is weakly transitive.
2. AIg A is semisimple.
3. AIg A is a discriminator variety.
PROOF. By L e m m a 4.3.19 (1) implies b o t h (2) and (3), and by Proposi-

tion 4.3.11 (3) implies (2). It remains to be shown that (2) implies (1). So,
suppose that AIg A is semisimple. By Theorem 4.3.17 A is weakly transitive.
W h a t remains to be shown is that if AIg A is semisimple, A must be cyclic.
Since A is weakly transitive and n finite, there is a maximal compound modal-
ity, []. Suppose that A is not cyclic. Then p --+ 7H~ [] ~p ¢ A. Hence there
exists a subdirectly irreducible algebra 91 such that 91 ~ p -4 ~ [] ~p. So
there exists a b E A such that b A - [] - [] - b # 0. Consider the open filter F
generated by - b . F = {c : c > [ ] - b } , by the assumption t h a t [] is a strongest
compound modality. Suppose that • - b = 0. Then - • - [ ] - b = -•-0 = 0,
a contradiction. So, F ¢ A. Suppose that [] - b = 1. Then, since [] - b ~ - b ,
- b = 1. It follows that b = 0, again a contradiction. So, F # {1}. We
conclude that 92 is not simple. This contradicts our assumption. So, A must
be cyclic. []
Notes on this section. The complex of ideas surrounding the property of

equationally definable principal congruences has been studied in the papers
[28], [26], [30] and [27] by WIM BLOK, DON PIGOZZI and PETER KOHLER.
In [26] it is proved that if a congruence-permutable variety is semisimple it
has E D P C iff it is a discriminator variety.
E x e r c i s e 143. Let 91 be a modal algebra. Denote by ~p(91) the semilattice

of compact congruences. Show that if Th 91 is weakly transitive and cyclic
then ¢p(91) is the semilattice reduct of a boolean algebra.
E x e r c i s e 144. Show that a finite tense algebra is semisimple.

E x e r c i s e 145. An algebra 91 is called h e r e d i t a r i l y s i m p l e if every subal-

gebra of 92 is simple. Show that every simple modal algebra is hereditarily
simple.
E x e r c i s e 146. Show that a discriminator variety is congruence distributive.

Hint. First show that it has permuting congruences. T h e n show t h a t the
variety is congruence distributive.
4.4. Stone Representation and Duality

This section provides the rudiments of representation theory and duality
theory. For a proper understanding of the methods it is useful to learn a
bit about category theory. In this section, we provide the reader with the
essentials. More can be found in [143], [101] or [80]. The basic notion is
that of a category. A c a t e g o r y is a structure C = (Ob, Mot, dom, cod, o, id)
where 0 b is a class, called the class of o b j e c t s , Mot another class, the class of
m o r p h i s m s , dora, cod : Mot ~ 0 b two functions assigning to each m o r p h i s m
a domain and a c o d o m a i n , o : Mot x Mot ~ Mot a partial function, assigning
to suitable pairs of morphisms their c o m p o s i t i o n , and id : Ob --~ Mot a
function assigning to each object a morphism, the identity on t h a t object.
We write f : A ~ B or A ~ B to state t h a t f is a morphism with domain A
and codomain B. We also say that f is an a r r o w from A to B. We require
the following.
(1.) For morphisms f, g the composition f o g is defined iff cod(g) = d o m ( f ) .
(2.) For every object A, dom(id(A)) = cod(id(A)) = A.
(3.) For every morphism f : A --* B we have f o id(A) = f and id(B) o f = f.
(4.) Iff:A~B,g:B~Candh:C-~Dthenho(gof)=(hog)of.
For example, take the class of sets with the functions defined as usual, this
gives rise to a category called Set. To give another example, let L be a
language of signature f~ and T an equational theory over f~, then the class
AIg T of f/-algebras for T form a category with the morphisms being the f~-
homomorphisms. An arrow f : A -~ B is an i s o m o r p h i s m if there is a
g : B -I+ A such that g o f = id(A) and f o g = id(B). If f : A -~ B is
an isomorphism, A and B are said to be i s o m o r p h i c . A pair A = ( O , M ) ,
where O is a class of objects and M a class of morphisms of C such t h a t for
each p C M we have dom(p), cod(p) E 0 is called a d i a g r a m . A diagram
c o m m u t e s if for any three arrows A ~ B, B -~ C and A h C of M we have
h=gof.
4.4. Stone Representation and Duality 191
h=go~~g/
A ,B
DEFINITION 4.4.1. Let C be a category. Put 0 b °p := 0b, Mor °p := Mor,

id °p := id; moreover, put dom °p := cod and cod °p := dom, and finally f o °p
g := g o f . Then C°p defined as C°p := (0b °p, Mor °p, dora °p, cod °p, o °p, id °p)
is called the d u a l or o p p o s i t e category.
The dual category arises by reversing the direction of the arrows. So if for
example f : A --+ B is a map, then there is a dual m a p from B to A in
C°p, which is usually also called f. The notation is rather unfortunate here
since it uses the same name for the objects and for the arrows and only
assigns a different direction to the arrow in the opposite category. So, the
arrow exists in the category as well as in the dual category and which way it
goes is determined by the context. To remove this ambiguity we will write
fop : B --+ A to distinguish it from the corresponding arrow f : A -+ B of C.
DEFINITION 4.4.2. Let C and © be categories, F both a map from the

objects of ~ to the objects of © and from the morphisms of C to the morphisms
of ~3. Then F is called a c o v a r i a n t f u n c t o r if
1. F ( d o m ( f ) ) = d o m ( F ( f ) ) , F ( c o d ( f ) ) = c o d ( F ( f ) ) ,
2. F ( g o f ) = F(g) o F ( f ) and
3. F ( i d ( A ) ) = i d ( F ( A ) ) .
F is called a c o n t v a v a v i a n t f u n c t o v if
1. F ( c o d ( f ) ) = d o m ( F ( f ) ) , F ( d o m ( f ) ) = c o d ( F ( f ) ) ,
2. F ( g o f ) = Y ( f ) o F(g) and
3. F ( i d ( A ) ) = i d ( F ( A ) ) .
Thus, a covariant functor maps f : A --+ B into F ( f ) : F ( A ) --4 F ( B )

while a contravariant functor maps f into F ( f ) : F ( B ) -+ F ( A ) . There-
fore, as the direction of the arrows is reversed, the composition must also
work in reverse order. Now, an essential feature of modal duMity theory is to
find well-behaved functors between the various categories t h a t arise in modal
logic. Typically, in the standard algebraic tradition one would be content
to define just representation theories for the objects of t h a t category (e. g.
of finite boolean algebras as algebras of subsets of a set). However, from
the categorial perspective the most natural representation is that which is in
addition also functorial. The advantage is for example that the characteriza-
tion of modally definable classes of algebras (namely varieties) is transferred
to a characterization of modally definable classes of frames. For we do not

only know something about the objects, we also learn something about the
maps between them. An instructive example is the representation theorem for
modal algebras. There is a function mapping the modal algebras to descrip-
tive frames and functions between modal algebras to p-morphisms; in other
words we have a functor between the respective categories. We will show
that this functor is contravariant and has an inverse. So, every p - m o r p h i s m
between frames is the image of a homomorphism between the corresponding
algebras under the functor. This has far reaching consequences. A similar
example is Theorem 4.4.10 of this section. If F is a functor from C to © and
G a functor from © to £ then G o F is a functor from C to £. G o F is covariant
if F and G are both covariant or both contravariant; G o F is contravariant if
exactly one of F and G is contravariant.
The following construction will be a major tool in representation theory,
both in this section and in Section 7.4. For the category C let home(A, B)
denote the class of arrows from A to B. A category C is called l o c a l l y
s m a l l if home(A, B) is a set for all objects A, B. All categories considered
in this book will be locally small. If C is locally small then for every object
A the map h o m e ( - , A) can be turned into a contravariant functor HA from
C into the category Set of sets and functions. Namely, for an object B we
put HA(B) := home(B,A) and for a function f : C --+ B we put H A ( f ) :
HA(B) -+ HA(C) : g ~+ g o f. This is well defined. It is a functor; for let
e : D --+ C, f : C -+ B and 9 : B --+ A. Then H A ( f o e) : 9 ~-~ 9 o ( f o e)
and Ha(e) o HA(f) : g ~+ (9 o f) o e. These two functions are the same, by
definition of a category. Furthermore, if f = id(B) then HA(f) = id(HA(B)),
as is easily verified. This shows that we have a contravariant functor.
PROPOSITION 4.4.3. Let ¢ be a locally small category and A an object in
e. Define H A by HA(B) := home(A,B) and H A ( f ) : HA(B) --+ HA(C) :
g ~ f ° 9 , where f : B ~ C is an arrow inC. Also, define HA byputting
HA(B) := home(B,A) and H A ( f ) : HA(C) --+ HA(B): g ~ 9 ° f . Then H A
is a covariant functor from C into Set and HA a contravariant functor from
C into Set.
Usually, H A is called the c o v a r i a n t h o m - f u n c t o r and HA the c o n -
t r a v a r i a n t h o m - f u n c t o r . Let us apply this to boolean algebras. By BA
we denote the category of boolean algebras with boolean homomorphisms.
It is a locally small category as is easily verified. By Proposition 1.7.10 an
ultrafilter U on an algebra 92 defines a homomorphism fu : 92 --~ 2 by putting
fv(a) = 1 iff a C U. Conversely, for every map f : 92 --~ 2 the set f - 1 ( 1 ) is an
ultrafilter on 92. Finally, every map f : 92 -~ 2 is surjective, since f ( 1 ~) = 1
and f ( 0 ~) = 0. We will henceforth call maps 92 -~ 2 p o i n t s of the algebra 92.
The contravariant hom-functor H2 will now be denoted by ( - ) . . To give the
reader a feeling for the construction we will repeat it in this concrete case.
We denote the set of points of a boolean algebra 92 by 92,. Moreover, for any
algebra ~ and boolean h o m o m o r p h i s m f : ~ -+ 92 the m a p f , : 92, -+ ~ , is
defined by f , (g) : = g o f .
f
,92
Let us investigate this a bit closer. We will show t h a t f is surjective iff f , is

injective; and t h a t f is injective iff f , is surjective. Assume f is surjective.
Let g and h be two two different points of 92. T h e n there is an a such t h a t
g(a) • h(a). Therefore, there is an element b E f - l ( a ) with go f(b) ¢ ho f(b).
Hence f , : P2, ,--, ~ , . Now assume t h a t f is not surjective. T h e n let im[f]
be the direct image of f in 92. Since this is not the whole algebra, there is
an element a such t h a t neither a nor - a is in im[f]. Now take an ultrafilter
U on 92. Let V : = UNim[f] be the restriction of U to im[~]. T h e n V
is an ultrafilter on im[f] (can you see why this is so?). Now b o t h the set
V [2 {a} and V [2 { - a } have the finite intersection p r o p e r t y in 92, hence can be
extended to ultrafilters Y 1 and V 2. T h e n f - l [ U ] = f-l[V1] = f-l[y2], and
so f , is not injective since V1 ~ V2. This concludes the case of surjectivity of
f. Now assume t h a t f is injective. We will show t h a t f , is surjective. So let
g : ~ -~ 2. P u t F : = f[g-~(1)], the direct image of the ultrafilter defined by g.
This generates a filter in 92 and can be extended to an ultrafilter. Hence f , is
surjective. Now let f be not injective. T h e n there exist elements a and b such
that a ~ b but f(a) = f(b). This means f(a ++ b) = 0, but a ++ b ~ 0. T h e n
either a N - b ¢ 0 or b n - a ¢ 0. Hence there is an ultrafilter containing one
but not the other. This ultrafilter is not of the form f - 1 [U] for any ultrafilter
o n 92.
THEOREM 4.4.4. H2 : 92 ~-~ 92, is a contravariant functor from the cate-

gory of boolean algebras to the category of sets. Moreover, f, is injective iff f
is injective, and f, is surjective iff f is injeetive.
Now let us try to define an inverse functor from Set to 8A. Take the set
2 -- {0, 1} and look at homset(X, 2). It is not difficult to t u r n this into a
boolean algebra. Namely, observe t h a t for a function f : X ~ 2 each fibre
f - l ( 1 ) is a subset of X , and every subset A C X defines a function )CA by
XA(X) = i iff x G A. This function is known as the c h a r a c t e r i s t i c f u n c t i o n
of A. Now define
0 := Xo
1 :-- XX
--XA :z XX-A
X.A ('~ X B :~-- )(.AAB
X A l.j X B :z XAOB
So, for X we let X* be t h e b o o l e a n a l g e b r a defined on t h e set homset(X, 2).
A n d for f : Y -+ X let if(g) : = g o f . f
Y ,X
f * ( g ) := g
THEOREM 4.4.5. The map X ~ X* is a contravariant functor from the

category of sets into the category of boolean algebras. Moreover, f* is injec-
tire iff f is surjective and f* is surjective iff f is injective, f* is called the
powerset functor.
T h e p r o o f is left to t h e reader. Now, we have a c o n t r a v a r i a n t f u n c t o r f r o m BA
to Set a n d a c o n t r a v a r i a n t f u n c t o r from Set to BA. C o m b i n i n g t h e s e f u n c t o r s
we get a c o v a r i a n t functor from BA to BA a n d from Set to Set. F o r e x a m p l e ,
s t a r t i n g w i t h a b o o l e a n a l g e b r a 92 we can form t h e set 92. a n d t h e n t u r n
this into a b o o l e a n a l g e b r a (92.)*. ( T h e l a t t e r o p e r a t i o n we s o m e t i m e s refer
to as raising a set into a b o o l e a n algebra.) Likewise, we c a n s t a r t w i t h t h e
set X , raise this to a b o o l e a n a l g e b r a X* a n d form t h e p o i n t set ( X * ) . .
U n f o r t u n a t e l y , in n e i t h e r case we can e x p e c t to have a n i s o m o r p h i s m . T h i s
is a n a l o g o u s to t h e case of vector spaces a n d t h e i r biduals. I n t h e finite
d i m e n s i o n a l case t h e r e is a n i s o m o r p h i s m b e t w e e n a v e c t o r s p a c e a n d its
b i d u a l , b u t in t h e infinite d i m e n s i o n a l case t h e r e is only a n e m b e d d i n g of t h e
former into t h e l a t t e r . T h e s a m e holds here. T h e m a p x ~-~ U~ = {Y C_ X :
x C Y} e m b e d s X into t h e p o i n t set of X*. T h e m a p a ~-~ 3 = { f : 92 -+ 2 :
f(a) = 1} e m b e d s 92 into t h e p o w e r s e t - a l g e b r a over t h e p o i n t s of 92.
T h u s t h e s i t u a t i o n is n o t o p t i m a l . Nevertheless, let us p r o c e e d f u r t h e r
along t h i s line. F i r s t of all, if i n t u i t i v e l y b o o l e a n logic is t h e logic of sets,
t h e n in a way we have succeeded, b e c a u s e we have shown t h a t a n y t h i n g t h a t
satisfies t h e laws of b o o l e a n a l g e b r a s is in effect a n a l g e b r a of sets. W e c a s h
out on this as follows. Let 92 be an algebra, X a set. We call a b o o l e a n
h o m o m o r p h i s m f : 92 ~-~ X* a r e a l i z a t i o n of 92. A r e a l i z a t i o n t u r n s t h e
e l e m e n t s of t h e a l g e b r a into s u b s e t s of X, a n d i n t e r p r e t s t h e o p e r a t i o n s on
91 as t h e n a t u r a l ones on sets. ( A c t u a l l y , X* was c o n s t r u e d v i a t h e h o m -
functor, so we have on t h e right h a n d S{de t h e c h a r a c t e r i s t i c f u n c t i o n s r a t h e r
than the sets, but this is inessential for the argument here, since the two
algebras are isomorphic.) What we have proved so far is that every boolean
algebra can be realized. Now, in order to recover 92 in X* we do not need
the operations -- because they are standard. All we need is the collection
f[A] = {f(a) : a 6 A}. So, we can represent 92 by the pair (X, f[A]}. There is
a structure in mathematics which is almost of that form, namely a topological
space. Recall t h a t a t o p o l o g i c a l s p a c e is a pair :~ = (X, ~[) where • is
a collection of subsets of X , called the set of o p e n s e t s , which contains
~, X and which is closed under finite intersections and a r b i t r a r y unions.
Maps between topological spaces are the c o n t i n u o u s f u n c t i o n s , where a
function f : X --+ Y is a continuous function from (X, J~} to (Y, Y} if for
every A 6 3f, f - l [ A ] 6 X. Alternatively, since the o p e n sets of ~ form a
locale ~2(2~) : = (X, Q, U } with finite meets and infinite joins, we can say t h a t
a function is continuous if the function f t ( f ) : f t ( ~ ) --+ ft(~) defined by
f t ( f ) ( A ) :-- f - l [ A ] is a h o m o m o r p h i s m preserving finite meets and infinite
joins. Let ~ = (X, J~} be a topological space. A set A C_ X is c l o p e n in the
space ~ = (X, X} if b o t h A and X - A are open. :~ is d i s c r e t e if every subset
is open. ~ is discrete iff for every x 6 X the singleton set {x} is open. X is
called c o m p a c t if for every union I-ix x~ = X of open sets xi we have a finite
subset J C_ I such that Uj x i = X . Finally, a subset B of X is called a b a s i s
of the t o p o l o g y if every open set is the (possibly infinite) union of m e m b e r s
of •.
DEFINITION 4.4.6. A topological space is z e r o - d i m e n s i o n a l if the clo-
pen sets are a basis for the topology.
Let 92 be a boolean algebra. P u t X : = pt(gA). For a 6 A p u t
:={p6X:p(a)---1}.
Let A := {3 : a 6 A}. A is closed under all boolean operations. Now let
X be the set of all unions of members of A. Alternatively, X is the smallest
t o p o l o g y induced by A on X . P u t 9-1o := (X, X}. 9.1o is a zero-dimensional
topological space. Before we set out to s t u d y the functorial properties of this
map, let us t u r n to a f u n d a m e n t a l problem of this construction, n a m e l y how
to recover the set A when given the topology X. T h a t we can succeed is not at
all obvious. So far we have a set X and a collection of clopen subsets forming
a boolean algebra which is a basis for the topology. To show t h a t the clopen
sets are not always reconstructible take X = w, B the collection of finite and
cofinite sets and C the collection of all subsets of X . B o t h B and C are a
basis of the same topology, namely the discrete topology. In this topology,
every subset is clopen.
PROPOSITION 4.4.7. Let ( X , X } be a compact topological space. Assume
No is a basis for the topology and that No is closed under complements and
finite unions. Then for x C X we have x 6 No iff x is clopen.
PROOF. If x is in NO then it is clearly clopen, since X - x E NO as

well. Now assume that x is clopen. Let x = U z y i , x - x = U j z j be
two representations such that yi,zj 6 NO for all i E I and j E J. Now
X = x U (X - x) = (Ui yi) u ( U j zj). Thus there is a finite g C_ I and a finite
L c Z such that Z = (UK Y~) u (UL z~). Then x = x n ((UK Y~) u (UL z~)) =
U K ( x n yd u UL(x n zj) = UK x n y~ = UK Y~ • ~ . []
As before, a map f : ~3 --+ 92 induces a continuous function fo : 920 -+ ~ o

via fo(g) := 9 o f. For first of all we have the set-map f . : 92. --+ ~3. and
we simply define fo(Ui xi) := Ux fo(x~). (The reader is asked to verify that
this definition is consistent. This is not entirely harmless.) This defines the
first functor. For the opposite direction, take a topological space X and define
the set of points by hom-rop(2~, 2), where 2 is the discrete topological space
with 2 elements. (So, 2 = ({0, 1}, {~, {0}, {1}, {0, 1}}).) Such functions are
uniquely characterized by the set on which they give the value 1. So they are
of the form Xx for a set x. x must be open, being of the form f - 1 ( 1 ) , and
closed, being the complement of f-l(0). Thus, we get as before a boolean
algebra of functions X~, namely for all clopen elements. We call this algebra
.~o.
Now, do we have 92 ~ (920)° as well as (~°)o? Let us begin with the first
question. We have defined 920 on the set of points, or - - equivalently - - on
the set of ultrafilters. We will show that this space is compact, allowing us to
recover the original boolean algebra as the algebra of clopen sets. Since this
is the algebra we will get when raising via o, we do in fact have the desired
isomorphism. A space is compact if for any intersection ~lr xi of closed sets
there is a finite J C I such t h a t n J Xi : fZJ. Alternatively, consider a family
of sets {xi : i E I) such that any finite subfamily has n o n - e m p t y intersection.
This is the finite intersection property defined in Proposition 1.7.11. If a space
is compact, such a family must have n o n - e m p t y intersection. Now consider
a set {Si : i E I} of closed sets in 92o with the finite intersection property.
Each Si is an intersection of sets of the form 3. Without loss of generality we
may therefore assume that we have a family (3j : j C J) of elements 3j with
the finite intersection property. Then there is an ultrafilter U containing that
family. There is a function fu such that f - ~ ( 1 ) = U. Hence, fu C nj~@
and so the intersection is not empty. So, 920 is a compact space. Thus, by
Proposition 4.4.7, 92 - (920)°. In general, for a compact topological space
~ (~°)o does not hold. Thus, we must restrict the class of topological
spaces under consideration. This leads to the following definition.
DEFINITION 4.4.8. A topological space is called a S t o n e space if it is

compact and for two different points there is a clopen set containing one but
not the other. The category of Stone spaces and continuous maps between
them is denoted by StoneSp.
Consider this in contrast to the separation axioms To and T2. A topo-

logical space is a Hausdorff space or T2-space if whenever x and y are
distinct there is an open set containing x but not y. A space is a To-space
if for every given pair x, y of points there exists an open set A such that
~(A • {x,y}) = I. (More about T0-spaces in Section 7.4.) These two con-
ditions are not the same. For example, the topological space over 0, 1 with
the sets O, {0} and ~0, I} satisfies the condition that there is an open set
containing 0 but not i; but there is no open set containing 1 and not 0. So it
is a T0-space but not a T2-space. (This space is known as the Sierpifiski-
space.) Now consider requirement in the above definition. It is almost like
the T0-axiom but requires not just an open set but a clopen set. This however
means that it is the same as a T2-axiom with respect to the clopen sets. For
if a contains x and not y and a is clopen, then X - a is clopen as well, and it
contains y but not x.
DEFINITION 4.4.9. Two categories ¢ and © are e q u i v a l e n t if there exist

covariant functors F : C --+ © and G : ~D --+ C such that for each A G Ob(C)
we have A ~- G ( F ( A ) ) and for each B • Ob(©) we have B ~- F ( G ( B ) ) .
THEOREM 4.4.10 (Stone). The category BA of boolean algebras is equiva-

lent to the category StoneSp °p, the dual category of the category of the Stone-
spaces.
PROOF. We know t h a t we have a contravariant functor ( - ) o : 3A --+

StoneSp and a contravariant functor ( _ ) o : StoaeSp -~ BA. These can be
made into covariant functors by switching to the opposite category StoneSp °p.
T h e remaining bit is to show t h a t :~ ~ (:~°)o. Now, for a point x C X p u t
Us : = {a C X : a clopen, x C a}. This is an ultrafilter on the algebra of clopen
sets. We show t h a t the m a p u : x ~-~ Us is a topological isomorphism. It is
injective by the definition of a Stone space; for if x ~ y t h e n there is a clopen
set a such t h a t x G a but y ~ a. The m a p is also surjective, by construction.
Finally, a clopen set of (:~°)o is a set of the form ~ : = {U : a • U}, where U
ranges over the ultrafilters of :~o and a is clopen in :~. Now let a be a clopen
subset of :~. T h e n
u[a] = {ux:x • a} : {Ux:a • Us}

={U:a•U} =3
(Here we use the fact t h a t every ultrafi]ter is of the form Us for some x.)
Hence, u induces a bijection of the clopen sets. Therefore, it induces a bijec-
tion between the o p e n sets. Hence u is a topological isomorphism. []
E x e r c i s e 147. Show t h a t the dual category fop of a category e is a category.

Show also t h a t if F : C ~ © is a contravariant functor, then F °p : C -+ ©op
is a covariant functor, where F ° P ( A ) : = F ( A ) and F ° P ( f ) : = F ( f ) °p.
E x e r c i s e 148. Show Theorem 4.4.5.
E x e r c i s e 149. Show that a topological space is a Stone space iff it is compact

and zero-dimensional.
E x e r c i s e 150. Let J C R be a subset of R, and 5 the set of sets of the form

O N J , O open in R. Now let ~ be the topological space <J, 5). Show t h a t
points in 3 are nothing but Dedekind cuts.
E x e r c i s e 151. Show that intervals I = Ix, y] c R endowed with the relative

topology of R, have no points. Show that Q is not a Stone-space.
E x e r c i s e 152. Let the real numbers between 0 and 1 be presented as 3-

adic numbers, that is, as sequences <a~ : 0 < i E w) such t h a t a C {0, 1, 2}.
Such a sequence corresponds to the real number ~ i = l ai • 3 - i . To make this
correspondence one-to-one, we require t h a t there is no i0 such t h a t a i = 2 for
all j > i0. Now let C C [0, 1] be the set of reals corresponding to sequences
in which no ai is equal to 1. This set is known as the Cantor-Set. Show t h a t
this set, endowed with the relative topology of the real line, is a Stone-space.
* E x e r c i s e 153. Let X be countably infinite. Construct a compact To-space

over X. Moreover, show that no T2-space over X can be compact.
4.5. A d j o i n t F u n c t o r s a n d N a t u r a l T r a n s f o r m a t i o n s
In this section we will prove a representation theorem for frames t h a t
makes use of the topological representation developed by MARSHALL STONE.
DEFINITION 4.5.1. Let C and 9 be categories and F, G : C --+ ~D functors.

Let rl : 0b(C) --+ Mot(©) be a map such that for all C-objects A we have
G(A) = ~?(F(A)) and that for all C-arrows f : A --~ B we have G ( f ) o ~(A) =
~(B) o F ( f ) . Then rl is called a natural transformation from F to G.
The last condition can be presented in the form of a commutative diagram.
F(A)
q(A), G(A)
F(I)[ [a(f)
F(B) , a(B)
It is important in category theory that everything is defined not only with
respect to objects but also with respect to morphisms. The latter requirement
is often called naturalness. Thus, a natural transformation is natural because
4.5. Adjoint Functors and Natural Transformations 199
it conforms with the arrows in the way indicated by the picture. Another
example of this naturalness condition is in the definition of adjoint funetors
given below.
Let C and © be categories, and F, G, H : C --+ © be functors, 77 a natural
transformation from F to G and 0 a natural transformation from G to H.
Define a map 0 • r/: Oh(G) --+ Mor(©) by (0 • rl)(A ) := O(A) o ~(A). This is
a natural transformation. For let A ~ B. Then r/(B) o F ( f ) = G ( f ) o rl(A )
and O(B) o G ( f ) = H ( f ) o O(A) by assumption that r/ and 0 are natural
transformations. Then
(0.~)(B)oF(f) = 0(B) o~(S)oF(f)
= O(B) o G ( f ) o rl(A )
= H(f) o O(A) o~?(A)
= H(f) o (O.rI)(A)
PROPOSITION 4.5.2. Let C and © be categories. Then the functors from
C to © form the objects of a category with arrows being the natural transfor-
mations.
DEFINITION 4.5.3. Let C and © be categories, and F : C -+ © and let
G : ~) -+ C be funetors. F is called left adjoined to G, in symbols F -~ G,
if there exists for every C-object A of C and every ~D-object B a bijeetion
13AB : hom~)(F(A),B) --+ home(A,G(B)). Moreover, /3AB must be natural in
both arguments; this means that for arrows f : A -+ A I and g : B -+ B' we
have
/3ABOHaB(f) = HB(F(f))o~A,B
13AB' o HA(G(g)) = HFA(g) O/3AB
The definition of naturalness is best understood if presented in the form
of a picture. Here is the picture corresponding to the first of these conditions.
/3A,B
hom~)(A',GB) , home(FA',B)
HGB(f)I t HB(F(f))
hom~(A, a B ) , home(FA, B)
/3AB
Assume now that F is left adjoined to G. Then there exists a bijection
~A,FA from hom(FA, FA) to horn(A, GF(A)). In particular, /3A,FA(id(A)) :
A --+ GF(A). The map rI : A ~-+ flA,FA(id(A)) is a natural transformation
from the identity functor on C to the functor GF. Similarly, starting with
a bijection between hom(GB, GB) and hom(FG(B), B) we obtain a natural
transformation from the functor FG to the identity functor on ©. These
two transformations are called the u n i t (7/: le --+ GF) and the c o u n i t (0 :
FG -+ 19) of the adjunction. Moreover, the following so-called t r i a n g u l a r

i d e n t i t i e s hold for all objects A of C and B of ©:
O(FA) oF(~(A)) = id(FA)
G(~(B)) ou(GB) = id(GB)
It can be shown that the existence of a natural transformation r/: 1¢ -+ GF
(the unit) and a natural transformation 0 : FG --+ 1~) (the counit) is enough to
ensure that two functors are adjoint. Namely, consider the following schemata.
g A --+ G(B) f F(A) --+ B
F(g) F(A) -+ FG(B) G(f) GF(A) -+ G(B)
O(B) o F(g) F(A) --+ B G(f) o 7I(A) A -+ G(B)
So, ~3AS : g ~-~ F(g) o ~I(A) is bijective. It is not hard to show that if the
triangular identities hold then these bijections are natural in both arguments.
The existence of a unit and a counit satisfying the triangular identities is
typically somewhat easier to verify.
To understand this terminology, let us look at a particular case. Let V
be a variety of fl-algebras for a given signature ft. They form a category,
which we also denote by 17. F : ~? -+ Set : 92 ~-~ A the functor sending an
algebra to its underlying set, and a homomorphism between algebras to the
corresponding set-map. F is called the f o r g e t f u l f u n c t o r . Consider the
functor G : Set --+ ~? : X ~-> ~ v ( X ) , sending a set X to the algebra freely
generated by X in ~?. We call it the free f u n c t o r . We claim that G q F.
For let f : X -+ F(92) be a map. Then, by definition of a free algebra there
exists an extension f : ~ v ( X ) -+ 92; the correspondence between f and f is
bijective. It is a matter of straightforward but tedious calculations to verify
that this bijection is natural. We conclude the following theorem.
THEOREM 4.5.4. Let V be a variety of ~-algebras. Then the forgetful

funetor has a right adjoint, the free functor.
The next example appears also in many guises throughout this book. A
poser (P, ~) can be regarded as a category as follows. P u t Ob(C) := P,
Mor(C) := <_, id(x) := (x,x), and (x,y) o (y,z) := (x,z). The domain of
(x,y) is x, and the codomain is y. We call a category a p o s e t c a t e g o r y if
for every pair A, B of objects there exists at most one arrow from A to B.
In a poset category it is unnecessary to name arrows. So we simply write
A -+ B to denote the unique arrow from A to B; and we write A < B to
state that there exists an arrow from A to B. Let now e and © be poset
categories. A functor F : C -~ :D is uniquely determined by its action on
the objects. For if f : d --+ B then F ( f ) : F(A) --+ F(B) is an arrow and
so is uniquely determined. The requirement that F is a functor is equivalent
to the condition that F be i s o t o n i c , that is, if A _< B then F(A) < F(B).
Suppose now that ~ and ~ are posets and f : P -+ Q, g : Q --+ P be isotonic
4.5. A d j o i n t F u n c t o r s a n d N a t u r a l T r a n s f o r m a t i o n s 201
maps. We may think of the maps as functors between the corresponding poset
categories. We claim that f is left-adjoint to g iff the following holds for all
x•Pandally•Q:
x < g(y)
f(=) _< y
(This is read as follows. The situation above the line obtains iff the situation
below the line obtains.) The proof is straightforward. There exists a bijection
between the horn-sets, and this bijection is uniquely defined by the fact that
the horn-sets contain only one member. The naturalness is also immediately
verified. Now let x • P , y • Q. T h e n f r o m g ( y ) < g(y) we get f g ( y ) < y,
a n d from f ( x ) < f ( x ) we get x < g f ( x ) . A s s u m e now t h a t f : q3 --+ ~ a n d
g : t.~ --+ ~ such t h a t fg(y) < y for all y • Q a n d x < g f ( x ) for all x • P .
T h e n from y <_ f ( x ) we d e d u c e g(y) < g f ( x ) . Since g f ( x ) < x we have
g(y) < x. If g(y) <_ x t h e n fg(y) < f ( x ) . T o g e t h e r w i t h y < f g ( y ) we get
y <_ f ( x ) .
W e will now e x t e n d the results of d u a l i t y t h e o r y to m o d a l algebras. T h i s
can b e done by p u s h i n g t h e t o p o l o g i c a l d u a l i t y further, as o u t l i n e d in GIO-
VANNI SAMBIN a n d VIRGINIA VACCARO [186]. We will sketch this a p p r o a c h ,
p r o v i n g only p a r t of t h e results. Some t e c h n i c a l d e t a i l s have to b e a d a p t e d
when lifting this a p p r o a c h to p o l y m o d a l algebras. T h i s is t h e r e a s o n why
we do n o t s i m p l i f y t h e e x p o s i t i o n to m o n o m o d a l algebras; o t h e r w i s e it looks
overly c o m p l i c a t e d . T h e key is to r e g a r d p o ] y m o d a l a l g e b r a s as functors
from a special d i a g r a m into t h e c a t e g o r y of b o o l e a n a l g e b r a s w i t h h e m i m o r -
p h i s m s as functions. R e c a l l t h a t a h e m i m o r p h i s m from a b o o l e a n a l g e b r a
9.1 to a b o o l e a n a l g e b r a ~3 is a m a p r : A --+ B such t h a t T(1) = 1 a n d
r(a • b) = r(a) fl r(b) for all a, b • A. We w r i t e r : 92 ~ ~ for t h e fact t h a t
r is a h e m i m o r p h i s m . A e o - h e m i m o r p h i s m is a m a p c~ : A -~ B such t h a t
a ( 0 ) = 0 a n d c~(a U b) = or(a) U or(b) for all a, b • A. If r is a h e m i m o r p h i s m ,
c~(a) : = - r - (a) is a c o - h e m i m o r p h i s m , a n d if cr is a c o - h e m i m o r p h i s m t h e n
r ( a ) : = - r - (a) is a h e m i m o r p h i s m . T h e c a t e g o r y of b o o l e a n a l g e b r a s as ob-
j e c t s a n d h e m i m o r p h i s m s as arrows is d e n o t e d by Bal. T h e d u a l of a b o o l e a n
a l g e b r a is a S t o n e - s p a c e ; t h e d u a l of a h e m i m o r p h i s m t u r n s o u t to be a rela-
t i o n b e t w e e n t h e p o i n t s of t h e spaces. Namely, if r : 92 ~ ~3 a n d f : 92 --+ 2,
g : ~3 --+ 2 a r e p o i n t s , t h e n p u t g <1 f if for all a • A, g(r(a)) = 1 i m p l i e s
f ( a ) = 1.
C o n s i d e r a r e l a t i o n < C X × Y. G i v e n a s e t S C X a n d a s e t T_C Y,
write
<~T := {z • z : (~y)(x < y and y e T)}
[]T := {z • X : (Vy)(if x < y then y C T)}
<>S := {y • Y : (~x)(x <~y and x • S)}
[]S :-- {y • Y : (Vx)(if x <l y then x • S)}
Then <>, [ ] : p ( X ) --+ ~o(Y) and @, [ ] : ~o(Y) -+ p ( X ) . Moreover, the following

laws of adjunction hold.
S c []T ~S D T
<>S c_ T S D_ e T
These laws are reflected in the postulates p -+ []<>p and p -+ []<gp of tense
logic. (In fact, the latter encode that []<> and ~n~ are the unit and counit
of this adjunction.) Let 36 and ~ be topological spaces and f : 9[ -+ Y be a
function, f is called o p e n if for every open set S C X , f[S] is open in f .
Likewise, f is called c l o s e d ( c l o p e n ) if the direct image of a closed (clopen)
set is closed (clopen). In general a m a p t h a t is both open and closed is also
clopen. The converse is generally false.
DEFINITION 4.5.5. Let 26 and ~3 be topological spaces, and ~ C_ X x Y be
a relation. <~ is called a c o n t i n u o u s r e l a t i o n from 26 to ~ if <~ is a clopen
map from ~ to ~. The category of topological spaces and continuous relations
as maps is denoted by Spa. <~ is called closed if <> is a closed map.
A relation is continuous iff [3 is clopen. The reader is warned t h a t [ ] H
is not the inverse image of H under <~. The latter is <~H. The two coincide
just in case <~ is a surjective function. Let 260 denote the boolean algebra of
clopen subsets of X.
260 := ( { H : H clopen in 26}, 1, - , V/>
Let <1 C X x Y be a continuous relation. Then <) commutes with a r b i t r a r y
unions. So ~ : ~ o __+ 260 is a co-hemimorphism and [] : gjo _+ 32o is a
hemimorphism. Now let • C_ Y x Z be a continuous relation from ~ to 3.
Then <l o • C_ X × Z is a continuous relation from X to Z.
THEOREM 4.5.6. (_)o is a contravariant functor from Spa to Pal.
Now return to the case of a hemimorphism ~- : 91 ~ ~ . We define 91° to
be simply the space of points endowed with the topology generated by the
sets 3 = {U E pt(Pl) :a E U}. Its sets can be characterized as follows.
PROPOSITION 4.5.7. Let 26 be the Stone space of Pl. A set S is closed in
32 iff it is of the form {U : U D F } , where F is a filter in P2. A set is clopen
iff it is of the form {U : U D F}, where F is a principal filter.
PROOF. We prove the second claim first. Assume S is clopen. T h e n
S=~= { U : a C U} for s o m e a C A. Now put F = { b : b > _ a } . T h e n S =
{U : U _D F}. F is principal. Conversely, if F is principal, say F = {b : b >_ a}
for some a, then S = 3, hence S is clopen. Now assume t h a t S is closed. T h e n
S = A~e~ R~ for some family of clopen sets R~. Let R~ = {U : U _D F~}, i C I ,
where each Fi is a principal filter. Let G be the filter generated by the Fi.
We have S = {U : U _D G}. Conversely, assume that there exists a filter G
4.5. Adjoint Functors and Natural Transformations 203
such that S = {U : U _D G}. Then a is generated by a family (F~ : i C I) of

principal filters. It follows that S = ~ i c z R~, where R~ = {U : U _D Fi}. Each
Ri is clopen, and so S is closed. []
Now for U C pt(~3) and V E pt(92) we put U % V ifffor every a E U, ~-a C U
implies a E V. The latter is the same as: a E V implies or(a) E U, where
a(a) :-- -7- - a E U. With <1 understood to be To, <?, ~ etc. are properly
defined. Therefore, by definition
U~oV ** vc_~[u]
U~-oV ** u c_ cr[v]
DEFINITION 4.5.8. Let ~ and fO be topological spaces. A relation <1 C
X x Y is called p o i n t closed, if for every x E X , ~{x} is closed in ~ .
PROPOSITION 4.5.9. Suppose ~- : 92 ~ ~3. Then % is a continuous, point
closed relation from ~3o to 920.
PROOF. It is not hard to see that ro is point closed. For let U C pt(%).
Then v - l [ U ] is a filter of 92; hence H := {T • pt(92) : T D_ T-I[U]} is closed.
Moreover, H = {T : T % U } . Now for the first claim. Let C be a clopen set,
C = ~. Then S • ~ C i f f S T o T for some T • 3 i f f S r o T for some T 9 a
iff - T - a C S. The latter is a clopen set. Hence, ~ maps clopen sets onto
clopen sets. []
THEOREM 4.5.10. (--)0 is a eontravariant functor from Bal into Spa.
PROOF. Let T : 92 ~ £3 and v : ~3 ~ ~. We have to show that
(vor)o =roOVo.
C l a i m . Let 2~ and £3 be Stone spaces. If <~ is a continuous and point closed
relation from 2~ to fO then ~ is closed. For a proof let D be a closed set of ~2.
We show that if y ¢ ~ D then there exists a clopen C C Y such that <>D _C C
but y ¢ C. This suffices for a proof. So let y ¢ q D . For every x • D,
y ~ ~{x}, and since ~{x} is closed there exists a clopen set Cx such that
y ¢ Cx but ~{x} _C C~. (This follows from the fact that each closed set o f ~ is
the intersection of clopen sets.) Hence, x • [3C~. Therefore D C_ U~CD []C~.
This is an open cover of D, and by compactness of ~ there exists a finite set
S such that D C U~s []C~. Then also ~ D C_ Ux~sC~. P u t C := U~esC~.
C is clopen and does not contain y. This proves the claim.
From this fact we deduce that ~-o o Vo is point-closed. For, being the
composition of closed maps, it is closed, and a f o r t i o r i point-closed. We
finally need to verify that this map is based on the same continuous relation
as (voT)o. The proof is essentially the same as in Theorem 3.2.8; if U ( r o v ) o V
then one has to find a point Z such that U % Z and Z vo V. Alternatively, two
maps between Stone spaces are identical iff they are identical on the clopen
sets. We leave the verification of that fact to the reader. []
Finally, consider having n - m a n y modal operators. Let ~(n) be the cate-

gory consisting of a single object, denoted here by Q. The set of morphisms
consists in finite sequences from n. If a and ~- are such sequences, then
cr(~-(Q)) = (cr~T)(Q). The identity on Q is the m a p e, where e is the e m p t y
sequence. (Moreover, dom(f) = cod(f) = Q for every arrow f . ) This defines
the category J(n). Consider a functor F : ~(n) --+ Bal. Then F ( Q ) = 91 for
some boolean algebra 91 and for each j < n, F ( j ) : 92 ~ 91. For each sequence
or, F(cr) : 9.1 -~ 91 is uniquely determined by the F(j), j < n. Hence, the
functor can be viewed as a n - m o d a l algebra. Next, consider another functor
G : J(n) --+ Bal. Suppose t h a t r1 is a natural transformation from F to G.
This means that r/(Q) : 91 ~ ~3 and that
r/(e ) o F ( j ) = G(j) o rl(e )
A natural transformation r] is b o o l e a n if ~(Q) is a boolean homomorphism.
In t h a t case, the conditions on r1 being a natural transformation are equivalent
to ~?(Q) being a homomorphism.
THEOREM 4.5.11 (Sambin & Vaccaro). The category of~-modal algebras
is equivalent to the category of functors from J(~) to Bal, with arrows being
the boolean natural transformations. This category is denoted by Main.
In a similar way we can introduce functors from J(~) to Spa. T h e y corre-
spond to n - m o d a l frames. In other words, a n - m o d a l frame is a topological
space ~ endowed with a family of continuous relations from 3: to ~ indexed
by ~. For the moment, however, frames are functors. Let F and G be two
such functors. A w e a k c o n t r a c t i o n from F to G is a continuous relation c
from ~ := F ( Q ) to ~ := G(Q) such that (1.) For every clopen set H of ~ ,
c - i [ H ] is clopen in ~, (2.) For every clopen set H of ~ , <)c-l[H] = c - l [ O H ] .
A c o n t r a c t i o n is a function (!) and a weak contraction t h a t satisfies (2.)
for all sets {x} (and so for all subsets of Y). Hence a contraction is a weak
contraction. It is tempting to conclude that the category of frames is the same
as the category of functors from J(~) into the category of topological spaces
together with weak contractions. This is not so, however. Rather, this result
can be true only if we take zero-dimensional topological spaces.
THEOREM 4.5.12 (Sambin & Vaccaro). The functors from ~(~) to the ca-
tegory of zero-dimensional compact spaces with contractions as arrows form a
category denoted by Fray. Fra~ is equivalent to the category of n-modal frames
and p-morphisms.
The proof of this theorem is left as an exercise. The reader should con-
vince himself that this theorem is essentially a reformulation of the notion of
a general frame and a p - m o r p h i s m between such frames into category the-
ory. Now consider the category of Stone spaces, denoted by StSpa~, and the
category of functors from J(~) with point closed contractions as arrows.
4.6. Generalized Frames and Modal Duality Theory 205
THEOREM 4.5.13. StSpa~ is dual to the category Mal~.
E x e r c i s e 155. Show that if f : X -+ Y is a function, the map f : ~(X) --+

p ( Y ) : A ~ f[A] has both a left adjoint and a right adjoint (if viewed as a
map between posets).
4.6. Generalized Frames and Modal Duality T h e o r y

We have seen that a boolean algebra can be realized as a set algebra i. e. as
a suba]gebra of a powerset-algebra. Also, we have seen that boolean algebras
can be repesented by certain topological spaces. So, we can either choose a
topological representation or a representation of boolean algebras by so-called
boolean spaces (which are pairs (X, X) where X is closed under complement
and union). In a subsequent section we have developed the topological repre-
sentation of modal algebras. In a second step we restrict the topological space
to the clopen sets. In this way we get the standard representation of modal
algebras as certain general frames.
DEFINITION 4.6.1. Let 92 be a a-modal algebra. Then the generalized

frame 92+ is defined as follows. The w o r l d s are the ultrafilters of 92, and
U <~j V iff for all b E V we have $jb C U. Furthermore, the i n t e r n a l sets
are the sets of the form b = { U : b E U}. For a homomorphism h : 92 -+
we let h + : ~ + -+ 92+ be defined by h+(V) := h-l[U] = {b: h(b) e U}.
The converse direction is harmless as well.
DEFINITION 4.6.2. Let ~ = (~,F} be a generalized frame. Then the modal

algebra 5+ is defined as follows. The e l e m e n t s are the i n t e r n a l sets, and
the o p e r a t i o n s are intersection, complement, and $j, j < ~, defined by
ejb := { w : (3x)(w <lj x and x C b)}.
If p : ~ --+ ® is a p-morphism, then p+ : ®+ --+ 5+ is defined by p+(b) :=
THEOREM 4.6.3. (--)+ is a contravariant functor from the category Mal of

modal algebras to the category Frm of general frames; ( - ) + is a contravariant
functor from Frm to Mal. Moreover, for every modal algebra 92++ ~ 9J.
We entrust the proof of the fact that ( - ) + and ( - ) + are functors onto
the reader. Again we are faced with the problem of the converse, namely, to
say when for a general frame 5+ + ~ 5. In order to state this condition, let
us give another definition.
DEFINITION 4.6.4. A frame is dii~ferentiated if for every pair x and y

of different worlds there is an internal set containing x but not y. A frame is
tight if for every pair x and y of worlds such that x ~j y there is an internal
set b such that x E I j b bury ¢ b. A frame is c o m p a c t if every family of
internal sets with the finite intersection property has a nonempty intersection.
A frame is refined if it is both differentiated and tight, and it is d e s c r i p t i v e
if it is refined and compact. A frame is a t o m i c if for every world x the set
{x} is internal and full, if every subset is internal.
We introduce also abbreviations for classes. ~:p denotes the class of
Kripke-frames, (5 the class of generalized frames, ~ f the class of differenti-
ated frames, ~i the class of tight frames, 9~ the class of refined frames, Crop
the class of compact frames, ~ the class of descriptive frames and E the class
of canonical frames. Atomic frames will play only a marginal role, although
atomicity is a rather desirable property of frames, playing a fundamental role
in the completeness proofs for fusions (see Chapter 6). The definition of com-
pactness is equivalent to the definition of compactness of the space generated
by the sets of the algebra. Moreover, the postulate of differentiatedness is
the separation postulate for Stone-spaces. We may say informally, that a
compact frame has enough worlds, a differentiated frame has enough internal
sets for identity and a refined frame has enough sets for the basic relations.
The properties of frames have a topological counterpart.
PROPOSITION 4.6.5. Let ~ be a a-modal frame. (i.) ~ is differentiated iff
{x} is closed for all x E f iff the corresponding topological space is Hausdorff.
(ii.) ~ is refined iff the space is Hausdorff and <~j is point closed for all j < a.
(iii.) ~ is compact iff the corresponding topological space is compact.
We remark here that the property of tightness also is a topological sep-
aration property. It says, namely, that for every x and j < a the set
sucj(x) = {y : x ~j y} can be separated by a clopen neighbourhood from
any point not contained in it.
THEOREM 4.6.6. For a descriptive frame, ~ ~- 5+ +.
PROOF. Consider the map x ~-~ Us := {b : x C b}. This map is injective,
since the frame is differentiated. It is surjective, since the frame is compact.
Now assume that x <lj y. Then if b C Uy, Ojb C Us. Hence Us <~j Uy, by
definition of the relation ~j. Now assume x :~j y. Then, since the frame
is tight, there is a b E Uy such that • - b E Ux, that is, 0b ~ Ux. Hence
u~ ~ y~. []
COROLLARY 4.6.7 (Duality Theorem). The categories Mal~ of (a-)modal
algebras and DFrm~ of descriptive n-frames are dually equivalent. Moreover, a
homomorphism of modal algebras is surjective iff its dual map is an embedding;
it is injeetive iff its dual map is a contraction.
4.6. Generalized Frames and Modal Duality T h e o r y 207
T h e duality t h e o r e m distinguishes the descriptive frames from the other frames

and will be used quite effectively. From a theoretical point this is a very sat-
isfactory result, allowing us to transfer algebraic proofs into geometric proofs
and conversely. However, as it turns out, descriptive frames are hard to con-
struct. We do not have such a good intuition a b o u t t h e m . Typically, it is
much easier to construct refined frames t h a n to construct descriptive frames.
So, we will quite often work with refined frames instead. As a last point notice
t h a t there exists a m o d a l algebra based on a single element. This algebra is
denoted by 1. A p p l y i n g the representation theory we obtain t h a t 1 + is the
e m p t y frame. It is this fact to allow frames to have no worlds at all.
In addition to (general) frames we also have K r i p k e - f r a m e s and there
is a rather easy way to turn a frame into a K r i p k e - f r a m e , n a m e l y by just
forgetting the internal sets. So, given a frame ~ = ([, F) we put ~ := f, and
given a p - m o r p h i s m 7r : ~ --+ ® we put ~r~ := 7r. T h e n 7r~ : ~ ~ ¢~. This is a
functor, as is easily checked. This is a kind of forgetful functor. Conversely,
given a Kripke frame [, let ~ := ([, 2f). In analogy, we d u b this the r e c o v e r y
f u n c t o r . Actually, the same terminology can be applied to the previous case.
To pass from a general frame to a m o d a l algebra is practically a forgetful
operation, and to get a frame from the algebra is a recovery process. T h e r e
is a complete analogy here, because forgetting w h a t has been reconstructed
results in the s a m e structure; we have b o t h 9.1++ -~ 9.1 and ~ ~ f. B u t the
converse need not hold; we are not sure to be able to reconstruct w h a t we
have forgotten.
THEOREM 4.6.8. The map (-)~ is a covariant functor from the category
Frm of frames into the category Krp of Kripke frames. The map (-)~ is a
covariant functor from Krp into Frm. Moreover, for a Kripke-frame ~ ~ ~ [.
Evidently, ~ is full iff ~ = ~ . So, the category of full frames is equivalent

to the category of K r i p k e - f r a m e s . (This is not as difficult as it sounds.) A nice
consequence is the following construction, originally due to BJARNI JONSSON
and ALFRED TARSKI. Take a n - m o d a l algebra 92. It is called c o m p l e t e if
the lattice reduct is complete. T h e c o m p l e t i o n of 9~, ffrn(92), is a m o d a l
algebra together with an e m b e d d i n g c : 9.1 ~ ~m(92), which is complete and
satisfies t h a t for every complete ~3 and h o m o m o r p h i s m h : 9.12--+ ~3 there exists
a k : ~rn(9.1) -+ ~3 with k o c = h. It is easy to see t h a t a complete m o d a l
algebra is isomorphic to an algebra of the form 9Jta(f) for some K r i p k e - f r a m e
f (where 9)~a(f) as defined earlier can now be redefined as f~+). Simply take
as elements of f the a t o m s of 9.1, and put b <~j c for a t o m s b and c, iff b _< 0jc.
Now if 9.1 is a m o d a l algebra, then let ~ra(92) := (92+~)~+. This m e a n s the
following. We pass from 9.1 to the dual general frame. Next we pass to the
corresponding full frame by taking all sets of the frame as internal sets (this is
the actual completion). Finally, we take the algebra of sets of this full frame.
208 4. U n i v e r s a l A l g e b r a a n d D u a l i t y Theory
T h e identity on pt(9~) is a p - m o r p h i s m i : ( ~ + ) ~ ~-~ ~ + . It is surjective,

and so i + : ~ ~-- ~m(91). We show t h a t c : = i + has the required property.
To t h a t end, let h : 92 --+ % be a h o m o m o r p h i s m and % complete. T h e n
h + : if3+ --+ 9.1+ is a p - m o r p h i s m of frames. Since if3+ is complete, we can
actually factor h + t h r o u g h i and a function j : if3+ -+ (91+)~ ~, by extending
h + to all subsets of the frame. We have h + = i o j. Now switch back to the
m o d a l algebras. By duality, this gives h = j + o i + = j + o c. P u t k : = j + .
This concludes the proof.
THEOREM 4.6.9. For every modal algebra, the natural embedding map
9.1 ~ ~ m ( ~ ) : = (gA+~)~+ is a completion.
Now, w h a t are the relationships between all these classes? First of all, a
full frame is b o t h differentiated and refined. A full frame is c o m p a c t iff it is
finite. For consider the family of sets f - {x} for x C f . If f is infinite, this
family has the finite intersection property. Yet the intersection of all these sets
is empty. A differentiated c o m p a c t frame is also tight and hence descriptive.
(See exercises; a proof using quite different m e t h o d s will be given in the next
chapter.) T h e r e are tight, c o m p a c t frames which are not differentiated. For
example, take a descriptive frame 5. F o r m a new frame ~a from ~ by t a k i n g
twins w 1 and w 2 for each w C f . P u t x i <lj yk iff i = k and x <~j y; finally, the
i n t e r n a l s e t s are t h e s e t s o f t h e f o r m a 1 U a 2 = { x 1 : x C a } U { x 2 : ~ C a } .
This is a frame; it is tight, and compact. But it is not differentiated.
Now for the difference between tightness and differentiatedness. We have
seen already t h a t there are tight frames which are not differentiated. For the
converse we have to work harder. A finite frame will not do here. Define
a general frame 9~ : = (r,]~), where ~ : = ( w , ~ ) and ~ is the set of all finite
unions of sets of the form
r(i,j)={k:k-j (modi)}
where 0 < j < i. Since - r ( i , j ) = U j , # j r ( i , j ' ) , ~ is closed u n d e r comple-
ments. ]R closed under intersection, too, by the Chinese R e m a i n d e r T h e o r e m .
Finally, Ca = w iff a ~ O, and ¢ O = O, so 9~ is a frame. It is differentiated.
For let i # j, say j < i. T h e n i E r(i + 1 , - 1 ) but j ~ r ( i + 1 , - 1 ) . Now 9~ is
not tight. For i ¢ j iff j < i. But there is no set b such t h a t i C I b b u t j ~ b.
For either b ~ I and then i ~ rob(= ~ ) , or b = w and t h e n j E b.
This is actually an instructive frame and we will prove s o m e t h i n g more
t h a n necessary right now.
THEOREM 4.6.10. Th(9~) = $5. Moreover, every finite Kripke-/rame for
$ 5 is a p-morphic image of iR.
PROOF. We prove the second claim first. Consider an S 5 - f r a m e with i
elements. T h e n the m a p p : j ~+ j (rood i) is a p - m o r p h i s m . We take as <~
the direct image of ~_ u n d e r p . Now i f k < i then for some r C ~ we have
4.6. Generalized Frames and Modal Duality Theory 209
j < r.i+k, so that p ( j ) <lp(k). Thus q = i × i, and so we have a p -

morphism of Kripke-frames. Next, take any subset J C { 0 , . . . , i - 1}. Then
p - l [ j ] = U j e j r ( i , j ) , which is internal. Finally, by the fact that S5 has the
finite model property we have Th(9~) C S5. However, ~ ~ p --+ O0p. For if
/3(p) = ;~ then t3(p) C ~ ( O 0 p ) . And if 13(p) ~ O then ~(E]0p) = I 1 = 1. []
It seems at first sight that a flame which is not differentiated can be made
into a differentiated frame by taking the map defined by x ~-~ Us. We call
this the refinement map. It maps two points x and x ~ onto the same target
point if Us -- Ux,. If U~ -- Us, we also write x ~ x ~. Unfortunately, this
generally is not a p-morphism. We need an extra condition to ensure this.
One possibility is to require tightness. For then, ifp(x)<lp(y), that is, Us <1Uy
then for every x ~ ~ x there is a y~ such that x ~ <~y~ and y~ ~ y. In the case of
tightness we can even show something stronger, namely that x ~ <~ y. Namely,
let y C b, that is, b 6 Uy. Then 0b C Us = Us,, that is, x ~ C 0b. Hence, by
tightness, x ~ <~ y. However, tightness is stronger than necessary.
PROPOSITION 4.6.11. If 3 is tight, the refinement map p : x ~-+ Us is a

p - m o r p h i s m . Moreover, if x <S y, x ~ x ~ and y ~ yl then x ~ <1 y~.
We will now turn to some important questions concerning the relationship

between geometrical properties of a frame 3 and properties of the algebra of
sets. In particular, we will be concerned with the question of whether 3 is
rooted corresponds to 3+ is subdirectly irreducible. The material presented
here is based on [185]. We will give two examples, showing that in fact
neither implication holds.
EXAMPLE. Consider the frame 3 = (Z, -<, F) where x -< y iff lx - y} = 1
and F is the set of finite and cofinite subsets of Z. This is well-defined. The
frame is connected and the only open filters are the trivial filters. (Let F be
an open filter. If F ~ F, then F does not contain a finite set. For if a is
finite, there is an n E w such that I n a = ~. Now, assume that F ~ {1}.
Then for some z, Y := Z - {z} E F. Then if Iz' - z I <_ n, z' ¢ I n y C F,
whence Z - {z ~} E F for all z ~. Any cofinite set is an intersection of such
sets. Hence, F is the filter containing all cofinite sets.) Thus, C o n ( 3 + ) ~ 3.
This shows that the algebra of 3 is subdirectly irreducible. Now consider the
bidual 3+ +. A point of 3+ + is an ultrafilter of 3+. If U is not principal then it
contains only infinite sets, hence only cofinite sets. Moreover, it must contain
all cofinite sets. So, there exists exactly one nonprincipal ultrafilter, which we
denote by V. Also, Us := {a E IF: x C a}. Then for all x, V ~ Us, since 0{x}
is finite and so not in V. Likewise Us ~ V. For let b := {y: ] x - y I > 1}. Then
b C V but x ¢ 0b. It is not hard to verify that V < ~ V . So, t h e b i d u a l o f 3
has as its underlying frame the disjoint union of (X, -<) and [ ~ . It is therefore
not rooted, while its algebra is subdirectly irreducible. This example is quite
similar to the algebra of finite and cofinite subsets of the infinite garland in
Section 7.9.
EXAMPLE. Consider the frame f~ := (co, > , © ) , where © is the set of fi-
nite and cofinite sets. Here, f~+ is not subdirectly irreducible, but f~++ is
rooted. To verify the first claim, notice that f~+ has an infinite descending
chain of congruences whose intersection is the diagonal. These congruences
correspond to the finite generated subframes of ft. Hence, f~+ is not subdi-
rectly irreducible. On the other hand, f~++ has one more element t h a n f~,
consisting of the ultrafilter V of cofinite sets. We show that for all ultrafilters
U, V<1U. For let b E U. T h e n n C b for some natural number n, and so
0b _D {y : y > n}, which is cofinite. Thus 0b C V, and therefore V < U, by
definition of <.
THEOREM 4.6.12 (Sambin). There exist descriptive frames 3 which are

rooted such that 3+ is not subdirectly irreducible. There exist algebras 92 which
are subdirectly irreducible such that 92+ is not rooted.
Let us stay a little bit with the theory of descriptive frames. JOHAN VAN
BENTHEM [12] has investigated the structure of some ultrafilter extensions of
frames. We will show that though the structure of ultrafilter extensions gives
some evidence for the structure of biduals, it is by no means complete. Before
we give examples we will develop some terminology. We will simplify the
discussion by restricting ourselves to a single operator, [~. Notice, however,
t h a t when we have a Kripke-frame {f, <~}, then automatically we have two
operators on the algebra of sets, namely • and • . We will use this notation
throughout this section. Notice on the other hand that a (general) frame for
the monomodal language need not be a frame for the bimodal language, since
the algebra of internal sets need not be closed under • . Nevertheless, we will
use • , keeping in mind that it may result in noninternal sets. Moreover, we
will reserve • for the operator on 92, and use • and • for the operations on
92+. Finally, in a frame 3, a set is denoted by a lower case letter only if it is
an internal set. Upper case letters stand for sets which may also be external.
A reminder on the use of topology: internal sets are also clopen sets. We will
switch freely between these two characterizations.
DEFINITION 4.6.13. An element of a modal algebra 92 is called e s s e n t i a l

if the open filter generated by it is a minimal open filter of 92 distinct from
{1}. Ea denotes the set of essential elements of 92.
The reader m a y verify that an element is essential in 92 iff it is an opre-

mum.
LEMMA 4.6.14. Suppose that E~ is not empty. Then E~ U {1} is the

smallest open filter. So, E~ is nonernpty iff 92 is subdirectly irreducible.
Now we start to investigate the nature of 92+. Let us call a set X in a

frame ~ a t r a n s i t if it is successor closed. Furthermore, X is closed if it is an
intersection of internal sets (iff it is a closed set of the t o p o l o g y induced by F
on f ) . It is not hard to see t h a t the closed transits are closed under a r b i t r a r y
intersection and union, and that they form a locale,
THEOREM 4.6.15 (Sambin L= Vaccaro). The locale of open filters of 92 is
anti-isomorphic to the dual locale of closed transits of 92+.
PROOF. Consider the m a p C defined by
F ~-+ C(F) : = {Y • pt(92) : F C_ U} = a • F}
This m a p is a bijection between closed sets of 92+ and filters of the boolean
reduct of 92. We need to show t h a t F is open iff C(F) is a transit.
(Va)(a C F ==~ ma 6 F )
¢=~ (Va)(C(F) C "d ~ C(F) C_ . a ) since a C F iff C(F) C_3
¢:~ (Va)(C(F) C "~ ~ C(F) C_ • ' d ) since l a = • ~
(Va)(C(F) C "~ ~ • ' d C "d) by adjunction
¢, • C(F) c C(F) since C(F) and • C(F) are closed
¢* C(F) is a transit
[]
COROLLARY 4.6.16. 92 is subdirectly irreducible iff92 + has a greatest closed

transit.
For frames, a similar terminology can be set up. Notice first of all the follow-
ing.
PROPOSITION 4.6.17. Let ~ be a frame and C C_ f a set. Then the small-

est transit containing C is the set T(C) : = [Jke~ • kC" It is open if C is
internal. The largest transit contained in C is the set K ( C ) :-- [ ~ k ~ • kC"
It is closed if C is internal.
DEFINITION 4.6.18. Let ~ be a frame. Let I~ denote the set of all points
x such that the transit of x is 5. Furthermore, put H~ : = f - I~.
LEMMA 4.6.19. For every frame, H~ is a transit. Moreover, if ~ has a
greatest nontrivial transit C then C = H~. Otherwise, f = H~. So ~ is rooted
iff it has a greatest nontrivial transit.
LEMMA 4.6.20. For every frame ~ and every set C, if H~ C C and C ~ f
then H~ = K(C).
Recall t h a t in a topological space :~, a set S is called d e n s e if the closure
in ~ is the entire space. A set is of m e a s u r e z e r o if it does not contain any
open set iff its open kernel is e m p t y iff its complement is dense.
LEMMA 4.6.21. For every frame, H~ is either dense or closed.

PROOF. Suppose H;~ is not dense. Then there exists a clopen set C such
t h a t H~ C_ C C f. (For every closed set is the intersection of clopen subsets,
and the closure of H~ is not f.) Therefore, by L e m m a 4.6.20, H~ = K ( C ) .
By Proposition 4.6.17, H~ is closed. []
PROPOSITION 4.6.22. For every algebra 9A, if I~+ is not of measure zero,
then 92 is subdirectly irreducible.
PROOF. Let I~+ be not of measure zero. Then H~+ is not dense. There-
fore it is closed, by L e m m a 4.6.21. By Corollary 4.6.16, 92 is subdirectly
irreducible. []
It might be deemed unnecessary to invoke the topology when we want to

speak a b o u t Kripke-frames. For example, we know that for a K r i p k e - f r a m e
~, 9)ta(f) is subdirectly irreducible iff f is rooted. But this is almost the only
result t h a t does not make use of the topological methods. So, let us deal first
with the question of open filters in the algebra 9~la([) of a Kripke-frame. At
first blush they seem to correspond simply to the transits of 5. As a particular
corollary, i f ~ is connected, 9Jla(f) should be simple. But this is wrong. For let
be countably infinite. Then 991a(I) has at least 2 ~° elements. It is easy to
show, however, that such an algebra cannot be simple. For take any element
b ~ 1. The least open filter containing b is countable. Hence it is not {1}
and not the entire algebra. Thus, the conjecture that ffJta([) is simple if every
world of f is a root is easily refuted. Nevertheless, it is instructive to see a
concrete counterexample.
EXAMPLE. Let 3 = (Z, <~} be the frame with Z the set of integers and
x <l y iff Ix - Yl = 1. Every world of Z is a root. So, H;~ = 2~, which is a clopen
set. So, the algebra of sets is subdirectly irreducible (which also follows from
the fact that Z is rooted). Therefore, E 3 is not empty. T h a t is, we should
have essential elements.
LEMMA 4.6.23. E~ consists exactly of the cofinite sets 7£ Z.
PROOF. Suppose that A is cofinite. Then there exists a k C w such
t h a t Z - A C_ [ - k , k ] . Let B C w. T h e n for some m E Z, m ~ B. So,
[m - n, m + n] C_ . n B . Hence, I - k , k] _D mm+kB, from which follows t h a t
A D_ Rm+kB. So, A is essential. Now let A not be cofinite and let B := w - { 0 } .
Then for no k E w, mkB C_ A, since mkB is cofinite. []
We remark here t h a t the converse of Proposition 4.6.22 is false.

EXAMPLE. Let 3 := (6, ©) where © consists of all finite unions of sets of
the form o(k, a) := { n . 2 k ÷ a : n E Z}, where k is a natural number and a
an integer. The algebra PA := 3+ is simple. For any set of © is of the form
b = [Ji<p o(k, ai) for some p and k and some ai. Let {bj : j < q} be the set of
n u m b e r s such t h a t for every j < q there exist i, i ~ < p such t h a t bj - ai 4- 1

( m o d 2 k) and b -~ ai, - 1 ( m o d 2 k ) . T h e n lib = Uj<qO(k, bj). It follows
t h a t . 2 k b ---- ~ if[ b ~ 1. By the criterion for subdirect irreducibility, 92 is
subdirectly irreducible with 0 an opremum. Hence 92 is simple. We will show
t h a t the frame underlying 92+ is decomposable into a disjoint union of at least
two frames, from which follows t h a t I~+ = ~. To see that, take an ultrafilter
U in 92. We define a sequence J ( U ) = (jk : k G co) as follows. For each k G co
let jk be the (unique) remainder (mod 2 k) such t h a t o ( k , j k ) C U. J ( U ) is a
sequence satisfying
(*) jk+l = jk or jk+l = jk + 2 lk
Conversely, let S = (sk : k C co) be a sequence satisfying (t). Let U ( S ) be

the ultrafilter containing o(k, sk). It is easy to see t h a t this ultrafilter exists
(this collection has the finite intersection property, by (*)), and t h a t U ( S ) is
indeed uniquely defined. Now, there are evidently 2 a° m a n y such sequences.
Hence, 92+ has u n c o u n t a b l y m a n y worlds.
3 satisfies alt2 and B and so 9.1+ satisfies alt2 and B as well. Hence,
the underlying frame has the p r o p e r t y t h a t (a) each point sees at m o s t two
points, (b) if x sees y, y sees x. (This follows from the results of C h a p t e r 5,
but can also be established directly.) It follows t h a t a connected c o m p o n e n t
is generated by any of its points, and t h a t it is countable. Since the frame is
uncountable, it is has more t h a n two connected components. Consequently,
J9.1+ ~- ~ .
E x e r c i s e 156. Show t h a t a finite differentiated frame is f u l l
E x e r c i s e 157. A n y finite frame is compact. Hence a finite differentiated

frame is descriptive.
E x e r c i s e 159. (This example is due to FRANK WOLTER. It is very closely

related to the frame considered in Section 3.5.) Take the set co + 1 and put
i <~j iff i < j < co or i = j = co. T h e internal sets are all sets which are finite
and do not contain ca or else are cofinite and contain co. Show t h a t this frame
is c o m p a c t , differentiated but not tight.
E x e r c i s e 160. Let V be a variety, n C co and 92i C V, i < n. A subalgebra

~-~ H i < ~ 92i is called s k e w - f r e e if for every congruence O on ~3 there exist
q~i E Con(92i) such t h a t 0 = ( X i < , ~ ) N B 2 (see Section 1.3 for definitions).
Show t h a t every subalgebra of a direct p r o d u c t of m o d a l algebras is skew-free.
Hint. Use duality.
E x e r c i s e 161. Show that there exist simple modal algebras which do not
generate semisimple varieties.
4.7. F r a m e C o n s t r u c t i o n s I I I
Our aim in this section is to translate Birkhoff's Theorem on varieties into
general frames. This will allow us to say which classes of frames are classes
of frames for a logic. In the form that this theorem takes here, however, it is
not very surprising, but we will transform it later into stronger variants that
allow deep insights into the structure of such classes. Recall that Birkhoff's
Theorem says that a class is equationally definable iff it is closed under pro-
ducts, subalgebras and homomorphic images. In the context of modal algebras
equational classes coincide with modally definable classes, where a class :K of
algebras is called m o d a l l y d e f i n a b l e if there is a set (I) of modal formulae
such that K contains exactly the algebras satisfying ~. Likewise, a class K of
frames is m o d a l l y d e f i n a b l e if • is the class of frames for some (I). This can
be relativized to some class, e. g. the class of refined frames, descriptive, full
frames etc. Of particular interest is the class of modally definable Kripke-
frames.
One immediate problem is that we have to translate the notions of prod-
uct, subalgebra etc. into frames. Here, category theory is of good use since
it provides canonical definitions of these things. A good illustration is the
notion of a product.
DEFINITION 4.7.1. Let C be a category, and Bi, i E I, an indexed col-
lection of C-objects. A pair (A,{p~ : i E I } ) , where A is a C-object and
Pi : A -+ Bi C-arrows, is called a p r o d u c t of the Bi if for each object C
and arrows q~ : C --+ B~ there is a unique m o r p h i s m rn : C -+ A such that
qi = m o p i for all i E I. The maps p~ are called p r o j e c t i o n s . (A, {Pi : i E I}}
is called a c o p r o d u c t of the Bi if (A, {(pi)°P : i e I } l is a product in the dual
category, C °p .
, ]32
C _ _ _ . . " _ _
~ ]31
Usually, only the object A in the pair (A, {p~ : i C I}} is referred to as the
product. (This at least is common usage in algebra.) So, a product is an
object for which projections p~ that make (A, {p~ : i C I}) a product in the
sense of the definition. However, notice that the map denoted by '!' in the
4.7. Frame Constructions I I I 215
picture above is not uniquely defined by A and C alone but only given the
pairs ( A , { p l , p 2 } ) and (C, {ql,q2}). We can view a p r o d u c t as a solution
to a special d i a g r a m (consisting of two objects and identity arrows). This
solution consists in an object and arrows from t h a t object into the objects of
the diagram. W h a t makes such a solution a p r o d u c t is a condition t h a t is
usually referred to as the universal property. (See the exercises.) T h e reader is
advised to spell out the definition of a coproduct in detail. Before we proceed
to examples, let us note one i m p o r t a n t fact a b o u t p r o d u c t s and coproducts.
THEOREM 4.7.2. L e t C and D be products o f the B i , i E I . T h e n C is

isomorphic to D . Moreover, i f C is a product and i s o m o r p h i c to D , then D
also is a product o f the B i , i C I .
PROOF. By a s s u m p t i o n there are maps Pi : C ~ Bi and qi : D -+ B{ such

t h a t for any E with m a p s ri : E --+ B i we have m : E --+ C and n : E --+ D
such t h a t r i: Pi o m and ri : qi o n for all i E I. We apply the universal
p r o p e r t y for C to D and get a m a p f : D ~ C such t h a t qi = P i o f for all i. We
apply the universal p r o p e r t y of D to C and obtain likewise a m a p g : C --~ D
such t h a t Pi = qi o g for all i. T h e n we have qi : Pi o f = (qi o g) o f = qi o (g o f )
as well a s p i = q i o g = (Piof) og=pio(fog), again for a l l i . S i n c e C i s
a product, there is only one m a p i : C -+ C satisfying Pi = Pi o i. Since the
identity i d ( C ) on C has this property, we conclude t h a t f o g = i d ( C ) . A n d
analogously we get t h a t g o f = i d ( D ) . So C and D are isomorphic.
For the second claim let C be a p r o d u c t with projections Pi, i C I . Let
h : D ~ C be an isomorphism with inverse k : C - ~ D. T h e n we claim
t h a t D is a p r o d u c t with projections qi : : Pi o h. For let E be given and
ri : E --+ B i , i E I. T h e n there exists a unique m a p f : E --+ C such t h a t
ri=piof for a l l i c I . Letg::kof. Thenri=piof=qiokof:qiog
for all i C I. Furthermore, g is unique. For if g' also satisfies ri = qi o g' for
all i, then ri : Pi o h o g' for all i, from which h o g' = h o g. This implies
k o h o g' = k o h o g, which is the same as g = g'. []
T h e p r o d u c t of two algebras, defined earlier, is a p r o d u c t in the categorial

sense. In general, let f~ be a signature, T an equational t h e o r y in f~, and
let A I g T be the category of f~-algebras for T. This c a t e g o r y has products.
( n i c I f21i,Pi), where Pi is the projection onto the ith c o m p o n e n t , is a p r o d u c t
~f the family {9.1i : i E I}. This fact is used later. We p e r f o r m the a r g u m e n t
with I : = {1,2}. Let fl31, fl32 be given. P u t 92 : = ~31 x fl32, and let the
projections be Pl : (bl,b2) ~ bl and p2 : (bl,b2) ~ b2. Take any ¢ with
h o m o m o r p h i s m s qi : ff --~ fl3i. P u t m : ~ --+ fl31 × ~ 2 : c ~+ (ql(c),q2(c)).
t h e r e is no other choice; for assuming re(c) = @1, z2) we get
q (e) = o : :
q2(c) : (Pl o m)(C) : p2(<xl, X2>) : X2

So m is unique, and we only have to show t h a t it is a homomorphism. We

trust t h a t the reader can fill in the proof. From the previous theorem we
get t h a t products are unique up to isomorphism. Let us cash out on this
immediately. Say that a category h a s p r o d u c t s ( h a s c o p r o d u c t s ) if for
any family of objects the product (coproduct) of that family exists. The
category of algebras in a variety has products.
THEOREM 4.7.3. Let A be a polymodal logic. The category of descriptive
A-frames has coproducts.
PROOF. The proof is by duality of the category of descriptive frames and
the category of modal algebras. Suppose that ~i, i E I, is a family of frames;
put
II~i :: (I](~a+) ÷.
iCI iEI
We claim t h a t this a coproduct. By the Duality Theorem it is enough to show
that ( l J i e I ~i)+ is a product. However, it is isomorphic to l-Iiez(~i)+" The
latter is a product of the (~i)+. []
There also is a notion of a coproduct of frames, called the disjoint union.
Let ~i = (f/, Fi) be frames. The disjoint union, (~)ieI ~i, is defined over the
disjoint union of the sets fi with relations being the disjoint union of the
respective relations. The sets are of the form b = [-Jicz bi where bi C Fi. Since
the fi are disjoint, so are then the bi, and it turns out t h a t we get
iEl iEI
The projections are Pi : a F-+ a V/fi. The so defined sets form a modal algebra.
To see this, consider the m a p 7r : b F-+ [.Jie~ bi. This is readily checked to be a
h o m o m o r p h i s m and bijective. The following theorem has been shown in the
special case f = {1, 2} and Kripke-frames in Theorem 2.4.4.
THEOREM 4.7.4. (~)iCI ~i is a coproduct of the frames ~i in Frm.
PROOF. P u t (5 := ( ~ i c I ~i. Let .~ be a frame and hi : ~i --+ -9. We define
m : ~ i e I ~i -+ ~ by re(x) := hi(x) if x E f i . Since the fi are all disjoint, this
definition by cases is unambigous, m is a p - m o r p h i s m , for if x <lj y in (5 and
x C fi then also y in fi and x <lj y already in fi- Since hi is a p - m o r p h i s m ,
re(x) = hi(x) <~j hi(y) = re(y). Now assume re(x) <lj u. Let x e f~. T h e n also
hi(x) <~j u, and we get an y C fi such that hi(y) = u and x <~j y. But then
re(y) = hi(y) = u, and this proves the second condition. Finally, let b E H be
a set. We have ci := h~l[b] C Fi. P u t c := Ui~i el. c is an internal set of (5
and m[c] = b. []
Now we have two definitions of coproducts, one for descriptive frames and
one for arbitrary frames. Since the category DFrm is a subcategory of Frm in
FIGURE 4.2
the sense that it contains a subclass of the objects of Frm but all Frm-arrows
between them, it can be shown that for descriptive frames ~ , i C I, there
exists a m a p ~ I ~i "-~ H ~ s ~i. As it turns out, the two coproducts are not
always the same. For example, take the full frame ¢ ~ , c~n :-- (n, >). These
frames are descriptive. P u t ~ :-- ~,~c~
IT c[~~ , and ~ :-= ~ n ~ c ~ . There exists
an arrow ~ --~ ~ . Furthermore, the algebras ~+ and ~ + are isomorphic, since
they are (isomorphic to the) the product of ( c ~ ) + . Nevertheless, there is no
isomorphism between these frames. One way to see this is as follows. The
frame ~ contains a world which has no finite depth (or no depth at all),
while each point in ~ has a depth. Another argument is the following. ~ has
countably m a n y worlds, the set of worlds being a countable union of finite
sets. But ~ has uncountably m a n y points, as there are uncountably m a n y
ultrafilters in the algebra of sets. (This is given as an exercise below.) It
appears to be paradoxical that we should end up with several versions of a
coproduct, but this is due to the fact that a coproduct is a notion that makes
sense only within a category; it has no absolute meaning. Since the category
Frm has more objects, we end up with different coproducts. Finally, it is
clear that we shall end up with copoducts of frames rather t h a n products, by
duality, or simply the fact that arrows go the opposite way. The reader may
however also check that the categories of frames and descriptive frames have
no products. Namely, take ~1 to be the one point reflexive frame (n = 1),
52 the one point irreflexive frame. Then there is no product of these two in
either category.
Different problems arise with subalgebras and homomorphic images. It
appears at first sight that subalgebras translate into p - m o r p h i c images and
homomorphic images into generated subframes. But the terminology on the
frames is more subtle here. We know that a s u b a l g e b r a - m a p h : 92 --~
defines a unique h + : ~ + -~ 92+ and conversely. But we write f : ~ --- ¢~ for
frames if f is surjective on the worlds only. Now consider case of Figure 4.2.
We have ~ ~ (5, and we have an isomorphism between the algebras of internal
sets. We expect therefore that there is a surjective p - m o r p h i s m from ~ to ®.
But there is none. The duality theory cannot be used in this case. A similar
counterexample can be constructed for generated subframes. Consider namely
the frame 5). Although the algebras of sets is isomorphic to t h a t of @, none
FIGURE 4.3
is a generated subframe of the other.

Moreover, notice that there is a distinction between embeddings and sub-
frame embeddings. An embedding is just an injective p-morphism, while a
map i : 3 "-~ ~5 is a subframe embedding if i -1 : G -4 IF is surjective. In
the latter case 3 is required to have no more sets than necessary to make i a
p-morphism. If e : 3 -4 ~5 is an embedding, take F2 = {e-l[b] : b C qh}. T h e n
the identity is an embedding: id: (~,F2) ~-~ ([,F) and the image of (f, F2)
under e is a generated subframe.
Now let us proceed to the promised characterization of modally definable
classes of frames. The easiest case is that of classes of descriptive frames. In
descriptive frames, the correspondence is as exact as possible.
THEOREM 4.7.5. A class of descriptive frames is modally definable iff it
is closed under coproducts, p-morphic images and generated subframes.
PROOF. This follows from the Duality Theorem as follows. Suppose that
3: is modally definable. Then 3:+ is equationally definable, hence a variety.
Therefore 3: is closed under coproducts. For if 3i, i E f, is a family of
descriptive frames from 3: then ( ~ i e s 3i)+ ~ 1-[i~i(3i)+. Since 3:+ is closed
under products the latter is in 3:+, and thus is the former, by closure under
isomorphisms. But since (3:+)+ ~ 3: we also have ~ i e I 3i C 3:. Analogously
it is shown that 3: is closed under generated subframes since 3:+ is closed
under homomorphic images, and that 3: is closed under contractions since 3:+
is closed under subalgebras. Conversely, if a class 3: of descriptive frames is
closed under coproducts, 3:+ is closed under products. If 3: is closed under
generated subframes, 3:+ is closed under homomorphic images, and if E is
closed under contractions, 3:+ is closed under subalgebras. []
We know that two frames 3 and ® whose algebras are isomorphic have
the same modal theory. Hence, if we want to see what classes of frames are
modally definable, we just have to close it under the operation B : 3 ~-+ (3+) +
as well as B -1. (3+) + is known as the b i d u a l of 3.
THEOREM 4.7.6. A class of frames is modally definable iff it is closed
under disjoint unions, p-morphic images, generated subframes, and if it and
its complement are closed under biduals.
PaOOF. Suppose t h a t X is modally definable. Then it is closed under

biduals, and its complement is closed under biduals as well. By the previous
theorem, X is closed under p - m o r p h i c images and generated subframes. Fur-
thermore, the bidual of ~ i ~ I ~i is nothing but the coproduct in DFrm. Since
9~ is closed under coproducts and inverse biduals, :~ is closed under disjoint
unions. Suppose then that :~ has all the required closure properties. Then
consider :~d := : ~ A ~ , the class of descriptive frames contained in 9C. We have
:~ -- B-l[~:d], by closure and inverse closure under B. Therefore, the modal
theory of ~ is the same as the modal theory of 3~d, and :~ is the class of frames
satisfying the modal description of :~d. So, we only need to show that 9Cd is
modally definable. ~d is closed under coproducts since the coproduct in DFrm
is the bidual of the disjoint union. ~d is closed under p - m o r p h i c images and
generated subframes by assumption on X. Thus ~d is modally definable. []
E x e r c i s e 162. Recall the notion of a net extension of Chapter 2.4. Prove

that given [, g and 0, an embedding e : [ ~ ~ and a contraction c : f --~ 0,
there exists a e and maps p : ~ -+ e, q : ~ --+ e such t h a t (i) p o e = q o c and
(ii) for all e', p' : g --+ e' and q' : b --+ e' satisfying (i) there exists a unique
i : e ~-~ e~ such t h a t p~ = iop, and q~ = ioq. Prove also that p is a contraction
and q an embedding.
E x e r c i s e 163. Generalize the setting of the previous exercise as follows. Let

A, B and C be objects and morphisms p,q as in the picture below to the
left. This is a special diagram. Say that D together with morphisms B --+ D,
C -+ D is a p u s h o u t of this diagram in a category C if for any other D I with
maps B --+ D/, C --+ D ~ there exists a unique m a p D --+ D ~ making the entire
diagram to the right commute. D together with the maps into D is called a
co-cone of the diagram to the left. The dual notion is t h a t of a p u l l b a c k . It
is based on the opposite or dual of the diagram to the left. Try to formulate
the notion of a pullback. Show that any two pushouts of a given diagram are
isomorphic.
P P
A "B A ~B
C ~ D ~
E x e r c i s e 164. Let A = (Ob, Mot) be a diagram in a category C. A c o n e

for A is a pair 2. = (L, (Pc : C COb)) where L is an object and P c : L --+ C
such that for any pair C, D E Ob and C ~ D E Mor we have PD = f o pc. A
220 4. U n i v e r s a l A l g e b r a and D u a l i t y Theory
limit for A is a cone t2 such t h a t for any cone t2' for A we have a uniquely
defined m a p i : L ~ --+ L such t h a t p ~ = P c o i. Now show t h a t p r o d u c t s and
pullbacks are limits for special kinds of diagrams. Dualize to define the notion
c o - c o n e and c o - l i m i t and see where you can find instances of c o - l i m i t s .
Exercise 165. Show t h a t there exist 2 ~° ultrafilters on p(co). (In fact, t h e r e

are 22~° many.) Hint. Consider the sets In := { i : i = 0 ( m o d Pn)}, where Pn
is the n t h prime number. For each M C_ a~ let UM be an ultrafilter containing
[n iff n C M .
4.8. Free Algebras, C a n o n i c a l Frames and D e s c r i p t i v e Frames

In this section we will discuss the difference between canonical frames and
descriptive frames. Before we do so, let us reflect on duality in connection
with canonical frames. We have shown in Section 1.3 t h a t in any given variety
V, for any cardinality a , 3? contains freely a - g e n e r a t e d algebras. F u r t h e r m o r e ,
if v : a --~/3 then ~ : ~ r v ( a ) ~ ~ v ( ~ ) and if w : a -~ Z t h e n ~ : ~ v ( a ) -~
~ r v ( ~ ) - Now let ~ be the variety of ®-algebras. T h e n d a n e ( a ) t u r n s out
to be the dual of ~ r v ( a ) . Moreover, ~+ : ~ano(13) --~ ~ a n o ( a ) and also
~ + : ~ane(13) --~ ~ a n o ( a ) . T h e difference is t h a t while the worlds of C a n e ( a )
are m a x i m a l l y consistent sets the worlds of the dual of 3 t o (a) are ultrafilters.
The difference, however, is negligeable. Let us elaborate on this a little bit.
For given cardinal n u m b e r a let V~ := {pz : /3 < a}. Now let ~ -- X iff
++ X E (9. This is a congruence. Moreover, ~ v ( a ) = T m ( V ~ ) / ~ . T h e
h o m o m o r p h i s m corresponding to =- is denoted by h - and the congruence class
of ~ is denoted by [~].
PROPOSITION 4.8.1. A set A is a maximally consistent set of V ~ - t e r m s
iff there exists an ultrafilter g in ~ r v ( a ) such that A = h:=~[U].
PROOF. Let A be m a x i m a l l y consistent. T h e n it is deductively closed, as
is easily seen. In particular, if ~ E A and ~ = X, then X C A. So we have A =
h_-l [h_=[A]]. Furthermore, the image g of A under h--_ is deductively closed
and so it is a filter, by Proposition 1.7.8. To show t h a t U is an ultrafilter, take
b ~ U. T h e n b = [~] for some ~ ¢ A. By m a x i m a l consistency of A, ~ C A
( L e m m a 2.8.2). Moreover, [-~] = - [ ~ ] = - b . Hence - b C U. If - b c U,
t h e n b ¢ U, otherwise A is inconsistent. So, U is an ultrafilter. Conversely,
assume t h a t U is an ultrafilter, and p u t A := h_-_1 [U]. A is deductively closed.
For let ~ C A and qo -+ X E A. T h e n [~] C U and [~] -+ IX] = [~ -+ X] C g .
U is deductively closed, so [X] C U, from which X E A. A is m a x i m a l . For
given ~, [~] ~ U or [=T] = - [ ~ ] ~ U. Hence T ~ A or - ~ ~ A. B u t not b o t h ,
since not b o t h [~] ~ U and -[qo] ~ U. []
PROPOSITION 4.8.2. The map h51 is an isomorphism f r o m ~ r ~ ( a ) + onto
Cane(a).
4.8. Free Algebras, Canonical Frames and Descriptive Frames 221
The proof of this proposition is left as an exercise. Now let v : a -4 fi

be a function. Denote by v also the function p~, ~-~ Pv(~,), # < c~. Then
: ~ r o ( a ) -+ ~re(fl)- Furthermore, if v is injective iff g is, and v is surjective
iff g is. T h e n F + : ~ro(fi) + --9 ~ r e ( a ) +, by duality. Furthermore, F+ is
injective iff g is surjective, and surjective iff g is injective. Recall also how
g+ is defined; given an ultrafilter U of ~re(fi), ~+(U) := g - l [ U ] . Now recall
from Section 2.8 the m a p Xv : ~m(Vz) --+ !rn(Vz), taking a f l - t e r m to its
preimage under v. This m a p can be shown to m a p maximally consistent sets
onto maximally consistent sets. Namely, given A, put
Yv(~) := h - o ~ o h_=[Z~]
LEMMA 4.8.3. For a maximally consistent set A, yv( A ) = X v ( A ).
The proof is straightforward but rather unrevealing. The T h e o r e m 2.8.11

follows in this way from the duality theory developed in this chapter. On the
one hand duality theory is more general, since there are descriptive frames
which are not canonical for any logic. On the other hand, we will show below
that descriptive frames are generated subframes of canonical frames. This
result of duality theory therefore follows already from T h e o r e m 2.8.11.
DEFINITION 4.8.4. Let A be a modal logic and o~ a cardinal number. A

frame ~ is called a c a n o n i c a l f r a m e f o r A if ~ is an c~-canonical frame for
some c~; and ~ is c a n o n i c a l simpliciter if it is a canonical frame for some
logic.
Recall t h a t the c~-canonical frame for A is isomorphic to ~rh(C~) +. It m a y

appear at first blush that canonical frames are the same as descriptive frames.
This is not so as we will show below. However, every descriptive frame is a
generated subframe of a canonical frame. Namely, fix any logic A and let
be a descriptive frame for A. Then ~ + is a A-algebra. Every A-algebra is the
image of a free A-algebra by Theorem 1.3.5. Consequently, ~ ~-~ ~ r h ( a ) for
some c~. We will use this fact to derive a useful characterization of canonicity
of logics first shown in [186]
DEFINITION 4.8.5. A logic A is called d - p e r s i s t e n t if for every descrip-

tive frame ~ for A the underlying Kripke-frame, ~ , is a A-frame as well.
THEOREM 4.8.6 (Sambin & Vaccaro). A logic is canonical iff it is d-per-

sistent.
PROOF. Since every canonical frame is descriptive, d-persistence implies

canonicity. So let us assume that A is canonical and ~ a descriptive frame
such t h a t ~ ~ A. Then ~ ~ ~ for some canonical ~ frame for A. Because
~ A, we have ~ ~ A, by the fact that A is canonical. But then ~ ~ A,
since ~ ~-~ ~ . []
FIGURE 4.4
..s S'"
x z
We can cash out here a nice result concerning finite model property. A
variety is said to be l o c a l l y finite if every finitely generated algebra is finite.
Obviously, a variety is locally finite if all finitely generated free algebras are
finite. By duality, they are isomorphic to the full algebra of sets over a finite
Kripke-frame. Now say that a logic is l o c a l l y finite if its corresponding
variety is. Say that a logic A has a property ~ e s s e n t i a l l y if every extension
of A has 9 .
THEOREM 4.8.7. If 0 is locally finite then every extension of 0 is weakly
canonical and has the finite model property essentially.
Now let us return to the question whether descriptive frames are also
canonical. The question is whether modal algebras are free algebras in some
variety. For example, vector spaces are freely generated by their basis. How-
ever, not all boolean algebras are free, for example the algebra 2 3. (A finite
boolean algebra is freely generated by n elements iff it has 2 2n elements.)
Nevertheless, it is interesting to approach the question to see clearly what
the relation between the two notions is. Let ~ be a descriptive frame and
put 0 := T h ( ~ ) . First, if ~ ~ ~anh(a) for some a and some A, then
~- ~ a n o ( a ) . In other words, ~ is a-canonical in its own variety iff it is
a-canonical. For by duality, if an algebra 91 is freely a-generated in a variety
~?, we have 91 E ~? and so the variety generated by 91 is included in V. How-
ever, 91 is then freely a-generated in any smaller variety containing it, and so
freely a - g e n e r a t e d in the least such variety.
PROPOSITION 4.8.8. Let ~ be a frame. ~ is a-canonical for some logic
iff it is a-canonical for Th 3.
We will need the distinction between an algebra generating a variety and
an algebra free in that variety later on in the connection with splittings. It
is necessary for the understanding of the results proved there to have seen an
explicit construction of a-free algebras and we will provide such an example
now. Consider the frame of Figure 4.4 This frame is not 0-generated; but
it is 1-generated, for example, by the set {y}. Therefore it is a canonical
frame iff it is also freely 1-generated in its own variety. Now, how does the
4.8. Free Algebras, Canonical Frames and Descriptive Frames 223
1-generated canonical frame look like? To t h a t effect recall t h a t the freely

n - g e n e r a t e d algebra in a variety V is a subalgebra of the direct p r o d u c t of
the m e m b e r s of ~?, indexed by n - t u p l e s of elements in the algebras. In the
present context, where :X = {92}, this reduces to the direct p r o d u c t HbEA92.
The freely o n e - g e n e r a t e d algebra is c o m p u t e d as the subalgebra generated by
the function which picks b in the c o m p o n e n t indexed by a. (Recall t h a t the
p r o d u c t is indexed over A, so t h a t each factor takes an index b E A. In each
factor, s takes a value, in this case s(b) = b.) T h e reason why this is so lies
in the following. Let V be generated by :K. Let us call 92 a - f r e e for ~K if
there is a subset X _C A of cardinality a such t h a t for every ~ E :X and m a p s
v : X --+ B there is a h o m o m o r p h i s m ~ : 92 --+ ~ . T h e difference with the
concept of a freely generated algebra is t h a t 92 E :K is not required; moreover,
92 need not be generated by X . But the following holds.
PROPOSITION 4.8.9. Let 92 be a-free for tK. Then 92 is a-free for HSP(K).
PROOF. Let 92 be a - f r e e for :K. We show t h a t t h e n 92 is a - f r e e for the
classes H(tK), S(:K) and P(tK). To simplify the a r g u m e n t a t i o n , let us first
remark t h a t if 92 is a - f r e e for :K, then it is a - f r e e for I(K), the closure of :K
under taking isomorphic copies.
(1.) Let C C H(tK). T h e n there is a ~ E :X and a h o m o m o r p h i s m h : ~ --~ ~.
Now take a m a p m : X --+ C. There exists a m a p n : X -+ B such t h a t
m = h o n . (Just let n ( x ) : = a for some a e h - l ( m ( x ) ) . ) B y a s s u m p t i o n there
exists a h o m o m o r p h i s m ~ : P2 -+ ~ extending n. T h e n h o ~ : 92 --+ ~ is a
h o m o m o r p h i s m extending m.
(2.) Let ~ E S(:K). T h e n there is a ~ E K such t h a t ~ _< ~ and so C _C B.
Let i : C -+ B be the inclusion map. Let m : X --+ C be a map. T h e n
i o m : X --+ B and by assumption there is an extension i o m : 92 --+ ~ .
However, the image of this m a p is contained in C, and so restricting the
target algebra to ~ we get the desired h o m o m o r p h i s m ~ : 9/--+ ~.
(3.) Let ~ E P(tK). T h e n there are ~3i, i E I , such t h a t ~ = 17Iiex ~ i - We
m a y assume E ---- 1-L~I ~ i . Now take m : X --~ C. T h e n for the projections
Pi we have Pi o m : X -+ Bi, and by assumption there are h o m o m o r p h i s m s
Pi o m : 9 / - + ~ i . By the fact t h a t the algebraic p r o d u c t is a p r o d u c t in the
categorial sense there is a unique f : 92 --+ r l i ~ i ~3i such t h a t Pl o f = pi o m .
T h e n f extends m. []
~Fhe present algebra, the algebra of sets over the frame [, has eight elements.
Thus, the freely o n e - g e n e r a t e d algebra is a subalgebra of 92s. T h e eight choices
• re d i a g r a m m e d in Figure 4.5 below; in each copy the set of elements which
~re values o f p are put into a box. All we have to do is to calculate the algebra
generated in this complicated frame. However, we are helped by a n u m b e r
of facts. First, [ a d m i t s an automorphism, n a m e l y x ~4 x, y ~4 z, z ~+ y.
By this a u t o m o r p h i s m , III is m a p p e d into IV and v into vI. All other models
FIGURE 4.5
I II III
< < j-
IV V
VII
are mapped onto themselves. This fact has as a consequence that the algebra
induced on Ill and IV jointly (on the underlying frame ~ O [) is isomorphic
to the one nI induces on ~ (and isomorphic to the one induced on IV on its
copy of ~). Hence, we can drop IXI and v in the direct sum. Next, the frames
induced on I, II, VII and viii are not refined. I and II are actually isomorphic
to a one-element frame, vii and VIII to a two-element chain. This gives a
reduced representation of the underlying frame as the direct sum of two one-
element chains, two two-element chains and two copies of ~. The general
frame is still not refined. Its refinement is the frame shown in Figure 4.6.
(The frame is shown to the left. To the right we repeat the frame, with
worlds being numbered from 1 to 6.) It might be surprising that the frame
just constructed should indeed be canonical, as it contains more points. But
notice that this frame is rather regular, having 8 automorphisms, and that
Pgut(Cano (1)) ~ Z2 xZ2 xZ2, generated by the permutations (1 4)(2)(3)(5)(6),
(1)(2 3)(4)(5)(6) and (1)(2)(3)(4)(5 6). The orbits are {1, 4}, {2, 3} and {5, 6}.
There is up to automorphisms of the frame only one generating set, containing
one world from each orbit of the group.
4.9. Algebraic Characterizations of Interpolation 225
FIGURE 4.6. ~aao(1)
-~ p 2
-~p 3
N o t e s o n t h i s s e c t i o n . Canonical frames have been heavily used for ob-

taining results, such as completeness results. Yet, their structure is not well-
understood. Some new results can be found in the thesis by TIMOTHY SUREN-
DONK ([204]). It is not known, for example, whether there exists a cardinal
number a such that if a modal logic is a-canonical it is also ~-canonical for
any ~.
* E x e r c i s e 168. Let S be a set, and let ~ be a group of permutations of S.

We say that ~5 is t r a n s i t i v e on S, if for every given pair of points (x, y) C S 2
there exists a g c G such that g ( x ) = y. We say that ~ is s h a r p l y tran-
sitive if at most such g exists given (x,y). Now let ~ A ( A ) be the algebra
freely generated by A m a n y elements. Call a function f : a --+ FrA(A) an
M - s y s t e m if f[a] is a maximal independent subset of ~rA(A). (A set H _C A
is called i n d e p e n d e n t in an algebra 2[ if for all a C H , a is not contained in
the subalgebra generated by H - {a} in P2.) Show that 92ut(~rA(A)) is sharply
transitive on the set of M - s y s t e m s of ~rA (~).
E x e r c i s e 169. Let (9 be a consistent modal logic and n a natural number.

Show that 92ut(~re(n)) has a subgroup of size n ! . 2%
4.9. Algebraic Characterizations of Interpolation

Categories of frames have coproducts since they are dual to the algebras.
However, to have coproducts is a rather rare property. This accounts in a way
For the fact that properties such as interpolation and Ha]ld~n-completeness are
rather rare. We will prove in this section two standard results on interpolation,
both shown by LARISA MAKSIMOVA in a series of papers, and then derive some
useful characterizations of Halld~n-completeness.
DEFINITION 4.9.1. A variety V of polymodal algebras is said to have the

a m a l g a m a t i o n p r o p e r t y if for any triple 920, 921 and 922 of algebras in V
and embeddings il : 91o "-~ 921 and i2 : 920 ~ 922 there exists an algebra
923 ff V and maps el : 921 ~ 923, e2 : 922 ~-* 923 such that el o i l = e2 oi2- 52 is
said to be have the s u p e r a m a l g a m a t i o n p r o p e r t y if in addition the ei can
be required to have the property that whenever el(a~) < e2(a2) for al C A1,
a2 C A2 there exists an ao E Ao such that al <_ Q(ao) and i2(a0) _< a2.
921.,.
i "'..el
/ ".,
920 923
/2~ .'" .."e2

...,e
92
THEOREM 4.9.2 (Maksimova). Let A be a poIymodal logic. Then the fol-
lowing are equivalent.
1. A has local interpolation.
2. The variety of A-algebras has the superamalgamation property.
PROOF. Assume t h a t A has local interpolation. Let 920, 9-11 and 912 be
A-algebras and ij : 920 "-~ 92j (J E {1, 2}) be embeddings. W i t h o u t loss of
generality we can assume that A0 C A1M A2. For each element a E A1U A2 fix
a variable Xa. We assume all these variables are distinct for distinct elements.
Denote by 3i, i e {0, 1,2}, the A-algebra freely generated by {xa : a E A i } .
Denote by 33 the algebra generated by {Xa : a C A1 U A2}. There are natural
embeddings 3o ~-~ 3i ~-~ 33, i C {1,2}. Also there are h o m o m o r p h i s m s
bl : 3i --~ 91i, i E {1,2} defined by bl : xa ~+ a and b2 : Xb ~-~ b. The m a p s
assign the same value to each Xa where a C A0. Now let T1 := { p : bl(~) = 1}
and T2 := { T : b2(~) = 1}. P u t
T := {X : T1 U T2 I~-, X}.
Now we show that the following holds for { i , j } = {1,2}
T IVA ~ --+ ¢ ¢~ (Bx C Fo)(p --+ x E Ti & x --~ ¢ ~ Tj) .
From right to left is clear. Now assume t h a t T it-A ~ --+ %b. T h e n for some
finite set F1 C_ T1 and some finite set F2 C T2 we have F1;F2 IbA ~ ~ ¢ and
so for some compound modality [] we get • F 1 ; []F2 PA ~ -+ ¢- Thus
PA ~o A raP1. --+ . [ ] r 2 -+%b.
There exists now a local interpolant X, that is, a formula using the c o m m o n
variables of ~ and ¢ and bh (mF1 A ~) --+ X as well as ~-a X -+ ([]F2 -+ ¢).
With this X we have F1 I~-A ~ --+ X and F2 J~-A X -+ ¢. Thus we have

~ - - + X C T1 and X - - + ¢ C T2.
Now define W® ¢ by T I~-A ~ /-+ ¢ . This defines a congruence on 33, since

it defines an open filter. Now put 9/3 := 3 3 / 0 . There is a h o m o m o r p h i s m
b3 : 33 --~ 9/3- We will show that the filters defined by O on the algebras
31 and 32 are exactly T1 and T2. (See picture above.) This means that
ker(b3) r Fi = ker(b~) (i E {1,2}). Namely, for ~p C F1 we have ~ O 1 iff
T IF-A T --+ ~. This means that for some X in the common variables we have
T --+ X E T2 and X --+ ~ E T1. Then X is constant and since T C T1 we
also have T --+ X C T1 showing T --+ ~ C T1, that is, ~ C T1. Likewise we
show that ¢ ® 1 implies ¢ E T2 for ~ C F2. Denote by ci the restriction of
b3 to 3i, i = 1,2. So ci : ~i --+ 9/3. We have shown that ker(ci) = ker(bi).
There now exist maps e~ : 9/~ ~ 9/3 (i G {1, 2}) such t h a t ci factors through
ei, and el o bl = cl as well as e2 o b2 = c2. Finally, let el (a) _< e2(b). Then
e l ( a ) -+ e2(b) = I and so x~ -+ xb ® l which means t h a t T I~A x~ --+ Xb.
From this we get a X E Fo such that x~ --+ X E T1 and X --+ Xb E T2, and so
a = bl(x~) <_ b l ( x ) as well as b2(x) _< b2(xb) = b. Moreover, b l ( x ) = b2(x),
since bl and b2 assign the same value in Ao to X, which is a common subalgebra
of b o t h 9/1 and 9/2. This shows that the superamalgamation property holds
for V.
Now assume t h a t the variety of A-algebras has the s u p e r a m a l g a m a t i o n
property. Let p = ~(iff,r~ and ¢ = ¢(~', q~ be formulae such that for no X
based on the variables ~ we have F-A ~ --+ X; X --+ ¢. Let 30 be the algebra
freely generated by ~, 31 the algebra freely generated b y / ~ and ~, 32 the
algebra freely generated by g and ¢ and 33 the algebra freely generated by
p~, g and ~'. Let U1 be an ultrafilter on 33 containing ~, and let V be the
ultrafilter induced by U1 on the subalgebra 30. Then V U { - ~ } has the finite
intersection property in 33 and so there exists an ultrafilter U2 containing it.
We then have U2 n Fo = U1 n Fo. Let @1 be the largest congruence contained
in U1, and O2 be the largest congruence contained in U2. P u t • := 01 U 02.
(The largest congruence induced by a filter F is the same as the congruence

induced by the largest open filter contained in F; and this is equal to the set of
elements a such t h a t [~a C F for all c o m p o u n d modalities [~.) For elements of
F0, u 0 1 v iff for all c o m p o u n d modalities [~(u ++ v) C U1 iff for all c o m p o u n d
modalities 7~(u ~ v) C V iff for all c o m p o u n d modalities [](u ++ v) C U2 iff
u 02 v. P u t 0 o := ¢I,N (Fo × F0) and 92o := 3o/®o. T h e n we get an e m b e d d i n g
il : 920 ~-~ 9.11 as well as i2 : 920 "-~ 922. Assume now t h a t we have an element
X C F0 such t h a t ~ -+ X 0 1 1 and X --+ ¢ 8 2 1 . T h e n X E U1 since ~ C U1,
and t h e n also ¢ C U2. But this contradicts our choice of U1 a n d U2. Hence no
such element exists. By s u p e r a m a l g a m a t i o n , however, we get an a l g e b r a
and m a p s e~: 92~ ~-~ ~ such t h a t el o il -- e2 o i2 and el([T]81) :~ e2([¢]82).
Now define a m a p b : 33 --+ ~ by b(pi) := el([pi]81), b(qj) := e2([qj]82) as
well as b(rk) = el(Irk]81) = e2([rk]82). This is uniquely defined and we have
b(~) ~ b(¢) from which FA ~ -~ ¢. []
THEOREM 4.9.3 (Maksimova). Let A be a polymodal logic. Then the fol-
lowing are equivalent.
1. A has global interpolation.
2. The variety of A-algebras has the amalgamation property.
PROOF. T h e p r o o f of a m a l g a m a t i o n from global interpolation is a c t u a l l y
analogous to the previous one. So let us prove t h a t a m a l g a m a b i l i t y implies
global interpolation. We assume t h a t ~ tFA ~b but no interpolant exists. Define
~:0 to be algebra freely generated by the c o m m o n variables of ~ and ~b, and 33
the algebra generated by all the variables of T and ~. Let O1 be the o p e n filter
generated by T, and 02 an open filter containing - ~ and O1NF0. Such a filter
exists. T h e n O1 Vl F0 = 02 A F0. Let 8 i be the congruence associated with
Oi. Now p u t 921 := 3 3 / 8 1 and 922 := ~ 3 / 8 2 . T h e n as before for elements
of F0, u 81 v iff u 02 v, and hence 81 and 02 induce the s a m e congruence on
30. Therefore, we have an embedding il : 920 ~-~ 921 and i2 : 920 ~-~ 922. Thus,
assuming the a m a l g a m a t i o n property, there is an algebra ~3 and m o r p h i s m s
ei, i = 1,2, satisfying el o il = e2 o i2. Define v by v(p) := el([p]81) if
p C var(~) and v(p) := e2([p]82) if p C var(¢). This is noncontradictory.
Since we have T E O1 we also have ~([~T) = 1 for all c o m p o u n d modalities.
Since -~b E 02 we have ~(~b) ~ 1, and so ~ ~A ~. []
T h e results on a m a l g a m a t i o n p r o p e r t y can be i m p r o v e d by showing t h a t
923 enjoys a so-called universal property. This m e a n s t h a t given 920, 921 and
922, e m b e d d i n g s ij : 920 "-~ 92~ (J E {1,2)) there exists an 923 a n d m a p s
ej : 92j ~ 923 such t h a t el o il = e2 o i2 and for every algebra ~3 t o g e t h e r
with m a p s dl : 921 --+ %, d2 : 922 ~ ~3 with dl o i l = d2 o i2, t h e n there
exists a unique h o m o m o r p h i s m h : 923 --+ ~3 such t h a t dl = h o el and
d2 = h o e2. Namely, consider the m a p s Vl : x~ ~-+ dl o bl(x~) = d l ( a ) ,
v2 : x~ ~-4 d2 0 b2(xa) -= d2(a). T h e n Vl and v2 agree on the elements in Ao,
and so v := vl U v2 is well-defined. It extends to a unique h o m o m o r p h i s m

: 3 3 ---} ~ . We have that O1 := ker(~) [ 31 = ker(dl o 51) and O2 :=
ker(~) [ 32 = ker(d2 o 52). P u t O := O1 U 02. As in the previous proof it
is shown that (~ I F 1 = O1 and O I F 2 = 02. Moreover, O is the kernel of
the m a p ~ and includes ker(b3). Thus it can be factored uniquely through
b3, yielding a h o m o m o r p h i s m h : 9/3 --+ ~ with the desired properties. In the
sense of the definitions in the exercises of Chapter 4.7 what we have shown
is that a variety of modal algebras that has a m a l g a m a t i o n also has pushouts
for those diagrams in which both arrows are injective.
Now we turn to local and global Halld~n-completeness. Recall that if a
logic is locally or globally Halld~n-complete it has trivial constants; another
way of saying this is that the freely zero-generated algebra contains exactly
two elements (if A is consistent). Then for every pair of nontrivial algebras
921 and 922 we can find an 920 and injections i0 : 92o ~ 921 and il : 920 ~-~ 922.
Simply take the zero-generated subalgebra; if the algebras have more than
one element, this algebra is isomorphic to 3rA(0).
DEFINITION 4.9.4. A variety V has f u s i o n if for every pair 921,922 C V of

nontrivial algebras there exists an algebra ~ and embeddings el : 921 >-* ~ and
e2 : 922 ~-* ~3. V has s u p e r f u s i o n if for every pair 921,922 C V of nontrivial
algebras there exists an algebra ~ and embeddings e~ : 92i ~ ~ such that for
every a • A1 - {0} and every b • A2 the inequation il(a) < i2(b) does not
hold.
The following theorem can be found in a slightly different form in [153].
lowing are equivalent
1. A is locally Hallddn-complete.
2. The variety of A-algebras has superfusion, and the zero-generated al-
gebra contains at most two elements.
The global version is as follows:

lowing are equivalent
1. A is globally Hallddn-complete.
2. The variety of A-algebras has fusion, and the zero-generated algebra
contains at most two elements.
For a proof, let A be (locally/globally) Halld~n-complete. We m a y assume

that A is consistent; otherwise the equivalence is clearly valid. Take two
algebras 921 and 922- We can embed 3rA(0) into both of these algebras. Now
follow the proofs of the Theorems 4.9.2 and 4.9.3. We obtain an algebra ~3
and embeddings ei : 92i ~-~ ~ . The superfusion condition is verified for the
local case. For the converse, let the variety of A-algebras have (super)fusions,
and let the zero-generated algebra have two elements. Then enter the second
half of the proof of Theorem 4.9.2 with T and ¢ assuming that for no constant
proposition both ~ ~-A X and X ~-h ¢. But either X = T and then ~ZA ¢, or
X = ± and then T FA 5_. Performing the same argument as in the proof we
get that T ]/A ¢- As a consequence we get the following theorem of [15].
THEOREM 4.9.7 (van Benthem 8z Humberstone). Let A be a logic such
that f o r every pair (31, x i ) and (32, x2) of pointed frames there exists a pointed
frame (@,y) and two contractions Pi : ¢~ --~ 3 i p2 : ¢3 --~ 32 such that
p i ( x i ) = p2(x2) = y. Then A is locally Hallddn-compIete.
DEFINITION 4.9.8. A variety V has f i n i t e c o p r o d u c t s if f o r every pair
921,912 of algebras in V there exists a third algebra ~ and maps ii : 921 -+ ~ ,
i2 : 922 --+ ~ such that for every algebra ~ and every pair of maps ei : 911 --+
and e2 : 922 --+ C a unique h : ~ --+ ~ exists satisfying ei = h o ii and
e2 -~ h o i2. We denote % by 911 ® 922.
THEOREM 4.9.9. A logic is globally Hallddn-complete iff it has trivial con-
stants and the corresponding variety has finite coproducts.
E x e r c i s e 170. Give a characterization of those logics whose variety has co-

products, without any restriction on constant formulae.
E x e r c i s e 171. Let Ai, i < co, be logics which have (local/global) interpola-
tion. Suppose that Ai C Aj if i _< j. Show that I lie~Ai has (local/global)
interpolation.
E x e r c i s e 172. Give an example of logics As which have local interpolation,

such that As ~ Aj for i < j, such that N~c~ Ai fails to have interpolation.
CHAPTER 5
Definability and Correspondence
5.1. M o t i v a t i o n
Correspondence theory developed from a number of insights about the
possibility of defining certain elementary properties of frames via modal axi-
oms. For example, transitivity of Kripke-frames may either be characterized
by the first-order formula ( V x y z ) ( x <~y <~z. -+ .x <~z) or by the modal axiom
0 0 p --+ 0p. We therefore say that the axiom 4 corresponds to transitivity
on Kripke-frames. These insights have sparked off the search for the exact
limit of this correspondence. In particular, the following two questions have
been raised by JOHAN VAN BENTHEM in [10], who has also done much to give
complete answers to them.
* Which elementary properties of Kripke-frames can be characterized
by modal axioms?
* Which modal axioms determine an elementary property on Kripke-
frames?
Both questions were known to have nontrivial answers. Irreflexivity cannot be
characterized modally, so not all first-order properties are modally character-
izable. On the other hand, some modal axioms like the G - a x i o m determine
a non-elementary property of frames. Many people have contributed to the
area of correspondence theory, which is perhaps the best worked out subtheory
of modal logic, e. g. JOHAN VAN BENTHEM [8], ROBERT GOLDBLATT [77],
[79]. With HENDRIK SAHLQVIST's classical paper [183] the theory reached
a certain climax. There have been attempts to strengthen this theorem, but
without success. It still stands out as the result in correspondence theory.
Nevertheless, there has been a lot of improvement in understanding it. The
original proof was rather arcane and methods have been found to prove it in a
more systematical way (see VAN BENTHEM [10], SAMBIN and VACCARO [187]
and also KaACHT [121] and [124]).
Meanwhile, the direction of the research has also changed somewhat. Cor-
respondence theory as defined above is just part of a general discipline which
has emerged lately, namely definability theory. Definability theory is the ab-
stract study of definability of sets and relations in general frames. There are
a number of reasons to shift to this more abstract investigation. First, as
231
232 5. D e f i n a b i l i t y and Correspondence
has been observed already in SAMBIN and VACCARO [187], there is a real
benefit in raising the above questions not for Kripke-frames but for frames
in general and suitable classes thereof. For suppose that (as in SAHLQVIST'S
original proof) correspondence of modal axioms of a certain form with first-
order conditions is established not only for Kripke-frames but also for de-
scriptive frames. Then one can immediately derive that logics axiomatized
by such formulae must be complete. First-orderness on Kripke-frames is not
enough for that. Second, the success of canonical formulae of MICHAEL ZA-
KHARYASCHEV for logics containing K 4 (see Chapter 8) shows t h a t there can
be useful geometric characterizations of axioms which are not first-order in
general. Even though first-order properties are well-understood and there is a
powerful machinery for first-order logic, there is nevertheless much to be said
in favour of an a t t e m p t to characterize in whatever terms possible the geomet-
ric condition imposed on frames by an axiom. They cannot all be first-order
as we know, but they may in m a n y cases be simple, as the notorious example
of G shows. The third generalization concerns the use of relations rather t h a n
properties. T h a t is, we ask which relations are characterizable modally in a
given class of frames. This move has m a n y advantages, although we need to
clarify what we mean by a modal characterization of relations since modal
formulae are invariably properties of worlds rather t h a n sequences of worlds.
Nevertheless, we will develop such a theory here and it will be possible to say
for example that a frame is differentiated iff equality is modally definable. In
this way natural classes of frames are motivated by the possibility to define
certain relations.
Definability theory is thus the study of the following questions.
• Given a class ~ of frames, which relations between worlds are charac-
terizable modally in ~?
• Given a class Z of frames, what geometric properties of frames do
modal axioms impose on the frames of ~ ?
From there m a n y questions arise which have been treated in the literature with
sometimes surprising answers. These questions are for example the following.
• Which classes of frames can be characterized by modal axioms?
• W h a t is the relation between closure properties of classes and the syn-
tactic form of definable properties or relations?
5.2. The Languages of Description

Before we can enter the discussion on definability we will fix a suitable
language within which we derive formal results. Such a language is tradi-
tionally seen to be monadic second-order logic. Here, however, we will use a
notational variant, namely two-sorted first-order predicate logic. T h e special
language will be called the external language and denoted by L e. It is two-
sorted; t h a t means, there are two sorts of well-formed expressions, namely
5.2. T h e Languages of Description 233
modal propositions ( c a l l e d / - f o r m u l a e ) and formulae. Moreover, L e has two

sublanguages, L m, the modal language, for talking a b o u t m o d a l propositions
and L f, the f r a m e language, for talking a b o u t worlds. L m is in fact isomorphic
to the basic m o d a l language we are operating in. R e m e m b e r t h a t there is not
just a single m o d a l language, but a whole family of t h e m with varying n u m b e r
of m o d a l operators and varying n u m b e r of variables and constants. For the
m o m e n t we assume t h a t we have c o u n t a b l y m a n y propositional variables and
no constants, t h o u g h n o t h i n g depends on that. Of proposition variables and
constants there will otherwise be as m a n y as needed. Similarly, depending on
the cardinality a of basic m o d a l operators L m will have the following symbols
* proposition variables p - v a t = {pi : i C ~ } ,
* boolean connectives T, 9, A,
* m o d a l connectives [Sj, j < ~.
All symbols are standard, and their interpretation will be as well. Notice t h a t
0 j is not a primitive symbol. This is done to make the subsequent discussion
simpler. N o t h i n g hinges essentially on that. Now the frame language L f has
* world-variables w-var :-- {wl : i E ~},

* equality -'-, and relations <~i for i < a,
* logical j u n c t o r s t, 4, A and -%
* quantifiers V and 3.
(Here t is a constant, which receives the value true.) Finally, L e has two more
ingredients, n a m e l y
* a m e m b e r s h i p predicate e,
, proposition-quantifiers V and 3.
For a rigorous definition of formulae we need two sorts, propositions and
worlds, over which the two sorts of quantifiers m a y range. Moreover, we define
two sorts of well-formed expressions, i n t e r n a l f o r m u l a e ( i - f o r m u l a e ) and
e x t e r n a l f o r m u l a e ( e - f o m u l a e ) . T h e y are c o m p o s e d as follows
1. pi, i < w, T, ± are i-formulae.
2. If ~ and ¢ are i-formulae, so are ~ , ~ A ¢ , [ ] j ~ (j < ~).
3. If T is an i-formula and i < w then wi e ~ is an e-formula.
4. t is an e-formula.
5. wi - w j and wi <~k w j are e-formulae for all i, j < ~, k < n.
6. If ~ and U are e-formulae then so are -"4, ~ A U and ~ -+ U.
7. If ~ is an e - f o r m u l a and i < ~ then (Vwi)~ and (3w~)¢ are e-formulae.
8. If ~ is an e - f o r m u l a and i < ~ then (Vp~)~ and (3pi)~ are e-formulae.
An e - m o d e l is a triple (~,/3, c) with ~r a frame, and w i t h / 3 : p - v a r --~ G and

: w-var -+ g. Given an e - m o d e l (~,/3, c) and an e - f o r m u l a ~, the relation
(¢5,/3, c) ~ ~ is defined inductively as follows. (Here, given two functions f
234 5. D e f i n a b i l i t y a n d C o r r e s p o n d e n c e
and g, f ~ g abbreviates the fact that f ( x ) 7£ 9 ( x ) only if x = a.)
(¢3,/3,5) ~ w~ e cz ¢# 5(~) e ~(~)

(~,/3, 5) > w~ - w~ **
(~,/3, ~> ~ 9¢
(e,/3,5} ~ ¢ A r/ ¢:~ (~3,/3, L) > ~ and (~3,/3, 5) ~ U
from @5,/3, 5} D ¢ follows (qS,/3, 5} N 77
(e,/3, 5> ~ ( w , ) ¢ ¢, for all c' ~ , 5 (®,/3, #> > ¢
(e,/3, 5) ~ (~w~)¢ ¢, for some 5' ~-,~ 5 (¢5,/3, 5'} ~
(~,/3, 5> ~ (vpd¢ ¢, for all/3' ~p, /3 (®,/3', 5} ~
(~,/3, ~) ~ (~p~)¢ for some/3' ~"v~ /3 ((5,/3', 5} ~
Further, ~5 ~ ~ iff for all/3 and all 5 we have ((5,/3,5} ~ ~. Notice that we
have used/3(~) in the first clause. This is strictly speaking yet to be defined.
However we assume that/3(~) is computed as before by induction on ~. Notice
that in addition to equality there is one symbol whose interpretation is rather
special, namely e. It must always be interpreted as membership. The following
sentences are theorems of the L~-logic of generalized frames. They show that
the Li-connectives are in principle dispensable. (Here ~ = 7/ abbreviates
-+ rl. A .r/-* ~. Open formulae are as usual treated as if all free variables
were universally quantified.)
wl e ~o A ~b =-- wi e ~ A wi e ¢
wi e ~ ~ ¢ =_ wi e ~o. --+ .wi e ¢
w~ e Djw - (vwk)(w~ <j wk. -+ .wk e ~)
L e is not interesting for us because it defines a logic of structures, but because
it is a rather strong language within which we can express (almost) everything
we wish to say. For example, it contains both first-order properties for frames
and properties expressed by modal axioms. With respect to the latter only
the notation has changed somewhat. We can no longer write ;~ ~ ~o but must
instead write
a ~ (Wo)(Wo e ~).
Also, the so-called s t a n d a r d t r a n s l a t i o n of a modal formula is defined as
follows.
ST(p,x) := x e p
ST(-~,x) := ~ S T ( ~ , z )
ST(~ A ¢, x) := ST(V, x) A S T ( ¢ , x )
ST(E]j~, x) := (Vy)(x <j y. --+ .ST(~o,y))
(In the last clause, y must be a variable not already occurring in ST(~p, x).
By construction, x is always the unique free variable. Clearly, S T ( ~ , y ) is
5.2. The Languages of Description 235
then the same as ST(~o, x)[y/x].) Obviously we have in all frames

(Vx)(x ~ ~ - ST(~, x)).
To conclude this part we will introduce the language ~ . In this language the
quantifiers over worlds are replaced by so-called restricted quantifiers. They
are defined as follows.
(w; >~ ~ ) ~ := (w;)(w~ <k ~ . -+ .~)
(3~; >k ~ ) ~ := (~j)(w~ <~ ~j. A .¢)
The restricted quantifiers can be defined from the unrestricted quantifiers as
shown, but the converse does not hold in absence of weak transitivity. Syn-
tactically, we want to construe the restricted quantifier as follows. It takes an
e-formula { and two variables wl, w j and returns an e-formula. Hence, for
each i < a there is a distinct restricted quantifier. Notice that this quantifier
is said to bind only w j , and that wi is called a r e s t r i c t o r . Let ~ de-
note the language L ~ with restricted world quantifiers instead of unrestricted
ones. Likewise, 5~f denotes the language obtained from L I by replacing the
unrestricted quantifiers by restricted quantifiers. 9~~ can be construed as a
sublanguage of L ~, and :RI as a sublanguage of L f. The restricted languages
are expressively weaker, but the difference turns out to be inessential in the
connection with modal logic. On the other hand, 5~~ and its frame counter-
part :Rf have several advantages, as will be seen shortly. With the restricted
quantifiers we will define the following shorthand notation, corresponding to
compound modalities. Let ~r range over sequences of numbers < t~, and s, t
over sets of such sequences. (If s = {~}, we omit the brackets.)
(vx ~,~ w)~ := ~[w/x]
(3~u~)~ := ~[~/~]
(w ~ ~)~ := (vy >~ ~ ) ( w >~ y)~
( w u~'* w)4 := ( w u~ ~)~ A ( w ~' ~)~
(~x > ~ ~)¢ := (~y ~ ~)(3x ~ y)¢
(3~ ~s~ ~)¢ := (3x ~ w)¢ v (3~ ~ ~)¢
Here, it is assumed that y is not free in ~. Similarly the shorthand notation
x <s y is defined. It corresponds to the compound modality [~s. Sets of the
form (x : w <s x} are called cones. The restricted quantifiers range over
cones. The last language introduced is S f. It has in addition to the symbols
of :Rf all <~ as primitive symbols, where s is a finite union of finite sequences
over ~. Also, it has the following additional axioms
x<lO y = x-y
x <~ y ~ ,(x-x)
x <SUt y -- x <~ y. V .x <~t y
FIGURE 5.1. ~ (left) and ~ (right)
E x e r c i s e 173. Show t h a t no :R¢ and :RI formula can be a sentence, i. e. have

no free w-variables.
Exercise 174. Show t h a t :Rf has nontrivial constant formulae, in c o n t r a s t

to L / .
5.3. F r a m e Correspondence -- An Example

Before we s t a r t with the general t h e o r y we will look at an instructive
example. Let there be just one operator, for simplicity. Consider the a x i o m
a l t l := (}p A (}q. --+ .(}(p A q). It is quite easy to show t h a t a K r i p k e - f r a m e
satisfies this axiom iff the frame is quasi-functional, t h a t is, satisfies the first-
order a x i o m (Vx)(Vy0 E> x)(Vyl ~> x ) ( y o - Yz). Namely, if the f r a m e ~ = <f, <~}
is quasi-functional, and (~, 13, x / ~ (}p A (}q, then there is a successor Y0 ~ P
and a s u c c e s s o r Yl ~ q. But Y 0 - - YI, and so Yo ~ p A q from which we get
x ~ (}(p A q). On the other hand, if ~ is not quasi-functional there is an x E f
and x <~ Yo,Yl for distinct Yo,Yl. P u t ~(p) :--= {Y0} and ~(q) := (Yz}. T h e n
we have Y0 ~ P; -~q, Yl ~ q; -~P and for all other points z we have z ~ -~p; ~q.
W h e n c e x ~ (}p; (}q; -~(}(p A q).
Now lets suppose we have a general frame ~ = <[,IF>. Does it still hold
t h a t it satisfies altz iff it is quasi-functional? Well, one direction is u n c o m -
plicated. If the underlying K r i p k e - f r a m e is quasi-functional t h e n ~ ~ a l t l .
J u s t copy the proof given above. However, in the converse direction we en-
counter problems. W h e n x has two different successors we t o o k two special
sets for ~(p) and fi(q) and there is no guarantee t h a t we m a y still be able to
do so. In fact, the following frame shows t h a t the converse direction is false
in general. Namely, let ~ be based on three points x , y o , Y l with x <~ Yo,Yz.
Define IF := { ~ , { x } , { y o , y l } , { x , y o , y l } } . T h e n ~ := <~,IF} is a frame since F
is closed under boolean operations and under ¢, as is quickly c o m p u t e d . (See
Figure 5.1.) T h e valuation t h a t we used to show t h a t a l t l can be refuted is
now no longer available. Moreover, the m a p p : x ~-~ u, Y0 ~-~ v, Yz ~+ v is a p -
m o r p h i s m onto the quasi-functional frame ~t. Moreover, ]F is the p - p r e i m a g e
of the powerset of {u, v}. Thus, we actually have ~ ~ altz.
5.3. Frame Correspondence - - An E x a m p l e 237
We see t h a t the correspondence between first-order properties and m o d a l

properties m a y break down if we pass to a larger class of frames. This is
an i m p o r t a n t point. In most of the literature on correspondence t h e o r y only
the problem of correspondence with respect to K r i p k e - f r a m e s is discussed.
A l t h o u g h this is by itself a legitimate problem, it is for reasons discussed
earlier advisable to broaden the class of frames to be looked at. We can
also put the question on its head and ask how large the class is for which
the correspondence between a l t l and quasi-functionality can be shown. T h e
way to a p p r o a c h t h a t question is to look for sets which can serve as values
for valuations falsifying a given formula once the corresponding first-order
condition is not met. For example, if we have a violation of quasi-functionality
because x <~Y0, yl for Y0 ¢ Yl we want to be able to produce a valuation/3 such
t h a t (~,/3, x) ~ a l t l . Above we have chosen/3(p) = {Y0} a n d / 3 ( q ) = {Yl}, so
we conclude t h a t in atomic frames the correspondence will still hold. But this
is not such a g o o d result. Consider the frame ~ = ([~, ]H~ where h = {x} U { y i :
i C w} and x<Syi, but no other relations hold. Finally, H is the boolean algebra
generated by the sets of the form r(i,k) : = {Yn : (3e C w)(n = i . ~ + k)},
where k < i. T h u s H consists of finite unions of such sets possibly with {x}.
H as defined is closed under complements and unions, as an analogous frame
constructed in Section 4.6. T h e point a b o u t this frame is t h a t it is not atomic
but nevertheless the correspondence holds. For pick any Yi and yj with i ¢ j.
W h a t we need is sets a0, al such t h a t Yi C a0, yj C al, and Yk ~ a0 r'/al for
a l l k . A s s u m e i < j. T h e n put al : = r ( j , O ) - r ( j , i ) and a0 : = r(j,i). We
have al A a0 = ~ by construction and y~ 6 a0, yj E al. So, such sets exist for
all choices for offending triples x, Y0, Yl.
In general it is sufficient t h a t ~ be differentiated. For let us assume t h a t
x sees two distinct points Y0 and Yl. T h e n differentiatedness guarantees the
existence of a set a such t h a t Yo C a but Yl ¢ a. P u t t i n g / 3 ( p ) : = a,/3(q) :--- - a
we get the desired valuation proving ~ 1L a l t l .
PROPOSITION 5.3.1. A differentiated frame is a frame for K . a l t l iff the
underlying Kripke-frame is quasi-functional.
Now consider the frame ~ in Figure 5.2. This frame is not quasi-functional,
it is not differentiated, but it also does not satisfy a l t l . Thus, the result above
is still not optimal. T h e reason for this failure is easy to spot. On the one
h a n d we have x <~Yl and x <1Y2 and there is a set a such t h a t Yl C a b u t Y2 ~ a.
This alone suffices to establish the equivalence between failing a l t l and not
being quasi-functional. However, on the other h a n d we also have x <~ Y0 and
x <~ yl and Yo and Yl are not separable by a set. So, w h a t we have shown in
the cases above is t h a t no m a t t e r w h a t triple of points x, y, z we choose such
t h a t x < ~ y , z there always is a s e t a such t h a t y E a but z ¢ a. T h e frame
here does not have this property. To distinguish these two properties let us
say t h a t quasi-functionality corresponds c a s e w i s e to a l t l if for every choice
FIGURE 5.2
of points x <~y,z such that y ~ z there is a set a with y E a but z ~ a. And

let us say that quasi-functionality corresponds simply if from the failure of
quasi-functionality somewhere we can deduce the existence of triple x <~ y, z
and a set a such that y E a but z ~ a. Simple correspondence is clearly
weaker.
PROPOSITION 5.3.2. In the class of differentiated frames the property of

being quasi-functional corresponds casewise to the property of being a frame
for K. altl.
Casewise correspondence is not actually the same as local correspondence,
the notion we are ultimately interested in. Local correspondence is defined
only with reference to the root x. Namely, quasi-functionality locally corre-
sponds to a l t l in a frame ® if
® (vu)[(vy0 u)(vyl - yl) - (Vp)(Vq)(u
The frame ~ above satisfies this correspondence as well, showing that the
casewise correspondence as just defined is really weaker. Nevertheless, it
seems a rather artificial concept to begin with. The problem is, however, that
the schema above can be written down rather nicely. On the left hand side
we have a first-order formula satisfied at u, and on the right hand side we
have a modal formula satisfied at u. Notice that the ~ is ambigous here. On
the left we would have to construe the statement as being expressed in ~ f ,
the frame part of the external language, whereas on the right hand side we
have the ordinary ~ from the internal language modal logic, and not from the
modal fragment of the external language. It is now in principle possible to
rephrase the left hand side to state that we have a concrete triple violating
functionality.
(~,~)~x<yo A x<yl A Y0#Yl
On the left hand side we have several such statements:
5.4. The Basic Calculus of Internal Descriptions 239
Thus, we can define (strong) local correspondence by requiring t h a t the vio-

lation of quasi-functionality is equivalent to the simultaneous satisfaction of
three internal formulae, one at each of the worlds which constitute the triple
violating quasi-functionality. It is this latter formulation of correspondence
that we will take as the key definition. If we use L ~ here, we can write this
sequence of conditions as
(3,]~,b}~xe-10(pAq) A Y0gP A Y l e q
Furthermore, we can abstract from the valuation/3:
A yo A e q)
Exercise 175. Show that Y~ is a frame. In particular, show that the intersec-
tion of two sets r(i0, k0), r(i~, k~) is a finite union of sets of the form r(i, k).
Exercise 176. Show that for tight frames 3 and points x , y , z such t h a t
(3,~} ~ x <3 y <3 z ; x ~ z iff for some valuation x ~ ~(}p, y ~ -1- and z ~ p.
Hence, transitivity corresponds locally to .4 in the class of tight frames.
5.4. T h e Basic Calculus of Internal D e s c r i p t i o n s

Informally, we will say that a (first-order) relation ~ ( x 0 , . . . , xn-1) can
be internally described in a given class ~ of frames if we can find a sequence
( ~ 0 , . . . , ~n-1} of modal formulae such that for every frame 3 E :~ we have
(3, 5) > ff(Xo,..., xn-1) ¢~ for some fl (3, fl, c(xo)) D ~o,
• .., <3, 4x -1)>
We will rewrite the right-hand side into
(3,2~, b) ~ X0 ~ ~:~0 A . . . A Zn- 1 ¢ ~n-1
We use overstrike arrows in the following way. Let £ and ~ be n-long se-
quences. Then
A
i<n
Notice t h a t the number i does double duty in xi by both identifying xi and
assigning to it a modal formula ~. This will be rather cumbersome. Thus, to
make the association of the variables with the modal formulae independent
of the index i on the variables xi, we use the following notational device.
We write ([£] to denote (a pair consisting of) the formula ( and a sequence
(x~ : i < n} such that every free variable of ~ is identical to some x~, i < n. (It
is not required that xi is distinct from xj if i # j.) The numbers i < n of an
n-long sequence are also called slots. Given an n-long sequence ~ of modal
formulae, the slots of ([£] are in one to one correspondence with the slots of ~.
Notice that writing {[~] we may nevertheless have ~ ~ fvar(¢), a fact which
we will make use of. We call ~[2] a s l o t t e d f o r m u l a . We emphasize t h a t this
is just a piece of notation, nothing more. If the association between variables
and slots is clear (especially when there is just one variable), we m a y drop
the sequence. Occasionally we will also use subscripts 0, i etc. rather t h a n
the sequence of slots.
DEFINITION 5.4.1. Let X be a class of frames, 4[x0,... ,xn-1] a slotted
Lf-formula and ~ = (Po,..- , ~ n - 1 ) a sequence of length n. We say that
i n t e r n a l l y describes ~ in X if for all ~ E X and all 5, (5, 5) ~ ~(~) iff
(~, ~) ~ ~ ~ ~. Symbolically, we write ~[~] ~'~x ~, or simply ~[~] ~¢~ ~. Given
X and ~ we say that ~ is e l e m e n t a r y i n X if an ~ C L f exists which is
described by ~ in X.
Notice first of all that internal describability of ¢ itself is Le-definable.
Namely, (~, 5) ~ 2 e ~ is just a shorthand for the conjunction of xi e ~ . Now
we have the following equivalence
C 7 ¢* x - 7))
Here :ff collects all variables of ~5. A particularly interesting example of an
elementary sequence are those of length 1. In that case it is of the form ~0,
i. e. an ordinary proposition. ~0 is elementary in X iff there is an ~(x0) E L f
such t h a t
X ~ (Vxo)(¢(xo) - ( ~ ( x o ~ ~ o ) ) .
Alternatively,
X ~ ( V x o ) ( ~ ' ( ~ o ) -- . ( V ~ ( x o ~ ~ o ) ) •
If the latter holds we say that ~ 0 d e f i n e s ~ [ x 0 ] and is e l e m e n t a r y , and
that ~ is m o d a l l y d e f i n a b l e . The same definition could be generalized to
sequences, but this is of little benefit. We say t h a t a logic is X - e l e m e n t a r y
if all of its axioms are elementary in X. In addition to this definition of
elementarity, which for distinction will be called local, there is also a global
elementarity. Namely, we say that ~ is g l o b a l l y e l e m e n t a r y in X if there
exists an L f - s e n t e n c e ~ such that for all ~ E X we have ~ ~ ~ iff ~ ~ ~.
Global elementarity is weaker as we have seen earlier, and it will be of little
importance henceforth. Nevertheless, the following theorems can be stated
for global rather t h a n local elementarity.
PROPOSITION 5.4.2. Let X be closed under the map (~,F) ~-+ ~. Then if a
logic is globally X-elementary, it is complete with respect to the Kripke-frames
oyX.
Examples of such classes are the class of differentiated frames, of refined
frames, the class of canonical frames together with the class of Kripke-frames.
With respect to a class X we say that a logic h is p e r s i s t e n t if for all ~ C X
we can infer ~ ~ A from ~ ~ h. This is the general scheme. Moreover,
we have d-persistence, which is persistence with respect to ~ , r-persistence,

which is persistence with respect to 9~ etc.
PROPOSITION 5.4.3. If A is 2~-persistent and :~-complete, A is X~-com-

plete.
PROOF. Let ~ ~ A. Then since A is ~ - c o m p l e t e there is a frame ~ E :~
such that ~ ~ A but ~ ~ ~. Since A is :~-persistent, ~ ~ A. But ~ ]~ ~ as
well, showing A to be ~ - c o m p l e t e . []
PROPOSITION 5.4.4. If A is globally ~ U)~-elementary it is 2~-persistent.
PROOF. Let A = K~ ® A. Each ~ C A is X U : ~ - e l e m e n t a r y , whence

there is an elementary sentence ~ such that ~ ~ ~ i f f ~ ~ ~ for all ~ C .t~U.~I~.
Now assume ~ ~ A. Then ~ ~ A. By :~-elementarity of A, ~ ~ (~ for all
p E A, hence ~ ~ ~ , for all p C A since each (~ is an L / - s e n t e n c e . So
~ ~ A, by : ~ - e l e m e n t a r i t y of A. Consequently, ~ ~ A. []
Let us now turn to the problem of determining which statements of the

form '~ ~ - ~ ~5' are valid. In the sequel we will be concerned with five ba-
sic choices for :~, namely the class of all frames, the class of differentiated
frames, tight frames, refined frames and of descriptive frames. Of course if
~x ~ and ~ _~ ~ then also ~ ~'~g9 ~ as well, so not all work has to be
done separately. In this section we will introduce a calculus for deriving cor-
respondence statements in a quasi-logical style, with axioms and rules. This
calculus, called Seq, will be correct for ~ and hence for all classes. As starting
sequences one can in all cases take a constant formula. Recall namely, that
if ~ is constant, so is the standard translation ST(p, x). Consequently, the
standard translation is first-order. Now we always have
~ (V~(W)(x e p. - .ST(p, x))
In this special case, the propositional quantifier is superfluous since there are
no variables to be bound. This shows that we always have
(axiom.) ST(p, Xo)[x0] ~'~x P if var(p) =

Of course, this can be stated for sequences as well. So we do have some non-
trivial correspondences to start with. The next set of rules is rather obvious.
We m a y add, for example, an inessential variable. If (Ix0....-x,~_~] is a condi-
tion on the n - t u p l e ( x 0 , . . . , x~_~), we can nevertheless view it as a condition
~n n + 1-tuples ( x 0 , . . . , x~} (which we abbreviate by ~. xn). There is then no
condition on x~, and so on the modal side this corresponds to adding T at the
end of the sequence. Furthermore, we can permute sequences. Given a per-
mutation ~r : n -~ n, we can write ~r(£) for the sequence ( x ~ ( 0 ) , . . . , x~(n-z)},
and similarly for the modal side. Then the following rules are valid.
(exp.) (per.)
¢[~-xn] ~ ~. T ¢[~(~)] ~ ~(~)
In b o t h rules we used a double line separating the top row from the b o t t o m
row. This means that the rules can be applied t o p - t o - b o t t o m or b o t t o m - t o -
top, which in the case of (exp.) amounts to killing an unnecessary variable.
Next consider the operation of renaming the variables in ~ by a substitution
a. Obviously, if p~ is another variable and p ~-+ p~ is injective, t h a t is, no two
variables are identified, then this is just a harmless operation, with no bearing
on the property described by the sequence.
(ren.) ~ : p - v a r ~-~ p - v a r
Similarly, consider swapping ~p and p (denoted by p[~p ~ p]), or alterna-

tively, replacing p by ~p and killing double negation. Again, this does not
change the elementary property described.
(swap.)
Next consider the operation of replacing in ~(£) a variable, say x ~ - l , by

another, say x~-2. This amounts on the modal part to replacing ~ n - 2 by
~ n - 2 A ~ - 1 . Also consider iterating a condition on another variable.
(cnt.) ~(~,y,z)[2.y. z] ~ ~ . ~o~-2 " ~ - 1

¢ ( ~ , y , y ) [ ~ . y] ~ ~. ~_~ A ~_~
~(~,y)[~. y] ~ ~.~
(iter.)
(¢(~,y) A ~(~,z))[~. y. z] ~ ~.~.~
Suppose, namely, that ( [ ~ . y • z] ~ ~ . ~o~-2 • ~o~-1. Then for all n - t u p l e s
c f of worlds in ~ ~ X, we have ¢[~] i~ for some valuation ~, ~ ~ ~ ( ~ )
for all i < n. Now choose an n - 1-tuple ~ C f . By assumption, ~ [ ~ . w~-2]
iff for some valuation wi ~ ~(~0~) for all i < n - 2 and w~-2 ~ ~(~,~-2) as well
as w~-2 • ~ ( ~ - ~ ) , so that w~-2 • ~(~0~_~ A ~0~_~), and conversely. Thus
the rule is correct. Notice that we must reduce the number of arguments here.
We cannot conclude, for example, that ( A x e - 2 -- x n - z is describable, for this
would require simultaneous fixing of x~-2 and x,~_~ to the same value. The
rule (iter.) is likewise straightforward. For the statement of the following rule
let ~1 and ~2 be two sequences of length n. Then ~ A ~ := ( ~ 1 A ~ 2 : i < Tt).
(A-I.) (1[/] ~.~ 9j1 ¢2[~] ~., 152 if var(¢ 1) • var(152) = ;~

(¢1 A ¢z)[~] ~ 151 A 152
For a proof assume ((1 A (2)[~7]. By assumption we have a valuation/31 such

that w, c ( ~ ) for all i < n, and a valuation/35 such that w, C ~ 2 ( ~ ) for
all i < n. Define/3 as follows. 9(P) := /3'(P) if p • vat(15), /3(p) .'= ~2(p)
otherwise. This is well defined by the assumption that t51 and 152 are disjoint
in variables. Then w, • ~ ( ~ ) as well as wi • ~(~2) for all i < n, and so
wi • ~ ( ~ A ~2), as required. Conversely, if w~ • ~ ( ~ A ~2) for all i < n then
wi • ~ ( ~ ) as well as w~ • / 3 ( ~ i ) for all i < n and so by assumption (I[~U] as
-- 2
well as (2[~U].
(v-I.) (~[~. y] ~ 15.x (2[~. y] ~.~ 15. ¢

((1 v (~)[~. y] --~ 15.x v ¢
((~, y ) [ ~ . y] ~ 15- x
(o-~.)
(3z >~ y)((~, z)[~. y] ~ 15.0~x
To see the correctness of the first rule, assume the premisses hold, and that
~1 V ~2[u~] for an n + 1-tuple v~. Then either (l[u~] or ~2[v~]. Assume without
loss of generality the first. Then there is a valuation/3 such that wi • ~ ( ~ )
a l l / < n and w,, • ~(X). Then also wn • ~(X V ¢). Conversely, assume
that wi • ~(Pi) for all i < n and w~ • ~(X V ¢). Then either wn • ~(X) or
wn • ~(¢). Assume without loss of generality the first. Then, by the left hand
premiss of the rule (1[~], whence ((1 V (2)[u7]. Next the rule ( 0 - I . ) . Assume
((3z>iy)()(~, z). Then there is a sequence v~.v such that (3z>iy)([~.v], that
is, there is a u such that v <~i u and ~[u~. u]. By assumption, we can find a val-
uation/3 such that wi • ~ ( ~ ) for all i and u • ~(X). Then v • ~((}iX), which
had to be shown. Conversely, assume wi • ~(~i) for all i and v • ~(QX).
Then there is a u such that v <~i u and u • ~(X). Thus, by the premiss of the
rule, ( ( ~ - u). Then, however, ((~z >i v)((g, z ) ) [ ~ , v], as required.
Now let Seq, consist of the axioms ( a x i o m . ) and of all the rules (exp.),
(per.), (ren.), (swap.), (cnt.), (iter.), (A-I.), (V-I.) and ( 0 - I . ) . We call S e q
the b a s e calculus. Let ¢ be a calculus of internal descriptions, consisting
of axioms and rules. A statement ' ( ~ 15' is d e r i v a b l e in ~ if it can be
produced from the axioms with the help of the rules in finitely many steps.
We say that 15 is d e r i v a b l e in ¢ if there exists a ( • L ~ and a sequence
such that '([~] ~-~ 15' is derivable in Seq; and that ([~7] is d e r i v a b l e in ~ if
there exists a sequence 15 such that '([~] ~ 15' is derivable in ~. A calculus ¢
of correspondence statements is s o u n d for a class ~ if '~[~] ~ 15' is derivable
in ~ only if 15 internally describes ([~] in ~. ¢ is called c o m p l e t e for ~ if
whenever 15 internally describes ([~] in :~, '([~] ~-~ 15' is derivable in ~.
THEOREM 5.4.5. Seq is sound for all classes of frames.

An important consequence is the following.
THEOREM 5.4.6. Let 2~ be any class of frames. If ~(xo) is obtained from
formulae internally describable in 2~ with the help of conjunction, disjunction
or restricted existential quantification, then ~(xo) is internally describable in
PROOF. From Lemma 5.4.7 we conclude that the set of internally describ-
able ~(~) is closed under A. It is clearly also closed under restricted 3. Now
let ~(x0) be composed from internally describable formulae with conjunction,
disjunction and restricted existential quantification. Then, by some straight-
forward manipulations, ~(x0) is equivalent in predicate logic to a disjunction
of formulae ~(x0), i < n, each of which is made from describable formulae
using only A and restricted 3. Then, for all i < n, Ui(x0) is describable in :~,
and by Lemma 5.4.8, ~(x0) is describable in ~. []
LEMMA 5.4.7. Let 2~ be a class, and let ~(~) and U(~) be describable in
~. Then (~ A ~1)(~) is describable in ~.
PROOF. By assumption, there exists sequence ~ such that ~(~) ~
and a sequence ~ such that ~(~) ~-~x ~. Now let )~ result from renaming
variables of ¢ in such a way that they become disjoint to the variables of ~.
Then, by (ran.), ~/(~) ~ ¢. Finally, by (A-I.), ~ A ~ describes (~ A U)(~)
in :~. []
LEMMA 5.4.8. Let 2~ be a class, and let ~(xo) and ~?(xo) be describable in
2~. Then (~ V U)(x0) is describable in 2~.
The proof of this theorem is similar. Notice that in proving the correctness
of the rules we have always shown how to construct a valuation for a given
sequence of worlds. Now we consider two rules.
(~-I.) x0 # xl ~-~ p" -~p (~-I.) Xo ¢i xl ~'~ [~ip'-~p
THEOREM 5.4.9. Seq + (~=-I.) is sound f o r ~ f , the class of differentiated

frames.
The proof is an exercise.
THEOREM 5.4.10. Seq + ( ~ - I . ) is sound forTi, the class of tight frames.
Again the proof is an exercise. We now give some examples of the calculus.
EXAMPLE 1. K . T is ti-persistent. For a proof consider the following
derivation.
Xo ~ xl ~'~ Dp. ~p
Xo ¢ xo ~ [Zp A ~p
The first line is true in ~i. Hence []p -+ p locally defines x0 <~ x0 in the class
of tight frames.
EXAMPLE 2. K . t is ti-persistent.
xo 40 xl ~ [~oP'-'P
(~u ~1 xo)(u 4o xl) ~ <>l[]op. ~p
(3u ~1 xo)(u 4o xo) ~ Q D o p A ~p
Xo 41 xl ~ E]lp'-,p
(~U [:>0 X0)(U 41 Xl) ~ (~0[-]lP " ~P
(3u ~o xo)(u 41 ~o) ~ ~o[]~p A ~p
These two derivations show that in the class of tight bimodal frames the
formulae (}o[Z]tp --+ p and (}1[Slop --+ p locally correspond to (Vut>l xo)(u'%xo)
and (~u [2>0 xo)(u <:]1 xo), respectively.
EXAMPLE 3. K . a l t l is dr-persistent. The following derivation is a proof
of this fact.
xo~xl ~p'--p
(~xo ~ u)(u ¢ ~1) ~ ~p.~p
(~xo ~ ~)(3u ~ xo)(u ¢ ~) ~ Op A O~p

Notice that the axiom (}p A (}q. --+ .(}(p A q) is not derivable with the help of
the calculus for differentiated frames.
THEOaEM 5.4.11. Let ~(xo) C 52f be universal, restricted and positive.
Then ~(xo) is internally definable in 9l, the class of refined frames.
PROOF. ~ ( x 0 ) is negative, existential and restricted. Thus, it is com-
posed from formulae of the form xi ~ xj, xi 4k xj with the help of con-
junction, disjunction and restricted existential quantification. Whence it is
derivable in S e q + ( ¢ - I . ) + (4 - I . ) . []
Let us end this section by considering the question of the completeness of

the basic calculus. We will show in Section 9.5 that for a modal formula T it
is undecidable whether it axiomatizes the inconsistent logic. Hence the set of
sequences 'T ~-~ f' is undecidable. The idea behind setting up modal equiv-
alence not as a m a t t e r of theorems of second-order logic but as a deductive
calculus is that while the former is undecidable, the latter m a y be decidable,
however at the price of being incomplete. On the other hand, we will see
later (Theorem 5.8.6) that S e q is complete in the following weaker sense. If
'~ ~ ~(x)' holds in ® then there exist ¢ and 77 such t h a t K~ ® ¢ = K~ ®
and (Vx)~(x) - (Vx)r/(x) (in predicate logic) and ' ¢ ~ rj(x)' is derivable in
Seq. It is not clear whether this generalizes to arbitrary sequences, but that
seems to be the case. So, rather than generating all facts we generate at least
a representative class of them. We could in principle add rules that would

make the calculus complete (by closing under equivalence), but that would
make it undecidable and useless for practical purposes.
E x e r c i s e 177. Show that (iter.) is sound for all classes.
E x e r c i s e 178. Show that inequality is internally describable in :~ iff :~ is a

class of differentiated frames. This proves Theorem 5.4.9.
Exercise 179. Show that j-inaccessibility (i. e. :~j) is internally describable

in :~ for all j < n iff :~ is a class of tight frames. This proves Theorem 5.4.10.
E x e r c i s e 180. Show that the set of 3:-elementary logics is closed under finite
joins and finite meets in the lattice of all modal logics.
* E x e r c i s e 181. Show that [3(}[~E]p -+ (}(}[](}p is globally J~rp-elementary

and corresponds to (Vx)(3y)(x <~y). However, this axiom is not locally elemen-
tary. (The first half is not so difficult, only the failure of local elementarity.
For those who want to see a proof, it can be found in VAN BENTHEM [10],
page 82.)
5.5. Sahlqvist's T h e o r e m
Two classes of modal formulae will play a fundamental role, monotone
and A-distributive formulae. A formula ~(p, q~) is m o n o t o n e in p if F- ~(p A
r, q0 -+ ~(P, q~, where k- ~ abbreviates ~ E K~, and A - d i s t r i b u t i v e in p if
F- ~(pAr, q~. ++ .~(p, q~)A~(r, q~. ~ is called m o n o t o n e ( A - d i s t r i b u t i v e ) if it
is monotone (A-distributive) in all occurring variables. The notions a n t i t o n e
and V - d i s t r i b u t i v e are dual notions. T h a t is to say, ~ is a n t i t o n e (V-
distributive) in p if[ ~ is monotone (A-distributive) in p. All of these notions
can be characterized syntactically. We will give sufficient criteria here. Call
a formula ~(p, q0 p o s i t i v e in p if all occurrences of p are in the scope of an
even number of negations. Call a formula s t r o n g l y p o s i t i v e in p if p does
not occur in the scope of ~ (or any (}j for j < a if (}j is a primitive symbol).
is p o s i t i v e ( s t r o n g l y p o s i t i v e ) if it is positive (strongly positive) in all
occurring variables. A formula is n e g a t i v e ( s t r o n g l y n e g a t i v e ) if it can
be obtained from a positive (strongly positive) formula by replacing each
occurrence of a variable p by -~p. Notice that we can characterize a positive
formula also as follows.
PROPOSITION 5.5.1. 99(p,q~ is positive in p iff there exists a formula ¢

which is built from the letter p and formulae not containing p with the help of
A, V, [~, Q , i < ~, such that ~(p) ++ ¢ C K,~.
5.5. Sahlqvist's Theorem 247
The proof is an exercise. Notice that the occurring constant subformulae

can be arbitrarily complex. The formula fi]0(((}l(T A 9[]0±) V p) A 00P) is
positive. Likewise, []0(01T V [~0P) is strongly positive.
PROPOSITION 5.5.2. The following holds.
1. If W(p, ~ is positive in p it is monotone in p.
2. If ~(p, q-) is strongly positive in p it is A-distributive in p.
PROOF. (1.) By induction on ~(p, q]. By Proposition 5.5.1 we can assume
that p is built from the letter p and formulae not containing p with the help
of A, V, [~j and (}j, j < ~. The formula p is monotone in p; likewise a formula
not containing p is obviously monotone in p. Suppose ~ = ¢1 A ~b2. By
induction hypothesis, ¢1 (P A r, q~ F- ¢1 (P, ~ and ¢2 (P A r, q~ e ¢2 (p, q~. Hence
~ ( p A r , q~ e ¢l(p,q~ A ¢ 2 ( p , ~ ( = ~(p,q~). Similarly for ~ = ¢1 V¢2. Now
suppose that ~ = [~i¢1. Then by induction hypothesis ¢1(P A r, ~ e ¢1(p, ~ .
From this we get []i¢(pAr, @ e [~¢(p, q~, as required. Similarly for ~ = 0~¢1.
(2.) Again by induction. A formula not containing p is A-distributive in p
by the fact that e ¢ +~ ¢ A ¢ . Now let ~ = ¢1 A¢2, ¢1 and ¢2 strongly
positive. By induction hypothesis both are A-distributive in p. Then we have
e ¢~(p A r , ~ ~ ¢~(p,~ A ¢ i ( r , ~ and so
~(pAr,~ He ¢ l ( p A r , ~ A ¢ 2 ( p A r , ~
He ¢1 (P, q~ A ¢1 (r, q~ A ~b2(p, q~ A ¢2 (r, q~
Similarly for ~ --- Vl~¢. []

A sequence q5 of formulae is called a s p o n e if each of its members is either
strongly positive in all variables or negative in all variables. (The name spone
is an acronym from strongly positive and negative.) We will usually write a
spone in the form 77 • #, where 77 is the subsequence of the strongly positive
formulae and # the subsequence of the negative formulae. We will show that
in the class ~ p U ~ all spones are elementary, and this will prove Sahlqvist's
Theorem. In the basic calculus Seq not all spones can be derived. However,
in the class ~ p U ~ another rule is valid. The key to this rule is the following
lemma. To state this lemma properly recall from Section 2.9 the concept of
an upward directed family of sets. Let a frame ~ be given. Consider a set I of
indices which is partially ordered by ~, and for each pair i, j there is a k such
that i ~ k and j < k. A family over (I, _<) is simply a collection of sets ci,
i C I (or a function from I into F). This family is u p w a r d d i r e c t e d if i _< j
implies c~ C_ cj. Given such a family ~) = (di : i 6 I) we write [Zj:D for the
family ([]jdi : i E I). An upward directed family has a limit, lim~), which
is just the union Ui~i di. The notion of a downward directed family is dual,
that is, we require i _< j ~ di _~ dj instead. The following theorem is due to
LEO ESAKIA in [58].
LEMMA 5.5.3 (Esakia). Let ~ be a tight and compact frame and ~D = (di :
i E I) be an upward directed family of sets in F. Then
li lim© = limli©.
PROOF. ~D is upward directed, and so - © -- ( - d i : i C I ) is downward
directed. It will be sufficient to show that for a downward directed family ~,
l i l i m - ~ = l i m L - ~. This is the same as - l i l i m - ~ = - l i m l i - ~ since
lim commutes with - . This is finally equivalent to
0i lim ~ = l i m 0~8.
(For an upward directed family this is clear, but now g is downward directed.)
(C_.) Let x C 0~lim~. Then there is a y C lim~ = N ~ such t h a t x < ~ i y .
Thus for a l l j C I, x C Oidj and s o x C l i m 0 i ~ . (D.) S u p p o s e x ¢ 0 j l i m ~ .
Pick a y E lim $. Then x ~ j y. By tightness of ~ there is an internal set ay
such t h a t y E ay, but x E IIj - a y . Since the union of the ay contains lim ~,
there is a finite subset Y C_ lira ~ such that lim ~ C_ [.Jy~y ay. (This follows
from the compactness of the frame.) Let b := UyeY ay. Then lira ~ C b and
x E " j - b. Moreover, by compactness of ~ again, there is an e C ~ such t h a t
e C_ b. Then ~je C 0jb. Since x ¢ 0jb, x ¢ 0je. Therefore x ¢ l i m 0 j g . []
THEOREM 5.5.4. The rule (D-I.) is sound for J~rp U ~ .
([B-I.) ~[~" y] ~'~ fi" # (fi a spone, # negative)
PROOF. Assume that ~[2. y] ~-~ ft. # in the class ~qrp t2 ~ . Let ~ be of
length n. Take a frame ~ from J~rpU~ and a valuation/3 such t h a t wi E ~(pi)
for all i < n and v E ~([~j#). Then for all u such that v<~ju we have u ~ ~(#),
and so by assumption ~[~-u]. Hence ((Vy t>j v)~(~, y ) ) [ ~ , v], as required. For
the converse direction choose points w 0 , . . . , w~_~,v such that for all u with
v <~ u we have ~[u~-u]. Let A := {u: v <~j u}. By assumption, for each u ~ A
there is a valuation Z~ such that wi e ~ ( P i ) and u e ~ ( # ) .
C a s e 1. ~ is a Kripke-frame. Then put ¢?(p) := A~eA ~ ( P ) - Then wi E ~(Pl)
for all i. For either fli is negative (and the claim easily follows), or Pi is strongly
positive, in which case -~(Pi) = N ~ A ~-~(pl) and so w~ ~ ? ( ~ i ) . Furthermore,
u ~ ? ( > ) for all u ~ A, since # is negative. Thus v ~ ? ( [ ] j > ) , and so
everything is shown.
C a s e 2. ~ is a descriptive frame. Let I be the set of finite subsets of A. For
S E I let ~s(P) := N ~ e s ~ ( P ) - It is not hard to verify that wi ~ ~s(Pi) for all
i < n. Our aim is to find an S ~ I such t h a t v ~ ~ s ( [ ] j # ) . W i t h respect to C_,
I is ordered and (~s(P) : S ~ I) is a downward directed family of sets. Now,
u E ~ s ( # ) i f u ~ S. Hence A C_ l i m ~ s ( # ) and so v ~ I l i m ~ s ( # ) . By Esakia's
L e m m a v ~ I j l i m ~ s ( # ) = l i m l j ~ s ( # ) = l i m ~ s ( ~ j # ) . By compactness of
there is an S ~ I such that v E ~ s ( ~ j # ) ; this had to be shown. []
We define a new calculus S e q +, which is S e q enriched by (#-I.) and

( [ ] - I . ) and a new rule for conjunction, (A-I2.). To define it, let ¢1 and ¢2 be
two n-long sequences. Recall that ¢1 A ¢2 := ( ~ A W2: i < n).
(A-I2.) ~l[X] ~ 7~1" /~1 ¢2[X] ~ K2" Z72 for z~i" gi spones
A (el A e2)" (al A 72)
It is left as an exercise to show the soundness of this rule. Moreover, this rule
is sound in all classes, so it can actually be added to Seq. It is only for ease
of exposition that we have chosen to ignore this rule previously.
THEOREM 5.5.5 (Sahlqvist). Let X be a modal formula of the form
where [] is a compound modality. Suppose that

1. ¢ is positive.
2. ~ is composed from strongly positive formulae using only A, V and Oj,
j<a.
Then K~ • X is locally d-persistent and locally elementary in ~:p U ~ .
PROOF. It is enough if we can show that there is a ~ such t h a t ( ~-~ -~X-
Now -~X = -~[]-~(~A-~¢). By repeated use of ( 0 - I . ) and (V-I.) backwards this
can be reduced to showing that ~ A 7 ¢ is derivable. By ( c n t . ) it is enough
that ~ . - 7 ¢ is derivable. 9 ¢ is negative, and ~ is composed from strongly
positive formulae using A, V and 0j. By appealing to the rules (cnt.), (V-I.),
( 0 - I . ) this can be reduced to the problem of showing t h a t sequences of the
form z7 • v are derivable, where r7 is a sequence of strongly positive formulae
and v negative. The theorem below establishes this. []
LEMMA 5.5.6. All spones are derivable in Seq +.

PROOF. Let us concentrate on the modal formulae in S e q +. We proceed
in several steps.
S t e p 1. All spones zT. J are derivable where for some p, 7ri = p and ~j = -~p,
T or A_. Simply start with p - ~p and use (iter.) and (per.).
S t e p 2. All spones are derivable in which the ~i are constant or a negated
variable. Namely, by (swap.) it is enough to show that if all 7ri are variables
or constants and ~j are strongly negative, the spone is derivable. So there are
three possibilities, (i) ~j is constant, (ii) ~j = #1 V #2, (iii) ~j = 0 j # . We deal
with these cases in turn, assuming without loss of generality that j = n - 1,
the last entry of the list. So we have the spone ~ . ~ - 1 . C a s e (i). If ff is
derivable, so is ff-T. Moreover, Vn-1 is derivable, and by (exp.) also - ~ ' v ~ - l ,
and so by (A-I.) P ' ' ~ - I . C a s e (ii). If ~ ' # 1 and ft. #2 are derivable, so is
O" #1 V #2, by (V-I.). C a s e (iii). If g . # is derivable, so is ~- 0 j # , by ( 0 - I . ) .
S t e p 3. All spones are derivable. We may start from spones in which the
strongly positive part can be complex. The reduction is similar to t h a t in

Step 2, with two more cases to be considered, namely (iv) u = #1 A #2, (v)
u = [ ~ j # . Case (iv) is dealt with by using the rule (A-I2.), and Case (v) is
dealt with by using ([~-I.). []
To master the syntactic description of the theorem requires some rou-

tine. We note that for example the Geach formula (}~p --+ ~(}p satisfies the
conditions of the theorem while the McKinsey formula [3(}p --+ (}Dp does not.
EXAMPLE 1. The Geach formula is elementary in ~(tp U ~ .
Xo :~ xl ~ p" ~p
( w > =o)(~ # =1) ~ ::p- ~p
(w > ~l)(Vu > =o)(~ # ~) ~ rap. D~p
(3u' > xo)(W > = , ) ( w > ~')(~ # ~) ~ O::p. ~ p
( ~ ' > ~)(3u' > =o)(V~ > v')(vu > u')(u ¢ ~) ~ 0pp. $ ~ p
(3v' D xo)(3u' [:>xo)(Vv D v')(Vu b> u')(u ~£ v) e,,,,, O[]p A O[Z]~p
EXAMPLE 2. The axiom 0p A 0q --+ 0 ( p A q) defines a dr-persistent logic.
Nevertheless, we have seen that this cannot be shown inside the calculus for
differentiated frames. However, in the extended calculus it is derivable. T h e
derivation is somewhat contrived. First, from xo ~ x2 ~ p " T • --p and
t~T-T.qwegetxo ~x2 ~p.q.~p. Likewise, x1 ~ x 2 ~ p - q ' ~ q i s
derived. From this we get xo ~ x2 V xl ~ x2 ~ p - q • -~p V --q. Now we get
(vy > x~)(~0 # y v ~1 # y) ~ p. q- [](~p v ~q)

(3u > xo)(~v > xl)(vy > x2)(u # y v v # y) ~-~ op. Oq. :~(~p v ~q)
(Eu ~>xo)(~v E>xo)(Vy D xo)(v # y V u # y) e,,,,, Op A Oq A -,O(p A q)
The formula (qu ~> Xo)(~v [:>xo)(Vy D Xo)(u :~ y V v ~£ y) is equivalent in
predicate logic to (~u E>xo)(~v ~> xo)(u -:/: v). Hence 0p A 0q --+ 0 ( p A q)
corresponds to (Vu E> x0)(Vv C>xo)(u - v).
A formula satisfying the conditions of the theorem will be called a ( m o d a l )
S a h l q v i s t f o r m u l a . A logic axiomatizable by Sahlqvist formulae is called a
S a h l q v i s t logic. We note that formulae of the form qa --+ ¢, where b o t h
and ¢ are positive and free of ff], are all Sahlqvist formulae. There are some
stronger versions of this theorem. For example the following characterization
due to JOHAN VAN BENTHEM [10], which uses the notions of positive and
negative occurrences of a variable. These are defined as follows. Let ~a be
a formula with n occurrences of p. For ease of reference, we consider the
different occurrences as being numbered from 0 to n - 1. Replace each occur-
rence by a distinct variable Pl, and call the resulting formula W~. p o c c u r s
p o s i t i v e l y a t j in ~ if ~ is positive in pj. p o c c u r s n e g a t i v e l y a t j if
it does not occur positively at j. For example, in ~ -- p A (}~p, then the
first occurrence of p is positive and the second occurrence negative. We have

~x = Pl A (}~P2, which is positive in Pl and negative in P2.
DEFINITION 5.5.7. Call a formula to a S a h l q v i s t - v a n B e n t h e r n f o r -

m u l a if for all occurring variables p either (i) no positive occurrence of p in
is in a subformula of the form ~b A X or [~j¢ if that subformula is in the
scope of a (}k or (ii) no negative occurrence of p in ~ is in a subformula of
the form ¢ A X or Vlj¢ if that subformula is in the scope of some ~k.
Not every Sahlqvist formula is a Sahlqvist-van Bethem formula. For

example the formula (}(pA[3(}~p) --4 ((}[~pV[3E]~p) is Sahlqvist-van Benthem,
but not Sahlqvist. It was shown in [10] that logics axiomatized by a Sahlqvist-
van Benthem formula are canonical. In the next section we will establish
that the class of logics axiomatizable by Sahlqvist-van Benthem formulae
is actually not larger than the class of Sahlqvist logics, so that this result
actually immediately follows.
Here we will show a theorem to the effect that the class of Sahlqvist logics
can be axiomatized by simpler axioms than originally described by Sahlqvist.
Namely, one can dispense with the operator prefix. This is quite suitable for
certain applications. We should perhaps note that our definition of a Sahlqvist
formula is not exactly Sahlqvist's own. In fact, he allows only an operator
prefix of the form D k. (His proof is only for monomodal logics, but is easily
extended to polymodal logics.) We have allowed ourselves to define a slightly
larger class, which is also somewhat easier to define in a polymodal setting.
In view of the next theorem, these differences are rather marginal. The class
of Sahlqvist logics remains the same, no matter what definition we choose.
THEOREM 5.5.8. Let []s be a compound modality. Let X = Ds(~ -+ ¢) be

a Sahlqvist formula and q • var(x). Put X ° :-- (}S(q A ~) -+ DS(q V ¢). X ° is
Sahlqvist and K~ • X = K~ • X ° •
PROOF. It is clear that X° is Sahlqvist. Therefore, let us show the second

claim. ~ -+ ¢ is Sahlqvist. Let ~ . ~ ¢ describe ( ( x , y ) [ x . y]. Then ~ A ~ ¢
describes ((x, x), by (cat.), and (}~(p A ~¢) describes (By E>s x)((y, y). Hence
the elementary condition of X is (Vy E>s x ) ~ ( ( y , y ) . Now, q . ~q describes
(x ~ y)[x. y]. Hence, by (A-I.), q A p . ~ q A ~ ¢ describes the formula (((x, y)A
x ~ y)[x. y]. Therefore, 0~(q A ~ ) . (}~(-~q A -~¢) describes
y') .y].
(}~(q A p) A (}~(~q A -7¢) describes (~z' C>~ x)(3y' ~>x)(~(x', y') A x' =~ y'), by
(cat.). Hence X°, which is the negation, defines
' - y' y'))
This is the same as (Vx' ~>~ x ) ( ~ ( x ' , x')). []

252 5. Definability and Correspondence
E x e r c i s e 183. Name formulae which are monotone but not positive, and
formulae which are A-distributive without being strongly positive.
E x e r c i s e 184. Show the soundness of the rule (A-I2.).
E x e r c i s e 185. Suppose ~ contains only positive or only negative occurrences

of p. Show that either K~ e ~ = K~ ® ~[-t-/p] or K~ ® ~ = K~ • ~[_l_/p].
Thus, to be essential a variable must occur at least once positively and once
negatively.
E x e r c i s e 186. Show that if ~ is Sahlqvist there exists a Sahlvist formula

¢ such that K~ O ~ = K~ ® ¢, and every variable of ¢ occurs exactly once
positively and once negatively.
E x e r c i s e 187. Generally, modal algebras are not closed under infinitary in-
tersections and unions. This can be remedied as follows. Given a modal alge-
bra 92, the completion of 92 is defined by ~m92 := ((92+)~)+. (See Section 4.8 0
It consists of all sets which can be generated as arbitrary intersections and
unions of sets in 92. Show that if ~ is a Sahlqvist-formula then 92 ~ ~ implies
5.6. Elementary Sahlqvist Conditions

In this section we will characterize those elementary conditions which are
determined by axioms of the form considered in Sahlqvist's Theorem. It is
clear that all derivable sequents ' ( ~ ~' state a local correspondence and
( E :R/. We will for simplicity always assume that the situation never arises
that a world-variable v occurs both free and bound in a subformuta. Such
a formula is called clean. Not clean is x <~1 Y A (3y [>2 x)(y <1 x). Every
unclean formula is equivalent to a clean formula in the predicate calculus. In
a clean formula ( a variable y is inherently existential if (i) all occurrences
of y are free or (ii) ( has a subformula ~? = (By ~>k x)~ which is not in the
scope of a universal quantifier. Likewise y is inherently universal if either
all occurrences of y are free in ~ or ~ contains a subformula U = (Vy ~>k x)~
which is not in the scope of an existential quantifier.
Now recall the notation x <F y for sets s of sequences of indices < ~.
It means that x can see y through one of the sequences in s. The formula
x - y can be represented by x <~{~}y. We will change our language for frames.
Recall from Section 5.1 that the language S f is obtained from [Rf by adding
5.6. Elementary Sahlqvist Conditions 253
x <]s y as primitive expressions. This is only a technical move to simplify the

terminology somewhat. The theorem we will prove is the following.
THEOREM 5.6.1. Suppose ~(x) is a positive Zf-formula such that every
subformula x ~s y contains at least one inherently universal variable. Then
there exists a Sahlqvist formula ~ which locally corresponds to ~(x) in ~qrpU~.
Conversely, any Sahlqvist formula corresponds in Nrp U ~ to an elementary
formula of this kind.
A first-order g / - f o r m u l a in £,f satisfying the conditions of the theorem
will henceforth be called an e l e m e n t a r y or f i r s t - o r d e r S a h l q v i s t f o r m u l a .
Generalizing this somewhat, a formula 4(2) of gl is called S a h l q v i s t if it is
positive and every atomic subformula contains at least one inherently universal
variable. If ~(2) is Sahlqvist, ~ ( 2 ) will be called n e g a t i v e S a h l q v i s t . On
the way to prove Theorem 5.6.1 we will derive some useful facts.
LEMMA 5.6.2. Let ~(~,y) be a ~ f - f o r m u l a such that every a t o m i c sub-
formula contains at most one variable (t 2. Then there exists a clean r/(2, y)
such that (Vy)(V2)(~(2, y) ==_~( 2, y) ) in predicate logic and e v e r y subformula
of 77 has at most one free variable outside of 2.
PROOF. By induction on ~. We can assume that negation occurs only in
front of the atomic formulae. Moreover, we can assume that ~ is clean and
that a subformula r/is in the scope of a quantifier only if the quantifier binds
a free variable of r/. The claim holds by virtue of the assumptions for positive
and negative atomic subformulae. Now let ~(2, y) = r/l(2, y) A ~2(2, y).
By hypothesis, every atomic subformula of ~ has at most one free variable
¢ ~. This holds also of the r/i. By induction hypothesis there exist J1 (2, y)
and 52 (2, y) such that (Vy)(V~)(r/1 (2, y) _= J1 (2, y)) and (Vy)(V2) (r]2(2, y) _=
52(~,y)) in predicate logic and every subformula of 51 and 52 contains at
most one free variable ¢ ~. Then put O(~,y) := 51(~,y) A 52(~,y). This
satisfies the claim. 0 is clean if 5i and 52 are. For a subformula 0' of 0
either 0' = 0 or 0' is a subformula of 51 or J2. In the first case only y
is a free variable ¢ 2. In the second case we know that every subformula
of 0 / contains at most one free variable ~ ~. Similarly the case 4(2, y) =
r/l(~, y) V~/2(~, y) is treated. Next assume ~(2, y) = (Vw ~>k v)O(~, w, y). Then
v C 2 or v = y. Assume 0(~, w, y) = 51 (~, w, y) A 52 (~, w, y). Then distribute
the quantifier over the conjunction. The formula (Vw c>k v)(Jl(2, w , y ) ) A
(Vw Dk v)(J2(~,w,y)) is equivalent to ~ in predicate logic; it is clean and
every atomic subformula contains at most one variable ~ 2. Now back to the
case of ~ = r]l A r~2. Assume #(2, w , y ) = 81(2, w,y) V 82(~,w,y). We may
assume by induction hypothesis that every subformula of 81 and 82 contains
at most one free variable outside of i ; this variable is either w or y. Hence
several cases may arise.
C a s e 1. w ft fvar(51), w f~ fret(J2), where fvar(~) denotes the set of variables
occurring free in 4. Then put 77 := 01 (~, y)V02(~, y); 77fulfills the requirements.
C a s e 2. w C fvar(01) and w E fvar(02). Then consider an atomic subformula
containing the variable y. By assumption on { it must be of the form xi <1s y
or y <~s xi for some i and s. By assumption on 4 this cannot happen.
C a s e 3. w E fvar(01) and w • fvar(02). Then
y) := (vw w). v
Then since y --- v or y E ~, 77 fulfills the claim.
C a s e 4. w ¢ fvar(01) and w E fvar(02). Then
y) := y). v .(vw
This fufills the requirements. Proceed dually in the case of an existential
quantifier. []
The previous theorem shows that if the atomic subformulae are of the
form x <~ y with x C fvar(~), then ~ can be written in an essentially ' m o d a l '
way. If we do not require ( to be clean we can actually arrange t h a t ~ contains
very few variables by reusing bound variables every time they are no longer
needed. Two more variables than occur free in ~ are therefore needed; in
particular, for ~ = (Vx)77(x) with no free variables, then if ~ is of this form, it
has exactly three variables. This follows from the next theorem. It has been
shown by D o v GABBAY ([72]) (see the exercises).
PROPOSITION 5.6.3 (Gabbay). Let ~(:2) E :~f be Sahlqvist, ~ of length n.
T h e n ~ - 77 for an U which contains at most n + 2 variables.
LEMMA 5.6.4. Every ~(~,y) C S I which is negative and in which every

subformula has exactly one free variable outside of ~ is derivable in S e q + and
corresponds to a spone 4 . ~.
PROOF. By induction on ~(~, y), using the rules (A-I2.), (V-I.), ( 0 - I . )

and (K]-I.). The starting sequences are x0 18~ Y0 *'~ [2~P " ~P, which in t u r n
are derivable in S e q +. []
The proof of Theorem 5.6.1 is now quite short. Suppose t h a t ((x0) is a

negative Sahlqvist formula. According to T h e o r e m 5.4.6 it is enough to prove
the claim for negative Sahlqvist formulae of the form (Vy' t>k x~)7/(~, y'). The
latter formula results from ~ := (Vy' E>k Y0)~(£, Y') by applying (cnt.). Hence
it is enough to show the claim for the latter formula. By L e m m a 5.6.2, we
can assume t h a t every subformula of ~ contains at most one variable ~ £.
Those subformulae with free variables completely in ~ can be moved outside
the scope of any quantifier using laws of predicate logic. So the problem is
reduced to the case where every subformula of ~ contains exactly one extra
free variable. By L e m m a 5.6.4, those are derivable in S e q +. The proof of
T h e o r e m 5.6.1 is now complete.
Let us discuss the theorem to get a better insight into the classes of
formulae that are at issue here. Clearly, a somewhat more satisfying result
would be one in which we had only the restriction that ~ C 9~f and that
is positive. Better than that we can never do. (See the next sections.) The
really hairy part is the conditions on variables. Moreover, two questions arise.
(1.) Why have we chosen ~ / r a t h e r than L / to begin with? (2.) Why use S f
rather than :~f in S e q + ? The answer to both questions is: this is a matter
of utility. For example, it is hard to state what we mean by p o s i t i v e in
~ / inside L f without actually talking about restricted quantifiers in some
hidden form. The second question is somewhat harder to answer, but it
will become clear by discussing some cases. Suppose first that ~ C :~f and
that ~ contains no existential quantifiers; then the restriction on variables is
vacuous. By Theorem 5.4.11 we know already that all elementary conditions
are derivable. Next consider the formulae of the form V3 which are not V.
These formulae have existential quantifiers inside universal quantifiers, but
not conversely. In this case the condition says that in a subformula x <s y
not both x and y may be existentially bound variables. If we replace the
clause x <~s y by (~z ~>~ x)(z - y), then both z and y may be existentially
bound so the condition on the variables cannot be stated in the same way.
Finally, consider the case where 4 is of the form V3V. Here, if we rewrite the
clauses x ~s y the quantifier alternations increase; the resulting formula is of
the form V3V~. Moreover, the newly introduced existentials are innermost,
that is, closest to the atomic subformulae. Let us now assume that we have
eliminated the expressions x <]~ y. Then both x and y may be existentially
bound. But there is a difference between variables that are introduced from
rewriting the complex accessibility clauses and the original variables. The
former must be bound by an innermost existential which in turn must have
a restrictor which is inherently universal. No such restriction applies to the
other variables. Although the restriction can be restated in this way, it is
clear that this characterization is much less straightforward.
As an application, the following result will be proved.
THEOREM 5.6.5. Let ~ be a Sahlqvist-van Benthem formula. Then the
logic K~ ® ~ is locally d-persistent and locally elementary in ~rp 0 ~ . More-
over, there exists a Sahlqvist formula ¢ such that K~ • ~ = K~ • ¢.
PROOF. We shall show that ~ p corresponds to an elementary negative
Sahlqvist-formula. First we rewrite ~ so that it contains only variables,
aegated variables, 03, []j, A and V. Recall the definition of a Sahlqvist-
van Benthem formula. The negation of such a formula satisfies the dual of
Lhat condition. This is the following condition: for every variable p, either
(i) no positive occurrence of p is a subformula of the form ¢ V X or 0 j ¢ if
bhat subformula is in the scope of a Ilk or (ii) no positive occurrence of p is a
subformula of the form ¢ VX or 0 j ¢ if that subformula is in the scope of a Elk.
By substituting ~p for p, we can arrange it that for every variable p only (i)
obtains. Call such a formula g o o d . A formula is good if every subformula [~k)/
is either negative or strongly positive. Notice that the set of good formulae is
closed under subformulae. We will now show by induction on the constitution
of the formulae that each sequence ;~ = X0 • X1 " -.- " Xn-1 of good formulae
corresponds to a negative Sahlqvist formula. To start, assume every X~ is
either positive or negative. In that case, because the X~ are good, the positive
)/i are actually strongly positive. Hence ;~ is a spone, and it corresponds to
a negative Sahlqvist-formula. If this does not obtain, )~ contains a formula,
say X0, which is neither positive nor negative. Then one of the following
cases obtains. C a s e 1. X0 = (}iT. Then T - X 1 . . . . . X , _ I corresponds
by induction hypothesis to an elementary negative Sahlqvist formula ~(~).
Then ;~ corresponds to ((~y I>j Xo){[y/zo])[~], which is negative, elementary
Sah]qvist. C a s e 2. X0 = ~-1 V ~-2. By induction hypothesis and closure of
elementary negative Sahlqvist formulae under disjunction. C a s e 3. Xo =
~-1 A ~-2. Then by induction hypothesis 7-1 • ~-2 • X1 " -.- " X ~ - I corresponds to
some ~(~), which is elementary negative Sahlqvist. Then ~ corresponds to
~[Xl/XO], which is also elementary negative SaMqvist. []
For future reference we will introduce the S a h l q v i s t H i e r a r c h y to mea-

sure the complexity of descriptions given by Sahlqvist formulae in :Rf. This
will be measured roughly by the number of quantifier alternations occurring
in the formula. This can be defined as follows. Let 7 be an occurrence of a
subformula of ~. Then by replacing t by xi -- xi and by suitably renaming
the occurrences of the variables in ~ we can achieve it that each subformula
occurs only once. Let ~t denote a formula resulting from ¢ by this operation.
~t need not be unique. Under this condition, the following is well-defined
(and does not depend on a particular choice of ~t).
sq_rank(~t, ~t) = 0
sq-rank( ~t, -~7) = sq-rank(¢t, )
sq-rank(~ t, 771 A 72) = sq-rank(~t,71)
sq-rank(~ t, 71 A 72) = sq-rank(~t, ~2)
{ sq-rank(~*,~7)
if sq-rank(~*, (3y I>j x)7) is odd
sq-rank(~ t, (3y ~>jx)r]) = sq-rank(~t,rl) + 1
if sq-rank( t, is even
{ sq-rank(~t,V)
if sq-ra k( *, (vy is even
sq-rank(~ t, (Vy ~>j x)7)
= sq-rank(~t,7) + 1
if sq-rank(~ t, (Vy ~>j x)7) is odd
Call a formula c o n s t a n t if all atomic subformulae are of the form f or xl - xi.

Finally,
sq-rank(~) := m a x { s q - r a n k ( ~ f , ~ ) : ~ E sf(~t),~? not constant}
Here, the m a x i m u m over an empty set is defined to be 0. This defines the r a n k
o f a n e l e m e n t a r y f o r m u l a . For example, a universal restricted formula
comes out with rank 0, while an existential formula has rank 1. This is desired
even though the rank counts quantifier alternations. Let us note, namely,
that a :Rf formula must contain at least one free variable since all quantifiers
are restricted and the free variables are assumed to be quantified universally
(though by an unrestricted quantifier). Thus, all formulae invariably start
with a universal quantifier, and this causes the asymmetry. Notice further
that the rank does not increase in case of a constant formula (even though
this has not been noted explicitly in the informal definition, but should be
clear from the first clause). This is another deviation from classical quantifier
complexity. It is essential for m a n y reasons that constant formulae have zero
complexity. If for some reason we want to count the constant formulae as
well, we speak of the p u r e r a n k . It is defined by
sq-rank(~) := m a x { s q - r a n k ( ~ t , ~ ) : 77 E sf(~t)}
We denote by $q,~ the class of Sahlqvist formulae of rank n and the logics
axiomatizable by such formulae by ®qn" So, 0-1- C Sq 0, but it has pure rank
1. Likewise, a l t l C Sq 0, .4 E $q0- Finally, it is helpful to distinguish the rank
we obtain as above from a rank in which x <~2 Y is not a primitive formula,
but equal to (~z E>j x ) ( z - y). This we call the s p e c i a l r a n k . Notice that
the special rank of ~ is equal to the Sahlqvist-rank of ¢ rank if the latter is
odd or if the atomic formulae are of the form x - y (i. e. not using <~j), and
: sq-rank(~) + 1 otherwise. This is so, because for the special rank we only
have to eliminate the formulae x <lj y, introducing an existential quantifier.
Notes on this section. Let F O k be the set of expressions of predicate logic
in which at most k distinct variables occur. It is known t h a t F O 3 is generally
undecidable. However, M. MORTIMER has shown in [158] t h a t F O 2 has the
finite model property and is therefore decidable. Given ~, the size of model
of a minimal model for p is exponential in the length of ~. Hence polymodal
K is decidable. This does not extend to polymodal logics in general. For
there exist finitely axiomatizable Sahlqvist-logics without the finite model
property; the first system of this kind is the one of DAVID MAKINSON in
[145]. MAKINSON only shows that his logic is complete but does not possess
the finite model property. His paper predates that of SAHLQVISTand does
therefore not discuss the fact that this logic is Sahlqvist. See also [106] for
a discussion. Examples of Sahlqvist logics without the finite model property
can be found in this book in Section 9.4. These logics are elementary, but
the corresponding first-order property is in F O 3. It cannot be in F O 2, by
MORTIMER'S result.
E x e r c i s e 188. (GABBAY[72].) Show Proposition 5.6.3.
E x e r c i s e 189. Show that ®qn is closed under arbitrary unions and finite
intersections.
E x e r c i s e 190. Show that a Sahlqvist logic can be axiomatized by formulae

of the form [~(~ -~ ¢) where b o t h T and ¢ are positive and each occurring
variable occurs exactly once in T and exactly once in ¢. Moreover, if t h a t
axiom corresponds to ~, the number of variables is at most the number of
atomic subformulae in ~.
E x e r c i s e 191. A naive approach to correspondence is to take a formula and

regard the variables as referring to worlds rather than sets of worlds. For
example, in p --+ (}p or p --+ VT(}p we can get the desired first-order corre-
spondent by thinking of p as denoting a single world. This is not in general
a correct approach (for example, it fails with the Geach Axiom). Show how-
ever, t h a t if ~ --+ ¢ is Sahlqvist, and T contains no box-modalities and each
variable in the antecedent at most once, then pretending p to denote single
worlds yields a correct first-order correspondent. Hint. Consider valuations
of the form 13(p) = {w} and show that they can detect a failure of the axiom.
E x e r c i s e 192. Show that the Sahlqvist-van Benthem formulae are Seq + -

derivable.
5.7. P r e s e r v a t i o n Classes
The next two sections will require some techniques from model theory,
though quite basic ones. In contrast to the previous sections, which charac-
terized those statements '~ ~"~x ~' which are valid, we will now derive facts
about the correspondence statements t h a t are not valid for any ~. We will
prove in this section that if a logic is complete and closed under elementary
equivalence it is canonical, which is the same as being d-persistent, by The-
orem 4.8.6. Whether the converse holds is unknown and has resisted any
a t t e m p t to solve it. Moreover, in the next section we will elucidate the con-
nection between the syntactic form of elementary conditions and persistence
with respect to a class. Both questions receive only partial answers; for the
most interesting class, ~rp U ~ , they are in effect still unsolved. The first
result is t h a t when 9¢ contains all Kripke-frames and is L~-definable, any
X-persistent logic is elementary. Furthermore, we will show t h a t if :~ includes
the class of Kripke-frames and if ~ is internally describable then ~ is equivalent
to a positive and restricted formula. So all t h a t is needed in order to show
5.7. Preservation Classes 259
that Sahlqvist's Theorem is optimal is to derive the condition on variables

occurring in atomic subformulas. No one has solved t h a t yet.
Recall now from model theory the construction of an ultraproduct. In con-
nection with algebra this construction has been introduced in Section 4.1. An
ultraproduct can be defined for Kripke-frames and for generalized frames as
well. However, these are two different constructions. One is the ultraproduct
of Kripke-frames defined in the usual way. The other is the ultraproduct of
Kripke-frames viewed as general frames, i. e. as full frames. Take an index
set I, an ultrafilter U on I and a family ([i : i E I} of Kripke-frames. The
u l t r a p r o d u c t of the ~i modulo U, denoted by [ I v ~/ is defined as follows.
The worlds are equivalence classes of sequences u~ = (wi : i E I} modulo ~ ,
which is defined by g ~ v~ iff {i : vi = w~} E U. We write U~u for {~ : ~ ~ ~ }
and [iu fi for the set {~Yu : g C XiEIfi}. (Mostly, we will write z~ rather than
~ u if no confusion arises.) We put ~Yu <~j zgu iff {i : vi <~j wi} E U. This defi-
nition does not depend on the choice of representatives, as is easily checked.
Finally, [ i u fi := ( [ i u fi, (<aj : j < ~)}. Now to the ultraproduct of general-
ized frames. Let ~i, i E I, be a family of frames. The ultraproduct 1-Iu ~i of
the frames is defined as follows. The underlying K r i p k e - f r a m e of 1-Iu ~i is the
frame [ i u [i, where fi := (~i)~. The internal sets are as equivalence classes of
sequences d = {ai : i G I) modulo ~, where ~ ~ g i f t {i : ai = bi} E g . We
write du (or mostly simply 4) for the set {6: g ~ g}. Furthermore, Z~u E du
iff {i : wi C hi} C U. It is straightforward to verify that this definition does
not depend on the choice of representatives of the classes. This is the only
nontrivial fact here. Notice that elementhood really has to be defined, so we
had actually better use a different symbol here. In the same way as in model
theory we can conclude t h a t [ i u ~'i is a frame. Namely, by Proposition 5.7.1
below the so defined set of internal sets is closed under the operations.
PROPOSITION 5.7.1. Let ~i, i E I, be a family of frames, and U an ultra-

filter over I. Let [iu ~i be the ultraproduct of the ~i with respect to U. Then
(i) ~ E - 4 iff ~ ¢ 4, (ii) ~ E d;-lb iff v~ E d a n d r e b, (iii) ~ E Oj~ iff
there is a ff such that ~ <~j ff and g E 4.
The proof of this theorem is left as an exercise. Moreover, the identity
m a p is an isomorphism between (1-Iv ~i)+ and 1-Iv(~i)+. This is not hard to
see. A valuation on an ultraproduct can be seen as the equivalence class of a
sequence "~ (~ of valuations on the individual factors.
THEOREM 5.7.2. Let ~i be a family of Kripke-frames and ~ E L ~. Then

([Iv ¢ i# {i: ¢} u.
PROOF. Analogous to the elementary case. For example, let ~ = (~p)r/.
Then ([Iv ~'~,5, ~ e ¢ iff there is a valuation ~ = (7~: i E I} different from
7 at most in p such that {[iu 8~, 3'3, ~ ~ ~7 iff there is a valuation @ different
from 7 at most in p such that {i : (~,7~,Li} ~ ~7} C U iff {i : (~i,Ti,*~}

u. []
LEMMA 5.7.3 (Goldblatt). The ultraproduct l-Ig ~i is a generated subframe

of the ultrapower l-Iu((~ici ~i)-
PROOF. Take h : ~ ~-+ {(i, wi) : i E I}. Then ~<~j~iff {i : wi<~jvi} E U.
However, {i: wi < j vi} = {i: (i, wi} <j (i, vi}}, and so ~ < j g iff h(u~) <lj h(g).
Similarly it is shown that g = ~ iff h(V) = h(~), so h is injective. And
similarly for the other properties. It is easy to see that the map defines a
p-morphism, for if h(~) <~j g then for almost all i, (i, wl) <lj (i, v~). In that
case, define • by (i,v~} := (i, vi} if wi <~j vi and (i,v~} := (i, wi} otherwise.
We have ~ m ~7, so they are equal in the ultraproduct, and for g = (v~ : i E I)
we have h(~7) = ~ . []
COROLLARY 5.7.4. Any class of Kripke-frames closed under generated
subframes, disjoint unions, isomorphic copies and ultrapowers is closed under
ultraproducts.
A class of first-order structures is called e l e m e n t a r y if it is characterized
by a single sentence, and / k - e l e m e n t a r y if it is characterized by a set of
sentences. It is called E - e l e m e n t a r y if it is a union of elementary classes,
and E / k - e l e m e n t a r y if it is a union of/k-elementary classes. It can be shown
that a class is closed under elementary equivMence iff it is E/k-elementary.
Moreover, Theorem 6.1.15 of [45] states that two structures are elementarily
equivalent iff some ultrapowers of the structures are isomorphic. Hence a
class is elementary iff it and its complement are closed under ultraproducts
and isomorphisms. In the present case, the hierarchy collapses.
THEOREM 5.7.5 (van Benthem). Let 2~ be a Elk-elementary class of Krip-
ke-frames closed under generated subframes and disjoint unions. Then 3~
is ~k-elementary. Under the same conditions 3C is elementary if it is E -
elementary.
PROOF. 3~ is closed under elementary equivalence and hence under ultra-
powers. By Lemma 5.7.3 it is also closed under ultraproducts. By a standard
model theoretic result 5~ is/k-elementary. (For example, see 4.1.12(i) in [45].
There a class is called elementary if it is /k-elementary in our sense.) Now
assume that 3~ is E-elementary. Then its complement i s / k - e l e m e n t a r y hence
closed under ultraproducts as well. So, ~ is elementary. []
THEOREM 5.7.6. A modal formula globally corresponds in ~rp to an L I -
sentence iff it is preserved in ~rp under Lf-elementary equivalence iff it is
preserved under ultrapowers of Kripke-frames.
PROOF. If ~ corresponds to ~ C L f , ~ a sentence, then it is closed under
elementary equivalence. And if it is preserved under elementary equivalence
then it must be closed under ultrapowers. By the previous corollary we have

that the class of Kripke-frames for ~, Krp(~), in addition to being closed
under isomorphic copies and generated subframes is also closed under ul-
traproducts. Now, consider the complement of Krp(p). It is definable by
( 3 x 0 ) ( 3 ~ ( x 0 E ~). By 4.1.14 in [45] we have that the class of models of this
formula is closed under ultraproducts, too. Both classes are therefore closed
under ultraproducts and isomorphic images. Therefore Krp(~) is elementary.
Hence ~ corresponds to an L / - s e n t e n c e . []
We can immediately boost this up. Let :~ be class of general frames which
can be defined by a set ~5 of Le-sentences. Examples are the classes ®, ~ [ ,
~i, 9~. We can define a modal logic A to be ~ - p e r s i s t e n t if for all frames
such t h a t ~ ~ • we can infer f ~ A from ~ ~ A.
THEOREM 5.7.7 (Goldblatt). Let q~ a set of L c-sentences true in all Krip-
ke-frames. Then if a finitely axiomatizable logic A is g~-persistent, it is glo-
bally elementary in J~p.
PROOF. We want to proceed as before and show that the class of frames
for A and its complement are closed under ultraproducts. For the complement
there is no problem, we appeal again to 4.1.14 of [45]. For the class itself notice
that if we take the ultraproduct of the Kripke-frames as full frames then we
may from f~ ~ A still conclude I-[u ~ ~ A. The latter is not in general a
Kripke-frame. But we can use the fact that the underlying frame is in fact
1-Iu [~, the desired ultraproduct, and that it is in the class defined by ~, by
assumption. Furthermore, we have ~5-persistence, so [ I u [i ~ A. []
This shows t h a t g-persistent, df-persistent and ti-persistent logics are

A - e l e m e n t a r y and also the
COROLLARY 5.7.8 (Fine). Let A be r-persistent. Then A is A-elementary.
We half-complete the circle by showing the following famous theorem of
KIT FINE in [65] below. To prove it, let us recall some notions from model
theory. Let xi, i < n, be a set of variables, 9)l a first-order structure for a
first-order language L. Let F = {~/~(~) : i E I} be a set of formulae in the
variables x~, i < n. An n - t u p l e ff := (u~ : i < n) r e a l i z e s F if 9Jr ~ -),[g] for all
7 E F. F is f i n i t e l y r e a l i z a b l e in 93I if every finite subset of F is realized by
some ft. 9Jr is n - s a t u r a t e d if every finitely realizable set in n variables can
be realized. 9Jr is N 0 - s a t u r a t e d if it is n - s a t u r a t e d for every n < N0. It is
a well-known fact of model theory that for every L - s t r u c t u r e 9)I there exists
an elementary extension ® which is N0-saturated.
DEFINITION 5.7.9. A frame ~ is called m o d a l l y 1 - s a t u r a t e d if for every

set U C_ F with the finite intersection property, N U 7£ 2J. ~ is called m o d a l l y
2 - s a t u r a t e d if for every j < t~ and every set U C F such that x E OjU there
is a y ~>j x such that y E N U. 3 is m o d a l l y s a t u r a t e d if it is modally 1-

and 2-saturated.
Clearly, a frame is 1 - s a t u r a t e d iff it is c o m p a c t . Recall from Section 4.6
the notion of the refinement map. We show t h a t on m o d a l l y s a t u r a t e d frames
the refinement m a p is a p - m o r p h i s m , and the image is a descriptive frame.
LEMMA 5.7.10. Let 3 be modally saturated. Put Ux := {a E iF : x E a}.
Define ~ C_ f × f by x ~ y iff U~ -- Uy. Then ~ is a net o n 3 , and 3 / ~ is
descriptive. The algebra of sets of 3 is isomorphic to the algebra of sets of
3/~.
PROOF. Let v ~ v ~ and v <~j w. T h e n v E ['] 0 j U ~ . Then, by definition of
~ , v I C [-] 0 j U ~ . By 2 - s a t u r a t i o n , there exists a w ~ E ['] U~ such t h a t v ~<ljw ~.
Again by definition of ~ and of U~, w ~ ~ w. T h e algebra of sets over 3 / ~ is
defined as the set of sets [c] := ([x] : x C c}. Given two internal sets b, c C F
there exists an x such t h a t x C b, but x ~ c. T h e n [x] C [b] b u t Ix] ~ [c].
Hence b ~-+ [b] is bijective. This shows the last of the claims. In 3 / ~ we have
U~ -- Uy iff x -- y, and so it is refined. Moreover, it is c o m p a c t , since 3 is.
Finally, let Iv] ~ [w]. T h e n for no w' ~ w, v <~j w'. By 2 - s a t u r a t e d n e s s of 3,
therefore, v ~ ~ OjU~. So there exists a c E U~ such t h a t v ~ Ojc. []
THEOREM 5.7.11 (Fine). If A is ~rp-complete and J~tp-EA-clementary
then A is ~l-canonical.
PROOF. If ~ is A-consistent there is a A - m o d e l ( [ ~ , ~ , x ~ ) ~ ~. Let
:--- ( ~ [~ and /3 := (~w ~ " Now let F consist of all the sets of the f o r m
~ ( ¢ ) for some ¢. T h e n (~, iF) is a general frame, since iF is closed under the
usual operations. Also, 3+ ~ 3rA(~0)- For consider the m a p ~ ~-~ ~ ( ~ ) ; we
show t h a t it is an isomorphism. If 3rn(R0) ~ ~ ~ ¢ then A t- ~ ++ ¢ and so
3 ~ ~ ++ ¢, thus ~ ( ~ ) = ~ ( ¢ ) . Further, if 3tA(~0) ~z T ~ ¢ then A ~ ~ ++ ¢,
then X := ~ ( T ++ ¢) has a model (Ix, ~x, xx)" Hence the world (X, xx} E f is
a m e m b e r of ~(X), t h a t is, (3,/3) J~ -~X, or equivalently, (3, f~) / ~ ++ ¢. Now
adjoin for each c C 3 a u n a r y predicate c to L : . This defines the language
L / ( 3 ) . E x p a n d [ to an L Y ( 3 ) - s t r u c t u r e [o by interpreting c as the set c itself.
(This allows to forget the set structure on 3 for a while.) T h e r e exists an
e l e m e n t a r y extension of [° in the language L / (~), denoted by t~°, which is Ro-
s a t u r a t e d . P u t [_c_]:= {w e g : c__(w)}. T h e [_c_]are closed under intersection,
c o m p l e m e n t and I j . In particular, we have -[_c] = [-_~, [_c] V~ [d] = [c A d]
and I j [ c ] = [Ijc]. For these are e l e m e n t a r y s t a t e m e n t s which hold in [o,
thus they hold in ~°. T h e m a p k : c ~-+ [c] is therefore an i s o m o r p h i s m of the
algebras. Let ~ be the L : - r e d u c t of ~o. P u t G := ([_c] : c ~ iF}. We have
m a n a g e d to get a structure ® = (~, G) such t h a t ®+ ~ 3rh(R0) with ~+ ~ 0 -
s a t u r a t e d . We show t h a t ® is m o d a l l y s a t u r a t e d . Namely, if A C_ G has the
finite intersection property, then [-] A ~ ~. For under the given a s s u m p t i o n ,
{_C(Xo) : [c] E A} is finitely satisfiable; by saturatedness of 9 ° there exists a w

such that c(w) for all [c] C G. Hence w C N A. So, ® is 1-saturated. Second,
if A is a set such t h a t for every finite subset Ao, v E e j N A0, then there exists
a w such that v <~j w and w E c for all c C A. For by assumption on A, the
set {v <]j y A c(y) : [c] E A} is finitely satisfiable. Hence by saturatedness
there exists a w such that v ~ j w and c(w) for all [c] E A. So, ¢~ is modally
2-saturated. By L e m m a 5.7.10 the refinement m a p is a p - m o r p h i s m from ~5
onto a descriptive frame whose set algebra is isomorphic to ~rA(R0). Hence
there is a p - m o r p h i s m ® --, ~rA(R0). Now, for all W we had [~ ~ A and
so ~ ~ A. t~, being an elementary extension, also satisfies A, by assumption
that Krp A is closed under elementary equivalence. Finally, ~rh(R0)~ ~ A by
closure under p-morphisms. []
The proof works analogously for any cardinal a _> R1. Using this theorem
we can obtain a partial converse of Theorem 5.7.11. Namely, if we have a logic
which is RI-canonical and elementary, then we can get ~lZA(a ) in a similar
process from finite models. Moreover, the following holds as well.
THEOREM 5.7.12 (Fine). Suppose that A is ~rp-complete and ~ p - E A -

elementary. Then A is canonical.
The connection between canonicity and elementarity is a very delicate
one as FINE shows in [65]. Consider the logic O := K ® p where
:= 0[Zp --+ 0[](P A q) V ()[~(p A ~q)
We claim the following: (a) e is canonical, (b) Krp(O) is not E / k - e l e m e n t a r y
and (e) (9 is complete with respect to some elementary class of Kripke-frames.
So, canonicity does in general not imply elementarity. It is believed until today
that it does imply completeness with respect to some ( A - ) e l e m e n t a r y class of
frames. 0 is a case in point. Let us prove the stated properties of ®. Consider
the formula
e(x) := (Vy D x ) ( q z E>x)(Vu, v D z)(u -- v A y <~z)
LEMMA 5.7.13. Let ~ be a frame. If ~ ~ (Vx)¢(z) then ~ ~ ~.
PROOF. Assume ~ ~ (Vx)e(x) and let /3 and x be such t h a t (~,/3, x)

0f-lp. Then for some y D x we have (5,/3, Y) ~ r-lp. Now, by our assumptions,
there is a z D x such that z is either a dead end or has exactly one successor.
If z is a dead end, 71(pAq) is true at z and so 0[q(pAq) is true at x. So, let us
assume t h a t z actually has a successor, u. Then y<Zu. Since Vlp holds at y, p is
true at u. Now either q is true at u or not. In the first case, (~,/3, z) ~ E](pAq)
and so (~,/3, x) ~ 0[](p A q). In the second case (~,/3, x) ~ 0E~(p A ~q), as
required. []
LEMMA 5.7.14. Let ~ be a canonical O-frame. Then ~ satisfies (Vx)e(x).

PROOF. Since 3 is canonical, worlds are actually maximally consistent

sets of formulae. Let X be a point of 3. In case that X has no successor or
sees a dead end (that is, in case [Z± E X or (}[3± C X) ¢ holds trivially of
X. So, let us consider the case where this is not so. Then pick a successor Y.
We have to show that there is a Z with exactly one successor, call it U, such
that Y <3 U. Now choose an enumeration Xn of the language. Inductively we
define formulae an as follows, a0 := T. a n + l := a n A X.n if O[q(an A Xn) E X
and [3(an A Xn) E Y, and a n + l := a n A -'Xn if 0K](an A -~Xn) E X . Finally,
we let U be the MP-closure of the set {an : n C w}. By force of the axiom
~, this is actually well-defined and if <>[Zan ff X then also (>K]an+l E X .
Moreover, <>[]a0 C X, and so (>Dan E X for all n C w. It follows t h a t U is
consistent. For if not, for some n, an is inconsistent, that is, a n ~- ±. But
since <>[]an C X, also <>[:]i E X, which we have excluded. Furthermore, by
construction, for every n either Xn C U or ~Xn C U. So, U is a maximally
consistent set, and so a world of 3. Also, for every [qX E Y, X C U, again
by construction. It follows that Y <~ U. Finally, put A := {[]X : X C U}
and D := {X : DX ff X}. We claim that A U D i s consistent. If not, there
is 5 E D and [~a C A such that 5;K]a F- ±. (A and D are closed under
conjunction.) So, [Na ~- 95 and hence <>[]a ~- <>~5. Since (>[qa • X we also
have ~K]5 • X, against the definition of D. So, A U D is consistent and is
contained in a maximally consistent set Z. Then if []X • Z, X • U and so
Z <~ U. Furthermore, if Z < V then A C_ V, from which U C_ V. Since U is
maximally consistent, U = V. Finally, assume t h a t []X • X. Then X • D and
so X • Z, by construction. This shows that Z has the desired properties. []
COROLLARY 5.7.15. O is canonical and complete with respect to an ele-
mentary class of frames.
PROOF. Let 3 be some canonical frame for ®. Then by L e m m a 5.7.14
3 ~ (Vx)c(x), whence also 3~ ~ (Vx)E(x). By L e m m a 5.7.13, 3~ ~ P. Hence ®
is canonical. Clearly, this shows that ® is complete with respect to the class
of Kripke-frames satisfying (Vx)~(x). []
LEMMA 5.7.16. Krp(O) is not EA-elementary.
PROOF. Let F be a nonprincipal ultrafilter on w. Let O be the union
of w, F and {F}. It is easy to see that this union is disjoint. Finally, put
<l := {(x,y) • 0 2 : y • x} and ~ := (O,<~}. It is not hard to show t h a t
~t ~ ~. O is not countable. By the downward Lhwenheim-Skolem theorem,
~t has an elementary countable submodel p = (P, <~p}, where P C O and
<~p -- ~ N p2. We will show t h a t p ~ ~. This establishes the claim. The
following properties of p are easy to establish. (1) p contains the root of ft. (2)
P N F is infinite. (3) For any two M, M ' • P N F the set s u c c ( M ) N suec(M')
is infinite. Now let P N F = {]Vii : i • w} be an enumeration of P N F . We can
assume by (3) that there are sequences (a~: i • w) and (bi: i • w} of natural
numbers which are all distinct such that all ai, bi C Mo N P, and Mn <~an, bn.
P u t fl(Po) := M o N P and fl(Pl) := {ai : i E w}. Then (p,fl, F ) ~ 0E]po
since (p, 13, Mo) ~ [~Po- Now take an M such that F ~ M. Then for some n,
M = Mn. Then Mn ~ 0 ( p o A p l ) since Mn <~an and an ~ p o ; p l . On the
other hand bn ~ Po; ~Pl, and since M~ <~ b,~ we also have M~ ~ 0(po A -'Pl).
It follows that (p,/3, F ) ~ -@[3(p0 Apl); ~012(p0 A ~Pl). So, p ~ ~. []
Let us close with a characterization of classes of Kripke-frames which are

both modally definable and elementarily definable. Those classes must be
closed under the standard operations of generated subframes, disjoint unions
and p-morphic images. However, notice that we also have in the case of
generalized frames the closure under biduals for both the class and its com-
plement. The bidual of a Kripke-frame is not necessarily a Kripke-frame;
however, let us define for a Kripke-frame ~ the u l t r a f i l t e r e x t e n s i o n ue([)
to be the Kripke-frame underlying the bidual of ~, that is, ue(~) = ([~+)+~.
The following has been proved in [83 I.
THEOREM 5.7.17 (Goldblatt ~: Thomason). A class of Kripke-frames is

both modally and elementarily definable iff it is closed under generated sub-
frames, disjoint unions, p-morphic images and ultrafilter extensions, while its
complement is also closed under ultrafilter extensions.
If we analyse the proof of Theorem 5.7.11 we see that it proves that the
ultrafilter extension of a frame is a p-morphic image of some ultrapower. This
is a rather useful fact.
THEOREM 5.7.18. The ultrafilter extension of a Kripke-frame g is a con-

tractum of an ultrapower of t~.
Notes on this section. The theorem by KIT FINE was proved again by
BJARNI JONSSON in [112] using methods from universal algebra. Later, YDE
VENEMA has generalized the results as follows (see [222]). Call an elementary
condition a a pseudo-correspondent of X if c~ holds in the Kripke-frames
underlying each canonical frame for K O X, and every frame satisfying c~ also
satisfies X. Obviously, the formula (Vx)c(x) above is a pseudo-correspondent
of ~. Now let 7r(p) be a positive formula in one variable. Then the formula
7r(p V q) ++ 7r(p) V 7r(q) has a first-order pseudo-correspondent. Hence it is
canonical, complete with respect to an elementary class of Kripke-frame, but
its class of Kripke-frames is in general not EA-elementary. The formula of
KIT FINE falls into this class of formulae, as is easily demonstrated.

5.8. S o m e R e s u l t s f r o m M o d e l T h e o r y
In this section we will use some techniques from model theory which en-
able us to derive characterizations of modally definable elementary conditions.
For an extensive exposition the reader is referred to VAN BENTHEM [10] and
also for a survey to [11]. Here we will basically prove two of the results, which
we consider to be the most central.
DEFINITION 5.8.1. A first-order formula c~(Z) on frames is preserved
under generated subframes if for each model <g, L> ~ c~(~) and each ~ H
such that ~(~) C f also ([, L) > c~(r~), c~(Z) is reflected under generated sub-
frames if (~c~)(aT) is preserved under generated subframes. ~(r~) is i n v a r i a n t
under generated subframes if it is both preserved and reflected under generated
subframes.
The following theorem is an analogue of a theorem by GOLDBLATT (mo-
deled after FEFERMAN [59]).
THEOREM 5.8.2. A first-order formula a( ~) with at least one free variable
is invariant under generated subframes iff it is equivalent to a t3(~) C tRf.
PROOF. Surely, if c~ is equivalent to a restricted/3 in the same variables,
then c~ is invariant under generated subframes. The converse needs to be
established. Thus assume that c~(~) is invariant under generated subframes,
and let c~ be consistent. Moreover, c~ contains at least one free variable, say x0.
Let R(c~) :-- {~(~) :/3 E :X f and c~ ~/3}. Thus if we can show that R(~) ~
we are done. Namely, by compactness there is a finite set /k C R(~) such
that /k ~ ~ and so, taking/3 to be the conjunction of/k, we have found our
desired formula. Thus assume R(c~) has a model 3o = (~o, L). (From now on
we will suppress the explicit mentioning of the valuation; thus when we speak
of a model 30 we mean the frame endowed with a valuation.) The language
we are using is called L0, for future reference. Now form LI by adjoining a
constant ci for each variable xi occurring free in c~. Expand the model 3o to
a model 31 by interpreting the constants ci by ~(xi). We claim that the set
:= {5 C ~ : al ~ 5} U {o~[~'/:~]}
is consistent. Otherwise there is a finite set - - or indeed, a single formula
5(2), by closure of the set under conjunction - - such that a[g/2] ~ (-~5)[~'/~],
whence 75 E R(a). But this contradicts the definition of R ( a ) , by consistency
of a. So, E C_ L1 has a model, ®1. Moreover, every restricted L l - s e n t e n c e
is true in 31 i f f i t is is true in ~51. (We say briefly that 31 and ~1 are r -
equivalent.) For, if 5 is restricted and holds in 31 then it is in E and holds in
®1, too. And if 5 fails in 31, then ~5 holds in 31, so --5 holds in ~51 as well.
This is now the starting base. We have two models 31 and ®~ over a c o m m o n
language L~, such that ®~ is a model for a[c~/~] and b o t h are r-equivalent.
We will now construct sequences Li of languages, and 3i and ®i such t h a t
5.8. Some Results from Model Theory 267
1. ~i, 6 i are L~-models.

2. 6~ ~ ~[¢/~].
3. The same Li-sentences hold in 6 i and ~i- (~i and 6 i are r-equivalent.)
4. ~i-1 is an L~_l-elementary substructure of ~i and 6 i - 1 is an L i - 1 -
elementary substructure of 6i.
We have started the construction with i = 1. Now let the construction proceed
as follows.
C a s e 1. i is odd. First we define L~+I. Assume that we have a constant c
in L~ and that 6~ ~ c <j x for some x E 6i. Then adjoin a constant x__for
x. Do this for all x of this kind. This defines L~+I. 6~+1 is defined as the
expansion of 6 i in which _x is interpreted as x. Then 6~ is an Li-elementary
substructure of 61+1, which we abbreviate by 6~ 4~ 6i+1. Now form the set
This set is finitely satisfiable in ~ . To show this it is enough to see that every
formula 5 ~ E is satisfiable in ~ . Let 5 ~ E be given. Now retract the new
constants in Li+l as follows. Let x__~ L~+~ - Li occur in 5. Then there is a
c such that 6~ ~ c <~j x. Put 51 := (~y ~>j c)5[y/x_]. Continuing this process
until there are no free variables outside L~ left, we get a formula 5, G ~i.
Now since 6 i ~ 5, and ~ is r-equivalent to 6~, we get ~ ~ 5,. Hence 5 is
consistent, (~, being the existential closure of 5. Therefore, E is consistent. So
there is a model ~ + 1 in the language L~+I such that ~ 4~ ~ + 1 and 6~+1
and ~ + 1 are r-equivalent in L~+I.
C a s e 2. This step is dual. Now adjoin constants for worlds x such that ~
c <~ x, c ~ L~. Interchange the roles of ~ and 6~ in the above construction.
f
61 41 62 42 63 43 64 --" 6", 6°
L1 L2 L3 L4 • .- L~
Let L~ := U i L i - Now consider the structure ~* := ~Ji~i" Since ~i

is an Li-elementary substructure and also contained in ~ + 1 , we can invoke
Tarski's L e m m a on elementary chains to conclude that ~i is Li-elementary
in ~*. Likewise for 6*. Consider the substructures induced by the constants.
That is, let ~o be the subframe of ~* based on all interpretations of constants
c E L~ and likewise 6 ° the subframe of all interpretations for constants in
6*. We claim that ~° --~ ~* and 6 ° --~ 6*. Namely, if x E ~* is the value
of a constant c in L~, there is an i such that c C L~, and so c is interpreted
in ~ . Let x ,~j y for some y C fi- If i is even then there is a y E L~+I such
that y is interpreted in 3i+1, and its interpretation is y by construction. If i

is odd, then a y will be introduced in Li+ 2. Similarly for 6 °. Now, define a
m a p f : 3 ° --+ ®o as follows. We put f ( x ) := y if there is a constant c such
t h a t c is interpreted as x in 3 ° but as y in ®% We claim t h a t this function is
an isomorphism. (i.) It it is defined on 3 °. (ii.) If c and d are two constants
such that their interpretations coincide, then they coincide in a 3i for some
i, and so 61 ~ c - d, from which follows t h a t 6 " ~ c - d. So the definition
of f is sound. (iii.) f is onto, by definition of 6 °. (iv.) f is injective.
For if c and d are interpreted by different worlds in 3" then 3" ~ -~(c - d),
whence3~ ~ ~ ( c - d ) , for all i such that c, d E L~. Then 6~ ~ ~ ( c - - d ) , by
r-equivalence. And (v.) 3 ° ~ c .qj d iff 6 ° ~ c <~j d, by the fact t h a t the two
are generated subframes and the formula is restricted.
Thus, since 61 ~ a[~'/2], we also have 6 " ~ a[K/2] and therefore 6 °
a[K/~], since a was assumed to be invariant under generated subframes. Now
also 3 ° ~ a[~'/~], by the fact that the two models are isomorphic. Again
by invariance under generated subframes, we get 3" ~ (x[K/~] and finally
31 ~ a[~'/~], which is to say (~0, ~) ~ a. This had to be shown. []
DEFINITION 5.8.3. A first-order condition a(~) is said to be p r e s e r v e d

under contractions if whenever (~, ~) ~ a(~) and p : ~ --~ ~ is a p-morphism
then for t'(xi) := p(L(xi)) we have (~, L') ~ a(~). a(~) is reflected under con-
tractions if (~a)(:~) is preserved under contractions, and it is called i n v a r i a n $
under contractions if it is both preserved and reflected under contractions.
Below we will prove that a formula with at least one free variable is invariant
under generated subframes and preserved under p - m o r p h i s m s iff it is equiv-
alent to a positive R f - f o r m u l a . The proof is taken from VAN BENTHEM [10].
This shows that if a defines a modal class in ~rp U ~ then it is equivalent to a
restricted positive formula. To show that Sahlqvist's Theorem is the best pos-
sible result for local correspondence, two things need to be established. (1.)
Every locally ~ - p e r s i s t e n t logic is elementary, (2.) every restricted positive
a defining a modal class in ~ p U© is equivalent to a Sahlqvist formula. Both
are still open questions. Notice that for (1.) and (2.) the choice of the class
seems essential. We have seen (Corollary 5.7.8) t h a t (1.) holds if the class 9~
is chosen instead. If it holds for the class ~ we would have the converse of
T h e o r e m 5.7.11. It is unknown whether or not this holds. Also, with respect
to (2.) it is also not known whether it holds. In the exercises we will ask the
reader to show that if a E L f is at most V~ and modally definable, then a is
equivalent to some restricted formula.
THEOREM 5.8.4 (van Benthem). A first-order formula a(£) with at least
one free variable is invariant under generated subframes and preserved under
contractions iff it is equivalent to a positive restricted formula in the same
free variables.
PROOF. We proceed as in the previous proof. The construction is some-

what more complicated, though. It is clear that a positive 9~l-formula is
invariant under generated subframes and preserved under p-morphisms. The
converse is the difficult part of the proof. Let a = a(~) be a formula in xl,
i<n, withn>0. Put
RP(a) := {/3(Y):/3 E 9~I,/3 positive and a ~ / 3 } .
If we can show that RP(a) ~ a we are done. For then, by compactness of

first-order logic, there is a finite subset A of RP(a) such that A ~ a. The
conjunction 5 of the members of A is also in RP(a). It follows that a ~ 8 ~ a,
as desired. Let 3 be a model for RP(a). Let xi be interpreted by wi. The
language L1 is obtained from our initial language by adjoining a constant wi,
i < n. 31 is the expansion of 3 in which wi is interpreted by wi. Consider
now the set
E := {a[tg/~]} U {-./3:/3 a restricted positive Ll-sentence, 31 ~ -./3}-
E is finitely satisfiable. For if not, there is a finite subset E0 := {a[~/zg]} U

{=/3i : i < p} such that E0 ~ f. This however implies that a[~/g.] ~ Vi<p/3i.
So, a ~ (Vi<p/3i)[•/w_+](=:/3). (Here, bound variables of the form xi are
suitably renamed.) Thus /3 C RP(a). But then, by definition of 31, 31
/3[v7/i], in contradiction to the definition of E, because 31 ~ ~/31 for all i < n.
Hence, Eo is consistent. So, E is finitely consistent and therefore has a model,
(51 (for example, an R0-saturated extension). The following holds of ¢~1.
(i)
(ii) Every positive restricted L1-sentence which holds in (51 also holds in
31.
We will now construct a series of languages 2-i, i < n, such that 2.,i C "~i+1,
Li an expansion of L0 by constants, and two series of models, 3i and (5i, such
that f o r i > 1
1. ~i, ®i are Li-models.

2.
3. Every positive restricted Li-sentence which holds in (5i also holds in
3i-
4. 3i--1 is an L i _ l - e l e m e n t a r y substructure of 3i and (5i-1 is an L i - 1 -
elementary substructure of (5i.
We repeat the picture from the previous proof in a modified form.

~1 ~1 ~2 ~2 ~3 ~3 ~4 "'" ~** ~ o
71"
~1 ~1 ~2 ~2 ~3 ~3 ~4 "'" ~*'
L1 L2 L3 L4 ... L~
Assume that we have Ln and ~,~, ®n satisfying (0.) - (3.). We show how to
produce L~+I, ~ + 1 and ~ + 1 satisfying (0.) - (3.). Let c be a constant of
L~ and w C (5,~ such that ~n ~ (c <lj x)[w/x]. Then add a new constant w_
to L~. L 1 shall denote the language obtained by adding all such constants.
~,~ is expanded to an L~-structure by interpreting w by w. Let
A := {/3:/3 a restricted positive L~-sentence such that ®~ ~ 13}.
We show that each finite subset of A can be satisfied in a model which is an
expansion of ~ . Let namely A0 := {/3i : i < p} be such a set, and denote
its conjunction by 5. Assume that wj, j < k, are the constants of L I - L~
occurring in (f. By construction of L I there exist constants cj, indices j(i) <
and variables xi for all i < k such that
®n (3x0
This formula is an L,-sentence. By (2.) it holds in ~n- Hence we find
corresponding values for the constants w_j, j < k. There exists an ~1 such
that
(a) ~ I is an L~-structure.
(b) ~n is an Ln-elementary substructure of ~I-
(c) Every positive restricted L1-sentence which holds in ~51 also holds in
Now we turn to the dual step. For each constant c of L 1 and each w in
~ such that ~ ~ (c <j x)[w/x] add a new constant 7(c,j,w). Let Ln+l
be the language obtained by adding all such constants. Expand ~ I to an
~ , + l - s t r u c t u r e by interpreting ~/(c, j, w) by w. Let
:= {~fl :/3 a restricted positive L~+l sentence such that ~,~+1 ~ -~fl}
We show that each finite subset of ~ is satisfiable in an expansion of ~ . For

suppose that ~-0 is a finite subset of E. It consists of a set of -~/3i, i < s, from
the first set, and a set C := {cl <lj, 7(ci,jl,xi) : i < t}. The conjunction
of the ~/3i is denoted by ~. ~ is also in ~, so we may take ~ in place of the
first subset of ~0. We may additionally assume that C already contains all
j, w) n+1 - that appear in Now suppose that C U is not

satisfiable in any expansion of (5~. Then
(5~ N (YXo E>jo c0)(Vxl E>jl e l ) . . . (Vxt-1 :>j,_l ct-1)(~[xi/~/(ci,ji,wi): i < t])
This is a positive restricted L~-sentence, so it holds in 3~, by (c.). But
<5, :i < : i < t]
This is a contradiction. Hence, every finite subset of ~ is satisfiable in an

expansion of (51 and so there exists a (5,~+1 such t h a t
(a) (5n+1 is an L n + l - s t r u c t u r e .
(b) (5n+11 is an Lln-elementary substructure of (5~+1-
(c) Every positive restricted L~+l-sentence which holds in (5n+1 also holds
in 3~+l-
The claims (1.) - (4.) now easily follow for Ln+~, ~n+~ and (5n+1 using (a.),
(b.) and (c.). We have that ~n is a n L ~ - e l e m e n t a r y substructure of 3~+l-
By induction it follows that ~n is an L n - e l e m e n t a r y substructure of 3n+m for
every m. Similarly for (5~. Let L~ := [-Jie~ Li, and let 3" be the union of the
3i, (5* the union of the (5i. Then 3" and (5* are L ~ - s t r u c t u r e s , of which ~
and (5~ are respective L ~ - e l e m e n t a r y substructures. Furthermore, let 3 ° be
the transit of the wi in ~* and (5° the transit of the interpretation of the w i
in (5*. By construction, the constants of L ~ all take values in ~° ((5°), and
every element of 3 ° ((5°) is the interpretation of a constant from L~. Define
7r : (5° --+ 3 ° as follows. If x E (5°, let c be a constant with interpretation
x. Then 7r(x) := y, where y is the interpretation of c in 3 °. (1.) 7r is well-
defined. Suppose that c and d are interpreted by x. Then there is a n E w
such that c C Ln and d E Y-,~. Then (5~ ~ c -- d, and so (5* ~ c - d, being
an L ~ - e l e m e n t a r y superstructure. Hence ~* ~ c -- d, and this shows that the
interpretation of c is equal to the interpretation of d in 3" (and it is in 3°).
(2.) rr is onto. For since every element of 3 ° is in the interpretation of some
c. Take x E (5 ° such that x is the interpretation of c in (50. Then rr(x) = y.
(3.) rr is a p - m o r p h i s m . (a) Suppose x <~j y. Let x be the interpretation of
c in (5° and y the interpretation of d. Then (5* ~ c < j d. There is an n such
that c E L~ and d C Ln. For this n, (5,~ ~ c<ajd. H e n c e 3 n ~ c<~jd, and
so 3" ~ c <~j d. Therefore 7r(x) <~j rr(y). (b) Let lr(x) <~j u. There exists a
constant c such that c is interpreted by ~r(x) in 3*. T h e n consider 3@, J, u).
By construction, (5* ~ c <~ ?(c,j,u). Let y be the interpretation of ",/(c,j,u)
in (5*. The interpretation of 7(c,j, u) in ~* is just u. Therefore, ~r(y) = u,
and x <lj y, as required.
Finally, we know that (5~ ~ ct[~/~]. Hence, (5* ~ a[~/~.]. Since c~
is invariant under generated subframes, (5o ~ a[~/aF], a is also preserved
under contractions, and so ~° ~ a [ ~ / ~ ] . Again by invariance under generated
subframes, 3" ~ a [ ~ / ~ ] . Finally, since ~ is a L~-elementary substructure,
31 ~ a [ ~ / ~ ] . By construction, this is the same as 3 ~ a [ ~ / ~ ] . This is the

desired conclusion. []
The proof in MARCUS KRACHT [124], supposed to be a construction of the

equivalent formula, is actually incorrect. There is a plethora of similar results
concerning the interplay between syntactic form (up to equivalence) and in-
variance properties with respect to class operators. A particular theorem is
the following.
THEOREM 5.8.5. Call a formula ~(~) constant if no prime subformula
containing a variable is in the scope of a quantifier. Let a( Z) be a formula
with at least one free variable. Assume that a is invariant under generated
subframes and contractions. Then a is equivalent to a constant formula X(~).
PROOF. The direction from right to left is easy. So, assume t h a t a is
invariant under contractions and generated subframes. Let a ( ~ ) be given.
The initial language is ~ / . P u t
C(a) := {/3(~) :/3 constant, a ~ ~}
Clearly, a ~ C(a). We show that C(a) ~ a. To do that, let 3 ~ C(a), where
xi is m a p p e d to wi. Adjoin new constants w~. This yields the language/21.
Make 3 into a / 2 1 - s t r u c t u r e 31 by interpreting w i by wi. Now let
:= {a[~/~]} U { 5 : 5 a constant/21-sentence and 31 ~ 5}
is finitely satisfiable, and so there is a model (51. Let 3" be a Ro-saturated
expansion of 31 and (5* a Ro-saturated expansion of (51- Further, let 3 °
be the subframe of 3" generated by w~, i ~ n, and (5o the subframe of (5*
generated by the interpretation of w__/. Now, define a binary relation ~ on
3 ° by u ~ u ~ iff u and u ~ satisfy the same constant :R/-formulae. It is not
hard to show that this is a net. For if u ~ u' and u <lj v, let D(y) be
the set of constant :R/-formulae in y satisfied by v. Then let (}jD(y) :=
{(3y E>j x)5(y) : 5 C D(y)}. u satisfies (}jD(y). Hence u' satisfies OjD(y). By
saturatedness, there is a successor v ~ of u ~ satisfying D(y). Similarly, define
on (5 ° by u ~ u t i f f they satisfy the same constant ~ / - s e n t e n c e s . Now,
let ti interpret w__i in (51. Then ti satisfies the same constant formulae as
does wl. Therefore, it can be shown by induction on the depth t h a t 3 ° / ~ is
isomorphic to (5°/~. Now the conclusion is easily established. ®~ ~ a[~/Z].
By elementary embedding, (5* ~ a[~/~] and so (5° ~ a[~/~], (5o/~ ~ a[w__/:~]
by preservation under generated subframes and contractions. T h e n ~ ° / ~
a[w/~], since it is isomorphic to the latter structure. It follows t h a t 3~
a[~/~], and finally 3 ~ a[~/~]. []
THEOREM 5.8.6 (van Benthem). A logic K ® X is g-persistent iff there
exists a constant formula V(x0) such that the class of frames for that logic is
defined by 7(x0).
These theorems establish the weak form of completeness of the calculus S e q

for ¢~ discussed at the end of Section 5.4.
E x e r c i s e 195. (n < N0.) Let a(x) E L / be modally definable in .ff.I:p. We can

assume that a is in prenex normal form; furthermore, assume t h a t a contains
only V-quantifiers. Show that a is equivalent to a formula/3 in which every
universal quantifier Vy is replaced by a restricted quantifier (Vy ~>s x) for some
x and some finite set s of finite sequences over n. Hint. Let 8i(x) be obtained
by replacing each quantifier Vy by (Vy~>sn x), where s~ consists of all sequences
of length _< n. Then 8n(z) F- (~+l(x) in predicate logic, and 8n(x) ~- a ( x ) . It
remains to prove t h a t some n exists such that a(x) ~- 8n(X). Suppose t h a t no
such n exists; then the set {-,Sn(x) : n E w} (2 {a(x)} is consistent. It has a
model (f, L). Now use the fact that a(x) is closed under generated subframes.
E x e r c i s e 196. As above, but for a of the complexity V3. (Assume n < N0;
see [128] for a proof.)
E x e r c i s e 197. The same as the previous exercise, but without the restriction
CHAPTER 6
Reducing Polymodal Logic to Monomodal

Logic
6.1. Interpretations and Simulations

The main body of technical results in modal logic is within monomodal
logic, for example extensions of K4. The theory of one operator is compara-
tively speaking well-understood. Many applications of modal logic, be they
in philosophy, computer science, linguistics or mathematics proper, require
several operators, sometimes even infinitely many. Moreover, in the theory
of modal logic many counterexamples to conjectures can be found easily if
one uses logics with several operators. So there is a real need for a theory of
polymodal logic. On the other hand, if such a theory is needed and we have
developed a theory of a single operator, it is most desirable if we could so to
speak transfer the results from the one operator setting to several operators.
This, however, is not straightforward. It has often been deemed a plausible
thing to do but turned out to be notoriously difficult. Only fairly recently
methods have been developed that allow to transfer results of reasonable gen-
erality. They go both ways. It is possible to interpret a monomodal logic as
a polymodal logic, which involves axioms for one of the operators only. Let
us call these one-operator logics. They were introduced by S. K. THOMASON
[211] and systematically studied in KiT FINE and GERHARD SCHURZ [67] and
also MARCUS KRACHT and FRANK WOLTER [132]. If we fix an operator, we
have a natural embedding of monomodal logic into polymodal logic. We can
also study unions of one-operator logics, where the distinguished operators
differ. Such logics are called independently axiomatizable. For independently
axiomatizable logics there exist a number of strong transfer results. Basically,
all common properties of the one operator logics transfer to the union.
This direction is unsurprising, perhaps. Moreover, polymodal logics con-
tain monomodal logics. The converse, however, is prima faeie unplausible for
it suggests that we can model several operators with the help of a single one.
Yet, exactly this is the case. S. K. THOMASON has proved a lot of negative
results for monomodal logic by reducing polymodal logic and other sorts of
logics to monomodal logic. However, it has gone unnoticed that not only neg-
ative properties of logics such as incompleteness, undecidability and so on are
275
276 6. R e d u c i n g P o l y m o d a l L o g i c t o M o n o m o d a l Logic
transferred, but also positive ones. Once this is noticed, we derive a plethora
of strong results concerning monomodal logic. In this way we can gain insight
not only into polymodal logic but also into the theory of a single operator.
Before we enter the discussion of polymodal versus monomodal logic, we
need to make our ideas precise concerning reduction. We have mentioned
two cases of reduction, one from polymodal logic to monomodal logic, and
another from monomodal logic to polymodal logic. More precisely, these
reductions consist of a translation of one language into another. Moreover,
this translation reduces a logic in the first language to a logic in the second
language if it is faithful with respect to the deducibilities. This is made precise
as follows. Let L1 and L2 be two propositional languages with variables
drawn from var. An i n t e r p r e t a t i o n of L1 in L2 is a m a p which assigns to
the variables uniformly an expression of L2 and to each connective of L1 a
possibly complex functional expression of L2. This means that for I : L1 -+ L2
to be an interpretation it must satisfy
(f(~0, • .. , ~k-1)) I . (f(Po,

. . .. ,Pk-1))I[~I
. . /po, ,))k-1/Pk-1]
I
for all connectives f in 51, formulae ~ 0 , . . . , ~ k - 1 E 2oz, and variables Pi,

i < k; and for all variables p, q
= pi[q/p]
In brief, an interpretation is fixed by the t e r m it assigns to a simple expression;

interpretations are not to be confused with homomorphisms. First of all, an
interpretation is a map between languages with possibly different signatures.
Moreover, even when the signatures are not different, the concept itself m a y
still differ. For example, the duality m a p is an interpretation, though strictly
speaking not a homomorphism, because conjunction and disjunction m a y not
be interchanged by a homomorphism. Furthermore, an interpretation is free
to assign a complex term to a simple term. For example, we might choose
to interpret [] as R(} (for example, in interpreting non-classical m o n o m o d a l
logics as bimodal logics, see [133]).
The definitions have some noteworthy consequences. First of all, a vari-
able p is translated into an expression p l which contains at most the variable
p, that is, var(p I) C_ {p}. For if q ~ p we have p1 = qI[p/q], thus q f~ var(pI).
Likewise, for any expression p we have var(F I) C_ var(~).
Now consider two logics (2O1,~-1} and (L2,72) and an interpretation I.
Then ?2 s i m u l a t e s kl with respect to I if for all F C_ L1 and ~ C 131
F k 1 ~) iff F I ~-~ ~ J .
Denote by gI(~-l) the set of all consequence relations ~- over £,2 which simulate
F-1 with respect to I. It is readily checked t h a t SI(~-1) contains a minimal
element. The following is a fundamental property of simulations.
6.1. Interpretations and Simulations 277
PROPOSITION 6.1.1. Suppose that F-2 simulates ~-1 with respect to some
interpretation I. Then if ~-2 is decidable, so is k t.
For a proof just observe that by definition the problem F ~-1 ~ is equiv-
alent to F I ~-2 ~J- A priori, a connective can be translated by an arbitary
expression. However, under mild conditions the interpretation of a boolean
connective ® must be an expression equivalent to ®. In the case of modal
logics this means that under these conditions only the modal operators receive
a nontrivial interpretation. Call an interpretation I a t o m i c if p I = p for all
propositional variables p. In this case fI(c,a0,.. • , ~k-1) will be used instead
of f ( P o , . . . ,Pk-1)I[~o/Po,... , ~k-1/Pk-1]- The following then holds, as has
been oberserved in [133].
PROPOSITION 6.1.2 (Wolter). Write ~ =-2 X for qa ~-2 X and X ~-2 ~.

Suppose that A and ~ are both symbols of LI and L2, that the restrictions of
~-1 and k2 to expressions containing -~ and A equals the propositional calculus
over -~ and A. And suppose that we have the r e p l a c e m e n t rule for k2; that
is, if ~l - 2 ~2 then ¢[~al/p] =-2 ¢[~2/P]. Finally, suppose that I is an atomic
interpretation. Then if ~-2 simulates V1 with respect to I the following holds:
1. p A q = - 2 p A ~ q.
2. ~ p =-2 R I P •
PRooF. (1.) We have p A q ~ {;, q} ~ p A ~ q k~ {p, q} ~ p A q. (20 It

is readily checked that ~ is ~-l-inconsistent iff qoI is ~-2-inconsistent. Hence,
p ; - J p is k2-inconsistent, since p;-~p is ~-l-inconsistent. Hence ~Sp k2 ~p.
It remains to show that ~ 1p; ~p is k2-inconsistent. But this follows with
q t-2 ((P A q) V (~p A q)ff and (i) by
~_~lp A ~p
~-2 (PA I ( ~ - ~ I p A ~ p ) ) V I ( I p A I (~ I p A ~ p ) )
k2 (pA z ~p) V I ( ZpA z ~ ~p)
u~ ((p A ~p) v (p A ~p)y
and the ~-2-inconsistency of this last formula. []
It is worthwile to reflect on the notion of a simulation. We will use it

also to show undecidability of certain logics. The rationale will be to use
well-known undecidable problems, in this case facts about word-problems in
semigroups, and simulate these problems in polymodal logics. Furthermore,
as polymodal logics can themselves be simulated by monomoda] logics, this
yields undecidable problems for monomodal logic. We shall indicate here that
the notion of simulation is quite similar to a notion that is defined in RYSZARD
WOJCICKI [231].
6.2. S o m e P r e l i m i n a r y R e s u l t s
In the next sections we are dealing with the following standard situation.
We have a bimodal language L2, denoted here by L c l , and two m o n o m o d a l
fragments, Los and L i . Naturally arising objects such as formulae and con-
sequence relations are subject to the same notation, which we assume to be
clear without explanation. There are two possible interpretations of a single
operator - - denoted here by [] - - in bimodal logic over [] and • . We m a y
read it as [] or as • . Notice that these symbols are used in place of K]~.
Although from a technical viewpoint, if, say [] = []0 and • = []1, then [] and
[] are the same, we wish to make notation independent from an accidental
choice of interpretation for [] and • . Seen this way, we are now dealing with
three independent operators, [], [] and • .
Define two translations, z~ and ~ in the following way.
6o) := p -~@) := ;
"rl::](T) := T 7-11(T) := T
~'D( ~ ) := -:~-~(~) ~-"(~) := -~-"(@
~'r~(~ A X) := ~ ( ~ ) A ~-E:(X) ~ ( ~ A X) := ~-"(~) A ~ ( X )
~-~([]@ := []TD(~) ~-"([]~) := •~-"(X)
We speak of the translation ~-II as the dual of 7I:1, by which we want to imply
that the roles of • and [] are interchanged. (So, w[] likewise is the dual of ~'l-)
It is in this sense that we want to be understood when we talk a b o u t duality
in this chapter. This will frequently arise in proofs, where we will perform the
argument with one operator, and omit the case of the other operator. Given
two modal logics, A and (9, the f u s i o n is defined as in Section 2.5 by
A @ (9 := K2 69 ~-a [A] 69 Tm[®]

This defines an operation - @ - : (EK1) 2 -+ ~K2. @ is a U - h o m o m o r p h i s m
in b o t h arguments. Moreover, it is easy to see that
( E l 69 X) ® ( E l 69 Y) = K2 69 7]:][X] 69 Tm[Y]
(Namely, observe that ~'D and Tm translate valid derivations in K1 into valid
derivations in K2. So, if A derivable from X by (rap.), (mn.) and substitution
in K I , all formulae of Tls[A] are derivable from T[][X] by means of (rap.),
(mn.) and substitution in K2.) We call a bimodal logic E i n d e p e n d e n t l y
a x i o m a t i z a b l e if there exist A and (9 such that E = A ® (9. The following
theorem will be made frequent use of.
LEMMA 6.2.1. {9, q, 4, G} ~ A @ (9 iff {g, q, G} ~ A and (g, 4, G} ~ (9.
The easy proof is left to the reader. Moreover, it should be clear t h a t if

{g, q, 4, G} is a bimodal frame, (g, q, G} and (9, 4, G) are monomodal frames.
6.2. Some Preliminary Results 279
Given a bimodal logic A define

AD := ~I[A]
Am := r~l[A]
There are certain easy properties of these maps which are noteworthy. Fixing
the second argument we can study the m a p - @ (3 : 8K1 -+ 8 K : . This is a
L.]-homomorphism. The m a p - [ ] : 8K= + 8K1 : A ~-~ A[] will be shown to
be almost the inverse of - ® (3.
LEMMA 6.2.2. (1.) Let A be a normal modal logic. Then (A ® (3)[] _D A.
(2.) Let E be a normal bimodal logic E. Then ED @Em C_ E. Moreover, ~ is
independently axiomatizabIe iff ~ = (:a[2) @ (~:m).
PROOF. (g, <~, • , G } > E implies (g, <,G} ~ ED and (g, • , G } ~ 7~1. This
implies in turn t h a t (g, <~, • , G } ~ ~:~ @ .~.m. Consequently, if E is indepen-
dently axiomatizable then (g, <, • , G) is a general m-frame iff {g, <l, G} is a
general Eel-frame and {g, • , (;} is a general ~m-frame. []
Given a monomodal frame ¢5 := (g, ~<, G), put ~5° := {9, <1, • , G}, where
:= ~<, and • := O, and put 6 ° := {g,<, • , G } where we have <l := ~ and
• := {(x,x} : x E g}- It is easy to check that both ~5° and ~3° are bimodal
frames. (To see that, one only has to verify t h a t for b C G also " b E G; but
this is straightforward.) The following was shown in [211].
THEOREM 6.2.3 (Thomason). (A ® (3)[] = A iff 5_ ~ (3 or 5_ C A.
PROOF. ( ~ ) Suppose 5_ E (3 and 5_ ~ A. Then 5_ C A ® (3 and hence
5_ E (A ® (3)D, so hat A ¢ (A ® (3)~.
( ~ ) Suppose 5_ C A. Then 5_ E A ® (3 and so 5_ E (A ® (3)~ from which
A = (A @ (3)D. Now suppose 5_ ¢ A. Then 5_ ~ (3 and hence either [ ] ~ (3
or [ ] > (3. Let q5 -- (g,<~,G;} be a A-frame. Then put ~5* := ( g , < l , • , G }
as above with • := ~ and G ° -- ( g , < , • , G } with • := { ( x , x } : x C g}. If
[ ] > (3 then 6 ° is a A ® (3-frame and if [ ] ~ (3 then ~5° is a A ® (3-frame.
For WC A~ we have ® ~ ~¢:~ ~5" V ~ ¢ * ~ 5 ° ~ . Thus (A®(3)[] _C A a n d
therefore (A @ (3)~ = A. []
The theorem states that if 5_ ~ A or 5- ¢ (3 then A @ (3 is a conservative
extension of A. Thus given two logics A, (3 we have b o t h A = (A @ (3)t5 and
(3 = (A ® (3)~ iff 5- ~ A ¢:> 5- ~ ~. In all the theorems t h a t will follow the
case t h a t 5- ~ A or 5_ E (9 will be excluded. These cases are trivial anyway,
so nothing is lost. The way in which Makinson's theorem has been used to
build a minimal extension of a m o n o - m o d a l frame to a bimodal frame is worth
remembering. It will occur quite often later on. Although Makinson's theo-
rem has no analogue for bimodal logics as there are infinitely m a n y maximal
consistent bimodal logics, at least for independently axiomatizable logics the
following holds.
COROLLARY 6.2.4. Suppose that A is a consistent, independently axiom-

atizable bimodal logic. Then there is a A-frame based on one point.
THEOREM 6.2.5. Suppose that ± ff A, O. Then A ® (9 is finitely axioma-
tizable (recursiveIy axiomatizabIe) iff both A and (9 are.
PROOF. If A and (9 are recursively axiomatizable, so is clearly their fu-
sion. And if the fusion is, then the theorems are recursively enumerable, and
hence also (A ® (9)[] and (A ® (9)1. Thus A and (9 are recursively axiomati-
zable. Now for finite axiomatizability. Only the direction from left to right is
not straightforward. Assume therefore t h a t A ® (9 is finitely axiomatizable,
say A ® ( 9 = K 2 ( Z ) . Let Z and Y be such that A = K I ( X ) , (9 = K I ( Y ) .
Then Z C_ K2(T[][X] U zl[Y]). By the Compactness T h e o r e m we have fi-
nite sets X0 C_ X, Y0 C_ Y such that Z C_ K2(zI:?[X0] U T,,[Y0]). But then
A ® (9 = K2(7~[Xo] t2 7l[Yo]) = K I ( X o ) ® KI(Yo) and hence A = K I ( X o )
and (9 ----KI(Yo). []
THEOREM 6.2.6. Suppose that ± ~ A, (9. Then A ® (9 is r-persistent iff
both A and (9 are.
PROOF. ( ~ ) Suppose that both A and (9 are r-persistent. Further as-
sume t h a t ( g , < ~ , , , G ) ~ A ® ( 9 . Then (g,<~,G) ~ A a n d (g,~t,G) ~ (9. By
assumption, (g, 4) ~ A and (g, a ) ~ (9 and so (g, 4, a ) ~ A®(9. ( 3 ) Suppose
that £ ¢ A and that A is not r-persistent. We have to show t h a t A ® (9 is also
not r-persistent. We know that there is a A-frame ~5 = (g, <~, G) such t h a t
(g, <~) ~ A. On the condition that ®" and ~5° are both refined the theorem is
proved. For either ®" ~ A ® ( 9 or ®° ~ A ® ( 9 , but (g,<~,4) I~ A ® ( 9 since
(g,<) h.
Both ®" and ¢5° are differentiated and tight. T h a t ~5° satisfies tightness
for • is seen as follows. If x -- y then for all c G G, x E • c implies x C c since
• c----c. But i f x # y t h e r e i s a c G G s u c h t h a t xEc, y¢c. ThenxE •c,
y ~ c, as required. Similarly, ~5" satisfies tightness for • since for arbitrary
x, y there is c C G; with y ~ c. Moreover, x C • c , since • c = 1. []
THEOREM 6.2.7. Suppose that £ ~ A, (9. Then A ® (9 is d-persistent iff
both A and (9 are.
PROOF. As in the previous theorem. One only has to check t h a t if ~5
is descriptive, so are ®" and ~5°. We have seen that if ~5" and ~5° are b o t h
refined, so we only have to check compactness. So let U be an ultrafilter ®"
(~5°). Then U is also an ultrafilter of ®, since the underlying boolean algebras
are identical. Since ~5 is compact we have N U 7~ 2~, as required. []
COROLLARY 6.2.8. Suppose that £ ~ A, (9. Then A ® (9 is canonical iff
both A and (9 are.
PROOF. By Theorem 4.8.6. []
6.3. The Fundamental Construction 281
E x e r c i s e 198. Let O be a consistent logic. Define e : ~ K1 -+ £ K2 by

A ~+ A ® ® and let r : £ K 2 --+ EK1 be defined b y F ~ - + F [ ] . B o t h e a n d r
are not necessarily lattice homomorphisms. However, b o t h are isotonic (prove
this). Show that e is left adjoined to r. T h a t is, for monomodal logics A and
bimodal logics
e(h) C r ¢> a c_ T(F)
It follows t h a t ere(A) = e(A) as well as rer(P) = r(P). Moreover, show t h a t
e preserves infinite joins, while r preserves infinite meets. In b o t h cases, give
a direct proof and a proof using the adjoinedness of the maps.
E x e r c i s e 199. Show that d is a left adjoint of r, where d : ~ K1 -+ ~ K2 is

defined by d : A ~+ A @ A.
E x e r c i s e 200. Let t, the twist map, be defined as follows.

pt :___ p
(-~)t := ~t
(~Af)t :__ ~tA¢,

(D~) t := tt~ t
("~)' := [3~'
For a logic let A t := {~* : T E A}. Show that A t is again a logic, t h a t Att = A
and t h a t ( _ ) t : E K2 --+ £ K2 is an automorphism of the lattice of bimodal
logics.
E x e r c i s e 201. (Continuing the previous exercise.) Show that K 2 ( X ) t =

K 2 ( X t ) . Hence show t h a t for all bimodal logics F, A :-- F U F t is the smallest
logic containing P such that A t = A. Show also that i f F = A ® O then
Fur' : (A~e) ® (hu e).
E x e r c i s e 202. This exercise is given to motivate the notation of fusion as

@. Show namely t h a t @ distributes over arbitrary joins, thus behaves like a
meet operator. Show also that it distributes over arbitary meets!
E x e r c i s e 203. Show that the lattice ~ K2 @ ([~p ++ rap) is isomorphic to
6.3. T h e F u n d a m e n t a l Construction
In this section we will prove Theorem 6.3.6. It says that a consistent
bimodal logic A ® O is complete with respect to atomic frames iff both A
and ® are complete with respect to atomic frames. The proof is a successive
construction of a model, and it allows similar results concerning completeness
and finite model property. It allows to reduce the decision procedure in the
bimodal logic A ® O to a decision procedure in the logic A (or ®) given an

A ® O-oracle for formulae of smaller complexity. However, it is conditional
on the completeness of A and ® with respect to atomic frames. A proof of
this fact without this assumption will be proved in the next section.
For a proper understanding of the method some terminology needs to be
introduced. For each formula []¢, I1¢ E L u • we reserve a variable q[]¢ and
q • ¢ respectively, which we call the s u r r o g a t e of [3¢ (11¢). q[2¢ is called a
[ ~ - s u r r o g a t e and q•¢ a " - s u r r o g a t e . We assume that the set of surrogate
variables is distinct from our original set of variables. Any variable which is
not a surrogate is called a p - v a r i a b l e and every formula composed exclusively
from p-variables a p - f o r m u l a . A p-variable is denoted by P, P l , . . . , P i , . . .
and an arbitrary variable by q. Finally, if ~ is a formula, then varP(qo) denotes
the set of p-variables of T, and likewise the yarD(T), v a r • ( ~ ) denote the set
of [~-surrogates of qo and the set of " - s u r r o g a t e s . The set of p-variables in
L ~ • is assumed to be countably infinite.
DEFINITION 6.3.1. For a p-formula qo we define the [ ~ - e r s a t z qo[3 ~ L ~

as follows.
qU := q
A := A
([q~)r7 := [q~m
(llqo) [] :-- q•~
Notice that the ersatz of qo is computed outside in and not inside out,
which is typical for an inductive definition. For a set F of p - f o r m u l a e call
F t3 := {qoO : T E F} the IN-ersatz of F. Dually for II.
Now let T be composed either without I - s u r r o g a t e s or without [~-surro-
gates. T h e n we define the r e c o n s t r u c t i o n of qo, Sqo, as follows.
:=
top := ~(DqODO/qE]~o, .." , [-]V~-I/qE]~e_I,Po,'' " ,Pm-1)

Note that if 1`is defined on ~ it is also defined on I"¢; for if ~ was free of I I -
surrogates, 1`~ is free of [q-surrogates and vice versa. Now if ~ is a p - f o r m u l a
then ~D is free of rq-surrogates and therefore the reconstruction operator is
defined on ~. Also, if 1" is defined on ~ then for some n C w, 1`n+1~ =i.nqo
(where 1"n denotes the n t h iteration of 1") which is the case exactly if 1"n~ is
a p-formula. We then call 1`n~ the t o t a l r e c o n s t r u c t i o n of ~ and denote
it by qot. ~i" results from ~ by replacing each occurrence of a surrogate qx in
by X. Now let ~ be a p-formula. Then we put qo~ :=1`n(~[]). It is clear
that ( q g ) t = ~. The [ ~ - a l t e r n a t i o n - d e p t h of ~ - - adp~(~) - - is defined
by adp~(~) = m i n { n : T~ = ~}. For m > adpD(~), qOm = qo,~-l. T h e
m - a l t e r n a t i o n d e p t h , a d p " ( ~ ) is defined dually. Finally

adp(~) := (adp[](7~) + adp"(7~))/2 .
It is easy to show that ladpD(~) - adp•(~)t < 1. For example, if 7~ E LD
then adp[](T) = 0 and adp"(~) = 1 and so adp(7~) = 1/2. Conversely,
adp(~) = 1/2 implies ~ E LD U L,,. Finally, let us define the [ ] - d e p t h of a
p-formula T, dp[](~), as follows.
@E:(p) := o
@:~(m) := o
@s(~) := @s@)
@~(~A¢) := ma~{@D(@,@D(¢)}
dpC(~cp) := 1 -t- dpS(~p)
@~(m~) := @(@
Dually the m - d e p t h is defined.
DEFINITION 6.3.2. Let A be a (bimodal) logic and A C L[2• be a finite
set. For a set A C A let
:= A<x: x A> A A< x: x e A>.
The consistency s e t of A, C(~), is defined by
C(A) := {CA : A C_ A, ¢A is A-consistent}.
The consistency f o r m u l a E(A) of A (with respect to A) is defined by
r~(A) := V < x : x • c(~)>.
If A is an infinite set then we define
C(A) := U<C(A'): a ' C A,A' finite>
s(z~) := {r~(z~,): z~' c_ A, A' finite}
Note that the consistency formulae are A-theorems. T h e y depend of
course on A, but we write E(A) rather than EA(A). We abbreviate the
consistency formula for the set s f { ¢ : q¢ E var(~[5)} U varP(T) by ED(~).
THEOREM 6.3.3. Let A and ® be complete with respect to atomic frames.
Then A @ Q) is also complete with respect to atomic frames.
PROOF. In the proof of Theorem 6.3.3 we construct not ordinary mod-
els but partial models. If ¢~ is a frame and V a set of variables then /3 :
Y --+ {0,1,*} g is called a p a r t i a l v a l u a t i o n if/3-1(0), /3-1(1) a n d / 3 - 1 ( , )
are internal. Here, 0, 1 are called the s t a n d a r d t r u t h values and * is the
u n d e f i n e d or - - to avoid confusion - - the n o n s t a n d a r d t r u t h value. We
define the value of a formula according to the three-valued logic of 'inherent
undefinedness' (or Weak Kleene Logic). It has the following t r u t h tables
A 0 1 •
0 1 0 0 0 *
1 0 1 0 1 •
We put
~(~,x) := ~(~,x)
~ ( ~ A ¢, x) := ~(~, x) A ~ ( ¢ , ~)
~([]~,~) := A ( ~ ( ~ , y ) : ~ < y)
Note t h a t by definition E]~ and O~ receive a standard t r u t h value iff every
successor receives a standard truth value. We define the following order on
the t r u t h values
0 1
\/
In the sequel we will assume t h a t all valuations are defined on the entire set
of variables. In contrast to what is normally considered a partiM valuation,
namely a partial function from the set of variables, the source of partiality or
undefinedness is twofold. It may be local, when a variable or formula fails to be
standard at a single world, or global, when a variable or formula is n o n s t a n d a r d
throughout a frame. Our proof relies crucially on the ability to allow for local
partiality. The d o m a i n of a valuation/3 : V -4 (0, 1, *)g is the set of variables
on which/3 is not globally partial i. e. dora(~3) := {q: (3x • g)/3(q, x) ¢ *}.
If/3, 3' : V -4 {0, 1, *}g we define /3 _< 3" if/3(p, x) _< 7(P, x) for all p • V
and all x • g. It is easy to see that if/3 ~ 3" then for all x • g and all
with var(T) C_ V: -~(T,x) <_ 7(T,x). Hence if/3 and 3" are comparable
then they assign equal standard t r u t h values to formulae to which they b o t h
assign a standard t r u t h value. In the proof we will only have the situation
where a partial valuation/3 is nonstandard either on all [Z-surrogates or on
all " - s u r r o g a t e s . In the latter case we define for a point x • g and a set A of
formulae
:: {¢: ¢ • : 1}
u ¢ • : 0}
and call X~'~(x) the c h a r a c t e r i s t i c s e t of x in (g,/3). If X~'~(x)is finite

(for example, if A is finite), then X~'m (x) := A x ~ 'A (x) is the c h a r a c t e r i s t i c
f o r m u l a of x. And dually X~-'A (x) and X~'A(y) are defined. We call a set A
s f - f o u n d e d if for all X C A and ~- E sf(x) then either T E A or ~ - C A.
Now we begin the actual proof of the theorem. Assume I/[z• ~ and
adpm(~) = n. Let
Si :-- sf{%b: q¢ e var(~i)} U varP(v)
For i = 0 this is exactly the set of formulae on which the consistency formula
for ~o is defined. We will use an inductive construction to get a A (9 O - f r a m e
for ~. We will build a sequence ((~i,/3i,wo) : i • w} of models, which will
be stationary for i > adpm(T). The construction of the models shall satisfy
the following conditions, which we spell out for i = 2k; for odd indices the
conditions are dual.
[a]2k <®2k,fl2k,W0> h ~O2k
[b]2k dom(/32k) = var(SE2k)
[C]2k (g2k, < 2 k , G2k) : (g2k-2, <~2k-2, (~2k-2) @ J~ for some 59, and
~12k : 42k-1
[d]2k ~2k is an atomic frame and ~2k b A
[e]2k For x • g2k-l:
(1) /32k(p,x) =/32k-l(p,x), p • var(~)
(2) /32k(q•¢,x) </32k-l('¢m,x), q•¢ • var(Sm2k)
(3) /3 k-l(qD¢,x) _</32k([]¢[],x), q[]¢ • var(S k_l)
[f]2k X2k(x) := X~--~k'S~k(X) is A ® O-consistent and sf-founded for
x • 92k -- 9 2 k - I
We begin the construction as follows. Let ~o be given; put 6 := dpm(~o). Since
~o is A (9 O-consistent, so is ~<~Em(~o) E] A ~o[]. For Ec](~o) is a theorem of
A ( 9 0 . Afortiori, [Z-<~E~(~o) ~ A p ~ is A-consistent and has a model
such t h a t ( g o , < o , G o ) i s atomic and dom(7o ) = var(gDo). We put *'o := £%

this way, ~5o := (go, <~o, 4o, Go} is a bimodal frame. We m a y assume that
the frame is rooted at wo. Pltrthermore, we may assume that the valuation
is nowhere partial for q • dom(7o) , that is, 7o(q,x) # * for all x and all
q. Let now Io be the set of points for which X~ ° (x) is A ® O-inconsistent;
and let Ik := @ I o - (}<k-llo for all k < 5. Finally, put I~ := go - [.Jj<~ Ij.
All sets Ie are internal. (To see that, notice that the set of points with a
given consistency set is internal; Io is a finite union of such sets.) We define a
partial valuation/3o _< 7o as follows. /3o(q¢, x) := * iff for some k, x • I~ and
dp[~(¢) > k. Since Wo ~ I~, p ~ is defined at Wo and since/30 _< 70
(eo,/3o, o> (=
Therefore, [a]o, [b]o and [d]o hold. [C]o and [e]o are void, there is nothing to
show. For [f]o note that X°(x) C_ XeO(x); and the latter is consistent. To
286 6. R e d u c i n g Polymodal Logic to Monomodal Logic
show t h a t X°(x) is sf-founded notice that: (i) So is sf-founded; (ii) X ~ X°(x)

iff/3(X, x) 76 *, by definition. A n d so it is enough to show t h a t
($) I f / 3 ( x , x) # * and T ~ sf(X) then also/30(% x) # * .
This is, however, immediate; for if ~0(T, x) = * then there exists a k such t h a t
x ~ I~ and dp~(~ -) > k. From this follows dpm(X) > k, and so ~0(X,x) = *
as well.
T h e inductive step is done only for the case i = 2k > 0. For odd i the
construction is dual. Assume [a]2~-[f]~. For every point y ~ g2~ - g2~-i we
build a m o d e l
((hy, ,qv, H~), 7y, Y> ~ md~" (z~(t))Em(X2~(t))"; X2~(t) "
based on an atomic O - f r a m e ~ y := (by, 4y,]H~), where X2k(y) := X~2k's2~(y).

This is possible since all the characteristic formulae are A ® O - c o n s i s t e n t and
so their " - e r s a t z is O-consistent. We assume t h a t hy N hy, = ~ for y ~ y~,
hy N g2k = {Y} and hv = Tr"(y, hy). Now call x and y C - e q u i v a l e n t if
x2k(x) = X2k(y). We will assume t h a t if x and y are C-equivalent there is an
i s o m o r p h i s m L~y from the model (Y)~,7~, x> to the model ( ~ y , T y , y ) . (This
m e a n s t h a t L~y is an isomorphism of the frames, t h a t 7~(q, u) = ~y(q, L~y(u))
for all u E h~ and t h a t L~y(X) = y.) Finally, the following c o m p o s i t i o n laws
--1
shall hold for all x , y , z C g2k - g 2 k - l : ~z~ = id, ~y~ -- L~y, L~ -- L~y o Lye.
This is always possible. For let Y be the set of all points C - e q u i v a l e n t to a
given point Y0. For every y E Y let ny be an isomorphism from (Y)yo,Tyo,Y0}
o n t o (Y3y, Ty,Y)" T h e n p u t L~y := ny o ~i.
In case t h a t dp"(X2k(y)) = 0 we set in particular
hy :-- {y}
] {<y,y>} if[] o
4v
[ O otherwise
Clearly t h e n ~2k(q,Y) = 7y(q,Y) for q C var(F~12k). As before, Io is the set

of z such t h a t X~(z) is inconsistent. Let 6 := dpm(xT~(z)) a n d k < 6.
T h e n Ik := OkIo - 0-<k-110, I~ := hy - Uj<~Ij. We p u t ~y(q,x) :-- *
for q ~ var(X2k(y) I) and ~y(q[]¢,x) = * if there is a k such t h a t x E Ik
and dpl(¢) > k. In all other cases ~y(q,x) := 7y(q,x). Clearly, ~y <
7y. Now observe t h a t var(Sm2k) = var(F~12k+l) and therefore var(x2k(y) m) C_
var(Sl2k+i). We can conclude t h a t (1) X2k(y) " is defined at y in (2)y,3y) and
therefore ((hv, a v , Hy), 3y, y) ~ X 2k (y)m and t h a t (2) X ~ (x) is consistent and
sf-founded (using (:~)). Now let
g2k+l := g2k U U(hy : y E g2k - 92k-1)

"(]2k+l := <:]2k
~2k+1 := 42k U U ( ~ y : Y E g2k -- gek-1}
For the internal sets, some care is needed. P u t h := [.J(hy : y E g2k--g2k-1}. A
subset T C_ h is called h o m o g e n e o u s with s u p p o r t a if (1.) a C_ g2k - g 2 k - 1 ,
a E G2k, (2.) T = (-JxeaT A hx, (3.) all x , y E a are C-equivalent and
t~y ( T N hx) = T • hy. A set T is called h o m o g e n e o u s if there exists an a such
that T is homogeneous with support a. T is called s e m i h o m o g e n e o u s if it
is a finite union of homogeneous sets (not necessarily with identical support).
G2k+l :: {SUT : S C G 2 k - I , T semihomogeneous}

q52k+l := (g2k+l, <12k+l, 42k+l,G2k+l)
We have to show that ®2k+1 is a (bimodal) frame, q32k+l is clearly closed
under finite unions. To see that it is closed under complements it is enough
to show that the relative complement of a homogeneous set T in h, h - T, is
semihomogeneous, h - T is the union of the sets uy : : h y - T , y E g2k - g 2 k - 1 .
Let a be the support of T, U : : UycaUy and V := [..Jy~-aUy" T h e n U
is homogeneous with support a, and V is homogeneous with support - a .
Moreover, h - T = U t] V, and so the complement of T is semihomogeneous.
Furthermore, we have to show that the internal sets are closed under <) and ¢.
First 0. Let S C G2k+l. Then 0 S O ( S N g 2 k ) . g2k is the union ofg2k-1 and
:
g2k - g2k-1, g2k-1 is internal, g2k - g2k-1 is semihomogeneous with support

g2k --g2k-1, for by construction, all frames ~y are atomic, and so in particular
the set {y} is internal in ~)y. Hence 0 ( S ~ g2k) is an internal set of 152k and
so by the same argument an internal set of ®2k+1. Next, closure under 0
must be shown. Take a set S U T, where S E G2k-1 and T semihomogeneous.
Then O ( S U T ) = ( 0 S ) U (0T). By construction, 0 S C G2k-1. Moreover, it is
easy to see that if T ' is homogeneous with support a, so is 0 T ' . Thus, 0 T is
semihomogeneous, since T is. Finally, ¢~2k+1 is atomic. For if x E g2k-1 then
{x} is internal by [d]2k-1. If x ¢ g2k-~ then there exists a y such t h a t x C hy.
Then {x} is an internal set of ~)y, and since {y} is internal in ®2k, {x} is
homogeneous with support {y}. (Here we use the fact t h a t tyy(X) = x.)
Define /32k+1 as follows. P u t 132k+l(q,x) := /3y(q,x) for x C hy and
/32k+1(q,x) := /32k-l(q,z) for x E g 2 ~ - i , q C var(Slek+l); in all other cases
~2~+~(q, x ) : = *. By construction, [b]~+l holds. [c]2k+1 holds because
( ~ + 1 , ,*~+1, G~+I ) = (g2~-1, ~ k - ~ , O~-l) • s) ;
moreover, .~ is based on the disjoint union of the (h~, ~ ) , taking semihomo-
geneous sets as internal sets. And <~2~+1 = <~2~. [d]2k+~ is immediate from
the observation above t h a t the frame is atomic, and from [C]2k+~, [d]2k-1 and
288 6. R e d u c i n g P o l y m o d a l L o g i c t o M o n o m o d a l L o g i c
.~y ~ O. Now we show [e]2k+l.

A d (1). Let x C g2k-1. Then by [e]2k, /~2k(P,X) ---- ~ 2 k - I ( P , X ) =- ]~2k+i(p,x).
If X E g2k -- g2k-1 then
;~2k+1 (p, z) = 1
¢, Z~(p,x)= 1
.,=~t p E X 2 k ( x )
** 9~k(p,~) = 1
Here, ¢:~* is true since X2k(x) is sf-founded and dom(/~x) = v a r ( X 2 k ( x ) ) .
Similarly, fl2k+l(P, x) = 0 ~ f~2k(P, x) = 0 is shown.
A d (2). If z C g2k-1 we have fl2k+l(qc?¢,x) = ~2k-l(q[]¢,x) < fl2k([Z¢[5,x),
by [e]2k. If x E g2k - gek-1 then
32k+l(q~¢, x) = 1
¢* 9~(qD,,x) = 1
,, []¢ c x ~ k ( z )
,, Z~k([]¢ D, x) = 1
The argument continues as in (1).
A d (3). If x E g2k-1 the claim follows by [e]2k. Now let x C g2k - g2k-1- If
~ 2 k ( q l ¢ , x ) = * then there is nothing to show. However, if f~2k(qm¢, x) ~ *
then m%b C X2k(x) or ~m¢ E X2k(x) and thus
~2k+1 ( " ¢ ' , ~) = 1
,, ~ x ( m ¢ m, ~) = 1
** I¢ ~ x2k(z)
*V ~2k(qm¢,X) = 1
[f]2k+~ holds because of [c]2k+1 and by the definition of fl2k+1 and finally
because of (2) of [e]2k+l. [a]2k+l follows directly from [e]2k+l (1) and (3).
If n = adpC~(~o) we have g~+l = g,~ and dpm(xn(y)) = dprn(X~(y)) = 0
for all y since Sn = var(y)) and therefore dom(~n) = var(~), by [b]n. By
construction, the ~y are based on a single point and thus g~+l is based on
the same points as g~. Moreover, by [d]~ and [d]~+~, ®~+~ ~ A ® ® and by
[a]n+l, (~n+l,/3n+1, W0} ~ ~n+l (--~ ~)" Take any valuation 3' -> fln+l which
is standard for the p-variables. Then (~n+l,V, W0} ~ ~. []
COROLLARY 6.3.4. Suppose that A®® are complete with respect to atomic
frames. Let ~ be a bimodal formula, m >_ dpD(p), and n >_ d p " ( T ) . Then
the following are equivalent.
1. F-F~.. T .
2. ~-D [~-<mED(q°)D -+ ~ .
3. ~-.. I < - ~ E . . ( ~ ) " --+ ~ " .
6.3. T h e F u n d a m e n t a l C o n s t r u c t i o n 289
T h e p r o o f of the latter t h e o r e m is evident from the c o n s t r u c t i o n used in the

previous proof.
THEOREM 6.3.5. Suppose that A and ® are complete with respect to atomic
frames. Then A ® 0 is decidable iff both A and (9 are decidable.
PROOF. By induction on n :-- adp(p). If n = 0, p is b o o l e a n and since
± ¢ A ® O F[] • ~ iff p is a boolean tautology. Since the propositional calculus
is decidable, this case is settled. Now suppose t h a t for all ¢ with adp(¢) < n
we have shown the decidability of F-D • ¢. We know by T h e o r e m 6.3.4 t h a t
for m >_ max{ dp[~(~), d p • ( ~ ) }
l-r~• ~ ~ ~-n [3-<mE[](~) [] --4 ~[]
Therefore we can decide F~ • ~ on the condition that either ~D (~9) or ~ i (~)

can be constructed. B u t now either adp(r D( )) < n or <
This is seen as follows. Suppose t h a t adp[](~) ~ adp•(~). T h e n there is a
m a x i m a l chain of nested alternating modal•ties starting with [3. T h e n any
m a x i m a l chain of nested alternating modal•ties in E[] (p) starts with • (!) and
is a subchain of of such a chain in ~. Consequently, we have adp•(E[:](~,)) <
adpi(p) and with adpD(ED(~)) <_ adp[~(~) the claim follows. Now assume
t h a t adp(E[3(T)) < adp(~) is the case. T h e n
Consequently, E[z(T) can be constructed if only ~-D • 9¢c is decidable for all
c. But this is so because adp(~¢c) < n. []
Note t h a t for m > 1 we have adp~([~<-mED(~)) < adp~(~). However,
adp"([::]<-mED(p)) < adp"(~) + 1.

A case where the inequalities are sharp is given by p = • p . But in all these
cases adp~(~) > adp•(p) in which case we also have
<
and
adpi(E•(~)) < adp~(i<-mE•(~)) ~_ adp[~(p)
and therefore adp(•<mE•(~)) <_ adp(p).
COROLLAaY 6.3.6. Suppose L ~ A, O. Then A ® (9 is complete iff both A

and 0 are complete.
COROLLARY 6.3.7. Suppose that ± ¢ A , ® . Then A ® 0 has the finite
model property iff both A and ® have the finite model property.
N o t e s on this section. EDITH SPAAN [202] has investigated the complex-

ity of the fusion of modal logics. If A and 0 are b o t h C-hard, so is A ® O.
If A and 0 are both in PSPACE, then so is A ® O. The fusion m a y however
become P S P C A E - h a r d even if the individual factors are in NP. For example,
JOSEPH HALPERN and Y. MOSES [95] have shown that $5 ® $5 is P S P A C E
hard. However, S5 is itself in NP by a result of R. LADNER [137]. (See also
the exercises of Section 3.1.)
E x e r c i s e 204. Show with a specific example that there are m o n o m o d a l log-

ics A and O which are consistent and tabular such that A ® O is not tabular.
Show furthermore that A ® O has the finite model property and that it is
tabular iff one of A, O is of codimension i.
* E x e r c i s e 205. (WOLTER[237].) As the previous exercise has shown, there

are logics whose lattice of extension is finite, yet the lattice of extensions of
their fusion is infinite. Now show t h a t the lattice ~ K . T is isomorphic to the
lattice ~ ( K . a l t l ® $5). As we will see, each of the lattices ~ K . a l t l and ~ S5
is countable, but ~ ( K @ p --~ (}p) is uncountable.
* E x e r c i s e 206. Let A = K.alt1.E]3±. Show that ~ ( A ® A ) has 2 ~° elements.
* E x e r c i s e 207. The following three exercises will establish t h a t the re-

quirement of completeness for atomic frames can be lifted in the Consistency
Reduction Theorem when we consider only extensions of K 4 . The next sec-
tion will establish this for all logics but with a different technique. Call a
K 4 - f r a m e ~ s e p a r a b l e if for all x E f there exists a a C F such t h a t x <] y
and y C a implies y <] x. Show that every extension of K 4 is complete with
respect to separable frames. Hint. Take the canonical frame ~anA(n). Call
W eIiminable if for all formulae ~ such that W E ~ there exists a V E
such that W <~ V ~ W. Show that if W is eliminable and ~ E W then there
exists a noneliminable V such that W ,~ V and p C V. Now take the set N
of noneliminable points and let 9] be the frame based on N. (Since N is not
internal, this is not a subframe.) Show that Th 9~ = Th ~anA(n). (The t e r m
separable is taken from RAUTENBERC [169]. The completeness result is from
FINE
E x e r c i s e 208. (Continuing the previous exercise.) Call a frame ~ h o o k e d

if it is rooted, and that for the root w0 the set {w0} is internal. Show t h a t
the results of this section can be generalized to logics which are complete with
respect to hooked frames.
E x e r c i s e 209. (Continuing the previous exercise.) Show t h a t all extensions

of K 4 are complete with respect to hooked frames.
6.4. A General T h e o r e m for Consistency Reduction 291
6.4. A G e n e r a l T h e o r e m for C o n s i s t e n c y Reduction

T h e results and proofs of this section are from FRANK WOLTER [243].
T h e y make use of another type of completeness for logics, n a m e l y with re-
spect to algebras which are atomless. For simplicity we assume t h a t a < ~0
t h r o u g h o u t this section. Recall t h a t an element x of a b o o l e a n algebra is an
a t o m if for all y such t h a t 0 < y ~ x we have y -- x. A boolean algebra is
called a t o m l e s s if it has no atoms. A m o d a l algebra is called atomless, if
its boolean reduct is atomless. A n atomless algebra is either isomorphic to
1 or infinite. T h e following is a well-known fact a b o u t b o o l e a n algebras (see
[117]).
PROPOSITION 6.4.1. Let 91 and ~ be two countably infinite atomless boole-

an algebras. Then 2[ ~ ~ .
Moreover, let 91 be atomless and countable and a > 0. T h e n let 91a be

the algebra based on a]l sets b _< a, with the operations being intersection
and relative complement. T h e n 91a is also countable and atomless, and 91~ is
not isomorphic to 1. Hence, it is countably infinite and so by the previous
t h e o r e m isomorphic to P2.
DEFINITION 6.4.2. Let 91 be a boolean algebra. A family {ai : i E n} is a

p a r t i t i o n of 91 if (1.) ai • 0 for all i < n, (2.) ai n aj - ~ 0 if i ¢ j and (3.)
U(a : i < = 1.
LEMMA 6.4.3. Let 91 be a countably infinite atomless boolean algebra and

{ai : i < n} be a partition of 91. Then ~ : 91 -+ [L<n ~ defined by ~ ( x ) : =
( x N a i : i < n} is an isomorphism. Moreover, if {ai : i < n} is a partition
of P2 and {bi : i < n} a partition of ff~, then there exists an isomorphism
7- : 91 --+ ~ such that T(ai) = bi.
PROOF. T h e proof of the first claim is as follows. T h e m a p x ~+ x N a is a

h o m o m o r p h i s m from P2 onto 91~. Hence, a defined above is a h o m o m o r p h i s m .
Now suppose t h a t a ( x ) -- a(y). T h e n x N a i = yV~ai for a l l i < n. So we
have x A U i < n a i = yNUi<nai. Since U i < n a i = 1, x = y. Hence a is an
isomorphism. Now let {ai : i < n} be a p a r t i t i o n of P2, and {b~ : i < n} a
partition o f ~ . Let c~ : x ~ ( x A a i : i < n), and ~ : x ~+ ( x n b i : i < n). These
are isomorphisms. Furthermore, there exist i s o m o m o r p h i s m s hi : 91a~ -+ ~b~-
^-1
P u t v : = 1-L<~ hi. Finally, ~- := a o v o a. This is an isomorphism, and
= ov((o,...,o,l,o,...,o))
= ,o, 1 , o , . . . ,o))
---- b7
[]
292 6. R e d u c i n g P o l y m o d a l Logic to M o n o m o d a l Logic
PROPOSITION 6.4.4. (n < R1.) Let A be a n-modal logic and A an infinite

cardinal. Then ~rA(A ) is atomless. Moreover, ~A(R0) is countable.
If A is inconsistent then ~A(A) is finite and isomorphic to 1. In all other
cases it is infinite.
DEFINITION 6.4.5. Let A a n-modal logic. By Atg A we denote the set of
countably infinite atomless algebras 92 of AIg A.
The following is now immediate.
PROPOSITION 6.4.6. (n < ~0-) Let A be a consistent n-modal logic. Then
A = Th Atg A.
THEOREM 6.4.7 (Wolter, Global Consistency Reduction). (n < R~.) Let
A and 0 be consistent monomodal logics. Then the following are equivalent
i. ~ ~ID ¢
2. There exists A C_ C ~ ( ~ ; ¢ ) such that
(a) ~ ; (V Lx)~) ~ c ¢c
(b) for all X E A, ~[]; (V A) E ~zD ~X D and (V A) m ~zI -'X m
3. There exists A C_ CI(~;~#) such that
(a) ~I; (V A) m)~m cm
(b) for all X E A , TI; (V A) I ~ I ~X I and (V A) D J ~ ~x ~
If A satisfies 2., then ~; V A IF[3I ~ and V A ~F[~I ~X for all X E A .
PROOF. Obviously, it is enough to prove the equivalence of (1.) and (2.).
Moreover, (1.) implies (2.). For suppose that p ~F~I ¢. Then there exists a
bimodal algebra f13 and a valuation/3 such that ~(p) = 1 but ~(~) ~ 1. Let
A :-- {X E s f ~ ( ~ ; ¢ ) : ~(X) > 0}. Then A is as required for (2.). So, in the
remainder of the proof we will show that (2.) implies (1.). Assume therefore
that A C_ C ~ ( ~ ; ¢ ) exists satisfying the requirements under (2.). Then for
each X E A U {~b} there exists a ~ x ~ AtgA and a valuation/3 x such that
~ ( ( V z~)[] ~ ~[]) = 1 and ~(x[]) > 0.
Let
and
/3D := 11</3~ : x ~ ~ u {7¢})
Then ~ E AtgA, ~[](~G) = I, ~[](¢[]) ~ 1 and the set
{~D(x[]) : x c z~}
is a partition of ~. Since (V A) m ~i ~X m for a]l X E A we get in a similar
way a ¢ E Atg (~ and a valuation/3 m such that the set
6.4. A General Theorem for Consistency Reduction 293
is a partition of ¢. By Proposition 6.4.3 there is an isomorphism cr from the

boolean reduct of ~ onto the boolean reduct of ~3 such that for all X E A
~(~m(x')) = ~ ( x ~ ) .
Via this isomorphism we define the following algebra
92 := (B, 1, - , n, Q', I ' )
where [2'a := [~a and m'a := c~-l(mcr(a)). Clearly, 9.1 E Atg(A @ O). It holds
for all r] E sfC5(T; ¢) that
?s(@) = v(~rq(xrq) : x E A , r / a conjunct of X)
V(~'(x') : x c ~ , ~ a conjunct of X)
~'(~')
We now define a valuation 7 on the p-variables of ~ and ¢ by
~(p) := 9D(p) (= Z'(;)) •
We claim that for all 77 E s f D ( ~ ; ¢ )
~(~) = ~=(@) (= ~ ' ( ~ ' ) ) .
The proof is by induction on the complexity of r]. The claim holds by construc-
tion for the p-variables. Moreover, the steps for ~ and A are straightforward.
Now let r1 = D0. Then
~(D0) = []W(0)
= D'~D(0 c)
= []'~[](0)
= ~[](S0)
Analogously, the case r] = II0 is deaIt with. With the claim being proved,
we have 5(~) --= ~[](y)[]) = 1 and 7(¢) = ~[](¢D) ¢ 1. And that had to be
shown. []
THEOREM 6.4.8 (Wolter, Local Consistency Reduction). (n < Ill.) Let A

and 0 be consistent normal rnonomodal logics. Let ~ be a birnodal formula,
and rn > dpB(~), n > dpn(~). Then the following are equivalent
1. F-i~ ~.
2. ~-[] [Z-<roSa(@ G -+ ~D.
3. ~m i-<~En(~) n ~ ~ l
LEMMA 6.4.9. Suppose that [~-<mE[](~)[] -+ T[] ¢ A. Then there exists
92 E AtgA, a valuation ~[~ and a sequence (ai : 0 < i < rn) such that the
following holds:
294 6. R e d u c i n g P o l y m o d a l Logic to M o n o m o d a l Logic
(al) ai+~ ~ ai ~ [~ai for all i < m .

(a2) ai < ~([]_<iE~(~)[3), for all 0 < i < m .
(a3) am N~[:](~ [3) ~ 0.
(a4) The set { ~ ( X n) h a m : X • E~(~)} is a partition ofP2~ m.
(a5) The sets ( - ~ ( X ~) N (a, - ai+~) : X • E~(~)}
are partitions of 94(~_~+~), for all i < m .
PROOF. There exists a ~3 • Atg A and a valuation 3, such that

T ( ~ a A ~-<~rG(~) D) > 0.
We put f o r 0 < i < m ,
Take for each i _ m an algebra Ki • Atg A and valuations 5i such that
is a partition of ¢~. (For example, ffl := ~ A ( X ) , where X := v a r ( E [ s ( ~ ) 5 ) ,

and 5~ : p ~+ ~ the natural valuation, as defined in Section 2.8.) P u t
:= ~ × [ [ ( ¢ ~ : i < .~>, Z ~ := ~ × I I { ~ : i < m}

and define the ai by
a0 := (bo,1,1,...,1}
al := (bl,0,1,...,1}
a,~ := (bm,0,0,...,0,1}.
Then the elements {ai: 0 < i < m} and the valuation/3 5 satisfy (al) - (a5).
A d ( a l ) . Dai = {b~+I,[Z0, D 0 , . . . ,D0, 1,... , 1}. Therefore we have a~ n
[]al ---- (bi+l,0,... , 0 , 1 , . . . ,1} (i times 0). Thus ai+l ~ ai N Da~.
A d (a2). By definition of 5i, 5i(E~(~) c) = 1. Therefore ~[~([3-<~E[z(~) [2) =
(bi,1,... ,1} = hi.
Ad (a3). am N~[](-'~ Is) = {~(-'~8),0,... ,0,5i(-~[S)} > 0.
A d (a4). Clearly, ~[~(X~) N am and ~ ( X 2~) N am are disjoint for X1 ¢ )/2,
and the sum of all of these sets is = a,~. Moreover, by choice of a,~ and the
definition of E[~(p), these elements are all distinct from 0.
A d (ah). Again, the members {5i()/[]) : X E Ec(~)} are pairwise distinct
and their sum is ai - ai+l. Also, by choice of the hi, none of the elements is
=0. []
LEMMA 6.4.10. Suppose that E]-<rnE~(p) ~ --+ ~D • A. T h e n there exists
a 94 • AtgA ® O such that there are valuations 13[~ and ~ " and a sequence
(hi: 0 < i < rn} satisfying ( a l ) - (ah) and
6.4. A General Theorem for Consistency Reduction 295
(a6) j ~ ( X []) n a0 = ~m(X") A ao, for all X C C(~o).

(a7) For all O < i < rn and all b c A, ai rq mb = m(a~ rq b).
PROOF. For each i _< rn take a %4 C Atg O and a valuation h'i such that
{Ti()(m) : • E C(~o)}
is a partition of !Bi. In addition, take an arbitrary ~ - 1 and an arbitrary
valuation 9'-1 on 1B_1. Put
:= 1 - I ( ~ : - 1 < i < m>
m := 1-[<~:--l<i<m>
Now take ~ E Atg A and a valuation 7 [] such that the conditions (al) -
(a5) are satisfied. P u t a-1 :-- 1~ - a0. Without loss of generality we may
assume that a-1 # 0. There exist boolean isomorphisms a.~ : ~ m --+ ~a~ and
ai : !Bi -e ~a~-~+l such that for all i < rn
errn(~m(X")) ----~:3(X0 ) N am and o-i(Tm(Xm)) = 7rq(X rq) ~ (ai - ai+l) •
Fix an arbitrary boolean isomorphism c~-1 : 1B_1 --+ ~ _ ~ . Now the map
defined by a := 1-I(ai : - 1 < i < rn} is an isomorphism from ~ to ~ . Using
this isomorphism we define 92 as follows.
92 := {D, 1, - , n, [Z', m,}
[~'a := [~a
m'a :---- ~-~(mz(a))
P u t / 3 " ( p ) := a(Tm(p)) and/3D(p) := 7re(p). Then (a6) holds by construc-
tion. (aT) holds by virtue of the fact that 9.1 is isomorphic to a product
of " - a l g e b r a s based on the relative boolean algebras 9.i~ and 9~1_~, for all
0<i<rn. []
P r o o f o f T h e o r e m 6.4.8. The equivalence between (1.) and (3.) is proved
in the same way as the equivalence between (1.) and (2.). So we perform
only the latter. Moreover, (2.) implies (1.) by definition of the consistency
formulae. So, let us assume that (2.) fails. We will show that (1.) fails as well.
Suppose therefore that [-q<'~E:~(~) n -~ ~oG ~ A, where rn >_ dp~(~). Take
C Atg A ® O and valuations/3 D a n d / 3 " satisfying (al) - (a7). Using (a4),
(a5) and (a6) and the definition of E[u(~o) it is shown that for all r/C s f ~ ( ~ )
a0 n~:~(rP) = a0 N U < ~ ( x [ ] ) : x E E~(~o),r/a conjunct of X>

= a0 ~q U < ~ " ( x " ) : x E E[](~o),r/a conjunct of X)
= a0 n ~'(r/')
We define a valuation/3 on 92 by putting for all p-variables p of
/3(p) :=/3~(p).
We claim that for all 0 < k < m and for all 7 E sfD(~) such that dp~Ol) < k:
ak V1~(7) ----ak V1~[]0/[]) .

The proof is by induction on 7- If 7 is a p-variable, the claim is obviously
correct. The inductive steps for -7 and A are straightforward. So, we are left
with the cases U = IN0 and 77 = . 0 .
Assume that 7 = [~0. By induction hypothesis, ak V/~(0) = ak V1~[] (0~).
So,
aknDakn (7) = a nDakn (e)
--- ak N Eak N [ ~ [ ] ( 0 D)
= ak n Dak n a(7
(This makes use of the following. If a N c = aNb, then [ZaNcN~c = K]bNcN[~c.)
Since ak+~ __ ak N Dak (by (al)), the claim now holds for 7- Next assume
that 7 = nn~. We know by induction hypothesis that ak N~(0) = ak N ~ " ( 0 i ' ) .
Therefore
• (ak N ~ ( 0 ) ) : n ( a k 71 ~ i ( 0 i ) ) .
Using (a7) we conclude
ak N i ~ ( 0 ) = a~ N i ~ i ' ( 0 i l ) .
This gives ak N~(7) = a~ N ~ii(0in). It follows that ak N--~(0) = a~ N - ~ ( 0 ~)

since a~ n ~ ( 7 ~) = ak N ~ i ( ~ i ) . Finally, am N~(~) = am n/~ ( ~ ) > 0, by
(a3). Thus ~ ¢ A ® O. This concludes the proof.
6.5. M o r e P r e s e r v a t i o n Results
THEOREM 6.5.1. Let 5_ ~ A, O. Then A ® ® is compact iff both A and ®
are compact.
PROOF. Proceed as in the proof of Theorem 6.3.3. The only difference
is that we work with sets of formulae rather than a single formula. Call the
starting set A. We get a sequence of models (gk,/3k, w0) satisfying [a]k - [f]k
for ~kA[]. Now put
g~ := UkE~,gk
<~k := UkEw <~k
"qk : : UkEw " k
Furthermore, for a p-variable p put/3~ (p, x) := 1 if/3k (p, x) = 1 for almost
all k E w,/3~(p,x) := 0 if/3k(p,x) -- 0 for almost all k. It is checked that if
t3k(p, x) = 1 then also t3k+l(p, x) = 1, and if/3k(p, x) = 0 then/3k+l(p, z) = 0.
FinMly, let V be any standard valuation such that/3~ ~ 7. It is not hard to
show, using the properties [a] - [f], that (g~, V, w0) ~ A. []
6.5. More Preservation Results 297
THEOREM 6.5.2. Suppose 5_ ~ A, (9. Then A ® 0 is globally complete

(globally compact) iff both A and 0 are globally complete (globally compact).
For weak compactness this method fails to yield a transfer theorem. Indeed,
a counterexample can be contructed as follows. Take a m o n o m o d a l logic A
which is weakly compact but not compact, for example G r z . 3 . As is shown
in [66], G r z . 3 is weakly compact. To show this is an exercise, see below.
Earlier, in Section 3.2, we have shown that G r z . 3 is not ~1-compact. Now
we show that G r z . 3 ® K is not even 1-compact. For there exists a set A
which is G r z . 3 - c o n s i s t e n t but lacks a Kripke-model. Then A is based on
infinitely m a n y variables, namely var[A] = {Pl : i E w}; now let A result
from X by replacing the variable Pi by the formula Ii+15_ A -~Iih_ for each
i E w. Then var[~k] = ~ and so 5 is based on no variables. Clearly, 5 is
consistent; but if A has a model based on a K r i p k e - f r a m e then this allows a
direct construction of a Kripke-model for A. Thus G r z . 3 ® K is indeed not
1-compact. It is to be noted that this argument uses only the fact t h a t K
has infinitely m a n y constant formulae. Any other weakly compact logic with
these properties will do.
Now we will turn to interpolation and Halld~n-completeness. The proof
in both cases consists in a close analysis of the consistency formulae E D(~ V ¢)
and ED(~ -+ ¢). Since both are identical, it suffices to concentrate on the
latter. We can write ED(T) = V(~c : c E C) and ED(¢) = V{~d : d E
D). Then obviously ED(~ -+ ¢) is (up to boolean equivalence) a suitable
disjunction of ~c A ~d; namely, this disjunction is taken over the set E of all
pairs (c, d) such that ~ A ~d is consistent. Equivalently, we can write
En(~v -+ ¢) = ED(T) A Z[](¢) A A ( ~ ¢ -+ ~ ¢ d : (c,d) ¢ E)
We abbreviate the third conjunct by V(~; ¢) (or, to be more precise we would

again have to write V D ( T ; ¢ ) ) . Obviously, V ( T ; ¢ ) serves to cut out the
unwanted disjuncts. In some sense V(F; ~) measures the extent to which ~v
and ¢ are interdependent. So if V(p; ¢) = T both are independent. It is vital
to observe that all reformulations are classical equivalences.
THEOREM 6.5.3. Suppose that 5_ ¢ A, (9. Then A®(9 is Halldgn-complete

iff both A and (9 are.
PROOF. ( 3 ) Suppose ~V#) E A and var(gz)Mvar(¢) = Z. Then ~ V ¢ E
A ® ( 9 and so either ~ E A ® ( 9 or ¢ C A ® ( 9 and thus either ~ E A or
¢ E A, since A ® (9 is a conservative extension of A. ( ~ ) By induction on
n := adp(T V ¢). For n = 0 this follows from Theorem 1.7.14. Now assume
that n > 0 and that the theorem is proved for all formulae of alternation
depth < n. Take ~ V ¢ such that var(~) 0 vat(e) = Z and adp(~v V ¢) = n.
Assume adp~(EG(p V ¢)) < adp~(~ V ¢). (By the calculations following
T h e o r e m 6.3.5 we may assume that adpS(E[~(p V @)) < adp[~(p V ~) or t h a t

adp"(E,,(p V ¢)) < adp"(~ V ¢). We only show how to deal with the first
case; the other case is dual.) Then by Theorem 6.4.8,
for large m, by which

kcs []<mr~(@~ A [ 3 < ' ~ ( ¢ ) G A DS'~V(~;C)s ~ ~,s VCn
The crucial fact now is that V(@;@) = T. For if ~c and ~d are both
A ® e-consistent, then since var(~c) n var(Sd) C_ var(~) n vat(@ -- e and
adp(~a, adp(~d) < n, ~ A ~d is A ® O-consistent by induction hypothesis
Thus V(~;¢) is an empty conjunction. Consequently, we can rewrite the
above to
ks []<-'~sD(@ c~ A E3<m~S~(¢)Es~. ~,r~ V Cos
kn I 3 < ' ~ ( ~ ) s --~ ~s. V .D-<m~3~(@~ --~ ,¢c~
Now since A is Halld~n-complete, we have [ ~ < m E s ( ~ ) S --+ ~ S 6 A or
[3-<mE=(¢) [] -+ ¢ ~ C A from which by Theorem 6.4.8 ~ E A ® ® or
¢EA®®. []
THEOREM 6.5.4. Suppose that _L ~ A, O. Then A ® 0 has interpolation

iff both A and 0 have interpolation. Moreover, if ~ -+ @ ~ A @ 0 then an
interpolant ~ can be found such that adp~Oc) <_min{adpS(p), adp~(~) } and
adp"(X) < min{adp"(~), adp"(¢)}.
PROOF. ( ~ ) Let ~ --+ ~ ~ A. Then by hypothesis there is a X such t h a t
-+ X, X -+ ¢ ~ A @ ® based on the common variables of ~ and ~. Now,
by Makinson's Theorem, either O(p ++ I p ) or O ( I p ) is consistent. Let the
former be the case. Then let X° result from X by successively replacing a
subformula I ¢ by ~. Then X° ~ L ~ and X ++ X° ~ O(p ++ rap). Hence, as
-+x~A@O, thenalso~-~X ° EA@O(p+~mp). ButA®®(pe+ip) is
a conservative extension of A and therefore ~ -~ X° ~ A. In the case where
® ( I p ) is consistent, define X° to be the result of replacing subformulas of
type " ¢ by T. Then use the same argument as before.
( ~ ) By induction on n := adp(~ ~ ¢). The case n = 0 is covered by
Theorem 1.7.14. Now suppose that n > 0 and that the theorem has been
proved for all formulae of alternation depth < n. Let ~ --+ ¢ ~ A ® ~. We
may assume that adp~(p -+ ¢) ~ adp"(~ --+ ¢) and thus
adp"(E=(~ ~ ¢)) < adp"(~ -~. ¢)
(see the calculations following Theorem 6.3.5). Then adp(E=(~ ~ ¢)) <
adp(~ --~ ¢). By Theorem 6.4.8, for sufficiently large m,
(t) ks ~_<-~r~(~)~ A ~ A SS~V(~;¢) ~ ~ 8 - < ~ ( ¢ ) ~ ~ ¢~

6.5. More Preservation Results 299
Let ~c -+ ~¢d be a conjunct of V(~; ¢). By induction hypothesis and the fact
that adp(~c), adp(¢d) < adp( ~ --+ ¢) (since we have adpm(~c), adpm(~¢d) <
adpm(~ --+ ¢) and adpC](~c), adpn(~d) < adp~S(~ --+ ¢)) there is an inter-
polant Q~,d for ~c and Cal. Note that var(Q~,d) ---- varP(Qc,d) C var(~) N
var(¢) and t h a t
a@~(e~,~) <_~i~{~@D(~), a@D(~)} < mi~{a@C(~), a@C(¢)}
Likewise for I . Again by Theorem 6.4.8 we get
~ D<-mr~(~ ~. Q~,d)D A D<-'~(Q~,d --~ -£bd)D.
-~ .(~c -+ Q~,d) c A (Q~,~ -~ ~ ) c
and therefore with F := C × D - E (recall the definition of V)
AF SSmr~(~ _~ Q~,d)DA A~ D-<~r~(Q~,d -~ ~¢~)~
es v(~; ¢)a
Thus (~) can be rewritten modulo boolean equivalence into
~-<~r~(~)c ~ vn ~ A~ ~-<~r~(~c -+ Q~,~)[]

~ A~ ~-<~E~(Q~,~ ~ ~¢~)~ A n-<'m~(¢) ~ -~ ¢~
Abbreviate the formula to the left by ~?t and the one to the right by Ur. Then
adpCS(t ~) = max{adpS([3<-mEs(@)), adpS(@),

a@~(A~ s<~rc(C~ ~+ e~,~))}
= adpN(@),
since we have t h a t adpS([3<-mEs(~)) <_ adpS(p) by an earlier observation
and
a@c(A C~<~rs(~c -+ ec,d)) <--a@~(D<-mr~c(~)) •
F
By a similar argument
a@S(t vT) = .~a~{a@D(~S-<mSs(¢)),a@D(¢),

= a@C~(¢).
Likewise we argue with adp I. By assumption on A, there is an interpolant X

for Ut and ~r. By definition, X is based on the same surrogate variables as T]t
and ~]r. Therefore for the total reconstruction X $ of X
a@[](x ~) <_ min{a@E(t V~), adp[](V;)} = rain{a@[](@, a@D(¢)}

and similarly for adp • . It is easily verified that varP(x 7) C varP(~)A varP(¢).
Moreover, from r/e = r~ :3 ~-DJ" XG with Theorem 6.4.8 and the fact that the
consistency formulae are A ® ®-theorems we conclude t h a t ~ ~-~ • Xi" and
likewise t h a t X¢ F-:] • ¢. []
T h e o r e m 6.5.4 implies a stronger interpolation property for A ® O. Namely,

if ~ --+ ¢ E A ® O then an interpolant exists which is not only based on the
common variables but also contains only the modalities which occur in b o t h
and ¢.
E x e r c i s e 210. The logic H e r e defined by the axiom p -+ Dp is Sahlqvist.

It corresponds to (Vy E>x)(x = y) V (Vy ~> x)f. Define for a modal logic A the
logic A(c), which is the result of adding a new propositional constant c to the
language. Now define a translation from T~(c) to T~+I by translating c by
<>~T. (Further, the translation commutes with 7, A and (>A, ~ < ~.) Show
t h a t the translation induces an isomorphism from g A(c) onto E A ® H e r e .
E x e r c i s e 211. (FINE [63] and [66].) Show that G r z . 3 is weakly compact.

Hint. First of all, the canonical frame based on n generators is linear. Now
show t h a t there is no infinite properly ascending chain of points by showing
that the underlying Kripke-frame is isomorphic to (c~, >), where a is an or-
dinal number. To do that, let 3 be a G r z . 3 - f r a m e such that 3+ is generated
by the elements ai, i < n. Define the c h a r a c t e r of a point x to be the set of
all subsets C of n such that
x Na oN-a
icc i~c
A point is of m i n i m a l c h a r a c t e r in M C a iff for all y E> x, and y E M ,
y has the same character as x. Show that if x is of minimal character in M
and x <3 y for y E M then x = y. Let x ° be the point of minimal character
in c~. In general, let x ~ be the point of minimal character in the set M z :=
{ y : (V7 < ~)(y <~ x ~ and y # x~)}. Whenever M ~ is not empty, there exists
a point of minimal character. Now define as the new frame ~ the subframe
induced on the set g of all x ~. Show t h a t ~5+ is isomorphic to 3+.
6.6. T h o m a s o n Simulations
Now t h a t we have seen how to embed modal logics into polymodal logics,
we will turn to the question of simulating potymodal logics by m o n o m o d a l
logics. The results can be found also in [133], though sometimes with different
proofs. Again, it is useful to restrict the discussion to the case of bimodal
logics. It is a priori not clear that we can use a single operator to simulate
two operators but it will turn out that the situation is as good as possible.
Not only can we perform such a simulation, we can also m a p the whole lattice
6.6. T h o m a s o n Simulations 301
FIGURE 6.1. A Simulation
3 3 3° 3°
"I
2 bs
• --2 ~ ~ .
1 1 1° 1"
of extensions of K2 isomorphically onto the interval [Sim, Th [-g]], where S i m

is some finitely axiomatizable ( m o n o m o d a l ) logic. This will have n u m e r o u s
consequences for the theory of m o d a l logic, as we will see. T h e construction
works only for n o r m a l logics, and in the exercises we give some indication as
to why it fails for q u a s i - n o r m a l logics. T h e basic idea of simulations is due to
S. K. THOMASON, who used it in [209], [208] and [210] as a tool to derive
certain negative facts a b o u t m o n o m o d a l logics. We explain this simulation
first by using frames and then go over to algebras.
Take a b i m o d a l frame ~3 = (b, <1, 4 , ~ ) . T h e s i m u l a t i o n of ~3 is a frame
~3 s = (b s, ~<,1~s) for the logic with the o p e r a t o r [~. It is defined as follows.
We let V := { o , . , . } . P u t b s := { * } U b x { o , . } . We will denote (x,o) by
x ° and (x,*) by x °. Moreover, if x C b then x* will be a n o t h e r n a m e for ..
T h e n o t a t i o n is extended to sets of points. For a set A C_ b, A ° := A x {o},
A ° := A × {.} and A* := {*} i f A i s nonempty, and ~* := ~. We use the
symbols b and ~ as variables ranging over V ( = {% o, *}). T h e relation ~< is
defined as follows.
~=o,~=,,
or ~=o,~=o, x=y
x b ~y~ if or ~=°,~=o, x=y
or ~=o,~=o, x <~ y
or ~ = o , ~ = o, x • y
]B~ : = { c ' t A d ° t J e . : c , d , e c l B }
This defines ~3 ~. Proposition 6.6.1 asserts t h a t this is indeed a frame. Given a
K r i p k e - f r a m e b, b ~ := (b ~, <~). For example, let b consist of three points, 1, 2
and 3. Let 1 <l 1, 1 <~ 2 and 2 • 2 and 2 • 3. T h e n b ~ is shown in Figure 6.1.
If ~ is the e m p t y frame then ~3 ~ = rg]. This is a useful fact. It accounts
for the fact t h a t the simulation of the inconsistent b i m o d a l logic is not the
inconsistent m o n o m o d a l logic, but rather the logic of [-g].
PROPOSITION 6.6.1. Let ~ be a birnodal frame, and let ~ s be defined as

above. Then ~8 s is a frame.
P R O O F . C l o s u r e u n d e r c o m p l e m e n t a n d n e g a t i o n is clear. N o w we s h o w
c l o s u r e u n d e r (}. W e c a n r e d u c e this to a d i s c u s s i o n of t h r e e cases, n a m e l y
c=d*, c=d ° andc=d'. Let c=d*. Ifd=£~, (}d* = £ ~ = d * . Ifd~ O,
t h e n (}d* = b °. N e x t , let c = d °. T h e n (}c = d" tJ ((}d) °. F i n a l l y , let c = d °.
T h e n ~ c = d ° U ( $ d ) °. Since b o t h (}d a n d $ d are in B, we are d o n e . []
L e t us n o t e t h a t the sets {*}, b ° a n d b" are d e f i n a b l e b y constant f o r m u l a e .

T h i s will b e of e x t r e m e i m p o r t a n c e . N a m e l y , p u t
w := BJ_
a := ~B±
/3 := ~_[_A--,B±
W e will u s u a l l y also d e n o t e b y a , / 3 a n d w t h e sets of p o i n t s d e f i n e d b y a , / 3
a n d ~ in a given f r a m e . It is e a s y to v e r i f y t h a t w = {*}, a = b ° a n d / 3 = b °.
W e c a n n o w c o n c l u d e t h a t if we have ]~ t h e n B s is g e n e r a t e d b y t h e s e t s c ° U c °
u s i n g t h e o p e r a t i o n s - , r] a n d El. N o w let !~ a n d ~ b e b i m o d a l f r a m e s a n d
7r : ~ -+ C a p - m o r p h i s m . P u t
S(x b) :=
W e c l a i m t h a t 7r~ : !~ s -+ ~ . So, let us c h e c k t h e first c o n d i t i o n o n p -

m o r p h i s m s . A s s u m e x ~ ~ y~. (1.) ~ = o, ~ = . . T h e n ~r~(x ~) = ~-(x) °
7r(y)* = 7rS(y~). (2.) ~ = o, ~ = • a n d x = y. T h e n 7rS(x ~) = 7r(x) ° ~<
7r(x)" = 7r~(y~). (3.) ~ = . , ~ = o a n d x = y. A s in (2.). (4.) ~ = ~ = o.
T h e n x <3 y a n d so 7r(x) <3 7r(y) b y a s s u m p t i o n on 7r. T h e r e f o r e , 7r~(x ~) =
~'(x) ° ~< 7r(y) ° = 7r~(y~). (5.) ~ = ~ = - . A n a l o g o u s t o (4.). T h i s e x h a u s t s
all cases. N o w let us check t h e s e c o n d c o n d i t i o n . A s s u m e t h a t 7r(x) ~ ~< u ~
for s o m e u E c, c t h e u n d e r l y i n g set of ¢. B y d e f i n i t i o n of 7r~, 7r~(x ~) ~< u~.
Henceb #*. If~ =*,thenb = o a n d we m a y t a k e u : = x . Thenu~ =*,
a n d 7r~(x ~) = 7r(x) ° ~< 7r(x)* = u ~. So a s s u m e f r o m now on t h a t ~ # * a n d
b # *. A s s u m e first t h a t b = o. If also ~ = o t h e n we m u s t h a v e 7r(x) q u, b y
c o n s t r u c t i o n of ¢ ~. Since 7r is a p - m o r p h i s m t h e r e is a y such t h a t 7r(y) = u
a n d x < 3 y . T h e n x ~ ~< y~ a n d ~rS(y ~) = ~r(y)~ = u~, as r e q u i r e d . I f ~ = •
then = t h a t case, we h a v e x" a n d = = gain
as d e s i r e d . A s s u m e n e x t t h a t b = *. T h e a r g u m e n t a t i o n is p a r a l l e l t o t h e
first case. F i n a l l y , we m u s t show t h a t for e v e r y i n t e r n a l set d of ~ t h e set
(7r~)-~[d] is i n t e r n a l in ~ . So let d = d ~ U d ~ U d * for s o m e sets d~,db, do ~ C.
Then
(Tr~)-~[d] = 7r-~[dw] ° U 7c-~[db] ° U 7r-~[do] *
T h i s is a n i n t e r n a l set of ~ .
6.6. T h o m a s o n S i m u l a t i o n s 303
THEOREM 6.6.2. The map (_ )s is a functor from the category of bimodal

frames into the category of monomodal frames. Moreover, for a bimodal frame
~ , ( ~ ) ~ = ( ~ ) ~ and for a bimodal Kripke-frame b, (b~) ~ = (bS) ~.
PROPOSITION 6.6.3. Let ~ be a bimodal frame.
1. ~ s is differentiated iff ~ is.
2. ~ s is refined iff f~ is.
3. ~ s is compact iff ~ is.
PROOF. (1.) If ~ s is differentiated, t h e n surely ~ is differentiated. Now
assume t h a t ~ is differentiated. Since b°, b" a n d {*} are i n t e r n a l sets, x ~ a n d
y~ can be d i s c r i m i n a t e d by a n i n t e r n a l set if b ¢ ~. So, a s s u m e t h a t ~ = ~.
I n case b = * we have x = y by c o n s t r u c t i o n . So, finally, b -- ~ E {% .}. I n
t h a t case x ¢ y a n d there exists a set c E B such t h a t x C c b u t y ¢ c. T h e n
x ~ E c b b u t y ~ ¢ c ~.
(2.) Suppose t h a t ~ is tight. T h e n ~ is tight as well. For let x ~ y.
T h e n x ° ~ yO a n d so there is a set a such t h a t x ° C B a b u t yO ~ a. Let
a = a o1 U a ~ U a 3* for some i n t e r n a l sets a l , a2, a3 of ~ . T h e n x C [1al b u t
y ~ a l . Now let x • y n o t be the case. T h e n x ° ~ y ' . So there is a s e t a
such t h a t x" C [3a b u t y" ~ a. Let a = a~ U a~ 12 a~ for some i n t e r n a l sets a l ,
a2 a n d a3 of ~ . T h e n y ~ a l , b u t
x° e n B(Z aD).
F r o m this follows x E h a l , as desired. Now it follows by (1.) t h a t if ~
is refined, ~ is refined as well. Now assume t h a t ~ is refined. T h e n ~ s is
differentiated, by (1.). We have to show t h a t it is tight. Let x ~ y. We have
to find a such t h a t x C ~ a b u t y ~ a . C a s e 1. x = * or y = * . If x = * p u t
a : = ~ . If y = * t h e n x E/3. So p u t a : = - / 3 . C a s e 2. x - - u ° a n d y = v "
(or x = u" a n d y = v°). T h e n u ~ v a n d so there is a set c such t h a t u C c
b u t v ~ c. T h e n x E B ( - / 3 U c ' ) , since every successor which does n o t satisfy
13 m u s t be of the form u ° C c ' . Moreover, y ~ - / 3 [ 2 c ' . So, a : = - / 3 U c ' .
C a s e 3. x ° ~ y° or x ° ~ y ' . Let the first be the case. ( T h e other case
is dual.) Here, we can use the a s s u m p t i o n of tightness for ~ to get a set c
such t h a t x C [~c b u t y ~ c. T h e n x ° E B(c ° t2 - a ) b u t yO ~ - a U c ° . P u t
a : = - a [D c °.
(3.) A s s u m e ~ is compact, a n d let U be a n ultrafilter on ~ . P u t U ~ : =
{a ° : a ~ U}. U ~ is a n ultrafilter on a in ~ s . U ~ c a n be e x p a n d e d to a n
ultrafilter U + on ~ , t a k i n g a]] sets c such t h a t cMa ~ U ~. We have a ~ U +.
U + has n o n e m p t y i n t e r s e c t i o n by a s s u m p t i o n on ~ . Let x ~ ['1U +. T h e n
x ~ a , so x = u ° for some u. T h u s u ~ ['1 U, showing the c o m p a c t n e s s of ~ .
Now assume t h a t ~ is c o m p a c t a n d let U C_ B ~ be a n ultrafilter on ~ . We
have w U a (J/3 C U. T h u s three cases arise. C a s e 1. w ~ U. T h e n {,} ~ U
and so* ~U. C a s e 2. a ~ U . T h e n the t r a c e U a = { a V l a : a ~ U}isan
ultrafilter on a. Hence it corresponds to the ultrafilter V : = {v : v ° ~ Ua} on
~ . By assumption, there is a y such t h a t y C n V. T h e n yO E n u ~ = n u -

This completes the second case. C a s e 3. /3 C U. T h e n the trace UZ : =
{aN/3 : a E U} is an ultrafilter on/3. B u t then U' = { a N ~ z a : a C U~} is an
ultrafilter on a, since we replace sets of the form a" by sets of the form a °.
We t h e n get a y° C n u' and so y" • n u~ = n u , as required. []
T h e construction can be defined on algebras as well. Namely, assume t h a t

= (B, 1, - , N, D, , ) is a bimodal algebra. T h e n let B e : = B × B × 2 and
(bNGa, aNmb, 1 } i f c = l
[](a,b,c) : = (0, a N I b , 1) ifc 0
T h e n K~s := {B e, 1, - , N, []/ is a m o n o m o d a l algebra, where I : = (1, 1, 1}.

We use a, /3 and w for the value of a, /3 and a;, respectively, in 92. Given a
h o m o m o r p h i s m h : ~B -+ ~ of bimodal algebras, the m a p h x h × id(2) can be
shown to be a h o m o m o r p h i s m of m o n o m o d a l algebras.
THEOREM 6.6.4. ( _ ) s is a funetor from the category of bimodal algebras

to the category of monomodal algebras.
THEOREM 6.6.5. Let ~ be a bimodal algebra. Then (~+)~ ~ ( ~ ) + . Let
be a bimodal frame. Then (~+)~ ~ (C~)+.
PROOF. T h e second claim is rather straightforward. Let ff be given. T h e n
the m a p s d ~+ d N a, d ~-~ d N/3 and d ~+ d N w are projections of the b o o l e a n
algebra of internal sets of ff~ onto the boolean algebra of internal sets of ~,
and onto 2, respectively. T h e y induce an isomorphism C ~ ~ C × C × 2. T h e
operation [] on C ~ is checked to be exactly as in the definition of (C+) ~. This
shows the second claim. Now to the first claim. We define a bijection ~ from
the set of points of (~3~) + onto the set of points of (~B+) s. Take an ultrafilter
U of ~ . This is an ultrafilter in the boolean algebra B × B x 2, where
B = (B, 1 , - , N / . Now, (1, 1, 1} • U, and (1,1, 1} = (1,O,O/U(O ,1,O}U(O,O,1 /
and so either (1,0,0) • U, or (0,1,0) • U or {0,0,1) • U. In the first case,
there exists an ultrafilter V on B such t h a t U = {(a, b, c) : a • V, b • B, c • 2}.
We then p u t ~(U) : = V °. In the second case there exists an ultrafilter V on
S such t h a t V = { ( a , b , c } : a • B , b • V,c • 2}. T h e n we p u t ~(U) : = V °. In
the third case U = {{a,b, 1} : a,b • B}. T h e n we put L(U) = * or L ( U ) : = V*,
where V is any ultrafilter of B (by our n a m i n g convention this is the same).
By careful checking of cases it is shown t h a t L is an isomorphism. []
Given a class 9( of frames we write :Xs for the class {~B~ : ~ • :X} and
similarly for classes of algebras.
DEFINITION 6.6.6. Let (~ be a bimodal normal logic. Then we define

O~ : = Th (Frm E)) ~. This is called the (monomodal) s i m u l a t i o n of 0.
6.6. Thomason Simulations 305
Our aim is to show that the map O ~+ (98 actually is an isomorphism

from the lattice ~ K2 onto an interval in the lattice ~ K1. To show this we
first show how to axiomatize O s. Before we can give such an axiomatization,
we state a simple lemma.
LEMMA 6.6.7. Let 9)I be a monomodal flame. 9)I ~- ~3 s f o r some bimodal
frame iff
(A) every point satisfies either w, a or t3,
(B) there exists exactly one point in w,
(C) every point in ~ sees one point in w,
(D) no point of 13 sees a point in w,
(E) every point in a has exactly one successor in t3,
(F) every point in 13 has exactly one successor in a,
(G) for every point x in a and every point y in t3, x <. y iff y <. x.
It is clear that the properties (A), (C) - (G) are modally characterizable
and also dr-persistent. The problem lies in condition (B). However, notice
the following.
LEMMA 6.6.8. Let 9Y~ be a rooted monomodal frame. Let ~ satisfy the
properties (A), (C) - (G) of L e m m a 6. 6. 7. Then 9)I satisfies (B) iff it satisfies
(H).
(H) for all x and all y , y ' 6 w such that x <<3 y;y, we have y = y~.
PROOF. It is clear that (H) rules out that a point has more than one
(direct) w-successor. On the other hand, each point sees an w-point in at
most two steps. If that point is always the same for each point reachable from
the root, we have only one w-point. So, if w contains more than one point,
there must be a point x ¢ w seeing in at most two steps an w-point u, and a
E-successor y of x such that y sees in at most 2 steps an w-point u ~ different
from u. Then x sees in at most three steps two different w-points. This is
ruled out by (H), however. []
DEFINITION 6.6.9. The logic S i m is the extension of 1£1 by the axioms

(a) - ~ .
(a) wv~v/3
(b) ~. - ~ . e (w A p) -~ E~(w -~ p)
(e) o~.-4 .e,L/A ew
(d) ~. -~ . e ( ~ A p ) ~ ~ ( ~ - - ~ p )
(e) /3. -e .eolA =ew
(f) /3. --+. ~ (o~ A p) -+ E~(o~--+ p)
(9) ~(,~ -~ p))
~. ~ .p ~ ~ ( 9 ~
(h) e-<~(w/\p) ~E3-<~(w~p)
Some of the axioms are satisfied by choice of a, /3 and ~, but this is unim-
portant. The following is easy to verify.
PROPOSITION 6.6.10. Sirn is dr-persistent. It is Sahlqvist and of special
rank O.
(For a proof, notice that the axioms state properties that assert (i.) the
existence of successors or (ii.) the uniqueness of successors. These are of
special rank 0.) Before we prove that S i m is correctly defined let us introduce
some notation. P u t ~ X := ~ ( a A X), ½~X := ½ ( a --+ X) and likewise for/3
and w. T h e n
~x = Vbev~ = ~xve~xve.,x
½x = A~vsbx = ½~XA½,XAB~X
Clearly, ~ X is satisfied at a point x in a frame if there exists a successor y
which is in b and satisfies X. B~X is satisfied at a point x if all successors y of
which are in b satisfy X. This allows to rewrite some of the postulates.
(b') a. -~ . ~ p - + ½~p
(d') a. --+ .¢zp --+ Bzp
(f') ~. --+ .e~p ~ ½~p
(g') a. -+ .p --+ ½~ ~ p
Moreover, we have the following theorems, which follow from L e m m a 6.6.7
and 6.6.8, respectively.
COROLLARY 6.6.11. Let 9)t be a differentiated monomodal frame whose
theory contains (a) to (g). Let w have a single point as its extension in i)Y~.
Then for some bimodaI frame ~ we have 93I ~ f~s.
COROLLARY 6.6.12. Let 93I be a rooted, differentiated frame for Sire.
Then YY~~ ~3s for some (possibly empty) differentiated bimodal frame ~ .
The previous theorem shows t h a t S i m axiomatizes the right kind of
frames, assuming that we deal with differentiated frames. To obtain an axi-
omatization of (9 ~ on the basis of an axiomatization for O we must also find
a syntactic correlate of the frame simulation. The simulation of a bimodal
formula U is defined as follows.
pS :~ p
(~u) s := o, A ~(u s)
(uA,¢) ~ := u~A,¢ '
(OU) s := ot A e U s
($U) s :---- o~A e ( / 3 A e ( ~ A e(txAuS)))
PROPOSITION 6.6.13. If U is a Sahlqvist formula, so is Us.

6.6. T h o m a s o n Simulations 307
PROPOSITION 6.6.14. Let ~ be a bimodal f l a m e and 3" a valuation on ~3 s

such that 7(P) N b° = (/3(p)) °. Then
PROOF. By induction on ~. []
PROPOSITION 6.6.15. Let (~3,13, x) be a bimodal model. P u t ~S(p) : =

j3(p) °. T h e n the following holds.
¢. s
¢, -4
PROOF. T h e first is an immediate consequence of the previous lemma.

T h e second also follows, since x ~ satisfies a in ~3 s if[ b = o. T h e third is
proved thus. F r o m right to left is a consequence of the previous line. From
left to right is not entirely obvious. Pick a valuation 3' on if3s. T h e n let
13(p) : = {x : x ° E 3'(p)}. T h e n j3(p) ~ b ° = (3'(p))° and so by the previous
l e m m a (~,/3, x) ~ ~ iff ( ~ s , 3 ' , x ° ) ~ a -4 ~s. Moreover, (~3% 3`,x'} ~ a -4 ~
as well as ( ~ s , 3 ' , x * ) ~ a -4 (ps. So, ( ~ , f l ) ~ ~ i f f ( ~ s , 3 ' ) ~ a -4 Ts. Since 3'
was a r b i t r a r y and ~ ~ ~, we conclude t h a t ~3s ~ a -4 ~s. []
THEOREM 6.6.16. Let O be a bimodal logic and O = K2 ® X . Then

O s = SimG{a-4~:~EX}.
PROOF. Let A : = S i r e ® {a -4 p~ : ~ E X}. A differentiated rooted

frame for A is of the form ~3 ~ for some bimodal frame ~3 by Corollary 6.6.12.
~3~ ~ a -4 ps for all ~ E X. Hence ~ ~ ~ for all ~ E X . So, ~ is a O - f r a m e .
The converse also holds. Hence, A = O s. []
Now t h a t we have defined the simulation of a b i m o d a l formula let us

see whether there is a possibility of recovering from a b i m o d a l frame the
m o n o m o d a l frame of which it is a simulation; this we call unsimulating. We
will deal with unsimulations of frames t h a t are frames for S i r e and differ-
entiated. It is possible to generalize this s o m e w h a t (see the exercises) but
there is no benefit for the aims t h a t we are pursuing here. First, the previous
theorems guarantee t h a t for a rooted differentiated E l m - f r a m e 9)~ there is
exactly one ~ such t h a t %s ~ 9)I. We can make q3 unique by the following
construction.
DEFINITION 6.6.17. A s t a n d a r d s i m u l a t i o n f r a m e is a differentiated

m o n o m o d a l S i r e - f r a m e such that co has a single point as its extension. The
category of standard simulation frames and p - m o r p h i s m s is denoted by StSim.
Let 9)I be a standard simulation frame. Then 9Jts is defined as follows, ms :=
m N a, x < y iff x <. y and x *1 y iff there exist x I, yl E j3 such that x <~
x' ~ y' ~ y. Finally, M~ : = { a N ( ~ : a • M} = {a • M : a ~ c~}. Then

ff)~s : = <ms, <, ~,M[). ff)~s is called the u n s i m u l a t i o n of 97L
Let 8)I and 91 be standard simulation frames and ~r : 9)I -4 91 a p-

morphism. Then put Try(x) := 7r(x), x • a. It is not hard to verify that this
is a p-morphism from 9Y[ s to ff[~. Moreover, each p-morphism p : 9Y~ -+ 91 s
is actually of the form 7r~ for some ~T : ~ --4 ~Y[.
THEOREM 6.6.18. Let Dr2 denote the category of differentiated bimodal

frames. (-)~ is a functor from StSim to Dr2. Moreover, the two categories
are naturally equivalent; there is a natural transformation from the identity
functor on StSim to the functor ( ( - ) s ) s and a natural transformation from
the identity functor on Dr2 to ( ( - ) s ) s. In particular, for a bimodal frame
and a monomodal frame ~ we have (f~s)s ~- ~ , (gY~s)s ~ ~;L
PROOF. Let ~ be a bimodal frame. T h e n U ( ~ ) : x ~-~ x ° is an isomor-

phism from ~ to (~s)~. ~7is a n a t u r a l t r a n s f o r m a t i o n from the identity on Dr2
to ( ( - ) ~ ) s as is straightforward to check. Now let 9)I be a s t a n d a r d simulation
frame. T h e n e(ff2) is defined as follows, e(ff2)(x) : = x ° if x • a , e(g:~)(.) = ..
If x • / 3 , e(ff2)(x) := y ' , where y • a and x ~ y. Again, it is straightforward
to verify t h a t e(gY[) is an isomorphism from 9)I onto (gY~) ~ and e a n a t u r a l
t r a n s f o r m a t i o n from the identity functor on StSim to ((-)~)L []
Notice the following particular consequence. The c a t e g o r y of b i m o d a l

frames has coproducts, n a m e l y the disjoint union. So, by equivalence, the
c a t e g o r y of s t a n d a r d simulation frames must have coproducts, too. However,
it is not closed under disjoint unions. So, the c o p r o d u c t of frames exists b u t
is not the disjoint union. Since we have an equivalence, the frame o p e r a t i o n
defined below is a coproduct.
A
:_-
iCI i~I
There is an alternative way of defining ~ . Take a differentiated S i m - f r a m e

~ . Let ~ be defined by x ~ y i f f x , y E w or x = y. T h e n ~ is a n e t and
the n a t u r a l factorization f f 2 / ~ is denoted by 8)I.. It is easy to verify t h a t
~ , is a s t a n d a r d simulation frame. Moreover, for every point x, Th <9)~, x} =
Th <9)~,, [x]). (This follows from the fact t h a t the transit of x in ~ has a
unique w - p o i n t in it, and so is isomorphic to the transit of Ix] in ~ , . ) It is
easy to see t h a t
A
iEI icI
6.6. Thomason Simulations 309
Similar definitions can be made with respect to the coproduct of descriptive

frames.
i~I iEI
Call this construction the r e d u c e d c o p r o d u c t . Now recall from Theo-
rem 4.7.5 that a class of descriptive frames is modally definable iff it is closed
under coproducts, p-morphic images and generated subframes. It follows that
if a class of bimodal descriptive frames is modally definable, its simulation
image is closed under reduced coproducts, p-morphic images and generated
subframes. The converse also holds. A class of descriptive Sire-frames is
modally definable iff its intersection with the class of standard Sim-frames is
closed under reduced coproducts, generated subframes and p-morphic images.
Putting this together we obtain the following result.
THEOREM 6.6.19. The map 0 ~-~ 0 s is an isomorphism from the lattice

g K2 onto the interval [Sim, Th~-~] in g l .
DEFINITION 6.6.20. Let A be an extension of Sire contained in Th [ ~ .

Then let As be the unique bimodal logic (~ such that 0 s = A.
It is clear that the map A ~-+ As is a lattice isomorphism from the interval
[Sim,ThF~].
E x e r c i s e 212. Let ~ be a bimodal algebra. Show that S u b ( ~ s) ~- S u b ( ~ )
and C o n ( ~ s) ~ C o n ( ~ ) ÷ 1. (Here, ~ + 1 denotes the addition of a new
element at the top of the lattice £.)
E x e r c i s e 213. Call a monomodal algebra local if it has a unique maximal

congruence # V. Show that a Sim-algebra is local iff its dual is a standard
simulation frame. Moreover, show that every Sire-algebra has a largest local
subalgbra.
E x e r c i s e 214. Call a monomodal frame if2 local if (1.) ~ is differentiated,

(2.) the subframe ® of points x such that (5, x) ~ S i m is a standard simu-
lation frame. Define ( - ) t on local frames as follows. Let (D be the subframe
of points x such that (~,x) ~ Sim. Then ~t := ~s- Show that ( - ) t is a
functor. Now define ( _ ) t a functor from the category of bimodal frames into
the category of local frames simply by ~ t := ~s. Show that ( - ) t is right
adjoint to (-)~.
E x e r c i s e 215. Let (~ be a quasinormal bimodal logic. Then define
oq:= N Th(~',x °)ATh(~,x °)ATh(~,x*)

Show that the map 0 ~-+ e q maps ~ K2 isomorphically into 0 Sim. (This
map does not need to be onto!) Moreover, show that if 0 is normal, ®q = (9~.
E x e r c i s e 216. Show that there exists no lattice isomorphism ( - ) q from

0 Sire onto 0 K2 such that for a normal logic A, Aq = As. Show furthermore
that there is no homomorphism from 0 Sire into 0 K2 such that for normal
bimodal logics 0, Aq = 0 iff A = (9s. Hint. Consider the bimodal frame
~ := ({0,1}, <~, , ) where 4 = ~ and <~ = {(0,1)}. P u t (~ := Th[. Show
that there is no lattice isomorphism from 0 (9 s onto 0 (~ with the required
properties.
6.7. P r o p e r t i e s o f t h e S i m u l a t i o n
In this section we will investigate what properties of logics are preserved

under simulation and unsimulation. As on similar occasions, we say that a
property q3 is p r e s e r v e d under simulation if for a given bimodal logic A, A s
has q3 if A has ~3. q3 is reflected under simulation if h has ~ whenever A s has
q3; and q3 is invariant under simulation if it is both preserved and reflected
under simulation. As a first step we investigate the problem of axiomatizing
As on the basis of an axiomatization of A.
AXIOMATIZATION. In analogy to the construction of Ws from a bimodal for-
mula ~ we want to construct for a given monomodal formula ~ a bimodal
formula ¢ such that ~ is equivalent to Cs. This is not possible; just take the
formula a A ~}(/3 A p). The problem is that the bimodal frame is internally
reconstructed as the area defined by a. So the values under a valuation in
the /3 and w region cannot be distinguished. We can, however, mimick the
behaviour of a monomodal valuation as follows. For each variable p we intro-
duce three new variables, p*, p° and p ' . Given our original set V or variables,
we obtain three new sets, V b := {p~ : p C V}, b E {*, o, ,}. All four sets are
assumed to be pairwise disjoint.
In this section we are working in the language [3~, K~ and K]~ together
with their duals {}a, ~Z and ~ . Let ~ be a formula and X a subformula
of ~. Fix an occurrence of X in T. A m o d a l c o v e r of that occurrence of X
is a minimal subformula ~ of modal degree greater than X containing that
occurrence of X. We also say that this particular occurrence of ~ m o d a l l y
c o v e r s X. If X has a modal cover, it is unique and a formula beginning with
a modal operator. (We will often speak of formulae rather than occurrences
of formulae, whenever the context allows this.) Now let T be a formula of the
language with operators K~, ~}v, b E {a,/3, w}. Let us agree to say that an
occurrence of a formula X in T is b - c o v e r e d if it modally covered by a formula
of the form [3v¢ or ~b¢. Call ~ and X w h i t e - e q u i v a l e n t if a ~-Slm ~ ++ X
and b l a c k - e q u i v a l e n t if/3 F-sire T ++ X" Given a formula, we say that a
subformula occurs w h i t e if it is not in the scope of a modal operator or else
6.7. Properties of the Simulation 311
is a-covered. A subformula occurs b l a c k if it is/3-covered. If to occurs white

(black) in ¢, and ~ is white-equivalent (black-equivalent) to X, then that
occurrence of ~ may be replaced in ¢ by X preserving white-equivalence. By
axioms (c) and (d) of Sire, [~Z~- is white-equivalent to (~Z~- and by (e) and
(f), B ~ - is black-equivalent to ( ~ T .
LEMMA 6.7.1. Let ~ be a formula in the language with B~ and ~b, ~ C
{a, f~, w}. There exists a finite number n and formulae Xi and ¢i, i < n, such
that Xi is nonmodal for all i < n, and ¢~ is in the language with ~a,E]Z, ~a
and ~ for all i < n, and ~ is white-equivalent to the formula
i<n
PROOF. By a straightforward semantic argument one shows for every n
Take S C var(~). P u t
x ( s ) := A pA A
pES pEvar(~)-S
Then
scvar(~)
Now by the above consideration, any occurrence of (}~X(S) as a subformula
of ~ can be replaced by -7 in the formula ~A (}~x(S). It remains to be shown
that any subformula K]~X is of this type. Clearly, by some Sim-equivalences,
any subformula []~X can be transformed into a subformula B~X I where XI
is nonmodal. Moreover, we are al]owed to replace X~ by X(S) A X ~. The
latter reduces via some equivalences to either X(S) or to ±. Therefore, the
subformulae of the form B~2_ need to be treated. Several cases need to be
distinguished. (a) K ~ ± is w-covered. Then it can be replaced by -7. (b) B ~ ±
is a-covered. Then it can be replaced by ±. (c) K]~-7 is f~-covered. Then
it can be replaced by -7. (d) K ~ ± is not in the scope of an operator. Then
it can be replaced by I . All these replacements are Sire-equivalences. This
shows the lemma. []
DEFINITION 6.7.2. A monomodal formula ~ is called s i m u l a t i o n t r a n s -
p a r e n t if it is of the form p, ~p, ~Zp, -~}Zp, ~ p , -~(~,~p, p a variable, or
of the form ¢ A X, ¢ V X, ~ ~ ¢ , B~ ¢ , (~~ ~)~ ~ ~ ¢ or E]~ [~~ [~a ¢ where ¢ and
X are simulation transparent.
DEFINITION 6.7.3. Call a formula ~ w h i t e based if there do not exist
occurrences of subformulae Xo, X1, X2 and X3 such that Xo /3-covers X1, X1
~-covers X2, and X2 ~-covers X3.
LEMMA 6.7.4. For every formula ~ there exists a formula X which is

white-based and white-equivalent to ~.
PROOF. Suppose that there is a quadruple (X0,)il, X2, X 3 / o f occurrences

of subformulae such that X0/3-covers X1, X1/3-covers X2, and X2/3-covers X3-
Then there exists such a quadruple in which X0 occurs white. Now replace the
occurrence of X3 by ½~ [~Z)/3. Since X3 is black equivalent with []~ [3Z X3, this
replacement yields a formula ~ which is white equivalent to ~ . Now repeat
this procedure with ~ . It is not hard to see that this process terminates with
a white based formula. (For example, count the number of occurrences of
q u a d r u p l e s ( X o , X 1 , X 2 , X 3 / s u c h that Xo/3-covers X1, ~¥1/3-covers X2, and X2
fl-covers X3. It decreases by at least one in passing from ~ to T~. If it is zero,
the formula is white based.) []
LEMMA 6.7.5. Let ~ be a monomodal formula. Then there exists a simu-
lation transparent formula )l such that
OZ[-Sim ~ ~ X
PROOF. First we simplify the problem somewhat. Namely, by some stan-
dard manipulations we can achieve it that no operator occurs in the scope of
negation. We call a formula in such a form b a s i c . So, let us assume ~ to be
basic. By L e m m a 6.7.1 we can assume ~ to be a disjunction of formulae of the
form v A ¢, where v = ~ X , for a nonmodal X, and ~- contains only Kt~, ~ ,
[~Z and ~ . In general, if the claim holds for T1 and 92, then it also holds for
~1 V T2 and T1 A ~2. Therefore, we have two cases to consider: (i) T contains
no occurrences of ~ , B~, ~}~ or K~Z, or (ii) ~ contains no occurrences of ~
and K]~. In case (i), we know that ~}~ distributes over V and A, so t h a t we
can reduce ~ (modulo white-equivalence) to the form ~ p , and ~ - ~ p . Now
we have
So in Case (i) ~ is white-equivalent to a simulation transparent formula. From

now on we can assume to be in Case (ii). Furthermore, by L e m m a 6.7.4, we
can assume that T is white based, and (inspecting the proof of t h a t lemma)
that ~ is built from variables and negated variables, using A, V, and the modal
operators ( ~ , ~ , 7t~ and ½f~.
Let # ~ ( p ) denote the m a x i m u m number of nestings of black operators
( ( ~ , B~) in ~. Call ~ t h i n n e r t h a n X if either p~(~) < #Z(X) or ~ is a
subformuta of X- We will show that for given white based, basic T there
exists a simulation transparent formula X which is white-equivalent to ~ on
the condition that this holds already for all white based basic formulae T~
thinner t h a n ~.
If ~ = p we are done; for T is simulation transparent. Likewise, if ~ = ~p.
Suppose T = ~l A T2. ~ and T2 are thinner than ~. Therefore there exist
simulation transparent formulae X1 and X2 such that Xi is white-equivalent to

Ti, i C {1, 2}. Then X1 A X2 is white-equivalent to Wl A ~2. Similarly for ~ =
~1 V ~2. If T = ~ 1 there exists a simulation transparent X which is white-
equivalent to ~1. So a --+ T1 -~}-sim a --+ X, and therefore @~T1 -{~-Sim ~aX;
it follows t h a t @aX is white-equivalent to ~. Similarly for ~ = [ 3 ~ 1 . We
are left with tile case that ~ is either @Z~- or [~7-. By inductive hypothesis,
for every basic white based X such that #Z(X) < #Z(W) there is a simulation
transparent w such t h a t w is white-equivMent to X- As T occurs white, we
can distribute []~ and @~ over A and V, and so reduce 7- to the form p, ~p or
K]~p or ~}~p, with b E {a,/3} and p basic. This reduction does not alter #Z(~).
C a s e 1. T = p. Then W is simulation transparent.
C a s e 2. T = -~p. Observe that ~ Z = p is white equivalent to ~@~p. So we are
done.
C a s e 3. ~- = B~p or T = @~p. Then ~ is white-equivalent to p. The claim
follows by induction hypothesis for p.
C a s e 4. ~- = ~ p or 7 = ~Zp, p basic. Now let us look at p. p occurs black.
Furthermore, p is the result of applying a lattice polynomial to formulae of
the form p, ~p, [3~t*, @~v, [ ~ { , @ j / . (Here, a lattice polynomial is an
expression formed from variables and constants using only A and V, but no
other functions. It turns out that ~- and £ can be eliminated from this
polynomial as long as it contains at least one variable. It it does not contain
a variable, it is equivalent to either ~- or £ . Finally, -1- m a y be replaced
by p V ~p and ± by p A ~p, so we may assume that the polynomial does
not contain any occurrences of -1- and _I_.) However, as W is white based,
formulae of the form @Z# or []Zv do not occur. Furthermore, if #~(p) > 0, we
replace the unmodalized occurrences of p by Ba BZ p, and the unmodalized
occurrences of ~p by [ ~ K~Z ~p. This replacement does not change #Z(p).
(The case #~(p) = 0 needs some attention. Here we replace p by ~ K~ p.
Then ~ is white equivalent to either V~ ½~ [ ~ ~ p or @ ~ a ~]~ P, P
nonmodal. Now we are down to the case of a formula of the form ~ p . ~
commutes with A and V, which leaves the cases BZp and [~Z~p to consider.
These are immediate.) Now, after this replacement, p is a lattice polynomial
over formulae of the form ~ , ~a~7. The latter occur black, so ~ is
intersubstitutable with ~ { and ~ V is intersubstitutable with B ~ . Finally,
~- is intersubstitutable (modulo equivalence) with -1-. So p is without loss
of generality of the form f((~a~li < n)) for some lattice polynomial f . Thus,
p can be substituted by the formula ~ f ( ( 8 i l i < n)). Therefore, p can be
reduced to the form [ ~ J for some basic ~. (~ is white based. Therefore,
by induction hypothesis, ~ is white-equivMent to a simulation transparent
formula 0. ~- has the form ~ ( C a s e A) or ~ E]a~ ( C a s e B), so
is white-equivalent to either X1 := ~ ½Z ½ ~ (Case A) or ~ Z ~ ~
(Case B). The latter is white-equivalent to X2 := ~ Z ~ 0 . Both X1 and X2

are simulation transparent, by assumption on 0. []
LEMMA 6.7.6. Let ~ be a monomodal formula in the variables {pi[i < n}.
There exists a bimodal formula X in the variables {p~,p~,p*[i < n} such that
(*): OL]-Sim ~P /-+ X~[Pl/P~, eZP~/P~, e,,,Pi/P*Ii < n]
X is called an u n s i m u l a t i o n of !o.
PROOF. There exists a simulation transparent ~- which is white-equivalent
to ~. X is obtained from v by applying the following translation outside in.
(e~x)~ := <>x~ (B~x)~ := Dx~
(e~p)~ := p* (-Te~p), := ~P*

(e~p)~ := p" (~e~p)~ := ~p"
This concludes the proof. []

The formula X is of course not uniquely determined by ~-, but is unique up
to equivalence. The proof of L e m m a 6.7.5 is actually a construction of X, and
so let us denote by Xs the particular formula t h a t is obtained by performing
t h a t construction. Now take a set A of monomodal formulae; put As := {~8 :
qp 6 A}. Assume that 92t is a simulation frame and (93~,/3, x) ~ a; A. Then
we have
where c~(p°) := p, cr(p °) :-- ~ ; , ~r(p*) := ~ ; . Now define a valuation/3 ° of

the set {pO,p.,p. : p E var[A]} by
9°(p °) := 9(p) n f°, 9°(p .) := ebg(p) n f°, 9°(p *) := e ~ ( p ) n fo
By definition of 9 ° ,
( ~ , 9, x) ~ ~; ~(¢) ¢, ( ~ , 9 °, x) ~ ~; ¢
Thus we conclude
(~,9°,x) ~ a; (Z~s) s
Define a valuation 9' on 9~s by 7(q) :=/3°(q) • x is of the form y° for some
y C ms; in fact, by construction, yO = x. By the previous results,
( ~ , 9 °, x) ~ a; (As) s ~* (~, % x) ~ Z~.
It therefore turns out that the satisfaction of A in a simulation frame at a
white point is equivalent to the satisfaction of As in the unsimulation of the
frame. The satisfaction of A at a black point is equivalent to the satisfaction
of E]/3A := {E]/3T : T E A} at a white point. The satisfaction of A at f* is
likewise reducible to satisfaction of B ~ A at a white point, which is defined

analogously.
Now assume t h a t ~ is a t h e o r e m of A, A m o n o m o d a l and consistent. T h e n
w -4 ~, a -4 T a n d / 3 -4 T are theorems of A as well. By consistency, w --4
can only be a t h e o r e m if it is a tautology. /3 -4 p is globally equivalent in
S i r e to a -4 E]~p. Thus, we can always assume t h a t an axiom is of the form
:= a -+ T for some T. (This follows independently from the surjectivity
of the simulation m a p and the fact t h a t an axiomatization of this form for
simulation logics has been given above.) Now ~ is falsified in a m o d e l based
on 9~ iff ~ is rejected at a white point iff Ts is rejected in a model based on
~I~ s .
We summarize our findings as follows. Given a m o n o m o d a l rooted S i r e -
frame ®, a set A, a valuation 7, we define % by %(q) : = v~(q), where q is a
variable of the form p°, p° or p*.
(qh,'),, x) ~ a A A (¢~s, %, x} ~ A ~
(®s, %}
Two cases m a y arise. Suppose t h a t ® contains only one point. T h e n ¢~s is

empty, and no formula is satisfiable in it. This case has to be dealt with sepa-
rately. Else, let ® have more t h a n point, and be rooted. T h e n satisfiability of
A in ® is reducible to satisfiability of either a; As or a; [ ] ~ A s or a; E]zAs. All
problems are reducible to satisfiability of a set A in ®s. Global satisfiability
of A in (~5,V) is equivalent to the global satisfiability of { a -4 ~ : ~ E A + } ,
where A + : = A; E]~A; ~ A .
THEOREM 6.7.7. Let A be a monomodal logic in the interval [Sim, T h i e f ] .
Then A is axiomatizable as A : = K1 ® {a --4 5 : 5 C A} for some A . Further-
more, As = K2 • Z~s @ (D~A)s • ( B ~ A ) s .
T h e following properties are now shown to be invariant u n d e r simulation:
finite axiomatizabiIity, (strongly) recursive axiomatizability, O-axiomatizability.
DECIDABILITY AND COMPLETENESS.
PROPOSITION 6.7.8. Let A be a set of bimodal formulae, ~ a bimodal

formula, and ® a bimodal logic. Then
A i-e ~ 4=~ a -+ A s t-e8 a -4 ~s
A II-e p 4:~ a -4 A s I~-e= a --+ ~ s
PROOF. Assume A F-o ~. Let 9Y~ be a rooted and differentiated O s -

frame, 5 a valuation and x a world such t h a t (gJt, 5, x} ~ a --+ A s. T h e n
9Yt ~ ~ s for some rooted ® - f r a m e ~ . Two cases arise.
C a s e 1. (93q, 5, x) ~ --a. T h e n ( ~ , 5, x) ~ a -+ ~s.
C a s e 2. x ~ a. T h e n x = y° for some y E b. and a valuation V on ~ such t h a t
316 6. R e d u c i n g P o l y m o d a l Logic t o M o n o m o d a l Logic
7(p) ° = 5(p) A b°. Then (~3,7, Y) b A. By assumption, (~3,7, Y) b W and so

(gYt, 6, x) ~ a -+ ~*. Thus a --+ A s ~-e* a -+ Ts. Now assume that the latter
holds. Take a bimodal model (~3,7, y } ~ A. Then (~3s,7 s,y°) b a --+ A*
and so by assumption (~3s,7*,y °} ~ a --+ ~* and so (~3*,7",y °) ~ ~s. From
this we get (~3,7, y } ~ T. Hence A ~-o ~. Now for the global deducibility
assume A [~-o ~. Let (gYt,7} V a --+ A*. Then we can assume that 9Yt = ~3"
for some ~3 and that 7 = 5*. We then have (~3,5) ~ A. By assumption,
(~3,5} ~ ~. From this we get (~3",5 *} ~ a -+ ~. Hence a -+ A s I~-o. a --+ ~s.
Assume the latter holds and let (~3,5} ~ A. Then (~3",5 s) b a -+ A* and so
(~3s, 5") ~ a -+ ~*. From this we obtain (~3, 5) ~ ~. Hence A [~-o ~. []
PROPOSITION 6.7.9. Let A be a monomodal logic, A a set of monomodal

formulae and ~ a monomodal formula. Then
A ~-A ~ ¢:> A~ ~A~ ~*
and ([]3A)s
and (~A)s
PROOF. Let A ~-A W. Suppose that (~3,5, x} ~ As. Then ( ~ * , 5 * , x °)

a A A. By assumption, (~3*,hS,x °) ~ a A ~. Hence (~3,5, x) ~ ~s. This
shows As ~-h8 ~s. Similarly, ([33A)s ~-h~ (7~3A)s and ([]~A)s ~-h~ ([3~A)s
are proved. Now assume that all three obtain. Let (9)I, 7, x} ~ A.
C a s e 1. x = u °. Then (gXs,%,u) ~ A, and so (9~t,,%u) ~ ~s- Thus
°)
C a s e 2. x = u ' . Then (9)2, 7, u°} ~ ~3 A. As in Case 1. we get (gX, 7, u°)
E]3~. Therefore (93t, 7, u" ) > ~. C a s e 3. x -- u*. Then (gX, V,u °)
[ ~ A . As in Case 1. we get (9~t,V,u °) ~ [ 3 ~ . Hence (gJt, V,u* } ~ ~.
Hence, A ~-n ~- For the global deducibility let A I~-A ~. Assume (~,5}
As; (B~A),; (B~A),. Then (q3s,5 *} ~ A. By assumption, (~3s,5 *) ~ ~; this
gives (~, 5} > ~,; ([~3~),; ( ½ ~ ) s . Therefore
zx,; (B=zx)s l+a. (B=v)s.
Assume finally that the latter holds. Let (gYt,7 ) ~ A. Then (97t,,%)
A; (e~A)s; ( [ ~ A ) s . By our assumption, (gYts,%) ~ ~,; (~3~),; ( 7 ~ ) s . This
gives (ff)~,7) ~ ~. Hence A I~A ~. []
As a consequence of these two theorems, the following properties are
invariant under simulations: local decidability, global decidability, local com-
pleteness, global completeness, local finite model property, global finite model
property, weak compactness, strong compactness.
COMPLEXITY. Let ~ be a bimodal formula. Then ~s can be computed in
quadratic time. Moreover, its length is O(t~1 ). So, if O s is if-computable
(where ff is for example NP, PSPACE or EXPTIME, but any class invariant
under linear time reductions will do) then so is (9. Similarly, if (gs is globally
C-computable, so is (9. Conversely, let X be a monomodal formula. Going
carefully through the construction one can show that it takes polynomial time
to construct Xs from X- Moreover, tXsl is of size O(IxI ). It follows t h a t if (9 is
in e, so is (9~. Therefore, (9 and O ~ belong to the same complexity class, as
satisfiability problems are linearly interreducible. Hence the following proper-
ties of logics are invariant under simulation: local and global C-computability,
local and global C-hardness and local and global C-completeness.
PERSISTENCE. Assume that h is r-persistent. Then let if2 be a refined M -
frame. We know that 9~s is a A-frame, and is refined by Proposition 6.6.3.
So (gYt,)~ is a A-frame, by r-persistence of h. But (gJt,)~ = (ff2~),, so 9Jt~
is a AS-frame as well. The same reasoning establishes preservation of dr-
persistence, d-persistence and c-persistence. And analogously the reflection
of these persistence properties is shown. Only with properties such as a -
canonicity one has to be careful, since the unsimulation uses more variables.
Finally, if ~ is constant, so is T ' . And if ¢ is constant, so is Cs. This
shows invariance of g-persistence as well. The following properties have been
shown to be invarariant under simulation: g-persistence, dr-persistence, r-
persistence, d-persistence, ~-canonicity ( t; infinite).
ELEMENTARITY. The next l e m m a shows that a bimodal logic is ~rp-elemen-
tary iff its simulation is ~rp-elementary.
LEMMA 6.7.10. Let ~ be a class of bimodal Kripke-frames, £ a class of

standard simulation frames. Then
UpS' = (Up.~)"
UpS, = (up,~),
It is easy to define a translation for elementary properties under simula-

tion.
~(x) := (3y ~ ~)~(=)

Z(z) := ~ ( = ) A (vy ~ =)~(x)
t h e s e formulae define the sets i f , fw and f t in a S i r e - f r a m e . Now put
A <=).
=c.ear(,~)
(x-y)/ :-- x-y

(x<ly)S := z<~y
(~ A ~)~ := ~(A ~
(~)~ := ~(~s)
((~)~)~ := (~)(~(~) A ~ )
It is not difficult to show that for a sentence ~,
It is not hard to see that the class of S i m - K r i p k e frames is elementary. This

also follows from the fact that axioms for Sire are Sahlqvist.
The case of unsimulating elementary properties is more complex t h a n the
simulating part. Take a formula ~ in the first-order language for 1 - m o d a l
frames. We may assume (to save some notation) that the formula does not
contain V. Furthermore, we m a y assume t h a t the formula is a sentence, t h a t
is, contains no free variables. Finally, we do not assume t h a t structures are
nonempty. We introduce new quantifiers q~, ~ C {c~,~,w}, which are defined
by
(~bx)~(x) := (3x)(~(x) A ~(~))
Furthermore, for each variable x we introduce three new variables, x b, where
E (e,~9,~}. Now define a translation ( _ ) t as follows.
(~ A ¢)t := ~t Act
(~ v ¢)t := ~t v Ct
It is clear that ~ and ~t are deductively equivalent. In a next step replace

(qZx#)~,(x #) by
( ~ ) ( ~ z ~ # / > x")~(~ ~)
and (3'~z'~)~(x '~) by
(~)(~/> ~)~(~)
Call ¢~ the result of applying this replacement to ¢. It turns out t h a t (in the
first-order logic of simulation frames)
T h a t means that ¢~ and ¢ are equivalent on all frames ~ where ~5 is not

empty. In ¢~ the variables x ~ and x ~ are bound by a restricted quantifier
with restrictor x~; x ~ in turn is bound by 3 ~. To see whether such formulae
are valid in a frame we may restrict ourselves to assignments h of the variables
in which x ~ is in the b-region for each b and each x, and furthermore h(x ~) >~
h(xZ). In a final step, translate as follows
:=
:=
:=
:=
For atomic formulae, ~ is computed as follows:
x~y ~ w b t x~y ~ w b t
w x~y x&y b w x&y ± ±
$ b x~y x~y ± $ b ± x~y ±
t L ± ± t ± ±
Let ¢ be a subformula of Tt* for some sentence T. Let ~5 be a i - m o d a l

simulation frame. Then ~5 is isomorphic to ~s for some bimodal frame
(for example, ~ := ~hs) and therefore ~5 has exactly one point in the w-region.
Suppose ~5 ~ ¢[h]. Then we may assume that h(z ~) is in the a-region for each
b, and that h(x ~) -< h(x~). Now put k(x) := h(x~). It is verified by induction
on ¢ that ~5 ~ ~b[h] iff ~hs ~ ¢*[k]. On the other hand, if k is given, define h
as follows: h(x ~) := k(x), h(x ~) is the unique world u in the ~-region such
that k(x) >~ u, and h(x ~) is the unique world in the w-region, h is uniquely
determined by k, and again it is verified that ~5 ~ ¢[h] iff ~hs ~ ¢*[k]. In
particular, for ¢ = Wt* we get ~5 ~ ¢ iff ~hs ~ ¢~. Now we return to ~. We
have ~ -- ~,A (Vx)w(x). V .~ A (3x)~w(x). The first formula is either equivalent
to ± ( C a s e 1) or to (Vx)w(x) ( C a s e 2). C a s e 1. P u t We := (Wt*)*. Then
® ~ ~ iff ~hs ~ ~,. C a s e 2. Put ¢~ := ((Vx)~(x - x)) V (~t*)*.
PROPOSITION 6.7.11. Let )C be a class of simulation frames. If )C is ele-

mentary (A-elementary) so is :Es.
The following properties have been shown to be invariant under simula-
tion: ~rp-(A )-elementarity.
SAHLQVISTLOGICS. Likewise, by Proposition 6.6.13, if ® is Sahlqvist, so is
O ~. For the other direction, we need to be a little bit more careful. Assume
that A is a monomodal logic and Sahlqvist. Then by Theorem 5.5.8 it is
axiomatizable by formulae of the form ~ -+ ¢ where ¢ is positive and T is
composed from strongly positive formulae using A, V, and 0j. Now (~Z -). "¢)s
is the same as ~s -+ ~s- It is not hard to see that the unsimulation of a
positive formula is positive, and that the unsimulation of a strongly positive
formula is strongly positive. Moreover, ¢s is composed from strongly positive
formulae using A, V and Oj. So, it is Sahlqvist.
The rank as well as the special rank do not increase under simulation.
This can be shown by observing a few equivalences. Namely, note that the
following holds.
( ( v y E> x),~) s - (v~/>

((vy. =),~)s _ ( v ~ / > x)(V~/> y)(Vy /> ~)(/3(~) A 13(y) A o~(y). --4 .(~s)
((~y. =),5) ~ - (3~/> • )(3~/> y)(3y /> ~)(Z(~) A Z(~) A ~(y). -~ .a s)

(x•y) -- (V~>/
(x • ~) - (~/> x)(~"/> y)(/3(~) A/3(~') A ~ ~< ~')
So, we translate a restricted V by a sequence of restricted V and a restricted
by a sequence of restricted 3. The atomic formulae x <l y and x • y may
be translated by existential formulae or universal formulae, and so neither
the rank nor the special rank are changed under simulation. Thus the special
rank of t0s is equal to the special rank of ~. Conversely, let ¢ be given, a
monomodal Sahlqvist-formula of special rank n. Then it is likewise checked
that the unsimulation % is a Sahlqvist formula of rank at most that of ¢.
Moreover, tb is deductively equivalent to a formula ~ whose unsimulation
has rank equal to that of ¢. (The construction of the unsimulated formula
actually proceeds by producing ¢ and unsimulating ~ rather than ¢.) Now,
Sire is ~ - e l e m e n t a r y and of special rank 0. Hence, if K2 G A is a Sahlqvist
logic of (special) rank n, then its simulation S i m @ A s is of (special) rank
n, the rank being the maximum of the ranks of S i m and of A. (Notice that
the property of being Sahlqvist of rank n are here properties of logics, not of
the actual axiomatization; otherwise the claim would be false.) Therefore the
properties being SahIqvist of rank n and being Sahlqvist of special rank n are
invariant under simulation.
INTERPOLATION. Finally, we take a look at interpolation. It is easier to show
that interpolation is transferred from monomodal to bimodal logic using the
characterization via the amalgamation property than using the simulation
of formulae. However, this has the disadvantage that an interpolant is not
explicitly constructed. Take a bimodal logic ®. Assume that O s has global
interpolation. Let *1 : %0 "-~ ~1 and '2 : ~ o ~ !t32 be embeddings of O -
algebras. Then ,~ : ~ ~-* %~ and ~ : ~ ~-~ ~ are embeddings of the
simulations. Now, by the assumption that ®s has global interpolation we get
a 9.13 and embeddings ~1 : !I3~ ~-* 9~3, ~2 : !13~ ~-* 9/3 such that ~1 o ,~ = ~2 o ,~.
6.7. P r o p e r t i e s of the Simulation 321
Let ~B3 :-- (9-13)s and p u t ~1 := ((1)s and e2 :--- ((2)~- T h e n

~1 o ~1 = (¢1), o ( ~ ) .
-- ((1 o ~)~
= (~2 o 4 ) s
= ((2)~ o (4)~
([2 0 L2
Therefore, the variety of O - a l g e b r a s has a m a l g a m a t i o n . Suppose t h a t the

variety of OS-algebras has s u p e r a m a l g a m a t i o n . T h e n we can show t h a t the
variety of (9-algebras has s u p e r a m a l g a m a t i o n , too. For assume t h a t el(b1) _<
e2(b2). T h e n via the identification of elements with subsets of the o-region
we conclude (e~(b~)) ° < (e2(b2)) ° from which also e~(b~) < 4(b~). T h u s
there exists a ao such t h a t b~ _< (l(ao) and (2(ao) _< b~. Intersecting with
a we get b~ _< ( l ( a o ) NC~ and (2(ao) f l a _< b~. Now, p u t bo := a o A a .
T h e n (l(bo) = ¢1(a0) 71 a as well as (2(bo) = (2(ao), from which follows t h a t
b~ < ~l(b0) and ~(b0) < b~.
Assume t h a t O has local interpolation. P u t [- :-- F-e.. Assume ~, [- ¢.
T h e n (a) a --+ ~ ~- a --+ ¢. Moreover, we also have ( b ) / 3 --+ ~ ~-/3 --+ ¢ and
(c) w --+ ~ [- a; --~ ¢. (b) can be reformulated into (b') a --+ ~ [- a -+ ~Z¢"
T h e cases (a), and (b ~) receive similar t r e a t m e n t . Take (a). We have ~ ~-e ~ .
There exists by a s s u m p t i o n on ® an X such t h a t var(x ) C_ var(p~) ?1 var(¢~)
and ~ ~-e X [-e X~. T h e n a -+ (~s)~ ~- a --+ X ~ ~- a --+ ( ~ ) ~ . Now take the
substitution cr as defined above. T h e n
(~ ~ ) ~ -~ ~ ( ( ~ ) ~ ) ~ ~ -~ ~ ( ~ ) ~ ~ -~ ~ ( ( ¢ ~ ) ~ ) ( ~ ¢ ) .
P u t ( :-- cr(X~). T h e n we have var(a --+ () C var(~) ~ vat(C) and a -+
[- a -+ ( [- a -+ ¢. Similarly, we find a formula ~ such t h a t a --+ ~ ~-
a --+ ~ ~- a - + ~}Z¢" H e n c e / 3 --+ ~ [- ~ --+ ~}~U [- ~3-+ ¢. For (c) we can
appeal to the fact t h a t classical logic has interpolation to find a 0 such t h a t
a -+ Q~W ~- a -+ Q~0 [- a --~ ~}~¢. (For in the scope of Q~, W and ¢ reduce
to n o n m o d a l formulae.) T h e n put A :--- (a --+ () A (/3 --+ ~ ) A (w -+ 0). It
follows t h a t ~ k A [- ¢. Thus ~* has interpolation. E x a c t l y the same proof
can be used for global interpolation (no use of the deduction t h e o r e m has
been made). So, local interpolation and global interpolation have been shown
to be invariant under simulation.
THEOREM 6.7.12 (Simulation T h e o r e m ) . The simulation map A ~-+ A s is

an isomorphism from the lattice of bimodal normal logics onto the interval
[Sire, Th [~-]] in ~ K1 preserving and reflecting the following properties of log-
ics.
* finite, strongly recurs@e, recursive axiomatizability,
* codimension n,
* tabularity,
* local and global decidability,
* local and global C-computability (C-hardness, C-completeness),
* local and global completeness,
* local and global finite model property,
* local and global interpolation,
* strong and weak compactness,
* g-, dr-, r-, d- and c-persistence,
* being a Sahlqvist logic of (special) rank n.
E x e r c i s e 217. Show the remaining claims in the Simulation Theorem.
E x e r c i s e 218. Show by direct calculation t h a t ( - ) ~ is a [_]-homomorphism.
E x e r c i s e 219. Using the simulation, we can simulate the fusion operator

inside ~ K1 as follows. Define A ~ O := (A ® O) s. By the transfer results A ~ O
has finite model property (is complete etc.) iff A and O are. Now suppose
t h a t A and O are logics with finite model property such t h a t their join, i. e.
A U O, fails to have the finite model property. Show that A ~ O is a logic with
finite model property such that adding (Dp ++ rap) s fails to have the finite
model property. Clearly, S i m • ([3p ++ rap) s has the finite model property.
Thus we have shown that if finite model property is not preserved under joins,
there is a Sahlqvist logic of special rank 1 which fails to preserve joins. We
will later refine this result somewhat; but already it shows t h a t very simple
axioms can induce rather drastic effects if added to a logic, on the condition
that such effects can occur at all.
6.8. S i m u l a t i o n a n d T r a n s f e r - - S o m e G e n e r a l i z a t i o n s
We have proved the simulation and transfer theorems only for the case of
two monomodal logics. This suggests two dimensions of generalization. First,
we can generalize to the simulation and transfer of several logics with one
operator each, and of two logics with several operators; of course, the two can
be combined. Let us begin with the fusion. Obviously, if we can generalize
the transfer theorems to the case of two logics with several operators, we
have -- by induction -- a transfer theorem for the independent fusion of n
monomodal logics. The case of infinitely many logics needs to be discussed
separately. In fact, it does give rise to a number of exceptions which must be
carefully noted.
Now let us study the case of two polymodal logics A and O. The con-
structions of an ersatz is straightforwardly generalized. The only difficulty
we face is that the proof of Consistency Reduction makes use of MAKINSON'S
Theorem. Recall, namely, that in the model building procedure we want to
6.8. Simulation and Transfer - - Some Generalizations 323
insert certain small models at an internal set of nodes. W h a t we therefore

need is t h a t A ® (9 allows for at least one finite model, if for example transfer
of finite model p r o p e r t y is desired. This is true if only A and (9 b o t h have
a finite model. Namely, the following construction yields frames for the fu-
sion. Take a frame ~ = ([,F) for K~ and a f r a m e 6 = (g,G) for K~. We
define the K ~ + ~ - f r a m e ~ ® 6 as follows. T h e underlying set is f x g. For
O~ ( K, (Vl, Wl) < ~ (V2, W2) iff Wl = W2 and vl <~a v2. For /3 < A we put
<vl,wl> ~,~+~ ® (v2,w2) iffvI = v2 and wl <~ w2. Finally, F ® G is the set of all
unions of sets a®b = {(v,w) : v C a , w E b}, where a E F and b E G. It is easy
to check t h a t this is indeed a frame. For i f a < n then ¢ ~ ( a ® b ) = ( ¢ ~ a ) ® b
and if/3 < A then ¢~+~(a @ b) = a ® (¢zb). The frame ~ ® 6 is called the
t e n s o r p r o d u c t of ~ and 6 .
LEMMA 6.8.1. Let ~ be a A-frame and 6 be a (9-frame. Then ~ ® 6 is

a A ® (9-frame.
PROOF. T h e projections of a set a are defined by
La-Jl := {V:(~W)((V,W) E a ) }
ua~2 := {W:(3V)((V,W) Ga)}
Let /3 be a valuation on ~ ® 6 . Define a valuation ffl(p) : = {v : L/3(p)J1},
and 72(P) : = {w : L/3(p)J2}. By induction it is shown t h a t if ~ uses only
modalities D~, a < n, then L~(~)J~ = VT(P) and if p uses only modalities
n + / 3 then L~(~)J2 = ~ ( ~ ) . Hence, if ~ is a t h e o r e m of A, it is a t h e o r e m
of ~ ® 6 , and if it is a theorem of (9, then it is a t h e o r e m of A ® (9, under
suitable identification of modalities. This proves the theorem. []
Let us note t h a t if we have infinitely m a n y logics, this m e t h o d does not
allow to find a finite model for their fusion. However, suppose t h a t there is
a n u m b e r m such t h a t all logics have a frame of size at most m (e. g. if
t h e y are all m o n o m o d a l ) then we can construct a finite model for the infinite
fusion. T h e basic idea is t h a t there are only finitely m a n y choices of relations,
so r a t h e r t h a n taking the infinite p r o d u c t as above, we will collapse a lot
of relations into a single one. Let us first observe t h a t if A is a logic with
infinitely m a n y operators which has a frame ~ of size < n, t h e n the theory of
contains m a n y axioms of the form (}~p ++ (}~p, a,/3 < n, saying t h a t the
relations a a n d / 3 are identical. It is then possible to regard Th ~ as a finite
o p e r a t o r theory, and to compress ~ into a frame containing finitely m a n y
relations. Call an u n i n d e x e d f r a m e a pair (f, R) where R is a set of binary
relations over f . T h e m a p ( f , ( < ~ : a < n}} ~-~ ( f , { < ~ : a < n}) is called
the c o m p r e s s i o n m a p . Given n : = ~f there are finitely m a n y unindexed
frames of size n, for there are at most 2 "2 different binary relations, and a
compressed frame contains a subset of them. Notice also t h a t if [ and g are
finite frames of size n, ~ a frame for A and g a frame for (9, then there exists
an unindexed frame of size n for A ® O. For we can assume the underlying

sets to be identical. Then we interpret <~, a < t~ by <~ and <3~+~ by <~Z,
and then compress. Now, for the final result, take a set of logics Ai, i E I,
such that there is a number n such that each Ai has a frame of size < n. Then
take the unindexed tensor product. It is finite. Now uncompress the frame,
assigning each modality its relation.
LEMMA 6.8.2. Let Ai, i C I, be an indexed family of polymodal logics. Let
n be a number such that every Ai has a model of size ~_ n. Then ~ i c I Ai has
a model of size ~ n!.
DEFINITION 6.8.3. A property • of logics is said to be p r e s e r v e d u n d e r
f u s i o n if for each family Ai, i E I, of consistent logics, ~ i ~ I Ai has ~P when-
ever each Ai has T. If we can infer [P for the factors from that of the fusion
we say (P is reflected u n d e r f u s i o n . [P t r a n s f e r s u n d e r f u s i o n iff it is
both preserved and reflected under fusion.
THEOREM 6.8.4 (Transfer Theorem). The following properties are trans-
ferred under fusion.
• finite and recursive axiomatizability, provided the indexed collection is
finite,
• g-, df-, ti-, r-, d- and c-persistence,
• local and global decidability,
• local and global finite model property, under the condition that there
exists a uniform bound for the smallest model of the factors,
• local and global completeness,
• weak and strong compactness,
• local interpolation, local Hallddn completeness.
Of course, at some points things can be improved. Finite axiomatizability
is reflected under fusion even when I is infinite and the same holds for recursive
axiomatizability. Tabularity transfers exactly when almost all factors are
trivial.
Now let us turn to simulation. We can generalize in two directions. First,
we can simulate an extension of K~+~ by using only the modalities of K~.
(Notice that we cannot simply simulate K~+~ for different ~ and /k. But if
t~ ~/k, it is possible to adjoin to one of the logics a number of trivial operators
to make the reduction work in this case.) In that case we double up the frame
and pick one modality to play the complicated role of encoding the double
structure. Notice that the reduction can only be iterated a finite number of
times, unlike the fusion. But we can also wrap up the modalities in one step
as follows. For n modalities, f n consists of n different copies if, i < n, plus
n - 1 extra points , 0 , . . . , , n - 2 . Given x E f we write x i for the corresponding
point in f i Then the relation ~ j is coded among the x j. The points ,i are
related by *J ~< ,k iff k = j - 1. Furthermore we have x j <~ ,k iff j = k. This
serves to distinguish the x j from the x k, k ~ j. Finally, for j ¢ k we put

x j ~ yk iff x ----y. Notice t h a t x j ~< x j iff x <~j x. T h e n the formulae a j and
wj are defined as follows.
i
~d(j) := ~::~j+l~_ A (~ T j < n - 1
a(j) := ~w(j + l) A ~ w ( j ) j < n- 2
:= e (n-2)
a(n- 1) := h/<~-l~w(J) A-~a(j)
As before, the logic of simulation structures can be axiomatized. It is the best
s t r a t e g y to define the operators B~(j) and ½~(j). These o p e r a t o r s all satisfy
a l t l . Moreover, we have a ( j ) --+ ~ ( i ) T for i ¢ j. In total, we have the
following postulates.
a(i) -+ ~ a ( j ) i~ j
a(i) A ~ ( j ) p . --+. B~(j) p i ¢ j
e (i)
c~(i) -+ -~5)w(j) i ~ j
a(i) A ~ ( j ) p . -+ 5)~(k) ~ ( j ) P i ¢ k, j ¢ k
<4
- w(j) A p. --+. ~<4 (02(j) --~ p)
for j ¢ k. This defines Siin(n). T h e special postulates ensure t h a t we can

move from xJ to x k in one step. T h e price to be paid for this is t h a t the special
rank of S i r e ( n ) is 1 for n > 0. There is a simulation which avoids this (see the
exercises), but it has other disadvantages. We put I n c ( n ) := K . a l t l ® ~ ± ,
which is the same as ( K s ® ± ) s .
THEOREM 6.8.5 (Simulation T h e o r e m ) . The simulation map A ~-+ A s is

an isomorphism from the lattice of normal n-modal logics onto the interval
[Sire(n), Inc(n)] preserving and reflecting the following properties.
* finite, strongly recursive and recursive axiomatizability,
* tabularity,
* local and global decidability,
* local and global C-computability (C-hardness, C-completeness),
* local and global completeness,
* local and global finite model property,
* local and global interpolation,
* strong and weak compactness,
* g-, dr-, r-, d-, and c-persistence,
* being a Sahlqvist logic of (special) rank n,
* being weakly transitive,
* being of bounded alternativity.
To close this chapter we will generalize the conservativity result of this

chapter to consequence relations. Apart from being interesting in its own
right, this will also sharpen our understanding of the methods used previ-
ously. As before, the situation we look at is the fusion of two m o n o m o d a l
consequence relations and the simulation of a bimodal consequence relation
by a monomodal consequence relation. The results are general, however. The
simplification is made only to avoid baroque notation and to be able to con-
centrate on the essentials of the method. As is to be expected, the proofs
of some theorems are more involved. In particular, to show t h a t the fusion
of two consequence relations is conservative requires other methods because
an analogue of Makinson's Theorem is missing. (See Section 3.9.) For the
purpose of the next definition recall that if R is a set of rules, [-R denotes the
least modal consequence relation containing R. We have the two translations
and 7-m from T1 into T2, defined in Section 6.2. 7~ replaces [] by [] and
~-m replaces [] by I .
DEFINITION 6.8.6. Let [-1 = [-R and [-2 = [-s be two m o n o m o d a l con-
sequence relations. P u t T := 7-[][R] U ~'m[S]. Then [-1 ® [-2 is a bimodal
consequence relation defined by [-1 ® [-2 := [-T. [-1 ® [-2 is called the f u s i o n
of [-1 and [-2.
Let [- be a bimodal consequence relation. Define two consequence rela-
tions, ([-)G and ([-)m by
[-D := r~l[[-]
[-= := r,~l[[-]
We call [-[] and [-m the w h i t e and b l a c k r e d u c t of [-, respectively.
PROPOSITION 6.8.7. Let [-1 and [-2 be monomodal consequence relations.

Then [-1 ® [-2 is the least consequence relation such that ([-1 ® [-2)[~ : [-1
and ([-1 ® [-2)** = [-2"
Our first main result is the analogue of Theorem 6.2.3.
THEOREM 6.8.8. Let either ~-1 be inconsistent or else ~-2 be consistent.

Then
([-1 ® [-2)D = [-1
First, let us deal with the easy case. If [-1 is inconsistent then so is
[-1 ® [-2, and the fusion is clearly conservative. So, let us from now on assume
t h a t [-1 is consistent. In that case, [-2 is also assumed to be consistent. Now
suppose ff)I = (Pl, D) and 9~ = (!D, E} are matrices such t h a t [-1 C [-~ and
[-2 C [-~. Then construct a bimodal matrix 9:~ ® 9~ as follows. It is based on
the tensor product of boolean algebras. This tensor product is formed from
so-called b a s i c t e n s o r s . These are pairs (x, y} E A × B, denoted by x ® y.
T h e (boolean) tensor algebra is the boolean algebra generated by the basic

tensors with respect to the following rules.
(x~_ ® y~_) n (x2 ® y2) := (x~ n z:) ® @1 N Y2)

x@O := 0@0
0®y := 0®0
( ~ ® y) u (~: ® y) := ( ~ u ~:) ® y
- ( ~ ® y) := (-x) ® y u • ® (-y) u (-~) ® (-y)

It can be shown t h a t an element of 94 ® ~ is of the form Ui<n xl ® y~, where
xi > 0 as well as Yi > 0 for all i < n. (Moreover, if n = 0, t h e n this disjunction
denotes the tensor 0 ® 0.) For a basic tensor x ® y = 0 ® 0 iff x -- 0 or y = 0.
This defines the t e n s o r p r o d u c t of boolean algebras.
LEMMA 6.8.9. The following holds.

(A.) xl ® Yl = x2 ® Y2 iff
(1.) 0 e { x l , x 2 } and 0 e {Yl,y2} or
(2.) X l = X 2 and yl = Y2.
(B.) x l ® y l _ < x 2 ® y 2 iff
(1.) xl = 0 or
(2.) y l = 0 o r
(3.) Xl, x2, yl andy2 are all distinct from 0 a n d x l ~_ x2 a n d y l ~ Y2.
PROOF. Clearly, (A.) follows from the construction. Therefore we need

to show (B.). If xl ~ x2 and yl _~ y2, then
(x~ ® yl) n (x2 ® y2) = (xl n =:) ® (yl n y2) = =1 ® y,
So, x l ® y z ~_ x2®y2. Now let xz = 0 or Yl = 0. T h e n x l ® y l = 0 ® 0 ~ x2@Y2.
Let all elements be distinct from zero. T h e n xl ® Yl _< x2 ® Y2 implies
( x l ® y I ) n ( x 2 ® Y : ) = x l ® y l . So, (x, n x 2 ) ® ( y , n y : ) = x , ® y l . Now,
xl ¢ 0 and yl ¢ 0, hence xl n x2 -- xl and Yl ~ Y2 = Yl. This implies x2 ¢ 0
and Y2 ~ 0, and moreover xl _~ x2 and Yl _< Y2- []
T h e b o t t o m element is 0 @ 0, and the top element is 1 ® 1. T h e action of

[] and • is defined via their duals, 0 and ¢, in the following way.
0(~ @u) := (ez)ey
¢(~ ® u) := • ® (ey)
This defines now the f u s i o n of two m o n o m o d a l algebras. Finally, let D ® E
denote the least filter containing the elements x ® y, x E D and y C E . It can
be shown t h a t this is likewise the filter generated by x ® 1, x C D, and 1 ® y,

y E E. Finally, put
~ ® 9 ~ := ( 9 2 @ ~ B , D ® E )
The first fact to note is
LEMMA 6.8.10. Assume that 0 ~_ D and 0 ~ E. Then x ® y E D ® E iff
xED andyEE.
PROOF. From right to left holds by definition of D @ E. From left to
right, assume that x ® y E D @ E . Then there is x ~ E D and yt C E such
t h a t x @ y > x ~@y/. Since x ~ > 0 andy~ > 0 by assumption o n D and E,
x >_ x ~ and y _> y/ from which x E D and y E E, since D and E are filters,
and x ® y E D ® E, by definition of D ® E. []
LEMMA 6.8.11. The map x ~+ x ® t is an in injective homomorphism
from 92 into 92 ® ~ [ {1, - , G, V]}. Dually, the map y ~-~ 1 ® y is an injective
homomorphism from ~ into 92 ® f13 I {1, - , N, " } .
This is left as an exercise.
LEMMA 6.8.12. Assume that 9~ -- (~B,E) is a monomodal matrix and
0 q! E. Then (~-~®m)[] C_ F-~.
PROOF. Assume that 9)~ = (92, D). Let p = (A, ~) ¢ ~-~. T h e n 0 ¢ D.
We want to show that ~ [ p ] := (T~[A],z~(~)) d ~-~®~. To t h a t end, let/3
be a valuation such that ~[A] C_ D but ~ ( f ) ¢ D. Then ?(p) :=/3(p) @ 1 is a
valuation into ~7~ ® 9~. Moreover, by L e m m a 6.8.11, for any formula X E T1,
~(v[](Z)) -- ~(X) ® 1. It follows with L e m m a 6.8.10 that 7(TD(X)) E D ® E
iff ~(X) E D. Hence, ~[vD[A]] C D ® E , and 7(7~(~)) ¢ D ® E . So,
~-D[P] ¢ k0n®m. []
Recall from Section 3.5 the following notation. Given a modal algebra ~ ,
~B~ denotes the algebra of elements _< z. This is called the relativization of
LEMMA 6.8.13. The map r~ : ~ -+ ~ N (1 ® z) is a homomorphism with

respect to the operations 1, - , O, [~.
PROOF. Let ~ := 92 ® ~ z . We have 1~®~ = 1 ® 1, and 1 ¢ = 1 ® z =
( l ® l ) n ( l ® z ) . It is easy to check that rz commutes with N and U. The latter
is useful in order to derive that r ~ ( - ~ ) = -r~(~). Namely, we can reduce this
now to the case where ~ = x ® y. In that case,
r~(-(x®y)) = r~((-x)®yUxe(-y)U(-x)®(-y))
= ( - x ) ® (y n z) u x ® ( ( - y ) n z) u ( - x ) ® ( - y ) n z
= - 6 ( x ® ( y n z))
= -%~(x®y)
6.8. Simulation and Transfer -- Some Generalizations 329
Finally,
r=(O(x ® =
= (~x)® (~nz)
= O(x ® (~ n ~))
= O(r (x ®
[]
LEMMA 6.8.14. (~-~®~)[] D b~.
PROOF. Let p = (A,~) be a rule of monomodal logic. Assume that

TD[p] ~ F ~ ® ~ where (P2,D) and 92 = ( ~ , E ) . We have to show that p ~ F ~ .
To that end, note that there is a valuation 7 into ff)I®92 such that ~[A] C_ D Q E
but ~(~) ~ D ® E. Let the variables occurring in A and ~ be Pi, i < k. We
may assume that there exist elements cj, j < m, of ~ such that cj > 0 for all
j <rn, ciNcj =0iffi=jandforcertainaj i cA,
7(Pi) = U aj @ cj
j<rn
(Of course, some or all of the aji may be zero.) Now, put
i
/3y (Pi):= aj
Then 7(Pi) f~ej =/3j (p~) ®cj. It follows that ~(7"D(X)) Ncj = ~(X) ®cj for every
X E AU{~}. Finally, observe that rc~ [D®E] = DQEcj, where Ec~ is the filter
generated by y N c j , y E E. Notice that it may well happen that E~j = B, for
example if - c j E E. Now, for X E AU{~}, 7(7-[~(X))Ncj ~j(X)@Cj C D®Ec~
=
iff/3j (X) C D. Now, ~[v5 [A]] _C D ® E implies that/3j [A] C D for all j < m
and ~(~-~(~)) ¢ D ® E implies that for some j < m, /3j(X) ff D. Hence
COROLLARY 6.8.15. Let 92 = ( ~ , E I be a matrix such that 0 ff E. Then

(~-~®~)~ = ~ .
COROLLARY 6.8.16. Let ~ be a matrix for F-1 and 92 a matrix for ~-2.
Then ~ ® 92 is a matrix for ~-1 ® ~-2.
We are now in a position to prove Theorem 6.8.8. Assume that F-1 is con-
sistent. Then ~-2 is consistent, according to the assumptions of the theorem.
Hence there exists a matrix 92 = (~, E / such that ~-~ _D b~. Now assume that
p ~ ~-~. Then there exists a matrix 92 such that p ¢ ~ . Then T~[p] ¢ ~-~®~.
Thus, ~ [ p ] ~ ~-~ ® ~-~, since ~ ® 92 is a matrix for the fusion. It follows
that p ¢ (~-~ ® ~-2)[~. Hence (~-~ ® ~2)c C ~-~. The converse inclusion is
straightforward from the definition. The theorem is proved. As an immediate
consequence we obtain from it the fact that the fusion of logics is directly
connected with the related consequence relations.
THEOREM 6.8.17. Let A and (~ be monomodal logics. Then

[-A®O : [-A ® ~-0 [I-A®@: I~-A® H-O
E x e r c i s e 220. Prove Theorem 6.8.17.
E x e r c i s e 221. Is Halld6n-completeness reflected under simulations?
E x e r c i s e 222. Propose a different simulation by putting x j ~ x k iff j = k

(and x <~j x) or k = j + 1 (mod n)). Show that in this simulation the special
rank of the simulating logic is not increased even in the case it is zero.
E x e r c i s e 223. Discuss the tradeoffs of the various simulations in t e r m s of

the alternativity bounds versus transitivity bounds.
E x e r c i s e 224. Show that tabularity transfers under fusion exactly when all
but one factor only have frames of size 1.
E x e r c i s e 225. Recall that a logic is a - c o m p a c t if for every consistent set

based on < a variables there is a model based on a frame for t h a t logic. Show
t h a t for finite a, A is a - c o m p a c t only if A ® K is 1-compact.
E x e r c i s e 226. Show that a logic is 1-compact iff it has at least one K r i p k e -

frame.
E x e r c i s e 227. Call a property T o p e r a t o r c o m p a c t if it holds of an infinite

fusion of logics iff it holds of all finite sub-fusions. T h a t is to say, @i~l Ai
has ~P iff ~)iEJ Ai has [P for all finite Y C_ I. Find out which properties are
operator compact.
CHAPTER 7
Lattices of Modal Logics
7.1. T h e R e l e v a n c e of S t u d y i n g L a t t i c e s of Logics
We have seen in an earlier chapter that the lattices of (normal) modal log-
ics are m at least from a lattice theoretic point of view m quite well-behaved.
However, a typical fact of lattice theory is that the abstract properties of a
lattice give very little insight into the actual structure of a lattice. We al-
ways need a lot of special information about it. We will see that there is no
hope of deducing strong results about logics using these abstract properties;
however, there is a mixture between abstract theory and concrete work with
modal logic that yields surprisingly deep insights into both the structure of
the lattices and the properties of the logics. The main tool will be that of a
splitting, developed within logic by V. A. JANKOV [108~ 109] for intuition-
istic logic and in modal logic especially by WIM BLOK [25] and WOLFGANG
RAUTENBERG [168, 170]. These authors have also paid explicit attention to
the structure of lattices in their work. Elsewhere, lattice theory had only a
minor role to play. The motivation for introducing the concept of a splitting
has been solely the quest for studying the lattice of normal monomodal log-
ics as an object. However, it soon appeared that there are deep connections
between splitting results and intrinsic properties of logics. Here we will de-
velop the theory with hindsight, paying attention to the interplay between the
structure of the lattices of extensions of a logic and properties of the logics in
that lattice.
The objects of study are extensions of the logic K~ (or extensions thereof),
the minimal logic for Kripke-frames with ~ operators. Most results hold only
if ~ is finite, but they will be explicitly m a r k e d - as we have done before
on similar occasions. ~ K~ is the image of (p(T~), N, U> under the map
a : X ~-+ K~ @ X. This map commutes with infinite joins but generally not
with meets. (For example, let X1 := {p} and X2 := {q} with p # q. Then
X1 N X2 = O but obviously K @ X1 = K @X2 = K 9 J-.) The structure of the
lattice is fully determined by a. However, since p(~P~) is uncountable, we must
restrict ourselves to finite sets, or occasionally also recursive or recursively
enumerable sets. These restrictions occasionally cause complications; the set
of finitely axiomatizable logics is not closed under intersection, except for
331
332 7. L a t t i c e s of M o d a l Logics
weakly transitive logics, and generally not closed under infinitary unions. The
set of (strongly) recursively axiomatizable logics are closed under unions and
meets, but closed under infinitary unions only under special circumstances.
Mostly, we will consider finitely axiomatizable logics. Even here the situation
is not so favourable. We will show in Chapter 9 that for given (finite!) sets
X, Y it is undecidable whether K~ | X - K~ @ Y. So, not too much should
be expected from this approach.
In general we will be concerned with decidable subsets of s K~. A subset
E C_ s K~ is called d e c i d a b l e if for given finite X, the problem ' a ( X ) E E '
is decidable, a is a closure operator on T~. A closed element is of the form
a(Z). Take two elements, a ( Z ) and a(Y). Suppose that a ( Z ) is decidable
and Y finite. Then it is decidable whether or not a(Y) C_ a(X). Namely,
we only have to test all elements of Y whether they are in a(X). And since
'~ E a ( Z ) ' is decidable, so is therefore 'a(Y) C_ a(X)'.
PROPOSITION 7.1.1. (n < ~1.) K~ 9 X is decidable iff (A" A C_ K~ 9 X}
is decidable iff for all finite Y it is decidable whether or not K~ @Y C_ K~ G X.
We remark here that to say that a(X) decidable is not the same as to say
that { a ( X ) } is decidable. Now take the lattice s $4.3. We will see later that
all extensions are finitely axiomatizable and decidable. Hence, we can decide
for two logics $4.3 | X and $4.3 G Y, where X and Y are finite, whether
$4.3 9 X C $4.3 | Y and hence whether $4.3 | X - $4.3 @ Y. The same
holds for the lattices of extensions of K . a l t l and Kh. In general, however,
not only are there non-finitely axiomatizable logics, there are also finitely
axiomatizable, undecidable logics. In general, therefore, intrinsic properties
of logical systems and properties of logics in the lattice ~ K~ are unrelated,
with some important exceptions. One such exception are splitting pairs. We
say that (K~ 9 X, K~ | Z} is a splitting pair in the lattice ~ K~ if the lattice
is the disjoint union of the principal ideal generated by K~ | X, and the
principal filter generated by K~ | Z. Suppose that (K~ | X, K~ | Z} is a
splitting pair and both K~ @ X and K~ | Z are decidable. Then for each
finite set Y, 'K~ @ Y - K~ G X ' is decidable. For let Y be given; and
let Y be finite. Then, by the decidability of 'K~ @ Z', for every p E Y,
'K~ | p C_ K~ | Z' is decidable. Hence 'K~ | Y C_ K~ @ Z' is decidable.
Moreover, ' K ~ G Y ~ K ~ | is equivalent to ' K ~ | _C K ~ G X ' . The latter is
decidable. So, splitting pairs are ideal tools for getting a grip on the structure
of the lattice. Yet, it turns out that such pairs are quite rare. In general,
to have sufficiently many such pairs the base logic must be strengthened to
include at least an axiom of weak transitivity. There are many interesting
logics that are weakly transitive, especially in the case of a single operator.
But naturally arising logics in several operators will most likely not be weakly
transitive. (However, one can always add a universal modality to make a
logic weakly transitive.) This is on the one hand a rather negative fact, but
7.2. Splittings and other Lattice Concepts 333
on the other h a n d even the absence of splittings reveals s o m e t h i n g a b o u t the

s t r u c t u r e of the lattice.
7.2. S p l i t t i n g s a n d o t h e r L a t t i c e C o n c e p t s
Let (L, -[-, _k, N, U) be a b o u n d e d lattice. An element x is called R -

i r r e d u c i b l e if x ~= -[- and for every pair y, z of elements such t h a t y • z - x
either y - x or z - x. Dually, we say t h a t x is U - i r r e d u c i b l e if x ~= _1_ and
for every pair of elements y and z such t h a t x - y U z we have either x - y or
x - z. An element x is called R - p r i m e or m e e t p r i m e ( U - p r i m e or j o i n
p r i m e ) if x ~= T (x ~ _1_) and for every pair y and z of elements such t h a t
yNz <z (yUz>x) we have e i t h e r y < z o r z _ x (eithery> xorz>x).
The following is an easy fact a b o u t lattices.
PROPOSITION 7.2.1. Let ~ = (L,-[-, _J_,N, U) be a bounded lattice and x C
L. If x is N - p r i m e , x is N-irreducible. If x U - p r i m e it is also U-irreducible.
Suppose that 2. is distributive. Then x is N-irreducible iff x is R - p r i m e and
x is U-irreducible iff x is U-prime.
In finite lattices m e e t - i r r e d u c i b i l i t y is equivalent to having exactly one
u p p e r cover, and join-irreducibility is equivalent to having exactly one lower
cover. Now, in a finite lattice, every element is the m e e t of m e e t - i r r e d u c i b l e
elements, and every element is the join of j o i n - i r r e d u c i b l e elements. This
follows from the fact t h a t for every pair of elements x < y there exists a
m a x i m a l element p such t h a t p :> x but p ~ y. p m u s t be m e e t - i r r e d u c i b l e ,
for every element strictly larger t h a n p is above y, and if p - u R v for some
u, v > p t h e n u, v :> y and hence p - u N v >_ y, a contradiction. The reader
m a y now check t h a t any finite distributive lattice is isomorphic to the lattice
of u p w a r d closed sets of m e e t - i r r e d u c i b l e s , and dually also isomorphic
to the lattice of downward closed sets of join-irreducibles. This is the general
idea behind the topological r e p r e s e n t a t i o n t h a t we will develop later.
In a complete lattice we can define infinitary versions of these notions
as well. Call x ~ - i r r e d u c i b l e or s t r i c t l y m e e t - i r r e d u c i b l e if whenever
x - ~ ( y i " i C I) we have x - Yi for some i C I. Call an element x [ 7 -
p r i m e or s t r i c t l y m e e t - p r i m e if x > I-l(yi " i C I) implies x > yi for some
i C I. And analogously U - i r r e d u c i b i l i t y or s t r i c t j o i n - i r r e d u c i b i l i t y and
U-primeness or s t r i c t j o i n - p r i m e n e s s are defined. Again, primeness is
stronger t h a n irreducibility. A set Y - (yi " i C I) is called a s u b r e d u c t i o n
of x if ["]Y _< x while Yi ~ x for all i E I.
PROPOSITION 7.2.2. Let 2. be a complete and distributive lattice. I f 2. is

upper continuous then every U-irreducible element is U - p r i m e ; if ~ is lower
continuous then every ~ - i r r e d u c i b l e element is also ~ - p r i m e .
PROOF. Surely, the two s t a t e m e n t s are dual to each other, so let us prove
only the first. Let x be U - i r r e d u c i b l e and let x < U ( y i " i c I). T h e n
334 7. L a t t i c e s of Modal Logics
FIGURE 7.1
9 o 9 o ~
X " 9
x-- xnU(yi'i e I}. S i n c e x m U ( y i . i e I)- U(xnyii e I} we have

x -- U (x R yi 9 i c I}, so by the irreducibility of x, x - x R yi for some i C I.
T h u s for some i C I, x _ yi, as required. KI
T h e converse of these s t a t e m e n t s fails to hold. J u s t consider t h e lattice

1 + Z • Z + 1, o b t a i n e d by adjoining to the p r o d u c t of t h e linear lattice
(Z, min, max I with itself a new b o t t o m and a new top element. T h i s lat-
tice is not continuous, b u t distributive. No element is even simply m e e t -
or join-irreducible. So all strictly j o i n - i r r e d u c i b l e elements are also strictly
j o i n - p r i m e . Lattices of m o d a l logics are u p p e r continuous, so strictly j o i n -
irreducible logics are also strictly j o i n - p r i m e . To give the reader an i m p r e s s i o n
of lattices which are distributive b u t fail one of the large d i s t r i b u t i v e laws,
consider the lattice in Figure 7.1. T h e b o t t o m element is the intersection b o t h
of the lower descending chain and the u p p e r descending chain. Hence x is I-I-
irreducible b u t not [-]-prime. O t h e r examples are the lattices of o p e n subsets
of the topological spaces R n, and the lattice ~ K . a l t l (see C h a p t e r 7.6). T h e
following example illustrates t h a t ~ K . a l t l is not lower continuous. T h e the-
ory of the o n e - p o i n t reflexive frame, ~5-], contains the t h e o r y of the infinite
frame (w,-</ where i -< j iff j = i + 1. For the m a p n ~ 0 is a c o n t r a c t i o n
of (w,-<) onto IT]. T h e t h e o r y of (w,-</ is the intersection of t h e theories of
(n,-</, by the fact t h a t the n - t r a n s i t s of the respective roots are isomorphic.
So {Th (n,-<) : n C w} is a s u b r e d u c t i o n of the t h e o r y of [~-]. Strongly irre-
ducible elements have a nice c h a r a c t e r i z a t i o n from a lattice t h e o r e t i c p o i n t of
view. Given an element x C L we say t h a t y is an u p p e r c o v e r of x if x < y
b u t for no element z we have x < z < y. x is called a l o w e r c o v e r of y if y
is an u p p e r cover of x. We write x -< y to say t h a t x is a lower cover of y (or
t h a t y is an u p p e r cover of x). T h e following is not h a r d to see.
PROPOSITION 7.2.3 (Blok). In a complete lattice, x is strongly j o i n - i r r e -

ducible iff it has an upper cover y and .for all z > x we have z >_ y. Dually,
x is strongly meet-irreducible iff it has a lower cover y and f o r all z < x we
have z <_ y.
For finitely axiomatizable logics we can show they are never the limit of
an ascending sequence of logics (see [23]).
THEOREM 7.2.4. Let A be finitely axiomatizable. Then f o r every logic

0 ~ A there exists a lower cover (9 ~ of A such that 0 C_ 0 ~
PROOF. Consider the following property ~P. X has IP iff O @ X ~ A.

This is a property of finite character. For A - O @ E for some finite set E.
Thus if O @ X _D O @ E then there exists a finite subset X0 C_ X such that
E C_ O @ X0. Therefore [P has finite character. So O is contained in a maximal
set X ~ by Tukey's Lemma. P u t O ~ " - O @ X ~ By definition of X ~ 0 ~ is
a lower cover of A. For consider an extension O~ Y. If O ~ Y ~ A then
X~ Y must have [P, which by the maximality of X ~ means t h a t Y C_ X ~
Thus O ~ @ Y - O ~ as required. V7
Similarly the following theorem is proved. For the purpose of stating the
theorem, a pair x / y is called a q u o t i e n t if x _> y. The quotient x / y is p r i m e
if x > y and for no z, x > z > x.
THEOREM 7.2.5. Let Ao < A ~ Then the interval [Ao,A ~ contains a

p r i m e quotient.
PROOF. It is not hard to see that the interval contains a logic O # Oo

which is finitely axiomatizable over Ao. Now show as above t h a t there exists
a lower cover | for (9. [:3
The lattice of rational (or real) numbers in the closed interval [0, 1], with rain
and m a x as operations, contains no prime quotients. So this is a nontrivial
property of lattices of logics. Logics which are not finitely axiomatizable may
also have lower covers. Examples are the logics t ( M ) of Chapter 2.6, for
infinite M. The covers are not necessarily unique. Moreover, even if A has a
lower cover Ao it may happen that there exist logics O such that O C A but
OgAo.
The notion of a s u b l a t t i c e and homomorphic image of a lattice is defined
just as all the other algebraic concepts. (However, notice t h a t our lattices usu-
ally have infinitary operations. So the concepts must be extended to infinitary
operations; we trust that the reader understands how this is done.) If we just
consider lattices as objects, we can afford to be vague as to whether or not
we consider infinitary operations or the top and b o t t o m elements as being
primitve operations, because we can define them from the others if necessary.
When we consider homomorphisms of lattices, however, this makes a great
difference. A homomorphism which is faithful to the finitary operations need

not be faithful to the infinitary ones. (This is a point where the notion of
a category defined in Section 4.4 is helpful. Taking the class of complete
lattices with maps respecting the finitary operations results in a different cat-
egory than taking the class of complete lattices together with maps preserving
the infinitary operations.) Let us say, therefore, that a lattice only has the
finitary operations. Recall that locales are structures (L, ~, [_J) where N and
[_J are the finitary meet and infinitary join. (See Section 3.) Notions of sub-
object and homomorphism differ with respect to the top and bottom element,
depending on whether they are present in the signature. If they are they
must be preserved under embeddings. In boolean algebras, where the top and
bottom are definable, a subalgebra must contain the top and bottom of the
original algebra. For lattices this need not be so. For our purposes, we are
quite happy not to have the top and the bottom in the signature.
Recall from Section 1.1 the notation S X and $X. An upper cone is a
filter if it is closed under finite intersections; a lower cone is called an i d e a l
if it is closed under finite unions. A filter (an ideal) is p r i n c i p a l if it is of
the form ~{x} (${x}) for some x. We usually write $x instead of ${x} and
l"x instead of ${y}. If F is a principal filter and F = Sx = S y then x = y.
Similarly for ideals. Filters and ideals are special sorts of sublattices. If they
are principal, then they are also complete sublattices, or, for that matter,
sublocales.
Now we come to the main definition of this section, that of a splitting,
shown in Figure 7.2.
DEFINITION 7.2.6. A pair (p, q) of elements is said to split a lattice 2. if

for every element x C L we have x < p or x > q, but not both. Equivalently,
(p, q) is a splitting if L is the disjoint union of the ideal $p and the filter ~ q.
If that is so, p is said to split s and q is said to c o - s p l i t 2.. p and q are
s p l i t t i n g c o m p a n i o n s of each other. We write s or • for the splitting
companion of p.
The situation is depicted in Figure 7.2. We give some examples from

modal logic.
EXAMPLE. K 1 / T h [ - ~ ] - K1.D. For let A ~ Th[-~. Then [-1• ~ A, and so
0q- E A.
EXAMPLE. S4/Th[ o .~o [ = S5. Io .~o ] consists of two points, 0 and
1, and <l = {(0,0), (0, 1), (1,1)}. Assume that A is an extension of S4 such
that A ~ S5. We have to show that [~ is a frame for A. So, by assumption,
(}(p A (}ff]~p) V (}(-~p A (}if]p) is consistent with A and therefore there exists a
model (~,/3, x) ~ p; (}ff]~p. Put
v .- p o p ) A D(- p -+ DO- p))

FIGURE 7.2
p
,(
,(
<
,(
<
<
,(
,(
V is an internal set of ~ and nonempty. It is a generated subset. Furthermore,

every point not in V has a sucessor in V. To see this, define the s p a n of
y C f to be the set of formulae p C {p,-~p} such t h a t there exists a z E> y
with (~,/3, z) ~ ~a. A point y is of m i n i m a l s p a n if every successor of y has
the same span as y. It is easy to see that every point sees a point of minimal
span. V is the set of points of minimal span. Let W be the set of successors
of x which are not in V. W is not empty, for it contains x. (By the fact that
x has span {p,-,p}, and has a successor of span {--p}, x r V.) Define 7r by
7r(y) := 0 if y C W and ~r(y) := 1 if y C V. 7r is a contraction onto b of the
subframe generated by x in ~.
EXAMPLE. (See [241] for details.) Let O be the tense logic of the rational
numbers (Q, <) and A the tense logic of the real numbers (R, <). Then
......
A = O//Th[c -~o]. For let (f,<~) be a rooted frame for | T h e n <~ is

a linear dense order without endpoints, though not necessarily irreflexive.
Suppose t h a t there is a tense p - m o r p h i s m onto I c ~o I. Then there are sets
A, B C f such t h a t for every x C A and every y C B: x <~ y. Moreover, every
element in A has a successor in A, and every element in B has a predecessor
in B. Such pair of sets is called a gap. Hence, the frames for A are those
frames for O which contain no gaps. In particular, (Q, <) is not a frame for
A.
The following theorem gives a full characterisation of splitting elements
and splitting pairs.
THEOREM 7.2.7. An element p splits a complete lattice iff p is ["]-prime.

Dually, an element q co-splits a complete lattice iff q is [.J-prime. If p splits
s the companion is uniquely determined, similarly for co-splitting elements.
The map p ~ p, which sends a strictly meet-prime element to its splitting
companion is an isomorphism from the poser of ~ - p r i m e elements onto the
338 7. L a t t i c e s of M o d a l L o g i c s
poset of [.J-elements. Its inverse is the map q ~-~ q* which sends a co-splitting
element to its splitting companion.
PROOF. First, if p splits/~, then the complement of Sp is a principal filter

j'q for some q. q is certainly unique. Moreover, consider p _> [7 (yi " i C I).
Then, [-](y~ 9 i C I) ~ l"q. This can only be the case if for some i C I,
yi r S q, which means for that i, yi __ p. Thus p is [']-prime. Dually, we can
show that if q co-splits the lattice, then q is [.J-prime. Conversely, assume
that p is [-]-prime. Take q "- [ 7 ( Y ' Y ~ P). We have q ~ p. Moreover, if
x ~ p then by definition of q, x _> q. And if x _> q, then x _< p cannot hold.
So, we have a splitting pair (p,q). Dually for [ [ - p r i m e elements. Now for
the last claim. Take as a map the map that sends a splitting element to its
companion. This map is injective; for if Pl ~: p2 then $pl r 1"p2. T h e n also
$ (Pl), ~: $ (P2),, showing (pl), r (p2),. The map is surjective, because each
[ [ - p r i m e element has a companion. Now consider pl _< p2, and let (pl, ql)
and (p2,q2) be splitting pairs. Then $pl C_ $p2, whence 1"ql _D 1"q2, from
which ql _< q2. [3
If we have a splitting pair, then there arise two natural sublattices, namely
Sp and S q. Generally, in modal logic, we are interested in lattices of exten-
sions, so the lattices Sq are of special importance. It is generally the case that
if (~, q~ is a splitting pair of ~ and ~ :> q then ~ splits the lattice f../p. (In
case ~ ~ q the induced splitting is trivial.) The splitting companion is q LJ ~.
So we have the following equation.
The iterated splitting does therefore not depend on the order in which we
split. Thus for a possibly infinite set N of prime elements of ~ we define
~ / N "- U i ~ / p " p c N )
In Figure 7.3 the situation is shown where p and r are being split in succession,
with q the splitting companion of p and s the splitting companion of r. In
case we have a complete boolean algebra and a splitting pair (p, q), p is a
coatom and q an atom. (Only coatoms are N-irreducible, and only atoms are
U-irreducible.)
THEOREM 7.2.8. A complete lattice f, is isomorphic to 2 • ~ for some
complete lattice ~ iff there exists a splitting pair (p, q) such that p is a coatom
and q is an atom.
PROOF. Assume s ~- 2 • ~. We may actually assume that s - 2 • ~. Let
b be the b o t t o m element of ~, and t the top element of ~. (These elements
exist in ~ since it is a complete lattice.) Put t* . - (0, t) and b, := (1, b). The
pair (t*, b,) is a splitting pair of s For assume that (i, y) f t* - (0, t). Then
i f 0, since y _< t by choice of t. Hence i - 1. But then (i,y) >_ (1, b) - b,.
FIGURE 7.3. Iterated Splitting
r > p
>
>
>
>
>
\
And conversely. Now assume t h a t (p, q) is a splitting pair and t h a t p is a

coatom, while q is an atom in s We show t h a t the m a p x ~ x U q is a
bijection between the ideal Sp and the filter t q with inverse y ~ y N p. For
notice t h a t under the assumptions, p N q - _k and p U q - T. (For q f p and
so p N q < q. Therefore p N q - _1_, for q is an atom. Likewise, p U q > p and
so p U q - T, since p is a coatom.) Now for x ___p and y >_ q we have
,q)np- np)u (qnp) - xnp-
(y ~ p) U q - ( y U q ) [ ~ ( p U q ) - y U q -- y
This shows t h a t both maps are bijective. Moreover, they are isotonic and
hence isomorphisms. Ul
It follows, for example, t h a t a finite boolean algebra is isomorphic to a direct

product of some copies of 2. However, in general, in a splitting pair (p,q)
neither is p a coatom, nor is q an atom. An example is the six element chain
6 - (6, min, max}. In this lattice (2, 3) is a splitting pair, but 2 is not an
a t o m and 3 not a coatom. The reader may verify t h a t also the splitting
(Th [--~, K . D ) is an example.
We end this section with a useful application to the theory of consequence
relations proved in [172]. Its proof is not difficult considering the fact t h a t
O ~ ~o is an isomorphic embedding.
THEOREM 7.2.9 (Rautenberg). Let (O,A} be a splitting of the lattice of

extensions o r E . Then the pair (~-~,FA) is a splitting of ~ ( k ~ ) .
In t h a t case one can show that for every axiomatization R of kA over ~-~
there is a p E R such t h a t FA -- F +p.
COROLLARY 7.2.10. Let (0, A) be a splitting of s E. Let p be a rule. Then

p axiomatizes FA over ~ iff p is derivable in kA but not admissible for O.
For example, to know whether adding a rule p to ~-s4 axiomatizes ~s5

we have to check whether p is a derived rule of F-s5 and not admissible in the
logic of]o ~o I.
E x e r c i s e 229. Let s be a lattice, not necessarily complete. Show t h a t an

element r _l_,T is join-irreducible (meet-irreducible) iff it has exactly one
lower cover (exactly one upper cover).
E x e r c i s e 230. Let s be a complete lattice. Show that if p is [']-irreducible

in s ([']-prime) then it is also ~ - i r r e d u c i b l e ([-]-prime) in any complete
sublattice containing it.
E x e r c i s e 231. Show with a specific example that there are lattices s and
elements p such that p is not ["]-prime in/~ but ~ - p r i m e in the lattice s
for some set of ["]-prime elements N. Thus the notion of being a splitting
element is not stable under iterated splittings.
E x e r c i s e 232. A logic A is called e s s e n t i a l l y 1 - a x i o m a t i z a b l e over O

if for every axiomatization A = O(X) over 0 there exists a T C X such
that A = 0 ( ~ ) . Show that a modal logic A co-splits ~ 0 iff it is essentially
1-axiomatizable over |
7.3. Irreducible and P r i m e Logics

In this section we will investigate the possibility of characterizing irre-
ducibility and primeness of logics algebraically.
THEOREM 7.3.1. A E ~ O is [-]-irreducible only if A = Th 92 for a subdi-

rectly irreducible 92.
PROOF. Surely we have A - Th 92 for some P./. Suppose t h a t P2 is not
subdirectly irreducible. Then, by Theorem 4.1.3, 92 is a subdirect product
of (9./i : i C I), where every 92~ is subdirectly irreducible. Then we have
Th 92 = [7 (Th 9./~ : i C I). Hence there is a 92~ such that Th 92 = Th 92i = A. K]
The converse of Theorem 7.3.1 is generally false. For example, take the
S 4 . 3 - f r a m e 0 := (w + 1, >_). This frame is generated by v and therefore
the algebra 9~ta(o) of subsets of that frame is subdirectly irreducible. How-
ever, ThffJta(0) = G r z . 3 has the finite model property and is therefore not
irreducible in ~ K. A useful result is this.
PROPOSITION 7.3.2. Let A be [']-irreducible. Then A = Th 92 for an 92

which is finitely generated.
7.3. Irreducible and Prime Logics 341
PROOF. Let A - Th 92. We have 92 )# ~ iff there is a valuation/3 and a

U such that (92,/3, U) ~ -~p. Take ~ to be the subalgebra generated by/3(p),
p E var(w). Let Y " - U M C - {a c U " a C C}. We claim t h a t V is an
ultrafilter on C. For V is obviously upward closed in C. Moreover, it is closed
under intersection, since C is closed under intersection and U is as well. Also,
i f x C C, then also - x E C. Now either x C U o r - x C U. Hence either
x C V or - x E V. So, V is an ultrafilter in C. Since ~ is a subalgebra of 92,
we have Th ~ _~ Th 92. Also, (~, fl, V) V ~, since f l ( - ~ ) E C and f l ( - ~ ) C U.
We conclude that if there is a model based on 92, there is a model based on a
finitely generated subalgebra of 92. So
Th 92 - ('~(Th ~" ~ --~ 92, ~ finitely generated)
Consequently, if Th 92 is r'l-irreducible, there must be a finitely generated

--~ 92 such that Th ~ - Th 92. 71
THEOREM 7.3.3. A is irreducible in ~ 0 if and only if A - Th 92 for a

subdirectly irreducible algebra such that Th 92 is prime in ~Th 92.
PROOF. Observe that in a lattice an element x is ~ - i r r e d u c i b l e iff it is
r-I-prime in ~ x. 71
Thus, in order to find algebras whose logics are irreducible we need to answer
the question of which algebras are splitting algebras. This is by no means
trivial or more easy, so the last theorem is really of little practical significance.
By analogy, an algebra 92 or a frame ~ is called p r i m e in ~O or AIg (9 if Th 92
( T h e ) is prime in ~O. For the statement of our next theorems we shall
introduce the notion of a presentation of an algebra. We shall first give a
general definition and then specialize to modal algebras.
DEFINITION 7.3.4. Let ~? be a variety of ~-algebras and 92 E ~?. A pair

( X , E } is called a p r e s e n t a t i o n o f 92 if (i.) X is a set, E C T m n ( X ) x
T m a ( X ) , and (ii.) ~ r v ( X ) / O ( E ) ~- 92, where ~)(E) is the least congruence
in ~ r v ( X ) containing E. A presentation is f i n i t e if X and E are both finite.
92 is called f i n i t e l y p r e s e n t a b l e in V if 92 has a finite presentation in V.
Since every algebra is the homomorphic image of a free algebra in a variety,
every algebra of a variety is presentable in it. Simply let h 9 ~ m a ( X ) --- 92
be a surjective homomorphism. Then put E "- ker(h). The pair (X, E) is a
presentation of 92.
Now fix an isomorphism i. Then i o he(E) " X -+ A. So, we may actually
assume that X consists of elements of A. Then 92 is generated by X, and E
is a set of equations such that the minimal congruence containing E is the
kernel of the canonical map ~ r v ( X ) ~ 92. In the present context, we speak of
logics rather than varieties, and of open filters rather than congruences. We
write {A} for the open filter generated by A.
342 7. L a t t i c e s o f M o d a l L o g i c s
DEFINITION 7.3.5. Let 92 be a ~ - m o d a l algebra, 0 a ~ - m o d a l logic. A

p r e s e n t a t i o n of 92 over 0 is a pair ( X , A ) where X C_ A, is a set of variables
and A a set of terms over X such that the map pr" ~ r e ( Z ) / / ( A ) -+ 92 induced
by the identity assignment p r ( a ) " - a is an isomorphism. 92 is called f i n i t e l y
p r e s e n t a b l e if both X and A can be chosen finite. We call A a d i a g r a m of 92
over O. A is not uniquely determined. We also say that 92 is a - p r e s e n t a b l e
if~x - a.
If A can be chosen finite, we write 5 to denote the set A as well as A A.

Moreover, we may identify ~ a n e ( X ) and ~ a n e ( a ) . In order to appreciate this
definition, consider the dual frames 92+ and ~ a n e ( a ) . Of course, since 92 is
a - g e n e r a t e d there is a homomorphism h 9 ~ r e ( a ) ~ 92. Thus, h + 9 92+ --~
~ a n e ( a ) . So, 92+ is a generated subframe of the a - c a n o n i c a l frame for 0 . It
is clear t h a t there exists a possibly infinite A such t h a t ~ r e ( a ) / ( A ) ~- 92. We
also call this a d i a g r a m . Thus, viewing the canonical frame as equivalence
classes of formulae, there is a set of formulae characterizing the notion being in
the generated subframe underlying 92+ generated by a. 92 is finitely presentable
if there is a finite set of this kind. Now, recall t h a t even if O is the modal
theory of 92, there are many ways in which 92+ lies embedded in the canonical
O-frame, just because there are many ways to generate 92 by a elements. The
diagram of 9.1 completely determines the m a p h 9 ~ r o (a) --~ 92 ~ at least up to
isomorphisms of 92, which is the best we can hope for anyway. Now fix h, let
A be the associated diagram, and ~ is the h + - i m a g e of the frame underlying
92+. Thus in the canonical frame, a world x satisfies this diagram, x E N ~ A,
iff x is in ~.
If 92 is finite and the number ~ of primitive modalities is finite then for
every k C w, 92 is k-presentable iff it is k-generated. In particular, 92 is
k-presentable for k " - ~A"
PROPOSITION 7.3.6. (~ < b~0.) Let 92 be finite. Define
~(~) " - A(Pa A Pb t-+ Panb " a, b C A)

A A ( ~ P a ++ p_~ " a E A)
AA(~]iPa ~ Piia "a E A, i < ~)
T h e n 5(92) is a diagram for 92 over 0 .
PROOF. Consider e 9 ~ r o ( v a r ( 5 ) ) -+ 92 " Pa ~+ a. e is surjective and
factors through pr, that is, e - p r o ~ for a homomorphism ~. Since p r o ~ -
is surjective, pr is surjective. It remains to show t h a t pr is injective. This is
done by showing that ~(~) - ~(r iff ~ +-~ r C (5) and for every ~ there is
a a E A such that p ~ pa C (5). This is proved by induction on p. Hence
~im(~) - ~A and since A is finite, pr is injective.
This diagram uses as many variables as there are elements in the algebra.
However, it is possible to use far less variables. First, a finite algebra is
isomorphic to the set algebra over a finite Kripke-frame ~. Given ~ we may

replace the diagram by
a(f) .-
This is clearly an equivalent characterization of 92. Notice that by passing

from variables for sets of worlds to variables for worlds we have used k "- ~f
variables where previously we had 2 k. We can compress this diagram further.
Let t~ "- r2log k 7, the least integer such that 2e >__ k. Then there exists an
injection s" f --+ ~o(e). Now take variables qi, i < e. For S C_ { 0 , . . . , g - 1}
define
x(s) . - A q' A A
ics i~s
Then the diagram for [ can be replaced by the following formula
a(f).- V<x(s(v))-v c.f>
AA<x(s(v)) ~ O j X ( s ( w ) ) " v <:lj w , j < ~>
9 <
Thus the number of variables needed is at most logarithmic in the size of the
underlying frame. It is clear that a sharp bound is the number of generators of
92. We have shown, however, that this number is at most doubly logarithmic
in ~A.
DEFINITION 7.3.7. Let 92 be subdirectly irreducible. Let A be a diagram of
92 over 0 , ~ C AIg O. For a compound modality Rq and a finite subset 5 C A
we say that ~ is R q S - c o n s i s t e n t w i t h 92, if a valuation 13" vat[A] --+ B and
an ultrafilter U exists such that (!~,/~, U) ~ ~Pc; FRS, where Pc E var[A], c
being an opremum of 92. f8 is called w - c o n s i s t e n t w i t h 92 if a valuation
~" var[A] -+ B exists satisfying (!~,/3, U) ~ -~Pc; {mS" m compound, 5 C A}.
If ~ is EBS-consistent with 91 for every compound modality FR and finite subset
5 C A then ~8 is said to be w e a k l y c o n s i s t e n t w i t h 92.
THEOREM 7.3.8. Let 92 be subdirectly irreducible. Then the following as-

sertions are equivalent/or all ~ 6 AIM O"
(i) !~ is w-consistent with 92.
(ii) 92 C SH(~).
(iii) 9/C HS(~3).
PROOF. Let A be a diagram of 92 over O. Assume first (i). Let then

13" var[A]--+ B and U be such that /3(~pc) C U and for all compound
modalities El and formulae 5 C A we have/3(m5) E U. Let F be the open filter
344 7. Lattices of M o d a l Logics
FIGURE 7.4
var[ ]
92
;~ro(va
generated by t h e / 3 ( m 5 ) . Consider the induced m a p p i n g r ~ -+ ~ / F -"

and put 7 " - r 1 7 6 9 vat[A] -+ C. T h e n 7(m5) - 1 for all m and 5 E
A. T h e h o m o m o r p h i s m ~ factors t h r o u g h 77 9 ~re(var[A]) --+ 91, for 92 -~
~re(var[A])/(A}. (See Figure 7.4.) 92 is subdirectly irreducible and has a
m i n i m a l nontrivial congruence relation, which is generated by c. To show
t h a t the induced m a p p i n g r is injective it is therefore sufficient to show t h a t
r # 1 or, equivalently - - t h a t r # 0. But ~ ( - c ) - ~ o r/(-~pc) -
7(-~pc) - c o ~(-~Pc) > 0, because for every a C F there exists a K] and a finite
subset 5 C_ A such t h a t a > ~(K]5). T h e n ~(-~p~)n a > /3(-~pc)n ~(K]5) -
~(-~p~ AK]5) # 0. Hence 4 is injective and 92 E S(~) C_ SH(K~). This shows (ii).
T h a t (ii) implies (iii) we have seen in C h a p t e r 1.3. Finally, assume (iii) holds.
We will show (i). Let 92 C HS(K~). T h e n there is a E, an injective e" ~ ~-,
and a surjection p" ~ --~ 92. (See Figure 7.5.) Here, ~ " - ~ro(var[A]) and ~ is
the canonical mapping. ~ is free and thus ~ can be lifted over p to 7 " ~ --+ ~.
Then/3"- c o T . For every n E w, ~(~p~AK]5) - ~(-~p~) > 0 and since
is injective, ~(~p~ A ~5) -- e o 7(-~pc A K]5) > 0 for every K]5. Hence ~3 is
w-consistent with 92; (i) is shown. [i]
LEMMA 7.3.9. Let 92 be subdirectly irreducible. If ~ is weakly consistent

with 92 then there is a G G U p ( ~ ) which is w-consistent with 92.
PROOF. Let S be the set of formulae m5 such t h a t 5 is a finite subset of

A and K] is a c o m p o u n d modality. (We have K] - K]s for some finite set s of
sequences over ~.) For each a E S we have a valuation ~ "var[A] --+ B and
an ultrafilter U~ such that/3~(-~pc Act) E U~. Let s be a finite set of sequences
FIGURE 7.5
91
and 5 a finite subset of A.

E(~, ~ ) : : {D~a' 9 ~ c_ t, a c_ a'}
Next put
E "- {E(s,5)'s C ~*,5 C_ A,s, 5 finite}
E has the finite intersection property, for we have
E(s~, 51) e E(s2, 52) - E(s~ U s~,5~ u 5~.).
Therefore, E is contained in an ultrafilter, say V. Now define r "- 1-Iv ~ ,
7 := rig/3~. Then ~(-~pc A FFth) > 0 for every EB5 E S. The set
{T(-~pc A l-IsS) 9 s C a*, 5 C_ A}
has the finite intersection property and is contained in an ultrafilter. Let it
be U. Then (~, 7, U) ~ ~pc; S. Hence ~ is w-consistent with 91. E3
P u t t i n g our results together we get
THEOREM 7.3.10. Let 91 be subdirectly irreducible. Then the following

assertions are equivalent for every ~3"
(i) ~3 is weakly consistent with 92.
(ii) 92 E SHUp(~).
(iii) 92 E HSP(~3).
PROOF. (i) =~ (ii). Assume (i). By Lemma 7.3.9 there is a r E Up(~3)

which is w-consistent with 91. Hence by Theorem 7.3.8 92 E SH(~) G HSUp(~3).
(ii) ~ (i). If ~3 is not weakly consistent with 91 then there is a m5 such
that ~ ~ m5 -+ Pc. It follows that G ~ []5 -+ Pc for any G E Up(~).
Hence | is not w-consistent with 92 and thus 9.1 r SH(~). Since this is
valid for all | E Up(~3), 91 ~ SHUp(~). Now (ii)implies (iii) because
HSUp(!B) C_ HSP(!B), and (iii)implies (i). F1
Notice that we did not actually use Jhnsson's Lemma, but derived it. T h a t
this is possible was first observed in [234]. Notice that in the course of the def-
initions we have made crucial use of the fact that 92 is subdirectly irreducible.
The previous theorem shows quite clearly why this must be so. For if not,
all varieties are of the form HSUp(K) for a class :K. The following Splitting
Theorem can now be proved. The original version appeared in RAUTENBERG
[170] but only for finite algebras in weakly transitive logics. In KRACHT [120]
it was generalized to the case where 92 is finitely presentable. The following
version is fully general and was proved in WOLTER [234].
THEOREM 7.3.11 (Splitting Theorem). Let 92 be a subdirectly irreducible
modal algebra with diagram A. Then the following are equivalent.
1. Th 92 is prime in ~ 0.
2. There is a compound modality ~] and a finite ~ C A such that for all
E Alg 0-
( ~) If ~ is E]5-consistent with 92 then ~ is weakly consistent with 92.
3. ( ~ " 92 r HSP~} is a variety.
4. { ~ " 92 r HSP~} is closed under ultraproducts.
Moreover, if 92 and E]~ fulfill ( ~) we have ~ 0 /92 - 0 @ m5 -+ pc.
PROOF. Clearly, the first and the last are equivalent, for if we have a
splitting, the corresponding splitting partner axiomatizes a variety, the variety
of all algebras whose theory is not contained in the theory of 92, or those
algebras in whose variety 92 is not contained. Thus, we have to show the
equivalence of the last three. Let S be the set of all []5. Assume that (t)
fails for all a G S. For every a G S there is a ~ which is a-consistent
with 94 but not weakly consistent with 92. Hence by the preceding theorem,
for every a E S, 92 r H S P ( ~ ) that is T h ~ ~ Th92. However, for a
suitable ultraproduct ~ - 1--[~ES~ is weakly consistent with 92, and hence
Th 92 _D Th ~, from which follows that { ~ 9 92 r HSP~} is not closed under
ultraproducts and hence also not a variety. Also, Th 92 is not prime in ~O. Let
now (~) be fulfilled by some E](~. Then {(B" P./r H S P ~ } - {s s # E]~ -+ pc }
and hence it is a variety, and so closed under ultraproducts. Thus Th 92 is
prime, with splitting companion ~ @ ~5 --+ Pc. Eli
There are two important subcases of the Splitting Theorem, which are
each quite characteristic. One concerns the case of weakly transitive logics,
the other that of cycle-free frames. Recall that weakly transitive logics are
characterized by the existence of a strongest compound modality. Further-
more, each finite subdirectly irreducible algebra has a diagram. Hence (t) is
easy to satisfy in this case.
COROLLARY 7.3.12 (Rautenberg). (~ < ~0.) Let 0 be weakly transitive.
Then every finite subdirectly irreducible O-algebra splits ~ 0 .
7.4. Duality Theory for Upper Continuous Lattices 347
Let us say that a finite algebra is c y c l e - f r e e iff 92 ~ []'~A_ for some

m E w. It can be seen that 92 is cycle-free iff the corresponding Kripke-
structure contains no cycles. If 92 ~ [ ] < m - l • then ~<m-l_L is an opremum
of 92.
COROLLARY 7.3.13 (Blok). (~ < tq0.) Let 92 be finite subdirectly irre-
ducible and cycle-free. Then 9.1 is prime in every variety containing it. Then
there exists a finite number m such that ~ 0 / 9 2 - O | --+ ~[~m _L.
PROOF. Let [~m_L be an opremum of 92; let 5 " - 5(92) be a diagram of 92.
Then ~<m+15 satisfies (l) in Theorem 7.3.11. For if (~B,13, U) ~ ~]<m+15 A
[~m_j_ then for all r > 0 ( ~ , 13, U) ~ ~]<_m+r(~ A ~rn_j_ because 1 - - ~ ( [ ~ m _ j _ __~
[~m ~]~ 5) -- ~([~m_k --+ ~m+~5) for all r :> 0. So for all r > 0,-~(~]<m5 A
E x e r c i s e 233. Show that the number of variables needed to axiomatize

~A/92 over A is equal to the number of elements needed to generate 92.
E x e r c i s e 234. Let O - [__]icz~ A/fl3i where ~ i are splitting algebras. Sup-

pose that the Th ~ i are incomparable. Then the minimum number n of
variables such that O is n-axiomatizable is equal to the minimum number k
such that every ~Bi is k-generated.
E x e r c i s e 235. A variety is said to have the c o n g r u e n c e e x t e n s i o n p r o p -

e r t y (CEP) if for every algebra 92 and every subalgebra f13 < !t3 the following
holds. Every congruence 0 on ~B is the restriction to B of a congruence on
92. Hence, in a variety with CEP we have HS 9 2 - SH 92 for all P.I. Show that
Abelian groups have the CEP, but groups in general do not.
E x e r c i s e 236. Show that the variety of x-modal algebras has CEP for any
/~.
* E x e r c i s e 237. Show that the algebra 92 of the frame Ic ~o] splits the
lattice of extensions of K . a l t 3 . B . T of reflexive, symmetric frames with at
most three successors. However, show that in general it does not follow that
if P2 E SHIJp ~ then also 92 C SH ~ . The latter is in fact in many cases true,
and a counterexamples are rather difficult to construct. For example, ~ must
in any case be infinite. (Can you show this?)
7.4. D u a l i t y T h e o r y for U p p e r C o n t i n u o u s L a t t i c e s
We have seen that lattices of modal logics are locales also called upper con-
tinuous lattices and that they are generated by their join-compact elements,
because by Theorem 2.9.8 the latter coincide with the finitely axiomatizable
logics. In this chapter we want to go deeper into the structure theory of such
locales a n d see w h a t further p r o p e r t i e s of the lattices of logics we can derive.

T h e r e is a duality t h e o r y for locales (see [110]). A topological space always
gives rise to an u p p e r continuous lattice, also called a locale. Namely, let
- (X, X} be a topological space over the set X with o p e n sets X. T h e n let
gt(~) " - (X, N, U). By definition of a topological space all finite intersections
of o p e n sets are open, and all unions of open sets are open; so, ~ ( ~ ) is a locale.
Let f 9 9~ -+ ~ be a continuous map. T h e n let ~t(f) be the the r e s t r i c t i o n of
f - 1 . 2 y __+ 2 z . A ~ f - l [ A ] t o X . By continuity o f f this is a m a p f r o m
X to Y. It is not h a r d to see t h a t gt(f) 9 ~ ( ~ ) --+ ~t(~), i. e. t h a t gt(f) is a
h o m o m o r p h i s m of locales. It is easily seen t h a t ~t is a c o n t r a v a r i a n t f u n c t o r
from the category of topological spaces into the category of u p p e r - c o n t i n u o u s
distributive lattices. T h u s ~t can also be c o n s t r u e d as a covariant f u n c t o r from
the c a t e g o r y of topological spaces to the category of locales.
To m a k e the exposition analogous to t h a t of S t o n e - D u a l i t y , let us switch
from the category of locales to the dual category, t h a t of frames. A p o i n t of a
frame is a surjective f r a m e - m o r p h i s m p " ~ --- 2. To have a surjection onto 2
m e a n s t h a t L can be split into two sets, F " - p - l ( 1 ) and I " - p - l ( 0 ) , where F
is a filter a n d I an ideal, a n d I is closed u n d e r a r b i t r a r y joins. For if p ( a s ) - 0
for all s E S, t h e n p ( t J s ~ s a s ) - U s c s p ( a ~ ) - O. T h u s I - S x for some x.
Moreover, x m u s t be M-irreducible (since F is closed u n d e r intersection). So,
x is R - p r i m e . This c h a r a c t e r i z a t i o n is exact. For if x is N-irreducible t h e n S x
is an ideal a n d L - Sx is a filter. Namely, from y ~ x and y < z we d e d u c e
z ~ x, a n d from y , z ~ x also y M z ~ x. Let pt(2.) denote the set of points
of ~. T h e r e is a bijection b e t w e e n the set of points and the R - i r r e d u c i b l e
elements of ~, defined by x ~ p~, where p~ is defined by p~(y) - 0 iff y < x.
O n pt(2.) we take as open sets the sets of the form ~ - { p ' p ( x ) - 1). So,
since every p is of the form py for some y, we have ~ - {py 9 y ~ x } . Now
put | - (pt(~), {~'x E n}). In place of h o m o m o r p h i s m s we m i g h t
also take the m e e t - i r r e d u c i b l e elements as elements of | | is a
topological space, by the next t h e o r e m , if we add t h a t ~ - T a n d pt(2,) - _L.
LEMMA 7.4.1. The m a p ~ -- {p " p ( x ) - 1} c o m m u t e s with arbitrary j o i n s
and finite meets.
PROOF. Let z -- [.J~eixi. p E ~ iff p ( z ) -- 1 iff p ( x i ) -- 1 for at least
one i E I i f f p E xi for at least one i E I. So ~ - UicIXi" F u r t h e r m o r e ,
p E xl ~"-~2 iff p ( x l n x2) - 1 iff p ( x l ) - 1 and p ( x 2 ) - 1 iff p E x~ n ~'22.
Hence X 1 ~"X2 -- X'l n x2. []
T h e sets of the form l"x are closed sets. Now consider a h o m o m o r p h i s m

h : ~ -+ 9Y~. T h e n the m a p p ~-~ p o h m a p s a point p : ~ ~ 2 of if2 onto a
point of ~. T h u s put | : p ~ p o h.
THEOREM 7.4.2. The m a p | m a p p i n g a locale 2. onto | and a
h o m o m o r p h i s m h : ~ --+ if2 onto Gpc(h) : p ~-~ p o h is a covariant f u n c t o r
from the category of locales to the category of topological spaces and continuous
maps.
PROPOSITION 7.4.3. Let koc ~ be the opposite category of the category

koc of locales and homomorphisms; and let Top be the category of topological
spaces. Then ~ considered as a functor Top --+ koc ~ is left adjoined to Gpr
considered as a functor koc ~ --+ Top.
The proof of this fact is an exercise. Interesting for us are the unit and
counit of this adjunction. If we have a locale, the function x ~-~ ~ is a canonical
map s -~ ~(|163 This map is surjective, but in general not injective, see
the exercises below. Likewise, given a topological space :~, we have a map
:~ --+ Gpc(~t(~)). This map is surjective, but in general not injective. For
a point x, the set {x} is join-irreducible and so its complement is m e e t -
irreducible. Hence it gives rise to a point Px of ~(:~). There may exist x and
y such that x ~ y and px = py. This motivates the following definition.
DEFINITION 7.4.4. A locale 2. is called s p a t i a l if the map x ~-~ ~ is a

bijection from s onto gt(|163 A space X is called s o b e r if the map x ~-~ p~
is a bijection from X onto |
THEOREM 7.4.5. A locale is spatial iff every element is the meet of meet-
irreducible elements.
PROOF. Let ~ be spatial. Then there is a canonical map x ~-~ ~. This is

surjective iff ~ is spatial. [2
A characterization of sober spaces is harder to obtain. There are various

reasons why a space can fail to be sober. First of all, it can happen that in
two different elements, say x and y, are contained in the same open sets.
Then the locale of open sets is isomorphic to the locale of the space ~ which
differs from ~ only in that x has been taken away. ( ~ is the image of :~ by a
continuous map.) Therefore we define the following.
DEFINITION 7.4.6. Let :~ -- (X,:~} be a topological space. 2~ is a T o -

space if for every pair of elements x, y E X if x ~ y there exists an open set
0 such that ~(0 n {x, y}) -- 1.
The Sierpifiski-space is a T0-space, though not a T2-space. Now we define

a relation < on points of a T0-space by x _< y r {x} _D {y}. We call < the
specialization order. (In [110] the converse ordering is considered. The order
defined here has the advantage to make the sets Sx closed and coincide with
the upper set in fl(X).)
PROPOSITION 7.4.7. A topological space is a To-space if] the specializa-

tion order is a partial order.
PROOF. Let X be a T0-space. Since {x} _~ {x} w e have x < x. Moreover,

if x _< y and y <__z then {x} _~ {y} and {y} _~ {z}, from which {x} _~ {z}, or
x ~ z. Hence _< is a partial order. Now assume that 3~ is not a T0-space. Then
there exist x, y, x ~ y such that for all open sets 0 , either 0 n {x, y} -- O or
{x, y} C_ O. Hence, for all closed sets C, either {x, y} C_ C or C n {x, y} -- ~.
It follows that y E {x} and x E {y}. So, x < y and y ~ x, that is, < is not a
partial order. E]
LEMMA 7.4.8. A space is To iff the map x ~-4 p~ is injective.
The proof of this lemma is left as an exercise.
THEOREM 7.4.9. For any locale, the space Gpr is sober.
PROOF. Let ~ be a locale and A a meet-irreducible element of ft(Gpc(.~)).

Then A is of the form ~ for some x E L. Since the map y F-+ ~ preserves infinite
joins, there is a largest such x. We show that x is N-irreducible. For otherwise,
there are y and z such that x - y R z and so ~ - ~ N ~, a contradiction to the
assumption that A is meet-irreducible. Now let p~ be the point defined by x,
x [-]-irreducible. Then
- r, t (,c ) - .
For ~ - {py " p y ( x ) - 1} - {py " x ~ y}. Thus, ~ - {py "y r j'x}. Hence,
p t ( ~ ) - ~ - {py "y >_ x}. This is a closed set. Moreover, it is the least closed
set containing x. Thus the canonical map x ~4 ~ is surjective. Moreover, it is
injective, since for different points p, q we cannot have p ( x ) - q(x) for all x.
Thus ~ contains just one of p and q and so Gpc(~) is a T0-space, and so by
the previous lemma the canonical map is injective. K]
The specialization ordering on | can be determined easily in ~. For

notice that we have
p~ _< py
r {q. q(y)- 1} g {q. q(x)- 1}

r 1} c { z . 1}
r
r x <_y.
The interest in this duality of sober spaces with spatial locales for our purposes
lies in the possibility to describe the lattices of modal logics as certain sober
spaces. The way to approach the structure of a sober space is by first studying
the specialization ordering and then looking at the topology defined over it.
However, some care is needed. Just as with boolean algebras, the space of
irreducible elements alone cannot provide a complete description of the lattice.
For notice that in many cases there are at least two topologies for a given
specialization order. Namely, given (X, <) let Y ( X , <) be the set of all lower
closed sets, that is, sets of the form $ S for some S. This is the finest topology
we can define. Also, let ~ ( X , <) be the smallest topology that contains O
and all sets of the form
X - ( t X l [..J t x 2 [..J... [_J tXn)
9 (X, _<) is called the w e a k t o p o l o g y and Y ( X , <_) the A l e x a n d r o v topol-
ogy.
THEOREM 7.4.10. Let ~ -- (X, •) be a To-space with specialization order
<. Then
<_) c x c Y ( x , <_)
PROOF. First, all sets of the form $ x for a single element x must be closed
sets. Moreover, t x - {x}. For y C {x} iff every closed set containing x also
contains y iff {y} C {x} iff y _> x. This shows the first inequality. The second
follows, for Y ( X , <) contains all sets which are lower closed. KI
THEOREM 7.4.11. Let 2. be a spatial locale. 2. is continuous iff the topol-
ogy on Gpr with respect to the specialization order is the Alexandrov topol-
ogy.
PROOF. A locale is continuous iff every [-]-irreducible element is [ 7 -
prime. Now let x be [-]-irreducible. Assume that x is not El-prime. Then
we can find a sequence yi of elements such that ~ y i _< x while for all i we
have Yi ~ x. Since we have a spatial locale, we can choose the Yi to be ~ -
irreducible. Now, Yi _< x iff {p~} ~ {py, } iff py~ ~ {Px}. By the next theorem,
lim ~ exists, since Gpc(~) is sober. We have lim ~ < x by assumption. Hence
lim ~ e {p~}. This shows that the set [.J {py, } is not closed, even though it is
upward closed. Taking complements, we see that we have a downward closed
set which is not open. Hence the topology is not equal to the finest topology.
For the other direction, just reason backwards. [-7
It is a rather intricate matter to say exactly what specialization orders
admit a sober topology. We will prove a rather useful theorem, stating that
the specialization order on a sober space must be closed under lower limits.
THEOREM 7.4.12. Let )C be sober and ~ = (xi : i E w) be a descending
chain of points. Then lim 5 exists in ~.
PROOF. Let (xi : i E w / be a descending sequence of points. Then the
sequence ({xi} : i C w). is ascending, that is, {xi} C {xj} i f / < j. Let T
be the closure of its union. Assume that T - $1 U $2 for some closed sets $1
and $2. Then almost all xi are in $1 or in $2. Hence, by directedness, either
a l l x i are in $1 or a l l x i are in $2. Thus T = $1 or T = $2, s h o w i n g T t o
be indecomposable. By sobriety, T = {y} for some y. It is easy to see that
y - lim ~. D
Such elements corresponding to limits of chains are R-irreducible elements

which are [']-reducible. We have seen that in lattices of logics a [-]-irreducible
logic is the logic of a subdirectly irreducible algebra. By the decomposition
theorem, every logic is the intersection of the Th 92 for some subdirectly ir-
reducible 92. We have seen, however, that we are not entitled to conclude
that Th 92 is [']-irreducible if 92 is subdirectly irreducible. Nevertheless, the
following theorem can be established.
THEOREM 7.4.13. In the lattices E K~, every logic is the intersection of

-irreducible logics.
PROOF. Let 0 be a logic, and ~ r | Then let S~ - {A _D 0 9 ~ r A}.
S~ is not empty, and is closed under direct upper limits. For if for a chain
An C S~, An _D An+l we have limAn r S~, then there exists an no such
that p E An o. Hence we conclude that there must be a maximal element in
S~, which we denote by S$. (This element is not necessarily unique.) Now
every proper extension of S~ contains ~, while S~ does not. Hence S w is
O-irreducible. Now 0 - N ( S ; 9 ~a r 0). U]
COROLLARY 7.4.14. For any modal logic A, E A is spatial.

We can use Theorem 7.4.13 for a sharper variant of the representation
theorem. Above we have seen that in a sober space the limit of a descending
chains always exists. This limit is not ["]-irreducible. Under mild assumptions
on the space it is redundant in the representation. Let I r r ( s be the set
of ~-irreducibles; and ~ ( ~ ) : = (Irr(~), ~). Then define the topology as
before, namely putting & = {p e Irr(~) : p(x) = 1} = ~ A Irr(2.). Let
3|162163 : (Irr(s { k : x e L}).
DEFINITION 7.4.15. A topological space is called a T D - s p a c e if for every
point x the set {x} is relatively open in {x}.
THEOREM 7.4.16. The natural map x ~ 5c 9 2. --+ O(]Gpc(~)) is an
isomorphism iff every element is an intersection of N-irreducible elements.
We obtain that every lattice of extensions is representable by a TD-space.
Let us make this explicit. With a logic A we associate a locale ~ A, and two
spaces. The first is | A). It consists of the n-irreducible logics, ordered
by set inclusion. The topology has as closed sets the sets of the form ~ O, where
O 2 A. The second space is 3| A). Its members are the ~ - i r r e d u c i b l e
logics ordered by set inclusion. Again, the closed sets are the sets of the form
j'O for O _~ A. It is immediately clear that 3| A) is nothing but the
relativization of the space Gpc(E A) to the set of [']-irreducible elements, and
so uniquely defined. We have proved in addition that given 3Gpc(E A), the
space | A) is uniquely determined. It can in fact be constructed. (See
the exercises.)
7.5. Some Consequences of the Duality Theory 353
E x e r c i s e 239. Prove Lemma 7.4.8.
E x e r c i s e 240. Take the lattice ~ - 1 + (Z • 2) § 1, which is isomorphic to

the direct product of Z with the two-element chain plus a bottom and a top
element. Show that ~ is a locale, but not spatial.
E x e r c i s e 241. For a topological space :~, the s o b e r i f i c a t i o n is the space

| Show that for lattices of modal logics, Gpr is the soberification
of fl|162 (This shows how | can be recovered from ~|162
E x e r c i s e 242. Show with a counterexample that ~(X, ~) does not neces-

sarily contain only sets of the form
X - ( ~ ' x 1 [..J ~ ' x 2 U . . . [_J ~Zn)
(in addition to 0). Thus depending on the properties of <, the topology
9 (X, _<) may contain more sets.
7.5. S o m e C o n s e q u e n c e s of t h e D u a l i t y T h e o r y
Lattices of (normal) extensions of a logic are algebraic, with the compact
elements being the finitely axiomatizable logics; we have seen also that in the
lattice ~ A every element is the intersection of ~-irreducible elements. The
finitely axiomatizable logics are closed under finite union, just as the compact
elements. An infinite join of finitely axiomatizable logics need not be finitely
axiomatizable again. Likewise, the finite meet of finitely axiomatizable logics
need not be finitely axiomatizable. However, in the case of weakly transitive
logics, any finite intersection of finitely axiomatizable logics is again finitely
axiomatizable.
DEFINITION 7.5.1. A locale is c o h e r e n t if (i) every element is the join of
compact elements and (ii) the meet of two compact elements is again compact.
Coherent locales allow a stronger representation theorem. Let ~ be a
locale, K ( ~ ) be the set of compact elements. They form a lattice ~(~) "-
(K(~), R, U}, by definition of a coherent locale. Given ~(~), ~ is uniquely
identified by the fact that it is the lattice of ideals of ~(~).
LEMMA 7.5.2. A locale is coherent iff it is isomorphic to the locale of
ideals of a distributive lattice.
PROOF. Let ~ be coherent. Denote by Id(~(2.)) the set of ideals in ~(~);
moreover, let f l ~ ( ~ ) ' - (Id(~(~)), A, [.J'). Here, if Ic, c e C, is a family of
ideals, [.J'c~c I~ is the least ideal containing [-J~ec I~. This is a locale; the
compact elements are of the form $S where S is finite. For x E L let x*
be defined by x* := {y E K ( ~ ) 9 y _< x}. We show t h a t this m a p is an

isomorphism from ~ onto ~ ( ~ ) . First, x* is clearly an ideal. Moreover, if
x < y then x* C_ y*. Conversely, let I C_ j~(~) be an ideal; then p u t / . := [_]I.
If I C_ J then I . < J , . Both maps are order preserving; it is therefore enough
to show t h a t one is the inverse of the other. We show that x - (x*), and t h a t
I - (I.)*. For the first, observe that x > (z*), generally holds in a lattice.
Since by assumption every element is the union of join compact elements, we
also have (x*), >_ x. This gives x - (x*).. For the second claim, let y be
compact and I an ideal. Suppose that y E I. Then y < [.JI - I , and so
y C (I,)*. Conversely, if y C (I,)* then y _< I . - [..]I. By compactness, there
exists a finite subset I0 C_ I such that y < kJI0. Since I is closed under finite
joins, [_]I0 E I, and so y C I. Hence I and (I,)* contain the same compact
elements. So they are identical subsets of K ( ~ ) . Now let ~ be a distributive
lattice. Then ~ ( ~ ) - ( I d ( ~ ) , A , [.J') is coherent. An ideal is compact iff it
is principal, that is, of the form Sy for some y. Since Sy N Sz - $ ( y ~ z),
principal ideals are closed under meet. Furthermore, suppose t h a t I is an
ideal. Then I - [.J'x C I Sx " So 3 ( ~ ) is a coherent locale D
If we have a lattice homomorphism ~ ( ~ ) --+ ~i(gX) then this m a p can
be extended uniquely to a homomorphism of locales ~ -+ ff)l. Not all locale
homomorphisms arise this way, and so not all locale maps derive from lattice
homomorphisms. Hence call a map f 9 ~ --+ ~ c o h e r e n t if it maps compact
element into compact elements.
THEOREM 7.5.3. The category Dkat of distributive lattices and lattice ho-
momorphisms is dual to the category Cohkoc of coherent locales with coherent
maps.
Now if A is weakly transitive, the intersection of two finitely axiomatizable
logics is again finitely axiomatizable. Now, a logic is compact in ~ A iff it is
finitely axiomatizable over A. We conclude the following theorem.
PROPOSITION 7.5.4. Let A be weakly transitive. Then ~ A is coherent.
The converse need not hold. In Chapter 9 we will see t h a t ~ K . a l t l is
coherent (because every logic in this lattice is finitely axiomatizable) but the
logic is not weakly transitive.
The next property of lattices has already made an appearance earlier,
namely continuity. It is connected with a natural question about axiomatiz-
ability of logics.
DEFINITION 7.5.5. Let 2. be a complete lattice. A set X C_ L is a g e n e r -
a t i n g set if for every member of L is the join of a subset of X . ~ is said to
have a basis if there exists a least generating set. Moreover, X is a s t r o n g
basis for ~ if every element has a nonredundant representation, that is, for
each x there exists a minimal Y C X such that x - kJY.
THEOREM 7.5.6. Let 2, be a locale. 2` has a basis iff (i) 2, is continuous

and 5i) every element is the meet of [7-irreducible elements. 2` has a strong
basis iff it has a basis and there exists no infinite properly ascending chain of
- p r i m e elements.
PROOF. Assume t h a t (i) and (ii) hold. Let I be the set of ~ - i r r e d u c i b l e

elements. Since 2` is continuous, I is also the set of [-]-prime elements. Hence
every element of I splits 2`. Take an element x C L. Let J " - Sx N I. T h e n
put K " - I - J and xo " - 2 ` / K - - [-JyEK2`/Y" Since for no y C K it holds
t h a t x < y, we have x > 2`/y; and hence x >_ x0. We also have x0 < x
since x0 ~ q for all q E J , and so also x0 < [ - ] J - x. Hence x is a union of
elements of the form 2,/y, y [-]-irreducible. We have to show t h a t the set X
of U - i r r e d u c i b l e elements is minimal. The elements of X are also U - p r i m e
and hence compact. Moreover, let Y C X. Say, y C X - Y. T h e n for no
subset U C_ Y we can have [.JU - y, since y is [_J-irreducible. Thus X is a
basis. Now let 2` have a basis, X. Let x E X. Assume t h a t x - LaY for some
Y - {yi 9 i E I}. By assumption, each yi is the join of some set Xi, Xi C_ X .
P u t X0 " - [-JicIZi" T h e n x - [.JX0. Suppose x ~ X 0 . T h e n X - { x } isa
generating set, contradicting our assumption on X. Consequently, X consists
of U - i r r e d u c i b l e elements. 2` is a locale, and so X consists of the [.J-prime
elements. Hence, each element y E X has a splitting companion y,. Now
take x E L. P u t U := ( X - l"x). Consider the element x ~ : - N y c u Y , .
Clearly, x _< x~ for i f y E U t h e n x ~ y, s o y , _ x. Now, i f x r x ~ there
exists a y E X such t h a t y _ x ~ but y ~ x. T h e n x :> y,. Hence x ~ > y,.
Contradiction. So, x - x ~ Hence, (ii) is fulfilled. Moreover, each element is
the meet of prime elements. This implies t h a t (i) holds. For assume t h a t u is
~ - i r r e d u c i b l e and the intersection of a set of [-]-prime elements. T h e n this
set is o n e - m e m b e r e d , and u is therefore also [']-prime.
Now let us t u r n to strong bases. Assume t h a t 2` has a base and let
(x~" i e w) be a strictly ascending chain of ~ - p r i m e elements. T h e n (2`/xi "
i E w) is a strictly ascending chain of [_J-prime elements. Let y be its limit.
T h e n there exists no minimal set Y of [.J-primes whose join is y. Thus 2`
fails to have a strong basis. Assume on the other hand t h a t there exists no
such chain. T h e n for each upper closed set P of ~ - p r i m e elements the set
m a x ( P ) " - {x E P " (Vy E P ) ( y >_ x ~ y - x)} is well-defined and is the
least set Q such t h a t $ Q - P.
PROPOSITION 7.5.7. Let 2~ be a locale with a strong basis. Then the ele-
ments of 2, are in o n e - t o - o n e correspondence with antichains in ~Gpc(~).
PROOF. By T h e o r e m 7.2.7 the map p ~ p* " - 2,/p induces an order

isomorphism from the poset of [-]-primes onto the poset of U - p r i m e s , whose
inverse is q ~-+ q,. Let x be an element. T h e n let Y be a minimal set of
[.J-primes whose join is x. T h e n Y is an antichain. T h e n Y, is an antichain of
[-]-primes. Put Xo := FLY,. Conversely, given an antichain Z of [_J-primes,

let Z ~ := LJZ*. Then x = (x o) ~ as well as (Z~ so this is a bijection. D
THEOREM 7.5.8. Let A be a modal logic. Then E A has a basis iff E A is
continuous.
Since continuous lattices are the exception in modal logic, most extension
lattices do not have a basis. We can sharpen the previous theorem somewhat
to obtain stricter conditions on continuity.
COROLLARY 7.5.9. Let A be weakly transitive and have the finite model
property. Then the following are equivalent.
1. s A has a basis.
2. s A has a strong basis.
3. Every extension of A has the finite model property.
4. Every extension of A is the join of co-splitting logics.
5. Every join of co-splitting logics has the finite model property.
PROOF. The most interesting part is perhaps the fact that if we have a
basis, then we already have a strong basis. But this follows, because there is
no infinite strictly ascending chain of [']-irreducibles. For the [-]-irreducible
logics are the logics of finite frames, and for finite frame ~, g we have t h a t
Thg D Thf implies ~g _< ~f. If A is weakly transitive then every finite
subdirect irreducible algebra induces a splitting. On the other hand, if A has
the finite model property, then no more elements can induce splittings. The
equivalence now follows directly. D
COROLLARY 7.5.10. Let Alga be locally finite. Then s A is continuous.
PROOF. Since Alga is locally finite, every extension of A has the finite
model property (see Theorem 4.8.7). Moreover, the one-generated A-algebra,
~rh({P}), is finite. Now consider the elements E]p, 7H a compound modality.
Up to equivalence there exist only finitely many of them. So a largest com-
pound modality exists. So, A is weakly transitive and has the finite model
property. By Corollary 7.5.9, s A has a basis, and by Theorem 7.5.8 it is
therefore continuous. F]
The converse does not hold. The lattice s $4.3 is continuous but S4.3
fails to be locally finite. Now, finally, consider the question of finite axiom-
atizability. In lattices which have a basis, this can be decided rather easily.
Given a set S, a relation -< is called a w e l l - p a r t i a l o r d e r (wpo) if it is a par-
tial order and there are no infinite strictly descending chains, and no infinite
antichains. There is a rather famous theorem by J. B. KRUSKAL which says
that the set of finite trees under the embedding-ordering is a wpo ([135]).
THEOREM 7.5.11. Let s be locale with a basis. Then the following are
equivalent.
1. >_ is a well partial order on 3|

2. 2. has a strong basis and no infinite antichains exist in 3|
3. Every element is a finite union of U - p r i m e s .
4. s ~- ~(~2).
5. Every strictly ascending chain in 2, is finite.
PROOF. If 12 has a basis it is continuous and so the properties on the poset

of [-]-irreducible elements can equivalently be checked on the poset of U -
irreducible elements. Let (1.) be the case. Then (2.) holds by Theorem 7.5.6.
Now consider an element x of t2 and let Y be a minimal set of [ J - p r i m e s such
that x - LjY. Then Y is an antichain, and so Y is finite. Now, if every
element is a finite union of join-primes, then every element is compact. Now
let (xi " i C w) be an infinite strictly ascending chain. Then its limit cannot be
compact. Finally, assume that there are no infinite strictly ascending chains
in 12. Then there are no strictly ascending chains in Irr(2.). Furthermore, if
X is an infinite antichain, then we can choose an infinite ascending chain of
subsets of X corresponding to an infinite ascending chain of elements in 12.
Hence _> is a wpo, as required. 7]
COROLLARY 7.5.12. (n _< 1~0.) Let ~ A have a strong basis. Then the
following are equivalent.
1. Every extension of ~ A is finitely axiomatizable.
2. ~ A is finite or countably infinite.
3. There exists no infinite set of incomparable splitting logics.
There are lattices of modal logics in which there are infinite ascending chains
of irreducible elements and infinite antichains. The latter has been shown in
[61] (see exercises below). Another example is the logics of Chapter 2.6. Here
we will produce an infinite ascending chain of irreducible elements; a first
proof of this fact was given by BLOK [23]. Our example will allow to prove a
number of very interesting negative facts about modal logics in general. Let
and | be frames. Then let ~ Q (5 be the following frame.
h := S • {0}ug• {1}
<~h := 0), (y, 0 ) ) . y e S, 9 <s y}
U { ( ( x , l ) , ( y t l ) } 9 x , y e g , x <::]gy}
U {((x, O), (y, 1 ) ) ' x e f , y E g}
H 9= {ax {O}Ubx {1}.aEF, beG}
9- (h, <h,R)
Moreover, if a is an ordinal number, let g_ " - (a, 2). We are interested in
the logic of the frames of the form .@(g_ @ ~ ) @ 7 for infinite a and/3. In
m m
Figure 7.6 the frame o@ (w__| w__)@n is shown.

FIGURE 7.6. ,, (~ (w__~ w___() ~ n
~e "=~- 1 n...- 2 1 0
. . . A ~ A
v v w
Consider the following formulae.

~o "- poA~p~A[:](-~poA~p~)
~ "-- ~Po Apl A [](~po A ~p~)
The logic G.gt2 is defined as follows.
G.~t2 := K . G 9 OPo A (}Pl A OP2 -+ Vi<j<3 O(Pi A pj) V V i i i O(Pi A OPj)
9 0~o A 0 ~ -~ D-~(O~o A 0 ~ )
THEOREM 7.5.13. Every extension of G.~2 is complete with respect to
frames of the form 5) "(~ (w__| w__)~_, ~ <_ w, or (ii) 2, ~ <_ w.
The Theorem 7.5.13 is proved as follows. Every extension of G.~-~2is com-
plete with respect to simple noetherian frames, by Theorems 8.6.14 and 8.6.15
of Section 8.6. Moreover, it is easy to see that the Kripke-frames underlying
the reduced canonical frames for G.f~2 have the structure . @ (a_@/3) @7, for
certain ordinal numbers a,/3 and 7. Furthermore, if a and/3 are nonzero, they
must be infinite. Let A D G.fi2, and let T r A. Then there exists a model
(~,/3, x) ~ -,~ based on a generated subframe of a reduced weak canonical
frame. Let y E f; define C ( y ) : = {X e s f ( ~ ) ' y ~ X}. Call y ~ - m a x i m a l if
for every z E> y such that C(z) - C(y) also z <] y. Let D be the set of points
which are ~-maximal. D has <_ 2. ~sf(~) points. Consider the subframe
of ~ based on D. It is cofinal, that is, it contains all points of depth 0. Let
V(P) "-/3(p) N D. There exists a weak successor y e D of x and (~,/3, y) ~ ~.
can be partitioned into four possibly empty pairwise disjoint frames ~ , 94,
~ , 3, each linearly ordered by <1 such that ~ - 3@(92 | ~ ) @ ~ - Observe
that if 92 is empty, we can choose ~ in such a way that it is empty, too.
C a s e 1. 94 or ~ is empty. Then ~ is linear. This case is rather straightfor-
ward.
C a s e 2. Both P2 and ~ are nonempty. Then 3 is nonempty and so 3 ~ [ ~ .
Let r - 3 Q ((a_l @ 92) | (~----2G ~ ) ) Q (~----3Q ~ ) for certain ordinal numbers
a l , a2 and a3. r is a G.~2-frame. Let ~ be the subframe of (5 based on
the union W U A U B U Z. It is possible to extend the valuation ~y on ~ to a

valuation (~ on ~ such that (i) ~ ( p ) A E - ~/(p), (ii) each ~(p) is a finite union of
intervals and singletons, (iii) (~, 5, y) ~ ~ . (Namely, let x E 5(p) for x E a l
iff for the root Yl of 91, y E ~'(p). Let x E 5(p) for x E a2 iff for the root Y2 of
!t3, y E 7(P), and let x E ~(p) for x E a3 iff for the root y of ~ , Y3 EE 7(P).)
Now, the size of ~ depends on ~. However, it is always finite. Hence
we obtain that the logic of 9 (a_ @/3) Q T, where a <_ /3 is identical to the
logic of 9 ~ @/3') Q T ' , where a ' < / 3 ' if (i) a - a ' , / 3 - / 3 ' or a - a ' and
/3, ~' >_ w or a, a',/3,/3' > w, and (ii) 7 - 7' or 7, 7' > w. This completes the
proof of Theorem 7.5.13. We deduce the following.
THEOREM 7.5.14. 8 G.t22 ~- w + 2 + w ~
PROOF. It is a m a t t e r of direct verification (using some formulae) that

the logics Th 7, ~Y -< w, as well as Th 9 @ (w 9 w__)@~y, 7 -< w, are pairwise
m
distinct. Now, the theorem is established by the following facts.

1. m ,-+ n iff m < n.
2 . . @ ( ~ _ 9 ~ _ ) @ ~ -~ . | (~_ 9 ~_)| iff .~ > n.
3. xh (. @ (~ 9 ~) @_~) c Th (. @ (~ 9 ~) | ~).
4. n ~ ~ ~ . @ ( ~ e ~_)G~_.
This ends the proof. [2
Call a set A of formulae i n d e p e n d e n t if for every 5 E A we have

5 r g | ( A - {5}). (For example, a basis is an independent set.) A logic
A is i n d e p e n d e n t l y a x i o m a t i z a b l e if there exists an independent set A
such that A = K | A. Every finitely axiomatizable logic is independently ax-
iomatizable. It has been shown in CHAGROV and ZAKHARYASCHEV [42] that
there exists a logic which is not independently axiomatizable. Furthermore,
KRACHT [125] gives an example of a logic which is not finitely axiomatiz-
able, but all its proper extensions are. Such a logic is called p r e - f i n i t e l y
a x i o i n a t i z a b l e . Here is a logic that has both properties.
THEOREM 7.5.15. The logic of the frame 9 @ w_)@w__ is pre-finitely

axiomatizable. It splits the lattice of extensions of G.~t2. Moreover, it is not
axiomatizable by a set of independent formulae.
PROOF. The first two claims are immediate. For the last, let O " - Th 9
@ (w__@ w__)@w__. Let A be a set of formulae such that (9 - K • A. G.~I2
is finitely axiomatizable. Hence there exists a finite set A0 C A such that
G.t22 C K | A0 C O. Let ~, 5 ~ E A - A 0 be two different formulae. Then
eitherK~)A 0|174 0@~orKOAOG~ CKOA0| Hence the
set A is not independent. F-1
FIGURE 7.7. g G.~2
Th 1
I
Th2m
Th3I
G.3 I Thco
O Th 9 (~) (co @ co)@ w__
Th 9 (~ (co @ co) Q_3
Th 9 Q (co @ ~) G
Th 9 Q (co@ co)Q 1_
G.a2 Th 9 @ (~_9 ~ ) @ o
Moreover, with this logic we have an example of an algebra that induces

a splitting but is not finitely presentable9 This shows that the generality of
Theorem 7.3.11 is really needed.
THEOREM 7.5.16. Let 92 be the algebra generated by the singleton sets of

9 Q (co @ co)(~)co is not finitely presentable. Its logic splits ~ G.f~2.
PROOF. Take the freely 2-generated G.~t2-algebra, with the generators

a and b, and consider the open filter F generated by the set E.
E "- {a -+ - b n m ( - a n - b ) , b --+ - a n m ( - a n - b ) }
U {(a U b) --~ r n -r0)- n e w}
We wm how that that for every finit subset Eo of
E and filter F0 generated by E0, 92 is actually not isomorphic to the quotient
~re({a, b})/Fo. The second claim is easy. For if E0 C_ E is finite, for some n,
(aUb) -+ r r E0. It is consistent to add (aUb) nmn+10, that
is to say, adding that formula does not generate the trivial filter. So, no finite
subset is enough to generate the filter of E. Now we show the first claim. Let
us look at the reduced 2-canonical frame rather than the 2-canonical frame.
(See Section 8.6 for a definition.) Let W be the set of (noneliminable) points
satisfying - a n - b n m ( - a N -b). This set is the set of 0-definable points. It
is not hard to see that it is linearly ordered by <1. Moreover, it is not finite,
by choice of E. Moreover, we can show that every world of infinite depth is
eliminable, and so that (W, <~} is isomorphic to ~. To see that, it is enough
to show that for every formula p(p, q), ~(a, b) is either finite or cofinite. This
is shown by induction on p. Now, let u be of depth w in W, and let r b)
hold at a point of infinite depth in W. We may assume that ~ is strictly
simple (see Section 1.7 for a definition). Any set (Ox)(a,b) containing u is
cofinite, and so is every set ([~x)(a, b) containing u. Now, a nonmodal formula
#(a, b) containing u holds at all points of W. So, u is eliminable. However,
the frame consists only of noneliminable points. This shows that (W, <l) - w__.
Now, every world o f - W must see all worlds of W. Moreover, it must be
in a U b U Ca U Cb. Let A be the set of worlds in (a U C a ) A - r B the set
of worlds in (b U Cb) N - C a . Finally, Z "- - ( W U A U B). Every world of
Z is in Ca and in Cb. For otherwise it contains m - a or m _ b. Suppose it
contains both; then it is in W. Suppose it contains only one, say m _ b. Then
it does not contain m _ a, so it contains Ca. But then it is a member of A.
Contradiction. By the axioms of G.gt2, if x E Z, then x has no successor in
Z. So, Z has one member only and it generates the frame. Both A and B
are linearly ordered. This follows from two facts. (i) No member of A sees a
member of B, and no member of B sees a member of A. (ii) There are no
three incomparable points. Both A and B must be infinite, by the axioms of
G.gt2. Hence, by the same argument as for W, they are isomorphic to w_. [~
E x e r c i s e 243. Show that a coherent locale is spatial.
E x e r c i s e 244. Let Grz3 be the logic of G r z - s t r u c t u r e s of depth 3. Show

that Grz3 is locally finite but that ~ Grz3 is uncountable. Hint. The frames
for Grz3 are simply characterized by the fact that they are a poset in which
no strictly ascending chain of length 4 exists.
Exercise 245. Show that there exists a logic | such that all extensions of
O are finitely axiomatizable over O but not all extensions are finitely axiom-
atizable over K.
7.6. P r o p e r t i e s of L o g i c a l C a l c u l i a n d R e l a t e d Lattice P r o p e r t i e s
In this section we will map out the distribution of logics that have a certain
property in terms of their closure under lattice operations. Some easy facts are
the following. Given a class ~ of frames, the logics which are ~ - c o m p l e t e are
closed under (infinite) intersection. The set of logics which are ~ - p e r s i s t e n t
are closed under (infinite) union. Moreover, it can be shown t h a t logics which
are ~ - e l e m e n t a r y for some modal class 2E of frames form a sublocale in the
locale ~ A. We will show here a special case, the most i m p o r t a n t one, namely
that of the Sahlqvist logics (see [127]). The general fact is proved the same
way. Recall that | denotes the class of logics axiomatizable by a set of
Sahlqvist axioms of Sahlqvist rank n.
THEOREM 7.6.1. (~ < ~0.) The logics Gqn form a sublocale of the locale
K~ of x-modal logics.
PROOF. The only thing which is not straightforward is the closure un-
der meet. To this end take two elementary Sahlqvist formulae of rank n,
(Vx)c~(x)and (Vx)/3(x). We want to show that ( V x ) a ( x ) V (Vx)/3(x) is again
Sahlqvist of rank n. Define formulae "Yk, k E w, by
j<k j<k
where x <1j y iff there exists a path of length j from x to y. Observe now t h a t
~ (Vx)c~(z) V (Vx)/3(x) iff ~ ~ {'Yk: k e w}
For if ~ ~ (Vx)c~(x)V (Vx)/3(x) then ~ ~ (Vx)c~(x) or ~ P (Vx)/3(x). Assume
withouth loss of generality the first. Thus ~ ~ (Vx)(Vy C>j x)c~(y) for all j and
consequently ~ ~ "Yk for all k. For the converse assume ~ satisfies all "Yk- Take
a world w E ~. Then for an infinite number of k E w we have either
~ {(Vy C>j x)c~(y)[w]: j <_ k}
or
Let the first be the case. Denote by [5 be the subframe generated by w in ~.

Then r ~ (Vx)a(x). Consequently,
|
This holds for all generated subframes, and so it holds for ~ as well, since
c~(x),13(z) are restricted. Secondly, for every k, 0'k is Sahlqvist of rank n.
Two cases need to be distinguished. First case is n = 0. Then "Yk is constant
7.6. Properties of Logical Calculi and Related Lattice Properties 363
and so Sahlqvist of rank 0. Second case n > 0. Since the formula begins with
a universal quantifier and (Vy t:>j x) is a chain of universal quantifiers, the
rank of 7k is the m a x i m u m of the ranks of a and/3, hence at most n. [--1
As a particular application of the splitting theorem we will prove that

there are effective means to axiomatize tabular logics and that the tabular
logics form a filter in the lattice of logics. This is by no means trivial to
show, and requires some advanced methods, despite the seemingly simple
character of the generating algebra. In general, these results have been shown
to hold in congruence distributive varieties by BAKER in [2]. However, his
proof uses abstract algebra, whereas here we can make use of geometric tools.
Moreover, some results can be sharpened. Crucial to the analysis are two
families of logics, the family of weakly transitive logics and logics of bounded
alternativity. Recall from Theorem 3.2.12 that logics of bounded alternativity
are canonical. Consider now a logic which is m - t r a n s i t i v e and satisfies altn
for some m, n C w. Then each rooted subframe of the canonical frame has
am+ 1
at most n-1- 1 many points. For, by induction, the kth wave consists of at
most n k many points and the transit of a point is the union of the k-waves
for k < m, by m-transitivity. So, there are only finitely many non-distinct
frames in the canonical frame.
THEOREM 7.6.2. A logic is tabular iff it is of bounded alternativity and

weakly transitive.
PROOF. Let O be tabular, say, O = Th [ for some finite [. Let n = ~f.

Then O is n-transitive, and each point has at most n successors. Conversely,
let O be weakly transitive and of finite alternativity. Then O is canonical,
hence complete. Each rooted subframe for O is finite and bounded in size by
1 + n + n 2 + . . . + nm. So there are up to isomorphic copies finitely many
rooted frames for O. Their disjoint union, fl, is finite and O - Th g. [--]
COROLLARY 7.6.3. The tabular logics form a filter in the lattice of modal
logics.
How can we axiomatize the logic of a single frame [ if it is finite? First, we

know that there is an axiom of weak transitivity satisfied by [, say trsm.
Hence, we know that all finite frames for Tb [ induce a splitting of the lattice
of extensions of K . t r s m . Moreover, there is an axiom alt~ satisfied by [, so in
fact all rooted frames are finite. Hence we have the following representation.
Th J: - K . a l t , ~ . t r s n / N
where N is the set of logics of rooted frames 9 which are not generated sub-
frames of p - m o r p h i c images of [ and which are frames for K . a l t m . t r s n . There
are finitely many of them.
FIGURE 7.8
1
3 9 Th[-~
O
PROPOSITION 7.6.4. Tabular logics are finitely axiomatizable and oS finite

codimension.
The converse does not hold. We give an example in monomodal logic,
namely the logic of the veiled recession frame. It is of codimension 2. Its
lattice of extensions is depicted in Figure 7.8; it is isomorphic to (3, _<}. Let
r denote the frame defined by r " - (w, <~} where n <1 m iff n _< m + 1 (see
[25]). Let 91 be the algebra of the finite and cofinite subsets of w and O " -
Th ~3 - K(V-10p -+ 0[:]2p, [-lp -+ p, 0p A E](p -+ Dp) --+ p). 9l is 1-generated;
simply take {0} C_ w. It is not entirely simple to see that the logic of the
veiled recession frame is ~ - p r i m e in its own lattice. However, consider the
structure of the frame ~ro (1). It is constructed by taking a product of 91, each
factor representing a possible valuation of p into the algebra JR. The canonical
value of p in that frame is then the sequence of the values/3(p). Crucial is
the fact that a valuation sends p to a non-trivial set, that is, a set ~ O and
r, iff it does not send (}p A -~p to 0. Notice on the other hand that a set
generates ]R exactly if it is nontrivial. (This will be an exercise.) Now consider
an extension A ~ O. Then there is a nontrivial map h" ~ro(1) --~ ~rA(1).
By the fact that O is the logic of the veiled recession frame, ~rA(1) must not
contain any generated component looking like the recession frame. Hence, all
components have been killed by h, that is, h(Op A -~p) -- ~, since the latter
formula defines exactly the generated components looking like veiled recession
frames. To put this differently, h(Op -+ p) - i, that is, 0p -+ p 6 A, which
means that A contains the logic of the one-point reflexive frames. Hence 91
is prime in 80 and induces a splitting 80/0 - Th [~-]. It is no coincidence
that the counterexample is a logic of depth 2, by the Theorem 2.9.9. We will
show later that with two operators there are 2 ~~ many logics of codimension
i.
Our last example concerns the problem of closure under union of com-
pleteness properties. We have seen that tabular logics are closed under union.
BLOK [21] has first given an example of two logics which have the finite
model property such that their join does not. Here we will show that there
are complete logics such that their join is incomplete. The example is based
on [62], one of the first examples of an incomplete logic. (The article by FINE
appeared in the same edition of the journal where THOMASON published his
7.6. P r o p e r t i e s of Logical Calculi a n d R e l a t e d Lattice P r o p e r t i e s 365
FIGURE 7.9
i/
[207]. T h e two results have been i n d e p e n d e n t l y obtained.) Consider the
frame of F i g u r e 7.9. T h e frame is d e n o t e d by [. T h e first t h r e e rows of points
are of finite d e p t h , the lowest consists of points of infinite d e p t h . Notice t h a t
the points of infinite d e p t h form an infinite ascending chain. We refer to the
first two rows as the s p i n e of the frame, the points of the t h i r d row are called
n e e d l e s . T h e lowest row is the h e a d . Consider the algebra F g e n e r a t e d by
the finite subsets of the spine. This set contains all finite sets of needles. It
can be shown inductively t h a t each element of F has a m a x i m a l point. T h a t
is, if a C IF t h e n for every x there exists a y C a such t h a t x <~ y and for all
y < z E a we have z = y. This holds of the g e n e r a t i n g sets, a n d if it holds of
set a, b t h e n o f - a , of Ca, and of a U b. Hence ~ V G r z . O n the other h a n d
f~e G r z
THEOREM 7.6.5 (Fine). The logic of ~ is incomplete.
PROOF 9 Consider a formula ~p such t h a t p can be satisfied in ~ u n d e r a
v a l u a t i o n 3 at a point w0 iff there exist points zm, z2, z3, y a n d x such t h a t
(i.) w0 <~ x <l y <~ Zl; z2; z3 and z2 < z~, zl :~ z3, z3 :~ zl, z :~ w0, y :~ x, (ii.)
for all v such t h a t z2 <l v either v = z2 or v = z3, (iii.) for all v such t h a t y <l v
either y = v or Zl <1 v or z2 <l v. We leave it to the reader to c o n s t r u c t such a
formula p. T h e n ( ~ , 3 , w0) ~ ~ iff w0 is in the head of the frame. Moreover,
if ~ is a frame for Th ~ such t h a t | ~ -~p t h e n its u n d e r l y i n g K r i p k e - f r a m e
contains an infinite strictly ascending chain of points and so the K r i p k e - f r a m e
on which it is based is not a frame for Th ~. Hence there exists no m o d e l for
based on a K r i p k e - f r a m e . T h u s Th ~ is incomplete. Ill
COROLLARY 7.6.6. The logics Tb. f and G r z are complete9 Their union
is incomplete, however.
We s u m m a r i z e the facts in Table I. w e include here not only facts a b o u t
closure u n d e r finite union and finite intersection b u t also a b o u t u p p e r and
TABLE 1. Closure of Properties of Logics
sup inf U M
tabularity yes no yes yes
finite model property no ye~, no yes
completeness no ye~, no yes
compactness ? no no ?
canonicity yes no yes yes
A-elementarity yes no yes yes
finite axiomatizability no no yes ?
decidability no no no yes
subframe logic yes ye~, yes yes
Halld~n-completeness yes no no no
interpolation yes no no no
lower limits. Not all facts have been shown so far. T h a t the finite model
property and completeness may be lost under suprema is left as an exercise.
FRANK WOLTER in [240] gives an example of two compact logics whose join is
not compact. G . 3 is the infimum of logics which are compact, A - e l e m e n t a r y
and canonical, but G . 3 is neither. Nonpreservation of decidability under join
is shown in Section 9.4. Nonpreservation of decidability under suprema is
straightforward. Now consider the frames Pn (n § 1, <1) where i <l j iff
----
i - - j = 0 or i < j. Let P := {Pn : 0 < n C w}. These frames are frames

for K 4 . 3 . Clearly, the logic of finitely many such frames is decidable (it is
tabular), but also the logic of any cofinite set R. For this logic is simply
K 4 . 3 / Q , where Q = P - R. So, consider the map n sending each subset
of P to the logic of its frames. (This is similar to t ( M ) of Section 2.6.)
Clearly, the splitting formula for pn is satisfiable in n(M) iff n E M . So,
n(M) is undecidable for a nonrecursive set M and hence it is not finitely
axiomatizable. This logic is the supremum of finitely axiomatizable logics
with finite model property. These logics are decidable. Yet, n(M) is not
decidable. The logics n(M) are also useful in showing t h a t A - e l e m e n t a r y
logics as well as finitely axiomatizable logics are not closed under infima.
From Theorem 9.4.6 one can easily deduce that Halld~n-complete logics are
not closed under meets. Halld6n-complete logics are closed under suprema
just like logics with interpolation. The proof is rather straightforward. Logics
with interpolation are not closed under intersection or meet. This follows
from the classification of logics containing G r z with interpolation (see [148]).
7.6. Properties of Logical Calculi and Related Lattice Properties 367
T h a t subframe logics are closed under all these operations has been shown in
Section 3.5.
There is also the notion of a bounded property. A logic A is said to bound
a property T if A does not possess T but all its proper extensions do. For
facts about bounded properties see Section 8.8.
E x e r c i s e 246. Show Corollary 7.6.3.
E x e r c i s e 247. Show that a set in the veiled recession frame generates the
algebra of finite and cofinite sets iff it is not empty and not the full set.
Hint. Proceed as follows. Show that there must exist a set of the form
In) - { m ' n < m}. Then all [k) for k :> n exist. Then we have almost all
singletons {k}. Now observe that 0{n} - { n - 1, n,n + 1} and show that we
get all other singletons as well.
E x e r c i s e 248. Let [ be a finite frame and let Th [ split s Show t h a t there

is a formula in r2log(~f§ 1) 7 variables axiomatizing 0 / [ . Hint. Use a different
axiomatization than developed above. Instead of adding altn add an axiom
saying that the frames must have at most ~f points.
E x e r c i s e 249. Using the fact that $4.3 is axiomatized over S4 by splitting

the following two frames, show that there is no axiom for .3 using one variable
only. However, show that $4.3.t can be axiomatized over by axioms using a
single variable.
E x e r c i s e 250. Let 0 have the finite model property. Show t h a t any logic in
0 has a lower cover.
E x e r c i s e 251. Construct a formula ~ satisfying the requirements of the

proof of Theorem 7.6.5. Supply the remaining details of the proof!
* E x e r c i s e 252. Show that there is an ascending chain of logics which have

the finite model property such that their supremum fails to have the finite
model property. Hint. P u t Pn "- {P,~ " 0 < m _< n} with Pn as above.
Next put On " - K4/Pn and | " - K4/P. Clearly, | is the supremum of
the On. Show that (a) 0~ has the same finite models as G, (b) K 4 / p n is
axiomatizable by a constant formula. Deduce that 0,, has the finite model
property for all n. (c) O ~: G.
E x e r c i s e 253. (Continuing the previous exercise.) Show that O~ has in-

terpolation. (So, there exist logics with interpolation but without the finite
model property.)
E x e r c i s e 254. Show that there exist logics with interpolation which are un-
decidable.
E x e r c i s e 255. Let 0 be the modal logic of frames (N, <~1, 42, <~3, <::]4,<~5},
where (N, <~1} is isomorphic to the set of natural numbers and the successor
function, <]2 - <]1, <~3 - <~+, <]4 - <1+, and finally <]5 a subset of the
diagonal of N. Show that 0 has up to isomorphism only frames of the form
specified. Consider extensions of the form
0~.- Oe{[3~• ~ Ds• < m}.
Let O~ := sup{Ore "m C w}. Show that O~ is consistent and has no flames.
Clearly, all | are complete. Conclude that completeness is not preserved
under suprema.
E x e r c i s e 256. Show that 0-axiomatizability is preserved under join and

meet, and under suprema. (It is not known whether 0-axiomatizability is
preserved under infima.)
7.7. S p l i t t i n g s of t h e L a t t i c e s of M o d a l Logics a n d C o m p l e t e n e s s
We will start the investigation of splittings in the lattices of modal logics
by studying the lattice of extensions of the minimal modal logic. We will see
that for a logic A in order to split the lattice ~ K~, A must be the logic of a
finite cycle-free frame. There are several questions which come to mind. First,
what happens if we try to iterate the construction? Are there possibilities to
split the lattices by algebras which did not previously induce a splitting?
Secondly, what interesting properties do the resulting splitting logics have?
We will give quite complete answers to these questions. First of all, notice
that by the unravelling technique we know that if a formula ~ is consistent
with K~ then it has a model based on a cycle-free frame, in fact a totally
intransitive tree. This shows that only cycle-free frames can split ~ K~.
PROPOSITION 7.7.1 (Blok). (~ < No.) A logic splits ~ K~ iff it is the logic
of a finite rooted cycle-free frame.
PROOF. K~ - A ( T h f " f finite cycle-free}. So, an element 0 is prime in
K~ only if 0 _D Th f, for some finite f, that is, 0 - Th0 for some O which
is a p-morphic image of some generated subframe of f. If f is cycle-free, so is
7.7. Splittings of the Lattices of Modal Logics and Completeness 369
b. On the other hand, by Corollary 7.3.13 finite rooted cycle-frames induce

a splitting of E K~. []
We can derive the following fact about the structure of the lattice of extensions
from [23].
THEOREM 7.7.2 (Blok). ~ K~ has no elements of finite nonzero dimen-

sion. Hence ~ K~ is atomless.
PROOF. Suppose otherwise. Then there is an a t o m A in the lattice.
Atoms are LJ-irreducible and so U - p r i m e , by the fact t h a t lattices of ex-
tensions are upper continuous. Hence they are co-splitting, and so there
must be a splitting companion of A. Call it | We know from the previous
theorem that ~ - Th [ for some finite, rooted and cycle-free frame. It is easy
to see that there is a splitting logic O' C O. Namely, if f - (f, <~/with w0 the
root, let u ~ f and put <' " - < U {(u, w0/}. Finally, ~ " - (f U {u}, <f) and
| " - Th 0. ~ is rooted and cycle-free and so O' splits ~ K~. It has a splitting
partner A' for which A' C A. But then A' - K~, which is impossible.
We will show now another theorem by BLOK which says that any iterated split-
ting of K~ has the finite model property. However, our m e t h o d is markedly
different from that of BLOK, keeping with the spirit of the technique of con-
structive reduction. Moreover, the original proofs only deal with monomodal
logics and are redone here for polymodal logics.
THEOREM 7.7.3. (~ < ~0.) Adding an axiom of the form [k_L --+ ~ pre-
serves decidability, global completeness, and the global finite model property.
PROOF. The proof is based on the observation that for any m , n C w
there is a finite set S(m, n) of substitutions form the set of formulae over Pi,
i < m, into the set of fromulae over pi, i < n, such that for any finite set of
generators { a 0 , . . . , an-l} for the set algebra F of a refined frame ~ - (f,F)
for the valuation/3 "Pi ~ ai, i < n, any formula ~ in the variables Pi, i < m,
and any point x C f
For then it holds that for all X, r based on the sentence letters p 0 , . . . ,pn-1
orES(re,n)
From this we can deduce that if A is globally complete, so is now the logic A @
[]k_l_ --+ ~ and if A has the global finite model property, so does also the logic
A @ [k2_ --+ ~. For a proof of this consequence from (t) just check all models
on rooted refined frames ~ where the underlying set algebra is generated by
the values of ~ ( P 0 ) , . . . , fl(Pn-1). It is enough to show the theorem in the
class of refined frames.
Now for the proof of (~). From right to left holds for any set S(m, n).
So the difficult part is from left to right. We begin by constructing the
S(m, n). Consider the subframe ~ based on the set C of all points x such that
(~, x) ~ []ks ~ is a generated subframe of ~ and hence refined since ~ is.
By Theorem 2.7.14, C is finite, bounded in size by a function depending only
on n (and k). Hence ~ is full. Consider now the induced valuation on ~, also
denoted by ~. It is possible to show that any set T C C can be presented as
the extension of TT(aO,... , a n - l ) under ~ for a suitable 7T which is of modal
degree < 2k. Collect in S(m,n) all substitutions ~ ' P i ~-+ Ti(po,... ,Pn-1),
i < m, for formulas of depth ~ 2k. S(m,n) is finite. We show (=~) of (t)
with these sets. To that end, assume that ~ ~ ~ k • _+ T for some T such that
var(T) - {Pi" i < m}. Then there exist 7 and x such that
(~, ~, x) ~ ~ k • A ~
Then x E C and so we have by the fact that ~ is a generated subframe
There exist wi(~, i < m, such that ~(Pi) - ~(7-i(~). It follows that 7(~) -
m
~(~[~-~(~/Pi" i < m]). Put ~ "Pi ~ T~(~, i < m. Then ~(~) - ~(~).
Therefore
Z,
And so
This demonstrates (t). [::]
Since K~ has the global finite model property we conclude that split-
ting finitely many frames does not disturb the global finite model property.
However, we can conclude the following.
THEOREM 7.7.4 (Blok). (a < ~0.) All splittings ~ K~/N where N is a
set of finite rooted cycle-free frames have the local finite model property.
PROOF. The proof will be performed for the case •- I. The generaliza-
tion to arbitrary finite n is straightforward but somewhat tedious. Let N be
a set of finite rooted cycle-free frames. Suppose that N is finite. By Theo-
rem 7.3.13 the splitting formula is of the form []-<m+15(PA) -+ _~[]m • for some
m and PA. By repated use of Theorem 7.7.3, one can show that ~ K~/N has
the global finite model property. Now let N be infinite. Put A := ~ K~/N.
Suppose that ~ is a formula and d- dp(T). By induction on d we prove that
every A-consistent formula of degree _< d has a A-model based on a frame
of depth _< d. This certainly holds for d - 0. Let N(d) be the set of frames
of depth _< d contained in N whose powerset algebra is generable by at most
~var(T) elements. This set is finite by Theorem 2.7.14. It is enough to show
that if ~ is A-consistent then it has a model based on a frame of depth <_ d
not having a generated subframe that is contractible to a member of N(d).

Suppose that ~ is consistent with A. Then it is consistent with K~ as well.
So there exists a finite Kripke-frame [ such that [ ~ ~p. Two cases have to
be considered.
C a s e 1. W ~ ~]d_L_ (where F- - ~-g~). Then any model for W is based on
a cycle-free frame of depth _ d. Now ~ is consistent with A and so it is
consistent with K,~/N(d). Thus it has a finite model based on a frame 9 by
the fact that N(d) is finite. 9 must be a frame which is not reducible to any
member in N(d), hence ~ ~ N, since g is of depth < d.
C a s e 2. ~ / ~d~_. Then we can build a finite model of p A -7 [~d _j_ with root
w0. Let qo be in normal form and ~ - ~/i<~ ~ " Let ~i - #iA[S:]xiAAj<m, (}r
Pick i < n such that ~ is A-consistent. Let (iij,/3j, yj} ~ r AXi, tij -- (gj, <~j}.
Assume that the gj are disjoint, of depth _< d - 1 and rooted at yj; moreover,
let w0 r gi for all i. By inductive hypothesis we may assume that no ftj has
a generated subframe which can be contracted to a member of N. Now put
h := {w0} U Uj<mi gJ, < "- Uj<m, <lj u {(wo,yj) . j < mi}, ~ := (h, <1}. By
choice of i, #i is consistent in boolean logic. Let 5 be a valuation into 2 such
that 5(#i) - 1. P u t
{ Uj<m, Zj(p) u {w0} if 5(p) - 1

7(P) " - U j < m i I~J(P) if 5(p) - 0
Then (O, 7, w0} ~ qoi as can be verified immediately. Suppose that 0 has a
generated subframe t~ which can be contracted to a member of N. Then, by
construction of O, t~- [9. So, O is cycle-free. Suppose therefore that for every
i < n the construction must yield a cycle-free frame. Then qDF- ~m_l_ for some
m. It is not hard to show that m < d. This contradicts our assumptions,
however. Hence, there is an i < n and ftj, j < ni such that the frame O
constructed from the 9j has a cycle. (So, one of the 9j contains a cycle.)
Then O is not contractible to a member of N. E]
Here now is another result about lattices of extensions, which is rather peculiar
in nature.
THEOREM 7.7.5. The lattice ~ K.t is isomorphic to a direct product 2• ~,

where 2. has no splittings.
PROOF. First of all, E K . t has a splitting (A, O}, that is induced by the
one-point irreflexive frame, by the fact that it splits E K2, of which K . t is
a member. But notice that O has finite model property by Theorem 3.6.1,
because it is axiomatized by the constant formula <~T V ~-[-. The frames
for 0 are all frames which are different from the one point irreflexive frame.
Suppose now that O ~ C O. Then there must be a finite frame for 0 ~ which
is not a frame for O. The only choice is the one-point irreflexive frame; and
then O ~ - K.t. Thus 0 is an atom, and A is a coatom. Now Theorem 7.2.8
yields the decomposition into 2 • ~. Any splitting of ~ induces a splitting of

2 • ~, so we are done if we can show that no other splittings exist. This is
left as an exercise. [i]
We have seen in the preceding chapter that there are quite strong incom-
plete logics, so completeness is actually not a guaranteed property of modal
logics, as has been believed until the early seventies, before counterexamples
have been produced by S. K. THOMASON [207] and KIT FINE [62], later
also by JOHAN VAN BENTHEM [9]. Nevertheless, despite the fact t h a t there
are incomplete logics one might still believe that the phenomenon of incom-
pleteness is somewhat marginal. To get an insight into the whereabouts of
incomplete logics KIT FINE has proposed in [61] to study the degree of in-
completeness of a logic. This is defined to be the set of all logics sharing the
same Kripke-frames with a given logic. Of course, only one of these logics can
be complete, so the cardinality of this set gives an indication of how many
incomplete logics close to a given logic exist.
DEFINITION 7.7.6. The Fine-spectrum of a logic A is the set sp(A) :=

{O 9 Krp(O) - Krp(A)}. The cardinality of the Fine-spectrum of A is called
the degree o f i n c o m p l e t e n e s s o f A. If A has degree of incompleteness 1
then it is called i n t r i n s i c a l l y c o m p l e t e or s t r i c t l y complete.
We can give a lattice theoretic analogue of this definition. Suppose we
are given a set P c L in a lattice. Then the P - s p e c t r u m of an element is
the set
spp(x) "- {y 9 (VR C P)(x < p ~ y < p)}.
W i t h P being the set of complete logics (or the logics of rooted frames), we
get our definition of the Fine-spectrum. It is clear that each P - s p e c t r u m has
a maximal element, namely ~ (p" x < p C P). In general, there is however no
least element. Fine-spectra do not always contain a minimal element. Here
is an example. (See also Section 7.9.)
We are interested in the tense theory of w ~ P u t 0 := G . 3 . t . K 4 . 3 . D - .
It is not hard to show that the Kripke-frames for 0 are of the form a ~ where
a is a limit ordinal. For the axioms of G.3 guarantee <~ to be a conversely
well-founded relation. Adding K 4 . 3 - forces the relation to be a well-order.
Finally, adding D - a is forced to be a limit ordinal. The theory of w ~
contains 0. W h a t we will show is that this theory is a maximal consistent
and complete logic in the lattice of tense logics, and that it is the intersection
of countably many logics extending it. Since the extensions have no frame,
they are incomplete, and are all in the Fine-spectrum of the inconsistent tense
logic. Hence, in tense logic the Fine-spectrum of the inconsistent logic is not
an interval.
Let us define for subsets A, B of w, A ~ B iff for almost all n C w, n C A
i f f n C B. We say that in this case A and B are a l m o s t e q u a l . It is not
hard to show that ~ is a congruence in the algebra of sets over w, with the
operations
A "- {n'(Vm < n)(m e A ) }
I A "- {n'(Vm>n)(meA)}
Furthermore, let ([} be the set of finite and cofinite subsets of w. ([} is closed
under all operations, and therefore (w, >, O) is a general frame. @ contains
exactly the sets which are almost zero or almost one. For the purpose of the
next theorem, a p a r t i t i o n of a set M is a subset X of p ( M ) such t h a t (a)
Or (b) M - ( . J X , a n d ( c ) f o r a n y S , T e Z , i f S ~ = T t h e n S N T = ~ .
LEMMA 7.7.7. Let A c_ p(w) be a finite partition of w. Then the least

algebra containing A is the set of all sets which are almost identical to a
union of elements of A.
LEMMA 7.7.8. Let G "- {Si " i < n} be a finite set of subsets of w. Then
there exists H - {Tj " j < k}, which is a partition of g of cardinality < 2 n
such that G and H generate the same algebra of sets.
LEMMA 7.7.9. Let Q be an n-generated set of sets. Then (w, >, Q) satis-
fies x(2n), where x(k) is the following formula.
Vi<k f > ~ p ~
v V~<j<k ~ ( p ~ ++ p;)
v 4>F1V~<kP~
PROOF. Let ~ be a valuation and let ai " - ~ ( P i ) . Pick some point n C w.

We have to show t h a t ( ( w , > , Q ) , k , ~ ) ~ x(2n). To t h a t end, assume that
k ~ V i<k ~>E~pi. Then all ai are infinite. Assume furthermore that k t~
V~<j<k e>E(p~ ++ pj). Then for no different i , j we have a~ ~ aj. Hence no
two sets ai and aj are almost equal. We know by L e m m a 7.7.8 that in an
n-generated algebra there are at most 2 n sets such that no two are almost
equal. Moreover, their union is almost w. Hence k ~ f>E] Vi<kPi. T h a t had
to be shown. K]
LEMMA 7.7.10. No logic containing 6)@ x(k) has a nonempty Kripke-

flFame.
PROOF. Let [ be a (nonempty) Kripke-frame for O @ x(k). Then [ i s
isomorphic to a converse well-order a ~ where a is a limit ordinal, and so
not finite. But then x(k) is clearly not valid in ~. K]
THEOREM 7.7.11. The tense logic of wop is the downward limit of logics
which have no frame.
PROOF. Suppose that wop t~ -,~. Then there exists a model {w~ ~, k)
~. Now let Q be the algebra generated by/3(p), p E var(T). Clearly, we have
(w, >, Q) ~ --~. Q is finitely generated. Hence (w, :>, Q) satisfies an axiom
x ( k ) for some k. Hence ~ is consistent in Th f~ @ x(k). So,
Th~~ - N Th~;~ @ x(k)

kC~
This concludes the proof of the theorem. [2]
COROLLARY 7.7.12. The spectrum of the inconsistent tense logic is not

an interval.
THEOREM 7.7.13. The spectrum of the monomodal logic Th[o-~ is not an

interval.
This last theorem follows immediately from the Simulation Theorem. The
situation is different if we specialize on the set P of prime or splitting elements.
The P - s p e c t r u m of x is called the p r i m e - s p e c t r u m and denoted by psp(x).
PROPOSITION 7.7.14. Let ~ be a lattice. For every x E L there are xo

and x ~ such that psp(x) - [x0,x~ x ~ is P-complete and xo is a union of
co-splitting elements. Consequently, if x - xo and x is P-complete then x is
strictly P-complete.
PROOF. P u t x ~ := [ I ( P ' P C P,p >__x) and x0 := L.](2"/p'p e P,p ~ x).

Suppose that z c psp(x). Then clearly z ~ x ~ We have to show t h a t xo _< z.
Now for p C P we have p ~ x r p ~ z whence ,?,/p < x r ,?,/p < z.
Therefore z > L.](2,/q'p c P,p ~ x) - xo. Conversely, if x0 _< z < x ~ then
p > z r p _> x for all p C P and so z C psp(x). [-1
This shows among other that the splittings of K~ by cycle-free frames

are intrinsically complete. Thus incomplete logics cannot be 'close' to unions
of co-splitting logics. The difficult question remains as to where these logics
are. We will spend the rest of this section and the whole next section to prove
one of the most beautiful theorems in modal logic. We will call it Blok's
Alternative, because it says that logics have only two choices for their degree
of incompleteness" 1 or 2 ~~
THEOREM 7.7.15 (Blok's Alternative). A logic containing K1 is intrinsi-

cally complete if] it is an iterated splitting of K1. Otherwise, the logic has
degree of incompleteness 2 ~~
The proof is done in two steps. The first step consists in showing that
the coatoms have degree of incompleteness 2 ~~ In the second step we use the
frames produced in the first step to show that in fact any non-splitting logic
has degree of incompleteness 2 ~~ Let us begin by defining the frames ~ M ,
where M C_ w - {0}. The set of worlds of ~ M is the (disjoint) union of the
FIGURE 7.10. ~M, M - {1, 4 , . . . }
4o 1~
0
co+l
A ,.A
r w
co 6" 5" 4" 3" 2" I" T

T
sets { n ' ' n C w}, { , , c o , co + 1} and { n ~ C M}. We put
xE{m~176176176
or x =y=m ~
x <] y r or x=0",yC{,,co+l},
or xC{co, co+l},y=co,
or x = co, y C { n ' , n ~
This defines the frame PM. The frame corresponding to M = {1, 4 , . . . } is

shown in the picture. PM is intuitively obtained as follows. We have the set
of natural numbers ordered by <l = >. Each point in t h a t set is referred to
as n ~ Moreover, if n C M we also have a reflexive companion n ~ Notice
that the set {,} is the only nontrivial generated subframe. We let ~M be the
algebra of all finite sets not containing co and of all cofinite sets containing
co. This set is closed under all operations. Indeed, if we have a finite set
S not containing co, 9 or co + 1, or a cofinite set then 0 S is cofinite, and
contains co. Furthermore, we have r = r + 1} = {0 ~ which is finite
and does not contain co. If S is cofinite and contains co, so does OS. Thus
this is well-defined and so q3M := (PM,PM} is a frame.
LEMMA 7.7.16. ~M is O-generated.

PROOF. We will show two things. (1.) {,} is definable, and (2.) all
other sets are definable from {,}. (1.) is easy. We have I1~ - {,}. For
(2.) we need to do some work. First, I1{,} - {0"}. Moreover, {co + 1} -
0<2{0 ~ -0<1{0~ Now define by induction on n the polynomials in(p),
p. (p), pO(p).
in(p) "--- Vk<n(P~(P) V p~(p))
p](p) .= p
p$(p) .= •
Pn+I(P) "-- I--]in(p)

o
Pn+l (P) - O Pn(P)
" /~ Pn+l (p) A Pn+l (p)
T h e n io(p) - p. By induction it is verified t h a t p~({0"}) - { n ' } and t h a t
p~({0"}) - {n ~ if n E M, a n d - O otherwise. Hence in({0") - {k" 9
k < n } U { k ~ 9 k ~ n , k E M } . Finally, { 0 " } - toO, so we can define all
these sets if we replace p by D_l_. So all singleton sets except for {cx)} are
0-definable, and all finite sets not containing c~ and their complements are
also 0-definable. [::]
Notice t h a t the definition of the polynomials does not depend on the set
M C_ w. We will make use of t h a t in the next section. The next goal is to
show t h a t the logics of these frames are all of codimension 2. To see this
we can use the splitting theorem. The algebra is subdirectly irreducible and
the logic it generates is 4-transitive. Hence we are done if we can show this
algebra to be finitely presentable. But this is entirely obvious, since we have
just shown t h a t it is 0-generated, and so isomorphic to ~th (0). Hence, we can
take as a diagram simply -1-. To get an extension of this logic we simply add
the formula D_l_, for [31 is an opremum! Therefore, Th ~[~M has codimension
2. Finally, for different sets M, the logics of the frames are distinct, simply
by the fact t h a t the points n ~ are definable by constant formulae. We have
proved now not only that there are uncountably many incomplete logics with
K r i p k e - f r a m e [-~ but also t h a t the logic of the irreflexive point is co-covered
by them.
DEFINITION 7.7.17. Let s be a lattice, and x E L. Define the c o - c o v e r i n g

n u m b e r of x to be the cardinality of the set {y " y -~ x}.
THEOREM 7.7.18 (Blok). The logic of the irreflexive point is co-covered

by 2 ~~ incomplete logics. Hence it has co-covering number 2 ~~ and degree of
incompleteness 2 ~~ .
Next we need to deal with the frame [ ~ . The solution will be quite
similar. Namely, instead of the frame OM we define the frame qM consisting
of the same set, and the relation <] differs minimally in t h a t we now put 9 <~ ,.
The rest is the same. We put QM "-- ~M and this defines ~ M "
LEMMA 7.7.19. ~M is 1-generated. Moreover, a set generates L~M iff it

is Fs 0 and Fs qM.
FIGURE 7.11. i]M, M - {1, 4 , . . . }
4~ 1~
0
oc+l
A
o~ 6" 5" 4" 3" 2" I" *
PROOF. Since the frame is identical on - { , } , we can use (2.) of the

previous proof to show that QM is generated from {,} and so 1-generated.
For the second claim it suffices to show that {,} is definable from any other
nontrivial set. We claim that - { , } = r 11<4b iff b r O and - b r O.
Namely, suppose that b ~: O as well as - b r O. Then, because b r O,
r D - { , } . For if x C b for some x then by 4-transitivity all elements that
can at all reach x are in r Hence either 9 C b and then r - 1 or 9 ~ b
and then r - { , } . C a s e 1. 9 C b. Then there exists an x ~: 9 such that
x r b. Consequently, 11<4b = {,}. C a s e 2. 9 ~ b. Then r = - { , } , and
m-<4b- O, as required. E]
The algebra underlying i~U has two subalgebras, the two element algebra
and itself. It has three homomorphic images, itself, the trivial algebra and
the algebra of [-6-], the latter corresponding to the subframe generated by {,}.
Now take a logic 0 _D AM Th ~ M . We have h 9 ~tAM(1) ---- ~rO(1). We
" - -
will use an argument similar to the one before. However, this time we do
not have such a simple structure for the free frame. Namely, it consists of
countably many copies of L~M, each corresponding to a different generating
set. The generating sets are exactly the nontrivial sets. We now reason as
with the veiled recession frame of Section 7.6.
LEMMA 7.7.20. Let S be an internal set of i.~M and C " - S N r S.

C- ~ iff S - 0 , S - {,} or S - qM.
PROOF. Suppose that S - o. Then C - O. Suppose on the other hand

that - S - 0. Then C c_ r 0. Finally, if S - {,}, t h e n , r r
and so C - O as well. Now assume that neither of the three is the case. Let
S be cofinite. Then it contains oo. - S is finite, and r S contains oo if
- S contains at least n ~ or n ~ for some n. In that case, oc C C. Now let

-S C {co+l,,}. In this case 0" C C. This finishes the case where S is
cofinite. So, let now S be finite. It does not contain oc, so c~ § 1 C 0 - S.
Hence if co § 1 C S we are done. So, assume from now on co + 1 ~ S. Let n
be the smallest number such t h a t S n {n ~ n ~} r 0. n - - - 1 if such n u m b e r
does not exist. I f n > 0, 0" C - S and so { n ' , n ~ C_ 0 - S . In t h a t case,
Cr Now l e t n - 0 . Then, s i n c e o c § 0" c 0 - S , andso0 ~ cC.
Finally, let n - - 1 . Then S - {,}. But t h a t was excluded. [2]
So, h(p -4 Vlp A -~p --+ I]--p) = 1 iff h(p -+ lip) = 1 and h(-~p --+ []-,p) = 1
h(p) = o or = h(p) do s . o t full of i.t rn 1
sets iff h(p) generates a subalgebra isomorphic to the algebra of subsets of F ] .
Hence if p --+ []p r O, ~ r e ( 1 ) contains a generated subframe isomorphic to
QM, so t h a t 0 = AM. The following is now proved.
THEOREM 7.7.21 (Blok). The logic of the reflexive point is co-covered by

2 s~ incomplete logics. Hence it has co-covering number 2 s~ and degree of
incompleteness 2 s~ .
E x e r c i s e 257. Show t h a t there are 2 s~ iterated splittings of K1.
E x e r c i s e 258. Show t h a t the frame underlying the freely 0 - g e n e r a t e d K 1 -

algebra has cardinality 2 ~~ .
E x e r c i s e 259. Show that Th[-~ is co-covered by 2 s~ logics in the lattice

E K . t r s 4 . Hint. This should be entirely easy.
E x e r c i s e 260. Describe the structure of the canonical frame for one variable
in ThL~M.
* E x e r c i s e 261. Show that the lattice of tense logics has exactly one splitting.
Hint. Show first t h a t K . t is complete with respect to finite frames which
contain no forward cycle, and hence no backward cycle. Now let [ be such a
frame and let it have more t h a n one point. Create a sequence of frames In
such t h a t (i) they contain a reflexive point and so Th ~n ~ Th~ and (ii) t h a t
n T h fn c_ Th f. (See [123].)
E x e r c i s e 262. Let 92 be a finite n - g e n e r a t e d modal a-algebra. T h e n Th 91

can be axiomatized by formulae containing at most n + 1 variables.
7.8. B l o k ' s A l t e r n a t i v e
Now t h a t we have shown for the two coatoms t h a t they are m a x i m a l l y

incomplete, we proceed to a proof t h a t in fact all consistent logics which are
7.8. Blok's Alternative 379
FIGURE 7.12
2 3o 21
1
J o0
2~ *x ~ '
0 1~ 01
not joins of co-splittings of ~ K I have degree of incompleteness 2 ~~ Suppose

that O is a consistent logic and not identical to a logic K 1 / N , N a set of cycle-
free frames. Then O properly includes O0, the splitting of K i by all cycle-free
frames not belonging to Krp O. Since O0 has the finite model property, there
must be a finite frame f such that f ~ O0 but f ~ O. By the construction of O0,
f cannot be cycle-free. It is by playing with f that we obtain a set of frames
~M, where M C_ w and ~M is not a frame for O, but all Kripke-frames for
Th ~M are frames for @), so that O A Th (~M is an incomplete logic different
from O.
Consider now the frame f. The desired frames (~M will be produced in
two stages. First, we know that f contains a cycle, say co <~ c1 <1... <1 c~ - co.
P u t C " - {ci 9 i < V}. We assume the cycle to be minimal, t h a t is, ci <:1cj iff
j - i + 1 (mod V)- The proof of the existence of such a cycle is an exercise.
We produce a variant of [, fn, by blowing up this cycle. This variant is defined
as follows. We p u t / n .__ ( / _ C) [-J C x n . (Here, the union is assumed to be
disjoint.) We write x i rather than <x, i} for x C C and i < n. Also we write
C n " - {c} 9 i < n, j < 7}. So, fn consists in replacing C by n copies, so that
a point x C C is now split into x ~ , x n-1. 9 is defined as follows.
(1) z,y r C n andx<~y

or (2) x- c ij , y - c ik a n d k - j + l (mod7)
x9 or (3) x-c~,y-c~ +1,
or (4) x- c ij , y r C n , c j <~ y,
or (5) x r C n, y - 0 and x <~ cj
cj,
Figure 7.12 shows a duplication of a minimal four cycle according to the

previous construction.
PROPOSITION 7.8.1. Define 7r by 7r(c}) " - cj and 7r(y) := y f o r y ~ C n .

T h e n 7r " [n ~ [.
PROOF. This is a straightforward checking of cases. The first p - m o r p h i s m

condition is proved as follows. Assume x 9 y. Then if (1) holds, also x <~ y;
moreover, x - ~(x) and y - ~(y). If (2) holds, then x - cj and y - c k
withk--j+l ( m o d 7 ) . T h e n ~ ( x ) = c ~ a n d s ( y ) = ck. By the fact t h a t
C i s a c y c l e , cj <~Ck. Similarly for (3). Now let x - cji 9 y, y r C n . Then
~(x) = cj <~ y = ~(y). Similarly for (5). This concludes the proof of the first
condition. For the second condition assume ~(x) <~u. C a s e 1. x r C n. T h e n
~r(x) = x. Now either ( l a ) u r C; in this case x 9 u and ~(u) = u. Or ( l b )
u e C; in this case ~(c ~ - u for some c~ such that x <~c~. Then by (5) x <~c~. 0
C a s e 2. x ~ C ~ , x - c j , for some i j. Then either (2a) u r C in which case
we have u = ~(u) and x 9 u b y (4); or (2b) u e C. In this case we must
have u = ck for k -- j + 1 (mod 7), by the fact that the cycle was chosen to
be minimal. Hence if we put y - c k t h e n x 9 , and~(y)-u . []
We also note the following. If L : f ~-+ g, then ,[C] is a cycle of g, and

gn is defined in an analogous way. Then there is a p - m o r p h i s m Ln :[n H gn
defined by , n ( x ) " - L ( x ) i f x r C and J~(c~):= L(cj) i for cj E C. We call Ln
the l i f t i n g of L. Likewise, a p - m o r p h i s m ~ 9 [ --- g which is injective on C
defines a lifting ~n : in __, gn, with gn analogously defined.
Take two pointed frames (~,x) and (6, y). Assume that the underlying
sets of worlds are disjoint. Then the connected sum, denoted by ~ Y~ ~ - - or,
if the context allows this, by ~ _v ~ is defined as follows.
<h "= <S U <g U {(x, y)}
H "= {aUb'aCF, bCG}
._ (/ug,<h,n)
It is easy to see that whenever {x} is internal in ~, ~ _v~ r is a frame. For

Ch(a v b) - Csa U if y r b and Ch(a U b) - Csa U U else.
In what is to follow we will fix ~5 to be ~ M , M C_ w - {0}, and ~ will be
the full frame corresponding to a frame of the form in for a finite frame with
a cycle C. Given this choice, the construction always yields a frame. For this
cycle we assume to have fixed an enumeration C " - {ci " i < n} and so defines
a frame of the form in. Now define ~___~ M to be ~v_~+ 1 ~ M , where x -- c~ -1.
Notice t h a t the construction depends on several parameters, the number n,
the distinguished cycle, the order of the cycle, and the set M. Let us agree
to call a point of ~ __ ~M b a d if it is in g, and g o o d otherwise. There are
only finitely many good points, since ~ is finite. The following theorem holds
quite generally for arbitrary ~ in place of ~[~M (if glued at the same point x).
LEMMA 7.8.2. Suppose that ~ ~ ~ and that d > d p ( p ) . T h e n fdv_~JM ~ ~.
PROOF. Suppose that ([,~,x) ~ ~ . Fix the p - m o r p h i s m 7r 9 [d _~
of L e m m a 7.8.1. If x E C choose y := x, else y := x ~ Take the valuation
7(P) := ~-*[~(P)]. Then by Proposition 7.8.1, since d > d p ( ~ ) we have
(fd, ~, y) ~ ~ . Now let 5 be any valuation on [d_v~ such that 5 ( p ) n f d = ~/(p).

We claim that ([d_v 6 , 5, y) ~ - ~ . For that purpose it is enough to show t h a t
the d - 1-transit of y in [d _Vr contains no bad points. For then it is identical
to the d - 1-transit of x in [d. Since 9 R(f d x fd) = <ld, the two transits are
isomorphic as frames and the conclusion follows. So, suppose there is a chain
(yi : i < d) such that Y0 = Y and yi <lyi+l, i < d - 1. If it contains a bad point,
by construction of fd _V~M there is a io < d - 1 such that Yio = cd-1. Then
there must be a il < io such that Yil - Co d-2, a i2 < il such t h a t Yi~ - cd-3
etc. It follows finally that there is a id-1 such that Yid_l - co. But this means
that id-1 < O, contradiction. F-]
Let us get some more insight into the structure of [d V ~M- First, ~M
is a generated subframe of [d V ~M. Next, look at the polynomials pO, p~
and in. Since ~M is a generated subframe of [d V ~ M , p ~ ( 0 T A []21) is
satisfiable in fd V ~ M , and if n E M, p~ A F~2_L) is also satisfiable in that
frame. However, if n ~ M, then p~ is not satisfiable at n ~ Moreover,
p~(0TAF-]2_L) is satisfiable only at n" and p ~ only at n ~ i f n > ~fd.
From that we deduce
PROPOSITION 7.8.3. Let d C w and M, N be subsets of w. Suppose that

there exists a n e ( M -N) U ( M - N) such that ~fd < n. T h [ d V ~ M
Th fd _V~Y.
PROOF. Suppose M r N. Then there exists a n such that n C N - M
or n C M - N. Assume the first. Then pO ((}T A [:]21) is not satisfiable in
[d V ~M since n > ~fd. Hence ~p~ A El2/) E Th [d V ~M- However,
p~ A El2/) is satisfiable in f d v ~ g . Now assume n E M - N. Then
R~+1(0TAE]22) -~ (}p~ theorem o f [ d V ~ M but not of[dV~3g. E]
LEMMA 7.8.4. Let (f,y) be a finite pointed frame. Assume that H is a

subalgebra of the algebra of internal sets of Ivy
- oo-F1 ~ M Then either H is finite
or there is a p-morphism 7r 9 [vy--c~-F1 ~ M ~ J~ for some atomic frame ~, such
v~'(y) ~M
that ]HI is the 7r-preimage of K. Moreover, if {,} C H, then ~ -~ g--c~+l
for some contraction ~v : [ --~ g.
PROOF. Assume that ]E is infinite. Then it contains a nontrivial subset of

g}M. By the results of the previous section, the trace algebra induced by H on
~M is the entire algebra ]PM. Now put x ~ y iff for all a C ]E, x C a iff y C a.
I f x i s bad and ~=,, t h e n x ~ y o n l y i f y = x. Hence, [x]~ := { y : x ~ y }
is finite. [ v ~M is atomic; therefore there is a set as for every x such that
Ix]~ = as. Now x <] y iff x E 0ay and from this follows easily that ~ is a net,
and we have an induced p-morphism. If {,} e H, then [,]~ = {,}. It is easy
to see that the induced map comes from a p - m o r p h i s m 7r: [d __~ ~d. F-]
LEMMA 7.8.5. Let ([,y} be a finite pointed frame and r generated sub-
/rame of f --c~+
VY l ~ M Then
(a) ",~ VY ~ M , where g < [ and y C g, or

_
(b) r -~ ~ M , or
(c) |
(d) O is a generated subframe o f f and y r h.
W i t h ([, y} given, let [-YAc~-t-1~ M or simply [ Z ~ M denote the frame which

results from Ivy --oo+1 ~ M by factoring out the net ~ for which [.] ~ - {. ' z} for
some z C f and [x]~ - {x} for all x ~ { . , z } .
LEMMA 7.8.6. Let ([,y) be a finite frame. Let A be an extension of

Th f vY
-c~+ l ~ M Then A is the intersection of logics of the form
1. Th g v_ ~ M or Th g X ~ M , where g is a contractum of some generated

subframe of [ containing y,
2. TheM,
3. T h ~ and
4. Th g, where g is a contractum of a subframe of [ not containing y.
PROOF. The proof is by induction on ~f. We assume that the theorem

holds for all frames g such that ~g < ~f Let ~ -- [VY 9
~ M or ~- - c o + l
--
Aoo+l--Y~ M . There exists a formula ~ such that ( ~ , ~ , x ) ~ ~ iff the transit

of x is ~, and the map /3 is onto. Before showing the existence of ~, let
us see how it proves the theorem. Consider an extension A of Th ~. It is
proper iff it contains -~. Now the algebra ~l:h(n) is a subdirect product of all
algebras generated by valuations into ~ satisfying -~. By L e m m a 7.8.4 and
L e m m a 7.8.5 these algebras correspond to frames of the form (a) with g being
of smaller cardinality than f, (b) or (c). By induction hypothesis the desired
conclusion follows.
Now for the existence of ~. We perform the argument for ~ - f _v ~ M .
First, ~ is d-transitive with d "- ~f + 4. For each x C f let p~ be a variable.
Add p. as a variable. Let P0 be the variable of the root of [. Define now
r .__ ~3__.
Pn(V~P*)A K]D_s for n > ~f.
LEMMA 7.8.7. Let (~, ~, x) ~ r for some ~. Then x - 0".
The proof of this lemma is an exercise. Now put
p(n~ 9= Pn(r
p(n~ 9= s162
p(oo +1) := 00p(0")A 0p(0").

By L e m m a 7.8.7, (~, 13, y) ~ p(a), a e PM, iff x -- a.

p. -+ []2_
A p~ -+ <>p(oo + 1)
AV (p~:x ~ f u {,}} v O<~p(O")
AA (p~ --+ -~p~, : x -r ~'; ~, x' c / u {,})
AA (p~ -+ Opt, :x < x'; x, ~' c f u {,}}
AA (p~ -+-~Op~, : 9 r x'; ~, ~' c / u {,})
"-- P0 A [--1-(d5
Suppose that (~,/3, x} ~ ~. Then x ~ P0. It can be shown chasing successors
that x b o<_dpy, and so x ~ o<dOp(c~+ 1). Since p(cx~+ 1) can only be satisfied
at oc + 1, we have first of all that the transit of x contains ~ u . Moreover,
f~(p(oc + 1)) -- {c~ + 1}, and so the trace algebra of im[~] relative to ~ U is
PM. Using this one can show that the transit of x must contain exactly ~f
many points outside of ~ M . Finally, it is shown that f~(p~) - {o(x)} for some
function o" f --~ f, and that o is an isomorphism. It follows t h a t x generates
the whole frame, and that the algebra induced b y / 3 is the entire algebra of
internal sets. This concludes the proof. [3
LEMMA 7.8.8. Let ([,y) be a finite frame. Then Th ~ V ~ M has finite
codimension in s K.
The previous results also hold if ~ M is replaced by s However, some
adaptations have to be made. For example, r is now defined by
r . - 03p:(C]p,)A D p , .
Furthermore, in 5 we need to put p, --+ [3p, in place of p, -+ [32_. The ana-
logue of L e m m a 7.8.7 does not hold with this definition. The entire argument
goes through nevertheless, with some slight complication. We will not spell
out the full details. This is what we need to prove the next theorem.
THEOREM 7.8.9 (Blok). Let | be consistent and not equal to a splitting
K 1 / N by a set of cycle-free frames. Then 0 is co-covered by 2 s~ incomplete
logics.
PROOF. If 6) is consistent, then (9 C_ Th[~] or O C Th[T], by Makinson's
Theorem. Suppose that O C_ T h r ~ . Take a minimal [ such t h a t [ is not a
frame for 0 . Then there exists a formula ~ such that f ~ ~ but ~ C O. P u t
d := d p ( ~ ) + 1. Let M C_ w - {0}. Then by L e m m a 7.8.2, [d V ~ M ~ ~P, SO
AM "-- Th [d_V~U 4 ~" Let A ~ be a maximal extension of AM not containing
O. By L e m m a 7.8.6, the fact that O C_ Th[-~ and the choice of [, A ~ is
incomplete. Moreover, it is r]-irreducible, and so A ~ is of the form Th 9_V~M ,
Th ~ A ~ U or Th ~ U . O [-] A~/ co-covers O. For if O D O' D O 71 A ~ then
O I u A ~ 2 O, since @/L] A ~ properly contains A ~ . On the other hand O --

( O ' D A ~ ) D O - ( A ~ n O ) D O ' - O' ~ O. Contradiction. A ~ is incomplete.
and by minimality of f all finite frames of AM are frames for O. It remains to
be seen that there exist 2 ~o many different such logics. To that effect note t h a t
by Proposition 7.8.3 that there is a number c such that if (N - M ) U ( M - N)
contains an n _> c then Th g _v ~ M and Th g _v ~ N are incomparable, and
similarly Th g N ~M and Th g ~ ~3N are incomparable. Hence, for such sets A ~
and A~v are incomparable. Moreover, O A A ~ r O A A~v. Both are co-covers
of O. Hence, any set N such that N N { 0 , . . . , c - 1} -- M N { 0 , . . . , c - 1} would
have sufficed equally well. There are 2 ~~ many such sets. This concludes the
proof of the case O C_ Th I-g]. If O ~ Th [-~, use ]L~M is place of ~M. []
Little remains to be done to complete the proof of Blok's Alternative.
Namely, we need to show that there arise no new splittings except for K . D -
K/[~-], which has a splitting frame [-~, yielding the inconsistent logic as an
iterated splitting of K. Since the inconsistent logic was always included in
the theorems, they hold for that case as well. Hence, the following theorem
concludes the proof.
THEOREM 7.8.10 (Blok). Let A - K 1 / N for a set N of cycle-free frames.

Then either ~-~ C N and Th[-~ splits E A or [ ~ r N and O splits E A iff it
splits ~ K1.
PROOF. Let ~ r N. T h e n ~ i s a f r a m e for A. Assume t h a t [ i s not
cycle-free and minimally so, that is, for every frame g to which [ reduces
either g - [ or g is cycle-free. Then
(Th [d V ~ M " d C w) C_ Th [
But for no d Th fd V ~M ~ Th [. Hence Th [ is not prime in ~A. Yl
E x e r c i s e 263. Show that if a frame contains a cycle, it contains a minimal
cycle.
E x e r c i s e 264. Let ~ be a distributive lattice, x , y C L. Show t h a t the

interval [x R y, y] is isomorphic to the interval [x, x U y]. (In fact, distributivity
is stronger than necessary.)
E x e r c i s e 265. Show L e m m a 7.8.7. Hint. Show first that (~, ~, x) ~ p~ (p)

implies t h a t there is a chain (xi " i < n + 1) such that x - x0, xi <] xj iff i < j,
and that all xi are different.
7.9. T h e L a t t i c e of T e n s e L o g i c s
Tense logic is on the one hand an important area of modal logic when it
comes to its applications. On the other hand it has proved to be influential
also in the theoretical development of modal logic. It was here t h a t m a n y
7.9. The Lattice of Tense Logics 385
counterexamples to general conjectures have been found. Moreover, it has

been shown that the rich theory of extensions of K 4 is not the effect of it
being operator transitive, but rather of the combination of it being operator
transitive and weakly transitive. The first implies the second when there is
one basic operator, but already in tense logic this fails to be the case. The
consequences of this will be seen below. We begin this section by showing
that in the lattice of tense logic there exist incomplete logics of codimension
1. The first example is due to S. K. THOMASON [206]. (See also Section 7.7.)
We present it in the form given in RAUTENBERG [169]. Let
A := G.3.t @ K 4 . 3 . D - @ a]~>p --+ ~>Ep.
A Kripke-frame for G.3.t @ K 4 . 3 - is nothing but a converse well-order )~op.

A Kripke-frame for G.3.t @ K 4 . 3 . D - is a converse well-order Aop where
is a limit-ordinal. It is easy to show, however, that no such frame can be
a frame for []~p --+ ~>a]p. A therefore has no Kripke-frames. But it is not
inconsistent. For example, let 12 be the frame based on w with the finite and
cofinite subsets as internal sets. Then fl is a frame for A. It can be shown
that A = Th ~.
THEOREM 7.9.1 (Thomason). The lattice of tense logics has an incom-

plete logic of codimension 1.
Already, this allows to deduce that the lattice ~ K1 has an incomplete

logic of codimension 2. So, Theorem 2.9.9 is the best possible result. One
can play several variations on that theme. It is possible to produce 2 ~~ many
incomplete logics of any finite codimension in ~ K.t. It that way many results
of Section 7.8 can be shown using tense logic.
The previous section has illustrated that we can deduce important facts
about logics if we concentrate on certain sublattices. These sublattices also
had the best properties we could expect, namely to be continuous sublattices.
The general situation is far from being that good. As an illustration we will
study some basic properties of the lattice of tense logics. It is easy to see that
there are embeddings of the lattice of modal logic into the lattice of tense
logics, and that there are also maps in the opposite direction. However, these
maps usually behave well only with respect to one of the two operations. The
general situation is as follows. We have a pair of modal languages, L1 and
L2, and L1 is a sublanguage of L2. Then a logic 0 over L1 has a natural least
extension @e in the lattice of L2-1ogics. Moreover, each L2-1ogic gives rise
to an Ll-logic by restricting to the language L1. Since we are always dealing
with normal modal logics, this is satisfied.
We will study the extension lattices of K.t, K 4 . t and S4.t. The method
is uniform in all three cases and can be transferred to numerous other logics.
Each of these logics has the finite model property and therefore only finite
386 7. L a t t i c e s o f M o d a l Logics
FIGURE 7.13
algebras can induce splittings. Thus we can concentrate on the K r i p k e - f r a m e s

of those algebras. Here is an i m p o r t a n t fact a b o u t tense algebras.
PROPOSITION 7.9.2. For finite tense algebras 92 the following are equiva-
lent.
1. 9~ is subdirectly irreducible.
2. 92 is directly irreducible.
3. 92+ is connected.
4. 92 is simple.
This can easily be established by using the duality theory. It can also
be obtained using the m e t h o d s of Section 4.3. Notice t h a t this t h e o r e m fails
for infinite algebras; a counterexample will be given below in the exercises.
This is due to the fact t h a t there exist logics which are not weakly transitive.
Notice also t h a t K 4 . t and S4.t while being operator transitive are not weakly
transitive. This has to be borne in mind.
For the definition of subreducing frames we use the following construction.
Take two frames ~, g and let x C f, y C g. T h e n let ~ ly g denote the frame
(f ~iy g, <]) where f ~ l y g - f • {0} U g • {1} - {(y, 1)} and <] is defined by
(i) (a, 0)<~ (b, 0) iff a <If b
(ii) (a, 1} <] (b, 1} iff a <~9 b
(iii) (x, 0} <~ (b, 1} iff y <~ b
This is well-defined whenever x <1~ x ~=~ y <]g y. If [, g are transitive, <l will
be taken to be the transitive closure of the relation defined above. We call
[ ~ ly g a b o o k and [ and g the p a g e s . W h e n the choice of the points is clear
we write [ / g instead of [ ~ ly g. W i t h two points x <] y E f fixed we define '~[,
n>_ 1, by
2k+2[ .= 2k+~[yly
We distinguish the elements of different pages in n[ by indices 0 , . . . , n - 1.
The m a p ~ 9 xi ~-~ x is a p - m o r p h i s m ; for if xi <~ yj then either i - j and
thus x <] y by (i) and (ii) or i + 1 - j and x <] y by (iii). Now if ~(xi) <] y we
have y - ~(y~) and xi <] yi. Likewise for x <] ~(yi). The same can be shown
in the transitive case. By this we see t h a t any m a p r 9 n [ l g "~ [ satisfying
FIGURE 7.14. The garland r
0 2 2n- 2
" " "
1 3 2n- 1
r = x is n - 1-1ocalic with respect to any point x0 of the first page. This

suggests t h a t by taking a suitable ~ so that there is no p - m o r p h i s m from n[lg
to [ for any n E w, {Th nil ~ : n C w} is a subreduction of Th [. A particularly
important class of frames for our purposes and for illustration of the books let
us introduce the garlands. A garland is a zigzag frame shown in Figure 7.14.
Formally, we define g[~ as a frame (n + 1, <]) where i <] j iff i = j or i is
odd and j = i + l . Thus, gn has 3n arrows and n + l points. Note that
a garland is isomorphic to a book where each page is the frame I o --o I .
Garlands can be characterized modally. A frame [ = (g, <~) is called m e a g e r
if there are no two points x <~ y <] x. A connected frame [ is a garland iff it is
reflexive, meager, of alternativity 3 in both directions, and of depth 2. Thus
G a := Qrz.t.alt3+.alt3 as the reader may verify. We first prove an i m p o r t a n t
LEMMA 7.9.3. Suppose that [ is not a cluster. Then T h g splits ~ K . t ,
K 4 . t and E S 4 . t only if ~ is a garland.
PROOF. [ is finite and connected and ~f - n. Since f is not a cluster,
there are points x, y such that x <] y r x. Also [ - 2Yr,(x). Now consider
the frame m[ lo g[2n+s. This is well defined in case x and y are b o t h reflexive
points. In case one of them is not reflexive we take mf 20 g[2n+s instead, where
g[~n+s is identical to g[2n+s except that 0 r 0. We have to show t h a t if there
r(o) -+ f then (a) [ is reflexive
is a p - m o r p h i s m (in both relations) 7r" mf l0 ~'2n+s
and transitive, (b) [ is of alternativity _< 3, (c) [ is meager and (d) [ is of
depth < 2. For then ~ splits the lattice only if it satisfies (a) - (d) and thus
is a garland. The details of the proof are left to the reader. The crucial fact
to be shown is that if x C f there exists a z C gl2~+s such that p(z) - x. For
then 71-[~[2n+S] - [ , from which the rest follows. [i]
This considerably reduces the class of possible splitting frames. However, we
will also show that most of the garlands and clusters cannot split any of these
logics. We do this by establishing a lemma on splittings of E G a .
LEMMA 7.9.4. Th 9In splits E G a iff n <_ 1.

Once this l e m m a is proved it follows t h a t 0In c a n n o t split ~ K . t , ~ K 4 . t

nor ~ S4.t for n > 1 since all these lattices contain ~ G a .
PROPOSITION 7.9.5. T h e r e is a p - m o r p h i s m 7r 9 Olm ----} ~ [ n i f f n -- O,

m -- w or n divides m .
PROOF. (=~) Suppose 7r" g[m --+ 0In is a p - m o r p h i s m . A s s u m e n > 0.

Write i - j for 7r(i) - ~r(j). We now have the following
CLAIM 7.9.6. On the c o n d i t i o n that n > O, i f i - j t h e n i - j ( m o d 2).

Moreover, if i - j then i- l - j-1 or- j + l and i + l - j -1 or- j + l,
w h e n e v e r these p o i n t s exist.
For suppose i - j and t h a t i is even and j is odd. T h e n j <~j - 1, j a n d if

j + 1 < m t h e n also j <1j + 1. By the p - m o r p h i s m condition for <~ a n d the fact
thati<~kiffk-iwegetj-l-iandj+l-i. Similarly, i f i > 0 t h e n b y
the p - m o r p h i s m condition for c> we have i - 1 - j a n d if i < m also i + 1 - j.
C o n t i n u i n g this a r g u m e n t we get k - g for all k, g a n d hence n - 0, which we
have excluded. Now let again i - j. T h e n if b o t h are even, i - 1, i, i + 1 <~ i
a n d j - 1, j, j + 1 <~j whenever these points exist. By the second p - m o r p h i s m
condition, e i t h e r i - l - j - 1 ori-l-j ori-l-j+l. B u t s i n c e j is
even a n d i - 1 is odd, i - 1 - j c a n n o t hold. Likewise, i + 1 - j as well as
i - j - 1, j + 1 cannot occur. This proves Claim 7.9.6.
In order to prove t h a t m is a multiple of n we look at subsets C of 0Ira
which are c o n n e c t e d and on which 7r [ C is injective. Such sets are called
p a r t i a l s e c t i o n s . If 7r r C is also surjective, in other words, if ~C - n + 1
t h e n C is called s e c t i o n . We now prove
CLAIM 7.9.7. I f C is a p a r t i a l s e c t i o n a n d ~C > 1 t h e n C is c o n t a i n e d in
e x a c t l y one section.
To see this observe first t h a t since C is c o n n e c t e d a n d 7r is a <~-ho-

m o m o r p h i s m , 7r[C] is connected as well. Therefore, C C_ { 0 , . . . , m ) is an
interval as is 7r[C] C_ { 0 , . . . , n } . Ifi, i+l E C and 7r 9 i ~ k t h e n by
Claim 7.9.6, T r ( i + l ) - k + l orTr(i+l)-k-1. IfTr'i+l~k+lthen
7r 9 i + 2 ~ k + 2, for by the same a r g u m e n t 7r 9 i + 2 ~ k, k + 2 b u t 7r(i + 2) - k
c o n t r a d i c t s the injectivity of 7r F C. B u t if 7r 9 i + 1 ~+ k - 1 t h e n similarly
7r 9 i + 2 ~ k - 2. So, by induction, either 7r [ C is a strictly increasing or
strictly decreasing function. Now if C - { i , . . . , j ) a n d 7 r [ C ] - { k , . . . ,g} a n d
7r is m o n o t o n e increasing t h e n if k > 0, 7r(i) - k > 0 we get by the second
p - m o r p h i s m condition t h a t either i - 1 or i + 1 is in the p r e i m a g e of k - 1.
ButTr(i+l)-k+l. Thusi>0andTr(i-1)-k-1. So we a d d i - l t o C .
Likewise, 7r(j) - g and if g < n t h e n j < m and we add j + 1 to C. Similarly,
if Ir is decreasing. This proves Claim 7.9.7.
Now 0Ira contains exactly m subsets of the form {i, i + 1}. 7r is injective
on each of t h e m and they are all c o n t a i n e d in one and only one section. E a c h
section contains n § 1 points and thus n subsets {j, j § 1}. Hence n divides
morm-w.
(r I f n = 0 take the constant m a p j ~-+ 0. I f n > 0, gm must be covered
by sections as follows. If S, T are sections then S = T or ~(S n T) < 1. Each
section is an interval of n + 1 points and each pair {i, i + 1} is in exactly one
section. Hence the sections are Sk -- {nk, nk + 1 , . . . , n ( k + 1)}. On each
section 7r is bijective. Suppose that 7r is increasing on Si. T h e n 7r(n(k + 1)) -
n. Thus 7r must be decreasing on Si+l and vice versa. Thus let 7r be increasing
on all even n u m b e r e d sections S2i and decreasing on all odd n u m b e r e d sections
$2i+1. Thus 7 r ( i ) - s i f f i - 2 k n + s or i - 2 ( k + l ) n - s for some k. We
show t h a t 7r defined this way is a p - m o r p h i s m . We have i <~i and 7r(i) <~ 7r(i).
Moreover, i f i < ~ i + 1 or i < ~ i - i then i is odd. Now if 7r(i) - s then
eitheri- 2kn+s ori- (2k+2)n-s. In b o t h cases s i s odd as well and
s <1 s + 1, s - 1 and {s - 1, s + 1) - ~ [ { i - 1, i + 1}]. Similarly if i is even. Now
suppose 7r(i) <1~. If i E S, 7r(i) <l~, then take s C 7 r - l ( ~ ) N S . s is unique. Since
7r F S " (S, <~) -+ g[,~ is an isomorphism, i <~ s as well. Similarly if e <~ ~r(i). E]
W i t h this result in our hands we can probe quite deeply into the structure of
G a and also prove the desired theorem. We have t h a t G a - Th g[~ since
Th g[~ 2 Th gin for every n. Each logic containing G a must be complete. This
is due to the fact t h a t logics of b o u n d e d alternative are complete in general.
The ["]-irreducible elements are the Th g[,~ for n E w. Every proper extension
of G a which is not trivial is therefore an intersection [7 (Th g[,~ 9 n C F ) where
F C w is finite. For if F is infinite we immediately have
(Thgn'n e F) - G a ,
since g[o~ is contained in Up gin for a non-trivial ultrafilter U on F . (This
can also be shown without the use of ultrafilters. This is left as an exercise.)
Therefore every proper extension of G a is tabular while G a itself is not.
THEOREM 7.9.8. G a is pretabular.
It is now straightforward to m a p out the structure of the locale ~ G a . We
will establish here the structure of the corresponding TD-space. Recall t h a t
the latter consists of the [']-irreducible logics as points, and the closed sets
correspond to sets of the form I"A where A is a logic. Thus, by the results
obtained above, the points are logics Th gn, n e w. Let # := ( w - {0}, I) ~
where m I n iff m divides n. Now let # + 1 be the poset obtained by adding a
new top element. Recall that (I)(# + 1) denotes the weak topology on # + 1. In
this topology, an upper set is closed if it is (i) the entire space or (ii) a finite
union of sets of the form j'x.
THEOREM 7.9.9. 3Gpc(~ Ga) ~- (~(# + 1).
The corresponding space Gpc(~ G a ) can easily be determined. It has a
new element at the b o t t o m (corresponding to G a itself, which is R-irreducible
FIGURE 7.15. E G a
2 3
e" 4,3 dr 6 9 2,3, ~ ~
but not ["]-irreducible). The closed sets are those sets which are finite (and
hence do not contain G a ) or contain G a (and hence all other elements). Thus
the upper part of E G a is depicted in Figure 7.15. To the left of each node we
have written numbers n such that the node is the intersection of the logics of
the corresponding 0In.
PROPOSITION 7.9.10. For n > 3 there are infinitely many logics of codi-
mension n in E Ga.
The proof of Lemma 7.9.4 is now easy. Clearly, b o t h 0[0 and 011 split
the lattice. But for n > 1 observe that the sequence (0[p " P prime, p > n)
subreduces 0In. As we have noted, this implies also that none of the garlands
0In split ~ S4.t unless n < 1. It will turn out soon that we cannot improve
this result for ~ 84.t. But for ~ K . t and ~ K 4 . t even these cases are ruled
out. Look at the sequence (010 " n C w) where ~[: differs from ~[,~ in t h a t
n ~ n. The maps 7r and p defined b y T r " 010 ~ 011 " J ~ J ( m o d 2 ) and
P " 0[~ -~z 0[0 " J ~ 0 are n-localic with respect to 0. Thus this sequence
subreduces b o t h frames in E K 4 . t and in E K.t.
LEMMA 7.9.11. For no n, Th ~[n splits ~ K.t, ~ K~.t.
It now remains to treat the clusters. Here the situation is quite similar
to the situation of the garlands.
LEMMA 7.9.12. Suppose g is a cluster. Then Th 0 splits ~ K.t and ~ K ~ . t

only if 0 ~- [-~ and ~ S~.t only if 0 ~- ~ .
FIGURE 7.16. t~k
%/% %
PROOF. Let n := ~g > 1 and Ok = (hk, ~) with
hk = { 0 , . . . ,k} • { 1 , . . . , n } - {(k,n)}
and (i, j) <~ (i',j') iff (i) i is odd, i' = i + 1 or i - 1 or (ii) i is even and
i I = i. This can be visualized by Here, [] denotes a cluster with n points and
o a cluster with 1 point. There is no p - m o r p h i s m from I~k into 9 as there is
no way to m a p a point belonging to an n - 1-point cluster onto a n - p o i n t
cluster.
Now look at the k - t r a n s i t of (0, 0) in I]k; call it e. Let e be its underlying
set. Every point in e is contained in an n - p o i n t cluster since (i, j) C e iff i < k.
Thus there is a p - m o r p h i s m 7r : e --+ g. E x t e n d 7r to a m a p p+ : Ilk ~-* 9. P+
is k-localic with respect to (0, i) for every i. Hence l~k is k - c o n s i s t e n t with 9.
It follows t h a t (Th I]k: k C w) is a subreduction of Th 9. K]
Now we have collected all the m a t e r i a l we need to prove the splitting theorems.
Notice t h a t a splitting frame for any of these logics can only be o n e - p o i n t
cluster or a t w o - p o i n t garland. We will now show t h a t the frames not excluded
by the above l e m m a t a are indeed splitting frames.
THEOREM 7.9.13. A splits the lattice ESg.t iffA = Th g[1 = Th I o I

or A = Th g[0 = Th ~5-].
PROOF. (r The nontrivial part is g[1. We will show ~ $ 4 . t / g [ 1 = $ 5 . t

by proving t h a t (t) of the Splitting T h e o r e m holds for m = 1. Therefore let
92 be an algebra satisfying Th 92 ~ Sht. T h e n there is a set c C A of 92 such
t h a t 0 < c A b 9 - c . Consequently, in the underlying K r i p k e - f r a m e there
are two points s <1 t such t h a t s C c and t C B - c whence t r s. Now we
have (~(9[1) = (Pa ++ -~Pb) A <~PbA e~pa A ([Spb t-+ Pb) A (apa ++ p~). Suppose
t h a t under these circumstances we can construct valuations a n : {p~, Pb} ~ A
such t h a t s C ~n(P~ A 5<n5(9[1) ). T h e n (t) of the Splitting T h e o r e m , 7.3.11
(2), 911 splits E S4.t. To construct the a n we define inductively subsets an, bn
in 92 as follows:
(0) b0 := ~ - c
a0 := -b0
(A) aek+l := --b~k+l
b2k+l :-- b2k ~ "~ a2k
(B) a~+~ "- a~+ln ~-b~+~
a2k+2 :: --a2k+l
Furthermore define To := {s}, T2~+1 := r T2k+2 := CT2k+l. The Tn

are not necessarily internal. Since Tek C_ -~ ~T2k = ~ T 2 ~ + 1 and further-
more Tek+l C_ ~- -~ T2k+l = ~- Te~+I it follows that Tn C_ []<lTn+ 2. For
~]-<lT2k+2 := ~ - T 2 k + 2 ~ l T 2 k + 2 ~_ T2~+1 ~ l T 2 k + l ~_ T2k and dually for odd
n. Consequently, s ~ ~<-nTn+l. We now verify the following claims:
(I) an ~ bn =
an [J bn = 1
(II) an -- lien
bn = ~- bn
(III) Tn N an = Tn ~ an+l
Tn A b n = Tn A b n + l
(IV) T~ C ~-b,~~ -~ a,~
(I) holds by construction. ( I I ) - (IV) are verified by induction; for (II) we only
need to show an C_ l an and bn C B bn. By symmetry of (A) and (B) we may
restrict to (A). b 2 k + l - - b2k A "~ a2k -- m-b2k A "~ a2k C_ m-b2k A m- "~ "~ a2k --
B(b2k • @a2k) - - B b 2 k + l , a2k+l = - - b 2 k + l - - - B - b 2 k + l -- ~a2k+l C_
Ibba2k+l - l ba2k+l - -m - B b 2 k + l -- -m --b2k+l -- l a~k+l. This
shows (II). For (III) we now observe that T2k N b2k+l -- T2k A bek A "0 a2k --
T2k N b2k since T2k C_ "0 a2k by (IV). T2k N a2k+l -- T2k A a2k immediately
follows. To prove (IV) we observe that if y E T2k+l there is a x C T2k such
that x c> y. By induction hypothesis we have x E b b2k and therefore there is
a zE>x such that z C T2kAb2k. Now z C b2kN'Oa2k --b2k+l and, as y<~x<~z,
y C bbe~+~. To show y E ~a2~+1 we distinguish two cases: (a) y ~ a2k+l
and (fl) y ~ b2~+1. In case (a) we immediately have y ~ ~ a2~+1 and in case
(/3) we have y ~ ~ a2~. But as
a2~+1 - ~ (a2k U i b 2 k ) ~_ "~ a2k
we also have y C ~ a2k+l.
Now we put an "Pc ~ an+l,Pb ~+ bn+l. It remains to be shown that s C
~n(Pa A ~]<nh(g[1) ). Notice that (I) - (IV) together yield Tn+l C_ ~n(5(g[1) )
whence {s} C_ ~]<nT,~+l C_ ~n(~]<nh(~[1) ). And since by (III) and the fact
that s C Tn+l we have s C an(pa), everything is proved. K]
THEOREM 7.9.14. The following holds.

1. 0 splits ~ K.t iff O = Th Fg~.
2. 0 splits ~ K 4 . t iff O = Th ~g].
PROOF. [-~ is cycle-free and therefore splits E K.t. A fortiori it splits
K 4 .t. [::]
Notes on this section. The results of this section have mainly been established
in [123]. The connections between monomodal logics and their minimal tense
extensions have been investigated thoroughly in a series of papers by FRANK
WOLTER. He has shown that many properties are lost in passing from a
monomodal logic to its minimal tense extension; these are among other the
finite model property and completeness. (See [238], [239].) WOLTER has
also shown that the minimal tense extension of a modal logic need not be
conservative. For completeness results in tense logic see also [242] and [236].
E x e r c i s e 266. Complete the proof of Lemma 7.9.3.
E x e r c i s e 268. Show that S4.2.t, where S4.2 - S4 9 0[:]p -+ []0p, is 2-

transitive. Hence show that every finite connected frame splits ~$4.3.t and
S5.t.
E x e r c i s e 269. Show that ~ S4.t has exactly two elements of codimension 2.
E x e r c i s e 270. Let ~[~ the frame consisting of g[~ and the finite and cofinite
sets as internal sets. Describe the bidual (~[~)++. Show that its algebra of
sets is subdirectly irreducible but that the frame is not connected.
E x e r c i s e 271. Show without the use of ultrafilters that ~')nCM ~[n -- G a

whenever M C w is infinite.
CHAPTER 8
E x t e n s i o n s of K 4
8.1. T h e G l o b a l S t r u c t u r e o f s
The first general results in modal logic were obtained for transitive logics.
Moreover, transitive logics have for a long time been in the focus of interest of
modal logicians since the motivating applications very often yielded transitive
logics. Moreover, as we shall see, the structure of K 4 - f r a m e s is also by far
easier than the the structure of frames for nontransitive frames. The first
general result about a large classes of logics rather than individual logics was
undeniably ROBERT BULL'S [34] and the sequel [60] by KIT FINE. These
papers gave an exhaustive theory of logics containing $4.3. However, $4.3
is a very strong logic, and the result therefore did not cover many i m p o r t a n t
transitive logics. The starting point of a general theory of transitive logics was
the paper [63] by KIT FINE in which it was shown that the addition of an ax-
iom of finite width makes transitive logics complete. The paper also contained
useful results concerning weak canonical frames. The paper [66] was another
breakthrough. It introduced the notion of a subframe logic. It was shown
that all transitive subframe logics have the finite model property. Moreover,
characterizations were given of canonical subframe logics. Many important
logics turn out to be subframe logics. These results were discovered around
the same time by MICHAEL ZAKHARYASCHEV, who introduced the still more
general notion of a cofinal subframe logic. All transitive cofinal subframe lo-
gics were shown to have the finite model property. Moreover, each transitive
logic is axiomatizable by a set of so-called canonical formulae. These are for-
mulae which characterize the frames of the logic by some geometric condition.
This allows to deal with transitive logics by means of geometric conditions on
frames rather than axioms. These results have established a novel way of
thinking about transitive logics, and have also influenced the general theory
of modal logic through the concepts of subframe and cofinal subframe logic.
Before we plunge into the theory of K 4 - f r a m e s , we will introduce some no-
tation and recall some easy facts about K 4 - f r a m e s and p - m o r p h i s m s between
them. In transitive frames we say that y is a s t r o n g or s t r i c t s u c c e s s o r of
x if x <~ y r x and that y is a w e a k s u c c e s s o r of x if x <~ y or x - y. We
write x<~y if y is a strong successor of x and x ~ y if y is a weak successor
397
398 8. E x t e n s i o n s of K4
of x. A cluster is a maximal set C of points such t h a t given x, y C C then

x is a weak successor of y. Thus, for elements x, y of a cluster C, if x r y
then x <1 y. Consequently, if ~C > 1 then <1 M C 2 = C 2. On the other hand,
i f ~ C = 1, s a y C = {x}, then either x < 1 x (and so <1MC 2 = C 2) or x ;~ x
(and so <1 M C 2 = 0 ) . C is called d e g e n e r a t e if it contains a single point
and x r x, and i m p r o p e r i f C = {x} w i t h x < 1 x . I f ~ C > 1, C i s called
p r o p e r . The t y p e of a cluster C is the cardinality of the set {y : x <1 y} for
some x E C. This is clearly independent of the choice of x. If the type is at
least 2 then the cluster is proper, if the type is 1 the cluster is improper, and
the cluster is degenerate if the type is 0. Instead of saying t h a t a cluster has
type 0 we also say that it has type ~. The cluster containing a given point x
is denoted by C(x). If C and D are clusters, and for some x E C and y C D
we have x <1 y, then in fact for all x C C and all y C D holds t h a t x <1 y.
This justifies the notation C <1 D for clusters. Finally, if C and D are distinct
clusters and C <1 D, then all points of D are strong successors of all points
in C. Consequently, we may represent a transitive frame as a triple /P, ,l, v}
where (p, 4} is a partial order and ~ a function from p to w. ((p, ~} is the
ordering of clusters, and p(x) is the type of the cluster represented by x.) Let
= (f, <1} be a Kripke-frame. Recall t h a t the depth of a point x, in symbols
dp(x), is an ordinal number defined by
dp(x) := {dp(y) : xJ~y}
This requires t h a t dp(y) is defined for all strong successors y of x. In parti-
cular, x is of depth 0 if for every successor y of x, y <1 x. In other words, x
has no strong successors. We then say t h a t the cluster of x is final. Notice
t h a t the definition assigns depth only to points for which there is no infinite
ascending chain of strong successors, or to rephrase this, if the partial order
of clusters from x onwards has no ascending chains. The definition as given
also extends to infinite ordinals; for example, a point is of depth w if all strong
successors are of finite depth and, moreover, for each n E w there is at least
a point of depth n. It is left as an exercise to show t h a t if x is of d e p t h a
and a :>/3 then there is a strong successor of depth 13. We call the points of
a given depth c~ the c~-slice of the frame.
DEFINITION 8.1.1. Let ~ be a Kripke-frame and a an ordinal number.

Then [<a denotes the subframe of points of depth < c~. Likewise, ~>~ and ~
denote the subframes of points of depth > a and = c~, respectively. ~<~ is the
subframe induced by ~ on the Kripke-frame ~<~. Likewise for ~>~, if that is
well-defined.
DEFINITION 8.1.2. Let ~ be a frame. The depth of ~ is the set of all
ordinals a such that there exist points of depth c~. A frame is said to satisfy
the ascending chain condition (acc) iff there exists no infinite, strictly
ascending chain of points. Such frames are also called noetherian.
8.1. T h e Global S t r u c t u r e of ~ K 4 399
Notice t h a t a frame ~ is n o e t h e r i a n iff ~ = ~<~ for some c~. (In fact, any
ordinal greater or equal to the d e p t h of ~ suffices if ~ is n o e t h e r i a n . )
We begin by recalling the frame c o n s t r u c t o r 0 , s t a n d i n g for t h e disjoint
union. It takes two frames and places t h e m side by side. We say t h a t it places
the frames p a r a l l e l to each other. Now we i n t r o d u c e a second c o n s t r u c t o r ,
Q . R a t h e r t h a n placing frames parallel to each other it places the first before
the second. To be precise, let f = (f, <1/) and g = (g, <~g) w i t h f a n d g disjoint.
Then
:= / x gu
Thus, the u n d e r l y i n g set is the disjoint union of the sets of the individual
frames. A n d x < ~ y i f f e i t h e r x and y are in f a n d x < ~ / y or x in f and y
in g (and no other condition) or x and y are b o t h in g a n d x qg y. Now if
I = (I, <) is an ordered set and [~ = (f, <~i), i e I, an i n d e x e d family of
frames on disjoint sets t h e n
.= <U U <, x i,>

I I i<j
PROPOSITION 8.1.3 (Decomposition). Let [ - g Q O. (1.) A generated

subframe of ~ is of the f o r m b (~ ~ for a ~ H g or of the f o r m 0 for some b ,-~ O.
f..
(2.) If P " g --" g and q " b --~ t~ are contractions, so is p Q q " g ~ ~ -~ ~ ~.
Call a p - m o r p h i s m l a t e r a l if it is d e p t h preserving. A c o n t r a c t i o n m a p
" g ~ 0, g, ~) finite, is m i n i m a l if in any d e c o m p o s i t i o n ~ - p o p~ either p
or p~ is an isomorphism. Any c o n t r a c t i o n of a finite frame can be d e c o m p o s e d
into a sequence of m i n i m a l contractions a n d there is only a small n u m b e r of
such contractions. First of all, let us take a point v of m i n i m a l d e p t h which is
identified via ~ w i t h a n o t h e r point and let w be of m i n i m a l d e p t h such t h a t
~ ( v ) - ~(w). Now two cases arise. E i t h e r d p ( w ) - d p ( v ) o r dp(w) > dp(v).
In the latter case it is easily seen t h a t there exists a pair w and v ~ such t h a t
v' e C ( v ) a n d dp(w) - d p ( v ' ) + 1. We m a y assume t h a t v' - v. A n o n - l a t e r a l
p - m o r p h i s m collapses at least a point with an i m m e d i a t e successor. In t h a t
case, the successor c a n n o t be irreflexive. F u r t h e r m o r e , if w q v t h e n C ( w )
is m a p p e d onto C ( v ) . So we have the situation 7r 9 m G n ~ k. T h e n the
restriction of ~ to the cluster n is surjective. Hence k _~ n. Moreover, k < n
implies t h a t we can d e c o m p o s e ~ in the following way:
m(~n -~ m ( ~ k -~ n Q k .
This is left for the reader to verify. So, by m i n i m a l i t y and the fact t h a t we
have a n o n - l a t e r a l contraction, k - n. Likewise, m :> n c a n n o t occur, because
then we can d e c o m p o s e ~ into
So m < n. ~ 9 m Q n --- n is actually i n d e c o m p o s a b l e iff it is injective on

b o t h clusters. If m < n such a c o n t r a c t i o n can be constructed.
Now we come to the case where two points of identical d e p t h are identified.
Let d be the m i n i m a l d e p t h of points on which p is not injective. It is clear
t h a t if ~ is a contraction, and we let ~ be defined by ~(w) := w if w is of
d e p t h ~ d and ~(w) - ~(w) else, ~ is a c o n t r a c t i o n as well, and 7r factors
t h r o u g h ~. Thus by m i n i m a l i t y ~ is lateral. Let v and w be two distinct
points such t h a t ~(v) - ~(w). T h e n two cases arise. 1. C a s e . C(v) - C ( w ) .
T h e n ~ is of type m - - - n f o r m > n. 2. C a s e . C(v) # C(w). T h e n ~ i s
of the t y p e 7r 9 m @ n ~ k. T h e n either all three are d e g e n e r a t e or all t h r e e
are n o n d e g e n e r a t e . Again we find t h a t 7r m u s t be injective on b o t h clusters,
hence m - n, and t h a t k < n contradicts minimality. So, ~ 9 n (~ n --- n, and
7r is injective on the individual clusters.
THEOREM 8.1.4. ~ is a minimal contraction iff it is of the following types.
Lateral. (a.) n + l n, >__ 1, (b.) o 9 o (c.) n 9 n n.
Nonlateral. ~ n --, n or m Q n - , n with m < n. In all cases, 7r is
injective on the two individual subclusters.
Recall the notion of a local subframe. We know t h a t if @ is a local
s u b f r a m e of ~ t h e n any c o n t r a c t i o n on @ can be e x t e n d e d to a c o n t r a c t i o n
on ~ which is the identity on all points not in ~. For example, clusters are
always local in a K r i p k e - f r a m e . Notice t h a t this is so no m a t t e r how t h e y are
e m b e d d e d into the frame. So, clusters can always be contracted. Moreover,
there are additional situations which are useful to know. For example, in a
finite frame, if x has a single i m m e d i a t e successor, y, and y is reflexive, t h e n
the set {x, y} is local. For if x <~ z t h e n either x - z or y _ z by the fact t h a t
y is the only i m m e d i a t e successor, and so y <~z, since y is reflexive. And y <1 z
implies x <~ z. Thus the set is local, and there is a c o n t r a c t i o n reducing it to
a single point in the total frame. We call x in this s i t u a t i o n a u n i f i e r (for y).
In general, x is a unifier for a set X if X is the set of all i m m e d i a t e successors
of X.
Say f (or [) is t o t a l l y l o c a l if it is local and for all x E g - f a n d all
y, z C f we have x <~ y iff x <~ z. In case f is totally local, we write s o m e t i m e s
~[[] to denote the given, totally local occurrence of [. A r e p l a c e m e n t of [ by
a n o t h e r frame ~) is t h e n u n a m b i g o u s l y defined. We define ~[[~/[] to be a frame
with underlying set g - f + h and relation x ~ z iff (1.) neither x and z are in
andz<lzor (2.) x C g a n d z E h o r x E h a n d z C g o r (3.) b o t h x , z i n h
and t h e n x <~ z (in 0).
F r o m a theoretical viewpoint it has shown useful to i n t r o d u c e four pa-
r a m e t e r s to classify transitive logics.
D e p t h . Let ~ be a refined transitive frame. The d e p t h of ~ is the set of all
a such t h a t there exists a point of d e p t h a in ~. We write dp(~) to d e n o t e
the d e p t h of ~. The d e p t h of A _~ K 4 is the s u p r e m u m of all dp(~), where
8.1. The Global Structure of 8 K 4 401
is a refined A-frame - if such a supremum exists. Otherwise A is said to be

of unbounded depth. The logic of frames of depth ~ n can be axiomatized
for finite n. Namely, put
dp(> n) "- po AWl+Ai</<n+lpi --~ ~pj

AE]+ A~<j<n+l pi -+ 0pj
Aft]+ Ai<j<n+l PJ -+ -~Pi
(Here the convention E]+~ "- ~ A ff]~ is used.) dp(> n) is satisfiable in ~ if

there is a chain (xi 9 i < n) of length n such that xi ~ Xi+l for every i < n - 1 ;
we call such a chain p r o p e r l y a s c e n d i n g . Clearly, if no such chain exists,
is of depth < n, and conversely. Let dp(_< n ) : = -~dp(> n). This axiom is
also denoted by Jn in the literature. Then K4.dp(_< n) is the logic of frames
of depth _< n.
W i d t h . A set {x0,... ,Xk} of points is an a n t i c h a i n of length k + 1 if for
no two distinct points xi, xj we have xi <~ xj. A refined frame ~ has width
n if there is a point x and an antichain Y of length n such that x sees every
member of Y. We denote by wd(~) the width of ~. The width of A is the
supremum of all widths of refined A-frames (if such a supremum exists) and
denoted by wd(A). If a logic is of width n for some n < w it is of finite
w i d t h . A logic is of w i d t h n if it satisfies the axiom
n
n):= opt. V o(p A pj) v V O(p A (>pj)

i=0 i#j ir
This axiom is also denoted by In in the literature.
T i g h t n e s s . Let ~ be a refined transitive frame. Let x be a point and x <~ y.
Moreover, let ~ be a subframe such that no point of ~ is comparable with x.
Then if ~ is of depth a, ~ is said to be of t i g h t n e s s a. We write ti(~) - a in
that case. The t i g h t n e s s of A _D K 4 is the supremum of all ti(~), where ~ is
a refined A-frame, if such a supremum exists. Consider the following formula.
ti(> n) "- Po AD+ Ai<j<n+2pi --+ -~Pi

A[:]+P0 -+ OPn+l
AD+ Ai<j<n+l P~ --+ Opj
AD+ Aj<~<n+2 P~ --+ -~0pj
A[:]+ Ai<n+l P~+I ~ -~0Pi
A[::]+ A0<i<~+~ pi -+ ~0p~+~
This formula is satisfiable in a refined frame iff there is a point seeing a
properly ascending chain of length k + 1 and an incomparable point. Let
ti(< k ) " - -~ti(> k). Then ~ ~ ti(<_ k) iff ~ does not have tightness > k. So,
K 4 . t i ( < k) is the logic of frames of tightness _< k.
402 8. E x t e n s i o n s of K 4
F a t n e s s . For a cardinal number c~, the cluster ~ is the cluster containing c~

many points. We also write n for n_, n E w. We call the type of a cluster C
also its f a t n e s s and denote it by ft(C). The fatness of a refined frame, ft(~)
is the supremum of all ft(x), where x is a point of ~. The fatness of a logic A
is the supremum of the set of all ft(~), ~ a refined frame for A. Equivalently,
a frame fails to be of fatness n if the following formula is satisfiable
f-t(> n ) . - po AD]+ Ai<j<n+l pi -+ -~pj

AD+ Ai<j<n+l pi -+ Opj
The logic of frames of fatness O is the logic G = (}p --+ (}(p A M~p).
The significance of the choice of these logics will become clearer later
on. Let us say here only that the logics of finite depth Kd.dp(_< k) form an
infinite descending chain in the lattice ~(K4); likewise the logics K4.Ik (alias
K4.wd(_< k)) and K4.ti(_< k) as well as K 4 . f t ( < k). Moreover, the axiom of
linearity, known as .3 is the same as I1 and the same as wd(<_ 1). G r z is the
logic of reflexive frames of fatness 1, G the logic of frames of fatness Z. The
last two statements need a rigorous proof, but with the methods developed in
this chapter this is actually very easy.
E x e r c i s e 273. Show that m Q n --~ 1 Q n --~ n.
E x e r c i s e 274. Let f be a finite reflexive frame. Show that there is a p -

morphism [ -~ g such that ~g - ~ f - 1. So there always exists a truly minimal
p-morphism.
E x e r c i s e 275. Let [ - {f,<l} be a frame of depth >_ 2. Define x ,, y by

x r y a n d y r x. We call (f, ll} the a n t i f r a m e of[. Show that if f is @ -
decomposable the antiframe (f, ll} is disconnected. Show that the converse
need not hold.
8.2. T h e S t r u c t u r e of F i n i t e l y G e n e r a t e d K 4 - F r a m e s
Fundamental for the study of the lattice E K 4 is the fact that the structure
of the finitely generated algebras is rather well-behaved. Looking at the
underlying Kripke-frame for a refined frame, we can say that all these algebras
contain an upper part, consisting of points of finite depth, and that each point
of infinite depth is 'behind' this upper part. Moreover, this upper part is
atomic, and each level is finite. It should be said that the infinitely generated
algebras are not as nice as that, so the restriction to finitely generated algebras
is indeed necessary. But we know that logics are complete with respect to such
8.2. The Structure of Finitely Generated K 4 - F r a m e s 403
frames, so we do not loose anything by specializing to such frames. Before we

begin let us note a simple fact.
THEOREM 8.2.1. Let ~ be a transitive refined frame and F be n-generated.
Then the clusters have type 0 or t for t ~ 2 n.
PROOF. Let A = {ai : i < n} be a generating set and let x and y be two
points in a cluster. By induction on the sets generated from A it is shown
that x C a~c=~y C ai for a l l i < n implies that x C b c ~ y C b f o r a l l b c F.
Hence a cluster can have at most 2 n distinct points.
This idea of induction on the sets generated from A can be made precise
in the following way. Let us take a general frame ~ = ([, F / for K 4 which is
refined, such that the algebra of subsets is n-generated. Let { a 0 , . . . , a,~-l}
be a generating set. Let Pn := {Pi : i < n}. P u t v : P n --+ F :Pi ~-~ a~, i < n.
Then for any b C F there is a ~ with var(~) C_ Pn and b = ~(~). Therefore
we can prove a property of all internal sets by induction on the constitution
of formulae ~ such that var(~) C_ IPn. If ~(~) = b and dp(~) = k, we say t h a t
b is k - d e f i n a b l e . (Notice that 0-definable was also defined to mean that
the set is definable by a constant formula. Throughout this chapter we will
not use the term 0-definable in that latter sense.) Notice finally, that if ~ is
n-generated then any generated subframe is n-generated by the same set of
generators. Since ~ is refined, [ is transitive. We will now show that [ contains
a generated subframe [<~ such that each point of [<~ is of finite depth and,
moreover, any point not in [<~ sees a point of arbitrary finite depth. Such a
structure is called top-heavy.
DEFINITION 8.2.2. A frame ~ is called t o p - h e a v y if every point not in
~<~ has a successor of depth n for any n < w.
[<~ will be defined inductively. The generated subframe of ~ induced by
[<~ will be denoted by ~<~.
LEMMA 8.2.3. Let c~ be an ordinal and [ be a Kripke-frame. Then the
subframe ~<~ consisting of all points of depth < c~ is a generated subframe of
f.
PROOF. Let x be of depth ~. dp(x) is well-defined only if all strong
successors have a depth. By definition, if x <~ y then either y is not a strong
successor, in which case y belongs to the cluster of x and is of equal depth,
since x and y have the same successors and the same strong successors. Or
y is a strong successor and has a depth according to the definition of the
depth. [~
LEMMA 8.2.4. Let ~ - ([,F} be n-generated and refined. Then the fol-
lowing holds.
1. There are at most 2~(2 2n-1 + 1) points of depth O.
2. ~<1 is a full finite frame.

3. Every set a C_ [<1 is 2-definable.
4. Every point has a weak successor of depth O.
PROOF. Because ~<1 is a g e n e r a t e d s u b f r a m e of ~, it is refined a n d gen-

e r a t e d by the hi. So if (1.) holds, (3.) holds as well, since finite refined f r a m e s
are full. Let us s t a r t t h e n by a t t a c k i n g (2.). For S C_ {0, 1 , . . . , n - 1} define
:= A AV
ics i~s
P u t As "- ~(x(S)). Every node is in exactly one of the sets As. If x C As,
x(S) is called the a t o m of x. Let C be s y s t e m of subsets of {0, 1 , . . . , n - 1}.
Define
e-span 9- A < O x ( s ) . s c e) A A< Ox(s). s r e>

._ { E]J_
mspan
V V(x(S) A e-spanA A~)ce~<>9 S E e)
We say t h a t x is of s p a n C if x C ~(C-span). We say t h a t x is of m i m i n a l
s p a n if it is in ~(mspan). So, x is of m i n i m a l span iff it is w i t h o u t successor or
it has successors, a m o n g which one has the same a t o m as x and all successors
of x have the same span as x. T h e set of points of span C is 1-definable, the
set of m i n i m a l span is 2-definable.
(]~) Let x and z be points of identical span and identical atom. A s s u m e t h a t

x and z are of m i n i m a l span. T h e n x - z.
Let the span of x by C. We show by induction on ~ t h a t x C ~(~) iff z C ~(~).

It follows t h a t x C b iff z C b for all b C IF and by refinedness x - z. Now for
the inductive proof. Since x and z have the same a t o m , x C ~(p~) iff z C ~(p~)
for all i < n. The induction steps for -~ and A are straightforward. Now let
= <>7/. Assume t h a t x E ~(~). T h e n there exists a y such t h a t x <1 y and
y C ~(r]). Now y has m i n i m a l span (as it succeeds a point of m i n i m a l span).
Hence its span is equal to the span of x. Let x(S) be the a t o m of y. T h e n
Ox(S) is a conjunct of C-span. But z has span C, and so t h e r e is a point
u such t h a t z <] u and u has a t o m x(S). Also u has span C. By i n d u c t i o n
hypothesis, u E ~(r/). Hence z e ~((>~) = ~(~).
It follows now t h a t a point x is of d e p t h 0 iff it is of m i n i m a l span. F r o m
left to right is clear. Now assume t h a t x is of m i n i m a l span. If it has no
successors, the case is clear. Now let x have successors. Let y be a successor
of x. By a s s u m p t i o n , the span of y equals the span of x, so t h e r e is a successor
z of y satisfying the same a t o m as x, by definition on the formula mspan. By
(]~), z = x, since z and x have the same span. It follows t h a t y <] x. So, x is
of d e p t h 0.
Now we come to (1.). W h a t we need to specify for a point of depth 0 is

its span, which is then minimal, and hence defines the size of the cluster as
well as the atoms which are represented in it, unless the point is successorless,
in which case its own atom suffices for a characterization. There are at most
2 n degenerate clusters. If the cluster is nondegenerate, it has size between 1
and 2 n, for each size k there is a choice of (2kn) atoms. Summing this we get
~[l_<2n+Ek(2;) _2~+2 n 2~-lk --2n(22n-l+l)

k--1 k=O
Moreover, each point in that frame is 2-definable. Hence each subset is. This
shows (3.).
Next we show (4.). P u t Z0 (x) := { S : x C CAs}. Z0 (x) is finite. Hence
there is a weak successor y of x such that for all z ~ y, Zo(z) = Zo(y). Two
cases arise. C a s e 1. y is reflexive. Then y is of minimal span and so of
depth 0. C a s e 2. y is irreflexive. Then if y has no successors it is of minimal
span and so of depth 0. Finally, if y has successors, then let z > y. Since
Zo(z) = Zo(y) and since the atom of z is in Zo(y), the atom of z is in Zo(z).
It follows that z is of minimal span. []
Let the function 5(n, k) be defined inductively as follows
1) .= 2 (2 + 1)
5(n,k + 1) "- 5(n, 1)(2 ~(n'k) - 1)
THEOREM 8.2.5. Let ~ -- ([,F) be an n-generated refined K 4 - f f a m e .

Then the following holds
1. There are at most 5(n, k § 1) points of depth <_ k.
2. ~<k+l is a full finite frame.
3. Every subset of ~<k+l is 4k + 2-definable.
4. Every point in ~ not of depth < k has a weak successor of depth k.
PROOF. The case k - 0 is covered by the previous lemma. Now we do
the induction step for k. Again, (2.) follows from (1.) and (1.) follows
from (3.). For each point x of depth < k there exists a formula 7(x) in the
variables P,~ such that x C ~(7(x)). Let us consider the set P of formulae
7(x) where x E f<k. We know by induction hypothesis t h a t ~P <_ 5(n, k) and
that for each x, 7(x) if of degree < 4k - 2. Moreover, the disjunction of all
7(x) is a formula of degree 4k - 2 defining the points of depth < k. Denote
it by dp(< k). Let Q c_ P. Say that a point in the frame is of w i d t h Q i f
x E ~(Q-width) where
Q - w i d t h ' - -~dp(< k ) A A ( ( ~ C " C c Q) A A ( ~ ( } C " C e P - Q)
Clearly, a point in the frame is of width Q iff it is not of depth < k and sees
exactly the points defined by Q. Notice that the width can never be empty, in
406 8. E x t e n s i o n s o f K 4
fact must contain at least k - 1 formulae. Furthermore, a point is of m i n i m a l

w i d t h if it satisfies
mwidth'- V (Q-widthA A - ~ R - w i d t h )
QCP RCQ
A point is of minimal width if it is of width Q for some Q and no successor

can be of lesser width (unless it is of depth < k in which case it has no width).
It is easy to see that the points of minimal width together with the points of
depth < k form a generated subframe. However, this is not yet the desired
subframe. For now we have to start the same procedure as above, selecting
points of minimal span. As in the previous lemma, each point has an a t o m
As, S C {0, 1 , . . . , n - 1}. A point is of s p a n C within being of minimal width,
C C 2 '~, if it satisfies the formula
C-span "-- A(~(mwidthA X(S)) " S e C} A A ( - ~ ( m w i d t h A X(S)) " S r e}

It is easy to check that a point is of C-span within being of minimal width iff
it has exactly the successors with atoms x(S) of minimal width. Finally, the
formula
mspan :-- Ddp(< k) V V ( ( x ( S ) A e-span A A -~O9 " x(S) E e C 2 n)

D~c
defines the points of minimal span within being of minimal width. We claim
it defines the slice of points of depth k. It is easy to see that if a point is of
depth k it must be of minimal width and within that of minimal span. For
the converse, however, we prove the following.
(1:) Let x, z be of minimal width and of minimal span within being of minimal
width. Assume that x and z are of equal width, equal span and equal atom.
Then x - z.
By induction on the constitution of the formula ~ in the variables ]Pn we show
that for points x and z satisfying the assumptions of (1:), x E ~(~) iff z C ~(~).
This suffices for a proof. Now for variables this is so by the fact t h a t x and z
have the same atom. The steps for A and -~ are straightforward. Now assume
that ~ - (}7/and that x C ~(~). Let x have span Q and width C. Then there
exists a y such that x <l y and y C ~(~). Suppose that y is of depth < k. Then
(~/(y) is a conjunct of Q-span. So, z has a successor satisfying ~,(y). But only
y satisfies ~/(y). Therefore z <l y. Now assume that y is not of depth < k.
Then it has span Q and width C. Let X(S) be the atom o f y . Now x has a
successor of span Q, width E and atom x ( S ) . Then z has a successor u with
span Q, width C and atom X(S). By induction hypothesis, u E ~(~). Hence
z c -
Next we shall show that the points satisfying mspan are exactly the points
of depth k. Clearly, if a point is of depth k it satisfies mspan. So, assume t h a t
8.2. The Structure of Finitely G e n e r a t e d K 4 - F r a m e s 407
x satisfies mspan. If every successor of x is of depth < k we are done. So let

this not be case. T h e n the a t o m of x is in the span of every successor of x
which is not of d e p t h < k. Moreover, no successor of x which is not of d e p t h
< k has lesser w i d t h t h a n x. So, let x <1 y and y not of d e p t h < k. T h e n y
has the same width and the same span as x and the a t o m of x is in the span
of y, which means t h a t there exists a successor u of y not of d e p t h < k which
has the same a t o m as x. By (1:), u = x and so y <1 x. This shows t h a t x is of
d e p t h k.
The formula defining the points is of degree 4k § 2, since at most four
modal operators are stacked on top of formulae defining points of d e p t h < k.
Counting the points, we find t h a t if there are 5(n, k) points of d e p t h < k,
there can be at most 2 5 ( n ' k ) - - 1 n o n e m p t y subsets. For each subset C, there
is a (possibly empty) set of points having minimal width C. W i t h i n t h a t set,
we have counted in effect the number of points of d e p t h 0. This shows (1.).
Finally (4.). If x is not of d e p t h < k, then it has a width. If t h a t width is not
minimal, there is a successor v of minimal width, since the w i d t h is finite. As
in the proof of the previous t h e o r e m we see t h a t v has a weak successor w of
minimal span within being of minimal width, by the same argument. A weak
successor of w is a successor of x. If the width of x is minimal, t h e n argue
with x = w. This completes the proof. [3
Let us r e m a r k t h a t the b o u n d for the n u m b e r of points of d e p t h 0 is

exact. There is a frame (the frame underlying the free algebra) which has
5(n, 0) m a n y points. However, for greater d e p t h the b o u n d is too large. This
is so because not for every set S of points of d e p t h < k there exists a point
of width exactly S. For if x sees a point y C S and y <1 z t h e n x <1 z as well,
so suitable S are only those sets which are closed under successors. However,
to count their n u m b e r is more intricate and not of i m m e d i a t e interest for us.
As a first consequence we will show t h a t the u p p e r part of the lattice ~ K 4
is r a t h e r well-behaved. Recall from Section 4.8 the T h e o r e m 4.8.7 on locally
finite varieties.
THEOREM 8.2.6 (Segerberg). Let A D_ K4 | dp(< n). The variety of
A-algebras is locally finite. Consequently, A has the finite model property.
THEOREM 8.2.7 (Maksimova). An extension of K4 is of finite codimen-
sion iff it is tabular.
PROOF. The direction from right to left follows from Proposition 7.6.4.
Now assume t h a t A has finite condimension. Consider the sequence of logics
Th(~rA(n)), n C w. Clearly, A C Th(~rA(n)) for all n, and A = Th(~rA(n)).
Furthermore, T h ( ~ r h ( n + 1)) C_ T h ( ~ r h ( n ) ) . If equality holds then we have
A = T h ( ~ r h ( n ) ) . These facts together show t h a t since A has finite codimen-
sion there is an no such t h a t A = Th(~rh(n0)). Now consider the extensions
A(d) " - Th(~r~d(n0)) for d < co, where ~r~d(no) is the subframe of points of
408 8. E x t e n s i o n s of K 4
depth < d. Again, A = ~de~ A(d) and A(d + 1) _C A(d). If equality holds,
A - A(d). So there exists a do such that A - A(d0) - Th(~r<d~ This
frame is finite. Hence A is tabular. [i]
Thus, if one of the four parameters, the depth, is finite, we have more or
less good control over the situation. A logic is called p r e t a b u l a r if it is not
tabular, but all of its proper extensions are. By abstract arguments one can
show that any logic which is not tabular must be contained in a pretabular
logic. In the mid-seventies it was established by LEO ESAKIA and V. MESKHI
in [57] that ~$4 has exactly five such logics. (This fact has been proved
independently also by WOLFGANG RAUTENBERG and LARISA MAKSIMOVA.)
The proof is a real classic in the theory of K4. We prove it by playing with the
fundamental parameters of depth, fatness and width. Let H be a pretabular
logic in ~ S4. If its width, fatness and depth is finite, H is tabular. So, one
of the parameters is infinite. Suppose, then, that the depth of H is not finite.
Then as one can easily see, the n-element chains are models of that logic.
This follows from the following useful fact.
PROPOSITION 8.2.8. Let f be a noetherian Kripke-frame for S~ of depth

c~. The map x ~-~ dp(x) is a p-morphism onto (a, >}. Let f be a noetherian
Kripke-frame for G of depth a. The map x ~-+ dp(x) is a p-morphism onto
PROOF. Let }: be a Kripke-frame for S4. Call the map which sends x to
its depth ~. Clearly, if x <1 y then ~(x) > 7r(y), by definition of depth. Next,
assume that ~(x) :>/3. Then x is of depth at least/3. It follows by definition
of depth that there exists a successor of x which has depth/3. Likewise for
G. [i]
Namely, let II be of unbounded depth. By the fact that frames for H are
top-heavy, for each d < w there exists a H-frame f of depth d. Then, by the
previous theorem, (d, ___) is also a H-frame. The logic of all finite chains is
G r z . 3 . It is also the logic of the infinite chain (w, ~). It is clearly not tabular,
but each proper extension is of finite depth; since the logic is of fatness 1 and
of width 1, every extension is tabular. This concludes the case of infinite
depth. Now we may assume that the depth of H is finite. Let now the fatness
be unbounded. We distinguish two cases. C a s e 1. The depth of prefinal
clusters is unbounded. Notice that if n is a prefinal cluster, take the subframe
generated by that cluster. Then ~ --- n @ o . So, all these frames can be
mapped onto a frame of the form n --- o, called t a c k s in [57]. Let H be the
logic of the structure f~o - R0 Q o. H has depth 2 and width 2 but infinite
fatness. It is the logic of R0@o. Then II is not tabular. But every proper
extension must be of finite fatness and hence tabular. C a s e 2. The depth
of final clusters is unbounded. Then II - $5, the logic of the lonely clusters,
FIGURE 8.1. The five pretabular extensions of S4
chain tack cluster kite
i
fork
called clots. For in a refined frame with a final cluster of size n we find a
generated subframe isomorphic to the cluster n. So, H C $5. On the other
hand, $5 is not tabular, since it is not of finite fatness. Yet, every proper
extension must be.
Now we have reduced the investigation to logics of finite depth and finite
fatness. Since II is not tabular its width must be unbounded. Now, if the
depth and fatness is bounded then the number of immediate incomparable
successors of a point is also not bounded. (If it were, let H be of depth d and
fatness f, and let points have at most q immediate incomparable successors.
Then there are at most 1 + q + q2 + . . . + qd-1 clusters; each cluster has at
most f points. Hence H is tabular.) So for any n there is a refined frame ~,~
and a point x with n incomparable immediate successors. Take the subframe
~Sn generated by x. Then II is the logic of the ~Sn, as is not hard to see.
Furthermore, we may assume that ~3,~ are of fatness 1. (Otherwise, take the
set of all frames -~n resulting from r by collapsing each cluster to 1. This
set is of unbounded width, and so not tabular. Since H is pretabular, H is
410 8. E x t e n s i o n s of K 4
the logic of the ~ . ) Let gn - { x } U I n U N , ~ U Z ~ , where In is the set of

immediate successors of x of depth 0, Nn the set of immediate successors of
x not of depth 1, and Zn the set of the remaining points. C a s e 1. The set
{ ~ I n ' n C w} is unbounded. Let ~ := (k~, <~n), where
k, n "-- {x}UI n U{y}

<n -
Then bn ~- 1 G ~Di<n 1 with n - - ~ I n + 1; this frame is also called the n - f a n .

So [n is a ~ I n + l - f a n . Let p~ " g n - + kn be defined b y p ~ r { x } u i n - id,
and pn(w) - y for w C N~ U Z. This is a p-morphism. Now II is included
in the logic of all n-fans. The logic of n-fans is not tabular, but a proper
extension is, since it must be of finite width. This exhausts the first case.
C a s e 2. {~I~ 9 n C w} is bounded. Then {~Nn 9 n E w} is unbounded.
Put bn := 1(~)((~1<,~ 1 ) Q 1 , and call it the n - t o p or n - k i t e . There is a
p-morphism from ~n onto the ~Nn-kite. Namely, collapse In and z into a
single point. Hence II is contained in the logic of all kites. The latter is not
tabular. Hence II is equal to it.
THEOREM 8.2.9 (Maksimova, Esakia and Meskhi, Rautenberg). Let A be

a logic containing S4. A is pretabular iff A is one of the following five logics:
the logic of the chains, the logic of the clusters, the logics of the tacks, the
logics of the fans, the logic of the kites.
Notes on this section. The structure of finitely generated K4-algebras
and of finitely generated S4-algebras also called interior algebras has
been studied by a number of people, either directly or indirectly by means
of the canonical frame; the first to mention is KRISTER SEGERBERG [193].
For interior algebras, see WIM BLOK [22] and FABIO BELLISSIMA [4]. The
presentation here follows MARCUS KRACHT [121], who builds on KIT FINE
[66]. It can be shown that ~ G has ~0 many pretabular logics. s has 2 ~~
many pretabular logics, as was proved by WIM BLOK.
E x e r c i s e 276. Show that all pretabular logics of S4 are finitely axiomatiz-

able.
E x e r c i s e 277. Let [ be a finite, rooted, reflexive frame. Show that Th(f) is

of codimension >_ ~f. Hint. Show that for each cardinality < ~f there is a
rooted frame g with [ --- g.
E x e r c i s e 278. As before, but with [ not necessarily reflexive.
E x e r c i s e 279. Let A _D K 4 be pretabular. Show that ~A ~- 1 + w*. Hint.

Show first that the sublattice of points of finite codimension must be linear.
8.3. The Selection P r o c e d u r e 411
E x e r c i s e 280. A variety of K 4 - a l g e b r a s is locally finite iff it is of finite

depth. Hint. Show t h a t dp(~_ n) is expressible with one variable. Now look
at ~YA(1). It must be finite.
8.3. The Selection Procedure
This section will introduce one of the most important techniques in modal
logic, t h a t of the selection procedure, developed by KIT FINE ([66]) and
MICHAEL ZAKHARYASCHEV ([245]). The idea is t h a t if we have a general
frame refuting a formula ~a, we can extract a finite c o u n t e r m o d e l for ~a. We
cannot expect t h a t the new model will be based on a frame for the logic under
consideration. However, as has been noted, m a n y logics are actually closed
under this new operation, so t h a t for t h e m this will be a proof of the finite
model property. We begin by examining a special case. Suppose t h a t we have
a global model 9)t - (~,/3} and a formula ~. Let X "- sf (~a) and x C f . T h e n
there is a unique set Y C_ X such t h a t 9Yr ~ X iff X C Y. Now define
xEY xEX-Y
If co fusio ris s, writ mo( ) r th r th n is c ll d
the ~ - m o l e c u l e of y. We say t h a t y is # - m a x i m a l if it has molecule #
but no strict successor has molecule Y. Given ~, we say t h a t y is m a x i m a l
(for ~) if it is Y - m a x i m a l for some Y. If [ is n o e t h e r i a n it has no infinite
strictly ascending chain and so every point y which is not itself m a x i m a l has
a successor which is m a x i m a l for the molecule for y. Thus, given y there is a
weak successor which is m a x i m a l for the molecule of y; it is denoted by y~.
Let ~ be the subframe of m a x i m a l points.
LEMMA 8.3.1. A s s u m e ([,13, x) ~ ~a. Let g KK_[ be a subframe such that
every point in f has a weak successor in g with identical molecule, and let
x C g. Then ( g , ~ , x } ~ ~a.
PROOF. By a s s u m p t i o n there is a function y ~-~ yP such t h a t y and yP
have the same molecule in [ and yP is a weak successor of y contained in g.
We m a y assume t h a t x p - x. By induction on X C sf(~a) we show t h a t
x x
Let X - P, a variable. Then, since yP has the same molecule as y, they

satisfy the same variables of ~, irrespective of the s u b f r a m e they are in. So
the base case is treated. The steps for A and -- are straightforward. Now let
X - 0r If ([,/3, y) ~ 0 r then ([,/3, yP) ~ 0 r by definition of yR. T h e r e is
a successor z such t h a t ([,/3, z} ~ r By induction hypothesis, (~t,/3, z p) ~ r
Then, as yP q z, z p is a successor of yP, so (~t,/3, yP} ~ 0 r Conversely, assume
(~t,/3, yP) ~ (}r T h e n for some successor z, (~t,/3, z) ~ r We have t h a t z p
is a weak successor of z, so it is a successor of yP and of y. By i n d u c t i o n

hypothesis, (g,f~,z p) ~ r and so ([,/3, y) ~ 0 r This shows the first claim.
T h e second is similar. [i]
Let S be a subset of f. P u t
tS := {y'xqy} J:S := {y'x_qy}

$S := {y'yqx} ~:S := {y'ySx}
DEFINITION 8.3.2. Let (f,F) be a K 4 - f r a m e . a E F is called coj~nal if

$ a ~ a. @ is a co]inal subj~rame of ~ if it is of the form ~ M a for some
cofinal a C F.
Let ~o be given and ( [ , ~ , x ) ~ ~o. If [ is n o e t h e r i a n and g contains the
m a x i m a l points, then g is cofinal in [. For if y E g and z a weak successor of
y, t h e n z ~ is weak successor, which is in g.
We r e t u r n now to the general case. Recall the definition of a molecule.
Given ~, we have < 2 k molecules, where k "- ~sf(~). We call a m o l e c u l e also
the set of points satisfying one and the same molecule. Moreover, M ( T ) , or
simply M , is the set of all molecules with respect to ~. M forms a p a r t i t i o n of
the original frame. Let a C F. Call x c r i t i c a l for a if for no weak successor y
of x such t h a t y C a we have m o ~ ( y ) - m o ~ ( x ) . This leads to the following
definition
crit(x) :-- V /.t A --'X A E:](X -+ -~#)
ttEM
A point has m o l e c u l a r s p a n (or m - s p a n ) N if it is in the set
N-span := A O# h =0#
ttE N ttE M - N
T h e m o l e c u l a r d e p t h (or m - d e p t h ) of a point x is the length of a m a x i m a l

sequence x - x0 q Xl q . . . q xn-1 such t h a t xi+l has lesser molecular span
t h a n xi. (It follows t h e n t h a t xiJ~xi+l.) In particular, x is of m - d e p t h 0
iff it is of m i n i m a l m - s p a n . Notice t h a t x C crit(x) iff every successor of x
satisfying X has lesser m - s p a n (unless x is irreflexive). So, as can easily be
seen, the operation of taking critical points can only be nontrivially i t e r a t e d
finitely m a n y times, at most as m a n y times as there are molecules.
Now let 9 ) I - (~,/3) be a global model. We define a frame | a finite
K r i p k e - f r a m e ~(~o) and a p - m o r p h i s m rr" | -+ ~(~o) such t h a t the following
holds.
1. s(~o) is a cofinal subframe of ):, and 5(~) C_F.
2. 7r" e(~o) -~ a(~o) is the refinement map.
3. For every x E f there exists a weak successor x v C | of x such t h a t
(1.) x p has the same molecule as x;
(2.) x p has the same molecular d e p t h as x.
8.3. The Selection Procedure 413
4. For every x E s(~), x is of molecular depth d iff 7r(x) is of depth d in

3(~).
5. For every x E z(~), 7r-l(x) is definable by means of a formula of modal
degree 2 + d. dp(~), d the molecular depth of x.
The frames | and 3(~) will be defined in stages. We define Gd(~), 3d(~)
and maps 7rd" | ~ ~d(~). Here, | is the frame of points of |
of molecular depth < d; 71"d is the refinement map. ~d(~) is the set of points
of depth < d in 3(~). We will have the following facts" | is a generated
subframe of | consisting of the points x such that 7rd+l(X) is of depth
< d. Furthermore, 7rd+l r Sd(~) -- 7rd. 7rd is the refinement map. For x E z(~)
we denote by 7(x) the formula defining the set 7r-l(z). It is clear from the
previous remarks that 7(x) is independent of d.
We start with d - 1. | consists of all points of depth 0. These are
the points critical for I . Moreover, the formulae 7(x) are of degree 2. Now
we define the points of | on the basis of the points of | Let A be
the set of 7(x), x E Zd(~). Put 5* "-- V A. A point y is of m o l e c u l a r w i d t h
F C_ A if it satisfies the formula F-width defined by
r-width . - A o,,,,, A
7EF 7EA-F
As before, the notion of minimal molecular width is defined by
mwidth . - ~5" A DS*V V (r-width f A th +

r cA r~gr
Among the points of minimal molecular width we select the points of minimal
molecular span.
mspan
9-- -,6" A / V/~EM(~)# A [-1(-,#A 6*)
[ V yNC_M(~)(N-span A A O g N ( O - s p a n -+ 5*))
Finally
:= mspan A mwidth
Let now x E Sd+l(p) iff x E Sd(~) or x is of minimal molecular span and
molecular width. For x E Sd+l (~), let x have molecular span N and molecular
width F; then put
7(x) "- N - s p a n A F-width A (
This is a formula of depth kd+l "-- kd + dp(~) if ~ has depth at least 1. It is
clear that | is a generated subframe of Gd+~(p). Moreover, every point
of | has a strict successor in | As internal sets we take the sets
generated by ~(pi). Let 7rd+l be the refinement map of Gd+l(~). We show
that the image of ~rd+l is a finite frame of depth d + 1.
LEMMA 8.3.3. Suppose that x and y in Gd+I(~P) have identical width,

identical span and identical molecule in (~,~). Then in Gd(~), x E a iff
y E a for all internal sets. Hence ~d+l(x) --~d+l(Y). Consequently, 3d+1(~)
is finite.
PROOF. Every set is g e n e r a t e d by ~(p~), so we m a y reason by i n d u c t i o n
on the formula X ( ~ . For X - pi the case is clear, likewise the steps for -~ and
m
A. Now let X - 0~. Assume t h a t x E ~ ( 0 ~ ) . T h e n there is a successor u

of x satisfying ~. Two cases arise. C a s e 1. u E G d ( p ) . T h e n , as y has the
s a m e w i d t h as x, there is a point v such t h a t y <] v and v satisfies 77. C a s e
2. u ~ | T h e n u has the same width as x. It also has the s a m e s p a n as
x. T h e a t o m of u is the span of x. u is critical, t h a t is, it satisfies a molecule
t h a t no successor of layer < d can satisfy. As y is of equal span as x and of
equal m - s p a n as x (by virtue of being of equal width) there exists a point v
such t h a t y <] v and v satisfies the molecule of u. v cannot be of layer < d,
since its molecule cannot be satisfied there. Hence v is of layer d. This m e a n s
t h a t v has the same span as u, the same w i d t h as u and the s a m e a t o m as u.
By i n d u c t i o n hypothesis, v E/3(~). Hence y E ~(0r/) -- ~(X). A n d t h a t h a d
to be shown. [:]
Now we put for z E ~(T), 7 ( x ) : = 7(Y), where ~Vd+l(y) -- x. By definition

of 3(T), this is i n d e p e n d e n t of the representative. Now let x be a point of
molecular d e p t h d. By a s s u m p t i o n , if x <] y and y has lesser molecular s p a n
t h a n x, y is of molecular d e p t h < d. Moreover, x has a successor y of m o l e c u l a r
d e p t h d - 1 such t h a t the molecule of x is not satisfied at any weak successor
of y. By induction hypothesis, y has a weak successor yP in | which is
of d e p t h d - 1 in 3d(~). Clearly, x <1 yP. Therefore, 7(y p) is c o n t a i n e d in the
w i d t h of x. Consider a weak successor u of x which is of m i n i m a l w i d t h and
m i n i m a l span. Then, by choice of x, u has the same span and same molecule
as x, and 7(yP) is contained in its width. It follows t h a t u has m o l e c u l a r
d e p t h d. This shows the claims.
THEOREM 8.3.4 (Zakharyaschev, Fine). Let ~ be a formula and assume

that (~,:3, x) ~ qo. Let k "- ~sf(qo). Then there exist | and 3(qs) such that
the following holds.
1. | is a cofinal subframe of ([, G}, where C~ is a subalgebra of ~; the
of 3 (:).
2. 3(~) is finite and of depth < 2 k.
3. For every x E f there exists a x p E s(T) such that
(1.) x p is a weak successor of x;
(2.) the molecule of x in (~,~) is the same as the molecule of x p in
(| and G(qo) is called the T - e x t r a c t of (~,/3) and (3(qs),/3} as well as

3(~) the qo-reduct, x is called q u a s i - m a x i m a l if x E s ( : ) .
8.4. Refutation Patterns 415
Recall from Section 3.5 the definition of a subframe logic. It is a logic

whose class of frames is closed under taking subframes. Now call A a cofinal
s u b f r a m e logic if for every ~ ~ A and every cofinal ~ C ~ also ~ ~ A.
COROLLARY 8.3.5 (Zakharyaschev). Every cofinal subframe logic has the
finite model property.
COROLLARY 8.3.6 (Fine). Every subframe logic has the finite model prop-
erty.
We remark here that there is a simple criterion for a logic to see whether
it is a (cofinal) subframe logic.
THEOREM 8.3.7. Let A be a logic. If A has the finite model property and
the set of finite models is closed under taking (cofinal) subframes, then A is a
(cofinal) subframe logic.
The proof is simple. Let 0 be the cofinal subframe logic whose set of finite
frames equals the set of finite frames for A. Such a logic exists by assumption
on A. Since A has the finite model property, 0 - A. Hence, $4, $5, G and
G r z are subframe logics.
Exercise 281. Show that K 4 . D is a cofinal subframe logic but not a sub-
frame logic.
E x e r c i s e 282. Show that the logics of width 0, the logics of tightness 7,

the logics of fatness ~, the logics of depth 5, all are subframe logics, in all
combinations.
E x e r c i s e 283. Let A D K 4 be canonical and let the frames be determined

by a set of universal sentences. Then A is a subframe logic.
8.4. R e f u t a t i o n P a t t e r n s
This section is devoted to the so-called canonical formulae by MICHAEL
ZAKHARYASCHEV. Basically, for every formula ~ there exists a finite set
of geometrical configurations, called refutation patterns, such that a frame
refutes ~ iff it realizes one of the refutation patters. The exclusion of a
refutation pattern can be characterized by a modal axiom, and this axiom is
called a canonical formula. The refutation patterns for ~ can be constructed
from the frames underlying minimal countermodels for ~, by observing how
the selction procedure selects points and how the selected frame lies embedded
in the whole frame. For concreteness, let us take a frame ~. Let (~, ~, x) ~ - ~ .
In ~ lies cofinally embedded the ~-extract G(~). There is a contraction
~ : G(~) ~ 3(~) onto the reduct. Now let us take a look at the other points
of the frame and see how they lie with respect to G(~). More precisely, we
are only interested in the way they lie with respect to ~(~), because points
t h a t are being m a p p e d onto the same element in the reduct have the same
molecule. Call a subset V of points of ~(p) a v i e w if V - j" V. Now take a
point x e f. The set vw(x) = ~[~x A z(~)] is a view. It is called the ( ~ - ) v i e w
of z. A view Y is i n t e r n a l if there is an x E s(p) such t h a t Y -- v w ( x ) , and
e x t e r n a l if there is an z e f - s(p) such t h a t Y = v w ( x ) . Notice t h a t a
given view may be both external and internal; it may also be neiter internal
nor external. We say t h a t the frame r e a l i z e s a n e x t e r n a l ( i n t e r n a l ) v i e w
V if V is the external (internal) view of some point of f. We will see t h a t
for a given frame two factors determine whether or not a model for p can be
based on it. One is to which frames it is subreducible and the other is which
~-views it realizes. To make this precise, two more definitions are needed. We
say t h a t two points 0 - a g r e e if they satisfy the same n o n - m o d a l formulae,
and t h a t they p - a g r e e if they satisfy the same subformulae of p (iff they
have the same molecule).
PROPOSITION 8.4.1 (Agreement). Let (~,13} be a model, | the p -

extract, 3(P) the p - r e d u c t and 7r : G ( p ) ~ 3(P) the refinement map. Let
x and y have the same 3-view. If x and y O-agree, then they also p-agree.
PROOF. By induction on the subformula X it is shown t h a t for two points

x and y which 0-agree, x ~ X iff y ~ X. The only nonobvious case is X = (}r
Let x ~ 0 r Then there is a successor u of x such t h a t u ~ r Furthermore,
there exists a weak successor u p C s ( p ) of u such that u p ~ r We then have
z <]up as well. By assumption, there exists w such t h a t y <]w and ~r(w) = ~v(v).
It follows t h a t for all subformulae ~- of p, u p ~ ~- iff w ~ 7. Choose ~- = r
This implies w ~ r and thus y ~ X. [::]
PROPOSITION 8.4.2 (Homogenization). Let (~,13} be a model, x an ex-

ternal world, and Y a set of external worlds and let all points of Y U {x}
have identical view. A s s u m e that Y is an internal set. Define 7 as follows.
-y(p) :=/3(p) U Y if x C/3(p), and "y(p) := 13(p) - Y otherwise. Let y, y' E Y .
Then y and y~ p-agree in (~,'~). Moreover, for z ~ Y the molecule of z in
(~,'~} is the same as the molecule of z in (~,/3}.
PROOF. Since Y is internal, 7 is a valuation. Now, all internal points

have the same valuation as before, likewise all external points ~ Y. Now we
prove t h a t if z r Y, then for all X C sf (p) we have
The only non-obvious case is X = 0 r Let v be a successor of z such t h a t

(~, 7, v) ~ r v has a weak successor v p such t h a t (~,/3, v p} ~ r By induction
hypothesis, (~, 7, vP} ~ r Hence, (~,3, z} ~ (}r since z <:1v p. Now let the
latter be the case. Then, analogously, there is a v p such t h a t (~,/3, v p) ~ r
By induction hypothesis, (~, 7, vP) ~ r and so (~, 7, z) ~ (}r as required. [-1
8.4. R e f u t a t i o n P a t t e r n s 417
T h e last l e m m a says t h a t for external points we can m a k e the v a l u a t i o n quite

uniform, because in evaluating a formula we can 'skip' the e x t e r n a l points.
It is actually possible to derive the H o m o g e n i z a t i o n L e m m a from the first
l e m m a . Finally, we are also interested in the possibility to insert points. Let
(~,/3) be a model. We t h e n have a uniquely defined e x t r a c t | and the
refinement m a p 7r : G --~ 3. We fix all these elements. We see t h a t we can
tinker w i t h the external points quite drastically, t a k i n g t h e m away if we want,
and a d d new ones. However, w h a t we need to know w h e n we insert points is
t h a t its view (if external) m u s t already be realized in the frame by a point x.
T h e n we simply e x t e n d the valuation by giving the new point the v a l u a t i o n
of x a n d we have again a model of ~o. Thus, the s t r u c t u r e of the external
points is quite irrelevant as long as certain views are not realized. Notice
also t h a t if a view is realized t h e n we always have a witness x giving us a
v a l u a t i o n for the new point. Now go one step further. Take any frame
with a cofinal s u b f r a m e 05 a n d a p - m o r p h i s m 7r : 05 --- 3. T h e n we can define
3-views analogous to views. Namely, given a point x we p u t vw(x) = 7r[Tx]
and call it the 3 - v i e w of x. Notice t h a t views d e p e n d on 05 a n d 7r in a d d i t i o n
to 3. A n d we have the following t h e o r e m .
PROPOSITION 8.4.3. Let rift := (~,/3). Let | = | be the ~a-extract of
(~, Z) and 7r : G --, 3 the refinement map. Consider a frame 05 containing
a cofinal subframe ~ and p : ~ --~ 3, such that every external 3-view of 05
is realized in ~. Moreover, let every point of g have a weak successor in h
with identical view. Then there exists a valuation "y such that the ~a-reduct of
9I := (05, 7) is isomorphic to 3.
PROOF. By a s s u m p t i o n there is a function z ~-+ Z p : g ~ h such t h a t z p
is a weak successor of z with identical view. F u r t h e r m o r e , z p = z if z C h.
T h e v a l u a t i o n is as follows. (i) x e h. T h e n z e "?(p) iff t h e r e is a y e s(~a)
with 7r(y) = p(x) and y e /3(p). (ii) x r h. For each e x t e r n a l view Y take
an e x t e r n a l y y in ~ with identical 3-view. Yv exists by a s s u m p t i o n . T h e n let
x e "y(p) iff y y e ~(p). This is a valuation on 05. For 3'(P) is a u n i o n of sets
of the form p - l ( u ) , u C z(~a), and sets of points of indentical view. T h e s e are
all internal sets. We claim the following.
(,) if 7r(y) = p(x) t h e n mob(y) = morn(x)
(**) mo (z) = mo (z )
F r o m this it i m m e d i a t e l y follows t h a t the reducts are isomorphic. We show
(,) a n d (**) by s i m u l t a n e o u s induction on the subformulae of ~. For variables
b o t h hold by construction. T h e case of -- a n d A is clear. Now let X = 0 r
We show (,) for X. Let (~,/3, y) ~ 0 r T h e n there exists a z such t h a t
y <] z and (~,/3, z) ~ r We m a y actually assume t h a t z E s(~). T h e n there
exists a v e h such t h a t 7r(z) = p(v). Now, 7r(y) = p(x) a n d 7r(y)<~ 7r(z)
imply p(x) <~7r(z). Since p is a p - m o r p h i s m , there is a v such t h a t x <1 v and
418 8. E x t e n s i o n s o f K 4
p(v) - ~ ( z ) . By (.) for r (O,V,v) ~ r and so ((5,V,x) ~ 0 r Now let the

latter be the case. Then there exists a v such that x <~v and ((5, V, v) ~ r By
inductive hypothesis for (**), ((5, 7, vP) ~ r Now there exists a w such t h a t
y <1 w and ~(w) - p(vP), and so (~,/3, w) ~ r It follows that (~,/3, y) ~ 0 r
Now we show (**) for X - 0 r Assume that ((5, V,z p) ~ 0 r Then there
exists a y E>z p such that ((5, 7, Y) ~ r Hence ((5, ~, z) ~ 0 r since z <~z p <~y.
Now assume that ((5, V, z) ~ 0 r Then for some y ~> z, ((5, V, Y) ~ r Since
y<~yP, z<~yP. Now z and z p have the same b-view and so there is a u p
such t h a t zp <~ u p and p(uP) - p(yp). By (.) for r ((5,7, u p) ~ r Hence
The last theorem has in effect identified what the geometric condition
for ~ is. We need to know: (1.) the structure of the reducts and (2.) the
admissible external views for each reduct. Since views are upward closed sets,
that is, cones, the following definition emerges.
DEFINITION 8.4.4. Let b be a finite K r i p k e - f r a m e and 93 be a set of cones
of b. T h e n the pair ~ := (p, ~ ) is called a r e f u t a t i o n p a t t e r n . We say that
a f r a m e ~ s a t i s f i e s or r e a l i z e s ~ if there is a subframe (5 and a contraction
(5 --" b such that no external b-view is in 9J. ~ E 93 is called a c l o s e d d o m a i n
of I?. I f ~ does not realize ~ we say that it o m i t s ~.
Given ~, there exist only finitely m a n y reducts. On each reduct there exist
finitely m a n y closed domains, hence there are finitely m a n y refutation patterns
for ~. They can be calculated algorithmically. W i t h ~ given, enumerate all
models of size < 2 k, k " - ~ s f ( p ) . If necessary, reduce these models. This
enumerates all reducts. Next try inserting a new point x somewhere and
defining a valuation such that the reduct remains intact. If this is impossible,
the view of x is a closed domain.
The nonsatisfaction of a refutation p a t t e r n can be characterized axiomat-
ically. Given a refutation pattern (3, ~ / , where b is a frame based on the set
{0, 1 , . . . , n - 1} and with root 0, we take for each i a distinct variable Pi and
define
~al(3, ~ ) : = A(p, -+ ~pj: i r j) (r
A A(p~ -~ % : i ~ j) (<])
A A0~ -~ - ~ % : i ~ j) (~)
A h ( ~ v ~ ( ~ ) : ~ c ~} (~d)
~w (~):= A (0p~: i c 0)
A A(~pi:i<n)
Consider a valuation/3 : p i ~-~ ai into ~ and a world x such t h a t (~',/3, x)
Po A D+real(3, 93). Then the set Ui/3(pi) is a subframe which can be m a p p e d
onto 3. This follows from (:/:), (<]) and (r The formula evr(u) says t h a t
8.4. R e f u t a t i o n P a t t e r n s 419
a point has p-view v and is external. This is excluded by real(d, ~ ) . The

cofinality requirement need not be stated separately, for we have the following
fact.
PROPOSITION 8.4.5. Let (6,f~) be an embedding pattern. If r is a sub-

frame in ~ and r --~ 3 then r is cofinal iff the empty 3-view is not external.
Now define
:=
DEFINITION 8.4.6. A formula ~ is called a c a n o n i c a l f o r m u l a if ~a is

of the form ~/(IP) for some refutation pattern IP = (3, ~ ) .
PROPOSITION 8.4.7. Let IP be a refutation pattern. Then ~ ~ 7(IP) iff
satisfies ]P.
Now consider ~a again. We can specify all refutation p a t t e r n s for ~a. If
is an axiom, the frames satisfying the refutation p a t t e r n s are exactly those
frames which must be excluded. For they allow to define a valuation refut-
ing ~a. We have seen t h a t these p a t t e r n s can be axiomatized by canonical
formulae. Thus we have the following theorem.
THEOREM 8.4.8 (Zakharyaschev). There is an algorithm which for given

formula ~a returns a finite set of refutation patterns (3i, f~}, i < n, such that
g ~ | ~a = g ~ @ {~/(pi, ~ i ) : i < n}.
The consequences of this theorem are enormous and we will have to be

content to list a few of t h e m in the sequel. For the m o m e n t let us notice a few
details. In [245], the closed domains are defined as antichains r a t h e r t h a n
cones. T h e effect is the same, because an antichain generates a cone, and
each cone is generated by an antichain. However, there are more antichains
t h a n there are cones, so taking antichains introduces a r e d u n d a n c y here. The
formulae ~,(~, ~ ) are therefore not identical with the a - f o r m u l a e of [245]. We
will switch freely between a definition of refutation p a t t e r n via cones and via
antichains, taking whatever suits the purpose best.
Finally we have to discuss the case where external views are also internal,
t h a t is, when there is a point x E 3 such t h a t j" x = v for v C ~ . W h a t h a p p e n s
if we exclude such an external view? There are two cases. First, assume t h a t
x is reflexive. T h e n x C v, corresponding to an antichain {x}. T h e n nothing
changes if in the frame we take all points with view v as internal points,
changing their valuation into t h a t of x. Thus, forbidding such an external
view has no effect. If, however, x is irreflexive, then forbidding the view of x
to be external will have substantial effects.
PROPOSITION 8.4.9. Let (~,~} be a refutation pattern, and x a reflexive

point. Then g ~ | ~/(~, 93) = K ~ | 7(~, 9~ U ~ x).
420 8. E x t e n s i o n s of K 4
It should be emphasized that even though the formulae 7(3, r.8 tJ {Tx})
and 3'(3, ~/) are axiomatically equivalent over K4, they are not satisfied by the
same models and so not deductively equivalent. A valuation refuting the first
formula refutes the second; the converse does not necessarily hold. Namely,
to satisfy the first formula, the view l"x may not be external, while for the
second it is enough that j'x is internal, it may also be external. To prove
Proposition 8.4.9, assume that 9Y~ = (~,/3) satisfies 3'(3, ~ ) . Define a new
valuation 5 and a model fit := (~, 5) as follows. If x e /3(p) let 5(p) be the
union of/3(p) and all external points with view S x; if x r let 5(p) be
/~(p) minus the set of all external points with view J'x (check that this is an
internal set, so 5 is well-defined). In (~, 5) the view Sx is now internal. To
see that, note that the points of ff)~ which were external and had view S x
have the same molecule as x in fit, and it can be shown that they belong to
the extract of 9~. So they are internal. Nothing else changed. Therefore, the
~-reduct of 9l is the same as the ~-reduct of ff)~, and they realize the same
views except for l"x. Playing with this distinction will be helpful sometimes.
In addition to 3'(3, ~U) there is a formula 7~ ~U), which results from
3'(3, ~ ) by dropping the cofinality requirement. We then have
THEOREM 8.4.10. A logic A is a cofinal subframe logic iff for every 3'(3, ~U)
we have that from 3"(3, ~U) E A follows 3"(3, ~) C A. A logic is a subframe logic
iff 3"(3, 9J) e A implies 3"0(3, ~) e A. Hence a cofinal subframe logic can be
axiomatized by formulae of the form 7(5, 0), a subframe logic by axioms of
the form 3"0(3, ~).
PROOF. Let 3'(3, O) r A, that is, A admits the refutation pattern (3, ~ ) .
Suppose there is a a frame ~ with a cofinal subframe ~ such that ~ - - 3
respecting the external views. However, by assumption on A, ~ is itself a
frame for A and realizes the refutation pattern (3, O). Likewise for subframe
logics. E]
THEOREM 8.4.11. A logic K4 @ 3'(3, ~ ) is a splitting of K4 by 3 iff ?8 U

{$x : x <~x} contains the set of all cones of 3.
PROOF. Let ~ be a frame, ~5 C ~ a cofinal subframe such that ~ --- 3. If
| is a generated subframe, then no view is external. Hence no external view
is realized. On the other hand, let no external view be realized. Then ~ is a
generated subframe. [Z
The canonical formulae do not provide an axiomatic basis for extensions of

K4. This follows from the fact that there exists an extension of K 4 that does
not possess an independent axiomatization. In the case of (cofinal) subframe
logics the situation is different, though. Let A be a subframe logic and 3
rooted and finite. Denote by A/F 3 the smallest subframe logic containing A
not having 3 as a frame, and call A/F 3 the F i n e - s p l i t t i n g of A by 3. It
turns out that A/F 3 = A | ~(3, ~). Any subframe logic A is a FINE-splitting
K4/F G with G - {3 : 3 ~ K r p ( A ) , g rooted and finite}. Analogously, let A
be a cofinal subframe logic. The Z a k h a r y a s c h e v - s p l i t t i n g of A by 3 is the
least cofinal subframe logic containing A for which 3 is not a frame. It is
axiomatizable by A 9 7(3, ~). We write A/z 3. It follows from the fact that
the cofinal subframe logics have the finite model property that we have a
sublattice S K 4 of subframe logics and a sublattice C9"K4 of cofinal subframe
logics, that all (and only) the rooted finite frames induce splittings and the
splitting logics all have the finite model property. Under these circumstances
we conclude the following. P u t [ < F fl if g is a contractum of a subframe of [,
and [ < z ~ if 9 is a contractum of a cofinal subframe of [. Then < z C -~F and
in both cases the lattices are isomorphic to the lattice of upper sets of finite
rooted frames. Hence, on the finite frames the topology is the Alexandrov-
topology and so the lattices are actually continuous. Moreover, there is no
infinite strictly ascending chain of prime elements.
COROLLARY 8.4.12. Both S K~ and C~ K~ are continuous lattices and

have a strong basis. The natural embeddings S K~ C C~ K~ q E K~ are
continuous maps.
The continuity of the embeddings is a consequence of the fact that the

upper limits and the lower limits of chains coincide with the respective limits
in ~ K 4 . So any infinite intersection of (cofinal) subframe logics is again
a (cofinal) subframe logic. In addition, we can study the splittings of the
sublattices as induced splittings of the larger lattice. Notice namely, that if
we have a locale !~ and a point p : i~ -~ 2 such that p - l ( 0 ) is a principal ideal
and p - l ( 1 ) a principal filter, and we have a sublocale i : if2 --~ i~ such that
the embedding respects all limits, then i o p : 9)~ --~ 2 is a point with similar
properties. To put this concretely, if (p, q / i s a splitting of ~ then (pS, q$) is a
splitting of 9)I where
P~ := U(x:i(x) <_p)
q+ := Vli~:i(~)_>q>
pC is the largest if)I-logic below p, while ql" is the smallest ffJt-logic containing
q. In the present context, A 1" is the (cofinal) subframe closure, the logic
of all frames for which all (cofinal) subframes are frames for A, while A S is
the (cofinal) subframe kernel, the logic of all (cofinal) subframes of frames
for A. Figure 8.2 illustrates the situation. Take the distributive lattice to
the left. It is a sublattice of the lattice to the right, the top and b o t t o m
elements are the same. The splitting (p, q) of the larger lattice induces the
splitting (pC, q$) of the embedded lattice. The cofinal subframe logics provide
a rich class of natural logics. We present them here without proof. In the
exercises the reader is asked to supply some of the proofs, and to compare
422 8. E x t e n s i o n s of K 4
FIGURE 8.2
%"..~
i(q ~) P
q~ p$ ~176176176 o ~176
q
the F i n e - and Zakharyaschev-splittings with standard splittings. We have

K 4 . 1 := K 4 @ [3(}p -~ (}E]p, K 4 . 2 " - K 4 9 (}[-lp --+ (}[3p. K 4 n is the logic of
Kripke-frames in which there is no chain of points (xi " i < n + 1 / such t h a t
xiJlxi+l; K 4 . I n is the logic of Kripke-frames in which there is no antichain
of length n § 1.
$4 - K 4 / F I-g]
G - K4/F ~]
Grz - K 4 / F {[-~,[-&-]}
K4.1 - K4/z
K4.2 - K a / z F1
K4.3 - K a / F F1
K4~ -- K a / F L~
K4.In - K a / F Fn+l
The notation is as follows. Ln is the set of frames whose reflexive closure is
(n,->/ and Fn+l is the set of frames whose reflexive closure is the frame ~n+l
of Figure 8.3. The exercises below shed some light on the splittings of ~ K 4
as first outlined in[169].
Notes on this section. The idea to study lattices of subframe logics rather
than lattices of logics is a major theme of WOLTER [244]. This has turned
FIGURE 8.3. 0n+l
n+l
out to be a more promising line of research than the investigation of entire

lattices of logics, since the latter are too complex if the logics is not so strong
(e. g. in the case of $4.3, KS).
E x e r c i s e 284. Show that S4 - K4/F[-J]. Show also that $4 is obtainable

by splitting two frames from K4.
E x e r c i s e 285. Show that G r z is obtained by splitting two frames from $4,

and that G r z - S 4 / F [ - ~ . Verify that $4.3 - S4/F02 and t h a t S4.3 is
obtained by splitting two frames from S4.
E x e r c i s e 286. Show that G - K 4 / F ~ . N a m e a set U of frames such that

G - K 4 / z U . Show that there is no set N of frames such that G - K4/N.
E x e r c i s e 287. Show that K 4 . 1 and K 4 . 2 are cofinal subframe logics but

not subframe logics.
E x e r c i s e 288. Show that there exist extensions of K 4 which are not cofinal
subframe logics. Can you name one?
E x e r c i s e 289. (KRACHT [126].) Show that any splitting G/N for finite N
has the finite model property.
E x e r c i s e 290. Let ~t be a frame, ~ and ~ be sets of cones with !U C_ !~U.

Show that K 4 9 7(3, ~ ) C_ K 4 9 7(~, ~ ) .
E x e r c i s e 291. Obviously, a subframe logic is a cofinal subframe logic. Given

7 ~ (~t, ~U) compute explicitly a finite set C of cofinal subframe axioms such that
K 4 G 7 ~ (3, ~ ) - K 4 | C.
E x e r c i s e 292. (ROBERT BULL [34].) Show that every extension of $4.3

has the finite model property. Hint. Show that every extension is a cofinal
subframe logic.
424 8. E x t e n s i o n s of K 4
8.5. E m b e d d a b i l i t y P a t t e r n s a n d the Elementarity of Logics

Refutation patterns are second-order conditions; some of them cannot
be reduced to first-order conditions. An example is the logic G defined by
the exclusion of ~ as a subframe. Moreover, it is generally not decidable
whether a logic is first-order or not (see [44]). However, as have noted KIT
FINE and MICHAEL ZAKHARYASCHEV, for cofinal subframe logics there is a
close connection between the geometric conditions expressed by the refutation
patterns and the elementarity of the conditions they impose on frames. (See
[66] and [246].) We approach this circle of themes by defining a different
condition associated with a refutation pattern P = (6, ~U), called embedding
pattern or embeddability pattern.
DEFINITION 8.5.1. Let 3 be a finite frame and 93 a set of cones. Then
the gripke-frame [ satisfies the (coj~nal) embedding p a t t e r n F := (3, ~U) if
there is no (cofinal) embedding of ~ into ~ such that no view in f~ is external in
the frame generated by the image of 3 in [. We write e(3, ~ ) for the property
of satisfying the cofinal embedding pattern and e~ 93) for the property of
satisfying the embedding pattern P.
Clearly, e(~, ~ ) and eo(~, ~ ) are first-order. Moreover, they are describ-
able by a restricted V3-sentence. Namely, let ;3 - (z, <1) be given. Let
z - {0, 1 , . . . , n - 1} and 0 be the root. Now ~e(3, ~ ) is the condition
( Ai<j<n xi ~ x~
A Ai~j xi <~xj
A A ~ j xi r xj
x0)... =0) A (vy y < v y -
A (vy y -
Here, the first three formulae describe the fact that we have an embedding,
the fourth that this embedding is cofinal and the last that it satisfies the
closed domain condition (no view from ~U is external). As we can see, parts
of the conditions are positive; it is not clear that these conditions define at all
a modal class of frames. However, notice the following.
LEMMA 8.5.2. e~ O) is equivalent to a positive universal ~ f -sentence.
PROOF. Let f = {i : i < n}. Let A(n) be the set of atomic fomulae of
the formz~ " zj, i , j < n or xi <~xj, i , j < n. Let C C _ A ( n ) ; t h e n l e t r ( C )
be the frame (re, <~c} where f c is the factorization of f by the equivalence
relation generated by (i, j}, where xi - xj C C. Then put [i] <~c [j] iff there
is a g C [i] and an h C [j] such that zg <~Xh. We call C a b a r r i e r if for no
8.5. Embeddability Patterns and the Elementarity of Logics 425
D D C, r(D) is isomorphic to [. Let B be the set of barriers. Now let
(w0)(vx, x0)(To_, V A c)
CEB
Then g ~ a([) iff [ is not embeddable into g. a([) is positive, universal and
restricted. [:2
Similarly -~e~ ~3) is ~V. Now, just as 7(3, ~l/) need not be elementary,
c(3, ~l/) need not be modal. We will investigate situations in which the embed-
dability conditions are modal and situations where the refutation patterns are
elementary. In b o t h cases, we may either consider special sets of patterns or
special classes of frames. Let us first consider the general question of reduc-
ing the refutation patterns to first-order conditions in special classes, namely
noetherian frames.
THEOREM 8.5.3. I n noetherian frames of width 0 every (cofinal) refuta-

tion pattern is a conjunction of finitely many (cofinal) embedding patterns.
Moreover, a (cofinal) embedding pattern corresponds to a V3-sentence.
PROOF. Let [ be a noetherian frame, and (3, ~ ) be a refutation pattern.
We may assume that no closed domain is of the form T x, where x is reflexive.
In that case, if there is a cofinal subframe 9 of [ and a p - m o r p h i s m 7r 9 9 --~ 3
such that the closed domains are not external views, then let [ be the frame
consisting of all points maximal in 7r-1. 7r F k 9 t~ --~ 3 and no closed domain is
an external view. Moreover, no cluster of t~ is larger than the largest cluster
of ~. Now let 3 have t~ many points. Let the largest cluster have size c. Then
has at most g. c. 0 many points. Take the embedding p a t t e r n based on (~, ~ )
where ~3 consists of all closed domains V such that 7r[V] E ~ . Then ~ realizes
that embedding pattern. So, let B be the set of all embedding patterns ({~,~ )
based on frames with _< g. c. 0 points such that there exists a 7r 9 ~ --~ 3 and
consists of all closed domains V with 7r[V] C ~ . Then ~ realizes the refutation
p a t t e r n P iff it realizes some embedding pattern of B. [:2
Another i m p o r t a n t case is when we have no closed domains. Then the

refutation p a t t e r n defines a cofinal subframe logic. Even though the following
arguments extend to cofinal subframe logics we focus on subframe logics. Take
a finite frame [ with root z. Let Z :- C(z). Call a c l u s t e r s e q u e n c e of
a sequence (Ci : i < k) of clusters of [ such that Co = Z and Ci+l is an
immediate successor cluster of Ci. It is clear that the length of a cluster
sequence is bounded by the depth of [. Let u([) be the set of pairs (E, x) such
that E is a cluster sequence and x a member of the last cluster of E. P u t
( E , x ) 9 (F,y) if either (a) E = F and x <~ y or (b) E is a proper prefix of
F. Now put u ( ~ ) : = (u([), 9 and call it the c l u s t e r u n r a v e l l i n g . The map
d : (E, x) ~ x is a p-morphism. Now let ~ be any frame such that there exist
p - m o r p h i s m s ~ : u ( [ ) --~ 0 and p : 0 --~ ~:. T h e n 0 is called a d i s e n t a n g l i n g

of [. (Disentanglings are called d e s c e n d a n t s in [66].)
LEMMA 8.5.4. Let ~ be a finite frame and 9 a noetherian frame. Suppose

that 9 is subreducible to [. Then there is a disentangling ~ of ~ which is
embddable into 9.
PROOF. A s s u m e there is a s u b f r a m e ~ of 9 a n d a p - m o r p h i s m ~ : ~ --~ [.
is also n o e t h e r i a n . It suffices to show t h a t there is a disentangling of [ which
is e m b e d d a b l e into ~. We select a subset of h in the following way. We define
a m a p 7 : u([) -+ g by i n d u c t i o n on the l e n g t h of the cluster sequence. Let
E = (Z). Since O is n o e t h e r i a n , the set U : = 71"-l[z] contains a cluster W
isomorphic to Z. Namely, let w E U be such t h a t w <l v a n d v C U implies
v <1 w. T h e n two cases arise. C a s e 1. Z is degenerate. T h e n U is a disjoint
union of d e g e n e r a t e clusters, a n d we m a y pick any w C U. C a s e 2. Z is
n o n d e g e n e r a t e . T h e n a <~-maximal cluster contains at least as m a n y points
as Z. So, we pick a subset of size ~Z from such cluster. We m a y in fact pick
such a subset W on which 7r is injective. Finally, p u t T((E, x ) ) : = y, w h e r e
y C W a n d 7r(y) = x. Now let ~- be defined on all pairs ( E , x ) where E is
a cluster sequence of length < d. Let F := (E, C). Let x be an e l e m e n t of
the last cluster of E and y C C. P u t w := 7 ( ( E , x}). Select a cluster D from
the set 7r-1[C] M Sw. Such a cluster exists since 7r is a p - m o r p h i s m a n d ~ is
n o e t h e r i a n . Moreover, we let D be such t h a t 7r is injective on D. Now p u t
~-({F, y } ) : = v where v C D and 7r(v) = y. This defines w. Let i) be t h e image
of u([) u n d e r ~-. T h e n ~- induces a m a p from u([) to 0. This is easily seen to
be a p - m o r p h i s m . Finally, the restriction of 7r to 0 is a p - m o r p h i s m as well.
So, ~ is a disentangling of ~ and a s u b f r a m e of 9. 71
It is clear t h a t if 9 is cofinally subreducible to [ t h e n some disentangling of

is cofinally e m b e d d a b l e into 9. It follows
LEMMA 8.5.5. Let P = (3, Z) be a refutation pattern without closed do-
mains. Then there exists a finite set E of embedding patterns without closed
domains such that a noetherian frame realizes P iff it realizes some member
orE.
THEOREM 8.5.6. Every (cofinal) subframe logic is A - e l e m e n t a r y in the
class of noetherian frames.
This s t a r t is promising. So we have to s t u d y w h a t h a p p e n s if we lift t h e
condition t h a t the frames be n o e t h e r i a n . Here we face a problem. Now it is t h e
case t h a t a frame is subreducible to a finite frame [ w i t h o u t t h e r e being a finite
s u b f r a m e ~ which can be m a p p e d p - m o r p h i c a l l y to [. This s i t u a t i o n however
arises exclusively with clusters. A case in point is the chain w < - (w, < / . This
chain can be m a p p e d onto [ ~ , but ~ c a n n o t be e m b e d d e d in it; also t h e r e
does not exist any finite s u b f r a m e of ~ < which can be m a p p e d p - m o r p h i c a l l y
8.5. E m b e d d a b i l i t y P a t t e r n s and the E l e m e n t a r i t y of Logics 427
onto [-~]. A n o t h e r case is the chain w -< - (w, _<). It can can be m a p p e d onto
any cluster n. However, if n > 1, there exists no finite s u b f r a m e which can be
m a p p e d p - m o r p h i c a l l y onto n; moreover, n is not e m b e d d a b l e into w -<. These
two examples play a pivotal role here. Let k < " - (k, <) and k-< := (k, _<) for
every k C w. Let [ be a frame, and C a cluster. T h e n the subframe based
on C is totally local, and so the replacement of C by any other frame fl is
well-defined. ~ is an i m m e d i a t e v a r i a n t of fl if there exists a cluster C in
[ such t h a t either (a) C is improper and fl - ~[k</C] for some k C w, or
(b) C is proper and ~ - [[k < / C ] for some k E w. To s u m m a r i z e , i m m e d i a t e
variants are obtained from frames by replacing a n o n d e g e n e r a t e cluster by a
finite irreflexive chain or replacing a proper cluster by a finite reflexive chain.
A v a r i a n t of [ is obtained by iterating the process of forming i m m e d i a t e
variants. Now say t h a t a set S of frames is q u a s i - c l o s e d u n d e r v a r i a n t s
if for each ~ C S and each cluster C of ~, some variant ~[k < / C ] (~[k</C]) is
subreducible to a m e m b e r of S. A set T has the f i n i t e e m b e d d i n g p r o p e r t y
if a frame [ belongs to T exactly when each finite subframe belongs to T. Say
t h a t a logic has the f i n i t e e m b e d d i n g p r o p e r t y if its class of frames has
the finite e m b e d d i n g property.
THEOREM 8.5.7 (Fine). Let S be a set of finite frames. Then the set
K r p ( g 4 / g S ) has the finite embedding property iff S is quasi-closed under
variants. If either condition holds, the set of frames not subreducible to a
member of S equals the set of frames into which no disentangling of a member
of S is embeddable.
PROOF. Suppose t h a t S is not quasi-closed under variants. T h e n there
exists a ~ c S such t h a t no i m m e d i a t e variant is subreducible to a m e m b e r of
S. We m a y assume t h a t f is a frame of minimal size with this property. Take
a cluster C of [ and consider the frame fl " - [[w</C] if C is i m p r o p e r and
g := [[k < / C ] if C is proper. By assumption, no finite subframe is subreducible
to any m e m b e r of S. (Here we need the a s s u m p t i o n t h a t [ is of m i n i m a l size.)
However, ~ is subreducible to [. Hence, fl r K r p ( K 4 / F S ) while every finite
subframe ~ of fl is contained in t h a t class. So, Krp(K4/FS) does not have
the finite e m b e d d i n g property. Now assume t h a t S is quasi-closed under
variants. Let ~ be a frame which is subreducible to some [ C S. So, there
exists a subframe ~ and a p - m o r p h i s m ~ 9 ~ -~ [. Suppose there exists an
i m p r o p e r cluster C of [ such t h a t ~ - 1 [C] contains a cofinal subset of the form
w <. There is a variant 3 " - [[ k < / C ] which is subreducible to a m e m b e r of S.
fl is subreducible to 3, as can be shown. Hence fl fails the subframe condition
for S. Similarly if C is proper. It follows t h a t if 9 is subreducible to [, some
disentangling of [ is e m b e d d a b l e into ft. [i]
THEOREM 8.5.8 (Fine). For a subframe logic A D_ K4 the following are

equivalent.
428 8. E x t e n s i o n s of K 4
1. Krp(A) is definable by a set of universal, positive ~/-sentences.

2. A is r-persistent.
3. A is A-elementary.
4. A is canonical.
5. A is compact.
6. A has the finite embedding property.
PROOF. The implication from (1.) to (2.) follows from Theorem 5.4.11.
The implication from (2.) to (3.) is a consequence of Theorem 5.7.8. If
(3.) holds, then (4.) follows from Theorem 5.7.11 and the fact that sub-
frame logics are complete. A canonical logic is compact. This follows from
Proposition 3.2.7. Now, suppose that (5.) holds. We establish (6.). Take an
infinite frame ~ such that every finite subframe satisfies A. Let A(f) be the
diagram. This diagram is finitely satisfiable, since each finite subset involves
only a finite number of points. So, a finite subset describes a finite subframe,
which is a frame for A. So the whole diagram is satisfiable. Thus, A has
the finite embedding property. Now assume (6.). Then, let F be the class
of rooted finite frames not being A-frames. Consider the set of sentences
e~ O), 3 C F. This is a set of universal :~f-sentences. This set can be
turned into a set of universal positive ~f-sentences since F is inversely closed
under contractions. C]
COROLLARY 8.5.9 (Fine). A subframe logic K,~/FS is canonical iff S is
quasi-closed under variants.
So, S4 is canonical, since it is obtained by excluding [-~, while G is not
canonical, being obtained by excluding [-~. The latter set is not quasi-closed
under variants; no variants are subreducible to [ ~ . Likewise, S5 is canonical
while G r z is not. This concludes the discussion of subframe logics. Now in
special circumstances, other refutation patterns yield elementary logics. A
special case are logics of finite width. If we consider only noetherian frames,
then elementarity is guaranteed. But we do not always need to assume that
the frame is noetherian. For example if we consider splittings of irreflexive
frames in logics of finite width. In this special case the assumption that the
frame is noetherian can be dropped.
THEOREM 8.5.10. Each splitting condition by a finite irreflexive frame is
elementary in the class of frames of width 0.
In [12] it was claimed that this holds without assuming finite width. To
see that the condition of finite width is essential look at the following frame.
Let p be a prime number and Zp := {q} LJ {k*: k C p} U { k . : k C p}. We put
{(q,k*/. k < ;}
<p "- u { ( q , k , ) . k < p}
U {(k*,n,i'k,n<p, andn-korn-k+l (modp)}
8.5. Embeddability Patterns and the Elementarity of Logics 429
FIGURE 8.4. The frame 33
w v w
Finally, 3p "-- (Zp, <Jp). It is not hard to see that 3p cannot be contracted
onto 33 unless p - 3. Let U be an ultrafilter over w - {0, 1, 2} containing all
cofinite sets. P u t 3~ : - 1--[u3pn, where pn is the n t h prime number. We show
that 3~ can be m a p p e d onto 33 (in fact onto any 3p). Thus the condition
'does not contain a generated subframe which can be m a p p e d onto 33' is not
elementary. 3~ is of depth 3, and every point of depth 1 sees exactly two
points of depth 0, every point of depth 0 is seen by exactly two points of
depth 1 (these properties are elementary, hence preserved by passing to an
ultrapoduct). Let H be the set of points of depth 0 in 3~. Call a subset
C c_ H a c y c l e if it is closed under the operations
~+~,+m'-((vi).'vi-ui+k (modpi))
for each k C Z. Pick from each cycle C a representative r(C). Let Z be the
set of all cycles. Now let
Ho :- U c ~ z { r ( c ) + 3 k - k e z}
HI :- U ~ z { , ' ( c ) + 3k + 1. k e z}
H2 := Uc~z{r(C) + 3k + 2" k e Z}
Let
K0 .- {~*~,eH0}
K1 " - - {u* " ~, C H1 }
K2 "- {~*'~, CH2}
Define ~ : 3~ -~ 33 as follows. ~ ( ~ ) : = q, ~ ( K i ) : = i* and 7r(Hi):= i,. We
leave it to the reader to check that this is a p-morphism.
Notes on this section. FRANK WOLTER has shown in [242] t h a t the mini-
mal tense extension of a cofinal subframe logic has the finite model property iff
it is A-elementary. Nevertheless, the minimal tense extension of any finitely
axiomatizable cofinal subframe logic is decidable.
E x e r c i s e 293. Show that an extension of G is canonical iff it is of finite

depth. Show that an extension of G r z is canonical iff it is of finite depth.
430 8. E x t e n s i o n s o f K 4
E x e r c i s e 294. K 4 . 1 = K 4 G Vl()p --~ {)lip is the cofinal subframe logic

K 4 / z [ - ~ . Show that it is not canonical but b~0-canonical.
E x e r c i s e 295. (Continuing the previous exercise.) Say that a set of frames is

q u a s i - c l o s e d u n d e r n o n f i n a l v a r i a n t s if for each ~ C S and each nonfinal
cluster C some variant [[k </C] ([[k <-/C]) is cofinally subreducible to a mem-
ber of S. Show that if S is quasi-closed under nonfinal variants, the cofinal
subframe logic K 4 / z S is N0-canonical.
8.6. L o g i c s of F i n i t e W i d t h I
Before we begin with logics of finite width proper, we will think somewhat
more about how models can be made simple. A natural consequence will be
that logics of finite width satisfy one of the central properties t h a t allow to
make models countable. Though we have in principle settled the question
of characterizing the frames determined by an axiom, we will introduce here
another line of thinking, which considers the question of how to make a model
small without distorting it too much. As we have seen, there are ways to
obtain very small models, based on frames whose cardinality depends only
on ~. However, the price we have paid is that we do not know whether this
operation ends in a model for the logic under consideration. We have seen
that the logic must be a cofinal subframe logic if this is case generally. If not,
we are however not without tools. Still, we can make the model well-behaved
in certain respects. Some special cases are provided when we consider a local
subframe ~ of a Kripke-frame [. If ~ - - o, then we can contract ~ in [, by
locality, and so we can actually contract [ as soon as it contains no quasi-
maximal points. For we know that if in a given model ~ contains no quasi-
maximal points, then all points realize the same external view; contracting
them to a single point leaves the refutation pattern that is being realized
intact. On the other hand, since it is a contraction, it preserves the fact t h a t
the frame is a frame for the logic.
PROPOSITION 8.6.1. Let (~,/3, x) ~ ~ and 05 a local subframe containing

no quasi-maximal points for ~, and ~ a net on 05. Let ~ be the unique
extension on ~ and ~ : ~ -+ Y~ the corresponding p - m o r p h i s m . Then there
exists a valuation 7 and a y such that (~, V, Y) ~ ~.
PROOF. Let (~,/3, x) ~ ~ and ~ a net on 05, 05 a local subframe of

containing no quasi-maximal points. Denote by ~ the extension of ~. For
z e g Let [z] : - {y : z ~ y}. By assumption this is an internal set, and
by the locality of 05, all members of [z] have the same view. For each [z]
let r([z]) e [z]. Now put ~(y) : - /3(y) if y r g and ~(y) : - /3(r([y])) if
y C g. By the Homogenization Theorem, the molecule of y in (~,/3) is the
8.6. Logics of Finite W i d t h I 431
same as the molecule of y in (~,5). It follows that (~,5, x) ~ ~. The p -

morphism ~ corresponding to ~ is admissible for 5. Denote by V the valuation
V(~(x)) " - ~(x). Then (~, V, ~(x)) ~ ~. [i]
For example, we may eliminate points x such that for some y, xQy and
x ~ z implies y ~ z for all z, on condition that x is not quasi-maximal, and
we may eliminate all non quasi-maximal points in a proper cluster. Let ~" be
a frame, N C_ f a subset (not necessarily internal). We say that ~ results
from d r o p p i n g N if (5 - ~ • (f - N). So, dropping points is another way
of saying that we are passing to a subframe. In analogy to [125] we say that
dropping N is safe if Th((5) _~ Th(~). We say that dropping N is s u p e r s a f e
if N can be safely dropped even when ~ is a totally local subframe of another
frame (B. Evidently, if dropping a set is supersafe, it is also safe. The converse
need not hold. A first case of (super)safe dropping is dropping points from a
cluster. A more general case is dropping points which can be thought of as
being collapsed by a p-morphism. For example, in the kite o @ ( o | o ) G o we
can drop one (or both) of the intermediate clusters.
Let us return to the problem of reducing a model of ~. Clearly, not all
points are available for dropping; either because they are quasi-maximal or
because they are required to keep the "structure" intact. But if a point can see
another point of identical molecule, then it may be dropped. More generally,
N is called ~ - c o v e r e d if every point in N has a successor outside of N with
identical molecule.
LEMMA 8.6.2. Let (~,/3} be a model and N a ~-covered subset. Then
every point in f - g has the same molecule in ( ~ - N,/3) as it has in (~, ~).
PROOF. By Lemma 8.3.1. [::]
A note of caution is in order. In principle, the process of dropping can
be iterated as many times as one likes. However, repeating it infinitely often
requires careful argumentation. Just consider the frame (w, <). Dropping a
single point is safe - - the resulting frame is still isomorphic to (w, <). Hence,
we might consider dropping 0, then 1, then 2, etc. In the limit, however, the
whole structure is gone! There is an additional difference between dropping
from a frame and dropping from a model. When we drop from a frame we
are only interested in the fact that it is a frame for the logic. If on the other
hand we want to drop from a model for ~ we want to ensure in addition that
after dropping we still have a model for ~. Now let in our previous case be
- p and/3(p). Each finite set is p-covered, but their union is not.
A second critical case is provided by b~ " - o(~ (~iC~o o. Consider the
formula ~ = p A (}~p, and let/3(p) := {x0} where (x0} is the initial cluster.
Then dropping a single point from the model, even a finite set, is permissible
since every point is ~-covered. Again, if we drop all points, we no longer have
a model for ~p.
Evidently, if we want to avoid the dangers just described, we need to make

sure t h a t there are no infinite ascending chains and no infinite antichains.
DEFINITION 8.6.3. A frame has the f i n i t e a n t i c h a i n p r o p e r t y (lap) if

every antichain is finite.
DEFINITION 8.6.4 (Fine). A frame is said to have the f i n i t e cover pro-
p e r t y (fcp) if for every set N there is a finite C such that N C_ TC.
PROPOSITION 8.6.5. I f a frame is noetherian and has the finite antichain
property it also has the finite cover property.
PROOF. Now assume t h a t ~ is b o t h n o e t h e r i a n and has the finite antichain
property. Let S C_ f be a set. Let m a x S be the set of points w i t h o u t strict
successors in S. We have S C T max S. For if x - x0 is a point, and has a strict
successor in S, then xl be such a successor. If xl has a strict successor in S, let
x2 be such a successor. By the ascending chain condition, this process must
stop; and so we get a successor in S w i t h o u t a strict successor in S. m a x S is a
subframe which is of d e p t h 1, thus a direct sum of clusters, max S - (~icI Ci.
Pick from each cluster Cj a representative xj. T h e n D " - {x~ 9 i E I} is an
antichain, and so it must be finite. Moreover, TD D_ m a x S and so ~:D _D S,
by transitivity. [::]
The R0-kite o @ ( ( ~ i c ~ ~ o ) ( ~ o is an example of a frame with the finite cover

p r o p e r t y but without the finite antichain property. It is quite i m p o r t a n t to
have a geometrical intuition as to what points are eliminable from a frame.
A useful case is this one.
DEFINITION 8.6.6. A cluster is s l i m if it has type ~ or 1, that is, if it

contains a single point. A frame is s l i m if every cluster is slim. A frame
is a l m o s t s l i m if (1.) all clusters are finite and (2.) all but finitely many
clusters are slim.
THEOREM 8.6.7. Suppose that A D_ K g is complete with respect to noe-
therian frames. Then it is complete with respect to almost slim noetherian
frames.
T h e proof is straightforward. Recall for t h a t purpose the fact t h a t in
n o e t h e r i a n frames of b o u n d e d width all refutation p a t t e r n s can be replaced
by suitable embedding patterns. Consider a subset N of f such t h a t every
cofinal refutation p a t t e r n which can be realized on ~ at all can be realized
avoiding N . T h e n N can safely be dropped. For if ~ fails at x in [, x r N ,
there is a cofinal embedding p a t t e r n (for T) which can be realized in [. By
a s s u m p t i o n on N , the same p a t t e r n can be realized with no points in N .
T h e n N may add some external views, but it is harmless to drop them. Now
consider for the c o n v e r s e - t h a t T fails at x in [ - N . T h e n a cofinal
e m b e d d a b i l i t y p a t t e r n for ~ can be realized in [ - N . Now consider w h a t

h a p p e n s when N is added. T h e n there is the possibility t h a t some external
views are actually added. If N is initial, t h a t is, if [ - N is a g e n e r a t e d
subframe in [, t h e n these external views are harmless. However, N need not
be initial, and therefore a stronger notion is called for.
DEFINITION 8.6.8. Let [ be a frame. A set N C_ f a v o i d s all configu-

r a t i o n s if for all cofinal subframes m of ~ there is an isomorphism &9 m ~ ~,
where ~ is a cofinal subframe of [ such that ~ A N - 0 and for any external
m-view V which is realized in [ the ~ - v i e w c[V] is realized in [ - N .
PROPOSITION 8.6.9. Let ~ be noetherian and let N avoid all configurations

in [. Then dropping N is supersafe.
PROOF. Obviously, dropping a set which avoids all configurations is safe.

So it is enough to show t h a t it is safe when [ is e m b e d d e d as a totally local
subframe. So, let [ be a local subframe in ~ and let m be a cofinal subframe
of ~. T h e n n := m A f is a subframe of [. T h e n there exists an isomorphic
copy ~ in [ such t h a t all external n-views (for [) are realized for ~ in [. Let
9- ( m - n)t2 ~. We must ensure t h a t ~ is cofinal. This is evidently so if
[ is not itself cofinal. For t h e n every point in [ has a strict successor outside
of [, so m - n is cofinal, and thus ~ is as well. However, if [ is cofinal, n is
cofinal in [ and so is ~ by construction, and so ~ is cofinal in g. Fix a bijection
m ~-+ ~ and extend it to a bijection n ~-~ ~. Modulo this bijection, a - v i e w s
are t r a n s l a t e d into ~-views and back. Now let a view be realized in fl, say by
x. Two cases arise. First, x r f. T h e n the n-view of x is n or 0 , depending
on where x is s i t u a t e d with respect to f. If it is -- n, t h e n the ~ - v i e w is ~,
and if it is O t h e n the ~-view is O as well. Thus, the ~ - v i e w and the m - v i e w
of x correspond. (Recall t h a t f is assumed to be totally local.) Next, assume
t h a t x E f. T h e n by a s s u m p t i o n on N, there is an ~ realizing the ~ - v i e w in
- N corresponding to the n-view of x. Now, the m - n - v i e w of x and the
m - n - v i e w of ~ concide. For if x <] y and y r [ then also ~ <~y, and conversely,
by the fact t h a t f is totally local in ~. So, the ~ - v i e w of ~ corresponds to the
m - v i e w of x. A n d t h a t had to be shown. D
After all these theoretical results we will finally prove some useful, concrete
results.
THEOREM 8.6.10. (1.) Let o ~ ~ be a S~-frame of depth w + 1. If ~ is

of finite tightness then the initial cluster can be supersafely dropped. (2.) Let
9 Q [ be a G - f l a m e of depth w + 1. If f is of finite tightness then the initial
cluster can supersafely be dropped.
PROOF. Show that in the cases given the cluster avoids all geometrical
configurations. [:]
These are generic cases; one should not be mislead into easy generalizations,
see the exercises. However, as a guideline, if we have a frame [ of d e p t h w
which looks completely regular in shape, then in [ Q [ the first occurrence of
can be dropped. We will meet situations like this later on.
THEOREM 8.6.11 (Fine). Let A be a logic of finite width. Then the weak
canonical frames are noetherian and have the finite cover property.
PROOF. Let A be of width 0. Recall the structure theory for finitely

generated K 4 algebras. We have seen there how the points of layer a can
be constructed on the basis t h a t we have constructed the points of d e p t h
< a. However, in t h a t construction, a was finite. Here we will show t h a t
the procedure extends to all ordinals a. Let us assume then t h a t we have
constructed all points of d e p t h < a. We show t h a t each point sees a point of
minimal width in the set of points of d e p t h < a. For let x = x0 be a point
not of d e p t h < a. Suppose t h a t there is a chain <xi : i C w) such t h a t for each
i < w there is a wi such t h a t xi <~wi but Xi+l ~ wi. We cannot have Wi+l <~wi.
Otherwise Xi+l <l Wi+l <~wi which was excluded. So the sequence <wi : i C w)
is non-descending. Since each wi is incomparable with at most 0 points in the
chain, there must be an ascending subchain. Contradiction. Consequently,
there exists no infinite ascending chain <xi : i C w) with decreasing width, and
we must have a successor for x of minimal width. W i t h i n the set of points of
minimal width we choose again the set of points of minimal atomic span, and
this is the desired set of points of d e p t h a. Since the frame can be e n u m e r a t e d
with the ordinals, every point will eventually be assigned a depth. So there
are no ascending chains, and every point has a depth.
Now consider an arbitrary set N of points, not necessarily internal. For
each point x in N there is a weak successor of minimal d e p t h in N . T h e set
of these points will be called max N . max N is a sum of clusters and there are
at most 0 clusters. Pick a representative from each cluster. This is a cover
for N . Hence N has a cover of size < 0. []
Let ~ = ([, F) be a K 4 - f r a m e . Call a point x e l i m i n a b l e if for all a such

t h a t x C a there exists a strong successor y such t h a t y C a. Call ~ s e p a r a b l e
if there are no eliminable points.
THEOREM 8.6.12 (Fine). Every logic containing K g is complete with re-

spect to separable frames.
PROOF. Let ~ = <[,F) be a descriptive frame. Now drop from f all
eliminable points. Denote the resulting set by f ~ and the resulting K r i p k e -
frame by [~ ( f ~ is not an internal set.) We show t h a t the algebra induced by
F on [~ is isomorphic to the original one, and so their logics coincide. First,
the m a p a ~ a A f ~ is a boolean h o m o m o r p h i s m . Next, we have to show t h a t
it respects O. So let x C f ~ and x C Sd. T h e n x has a successor y C d; y has
a n o n e l i m i n a b l e weak successor y~ C d. For the set d contains an ascending

chain (xi : i C a) such t h a t Xi+l ~ xi which is cofinal in d. S u p p o s e this chain
is strictly ascending, t h a t is, xi<~Xi+l for all i < w. Consider the set U of all
sets containing almost all points from this chain. T h e n U is c o n t a i n e d in an
ultrafilter U*. T h e n U* C a, and moreover, U* is final in d. C o n t r a d i c t i o n .
T h e n x <1 y~ and so x E 0 ( d M f ~ T h e converse is easy. [:3
COROLLARY 8.6.13. Let A be of finite width and ~ a finitely generated,

refined, separable A-frame. Then ~ is atomic.
PROOF. Let x be a point. By separability there is a set d such t h a t no
strong successor of x is in d. C a s e 1. x ~ x. T h e n p u t e := d M i - d . This set
contains x a n d is an antichain. Consequently, it is finite. By refinedness, {x}
is an internal set. C a s e 2. x <~ x. Let N := {y : x ~ y} a n d let C := maxN,
the set of points y such t h a t y C N and if yJ~z t h e n z r N . For each z E C
let cz be a set such t h a t x E I - cz b u t x C cz (this exists by refinement a n d
the fact t h a t x r C) a n d b~ a set such t h a t z C b~ b u t for no strong successor
y, y C b z . P u t
nl- ,.zcC>
Let y C e. We claim t h a t x <1 y <~ x. First, if y C N t h e n t h e r e is a z C C such
thaty<~z. Ify-zthenyCC. Hencey~-by, soy~e;ifyr
for some z and so y ~ I - cz, from which y ~ e. Hence, e is disjoint w i t h N .
It follows t h a t for y E e, x ~ y cannot hold. So assume now t h a t x <1 y. If
y ~ x t h e n y r d and so y r e. Hence also y <~ x. It follows t h a t e is a subset
of the cluster of x. Since the algebra of sets is finitely g e n e r a t e d , e is finite.
T h e frame is refined, a n d so the set {x} is easily shown to be internal. [:3
THEOREM 8.6.14 (Fine). Every logic of finite width is Ro-canonical. Eve-

ry logic of finite width is complete with respect to countable frames with the
finite cover property, without ascending chains and of depth < ~, where ~ is
the smallest uncountable ordinal.
PROOF. Let ~ be a weak canonical frame for A. T h e n ~ is n o e t h e r i a n
a n d has the finite cover property. It is also atomic. Take a r e f u t a t i o n p a t t e r n
( r , ~ ) . We can assume t h a t there are no sets of the form S x in ~ for a
reflexive x. S u p p o s e t h a t (r, ~ ) can be realized in ~. T h e n t h e r e exists a
cofinal s u b f r a m e m a n d a p - m o r p h i s m ~ : m ~ r satisfying the closed d o m a i n
condition. Let n consist of the sets maxTr-l(x) for all x C r. ~ [ n : n --~ r.
T h e s u b f r a m e n is cofinal a n d satisfies the closed d o m a i n conditions for (r, ~}.
T h e n there are only finitely m a n y points in ~ - l ( x ) . We have j u s t seen t h a t
these sets are internal in a separable frame. Hence a d d i n g new sets does not
change the logic. This shows t h a t the logics are weakly canonical. Since
there are only c o u n t a b l y m a n y formulae in the language (the restriction to
finitely m a n y g e n e r a t o r s plays no role here), there can be only c o u n t a b l y m a n y
436 8. E x t e n s i o n s o f K 4
internal sets of the form {x}, hence only countably many points. The smallest
uncountable ordinal is ~, so if a frame is countable, it is of depth < ~. []
So all logics of finite width have Kuznetsov-index < b~0. To get a more
fine-grained view of the matter, let us define the o r d i n a l K u z n e t s o v - i n d e x
OKz(A) of a logic complete with respect to noetherian frames as follows.
OKz(A) is the supremum of all ordinal numbers a such that there is a formula
~ A and a A-frame of depth a refuting ~ but no A-frame of depth < a
exists refuting T. It is shown in [125] that for every a < w 2 there is a logic
A with OKz(A) - a. This can be extended up to w~. Whether the ordinal
Kuznetsov-index can be greater than w~ is unknown. We conjecture t h a t this
is false for logics of finite width.
We may cash out a useful result from the previous proof. Call a set a
an i n t e r v a l if it is of the form [x,y] := {z 9 x _~ z _~ y}, (x] := {y 9 y _~ x},
]x) := {y" xJ~y} or of the form {x}. Call a set s i m p l e if it is a finite union of
intervals. It is a matter of direct verification that in a noetherian frame with
the finite antichain property the simple sets are closed under all operations
and are the least such set containing all singleton sets.
THEOREM 8.6.15. Let A be a logic of finite width. Then A is complete

with respect to countable frames ~ of finite width which are atomic, separable
and such that the internal sets are the simple sets.
Theorem 8.6.14 deserves special attention. We have shown t h a t all logics
of finite width are complete. Moreover, they are complete with respect to
countable frames. We will state this explicitly once again, but we will be
more specific about the structure of the frames for these logics. Take a subset
of the form
s,~ - {y" d p ( y ) - wn +/30 for some/3o < w}
s~ is the disjoint sum of at most 0 many sets Fi which are G-indecomposable.
Such sets are called g a l a x i e s . We say a galaxy in s~ has d e p t h n, and a
point in that galaxy has g a l a c t i c depth n. The l o c a l depth of a point is just
the unique ~0 < w such that dp(x) - w . n + ~o. So, a point is determined
both by its galactic depth and its local depth. A logic has g a l a c t i c f i n i t e
m o d e l p r o p e r t y if it is complete with respect to frames of finite galactic
depth. In case of finite width this means that there are only finitely many
galaxies. Many notions are now generalized to galaxies, such as being linear.
Moreover, we can in some sense consider the frame as a frame of galaxies.
Namely, let gal(~) be the set of galaxies of ~. P u t F <]u A if for every x C F
there is a y E A such that x <~ y. Finally, let
U([) := (gal(~), <~)
U(~) is called the f r a m e of g a l a x i e s of ~. It is not necessarily a c o n t r a c t u m
of f.
PROPOSITION 8.6.16. Let ~ be a noetherian frame. Then U([) has fat-

ness 1.
PROOF. Let F <]u A <~ F. Let x C F. Then there exists a y ~ A such that
x <~y. There exist a no, nl and ~o, ~ such that the depth of x is who + ~o and
the depth of y is wn~ + ~ . Furthermore, nl < no or ~ _</30. From A <~v F
it follows that no = n~. Hence, F and A are not disconnected subsets of s~o,
and so they are identical. [3
We put U2([) := U(U([)). There is a different construction which runs

as follows. Let a h y p e r g a l a x y of depth n be a maximal @-indecomposable
subset of the set of points of depth w2n § w~l § ~0, for some/30, ~1. For two
hypergalaxies F2 and A2 put F2 < ~ A2 if for all x C F2 and all y C A2 we
have x <] y. Now let gal2([) be the set of hypergalaxies of [ and
U2(~) := (gal2([), <~>
It seems plausible that U2(~) ~- U2([). However, this is false. We leave a proof
of that as an exercise and indicate only why this is in fact not to be expected.
For note that for two hypergalaxies G, ID we have G <~2 I[} iff for all x E G and
all y C D, x <] y. On the other hand, let [G] be the set of galaxies contained
in G, and [ID] the set of galaxies contained in D. In U(U(~)), [((3] <] [D] iff for
all F E [((3] there exists a A E [D] such that for all z C F some y E A exists
such that x <~ y. The latter is clearly a stronger condition than the previous
one. However, there is an important case where the two coincide.
PROPOSITION 8.6.17. Let [ be a noetherian frame of finite tightness.

1. If F is not initial in U([), it is infinite.
2. U([) is galactically linear.
3. (f) u (f).
PROOF. Let F be noninitial. Then there exists an immediate predecessor
A of F. Suppose that there does not exist an infinite F' such that F ~
F' and F' r F. Then A, F are part of the same galaxy. Contradiction.
Hence there exists an infinite immediate successor F' of A. Suppose F' is
incomparable with F. Then the frame is not of finite tightness. So, F is the
only immediate successor of A. Hence U([) is galactically linear and of fatness
1, by Proposition 8.6.16. Moreover, take hypergalaxies G and I[}. Let G < ~ D.
Let F and A be galaxies a n d F C_ G, A C_ D. T h e n F < ~ A or A < ~ F , by
galactic linearity. Suppose now that G # D. Then F<~A. Suppose G - D.
Then for every x C G there is a y C G such that x <] x. Let A the galaxy of
least depth in G. Then A <~u A. Furthermore, for every F C_ G, F <lu A, by
galactic linearity. This shows that G <l~ D. [3
438 8. E x t e n s i o n s of K 4
This approach can be generalized as follows. Let ~ be an ordinal number.

For every ordinal ~ put
s(~,n) := {x" dp(x) - w~ . n + ~ , ~ < w ~}
Call a @-indecomposable subset F of s(~, ~) a ~ - h y p e r g a l a x y of depth ~.
Thus, g a l a x i e s are 1 - h y p e r g a l a x i e s and hypergalaxies are 2 - h y p e r g a l a -
xies. We let Gal~ (f) be the set of ~-hypergalaxies, and F <~ A iff for all x C F
and y e A we have x <~y. Finally, U~(f):= (Gal,(f), <~). We mention also an
interesting particular case, ~ = 0. It turns out that the 0-hypergalaxies are
exactly the clusters, and C <~0 D iff C <~ D. We call U0(f) the s k e l e t o n of f.
LEMMA 8.6.18. Let f be a noetherian frame. Then there is a p-morphism
f
The proof is very simple and is omitted. Notice that for ~ > 0 this need
not hold.
There is no restriction on ~. In principle, it is possible to define the notion
of an w - h y p e r g a l a x y . We believe, however, that logics of finite width are
complete with respect to frames of depth < w~. That is to say, we conjecture
that the ordinal Kuznetsov-index of logics of finite width is < w~. If that
is so, the need of considering w-hypergalaxies does not arise. At present,
however, this is only speculation. We will show that the conjecture holds for
extensions of S4 of finite width and finite tightness, and we will see later that
it also holds for all logics of finite width and finite tightness.
THEOREM 8.6.19. Every extension of S~ of finite width and finite tight-
ness has the galactic finite model property. Equivalently, it is complete with
respect to frames of hypergalactic depth 1.
PROOF. Let f be a frame for A, and let (f,/3, x / ~ ~. Consider the ~ -
extract. It is a cofinal subframe of finite depth hence finite. Hence only finitely
many galaxies contain (quasi-)maximal points. The p-morphism contracting
such a galaxy to a single reflexive point is locally admissible. So, we can
reduce such galaxies. If f had depth w./31 +/30, it now has depth/31. If after
that step we still have infinitely many galaxies left, i. e. if k - 1 > 1, we
iterate the procedure, infinitely often if necessary. After completion we have
reached a frame of finite galactic depth on which a model for ~ is based. [::]
E x e r c i s e 296. Give an example of a frame which has the finite cover pro-
perty but is not noetherian. Hint. There is a linear frame with this property.
E x e r c i s e 297. Show that all logics containing $4 of tightness 1 have the

E x e r c i s e 298. Show that all extensions of S4/k2 have the finite model pro-
perty. Hint. Only the 0-slice may contain a proper antichain.
8.7. Logics of Finite W i d t h II 439
E x e r c i s e 299. Let [ be of depth w. Show that when o Q [ is local but not

totally local, o may not be safely dropped.
E x e r c i s e 300. Show that in general for a frame ~ of depth w, o in o(~f

cannot be safely dropped. Likewise for 9 Q f.
E x e r c i s e 301. Show that in general, if ~ is an Sd-frame of depth w, the

initial cluster of n Q [ may not be safely dropped if n > 1.
E x e r c i s e 302. Show that no nonempty set in a finite flame avoids all config-
urations. Show that, nevertheless, there is safe dropping from finite models.
8.7. L o g i c s of F i n i t e Width II
Consider a logic containing K 4 . 3 . We know that it is complete with re-

spect to linear noetherian frames. Moreover, we know that we can assume
these frames to be almost slim. Let us see how we can simplify the struc-
ture of frames even more. Let (~,~) be a A-model for ~, f noetherian and
linear. Consider a segment F without maximal points. Then ~ ~ fl G F Q [ 2 .
C a s e 1. F ~- F ' Q n Q F " . Then there is a p - m o r p h i s m F -~ o G F " . By
Proposition 8.6.1, there is a model for ~ on the frame
f~ @ o @ r" @ f2
C a s e 2. F contains no reflexive point. Then O --- a~ a an ordinal. However,
if a > w it can be shown that all points of depth >_ w in a ~ can be supersafely
dropped. It follows from this consideration that logics containing K 4 . 3 are
complete with respect to frames of the following form
fk |174 | ... @~o~ @f ~
where each [i is finite and linear. It follows that every extension of $4.3 has
the finite model property.
THEOREM 8.7.1 (Bull, Fine). Every extension of $ 4 . 3 has the finite mo-
del property and is finitely axiomatizable. Hence there are only countably
many such extensions, and all are decidable.
PROOF. We know already that all extensions have the finite model prop-
erty. An extension is characterized by a set of canonical axioms ~/(m, ~U), where
m is linear. Since all closed domains of the form N'x, we can dispense with
the closed domains entirely. For we have $4.3 @ ~/(m, ~U) - s4.3 9 7(m,
(Hence all extensions of $4.3 are cofinal subframe logics, from which it fol-
lows once again that they have the finite model property.) We have to show
that every extension is finitely axiomatizable. We m ~-~ n iff n -* m iff m is
subreducible to n. Consider the partial order ~ on chains of clusters defined
440 8. E x t e n s i o n s o f K 4
by n -< m iff m is a subframe of a. This ordering can be construed as chains-

over-(w, >). The latter is a well-partial order (see [135]) and so is then the
order -~. Now consider [ Q m _E 9 Q n iff ~ -~ [ and m _> n. Then this actu-
ally defines the order of being-a-cofinal-subframe-of, and thus the converse of
being-reducible-to, the one we are interested in. Now this order is a product
of two well partial orders, and hence itself a well partial order. So there are
no infinite antichains, and hence every logic is finitely axiomatizable. [:]
This more or less finishes the case of $4.3. If we drop reflexivity, things
get a bit more awkward. First of all, the frames o Q c ~ ( n ) where r is an
irreflexive chain of length n, form an infinite antichain.
PROPOSITION 8.7.2. There are 2 ~~ many logics in ~K~.3.
Call a frame a l m o s t i r r e f l e x i v e if all but finitely many clusters consist
of a single irreflexive point.
THEOREM 8.7.3. Every extension of K ~ . 3 has the galactic finite model
property. Moreover, it is complete with respect to finite chains of almost
irreflexive galaxies.
COROLLARY 8.7.4. G.3 has the finite model property. It is weakly canon-
ical but not canonical. Every proper extension of G.3 is tabular.
PROOF. The first statement actually follows from the fact that G . 3 is
a subframe logic, but can be shown also by showing that every finite G . 3 -
configuration can be realized on a finite chain. A proper extension must
therefore have one of the finite chains not among its models. But then almost
all of them are not among the models. The logic is then the logic of a finite
chain, hence tabular. G.3 is weakly canonical, being of finite width. But it
is not canonical. Consider namely the following set
Each finite subset is satisfiable on a finite frame, so the set is consistent. A

Kripke-frame underlying a model for this set must have a strictly ascending
chain, so cannot be a G.3-frame. D
PROPOSITION 8.7.5. The logic of the frame o ~ w ~ is not finitely axiom-

atizable and fails to have the finite model property. T h ( o Q w ~ is the lower
cover of G. 3 in ~K4.3. Hence every proper extension is finitely axiomatizable
and has the finite model property.
PROOF. Add to K 4 . 3 the axiom St(< 1) and the subframe axiom saying
that a reflexive cluster must be strictly initial, that is, not properly preceded
by any other point. Call this logic Ref. Now add all splitting axioms for
o Q c O ( n ) , n C w. The resulting logic is not finitely axiomatizable, because
this axiomatization is independent. Moreover, the only infinite frames for
this logic are frames of the form a ~ or o G a ~ with a an ordinal. Now,

similar reasoning as in the Corollary 8.7.4 shows that for an infinite ordinal
a we have T h ( o @ a ~ - Th(o@w~ Thus we have axiomatized the logic
of Th(o@w~ Any proper extension does not have this frame among its
models, hence must be a logic of irreflexive chains. Consequently, every proper
extension contains G.3. [::]
Finally, we turn to the question of decidability. Recall from Section 2.6 that a
logic which is finitely axiomatizable is decidable if it is complete with respect
to an enumerable set of effective algebras. So, we have to show that for finitely
axiomatizable logics there exists such a set of algebras. We know that logics of
finite width are complete with respect to frames ([, F} where [ is a noetherian
frame of finite width and finite fatness and F is the algebra of simple sets. For
the purpose of the next definition, a computable function from Z -+ p(w),
where Z is a possibly infinite set, is a function that yields a finite set for each
z C Z and which can be computed.
DEFINITION 8.7.6. Let ~ -- (f, F} be a frame. Call ~ s i m p l e if it is of
finite width and finite fatness and F is the set of simple sets. Call ~ e f f e c t i v e
if f - w, ( A ) 'x <1 y ' is decidable for all x, y e w, (B) ~y) - 0 ' is decidable
for all y e w, and (C) there are computable functions a, c, e" w -+ p(w), and
d , p , q " w 2 -+ ~9(w) such that (1.) a(x) is an antichain with x ~ z for all
z e a(x), such that x ~ y implies that there is a z e a(x) with y <1 z (2.) e(x)
is the cluster containing x, (3.) c(x) is an antichain such that y Q x iff y has a
weak successor in c(x), (~.) d(x, y) is an antichain such that z <~x and z r y
iff u <~ z for some u C d(x, y), (5.) p(x, y) is an atichain such that for all z:
z <~w for some w C p ( x , x') iff z <~x and z <~y, and (6.) q(x, y) is an atichain
such that .for all z: w ~_ z for some w E q(z, x') iff x <~ z and y <~ z.
We remark here that this definition of effectiveness is only useful for noe-
therian frames of finite width and finite tightness. For other classes of frames,
other definitions have to be found.
THEOREM 8.7.7. Let ~ be simple. I f ~ is effective, the algebra ~+ is
effective.
PROOF. First of all, F is countable and there is a computable bijection
between the set of finite sets of intervals and w. (Basically, an interval is a
set [x, y] hence a pair of natural numbers, a set (x], a set Ix) or a set {x}.
Hence a set in F can be represented by a sequence (W, X, ]I, Z} where W is a
finite sequence of pairs of natural numbers and X, Y and Z finite sequences
of natural numbers. It is a standard fact that there is a computable bijection
between w and such quadruples.) We have to show how the operations can
be computed. Union is clear. Next intersection. It is easy to see t h a t since
Ix, y] - Ix)M (y], it is enough to show that we can compute the intersections
of two open intervals. By definition of p and q we have
[~)nly)- U [~)
~6q(x,y)
and
(z] N (y] -- U (w]
wep(x,y)
Now 0. If x q x then 0{x} - (x]. If x ~ x then 0 { x } - Uyec(x)(y].
0(x] - 0{x} and O[y,x] - 0{x}. Therefore the case 0 ]x) remains. Suppose
first t h a t Ix) is e m p t y (this is decidable). T h e n 0 ] x ) - O. So, suppose
t h a t ]x) r O. Let Y " - {y G a ( x ) " ] y ) r ;~}. By assumption (A) and the
c o m p u t a b i l i t y of a, Y is computable. We have
~,Ix) - (H u ]~) u U (Y]

y6Y
For let z C 0]x). Then there is a y such t h a t x q y and z <]y. If y - x, y C (x].

If x <1 z then z C]x). Assume therefore x 14 z. Then there exists a y C a ( x )
such t h a t z q a(x). Moreover, y must have a successor (and this must be a
point in ]x), by definition of a ( x ) ) . Hence y e Y. The converse inclusion is
established similarly. Finally, we treat the complement. This is by far the
most involved case. First the singletons. P u t L := f ( x ) - {x}.
-(~} - ] ~ ) u U {y} u U (y]

yEL yEa(x)
For if y ~= x then either (a) x <~ y <] x, or (b) x <1 y ~ x or (c) x ~ y. In case
(a), y e L, in case (b) y e]x). In case (c) there is a z e a ( x ) such t h a t y _ z.
So y c (z]. Since x is not contained in the right h a n d side, equality is shown.
Next the sets (x]. P u t D := { d ( y , x ) : y e a(x)} and A := a ( x ) - { x } . It is
checked t h a t
-(~] - U [z,y] u Ix) u U ]y)
z6D,y6A y6A
The sets - ] x ) are c o m p u t e d as follows.
-]~) - e(~)u U (z]

yea(x)
For if it is not the case that x Q y then either (a) x <~y q x or (b) x r y. In case
(a) y C e(x) in case (b) y e (z] for some z e a ( x ) . Finally the sets - I x , y].
u e - [ x , y] iff either x r u or u ~ y iff u e - ( y ] or u e - [ x ) . We can c o m p u t e
- ( y ] , so we need to know how to compute - [ x ) . It is easily seen t h a t
-I,)- yca(~)
U
Finally, we want to show that for two unions of intervals b, c, it is decidable

whether 'b -- c'. This is equivalent to the decidability of the emptiness of a
given union of intervals. This in turn is equivalent to the problem to decide
whether an interval is empty. (a) (x] is never empty. (b) {x} is never empty.
(c) Ix, y] is empty iff x ~- y and x r y. This is decidable by (A). (d) Whether
Ix) is empty is decidable by (B).
DEFINITION 8.7.8. Let ~ = (w, <~,F) be simple. ~ is called superej~fec-

tire if (a) it is effective and (b) there is a computable function j from the
set of embedding patterns into p(w) such that for every embedding pattern
c = c(r, ~3) the set j(e) is a finite set of points such that whenever ~ is realiz-
able in ~ there is an embedding p: r -+ j(e) realizing e in ~.
THEOREM 8.7.9. Suppose that ~ is supereffective. Then Th ~ is decidable.
PROOF. It is enough to show that for every embedding p a t t e r n it is decid-
able whether or not it is realizable. Take e = e(r, ~3). Then whether or not
is realizable in ~ can be checked by trying all embeddings p : r --+ j (e). There
are only finitely many of them. Hence it is enough to show t h a t for every
embedding p it is decidable whether p satisfies all closed domain conditions.
So take a closed domain u C ~ . We need to decide whether 0 is an external
view, that is, whether there is a point x not in p[r] such t h a t p[u] is the set of
points p(y) seen by x. For that we need to check whether the set b is empty,
where
b := -pk] n A *{p(y)} n A
yco y~,
b can be computed, since the frame is effective. Whether b is empty is decid-
able. [-1
THEOaEM 8.7.10. Let A be a logic of finite width. Suppose that A is

finitely axiomatizable and complete with respect to a recursively enumerable
class of frames which are supereffective. Then A is decidable.
PROOF. Let ~ be a recursively enumerable class of supereffective frames
such that A is ~-complete. Since A is finitely axiomatizable it is recursively
enumerable. Let A - K 4 | ~. It is decidable for ~ C ~ whether ~ ~ ~. Hence
the class of ~-frames for A is recursively enumerable. Let it be ~ . Hence A is
complete with respect to a recursively enumerable class of effective algebras,
and so co-recursively enumerable. [-1
K 4 . 3 is complete with respect to finite chains of the form

@gk_., @...
where each g~ is finite. It is not hard to see that these frames are supereffective.
However, this is not all we have to show. For we need to be able to determine
whether or not a given Kripke-frame [ is a frame for a logic. It remains to show
444 8. E x t e n s i o n s o f K 4
FIGURE 8.5
0 0 0 0
0/0 0 0
t h a t we can decide for such frames and a given refutation p a t t e r n whether it

is satisfiable in the frame. This is not hard to see. For let ~(~, !U) be given;
let ~ - k. For cofinal embeddability, we need to check only final segements
of w of depth < k. So, (~, !U) is satisfiable in this frame iff it is satisfiable
in a finite subframe which can be computed from the original frame and the
refutation pattern. The latter problem is now clearly decidable because we
have to check only finitely many cases. Now given an extension of K 4 . 3 by
means of finitely many canonical formulae, we can enumerate the frames of
the above form, because we can decide whether the refutation patterns are
satisfiable. Now let a (canonical) formula ~ be given. At the same time t h a t
we are checking these frames for whether they satisfy the canonical axioms
for the logic we can also check whether or not ~ can be refuted. Again, this
is decidable.
THEOREM 8.7.11 (Alekseev ~ Zakharyaschev). Every finitely axiomatiz-
able extension of K $ . 3 is decidable.
This theorem is actually the first general theorem on decidability of modal
logics without assuming the finite model property.
Logics of width 2 play a significant role in the lattice of extensions of
S4. We will concentrate on such extensions here. We begin by showing t h a t
S4.I2 - S4.wd(_< 2) has an exceptional status with respect to logics of finite
width in that it has a splitting representation.
THEOREM 8.7.12. S~.wd(< 2) is a finite splitting of S~ and hence of K~.
PROOF. Consider the frames in Figure 8.5. We will show t h a t there is

a finite set R of finite frames of width 3 such that for all refined frames
for S4 of width 2 there exists an ~ E R such that some generated subframe
g of [ is contractible to ~. Suppose that b3 - o ( ~ ( o 9 o 9 o) is a subframe
of a refined S4-frame ~. Then we have four points x, Y0, Yl and Y3 such
that x sees Yi for a l l i < 3 and yi<~yj i f f i - j. There are sets bi, i < 3,
such t h a t yi E bj i f f j - i . Put Z ' - - 0 b 0 N - 0 b l A - 0 b 2 . If Z is empty,
the first frame is cofinally embeddable. Now suppose that Z is not empty.
FIGURE 8.6
q23
Y2 }o Y2
q12
Z is a definable subset and the map contracting Z to a single point is a p -

morphism; for Z is successor closed. Thus without loss of generality we may
assume Z - {z}. Let B0 " - Cb0 N - r N-r B1 " - - r A-r and
B2 : - - r 1 6 2 ACb2. Then Yi C Bi. The map sending Bi to a single point
for each i < 3 is a p-morphism. Hence we may also assume that Bi - {yi}.
Now four cases arise. C a s e 1. z is incomparable with all Yi. Collapse z and
y2 into a single point. This is a p - m o r p h i s m and we now have an antichain of
size 3 in which each point is of depth. (First picture in Figure 8.5.) C a s e 2.
z is comparable with exactly one of the yi. Then without loss of generality
y2 <~z. (Second picture in Figure 8.5.) C a s e 3. z is comparable with exactly
two of the Yi. Then without loss of generality yl'~Z and Y2<~z. (Third picture
in Figure 8.5.) C a s e 4. All three Yi are comparable with z. Then yiQz for
all i < 3. (Last picture of Figure 8.5.)
The set R is produced as follows. Take a frame m from Figure 8.5. It
consists of an antichain Y " - {Y0, yl, y2 }, a root x and (with one exception)
also a point z of depth 0. I f i < j < 3 then let qij be a new point. Let
mo " - m U {qol, q02, q12} and let x~qij and q i j <~ Y k iff k - - i or k -- j. Then
m is a subframe of too. There exist 5 subframes which contain m and which
are contained in m0. R is the set of all such frames. (See Figure 8.6.)
We show that ~ contains a generated subframe which can be m a p p e d
onto some member of ~. We may assume that x generates ~ (otherwise take
the subframe generated by x). Let i < j < 3. Choose k such t h a t (1, 2, 3} -
{i, j, k}. Define sets Cij "- Cbk A -Obj A -r and X := Cb0 N Cbl N 0b2. The
sets Cij (if nonempty) can be mapped onto a single point. After t h a t X can
be m a p p e d onto a single point. After this process we have a frame of R.
There are a number of important frames which will play a fundamental role in
the theory of finite width. They exhibit a certain regular p a t t e r n connected
with their tightness. The simplest of them are shown in Figure 8.7.
FIGURE 8.7
~ O ~0 ~0 ~0 ~0 ~0 ~0
The first of the three galaxies is called photonic, the second leptonic and
the third mesonic. Their indecomposable generated subframes are the ele-
m e n t a r y particles from which the galaxies are composed. So, there is only
one photon, namely o, and there are two leptons, o and o @ o. There is an
infinite series of mesons. So the mesonic galaxy is itself a meson. It is immedi-
ately checked t h a t p~ is of tightness 0, A~ of tightness 1 and #~ of tightness 2.
In a similar fashion we can create 'heavier' particles. It is perhaps intuitively
clear how we will proceed, but let us introduce some notions t h a t will help
us with more complex structures as well. Let [ be a noetherian K r i p k e - f r a m e
of width 2 and d e p t h / 3 . Then there exist sequences c = (ca : a < f~) and
= (d~ : a </3} such that (1.) ca and d~ are clusters of [ of d e p t h a, (2.)
ca <1 ca, and d~ <1 d~, for all a > a ' , and (3.) every point is contained in some
ca or d~. Notice t h a t it is not required t h a t ca and d~ are distinct. We call
c and ~ the two s p i n e s of ~. The division of [ into spines is a r b i t r a r y but the
results are independent of it.
Let us restrict our attention to a single galaxy for the moment. Such
a galaxy is fixed in its structure by two things: (1.) the cardinality of the
clusters ca and d~, (2a.) the m a x i m a l local depth of a cluster in the ~-chain
seen by ca, (2b.) the maximal local depth of a cluster in the c-chain seen
by d~. Let us call the i n d e x of x C ca (or of ca itself for t h a t m a t t e r ) the
order type of S d~ - Sx and the index of y C d~ the inverse order type of
J'c~ - S y , which is the same as its depth as a frame. This is therefore an
ordinal number. This cumbersome definition takes care of the case where
there is no finite index. (This can happen! Think of two parallel chains of
galaxies.) The proof of the next proposition is left as an exercise.
PROPOSITION 8.7.13. The following holds.

1. ind(z) > O.
2. <
3. If y immediately precedes x then ind(x) <_ ind(y) < ind(x) + 1.

4. If y immediately precedes x and ind(y) - ind(x) + 1 then y is non-
branching.
LEMMA 8.7.14. Extensions of S4 of width 2 are complete with respect to
noetherian frames of bounded index.
PROOF. Consider a formula ~ rejected on a ( n o e t h e r i a n ) frame. We can
safely drop a n o n b r a n c h i n g point if it is n o n m a x i m a l . Thus, as there are at
m o s t ~sf(~) m a x i m a l points in a single spine, the index of a point is b o u n d e d
by ~sf (~).
To have a b o u n d e d index is in this case the same as being of b o u n d e d tightness.

It is now easily d e d u c e d t h a t extensions of S4 of w i d t h 2 are galactically
linear. Moreover, we know t h a t frames can be a s s u m e d to be a l m o s t thin.
Call a frame [~ (k, e ) - h a d r o n i e if there is a p a r t i t i o n into two spines such
t h a t every point of the first spine has index k, and every point of the second
spine has index g. Call a frame [~ almost (k, g ) - h a d r o n i c if there is a p -
m o r p h i s m 7r "0 -~ ~(~o such t h a t card(Tr-l(x)) - 1 for almost all x, and ~ is
(k, g ) - h a d r o n i c .
THEOREM 8.7.15. The set Y3 of reflexive frames of width 2 and finite

galactic depth in which all galaxies are almost hadronic is recursively enu-
merable. Moreover, all members of;3 are supereffective.
PROOF. We show t h a t the set of almost hadronic frames is e n u m e r a b l e
a n d t h a t these frames are supereffective. T h e general case is r a t h e r cumber-
some a n d not revealing. A slice is uniquely characterized by a q u a d r u p l e of
n a t u r a l n u m b e r s T -- (i, a, j, b), where i is the index of a point of the first
spine, a the size of the cluster of t h a t point, j the index of a point of the
second spine, a n d b the size of its cluster. Hence a reflexive frame of w i d t h
2 is uniquely characterized by a sequence (Tn "n e a}, Tn -- (in, an,jn, bn),
where a < w + 1. A frame is almost hadronic if there is a no < a , such
t h a t for all n > no, ~-no -- Tn -- (k, 1,g, 1). (Or, by switching the spines,
almost all types are - (g, 1,k, 1).) T h u s we need three things: k, t~, no a n d
the sequence T - (~-i " i < n0}. This is a finite set. W h a t needs to be shown
is t h a t given these four p a r a m e t e r s we can c o m p u t e the relations a n d opera-
tions of the frame. (This shows t h a t the frame is effective.) For t h a t , points
can be r e p r e s e n t e d as triples (i,n,p) where i = 1 or i - 2; n < w; p < an
if i - 1 a n d p < bn otherwise. For any triple it can be decided w h e t h e r
it is a m e m b e r of the frame (k,t~,n0,T}. N e x t it can be decided w h e t h e r
x := (i, n, p) <~y := (i', n', p') as follows. If i - i', x <~y iff n _ n'. If i - 1 and
i' - 2 t h e n x <~ y iff n' _< n - in, where in is given by 7n for n < no a n d by k
for n >_ no. If i - 2 and i' - 1 t h e n x <~ y iff n' < n - jn. F u r t h e r , w h e t h e r
]x) is e m p t y is equivalent to n - 0 a n d so decidable. If x - (1, n , p ) t h e n
c(x) = {(1, n,q} " q <_ a,~}; if x = (2, n,p) then c(x) = {(2, n,q) " q <_ bn}.
a(x) can be computed as follows. Let x - (1, n,p). If in - 0 then a(x) = {x}
otherwise a(x) - {x, (2, n - jn + 1, 0} } (if such a point exists) and a(x) - {x}
otherwise. Similarly for x - (2, n,p}. Similarly the existence of the other
functions is proved. This shows that the frames are all effective. Now we
show t h a t they are supereffective. Let e(~, ~ ) be an embedding p a t t e r n and
let it be realizable in ~. We show that it is realizable in the set of all points
of depth __< no + k . ~r if k _> ~ and < no + t~. ~r otherwise. W i t h o u t loss of
generality we assume that k >__e. For let q " r --+ f be an embedding. Take
W the set of all x C r such that q(r) is of depth > no and for all y such t h a t
xJ~y, the depth o f q ( y ) is_< q ( x ) - k . Let x b e a m e m b e r o f T . Then the
cluster of x is degenerate. Then let q' be defined by q'(y)"- q(y) for y ~ W
and q' (x) := (i, n - 1, 0) for x - (i, n, p) C W. We shall show that q' is an em-
bedding realizing e(r, ~ ) . Suppose that x <~y and x, y E W. Then q(x) <~q(y)
depends only on the difference between the depths of q(x) and q(y) (and the
spine of them) and so q' (x) <1q' (y) iff q(x) <~q(y). Or x, y r W and q(x) <~q(y)
iff q'(x) <~q'(y), since q'(x) - q(x) and q'(y) - q(y). Or x C W and y r W.
Then q(x) <~q(y), since the depth of q(x) exceeds that of q(y) by more t h a n
k. Then the difference between q'(x) and q'(y) is _> k and so q'(x)<~ q'(y).
Similarly it is shown that an external view is realized by q iff it is realized by
q'. Consequently, as long as W r O this operation is applicable and reduces
the m a x i m u m depths of points q(x), x C r, by 1. The procedure ends when
W = O. Then if x immediately precedes y, the depth of q(x) minus the depth
of q(y) is at most k. O
THEOREM 8.7.16. Extensions of S~ of width 2 are complete with respect

to frames with almost hadronic galaxies. Moreover, each finitely axiomatizable
extension of S~ of width 2 is decidable.
PROOF. Let A be of width 2. Since A is galactically linear, the index of
a point is always finite. We know moreover that it can be bounded. The se-
quences ( i n d ( x ~ ) ' x e ca} and ( i n d ( x ~ ) ' x ~ e d~) are bounded by a number
k. Moreover, in a galaxy there are for any model only finitely m a n y maxi-
mal points, and so by Proposition 8.7.13 we can assume that these sequences
are almost non-increasing. Hence, they must be stationary from some depth
a* onwards. P u t k* := ind(c~.) and e* " - ind(d~.). Then the galaxy is
almost (k*,g*)-hadronic. S4 has the galactic finite model property by The-
orem 8.6.19. Thus we can assume models to be finite sequences of almost
hadronic frames. The theorem now follows from the previous theorem. W]
Extensions of $4 of finite width in general enjoy a rather nice finiteness prop-

erty, namely that they are complete with respect to frames with finitely m a n y
segments. Namely, take any model for ~ based on a frame. Each segment
which does not contain a maximal point for p can be reduced p - m o r p h i c a l l y
to a single point. Thus, in analogy to the case of fatness, we can already

assume that almost all segments are one-membered. Now, take such a one-
point segment {x}. If it is of depth w. A, then it can be dropped. Otherwise
it precedes directly another segment. In almost all cases this segment is of
the form {y}, one-membered. In that case, we can collapse x into y, reaching
a further reduction. It is not hard to show that this leaves us with at most
twice as many segments as there are maximal points.
THEOREM 8.7.17. All extensions of S~ of finite width are complete with

respect to frames with finitely many indecomposable segments.
COROLLARY 8.7.18. All extensions of S,~ of tightness 2 have the finite
model property.
PROOF. Indecomposable segments consist of points of same depth. So,
the logics are complete with respect to frames of finite depth. Hence they
have the finite model property. [:]
We can push this result a little bit further. Consider a frame of width 2
and tightness 3. Then the points of infinite depth are all eliminable because
any embedding pattern using these points can be avoided.
THEOREM 8.7.19. All logics containing S4 of width 2 and tightness 3 have
the finite model property.
E x e r c i s e 303. A logic is called d e n s e if it contains the axiom (}p --+ (}(}p.

Show that this corresponds with the refutation pattern ~ ' ~
Show that all logics of dense linear frames have the finite model property.
E x e r c i s e 304. Show that there are countably many logics of dense linear
orders, each finitely axiomatizable.
E x e r c i s e 305. (SEGERBERG[194].) Show that all quasi-normal extensions

of $4.3 are normal.
E x e r c i s e 306. Show that all extensions of S4.3 of the form Th [ for a single
frame are Halld~n-complete.
E x e r c i s e 307. (SPAAN[202].) Show that for every consistent logic contain-

ing S4.3 satisfiability of a formula is NP-complete. Hint. NP-hardness is
clear. Show that given T, there is a model of size at most ITI + 1.
E x e r c i s e 308. Show Proposition 8.7.13
E x e r c i s e 309. Show that all dense logics of width 2 have the finite model
450 8. E x t e n s i o n s of K 4
property.
E x e r c i s e 310. Give a detailed proof of Theorem 8.7.19.
8.8. B o u n d e d P r o p e r t i e s a n d Precomplete Logics above $4

In GEORGE SCHUMM [190] the following definition is given. A logic 0
b o u n d s a property ~ of logics if 0 fails to have ~ while all proper extensions
of O have ~ . Often such a logic is said to have p r e - ~ . (For example: pretab-
ular, precomplete.) Let us call a property e s s e n t i a l l y b o u n d e d if every logic
without ~ is contained in a logic that bounds ~ . If the inconsistent logic has
then no consistent logic can bound ~ . So the concept of boundedness is
only interesting for properties which the inconsistent logic has.
THEOREM 8.8.1. Let 0 be finitely axiomatizable. Then the property 'con-
t a i n s 0 ' is bounded.
The proof of this theorem is left as an exercise. It uses only the fact that
the finitely axiomatizable logics are the compact logics, and that the lattice of
logics is algebraic. Notice that a lower cover of a finitely axiomatizable logic
need not be finitely axiomatizable again. For tabular transitive logics this
is correct, though. This is the deeper reason for correctness of the following
theorem.
THEOREM 8.8.2. The property 'is o f c o d i m e n s i o n less t h a n n ' is a
bounded property in the lattice ~ K4.
PROOF. There exist only finitely many logics of codimension < n in ~ K4.
They are tabular and finitely axiomatizable. Hence for each logic O of codi-
mension n there is a finite set L(O) of lower covers, such that every logic
properly contained in O is contained in a member of L(O). Let L be the
union of L(O), 0 of codimension n. Then any logic of codimension ~ n is
contained in a member of L. [i]
In this section we will examine completeness and incompleteness of logics

containing $4. In [62], KIT FINE has constructed a logic containing $4 which
is not complete. We have met this logic in Section 7.6. We will show that it is
also p r e - c o m p l e t e . The underlying frame of E is shown again in Figure 8.8.
The sets are the sets which are finite or cofinite subsets of the three upper
layers and finite or cofinite on the lower layer.
THEOREM 8.8.3. The logic of ~ is precomplete. Moreover, each proper
extension has the finite model property.
PROOF. We have Th ~ _~ Grz. The following embedding pattern is sat-
isfiable but cannot be satisfied on a frame. (The oval encloses the only closed
domain of this embedding pattern.)
8.8. Bounded Properties and Precomplete Logics above $4 451
FIGURE 8.8.
X7 X6 X5 X4 X3 X2 Xl X0
The reason is simply that the root point must always be a part of the head.
Hence it has a successor which is also a head part etc. There exists therefore
an ascending chain of points satisfying this embedding pattern. Hence the
underlying frame does not satisfy Grz.
Now we show that the logic of E is pre-complete, and that every proper
extension has the finite model property. Consider the subframe generated by
Xl. If the points of depth < 2 are contracted onto a single point (they are
enclosed in a box in Figure 8.8) then we get a p-morphism onto E. Hence
if a formula is satisfiable at x l it is satisfiable at x0. So, by induction,
is satisfiable at xi, i C ~, iff it is satisfiable at x0. Let ~ be such that
Th E G ~ is different from Th ~. Then ~ is not satisfiable at any point xi.
Then Th ~ 9 ~ _~ Th ~<~. This logic has the finite model property. [::]
The situation changes drastically when we shift to logics of finite width.

By Theorem 8.6.14, there are no pre-complete extensions of finite width.
What we will show is that there are 13 logics of finite width which have pre-
finite model property, and that finite model property is essentially bounded
in the lattice of logics of finite width. Let O be a logic of finite width that
does not have the finite model property. We will show that O is contained in
one of 13 logics that will be described below.
LEMMA 8.8.4. Let 0 be a logic containing S~ without the finite model

property. Then #~ G o is a frame for O.
452 8. E x t e n s i o n s o f K 4
PROOF. We may assume that O - Th ~ for some ~. Consider ~ : - ~ < ~ .

Suppose t h a t ~ is decomposable into infinitely many segments. Then
- ... @~3 @~2@~1 @~0
for some ~ , i C w. Then all Ri are finite. They are totally local, and
so the contraction of ~i --" o induces a p - m o r p h i s m of ~ reducing t h a t
component. Let k C ~. Denote by ~k the result of contracting the frame
9.. ~k+2 (~ ~k+l Q Rk onto a single point. Then Th ~ - ~ k Th ~k. Hence one
of the Th ~k fails to have the finite model property. We may assume therefore
t h a t ~ is not decomposable into infinitely many segments. Then it contains a
single infinite segment. Contract all other segments into a single point. Then
we have reduced ~ to a single indecomposable segment. Moreover, we can
assume that there is a single galaxy of depth 0. Otherwise, we contract all
but one galaxy into a single point of depth 0. Then ti(> 2) is satisfiable in
6 . Therefore there exist points x0, Y0, Yl and u such that uQxo, uQyl~yo,
and x0 ~ayl, x I~y0. Contract all points which do not see x0 or y0 into a single
point. Now we construct the points of depth 2. There exists a point Xl,~X0
of depth 2. We may assume that x l sees a point Yi. Since the segment is
indecomposable, i - 0. Now for the points os depth 3. There exist x2 and
y2 such t h a t X2~Xl and y2~yl. X2 sees a Yi. By indecomposability, i - 1.
Likewise, Y2 <3xj for some j < 2. Again j - 1. And so on. Hence, the segment
is isomorphic to #~. Now, #~ --~ #~ Q o. K]
The following is proved using Proposition 8.6.9.
LEMMA 8.8.5. Th #~ G #~ -- Th #~.
It follows that if O is without the finite model property, there is a O-frame

0 ( ~ # ~ Q o such that Th 0 ( ~ # ~ G o fails to have the finite model property.
Suppose that (A), (B), (C) and (D) hold.
(A) 0 is of fatness 1.
(B) 0 is of tightness 3.
(C) 0 does not contain p.
(D) 9 is of width 2.
Then the logic of 0 (~ #~ G o has the finite model property. For an indecom-
posable segment of f~ is either of tightness 1 and so of depth 1, or it is of
tightness 2 and then isomorphic to #~. We leave the verfication of this fact
to the reader. So, one of (A), (B), (C) and (D) must fail. Suppose (A) fails.
Then 9 contains a cluster of fatness > 1. Let it be C. Let b be the subframe
of 0 generated by C. Then ~) --- C ~ 2 or f) --- C Q o --- 2 Q o. In the first
case, T h 0 Q # ~ Q o C_ T h 2 G # ~ Q o . In the second case Th, I ~ Q # ~ Q o -
Th 2 Q o Q #~ Q o - Th 2 Q #~ Q o. The last equation follows from Theo-
rem 8.6.10. Now suppose that (B) fails; we assume for simplicity t h a t (D)
8.8. B o u n d e d P r o p e r t i e s and P r e c o m p l e t e Logics above 8 4 453
FIGURE 8.9. p
0~i ~0
;0
does not fail. Call u the frame consisting of a chain of l e n g t h 2 parallel to a

single point. This frame is e m b e d d a b l e into 0. 0 is n o e t h e r i a n . T h e r e exists
chain x4J]x3Qx2J]xlJ=]xo a n d a point y such t h a t x4~y a n d y II X0 (and so
y ii x l , y ii x2 a n d y ii x3 as well). It follows t h a t we m a y choose the points
in such a way t h a t xi i m m e d i a t e l y succeeds xi+l a n d also y i m m e d i a t e l y suc-
ceeds x4. Take the s u b f r a m e [~ of 0 g e n e r a t e d by x4. We m a y assume t h a t
all clusters have size 1. Take a point u different from y or xi; t h e n either
y <] u or x0 <] u. Therefore, we m a y contract all points different from y and
the xi onto a single point v. C a s e 1. x0 ~] v. T h e n c o n t r a c t y a n d v to
a single point. C a s e 2. y ~] v. T h e n contract x0 and v to a single point.
C a s e 3. B o t h y <1 v and x0 <] v. Hence, we m a y assume t h a t 0 is of the form
u G # ~ or u Q o Q # ~ . In the latter case we m a y use T h e o r e m 8.6.10 again and
drop the o. Hence we have shown t h a t the logic of the frame is c o n t a i n e d in
u Q #~. Now suppose t h a t (C) fails. T h e n as before we m a y assume t h a t we
have points x2J=]xlJ:]xo and x2J=]ylJ:]yo. For simplicity we assume again t h a t
the frame is of w i d t h 2. In the same way we reduce 0 to the form p G #~ or
P G o Q # ~ . Aagain, by T h e o r e m 8.6.10 the logic of the l a t t e r is identical to
the logic of the former. This leaves us with (D) to discuss. Suppose we have
an antichain X - {xo,xl,x2}. By d r o p p i n g i n t e r m e d i a t e galaxies (after con-
t r a c t i o n to a point and t h e n d r o p p i n g t h a t point supersafely) we can reduce
to a d e c o m p o s i t i o n of two galaxies 0 Q #~, where 0 contains an a n t i c h a i n of
size 3. We c o n t r a c t all points seen by either m e m b e r of X into a single point
z, if such points exist. This is a p - m o r p h i s m . F u r t h e r m o r e , consider a point
y i, X . T h e m a p c o n t r a c t i n g the interval [y, z] onto z is a p - m o r p h i s m . There-
fore we m a y assume from now on t h a t X is an antichain of m a x i m a l size. We
have now the following s t r u c t u r e at d e p t h 0 and 1 of 0 shown in Figure 8.10.
T h e first s i t u a t i o n can be reduced to the last by T h e o r e m 8.6.10. T h e t h i r d
can be r e d u c e d to the last as well by contracting x l and z. This leaves us w i t h
two frames. Now we consider the points i m m e d i a t e l y preceding m e m b e r s of
X . Given an antichain Z, call u a Z - u n i f i e r if for all z C Z u~z a n d if uJ~v
t h e n z <3v for some z C Z. If ~Z = k, call u a k-unifier. If Z - {z} a n d u a Z -
unifier, t h e n the m a p collapsing u into z is a p - m o r p h i s m . We m a y therefore
454 8. E x t e n s i o n s o f K 4
FIGURE 8.10
X2 o X2 o X20 X2 o
X1 o ~o Z X1 o ~aZ X1 o
/
XO X0 o ~ X0 o X0 o
FIGURE 8.11
assume that we have no 1-unifier. For a subset Y C X of cardinality at least

2 denote by YY the Y-unifier. For certain Y there may not exist such a y r .
C a s e 1. There is an X-unifier, y z . Then we take the subframe generated by
y x . (This gives two possibilities, depending on whether X is of depth w or
whether x0 and Xl are of depth w § 1, see Figure 8.10. For the first possibility
can be eliminated through dropping z and the third by a contraction.) C a s e
2. y z does not exist. C a s e 2a. Assume that all 2-unifiers yy exist and t h a t
there is a {Y01, y02, yl2}-unifier w. Then we take the subframe generated by
w. This gives 2 distinct frames, shown in Figure 8.11. C a s e 2b. For two sets
if, Z C_ X of cardinality 2 there exist yy and Yz and a {yy, yz}-unifier w.
We take the subframe generated by w. (This gives 3 possibilities, depending
on whether X is entirely of depth ~v or whether two points, say x0 and x l
are of depth w § 1. In the latter case we have to distinguish whether or not
{x0, xl} e {Y, Z}.) C a s e 2c. For only one set Y of cardinality 2 there exists
yy. Let xi E X - Y. Then {xi, y y } is an antichain and has a unifier, w. We
take the frame generated by w. (Again 3 possibilities.)
THEOREM 8.8.6. Finite model property is a bounded property in the lattice

of logics of finite width containing S4. Moreover, there are 3 logics of width
2 bounding finite model property, and 13 logics of width 3.
We close this section with an overview over bounded properties. For
most claims the proof is rather easy. G . 3 bounds the following properties:
finite codimension, tabularity, compactness and elementarity. Above we have
8.8. Bounded Properties and Precomplete Logics above S4 455
TABLE 1. Bounded Properties
property [bounded
finite codimension yes
tabularity yes
finite model property yes
completeness yes
compactness yes
finite axiomatizability yes
elementarity yes
decidability yes
interpolation yes
Halld~n-completeness yes
shown that completeness is a bounded property. We have shown in Theo-

rem 7.5.15 that there is a logic bounding finite axiomatizability. Now we turn
to decidability. Let a(0) := 9 and a(1) := 9 @ o. Take an infinite sequence
a :w --+ {0, 1}. Then let [~ be the following Kripke-frame
~:~ "--... (~a(a(3))(~a(a(2)) @a(a(1)) (~a(a(O))
It is easily seen that for any two different sequences a and/3 that Th [~ ~ Th [~.
Hence we obtain the following theorem.
THEOREM 8.8.7 (Blok). There are 2 ~~ many pretabular logics in E K#.
Now it is obvious that there must be also pretabular logics which are
undecidable. These logics bound decidability since all proper extensions are
tabular and so decidable, by Theorem 1.6.1. The logic of the two one-point
frames (axiomatized by (}p ~ p) is not Halld~n-complete. For 0 T V if]A_ is a
theorem, but neither 0 T nor C]_k is a theorem. However, all proper extensions
are Halld~n-complete. In order to show that interpolation is bounded it is in
fact enough to show that there is a logic of $4 of finite codimension which
fails to have interpolation. This is left as an exercise.
E x e r c i s e 311. (WIM BLOK [21].) Let b3 := o @ ( o | and b := b3 Q # ~ .

Show that the logic of b fails to have finite model property. Show that it is
the union of two logics which have finite model property. Hint. Define the
subframe logic which consists of frames of fatness 1, width 2, tightness 4, such
that there exists at most one chain of length three parallel to a point. Show
that the theory of b can be obtained by splitting two frames from this logic,
and that splitting either frame results in a logic with finite model property.
456 8. E x t e n s i o n s of K 4
E x e r c i s e 312. (Continuing the previous exercise.) Show that Th[) is a

splitting of K 4 by finitely many frames. Hence, splitting does not always
preserve the finite model property.
E x e r c i s e 313. Let ~ be the irreflexive counterpart of ~ of Figure 8.8. Show

that the logics Th ~ @ a ~ for c~ < w form an ascending chain of logics. The
limit of this chain is Th ~ @ w ~
E x e r c i s e 314. Prove Theorem 8.8.1.
E x e r c i s e 315. Let (~ be the logic of the frame # Q A ~ where # - o@ (o @

(o @ o)). Show that 0 cannot be finitely axiomatized. Show that any proper
extension can be finitely axiomatized, however. (See [125].)
*Exercise 316. Show that there are 2 ~~ many pre-complete logics in the
lattice ~ Grz.
E x e r c i s e 317. Show that if c~ ~= /3 are infinite sequences of 0 and 1 then

Th f~ # Th f~.
E x e r c i s e 318. Let A bound decidability. Show that (a) A is R-irreducible,

(b) if A is recursively enumerable then it is not the lower limit of a chain of
logics.
E x e r c i s e 319. Show that there exists a tabular logic in ~ $4 which fails to

have interpolation. Hint. Show that there exists a logic which fails to have
superamalgamation.
E x e r c i s e 320. Show that the logic of o Q 9 bounds 0-axiomatizability. Hint.

o @ (. @ 9 satisfies the same constant formulae as o @ ..
8.9. Logics of F i n i t e T i g h t n e s s
We advise the reader to study Section 10.4 before entering this section.
We will show here that logics of finite width and finite tightness are decidable
if finitely axiomatizable. The proof uses methods from the theory of finite
automata, which are provided in Section 10.4.
LEMMA 8.9.1. Let A be a logic extending K ~ and let ~ be a formula.

Suppose that ~var(~) < k. Then ~ C A iff ~ C h . f t ( < 2k).
PROOF. Clearly, if ~ E A then ~ C A.ft(< 2k). Now suppose that ~ r A.

Then ~anA(k) ~ ~. Now, ~anh(k) is of fatness < 2k; so it is a frame for
A.ft(_< 2 k). Hence ~ r A.ft(<__ 2 k). [i]
8.9. Logics of Finite Tightness 457
Let 0 be fixed. A 0 - a c c e s s i b i l i t y m a t r i x is a 0 • 0 - m a t r i x over w such

that ass = 0 for all i < 8. A 0 - f a t n e s s v e c t o r is a 0-vector of elements in
w. A 0 - t y p e is a pair (A, v), where A is a 8-accessibility m a t r i x and v a
0-fatness vector such that if asj = 0 for some i # j then vi = vj and for all
k < 8 ask = ajk and aki = akj. (It follows that asj = asi = ajs = ajj = 0.) A
(0, ~, 7 r ) - t y p e is a type such that (i) aij < T for all i, j < 0, (ii) vs < 7r for all
i < 0. The set of (0, ~-, 7r)-types is finite and denoted by Tp(0, T, 7r) or simply
by Tp. Types are intended to code the structure of frames. A sequence of
(0, T, 7r)-types (t s : c~ < 7) is called a f r a m e s e q u e n c e if the following holds.
I f t s - ( A s , v s) t h e n a ~ < a ~ j +1 < a ~ + l f o r a l l i , j<Ssuchthati#-j. A
Z _0foralli
sequence is called r o o t e d if 3' - 13 + 1 and aij j<0.
For a rooted noetherian Kripke-frame f of width 0, tightness ~- and fatness
a rooted frame sequence of (8, T, T)-types is constructed as follows. Let 7
be the depth of ~'. Then there exist 8 many sequences E(i) = (C~ : a < 7)
for i < 8 and a map tr from the set of clusters of ~ into the set 8 such that (i)
the restriction of n to the clusters of depth a is injective, (ii) for a cluster C
of ~ of depth a, C = C~ (C) (iii) C~ immediately succeeds C~+1 a + 1 < 7
, , 9
Now a sequence (t s : a < 7} of types is defined as follows, t s = (A s,v~},

where v~ is the type of C~, and a~ is the least number k such t h a t a - 13 + k
and C~ <3 C~. (Since the frames are of tightness T, k exists, is finite and < T.
Clearly, ass - 0 by choice of the clusters. Each t s is a type. For let a~ - 0.
Then C~ <3 C~. Hence the clusters, being of identical depth, are equal. So,
ajs - 0. Furthermore, ask - ajk and aks - akj for all k. Finally, a~.+1 - a~
or a~ +1 - a~ -4- 1. Now suppose that ~ is rooted with root cluster C. Let C
be of depth ft. Then 7 = 13+ 1. Moreover, all clusters of depth fl are identical;
so, a~ - 0 for all i, j < 8. So, the frame sequences is rooted. It should be
noted that the type sequence of f depends on the division into sequences of
clusters. The results are independent of this, however. Denote by Seq([) the
set of all possible type sequences for [. Given a type sequence, let Fr(E) be
the frame corresponding to E. This is unique.
PROPOSITION 8.9.2. For every rooted frame sequence E of (0, T, 7r)-types

there exists a rooted noetherian frame [ of width O, tightness ~" and fatness 7r
such that E C Seq([).
PROOF. Now let E - (t s " a < 71 be a rooted frame sequence. Let F be

the set of pairs (a, i / such that a < V and aki # 0 for all k < i. F represents
the set of clusters of ~. For C e F has type v~. Now put (a, i) <3 (Z, J) if (1)
i-j and (la) a > / 3 o r (lb) a - 1 3 a n d v ~ # O , (2) i # j anda>fl+aij.
(Note t h a t in case (2), asj ~ 0 by choice of the set of clusters.) We will leave it
to the reader to show that <3 as defined is transitive, and that E C Seq(f). [:]
458 8. E x t e n s i o n s of K 4
Now let E - (t ~ 9 a < 7). Let ~ be an ordinal such that w~ _< 7. P u t

E(~) := <t~ " w ~ < a < w ( ~ + 1))
E(n) corresponds to the galaxy of depth n in ~3(~). Hence we call E(n) a
galaxy sequence. Put
u(r ) 9 _<
Just as with frames, we may study E be means of the sequence of galaxy

sequences.
DEFINITION 8.9.3. A class 2C of (noetherian) frames is type regular if

there exists a regular language R such that [ E :~ iff for all ~ such that w~ <
dp(f), E(~) E R.
LEMMA 8.9.4. The class of all noetherian frames of width 0, tightness T
and fatness 7r is regular. For regitlar classes ~ and ~ , ~ U ~ , 2C A f~ and
2C- f~ are regular.
The proof of this theorem is left as an exercise. We note t h a t if T < T'
and ~r _< 7r' then Tp(0, T, 7r) C_ Tp(0, T', ~'). A regular class X of (0, ~-, 7r)-
frames is therefore also a regular class of (0, ~-', 7d)-frames. Moreover, if R
is a regular expression defining ~ as a class of (0, 7, 7r)-frames, then R is a
regular expression defining :~ as a class of (0, ~-', ~d)-frames. This is useful
to know. Now let us see why we can limit ourselves to type regular frames.
Suppose we want to know which frames can refute ~. Then, alternatively, we
can ask ourselves which refutation patterns can be realized. Finally, since we
have bounded width and no ascending chains, we can limit ourselves to em-
bedding patterns. Take an embedding p a t t e r n e(D, ~ ) . Let us for the m o m e n t
ignore the closed domains. Let O - Dn-1 Q . . . (~O0 be a decomposition into
indecomposable parts. Then L" b --+ ~ is a cofinal embbedding into ~ if t F O0
is a cofinal embedding, L F bi is an embedding for 0 < i < n and L preserves
the decomposition, that is,
: . . .
Now, let us consider the simplest possible case, that of the embeddability of an
indecomposable part, cofinal or not. Recall the definition of the antiframe. We
have defined x i, y by x @ y ~ y, and defined the antiframe of [ - {f, <]) to be
the frame (f, I,). Now, if f is indecomposable then its antiframe is connected.
Moreover, if ): is a subframe of g, then the antiframe of [ is embedded as
a subframe in the antiframe of g. Consequently, in our case t[bi] must be
connected in (f, I~). However, by the fact that the target frame f is of tightness
T, the points of bi cannot be too far apart. Namely, if x ,, y then L(x)~ L(y);
but the latter can only be if L(x) and L(y) are in the same galaxy (recall t h a t
f is galactically linear) and their local depths differ by at most 7. If b is
indecomposable, the antiframe is connected, and for each pair x, y of points
there is a path of length at most ~0 connecting x to y in the antiframe. This

shows the following lemma.
LEMMA 8.9.5. Let ~ be an indecomposable frame of cardinality e and let
L 9 0 --+ ~ be an embedding into a noetherian frame of tightness 7. Then for
any two points x and y of 0 the points L(x) and ~(y) are in the same galaxy,
and the local depths differ by at most ~'. e.
LEMMA 8.9.6. Let ~ be indecomposable, and e(~ 93) a (cofinal) embed-
ding pattern. The class of frames that realize e(O, f~) is type regular.
PROOF. We show the criterion of embeddability first. Let E C Seq([).
Clearly, e(~ ~ ) is realizable iff e(~ ~ ) is realizable in a galaxy of ~ iff
e (~ (~, ~ ) is realizable in a subframe of points consisting of the slices a, a + 1,
. . . , a + T. t~, for some a. Hence there exists a finite set T of sequences of
length _ T. e such that e(~ ~ ) is realizable in [ iff no m e m b e r of T is a
subword of some E(n). For each n, the language of types not containing a
member of T as a subword, is a regular language. Namely, for each t C T we
can write a term t'defining just the language {t}. Then the t e r m Tp*. i'. Tp*
defines the language of sequences containing t. Now take the union of all
these terms (this is well-defined since T is finite). The intersection of the just
defined language and the language of frame sequences is again regular and
can be described by a regular term. For cofinal embeddability there exists a
finite set S of sequences of length _ T. t~ such that e(~, ~ ) is realizable iff E(0)
does not start with a member of S. This is a regular language, as is shown
similarly. D
So, the next step is to consider embedding patterns consisting of a de-
composable frame. Here, we meet a subtle problem. If 0 = m Q n we may
not necessarily conclude that an embedding p a t t e r n for 0 can be realized by
embedding m before n. Whether this division can be made depends on the
closed domains of t h a t embedding pattern. Recall that a closed domain can
be viewed as a cone, i. e. an upper set, or alternatively as an antichain. It is
the latter that suits our purpose best here. Let
Each antichain u is contained in one and only one segment of 0. An antichain

is called an o u t e r a n t i c h a i n of 0i if 0i = J:u, else it is called an i n n e r
a n t i c h a i n . An antichain is an outer (inner) antichain of 0 if it is an outer
(inner) antichain of some segment (which is unique given the antichain). The
outer chains glue the segments to each other in the following way.
LEMMA 8.9.7. Let ~ : 01 ~ 0 o --+ f be an embedding and t~ an outer an-
tichain of 00. If ~ satisfies the closed domains condition for ~ then 01 and 0o
are in the same galaxy and the depth of a maximal point of ~[01] is at most ~-
larger than the maximal depths in L[00].
460 8. E x t e n s i o n s of K 4
PROOF. Let z<~L[v]. Then z may not be an external point; hence z C e[~l].
On the other hand, there is such a z at depth _< max dp[~(tJ)] + T. [--1
By a c o m p a r t m e n t of ~ with respect to ~U we understand a maximal chain

of the form ~i+j-1 @ . . . @~i of segments of ~) such that for every k < j - 1
there is an outer antichain for ~i+k in ~U.
LEMMA 8.9.8. Let ~ -- ~k @ ... @~o be a decomposition of ~ in compart-

ments with respect to ~U. Then the embedding pattern ~(0, f~) can be satisfied
iff there is an embedding L 9 ~ --~ ~ which satisfies all embedding patterns
(~i, 93 [ ~i) considering points of depth < m a x { d p ( x ) " x C ~} + T.
PROOF. If L satisfies the embedding pattern for ~ then it satisfies the
embedding patterns for the compartments. For the converse, let L respect
the decomposition and satisfy the embdeding patterns for the compartments.
Then L is an embedding for ~. Furthermore, let v C !U. Then v C_ Oi for
some i. Let z <~ ~[~]. We have to show that z is in ~[~]. Suppose t h a t v is
an outer antichain; then, by definition of a compartment, it is not an outer
antichain of the compartment, so all antichains are inner antichains of the
compartments. This means that there is no antichain ~ in a c o m p a r t m e n t
Oi such that ~z N J:Oi - J:v for an internal point not belonging to Oi itself.
Or, to put it another way, internal points outside of Oi cannot realize views
forbidden by the closed domain conditions for Oi. So if such a view is realized,
it is externally realized. But we know that we need not look very far for such
points, namely only up to depth maxdp(L[v]) + 7. [3
Notice that it is not enough to satisfy all local patterns for compartments,
since they are formulated such that we need only consider points of the sub-
frame generated by these compartments. For the overall embedding, inter-
mediate points have to be checked as well. If a decomposition of the original
frame is given, for example into galaxies, then the condition is satisfied since
it does not have to be enforced that the decomposition is respected. The proof
of the next proposition is now obvious.
LEMMA 8.9.9. Let e(~, ~ ) be a cofinal embedding pattern. Let f - (~ ~<~g~

be an arbitrary frame. The embedding pattern can be satisfied in [ iff there
is a decomposition of ~ - O k - 1 Q . . . ~ o respecting the compartments and a
monotonic sequence of numbers (ai " i < k) such that ao - 0 and the coil-
nal embedding pattern (~o, f~ I Oo) is satisfiable in g0, the embedding patterns
e(~, ~U I ~ ) are satisfiable in ga,, 0 < i < k.
The consequence is that we have a preservation theorem concerning de-
compositions. Namely, if e(O, ~U) is satisfiable in [(~ g then it is also satisfiable
in [(~0 Q g, simply by checking that the same embedding modulo identifica-
tion of the parts does the job. This implies among other the following.
THEOREM 8.9.10. Let A be of finite width and finite tightness. Then drop-
ping nonfinal galaxies is supersafe. Hence A has the galactic finite model
property.
Having progressed this far we now need a notational system for frames
which consist of a single hypergalaxy. The essential idea is the following. Each
regular language is defined by a regular expression. Let us take a set R of
regular expressions over the set of types. R is called a b l o c k if (0) R is finite,
(i) every frame sequence belongs to some member of R, and (ii) no frame
sequence belongs to different members of R. In other words, the members of
R partition the set of frame sequences. A g a l a c t i c R - e x p r e s s i o n is a regular
expression ~ over R. A sequence ( E ( k ) : k < c~/ (c~ < w) is a c c e p t e d by
if there is a sequence (Ri : i < a/ of terms in R which is accepted by ~ and
P~ accepts E(i) for all i < c~. Likewise, a frame [ of hypergalactic depth 1
is accepted if ~ ( E ) is accepted by ~ for some E E Seq([). (We note that in
principle, even if E C Seq([) is accepted, some other E' C Seq([) might not be
accepted. This is however harmless.) A h y p e r g a l a c t i c a l l y r e g u l a r class of
frames is a class ~ of frames such that there exists a block R and a galactically
regular expression ~ over R such that [ C 9~ iff for each hypergalaxy F2 of f,
accepts F2. We show that the class of frames of frames of hypergalactic depth
1 satisfying a given embedding pattern is hypergalactically regular. From this
everything follows. We already have completeness with respect to noetherian
frames of hypergalactic depth 1.
PROPOSITION 8.9.11. Let ~ and ~ be hypergalactically regular classes of

noetherian (~, r, ~)-frames. Then ~ M ~ , ~ U ~ and ~ - ~ are also hypergalac-
tically regular. The class of all frames is hypergalactically regular. Moreover,
2C is also hypergalactically regular as a class of (8, r', )-frames for r < r ~
and ~ < ~'.
m
PROOF. First, let ~ define ~ and ~ define ~ . If ~ and ~ are regular

expressions over the same block R, the claim is clear. So, let ~ be defined
over R and ~ defined over S. Let T the set of all expressions defining the
languages L(R)M L(S), R E R and S C S, if nonempty. These expressions
exist, since regular languages are closed under intersection. T as just defined
is a b l o c k . Let R e 1~. Define R t by R t "- L.J(T" T e S , L ( T ) C_ L(R)}.
The latter is well-defined, since the union is finite. By the definition of T,
L(R) - L(Rt). Call ~t the result of replacing each occurrence of R by R t, for
each R C R. ~t is a regular expression over "IF. The class of frames accepted
by ~t is exactly :~. Similarly we can produce a term ~$ over T defining ~ .
Now the first claim is clear. For the second claim, note that the class of all
frames corresponds to the term ([-JRC~ R)*. For the last claim, note that
accepts ~ also as a class of (0, r', rr')-frames for r _< r ' and rr _< rr'. [~
462 8. E x t e n s i o n s o f K 4
THEOREM 8.9.12. The set of noetherian frames of width 0, tightness T

and fatness 7r and depth < w 2 realizing a given embedding pattern is hyper-
galactically regular.
PROOF. Let 0 -- 0k-1 @ 0 k - 2 @ . . . @01 @00 be a division of 0 into com-
partments. Then define I~ as follows. For each pair i, j < k such t h a t i < j
let O[i,j] " - 0 j @ 0 j _ l Q . . . @0i and E[i, j] be the regular expression defin-
ing the class of galactic frames in which e ~ (0[j, i], ~U F0[i, j]) is realizable, for
i > 0, and e(0[j, i], ~U r0[j, i]) if i - 0. Let X define the class of frames into
which no 0[i, j] is embeddable (confinally, if i - 0). Take all possible intersec-
tions of those Eli, j] which are nonempty. The set of the terms thus defined
together with X forms a block. Call it R. Let the union of all terms of R be
R. Let H := (is 9 x < r) be a strictly ascending sequence of natural numbers
< k - 1. P u t
~[H] := E[O, i o ] . R * . E [ i o + l , i l ] . . . . . E [ i r _ 2 + l , i ~ _ l ] . R * . E [ i T _ l + l , k - 1 ] . R *
Now let ~ be the union of all ~[H]. Since there are only finitely m a n y se-
quences H, this is well-defined. It is not hard to check that ~ admits a frame
[ consisting of a single hypergalaxy iff e(~, ~ ) is realizable. K]
COROLLARY 8.9.13. Let A be a finitely axiomatized logic of width 0, tight-
ness 7 and fatness ~. Then the class of rooted frames for A of hypergalactic
depth 1 is hypergalactically regular.
COROLLARY 8.9.14. Let A be of finite width and finite tightness. If A is
finitely axiomatizable, it is decidable.
PROOF. Let ~ be given. We want to decide whether ~ C A. For ~r large
enough, this is equivalent to the problem '~ C A.ft(_ ~)'. A is complete with
respect to noetherian frames of hypergalactic depth 1. It determines a hyper-
regular set in that class. Likewise, the set of frames of hypergalactic frames
in which ~ is satisfiable is a hyperregular set. Now their intersection is also
hyperregular. Moreover, terms representing these sets can be algorithmically
computed. Hence we are done if we can decide whether for a given ~ we
can decide whether or not there exists a frame accepted by ~. To t h a t end,
we decide whether there are nonempty R-sequences falling under ~. This is
decidable on the basis of ~. If there exists such a sequence, there exists a
frame accepted by ~. []
E x e r c i s e 321. Show L e m m a 8.9.4.
CHAPTER 9
Logics of Bounded Alternativity
9.1. The Logics Containing K.altl

Throughout this chapter we will deal with logics with one or several oper-
ators each of which satisfies an axiom of the form altn for some n. These are
the logics of bounded alternative or logics of bounded alternativity. The study
of such logics has been initiated by several insights and questions. First,
any relation on a set can be seen as the union of sufficiently many partial
functions. Hence, any monomodal Kripke-frame can be viewed as a frame
for polymodal K . a l t l where the distinction between the operators has been
lost. We will study this interpretation under the name of colouring and de-
colouring of polymodal frames. The second motivation comes from studying
attribute-value formalisms, as used in computer science and linguistics. It
has been realized that modal logics containing K . a l t i provide an ideal basis
for studying such formalisms. It quickly turned out that the quasi-functional
logics bear close resemblance to Thue-systems and this connection has been
exploited to derive numerous undecidability results in modal logic. Many of
these results were known already to ALBERT A. MUCHNIK, but only few got
published (see [55] and [56]). KRISTER SEGERBERG in [197] has investigated
the lattice E K . a l t i . Most of his results were reproved by FABIO BELLISSIMA
in [5], who added some facts about the lattice ~ K.alt2. The logics of bounded
alternative form a neglected area of modal logic, with most results generally
being unknown. However, as we will see, they provide an ideal source of many
powerful negative results on modal logic, as well as an ideal tool for modelling
certain mathematical structures.
We start the investigation by classifying the extensions of K . a l t i . We
have seen in Section 3.2 that extensions of K . a l t l are canonical, and it is
not hard to see that they all are dr-persistent. We derive from this the fact
that all extensions are complete, and so we can classify the extensions just
by looking at the frames for K . a l t i . Furthermore, it is enough to look at the
rooted frames. We can view them as ordinary frames (f, <l) or as structures
(f, s}, where s is a partial function such that s(x) is the <l-successor of x, if
such a successor exists, and undefined else. A map 7r : f --+ g is a p-morphism
from ( f , s ) to (g,s') iff for all x E f, ~(s(x)) exists iff s'(~(x)) exists, and
463
464 9. L o g i c s o f B o u n d e d Alternativity
then both are equal. To say t h a t f is rooted is to say t h a t there is a point

x e f such t h a t f = { s n ( z ) : n e w}. There are three possibilities. C a s e
1. There is an n E w such t h a t s~(x) is not defined. Assume t h a t s n - l ( x )
exists. In t h a t case we have a chain of length n. We abbreviate this frame
by con. C a s e 2. sn(x) exists for all n. In t h a t case, there are two further
possibilities. C a s e 2a. All sn(x) are different. Then our frame is isomorphic
to the n a t u r a l numbers with the successor function. We denote this frame
by cf),~. C a s e 2b. There are m , n such t h a t sin(x) = s~(x). Let m , n be
minimal such t h a t m < n. Then put p : - n - m and y := sm(x). We have
sP(y) = y, and consequently, for all natural numbers, si+P(y) = si(y). Hence,
y is in a cycle of length p (since n was chosen minimally). The elements
z , s ( x ) , s 2 ( x ) , . . . , s m - l ( x ) form a chain leading into the cycle. We denote
this frame by ct)%~,p.
PROPOSITION 9.1.1. (i.) c~)m is a generated subframe of con iff m < n,

but not a generated subframe of either ct)~ or cr)cn,p. No frame for K . a l t l
maps p-morphically onto C[~m, except for c~)m. (ii.) ct)Cm,p is a generated
subframe of cr)cn,q iff p - q and m < n, and a p-morphic image of cr)Cn,q iff
m < n and p divides q. cr)Cm,p is a p-morphic image of ct)~, but no image of
COn for any n.
The proof of these facts is easy in all cases except for the p - m o r p h i c images
of r p. In t h a t case we have t h a t if ~(x) -- u then ~(s(x)) = s(u), from
which above claims can be derived. We now have a full overview of the relation
between these frames. Furthermore, for finite f and ~ we have Th f C Th g iff
g is a p - m o r p h i c image of a generated subframe of f. Finally, let us observe
t h a t all frames except for cow are finite. However, Th r has the finite model
property. For consider a formula ~ and ct)~ ~ ~. W i t h o u t loss of generality
we can assume t h a t ~ does not hold at the root. Then for n + p > dp(~) we
get r ~ ~. This is so, because the d p ( ~ ) - t r a n s i t s of the roots of r and
cr)cn,p are isomorphic. Likewise it is shown t h a t ct)~ - ~ ( T h ct)n " n C w).
PROPOSITION 9.1.2.
Thcb~ = K.altl @ ~n~-

ThcO~ = K.altl.D
Th C~Cn,q = K . a l t l . D G ~]n (p ~ ~qp)
PROOF. By correspondence theory. Notice t h a t we can describe the struc-

ture of these frames by first-order restricted sentences, namely in the case of
COn by (VyC>n+l x)f, in the case of c[~ by (3yc> x)t, and in the case of or)ca,q by
(VyE>n x)(Vz E>q y)(z - y). All these are clear instances of Sahlqvist-formulae.
The axioms above are the modal correspondents. D
9.1. The Logics Containing K . a l t l 465
Notice also t h a t the conditions expressed by these formulae are derivable

in ~ [ , the class of differentiated frames, another proof t h a t these logics are
dr-persistent.
THEOREM 9.1.3 (Segerberg). K.altl is the logic of the chains C~n, n C w.
Every extension of K.altl is dr-persistent and has the finite model prop-
erty. Moreover, K . a l t l . D is pretabular. Each proper extension of K.altl
not containing K . a l t l . D is of the form A R Th r for some n C w, where
A C ~ K . a l t l . D . Hence in ~ K . a l t l there are only countably many elements.
PROOF. For the first claim let T be a formula of modal depth d and let
it have a model on a rooted frame f. If ~ is not a finite chain then it is a
p - m o r p h i c image of r so a model for ~ can be based on cO~. However,
the d - t r a n s i t of the root of Cbd is isomorphic to the d - t r a n s i t of any point in
cO~, so a model can be based on COd as well. Consider a proper extension A
of K . a l t l . D . Suppose there are infinitely many or)on,p which are frames for
A. T h e n the set {n + p 9 or)on,p ~ A} is unbounded. Hence every formula
satisfiable in c ~ is satisfiable in a frame for A, so t h a t A C_ Th c ~ . Hence all
cr)Cn,p are frames for A, a contradiction, since then A - K . a l t l . D . Every logic
containing K . a l t l is determined by its frames of the form Olin, n E w, and
the frames of the form or)on,p. Of the first there are finitely many, since the
extension is proper. T h e n there is a largest chain con contained in the set and
the set of chains for A is the set of chains of the form corn , m < n. The logic of
the cycles for A is an extension of K . a l t l . D and this proves the claim. Finally,
to see that the lattice of extensions is countable, observe t h a t K . a l t l . D has
at most as many proper extensions as there are finite subsets of w • w. w • w
is countable, and the set of finite sets of a finite set is countable again. So
K . a l t l . D is countable. Now, an extension A of K . a l t l is characterized by
the number of the largest chain admitted by it and the logic A LJ K . a l t l . D .
So there are at most w • w logics, again a countable number. [::]
From this theorem we can get a complete overview of the lattice ~ K . a l t l
by describing the corresponding TD-space. Recall that if | is a poset then
9 (| is the space of the topology having as closed sets all unions of sets of
the form Tx. Now, given two topological spaces 3~ and ~ , define X >~ if] as
follows. A set is closed iff it is a finite union of sets A • B where (i) A ~= X
or (ii) A = X and B = Y. It is immediately verified t h a t this is a topological
space.
THEOREM 9.1.4. Let a := (w, <_}op, and # := (w - {0}, I}~
3| K.altl.D) ~- ~ ( a • #)
3| K.altl) ~ r x r • #)
a derives from the additive structure over w (n < m iff there exists a k
such t h a t n + k - m), and # from the multiplicative structure over w - {0}
466 9. Logics o f B o u n d e d A l t e r n a t i v i t y
FIGURE 9 . 1 . ~ K . a l t l
L2 (-- T h (COo))
K.altl.D z "~ f ~ ~ ~ ,1
............ -. . . . . . . . . . 9 ~ Th (r
I i ~ ~ rh(cr)co,1)nT~C~l)~~ /
T h (cD~) A T h ( .............
9 T h (cr)co,1 ) N T h ( c O n ) ~
i --- . . . . . . . ~ T h ( c D n )
\
,\
T h ( ct)~ ) N T h ( ci)n ~ ~( .~. -- --
(with n I m iff there exists a k > 0 such that n k = m ) . Note that a = (w, _},
but for stating the theorem it is better to display the similarity between a and
#. For a proof of the first fact notice that every proper extension is tabular,
and so its corresponding set in the space is finite. Hence the topology is
indeed the weak topology. Now, an extension of K . a l t l is an intersection of
an extension of K . a l t l . D with a logic of the form Th r This representation
is unique. Hence, any logic extending K . a l t l corresponds to a closed set in
~| and a closed set in (I)(a). However, if the latter is the entire
space, the extension is not proper, and conversely. In that case, the first set
is the entire space as well. This shows the correctness of the representation.
THEOREM 9.1.5. A l l logics in ~ K . a l t l are f i n i t e l y a x i o m a t i z a b l e , have

the global f i n i t e m o d e l p r o p e r t y a n d are globally decidable.
9.1. The Logics Containing K . a l t l 467
PROOF. If A = K . a l t l . D , the claim is clearly true, so let A be a proper

extension. Every extension of K . a l t l . D is either this logic itself or tabular.
In that case, all three claims follow. The logics Th r also have all these
properties. Any finite intersection of logics which have the global finite model
property also has the global finite model property, so we need to show that the
intersection is finitely axiomatizable. This is unproblematic if A D K . a l t l . D ,
since in that case A is tabular. Thus, we have to study logics of the form
K . a l t l . D N Th c[~. In that case A = K . a l t l | ~(}~-F V 0~+l-F.
PROPOSITION 9.1.6. Every quasi-normal extension of K . a l t l . D is nor-

mal.
PROOF. A quasi-normal extension is the theory of some set of pointed
frames ([, w0), where w0 is the root of [. We will show for every x C f we have
Th ([, w0} C Th ([, x). It follows that Th ([, w0) - Th [. So, take an [. C a s e 1.
[ - cow. Then the claim is immediate. C a s e 2. [ - cr)c,~,p. Then the transit
of x in [ is of the form cr)cn,p for n < m. By Proposition 9.1.1 there exists a
map cr)c,~,p -~ ct]cn,p. It maps the root of the former frame onto the root of
the latter frame. Hence Th ([, w0) C_ Th ([, x).
PROPOSITION 9.1.7. An extension of K.altl is Hallddn-complete iff it is

the logic of r cO~ or ct]Cn,p for some n,p G w.
PROOF. A logic is Halld~n-complete only if it is the logic of the one-
element chain or contains the axiom D. Let the latter be the case. In case
A is the theory of a single frame cr)%,v , it is the quasi-normal theory of
(ct)%,v,x), where x is the generating point. Hence, using Theorem 1.6.5 we
see that the logic is Hallddn-complete. If it is not the logic of a single frame,
then there exist frames cOcm,p and ct]%,q such that there is no frame for A
having both of t h e m as its p-morphic image. Call these frames m i n i m a l .
There exists ~ such that ~ is refutable on ct]cm,p but not on r and a r
which is refutable on cr)Cm,q but not on cO%,p. We can assume that ~ and r
are disjoint in variables. Then ~ V r is not refutable on either frame, or of a
p-morphic image of the two. So, if A is characterized by these two minimal
frames, p V r cannot be refuted on any frame for A, though ~a and r both
are refutable. Likewise we can proceed if A is characterized by n minimal
frames, picking a formula ~n for each minimal frame. (Choose, for example,
the diagrams of each of these frames, prefixed by a sufficiently large ff]-<~.) if]
E x e r c i s e 322. Show that K . a l t l has 2 s~ many quasi-normal extensions.
E x e r c i s e 323. Show that K . a l t 2 has 2 s~ many normal extensions. Hint.

These extensions can even be axiomatized by constant axioms.
E x e r c i s e 324. Show that K . a l t l . D is the only pretabular logic in s K . a l t l .

468 9. L o g i c s of B o u n d e d A l t e r n a t i v i t y
Exercise 325. Show that A splits s K . a l t l . D iff A - Th ct)n for some n C w.

The logics K . a l t l / r are called Chellas-Hughes-Logics in [197]. Show t h a t
they can be axiomatized by constant formulae.
Exercise 326. Show that ~ K . a l t l . D has exactly one splitting, namely the
reflexive, one-point frame.
Exercise 327. Show that every element in ~ K . a l t l . D has N0 many lower

covers.
Exercise 328. Show that E K . a l t l . D has 2 s~ automorphisms.
Exercise 329. (SPAAN [202].) Show that K . a l t l is NP-complete. Hint.

Estimate the size of models.
9.2. P o l y m o d a l L o g i c s w i t h Q u a s i - F u n c t i o n a l Operators
In this section we will turn to the logics containing @i<~ K . a l t l . The
principal aim is to define a set of formulae with which any extension can be
axiomatized and which is geometrically perspicuous.
DEFINITION 9.2.1. Let 0 be a modal logic, and ~, r formulae. We say
that ~ and r are a x i o m a t i c a l l y e q u i v a l e n t o v e r (9 if 0 | ~ - 0 (~ r
Obviously, if ~ and ~ are deductively equivalent in O then they are also
axiomatically equivalent. The converse does not hold, however. For example,
i f r = p and ~ = q, p ~- q, then | = O@l_ = OOr But if O is
consistent, ~p and ~ are not deductively equivalent. For then p +-~ q C O, from
which p ++ ~p C (9, hence _[_ C O.
Recall that a model for polymodal K . a l t l is thought of as a frame or
alternatively as a partial algebra with n many partial unary functions. An
extension of polymodal K . a l t l . D is complete with respect to semigroup mod-
els. To be precise, take a frame (f, (<~j: j < n)). Let x C f and j < n. Then
if x has a j-successor y we put j ( x ) : = y and let j ( x ) be undefined otherwise.
For a sequence a we define inductively or(x) to be that element which can
be reached from x 'following' the path defined by a, if that p a t h exists. In
extensions of polymodal K . a l t l . D all paths exist, and they form a semigroup
under concatenation. Now, our first aim is to derive a special normal form
for formulae. Notice, namely that we have for each operator the theorems
Ojp. ++ .OjT A [:]jp , Oj(p A q). ++ .Ojp A Ojq ,
Dip. ++ .[]j J_ V (}jp , Dj(p V q). ++ .K]jp V W]jq .
If we apply the standard algorithm for obtaining normal forms of Section 2.7
we observe that a normal form of degree n + 1 is a disjunction of formulae
of some r i < n, where each r is a conjunction of (i) a normal form # of
9.2. Polymodal Logics with Quasi-Functional Operators 469
degree 0 and (ii) for each j, either Fqj_L or KIjX A 0iX, where X is some normal
form form of degree n. The following definition was proposed in [91].
DEFINITION 9.2.2. A formula ~ is a s t r i c t c a n o n i c a l a l t l - f o r m u l a if

it is of the following form
crEW T a ( t W -L i<k
Here, cri,Ti are finite sequences of indices, and W T, W -L finite sets of such
sequences. A formula ~ is a c a n o n i c a l a l t l - f o r m u l a if it is a disjunction of
strict canonical formulae ~i, i < n, such that var(~i) r~ var(~j) = ;~ if i r j.
PROPOSITION 9.2.3 (Grefe). There is an effective algorithm reducing each

formula ~p into a canonical ~ K . a l t l - f o r m u l a r which is axiomatically equiv-
alent with ~ over ~ K . altl .
PROOF. S t e p 1. Let ~ be given. Eliminate all symbols different from

variables, T, 7, A, V and 0j by their respective definitions. Next move
unary operators into the scope of binary ones, i. e. negation into the scope
of conjunction and disjunction, and 0j into the scope of conjunction and
disjunction as well. The latter is admissible, by the laws of distribution given
above. Next, replace subformulae of the form 0j-~r by -~0jr A OjT. Finally,
move conjunction out of the scope of disjunction. If none of these operations
are applicable, we have obtained a formula p(1) which is deductively equivalent
to ~ and of the form Ai<m ~}1), where for each i there exist finite sets of finite
sequences over n, W + and W - , such that
~(1) o- o-
,:: V ooTv V V
nEW T n E W -L nEW + nEW-
The variables P~,i need not be distinct for distinct cr or i. S t e p 2. There

exists a renaming of the variables which makes the formulae 9 (1) disjoint in
variables. In general, the axiom a A/3 is equal in force with the axiom a J~/3,
which results from making a and/~ disjoint in variables. Now assume that P~,i
occurs only positively in ~I 1). Then we can replace 0~p~,i by 2. Likewise, if
p~,~ occurs only negatively. (That this results in an axiomatically equivalent
formula has been the content of an exercise of Section 5.5.) Now take E + to
be the set of paths in W + with associated variable p = p~,i and E - the set
of paths in W - with associated variable p. By construction, E + and E - are
not empty. Let X be the disjunction
V v V
o-EE+ o'EE-
470 9. Logics of B o u n d e d A l t e r n a t i v i t y
Replace X by Xr which is defined by
X' := V O'~po-,.rV--,O'rp,~,r
crEE+,~-EE-
We show that X and XI are axiomatically equivalent. Since X is a substitution
instance of X!, we must have ~ ) K . a l t l @ X C_ (~ K . a l t l @ X~. The converse
inclusion remains to be shown. Now let ~ ~ X~, say (g,x / ~ X~. Then all
paths 7 from E - must exist. Furthermore, assume a path a C E + exists
starting from x. Then its endpoint is distinct from all endpoints of paths ~-
starting from x. Let ~ ( p ) : = {~(x) : ~ C E - } . Then (~,/3, x) ~ X, as required.
(2) _(1)
Let ~i result from ~i by performing the substitution just mentioned. A
slight modification of the previous argument shows that p~2) is axiomatically
equivalent to ~1). This concludes the second step. Put p(2) ._ Ai<m ~2).
~(2) is a conjunction of formulae of the type ~2) with the distinctive property
that each variable occurs in at most one p~2) and there exactly once positively
and once negatively. S t e p 3. For each variable p of ~(2) there exists an i < m
(2)
and two subformulae of ~2) of the form (}~p and -~0"p. ~i is therefore
deductively equivalent to a formula r V 0~p V -~0~p for some r not containing
(2)
p. Replace ~i by r V ~(}~-[- V (0~p ++ 0"p). This results in a deductively
equivalent formula. Perform this operation for every occurring variable. The
resulting formulae are denoted by ~p~3). Rewrite ~3) by moving disjunction
out of the scope of conjunction. After that ~3) is a disjunction of formulae
r which are conjunctions of formulae which are either constant and of the
form (}~-V or -~0~-[- or they are of the form (}~p ++ 0~p. Moreover, ~(3) is
deductively equivalent to the disjunction of the r for some suitable set E of
pairs (i, j}. In each r replace all variables by a single one among them. This
yields the formula 7-i,j. By choice of the variables, var(T~,j) A var(Tm,n) = 0
for (i, j) ~ (m, n I. It is easy to check that ~-i,j and r are axiomatically
equivalent and so is their disjunction. Finally, let ~(4) be the disjunction
of the Ti,j, (i, jl E E. ~(4) is axiomatically equivalent to ~(3). p(4) is in
alt 1-canonical form. [-]
PROPOSITION 9.2.4. Every extension of polymodal K.altl can be axiom-

atized by formulae of the form
xvy,,(
A A" ~,~-~0"TA)A(0:"'~jTA0"<~JTA0~'
k ~Pi~0"~Pi)
where vat(x) = ~ and crij # ~-ij.
COROLLARY 9.2.5. Every extension of ~ K . a l t l . D is axiomatizable by
formulae of the form Vi<m Oa'Pi ++ ~riPi for some finite sequences ai, ~-i.
Moreover, these axioms correspond in the class of differentiated frames to

(w)[ V (vy x)(Vz x)(y 9 z)].
i<m
COROLLARY 9.2.6. Each extension of polymodal K.altl is of special SaM-

qvist rank O. Every extension of polymodal K.altl.D is of pure, special SaM-
qvist rank O.
We can cash out on this characterization as follows. Call a statement of the
form
O~T A OfT A O~p <-->,Orp
a p a t h - e q u a t i o n , and denote it by a ~ T. A path equation states that both
paths exist and that they are equal:
(3y A (3y z)t A (vy )(Vz x)(y " z)
An extension of polymodal K . a l t l is then characterized by a disjunction of
conjunctions of path equations and constant formulae. The constant formulae
are either existence or nonexistence statements for certain paths. An extension
of ~) K . a l t l D is axiomatizable by a set of disjunctions of path equations.
Finally, recall that an elementary class of frames is modally definable
iff it is closed under contractions, generated subframes, disjoint unions and
ultrafilter extensions, while also the complement is closed under ultrafilter
extensions. In the present context the last restriction can be dropped.
PROPOSITION 9.2.7. (~ < NO.) A class of frames for logics of bounded

operator alternative is modally definable iff it is closed under contractions,
generated subframes, disjoint unions and ultrafilter extensions.
PROOF. We have to show that the class is inversely closed under ultra-
filter extensions. We show that g is isomorphic to a generated subframe of
uP(g), which will establish the theorem. The elements of g correspond to the
principal ultrafilters of uP(g). Thus we have to show that no principal ultra-
filter can see a non-principal ultrafilter. So let Ux be the ultrafilter of sets
containing x and let T be non-principal. The set of points seen in one step
by x via j is finite, consisting, say, of y 0 , . . . ,yn-1, where n _~ a. For each
i < n there exists a set ai such that yi C ai and ai ~ T. P u t b " - Ui<~ai.
Then b r T, since T is an ultrafilter. On the other hand, x C mjb. Thus
U~jT. D
COROLLARY 9.2.8. (~ < NO.) A class of frames of bounded operator al-
ternativity is modally definable iff it is closed under contractions, generated
subframes, disjoint unions and ultraproducts.
PROOF. Let :K be modally definable. Then it is elementary, since its logic
is r-persistent. It follows that :K is closed under ultraproducts. Conversely,
assume t h a t 9C is closed under ultraproducts. By Theorem 5.7.18, the ultra-

filter extension of a frame ~ is a c o n t r a c t u m of an ultrapower of g. Therefore
K is closed under ultrafilter extensions as well. [i]
The lattice of extensions of 8 (~)j<~ K . a l t l is quite complex despite the

fact t h a t the extensions are axiomatizable by such simple axioms. We will
restrict our attention to ~ = 2, but it is clear t h a t the results can be general-
ized.
PROPOSITION 9.2.9. ~ (K. altl | K. altl ) contains countably many tabular

logics of codimension 1. ~ (K. altl.D | K. altl.D) contains countably many
tabular logics of codimension 2.
PROOF. Consider the flames [--C~)C n defined on the set 2n + 1 where i <l j
iffi+l = j a n d i even, or i = 2 n - 1 a n d j = 2n, a n d i 4 j i f f j = i + l
and i is odd or i = 2n and j = 0. The frames [-cr)cn are almost like a cycle
where the relations between neighbouring elements alternate. There is one
position where this regularity is disturbed and this accounts for the fact t h a t
[-cr)cn has no proper contracta. Since it has no proper generated subframes
either, its logic is of codimension 1. In bimodal K . a l t l . D we have exactly
one logic of codimension 1, which is the logic of the one-point frame, which is
reflexive with respect to all relations. Now for this case take a prime n u m b e r
q # 2 and define on q the relations <l and 9 as follows. We let 0 <1 0 and
q - 1 9 q - 1. For even numbers i we put i 9 i + 1 9 i and for odd numbers
i we put i <~ i + 1 <~ i. No other relations hold. This defines the frame O-C~Cq.
Now define a function f : q --+ q by taking f(i) to be the <]-successor of the
9 -successor of i. It is verified in the case of q = 5 t h a t f(0) = 2, f ( 2 ) = 4,
f ( 4 ) = 5, f(5) = 3, f(3) = 1 and f(1) = 0. In general, we have fq(x) = x but
f~(x) # x for r < q. The powers of f form a cyclic group of order q under
composition. Now take a p - m o r p h i s m ~ of D-cr)cq onto a frame e. T h e n
the function f induces a function g on e defined as follows; g ( x ) : = ~ ( f ( y ) )
where y E ~ - l ( x ) . (By the conditions on p - m o r p h i s m s this is well-defined.)
Moreover, the powers of g under composition must form a group, too. For g is
invertible, since gq(x) - x. However, the cyclic group of order q has only two
images, itself and the group of order 1. In the first case p must be injective,
in the second case e has one element. For O-c~cq is connected under f , t h a t
is, for any two elements i , j there exists an r such t h a t j = f~(i), e must
have the same property with respect to g. So e has one element, and thus the
theory of 0-cr)cq has codimension 2. [:]
Since the finite frames are countable for finite ~, this is the best possi-
ble result. Since all extensions of polymodal K . a l t l are complete this leaves
us with the possibility of maximal consistent logics which are complete but
not tabular, t h a t is, which are determined by a single, infinite frame. The
key to such logics lies in the existence of infinite sequences over a given al-
phabet which are nonrepeating. These are sequences in which no subword is
eventually repeated over and over. By a simple cardinality argument there
are 2 ~~ sequences which are nonrepeating. However, the result t h a t we are
going to prove is much stronger; it shows that there are 2 ~~ many infinite se-
quences over the alphabet {0, 1} such that no subword is repeated five times
consecutively.
DEFINITION 9.2.10. For a n a t u r a l n u m b e r k and a f i n i t e sequence cr, let

k x cr denote the sequence consisting of the k - f o l d repetition o f ~r. L e t a : w -4
{0, 1} be an infinite sequence and cr = s o s i . . . S n - i a f i n i t e s t r i n g over {0, 1}.
The i n d e x o f cr in a , Ha(or) is the m a x i m a l ordinal n u m b e r j < w + 1 such
that a c o n t a i n s a subword o f the f o r m j x or. T h a t is to say, H a ( e ) = j i f f
there is an r such that f o r all i < j and e < n, a ( r + n . i + e) = se.
For an infinite sequence a let g~ be the frame based on w and i <] j iff
j=i+landa(i)=0andi 9 If H a ( s ) = w, t h e n
there is a point from which the sequence a repeats the sequence cr periodically.
P u t E := { a : H a ( a ) E w , f o r all cr E {0, 1}*}. Thus E is the set of sequences
in which every subword has finite index. It is clear that E has the cardinality
2 ~~ It is the logics of ga for a E E that we want to use as examples of logics
of codimension 1. However, there are two obstacles. The difficulty lies in
showing that (i) they are of codimension 1 and (ii) that they all give rise to
distinct logics. We will solve these problems as follows. We show first that
even if the theory of g~ is not maximal, it has no finite frames. The second
problem is solved by constructing special sequences, in which even Ha (or) _< 4
for all (r.
LEMMA 9.2.11. L e t cr be a n o n - e m p t y f i n i t e string with i n d e x m . Then

ga ~ ~0(m+i)X~T.
LEMMA 9.2.12. L e t ga be a f r a m e such that a E E a n d let f be a f i n i t e

frame. Then Th ga ~ Th f.
PROOF. Let ~ - (f, <~, 4}. It is enough to show this for such finite f for
which the theory is of codimension 1. In that case [ has no proper generated
subframes. C a s e 1. [ is the one-element frame with empty relations. Then
0TVeT E Thga-Thf. C a s e 2. Not both <~ and 9 are empty. Let [
have n elements. Then f is a full cycle, that is, there is a sequence cr -
sosi...sn-1 over {0, 1} and a world w0 such that wi+i - s i ( w i ) , i < n - 1
and w0 - crn-i(Wn-1). P u t m := Ha(or). Then -~0(m+l)X~T E T h g a - T h [ ,
as required. [3
DEFINITION 9.2.13. Let X C w. We define a sequence a x to be the con-

c a t e n a t i o n o f the words z nz , which are defined as follows. W e p u t z nx " - 1nx r nx .
1x and r nx -- or simply In and rn from now on - - are defined inductively by
10 := O, r0 := ~, if ocx,
10 := 00, r0 := 11, if 0 C X,
ln+l "-- Inrn, rn+l "-- rnln, /fn+l CX,
ln+l "-- lnrnlnrn, rn+l := rnlnrnln, if n + 1 C X .
In order to u n d e r s t a n d this definition, let us note some properties of this
sequence. First, it can be divided from left to right into blocks of equal length
of the form loro or rolo, and starting from the first occurrence of ln, it can be
divided into blocks of the form l n r n and rnln. These blocks have length 2 k for
some k. Let us define a new sequence from a z by forgetting z x and dividing
the remaining sequence into parts of the form loro and rolo, as indicated.
The sequence loro is replaced by 0 (its first symbol) and rolo is replaced by 1
(again its first symbol). The resulting sequence is a sequence of the form a y
where Y - {n 9 n + 1 C X}. The proof is by induction. So, if the sequence
taken from a certain point onwards and is thinned by taking only the 2 k-
next symbol, we get a sequence of type a y . In particular, if 0 C X t h e n we
may replace in the entire sequence blocks of the form loro - 0011 by 0 and
blocks of the form rolo - 1100 by 1. T h e n we get a sequence of the form
ay with Y - {n " n + 1 C X}. The point where the sequence ln rn a p p e a r s
for the first time, will be denoted by On. Furthermore, if 0 ~ X t h e n the
subsequence ( a z ( 2 n ) , a z ( 2 n + 1)) contains a zero and a one, and if 0 e X
then ( a z ( 4 n ) , a x ( 4 n + 1 ) , a x ( 4 n + 2 ) , a z ( 4 n + 3 ) ) i s either 0011 or 1100.
Thus, in any subsequence the n u m b e r of zeros may differ from the n u m b e r of
ones by at most four. (For each subsequence of length can be divided into a
sequence g . ~0 9U~l . . . . " V~n-1 " g' where the u~i are equal to 0011 or 1100,/7 a
postfix of 1100 or 0011 of length < 3, and g' a prefix of length < 3 of 0011 or
1100. Each u~i contains two ones and two zeros, so it is balanced. I n / 7 and g"
the n u m b e r of ones and zeros can differ at most by 2.)
LEMMA 9.2.14 (Grefe). For all X C_ co and all finite strings cr Hax (or) <
4. In particular a x E E.
PROOF. Let Ing(a) denote the length of a string. Let cr have the length
2 k . u , u an odd number. Suppose, a z contains the subsequence 5 • or. C a s e
1. T h e r e is an n such t h a t Ing(lnrn) - 2 k. Let 5 • cr begin at the n u m b e r r
and let t C co the unique number such t h a t on + ( t - 1). 2 k < r < On + t" 2 k.
(This n u m b e r exists since on < lng(lnrn) - - 2k.) T h e n a x ( o n + (t + i ) .
2 k) -- a z ( o n + i + j ' u ) ' 2 k) for a l l i < Ing(s) and j < 5. We m a y t h e n
form the sequence b ( i ) " - a(On + 2 k " i) and get b - a y for some Y. T h e n
(b(t), b(t + 1 ) , . . . , b(t + u - 1)) has uneven length. So it has one more zeros
t h a n ones ore one more ones t h a n zeros. Since this sequence repeats five times
consecutively, we have found a subsequence in which the n u m b e r of ones and
9.3. Colourings and Decolourings 475
zeros differs by at least 5, in contradiction to the fact t h a t b is of the form

ay. C a s e 2. There is no n such that lng(lnrn) = 2 k. Then there is an n such
that Ing(l,~rnlnr~) - 2 k. Now proceed as in Case 1. [-7
THEOREM 9.2.15 (Grefe). The lattice ~ ThcO4| r has 2 ~~ many log-

ics of codimension 1.
PROOF. The monomodal fragments of Th g~x contain the theory of the
four-element chain, since neither 0 nor 1 may be iterated five times. Th ~ x
is contained in a maximal consistent logic. It is enough to show t h a t for
X r Y, Th O~x U Th ~ar is inconsistent. Let n be the smallest number on
which X and Y disagree, say, n C X - Y. Now, assume t h a t n - 0. Then
g~r ~ DE]D_I_. However, r2x contains four consecutive zeros. Moreover, 1x
may not be repeated more than four times, so that wherever we are, after at
most k = 5. lag(1 X) steps r2x must occur, and hence a sequence with three
consecutive zeros. Thus 9 ~ ~ ~ []<k 000-[-. For n - m § 1 the proof is
performed with the subword Im in place of 0 and rm+
2 X in place of r2x . [-1
Notes on this section. In EDITH SPAAN [202] is is shown t h a t K . a l t l is lo-

cally N P - c o m p l e t e and globally PSPACE-complete, while polymodal K . a l t l
is locally N P - c o m p l e t e and globally E X P T I M E - c o m p l e t e .
E x e r c i s e 330. W i t h this set of exercises we will shed some more light on the
results by WIM BLOK of the preceding chapter. First, show t h a t any tabular
n - m o d a l logic is co-covered by 2 ~~ complete logics, n > 1. Hint. Use the fact
that the inconsistent logic has this property.
E x e r c i s e 331. Show that any tabular monomodal logic contained in Th [ ~

is co-covered by 2 ~~ complete logics. Hint. Use the simulation theorem.
E x e r c i s e 332. Let the degree of nonfinite model p r o p e r t y of a logic A be

defined to be the cardinality of the spectrum of logics with the same class of
finite models as A. Show that a tabular n - m o d a l logic has degree of nonfinite
model property 2 ~~ unless n = 1 and A = K1 9 •
E x e r c i s e 333. (SPAAN [202].) Show that polymodal K . a l t l is NP-complete.
9.3. Colourings and Decolourings

A bimodal flame in which each operator satisfies a l t l can be seen as
a kind of monomodal alt2-frame in which the successors are discriminated
by the modalities. We can define two operations on these frames going in
opposite directions; we call them colouring and decolouring. Decolouring is
the process of forgetting the distinction between the relations, and colouring
is the (re)introduction of the distinction. Any monomodal Kripke-frame can

in principle be treated in this way, given enough additional relations. If each
point of the frame has at most ~ many successors then given ~ many operators
the successors can be distinguished by the new accessibility relations. Each
of the new accessibility relations satisfies K . a l t l and may thus be interpreted
as a partial function on the frame. In this way, modal logic seems generally
reducible to the study of (polymodal) K . a l t l . Moreover, alternative semantics
in terms of functions on frames rather than relations as proposed e. g. in
VAN BENTHEM [13] can be studied as well. Therefore, in this section we
will show how to perform such a reduction and where its limits are. Basically,
the limitation is to logics which are of bounded alternativity. The reason is
that otherwise the logics are not necessarily complete and the operation of
colouring may not be definable on enough frames to guarantee t h a t the logic
is recoverable from its coloured counterpart.
There are two approaches to colouring and decolouring. One is to view
this as a relation between bimodal and monomodal frames. The other is
to view it as a relation between bimodal and trimodal frames. In the first
approach the monomodal frame contains the relation that is 'split' in its
bimodal coloured version. In the second approach the three modalities are
merged. So, the three-modal frame is obtained by a fusion of the bimodal
coloured frame and its uncoloured variant. (This has been brought to my
attention by KIT FINE.) Both approaches have their merits. P r i m a facie
colouring seems to be an operation that is always defined. Unfortunately,
we have to introduce two restrictions. Namely, we must restrict ourselves
to Kripke-frames, and moreover to Kripke-frames in which a point has a
<~-successor iff it has a ,-successor.
DEFINITION 9.3.1. Let ~ - (f, <~,,) be a Kripke-frame for K . a l t l |
g . a l t l ( ~ T ++ CT). Then D(~:) := (f, <]U , ) is called the d e c o l o u r i n g of
[. Conversely, if g - (g, <~) is a monomodal Kripke-frame then any f such
that O(f) - 9 is called a c o l o u r i n g of 9. We denote by C(9) the set of colour-
ings of 9, and for a class 9( of monomodal Kripke-frames, C K := D-I[K].
Let us first present an example of an uncolourable frame. Take the set
f-w x 4. Let ~ - ( f , ~<) where
~< := {((n,O), ( n , O ) ) ' n C w}
m {((n, 0), ( n , l > ) ' n e w}
m {{(n, 2>,(n, 3 > > ' n e w}
U {((n,2>, <n+ 1,3>>'n e w}
P u t Y " - (w • {0})U(w • {2}) and X := (w • {1}) • (w • {3}). The sets X and
Y are definable by constant formulae, for example 7 t l and @ V] _1_. Finally,
let Y be the algebra of sets generated by the finite sets. This is exactly the
algebra A of sets a such t h a t a 71X is finite or cofinite in X and a n Y is finite

or cofinite in Y. To show t h a t A - F, let us first note t h a t A contains all finite
sets. Moreover, A C IF because if we have a finite set a, we have a D X E
and a n Y E F, as well as Z A ( - a ) and Y n ( - a ) . So we can compose any
set which is finite or cofinite in X and finite or cofinite in Y. IF C A is clear.
It is also routine to check that A is closed under the n a t u r a l operations.
PROPOSITION 9.3.2. ~ "-- ([, A) is not colourable.
PROOF. Let (f, <l, 4, A) be a colouring of (f, ~<, A). T h e n ~< = <l U 4.
Any point in w x {2} sees exactly one point in w x {3} in each direction.
Hence the sets (}(w x {3}) and 0(w x {3}) are both cofinite in w x {2}. But
(}(w x {1}) and 0(w x {1}) cannot both be cofinite in w x {0}. W i t h o u t
loss of generality we assume that w x { 0 } - (}(w x {1}) is infinite. Then
(}X = (}(w x {1}) U (}(w x {3}) is neither finite nor cofinite in Y, and so not
in A. Thus no colouring exists on ~. [D
The next question is the restriction to frames which validate the ax-
iom 0 T ++ (}T. If we are interested in the mere operations of colouring
and decolouring such a restriction is strictly speaking unnecessary. However,
consider the monomodal frame m := ({0,1}, <) where ~< := {(0,1}, (1,1}}.
Consider the bimodal frame b := ({0,1}, <l, 9 where <l = {(0,1/} and
9 = {(1, 1/}. Then m has a one-point contractum, but b does not. This is
not a good situation; we want that a colouring of m must be a frame c such
t h a t each p - m o r p h i c image n of m can be coloured in such a way t h a t it is a
c o n t r a c t u m of c. This gives rise to the restriction t h a t a point must have a
<]-successor iff it has a 9
DEFINITION 9.3.3. Let A be an extension of K.alt2. Then A, defined by

A c := Th (C [Krp A]) is called the c o l o u r i n g o f A. Let 0 be an extension of
the logic
K . a l t l G (}T ++ O T .
2
Then 0 d "- Th (D [Krp O]) is called the d e c o l o u r i n g k e r n e l o f O.
In this section we will be concerned with the algebraic properties of these

maps, with a syntactic description of the operations of colouring and decolour-
ing as well as the properties of the corresponding semantic maps. First, as it
is easy to see, colouring and decolouring are isotone maps between t h e poset
reducts of the lattices of logics.
THEOREM 9.3.4. The colouring map (as a map between posets) is left
adjoined to the decolouring kernel map. Moreover, for every extension 0 of
K. alt2 , 0 - - 0 cd.
478 9. L o g i c s of B o u n d e d A l t e r n a t i v i t y
PROOF. We have DD -1 [K] -- K for all classes of monomodal Kripke-

frames. Putting 9(: "- KrpO shows | - o~d. For by Kripke-completeness,
O - Th Krp O. Let A be a bimodal logic, O a monomodal logic. Then A D |
iff Krp A C_ Krp O c - C [Krp O]. The latter implies D[Krp A] C_ DD -1 [Krp O]
and so D Krp A C_ Krp O, since DD -1 [9(:] - * for all classes of Kripke-frames.
The latter is equivalent to A d D O. It is checked that the converse holds as
well. []
The connection between frames and colourings is introduced by the fol-
lowing translation.
DEFINITION 9.3.5. The map c " [Pl --+ ~2 is defined by
pC 9--- p
9- ~o~ A r
9 ---- --1(10 c
:=
Moreover, A c := { ~ c . ~ C X} f o r a set A C T1.
PROPOSITION 9.3.6. Let b be a K r i p k e - f r a m e f o r K . a l t l | K . a l t l ((}X ++

0 T ) , x C b and 13 a valuation on b. T h e n
(b,/3, x} ~ cp~ ~ (D(b),/3, x) ~
Moreover, b ~ ~c iff D(b) ~ ~.
PROOF. By induction on ~p. Only the step for ~ - (~r is nonobvious.
iff (b,/3, x) ~ (}r or (b,/3, x) ~ 0 r c

iff for s o m e y C > x (b,/3, y ) ~ r (b,/~,y)~r
iff for some y(c>U , ) x (b,/3, y) ~ r
iff for s o m e y > / x (D(b),f~,y)~r
iff (D(b),/3, x) ~ ~.
The second claim follows immediately. If b ~ ~ then (b,/3, x) ~ (-~)~ for

some g, x. Hence (D(b),/3, x) ~ - ~ and so D ( b ) ~ ~. And conversely. [:]
The syntactic correlate of colouring of a bimodal logic is also rather easily
defined.
PROPOSITION 9.3.7. Suppose A - K.alt2 @ A. Then A c - ~)2 K . a l t l @
GT ~ OT @ AC.
PROOF. Let m be a monomodal frame for K.alt2. From Proposition 9.3.6

we deduce that m ~ f iff C(m) ~ ~oc. Alternatively, rn E Krp K.alt2 @ ~o iff
C(rn) c_ Krp (~)2 K . a l t l @ 0 T ++ 0 T | ~;c. From this the theorem follows, r-3
THEOREM 9.3.8. The colouring map is a homomorphic embedding of the

locale E K.alt2 into the locale E ((~2 K . a l t l | 07- ++ 07-).
PROOF. Using the results of Section 2.9 it is not hard to see that colouring
commutes with (infinite) joins. To see that it commutes with meets, suppose
that A1 - K . a l t l | A1 and A2 - K.alt2 @ A2. Then
A~ ~ A~
= (~)2 K . a l t l | 07- ++ 07- | A~) R (~)2 K . a l t l 9 07- ++ 0 T (3 A~)
= (~2 K . a l t l | 07- ++ 0 T @ {[]---mqa V []<_me. ~ E A~, r E A~, m ~ o.)}
= (~)2 K . a l t l | 0 T ++ 0 T | {(~;~<_m~ ~/ ~]_<mr . ~ ~ i l ' r ~ A2 ' m ~ 03}
= (A1 N A2) c
(Note that the definition of ~ depends on the language. Hence the step from
the second to the third line is correct.) Finally, we have to show that the
colouring is injective. So, let A " - K.alt2 | A and 0 " - K . a l t 2 @ A' and
A r O. W i t h o u t loss of generality there exists a Kripke-frame m such that
m ~ A, but m ~ ~ for some ~ C A'. Then there exists a colouring b of m such
that b ~ r b ~ A ~. This shows A c # 6)c. [:]
PROPOSITION 9.3.9. Let 0 be an extension of ~ 2 K . a l t l @ OT ++ 07-.

Then O d - {qa 9 qa~ C O}.
PROOF. 99 E 6) d iff qa C Th D[Krp 0] iff for all b C Krp 0 we have D(b) ~ qo

iff for all b C Krp 19 we have b ~ ~ iff r C 19. V]
THEOREM 9.3.10. The decolouring operation D commutes with the class

operators of taking generated subframes, contraction images, disjoint unions
and ultraproducts.
The proof is routine and left as an exercise. We note that from Corollary 9.2.8
we derive
COROLLARY 9.3.11. For any class K of bimodal Kripke-frames for the

logic ~ 2 K ' a l t l | OT ~-~ OT it holds that D M K - M D K. Here, M iK "-
Krp Th :K.
THEOREM 9.3.12 (Grefe). The colouring map map leaves invariant f i n i t e

a x i o m a t i z a b i l i t y and local finite model property.
PROOF. Finite model property is clear from the definition and the con-
struction of decolouring. Moreover, if A is finitely axiomatizable, so is A c, by
Proposition 9.3.7. For the converse, let A c be finitely axiomatizable. Then
Krp A ~ is an elementary class. Therefore Krp A = D Krp A c is an elementary
class, since D commutes with Up, by Theorem 9.3.10. [:3
THEOREM 9.3.13. C commutes with taking generated subframes and dis-

joint unions of Kripke-frames. Moreover, for contraction images Cn, Cn C :K c
C Cn :K.
PROOF. G e n e r a t e d S u f r a m e s . Let g be a K r i p k e - f r a m e , a n d ~ ~-, g.
T h e n any colouring of g induces a colouring of ~. Any colouring on ~ can be
e x t e n d e d to a colouring of g. D i s j o i n t U n i o n s . Let gi be K r i p k e - f r a m e s ,
i C I. Pick a colouring [~i on each gi. It is easy to check t h a t (~i~I[~i
is a colouring of ( ~ i ~ I gi. Also, every colouring on ( ~ i ~ I gi gives rise to a
colouring of each individual gi. C o n t r a c t i o n I m a g e s . Assume t h a t g and [~
are m o n o m o d a l K r i p k e - f r a m e s , and ~ 9 g -~ [~. Let ~)c be a colouring of [~. We
have to show t h a t there is a colouring gC of g such t h a t ~ " g c -~ t~~. So let
x, y C g and x ~ y. we have to decide w h e t h e r x <l y or w h e t h e r x 9 y. T h r e e
cases arise. (i.) x has a single successor. This successor is y. T h e n we p u t
x <~ y and x 9 y. (ii.) x has two successors y and z and 7~(y) r ~(z). T h e n
x <] y iff ~(x)<] ~(y) and x 9 y iff ~(x) 9 ~(y); and similarly for z. (iii.) x has
two successors y and z but ~(y) - ~(z). In this case we choose as to w h e t h e r
x <~ y and so x 9 z or x 9 y and so x <] z. W i t h this definition m a d e , let us
show t h a t ~ " g ~ --~ [~. So, let x < ~ y . T h e n x ~ y and so ~(x) ~ ~(y). It
r e m a i n s to be seen t h a t 7r(x) <] 7r(y). If ~v(y) is the only ~ - s u c c e s s o r of ~v(x),
t h e n ~(x) <] ~(y) by virtue of definition of a colouring. So a s s u m e t h a t ~ ( x )
has two successors, ul and u2. By construction, if x <~ul t h e n ul - ~(y), and
so ~ ( x ) <] ~(y). Similarly the case x 9 y is dealt with. This shows the first
condition on p - m o r p h i s m s . For the second condition, assume ~(x) <] u. T h e n
~(x) ~ u and for some y we have x ~ y and ~(y) - u . T h e n x 9 y or x <~ y.
Clearly, x has a ~ - s u c c e s s o r . C a s e 1. x has exactly one successor. T h e n
b o t h x 9 y and x < ] y and we are done. C a s e 2. x has two successors and
their images are distinct. T h e n x <] y by definition of <~. C a s e 3. x has two
successors y~, Y2, and b o t h are being m a p p e d onto u. T h e n we have x <~ y~
or x <] Y2, by construction. This concludes the proof of the fact t h a t ~ is a
p-morphism. [::]
LEMMA 9.3.14. Let ~ be a class of Kripke-frames for K . a l t 2 . Then

UpC~ c_ CUp~.
PROOF. Let gi, i E I, be K r i p k e - f r a m e s in 9<:, and choose a colouring g~
of each frame. Let U be an ultrafilter on I. We want to show t h a t 1-[u g~ is
isomorphic to a colouring of an u l t r a p r o d u c t of the g~. We take ~) " - 1-[u gi.
We let x u <] Yu iff { i ' x i <~yi} ~ U. This definition does not d e p e n d on the
choice of the representatives. Moreover, it shows t h a t the identity m a p is an
i s o m o r p h i s m from 1-Iu g~ onto ~)~. []
THEOREM 9.3.15 (Grefe). Let :K be a modally definable class of Kripke-

frames for K.alt2. Then C ~ is a modally definable class of Kripke-frames
for K.alt~ | K.alt~(~T ~ O-V).
PROOF. By the previous theorems, C[K is closed under generated sub-

frames, contractions, disjoint unions and ultraproducts. By Corollary 9.2.8
we get that M C 9C C C M K, since M consists in taking generated subframes
of contractions of ultraproducts. If 9C - M K then C ~ C M C 9C C C M , 5C -
C JC. And so equality holds. [-]
THEOREM 9.3.16. Th Up C ~ - Th C Up K.
PROOF. The inclusion 'D' follows from Lemma 9.3.14. For the other
inclusion assume ~ r Th C Up K. Then for some gi there exists a colouring
l~ of 1-Iv iti such that l] ~ ~, say ([~,g, 5u) ~ - ~ . Let 5 be the modal depth of
~. Now look at the (f-transit of 5u in ~. It is finite, and hence
A := (i 9 c u
For each i C A fix an isomorphism Li from the (i-transit of xi in fti onto the
5-transit of x u in ~. We colour the 5-transit of xi in 9i in the way prescribed
by ~. That is, we put xi <] yi iff ~,~(xi) <] ~,~(yi). This defines for each i C I
a partial colouring (if i r A, then nothing is prescribed so far). Choose any
colouring on the 9i that extends the partial colouring. This defines fl~. Then
let t~ "- riG 9~. It is routine to show that the (f-transit of x u in t~is isomorphic
to the (f-transit of xu in t~. Hence t~J~ ~. And t~ E Up C ~K. [-]
THEOREM 9.3.17. For any class • of K . a l t 2 - K r i p k e - f r a m e s
MC~C- CM~C
PROOF. Both are modal classes. And their theory is identical. Hence
they are identical. D
E x e r c i s e 334. Suppose A c is defined on the basis of an axiomatization as in

Proposition 9.3.7. Verify syntactically that this definition is independent of
the chosen axiomatization of A.
E x e r c i s e 335. Let 0 be an extension of ~)2 K . a l t l @ 0 T ++ 0T. Show

syntactically that {~" ~pc C O} is a logic.
E x e r c i s e 337. Show with a particular counterexample that C Cn K C_ Cn C K

does not hold in general.
* E x e r c i s e 338. Generalize the results of this section to polymodal logics.

Show that extensions of n-modal K . a l t l can be interpreted as extensions of
K.altn.
482 9. L o g i c s of B o u n d e d Alternativity
9.4. Decidability of Logics

Consider a finite set A of symbols, called a l p h a b e t . By A* we denote the
set of finite strings over A, including the empty string, denoted by c. Strings
will be denoted by a vector arrow, e. g. ~, ff etc. An e q u a t i o n is a pair
(v, w/ C A* • A*, written g ~ u~. A T h u e - p r o c e s s over A is a finite set of
equations. Given a Thue-process F we write ff ~ % iff there exist ~, d C A*
and g ~ ~ E F s u c h t h a t i f - f f . g . d a n d % - 5 . ~ . d o r if- ff.~.dand
Z - ff.g.d. We say also that 5' is o n e - s t e p d e r i v a b l e from ft. Thus, Z is one-
step derivable from ff iff it can be produced from ff by replacing a substring
matching one side of an equation in F by the other side of that equation. We
define ff ~ 5' inductively by ff ~o Z iff f f - Z and ff ~ + 1 F iff there exists
such that ff ~ ~ ~ %. Finally, ff ~ % iff there exists an n C w such t h a t
~ 2'. We say that % is n - s t e p d e r i v a b l e f r o m ff if ~ ~ ~ and t h a t it
is d e r i v a b l e if ff ~ ~. A Thue-process is d e c i d a b l e if for given if, % it is
decidable whether or not % ~ ff is derivable in it. A Thue-process is called
trivial if 2 ~ ~ is derivable for all 2 C A* iff a ~ e is derivable for all a ~ A.
The following facts will be used in sequel.
THEOREM 9.4.1 (Post, Markov, Rabin). Let A contain at least two sym-
bols.
1. There exists undecidable Thue-processes.
2. The set of decidable Thue-processes is undecidable.
3. The set of trivial Thue-processes is undecidable.
These statements were shown in [1631, [156] and [167], respectively.
Proofs of these fact can also be found in many textbooks. A Thue-process can
be seen as a presentation of a semigroup. Recall from Section 1.3 the notion
of a presentation. Let SG the theory of semigroups in the language 1, .. For
example, SG - {x ~ x. 1, x ~ 1 . x , (x. y). z ~ x. (y. z)}. The free semigroup
over A, ~tsG(A) is isomorphic to (A*, c, .>. Given a Thue-Process P, (A, F> is
a presentation of a semigroup denoted by ~ t s c ( A ) / F , and we have ~ ~ ff iff
and ff are identical elements of ~tsG(A)/F. This is the most standard way
of thinking about Thue-processes. We can derive two rather useful results.
First, take the first-order theory of two unary functions, f0 and fl. These
functions are interpreted on a semigroup generated by two elements, say 0
and 1, as concatenating a word with one of the two generators. Namely, we
put f o ( x ) " - x.O and fl ( x ) " - x. 1. Since 0 and 1 generate this group, for any
word ff the function f~ 9 x ~ x. ff can be composed from these two functions.
Hence an equation ff ~ g can be rewritten into the statement
Now, a semigroup satisfies the equation z~ ~ ~ iff it satisfies the first-order

condition (Vx)(f~(x) " f~(x)). Thus, the first-order theory of the semigroup
9.4. Decidability of Logics 483
generated by a Thue-process P is exactly axiomatized by the the axioms (1:)

for (~7,~) G ~. So, we can encode the derivability of the Thue-process into
first-order logic over two unary functions. The unary function f0, on the other
hand, can be viewed as a binary relation <~0 which satisfies two postulates. (i)
(Vxyz)(x <~oy A x <~oZ. -+ .y -- z) and (ii) (Vx)(3y)(x <~oY). Similarly with fl.
Hence we conclude that the V3-theory (i. e. the set of sentences of that theory
which are of complexity at most VS) of two binary relations is undecidable.
Given a Thue-process we build a canonical process-frame c~ as follows. ~ is
a congruence on A*. The equivalence class of ~ in this congruence is denoted
by [~]. So, [~] = {~: ~ ~ ~}. Then c~ := {[~]: ~ e A*}. We put A := {w, b}
(w stands for white and b for black.)
A*/P 9- {[2] 9 2 E A*}

<3 9= {([~, [Z. w]}" ~'E A*}
9= {([z-I,[Z. b])" ~C A*}
cp 9= (A*/F,<~, 9
In what is to follow we will not always distinguish between sequences and
the equivalence classes of these sequences modulo IP. Fix a sequence ~. The
map 7r "[g] ~ [~]. [~ is a p-morphism of c~ onto the transit of [~]. For if
[~] <J [~ then [~ - [~. w], and then also [~]. [~ - [~]. [~7" w], from which
7r([~) - [~]. [y~ <~ [ ~ ] - [ ~ - 7r([~). Next, if 7r([~) - [~]. [~ <l [~7], then
[d] - [ 2 ] . [~. w]. So, let [g] "- [~. w]. Then 7r([g]) - [2]. [~. w ] - [~7]. And
[~] <~ [d]. Similarly for 9
With a Thue process we associate two logics, namely
E~ - K . a l t l | K . a l t l 9 {0~p ++ K]~p 9 g ~ u7 C P}
A~ - K.altl.D @ K . a l t 1.D @ {()~p e+ K]~p 9 g ~ u~ C P}
= Ee | 0 T | 0 T
The following theorem is easy to prove with the help of Theorem 3.5.3.
PROPOSITION 9.4.2. Let P be a Thue-process. E~ is a subframe logic. It

is of (pure) Sahlqvist rank 0 and its frames are characterized by the properties
.., .., .
(Vx)(Vy )(Vz z)(y z), e

Ee has the finite model property and is locally decidable.
PROPOSITION 9.4.3. Let g be a rooted Kripke-frame for Ae. Then there

exists a contraction r --~ g.
PROOF. Let w0 be the root of g. Then let s 9 A* --+ g be defined by

s(~) "- w0. Further, s ( ~ . w) is the unique element y such that s(~)<~ y.
Similarly, s ( ~ . b) is the unique element y' such that s(~) 9 y'. It is checked
that if ~ ~ ~ then s(~) - s(g). Hence, the function defines a function
484 9. Logics of B o u n d e d Alternativity
t([~]) := s[([~])]. It is readily verified that t is a p-morphism. It is onto by

the fact that g is rooted at w0. [:]
PROPOSITION 9.4.4. Th (c~, [~]) = Th ca = Aa.
PROOF. We know that Th (ca, [~]} _~ Th (c~, [r from the fact that the
transit of [~] is the p-morphia image of c~. Hence Th (ca, [c]) is normal and
identical to Th c?. Now, A? and Th ca are extensions of K . a l t l . D | K . a l t l . D .
First of all, ca is a frame for Aa. Hence Th c? ~ A?. Also, let g be a rooted
Kripke-frame for A?. Then 9 is a contractum of c?. So, Th c? and A? have the
same rooted Kripke-frames. Hence being c o m p l e t e - they are identical.
[]
THEOREM 9.4.5. Aa is decidable iff F is decidable.

PROOF. Suppose that IP is undecidable. Then the problem '~ ~ ~' is
undecidable. Hence the problem '0~p ++ 0Yp C Aa' is undecidable (since it is
equivalent to 'Z ~ y-'). Now suppose that IP is decidable. Given a formula
~, we want to be able to decide whether it is satisfiable in a A~-frame. By
Proposition 9.4.4, it suffices to be able to decide whether p is satisfiable
in (c~, [e]). It is possible to convert ~ algorithmically into a disjunction of
formulae X which have the form
i<n
where xi C A* and #i a conjunction of variables or negated variables. (More-

over, no variable occurs both simply and negated.) It is enough if we are able
to decide whether X is satisfiable in (c~, [e]). Now consider a valuation 7 such
that (c~,7, [e]) ~ X. Then (c~,7, [xi]) ~ #i, for a l l / < n. And conversely. 7
exists iff there do not exist i and j and a variable p such that [~i] - [a3j] and
p is a conjunct of #i, -~p a conjunct of pj. Since ~ is decidable, the problem
'[:~i]- [:~j]' is decidable. Hence we can decide whether 7 exists. [~
We can extract the following characterization of Halld~n-completeness.
THEOREM 9.4.6 (Grefe). A n extension A of bimodal K . a l t l is Hallddn-

complete iff
1. h - K(D_L) | g([~_l_) or
2. there are s, t C w such that A = K ( [ 3 • | Th cr)cs,t or A = Th cr)cs,t |
K ( [ 3 Z ) or
3. A = (possibly infnit ) Thu -p oc .
PROOF. Since Halld~n-completeness transfers under fusion the logics of
the first two types are Halld~n-complete. The logics of the third kind, how-
ever, are also Halld~n-complete since they are determined by a single matrix
by Theorem 1.6.5. This matrix corresponds to (c~, [el). Now let A be a logic
9.5. Decidability of Properties of Logics I 485
containing bimodal K . a l t l . If it contains the axiom [::]_[_, then it is of the

form K . a l t l ([:]_1_) | 0 with O Halld~n-complete. Hence, it is of the first or
the second type. Likewise if it contains the axiom m• If neither is the case,
A is an extension of bimodal K . a l t l . D . Let p be an axiom for A. We can put
it into canonical form; it is then a disjunction of formulae disjoint in variables.
Hence, by Halld~n-completeness the disjunction is trivial, and so ~ is actually
a conjunction of path-equations. And so A is of the form A~ for some possibly
infinite P. [-1
A~ extends E~ by two constant axioms, E~ does not have the global finite
model property, by constructive reduction. Similarly for decidability.
PROPOSITION 9.4.7. Let P be a Thue-process. (i) I f P presents an infinite
semigroup then E~ fails to have the global finite model property. (ii) If P is
undecidable then E~ is globally undecidable.
E x e r c i s e 339. Show that it is undecidable whether or not given a Thue-

process ]? the semigroup presented by P is finite.
E x e r c i s e 340. Show that it is undecidable for every n whether the semigroup

presented by P has _ n (exactly n) elements.
E x e r c i s e 341. Generalize Theorem 9.4.6 to arbitrarily many operators.
E x e r c i s e 342. Show that the first-order theory of a single binary relation is

undecidable. Hint. Use the modal simulation theorem.
E x e r c i s e 343. Let P be a Thue-process. Show that E~ is in NEXPTIME.
9.5. D e c i d a b i l i t y of P r o p e r t i e s of Logics I
In the remaining sections of this chapter we will discuss questions of
decidability of properties of logics. Recall that we study these questions in
the following general setting.
DEFINITION 9.5.1. Let P be a subset of ~ K~. P is decidable if for every
finite set A of formulae in ~,~ the problem 'K~ G A C P ' is decidable.
In general sets P will be determined by certain properties. So, we will
say that a property of logics is u n d e c i d a b l e if the corresponding subset of
the lattices have that property. It will turn out that properties of logics are
undecidable in the overwhelming number of cases. There might be a general
reason for this, but right now we will just walk through a number of properties
and discuss their decidability. The first property we will discuss is identity
to a given logic, and related to it inclusion in a given logic. Recall from
Section 7.1 the following theorem.
486 9. Logics of Bounded Alternativity
PROPOSITION 9.5.2. A is decidable iff the problem 'K~ @ ~ C A ' is de-

cidable. In other words, A is decidable iff the set S A is a decidable subset of
~K~.
THEOREM 9.5.3. For monomodal logics, c o n s i s t e n c y is decidable.
PROOF. By Makinson's Theorem, A - K | l iff A ~ Th[~-~ and A

Th [7]. The latter is decidable. [i]
This is an exceptional fact of monomodal logic, just in the same way as Makin-
son's Theorem is unique for monomodal logics. Consistency is undecidable as
soon as we have two modal operators. The way to see this is as follows. First,
by the results of the previous section it is undecidable whether a bimodal
logic is equal to Th I olo I' where I olo l is the one-point bimodal frame which
is reflexive in both relations. For example, take the logics E~ corresponding
to Thue-systems. The logics E~ can have the following one-point frames as
models, nono l, I ~1761, I ~1761, I ~1761, standing for the one point frame with the
two relations being either reflexive or irreflexive. Now, P ~ w ~ e iff neither
I.I. I nor is a o for Seine
o := Th I.I. In Thl.lo I
O is a s u b f r a m e logic. T h e n P ~ w ~ e i f f | is inconsistent. It is not
decidable whether P > w ~ e, otherwise it is decidable whether a Thue-process
is trivial. Notice that 0 U E~ is a subframe logic.
THEOREM 9.5.4. (n :> 2.) Decidability is undecidable. Moreover, con-

sistency of elementary subframe logics is undecidable. (t~ - 1.) { T h [ ~ } is
undecidable.
This theorem has a number of consequences. The first concerns the prop-
erty of being a subframe logic. The argument is shown for n = 1, but can
easily be lifted to a logics with several operators.
COROLLARY 9.5.5. It is undecidable whether a given logic is a subframe

logic.
PROOF. (Version 1.) Let O be a bimodal logic. Consider O s. We show

that it is closed under subframes iff O is inconsistent. This is undecidable.
Suppose that | is a subframe logic. Let ~ be a O-frame. Then let ~ be the
subframe over the points satisfying a V/3. This is not a frame for S i m unless
is empty. So, ~ is either empty or ~ ~ [-~. Hence O is inconsistent. Now
suppose that O is inconsistent. Then the frames of O ~ are the empty frame
and [-~. So, 0 ~ is a subframe logic. [i]
PROOF. (Version 2. (~ > 1.)) Consider the logics A~. They are complete
and elementary. Now, A~ is a subframe logic iff r contains only a single point.
Namely, consider the subframe r based on [e]. Either for all <lj, [c] <~j [~] and
then r has only one point, or [e] r [e] for some j, and then r 0j-I- while
c~l= ~j-F. M
Furthermore, tabularity and weak transitivity are undecidable. For notice

that the two are equivalent in logics of bounded alternativity. So suppose that
weak transitivity is decidable. First we decide whether or not A~ is weakly
transitive. If not, it I? unequal to the trivial process. But if E~ is weakly
transitive, the canonical Thue-frame c~ is finite, and A~ is decidable, and we
can check whether it contains the equations e ~ a, a E A. So we can decide
of ]? whether it is identical to the trivial process. Contradiction.
THEOREM 9.5.6. T a b u l a r i t y and w e a k t r a n s i t i v i t y are undecidable.
THEOREM 9.5.7. It is undecidable whether or not a monomodal logic con-

tains K ~ .
PROOF. Consider the logic (E~ ,1 O) s, the simulation of E~ L] O. It is

transitive iff I? is trivial. [2
Now let us deal with decidability. If we have more than one operator,
then local decidability is undecidable, because A~ is decidable iff I? is (The-
orem 9.4.5). Now, E~ is locally decidable. Let us assume we can decide
whether or not E~ is globally decidable. Then we can actually decide whether
I? is trivial. This goes as follows. Take I? and check first whether E~ is glob-
ally decidable. If not, I? is not trivial. If E~ is globally decidable, then decide
whether or not ~-C; r p ++ ~p;p ++ Cp. I? is trivial iff both formulae
are derivable. This is decidable. Similarly we can show that it is undecid-
able whether E~ has the global finite model property. Namely, assume it
is decidable whether or not E~ has the global finite model property. Take a
Thue-process I? and check whether E~ has the global finite model property. If
not, I? cannot be trivial. If it does, however, A~ is decidable, and we can then
decide whether t? is trivial. We leave it to the reader to supply the argument
that it is undecidable whether A~ has the local finite model property.
THEOREM 9.5.8. (~ _> 2.) It is undecidable whether or not a logic is lo-

cally decidable. Moreover, it is undecidable whether or not a logic is globally
decidable even when it is known that it is locally decidable.
THEOREM 9.5.9. (~ _> 2.) It is undecidable whether or not a logic has the
local finite model property. Moreover, it is undecidable whether or not a logic
has the global finite model property even when it is known that it has the local
488 9. L o g i c s of B o u n d e d Alternativity
The next result has first been obtained by LILIA CAGROVA in [44]. The
present proof appeared first in CARSTEN GREFE [91].
THEOREM 9.5.10. (n > 2.) It is undecidable whether a first-order condi-
tion is modally definable.
PROOF. Let T1 be the elementary theory of two binary relations. 211 is
universal and undecidable. The formula
o~o := (w:)[(vy t> =)(v # x) A (vv 9 ~)(y # ~)]
expresses irreflexivity and is not modally definable, because it is not positive.
Now consider the formula/3 : - a0 V 7, where ? is arbitrary. Then if for a
Kripke frame [ we have f ~ a0, then also f ~/3. Now suppose that/3 is modally
definable. Then it must hold in all Kripke frames for K . a l t l | K.altx by the
fact that every frame is the p-morphic image of an irreflexive frame (using
unravellings). Thus T1 k a0 V V, whence T1;--a0 k 7. Suppose now t h a t / 3
is not modally definable. Then T1 Y/3, that is, 2 ' 1 ; - ~ a 0 / 7 , for otherwise/3
holds in all frames and is therefore modally definable (for example by the true
constant). Hence if we are able to show that T2 := T1 U {-~a0} is undecidable,
we have succeeded in showing that modal definability (of/3) is undecidable.
Now consider the theory T3 := T1 U { a l } with
Sl :-~
(3x) [(vv u ~)(y " ~) A (vy 9 ~)(y " ~)
A (Vz){(vv u z)(v # ~) A (Vv * z)(v # x)}]
Since F- a l --+ ~a0 we have that if 2'2 is decidable, so is {r 9 2'2 ~- a l -+ r
= { ~ ' T 1 ; - ~ s 0 ; s l F- ~ } - { r F- ~ } - { ~ ' T 3 ~ } . So we are done if
we have shown that T3 is undecidable. Now, consider a frame f for T1. If we
add an inaccessible, reflexive point, that is, if we form the disjoint union f 9 r,
where r is the one-point, reflexive frame, then we have a T3-frame. And a
T3-frame is a Tl-frame. Denote by T v the set of sentences of T which have
the complexity V. By standard model theory, T3v - T v - 2"1. Hence, since T1
is undecidable, so is Tav and a fortiori T3. El
COROLLARY 9.5.11. (n :> 2.) It is undecidable for Sahlqvist logics of rank

0 whether they are of pure rank O.
THEOREM 9.5.12. (n > 2.) It is undecidable whether a logic is a fusion
of monomodal logics.
PROOF. Suppose it is decidable whether a n - m o d a l logic is fusion of
monomodal logics. Let F be a Thue-process over n. Then it is decidable
whether A~ is a fusion of monomodal logics. We show that it is decidable
whether F is trivial. For assume that A~ is in fact a fusion of monomodal
logics. Then by Theorem 9.1.3 it has the finite model property and is decid-
able. So, it is decidable whether or not A~ is the logic of a one-point frame.
Therefore it is decidable whether P is trivial. Now assume t h a t A~ is not the

fusion of monomodal logics. Then I? is not trivial. [i]
Finally, by invoking the Simulation Theorem, the following is proved with-
out any condition on the number of operators.
THEOREM 9.5.13. The following properties of finitely axiomatizable logics
are generally undecidable on the basis of a finite axiomatization
9 identity to and inclusion in a given tabular logic,
9 independent axiomatizability (for ~ > 1),
9 tabularity, weak transitivity,
9 local decidability,
9 global decidability, with local decidability given,
9 local finite model property,
9 global finite model property, with local finite model property given,
9 modal definability for elementary conditions,
9 being a subframe logic.
The first claim should be read as follows. There is no general algorithm
that decides, given a tabular logic A and a formula ~, whether or not K @ ~ C_
A, and whether or not K | - A. In particular, there is no algorithm deciding
these problems for A - T h [ ~ in case n - 1 and A - K~ @ _1_ for n > 1. The
application of the Simulation Theorem is in each cases straightforward. Notice
that the simulation of a subframe logic is not a subframe logic, so in this case
nothing follows for n - 1. Likewise, Corollary 9.5.11 has no analogue, even
though a somewhat less interesting variant could be formulated. The reader
may find it useful to describe the fact about monomodal logics that this
theorem gives rise to.
Notes on this section. Many results of this section have been obtained by
ALEXANDER CHAGROV, LILIA CHAGROVA,and MICHAEL ZAKHARYASCHEV,
see for example [44], [41]. The method in these papers is mainly simulating
the action of a Minsky machine in a logical frame. This construction is rather
complicated compared to the methods of this section. However, it achieves
stronger results (whenever possible) namely for extensions of K 4 and for
intermediate logics.
E x e r c i s e 344. Show that being n-transitive is undecidable for any n _> 1.
E x e r c i s e 345. Show that an extension of G is canonical iff it contains an

axiom of the form [:in• Show that the latter is equivalent to the logic being
contained in G.3.
E x e r c i s e 346. Show that being of codimension n is undecidable for any given

n, with the exception of ~ - 1 and n - 0.
9.6. D e c i d a b i l i t y of P r o p e r t i e s of Logics I I
The previous results have been more or less straightforward consequences
of the classical theorems on Thue-processes. For many properties, however,
there is a rather effective tool for establishing their undecidability. It is due
to S. THOMASON [212]. Suppose we are interested in the decidability of a
property ~ . Suppose further that the inconsistent polymodal logics have
and that we have found a finitely axiomatizable logic A which lacks ~ . If
transfers under fusion, then undecidability of ~ follows simply from the fact
that consistency is undecidable for bimodal logics. For consider logics of the
form A @ O, O finitely axiomatized. This logic has ~ iff O is inconsistent.
For if O is consistent, so is A @ O, and since A fails to have ~ , A | O fails
to have ~ , too. If, however, O is inconsistent, so is A | O and has ~ by
assumption. We refer to this argument as Thomason's Trick. Clearly, if
also transfers under simulation, then undecidability of ~ can be shown for
monomodal logics. An impressive list of properties can be treated in this way.
The following list is by no means exhaustive.
THEOREM 9.6.1. The following properties of logics are undecidable.

1. Y~p-elementarity,
2. c~-compactness,
3. c~-canonicity,
4. global completeness given local finite model property,
5. (local/global) interpolation,
6. (local/global) Hallddn-completeness.
PROOF. Using Thomason's Trick. The inconsistent logic is J~rp-elementa-
ry, a-compact, a-canonical, has the global finite model property, is Hallddn-
complete and has interpolation. Hence, single negative examples must be
found for all of these properties. G is a logic that is not ~rp-elementary
(1.), G | g is not 1-compact (see Section 6.5) (2.) and therefore not 1-
canonical (3.). (4.) follows from the conjunction of Theorem 9.6.3 and 9.6.4.
For interpolation and Halld~n-completeness we need to ensure that a A exists
not having these properties which is also complete with respect to atomic
frames. For (local/global) Halld@n-completeness take K. For interpolation
take $4.3. E]
This method can be adapted to bounded properties as well. Notice that

bounded properties generally fail to transfer under fusion. For example, if A
is pre-tabular, then A | O need not be pre-tabular even if | is tabular. How-
ever, if O - Th[-~, or ~ - Th [-~, pre-tabularity is preserved and reflected.
Moreover, for any other logic A | O is not pre-tabular, by Makinson's The-
orem. To show that pre-tabularity is undecidable it just remains to show
that the set {Th [-~, Th [ ~ } is an undecidable subset of ~ K. This is relatively
9.6. Decidability of Properties of Logics II 491
straightforward given the results of the previous section and is therefore given
as an exercise.
We still have to show t h a t there exist logics which have the local finite
model property but are globally incomplete. Here is such an example, the
logic | It is a 3 - m o d a l logic based on []i, i < 3. We write ff] for []]0, I for
fill and E] for []2.
Oo := K4.3 | K.altl | K.altl
9 p--+IOp
| QCqVCQq-+qVOq
| Op-+Op
| FT89
| FIMF~_L
9 OOP A Oq --+ O0(P A Oq)
e := eo 9 Op -+ O(p A-.Op)
Oo is Sahlqvist and therefore ~rp LJ~ - e l e m e n t a r y . O contains in addition the
axiom G for D. A Kripke-frame (f, <l, 9 ~<) satisfies the axioms of O iff the
following conditions are met.
(a) <~ is a linear irreflexive order such t h a t there are no infinite ascending
chains.
(b) If x 9 yl,y2 then yl = y2.
(c) <~ contains the converse of 9
(d) I f x < ~ y 9 1 4 9
(e) If x ~ ym and x ~< y2 then Yl = Y2.
(f) I f x ~ y t h e n x < ~ y .
(g) If x ~ z and z = y or z <] y then y has no ~-successor.
(h) I f x < ] y , x ~ < x + and y ~< y+ , then y+ <~x +.
The following is easily checked by inspection of all axioms.
PROPOSITION 9.6.2. Oo and 0 are subframe logics.
Now suppose we have a formula T and a global model (~', fl}. P u t X :=

sf(~). The T - s p a n of a point x E f is the set of all subformulae 0 r of
such t h a t (~,fl, x) ~ 0 r If (~:,/3, x} ~ 7 for some ~- E X then there exists
a point y such that x = y or x <~ y and y has least T - s p a n among all points
satisfying 7. This point is necessarily irreflexive. (For if it is reflexive, then
it also satisfies 0~-, and so it satisfies 0(T A ~0T). Thus, it has a successor z
satisfying ~- and 7 is not in the T - s p a n of z. Since ~- is in the T - s p a n of y, y
has not been chosen minimal. Contradiction.)
We now assume t h a t T is in normal form and t h a t (~,/3, w0) ~ ~. We
can actually assume t h a t T is not in the T - s p a n of w0. Moreover, we assume
to be descriptive; hence its underlying Kripke-frame is a Oo-frame. We
now define a selection procedure on points as follows. In each step a point is

selected by a so-called r e q u e s t . A request is a pair (y, r where (~,/3, y} ~ r
and r C X. In the selection step we answer the request by selecting for
(y, r a point z. D e p e n d i n g on r this m a y give rise to new requests. After
the selection, (y, r is removed from the list of requests. T h e s t a r t i n g set of
requests is {(wo, ~)}. The answers on the request ( y , r are as follows. If
r = ~-1 A ~'2, then the request is answered by adding the requests (y, ~-1) and
(y, 72). (y is then said to be selected for ~-1 and T2 from y.) If r = ~-1V T2, t h e n
if ~-1 holds at y the answer is (y, 71 } and y is selected for 71 from y. Otherwise
~-2 holds at y, and we answer with (y, T2}. If r = 071, then a z is chosen such
t h a t y <~ z and (~,/3, z) ~ ~-1, and the request is answered with (z, 71). z is
said to be selected for ~-1 from y. Analogously with r = r and r = (~'1.
If r = U]~-I or = m71 or = []71, then the request is answered by dropping it,
i. e. no new request is being made. Similarly if r = p, a variable. We take g
the set of all points t h a t have been selected. It is easy to see t h a t g is finite.
g is not necessarily an internal set. However, 0o is a subframe logic and ~ is
a descriptive 0 frame. Hence its underlying K r i p k e - f r a m e , ~, is a O o - f r a m e ,
and so the subframe g defined by g is actually a Oo-frame. Moreover, if we
let 7(P) := 13(p)A g it is established by easy induction t h a t each point satisfies
the formulae it has been selected for. In particular, (9, 7, x) ~ ~.
We now massage t~ into a O - f r a m e . To do this, we first analyse w h a t could
have gone wrong in case g ~ 0 . In t h a t case g m a y contain <~-self-accessible
points ( t h a t is a point x such t h a t x <~x). Let us call t h e m i m p r o p e r p o i n t s .
I m p r o p e r points can be avoided as an answer to a request of the form (y, 0~-)
as we have seen above. Moreover, if x :~ x and x 9 y then y r y, so i m p r o p e r
points are not selected from proper points by answering a request of the form
r (To see this, notice t h a t by (d), if x is proper, and x 9 y t h e n y is the
i m m e d i a t e <]-predecessor of x. It is not difficult to verify t h a t in t h a t case y
is also proper.) Therefore, i m p r o p e r points can only arise t h r o u g h a request
of the form (~T. Thus, let y be selected to answer a request (x, (~-). T h e n
we claim t h a t in the frame generated by y in ~, no point has a ~<-successor.
This holds for all <]-successors of y, by (f). Now let y 9 z. T h e n z = y or
z <] y, by (d). Hence, z is already in the transit of y. Moreover, z <~ z, and
the same a r g u m e n t can be repeated with z. (To see this, notice t h a t z <] y.
since y 9 z, by (c). Second, y <~ y 9 z, and so y = z or y <l z. In the first case
z <~ z since y <l y. In the second case, z <~ z follows from z <] y and y <~z, by
t r a n s i t i v i t y of <].)
Now take an improper point x C g. Let C(x) := {y C g : x < ] y < ~ x } .
We can assume t h a t (i) C ( x ) = { x i : i < n} and xi 9 xj iff j _-- i ( m o d n)
and t h a t (ii) y ~< x for some y C g. We let ~tn := ( Z , < ] a , 9 where
< ~ := >, 9 := {(i,i + 1} : i ~ E} and ~<a := ~. This is a O o - f r a m e . We
assume its set of points to be disjoint from t h a t of g. Now define the cluster
substitution for C(x) with respect to y by
g+ := (g - c ( ~ ) ) u z
< n (g- c ( ~ ) ) ~ u <,~
<]+ := u { ( ~ , v ) 9~ e g , v ~ z , ~ < ~ }
u { ( v , ~ ) 9~ ~ g , v c z , x < ~ }
9 := .~ n ( g - c(x)) 2 u .~
~<+ := <~ n ( g - C ( z ) ) 2 U{<y,O)}
This defines the flame g+. As is easily checked, it is a Oo-frame. Furthermore,

the map 7r 9 x ~-~ x for x r Z, 7r 9 k . n + i ~ x~ is a p - m o r p h i s m 7r" g+ ~ g.
(This is clear for the relation <], and easy to check for the relation ~. In
the case of 4, one only has to observe that the cluster C(x) is a connected,
isolated component of the graph (g, 4). This, however, follows from (c).) P u t
~+(p) " - 7r-l[7(p)]. Then (g+,V+,wo) ~ ~. We perform this substitution
successively for all clusters of g containing an improper point. This yields a
frame [~ and a valuation ~ such that (~, 5, To) ~ p. It is a Kripke-frame for
0o. On the basis of our original selection procedure we define a new selection
procedure on (~, 5} that will yield a finite model on a frame for O.
Assume that the selection procedure began with (To, p). wo is proper.
Let y be the first improper point selected. Then y is in a cluster which gets
replaced by ~tn. We select 0 in the new model, instead of y. Now we consider
the effect this can have on our future selection. For each point we define the
s e l e c t i o n h i s t o r y to be a sequence of subformulae of ~ by induction on the
selection process as follows, x is assigned (~p>. If y is assigned (tbi " i < n>
and z is selected by (y, r with answer (z, ~'1) ({(z, 71), (z, T2)}), then z
is assigned (tbo,... , r ~-1) ( ( r ,r TI> and ( r ,r ~-2)). z
may of course be identical to y, so that points may have several histories;
but each history cr is assigned to only one point, denoted by a(cr). Now we
define a function b from selection histories into h as follows. If a(cr) is proper,
then b(cr) := a(a). Otherwise, consider the shortest history cr leading to an
improper point z. Then b(a) " - 0, where 0 is the distinguished in the frame
replacing the cluster of a(cr). We claim that any history such t h a t a(cr) is
improper extends a shortest history of this kind only by A, V and 0 moves.
To show this claim, assume that Uo is the first improper point in a history
or. We have to show that all improper points of cr are obtained by a series of
A, V and 0-moves from uo. Let us observe the following. The last move in
a history that leads to an improper point is not a (}-move. For they select
(by our choice) proper points. Moreover, if x is proper and x 9 y then y is
also proper. Hence, an improper point arises either by making a ~<-move (A),
or from an improper point by making a 9 (B). Consider (B) first. We
have the situation that x is improper and that x 9 y. Then by (d), since x <Ix
we deduce t h a t x <] y or x = y (and t h e n also x <] y). Moreover, by (c), y <~ x.

Hence, x and y are contained in the same <~-cluster. So y is i m p r o p e r , too.
This finishes the case (B). Now let (A) be the case. We show t h a t once (A)
has o c c u r r e d in a selection sequence it will not h a p p e n again. To t h a t end let
us take u0 and a sequence <ui : i < n + 1> such t h a t ui+i is a successor via
one of the relations. Assume t h a t no <~-successor is i m p r o p e r . First of all,
note t h a t u0 <l ui for all i < n + 1. For u0 is i m p r o p e r , and so u0 9 ui implies
u0 <1 u i , as we have seen. F u r t h e r m o r e , u0 ~< ui implies u0 <~ u i , by (f). Now
consider the a point uk+i. A s s u m e t h a t u0 <1 uk. T h e n uk <~ Uk+l c a n n o t
occur, by (g). If uk <1 uk+i t h e n u0 <~ uk+i by t r a n s i t i v i t y of <l. Finally, let
uk 9 uk+i. T h e n uo <~ uk 9 Uk+l yield u0 = uk+l or Uo <~ Uk+l. Since Uo is
i m p r o p e r , u0 <~ uk+i in b o t h cases. This shows, by (g), t h a t no uk has a ~ -
successor. It follows t h a t if uk+i is i m p r o p e r , uk 9 uk+i a n d uk is i m p r o p e r
too. A n d this proves our claim t h a t an i m p r o p e r point is in the a - t r a n s i t of
u0, the first i m p r o p e r point in a history.
Thus, let T be i m p r o p e r and a the shortest s u b h i s t o r y such t h a t a(~-)
is i m p r o p e r . We have defined b(~-), b(cr) is now defined by i n d u c t i o n on its
length. (In each of the steps the choice of value u n d e r b is unique.) Finally, let
k be the set of points selected by the new procedure, k is finite, a n d defines
a s u b f r a m e of ~, 2. Let ~(p) := k A 5(p). T h e n as before we conclude t h a t
<[, e, w0> ~ ~. Moreover, t is a frame for O. We have shown the following.
THEOREM 9.6.3. | has the local finite model property.
Now we show
THEOREM 9.6.4. O is not globally complete.
PROOF. We claim t h a t (i) (~T --+ $ ( ~ T ~ o ~(~T and (ii) no m o d e l in

which (~ T -+ $(~ T holds globally b u t (~ T is locally satisfied is b a s e d on a
K r i p k e - f r a m e for O. T h a t (ii) is the case is r a t h e r easy to see. S u p p o s e we
have a O - f r a m e ~ such t h a t ~ ~ (~ T -+ $ ~ T and t h a t there exists a point
x such t h a t <~, x> ~ (~T. T h e n x ~ $ ( ~ T and so there exists a xi such t h a t
x 9 xi a n d xi ~ ~ T . Inductively one can show t h a t there exists a chain
<xi 9 i E w> such t h a t x0 - x and xi 9 xi+i for all i. Moreover, t h e r e exist
points yi such t h a t xi ~ yi, by (h). By the p o s t u l a t e s of O we have Yi <1 yi+i
for all i. This m e a n s t h a t ~ contains an ascending chain w i t h respect to <~.
Hence, the u n d e r l y i n g K r i p k e - f r a m e violates the axiom G for <~, t h u s is not a
O - f r a m e . Now we still have to show (i). Take a copy of the n a t u r a l n u m b e r s
N, m e m b e r s of which we denote by underlining, e. g. 5; and a (disjoint) copy
of Z, m e m b e r s of which are d e n o t e d by overstriking, e. g. - 5 . We p u t k <~
for all k E Z and e E N; k <l g iff k < t~; and _k <l ~ iff k > t~. So, with respect to
<1 we have placed Z 'before' N ~ . We p u t k_ 9 ~ iff t~ - k + 1; k 9 t~ iff g - k - 1
a n d no other points are related with respect to 9 Finally, for k > 0 we p u t
for - k ~< k, and no other relations shall hold. We take as the algebra @ of sets
the 0 - g e n e r a t e d algebra of sets. This defines the frame Ft. Now observe t h a t
Ft~ satisfies all postulates except for G. Hence everything depends on finding
a set of internal sets such t h a t G is also valid. To show t h a t ~ fulfills G, let us
prove t h a t O is nothing but finite unions of intervals with respect to <1, where
an interval is a set of the form [x, y] " - {z" x <] z <] y} U {x, y}, x, y E f, or of
the form [ ~ , x] " - {y" y <l x} U {x}, where oc is just an artificial symbol. We
claim t h a t these sets are 0-definable. (A set c is 0-definable in ~ if there exists
a variable free formula C such t h a t (~, x} ~ C iff x C c. It is easy to see t h a t
c is 0-definable iff it is contained in the 0 - g e n e r a t e d subalgebra of ~+.) To
t h a t end, observe t h a t Ice, k] is 0-definable, by the fact t h a t [c~, 0] is defined
by T and 0[c~,_k] - [ c ~ , k + 1]. Furthermore, we have [cx~,---1] - (~[oo,0],
[cc, k + 1] - . I c e , k], and [ c ~ , k - 1] - 0lot, k]. Thus all sets of the form
[cxD,x] are 0-definable, and from t h a t follows t h a t all finite unions of intervals
are 0-definable. Secondly, we show t h a t this algebra is already closed under
all operations. The set operations are clear. Now 0Ix, y] - [cxD,y+], where
y 9 y+, and O[x,y] - [ x - , y - I , where x - 9 x, y - 9 y, if these points
exist. (The other cases are also straightforward.) Next 0. If x - y - 0 then
. I x , y] - ~. Otherwise, . [ o c , y] - [oc, y - ] . Now finally ~. Consider Ix, y]. If
x - k_ then (~[x, y] - Z. Hence let x - k. We m a y assume k > 0, otherwise
(~ Ix, y] - (~ [1, y]. If y - ~_, then ~ Ix, y] = (~ [k, _e] - [+oo, - k ] . If y - ~ then we
can assume y > x and so (~ [x, y] - (~ [k, ~] - [ - g , - k ] . Indeed, the set of sets is
closed under all operations. We have to show now t h a t ~ ~ 0p --+ 0(p A ~ 0 p ) .
Take a valuation assigning an internal set c to p. T h e n this set has a largest
element with respect to <l. For it is a union of intervals [xi, yi]. A m o n g the
Yi there is a largest with respect to <], say y0. T h e n if (~,/3, x} ~ 0p we have
x <l Y0 and so (f~,/~, x) ~ O(p A ~OP), since (f~,~,Y0} ~ P A-~Op by choice of
Y0. To end the proof, we notice t h a t f~ ~ (~T --4 0(~T but f~ ~ ~(~T. For the
first note t h a t (Ft, x} ~ (~T i f f x - k for some k < 0. T h e n k 9 k - l , and
(Ft, k - 1) ~ (}T. For the second notice t h a t (gt,---T) }L -~(}T. Ul
COROLLARY 9.6.5. O is locally complete. Its extension by a universal
modality, 0 I , is not locally complete.
This follows i m m e d i a t e l y with T h e o r e m 3.1.13.
E x e r c i s e 347. This is another set of exercises t h a t deal with Blok's Alter-

native. This time we deal with incompleteness phenomena. Show t h a t any
of the logics of Section 9.2 based on sequences of b o u n d e d index could have
been taken. Show based on these considerations t h a t the lattice of 3 - m o d a l
logics has 2 ~~ c o - a t o m s , which are incomplete.
E x e r c i s e 348. Show t h a t in the lattice of 3 - m o d a l logics any logic of finite

co dimension has degree of incompleteness 2 ~~
E x e r c i s e 349. Show that in the lattices of x-modal logics, ~ ~ 0, any con-

sistent logic of finite codimension has degree of incompleteness 2 ~~ W h a t if
the logic is inconsistent?
E x e r c i s e 350. Show based on the preceding considerations that no consis-

tent modal logic of finite codimension can be obtained by iterated splittings
of K~. Hint. Otherwise, show that it can be co-covered by at most countably
many logics.
E x e r c i s e 351. Let C be any of the following complexity classes: NP, PSPACE,

EXPTIME. Show that given a monomodal formula T it is undecidable whether
or not K1 9 ~ is C-computable. Likewise, show that it is undecidable whether
or not K1 9 ~p is C-hard (C-complete).
E x e r c i s e 352. Show that {Th [-~, Thl~-I} is an undecidable subset of ~ K.
E x e r c i s e 353. (Continuing the previous exercise.) Show that pre-tabularity,

pre-finite model property and pre-completeness are undecidable.
E x e r c i s e 354. Show that the set of pre-tabular logics of ~ $4 is decidable.

C H A P T E R 10
Dynamic Logic
10.1. P D L A Calculus of C o m p o u n d Modalities

Propositional Dynamic Logic, P D L for short, can be seen as a special
kind of polymodal logic and this i s - - at least i m p l i c i t l y - the way we will
handle it here. This has the advantage t h a t we can use the results of the
previous chapters to a large extent. Nevertheless, P D L has a different syntax.
P D L concentrates on compound modalities; the idea is t h a t it is worthwile
to investigate the structure of compound modalities separately, because many
compound modalities arise naturally from the relational interpretation. For
example, if we have an operator W1based on the relation <~ and an operator I
based on the relation 4, then the compound modality D p A I p is based on the
union <~ U 4 of the two, and D i p is based on 9 o <l. The perspective from
which P D L views these things is from the perspective of relations. To be
able to define an operator based on the union or composition of two relations
is important. For the main interest for doing P D L is in reasoning about
computer programs, or more generally, about actions. A computer program
can be seen simply as a relation between memory states of a computer. For
example, if our computer has three memory cells, x, y and z, storing integer
numbers, the program z := x § y is a relation between triples of integer
numbers. In order not to confuse a program with the relation we call the latter
the extension of a program on a given computer. Thus, <<3,-4, 7>, < 3 , - 4 , - 1 > >
is in the extension of the program z := x + y, but ((1, 1, 0), (0, 1, 2)> is not.
The particular power of programming languages comes from the fact t h a t
programs can be combined. We can namely
Compose Programs: From two programs a and/3 we form the compo-

sition a; ~ which is the program defined by executing first a and then
T e s t : We can ask whether certain facts hold such as 'x = y?', by means
of which we can ask whether the number assigned to x is identical to
the number assigned to y.
Combine T e s t s : We can use standard boolean connectives such as true,
not and and to build more complex tests.
497
498 10. D y n a m i c Logic
Combine Programs L o g i c a l l y : T h a t is, we can use logical gates such

as
if ~o then a else/3 fi 1
where ~a is a s t a t e m e n t and c~ a n d / 3 programs. This m e a n s t h a t we
execute first a test for ~. If ~a holds, a is executed; and if not, /3 is
executed.
I t e r a t e P r o g r a m s : For a s t a t e m e n t T and a p r o g r a m a, the p r o g r a m
while ~ do c~ od
will execute a and continue to do so as long as the condition ~ is

m o m e n t a r i l y satisfied, and the p r o g r a m
until ~o do c~ od
will execute a as long as ~a is m o m e n t a r i l y not satisfied.

This is the basic inventory of s t a n d a r d p r o g r a m m i n g languages like Algol,
Pascal and their derivatives, with those p a r t s s t r i p p e d away which belong to
specific i n t e r p r e t a t i o n s of the symbols (such as real or integer n u m b e r s , or sets,
or lists) and the usual i n p u t / o u t p u t routines. T h e y can be r e i n t r o d u c e d as a
set of basic (or e l e m e n t a r y ) p r o g r a m s , from which the c o m p l i c a t e d routines
are built up. To be realistic, we would need to talk at this point a b o u t
m e m o r y cells and assignments. So in fact, a real model of a c o m p u t e r would
have to use some first-order logic. But the purely propositional p a r t is not
only interesting (and decidable), but already very powerful.
Now a p r o g r a m is in some sense a relation of states in a c o m p u t e r , and the
list of constructions we have just given can be p r o d u c e d from a small list of
basic constructions. P D L s t a r t s with two different sorts of symbols, propo-
sitions and programs. T h e r e are at the basic level propositional variables,
propositional constants and p r o g r a m constants. The set of basic p r o g r a m
c o n s t a n t s is d e n o t e d by II0 and consists of a0, a l , . . . . The n u m b e r of these
c o n s t a n t s varies. There are no p r o g r a m variables. P r o p o s i t i o n s can be com-
posed as before with the boolean connectives. P r o g r a m s are c o m b i n e d using
the connectives
a U/3 set theoretic union
a;/3 relational composition
a* reflexive transitive closure
Any proposition p can be converted into a p r o g r a m , by using the q u e s t i o n
m a r k '?', called test. If we test for p, we write '~a?' and read t h a t ~a test.
The p r o g r a m p? o p e r a t e s as follows. If ~ is the case - - we say t h e n t h a t the
1To write an if-then-else-clause with the help of fi is standard practice, though not
always explained. The word fi has actually no meaning. It is used simply to mark the end
of the clause. This saves brackets while writing and makes a program optically perspicuous.
The same applies to the pair do ... od below.
10.2. Axiomatizing P D L 499
program ~? succeeds the next clause is carried out. If ~ is not the case
we say the program ~? fails then no subsequent programs will be carried
out. For example, the program (~?; c~)U ( - ~ ? ; Z) performs a if ~ is the case
and fi if ~ is not the case. Thus it is identical to if ~ then c~ else 13 fi. Notice
that because we use the union, we save the entire program from failing, even
though some of its parts may fail individually. Thus, P D L makes crucial use
of the fact that programs are allowed to be nondeterministic. This sounds
absurd if we think of numerical calculations, where a correct program should
yield a definite result, but is really rather useful in connection with inbuilt
choices, as we have in fact just seen. Finally, any program c~ defines a modMity
[a] and its dual (a). Basically, the extension of the program is the relation on
which the modality is based. The statement [a]~ can be read as at the end
of all possible computations of a, ~ holds. Dually, the statement (a)~p means
there is a computation for a at the end of which ~ holds.
One can think of various fragments, extensions and refinements of P D L .
The fragment of P D L that does not use the star, is called E P D L , which
is short for e l e m e n t a r y P D L . It turns out to be a notational variant of
polymodal logic. Second, there is test free P D L , of which we will make
certain use. A particularly interesting extension is the addition of the converse
operator. Given a program a, a ~ will denote the converse program, i. e.
the backward execution of a. Although for the standard interpretation this
makes little sense, because standard computer languages do not need such an
operation, it makes sense in reasoning about the behaviour of programs and
actions. Also, dynamic logic is increasingly used in the semantics of natural
language (see for example JEROEN GROENENDIJK and MARTIN STOKHOF
[92], JAN VAN EIJCK and PER-JAN DE VRIES [218]), and in reasoning about
actions (see VAUGHAN PRATT [165] and BRIGITTE PENTHER [160]), to name
just a few. Many connectives in natural languages make reference to the
converse, such as until, since etc., though mostly in connection with tense
only.
10.2. A x i o m a t i z i n g P D L
In this section we are going to axiomatize P D L and prove the correctness

of this axiomatization. As it turns out, the newly added program construc-
tors are pretty harmless, with the exception of the Kleene Star. The latter,
however, is a relatively difficult operator. We will see that there is no axiom
system that can guarantee that c~* is in all cases the reflexive transitive closure
of c~. However, it is possible to give an axiomatization such t h a t at least in
Kripke-frames c~* has this property. Thus let us start with the latter. Let H0
be given; a member of H0 is denoted by ~ or rl. A d y n a m i c K r i p k e - f r a m e
over H0 is a pair [ = (f, or), where ~ : I I 0 -+ 2/• / assigns to each basic pro-
gram a binary relation on f. A v a l u a t i o n is a function 7 : vat U cons --~ 21
and the satisfaction clauses are as follows.
(f, 7, z) ~ [:?]r r (f, 7, ~) ~ : -~ r

(f, 7, x) ~ [~]: r for all y such that (x, y) E cr(~) (f, 7, y) ~ ~o
(f, 7 , ~ ) ~ [~ u ~]~, . (f, 7, ~ ) ~ [~]~; [~]:

(~, ~, ~) ~ [~*]~ r (f, 7, ~) ~ :; [~]:; [ ~ ] : ; [ ~ ] ~ ; . . .
Alternatively, given/3 and or, we can extend cr to a map ~ from all programs
to binary relations over f a n d / 3 to a m a p / 3 from all propositions to subsets
of S.
9(-~) :: f -/~(:)
/3(: A r :-- /3(:) n/~(r
Z([~]:) := {~. (vy)((~,y) e ~(~) ~ y e Z(:))}
~(c) := ~(C)
~(~ u z) := ~(~) u ~(z)

~(~;Z) := ~(~) o ~(Z)
~(~.) := ~(~)*
In order to make the notation perspicuous we often write x -~ y if (x, Y/ C

C~
~(a); in other words, we write x --+ y to state that from x there exists an
a - t r a n s i t i o n to y. This is a popular notation in computer science. The reader
may verify that the two definitions of acceptance of formulae are one and
the same, that is, we have (f,/3, x) ~ : iff x C /3(:) for all propositions.
Now we proceed to the promised axiomatization. As usual, the only rule of
inference is modus ponens (in the local case), and the rules of substitution and
necessitation are admissible. The axioms are in addition to those of boolean
logic the following.
(bd.)
(df?.) k- [p?]q. ++ .p ~ q
(dfU .) ~- [~ u 9];. ~ .b]p A [9];
(df; .) [~; ~]p. ++ .[~][~];
(cir.)
(ind.) ~- [~*](p ~ [~]p). ~ .p ~ [~*]p
This logic is standardly known as P D L . These axioms are due to KRISTER
SEGERBERG, see for example [195]. Notice that a and/3 are not variables of
the logic, but meta-variabes for programs. Hence the above postulates are not
axioms but schemes of axioms. P D L is in general not finitely axiomatizable;
the axiom system is recursive (i. e. decidable). Our first theorem concerns
the correctness of this axiomatization. This means in informal terms that if
we view the programs of P D L as separate and consider the above axioms as
restrictions on the definition of these programs, then it will turn out that any
Kripke-frame for the above axioms reduces to a dynamic Kripke-frame. To
state this precisely, let us introduce the notion of a P D L - K r i p k e - f r a m e .
This is a pair (f, 7} which satisfies the axioms of P D L not involving test. A
PDL-Kripke-model is a triple (f, %/3) such that (f, T) is a P D L - K r i p k e -
frame and the axioms (df?.) are satisfied. (In this case we say that T and/3
are c o m p a t i b l e . )
THEOREM 10.2.1. Let (f,T) be a PDL-Kripke-frame. Then

1. ~-(~ u ~) = ~-(~) u ~-(~).
2. -,-(~; 9) = ~(o~) o "-(9).
3. T(a*) = T(a)*.
PROOF. Suppose, (f, ~-) ~ [a U 13]p ++ [c~]p A [/3]p. We know that this
axiom is Sahlqvist, and it is easily checked that it corresponds to the following
property.
(W)(Vy)(x %~ y ~ x ~ y V x s y)
Thus, (dtLJ.) holds iff ~-(a U/3) = w(a) U w(13). Similarly it is shown that (df;.)
holds iff T(a;/3) = ~-(a)o ~-(13). The axiom (cls.)
[~*]; ~ p A b][~*]p
has as its dual
p v (~; ~*);.-+ .(~*)p
which is a conjunction of p --+ (a*)p, corresponding to the reflexivity of a*,
and (a; a*)p. --+ .(a*)p. The latter is also Sahlqvist and corresponds to the
condition
(w)(vv 3- x)(Vz ~ v)(x -~ z)
In other words, the relation T(a*) is reflexive and successor closed with respect
to r(c~). It therefore contains the reflexive transitive closure of ~-(c~). T h a t it
is exactly the closure is the effect of the induction axiom (ind.). Namely, let
x be given. Assume
CI~ n
P u t / 3 ( p ) " - {y " (3n E w)(x -+ y)}. Then

Z, x) v; -+ [4p)
C~ n
and so ([,/3, x) ~ [a*]p. Hence x ~ y implies x -+ y for some n 6 w. [i]
The last theorem established that (f, r) is effectively a dynamic frame if

we concentrate on test-free propositions. The test presents a problem, not
so much for the axiomatization as for a proper formulation of the result. For
suppose t h a t in the last theorem we had stated that ( f , r } satisfies (df?.).
Then since ~- is given,/3 is actually fixed. For we have
(y, [p?]• •
which is the same as
(f, "r)~ (p?)T. ++ .p
Thus /3(p?) = { x : (3y)((x,y) E r(p?))}. However, intuitively we want the
assignment of relations to tests be regulated by/~ and not by T. This problem
can be solved as follows. We a b a n d o n / 5 and define it as above. Therefore,
the whole information is already in the map r. Then the pair (f, r) can serve
both as a frame and as a model. The variability of/3 as opposed to t h a t of
cr will be accounted for by the notion of variants. Given (f, T) we call (f, ~)
a v a r i a n t if ~ [ H0 = r [ II0. We now say that the f r a m e ([, r) satisfies
iff for all variants ~ of ([, r) we have ([, ~) ~ qo. Now what happens if the
frame (f, r) satisfies (dr?.)? Take any point x, and assume t h a t there is a y
such that (x, y) E ~(p?) for some variant ~. Then /3(p) contains x, so p is
true at x. Assume furthermore ~ ( q ) = {x} (that is to say, ~(q?) = {(x,x)}).
Then (f, ~, x) ~ p -+ q and so (f, ~, x) ~ [p?]q, so that (f, ~, y) > q, which is
to say, y = x. Hence if the frame (f, r) satisfies (dr?.) then r(p?) C_ A f , the
diagonal on f. Moreover, given the definition of ~ we get that (x, x) C T(p?)
iff X C /~(p), since x has a p?-successor iff x C /3(p). This finally justifies to
restrict the attention to dynamic frames. To recapitulate, the problem lies in
the fact that the basic programs are actually constants and so are all test-free
programs. The test p?, however, behaves like a variable, though not induced
by the assignment cr :ri0 ~ 2 / • / but by the basic assignment ~.
Let us turn now to generalized frames. In principle, they are defined in full
analogy to the modal case, if we just take dynamic logic to be a particular
case of modal logic, where we have an infinite stock of operators with the
postulates as above. This is not so straightforward for reasons mentioned
earlier. Nevertheless, we can form the canonical structure EanpDL(var) in

the following way. The worlds are as usual maximally consistent sets, and the
accessibility relation is as in Section 2.8, namely U -% V iff for all [a]T C U
we have T C V.
PROPOSITION 10.2.2. In the canonical structure, ~anpDL(Var), T(a U
~) = ~-(a)O T(~), T ( a ; ~ ) = ~-(a)o T(~), and T(~?) = { ( U , U ) : ~ e U}.
Moreover, T(a*) D T(a)*.
PROOF. The first two and the last are straightforward from the fact that
the postulates (dfO.) and (df;.) are elementary and correspond to the fact
that the relation for a O fl (a; fl) is the union (composition) of the relations
for a and ft. Similarly, the last claim follows from (cls.). For ~? observe that
U ~-~ V ifffor allff if[p?]~ E U w e have ~ C Y ifffor all ~ i f ~ - - + ff C U
we have ~ C V. Now two cases arise. First, i f ~ C U, then ~ - + ff C U i f f
C U, so that we get U ~ V iff V _D U iff U - V, since b o t h are maximally
consistent. The second case is ~ r U. Then [ ~ ? ] l C U, and so no world is
~?-accessible. [:3
The problem with the canonical structure is the bad behaviour of the star.
We only know that the relation for a* is at least the transitive closure of the
relation for a, and is successor closed, but we really cannot say more. This
is no accident, for P D L fails to be frame-compact. Hence, a completeness
proof for P D L is not straightforward. In fact, many authors have overlooked
this rather subtle fact, and it took some time until a correct proof appeared
that did not make the assumption that the canonical model has an underlying
Kripke-frame which also satisfies P D L . We add here that P D L is also not
compact. It is in fact not even 1-compact, which means t h a t there is a set of
constant formulae that have no model based on a P D L - K r i p k e - f r a m e . This
is left to the reader as an exercise.
Standardly, P D L has infinitely many variables and infinitely many pro-
grams. However, as in the modal case we can restrict ourselves as well as to
finite variable fragments as well as to finitely many basic operators. Let us
fix the set of variables Pi, i E w, and ~i, i C w. Let us denote by P D L ( m , n)
the restriction of P D L to the subset of variables pi, i < m, and programs
~i, i < n. Similarly, the notation P D L ( w , n) and P D L ( m , w ) is used. Then
P D L = P D L ( w , w) and
PDL- 0 PDL(m,n)
m~?I~o2
It turns out that P D L ( w , n) is weakly transitive for n < w, while P D L is

not. The master modality of P D L ( w , n) is 7* with 7 := [-J~<n ~"
Let us now turn to P D L ~, the extension of P D L by the converse. There
is an axiomatization of the converse as follows.
504 10. D y n a m i c L o g i c
(df+.) ; A -+ A
(df_~.)
This is not exactly the way in which the converse is standardly axiomatized.
In tense logic, the postulates p -+ [a](a~)p and p --+ [a~](a)p are normally
used. This is a matter of convenience. It is also possible to replace the second
postulate by ( a ~ ) p ++ (a)p. All these boil down to the same. For notice
that the postulates (df~.) are r-persistent, so that in the canonical frame we
do have ~-(a ~) - 7(a) ~. The other postulates have the same effect on the
canonical structure.
Finally, there are important program constants that we can define, namely
skip and fail, defined by T? and 2_?, respectively, skip is interpreted by the
diagonal, and fail by the empty relation. There are many more operators that
get studied with respect to computer application, such as since and until, both
binary connectives for formulas, nominals (see SOLOMON PASSY and TINKO
TINCHEV [159]) and many more. An overview of some of these different logics
is given in ROBERT GOLDBLATT [81]. We will study some of these in detail. A
very important logic from a practical point of view is the logic D P D L , which
in addition to P D L contains the axiom (~)p -+ [~]p for all basic programs
~. This codifies the fact that an execution of a program yields at most one
outcome. We say that the program is deterministic, and this is why D P D L
contains the additional letter 'D'. Notice that it is only the basic programs
which are required to be deterministic; for even if c~ is deterministic, a* as
well as a U (a; a) need not be.
E x e r c i s e 355. Formulate P D L instead of * with the operator + of transitive

closure.
E x e r c i s e 356. Show that the boolean connectives can all be dropped from
P D L without loss of expressivity, retaining only the propositional variables
and constants, including T and 2_.
E x e r c i s e 357. Show that the postulates
p --+
--+
[ 4 p ) - + (p--+
are all independent. That means, show that no logic axiomatized by two
together contains the third.
E x e r c i s e 358. Show that P D L is not 1-compact. Hint. The only source for
failure of 1-compactness can be the induction axiom.
10.3. The Finite Model Property 505
E x e r c i s e 359. Recall that E P D L is the fragment of P D L without the star.

Show t h a t in E P D L every formula is deductively equivalent to a formula
without program connectives, that is, without U, ;, *, and ?.
E x e r c i s e 360. Show that P D L - Um,ne~ P D L ( m , n ) . Hint. Use the

compactness of the derivability in P D L .
10.3. T h e F i n i t e M o d e l P r o p e r t y
One of the earliest results on P D L was the proof by FISCHER and LADNER
[69] that P D L has the finite model property. The proof is interesting in two
respects; first of course for the result itself, and second for the notion of
downward closure t h a t it uses. Notice that one of the first problems that one
has to solve is t h a t of defining a good notion of subformula. Of course, there is
a syntactic notion of a subformula, but it is of no help in proofs by induction
on subformulae. For what we need is a notion that allows us to assess the t r u t h
of formula in a model by induction on its subformulas. Consider, for example,
the formula (c~;3)P. The only proper subformula, syntactically speaking, is p.
So, all we can say is that (a;/3)p is true at a point iff there is a c~; ~-successor
at which p holds - - period. But c~;/3 is a complex program and we would
like to do an induction using only the basic programs to start with. Hence,
let us take (c~)(3)P also as a subformula. Then we can break down the t r u t h
definition of (c~;f~)p into two parts. (c~;f~)p is true at a node x if there is
a c~-successor y at which (/3)p holds. The latter is the case if there is a 3 -
successor z of y such that p holds at z. It is these considerations t h a t lead to
the following definition, taken from [69].
DEFINITION 10.3.1. Let X be a set of PDL-formulas. The F i s c h e r -

Ladner closure of X, denoted by FL(X), is the smallest set containing X
such that
(rid.) FL(X) is closed under subformulae.
(fl?.) If [~?]~ C EL(X) then also ~, r C FL(X).
(fl;.) If [c~;3]~ E FL(X) then also [al[3lp C FL(X).
(flU.) If [c~U 3]~ e FL(X) then also [c~]~, [f~]~ e EL(X).
(fl*.) If [a*]~ e FL(X) then also [c~][a*]~ e EL(X).
LEMMA 10.3.2. If X is finite, so is FL(X).
The proof of this theorem is left to the reader. An effective bound on

FL(X) a priori can also be given (namely, the sum of the lengths of the
formulae contained in X). The following proof of the finite model property
of P D L is due to ROHIT PARIKH and DEXTER KOZEN. The original proof of
M. J. FISCHER and R. E. LADNER procedes from the tacit assumption that
P D L is complete (and therefore that the interpretation of a * is the reflexive

transitive closure of the interpretation of a).
THEOREM 10.3.3 (Fischer & Ladner, Kozen & Parikh). P D s has the lo-
cal finite model property.
PROOF. Take a formula ~ and let S(~) be the collection of all those atoms
in the boolean algebra generated by FL(T) which are P D L - c o n s i s t e n t . These
can be represented as subsets of FL~(~), which is defined by FL(T)U{-,X: X E
FL(T)}. We write W both for the set and the conjunction over its members.
A dynamic Kripke-frame is based on S(~) via
V -~ W ~=~ COnpDLV A (0~>W

Moreover, we put /3(p):= {W : p C W}. If p is consistent there is a W E
S(T) such that ~ C W. This concludes the definition of the dynamic frame,
which we denote by | and the valuation 13. That (| 13, W) ~ ~ is a
consequence of the next three lemmas. U]
LEMMA 10.3.4. For any program a, if ConpDLV /~ (a}W then V --~ W

in (| f~).
PROOF. We do an induction on a. The basic case is covered by the
definition. Also, if a = r and V A (r is consistent, then by the axiom
(df?.) we have that V A W is consistent. V and W, being atoms, must
be equal, and r C V, and so V r W, as required. Next, let a - /3U-7.
Let Y A (/3 U "),}W be consistent. Then by (dfl2.) V A ((13)W V (7)W) is
consistent. By induction hypothesis therefore V ~-+ W oder V --~ W. Third,
assume a = 13; 7 and Y A (13;"y)W is consistent. Then by (df;.), V A (Z} (')')W
is consistent, which is to say, Y A (13}(T A (7}W. Now, T - V z ~ s ( ~ ) z , so
that we can rewrite the latter into the disjunction of all V A (/3)(Z A (7)W),
Z C S(~). Thus, by induction hypothesis there exists a Z C S(~) such that
Y ~ Z and Z -~ W. Finally, let a - 13". Assume that V A (13*)W is
consistent. Let ~ be the closure of {V} under 13-successors. We wish to show
that W C ~ . To see that put X := V z e ~ z. Then X A (/3)-X is equivalent
to the disjunction of all T A (/3)U, where T C ~3 but U r ~B. Each disjunct is
individually inconsistent by induction hypothesis, for T ~ U does not hold.
Thus, X A (/3)--X is inconsistent, from which ~-PDL X --+ [/3]X. By necessitation,
~-PDL [/~*](X --~ [~3]X), and so
~PDL V --~ [~3*](X ~ [~3]X).
By V C ~ we also have [-PDL V ~ X, and thus finally [--PDL V ~ [~*]~. This
shows that V A {13*>-'X is inconsistent, thus W C ~ , since we have assumed
that V A (13"}W is consistent. [2]
10.3. The Finite Model Property 507
LEMMA 10.3.5. For every (a>X C FL(~p) and V C S(~), (a)X 6 V iff
there exists a W C S(~) such that V -~ W and X C W.
PROOF. From left to right follows from the previous theorem. Now for
the other direction. Again, we do induction on a. If a is basic, and V 2+ W
with X C W, then ( a ) W F (a)X. Since Y A ( a ) W is consistent by definition
of a, so is V A (a}x. By definition of S(~), (a}X C V. The case of test is easy.
Now let a - 13U'y. If V ~Y+~W then either Y s W or V -~ W, from which by
induction hypothesis and the fact that (fl)X e FL(~) and (')')X 6 EL(?)) we
have (j3)X e W or (')')X e W. By (dfi_J.) we have in either case (13U'y)X 6 W.
Next let a - fi;-),. If Y ~5~ W and (13;~/)X 6 FL(~)) then ('Y)X 6 EL(Q) as
well as X E FL(~). Assume X C W. By assumption on the model there is
a Z such that Y ~-~ Z -~ W. Then by induction hypothesis (')')X C Z and
(j3)('y)X C V from which by use of (df;.) we get (j3;')')X 6 Y. Finally, let
a - / 3 * . If V ~ W there exists a chain
V - Vo g V, g g ... g v n - w.
Now, assuming X E Vn we get (f3*}X C Vn by (cls.) and so (13}(~/*}X C Vn-1,
since the latter formula is in FL-~(~)). By (cls.), (13*}X C Vn-1. Inductively,
we get (fi*)X C 17/ for all i, and so (fi*)X C V. [-]
LEMMA 10.3.6. For all X C FL-~(~) and all V C S(~), (| V} ~ X

i f f x E V.
PROOF. By induction on X- If it is a variable, then this is true by defini-
tion of j3. The induction steps for the boolean connectives are straightforward.
There remains the case X = (a)r (G(r V) ~ ( a ) r iff V -% W for some
W such that (| ~, W) ~ r By induction hypothesis, this is the case iff
r C W, and by the previous lemma the latter holds iff ( a ) r C V. [2
COROLLARY 10.3.7. PDI, has the global finite model property.
PROOF. Assume T J~PDL r Take H ~ to be the set of basic programs

occurring in T or r and put -y := U~eii, a. Then [-y*]?) ~PDL r By The-
orem 10.3.3 there is a finite model (f,/3, x} ~ [-),*]r162 based on a dynamic
frame f for P D L ( H r ) . We may assume that [ is rooted at x and that every
node is a ~/*-successor of x. Now interpret the programs of H 0 - H I arbitrarily.
This gives a dynamic frame f+ for P D L ( I I 0 ) such that (f+,13, x} ~ [7"]?); 7 r
However, by construction this means (f+,fi/ ~ ~) and so we have found a
countermodel. [2
Now let us turn to P D L ~. Start with the same definition of V -~ W.

We have to show only that V ?-+ W iff W -% V. Assume the first. Then
508 10. D y n a m i c L o g i c
V A (a~>W is consistent. From this follows with (df_~.)

v A A
This last formula must be consistent too, and so must be W A (a>V. Hence
W -% V. Conversely, assume that W 2+ V. Then W A (a> V is consistent and
so is
W A <a> (V A (a~>W)
v
by (df+.). Hence, V A (a ~ ) W is consistent and so W ?-> V. The following is

due to DIMITER VAKARELOV [217].
THEOREM 10.3.8 (Vakarelov). PDI~ ~ has the (global) finite model prop-
erty.
Notes on this section. It is known that P D L is E X P T I M E - c o m p l e t e .
T h a t it is E X P T I M E - h a r d has been shown by in M. J. FISCHER and R. E.
LADNER [69]); this follows however also immediately from the fact t h a t the
problem '~ I~-g r is equivalent to '[a*]~ --+ r C P D L ? ' . The E X P T I M E
upper bound has been shown in VAUGHAN PRATT [164]. GIUSEPPE DI GI-
ACOMO [75] has shown that P D L ~ can be constructively reduced to P D L .
This can be used to show that if P D L has interpolation, so does P D L ~. This
is given here as an exercise. It was shown by MOSHE VARDI and P. WOLPER
in [219] that satisfiability in P D L ~ can be computed in O ( 2 ~ ) - t i m e . This
bound is also easily established using constructive reduction.
E x e r c i s e 361. Show Lemma 10.3.2.
E x e r c i s e 362. Give a model theoretic proof of the fact that P D L -

[.Jm,n~ P D L ( m , n) 9
E x e r c i s e 363. Show that P D L - has interpolation if P D L has interpolation.
E x e r c i s e 364. Let a (n) denote the union of the programs a i for i < n. Let
P D L n be the extension of P D L by all axioms of the form (a*>p ++ (a(n)>p.
Show that P D L n has the finite model property without using the the finite
model property of P D L . Show furthermore that from the fact t h a t P D L has
the finite model property one can deduce that
PDL- N PDL~
nE~a
10.4. Regular Languages

The specific power that P D L has is that of the star. Without it, no gain
in expressive power is reached, and we just have a definitional extension. It
10.4. Regular Languages 509
is useful to know some basic facts about the star when dealing with P D L .
There is a well-known theorem by STEPHEN KLEENE saying that a language
is regular iff it can be evaluated using a finite state automaton. Since this
result is of great significance, we will prove it now, in full generality. For
general literature on languages and a u t o m a t a see [99], [105] or [176]. First,
let us be given a finite alphabet A, consisting of symbols, denoted here by
lower case letters such as a, b etc. A language is simply a set of strings
over A. An alternative formulation is as follows. Consider the free semigroup
generated by A, denoted by ~ s c ( A ) . A language is a subset of this semigroup;
alternatively, it is a set of terms in the language 9 (standing for concatenation)
and ~ (standing for empty word) modulo the following equations.
X'g ~ X
g'X ~,~ X
We consider the following operations on languages. For two languages L and

M, we use the set union LUM, the composition L . M - {~..~ " ~ E L, ~] E M},
and the Kleene-star L* - (-JnE~ Ln' where L n is defined inductively by L ~ -
{~} and L n+* - L . L ~. A language is called r e g u l a r if it can be obtained
from the languages O, {a}, where a E A, by use of the operations union,
concatenation and star. Another way to express this is as follows. Take the
language with the function symbols O, ~, tO, 9 and *. A r e g u l a r expression
is an expression of that language. With each regular expression R over A we
associate a language L(R) as follows.
L(O) "=
L(r "= {~}
L(a) "= {a}
L(RtO S) := L(R) tOL(S)
L(R. S) := L(R) . L(S)
L(R*) := L(R)*
We will not always distinguish in the sequel between a regular expression and
the language associated with it.
A finite a u t o m a t o n is a quadruple 9 2 - (S, So, F, T), where S is a finite
set, the set of s t a t e s , So E S the i n i t i a l s t a t e , F a_ S the set of a c c e p t i n g
s t a t e s , and T 9 S • A ~ 2 s a function, the (nondeterministic) t r a n s i t i o n
f u n c t i o n . We define the transition function inductively by
a) "= a)
:=
92 a c c e p t s ~ if~(S0,~) A F ~: O, otherwise 92 r e j e c t s ~. Obviously, the set of

accepted words of 92 is a language, denoted by L(92). If ~(S, ~) is a singleton
set for any input ~ and state S, then the a u t o m a t o n is called d e t e r m i n i s t i c .
In t h a t case the transition function can be constructed as a function from
pairs of states and letters (or words) to states. A triple (S, a, T}, also written
S -~ T, such that T E ~'(S, a) is called a t r a n s i t i o n of the a u t o m a t o n . ~- is
completely determined by its transitions.
LEMMA 10.4.1. For every finite state automaton 92 there exists a deter-
ministic automaton ~ such that k ( ~ ) - [_(92).
PROOF. Let 9 4 - (S, So,F,~-) be given. Let ~1" " - 2 s, To := {So}, G :=
{ H C_ S" H N F ~: O} and for E C_ S and a C A let c r ( E , a ) : = U(w(S,a)" S c
E}. It is verified by induction that for every word 2 C L* also
- U
SEE
Put ~-- (V, T0, G, cr). It is then clear by the definitions that [_(~3) - L(92).
V]
For any regular language L an automaton 92 can be constructed such

that L - L(Pl). Namely, for the one-letter languages {a} construct a three-
state automaton based on S- {So, S,, $2}, where F "- {S,} and such that
r(So, a ) " - - S,, but T(Si, ~):= $2 if g ~: a or Si ~: So. For the union assume
that 92i _ (Si, S i , F i , ~_i> are a u t o m a t a such that L(92i) - L ~, i - 1, 2. Then
construct the a u t o m a t o n
:= {s • s (So*, F • S u S • F •
where T 1 • T 2 ( ( S , T > , a ) " - ('rl(S,a), ~-2(T, a)}. It is easy to check t h a t [-(~) -
L1UL 2. For the composition L 1 .L 2 of the two languages perform the following
construction. For each T C F 1 add a copy 942 of 9.12 to the a u t o m a t o n ,
identifying the state T with (So)T as follows. For every transition into T add
a transition into (So)T and finally remove T. The initial state is S~ and the
new accepting states are all states from F~, T C F 1. Finally, for the star
just add for each transition S 2~ T where T E F a transition S 2~ So, and
remove all T C F different from So. The new set of accepting states is {So}.
This new a u t o m a t o n accepts (L1) *. This shows that all regular languages are
languages accepted by some finite state automaton.
THEOREM 10.4.2 (Kleene). Let A be a finite alphabet. A language L C_

A* is regular iff there exists a finite state automaton 94 such that L - [-(91).
PROOF. Let 94 - (S, So, F, 7} be a finite state automaton. We can assume
that it is deterministic. Let for simplicity be S - { 0 , 1 , . . . , n - 1}, and
So - 0. Take a set Xi of variables ranging over subsets of A*. The intended
interpretation of Xi is t h a t it denotes the set of words 2 such t h a t 7(0, 2) -- i.

The a u t o m a t o n is determined by a set of equations over these variables of
the following form. For e a c h i < n a n d j < n l e t aij - { a " i - - ~ j , a E A } .
Each aij is a regular language, being a (possibly empty) union of one-letter
languages. We then have
Xo = e U Xo "aoo U Xl 9alo U . . . U X n - 1 9an-l,o

X1 = Xo "aol U X1 9 U . . . U X n - 1 9a~-1,1
. . . .
Xn-i = Xo "ao,n-i U X i 9 a i , n - 1 U . . . U X n - - i " an--i,n--i
It is easy to see t h a t any solution Lo, L~ etc. for this system of equations
has the property t h a t the set of words ~ such that T(O, ~) - i is exactly the
solution for Xi. Furthermore, there can be only one such solution. To see this
suppose t h a t we have an equation of the form
X-X.RUY
where R is a regular expression over A containing no variables and Y a regular

expression not containing X. This equation has the same solutions as
X-Y.R*
Namely, take a word from Y . R*. It is of the form ~ . r'0" r'l 9 r'2.., r'n-1,
r-'/ E R and ~ C Y. T h e n by the first equation g C X, since Y C_ X, and then
g.r-'0 C X, s i n c e X . R C_ X. S o g . r ' 0 . r ' l C X, and so on. H e n c e 2 C X.
Conversely, let ~ E X. Then either ~ C Y, or s E X . R , t h a t is to say,
= 2~'' ~'0 for some ~'0 E R and some 2~' C X. Again, either ~' C Y or it is of
the form ~ " . r'l, ~'1 C R and ~" E X. This process must come to a halt, and
this is when we have a decomposition of 2 into ~. rvn_l 9r'n-2" . . . " r'0, r/ C R
and ffC Y. Thus s E Y . R*.
A r m e d with this reduction we can solve this system of equations in the
usual way. We start with Xn-1 and solve the equation for this variable. We
insert this solution for Xn-1 into the remaining equations, obtaining a new
system of equations with less variables. We continue this with X n - 2 and forth
until we have an explicit solution for X0, containing no variables. This we
now insert back again for X1 obtaining a regular expression for X1, and so
on. Finally, we have L(92) U i E F X i , and inserting the concrete solutions
-
gives us a regular expression for the language accepted by 92. []
This theorem has numerous consequences. Let us list a few. Given a word
e - a(0). a ( 1 ) . . . . , a ( n - 1), we put e "~ " - a ( n - 1 ) . . . . . a(1). a(0) and call it
the t r a n s p o s e of 2~. For a language L we write L "~ := {~'~ 9 ~ C L} and call
that the t r a n s p o s e of L.
THEOREM 10.4.3. The intersection of two regular languages is a regular

language. The transpose of a regular language is a regular language.
The proof is left as an exercise. Given a finite a u t o m a t o n , and two states
i and j, let k[i,j] := { 2 : T ( ~ , i ) - - j}. Then the proof m e t h o d establishes
t h a t all k[i, j] are regular. The languages k[i, j] satisfy the equations
L[i,j] - U L[i,k]. L[k,j].

k
Thus L[i, k].L[k,j] C L[i,j]. For a deterministic a u t o m a t o n L[i,j]ML[i,j'] = 0
whenever j ~: j'. Now, define the p r e f i x c l o s u r e L p of a language L to
be the set { ~ : ( 3 ~ ) ( ~ . g e L)} and the suffix c l o s u r e L s to be the set
{ ~ : (3~)(~. 2 e L)}. If L is regular, so are L p and L s. Namely, take an
a u t o m a t o n P2 = (S, So, F, 7} recognizing L. L p is the union of all L[S0, i] such
t h a t L[i, j] is not empty for some j E F. This is regular. For L s, observe t h a t
L s is the union of L[i,j] such t h a t j C F. All of the latter are regular, and
so is L s. There is an explicit way to compute the suffix closure of a regular
expression, which runs as follows.
a s "-- ~ Ua
.= R .SUS
(RUS) .-
(R*) s .= RS.R *
The correctness of this definition is seen by induction on the definition of the
regular expression. Clearly, a suffix of a, a C A, is either a itself or the e m p t y
word, so a s is correctly defined. Now let ~ be suffix of a word g of L U M .
Then i f g E L , u T E L s, and i f g c M t h e n u T C M s. Next l e t ~ C ( L - M ) s,
t h a t is, for some g, g . ~ C L . M . There exist ~ C L a n d ~ C M s u c h t h a t
g . ~ - ~ . ~. T h e n either g is a prefix of ~ or ~ is a prefix of g. If the first is
the case we have a decomposition 2 - g. 2' and so g. ~ - g. 2 ' . ~ from which
- ~ ' . ~. This means t h a t ~ C L s 9M. Now let ~ be a prefix of g. T h e n we
have a decomposition ~7- ~. g' and so ~. ~ 9u7 - 2 . ~ from which g' 9~ - ~.
Thus u7 C M s.
Recall from the previous section the definition of the Fischer-Ladner clo-
sure. A p a r t from the usual syntactic closure it involves a certain kind of
closure under subprograms. In a sense to be made precise now, the Fischer-
Ladner closure of a formula p involves the closure under suffixes of each pro-
gram c~ occurring in ~.
P R O P O S I T I O N 10.4.4. Let (a)p be a P D L - f o r m u l a . Then (aS}p is deduc-

tively equivalent to the disjunction of all formulae in FL((a)p).
PROOF. We do induction on a. Suppose a is elementary. T h e n by (fld.)
we have V FL((a}p) - (a)p V p - (aS)p. Next suppose t h a t a - 3; 0'. T h e n
V FL((a)p) is equivalent to V FL((~)(Tip). Furthermore, FL((~)(7)P) =

(FL((~)q))[(V)p/q] t2 FL((v)p). By induction hypothesis, the disjunction over
the first is nothing but ((~S)q)[(7)p/q] = (f~s)(7)p, which is equivalent in
P D L to (f~s; V)P. The second is nothing but (7S)p. The disjunction is equiv-
alent to ((~s; V)U VS)p and that had to be shown. Next consider c~ = ~ t2 7.
Then V FL((a)p) is equivalent to V FL((f~)p) v V FL((v)p). By induction
hypothesis the latter is equivalent to (~S)p V (7S)p, and so equivalent to
(aS}p, as promised. Finally, we turn to a = f~*. We have that V Fi((~*)p)
is equivalent to the disjunction of FL((~*)p) and FL((~i(~*)p). The first
does not reduce again except with the rule (fl*.) so that this disjunction
is equivalent to (~*)p V V FL((~)(~*)p). Now notice that FL((~)(f~*)p) =
FL((~)q)[(~*)p/q] U {p}. Therefore, the last of the two is equivalent by in-
duction hypothesis to the disjunction of (~s)i~.)p and (f~*)p. Thus the whole
disjunction is equivalent to the formula ((f~s; ~*)U ~*)p, which is nothing but
E x e r c i s e 365. Show that the reduction algorithm in the proof of Kleene's

Theorem does not depend on the assumption that 92 is deterministic. Thus,
another proof is found that languages accepted by nondeterministic finite
state automata are the same as those accepted by a finite state deterministic
automaton.
E x e r c i s e 366. A regular automaton is said to be an automaton with e-

transitions if there are transitions of the form S -~ T. Show that any language
accepted by a finite state automaton with e-transitions is regular.
E x e r c i s e 367. Spell out in detail the constructions to show that L1. L2 and
L~ are languages accepted by a finite state automaton whenever L1 and L2
are.
E x e r c i s e 368. Use the method of solving equations to show that all regular
expressions define languages accepted by some finite state automaton. Hint.
Let R be a regular language and R the term. Then start with the equations
X1 - X0" R, X0 - e. Now reduce the regular expression R by introducing
new variables. For example, if R = R I. R ' , then Xi = R ~. R ' . Xj U Y.
Introduce a variable Xj and add instead the equations Xi = RI.X~ U Y,
Xj - R ' . Xj. Finally, if the system of equations is such that the expressions
Xi 9 R are of the form where R is simply of the form a, a C A, or a union
thereof, see how you can define a finite state automaton accepting R.
E x e r c i s e 370. Try to define the intersection and transpose of a language

explicitly on the regular expression. Hint. This is complicated for the inter-
section, but should be relatively easy for the transpose.
10.5. A n E v a l u a t i o n P r o c e d u r e
The problem that we are now going to attack is that of evaluation of
P D L - f o r m u l a e in a given model. This will provide the basis for some in-
teresting theorems in dynamic logic. Before we start let us see how we can
simplify PDL-formulae. What is particularly interesting is to obtain some
canonical form for the programs. First, we define the dynamic complexity of
a formula as follows.
DEFINITION 10.5.1. A formula ~ is of d y n a m i c c o m p l e x i t y 0 if it is

a boolean combination of variables. ~ is of d y n a m i c c o m p l e x i t y d + 1 if
it is not of dynamic complexity <_ d and a boolean combination of formulae
of the form (c~)r where r is of dynamic complexity <_ d, and a is composed
from basic programs and tests over formulae of dynamic complexity <_ d using
program union, composition and star. We denote the dynamic complexity of
a formula ~ by dc(~).
Call a program 7 a c h a i n if it is composed from basic programs and tests
using only composition. A chain is s e m i r e g u l a r if it is of the form ~ or c~; ~,~,
where 7~ is a chain and where ~ is a basic program, not a test. A chain is
called r e g u l a r if it is of the form ~; ~? for some basic program ~, or of the
form ~; ~?; 7~ for some regular chain 7~. (This is obviously well-defined, since
~ is a chain of smaller length than 7.) The first lemma concerns the reduction
of star-free programs to regular chains.
LEMMA 10.5.2. Suppose X = (c~)~ is a formula where c~ is free of star.
(1.) There exist semiregular chains ~/i, i < n, such that X is equivalent to
a disjunction of the formulae (~/i)~. (2.) There exist regular chains 5i, i <
m, and formulae r of dynamic complexity less than de(x), such that X is
equivalent in P D L to the disjunction of r A (5i)~.
PROOF. First, by (diD.) and (dr;.) we are allowed to convert (c~; (Zl t_J
Z2); ~/)~ into the disjunction of (c~;/31; ~')~ and (a;132; ~)~. Thus, we can
replace a star-free program by the disjunction of chains. Next, we can replace
(O~;~/)17;r by (a; (r A r ~)~p. Finally, we can rewrite (r a ) ~ by
r A (a)~. Perform these reductions in succession and this gives the desired
disjunction into formulae using regular chains. Z]
The star operator introduces some complications. First of all, however,
notice the following identities. (Recall that skip := -V?.)
LEMMA 10.5.3. The following holds.
1. (a U/3)* - (a*;/3*)*.
10.5. An Evaluation Procedure 515
2. (c~;/3)* -- skip U c~; (/3; c~)*;~.

3. (~p?)* - skip.
Using these identities we can do the following. Take the smallest subfor-
mula of the form c~*. T h e n c~ can be assumed to be a disjunction of semiregular
chains. By (1.) of the above proposition, we can remove the disjunction in a*,
so t h a t we are down to the case where a contains a single semiregular chain.
If it is not already regular, then it is of the form r or it the form r The
latter is dealt with using (3.). It can be replaced by skip. Now consider the
first case. Use (2.) to rewrite (r into skip U r (7;r 7 begins
with a basic program, so it is a regular chain. After these manipulations we
have t h a t the smallest starred subformulae are of the form c~* with c~ a reg-
ular chain. This we rewrite again into skip U c~+. Now we continue with the
smallest subprogram of the form ~* in the formula thus obtained. It may now
contain a program of the form/3 +. However, the same manipulations can be
performed. Iterating this we get a formula which we call proper. A definition
recursive in the dynamic complexity runs as follows.
DEFINITION 10.5.4. A program c~ is called p r o p e r if it is composed from

regular chains using tests on proper formulae, with the help of composition,
union, and +. A formula is proper if it is a boolean combination of formulae
(c~)r where c~ and r are proper.
PROPOSITION 10.5.5. Every formula of P D L is PDL-equivalent to a

proper formula.
The use of proper formulae is explained by taking a look at p a t h sets
defined by programs. First, notice that programs do not simply define paths
in a model, but they define path sets since they are regular expressions. We
give a syntactic description in terms of chains.
ch (V) "= {7} if 7 is a chain

ch(v U S) "- ch(7) U ch(5)
= ch(
To simplify the notation, we can assume that a regular chain is a composition

of programs of the form 7i; ~i?. If not, then add suitable tests T? in between
basic programs. W i t h such simplification, take a model (f, a, ~} and a point
x0. A p a t h of length n starting at x0 is a sequence ~ - (xi 9 i < n + 1) such
t h a t 7r(0) - x0 and for all i < n there exists a basic 7i such t h a t xi 24 xi+l.
Now, a p a t h 7r is said to fall u n d e r a regular chain 5, where 5 is of the form
(~--,~0;r162 ; ~n; ~)n? ,
if x~ 24 xi+l for all i < n, and ( f , a , 3, xi+l) ~ r If 7r falls under 5 we

also say t h a t 7r is a c o m p u t a t i o n t r a c e of 5. (This is called a computation
sequence in [119].) This notion is extended to all proper programs as follows.

A p a t h falls u n d e r or is a c o m p u t a t i o n t r a c e of a proper program c~ if it
falls under any one of ch(a). Finally, if one wishes, the notion can also be
extended to arbitrary a, by saying that ~ falls under a if it falls under any
which is proper and equivalent to a.
A first application is the proof that D P D L has the finite model prop-
erty. We will show that D P D L is constructively reducible to P D L using the
following global reduction function
XD(A) := {-'[r -+ [r : [r r FL[A], r r IIo}.
These reduction sets split. Now assume that we have a finite model (f, wo)
such t h a t XD(T) holds globally while ~ holds at wo. Starting with this con-
figuration, we will produce a finite D P D L - m o d e l for ~, based on the basic
programs occurring in ~. (So, we assume rio to be finite.) Moreover, we will
assume that ~ to be proper. To see that this is harmless even for a proof of
interpolation observe that by Proposition 10.5.5 we can turn a into a proper
program by some equivalence transformations and insertion of some tests
T?. The construction is essentially due to MORDECHAI BEN-ARI, JOSEPH
HALPERN and AMIR PNUELI [7]. First, we will unravel [ at x. The result-
ing frame is called n. The unravelling is strictly speaking unnecessary, we
do it only for technical convenience. The construction of the actual model is
inductive. The intermediate objects are not models, however, but what we
call pseudomodels. A p s e u d o m o d e l is a pair ([, b) where b assigns a set of
formulae to each x C f. (f, b, x) ~ ~ is defined inductively. However, it is a
priori not excluded that the valuation is inconsistent. So, we work with sets
of t r u t h values. Given ~Y~ := (f, b) we define v~(x, x) C_ {0, 1, .} inductively.
1C v~(p,x) r p E b(x),
O C v~(p,x) r -~p C b(x)
For X = -~X', let 1 C v ~ ( x , x ) iff 0 C v~(x',x) and 0 C v ~ ( x , x ) iff 1 C
v~(x',x). For X = x I A X 2 put v ~ ( x , x ) := {~M~ : ~ E v~(Xl,X),~ C
v~()(.2,x)}. Finally, let Z = (a)X'. We put 1 e v ~ ( x , x ) iff there exists 13 and
~/such t h a t (#; ~,)p ~ <a>p C P D L and a path from x to y falling under 13,
such t h a t <7)Z' e b(y). Otherwise, 0 e v~(x,x). (Clearly, for [a]Z' the dual
of this condition is chosen.) Finally, we write ~ ~ X if 1 E v ~ ( x , x ) . Call
a formula a s t r i c t d i a m o n d f o r m u l a if it is not of the form -~-~, ~ A r
V r (~?)r or [a]r Say that a pseudomodel is p r o p e r if b(x) always is a
set of strict diamond formulae.
LEMMA 10.5.6. Let ~ = (~, b) be proper. Then the sets A(x) := {X :
v ~ ( x , x ) = 1} are consistent.
For a proof of this lemma notice that a strict diamond formula never gets
the value {0, 1}, by definition. So A(z) can be inconsistent iff there exists
a variable p such that p,-~p E b(z). If b(x) only contains variables for each
x E f, define 3 ( p ) : - { x : p e b(x)}. Then ([, b,x) ~ ~p iff ([,/3, x) ~ ~a, as is
easily verified.
Now we start the construction. P u t P0 : - {To}, where wo is the root of
n. Furthermore, let P0 be the subframe based on P0 (that is, all relations are
empty); and finally,
bo (wo) " - {X E FL-~(~) "X a variable or strict diamond formula

and (n, ~, (wo)> ~ X}.
Starting with (P0, bo) we will construct a sequence of pseudomodels (p~, bn)
and sets Ln such that
1. Pn is based on a finite subset P,~ of n.
2. bn(x) C_ FL(~a) for all x C Pn.
3. L,~ is the set of all x for which there exists i < n such t h a t x is a leaf
of Pi.
4. (p~,bn} is proper. Moreover, bn(x) contains formulae which are not
variables only if x has no successors.
5. For X e FL(~) we have (p~, b~, x} ~ X iff (n,Z,x} ~ X.
6. Pn is a D P D L - f r a m e .
These claims are immediate for <P0, b0>. We will also construct auxiliary sets
Ln. L o : - {w0}. Now let (pn, bn} and Ln already be constructed. Pick a
node x in Pn without successors. C a s e 1. Along a path from w0 to x there
exists a y E Ln and y # x such that (Pn, bn,y} ~ X iff (Pn, bn,z} ~ X for all
X C FL(~). Then y E n n - 1 . P u t Pn+l "- P n - {x}. For r e rio put z ~ z'
in the new frame if either z _~4 z' in the old frame, or z' - y and z ~ x in the
old frame, bn+l : - bn F Pn+l. Ln+l : - L n - {x}. In this case, the properties
(1.) - (4.) and (6.) are immediate. We will verify (5.) in L e m m a 10.5.8.
C a s e 2. Along no path from wo to x there is a y C Ln different from x such
that y satisfies the same set of formulae in FL(~a) as does x. Pick a formula
-~[a]r in bn (x). There exist points si in n, i < k + 1, and a computation trace
TO7"; VO; 7"17'; V1; . . . Tk-17'; V k - 1
falling under a; 9 r such that so - x, si 74 Si+l in n, (n,#, si) ~ 7i for all

i < k, and (n,/3, Sk} ~ r (It is here that we need ~a to be proper. This
assumption is not essential, but allows for simpler statement of the facts. If
is not proper there might not be a ~-i? for certain i, but the proof does in
no way depend on that.) For each i < k and each basic ~ we let t(i, ~) be a
point of n such that si ~ t(i, ~). If ~ - 7i, then t(i, ~):= Si+l. P u t
T(x,-~[a]r " - {si" i < k + 1} U {t(i,~)" i < k,r C H o } .

518 10. Dynamic Logic
We call T(x,-~[a]r the t h o r n sprouting at x for -~[a]r We let pn(x,-~[a]r

be the result of adding the thorn sprouting at x for ~[a]r to Pn. A map c
is defined by c(x) : - bn(x) - {-~[a]r c(z) : - bn(z) if z e Pn -- {X}, and
c(z) = {-~[a']r :-~[a']r C FL(~), (n,~,z} ~ ~} for z e T(x,-,[a]r a leaf,
c(z) := { p : p e var(~),z e 13(p)}, z not a leaf. As one can show, the new
pseudomodel satisfies all properties but (4.) in the above list. Therefore, we
would like to add all thorns sprouting at x for a (strict diamond) formula of
the form -~[a]r However, the resulting frame will not be a D P D L - f r a m e .
Therefore, some care has to be exercised in defining the thorns.
LEMMA 10.5.7. Let ~ be a formula, and

([,/3) ~ {-~[(]X ~ [~]-~X : [r e fL(qa), ~ e II0}.
Let-~[a]r be in fL-~(~a) and ([,/3, x) ~ -~[a]r x r y. If there is a path

starting with x ~-~z falling under a computation trace for a; ~r then there
is a formula -~[a']r in FL-~(~p) such that for any path 7r starting at z falling
under a computation trace for a';-~r the path x~Tr falls under a computation
trace for a; 9 r Furthermore, ([,13, y) ~ -~[a']r and there is a path falling
under a computation trace for a;-~r starting at x leading through y.
Thus, let ~ [ a ' ] r be a formula that holds at x in n. By repeated use of
the lemma one can show that there exists a path falling under a computation
trace for a; 7 r that is either fully contained in (i) T(x,-~[a]r or (ii) leads
through a leaf y of T(x,-~[a]r In case (i), we do nothing. In case (ii), rather
than sprouting a thorn for ~ [ a ' ] r at x in Pn, we sprout a thorn at y for some
suitable -~[a"]r in pn(x,-~[a]r We perform this sprouting for all formulae
-~[a']r in b,~(x). Let T(x) be the union of the thorns. Let y ~ z in T(x)
iff y r z in n. This defines the frame t(x). It is a D P D L - f r a m e ; it is a
tree. We put Pn+l := P n U T ( x ) , and y-~r z iffy, z C P,~ and y-~r z i n p n
or y , z e T(x) and y ~ z in t(z). Finally, b n + l ( y ) " - b~(y) if y e Pn - {x},
bn+l(x) := {p: p e bn(z)}, b n + l ( y ) : = {p: (n,13, y) ~ p} if y is in T(x) but
not a leaf of t(z), and finally b~+l (y) := {~[a]r e FL~(~) : (n, Z, Y} ~ -~[a]r
if y is a leaf of t(x). Finally, n , + l := n,~ U {y: y a leaf of t(x)}.
Now, Pn+l is finite. Moreover, one can estimate the size of T(x) in the
following way. For each -~[a]r we have added a computation trace. Since
n is the unravelling of [, a bound on a computation trace for a; r can be
given that depends only on the size of f and ~[a]r b,~+l(y) is a subset
of FL-~(~a) by construction; moreover, if y is not a leaf, then bn+l(y) only
contains variables, bn+l(y) contains by construction only formulae of the
form -~[a]r Furthermore, P,~+I is a D P D L - f r a m e , as is straightforward to
verify. It remains to be seen that (Pn+l, bn+l, Y> ~ X iff (n,/3, y> ~ X for all X
in FL-~(~a). This follows from the lemma below.
LEMMA 10.5.8. For all n, all X E FL~(~) and all y E Pn:

($) y> x bn, y> x
PROOF. By induction on the dynamic complexity of X. For variables this
holds by construction, and the only problematic step is for the strict diamond
formulae. Let c be the dynamic complexity of -~[a]X, c > 0, and let the claim
have been shown for formulae of complexity < c. In the definition of Pn+l
two cases have been distinguished. C a s e 1. Pn+l is obtained by removing
a point (and adding a transition). Let (Pn+l, bn+l, Y) ~ -~[a]X. And assume
that Pn+l - P n - {x} and that the transition v ~ w has been added for some
w. Then in Pn, v -~r x, and x and w satisfy the same strict diamond formulae
and the same variables. Then y ~ x. Consider a path from y falling under a
computation trace for a; ~Z?.
C a s e l a . The path does not go through v. Then it is a path in Pn, by
construction. By inductive hypothesis, since the computation trace involves
tests for formulae of dynamic complexity < c, the path falls under the same
computation trace in (Pn, bn). In this case we are done, for then clearly
(p,~, b~, y) ~ ~[a]Z. Now assume (pn, b~, y) ~ -~[a]Z. Consider a path 7r falling
under a computation trace in (pn, bn>. Then it is a path in P,~+I, except if
the path ends in x. In that case, let 7r' differ from 7r only in that the endpoint
(which is x) is replaced by w. Otherwise, 7r' := 7r. By inductive hypothesis,
the path falls under a computation trace for a; ~X? in (P,~+I, bn+l).
C a s e l b . Suppose that the path goes through v. Then (Pn+l, bn+l, y)
~[a]X by virtue of the fact that either we have (i) (Pn+l, bn+l, v) ~ ~X or (ii)
(P,~+l, b,~+l, v) ~ -~[a']X for some strict diamond formula -~[a']X in FL~(~v).
Here as in sequel, a ~ is such that there exists a computation trace cr such that
(i) the path from y to v falls under a and (ii) for every computation trace 7
for a' or; T is a computation trace for a. In case (i) we are done by induc-
tive hypothesis. In case (ii) observe that then (Pn+l,b,,+l,W) ~ r
for some r of dynamic complexity < c and some a ~' such that for every
computation trace ~- for a ~' the trace ~; r T is a computation trace for a'.
(a ~' may be identical to skip), w E Ln-1. Therefore, by construction of
the sequence (Pi, bi/, there exists a path falling under a computation trace
for a'~;-~X? inside Pn. By inductive hypothesis, since it involves tests of de-
gree < c and the fact that the path is a path in p~, this is a path falling
under a computation trace in Pn. So, (Pn, ha, w) ~ ~[o~"]X. Furthermore,
(p,~,bn,w) ~ r It follows that (p~,b~,x) ~ r by construction of
(Pn+l, b,~+l). Hence (Pn, b~,v) ~ ~[a']X, since v ~ x. From this follows
finally that (p,~, bn, y) ~ -~[c~]X.
C a s e 2. Let (Pn+l, b~+z,y) ~ -~[a]X. Consider a path 7r falling under a
computation trace for a; X?. If that path is inside Pn then by inductive hypoth-
esis, (p=, b=, y) ~ ~[a]X. Otherwise, if the path is inside t(x), then (pn, bn, y)
--[a]X be definition, using the previous lemma. Finally, if the path is not prop-
erly contained in either, then it leads through x. So, (Pn+z, bn+z, y) ~ ~[a]X
holds by virtue of the fact that (Pn+z, bn+z,x) P ~[a']X for some ~[c~']X in
FL"(~). The latter is equivalent to (t(x),bn+z [T(x),y> ~ ~[a']X which in
turn means that ~[a']X C b~(x). By (5.) since (p~, bn,x> ~ -~[a']X, we have
(Pn, b,~, y) ~ ~[a]X. (Induction on the length of the path leading from y to
x.) Now, assume that (Pn, b~, y> ~ -"[a]X. Then if y is not a leaf, there is
a path in Pn falling under a computation trace for a;--Z?. This path falls
under the same computation trace for a;--X? in (P~+z, bn+l}. By inductive
hypothesis, therefore, (Pn+l, bn+z, y> ~ -~[a]X. If y is a leaf different from x,
then ~[a]X e bn(y) and so also -~[~]X e bn+l (y). Finally, if y = x the claim
holds by construction of t(x) and Pn+z. []
THEOREM 10.5.9 (Ben-Ari &: Halpern & Pnueli). DPDL has the finite
model property.
Now let us return to the question of evaluating a formula in a model.
We will propose a special procedure, which we will use in Section 10.7. It is
effective, but we make no claims about its efficiency. We may assume that the
formula is proper. Thus take a formula ~ and a model 9Y~ = (f, or, ~). First,
put ~ into the form
VA
i<k j<k~
where each r j) is of the form (a(i, j))x(i, j), or [a(i, j)]x(i, j). Moreover,
we can rewrite these formulae into (a(i,j); x(i,j)?)T and [a(i,j);-,x(i,j)?]_L.
Without loss of generality, we can assume that r j) = (a(i, j)> T or r j) =
-~(a(i, j)>T. Thus, ~ turns into a boolean combination of existence statements
for paths. The strategy we are going to use is simply put the following.
Enumerate all possible paths in the model, and see whether they fall under
one of the descriptions. To make this work, several things have to be assured.
First, that we can in principle enumerate all the paths, second that we are
able to see whether they fall under one of these descriptions, that is, whether
they are of the form a(i, j); and third, since the a(i, j) might use the star, we
must find a way to make the procedure finite if the model is finite as well.
We will deal with these problems in the following way. Suppose that we
have a path 7r of length n. Then there are basic programs ~(i) such that
xi r Xi+l. Given a(i,j), under what condition does 7r fall under a(i,j)? If
a(i, j) is free of stars, this is easy. For then a(i, j) is a disjunction of chains,
each chain being composed from basic programs and tests. Moroever, we can
assume that it is a composition of programs of the form ~i; r where (i is
basic. Then check whether xi -~ xi+z for all i < n, and whether or not
(f, ~,/3, xi+z} ~ ~ . The latter is a task similar to the evaluation of the main
formula ~; however, r has dynamic complexity less than the complexity of
~. Assuming that the latter task can be achieved by the same method as we
describe now at least for formulae of lesser dynamic complexity that ~p, we
have succeeded. Of course, the case that the r have dynamic complexity 0 is
granted to us. We just need to see whether a boolean combination of variables
holds at a node. We use/3 to tell us so. Now, if the program contains stars,
we can do the following. Let us assume that the frame has at most p points.
Then any pair of points related via a* can be related via a <p, which is the
union of all a n such that n _< p. For any sequence of length > p must contain
a repetition, which we could have avoided. Thus, we can simply replace the
star by a suitable disjunction. This solves the problem of deciding whether a
path falls under a path description.
The next problem is that of enumerating the paths. Recall that we were
able to replace the star, so we are down to a finite disjunction of chains. This
problem is therefore solved if for given maximum length c of a chain we can
successfully enumerate all possible paths of length _< c. The choice of c makes
sure we really enumerate all paths that possibly stand a chance of falling
under a description a(i, j). Now to enumerate these paths starting at a given
point x0, let us first of all assume that given a node w and a basic program
~, the ~-successors of w are ordered, or ranked. Moreover, we consider the
basic programs ordered, say by their enumeration as ~0, r etc. Then, the first
degree successors of w are ordered as follows, y precedes z if either w -~ y
and there is no j _< i such that w -~ z, or else w -~ y, z and y is ranked higher
than z as a ~i-successor. Finally, we rank the paths starting at x0 as follows.
7r ~ p if either (i) p is a prefix of 7r or (ii) there exists a largest i such that
7r(i) ---- p(i) and 7r(i § 1) precedes p(i § 1) as a first degree successor of rr(i)
(= p(i)). This ranking corresponds to a depth-first search. Starting at x0 we
always pick the successors of highest priority, going as deep as we can, but at
most c steps deep. After that we do what is known as backtracking. To get
the next path we go back to the last point 7r(i) (in the numeration of the path,
i. e. we choose the highest i with this property) where we could have chosen
a successor of lower priority rather than 7r(i + 1) and there we go instead to
the point immediately lower in priority than 7r(i + 1). After this choice we
start picking successors of highest rank again, up to depth c, or as deep as we
can otherwise. If there exists no point lower in priority t h a t 7r(i + 1), the new
path ends at 7r(i), that is, has length i. In this way we really enumerate all
paths starting at x0 of length _< c.
The problem of deciding whether qD is accepted at x0 in a model of size
p is now solved as follows. Replace all programs a* by a -<p. Compute the
constant c, the m a x i m u m nesting of basic programs in qD. Now open a table
for qD. We understand a table for qD to be a bit-vector consisting of one bit
per subformula ~b(i,j) - ( - , ) ( a ( i , j ) ) T . The table is initialized by putting
t ( i , j ) - O, if r is of the form ( a ( i , j ) ) T , and by putting t ( i , j ) - 1 if
522 10. D y n a m i c L o g i c
r is of the form -~<c~(i,j)>T. Now start generating all paths of length

< c and check whether they fall under one of the c~(i, j). This may require a
recursive step, that is, opening up tables for formulae of lower rank, but this
is immaterial. Now suppose we have generated a new path 7r. Then for each
index (i, j) we check whether or not the path falls under a(i,j). If so, we
overwrite the entry t(i,j) to 1; if not, t(i,j) remains untouched. At the end
of this procedure, <a(i,j)>T comes out true if t(i,j) - 1, because the latter
means t(i,j) has been overwritten at least once, so must have found a p a t h
falling under a(i, j). It comes out false if t(i, j) - 0, because t h a t means we
have never touched t(i, j), which in turn means that no path has been found.
Then for -~<a(i, j)>T we also know whether or not it is accepted. The number
of tables that we need to keep at a given moment of time is bounded by the
dynamic complexity of ~. This means that apart from the procedure that
generates paths we only need a finite memory to do the bookkeeping.
If ~ is proper we can deduce a number of useful properties of the com-
p u t a t i o n procedure. Suppose that we want to evaluate ~ at x0. The p a t h
generator starts with generating paths going through the successor of p with
highest priority. Call it x i. Then it chooses a successor of highest priority,
x2. And so on. This defines first of all a linear order of the paths generated,
which we can also represent as an enumeration function g 9 j -+ f, j some
natural number. Say that the computation e x i t s u at i, i < j, if the end
point of g(i) is u, and for no larger i' u is the endpoint of g(i'). Now, define a
priority E over the transit of x0 by saying that v E w if the c o m p u t a t i o n exits
v at i and the computation exits w at i' and i < i'. This is defined only for
the main computation, generating paths from the point x0, at which we want
to evaluate the main formula ~. However, in the course of evaluating ~ the
procedure generates a path with intermediate points and calls on subroutines
to check certain subformulae of lower dynamic complexity at certain nodes u.
This procedure may by itself also start generating paths. The set of a c t i v e
n o d e s of a computation at time t is the set of nodes at the main p a t h active
at t, together with the set of nodes active in the subroutine just at work at
time t, plus whatever subroutines are called at t. The set of active nodes is a
union of at most d paths, d the dynamic complexity of ~. Now say - - finally
J that the overall computation t o t a l l y e x i t s x at t if x is active in it at t,
but for no point of time t' > t, x is active at t'.
Notes on this section. Already in the abovementioned paper, BEN-ARI,
HALPERN and PNUELI have shown that D P D L is E X P T I M E - c o m p l e t e . This
can be established using constructive reduction. The procedure described at
the end of this section is called a model checking procedure. Model checking,
especially its complexity, is an important topic in temporal logic (see E. A.
EMERSON [54]). In general, the model checking problem is as follows" given
a finite model if2 and a formula ~, is 9Y~ a model for p? Of course, this is in
10.6. The Unanswered Question 523
general a decidable problem, so one is interested in the number of steps needed

to compute the answer, or alternatively, in the amount of storage space. This
number may depend both on the size of ff)l as well as the length of T. The
procedure described above is not very efficient. If we want to save space, we
should precompute the relations corresponding to the programs occurring in
T. This can be done iteratively, by induction on FL(T ). In tandem, we shall
evaluate the assignments of the subformulae. All this needs space quadratic
in the size of the model and linear in the size of FL(T ). It is easy to see
that card(EL(p)) is linear in the length of T. The time is a polynomial in
the sum of the size of the model and the length of the formula (this was
already observed in the paper by M. J. FISCHEa and R. E. LADNER [69]).
Notice that despite these low bounds, satisfiability of formulae in P D L takes
exponentially many steps to compute (see Section 10.3). For the models to
be built are in the worst case exponential in the length of T.
E x e r c i s e 373. Show that D P D L with converse does not possess the finite
model property. (This is due to JOSEPH HALPERN.)
10.6. T h e U n a n s w e r e d Q u e s t i o n
The remaining two sections will deal mainly with the problem of interpo-
lation for P D L . This is one of the major open problems in this area. Twice a
solution has been announced, in [1381 and in [33], but in neither case was it
possible to verify the argument. The argument of LEIVANT makes use of the
fact that if T ~-PDL ~) then we can bound the size of a possible countermodel
so that the star a* only needs to search up to a depth d which depends on T
and ~. Once that is done, we have reduced P D L to E P D L , which definitely
has interpolation because it is a notational variant of polymodal K. However,
this is t a n t a m o u n t to the following. Abbreviate by P D L n the strengthening
of P D L by axioms of the form [a*]p +~ [a<n]p for all a. Then, by the finite
model property of P D L , P D L is the intersection of the logics P D L n. Un-
fortunately, it is not so that interpolation is preserved under intersection. A
counterexample is the logic G.3, which fails to have interpolation while all
proper extensions have interpolation, since they have all constants, by Theo-
rem 1.6.4. We have not been able to decide the question of interpolation for
P D L . But some answers can be given that point to the fact that P D L does
indeed have interpolation. Also, we wish to show that no significant fragment
of P D L has interpolation. The picture that emerges is this. If we start with
a polymodal language Kn, then interpolation obtains, because the language
524 10. D y n a m i c L o g i c
is not so strong. As soon as we add just one more operator, the star closure
of the basic programs, we can regain interpolation only if we add at least all
test-free programs of P DL. We believe that this latter fragment of P D L does
in fact have interpolation, and show moreover that if it does, P D L must as
well have interpolation.
Let us have the basic programs {0, ~1, . . . , r P u t 7 : - {0 U ~1 U
... U ~n_ 1. We will show first that if a fragment of P D L contains at least the
program 7", then it has interpolation only if it is closed under union, compo-
sition and star. This generalizes an observation of MAKSIMOVA in [152]. To
understand this result, let us call a f r a g m e n t of P D L a modal logic which
contains some subset of the programs definable from H0 plus the relevant ax-
ioms. There are various interesting fragments of P D L . One is the fragment
consisting of the basic programs and the star closure of the basic programs,
another is test-free P D L , where we close II0 only under union, composition
and star.
THEOREM 10.6.1. Let P D L - be a fragment of P D L containing at least

the star closure of the basic programs. Then it has interpolation only if it is
at least the fragment of test-free P D L .
PROOF. We show that it is possible to give an implicit definition of any

given regular language, so that P D L - can have the global B e t h - p r o p e r t y
only if it contains the closure of the basic programs under union, composition
and star. For the proof, let L be a regular language, definable by a regular
expression f~ over the basic modalities, ~i, i < m. There is a finite a u t o m a t o n
with m states recognizing L ~. As in the proof of Kleene's Theorem we can
describe the a u t o m a t o n recognizing L with a system of equations.
X0 = cU X0 9a00 U Zl " al0 U . . . [_JX n - 1 " an-l,o

Xl - Z 0 9 a01 U X 1 9 all U... U Xn_ 1 9 an-l,1
Xn-1 = Zo 9 ao,n-1 U X1 9al,n-1 U . . . U Xn-1 " an--l,n--1
If F is the set of accepting states, L - UieF Z i " Now, take a propositional let-
ter qi for each state i, and one more letter, q*. Let A(q, q*) be the conjunction
of the following set of formulae.
qo <-+ bo V <aoo>qo V <alo>ql V... V <an-l,o)qn-1

ql ++ bl V <aol)qo V <all)ql V... V <an-l,l)qn-i
q'n,--1 <--)' bn-1 V <ao,n_l>qo V <al,n_l>ql V . . . V <an_l,n-l>qn-1

Here, bi := q* if i E F and bi := • else. In addition, let B(~,q*) be the

conjunction of
q, A V q,, A A [qq,).
i~=j i<n j <m i<n
Put C(~, q*) := A(q, q*)A B(q, q*). The proof is complete if we prove the
following three things. (i) C(~, q*) is a global implicit definition of q0, (ii) An
explicit definition is q0 ++ (/3)q*, (iii) No explicit definition can be found if/3 is
not definable by a formula in P D L - . For (i), notice that since P D L has the
finite model property, so does P D L - . Now take a finite model for C(0", q*).
We show that the values of the qi are completely determined by the values of
q*. First of all, B(~, q*) is chosen so that if q* is true at a point, it remains
true throughout the transit of that point. Second, if there is a point x and a
one-step ~i-successor y satisfying qi, then all one-step successors satisfy qi.
Now we show the following, x ~ qi iff there exists a path w from x to y where
y ~ q*, and u~ C k[j, i] ~ for some accepting state j. First of all, observe that
for each x there exists a i such that x b qi, and this qi is unique. Now we
do induction of the smallest path z~ from x to a point y ~ q*. Suppose, z~
is of l e n g t h 0 . T h e n x = y a n d s o x ~ q* i f f y ~ q*, and e E L[j,j]~ for an
accepting state j. Now let z~ - ~k" z~' and assume x -~ x' --+ y for some
y ~ q*. Then x' ~ qs for some s. Then any ~k-SUCCessor of x satisfies qs,
and among them we choose the one through which the minimal path u~' goes.
By induction hypothesis there exists a path g to a point y' such that y' ~ q*
and ~Y C L[j, s] ~ for some accepting state j. Then ~k" g E L[i, s] ~, by the
k
fact that the a u t o m a t o n has a transition s --+ i. Hence, the qi are implicitly
defined, and equivalent to {/3i)q*, where/3 is the regular expression belonging
to U j e F k[i,j]~. The claim follows in the particular case of i = 0. Finally,
for (iii) notice that there exist the following models. Take any word w in the
alphabet A = {~i : i < n}. Let the frame consist of the prefixes of u~ and put
~Y-~ g' iff ~Y= g'. ~i. The frame codes nothing but u~ in reverse order. On this
frame, put/3(q*) := {e}, e the empty word. Then there is a unique valuation
/3+ making C(~, q*) globally true. At any point g in this frame, g ~ qi exactly
if g~ E k[j, i] for some accepting state j. Therefore, no simpler definition for
q0 can be given. []
On other hand, if test-free P D L has interpolation, then full P D L also

has interpolation. This is the content of the next theorem.
THEOREM 10.6.2. P D L has interpolation if test-free P D L has interpo-

lation.
PROOF. We use constructive reduction. The argument is therefore of a

more general character. First of all, we can restrict ourselves to finite H0, so
that we have weak transitivity. Then global reduction sets can be reduced to
local reduction sets. Take a set A and let
X?(A) " - {(r +4 .r A X 9 (r e FL(A)}.
It is enough to show that these sets are global reduction sets. For they split,
and therefore interpolation can be deduced for P D L as follows. Full P D L
is the logic which is obtained from test-free P D L with infinitely m a n y basic
programs by adding the test axioms for all programs [T?]. (This is to say,
from the basic modalities we select some to play the role of tests, and call
t h e m [~?]. For exactly those modalities, the test axioms are added.) Now
take a model ([,/3} for ~ in which the reduction formulae hold globally. We
can assume [ to be a dynamic frame. The only problem with t h a t frame is
that the t e s t - p r o g r a m s are interpreted freely. Let 0 differ from [ in t h a t x r y
iff x - y and <[,/~} ~ r for all r C FL(cp). We show that for all formulae
X C FL(~), and all x,
(J;) (f,/3, x) ~ X iff (~t,/3, x) ~ X
After t h a t we can actually drop the assignment of the test programs, and
obtain a full dynamic model for ~o. But now for the proof of (1:). Clearly, for
variables and boolean junctors there is nothing to show. So, let us take the
case of a formula (a>w. If a - / 3 tO-y, or a - ~; ~/or a - / 3 * , then we can also
use the induction hypothesis in a straightforward way. There remain the cases
a - (i and a - r The first is also straightforward since the interpretation of
the ~ has not changed. So let us proceed to the really critical case, X - (~b?}w.
Here, if ([,/3, x> ~ (r then also (~,/3, x> ~ ~b; w. By induction hypothesis,
<~t,13, x> ~ r w and so (~t,/3, x> ~ <r And conversely. V1
We will now present some particular cases of formulas where interpolants

can be found. In view of the preceding result it is enough to consider test-free
formulae, and this is what we will do now. The method is a rather explicit
construction of the interpolant, using a m e t h o d analogous to the one described
at the end of Chapter 3.8. We will illustrate it with the case where ~ ~ r and
is of dynamic complexity 1, that is, of the form V ~i, each ~i a conjunction
of fomulas of the form (/~>X, [/3]X, where X is nonmodal. In this case we must
have ~i ~- r for all i, so it is enough if interpolants can be produced for each
~i individually. Now recall the strategy of the Chapter 3.8. Let us define
~7- to be the result of replacing p by T where it occurs positively, and by _1_
where it occurs negatively. Then we have p ~- ~7- by construction. It remains
to be verified that ~7- ~_ r To have that it is enough to show that if we have
a model for ~oq-;~ b then we can also produce a model for ~;-~b. Since the
latter does not obtain, ~oq- k- r is proved. To make that work, ~o has to be
carefully prepared before the elimination of the variable p takes place. The
reason is that forgetting p can lead to a failure in the strategy, because ~oT
now allows for essentially more models than ~ itself. A case in point is when
we have subformulae of the form (>p A [3-~p, or [-]p A []-~p. In the first case
we have ((>p A [3~p)T _ (>T A [3T, which is equivalent to (>T. On the other
hand, the original formula is simply false, that is, deductively equivalent to
_L. It is the latter formula that must be chosen as an interpolant, and not <>T.
This problem does not arise with the formula (>p A (>~p. W h y is this so? The
reason is that in the first two cases we have a formula t h a t speaks over all
successors of a point. Hence, if another formula also speaks about successors,
then the valuation on the successors must be matched with the requirements
of the first formula. If a point accepts [3p, then all successors must satisfy p,
and so (>-~p cannot hold at that point. However, if we have formulas (>p A (>~p
then no conflict arises, because we can always arrange it that a point has two
different successors, one satisfying p, the other satisfying ~p.
Now return to the case where ~ is a conjunction of formulae of the form
(f~)X or [/9Ix, X nonmodal, f~ test-free. We want to rewrite ~ in a similar way
as we have done with polymodal formulae. However, this time m a t t e r s are
even more complex. For example, the programs (a2) + and (a3) +, although
different, give rise to subtle interactions. From [(a2)+]p and ( ( a 3 ) + ) ~ p we can
deduce that -~(a 6}T. Hence we must reckon beforehand with a new operator,
a 6. To care for this, we analyse the possible intersections of programs. Recall
that the regular languages over a finite alphabet are closed under intersection.
Therefore, let us take a second look at ~. Suppose that the regular expressions
occurring in ~ are/3i, i < n. Then for each subset S C_ n we let 7s be the
regular expression corresponding to
i6S i~s
7s exists by the results of Section 9.4. The following is then clear. The
languages corresponding to the 7s are mutually disjoint, and
~ - U 7s
i6S
We now change the 'program basis' in ~p by replacing talk of ~i by talk of
7s. Hence, we can assume that p is a conjunction of formulae of the form
(Ti>X, [Ti]X, X nonmodal, 7i test-free and mutually disjoint. Moreover, we
assume that for no i, e falls under 7i, that is, we assume that the programs
are proper. W i t h this given, we can compute the interpolant in the same way
as for polymodal K. In fact, let us make the reduction as follows. Each [ffi] is
regarded now as a primitive modality, with dual operator (7i). Then compute
the interpolant ~T as if working in polymodal K, letting (>i replace (7i). This
is possible for the following reason.
LEMMA 10.6.3. Let IIo be finite, and "yi, i < n, be regular test-free pro-
grams over IIo such that 5.) no path falls under both "7i and 7j, i r j, 5i.)
e does not fall under anyTj, J < n. Let ~ ~ Kn be a formula of degree 1,

and p(~) be the result of replacing each occurrence of Di by [7i], for all i < n.
Then ~ ~ Kn iff p(~) ~ P D L .
PROOF. Assume that ~ ~ Kn; then there is a finite Kripke-frame ~ and
and wo such that ([,f~,wo} ~ -~p. Since p is of degree 1, it is enough to
assume that f consists of the 1-transit of Wo. Now fix for every i < n a finite
sequence or(i) C_ H~) falling under 7i. Let or(i) have length t~(i). Now define
fP "-- {wo} U {(w,i,x} 9 ~- a prefix of cr(i),wo qi x}.
Further, let (e, i, x} := wo. Then
i, X) (7', i', x') iff { or Tx_- e,X"T'i--_ i',4 T' -- ~-~a
(Here, ~ denotes as usual an elementary program.) Fix for i < n a world y(i)
such that w0 <i y(i). This defines the dynamic Kripke-frame [P. Now
.- i, x ) . x c Z(p)}
u eZ(p)}
U {(7, i, X i ' T falls under 7i, T # o(i), y(i) e f~(p)}
This is well-defined since the i such that T falls under 7i is unique if it exists.
Moreover, e does not fall under any 7i. We claim that (fP, 7, wo) b p(~o). To
that end, we prove that (1.) for a formula K]i#, # nonmodal, ([P, 7, wo) ~ D~#
iff ([,/~,wo) ~ [7i]#; that (2.) for a nonmodal formula #, (~P,7, wo) ~ #
iff ([,3, wo) ~ #. (2.) is immediate from the definition of 7. For (1.) let
([P, 7, wo) ~ K]i#. Take a x such that wo <li x. Then (cr(i),i,x) C ~(#) and
so, by definition of 7, x E ~(#). Hence (~,/3, wo) ~ Di#. Assume ([,/~, wo}
Vli#. Take a point (v,i,x} such that wo 75 (%i,x}. Then ~- ~: e. C a s e 1.
7 - or(i). Then wo qi x and by definition of 7, (a,i,x} E ~(#). C a s e 2.
7 ?t a(i). Let ~- fall under 7j. Then (%i,x} C ~(#) iff (o(j), j, x) C ~(#) iff
y(j) C ~(#). By choice of y(j), wo qj y(j). Hence y(j) C f~(#) and therefore
(or(j), j, x) C 7(#) and this shows that (T, i, x) E 7(#). Hence, ([P, 7, wo}
and so r P D L . Conversely, assume that p(~o) r P D L . Then
there exists a finite dynamic Kripke-frame it and a model (9, 7, wo} ~ ~p(~o).
Put y z iff y - wo and wo -74 z. This defines ~. Let 3(p) "- 7(P). We
claim that (f, f~, wo} ~ -~o. To that end, observe that for a nonmodal formula
#, ([, 13, wo) ~ p iff (g, 7, wo) ~ p. Moreover, for a formula Z]~#, # nonmodal,
(~,/3, wo) ~ Vii# iff (g, 7, wo) ~ [7~]#. For wo q~ x iff wo -~ x. [:3
PROPOSITION 10.6.4. P D L has interpolants for ~ F r if ~ and r are

dynamic complexity ~ 1.
10.7. The Logic of Finite Computations 529
PROOF. It is enough to show this for formulae without test, by Theo-

rem 10.6.2. Let ~ ~- r and let ~ and r be of dynamic complexity < 1. Then
A
and r are locally equivalent to formulae ~ and r which are of dynamic

complexity ~ 1 and a boolean combination of nonmodal formulae and formu-
lae [7i]#, i < n, such that # is nonmodal; moreover, the 3'i do not contain
the empty path, and the path sets of q'i and 7j are disjoint for i r j. Then
- p(T) and r - p(v) for some 7, v e K,~. By the previous lemma T ~- V.
Both 7 and v are of degree _< 1. There exists an interpolant p in Kn of modal
degree _< 1. Then, again by the previous lemma, p(T) ~- p(p) and p(p) ~ p(v).
So, p(T) is an interpolant of ~ and r since var(p(p)) -- var(p); for the ~/i are
free of tests. []
Notice that the same argument does not work if we iterate the programs.
There are interactions between the 7i, but they are not noticeable in the first
iteration. For example, if we have a single program, a*, then taking this as a
primitive program is fine unless we study iterations, such as a * ; a * , which is
- - n a m e l y - the same as a*.
E x e r c i s e 374. Show that P D L ( w , w ) has interpolation if it holds for every

n that P D L ( w , n) has interpolation.
10.7. T h e Logic of Finite Computations

Finally, we want to study the logic of finite computations, which we call
P D L . f . It is the logic of all those structures in which no computation can
run forever. We have encountered in the monomodal case an axiom that
ensures such a property, the axiom G. Suppose that we have finitely many
basic programs, H0 := {~i : i < n}, then putting 3, := ~ i < n ~i we obtain
P D L . f ( w , n) by adding to P D L ( w , n) the axiom
p) --+ [7+]p.
If we have infinitely many basic programs, P D L . f is obtained by adding all
postulates of its weak fragments. Thus, P D L . f " - ~ , ~ P D L . f ( w , n ) . We
will prove two facts in this section. First, that P D L . f has the finite model
property, from which it follows by constructive reduction that D P D L . f has
the finite model property as well. Second, that D P D L . f does not have the
global Beth-property, from which the same follows for P D L . f . The argument
will be a constructive reduction to plain P D L . It does not use splitting reduc-
tion sets, and so the negative example on the Beth-property does not transfer
to P D L . It follows from the fact that P D L . f has the finite model property
that it is the logic of finite computations, whence the title of this section. It
also follows that D P D L . f is the logic of finite computations of deterministic
basic programs.
530 10. D y n a m i c L o g i c
Before we enter the first proof, let us get some intuition about P D L . f .
Clearly, if [7 +] satisfies the G - p o s t u l a t e , then the 7+-transitions in a K r i p k e -
frame must be without circles. Since 7 is the union of all ~i, it follows t h a t
the (i must be irreflexive. The converse does not hold, however. Now, take
a formula ~. Let ~[(~) be the collection of all atoms in the boolean algebra
generated by FL(~). We propose as reduction sets the set of all formulae
[7"]((7+)A -+ ('),+)(A A-~(7+)A))
where A C ~[(~). Now let ([,~,x} be a finite P D L - m o d e l for ~ and the

reduction formulas above. Define a new model in as follows. Call a point
x m a x i m a l if there exists an atom A such that x ~ A A -~(v+)A. Then
g is defined to be the set of all maximal points. Moreover, we put x -~ y
for two maximal points if there exists a y - in the original model such t h a t
x -~ y - ~ y and y - and y satisfy the same atom in ~[(~). This defines the
frame 9. This new frame contains no cycles. For each point has been chosen to
satisfy a formula A A--(7+)A in [, so each chain Y0 7_~ Yl 7_~ Y2... Yn-1 ?-~ yn
is without cycles.
Now ~ is based on a subset of f, so we take the valuation 13 restricted to
g, denoted also by/3. We show now
(t)
for all X E FL(cp) and y C g. The proof is by induction on the number
of V+-successors of x in 9. Assume x has no 7+-successors in 9. Then it
actually has no 7+-successors in [ either. Namely, if x has a successor in [,
y, then y satisfies an atom A that x does not satisfy (by choice of x) and has
a maximal successor y+ satisfying A, and so x ~-~ y+ in [ and so x ~-~ y+
in 9. Thus, the generated subframe of x in ~ is isomorphic to the generated
subframe generated in [; the models defined on them are as well. Hence, the
two models satisfy the same formulae. Now let us assume t h a t all successors
of x satisfy (~). Then (~) holds for x and X = P, P a variable. The induction
steps for A and ~ are straightforward. The only critical step is X = ( a ) r
Again, many cases are easy. I f a = / 3 U 7 , a = r a = /3;')' or a = f3*
then we can reduce the problem once more. The case a = (i C II0 remains.
From left to right let (~,13, x) ~ (~i)X- Then there is a (i-successor y of x
in ~ such that (9,/3, y) ~ X. By induction hypothesis, since y has less ")'+-
successors, ([,/3, y) ~ X. By construction, there exists a y - such t h a t in
x -~ y - ~ y, and y and y - satisfy the same atom. Hence ([,/3, y - ) ~ X
and so ([,/3, x) ~ (~i)X. This shows one direction. For the other, assume t h a t
([,/3, x) ~ ((i/X. Thus there exists a y such that x ~ y and ([,/3, y) ~ X. Either
y is already maximal for its atom A or else m by choice of the reduction sets
(f,/3, y) ~ (V+)(A A-~(7+)A). Hence we can find a y+ such that y Z~ y+

satisfying A and being maximal for A. By construction of g we have x -~ y+ in
g. By induction hypothesis, since y+ has less sucessors than x, (g,/3, y+) ~ A,
and so (g,/3, y+) ~ X. From this we deduce (g,/3, x ) ~ (r as desired.
THEOREM 10.7.1. The logic PDL.J' has the finite model property.
This proof method is basically the same as in the monomodal case for G,
and so the expectation is that it can be generalized. In the exercises, some of
these generalization are considered. One consequence is
COROLLARY 10.7.2. The logic D P D L . f has the finite model property.
The proof is by constructive reduction. Use the formulae as described in
the last section. The proof procedure works exactly in the same way. This
last result will be used in the remaining part, where we show that P D L . f
fails to have the Beth property. For notice that the reduction sets for D P D L
split, so that the failure of D P D L . f is passed down to P D L . f .
Now we proceed to the demonstration that the Beth-property fails for
the logics of finite computation. The formula we consider is the following. It
uses three elementary programs, r r and ~2, and V := r U r U r
(r A /
A(p, q ) ' - p. ++ .(q A [71/) V V (r A (~2)P
v A (r
LEMMA 10.7.3. A(p, q) is an implicit global definition of p in DPDL.f.
PROOF. We have to show that
A(p, q); A(r, q)I~ p ~ r
To see that, we have to check only the finite models of D P D L . f . Now, suppose
that we have a global model for A(p, q) rooted at x0. We show the uniqueness
of the valuation for p, given the valuation on q. Namely, the valuation of p can
be reconstructed by induction on the number of V+-successors as follows. At
points without V+-successors, p is true if and only if q is true. Now at a point x
with V+-successors, p is true iff it is true at exactly two immediate successors.
Since the immediate successors of x have less 7+-successors, this is actually
well-defined and unique. Therefore we have an implicit definition. D
THEOREM 10.7.4. The logic D P D L . f fails to have the global Beth-pro-

perty.
PROOF. We aim to show that there is no explicit definition for p in A(p, q).
In view of the results of Section 10.5 it is enough if we show that there exists
no evaluation procedure using finite memory which can tell us whether or not
p holds at a given point. For any formula B(q) can be evaluated with such
532 10. D y n a m i c L o g i c
a procedure. We work on special frames, which are ternary branching trees

of depth 5, coded as sequences of numbers 0, 1, 2 of length < 5. The root is
the empty sequence, e. Moreover, for sequences sl and s2 we put sl -~ s2 iff
s2 - sl .i. An evaluation procedure for a formula B(q) will start enumerating
the paths giving priority to 0 over 1 and to 1 over 2. It generates the path
(c, 0, 0 0 , . . . ) up to length 5. Let 5 - 3. Then after (e, 0, 00,000) the next
path is (e, 0, 00,001) and then (e, 0, 00,002) after which comes (e, 0, 01,010).
And so on. Recall that in addition B(q) calls subroutines which themselves
generate paths. The set of active nodes, however, is connected. Consider now
what happens if the main subroutine exits the point 0. Then it next goes
to (e, 1, 10,100) and starts working there. No subroutines will ever look into
the structure generated by 0. So the computation totally exits that structure.
Therefore, since the valuation of p at the root (= the truth value of B(q) at
e) depends on the value of p at 0, we need to remember this value. Thus we
need a memory of bit-size at least 1. Next consider the point 10. When the
computation exits 10 it needs to store the result that it computed for p at 10
to correctly compute the value of p at 1. Hence we need an additional bit of
memory, raising the size to 2. Likewise, for the points 110, 1110 and so on.
Since the size of the tree is not bounded a priori, we cannot say how large
the memory is that we need. To put it differently, if we have a memory of
# states, then the computation will give wrong results for trees whose depth
exceeds log 2 #. V-]
The argument actually has an information theoretic flavour. It does not
really m a t t e r whether the memory encodes exactly the fact that the compu-
tation yielded the value 'yes' or 'no' at the node 0. W h a t is important is
how many computation histories the memory can discriminate, if we have
a memory of size # then at most # histories can be discriminated. We have
shown, however, that in the case at hand no upper bound can be given on
the number of computation histories that must be distinguished. This proves
the theorem. We can note a number of consequences. First of all, A(p, q) is
of the form p ~ D(p, q) where p occurs inside a modal operator. Thus, we
have no analogue of the fixed-point theorems of Chapter 3.7 for P D L . f . The
strategy of this example is interesting in many respects. It has been shown by
HARVEY FRIEDMAN in [71] that an unbounded memory really increases the
power of algorithmic procedures. One example is the task of generating uni-
form two-branching trees, which cannot be achieved with bounded memory.
This has been used by JERZY TIURYN [213] to show that the expressive power
of programming languages is increased if an unbounded program memory is
allowed. The example that we have produced here uses only propositional
dynamic logic, not first-order dynamic logic, but the same result is obtained,
using a variable q. The property is true at a binary branching subtree whose
leaves are q cannot be checked using finite memory. However, as the proof
indicated, there is way to check this property using a memory stack of num-
bers from 0 to 2.
E x e r c i s e 375. Let P D L . f ~ be the extension of P D L . f by the axiom of

finite computations in one direction. Show that P D L . f ~ fails to have the
E x e r c i s e 376. Let P D L ~ . f + be the extension of P D L . f with converse oper-

ator, where both the programs V+ and Vv+ satisfy the G-axiom. Show that
this logic fails to have the finite model property.
E x e r c i s e 377. (Continuing the previous exercise.) Add to P D L ~ . f • the

axiom (7~)p -~ [V~]P. Show that this logic has the finite model property.
What finite frames does this logic describe?
E x e r c i s e 378. The example formula A(p,q) given above is not the most
economical one. Name an implicit definition of p that uses only two basic
modalities. This reduces the failure to the case of binary branching trees.
This seems to be the best possible result. Unary branching trees are just
strings, so here the argument must break down.
E x e r c i s e 379. Let P D L ~ . f be the logic of finite computations, where pro-

grams are allowed to be executed forwards and backwards. Show that this
logic is decidable. Hint. Do not take this too seriously.
Index
List o f S y m b o l s ded( A, ~o), 28

OF, F e , 34
~(S), 2 s, l~S, 3 =:>T, =:~~, ~ =:=>T~
* 38
f . A ,-+ B, f . A - + B, 3 p~(~), 42
f[S], f - l [ T ] , f o g, 3 ~ , 46
im[f], f I C, 3 ~P~, 47
a+ 4 I1~,~11, 47
S X , T X , Sx, Tx, 4 dp(~o), 48
D, LJ, 4 ~od, 49
s 5 +, @, 50
O, 1, - , n, u, 5 K]~ K](/) [~-<~ [] 52
c(x), 6 A s 53
T, _L, -~, A, V, --+, ++, 6 AF, sa
R oS, R*,R +,Rv 7, D<J>, D ~, [~, 54
A* ~ , e ~ 7
i~1 < h ~ 2 , ~ - - h ~ 2 , 5 4
I~1, r ~1 U[~2, ~1 o ~ 2 , 54
~, 8 IIi, 55
Tmo(X), 8 Th, hlg, 57
sf(~o), var(~o), 9
--A, 59
Clo(P2), Pol(ia), 10
~rA (vat), 59
~nb(g2), 92ut(92), 11
e, o, 60
ker(h), 11
[x]O, [D]@, 11 x J~ y, x Z+ y, 61
92/(9, 11 Th if)I, 63
O(E), 12 Frm(A), Krp(A), 63
Con(91), 12 f//~, 66
h, 13 @, 67
92<~,14 a<_~,68
*-+, --~, 14 a / ~ , 70
O I'A, 14 | 74
1-I~el P&, 15 9 , 74,
H, S, P, 15 [3, ~], 74
a~v(x), a~v(~), 16 A.t, 75
t-, 18 [~]k, 89
Taut(I-), 20 ~, 92
b +p, E(b), 21
T(b), 21 A ~ n, 95
E ~- 22 ~9r
(mp-+.), (mp.), 27 E A, Q A, 100
535
536 Index
NExt A, Ext A, 100 i~, f~, 207

I~-A, 103 ~m(~), 207
92/F, 105 E~, 210
l , A " , 107 T(C), g ( c ) , 211
e5 _~ 1:, 116 I~, H~, 211
Wave~(S), Try(S), 117 HiE1 ~i, 216
Cy(x), 117 B(~), 218
dp(x), 117 (r ~, L), 233
ur (t:, wo), 120 ", t, -~, A, -+, V, 3, e, 233
g(r), 120 ST(~, x), 234
~ $ p , 126 r - ~?, 234
qo~f, Asf , A sl, 126 (w~ >5 ~k), (w~ >~ ~k), ( 3 ~ >j ~k),
SCTK,~, 126 (3wi ~ Wk), 235
K~(h), 19,9 ~e, ~ f , 235
X ( A ) , X(~o), 133 Sf, 235
ffK, 146 er r 239
~atx , 149 r ~-~ ~, 240
X a X c X i 151 ( a x i o m . ) , 241
qoT(p) , V~T, 153 (exp.), (per.), (ren.), (swap.), (cnt.),
~ ' 157 (iter.), 242
CRel(A), 157 (A-I.), (V-I.), (0-I.), 243
TSp(A), 159 Seq, 243
I-IF, I-Iu, Up, 172 (~-I.), (r 244
ho, F o , OF, 173 ([:]-I.), 248
(C), 173 (A-I2.), 249
Z2, 174 S e q + , 249
ffiRa([), 174 ~t, 256
Eqx(V), Eq(V), 175 sq-rank(~), 256
~-v, 177 sq-rank, Gqn , Sqn , 257
Mal, 179 ue(f), 265
Th(I'), Eq(A), 180 RP(a), 269
Cp(P2), 189 C(a), 272
Set, 190 pZ, fI, 276
(_)op, 191 T~, ~-I, 278
9A., 192 | AE], All, 278
h o m g ( - , - ) , 192 r r ~ 279
BA, 192 ~a,-~ (~), ~ar D (~), va,~" (~,), 282
X* 194 ~o[:], ~o", 282
P2o, 195 $~0, ~o$, adp[:](~o), adpl(~o), adp(~o), 282
a(X), 195
dp[:](V), dp'(v), 283
StoneSp, 196
c(~), co(~), c . ( ~ ) , 28a
F -~ G, 199
~(~), ~D(~), r,.(~), 283
~ , [~, ,~, ~ , 201
x Sf~' A , (x), Xm (x), 284
Bal, 201
~o, 202 Atg A, 292
~J(a), ~, 203 V[2(~; r 297
StSpa~, 204 V , . , o , . , 301
92+, 205 if3s b s 301
~+, 205 w, a,/3, 302
C5, ~f, 2~i, 9~, ~rp, ~mp, ~ , ~, 206 ~3~ 5C~ 304
Mal, DFrm, 206 F~,~, (~ ~, FZ]f~, ~ Z, FO.,, ~ ,o, 306
Index 537
~s, 306 e(3, ~), eo (~, c2) ' 424

StSim, 307 A/F 3, 420
9~8, 307 A/z 3, 421
Dr2,308 S K4, C9-K4, 421
A
(~, ff~., 308 pt, p$, 421
As, 309 k < k -< 427
pO, p., p., 314 OKz(A), 436
Xs, 314 U(f), U2(f), 436
r r r 318 Ogz(A), 436
r 319 ~ , A~, tt~, 446
| ~, 323 ind (x ) , 446
~-O, ~-", 326 Tp(0, T, ~r), 457
a, 331 F~(~), S~q(f), 457
-<, 334 U(E), 458
x/y, 335 con, c ~ , coCm,p, 464
A, 5, 342 ~ ~, 465
5(~), 342 Ha(q), j x s, 473
5(f), 343 D(f), C(g), 476
pt(n), ~, ~pr 348 A c, O d, 477
Y(X, <_), (~(X, <), 351 ~c, 478
I~(~), ~::(n), ~pr 3~2 M 9~, 479
~(~), 3~3 c~, 483
DLat, CohLoc, 354 Ep, A~, 483
~ @ ~ , 357 a;/3, a U/3, a*, ~?, 498
sp(A), 372 [a], (a), 499
ot
in (p), p~ (p), p~ (p), 375 x -+ y, 500
PM, ~[JM, 375 skip, fail, 504
qM, L~M, 376 EL(X), 505
fd, 379 L(R), 509
~_~ ~, 380 L(P2), 510
~ -A~o ~ + I QM, 381 Zv, 511
<~, <~, 397 L p 512
f<~, f~, f>~, 398
L s 512
k[i, j], 512
i~@O, 399
dc(~), 514
g[O/f], 400
ch(~), 515
[-1+ ~, 401
~[(~), 530
~,, 402
Pn, 403 List of Logics
X(S), ~ - span, mspan, 404
Q - width, 405 altn, t r s n , 72
mwidth, ~ - span, mspan, 406 D P D L , 504, 516, 520, 522
mo~(x), 411 D P D L . f , 531
yU, ~, 411 D P D L v, 523
t s , Ss, :l:s, Ss, 4~2 E~, 50
crit(x), N-span, 412 E P D L , 499, 505, 523
~(~), 3(~), 412 G, 112,113, 116,125,134, 135,137, 143-
mspan, mwidth, 413 145, 153, 155, 156, 231, 232, 367,
real(3,3), err(3, ~3), 418 402, 410, 415, 422, 423, 428, 429,
")'(3, ~3), 419 489, 490, 529
"7~ 420 Ga, 387, 389, 392
538 Index
GL, 72 K4.ft(_< n), 402, 456

G.3, 130, 161, 366, 440, 454, 489, 523 K4.G, 72
G.3.t.K4.3.D-, 372 K4.I2,449
C-:~'.~"~2,358, 359 K4.In, 401,402, 422,430, 434-436,443
G | K, 490 K4.Jn, 401, 402
Grz, 113, 134, 136, 137, 143, 153, 155, K4.ti(_< n), 401,402,456
156, 365, 366, 402, 415, 422, 423, K4.t, 385, 386, 393
428, 429 K4.wd(_< n), 401,402
Grz.3, 112, 297, 300, 340, 408 K4.wd(_< n).ti(<_ m), 462
Grz.3 | K, 297 K4n, 422
Grz3, 361 K5, 72, 113, 332
Here, 300 M,~, 50
Inc(n), 325 PDL, 497, 504, 507, 508, 515, 525, 528
K, 157, 163, 321,374, 378, 383, 490, 496 PDL.f, 529, 531
K2 (D [~p <-> rap, 281 P D L . f V, 533
Kn, 527, 528 PDL(w,w), 503, 529
K~ @ [:]Sp _+ [:]tp, 114 PDL(m, n), 503, 505, 508
K~, 50, 85, 106, 109, 120, 122, 149, 152, PDL(m, w), 503
155, 368-370 PDL(w, n), 503, 529
K.1, 72, 250 P D L n, 508
K.2, 72, 250, 258 P D L - , 524
K.3, 72 P D L ~, 503, 507, 508
K.B.T, 143
S4, 72, 143, 150, 155, 336,340, 410, 415,
K.B, 72, 134, 143
422, 428, 433, 438, 450, 451, 454,
K.D, 72, 112, 113, 336, 339
456,496
K.G, 72
S4.2.t, 393
K.Grz, 72
S4.3, 73, 113, 332, 356, 367, 397, 423,
K.t, 74, 132, 245, 371, 385, 393
439, 449, 490
K.T, 72, 134, 143, 153, 244, 290
S4.3.t, 393
K.altl, 72, 89, 109, 113, 134, 143, 153,
S4.Grz, 72
236-239,245,290,332, 334,354,463,
$4.I2,444, 444-448
465-475,483, 484
S4./2. ti (< 3), 449
K.altl.D, 464-472, 483, 485
K.altl 9 Vt3• 290 S4.In, 449,454
K.alt3.B.T, 347 S4.In.ti( <_m), 438
K.altn, 72, 363 S4. ti(_< 1), 438
K.altn.trsm, 363 S4.ti(<_ 2), 449
K.altl | K.altl, 472 $4. wd (< 2), 444-448, 454
K.altl | K.altl (B <~p <-+ 0p, 476-481 S4.wd(<_ 3), 454
K.alt 1.D | K.alt 1.D, 472 S4.t, 74, 367, 385, 386, 391
K.alt2, 477-481 $5,72, 109, 113, 130, 143, 208, 290, 336,
K.trs2, 73 340, 408, 415,428
K.trsn, 72, 74, 101, 189, 363 S5.t, 74, 393
K4, 72, 113,131,143,153, 156,231,232, S5 | $5,290
290,367,410, 415,419-422,427, 428, Sire, 305-310, 315, 320-322
433,434, 450, 455,489 Sim 9 Dp ++ rap, 322
K4.1, 422, 423, 430 Sim(n), 325
K4.2,422, 423 ThF--], 73, 101, 160, 301, 309, 315, 321,
K4.3, x, 366, 402, 422,439, 440, 444 336, 376, 378, 384, 422, 428, 475,
K4.D, 415 486,489,490, 496
K4.dp(_< n), 401,402,411 Th m +.Th 393
Index 539
Th[-~'], 73, 101, 160, 334, 364, 384, 422, Gleit, Zachary, 144
428, 486, 490, 496 GSdel, Kurt, 72
Th ~ ] + . T h [~-] - , 391 Goldblatt, Robert I., viii, 231, 260, 261,
Th[e ~o l 336, 347
265,266, 504
_
Goldfarb, Warren, 112, 116, 144
-hl o -~ r.T l ~ o 336 391 Goranko, Valentin, 77, 107, 108
Th f~o+l, 128 Gor~, Rajeev, 155
Th (Q, <), 337 Grefe, Carsten, x, 469, 474, 479, 480,
Th (]~,<), 337 484, 488
Groenendijk, Jeroen, 499
List o f N a m e s
Grzegorczyk, Andrzej, 72
Alekseev, Alexander, 444 Halld~n, SSren, 29
Amerbauer, Martin, 155 Halpern, Joseph, x, 290, 516, 520, 522,
Baker, K. A., 363 523
Balbiani, Philippe, 135 Harrop, A., 26, 80
Bellissima, Fabio, x, 115, 410, 463 Herrmann, Burghard, 22, 180
Ben-Ari, Mordechai, x, 516, 520, 522 Herzig, Andreas, 135
van Benthem, Johan, viii, 210, 230, 231, Hintikka, Jaakko, 155
246,250,260, 266,268,272,372,476 Humberstone, L. L., 230
Beth, E. W., 139 Isard, Stephen, 81
Birkhoff, Garreth, 167, 169, 176, 177 Jankov, V. A., 331
Blok, Wim, ix, 25, 102, 123, 180, 183, Johnson, David S., 38
189, 331, 335, 347, 357, 364, 368- de Jongh, Dick, vii, 75, 143
370, 374, 376, 378, 383, 384, 410, J6nsson, Bjarni, 64, 172, 207, 265
455 Kanger, Stig, 64
Boolos, George, 72 KShler, Peter, 183, 189
Bull, Robert, 397, 423, 439 Kozen, Dexter, x, 505, 506
Carnap, Rudolf, 64 Kowalski, Tomasz, 186, 189
Chagrov, Alexander, ix, x, 130, 359,489 Kleene, Steven C., 39, 509, 510
Chagrova, Lilia, x, 488,489 Kracht, Marcus, viii, 81, 102, 109, 130,
Chellas, Brian, 163,468 231,272,275,346, 359,410, 423
Church, Alonzo, 39 Kripke, Saul, 50, 64
Cook, S. A., 45 Kruskal, J. B., 356
Czelakowski, Janusz, 25 Ladner, R. E., 109, 150, 290, 505, 506,
van Eijck, Jan, 499 508, 523
Emerson, E. A., 522 Leivant, Daniel, 523
Esakia, Leo, 247, 248, 408, 410 Lewis, Clarence Irwing, 72
Feferman, Solomon, viii, 266 Lindenbaum, 90
Fine, Kit, viii, ix, 82, 124, 125, 129, 261- LSb, M. H., 72
265, 275, 290, 300, 364, 365, 372, Los, J., 30
397, 410, 414, 415, 424, 427, 428, Lukasiewicz, Jan, 25
434, 435, 439, 450, 476 Lyndon, Robert C., 180
Fischer, M. J., 505, 506, 508, 523 Maehara, S., 155
Fitting, Melvin, 155 Makinson, David, 101,257
Frege, Gottlob, 27 Maksimova, Larisa, viii, 139, 144, 146,
Friedman, Harvey, 162, 532 225-229, 407, 408, 410, 524
Gabbay, Dov, 254, 258 Malcev, A., 170
Garey, Michael A., 38 Markov, A. A., 482
Geach, Peter, 73 McKinsey, 73
Ghilardi, Silvio, 155 Meskhi, V., 408, 410
di Giacomo, Giuseppe, 508 Meyer, A. R., 45
540 Index
Mortimer, M., 257 Vardi, Moshe, 508

Moses, Y., 290 Venema, Yde, 265
Muchnik, Albert A., 463 Visser, Albert, 155
Parikh, Rohit, x, 505, 506 de Vries, Fer-Jan, 499
Passy, Solomon, 107, 108, 504 Williamson, Timothy, 163
Penther, Brigitte, 499 W6jcicki, Ryszard, 24-26, 30, 277
Pigozzi, Don, 25, 123, 180, 183, 189 Wolper, P., 508
Pitts, Andrew, 155 Wolter, Frank, viii, ix, 22, 109, 111, 113,
Pnueli, Amir, x, 516, 520, 522 124-128,155,156,213, 275,277,290-
Pogorzelski, W. A., 26 293, 346, 366, 393, 422,429
Post, Emil, 21,482 Wrofiski, Andrzej, 181
Pratt, Vaughan, 499, 508 Zakharyaschev, Michael, ix, x, 79, 130,
Prucnal, T., 181 232, 359, 397, 411, 414, 415, 419,
Rabin, Michael O., 482 424, 444, 489
Rautenberg, Wolfgang, ix, 102, 146, 174, Zawadowski, M., 155
180, 290, 331,339,346, 385,408,410
Reidhaar-Olson, Lisa, 144 List of Terms
de Rijke, Maarten, 109
Rybakov, Vladimir, 162 accessibility matrix, 457
Sahlqvist, Hendrik, viii, 231, 232, 249, accessibility relation, 60
257 associated, 60
Sambin, Giovanni, vii, viii, 72, 75, 143, adjunction
144, 201, 204, 210, 211, 221, 231, counit, 199
232 unit, 199
Savitch, Walter, 41 algebra, 10
Schumm, George, 450 absolutely free, 16
Schurz, Gerhard, viii, 275 boolean, 5, 32
Segerberg, Krister, x, 163, 407, 410, 449, congruence distributive, 170
463, 465, 501 cycle-free, 347
Shehtman, Valentin, 81 directly irreducible, 169
Smoryfiski, Craig, vii, 144 directly reducible, 169
Solovay, Robert, 72 effective, 79
Spaan, Edith, 81, 109, 290,449,468,475 finitely presentable, 341,342
Stockmeyer, L. J., 45 freely A-generated, 16
Stokhof, Martin, 499 functionally complete, 28
Stone, Marshall H., 167, 197 hereditarily simple, 190
Surendonk, Timothy, 225 isomorphic, 14
Suszko, Roman, 25, 30 A-generated, 16
Tarski, Alfred, 25, 90, 207 modal, 56
o
Teichmiiller, Oswald, 36 fi-,-,.,, 10
Thomason, S. K., ix, 64, 265, 275, 279, polynomially complete, 28
301,364, 372, 385, 490 prime, 341
Tinchev, Tinko, 504 realization, 194
Tiuryn, Jerzy, 532 semisimple, 184
Tokarz, M., 21 simple, 12
Tukey, J. W., 36 subdirectly irreducible, 167
Turing, Alan, 39 alternation depth, 282
Urquhart, Alasdair, 80 amalgamation, 226
Vaccaro, Virginia, viii, 201,211,204,221, antichain, 401
231,232 inner, 459
Vakarelov, Dimiter, x, 508 outer, 459
Index 541
antiframe, 402 locally small, 192

arrow, 190 opposite, 191
ascending chain condition, 398 poset, 200
assignment, 23 C-equivalence, 286
atom, 5, 404 chain, 4, 100, 514
automaton properly ascending, 100, 401
deterministic, 510 regular, 514
finite, 509 semiregular, 514
automorphism, 11 character, 300
avoiding all configurations, 433 finite, 36
axiom, 20 minimal, 300
axiomatizability class
n-, 95 A-, E-, E/X-elementary, 260
finite, 77 elementary, 260
independent, 278 modally definable, 214
recursive, 77 clone, 10
strongly recursive, 77 closed domain, 418
closure
barrier, 424 prefix, 512
base calculus, 243 suffix, 512
basis, 31, 195, 354 closure operator, 6
strong, 354 clot, 409
Beth-property cluster, 117, 398
global, 139 degenerate, 398
local, 139 final, 398
bidual, 218 slim, 432
block, 160, 461 cluster sequence, 425
blowing up, 118 co-cone, 220
book, 386 co-covering number, 376
boolean algebra, 5, 32 co-hemimorphism, 201
atomless, 291 co-limit, 220
expanded, 55 co-splitting, 336
coatom, 5
calculus codimension, 101
complete, 243 codomain, 190
Hilbert-style, 27 colouring, 476
o
m p - ~ , 27 compactness, 110
sound, 243 global, 111
canonical altl-formula, 469 strong, 110
canonical formula, 419 weak, 110
canonical frame, 92 complement, 33
o
a - ~ , 93 relative, 33
weak, 95 completeness, 110
canonical model, 92 global, 104
local, 92 intrinsic, 372
canonicity, 110 local, 104
weak, 110 strict, 372
cardinal number, 4 compound modality, 53
category, 190 normal, 53
dual, 191 compression map, 323
equivalent, 197 computability, 38
542 Index
1-step, n + 1-step, 38 correspondence, 114

deterministic, 40 casewise, 237
computation, 38 simple, 237
halting, 38 cosets, 11
nondeterministic, 38 cover, 334
computation trace, 515 lower, 334
conclusion, 20 modal, 310
condition upper, 334
invariant, 266 qo-covered, 431
preserved, 266 Craig Interpolation Property (CIP), 138
reflected, 266 cycle, 117, 429
cone, 219, 235
lower, 4 d-persistence, 221
upper, 4 decidability, 41
congruence, 11 global, 104
equationally definable principal, 182 local, 104
filtral, 171 decidable
fully invariant, 1TT o set, 332
matrix, 24 decolouring, 476
o
permuting ~s, 169 deduction theorem, 27
principal, 169 deductively closed set, 22
congruence extension property (CEP), 34T definability, 403
o
conjunct, 83 k-H, 403
conjunction, 28 defines, 240
consequence relation, 19 definition
compact, 19 implicit, 139
finitary, 19 degree, 48
o
fusion of ~s, 326 o of incompleteness, 372
global, 103 depth, 11T
o
modal, 156 [:]-~, m - ~ , 283
modal classical, 157 o of a frame, 398, 400
modal monotone, 15T alternation ~, o 282
Post-complete, 21 local, 436
o
reduct of a ~, 326 molecular, 412
structurally complete, 21 derivability, 243
consistency, 19 descendant, 426
mS-, w - ~ with, 343 description
weak o with, 343 internal, 240
consistency formula, 283 diagonal, 7, 12
consistency set, 283 diagram, 190, 342
constant, 9 commuting, 190
constructive reducibility, 134 dimension, 101
continuous function, 195 directed system, 96
contraction, 65, 69, 204 discriminator term, 183
minimal, 399 disentangling, 426
weak, 204 disjoint union, 67, 70
contractum, 65 domain, 190
converse, 7 closed, 418
coproduct, 214, 230 DPOF, 182
reduced, 309 dropping, 431
Index 543
safe, 431 Fischer-Ladner closure, 505

supersafe, 431 fixed point, 143
dynamic complexity, 514 formula
dynamic logic (PDL), 497 antitone, 246
axiomatically equivalent, 468
EDPC, 182 ~'-boxed, 144
element canonical a l t l - ~ 469
deep, 185 characteristic, 284
dense, 185 clash-free, 85
embedding, 68 clean, 252
embedding number, 44 constant, 257, 272
embedding pattern, 424 o
depth of a ~, 48
cofinal, 424
V-distributive, 246
endomorphism, 11
A-distributive, 246
equivalence
dual, 49
axiomatic, 468
elementary Sahlqvist, 253
black, 310
explicit, 84
deductive, 48
external, 233
local, 48
internal, 233
white, 310
monotone, 246
equivalential term
o negative, 246
set of ~s, 181
negative Sahlqvist, 253
ersatz, 282
plain, 107
essential, 210
positive, 246
EXPTIME, 41
Sahlqvist, 250
extender set, 21
Sahlqvist-van Benthem, 251
qp-extract, 414
simple, 82
factor algebra, 11 simulation transparent, 311
family slotted, 240
o standard, 82
limit of a ~, 247
strict diamond, 516
upward directed, 247
strictly simple, 82
fan, 410
strongly negative, 246
fatness, 402
fatness vector, 457 strongly positive, 246
fibre, 11 thinner, 312
filter, 5, 34, 336 unleashed, 84
definable principal open, 182 white based, 311
improper, 34 frame, 62
open, 105 o for a logic, 63
principal, 5, 336 a-canonical, 93
trivial, 34 almost slim, 432
,0
filtration, 120 antl~, 402
finite antichain property (fap), 432 atomic, 206
finite character, 36 canonical, 221
finite cover property (fcp), 432 compact, 206
finite embedding property, 427 cyclic, 117
finite intersection property, 36 descriptive, 206
finite model property (fmp), 79 differentiated, 206
galactic, 436 effective, 441
global, 104 full, 206
544 Index
hadronic, 447 global, 141

hooked, 290 local, 141
o
A - ~ , 63 halting string, 38
local, 309 head, 365
meager, 387 hemimorphism, 56
modally saturated, 262 Hilbert-style calculus, 27
noetherian, 398 hom-functor, 192
one-generated, 117 homomorphic image, 14
pointed, 62 homomorphism, 10
refined, 206 hypergalaxy, 437
rooted, 117 ~_o, 438
separable, 290, 434
simple, 441 ideal, 5, 336
slim, 432 principal, 5, 336
o
standard simulation ~, 307 immediate variant, 427
supereffective, 443 in conjunction with, 84
o
tensor product of ~s, 323 incompleteness
o
theory of a ~, 63 degree of o, 372
tight, 206 index, 446, 473
top-heavy, 403 infix notation, 8
o inner antichain, 459
ultraproduct of ~s, 259
instruction, 38
unindexed, 323
interpolation, 28
frame sequence, 457
global, 138
rooted, 457
local, 138
function
characteristic, 193 uniform, 153
interpretation, 276
clopen, 202
interval, 436
closed, 202
irreducibility
coherent, 354
open, 202 join, 333
o meet, 333
reduction ~, 133
strict join o , 333
functor o
adjoint, 199 strict meet ~, 333

contravariant, 191 isomorphism, 14
covariant, 191
join compactness, 5
forgetful, 200
free, 200 kernel, 11
powerset, 194
kite, 410
recovery, 207 Kripke-frame, 60
fusion, 74, 229, 278 dynamic, 499
o of consequence relations, 326
P D L - ~~ , 501
galactic regular expression, 461 pointed, 61
galactically regular class, 461 Kripke-model, 61
galaxy, 436 PDL - ~ 501
galaxy sequence, 458 Kuznetsov-index, 129
generated subframe, 64, 68 ordinal, 436
generating set, 354
group, 7 language
regular, 509
Halld6n-completeness, 29 weak, 8
Index 545
lattice, 4 compact, 110

algebraic, 5 complete, 110
bounded, 5 complex, 110
complete, 5 connected, 75
continuous, 97 cyclic, 74
distributive, 5 d-persistent, 221
lower continuous, 97 dense, 449
upper continuous, 97 elementary, 240
limit, 220 globally compact, 111
locale, 97 independently axiomatizable, 278, 359
coherent, 353 intrinsically complete, 372
spatial, 349 locally finite, 222
localization, 126 monotone, 48
logic, 19 normal, 48
consistent, 19 operator transitive, 74
decidable, 26 pre-complete, 450
equivalential, 181 pre-finitely axiomatizable, 359
essentially 1-axiomatizable, 340 quasi-normal, 50
finitely equivalential, 181 Sahlqvist, 250
inconsistent, 19 strongly compact, 110
o
rule base of a ~, 20 subframe, 125
tabular, 79
M-system, 225 weak, 94
map weakly canonical, 110
isotonic, 200 weakly compact, 110
n-localic, 120 weakly operator transitive, 74
master modality, 74 weakly transitive, 74
matrix, 23 modality
o compound, 53
f~-~, 23
master, 74
reduced, 25
universal, 74
modal algebra, 56
model
complete, 207 o
e - ~ , 233
effective, 78
o algebraic, 57
fusion of ~s, 327 canonical, 92
local, 309 direct, 86
o
presentation of a ~, 342 geometric, 63
modal consequence relation, 156 global, 103
modal cover, 310 9 O
K n p k e - ~ , 61
modal formula local, 103
elementary, 240 local canonical, 92
modal logic, 48 local extension, 103
o of bounded alternativity, 74 saturated, 261
o of bounded operator alternativity, modus ponens, 27
74 molecular depth, 412
C-complete, 76 molecule, 411
C-computable, 76 monoid, 7
C-hard, 76 monolith, 168
c-persistent, 110 morphism, 190
canonical, 110 mp-calculus, 27
classical, 48 multiplication, 119
546 Index
natural transformation, 198 precluster, 117

boolean, 204 prefix, 7
needle, 365 prefix notation, 8
net, 66, 70 premiss, 20
N E X P T I M E , 41 presentation, 341
normal form, 86 preservation, 310
NP, 40 primeness, 333
join ~,o 333
object, 190 meet o , 333
occurrence strictly join o , 333
black, 311
strictly meet o , 333
negative, 250
problem, 41
positive, 250
C-complete, 41
string, 38
C-hard, 41
white, 310
trivial, 41
operator
process
classical, 48
Semi-Thue, 38
dual, 47
T h u e - ~~ , 482
monotone, 48
product, 15, 214
normal, 48
subdirect, 167
operator compactness, 330
projection, 214
opremum, 174
proof tree, 20
ordinal number, 3
property
outer antichain, 459
bounded, 450
P, 40 essential, 222
p-formula, 282 essentially bounded, 450
o
p-morphism, 65, 69 transfer of a ~, 324
admissible, 66 propositional language, 8
lateral, 399 pseudomodel, 516
p-variable, 282 proper, 516
packed representation, 45 pseudo relevance property, 141
page, 386 PSPACE, 41
partial order, 4 pullback, 219
partition, 373 pushout, 219
path, 120, 515 quasi-maximal point, 414
o
n-H, 121 quotient, 335
path equation, 471 prime, 335
persistence, 240
c-, d-, df-, g-, r-, ti - ~ 241 rank, 257
point, 192, 348 pure, 257
bad, 380 special, 257
eliminable, 434 reconstruction, 282
good, 380 total, 282
improper, 492 ~-reduct, 414
maximal, 137, 411,530 reduction
~-maximal, 137, 358 o set, 133
quasi-maximal, 414 o function, 133
Polish Notation, 8 reduction set, 133
polynomial, 10 splitting, 142
postfix, 7 refinement map, 209
Index 547
refutation pattern, 418 semihomogeneous, 287

omission, 418 sf-founded, 284
realization, 418 simple, 436
regular signature, 8
o language, 509 computable, 43
o expression, 509 simulation, 276, 304
relation atomic, 277
closed, 202 simulation transparency, 311
continuous, 202 situation, 62
point closed, 203 slice, 398
relation composition, 7 slotted formula, 240
relativization, 124 span, 337, 404
replacement, 177 minimal, 337, 404
request, 491 molecular, 412
o
residuation, 32 ~ - ~ , 491
restrictor, 235 spectrum, 372
9 O
root, 117 F m e - ~ , 372
rooted frame sequence, 457 prime, 374
rule, 20 T - ~ 159
admissible, 21, 162 spine, 365, 446
derived, 20 splitting, 336
instance, 20 o companion, 336
n-ary, 20 9 O
F m e - ~ , 420
proper, 20
Zakharyaschev - ~ 421
rule base, 20
spone, 247
Sahlqvist Hierarchy, 256 standard simulation frame, 307
Sahlqvist-van Benthem formula, 251 standard translation, 234
satisfiability, 20 state
saturated, 261 accepting, 509
downward, 148 initial, 509
section, 388 Stone space, 196
partial, 388 strict diamond formula, 516
selection history, 493 string, 7
semigroup, 7 length, 7
set occurrence, 43
characteristic, 284 string handling machine, 38
(co-)recursively enumerable, 76 deterministic, 40
clopen, 195 subalgebra, 14
decidable, 332,485 skew-free, 213
dense, 211 subformula, 9
downward saturated, 148 subframe, 116
external, 62 o logic, 125
generating, 354 local, 118
homogeneous, 287 subframe logic, 125
independent, 225, 359 sublattice, 335
internal, 62 subreduction, 333
measure zero, 211 substitution, 13
open, 64, 195 substring, 7
recursive, 76 o
occurrence, 38
reduction o , 133 successor
548 Index
o
strong, 397 m - ~ , 74
weak, 397 weakly ~, o 74
superamalgamation, 226 translation, 10
superfusion, 229 transpose, 511
surrogate, 282 triangular identities, 200
symbol count, 42 trivial constants, 88
truth value, 23
T-spectrum, 159 designated, 23
tableau, 147 Tukey's Lemma, 36
closing, 147 type, 457
good, 149 type regular, 458
tack, 408
tautology, 20 ultrafilter, 34
tense logic, 74 ultrafilter extension, 265
tensor product, 327 ultraproduct, 172, 259
term, 8 underlying set, 10
discriminator, 183 unit, 7
equivalential, 181 unital semantics, 25
term-function, 10 universal modality, 74
termalgebra, 10 unravelling, 120
theory, 22, 63 unsimulation, 308, 309, 314
consistent, 25
valuation, 23, 61, 62, 499
equational, 177 o
maximally consistent, 25 domain of a ~, 284

natural, 92
Thomason
partial, 61,283
~ Trick, 490
variable, 9
thorn, 518
inherently existential, 252
Thue-process, 482
inherently universal, 252
decidable, 482
variant, 427, 502
trivial, 482
immediate, 427
tightness, 401 o
topological space, 195 quasi-closed under ~s, 427

o
T0-space, 197, 349 quasi-closed under nonfinal ~s, 430

T2-space, 197 variety, 15
TD-space, 352 o with constructible free algebras, 79
basis, 195 congruence distributive, 170
compact, 195 congruence permutable, 170
discrete, 195 discriminator ~, o 183
Hausdorff, 197 locally finite, 222
Sierpirlski, 197 semisimple, 184
sober, 349 view, 416
soberification, 353
wave, 117
zero-dimensional, 195
weakening, 147
topology
weight, 42
Alexandrov - ~ 351
well-partial order (wpo), 356
weak, 351
width
trace algebra, 116
molecular, 413
transit, 117, 211
world, 60, 90
transition, 61,510
transition function, 509 X-string, 8
transitivity
Bibliography
[1] M a r t i n A m e r b a u e r . C u t - f r e e t a b l e a u calculi for s o m e p r o p o s i t i o n a l m o d a l logics. Stu-
dia Logica, 57:359 - 371, 1996.
[2] K. A. Baker. Finite e q u a t i o n a l bases for finite algebras in a c o n g r u e n c e - d i s t r i b u t i v e
e q u a t i o n a l class. Advances in Mathematics, 24:207 - 243, 1977.
[3] Philippe Balbiani a n d A n d r e a s Herzig. A t r a n s l a t i o n from t h e m o d a l logic of prov-
ability into K 4 . Journal of Applied Non-Classical Logics, 4(1):73 - 78, 1994.
[4] Fabio Bellissima. A n effective r e p r e s e n t a t i o n for finitely g e n e r a t e d free interior alge-
bras. Algebra Universalis, 20:302 - 317, 1985.
[5] Fabio Bellissima. O n t h e lattice of e x t e n s i o n s of t h e m o d a l logics K.Altn. Archive for
Mathematical Logic, 2 7 : 1 0 7 - 114, 1988.
[6] Fabio Bellissima. P o s t c o m p l e t e a n d 0 - a x i o m a t i z a b l e m o d a l logics. Annals of Pure
and Applied Logic, 47, 1990.
[7] M. Ben-Ari, J. I. Halpern, a n d A. Pnueli. D e t e r m i n i s t i c p r o p o s i t i o n a l d y n a m i c logic;
finite models, c o m p l e x i t y a n d completeness. Journal of Computer and System Sci-
ences, 25:402 - 417, 1982.
[8] J o h a n van B e n t h e m . M o d a l r e d u c t i o n principles. Journal of Symbolic Logic, 41:301
- 312, 1976.
[9] J o h a n v a n B e n t h e m . S y n t a c t i c a s p e c t s of m o d a l i n c o m p l e t e n e s s t h e o r e m s . Theoria,
45:63 - 77, 1979.
[10] J o h a n v a n B e n t h e m . Modal and Classical Logic. Bibliopolis, 1983.
[11] J o h a n v a n B e n t h e m . C o r r e s p o n d e n c e theory. In Dov M. G a b b a y a n d Franz G u e n -
t h n e r , editors, Handbook of Philosophical Logic, v o l u m e 2, p a g e s 167 - 247. Reidel,
1984.
[12] J o h a n v a n B e n t h e m . N o t e s on M o d a l Definability. Notre Dame Journal of Formal
Logic, 39:20 - 39, 1989.
[13] J o h a n v a n B e n t h e m . B e y o n d Accessibility: F u n c t i o n a l Models for M o d a l Logics. In
M a a r t e n de Rijke, editor, Diamonds and Defaults, p a g e s 1 - 18. Kluwer, Dordrecht,
1993.
[14] J o h a n v a n B e n t h e m a n d W i m Blok. T r a n s i t i v i t y follows f r o m D u m m e t t ' s axiom.
Theoria, 4 4 : 1 1 7 - 118, 1978.
[15] J o h a n v a n B e n t h e m a n d I. L. H u m b e r s t o n e . H a l l d 6 n - c o m p l e t e n e s s by gluing K r i p k e -
frames. Notre Dame Journal of Formal Logic, 24:426 - 430, 1983.
[16] E. W. B e t h . O n p a d o a ' s m e t h o d in t h e t h e o r y of definition. Indagationes Mathemat-
icae, 15:330 - 339, 1953.
[17] G a r r e t h Birkhoff. Subdirect p r o d u c t s in universal algebra. Bulletin of the American
Mathematical Society, 50:767 - 768, 1944.
[18] G a r r e t h Birkhoff. Lattice Theory. A m e r i c a n M a t h e m a t i c a l Society C o l l o q u i u m P u b -
lications, R h o d e Island, 1973.
549
550 Bibliography
[19] Patrick Blackburn, Wilfried Meyer-Viol, and M a a r t e n de Rijke. A P r o o f S y s t e m for

Finite Trees. In H. Kleine Brining, editor, Computer Science Logic '95, n u m b e r 1092
in Lecture Notes in C o m p u t e r Science, pages 86 - 105. Springer, 1996.
[20] W i m BIok. P r e t a b u l a r varieties of modal aIgebras. Studia Logica, 39:101 - 124, 1880.
[21] W i m Blok. Varieties of interior algebras. P h D thesis, University of A m s t e r d a m , 1976.
[22] W i m Blok. T h e free closure algebra on finitely many generators. Indagationes Math-
ematicae, 39:362 - 379, 1977.
[23] W i m Blok. The lattice of modal algebras is not strongly atomic. Algebra Universalis,
11:285 - 294, 1980.
[24] W i m Blok. The lattice of modal logics: A n algebraic investigation. Journal of Sym-
bolic Logic, 45:221 - 236, 1980.
[25] W i m Blok. O n the degree of incompleteness in m o d a l logics and the covering relation
in the lattice of modal logics. Technical R e p o r t 78 - 07, D e p a r t m e n t of M a t h e m a t i c s ,
University of A m s t e r d a m , 1980.
[26] W i m Blok, P e t e r KShler, and Don Pigozzi. On the structure of varieties with equa-
tionally definable principal congruences, II. Algebra Universalis, 18:334 - 379, 1984.
[27] W i m Blok, P e t e r KShler, and Don Pigozzi. On the structure of varieties w i t h equa-
tionally definable principal congruences, IV. Algebra Universalis, 31:1 - 35, 1994.
[28] W i m Blok and Don Pigozzi. On the structure of varieties with equationally definable
principal congruences, I. Algebra Universalis, 15:195 - 227, 1982.
[29] W i m Blok and Don Pigozzi. Algebraizable Logics. N u m b e r 396 in M e m o i r s of the
A m e r i c a n M a t h e m a t i c a l Society. 1989.
[30] W i m Blok and Don Pigozzi. O n the structure of varieties with equationally definable
principal congruences III. Algebra Universalis, 32:545 -608, 1994.
[31] George Boolos. The Logic of Provability. Cambridge University Press, Cambridge,
1994.
[32] George Boolos and Giovanni Sambin. Provability: The Emergence of a M a t h e m a t i c a l
Modality. Studia Logica, 50:1 - 23, 1991.
[33] Manfred Borzechowski. Tableau-KalkriI frir P D L und Interpolation. M a s t e r ' s thesis,
D e p a r t m e n t of M a t h e m a t i c s , F U Berlin, 1988.
[34] R o b e r t A. Bull. T h a t all normal extensions of $4.3 have the finite model property.
Zeitschrift fii mathematische Logik und Grundlagen der Mathematik, 12:341 - 344,
1966.
[35] R o b e r t A. Bull and Krister Segerberg. Basic m o d a l logic. In Dov M. G a b b a y and
Franz Guenthner, editors, Handbook of Philosophical Logic, volume 2, pages 1 - 88.
Reidel, 1984.
[36] J o h n P. Burgess. Basic tense logic. In Dov M. G a b b a y and Franz G u e n t h n e r , editors,
Handbook of Philosophical Logic, volume 2, pages 89 - 133. Reidel, 1984.
[37] Stanley Burris and Sankappanavar H. P. A Course in Universal Algebra. N u m b e r 78
in G r a d u a t e Texts in Mathematics. Springer, 1981.
[38] Rudolf Carnap. Meaning and Necessity. T h e University of Chicago Press, Chicago,
1947.
[39] Alexander Chagrov. T h e undecidability of the tabularity problem for m o d a l logic. In
Logic Colloquium '9~, page 34, 1994.
[40] Alexander Chagrov and Michael Zakharyaschev. T h e Disjunction P r o p e r t y of Inter-
m e d i a t e Propositional Logics. Studia Logica, 50:189 - 216, 1991.
[41] Alexander Chagrov and Michael Zakharyaschev. T h e undecidability of the disjunction
p r o p e r t y and other related problems. Journal of Symbolic Logic, 58:967 - 1002, 1993.
[42] Alexander Chagrov and Michael Zakharyaschev. O n the I n d e p e n d e n t AxiomatizabiI-
ity of Modal and I n t e r m e d i a t e Logics. Journal of Logic and Computation, 5:287 -
302, 1995.
Bibliography 551
[43] Alexander Chagrov and Michael Zakharyaschev. Modal Logic. Oxford University
Press, Oxford, 1997.
[44] Lilia A. Chagrova. Undecidable problems in correspondency theory. Journal of Sym-
bolic Logic, 56:1261 - 1272, 1991.
[45] C. C. Chang and H. Jerome Keisler. Model Theory. N o r t h - H o l l a n d , A m s t e r d a m , 3
edition, 1990.
[46] Brian F. Chellas and Krister Segerberg. Modal Logics with t h e M a c i n t o s h Rule.
Journal of Philosophical Logic, 23:67 - 86, 1994.
[47] S. A. Cook. The complexity of t h e o r e m proving procedures. In Proceedings of the 3rd
Annual ACM Symposium on Theory of Computing, pages 151 - 158, New York, 1971.
Association of the C o m p u t i n g Machinery.
[48] Max Creswell. A n incomplete decidable logic. Journal of Symbolic Logic, 49:520 -
527, 1984.
[49] Janusz Czelakowski. Equivalential logics (I). Studia Logica, 40:227 - 236, 1981.
[50] Janusz Czelakowski. Equivalential logics (II). Studia Logica, 40:335 - 372, 1981.
[51] Marcello d'Agostino. Are Tableaux an Improvement on T r u t h - T a b l e s ? C u t - f r e e
Proofs and Bivalence. Journal of Logic, Language and Information, 1:235 - 252,
1993.
[52] B. A. Davey and H. A. Priestley. Lattices and Order. Cambridge University Press,
Cambridge, 1990.
[53] Dick de J o n g h and Albert Visser. Explicit Fixed Points in Interpretability Logic.
Studia Logica, 50:39 - 49, 1991.
[54] E. A. Emerson. T e m p o r a l and Modal Logic. In J a n van Leeuwen, editor, Handbook
of Theoretical Computer Science, Vol. B, Formal Models and Semantics, pages 996
- 1072. Elsevier, A m s t e r d a m , 1990.
[55] N. M. Ermolaeva and Albert A. Muchnik. Modal logics defined by e n d o m o r p h i s m s of
distributive lattices (Russian). In A. D. Bochvar and Grishin, editors, Issledovaniya
po Teorii Mnozhestv i Neklassicheskim Logikam. Sbornik Trudov. (Investigations on
Set-Theory and Non-Classical Logics), pages 229 - 246. 1976.
[56] N. M. Ermolaeva and Albert A. Muchnik. P r e t a b u l a r t e m p o r a l logic (Russian). In
A. I. Mikhailov, editor, Issledovaniya po Neklassicheskim Logikam i Teorii Mnozhestv
(Investigations on Non-Classical Logics and Set-Theory), pages 288 - 297. 1979.
[57] L. Esakia and V. Meskhi. Five critical systems. Theoria, 40:52 - 60, 1977.
[58] Leo Esakia. Topological Kripke Models (Russian). Soviet Mathematical Doklady,
1 5 : 1 4 7 - 151, 1974.
[59] Solomon Feferman. Persistent and invariant formulas for outer extensions. Compositio
Mathematica, 20:29 - 52, 1969.
[60] Kit Fine. T h e logics containing $4.3. Zeitschrift fiir Mathematische Logik und Grund-
lagen der Mathematik, 17:371 - 376, 1971.
[61] Kit Fine. A n ascending chain of $4 logics. Theoria, 40:110 - 116, 1974.
[62] Kit Fine. A n incomplete logic containing $4. Theoria, 40:23 - 29, 1974.
[63] Kit Fine. Logics containing K4, P a r t I. Journal of Symbolic Logic, 39:229 - 237, 1974.
[64] Kit Fine. Normal forms in modal logic. Notre Dame Journal of Formal Logic, 16:31
- 42, 1975.
[65] Kit Fine. Some connections between elementary and m o d a l logic. In Stig Kanger, ed-
itor, Proceedings of the Third Scandinavian Logic Symposium, pages 15 - 31. N o r t h -
Holland, 1975.
[66] Kit Fine. Logics containing K4, P a r t II. Journal of Symbolic Logic, 50:619 - 651,
1985.
[67] Kit Fine and G e r h a r d Schurz. Transfer theorems for stratified m o d a l logics. In Pro-
ceedings of the Arthur Prior Memorial Conference, Christchurch, New Zealand, 1991.
552 Bibliography
f68] Marcello Finger and Dov M. Gabbay. Adding a T e m p o r a l Dimension to a Logic

System. Journal of Logic, Language and Information, 1:203 - 233, 1993.
[69] M. J. Fischer and R. E. Ladner. Propositional d y n a m i c logic of regular programs.
Journal of Computer and System Sciences, 18:194 - 211, 1979.
[70] Melvin Fitting. Proof Methods for Modal and Intuitionistic Logic. N u m b e r 169 in
Synthese Library. Reidei, Dordrecht, 1983.
[71] Harvey Friedman. Algorithmic procedures, generalized Turing algorithms, and ele-
m e n t a r y recursion theory. In Logic Colloquium 1969, pages 361 - 390, A m s t e r d a m ,
1971. North-Holland.
[72] Dov M. Gabbay. Expressive Functional Completeness in Tense Logics. In Uwe
MSnnich, editor, Aspects of Philosophical Logic, pages 91 - 117. Reidel, Dordrecht,
1981.
[73] Michael R. Garey and David S. Johnson. Computers and Intractability. A Guide to
the Theory of NP-Completeness. W. H. F r e e m a n and Company, New York, 1979.
[74] Silvio Ghilardi and M. Zawadowski. Undefinability of propositional quantifiers in
m o d a l system s4. Studia Logica, 55:259 - 271, 1995.
[75] Giuseppe di Giacomo. Eliminating "Converse" from Converse PDL. Journal of Logic,
Language and Information, 5:193 - 208, 1996.
[761 Zachary Gleit and Warren Goldfarb. C h a r a c t e r s and fixed points in provability logic.
Notre Dame Journal of Formal Logic, 31:26 - 26, 1990.
[77] R o b e r t I. Goldblatt. F i r s t - o r d e r definability in modal logic. Journal of Symbolic Logic,
40:35 - 40, 1975.
[78] R o b e r t I. Goldblatt. M e t a m a t h e m a t i c s of modal logics, P a r t I. Reports on Mathe-
matical Logic, 6:41 - 78, 1976.
[79] R o b e r t I. Goldblatt. M e t a m a t h e m a t i c s of m o d a l logics, P a r t II. Reports on Mathe-
matical Logic, 7:21 - 52, 1976.
[80] R o b e r t I. Goldblatt. Topoi - - The Categorial Analysis of Logic. N u m b e r 98 in Studies
in Logic and the Foundations of M a t h e m a t i c s . North-Holland, A m s t e r d a m , 1984.
[81] R o b e r t I. Goldblatt. Logics of Time and Computation. N u m b e r 7 in CSLI Lecture
Notes. CSLI, 1987.
[82] R o b e r t I. Goldblatt. Varieties of compIex algebras. Annals of Pure and Applied Logic,
38:173 - 241, 1989.
[83] R o b e r t I. Goldblatt and S. K. T h o m a s o n . Axiomatic classes in propositional m o d a l
logic. In J. Crossley, editor, AIebra and Logic, n u m b e r 450 in Springer Lecture Notes
in M a t h e m a t i c s . Springer, 1974.
[84] M a r t i n Goldstern and Haim Judah. The Incompleteness Phenomenon. A K Peters,
Wellesley, Massachussetts, 1995.
[85] Valentin Goranko. Refutation Systems in Modal Logic. Studia Logica, 53:299 - 324,
1994.
[86] Valentin Goranko and George Gargov. Modal logic with names. Journal of Philo-
sophical Logic, 22:607 - 636, 1993.
[87] Valentin Goranko and Solomon Passy. Using the universal modality: Gains and Ques-
tions. Journal of Logic and Computation, 2:5 - 30, 1992.
[88] Rajeev Gor6. C u t - f r e e sequent and tableau systems for propositional Diodorean
modal logics. Studia Logica, 53:433 - 457, 1994.
[89] George Grgtzer. Universal Algebra. van Nostrand, New York, 1968.
[90] George Gr/itzer. Lattice Theory: First Concepts and Distributive Lattices. Freeman,
1971.
[91] C a r s t e n Grefe. Modale Logiken funktionaler Frames. Master's thesis, D e p a r t m e n t of
M a t h e m a t i c s , F U Berlin, 1994.
Bibliography 553
[92] Jeroen Groenendijk and Martin Stokhof. Dynamic Predicate Logic. Linguistics and
Philosophy, 14:39 - 100, 1991.
[93] S5ren Halld@n. On the semantic non-completeness of certain Lewis calculi. Journal
of Symbolic Logic, 16:127 - 129, 1951.
[94] Paul Halmos. Algebraic logic I. Monadic boolean algebras. Compositio Mathematica,
12:217 - 249, 1955.
[95] Joseph Halpern and Y. Moses. A guide to the modal logics of knowledge and belief.
Artificial Intelligence, 54:319 - 379, 1992.
[96] Bengt Hansson and Peter G~irdenfors. Filtrations and the finite frame property in
boolean semantics, in Stig Kanger, editor, Proceedings of the Third Scandinavian
Logic Symposium, pages 32 - 39. North-Holland, 1975.
[97] D. Harel. Dynamic logic. In Dov M. Gabbay and Franz Guenthner, editors, Handbook
of Philosophical Logic, volume 2. Reidel, 1984.
[98] David Harel, Amir Pnueli, and Jonathan Stavi. Propositional dynamic logic of non-
regular programs. Journal of Computer and Systems Sciences, 26:222 - 243, 1983.
[99] Michael A. Harrison. Introduction to Formal Language Theory. Addison Wesley,
Reading (Mass.), 1978.
[100] R. Harrop. On the existence of finite models and decision procedures for propositional
calculi. Proceedings of the Cambridge Philosophical Society, 54:1 - 13, 1958.
[101] Horst Herrlich and George E. Strecker. Category Theory. Allyn and Bacon, 1973.
[102] Burghard Herrmann and Wolfgang Rautenberg. Finite replacement and finite Hilbert-
style axiomatizability. Zeitschrift fiir mathematische Logik und Grundlagen der Math-
ematik, 38:327 - 344, 1992.
[103] Burghard Herrmann and Frank Wolter. Representations of algebraic lattices. Algebra
Universalis, 31:612 - 613, 1994.
[104] Jaakko Hintikkm Knowledge and Belief. An Introduction into the logic of the two
notions. Cornell University Press, Ithaca, 1962.
[105] John E. Hopcroft and Jeffrey D. Ullman. Formal Languages and their Relation to
Automata. Addison Wesley, Reading (Mass.), 1969.
[106] George E. Hughes and Max J. Cresswell. A Companion to Modal Logic. Methuen,
London, 1984.
[107] Stephen Isard. A finitely axiomatizable undecidable extension of K. Theoria, 43:195
- 2O2, 1977.
[108] V. A. Jankov. The construction of a sequence of strongly independent superintuition-
istic propositional calculi. Soviet Mathematics, 9:806 - 807, 1968.
[109] V. A. Jankov. Conjunctively indecomposable formulas in propositional calculi.
Izvestija Akad. Nauk. SSSR, 33:18 - 33, 1969.
[110] Peter T. Johnstone. Stone Spaces. Number 3 in Cambridge studies in advanced math-
ematics. Cambridge University Press, 1982.
[111] Bjarni JSnsson. Algebras whose lattice of congruences is distributive. Mathematica
Seandinaviea, 21:110- 121, 1967.
[112] Bjarni JSnsson. On the canonicity of Sahlqvist identities. Studia Logica, 53:473 - 491,
1994.
[113] Bjarni JSnsson and Alfred Tarski. Boolean algebras with operators. American Journal
of Mathematics, 73:891 - 939, 1951.
[114] Winfried Just and Martin Weese. Discovering Modern Set Theory. I. The Basics,
volume 8 of Graduate Studies in Mathematics. AMS, 1996.
[115] Winfried Just and Martin Weese. Discovering Modern Set Theory. H. Set-Theoretic
Tools for Every Mathematician, volume 18 of Graduate Studies in Mathematics.
AMS, 1997.
554 Bibliography
[116] Steven C. Kleene. Introduction to Metamathematics. N o r t h - H o l l a n d , A m s t e r d a m ,

1964.
[117] Sabine Koppelberg. General theory of boolean algebras. In J. Monk and S. Koppel-
berg, editors, Handbook of Boolean Algebras. I. Elsevier, 1988.
[118] Dexter Kozen and Rohit Parikh. A n elementary proof of the completeness of PDL.
Theoretical Computer Science, 14:113 - 118, 1981.
[119] Dexter Kozen and Jerzy Tiuryn. Logics of Programs. In J a n van Leeuwen, editor,
Handbook of Theoretical Computer Science, Vol. B, Formal Models and Semantics,
pages 789 - 840. Elsevier, A m s t e r d a m , 1990.
[120] Marcus Kracht. An almost general splitting t h e o r e m for modal logic. Studia Logica,
49:455 - 470, 1990.
[121] Marcus Kracht. Internal Definability and Completeness in Modal Logic. P h D thesis,
F U Berlin, 1991.
[122] Marcus Kracbt. A solution to a problem by Urquhart. Journal of Philosophical Logic,
20:285 - 286, 1991.
[123] Marcus Kracht. Even more about the lattice of tense logics. Archive of Mathematical
Logic, 31:243 - 357, 1992.
[124] Marcus Kracht. How completeness and correspondence theory got married. In
M a a r t e n de Rijke, editor, Diamonds and Defaults, Synthese, pages 175 - 214. Kluwer,
1993.
[125] Marcus Kracht. Prefinitely Axiomatizable Modal and I n t e r m e d i a t e Logics. Mathe-
matical Logic Quarterly, 39:301 - 322, 1993.
[126] Marcus Kracht. Splittings and the finite model property. Journal of Symbolic Logic,
58:139 - 157, 1993.
[127] Marcus Kracht. Highway to the danger zone. Journal of Logic and Computation, 5:93
- 109, 1995.
[128] Marcus Kracht. Power and Weakness of the Modal Display Calculus. In Heinrich
Wansing, editor, Proof Theory of Modal Logic, pages 95 - 120. Oxford University
Press, 1996.
[129] Marcus Kracht. Modal Logics t h a t Need Very Large Frames. m a n u s c r i p t , 1998.
[130] Marcus Kracht. Reducing Modal Consequence Relations. manuscript, 1998.
[131] Marcus Kracht. Lattices of Modal Logics and Their Groups of A u t o m o r p h i s m s . Jour-
nal of Pure and Applied Logic, 1999. to appear.
[132] Marcus Kracht and Frank Wolter. Properties of independently axiomatizable b i m o d a l
logics. Journal of Symbolic Logic, 56:1469 - 1485, 1991.
[133] Marcus Kracht and Frank Wolter. Normal m o n o m o d a l logics can simulate all others.
Journal of Symbolic Logic, 64, 1999.
[134] Saul Kripke. Semantic analysis of modal logic, P a r t I. Zeitschrift fiir mathematische
Logik und Grundlagen der Mathematik, 9:67 - 96, 1963.
[135] J. B. Kruskal. WeIl-quasi-orderings, the Tree T h e o r e m and Vazsonyi's Conjecture.
Transactions of the American Mathematical Society, 95:210 - 225, 1960.
[136] A. H. Lachlan. A note on T h o m a s o n ' s refined structures for tense logics. Theoria,
4 0 : 1 1 7 - 120, 1974.
[137] R. E. Ladner. T h e c o m p u t a t i o n a l complexity of provability in systems of m o d a l logic.
SIAM Journal of Computing, 6:467 - 480, 1977.
[138] Daniel Leivant. P r o o f theoretic methodology for propositional d y n a m i c logic. In For-
realization of Programming Concepts, Proceedings of the International Colloquium,
Peniscola, number 107 in Lecture Notes in C o m p u t e r Science, pages 356 - 373.
Springer, 1981.
[139] E. J. Lemon. A note on Halld@n-incompleteness. Notre Dame Journal of Formal
Logic, 7:296 - 300, 1966.
Bibliography 555
[140] J. Los a n d R o m a n Suszko. R e m a r k s on sententical logics. Indagationes Mathematicae,

2 0 : 1 7 7 - 183, 1958.
[141] J. Lukasiewicz a n d Alfred Tarski. U n t e r s u c h u n g e n fiber d e n aussagenkalk/il. Comptes
Rendues des Sdances de la Societ@ des Sciences et des Lettres de Varsovie, 23:30 -
50, 1930.
[142] R o b e r t C. L y n d o n . Identities in 2 - v a l u e d calculi. Transactions of the Americal Math-
ematical Society, 71:457 - 465, 1951.
[143] S a u n d e r s M a c L a n e . Categories for the Working Mathematician. N u m b e r 5 in G r a d -
u a t e T e x t s in M a t h e m a t i c s . Springer, Heidelberg, 1971.
[144] S. M a e h a r a . O n t h e i n t e r p o l a t i o n t h e o r e m of Craig (japanese). Sugaku, 12:235 - 237,
1960.
[145] D a v i d M a k i n s o n . A n o r m a l m o d a l calculus b e t w e e n T a n d $4 w i t h o u t t h e finite m o d e l
property. Journal of Symbolic Logic, 34:35 - 38, 1969.
[146] D a v i d C. M a k i n s o n . Some e m b e d d i n g t h e o r e m s for m o d a l logic. Notre Dame Journal
of Formal Logic, 12:252 - 254, 1971.
[147] L a r i s a L. M a k s i m o v a . P r e t a b u l a r e x t e n s i o n s of Lewis $4. Algebra i Logica, 14:28 - 55,
1975.
[148] Larisa L. M a k s i m o v a . I n t e r p o l a t i o n t h e o r e m s in m o d a l logic a n d a n d a m a l g a m a t e d
varieties of t o p o b o o l e a n algebras. Algebra and Logic, 1979.
[149] Larisa L. M a k s i m o v a . Definability t h e o r e m s in n o r m a l e x t e n s i o n s of t h e provability
logic. Studia Logica, 48:495 - 507, 1989.
[150] Larisa L. M a k s i m o v a . T h e B e t h Properties, interpolation, a n d a m a l g a m a b i l i t y in
varieties of m o d a l algebras. Soviet Mathematical Doklady, 44:327 - 331, 1992.
[151] L a r i s a L. M a k s i m o v a . Definability a n d i n t e r p o l a t i o n in classical m o d a l logics. Con-
temporary Mathematics, 131:583 - 599, 1992.
[152] Larisa L. M a k s i m o v a . T e m p o r a l logics with "the n e x t " o p e r a t o r do n o t have Interpo-
lation or t h e B e t h Property. Siberian Mathematical Journal, 32:989 - 993, 1992.
[153] Larisa L. M a k s i m o v a . O n variable s e p a r a t i o n in m o d a l a n d s u p e r i n t u i t i o n i s t i c logics.
Studia Logica, 55:99 - 112, 1995.
[154] Larisa L. M a k s l m o v a a n d V l a d i m i r V. Rybakov. L a t t i c e of n o r m a l m o d a l logics. Al-
gebra i Logica, 13:188 - 216, 1974.
[155] A. I. Malcev. O n t h e general t h e o r y of algebraic s y s t e m s . Math. Sbornik, 35:3 - 20,
1954.
[156] A. A. Markov. I m p o s s i b i l i t y of C e r t a i n A l g o r i t h m s in t h e T h e o r y of Associative Sys-
t e m s ( R u s s i a n ) . Doklady Akademia Nauk SSSR, 77:953 - 956, 1951.
[157] M a a r t e n M a r x , L£szl6 Pdlos, a n d Michael M a s u c h . Arrow Logic and Multi-Modal
Logic. Studies in Logic, L a n g u a g e a n d I n f o r m a t i o n . CSLI, S t a n f o r d , 1996.
[158] M. M o r t i m e r . O n l a n g u a g e s with two variables. Zeitschrift fiir mathematische Logik
und Grundlagen der Mathematik, 21:135 - 140, 1975.
[159] S o l o m o n P a s s y a n d T i n k o Tinchev. A n essay in c o m b i n a t o r y d y n a m i c logic. Infor-
mation and Computation, 93:263 - 332, 1991.
[160] Brigitte P e n t h e r . A D y n a m i c Logic of Action. Journal of Logic, Language and Infor-
mation, 3:169 - 210, 1994.
[161] A. P i t t s . O n a n i n t e r p r e t a t i o n of second order quantification in first order intuitionistic
logic. Journal of Symbolic Logic, 57:33 - 52, 1992.
[162] W. A. Pogorzelski. S t r u c t u r a l c o m p l e t e n e s s of t h e p r o p o s i t i o n a l calculus. Bulletin
de l'Acaddmie Polonaise des Sciences, Sdries des sciences mathdmatiques, as-
tronomiques et physiques, 19:439 - 451, 1971.
[163] Emil L. Post. Recursive Undecidability of a P r o b l e m of T h u e . Journal of Symbolic
Logic, 12:1 - 11, 1947.
556 Bibliography
[164] Vaughan Pratt. Models of program logics. In Proceedings of the 20th IEEE Symposium
on Foundations of Computer Science, pages 115 - 122, 1979.
[165] Vaughan Pratt. Dynamic Algebras: Examples, Constructions, Applications. Studia
Logica, 50:571 - 605, 1991.
[166] T. Prucnal and A. Wrofiski. An algebraic characterization of the notion of structural
completeness. Bulletin of Section Logic of the Polish Academy of Sciences, 3:20 - 33,
1974.
[167] Michael O. Rabin. Recursive Unsolvability of Group Theoretic Problems. Annals of
Mathematics, 67:172 - 194, 1958.
[168] Wolfgang Rautenberg. Der Verband der normalen verzweigten Modallogiken. Mathe-
matische Zeitschrift, 156:123 - 140, 1977.
[169] Wolfgang Rautenberg. Klassisehe und nichtklassische Aussagenlogik. Vieweg, Braun-
schweig/Wiesbaden, 1979.
[170] Wolfgang Rautenberg. Splitting lattices of logics. Archly fiir Mathematisehe Logik,
20:155 - 159, 1980.
[171] Wolfgang Rautenberg. Modal tableau calculi and interpolation. Journal of Philosoph-
ical Logic, 12:403 - 423, 1983.
[172] Wolfgang Rautenberg. Applications of Weak Kripke Semantics to Intermediate Con-
sequence Relations. Studia Logica, 45:119 - 134, 1986.
[173] Wolfgang Rautenberg. Axiomatizing Logics Closely Related to Varieties. Studia Log-
ica, 50:607 - 622, 1991.
[174] Wolfgang Rautenberg. Strongly finitely based equational theories. Algebra Univer-
salis, 28:549 - 558, 1991.
[175] Lisa Reidhaar-Olson. A new proof of the fixed-point theorem of provability logic.
Notre Dame Journal of Formal Logic, 31:37 -43, 1990.
[176] GySrgy E. R@v~sz. Introduction to Formal Languages. McGraw-Hill, New York, 1985.
[177] Maarten de Rijke. Extending Modal Logic. PhD thesis, Universiteit van Amsterdam,
1993.
[178] Vladimir V. Rybakov. Admissible rules of pretabular modal logics. Algebra and Logic,
20:291 - 307, 1981.
[179] Vladimir V. Rybakov. Problems of substitution and admissibility in the modal sys-
tems grz and intuitionistic calculus. Annals of Pure and Applied Logic, 50:71 - 106,
1990.
[180] Vladimir V. Rybakov. Rules of inference with parameters for intuitionistic logic. Jour-
nal of Symbolic Logic, 57:912 - 923, 1992.
[181] Vladimir V. Rybakov. Criteria for Admissibility of Inference Rules. Modal and Inter-
mediate Logics with the Branching Property. Studia Logica, 53:203 - 226, 1994.
[182] Vladimir V. Rybakov. Admissibility of Inference Rules. Number 136 in Studies in
Logic. Elsevier, Amsterdam, 1997.
[183] Hendrik Sahlqvist. First and second order semantics for modal logic. In Stig Kanger,
editor, Proceedings of the Third Scandinavian Logic Symposium, pages 15 - 3 1 . North-
Holland, 1975.
[184] Giovanni Sambin. An effective fixed-point theorem in intuitionistic diagonalizable
algebras. Studia Logica, 35:345 - 361, 1976.
[185] Giovanni Sambin. Subdirectly irreducible modal algebras and initial frames. Studia
Logica, 62, 1999.
[186] Giovanni Sambin and Virginia Vaccaro. Topology and duality in modal logic. Annals
of Pure and Applied Logic, 37:249 - 296, 1988.
[187] Giovanni Sambin and Virginia Vaccaro. A new proof of Sahlqvist's theorem on modal
definability and completeness. Journal of Symbolic Logic, 54:992 - 999, 1989.
Bibliography 557
[188] Giovanni Sambin and S. Valentini. The modal logic of provability, the sequential
approach. Journal of Philosophical Logic, 11:311 - 342, 1982.
[189] Walter J. Savitch. Relationship between nondeterministic and deterministic t a p e com-
plexities. Journal of Computer and Systems Sciences, 4:177 - 192, 1970.
[190] George Schumm. B o u n d e d properties in modal logic. Zeitschrift fiir mathematische
Logik und GrundIagen der Mathematik, 27:197 - 200, 1981.
[191] George F. Schumm. W h y does Halld~n-completeness m a t t e r ? Theoria, 59:192 - 206,
1993.
[192] Dana Scott. Advice on Modal Logic. In Lambert, editor, Philosophical Problems in
Logic. Some Recent Developments, pages 143 - 174. Reidel, Dordrecht, 1970.
[193] Krister Segerberg. A n essay in classical modal logic. Mimeograph, Uppsala, 1971.
[194] Krister Segerberg. T h a t all extensions of $4.3 are normal. In Stig Kanger, editor,
Proceedings of the Third Scandinavian Logic Symposium, pages 194 - 196. North-
Holland, 1975.
[195] Krister Segerberg. A completeness t h e o r e m in the m o d a l logic of programs. Notices
of the AMS, 24(6):A - 522, 1977.
[196] Krister Segerberg. Classical Propositional Operators. N u m b e r 5 in Oxford Logic
Guides. Oxford University Press, 1982.
[197] Krister Segerberg. Modal logics with functional alternative relations. Notre Dame
Journal of Formal Logic, 27:504 - 522, 1986.
[198] Valentin Shehtman. Undecidable propositional calculi (russian). In Problemy kiber-
netiki. Neklassicheskije logiki i ikh primenenije. Akademia Nauk SSSR, 1982.
[199] Joseph A. Shoenfiled. Mathematical Logic. Addison Wesley, Reading, Massachussetts,
1967.
[200] Craig Smoryfiski. B e t h ' s T h e o r e m and Self-referential sentences. In Logic Colloquium
77, pages 253 - 261, A m s t e r d a m , 1978. North-Holland.
[201] R o b e r t Solovay. Provability interpretations of modal logic. Israel Journal of Mathe-
matics, 25:287 - 304, 1976.
[202] E d i t h Spaan. Complexity of Modal Logics. P h D thesis, D e p a r t m e n t of M a t h e m a t i c s
and C o m p u t e r Science, University of A m s t e r d a m , 1993.
[203] L. J. Stockmeyer and A. R. Meyer. Word problems requiring exponential time. In
Proceedings of the 5th Annual ACM Symposium on Theory of Computing, pages 1 -
9, New York, 1973. Association of C o m p u t i n g Machinery.
[204] T i m o t h y Surendonk. Canonicity for Intensional Logics. P h D thesis, A u t o m a t e d Rea-
soning Project, Australian National University, 1998.
[205] Oswald Teichmiiller. Braucht der algebraiker das auswahlaxiom? Deutsche Mathe-
matik, 4:567 - 577, 1939.
[206] S. K. T h o m a s o n . Semantic analysis of tense logic. Journal of Symbolic Logic, 37:150
- 158, 1972.
[207] S. K. T h o m a s o n . A n incompleteness t h e o r e m in modal logic. Theoria, 40:30 - 34,
1974.
[208] S. K. T h o m a s o n . Reduction of tense logic to modal logic I. Journal of Symbolic Logic,
39:549 - 551, 1974.
[209] S. K. T h o m a s o n . Reduction of second-order logic to modal logic. Zeitschrift fiir math-
ematische Logik und Grundlagen der Mathematik, 21:107 - 114, 1975.
[210] S. K. T h o m a s o n . Reduction of tense logic to modal logic II. Theoria, 41:154 - 169,
1975.
[211] S. K. T h o m a s o n . I n d e p e n d e n t propositional modal logics. Studia Logica, 3 9 : 1 4 3 - 144,
1980.
558 Bibliography
[212] S. K. Thomason. Undecidability of the completeness problem of modal logic. In Uni-

versal Algebra and Applications, Banach Center Publications Vol. 9, Warsaw, 1982.
P N W - P o l i s h Scientific Publishers.
[213] Jerzy Tiuryn. Unbounded Program Memory Adds to the Expressive Power of F i r s t -
Order Programming Logic. Information and Control, 60:12 -35, 1984.
[214] M. Tokarz. Connections between some notions of completeness of structural proposi-
tional calculi. Studia Logica, 32:77 - 89, 1973.
[215] J. W. Tukey. Convergence and Uniformity in Topology, volume 2 of Annals of Math-
ematical Studies. Princeton University Press, Princeton N J, 1940.
[216] Alasdair Urquhart. Decidability and the finite model property. Journal of Philosoph-
ical Logic, 10:367 - 370, 1981.
[217] Dimiter Vakarelov. Filtration theorem for dynamic algebras with tests and inverse
operator. In A. Salwicki, editor, Logics of programs and their applications (Poznan
1980), number 148 in Springer Lecture Notes in Computer Science, pages 314 - 324,
1983.
[218] J a n van Eijck and Fer-Jan de Vries. Reasoning about update logic. Journal of Philo-
sophical Logic, 24:19 - 45, 1995.
[219] Moshe Y. Vardi and P. Wolper. A u t o m a t a Theoretic Techniques for Modal Logics of
Programs. Technical report, IBM Research Report, 1984.
[220] Robert L. Vaught. Set Theory. An Introduction. Birkh£user, Basel, 2 edition, 1995.
[221] Yde Venema. Many-Dimensional Modal Logics. PhD thesis, Universiteit van Amster-
dam, 1991.
[222] Yde Venema. Canonical Pseudo-Correspondence. In Maarten de Rijke, Heinrich
Wansing, and Michael Zakharyaschev, editors, Advances in Modal Logic 98. CSLI,
1999. to appear.
[223] Albert Visser. Uniform interpolation and layered bisimulation. In Peter Hajek, editor,
GSdel 96, pages 139 - 164. Springer, 1996.
[224] Heinrich Wansing, editor. Proof Theory of Modal Logic, number 2 in Applied Logic
Series. Kluwer, 1996.
[225] Titus Welder. Simulation kristalliner Strukturen unter dem EinJ~ufl ether Scher-
strSmung. Wissenschaft und Technik Verlag, Berlin, 1996.
[226] Timothy Williamson. An alternate rule of disjunction in modal logic. Notre Dame
Journal of Formal Logic, 33, 1992.
[227] Timothy Williamson. Some Admissible Rules in Nonnormal Modal Logics. Notre
Dame Journal of Formal Logic, 34:378 - 400, 1993.
[228] Timothy Williamson. Non-genuine Macintosh Logics. Journal of Philosophical Logic,
2 3 : 8 7 - 101, 1994.
[229] Ryszard W6jcicki. Locical matrices stronlgy adequate for structural sentential calculi.
Bulletin de l'Aeaddmie Polonaise des Sciences, Sgries des sciences mathdmatiques,
astronomiques et physiques, 6:333 - 335, 1969.
[230] Ryszard W6jcicki. Some remarks on the consequence operation in sentential logics.
Funadmenta Mathematicae, 68:269 - 279, 1970.
[231] Ryszard W6jcicki. Matric approach in methodology of sentential calculi. Studia Log-
ica, 32:7 - 37, 1973.
[232] Ryszard W6jciekl. Theory of Logical Calculi. Kluwer, Dordrecht, 1988.
[233] Frank Wolter. Eigenschaften u n a b h i n g i g axiomatisierbarer polymodaler Logiken.
Master's thesis, II. Mathematisches Institut, FU Berlin, 1991.
[234] Frank Wolter. Lattices of Modal Logics. P h D thesis, FU Berlin, 1993.
[235] Frank Wolter. Solution to a problem of Goranko and Passy. Journal of Logic and
Computation, 4:21 - 22, 1994.
[236] Frank Wolter. Tense Logic, Part II: Decidability. unpulished, 1994.
Bibliography 559
[237] Frank Wolter. W h a t is the Upper P a r t of the Lattice of Bimodal Logics? Studia
Logica, 53:235 - 242, 1994.
[238] Frank Wolter. T h e finite model property in tense logic. Journal of Symbolic Logic,
6 0 : 7 5 7 - 774, 1995.
[239] Frank Wolter. A Counterexample in Tense Logic. Notre Dame Journal of Formal
Logic, 37:167 - 173, 1996.
[240] Frank Wolter. Properties of tense logics. Mathematical Logic Quarterly, 42:481 - 500,
1996.
[241] Frank Wolter. Tense logic without tense operators. Mathematical Logic Quarterly,
42:145 - 171, 1996.
[242] Frank Wolter. Completeness and Decidability of Tense Logics closely related to Logics
above K4. Journal of Symbolic Logic, 62:131 - 159, 1997.
[243] Frank Wolter. I n d e p e n d e n t fusions of modal logics revisited. In Marcus Kracht,
M a a r t e n de Rijke, Heinrich Wansing, and Michael Zakharyaschev, editors, Advances
in Modal Logic, pages 361 - 379, Stanford, 1997. CSLI.
[244] Frank Wolter. T h e structure of lattices of subframe logics. Annals of Pure and Applied
Logic, 8 6 : 4 7 - 100, 1997.
[245] Michael Zakharyaschev. Canonical formulas for K4, P a r t I: Basic results. Journal of
Symbolic Logic, 5 7 : 1 3 7 7 - 1402, 1992.
[246] Michael Zakharyaschev. Canonical formulas for K4, P a r t II: Cofinal Subframe Logics.
Journal of Symbolic Logic, 61:421 - 449, 1996.
[247] Michael Zakharyaschev and Alexander Alekseev. T h a t All Finitely Axiomatizable
Normal Extensions of K 4 . 3 are Decidable. Mathematical Logic Quarterly, 41:15 - 23,
1995.

Techniques

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Techniques

Hochgeladen von

Copyright:

Verfügbare Formate

A b o u t this B o o k

Berlin, March 1999

understanding higher modal logic. In Chapter 4 we develop duality theory

of extensions of $4. The first result is a substantial generalization of [247],

Algebra, Logic and Deduction

In this section we will briefly explain the n o t a t i o n as well as the basic

of Choice is equivalent to t h e s t a t e m e n t t h a t every set can b e w e l l - o r d e r e d .

:-- {y: e X ) ( y _ x)}

T h e s e laws are referred to as t h e laws of associativity, commutativity, idem-

follows t h a t ~ is a partial order, and x R y is the greatest lower bound (glb) of

We also use the notation n , c x x or n x , and likewise for U, as for lattices.

G R o u P s AND SEMIGROUPS. Given a set G and a binary operation • on

Suppose we have an additional operation -1 : G ---} G such t h a t the following

T h e n the structure (G, 1, -1 , .} is called a group.

1.2. Propositional Languages

f~Q.. ~tn(:)_l • Tmn(Z)

we will write / ( t o , . . . , t n - , ) instead of f ~ t o t ~ . . ~ t~-l. If we do not want

each f C F an f ~ ( f ) - a r y function from A to A. This definition is somewhat

This is called the a l g e b r a o f f ~ - t e r m s o v e r X or simply the t e r m a l g e -

fig0,..., gk-1](S) := f(g0(~),..., gk_l(s))

h : A --+ B such t h a t for all f E F and all elements aj C A, j < ~ ( f ) ,

THEOREM 1.2.2. Let t2 be a signature and 92 an ~-algebra. Put

A m a p h : A --+ B induces an equivalence relation ker(h) on A via ix, y) E

It is the fact t h a t O is a congruence relation t h a t makes this definition in-

PROPOSITION 1.2.3. Let h : 92 ~ ~ be a surjective homomorphism. Then

Let E C_ A × A. The smallest congruence relation containing E is denoted

PROPOSITION 1.2.4. Let 92 be an algebra. A binary relation on A is a

PROPOSITION 1.2.5. Let Q)~, i C I, be congruences of a given algebra 92.

PROOF. [-JiEI Oi is symmetric and reflexive; so is its transitive closure, q .

Now let f E Poll (9/) be a translation. T h e n

We derive the following useful consequence. Let (a, b) E [_] (Oi : i C I ) .

T e r m algebras have the important property that for any function v : X -+

1. For x E X we have ~(x) = v(x).

E x e r c i s e 1. Let A be a set, and E C_ A × A. Show t h a t the least equivalence

E x e r c i s e 2. (Continuing the previous exercise.) Show t h a t if E is reflexive,

Next we define the product of algebras. Let 92j, j C J, be a family of algebras.

Finally, PS(gC) C S P ( ~ ) since the product of subalgebras is a subalgebra of

the algebra freely generated by X in V; and for a cardinal number A we denote

E x e r c i s e 4. Show that either a variety contains up to isomorphism only the

E x e r c i s e 5. Let L be a language with signature ~. Let 7 be the cardinality

E x e r c i s e 6. Show with a specific example that the claim of the previous

exercise need not hold for finite 5.

E x e r c i s e 7. Show that for any class of algebras, HH(gC) C H(:E) as well as

relation E F- ~ says t h a t for any concrete instances of the occurring variables,

DEFINITION 1.4.1. Let L be a propositional language and F-C p ( C L ) × L.

Notice the following. In contrast to the s t a n d a r d literature, we require

(~ Y X). T h e terms ~ satisfying ~ ~- p are called the t a u t o l o g i e s of the logic.

DEFINITION 1.4.2. Let L be a language and R a set of rules over L . T h e n

To characterize ~_R define an i n s t a n c e of p = (E, ~} to be any pair

R - p r o o f of ¢i from E, i < n. Now let

It is straightforward to check that ( i s a proof of ~ from E. []

It will be proved later that 2-valued logics with 7 , -~ and /~ is b o t h

Another characterization of logics is via consequence operations or via

E x e r c i s e 8. Give a detailed proof of Theorem 1.4.3.

E x e r c i s e 9. Let R be a set of axioms or l - d r y rules. Show that A F R ~ iff

E x e r c i s e 10. Show that every consistent logic is contained in a P o s t -

E x e r c i s e 11. Show that in 2-valued logic

Thus if ~ =- ¢ is defined by F ~ +4 ¢, then = is a congruence relation. W h a t

E x e r c i s e 12. Show that there is no t e r m p in T, V, A such t h a t ~ F -~p F

E x e r c i s e 13. Show that if (L, F) is a logic, the m a p E ~-+ E e is a finitary

E x e r c i s e 14. Show that if E ~-+ E e is a finitary closure operator, the system

1.5. C o m p l e t e n e s s of Matrix Semantics

DEFINITION 1.5.1. A n ~ - m a t r i w for a signature ~ is a pair ~ = (92, D)