Arch. math.

Logik 22 (1982), 141-158

ON ORDE RI NGS OF THE F AMI L Y OF ALL LOGI CS *'1
Michak Krynicki and Jouko V~i~in~inen
The pur pos e of this paper is t o exami ne the st r uct ur al compl exi t y of the subl ogi c
rel at i on bet ween abst r act logics.
Let ~( ") denot e n 'h or der logic. Then cp(,) is a pr oper subt ogi c of A (" 1~ for each
n < co, and we have t he chai n
~ ( 1 ) < ~ ( 2 ) < ... <~cp(,)< . . . .
The quest i on nat ur al l y arises, what ot her ki nds of chai ns or part i al orderi ngs we
can have a mong sufficiently regul ar abst r act logics. Can one have a fami l y of logics
or der ed in the or der t ype of the rat i onal s (or per haps the reals)?
In Chapt er 2 we pr ove t hat any di st ri but i ve lattice, and t herefore any part i al
orderi ng, can be embedded i nt o the subl ogi c rel at i on a mong what we call nor mal
logics. Chapt er 3 is devot ed t o a subl ogi c rel at i on defined usi ng PC-cl asses r at her
t han el ement ar y classes. In t he last chapt er we restrict oursel ves t o the very special
logics of the f or m ~( Q1, - - - , Q,)- The si t uat i on becomes mor e pr obl emat i c but we
can still pr ove t hat any count abl e part i al orderi ng is embeddabl e i nt o the subl ogi c
rel at i on of these logics.
We confine oursel ves to single-sorted st ruct ures, but ma ny of the results are
equal l y t rue of many- s or t ed structures.
1. Introduction
The basi c i dea of an abst r act logic is very general and easily f or mul at ed : we have a
class s~ of obj ect s called al phabet s and a ma ppi ng ~ which relates every el ement t
of d with a triple
~, =(5' ~t , o~, ~-~ ) , where ~ - ' ~ t C~ x ~ .
Int ui t i vel y,
is the class of A,-structures,
is the class of ~cs ent ences ,
is the ~t - sat i sf act i on rel at i on.
* Eingegangen am tl.2.1980.
t Some of the results of this paper were contained in the Ph.D thesis of the first author. He
wishes to thank his supervisor Professor A. Lachlan for his help and encouragement.
142 Michak Krynicki and Jouko I/'diin~nen
Thi s definition is all one needs to formul at e such f undament al propert i es of
abst ract logics as compact ness and its generalizations (see [7]). But al ready the
f or mul at i on of t he vari ous LSwenhei m-Skol em propert i es is uncl ear due t o t he
lack of any not i on of cardi nal i t y of a model.
It is usually assumed t hat t he al phabet s are sets of predicate-, function-, and
const ant -symbol s, and t hat ~ is the class of all first or der st ruct ures over the
al phabet t. Thi s simplifies mat t ers, but rules out all non-classical logics. For a
t r eat ment coveri ng non-classical logics as well, see ['5]. However, once this step of
consi deri ng usual first or der st ruct ures is taken, a vari et y of model t heoret i c
not i ons become available. Wha t is mor e i mport ant , the general idea of an abst ract
logic, sket ched above, can be t ur ned into a ri gorous definition, which enabl es us to
carry out non-t ri vi al and i nt erest i ng model t heoret i c const ruct i ons on abst ract
logics. St andar d references for such definitions are [3] and [ i ] . A discussion on
the mut ual relationships bet ween these definitions in cont ai ned in [8].
In a cont ext where the al phabet s and structures are fixed, the following nat ural
subl ogi c-rel at i on suggests itself:
~=<~e' if and onl y if Vt e d ( ~ c _ ~ ' and g ~=~- ~' c ~( ~x ~) ) .
The st udy of this rel at i on is the mai n purpose of this paper.
The propert i es of t he rel at i on =< cert ai nl y depend heavily on t he part i cul ar
definition of an abst ract logic one uses. A t oo general definition woul d al l ow non-
st andar d compl et el y uni nt erest i ng logics to mess up the ordering, whereas a t oo
nar r ow definition mi ght excl ude logics which are relevant for t he orderi ng. We
shall st art with a fairly general class of logics we call normal. Nor mal i t y is all one
uses in such st andar d results of abst ract model t heory as Li ndst r 6m' s t heorem. As
an exampl e of a mor e restricted class of logics, we consi der t he class of logics
generat ed by a finite number of generalized (first-order) quantifiers, the so called
f i ni t el y generated logics. In sharp cont rast to normal logics, these logics have a
simple and well-defined syntax. Thei r ordering, as well, t urns out t o be very
different from t hat of nor mal logics.
Ther e are obvi ous set -t heoret i cal difficulties in any at t empt to give a ri gorous
definition of abst ract logics and their ordering. It is not onl y t hat an abst r act logic
itself is a map relating sets wi t h classes, but t he orderi ng < is defi ned bet ween
these maps. The reason why difficulties hardl y ever arise in pract i ce is t hat the
embeddi ngs of one logic i nt o anot her have a nice uni form nat ure.
In t he present cont ext we over come t he difficulty by simply switching t o class
t heory, mor e exactly to t he Most owski - Kel l ey- Mor se t heor y of classes.
In class t heor y families of classes can be coded i nt o classes in t he fol l owi ng well-
known way: If ~" is a class, the domain of J r , dom( ~") , is t he class
{xl 3y( ( x, y) ~gf f ) }. I f x is any set, oY" x is used t o denot e t he class { y l ( x , y ) ~gf f } .
Now we can t hi nk of 9ff as a code of the family {~Y'xlxedom(.Yf)} of classes.
Accordi ngl y we use the not at i on
J/ ~ ~f for 3 x( J# = Jfx)
X__c~t ~' for V~r ( Xe J l --* 2fe~ff' ).
On orderings of the family of all logics 143
We shall now proceed t o t he definition of an abst ract logic. Our present at i on
closely follows [3].
An element of the set T= U a)" is called a t ype. A st ruct ure of type t, 9.I, consists
n<t o
of a domai n t oget her with a t(i)-ary rel at i on for each i adom(t ). The class of all
structures of type t is denot ed by Mod t. Type t' is an ext ens i on of t ype t, if
d o m( t ) Cd o m( t ' ) and t'ldo,,)----t. If ~f" is a class of structures of type t and t' is an
extension of t, then the f r e e expans i on of ~ to type t' is the class ~fft, of 9.Ia Mod t,
such t hat the reduct Nit of 9.i to t is in .,Y'. Two types t and t' are equi val ent , t ~- t', if
dom(t ) = dom(t ' ) = k and t(i) = t'(jl) for some per mut at i on (Jo . . . . , Jk- 1) of k. I f X is a
class of st ruct ures of t ype t and t = t', with t ( i ) = t'(ji) t hen o,Y "`'t' is the class
,.g-t.t, = {oi12 [ = ( A, Rio . . . . , ei~ - , ) , (A, R0, . . . , R k_ 1) a ,X#}.
A subclass of Mod t, which is closed under isomorphisms, is called a model cl ass o f
t ype t. A f a mi l y o f model cl asses is a class J{" such t hat dom(~,") is a set and for
each i a d o m( X) , ~ is a model class.
Def i ni t i on l . A quasi l ogi c is a class P with domai n T such that L t is a family of
model classes of type t, for each t a T.
I f ~ is a quasilogic and t a T, we can also t hi nk of~t as a triple ( ~ , -~, 3-~,) where
~ = Mo d t , o~= {JY'l~a,~} and T~(9I, K) ~ 9.1aK, for ~21[a~ and ~,Y'a -o~ t. Thus
our definition is consistent with the general idea present ed in the beginning. Not e
t hat we do not allow a quasilogic to have a pr oper class of sentences (as e.g. ~ ) .
This restriction onl y plays the role of simplifying the formul at i on of cert ai n results.
A class 2,o which meets the ot her parts of Defi ni t i on 1, except t hat dom(L,t) may be
a pr oper class for some t a T, is called a cl ass- l i ke quasilogic.
Abst ract logics will be defined as quasilogics satisfying certain simple axioms.
Not e t hat any (it seems) pr oper t y of an abst ract logic can al ready be formul at ed
for quasilogics.
Def i ni t i on 2. An abst ract logic is a quasilogic satisfying the following three
condi t i ons :
(t) For all t a T , if Jg',~t[a.LP t, t hen J~fc~Jga~q~, and Mo d e - J' fa&,.
(2) For every t a T and for every extension t' of t, ~ffa-,q t implies J~.f"'a2't,.
(3) For all t , t ' a T such t hat t ~ - t ', oY' aSf , implies ~P"t' aLPt,.
This definition is equi val ent to the definition of a "generalized first or der logic"
in ['3].
The effect of axi om (1) is simply, t hat .L~ is closed under the usual finitary
proposi t i onal operat i ons. Axi oms (2) and (3) are the crucial st ruct ural axioms t hat
really distinguish an abst r act logic from a mere family of families of model classes.
Axi om (2) says t hat the t r ut h of a sentence onl y depends on the symbols
"occurri ng" in it (if 3ffaL-W, we t end to t hi nk t hat onl y the non-logical symbols
det ermi ned by t "occur " in 3if). Axi om (3) says t hat we can execute the simplest
possible subst i t ut i on oper at i ons: replacing of a predi cat e-symbol by anot her.
144 M ichak Kr yni cki and Jouko l, rdiiniinen
Fol l owi ng common usage, we call a class ~" of model s an e l e me n t a r y c l as s o f 2' , in
symbol s ~ e EC~, or . , ~s La, i f t here is a t ype t(.,~ r) such t hat JTE ~ett~. ~. We use off"
t o denot e Mod~(~c~--~f'.
Observe t hat an abst ract logic need not cont ai n any nonvoi d el ement ar y classes at
all : the abst ract logic .~o such t hat ~t = 0 for all t s T is onl y able to charact eri ze the
empt y model class. Some new axi oms are called for t o guarant ee non-t ri vi al
expressive power.
De f i n i t i o n 3. The logic ~a is non- t r i v i al if for each n~c9 the class
{ ( A , R , a 1 . . . . , a n ) I N C A n, ( a l , . . . , a n g e R }
and the class
{ ( A , a , b ) [ a = b }
are el ement ary classes of ~a.
Non-t ri vi al i t y is not yet enough to give all the el ement ary classes of ~, , ~, ; for t hat
we woul d need quantifiers. It proves useful t o i nt roduce t he following general
f r amewor k: Recall t hat a model class of t ype t is called a ge ne r al i z e d q u a n t i f i e r o f
t y p e t, or j ust a q u a n t i f i e r o f t y p e t [2].
De f i n i t i o n 4. Suppose Q is a quant i fi er of t ype t~ co". Let k = max {t(i)li < n}, let t' be
t concat enat ed with k zeros, and let t ~ be t concat enat ed with t(i) zeros. A logic L~ is
cl os ed unde r t he q u a n t i f i e r Q if for all classes 3 f o . . . . . J r , _ 1 ~ L/'t' such t hat Y/ = , g~'
for some J/ti~ ~-c-et,, t he class
where
{92~ Modc](lgl], R:~ . . . . . R~._"~ ' ) ~ Q } ,
R ~ ' = ( ( a o . . . . . at(i~ - 1 )1(9-I, a o . . . . . at~i~ - 1 ) ~ i l l } ,
is in ~c -
The most rel evant quant i fi er is, of course, the e x i s t e n t i a l q u a n t i f i e r
{(A, B ) t B ~_ A, B ~ 0}. It is easily seen t hat t here is a smallest non-t ri vi al abst r act
logic closed under t he existential quantifier, and the el ement ary classes of this logic
are exactly the el ement ary classes of Lo, o,.
It woul d seem like a very desirable pr oper t y of any abst ract logic t hat it be closed
under every quantifier which is its own el ement ary class. This is a st r ong f or m of a
s u b s t i t u t i o n a x i o m. Our axi om (3) legitimates onl y subst i t ut i on of predi cat e
symbol s t o predi ct symbols, but this st ronger axi om would legitimate subst i t ut i on
of any formul ae (under cert ai n coher ency condi t i ons) to predi cat e symbol s. Ther e
is a weal t h of different subst i t ut i on axi oms of varyi ng strength, and the one we are
now going to define is among the weakest. Thi s version is rel at ed to codi ng of
predicates i nt o one, somet hi ng one does in the well-known quant i fi er
mani pul at i on
V x 3 R q b ( R ( . ) , x ) ~ 3 R ' V x ~ 2 ( R ' ( x , . ) , x ) .
On orderings o f the f a mi l y o f all logics 145
Suppose ~" is a model class of t ype t~ co ~. Suppose i o . . . . . i k e n such t hat t ( i j ) : # 0 for
j = 0 . . . . , k. We define t' as follows:
I t ( i ) if i e n - - { i o , . . . , i k } ,
t ' ( i ) = I t ( o i ) + l ifif i=n.i~{io''"'ik}'
A model class ~o. . . i k of t ype t' is defined by
. . . . ik = { ( A , R o . . . . . R._ 1, a) l ( A, e(o ") . . . . . e(.~)_ 1 )~ oU} ,
where
( , ) _ l { ( a o . . . . . a t ( i ) ) l ( a o . . . . . a t ( i ) , a ) ~ R i } if i~ {i o . . . . , i k } ,
Ri - "[R~, ot herwi se.
D e f i n i t i o n 5. A logic 2 ' is AI R (admits i ndi zat i on of relations) if ,~/o...i ~2"
whenever ,gCs~ and i o . . . i , ~ t ( J { ' ) such t hat t (~f): # 0 for j = 0 . . . . , k.
For logics closed under the existential quantifier, AI R implies the pr oper t y of
being "st r ong" in the sense of [3].
D e f i n i t i o n 6. An abst ract logic is n o r m a l if it is non-trivial, AI R and closed under
the existential quantifier. N L denot es the collection of all normal logics. The
collection of all normal class-like logics is denot ed by N L c.
We feet t hat nor mal i t y is a stable and interesting pr oper t y of an abst ract logic, and
cannot be weakened (apart from omi t t i ng closure under negation) wi t hout an
essential decline in the ability t o carry out model t heoret i c const ruct i ons. The
definition of N L does not take place in class t heory itself, but on a metalevel:
nor mal i t y is a definable pr oper t y of classes.
2. Th e Or d e r i n g o f N o r ma l Lo g i c s
In this chapt er we shall discuss the usual sublogic rel at i on between nor mal logics.
Thi s rel at i on t urns out t o provi de N L with a non- modul ar l at t i ce-orderi ng into
which every distributive lattice can be embedded.
D e f i n i t i o n 7. Let 2" and 2 ' ' be nor mal logics. We say t hat 2" is w e a k e r t h a n 5 f ' , in
symbol s 2"_---2'', if for all t s T, 2", =c ext2"[- The logics 2" and 2 " are e q u i v a l e n t ,
~ = 2 ' ' , if 2 " < 2 ' ' and 2' ' __<~.
Not e t hat __<is a definable pr oper t y of classes, thus defined on the metalevel only.
The meani ng of the definition
jV'&a = ( NL, __<>
is t o be under st ood in the same way. A wor d about i dent i t y is here in order. Two
classes are identical if they have the same sets as elements. For classes which are
146 Michat Krynicki and Jouko Vdiiniinen
quasilogics we have the weaker not i on of equivalence, defined above. It does not
seem t o be plausible t o distinguish t wo logics with the same model classes, j ust
because t hey happen t o have a different definition. This is in accor dance wi t h our
general suppressi on of syntax. Wi t h this in mi nd we adopt the not i on of
equi val ence as the not i on of i dent i t y bet ween logics, and consi der , t / ' ~ as a
st ruct ure with this i dent i t y-rel at i on and the i ndi cat ed bi nar y rel at i on <. Havi ng
made this convent i on, it is emi nent l y obvi ous t hat ,-&o is a part i al l y or der ed
st ruct ure with a least element. We shall proceed to provi ng t he following st r onger
resul t :
Theorem 1. j . ~o is a compl et e lattice wi t h a least element.
Proof . Suppose { 2, i ] i 6I } is a family of logics and I is a set (such t hat I n t o =0 ) . We
can define a new quasilogic 2, by letting for each t~ T:
2 , , = f i
i ~l
It is readily verified t hat 2, is the infimum of {2,~1i~I} in J V' ~. Thus we can
concent rat e on the pr obl em of finding the supr emum of {P~Ii~I}. An obvi ous
candi dat e is the quasilogic 2, * defined for t~ T by
2, * not necessarily being a nor mal logic, we ext end it successively to one. For this
end, we shall consi der four nat ur al operat i ons on quasilogics. In the following 2, is
an ar bi t r ar y quasilogic.
Let Bool(~C~) be the quasilogic 2, ' defined as follows:
} ~CP;= s 0 ~je(i.l)le~2"", ~ f i L P , , X, ij =J{ilj, Jut~iil = & j , n , m<c o .
Fact 2. Bool ( 2, ) is t he least quasilogic containing 2 ' and closed under Bool ean
operations.
Let Ext ( 2, ) be the quasilogic 2, ' defined as follows:
2, ~= {~VI t here is an extension t 1 of t such t hat o~f'= J t '~
for some ~' E2,~1 }.
Fact 3. Ext ( 2, ) is t he least quasilogic containing ~ and closed under expansi ons.
Suppose ~f ~2, , and for some i o <k =d o m( T) , t(io)=0. Let t ' (j )=t (j ) i f j < i o and
t' (])= t ( j + 1) if i o < j < k - 1 . We let 3xW be t he class
{ (A, Ro, . . . , Rio_ 1, Rio + 1 . . . . . RkSt for some a t A (A, R o . . . . , Rio- 1, a, . . . , Rk) ~ ~,T~}.
On orderings of the family of all logics
Let Eq(& ) be the quasilogic defined by
,Lt ' = {3x 1 ... 3x, J' f [ of ' ~. ~ and t(3x 1 ... 3x, ~{' )= t }.
147
Fact 4. Eq (So) is the least quasilogic containing f and closed under the existential
quantifier.
The quasilogic Ind(~C,) is defi ned as the P' such t hat
5 t' = {JY'It(JY ~) = t and for some ~'-1 ~ ~ and i o. . . i k e dom(t(~{1))
such t hat t(~Yl) (i~) :~ 0 for j__< k,
X=~o...i~}~2',.
Fact 5. Ind(, ) is the smallest AI R quasilogic containin 9 .o c&
Put t i ng the above operat i ons t oget her we obt ai n a useful operat i on Nor as follows.
Let {~"[ n<o) } be a family defined by
5 =5
~ , + 1 = Ind(Eq(Ext(Bool(~")))).
Let
Nor ( 2' )
be the quasilogic 2,~' defined by
n<O
Fact 6. I f ~ is a non-trivial quasilogic, then Nor ( 2' ) is the smallest normal logic
stronger than ~ .
Let us now ret urn to the pr oof of Theor em 1. Let ~v = Nor(&a,). The abst ract logic
P is the desired nor mal supr emum of {pi[is I}.
The above pr oof also shows t hat ~/-&o is a compl et e sublattice of.AFoL~v~ (defined as
( NL c, < ) ) and that/' &v c has the following st ronger completeness pr oper t y : every
i ndexed family { ~ l i ~ J } , J a class, has a supr emum and an i nfi mum in jg-&o.
In the sequel we use U and 0 to denot e the lattice operat i ons of ~/V& .
What ki nd of sublattices does Jlr& a have? We have the following general resuit
concerni ng distributive lattices. Non-di st ri but i ve lattices will be discussed later.
Theor em 7. Every distributive lattice (a set) can be embedded as a sublattice into
y&t ' .
Proof The proof, as many subsequent ones, depends on the following special
family of nor mal logics: For any ordi nal c~ let ~f-o(yfl, ~)ffz) denot e the class of
148 M icha~ t Krynicki and J ouko I/'dfin~nen
structures of the empt y type and of cardi nal i t y c0, (at most o),, at least c%
respectively). If K is any family of model classes, we let Nor(K) denote the logic
Nor ( 2' ) where ~ is the smallest non-trivial quasilogic havi ng every ~ f s K as an
el ement ary class. If K= {~ffo, .--, ,T',}, we write Nor ( ~" o . . . . . ,'F~) for Nor(K). For the
rest of the proof, K denotes a subfamily of {~f~lde 3, e e On}.
Fact 8. Every elementary class of Nor(K) is of the form
U (Mod,(~b~)c~a(O). a~ffc~gan ... n a ( n - 1) - ~' 2: ~' t ) ,
where I ~_ 2", Yt"~ ~ K, ~ is a sentence of -~o, and
i f e=O,
e' ~"' =t~/"--' -7 if e = l .
The fact is proved by observing the inductive definition of Nor(K) and by verifying
t hat the quasilogic of model classes of the form displayed is closed under the
const i t uent s Bool, Ext, Exp, Ind of Nor.
An i mmedi at e consequence of the above fact is:
Fact 9. If the family K contains model classes of the form o(o only, then for any ~ we
have
JfceNor(K) ~ ~ e K.
Let us now return to the proof of Theorem 7. Suppose D = (D, < ) is a distributive
lattice. We may assume t hat D is a lattice of subsets of On. For d~D, let
Ka={~c]~sd}. We define a mappi ng F:D --* NL by letting
F(d) = Nor (Ka).
I f a<b in D, then K, ~K b, whence F(a)<F(b). On the ot her hand, if not a<b, and
a~a- b, then j,f-o is an elementary class of F(a), but, by Fact 9, not of F(b). Thus
a<b in D is equivalent t o F(a)<F(b) in Jtr& . Moreover, by Fact 8, F(a)nF(b)
= Nor ( K. ~Kb) = Nor(K~,b) = F(anb), and similarly F(a)wF(b) = F(awb). []
The pr oof of Theorem 7 act ual l y establishes the following st ronger concl usi on:
Theorem 10. The complete lattice of atl sets of ordinals can be completely embedded
into JV't ~.
Corollary 11. Every partially ordered class every initial segment of which is a set, can
be embedded into ~/'P.
For the not i on W~ stronger embedding properties can be proved:
On orderings of the family of all logics 149
Theor em 12. The complete (in the same sense as JV;~-q~c) lattice of all classes of
ordinals can be completely embedded into ~/'~*q~. Every partially ordered class can be
embedded into Jf ' ~c.
Anot her way of i mprovi ng Theor em 7 is the following: Suppose 2 ' is a nor mal
logic. Let f be the L6wenhei m number of Z~o, t hat is,
# = s u p { mi n { I g . t l l g X~ } [ ~ 2 ' & ~ 4 : 0 } .
We can carry out the pr oof of Theor em 7 using classes d for a > E onl y and
addi ng t hem t o ~ rat her t han t o ~'~oo,. In this way we obt ai n the following result :
Theorem 13. Let ~ be a normal logic. The complete lattice of all sets of ordinals can
be completely embedded into ~A/'~ above . ~. Hence every distributive lattice and
every partial ordering every initial segment of which is a set, can be embedded into
~A/'~ above .ZP.
Thus for example, every ~ E JV;LP has a pr oper class of i ncompar abl e extensions
and a pr oper linearly or der ed class of extensions in JV'& .
How about non-di st ri but i ve lattices ? Here the situation is very unclear. We do not
even know if there is any lattice at all which cannot be embedded i nt o jC'cL.
However, the most common non-di st ri but i ve lattices are embeddabl e as the
following result indicates:
Proposition 14. The lattices
e h
:& <5
a
a
can be embedded as sublattices into ~ff~c&
Proof. Let a =~ o , , b - - No r ( Xl ) , c = No r ( d l , df l ) , d = No r ( g f ", JUwd{' ), and
e = No r ( X , d l I, WOa, X ) By Fact 8 we may i nt er t hat c ~d = a. It can be pr oved
t hat e =bwd. This establishes the first claim. For t he second lattice we t ake
f =No r ( ~f f ) , g =No r ( d l 2 ) , and h=Nor(dfl,~). The equat i ons b<__fug,
f <=bug, g<=bu f follow easily. Trivially b~ f =f c ~g=gnb=a. []
Corol l ary 15. JV'~L# is not modular, and therefore not distributive.
Thus the converse of Theor em 10 fails: ,/V'~CP cannot be embedded i nt o the lattice
of all sets of ordinals.
150 Micha~ Krynicki and Jouko Vdi~ngmen
3. The Ordering Based on PC-Definability
In this chapt er we shall consi der a sublogic rel at i on based on PC-classes. This
or der i ng of NL is not as nat ur al as <, but it is in fact mor e frequent a nd mor e
useful in practice. The mai n results concerni ng this new or der i ng follow t he same
pat t er n as t he results of t he previ ous chapter.
Let us st art with some definitions. The reduct of a model 9.I to t ype t is denot ed by
9,IIt. The r educt of a model class ~ t o t ype t is defined as {~I/tig, I e cU} and denot ed
by ~,Ult. Let ~ be a quasilogic. A PC-class of, is a model class which is t he reduct
of an el ement ar y class of 2".
Definition 8. If 2" and 2 ' ' are quasilogics, we say t hat 2" is PC-weaker than 2' ' , if
every el ement ar y class of 2" is a PC-class of 2' ' , in symbol s : 2" <pc~ '. We say t hat
2 ' and 2 ' ' are PC-equivalent, if 2" < pcC~ ' and 2 ' ' < Pc2", in symbol s : 2" = Pc2' ' .
We can now give NLa new st ruct ure by letting =Pc be t he i dent i t y and <pc the
ordering. Let us denot e this new st ruct ure by .A/'&~. The st r uct ur e , t/' &o~, is
defined similarly.
Theorem 16. JV' A~ is a complete lattice with a least element.
Proof Let I be a set (Ic~co=0) and {2"i [i sI } a family of nor mal logics. The
st rai ght forward intersection of the family does not necessarily yield t he infimum,
but let
2" = No r ( { ~( l J l and J~" are PC-classes of 2"i for every i e I}).
As it is cl ear t hat 2 ' ' _-< 2" for any l ower bound 2 ' ' of {2"i[i~ I}, we can concent r at e
on provi ng
2 " <p c 2 " i for all i e t .
The definition of Nor gives a represent at i on of 2" as a uni on Q) 2"n. Tri vi al l y
n<~tO
2" 0< ~c2"~ for all i e I. The rel evant i nduct i on step follows from t he fol l owi ng eight
equat i ons:
Suppose s t and N' are model classes and J {' = ~' [t, ~ ' = ~ [ t . Then:
(1) J g ' w~ ' = ( d " u ~ " ) l t , where t' is some extension of t ( d ) and t (~),
(2) ~' ~c ~J l =( dt ' c ~" ) l t , where t' is as above,
(3) X ,1 =dr ' t i l t 1, where t' is some extension of t l and t ( d) , and t'~-t'~,
(4) o~U ' =d" ' t i [ t ~, where t and t'~ are as above,
(5) 3x X=( 3x A) l t ' , where t"O=t,
(6) Vx X = Va Vx ( x . a v (~yt ... y, d) nl ..... + p)It', where t "0 = t and n 1 = dom(t ),
(7) ~f~io...ik=S~io...Jtl, where t 1 =t (~o. . . i k),
( 8 ) ~ = ( ~ ) i o . . . , k .
On orderings o f the f ami l y o f all logics 151
All these equat i ons are ent i rel y trivial, with the except i on of (6), which can be
verified as follows:
d ~ Vx ~ r ~-~ Va ( ~ , a ) ~ '~ ~ Va3R3b( gX, a , R , b ) ~ *-~ V a 3 R ( 9 i , a , R ) ~ 3 y d
3 R V a ( 9.I, R, a ) ~ V x ( x +-a v ( 3 y s J ) . , . . . . 1+ v)
*-+ 3R( 9. I , R ) ~ Va Vx ( x * a v ( 3 y s ) . , . . . . 1 + p)
, - . 9.I~ Va Vx ( x 4: a v ( 3y ~ ) . . . . . . 1 + p)lt' .
The Equat i ons (1)-(8) give i mmedi at el y the i nduct i on step for the pr oof of
~ " < FcZP i.
This ends t he pr oof t hat , is t he infimum of {#ti~ t}. The pr oof t hat {L.9~lie I}
has a supr emum goes al ong t he same lines: Let
= Nor({2' ~l i e I}).
Using (1)-(8) one readily shows t hat Za is the desired supremum.
Theorem 17. Ev e r y di s t r i but i v e l at t i ce ( a s e t ) can be e mb e d d e d as a s u b l a t t i c e i nt o
jV'C~~. Ev e r y par t i al l y or de r e d cl ass can be e mb e d d e d i nt o ~N'c~o ~. Bo t h e mbe d-
di ngs can be done abov e a n y gi v e n ~ 6 ~ 4 r ~ g i.
P r o o f Ther e are no essential differences to the pr oof of Theor em 7. Not e t hat if K
is as in Fact 9, then Jt " is a PC-class of Nor ( K) iff dg'fle K.
Also the pr oof of Pr oposi t i on 14 carries over :
Proposition 18. JV' A~ and JV2~q~ ar e non- modul ar .
Closely related to < PC is the so called A-operation. We say t hat an abst ract logic
has the Sousl i n-Kl eene-i nt erpol at i on pr oper t y if j~/-~ ~o whenever ,~" and
are PC-classes of ,e. Let N L I denot e the collection of all normal logics with the
Sousl i n-Kl eene-i nt erpol at i on propert y.
As a pr oper t y of an abst r act logic, Souslin-Kleene i nt erpol at i on is clearly
desirable, but unf or t unat el y fairly rare. The most famous examples of logics with
this pr oper t y are A~,,~ and 2, 0, , o. However, its i mpor t ance is largely due t o t he
following notion. For any quasilogic L~ a let
A( ~ ) = Nor({J~fflof and , ~ are PC-classes of .~}).
Clearly, A(~ a) is the smallest extension of J;e to a normal logic with the Souslin-
Kl eene-i nt erpol at i on propert y. For details concerni ng this oper at i on t he reader is
referred to [4], where also a pr oof of the following l emma can be f ound :
Lemma 19.
(1) l f X e A ( ~ ' ) , t hen ~ i s a PC- c l a s s o f C~,
(2) ~ has SKI i f and o n l y / f Z P =A(o~P),
(3) ~v = vcA(~),
(4) 2 ' = p c -v' i f and o n l y / f A ( ~ ) = A ( Z ~ a ' ) .
152 Micha4 Krynicki and Jouko Vdgmiinen
Corollary 20. The st ruct ures ( N L I , <=) and ( N i l , <= PC) coincide and are isomor-
phi c to ( N L , <= Pc)-
It follows from this cor ol l ar y t hat ( N L I , < ) is a compl et e lattice. It is not a
sublattice or ( N L , < ) , t hough, and we onl y have the resul t :
Corollary 21. ( N L, <PC) can be embedded as a suborderi ng into ( N L, ~ ) .
As a final observat i on on , 4r&~ we have the following appl i cat i on of Li ndst r6m' s
t heor em [3] :
Proposition 22. Le t 5f ( Qo) be t he logic wi t h t he quant i f i er "t here exi st s i nf i ni t el y
many". The logic d(~q'(Qo) ) is an at om o f ~r , ~q~.
Pr o o f Suppose ~,~,~ < pc.LP< pcA(L~O(Qo)). Then every non- empt y q~s..,~o cont ai ns a
count abl e model. Thus by Theor em3. 1 of [3], the class of model s
{(A, < >i(A, < ) -- (co, < )} is PC-defi nabl e in .~e. Hence indeed
~ =pc~( Oo) . [ ]
4. The Ordering of Finitely Generated Logics
In this chapt er we shall consi der logics which are generat ed by a finite number of
generalized quantifiers. These logics const i t ut e a r at her special but highly interest-
ing subfamily of NL. The main result is, t hat every count abl e part i al or der i ng can
be embedded into the sublogic rel at i on of these logics.
Def i ni t i on 9. Suppose Q is a quant i fi er of type t ee)". Let t' and t i be as in
Definition 3. We define a logic &O(Q) as follows: LP(Q)= U 5f,, where
~ o =Nor ( Q) , "~'
'f,+ 1 = Nr({M(gfo . . . . . J r , - 1)lgfo . . . . . aY-,-i ~ ~,, ~ = "#[', . . ~ (Le,),,}),
..., Modcl (I I, Ro , ,
M( So, Sf. _ 1) = {9.i6 9.I x , .... R f y { ' ) ~ Q }
and
Rf ' = {( a o . . . . . %0- 1 )1(9-I, a o . . . . . a,,)_ 1 ) ~ ~Pl,}.
I f Q1 . . . . , Q, are quantifiers, a logic ~a(Q1 . . . . . Q,) can be defined in an anal ogous
way, closing ,0~o ~ inductively under each Qi-
Not e t hat ~(Q~, . . . , Q, ) is the smallest nor mal logic which is closed under
quant i fi ers Q1 . . . . . Q,. Logics of this ki nd were i nt r oduced in [1].
Def i ni t i on 10. A logic ~ is f i ni t el y generated, if t here are quantifiers Q1 . . . . . Q, such
t hat ~ = 2' (Q1, ..., Q,).
On orderings of the f ami l y of all logics 153
If 2"is finitely generated, one single quantifier Q can be found such t hat 2 ' = 2"(Q).
In a simple case, suppose 2"= ~(Q1, Q2, Q3) where Q1, Q2, and Q3 are arbi t rary
quantifiers of type ( 2) . Then ~ = 2"(Q), where
Qx y z A ( x ) B ( y , z) ~ (IA(-)[ = 1 a Ql y z B ( y , z)) v
([A(-)[ = 2 & Q2yz B( y, z)) v
([A(.)I=3 & QAy z B( y , z ) ) .
Thus it is nat ural to t ake logics of the form ~(Q) as canonical representatives of
finitely generated logics. It is not true t hat every logic is finitely generat ed; for
example, second order logic and the A-closure of weak second order logic are
not.
The problem whether a given logic is finitely generated seems interesting and has
certainly wider relevance, but it will not be treated here.
Let us use N L Q to denote the family of all (necessarily normal) logics of the form
2"(Q), and JV'&.~ to denot e the structure ( N L Q , < ) .
By virtue of what has been said above, it is obvious t hat :
Pr o po s i t i o n 23. JV:~~ is an upper semi l at t i ce.
Problem 24. I s J V' e ~ a l at t i ce?
The structure of . /trf~ seems much harder to reveal t han t hat of Jtr& and many
open problems immediately suggest themselves. Our main result concerning
.A/'2"N is the following universality propert y:
The o r e m 25. Ev e r y c ount abl e di s t r i but i ve l at t i ce can be e mbe dde d as an upper
s ubs emi l at t i ce i nt o JV:~q'~.
P r o o f Suppose D = (D, N ) is a countable distributive lattice. We may assume t hat
D is a lattice of non-empt y subsets of o~-{0}. Let
For d , e D, let
and
D = {d. l n < o} .
Q. = {(M, A, B)I IAI = me d. and IBI = w,.}
Q = { < M , A , B > I I A I = n and IBled.}.
We obt ai n the desired embedding by letting
2". = 2"(Q., Q).
Suppose at first d , < din. Now Q, x y A ( x ) B ( y ) is equivalent to
154 Micha-~ Krynicki and Jouko V'dgmi~nen
whence .oq~, < Sam . Not e also t hat if d~ v dm= d k, t hen San v Sam = LPk because on the
one hand
QkxyA(x)B(y) ~ QnxyA(x)B(y) v QmxyA(x)B(y)
and on the ot her hand La, < Sak, Sam < Sak" Thus t o end the proof, we j ust have t o
pr ove t hat ~ , < Lain implies d,<dm. For this end, suppose not dn<dm, and
i edn- d, , . Let 92=( 09) and ~3=( a) i ) (note t hat i4:0).
Lemma 26. For any formula qS(x 1, ...,Xk) of Sam and any a 1, ...,akeog,
92~qS(a 1 . . . . ,ak) i f and only i f ~ b ( a 1 . . . . ,ak).
Proof. For any a and b in e) i t here is an aut omor phi sm of 92 which i nt erchanges a
and b, but leaves the ot her elements fixed. The st ruct ure ~ has the same pr oper t y.
Hence the Bool ean al gebra of subsets of 1 9 2 1 ( 1 ~ 3 t ) which are Sa(Qm)-definable using
a given sequence a~, . . . , a, of paramet ers, is generat ed by t he el ement s
{a 1 } . . . . , {an}. Usi ng this fact, the claim is easily proved by i nduct i on on t he length
of qS(xl . . . . . x,). [ ]
Let us ret urn to the pr oof of Theor em 25. The structures 92 and ~3 can be
separat ed by the following Sa,-sentence
3xi ""3Xn(,<_k/~<t<_iXk:~XtAQx.XY(k~_iX=Xg)(Y=Y)) _
On the ot her hand, Lemma 26 implies t hat these models cannot be separ at ed by
any sentence of L,e m. Ther ef or e San ~ Sam and the pr oof is complete. [ ]
Corol l ary 27. Every countable partial ordering can be embedded into J i f f Y .
The embeddi ng of Theor em 25 Can also be done above any given Le NLQ - one
j ust has to choose the model s 92 and ~3 mor e carefully. The result cannot be
generalized t o uncount abl e lattices because W& ~ is e)l-like, t hat is, every
SaeNLQ has but count abl y many predecessors. Whet her every ~Ol-like part i al
orderi ng can be embedded i nt o . /ff~2~ remai ns open.
Let us t hen t ur n i nt o non-di st ri but i ve lattices. Again the si t uat i on is even mor e
uncl ear t han with J t r ~ . It is obvi ous t hat t he five-element lattices of Pr opos i t i on
14 can be embedded into .A/'&~'.~ such t hat sups are preserved. But t o preserve infs
mor e powerful met hods seem to be needed:
For any classes ~g of cardinals let
Q~eC= {<A, B)IB is a linear orderi ng of its field with a dedeki nd cut (B 1, B2) such
t hat the cofinality of B~ and the co finality of the reverse or der i ng of
B 2 are in (g}.
On orderings of the family of all logics 155
Theorem 28 (S. Shel ah [6]). I f there is a weakl y compact cardinal, or a cert ai n
combi nat ori al principle ( t r ue in L ) holds, t hen there is a class cg o f regul ar cardinals
including o3 such t hat the logic ~q~(Q~) is f ul l y compact.
Corollary 29. Under t he hypot hesi s o f Theorem 28, the pent agon lattice ( as in
Proposi t i on 14) can be embedded into ~/'~q'.~ such t hat sups, infs, and least elements
are preserved.
Proof . Usi ng the not at i on of Pr oposi t i on 14, let a = ~o~,~, b = ~( Qo) , d = 2g( r~aq
e = ~( Qo, Q~9, and c = A(Qo, Q), where
Q = {(A, B, F ) [ Ther e are an a e A and an c~eOn such t hat
( A, B, a ) "- ( a, <, co) and F has the pr oper t i es
(1) Vx y z u ( F( x , y , z ) & F( x , u , z ) ~ y = u )
(2) V x 3 y V z 3 u ( B ( z , x ) ---, B(u, a) & F( y , z , u) ) .
A si mpl e Fuhr ken- t ype r educt i on- ar gument shows t hat c satisfies the L6wenhei m-
Skol em t heor em down to co. As d is compact , we obt ai n f r om Li ndst r 6m' s t heor em
the result t hat d A C exists and equal s a. Thus it suffices to pr ove t hat c < b v d. But
consi der the conj unct i on qS(B, F) of the following sentences of e :
(1), (2) as in the definition of Q,
(3) "B is a strict l i near orderi ng with a least el ement such t hat every non- maxi mal
el ement b has an i mmedi at e successor b + 1",
(4) 3 x V y ( B ( y , x ) ~ - Qo z B ( z , y ) ) ,
(5) - Q~x y B( x , y).
I f an infinite model of ~b(B, F) cont ai ns a sequence (b.). <o~ such t hat B(b.+ t, b.) for
all n < co, t hen the sets
Al ={ x l V n <c o : B ( x , b . ) } , A 2 = { x l 3 n < c o : B ( b . , x ) }
const i t ut e a Dedeki nd cut. By (1), (2), and (4) the model is count abl e. As co~ C, the
cut (A1,A2) cont r adi ct s (5). [ ]
We have obser ved al r eady t hat ,#-&o~ cont ai ns count abl e chai ns only. But the
following trivial result shows t hat there is no bound on the size of ant i - chai ns:
Proposition 30. Le t Q~={ ( A, B) [ BC=A, IBl~co~}. The logics -_~(Q~), a ~On, are
incompatible in ~4/'~P~ (even in ./V'~.~ ).
"Proof. Suppose A__< A(Q~) and A=< A(Qp) where a <f l . Let ~b~ ~. As ~be ~(Q~),
we can find a ~E .L,?~ such t hat q~ and tp have the same model s of power <co~. Ifg.I
is any model , t here is an el ement ar y subst r uct ur e ~3 of 92[ of power _< co~ wi t h
respect t o the logic ~(Q~). Now we have
whence ~b and tp have the same models. Thi s shows t hat A= ~eo~o~. [ ]
156 Mi cha~ Kryni cki and Jouko V'dgmgmen
I t is also easily pr oved t hat any L~'s ~4:.L:'2/has a pr oper class of i ncompar abl e
ext ensi ons in , f-&o~. The following new not i on leads us t o a st udy of t he l ocal
st r uct ur e of J{:M' ~.
De f i n i t i o n l 1. Let x be a cardinal. A generalized quant i fi er Q is b o u n d e d b y ~c, if
every st ruct ure in Q has cardi nal i t y < x. A logic L# is b o u n d e d b y x , if L: = e(Q) for
some Q which is bounded by x.
The most obvi ous example is the quant i fi er "t her e exists at most co~ many", which
is bounded by co,. Trivially, t here are at most 22. non-equi val ent logics bounded
by x.
Theor em 31. L e t x be a c ar di nal > o2. T h e r e ar e e x a c t l y 22~ i n c o mp a r a b l e f i n i t e l y
g e n e r a t e d l ogi cs bounde d b y ~c.
Pr o o f . For any x_~x, let 9.I x be the st ruct ure Q, <x> where the or der t ype of <x
results from t hat of x by replacing every ct sx by a copy of r/(x* + K) (r/is the or der
t ype of the rationals), and every e ~x by a copy of r / ( x*+K+ 1). Thus 9 A~ Ny
whenever x 4= y.
For o~C~(K), let
Q~ = {9.I19.I ~ ~I~ for some x~o~}.
Let J~ and f# be different subsets of ~ ( x ) - {x}. We claim t hat e(Q~) and F(Q~)
are i ncomparabl e. Indeed, suppose Qs* is definable by a sentence qb of #(Qe). Now
we need a simple l emma:
Lemma 32. F o r e v e r y f o r mu l a ~)(x 1 . . . . . x , ) o f 2/~(Q~) t he r e is a q u a n t i f i e r f r e e
f o r mu l a l p(x 1, . . . , x , ) s uch t hat f o r al l x ~ f : ,
% ~ c b ( x l , . . . , x , ) ~ ~ o ( x l , . . . , x , ) .
Pr o o f . We use i nduct i on on t he length of ~b(x p . . . , x,). As t he t heor y of dense linear
or der admi t s el i mi nat i on of quantifiers, we onl y have t o consi der the i nduct i on
step for t he quant i fi er Q. Suppose t herefore
4) ( x , , . . . , x. ) = O ~, uvO( u, v, x p . . . , x. ) ,
where O(u, v, x 1 . . . . . xn) is quantifierfree. If a 1 . . . . . a , e x and x~f#, the or der t ype of
{ ( b, c >l g. I x DO( b, c , al . . . . . an)} is never t hat of an 9.1y, yef~. Therefore, in such
model s O( Xl , . . . , Xn ) is equi val ent to x 14=xx. [ ]
We are r eady t o cont i nue t he pr oof of t he t heorem. For the chosen qb we can find a
quant i fi erfree ~0 such t hat for x~f f ,
%~4, ,--, ~.
But now it follows t hat 9.I~qb, whence t <e ~, a cont radi ct i on. [ ]
On orderings of the family of all logics 157
The previ ous t heor em was one of the very few places so far where we have made
use of non- monadi c (of t ype t~'2) quantifiers. This suggests the st udy of the
subst ruct ure of .Ar&O~ consisting of monadi c quantifiers only.
Let MLQ be the collection of all ~( Q) such t hat Q is a general i zed quant i fi er of a
monadi c t ype t sn2 (neco). Summi ng up t he results the proofs of which do not use
non- monadi c quantifiers, yields:
Theorem 33. a) The structure ~/l~q'~ = ( MLQ, <) is an upper subsemilattice of
b) Every ~ ~/tl~.~ has a proper class of incomparable extensions.
c) Every countable distributive lattice can be embedded as an upper subsemilattice
into ~ l ~ . ~ .
As a compl ement to these similarities bet ween ,4r&N and ~/ &o~ we have the
following local difference:
Theorem 34. Let ~ be a cardinal ~co and 2=] {#<~1~ a cardinal}l. There are
exactly 24 incomparable logics of the form Cp(Q), where Q is a monadic quantifier
bounded by K.
Proof. Suppose at first 2 > co. For any set X of infinite cardinals =< x, let
Qer = {( A, B)IBC=A and I BI sX}.
The number of different such sets X is 2 z and different sets X give rise to
i ncompar abl e logics 2f(Q=) (this is seen as in the pr oof of Theor em 27). Suppose
t hen ;~ =o9. For X =c co, let Qer be defined as above. Suppose L(Q=)<= L(Qe). Then
t here is an i dent i t y sentence ~b of L(Qe,) such t hat ~b is equi val ent t o Q=v (v = v).
Thus
n s X if and onl y if ( n ) ~ b .
It follows t hat X is Tur i ng reducible t o @ Now the claim follows from the fact t hat
t here are 2 ~ i ncompar abl e degrees of unsotvability. [ ]
REFERENCES
[1] Barwise, J. : Axioms for abstract model theory. Ann. Math. Logic 7, 221-265 (1974).
[2] Lindstr6m, P.: First order logic and generalized quantifiers. Theoria 32, 187-195
(1966).
[3] Lindstr6m, P.: On extensions of elementary logic. Theoria 35, 1-11 (1969).
[-4] Makowsky, J., Shelah, S., Stavi, J. : A-logics and generalized quantifiers. Ann. Math.
Logic 10, 155-192 (1976).
[5] Manders, K. : First order logical systems and set-theoretical definability (to appear).
[.61 Shelah, S. : Generalized quantifiers and compact logic. Trans. Am. Math. Soc. 204,
342-364 (1975).
158 Micha4: Krynicki and Jouko Vdi~ni~nen
[7] Stavi, J. : Compactness properties of infinitary and abstract languages. I. Logic
Colloquium '77 (North-Holland, Amsterdam 1978), pp. 263-276.
[8] Westerst~ihl, D. : Some philosophical aspects of abstract model theory. Philosophical
Communications No. 2. University of Gothenburg, 1976.
Michat Krynicki
Institute of Mathematics,
University of Warsaw,
PKiN IX p.
PL-00-901 Warszawa, Poland
Jouko V~i~in~inen
Department of Mathematics,
University of Helsinki,
Hallituskatu 15,
SF-00100 Helsinki 10, Finland