THERE ARE NO INFINITE ORDER POLYNOMIALLY COMPLETE LATTICES AFTER ALL

MARTIN GOLDSTERN AND SAHARON SHELAH Abstract. If L is a lattice with the interpolation property whose cardinality is a strong limit cardinal of uncountable conality, then some nite power Ln has an antichain of size . Hence there are no innite opc lattices. However, the existence of strongly amorphous sets implies (in ZF) the existence of innite opc lattices.

0. Introduction We call a lattice L n-order polynomially complete (opc) if every monotone function Ln L is induced by a lattice polynomial, and we say that L is order polynomially complete if L is n-order polynomially complete for every n. This denition is from Schweigerts Ph.D. thesis [4]. The survey [3] gives several results and bibliographical references for results on order polynomially complete lattices. While the nite opc lattices are now well understood, the main question on innite opc lattices: are there any? has remained open until now. Haviar and Ploscica [2] showed that there is no opc lattice of size < . We showed in [1] that the size of an innite opc lattice (if one exists at all) must be a strongly inaccessible cardinal. We now complement this result by showing (in ZFC) that the cardinality of an opc lattice cannot be a strongly inaccessible cardinal. Hence there are no innite opc lattices. Again the proof is not algebraic in nature, but based on a counting argument. Unlike our previous proof, which employed the heavy machinery of partition calculus, this paper uses only very basic set theory (the notions of conality and strong limit) and some baby model theory (the notion of type). We also point out that some version of AC (the axiom of choice) is necessary for our result, since under a strong negation of AC there are
Date: Jan 1999. 1991 Mathematics Subject Classication. Primary 06A07; secondary 08A40, 06B99, 03E55. Key words and phrases. polynomially complete, lattice, interpolation property, inaccessible cardinal, amorphous set, axiom of choice. The second author is supported by the German-Israeli Foundation for Scientic Research & Development Grant No. G-294.081.06/93. Publication number 688.
1
modified:1999-09-02

688

revision:1999-08-31

MARTIN GOLDSTERN AND SAHARON SHELAH

pathological sets, which (while being innite) are suciently similar to nite sets that it is still possible to build an opc lattice on them. 1. No opc lattices 1.1. Denition. We say that a lattice L has the unary interpolation property (1-IP), if every monotone function from L into L can be interpolated by a polynomial on any nite set. (Equivalently: Whenever a, b, c, d L, b a, c d, then there is a polynomial p with p(a) = c, p(b) = d.) 1.2. Theorem. Assume that L has the 1-IP. Let be the cardinality of L and assume that is a strong limit cardinal of uncountable conality. Then for some natural number n there is an antichain A Ln of cardinality . Proof. Let L be a lattice satisfying the assumptions of the theorem, and pick any two distinct comparable elements of L. We will call them 0 and 1, where 0 < 1. We will dene sequences (Li : i < ), (ai : i < ), (bi : i < ), (i : i < ), (ni : i < ), (i : i < ) such that the following are c satised for all i, j < : (1) {0, 1} Li L, |Li | < (2) If i < j, then Li Lj . (3) ai and bi are in L and realize the same (quantier-free) type over Li , i.e.: whenever (x) and (x) are unary polynomials with coecients in Li , then (ai ) (ai ) i (bi ) (bi ). (4) bi ai (5) ni is a natural number (6) i is an ni + 1-ary term (7) ci = (c1 , . . . , cni ) Lni i i (8) i (ai , ci ) = 0, i (bi , ci ) = 1. (9) For all i < j we have ai , bi , c1 , . . . , cni Lj . i i The sequences are dened by induction on i < . In stage i we let Li be the set of everything used so far: Li = {0, 1, ak , bk , ck : k < i, nk }. Let i = max(|Li |, 0 ). Since every type over Li can be represented as a set of pairs of polynomials with coecients in Li , there are at most 2i many possible types over Li . By our assumption, i < implies 2i < , so we can nd two dierent elements ai , bi which have the same type over Li . Wlog ai > bi . Now the function that maps the set {x : x ai } to 0 and everything else (including bi ) to 1 is monotone, so it is realized by a polynomial pi (x). Let ni be the number of coecients of pi , so we can write pi (x) as i (x, c1 , . . . cni ), where i (x, y1 , . . . , yni ) is an (ni + 1)-ary term. i i This concludes the construction of our sequences. Note that if (ki : i < ) is the increasing enumeration of a -size subset of , then (Lki : i < ), (aki : i < ), etc., also have all the properties listed above.

688

revision:1999-08-31

modified:1999-09-02

NO OPC LATTICES

Since has uncountable conality, there must be some natural number n such that {i : ni = n} has cardinality , so wlog (after thinning out our sequence, if necessary) we may assume that ni = n for all n. Similarly, since there are only countably many n + 1-ary terms, we may assume all i are equal to some xed term . Now let di = (ai , bi , c1 , . . . , cn ). We claim that (di : i < ) is an i i n+2 antichain in L . Indeed, pick any i < j and assume that either di dj or di dj . Since is monotone in each argument, we either have 0 = (ai , ci ) (aj , ci ) (aj , cj ) = 0 or the converse inequality, so in any case (aj , ci ) = 0. Similarly we get (bj , ci ) = 1 However, since aj and bj have the same type over ci , 0, 1, the equa tion (aj , ci ) = 0 implies (bj , ci ) = 0. This is a contradiction, so we conclude that di and dj are incomparable. 1.3. Conclusion. There is no innite opc lattice Proof. Assume that L is opc. A fortiori, L has the 1-IP. Let = |L|. Since L is opc, we know from [1] that must be a strongly inaccessible cardinal, so in particular is a strong limit cardinal of uncountable conality. By our theorem, there is an antichain A L of cardinality . But this easily implies that there are 2 > many monotone functions from Ln to L, and at most many of them can be polynomials.
modified:1999-09-02

2. The role of AC 2.1. Denition. An innite set A is called strongly amorphous if, for all natural numbers n, all n-ary relations on A are rst order denable (with parameters) in the language of equality. Equivalently, A is amorphous if all sets R An are in the Boolean algebra generated by the sets {(x1 , . . . , xn ) : xi = xj }, {(x1 , . . . , xn ) : xi = a} (a A). While the axiom of choice (in fact, already a very weak version of AC) clearly implies that there are no innite strongly amorphous sets, it is well known that the theory ZF + there is an innite strongly amorphous set is equiconsistent with ZFC. That is, ZF cannot refute the existence of innite strongly amorphous sets. Hence (as we will see below), ZF cannot refute the existence of innite opc lattices. For the rest of this section we promise not to use the axiom of choice. 2.2. Theorem (ZF). For every innite set L there is a bounded lattice (L, , , 0, 1, ) such that: For all natural numbers n, for all monotone functions f : Ln L: If f is denable in (L, , , , 0, 1), then f is induced by a polynomial.

688

revision:1999-08-31

MARTIN GOLDSTERN AND SAHARON SHELAH

2.3. Remark. (1) The ZF above means that this theorem is proved in the usual framework of mathematics (such as given by the Zermelo Frnkel axioms for the underlying set theory) but witha out invoking the axiom of choice. (2) By denable we mean here: as a relation f Ln+1 , f is denable by a rst order formula (with parameters from L) in the language of lattice theory. Since all our lattices will be bounded, it will be convenient to include the constants (or 0-ary operations) 0 and 1 into the language of lattice theory. Thus, denable with parameters c1 , . . . , ck will mean the same as denable with parameters c1 , . . . , ck , 0, 1. 2.4. Corollary (ZF). Assume that there is an innite strongly amorphous set. Then there is an innite order polynomially complete lattice. 2.5. Construction. Let L be an innite set, 0 and 1 two distinct elements of L. Dene a lattice structure on L by requiring 0 a 1 for all a L. 2.6. Fact. Let (L, , , 0, 1, ) be a lattice as in construction 2.5. Then every subset R Ln which is denable in (L, , , 0, 1, ) with parameters c1 , . . . , ck L is also denable (with parameters c1 , . . . , ck , 0, 1) in the language of equality [i.e, in (L, =)]. We will abbreviate a situation as in fact 2.6 by writing R is denable from (c1 , . . . , ck ). Functions f : Ln L will be treated as relations f Ln+1 . 2.7. Notation. We indicate formal variables or indeterminates by a special typeface, e.g., x, t1 , etc. We abbreviate tuples (c1 , . . . , ck ) and (x1 , . . . , xn ) as ck and xn , or sometimes c and x. We abbreviate (, c1 , . . . , ck ) by (, c) or sometimes , c. M on(L, L ) is the set of all monotone maps from L to L . T is the set of all lattice-theoretic terms in the variables x1 , x2 , . . . , s1 , s2 , . . . , t1 , t2 , . . . (We include the constants 0 and 1 among lattice-theoretic terms) 2.8. Denition. Assume that L, L are isomorphic lattices, with isomorphism : L L . We extend canonically to an isomorphism : Ln Ln . For any f : Ln L we write the conjugate function as f or (f ): for all a Ln : (f )() = (f ()) a a 2.9. Fact. If f : Ln L is denable from c = (c1 , . . . , ck ), : L L an automorphism which satises (cj ) = cj for j = 1, . . . , k, then f = f . 2.10. Denition. Let (L, ) be a partial order, A L. The monotone characteristic function of A is the function A : L {0, 1}

688

revision:1999-08-31

modified:1999-09-02

NO OPC LATTICES

dened by A (x) = 1 if a b for some b A 0 otherwise,

i.e., A is the customary characteristic function of the upward closure of A. From now on L will be a lattice as in 2.5 2.11. Denition. Let {c1 , . . . , ck } L, d = (d1 , . . . , dm ) Lm . We say that d is independent over {c1 , . . . , ck } (or: over (c1 , . . . , ck )) i all di are distinct, and no di is in {c1 , . . . , ck , 0, 1}. [This is a special case of the usual model-theoretic notion of independence.] 2.12. Denition. (1) Let (s, x, t1 , t2 , t3 ) be the term [(x s) t1 ] [(x s) t2 ]. (2) For L \ {0, 1} let (x, t1 , t2 , t3 ) = (, x, t1 , t2 , t3 ). c (3) If A {c1 . . . , ck , 0, 1} L, then we dene A (k , x, 3 ) (the s t monotone characteristic function of A, given the parameters c) as follows: c (a) If A = , then A (, x, t1 , t2 , t3 ) is the constant term 0. s c (b) If A = {1}, then A (, x, t1 , t2 , t3 ) = (xt1 , xt2 , xt3 ), s where is the following majority term: (x, y, z) = (y z) (z x) (x y)
modified:1999-09-02

c s (c) If 0 A, then A (, x, t1 , t2 , t3 ) is the constant term 1. c (d) Otherwise we let I = IA = {i : ci A \ {0, 1}} and we let c A (x, t) = iI c (4) If A is conite, L \ A {c1 , . . . , ck }, then we dene A (, x, ) s t similarly (dually), such that fact 2.13 below holds. We leave the details to the reader.

ci (x, t)

revision:1999-08-31

2.13. Fact. If A {c1 , . . . , ck , 0, 1} or L \ A {c1 , . . . , ck , 0, 1}, and d = (d1 , d2 , d3 ) is independent over c1 , . . . , ck , then the function
c a A (, a, d1 , d2 , d3 ) c

is the monotone characteristic function of A. Proof. Easy computation. 2.14. Lemma. Let L be as in construction 2.5. Then there is a function

p:
n=0

M on(L , L)
k=1

Lk

c that assigns to each pair (f, c) a polynomial pf ; T such that the following hold:

688

MARTIN GOLDSTERN AND SAHARON SHELAH

c (1) If c = (c1 , . . . , ck ) and f : Ln L, then pf ; is a term using (at most) the variables s1 , . . . , sk , x1 , . . . , xn , t1 , . . . , t3n . We will c c write pf ; as pf ;(; x; t). s (2) If c and f are as above, and f is denable from c1 , . . . , ck , then we have: Whenever d = (d1 , . . . d3n ) is independent over {c1 , . . . , ck }, c then for all a Ln : f () = pf ;(; a; d), a c c that is, the function f is induced by the lattice polynomial pf ;(; x; d). c f ; c (3) Moreover, the term p depends only on the isomorphism type of (f, c). That is, whenever : L L is an isomorphism, then f ; c c p (, x, t) = pf,() (, x, t). s s

2.15. Remark. Let L be as above. Then the following is not provable in ZF: () There is a map p which assigns to each denable monotone function f : L L a polynomial pf (x) with coecients in L such that pf (a) = f (a) for all a L. This explains why we have to explicitly mention the parameters c in lemma 2.14. Proof of lemma 2.14. We dene the map p by induction on n (the arity of f ). We leave the case n = 0 to the reader. Let f : Ln+1 L, c = (c1 , . . . , ck ) Lk . (Wlog assume that f is c denable from c, otherwise set pf ; = 0.) Let A = L \ {c1 , . . . , ck , 0, 1}. We will now apply the induction hypothesis to the functions f : n L L, where f () = f (, ). For each L we thus get a term a a = pf . We will show that only nitely many dierent terms actually appear, and that they can be combined to yield a term pf . The cases A and A have to be treated in dierent ways. / To save us some cases distinctions, we agree that ck+1 = 0, ck+2 = 1. For each {c1 , . . . , ck+2 } the function f : Ln L, f () = a f (, ) is denable from c = (c1 , . . . , ck ). Let a
f c ;(n ; sk ; t3 x
n+1

modified:1999-09-02

c ) = pf ,(n ; sk ; t3 ) x

revision:1999-08-31

f, (so the indeterminates t3n +1 , . . . , t3n+1 do not appear in c .) By our inductive assumption we know for all a Ln : n+1 c (-1-) f ;(n ; ck ; d3 ) = f () = f (, ) a a a

for all {c1 , . . . , cm+2 }, whenever d is independent over c. Let

m+2
n+1 f ; c constant (n , xn+1 ; sk ; d3 ) x

=
i=1

f c ci (xn+1 , 3 ) ci;(n ; sk ; t3 t x

n+1

) .

Hence we have for all a Ln : (-2-) f (, ) a n+1 constant (n , ; ck ; d3 ) = a f (, 0) a

if {c1 , . . . , ck , 0, 1} if A = L \ {c1 , . . . , ck , 0, 1}.

688

NO OPC LATTICES

whenever d is independent over c. This follows easily from equation (-1-) and the monotonicity of f . For each A the function f : Ln L, f (n ) = f (n , ) is dena a n f ;, n c able from (, c). Hence the polynomial p ( ; xn+1 , sk ; t3 ) satises x for all a Ln c (-3-) pf ;,(; , ck ; d) = f (, ) a a for all A, whenever d is independent over , c. f ;, c We claim that the function p (; xn+1 , s; t) is actually conx stant on A. The reason is that A is disjoint to the set of parameters from which f is dened.
c c More formally, let , A. We will show pf ;, = pf ;,. Let : L L be an automorphism that xes {c 1 , . . . , ck , 0, 1} pointwise, and maps to . Then f = f , by fact 2.9. Hence c c pf ;, = p(f );(),() c = p(f )() ;, c = pf ;,

by induction hypothesis, lemma 2.14(3) since (f ) = (f )() since f = f

f ; c We will write raw (, xn+1 ; s; t) for the common value of this function. x So we can rewrite equation (-3-) as: for all a Ln , A: c (-4-) f ; (, ; c; d) = f (, ) a a raw

whenever d is independent over c. The restriction that d has to be independent not only of c but also of is inconvenient. We get rid of it with the following error correction device: Let f c f ; c 1 ;(n , xn+1 ; s; t1 , . . . , t3n+1 ) = raw (n , xn+1 ; s; t1 , . . . , t3n ) x x f ; n c f ; n c 2 ( , xn+1 ; s; t1 , . . . , t3n+1 ) = raw ( , xn+1 ; s; t3n +1 , . . . , t23n ) x x f ; n c f ; n c 3 ( , xn+1 ; s; t1 , . . . , t3n+1 ) = raw ( , xn+1 ; s; t23n +1 , . . . , t33n ) x x and let f ; c f c f c f c clean = (1 ;, 2 ;, 3 ;), where (x, y, z) is a majority term, i.e., satises (-5-) (a, a, b) = (a, b, a) = (b, a, a) = a for all a, b L, see denition 2.12. Let
f ; c f ; c f c c bounded (n , x; s; t) = clean (n , x; s; t) 1 ;(n ; s; t) A (x, t3 ). x x x f ; c We claim that bounded (n , xn+1 ; sk ; t3 ) satises x f (n , ) A a n+1 f ; c n k 3 )= 0 (-6-) bounded ( , ; c ; d a {c1 , . . . , ck , 0} f (, 1) = 1 a
n+1

revision:1999-08-31

modified:1999-09-02

for any a Ln , L, whenever d is independent over c.

688

MARTIN GOLDSTERN AND SAHARON SHELAH

f c This is clear if {c1 , . . . , ck , 0}. For = 1 recall that 1 ;(; c; d) = a f (, 1). a Finally, let A, and assume that d is independent over c. Let d1 = (d1 , . . . , d3n ) d2 = (d3n +1 , . . . , d23n ) d3 = (d23n +1 , . . . , d33n )

Now since all the dj are distinct, at least two among d1 , d2 , d3 do not contain and hence are independent over {c1 , . . . , ck , }. So by equation (-4-), at least two of the equations n+1 c f () = f ;(, ; c; d3 ) a a = 1, 2, 3 are true, hence we have (by the property (-5-)) c f ; (, ; c; d) = f () a a
clean

so = f () 1 (; c; d) = f (, ) f (, 1) = f (, ) a a a a a for all A. f ; c This concludes our discussion of the term bounded . c We can now dene pf ; as
c pf ;(n , xn+1 ; sk ; t3 x
n+1

f ; c bounded (, ; c; d) a

f ; c f ; c ) = constant (n , xn+1 ; s; t) bounded (n , xn+1 ; s; t) x x

From (-2-) and (-6-) we now get the desired property c pf ;(, ; c; d) = f (, ) a a whenever d is independent over c. Theorem 2.2 now immediately follows from lemma 2.14.
modified:1999-09-02

References
[1] Martin Goldstern and Saharon Shelah. Order-polynomially complete lattices must be LARGE. Algebra Universalis, 39:197209, 1998. [2] Miroslav Haviar and Miroslav Ploica. On order-polynomial completeness of sc lattices. Algebra Universalis, 39, 1998. [3] Hans Kaiser. Interpolation and order polynomially complete lattices. volume 9 of Contributions to general algebra. Hlder-Pichler-Tempsky, Wien, 1995. o [4] Dietmar Schweigert. Zur Theorie der Verbandspolynome. PhD thesis, Technische Universitt Wien, 1972. a Institut fur Algebra, Technische Universitat Wien, Wiedner Hauptstrae 8-10/118.2, A-1040 Wien, Austria E-mail address: Martin.Goldstern@tuwien.ac.at Department of Mathematics, Hebrew University of Jerusalem, Givat Ram, 91904 Jerusalem, Israel E-mail address: shelah@math.huji.ac.il

688

revision:1999-08-31