ilijas farah and saharon shelah- a dichotomy for the number of ultrapowers

8/3/2019 Ilijas Farah and Saharon Shelah- A Dichotomy for the Number of Ultrapowers

1/35

954

re

vision:2010-11-24

modified:2010-11-24

A DICHOTOMY FOR THE NUMBER OF ULTRAPOWERS

ILIJAS FARAH AND SAHARON SHELAH

Abstract. We prove a strong dichotomy for the number of ultrapow-ers of a given model of cardinality 20 associated with nonprincipal

ultrafilters on N. They are either all isomorphic, or else there are 220

many nonisomorphic ultrapowers. We prove the analogous result formetric structures, including C*-algebras and II1 factors, as well as theirrelative commutants and include several applications. We also show thatthe C*-algebra B(H) always has nonisomorphic relative commutants inits ultrapowers associated with nonprincipal ultrafilters on N.

1. Introduction

In the following all ultrafilters are nonprincipal ultrafilters on N. In par-ticular, all ultrapowers of A always stands for all ultrapowers associatedwith nonprincipal ultrafilters on N.

The question of counting the number of nonisomorphic models of a giventheory in a given cardinality was one of the main driving forces behindthe development of Model Theory (see Morleys Theorem and [19]). On the

other hand, the question of counting the number of nonisomorphic ultrapow-ers of a given model has received more attention from functional analyststhan from logicians.

Consider a countable structure A in a countable signature. By a classicalresult of Keisler, every ultrapower

U

A is countably saturated (recall thatUis assumed to be a nonprincipal ultrafilter on N). This implies that theultrapowers of A are not easy to distinguish. Moreover, if the ContinuumHypothesis holds then they are all saturated and therefore isomorphic (thisfact will not be used in the present paper; see [5]).

Therefore the question of counting nonisomorphic ultrapowers of a givencountable structure is nontrivial only when the Continuum Hypothesis fails,and in the remaining part of this introduction we assume that it does fail. If

Date: November 24, 2010.1991 Mathematics Subject Classification. Primary: 03C20. Secondary: 46M07.The first author was partially supported by NSERC and he would like to thank Takeshi

Katsura for several useful remarks.The second author would like to thank the Israel Science Foundation for partial support

of this research (Grant no. 710/07). Part of this work was done when the authors visitedthe Mittag-Leffler Institute. No. 954 on Shelahs list of publications.

We would like to than Gabor Sagi and the anonymous referee for a number of usefulcomments on the first version of this paper.

1


2/35

954

re

vision:2010-11-24

modified:2010-11-24

2 ILIJAS FARAH AND SAHARON SHELAH

we moreover assume that the theory ofA is unstable (or equivalently, that ithas the order propertysee the beginning of 3) then A has nonisomorphic

ultrapowers ([19, Theorem VI.3] and independently [6]). The converse, thatif the theory of A is stable then all of its ultrapowers are isomorphic, wasproved only recently ([10]) although main components of the proof werepresent in [19] and the result was essentially known to the second author.

The question of the isomorphism of ultrapowers was first asked by oper-ator algebraists. This is not so surprising in the light of the fact that theultrapower construction is an indispensable tool in Functional Analysis andin particular in Operator Algebras. The ultrapower construction for Banachspaces, C*-algebras, or II1 factors is again an honest metric structure of thesame type. These constructions coincide with the ultrapower constructionfor metric structures as defined in [2] (see also [10]). The DowShelah resultcan be used to prove that C*-algebras and II

1factors have nonisomorphic

ultrapowers ([14] and [9], respectively), and with some extra effort this con-clusion can be extended to the relative commutants of separable C*-algebrasand II1 factors in their utrapowers ([8] and [9, Theorem 5.1], respectively).

However, the methods used in [14], [8] and [9] provide only as manynonisomorphic ultrapowers as there are uncountable cardinals c = 20

(with our assumption, two). In [15, 3] it was proved (still assuming onlythat CH fails) that (N,


3/35

954

re

vision:2010-11-24

modified:2010-11-24

A DICHOTOMY FOR THE NUMBER OF ULTRAPOWERS 3

Theorem 3. Assume CH fails. If A is a metric structure of character den-sity c such that the theory ofA is unstable, then there are 2c isometry types

of models of the formUA, where U ranges over nonprincipal ultrafilters

onN.

The proof is a modification of the proof of Theorem 1 and it will beoutlined in 6. Although Theorem 3 implies Theorem 1, we chose to presentthe proof of Theorem 1 separately because it is the main case and becausesome of the main ideas are more transparent in the discrete case.

Corollary 4. For a metric structure A of character density c with acountable signature either all of its ultrapowers are isomorphic or there are 2c

isomorphism types of its ultrapowers.

Proof. We may assume A is infinite. If the theory of A is stable, then U Ais saturated and of character density c and therefore all such ultrapowers areisomorphic ([10, Theorem 5.6]). If the Continuum Hypothesis holds then allultrapowers are isomorphic by the analogue of Keislers theorem for metricstructures ([2]). In the remaining case, when the Continuum Hypothesisfails and the theory of A is unstable use Theorem 3.

Important instances of the ultraproduct construction for metric spacesinclude C*-algebras, II1 factors (see e.g., [10, 2.3.1 and 2.3.2]) and metricgroups (see [18]).

Organization of the paper. The proof of Theorem 1 uses ideas from [19,VI.3], [15, 3] and [21, 3] and it will be presented in 2, 3, 4 and 5.Theorem 3 is proved in 6, and some applications will be given in 8. In 7we prove local versions of Theorem 1 and Theorem 3, and in Proposition 8.5we use the latter to prove that B(H) always has nonisomorphic relativecommutants in its ultrapowers associated with nonprincipal ultrafilters on N.Sections 2 and 3 are essentially a revision of [19, 3], and 4 has a small,albeit nonempty intersection with [15, 3] (and therefore with the latter halfof [19, VI.3]).

Notation and terminology. If A denotes a model, then its universe isalso denoted by A and the cardinality of its universe (or any other set A) isdenoted by |A|. Hence what we denote by A is denoted by A or by |A| in

[19] and [21], and what we denote by |A| is denoted by ||A|| in [19] and [21]if A is a model. We also dont distinguish the notation for a formula (x)and its evaluation [a] in a model. It will always be clear from the context.

Letters I and J, possibly with subscripts or superscripts, will alwaysdenote linear (i.e., total) orders. The reverse of a linear order I will bedenoted by I. The cofinality of a linear order I, cf(I), is the mininalcardinality of a cofinal subset of I. By I + J we denote the order withdomain I J in which copies of I and J are taken with the original orderingand i < j for all i I and all j J. If J and Ij, for j J, are linear


4/35

954

re

vision:2010-11-24

modified:2010-11-24


orders then

jJ Ij denotes the order with the underlying set

jJ{j} Ij

ordered lexicographically.

Following the notation common in Model Theory, an ultrapower of Aassociated with an ultrafilter U will be denoted by

U

A, even in the case

when A is an operator algebra, where the notation AU for the ultrapower isstandard. We refrain from using the symbol in order to avoid confusion.

By m we denote the quantifier for all large enough m N. Moregenerally, if D is a filter on N then by (Dn) we denote the quantifier as ashortcut for the set of all n such that. . . belongs to D.

An n-tuple of elements of A is always denoted by a.For k 1 by [X]k we denote the set of all k-element subsets of X.A cardinal will be identified with the least ordinal of cardinality ,

as well as the linear order (,


5/35

954

re

vision:2010-11-24

modified:2010-11-24


Let D(, 1) be the filter on dual to the ideal generated by the nonsta-tionary ideal and the set { < : cf() 0}. Define f: Card by

f() =

, if 10 if 0.

If the set { : f() = 0} belongs to D(, 1) then let inv1(I) be the equiva-

lence class of f modulo D(, 1), or in symbols

inv1(I) = f /D(, 1).

Otherwise, inv1(I) is undefined.Assume m 1 and invm(J) is defined for all linear orders J (allowing

the very definition of invm(J) to be undefined). Assume I and I, for < = cf(I), are as in the case m = 1. Define a function gm with

domain viagm() = inv

m((I \ I)).

If{ : gm() is defined} belongs to D(, 1) then let invm+1(I) be the equiv-

alence class of gm modulo D(, 1). Otherwise invm+1(I) is undefined.

This defines invm(I) for all I. For a (defined) invariant d we shall writecf(d) for cf(I), where I is any linear order with invm(I) = d. We also write

|d| = min{|I| : d = invm(I) for some m}.

Our invariant invm(I) essentially corresponds to invm1(I) as defined in [21,Definition 3.4]. Although inv can be recursively defined for every ordinal ,we do not have applications for this general notion. As a matter of fact, onlyinvm for m 3 will be used in the present paper.

Example 2.1. Assume throughout this example that is a cardinal withcf() 1.

(1) Then inv0() = cf() and inv1() is undefined.(2) If is a cardinal with cf() 1 then inv

0( ) = cf() andinv1( ) is the equivalence class of the function on cf() everywhereequal to cf(), modulo D(cf(), 1).

(3) If invm(I) is defined for all < and is regular then with I =


6/35

954

re

vision:2010-11-24

modified:2010-11-24


Proof. This is cases (13) of [21, Lemma 3.8], with = 1 but we reproducethe proof for the convenience of the reader.

(1) If 2 is regular, then the set { < : cf() 1} can be parti-tioned into disjoint stationary sets (see [19, Appendix, Theorem 1.3(2)] or[16, Corollary 6.12]). Denote these sets by S, for < . For Z definea linear order LZ as follows. For < let

() =

1 if

Z S

2 if

/ZS

1 if cf() 0.

Let LZ =

2. Then inv

1(LZ) is clearly defined. A stan-

dard argument using the stationarity of S for any ZY shows thatinv1(LZ) = inv

1(LY) if Z = Y.(2) Now assume is singular. Pick an increasing sequence of regular

cardinals i, for i < cf(), such that

i


7/35

954

re

vision:2010-11-24

modified:2010-11-24


2.2.3. Assume is singular and 1 cf(). This case will require extrawork. Like above, fix an increasing continuous sequence of cardinals , for

< cf(), such that = sup h()} Dh(2) then let

invm,(J) = d : < 2/Dh(2)

and undefined otherwise.

Since Dh(2) extends D(2, 1), this invariant is well-defined.Definition 2.4. Given a cardinal 2 and m N, an m, -invariant isany invariant invm,(J) for a linear order J of cardinality that is not equalto undefined.

Two representing sequences d : < and e : < of invariants ofthe same cofinality are disjoint if d = e for all . Note that this is not aproperty of the invariants since it depends on the choice of the representingsequences.

Lemma 2.5. For every cardinal 2 there exist m N and 2 disjoint

representing sequences of m, -invariants of linear orders of cardinality .

Proof. Assume first is regular. By Lemma 2.2 there are 2 linear ordersof cardinality andwith cofinality equal to , listed as I for < 2

, withdistinct (and defined) invariants inv1(I). Let I = J

. Then |I| = ,

inv2,(I) is defined since cf(I) = for all and it has constant representingsequence. Therefore all these representing sequences are disjoint.

Now assume is singular. By Lemma 2.2 for every regular < thereare 2 linear orders, J, , for < 2

, of cardinality , cofinality , and with

distinct and defined invariants inv2(J,).


8/35

954

re

vision:2010-11-24

modified:2010-11-24


(a) Assume furthermore that cf() 2. Fix an increasing continuoussequence , for < cf() with the supremum equal to , as in 2.2.2. Now

fix an increasing sequence , for < cf(), of regular cardinals with thesupremum equal to and such that > for all . For < 2

let

I =


9/35

954

re

vision:2010-11-24

modified:2010-11-24


3.2. Combinatorics of the invariants. The following is a special case ofthe definition of weakly (, )-skeleton like where is an arbitrary cardinal

and is set of formulas as given in [21, Definition 3.1]. Readers familiarwith [21] may want to know that we fix = 1 and = {, } where(x, y) stands for (y, x).

Definition 3.1. A -chain C is weakly (1, )-skeleton like inside A if forevery a An there is a countable Ca C such that for all b c in C with[b, c] disjoint from Ca we have

A |= (b, a) (c, a)

and

A |= (a, b) (a, c).

Remark 3.2. One can weaken the definition of weakly (1, )-skeleton likeby requiring only that (with a, Ca, b and c as in Definition 3.1)

a b if and only if a c

and

b a if and only if c a.

All the statements about the notion of being weakly (1, )-skeleton like,except Lemma 3.7, remain true for the modified notion. As a matter of fact,it is transparent that even their proofs remain unchanged.

Remark 3.3. For a Ak and b An define

tp(a/b) = {(x, b) : is a k + n-ary formula and A |= (a, b)}.

One may now consider a stronger indiscernibility requirement on a -chain Cthan being weakly (1, )-skeleton like, defined as follows.

(*) For every k N and a Ak there is a countable Ca C such thatfor all b c in C with [b, c] Ca = we have that

tp(a/b) = tp(a/c).

The proofs of Theorem 1 and Theorem 3 can be easily modified to providean ultrafilter U such that for a given linear order I the ultrapower

U

Aincludes a -chain C isomorphic to I and satisfying (*). See Remark 4.5 and

Remark 6.9.

The nontrivial part of the following is a special case of [21, Claim 3.15]that will be needed in 3.3.

Lemma 3.4. Assume C is a -chain that is weakly(1, )-skeleton like inA.Then C is weakly (1, )-skeleton like inside A, and every interval of C isweakly (1, )-skeleton like inside A. If E C is well-ordered (or converselywell-ordered) by then E is weakly (1, )-skeleton like in A.


10/35

954

re

vision:2010-11-24

modified:2010-11-24


Proof. Only the last sentence requires a proof. For b An define Eb E asfollows.

Eb = {min(E [c, )) : c C b}.

Each Eb is countable since every c Cb produces at most one element of Eb.For a c in E such that [a, c] Eb = we have that [a, c] Cb = andtherefore tp(a/b) = tp(c/b).

If C and E are -chains in A then we say C and E are mutually cofinalif for every a C we have a b for all large enough b E and for everyb E we have b a for all large enough a C .

Lemma 3.5. Assume C and E are mutually cofinal -chains in A. Thencf(C) = cf(E).

Of course this is standard but since is not assumed to be a partialordering on A we shall prove it. Also note that if the condition for everya C we have a b for all large enough b E is replaced by for everya C we have a b for some b E and the condition for every b Ewe have b a for all large enough a C is replaced by is replaced byfor every b E we have b a for some a C then we cannot concludecf(C) = cf(E) in general.

Proof of Lemma 3.5. Assume = cf(C) < cf(E) = and fix a cofinal X Cof cardinality . For each a X pick f(a) E such that a b for all bsuch that f(a) b. The set {f(a) : a X} is not cofinal in E and we can

pick b E such that f(a) b for all a X. Now let a C be such thatfor all c C such that a c we have b c. But there is c X is suchthat a c, and this is a contradiction.

The following is [21, Lemma 3.7] in the case = 1. We reproduce theproof for the convenience of the reader.

Lemma 3.6. Assume C0, C1 are increasing, weakly (1, )-skeleton like,-chains in A. Also assume these two chains are mutually cofinal and mis such that both invm(C0) and inv

m(C1) are defined. Then invm(C0) =

invm(C1).

Proof. The proof is by induction on m. If m = 0 then this is Lemma 3.5.Now assume the assertion has been proved for m and all pairs C0 and C1.Fix C0, C1 satisfying the assumptions for m + 1 in place of m and let =cf(C0) = cf(C1). Since inv

m(C0) is defined, 1. Since invm+1(C0) is

defined, D(, 1) is a proper filter and 2.For an elementary sumbodel N of (A, C0, C1) consider

C0N = {b C0 : A |= c b for all c Nn C0}, and

C1N = {b C1 : A |= c b for all c Nn C1}.


11/35

954

re

vision:2010-11-24

modified:2010-11-24


By our assumption that invm+1(C0) and invm+1(C1) are defined we have

that for any regular < the set of N (A, C0, C1) of cardinality such

that cf(Nn C0) 1 implies invm(C0N, ) is defined includes a club.In particular, for club many N of size such that cf(Nn C0) 1 wehave cf(C0N, ) 1. Similarly, for club many N of size such thatcf(Nn C1) 1 we have that inv

m(C1N) is defined and cf(C1N, ) 1.

Now pick N A such that cf(Nn C0), cf(Nn C1), cf(C

0N, ) and

cf(C1N, ) are all uncountable and invm(C0N, ) and inv

m(C1N, ) aredefined. We shall prove that in this case (C0N, ) and (C

1N, ) are mu-

tually cofinal.By the elementarity Nn C0 and N

n C1 satisfy the assumptions ofLemma 3.5, and in particular cf(Nn C0) = cf(N

n C1). Pick a C0N.

Since Nn C1 and Nn C0 are mutually cofinal, by elementarity for all

c Nn

C1 we have that c a.Let Ea C1 be a countable set such that for all b and c in C1 satisfying b c and [b, c] Ea = we have that A |= (b, a) (c, a) and A |= (a, b) (a, c). Since Ea is countable, by our assumptions on the cofinalities ofNn C1 and (C

1N, ) for cofinally many c N

n C1 and for -cofinally many d C1N we have

A |= c a d a.

Therefore for -cofinally many d C1N we have d a, i.e., a d.

An analogous proof shows that for every e C1 and -cofinally manyd C0 we have e d. We have therefore proved that the -chains

(C

0

N, ) and (C

1

N, ) are mutually cofinal. They are both obviouslyweakly (1, )-skeleton like, and by the inductive hypothesis in this case wehave invm(C0N, ) = inv

m(C1N, ) if both of these invariants are defined.By the inductive hypothesis we have invm+1(C0) = inv

m+1(C1).

3.3. Defining an invariant over a submodel. Assume Z is an elemen-tary submodel of A. By tp(a/Z) we denote the -type of a A

n in thesignature {} over Z, or in symbols

tp(a/Z) = {(x, b) : b Z, A |= (a, b)} {(b, x) : b Z, A |= (b, a)}.

IfB A (in particular, ifB is an elementary submodel of A) we shall writetp(a/B) for tp(a/B

n). Write tp(a/e) for tp(a/{e}).

Lemma 3.7. A -chain C in A is weakly (1, )-skeleton like in A if andonly if for every a An there exists a countable Ca C with the propertythat for c and d in C the condition

Ca (, c] = Ca (, d]

implies tp(a/c) = tp(a/d).

Proof. Immediate from Definition 3.1.


12/35

954

re

vision:2010-11-24

modified:2010-11-24


Definition 3.8. Assume B is an elementary submodel of A, m N, and dis an m-invariant. We say that c An \ Bn defines an (A,B,,m)-invariant

d if there are(1) (nonempty) linear orders J and I, and(2) aj B

n for j J and ai An \ Bn for i I,

such that

(3) ai : i J+ I is a -chain in A that is weakly (1, )-skeleton like

in A,(4) tp(ai/B) = tp(c/B) for all i I,(5) d = invm(I), and(6) if J, I, ai for i J

I and d satisfy conditions (1)(5) theninvm(d) = invm(d).

Let INVm(A,B,) denote the set of all m-invariants d such that some c

defines an (A,B,,m)-invariant d.

The point of Definition 3.8 and the conclusion of the Lemma 3.9 is that,once A, , and m are fixed, the invariant d depends only on the submodel Band the element c outside of this submodel, and not on the -chain C.Conditions (1)(5) of Definition 3.8 imply (6) of Definition 3.8. This isa consequence of Lemma 3.10 and the fact that cofinalities occurring ininvariants that are defined in the sense of 2.1 or 2.2 are uncountable.

The following notation will be useful. Assume C is a -chain that isweakly (1, )-skeleton like in A and B is an elementary submodel of A.For c C \ Bn let

C[B, c] = {a C : (b B

n

C)c b a

b}.

We shall always consider C[B, c] with respect to the reverse or-der, .

Lemma 3.9. Assume C = ai : i I is a -chain that is weakly (1, )-skeleton like inA. Assume B is an elementary submodel of A and c C \Bn

are such that

(1) Cb C[B, c] (, c] = for all b Bn, and

(2) d = invm(C[B, c], ) is well-defined.

Then c defines the (A,B,,m)-invariant d.

Proof. Let J0 be a well-ordered -cofinal subset of

{i I: ai Bn and ai c}

of minimal order type. By Lemma 3.4 the -chain ai : i J0 is weakly(1, )-skeleton like in A. Let I0 = {i I: ai C[B, c]}. We need to checkthat I0, J0 and ai : i J0 + I

0 satisfy (1)(6) of Definition 3.8.

Clauses (1)(2) are immediate. As an interval of a weakly (1, )-skeletonlike order, ai : i I0 is weakly (1, )-skeleton like. Therefore clauses (3)follows. In order to prove (4) pick b Bn and d C[B, c] (, c]. Then[d, c] Cb = , hence tp(c/b) = tp(d/b). Sine b B

n was arbitrary, we


13/35

954

re

vision:2010-11-24

modified:2010-11-24


have tp(c/B) = tp(d/B) and we have proved (4). Clause (5) is automatic,and (6) follows by Lemma 3.10 below.

Lemma 3.10. Assume I0, I1, J0, J1 are linear orders and ai : i J0 + I0

and bi : i J1 + I1 are weakly (1, )-skeleton like -chains in A such that

(3) ai Bn if and only if i J0 and bi Bn if and only if i J1,(4) tp(ai/B) = tp(bj/B) for all i I0 and all j I1,(5) each of cf(I0), cf(I1), cf(J0), and cf(J1) is uncountable.

If invm(I0) and invm(I1) are both defined then inv

m(I0) = invm(I1).

Proof. Pick i(0) I0. Since tp(ai(0)/B) = tp(bj/B) for some (any) j I1,

we have that bi ai(0) for all i J1. Since cf(J1) and cf(I1) are bothuncountable and since bi : i J1 + I

1 is weakly (1, )-skeleton like, we

conclude that for large enough i I1 we have ai(0) bi.

The analogous argument shows that for every i(1) I1 and all largeenough i I0 we have ai(1) bi. Then ai : i I0 and bi : i I1 are,when ordered by , mutually cofinal.

By Lemma 3.6 we have that invm(I0) = invm(I1) if both of these invari-

ants are defined, and the claim follows.

3.4. Representing invariants. In addition to A, and m fixed in 3.1we distinguish = |A|. A representation of A is a continuous chain ofelementary submodels A, for < cf(), of A such that |A| < |A| for all and

1. Then d is an m, -invariant ofA, if d is an m, -invariant and for every representation A, < cf() ofA we have

{ : d INVm(A, A, )} D(cf(), 1).

3.4.3. Assume is singular and 1 cf(). Fix h : 2 cf() as inLemma 2.3. Then d is an m, -invariant of A, if d is an m, -invariantand for every representation A =

|A|} Dh(2).

Lemma 3.11. Assume A,,m and = |A| are as above. Also assumeC = aj : j J is a -chain in A that is weakly (1, )-skeleton like in A.

If invm,(J) is defined then invm,(J) INVm,(A).


14/35

954

re

vision:2010-11-24

modified:2010-11-24


Proof. This is really three lemmas wrapped up in one. We prove each ofthe three cases, depending on the cofinality of (3.4.1, 3.4.2 and 3.4.3)

separately.

3.4.4. Assume is regular. Fix a representation A, < , ofA. Let C be the club consisting of all such that for every a An we have Ca A

n .

By the assumption cf(J) = we may clearly assume m 1. Let

d = d : < /D(, 1).

Fix C such that cf() = cf(C An ) 1 and d is defined. Since

cf(J) = by 2.2.1 the set of such belongs to D(, 1). It will thereforesuffice to show that for every such some c defines the (A, A, , m 1)-invariant d.

Pick c C such that (, c] An

C An

. Let I

be the order withthe underlying set {i J: ai C [A, c]}, so that inv

m1(I) = d. Then

cf(C An ) = cf() 1

and

cf(C[A , c], ) = cf(d) 1.

Since a An implies Ca An , Lemma 3.9 implies that c defines the

(A, A, , m 1)-invariant d.

3.4.5. Assume is singular and 1 < cf(). Fix a representation A, j(0)} (An

{Ca : a An }) = .

Let c = aj(0). Then

cf(An C (, c], ) = cf() 1


15/35

954

re

vision:2010-11-24

modified:2010-11-24


andcf(C[A, c]) = cf(d) 1.

By Lemma 3.9 we have that c defines the (A, A(0), , m 1)-invariant d.

3.4.6. Assume is singular and cf() 1. Fix a representation A, |A(0)|, we canpick j(0) J such that

{ai : i J, i > j(0)} (An(0)

{Ca : a A

n(0)}) = .

Let c = aj(0). Then

cf(An(0) C (, c], ) = cf() 1

andcf(C[A(0), c], ) = cf(d) 1.

By Lemma 3.9 we have that c defines the (A, A(0), , m 1)-invariant d.This exhausts the cases and concludes the proof of Lemma.

3.5. Counting the number of invariants of a model. We would like toprove the inequality | INVm,(A, )| |A| for every model A of cardinality 2. Instead we prove a sufficiently strong approximation to this inequaity.As a courtesy to the reader we start by isolating the following triviality.

Lemma 3.12. For every cardinal and every X P() of cardinality > there is < such that |{x X: x}| > .

Proof. We may assume |X | = + and enumerate X as {x : < +}. If the

conclusion of lemma fails then f() = sup{ < + : x} defines a cofinalfunction from to +.

See the paragraph before Lemma 2.5 for the definition of disjoint repre-senting sequences.

Lemma 3.13. For A,,m as usual and = |A| every set of disjoint repre-senting sequences of invariants in INVm,(A, ) has size at most .

Proof. Let us prove the case when is regular. We may assume m 1since the case m = 0 is trivial. Assume the contrary and let d(), for < +, be disjoint representing sequences of elements of INVm,(A, ). Letd() = d() : < /D(, 1). Fix a representation A, for < , of A.


16/35

954

re

vision:2010-11-24

modified:2010-11-24


For each < + fix S D(, 1) such that for every S some cdefines an (A, A , , m)-invariant d() . By Lemma 3.12 there is <

such that + distinct (A, A , , m)-invariants are defined by elements ofAn.Since |A| = , this is impossible.

The proofs of the two cases when is singular are almost identical to theabove proof and are therefore omitted.

The following application, proved in [21], is not concerned with ultraprod-ucts.

Proposition 3.14. Assume 2 andK is a class of models of cardinal-ity. If there are n and a2n-ary formula such that for every linear orderIof cardinality there exists a model A K such that I is isomorphic to aweakly (1, )-skeleton like -chain in A

n, then there are 2 nonisomorphic

models inK.Proof. Let I be a linear order and let A be a model such that I is isomorphicto a weakly (1, )-skeleton like -chain in A. By Lemma 3.11, inv

m,(I) INVm,(A) and by Lemma 3.13, INVm(A) has cardinality at most forevery A K. By Lemma 2.5 there are 2 disjoint representing sequences ofm, -invariants of linear orders of cardinality . By the pigeonhole principlethere are 2 nonisomorphic elements ofK.

4. Construction of ultrafilters

The main result of this section is Proposition 4.2 below. Its version in

which Mi = (N,


17/35

954

re

vision:2010-11-24

modified:2010-11-24


For h and h in FI(G) say that h and h are incompatible if h h is nota function. Note that if G is independent mod D then h h if and only if

Ah Ah =D and Ah D Ah whenever h h, for all h and h in FI(G).A standard -system argument (see [19] or [16]) shows that every family

of pairwise incompatible elements of FI(G) is countable. Lemma 4.1 below isa special case of [19, Claim VI.3.17(5)]. We include its proof for convenienceof the reader.

Lemma 4.1. Assume D is a filter on and G N is a family of functionsindependent mod D. Furthermore, assume D is a maximal filter such that Gis independent mod D. Then for every X there is a countable subsetA FI(G) such that

(1) For every h A either Ah D X or Ah X =

D .(2) For every h FI(G) there is h A such that Ah Ah =

D .

Proof. Assume for a moment that for every h FI(G) there is h h suchthat

(*) Ah D X or Ah X =

D .

Let A be a maximal family of incompatible elements of FI(G) such that (*)holds. Then A is countable and it satisfies (1). By our assumption and themaximality of A, it satisfies (2) as well.

Now assume there is h such that for every h h in FI(G) we have bothAh \ X =

D and Ah X =D . Let D be the filter generated by D and

the complement of X Ah. Since Ah X =D , for h, we have that D is a

proper extension ofD. Since Ah \ X =D for every h extending h, we have

that G is independent modulo D

. This contradicts the assumed maximalityof D.

Lemma 4.1 implies that for every X N there is a countable G0 Gsuch that A satisfying the above conditions is included in FIs(G0). In thissituation we say X is supported by G0. The question of uniqueness of asupport G0 for given X is irrelevant for us and it will be ignored.

Proposition 4.2. Assume (x, y) is a formula and Mi, for i N, aremodels of the same signature such that in Mi there is a-chain of lengthi.Then for every linear order I of cardinality c there exists an ultrafilter Uon N such that

U

Mn includes a weakly (1, )-skeleton like -chain Cisomorphic to I.

Proof. In order to simplify the notation and release the bound variable nwe shall assume that is a binary formula and hence the elements of the-chain C will be elements of A instead of n-tuples of elements from A. Letai(n), for 0 i < n, be a -chain in Mn. For convenience of notation, wemay assume

ai(n) = i

for all i and n, and we also write ai(n) = n 1 if i n. Fix an in-dependent family G of size c of functions f: N N (see [19, Appendix,


18/35

954

re

vision:2010-11-24

modified:2010-11-24


Theorem 1.5(1)]). Fix a filter D on N such that G is independent withrespect to D and D is a maximal (under the inclusion) filter with this prop-

erty. Let FI(G), Ah for h FI(G) and FIs(G) be as introduced beforeLemma 4.1. The following is an immediate consequence of Lemma 4.1 (i.e.,of [19, Claim VI.3.17(5)]).

Claim 4.3. For every g

nN Mn there is a countable set Sg I suchthat for all l N both sets

Xg,l = {n : Mn |= (al(n), g(n))}

Yg,l = {n : Mn |= (g(n), al(n))}

are supported by {fi : i Sg}.

Proof. Apply Lemma 4.1 to each Xg,l and each Yg,l and take union of the

supports. Fix an enumeration of G by elements of I and write G = {fi : i I}. For

i < j in I write [i, j]I for the interval {k I: i k j}. For elementsa b in a model M write

[a, b] = {c M: a c and c b}.

Since is not necessarily transitive, this notation should be taken with agrain of salt. For i < j in I write

Bij = {n : fi(n) fj(n)}.

Note that by our convention about ai(n) we have that fi(n) fj(n) isequivalent to fi(n) fj(n). For g nN Mn and i < j in I such that[i, j] Sg = let

Cgij = {n : Mn |= (fi(n), g(n)) (fj(n), g(n))

and Mn |= (g(n), fi(n)) (g(n), fj(n))}.

In other words, Cgij = {n : tp(fi(n)/g(n)) = tp(fj(n)/g(n))}, with tpas computed in Mn.

Claim 4.4. The family of all sets Bij for i < j in I and Cgij for g nN Mn and i < j in I such that [i, j]I Sg = has the finite intersection

property.

Proof. It will suffice to show that for k N, i(0) < < i(k 1) in I, and

g(0), . . . , g(k 1) in

nN Mn the setl


19/35

954

re

vision:2010-11-24

modified:2010-11-24


Write T = {i(k) : k < k}, also TG = {fi : i T } and SG = {fi : i S}.

Pick hm, for m N, in FI(SG \ TG) so that

(1) hm hm+1 for all m and(2) For all h FI(TG), all l N and all k < k, for all but finitely many

m we have either(iX) (

Dn Ahmh)Mn |= (al(n), g(k)(n)), or(iiX) (Dn Ahmh)Mn |= (al(n), g(k)(n))

and also either(iY) (

Dn Ahmh)Mn |= (g(k)(n), al(n)), or(iiY) (

Dn Ahmh)Mn |= (g(k)(n), al(n)).

The construction ofhm proceeds recursively as follows. Enumerate all triples(h,k,l) in FI(TG) k N by elements ofN. Let h0 = . If hm has beenchosen and (h,k,l) is the m-th triple then use the fact that Xg(k),l and

Yg(k),l are supported by S (Claim 4.3) to find hm+1 FI(SG \ TG) such thatAhm+1h satisfies one of (iX) and (iiX) and one of (iY) or (iiY). Then thesequence of hm constructed as above clearly satisfies the requirements.

In order to complete the proof we need to show that there exist h FI(TG)and n such that

(1) Ahnh D

l


20/35

954

re

vision:2010-11-24

modified:2010-11-24


Pick m large enough to decide the k-type of each p [h(i(l)), h(i(m))]Z. Since there are only four different k-types, by the pigeonhole prin-

ciple there are N elements of [h(i(l)), h(i(m))] Z with the same k-type. There is therefore t [Z]k such that ht extends h

and all N ele-ments of t [h(i(l)), h(i(m))] have the same k-type. This means thathm ht

D Cg(k),i(l),i(m), contradicting (t) = (k, l, m).

Therefore there exists t [N]k such that for every k < k and all l < m < ksuch that [i(l), i(m)]I Sg(k) = for some n = n(k,l ,m) we have

Ahnht D Cg(k),i(l),i(m).

Then ht and n = maxk,l,m n(k,l ,m) satisfy (1) and this completes the proof.

By Claim 4.4 we can find an ultrafilter U such that the sets Bij for i < jin I and Cgij for g

nN Mn and i < j in I such that [i, j] I = all belong to U. Let ai be the element of the ultrapower

U Mn with the

representing sequence fi if i I and with the representing sequence ai(n),for n N, if i N. Since the relevant Aki and Bij belong to U we have thatai, i I, is a -chain in the ultraproduct.

In order to check it is weakly (1, )-skeleton like fix g U

Mn anda representing sequence g

n Mn of g. Let Jg = {fi : i S}. If i < j

are such that [i, j]I Jg = , then Cgij U, which implies thatU

Mn |=(ai, g) (aj, g) and

U

Mn |= (g, ai) (g, aj), as required.

Remark 4.5. As pointed out in Remark 3.3, the proof of Proposition 4.2

can be easily modified to obtain U such thatU Mi includes a -chain C

isomorphic to I that satisfies the indiscernibility property (*) stronger thanbeing weakly (1, )-skeleton like stated there. In order to achieve this,we only need to add a variant Dijg of the set Cijg to the filter basis from

Claim 4.4 for every k N, every k +n-ary formula (x, y) and every g Ak.Let

Dijg = {n : Mn |= (fi(n), g(n)) (fj(n), g(n))}.

The obvious modification of the proof of Claim 4.4 shows that the augmentedfamily of sets still has the finite intersection property. It is clear that anyultrafilter U extending this family is as required.

5. The proof of Theorem 1

Fix a model A of cardinality c whose theory is unstable. By [19,Theorem 2.13] the theory of A has the order property and we can fix in the signature of A such that A includes arbitrarily long finite -chains.Therefore Theorem 1 is a special case of the following with Ai = A for all i.

Theorem 5.1. Assume CH fails. Assume (x, y) is a formula and Ai, fori N, are models of cardinality c such that in Ai there is a -chain of


21/35

954

re

vision:2010-11-24

modified:2010-11-24


length i. Then there are 2c isomorphism types of models of the form

U

An,where U ranges over nonprincipal ultrafilters onN.

Proof. Since |Ai| i for all i, the ultrapowerU

A has cardinality equalto c whenever U is a nonprincipal ultrafilter on N. By Lemma 2.5 , thereare 2c linear orders I of cardinality c with disjoint representing sequencescorresponding to (defined) invariants invm,c(I) (with m = 2 or m = 3 de-pending on wheher c is regular or not). Use Proposition 4.2 to construct anultrafilter U(I) such that I is isomorphic to a weakly (1, )-skeleton like-chain C in

U(I) Ai. The conclusion follows by Proposition 3.14.

6. Ultrapowers of metric structures

6.1. Metric structures. In this section we prove a strengthening of Theo-rem 3 which is the analogue of Theorem 5.1 for metric structures. First we

include the definitions pertinent to understanding the statement of Theo-rem 3. Assume (A,d,f0, f1, . . . , R0, R1, . . . ) is a metric structure. Hence d isa complete metric on A such that the diameter ofA is equal to 1, each fi is afunction from some finite power of A into A, and each Ri is a function froma finite power of A into [0, 1]. All fi and all Ri are required to be uniformlycontinuous with respect to d, with a fixed modulus of uniform continuity (see[2] or [10, 2]). In the interesting cases, such as (unit balls of) C*-algebras,tracial von Neumann algebras, and Banach spaces, this requirement followsfrom the uniform continuity of algebraic operations on bounded balls.

If U is an ultrafilter on N then on AN we define a quasimetric dU byletting, for a = (ai)iN and b = (bi)iN,

dU(a, b) = limiU

d(ai, bi).

Identify pairs a and b such that dU(a, b) = 0. The uniform continuityimplies that fn(a) = (fn(ai))i and Rn(a) = limiURn(ai) are uniformlycontinuous functions with respect to the quotient metric. The quotientstructure is denoted by

U

(A , d , . . . ) (or shortlyU

A if the signature isclear from the context) and called the ultrapower of A associated with U.An ultraproduct of metric structures of the same signature is defined anal-ogously.

The assumption that the metric d is finite is clearly necessary in order tohave dU be a metric. However, one can show that the standard ultrapowerconstructions of C*-algebras and of II1 factors can essentially be consideredas special cases of the above definition (see [10, 4] for details). These twoconstructions served as a motivation for our work (see 8).

More information on the logic of metric structures is given in [2], and [10]contains an exposition of its variant suitable for C*-algebras and II1 factors.

Let A = (A , d , . . . ) be a metric structure. Interpretations of formulas arefunctions uniformly continuous with respect to d, and the value of an n-aryformula at an n-tuple a is denoted by

(a)A.


22/35

954

re

vision:2010-11-24

modified:2010-11-24


We assume that the theory of A is unstable, and therefore by [10, The-orem 5.5] it has the order property. Fix n and a 2n-ary formula that

witnesses the order property of the theory of A. Define the relation onevery model such that is a formula in its signature by letting a b if andonly if

(a, b) = 0 and (b, a) = 1.

A -chain in A is a subset of An linearly ordered by . Theorem 3 is aconsequence of the following.

Theorem 6.1. Assume CH fails. Assume (x, y) is a formula and Ai, fori N, are metric structures of cardinality c of the same signature suchthat in Ai there is a -chain of length i. Then there are 2

c isometry typesof models of the form

U

An, where U ranges over nonprincipal ultrafiltersonN.

The proof proceeds along the same lines as the proof of Theorem 5.1 andwe shall only outline the novel elements, section by section.

6.2. Combinatorics of the invariants. For a An and b An write

tp(a/b) = (a, b)A, (b, a)A.

For a An and X An, let tp(a/X) be the function from X into [0, 1]2

defined bytp(a/X)(b) = tp(a/b).

The notation and terminology such as [a, b] have exactly the same inter-pretation as in 3.1.

Definition 6.2. A -chain C is weakly (1, )-skeleton like in A if for everya An there is a countable Ca C such that for all b and c in C satisfying

[b, c] Ca =

we have tp(a/b) = tp(a/c).

Note that (C, ) is an honest (discrete) linear ordering. Because of this anumber of the proofs in the discrete case work in the metric case unchanged.In particular, Lemma 3.4, Lemma 3.5, Lemma 3.6 and Lemma 3.7 are truewith the new definitions and the old proofs. Definition 3.8 and the defini-tion of C[B, c] are transferred to the metric case unmodified, using the newdefinition of tp. As a matter of fact, the analogue of Remark 3.2 applies

in the metric context. That is, even if weakly (1, )-skeleton like is definedby requiring only that (with a, Ca, b and c as in Definition 6.2) we only have

a b if and only if a c

andb a if and only if c a

then all of the above listed lemmas remain true, with the same proofs, inthe metric context. However, Lemma 6.5 below requires the original, morerestrictive, notion of weakly (1, )-skeleton like.


23/35

954

re

vision:2010-11-24

modified:2010-11-24


6.3. Defining an invariant over a submodel. Definition 3.8 is unchanged.The statement and the proof of Lemma 3.9 remain unchanged. However,

in order to invoke it in the proof of the metric analogue of Lemma 3.10 weshall need Lemma 6.3 below. For a metric structure B its character den-sity, the smallest cardinality of a dense subset, is denoted by (B). Notethat (A) |C| for every -chain C in A, since each -chain is necessarilydiscrete.

Lemma 6.3. Assume C = ai : i I is a -chain that is weakly (1, )-skeleton like in a metric structure A. Assume B is an elementary submodelof A and a C \ Bn is such that

cf(C[B, a], ) > (B).

Then there is c C[B, a] such that for all d C[B, a] (, c] we have

tp(d/B) = tp(c/B).

Proof. Pick a dense B0 B of cardinality (B). Let c C[B, a] be suchthat

C[B, a]

{Cb : b Bn0 } (, c] = .

Then for every d C[B, c] (, c] and every b Bn0 we have that

[d, c] Cb = , and therefore tp(c/b) = tp(d/b). Since the maps x

tp(c/x) and x tp(d/x) are continuous, they agree on all of Bn and

therefore tp(c/B) = tp(d/B).

6.4. Representing invariants. The definition of INVm,(A, ) from 3.4transfers to the metric context verbatim, and Lemma 3.11 and its proof are

unchanged.

6.5. Counting the number of invariants over a model. Lemma 3.12is unchanged but Lemma 3.13 needs to be modified, since the right analogueof cardinality of a model is its character density.

Lemma 6.4. For A,,m as usual every set of disjoint representing se-quences of invariants in INVm,(A)(A, ) has size at most (A).

Proof. In this paper we shall only need the trivial case when (A) = |A| = c,but the general case is needed in [12]. It will follow immediately from theproof of Lemma 3.13 with Lemma 6.5 below applied at the right moment.

Lemma 6.5. ForA,,m as usual and an elementary submodel B ofA thereare at most (A) distinct (A,B,,m)-invariants.

Proof. Let = (A). Let h : R [0, 1] be the continuous function suchthat h(x) = 0 for x 1/3, h(x) = 1 for x 2/3, and h linear on [1/3, 2/3].Let = h .

Note that every -chain is a -chain. Also, (a1, b1) = (a2, b2) implies(a1, b1) = (a2, b2), and therefore every weakly (1, )-skeleton like -chain is weakly (1, )-skeleton like, with the same witnessing sets Ca. Thisimplies the following, for every elementary submodel B of A and m N.


24/35

954

re

vision:2010-11-24

modified:2010-11-24


(*) If c An defines the (A,B,,m)-invariant d then c defines the(A,B,,m)-invariant d.

Denote the sup metric on An by dn. Since A is a uniformly continuousfunction, there is > 0 sufficiently small so that dn(c1, c2) < implies|(a, c1) (a, c2)| < 1/3 for all a. Therefore we have the following.

(**) For every a An we have that a c1 implies a c2, and a c2implies a c1.

Assume B is an elementary submodel of A and ci defines the (A,B,,m)-invariant di, for i = 1, 2. By (*) we have that ci defines the (A,B,,m)-invariant di, for i = 1, 2. Ifd

n(c1, c2) < then (**) implies d1 = d2.

Proposition 3.14 applies in the metric case literally.

6.6. Construction of ultrafilters. It is the construction of the ultrafilterin 4 that requires the most drastic modification. Although the statementof Proposition 4.2 transfers unchanged, the proof of its analogue, Proposi-tion 6.6, requires new ideas.

Proposition 6.6. Assume (x, y) is a formula and Mi, for i N, aremetric structures of the same signature such that in Mi there is a -chainof length i. Assume I is a linear order of cardinality c. Then there is anultrafilter U onN such that

U

Mn includes a -chain {ai : i I} that isweakly (1, )-skeleton like.

Proof. Like in the proof of Proposition 4.2, we assume is a binary formulain order to simplify the notation. Fix a -chain ai(n), for 0 i < n, in Mn.Like in 4 fix an independent family G of size c and a filter D such that G isindependent with respect to D and D is a maximal filter with this property.Define G, FI(G) and FIs(G) exactly as in 4. Since the diameter of each Mnis 1, each element of

n Mn is a representing sequence of an element of

the ultrapower. Claim 4.3 is modified as follows.

Claim 6.7. For every g

nN Mn there is a countable set Sg I suchthat for all l N and all r Q [0, 1] all sets of the form

Xg,l,r = {n : (al(n), g(n))Mn < r}

Yg,l,r = {n : (g(n), al(n))Mn

< r}

are supported bySg.

Proof. Since there are only countably many relevant sets, this is an imme-diate consequence of Lemma 4.1.

For i < j in I the definitions of sets

Bij = {n : fi(n) fj(n)}


25/35

954

re

vision:2010-11-24

modified:2010-11-24


is unchanged, but we need to modify the definition of Cgij . For g

nN Mn, i < j in I such that [i, j]i Sg = and > 0 letCgij = {n : |(fi(n), g(n))

Mn (fj(n), g(n))Mn | <

and |(g(n), fi(n))Mn (g(n), fj(n))

Mn| < }.

Claim 6.8. The family of all sets Bij for i < j in I and Cgij for g nN Mn, i < j in I such that [i, j]i Sg = and > 0 has the finite

intersection property.

Proof. It will suffice to show that for k N, i(0) < < i(k 1) in I, andg(0), . . . , g(k 1) in

nN Mn and > 0 the set

l


26/35

954

re

vision:2010-11-24

modified:2010-11-24


the sequence {hn} that

Ahnht Cg(k),i(l),i(m), =

D

.For t [N]k let (t) be the lexicographically minimal triple (k,l ,m) suchthat this holds for a large enough n. By Ramseys theorem, there are an

infinite Z N and a triple (k, l, m) such that for every t [N]k we haveAhnht Cg(k),i(l),i(m), =

D .

Let N = |[i(l), i(m)]I T | and find t [Z]k such that the set

[ht(i(l)), ht(i(m

))] Z

has at least (M2 + 2M)N+1 elements. Let h = h (TG SGg(k)

). Then for

each p N there are a large enough m = m(p) such that for some r = r(p)and s = s(p) we have

(Dn Ahmh)|(al(n), g(k)(n))Mn r/M| < /2

and

(Dn Ahmh)|(g(k)(n), al(n))Mn s/M| < /2.

We say that such m decides the k-type of p. Pick m large enough to decidethe k-type of each p [h(i(l)), h(i(m))] Z. Since there are only (M +1)2 different k-types, by the pigeonhole principle there are N elements of[h(i(l)), h(i(m))] Z with the same k-type. There is therefore t

[Z]k such that ht extends t and all N elements of t [h(i(l)), h(i(m))]

have the same k-type. This means that hn ht D Cg(k),i(l),i(m),,

contradicting (t

) = (k

, l

, m

).Therefore there exists t [N]k such that for every k < k and all l < m < k

such that [i(l), i(m)]I Sg(k) = for some n = n(k,l ,m) we have

Ahnht D Cg(k),i(l),i(m),.

Then ht and n = maxk,l,m n(k,l ,m) satisfy (2).

Let U be any ultrafilter that extends the family of sets from the statementof Claim 6.8. Since Mn are assumed to be bounded metric spaces, each fi isa representing sequence of an element of the ultraproduct

U Mn. Denote

this element by ai and let C denote ai : i I. Since Bi,j U for all i < jin I, C is a -chain isomorphic to I. For b UMn fix its representingsequence g and let Cb C be {ai : i Sg}. Since Cg,i,j, U whenever[i, j] Sg = and > 0, we conclude that C is a weakly (1, )-skeleton like-chain as in the proof in 4.

6.7. The proof of Theorem 6.1. Compiling the above facts into the proofof Theorem 6.1 proceeds exactly like in 5.

Remark 6.9. Remark 4.5 applies to Proposition 6.6 in place of Proposi-tion 4.2 verbatim.


27/35

954

re

vision:2010-11-24

modified:2010-11-24


7. Types with the order property

In this section we prove local versions of Theorem 1 and Theorem 3 inwhich the -chain is contained in the set of n-tuples realizing a prescribedtype t (the definition of a type in the logic of metric structures is givenbelow). We will make use of this in case when t is the set of all n-tuplesall of whose entries realize a given 1-type, and the set of these realizationsis a substructure. In order to conclude that a -chain is still a -chainwhen evaluated in this substructure, we will consider a formula that isquantifier-free. Throughout this section we assume A is a model, (x, y) isa 2n-ary formula in the same signature and t is an n-ary type over A.

Although the motivation for this section comes from the metric case, weshall first provide the definitions and results in the classical case of discretemodels. An n-ary type t over A has the order property if there exists a

2n-ary formula such that for every finite t0 t and for every m N thereexists a -chain of length m in A all of whose elements realize t0.

Proposition 7.1. Assume A is countable and type t over A has the orderproperty, as witnessed by . Assume I is a linear order of cardinality c.Then there is an ultrafilter U onN such that

U

A includes a weakly (1, )-skeleton like -chain isomorphic to I consisting of n-tuples realizing t.

Proof. Since t is countable we may write it as a union of finite subtypes,t =

iN ti. Let ai(k), for 0 i < k, be a chain in A of elements

realizing tk. Let G be an independent family of functions of cardinality c.Unlike the proof of Proposition 4.2, we cannot identify G with functions in

k{ai(k) : i < k}, since we cannot assume ai(k) = ai(l) for all i < min(k, l).Therefore to each g G we associate a function g such that

g(k) = ag(k)(k)

if g(k) < k and g(k) = ak1(k), otherwise. Then by the FundamentalTheorem of Ultraproducts g is a representing sequence of an element thatrealizes t. The rest of the proof is identical to the proof of Proposition 4.2.

In order to state the metric version of Proposition 7.1 we import somenotation from [8] and [9]. Given 0 < 1/2 define relation , on A

n via

a1 , a2 if (a1, a2) and (a2, a1) 1

Note that ,0 coincides with . A , -chain is defined in a natural way.We shall now define a type in the logic of metric structures, following [2]

and [10, 4.3]. A condition over a model A is an expression of the form(x, a) r where is a formula, a is a tuple of elements of A and r R.A type t over A is a set of conditions over A. A condition (x, a) r is-satisfied in A by b if (b, a)A r + . Clearly a condition is satisfied by bin A if and only if it is -satisfied by b for all > 0. A type t is -satisfiedby b if all conditions in t are -satisfied by b.


28/35

954

re

vision:2010-11-24

modified:2010-11-24


An n-ary type t over a metric structure A has the order property if thereexists a 2n-ary formula such that for every finite t0 t and for every

m N there exists a , 1/m-chain of length m in A consisting of n-tupleseach of which 1/m-satisfies t0.

Proposition 7.2. Assume A is separable metric structure and type t overAhas the order property, as witnessed by . Assume I is a linear order ofcardinality c. Then there is an ultrafilter U onN such that

UA includes

a weakly (1, )-skeleton like -chain isomorphic to I and consisting of n-tuples realizing t.

Proof. For elements a and b ofU

A and their representing sequences(ai)iN and (bi)iN we have a b in

U

A if and only if {i : ai , bi} Ufor every > 0. Modulo this observation and replacing t with its restric-tion to a countable dense subset of A, the proof is identical to the proof ofProposition 7.1.

In order to prove versions of Proposition 7.1 and Proposition 7.2 for un-countable (respectively, nonseparable) structures we shall need the followingwell-known lemma.

Lemma 7.3. Assume D is a meager filter onN extending the Frechet fil-ter. Then there is a family GD of cardinality c of functions inN

N that isindependent mod D.

Proof. Let G be a family of cardinality c that is independent mod the Frechetfilter ([19, Appendix, Theorem 1.5(1)]). Since D is meager there is a sur-

jection h : N N such that the h-preimage of every finite set is finite

and the h-preimage of every infinite set is D-positive (see e.g., [1]). ThenGD = {h f: f G} is independent mod D because the h-preimage of everyinfinite set is D-positive.

Again A, and t are as above and A


29/35

954

re

vision:2010-11-24

modified:2010-11-24


The construction described in the proof of Proposition 4.2 results in Usuch that all elements of the resulting -chain ai, for i I, realize t.

The proof of the following metric version is identical to the proof of Propo-sition 7.4. Note that A is not assumed to be separable.

Proposition 7.5. Let A be a metric structure and let t is a type over A.Assume there is a function h

kN A

kn such that the sets

X[t0, k] = {i : h(i) is a , 1/k-chain consisting of n-tuples 1/k-satisfying t0}

for t0 t finite and k N generate a meager filter extending the Frechetfilter.

Assume I is a linear order of cardinality c. Then there is an ultrafilterU onN such that

UA includes a-chain{ai : i I} that is weakly (1, )-

skeleton like and consists of elements realizing t.

8. Applications

Recall that Alt(n) is the alternating group on {0, . . . , n1}. The followingis the main result of [7] (see also [23]).

Theorem 8.1 (EllisHachtmanSchneiderThomas). If CH fails then thereare 2c ultrafilters on N such that the ultraproducts

U

Alt(n) are pairwisenonisomorphic.

Proof. Let (x1, x2, y1, y2) be the formula asserting that x1y2 = y2x1 andx2y1 = y1x2. It is then easy to see that for all natural numbers k 2n + 4the group Alt(k) includes a -chain of length n. Therefore the conclusion

follows by Theorem 5.1.

8.1. Applications to operator algebras. Theorem 3 and Theorem 6.1were stated and proved for the case of bounded metric structures. However,the original motivation for the present paper came from a question aboutthe of ultrapowers of C*-algebras and II1 factors stated in early versions of[9] and [10]. An excellent reference for operator algebras is [4].

In the following propositions and accompanying discussion we deal withthe ultrapower constructions for C*-algebras and II1-factors, as well asthe associated relative commutants. Although Theorem 3 was proved forbounded metric structures, it applies to the context of C*-algebras and II1factors. Essentially, one applies the result to the unit ball of the given alge-bra. All the pertinent definitions can be found in [9] or [10].

The classes of C*-algebras and of II1 factors are axiomatizable in the logicof metric structures. Both proofs can be found in [10, 3], and the (muchmore difficult) II1 factor case was first proved in [3], using a rather differentaxiomatization from the one given in [10]. Extending results of [14] and [8],in [9, Lemma 5.4] it was also proved that the class of infinite dimensionalC*-algebras has the order property, as witnessed by the formula

(x, y) = xy x.


30/35

954

re

vision:2010-11-24

modified:2010-11-24


Assume ai, i N, is a sequence of positive operators of norm one such thataiaj is positive and of norm one whenever j < i. Then this sequence forms

a -chain. Such a sequence exists in every infinite-dimensional C*-algebra(see the proof of [9, Lemma 5.2]). Note that it is important to have this-chain inside the unit ball of the algebra. In [9, Lemma 5.4] it was alsoproved that the relative commutant type (see below for the definition) ofevery infinite-dimensional C*-algebra has the order property, and that thisis witnessed by the same as above.

In [9, Lemma 3.2 (3)] it was proved that the class of II1 factors has theorder property, as witnessed by the formula

(x1, y1, x2, y2) = x1y2 y2x12.

Unlike the case of C*-algebras, the relative commutant type of some II1factors does not have the order property. For a II1 factor N, the relativecommutant type having the order property is equivalent to having a non-abelian relative commutant in some (equivalently, all) of its ultrapowersassociated with nonprincipal ultrafilters on N ([9, Theorem 4.8]). Such II1factors are called McDuff factors. We emphasize that, similarly to the caseof C*-algebras, an arbitrarily long finite -chain can be found inside the unitball of a II1 factor. This is necessary in order to have the proof work. Notethat without this requirement even C includes an infinite -chain, althoughC clearly does not have the order property.

Recall that two C*-algebras are (algebraically) isomorphic if and only ifthey are isometric, and that the same applies to II1 factors. The followingis a quantitative improvement to the results of [14], [8] (for C*-algebras)

and [9] (for II1 factors).

Proposition 8.2. Assume A is a separable infinite-dimensional C*-algebraor a separably acting II1-factor. If the Continuum Hypothesis fails, then Ahas 2c nonisomorphic ultrapowers associated with ultrafilters onN.

In Proposition 8.2 it suffices to assume that the character density of A is c. This does not apply to Proposition 8.4 below where the separabilityassumption is necessary (cf. the last paragraph of [10, 4] or [13]).

Proof of Proposition 8.2. Since by the above discussion both classes are ax-iomatizable with unstable theories, Theorem 3 implies that in all of thesecases there are 2c ultrapowers with nonisomorphic unit balls. Therefore the

result follows.

In the light of Proposition 8.2, it is interesting to note that the theory ofabelian tracial von Neumann algebras is stable (this is a consequence of [9,Theorem 4.7] and the characterization of stability from [10, Theorem 5.6]).More precisely, a tracial von Neumann algebra M has the property that ithas nonisomorphic ultrapowers (and therefore by Theorem 3 it has 2c noniso-morphic ultrapowers) if and only if it is not of type I. This is a consequenceof [9, Theorem 4.7].


31/35

954

re

vision:2010-11-24

modified:2010-11-24


The following is a quantitative improvement of [9, Proposition 3.3], con-firming a conjecture of Sorin Popa in the case when the Continuum Hy-

pothesis fails. The intended ultrapower is the tracial ultrapower, and theanalogous result for norm ultrapower is also true.

Proposition 8.3. Assume the Continuum Hypothesis fails. Then thereare 2c ultrafilters onN such that the II1 factors

U

Mn(C) are all noniso-morphic.

Proof. This is a direct application of Theorem 3, using -chains obtainedin [9, Lemma 3.2] .

Assume M is a C*-algebra or a II1 factor and U is a nonprincipal ultra-filter on N. Identify M with its diagonal copy inside

U

M. The relativecommutant of M inside its ultrapower is defined as

M U

M = {a U

M: (a M)ab = ba}.

Thus the relative commutant is the set of all elements ofU

M realizingthe relative commutant type of M, consisting of all conditions of the formxb bx = 0, for b M. (Here stands for 2 in case when M is a II1factor.) The relative commutant is a C*-algebra (II1 factor, respectively)and it is fair to say that most applications of ultrapowers in operator algebrasare applications of relative commutants. A relative commutant is said to betrivial if it is equal to the center of M. From a model-theoretic point ofview, a relative commutant is a submodel consisting of all realizations of adefinable type over M.

The original motivation for the work in [8], [9] and [10] came from thequestion whether all relative commutants of a given operator algera in itsultrapowers associated with ultrafilters on N are isomorphic. This was askedby Kirchberg in the case of C*-algebras and McDuff in the case of II1-factors.Here is a quantitative improvement to the answer to these questions givenin the above references.

Proposition 8.4. Assume A is a separable infinite-dimensional C*-algebraor a separably acting McDuff II1-factor. If the Continuum Hypothesis fails,then A has 2c nonisomorphic relative commutants in ultrapowers associatedwith ultrafilters onN.

Proof. In [9, Theorem 4.8] it was proved that the relative commutant type

of a McDuff factor has the order property, witnessed by given in theintroduction to 8.1. In [9, Lemma 5.4] it was proved that the relative com-mutant type of any infinite-dimensional C*-algebra has the order property,witnessed by given in the introduction to 8.1. Hence applying Proposi-tion 7.2 concludes the proof.

By B(H) we shall denote the C*-algebra of all bounded linear operators onan infinite-dimensional, separable, complex Hilbert space H. In [13] it wasproved that that for certain ultrafilters on N the relative commutant ofB(H)


32/35

954

re

vision:2010-11-24

modified:2010-11-24


in

U

B(H) is nontrivial. These ultrafilters exist in ZFC. It was also provedin [13] that the relative commutant of B(H) in an ultrapower associated to

a selective ultrafilter is trivial. Therefore CH implies that not all relativecommutants of B(H) in its ultrapowers associated with ultrafilters on N areisomorphic. This fact motivated Juris Steprans and the first author to askwhether this statement can be proved in ZFC. Since B(H) is not a separableC*-algebra, the following is not a consequence of Proposition 8.4.

Proposition 8.5. Assume that the Continuum Hypothesis fails. ThenB(H)has 2c nonisomorphic relative commutants associated with its ultrapowers.

Proof. We shall apply Proposition 7.5. The following construction borrowssome ideas from the proof of [13, Theorem 3.3 and Theorem 4.1]. Let F


33/35

954

re

vision:2010-11-24

modified:2010-11-24


a finite subset of t, let > 0, and let a0, . . . , ak1 list all elements of B(H)occurring in t0. Let = /6. [13, Lemma 4.6] implies that there are g0 and

g1 such that for each i < k we can write ai = a0i + a1i + ci so that(1) a0i commutes with ah for every h that is constant on every interval

of the form [g0(m), g0(m + 1)),(2) a1i commutes with ah for every h that is constant on every interval

of the form [g1(m), g1(m + 1)), and(3) ci < .

Then for i < k , j Xg0, Xg1,, and h an entry of h(j) we have

[ai, ah] = [a0i , ah] + [a

1i , ah] + [ci, ah]

and since a0i , a1i and ah are all 1 we conclude that [ai, ah] < 6.

Therefore ah realizes t0, and Proposition 7.5 implies that for every linear

order I of cardinality c there is an ultrafilter U such thatUB(H) contains

a -chain C isomorphic to I which is (1, )-skeleton like and included in therelative commutant of B(H). Since is quantifier-free, C remains a -chainin the relative commutant B(H)

UB(H). Since C is (1, )-skeleton like

inU

B(H), it is (1, )-skeleton like in the substructure. Using Lemma 2.5,Lemma 3.11, Lemma 3.13 and a counting counting argument as in the proofof Theorem 5.1 we conclude the proof.

8.2. Concluding remarks. Before Theorem 1 was proved the followingtest question was asked in a preliminary version of [10]: Assume A and Bare countable models with unstable theories. Also assume U and V are ultra-filters on N such that U A =VA. Can we conclude that U B =VB?A positive answer would, together with [15, 3], imply Theorem 1. However,the answer to this question is consistently negative. Using the method of [20]one can show that in the model obtained there there are countable graphsG and H and ultrafilters U and V on N such that

U

G,

VG and

V

Hare saturated but

U

H is not. This model has an even more remarkableproperty: Every automorphism of

U

H lifts to an automorphism of HN.An interesting and related application of [20] was recently given in [17].

Methods of the present paper was adapted to the class of all approximatelymatricial (shortly AM) C*-algebras in [12]. A C*-algebra is AM if andonly if it is an inductive limit of finite-dimensional matrix algebras ([11]).In [12] it was proved that in every uncountable character density there

are 2

nonisomorphic AM algebras. Unlike the classes of C*-algebras andII1 factors, the class of AM algebras is not elementary. This is becauseAM algebras are not closed under taking ultrapowers (by the proof of [10,Proposition 6.1]).

Results related to our 6 were proved in [22], where it was shown that anunstable theory in logic of metric structures has maximal possible number ofmodels in every uncountable cardinality. In the general case, treated in [22],there is a distinction between isomorphic and isometric. For C*-algebrasand II1 factors treated here the two notions are equivalent.


34/35

954

re

vision:2010-11-24

modified:2010-11-24


References

[1] T. Bartoszynski and H. Judah, Set theory: on the structure of the real line, A.K.

Peters, 1995.[2] I. Ben Yaacov, A. Berenstein, C.W. Henson, and A. Usvyatsov, Model theory for

metric structures, Model Theory with Applications to Algebra and Analysis, Vol. II(Z. Chatzidakis et al., eds.), Lecture Notes series of the London Math. Society., no.350, Cambridge University Press, 2008, pp. 315427.

[3] I. Ben Yaacov, W. Henson, M. Junge, and Y. Raynaud, Preliminary report - vNAand NCP, preprint, 2008.

[4] B. Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences, vol. 122,Springer-Verlag, Berlin, 2006, Theory of C-algebras and von Neumann algebras,Operator Algebras and Non-commutative Geometry, III.

[5] C. C. Chang and H. J. Keisler, Model theory, third ed., Studies in Logic and theFoundations of Mathematics, vol. 73, North-Holland Publishing Co., Amsterdam,1990.

[6] A. Dow, On ultrapowers of Boolean algebras, Topology Proc.9

(1984), no. 2, 269291.[7] P. Ellis, S. Hachtman, S. Schneider, and S. Thomas, Ultraproducts of finite alternatinggroups, RIMS Kokyuroku No. 1819 (2008), 17.

[8] I. Farah, The relative commutant of separable C*-algebras of real rank zero, Jour.Funct. Analysis 256 (2009), 38413846.

[9] I. Farah, B. Hart, and D. Sherman, Model theory of operator algebras I: Stability,preprint, arXiv:0908.2790, 2009.

[10] , Model theory of operator algebras II: Model theory, preprint, arXiv:1004.0741,2009.

[11] I. Farah and T. Katsura, Nonseparable UHF algebras I: Dixmiers problem, Advancesin Math. 225 (2010), 13991430.

[12] , Nonseparable UHF algebras II: Classification, in preparation, 2010.[13] I. Farah, N.C. Phillips, and J. Steprans, The commutant of L(H) in its ultrapower

may or may not be trivial, Math. Annalen 347 (2010), 839857.

[14] L. Ge and D. Hadwin, Ultraproducts ofC-algebras, Recent advances in operator the-ory and related topics (Szeged, 1999), Oper. Theory Adv. Appl., vol. 127, Birkhauser,Basel, 2001, pp. 305326.

[15] Linus Kramer, Saharon Shelah, Katrin Tent, and Simon Thomas, Asymptoticcones of finitely presented groups, Advances in Mathematics 193 (2005), 142173,math.GT/0306420.

[16] K. Kunen, Set theory: An introduction to independence proofs, NorthHolland, 1980.[17] P. Lucke and S. Thomas, Automorphism groups of ultraproducts of finite symmetric

groups, preprint, 2009.[18] V. Pestov, Hyperlinear and sofic groups: a brief guide, Bulletin of Symbolic Logic 14

(2008), 449480.[19] Saharon Shelah, Classification theory and the number of nonisomorphic models, Stud-

ies in Logic and the Foundations of Mathematics, vol. 92, North-Holland Publishing

Co., Amsterdam, xxxiv+705 pp, 1990.[20] , Vive la difference III, Israel Journal of Mathematics 166 (2008), 6196,math.LO/0112237.

[21] , General non-structure theory, preprint, arXiv:1011.3576, 2010.[22] Saharon Shelah and Alex Usvyatsov, Unstable Classes of Metric Structures, preprint.[23] S. Thomas, On the number of universal sofic groups, preprint, 2009.


35/35

re

vision:2010-11-24

modified:2010-11-24


Department of Mathematics and Statistics, York University, 4700 Keele

Street, North York, Ontario, Canada, M3J 1P3, and Matematicki Institut,

Kneza Mihaila 35, Belgrade, SerbiaURL: http://www.math.yorku.ca/ifarahE-mail address: [email protected]

The Hebrew University of Jerusalem, Einstein Institute of Mathematics,

Edmond J. Safra Campus, Givat Ram, Jerusalem 91904, Israel, and, Depart-

ment of Mathematics, Hill Center-Busch Campus, Rutgers, The State Uni-

versity of New Jersey, 110 Frelinghuysen Road, Piscataway, NJ 08854-8019

USA

E-mail address: [email protected]: http://shelah.logic.at/

ilijas farah and saharon shelah- a dichotomy for the number of ultrapowers

Documents