arXiv:1603.08134v6 [math.LO] 10 Mar 2021 hoei at about facts theoretic eut ftefis at egv oefcsaotgnrlBan general about facts th and some spaces Banach give (1) some we in lines part, dividing first model-theoretic the of results pt ml etrain embeds perturbations small to up pedxB eako oeta’ ihtm 26 dichotomy Rosenthal’s on remark A B: Appendix 8 iiiglnsi aahsae n oe hoy20 theory A: model Appendix and spaces 7 Banach in lines Dividing 6 h cniuu)fis-re hoyof theory first-order (continuous) the pc,nnidpnec rpry perturbed property, independence non space, 5 NIP and space Tsirelson’s 4 Types 3 analysis functional from Preliminaries 2 Introduction 1 Contents isomorphism. 1 c M Abstract. References M ujc lsicto:4B4 62,0C5 03C98. 03C45, 46B25, 46B04, classification: subject AMS Keywords atal upre yIMgat 3302ad93030059 and 93030032 grants IPM by supported Partially 0 and - T ℵ a the has 0 ctgrclBnc pcscontain spaces Banach -categorical ℓ p sbpcso perturbed of -subspaces srlo’ pc,cniuu oi,srnl separable strongly logic, continuous space, Tsirelson’s : o needneproperty independence non hspprhstreprs is,w sals oeo h ba the of some establish we First, parts. three has paper This ℵ 0 ctgrct,continued -categoricity, M T h srlo pc fFge n ono F7] eod u Second, [FJ74]. Johnson and Figiel of space Tsirelson the , ..Bx31517,Aa,Iran; Arak, 38135-1177, Box P.O. rkUiest fTechnology, of University Arak -al [email protected] e-mail: c 0 eateto science, of Department or M ai Khanaki Karim T ℓ ℵ p NP;()eeyBnc pc htis that space Banach every (2) (NIP); osntcharacterize not does 0 (1 ctgrclsae 14 spaces -categorical ℵ 1 0 6 -categorical < p 1 ∞ lotioercly consequently isometrically; almost ) i hois npriua,w show: we particular, In theories. eir c pcs hr,w study we Third, spaces. ach M T pt lotisometric almost to up , tbesae Rosenthal space, stable , ℓ p or ℵ c 0 -categorical 0 i model sic igthe sing 28 25 9 6 4 2 1 Introduction

A famous conjecture in Banach space theory had predicted that every Banach space con- tains at least one of the classical spaces c0 or ℓp for some 1 6 p< ∞. Tsirelson’s example [Tsi74] was the first space not containing isomorphic copies any of the classical sequence spaces, and the first space whose norm was defined implicitly rather than explicitly. This phenomenon, i.e. implicit definability, later played an important role in Banach space theory when new spaces with implicitly defined norms yielded solutions to many of the most long standing problems in the theory (see [Gow95]). Given the fact that all spaces whose norms are implicitly defined do not contain any of the classical sequences, some Banach space theorists, such as Gowers [Gow95, Gow09] and Odell [Ode97], asked the following question:

(Q1) “Must an ‘explicitly defined’ Banach space contain ℓp or c0?” To provide a positive answer to this question one must first provide a precise definition of ‘explicitly defined space’. Second, he/she must show that explicitly defined spaces contain ℓp for some 1 6 p < ∞, or c0. Third, all classical Banach spaces such as Lp spaces, Lorentz spaces, Orlicz spaces, Schreier spaces, etc., are explicitly defined but the Tsirelson example and similar spaces are not (see [Gow09] for more discussions). On the other hand, the notion of definability plays a basic and key role in model theory and its applications in algebraic structures. Recall that a subset A of a first-order structure M is definable if A is the set of solutions in M of some formula φ(x). However, the problem there is that the usual first-order logic does not work very well for structures in Analysis. Chang and Keisler [CK66] and Henson [Hen76] produced the continuous logics and the logic of positive bounded formulas, respectively, to study structures in analysis using model theoretic techniques. Since then, numerous attempts to provide a suitable logic has been going on (e.g. see [HM83, HI02, Ben03]), and eventually led to the creation of the first order continuous logic (see [BU10, BBHU08] for more details). In many ways the latter logic is the best and ultimate. One of the main purposes of this article is to take a step towards answering Gowers’s question using logical tools presented in [BU10, BBHU08] and to prove that Banach spaces whose norms are determined by first order statements (i.e. their first order theories 2 have exactly one separable model) contain ℓp for some 1 6 p < ∞ or c0. Since the dichotomy stable/unstable (in both logics, classic and continuous) has been widely studied in literature, and some model theoretic properties, such as IP and SOP, in continuous logic (or Banach space theory) have received less attention, the second goal of this paper is to provide examples of these concepts. The third and final purpose is to get a better understanding of communication between both fields (model theory and Banach space theory) so that techniques from one field might become useful in the other.

2In this paper, when we say a Banach space is determined by first order statements (or its theory) we mean its continuous theory is ℵ0-categorical in the sense of Definitions 5.13 or 7.1 below. That is, the first order axioms determine a unique separable space up to isometry.

2 It is worth pointing out that some of the results in the present paper, such as Corol- lary 5.14, seem to be folklore to experts in the field, including Ward Henson and Alexander Usvyatsov. Of course, we are not sure that this observation itself appears somewhere in the literature. On the other hand, some results, such as Corollary 4.19 and Theorem 5.10 and others, seem to be new even to experts. In fact our main result is stronger than the name of the paper. To summarize the results of this paper, in the first part (Section 4), we define the notion of ‘NIP on a model’ and study the example of Tsirelson as a guidance. Since, by the Krivine-Maurey theorem which asserts that every stable space embeds some ℓp almost isometrically, Tsirelson’s space is not stable, so we consider a weaker property and ask the following question: Does Tsirelson’s space have NIP?3 By a result due to E. Odell [Ode82] which asserts that the type space of Tsirelson’s space is strongly separable, we show that the answer to the latter question is positive.4 On the other hand, it is easy to show that 5 an ℵ0-categorical space is stable if and only if its type space is strongly separable. Using these observations, we conclude that the theory of Tsirelson’s space (in any countable language) is not ℵ0-categorical; equivalently, this space can not be determined by first order axioms in continuous logic. This fact leads us to the following question:

(Q2) Does an ℵ0-categorical Banach space contain ℓp or c0? In the second part of paper (Section 5), we answer this question. In fact, the answer is “yes”, and this is a step towards answering the question (Q1) above; furthermore we explain why ℵ0-categoricity (up to small perturbations) is in some ways the appropriate notion for our purpose and why this notion is not a complete answer to the question (Q1). We prove that separable structures which are ℵ0-categorical up to small perturbations contain ℓp or c0. We will discuss this subject in detail and offer other observations that could be interesting in themselves. In the third part of the paper (Sections 6), we first remark some results of [Kha18a] about correspondences between dividing lines in model theory and Banach space theory which will be used in the paper. Then, we study model theoretic dividing lines in some Banach spaces and their theories. In Appendix A, we revisit ℵ0-categoricity and in Appendix B, we remark an obser- vation about Rosenthal’s dichotomy communicated to us by Michael Megrelishvili which asserts that Fact 8.4 below is not a dichotomy in noncompact Polish case. To our knowl- edge this observation itself does not appear somewhere in literature in a clear form. To simplify to read through the paper, we list some more important observations and results: Theorem 5.10, Observations 4.6, 4.17, Corollary 4.19, 5.14, 7.5, Examples 4.5, 6.6, 8.8. Corollary 4.19 and Theorem 5.10 are new. It is worth recalling another line of research. In [Iov05], Jose Iovino studied the connection between the definability of types and the existence of ℓp and c0 in the framework of Henson’s logic, i.e. the logic of positive bounded formulas. Theorem 1.1 of [Iov05]

3See Definition 4.1 for the definition of NIP on a model. 4See Definition 4.15 for the definition of strongly separable space. 5 See Definitions 5.13 and 7.1 for the definition of ℵ0-categorical structures.

3 asserts that the existence of enough definable types guarantees the existence of ℓp or c0 as a subspace. After preparing of the present paper we came to know that, independently from us, Alexander Usvyatsov [Usv] also observed that ℵ0-categorical spaces contain some classical sequences. This paper is organized as follows: In the second section, we briefly review the notions of types in model theory and Banach space theory. In the third section, we recall the definition of Tsirelson’s space and some facts which are used later. Also, we present the notion NIP on a model and show that every Banach space with strongly separable types, as well as Tsirelson’s space, has NIP. In the forth section, we prove that every perturbed ℵ0-categorical space embeds c0 or ℓp almost isometrically (Theorem 5.10). In the fifth section, we briefly review some model theoretic dividing lines and their connections to some functional analysis notions, and study some examples and investigate their model theoretic dividing lines. In Appendix A, we revisit ℵ0-categoricity and in Appendix B, we give some remarks on Rosenthal’s dichotomy.

Acknowledgements. I am very much indebted to C. Ward Henson for his kindness and his helpful and detailed comments. I want to thank Alexander Berenstein, Gilles Godefroy, Timothy Gowers, Richard Haydon, Tom´as Ibarluc´ıa, Michael Megrelishvili and Alexander Usvyatsov for their valuable comments and observations. I am grateful to Thomas Schlumprecht for sending a copy of [Ode82] to me. I thank the anonymous referee for his detailed suggestions and corrections; they helped to improve significantly the exposition of this paper. I would like to thank the Institute for Basic Sciences (IPM), Tehran, Iran. Research partially supported by IPM grants 93030032 and 93030059.

2 Preliminaries from functional analysis

In this section we review some basic notions from functional analysis. Further details can be found in [AK16]. A Banach space M = (M, k·k) is a complete normed linear space. The dual (resp. bidual) space of M is denoted by M ∗ (resp. M ∗∗). Recall that a Banach space X is reflexive if the mapping : X → X∗∗, given byx ˆ(x∗) = x∗(x), is an isometry of X onto X∗∗. Let 1 6 p < ∞. ℓp bis the Banach space of all sequences x = (xn) of reals so that ∞ p 1/p kxkp = ( 1 |xn| ) < ∞. c0 is the Banach space of all null sequences (xn) under k(xn)k∞ = sup |xn| = maxn |xn|. The (nonseparable) Banach space ℓ∞ is the space of P n all bounded sequences (xn) under k(xn)k∞ = supn |xn|. Let M be a Banach space. The weak topology of M is the weakest topology on M ∗ such that each x ∈ M is continuous. That is, any net (xα) converges weakly to x0 ∈ M ∗ ∗ ∗ ∗ if and only if for each x ∈ M , x (xα) → x (x0). Let X be a compact Hausdorff space. The space of continuous real-valued functions on X is denoted by C(X). Since X is a compact space, every f ∈ C(X) is bounded and

4 C(X) is a complete normed linear space with the supremum norm. A net (fα) ∈ C(X) X converges pointwise to a function f ∈ R if and only if for each x ∈ X, |fα(x)−f(x)|→ 0. It is a well-known fact that for each compact topological space X, the weak topology and the pointwise convergence topology on norm-bounded subsets of C(X) are the same (see [Fre06, 462F]). Definition 2.1. Let M be a Banach space and A ⊆ M. (i) A is relatively weakly compact if it has compact closure in the weak topology on M. (ii) A is relatively weakly sequentially compact if every sequence (xn) of it has a subsequence (yn) such that it weakly converges to an element of M. Let X be compact and A ⊆ C(X) norm-bounded. Then, as the weak topology and the poinwise convergence topology on A are the same, A is relatively weakly compact if it has compact closure in the pointwise convergence topology on C(X). A ⊆ C(X) is relatively weakly sequentially compact if every sequence of it has a pointwise convergent subsequence such that its limit is a continuous function. Definition 2.2. Let M be Banach space and A norm-bounded subset of it. (i) A is Rosenthal if every sequence in A has a weak Cauchy subsequence. M is called Rosenthal if its unit ball is Rosenthal. In particular, if M = C(X) (where X is compact) and A ⊆ C(X) is Rosenthal, we say that A is relatively (weakly) sequentially compact in RX (short RSC). That is, every bounded sequence in A has a pointwise convergent subsequence in RX . (ii) A is said to be weakly sequentially complete (short WS-complete) if every weak Cauchy sequence of A has a weak limit (in M). M is called weakly sequentially complete if its unit ball is weakly sequentially complete. In particular, if M = C(X) (where X is compact) and A ⊆ C(X), we say that A has the weak sequential completeness property (or short SCP) if the limit of each pointwise convergent sequence (fn) ⊆ A is continuous. Fact 2.3 (Eberlein–Smulianˇ Theorem). Let M be Banach space and A ⊆ M norm- bounded. Then the following are equivalent: (1) A is relatively weakly compact in M. (2) The following two properties hold: (i) A is Rosenthal, and (ii) A is weakly sequentially complete. Explanation. See [Whi67] for a short proof of the Eberlein–Smulianˇ Theorem. (See also Theorem 1.6.3 of [AK16].) Notice that, (2) is precisely the condition B in the main theorem of [Whi67]. Indeed, every sequence (xn) has a subsequence (yn) such that it weakly converges to an element of M (i.e. relatively weakly sequentially compact) if and only if (i) each sequence has a weakly Cauchy subsequence (i.e. Rosenthal), and (ii) every weak Cauchy sequence has a weak limit in M (i.e. weakly sequentially complete). Clearly, (1) is the condition A in [Whi67].

5 3 Types

We assume that the reader is familiar with continuous logic (see [BBHU08] or [BU10]). We will study the notion of type in continuous logic and its connection with the notion of type in Banach space theory. In this paper, local types are more important for us.

Types in model theory Although in the current paper we study unbounded metric structures (i.e. Banach spaces) but without loss of generality we can assume that all structures are bounded and so usual framework of continuous logic is sufficient for our purposes. Convention 3.1. In this paper, we study bounded metric structures. More precisely, we study the unit balls of Banach spaces. Furthermore, each formula can have an arbitrary bound, although our focus is on the formula φ(x, y) = kx + yk. So, we can assume that the atomic formulas are [0, 2]-valued. As mentioned model theory notation is standard [BBHU08]. We fix an L-formula φ(x, y) and a complete L-theory T . We let φ˜(y, x) = φ(x, y). Let M be an L-structure, A ⊆ M and TA = T h(M, a)a∈A. Let p(x) be a set of L(A)-statements in free variable x. We shall say that p(x) is a type over A if p(x) ∪ TA is satisfiable. A complete type over A is a maximal type over A. The collection of all such types over A is denoted by SM (A), or simply by S(A) if the context makes the theory TA clear. The type of a in M over A, denoted by tpM (a/A), is the set of all L(A)-statements satisfied in M by a. If φ(x, y) is a formula, a φ-type over A is a maximal consistent set of statements of the form φ(x, a) > r and φ(x, a) 6 s, for a ∈ A and r, s ∈ R. The set of φ-types over A is denoted by Sφ(A).

Similarly, we define Sφ˜(A).

The logic topology and ρ-metric We now give a characterization of complete types in terms of functional analysis. Let M LA be the family of all interpretations φ in M where φ is an L(A)-formula with a free variable x. Then LA is an Archimedean Riesz space of measurable functions on M (see [Fre04]). Let σA(M) be the set of Riesz homomorphisms I : LA → R such that I(1) = 1. The set σA(M) is called the spectrum of TA. Note that σA(M) is a weak* ∗ compact subset of LA. The next proposition shows that a complete type can be coded by a Riesz homomorphism and gives a characterization of complete types. In fact, by M M compactness theorem, the map S (A) → σA(M), defined by p 7→ Ip where Ip(φ )= r if φ(x)= r is in p, is a bijection. Remark 3.2. For an Archimedean Riesz space U with order unit 1, write X for the set of all normalized Riesz homomorphisms from U to R, or equivalently, the positive extreme points of unit ball of the dual space U ∗, with its weak* topology. Then X is compact by Alaoglu’s theorem and the natural map u 7→ uˆ : U → RX defined by setting uˆ(x) = x(u) for x ∈ X and u ∈ U, is an embedding from U to an order-dense and norm-dense embedding subspace of C(X).

6 M Now, by the above remark, LA is dense in C(S (A)). It is easy to show that:

Fact 3.3. Suppose that M, A and TA are as above.

M (i) The map S (A) → σA(M) defined by p 7→ Ip is bijective. (ii) p ∈ SM (A) if and only if there is an elementary extension N of M and a ∈ N such that p = tpN (a/A).

M ∗ We equip S (A) = σA(M) with the related topology induced from LA. Therefore, SM (A) is a compact and Hausdorff space. For any complete type p and formula φ, we let M M φ(p) = Ip(φ ). It is easy to verify that the topology on S (A) is the weakest topology in which all the functions p 7→ φ(p) are continuous. This topology sometimes called the logic topology. The same things are true for Sφ(A).

Definition 3.4 (The ρ-metric). The space Sφ(M) has another topology. Indeed, define ρ(p, q) = supa∈M,kak≤1 |φ(p, a) − φ(q, a)|. Clearly, ρ is a metric on Sφ(M) and it is called the ρ-metric on the type space Sφ(M).

Remark 3.5. For unbounded logics, we define ρu(p, q) = supa∈M |φ(p, a) − φ(q, a)|. The ρu-metric sometimes is called the uniform topology. In this case, there is another topology k 1 on the type space, namely the strong topology. Indeed, define ρs(p, q)= 1 2k supkak6k |φ(p, a)− φ(q, a)|. For the formula φ(x, y)= kx + yk, R. Haydon has informed us that he has con- structed a Banach space M for which the space of φ-types over M isP strongly separable but not uniformly separable. Despite this, for bounded continuous logic, and hence in this paper, the ρ-metric and the strong and uniform topologies are the same.

Definition 3.6. A (compact) topometric space is a triplet hX, T, di, where T is a (com- pact) Hausdorff topology and d a metric on X, satisfying: (i) The metric topology refines the topology. (ii) The metric d : X2 → [0, ∞] is lower semi-continuous, that is, for all r ∈ R+ the set {(a, b) ∈ X2 : d(a, b) ≤ r} is closed in (X2, τ × τ), where τ × τ is the product topology.

Fact 3.7. Let M be a structure and φ(x, y) a formula. The triplet hSφ(M),τ,ρi is a compact topometric space where τ and ρ are the logic topology and the ρ-metric on the type space Sφ(M), respectively.

Proof. Since (Sφ(M), τ) is compact, it suffices to show that ρ is lower semi-continuous. (See Lemma 1.4 in [Ben08b].) Indeed, suppose that κ is a cardinal number and (pα, qα) ∈ 2 (Sφ(M)) , α < κ, such that ρ(pα, qα) ≤ r for all α < κ. Let U be an ultrafilter on κ. It is a well-known fact that, since Sφ(M) is τ-compact, there are p, q ∈ Sφ(M) such that the U-limit of (pα, qα),α<κ is (p, q). (Recall that the U-limit of (xα)α<κ is x if for every neighborhood V of x, the set {α : xα ∈ V } is in U. In this case, we write limU xα = x.)

7 Therefore,

ρ(p, q) = sup |φ(p, b) − φ(q, b)| kbk≤1

= sup lim |φ(pα, b) − φ(qα, b)| kbk≤1 U

≤ lim sup |φ(pα, b) − φ(qα, b)| ≤ r. U kbk≤1

As κ and U are arbitrary, ρ is lower semi-continuous. 

Note that (Sφ(M), τ) is the weakest topology in which all functions p 7→ φ(p, a), a ∈ M, are continuous, and all functions p 7→ φ(p, a), a ∈ M, are uniformly continuous relative to ρ with a modulus of uniform continuity ∆φ (see [BBHU08]).

Types in Banach space theory

Let M be a Banach space and a be in the unit ball BM = {x ∈ M : kxk 6 1}. In Banach space theory, a function fa : BM → [0, 2] defined by x 7→ kx + ak is called a trivial type over the unit ball of M. A type on the unit ball of M is a pointwise limit of a family of trivial types. The set S(M) of all types over the unit ball of M is called the space of types. In fact S(M) ⊆ [0, 2]M is a compact topological space with respect to the product topology. This topology is called the weak topology on types, and denoted by τ ′. ′ S(M) has another topology. Define ρ (f,g) = supa∈M,kak≤1 |f(a) − g(a)| for f,g ∈ S(M). Clearly, ρ′ is a metric on S(M). Remark 3.8. Notice that, in the original defintion of types in the Banach space theory, types are on the entire space (not just the unit ball), and to be realized by anything in the space not just in the unit ball. However, this does not affect the results of this paper, because we use positive results for the original definition. As an example, strong separability in the sense of [Ode82] implies strong separability in Definition 4.15 below. Fact 3.9. The triplet hS(M), τ ′, ρ′i is a compact topometric space. Proof. The proof is a straightforward adaptation of the proof of Fact 3.7. 

Let φ(x, y)= kx + yk. Then there is a correspondence between Sφ(M) and S(M). In- deed, let a ∈ BM , and consider the quantifier-free type a over BM , denoted by tpqf (a/BM ). It is easy to verify that tpqf (a/BM ) corresponds to the function fa : BM → [0, 2] defined by x 7→ kx + ak. Also, there is a correspondence between tpqf (a/BM ) and the Riesz M homomorphism Ia : LM → [0, 2] defined by Ia(φ(x, y)) = φ (x, a) for all x ∈ BM . To summarize, the following maps are bijective:

tpqf (a/BM ) fa

tpqf (a/BM ) Ia Now, it is easy to check that their closures are the same:

8 Fact 3.10. Assume that M is a Banach space and φ(x, y) = kx + yk. The compact ′ ′ topometric spaces hSφ(M),τ,ρi and hS(M), τ , ρ i are the same. More exactly, (Sφ(M), τ) ′ ′ and (S(M), τ ) are homeomorphic, and (Sφ(M), ρ) and (S(M), ρ ) are isometric.

4 Tsirelson’s space and NIP

In this section we show that the Tsirelson space has a property weaker than stability, namely the non independence property (NIP). First we remind the notion NIP on a model from [Kha19b].

NIP on a model Definition 4.1. (i) Let M be a structure, and φ(x, y) a formula. We say that φ(x, y) has the independence property (short IP) on M if there are a sequence (an) ⊆ M, and r>s, such that for each finite disjoint subsets E, F of N:

M M b ∈ M : φ (an, b) 6 s ∧ φ (an, b) > r 6= ∅. n∈E n∈F n ^
^
o M In this case, we say that (φ (an,y): i<ω) is an independent sequence. We say that φ(x, y) has NIP on M if it does not have IP on M. (ii) A complete theory T has IP if there are a formula φ and a model M of it such that φ has IP on M, and otherwise it is said that T has NIP. Lemma 4.2. Let M be a structure, and φ(x, y) a formula. Then the following are equiv- alent. (i) φ(x, y) has NIP on M.

(ii) For each sequence (an) ⊆ M, each elementary extension N  M, and r>s, there are some possibly infinite disjoint subsets E, F of N such that

N N b ∈ N : φ (an, b) 6 s ∧ φ (an, b) > r = ∅. n∈E n∈F n ^
^
o

(iii) For each sequence φ(an,y) in the set A = {φ(a, y): Sφ˜(M) → R | a ∈ M}, where ˜ Sφ˜(M) is the space of all complete φ-types on M, and r>s there are some finite disjoint subsets E, F of N such that

q ∈ Sφ˜(M): φ(an, q) 6 s ∧ φ(an, q) > r = ∅. n∈E n∈F n ^
^
o (iv) The condition (iii) holds for arbitrary disjoint subsets E, F of N.

(v) Every sequence φ(an,y) in A has a pointwise convergent subsequence.

9 Proof. See Lemma 3.12 in [Kha19b] for a proof.  Remark 4.3. (i) A formula φ has NIP for a theory T (in the obvious sense) iff it has NIP on every model M of T iff it has NIP on some approximately ω-saturated model of T . (See Definition 1.3 in [BU10] for the definition of approximate ω-saturation.) M (ii) Note that for each separable structure M, every sequence φ (an,y): M → R has a convergent subsequence (see Lemma 4.16 below), but this does not imply that every

sequence φ(an,y): Sφ˜(M) → R in A has a convergent subsequence (see Examples 4.5 and 8.8 below). (iii) The notion of NIP on a model was introduced in [Kha19b]. Recently, in [KP18] some equivalences of this notion are given. For example, it is shown that a formula φ has NIP on M iff every coheir of a φ-type over M is Borel (Baire 1) definable. Definition 4.4. A Banach space (or Banach structure) M has NIP (resp. IP) if the formula φ(x, y)= kx + yk has NIP (resp. IP) on M.

The following example shows that c0 has IP. Also, we will see later that this example shows that Fact 8.4 below is not dichotomy in noncompact spaces (see Example 8.8 below). We thank Michael Megrelishvili for communicating to us the example.

Example 4.5. Let c0 = {x = (xn)n ∈ ℓ∞ : limn xn = 0} with kxk = supn∈N |xn|. Let Bc0 be the unit ball of c0, i.e. Bc0 = {x ∈ c0 : kxk 6 1}. For each a ∈ Bc0 , define

fa : Bc0 → [0, 2] by x 7→ kx + ak. Let Fc0 = {fa : a ∈ Bc0 }. Then the family Fc0 contains an independent sequence. Indeed, for each n, define fn(x) = kx + enk for all x where (en)n is the standard basis of c0. We show that (fn)n is an independent sequence: for every finite disjoint subsets I and J of N (the naturals) define a binary vector x ∈ c0 as the characteristic function of J: xj = 1 for every j ∈ J and xk = 0 for every k∈ / J (in particular, for every i ∈ I). Then x ∈ c0, kxk = 1 and fi(x) = 1 for every i ∈ I, fj(x)=2 for every j ∈ J. Therefore, (fn)n is an independent sequence. Observation 4.6. Suppose that X is a Banach space structure containing a subspace isomorphic to c0.Then X has the independence property (IP).

Proof. Recall that every space isomorphic to c0 has subspaces that embed c0 almost isometrically. (See Proposition 2.e.3 in [LT77].) So, the sequence (fn)n in Example 4.5 works well.  The converse does not hold, and we will give a counterexample in a future work.

We now recall the notion of non order property (NOP) or stability on a model. Definition 4.7. Let M be a structure, and φ(x, y) a formula. The following are equivalent and in any of the cases we say that φ(x, y) is stable on M (or has NOP on M).

(i) Whenever an, bm ∈ M form two sequences we have

lim lim φ(an, bm) = lim lim φ(an, bm), n m m n when the limits on both sides exist.

10 (ii) The set A = {φ(x, b): Sx(M) → R |b ∈ M} is relatively weakly compact in C(Sx(M)). (Cf. Definition 2.1 above.) Remark 4.8. (i) The equivalence (i) ⇔ (ii) of Definition 4.7 is called the Eberlein– Grothendieck criterion (see [Kha19b]). In [KM81], stability is only defined for the formula kx+yk, and it is called stability for the whole space (and not NOP for the formula kx+yk). Definition 4.7(i) was based on Shelah general notion of NOP or stability of a formula. The general analogue of NOP in continuous setting appeared for the first time in [BU10]. (ii) A formula φ has NOP for a theory T (in the obvious sense) iff it has NOP on every model M of T . A theory has NOP (or is stable) if every formula has NOP for the theory. Definition 4.9. A Banach space (or Banach structure) M is stable (or has NOP) if the formula φ(x, y)= kx + yk is stable on M. The following is a consequence of the well-known compactness theorem of Eberlein and Smulianˇ (Fact 2.3). Indeed, if the set A above is relatively weakly compact in C(Sx(M)), then every sequence in A has a pointwise convergent subsequence in RSx(M), namely A is relatively sequentially compact in RSx(M). (Cf. Definition 2.2(i).) Fact 4.10 ([Kha19b], Fact 3.4). Let M be a structure, and φ(x, y) a formula. If φ is stable on M then φ has NIP on M. In particular, every stable Banach space has NIP.

As Lp[0, 1] (1 6 p < ∞) embeds ℓp, and the former is stable (cf. [BBHU08, Theo- rem 17.11]), so ℓp is stable. Therefore, by the above fact, ℓp (1 6 p< ∞) has NIP.

Tsirelson’s space Although we only use some known properties of Tsirelson’s example [Tsi74] or actually the dual space of the original example as described by Figiel and Johnson [FJ74], for the sake of completeness, we present Tsirelson’s space and remind some its properties which are used in this paper. The key property for our purpose will be presented in Fact 4.18. n A collection (Ei)1 of finite subsets of natural numbers is called admissible if n 6 E1 < E2 < ··· < En. By E < F we mean max E < min F and n 6 F means n 6 min F . For x ∈ c00 and E ⊆ N by Ex we mean the restriction of x to E, i.e. Ex(n)= x(n) if n ∈ E and 0 otherwise. Now we recall the definition MT (the Tsirelson of [FJ74]). For all x ∈ c00 set 1 n kxk = max kxk , sup kE xk :(E )n is admissible . ∞ 2 i i 1 i=1  n X o MT is then defined to be the completion of (c00, k·k). Note that this norm is given implicitly rather than explicitly. One can define this norm on c00 by induction. Indeed, for x ∈ c00, set kxk0 = kxk∞ and inductively 1 n kxk = max kxk , sup kE xk :(E )n is admissible . n+1 n 2 i n i 1 i=1  n X o 11 Then kxk = limn kxkn is the desired norm. Fact 4.11 ([AK16], Theorem 11.3.2). Tsirelson’s space has the following properties.

(i) MT is reflexive.

(ii) MT does not contain a subspace isomorphic to c0 or ℓp (1 6 p< ∞).

In addition, it is easy to verify that (en) is a normalized unconditional basis for MT and any spreading model of MT is isomorphic to ℓ1. See [AK16] for definitions and proofs. The following deep result, due to Krivine and Maurey [KM81], gives a partially answer to the main question of the present paper. First, recall that: Definition 4.12. Let X,Y be two Banach spaces. (i) X and Y are λ-isomorphic (for some λ > 1), if there is a linear isomorphism T from X onto Y such that for all x ∈ X, λ−1kxk≤kT (x)k ≤ λkxk. (ii) X and Y are called almost isometric if for each λ> 1, they are λ-isomorphic. (iii) X embeds Y almost isometrically, if for each λ> 1 there is a subspace Xλ of X such that Xλ and Y are λ-isomorphic.

Fact 4.13 (Krivine–Maurey theorem). Every (separable) stable Banach space embeds ℓp almost isometrically, for some 1 6 p< ∞.

Corollary 4.14. MT is not stable. Proof. Immediate by Facts 4.11 and 4.13.  A Banach space is weakly stable if the condition (i) in Definition 4.7 holds whenever (an) and (bm) are both weakly convergent. It is proved that every weakly stable space embeds c0 or ℓp, for some 1 6 p< ∞, almost isometrically (see [ANZ87]). To summarize, MT is not stable, even worse, it is not weakly stable.

Strong separability We will show that Tsirelson’s space has NIP. Indeed, we show that every (separable) Banach structure with strongly separable space of types has NIP. Then the desired result will be achieved from a deep result of Odell (Fact 4.18 below). Definition 4.15. We say a (separable) Banach space is strongly separable, or equivalently its type space is strongly separable, if (Sφ(M), ρ) is separable where φ(x, y)= kx + yk and ρ is the ρ-metric on φ-types. Recall that, by the Krivine–Maurey theorem, the Tsirelson’s space is not stable (see Corollary 4.14 above). By Odell’s result below, it easy to show that this space has a weaker property, namely NIP. For this, we recall some notion and fact. Recall that a function f form a metric space X to a metric space Y is k-Lipschitz (k > 0) if for all x, y in X, d(f(x), f(y)) 6 k · d(x, y). Now we give the following easy lemma.

12 Lemma 4.16. Assume that (X,d) is a metric spaces, and F a bounded family of 1- Lipschitz functions from X to R. If X is separable then every sequence in F has a (pointwise) convergent subsequence.

Proof. The proof is an easy diagonal argument. 

Observation 4.17. Assume that M is a separable Banach structure and φ(x, y)= kx+yk. If the type space Sφ(M) is strongly separable, then φ has NIP on M.

Proof. Recall that (Sφ(M),τ,ρ) is a compact topometric space where (Sφ(M), τ) is com- pact and Polish, and (Sφ(M), ρ) is a complete metric space. (Here, τ is the logic topology and ρ is the ρ-metric.) Note that ρ is separable since the type space is strongly separable. Also, the functions of the form φ(·, a), a ∈ M, are 1-Lipschitz with respect to ρ. Now, use Lemma 4.16 and the equivalence (i) ⇔ (v) of Lemma 4.2.  In [Ode82, Proposition 1], it is shown that the type space of a separable stable Banach spaces is separable with respect to the metric ρs (cf. Remark 3.5). For the unit ball of a separable and stable space M, this is actually a consequence of the separability of M and the definability of types in stable models (see [Ben14] and [Kha18a]). Edward Odell showed that the converse does not hold:

Fact 4.18 (Odell, [Ode82]). Tsirelson’s space is strongly separable. Furthermore, its type space in unbounded continuous logic is uniformly separable (see Remark 3.5 above).

Let M be a Banach space, φ(x, y) = kx + yk, X = Sφ˜(M) the space of complete φ˜-types on M. For a ∈ M, define φ(a, y): X → [0, 2] by q 7→ φ(a, q). (Notice that φ(a, y): X → [0, 2], a ∈ M, is continuous.) We say that the space of definable predicates on M is weakly sequentially complete if the set F = {φ(a, y): X → [0, 2] |a ∈ M} has the weak sequential completeness property (cf. Definition 2.2(ii)). The following observation is new, though it is really just a corollary of Odell’s deep result above.

Corollary 4.19. (i) Tsirelson’s space has NIP. (ii) The space of definable predicates on Tsirelson’s space is not weakly sequentially com- plete.

Proof. (i) Immediate, since the type space of Tsirelson’s space is strongly separable (see Fact 4.18 above). (ii) Note that, by the Eberlein–Smulianˇ theorem (Fact 2.3), NIP and the weak sequential completeness property is equivalent to stability. (See also Theorem 4.3 of [Kha19b].)  It seems to be open whether the theory of Tsirelson’s space has NIP.

Remark 4.20. Note that one can not expect a converse to Observation 4.17. There are some indications that the tree Banach spaces can be counterexamples. With the previous observations, we divide separable Banach spaces to some classes such as stable, containing

13 some ℓp (1 6 p < ∞), with strongly separable type space, with NIP, not containing c0, and these classes form a hierarchy

stable strongly separable ⊂? NIP ⊂? not containing c0

Two questions arise. Must a space not containing c0 have NIP? Must a space with NIP have strongly separable types? These questions seem to be open, although we know that the answer to one of these questions is certainly negative. Indeed, in the 90’s Gowers [Gow94] has constructed a space XG not containing ℓp and no reflexive subspace. By the result of Haydon and Maurey [HM86] which asserts that the spaces with strongly separable types contain ℓ1 or have a reflexive subspace, the type space of XG is not strongly separable and this space does not contain c0, so {strongly separable} {not containing c0}. Also, we strongly believe that the answers to both questions are negative. We will give sharper answers in a future work.

5 c0- and ℓp-subspaces of perturbed ℵ0-categorical spaces As we saw earlier the norm of Tsirelson’s space is given implicitly rather than explicitly. An easy observations (see Corollary 7.5 below) shows that the norm of Tsirelson’s space can not be characterized by axioms in continuous logic. This and the fact that Tsirelson’s space does not contain c0 or ℓp lead us to a natural question as Gowers and Odell asked: Must an “explicitly defined” Banach space contain c0 or ℓp? (See [Gow95, Gow09, Ode97].) Now, we give an explicit definition of a similar alternative of an “explicitly defined” Banach space, and give a positive answer to the above question.

Convention 5.1. In this section the language is the usual language of Banach spaces (in continuous logic). Otherwise, we explicitly state what is our desired language. c0- and ℓp-types Definition 5.2. If p ∈ [0, ∞) and ǫ > 0, the following set of statements with countable variablesx ¯ = (x0, x1,...) will be called the ǫ − ℓp-type: for every natural number n, and scalars r0,...,rn,

n n 1 n −1 p p (1 + ǫ) rixi 6 |ri| x0 6 (1 + ǫ) rixi . 0 0 0 X X
X

The following set of statements with countable variablesx ¯ =(x0, x1,...) will be called the ǫ − c0-type: for every natural number n, real number ǫ> 0 and scalars r0,...,rn,

n n −1 (1 + ǫ) rixi 6 max |ri| x0 6 (1 + ǫ) rixi . 06i6n 0 0 X
X

14 Remark 5.3. The notion of ǫ−ℓp-type (ǫ−c0-type) is a formal definition, and is different from the notion of types in model theory or Banach space theory. Recall that if T is a complete theory and there exists a model M of T such that the ǫ − ℓp-type (ǫ − c0-type) is realized in M, then it is a type in the sense of model theory, i.e. it belongs to Sx¯(∅). However, the following deep classical theorem, due to Krivine, guarantees that for every Banach space there exists some p such that the notion of ǫ − ℓp-type is really a type in the sense of model theory.

Krivine’s theorem We recall the well-known theorem of Krivine.

Definition 5.4. Let p ∈ [1, ∞] and (xi) a sequence of elements of some Banach space. We say that ℓp (resp. c0) is block finitely represented in (xi) if p ∈ [1, ∞) (resp. p = ∞) and for every ǫ> 0 and positive integer n, there are n + 1 finite subsets F0,...,Fn of the positive integers N with max Fj < min Fj+1, for all 0 6 j 6 n − 1 and elements y0,...,yn with yj in the linear span of {xi : i ∈ Fj} for all i so that for all scalars r1,...,rn,

n 1 n n 1 p p p p (1 − ǫ) |ri| y0 6 riyi 6 (1 + ǫ) |ri| y0 0 0 0 X
X X
n 1 p p (where ( 0 |ri| ) = sup |ri|, if p = ∞). In theP year in which Tsirelson’s example appeared in print [Tsi74], J.-L. Krivine [Kri76] publised his celebrated theorem.

Fact 5.5 (Krivine’s theorem). Let (xn) be a sequence in a Banach pace with infinite- dimensional linear span. Then either there exists a 1 6 p< ∞ so that ℓp is block finitely represented in (xn) or c0 is block finitely represented in (xn).

Main theorem In this section, we prove the main theorem of the paper; that is, any (suitable) perturbed ℵ0-categorical space contains some classical sequences c0 or ℓp. Let M, N be two Banach spaces and a ∈ M and b ∈ N. Recall from Definition 4.12 that, for ǫ> 0, a (1+ǫ)-isomorphic between (M, a) and (N, b) is a linear map f : M → N such that f(a)= b and kfk, kf −1k 6 1+ ǫ. Let us fix a complete theory T (in a countable language of Banach spaces) and a monster model U of it. In the following Sn(T ) denotes the set of n-types over ∅. Definition 5.6 (Banach–Mazur perurbation). The Banach–Mazur perturbation system

pBM (for theory T ) is the perturbations system defined by the Banach–Mazur distance

dBM in the sense of Definition 1.23 in [Ben08a], that is, for each ǫ> 0, pBM (ǫ)= {(p, q) ∈ 2 Sn(T ) : n<ω, dBM (p, q) < ǫ} where there is a (1+ ǫ)-isomorphism between (U, a) dBM (p, q) = inf log(1 + ǫ): ( and (U, b), and U  p(a), q(b) )

15 See also Iovino’s paper [Iov99].

Definition 5.7 (pBM -isomorphism). (i) Two separable models M, N are called pBM - isomorphic if for each ǫ > 0 there is a bijective map f : M → N such that M N n (tp (a), tp (f(a))) ∈ pBM (ǫ) for all a ∈ M , n<ω.

(ii) A theory T is called pBM -ℵ0-categorical if every two separable models M, N  T

are pBM -isomorphic. A separable Banach space is called pBM -ℵ0-categorical if its

continuous first order theory (in the usual language) is pBM -ℵ0-categorical.

Clearly, every two models of a pBM -ℵ0-categorical theory are almost isometric (cf. Definition 4.12 above). We want to give a general result to any perturbation system, so we generalizes the above notions. Let p be an arbitrary perturbation system in the sense of [Ben08a, Def- inition 1.23]. The notions p-isomorphism and p-ℵ0-categoricity are defined similar to

Definition 5.7, by replacing pBM with p. Similarly, a separable Banach space is called p-ℵ0- categorical if its continuous first order theory (in the usual language) is p-ℵ0-categorical. The following terminology is not standard, so we single it out. Before that, we recall some notations. If p(x) is any partial type and ǫ > 0 then p(xǫ) denotes the partial type ∃x′(p(x′) ∧ d(x, x′) ≤ ǫ). Ifx ¯ is a countable tuple of variables then p(¯xǫ) means ǫ ǫ p(x0, x1,...), where the metric is supermum metric on countable tuples. For a perturba- tion system p, the p(ǫ)-neighbourhood around p(¯x) is defined as follows: pp(ǫ)(¯x)= {q(¯x): (p(¯x), q(¯x)) ∈ p(ǫ)}. (Cf. Definition 1.1 of [Ben08a].) Definition 5.8 (p-saturation). Let p be a perturbation system. We will say that a structure M is p-saturated if for every type p(¯x) ∈ Sω(∅) (in countable variablex ¯), and ǫ> 0, the partial type pp(ǫ)(¯xǫ) is realised in M.

Notice that the notion p-saturation above is weaker than p-approximately ℵ0-saturation in the sense of [Ben08a, Definition 2.2]. Indeed, recall from Remark 2.4 of [Ben08a] that the latter notion can be defined for ω-types in Sω(∅) (i.e, types with ω-tuples). Definition 5.9 (Perturbed embedding). Let p be a perturbation system, M a Banach space and p(¯x) the 0 − ℓp-type (or 0 − c0-type). We say that M embeds ℓp (or c0) p- approximately if for every ǫ> 0, the partial type pp(ǫ)(¯xǫ) is realised in M.

Now we are ready to give the main result of the paper.

Theorem 5.10 (Main theorem). Let p be perturbation system and M p-ℵ0-categorical space. Then, M embeds c0 or ℓp (1 6 p < ∞) p-approximately. Moreover, M embeds ℓ2 p-approximately. In particular, if p is the Banach–Mazur perturbation system pBM , then M embeds ℓ2 almost isometrically.

Proof. Let M be a separable model of T . By Krivine’s theorem, some ℓp (or c0) is block finitely representable in M. So, for every ǫ > 0, the ǫ − ℓp-type is a type in Sω(T )(=Sω(∅)). (Note that in this case the ǫ − ℓp-type is a partial type, but we can

16 expand it to a complete type in Sω(T ).) As ǫ is arbitrary, the 0 − ℓp-type is a complete type. Let p(¯x) be this complete type. By Lemma 3.4 of [Ben08a], M is p-saturated. (Indeed, recall that the notion p-approximately ℵ0-saturation in [Ben08a, Definition 2.2] is stronger than p-saturation.) By the p-saturation, for every ǫ > 0, pp(ǫ)(¯xǫ) is realised in M, and so M embeds ℓp p-approximately. Moreover, it is known that ℓ2 is finitely representable in c0 and ℓp. (Indeed, in the case of c0, every Banach space is finitely representable in c0 (see [AK16, Example 12.1.2]). If ℓp is finitely representable, then so is Lp ([AK16, Proposition 12.1.8]), and since ℓ2 is isometric to a subspace of Lp ([AK16, Theorem 6.4.12]), ℓ2 is finitely representable.) So ǫ − ℓ2-type is a partial type in the sense of model theory, and so M embeds ℓ2 p-approximately.

In particular, if p is the Banach–Mazur perturbation system pBM , M embeds ℓp (or c0) pBM -approximately. That is, for each ǫ> 0 there is a sequencea ¯ ∈ M such that it realise p (ǫ) ǫ the partial type p BM (¯x ). This means that there is a sequence ¯b in an elementary exten- p (ǫ) sion of M such that d(¯a, ¯b) ≤ ǫ and |= p BM (¯b). So, by the definition, dBM (tp(¯b),ℓp) ≤ ǫ. As ǫ is arbitrary, there is a sequencea ¯ ∈ M (for ǫ> 0) such that dBM (tp(¯a),ℓp) ≤ ǫ. This means that there is a sequence in M which is (1 + ǫ)-isomorphic to ℓp (or c0). Again, as ℓ2 is finitely representable in c0 and ℓp, M embeds ℓ2 almost isometrically.  Remark 5.11. Recall that for two perturbation systems p, p′, the perturbation system p is stricter than p′, denoted by p < p′, if for each ǫ> 0 there is δ > 0 such that p(δ) ⊂ p′(ǫ). Clearly, if p < p′ and N, N are p-isomorphic then they are p′-isomorphic too. Therefore, if p be any perturbation system stricter than pBM , then the “almost isometrically” version of the above theorem holds for p.

Example 5.12. (i) Let T = T h(AN ⊆[1,r]) be the theory of Nakano spaces presented in [Ben09]. The theory T is not ℵ0-categorical, but it is λ-stable for λ > c (see Theo- rem 3.10.9 in [poi06]). Ben Yaacov proved that T is ℵ0-categorical and ℵ0-stable up to small perturbations of the exponent function. This perturbation system is stricter than pBM in the language of Banach spaces (see Proposition 4.6 in [Ben09]), and so the theory of Nakano spaces is pBM -ℵ0-categorical. (ii) (non-example) The theory of the 2-convexification of Tsirelson space, is denoted by

MT ,2, as presented by B. Johnson (in the language of Banach spaces) is not pBM -ℵ0- categorical because this space does not contain any ℓp or c0. It is known that every ultrapower of this space is linearly homeomorphic to a canonical direct sum of the space and a Hilbert space of suitable dimension. Indeed, see Proposition 2.4.a of [JLS96] for the proof; of course this depends on the argument of Pisier that is pointed to in [JLS96] in the paragraph just before Proposition 2.4.a (from the book [Pis89]). Ward Henson suggested the following argument that there are separable models N of the theory of MT ,2 such that N and MT ,2 ⊕ ℓ2 are isomorphic. Indeed, for an ultrafilter U, let H be a subspace of (MT ,2)U isomorphic to a Hilbert space for which (MT ,2)U is the direct sum ′ internally of MT ,2 and H (which is possible by the result pointed above). Let H be any separable, infinite dimensional subspace of H and let N be a separable infinite dimensional ′ elementary subspace of (MT ,2)U that contains MT ,2 and H . Letting H0 = Y ∩H it is easy to show that N = MT ,2 ⊕ H0 (again, the sum ⊕ is the one internal to (MT ,2)U ). Of course

17 this says essentially nothing about the precise way in which the norm on N = MT ,2 ⊕H0 is defined – and it seems likely that there will be a large number of different (non-isometric)

possibilities. To summarize, MT ,2 and MT ,2 ⊕ H0 are two non pBM -isomorphic separable models of this theory.

Definition 5.13. A separable Banach space is called ℵ0-categorical if it is pid-ℵ0-categorical, where pid is the strictest perturbation system. (See also Definition 7.1 below.)

Corollary 5.14. Every ℵ0-categorical Banach space M contains isometric copies of c0 or ℓp (1 6 p< ∞). Moreover, M contains an isometric copy of ℓ2. Proof. This is a consequence of [Ben08a, Lemma 2.5] and the main theorem (Theo- rem 5.10) assuming the identity perturbation system pid. Indeed, by Theorem 5.10, for pid(ǫ) ǫ every ǫ> 0, p (¯x ) is realised in M, where p(¯x) is the 0 − ℓp-type (or 0 − c0-type). By pid(ǫ) Lemma 2.5 of [Ben08a], for every ǫ> 0, p (¯x) is realised in M. As pid is the identity pid(ǫ) system, i.e. p = {p} for all ǫ > 0, there is a sequence in M which is isometric to ℓp (or c0). Similarly, M contains an isometric copy of ℓ2. 

Remark 5.15. (i) The spaces Lp[0, 1] (1 ≤ p < ∞) are ℵ0-categorical (see [BBHU08]). The Gurarij space (a universal, ultrahomogeneous Banach space) is also ℵ0-categorical (see [BH12]). (ii) Note that since the nature of our question, i.e. the existence of good subspaces, is local and ℵ0-categoricity is a global condition on norm, one can not expect a converse to Theorem 5.10. For example, the space MT ⊕ ℓp has a complex norm (since it contains the Tsirelson space MT ), although it contains a good subspace, i.e. ℓp. (iii) As we mentioned earlier, Corollary 5.14 is folklore to experts in the field. Ward Henson informed us that one can prove that every ℵ0-categorical space contains ℓ2 using Dvoretzky’s Theorem.

Remark 5.16. Note that sequence spaces ℓp (p 6= 2) and c0 are not ℵ0-categorical. Despite this, ℓp (p 6= 2) is almost categorical, i.e. its theory has exactly two separable models ℓp and Lp[0, 1] ⊕p ℓp. This fact is folklore to experts in field, including Alexander Berenstein, Ward Henson and Jose Iovino. We thank C. Ward Henson for communicating to us the following argument and references. The main point behind this fact is the classification of Lp(µ) spaces under elementary equivalence in Theorem 2.2 of [Hen76]. The proof has four steps. First, it was shown in [HM74, Corollary 2.5] that for a Banach space M with nonstandard hull M:

M is an Lp-space if and only if M is an Lp-space. c Therefore, if two Banach spaces ℓp and M have isometric nonstandard hulls, then M is c an Lp-space. (Recall that, by Kakutani representation theorem, every Lp-space M is of the form Lp(µ) for some measure µ.) Second, by the Keisler-Shelah Theorem for metric structures ([Hen76, Theorem 1.13]), two metric structures are elementarily equivalent iff they have linearly isometric ultrapowers iff they have linearly isometric nonstandard hulls. (We believe this theorem should be credited to C. Ward Henson [Hen76] and to Jacques

18 Stern [Ste78] equally and independently.) Using this theorem, if ℓp and M are elementary equivalent, then M is of the form Lp(µ) for some measure µ. Third, by Theorem 2.2 of [Hen76], if ℓp and Lp(µ) are elementary equivalent, then the measure space of µ has infinitely many atoms. Fourth, if M = Lp(µ) is separable and the measure space of µ has infinitely many atoms, then either (1) the atomless part of the measure space is {0} (so M is linearly isometric to ℓp), or (2) the atomless part of the measure space is isomorphic to the measure space of [0, 1] with Lebesgue measure (so M is linearly isometric to ℓp ⊕p Lp[0, 1]). (The latter statement is a well-known fact due to Carath´eodory [Lac74, Section 14, Theorem 5].) So putting everything together we see that the complete theory of ℓp (p 6= 2) has exactly two separable models ℓp and Lp[0, 1] ⊕p ℓp up to isometry. This fact suggests the following conjecture: Do all models of a theory with finitely many separable models contain some ℓp or c0? Still, one can say more:

Remark 5.17. (i) Let M be an ℵ0-categorical Banach space. Then M embeds (isomet- rically) every Banach space with a basis which is finitely representable in M. For this, repeat the argument of the proof of Theorem 5.10. (ii) The spectrum Σ(X) of a Banach space X consist of all p ∈ [1, ∞] for which X contains n ℓp ’s uniformly, that is, there exists K > 0 such that for all n, there is a subspace of X n which is K-isomorphic to ℓp . (See [Schw81, p. 57]) By Krivine’s theorem, the spectrum Σ(X) consist of all p ∈ [1, ∞] for which ℓp is finitely representable in X (see [Gue92], Theorem II.5.13). Dvoretzky’s theorem says that 2 ∈ Σ(X). It is known that Σ(X) is a closed subset of [1, ∞], and Σ(X) ∩ [1, 2] is always an interval; for example, Σ(Lp) = [p, 2] if 1 6 p 6 2, Σ(Lp) = {2,p} if 2 6 p < ∞, and Σ(L∞) = [1, ∞]. Therefore: Every ℵ0-categorical Banach space M contains isometric copies of ℓp for all p ∈ Σ(M).

Concluding remarks (1) As previously mentioned, Krivine and Maurey [KM81] proved that every stable Banach space contains a copy of ℓp for some 1 6 p < ∞. Also, it was noticed that the type space of every separable stable Banach space is strongly separable. Later, Haydon and Maurey [HM86] showed that a Banach space with strongly separable types contains either a reflexive subspace or a subspace isomorphic to ℓ1. (2) In [Cai17], Caicedo showed that for any suitable extension L of continuous logic with the compactness and the separable downward L¨owenheim-Skolem theoems, each sentence of L is equivalent to a sentence of continuous logic. On the other hand, using these two theorems (in continuous logic), one can give a proof of Corollary 5.14. Indeed, suppose that M is ℵ0-categorical. By Krivine’s theorem, some ℓp (or c0) is finitely representable in M, and so, using the compactness theorem, there is an elementary extension N of M such that it contains ℓp (or c0). By the separable downward L¨owenheim-Skolem theorem, ′ there is a separable elementary substructure M of N such that it contains ℓp (or c0). By ′ the ℵ0-categoricity, M and M are isometric, and so M contains some classical sequence space. As ℓ2 is finitely representable in ℓp and c0, it is easy to see that M contains ℓ2.

19 Therefore, by Caicedo’s result, the above argument does not work in a logic stronger than continuous logic. (3) There is still another stronger property: a separable Banach space X is said to be finitely determined if for each separable space Y such that X is finitely representable (f.r.) in Y and Y is f.r. in X then Y is isometric to X. Clearly, every finitely determined space is ℵ0-categorical, but not converse. In [Kha18b], a more direct proof of existence of classical sequence spaces in finitely determined spaces is given.

(4) One might expect that ℵ0-categoricity implies that a large number of ℓp, p ∈ [1, ∞), or c0, are involved in the space. This is not true; for example, the theory ℓ2 ∼= L2 = L2([0, 1],λ) is ℵ0-categorical (see [BBHU08]), but ℓ2 does not have any subspace isomor- phic to c0 or ℓq for all q 6= 2 (see [AK16], Corollary 2.1.6). In other words, a direct proof of Theorem 5.10, without using Krivine’s Theorem, does not imply a stronger result. (5) As previously mentioned, Theorem 1.1 of [Iov05] asserts that the existence of enough definable types guarantees the existence of ℓp subspaces. Note that our approach in this paper is different. In fact we counted separable models and the main result (Theorem 5.10) is about complete theories.

(6) Although ℵ0-categoricity and ℵ0-stability do not have any connection in general, but to our knowledge the most examples of ℵ0-stable theories are studied in continuous model theory are ℵ0-categorical (see [BBHU08]). (7) (A Krivine-Maurey type theorem). We believe that the following statement holds: For any separable NIP space X there exists a spreading model of X containing c0 or ℓp for some 1 6 p< ∞. The proof used Borel definability of types and Krivine’s Theorem. This result provides answers to the questions in Remark 4.20. Full proof will be presented elsewhere.

6 Dividing lines in Banach spaces and model theory

This section aims to provide a classification of Banach spaces similar to Shelah’s classi- fication in classical logic. For the sake of completeness, we recall some facts which were observed in [Kha18a], and then we use them to give some examples which are (in our view) very illuminating.

Banach space for a formula Let M be an L-structure, φ(x, y): M × M → R a formula (we identify formulas with real-valued functions defined on models). Let Sφ(M) be the space of complete φ-types over M and set A = {φ(x, a), −φ(x, a) ∈ C(Sφ(M)) : a ∈ M}. The (closed) convex hull of A, denoted by (conv(A)) conv(A), is the intersection of all (closed) convex sets that contain A. conv(A) is convex and closed, and kfk 6 kφk for all f ∈ conv(A). So, by normalizing we can assume that kfk 6 1 for

all f ∈ conv(A). Set B = conv(A) and V = λ>0 λB. It is easy to verify that V is a S 20 Banach space with the normalized norm and B is its unit ball. This space will be called the space of linear φ-definable relations on M.

Fact 6.1 ([Kha18a]). Assume that φ(x, y), M, B and V are as above. Then the following are equivalent:

(i) φ is stable on M.

(ii) B is weakly compact.

(iii) The Banach space V is reflexive.

Recall that for an infinite compact Hausdorff space X, the space C(X) is not reflexive.

Fact 6.2 ([Kha18a]). Assume that φ(x, y), M, and V are as above. Then the following are equivalent:

(i) φ has NIP on M.

(ii) V is Rosenthal Banach space.

(iii) V does not contain an isomorphic copy of ℓ1. Definition 6.3. Let M be a model (of theory T ) and φ(x, y) a formula. We say φ(x, y) has the strict order property on M (short SOP on M) if there exists a sequence (aibi : i<ω) in M and ǫ> 0 such that for all i < j,

φ(M, ai) 6 φ(M, ai+1) and φ(bj, ai)+ ǫ<φ(bi, aj).

The acronym SOP stands for the strict order property and NSOP is its negation. A complete theory T has SOP if there are a formula φ and a model M of it such that φ has SOP on M, and otherwise it is said that T has NSOP.

Remark 6.4. (i) Let U be a monster model (of theory T ), φ(x, y) a formula and M a small model. Suppose that φ has SOP on M. Then it is easy to verify that φ has SOP on the moster model U. (Indeed, suppose that p is a φ-type over M, (bj) ∈ M and tp(bj/M) → p. By SOP on M, φ(bj, ai) 6 φ(bj, ai+1), and since φ(x, ai)’s are continuous, so φ(b, ai) 6 φ(b, ai+1) for all b |= p. This means that φ(b, ai) 6 φ(b, ai+1) for all b ∈ U.) (ii) SOP strictly implies instability. In classical ({0, 1}-valued) logic, this is a known fact. (iii) Suppose that M = U and V is as above. If V is weakly sequentially complete, i.e. every weak Cauchy sequence has a weak limit (in V ), then φ has NSOP. This is a consequence of the Eberlein–Smulianˇ theorem (Fact 2.3). (Cf. [Kha19a, Remark 3.4(ii)].) (iv) Recently in [Kha19a] it is shown that the converse of (iii) above does not hold (for classical logic). That is, if φ has NSOP, we cannot conclude that V is weakly sequentially complete (cf. [Kha19a, Example 3.5(ii)]). On the other hand, we strongly believe that there is no perfect analog of Shelah’s theorem (i.e. a complete theory is stable if and only if it has NIP and NSOP) in continuous logic. This is discussed in detail in [Kha21].

21 For the sake of completeness we give the diagram of these observations for classical logic [Kha19a]:

Shelah Stable ⇐⇒ NIP & NSOP

m Eberlein–Grothendieck ⇓ ⇑

Eberlein–Smulianˇ Reflexive ⇐⇒ Rosenthal & WS-complete

Dividing lines in classical spaces In this subsection, we list some classical Banach spaces and some theories and point out their model theoretic dividing lines. Many of the following examples have been studied previously (mostly by Henson and co-authors in the 70’s), and much of what is said about them is well-known (see [HI02] and references therein). Nevertheless, there is a definite merit to gathering all of these facts together in one place. It should be noted that since the notions NIP in a model (in both classic and continuous logics) and SOP for continuous logic are new, the following observations on these notions do not appear somewhere in literature. In fact, in our view, the examples form the most interesting and illuminating part of the article.

Example 6.5. (i) Lp Banach lattices (1 6 p < ∞) are reflexive and so Rosenthal and weakly sequentially complete. By Kakutani representation theorem of L-spaces ([AB85, 4.27]), it was proved that the classes of Lp Banach lattices are axiomatiz- able (in a suitable language) and their theories, denoted by ALpL, are stable (see [BBHU08], Section 17).

(ii) L1 Banach lattices are neither reflexive nor Rosenthal in general (e.g. ℓ1). Al- though, they are weakly sequentially complete. Nevertheless, their theory, AL1L, is stable (see (i) above). However, it is not established in some extensions of the language. For example, Alexander Berenstein showed that L1(R) with convolution is unstable (see [Ber]), later did it appear that ℓ1(Z, +) with convolution is unsta- ble (see [FHS12a], Proposition 6.2). In fact it has SOP (see Example 6.6 below). On the other hand, since the theory atomless probability measure algebra, denoted by AP A, is interpretable in the theory L1 Banach lattices, the prior also is stable ([BBHU08]).

(iii) Hilbert spaces are reflexive. The theory infinite dimensional Hilbert spaces, as described in [BBHU08], is stable.

22 (iv) c0 is neither reflexive nor weakly sequentially complete, but it is Rosenthal. Let ∞ ∞ (en)1 be standard basis for c0 and sn = e1 + ··· + en. (sn)1 is called summing basis for c0. Then kem + snk = 2 if m 6 n and = 1 if m > n. So, the formula ∞ ∗ φ(x, y) = kx + yk has order property. Now let x = (an)1 ∈ ℓ1 = (c0) then ∗ ∞ ∞ x (sn) → 1 ak. So(sn)1 is a weak Cauchy sequence with no weak limit. It ∗ ∗ is trivial that c0 is Rosenthal, because (c0) = ℓ1 is separable but (ℓ1) = ℓ∞ is P nonseparable. Let ψ(x, sn) = max(kx + snk, kx − snk). Then ψ(x, sm) 6 ψ(x, sn) and ψ(en,sm)+1 6 ψ(em,sn) for all m < n. So ψ(x, y) has SOP on c0.

By Example 4.5 above, φ(x, y) has IP on c0. Therefore, if c0 is a model of a theory T , then T has IP. (Also, see Example 8.8 for more information.)

(v) ℓ∞ is neither stable nor Rosenthal or weakly sequentially complete because c0 and ℓ1 live inside ℓ∞ = C(βN) (see below). The formula ψ as described above has SOP in ℓ∞. This fact explains why the case p = ∞ is excluded from (i) above. Since c0 has IP, ℓ∞ has IP. (vi) C(X, R) space for an (infinite) compact Hausdorff space is neither reflexive nor Rosenthal nor weakly sequentially complete (see (vii) below). By Kakutani repre- sentation theorem of M-spaces, the class of C(X)-spaces is axiomatizable in the language of Banach lattices. Indeed, replace the axioms of abstract LP -spaces in [BBHU08] by the M-property kx+ + y+k = max(kx+k, ky+k). Now we add a con- stant symbol 1 to our language as an order unite, i.e. for all x, kxk = 1 implies + . + that x 6 1; equivalently, supx((kxk − 1) −kx ∨ 1 − 1k). Note that k·k is an order unite norm if for all x and α, x+ 6 α1 implies that kxk 6 α; equivalently + . supx(kx ∨ α1 − 1k − (kxk − α)). Since ℓ∞ has SOP, the theory of C(X)-spaces has SOP. Since ℓ∞ has IP, then the theory of C(X)-spaces has the independence property.

(vii) C0(X, C) for an (infinite) locally compact space X is neither reflexive nor Rosenthal nor weakly sequentially complete. Note that, by Gelfand-Naimark representation ∗ theorem of Abelian C -algebras, the class C0(X, C)-spaces is axiomatizable in a suitable language (see [FHS12a] for the noncommutative theory). First, it is known that all Banach spaces, in particular ℓ1 and c0, live inside C(K)-spaces where K’s are compact. So C(K) is neither reflexive nor Rosenthal or weakly sequentially complete in general. Second, by an example of [FHS12a], we will show that the theory C0(X)-spaces has SOP and IP. Indeed, in X find a sequence of distinct xn that converges to some x (with x possibly in the compactification of X). For each n find a positive an ∈ C0(X) such that kank = 1 and an(xi) = 1 if i 6 n and = 0 if i > n. By replacing an with maximum of aj for j 6 n, we may assume that this sequence is increasing. Also, we can assume that the support Kn of each an is compact and am(Kn) = 1 for n < m. So aman = an if n 6 m. Now let φ(x, y)= k(x − 1)yk. Then φ(x, ai) 6 φ(x, aj) and φ(aj, ai)+1 6 φ(ai, aj) for i < j. So φ has SOP. Since c0 has IP, this theory has IP.

23 (viii) C∗-algebras are neither reflexive nor Rosenthal nor weakly sequentially complete be- cause some infinite-dimensional Abelian ∗-subalgebra lives inside a C∗-algebra and so by Gelfand-Naimark theorem it has a C0(X)-subalgebra (see [FHS12a], Lemma 5.3). Note that the Gelfand-Naimark-Segal theorem ensure that this class is axiomatiz- ∗ able ([FHS12a]). So, the theory of C -algebras has SOP. Since c0 has IP, this theory has IP.

(ix) The class of (Abelian) tracial von Neummman algebra is axiomatizable (see [FHS12b]). A tracial von Neumman algebra is stable if and only if it is of type I. The theory of Abelian tracial von Neumann algebras is stable because it is interpretable in the theory of probability measure algebras (see above and also [FHS12a], Lemma 4.5).

To our knowledge the following observation itself does not appear somewhere in liter- ature.

Example 6.6 (ℓ1(Z, +) with convolution has SOP). In [FHS12b] it is shown that the formula φ(x, y) = infkzk61 kx ∗ z − yk has order property in ℓ1(Z, +). We show that φ also has SOP. Indeed, we need to show that: (i) φ(xi,y) 6 φ(xj,y) for i 6 j, and (ii) φ(xi, xi) < φ(xj, xi) for i < j, where xi’s are elements of ℓ1 as described in [FHS12b, Proposition 6.2]. It is easy to verify that (ii) holds. We check that (i) also is true. Indeed,

for any z, kzk 6 1 and any y, we have |kyk−kzk| 6 supt |z(t) − y(t)| 6 kyk + 1 where k·k is uniform norm on C(T). So, if we assume that kyk > 1, then infkzk61 kz −yk = |kyk−1|. (Note that we can replace y with max(y, 1) and so assume that kyk > 1. In fact we define 2i φ(x, y) = infkzk61 kx ∗ z − max(y, 1)k.) Therefore if i 6 j, then infkzk61 k(cos t) z − yk 6 2j infkzk61 k(cos t) z − yk. Equivalently, φ(xi,y) 6 φ(xj,y). To summarize, φ has SOP. A question arises: Does ℓ1(Z, +) with convolution have NIP? The following interesting fact affirms some of our observations.

Fact 6.7 ([Gue92], Theorem III.5.1). Every stable (separable) Banach space is weakly sequentially complete.

So, every Banach space containing c0 is unstable. This directly implies that the theories ∗ C(X)-spaces, C0(X, C)-spaces, and C -algebras are not stable (see (vi)–(viii) above). (There is even something stronger (see Proposition 4.6).) Of course, the converse does not hold, e.g. Tsirelson’s space and its dual are unstable but they are reflexive and so weakly sequentially complete.

Remark 6.8. In [Mar97] the author proved that for 1 6 p < ∞, p 6= 2, the non- comutative Lp space Lp(M) is stable iff M is of type I. (See also [JL03, p. 1479].) Later, in [FHS12a] the authors showed that a tracial von Neumman algebra is stable if and only if it is of type I. Two questions arise: Which tracial von Neumman algebra (or non-comutative Lp spaces) have NIP? Which of them have NSOP?

24 7 Appendix A: ℵ0-categoricity, continued

In this appendix, we revisit the notion ℵ0-categoricity. Some observations are not new for model theorists, but one reason for recalling them is to make the paper more accessible to other interested readers. Recall that the density character of a topological space X, is the least infinite cardinal number of a dense subset of X. When measuring the size of a structure we will use its density character (as a metric space), denoted kMk, rather than its cardinality. Similarly, for a separable structure M, since Sφ(M) is a metric space, we measure the size Sφ(M) by its density character kSφ(M)k.

Definition 7.1. A complete theory T in a countable language is ℵ0-categorical (or ω- categorical) if it has an infinite model and any two models of size ℵ0 are isomorphic. Equiv- alently, T has an infinite model and separable models are isometric. An ℵ0-categorical (or ω-categorical) structure is a separable structure M whose theory is ℵ0-categorical. (Com- pare with Defintition 5.13.) Let L be a language and T a complete L-theory. Suppose that M is a model of T and A ⊆ M. Denote the L(A)-structure (M, a)a∈A by MA, and set TA to be the L(A)-theory of MA. Remark 7.2. It is easy to check that for a separable model M, and countable subsets A ⊆ B of M, if TB is ℵ0-categorical then TA is also ℵ0-categorical (see [BBHU08], Corol- lary 12.13); however, the converse does not hold in a strong form (see Remark 12.14 in [BBHU08]). The following fact is folklore, and we remove the proof. Recall that a formula φ is stable for a theory T if for every model M of T , φ is stable on M; equivalently, φ has not the order property on every model M of T . Fact 7.3 ([Ben14], Corollary 4). Let T be a countable theory and φ(x, y) a formula. Then the following are equivalent.

(i) For every separable model M of T , Sφ(M) is strongly separable. (ii) φ is stable on every separable model of T .

(iii) φ is stable for T .

(iv) For every model M of T , kSφ(M)k 6 kMk.

Corollary 7.4. Let M be an ℵ0-categorical structure. The complete theory T of M is stable if and only if for every formula φ, the space Sφ(M) is strongly separable.

Proof. If T is stable, every φ-type is definable ([Ben14, Corollary 4]), and so Sφ(M) is strongly separable. The converse follows from the ℵ0-categoricity and the equivalence (i) ⇔ (iii) of Fact 7.3 above. 

25 The fact that Tsirelson’s space can not be ℵ0-categorical is a consequence of Corol- lary 5.14. Alternatively, it is clear by instability and the above fact:

Corollary 7.5. Suppose that T is the complete theory of Tsirelson’s space MT (in a countable language). Then the followings hold: (i) T is not ℵ0-categorical. (ii) There exists a separable model M of T which its type space is not strongly separable.

Proof. (i): Suppose, if possible, that T is ℵ0-categorical. Then, since Sφ(M) is ρ-separable where φ(x, y)= kx + yk, so φ(x, y) is stable on MT . By Corollary 4.14, this is a contra- diction. (ii): Immediate from Fact 7.3 and Corollary 4.14. (Clearly, this model is different from Tsirelson’s space by Fact 4.18.) 

Remark 7.6. Note that the ℵ0-categorical assumption in Corollary 7.4 is too strong and a weaker assumption is sufficient. Indeed, suppose that M is a pBM -saturated structure, where pBM is the Banach–Mazur perturbation system. (See Definitions 5.6, 5.8 above.) Then, it is easy to show that, M is stable if and only if for every formula φ the space Sφ(M) is strongly separable.

8 Appendix B: A remark on Rosenthal’s dichotomy

Note that a well-known result of Rosenthal (Lemma 8.3 below) is used in the proof of the direction (i) ⇒ (v) of Lemma 4.2 (see Lemma 3.12 in [Kha19b]). So it is worthwhile to study it more carefully.

Definition 8.1. (i) A sequence {fn} of real valued functions on a set X is said to be independent if there exist real numbers s

−1 −1 fn (−∞,s) ∩ fn (r, ∞) 6= ∅ ⊠ n\∈P n\∈M for all finite disjoint subsets P, M of N. A family F of real valued functions on X is called independent if it contains an independent sequence; otherwise it is called strongly (or completely) dependent.

(ii) We say that the sequence {fn} is strongly (or completely) independent if ⊠ holds for all infinite disjoint subsets P, M of N. A family F of real valued functions is called strongly (or completely) independent if it contains a strongly independent sequence; otherwise it is called dependent. It is an easy exercise to check that when X is a compact space and functions are continuous the above two notions are the same, but this does not hold in general (see Example 8.8 below). Lemma 8.2. Let X be a compact space and F ⊆ C(X) a bounded subset. Then the following conditions are equivalent:

26 (i) F is dependent.

(ii) F is strongly dependent.

Lemma 8.3 (Rosenthal’s lemma). Let X be a compact space and F ⊆ C(X) a bounded subset. Then the following conditions are equivalent:

(i) F does not contain an independent subsequence.

(ii) Each sequence in F has a convergent subsequence in RX .

Rosenthal [Ros74] used the above lemma for proving his famous ℓ1 theorem: a sequence in a Banach space is either ‘good’ (it has a subsequence which is weakly Cauchy) or ‘bad’ (it contains an isomorphic copy of ℓ1). We will shortly discuss this topic (see below).

Fact 8.4 ([Ros74]). If X is a compact space and {fn} a pointwise bounded sequence in C(X), then

(1) either {fn} has a (pointwise) convergent subsequence, or

(2) {fn} has a ℓ1-subsequence, equivalently, {fn} has an independent subsequence. Note that for non-compact spaces, Fact 8.4 is not a dichotomy. Michael Megrelishvili informed us with details about this observation (see Example 4.5 above and Example 8.8 below) and also pointed out to an Example which is in [Dul89]. Of course, we are not sure that the observation itself (that Fact 8.4 is not always dichotomy in noncompact Polish case) appears somewhere in the literature in a clear form. For compact space, independence and strong independence are the same.

Fact 8.5 (Rosenthal’s ℓ1-theorem). If (xn) is a bounded sequence in a Banach space V then

′ (1) either (xn) has a weakly Cauchy subsequence, or

′ (2) (xn) has a ℓ1-subsequence. Fact 8.5 is a consequence of Fact 8.4. Indeed, let B∗ be the unit ball of V ∗, and define ∗ ∗ ∗ ∗ ∗ fn : B → R by x 7→ x (xn) for all x ∈ B .

Fact 8.6 (Rosenthal’s dichotomy, [Tod97]). If X is a Polish space and {fn} a pointwise bounded sequence in C(X), then

′′ (1) either {fn} has a (pointwise) convergent subsequence, or

′′ X (2) {fn} has a subsequence whose its closure in R is homeomorphic to βN, equivalently, it has a strong independent subsequence.

27 Remark 8.7. Note that (1)′ ⇒ (1)=(1)′′, but (1)′′ ; (1)′ in general. Also, (2)′′ ⇒ (2)′, but (2)′ ; (2)′′ (see below). In fact (2)′ is equivalent to independence property. For Polish ′′ space X, (2) holds iff there is a compact subset K ⊆ X such that {fn|K} contains a strong independent subsequence, equivalently, {fn} has a strong independent subsequence (see [BFT78], Lemma 2B, Theorem 2F and Corollary 4G).

Example 8.8. We revisit Example 4.5. Recall that the family Fc0 (in Example 4.5) con- Bc0 tains an independent sequence. (i) The family Fc0 is weakly precompact in R . Indeed, since c0 is separable and the functions fa are 1-Lipschitz, so by a diagonal argument (see Lemma 4.16), every sequence has a point-wise convergent subsequence. This shows that Fact 8.4 is not a dichotomy for non-compact spaces. (ii) By Fact 8.5, the sequence (fn) has not a weakly Cauchy subsequence. (iii) By Fact 8.6, (fn) has not a strong independent subsequence.

Remark 8.9. (i) Define fn : Bc0 → [0, 2] by fn(x) = kx + snk. Then fn converges to the continuous function f(x)=1+ kxk. Now, let c0 be the Cech-Stoneˇ compactification ˆ ˆ ˆ ˆ of c0 and fn, f be the extensions of fn, f respectively. Then fn 9 f. Indeed, note that ˆ ˆ ˆ ˆ fn|c0 = fn and f|c0 = f, and f(e) = limm f(em) =b limm f(em) = limm limn fn(em)=2 6= 1 = limn limm fn(em) = limn limm fˆn(em) = limn fˆn(e) where (em) is the standard basis of c0 and e is a cluster point of (em) in c0. Note that c0 is dense in c0 and fn → f, but fˆn 9 fˆ. (ii) Define gn(x) = kx + enk. Then gn → g where g(x) = max(kxk, 1). But (ˆgn) does not contain a convergent subsequence;b because (gn) has independenceb property (see Example 8.8 above). (iii) (gn) is a poinwise convergent sequence but it has not a weakly Cauchy subsequence. Because, by Rosenthal ℓ1-theorem, either a sequence has a weakly Cauchy subsequence, or it has a ℓ1-subsequence (equivalently, it contains an inde- pendent subsequence). Note that weakly Cauchy is stronger than pointwise convergence, in general. But, for locally compact spaces, they are equivalent.

References

[AK16] F. Albiac and N.J. Kalton, Topics in Banach space theory, Graduate Text in Mathematics 233, Springer, Second Edition, 2016.

[AB85] C. D. Aliprantis and O. Burkinshaw, Positive Operators (Academic Press, New York, 1985).

[ANZ87] S. Argyros, S. Negrepontis and Th. Zachariades, Weakly stable Banach spaces, Israel J. Math. 57 (1987), 68-88.

[Ben14] I. Ben-Yaacov, Model theoretic stability and definability of types, after A. Grothendiek, Bulletin of Symbolic Logic, (2014), 20, pp 491-496.

[Ben09] I. Ben-Yaacov, Modular functionals and perturbations of Nakano spaces, J. Log. Anal. 1 (2009), Paper 1, 42 pp.

28 [Ben08a] I. Ben Yaacov, On perturbations of continuous structures, Journal of Mathe- matical Logic 8 (2008), no. 2, 225-249

[Ben03] I. Ben Yaacov, Positive model theory and compact abstract theories, Journal of Mathematical Logic 3 (2003), no. 1, 85-118. MR1978944 (2004c:03040)

[Ben08b] I. Ben-Yaacov, Topometric spaces and perturbations of metric structures, Logic and Analysis 1 (2008), no. 3-4, 235-272.

[BBHU08] I. Ben-Yaacov, A. Berenstein, C. W. Henson, A. Usvyatsov, Model theory for metric structures, Model theory with Applications to Algebra and Analysis, vol. 2 (Z. Chatzidakis, D. Macpherson, A. Pillay, and A. Wilkie, eds.), London Math Society Lecture Note Series, vol. 350, Cambridge University Press, 2008.

[BH12] I. Ben-Yaacov and C. W. Henson, Generic orbits and type isolation in the Gurarij space, Fundamenta Mathematicae, 237, no. 1, 47-82, 2017.

[BU10] I. Ben-Yaacov, A. Usvyatsov, Continuous first order logic and local stability, Transactions of the American Mathematical Society 362 (2010), no. 10, 5213- 5259.

[Ber] A. Berenstein, On adding the convolution to L1(R), unpublished, http://matematicas.uniandes.edu.co/∼aberenst/convolution.pdf

[BFT78] J. Bourgain, D. H. Fremlin, and M. Talagrand. Pointwise compact sets of baire-measurable functions. American Journal of Mathematics, 100(4):pp. 845-886, 1978.

[Cai17] X. Caicedo. Maximality of continuous logic, Beyond First Order Model Theory (J. Iovino, ed.) Chapman and Hall/CRC Press, 2017.

[CK66] C. C. Chang, H. J. Keisler, Continuous Model Theory. Princeton University Press, Princeton (1966).

[Dul89] D. van Dulst, Characterizations of Banach spaces not containing ℓ1. Centrum voor Wiskunde en Informatica, Amsterdam, 1989.

[FHS12a] I. Farah, B. Hart and D. Sherman, Model theory of operator algebras I: Sta- bility, Bulletin of London Math. Soc., 45 (2013) 825-838.

[FHS12b] I. Farah, B. Hart and D. Sherman, Model theory of operator algebras II: Model Theory, Israel J. Math. 201 (2014) 477-505.

[FJ74] T. Figiel and W. B. Johnson. A uniformly convex Banach space which contains no lp. Compositio Math., 29:179-190, 1974. [Fre04] D. H. Fremlin, Measure Theory, vol.3, (Measure Algebras, Torres Fremlin, Colchester, 2004).

29 [Fre06] D. H. Fremlin, Measure Theory, vol.4, (Topological Measure Spaces, Torres Fremlin, Colchester, 2006).

[Gow94] W. T. Gowers. A Banach space not containing c0, ℓ1 or a reflexive subspace, Transactions of the American Mathematical Society 344 (1994) no. 1, 407-420.

[Gow95] W. T. Gowers. Recent results in the theory of infinite-dimensional Banach spaces. In Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Z¨urich, 1994), pages 933-942, Basel, 1995. Birkh

[Gow09] W. T. Gowers. https://gowers.wordpress.com/2009/02/17/must-an- explicitly-defined-banach-space-contain-c0-or-ℓp/ [Gue92] S. Guerre-Delabri`ere, Classical Sequences in Banach Spaces, Dekker, New York-Basel-Hong Kong, 1992.

[HM86] R. Haydon and B. Maurey. On Banach spaces with strongly separable types. J. London Math. Society, 33:484-498, 1986.

[Hen76] C. W. Henson. Nonstandard hulls of Banach spaces. Israel J. Math., 25(1- 2):108-144, 1976.

[HI02] C. W. Henson and J. Iovino, Ultraproducts in analysis pp. 1-110 in Analysis and Logic (Mons, 1997), vol. 262 of London Mathematical Society Lecture Note Series, Cambridge University Press, Cambridge, 2002. Zbl 1026.46007. MR 2004d:03001. 21, 22

[HM74] C. W. Henson and L C Moore, Jr.: Nonstandard hulls of the classical Banach spaces, Duke Math. J. 41 (1974), 277-284.

[HM83] C. W. Henson and Jr. Moore, L. C. The Banach spaces ℓp(n) for large p and n. Manuscripta Math., 44(1-3):1-33, 1983.

[Iov05] J. Iovino. Definable types over Banach spaces. Notre Dame Journal of Formal Logic, Volume 46, Number 1, 19-50, 2005.

[Iov99] J. Iovino, Stable Banach spaces and Banach space structures, I: Fundamentals, In Models, algebras, and proofs (Bogot´a, 1995), Lecture Notes in Pure and Appl. Math., vol. 203, Dekker, New York, 1999, pp. 77-117.

[JL03] W.B. Johnson, J. Lindenstrauss, Handbook of the geometry of Banach spaces, volume II, North Holland; 1 edition (May 20, 2003).

[JLS96] W.B. Johnson, J. Lindenstrauss and G. Schechtman, Banach spaces deter- mined by their uniform structures, GAFA, Volume 6, Issue 3, pp 430-470 (May 1996)

30 [Kri76] J.-L. Krivine. Sous-espaces de dimension finie des espaces de Banach r´eticul´es. Ann. of Math., 104:1-29, 1976.

[KM81] J.-L. Krivine and B. Maurey, Espaces de Banach stables. Israel J. Math., 39(4):273-295, 1981.

[Kha18a] K. Khanaki, Correspondences between model theory and Banach space theory, unpublished note, (on arXiv 2015) arXiv:1512.08691v2, 2018.

[Kha19a] K. Khanaki, Dividing lines in unstable theories and subclasses of Baire 1 func- tions, arXiv:1904.09486v6, 2019.

[Kha21] K. Khanaki, On classification of continuous first order theories, submitted.

[Kha18b] K. Khanaki, Remarks on Banach spaces determined by their finite dimensional subspaces, unpublished note, arXiv:1804.08446v1, 2018.

[Kha19b] K. Khanaki, Stability, the NIP, and the NSOP: Model Theoretic Properties of Formulas via Topological Properties of Function Spaces, Math. Log. Quart. 66, No. 2, 136-149 (2020) / DOI 10.1002/malq.201500059

[KP18] K. Khanaki, A. Pillay, Remarks on NIP in a model, Math. Log. Quart. 64, No. 6, 429-434 (2018) / DOI 10.1002/malq.201700070

[Lac74] H.E. Lacey, The isometric theory of classical Banach spaces, Grundlehren der mathematischen Wissenschaften, 1974

[LT77] J. Lindenstrauss, L. Tzafriri, Classical Banach Spaces. I. Sequence Spaces (Springer, Berlin, 1977)

[Mar97] J. Marcolino, La stabilit´edes espaces Lp non-commutatifs, Math. Scand. 81 (1997), 212-218.

[Ode97] E. Odell. On Subspaces, Asymtotic structure, and Distortion of Banach Spaces; Connections with Logic, pp. 189-264 in Analysis and Logic (Mons, 1997), vol. 262 of London Mathematical Society Lecture Note Series, Cam- bridge University Press, Cambridge, 2002. Zbl 1026.46007. MR 2004d:03001. 21, 22.

[Ode82] E. Odell, On the types in Tsirelson’s space., In Texas functional analysis semi- nar 1982-1983 (Austin, Tex.), Longhorn Notes, pages 49-59, Austin, TX, 1983. Univ. Texas Press.

[Pis89] G. Pisier, The Volume of Convex Bodies and Banach Space Geometry, Cam- bridge University Press, 1989.

[poi06] L. Pedro Poitevin, Model theory of Nakano spaces, Ph.D. thesis, University of Illinois at Urbana-Champaign, 2006.

31 [Ros74] H. P. Rosenthal, A characterization of Banach spaces containing ℓ1, Proc. Nat. Acad. Sci. U.S.A. 71 (1974), 2411-2413.

[Schw81] J.T. Schwartz, Geometry and Probability in Banach Spaces, Lecture Notes in Math. 852, Springer, Berlin-Heidelberg-New York, 1981.

[Ste78] J. Stern , Ultrapowers and local properties of Banach spaces, Trans. Amer. Math. Soc. 240 (1978), 231-252

[Tod97] S. Todorcevic, Topics in topology, Lecture Notes in Mathematics 1652, Springer, 1997.

[Tsi74] B. S. Tsirel’son. It is impossible to imbed lp or c0 into an arbitrary Banach space. Funkcional. Anal. i Priloˇzen., 8(2):57-60, 1974.

[Usv] A. Usvyatsov, Special stable type and ℓp subspaces of a Banach space, Incom- plete manuscript.

[Whi67] R. Whitley, An elementary proof of the Eberlein–Smulianˇ theorem, Mathe- matische Annalen, vol. 172 (1967), pp. 116-118.

32