Eventual uniqueness of large models, without a prescribed threshold

← All digests

Family 240 · Mathematical logic
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-21

1. The question

Some axioms determine a mathematical structure completely once its size is specified. For example, vector spaces over a fixed field are classified by the cardinality of a basis. Model theory asks when a similar uniqueness principle holds for classes of structures described by much more general rules.

A class is categorical in a cardinal if it has exactly one model of that size up to isomorphism. Shelah’s eventual categoricity conjecture predicts that uniqueness at one sufficiently large size forces uniqueness at every size beyond a uniform threshold.

Family 240 contains two manuscripts dated 24 September 2026. M2 claims the qualitative conjecture in ZFC, with a threshold depending only on a measure of the class’s complexity. M1 constructs, assuming the continuum hypothesis, an obstruction to one specific proposed threshold. These claims are compatible: a threshold can exist without equalling the proposed explicit bound.

2. Background and essential definitions

Models and categoricity

A structure consists of a set together with named relations and operations. A group, an ordered set and a field are examples. An isomorphism is a bijection preserving that structure.

For a class $K$, let $I(K,\kappa)$ count the isomorphism types of models with cardinality $\kappa$. Categoricity means

$$I(K,\kappa)=1.$$

Existence is included. Having no models of a size is not categoricity, even though there are then no two models to distinguish.

An abstract elementary class, or AEC, is a class of structures together with a designated strong-substructure relation $\leq_K$. It is closed under isomorphism and coherent increasing unions. The relation must agree with ordinary substructure in the forward direction and satisfy a coherence condition among nested models.

The Löwenheim–Skolem axiom provides small strong submodels. There is an infinite cardinal $\ell$ such that every subset $A$ of a model $M$ lies in a strong submodel $N\leq_KM$ with

$$|N|\leq|A|+\ell.$$

The least such bound, subject to the language-size convention, is $\operatorname{LS}(K)$. It measures how much structure must accompany a chosen set of elements. AECs extend the first-order setting without assuming its compactness theorem.

Additional structural properties

Joint embedding means that any two models strongly embed into a common model. Amalgamation asks for such embeddings over a common strong submodel, agreeing there. Tameness says, roughly, that differences between types can be detected over small pieces of their parameter sets.

These are powerful hypotheses in categoricity theorems, but they are not part of the bare AEC axioms used by the new manuscripts.

Thresholds and the beth hierarchy

The cardinals $\beth_\alpha$ start with $\beth_0=\aleph_0$, satisfy $\beth_{\alpha+1}=2^{\beth_\alpha}$, and take suprema at limit ordinals. They iterate the power-set operation through transfinite stages.

A proposed threshold is

$$h(\kappa)=\beth_{(2^\kappa)^+},\qquad H(K)=h(\operatorname{LS}(K)),$$

where the superscript $+$ means the next cardinal, used as an ordinal index for $\beth$. It is not an added exponent.

The continuum hypothesis, CH, says $2^{\aleph_0}=\aleph_1$. Under CH, a countable Löwenheim–Skolem number gives $H(K)=\beth_{\omega_2}$, where $\omega_2$ is the initial ordinal of cardinality $\aleph_2$.

3. Development history and previous partial results

Morley’s 1965 theorem is the first-order prototype: a complete theory in a countable language that is categorical in one uncountable cardinal is categorical in every uncountable cardinal. Baldwin and Lachlan’s work on strongly minimal sets explains a central geometric mechanism, in which models are classified by bases and dimensions.

Shelah introduced AECs to organize nonelementary classes, including models of infinitary logics. In that setting, types, independence and amalgamation cannot simply be imported from first-order compactness. Eventual categoricity became a broad test of how much classification theory survives.

Major partial results imposed structural or set-theoretic hypotheses. Makkai and Shelah obtained transfers for infinitary theories using a strongly compact cardinal. Grossberg and VanDieren proved upward transfer from sufficiently large successor categoricity under tameness, amalgamation, joint embedding and arbitrarily large models. Boney derived tameness from large-cardinal assumptions and obtained further eventual-categoricity consequences.

Vasey proved ZFC transfer for universal classes without assuming amalgamation, and other results treated tame classes with primes. Shelah and Vasey developed categoricity through multidimensional diagrams for excellent classes, with broader conclusions under additional hypotheses. These results show why both the structure of the class and the kind of starting cardinal matter.

The literature also distinguishes an unspecified eventual threshold from a prescribed one. An existence bound can ensure that a class has arbitrarily large models without ensuring uniqueness there or guaranteeing that every particular model extends arbitrarily far.

Both manuscripts discuss earlier public claims by Espíndola of general eventual categoricity and explicit-threshold transfer. M2 describes its construction as independent of those preprints. M1 addresses the prescribed-threshold statement and a specific step in the cited argument. These are attributed comparisons in the sources, rather than an independent determination of priority or proof correctness.

4. What the new manuscripts claim

M2: one uniform threshold exists

For every infinite cardinal $\lambda$, M2 asserts that there is a cardinal $\mu(\lambda)\geq\lambda$ such that every AEC $K$ with $\operatorname{LS}(K)\leq\lambda$ satisfies

$$\begin{aligned} &K\text{ categorical in some }\kappa\geq\mu(\lambda)\\ &\qquad\Longrightarrow K\text{ categorical in every }\kappa'\geq\mu(\lambda). \end{aligned}$$

The same threshold appears in the premise and conclusion, and it depends only on the Löwenheim–Skolem bound, not on the particular class.

The theorem is stated in ZFC, without amalgamation, joint embedding, tameness, absence of maximal models or arbitrarily large models as extra assumptions. The initial categoricity cardinal may be a successor, a regular limit or a singular cardinal.

M2 also states an equivalent qualitative form: an AEC categorical in unboundedly many cardinals is categorical on a tail, meaning every cardinal above some bound. It does not give a numerical formula for the uniform threshold.

M1: the proposed endpoint can fail under CH

Assuming CH, M1 constructs an AEC in a countable finitary relational language with $\operatorname{LS}(K)=\aleph_0$, at least two nonisomorphic models of size

$$H(K)=\beth_{\omega_2},$$

and categoricity in every cardinal at least

$$\Lambda=\beth_{(2^{\aleph_1})^+}.$$

This $\Lambda$ is larger than $H(K)$. Thus categoricity high above the proposed endpoint does not transfer down to that endpoint. The same class is nevertheless categorical on a higher tail.

Using the relative consistency of CH, the manuscript concludes that, if ZFC is consistent, the specified universal threshold assertion is not provable in ZFC, for the precise formulations discussed there. This is one-sided nonprovability. It is not a claim of independence in both directions, and it does not refute qualitative eventual categoricity.

5. Why the advance matters

Importance: 5/5, provisional. M2 claims a general solution of one of the central classification problems for AECs, removing the structural and starting-cardinal restrictions present in major earlier theorems.

The uniform quantifier is particularly strong. It says that a single threshold works for every class with a bounded Löwenheim–Skolem number, even when their internal structures differ greatly.

M1 contributes a complementary warning about quantitative formulations. A threshold motivated by existence theory need not be the correct threshold for uniqueness. Its example would force any general explicit-bound proposal to account for models that coexist at one large size but cannot continue into the final categorical regime.

The importance score concerns those claims together. It does not turn the unspecified bound in M2 into an effective estimate or treat M1’s CH-dependent construction as an unconditional ZFC counterexample.

6. The main idea

M2 constructs geometry from eventual uniqueness

The positive argument first reduces to a fixed AEC categorical in unboundedly many cardinals. It builds models indexed by linear orders, with each element described using a label and finitely many order positions. Order embeddings then induce strong maps.

Specially chosen orders control the possible records of short tuples. The manuscript constructs reversals of strong maps on prescribed bounded tuples directly, rather than assuming an unrestricted syntactic characterization that the cited literature had identified as problematic.

At selected sufficiently closed cardinal levels, these comparisons stabilize. They support amalgamation, homogeneity and an independence calculus for the working models. Here these properties are outputs of the construction at controlled levels, not assumptions on the original class.

The paper then extends the calculus to finite diagrams of models, retaining all specified boundary maps. Completing a boundary in a larger common model and extending an isomorphism onto a prescribed target are treated as separate requirements.

A minimal type supplies a pregeometry, an abstract dependence structure with a basis notion analogous to linear independence. The crucial claim is that a proper strong extension must add a new point of this geometry. Above a threshold, basis cardinality therefore measures the cardinality of the model itself.

The final induction classifies models by basis maps, using coherent unions at limit stages and separate cardinal bookkeeping for successor, regular-limit and singular sizes. This gives a categorical tail for the fixed class.

Uniformity comes from a set of codes

A separate coding argument passes from individual tails to one threshold for all classes with bounded Löwenheim–Skolem number. Up to the relevant coding, there are set-many such AECs.

For each code, either the categoricity spectrum is bounded, or the qualitative theorem gives a categorical tail. Taking a common bound above all bounded spectra and all tail thresholds yields $\mu(\lambda)$. A class categorical above that common bound must belong to the second case.

This is an existence argument for a bound. It does not compute a convenient beth expression for it.

M1 keeps two kinds of models alive at the endpoint

The obstruction uses an index sort $I$, controlling total size, and a base sort $B$, a dense linear order without endpoints whose bounded intervals are countable. This forces $|B|\leq\aleph_1$.

Attached row and column fibers are torsors: they behave like groups with no distinguished origin. Choosing origins displays binary matrices. Changing row origins alters finitely many entries in each row, while changing column origins alters finitely many in each column.

When $B$ is countable, these changes absorb every matrix, leaving one isomorphism type at a fixed large index size. When $B$ is uncountable, residual binary-sequence colors remain modulo finite changes. They encode countable extensional relations with ordinal rank labels below $\omega_1$.

Under CH, local rank approximations permit an uncountable-base model at $\beth_{\omega_2}$, alongside a countable-base model of the same size. The base-sort sizes distinguish them.

At sizes at least $\Lambda$, the color tests would force a coherent global pattern of bounded-rank extensional relations too large to exist. This eliminates uncountable-base models. Only the countable-base regime remains, giving categoricity on the higher tail.

Strong inclusions control newly added columns on old tuples. That detail permits small strong submodels and directed unions even though each tuple can have its own countable exceptional set of columns.

7. Scope, limitations, and open questions

The positive theorem concerns bare AECs with set-sized finitary languages and the stated coherent-union axioms. Categoricity includes existence at every size in the resulting tail.

The negative construction assumes CH and fails both joint embedding and amalgamation. It therefore does not contradict transfer theorems that assume those properties. Nor does the existence of arbitrarily large models somewhere in the class imply that its uncountable-base endpoint model has arbitrarily large extensions.

M2 does not specify the least possible uniform threshold. The classical prescribed $\beth_{\omega_1}$ question for countable infinitary sentences is also kept separate in the manuscript. M1’s obstruction targets the displayed AEC bound $H(K)$, not every proposed threshold in every restricted setting.

The two manuscripts should therefore be read as a qualitative existence claim paired with a limitation on one quantitative proposal. Their constructions are independent: M1’s higher categorical tail is established within its own argument, without invoking M2.

8. Sources and further reading

Morley, Categoricity in power, Transactions of the American Mathematical Society 114 (1965), 514–538, is the first-order prototype. Baldwin and Lachlan, On strongly minimal sets, Journal of Symbolic Logic 36 (1971), 79–96, develops the geometric classification background.

Shelah, Classification Theory for Abstract Elementary Classes (2009), formulates the broader program. Boney and Vasey, A survey on tame abstract elementary classes, provides background on tameness and transfer. Their Categoricity and infinitary logics explains the version-sensitive order-presentation issue cited by M2.

Shelah and Vasey, Categoricity and multidimensional diagrams, Journal of the European Mathematical Society 26 (2024), 2301–2372, supplies a major predecessor for the finite-system methods. Espíndola’s eventual-categoricity preprint and explicit-threshold preprint are the earlier public claims discussed in the source manuscripts.

Source manuscripts

• M1. OpenAI, A CH obstruction to a prescribed categoricity threshold (September 24, 2026).
• M2. OpenAI, Eventual categoricity for abstract elementary classes (September 24, 2026).

Sources consulted on 21 October 2026 at repository revision 1581d4c. Download the manuscript citations.