Shelah's Categoricity Conjecture for L_omega1,omega

Shelah's Categoricity Conjecture for L_omega1,omega

OpenAI's claim

OpenAI presents the following claims relevant to this question, with manuscript numbers following paper order in Sources. In manuscript 1 (family 240; 2026-09-24), the asserted result is that in ZFC, each bound on Löwenheim--Skolem number has a uniform threshold above which categoricity in one cardinal implies categoricity throughout the same tail.

Mathematical statement of the claimed result

Manuscript 1.

For every infinite cardinal $\lambda$, the claim gives a uniform cardinal threshold $\mu(\lambda)\ge\lambda$ for all abstract elementary classes $K$ of Löwenheim--Skolem number at most $\lambda$: \[ (\exists\kappa\ge\mu(\lambda):I(K,\kappa)=1) \Longrightarrow(\forall\kappa'\ge\mu(\lambda):I(K,\kappa')=1). \] Here $I(K,\kappa)$ is the number of model isomorphism classes of cardinality $\kappa$. The same threshold occurs in the premise and conclusion. A second formulation asserts that categoricity in unboundedly many cardinals forces categoricity on a tail, without adding amalgamation or tameness assumptions.

Outline of the argument

For manuscript 1 (family 240; 2026-09-24), fixed order-indexed presentations encode short model orbits. A direct reversal theorem for strong maps preserves bounded tuple records; elimination and stabilization of strong-map classes then create the stable diagrams and closed-level amalgamation needed for tail transfer.

Formal verification

No Lean formalization is listed at this snapshot for manuscript 1 (family 240; 2026-09-24).

This site has not executed the Lean code. Described statement coverage is not an independent proof check.

Status caveat

The manuscripts were produced by an internal OpenAI model. OpenAI cautions that results without formalization can contain mistakes. This site has not independently checked the proofs; the label records the scope of OpenAI's claim.

Sources

• OpenAI, Eventual categoricity for abstract elementary classes, 2026-09-24; citation key OAI:Eventual-Categoricity-for-Abstract-Elementary-Classes-September-24-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: main:uniform; Statement source: main:qualitative.
• Family 240: Lean scope.
• OpenAI catalogue entry.

Shelah's Categoricity Conjecture for L_omega1,omega

Definitions and mathematical statement

The logic $L_{\omega_1,\omega}$ permits countable conjunctions and disjunctions but only finite strings of quantifiers. A sentence $\psi$ is categorical in $\lambda$ if it has exactly one model of cardinality $\lambda$ up to isomorphism. Source notation correction: Conjecture 1.1 of Shelah--Vasey prints the threshold $\beth_{\omega_1}$, not the undefined "delta_omega1" of the JSON transcription. Define $\beth_0=\aleph_0$, $\beth_{\alpha+1}=2^{\beth_\alpha}$, and $\beth_\lambda=\sup_{\alpha<\lambda}\beth_\alpha$ at limit ordinals. The cited strong conjecture is \[ \left[\exists\lambda\ge\beth_{\omega_1}:\psi\text{ categorical in }\lambda\right] \Longrightarrow \left[\forall\mu\ge\beth_{\omega_1}:\psi\text{ categorical in }\mu\right]. \] The correction is based on the source's displayed statement. Allowing an unspecified larger eventual threshold would weaken this question.

Formulation and concept references: [S1].

Short English statement

If an infinitary sentence has a unique model at one cardinal above the specified Hanf threshold, must it have a unique model at every cardinal above that threshold?

Research attempt

The exact threshold and the source audit (27 September 2026).

The target is the strong version with the fixed threshold $\beth_{\omega_1}$, not eventual categoricity above an unspecified larger cardinal. The supplied Shelah--Vasey paper [S1], Conjecture 1.1 and the following discussion, were inspected. Its Corollary 14.7 proves an eventual statement under a strongly compact cardinal hypothesis and with a threshold depending on that cardinal. Those extra assumptions and the different threshold cannot be erased to solve the selected ZFC statement.

We use the usual countable vocabulary convention for the infinitary sentence. The countably many symbols and subformulas actually appearing in a sentence are the relevant syntactic data; arbitrary expansions by unrelated new symbols are not part of the categoricity assertion. Categoricity includes existence: zero models in a cardinal is not categoricity there.

The attempt develops three checks. First, a complete vector-space family illustrates a successful dimension argument. Second, explicit infinitary sentences show what countable categoricity and compactness cannot give. Third, a fragment-hull construction and a precise back-and-forth lemma isolate the missing extension property. None of the counterexamples below satisfies the large-cardinal premise of the conjecture; they test proposed shortcuts, not the conjecture itself.

A fully proved categorical family in every infinite cardinal.

Fix a finite or countably infinite field $F$. Use symbols for addition, zero, and multiplication by each scalar in $F$. The vector-space axioms form a countable set of first-order sentences, so their conjunction is an $L_{\omega_1,\omega}$ sentence. Conjoin, for each positive integer $n$, the assertion that there are $n$ linearly independent vectors. For a countable field, independence of a fixed finite tuple is expressed by the countable conjunction \[ \bigwedge_{(a_1,\ldots,a_n)\in F^n\setminus\{0\}} a_1x_1+\cdots+a_nx_n\ne0. \] Only $n$ existential quantifiers occur in that assertion; the full sentence uses a countable conjunction of these finite strings, not an illegal infinite quantifier string. Denote the resulting sentence by $\psi_F$.

Its models are exactly the infinite-dimensional $F$-vector spaces. Every vector space has a basis by the usual Zorn argument: order linearly independent subsets by inclusion, take a maximal one, and note that a vector outside its span could be adjoined. A bijection between bases extends uniquely by finite linear combinations to a vector-space isomorphism. Thus dimension determines the isomorphism type.

If the basis has infinite cardinality $\delta$ and $F$ is countable, the set of finite linear combinations has cardinality $\max(\delta,\aleph_0)=\delta$. The upper bound follows by a countable union over finite tuple lengths, and the lower bound follows from containing the basis. Consequently a model of infinite cardinality $\lambda$ has dimension $\lambda$, and the direct sum $F^{(\lambda)}$ supplies one. Therefore \[ I(\psi_F,\lambda)=1\qquad\text{for every infinite }\lambda. \tag{370.1} \] Here $I$ counts isomorphism types of models of the indicated size. This verifies the selected conclusion for an explicit class, including all cardinals above the requested threshold.

The argument has a genuine invariant, basis cardinality, together with an extension theorem for maps of bases. Merely calling an arbitrary model's cardinality its "dimension" would not reproduce either step.

Why finite dimension over a countable field needs care.

If $F$ is countably infinite, a nonzero finite-dimensional vector space is already countably infinite as a set. Omitting the infinite-dimensional conjunct would therefore give several nonisomorphic countable models, although the uncountable models are still classified by their cardinality. The sentence above includes the conjunct explicitly so that $370.1$ is valid also at $\aleph_0$.

This small example matters when using cardinal arithmetic in a general argument: the formula $|V|=\dim_FV$ was proved only for infinite basis cardinality. A countable language also does not force all countable models of a theory to be isomorphic.

A sentence categorical only in the countable cardinal.

Take countably many constant symbols $c_0,c_1,\ldots$, with no other nonlogical symbols, and form \[ \chi=\left(\bigwedge_{i<j}c_i\ne c_j\right) \wedge\forall x\left(\bigvee_{n<\omega}x=c_n\right). \tag{370.2} \] Every model has exactly the elements named by the distinct constants, so it is countably infinite. Between two such models the unique map sending $c_n$ to $c_n$ is a bijection and an isomorphism. Thus $\chi$ is categorical in $\aleph_0$ and has no uncountable models.

This does not challenge the conjecture, whose premise begins at $\beth_{\omega_1}$. It directly refutes an argument that transfers categoricity from an arbitrary infinite cardinal merely by adding unnamed elements. Formula $370.2$ prohibits those elements in one countable disjunction.

It also demonstrates failure of compactness in the required logic. Add a new constant $d$ and the sentences $d\ne c_n$ for all $n$. Every finite part of this theory has a model: retain $\chi$ and interpret $d$ as a constant not among the finitely many forbidden ones. The entire theory has no model. First-order compactness cannot be applied to this collection as if its infinitary disjunction were a finitary formula.

Even existence in all cardinals does not repair a countable argument.

The first-order theory of dense linear orders without endpoints is categorical in $\aleph_0$. A direct proof is the classical alternating extension construction: enumerate the two countable orders, and extend a finite order-preserving partial bijection alternately to include the next unused element of each. Density and the absence of endpoints supply an image in the required finite interval. The union is an isomorphism.

For every uncountable cardinal $\lambda$, that same theory has nonisomorphic models of size $\lambda$. Here is an explicit pair. Let $J=\lambda\times\mathbb Q$ with lexicographic order. It is dense and has no endpoints: within a block use density of $\mathbb Q$, and between different blocks a larger rational in the earlier block lies between the two specified points. Put \[ L_0=\mathbb Z\times J,\qquad L_1=\omega_1\times J \] with lexicographic orders. Both are dense without endpoints and both have cardinality $\lambda$.

The cofinality of $L_0$ is $\aleph_0$: choose one point in each positive integer block to obtain a countable cofinal sequence, and no finite set is cofinal in an order with no last point. The cofinality of $L_1$ is $\aleph_1$. Any countable set has its first coordinates bounded in $\omega_1$, so is not cofinal. Choosing one point in each block is cofinal and has size $\aleph_1$. Cofinality is preserved by order isomorphisms, so these models are not isomorphic.

The finite-partial-map proof for countable orders failed because at uncountable stages one must realize cuts determined by potentially infinite earlier domains. Density only promises points between finitely many bounds. This is a concrete model of the extension problem a transfinite categoricity proof must solve; increasing the length of the enumeration does not automatically upgrade the available extension property.

A fragment version of downward Löwenheim--Skolem can be proved directly.

Let $\mathcal F$ be a countable collection of $L_{\omega_1,\omega}$ formulas containing the sentence $\psi$ and closed under subformulas and the finitely many-variable operations needed for the following witness argument. Such a countable fragment can be chosen for one sentence by recursively adjoining its subformulas and closing under countably many syntactic operations. Let $N\models\psi$ and $A\subseteq N$.

Build sets $A_0\subseteq A_1\subseteq\cdots$ starting with $A$ and the interpretations of constants. At each step close under all function symbols, and for each existential formula $\exists x\,\varphi(x,\bar a)\in\mathcal F$ with parameters in the current set that is true in $N$, choose one witness in $N$ and adjoin it. There are at most $|A_i|+\aleph_0$ choices, since the fragment is countable and each parameter tuple is finite. Put $M=\bigcup_{i<\omega}A_i$ with the induced structure. Then \[ |M|\le |A|+\aleph_0. \tag{370.3} \]

Induction on formulas in the fragment shows that $M$ and $N$ agree on their truth values for parameters from $M$. Atomic formulas agree because it is a substructure. Negation and finite Boolean operations preserve agreement. For a countable conjunction, agreement holds for each conjunct by induction, so holds for their conjunction; the same applies to disjunction. For an existential formula true in $N$, its finitely many parameters appear together at some finite construction stage, and the next stage supplies a witness in $M$. The converse follows from agreement for the witnessing subformula. Thus $M\preccurlyeq_{\mathcal F}N$, and in particular $M\models\psi$.

For an infinite $\mu\le|N|$, starting with a set $A$ of size $\mu$ therefore gives a model of $\psi$ of size exactly $\mu$. This yields existence below a known model. It neither constructs larger models nor shows that the models of a given smaller size are isomorphic. The submodel relation is elementary for the specified fragment, not for every formula in the entire infinitary language.

Unions require a compatible elementary-chain hypothesis.

The same fragment induction proves a useful chain statement. If $M_i\preccurlyeq_{\mathcal F}M_j$ for $i<j$ in an increasing chain, then each $M_i$ is $\mathcal F$-elementary in the union. In the existential step, a witness in the union belongs to some later stage, and elementarity transfers the existential assertion back to the stage containing its parameters. A witness at that stage also works in the union. The Boolean and countable-conjunction steps are pointwise.

But arbitrary embeddings between models of $\psi$ need not be $\mathcal F$-elementary. Nor does a sequence of isomorphism types come equipped with chosen compatible embeddings. Building a directed union without specifying a relation that guarantees preservation of $\psi$ is therefore not a proof of existence, much less categoricity, at a limit cardinal. This is one reason the model-theoretic source works with controlled classes and embeddings rather than only raw cardinality counts.

A precise sufficient extension principle for uniqueness.

Lemma 370.1. Let $M,N$ be structures of the same infinite cardinality $\kappa$. Suppose there is a nonempty family $\mathcal I$ of partial isomorphisms from $M$ to $N$, each of domain and range size less than $\kappa$, with the following properties. The family is closed under restrictions. Any member can be extended in $\mathcal I$ to include any prescribed element of $M$ in its domain and any prescribed element of $N$ in its range. Unions of increasing chains from $\mathcal I$ whose union has size less than $\kappa$ also belong to $\mathcal I$. Then $M\cong N$.

Proof. Enumerate each universe in length $\kappa$. Recursively build an increasing sequence of partial maps, at successor stages including the next listed domain and range elements by the two extension properties. At a limit stage below $\kappa$, take the union. Start with the empty map, available by restriction closure. After each extension restrict back to the old domain and the at most two new elements needed for the requests. At stage $\alpha$ the domain has size at most $2|\alpha|$, hence is finite when $\alpha$ is finite and is less than $\kappa$ for every $\alpha<\kappa$. The union hypothesis therefore applies, even for singular $\kappa$.

At the final stage the union is a total bijection. Every atomic relation and function instance involves finitely many elements, hence is covered by one earlier partial isomorphism. The union is an isomorphism. $\square$

The restriction hypothesis is essential to the displayed size control. An extension between whole submodels may add arbitrarily many elements below $\kappa$; an increasing sequence of such sizes can become cofinal in a singular $\kappa$ too early. The lemma is a sufficient criterion for the precisely specified partial-map family, not a claim that every family of submodel embeddings satisfies its hypotheses.

Where applying this lemma to the conjecture stops.

The vector-space example supplies such maps by mapping small independent sets and extending maps of bases. The countable dense-order proof supplies finite maps, enough for $\kappa=\aleph_0$. A general sentence categorical at one huge cardinal does not visibly provide any such family for arbitrary models at each other huge cardinal.

An isomorphism between two large models is also not automatically an isomorphism fixing a specified common submodel pointwise. The latter requires homogeneity over that submodel, or an amalgamation/uniqueness argument controlling the embeddings. Existence of an isomorphism type and compatibility of diagrams are distinct assertions. The multidimensional diagrams in [S1] address precisely stronger structural information; their hypotheses cannot be replaced with an informal appeal to cardinality.

Even an eventual theorem beginning above a larger cardinal $\theta$ would leave the interval $[\beth_{\omega_1},\theta)$ untreated if $\theta>\beth_{\omega_1}$. The downward fragment construction gives models in that interval but no uniqueness theorem there. Thus it cannot repair the difference between the source's conditional eventual result and the sharp threshold required in this record.

Outcome and first missing step.

The arguments were checked against explicit vector-space bases, the countable-only sentence $370.2$, and the cofinality distinction between the two dense orders. These are exact logical examples. Finite enumeration of small structures cannot test categoricity at the beth threshold and is not presented as verification of such a claim.

Outcome: UNRESOLVED. The complete positive family has a genuine dimension-and-extension theorem, and the fragment construction supplies controlled downward existence and elementary unions. For an arbitrary sentence, the missing step is to derive sufficient structural extension and uniqueness properties from the one-cardinal categoricity premise, with the exact threshold and without additional large-cardinal axioms. None of compactness, countable back-and-forth, or a larger eventual threshold supplies that step as written.

Sources

• S1 Shelah and Vasey, Categoricity and multidimensional diagrams, JEMS, 2024. https://shelah.logic.at/files/241811/10.4171-jems-1477.pdf. Formulation source. Consulted 24 September 2026: Conjecture 1.1, printed page 2302: the threshold is beth_{omega_1}, not delta_{omega_1}..