Episode 297, October 2026

One irrep, not compact?

A second ZFC counterexample to Naimark’s problem, two days after Tanaka’s

Covers result 297 in OpenAI’s catalog: its one paper. See its connections to other results and the work it cites.

Play Episode 297: One irrep, not compact? Six blocks glued into one class by arcs.

4:12. English captions available. The video loads from YouTube when you press play.

Chapters

The claim, in plain words

The compact operators are the infinite matrices that are almost finite, and up to a change of basis they act irreducibly in only one way. In 1951 Naimark asked the converse: is a C*-algebra with only one irreducible representation, up to equivalence, always the compact operators? Rosenberg answered yes in 1953 when that representation acts on a separable Hilbert space, which covers every separable algebra, so any counterexample has to be nonseparable. In 2004 Akemann and Weaver built a counterexample assuming Jensen’s diamond principle, an extra axiom independent of ZFC, and whether ZFC alone gives one stayed open.

On September 22, 2026, a preprint by Ryotaro Tanaka gave a ZFC counterexample; its AI-use statement says the construction and proofs were developed predominantly by OpenAI’s ChatGPT, supervised and reviewed by the author. Result 297 is one OpenAI paper, 23 pages, dated September 24, 2026, and published October 6. It credits Tanaka and gives an alternative proof: in ZFC, a unital infinite-dimensional simple C*-algebra with a faithful tracial state, whose nonzero irreducible representations are all equivalent, so it is not the compact operators.

What it doesn’t claim

Tanaka’s theorem says more: a unique trace, a faithful representation on a separable Hilbert space, and a precise size. What the release adds is a different method; it claims no exact size for its algebra, only an upper bound.

Where it stands

Posted as a preprint: OpenAI, 23 pages, dated September 24, 2026
Peer reviewed: Not yet
Formalized in Lean: The main theorem (docs)

Last checked October 19, 2026. The paper has not been revised. In Lean, the main theorem is formalized and matches the paper, in about 26,000 lines, with no unproved steps in our scan (we did not build it). Lean’s logic is somewhat stronger than ZFC, a point raised on the Lean Zulip and relayed on MathOverflow; Nik Weaver, having read the manuscript, did not think it an issue. OSVerify has not reviewed it. On MathOverflow, Weaver wrote that it was “probably fair to say that the conventional wisdom was that a positive solution would also be consistent with ZFC.”

How the proof goes

Gluing, one pair at a time. The construction starts from the CAR algebra and adjoins unitaries one at a time by reduced HNN extensions. Each new unitary merges exactly one chosen pair of classes of pure states, and no other pair of old classes.
A schedule instead of a guess. Each stage can hold continuum many pure states, and gluing can create new ones, so the gluing runs for the successor of the continuum many steps, on a fixed schedule that never treats a state before it exists. Every pure state is determined on a separable piece, so the schedule reaches it and joins it to one reference class; that replaces the guessing of the diamond principle.

The numbers in the video

Ours: of 1,497 finite-dimensional C*-algebras up to isomorphism with dimension at most 50, exactly 7, the full matrix algebras, have only one irreducible representation up to equivalence.
Ours: gluing six 2×2 blocks with five unitaries takes six classes down to one, ending at the 12×12 matrices.

Read the source

A Counterexample to Naimark’s Problem in ZFC (PDF), 23 pages.
Tanaka, “A separably representable counterexample to Naimark’s problem in ZFC” (arXiv, September 22, 2026).
Nik Weaver’s MathOverflow answer (October 8, 2026).
The Lean formalization notes.
The collection’s catalog, where this is result 297.

The paper

OpenAI groups related papers into one result; this one has a single paper.

A counterexample to Naimark's problem in ZFC

Transcript

Read the full script

0:00The question

The compact operators are the infinite matrices that are almost finite, and up to a change of basis they act irreducibly in only one way. Our computer checks the finite picture: every unital representation of a full matrix algebra splits into copies of the standard one, and among 1,497 finite-dimensional C*-algebras of dimension at most 50, up to isomorphism, exactly 7, the full matrix algebras, have only one irreducible representation up to equivalence. In 1951 Naimark asked the converse: is a C*-algebra with only one irreducible representation, up to equivalence, always the compact operators? Rosenberg answered yes in 1953 when that representation acts on a separable Hilbert space, which covers every separable algebra, and Glimm’s work in 1961 gave another route. So any counterexample has to be nonseparable. In 2004, Akemann and Weaver built a counterexample assuming Jensen’s diamond principle, an extra axiom independent of ZFC, and showed that a counterexample generated by ℵ₁ elements is independent of ZFC. Whether ZFC alone gives a counterexample stayed open; a December 2025 draft of Farah’s book still said it is not known.

1:25The claim

On September 22, 2026, a preprint by Ryotaro Tanaka gave a ZFC counterexample; its AI-use statement says the construction and proofs were developed predominantly by OpenAI’s ChatGPT, supervised and reviewed by the author. Result 297 is one OpenAI paper, 23 pages, dated September 24 and published October 6. It credits Tanaka and gives an alternative proof: in ZFC, a unital infinite-dimensional simple C*-algebra with a faithful tracial state, whose nonzero irreducible representations are all equivalent, so it is not the compact operators. Tanaka’s theorem says more: a unique trace, a faithful representation on a separable Hilbert space, and a precise size. What the release adds is a different method. It claims no exact size for its algebra, only an upper bound.

2:24How the proof goes

The construction starts from the CAR algebra and adjoins unitaries one at a time by reduced HNN extensions. Each new unitary merges exactly one chosen pair of classes of pure states, and no other pair of old classes. Our computer shows the finite version: gluing six 2×2 blocks with five unitaries takes six classes down to one, ending at the 12×12 matrices, merging exactly the chosen pair each time. In the infinite case each stage can hold continuum many pure states, and gluing can create new ones, so the gluing runs for the successor of the continuum many steps, on a fixed schedule that never treats a state before it exists. Our computer checks such a schedule on a 60×60 grid. Every pure state is determined on a separable piece, so the schedule reaches it and joins it to one reference class; that replaces the guessing of the diamond principle.

3:19Where it stands

The paper is unrefereed and unchanged since release. In Lean, the main theorem is formalized and matches the paper, in about 26,000 lines; our scan finds no unproved steps, but we did not build it. Lean’s logic is somewhat stronger than ZFC, a point raised on the Lean Zulip and relayed on MathOverflow; Nik Weaver, having read the manuscript, did not think it an issue. On MathOverflow, Weaver wrote that the conventional wisdom, he thought, was that a positive answer would also be consistent with ZFC. If it holds, ZFC alone gives a C*-algebra with only one irreducible representation up to equivalence that is not the compact operators, now shown twice: by OpenAI, and by Tanaka with ChatGPT.

Spot a mistake?

This episode simplifies on purpose, but it shouldn’t be wrong. If something is, open an issue on GitHub or leave a comment on YouTube. Corrections are noted here with a date.