Every triple becomes perfect?
An AI preprint claims Donaldson’s hypersymplectic deformation conjecture
Covers result 341 in OpenAI’s catalog: its one paper. See its connections to other results and the work it cites.
4:00. English captions available. The video loads from YouTube when you press play.
Chapters
The claim, in plain words
A four-dimensional space can carry three closed two-forms at once. If at every point their pairwise wedge products form a positive definite 3 × 3 matrix, the triple is called hypersymplectic. Our computer finds that a random triple at a point passes this test only about 1 time in 90. The ideal case is hyperkähler: the matrix is the identity times a volume form everywhere, and the three forms come from a hyperkähler metric. Can every hypersymplectic triple be deformed into a hyperkähler one?
In 2006 Donaldson asked whether a compact four-manifold carrying such a triple must admit a hyperkähler structure. In 2017 Fine and Yao stated a stronger version as a conjecture of Donaldson: normalize the triple, then deform it through hypersymplectic triples, keeping all three cohomology classes fixed, to a hyperkähler triple. Their hypersymplectic flow became the main tool, and convergence was known only in very restricted cases, such as symmetric triples on the four-torus. Simply interpolating fails: on the four-torus our computer finds a straight line between two cohomologous symplectic forms that stops being symplectic in the middle.
Result 341 is one OpenAI paper, 72 pages, first dated September 23; a revision dated October 7 changed only the introduction. It claims the conjecture in full: every normalized hypersymplectic triple on a closed, connected, oriented four-manifold deforms, with all three cohomology classes fixed, to a hyperkähler triple. A corollary: any such manifold is diffeomorphic to a K3 surface or the four-torus. Before, this was known only in special cases: up to homeomorphism for simply connected manifolds, or for triples with a circle symmetry.
What it doesn’t claim
Not claimed: that the hypersymplectic flow converges, since the proof does not use it; a classification of all symplectic Calabi–Yau surfaces; or a machine-checked proof, since this paper has no Lean.
Where it stands
- Posted as a preprint
- OpenAI, 72 pages, dated September 23, 2026; revised October 7 (introduction only)
- Peer reviewed
- Not yet
- Formalized in Lean
- Not for this paper (its first step is result 342’s theorem, which has Lean)
Last checked October 19, 2026. The paper is unrefereed. Its first step, tamed implies compatible, is the companion paper’s theorem, which has a Lean proof; the rest has none. The argument leans on deep inputs, including Taubes’s work relating Seiberg–Witten invariants to curves, and Li and Liu’s wall crossing. We found no specialist comment, no reported error, and no review by OSVerify. The conjecture’s name is the paper’s own; Donaldson’s original question asked only about existence. OSVerify has not reviewed it.
How the proof goes
- Taming, then curves. Two of the forms define an almost complex structure, and the third must tame it, as our computer confirms in 4,000 random cases. Tamers of a fixed structure form a convex set, unlike symplectic forms in a class. The paper first proves that tamed implies compatible, the companion result 342, then gives a curve criterion for when a cohomology class still contains a tamer. Seiberg–Witten theory over the sphere of combinations of the forms produces pseudo-holomorphic curves through every point.
- A torus fibration, collapsed. With a carefully chosen lattice class, these curves become the fibers of a torus fibration with nodal fibers. Shrinking the fibers with a semi-flat model, like the collapse our computer checks, where fiber area shrinks while the base grows, the paper builds an auxiliary form, and two applications of Yau’s theorem finish the deformation.
The numbers in the video
- Ours: about 1 in 90 random triples of two-forms at a point is positive (2,225 of 200,000).
- Ours: a third form completes a positive triple exactly when it tames one of the two signs of the almost complex structure, in 4,000 of 4,000 random cases.
- Ours: on the four-torus, a straight line between two cohomologous symplectic forms stops being symplectic in the middle (minimum of ω∧ω/2 down to −0.95).
- Ours: in the semi-flat model, fibre area ε and base area 0.0908/ε: the fibres collapse while the volume stays fixed.
Read the source
- Deforming hypersymplectic four-manifolds to hyperkähler triples (PDF), 72 pages.
- The collection’s catalog, where this is result 341.
The paper
OpenAI groups related papers into one result; this one has a single paper.
Transcript
Read the full script
0:00The question
A four-dimensional space can carry three closed two-forms at once. If at every point their pairwise wedge products form a positive definite 3 × 3 matrix, the triple is called hypersymplectic. Our computer finds that a random triple at a point passes this test only about 1 time in 90. The ideal case is hyperkähler: the matrix is the identity times a volume form everywhere, and the three forms come from a hyperkähler metric. Can every hypersymplectic triple be deformed into a hyperkähler one? In 2006 Donaldson asked whether a compact four-manifold carrying such a triple must admit a hyperkähler structure. In 2017 Fine and Yao stated a stronger version as a conjecture of Donaldson: normalize the triple, then deform it through hypersymplectic triples, keeping all three cohomology classes fixed, to a hyperkähler triple. Their hypersymplectic flow became the main tool, and convergence was known only in very restricted cases, such as symmetric triples on the four-torus. Simply interpolating fails: on the four-torus our computer finds a straight line between two cohomologous symplectic forms that stops being symplectic in the middle.
1:17The claim
Result 341 is one OpenAI paper, 72 pages, first dated September 23; a revision dated October 7 changed only the introduction. It claims the conjecture in full: every normalized hypersymplectic triple on a closed, connected, oriented four-manifold deforms, with all three cohomology classes fixed, to a hyperkähler triple. A corollary: any such manifold is diffeomorphic to a K3 surface or the four-torus. Before, this was known only in special cases: up to homeomorphism for simply connected manifolds, or for triples with a circle symmetry. Not claimed: that the hypersymplectic flow converges, since the proof does not use it; a classification of all symplectic Calabi–Yau surfaces; or a machine-checked proof, since this paper has no Lean.
2:12How the proof goes
Two of the forms define an almost complex structure, and the third must tame it, as our computer confirms in 4,000 random cases. Tamers of a fixed structure form a convex set, unlike symplectic forms in a class. The paper first proves that tamed implies compatible, the companion result 342, then gives a curve criterion for when a cohomology class still contains a tamer. Seiberg–Witten theory over the sphere of combinations of the forms produces pseudo-holomorphic curves through every point. With a carefully chosen lattice class, these curves become the fibers of a torus fibration with nodal fibers. Shrinking the fibers with a semi-flat model, like the collapse our computer checks, where fiber area shrinks while the base grows, the paper builds an auxiliary form, and two applications of Yau’s theorem finish the deformation.
3:08Where it stands
The paper is unrefereed. Its first step, tamed implies compatible, is the companion paper’s theorem, which has a Lean proof; the rest has none. The argument leans on deep inputs, including Taubes’s work relating Seiberg–Witten invariants to curves, and Li and Liu’s wall crossing. We found no specialist comment, no reported error, and no review by OSVerify. The conjecture’s name is the paper’s own; Donaldson’s original question asked only about existence. If it holds, every closed, connected four-manifold with a hypersymplectic triple is really a K3 surface or a four-torus, and the triple, once normalized, can be deformed to a hyperkähler one without changing its cohomology classes.
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.