OpenAI proves the spin-one Haldane gap (periodic chains)

Your AI and this story

Key facts

• Theorem 1.1 (periodic paper, verbatim in substance): ‘the ground state is unique for every even L ≥ 60, and γ_L > (4/105) log(80/79) for every even L ≥ 60, γ_L > log 20/784 for every even L ≥ 2304. Consequently Δ₁ ≥ log 20/784 > 0.’ The paper says this ‘resolves the spin-one Haldane gap conjecture positively in the even-periodic finite-volume formulation Δ₁ > 0’
• Companion (boundary field): uniform gap for odd open chains of 2L+1 sites with the same field h = 3/5 at both ends, ‘a gap greater than log(10)/392 on every chain … with L ≥ 960’; ‘Tasaki’s index theorem then gives index −1’ for boundary-selected limit states (openai/math CONTENTS, family 268)
• Background: Haldane’s prediction (1981, published 1983) that integer-spin Heisenberg chains are gapped while half-integer ones are not underlies the 2016 Nobel Prize in physics. Before this, a gap was proved only for special models such as AKLT (Affleck–Kennedy–Lieb–Tasaki 1987)
• Expert reading: Hal Tasaki (Gakushuin University), on his website: ‘I carefully studied the analytic argument in the first paper, which proves the Haldane conjecture for the S = 1 antiferromagnetic Heisenberg chain in the standard periodic-chain formulation. I found the argument remarkably simple and elegant … The complete proof also uses computer-assisted estimates, which I have not independently checked.’ Two 34-page expository notes (continuous- and discrete-time versions, 18–19 Oct)
• Tasaki on X (11 Oct, Japanese, 21.6k views; our translation): he feels ‘in a nightmare, sweet enough to melt yet appalling’, reading OpenAI’s Haldane-gap proof with Berlioz’s Symphonie fantastique playing and rewriting it in his own words, and calls ‘the paper’s writing the worst’. In the v2 revision note of arXiv 2609.34119 (10 Oct): ‘I have so far found no logical gap in the portions I have examined, but my verification is not yet complete.’
• Other reactions: Masaki Oshikawa (Univ. of Tokyo, 6 Oct, 57k views) asked Tasaki whether this proves the Haldane gap and that the S=1 Haldane phase is SPT, calling it ‘a statement established among physicists but hard to prove mathematically’; Isaac Kim listed the ‘spin-one Haldane gap’ among five results in a ‘shocking list of problems’ in his field; a reader (@dualm, 10 Oct) found ‘an interesting new idea (or trick) … followed by a huge, almost brute-force calculation’
• Follow-up (Tasaki, arXiv 2610.34330, 25 Oct, v2 27 Oct): ‘Using the analytic machinery and numerical inputs of OpenAI’s proofs of the Haldane gap, we prove uniqueness of the translation-invariant ground state of the infinite S = 1 Heisenberg antiferromagnetic chain and of the J–D model for sufficiently small |D|.’ AI statement: ‘The author formulated the problems and proposed the core strategies. ChatGPT Astra made substantial contributions to developing the proofs and preparing the manuscript’; ChatGPT ‘found the separated-insertion argument’
• Verification: unrefereed OpenAI preprints with a Python/NumPy finite verifier (‘These finite checks do not replace the analytic uniform-gap argument’); no Lean formalization (community AGENT_NOTES: ‘None found’). An unreviewed GPT-6 Astra Max note (amachino, GitHub) claims a periodic gap > 0.1515 for every length ≥ 24