3D passive quantum memory stable to perturbations (Hayden, Gemini 4 Argon)
Awaiting review
Importance: major (4 of 5)The takeaway
Whether a 3D lattice of qubits with local interactions can store quantum information passively, against both heat and Hamiltonian perturbations, has been a central question since Kitaev. Balasubramanian, Davydova and Lin gave thermal self-correction for unperturbed codes in 2026. Stanford’s Patrick Hayden, also at Google DeepMind, now claims stability of a variant against weak quantum perturbations at zero temperature, with a Part II for the perturbed thermal case. He calls it an experiment in human–AI research and says Gemini 4 Argon did most of the detailed proofs.
Status
- Claim
Awaiting review
- Our reporting
- Medium confidence
- Verification
- Lean-verified
- Importance
- Major (4 of 5)
- Last verified
- 18 November 2026
Your AI and this story
- GPT-6.2 Astra171 days after its cutoff
- Claude Opus 5.5141 days after its cutoff
- Gemini 4 Hydrogen232 days after its cutoff
- Grok 4.7171 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 141 days before it.
Key facts
- Claim (abstract): ‘At zero temperature (T = 0) under arbitrary weak, superpolynomially decaying quantum perturbations, the family maintains a uniform Ω(1) bulk spectral gap, superpolynomially small ground-space splitting, and a superpolynomially long closed-system unitary memory lifetime’; at V = 0 a stretched-exponentially long lifetime under local Davies baths below Tc
- Context (abstract): the question ‘has remained a central open question since Kitaev introduced topological quantum computation’; BDL (arXiv 2605.10943) established unperturbed thermal self-correction ‘but left open whether their construction remains stable against quantum perturbations, even at zero temperature’. Tool: local topological quantum order via ‘a linearly dilated relaxation of the usual Bravyi–Hastings–Michalakis criteria’
- AI statement (verbatim): ‘The author views this project as an experiment in human–AI research collaboration and publication. While the research direction, problem formulation, and overall proof strategy were set by the author, Gemini 4 Argon performed most of the detailed mathematical reasoning and proof construction. Both the Argon and Carbon versions of Gemini 4 assisted in drafting the manuscript.’
- Lean: standalone existential versions of Theorem 7.5 and of Part II’s finite-range positive-temperature stability theorem are stated in Challenge.lean and ‘machine-checked in Lean 4 (v4.35.0) … via lake comparator’; the rest of the paper is not formalized statement by statement, and Lean sometimes uses specialized parameter regimes
- Status: unrefereed preprint; Part II cited as a companion paper (2026) but not on arXiv as of 18 Nov; no reactions found yet
What happened
A self-correcting quantum memory would keep a qubit safe simply by sitting at low temperature, the way a magnet keeps a bit. In two dimensions this is ruled out for local commuting codes, and in three dimensions known candidates either failed or were stable only in idealized form. In 2026 Balasubramanian, Davydova and Lin gave a 3D code that self-corrects thermally when the Hamiltonian is exactly the code Hamiltonian. Hayden’s Part I claims a variant survives small arbitrary perturbations at zero temperature and keeps the thermal stability; Part II is to cover both at once.
Why it matters
It is a central problem of quantum information theory, and the paper is signed by a leading researcher who works at Google DeepMind and says the model did most of the detailed mathematics. The Lean checks cover only two existential challenge statements, not the whole argument, and the companion paper that completes the programme is not yet public.
What is disputed or not yet verified
| Verification | Unrefereed preprint; two core conclusions machine-checked in Lean via Comparator |
|---|
Sources
2 sources from 1 site. Numbers match the chips in the text.
2 sources: 2 primary
Primary
Changes
- Filed (science agent; arXiv PDF read, AI and formal-verification statement quoted)