VibeMathedMath problems solved by AI

Polynomial-Time MIMO Detection at the ML Threshold

In the square Gaussian binary MIMO model y=ρ/NHx+wy = \sqrt{\rho/N}\,Hx^\star + w, exhaustive maximum-likelihood detection recovers xx^\star once ρ>2logN\rho > 2\log N, while sphere decoding at that threshold scale costs exp{Θ(N/logN)}\exp\{\Theta(N/\log N)\}. Whether any polynomial-time detector reaches the same first-order threshold, or whether a computational-statistical gap separates them, was open. The claim: rounded linear MMSE followed by steepest single-bit descent recovers xx^\star with failure probability tending to zero, uniformly over every transmitted word, in O(N3)O(N^3) operations.

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI-discovered
Method
Argument
Field
Information theory; average-case complexity
Posed by
Year posed
Years open
Solved
2026-08-08
Model
GPT-5.6, Claude Fable 5
Vendor
OpenAI, Anthropic
Collaborators
Dimitris Papailiopoulos
Verification
Unreviewed
Publication
Announced
Significance
18 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

An average-case claim about the Gaussian model, not a contradiction of the worst-case NP-hardness of integer least squares. If it holds, no computational-statistical gap separates polynomial-time detection from exhaustive maximum likelihood at first order in this model.

What the AI did

The manuscript's footnote in full: "The results in this paper were proved by GPT-5.6 and Claude Fable 5, which also drafted the initial manuscript. The author posed the problem, directed several rounds of proof simplification, verified all mathematical arguments, edited the manuscript, and takes full responsibility for its content." By the author's public account, Claude proposed the algorithm (signed LMMSE plus greedy bit flips) and GPT repaired and simplified the proof over several days of directed iteration.

Verification

A 46-page manuscript posted to the author's own site and announced on X, days old, with no independent review and no formalization. The author states he checked every argument line by line over about five days, and he has published previously on polynomial-complexity ML detection, so the domain expertise is real; neither fact is independent scrutiny, which is why this is a candidate.

Sources

Author manuscript

Submitted by VibeGene on

Changelog2 changes
  • Rasmus Lindahlapproved this entry
  • GoldenQuokka142submitted this entry

Discussion