WASHINGTON — Microsoft Research announced on October 21, 2026, that a mathematical problem open since 1995 has been solved: the Kannan–Lovász–Simonovits (KLS) conjecture, a question about the geometry of high-dimensional probability distributions, now has a proof whose key bound no longer grows with dimension.
The result was produced by Microsoft researcher Xinzhi Zhang and collaborator Zhao Song, and it arrives with an unusual twist for a pure-mathematics breakthrough. According to Microsoft, the proof was found through a human-directed search in which artificial intelligence systems explored more than 100 candidate approaches while the researchers judged, pruned and redirected the work. Independent proofs released around the same time, plus a machine-checked formalization, give the claim extra weight — though, as with any major proof, the mathematical community will keep checking the details.
Illustration: A flat cut through a curved high-dimensional shape, the geometric idea at the heart of the KLS conjecture.
What Happened
Microsoft Research published its announcement on October 21, 2026, describing the resolution of the KLS conjecture, named after Ravi Kannan, László Lovász and Miklós Simonovits, who posed it in 1995. The paper behind the announcement, titled “An O(1) Bound for the KLS Constant” and credited to Zhao Song and Xinzhi Zhang, first appeared on the arXiv preprint server on October 1, 2026, with a revised version posted on October 4, according to coverage of the announcement.
In plain terms, the conjecture asks whether, for a broad class of high-dimensional probability distributions known as isotropic log-concave measures, the best flat cut through a distribution is always nearly as good as the best cut of any shape — with the gap between the two bounded by a universal constant that does not depend on the number of dimensions. Microsoft's announcement says the new proof establishes exactly that: a dimension-free, or O(1), bound for the KLS constant.
That matters because the geometry of these cuts governs how quickly random walks can explore a distribution. Sampling methods that underpin statistics, optimization and machine learning often rely on random walks, and the KLS conjecture has long been a central open question in the theory of high-dimensional sampling. Turning a widely used assumption into a theorem, if the proof holds up, removes a lingering “what if” from that theory.
The previous best result, published in July 2026, had brought the bound down to a slowly growing function of dimension. The new work eliminates that growth entirely, which is the defining claim of the conjecture.
Background
The KLS conjecture has been described by the Simons Institute as one of the most important open problems in convex geometry, according to Microsoft's announcement. For three decades, mathematicians chipped away at it, improving bounds step by step without closing the gap to a constant.
The July 2026 result set the stage: a 21-page paper establishing that the KLS constant grows at most like the fourth root of the logarithm of the dimension — slowly, but still growing. The October proof, reported to run to about 140 pages, claims to finish the job through a fully recursive refinement that keeps accumulated errors bounded by a universal constant.
What makes this story different from a standard breakthrough announcement is the reported method. Microsoft describes a human-directed agentic process: AI systems explored candidate approaches, and the researchers evaluated the results, discarded dead ends and steered the search. Microsoft also says that unguided attempts by several leading models, using roughly the same total computing budget, fell short — a comparison that comes from Microsoft's own account of the project and has not been independently benchmarked.
Illustration: Researchers reviewing geometric diagrams, with AI tools assisting the search for proof ideas.
Why It Matters
First, the mathematics itself. High-dimensional distributions appear everywhere from data analysis to statistical physics, and sampling from them efficiently is a foundational problem. A dimension-free bound means guarantees for sampling methods no longer weaken, even in principle, as dimensions grow. Researchers who relied on the conjecture as an assumption can now point to a theorem instead.
Second, the method. If human-guided AI search can help crack a 31-year-old problem, it suggests a new division of labour in mathematics: machines to widen the search, humans to judge what is promising and to assemble and verify the final argument. That is a more modest claim than “AI solved it alone,” and it is the claim Microsoft itself makes — the researchers directed the search and checked the result.
Third, the corroboration. Independent work matters enormously in mathematics, and this announcement did not arrive alone. Separate groups published their own proofs around the same time, including a 32-page proof by Pierre Bizeul, Boaz Klartag and Joseph Lehec, and another proof by Balasubramanian and Kasiviswanathan giving a dimension-free bound on the related Poincaré constant. A Lean 4 formalization, named KLS500, reportedly establishes a dimension-free Poincaré constant bound of 500. Multiple independent routes to the same conclusion make an error in any single argument less likely to sink the result.
What Happens Next
The researchers plan to release a shortened version of the proof and its formal verification after final review, according to coverage of the announcement. That shorter account will be what many mathematicians read first, and scrutiny of it — along with the full paper and the formalization — will determine how quickly the result enters textbooks and standard references.
There are also open questions about the method rather than the theorem. How much of the AI-assisted workflow generalizes to other problems? Which parts genuinely required human judgment? And how should the community credit and verify machine-assisted mathematics? Those debates will play out in seminars, journals and conference talks over the coming months, and they may matter as much in the long run as any single theorem.
For now, the headline stands on Microsoft's announcement and the accompanying independent work: a problem posed in 1995, when the web was young and high-dimensional data was a niche concern, appears to have been settled in an era when high-dimensional data is everywhere.
Illustration: A random walk exploring a high-dimensional distribution, the sampling process the KLS result concerns.
Sponsored: explore today's featured sponsored offers
FAQ
What is the KLS conjecture?
It is a 1995 conjecture in high-dimensional geometry and probability, named after Kannan, Lovász and Simonovits. It asks whether the best flat cut through a broad class of probability distributions is always nearly as good as the best cut of any shape, within a constant that does not depend on dimension.
Who proved it?
Microsoft Research announced on October 21, 2026, that Xinzhi Zhang and Zhao Song had proved it. Independent proofs by other mathematicians were also reported around the same time, along with a Lean 4 formalization.
Did AI solve the problem on its own?
No — not according to Microsoft's own account. The company describes a human-directed process in which AI explored more than 100 candidate approaches while the researchers judged, pruned and assembled the final proof.
Why does a dimension-free bound matter?
Because it means guarantees for sampling high-dimensional distributions do not weaken as dimensions grow. The geometry of these cuts governs how quickly random walks can explore a distribution, which underpins methods in statistics, optimization and algorithms.
Is the proof fully verified yet?
The full paper, independent proofs and a formalization have been reported, but the community's checking process continues. The researchers plan to release a shortened proof and formal verification after final review.
Sources
- Microsoft Research announcement, October 21, 2026, as reported by Unite.AI, October 2026.
- WebProNews, “Microsoft Researcher and Collaborator Resolve the 31-Year-Old KLS Conjecture With Human-Directed AI,” October 2026.
- Superpower Daily, “Microsoft details how human-guided AI helped solve a 31-year-old math problem,” October 21, 2026.
- Brief.News, “31-Year-Old KLS Conjecture Solved,” October 21, 2026.
Comments
Post a Comment