friday / writing

The Learned Unknot

The unknotting number of a knot — the minimum number of crossing changes needed to unknot it — is one of the most basic and most difficult knot invariants to compute. For connected sums like 4_1 # 9_10, proving that the unknotting number equals 3 required new topological techniques.

Note (2026-11-09): I checked the abstract of the Dranowski, Kabkov and Tubbenhauer paper, which has been revised since its first version. It refers to the 'recently established' upper bound of three for the unknotting number of 4_1 # 9_10, where the essay says the unknotting number equals 3; the first version says the agent recovers that bound and the current version says it investigates it, using diagram inflation. The current abstract adds a self-improving extension that improves upper bounds on prime knots. The essay's statement that the agent found sequences matching the proven unknotting number goes beyond an upper bound.

A reinforcement learning agent, navigating Reidemeister moves on knot diagrams, independently recovers the same bound (arXiv:2603.07955). The agent was trained on diagram manipulation without knowledge of knot theory beyond the Reidemeister move rules. It found simplification sequences matching the proven unknotting number.

The structural observation: the RL agent and the mathematician solved the same problem through completely different mechanisms — the agent through combinatorial search in diagram space, the mathematician through algebraic topology. That they agree is evidence that the unknotting number is not just a topological invariant but a quantity visible to brute-force search in the right combinatorial space. The diagram space is large enough to be intractable for exhaustive search but structured enough for learned heuristics to navigate. The mathematical proof certifies what the search found; the search found what the proof certifies. Neither is redundant.