Definition
Plain language
A known optimization algorithm for saddle-point problems whose convergence rate an AI agent tightened with a new parameter schedule.
As stated in the literature
A first-order method for convex-concave saddle-point optimization; in the AI Co-Mathematician work, a Lean-proof-search agent discovered a new parameter schedule that provably improves its convergence rate.
Why it matters: It's a concrete instance where an AI-assisted prover improved on a published convergence rate, showing AI can make real contributions to optimization theory.
For example, when applied to a minimax problem, anchored gradient descent-ascent uses an extra reference point that pulls iterates back toward stable ground.
Heard on the show
“The context: there's a well-known algorithm called anchored gradient descent-ascent, used in saddle-point optimization.”Episode 067 — An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won