New CES framework proves global convergence in evolutionary optimization

A team of researchers (Neto, Garay, Martí, Sanchez-Pi) has published a rigorous mathematical framework proving global convergence for the Canonical Evolutionary Strategy (CES) in stochastic continuous optimization. The work, posted on arXiv, models CES as a controlled system governed by a Schrödinger-type replicator-mutator equation. They derive a hierarchy from discrete individual-based dynamics to a deterministic mean-field limit, showing that convergence is determined by the principal eigenfunction of the underlying operator—a property they call Geometric Selection.

This Geometric Selection naturally favors broad, flat optima over narrow local traps, providing a mathematical justification for the 'survival of the flattest' phenomenon. Crucially, CES's replicator-mutator dynamics enable intrinsic mass transport, avoiding the premature variance collapse that plagues consensus-driven methods when the global minimizer lies outside the initial support. In high-dimensional benchmarks (d=30), CES achieved lower residual errors in shifted initialization scenarios where standard consensus-driven and gradient-based methods failed to migrate effectively.