Mirror Polyak and a Primal-Dual Lifting

arXiv:2608.17252 2026 Optimization 1 ideas extracted · analyzed Sep 1, 2026

What the math gives to ML

The paper develops a Polyak-style step rule for mirror descent, replacing Euclidean gradient geometry with a problem-specific Bregman geometry. The transferable asset is objective-gap adaptation combined with mirror maps that encode positivity, simplex constraints, entropy geometry, or other non-Euclidean structure. A practical neural-network experiment is to use this rule for constrained parameters such as mixture weights, MoE router probabilities, or positive gates, solving a one-dimensional scalar equation for each update. Direct use in fully nonconvex stochastic training should begin with conservative lower-bound estimates and step clipping.

Ideas from this paper

Failed on benchmark 2026

Bregman-Projection Polyak Optimizer

Replace the Euclidean Polyak step in an optimizer with a mirror-descent step whose length is chosen by projecting onto the current affine lower-bound halfspace in Bregman geometry. This permits entropy geometry for simplex-valued router probabilities, log geometry for positive parameters, and other mirror maps without reducing the method to a norm-based learning-rate rule.

Useful7/10
Difficulty5/10
Novelty7/10
Paper: Mirror Polyak and a Primal-Dual Lifting arXiv:2608.17252