Paperlane

no. 14 · Friday 13 November 2026

Dual-Inertial Bregman Methods with Energy-Based Safeguarding

Kangming Chen

math.OC · v3

Abstract

We study dual-inertial (DI) Bregman proximal methods for convex optimization via an exact representation of the signed energy drift between dual-coordinate and primal affine interpolation. For unrestarted DI, bounded one-sided directional curvature gives a last-iterate objective-gap rate $\mathcal{O}((1+\log N)/N^2)$ after $N$ iterations, which improves to $\mathcal{O}(N^{-2})$ under weighted summability. The energy balance also provides a computable acceptance test, which we use in two safeguards. The restart safeguard attains a global $\mathcal{O}(T^{-1})$ rate after $T$ proximal calls, including rejected trials. Under local quadratic growth and kernel regularity, it restarts only finitely often and eventually converges linearly. The non-restarting safeguard preserves the quadratic weight recurrence via dual substeps. Every fallback terminates without triangle scaling, yielding a $\mathcal{O}(N^{-2})$ rate in outer iterations and an $\mathcal{O}(T^{-1})$ rate in calls, which improves to $\mathcal{O}(T^{-2})$ when the cumulative fallback cost is $\mathcal{O}(N)$. Experiments on relative-entropy regression and nonnegative least squares illustrate the results.

fields
Optimization and Control
pdf ↗ page arXiv:2610.20420 ↗