Adaptive Accelerated Mirror Descent in Primal and Dual Spaces

Zeyi Xu, Long Chen · 2026

Read Adaptive Accelerated Mirror Descent in Primal and Dual Spaces by Zeyi Xu, Long Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02787.txt.

Description

Xu and Chen propose a new first-order optimization method. It blends three tools: preconditioning, acceleration, and adaptivity, all in mirror geometry. A flow-based view structures the method. The key device is a budget for Lyapunov perturbation. Local steps that fail to descend are allowed while the total budget stays nonpositive. Line search runs only when stability is at risk.

Convergence at the accelerated rate is proved. The proof assumes dual relative smoothness and convexity, plus a compatibility condition on the mirror geometry. For convex objectives, the rate is one over k squared. Homotopy gives that rate under a bounded-sublevel-set assumption. Experiments on relative-smoothness problems show the value of the full structure. Combining all three tools beats methods that use only part of it. The work refines optimization for non-Euclidean problems.

Catalog record

CREATORZeyi Xu, Long Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.02787

Browse the full library · About the Tower of Babel