{"id":"arxiv/2609.34773","url":"https://arxiv.org/abs/2609.34773","tags":["math.OC"],"score":0,"title":"Accelerated Bregman Proximal Gradient Methods from Dual Geometric Perspectives","domain":"math.OC","source":"arxiv","authors":["Yuya Yamashita","Shota Takahashi","Akiko Takeda"],"reviews":0,"subject":"math.OC","summary":"We study Bregman proximal gradient (BPG) algorithms under relative smoothness for convex, relatively strongly convex, and nonconvex objectives. Existing accelerated BPG algorithms for convex objectives typically require additional assumptions on Bregman divergences, most notably triangle-scaling conditions, which can lead to slower convergence rates. We propose a family of geometry-accelerated BPG algorithms that exploit Bregman geometry in both proximal-gradient and mirror-space updates, without imposing additional geometric conditions such as triangle-scaling conditions. Our methods adapt…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-28","reactions":{}}