The adaptive integrated mean: a novel perspective for acceleration and its extension to Bregman first-order methods

Applied Nanoscience · Springer Nature · 2026-09-23 · DOI: 10.1007/s10107-026-02423-4

Аннотация

We introduce the Adaptive Integrated Mean (AIM), a novel perspective on acceleration in first-order methods that interprets fast convergence as the result of an adaptive mean integrated directly into the optimization step. Within this framework, we first derive a streamlined and transparent analysis of FISTA, recovering its optimal rate $$O(1/k^{2})$$. Building on this insight, we extend the AIM principle to study acceleration for composite convex optimization in non-Euclidean settings under the smooth adaptability ( L -smad) framework, where classical acceleration techniques are known to fail in general. We derive the Bregman Accelerated Proximal Gradient (BAPG) method, a natural extension of FISTA, which is geometrically consistent with smooth adaptability and does not require strong convexity of the Bregman kernel. While acceleration is impossible for general Bregman geometries, we show that for the Boltzmann–Shannon entropy kernel (which generates the Kullback-Leibler divergence), BAPG achieves a convergence rate of $$O(\log k/k^{2})$$ under a mild additional assumption. This establishes the first accelerated rate for this fundamental non-Euclidean geometry in the smooth adaptable setting. Numerical examples illustrate and confirm the theoretical analysis.

Читать у издателя →

Похожие статьи