The algorithm achieves faster convergence by separating constant and gap-dependent curvature effects, with a unified update structure that handles non-Euclidean geometries (p>2) as easily as standard Euclidean optimization (p=2).
This paper presents A²ODA, a new optimization algorithm that combines acceleration techniques with adaptive preconditioning based on p-norm geometry. The method is designed to efficiently solve large-scale convex optimization problems without requiring line searches, using a simpler unified approach that works across different norm geometries.