A Global Convergence Analysis of Consensus ALADIN for Convex Optimization

Abstract

Distributed optimization problems are pervasive in machine learning and optimal control. In this paper, we study smooth strongly convex distributed consensus optimization problems. We present a distributed optimization algorithm for consensus problems based on the Consensus Augmented Lagrangian Alternating Direction Inexact Newton (C-ALADIN) framework. Our algorithm uses an auxiliary variable to decide when to update second-order information, enabling curvature exploitation without sacrificing global convergence. This contrasts with existing C-ALADIN methods, which require constant Hessian approximations and thus lose numerical advantages. Under smooth strong convexity, the algorithm converges globally, and the auxiliary variable converges sublinearly. Numerical experiments on logistic regression show that our algorithm outperforms baseline methods that use either fixed or updated Hessian information.