OCCVLGNADec 24, 2020

Global Convergence of Model Function Based Bregman Proximal Minimization Algorithms

arXiv:2012.13161v1
AI Analysis

This work provides a more general theoretical framework for designing optimization algorithms for a wider range of nonconvex nonsmooth problems, which is significant for researchers and practitioners dealing with complex optimization tasks.

The paper addresses the limitation of the L-smad property in handling nonsmooth functions in optimization by introducing the MAP property, which is applicable to a broader class of nonconvex nonsmooth composite problems. Based on this, they propose Model BPG, a globally convergent algorithm, demonstrating superior performance on phase retrieval and Poisson linear inverse problems compared to a state-of-the-art method.

Lipschitz continuity of the gradient mapping of a continuously differentiable function plays a crucial role in designing various optimization algorithms. However, many functions arising in practical applications such as low rank matrix factorization or deep neural network problems do not have a Lipschitz continuous gradient. This led to the development of a generalized notion known as the $L$-smad property, which is based on generalized proximity measures called Bregman distances. However, the $L$-smad property cannot handle nonsmooth functions, for example, simple nonsmooth functions like $\abs{x^4-1}$ and also many practical composite problems are out of scope. We fix this issue by proposing the MAP property, which generalizes the $L$-smad property and is also valid for a large class of nonconvex nonsmooth composite problems. Based on the proposed MAP property, we propose a globally convergent algorithm called Model BPG, that unifies several existing algorithms. The convergence analysis is based on a new Lyapunov function. We also numerically illustrate the superior performance of Model BPG on standard phase retrieval problems, robust phase retrieval problems, and Poisson linear inverse problems, when compared to a state of the art optimization method that is valid for generic nonconvex nonsmooth optimization problems.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes