Core Concepts
The existence of spurious stationary points, which satisfy the zero extended Bregman stationarity measure but are not true stationary points, poses fundamental challenges to the analysis and design of Bregman proximal-type algorithms. These spurious points can trap the algorithms within a small neighborhood, even for convex problems, highlighting the inherent distinction between Euclidean and Bregman geometries.
Abstract
This paper investigates the reliability of existing stationarity measures based on Bregman divergence in distinguishing between stationary and non-stationary points. The key findings are:
The authors introduce an extended Bregman stationarity measure that is well-defined and continuous over the entire domain, including the boundary. This measure unifies and extends previous stationarity measures.
The authors prove that the extended stationarity measure being zero is necessary for a point to be stationary. However, they show that this condition is not sufficient, as there exist "spurious stationary points" that satisfy the zero stationarity measure but are not true stationary points.
The authors establish a hardness result: Bregman proximal-type algorithms are unable to escape from a spurious stationary point in finite steps when the initial point is unfavorable, even for convex problems. This highlights the inherent challenges in the design and analysis of Bregman divergence-based methods.
The authors provide simple convex and non-convex counter-examples to illustrate the existence of spurious stationary points. They also show that spurious points generally exist for a broad class of convex problems with polytopal constraints.
The paper's findings introduce both fundamental theoretical and numerical challenges to the machine learning and optimization communities, calling for new principles in algorithm design to address the issue of spurious stationary points.