Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper establishes a centre manifold theorem specifically for maps defined along manifolds with boundaries of fixed points, which is a significant extension of existing theoretical frameworks. The authors apply this theorem to analyze the dynamics of gradient descent with large step sizes in the context of two-layer matrix factorization problems, revealing critical insights into convergence behavior. The key result demonstrates how the structure of the centre manifold influences the stability and convergence properties of the optimization process in this setting.
Large step sizes in gradient descent can lead to unexpected convergence dynamics, as revealed by a new centre manifold theorem tailored for fixed-point manifolds.
We prove a centre manifold theorem for a map along a manifold-with-boundary of fixed points, and provide an application to the study of gradient descent with large step size on two-layer matrix factorisation problems.