A Geometry-Adaptive Regularized Newton-Type Method for Manifold-Affine Intersection Problems

We propose a regularized algorithm, Regularized Newton-SLRA (RN-SLRA), for local manifold--affine intersection problems under weak intersection conditions, motivated in particular by structured low-rank approximation (SLRA).

Newton-SLRA is an efficient method for manifold--affine intersection problems, but its well-definedness relies on the transversality condition between the manifold and the affine subspace, a condition that may fail in practice.

RN-SLRA overcomes this difficulty by introducing a regularization term.

We prove that, under the intrinsic transversality condition, RN-SLRA converges linearly to the intersection, while under the transversality condition it achieves higher-order convergence, including quadratic convergence for a suitable choice of the regularization parameter.

We also study an inexact-projection variant, in which the projection onto the manifold is computed approximately, and show that the same local linear and quadratic convergence properties are preserved under the corresponding assumptions.

Numerical experiments on constructed degenerate instances and Hankel-structured examples illustrate improved robustness in settings where Newton-SLRA may fail.