A Proximal Newton Framework with Local Superlinear Convergence for Simple Bilevel Optimization

In this paper, we study a penalty method for simple bilevel optimization in which the lower-level objective has a convex composite structure. Our approach is based on a vanishing-penalty reformulation, which gives rise to a sequence of penalized problems as the penalty parameter decreases. A key difficulty is that metric subregularity of the lower-level optimality mapping does not, in general, yield a uniform error bound for the penalized problems as the penalty parameter vanishes, preventing a direct application of standard proximal Newton theory. To address this difficulty, we first analyze a regularized proximal Newton step by exploiting the specific structure of simple bilevel problems, and establish the local estimates that underpin superlinear convergence. Building on this local analysis, we propose SiB-Newton, which incorporates the proximal Newton step into a globalization strategy. We establish convergence guarantees for SiB-Newton and show that, under suitable local conditions, Newton steps are eventually accepted and the iterates converge at a superlinear rate. Numerical experiments on synthetic examples verify the predicted local superlinear convergence rate.