{"id":"arxiv/2609.35665","url":"https://arxiv.org/abs/2609.35665","tags":["math.OC","cs.LG"],"score":0,"title":"Learned Preconditioning for a Primal-Dual Interior-Point Method","domain":"math.OC","source":"arxiv","authors":["Abhinav Madabhushi","Jialin Liu","Minxin Zhang"],"reviews":0,"subject":"math.OC","summary":"Interior-point methods (IPMs) are among the most widely used algorithms for constrained optimization, yet their Newton-based search directions require costly second-order information and large linear-system solves. Learning to optimize offers cheaper updates learned from data, but the singular behavior of logarithmic barriers near constraint boundaries makes IPMs highly sensitive to perturbations, complicating both warm starting and learning reliable updates. We introduce pdLIP, an IPM for smooth nonlinear programs that integrates learned preconditioning with pdProj, an all-shifted…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-09-28","reactions":{}}