Adaptive Cubic Regularisation Methods under Dynamic Inexact Hessian Information for Nonconvex Optimisation

ARC methods are Newton-type solvers for unconstrained, possibly nonconvex, optimisation problems. In this context, a novel variant based on dynamic inexact Hessian information is discussed. The approach preserves the optimal complexity of the basic framework and the main probabilistic results on the complexity and convergence analysis in the finite-sum minimisation setting will be shown. At the end, some numerical tests on machine learning problems, ill-conditioned databases and real-life applications will be given, in order to confirm the theoretical achievements. Joint work with Stefania Bellavia and Benedetta Morini (Università di Firenze).