On the use of iterative methods in cubic regularization for unconstrained optimization

In this paper we consider the problem of minimizing a smooth function by using the Adaptive Cubic Regularized (ARC) framework. We focus on the computation of the trial step as a suitable approximate minimizer of the cubic model and discuss the use of matrix-free iterative methods. Our approach is alternative to the implementation proposed in the original version of ARC, involving a linear algebra phase, but preserves the same worst-case complexity count. Further we introduce a new stopping criterion in order to properly manage the ``over-solving'' issue arising whenever the cubic model is not an adequate model of the true objective function. Numerical experiments conducted by using a nonmonotone gradient method as inexact solver are presented. The obtained results clearly show the effectiveness of the new variant of ARC algorithm.