Convergência completa do método do gradiente com busca linear exata e inexata

Abstract

In this work we use the gradient method to minimize, without restrictions, convex and pseudoconvex continuously differentiable functions. An important theme considered is the path length determination. We have that, when minimizing pseudoconvex functions, the linear search is exact. In this case, we present the first algorithm to obtain the path length, where will be included a quadratic regularization term, in the proximal point method sense. When dealing with the minimization of convex functions case, we have that the linear search is not exact. To obtain the path length, two algorithms are presented: the former needs that the gradient of the objective function satisfies a Lipschitz condition with a known constant L > 0. The latter is based on the work of Dennis-Schnabel (see [4]). The three process are based on the quasi-Fejér convergence principle. Although these descent methods need that the objective functions to be minimized have bounded level sets, in order to establish that the limit points are stationary, this approach guarantees the complete convergence of every sequence to a minimizer of the function without the hypothesis of bounded level sets.