The cluster problem in multivariate global optimization (Q1338540)

If interval arithmetic branch-and-bound algorithms are applied to global optimization problems, an increase of the number of boxes which are near the solution and which have not been discarded yet, can be observed as the mesh width of the underlying subdivision tends to zero. This phenomenon is investigated under the assumptions that no acceleration devices such as Newton methods are used and that the Hessian matrix of the objective function is positive definite. It turns out that this number of boxes depends mainly on the order of the inclusion function which has been chosen for the objective function, on the smallest eigenvalue of the Hessian, on the mesh width and on the dimension of the space.