Stopping rules and backward error analysis for bound-constrained optimization (Q639367)

The authors consider the problem of minimizing a nonlinear differentiable function on a set \(F= \{x=(x_1,\dots,x_n)\); \(L\leq x\leq U\}\), where \(L\), \(U\) are given lower and upper bounds on \(x\). The aim is to find a first-order critical point \(x^*\) of the objective function. The paper deals with the problem of finding appropriate termination criteria for iterative solution methods of the original problem. The authors suggest to replace the question ``How far from point \(x^*\) is a current approximation \(x^k\)?'' by the question ``If there exists a perturbed minimization problem (P), for which \(x^k\) is the first-order solution, how far from the original problem is problem (P)?'' To answer the latter question, the authors propose three different norms appropriate for measuring the distance between the optimization problems. The problem of finding appropriate termination criteria is then solved using both a scalarization and a multicriteria approach. Theoretical results are illustrated by numerical examples in the concluding part of the paper.