An inexact algorithm for composite nondifferentiable optimization (Q1823153)

The paper presents an inexact version of Fletcher's trust-region algorithm with second-order corrections for minimizing composite non- smooth functions. The composite non-smooth optimization problem is: \(\min \{f(x)+h(c(x))\}\), where f and c are twice continuously differentiable and h is a polyhedral convex function of the form \(h(c)=\max h^ T_ ic+b_ i.\) This paper modifies the second-order correction method of \textit{R. Fletcher} [Numerical analysis, Proc. 9th Bienn. Conf., Dundee/Scotl. 1981, Lect. Notes Math. 912, 85-114 (1982; Zbl 0476.65048)] to allow inexactness in the solution of the subproblem at each iteration. The modified algorithm is shown to retain the global and local convergence properties of the original method under certain conditions.