“…For standard L-BFGS the use of a diagonal seed matrix other than a scaled identity has been studied in [37,24,51,49,42,32,12,16,35,8], but convergence is usually shown for strongly convex objectives or not at all, except in [35], where global convergence is proved for the non-convex case in the sense that lim inf k→∞ ∥∇J (x k )∥ = 0. In contrast, we have lim k→∞ ∥∇J (x k )∥ = 0 in that case and a linear rate of convergence, cf.…”