The global convergence of partitioned BFGS on problems with convex decompositions and Lipschitzian gradients (Q1176574)

This paper extends \textit{M. J. D. Powell's} global convergence result [Nonlin. Program., Proc. Symp. Appl. Math. New York City 1975, 53-72 (1976; Zbl 0338.65038)] to the partitioned BFGS method of the author and \textit{Ph. L. Toint} [\ {}Numer. Math. 39, 119-137 (1982; Zbl 0482.65035)]. The objective function \(f(x)\) is assumed to be convex on some convex domain \({\mathcal D}\subset\mathbb{R}^ n\); the gradient \(g(x)=\nabla f(x)\) is only required to be Lipschitz continuous on \(\mathcal D\) rather than differentiable. The paper analyses the global and then local convergence properties of the partitioned BFGS method in the case of the objective function having the form \(f(x)=\sum f^{(i)}(x)\), where the convex element functions \(f^{(i)}(x)\) are uniformly Lipschitz continuous on \(\mathcal D\), an open convex neighbourhood of a bounded level set \({\mathcal L}=\{x\in\mathbb{R}^ n:\;f(x)\leq\bar f\}\). Under these conditions it is shown that inexact BFGS with damped updating of all approximating Hessians whose element functions are not uniformly convex converges to the minimizer \(x_ *\in\mathcal L\) of \(f\) from any starting point in \(\mathcal L\). The corresponding value of \(f\) is \(f_ *=\inf f\) and that of the gradient \(g\) is \(g_ *=0\). Further, if the jump discontinuities in the element Hessians are zero, it is shown that \(R\)-superlinear convergence is obtained.