Quasi-convex sets and size \(\times\) curvature condition, application to nonlinear inversion (Q1179160)

The motivation of this paper is to develop mathematical tools to study the local minima and wellposedness of nonlinear least square problems which arise in parameter estimation and other inverse problems. The first part of the paper introduces a class of subsets of a Hilbert space, called the ``quasi-convex sets'', which possess a neighborhood on which the projection exists (when the set is closed), is unique, Lipschitz continuous, and local minima, though not excluded, can be distinguished from the global one by their value. The characterization relies on properties which are checked on a family of paths which play for quasiconvex sets the role of segments for a convex set. The proof follows very closely the classical proof for the projection on convex set, to which it reduces when the set is convex and equipped with the family of segments. The second part of the paper gives a constructive sufficient condition to recognize quasiconvex sets, which procedures a cylindrical neighborhood where the projection is well-behaved. This part has lost most of its interest since a more precise sufficient condition has been given [see the author, SIAM J. Control and Optimization 29, No. 6, 1348-1372 (1991; Zbl 0745.49008)], which moreover exludes the possibility of local minima. In the last part, rather academic applications to wellposedness of nonlinear least square problems are presented. A more realistic application can be found in a paper of the author and \textit{K. Kunisch} [``A geometric theory for the \(L^ 2\) stability of the inverse problem in a \(1-D\) elliptic equation from and \(H^ 1\) observation'', J. Appl. Math. Optimization (to appear)].