First-order rules for nonsmooth constrained optimization

Nonsmooth analysis (49J52) Set-valued and variational analysis (49J53) Optimality conditions for problems in abstract spaces (49K27) Programming in abstract spaces (90C48)

Review text

The paper is concerned with first order rules, meaning first order necessary conditions for local minima of lower semicontinuous (l.s.c.) functions, covering also the constrained minimization case. In fact, the author treats the equivalent case of the minimization of sums of l.s.c. functions, giving necessary conditions for the local minima of these sums in terms of the subdifferential of each function. The subdifferential \(\partial f(x) \subset X\times X^ *, \; X \,\) a Banach space, \(X^ *\) its conjugate, and \(f:X\to \mathbb R\cup \{\infty\}\), is defined axiomatically such that it agrees with the usual subdifferential when \(f\) is convex and \(\partial F(x,y) \subset \partial f(x) \times \partial g(y),\) for \(F = f+ g: X\times Y \to \mathbb R \cup \{\infty\}.\) One puts \(\partial f = \{(x,x^ *) : x^ * \in \partial F(x)\}\). The author proves the equivalence of seven first order rules, labeled (R1) to (R7), extending and synthesizing a lot of previous results obtained by J. Borwein and Q. J. Zhu, R. Deville and E. M. El Haddad, A. Ioffe, D. Zagrodny, F. H. Clarke and Yu. S. Ledyaev. For instance, the rule (R1) says that if \(f+\varphi\) attains a strong local minimum at \(x_ 0\) and \(\varphi\) is convex and Lipschitz, then there exist sequences \((\tilde x_ n,\tilde x_ n^ *)_ n \in \partial f\) and \((\tilde y_ n,\tilde y_ n^ *)_ n \in \partial \varphi\), such that \(\tilde x_ n\to x_ 0,\; \tilde y_ n\to x_ 0,\; f(\tilde x_ n) \to f(x_ 0)\) and \(\tilde x_ n^ * + \tilde y_ n^ * \to 0.\) It is clear that in order that these rules hold, some smoothness conditions on the Banach space \(X\) are required. A key notion is that of uniform infimum defined, for \(A\subset X,\) by \(r_ A(x) := \lim_{\delta \to 0} \inf \{f(x) : d_ A(x) \leq \delta\},\) where \(d_ A\) stands for the distance function. The author gives also several conditions under which the uniform infimum \(r_ A(\sum_ {i=1}^ kf_ i)\) agrees with the usual (called algebraical) infimum \(\inf f(A)\).

Recommendations

Cites work

Cited in

(16)

This page was built for publication: First-order rules for nonsmooth constrained optimization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q5940175)