Whitney differentiability of optimal-value functions for bound-constrained convex programming problems. (4th March 2019)

• Record Type: Journal Article
• Authors: Khan, Kamil A.
• Is Part Of: Optimization. Volume 68:Number 2/3(2019)
• Page Start: 691
• Page End: 711
• Publication Date: 2019-03-04
• Subjects: Parametric optimization -- convex underestimators -- smoothing methods; 26B05 -- 90C31 -- 90C25; Mathematical optimization -- Periodicals; 519.7
• DOI: 10.1080/02331934.2018.1534108
• Languages: English
• ISSNs: 0233-1934

Abstract

ABSTRACT: In the spirit of the Whitney Extension Theorem, consider a function on a compact subset of Euclidean space to be 'Whitney-differentiable' if it is a restriction of a continuously Fréchet-differentiable function with an open domain. Whitney-differentiable functions have been shown to have useful (yet possibly nonunique) derivatives and calculus properties even on the boundaries of their domains. This article shows that optimal-value functions for bound-constrained convex programmes with Whitney-differentiable objective functions are themselves Whitney-differentiable, even when the linear-independence constraint qualification is not satisfied. This result extends classic sensitivity results for convex programmes, and generalizes recent work. As an application, sufficient conditions are presented for generating continuously differentiable convex underestimators of nonconvex functions for use in methods for deterministic global optimization in the multivariate McCormick framework. In particular, the main result is applied to generate Whitney-differentiable convex underestimators for quotients of functions with known Whitney-differentiable relaxations.