On the existence of affine invariant descent directions. (2nd September 2020)

• Record Type: Journal Article
• Authors: Dai, Yu-Hong; Jarre, Florian; Lieder, Felix
• Is Part Of: Optimization methods and software. Volume 35:Number 5(2020)
• Page Start: 938
• Page End: 954
• Publication Date: 2020-09-02
• Subjects: Affine invariance -- descent direction -- Newton direction; 90C30 -- 90C26; Mathematical optimization -- Periodicals; Algorithms -- Periodicals; 519.7
• DOI: 10.1080/10556788.2020.1740221
• Languages: English
• ISSNs: 1055-6788

Abstract

Abstract : This paper begins with a brief review of affine invariance and its significance for iterative algorithms. It then explores the existence of affine invariant descent directions for unconstrained minimization. While there may exist several affine invariant descent directions for smooth functions at a given point, it is shown that for quadratic functions, there exists exactly one invariant descent direction in the strictly convex case and generally none in the case where the Hessian is singular or indefinite. These results can be generalized to smooth nonlinear functions and have implications regarding the initialization of minimization algorithms. They stand in contrast to recent works on constrained convex and nonconvex optimization for which there may exist an affine invariant 'frame' that depends on the feasible set and that can be used to define an affine invariant descent direction.