Spectral functions and smoothing techniques on Jordan algebras
Successful methods for a large class of nonlinear convex optimization problems have recently been developed. This class, known as self-scaled optimization problems, has been defined by Nesterov and Todd in 1994. As noticed by Guler in 1996, this class is best described using an algebraic structure known as Euclidean Jordan algebra, which provides an elegant and powerful unifying framework for its study. Euclidean Jordan algebras are now a popular setting for the analysis of algorithms designed for self-scaled optimization problems : dozens of research papers in optimization deal explicitely with them.
This thesis proposes an extensive and self-contained description of Euclidean Jordan algebras, and shows how they can be used to design and analyze new algorithms for self-scaled optimization.
Our work focuses on the so-called spectral functions on Euclidean Jordan algebras, a natural generalization of spectral functions of symmetric matrices. Based on an original variational analysis technique for Euclidean Jordan algebras, we discuss their most important properties, such as differentiability and convexity. We show how these results can be applied in order to extend several algorithms existing for linear or second-order programming to the general class of self-scaled problems. In particular, our methods allowed us to extend to some nonlinear convex problems the powerful smoothing techniques of Nesterov, and to demonstrate its excellent theoretical and practical efficiency.
School:Université catholique de Louvain
Source Type:Master's Thesis
Keywords:spectral functions jordan algebras convex optimization
Date of Publication:09/22/2006