Extracting optimisers by non-commutative GNS construction is robust
ID Povh, Janez (Author), ID Volčič, Jurij (Author), ID Klep, Igor (Author)

This document has no files.
This document may have a phisical copy in the library of the organization, check the status via COBISS. Link is opened in a new window

Extracting eigenvalue optimisers in optimization of non commutative polynomials can be done efficiently by Gelfand Naimark-Segal (GNS) construction if the dual (moment) problem has flat optimum solution. However, in numerical computations the flatness is always subject to rounding threshold, i.e., often we can find only approximately flat dual solutions. In this talk we present how to apply GNS construction to approximately flat data and present sensitivity analysis results. We show that if the optimum of the dual problem is close to a flat solution then it yields a solution on the primal side that is close to an optimum solution. The distance to the optimum solution on the primal side can be expressed by the distance to a flat solution on the dual side. Similarly we can express for constraint optimization how close is the solution given by GNS construction to primal feasible and/or optimum solution in terms of the distance of the dual optimum to flatness. With extensive numerical evaluations we show that the established relations are usually very tight when we deal with random non-commutative polynomials. The focus will be on the (constrained) eigenvalue optimization for noncommutative polynomials, but we will also explain how the main results pertain to commutative and tracial optimization

Keywords:high performance computing, polynomial optimization, optimum extraction, flat extension, robustness
Typology:1.12 - Published Scientific Conference Contribution Abstract
Organization:FS - Faculty of Mechanical Engineering
Number of pages:Str. 244
PID:20.500.12556/RUL-101998 This link opens in a new window
COBISS.SI-ID:16150555 This link opens in a new window
Publication date in RUL:19.07.2018
Copy citation
Share:Bookmark and Share

Record is a part of a monograph

Title:ISMP 2018
Place of publishing:Bordeaux
Publisher:University of Bordeaux
COBISS.SI-ID:16150299 This link opens in a new window

Secondary language

Keywords:visoko zmogljivo računalništvo, polinomska optimizacija, ekstrakcija optimuma, ploske razširitve, robustnost

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections: