Skip to main content


eCommons@Cornell

eCommons@Cornell >
College of Engineering >
Computer Science >
Computer Science Technical Reports >

Please use this identifier to cite or link to this item: http://hdl.handle.net/1813/6687
Title: Computing a Trust Region Step for a Penalty Function
Authors: Coleman, Thomas F.
Hempel, Christian
Keywords: computer science
technical report
Issue Date: Jul-1987
Publisher: Cornell University
Citation: http://techreports.library.cornell.edu:8081/Dienst/UI/1.0/Display/cul.cs/TR87-847
Abstract: We consider the problem of minimizing a quadratic function subject to an ellipsoidal constraint when the matrix involved is the Hessian of a quadratic penalty function (i.e., a function of the form $p(x) = f(x) + \frac{1}{2\mu} c(x)^{T} c(x))$. Most applications of penalty functions require $p(x)$ to be minimized for values of $\mu$ decreasing to zero. In general, as $\mu$ tends to zero the nature of finite precision arithmetic causes a considerable loss of information about the null space of the constraint gradients when $\nabla^{2}p(x)$ is formed. This loss of information renders ordinary trust region Newton's methods unstable and degrades the accuracy of the solution to the trust region problem. The algorithm of More and Sorenson [1983] is modified so as to be more stable and less sensitive to the nature of finite precision arithmetic in this situation. Numerical experiments clearly demonstrate the stability of the proposed algorithm.
URI: http://hdl.handle.net/1813/6687
Appears in Collections:Computer Science Technical Reports

Files in This Item:

File Description SizeFormat
87-847.pdf2.12 MBAdobe PDFView/Open
87-847.ps629.59 kBPostscriptView/Open

Refworks Export

Items in eCommons are protected by copyright, with all rights reserved, unless otherwise indicated.

 

© 2014 Cornell University Library Contact Us