Unconstrained Minimization by Interpolation: Rates of Convergence.
Abstract
The main attraction of Newton's method for finding a stationary point is its quadratic convergence. However, it necessitates computation and inversion of the second order derivative matrix. Common minimization algorithms approximate the Hessian or its inverse by first order (i.e., gradient) information. First order information algorithms in common use are known to have at best superlinear rate of convergence. We analyze the rate of convergence of a new class of algorithms based on n-dimensional interpolation. In particular, we present a class of algorithms which use first order information only, while maintaining quadratic convergence.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1980
- Accession Number
- ADA095025
Entities
People
- J. Barzilai
Organizations
- University of Texas at Austin