Misconvergence in the Lanczos Algorithm
Abstract
The Lanczos algorithm generates Ritz values in order to approximate eigenvalues. If some eigenvalues are clustered then a Ritz value may hover at a wrong value for a good number of steps. We study this phenomenon and focus on the point of discovery, the first step at which it is certain that there is a hidden eigenvalue in the vicinity of stabilized Ritz values. Both before and after this point the Ritz value behavior is routine - but for different eigenvalue configurations. The effective spread at step j is an interval guaranteed to contain all unknown eigenvalues. The notion of Ritz intervals leads to a computable counterpart to the exact theory.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1987
- Accession Number
- ADA206862
Entities
People
- Beresford N. Parlett
Organizations
- University of California, Berkeley