On Eigenvector Bounds

We show under very general assumptions that error bounds for an individual eigenvector of a matrix can be computed if and only if the geometric multiplicity of the corresponding eigenvalue is one. Basically, this is true if not computing exactly like in computer algebra methods. We first show, under general assumptions, that nontrivial error bounds are not possible in case of geometric multiplicity greater than one. This result is also extended to symmetric, Hermitian and, more general, to normal matrices. Then we present an algorithm for the computation of error bounds for the (up to normalization) unique eigenvector in case of geometric multiplicity one. The effectiveness is demonstrated by numerical examples.This revised version was published online in October 2005 with corrections to the Cover Date.