     Next: Overview of Available Algorithms. Up: Hermitian Eigenvalue Problems Previous: Hermitian Eigenvalue Problems   Contents   Index

# Introduction

In this chapter we treat the standard Hermitian, or most often real symmetric, eigenvalue problem (HEP), (16)

where . It is the most commonly occurring algebraic eigenvalue problem, for which we have the most reliable theory and the most powerful algorithms available.

A summary of the basic theory about the Hermitian eigenvalue problem (4.1) is given in §2.2. It is known that the problem (4.1) has real eigenvalues , which we may order increasingly so that . Note that several eigenvalues may coincide. In many applications the matrix is positive definite, (or positive semidefinite, ). Even in the cases where positive definiteness can be used to advantage, we choose to treat Hermitian with a general distribution of eigenvalues in this chapter.

When is Hermitian, it is always possible to find mutually orthogonal eigenvectors, , so that (17)

where , , and . The eigenvectors are not unique; what is unique is the invariant subspace for each different eigenvalue. For a Hermitian matrix , the invariant subspace is of the same dimension as the multiplicity of the eigenvalue. It is important to keep in mind that when certain eigenvalues coincide, their eigenvectors lose their individuality: there is no way of saying that one set of vectors are the eigenvectors of a multiple eigenvalue. Two different algorithms, and even two different runs with the same algorithm, will give different representative eigenvectors in the invariant subspace. On the other hand, a user is entitled to demand that an algorithm produce eigenvector approximations that are orthogonal to each other, even if the matrix has multiple eigenvalues.

The eigenvalues of are always well-conditioned. If a perturbation is added to the matrix , then each of the eigenvalues is perturbed at most as far as the spectral norm , (18)

There are several stronger results available, like the Wielandt-Hoffman inequality that says that the sum of squares of the differences between the eigenvalues is majorized by the sum of squares of the elements of ; see §4.8 and references therein. In many cases, one is interested in eigenvalues that are small compared to the norm , and for such eigenvalues the inequality (4.3) is not very satisfactory, since it allows large relative perturbations for such small eigenvalues. It is possible to develop perturbation bounds for relative perturbations under certain additional assumptions.

There is no result as simple as (4.3) available for bounding the perturbation of eigenvectors: one has to add an assumption of separation between the eigenvalues. Let us cite the venerable theorem of Davis and Kahan from .

Let us assume that is an approximation of the eigenpair of , where is normalized so that . The best'' corresponding to is the Rayleigh quotient , so we assume that has this value. Suppose that is closer to than any other eigenvalues of , and let be the gap between and any other eigenvalue: . Define the residual vector as ; then we have (19)

Under the same assumptions, we have the improved bound on the eigenvalue approximation, |-_i|{||r||_2,||r||_2^2}. The first alternative in this minimum is equivalent to the previous bound (4.3), since is an eigenvalue of the perturbed matrix with a rank-one perturbation of norm (it is not necessary for to be Hermitian in the bound (4.3)). The reader is referred to §4.8 for further details.

Subsections     Next: Overview of Available Algorithms. Up: Hermitian Eigenvalue Problems Previous: Hermitian Eigenvalue Problems   Contents   Index
Susan Blackford 2000-11-20