The shift and invert spectral transformation is used to enhance
convergence to a desired portion of the spectrum.
If (**x**,) is an eigenpair for

(,) and
then

( - )
^{-1}**x** = **x**, where
= 1/( - )

_{j} = +
1/_{j},

In general, will be non-Hermitian even if and are both Hermitian. However, this is easily remedied. The assumption that is Hermitian positive definite implies that the bi-linear form

is an inner product . If is positive semi-definite and singular, then a semi-inner product results. We call this a weighted -inner product and vectors are called -orthogonal if . It is easy to show that if is Hermitian (self-adjoint) then is Hermitian (self-adjoint) with respect to this -inner product (meaning for all vectors ). Therefore, symmetry will be preserved if we force the computed basis vectors to be orthogonal in this -inner product. Implementing this -orthogonality requires the user to provide a matrix-vector product on request along with each application of . In the following sections we shall discuss some of the more familiar transformations to the standard eigenproblem. However, when is positive (semi) definite, we recommend using the shift-invert spectral transformation with -inner products if at all possible. This is a far more robust transformation when is ill-conditioned or singular . With a little extra manipulation (provided automatically inShift-invert spectral transformations are very effective and should even be used on standard problems () whenever possible. This is particularly true when interior eigenvalues are sought or when the desired eigenvalues are clustered. Roughly speaking, a set of eigenvalues is clustered if the maximum distance between any two eigenvalues in that set is much smaller than the maximum distance between any two eigenvalues of .

If one has a generalized problem (), then one must provide a way to solve linear systems with either , or a linear combination of the two matrices in order to use ARPACK. In this case, a sparse direct method should be used to factor the appropriate matrix whenever possible. The resulting factorization may be used repeatedly to solve the required linear systems once it has been obtained. If an iterative method is used for the linear system solves, the accuracy of the solutions must be commensurate with the convergence tolerance used for ARPACK. A slightly more stringent tolerance is needed for the iterative linear system solves (relative to the desired accuracy of the eigenvalue calculation). See [18,32,30,40] for further information and references.

The main drawback with using the shift-invert spectral transformation is that the coefficient matrix is typically indefinite in the Hermitian case and has in the interior of the convex hull of the spectrum in the non-Hermitian case. These are typically the most difficult situations for iterative methods and also for sparse direct methods.

The decision to use a spectral transformation on a standard
eigenvalue problem () or to use one of the simple modes
described in Chapter 2 is problem dependent. The simple
modes have the advantage that one only need supply a
matrix vector product . However, this approach is
usually only successful for problems where extremal non-clustered
eigenvalues are sought. In non-Hermitian problems, extremal means
eigenvalues near the boundary
of the convex hull of the spectrum of .
For Hermitian problems, extremal means
eigenvalues at the left or right end points of the spectrum of .
The notion of non-clustered (or well separated)
is difficult to
define without going into considerable detail. A simplistic notion
of a *well-separated* eigenvalue for a Hermitian problem would
be
for all with
Unless a matrix vector
product is quite difficult to code or extremely expensive computationally,
it is probably worth trying to use the simple mode first if you are
seeking extremal eigenvalues.

The remainder of this section discusses additional transformations that may be applied to convert a generalized eigenproblem to a standard eigenproblem. These are appropriate when is well conditioned (Hermitian or non-Hermitian).