This package aids in the computation of the action of a matrix
polynomial on a vector, i.e. p(A)v, where A is a (square) matrix
(or a linear operator) that is supplied to the polynomial p. The
matrix polynomial p(A) is never formed explicitly, instead only its
action on v is evaluated. This is commonly used in time-stepping
algorithms for ordinary differential equations (ODEs) and discretized
partial differential equations (PDEs) where p is an approximation of
the exponential function (or the related φ functions:
φ₀(z) = exp(z), φₖ₊₁ = [φₖ(z)-φₖ(0)]/z, φₖ(0)=1/k!) on the
field-of-values of the matrix A, which for the methods in this
package needs to be known before-hand.
Other packages with similar goals, but instead based on matrix polynomials found via Krylov iterations are
Krylov iterations do not need to know the field-of-values of the
matrix A before-hand, instead, an orthogonal basis is built-up
on-the-fly, by repeated action of A on test vectors: Aⁿ*v. This
process is however very sensitive to the condition number of A,
something that can be alleviated by iterating a shifted and inverted
matrix instead: (A-σI)⁻¹ (rational Krylov). Not all matrices/linear
operators are easily inverted/factorized, however.
Moreover, the Krylov iterations for general matrices (then called
Arnoldi iterations) require long-term recurrences with mutual
orthogonalization along with inner products, all of which can be
costly to compute. Finally, a subspace approximation of the polynomial
p of a upper Hessenberg matrix needs to computed. The
real-symmetric/complex-Hermitian case (Lanczos iterations) reduces to
three-term recurrences and a tridiagonal subspace matrix. In contrast,
the polynomial methods of this packages two-term recurrences only, no
orthogonalization (and hence no inner products), and finally no
evaluation of the polynomial on a subspace matrix. This could
potentially mean that the methods are easier to implement on a GPU.