AMRVW. Fast and backward stable computation of roots of polynomials

Implementation of core-chasing algorithms for finding eigenvalues of factored matrices. FORTRAN code for such methods is provided in the eiscor repository.

This repository provides a Julia package implementing the methods, as applied to the problem of finding the roots of polynomials through the computation of the eigenvalues of a sparse factorization of the companion matrix in:

• Fast and Backward Stable Computation of Roots of Polynomials. Jared L. Aurentz, Thomas Mach, Raf Vandebril, and David S. Watkins SIAM J. Matrix Anal. Appl., 36(3), 942–973. (2015) https://doi.org/10.1137/140983434

• Fast and backward stable computation of roots of polynomials, Part II: backward error analysis; companion matrix and companion pencil. By Jared L. Aurentz, Thomas Mach, Leonardo Robol, Raf Vandebril, David S. Watkins; arXiv:1611.02435

The methods are summarized in monograph format:

Core-Chasing Algorithms for the Eigenvalue Problem; by Jared L. Aurentz, Thomas Mach, Leonardo Robol, Raf Vandebril, and David S. Watkins; https://doi.org/10.1137/1.9781611975345

As well, the twisted algorithm from "A generalization of the multishift QR algorithm" by Raf Vandebril and David S. Watkins; https://doi.org/10.1137/11085219X is implemented here.

The core-chasing algorithms utilize Francis's QR algorithm on sparse factorizations of the respected companion matrix. For polynomials with real coefficients, the storage requirements are O(n) and the algorithm requires O(n) flops per iteration, or O(n^2) flops overall. The basic QR algorithm applied to the full companion matrix would require O(n^2) storage and O(n^3) flops overall.

julia> import AMRVW; const A = AMRVW
julia> p4 = [24.0, -50.0, 35.0, -10.0, 1.0]  # (x-1) * (x-2) * (x-3) * (x-4)

5-element Array{Float64,1}:
  24.0
 -50.0
  35.0
 -10.0
   1.0

julia> A.roots(p4)
4-element Array{Complex{Float64},1}:
 0.9999999999999996 + 0.0im
 2.0000000000000027 + 0.0im
 2.9999999999999876 + 0.0im
 4.000000000000012 + 0.0im

By means of comparison, using the Polynomials package:

julia> using Polynomials
julia> roots(Polynomial(p4))

4-element Array{Float64,1}:
 1.000000000000002
 1.9999999999999805
 3.0000000000000386
 3.9999999999999822

The advantage of the methods comes from the fact that this algorithm can be used for much larger polynomials.

• Compared to the roots function of the Polynomials package, the methods are faster once the degree is around 75, and much faster as this grows. These methods are O(n) in storage and O(n^2) in time, whereas roots is O(n^2) in storage (the full companion matrix is created) and O(n^3) in time (the running time of a generic eigenvalue solver). As well, the roots function only computes over Float64 values, not generic floating point values.

• Compared to the roots function of the PolynomialsRoots package, these methods are a bit slower, and perhaps a bit less accurate. This roots function is O(n) in storage and also appears to be O(n^2) in time. This roots function works over generic floating point values. However, this roots method will run into numeric issues for polynomials of degree n larger than 350 or so.

The backward stability of the methods is shown in the second paper to grow linearly in the norm of the coefficients, so the following should be quite accurate and is computable in a reasonable time:

## by DOI:	10.1142/S0219199715500522, this should have expected value ~ 2/pi*log(n) + .625738072 + 2/(pi*n) ~ 6.48
julia> rs = rand(Float64, 10_000) .- 1/2
julia> @time rts  = A.roots(rs)
julia> sum(isreal, rts)
 15.955615 seconds (35 allocations: 1017.297 KiB)
5

As this is relatively speedy, statistics can be generated, albeit the following will take some time to finish:

julia> xs = [sum(isreal, A.roots(randn(3000))) for _ in 1:3000]
julia> using StatsBase
julia> xbar, s = mean_and_std(xs)
julia> n = 3000
julia> xbar .+ 1.96*s/sqrt(n) * [-1,1], 2/pi*log(n) + .625738072 + 2/(pi*n)
 ([5.67865426156726, 5.820012405099407], 5.7229621769994745)

There are no exported functions, as of now. But the internal functions may be of interest. For example, the paper Fast and stable unitary QR algorithm discusses a situation where a matrix A is unitary Hessenberg, and so is factored in terms of a descending chain of rotators. To fit this matrix into the current framework, we have, for example:

using LinearAlgebra
T = Float64
const A =  AMRVW
Qs = A.random_rotator.(T, 1:10)
Q = A.DescendingChain(Qs)
QF = A.q_factorization(Q)
F = A.QRFactorization(QF)
eigvals(F)

Which can be compared with:

MI = diagm(0 => ones(T, 11))
Qs * MI |> eigvals