Chebyshev approximation via polynomial mappings and the convergence behaviour of Krylov subspace methods (Q5942488)

Let \(\varphi_m\) be a polynomial of degree \(m\) which maps a set \(R\subset{\mathbf C}\) to the set \(S\subset{\mathbf C}\). Then, under certain conditions, it is shown that if \(p_{n-1}\) is the best (in the sense of supremum norm) polynomial approximation of degree \(n-1\) on the set \(R\) to a function \(f\in C(R)\), then \(p_{n-1}\circ\varphi_m\) is the best polynomial approximation of degree \(nm-1\) to \(f\circ\varphi_m\) on \(\varphi_m(R)=S\). Also the extremal signatures for the two problems are explicitly related. Similar results hold for polynomial approximants of degree \(n\), but which are ``constrained'' by fixing one of its zeros. This is related to Krylov subspace iteration because the spectrum inclusion set of the original and the preconditioned system are often polynomially related like the sets \(R\) and \(S\) above. The speed of convergence of the Krylov method can be measured by the supremum norm of the optimal polynomial on the inclusion set. Thus the problem is to find the optimal polynomial which attains a minimal norm under the condition that it is equal to 1 in the origin. The latter problem can be brought into the previously discussed form and this allows to compare the speed-up for the rate of convergence obtained by the preconditioning. This is applied to matrices \(\mathcal A\) and preconditioners \(\mathcal M\) of the following block form \[ {\mathcal A}= \left[ \begin{matrix} A & B\cr B^T & 0 \end{matrix}\right], \] \[ {\mathcal M}= \left[ \begin{matrix} A & F\cr 0 & M \end{matrix} \right]. . \]