Optimal expansion of subspaces for eigenvector approximations (Q2469518): Difference between revisions

The author considers the problem of optimal expansion of a subspace for approximating an eigenvector. More theoretically, the problem is, given a subspace \(\mathcal{S}\), what vector \(v \in \mathcal{S}\) is such that \(\mathcal{S} + \mathrm{span}\{ Av \}\) is optimal for approximating a given unit eigenvector \(x\). The quality of the subspace \(\mathcal{S}\) is measured by the angle \(\cos(\theta(x,\mathcal{S}))\) defined by \[ \cos(\theta(x,\mathcal{S})) := \max_{z \in \mathcal{S}} \cos(\theta(x,z))= \max_{0 \neq z \in \mathcal{S}} \frac{| x^T z| }{\| x\| \| z\| }. \] The optimal subspace expansion problem is to find \(v \in \mathcal{S}\) that maximizes \(\cos(\theta(x,\mathcal{S} +\mathrm{span}\{ Av \}))\). The author presents a theoretical solution to that problem. This solution is not computable since it depends on the unknown eigenvector \(x\). Nevertheless, the author presents some hints that may suggest that the Ritz vector is a good choice as a vector for expansion.