Power method with a dominant eigenvalue cluster (source code)

= Power method with a dominant eigenvalue cluster

Suppose a finite-dimensional <matrix> $A$ is <diagonalizable> and its <eigenvalues> of largest modulus $r>0$ form a cluster, separated in modulus from the remaining <eigenvalues>. If the initial <vector> has a nonzero projection onto the sum $E$ of their <eigenspaces>, normalized <power method> iterates approach $E$ in distance, but need not converge to one vector.

For example, $A=\operatorname{diag}(1,-1,1/2)$ and initial vector $(1,1,1)$ give alternating limiting directions in the first two coordinates. If the initial vector has no component in $E$, it never discovers that cluster. Without diagonalizability, <Jordan blocks> introduce <polynomial> growth factors, so a statement about the dominant subspace requires additional care.