Backward error of an approximate eigenpair (source code)

= Backward error of an approximate eigenpair

Let $r=A\widetilde v-\widetilde\lambda\widetilde v$ be an <eigenpair residual>, with $\|\widetilde v\|_2=1$. Among unrestricted <matrix> perturbations $E$, the smallest <spectral norm> making $(\widetilde\lambda,\widetilde v)$ an exact <eigenpair> of $A+E$ is $\|r\|_2$.

Indeed, exactness requires $E\widetilde v=-r$, so the <induced matrix norm> inequality gives $\|E\|_2\ge\|r\|_2$. Equality is attained by the rank-one <matrix>
$$
E=-r\widetilde v^*,
$$
where $*$ is <conjugate transpose>. Its <spectral norm> is $\|r\|_2\|\widetilde v\|_2=\|r\|_2$. For real data, use ordinary <transpose> instead.

This is an unstructured <backward error>: requiring $E$ to preserve a <symmetric matrix> structure or a sparsity pattern is an additional constraint. A small <backward error> also need not give a comparably small <eigenvalue> error for a nonnormal <matrix>.