Computing stable eigendecompositions of matrices (Q1092606): Difference between revisions

Let \(T_ 0\) be a complex \(n\times n\) matrix, \(\epsilon >0\) be a tolerance and \(T(\epsilon)=\{T: \| T-T_ 0\| \leq \epsilon \}\) where the norm is the Frobenius norm. Eigendecompositions of the matrices are written as \(T=SJS^{-1}\) where \(J=diag(J_ 1,...,J_ r)\) is block diagonal. The author explores the problem how the spectrum \(\sigma (T)=\cup^ r_{i=1}\sigma (J_ i)\) varies as T varies within T(\(\epsilon)\). A crucial assumption for the investigation is that the eigendecomposition is stable, which means that the correspondences \(T\mapsto S\) and \(T\mapsto J\) are continuous and that the dimensions of the blocks \(J_ i\) remain constant as long as \(T\in T(\epsilon)\). Let further \(\sigma_ i=\sigma (J_ i)\). Related to the problem of keeping the eigendecomposition stable is the concept of a dissociation of \(\sigma_ i\) and \(\sigma_ j\), which is the smallest norm of a perturbation of T such that a \(\lambda_ i\in \sigma_ i\) coalesce with some \(\lambda_ j\in \sigma_ j\). The main results are concerned with upper and lower bounds on the dissociation and with bounds on the condition number of S as \(T\in T(\epsilon)\).