Preserving spectral properties of structured matrices under structured perturbations (Q1979367)

In applications it is often desired to change/perturb a given matrix \(A\in M_n(\mathbb{C})\) so that its ``bad'' attributes are replaced by the desired ones, while its ``good'' attributes remain unchanged. Such perturbations are called \textit{non spill-over}. The main results of the paper give parametric formulae to find all non spill-over perturbations in the case of two different attributes. This is done for unstructured matrices as well as for two types of structured matrices: the Lie type, \[\mathbb{L}:=\{A;\;\;A^{[\ast]}=-A\},\] and the Jordan type, \[\mathbb{J}=\{A;\;\;A^{[\ast]}=A\},\] where either \(A^{[\ast]}=H^{-1}A^{T} H\) or \(A^{[\ast]}=H^{-1}A^{\ast} H\) is the adjoint with respect to a bilinear form \((x,y)\mapsto x^THy\) or a sesquilinear form \((x,y)\mapsto x^\ast Hy\), respectively; in both forms \(H\) is a unitary (skew)Hermitian matrix. More specifically, all perturbations \(\Delta A\) for which a given subspace becomes (or remains) an invariant subspace of \((A+\Delta A)\) are found for unstructured matrices and structured matrices of Lie and Jordan type. Similar results are obtained also for all perturbations such that \((A+\Delta A)\) leaves invariant a given complementary pair of invariant subspaces. Moreover, all (non spill-over) structured perturbations which change specific eigenvalues but leave invariant the corresponding generalized eigenvectors are found. The formulae typically require computing the Moore-Penrose inverse, sometimes also the eigensystem, and are too long to be given in this review. Here is an example: if \(\mathcal{X}\subseteq\mathbb{C}^n\) is a \(p\)-dimensional invariant subspace for a matrix \(A\), then each perturbation \(\Delta A\) such that \(\mathcal{X}\) remains invariant for \((A+\Delta A)\) takes the form \[\Delta A=X(\Lambda_a-\Lambda)X^\dagger+Z(I-XX^\dagger).\] Here, \(X\) is an \(n\)-by-\(p\) matrix whose columns span \(\mathcal{X}\), \(X^\dagger\) its Moore-Penrose inverse, \(\Lambda\in M_p(\mathbb{C})\) is such that \(AX=X\Lambda\), while \(\Lambda_a\in M_{p}(\mathbb{C})\) and \(Z\in M_n(\mathbb{C})\) are arbitrary.