Matrix nearness problems: Do real inputs admit real solutions?

Abstract

Given a real square matrix $A$ and a nonempty closed target set of complex matrices $\mathcal{E}$, invariant by complex conjugation and containing real matrices, does the subset of $\mathcal{E}$ consisting of the matrices nearest to $A$ in the Frobenius distance always contain a real matrix? We give negative answers when $\mathcal{E}$ is either the set of normal matrices (solving a question posed by N. Higham) or the set of matrices whose eigenvalues lie in the closed left half-plane (solving a question posed by the first author and F. Poloni). For the latter problem, we also argue that the ratio between the real and complex distances is unbounded in every dimension $n\ge 3$, and this holds for every distance induced by a unitarily invariant norm. For both problems and $n\ge 3$, we show that a real input $A$ uniformly drawn from the unit sphere $\Vert A{\Vert }_F=1$ has no real minimizer over $\mathcal{E}$ with probability strictly between $0$ and $1$. The paper is complemented by some further results that are valid for more general target sets $\mathcal{E}$.