Stable perturbations of nonsymmetric matrices (Q1188440)

The investigation in this paper is to scrutinize the problem what perturbations to the matrix are associated with stability when a matrix \(A\) is given on the boundary of the set of stable matrices, i.e., with no eigenvalue having positive real part and one or more eigenvalues lying on the imaginary axis. To be more specific, suppose that \(A^{(0)}\) is an \(n\times n\) Jordan block \(J_ 0\), with zeros on the diagonal and ones on the main superdiagonal, and consider a real perturbation \(A(\varepsilon)\), with \(A(0)=J_ 0\). It is shown that if \(A(\varepsilon)\) is a real analytic matrix function of the form \(A(\varepsilon)=J_ 0+\varepsilon B+0(\varepsilon^ 2)\), then a necessary condition on \(B\) for the stability of \(A(\varepsilon)\) on an interval \((0,\varepsilon_ 0)\) and a sufficient condition on \(B\) for the existence of such a family \(A(\varepsilon)\) is: (i) \(\text{tr} B\leq 0\); (ii) the sum of the elements on the first subdiagonal of \(B\) has nonpositive real part and zero imaginary part; (iii) the sum of the elements on each of the other subdiagonals of \(B\) is zero. This result is extended to matrices with any number of nonderogatory eigenvalues on the imaginary axis, and to a stability definition based on the spectral radius. A generalized necessary condition, though not a sufficient condition, applies to arbitrary Jordan structures. The proof uses two powerful techniques: The Puiseux-Newton diagram for describing roots of polynomials with analytic coefficients, and the Arnold normal form for perturbation of multiple eigenvalues. All the results have a straightforward extension to the case of reducing the radius of a matrix with multiple eigenvalues. Finally, practical implications are discussed.