On condition numbers and the distance to the nearest ill-posed problem

Summary

The condition number of a problem measures the sensitivity of the answer to small changes in the input. We call the problem ill-posed if its condition number is infinite. It turns out that for many problems of numerical analysis, there is a simple relationship between the condition number of a problem and the shortest distance from that problem to an ill-posed one: the shortest distance is proportional to the reciprocal of the condition number (or bounded by the reciprocal of the condition number). This is true for matrix inversion, computing eigenvalues and eigenvectors, finding zeros of polynomials, and pole assignment in linear control systems. In this paper we explain this phenomenon by showing that in all these cases, the condition number κ satisfies one or both of the diffrential inequalitiesm·κ²≤∥Dκ∥≤M·κ², where ‖Dκ‖ is the norm of the gradient of κ. The lower bound on ‖Dκ‖ leads to an upper bound 1/mκ(x) on the distance. fromx to the nearest ill-posed problem, and the upper bound on ‖Dκ‖ leads to a lower bound 1/(Mκ(X)) on the distance. The attraction of this approach is that it uses local information (the gradient of a condition number) to answer a global question: how far away is the nearest ill-posed problem? The above differential inequalities also have a simple interpretation: they imply that computing the condition number of a problem is approximately as hard as computing the solution of the problem itself. In addition to deriving many of the best known bounds for matrix inversion, eigendecompositions and polynomial zero finding, we derive new bounds on the distance to the nearest polynomial with multiple zeros and a new perturbation result on pole assignment.