Stability of the Lanczos algorithm on matrices with regular spectral distributions (Q6185983)

The authors consider the problem of stability of the Lanczos algorithm, being specifically run, on problems whose eigenvector empirical spectral distribution is close to a reference measure with well-behaved orthogonal polynomials. A backward stability result is presented. This can be upgraded to a forward stability result when the reference measure has a density supported on a single interval with square root behavior at the endpoints. An inference that could be drawn, in particular, is that the Lanczos algorithm run on many large random matrix models, even when computations are carried out in finite precision arithmetic, is actually forward stable, and therefore is nearly deterministic. Since, in general, the Lanczos algorithm is not forward stable, the investigation carried here, implies that care must be taken when using them to test numerical algorithms.