Paige-style finite precision stability analysis for block Lanczos

Establish a Paige-style finite precision stability result for the block Lanczos algorithm by proving that a perturbed version of the exact block three-term recurrence AQ_k = Q_k T_k + E_{k-1} Q_k e_1^T holds when the algorithm is executed in floating point arithmetic. Derive explicit bounds on the perturbation and related quantities, analogous to the classical stability guarantees known for standard (single-vector) Lanczos.

Background

The monograph highlights that, unlike the well-understood finite precision behavior of the standard Lanczos algorithm, the block Lanczos method lacks a Paige-style analysis guaranteeing perturbed recurrence relations and quantifying stability. This gap is notable given widespread practical use of block methods for matrix functions.

Providing such a result would unify theory between single-vector and block variants, offer rigorous guarantees for block Krylov subspace computations, and help guide robust implementations that handle deflation, rank deficiencies, and loss of orthogonality in finite precision arithmetic.

References

As far as we are aware, there is no Paige style analysis (see Section 4.1) which guarantees that a perturbed version of (9.3) holds in finite precision arithmetic.

The Lanczos algorithm for matrix functions: a handbook for scientists (2410.11090 - Chen, 14 Oct 2024) in Section 9.1.1