Work

Certifying Degenerate Spectral Subspaces in Stateful Scientific Computation.

Year: 2026
Type: preprint
Abstract: A small eigenpair residual does not certify recovery of a complete spectral subspace when the requested cutoff intersects a multiple eigenvalue. We present a finite qualification procedure combining an independently established spectral count and excluded-spectrum separation with exact arithmetic on saved binary Ritz data. A residual-and-Gram estimate bounds the Frobenius error of the saved low-rank projector without forming a dense projector. An archived recursive-graph case study contains 13 source matrices of orders 324, 972, 2916 and 8748, 22 prescribed payloads, two invocation histories per payload and 19 negative controls. The complete target contains 25 nonconstant dimensions, including a threefold boundary eigenspace. The largest recorded error bound is below 1.034 × 10⁻⁹, against the predeclared 10⁻⁶ ceiling. During manuscript preparation all 66 saved frames were re-evaluated with the supplied exact checker and reproduced their original rational certificates. This replay checks saved outputs; it is not a new eigensolver run or an independent external implementation. The contribution is a reproducible integration of established subspace estimates, exact graph spectra and explicit execution-history coverage, with a precise boundary between finite qualification and unrestricted reproducibility claims.
Language: English
Open Access status: green
OpenAlex ID: W7235575021
Cites:
Cited by:
Related to: 10