A validated enclosure of Robertson's stiff problem on [0,4000] and the true error of a Rust port of CVODE against it
We compute a validated enclosure of Robertson's stiff kinetics problem at ten output times from t = 0.1 to t = 4000, with an a-posteriori logarithmic-norm integrator in 160-bit Arb ball arithmetic. The enclosure is rigorous under stated assumptions (Arb correct; the step theorem is proved in prose, not formalised): the required relative width of 1e-6 is met everywhere, the reported widths are at most 6e-10. It was independently re-derived by an auditing AI agent's different method (MPFI Lohner) at t = 1, 10, 40 only. Against the enclosure we measure the true global error of the rusty-SUNDIALS cvode crate, a Rust port and not C CVODE, for rtol 1e-4 to 1e-10. For y2 the apparent absence of tolerance violations is only because atol dominates: relative to rtol*|y2| the error is up to 169 times larger. No claim for general stiff problems or t ≥ 1e4. Producer and auditor are AI agents; no human audit; no Lean.
Authors
- Xavier Callens
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24216836
- Primary Topic
- Numerical Methods and Algorithms
- Type
- preprint