Release pysib v0.2.5 · diegoeck/pysib

v0.2.5 — 2026-10-13

Optimizer robustness release, prepared during the revision of the IFAC Journal
of Systems and Control article.

The Gauss--Newton system of the shared C optimizer is now solved with LAPACK
dgelss (minimum-norm solution through the SVD, singular values below
machine precision treated as zero) instead of a Cholesky factorization
(dposv). Singular or ill-conditioned Gauss--Newton matrices, typical of
high-order models with clustered poles, no longer make every Gauss--Newton
step fail. On an eighth-order benchmark the median relative simulation error
of oe dropped from about 2.5% to 0.2%; estimates for well-conditioned
problems are unchanged.
The Gauss--Newton phase now stops when none of the ten backtracking
candidates decreases the cost, instead of repeating the same failed step
until the 1000-iteration budget is exhausted. Estimates are unchanged; the
phase typically ends after 150--200 iterations.
oe, armax, bj and their filtered variants issue a RuntimeWarning
when a predictor polynomial of the returned model (F for OE, C for
ARMAX, C and F for BJ) has a root on or outside the unit circle; the
warning of the unfiltered estimators suggests the filtered variant.

Added experiments/ifac-jsc/ with the scripts, data caches, and results of
the numerical experiments of the IFAC Journal of Systems and Control article,
which compare pysib with SIPPY and pysid.

Updated package, citation, and user-manual metadata to version 0.2.5.
Described the Gauss--Newton solver, the stopping rules, and the stability
warning in the user manual.