Changelog • logcumulant

logcumulant 0.1.1

Corrected the closed-form Fisher information matrices returned by fisher_closed() for two families, to match the numerical observed information used by default in mle_fit():
- Frechet now uses the correct off-diagonal (cross) term +(1 - gamma)/b. Weibull and Frechet share the same diagonal entries but differ in the sign of the cross term (log-inversion duality); the previous version returned the Weibull cross term for both.
- Log-Logistic now returns diag((pi^2 + 3)/(9 a^2), a^2/(3 b^2)). The previous version used diag(pi^2/(3 a^2), 1/b^2), whose (1,1) entry is the second log-cumulant Var(log X) rather than the shape information, and whose (2,2) entry omitted the factor a^2/3.
Added a regression test (tests/validate_fisher.R) comparing fisher_closed() against the Monte Carlo observed information for all six families.
Note: default inference (T2_bootstrap, size_study, power_study) is unaffected, since the covariance is estimated from the observed information (mle_fit()), not from these closed forms; fisher_closed() is only used as a fallback when the numerical Hessian is not invertible.

CRAN release: 2026-06-12

Initial release.
Three nested Hotelling-type T-squared goodness-of-fit statistics (T2_all, T2_bootstrap) with parametric-bootstrap calibration.
Three diagnostic diagrams (log_cumulant_diagram, kurtosis_diagram, cv_diagram), a combined panel (three_diagrams), a multi-dataset overlay (multi_lc_diagram), and the quick plot_lc interface.
Maximum-likelihood fitting for six reliability families with a C++ core.
Anderson–Darling and Cramer–von Mises tests and AIC via gof_compare_all.
Size and power simulation utilities (size_study, power_study).
Nine bundled reliability datasets (reliability_datasets).