AdMit

• The value returned by 'G' in 'AdMitIS' is validated: wrong length, wrong type and NA/NaN now give an informative error instead of a cryptic one from the matrix algebra.
• 'dMit' and 'rMit' no longer describe different mixtures when the probabilities do not sum to one: 'dMit' used them as given while 'rMit' normalised them through 'sample'. They are now normalised once, with a warning; a mixture that already sums to one is left untouched.
• Mixture components with a non-finite, non-symmetric or non-positive definite scale matrix are rejected instead of producing infinite densities and degenerate draws.
• The acceptance rate reported by 'AdMitMH' is now the number of accepted proposals divided by N-1. The chain makes N-1 proposals, the first draw being its initial state, so a candidate equal to the target reported (N-1)/N instead of one.
• 'fn.CV' returns zero for a constant weight vector instead of stopping. A candidate that reproduces the target exactly is the ideal case, not an error, and the adaptive loop now stops there rather than dividing by it.
• 'fnKernelMixtureArch_C', the kernel of the package demo, combines its two log-densities around the larger of the two. Forming the exponentials first returned -Inf for large observations or small conditional variances.
• 'N' is rejected above .Machine$integer.max, where the coercion in '.C' would have produced NA. 'AdMitIS' requires N >= 2, as its NSE and RNE are not defined for a single draw. 'mu0' must be finite, and 'ISscale' must be strictly positive, as the documentation already stated.
• Documentation: the degrees-of-freedom domain of 'mit$df' is stated as a positive real number, matching the implementation; the acceptance rate and the RNE = NA convention are documented.
• Declared 'testthat (>= 3.1.5)', the version that introduced 'expect_no_warning' and 'expect_no_error', which the test suite uses.

This release does change results, unlike 2.1.10 and 2.1.11. The acceptance

rate of 'AdMitMH' is larger by a factor N/(N-1) by design. Centring the

log-kernel perturbs the fitted mixture at the level of 1e-8 relative, and

the draws and estimates that follow from it accordingly. Nothing else moved:

the demo reproduces its mode, coefficient of variation path and posterior

mean unchanged.