NEWS

---

RTMB 1.9 (2026-03-20)

• Special functions have been added: dchisq(), pchisq(), qchisq(), pnorm() now supports all arguments (i.e. 'log.p' and 'lower.tail').
• Special functions now pass argument attributes on to the result.
• New autodiff features: - AD adaptive integration via integrate(). - AD root finding via uniroot(). - Fast AD version of Vectorize(). - Atomic sparse matrix times vector available via TapeConfig(). - Experimental atomic branching via ADoverload('if'). - AD order() now works.
• Change default to TapeConfig(matmul='plain') because 'atomic' could slow down too many examples.
• New 'tips' vignette.

RTMB 1.8 (2025-10-14)

• Special functions have been added: pnbinom(), lbeta(), dcauchy()
• Special functions have been improved: Add 'expon.scaled' argument to Bessel functions. dmvnorm() now allows zero-dim covariance. Fixed dgamma() rate argument.
• Some discrete methods now work with autodiff: Modulus ('%%'), sort(), findInterval() and taped subsetting by an AD index using '['. Math group: Added floor, ceiling, trunc.
• Improved splinefun(): Spline derivatives have been generalized to work for *all* parts of the spline including evaluation pooints and knots. In addition, the spline object now has a 'deriv' argument (like stats).
• Methods that now work for 'adsparse' class: band(), tril(), triu(), kronecker().
• Sparse matrix exponential expAv() now applies internal rescaling to gain robustness. In addition, caching can now improve memory efficiency when the same generator is used multiple times.
• New methods for the 'Tape' class: reorder(), timer(), force.update().
• MakeADFun: Model object now has a force.update() method for better interoperability with the DataEval() function. An example is given on the '?DataEval' help page.
• RTMB can now be built with preprocessor flag 'EIGEN_USE_BLAS' enabled.
• ADjoint() no longer crashes by errors thrown by user's forward or reverse mode code.