                 Changes in version 1.9.25 (2026-08-22)                 

    o   add PROTECT/UNPROTECT to protect against false-positive rchk warning #418.

    o   Finalize Eigen 5 compatibility.

    o   Some internal TMB namespaces have been cleaned up.

                 Changes in version 1.9.22 (2026-07-21)                 

    o   Fix clang-14 -std=c++20 RTMB build.

    o   Add partial Eigen 5.0.0 compatibility.

    o   Add Conway-Maxwell-Binomial distribution (thanks to Julius Bogomolovas).

    o   runSymbolicAnalysis() now supports an 'incomplete' Cholesky factorization.

                 Changes in version 1.9.21 (2026-03-23)                 

    o   Package restructure:
	- Github version == CRAN version [no more patching for CRAN]
	- 'Eigen' + 'unsupported' removed [we now completely rely on RcppEigen]
	- All examples wrapped in 'dontrun' [we no longer ship the 'simple' build]

                 Changes in version 1.9.20 (2026-03-17)                 

    o   Avoid non-API 'Rf_findVar'.

                 Changes in version 1.9.19 (2025-12-15)                 

    o   This release adds changes related to the 'TMBad' autodiff component:
	- Sparse plus low-rank Hessian now has an 'LDL' option to handle cases
	where the sparse contribution is not positive definite.
	- Tape optimizer can now recognize identical checkpoint operators as
	well as identical references to outer contexts.
	- Adaptive dynamically retaping checkpoint operators can now safely
	share their derivative table.
	- A bug in the 'expm1' operator has been fixed.

                 Changes in version 1.9.18 (2025-10-13)                 

    o   This release adds changes related to the 'TMBad' autodiff component:
	- Very slow `sdreport()` has been fixed.
	- Flag `TMB::config(tmbad.atomic_sparse_log_determinant=TRUE)` is no
	  longer ignored by `GMRF_t`.
	- `pow()` operator is now optimized to handle constant operand special
	  cases.
	- The experimental Hessian compression algorithm has been improved
	  and can now fix many more cases of memory explosion caused by the
	  sparse Hessian tape.
	  (Enabled using `TMB::config(tmbad.sparse_hessian_compress=TRUE)`)

                 Changes in version 1.9.17 (2025-03-10)                 

    o   splinefun fixes:
	Now handles parameter dependent x values and knots correctly.

    o   New atomic functions: findInterval, subset, order, sort, fmod.

    o   Replace non-API 'ENCLOS' by 'R_ParentEnv' (GH #405).

    o   Fix CRAN ASAN issues reported for the 'tmbstan' package.

    o   NA integer handling has been postponed as it broke parallelization.

                 Changes in version 1.9.16 (2025-01-08)                 

    o   Improved behavior for sdreport() when using 'bias.correct=TRUE'
	without any random effects (GH #403).

    o   dnbinom() impoved accuracy (GH #402).

    o   Handle integer NAs correctly for DATA_I* macros (GH #400).

                 Changes in version 1.9.15 (2024-09-09)                 

    o   Add missing std::numeric_limits for 'TMBad' #398 (fixing undefined
	behaviour in some cases)

    o   Enable bitwise reproducible results for 'TMBad' by default

                 Changes in version 1.9.14 (2024-07-03)                 

    o   Fix C++20 compiler warning

                 Changes in version 1.9.13 (2024-06-28)                 

    o   Fix isNullPointer #395

                 Changes in version 1.9.12 (2024-06-19)                 

    o   Fix clang-19 build issue

    o   use Matrix API for SEXP->(cholmod_factor *) coercion (#393)

    o   Add TMB.Version() #387

    o   Fix protection bug in REPORT(SparseMatrix)

                 Changes in version 1.9.11 (2024-04-03)                 

    o   Fix conflict between TMB and Eigen parallelization (GH #390).

    o   Matrix ABI check no longer writes external file during installation
	(GH #392).

    o   Improved diagnostic messages when data of invalid storage mode is
	passed to TMB (GH #391).

                 Changes in version 1.9.10 (2023-12-12)                 

    o   precompile(): Translation unit changes that should not be visible to
	the user.

                 Changes in version 1.9.9 (2023-11-28)                  

    o   Fix 'Wformat' warnings

                 Changes in version 1.9.7 (2023-11-21)                  

    o   Fixed tweedie integer overflow for extreme parameters.

    o   Added new argument 'adaptive' to 'tmbprofile'.

    o   Added new atomic matrix functions 'sqrtm' and 'absm'.

    o   Avoid unnecessary warnings on 'Matrix package version inconsistency'.

                 Changes in version 1.9.6 (2023-08-11)                  

    o   Fix UBSAN observed for nested AD contexts.

                 Changes in version 1.9.5 (2023-07-18)                  

    o   R: Prepare determinant() change from Matrix version 1.6-0.

    o   R: oneStepPredict() now allows mixed discrete/continuous distributions.

    o   C++: New atomic 'fft'.

                 Changes in version 1.9.4 (2023-04-18)                  

    o   Fix internal bug causing wrong 'sparse_matrix_exponential' derivatives.

    o   Prepare for upcoming Matrix_1.5-5 (Fix misuse of Matrix C-API).

                 Changes in version 1.9.3 (2023-03-28)                  

    o   Prepare for Matrix 1.5-4 deprecations.

    o   checkConsistency(): New argument 'estimate' enables full simulation
	study and another argument 'observation.name' can be used to not
	retape between simulation replicates.

    o   oneStepPredict(): Eliminate nuisance warnings from generic method.

    o   C++ side: Added preprocessor flag 'TMB_ABORT' to better control
	behaviour in 'TMB_SAFEBOUNDS' mode (e.g. signal normal R error
	rather than abort). In addition, 'TMB_CATCH' can now be configured
	to catch more than just std::bad_alloc.

                 Changes in version 1.9.2 (2023-01-23)                  

    o   Improved interoperability with Rcpp (conflicts resolved).

    o   Fix some broken sparse matrix coercions that used to work.

    o   New features on C++ side:
	- Atomic sparse matrix exponential.
	- Atomic vectorized operations mainly for internal use.
	- Faster matrix multiply for higher order AD (see example 'hmm_filter').
	- Utility for parameter object concatenation.

    o   Fix warnings from CRAN check page.

                 Changes in version 1.9.1 (2022-08-16)                  

    o   Prepare for Matrix 1.4-2 deprecations.

    o   Fix html version of sdreport docs (thanks to Kurt Hornik).

    o   Constant input test was missing for some special functions.

    o   Added argument 'max.order' to compile().

    o   oneStepPredict() CDF method was stabilized.

    o   2D interpolation operator was added on C++ side.

                 Changes in version 1.9.0 (2022-05-27)                  

    o   Fix PROTECT bugs (thanks to Tomas Kalibera)

    o   Behavior if TMB parallization has been changed:
	- Setting the number of threads of a TMB model using TMB::openmp() is now portable across platforms.
	- Number of TMB threads no longer interferes with number of threads used by other packages or BLAS/LAPACK.
	- By default, unless explicitly requested, models run with a single thread.

    o   precompile() can now auto generate 'TMB.h' and 'TMB.cpp' that may be used by R packages to split in several compilation units.

    o   MakeADFun() now checks that the DLL is loaded.

                 Changes in version 1.8.1 (2022-03-23)                  

    o   Fix some clang-14 compiler warnings.

                 Changes in version 1.8.0 (2022-03-07)                  

    o   Experimental support for new AD library 'TMBad' - see ?compile.

    o   Option to discard nuisance warnings from Eigen - see ?compile.

                 Changes in version 1.7.22 (2021-09-28)                 

    o   Fortran character strings: use 'FCONE' macro

    o   Add .onUnload() (GH #353)

    o   Fix tmbutils::array ASAN issue triggered by at least clang-13

                 Changes in version 1.7.21 (2021-09-06)                 

    o   Fix CRAN build issue triggered by clang 13.0.0 (omp.h)

    o   Fix tweedie thread safety on Windows (https://github.com/glmmTMB/glmmTMB/issues/714)

    o   Fix as.list(.) for class 'sdreport' (https://github.com/glmmTMB/glmmTMB/issues/692)

                 Changes in version 1.7.20 (2021-04-08)                 

    o   Make parList() work in more edge cases.

    o   Make tiny_ad variables work (again) with Eigen types.

    o   Avoid header issues in Matrix 1.3-0 due to SuiteSparse update (#340)

                 Changes in version 1.7.19 (2021-02-05)                 

    o   'sdreport()' failed in a special case (GH #333)

    o   Improved implementation of 'rtweedie()'

    o   Fixed thread safety issues (GH #330)

    o   Fixed 'tmbroot()' and 'tmbprofile()' for single-parameter models

    o   Function 'compile()' now quotes more paths (GH #323)

                 Changes in version 1.7.18 (2020-07-27)                 

    o   Fix clang-ASAN issue caused by memory management changes

                 Changes in version 1.7.17 (2020-07-23)                 

    o   New function 'FreeADFun()' gives the user more control with memory
	deallocation. In addition memory management of external pointers
	have been improved - see '?FreeADFun'.

    o   Documentation errors in '?sdreport' have been corrected.

    o   The function 'oneStepPredict()' now calculates residuals in reverse
	order because it is more stable. Old behaviour can be obtained by
	passing 'reverse=FALSE'.

                 Changes in version 1.7.16 (2020-01-15)                 

    o   Fixed memory leaks in sparse hessian.

    o   logspace_add: Added some edge case reductions.

    o   'oneStepPredict' no longer overwrites mapping of random effects.

    o   Fix isfinite, isinf, isnan etc (github #297).

    o   Add support for other object names to 'gdbsource'.

    o   dbinom: Handle edge cases 0*log(0).

    o   Add segment method to vector_indicator.

    o   data_indicator: Add default CTOR.

    o   splinefun: fix memleak + add vector evaluation.

    o   Saved models from old TMB versions should now be usable in new TMB
	versions without giving the INTEGER(...) error.

    o   checkConsistency bugfix: did not work for mapped random effects.

                 Changes in version 1.7.15 (2018-11-09)                 

    o   Allow report array of any type (similar to vector case)

    o   Solve 'rck' issue from CRAN

    o   Fix bug in separable simulate

                 Changes in version 1.7.14 (2018-06-23)                 

    o   Fix performance bug triggered by R-devel.

    o   Fixed bug in sparsity detection algorithm.

    o   Tweedie atomic function speedup.

    o   Fix save/reload model object. Loaded object no longer overwrites
	fitted parameters when running e.g. obj$report() or obj$simulate().

    o   New function 'tmbroot' computes likelihood profile confidence intervals.

                 Changes in version 1.7.13 (2018-03-22)                 

    o   Redundant warnings from CHOLMOD have been silenced.

    o   ADREPORT now tracks object dimensions. More infomation in
	'?TMB::as.list.sdreport'.

    o   New feature to do process normalization from R. See
	'?TMB::normalize'.

    o   New feature to update data from R without re-taping the
	computational graph. Enabled on C++ side using 'DATA_UPDATE'.

    o   bugfix C++ namespace 'autodiff': Jacobian had wrong dimension.

                 Changes in version 1.7.12 (2017-12-11)                 

    o   Fix CRAN rchk tool warnings (although false alarm)

    o   New macro TMB_OBJECTIVE_PTR for more flexible use of DATA and
	PARAMETER objects.

    o   Fix 'oneStepPredict' when no random effects.

    o   C callable routines 'tmb_forward' and 'tmb_reverse' added.

    o   Fix bug in empty array assignment.

    o   Fix slow 'oneStepPredict' caused by matrix inversion in
	'tmbprofile'.

    o   TMB:::setupRstudio()

    o   compile: Add flag for tracing forward sweeps.

                 Changes in version 1.7.11 (2017-08-09)                 

    o   Allow logical data objects

    o   Fix conversion R <-> C++ of long objects
	- Can now pass long vector (length >= 2^31) as DATA_VECTOR.
	- Can now REPORT a long vector.
	- Can now calculate long jacobian matrices (m*n can be greater than
	  2^31 where n is the number of parameters and m the number of
	  ADREPORTed variables).

    o   Fixed a possible segmentation fault in sparsity detection algorithm.

    o   New flag to sdreport() reduces memory usage when bias correcting a
	large number of variables.

    o   New experimental function checkConsistency() to check the Laplace
	approximation and correctness of the user template.

                 Changes in version 1.7.10 (2017-05-05)                 

    o   Fix broken Solaris build caused by Eigen 3.3.3 update

                 Changes in version 1.7.9 (2017-04-12)                  

    o   Fix PROTECT bugs reported by Tomas Kalibera.

    o   Now possible to use 'autodiff::hessian' in conjunction with atomic
	functions.

    o   Fix 'testthat:::expect_equal(obj, obj)' crash for non-random effect
	models.

    o   Fix compatibility issues of 'tmbutils::vector' with eigen 3.3.2.

                 Changes in version 1.7.8 (2017-02-08)                  

    o   Internal bug fixes.
	- Fix bug in GMRF constructor from grid array.
	- Fix printing of Eigen/AD types (broken by version 1.7.7)

                 Changes in version 1.7.7 (2017-02-01)                  

    o   Internal modifications:
	- Fix Eigen 3.3.2 compatibility issues.
	- Disable dynamic symbols.

                 Changes in version 1.7.6 (2017-01-15)                  

    o   Add Conway-Maxwell-Poisson distribution (dcompois, rcompois).

    o   Add lfactorial.

    o   Move MCMC to separate package.

    o   Add simulators for exponential, beta, F, logistic, Student T, and
	Weibull distributions.

    o   Fix a bug in function 'split'.

                 Changes in version 1.7.5 (2016-11-21)                  

    o   Atomic logspace_add and logspace_sub.

    o   Numerically robust binomial and negative binomial density functions.

    o   New macros:
	- DATA_STRING to pass strings from R.
	- SIMULATE to mark blocks of simulation code.

    o   Simulation methods implemented for density classes:
	- MVNORM, AR1, SCALE, VECSCALE, SEPARABLE, GMRF, ARk.

    o   R-style simulators now available from the template:
	- rnorm, runif, rpois, rbinom, rgamma, rnbinom and rnbinom2.

                 Changes in version 1.7.4 (2016-09-21)                  

    o   Add onLoad check on 'Matrix' package version.

    o   runExample: Fixed bug when running all examples from a specified
	example folder.

    o   sdreport: Fixed bug in the case with exactly one random effect.

    o   sdreport: Reduced size of output.

                        Changes in version 1.7.3                        

    o   New atomic functions:
	- pbeta
	- qbeta
	- dtweedie
	- besselK (derivatives wrt both arguments)
	- besselI (derivatives wrt both arguments)
	- besselJ (derivatives wrt both arguments)
	- besselY (derivatives wrt both arguments)

    o   Adaptive numerical integration:
	- gauss_kronrod::integrate     (one dimension)
	- gauss_kronrod::mvIntegrate   (multiple dimensions)

    o   sdreport: Fix bug that caused bias.correction to break when
	parameter maps were in use.

                 Changes in version 1.7.2 (2016-08-09)                  

    o   sdreport: Allow bias correct in chunks to reduce memory.

    o   Experimental TMB:::install.contrib to install user contributed cpp
	code from github.

    o   Disable CHOLMOD warnings from inner problem when silent=TRUE.

    o   sdreport: Keep running when solve(hessian) fails.

    o   sdreport: Fix bug that caused summary(sdreport(.)) to fail when
	nothing was ADREPORTed.

                        Changes in version 1.7.1                        

    o   Lots of minor performance optimizations.

    o   sdreport: New argument getReportCovariance.
	- Reduces memory usage when many variables are ADREPORTed.

    o   bugfix: numeric(0) in random list of MakeADFun caused wrong
	parameter list.

    o   New atomic: convol2d

                 Changes in version 1.7.0 (2016-03-25)                  

    o   CITATION update with JSS publication.

    o   New function 'as.list.sdreport' converts from sdreport format to
	original parameter list structure.

    o   MCMC:
	- Fix small bug in transforming intial values in NUTS.
	- Rename mcmc -> run_mcmc to avoid conflict with coda package.

                 Changes in version 1.6.6 (2016-01-30)                  

    o   oneStepPredict: Add one-step mean to output for methods
	'oneStepGaussianOffMode' and 'oneStepGeneric'.

    o   Array class changes:
	- More explicit array assignment from other class:
	Can now assign from matrix and un-evaluated expression template
	E.g. a = a.matrix() * a.matrix();
	- Documentation: warning about the 2D case
	- Re-implement vec() operator (avoid loop)
	- Add method matrix() to array class: Keeps first dimension and
	  collapses remaining dimensions.

    o   Clean up potential naming conflicts caused by Rmath macros

    o   Atomic functions general speedup

    o   bugfix: precompile broke REGISTER_ATOMIC

                 Changes in version 1.6.5 (2015-12-03)                  

    o   Fix bug that caused sdreport to fail for models using the 'profile'
	argument to MakeADFun.

    o   Robustify marginal likelihood for extreme parameters (to help MCMC).

    o   MCMC: Improved stability of adaptive step size algorithm.

    o   Bias correction documentation and test example. Results are now part
	of the summary output.

                        Changes in version 1.6.4                        

    o   precompile() makeover:
	- Works on all platforms.
	- Works with multiple models in same R instance.
	- Frequently used classes MVNORM_t and GMRF_t added to precompilation.

                        Changes in version 1.6.3                        

    o   Fix array rows() method (The method inherited from Eigen would
	report the length of the underlying vector).

    o   Eliminate std::cout in CRAN version.

    o   MCMC samplers (HMC, NUTS and RWM) added by Cole Monnahan.

                        Changes in version 1.6.0                        

    o   Major changeset to get closer to CRAN submission
	- Added 'dontrun' to most examples. Reason: All useful TMB examples
	take too long time to run because they require compilation.
	- Removed TMB-internals.
	- Added Roxygen for remaining functions: dynlib()
	runSymbolicAnalysis() config() plot.parallelBenchmark()
	summary.sdreport() print.sdreport() print.backtrace()
	plot.tmbprofile() confint.tmbprofile()
	- Got rid of global configuration, e.g flagsDefaults(),
	newtonDefaults() etc.
	- Reworked newtonOption() and adapted all examples to new syntax.
	(NOTE: not entirely backwards compatible since to old version
	modified global settings, which we do not allow anymore).
	- Fixed several typos, e.g. c++ -> C++
	- CITATION: Added arxiv reference.
	- DESCRIPTION: Changed title and added URL.
	- NAMESPACE: Cleanup

    o   CppAD jacobian: prefer reverse mode over forward mode if work is equal
	- TMB only requires 1st order reverse (not forward) mode to be
	implemented for atomic functions. This can result in the error
	'order not implemented' if CppAD automatically selects forward over
	reverse mode.
	- In cases where the work of forward and reverse mode are equal it
	is therefore more natural (from TMB perspective) to select reverse
	mode.

    o   Document parallel_accumulator #122

    o   Lots of doxygen documentation updates

                        Changes in version 1.5-1                        

    o   From now on the github version follows the R version

    o   Added NEWS file

    o   Emacs mode version 3.0 from Arni Magnusson

    o   Robustify tmbprofile() #137

    o   Experimental oneStepPredict() now handles discrete case as part of a
	general 'oneStepGeneric' method.

    o   tmbprofile() multiple changes
	- Allow specification of a valid range of the parameter.
	- Get more detail in the center of the distribution.
	  (does not affect cases where likelihood has been optimized)

    o   C++: New namespace 'autodiff' makes it easy to use AD from the user
	template.

    o   C++: New namespace 'romberg' with wrappers around CppAD integration
	routines.

    o   Automatic retape when loading saved ADFun objects.

    o   Keep attribute 'check.passed' when subsetting parameter list

    o   Fix 32 bit

    o   match pattern: Improvement for large problems. As a result it is now
	be possible to handle 2^31-1 nonzeros in Cholesky factor. Old method
	would break down around 2^30.

    o   MVNORM_t: Method to evaluate marginals.

    o   oneStepPredict() now works with maps #108.

    o   Lots of improvements and cleanups from Martin Maechler.

                        Changes in version 1.5-0                        

    o   Add pnorm/qnorm one-argument vectorized versions.

    o   bugfix: Vectorized math operations crashed in the zero-length case.

    o   Add missing bounds check to array class #100 - Now checks each index
	component (not just the re-mapped univariate index).

    o   BesselK atomic + Matern correlation function.

    o   asSparseMatrix and asSparseVector fix: Conversion from dense to sparse
	should not drop zero-entries that could potentially become nonzero
	(i.e. parameter dependent entries).

    o   CppAD::Variable now works for TMB models.

    o   Script to enable 'runSymbolicAnalysis' on all Linux and OS X systems.

    o   New experimental function 'oneStepPredict' for OSA residuals with
	ar1xar1 as an example.

    o   New atomic function 'ppois'.

    o   REPORT: Allow report vector of 'anything'.

    o   getUserDLL: Improve guess of user's DLL code (when DLL not given to
	MakeADFun).

    o   Eliminate warning with '-Wpedantic'.

    o   sdreport new option: ignore.parm.uncertainty.

    o   New example: HMM filter.

    o   General 1D likelihood profile function + method to get profile
	confidence intervals + plot method.

    o   MakeADFun: New option to move outer parameters to the inner problem
	(purpose: get same speed as the REML trick for linear fixed effects
	without loosing the ML interpretation).

    o   MakeADFun: New argument 'silent' to disable all tracing output.

                        Changes in version 1.4-0                        

    o   CppAD updated to development version. Reduced peak memory usage with
	up to 50% for large models.

    o   New configuration variables.

    o   Avoid crashes due to memory allocation errors.

    o   Print index errors to error stream.

    o   Allow compilation with -std=c++11.

    o   sdreport: bias.correct optimizations.

    o   Implement up to 7d array (like admb).

    o   Allow lists as data object with R-inla SPDE structures as an example.

                        Changes in version 1.0-0                        

    o   Initial release.
