Add invchol.syMatrices, cov2invchol, and
cov2chol.
Avoid swapping data (row <-> column-wise storage) in a number of places.
Cleaned up the class structure and corrected constructors for
syMatrices.
After more than 15 years on R-forge, mvtnorm is now hosted at codeberg. Thank you, WU Vienna, for the long-term support!
More careful checks of inputs.
logLik wasn't able to handle missing, continuous, and
interval-censored observations at once.
lpmvnorm failed when a non-missing invcholmean wasn't a matrix
with the same dimensions as lower and upper.
Minor performance improvements in sldmvnorm.
Speed-up switching from row- to column major order and vice versa.
Add invcholmean argument where possible.
Improve documentation regarding marginal and conditional distributions.
Ship nuweb source file in inst/nuweb
simulate ignored its seed argument.
sldmvnorm returned an incorrect log-likelihood
in the presence of means.
Observations with very small log-probabilities
are now treated the same in lpmvnorm and
slpmvnorm.
This reduced differences observed on machines w/o long doubles considerably; still there were minor diffs in the package vignette which we now prevent from showing up.
Use REFERENCES.bib.
mvnorm gains a invcholmean argument allowing a
parameterisation leading to convex optimisation problems.
Added zero.print = "." argument for printing lower
triangular matrices.
condDist() sometimes failed to return a
scale slot.
Explain how to use condDist() to obtain regression
coefficients for an example involving means.
Miwa(steps = 4097) are allowed in R, the
C code assumed steps <= 4096, now
steps == 4097 is possible. Triggered by
BDR's gcc-UBSAN check on package
OptimalGoldstandardDesign.
Explain and illustrate that omitting data dimensions produces marginal log-likelihoods in Chapter 7.
Add low-level functionality for reduced rank covariances in new Chapter 8.
invcholD and Dchol avoided very small diagonals;
now they also avoid very large diagonals (such that inverting the result
doesn't give too small diagonals).
Fix partial argument matches, reported by KH.
Challenging numeric test of score function at ML estimate failed on Mac M1.
Add logdet() function for computing log-determinants from
lower triangular matrices.
Speed up solve(<ltMatrices>, ...).
Avoid unnecessary copying of large data objects.
-1/2 -> -1 for det(C) in lmvnorm_src vignette; spotted by Kurt Hornik
Use M\_PI.
Update the Using mvtnorm package vignette and references therein.
Speed up Mult(<ltMatrices>, transpose = TRUE).
Speed up ldmvnorm().
Add constructor syMatrices() for multiple symmetric
matrices and as.syMatrices() for coercion from ltMatrices.
Fix segmentation fault or unnecessary error or warning + approximation for
algorithm = Miwa() or TVPACK(), in case dimension
reduction to one-dimensional is possible, e.g., for
pmvnorm(lower = rep(-Inf,3), upper = c(-1, Inf, Inf),
sigma = diag(3), algorithm = Miwa())
# or
pmvnorm(lower = c(-Inf,-Inf),
upper = c(- 1, Inf), sigma=diag(2), algorithm = TVPACK())
Remove empty print statement from miwa.c
Allow to change rnorm in rmvnorm, feature request
by Ralf Stubner.
Fix variable declarations in tvpack.f as reported by Intel
compilers icx/ipcx/ifx from oneAPI 2023.2.0 and oneMKL 2023.2.0, thanks to BDR
Fix overflow problem reported by ASAN.
Be even more careful inverting / computing Cholesky factors for hessians (M1 and macos-arm64).
Avoid attempts to allocate zero length arrays for one-dimensional problems.
Catch M1mac problems with inverting hessians.
New lpmvnorm() function for computing multivariate normal log-likelihoods for interval-censored observations.
New slpmvnorm() function for computing the corresponding score function for interval-censored observations.
New ldmvnorm() function for computing multivariate normal log-likelihoods for exact observations.
New sldmvnorm() function for computing the corresponding score function for exact observations.
New ldpmvnorm() function for computing multivariate normal log-likelihoods for a mix of exact and interval-censored observations.
New sldpmvnorm() function for computing the corresponding score function for a mix of exact and interval-censored observations.
New class ltMatrices representing multiple lower triangular matrices, with some useful methods.
New package vignette lmvnorm_src describing these new features.
pmvnorm(), qmvnorm(), pmvt(), and
qmvt() gain a seed argument.
Regression tests were improved.
Internal standardization is described better in the docs, suggested by Chris Wymant.
New NEWS file, the old file containing information up to version 1.1-3 is available as NEWS.old.