Changes in version 1.0-7 (2018-01-25)
o pmvt(..., df = ) is scalar only
Changes in version 1.0-6 (2017-03-01)
o use registered C routines
Changes in version 1.0-5 (2016-02-02)
o improvements in quantile estimation
Changes in version 1.0-4 (2016-01-19)
o a new algorithm for quantile estimation, again. Quantiles are
now computed by a stochastic root finding algorithm. Note
that f.quantile in the output of qmv{t,norm} is now (again)
the squared difference between the cdf evaluated at the
quantile and the level. The \code{interval} argument to the
quantile functions is used as a starting value for the
root finder when available.
o clean-ups in C and R code and documentation
Changes in version 1.0-3 (2015-07-21)
o new algorithm for quantile estimation. Quantiles are
now computed by minimising the squared distance between the
distribution function and the probability whereas previous versions
used uniroot(). The procedure is now performed multiple times with
difference random seeds in order to stabilise the results. The
\code{interval} argument to the quantile functions is IGNORED now.
Changes in version 1.0-2 (2014-12-16)
o start providing C interfaces to the underlying algorithms.
mvtnorm_C_mvtdst() directly calls Alan's FORTRAN code and
can be used in other packages via LinkingTo. See
mvtnorm/inst/C_API_Example for an example very much inspired
by the example provided in package xts.
o provide .C interfaces to the FORTRAN routines and allow switching
on/off of R' RNG.
Changes in version 1.0-1 (2014-11-11)
o replace internal MVCHNV FORTRAN FUNCTION with
R's sqrt(qchisq(p, n, FALSE, FALSE). This fixes a
problem where NaN was returned as reported by
David Charles Airey
Changes in version 1.0-0 (2014-07-08)
o After 14 years, we now feel safe enough to publish mvtnorm 1.0-0.
Many packages depend, import, or suggest mvtnorm, so this version
change also indicates that the package is now stable and, to a very
large extent, the API is frozen. We will of course continue to fix
bugs or other problems but new features are unlikely to go into this
package.
o use Authors@R in DESCRIPTION
o switch to standard NEWS format
Changes in version 0.9-99992 (2014-05-03)
o cleanups by MM
Changes in version 0.9-99991 (2014-04-25)
o version 0.9-9999 introduced new bug in dmvnorm
Changes in version 0.9-9999 (2014-04-17)
o dmvnorm (again) returns NaN in case sigma is not decomposable
Changes in version 0.9-9998 (2014-03-21)
o faster code for dmvnorm by Matteo Fasiolo
Changes in version 0.9-9997 (2014-01-17)
o T. Miwa fixed a runtime error in miwa.c reported by UB sanitizer
Changes in version 0.9-9996 (2013-09-16)
o documentation updates/corrections/examples by Marius Hofert
o df = Inf
o rmvt(): argument 'mean' not allowed anymore (prone to misuse)
o pmvnorm(lower=c(-Inf, 0, 0), upper=c(0, Inf, Inf),
mean=c(0, 0, 0), sigma=S, algorithm = Miwa())
returned NaN, fixed by Xuefei Mi
Changes in version 0.9-9995 (2013-05-29)
o update to version 2013-06-29 of mvtdst.f from Alan's website
(fixed a bug for 2-dim pmvt)
Changes in version 0.9-9994 (2012-12-06)
o set.seed(29)
rmvnorm(10, ...)
produces the same first ten rows as
set.seed(29)
rmvnorm(100, ...)
o as suggested by Paul Johnson . There is a new
argument pre0.9_9993 for changing back to the `old' output. This
_DOES NOT_ apply to rmvt.
Changes in version 0.9-9993 (2012-10-22)
o sigma is called `scale' matrix of {dpq}mvt, thanks to
Richard Boys for the hint
Changes in version 0.9-9992 (2012-01-19)
o qmvt(..., df = 0, ...) gave NaN
o R CMD check works on x64 with GCC 4.6.2 (Debian 4.6.2-11)
Changes in version 0.9-9991 (2011-06-10)
o tvpack wasn't working on 64bit machines. The reason
was that the wrapper SUBROUTINEs around the original
FUNCTIONs didn't return the appropriate double precision
for unknown reasons. Redefining TVTL and BVTL as FUNCTIONs
fixed the problem.
Changes in version 0.9-999 (2011-04-26)
o still problems in fix approx_interval (when !is.null(sigma));
disable for the time being
Changes in version 0.9-99 (2011-04-21)
o fix bug in approx_interval spotted by Ravi Varadhan
Changes in version 0.9-98 (2011-04-19)
o allow ... to pass arguments to rmvnorm in rmvt
Changes in version 0.9-97 (2011-01-31)
o use check.attributes = FALSE in isSymmetric calls
(requested by Nick Sabbe )
Changes in version 0.9-96 (2011-01-28)
o use fixed interval when sigma is specified in qmvt
Changes in version 0.9-95 (2010-11-18)
o q{mvt,mvnorm} shall always return a list, not a vector
checks for correlation matrices are less picky now
Changes in version 0.9-94 (2010-11-16)
o allow for two different noncentral t distributions (via type
argument)
o add support for one-dimensional quantiles (requested by Jerry Lewis
)
o documentation fixes for problems spotted by
Jerry Lewis
o interface to Alan's TVPACK algorithms for 2- and 3-d
probabilities by Bjoern Bornkamp added.
Changes in version 0.9-92 (2010-07-06)
o update to new mvtdstpack.f (7/10) by Alan. Fixes
potential bias problem in higher dimension.
Changes in version 0.9-91 (2010-04-13)
o better search interval for uniroot in qmv{t,norm} speeds up
quantile estimation; suggestion by Björn Bornkamp
Changes in version 0.9-9 (2010-01-27)
o document ... in pmvt.Rd
Changes in version 0.9-8 (2009-10-27)
o add citation entry
Changes in version 0.9-7 (2009-05-22)
o make sure `error' is not NA
Changes in version 0.9-6 (2009-03-25)
o update Alan's FORTRAN code
Changes in version 0.9-5 (2009-03-17)
o fix FORTRAN bug spotted by Alex Lenkoski
Changes in version 0.9-3 (2008-12-22)
o update meta data
Changes in version 0.9-2 (2008-07-08)
o be a little more liberal (tol = sqrt(.Machine$double.eps))
when testing for symmetry of covariance matrices
(and make R CMD CHECK monomvn happy again)
Changes in version 0.9-1 (2008-07-02)
o better check for covariance matrices, suggested by
James Rogers
Changes in version 0.9-0 (2008-04-01)
o add support for the multivariate normal distributions in small
dimensions by Miwa's method thanks to Tetsuhisa Miwa and Xuefei Mi;
both have been added as `authors'.
o new argument `algorithm' defaulting to `GenzBretz()' with
`Miwa()' being the alternative. Those two functions are now used
to specify hyper parameters such as `abseps'.
o internal function `mvt' is no longer exported.
Changes in version 0.8-3 (2008-02-19)
o make sure rmvnorm(1, sigma = matrix(0.5, 1, 1)) works
(reported by Kurt Hornik)
Changes in version 0.8-2 (2008-02-10)
o rmvnorm() now issues a warning for non-symmetric sigma and uses
the eigenvalue decomposition as default.
o make gfortran 4.3 happy
Changes in version 0.8-1 (2007-07-24)
o Orion Poplawski spotted a meaningless check in
the regression tests
Changes in version 0.8-0 (2007-07-23)
o upgrade to 7/7 version of MVTDST (includes better support for
dimensions > 100). Thanks to Karen Conneely
for motivating the update and for checking the new version.
o rmvnorm() now can also use a Cholesky decomposition to compute
the root of sigma (thanks to Fabian Scheipl)
Changes in version 0.7-5 (2006-09-15)
o fix problem reported by valgrind
Changes in version 0.7-4 (2006-09-08)
o add long requested `dmvt'
o call RNG functions only one time
o make sure unifrnd is double precision
Changes in version 0.7-3 (2006-08-23)
o make sure pmvnorm(lo=c(-Inf,-Inf), up=c(Inf,Inf), mean=c(0,0) == 0
Changes in version 0.7-2 (2005-08-29)
o make gfortran happy (a warning about unused variable NF remains)
Changes in version 0.7-1 (2004-11-18)
o use #!/bin/sh
Changes in version 0.7-0 (2004-10-14)
o a coding session with Frank produced `qmv{t,norm}'.
try to check if the support specified by `lower' and `upper' is
empty (problem spotted by Peter Thomson )
Alan's fix prevents negative values to be returned.
o some cosmetics
Changes in version 0.6-8 (2004-06-03)
o EXIT statements are not supported by `f2c', Alan added GOTO
statements to `MVCHNC'
Changes in version 0.6-7 (2004-05-27)
o Alan's fix to MVCHNC solves problems with large degree of freedom
Changes in version 0.6-6 (2004-01-22)
o `La.eigen' is deprecated and `eigen' replaces it in R-1.9.0
Changes in version 0.6-5 (2003-11-14)
o check if covariance matrix is pd in rmvnorm (by Fritz Leisch)
Changes in version 0.6-4 (2003-10-06)
o use new base function `cov2cor'
Changes in version 0.6-3 (2003-07-21)
o Alans changes were restricted to N <= 100, now N <= 1000
are possible again
Changes in version 0.6-2 (2003-06-25)
o Alan's recent changes to `mvt.f' make `g77 -pedantic -Wall'
happy
Changes in version 0.6-1 (2003-06-18)
o pmvt(..., df = 0, ...) will return normal probabilities for both
the univariate and multvariate problem
Changes in version 0.6-0 (2003-06-17)
o Fortran code in `mvt.f' updated to recent version by Alan and Frank.
This fixes problems with `pmvt' and large degrees of freedom.
Changes in version 0.5-15 (2003-06-16)
o a note on one-sided probabilities in `pmvt'
correlation matrices in cats example a little bit nicer
Changes in version 0.5-14 (2003-05-06)
o the package owns a vignette based on the paper in RNews 1(2)
Changes in version 0.5-12 (2003-05-08)
o allow df=0 for pmvt
Changes in version 0.5-11 (2003-04-29)
o package npmc trys to use 'mvt' which is internal: export it anyway
Changes in version 0.5-10 (2003-04-23)
o mvtnorm is now in a NAMESPACE
Changes in version 0.5-9 (2003-02-13)
o log argument added to dmvnorm, thanks to
Jerome Asselin
Changes in version 0.5-8 (2003-01-21)
o fixed bugreport PR#2478: sigma for univariate probabilities
Changes in version 0.5-7 (2002-11-27)
o use R's random number generator in the FORTRAN code:
set.seed has now has the desired impact.
Changes in version 0.5-6 (2002-10-07)
o rmvt added
Changes in version 0.5-5 (2002-07-03)
o use .Fortran(..., PACKAGE="mvtnorm")
Changes in version 0.5-4 (2002-04-09)
o correlation matrices for sigma with unequal variances incorrectly
computed, added `sig2corr' for that propose, tol argument removed,
fix by Alan to mvt.f
Changes in version 0.5-2 (2002-03-22)
o Frank added `tol' argument to MVTDST, now in mvtnorm
Changes in version 0.5-1 (2002-01-24)
o pmvt(0,1) works now
Changes in version 0.5-0 (2001-12-10)
o release for R-1.4.0
Changes in version 0.4-4 (2001-12-06)
o bugfix
Changes in version 0.4-3 (2001-12-05)
o the length of lower, upper and mean (delta) is now recycled to the
length of the largest, i.e. it is possible to say
pmvnorm(lower=-Inf, upper=1, mean=rep(1,10), corr=diag(10))
Changes in version 0.4-2 (2001-12-04)
o several typos, man-pages improved
Changes in version 0.4-1 (2001-12-04)
o interface changed: sigma (covariance matrix) can be specified as
well
o {rd}mvnorm added from package e1071 (thanks to Fritz!)