[x] Use the rchk tool to check for possible memory errors/problems before submitting to the CRAN.
[ ] Make NSGEV and potomax rely on nieve for the distribution functions, and discard the implementation of the probability functions therein.
[ ] Make sure that the the inst/computing directory and the four
vignettes GEV.Rnw, GPD2.Rnw, PP2PoisGP.Rnw, PoisGP2PP.Rnw are
not provided to the CRAN since the package may then fail to conform to
CRAN rules. Nearly 1MO of pdf files:( Suitable links should be found in
the package manual. Maybe keep nieve.Rnw as a package vignette with a
few examples?
[ ] Remove any kind of dependence to Renext?
[x] Should we keep the R implementation for the GEV distribution? no.
[x] Inquire about the extRemes package and report it in the table
of the vignette nieve.Rnw. The package does not aim at exporting
probability functions which are rather seen as being for internal use.
[x] Add a .bib file with the citations for the packages and cite the
packages in the vignette.
[ ] In GPD2.tex, recall how can the derivatives of the distribution
function F are related to those of the cumulated hazard H.
[x] In GEV.tex, give the formula for the 2-nd order derivatives of
the distribution function.
[ ] In GEV.tex, the expressions related to the log-likelihood
log f (and especially those for the 2-nd order devrivatives) could be
made much simpler by using the quantities W, T as defined in the
section devoted to the distribution function F.
[x] Implement the Taylor approximation for ξ ≈ 0 in the
PP2poisGPand poisGP2PP (C) functions.
[ ] Add the Extended GPD distributions of Papastathopoulos & Tawn. Since the distribution functions or survival are obtained by composition of functions, the derivatives should come by chain rule.
[ ] Should there be a Gumbel2 distribution?
[ ] Use a global variable for eps in order to share a value across
functions? Or add a new argument for each function?
[x] Add tests for the specific case ξ ≈ 0 (more precisely,
ξ < ϵ) in the transformations PP2poisGPand poisGP2PP.
[ ] Add a test checking that the transformation PP2poisGPand
poisGP2PP are reciprocal to each other (in both directions).
[x] Make the quantile functions work correctly for p = 0.0 and
p = 1.0. The corresponding end-point of the distribution (be it finite
or infinite) must be returned.
[x] (Except for exp1) Make sure that the distribution functions work
correctly with q = -Inf and q = Inf. The value 0 or 1 must be
returned.
[ ] Add the formal argument log.p to the distribution functions
pGPD2 and pGEV to better conform to what is done for distribution
functions in the stats package.
[ ] Add the formal argument log.p to the quantile functions qGPD2
and qGEV to better conform to what is done for distribution functions.
This of course implies changing the derivative and the Hessian.
[x] Make sure that the density functions are zero (not NA) outside
of the support.
[x] Make sure that the distribution functions are OK outside or the support: zero or one.
[ ] Make sure that the probability functions return NA when an
incorrect parameter is used (negative scale).
[ ] Test the functions related to the Exp1distribution.
[ ] When checking a Hessian , one can get a “false negative” ressult
because the numerical Hessian is wrong. Typically, it contains some
NA. Find something to do then: change the options numDeriv::hessian,
discard the test, …
[x] Add GitHub action to check on the three main platforms.
[ ] Add Github action to produce precompiled files as artifacts?
[ ] Add code coverage in a github action.