NEWS
stpphawkes 0.2.4.9000
Correctness fixes
- The compensator in the temporal Hawkes log-likelihood was short by one
alpha
on every call (the truncated tail contributed min_i terms where min_i + 1
were skipped), which biased the estimates of alpha. The corrected version now
matches an untruncated reference implementation exactly.
homog.STPP()'s C++ twin drew the background from rpois(mu * t_area), ignoring
the area of the region the points are laid down on. It now uses
rpois(mu * s_area * t_area), matching the R implementation.simulate_hawkes_stpp() never clipped its output to the study polygon, so it
returned the background points generated on the enlarged region used to avoid
edge effects.mcmc_temporal(..., t_mis =, branching = TRUE) did not seed its generator from
R's RNG, so it ignored set.seed().- The inverse-gamma prior term in
sig_posterior() used -sig/scale instead of
-scale/sig. The Metropolis update now targets the same posterior as the
conjugate sample_sig_gibbs().
mcmc_stpp_nonunif() passed t_max positionally into stpp.mle.nonunif()'s
poly argument, so a user-supplied t_max was silently replaced by max(data$t).- In the branching missing-data sampler the initial missing times were drawn
against an empty history instead of the observed times.
- The offspring burn-in window used
-b * log(fraction) instead of
-log(fraction) / b, so it scaled the wrong way with b.
- A stray mark code in the categorical-mark sampler hit a bare
throw;, which
terminated the R session; it now raises an R condition.
- The GSL generator used for the Dirichlet draws was never seeded, so those draws
ignored
set.seed(); it is now seeded from R's RNG and freed on every exit path.
sample_alpha()'s rejection loop could run forever; it now falls back to an exact
inverse-CDF draw from the truncated gamma after a bounded number of attempts.pip() is exported (its @export tag was indented and had been folded into
@return).predict_hawkes_t() squared beta by pre-multiplying alpha, which
simulate_temporal() already does internally.intensity_stpp() treated sig as a standard deviation; the package convention
is that it is the variance. Same fix in condInt_mcmc_stpp.cpp.inhomog.PPP() called the temporal homog.PPP() and then indexed the result as a
matrix; it now calls homog.SPPP().simulate_hawkes_stpp_inhom() checked ndims(params$mu) rather than ndims(mu),
and dropped the whole background catalog from its output.- Reading
.Random.seed before any RNG draw no longer errors in a fresh session.
DFtoMat() skips non-numeric columns, so a homog.STPP() data frame can be passed
as history; histories with a different number of columns than the background are
lined up instead of failing in join_cols.- Guarded several loops and index-0 writes against empty inputs.
Packaging
- Fixed the malformed
Date field in DESCRIPTION.
src/Makevars.win uses RcppGSL:::CFlags()/LdFlags() instead of the obsolete
LIB_GSL variable, which is unset on current Rtools..Rbuildignore patterns corrected; src/celero/ and src/tests/ no longer ship
in the tarball.NAMESPACE regenerated in roxygen's own format.
stpphawkes 0.2.4 (2026-09-18)
stpphawkes 0.2.2 (2025-03-10)
- bugfixes
- add additional outputs to mcmc to include missing data and branching structure
stpphawkes 0.2.1 (2023-08-15)
stpphawkes 0.2.0