cfsgn() and trig_moments() were
inaccurate for shapes below about 0.6: the Simpson rule on
[0, 37^(1/beta)], capped at two million panels, could not
resolve the long range, and cfsgn(0) was 40624 instead of 1
at beta = 0.1. The characteristic function is now evaluated
on a ray in the upper half-plane by the trapezoid rule in log-radius,
which removes both the oscillation and the long range, for every
beta > 0. Against 120-digit values of the convergent
series at t = 1, 2, 3, lam = 0.3, two scales
(0.4 and sqrt(0.21)) and shapes 0.1 to 3, the relative
error is below 1e-14. At large |t| the two
halves of the characteristic function cancel, so its error there is
absolute, of order tol: about 1e-15 to
3e-15 for shapes up to about 3, a few times
1e-15 for shapes 5 to 20, and up to about
1e-14 for shapes up to 50, at the default tol,
as the help page now says. The tolerance is a new argument
tol (default 1e-12), relative to the size of
each half, and method = "series" sums the convergent
descending (beta < 1) or ascending
(beta > 1) series as a check, refusing (NaN
with a warning) a half whose estimated rounding error exceeds
tol.
dwsgn(), pwsgn() and
qwsgn() truncated the image sum at K = 50
images, which was inaccurate for small shapes (relative error 0.14 at
beta = 0.3, and most of the mass missing below). The
remainder beyond the images summed directly is now added by the
Euler-Maclaurin formula with the exact incomplete-gamma integral.
Against 40-digit image sums at six angles, lam = 0.3, two
scales (0.4 and sqrt(0.21)) and shapes 0.1, 0.2, 0.3, 0.5,
1 and 2, the relative error of dwsgn() is below
1e-14 with the default K = 50; K
is kept as an argument, and the result does not depend on it. The tail
is kept in log space, so shapes far below 0.01 are covered (the density
is correct to 1e-13 at beta = 0.001).
pwsgn(2 * pi) is now one and qwsgn() no longer
warns about missing mass. Short arcs keep their relative accuracy, and
the incomplete gamma functions work from the logarithm of their
argument, so small probabilities and their quantiles are accurate down
to the underflow limit at every shape, also next to the location (for
instance qwsgn(1e-300)).
cfsgn() and trig_moments() validate
tol (one finite number in (0, 1)) and raise values below
1e-14, beyond the rounding of the quadrature, to
1e-14 with a warning. Log-likelihoods at shapes below about
0.6, including those of fit_wsgn_mle(), change
accordingly.
The help pages called the reciprocal-rate subfamily
(sigma = sqrt(lam (1 - lam))) “the identified
parameterisation”. That was wrong: the full four-parameter wrapped
distribution is identifiable, and the subfamily is a three-parameter
restriction of it. The documentation now says so.
The wrapped log-density put an observation within half a rounding
step (about 4e-16) of the location on the wrong side of it,
scoring it as sitting on the location. The signed distance is now
reduced exactly (remainder()), so the side is read from its
sign. This changed only likelihoods in which an observation lies that
close to the fitted location, for instance
profile_wsgn(1 + (-3:3) * 1e-14, 2, beta_floor = 0.1),
which is now 211.601 (was 211.673), and a fit whose location is within
rounding of an observation: the profile at the fit’s own shape now
equals the fit. Fits on ordinary data are unchanged. The Nelder-Mead
search now also stops when its simplex has collapsed to a few units in
the last place in every coordinate, which the corrected likelihood next
to an observation would otherwise leave running to its evaluation cap
(three to four times the run time at shapes below one).
fit_wsgn(): at_one_sided now also
applies to grouped fits (judged from the log-rates, as for exact data),
with the same message and the same refusal of
se = "information", which now names the one-sided case.
at_upper is flagged within a relative 1e-5 of
the upper bound (was 1e-6). ngrid is limited
to 1e8, with an error above it. The help page now says that
the lam standard error of a symmetric fit is
NaN (lam is fixed), and that the 1.25 and 1.6
spike shapes are skipped when they reach half the upper bound (floor
6.25 or more).
A one-sided fit could stop short of the maximum. When the fitted
location sits on an observation at the cusp of a one-sided likelihood,
the final free polishes collapsed along the location before the other
parameters had moved: with
set.seed(101); x <- rwsgn(40, 1, 3, 0.5, sigma = 0.01); th <- signif(1 + abs(x - 1), 15),
fit_wsgn(th) returned 150.39405 at shape 19.68 while
profile_wsgn(th, 20) gives 150.40293. The fit now ends with
one more polish for fits whose log-rates differ by more than
log(1e12): the location is held at its fitted value, the
larger log-rate is pulled back to the smaller plus 45, and the result is
kept only if the log-likelihood is higher by more than
1e-9. A free-shape fit that ends on the upper bound of the
shape is also compared with the search at that bound (the one
profile_wsgn(th, 20) runs) and the higher of the two is
kept, so that the result does not depend on whether the compiler fuses
multiply-adds; nfev of such fits rises. The reproducer now
gives 150.4029337676 at shape 20 (at_upper and
at_one_sided are TRUE) on every platform
tested. Of 26 test sets (20 simulated, the periwinkles at five floors
and the ants), compared to eight decimals, one other fit changed: a
one-sided data set of the same kind rose by 9e-7 (to its profile value
at shape 20); the other 25 did not.
fit_wsgn() and profile_wsgn() run the
independent trials inside a fit in parallel on ncores
threads (profile_wsgn() gains the argument; the default is
1 in both, as CRAN requires): the starting values, the screen of
candidate locations, the spike branches at the smallest shapes and their
free polishes, and, in profile_wsgn(), the shapes. The
package must be built with OpenMP (4.5 or later); otherwise
ncores is accepted and ignored. Each trial is a function of
its own inputs, writes only to its own slot, and the best is chosen
afterwards in index order (ties to the lowest index), so every result,
including the evaluation count, is identical to the last bit for every
number of threads, and the same as the serial fit. On a laptop with four
performance and six efficiency cores, busy with other work, a fit of
1500 observations took 160 s on one thread, 86 s on two, 57 s on four
and 55 s on eight (400 observations: 36, 18, 11 and 13 s); the gain
stops near four threads because the first pass of the location search,
the Nelder-Mead runs and the final polish still run in sequence and the
extra threads run on slower cores. A likelihood evaluation with 128 or
more observations (or cells) is split into chunks of 64 over the threads
and summed in index order, so it stays bit-identical.
ncores above 256 is reduced to 256 with a message, and the
log-gamma function is called in its reentrant form
(lgamma_r) where available, so that concurrent fits share
no state. Inside the parallel bootstrap each refit stays on its own
thread. The Python package has the same argument and a flag
wsgnstat.HAS_OPENMP; its Linux wheels are built with
OpenMP, its macOS wheels are serial (see its README).
dwsgn(), pwsgn(), qwsgn(),
rwsgn() and trig_moments() take the scale
sigma as an optional argument, giving the full
four-parameter wrapped distribution (mu, sigma, lam, beta);
without it they use the reciprocal-rate subfamily as before.
dwsgn() gains log.
fit_wsgn() fits the four-parameter wrapped
distribution by maximum likelihood over
beta_floor <= beta <= 20 (default floor 0.1),
searching both the regular branch, with the location over the
observations when the shape is at most one, and the spike branch at
three shapes, the floor times 1, 1.25 and 1.6, with the location over
every observation, and returning the best; or by the grouped likelihood
for rounded data (resolution). The regular branch also has
one-sided starts, and every location search screens each candidate
observation after a short optimisation of the other parameters (all
candidates when there are at most 200, otherwise a structured subset of
them, as the help page describes); profile_wsgn() uses the
same search, so the profile at the floor equals the fit’s floor search.
Fits on the shape floor, the upper bound and the rate bound are flagged
(at_floor, at_upper,
at_rate_bound), and the exact log-rates are returned
(log_rates), since lam can round to 0 or 1 on
a degenerate fit. With se = "auto", an interior fit with
shape above one gets standard errors from the expected Fisher
information (relative difference steps, a grid concentrated at the
location); otherwise no standard errors are returned, with a message
saying why. A parametric bootstrap, drawn from the fitted rates, runs
only with se = "bootstrap". Degenerate data (fewer than
three distinct angles, or one occupied cell) and non-finite angles are
refused with an error. Every step, the optimiser and the bootstrap
refits included, runs in C; the refits run in parallel when the package
is built with OpenMP. print(), coef(),
vcov() and logLik() methods are
provided.
profile_wsgn() gives the profile log-likelihood in
the shape, grouped_loglik() the grouped log-likelihood of
rounded data, and trig_decay() the constant of the exact
decay rate C p^-(1 + beta) of the trigonometric
moments.
fit_wsgn_trig() and fit_wsgn_mle() are
kept for backward compatibility; they fit the reciprocal-rate
subfamily.
The example of fit_wsgn() now uses 40 observations,
and its multi-thread comparison is in \donttest, to keep
its CPU time well under the CRAN limit.
The title and description now describe the distribution as the wrapped two-piece exponential power distribution, treated as a uniform scale mixture with generalised gamma mixing; the package name follows the authors’ earlier usage, wrapped skew generalised normal.
The vignette is rewritten around the mixture representation, the characteristic function for small and large shapes, the accuracy of the wrapped density, the four-parameter fit with its shape floor, and the grouped likelihood for rounded data.
dsgn(), psgn(), cfsgn(),
rsgn()).dwsgn(), pwsgn(),
qwsgn(), rwsgn()), and its trigonometric
moments (trig_moments()).fit_wsgn_trig()) and by maximum
likelihood with parametric-bootstrap standard errors
(fit_wsgn_mle()).