The scipy compatibility layer¶
accudist.compat is a deliberately partial drop-in for scipy.stats. It covers
the evaluation methods where tail precision matters and nothing else.
from accudist.compat import binom, poisson, norm, gamma
binom.logsf(900, 1000, 1/6) # -1312.688 (scipy: -inf)
binom(1000, 1/6).logsf(900) # frozen form
poisson(0.1).logsf(200) # -1331.454
gamma.isf(1e-300, 2.5, scale=2.0)
Supported¶
| methods | |
|---|---|
discrete (binom, poisson, nbinom, geom, hypergeom) |
pmf logpmf cdf logcdf sf logsf ppf isf rvs |
continuous (norm, gamma, beta, chi2, t, f, expon, weibull_min, lognorm, cauchy, logistic, uniform) |
pdf logpdf cdf logcdf sf logsf ppf isf rvs |
Distributions take scipy's parameters, including loc and scale, and map
them onto accudist internally. sf/logsf/isf go through R's direct upper-tail
algorithms instead of 1 - cdf; that is the whole point.
Absent by design: fit, expect, moment, stats, entropy, interval,
median, mean, var, std, nnlf, support. These are not precision
problems. Accessing one raises NotImplementedError naming the R-flat
replacement and the scipy original.
Mapping traps handled¶
| scipy | accudist | note |
|---|---|---|
geom(p) |
dgeom(k - 1, p) |
scipy counts trials (support from 1); R counts failures (from 0) |
hypergeom(M, n, N) |
dhyper(k, m=n, n=M-n, k=N) |
(population, successes, draws) vs (white, black, drawn) |
lognorm(s, loc, scale) |
dlnorm((x-loc)/scale, 0, s) |
scipy's scale is exp(meanlog) |
expon(loc, scale) |
dexp((x-loc)/scale, 1) |
|
gamma(a, loc, scale), chi2(df, loc, scale), t(df, loc, scale), ... |
standardised y = (x-loc)/scale, density divided by scale |
rvs(size=...) draws from accudist.default_rng(); it does not accept
random_state.
Tests¶
- Agreement: where scipy is accurate, every method agrees to
rtol=1e-12for every mapped distribution, frozen and unfrozen. - Improvement: on the gap cases accudist is finite and correct where scipy is
not, and the test also asserts scipy's
-infso a scipy fix shows up. - Absence: unimplemented methods raise with the documented message.