R to accudist¶
Naming rules¶
- Distribution functions use R's user-level names:
ppois,dbinom,qnorm,rgamma,ptukey,pwilcox,psignrank. - Special functions use Rmath's C names:
gammafn,lgammafn,digamma,trigamma,psigamma,beta,lbeta,choose,lchoose,bessel_i/j/k/y. R calls thesegamma,lgamma,besselI, ...; the C names avoid shadowingmath.gammaand sitting confusingly next todgamma. - Parameter names mirror R exactly, with one exception forced by Python:
lambdabecomeslambda_. log=is the only spelling of the log-scale flag, ford,p,qalike. R useslog=for densities andlog.p=forp/q.log_p=raisesTypeError.out=is keyword-only;lower_tailandlogmay be positional, as in R.
Signature shape¶
d<dist>(x, <params...>, log=False, *, out=None)
p<dist>(q, <params...>, lower_tail=True, log=False, *, out=None)
q<dist>(p, <params...>, lower_tail=True, log=False, *, out=None)
r<dist>(n, <params...>)
Quantiles of discrete distributions are returned as doubles (ad.qpois(0.5, 4.0)
is 4.0), and r* return float64 arrays, exactly as R does.
Dispatch¶
ncp=None selects the central algorithm¶
R decides between the central and non-central code paths by missing(ncp), not by
its value: pchisq(q, df, ncp = 0) calls pnchisq. accudist reproduces this with
None as the sentinel.
ad.pchisq(3, df=5) # C pchisq (central)
ad.pchisq(3, df=5, ncp=0.0) # C pnchisq (non-central with ncp 0; may differ in the last digits)
ad.pchisq(3, df=5, ncp=1.5) # C pnchisq
Applies to beta, chisq, f, t for d, p, q, r. Only rchisq has a
C non-central sampler; rbeta, rf, rt with ncp are composed from rchisq
and rnorm in the same order R composes them.
prob xor mu (negative binomial)¶
ad.pnbinom(q, size, prob=0.3) # C pnbinom
ad.pnbinom(q, size, mu=7.0) # C pnbinom_mu
ad.pnbinom(q, size, 0.3, 7.0) # TypeError: 'prob' and 'mu' both specified
ad.pnbinom(q, size) # TypeError: requires one of 'prob' or 'mu'
rate xor scale (gamma)¶
R warns if both are given but consistent (rate * scale within 1e-15 of 1) and
errors otherwise; accudist does the same (AccudistWarning / TypeError).
ad.pgamma(q, shape, rate=2.0) # scale = 0.5
ad.pgamma(q, shape, scale=0.5) # same
ad.pgamma(q, shape, rate=2.0, scale=0.5) # warning
ad.pgamma(q, shape, rate=2.0, scale=9.0) # TypeError
Exponential: rate¶
The C library takes a scale; accudist passes 1 / rate exactly as R's wrapper does.
Argument-order hazards handled for you¶
ptukey/qtukey: R presents(q, nmeans, df, nranges)but the C symbol takes(q, nranges, nmeans, df). accudist keeps R's order.hyper:phyper(q, m, n, k)wherenis the number of black balls, so the draw count ofrhyperis callednn:rhyper(nn, m, n, k). Same forrwilcox(nn, m, n)andrsignrank(nn, n).bessel_i/bessel_k:expon_scaled=Truemaps to the Cexpo = 2.
Translation table¶
| R | accudist |
|---|---|
ppois(q, lambda, lower.tail=FALSE, log.p=TRUE) |
ad.ppois(q, lambda_, lower_tail=False, log=True) |
dbinom(x, size, prob, log=TRUE) |
ad.dbinom(x, size, prob, log=True) |
qt(p, df, ncp) |
ad.qt(p, df, ncp=ncp) |
pgamma(q, shape, rate=2) |
ad.pgamma(q, shape, rate=2.0) |
dexp(x, rate) |
ad.dexp(x, rate) |
ptukey(q, nmeans, df, nranges) |
ad.ptukey(q, nmeans, df, nranges) |
pwilcox(q, m, n) |
ad.pwilcox(q, m, n) |
gamma(x) / lgamma(x) |
ad.gammafn(x) / ad.lgammafn(x) |
besselI(x, nu, expon.scaled=TRUE) |
ad.bessel_i(x, nu, expon_scaled=True) |
psigamma(x, deriv=2) |
ad.psigamma(x, deriv=2) |
signif(x, d) / round(x, d) |
ad.fprec(x, d) / ad.fround(x, d) |
set.seed(s) under RNGkind("Marsaglia-Multicarry") |
ad.set_r_seed(s) |
Escape hatches¶
| namespace | contract |
|---|---|
accudist |
the supported API: defaults, dispatch, errstate, docstrings |
accudist.rmath |
1:1 with Rmath.h: C argument order, 0/1 flags, no dispatch, no error translation; where pnchisq, dbinom_raw, dpois_raw live |
accudist._ufuncs |
the raw ufuncs themselves; private, fastest |
from accudist import rmath
rmath.pnchisq(3.0, 5.0, 1.5, 1, 0) # (x, df, ncp, lower_tail, log_p)