refactor: Hypothesis Testing

[huangziwei]

The main changes this round of refactoring is to bring randomization test (for p-values) to more hypothesis tests.

New

Resampling-based p-values aka randomization test

previously only rayleigh_test has a resampling p-value, we now roll it out to most of the hypothesis testing (some still lack it because it doesn't make a lot of sense).

rayleigh_test(alpha)                      # analytic, method="asymptotic"
rayleigh_test(alpha, n_resamples=9999)    # method="monte_carlo"
concentration_test([a, b], n_resamples=9999, seed=1)   # method="randomization"
watson_test(a, dist="vonmises", n_resamples=999)       # method="parametric_bootstrap"

Check docstring for usage. But in short:

• most hypothesis tests use n_resmpling=0 for analytical p-value and n_resampling>0 for simulation-based one.
• but for symmetry_test(method="wilcoxon"|"pewsey") , where wilcoxon has no simulation based p-value, only pewsey has; watson_test(method="uniform"|"vonmises"), where uniformonly work for both n_resamples=0 or >0, butvonmisesonly works forn_resamples>0`.

new hypothesis tests

• kuiper_two_test, which matches CircStat and pycircstat's two sample kuiper and circ_kuipertest . The original kuiper_test matches R's CircStats and circular's kuiper and kuiper.test
• watson_test(dist="vonmises") matches circular's watson.test(dist="vonmises")

Bug fixes

• omnibus_test: denim can be <=0 as m approached n/2, and we wrongly reported the pval=0 which should be 1.
• watson_u2_test: cumfreq assumed input data are sorted but we didn't
• wheeler_watson_test: didn't consider ties
• wallraff_test: didn't work for grouped data

all fixed.

Others

typos, style, lints, api harmonization, etc. most changes should be backward compatible.