Dear R developers,
At the same time I discovered a flaw in Marsaglia-Multicarry +
Kinderman-Ramage, I found another in Marsaglia-Multicarry + Ahrens-Dieter.
It is less obvious than for Kinderman-Ramage; so I created a new thread for
this bug.
The following code shows the problem (tested on R 4.1.1 x86_64 for Windows 10):
== start of code sample ==
set.seed(1, "Marsaglia-Multicarry", normal.kind="Ahrens-Dieter")
v=rnorm(1e8)
q=qnorm(seq(0.01, 0.99, 0.01))
cv=cut(v, breaks=c(-Inf, q, +Inf))
observed=table(cv)
chisq.test(observed) # p < 2.2e-16
== end of code sample ==
The chisq.test returns a P-value < 2.2e-16 while it was expected to return a
non-significant P-value.
The additionnal code below, shows severe irregularities in the distribution of
quantiles:
== continuation of code sample ==
expected = chisq.test(observed)$expected
z = (observed - expected)/sqrt(expected)
mean (abs(z) > 6) # 58% of z-scores are greater than 6 while none should be
== end of code sample ==
The bug is specific to the combination Marsaglia-Multicarry + Ahrens-Dieter.
There is no problem with Marsaglia-Multicarry + Inversion or Mersenne-Twister +
Ahrens-Dieter
I would expect at least a warning (or an error) from R for such a buggy
combination.
--
Sincerely
Andr� GILLIBERT
[[alternative HTML version deleted]]
______________________________________________
[email protected] mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
