Closed infinity0 closed 2 years ago
I've added another commit that adds constructors for those distributions in terms of the mean and stddev, rather than their formal parameters. Lemme know if you'd rather have them as a separate PR, I figured it was OK here since they're quite small.
The weibull one is not very satisfactory but it was the best one I could find online, and I still don't understand where that -1.086 number comes from, the paper I cite just plucks it out of thin air with no explanation...
Thanks! Yes it was good to run them, I missed them at first. Caught a few typos and corner cases. :)
I finally got around to merge this PR. Thanks a lot and sorry for the wait
Thanks for the quick review! I've changed to using
expm1
andlog1p
as you said, and added some tests.Since lognormal generates very large values in
quantile
I had to tweakcomplQuantileCheck
to make it pass, I figured since the relative error was OK it was worth keeping the test enabled, but you should take a look. Likewise I had to setprec_quantile_CDF _ = (64,64)
,48
was insufficient, I hope this is OK too.I also changed
lognormalDistrE
to forbid too-large values of sigma (in practise, > about 20) because these cause the variance to be> m_huge
.