Hi Ray,
> I think I'd rather have accurate values for atanh near zero. Until
> this came up here amongst us, there was not a single report (that I
> can remember) complaining that atanh had the wrong branch cuts. That
> tells me that most users don't care, but I think most users would be
> annoyed if atanh near zero were not very accurate, because that's
> something they can verify easily themselves.
Well, cancellation near zero can be avoided by invoking
(log(1 + x) - log(1 - x))/2 only when x falls on a branch cut,
and otherwise (imagpart(x) != 0 or x real and abs(x) < 1)
punting to cl:atanh.
best,
Robert