Feb. 16, 2010
11:06 p.m.
On Tue, Feb 16, 2010 at 10:46 PM, Steven D'Aprano <steve@pearwood.info>
What's the justification for that convention? It seems wrong to me.
It's difficult to do better than to point to Kahan's writings. See http://www.eecs.berkeley.edu/~wkahan/ieee754status/IEEE754.PDF and particularly the discussion on page 8 that starts "Were there no way to get rid of NaNs ...". I don't think it covers hypot, but the same justification given for having nan**0 == 1 applies here. Interestingly, he says that at the time of writing, 1**nan == nan is the preferred alternative. But since then, the standards (well, at least C99 and IEEE 754-2008) have come out in favour of 1**nan == 1. Mark