@charris wrote on 2008-05-17

I will fix this in 1.2, not sooner. I have cleanups on queue for umathmodule and I don't want to complicate the merge.

Attachment added by trac user brorson on 2008-05-17: NumPy_NAN.diff

Milestone changed to 1.2.0 by @charris on 2008-05-17

@charris wrote on 2008-07-20

Applied, but NAN doesn't work on MSVC. I replaced it with x when isnan(x) tested true.

@charris wrote on 2008-07-21

This patch requires fixing some conversions of NaN to at least integer types. I've reverted the fix until all the implications can be worked out.

@cournape wrote on 2009-03-02

Will you have time to look at it for 1.3.0 ? If not, change to what you think is best

@charris wrote on 2009-03-10

IIRC, there was some discussion of what should be returned as the sign of nan. Currently it is zero for all inexact types. We could also return the sign bit, or just return nan. Maybe we should bring this up on the list again.

@charris wrote on 2009-03-10

As of r6621,r6624 the sign of real or complex nans is nan.
The Logical value of nan is True. Note that the integer value of nan remains 0 which conflicts with the logical value of nan. We may want to start raising exceptions when nan is cast to an integer.