On Mar 7, 2:14 pm, Christian Heimes <li...@cheimes.de> wrote: > Steven D'Aprano wrote: > > Yes. Floating point NANs are required to compare unequal to all floats, > > including themselves. It's part of the IEEE standard. > > As far as I remember that's not correct. It's just the way C has > interpreted the standard and Python inherited the behavior. But you may > proof me wrong on that.
Steven's statement sounds about right to me: IEEE 754 (the current 2008 version of the standard, which supersedes the original 1985 version that I think Robert Kern is referring to) says that every NaN compares *unordered* to anything else (including itself). A compliant language is required to supply twenty-two(!) comparison operations, including a 'compareQuietEqual' operation with compareQuietEqual(NaN, x) being False, and also a 'compareSignalingEqual' operation, such that compareSignalingEqual(NaN, x) causes an 'invalid operation exception'. See sections 5.6.1 and 5.11 of the standard for details. Throughout the above, 'NaN' means quiet NaN. A comparison involving a signaling NaN should always cause an invalid operation exception. I don't think Python really supports signaling NaNs in any meaningful way. I wonder what happens if you create an sNaN using struct.unpack(suitable_byte_string) and then try to do arithmetic on it in Python... Mark -- http://mail.python.org/mailman/listinfo/python-list
