Re: NaN as Integer value
wren ng thornton <wren <at> freegeek.org>
2013-04-17 17:11:08 GMT
On 4/14/13 8:53 PM, Kim-Ee Yeoh wrote:
> On Sun, Apr 14, 2013 at 3:28 PM, wren ng thornton <wren <at> freegeek.org>
>> Whereas the problematic
>> values due to infinities are overspecified, so no matter which answer you
>> pick it's guaranteed to be the wrong answer half the time.
>> Part of this whole problem comes from the fact that floats *do* decide to
>> give a meaning to 1/0 (namely Infinity).
> I'm not sure what you mean about overspecification here, but in
> setting 1/0 as +infinity (as opposed to -infinity), there's an easily
> overlooked assumption that the limit is obtained "from above" as
> opposed to "from below."
Setting 1/0 = +inf isn't the problem, or at least not the main one. Of
course, there's always the question about which completion of the reals to
use--- i.e., whether we have +inf vs -inf, or whether we just have a
single point infinity.
Rather, the NaN problem comes from trying to resolve things like:
inf - inf
inf * 0
inf / inf
0 / 0
The overspecification problem I mentioned is that each of these
expressions has "too many" solutions. For example, we have the following
x * 0 == 0 -- where x is finite
inf * y == inf * signum y -- where y /= 0
inf * 0 == ??
x - inf == -inf -- where x is finite
inf - y == inf -- where y is finite
inf - inf == ??
x / 0 == inf * signum x -- where x /= 0
0 / y == 0 -- where y /= 0
0 / 0 == ??