inexact->exact

[Jim White]

Jim White wrote:
> Marco Vezzoli wrote:
>> jim@pagesmiths.com wrote:
>>>> ...
>>>> call to log (0.0)
>>>> return from log => #i-1/0
>>>
>>> That is the correct result.  The logarithm of zero is not a number. 
>>
>> You are right, I simply copied the routine and the examples the author
>> gived which included tests with zeroes. Maybe the scheme implementations
>> where the routine was written are more permissive (or simply wrong) with
>> the log arguments.
> 
> As far as a the specification goes, since the result of (log 0) is not 
> defined, there isn't any way to be wrong.  Other implementations are 
> probably doing what Java does, which is to return a negative infinity 
> float.  So actually I should have said it is undefined for Scheme (and 
> CL) specifications because IEEE does define -infinity as a number.
> 
> Same goes for inexact->exact's use of an exception for an invalid 
> number.  Friendlier though would be to return infinity (negative or 
> positive depending on the numerator's sign) rather than an exception at 
> that point.

Whoops.  I just checked out how infinities are handled, and Kawa's 
gnu.math is actually doing everything quite nicely.

(log 0)   ->  #i-1/0

Which is negative infinity (inexact).

(inexact->exact (log 0))  ->  -1/0

Which is negative infinity (exact).

The failure is actually due to the computation in sci where it is 
multiplying negative infinity (-1/0) by zero.

(* 0 (log 0))  ->  #i0/0

#i0/0 is the display for floating NaN.

I believe that is the correct result for IEEE fp.

Jim