Subject: SF bug [ 2013654 ] bfloat(NaN) => finite number
From: Richard Fateman
Date: Mon, 21 Jul 2008 06:18:20 -0700
There are ways of doing some computations with infinity that also work for
certain other kinds of objects like indeterminates.
For example, generate an indexed infinity for each use. that is inf[1],
inf[2] ....
so that you don't get infinity-infinity --> 0, but inf[1]-inf[2]. Which
you can, if you wish, relate to (say) undefined[1], by further processing,
perhaps by limit().
This works for intervals too, e.g. interval(a,b,index).
interval(0,1,ind1) -interval(0,1,ind2) can be simplified to 0 if
ind1=ind2), otherwise interval(-1,1,ind3).
This does not solve 0*infinity issues.
RJF
> -----Original Message-----
> From: maxima-bounces at math.utexas.edu
> [mailto:maxima-bounces at math.utexas.edu] On Behalf Of Robert Dodier
> Sent: Sunday, July 13, 2008 12:50 PM
> To: Maxima List
> Subject: Re: [Maxima] SF bug [ 2013654 ] bfloat(NaN) => finite number
>
> Hello,
>
> I have some code (still experimental) to detect floating
> point infinity
> and not-a-number. The immediate use is to disallow bigfloat conversion
> for those values, but I am guessing there could be other uses.
>
> I tried to implement tests in terms of operations that are valid for
> all non-special floating point values. Also I tried to use properties
> of floating point arithmetic rather than inspecting bits since I don't
> know a bit-inspection method which works on special values.
>
> Context:
> http://sourceforge.net/tracker/index.php?func=detail&aid=20136
54&group_id=4933&atid=104933
>
> HTH
>
> Robert Dodier
>
>
> PS.
>
> $ cat /tmp/disallow-bfloat-nan-inf-patch
> Index: src/float.lisp
> ===================================================================
> RCS file: /cvsroot/maxima/maxima/src/float.lisp,v
> retrieving revision 1.39
> diff -u -r1.39 float.lisp
> --- src/float.lisp 14 Feb 2008 01:31:37 -0000 1.39
> +++ src/float.lisp 13 Jul 2008 19:31:41 -0000
> @@ -258,10 +258,41 @@
> (list '(rat simp) 1 (exptrl 2 (1- fpprec)))
> $ratepsilon)))))
>
> +(defun float-nan-p (x)
> + (and (floatp x) (not (= x x))))
> +
> +(defun float-inf-p (x)
> + (and (floatp x) (not (float-nan-p x)) (beyond-extreme-values x)))
> +
> +(defun beyond-extreme-values (x)
> + (multiple-value-bind (most-negative most-positive)
> (extreme-float-values x)
> + (cond
> + ((< x 0) (< x most-negative))
> + ((> x 0) (> x most-positive))
> + (t nil))))
> +
> +(defun extreme-float-values (x)
> + ;; BLECHH, I HATE ENUMERATING CASES. IS THERE A BETTER WAY ??
> + (case (type-of x)
> + (short-float (values most-negative-short-float
> most-positive-short-float))
> + (single-float (values most-negative-single-float
> most-positive-single-float))
> + (double-float (values most-negative-double-float
> most-positive-double-float))
> + (long-float (values most-negative-long-float
> most-positive-long-float))
> + ;; NOT SURE THE FOLLOWING REALLY WORKS
> + ;; #+(and cmu double-double)
> + ;; (kernel:double-double-float
> + ;; (values most-negative-double-double-float
> most-positive-double-double-float))
> + ))
> +
> ;; Convert a floating point number into a bigfloat.
> (defun floattofp (x)
> + (if (float-nan-p x)
> + (merror "Attempted float-to-bigfloat conversion of floating point
> NaN (not-a-number).~%"))
> + (if (float-inf-p x)
> + (merror "Attempted float-to-bigfloat conversion of floating-point
> infinity.~%"))
> (unless $float2bf
> (mtell "Warning: Float to bigfloat conversion of ~S~%" x))
> +
> (multiple-value-bind (frac exp sign)
> (integer-decode-float x)
> ;; Scale frac to the desired number of bits, and adjust the
> Index: tests/rtest15.mac
> ===================================================================
> RCS file: /cvsroot/maxima/maxima/tests/rtest15.mac,v
> retrieving revision 1.68
> diff -u -r1.68 rtest15.mac
> --- tests/rtest15.mac 5 Apr 2008 19:38:27 -0000 1.68
> +++ tests/rtest15.mac 13 Jul 2008 19:31:44 -0000
> @@ -983,3 +983,23 @@
>
> (reset (float_approx_equal_tolerance), 0);
> 0;
> +
> +/* disallow bigfloat conversion for floating point infinity
> and not-a-number.
> + * SF bug [ 2013654 ] bfloat(NaN) => finite number
> + */
> +
> +(foo : ?most\-positive\-double\-float,
> + bar : ?most\-negative\-double\-float,
> + block ([ratepsilon : 0], [bfloat (foo), bfloat (bar)]));
> +/* might need to replace the expected result w/ a rat-ified value to
> ensure equality ... */
> +[1.797693134862316b308, - 1.797693134862316b308];
> +
> +errcatch (bfloat (2*foo));
> +[];
> +
> +errcatch (bfloat (2*bar));
> +[];
> +
> +errcatch (bfloat (2*foo + 3*bar));
> +[];
> +
> _______________________________________________
> Maxima mailing list
> Maxima at math.utexas.edu
> http://www.math.utexas.edu/mailman/listinfo/maxima
>