24 Oct
2011
24 Oct
'11
8:24 a.m.
On 19/10/11 4:19 PM, Fred Cisin wrote:
That "count" wouldn't mean much. But you can track, at least through
simple operations, how much of the value is significant. Compared to the
agonies of C, this would be a straightforward addition to any Lisp
numeric tower.
The idea has been around for a while - usually named "significance
arithmetic". In particular, Mathematica implements a related system (fwiw).
* http://mathworld.wolfram.com/SignificanceArithmetic.html
* http://www.wolfram.com/technology/guide/SignificanceArithmetic/
*
http://www.mathkb.com/Uwe/Forum.aspx/mathematica/21243/remarks-on-significa…
* and some of the Usenet links I posted earlier. Steve Richfield has
advocated this proposal as an important refinement of IEEE-754.
--T
On Wed, 19 Oct 2011, David Cantrell wrote:
Or rather, *how much* of the value is significant.
I would like to be able to:
* mark a variable as being an approximation, and with how precise it is;
THAT is already present. "float" in C means that it is an approximation
to 24 binary significant digits.
It does NOT, admittedly, however, tell you that any numbers that it was
based on were also only approximate. * for those markers to propagate to any result
that arises from a
calculation involving that variable;
THAT would be NICE!
I'm curious how you would implement it. I guess that you could create a
"floatpt struct" that contains a float and an int count of how many floats
were used to create the value. That count could "approximate" how many of
the least significant bits are meaningless.
--
Grumpy Ol' Fred cisin at xenosoft.com