Re: Switch rounding mode

[Jason Riedy]

Przemek Klosowski a écrit :
> If you prefer another example [c]=[a]+[b]
> > c.min = a.min \+/ b.min
> > c.max = a.max /+\ b.max
> 
> \+/ and /+\ are + operator with rounding downward for the first and
> upward for the second.

I recommend looking into Siegfried Rump's INTLAB for Matlab™ [1],
and also some of his relevant publications.

Note that IEEE-754 rounding modes sometimes are ignored by
compilers; and sometimes the C99 functions are there but they don't
actually do anything.  It's really, really annoying…

There is a somewhat portable option that may or may not be slower
(depending on the platform).  All operations have a form with an
error representation.  The two you picked just happen to be "easy".
A doubled double representation will produce a result + an error,
and you can use that error to round the result appropriately.
Multiplication and division are a little different but possible.

See Rump's pubs on "error free transformations" and accurate dot
products; you'll definitely want the latter for some matrix
operations.  Another handy reference is Bailey, Hida, et al.'s
quad-double [2] and higher precision implementations [3].  

Note that some interval algorithms for "simple" problems like
solving Ax=b for the smallest enclosure x are NP-hard… Don't get
your hopes too high for interval work.  Intervals tell you when you
might have a problem, but they don't necessarily help you find a
better answer without further analysis.

I have a really, really bad doubled-double implementation using
Octave's new object support.  Mail me if you want it, but you'd be
better off working from scratch.

Jason

Footnotes: 
[1]  http://www.ti3.tu-harburg.de/rump/intlab/

[2]  http://www.cs.berkeley.edu/~yozo/papers/LBNL-46996.ps.gz

[3]  http://crd.lbl.gov/~dhbailey/mpdist/