Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Motion P1788/M0014.01: 6.1_and_6.2_of_document: up for discussion



On 2010-04-27 11:36:39 -0700, Dan Zuras Intervals wrote:
> 	First, there is the point that mid-rad intervals form
> 	a different subset of real intervals than do inf-sup
> 	intervals.
> 
> 	I believe that all inf-sup intervals of finite width
> 	can be represented in mid-rad form of sufficient
> 	precision.  That is the common subset.
> 
> 	(There is a bit of a problem when the sum of the ULPs
> 	of the inf & the sup is odd but let's sweep that under
> 	the term 'sufficient precision' for now & ignore it.)

I think that the most important point is to have a way to represent
and compute intervals with a single full-precision value. The midrad
format would not be a goal, but just a mean to implement efficient
multiple-precision interval arithmetic. The (mid,del1,del2) format
allows that too (applications could also use alternative formats
such as (mid,del) with del = del2 = -del1 if this is interesting).
However, in the former case, I wouldn't say "mid" but "approx".

Moreover it is not obvious that one would have

  lowerBound = mid + del1 & upperBound = mid + del2 EXACTLY

In fixed-point arithmetic, this may bo OK. But if one requires that
lowerBound, upperBound and mid are floating-point numbers in some
format (possibly multiple precision) and if the implementation
represents del1 and del2 as floating-point numbers in some other
floating-point format (with less precision, for efficiency), then
there's a problem.

> 	But the semi infinite intervals, [2,+oo], [-oo,3], &
> 	the like, live in the inf-sup form & not the mid-rad
> 	form.

In most cases where midrad would be used, I'd say that all
practical information is lost when one of the bounds is
infinite.

> 	On the other hand, mid-rad intervals of the form
> 	veryLarge +/- verySmall exist even in precisions
> 	substantially less precise than verySmall/veryLarge.
> 	These have no counterpart in the inf-sup world.

You mean intervals where verySmall is significantly smaller than
the ulp of veryLarge?

> 	Still, since such intervals form a non-overlapping
> 	set surrounding only the representable veryLarge
> 	numbers, I think they are of limited utility in
> 	converging on representable solutions.

Yes.

> 	But the computational advantage of mid-rad forms is
> 	so great in some applications that I believe it will
> 	be used anyway.
> 
> 	After all, it is sufficient if inf-sup intervals are
> 	converted on input, the entire calculation performed
> 	in mid-rad form, & the results are converted back to
> 	inf-sup form.  The final result may be more or less
> 	accurate than if it were done all in inf-sup form.

Yes.

-- 
Vincent Lefèvre <vincent@xxxxxxxxxx> - Web: <http://www.vinc17.net/>
100% accessible validated (X)HTML - Blog: <http://www.vinc17.net/blog/>
Work: CR INRIA - computer arithmetic / Arénaire project (LIP, ENS-Lyon)