Re: Request for motion Fwd: Input from IFIP WG 2.5 to IEEE Interval Standards Working Group

[Dan Zuras Intervals <intervals08@xxxxxxxxxxxxxx>]

> Date: Fri, 11 Sep 2009 04:41:18 -0400
> To: stds-1788                           <stds-1788@xxxxxxxxxxxxxxxxx>
> From: Michel Hack                                 <hack@xxxxxxxxxxxxxx>
> Subject: Re: Request for motion Fwd: Input from IFIP WG 2.5 to IEEE Interval Standards Working GroupÙ
> 
> In reply to Dan Zuras, Arnold Neumaier wrote:
> > It is impossible to choose an element from an empty set, ...
> 
> I am sure Dan was referring to a *vector* element, i.e. we are dealing
> here with vectors of intervals.
> 
> The situation is analogous in some ways to a two-operand operation
> where one operand is Empty or NaI -- but in that case we have clear
> propagation rules.  Those rules may however be INAPPROPRIATE for
> operations involving possibly large vectors, for a number of reasons.
> One of these is that implementations should be allowed to partition
> the argument set and parallelise without affecting the final result
> (unlike reduction operations, which give some latitude here).  Another
> reason is to avoid throwing away a large amount of work if one of the
> operands has a problem -- and to avoid the need to pre-scan operands.
> 
> So I would be careful to describe the 754-2008 rules as "buggy".  I
> would have to go through the voluminous stds-754 correspondence to
> remind myself of what the arguments were for the exceptional rules,
> so I won't comment further (other than to agree that I was surprised
> too when I came across them).
> 
> Michel.
> ---Sent: 2009-09-11 09:23:40 UTC

	Folks,


	Michel is correct that I was discussing an interval vector
	element.

	And yet Arnold's point, that the 754 description of reduction
	operations makes them ill defined at best, is also correct.
	As I believe I mentioned earlier.

	Also, as he says in another note, that the members of 754
	were not all in agreement that these operations should even
	be in the standard & those that did could not agree on their
	features.

	Thus my reason for urging caution as we approach this.

	You can consider linear algebra operations to be something
	of a test case for applications of a well defined interval
	standard.

	It is a test we should arrange to pass.  And, not having
	written a standard yet, I believe we are well suited to
	make such an arrangement. :-)

	I advocate no cause here other than that.

	Yours,

				   Dan