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

Re: Motion 4: P1788 on non-754?



> Date: Mon, 13 Apr 2009 15:00:37 +0200
> From: Vincent Lefevre <vincent@xxxxxxxxxx>
> To: "stds-1788@xxxxxxxx" <stds-1788@xxxxxxxx>
> Subject: Re: Motion 4: P1788 on non-754?
> 
> On 2009-04-12 16:10:18 -0500, Corliss, George wrote:
> > At one level, how something may be implemented is not the concern of the
> > standard.  At that level, the underlying hardware is irrelevant.  At another
> > level, if something is impossible to implement (efficiently, depending on
> > one's point of view), there is no point putting it into a standard.  Hence,
> > discussions leading to a standard SHOULD sometimes include discussions about
> > implementations.  That is messy enough if the discussions are about possible
> > implementations in 754.  The discussion complexity is probably unbounded if
> > we admit discussions of possible implementations on ANY architecture, known
> > or imagined.
> 
> Not necessarily. IMHO, an arithmetic (not necessarily floating-point)
> with exceptions and directed rounding (possibly not correct rounding,
> but an implementation should provide error bounds, so that the user
> can know how tight his intervals will be) should be sufficient to
> define an interval arithmetic.
> 
> Let's take an example: the double-double arithmetic, which can be
> implemented thanks to IEEE 754 BTW (so, on most architectures), and
> which is often used to extend the precision. I wonder if directed
> rounding is also used in practice with it, but anyway, this is
> possible. Should P1788 ignore such arithmetics? Is there any reason
> why such arithmetics would make IA definition significantly more
> complicated?
> 
> -- 
> Vincent Lefèvre <vincent@xxxxxxxxxx> - Web: <http://www.vinc17.org/>

	There is much here worthy of comment.

	First, George is quite correct that, while we may prefer to
	discuss our important interval issues in the abstract, it is
	useful to know how those abstractions will be made manifest.
	Or at least to have in the mind's eye how such a manifestation
	might appear without working out the gritty details while
	thinking of the abstraction.

	Thus the focus on 754 which is the subject of Motion 4.

	Next, Vincent is also correct to point out that all the
	features of 754 are not, strictly speaking, necessary for our
	interval work.  Correctly rounded arithmetic is sufficient but
	'incorrect' or approximate arithmetic is also sufficient so
	long as you know of & can apply strict error bounds.

	He uses double-double as an example.

	Here is where we part company.

	While double-double can be used as a starting point for a
	larger & more general arithmetic, as a basis for an arithmetic
	about which we wish to prove theorems it leaves much to be
	desired.  Let me point out just one problem: the density of
	double-double numbers varies with the value of the number in
	an inconsistent way.  So while 1/3 may be approached from
	either side by numbers near 1/3 +/- eps^2 (where eps ~ 2^-53),
	we have that 1/2 is surrounded by a dense group of numbers
	down to the limit of the exponent range (1/2 +/- 2^-1074).

	Thus double-double is, in fact, an excellent example of an
	arithmetic we do NOT want to consider as a basis for our
	interval standard.  Not because its bad.  Its not.  Not
	because its inaccurate.  It isn't.  But because it has very
	unusual properties.  Properties that we would have to take
	into account every time we made a decision about: containment,
	acceptable roundoff error, exceptions (one part NaN & the
	other not or one part over/underflowed & the other not), &
	on & on.

	It is best if we don't have to consider such things.  And in
	this case, by 'best' I mean 'simplest'.

	Or at least as simple as we can make it & no simpler.

	As Vincent is no doubt aware, this will cause difficulties
	for MPFR as a basis for intervals.  It is not because it is
	a variable precision arithmetic.  We made sure to account for
	such things in 754-2008.  It is because it is an arithmetic
	that has made decisions to back off on things like correct
	rounding for very good & practical reasons.

	Still, if Motion 4 passes, it does not automatically rule
	out MPFR.  After all, even though MPFR does not meet 754-2008
	there is nothing YET in 1788 that MPFR cannot do.  Those
	decisions are in the future.  We are not yet ready to make
	them.

	And things can happen between now & then.  We could decide
	that very few requirements of 754 are necessary for 1788.

	Or the MPFR folks could decide to make a 754-2008 compliant
	version.  They have the incentive, at least now they do.

	We have the time.

	If we use it wisely, that is.


				   Dan