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



> From: "Svetoslav Markov" <smarkov@xxxxxxxxxx>
> To: "R. Baker Kearfott" <rbk@xxxxxxxxxxxxx>,
>  stds-1788 <stds-1788@xxxxxxxxxxxxxxxxx>
> Date: Thu, 22 Apr 2010 12:48:54 +0300
> Subject: Re: Motion P1788/M0014.01: 6.1_and_6.2_of_document: up for discussion
> 
> P-1788:
> 
> can somebody explain what for is the line of text in 6.1 that says:
> 
> "provided that it shall be possible to retrieve the bounds of x exactly."
> 
> What is the reason and motivation for this requirement? . 
> I cannot think of any reason why "exactly" should be required. 
> One would think inclusion isotonicity is enough.  Besides,
> it is not clear in what sense the word  "exactly" is used -
> in the sense of real or machine arithmetic (that is at most 
> 1 ulp lost)? 
> 
> Regards,
> 
> Svetoslav
> 

	Well, I did not write this text but I'll tell you what
	it means to me.

	I interpret 'exactly' in this context as stronger than
	isotonicity.  Stronger even than 1 ULP lost.  I interpret
	'exactly' as meaning that no information is lost in the
	recovery of the bounds from whatever internal format is
	used.

	Formally, exactly means that the implementer's format
	together with the method of recovering the bounds is
	indistinguishable from the ordered pair [lowerBound,
	upperBound] in some given floating-point format.  Stated
	slightly differently, the set of intervals representable
	must be indistinguishable from the set of intervals
	representable by [lowerBound, upperBound] where both
	bounds are numbers in a given floating-point format.

	Assuming that there exists -x for all x in a given
	floating-point format we have that all of [lowerBound,
	upperBound], [-lowerBound,upperBound], & [lowerBound,
	-upperBound are indistinguishable in this sense.  (So
	is [-lowerBound,-upperBound] for that matter but it
	doesn't seem to have any merit WRT fixing the rounding
	mode problem. :-)

	One subtle point that is not made clear by this is that
	it requires that a representation of the form (mid,del1,
	del2) which represents the interval [mid+del1,mid+del2]
	must do so exactly.  That is lowerBound = mid + del1 &
	upperBound = mid + del2 EXACTLY, with no rounding errors.

	This is not so much of a problem but it does mean that
	any system choosing to use this mid-rad form in an
	attempt to speed things up in the |rad| << |mid| case
	must find some other way of representing semi-infinite
	intervals such as [2,+oo] & [3,+oo] which don't fit in
	this format.

	Another thing one should note is that while the constraint
	to recover 'the bounds of x exactly' is REQUIRED (under
	the verb 'shall') the entire discussion of possible formats
	is within a note (that is, informative ONLY & not at all
	required).

	I compliment the author on being careful in how this is
	written.

	So, does that clarify things or just make it harder?

	Oh, well.  I tried... :-)


				Dan