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

Re: Comparisons and decorations, part 2



John Pryce wrote:
Dan, Arnold, Nate, P1788

On 26 Sep 2010, at 13:54, Dan Zuras Intervals wrote:
Arnold Neumaier wrote:
Dan Zuras Intervals wrote:
On 25 Sep 16:48 John Pryce wrote:
On 24 Sep 2010, at 11:48, Arnold Neumaier wrote:
Nate Hayes wrote:
special cases for
 A \interior B
such as when A and B are both Entire or A=[1,Infnity]
and B=[0,Infinity]. In these cases, the interior operator
would need to return a different result than when A and B
are compact intervals, such as A=[1,100] and B=[0,200].
??? There is a uniform formula in 754:

[al,au] interior [bl,bu] iff ~(bl-al>=0) and ~(au-bu>=0).
Looking at 754-2008 §5.3.1, I think
 (nextDown(al) >= bl) and (nextUp(au) <= bu)
also works, since nextDown(-oo) = -oo and nextUp(+oo) = +oo.

		interior([al,au],[bl,bu]) ==
			((bl < al) || (bl == -infinity)) &&
			((au < bu) || (bu == +infinity))
None of the three formulas works correctly in some cases where
A or B or both are Empty, represented as a pair of NaN's,
(Note that Empty is interior to every interval.)

So some additional patching seems unavoidable anyway.

Arnold Neumaier

	I guess it depends on what we store in the interval
	part when we are trying to represent empty.

What I get from this thread is that
- We have had at least four valid formulas presented for the topological "A interiorTo B" when A and B are nonempty, all of which, in principle, seem to allay Nate's worst fears of performance hit. My nextDown/nextUp method is clearly bad if these are done in software, but they are such trivial operations given the 754 number encoding, and so useful, that I would hope a bit of chip area would be devoted to hard-wiring them in the near future.

- I am very sceptical that anyone could predict which of the 4 will work best on a given commonly used architecture, and that probably depends strongly on how vectorised the code is.

But it doesn't matter very much since interval comparisons are used rarely in typical algorithms for verified computing. Thus their efficiency is much less important than the efficiency of the elementary
arithmetic operations.


- An "overall best way" to represent Empty in an inf-sup interval type depends on many things and I am equally sceptical that we can predict it. So, reluctantly, I believe P1788 should NOT standardise a representation of Empty this time around. It can do so at the next revision of 1788, if existing implementations show a "best way" has emerged.

I think the representation should be decided by what makes the binary
elementary arithmetic operations most efficient, since this is what determines the speed of most algorithms for verified computing.

This makes the representation of Empty by [NaN,NaN] stand out uniquely.


Arnold Neumaier