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

Re: Fw: Overflow and Inf



If I may be permitted to phantasize a bit further, it is actually
conceivable that the decorated-inifinity concept could be adapted
to BFP.

In 2008 we were still unable to define the Q-bit (QNaN vs SNaN)
for BFP -- only to recommend a choice.  Eventually we may reach
the point where everybody agrees.  (For DFP it was nailed down.)

Suppose we reserve TWO high-order fraction bits instead of one
for the highest-biased-exponent (all ones) case.  Then we could
have:
  00     Undecorated Infinity; remaining bits assumed zero
  01     Signalling NaN (with payload)
  10     Decorated Infinity (further decoration details?)
  11     Quiet NaN (with payload)

Unfortunately that doesn't work; in fact, no 2-bit method stands
a chance to be compatible: to be useful, Decorated Infinity would
have to be a Signalling NaN on old equipment.  But Signalling NaNs
should also be convertible to QNaNs by a simple bit-OR.  Would three
bits work?

There is another problem, already discussed in the earlier discussions
of this topic:  the inverse of a decorated infinity needs to be a
decorated zero, or (as I called it previously) an inexact signed zero,
i.e. the equivalent of Underflow.  This may not be necessary for IA,
but an FP standard needs to think about it -- and it's not pretty.

**********

In conclusion, I think this issue needs to be resolved strictly in
the Interval Arithmetic context.  There is no sensible FP equivalent.
We may need FP-primitive assistance along the lines discussed by
Arnold Neumaier and Ian McIntosh -- but we must not try to assign
FloatingPoint semantics to it.

Michel.
---Sent: 2010-09-29 21:14:16 UTC