| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
John Pryce wrote:
Nate On 26 Oct 2010, at 01:08, Nate Hayes wrote:Q3. I see no good reason why the different kinds of NaI that might be found useful in applications should be in 1-1 correspondence with the different decoration values used in whatever exception handling scheme we finalise. Justification, anyone?I don't fully understand the question.In Dan's experience, various NaN-like values were proposed, arising from user needs: "missing data", "finite real of unknown value", "arbitrarily large value", etc. though none made it to the final standard. In the present situation, the different "NaIs" arise not from user needs but from choices internal to P1788: e.g. if we go with Arnold's scheme there'll be 5 of them. Not a criticism, just a thought.
The 5 decorations I chose were just the minimum needed to make interval arithmetic safe for applications, and they have well-defined semantics. So this makes for a good object of standardization. In principle one could extend this set in many ways, at the cost of complicating the propagation rules, which are now quite simple since the 5 decorations form a canonically linearly ordered set. If one uses 3 bits for the minimum decorations, one can still use the 5 other bits in a byte to accommodate 5 more binary properties of the type you mention. But I think their meaning should not be enforced by the standard, so that different implementations have maximal flexibility. Arnold Neumaier