> On 31 Oct 2017, at 16:46, joe darcy <joe.da...@oracle.com> wrote: >>> >> In that case we need to double (sorry) down on the NaNs and include sum as >> well: >> >> * <li>{@code (min <= max && !isNaN(sum)) || (isNaN(min) && isNaN(max) && >> isNaN(sum))} > > A more complete test for the numerical consistency conditions would be > something like > > min <= sum/count <= max > > However, that would require a bit of thought due to potential round-off so > this might not be worth the complexity trade-off. (If any of min, sum, and > max were NaN, the comparisons would be false.) >
Yes, my recollection from the discussions we concluded not to do such checks. Paul.