On 09/20/2011 04:42 PM, Michael D Godfrey wrote:
On 09/20/2011 04:28 PM, Ben Abbott wrote:
Rik,
I'm happy to see you complete this effort.
With this change, I see 5 failures. They all appear to be due to tolerances. All tests pass if I use a tolerance of eps(). My diff is also below.
Is adding tolerance appropriate? If so, I'm happy to push a changeset.
Ben
Running on Linux, I see
no failures. So, this may need some checking.
Michael
Ben,
I'm not too concerned with the tpdf instance so I pushed a changeset
to add an eps tolerance in that case.
Like Michael, I'm running Linux and I don't see any issue with
binopdf. My guess is that it is the function gammaln() which is
returning ever so slightly different values. It shouldn't be too
hard to track down where the first difference arises in the
calculation chain. However, after that I'm not sure what we would
do since I don't think you have ready access to a different version
of gammaln(). Thus, I'm tempted to just add eps to the end of all
of the tests. It's a bit of a shame as the values in question, [1/4
1/2 1/4], are simple and exact.
I'll let Michael comment on whether there is a better approach.
--Rik
|