Boost C++ Libraries: Ticket #9183: quantile for poisson distribution - precision issue?

anonymous — Mon, 30 Sep 2013 18:14:49 GMT

I'm not surprised that that doesn't work as you expected - as you said it's a rounding issue - the calculated cdf is necessarily an approximation, and if you then try to invert that the chances of getting exactly 66 (or some other integer) back are basically nil. I guess we could check to see if the real-valued result is very near to an integer and then double check to see if that integer still fulfills the specified requirements, but I'm betting that might not work in this case either...

attachment set

John Maddock — Wed, 02 Oct 2013 17:23:22 GMT

attachment → math-discrete-distro.patch

John Maddock — Wed, 02 Oct 2013 17:26:03 GMT

Can you try the attached patch? I think it fixes this and gives somewhat more sane results in the corner cases. Note however that it's still not possible to round trip in the general case - particularly when the probability is very close to 1 (within a few epsilon) as in that situation there may be multiple values for the random variable which all give the same CDF value.

anonymous — Thu, 03 Oct 2013 13:01:04 GMT

Dear John,

Thank you very much for the fixes. Yes, things look good here (I will run through even more tests later).

By "in the general case", do you mean when x (in the example above) is real, or that the "round trip" is also "bad" even when x is integral? Only the latter case is important to us.

Best regards, HS

anonymous — Thu, 03 Oct 2013 16:03:02 GMT

By "in the general case", do you mean when x (in the example above) is real, or that the "round trip" is also "bad" even when x is integral? Only the latter case is important to us.

When x is integral and large there may be multiple (all integral) values of x which all yield the same cdf, this usually occurs when the cdf is a couple of epsilon less than 1, but I found a few cases for very small cdf values as well. There's nothing that can be done about those cases - other than moving to a type with more precision, or, in the case where the cdf is near one, using complements instead.

status changed; resolution set

John Maddock — Tue, 08 Oct 2013 17:17:29 GMT

status new → closed
resolution → fixed

(In [86205]) Improve quantiles of discrete distributions to round trip integers more often. Fixes #9183.

John Maddock — Sat, 12 Oct 2013 09:41:01 GMT

(In [86259]) Apply patch for issue #9183. Refs #9183.