Geometric, negative binomial and poisson fail for extreme arguments
Josef, This is ticket #896 from two years ago. IIRC, there was some more recent discussion on the list of some of these. Do you know what the current state of these distributions is? Chuck
On Tue, May 25, 2010 at 10:34 PM, Charles R Harris <charlesr.harris@gmail.com> wrote:
Josef,
This is ticket #896 from two years ago. IIRC, there was some more recent discussion on the list of some of these. Do you know what the current state of these distributions is?
I don't have any information on these and I don't remember any discussion (and a quick search didn't find anything). I never looked at the integer overflow problem, besides reading the ticket. All 3 distributions are used in scipy.stats and tested for some regular values. (my not very strong opinion: for consistency with the other distributions, I would go with Robert's approach of rejecting overflow samples. I don't know any application where the truncation would have a significant effect. In scipy.stats I switched to returning floats instead of integers for ppf, because we need inf and nans.) BTW: If you are fixing things in np.random, then depreciating and renaming pareto as we discussed recently on the list would help reduce some confusion. I don't think we filed a ticket. Josef
Chuck
_______________________________________________ NumPyDiscussion mailing list NumPyDiscussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpydiscussion
On Tue, May 25, 2010 at 9:20 PM, <josef.pktd@gmail.com> wrote:
On Tue, May 25, 2010 at 10:34 PM, Charles R Harris <charlesr.harris@gmail.com> wrote:
Josef,
This is ticket #896 from two years ago. IIRC, there was some more recent discussion on the list of some of these. Do you know what the current state of these distributions is?
I don't have any information on these and I don't remember any discussion (and a quick search didn't find anything). I never looked at the integer overflow problem, besides reading the ticket.
All 3 distributions are used in scipy.stats and tested for some regular values.
(my not very strong opinion: for consistency with the other distributions, I would go with Robert's approach of rejecting overflow samples. I don't know any application where the truncation would have a significant effect. In scipy.stats I switched to returning floats instead of integers for ppf, because we need inf and nans.)
BTW: If you are fixing things in np.random, then depreciating and renaming pareto as we discussed recently on the list would help reduce some confusion. I don't think we filed a ticket.
OK, but it would help if you did file a ticket. And if you think truncation is the way to go on the #896 could you post a note there also? Chuck
participants (2)

Charles R Harris

josef.pktd＠gmail.com