[
https://issues.apache.org/jira/browse/STATISTICS-25?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17220111#comment-17220111
]
Andreas Stefik commented on STATISTICS-25:
------------------------------------------
Hi there Gilles,
First, thanks so much for checking and responding so quickly. I appreciate it.
I'm not a Python expert, but scipy's continuous distributions are here:
[https://github.com/scipy/scipy/blob/v1.5.3/scipy/stats/_continuous_distns.py]
My best guess is that the distributions for student's t start around line 5893,
with_cdf as the function. That seems to call down to their special functions
library, (e.g., sc.). So far as I can tell, that connects down to the CEPHES
library in C. I was actually able to track that down here:
[https://github.com/scipy/scipy/blob/master/scipy/special/cephes/stdtr.c]
I haven't looked in detail at their algorithm, but on brief inspection it looks
like a different approach. How accurate it is I couldn't say, but hopefully
that helps.
> T Distribution Inverse Cumulative Probability Function gives the Wrong Answer
> -----------------------------------------------------------------------------
>
> Key: STATISTICS-25
> URL: https://issues.apache.org/jira/browse/STATISTICS-25
> Project: Apache Commons Statistics
> Issue Type: Bug
> Reporter: Andreas Stefik
> Priority: Major
>
> Hi There,
> Given code like this:
>
> import org.apache.commons.math3.analysis.UnivariateFunction;
> import org.apache.commons.math3.analysis.solvers.BrentSolver;
> import org.apache.commons.math3.distribution.TDistribution;
> public class Main {
> public static void main(String[] args) {
> double df = 1E38;
> double t = 0.975;
> TDistribution dist = new TDistribution(df);
>
> double prob = dist.inverseCumulativeProbability(1.0 - t);
>
> System.out.println("Prob: " + prob);
> }
> }
>
> It is possible I am misunderstanding, but that seems equivalent to:
>
> scipy.stats.t.cdf(1.0 - 0.975, 1e38)
>
> In Python. They give different answers. Python gives 0.509972518193, which
> seems correct, whereas Apache Commons gives Prob: -6.462184036284304E-10.
> That's a huge difference.
> My hunch is that as you get closer to infinity it begins to fail, but I
> haven't checked carefully. For calls with much smaller degrees of freedom,
> you get answers that are basically the same as Python or online calculators.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
