CDF and PMF of binomial function not same with extreme values

[mborland]

See: scipy/scipy#17809

[jzmaddock]

This may not be us:

#include <boost/math/distributions.hpp>
#include <vector>
#include <iostream>

int main() 
{
   double n = 25.0e21;
   double p = 1.0e-21;
   double r = 0;

   boost::math::binomial_distribution<double> d(n, p);

   std::cout << pdf(d, r) << std::endl;
   std::cout << cdf(d, r) << std::endl;

   return 0;
}

Outputs:

1.38879e-11
1.38879e-11

Or have I misread the OP?

[mborland]

I don't think this is us now that you confirmed. I got the same results with the following.

#include <boost/math/distributions/binomial.hpp>
#include "math_unit_test.hpp"

// See: https://github.com/scipy/scipy/issues/17809

int main()
{
    constexpr double n {25.0e21};
    constexpr double p {1.0e-21};
    constexpr double r {0};

    // In R the PDF and CDF are both 1.388794386496407e-11
    // In boost 1.81 CDF was 1.388794386496407e-11 but PDF was 1.0

    const auto binom_dist = boost::math::binomial_distribution<double>(n, p);

    CHECK_ULP_CLOSE(1.388794386496407e-11, boost::math::cdf(binom_dist, r), 1);
    CHECK_ULP_CLOSE(1.388794386496407e-11, boost::math::pdf(binom_dist, r), 1);

    return boost::math::test::report_errors();
}

[mborland]

I ran this against master too, and we still get the correct results.

[mborland]

Confirmed this fails using 1.80 but was fixed in 1.81. Scipy will update their packaged version of Boost.Math for next release.