High exponent numbers with scipy.stats functions

Question

I have a set of number that can get very small, from 1e-100 , to 1e-700 and lower. The precision doesn't matter as much as the exponent. I can load such numbers just fine using Pandas by simply providing Decimal as a converter for all such numeric columns.

The problem is, even if I use Python's Decimal , I just can't use scipy.stats.chi2.isf and similar functions since their C code explicitly uses double . A possible workaround is that I can use log10 of the numbers. The problem here is that although there is logsf function, for chi2 it's implemented as just log(sf(...)) , and will, therefore, fail when sf returns 0 where it should've returned something like 1e-600 . And for isf there is no such log function at all.

I wanted to know if there is any way to work with such numbers without resolving to writing all these functions myself for Decimal.

Answer 1

You could just write the expression for logsf directly using logs of gamma functions from scipy.special (gammaln, loggamma). And you could send a pull request implementing the logsf for the chi-square distribution.