[WP34s] Regularized incomplete Beta function

[Paul Dale]

(05-11-2014 01:23 PM)Dieter Wrote:  The CF branch uses a modified Lentz algorithm where the value for the "tiny" element is set to 1E-37. This is a common value for single precision where the working range ends right below this. The examples in the NR book use a similar value since they obviously assume single precision as well. However, in our case I feel this value should be more like 1E-6100. What do you think?

I've dropped the threshold to 1e-10000. We're using the high precision internal reals here and exponents are very wide. I also noticed a bug where I didn't take the reciprocal properly. The Chi² problems you found for high degrees of freedom, aren't impacted by this change. The series approximation is used there.

Quote:I do not have a solution yet. On the other hand I found some improvements for the Chi² quantile. The new code is running quite fast and it even handles cases where a simple Newton solver would fail. But it still needs some adjustments for high dof values with probabilities close to zero.

Sounds useful of course.


Pauli