Alex Herbert created STATISTICS-37:
--------------------------------------

             Summary: Better inverse Erfc function for the Normal distribution
                 Key: STATISTICS-37
                 URL: https://issues.apache.org/jira/browse/STATISTICS-37
             Project: Apache Commons Statistics
          Issue Type: Improvement
          Components: distribution
    Affects Versions: 1.0
            Reporter: Alex Herbert


The Normal distribution uses InverseErfc for the inverse CDF. This is a wrapper 
for InverseErf and as such it is limited in precision as p -> 0 since the 
method computes using (2p - 1). When p is < 1e-16 then the inverse CDF returns 
-infinity. The smallest z is around -8.2
as computed by matlab:

norminv(1e-16)
-8.222082216130437

scipy, R, matlab, octave, Mathematica all have methods to invert small p. Here 
is output from matlab for norminv(5e-324): -38.472346342768788. It is accurate 
down to the minimum value for a 64-bit float.

There are two free libraries that seem to be used for this: Boost and Cephes 
(function ntdri). The Boost licence is compatible with Apache and there are 
some Boost derived works already in Commons Numbers. Incorporating the Boost 
Erf functions into numbers would be useful.

The inability to invert the Erfc for all p affects the Normal, Truncated normal 
and Levy distributions.

Boost error function inverses:
[https://www.boost.org/doc/libs/1_77_0/libs/math/doc/html/math_toolkit/sf_erf/error_inv.html]



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Reply via email to