Sean Owen created MATH-1058: ------------------------------- Summary: Beta, LogNormalDistribution, WeibullDistribution give slightly wrong answer for extremely small args due to log/exp inaccuracy Key: MATH-1058 URL: https://issues.apache.org/jira/browse/MATH-1058 Project: Commons Math Issue Type: Bug Affects Versions: 3.2 Reporter: Sean Owen Priority: Minor
Background for those who aren't familiar: math libs like Math and FastMath have two mysterious methods, log1p and expm1. log1p(x) = log(1+x) and expm1(x) = exp(x)-1 mathetmatically, but can return a correct answer even when x was small, where floating-point error due to the addition/subtraction introduces a relatively large error. There are three instances in the code that can employ these specialized methods and gain a measurable improvement in accuracy. See patch and tests for an example -- try the tests without the code change to see the error. -- This message was sent by Atlassian JIRA (v6.1#6144)