[
https://issues.apache.org/jira/browse/MATH-1253?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14702114#comment-14702114
]
Bill Murphy commented on MATH-1253:
-----------------------------------
I agree with Otmar's comments entirely.
I do not yet know a better suggestion. Even a slight extension of the range of
correct results seems to me to be an improvement. I would still prefer going
forwards with these mods as suggested, but a true fix as outlined by Otmar
would be hugely better. Does someone else know enough numerical analysis stuff
to provide a mostly-correct approximation over the range of doubles? It is
sadly beyond me as of today.
> Skewness could get more precision from slightly reordered code.
> ---------------------------------------------------------------
>
> Key: MATH-1253
> URL: https://issues.apache.org/jira/browse/MATH-1253
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.5
> Reporter: Bill Murphy
> Priority: Minor
>
> In Skewness.java, approx line 180, there is code like:
> {noformat}
> double accum3 = 0.0;
> for (int i = begin; i < begin + length; i++) {
> final double d = values[i] - m;
> accum3 += d * d * d;
> }
> accum3 /= variance * FastMath.sqrt(variance);
> {noformat}
> If the division was moved into the for loop, accum3 would be less likely to
> overflow to Infinity (or -Infinity). This might allow computation to return a
> result in a case such as:
> {noformat}
> double[] numArray = { 1.234E11, 1.234E51, 1.234E101, 1.234E151 };
> Skewness skew = new Skewness();
> double sk = skew.evaluate( numArray );
> {noformat}
> Currently, this returns NaN, but I'd prefer it returned approx
> 1.154700538379252.
> The change I'm proposing would have the code instead read like:
> {noformat}
> double accum3 = 0.0;
> double divisor = variance * FastMath.sqrt(variance);
> for (int i = begin; i < begin + length; i++) {
> final double d = values[i] - m;
> accum3 += d * d * d / divisor;
> }
> {noformat}
> Thanks!
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
