[
https://issues.apache.org/jira/browse/MATH-418?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13975027#comment-13975027
]
Ted Dunning commented on MATH-418:
----------------------------------
As a clarification, the median and the mean are never separated by more than a
single standard deviation. The failure of binapprox on skewed data occurs
because essentially all of the data can be in the range around the mean. A
secondary failure is on sorted data.
For example, for Gamma(0.1, 0.1), the mean is 1, the sd is 3.15 and the median
is 0.006. About 93% of the distribution is in the range [mean-sd, mean+sd]
> add a storeless version of Percentile
> -------------------------------------
>
> Key: MATH-418
> URL: https://issues.apache.org/jira/browse/MATH-418
> Project: Commons Math
> Issue Type: New Feature
> Affects Versions: 2.1
> Reporter: Luc Maisonobe
> Fix For: 4.0
>
> Attachments: patch
>
>
> The Percentile class can handle only in-memory data.
> It would be interesting to use an on-line algorithm to estimate quantiles as
> a storeless statistic.
> An example of such an algorithm is the exponentially weighted stochastic
> approximation described in a 2000 paper by Fei Chen , Diane Lambert and
> José C. Pinheiro "Incremental Quantile Estimation for Massive Tracking" which
> can be retrieved from CiteSeerX at
> [http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.105.1580].
--
This message was sent by Atlassian JIRA
(v6.2#6252)
