[
https://issues.apache.org/jira/browse/MATH-1146?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14099667#comment-14099667
]
Gilles commented on MATH-1146:
------------------------------
By "policy" here, I don't mean best coding practices (which includes good
documentation).
But rather, can we state a general policy on handling special values instead of
repeating all over the place (and forgetting to do so) that results could be
garbage if NaN or infinities are involved?
There are several places in CM where the implementations do not behave as one
would expect from the plain math definition, and it has led to bug reports
(recall e.g. "Complex"). I don't dispute that there may be good reasons, and
that the doc may state it clearly enough (although I don't think it that the
doc should contain what could be deemed an implementation detail, such as the
actual computation used to maintain the mean); I find it preferrable to state
once and for all that if users want to manipulate NaN and infinities, they
should not expect that CM will deliver correct results. The rationale being:
bq. \[CM is\] designed for fast, accurate computations \[... and\] we do not
add a performance hit for users with standard (non NaN, non-Inf) data \[... in
order to\] return meaningful results in the presence of INFs.
which is a quite well stated and perfectly acceptable policy IMO.
> class Mean returns incorrect result after processing an Infinity value
> ----------------------------------------------------------------------
>
> Key: MATH-1146
> URL: https://issues.apache.org/jira/browse/MATH-1146
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.3
> Reporter: david cogen
> Attachments: MATH-1146.patch
>
>
> 1. Create a Mean object.
> 2. call increment() with Double.POSITIVE_INFINITY.
> 3. Call getResult(). Result is INFINITY as expected.
> 4. call increment() with 0.
> 5. Call getResult(). Result is NaN; not INFINITY as expected.
> This is apparently due to the "optimization" for calculating mean described
> in the javadoc. Rather than accumulating a sum, it maintains a running mean
> value using the formula "m = m + (new value - m) / (number of observations)",
> which unlike the "definition way", fails after an infinity.
> I was using Mean within a SummaryStatistics. Other statistics also seem to be
> affected; for example, the standard deviation also incorrectly gives NaN
> rather than Infinity. I don't know if that's due to the error in Mean or if
> the other stats classes have similar bugs.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
