[
https://issues.apache.org/jira/browse/NUMBERS-148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17162337#comment-17162337
]
Mohammad Rezaei commented on NUMBERS-148:
-----------------------------------------
Simplex has very serious issues with *small* numbers (e.g. 1e-20), let alone
subnormals. There is a settable "epsilon" value (defaults to 1e-6), but even
setting that doesn't really help with small numbers, because the rest of the
matrix has relatively large numbers (1 and -1), which when
dividing/subtracting, simply blow away the small numbers.
The scaling in MATH-1549 brings the range of solvable problems quite a bit
forward, so long as all the inputs are in some way scalable. For example, if
the input has 1e-20 and 1e20, it can't do anything.
Because of the subtraction heavy algorithm, it's best to think about the number
of digits in the mantissa, which for a double is roughly 15-17, and if we want
to converge on a good solution to 6 or so digits, that effectively means
Simplex can only deal with numbers that are within 10 or so orders of magnitude.
TL;DR: the concerns are valid for general stuff, but don't affect Simplex
because other issues will dominate first.
> Extract double exponent manipulation routines from commons-math to numbers
> --------------------------------------------------------------------------
>
> Key: NUMBERS-148
> URL: https://issues.apache.org/jira/browse/NUMBERS-148
> Project: Commons Numbers
> Issue Type: Improvement
> Components: core
> Reporter: Mohammad Rezaei
> Priority: Minor
> Time Spent: 20m
> Remaining Estimate: 0h
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
