You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@commons.apache.org by "Gilles (JIRA)" <ji...@apache.org> on 2016/09/19 17:12:22 UTC
[jira] [Commented] (MATH-1386) EPSILON value in
org.apache.commons.math.util.MathUtils seems like half of what it should
be
[ https://issues.apache.org/jira/browse/MATH-1386?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15504032#comment-15504032 ]
Gilles commented on MATH-1386:
------------------------------
bq. When dealing with double I've always used 2.220446049250313E-16 for epsilon.
What is the definition of {{epsilon}} in the above?
bq. You have it defined in org.apache.commons.math.util.MathUtils as 1.1102230246251565E-16
In recent versions of Commons Math, this constant has been defined in class {{Precision}}.
The Commons Math documentation agrees with the following:
{noformat}
eps = 1.1102230246251565E-16
1 + eps = 1.0
1 + 2 * eps = 1.0000000000000002
1 + nextUp(eps) = 1.0000000000000002
{noformat}
See also MATH-843.
> EPSILON value in org.apache.commons.math.util.MathUtils seems like half of what it should be
> ---------------------------------------------------------------------------------------------
>
> Key: MATH-1386
> URL: https://issues.apache.org/jira/browse/MATH-1386
> Project: Commons Math
> Issue Type: Bug
> Environment: win 7
> Reporter: Brian S. McCormick
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> I've always used 1.1920929E-7 for dealing with float epsilon values. When dealing with double I've always used 2.220446049250313E-16 for epsilon.
> You have it defined in org.apache.commons.math.util.MathUtils as 1.1102230246251565E-16 which is half what I think it should be.
> I come up with these numbers using the following:
> float fEps = Float.intBitsToFloat(Float.floatToIntBits(1f) + 1) - 1;
> double dEps = Double.longBitsToDouble(Double.doubleToLongBits(1) + 1) - 1;
> Am I correct? I don't really know. I do know that float epsilon in every legacy C/C++ compiler etc I have ever used is defined as about 1e-7 and this is the value using the formula for fEps above. When I started doing comps using doubles instead of floats I started using the formula for dEps above which looks to me like the equivalent for double numbers.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)