On Fri, Sep 16, 2022 at 3:18 PM Henry Rich <[email protected]> wrote:
> The actual test as I implemented it uses 1.0-ct rather than ct and is ever
> so slightly different at the extremes of the tolerance interval but ever so
> much faster. See the TEQ macro for details.

Hmm...

Conceptually, testing based on 1-epsilon instead of 1+epsilon should
actually increase the magnitude of the region contained by epsilon
(they're approximately reciprocals, but not quite). But that's not
happening here, I think because the limiting value on epsilon is too
small for this theoretical discrepancy to manifest. Though possibly
I'm not thinking things through adequately (I remember tolerant
equality having some non-intuitive properties when epsilon approached
1.)

Anyways, it looks like the critical issue is that you're excluding the
epsilon boundary from the region contained by epsilon. That's fine, I
think )assuming I've understood this change). But we need to figure
out how to change the documentation.  I'm Cc'ing the wiki list to
either help me remember to circle back on this (or maybe someone else
will pick this up).

Thanks,

-- 
Raul
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to