Brian -
This of course depends on _both_ the no of elements and their size
(the smaller the elements, the higher you can go)
(gm0,gm1,gm2) 441#5
5 5 5
(gm0,gm1,gm2) 442#5
_ 5 5
(gm0,gm1,gm2) 396#6
6 6 6
(gm0,gm1,gm2) 397#6
_ 6 6
or, using Raul's argument, it fails at the same order of magnitude
5^ 441 442
1.76105e308 _
6^ 396 397
1.40571e308 _
-M
At 2016-08-26 21:38, you wrote:
Martin,
Your examples of gm0 failing are not clear to me because you use random
data.
I tried the following nonrandom data, but a sort of random case where the
number of data values seems to be the issue because for samples of 441 and
less the failure is not present, but is present for 442 and greater.
Does this example suggest any more general pattern to other folks?
(gm0,gm1,gm2) 442#5
_ 5 5
(gm0,gm1,gm2) 441#5
5 5 5
On Fri, Aug 26, 2016 at 7:38 AM, Martin Kreuzer <[email protected]> wrote:
> Hi all -
> Sharing something I observed ...
> While the definitions (for arithmetic and geometric mean)
> amean=: +/ % #
> and
> gmean=: */ %:~ #
> are nice to look at as a pair (because of symmetry) I prefer this
> definition of geometric mean
> gmean=: [: */ # %: ]
> or
> NB. using the Log domain
> gmean=: [: ^ [: amean ^.
> because, given these (intermediate) definitions
> NB. often found
> gm0=. */ %:~ #
> NB. taking the roots first, then multiply
> gm1=. [: */ # %: ]
> NB. definition found on (&.:) page of NuVoc
> gm2=. amean &.: ^.
> I got these results (see gm0 failing while the other two hold)
> (gm0,gm1,gm2) >: ? 6 $~ 637
> 3.04409 3.04409 3.04409
> (gm0,gm1,gm2) >: ? 6 $~ 638
> _ 3.05264 3.05264
> (gm0,gm1,gm2) >: ? 6 $~ 2^24
> _ 2.99357 2.99357
> under J v8.04 on a 32-bit Vista machine.
> Thanks
> -M
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
--
(B=) <-----my sig
Brian Schott
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
