[Libreoffice] [REVIEW] fix for fdo#37218: tinv only takes values up to 10000 as second parameter

[Markus Mohrhard]

Hello Regina, Eike,

Do you have valid reference values for the accuracy? Are the values of TINV
> worse or better than approximating with standard normal distribution?
>

it seems that the method is quite accurate up to 1E11 and with a bit more
difference up to 1E13 or 1E14 and above this goes badly wrong. But it
depends very much on the first paramter.


>
> There are some loops involved. Do you test the performance with huge
> values?
>

I haven't noticed any performance problems witht his method even that i
created a 14 times 10 table with very large numbers. The originla code is
from before 2001 so if there were any performance problems back then we
should not have them with modern computer.

>
> Kind regards
> Regina
>
> I have no definite opinion on the upper limit of the second parameter. It
is "degree of freedom" and 1E5 covers already exploring 5000 records. I
don't know whether a larger limit is needed. But the user should get an
information about problems with huge numbers, perhaps a remark in the help,
that it is not recommended to use values above ...

We have a bug report that asks for a larger upper limit and I think as long
as Excel has one and there is no technical reason to deny this we should
increase it. I think at least the 1E10 that Excel allows is something we
should do, everything above this is something for discussions.

Regards,
Markus
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/libreoffice/attachments/20110721/9c0debaf/attachment.html>