Re: Do we need multiple forms of the SQL2003 statistics

On Sat, 29 Jul 2006, Peter Eisentraut wrote:

> Sergey E. Koposov wrote:
>> I think since we are supporting the numeric type as a special
>> high-precision type, Postgres must have the high-precision
>> versions of all computational functions. Just my opinion.
>
> Another way to look at it is whether you want to have accurate
> computations (numeric) or approximate computations (float). I'm not a
> statistician, so I don't know what most of these functions are used
> for. From a mathematician's point of view, however, some of these
> functions normally produce irrational numbers anyway, so it seems
> unlikely that numeric will be useful. But looking at the definition
> of, say, regr_avgx(Y, X), if all the input values are integers, it
> might be useful if I could get an exact integer or rational number as
> output, instead of a float, that is.

Exactly from the statistical point of view, there is no need to have the
integer output of those 2-arg. aggregates. For example corr(), regr_*()
are by definition not integer (they just don't have any sense as
integers...)( -1<= corr(Y,X)<=1 ) (for example the stddev(int) do
not produce int also, because it does not have any sense)

So it's perfectly fine that they are producing only floating numbers...

Regards,
Sergey

*******************************************************************
Sergey E. Koposov
Max Planck Institute for Astronomy/Sternberg Astronomical Institute
Tel: +49-6221-528-349
Web: http://lnfm1.sai.msu.ru/~math
E-mail: math(at)sai(dot)msu(dot)ru