From: Morten Welinder <terra_at_gnome.org>

Date: Tue 26 Oct 2004 - 02:04:08 EST

R-devel@stat.math.ethz.ch mailing list

https://stat.ethz.ch/mailman/listinfo/r-devel Received on Tue Oct 26 02:42:24 2004

Date: Tue 26 Oct 2004 - 02:04:08 EST

A little code study, formula study and experimentation reveals that the situation is mostly fixable:

- Get rid of the explicit alpha limit. (A form of it is implicit in (2) and (3) below.)
- Use the series formula when

(x < alph + 10 && x < 0.99 * (alph + 1000))

This guarantees that the sum converges reasonably fast. (For extremely large x and alpha that will take about -53/log2(0.99) iterations for 53 significant bits, i.e., about 3700 iterations.)

3. Use the continued fraction formula when

(alph < x && alph - 100 < 0.99 * x)

Aka, you don't want to use the formula either near the critical point where alpha/x ~ 1 unless the numbers are small.

4a. Go to a library and figure out how Temme does it for alpha near x,

both large. In this case the 0.99 from above could probably be lowered a lot for faster convergence.

or

4b. Use the pnorm approximation. It seems to do a whole lot better for

alpha near x than it did for the 10:1 case I quoted.

Comments, please.

Morten

R-devel@stat.math.ethz.ch mailing list

https://stat.ethz.ch/mailman/listinfo/r-devel Received on Tue Oct 26 02:42:24 2004

*
This archive was generated by hypermail 2.1.8
: Wed 03 Nov 2004 - 22:45:25 EST
*