Let me add what may be obvious, but I cannot be
sure. When generating Poisson random variates using the
routines we have been discussing, people are
usually interested in cases for which the mean, lambda, is
quite modest, because when lambda gets larger, the
Poisson distribution becomes so symmetric (unskewed) as to
be almost identical to the normal distribution. So it is
with small lambda, that the skewness of the distribution
would suggest threshholding as Raul or Fishman are
suggestiong and I think small lambda is also the case for
which "double precision" is needed, not for large lambda.
(Fishman's book is from 1978, and I don't know the
meaning of "double precision" in that context relative to
ours.)
So I wonder if more could be done to study both the
threshholding and the precision requirements for
small lambda.
Would the use of x: produce more precision and could
that be done exclusively in the definition of possible? If
so, how?
possible=: (^-y) +/\@:* 1 */\@, y&%
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
