I think that our algorithm in CDFLIB is better. Info on cdflib and
all code that we distribute can be found at URL
http://odin.mdacc.tmc.edu/anonftp/
If I recall correctly (it has been a while), Algorithm 243 computes
exp(-lamda*lambda/2) or some such on the first term of the series to
be summed where lambda is the noncentrality parameter. This makes the
answer zero for moderately large lambda.
Again, if I recall correctly, this algoritm could be greatly improved
by computing the log of this term (in toto) and then exponentiating.
Barry W. Brown
Department of Biomathematics,
Box 237
University of Texas M. D.
Anderson Cancer Center
1515 Holcombe Blvd
Houston, TX 77030
[log in to unmask]
|