R devel - Dec 2009 - dchisq tail accuracy (PR#14105)

Full_Name: Jerry W. Lewis
Version: 2.10.0
OS: Windows XP Professional
Submission from: (NULL) (198.180.131.16)


dchisq is inaccurate in the extreme tails.  For instance, dchisq(1510,2,952)
returns 4.871004e-18 which is almost 15 times smaller than the correct value of
7.053889e-17.  A better approach for ncp>0 would be
  besselI(sxn,d2,TRUE) * (x/ncp)^(d2/2) * exp(sxn-(x+ncp)/2)/2
where logs may prevent over/underflow in extreme calculations.

The undefined variables in the original post are
   d2 <- df/2-1
   sxn <- sqrt(x*ncp)

Jerry
	[[alternative HTML version deleted]]