I don't think the presence or absence of atlas should affect the
results from integrate. The atlas library is an enhanced, faster
version of the basic linear algebra subroutines (blas) and a few of
the lapack subroutines. Unless the C code in integrate is calling
these subroutines explicitly, atlas should not affect the results.
An example for the different behaviour is the following:
"prds" <- function (s,r,sn=1)
1/(pi*sqrt(1-s^2))/(sqrt(2*pi)*sn)*exp(-(r-s)^2/(2*sn^2))
UNDER R1.2:
===========
integrate(prds, -1,+1,minpts=100,maxpts=NULL,eps=0.01,r=0)
value relerr minpts lenwrk ifail
2.735080e-269 7.38275e-08 165 73 0
UNDER R.1.1:
============
integrate(prds, -1,+1,minpts=100,maxpts=NULL,eps=0.01,r=0)
value relerr minpts lenwrk ifail
0.3146147 0.009182386 1023 283 0
.... which makes more sense.
I'm wondering why the same integrate package under R1.1.1 uses atlas,
but under R1.2 it doesn't.