CCL: negative bsse content

[Serguei Patchkovskii <ps*_*ned.sims.nrc.ca>]

 Sent to CCL by: Serguei Patchkovskii [ps,ned.sims.nrc.ca]
 On Wed, 31 May 2006, Jozsef Csontos jozsefcsontos|a|creighton.edu wrote:
 > BSSE = E_{A}^{AB}(AB) + E_{B}^{AB}(AB) - E_{A}^{A}(AB) - E_{B}^{B}(AB)
 [...]
 > So, apart from the signs, I should always get results on the same side
 > of zero, but I didn't.
 Given the above definition, you can expect BSSE to be non-positive as long
 as the method used to calculate the energies is variational. The argument
 goes as follows: A bigger variational space (A in the combined A+B basis set)
 can only lead to lower energies compared a smaller variational space (A in
 the A basis set alone). Therefore, the difference E_{A}^{AB}(AB)-E_{A}^{A}(AB)
 must be negative (or zero). Same holds for the energy of B - so the sum of
 the two quantities must be non-positive as well.
 This argument breaks down if:
 a) the method you are using is not variational
 or
 b) the basis set alone does not represent the entire variational space of
    the method
 or
 c) there is numerical noise in the calculations, which exceeds the magnitude
    of the BSSE
 Because MP2 energy is not variational, there is no reason to expect a
 definite sign of BSSE for MP2 (either canonical or localized).
 In the case of DFT, there is both a "hidden" variational space
 (numerical
 integration grid) and a possibility of having significant numerical noise.
 Both effects should decrease with the use of better grids.
 Serguei