RE: QCISD energy lower than QCISD(T) energy!

[Jan Hrusak <jan.hrusak "-at-" jh-inst.cas.cz>]

 I would like to add also my contribution to the recent discussion on
 QCISD/QCISD(T). I have been using these methods for a quite long
 time. They became very popular due to the implementation into all
 (or almost all) of the QC programs. The wide use of the QCI methods
 is also due to the implementation and particular success of the G1/G2
 methods. For many (gas phase) experimentalists the numbers produced by
 using the GAUSSIAN G2 keyword are a kind of final answer if the experimental
 data are not listed in standard data compilations. I am not saying, that G2
 does not produce "good" numbers in most of the cases. However,
 examples
 documented (L. Radom) where even G2 energies are quite unreasonable.
 This could be atributed to the multireference character of the WF,
 wrong orbitals (T1 not close to 0), and/or to a simple failure of the QCISD
 method. The QCI is in the limit (i.e. T1=O) equal to the
 coupled cluster method (CC). For T1~0 the CC method is more stable than
 QCI. However, one should always check the T1 and T2 amplitudes and the
 T1 norm. Recently, we have investigated one of the cases where QCI gives
 unreasonable results while CC is still OK. (JCP 106 (1997) 7185) These
 cases demonstrate themself in big T1 (T2) amlitudes (and positive energy
 contributions of the pertubative triples in QCI). The QCISD gives seemingly
 correct results for the structure, energy, freqs etc. but the QCISD density
 (dipole moments!!) are qualitativelly wrong. QCISD(T) breaks down completely.
 In most of the QC programs the QCISD ist only marginaly more expensive
 than the corresponding CCSD and thus I do not see any reason for using
 the QCI method. In addition, in cases where the WF is not a strict single
 reference one the convergency of CCSD is superior to QCISD.
 Jan Hrusak