A new version of the multireference averaged coupled-pair functional (MR-ACPF-2)

The averaged coupled-pair functional (ACPF), as found in R. J. Gdanitz and R. Ahlrichs, Chem. Phys. Lett. 143 (1988), is probably the most successful method to approximate full configuration interaction (CI) on the multireference (MR) level of theory. However, ACPF has a tendency to slightly overestimate the effect of higher than double substitutions, which, when the zeroth-order wavefunction is of poor quality, may deteriorate the accuracy or even create instabilities. Since the properties of the ACPF and similar methods have apparently not always been correctly described in the literature, we repeat the derivation of this method in some detail. We analyze the connection between the (original) ACPF and the similar averaged quadratic coupled cluster (AQCC) method (which may be regarded as a damped ACPF), on one hand, and the different versions (0-3) of the coupled electron-pair approximation (CEPA), on the other hand. We find that ACPF and AQCC may be regarded as CEPA-1, respectively, CEPA-3, where the shifts of the Hamiltonian are substituted by a single averaged shift As CEPA-3 considerably underestimates correlation effects, AQCC shows the same behavior. However, when the zeroth-order wavefunction is of poor quality AQCC may be more stable and thus more accurate than ACPF By analyzing the role of the single substitutions, we find that ACPF may especially overestimate their contribution to unlinked clusters like 1/2 (T) over cap (2)(1). We therefore propose a new version, called ACPF-2, where (in contrast to AQCC) only the renormalization. factor, g, that corresponds to the singles, is damped; i.e., we have (4/N)[1 - 1/(2N - 2)]. In the limit of a large number of electrons, N, this, factor becomes two times as large as in the (original) ACPF, where 2/N is used. In order to test the new ACPF-2 method, we perform numerous comparisons with full CI, as well as calculations including terms that are linear in the interelectronic distances, r(ij). We find that in difficult cases, ACPF-2 is of similar stability as is AQCC and is thus considerably more accurate.