Bridging single- and multireference domains for electron correlation: spin-extended coupled electron pair approximation

TL;DR

This paper introduces a size-consistent, spin-projected coupled electron pair approximation that effectively handles static correlation and improves upon ECISD, with applications to molecular dissociation, splitting, and hyperfine interactions.

Contribution

It develops a size-consistent generalization of ECISD incorporating quadruples via coupled electron pair approximation, addressing spin-projection complexities.

Findings

Better results than ECISD in molecular dissociation and properties.

Approximate size-consistency due to symmetry projection.

Effective handling of static correlation in multireference systems.

Abstract

We propose a size-consistent generalization of the recently developed spin-extended configuration interaction with singles and doubles (ECISD), where a CI wave function is explicitly spin-projected. The size-consistent effect is effectively incorporated by treating quadruples within the formulation of coupled electron pair approximation. As in coupled-cluster theory, quadruple excitations are approximated by a disconnected product of double excitations. Despite its conceptual similarity to the standard single- and multireference analogues, such a generalization requires careful derivation, as the spin-projected CI space is non-orthogonal and overcomplete. Although our methods generally yield better results than ECISD, size-consistency is only approximately retained because the action of a symmetry-projection operator is size-inconsistent. In this work, we focus on simple models where exclusion-principle-violating terms, which eliminate undesired contributions to the correlation effects, are either completely neglected or averaged. These models possess an orbital-invariant energy functional that is to be minimized by diagonalizing an energy-shifted effective Hamiltonian within the singles and doubles manifold. This allows for a straightforward generalization of the ECISD analytical gradients needed to determine molecular properties and geometric optimization. Given the multireference nature of the spin-projected Hartree--Fock method, the proposed approaches are expected to handle static correlation, unlike single-reference analogues. We critically assess the performance of our methods using dissociation curves of molecules, singlet-triplet splitting gaps, hyperfine coupling constants, and the chromium dimer. The size-consistency and size-extensivity of the methods are also discussed.