Cumulants as the Variables of Density Cumulant Theory: A Path to Hermitian Triples

TL;DR

This paper introduces OλDCT, a new density cumulant theory approach that overcomes previous limitations, accurately describes molecular dissociation, and outperforms existing methods with similar computational scaling.

Contribution

The paper presents a novel orbital-optimized density cumulant theory with a new parameterization that is free of near-zero denominators and more accurate than prior methods.

Findings

Solves issues of previous density cumulant theories.

Accurately describes H2 dissociation and size-extensivity.

Outperforms CCSD(T) and CCSDT with similar computational cost.

Abstract

We study the combination of orbital-optimized density cumulant theory and a new parameterization of the reduced density matrices in which the variables are the particle-hole cumulant elements. We call this combination O$\lambda$DCT. We find that this new ansatz solves problems identified in the previous unitary coupled cluster ansatz for density cumulant theory: the theory is now free of near-zero denominators between occupied and virtual blocks, can correctly describe the dissociation of H$_2$, and is rigorously size-extensive. In addition, the new ansatz has fewer terms than the previous unitary ansatz, and the optimal orbitals delivered by the exact theory are the natural orbitals. Numerical studies on systems amenable to full configuration interaction show that the amplitudes from the previous ODC-12 method approximate the exact amplitudes predicted by this ansatz. Studies on equilibrium properties of diatomic molecules show that even with the new ansatz, it is necessary to include triples to improve the accuracy of the method compared to orbital optimized linearized coupled cluster doubles. With a simple iterative triples correction, O$\lambda$DCT outperforms other orbital-optimized methods truncated at comparable levels in the amplitudes, as well as CCSD(T). By adding four more terms to the cumulant parameterization, O$\lambda$DCT outperforms CCSDT while having the same $\mathcal{O}(V^5 O^3)$ scaling.