Tensor Decomposed Distinguishable Cluster. I. Triples Decomposition

Abstract: We present a cost-reduced approach for the distinguishable cluster approximation to coupled cluster with singles, doubles and iterative triples (DC-CCSDT) based on a tensor decomposition of the triples amplitudes. The triples amplitudes and residuals are processed in the singular-value-decomposition (SVD) basis. Truncation of the SVD basis according to the values of the singular values together with the density fitting (or Cholesky) factorization of the electron repulsion integrals reduces the scaling of the method to $N^6$, and the DC approximation removes the most expensive terms of the SVD triples residuals and at the same time improves the accuracy of the method. The SVD basis vectors for the triples are obtained from the approximate CC3 triples density matrices constructed in an intermediate SVD basis of doubles amplitudes. This allows us to avoid steps that scale higher than $N^6$ altogether. Tests against DC-CCSDT and CCSDT(Q) on a benchmark set of chemical reactions with closed-shell molecules demonstrate that the SVD-error is very small already with moderate truncation thresholds, especially so when using a CCSD(T) energy correction. Tests on alkane chains demonstrated that the SVD-error grows linearly with system size confirming the size extensivity of SVD-DC-CCSDT within a chosen truncation threshold.