Stochastic Tensor Contraction for Efficient MP2 Exchange

Abstract: Second-order Moller--Plesset perturbation theory (MP2) is one of the simplest correlated wave-function methods, but its conventional $O(N<sup>5)$ cost limits its application to large systems. Stochastic tensor contraction (STC) has recently appeared as a general technique to evaluate high-order tensor contractions in quantum chemistry. Here, we apply STC to the exchange contribution of Laplace-transformed MP2, which is the source of $O(N<sup>5)$ scaling in the formulation. The resulting STC exchange algorithm has an $O(N<sup>3)$ deterministic setup cost and an $O(N<sup>2)$ stochastic cost at fixed absolute error. The estimator is unbiased and provides a way to specify the target stochastic error by estimating the number of required samples before the full calculation. We implement the algorithm using a hybrid deterministic--stochastic evaluation strategy, with grouped index sampling, to reduce the computational prefactor. Over a range of benchmark molecules containing up to $\sim 7000$ basis functions, the stochastic exchange evaluation takes as little as $1/270$ of the time of a complete DF-MP2 calculation on the same system. Within the Laplace-transformed formulation, the $O(N<sup>4)$ scaling direct contribution is thus the only significant cost. Our results further substantiate the power of STC to serve as a general tensor-contraction engine for quantum chemistry.