Approaching the Full Configuration Interaction Ground State from an Arbitrary Wavefunction with Gradient Descent and Quasi-Newton Algorithms

TL;DR

This paper introduces gradient descent and quasi-Newton algorithms to efficiently approximate the FCI ground state energy from arbitrary initial states, enabling larger system calculations without storing intermediate wavefunctions.

Contribution

It presents a novel optimization approach that evaluates energies via expectation values, allowing larger system analysis compared to traditional FCI methods.

Findings

Efficient energy evaluation along optimization paths.

Ability to handle larger systems than standard FCI.

Application with non-orthogonal determinant reference states.

Abstract

We consider gradient descent and quasi-Newton algorithms to optimize the full configuration interaction (FCI) ground state wavefunction starting from an arbitrary reference state $|0 \rangle$. We show that the energies obtained along the optimization path can be evaluated in terms of expectation values of $|0 \rangle$, thus avoiding explicit storage of intermediate wavefunctions. This allows us to find the energies after the first few steps of the FCI algorithm for systems much larger than what standard, deterministic FCI codes can handle at present. We show an application of the algorithm with reference wavefunctions constructed as linear combinations of non-orthogonal determinants.