Efficient initialization of dynamical mean-field theory

Dynamical mean-field theory (DMFT) is one of the leading tools for studying strongly correlated systems. However, each DMFT step can be computationally demanding and many iterations are required to achieve convergence. In this work, we initialize the DMFT self-consistent loop with the solutions of cheaper and more approximate methods in order to accelerate convergence. The different starting points we compare are the non-interacting, Hartree-Fock, and Hubbard-I solutions, as well as the rotationally invariant slave bosons (RISB) quantum embedding method and its "ghost" generalization (g-RISB). We find that g-RISB significantly outperforms the others, yielding accurate energies, quasiparticle weights, fillings, chemical potentials and spectral functions after a single DMFT iteration, everywhere except precisely at the metal-insulator transition. Our work suggests that initializing a fully self-consistent DMFT calculation with g-RISB significantly reduces the time to converge in all parameter regimes.