Massively parallel exact diagonalization with symmetries and minimal communication

Scalable, distributed-memory approaches to exact diagonalization require that functions of the Hamiltonian can be applied efficiently to the wavefunction. We present a method that requires a single collective (all-to-all) communication per Hamiltonian application while making use of conserved quantities. The approach is applicable to a wide class of quantum lattice systems with U(1) symmetries and finite-ranged interactions as well as multi-orbital impurity problems. It is less communication intensive than existing approaches, and its parallel scaling is entirely determined by that of an all-to-all communication on the computing platform.