Bath optimization in the cellular dynamical mean-field theory

In the cellular dynamical mean-field theory (CDMFT), a strongly correlated system is represented by a small cluster of correlated sites, coupled to an adjustable bath of uncorrelated sites simulating the cluster's environment; the parameters governing the bath are set by a self-consistency condition involving the local Green's function and the lattice electron dispersion. Solving the cluster problem with an exact diagonalization method is only practical for small bath sizes (eight sites). In that case the self-consistency condition cannot be exactly satisfied and is replaced by a minimization procedure. There is some freedom in the definition of the "merit function" to optimize. We use Potthoff's self-energy functional approach on the one- and two-dimensional Hubbard models to gain insight into the best choice for this merit function. We argue that several merit functions should be used and preference given to the one that leads to the smallest gradient of the Potthoff self-energy functional. We propose a merit function weighted with the self-energy that seems to fit the Mott transition in two dimensions better than other merit functions.