Density-matrix renormalization using three classes of block states

An extension of the the density matrix renormalization group (DMRG) method is presented. Besides the two groups or classes of block states considered in White's formulation, the retained $m$ states and the neglected ones, we introduce an intermediate group of block states having the following $p$ largest eigenvalues $\lambda_i$ of the reduced density matrix: $\lambda_1 \ge >... \lambda_m \ge \lambda_{m+1}\ge ... \ge \lambda_{m+p}$. These states are taken into account when they contribute to intrablock transitions but are neglected when they participate in more delocalized interblock fluctuations. Applications to one-dimensional models (Heisenberg, Hubbard and dimerized tight-binding) show that in this way the involved computer resources can be reduced without significant loss of accuracy. The efficiency and accuracy of the method is analyzed by varying $m$ and $p$ and by comparison with standard DMRG calculations. A Hamiltonian-independent scheme for choosing $m$ and $p$ and for extrapolating to the limit where $m$ and $p$ are infinite is provided. Finally, an extension of the 3-classes approach is outlined, which incorporates the fluctuations between the $p$ states of different blocks as a self-consistent dressing of the block interactions among the retained $m$ states.