An efficient data compression method for the Davidson subspace diagonalization scheme

A flexible and efficient compression scheme for the expansion and product vectors Hamiltonian matrix times expansion vectors is presented within the Davidson diagonalization method. Our approach is based on an error analysis of the energy in terms of the aforementioned vectors and on a compression scheme for representing floating point numbers with a variable length mantissa. For a selection of typical quantum chemical test cases total saving factors of up to ten are reported. The method is expected to work especially well for extended multireference CI and full CI cases. As a general outcome of our analysis we obtain limits of possible sizes of a CI expansion within the Davidson procedure in relation to the energy and the desired accuracy of the energy assuming the usual IEEE floating point standard.