An accelerated conjugate gradient algorithm to compute low-lying eigenvalues - A study for the Dirac operator in SU(2) lattice QCD

The low-lying eigenvalues of a (sparse) Hermitian matrix can be computed with controlled numerical errors by a conjugate gradient (CG) method. This CG algorithm is accelerated by alternating it with exact diagonalizations in the subspace spanned by the numerically computed eigenvectors. We study this combined algorithm in case of the Dirac operator with (dynamical) Wilson fermions in four-dimensional SU(2) gauge fields. The algorithm is numerically very stable and can be parallelized in an efficient way, On lattices of sizes 4(4)-16(4) an acceleration of the pure CG method by a factor of 4-8 is found.