Cramer-Rao lower bounds for the discrete-time nonlinear state estimation problem are treated. The Cramer-Rao bound for the mean-square error matrix of a state estimate is particularly important for quality evaluation of nonlinear state estimators as it represents a limit of cognizability of the state. Recursive relations for filtering, predictive, and smoothing Cramer-Rao bounds are derived to establish a unifying framework for several previously published derivation procedures and results. Lower bounds for systems with unknown parameters are newly provided. Computation of filtering, predictive, and smoothing Cramer-Rao bounds, their mutual comparison and utilization for quality evaluation of some nonlinear filters are shown in numerical examples.