Posterior variance predictions in sparse Bayesian learning under approximate inference techniques