Convergence rate estimation through subsampling