Finite-sample and asymptotic analysis of generalization ability with an application to penalized regression