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

In this paper, we study the generalization ability (GA)---the ability of a model to predict outcomes in new samples from the same population---of the extremum estimators. By adapting the classical concentration inequalities, we propose upper bounds for the empirical out-of-sample prediction error for extremum estimators, which is a function of the in-sample error, the severity of heavy tails, the sample size of in-sample data and model complexity. The error bounds not only serve to measure GA, but also to illustrate the trade-off between in-sample and out-of-sample fit, which is connected to the traditional bias-variance trade-off. Moreover, the bounds also reveal that the hyperparameter K, the number of folds in $K$-fold cross-validation, cause the bias-variance trade-off for cross-validation error, which offers a route to hyperparameter optimization in terms of GA. As a direct application of GA analysis, we implement the new upper bounds in penalized regression estimates for both n>p and n<p cases. We show that the L2 norm difference between penalized and un-penalized regression estimates can be directly explained by the GA of the regression estimates and the GA of empirical moment conditions. Finally, we show that all the penalized regression estimates are L2 consistent based on GA analysis.