SUMMARYTwo important goals of high-dimensional modeling are prediction and variable selection. In this article, we consider regularization with combined L 1 and concave penalties, and study the sampling properties of the global optimum of the suggested method in ultra-high dimensional settings. The L 1 -penalty provides the minimum regularization needed for removing noise variables in order to achieve oracle prediction risk, while concave penalty imposes additional regularization to control model sparsity. In the linear model setting, we prove that the global optimum of our method enjoys the same oracle inequalities as the lasso estimator and admits an explicit bound on the false sign rate, which can be asymptotically vanishing. Moreover, we establish oracle risk inequalities for the method and the sampling properties of computable solutions. Numerical studies suggest that our method yields more stable estimates than using a concave penalty alone.Some key words: Concave penalty; Global optimum; Lasso penalty; Prediction and variable selection.1. INTRODUCTION Prediction and variable selection are two important goals in many contemporary large-scale problems. Many regularization methods in the context of penalized empirical risk minimization have been proposed to select important covariates. See, for example, Fan & Lv (2010) for a review of some recent developments in high-dimensional variable selection. Penalized empirical risk minimization has two components: empirical risk for a chosen loss function for prediction, and a penalty function on the magnitude of parameters for reducing model complexity. The loss function is often chosen to be convex. The inclusion of the regularization term helps prevent overfitting when the number of covariates p is comparable to or exceeds the number of observations n.Generally speaking, two classes of penalty functions have been proposed in the literature: convex ones and concave ones. When a convex penalty such as the lasso penalty (Tibshirani, 1996) is used, the resulting estimator is a well-defined global optimizer. For the properties of L 1 -regularization methods, see, for example, Chen et al. (1999), Efron et al. (2004), Zou (2006), Candès & Tao (2007, Rosset & Zhu (2007), and Bickel et al. (2009). In particular, Bickel et al. (2009 proved that using the L 1 -penalty leads to estimators satisfying the oracle inequalities under the prediction loss and L q -loss, with 1 ≤ q ≤ 2, in high-dimensional nonparametric regression models. An oracle inequality means that with an overwhelming probability, the loss of the regularized estimator is within a logarithmic factor, a power of log p, of that of the oracle estimator, with the power depending on the chosen estimation loss. Despite these nice properties,