Influence functions for penalized M-estimators (Q2405153)

The classical definition of influence function for a regular M-estimator is useful in the assessment of relative influence of individual observations towards the value of an estimate and its asymptotic properties of an estimate. This definition is, however, difficult to apply to penalized M-estimators which are obtained by minimizing a regularized risk since the penalty function in the definition of the risk function may not be differentiable. In this paper, a new definition of influence function is proposed as the limit of the influence functions for a sequence of penalized M-estimators defined by a sequence of smooth penalty functions which approach the potentially non-smooth penalty function. It is shown that this limit is uniquely defined which is unrelated to the chosen limiting sequence. This definition has been used to derive the influence functions of penalized M-estimators defined by Lasso and group Lasso type penalties and extended two-stage penalized M-estimators such as those defined by adaptive Lasso type penalties.