Consistency of the likelihood depth estimator for the correlation coefficient (Q2442676)

In statistics there is the notion of data depth. This notion, together with the likelihood function, brings the concept of likelihood depth. Estimators maximizing the likelihood depth turn out to be a robust counterpart to the usual maximum likelihood estimators. It has been shown that maximum likelihood depth estimators exhibit high robustness against contamination with other distributions. In [J. Stat. Plann. Inference 142, No. 9, 2501--2517 (2012; Zbl 1253.62042)] the authors studied the breakdown point and the consistency of estimators and tests based on the general depth notion. Such consistency is proven under the hypothesis of uniform convergence of the depth measure. In another previous paper the authors studied robust estimation for copulae parameters. For the particular case of a Gaussian copula, they developed a robust estimator for the correlation \(\rho\) if the bivariate normal distribution is considered i.e., using normal margins. The reviewed paper provides a proof for the uniform convergence of the depth measure and, hence, the consistency of the maximum likelihood depth estimator for the correlation coefficient of a bivariate normal distribution \(\rho\) if \(\rho \neq 0\). The uniform convergence of the depth measure is proven using the Vapnik-Červonenkis theorem, an extension of the Glivenko-Cantelli theorem. For the bivariate normal distribution almost everything is computable, so that, apparently, showing consistency of a related estimator sounds to be an easy task. However, the paper shows that this is not the case for the maximum likelihood depth estimator of \(\rho\). The proof shows up tricky, requiring elaborated tools as the VC classes.