High-dimensional robust precision matrix estimation: cellwise corruption under \(\epsilon \)-contamination (Q1753147): Difference between revisions

Consider the following cellwise contamination model. For each \(n\) and \(p\) in \(\{1,2,\dots\}\) and \(k\in \{1,\dots,n\}\), \(X_{k}\) is a real \(p\)-dimensional random vector defined by \[ X_{k}=(I-B_{k})Y_{k}+B_{k}Z_{k}, \] where \(B_{k}\), \(Y_{k}\) and \(Z_{k}\) are independent random fields, \(B_{k}\) is a diagonal \((p\times p)\)-matrix \[ B_{k}=\left[ \begin{matrix} B_{k1} & 0 & \cdots & 0 \\ 0 & B_{k2} & \cdots & 0 \\ \vdots & \vdots & \vdots & \vdots \\ 0 & 0 & \cdots & B_{kp} \end{matrix} \right] , \] where \(B_{k1}\), \(B_{k2}\) and \(B_{kp}\) are independent random variables with \[ P(B_{ki}=1)=\varepsilon_{i}, \quad i\in \{1,\ldots,p\}. \] In this paper it is assumed that \(Y_{k}\thicksim \mathcal{N}(\mu ,\Sigma^{\ast})\), and \(Z_{k}\thicksim \mathcal{H}^{\ast}\) (an unspecified distribution). The objective of the paper is the estimation of \(\Sigma^{\ast}\), and \(\Omega^{\ast}=(\Sigma^{\ast})^{-1}\). If the data are uncontaminated \[ \varepsilon_{i}=0, \quad i\in \{1,\ldots ,p\}, \tag{1} \] then the classical sample covariance estimator \(\widehat{\Sigma}= \left[ \widehat{\Sigma}_{ij}\right]_{1\leq i,j\leq p}\) defined as \[ \widehat{\Sigma}_{ij}=\frac{1}{n-1}\sum_{k=1}^{n}(X_{ki}- \overline{X}_{i})(X_{kj}-\overline{X}_{j}), \quad 1\leq i,j\leq p \] where \[ \overline{X}_{i}=\frac{1}{n}\sum_{k=1}^{n}X_{ki}, \quad 1\leq i\leq p \] is used. It is well known that if (1) is not met then \(\widehat{\Sigma}\) may be very inefficient. Several authors have proposed estimators of the covariance matrix under various ``pure'' model and contamination assumptions. Among them are the so-called GLasso and CLIMA. Recently, good properties of these estimators have been tested from the point of view of the ``breakdown point'' (maximum proportion of contamination that an estimator can admit). The authors of this paper propose new estimators modifying the aforementioned ones incorporating robust estimators of the covariance matrix. The results reported encourage to follow and extend the proposed estimators to other ``pure'' and contaminated models. A detailed and updated review of the ``state of the art'' in the subject is made. It is worth highlighting the list of notations that are used in the paper, which undoubtedly facilitates their reading. Although the theorems are enunciated and demonstrated with detail and rigor, the comments included in the work facilitate the understanding and application of them. I believe that the study of this work will be very useful for mathematicians and statisticians who are interested in the unavoidable topic of robust estimation in statistics.