An efficiency upper bound for inverse covariance estimation (Q2351738)

The author proves a non-existence result which is related to the problem of estimating elements of the inverse covariance matrix of a \(d\)-dimensional Gaussian random vector \(Y\) based on a sample of \(n\) independent and identically distributed random vectors \(Y_1, \hdots, Y_n\), where \(Y_1\) is in distribution equal to \(Y\). Let \(A\) denote the covariance matrix of \(Y\) and define the numbers \((\alpha_{i,j})_{1 \leq i, j \leq d}\) by \[ \alpha_{i, j} = \lim_{\epsilon \to 0} \mathbb{E}[Y_iY_j \; | \; |Y_k| < \epsilon, \forall k \not \in \{k_1, k_2\}], \] for all \(i, j \in \{k_1, k_2\}\). It is shown that for \(n < d/3\) there exists no function (estimator) which maps \((Y_1, \hdots, Y_n)\) to \(\{0, 1, 2\}\) such that, uniformly over all positive semi-definite \(A\), the probability that the rank of \(\begin{pmatrix} \alpha_{1, 1} & \alpha_{1, 2}\\ \alpha_{2, 1} & \alpha_{2, 2} \end{pmatrix}\) is estimated correctly exceeds \(0.9\). Some connections of this result to previous (positive and negative) results about the attainable precision of (inverse) covariance estimation are also drawn.