The affine equivariant sign covariance matrix: Asymptotic behavior and efficiencies. (Q1426351)

Let \(X\) be a \(k\)-dimensional random vector with the covariance matrix \(\Sigma=P\Lambda P^T\), where \(P\) is the matrix with eigenvectors of \(\Sigma\) and \(\Lambda\) is a diagonal matrix with the corresponding eigenvalues \(\lambda_1,\dots,\lambda_k\) as diagonal elements. This spectral decomposition of the covariance matrix may be represented in the form \(\Sigma=\lambda P\Lambda^* P^T\), where \(\lambda=(\lambda_1\cdots\lambda_k)^{1/k}\) is the geometric mean of the eigenvalues and \(\Lambda=\lambda\Lambda^*\). The terms scale, shape and orientation are used for the items \(\lambda\), \(\Lambda^*\) and \(P\). The authors consider the affine equivariant sign covariance matrix (SCM) which can be used to estimate the shape \(\Lambda^*\) and the orientation \(P\) of the covariance matrix \(\Sigma\). Under a specified elliptic model distribution a consistent estimate of \(\Sigma\) may be obtained. The SCM estimator [proposed by \textit{S. Visuri, V. Koivunen} and \textit{H. Oja}, J. Stat. Plann. Inference 91, 557--575 (2000; Zbl 0965.62049)]) is based on the concept of affine equivariant signs. For a review of multivariate signs and ranks see \textit{H. Oja}, Scand. J. Stat. 26, 319--343 (1999; Zbl 0938.62063). The main contribution of the paper is the derivation of the influence function and limiting distribution of the SCM. The asymptotic behaviour of the eigenvectors and standardized eigenvalues of the SCM are derived. It is shown that the proposed estimates are highly efficient in the multivariate normal case and perform better than estimates based on the sample covariance matrix for heavy-tailed distributions. A simulation study confirmed these findings for finite-sample efficiencies.