Robust canonical correlations: a comparative study (Q2488393): Difference between revisions

If \(x\) and \(y\) are random vectors, then the first pair of canonical vectors is \[ (\alpha_1,\beta_1)=\arg\max_{a,b}\,\text{Corr}\,(a^Tx,b^T,y). \] The authors consider some robust versions of canonical vectors (correlations) based on the following two main ideas. The first one is to use some robust measure of bivariate dependence instead of Pearson's correlation. It can be Spearman's rank correlation, a bivariate minimum covariance determinant estimator, or a bivariate M-estimate for the correlation. The second idea is to use a robust version of the alternating regression technique. The higher order canonical vectors can be also derived using these approaches. Different estimates are compared via simulations. Applications to the independence testing are considered.