Order Determination in Second-Order Source Separation Models Using Data Augmentation

Abstract

We propose a robust estimator for the number of latent components in an internal noise model within the second-order source separation (SOS) framework. Our approach utilizes a data augmentation strategy in conjunction with the robust SOS approach eSAM-AMUSE, which combines information from eigenvalues and variations of eigenvectors of eSAM-AMUSE. The resulting dimension estimate can be visualized using a ladle plot. Through a simulation study, we demonstrate the superior properties of the new estimator, which outperforms the bootstrap-based AMUSEladle estimator.