NEW CUMULANT-BASED APPROACHES FOR NON-GAUSSIAN TIME-VARYING AR MODELS

In this paper, we present two new cumulant-based methods for time-vary ing AR parameters estimation: a batch-type evolutive method and an ada ptive gradient-type algorithm. The evaluation of these techniques is p erformed through simulations on synthetic non-Gaussian signals contami nated by an additive, zero-mean, white Gaussian noise. We compare them to their autocorrelation-based counterparts. The obtained results sho w, using an appropriate criterion, the superiority of our cumulant-bas ed evolutive method over both its autocorrelation-based version and th e proposed cumulant-based gradient-type algorithm at the expense of mo re computational complexity.