Line spectrum pairs in pattern recognition

The optimal Bayesian Classifier is often difficult to implement because of its complexity. For Gaussian parameters, the Bayes decision rule reduces to a simple centroid distance rule. However, the centroid distance rule fails for non-Gaussian parameters with non-convex probability density functions (p.d.f.). This paper studies some statistical properties of Line Spectrum Pairs (LSP). These statistical properties can be used to study the convexity of LSP point clusters in pattern recognition applications.