Outlier detection through mixtures with an improper component

The  paper  investigates the use of a finite  mixture model with an additional uniform density for outlier detection and robust estimation.  The main contribution of this paper lies in the  analysis of the properties of the improper component and the introduction of a modified EM algorithm  which, beyond providing the maximum likelihood estimates of the mixture parameters, endogenously provides a numerical value for the density of the uniform distribution used for the improper component. The mixing proportion of outliers may be known or unknown.  Applications to robust estimation and outlier detection will be discussed with particular attention to the normal mixture case.