A non-parameteric ML estimator with unknown channel order

A recently proposed non-parameteric maximum likelihood (NPML) channel estimator shows superior performance to the least square (LS) estimator in presence of non-Gaussian noise. The derivation of the NPML estimator assumed perfect knowledge of the channel order, which, however, does not comply with most applications. In this paper, first we study the effects of the inaccurate order assumption on the NPML estimator, and then show that the traditional order selection criteria like the AIC are unreliable to apply for the NPML estimator. Finally we propose a simple method to trace the channel order where the order selection and channel estimation are carried out simultaneously.