Model Order Selection in Presence of Unknown Colored Noise
Farzad Talebi, Thomas G. Pratt · 2014
We consider the problem of detecting a number of complex sinusoids in unknown colored noise based on observations from an unstructured array. Based on extreme value theory, we first present a new model complexity penalty term for the log-likelihood function that outperforms both Minimum Description Length (MDL) and Akaike Information Criterion (AIC) in different array sizes. Second, we derive a new signal to noise ratio (SNR) threshold defining the breakdown point associated with the Maximum Likelihood (ML) estimator.