Impact of covariance mismatched training samples on constant false alarm rate detectors

TL;DR

This paper analyzes how covariance mismatches in training samples affect the performance of constant false alarm rate detectors in Gaussian noise, providing a statistical framework and mitigation strategies.

Contribution

It introduces a statistical representation of detector statistics under covariance mismatch and explores mitigation methods to reduce false alarm rate variations.

Findings

Covariance mismatch causes significant false alarm rate variations.

The paper provides a statistical model for detector statistics under mismatch.

Mitigation strategies can help stabilize false alarm rates.

Abstract

The framework of this paper is that of adaptive detection in Gaussian noise with unknown covariance matrix when the training samples do not share the same covariance matrix as the vector under test. We consider a class of constant false alarm rate detectors which depend on two statistics $(\beta,\ttilde)$ whose distribution is parameter-free in the case of no mismatch and we analyze the impact of covariance mismatched training samples. More precisely, we provide a statistical representation of these two variables for an arbitrary mismatch. We show that covariance mismatch induces significant variations of the probability of false alarm and we investigate a way to mitigate this effect.