Optimal and Suboptimal Detection of Gaussian Signals in Noise: Asymptotic Relative Efficiency

TL;DR

This paper analyzes the performance of simple quadratic detectors versus optimal detectors for Gaussian signals in noise, showing that at high SNR, the simple detector performs nearly as well as the optimal one across various signal correlations.

Contribution

It introduces the asymptotic relative efficiency framework to compare simple and optimal detectors for Gaussian signals, highlighting conditions where simple detectors are nearly optimal.

Findings

Asymptotic efficiency approaches one at high SNR.

Simple quadratic detector performs well across many correlation values at high SNR.

Performance gap diminishes as SNR increases, regardless of signal correlation.

Abstract

The performance of Bayesian detection of Gaussian signals using noisy observations is investigated via the error exponent for the average error probability. Under unknown signal correlation structure or limited processing capability it is reasonable to use the simple quadratic detector that is optimal in the case of an independent and identically distributed (i.i.d.) signal. Using the large deviations principle, the performance of this detector (which is suboptimal for non-i.i.d. signals) is compared with that of the optimal detector for correlated signals via the asymptotic relative efficiency defined as the ratio between sample sizes of two detectors required for the same performance in the large-sample-size regime. The effects of SNR on the ARE are investigated. It is shown that the asymptotic efficiency of the simple quadratic detector relative to the optimal detector converges to one as the SNR increases without bound for any bounded spectrum, and that the simple quadratic detector performs as well as the optimal detector for a wide range of the correlation values at high SNR.