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 a simple quadratic detector that is optimal for independent and identically distributed (i.i.d.) signals. Using the large deviations principle, the performance of this suboptimal detector is compared with the optimal detector for correlated signals via the asymptotic relative efficiency (ARE). 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, suggesting that simple detectors are highly effective in high-SNR regimes.
Introduction
In modern signal processing applications, such as wireless sensor networks, there is a constant tension between detection accuracy and computational efficiency. Detecting a stochastic Gaussian signal optimally requires a quadratic detector that accounts for the complete correlation structure of the signal across all samples. While mathematically superior, this approach is computationally demanding, scaling with $O(n^2)$ complexity.
A common alternative is the "simple" quadratic detector, which assumes the signal is independent and identically distributed (i.i.d.). This reduces complexity to $O(n)$, making it highly attractive for low-power nodes. This paper quantifies the performance trade-off of this simplification using the Large Deviations Principle (LDP), specifically asking: at what point does the computational complexity of the optimal detector cease to provide a meaningful advantage?
Methodology
The study employs the Large Deviations Principle (LDP) to analyze the "error exponent"—the rate at which detection error probability decays as the number of samples ($n$) increases. Three detector architectures are compared:
- Optimal Detector: Utilizes the full signal covariance matrix.
- Simple Quadratic Detector: Assumes an i.i.d. signal, ignoring correlations.
- Banded-Quadratic Detector: A middle-ground approach considering only local correlations (bandwidth $m$).
The primary performance metric is Asymptotic Relative Efficiency (ARE), defined as the ratio of sample sizes required by two detectors to achieve identical error performance in the large-sample limit. The analysis focuses on standard stochastic models like Gauss-Markov signals, where correlation decays exponentially.
Performance Analysis and Results
The core finding of this research is the "High SNR Equalizer" effect. As the Signal-to-Noise Ratio (SNR) increases, the performance gap between the simple and optimal detectors narrows significantly.
As shown in Figure 1, the error exponent for the simple detector is generally lower than that of the optimal detector, particularly as signal correlation (parameter 'a') increases. The simple detector's error rate is primarily driven by "miss" probabilities, as it is effectively optimized for false alarm events regardless of actual signal correlation.
However, the impact of SNR is transformative. The authors provide a formal proof that the ARE of the simple quadratic detector relative to the optimal detector converges to unity for any bounded signal spectrum as SNR goes to infinity.
Figure 2 illustrates this convergence. At high SNR (e.g., 30 dB), the ARE remains near 1.0 across almost the entire range of correlation strengths. This implies that in high-SNR environments, the specific correlation structure of the signal becomes negligible for detection purposes, and the $O(n)$ simple detector performs nearly as well as the $O(n^2)$ optimal detector.
Conclusion
This research demonstrates that complex signal processing is often unnecessary in high-SNR regimes. For a wide range of practical scenarios, replacing an $O(n^2)$ optimal detector with an $O(n)$ simple quadratic detector results in negligible loss in detection accuracy. Limitations remain in low-SNR environments, where the simple detector's performance degrades significantly if the signal is highly correlated. Future work may extend these findings to non-stationary signals or non-Gaussian noise environments where parameters must be estimated.
Summary updated