Abstract
In frequency-selective channels, linear receivers offer significantly reduced complexity compared to maximum likelihood (ML) receivers, albeit at the cost of potential performance degradation. This degradation often manifests as a loss of inherent frequency diversity or reduced coding gain. This paper demonstrates that the minimum mean-square error (MMSE) symbol-by-symbol linear equalizer incurs no diversity loss compared to ML receivers. Specifically, for a channel with memory $\nu$, the MMSE equalizer achieves the full diversity order of $\nu+1$, whereas the zero-forcing (ZF) symbol-by-symbol linear equalizer consistently achieves a diversity order of only one.
Introduction
In broadband wireless communication, signals typically traverse multiple paths, resulting in frequency-selective fading and Inter-Symbol Interference (ISI). While ISI is traditionally viewed as a distortion to be mitigated, it also provides "frequency diversity"—multiple independent versions of the transmitted signal—which can be exploited to enhance link reliability.
The optimal method for managing ISI is Maximum Likelihood Sequence Detection (MLSD). However, the computational complexity of MLSD grows exponentially with the channel's memory and spectral efficiency, making it impractical for many high-speed applications. Consequently, engineers often turn to simpler linear equalizers, such as Zero-Forcing (ZF) and Minimum Mean-Square Error (MMSE) filters. Historically, it was unclear whether these lower-complexity receivers sacrificed the channel's inherent diversity. This research provides a rigorous mathematical characterization of the diversity order achieved by these linear equalizers.
Theoretical Framework and Methodology
The study employs a quasi-static ISI channel model with memory length $\nu$. By representing the system in the $D$-domain (transform domain), the authors analyze the signal-to-noise ratio (SNR) at the output of the equalizer. Diversity order is defined by the rate at which the outage probability (or error probability) decays as the SNR increases.
To prove the performance of the MMSE equalizer, the authors utilize Hardy’s Inequality, a sophisticated tool from mathematical analysis, to establish a lower bound on the $L_1$ norm of exponential sums. This allows for a precise bounding of the MMSE equalizer's performance. Furthermore, the authors use the Monotone Convergence Theorem to extend their proof from low data rates to all possible spectral efficiencies, demonstrating that diversity order is a fundamental property of the equalizer rather than the transmission rate.
In contrast, the analysis of the ZF equalizer focuses on the properties of Gaussian random variables. The authors demonstrate that the noise enhancement inherent in ZF equalization—caused by the inversion of the channel response—effectively nullifies any diversity gain beyond the first order.
Key Insights: MMSE vs. Zero-Forcing
The MMSE Advantage
The core finding is that MMSE equalizers achieve the full frequency diversity order of $\nu+1$. This success stems from the MMSE criterion's ability to strike an optimal balance between ISI suppression and noise enhancement. While MMSE equalizers do not match the "coding gain" of MLSD (meaning the performance curve is shifted to the right), they match the slope of the MLSD error curve, ensuring the same level of robustness against fading at high SNR.
The Failure of Zero-Forcing
Zero-forcing equalizers are shown to be fundamentally limited, achieving a diversity order of exactly 1, regardless of the number of channel taps. Because ZF equalizers attempt to completely eliminate ISI, they significantly amplify noise at frequencies where the channel exhibits "notches" (low gain). This noise enhancement is so severe that it prevents the receiver from benefiting from the multipath components of the channel.
Rate Independence
A significant contribution of this work is the proof that diversity order is independent of spectral efficiency. Whether the system is operating at 1 bit/sec/Hz or 4 bits/sec/Hz, the MMSE equalizer maintains its $\nu+1$ diversity order, while the ZF equalizer remains trapped at a diversity order of 1.
Validation and Conclusion
The theoretical proofs were validated using Monte Carlo simulations of synthetic Rayleigh fading channels with two-tap ($\nu=1$) and three-tap ($\nu=2$) configurations. The results confirmed that for MMSE equalizers, the Bit Error Probability (BER) curves steepened as $\nu$ increased, whereas the ZF curves maintained a constant, shallow slope regardless of the channel memory.
While this paper establishes a strong theoretical foundation for linear equalizers, it assumes infinite-length filters and additive white Gaussian noise (AWGN). Future research may explore the impact of finite impulse response (FIR) constraints and the performance of these equalizers in the presence of non-Gaussian interference or within Decision Feedback Equalizer (DFE) architectures.
Summary updated