Abstract
In massive MIMO Frequency-Division Duplex (FDD) systems, acquiring accurate Channel State Information (CSI) is a significant challenge due to the high overhead of pilot signals. This paper examines the asymptotic performance of Zero-Forcing (ZF) precoding when the number of pilots is less than the number of base station antennas. We present a counter-intuitive finding: simple Least Squares (LS) channel estimation can outperform the statistically optimal Linear Minimum Mean-Square Error (LMMSE) estimator in terms of achievable sum rate at high signal-to-noise ratios (SNR). While LMMSE minimizes estimation error, it introduces a bias that prevents perfect interference nulling, leading to a performance saturation. Conversely, LS-based precoding asymptotically eliminates inter-user interference, provided the pilot count exceeds the number of users.
Introduction
Massive MIMO is a key technology for modern wireless communications, offering significant gains in spectral efficiency. However, in FDD mode, the base station (BS) requires CSI feedback from users. For a system with $M$ antennas, traditional estimation requires at least $M$ pilots, which is often impractical. When fewer pilots ($T_{dl} < M$) are used, the channel observations are undersampled. This study investigates whether the standard goal of minimizing Mean Square Error (MSE) through LMMSE estimation translates to optimal system performance when using ZF precoding.
The "Better Estimate" Fallacy
It is widely accepted that LMMSE provides a superior channel estimate in terms of MSE. As shown in Figure 1, LMMSE consistently maintains a lower MSE than LS across a wide range of transmit powers. This is because LMMSE utilizes spatial covariance matrices to filter out noise and interference, whereas LS is a simple projection of the received signal onto pilot sequences.
However, a lower MSE does not necessarily imply a higher data rate. The LMMSE estimator's reliance on channel statistics introduces a statistical bias. In the high-SNR regime, this bias prevents the ZF precoder from perfectly orthogonalizing the user channels. The resulting residual interference eventually dominates the signal, causing the sum rate to plateau.
Results and Asymptotic Performance
The core insight of this work is that LS-based ZF precoding is "interference-blind" in a beneficial way. The estimation errors in LS reside in a specific subspace that is naturally nulled by the ZF operation. Consequently, as transmit power increases, the interference term vanishes for LS but remains a non-zero constant for LMMSE.
As illustrated in Figure 2, for a system with $M=32$ antennas and $K=5$ users using only $T_{dl}=16$ pilots, the LS-based ZF sum rate continues to grow linearly with SNR, matching the slope of a genie-aided system with perfect channel knowledge. In contrast, the LMMSE-based system hits a rate floor above 20 dB. This demonstrates that LS-based precoding can achieve full degrees of freedom even with limited pilot overhead, provided the number of pilots is at least equal to the number of users ($T_{dl} \ge K$).
Conclusion
This research highlights a critical trade-off in massive MIMO precoding. While LMMSE is preferable at low SNR due to its noise-reduction capabilities, LS estimation is superior at high SNR for interference-limited scenarios. LS is not only more robust in these regimes but also computationally simpler as it avoids the need for complex covariance matrix estimation. These findings suggest that future "interference-aware" estimators should prioritize nulling accuracy over raw MSE to optimize performance in high-power massive MIMO systems.
Summary updated