Massive MIMO with imperfect channel covariance information

Emil Björnson, Luca Sanguinetti, Mérouane Debbah

Introduction

Massive MIMO systems rely on base stations (BSs) equipped with large antenna arrays to serve multiple users simultaneously. Optimal performance, typically achieved via Minimum Mean Squared Error (MMSE) estimation, requires precise knowledge of the channel covariance matrix—a statistical representation of signal correlations across the array. However, in practical deployments, these matrices are high-dimensional and must be estimated from limited pilot signals, leading to a "curse of dimensionality." This paper investigates the impact of imperfect covariance information and proposes a novel estimation scheme to bridge the performance gap.

Methodology: The "Via Q" Approach

The authors propose a two-stage estimation framework to overcome the limitations of standard sample covariance methods:

  1. Regularized Sample Covariance: To handle cases where observations are fewer than antennas, a shrinkage method is applied, combining the sample covariance with a diagonal matrix to ensure stability and full-rank estimates.
  2. "Via Q" Estimation: Rather than estimating a user's covariance directly from isolated pilots ("R direct"), the "Via Q" method estimates the total received pilot covariance ($Q$) first. By subtracting known interference statistics, it derives a cleaner estimate of the desired user's covariance ($R$), effectively leveraging a larger pool of observations.
  3. Capacity Bounds: The study introduces a "use-and-then-forget" capacity lower bound to evaluate spectral efficiency (SE) under imperfect covariance knowledge, allowing for fair comparisons between combining schemes like Maximum Ratio Combining (MRC) and Regularized Zero-Forcing (RZF).

Simulation Environment

The performance is evaluated using a standard theoretical benchmark:

  • Model: A 7-cell hexagonal layout with 100 antennas per BS and 10 users per cell.
  • Channel: The "One-Ring" spatial correlation model with 3GPP-inspired pathloss.

Performance Analysis

The proposed "Via Q" method demonstrates significant improvements in both estimation accuracy and system throughput compared to existing baselines.

Normalized MSE comparison showing the superior accuracy of the "Via Q" method over the "R direct" baseline as the number of extra pilots increases.

As shown in the Mean Squared Error (MSE) analysis, the "Via Q" approach consistently achieves lower error than the "R direct" method. Even with a small number of extra pilots, it significantly outperforms the interference-unaware Least Squares (LS) baseline, which requires no covariance knowledge but suffers from high error.

Average sum spectral efficiency vs. pilot overhead, illustrating that the "Via Q" method reaches 95% of ideal performance with significantly less overhead than previous methods.

In terms of spectral efficiency, the "Via Q" method reaches 95% of the ideal (perfect covariance) performance with only 100 extra pilots. In contrast, the older "R direct" method requires 250 pilots to reach the same threshold. The gains are particularly pronounced for RZF combining, which relies heavily on accurate covariance for effective interference suppression.

Conclusion

This research demonstrates that while perfect covariance knowledge is unattainable, Massive MIMO systems can still achieve near-optimal performance using intelligent statistical estimation. The "Via Q" method provides a robust path forward by maximizing the utility of available pilot data. Future work will need to address challenges such as pilot contamination in extremely dense networks and rapid statistical changes in high-mobility scenarios.

Summary updated

BibTeX
@inproceedings{bjornson2016massive,
  author = {Emil Björnson and Luca Sanguinetti and Mérouane Debbah},
  title = {{Massive MIMO with imperfect channel covariance information}},
  booktitle = {ACSSC},
  year = {2016},
  doi = {10.1109/acssc.2016.7869195},
}
Figure