Approximate Gram-Matrix Interpolation for Wideband Massive MU-MIMO Systems

Charles Jeon, Zequn Li, Christoph Studer

TL;DR

This paper tackles the computational bottleneck of Gram-matrix computation in massive MU-MIMO systems, which is crucial for efficient data detection and precoding. The authors propose novel approximate interpolation methods that significantly reduce complexity while maintaining near-optimal performance, even in the presence of imperfect channel information.

What Problem Does This Paper Address?

Massive multi-user multiple-input multiple-output (MU-MIMO) systems, a cornerstone of 5G and beyond, promise significantly improved spectral efficiency by equipping base stations with hundreds of antennas to serve tens of users. However, many efficient data detection and precoding algorithms for these systems, particularly those operating in wideband scenarios using OFDM or SC-FDMA, rely on computing the Gram matrix ($G = H^H H$) for each active subcarrier. This computation is computationally prohibitive, especially in wideband systems with many subcarriers and large antenna arrays. Existing methods to mitigate this either ignore crucial practical aspects like imperfect channel state information (CSI) or don't scale well to massive MIMO. This paper aims to bridge this gap by developing computationally efficient Gram-matrix computation techniques that are robust to practical impairments.

Methods

The core contribution of this paper lies in approximate Gram-matrix interpolation. Building on the idea that Gram matrices are "smooth" across subcarriers in wideband systems, the authors propose two methods:

  • Oth Order Interpolation (Nearest Neighbor): This method computes the Gram matrix at a few selected "base points" and then assigns the Gram matrix of the nearest computed base point to all other subcarriers. This is computationally very cheap.
  • 1st Order Interpolation (Linear Interpolation): This method also computes Gram matrices at base points but then uses linear interpolation between adjacent base points to estimate the Gram matrix for subcarriers in between. This offers a better trade-off between complexity and accuracy.

Crucially, the authors provide analytical frameworks to characterize the approximation error (Mean Squared Error - MSE) for both methods, considering factors like channel delay spread, BS-antenna correlation, and imperfect CSI. They also analyze the computational complexity of their methods and compare them to brute-force computation and exact interpolation.

Datasets

Not discussed in the paper. The evaluation is based on simulations using synthetic channel models (i.i.d. Rayleigh fading) and a realistic channel model (QuaDRiGa).

Key Insights and Ideas

  • Channel Hardening is Key: The paper leverages the "channel hardening" phenomenon in massive MU-MIMO, which makes the channel more predictable and less susceptible to deep fades, enabling effective interpolation even with undersampling.
  • Interpolation Outperforms Exact Methods Under Imperfection: While exact interpolation is ideal with perfect CSI, the proposed approximate methods show superior robustness and better error-rate performance when CSI is imperfect or BS antennas are correlated. This is because exact interpolation matrices can become ill-conditioned under these conditions.
  • 1st Order Interpolation Offers a Sweet Spot: The 1st order interpolation method strikes a good balance, achieving performance very close to brute-force computation with significantly reduced complexity, especially compared to exact interpolation.
  • Complexity Scales Favorably: The proposed methods drastically reduce the computational burden, making massive MU-MIMO systems more practical to implement.

Key Results

  • Significant Complexity Reduction: The proposed Oth and 1st order interpolation methods achieve substantial complexity reductions compared to brute-force computation. For instance, with a small fraction of base points, the complexity is less than half of the brute-force approach.
  • Near-Optimal Error-Rate Performance: The approximate interpolation methods achieve error-rate performance very close to brute-force computation, especially when a sufficient number of base points are used.
  • Robustness to Imperfect CSI: In scenarios with imperfect CSI, the approximate methods demonstrate a significantly lower error floor and better resilience compared to exact interpolation, which suffers from ill-conditioned interpolation matrices.
  • 1st Order Interpolation Advantage: The 1st order interpolation consistently outperforms the Oth order method, especially for larger channel delay spreads, and approaches the performance of brute-force with only a fraction of its complexity. For example, it can achieve near brute-force performance with only 45% of the complexity.

Limitations and Future Work

  • Higher-Order Interpolation: The authors suggest that higher-order approximate interpolation schemes could offer even better performance at the cost of increased complexity. Analyzing these is left for future work.
  • Realistic Channel Models: While the QuaDRiGa channel model was used, further investigation into more realistic scenarios, including frequency and timing offsets, and more sophisticated receive-side correlation models, is suggested.
  • Generalization to Other Algorithms: The paper focuses on Gram-matrix interpolation for data detection and precoding. Extending these approximate interpolation concepts to other matrix operations like matrix inversion and QR decomposition in massive MIMO systems is also a promising avenue.
  • Analytical Bounds: While the paper provides analytical bounds, further refinement of these bounds, particularly for the 1st order interpolation, is mentioned as a potential area for future research.

Summary updated

BibTeX
@article{jeon2020approximate,
  author = {Charles Jeon and Zequn Li and Christoph Studer},
  title = {{Approximate Gram-Matrix Interpolation for Wideband Massive MU-MIMO Systems}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2020},
  doi = {10.1109/tvt.2020.2983025},
}
Figure