Abstract
Recently, the decentralized baseband processing (DBP) paradigm and relevant detection methods have been proposed to enable extremely large-scale massive multiple-input multiple-output (MIMO) technology. Under the DBP architecture, base station antennas are divided into several independent clusters, each connected to a local computing fabric. However, current detection methods tailored to DBP only consider ideal white Gaussian noise scenarios, while in practice, the noise is often colored due to interference from neighboring cells. This paper addresses the challenge of decentralized linear minimum mean-square error (LMMSE) detection under colored noise scenarios for both star and daisy chain DBP architectures. We propose decentralized equalizers based on dimensionality reduction for star architectures and an optimal decentralized equalizer using the block coordinate descent (BCD) method for daisy chain architectures. Simulation results demonstrate that our proposed methods achieve excellent detection performance with significantly reduced communication bandwidth.
Introduction: The Front-haul Bottleneck
Massive MIMO is a cornerstone of 5G and 6G, utilizing hundreds of antennas to serve multiple users simultaneously. However, traditional Centralized Baseband Processing (CBP) faces a significant "front-haul bottleneck": transferring raw data from hundreds of antennas to a single central processor requires throughput that can exceed 1 Tbps. Decentralized Baseband Processing (DBP) offers a solution by partitioning antennas into clusters, each managed by a local Distributed Unit (DU).
The Colored Noise Challenge
While DBP works well in ideal scenarios, most existing methods assume "white noise," where interference is random and uncorrelated across antennas. In real-world deployments, noise is often "colored" due to interference from neighboring cells, resulting in a complex, non-diagonal noise covariance matrix.
Calculating an optimal LMMSE equalizer with colored noise typically requires every DU to share all its noise samples, which creates a massive data explosion and defeats the purpose of decentralization. As shown in the figure above, simply computing local noise covariance matrices ($\mathbf{R}_{cc}$) results in a block-diagonal approximation that ignores the off-diagonal correlations ($\mathbf{R}_{ij}$) present in the true global noise covariance matrix $\mathbf{R}$.
Proposed Decentralized Architectures
The research proposes solutions for two primary network topologies:
1. Star Architecture
In a star topology, DUs are connected to a Central Unit (CU). The authors introduce two Dimensionality Reduction (DR) methods:
- sDR-MMSE (Superimposed): DUs compress local information into a low-dimensional space, and the CU sums the results.
- cDR-MMSE (Concatenated): The CU concatenates compressed data from DUs, offering higher performance at the cost of increased complexity.
A key finding is that a specific compression matrix ($\mathbf{H}^H \mathbf{R}^{-1}$) is "lossless," allowing for significant data reduction without sacrificing LMMSE equalization accuracy.
2. Daisy Chain Architecture
For networks without a central hub, a daisy chain architecture is employed where DUs are connected in sequence.
To optimize this architecture, the authors developed:
- BCD-MMSE: An iterative algorithm based on Block Coordinate Descent. Each DU updates its portion of the solution and passes a small set of "communication variables" to the next DU.
- Low-Rank Decomposition (LRD): By exploiting the fact that interference usually originates from a few sources, the noise covariance is treated as a low-rank matrix. This allows replacing thousands of noise samples with a few representative vectors, drastically saving bandwidth.
Key Insights and Results
The proposed methods achieve near-centralized performance while maintaining communication overhead independent of the number of base station antennas ($M$). Instead, overhead depends on the number of users ($K$) and interference sources ($r$), ensuring scalability as antenna counts grow.
Simulation results using the high-fidelity QuaDRiGa channel model show that the BCD-MMSE algorithm achieves a Symbol Error Rate (SER) nearly identical to centralized LMMSE. Furthermore, the BCD approach converges rapidly, requiring only 2 to 3 iterations to reach near-optimal performance, making it highly practical for real-time hardware implementation.
Conclusion
By recognizing that colored noise is structured (low-rank), this work transforms a mathematical challenge into a data-compression advantage. The proposed decentralized equalizers enable massive MIMO systems to handle interference-heavy environments efficiently, overcoming the front-haul bottleneck while maintaining high detection accuracy.
Summary updated