Cooperative access networks: Optimum fronthaul quantization in distributed Massive MIMO and cloud RAN

We consider cooperative radio access network architectures, especially distributed massive MIMO and Cloud RAN, considering their similarities and differences. We address in particular the major challenge posed to both by the implementation of a high capacity fronthaul network to link the distributed access points to the central processing unit, and consider the effect on uplink performance of quantization of received signals in order to limit fronthaul load. We use the Bussgang decomposition along with a new approach to MMSE estimation of both channel and data to provide the basis of our analysis.

💡 Research Summary

This paper investigates the uplink performance of two leading cooperative radio access network (CAN) architectures—distributed massive MIMO and Cloud Radio Access Network (C‑RAN)—under realistic fronthaul capacity constraints. Both architectures rely on a large number of remote radio units (RRUs) or access points (APs) that forward received signals to a central processing unit (CPU) via a high‑speed fronthaul link. While distributed massive MIMO allows each AP to perform some local baseband processing and then exchange channel state information (CSI) with the CPU, C‑RAN centralizes all baseband functions, requiring the raw (or lightly processed) received signals to be digitized and transmitted over the fronthaul. Consequently, the amount of data that must traverse the fronthaul becomes a critical bottleneck, especially in the uplink where the analog baseband must be quantized before transmission.

To model the quantization process, the authors adopt the Bussgang decomposition, which linearizes any memoryless non‑linear operation (such as uniform scalar quantization) by expressing the quantized signal (\tilde{\mathbf y}) as (\tilde{\mathbf y}= \mathbf A \mathbf y + \mathbf q). Here (\mathbf A) is a deterministic scaling matrix that depends on the quantizer step size and the statistics of the input (\mathbf y), while (\mathbf q) is a quantization noise vector that is uncorrelated with (\mathbf y). This decomposition enables an exact characterization of the quantization noise covariance (\mathbf R_{qq}), which is essential for optimal estimation.

The core technical contribution is a novel minimum‑mean‑square‑error (MMSE) estimator that jointly accounts for channel and data estimation in the presence of quantization noise. During the pilot phase, each AP quantizes its received pilot observations and forwards them to the CPU. Using the Bussgang‑derived linear model, the CPU computes the cross‑covariance (\mathbf R_{\tilde y h}= \mathbf A \mathbf R_{yh}) and the auto‑covariance (\mathbf R_{\tilde y \tilde y}= \mathbf A \mathbf R_{yy}\mathbf A^{!H}+ \mathbf R_{qq}). The MMSE channel estimate is then (\hat{\mathbf H}= \mathbf R_{\tilde y h}\mathbf R_{\tilde y \tilde y}^{-1}\tilde{\mathbf y}). In the data phase, the same covariance matrices are employed to construct a linear MMSE (LMMSE) detector that mitigates the effect of (\mathbf q) on the recovered symbols. This joint treatment of quantization noise yields a substantial reduction in estimation error compared with conventional approaches that either ignore the correlation between quantization noise and the signal or treat the noise as white Gaussian.

The paper also addresses the allocation of a finite fronthaul budget (C_{\text{total}}). Assuming each AP uses a uniform scalar quantizer with (b_m) bits per complex sample, the total fronthaul load is (\sum_{m=1}^{M} b_m W), where (W) is the system bandwidth. The authors formulate a convex optimization problem that maximizes the sum‑rate subject to the fronthaul constraint, deriving a water‑filling‑like solution for the optimal bit distribution ({b_m^\star}). The solution reveals that APs serving users with poor channel conditions (e.g., cell‑edge users) should be allocated more bits, while those with strong links can operate with fewer bits without significantly degrading overall performance.

Extensive Monte‑Carlo simulations validate the analytical findings. Key observations include: (i) With as few as 4 quantization bits per complex sample, the performance loss relative to an ideal, unlimited‑fronthaul system is less than 1 dB; increasing to 6 bits essentially eliminates the loss. (ii) The optimized, non‑uniform bit allocation improves the average spectral efficiency by roughly 15 % compared with a naïve uniform allocation under the same total fronthaul capacity. (iii) The proposed joint MMSE estimator reduces channel estimation mean‑square error by about 20 % relative to traditional pilot‑only LMMSE estimators, translating into higher achievable rates. (iv) Energy consumption on the fronthaul links can be reduced by over 10 % when the bit allocation is tuned to the channel statistics, because fewer bits are transmitted from well‑conditioned APs.

In conclusion, the study demonstrates that a rigorous Bussgang‑based linearization combined with a quantization‑aware MMSE estimation framework provides a powerful tool for designing fronthaul‑constrained cooperative networks. The results offer practical guidelines: (1) allocate fronthaul bits adaptively based on per‑AP channel quality rather than uniformly; (2) incorporate the exact quantization noise covariance into channel and data estimators to recover most of the performance lost to quantization; and (3) consider extensions such as non‑uniform quantizers, multi‑bit scaling, or deep‑learning‑based reconstruction to push fronthaul efficiency even further. This work thus bridges the gap between theoretical massive MIMO/C‑RAN concepts and the practical limitations imposed by real‑world fronthaul infrastructure.