Asymptotically Optimal Local Receiver in Uplink CF-mMIMO: A Functional-Variational Analysis

Technical Summary: Asymptotically Optimal Local Receiver in Uplink CF-mMIMO

Problem Statement
In Cell-Free Massive Multiple-Input Multiple-Output (CF-mMIMO) systems, the standard approach for uplink local receiver design relies on the "use-and-then-forget" (UatF) bounding technique. This framework typically employs a local Minimum Mean Square Error (LMMSE) receiver at each Access Point (AP) followed by Large-Scale Fading Decoding (LSFD) at the Central Processing Unit (CPU). While analytically tractable, the authors identify three fundamental limitations in this paradigm:

1. Performance Gap: The UatF lower bound is a conservative surrogate for the true ergodic achievable rate. The gap between the bound and the true rate can be significant, particularly in interference-limited regimes or when channel hardening is not fully realized.
2. Overhead and Complexity: The LSFD stage requires the CPU to collect local estimates and compute optimal combining coefficients based on long-term channel statistics. This introduces additional fronthaul overhead and computational burden at the CPU.
3. Lack of Rigorous Optimality Proof: While the LMMSE-LSFD structure is widely adopted, the paper notes that a rigorous proof establishing its optimality specifically under the UatF objective has not been formally reported in prior literature.

The core problem addressed is whether local receivers can be designed directly from the true ergodic-rate objective (rather than the UatF surrogate) to improve performance while avoiding the computational and signaling costs associated with LSFD.

Methodology
The paper proposes a direct optimization framework for local receiver design under the assumption of perfect local Channel State Information (CSI) at each AP. The methodology proceeds as follows:

1. Problem Formulation: The authors formulate the receiver design as a functional optimization problem aimed at maximizing the ergodic rate, $E[\log_2(1 + \text{SINR}_{\text{INS}})]$. This is a challenging fractional programming problem involving an expectation of a logarithm over an infinite-dimensional function space.
2. Large-System Approximation: To render the problem tractable, the authors leverage large-system random matrix theory. They introduce an expectation-based surrogate constraint to replace the pointwise normalization of the instantaneous Signal-to-Interference-plus-Noise Ratio (SINR). In the large-system limit (where the total number of antennas $NM$ is large), random quantities in the SINR denominator concentrate around their expectations, allowing the pointwise constraint to be relaxed to an expectation-based constraint.
3. Functional-Variational Analysis: Using functional derivatives, the authors derive the first-order stationarity conditions for the relaxed problem. This involves constructing a Lagrange dual functional and analyzing perturbations.
4. Decoupling of Cross-AP Terms: A critical step involves analyzing the cross-AP coupling terms (interference from other APs). Using random matrix theory, the authors prove that in the large-system limit, the conditional expectation of these coupling terms vanishes. This decouples the local receiver design across APs.
5. Derivation of Q-LMMSE: The analysis yields a "Quasi-LMMSE" (Q-LMMSE) receiver. The authors demonstrate that this receiver shares the same directional vector as the conventional LMMSE receiver but differs only by an instantaneous, CSI-dependent scalar factor.

Key Contributions
The paper makes the following specific contributions:

• Functional-Variational Framework: The authors develop a novel framework for direct ergodic-rate-oriented receiver design in CF-mMIMO, moving beyond the UatF surrogate.
• Asymptotically Optimal Receiver: They derive the Q-LMMSE receiver in closed form, which is shown to be asymptotically optimal under the ergodic rate objective in the large-system regime.
• Formal Optimality Proof: The paper provides a rigorous variational analysis proving that the local MMSE combining followed by LSFD is indeed the optimal structure when the UatF bound is adopted as the performance metric.
• Complexity Reduction: By exploiting the matrix inversion lemma (specifically the Sherman-Morrison formula), the authors show that the Q-LMMSE receiver incurs the same per-AP computational complexity as the standard LMMSE receiver. Crucially, it eliminates the need for CPU-side LSFD coefficient calculation and the associated fronthaul signaling for statistical parameters.
• Performance Gain: The Q-LMMSE receiver employs equal-gain combining (direct summation) at the CPU. The instantaneous scalar coefficients vary across APs, implicitly providing adaptive weighting that better exploits instantaneous spatial degrees of freedom compared to the static statistical weighting of LSFD.

Numerical Results
The authors validate their theoretical findings through numerical simulations comparing the proposed Q-LMMSE receiver against the conventional LMMSE-LSFD benchmark and a centralized MMSE (C-MMSE) upper bound. Key findings include:

• Ergodic Rate Gain: The Q-LMMSE receiver consistently outperforms the LMMSE-LSFD benchmark in terms of true ergodic rate. The gain is approximately 5% when the number of antennas per AP is low (e.g., $N=8$), where channel hardening is less pronounced.
• Convergence: As the number of antennas per AP ($N$) increases, the performance gap between Q-LMMSE and LMMSE-LSFD narrows, consistent with the channel hardening phenomenon making the UatF bound tighter.
• Scalability: The proposed receiver maintains superior performance across varying numbers of APs ($M$) and users ($K$), while operating with strictly lower system-level complexity by removing the CPU-side LSFD processing.

Significance and Claims
The paper claims that the proposed Q-LMMSE receiver offers a compelling trade-off between performance and complexity. By deriving the receiver directly from the ergodic rate objective, it achieves a higher true data rate than the standard UatF-based approach without incurring the additional computational overhead of LSFD at the CPU. The authors emphasize that the improvement stems not from a more complex local processing algorithm (which remains $O(N^3)$), but from a more efficient fusion strategy at the CPU that implicitly adapts to instantaneous channel conditions rather than relying on long-term statistics.

The work is presented as a theoretical advancement in the design of distributed receivers for beyond-5G and 6G systems, providing a rigorous alternative to the dominant UatF framework. The authors note that future work will extend this analysis to imperfect CSI scenarios and investigate robustness under hardware impairments.