Distributed Multi-User Beamforming (D-MUBF)

• Distributed Multi-User Beamforming (D-MUBF) is a set of techniques where multiple base stations or access points work together by computing beamformers and controlling power in a distributed manner.
• It leverages methods such as dual decomposition, convex approximations, reinforcement learning, and consensus optimization to tackle nonconvex interference management problems in multi-cell, cell-free, massive MIMO, and mmWave systems.
• These approaches balance precise convergence guarantees with runtime efficiency, reducing signaling overhead and adapting effectively to dynamic channel conditions.

Distributed Multi-User Beamforming (D-MUBF) denotes a class of downlink beamforming and power-control methods in which multiple base stations (BSs) or access points (APs) coordinate interference management while computing beamformers in a distributed manner. In the cited literature, this includes exact distributed power minimization under long-term CSI in multi-cell networks (Shen et al., 2013), distributed weighted sum-rate maximization in multicell MU-MIMO OFDMA downlink (Kibria et al., 2017), deep-reinforcement-learning-based distributed dynamic coordinated beamforming in massive MIMO cellular networks (Ge et al., 2023), ADMM-based downlink beamforming in cell-free massive MIMO (Zafari et al., 2024), and distributed multi-agent beam selection in mmWave MIMO networks (Wang et al., 2020). The topic therefore spans dual decomposition, semidefinite relaxation, consensus optimization, and multi-agent learning, with markedly different assumptions on CSI, signaling, synchronization, and optimality.

1. Architectural scope and canonical system models

A central architectural distinction in D-MUBF is whether each user is served by only one transmitter or by many transmitters simultaneously. In the multi-cell formulation of "Asynchronous Distributed Downlink Beamforming and Power Control in Multi-cell Networks" (Shen et al., 2013), the network has $M$ cells, each cell has one BS with $N$ antennas, each cell serves $K$ single-antenna users, and each user is associated with only one BS. The channel model is flat fading, only long-term CSI is available, and the long-term CSI is represented by spatial correlation matrices $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$ with $\operatorname{rank}(R_{m,n,i})>1$, which makes the beamforming problem non-convex (Shen et al., 2013).

In multicell MU-MIMO OFDMA downlink, "Distributed Weighted Sum-Rate Maximization in Multicell MU-MIMO OFDMA Downlink" (Kibria et al., 2017) considers $M$ cells, $K$ users per cell, $N_t$ transmit antennas at each BS, $N_r \ge 2$ receive antennas at each user, $N$ OFDMA subcarriers, 1-cell frequency reuse, and non-overlapping subcarrier allocation within each cell, so there is no intra-cell interference. The distributed beamforming viewpoint is that each BS optimizes only its own beamformers while treating other BSs’ beamformers as fixed (Kibria et al., 2017).

Massive-MIMO and cell-free variants broaden this scope. In "Deep Reinforcement Learning for Distributed Dynamic Coordinated Beamforming in Massive MIMO Cellular Networks" (Ge et al., 2023), the system is a TDD massive MIMO cellular network with $N$0 cells, one BS per cell, a URA with $N$1 antennas at each BS, and $N$2 single-antenna UEs per cell. In "ADMM for Downlink Beamforming in Cell-Free Massive MIMO Systems" (Zafari et al., 2024), $N$3 distributed APs, each with $N$4 antennas, jointly serve $N$5 single-antenna users on the same time-frequency resource; here every AP can jointly serve every user, and the received signal is a superposition of contributions from all APs (Ge et al., 2023, Zafari et al., 2024).

The mmWave literature adds codebook-constrained and mobility-driven distributed formulations. "Multi-Agent Double Deep Q-Learning for Beamforming in mmWave MIMO Networks" (Wang et al., 2020) studies multiple BSs and multiple mobile UEs, with each BS as one RL agent selecting one analog beamforming vector from a finite codebook. The largest received power association criterion determines which BS serves each UE at each time step (Wang et al., 2020).

Paper	Setting	Distributed unit
(Shen et al., 2013)	Multi-cell downlink with long-term CSI	Each BS computes beamformers and dual updates
(Kibria et al., 2017)	Multicell MU-MIMO OFDMA downlink	Each BS optimizes its own beamformers locally
(Ge et al., 2023)	TDD massive MIMO cellular network	Each BS is a DDPG agent
(Zafari et al., 2024)	Cell-free massive MIMO downlink	Each AP solves a local ADMM subproblem
(Wang et al., 2020)	mmWave MIMO with multiple BSs	Each BS is a DDQN beam-selection agent

This suggests that D-MUBF is not a single system model but a family of distributed interference-management problems whose coupling structure is determined by architecture: cell association in multi-cell networks, subcarrier reuse in OFDMA, joint service in cell-free systems, or codebook-based beam selection in mmWave networks.

2. Optimization objectives and mathematical formulations

A canonical D-MUBF objective is total downlink power minimization subject to user SINR constraints. Under long-term CSI, (Shen et al., 2013) formulates the multi-cell PMBP as

$N$6

with mean SINR

$N$7

The non-convexity arises because the long-term CSI matrices are not rank-one (Shen et al., 2013).

Weighted sum-rate maximization is a second canonical formulation. In multicell MU-MIMO OFDMA, (Kibria et al., 2017) maximizes the weighted sum-rate across all cells, users, and subcarriers subject to a per-BS transmit power constraint,

$N$8

The rate is $N$9, and the resulting WSRM problem is NP-hard in general (Kibria et al., 2017).

Massive-MIMO and cell-free formulations preserve the same objective structure while changing the information model and coupling pattern. In (Ge et al., 2023), the problem is sum-rate maximization over downlink beamformers subject to per-BS transmit power constraints, with the beamformer decomposed as $K$0. In (Zafari et al., 2024), the objective is

$K$1

subject to per-user SINR requirements

$K$2

In (Wang et al., 2020), the objective becomes codebook-constrained network sum-rate maximization under constant-modulus analog beamforming constraints (Ge et al., 2023, Zafari et al., 2024, Wang et al., 2020).

Across these formulations, D-MUBF alternates between two dominant goals: minimizing total transmit power for prescribed QoS, or maximizing network rate under per-transmitter power budgets. A plausible implication is that most algorithmic differences in D-MUBF arise less from the high-level objective than from how interference coupling is exposed to distributed computation.

3. Distributed solution paradigms

A classical exact route is dual decomposition plus fixed-point iteration. In (Shen et al., 2013), the non-convex PMBP is cast into the dual decomposition framework, the semidefinite dual constraints are transformed by matrix pencil theory into scalar bounds via the minimum non-negative generalized eigenvalue, and the dual variables are updated through the fixed-point mapping

$K$3

The convergence proof relies on showing that $K$4 is a standard interference function with positivity, monotonicity, and scalability, which then yields convergence of both synchronous and asynchronous iterations to the same unique fixed point. Beamforming vectors are recovered from the null spaces of the KKT matrices, and powers are recovered by a classical interference-based power update that can also be run asynchronously and distributively (Shen et al., 2013).

A second route is approximation-based local convexification. In (Kibria et al., 2017), WSRM is handled by a two-stage optimization process. The first stage uses rank-constrained rank minimization for IA initialization. The second stage updates receive filters, solves a convexified semidefinite relaxation in covariance variables $K$5, and recovers beamformers from the EVD of $K$6. The convex approximation is based on high-SINR decomposition, IA-based small-leakage approximation, and the trace approximation of $K$7 when $K$8 has small eigenvalues (Kibria et al., 2017).

Learning-based D-MUBF replaces explicit online optimization by distributed policy inference. In (Ge et al., 2023), the method uses DDPG in a multi-agent setting, with each BS as an agent. The action is not the complex beamformer itself but a low-dimensional parameter vector containing user power splits, a total transmit power fraction, interference leakage control factors, and a background noise control factor. The beamformer is then recovered through a WMMSE-derived structure

$K$9

so the network learns scalar parameters rather than a high-dimensional complex vector (Ge et al., 2023).

Consensus optimization provides a third route. In (Zafari et al., 2024), the original cell-free SINR constraints are relaxed by a triangle-inequality bound in the numerator and a generalized Minkowski inequality bound in the denominator, which yields separable local interference variables $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$0. Each AP solves a local convex subproblem, sends a real-valued interference vector of size $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$1 to a central node, and the central node updates a consensus variable by averaging. Dual variables are then updated in a standard ADMM step (Zafari et al., 2024).

In codebook-constrained mmWave D-MUBF, (Wang et al., 2020) adopts distributed multi-agent DDQN. Each BS observes a local history state from associated UEs, selects a beam index from a finite codebook via $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$2-greedy DDQN, applies the beam, receives a reward tied to achieved rate, and updates its online and target Q-networks. The practical problem is therefore beam selection rather than continuous beamformer optimization (Wang et al., 2020).

4. CSI models, signaling regimes, and distributed information exchange

CSI assumptions are decisive in D-MUBF because they determine both tractability and signaling overhead. The long-term-CSI formulation of (Shen et al., 2013) replaces instantaneous channels by spatial correlation matrices. The paper explicitly notes that the rank $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$3 structure makes the beamforming problem non-convex, unlike the instantaneous CSI case. The practical implication stated in the paper is that the signaling burden is lower than instantaneous-CSI methods, and that robustness to stale information makes the method suitable for practical backhaul-limited deployments (Shen et al., 2013).

In (Kibria et al., 2017), each BS optimizes only its own beamformers, uses its locally known channels and interference-related quantities relevant to its own optimization, and treats other BSs’ beamformers as fixed during its local update. The paper emphasizes that BSs do not need to report additional information each iteration, which reduces backhaul/signaling overhead, implementation complexity, and coordination latency (Kibria et al., 2017).

The distributed dynamic coordinated beamforming framework of (Ge et al., 2023) uses a hybrid information model: each BS operates with only local CSI and some historical information from other BSs. The exchanged quantities include compressed CSI of interferer BSs, power allocations, interference powers $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$4, achievable rates of interfered UEs, and interference-plus-noise power, transmitted with delay over interfaces such as LTE X2. The rationale is that the channel is time-correlated, so information from the previous slot remains useful for the current slot (Ge et al., 2023).

Cell-free ADMM in (Zafari et al., 2024) keeps full CSI local at each AP and exchanges only a real-valued interference vector of size $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$5. The paper contrasts this with centralized CSI exchange on the order of $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$6, whereas the proposed method exchanges on the order of $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$7. Because $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$8 is large in massive MIMO, this is presented as a substantial fronthaul reduction (Zafari et al., 2024).

In mmWave DDQN beamforming (Wang et al., 2020), the serving BS only uses local information from its currently associated UEs. The state is a memory vector of omni-directional rates, with optional location information, and association is redefined at every time step by the largest received power criterion. This is a weaker-information regime than continuous-CSI beamforming, and it shifts the distributed problem toward adaptive beam selection under mobility (Wang et al., 2020).

5. Optimality, convergence, and empirical behavior

The strongest exactness result in the cited literature appears in (Shen et al., 2013). Starting from any feasible initial $R_{m,n,i}=\mathbb{E}[h_{m,n,i}h_{m,n,i}^H]$9, the synchronous update $\operatorname{rank}(R_{m,n,i})>1$0 converges to a unique fixed point $\operatorname{rank}(R_{m,n,i})>1$1, which is the optimal dual solution. Under the asynchronous update rule with delayed or stale information, the sequence still converges to the same unique $\operatorname{rank}(R_{m,n,i})>1$2. Because strong duality holds for the PMBP, the optimal dual variables yield the same optimum as the centralized solution, and the overall asynchronous distributed beamforming and power-control method converges to the centralized optimum (Shen et al., 2013).

Approximation-based distributed WSRM is more limited analytically. In (Kibria et al., 2017), the original WSRM problem is nonconvex and globally optimal solutions are not guaranteed. The proposed method is explicitly a heuristic convex approximation based on high-SINR and IA assumptions, and no formal convergence proof is provided. Numerically, the authors report convergence in a small number of iterations in all tested scenarios, and the simulations show that IA-based initialization yields a much better convergence curve than random initialization, with the sum-rate gap widening as transmit power increases (Kibria et al., 2017).

The cell-free ADMM method of (Zafari et al., 2024) is near-optimal but slightly conservative because of the SINR relaxation. In simulation, the algorithm consistently converged in up to 10 ADMM iterations over more than 100 channel realizations. For $\operatorname{rank}(R_{m,n,i})>1$3 dB and $\operatorname{rank}(R_{m,n,i})>1$4 dB, the achieved SINR with $\operatorname{rank}(R_{m,n,i})>1$5 stayed above about 14.5 dB and 23.5 dB, respectively, but with less than 20% outage in the first case and around 40% outage in the second; the paper reports that outage can be eliminated by increasing the safety factor $\operatorname{rank}(R_{m,n,i})>1$6, with $\operatorname{rank}(R_{m,n,i})>1$7 and $\operatorname{rank}(R_{m,n,i})>1$8 sufficient for the two cases (Zafari et al., 2024).

Learning-based distributed beamforming reports empirical competitiveness rather than exact optimality. In (Ge et al., 2023), the proposed DDPG-based method converges after about 4000 time slots, achieves performance close to the direct FP upper bound, outperforms WMMSE and closed-form FP, and for $\operatorname{rank}(R_{m,n,i})>1$9, $M$0, $M$1 has a reported runtime of about 1.5 ms, versus about 40 s for WMMSE and more than 800 s for direct FP. In (Wang et al., 2020), distributed multi-agent DDQN significantly outperforms random beam selection and achieves performance comparable to exhaustive search while operating at much lower complexity; including location information improves performance only slightly, and sum-rate grows almost linearly with the number of UEs (Ge et al., 2023, Wang et al., 2020).

These results support a sharp distinction within D-MUBF. Some methods are exact distributed solvers with formal convergence and centralized-optimum recovery; others are practical distributed approximations or learned policies whose advantages lie in overhead reduction, runtime, or mobility adaptation rather than analytical optimality.

6. Boundaries of the concept and relation to adjacent beamforming literatures

Not every multi-user beamforming method is D-MUBF in the strict sense of geographically distributed transmitters coordinating a common downlink. "Beamforming Algorithm for Multiuser Wideband Millimeter-Wave Systems with Hybrid and Subarray Architectures" (Viteri-Mera et al., 2019) is explicitly described as not a direct D-MUBF algorithm; it studies one AP serving multiple users in a wideband OFDM mmWave system, with training-based analog beam selection and digital block diagonalization at a centralized AP. "Multiuser Beamforming for Partially-Connected Millimeter Wave Massive MIMO" (Qi et al., 2023) is likewise not a distributed multi-transmitter beamforming paper in the classic D-MUBF sense; all beamforming is done at a single BS, even though the imperfect-CSI analog-only design uses an SLNR-like beam-nulling formulation. "Harnessing Multimodal Sensing for Multi-user Beamforming in mmWave Systems" (Patel et al., 2024) proposes sensor-aided MU beam training and an MU beamforming strategy that prevents transmission to multiple users over the same directions, but the BS remains responsible for beam training and beam selection, so the method is centralized rather than distributed across transmit points (Viteri-Mera et al., 2019, Qi et al., 2023, Patel et al., 2024).

A common misconception is that “distributed” means “no coordination.” The cited literature does not support that interpretation. The asynchronous long-term-CSI method of (Shen et al., 2013) explicitly uses local and neighboring information; the DDPG framework of (Ge et al., 2023) uses delayed historical information from other BSs; and the cell-free ADMM method of (Zafari et al., 2024) uses a central averaging step over low-dimensional interference vectors. What changes across methods is not the existence of coordination, but the dimensionality, freshness, and semantic content of the exchanged variables.

A second misconception is that learning-based D-MUBF directly predicts beamforming vectors. In (Ge et al., 2023), the neural network predicts scalar parameters $M$2 that are mapped to beamformers using a WMMSE-inspired structure. In (Wang et al., 2020), each BS selects a beam index from a finite codebook. The latter paper also contains an inconsistency in the reward definition: the explanatory text describes the reward as beamformed rate relative to omni-directional rate, while the algorithm block explicitly uses $M$3 (Ge et al., 2023, Wang et al., 2020).

The practical domains explicitly identified for D-MUBF include CoMP-like multi-cell networks, femtocells/small cells, interference-managed spectrum sharing, cell-free massive MIMO, massive MIMO cellular networks, and mmWave MIMO networks (Shen et al., 2013, Zafari et al., 2024, Ge et al., 2023, Wang et al., 2020). Taken together, the literature suggests that D-MUBF is best understood as a distributed interference-coupled downlink design paradigm whose concrete realization depends on architecture, CSI regime, and the acceptable balance between exactness, overhead, and online latency.