As wireless networks transition toward 6G, two architectural paradigms dominate the quest for higher spectral efficiency: Massive MIMO (Multiple-Input Multiple-Output) and Extremely Large-Scale Antenna Arrays (ELAA). By packing hundreds or thousands of antennas onto a base station or distributing them across a geographic area, networks can serve dozens of users simultaneously on the same frequency band.
However, scaling up the number of antennas introduces a severe physical challenge: the data deluge. In a traditional centralized architecture, raw digital signals (I/Q samples) from every single antenna element must be routed to a central baseband processing unit. When dealing with hundreds of antennas operating at wide bandwidths, this internal data transfer rate quickly exceeds Terabits per second, creating an unsustainable input/output (I/O) and fronthaul bandwidth bottleneck (Kaipeng Li et al.; Charles Jeon et al.). Furthermore, standard centralized baseband algorithms—such as Minimum Mean-Square Error (MMSE) equalization or symbol-level precoding—require matrix inversions that scale cubically with the number of antennas or users, hitting a computational wall.
Decentralized Baseband Processing (DBP) resolves this bottleneck by breaking the centralized processor apart. Instead of hauling raw data to a single brain, DBP partitions the antenna array into independent clusters, processing signals locally at the edge to dramatically lower interconnect bandwidth and distribute the computational load.
The Core Mechanism: Array Partitioning
The fundamental step in decentralized baseband processing is dividing a massive antenna array into $C$ smaller, distinct clusters (Yanqing Xu et al.; Kaipeng Li et al.). Each cluster is equipped with its own local processing unit (LPU) and operates using only its local Channel State Information (CSI).
Centralized Architecture:
[Antenna 1] \
[Antenna 2] --> (Raw I/Q Data) --> [Central Baseband Processor] (Cubic Complexity Wall)
[Antenna 3] /
Decentralized Architecture:
[Cluster 1: Antennas 1-3] --> [LPU 1] \
[Cluster 2: Antennas 4-6] --> [LPU 2] --> (Low-Dimensional Updates) --> [Fusion Step]
[Cluster 3: Antennas 7-9] --> [LPU 3] /
By localizing the initial computations, the system achieves modular scalability (Yanqing Xu et al.). If an operator wants to add more antennas to a base station, they do not need to redesign or upgrade a central processor; they simply add more self-contained antenna clusters.
Why It Works: The Dimensionality Reduction Trick
To understand why DBP successfully reduces interconnect traffic, it helps to look at how data is transformed within the clusters. In a single-user or multi-user setup, let $M$ be the total number of base station antennas, and $K$ be the number of users being served. Crucially, in massive MIMO systems, $M \gg K$ (e.g., 256 antennas serving 16 users).
In a single single-carrier block, a centralized processor must ingest an $M \times 1$ vector of raw received signals. As analyzed by Kaipeng Li, decentralized architectures can be split into two primary philosophies:
- Partial Decentralization: Each cluster $c$ manages a subset of antennas $M_c = M/C$. Instead of streaming raw signals, the local unit computes a local Gram matrix $\mathbf{G}_c = \mathbf{H}_c^H \mathbf{H}_c$, where $\mathbf{H}_c$ is the local channel matrix of size $M_c \times K$. Notice the dimensions: $\mathbf{G}_c$ is a $K \times K$ matrix. Because the number of users $K$ is small, transmitting this $K \times K$ matrix to a central hub requires significantly less bandwidth than streaming raw high-rate I/Q samples over time (Yue Dai et al.). The central hub then simply sums the local matrices ($\mathbf{G} = \sum \mathbf{G}_c$) to perform global detection.
- Full Decentralization: Local clusters perform data detection or precoding entirely at the edge, computing local estimates of the transmitted symbols based solely on their own antennas. They then pass only these final, low-bandwidth symbol estimates to a fusion fabric, bypassing the central processor entirely.
Algorithmic Strategies: Consensus vs. Feedforward
Executing baseband tasks across isolated clusters without a central coordinator risks losing signal processing accuracy. To close the performance gap between decentralized and centralized systems, researchers utilize two main algorithmic paradigms.
Iterative Consensus (ADMM and BCD)
To find the globally optimal signal estimate without sharing raw data, clusters can coordinate iteratively using consensus optimization. Kaipeng Li et al. demonstrated that the Alternating Direction Method of Multipliers (ADMM) can be mapped onto decentralized antenna clusters. Clusters compute local symbol estimates, share their results with neighbors to form a "global consensus" vector, and iteratively refine their local calculations. Within a few iterations, this consensus approach achieves a signal-to-noise ratio (SNR) virtually identical to a centralized processor.
Alternatively, this communication can happen sequentially. Xiaotong Zhao et al. proposed a Block Coordinate Descent (BCD) algorithm implemented over a "daisy-chain" architecture. Instead of broadcasting to a central hub, antenna clusters are organized in a sequential loop; each unit performs local MMSE equalization under colored noise and passes its updated estimate to the next neighbor, minimizing interconnect routing overhead. For downlink transmission, Ang Li 0003 et al. leveraged an inversion-free parallel ADMM approach to distribute symbol-level precoding (SLP), eliminating the need for costly centralized matrix inversions.
Feedforward Architectures
While consensus methods are mathematically elegant, their iterative nature introduces latency—units must wait for updates from other clusters before proceeding. To achieve gigabit-per-second throughputs with sub-millisecond latency, Charles Jeon et al. and Kaipeng Li developed non-iterative, feedforward architectures.
In a feedforward setup, clusters process incoming signals in parallel in a single pass, computing local un-fused symbol estimates. These estimates are then combined in a final, single-step fusion network. By optimizing the cluster mapping directly onto hardware accelerators (like multi-GPU clusters connected via high-speed fabrics), Kaipeng Li showed that feedforward architectures can eliminate iterative communication delays entirely, accepting a negligible SNR penalty in exchange for massive throughput scaling.
Extension to Cell-Free Massive MIMO
The principles of decentralized baseband processing become even more critical when antennas are physically separated across a geographic area rather than co-located on a single tower—a framework known as Cell-Free Massive MIMO (Hengtao He et al.). Here, traditional cellular boundaries are erased, and a massive number of distributed Access Points (APs) cooperate to serve users.
Centralized C-RAN:
[AP 1] --(Raw Quantized Signals)--> [Central Cloud] (Fronthaul Bottleneck!)
[AP 2] --(Raw Quantized Signals)--> [Central Cloud]
Decentralized Cell-Free:
[AP 1] --(Local Weighting/Combining)--> [Low-Bandwidth Estimates] --> [Fusion Hub]
[AP 2] --(Local Weighting/Combining)--> [Low-Bandwidth Estimates] --> [Fusion Hub]
In this domain, the fronthaul link connecting the APs to a central cloud is the primary bottleneck (Mugen Peng et al.). Kanapathippillai Cumanan et al. showed that forcing APs to perform local signal weighting and combining before quantizing and transmitting data over the backhaul drastically reduces network capacity requirements.
Similarly, Peng Zhang 0065 et al. demonstrated that decentralizing active precoding calculations allows individual APs to execute local matrix inversions. This shifts the computational complexity from a cubic relationship with the total number of network antennas down to a cubic relationship with only the antennas per AP. For massive unsourced multiple access, Kaan Okumus et al. proved that distributed likelihood processing at the APs preserves the macro-diversity gains of large arrays while protecting the core network from data congestion.
Hardware-Friendly Interference Rejection
Beyond solving the data deluge, decentralized processing architectures offer unique advantages for hardware-constrained systems, particularly in millimeter-wave (mmWave) deployments that utilize low-resolution Analog-to-Digital Converters (ADCs) to save power. Low-resolution ADCs are highly vulnerable to near-far interference or intentional jamming, which can easily saturate the quantizers and blind the receiver.
Gian Marti et al. and Gian Marti demonstrated that by partitioning a large array into localized clusters, engineers can apply hardware-friendly analog spatial transforms (such as "beam-slicing") within each cluster before the signals hit the digital quantizers. This modular approach isolates high-energy jammer interference into just a few localized data converters, leaving the remaining clusters free to capture clean user data without requiring dense, full-array matrix multiplications.
To handle extreme variations in user signal strengths, Victoria M. T. Palhares et al. extended this concept by implementing localized Householder reflections within independent antenna subsets. This prevents dominant, nearby signals from drowning out weaker, distant users in quantization noise, proving that massive antenna arrays do not require global, centralized connectivity to manage complex spatial interference patterns effectively.
Key Foundational Work
Distributed MIMO radar using compressive sampling (2008) — contributed 1 concept elements to this direction.
Fronthaul-constrained cloud radio access networks: insights and challenges (2015) — contributed 1 concept elements to this direction.