Graph Neural Network Beamforming

The fundamental innovation is modeling the wireless network as a heterogeneous graph where:

• Nodes represent transmitters (base stations) and receivers (user equipment)
• Edges represent interference links, with edge features encoding path loss, channel gain, or distance
• Node features include local CSI, buffer status, and QoS requirements

This graph structure explicitly captures the topology of interference, allowing the GNN to reason about spatial relationships that are invisible to traditional per-link optimizers.

GNN beamforming uses neural message passing to enable distributed coordination:

• Each node iteratively aggregates information from its neighbors (interferers)
• The aggregation function is learned, not hand-crafted
• After multiple message-passing rounds, each node computes its own precoding vector

This mirrors the distributed optimization of classical WMMSE algorithms but replaces explicit convex solvers with a learned, feed-forward inference pass that executes in microseconds.

A critical architectural property: the beamforming policy is permutation equivariant to the ordering of users and permutation invariant to the ordering of antennas.

This means:

• If you re-index the users, the output beamforming vectors permute identically
• The GNN respects the set-based nature of multi-user systems
• The model generalizes to arbitrary network topologies without retraining

This is enforced through proper choice of aggregation functions (sum, max, mean) and avoids the brittle input-ordering dependencies of standard MLPs or CNNs.

Unlike fully-connected neural networks that require fixed input dimensions, GNN beamforming scales to arbitrary numbers of transmitters and users:

• The same trained model can be deployed in a 4-user pico-cell or a 64-user massive MIMO scenario
• Computational complexity scales linearly with the number of edges, not quadratically
• Training on small networks transfers to larger deployments via the inductive bias of the graph structure

This is the key advantage over centralized optimization (e.g., WMMSE), which becomes computationally prohibitive beyond a few dozen links.

GNN beamformers are typically trained end-to-end without labeled data using a loss function that directly maximizes network utility:

• Sum-rate maximization: Loss = negative sum of user rates
• Minimum rate constraints: Penalize violations of QoS thresholds
• Power constraints: Enforce per-transmitter power budgets via projection layers

Gradients flow through the GNN, the beamforming computation, and the channel model. This unsupervised learning paradigm eliminates the need for optimal precoding solutions as training targets—which are often unavailable for large networks.

A powerful hybrid approach combines GNNs with deep unfolding of classical algorithms:

• The WMMSE algorithm is unrolled into a neural network with a fixed number of iterations
• Each iteration's update steps are replaced with small GNN modules
• The resulting unfolded GNN retains the convergence guarantees of WMMSE while learning optimal step sizes and interference weights

This bridges the gap between purely data-driven and purely model-based beamforming, offering robust performance even when channel models mismatch.

The direct synthesis of a multi-antenna precoding matrix by a neural network. Unlike GNN beamforming, which explicitly models the network topology as a graph, neural precoding learns the mapping from channel state information to the optimal precoder in a purely data-driven manner, often treating the problem as a black-box function approximation. It avoids explicitly solving complex convex optimization problems like weighted minimum mean square error (WMMSE) in real-time.

• Key Distinction: Operates on a single link or cell, whereas GNN beamforming scales across the entire interference network.
• Architecture: Typically uses a multi-layer perceptron (MLP) or convolutional neural network (CNN) to process channel matrices.
• Training: Trained offline using supervised learning on solutions from traditional optimizers.

A deep learning methodology that unrolls the iterations of a classic optimization algorithm, such as the WMMSE algorithm for beamforming, into a neural network. Each layer of the unrolled network corresponds to one iteration of the algorithm, but hand-crafted parameters are replaced with learnable variables. This provides the interpretability of model-based methods with the performance of deep learning.

• Relation to GNN: Both are hybrid model-based/data-driven approaches; GNNs model the network structure, while unfolding models the algorithmic structure.
• Advantage: Requires significantly fewer layers and parameters than a generic neural network.
• Example: Unfolding the iterative shrinkage-thresholding algorithm (ISTA) creates Learned ISTA (LISTA).

A privacy-preserving distributed learning paradigm where multiple base stations collaboratively train a shared beamforming model without centralizing sensitive user channel data. Each base station trains a local model on its own data and only shares encrypted model updates (gradients or weights) with a central server for aggregation.

• Synergy with GNNs: GNN beamforming is inherently distributed, making it a natural fit for federated learning. Each node (base station) can update its local GNN parameters and share them.
• Privacy: Raw channel state information (CSI) never leaves the local site.
• Challenge: Non-IID data distributions across base stations can slow convergence.

A planar metasurface composed of many passive or semi-passive elements that can dynamically tune the phase, amplitude, or polarization of impinging electromagnetic waves. Neural networks, including GNNs, are used to optimize the reflection coefficients of an RIS to create favorable propagation paths for beamforming.

• Graph Modeling: An RIS-aided network is naturally modeled as a graph where RIS elements are additional nodes that influence edge weights (channel gains).
• Optimization: GNNs can jointly optimize the precoding at the base station and the phase shifts at the RIS to maximize sum-rate.
• Hardware: Operates with near-zero power consumption compared to active relays.

A beamforming architecture that employs the attention mechanism from transformers to dynamically weigh the importance of different propagation paths or antenna elements. Instead of modeling the network as a graph with explicit edges, the attention mechanism learns pairwise relevance scores between all nodes in a fully connected fashion.

• Comparison to GNN: GNNs use a sparse, pre-defined graph structure based on known interference links, while attention-based methods learn the connectivity pattern from data.
• Scalability: The quadratic complexity of self-attention can be a bottleneck for very large networks, which GNNs handle more efficiently.
• Robustness: Excels in highly scattering environments where path importance is complex.

A neural network architecture where weights, biases, and activations are complex numbers, and backpropagation is performed using Wirtinger calculus. This inherently preserves the phase information critical for coherent wireless signal processing tasks like beamforming.

• Relevance to GNN Beamforming: Wireless channel coefficients and precoding vectors are inherently complex-valued. Using a complex-valued GNN avoids the information loss from splitting complex numbers into real and imaginary parts.
• Activation Functions: Uses complex-specific activations like modReLU or complex cardioid.
• Performance: Often achieves superior performance with fewer parameters compared to real-valued equivalents for physical layer tasks.