Abstract: We consider the problem of resource allocation in large scale wireless networks. When contextualizing wireless network structures as graphs, we can model the limits of very large wireless systems as manifolds. To solve the problem in the machine learning framework, we propose the use of Manifold Neural Networks (MNNs) as a policy parametrization. In this work, we prove the stability of MNN resource allocation policies under the absolute perturbations to the Laplace-Beltrami operator of the manifold, representing system noise and dynamics present in wireless systems. These results establish the use of MNNs in achieving stable and transferable allocation policies for large scale wireless networks. We verify our results in numerical simulations that show superior performance relative to baseline methods.
1. Introduction
Modern wireless communication systems are characterized by an increasing number of interconnected devices, making resource allocation—such as power control and interference management—a high-dimensional and non-convex optimization challenge. While traditional algorithms like Weighted Minimum Mean Square Error (WMMSE) are effective, they often struggle with computational complexity as the network size grows. Recently, Graph Neural Networks (GNNs) have emerged as a powerful alternative, learning heuristics that generalize across different topologies. However, a significant theoretical gap remains: ensuring that these learned policies are stable under environmental noise and transferable to networks much larger than those used during training.
2. From Discrete Graphs to Continuous Manifolds
To address the limitations of discrete graph representations in the limit of massive networks, this work models the wireless system as a continuous manifold $\mathcal{M}$. In this framework, the discrete network of $m$ transmitter-receiver pairs is viewed as a sampling from an underlying geometric structure. The central operator in this continuous space is the Laplace-Beltrami operator ($L$), which serves as the manifold equivalent of the graph Laplacian. It characterizes the propagation of signals and interference across the network's surface.
3. Manifold Neural Networks and Stability
We propose Manifold Neural Networks (MNNs) as a policy parametrization for resource allocation. MNNs consist of a cascade of manifold convolutional filters followed by non-linearities. To ensure the policy remains robust against the inherent "noise" of wireless environments—such as channel measurement fluctuations and topological shifts—we utilize Frequency Difference Threshold (FDT) filters. These filters group eigenvalues of the Laplace-Beltrami operator that are within a threshold $\alpha$ of each other, treating them similarly.
The core theoretical contribution is the proof of stability (Theorem 1). We demonstrate that MNNs using FDT filters are stable under absolute perturbations to the manifold operator. This provides a rigorous foundation for the observation that these models can maintain performance even when the underlying system dynamics change slightly.
4. Numerical Evaluation
The proposed MNN-based policies were evaluated in simulated 2D wireless environments. A policy was trained using unsupervised learning on a network of only $m=50$ transmitter-receiver pairs and subsequently tested on networks scaling up to $m=1,000$ pairs.
4.1 Transferability to Large-Scale Networks
As illustrated in Figure 2, the MNN policy (implemented as a GNN for discrete execution) consistently outperforms standard baselines, including WMMSE, equal power allocation, and random allocation. Notably, the performance gap remains significant as the number of transmitters increases, demonstrating exceptional transferability from small-scale training to large-scale deployment.
4.2 Robustness to Perturbations
To verify the stability results, we introduced log-normal noise to the channel states and measured the "Difference of Ratios"—a metric representing the deviation in performance between the original and perturbed settings. Figure 3 shows that while the error varies slightly with the number of layers or the number of filters ($F$), the overall difference remains bounded and small. This validates the stability bounds established in our theoretical analysis.
5. Conclusion
By leveraging the manifold limit of large-scale wireless networks, we have developed a framework for resource allocation that is both theoretically stable and practically transferable. Manifold Neural Networks, equipped with FDT filters, bridge the gap between small-scale training and massive-scale operation, offering a robust solution for the next generation of wireless systems.
Summary updated