Generative Graph Diffusion Model for Wireless Power and Fronthaul Bandwidth Allocation in Low-Altitude Assisted Cell-Free Networks

Low-altitude assisted cell-free networks provide a promising architecture for cooperative aerial-ground coverage through centralized control. However, the dynamics of low altitude network topology and aerial channels demand adaptive and generative resource management, while the distinct spectrum bands and limited capacity of access and fronthaul links make their joint allocation challenging. Existing discriminative model based methods offer only deterministic predictions with limited generative ability and struggle to adapt to such dynamics. To address these challenges, this paper investigates joint wireless power and fronthaul bandwidth allocation for spectrum efficiency maximization in low-altitude assisted cell-free networks. We first design a successive convex approximation (SCA) based algorithm to obtain near-optimal solutions as training samples. We then represent the network as a heterogeneous graph and propose a graph U-Net based diffusion (GUN-Diff) model, which learns the sample distribution to generate high-quality solutions via forward diffusion with Gaussian noise and reverse diffusion through a customized graph U-Net. We further prove that GUN-Diff possesses permutation equivariance, which theoretically supports its scalability and generative ability across varying network con figurations. Simulation results show that GUN-Diff approaches the spectral efficiency of the SCA-based method with polynomial computational complexity and generalizes well across diverse low altitude network environments.