Spectral Graph Convolution

The foundational operation that decomposes a graph signal into its frequency components. The eigenvectors of the graph Laplacian matrix serve as the Fourier basis, with corresponding eigenvalues representing frequencies. A signal is transformed by projecting it onto these eigenvectors: x̂ = Uᵀx, where U contains the eigenvectors. Low eigenvalues correspond to smooth, slowly-varying signals across the graph, while high eigenvalues capture rapid oscillations between neighboring nodes. This spectral representation enables frequency-selective filtering directly on irregular graph topologies.

The central mathematical object enabling spectral analysis, defined as L = D - A, where D is the degree matrix and A is the adjacency matrix. The Laplacian is a symmetric positive semi-definite operator that encodes fundamental topological properties:

• Its eigendecomposition L = UΛUᵀ yields the spectral basis
• The multiplicity of the zero eigenvalue equals the number of connected components
• The second smallest eigenvalue (algebraic connectivity) measures graph cohesion
• The normalized Laplacian L_norm = I - D^(-1/2)AD^(-1/2) is often preferred for stability

Direct eigendecomposition is computationally prohibitive for large graphs at O(n³). Modern spectral GNNs circumvent this by approximating filters using polynomials of the Laplacian:

• Chebyshev polynomials (ChebNet): Efficient recursive computation that localizes filters to K-hop neighborhoods without explicit eigenvector calculation
• First-order approximation (GCN): The seminal simplification by Kipf & Welling that limits the filter to 1-hop neighborhoods, achieving linear scaling while retaining competitive performance
• Cayley polynomials: Complex-valued filters that can capture directed and non-symmetric graph structures

Spectral filters act as frequency-selective operators on graph signals. A low-pass filter attenuates high eigenvalues, smoothing signals by averaging neighboring node features—critical for denoising interference patterns in cellular topology graphs. A high-pass filter amplifies differences between adjacent nodes, useful for edge detection and identifying boundaries between network clusters. The filter's frequency response function g(λ) defines its behavior: designing g(λ) as a learnable function enables the GNN to adaptively select which frequency components are relevant for tasks like resource allocation or anomaly detection.

A fundamental tension exists in spectral filter design:

• Localized filters (small polynomial order K) aggregate information from immediate neighborhoods, offering computational efficiency and spatial interpretability but limited spectral expressivity
• Global filters (high polynomial order or full eigendecomposition) achieve sharp frequency discrimination but lose spatial localization and scale poorly
• Spectral GNNs inherently balance this trade-off: the polynomial order K determines both the receptive field size and the filter's capacity to approximate complex frequency responses. For cellular topology graphs, K=2 or K=3 typically captures sufficient interference context without over-smoothing.

Spectral and spatial graph convolutions are mathematically equivalent under certain conditions. A K-th order polynomial spectral filter corresponds exactly to a spatial convolution aggregating features from the K-hop neighborhood. The key distinction is the design perspective:

• Spectral approach: Design filters in the frequency domain, then approximate spatially for efficiency
• Spatial approach: Define aggregation functions directly on neighborhoods (mean, sum, attention)
• Unification: Both paradigms converge in frameworks like Message Passing Neural Networks, where the choice between spectral and spatial design becomes a matter of inductive bias rather than fundamental architecture

The fundamental operator for spectral analysis, defined as L = D - A (degree matrix minus adjacency matrix). Its eigendecomposition L = UΛUᵀ yields the graph Fourier basis U and eigenvalues Λ that represent frequencies. The normalized variant L_norm = I - D⁻¹ᐟ²AD⁻¹ᐟ² is preferred for stability. In cellular networks, the Laplacian encodes interference topology, with eigenvalues indicating connectivity smoothness.

Projects a graph signal x onto the eigenbasis of the Laplacian: x̂ = Uᵀx. This transforms node features from the spatial domain to the spectral domain, where eigenvectors corresponding to small eigenvalues represent low-frequency, smooth variations across the graph, and large eigenvalues capture high-frequency, localized differences. Enables filtering of interference patterns by their spectral components.

A computationally efficient method to approximate spectral filters without explicit eigendecomposition. The filter g_θ(Λ) is approximated as a truncated expansion of Chebyshev polynomials T_k(Λ̃) up to order K, making convolution K-localized. This avoids the O(N³) eigendecomposition cost and enables spatially localized filtering. The ChebNet architecture popularized this approach for semi-supervised node classification.

A first-order approximation of spectral convolution that simplifies ChebNet to K=1. The layer-wise propagation rule is H^(l+1) = σ(D̃⁻¹ᐟ²ÃD̃⁻¹ᐟ² H^(l) W^(l)), where à = A + I adds self-loops. This operates as a low-pass filter, smoothing node features across neighborhoods. While computationally efficient for large cellular graphs, stacking multiple layers causes over-smoothing, limiting depth.

Two paradigms for graph convolution:

• Spectral: Defined in the Fourier domain via Laplacian eigendecomposition. Mathematically principled but basis-dependent, making it transductive—filters learned on one graph structure don't transfer to another.
• Spatial: Defined directly on neighborhoods via message passing. Inductive and transferable across graphs. Modern architectures favor spatial methods for dynamic cellular topologies where base station deployments change.

Spectral filters enable explicit control over which frequency components of a graph signal are amplified or suppressed:

• Low-pass filters smooth traffic load estimates across neighboring cells, useful for load balancing
• High-pass filters highlight anomalous interference spikes at specific nodes
• Band-pass filters isolate mid-frequency patterns like cluster-level congestion This frequency-aware design directly maps to physical wave propagation and interference patterns in radio networks.