Spectral Graph Convolutions

Updated 4 April 2026

Spectral graph convolutions are a set of operations that extend traditional convolution by filtering graph signals in the Fourier domain using the graph Laplacian.
Polynomial and rational parameterizations, such as Chebyshev expansions and Cayley filters, enable efficient localized computation and control filter smoothness.
Advanced methods generalize these convolutions to directed, heterogeneous, and complex graphs, enhancing transferability and scalability for deep learning applications.

Spectral graph convolutions are a class of graph signal processing operations that generalize classical convolutions to irregular graph domains by operating in the eigenbasis of the graph Laplacian. These techniques provide a foundation for graph convolutional neural networks (GCNs) and various spectral graph neural architectures. Spectral graph convolution leverages the spectrum of Laplacian operators to define filters, supports polynomial and rational parameterizations, facilitates efficient localized computation, extends naturally to heterogeneous and directed graphs, and underpins key advances in multi-graph transferability and scalable deep learning for general non-Euclidean domains.

1. Mathematical Formalism: Graph Laplacian, Fourier Transform, and Spectral Filtering

Let $G = (V, E)$ denote a graph with $|V| = n$ nodes, adjacency matrix $A \in \mathbb{R}^{n \times n}$ (possibly weighted), and degree matrix $D = \mathrm{diag}(d_1, \dots, d_n)$ where $d_i = \sum_j A_{ij}$ . The most common operators are the combinatorial Laplacian $L = D - A$ , or the symmetric normalized Laplacian $L_{\rm sym} = I - D^{-1/2} A D^{-1/2}$ (Bruna et al., 2013, Henaff et al., 2015, Chen, 2020, Ye et al., 9 Mar 2025).

Since $L$ is real symmetric and positive semidefinite, it admits an eigendecomposition

$L = U \Lambda U^T, \quad U \in \mathbb{R}^{n \times n}, \quad \Lambda = \mathrm{diag}(\lambda_1,\dots,\lambda_n)$

where $\lambda_i \ge 0$ are eigenvalues ("graph frequencies"), and $|V| = n$ 0 is orthonormal.

For any vector-valued graph signal $|V| = n$ 1, its graph Fourier transform is $|V| = n$ 2, and the inverse is $|V| = n$ 3. A spectral filter is defined as $|V| = n$ 4, with $|V| = n$ 5 a diagonal frequency response. The spectral graph convolution is then

$|V| = n$ 6

This recovers the classical convolution theorem: multiplication in the frequency domain is convolution in the vertex domain (Bruna et al., 2013, Chen, 2020, Ye et al., 9 Mar 2025).

2. Parameterizations: From Full-Basis to Polynomial and Rational Filters

The original spectral CNN formulation learns $|V| = n$ 7 independent filter weights (frequency responses), but this is computationally and statistically inefficient for large $|V| = n$ 8 and results in non-local filters (Bruna et al., 2013). Spatial localization and parameter efficiency are achieved by restricting the spectral filter to be a low-degree polynomial or rational function of $|V| = n$ 9.

Chebyshev polynomial parameterization: The filter response is expanded as $A \in \mathbb{R}^{n \times n}$ 0, where $A \in \mathbb{R}^{n \times n}$ 1 is the Chebyshev polynomial, and $A \in \mathbb{R}^{n \times n}$ 2 rescales the spectrum to $A \in \mathbb{R}^{n \times n}$ 3. In the vertex domain, this yields a $A \in \mathbb{R}^{n \times n}$ 4-hop localized filter:

$A \in \mathbb{R}^{n \times n}$ 5

This enables $A \in \mathbb{R}^{n \times n}$ 6-time evaluation, without explicit eigendecomposition (Bruna et al., 2013, Henaff et al., 2015, Chen, 2020, Nilsson et al., 2020, Ye et al., 9 Mar 2025).

General polynomial and basis expansions: Filters using monomials, Jacobi, or Bernstein polynomials (e.g., $A \in \mathbb{R}^{n \times n}$ 7) have been explored for expressivity and numerical stability (Huang et al., 2024).
Rational (Cayley) filters: Spectral multipliers are parametrized as $A \in \mathbb{R}^{n \times n}$ 8 with $A \in \mathbb{R}^{n \times n}$ 9. These specialize to frequency bands of interest, capture narrow-band phenomena, and maintain spatial localization via an exponential decay property in the graph domain (Levie et al., 2017).

The choice of polynomial or rational basis controls the trade-off between locality, frequency selectivity, and computational complexity. Chebyshev parameterization is the workhorse due to its orthogonality properties and efficient recurrence relations (Henaff et al., 2015, Nilsson et al., 2020, Ye et al., 9 Mar 2025).

3. Extensions: Heterogeneous, Directed, and Signed Graphs

Spectral graph convolution machinery has been extended to various non-standard graph topologies by recharacterizing the underlying operator and domain of functional calculus.

Directed graphs: Classical Laplacian-based spectral convolutions were restricted to undirected graphs. Several approaches generalize spectral GCNs for directed graphs:
- Redefining Laplacians using a symmetrized version via the Perron vector of the transition matrix, maintaining real symmetricity and positive semidefiniteness (Ma et al., 2019).
- Holomorphic functional calculus extends convolution to non-normal operators on arbitrary directed graphs, enabling polynomial, resolvent, and general holomorphic spectral filters (Koke et al., 2023).
- The complex Hermitian (magnetic) Laplacian approach for signed and directed graphs guarantees PSD spectra, enabling Chebyshev or rational spectral filtering in the complex domain (Ko et al., 2022).
Heterogeneous graphs: Spectral filtering is formulated using positive noncommutative polynomials over a set of normalized adjacency or Laplacian operators, corresponding to different relation types. The filter is constrained to be positive semidefinite via sum-of-squares constructions, enabling both expressivity and theoretical guarantees (He et al., 2023).
Coefficient tensorization: For multi-channel inputs/outputs, the coefficient tensor of the polynomial basis can be decomposed (e.g., CP/Tucker) to balance parameter sharing and expressivity, subsuming a range of prior spectral GCN architectures (Huang et al., 2024).
Multi-resolution and wavelet bases: General spectral wavelet convolutions employ multi-scale bases via admissible Chebyshev polynomial splits, allowing the construction of band-pass and low-pass filters, and the decoupling of short/long-range receptive fields (Liu et al., 2024).

4. Algorithmic Acceleration and Quantum Enhancements

Several strategies reduce the bottlenecks of $D = \mathrm{diag}(d_1, \dots, d_n)$ 0 eigendecomposition and $D = \mathrm{diag}(d_1, \dots, d_n)$ 1 matrix-vector multiplications:

Chebyshev recursion: Each filter power can be recursively computed, fully avoiding eigendecomposition and enabling application to large sparse graphs in $D = \mathrm{diag}(d_1, \dots, d_n)$ 2 time (Bruna et al., 2013, Chen, 2020, Henaff et al., 2015).
Iterative rational filtering: Jacobi or similar iterative schemes are used to compute $D = \mathrm{diag}(d_1, \dots, d_n)$ 3 for rational filters, maintaining linear complexity in the number of edges for fixed filter degree and iteration count (Levie et al., 2017).
Decoupled and precomputed propagators: In the heterogeneous setting, monomial compositions are precomputed and stored, minimizing runtime overhead (He et al., 2023).
Quantum acceleration: Quantum phase estimation and Hamiltonian simulation diagonalize $D = \mathrm{diag}(d_1, \dots, d_n)$ 4 in poly( $D = \mathrm{diag}(d_1, \dots, d_n)$ 5) time, enabling exponential speedup in the node dimension for spectral convolutional layers at the cost of approximation error $D = \mathrm{diag}(d_1, \dots, d_n)$ 6 (Ye et al., 9 Mar 2025).

Algorithmic Strategy	Classical Complexity	Quantum/Enhanced Complexity
Naive eigendecomp	$D = \mathrm{diag}(d_1, \dots, d_n)$ 7
Chebyshev (poly $D = \mathrm{diag}(d_1, \dots, d_n)$ 8)	$D = \mathrm{diag}(d_1, \dots, d_n)$ 9
Rational (iterative)	$d_i = \sum_j A_{ij}$ 0
Quantum phase est.		$d_i = \sum_j A_{ij}$ 1 (Ye et al., 9 Mar 2025)
Heterogen. (precomp)	$d_i = \sum_j A_{ij}$ 2

5. Transferability and Stability Across Graphs

A pervasive misconception held that spectral graph filters do not transfer across graphs of different topology or size due to potential instability in eigenbases. Recent theoretical and empirical results refute this:

Stability to graph perturbations: Spectral filters parametrized as smooth functions of the Laplacian (Chebyshev, Cayley, low-degree polynomial) are stable under small Laplacian perturbations, ensuring that $d_i = \sum_j A_{ij}$ 3 if $d_i = \sum_j A_{ij}$ 4 is Lipschitz (Nilsson et al., 2020, Levie et al., 2019).
Transferability across discretizations: If two graphs discretize the same underlying domain, a fixed spectral filter has similar repercussions on both up to a consistency term vanishing as graph sizes grow. This holds for both small and large topology changes (Levie et al., 2019, Nilsson et al., 2020).
Empirical evidence: ChebNet and CayleyNet match or outperform spatial GNNs on multigraph tasks (classification, regression, property prediction) using a single set of spectral coefficients without retraining per graph (Nilsson et al., 2020).

Transfer Setting	Transferability Guarantee
Small perturb/graph edit	Filter error $d_i = \sum_j A_{ij}$ 5
Large topology change (common domain)	Error $d_i = \sum_j A_{ij}$ 6
Multi-graph ChebNet filter	Consistent performance across diverse graphs

The only scenario causing instability is naive indexing of frequency responses by eigenvalue order rather than using smooth spectral functions (Levie et al., 2019).

6. Recent Advances: 2-D Convolutions, Set-to-Set Filters, and Universality

Spectrum-based convolutions have seen substantial recent generalization beyond traditional vector filters:

Two-dimensional graph convolutions: Classical spectral GNNs operate on features with independent or shared filters per channel. The 2-D paradigm parameterizes the full $d_i = \sum_j A_{ij}$ 7 convolution with one spectral operator per input-output channel pair. ChebNet2D uses Chebyshev interpolated filters reduced to $d_i = \sum_j A_{ij}$ 8 parameters, yielding universality—the ability to construct arbitrary output for any input—and unifying all prior paradigms (Li et al., 2024).
Set-to-set spectral filters: Specformer models the entire spectrum as a set and applies self-attention, constructing a set-to-set spectral filter equivariant to permutations, deeply expanding the flexibility/bandwidth of spectral GNNs (Bo et al., 2023).
Wavelet and multi-resolution filters: WaveGC uses Chebyshev-split wavelet constructions to create admissible band- and low-pass filters, paired with matrix-valued (transformer-style) spectral kernels, achieving flexible decoupling of local/global behavior (Liu et al., 2024).

7. Practical Implications, Applications, and Theoretical Guarantees

Spectral graph convolution is fundamental to a broad class of architectures and tasks:

Node and graph classification/regression: State-of-the-art results for node classification, link prediction, and graph-level regression tasks using polynomial/rational spectral filters (Nilsson et al., 2020, Henaff et al., 2015, He et al., 2023, Parisot et al., 2017).
Heterogeneous/multigraph settings: Positive noncommutative spectral filters (PSHGCN) achieve SOTA on large real-world multi-relational graphs (He et al., 2023).
Biomedical applications: Spectral CNNs for cortical surface parcellation deliver near state-of-the-art accuracy with orders-of-magnitude faster runtime than classical methods (Gopinath et al., 2018).
Oversmoothing control: Corrected graph convolution operators (removing the principal eigenvector) rigorously limit the adverse "oversmoothing" effect, enabling arbitrarily deep spectral filtering without collapse of discriminative information and provable exponential error decay up to optimal information-theoretic thresholds (Wang et al., 2024).
Quantum computation: Quantum GCNs using phase estimation and quantum parallelism achieve exponential per-layer complexity reductions, suggesting future directions for large-scale graph analytics (Ye et al., 9 Mar 2025).

Spectral graph convolution is established as a mathematically principled, highly expressive, and transferrable building block, forming the backbone of classical GCNs, ChebNet/CayleyNet, spectral Transformers, and recent universal-approximation GNNs. Its extensions to heterogeneous, directed, and complex-valued settings have resolved key barriers to generalization, transfer, and applicability on real-world, large-scale, and non-Euclidean data.