Graphon Signal Processing

• Graphon Signal Processing is an operator-theoretic framework that extends classical GSP to dense networks via graph limit theory and continuum models.
• It employs spectral theory and functional calculus to design scalable filters and provide rigorous convergence and sampling guarantees.
• Applications include neural data analysis, large-scale recommendation systems, and network stability, demonstrating computational efficiency and robust performance.

Graphon Signal Processing (GnSP) is an extension of classical graph signal processing into the regime of large, dense networks via graph limit theory. It provides a rigorous, operator-theoretic framework for representing, filtering, sampling, and analyzing signals on families of graphs characterized by common structural patterns. Central to GnSP is the concept of a graphon—a symmetric, measurable function $W : [0,1]^2 \to [0,1]$ that encapsulates the limiting behavior of graph sequences and enables transferability and scalability of signal processing methods. GnSP has established foundational results in spectral theory, filter design, stability analysis, and sampling, and has demonstrated robust performance across applications from generative random networks to biological neural data.

1. Mathematical Foundations: Graphons and Integral Operators

A graphon is a symmetric, bounded, and measurable function $W(u,v): [0,1] \times [0,1] \to [0,1]$ ($W(u,v) = W(v,u)$, $\int_0^1 \int_0^1 |W(u,v)|^2 du dv < \infty$) that serves as the limit of convergent graph sequences (Parada-Mayorga et al., 2024). Functions defined on $[0,1]$ represent "graphon signals" ($x \in L^2([0,1])$).

The graphon shift operator $T_W : L^2([0,1]) \to L^2([0,1])$ is defined by

$(T_W x)(u) = \int_0^1 W(u,v) x(v) dv,$

and possesses the properties of compactness and self-adjointness, admitting a countable real spectrum $\{\lambda_i(T_W)\}$ and orthonormal eigenfunctions $\{\phi_{W,i}(\cdot)\}$ (Ruiz et al., 2020). The spectral theorem yields the Hilbert-Schmidt expansion

$$W(u,v) = \sum_{i=1}^\infty \lambda_i \phi_i(u) \phi_i(v).$$

2. Spectral Theory and the Graphon Fourier Transform

Signals $x \in L^2([0,1])$ can be decomposed via the graphon eigenbasis:

$x(u) = \sum_{i=1}^\infty \hat{x}_i \phi_{W,i}(u),$

where $$\hat{x}_i = \langle x, \phi_{W,i} \rangle_{L^2} = \int_0^1 x(u) \phi_{W,i}(u) du$$ (Ghandehari et al., 2021). This parallels discrete GSP, but employs the continuum spectrum of $T_W$.

Bandlimited signals are defined for cutoff $\omega > 0$ as

$$PW_\omega(W) = \text{span}\{\phi_{W,i} : |\lambda_i(T_W)| \geq \omega\}.$$

For graphon signals $x \in PW_\omega(W)$, $\hat{x}_i = 0$ for $|\lambda_i(T_W)| < \omega$.

A noncommutative extension organizes the Fourier transform by orthogonal projections onto eigenspaces:

$\widehat{x}(\mu) = P_\mu(x),$

where $P_\mu$ projects onto $\ker(T_W - \mu I)$ (Ghandehari et al., 2021). This is analytically equivalent to the operator-valued Fourier transform familiar in group harmonic analysis.

3. Filters, Functional Calculus, and Convergence Results

Graphon filters are constructed via functional calculus:

$H(T_W) = \sum_{k=0}^M a_k T_W^k,$

where each polynomial filter $H$ is equivalent to pointwise spectral multiplication, acting on the $k$th Fourier mode as $H(\lambda_k) \hat{x}_k$ (Ruiz et al., 2020, Beck et al., 2023). Arbitrary continuous functional envelopes $h(\lambda)$ can be applied to the spectrum, generalizing GSP filter families.

Convergence theorems establish that graph Fourier transforms and filters of graph sequences $G_n$ (with adjacency matrices $A_n$) approach their graphon counterparts (Ruiz et al., 2020):

• For $G_n \to W$ (in cut-norm), eigenvalues and eigenvectors satisfy $\lambda_i(A_n)/n \to \lambda_i(T_W)$.
• For bandlimited signals, the discrete GFT coefficients converge to the graphon Fourier coefficients; filtered outputs converge in $L^2([0,1])$.
• The error decays as $O(\|W - W_{G_n}\|_\square^{1/2})$ or $O(n^{-1/2})$ for random graphs.

A Fourier–Galerkin approach constructs low-dimensional surrogates for $T_W$ by projecting onto an orthonormal basis, enabling finite-dimensional implementations that approximate graphon filters (Morency et al., 2020).

4. Sampling Theory and Transferability

GnSP provides a rigorous sampling theory that generalizes the analysis of uniqueness sets for signals on graphs (Parada-Mayorga et al., 2024):

• $\Lambda$-removable set $S^c \subset [0,1]$: There exists constant $A_{S^c}$ such that $\|T_W x\|_{L^2}^2 \leq A_{S^c} \|x\|_{L^2}^2$ for all $x$ supported in $S^c$.
• If $x \in PW_\omega(W)$, $\|T_W x\|_{L^2}^2 \geq \omega^2 \|x\|_{L^2}^2$.
• Sampling theorem: If $S^c$ is $\Lambda$-removable with constant $A_{S^c}$, then $\omega > A_{S^c}$ implies uniqueness (any two $\omega$-bandlimited signals matching on $S$ are identical).

Comparisons of sampling sets across graphs are made via their graphon representations, with removable constants scaling $A_{S_G}^G = N \cdot A_{S_{W_G}}^{W_G}$. Transfer of optimal sampling sets from small to large graphs is facilitated via induced subsets in $[0,1]$, yielding scalable algorithms robust to changes in graph size and node labeling.

Numerical results show that graphon-based sampling set selection substantially outperforms random sampling and closely tracks the optimality of combinatorial greedy approaches as graph size grows (Parada-Mayorga et al., 2024).

5. Application Domains and Computational Advantages

GnSP has demonstrated robust applicability in:

• Spiking neural networks and calcium imaging data, where graphon spectral projections yield trial-invariant, low-dimensional embeddings for stimulus identification (SIP). This framework confers model-based stability to both synthetic networks and biological data, outperforming PCA and discrete GSP in generalization accuracy (Sumi et al., 24 Aug 2025).
• Large-scale recommendation systems, where signal representations and graph neural network architectures derived from graphon models demonstrate stability under network perturbations and scalability as $n \to \infty$ (Ruiz et al., 2020).
• Dynamic and error-prone networks, with group-based Cayley graphons enabling block-diagonal harmonic analysis and convolutional filter design (Beck et al., 2023, Ghandehari et al., 2021).

Computational complexity is dominated by operations on low-rank block matrices (in block-structured graphons), eigendecomposition of small systems, and $O(nk)$ per-sample filtering/projection, rather than $O(n^2)$ or $O(n^3)$ discrete linear algebra (Sumi et al., 24 Aug 2025). This confers scalability for very large networks and stability with respect to network noise and stochastic variability.

6. Stability Theory and Neural Network Extensions

Graphon neural networks (WNNs) generalize classical GNN layers via compositions of graphon filters and Lipschitz nonlinearities. Stability bounds have been established under operator-norm perturbations of the underlying graphon:

$$\|Y' - Y\| \leq L F^{L-1} \left( A_2 + \frac{\pi n_c}{\delta_c} \right) \varepsilon \|X\|,$$

where $L$ is layer depth, $F$ is feature width, $n_c$ is the number of large-magnitude eigenvalues, $\delta_c$ the minimum eigengap, and $A_2$ the Lipschitz constant of the filter (Ruiz et al., 2020). For graphs instantiated from graphons, approximation error decays as $O(1/\sqrt{n})$, yielding uniform stability as network size increases.

A plausible implication is that deeper or wider graphon neural architectures incur proportionally larger stability constants but retain asymptotic robustness, suggesting practical utility for large, uncertain network deployments.

7. Group Symmetries, Cayley Graphons, and Harmonic Analysis

Cayley graphons, defined on compact groups $G$ with Haar measure, use $W(g,h) = \phi(gh^{-1})$ for a class function $\phi : G \to [0,1]$. Their shift operators are convolution operators:

$$T_W f(g) = \int_G \phi(gh^{-1}) f(h) dh = (f * \phi)(g),$$

diagonalizable via the Peter–Weyl theorem into irreducible representations (Beck et al., 2023). This renders group-based graphon analysis compatible with classical Fourier methods in both commutative and noncommutative cases, and provides explicit eigenbases and spectra for structured networks such as stochastic block models and spatially embedded systems.

In summary, Graphon Signal Processing offers a unifying, operator-based theory for signal analysis on large, dynamic, or random networks and provides rigorous convergence, sampling, and stability guarantees. It bridges asymptotic analysis and practical algorithmic design, enabling robust transfer of signal processing pipelines across network instances and scales (Parada-Mayorga et al., 2024, Ruiz et al., 2020, Ghandehari et al., 2021, Beck et al., 2023, Sumi et al., 24 Aug 2025, Ruiz et al., 2020).