Graphon-Signal Analysis Overview

• Graphon-signal analysis is a framework that defines graphons as continuum limit objects for dense graph sequences, enabling unified spectral analysis and filtering.
• It leverages operator theory to transition finite graph signals into a Hilbert-space setting, ensuring spectral convergence and consistent filter design.
• Extensions to sparse graphs and graph neural networks highlight its practical impact on stability, transferability, and scalable signal processing.

Graphon-signal analysis studies signals on large graphs by passing from finite, labeled networks to continuum limit objects—graphons, graphon signals, and graphon-induced operators—so that spectral analysis, filtering, sampling, and learning can be formulated on a common domain across changing graph sizes, labelings, or realizations. In the foundational formulation, a graphon is a bounded, symmetric, measurable function $W:[0,1]^2\to[0,1]$, a graphon signal is a function $X\in L^2([0,1])$, and the central operator is the integral transform

$(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$

whose spectrum defines the graphon Fourier transform and graphon filters (Ruiz et al., 2020). Subsequent work extended this operator-theoretic viewpoint to finite-to-continuum spectral convergence, graphon-level filter design, sparse generalized graphons, sampling theory, and graph neural networks on graphon-signal spaces (Ruiz et al., 2019).

1. Graphons, graphon signals, and induced continuum representations

A graphon is used in two complementary ways. First, it is the limit object of a convergent sequence of dense graphs. For a finite motif $F=(V',E')$, graph convergence is expressed through homomorphism densities: $\lim_{n\to\infty} t(F,G_n)=t(F,W),$ where $t(F,G_n)$ is defined from the adjacency matrix of $G_n$ and $t(F,W)$ is the corresponding integral over $[0,1]^{|V'|}$ (Ruiz et al., 2019). Second, the same graphon defines a random graph model by sampling latent labels $u_i\in[0,1]$ and drawing edges with probabilities $X\in L^2([0,1])$0, so a single continuum object can simultaneously encode a graph family and generate its finite instances (Ruiz et al., 2020).

A graphon signal is the continuum analogue of a graph signal. In the classical formulation, a graph signal is a pair $X\in L^2([0,1])$1 with $X\in L^2([0,1])$2, whereas a graphon signal is a pair $X\in L^2([0,1])$3 with $X\in L^2([0,1])$4 (Ruiz et al., 2019). This shift from vectors to square-integrable functions is structurally important: it puts graph-domain signal analysis into a Hilbert-space setting and makes graph size variable by construction.

Finite graphs and signals embed canonically into that continuum setting by step functions. If $X\in L^2([0,1])$5 has $X\in L^2([0,1])$6 nodes and the unit interval is partitioned into regular cells $X\in L^2([0,1])$7, then the induced graphon and induced graphon signal are block-constant functions: $X\in L^2([0,1])$8 up to the notational variants used across the literature (Ruiz et al., 2020). This induced representation is the basic mechanism by which graph signals of different sizes become comparable in $X\in L^2([0,1])$9.

A related extension treats graph-signal pairs themselves as points in a graphon-signal space. In that framework, a graphon-signal is a pair $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$0 with $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$1 a graphon and $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$2, and the graphon-signal cut distance compares both kernel and signal modulo a common measure-preserving relabeling (Levie, 2023). This suggests a broader usage of “graphon-signal analysis”: not only spectral signal processing on a fixed graphon, but the study of graph-signal families in a relabeling-invariant continuum state space.

2. Operator-theoretic and spectral foundations

The foundational operator is the graphon shift operator

$(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$3

Because $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$4 is bounded and symmetric, $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$5 is Hilbert–Schmidt, compact, and self-adjoint on $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$6 (Ruiz et al., 2020). As a result, it admits a spectral decomposition

$(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$7

with real eigenvalues ordered by sign and decreasing magnitude, and an orthonormal eigenbasis $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$8 of $(T_WX)(v)=\int_0^1 W(u,v)X(u)\,du,$9 (Ruiz et al., 2019).

The graphon Fourier transform (WFT) is defined by projecting a graphon signal onto that eigenbasis: $F=(V',E')$0 This is the direct graphon analogue of the finite graph Fourier transform $F=(V',E')$1 (Ruiz et al., 2019). Bandlimited graphon signals are then those whose coefficients vanish for all eigenvalues with magnitude below a threshold $F=(V',E')$2, which yields a finite-dimensional active spectrum because nonzero eigenvalues of compact operators accumulate only at $F=(V',E')$3 (Ruiz et al., 2020).

Graphon filters are polynomial functions of the graphon shift operator: $F=(V',E')$4 Thus filtering is diagonal in the WFT domain: $F=(V',E')$5 This reproduces the standard graph-spectral multiplier picture on a continuum domain (Morency et al., 2020).

A later reformulation replaces scalar WFT coefficients by eigenspace projections, which is essential when graphon eigenvalues have multiplicity greater than one. For a graphon represented on a standard probability space $F=(V',E')$6, the graphon Fourier transform is redefined as

$F=(V',E')$7

where $F=(V',E')$8 projects onto the eigenspace associated with the distinct eigenvalue $F=(V',E')$9 (Ghandehari et al., 2021). This removes the non-derogatory assumption that earlier convergence theorems required and makes the transform stable under repeated eigenvalues.

The same projection-based viewpoint becomes especially explicit for Cayley graphons. If $\lim_{n\to\infty} t(F,G_n)=t(F,W),$0 on a second countable compact group $\lim_{n\to\infty} t(F,G_n)=t(F,W),$1, then noncommutative harmonic analysis diagonalizes the graphon operator through the finite-dimensional matrices $\lim_{n\to\infty} t(F,G_n)=t(F,W),$2 indexed by irreducible representations $\lim_{n\to\infty} t(F,G_n)=t(F,W),$3, and graphon eigenvectors are built from Peter–Weyl coefficient functions (Ghandehari et al., 2021). This connects graphon-signal analysis to noncommutative Fourier analysis and shows that graphon “frequencies” can be group-theoretic as well as purely operator-theoretic.

3. Finite-to-continuum convergence and the corrected norm bridge

The central convergence question is whether graph-based spectral objects converge to graphon-based ones. For induced step graphons, the scaling laws are exact: $\lim_{n\to\infty} t(F,G_n)=t(F,W),$4 These identities show that graph eigenvalues must be normalized by $\lim_{n\to\infty} t(F,G_n)=t(F,W),$5, and graph Fourier coefficients by $\lim_{n\to\infty} t(F,G_n)=t(F,W),$6, before comparison with graphon quantities (Ruiz et al., 2019).

Under graph-sequence convergence and signal convergence after suitable permutations, the finite graph Fourier transform converges coefficientwise to the graphon Fourier transform for $\lim_{n\to\infty} t(F,G_n)=t(F,W),$7-bandlimited limit signals, provided the limiting graphon is non-derogatory (Ruiz et al., 2019). The broader graphon signal processing framework then proves convergence of graph filters to graphon filters, first spectrally and then in the vertex domain. For arbitrary graphons, vertex-domain convergence is recovered through eigenspace projection arguments, even when individual eigenvectors are not canonical (Ruiz et al., 2020).

These convergence results depend on a norm bridge between graphon convergence and operator convergence. The relevant statement is Proposition 4: $\lim_{n\to\infty} t(F,G_n)=t(F,W),$8 This inequality links the cut-norm topology, natural in dense graph limit theory, to the $\lim_{n\to\infty} t(F,G_n)=t(F,W),$9 operator norm required for spectral perturbation arguments (Ruiz et al., 2024). The lower bound passes through the $t(F,G_n)$0 operator norm, while the upper bound follows from Janson’s estimate

$t(F,G_n)$1

together with the equivalence between the type-1 and type-2 cut norms (Ruiz et al., 2024).

The proof of this proposition in the original graphon signal processing paper was incorrect, although the proposition itself was correct. The invalid step was the identification of the one-edge homomorphism density

$t(F,G_n)$2

with the Hilbert–Schmidt norm squared

$t(F,G_n)$3

which are different quantities in general (Ruiz et al., 2024). The repair leaves the substantive theory unchanged: spectral convergence of graphs to graphons, convergence of the GFT to the WFT, and convergence of graph filters to graphon filters remain valid (Ruiz et al., 2024).

4. Sparse graph sequences and generalized graphons

Classical graphon signal processing is a dense-graph theory. For a sparse graph sequence, the canonical graphons satisfy

$t(F,G_n)$4

so the sequence converges to the zero graphon in the usual cut distance, making the limiting operator and limiting spectrum trivial (Carvalho et al., 2023). This is the main obstruction to extending dense graphon-signal methods to sparse networks.

To avoid that collapse, sparse graphon-signal analysis replaces graphons on $t(F,G_n)$5 by generalized graphons on $t(F,G_n)$6: bounded, symmetric, nonnegative, integrable kernels $t(F,G_n)$7 with $t(F,G_n)$8 (Carvalho et al., 2023). The decisive normalization is the stretch

$t(F,G_n)$9

and graph comparison is performed in the stretched cut distance $G_n$0 rather than the classical cut distance (Carvalho et al., 2023).

The corresponding graphon operator is

$G_n$1

which remains bounded, self-adjoint, compact, and Hilbert–Schmidt on $G_n$2 (Carvalho et al., 2023). Sparse graph sequences sampled from generalized graphons then converge almost surely in stretched cut distance, and operator convergence in $G_n$3 follows from stretched cut convergence (Carvalho et al., 2023).

This framework changes the natural spectral scaling. For sampled sparse graphs with adjacency matrices $G_n$4, the correct normalization is

$G_n$5

rather than the dense normalization by graph size (Carvalho et al., 2023). Polynomial and continuous spectral filters built from $G_n$6 then converge nontrivially, restoring a usable signal-processing theory for sparse graph sequences.

A practical implication is that graphon-signal analysis does not reduce to one asymptotic regime. Dense sequences are naturally handled by classical graphons on probability spaces, while sparse sequences require generalized graphons on $G_n$7 together with stretch normalization. This suggests that the phrase “graphon-signal analysis” names a family of operator-limit methods rather than a single topology.

5. Sampling, uniqueness, pooling, and graphon-level computation

Sampling theory enters graphon-signal analysis through removable and uniqueness sets. For a graphon $G_n$8, an open set $G_n$9 is removable if there exists $t(F,W)$0 such that

$t(F,W)$1

and the infimum of such constants is denoted $t(F,W)$2 (Parada-Mayorga et al., 2024). If $t(F,W)$3 is removable and $t(F,W)$4, then $t(F,W)$5 is a uniqueness set for $t(F,W)$6: any two $t(F,W)$7-bandlimited graphon signals that agree on $t(F,W)$8 agree on all of $t(F,W)$9 (Parada-Mayorga et al., 2024). Through induced graphon representations of finite graphs and finite sampling sets, the same subset of $[0,1]^{|V'|}$0 can be compared across graphs with different node counts and labelings, and removable constants converge along graph sequences whose induced operators converge (Parada-Mayorga et al., 2024).

The same continuum viewpoint underlies graphon pooling. In graphon pooling, a graph is lifted to a graphon, then reduced by one of three procedures: Method 1 (M1, regular integration), Method 2 (M2, irregular integration), or Method 3 (M3, irregular sampling) (Parada-Mayorga et al., 2022). M1 and M2 generate reduced graphs by integrating the graphon over partition cells; M3 samples the graphon pointwise. Reduced signals are obtained by local interpolation in $[0,1]^{|V'|}$1. For M1 and M2, the reduced graphons converge to the underlying graphon almost everywhere, and the corresponding operators satisfy

$[0,1]^{|V'|}$2

The paper further proves filter perturbation bounds, graphon-neural-network mapping error bounds, and a stability theorem for M1 under graphon approximation error (Parada-Mayorga et al., 2022).

Graphon-level computation also includes direct filter design from the graphon rather than from a sampled graph. One route defines the graphon filter

$[0,1]^{|V'|}$3

as the continuum limit of finite FIR graph filters, and then introduces a Fourier-Galerkin shift operator obtained by projecting the graphon operator onto a finite Chebyshev basis (Morency et al., 2020). The resulting finite-dimensional matrix depends only on the graphon, not on a particular graph realization, and filter taps are designed by solving a least-squares problem

$[0,1]^{|V'|}$4

with truncated SVD recommended for numerical stability (Morency et al., 2020). This makes graphon filters computationally explicit and gives one concrete realization of graphon-level transferability.

Taken together, these results show that graphon-signal analysis is not limited to asymptotic existence theorems. It also provides a design language for sampling sets, coarsened operators, and transferable filters in large graph families.

6. Graph neural networks on graphon-signal spaces

A major extension of graphon-signal analysis treats graph neural networks as functions on graphon-signal spaces. In the message-passing setting, a graphon-signal is a pair $[0,1]^{|V'|}$5, and the graphon-signal cut distance is defined by

$[0,1]^{|V'|}$6

where the same measure-preserving relabeling $[0,1]^{|V'|}$7 acts on both graphon and signal (Levie, 2023). With this metric, finite graph-signals become dense in a compact quotient space, and MPNNs with normalized sum aggregation and Lipschitz message and update maps are Lipschitz continuous on that space (Levie, 2023). The resulting framework yields both a generalization bound and a subsampling stability theorem for MPNNs on arbitrary graph-signal distributions (Levie, 2023).

A later refinement extends this program to multidimensional signals, MPNNs with readout, improved robustness-style generalization bounds, and non-symmetric graphons and kernels, thereby covering directed graphs and non-symmetric interactions (Rauchwerger et al., 25 Aug 2025). This addresses several practical limitations of the original theory and brings graphon-signal analysis closer to contemporary graph machine learning practice.

At the architectural level, graphon neural networks (WNNs) are obtained by replacing finite graph filters with graphon filters inside a layered nonlinear architecture: $[0,1]^{|V'|}$8 Under Lipschitz filter responses, non-amplifying spectral multipliers, and 1-Lipschitz activations, WNNs are stable to graphon perturbations, and the corresponding finite GNNs on deterministic and stochastic sampled graphs inherit bounds whose perturbation terms decay with graph size (Ruiz et al., 2020). This is the operator-theoretic route from graphon uncertainty to finite GNN stability.

Higher-order graph neural networks require a different graphon treatment. In the higher-order setting, the relevant state space is again the graphon-signal space, but expressivity is characterized by a graphon-signal extension of the $[0,1]^{|V'|}$9-WL test via signal-weighted homomorphism densities

$u_i\in[0,1]$0

(Herbst et al., 18 Mar 2025). On that basis, the paper introduces Invariant Graphon Networks (IWNs), a bounded-operator subclass of graphonized IGNs, and proves that order-$u_i\in[0,1]$1 IWNs are at least as powerful as the $u_i\in[0,1]$2-WL test and are universal approximators on compact subsets in $u_i\in[0,1]$3 topologies (Herbst et al., 18 Mar 2025).

The same paper also establishes a central limitation: typical higher-order graphon models are discontinuous with respect to cut distance, and this discontinuity is tied to the very mechanism that gives them $u_i\in[0,1]$4-WL power (Herbst et al., 18 Mar 2025). The important qualification is that transferability can still hold. Even when graph-to-graphon convergence fails for higher-order models, graph-to-graph transfer between independently sampled graphs from the same graphon-signal can remain valid (Herbst et al., 18 Mar 2025). This sharply separates cut-distance continuity from size transferability.

7. Applications, limitations, and broader implications

Graphon-signal analysis was motivated by transferability across graph families, and that theme persists across applications. Early examples include recommendation systems on user-similarity graphs and spectral analysis of air-pollution signals across sensor networks in different cities, where the practical issue is that graph-specific eigensystems are expensive or unstable across related graphs (Ruiz et al., 2019). Graphon filters were then applied to source localization, point cloud classification, and recommendation, with graphon pooling reported to perform significantly better than other pooling approaches when dimensionality reduction ratios between layers are large (Parada-Mayorga et al., 2022).

The framework has also moved beyond classical GSP tasks. In stochastic games with memory, continuum-agent interactions are modeled by a graphon, and the Nash equilibrium is solved by combining graphon spectral decomposition with stochastic Fredholm resolvents; equilibria of finite graph games converge to graphon-game equilibria both for cut-norm-convergent graph sequences and for graph sequences sampled from a graphon (Neuman et al., 2024). Although game-theoretic rather than signal-processing in emphasis, this work shows that graphon-domain spectral methods extend naturally to non-Markovian stochastic dynamics.

A more direct signal-analysis application appears in spiking and biological neural networks. There, graphon-based spectral projections are used as trial-invariant low-dimensional embeddings for the stimulus identification problem, with simulations on stochastic block neural networks and calcium-imaging experiments on modular cultured neuronal networks (Sumi et al., 24 Aug 2025). On the biological dataset reported in that work, graphon-based classification accuracy was $u_i\in[0,1]$5, compared with $u_i\in[0,1]$6 for PCA and $u_i\in[0,1]$7 for a reservoir-computing baseline, while the authors also note that the experimental classification improvement is statistically underpowered (Sumi et al., 24 Aug 2025). This suggests that graphon-signal analysis can serve as a robust architectural prior when the graph realization itself fluctuates across trials.

The main limitations are equally clear in the literature. Much of the foundational theory assumes dense graph sequences; sparse networks require the generalized graphon extension. Many constructions presuppose that the graphon is known or meaningfully specified, whereas graphon estimation itself can be computationally difficult. Low-degree computational lower bounds provide evidence for a statistical–computational gap in graphon estimation: for dense SBM graphons with $u_i\in[0,1]$8, low-degree polynomial estimators cannot beat an error of order $u_i\in[0,1]$9, matching the known polynomial-time frontier up to logarithmic factors, and for Hölder graphons the low-degree rate is polynomially slower than the minimax benchmark for $X\in L^2([0,1])$00 (Luo et al., 2023). A plausible implication is that graphon-signal methods are strongest when the latent graphon is structurally available, well estimated, or imposed by design, rather than inferred from a single arbitrary network without additional assumptions.

Across these strands, graphon-signal analysis has evolved from a continuum limit theory for graph Fourier transforms into a broader operator-based framework for signals on graph families. Its unifying principle is that graph structure, graph signals, and sometimes even graph-learning architectures can be lifted into a shared continuum representation where convergence, transfer, and stability become questions about kernels, operators, and function spaces rather than about any single finite adjacency matrix.