Full-Spectrum GNN (FSpecGNN) Overview

• FSpecGNN is a second-order generalization of spectral GNNs that lifts node signals to the node-pair domain using bivariate eigenvalue filters.
• The model overcomes the 1-dimensional Weisfeiler-Lehman limitation by extending expressivity with low-rank polynomial approximations.
• Empirical evaluations show FSpecGNN’s advantages in heterophilic node classification and substructure counting across diverse benchmark datasets.

Searching arXiv for the named paper and closely related spectral GNN context. Full-Spectrum Graph Neural Network (FSpecGNN) is a second-order generalization of classical spectral graph neural networks in which graph signals are lifted from the node domain to the node-pair domain, and univariate spectral filters over Laplacian eigenvalues are replaced by bivariate filters over eigenvalue pairs. In the 2026 formulation, classical spectral GNNs arise as a diagonal special case of this construction. The model is motivated by a specific limitation of classical spectral GNNs: although they can universally approximate node signals, their expressive power remains bounded by the 1-dimensional Weisfeiler-Lehman test and is mirrored in a lack of universality for higher-order signals. FSpecGNN is introduced to extend spectral filtering beyond that bound while retaining scalable implementations through low-rank polynomial approximation, with reported advantages on heterophilic graph learning and substructure-counting tasks (Wang et al., 7 May 2026).

1. Mathematical formulation

For a graph $G=(V,E)$ with $|V|=n$ and Laplacian $L=U\Lambda U^{\top}$, classical spectral filtering acts on a node signal $x\in\mathbb{R}^{n}$ by first taking the graph Fourier transform $\hat x = U^{\top}x$ and then applying a univariate filter $g:\mathbb{R}\to\mathbb{R}$:

$g(L)x = U\,g(\Lambda)\,U^{\top}x,$

where $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$ (Wang et al., 7 May 2026).

FSpecGNN lifts this construction to node-pair signals. A node-pair signal is written as $E\in\mathbb{R}^{n\times n}$, with vectorization $e=\operatorname{vec}(E)\in\mathbb{R}^{n^2}$. The pair-domain Laplacian operators admit Kronecker eigendecompositions

$|V|=n$0

$|V|=n$1

and the basis vectors $|V|=n$2 diagonalize both operators. The central object is therefore a bivariate spectral filter $|V|=n$3, yielding the convolution

$|V|=n$4

When reshaped back to an $|V|=n$5 matrix $|V|=n$6, the entries are

$|V|=n$7

This construction formalizes the two defining moves of FSpecGNN: lifting from nodes to node-pairs, and extending from eigenvalue-wise filtering to eigenvalue-pair filtering. The paper states that classical spectral GNNs are recovered as a diagonal special case, so the framework generalizes rather than replaces the conventional node-domain operator.

2. Approximation properties and signal class

The approximation-theoretic claim of FSpecGNN is formulated at the level of node-pair signals. Under Theorem 3.4, if $|V|=n$8 has distinct eigenvalues and the input $|V|=n$9 has nonzero projection on every basis element $L=U\Lambda U^{\top}$0, then the linear FSpec layer

$L=U\Lambda U^{\top}$1

can be made to match any target one-dimensional signal $L=U\Lambda U^{\top}$2 by choosing

$L=U\Lambda U^{\top}$3

using a bivariate Lagrange polynomial $L=U\Lambda U^{\top}$4 (Wang et al., 7 May 2026).

The proof outline has two parts. First, $L=U\Lambda U^{\top}$5 is expanded in the joint eigenbasis and $L=U\Lambda U^{\top}$6 is chosen so that no frequency coefficient vanishes. Second, a bivariate polynomial $L=U\Lambda U^{\top}$7 is interpolated so that

$L=U\Lambda U^{\top}$8

The theorem is stated as universal approximation on node-pair signals rather than on ordinary node signals. This suggests that the core gain of the model lies in moving the target function class from first-order graph signals to second-order graph signals.

The abstract further emphasizes that this is particularly beneficial for heterophilic graph learning. In that framing, universality on node-pair signals is not treated as a purely formal extension; it is tied to graph regimes in which pairwise interactions, rather than neighborhood averaging on nodes alone, are important.

3. Expressivity relative to Local 2-GNN

FSpecGNN is explicitly situated relative to Local 2-GNN. Local 2-GNN is defined as an iterative refinement procedure on ordered pairs $L=U\Lambda U^{\top}$9 in which the current label of $x\in\mathbb{R}^{n}$0 is hashed with the multisets of labels of its “row” neighbors $x\in\mathbb{R}^{n}$1 and “column” neighbors $x\in\mathbb{R}^{n}$2. The comparison is not informal: the paper gives both an upper-bound theorem and a matching theorem under assumptions (Wang et al., 7 May 2026).

Theorem 3.7 states that any bivariate polynomial filter of total degree $x\in\mathbb{R}^{n}$3 yields representations that cannot distinguish more pairs than $x\in\mathbb{R}^{n}$4-step Local 2-GNN. The proof sketch shows that $x\in\mathbb{R}^{n}$5 can be unrolled into a shared-weights aggregation over depth-$x\in\mathbb{R}^{n}$6 neighborhoods in the pair graph, paralleling $x\in\mathbb{R}^{n}$7 rounds of Local 2-GNN. This is an important qualification: the model is not presented as unconstrainedly exceeding higher-order local message-passing.

Theorem 3.8 then gives an expressivity match. Under simple spectrum and non-degeneracy of the input spectral components, there exists a polynomial $x\in\mathbb{R}^{n}$8 such that $x\in\mathbb{R}^{n}$9 separates pairs at least as well as the infinite-round, stable Local 2-GNN. The proof sketch relies on three points: the family $\hat x = U^{\top}x$0 spans $\hat x = U^{\top}x$1; distinct outputs can be assigned to each 2-GNN color class by Lagrange interpolation over the $\hat x = U^{\top}x$2 grid of $\hat x = U^{\top}x$3; and this enforces a polynomial $\hat x = U^{\top}x$4 that achieves the required separation.

Taken together, these theorems place FSpecGNN in a narrow but technically precise position. The model is at most as expressive as Local 2-GNN as an upper bound, while also being able, under the stated spectral conditions, to match the stable Local 2-GNN partition. A common misconception is that the move to node-pair spectral filtering automatically implies unrestricted higher-order expressivity; the stated theory is more specific and more bounded.

4. Scalable implementation and low-rank factorization

The direct implementation of pair-domain filtering is exact but expensive. The paper describes Route I as learning $\hat x = U^{\top}x$5 with an MLP on pairs $\hat x = U^{\top}x$6 and implementing it through full eigendecomposition together with $\hat x = U^{\top}x$7 transforms, which is described as impractical for $\hat x = U^{\top}x$8 (Wang et al., 7 May 2026).

The alternative is Route II, a polynomial parameterization

$\hat x = U^{\top}x$9

for which

$g:\mathbb{R}\to\mathbb{R}$0

This costs $g:\mathbb{R}\to\mathbb{R}$1 to compute all $g:\mathbb{R}\to\mathbb{R}$2 with sparse multiplications when $g:\mathbb{R}\to\mathbb{R}$3 is per-feature-channel.

A further reduction uses low-rank decomposition. For a bivariate polynomial filter $g:\mathbb{R}\to\mathbb{R}$4 of degree at most $g:\mathbb{R}\to\mathbb{R}$5 in each variable with coefficient matrix $g:\mathbb{R}\to\mathbb{R}$6, Proposition 3.9 gives a decomposition into $g:\mathbb{R}\to\mathbb{R}$7 univariate terms:

$g:\mathbb{R}\to\mathbb{R}$8

Because each term acts on $g:\mathbb{R}\to\mathbb{R}$9 as $g(L)x = U\,g(\Lambda)\,U^{\top}x,$0, the implementation avoids explicit $g(L)x = U\,g(\Lambda)\,U^{\top}x,$1 operations. For a feature matrix $g(L)x = U\,g(\Lambda)\,U^{\top}x,$2, the per-channel implementation is

$g(L)x = U\,g(\Lambda)\,U^{\top}x,$3

$g(L)x = U\,g(\Lambda)\,U^{\top}x,$4

$g(L)x = U\,g(\Lambda)\,U^{\top}x,$5

The total per-layer cost is $g(L)x = U\,g(\Lambda)\,U^{\top}x,$6, and the paper states that for $g(L)x = U\,g(\Lambda)\,U^{\top}x,$7 this matches classical polynomial filters.

The layer pseudocode follows the same pattern. Given $g(L)x = U\,g(\Lambda)\,U^{\top}x,$8, adjacency $g(L)x = U\,g(\Lambda)\,U^{\top}x,$9, polynomial filters $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$0, mixing weight $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$1, and an output MLP, the layer computes the normalized Laplacian

$$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$2

applies $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$3 to $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$4, mixes with the original features, applies $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$5, sums the rank-$$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$6 components, and feeds the result to a final MLP. A common misunderstanding is that “full-spectrum” necessarily entails explicit node-pair-level computation; the implementation section is designed precisely to avoid that requirement.

5. Empirical profile

The empirical evaluation emphasizes heterophilic node classification and substructure-counting tasks (Wang et al., 7 May 2026). The heterophilic datasets are organized by edge homophily $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$7–$$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$8. The small datasets are Texas and Wisconsin, described as 2K–22K nodes; Chameleon and Squirrel, described as 10K–12K nodes; and the large datasets Roman, Minesweeper, Tolokers, and Questions, described as 50K–150K nodes. The setup uses sparse stratified $$g(\Lambda)=\operatorname{diag}(g(\lambda_1),\dots,g(\lambda_n))$$9 train/validation/test splits, two FSpecGNN variants using Cheb, ChebII, or Bern polynomial bases with $E\in\mathbb{R}^{n\times n}$0, $E\in\mathbb{R}^{n\times n}$1–$E\in\mathbb{R}^{n\times n}$2, hidden dimension $E\in\mathbb{R}^{n\times n}$3–$E\in\mathbb{R}^{n\times n}$4, and early stopping with 200 patience. The metric is Accuracy on small datasets and ROC-AUC on large datasets.

The reported test performance for the two FSpecGNN variants is as follows.

Dataset	FSpecGNN(Cheb)	FSpecGNN(Bern)
Texas	55.3±4.6	56.1±0.5
Wisconsin	49.9±8.3	54.6±9.5
Chameleon	33.1±0.9	37.9±3.9
Squirrel	39.6±0.7	37.6±1.3
Roman	54.4±0.7	56.2±1.2
Minesweeper	88.3±0.8	84.2±1.0
Tolokers	76.9±0.9	74.5±0.7
Questions	75.9±0.4	77.1±0.3

The same evaluation table also reports ChebNet and BernNet baselines. The abstract summarizes the outcome as validating the predicted expressivity and delivering strong performance on heterophilic benchmarks.

A second empirical track studies substructure-counting benchmarks, including homomorphism and (chordal) cycle counts at graph, node, and edge level, evaluated with normalized MAE. The reported highlights are that FSpecGNN matches or outperforms local 2-GNN and FGNN in most tasks, with MAE on the order of $E\in\mathbb{R}^{n\times n}$5–$E\in\mathbb{R}^{n\times n}$6, while runtime and memory are described as approximately $E\in\mathbb{R}^{n\times n}$7 faster per iteration together with the lowest peak GPU memory versus Local 2-GNN and FGNN.

6. Relation to adjacent spectral and spectral-spatial GNN lines

FSpecGNN belongs to a broader family of models that use Laplacian eigenstructure, but neighboring papers instantiate different mechanisms. “Graph Networks with Spectral Message Passing” introduces the Spectral Graph Network, which maintains two parallel graphs on each example: a spatial graph on the original nodes and a fully connected spectral graph whose vertices correspond to Laplacian eigenvectors. The two domains are coupled by eigenpooling, $E\in\mathbb{R}^{n\times n}$8, and eigenbroadcasting, $E\in\mathbb{R}^{n\times n}$9, with message passing applied in both domains. That model reports results on Graph-MNIST, MoleculeNet-HIV, QM9, and shortest-path problems, and emphasizes efficient training, robustness to edge dropout, and the use of a small number of eigenvectors such as $e=\operatorname{vec}(E)\in\mathbb{R}^{n^2}$0 or $e=\operatorname{vec}(E)\in\mathbb{R}^{n^2}$1 (Stachenfeld et al., 2020).

By contrast, “Bridging the Gap between Spatial and Spectral Domains: A Unified Framework for Graph Neural Networks” is a survey and taxonomy rather than a concrete FSpecGNN instantiation. It divides existing models into spatial families A-1, A-2, A-3 and spectral families B-1, B-2, B-3, establishes correspondences such as A-1$e=\operatorname{vec}(E)\in\mathbb{R}^{n^2}$2B-1, A-2$e=\operatorname{vec}(E)\in\mathbb{R}^{n^2}$3B-2, and A-3$e=\operatorname{vec}(E)\in\mathbb{R}^{n^2}$4B-3, and explicitly does not define, optimize, or evaluate a single architecture called “FSpecGNN” (Chen et al., 2021).

A later related direction is SpecSphere, a dual-pass spectral-spatial GNN with a Chebyshev-polynomial spectral branch, an attention-gated spatial branch, fusion by a lightweight MLP, a cooperative-adversarial min-max objective, closed-form robustness certificates, and a theorem claiming expressivity strictly beyond 1-WL while retaining linear-time complexity (Choi et al., 13 May 2025). Relative to these nearby lines, FSpecGNN is distinguished by its second-order, node-pair-domain formulation and by its use of bivariate spectral filters over eigenvalue pairs rather than either a coupled spatial/spectral branch architecture or a survey-level equivalence framework.