Spectral Graph Learning: Origins and Theoretical Limitations

1. Introduction

Graph theory is commonly traced to Euler’s solution of the Seven Bridges of Königsberg problem in 1736 [1]. Spectral graph theory emerged with Kirchhoff’s introduction of the graph Laplacian and matrix-tree theorem [2], before developing through Fiedler’s notion of algebraic connectivity [3] and the modern treatment of Chung [4]. More recently, spectral graph neural networks (Spectral GNNs) have become an influential research direction in graph representation learning, expanding to directed graphs by treating the graph Laplacian as a Hermitian matrix [5], which corresponds to the treatment of directed graphs in [6]. Recent work has questioned both the theoretical foundations and empirical effectiveness of Spectral GNNs [7,8].

This paper takes up that challenge from a different angle. Rather than evaluating Spectral GNNs empirically, we examine the historical development of the concepts they rest on, and trace the origins of the Graph Fourier Transform in particular—the construct that gives "spectral" methods their claim to a Fourier-analytic meaning.

Although the word "spectral" appears in both spectral graph theory and Fourier analysis, the two notions arose from different mathematical traditions. In classical Fourier analysis, a function is decomposed into sinusoidal modes associated with translation symmetry, and the notion of frequency derives directly from these oscillatory patterns. Fourier analysis is consequently tied to translation-invariant domains and to frequencies with a clear geometric interpretation. Spectral graph theory, by contrast, studies graphs through the eigenvalues and eigenvectors of graph-associated matrices such as the Laplacian [9]—a framework that, on its own, carries no such translation structure.

Beginning in the early 2000s, several researchers sought to extend concepts from signal processing, including Fourier and wavelet analysis, to graphs and other irregular domains [10]. This effort culminated in the Graph Fourier Transform, which became the conceptual foundation for three major research programmes: Discrete Signal Processing on Graphs, Graph Signal Processing, and Spectral Graph Neural Networks. Table 1 summarises a general historical progression of Spectral GNNs. We argue that the Graph Fourier Transform arose through a sequence of conceptual generalisations that progressively detached Fourier terminology from the mathematical structures that originally defined it.

2. Limitations of Spectral Graph Theory

Spectral graph theory studies the eigenvalues and eigenvectors of graph-associated matrices, particularly the graph Laplacian L. Two fundamental results are that (i) the multiplicity of the zero eigenvalue equals the number of connected components, and (ii) the second-smallest eigenvalue ${\lambda }_2$ (the Fiedler value) measures graph connectivity [3].

These two results also expose the field’s scope. Because its strongest theorems concern extremal spectral quantities(near-zero eigenvalues and ${\lambda }_2$ in particular), spectral analysis tends to yield qualitative characterisations and bounds rather than fine-grained descriptions of arbitrary graphs. Many of its classical results, moreover, apply cleanly only to structurally special graph families: cycles, stars, complete graphs, bipartite graphs, and expander graphs [4,20]. And critically for its later use in graph learning, spectral graph theory operates purely on connectivity: its input is the graph structure alone, with no mechanism for incorporating node features [15]. Table 2 summarises these results and limitations.

In short, spectral analysis is constrained by the limited information a graph’s structure alone can encode. For special graph families it produces elegant, exact results; but these results are typically derived for idealised settings and do not straightforwardly transfer to the complexity of real-world graphs. This is the theory later work built on — and its limitations set the stage for what follows in Section 4. Further discussion of spectral graph theory is given in Appendix A.

3. Graph Fourier Transform

• Ref. [14] extend traditional discrete signal processing theory to structured datasets by viewing them as signals represented by graphs. they presented GRAPH FOURIER TRANSFORM.
• In graph signal processing, spectral graph theory has been leveraged to define frequency spectra and to construct graph Fourier transforms based on the eigen-decomposition of the graph Laplacian [15]. However, spectral graph theory does not justify treating the eigenvectors of L as equivalent to the classical Fourier basis, and extending this interpretation to graph Fourier transforms leads to fundamental mathematical inconsistencies. Here is a review of graph signal processing [21].
• The graph Fourier transform underpins early spectral GNNs [16]. In addition to the conceptual issues above, these methods are computationally expensive, requiring full eigendecomposition of L. however, referring to this as a “polynomial approximation” is misleading, as it can be more precisely understood through a Vandermonde system interpretation.
  The resulting models are better understood as learned message-passing schemes [8,22].

4. Origins of the Graph Fourier Transform

Having set out the limits of spectral graph theory as a self-contained field, we now turn to how a second, unrelated tradition(Fourier and harmonic analysis), came to be layered on top of it. This section traces that lineage and identifies the precise points at which the term Fourier transform was decoupled from the mathematical content that gives it meaning. Table A1 (Appendix B) summarises the full timeline.

4.1. What Makes a Fourier Transform

The classical Fourier transform on $\mathbb{R}$ decomposes a function $f:\mathbb{R}\to \mathbb{C}$ into sinusoidal components $e^{i\xi x}$. These components represent oscillations that are globally defined across the domain and are naturally associated with translation symmetry. Consequently, the notion of frequency in classical Fourier analysis is tied to translation invariance and to the decomposition of signals into oscillatory modes. Historically, extensions of Fourier analysis were developed in settings that retain some analogue of these underlying structures. In the following sections, we examine the extent to which they are preserved in the constructions that eventually led to the Graph Fourier Transform.

4.2. The First Misapplication: Fourier on Finite Groups in Statistics (1988–1990)

The mathematical roots of this line of work predate the Fourier label. Representation theory developed independently of Fourier analysis [23,24,25], as discussed in Appendix B. The first explicit use of the Fourier name in this setting, as far as we know, is Diaconis [26], who defined

$$\widehat{f}\left(\rho \right)=\sum _{g\in G}f\left(g\right)\rho \left(g\right),$$

(1)

where $\rho$ ranges over the irreducible representations of a finite group G, and called $\widehat{f}\left(\rho \right)$ the Fourier transform of f. The construction was motivated by statistical applications such as ranked data and card-shuffling on $S_n$. The name is not arbitrary: in the abelian case the construction reduces exactly to the discrete Fourier transform and retains identities such as inversion and Plancherel. But outside that special case, the construction departs from classical Fourier theory on every count that matters—its outputs are matrices rather than frequency coefficients, its basis objects are representations rather than sinusoids, and no general notion of frequency exists on a non-abelian finite group. It therefore lacks the geometric and analytic content associated with classical Fourier theory, and, notably, it emerged from probability and statistics rather than from harmonic analysis or signal processing.

4.3. The Second Misapplication: Graph Fourier Transform (2013)

Shuman et al. [15] and Sandryhaila and Moura [14] pushed the generalisation further, introducing the Graph Fourier Transform (GFT) by projecting a graph signal $\mathbf{f}$ onto the eigenvectors of the graph Laplacian: $\widehat{\mathbf{f}}={\mathbf{U}}^{\top }\mathbf{f}$, where $\mathbf{L}=\mathbf{U}\Lambda {\mathbf{U}}^{\top }$ is the eigendecomposition of the graph Laplacian. The justification offered is analogy: since these Fourier modes are also eigenfunctions of the continuous Laplacian, the eigenvectors of the graph Laplacian are proposed as analogous basis functions on a graph. However, unlike the continuous setting, a general graph possesses neither a translation symmetry nor a corresponding notion of frequency. Graph Laplacian eigenvectors are simply eigenvectors of a connectivity matrix, not sinusoids. The term "graph frequency" is consequently an analogy rather than a mathematically derived notion — it borrows the name frequency without inheriting anything that would let it function as one.

4.4. The Third Misapplication: Harmonic Analysis on Graphs (2014)

Bruna et al. [16] extended the GFT framework a stage further, describing it as "Harmonic Analysis on Weighted Graphs." This adds a third conceptual layer on top of an already strained analogy. Classical harmonic analysis requires a locally compact group, translation operators, and a Haar measure [27]; analogous theories can be built on smooth manifolds via the Laplace–Beltrami operator, but a general weighted graph has none of these prerequisites. Its Laplacian $\mathbf{L}=\mathbf{D}-\mathbf{A}$ is a purely combinatorial matrix encoding connectivity, with no group action to speak of. As a result, the term harmonic analysis is transferred to a setting that lacks the underlying structures that give the concept its mathematical meaning.

Taken together, these three stages show a widening gap between name and structure: each step retains the vocabulary of the previous framework while discarding more of its mathematical substance. Whether the resulting analogies are close enough to be mathematically useful, despite lacking rigorous justification, is the central point of debate.

5. Conclusions

This paper made two points. First, spectral graph theory has fundamental limitations: it relies purely on graph structure, ignores node features, and its strongest results, near-zero eigenvalues and the Fiedler value, hold mainly for special graph families rather than graphs in general. Second, combining spectral graph theory with Fourier analysis to produce the Graph Fourier Transform is not mathematically justified: graph Laplacian eigenvectors are eigenvectors of a connectivity matrix, not sinusoids or characters, so "graph frequency" is an analogy rather than a derived concept. Spectral GNNs inherit both problems: the structural limits of spectral graph theory, and an unjustified conceptual leap in calling its eigendecomposition a Fourier transform.

Appendix A.1. Main Applications

The principal applications of spectral graph theory fall into two categories.

Graph Partitioning.

The most prominent application is graph partitioning—referred to variously as graph cutting [4], graph clustering [29], spectral clustering [30] or node clustering [31] in different literatures—which seeks to cut edges in regions of low density. In this thesis, to avoid confusion between node-level and graph-level clustering [29], we use the term graph partitioning to refer to node-level partitioning tasks. Related problems include finding a large hidden clique [32] and graph sparsification. Von Luxburg [30] shows that the multiplicity of the zero (or near-zero) eigenvalue of L L corresponds to the number of connected components, making low-order eigenvectors natural indicators of coarse connectivity structure. However, this framework is primarily aligned with homophilic assumptions, where labels correlate with graph connectivity. It becomes less appropriate in settings where node labels are driven by features or exhibit heterophily, where connected nodes may belong to different classes. Another highly similar branch that does not concentrate on the data but rather its structure is Network Science [33]. This area addresses issues such as uncovering community relations, perceived alliances, quantifying connectedness, or determining the relevance of specific agents . It determines for example the size of the giant component, distribution of component sizes, degree and clique distributions, clustering coefficients, betweeness and closeness centralities, path length, and network diameter.

Appendix B.1. The First Abstraction: Representation Theory of Finite Groups (1896–1927)

Ref. [23] initiated the character theory of finite non-abelian groups, defining characters as traces of irreducible matrix representations.

Ref. [24,34] extended these results to compact Lie groups, proving that every square-integrable function on a compact group decomposes into a series of matrix coefficients of irreducible unitary representations.

At this stage the Fourier name was not applied to these decompositions. The connection to Fourier analysis was structural and algebraic: on an abelian group, irreducible representations are one-dimensional characters that coincide exactly with classical sinusoids. On a non-abelian group, irreducible representations are matrix-valued and carry no notion of frequency, phase, or oscillation.

Table A1. Timeline of the historical development of the Graph Fourier Transform. Highlighted rows indicate major conceptual extensions discussed in this paper.

Year	Progress	Flaw / Overreach	Field
1822 [35]	Fourier series on $\mathbb{R}$; sinusoidal decomposition	None; foundation is rigorous	Mathematical physics
1896 [23]	Character theory of finite non-abelian groups; orthogonality of characters	No Fourier name claimed; decomposition is representation-theoretic	Algebra
1927 [24]	Irreducible unitary representations of compact groups; completeness theorem	No Fourier name claimed; matrix representations replace characters	Algebra / Analysis
1934 [25,27]	Duality theory for locally compact abelian groups; abstract characters	Characters on non-real groups lose geometric (sinusoidal) content	Abstract harmonic analysis
1988 [26]	Defines $\widehat{f}\left(\rho \right)={\sum }_{g}f\left(g\right)\rho \left(g\right)$ and calls it the Fourier transform on finite groups; applies to ranked data	First explicit misapplication: matrix decomposition named Fourier; no frequency, no sinusoids, no translation geometry	Statistics
1990 [36]	Fast algorithms for the finite-group “Fourier transform”	Inherits and entrenches the 1988 misapplication; imports FFT terminology into non-abelian algebra	Statistics / Computation
2013 [15], [14]	Define GFT as $\widehat{\mathbf{f}}={\mathbf{U}}^{\top }\mathbf{f}$; name it Graph Fourier Transform	Second misapplication: group structure abandoned entirely; Fourier name applied to pure matrix projection; no translation/convolution/frequency	Signal processing
2014 [16]	Spectral GNN built on GFT; Section 3.1 explicitly titled "Harmonic Analysis on Weighted Graphs"; claims graph Laplacian "provides a harmonic analysis on the graphs"	Third misapplication: harmonic analysis requires a locally compact group, translation operator, and Haar measure — none of which exist on a general graph; term borrowed from Riemannian geometry and applied to a combinatorial matrix with no underlying symmetry structure	Spectral GNN

Table A1. Timeline of the historical development of the Graph Fourier Transform. Highlighted rows indicate major conceptual extensions discussed in this paper.

Year	Progress	Flaw / Overreach	Field
1822 [35]	Fourier series on $\mathbb{R}$; sinusoidal decomposition	None; foundation is rigorous	Mathematical physics
1896 [23]	Character theory of finite non-abelian groups; orthogonality of characters	No Fourier name claimed; decomposition is representation-theoretic	Algebra
1927 [24]	Irreducible unitary representations of compact groups; completeness theorem	No Fourier name claimed; matrix representations replace characters	Algebra / Analysis
1934 [25,27]	Duality theory for locally compact abelian groups; abstract characters	Characters on non-real groups lose geometric (sinusoidal) content	Abstract harmonic analysis
1988 [26]	Defines $\widehat{f}\left(\rho \right)={\sum }_{g}f\left(g\right)\rho \left(g\right)$ and calls it the Fourier transform on finite groups; applies to ranked data	First explicit misapplication: matrix decomposition named Fourier; no frequency, no sinusoids, no translation geometry	Statistics
1990 [36]	Fast algorithms for the finite-group “Fourier transform”	Inherits and entrenches the 1988 misapplication; imports FFT terminology into non-abelian algebra	Statistics / Computation
2013 [15], [14]	Define GFT as $\widehat{\mathbf{f}}={\mathbf{U}}^{\top }\mathbf{f}$; name it Graph Fourier Transform	Second misapplication: group structure abandoned entirely; Fourier name applied to pure matrix projection; no translation/convolution/frequency	Signal processing
2014 [16]	Spectral GNN built on GFT; Section 3.1 explicitly titled "Harmonic Analysis on Weighted Graphs"; claims graph Laplacian "provides a harmonic analysis on the graphs"	Third misapplication: harmonic analysis requires a locally compact group, translation operator, and Haar measure — none of which exist on a general graph; term borrowed from Riemannian geometry and applied to a combinatorial matrix with no underlying symmetry structure	Spectral GNN

Appendix B.2. Later Explicit Misapplications

Table A1 summarises the chain of misapplications traced in this section. The Graph Fourier Transform and its deployment in spectral GNNs are not the product of a single conceptual error but of three successive generalisations, each of which retained a mathematical name while discarding the structural prerequisites that gave that name its meaning.

The first misapplication originates with [26] in the statistical analysis of permutation data, where the decomposition of a function on a finite non-abelian group into irreducible matrix representations—an object with no sinusoids, no frequency, and no translation geometry—was explicitly called a Fourier transform. This identification was carried forward computationally by [36], entrenching the name in the literature.

The second misapplication is due to [14,15], who extended the Fourier name further to arbitrary graphs by identifying the eigendecomposition of the graph Laplacian with a Fourier transform. At this stage even the representation-theoretic justification—which at least required a group—is abandoned. The graph Laplacian eigenvectors are not characters, not representations, and not sinusoids in any sense; they are eigenvectors of a connectivity matrix.

The third misapplication is due to [16], who additionally invoked the term harmonic analysis to justify the spectral graph neural network construction. Classical harmonic analysis requires a locally compact group, a translation operator, and a Haar measure—none of which exist on a general graph. The use of this term imports authority from a rigorous mathematical tradition and applies it to a combinatorial object that satisfies none of that tradition’s prerequisites.

Taken together, the three misapplications show a consistent pattern: a term is borrowed from a setting in which it is rigorously grounded, the structural prerequisites are quietly dropped, and the name is retained by analogy. The result is that Fourier analysis, harmonic analysis, and convolution all appear in the spectral GNN literature as inherited labels whose mathematical content no longer applies to the objects they describe.