Graph Spectral and Laplacian Equivalence

Updated 13 April 2026

Graph spectral and Laplacian equivalence is the study of how Laplacian spectra capture graph structure, connectivity, and topological invariants.
By leveraging affine relations and spectral invariants, the framework distinguishes properties such as regularity, bipartiteness, and isomorphism classes.
This equivalence underpins spectral algorithms and learning techniques used in clustering, quantum graphs, and manifold approximations.

Graph spectral and Laplacian equivalence concerns to what extent the spectrum of graph matrices—most importantly, the combinatorial and normalized Laplacians—determines the structure and properties of a graph, and under what conditions two graphs or operators are spectrally indistinguishable. This topic straddles algebraic graph theory, spectral geometry, spectral algorithms, learning, quantum graphs, and graph isomorphism testing, with profound connections to topological invariants, functional analysis, and applied machine learning.

1. Standard Laplacians, Operator Families, and Spectral Equivalence

Let $G = (V, E)$ be a graph with degree matrix $D$ and adjacency $A$ . The unnormalized Laplacian is $L = D - A$ , with all eigenvalues real and non-negative; for the normalized Laplacian $\mathcal L = I - D^{-1/2} A D^{-1/2}$ , the spectrum lies in $[0,2]$ (Lutzeyer et al., 2017). Both are symmetric, positive semidefinite, and foundational to spectral theory.

Two graphs are Laplacian-cospectral (L-cospectral) if their Laplacians have identical spectra. Spectral equivalence classically asks when L-cospectral graphs are isomorphic ("Spectra determine the graph"). For connected graphs, the multiplicity of eigenvalue zero equals the number of components; thus the spectrum encodes basic connectivity. For properties such as regularity and bipartiteness, specific spectral signatures exist. For instance, the Laplacian spectrum distinguishes a $k$ -component graph via $k$ zero eigenvalues, and for bipartite graphs, the normalized Laplacian spectrum is symmetric about 1 (Kumar et al., 2019).

Affine transformations relate the spectra of adjacency, unnormalized Laplacian, and normalized Laplacian operators. In $d$ -regular graphs these relations are exact: if $\mu_i$ are adjacency eigenvalues, then Laplacian eigenvalues satisfy $D$ 0, normalized Laplacian $D$ 1 (Lutzeyer et al., 2017). Away from regularity, tight bounds quantify deviations in terms of degree extremes ( $D$ 2), showing that large degree heterogeneity leads to significant spectral divergence, which can affect downstream inference and clustering (Lutzeyer et al., 2017).

2. Rigidity, Isospectrality, and Characterization Results

A central question is whether the Laplacian spectrum uniquely determines graph structure (DLS property, "determined by its Laplacian spectrum"). The conjecture of van Dam and Haemers posits that “almost all graphs” are DLS, though explicit families where DLS is provable are rare (Abdian et al., 2019).

For path-friendship graphs—constructed as a coalescence of a friendship graph $D$ 3 (a set of $D$ 4 triangles joined at a central vertex) and a starlike tree at their unique high-degree vertex— it has been shown that the Laplacian spectrum completely determines the isomorphism class. The proof leverages:

Spectral invariants such as vertex count, edge count, and degree sequence rigidity.
Triangle counts and high-degree vertex uniqueness, encoded in the third characteristic polynomial coefficient.
Uniqueness of the arm-lengths in the starlike component, via largest Laplacian eigenvalue rigidity (Guo’s lemma, Feng–Yu–Omidi–Tajbakhsh result).

In these graphs, the Laplacian spectrum pins down the structure so tightly—number of triangles, unique degree sequence, and exact subgraph configuration—that no non-isomorphic mate can exist (Abdian et al., 2019).

For Laplacians on quantum graphs (metric graphs with δ/δ′-vertex couplings), under rational independence of edge lengths, the spectrum almost always determines both the graph and the vertex couplings uniquely; the only strict exception is the three-vertex chain (A3) under mixed δ–δ′–δ couplings (Ershova et al., 2014). This follows from the joint analyticity of the spectral determinant and the M–function method, exploiting full cycle–sum expansions.

In bipartite metric graphs, the standard (Kirchhoff) and anti-standard (anti-Kirchhoff) Laplacians are isospectral up to a shift in the zero eigenvalue multiplicity by the first Betti number—their positive spectra coincide exactly (Kurasov et al., 2019). The underlying proof is an operator factorization: the standard Laplacian is $D$ 5, anti-standard is $D$ 6; their nonzero spectra coincide.

3. Equivalence Under Transformations and Algorithmic Implications

Affine relationships between spectra of the adjacency matrix, Laplacian, and normalized Laplacian facilitate direct eigenvalue and eigengap comparisons, crucial for robust application in graph signal processing and spectral clustering (Lutzeyer et al., 2017). Tightly quantified bounds ensure that, in low degree-heterogeneity regimes, distinct Laplacians yield nearly identical spectral information and thus similar algorithmic outcomes (clustering, denoising, signal representation). With increasing degree variability, the choice of operator becomes critical, potentially leading to divergent algorithmic output.

The general inner-product Laplacian framework further unifies the combinatorial and normalized Laplacians, hypergraph Laplacians, and directed graph analogues as special cases of a single abstract Laplacian defined via arbitrary positive-definite vertex and edge inner products (Aksoy et al., 14 Apr 2025). This framework subsumes classical results and enables generalized isoperimetric inequalities (Cheeger, expander mixing), as well as encoding domain-specific knowledge via the choice of inner product.

In learning, spectral constraints on the Laplacian spectrum directly encode structured graph families (e.g., enforcing a given clustering or regularity), transforming combinatorial constraints into tractable algebraic spectral constraints (Kumar et al., 2019).

4. Topological and Functional Equivalence: Hodge, Betti, and Quantum Graphs

Spectral equivalence is tightly linked to topological graph invariants. For instance, in Barycentric-refined graphs (simplicial complexes), the (combinatorial) Laplacian spectrum and the spectrum of the connection Laplacian are similar to the Hodge Laplacian: the Betti numbers (numbers of connected components and cycles) correspond to the multiplicities of ±1 eigenvalues. Explicitly, the nullity of the Laplacian equals the 0-th Betti number ( $D$ 7), and for 1-dimensional complexes, the spectrum of the connection Laplacian $D$ 8 and the Kirchhoff Laplacian are connected via the “Hydrogen = Hodge” similarity $D$ 9 (Knill, 2018).

In metric or lattice graph settings, the kernel of the Laplacian counts components, and the structure of Dirac-type operators reflects higher Betti numbers, as in the Wilson fermion case, which physically removes unwanted zero modes (species doublers) except those mandated by the underlying Betti structure (Yumoto et al., 2023).

For quantum graphs, Laplacian spectrum determines not only the graph connectivity but also, generically, the precise vertex couplings, except in a single degenerate configuration; this sharpens the notion of spectral equivalence from a purely combinatorial to a functional-analytic one (Ershova et al., 2014).

5. Spectral Equivalence, Graph Isomorphism, and the Weisfeiler–Leman Hierarchy

Laplacian (or adjacency) cospectrality is strictly weaker than graph isomorphism but strictly finer than distinguishability by the 2-dimensional Weisfeiler–Leman (2-WL) test: every pair of 2-WL-indistinguishable graphs are necessarily cospectral for all standard graph matrices, but the converse fails—cospectral pairs that are not isomorphic exist, and most are separated already by 2-WL (Rattan et al., 2021).

Moreover, individualizing a vertex and running 1-WL (the (1,1)-WL test) is already sufficient to distinguish all classical spectral invariants. For higher-order distinctions, families of spectral invariants capturing $A$ 0-WL indistinguishability are constructed by considering spectra of higher-order adjacency-type matrices; simultaneous similarity of these operator families exactly characterizes $A$ 1-WL indistinguishability (Rattan et al., 2021).

6. Nonlinear Spectral Equivalence and Graph Cut Objectives

Recently, the equivalence between certain nonlinear spectral minimization problems and combinatorial graph cuts (Cheeger cut, maxcut, mincut, anti-Cheeger, dual Cheeger) has been established via 1-Laplacian (and signless 1-Laplacian) operators: the critical values of associated Rayleigh-type quotients or variational eigenvalue problems coincide exactly with cut constants (Shao et al., 2024). For each such problem, an exact spectral reformulation is available, and Dinkelbach-type schemes yield exact solutions (up to uniqueness and multiplicity). Nodal domain theorems extend to this nonlinear setting.

Higher-order Cheeger-type inequalities generalize these links: the $A$ 2-th 1-Laplacian eigenvalue bounds multiway Cheeger constants, paralleling the classical (linear) higher-order Cheeger relations (Shao et al., 2024).

7. Graph Laplacian Equivalence in Approximation, Learning, and Random Graphs

Polynomial filtering of Laplacian operators—a key step in spectral GNNs and scalable graph learning—can be equivalently realized on Laplacian sparsifiers with guarantees on operator norm differences. For any polynomial $A$ 3, one constructs a sparsifier $A$ 4 such that $A$ 5 closely approximates $A$ 6, with spectral error bounds governed by the spectral approximation parameter $A$ 7 and filter degree. Such equivalence enables efficiency and scalability in large-scale spectral learning tasks (Ding et al., 8 Jan 2025).

In random graph models, the unnormalized/normalized Laplacian spectrum—when constructed from samples of a submanifold—converges (at a quantified rate) to the spectrum of the Laplace–Beltrami operator, establishing a spectral equivalence principle in manifold learning and justifying spectral clustering from sampled data (Wang, 2015).

In discrete Schrödinger settings, spectral difference identities quantify when two Laplacian-type operators have “near equivalent” spectra (regarding potential perturbations), with trace formulae and local Weyl laws yielding precise quantitative and asymptotic equivalences (Bifulco et al., 2024).

In summary, graph spectral and Laplacian equivalence comprises a rigorous suite of operator-theoretic, combinatorial, and analytic results showing when, and in what sense, the spectrum of graph Laplacians (or related matrices) can serve as a unique or near-complete fingerprint for graph structure, topology, and function. This framework spans explicit isomorphism-determining results, equivalences up to finer functional invariants, tight operator-theoretic bounds, algorithmic consequences, and topological interpretations, and underpins contemporary algorithmic and theoretical work in spectral graph theory, learning, and signal processing.