Spectral Theory of Finite Graphs

• Spectral definitions on finite graphs are a framework for associating eigenvalues and spectral measures with graph operators, capturing both local and global structural properties.
• They enable graph comparison, quadrature rules, and sampling methods while revealing robust features such as expansion and spectral classes under perturbations.
• Spectral analysis links Cheeger constants, eigenfunction designs, and structural correspondences, offering actionable insights into graph regularity and network design.

A spectral definition on finite graphs refers to the assignment of spectral data—primarily eigenvalues and associated measures—of linear operators canonically constructed from the graph’s combinatorics. This spectral viewpoint is foundational in modern finite graph theory, underpinning graph comparison metrics, quadrature and sampling theory, signal processing on graphs, structural correspondences, and expansion theory. The spectral objects of core interest are the spectra of (possibly normalized) Laplacians and adjacency operators, their associated projection measures, the relevant eigenfunctions, and functionals derived from these quantities.

1. Operators and Spectra Associated to Finite Graphs

Let $G = (V, E)$ be a finite, simple, undirected, unweighted graph with $n = |V|$. Central to spectral analysis are two operator families:

• Adjacency operator $A: \ell^2(V) \to \ell^2(V)$ with $(Af)(v) = \sum_{u : (u,v) \in E} f(u)$.
• (Combinatorial/normalized) Laplacian. The normalized Laplacian,

$L_G = I - D^{-1}A,$

where $D = \mathrm{diag}(d_x)$ is the diagonal degree matrix ($d_x$ the degree of $x$), acts by

$$(L_G f)(x) = f(x) - \frac{1}{d_x} \sum_{y \sim x} f(y)$$

for functions $f : V \to \mathbb R$.

The adjacency and Laplacian operators are real symmetric (hence self-adjoint, diagonalizable on $\ell^2(V)$). Their spectra (collections of real eigenvalues) are algebraic invariants of $G$.

For the normalized Laplacian $L_G$, the eigenvalues can be ordered as

$$0 = \lambda_0(G) \leq \lambda_1(G) \leq \dots \leq \lambda_{n-1}(G) \leq 2.$$

For the (unweighted) adjacency operator, the eigenvalues $\sigma(A) = \{\mu_1, ..., \mu_s\}$ are contained in $[ -\|A\|, \|A\| ]$.

2. Spectral Measures and Local/Global Structure

Given the self-adjoint operator $A$, the spectral theorem yields a unique projection-valued measure $E_A$ such that

$A = \int_{\mathbb R} t \, dE_A(t).$

For any $\xi \in \ell^2(V)$, the spectral measure

$\mu_\xi(B) = \langle E_A(B) \xi, \xi \rangle$

is a positive measure on $\sigma(A)$. For the vertex basis vector $\delta_v$, this defines the vertex (local) spectral measure $\mu_v$, encoding the local spectral profile at site $v$.

In the finite case, these are atomic measures: $$\mu_v = \sum_{j=1}^s \|P_j \delta_v\|^2 \delta_{\lambda_j},$$ with $P_j$ the orthogonal projector onto the $\lambda_j$-eigenspace. The k-th moment $$\int t^k\, d\mu_v(t) = \langle A^k \delta_v, \delta_v \rangle$$ counts closed walks of length $k$ at $v$ (Bruchez et al., 2022).

3. Spectral (Pseudo-)Distance and Asymptotic Stability

The spectrum admits a probabilistic encoding: define the empirical measure

$$\mu(G) = \frac{1}{n} \sum_{i=0}^{n-1} \delta_{\lambda_i(G)}$$

on $[0,2]$ for the normalized Laplacian. To enable quantitative comparison, Gu et al. introduce the Gaussian-smoothed spectral density

$$\rho_G(x) = \frac{1}{n} \sum_{i=0}^{n-1} \frac{1}{\sqrt{2\pi}\sigma} \exp\left(-\frac{(x - \lambda_i(G))^2}{2\sigma^2}\right)$$

with fixed kernel width $\sigma > 0$.

For two graphs $G, H$, define the spectral pseudo-distance

$D(G,H) = \int_0^2 | \rho_G(x) - \rho_H(x) | dx.$

This is a pseudometric on isomorphism classes, vanishing exactly for cospectral graphs (Gu et al., 2014).

Key stability: For graph sequences differing by at most $O(1)$ edits (edge insertions/deletions), $D(G_n, G'_n) = O(1/n)$ as $n \to \infty$ (Gu et al., 2014). Thus, the spectral comparison is robust to bounded perturbations in large graphs.

4. Spectral Classes and Limit Measures

A sequence of graphs $G_n$ with $n = |V(G_n)|$ is said to belong to spectral class $\rho$ — a probability measure on [0,2] — if their empirical spectral measures converge weakly-*: $\mu(G_n) \rightharpoonup \rho,$ i.e., $\int f\,d\mu(G_n) \to \int f\, d\rho$ for all continuous $f$ (Gu et al., 2014).

Canonical families:

• Complete, complete-bipartite graphs, and hypercubes have $\rho = \delta_1$.
• Paths and cycles converge to the measure with density $g(x) = \frac{1}{\pi} \frac{1}{\sqrt{2x - x^2}}$ for $x \in (0, 2)$.
• Certain star-like (petal) graphs have atomic limit measures.

The main theorem (Gu et al. Thm 3.2): any two families differing by $O(1)$ edits and with convergent spectral measures necessarily converge to the same spectral class.

5. Spectral Designs and Quadrature on Graphs

A spectral definition of graphical designs mirrors the spherical harmonics setting. Given eigenfunctions $\{\varphi_k\}$ of a Laplacian (or similar operator), a weighted subset $(S, \{a_w\}_{w \in S})$ is a $K$-graphical design if

$$\sum_{w \in S} a_w \varphi_k(w) = \frac{1}{|V|} \sum_{v \in V} \varphi_k(v), \quad\forall k\leq K.$$

For $k>1$ (non-constant eigenfunctions), the RHS vanishes.

The main result is a dichotomy: either graphical designs comprise many points or the neighborhoods of $S$ exhibit exponential volume growth as a function of graph distance. High-symmetry graphs support small designs integrating large eigenspaces, analogous to spherical $t$-designs (Steinerberger, 2018).

For smooth functions (energy concentrated in low-frequency modes), graphical designs yield quadrature rules with controlled error.

6. Cheeger Constants, Expansion, and Spectral Bounds

The edge-isoperimetric number (Cheeger constant) $h(G)$ is defined as

$$h(G) = \min_{\emptyset \neq S \subsetneq V} \frac{|E(S, \bar S)|}{|S|}.$$

Spectral bounds connect $h(G)$ to Laplacian eigenvalues: $$\frac{\mu_2}{2} \leq h(G) \leq \sqrt{\mu_2 (2 d_{\max} - \mu_2)}$$ where $\mu_2$ is the second-smallest Laplacian eigenvalue and $d_{\max}$ the maximal degree (Abiad et al., 24 Jan 2026).

Interlacing techniques produce further bounds, particularly for regular graphs. For graph powers $G^t$, polynomial interlacing machinery produces lower and upper bounds for $h(G^t)$ in terms of the spectrum of $G$.

Families with high algebraic structure such as distance-regular graphs often attain the spectral bounds exactly, with tight sets achieving equality: $h(G) = \frac{1}{2} \mu_2$ whenever the subset size is $n/2$ (Abiad et al., 24 Jan 2026).

7. Transfer Operators, Correspondences, and Structural Regularity

Finite graph spectral theory extends beyond Laplacians and adjacency operators. On graphs without dead ends, structural correspondences link:

• The transfer operator $T$ on non-backtracking paths,
• The averaging operator $A$ on vertices,
• Dual transfer operators on finitely additive measures,
• Poisson transforms on the boundary of universal covering trees.

Spectra and eigenspace structure of $T$, edge-based operators, and $A$, vertex-based operators, are canonically related: for spectral parameters outside a small exceptional set, eigenfunctions (resonant states) correspond bijectively and exhibit strong regularity, being locally constant on edges (Bux et al., 2023). This suggests that Laplacian/adjacency spectra capture not merely static properties but encode the functional-analytic landscape of flows and paths in the finite graph.

---

The spectral framework for finite graphs thereby provides a unifying analytic structure connecting local, global, asymptotic, and functional properties of graphs. Its axioms and methods are central for comparing graphs, quantifying structural robustness, designing optimal sampling, and bounding expansion. This versatility motivates continuing use and further generalization of spectral definitions in modern graph theory and network science (Gu et al., 2014, Bruchez et al., 2022, Steinerberger, 2018, Abiad et al., 24 Jan 2026, Bux et al., 2023).