Spectral Moments: Analysis & Applications

• Spectral moments are weighted power integrals of spectral data that capture key characteristics like dispersion, skewness, and higher-order statistics.
• They are applied across disciplines—from random matrix theory to time-frequency analysis—to quantify spectra and reveal structural and dynamical properties.
• Their trace representations allow practical computation and inference of spectral properties, aiding in signal processing, graph analysis, and quantum information theory.

Spectral moments are weighted power sums or power integrals of spectral data. Depending on context, the “spectrum” may be an eigenvalue set, a spectral density, a one-point density of a random-matrix ensemble, an analytic power spectral density, or the local spectrum of a time-varying multicomponent signal. In each case the common operation is to weight the spectral variable by a power and average, often with an equivalent trace representation such as $m_k = N^{-1}\mathrm{Tr}\,H^k$. Across the literature, these quantities serve as descriptors of width, kurtosis, return probabilities, circulation, entanglement measures, and finite-energy sum rules, while also providing a bridge between combinatorial, geometric, and operator-theoretic descriptions of spectra (García-García et al., 2018, Zhong et al., 30 Apr 2026, Lilly, 2012, Gubler et al., 2015).

1. Definitions and principal conventions

A standard definition starts from a spectral density $\rho(E)$ and sets the $k$-th moment to

$m_k=\int dE\,\rho(E)\,E^k.$

For a finite-dimensional Hamiltonian $H$ with Hilbert-space dimension $D$, the same quantity can be written as

$M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$

so moments become normalized traces of powers of the operator (García-García et al., 2018). In Gaussian $\beta$-ensembles the same power-sum viewpoint appears as

$m_k=\sum_{i=1}^N \lambda_i^k,$

with $m_1=\mathrm{tr}\,M$ and $\rho(E)$0 (Maciążek et al., 2015).

On graphs and networks, the operator is typically a Laplacian or a random-walk transition matrix. For an undirected connected graph with Laplacian $\rho(E)$1,

$\rho(E)$2

while centralized moments are defined by

$\rho(E)$3

For a graph or $\rho(E)$4-uniform hypergraph with transition matrix $\rho(E)$5,

$\rho(E)$6

which gives a probabilistic interpretation in terms of average return probabilities (Preciado et al., 2010, Tian et al., 29 May 2025).

The literature also uses several specialized variants. In inverse hard-edge problems one studies

$\rho(E)$7

so the moments emphasize the smallest eigenvalues (Maltsev et al., 20 Apr 2026). In non-Hermitian random-matrix theory one uses mixed moments

$\rho(E)$8

which separately track holomorphic and anti-holomorphic structure (Akemann et al., 18 May 2025). In time-frequency analysis one passes from global moments to instantaneous spectral-moment matrices of an analytic signal, so the moments become time-local Hermitian matrices rather than scalar averages (Lilly, 2012).

Low-order moments often have direct statistical interpretations. In the SYK setting, low-order moments characterize the width, kurtosis, skewness, and related features of the energy distribution, while high moments probe spectral tails and edges (García-García et al., 2018). In graph and hypergraph settings, $\rho(E)$9 and $k$0 are linked to degree heterogeneity and triangle-like or clustering structure through return probabilities (Tian et al., 29 May 2025). This recurring dual role—as both algebraic traces and structurally interpretable observables—is one of the main reasons spectral moments reappear across otherwise unrelated fields.

2. Instantaneous spectral moments and local spectral geometry

For a vector-valued signal $k$1 with $k$2 or $k$3, the analytic signal is defined componentwise by

$k$4

and the first three instantaneous spectral-moment matrices are

$k$5

$k$6

$k$7

These matrices are Hermitian, and their time integrals recover the corresponding one-sided spectral-moment matrices. Their scalar traces define the instantaneous signal power, instantaneous frequency, and instantaneous bandwidth through

$k$8

(Lilly, 2012).

The distinctive result in this setting is the exact identification of these kinematic moments with the physical moments of a hypothetical evolving ellipse traced by the signal. If $k$9 is regarded as a continuum of particles on the periphery of a slowly evolving ellipse, one can define instantaneous circulation, angular momentum, and moment of inertia by integrals around the closed curve. The unity theorem states that, for the canonical ellipse assigned by the analytic signal, the instantaneous spectral-moment matrices and the corresponding physical-moment decompositions coincide exactly. Most significantly, the paper summarizes the result as

$m_k=\int dE\,\rho(E)\,E^k.$0

that is, “circulation = frequency $m_k=\int dE\,\rho(E)\,E^k.$1 amplitude$m_k=\int dE\,\rho(E)\,E^k.$2” (Lilly, 2012).

This identification gives the moments a geometric content beyond formal time-frequency descriptors. The quantity $m_k=\int dE\,\rho(E)\,E^k.$3 is the phase-averaged squared radius of the ellipse, $m_k=\int dE\,\rho(E)\,E^k.$4 decomposes into orbital frequency together with precession and tilt contributions, and the angular momentum need not be normal to the instantaneous plane of motion when that plane tilts (Lilly, 2012). A plausible implication is that instantaneous spectral moments furnish a language in which nonstationary oscillations can be analyzed either as signals or as evolving mechanical geometries without changing the underlying invariants.

A related local formulation arises in mode-stirred reverberation, where the $m_k=\int dE\,\rho(E)\,E^k.$5-th spectral moment of the normalized single-sided analytic power spectral density $m_k=\int dE\,\rho(E)\,E^k.$6 is

$m_k=\int dE\,\rho(E)\,E^k.$7

and equivalently

$m_k=\int dE\,\rho(E)\,E^k.$8

through derivatives of the autocovariance at zero lag. The scaled spectral kurtosis is

$m_k=\int dE\,\rho(E)\,E^k.$9

For the exponential one-pole model $H$0, the moments are $H$1 and the spectral bandwidth is $H$2. The same literature analyzes finite-difference bias, aliasing, decimation, noise, EMI, and understirring, with the finite-difference bias depending, to leading order, quadratically on the sampling interval times the stir bandwidth (Arnaut et al., 2024).

3. Random-matrix theory and exact moment problems

In random-matrix theory, spectral moments are used both to define limiting spectral laws and to organize exact finite-$H$3 calculations. For Wigner matrices, the empirical moments

$H$4

converge to the semicircle moments, with $H$5 equal to the Catalan numbers and odd moments equal to zero. The same work proves a central limit theorem for moments of spectral measures, using combinatorial classifications of closed words and “CLT pairs” (Duy, 2014). For symmetric matrices with a rank-one variance profile, the limiting odd moments vanish and the even moments admit the closed form

$H$6

where $H$7; the coefficients are Narayana numbers (Preciado et al., 2014). For sparse random block matrices, the $H$8-th moments are expanded over closed tree-walks, and irreducible partitions lead to the functional relation $H$9, with explicit moment formulas for finite-$D$0 GOE blocks and algebraic equations in the $D$1 limit (Cicuta et al., 2021).

In many-body chaos, the SYK model provides a detailed instance of how moment expansions encode nontrivial combinatorics. The even moments $D$2 are summed over Wick pairings represented by rooted chord diagrams and their intersection graphs. To order $D$3, the moments coincide with those of the continuous Q-Hermite weight, while the first nontrivial $D$4 correction is proportional to the number of triangles in the intersection graph: $D$5 This maps the $D$6 moment problem to triangle counting and leads to explicit formulas for all moments through that order (García-García et al., 2018).

Non-Hermitian ensembles extend the same methodology in several directions. For the real Ginibre ensemble, the real-eigenvalue moments

$D$7

satisfy a second-order difference equation induced by a third-order ODE for the one-point density, and also admit hypergeometric representations and large-$D$8 expansions involving both integer and half-integer powers of $D$9 (Byun et al., 2023). For complex and symplectic non-Hermitian ensembles, mixed moments $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$0 and $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$1 are expressed in terms of planar orthogonal-polynomial data, and in the Hermitian limit the pure holomorphic moments satisfy

$M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$2

The symplectic holomorphic moments decompose into a “complex part” plus an explicit correction term (Akemann et al., 18 May 2025).

The first two moments themselves can have an exact joint law. In Gaussian $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$3-ensembles, the joint PDF of $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$4 is proportional to

$M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$5

with support

$M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$6

so the admissible region is a parabola-bounded domain in the $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$7-plane (Maciążek et al., 2015). At the hard edge of Laguerre $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$8-ensembles, inverse moments probe the smallest eigenvalues, and in the low-temperature limit $M_k=\langle \Tr H^k\rangle,\qquad m_k=\frac{M_k}{D},$9 with $\beta$0, the rescaled moments converge to even Bessel zeta values,

$\beta$1

linking random-matrix spectral moments to Rayleigh functions (Maltsev et al., 20 Apr 2026).

4. Graph, hypergraph, and lattice spectra

For graphs, spectral moments often encode structural information while remaining directly computable from local data. In the Laplacian setting, the first four moments can be written as averages of node-local functions of degrees, triangle counts, quadrangle counts, and neighboring degrees, all measurable from a 2-hop neighborhood. This permits fully distributed estimation of $\beta$2 by average consensus, and also decentralized control of these moments by edge additions and safe deletions. The resulting algorithm monotonically decreases the moment-error objective

$\beta$3

and terminates in finite time at a locally optimal topology while preserving connectivity (Preciado et al., 2010).

For higher-order networks, the same trace definition is applied to random-walk operators on uniform hypergraphs. If $\beta$4 is the transition matrix of the natural walk that selects a hyperedge containing the current vertex and then moves to a uniformly random other member, then

$\beta$5

Since $\beta$6 is the $\beta$7-step return probability to $\beta$8, $\beta$9 is the average return probability over uniformly chosen starting vertices. In this framework, $m_k=\sum_{i=1}^N \lambda_i^k,$0 relates to degree heterogeneity and $m_k=\sum_{i=1}^N \lambda_i^k,$1 senses triangle-like structure and clustering. The paper extends these relations to $m_k=\sum_{i=1}^N \lambda_i^k,$2-uniform hypergraphs through weighted dyadic projections $m_k=\sum_{i=1}^N \lambda_i^k,$3, with explicit formulas for second and third moments in terms of hyperedge counts, subset degrees, and hyper-triads (Tian et al., 29 May 2025).

Finite non-Hermitian lattices add a different structural phenomenon: spectral moments can be boundary-robust even when the spectrum itself is not. For an $m_k=\sum_{i=1}^N \lambda_i^k,$4 Hamiltonian $m_k=\sum_{i=1}^N \lambda_i^k,$5 with complex eigenvalues $m_k=\sum_{i=1}^N \lambda_i^k,$6,

$m_k=\sum_{i=1}^N \lambda_i^k,$7

The “Universal Spectral Moment Theorem” states that, as the linear size $m_k=\sum_{i=1}^N \lambda_i^k,$8 becomes large, the normalized moments converge to bulk-only quantities independent of boundary geometry. The trace $m_k=\sum_{i=1}^N \lambda_i^k,$9 is interpreted as a sum over closed loops of length $m_1=\mathrm{tr}\,M$0, and only sites within roughly $m_1=\mathrm{tr}\,M$1 hops of the boundary lose loops. Consequently,

$m_1=\mathrm{tr}\,M$2

The same work develops a loop-counting theory for finite-size corrections, verifies $m_1=\mathrm{tr}\,M$3 scaling experimentally in one-, two-, and three-dimensional acoustic lattices, and shows that moments up to $m_1=\mathrm{tr}\,M$4 collapse across distinct boundary geometries even when the complex spectra differ strongly (Zhong et al., 30 Apr 2026).

That invariance also has dynamical consequences. A short-time bulk evolution kernel can be expanded in local matrix powers, and the cited work identifies a dispersive-to-proliferative transition governed by moment structure rather than spectral boundary sensitivity. In the reported $m_1=\mathrm{tr}\,M$5-symmetric 1D model, the transition is associated with the sign change of $m_1=\mathrm{tr}\,M$6, and the experiment finds $m_1=\mathrm{tr}\,M$7 (Zhong et al., 30 Apr 2026). This suggests that, in finite non-Hermitian systems, moments can remain faithful bulk descriptors even when eigenvalue loci are strongly geometry dependent.

5. Quantum information and correlated-electron formalisms

In random quantum-state ensembles, spectral moments are directly tied to entanglement observables. For the Bures–Hall ensemble, the moments

$m_1=\mathrm{tr}\,M$8

of the unconstrained ensemble are derived from Pfaffian kernels built from Cauchy–Laguerre biorthogonal polynomials. Christoffel–Darboux formulas for these kernels lead to a three-term recurrence valid for arbitrary real $m_1=\mathrm{tr}\,M$9,

$\rho(E)$00

and differentiating this recurrence in $\rho(E)$01 yields a companion recurrence for entropy-related linear statistics $\rho(E)$02. This framework re-derives the average von Neumann entropy and quantum purity formulas conjectured by Sarkar and Kumar (Wei et al., 1 Feb 2026).

A dynamical extension appears in the ensemble generated by non-intersecting squared Bessel processes. If $\rho(E)$03 are the path positions at time $\rho(E)$04, the unnormalized moment is

$\rho(E)$05

while normalized eigenvalues $\rho(E)$06, $\rho(E)$07, define the reduced-density-matrix moment

$\rho(E)$08

The cited work derives exact finite-$\rho(E)$09 recurrences for $\rho(E)$10 from multiple orthogonal polynomials, Christoffel–Darboux identities, and integration by parts. It also gives a closed formula for the average purity and shows that the limit $\rho(E)$11 recovers the Hilbert–Schmidt ensemble (Huang et al., 18 Jan 2026).

In correlated-electron theory, spectral moments are used as constraints on the one-particle spectral function $\rho(E)$12. The $\rho(E)$13-th moment is

$\rho(E)$14

Within moment-functional based spectral density functional theory (MFbSDFT), the first four moments are emphasized: $\rho(E)$15, $\rho(E)$16, and the deviations in $\rho(E)$17 and $\rho(E)$18 are encoded by $\rho(E)$19 and $\rho(E)$20. These extra matrices represent correlation effects beyond the noninteracting picture and allow Hubbard bands, bandwidth corrections, and satellite peaks to be described. The same formalism generalizes maximally localized Wannier functions to “maximally localized spectral moment Wannier functions” constructed from the first four moments, and applies them to Wannier interpolation of the anomalous Hall effect in fcc Ni (Freimuth et al., 2023).

A plausible implication is that, in quantum-information and correlated-electron settings alike, moment hierarchies act as compressed surrogates for full spectral data: they are rich enough to recover entropy, purity, Hubbard-band structure, and interpolation matrices, yet structured enough to admit recurrences and finite-dimensional auxiliary Hamiltonians.

6. Sum rules, spectral inference, and known limitations

In hadronic physics, spectral moments are central to finite-energy QCD sum rules. For the $\rho(E)$21-meson correlator, the $\rho(E)$22-th moment up to the continuum threshold $\rho(E)$23 is

$\rho(E)$24

and matching these moments to the operator product expansion yields the first three finite-energy sum rules in terms of the Wilson coefficients $\rho(E)$25. Using an improved vector-dominance model constrained by $\rho(E)$26 data, the vacuum continuum onset is found near $\rho(E)$27, while in nuclear matter the zeroth and first moments remain compatible with the sum rules even as the $\rho(E)$28 spectral function undergoes strong broadening and asymmetric deformation. The second moment constrains strange four-quark condensates and indicates strong violation of naive factorization, with $\rho(E)$29 in vacuum (Gubler et al., 2016).

The same general idea—inferring global structure from finitely many moments—appears in other fields, but with sharply different mathematical guarantees. In the SYK analysis, Carleman’s condition is used to argue that the spectral density is uniquely fixed by its moments (García-García et al., 2018). By contrast, in the Gaussian $\rho(E)$30-ensemble only the joint law of the first two moments is obtained exactly, and extending the method to higher moments is obstructed by the complexity of the discriminant domain and its algebraic constraints (Maciążek et al., 2015). This suggests that moment-based reconstruction is highly context dependent: in some ensembles it is complete, while in others even the admissible integration region becomes intractable.

Practical estimation also has domain-specific limitations. In mode-stirred reverberation, finite differencing, aliasing, decimation, additive noise, EMI, and unstirring all perturb the estimated moments and kurtoses; the cited work therefore recommends explicit de-biasing, sufficient oversampling, and high signal-to-noise ratios for accurate extraction (Arnaut et al., 2024). In finite non-Hermitian lattices, moment invariance is asymptotic rather than exact, with finite-size deviations controlled by missing boundary loops and scaling as $\rho(E)$31 (Zhong et al., 30 Apr 2026). These examples show that spectral moments are rarely “just traces”: their value as observables depends on normalization, sampling protocol, operator class, and the regime in which moment truncations or asymptotics are invoked.