Spectral Sample: Theory and Applications

• Spectral sample is a method defined via spectral data—eigenvalues, Fourier coefficients, or spectral weights—applied across statistics, graph processing, and turbulence simulation.
• It leverages constructions such as Haar-distributed unitary matrices and Fourier-domain operations to mimic classical sampling behaviors and yield unbiased estimators.
• Applications range from free-cumulant asymptotics and temporal decorrelation in Voronoi percolation to improved phase statistics in optical turbulence.

“Spectral sample” is a field-dependent term used for constructions in which sampling is defined through spectral data rather than by direct coordinate selection. In algebraic-combinatorial statistics, it denotes the eigenvalue vector of a principal submatrix of a Haar-randomly conjugated diagonal matrix; in continuum percolation, it denotes a random finite subset of the plane whose law is given by squared Fourier coefficients averaged over random geometry; in graph signal processing, it denotes down- and upsampling performed directly on graph Fourier coefficients; and in optical turbulence simulation, it denotes randomized sampling of spatial frequencies in Fourier space to improve low-frequency phase statistics (Nardo et al., 2013, Vanneuville, 2019, Tanaka, 2017, Paulson et al., 2019).

1. Spectral sampling as an eigenvalue analogue of simple random sampling

Di Nardo, McCullagh, and Senato define spectral sampling from a finite population of real numbers

$$x=(x_1,\dots,x_n), \qquad X=\operatorname{diag}(x),$$

by first applying a Haar-distributed unitary matrix $H$, so that $HXH^\dagger$ is freely randomized, and then taking the leading $m\times m$ principal submatrix

$Y=(HXH^\dagger)_{[m\times m]}.$

The spectral sample of size $m\le n$ is the eigenvalue vector

$y=(y_1,\dots,y_m).$

For $m=n$, $y$ is a random permutation of $x$; for $H$0, the $H$1 are typically nonatomic and do not coincide with the original $H$2 (Nardo et al., 2013).

This construction is presented as the matrix or eigenvalue analogue of ordinary simple random sampling. If the unitary group is replaced by the permutation group $H$3, then $H$4 becomes an ordinary simple random sample from $H$5. The associated notion of a natural statistic is likewise transferred from the classical setting: a statistic $H$6 is natural relative to spectral sampling if

$H$7

In this sense, the statistic is inherited on the average. The paper places this viewpoint alongside Tukey’s classical natural symmetric polynomial statistics, including the normalized augmented symmetric functions

$H$8

with $H$9 (Nardo et al., 2013).

2. Natural statistics, spectral $HXH^\dagger$0-statistics, and free-cumulant asymptotics

The same paper develops the corresponding analogue of Fisher’s $HXH^\dagger$1-statistics and Tukey’s polykays. The central observables are trace powers of a random Hermitian matrix $HXH^\dagger$2,

$HXH^\dagger$3

where $HXH^\dagger$4 are the eigenvalues of $HXH^\dagger$5. Using umbral calculus, cumulants $HXH^\dagger$6 are defined from moments $HXH^\dagger$7 by

$HXH^\dagger$8

and the cumulant umbra satisfies

$HXH^\dagger$9

This symbolic machinery is extended to multivariate sequences and to eigenvalue data (Nardo et al., 2013).

For a permutation $m\times m$0, the paper defines

$m\times m$1

where $m\times m$2 is the set of cycles of $m\times m$3 and $m\times m$4 is the cycle length. The inversion

$m\times m$5

produces the spectral analogue of the classical moment–cumulant relation. The resulting matricial polykays, or spectral $m\times m$6-statistics, are unbiased estimators of products of cumulants of trace powers. For a partition $m\times m$7, the corresponding quantity $m\times m$8 is constructed so that its expectation equals a product of cumulants of the trace process (Nardo et al., 2013).

The asymptotic normalization connects spectral sampling to free probability. The normalized spectral $m\times m$9-statistics are defined by

$Y=(HXH^\dagger)_{[m\times m]}.$0

and satisfy

$Y=(HXH^\dagger)_{[m\times m]}.$1

where $Y=(HXH^\dagger)_{[m\times m]}.$2 are the free cumulants of the limiting noncommutative distribution. Generalized spectral polykays are then obtained by Möbius inversion on set partitions,

$Y=(HXH^\dagger)_{[m\times m]}.$3

with normalized version

$Y=(HXH^\dagger)_{[m\times m]}.$4

A plausible implication is that spectral sampling provides a finite-$Y=(HXH^\dagger)_{[m\times m]}.$5 combinatorial bridge between symmetric-function statistics and free-cumulant asymptotics (Nardo et al., 2013).

3. Annealed spectral sample in Voronoi percolation

In critical Voronoi percolation, the term refers to a different object. Given a bounded measurable Boolean function $Y=(HXH^\dagger)_{[m\times m]}.$6 on colored Poisson–Voronoi configurations $Y=(HXH^\dagger)_{[m\times m]}.$7, and a fixed realization $Y=(HXH^\dagger)_{[m\times m]}.$8 of the underlying point process, the quenched Fourier expansion is

$Y=(HXH^\dagger)_{[m\times m]}.$9

Rather than keeping the quenched spectral measure $m\le n$0, the paper defines the annealed version by averaging over $m\le n$1: $m\le n$2 and the normalized law

$m\le n$3

A random variable with law $m\le n$4 is the annealed spectral sample $m\le n$5. It is a random finite subset of $m\le n$6, i.e. a continuous point process (Vanneuville, 2019).

A crucial identity is the covariance formula for frozen dynamical Voronoi percolation,

$m\le n$7

which diagonalizes temporal decorrelation in terms of spectral-sample cardinality. For the left-right crossing event $m\le n$8 of $m\le n$9, the paper proves

$y=(y_1,\dots,y_m).$0

and

$y=(y_1,\dots,y_m).$1

For the one-arm event $y=(y_1,\dots,y_m).$2, it proves

$y=(y_1,\dots,y_m).$3

together with a clustering estimate showing that if the spectral mass is small, it tends to be spatially concentrated near the origin (Vanneuville, 2019).

The paper’s application is dynamical critical Voronoi percolation. The lower-tail and clustering bounds are used to show that there exist almost surely exceptional times with an unbounded monochromatic component, both for frozen color-resampling dynamics and for dynamics in which Voronoi centers move according to long-range stable Lévy processes (Vanneuville, 2019).

4. Spectral-domain sampling of graph signals

In graph signal processing, spectral sampling is defined directly in the graph Fourier domain rather than in the vertex domain. For a graph Laplacian

$y=(y_1,\dots,y_m).$4

the graph Fourier transform of a signal $y=(y_1,\dots,y_m).$5 is

$y=(y_1,\dots,y_m).$6

The paper argues that conventional vertex-domain sampling preserves retained values, but does not in general produce the classical frequency-domain effects of bandwidth expansion, aliasing, compression, or imaging on graphs (Tanaka, 2017).

For downsampling by a factor $y=(y_1,\dots,y_m).$7, the index-based spectral downsampling operator (GD2) is

$y=(y_1,\dots,y_m).$8

or, in matrix form,

$y=(y_1,\dots,y_m).$9

Because graph eigenvalues are generally nonuniformly spaced, the paper also defines the spectrum-based version (GD3),

$m=n$0

together with symmetrized variants $m=n$1 and $m=n$2 designed to better mimic classical aliasing behavior (Tanaka, 2017).

Upsampling is defined analogously through spectral repetition. The index-based operator (GU2) repeats the original spectrum $m=n$3 times,

$m=n$4

and the spectrum-based operator (GU3) interpolates the repeated spectrum according to the enlarged graph. The key point is that these operators preserve frequency-domain behavior rather than vertex-domain identity. The paper shows that on ring graphs they reduce to classical time-domain sampling, while on path graphs and more general graphs the graph-dependent eigenstructure materially changes the sampling effect. Fractional sampling and Laplacian pyramid representation are presented as applications (Tanaka, 2017).

5. Randomized spectral sampling in optical turbulence simulation

In optical turbulence, randomized spectral sampling is an FFT-based phase-screen synthesis method that randomizes the sampled spatial frequencies instead of evaluating spectral coefficients only on the rigid DFT lattice. The standard synthesis uses

$m=n$5

with

$m=n$6

The randomized method draws

$m=n$7

and replaces the fixed lattice by a randomly shifted one (Paulson et al., 2019).

Using the Fourier transform shift theorem, the inverse transform can be implemented as a standard inverse FFT followed by a phase ramp,

$m=n$8

$m=n$9

Equivalently,

$y$0

This produces phase screens that are no longer strictly periodic and that contain domain-wide low spatial frequency distortions, including tip/tilt-like structure, without introducing a separate subharmonic subgrid (Paulson et al., 2019).

The principal validation metric is the phase structure function. For isotropic turbulence,

$y$1

and the paper compares simulation to theory using RMS relative error

$y$2

For bounded spectra with finite outer scale, it reports that for

$y$3

the randomized method kept the half-domain RMS error between about $y$4 and $y$5, whereas the traditional unrandomized FFT grid ranged up to about $y$6. For generalized non-Kolmogorov power laws with $y$7, the paper proposes a hybrid method

$y$8

supplemented, when needed, by white-noise terms computed from unresolved spectral energy outside the sampled region (Paulson et al., 2019).

6. Terminological scope, later extensions, and recurrent themes

The cited literature does not use “spectral sample” for a single universal mathematical object. Instead, the term is attached to several constructions whose common feature is that the sampling operation is defined on eigenvalues, Fourier coefficients, or spectral weights rather than on raw coordinates, vertices, or pixels.

Setting	Spectral sample object	Defining mechanism
Finite-population matrix statistics	Eigenvalue vector $y$9	Principal submatrix of $x$0
Voronoi percolation	Random finite subset $x$1	Squared Fourier coefficients averaged over $x$2
Graph signal processing	Down-/upsampled graph spectrum	Sampling directly on graph Fourier coefficients
Optical turbulence	Randomized Fourier-grid phase screen	Cell-wise random frequency shifts

A common misconception is to treat these usages as interchangeable. They are not. The spectral sample of Di Nardo, McCullagh, and Senato is an eigenvalue sample associated with unitary conjugation; the annealed spectral sample of Voronoi percolation is a Fourier-analytic random point process; the graph-signal and turbulence literatures use spectral sampling to mean direct manipulation of spectral coefficients or sampled frequencies (Nardo et al., 2013, Vanneuville, 2019, Tanaka, 2017, Paulson et al., 2019).

Later work extends the phrase in yet other directions. In finite-sample spectral graph inference, one observed graph is treated as a spectral sample: a single adjacency matrix $x$3 obtained through a finite random sampling mechanism, with confidence regions constructed for latent eigenspaces and propagated to clustering, centrality, fairness-constrained post-processing, and topological summaries (Gokavarapu et al., 11 Feb 2026). In large-language-model fine-tuning, GradSentry scores each training example independently by the normalized spectral entropy of its per-sample gradient,

$x$4

where

$x$5

This is a per-sample spectral diagnostic rather than a classical spectral sample, but it preserves the same underlying idea that sample-level structure can be exposed through the spectrum of an associated operator (Zhao et al., 26 May 2026).

Taken together, these usages suggest a stable conceptual core: a spectral sample is a sample defined, weighted, or analyzed through spectral information, with the exact object determined by the ambient theory.