Canonical Isotropic Random Fields

• Canonical isotropic random fields are defined by invariance under the domain's symmetry group, employing intrinsic harmonic functions and irreducible representations.
• They decompose complex random processes into spectral components using tools like Gegenbauer polynomials on spheres and invariant tensor bases in Euclidean settings.
• This framework enables precise statistical, computational, and excursion analyses across high-dimensional Gaussian landscapes and non-Gaussian stable fields.

Searching arXiv for recent and foundational papers on isotropic random fields, spherical fields, tensor-valued isotropic fields, and manifold excursion theory. {"query":"all: isotropic random fields sphere tensor-valued manifolds Schoenberg Bochner Godement excursion probabilities", "max_results": 10} Canonical isotropic random fields are random fields whose laws, covariance kernels, or higher-order structures are organized by the symmetry group of the parameter space, so that the admissible representations are built from the intrinsic harmonic objects of that space: Fourier characters on Euclidean domains, Gegenbauer or Jacobi polynomials on spheres and compact two-point homogeneous spaces, spherical functions on symmetric spaces, or irreducible tensor components for vector- and tensor-valued fields. The phrase is not used in a completely uniform way across the literature. One paper states that it does not introduce a single distinguished “canonical” construction by that name, while others treat canonicality as the decomposition dictated by irreducible representation theory, invariant tensor bases, or intrinsic covariance expansions (Cheng, 2015, Leonenko et al., 2017, Bingham et al., 2021).

1. Symmetry, isotropy, and the scope of canonicality

At the most general level, isotropy is a symmetry statement. On $\mathbb R^d$, it is tied to invariance under translations and rotations; on the sphere $S^d$, it means that covariance depends only on geodesic distance; on a symmetric space $M=G/K$, isotropic covariance kernels are $K$-bi-invariant positive-definite functions (Bingham et al., 2021). For tensor-valued Euclidean fields, isotropy is not absolute but relative to a fixed orthogonal representation $U$ of $O(3)$ on the value space, so that after rotating the argument one also transforms the field values by $U(g)$ (Leonenko et al., 2017). For random cross-sections of homogeneous vector bundles over $G/H$, isotropy is formulated by $G$-invariance of the one-point and two-point tensors (Malyarenko et al., 2021).

This symmetry-based viewpoint explains why the adjective “canonical” usually refers not to a single probabilistic model but to a symmetry-determined parameterization. In Euclidean tensor theory, the canonical ingredients are irreducible $O(3)$-components, invariant tensor bases, and finitely many scalar spectral channels (Leonenko et al., 2017). On spheres and other compact homogeneous spaces, canonicality is expressed through zonal harmonics, spherical harmonics, spin-weighted harmonics, or Jacobi polynomials (Ma et al., 2018, Malyarenko et al., 2021). On general manifolds, Laplace-based Matérn constructions remain natural, but full isotropy in the homogeneous-space sense need not be available globally; this distinction is explicit in the discussion of manifolds and graphs (Bingham et al., 2021).

A recurrent misconception is that isotropy is always equivalent to “radial dependence on distance.” The literature is more precise. On highly symmetric spaces, radial dependence is the manifestation of group invariance; on tensor-valued fields, one must specify how the value space transforms; and on generic manifolds, a Laplace-based field may be canonical without being globally isotropic in the strict homogeneous-space sense (Bingham et al., 2021, Leonenko et al., 2017).

2. Harmonic and spectral foundations

The principal canonical representation theorem in the continuous setting is the Bochner–Godement theorem. For a symmetric space $S^d$0 with spherical dual $S^d$1, every positive-definite isotropic covariance kernel has the form

$S^d$2

where $S^d$3, $S^d$4 is a probability measure on $S^d$5, and $S^d$6 are the positive-definite spherical functions (Bingham et al., 2021). This simultaneously extends Bochner’s theorem on $S^d$7 and the Bochner–Schoenberg theorem on spheres.

On $S^d$8, the corresponding canonical basis is given by Gegenbauer polynomials. A scalar isotropic covariance admits the expansion

$S^d$9

with nonnegative coefficients satisfying the standard summability condition, and the matrix-valued extension replaces the scalars $M=G/K$0 by positive semidefinite matrices $M=G/K$1 (Alegría et al., 2020). On compact two-point homogeneous spaces, the Gegenbauer basis is replaced by Jacobi polynomials $M=G/K$2, with $M=G/K$3 determined by the geometry of the space (Lu et al., 2019).

The same architecture survives in infinite-dimensional codomains. For Hilbert-valued spherical random fields, the scalar angular power spectrum becomes a sequence of nuclear covariance operators $M=G/K$4, and the field has the harmonic expansion

$M=G/K$5

with

$M=G/K$6

while the covariance kernel becomes the operator-valued Schoenberg expansion

$M=G/K$7

The trace norms $M=G/K$8 play the role of reduced power spectra (Caponera, 2022).

This spectral viewpoint is the common backbone of canonical isotropic field theory. The basis functions are geometry-specific, but the structure is uniform: irreducible harmonic blocks carry independent or uncorrelated spectral content, and isotropy collapses the angular dependence to zonal objects determined by the symmetry group (Bingham et al., 2021).

3. Euclidean vector-, tensor-, and stable-field constructions

For homogeneous isotropic tensor-valued fields on $M=G/K$9, representation theory leads to finite-dimensional spectral reductions. A field taking values in a fixed tensor space $K$0 and isotropic with respect to a fixed orthogonal representation $K$1 of $K$2 is parameterized by finitely many scalar isotropic spectral densities; the number of such densities depends on the irreducible decomposition of $K$3 (Leonenko et al., 2017). In the scalar case there is one density, in the isotropic vector case exactly two, and in the worked symmetric rank-2 case five (Leonenko et al., 2017).

The covariance then splits into invariant tensor structures multiplied by radial transforms of the scalar spectra. For vector fields, the classical isotropic form

$K$4

is recovered, with $K$5 given by Yaglom-type formulas involving two finite measures (Leonenko et al., 2017). For symmetric rank-2 tensors, the covariance expands in the five basic covariants $K$6, such as $K$7 and $K$8, together with degree-2 and degree-4 tensor covariants (Leonenko et al., 2017).

A closely related spectral program is carried out through orthogonal random measures in the explicit Euclidean expansions for $K$9-valued and $U$0-valued homogeneous isotropic fields (Malyarenko et al., 2014). There the covariance measure $U$1 is operator-valued, and the field itself is expanded in spherical Bessel functions and real spherical harmonics with random coefficients determined by uncorrelated random measures. That paper also makes the link to finite-dimensional convex compacta explicit: the admissible isotropic spectral densities form convex compact sets, and their extreme points or components determine the number of spectral measures that appear in the canonical decomposition (Malyarenko et al., 2014).

The Euclidean theory also has a non-Gaussian stable branch. Random-time $U$2-stabilized subordination starts from an $U$3-fractional stable field $U$4 with kernel representation

$U$5

and an $U$6-field $U$7 with stationary increments in the strong sense. The subordinated field is

$U$8

and Theorem 2.4 proves that it is again an $U$9-fractional stable field with $O(3)$0 (Jung, 2011). This unifies Indicator FSMs and Substable or SubGaussian constructions within one procedure, and the invariance used is the strong Euclidean rigid-motion invariance of increments

$O(3)$1

In this sense, canonical isotropic stable fields arise from kernel invariance rather than covariance formulas (Jung, 2011).

4. Spherical and compact homogeneous formulations

On spheres, isotropic covariance theory is governed by the Schoenberg–Gegenbauer expansion. For scalar fields on $O(3)$2,

$O(3)$3

and for $O(3)$4-variate fields one has

$O(3)$5

with $O(3)$6 (Alegría et al., 2020). The coefficient sequences are canonical spectral parameters. The same paper derives closed-form Schoenberg coefficients for two benchmark isotropic covariance families on spheres of dimension $O(3)$7: the Chentsov model $O(3)$8 and the exponential model $O(3)$9 (Alegría et al., 2020).

For time-varying isotropic vector random fields on $U(g)$0, the covariance matrix function has the form

$U(g)$1

for $U(g)$2, where each $U(g)$3 is a stationary covariance matrix function on the temporal domain and the series is summable at $U(g)$4 (Ma, 2016). On $U(g)$5 the basis becomes $U(g)$6, and on $U(g)$7 it becomes $U(g)$8 (Ma, 2016). A parallel result extends this from spheres to all compact connected two-point homogeneous spaces $U(g)$9, replacing Gegenbauer polynomials by Jacobi polynomials: $G/H$0 with $G/H$1 determined by whether $G/H$2 is a sphere, real projective space, complex projective space, quaternionic projective space, or the octonionic projective plane (Ma et al., 2018).

A fixed-space matrix-valued characterization for compact two-point homogeneous spaces is given by

$G/H$3

where $G/H$4 are positive semidefinite matrices and $G/H$5 converges (Lu et al., 2019). The same paper identifies the smaller universal class valid on all compact two-point homogeneous spaces: $G/H$6 which collapses the geometry of all such spaces into a single basis (Lu et al., 2019).

The homogeneous-bundle formalism gives the sphere a second canonical description relevant to cosmology. Random cross-sections over $G/H$7 lead to the standard harmonic expansions of temperature and polarization fields, including spin-$G/H$8 decompositions and the $G/H$9-/$G$0-mode splitting

$G$1

which are canonical consequences of isotropy and representation theory on $G$2 or $G$3 (Malyarenko et al., 2021).

5. Riemannian manifolds and excursion asymptotics

For Gaussian fields on smooth Riemannian manifolds, isotropy can be defined intrinsically through geodesic distance. In the smooth global form,

$G$4

while in the local isotropic regime,

$G$5

with $G$6 and $G$7 (Cheng, 2015). The first is a global covariance symmetry; the second is a near-diagonal asymptotic condition sufficient to determine high excursion tails.

In the smooth isotropic case, the key simplification is that the field-induced metric is a scalar multiple of the background metric: $G$8 Consequently, the Lipschitz–Killing curvatures under the induced metric satisfy

$G$9

and the expected Euler characteristic of the excursion set $O(3)$0 becomes explicit: $O(3)$1 The corresponding excursion probability has the same full geometric expansion, up to a super-exponentially small error (Cheng, 2015).

In the non-smooth locally isotropic regime, manifold geometry survives only through volume. For an $O(3)$2-dimensional smooth compact submanifold $O(3)$3,

$O(3)$4

and for a $O(3)$5-dimensional smooth compact submanifold,

$O(3)$6

The local chart analysis shows that geodesic distance is asymptotically Euclidean with quadratic form given by the metric tensor, so the global manifold contribution reduces to the Riemannian volume element (Cheng, 2015).

This establishes a hierarchy between local and global canonicality. Local isotropy fixes the principal tail scale. Smooth isotropy, together with Morse-theoretic regularity, yields the full geometric expansion with boundary terms and Lipschitz–Killing curvatures (Cheng, 2015).

6. Critical points and high-dimensional landscape geometry

The canonical smooth isotropic Gaussian landscape also supports precise Morse-geometric results. In high dimensions on the sphere, isotropic Gaussian fields with covariance

$O(3)$7

admit an exact Kac–Rice reduction to GOE spectral statistics after a metric normalization of the Hessian (Fyodorov, 2013). The expected number of stationary points is governed by the GOE spectral density, while the expected number of minima is governed by the distribution of the largest GOE eigenvalue. In the $O(3)$8-spin spherical model, the magnetic field tunes a transition from exponentially many stationary points to the topologically minimal pair of one minimum and one maximum, a regime described as “topology trivialization” (Fyodorov, 2013).

The same paper shows that the critical crossover is controlled by GOE edge statistics. For minima, the Tracy–Widom law for the largest GOE eigenvalue determines the universal edge-scaled asymptotics near the transition, both for the spherical model and for the Euclidean model with parabolic confinement (Fyodorov, 2013). This makes isotropic random fields a source of canonical random-matrix universality classes.

A complementary local result concerns nearby critical points in smooth stationary isotropic Gaussian fields on $O(3)$9 with covariance

$S^d$00

Conditioning on two nearby gradients vanishing, the paper proves that the Hessian determinant at the nearby critical point is asymptotically equally likely to be positive or negative. It also proves that, at high levels, almost all critical points above the threshold are local maxima and saddle points with index $S^d$01. Consequently, closely paired critical points above a high threshold must asymptotically comprise one local maximum and one saddle point with index $S^d$02 (Marriott et al., 2023).

Together these results show that canonical isotropic Gaussian fields are not only classified spectrally; they also have sharply constrained critical-point geometry, from local max–saddle pairing in Euclidean space to global topology trivialization on high-dimensional spheres (Marriott et al., 2023, Fyodorov, 2013).

7. Statistical, sparse, and computational developments

Isotropy imposes strong constraints on statistical procedures. On $S^d$03, strong isotropy of a field is equivalent to invariance in law of each multipole block under the Wigner matrix action

$S^d$04

and this immediately explains why coefficientwise sparsification in the coordinates $S^d$05 breaks isotropy (Gia et al., 2018). The isotropy-preserving alternative is a hybrid $S^d$06-$S^d$07 penalty acting on whole multipole blocks,

$S^d$08

whose block soft-thresholding solution shrinks each $S^d$09-block by a scalar depending only on the rotationally invariant block norm $S^d$10. If the observed field is strongly isotropic, the regularized field remains strongly isotropic (Gia et al., 2018).

A different route to isotropy-compatible sparsity is the random-zonal-wave representation

$S^d$11

with $S^d$12. Every isotropic spherical random field admits such a representation with $S^d$13, and the construction is explicitly designed to preserve isotropy under sparsification (Greco et al., 29 Jan 2026). The same paper proves that no monochromatic isotropic Gaussian random field can be sparse, and consequently no Gaussian isotropic random field can be strongly sparse (Greco et al., 29 Jan 2026). This suggests that exact sparsity under isotropy is intrinsically non-Gaussian.

Simulation algorithms also exploit canonical spectral structure. The “turning arcs” method constructs scalar- or vector-valued Gaussian fields on $S^d$14 as sums of randomly oriented Gegenbauer waves, each varying along a random arc and remaining constant on orthogonal parallels (Alegría et al., 2020). The method is driven directly by the Schoenberg coefficients, has a Berry–Esseen analysis for choosing implementation parameters, and is reported to be much faster than spherical or hyperspherical harmonic simulation for irregular target grids (Alegría et al., 2020). In the Hilbert-valued setting, the operator-valued sample power spectrum

$S^d$15

is consistent in the high-frequency regime, and quantitative central limit theorems are proved for both $S^d$16 and the reduced spectrum estimator $S^d$17 (Caponera, 2022).

A terminological caution is also needed. In monochromatic isotropic random electromagnetic fields, “canonical” may refer not to a canonical covariance representation but to the canonical, or orbital, momentum density

$S^d$18

In that setting, orbital and spin momenta are found to be identically distributed in magnitude, while their correlation structures differ markedly (Gadeyne et al., 2023). This is a different but symmetry-related use of the term.

Taken together, these developments show that canonical isotropic random fields form a broad symmetry-governed family rather than a single model class. Their defining features are intrinsic harmonic parameterizations, invariant covariance or kernel structure, irreducible spectral blocks, and geometry-driven asymptotics. Across Euclidean spaces, spheres, compact homogeneous spaces, and manifolds, the canonical objects change with the domain, but the organizing principle remains the same: isotropy converts admissible random fields into structured mixtures of the elementary harmonics selected by the symmetry group.