Source-linked AI summary
Isotropic Gaussian random fields on the sphere: Regularity, fast simulation and stochastic partial differential equations
Annika Lang, Christoph Schwab
TL;DR
The paper addresses how isotropic Gaussian random fields on spheres can be characterized, analyzed for regularity, simulated, and used in stochastic partial differential equations. It develops spherical-harmonic and Karhunen–Loève methods, relates angular-spectrum decay to regularity, and proves spectral convergence results for the stochastic heat equation. The resulting framework covers S2 and corresponding higher-dimensional spheres, with applications motivated by environmental and cosmological modeling.
Problem
General Euclidean-space theory does not apply directly to practically relevant Gaussian random fields on spheres, motivating dedicated results on regularity, simulation, and stochastic partial differential equations.
Method
The paper uses spherical-harmonic Karhunen–Loève expansions to analyze isotropic Gaussian fields, construct isotropic Q-Wiener processes, and discretize the stochastic heat equation spectrally.
Results
The paper characterizes sample Hölder continuity and differentiability from angular-spectrum decay and proves strong convergence rates for spherical-harmonic spectral discretizations of the stochastic heat equation.
Takeaways & Limitations
Angular-power-spectrum decay provides a common basis for studying covariance smoothness, sample regularity, finite-expansion approximation, and stochastic-heat-equation discretization on spheres.
Takeaways & Limitations
For previously available estimates, the authors state that they cannot derive Lp or almost-sure convergence rates for truncated fields, and Borel–Cantelli applicability is unclear.
Abstract
from arXiv · showhide
Isotropic Gaussian random fields on the sphere are characterized by Karhunen-Loève expansions with respect to the spherical harmonic functions and the angular power spectrum. The smoothness of the covariance is connected to the decay of the angular power spectrum and the relation to sample Hölder continuity and sample differentiability of the random fields is discussed. Rates of convergence of their finitely truncated Karhunen-Loève expansions in terms of the covariance spectrum are established, and algorithmic aspects of fast sample generation via fast Fourier transforms on the sphere are indicated. The relevance of the results on sample regularity for isotropic Gaussian random fields and the corresponding lognormal random fields on the sphere for several models from environmental sciences is indicated. Finally, the stochastic heat equation on the sphere driven by additive, isotropic Wiener noise is considered, and strong convergence rates for spectral discretizations based on the spherical harmonic functions are proven.
1. Introduction.
The paper develops theory and numerical analysis for isotropic Gaussian random fields on spheres, motivated by applications where general Euclidean-space results do not apply directly. It connects covariance smoothness and angular-spectrum decay to sample regularity and studies spectral approximations of a stochastic heat equation.
- Motivation: General second-order random-field theory does not apply directly to spherical Gaussian random fields because its formulation requires a group structure on the realization space.The sphere’s invariance under group actions motivates a separate treatment.
- Scope: The paper develops sample-regularity, stochastic-partial-differential-equation, and numerical-analysis results for Gaussian random fields on S2 and higher-dimensional spheres.The main focus is S2, while Sections 2–6 also treat Sd−1.
- Contributions: It connects covariance-kernel smoothness with angular-power-spectrum decay and characterizes almost-sure sample Hölder continuity and differentiability.These results relate spectral behavior to regularity of isotropic fields.
- Contributions: The stochastic heat equation on S2 is represented using spherical harmonics and solved with isotropic Q-Wiener noise.The noise is additive and isotropic in the spectral formulation.
- Contributions: The fully discrete stochastic-heat-equation approximation has a convergence rate determined only by angular-power-spectrum decay, independently of space and time discretization.The paper also indicates fast simulation through spectral expansions and numerical examples.
2. Isotropic Gaussian random fields on the sphere.
The paper models isotropic Gaussian random fields on the sphere through spherical harmonics and Karhunen–Loève expansions. Isotropy yields structured random coefficients, while spherical harmonics provide the basis for representation, analysis, and simulation.
- Definitions: A real-valued random field on S2 is a measurable mapping from a probability space and the sphere into R.The paper focuses on real-valued fields after introducing the general framework.
- Definitions: Strong isotropy requires rotational invariance of all finite-dimensional distributions under SO(3).n-weak isotropy instead imposes finite nth moments at every point.
- Spherical harmonics: Spherical harmonic functions form an orthonormal basis of L2(S2), and every L2 function admits a spherical-harmonic series expansion.The spherical Laplacian has these functions as eigenfunctions with eigenvalues −ℓ(ℓ+1).
- Karhunen–Loève expansion: Every 2-weakly isotropic random field admits a convergent Karhunen–Loève expansion in spherical harmonics.The expansion converges in L2(Ω;R) for each point and places the field in L2(Ω;L2(S2)).
- Coefficient structure: For isotropic Gaussian fields, the nonnegative angular-power-spectrum sequence determines the variances and independence structure of the harmonic coefficients.The coefficients with nonnegative m are independent Gaussian variables, while negative-m coefficients are obtained from them.
- Higher-dimensional spheres: The same spherical framework extends from S2 to the unit sphere Sd−1 embedded in Rd for d≥2.The higher-dimensional formulation uses spherical harmonics indexed by their degree and multiplicity.
3. Decay of the angular power spectrum.
The paper establishes that angular power-spectrum decay is equivalent to covariance-kernel regularity, formalized through weighted Sobolev spaces and spectral sequence spaces. This equivalence extends from integer to noninteger regularity orders using interpolation theory and generalizes from S2 to Sd−1.
- Motivation: The error of a truncated Karhunen–Loève expansion is tied to angular power-spectrum decay, which is characterized by covariance-kernel behavior.The covariance kernel is often prescribed in applications, making this relationship useful for connecting model regularity and truncation accuracy.
- Weighted Sobolev framework: The paper formalizes kernel regularity through weighted Sobolev spaces V η(−1,1) and weighted sequence spaces ℓη.The spaces V η(−1,1) are defined for noninteger η by real interpolation, extending the integer-order construction.
- Interpolation: The equivalence between V η(−1,1) and ℓη for noninteger η follows by interpolation after establishing the result for integer orders.The proof uses the interpolation property of the spaces and corresponding interpolated sequence weights.
- S2 characterization: Theorem 3.1 states that u belongs to V η(−1,1) if and only if its associated weighted spectral sequence satisfies the corresponding summability condition.For integer n, this yields the equivalence between weighted 2-summability of angular power-spectrum coefficients and weighted square integrability of the nth weak derivative.
4. Sample H¨older continuity and differentiability.
This section connects angular-power-spectrum summability to covariance-kernel regularity and derives sample Hölder continuity and differentiability for isotropic Gaussian fields on spheres. The results use moment bounds and a sphere-specific Kolmogorov–Chentsov argument to obtain continuous modifications with explicit regularity exponents.
- Spectral criteria and covariance regularity: Angular-power-spectrum decay controls covariance-kernel Hölder behavior, linking spectral summability to mean-square regularity.The kernel estimate is obtained from weighted spectral criteria and underpins the sample-regularity results.
- From increments to sample regularity: Moment bounds for field increments in geodesic distance provide the input for a sphere-specific Kolmogorov–Chentsov theorem.The argument expresses increment moments through the covariance kernel and then applies the theorem on spherical charts.
- Hölder continuity: A summability exponent β yields a continuous modification that is Hölder continuous with every exponent γ < β/2.This is the principal sample Hölder-continuity conclusion on S2.
- Differentiability: For β > 0, the field has a C^γ-valued modification for every γ < β/2, with k-times continuous derivatives where k = ⌈β/2⌉−1.The kth derivatives are Hölder continuous with exponent γ − k.
- Higher-dimensional spheres: The same regularity characterization extends to isotropic Gaussian fields on S^(d−1) for d ≥ 2.The higher-dimensional result gives the same differentiability index and Hölder exponent relation.
- Higher-dimensional spheres: The higher-dimensional Hölder result improves an earlier theorem by removing its logarithmic factor from the angular-power-spectrum summability assumption.The stated improvement concerns the assumption required for Hölder continuity.
5. Approximation of isotropic Gaussian random fields.
The paper approximates isotropic Gaussian random fields by truncating their spherical-harmonic Karhunen–Loève expansions. Under algebraic angular-power-spectrum decay, it establishes convergence in mean-square, finite-p moment, and almost-sure senses, with rates depending on decay and, on higher-dimensional spheres, dimension.
- Construction: The approximation truncates the spherical-harmonic Karhunen–Loève series of isotropic Gaussian random fields.The truncated fields are introduced to enable implementation and convergence analysis.
- Convergence on S2: For angular power spectra decaying with order α > 2, truncations converge in L2(Ω;L2(S2)) with a bounded truncation error.The bound applies once κ ≥ ℓ0 under Aℓ ≤ C · ℓ^-α.
- Numerical validation: Numerical simulations for α = 3 and α = 5 match the theoretical convergence rates for mean-square and single-sample errors.The experiments use 1000 Monte Carlo samples for the mean-square approximation error and compare against a κ = 27 reference field.
- Convergence on S2: For any finite p ≥ 1, the same algebraic decay yields convergence in Lp(Ω;L2(S2)) with a truncation-error bound.The associated constant depends on p, C, and α.
- Convergence on S2: Almost-sure convergence holds with asymptotic truncation rates for every β < (α − 2)/2.The result uses the p-moment convergence together with Chebyshev’s inequality and the Borel–Cantelli lemma.
- Higher-dimensional spheres: On spheres S^(d−1), convergence in finite-p and almost-sure senses persists, with rates depending on sphere dimension d − 1.The almost-sure rate is stated for every β < (α + 1 − d)/2.
6. Lognormal isotropic Gaussian random fields.
The section shows that exponentiating an isotropic Gaussian random field preserves its sample regularity, including Hölder continuity and differentiability, under corresponding angular-spectrum decay conditions.
- Lognormal random fields are obtained pointwise as exp(T(x)) and model dust, feldspar, and ice-crystal structures.
- The lognormal field exp(T) has the same sample Hölder continuity properties as the underlying Gaussian field T.
- If A_ℓ≤Cℓ^-α with α>2, exp(T) has a modification Hölder continuous with every exponent γ<(α−2)/2, with γ≤1.
- The corresponding differentiability follows from compactness of the sphere, smoothness of the exponential, and the chain rule.
- If the spectrum satisfies the weighted summability assumption with β>0, exp(T) has a C^γ-valued modification for every γ<β/2.
7. Stochastic partial differential equations on the sphere.
The section formulates the stochastic heat equation on the sphere with isotropic Q-Wiener noise, solves it spectrally, and analyzes truncation errors in terms of angular-spectrum decay.
- Isotropic Q-Wiener processes on S2 are characterized spectrally by spherical harmonics, whose corresponding eigenvalues are determined by the angular power spectrum.
- The stochastic heat equation is solved by expanding in spherical harmonics, reducing the problem to stochastic ordinary differential equations for each mode.
- Spectral approximations truncate the solution expansion at κ and evaluate finitely many modes on discrete time and space grids.
- For A_ℓ≤C·ℓ^-α with α>0, the mean-square truncation error is bounded uniformly in time and independently of time discretization.
- Numerical experiments for α=1,3,5 match the theoretical convergence rates, including both Monte Carlo mean-square errors and single-path errors.
APPENDIX: INTERPOLATION SPACES
The appendix develops interpolation-space tools and establishes equivalence between interpolation norms and the spaces used for spherical spectral regularity analysis.
- The appendix introduces fractional-order spaces through the real method of interpolation between integer-order spaces.
- For integer m between k and n, the interpolation space Bm,(k,n) is equivalent to the corresponding space V^m.
- The proof reduces norm equivalence to weighted coefficient summability for Fourier–Legendre expansions.
- The argument uses decomposition into lower- and higher-regularity components, optimization over coefficients, and integral estimates.