A Deformation Approach to Axially Symmetric Kernels on the Sphere: Harmonic Coupling and Regularity

1 Introduction

• Axially symmetric kernels model a preferred direction, replacing isotropic diagonal harmonic structure with coefficient blocks; their harmonic description dates to.Related representations and kernel properties are studied in and.Source: §1 · p. 2
• Deforming the sphere and evaluating an isotropic kernel at transformed points does not generally provide explicit spherical-harmonic coefficients.Such coefficients are useful for truncated approximations and field simulations.Source: §1 · p. 2
• The paper’s normalized axial deformation yields an explicit spherical-harmonic series for the deformed kernel and closed-form expressions when available for the original kernel.The construction scales points along a chosen axis in R3 and normalizes them back to S2.Source: §1 · p. 2
• The paper studies positive definiteness, native spaces, harmonic coupling and decay, Sobolev regularity, interpolation bounds, and prediction under correctly specified and misspecified covariance models.The introduction describes comparisons with isotropic kernels in both covariance settings.Source: §1 · p. 2

2.1 Spherical harmonics and Sobolev spaces

• Spherical harmonics form an orthonormal basis of L2(S2) and are eigenfunctions of the Laplace–Beltrami operator; Sobolev spaces are defined spectrally, with negative orders defined by duality.The polar angle is L and longitude is λ.Source: §2.1 · p. 3

2.2 Isotropic kernels and native spaces

• A continuous isotropic positive definite kernel on S2 has a Legendre expansion with nonnegative coefficients, and its spectral coefficients characterize its native space.The cited criterion relates coefficient decay to Sobolev-equivalent native spaces under stated spectral conditions.Source: §2.2 · p. 4

2.3 Axial symmetry and Jones blocks

• Axial symmetry means invariance under rotations fixing e3, so the kernel depends on longitudes through their difference and has a block-structured harmonic representation.The paper calls the coefficient matrix for each order an m-th Jones block.Source: §2.3 · pp. 5–6

2.4 Gegenbauer polynomials and associated Legendre functions

• The preliminaries set up finite expansions connecting Gegenbauer polynomials and associated Legendre functions, including coefficients expressed using the Pochhammer symbol.The finite representation is related to identities cited in [12].Source: §2.4 · p. 6

3.1 Positive definiteness under pullback

• Pulling back a positive definite kernel preserves positive definiteness, and strict positive definiteness is preserved when the map is injective.The section presents these facts as basic properties of kernel pullbacks.Source: §3.1 · p. 7, §3 · p. 7

3.2 Native spaces under pullback

• For a bijective map, composition identifies the original and pullback native spaces isometrically.The proof uses the unique representation under a bijection to reduce the pullback quotient norm.Source: §3.2 · p. 7

3.3 Symmetry under pullback

• A G-invariant kernel remains G-invariant under pullback when the map is G-equivariant.For rotations fixing e3, this invariance is axial symmetry about e3.Source: §3.3 · p. 8

3.4 Sobolev stability

• Composition with a smooth diffeomorphism of S2 preserves Hτ(S2) with equivalent Sobolev norms.The section applies this stability together with the native-space isometry to the pulled-back kernel.Source: §3.4 · pp. 8–9

4.1 Geometry of the deformation

• For a smooth diffeomorphism Tβ, if Kiso is positive definite then Kβ is positive definite and axially symmetric about e3; strict positive definiteness is preserved when Kiso is strictly positive definite.For β = 1, Tβ is the identity. For β > 1, points away from the poles and equator move toward their hemisphere’s pole; for 0 < β < 1, they move toward the equator, while poles and equator remain fixed.Source: §4.1 · pp. 10–11

4.2 Harmonic mode coupling

• The pulled-back spherical harmonic expands in the original basis with no terms whose order m′ differs from m.This order-preserving coupling is expressed through integral connection coefficients.Source: §4.2 · p. 11

4.3 Jones blocks of the deformed kernel

• The deformed kernel has an L2 Jones-block expansion whose coefficients vanish whenever m differs from m′.The Jones coefficients are derived from the isotropic coefficients and the harmonic connection coefficients.Source: §4.3 · p. 12

4.4 Explicit formula for the connection coefficients

• The connection coefficients have an explicit finite-sum formula involving Beta and Gauss hypergeometric functions, and vanish when ℓ+n is odd.The formula applies for β>0.Source: §4.4 · pp. 13–14

4.5 Geometric decay for a fixed input mode

• For each fixed input mode, connection coefficients decay geometrically as output degree increases; the decay bound tends to zero as β approaches 1.At β=1 there is no degree coupling; the fixed-mode estimate alone does not give a uniform bound for an entire Jones block.Source: §4.5 · pp. 14–15

4.6 Native spaces and Sobolev regularity

• Under the stated hypotheses, deformation preserves the native space with an equivalent norm and the same Sobolev order, although covariance structure and harmonic coefficients may change.The section also derives Jones-block estimates with constants independent of m and the harmonic degrees.Source: §4.6 · p. 15

4.7 Extension to an arbitrary axis

• For an arbitrary axis u, rotating the adapted basis preserves the connection and Jones coefficients, and transfers the native-space and Sobolev results with constants independent of u.The resulting kernel is axially symmetric about u.Source: §4.7 · p. 17

4.8 Interpolation and error estimates

• When the isotropic kernel satisfies the stated coefficient-decay hypotheses for τ>1, interpolation with the deformed kernel transfers isotropic error estimates to transformed point sets.Interpolation with the deformed kernel is equivalent to isotropic-kernel interpolation at the transformed points.Source: §4.8 · pp. 17, 19
• For f ∈ Hτ(S2) and 0 ≤ σ ≤ τ, Corollary 4.10 gives a constant independent of u, X, and f.The supplied passage does not include the bound's formula.Source: §4.8 · p. 19
• The transformed estimates give the same convergence order as in the isotropic case, but do not guarantee a smaller interpolation error.The paper turns to a stochastic viewpoint in its numerical experiments to demonstrate when a transformed kernel can outperform the isotropic one.Source: §4.8 · p. 19

5.1 Simulation setup

• The simulations use 100 repetitions per scenario, with 120 of 180 uniformly sampled spherical locations for fitting and 60 for testing.Parameters are estimated by Gaussian maximum likelihood; axial fits estimate deformation strength and an unoriented axis in addition to variance and radial scale.Source: §5.1 · pp. 19–20

5.2 Correlation geometry and radial families

• Larger β narrows equatorial correlation contours north–south and expands high-correlation regions near the pole, changes an isotropic range adjustment cannot reproduce.Figure 1 shows axial Matérn correlation for equatorial and 80°N source points across β values.Source: §5.2 · p. 20, Figure 1 · p. 21

5.3 Effect of deformation strength

• For every tested non-isotropic deformation strength, the axial fit has positive mean gain; gains reach 9.7% at β0 = 2.5 and 14.5% at β0 = 3.5.At β0 = 1, where the generating covariance is isotropic, the mean gain is −1.5%; larger β0 values generally yield larger gains.Source: §5.3 · pp. 21–22, Figure 2 · p. 22

5.4 Prediction at matched mean correlation

• With mean pairwise correlation held fixed, mean gains are 6.2% at β0 = 0.5, 9.3% at β0 = 2.5, and 10.7% at β0 = 3.5.The mean gain is −1.1% at β0 = 1.Source: §5.4 · p. 22, Figure 3 · p. 23

5.5 Cross-family comparison

• Across all nine comparisons under axial truth with β0 = 2.5, mean relative RMSE gains are positive, ranging from 9.3% to 59.1%, with axial win rates from 0.84 to 1.00.The largest relative gains occur for Gaussian-generated data, especially when the fitted family is also Gaussian.Source: §5.5 · p. 22, Figure 4 · p. 23

5.6 Robustness to alternative axial geometries

• The generating deformation family gives the largest mean gain in each tested geometry, but Tβ also improves prediction for Gmid-generated data, with a 3.9% mean gain.For the banded example, Tβ gains 0.6% versus 15.6% for the flexible fit; the flexible family is not best in every case either.Source: §5.6 · p. 24, Figure 5 · p. 24
• In the illustrated Gmid realization, Tβ does not improve everywhere, but its area-weighted grid RMSE is lower overall than the isotropic model.Figure 6 selects, from ten separate realizations, the case whose relative grid-RMSE gain is closest to the average.Source: §5.6 · p. 24, Figure 6 · p. 25

6 Conclusion and outlook

• The deformation preserves the native-space Sobolev order and interpolation convergence order under suitable spectral assumptions, and predicts better on average in several tested axial cases.The authors also report better predictions when mean correlation is fixed or a different radial family is fitted; the banded example does not favor one global deformation for every axial covariance structure.Source: §6 · p. 25
• Future work could study estimating the axis and deformation strength from observations, as well as applications to larger datasets and global data.Source: §6 · p. 25
• Explicit bounds on the L2 error from finite kernel-series expansions could support simulation use of the representation.The authors identify this as a continuation, in the spirit of.Source: §6 · p. 25
• A further open question is how to add parameters to the deformation while retaining an explicit harmonic representation.Source: §6 · p. 25
• The transformed point set can also distribute points more favorably on the sphere as a variant of the fake nodes approach.Source: §6 · p. 25

Funding

• The work was supported by the Deutsche Forschungsgemeinschaft under project 461449252, “Kernel interpolation on Riemannian manifolds.”Source: Funding · p. 26

Competing interests

• The authors declare no competing interests.Source: Competing interests · p. 26

Data and code availability

• The numerical experiments use simulated data, and code, results, and figure-reproduction scripts are available online.Source: Data and code availability · p. 26

A Derivation of the explicit connection-coefficient formula

• The proof derives the coefficient formula through a finite associated-Legendre expansion, a parity rule, and evaluation of scalar integrals by Euler’s representation.Source: A Derivation of the explicit… · p. 28
• Substituting into the finite Gegenbauer expansion and applying the stated notation yields an intermediate expression for the derivation.Source: A Derivation of the explicit… · p. 29
• When ℓ+n is odd, the integrand is odd, so the parity argument eliminates that case from the remaining evaluation.Source: A Derivation of the explicit… · pp. 29–30
• The remaining scalar integral is evaluated using Euler’s representation, and substitution of the resulting expressions completes the theorem’s proof.The stated conditions include b > 0 and c − b = µ + 1 > 0.Source: A Derivation of the explicit… · p. 30

Gegenbauer coefficients

• The proof combines a coefficient estimate in the Gegenbauer convention with an associated-Legendre integral estimate and Bernstein-ellipse analyticity.Source: B Proof of geometric fixed-mode decay · p. 30, Gegenbauer coefficients · p. 31

B Proof of geometric fixed-mode decay

• For fixed input mode and deformation parameter, constants in the decay proof may depend on those choices but not on output degree n.Source: B Proof of geometric fixed-mode decay · p. 30

Wang’s estimate and its application

• Wang’s estimate bounds coefficients for functions holomorphic near a Bernstein ellipse; applied to the fixed transformed profile, it yields a bound independent of output degree apart from geometric decay.Source: Wang’s estimate and its application · pp. 31–32

Analyticity of the transformed profile

• The transformed profile is holomorphic inside ellipses up to the parameter set by zeros of qβ; applying the coefficient bound and letting ρ approach ρβ gives the fixed-mode decay rate.Source: Analyticity of the transformed profile · pp. 33–35

• On compact manifolds, smooth diffeomorphisms preserve Sobolev spaces up to equivalent norms through the local definition and smooth coordinate changes [29, Ch. 4].Source: C Proof of Sobolev stability · p. 35
• On S², the Fourier–Laplace Sobolev norm has an equivalent spectral description through spherical-harmonic coefficients [17, Ch. 1].Source: C Proof of Sobolev stability · p. 35
• The paper writes composition with T as C_T f := f ◦ T.Source: C Proof of Sobolev stability · p. 35

C.1 Integer Sobolev orders

• For every integer k ≥ 0, C_T extends uniquely to a bounded isomorphism on H^k(S²).Source: §C.1 · p. 35
• The proof transports a finite smooth atlas and its subordinate partition of unity through T to compare localized norms.Source: §C.1 · pp. 35–36
• Smooth coordinate changes preserve localized Sobolev norms up to constant factors, and the spherical-harmonic spectral norm on S² is equivalent to the atlas norm [29, Ch. 4; 17, Ch. 1].Source: §C.1 · p. 35
• Density extends the smooth-function estimate to all of H^k(S²), while applying the estimate to T⁻¹ supplies the reverse bound and invertibility [29, Ch. 4].Source: §C.1 · p. 36

C.2 Fractional Sobolev orders and duality

• For every τ > 0, composition with T is a bounded isomorphism on H^τ(S²).Source: §C.2 · p. 36
• Interpolation between integer-order Sobolev spaces establishes boundedness at fractional order, and applying it to T⁻¹ gives the bounded inverse.Source: §C.2 · pp. 36–37
• Remark C.3 introduces a Sobolev-duality consequence for use in Corollary 4.8, including when H^τ(S²) carries an equivalent norm.Source: §C.2 · p. 37