Kernel Spherical Score: Methods & Theory

• Kernel spherical score is a family of kernel-based constructions on spherical domains that computes intrinsic gradients for density estimation and drift-based generation.
• It leverages methods such as kernel-gradient drifting, Laplace–Beltrami spectral analysis, and RKHS embeddings to achieve robust, geometry-aware comparisons.
• The approach underpins applications in probabilistic forecasting, robust population comparison, and spherical image seam alignment.

Kernel spherical score is a family of kernel-based constructions on spherical domains rather than a single universally fixed object. In kernel-gradient drifting, it is the intrinsic gradient of the log kernel-smoothed density on $S^n$ and is the basic building block of the drift field that moves samples along the tangent space in the direction that most increases the kernel-smoothed data likelihood and decreases the kernel-smoothed model likelihood (Esteban-Casadevall et al., 11 May 2026). In robust comparison of spherical populations, a convenient “Kernel Spherical Score” consistent with the framework is the heat-decayed $L^2$ difference in Laplace–Beltrami spectral coefficients (Zhang et al., 2018). In probabilistic forecasting on $\mathbb{S}^d$, the kernel spherical score is the kernel score specialized to the sphere, with expected score differences equal to an RKHS discrepancy (Steinwart et al., 2017). Related computer-vision usage connects the phrase to a kernel-based seam alignment score of continuity across borders of 2D representations of spherical images (Christensen et al., 2024).

1. Common mathematical setting

The shared mathematical setting is the sphere equipped either with its Riemannian structure or with its Laplace–Beltrami spectral decomposition. For manifold smoothing, given a positive, normalized kernel $K(x,y)$ on a Riemannian manifold $M$, the kernel-smoothed density is

$(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$

and the intrinsic kernel-smoothed score is

$s_K^M[p](x):=\nabla_M \log (K*p)(x).$

On the unit sphere $S^n\subset \mathbb{R}^{n+1}$, the Riemannian gradient of a smooth function is obtained by projecting the ambient Euclidean gradient:

$\nabla_S f(x)=(I-xx^T)\nabla f(x).$

This makes the score an intrinsic tangent vector rather than an ambient displacement (Esteban-Casadevall et al., 11 May 2026).

A second common structure is the Laplace–Beltrami eigensystem. On $S^1$, $L^2$0 with eigenfunctions $L^2$1 and eigenvalues $L^2$2. On $L^2$3, $L^2$4 acts on spherical harmonics $L^2$5 with eigenvalues $L^2$6. Any sufficiently smooth density admits a spectral expansion in these eigenfunctions, and heat flow $L^2$7 smooths the density by exponentially damping higher-frequency coefficients (Zhang et al., 2018).

A third common structure is the RKHS mean embedding induced by a positive definite kernel $L^2$8. For a probability measure $L^2$9,

$\mathbb{S}^d$0

and the associated kernel score is expressed through $\mathbb{S}^d$1 and the reproducing kernel $\mathbb{S}^d$2 (Steinwart et al., 2017). This suggests a common theme: kernels are used either to smooth spherical densities, to define strictly proper scoring rules, or to quantify geometry-respecting discrepancies on spherical data.

Setting	Object called kernel spherical score	Representative expression
Kernel-gradient drifting	Intrinsic kernel-smoothed score on $\mathbb{S}^d$3	$\mathbb{S}^d$4
Spherical population comparison	Heat-decayed spectral surrogate	$\mathbb{S}^d$5
Probabilistic forecasting	Kernel score on $\mathbb{S}^d$6	$\mathbb{S}^d$7
Spherical image continuity	Kernel-based seam score	$\mathbb{S}^d$8

2. Intrinsic kernel-smoothed score on the sphere

In the kernel-gradient drifting formulation, the spherical version is explicit. On the unit sphere $\mathbb{S}^d$9, the kernel spherical score is defined by

$K(x,y)$0

with

$K(x,y)$1

The corresponding drift is the score difference

$K(x,y)$2

which lies in $K(x,y)$3 (Esteban-Casadevall et al., 11 May 2026).

Under mild regularity, the attraction and repulsion components satisfy $K(x,y)$4 and $K(x,y)$5, so the drift is their difference. This recovers the Gaussian case as a special instance but holds for general kernels, including spherical ones. The practical interpretation given in the paper is that the field moves samples in the direction that most increases the kernel-smoothed data likelihood and decreases the kernel-smoothed model likelihood (Esteban-Casadevall et al., 11 May 2026).

For the von Mises–Fisher kernel,

$K(x,y)$6

the ambient gradient is $K(x,y)$7 and the intrinsic spherical gradient is

$K(x,y)$8

Consequently,

$K(x,y)$9

where

$M$0

The paper characterizes this as a spherical mean-shift: the score points toward the weighted mean of nearby data, projected to the tangent space (Esteban-Casadevall et al., 11 May 2026).

For spectral kernels such as the heat kernel,

$M$1

the score can be written directly in spherical harmonics. In practice, the series is truncated to a finite bandwidth $M$2 and spherical-harmonic coefficients are precomputed. Updates remain on the sphere through either the exponential map

$M$3

or the first-order retraction

$M$4

Small steps are used to avoid overshooting and to ensure stability (Esteban-Casadevall et al., 11 May 2026).

3. Strict propriety, characteristic kernels, and identifiability

A central issue for any kernel spherical score is whether it determines the target distribution. In the RKHS formulation, the kernel score on a measurable space is

$M$5

On $M$6, the same formula defines the kernel spherical score. Its expected difference induces the maximum mean discrepancy

$M$7

and the kernel score is strictly proper if and only if the kernel is characteristic (Steinwart et al., 2017).

On compact spaces, continuous characteristic kernels admit a spectral characterization through positive Mercer coefficients. For isotropic kernels on spheres,

$M$8

or equivalently $M$9, Schoenberg or Gegenbauer coefficients determine whether the kernel is universal or characteristic. On $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$0, a kernel induced by $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$1 is characteristic iff the corresponding coefficients $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$2 are positive for all $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$3; it is universal iff they are positive for all $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$4 (Steinwart et al., 2017).

The kernel-gradient drifting formulation gives a closely related identifiability statement. If $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$5, then the score determines $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$6 up to a constant, and normalization fixes that constant. If $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$7 is characteristic on the relevant class of distributions, then equality of smoothed densities implies equality of the underlying densities. Formally, if $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$8 is characteristic and $(K*p)(x):=\int_M K(x,y)p(y)\,dvol_M(y),$9 for all $s_K^M[p](x):=\nabla_M \log (K*p)(x).$0, then $s_K^M[p](x):=\nabla_M \log (K*p)(x).$1. On compact manifolds such as $s_K^M[p](x):=\nabla_M \log (K*p)(x).$2, spectral kernels built from Laplace–Beltrami eigenpairs are characteristic when all spectral weights are strictly positive; manifold Matérn kernels and the heat kernel on $s_K^M[p](x):=\nabla_M \log (K*p)(x).$3 fall in this class and hence yield identifiability (Esteban-Casadevall et al., 11 May 2026).

A common misconception is to treat any geodesic radial kernel as automatically suitable for identifiability on curved manifolds. The cited analysis states the opposite for a prominent family: kernels of the form $s_K^M[p](x):=\nabla_M \log (K*p)(x).$4 are not positive-definite on curved manifolds for all $s_K^M[p](x):=\nabla_M \log (K*p)(x).$5 and thus do not guarantee identifiability globally. This motivates spectral or vMF-like kernels on spheres (Esteban-Casadevall et al., 11 May 2026).

4. Spectral smoothness equalization and the heat-decayed score

A distinct but related usage arises in robust comparison of spherical populations. The problem addressed there is that KDEs on spheres are notoriously sensitive to the kernel family, its bandwidth, and sample size. The proposed framework compares densities at matched smoothness, independent of the initial kernel and bandwidth that produced the KDEs and robust to sample size (Zhang et al., 2018).

The main construction uses the heat flow

$s_K^M[p](x):=\nabla_M \log (K*p)(x).$6

together with the smoothness functional

$s_K^M[p](x):=\nabla_M \log (K*p)(x).$7

Equalization is performed by mapping densities to the section

$s_K^M[p](x):=\nabla_M \log (K*p)(x).$8

and then computing the Riemannian geodesic distance $s_K^M[p](x):=\nabla_M \log (K*p)(x).$9 between equalized coefficient vectors on this ellipsoid. That geodesic distance is the paper’s principal score (Zhang et al., 2018).

Within the same framework, a convenient “Kernel Spherical Score” is introduced as the heat-decayed $S^n\subset \mathbb{R}^{n+1}$0 difference

$S^n\subset \mathbb{R}^{n+1}$1

The specializations stated in the paper are

$S^n\subset \mathbb{R}^{n+1}$2

and

$S^n\subset \mathbb{R}^{n+1}$3

Using a common $S^n\subset \mathbb{R}^{n+1}$4 across populations damps high frequencies uniformly and substantially attenuates bandwidth and sample-size effects; when $S^n\subset \mathbb{R}^{n+1}$5 is chosen so that $S^n\subset \mathbb{R}^{n+1}$6, $S^n\subset \mathbb{R}^{n+1}$7 tracks $S^n\subset \mathbb{R}^{n+1}$8 closely (Zhang et al., 2018).

The interpretation is resolution-dependent. Larger $S^n\subset \mathbb{R}^{n+1}$9 or $\nabla_S f(x)=(I-xx^T)\nabla f(x).$0 values indicate greater dissimilarity between the underlying spherical densities after equalizing smoothness, while values near zero indicate similarity at the chosen resolution or smoothness level. This usage is therefore comparative rather than generative: the object is a score or distance between spherical densities, not a tangent-vector field on the sphere (Zhang et al., 2018).

5. Estimation, numerical procedures, and one-step generation

For the intrinsic score on $\nabla_S f(x)=(I-xx^T)\nabla f(x).$1, Monte Carlo estimation is immediate from samples $\nabla_S f(x)=(I-xx^T)\nabla f(x).$2:

$\nabla_S f(x)=(I-xx^T)\nabla f(x).$3

For vMF kernels this becomes

$\nabla_S f(x)=(I-xx^T)\nabla f(x).$4

The model term is estimated analogously from generated samples $\nabla_S f(x)=(I-xx^T)\nabla f(x).$5. The reported computational cost per query $\nabla_S f(x)=(I-xx^T)\nabla f(x).$6 is $\nabla_S f(x)=(I-xx^T)\nabla f(x).$7 for the data term and $\nabla_S f(x)=(I-xx^T)\nabla f(x).$8 for the model term, with minibatches, neighbor truncation, or precomputed spectral features used to reduce cost. Denominator regularization, gradient clipping, and step-size control are recommended for stability (Esteban-Casadevall et al., 11 May 2026).

In one-step spherical generation, a generator $\nabla_S f(x)=(I-xx^T)\nabla f(x).$9 maps latent noise $S^1$0 to $S^1$1, then the sample is drifted to

$S^1$2

Training uses a stop-gradient loss

$S^1$3

In spherical geospatial experiments on $S^1$4, the intrinsic formulation improved performance, with exponential-map updates used to respect spherical geometry (Esteban-Casadevall et al., 11 May 2026).

For the spectral comparison framework, coefficient estimation can be performed directly from samples. On $S^1$5,

$S^1$6

and on $S^1$7,

$S^1$8

After estimating coefficients, one either solves for the equalization times $S^1$9 to reach the section $L^2$00 and computes $L^2$01 by path-straightening, or applies a common $L^2$02 and computes $L^2$03. The reported complexity is $L^2$04 on $L^2$05 with FFT-accelerated methods and $L^2$06 on $L^2$07 with precomputed spherical harmonics or spherical harmonic transforms (Zhang et al., 2018).

Bandwidth and truncation are substantive tuning parameters in both settings. Larger smoothing stabilizes estimates but blurs fine structure; smaller bandwidths increase estimator variance. In the spectral method, one avoids “deblurring” by choosing $L^2$08 so that equalization uses smoothing rather than negative time. In the drift-based method, spectral kernels with positive weights yield identifiability, while intrinsic projection mitigates bias from ambient displacements (Esteban-Casadevall et al., 11 May 2026, Zhang et al., 2018).

6. Related usages, interpretation, and limitations

In probabilistic forecasting, the kernel spherical score is explicitly distinct from the classical “spherical score” for discrete outcomes. The cited paper states that kernel spherical scores are RKHS-based and apply to continuous spherical domains; they generalize the CRPS and, on finite sets, include the Brier score as special cases of kernel scores, but they are distinct from the classical discrete spherical scoring rule (Steinwart et al., 2017).

A separate usage appears in spherical image evaluation. There, the relevant quantity is Discontinuity Score (DS), a kernel-based seam alignment score of continuity across borders of 2D representations of spherical images. Seam-centered windows are converted to grayscale, convolved with a horizontal Scharr kernel, and scored by a ratio of gradient magnitudes across the seam:

$L^2$09

with image-level aggregation over seams. DS is higher-is-worse and is designed to detect seam misalignment that standard FID and OmniFID do not penalize directly (Christensen et al., 2024). This suggests that “kernel spherical score” can also denote a local continuity metric on spherical image representations, rather than a density score or an RKHS scoring rule.

Several limitations recur across formulations. First, characteristicness or strict propriety does not imply control of total variation: if the space of finite signed measures is infinite-dimensional and $L^2$10 is characteristic, then distributions can be far apart in total variation while having arbitrarily small MMD. Spheres with their full measure spaces inherit this caveat, so kernel spherical scores based on RKHS embeddings cannot reliably distinguish every total-variation discrepancy (Steinwart et al., 2017). Second, on spheres and other curved manifolds, geodesic radial kernels are not positive-definite for all bandwidths, and cut loci can complicate their use; spectral kernels are therefore privileged when identifiability is required (Esteban-Casadevall et al., 11 May 2026). Third, all computational variants remain bandwidth- or truncation-sensitive in practice even when they are more robust than naïve KDE comparison: $L^2$11, $L^2$12, harmonic truncation, and neighborhood size determine the scale at which discrepancies are visible (Zhang et al., 2018).

Taken together, these formulations place kernel spherical score at the intersection of Riemannian density estimation, RKHS scoring rules, spectral statistics on spheres, and spherical-geometry-aware evaluation. The term does not designate a single invariant formula across fields; instead, it denotes kernel-based scores or score-like objects that are adapted to spherical structure, whether by intrinsic gradients, Laplace–Beltrami spectral decay, RKHS mean embeddings, or seam-local convolutional measurements (Esteban-Casadevall et al., 11 May 2026, Zhang et al., 2018, Steinwart et al., 2017, Christensen et al., 2024).