🔬 physics

Commutative Coarse-graining on the Sphere

This paper identifies that standard spatial filtering on the sphere fails to commute with differential operators due to intrinsic curvature, and proposes a unique generalized spherical convolution that ensures consistency with physics by commuting with radial projection, divergence, and curl, thereby eliminating spurious errors in coarse-grained dynamics.

Original authors: Chayut Teeraratkul, Hussein Aluie

Published 2026-10-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chayut Teeraratkul, Hussein Aluie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Earth is a sphere, yet for centuries, the mathematics used to describe the movement of air and water across its surface has relied on a simplifying assumption: that the ground beneath us is flat. This approximation works well for small regions, like a city or a local weather system, where the curve of the planet is barely noticeable. However, when scientists attempt to model the entire globe or simulate the complex, swirling interactions of the atmosphere and oceans, this flat-space intuition breaks down. The core challenge lies in how we simplify these massive, continuous systems into manageable pieces. To study large-scale weather patterns, researchers use a technique called coarse-graining, which is essentially a way of smoothing out tiny, chaotic details to reveal the bigger picture. In a flat world, this smoothing process works hand-in-hand with the laws of physics that govern how things move and change. But on a sphere, the traditional methods of smoothing have a hidden flaw: they distort the very physical laws they are meant to preserve, creating artificial errors that can grow to dominate the simulation.

A team of researchers at the University of Rochester and the University of California, Los Angeles, has identified and solved this long-standing geometric problem. They demonstrated that the standard way of averaging data on a globe—treating the east-west and north-south directions as if they were fixed, straight lines like on a piece of paper—fails to respect the curvature of the Earth. When scientists apply this traditional method to vector fields, such as wind velocity, the smoothing process inadvertently mixes the horizontal flow with the vertical direction. Imagine trying to average the direction of arrows drawn on a basketball; if you simply average the arrow components as if the ball were a flat sheet, the resulting average points slightly inward or outward, losing its original tangential nature. This "leakage" creates a false sense of divergence, suggesting that air is appearing or disappearing out of nowhere, which violates the fundamental physics of fluid flow. The researchers found that these errors are not just minor glitches; in some cases, the artificial errors generated by the old method were eighteen orders of magnitude larger than the actual physical signals they were trying to measure.

To fix this, the team developed a new mathematical framework for filtering data on a sphere that respects its true shape. Instead of sliding a filter across a flat grid, their method rotates the filter around the sphere, ensuring that the averaging process moves with the curvature of the surface. They proved that this new approach, which they call a generalized spherical convolution, is the only linear method that keeps the physics consistent. It ensures that if you smooth the data first and then calculate how much it is spreading out or spinning, you get the exact same result as if you calculated the spreading or spinning first and then smoothed the answer. This property, known as commutation, is essential for deriving accurate equations that describe how large-scale weather systems evolve over time. The researchers tested their method using a computer simulation of an unstable jet stream in the atmosphere. When they applied their new filter, the simulation remained perfectly consistent with the laws of physics, with errors so small they were indistinguishable from the limits of computer precision. In contrast, the traditional method produced massive, spurious errors that would have completely ruined the accuracy of any global climate or weather model.

The implications of this work extend far beyond a single simulation. The researchers showed that many popular techniques used in modern science, including certain types of machine learning algorithms that analyze global data, have been inadvertently using the flawed, component-by-component averaging method. By adopting their new framework, scientists can now process data from the atmosphere, the oceans, and even other planets without introducing artificial distortions. This allows for a more honest representation of how energy and motion transfer across different scales on a curved surface. The study confirms that the difficulty was never a matter of poor computing power or a specific choice of coordinates, but rather a fundamental geometric oversight. By correcting how we average data on a sphere, the researchers have provided a solid foundation for the next generation of global simulations, ensuring that the models we use to understand our planet's climate are built on a mathematically sound and physically consistent basis.

Technical Summary: Commutative Coarse-graining on the Sphere

Problem Statement
The paper addresses a fundamental inconsistency in the coarse-graining and filtering of vector fields on the spherical surface ( $S^2$ ), a critical operation for Large Eddy Simulations (LES), multiscale dynamics analysis, and data processing in geophysics and astrophysics. Standard filtering practices on the sphere typically rely on spatial convolution. However, a basic consistency requirement for deriving coarse-grained equations is that the filtering operator must commute with spatial differential operators (gradient, divergence, and curl).

The authors demonstrate that common filtering methods on the sphere violate this commutation property. This failure is not an artifact of numerical grid discretization, coordinate system choice, or metric terms, but is intrinsic to the curved geometry of the sphere. Specifically:

Componentwise Convolution: Treating spherical vector components (radial, zonal, meridional) as scalars and filtering them independently fails to respect the geometry. It incorrectly implies that a purely zonal flow with constant speed is scale-invariant, ignoring the intrinsic gradients and vorticity generated by the sphere's curvature.
Ambient Convolution: Performing convolution in the embedding 3D Euclidean space ( $\mathbb{R}^3$ ) by averaging Cartesian components accounts for the variation of basis vectors but fails to preserve the field's tangential or radial character for general flows. This leads to "radial-lateral mixing," where a tangent vector field acquires a spurious radial component after filtering, causing the filtered field to lose its physical properties (e.g., becoming non-divergent-free when it should be).

Methodology and Theoretical Framework
The authors build upon a generalized spherical convolution framework proposed in a prior work [1]. The core of their methodology involves:

Symmetry Analysis: Recognizing that while Euclidean convolutions commute with derivatives due to translation equivariance, spherical convolutions must rely on rotation equivariance ( $SO(3)$ ).
Spectral Decomposition: Utilizing the Edmonds basis for vector spherical harmonics. The authors show that the standard "ambient" convolution acts diagonally on this basis but with degree-shifted eigenvalues for certain components. This shift causes the mixing of radial, potential, and solenoidal modes, breaking commutation with differential operators for general flows.
Generalized Convolution Construction: To restore commutation, the authors define a unique linear operator, the generalized convolution ( $G \overrightarrow{\otimes}$ $G \otimes$ ), which applies spectrally shifted kernels to specific components of the vector field in the Edmonds basis.
- The operator is defined such that it preserves the eigenvalue $\dot{b}_G(n)$ for all vector spherical harmonic modes ( $Y_{n,j}, \Phi_{n,j}, \Psi_{n,j}$ ).
- This is achieved by defining shifted kernels $G^{(1)}$ and $G^{(2)}$ with Legendre coefficients $\dot{b}_G(n-1)$ and $\dot{b}_G(n+1)$ respectively, effectively "undoing" the degree shift introduced by the geometry.
Practical Implementation: The paper demonstrates that this generalized convolution can be computed efficiently without full spherical harmonic transforms by decomposing the vector field into Helmholtz scalars (radial, potential, and solenoidal) and applying the scalar convolution to each, followed by reconstruction.

Key Contributions

Identification of the Root Cause: The paper rigorously establishes that the failure of commutation is a geometric property of the manifold $S^2$ , distinct from issues related to coordinate singularities or grid resolution.
Uniqueness of the Operator: The authors assert that for a given scalar kernel, the generalized convolution is the unique linear operator that commutes with radial projection, divergence, and curl on the sphere. This implies that any consistent linear coarse-graining method (including spectral truncation, wavelet transforms, diffusion, or machine-learned convolutions) must take this form.
Quantification of Errors: The study quantifies the magnitude of errors introduced by non-commutative filters, showing they can exceed the physical signals themselves.

Numerical Results
The authors validate their theory using a simulation of an unstable zonal jet in the single-layer shallow water equations (SWE) on a sphere.

Divergence Test: For a flow that is strictly 3D divergence-free ( $\nabla_3 \cdot \mathbf{v} = 0$ ), the generalized convolution preserves this property to within numerical round-off ( $O(10^{-22} - 10^{-20})$ s $^{-1}$ ). In contrast, the ambient convolution produces a spurious divergence of $O(10^{-2})$ s $^{-1}$ , an error eighteen orders of magnitude larger and comparable to the flow's actual lateral divergence.
Gradient Test: When filtering the height gradient ( $\nabla h$ ), the generalized convolution maintains tangency to the sphere with a relative error of $O(10^{-12})$ . The ambient convolution introduces a radial component reaching $\sim 20\%$ of the gradient magnitude. Componentwise convolution yields errors comparable to the signal itself ( $O(1)$ ).
Vorticity Test: For solenoidal (divergence-free) flows, the ambient convolution does commute with the curl (vorticity). This is because the solenoidal basis functions ( $\Psi_{n,j}$ ) are eigenfunctions of the ambient convolution operator. Consequently, the vorticity of the filtered velocity computed via the ambient method agrees with the filtered scalar vorticity to round-off. However, this commutation is a special case limited to solenoidal fields and does not extend to general flows with divergent components, where the ambient convolution fails to preserve tangency and commutation with the gradient.

Significance and Claims
The paper claims that the generalized convolution provides the necessary mathematical framework for deriving consistent coarse-grained equations on the sphere. Without this framework, standard filtering introduces spurious sources and sinks (divergence errors) and gradient errors that contaminate the physics of the coarse-grained dynamics.

The authors emphasize that this construction is essential for:

Deriving LES equations in physical space for global flows.
Ensuring that data assimilation and inverse problems on the sphere preserve the underlying physical symmetries.
Correctly extending machine learning models (e.g., rotation-equivariant neural networks) to vector and tensor fields on the sphere, ensuring they do not fail outside their training intervals due to geometric inconsistencies.

The work serves as a correction to decades of practice in geophysical modeling and a guide for future applications in planetary physics, cosmology, and inertial confinement fusion where spherical geometry is intrinsic.

More like this