🔢 mathematics

Unifying spherical convolutions through lifting

This paper presents a unified framework for spherical convolutions by lifting spin fields to the rotation group $\text{SO(3)}$ , thereby reconciling existing definitions, establishing theoretical limits on equivariance, and providing efficient harmonic-space algorithms implemented in the open-source s2conv code.

Original authors: J. D. McEwen

Published 2026-10-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: J. D. McEwen

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

Imagine looking at the world not as a flat map, but as a perfect sphere. This is how scientists view the Earth, the sky, and the vast universe beyond. On this sphere, data is everywhere: the temperature of the air, the direction of the wind, the polarization of light from the early universe, and the subtle stretching of distant galaxies caused by invisible gravity. For decades, researchers have struggled to analyze this data because the sphere behaves differently than a flat sheet of paper. When you move a measurement from one point to another on a sphere, the local frame of reference—the "up" and "down" or "north" and "east" at that specific spot—rotates. This rotation changes the nature of the data itself. A simple temperature reading stays the same, but a wind vector or a polarization angle acquires a new phase, a mathematical twist, simply because the coordinate system has turned. This makes standard tools for analyzing data, like convolution, which is essentially a way of smoothing or filtering information, extremely difficult to apply correctly without introducing errors that depend on where you place your map's poles.

A new study by Jason D. McEwen offers a unified solution to this long-standing problem. The researcher developed a single, consistent method to perform these convolutions on a sphere by temporarily lifting the data off the surface and onto a higher-dimensional space known as the rotation group. Think of this rotation group as a complete catalog of every possible way to turn a sphere in three-dimensional space. By moving the data into this catalog, the complex problem of rotating frames on the sphere becomes a straightforward problem of shifting and mixing data within the catalog. Once the mathematical operation is performed there, the data is projected back down to the sphere. This approach allows the researcher to handle any type of data, from simple scalar values like temperature to complex spin fields like wind direction or cosmic polarization, with perfect mathematical precision. The method ensures that the result of the calculation does not depend on how the sphere was oriented when the calculation began, a property known as equivariance, which is crucial for scientific accuracy.

The paper reveals that many different methods previously proposed by scientists to handle these spherical calculations are actually just special cases of this single, unified framework. Some methods work well for simple, round filters but fail for directional ones; others handle direction but produce results on the rotation group rather than the sphere. McEwen's work shows that all these variations are connected. By carefully choosing which parts of the rotation group to use and how to project the data back, one can recover every known rotation-equivariant convolution. However, the study also draws a firm line in the sand: it proves that it is impossible to simultaneously have a directional kernel (one that can detect features in specific directions), an output that stays on the sphere, and perfect rotation equivariance for a fixed input and output spin. If you want to keep the output on the sphere and maintain perfect symmetry, you must give up the ability to use arbitrary directional kernels, or you must accept that the output will be a collection of different spin components rather than a single field.

To demonstrate the power of this new framework, the researcher applied it to real-world data from both the Earth and the cosmos. For the Earth, the method was used to analyze the topography, calculating the slope of mountains and the orientation of ridges without the distortions that usually occur near the poles. It was also used to study the Earth's gravity and magnetic fields, allowing for precise calculations of how these fields change with depth and how they relate to one another. In the realm of cosmology, the technique was applied to the polarization of the cosmic microwave background, the afterglow of the Big Bang. This allowed for a clean separation of the data into its fundamental components, known as E and B modes, which carry different physical information about the early universe. The study confirmed that these operations can be performed exactly and efficiently, with errors so small they are limited only by the precision of the computer itself.

The work also clarifies the deep geometric nature of these spin fields. The researcher showed that these fields, which acquire a phase when the local frame rotates, are not just functions on a sphere but are sections of specific geometric structures called line bundles. This insight links the physics of these fields directly to the mathematical theory of bundles, providing a rigorous foundation for why the rotation phase behaves the way it does. The phase is not an arbitrary addition but a necessary consequence of the geometry of the sphere and the way the local frame twists as one moves across it. By expressing this phase through the geometry of the bundle, the study provides a clear, unified language for understanding these complex fields.

All of these operators have been implemented in a new, open-source software package called s2conv. This tool allows scientists to perform these exact calculations on band-limited data, meaning data that has a finite resolution, without approximation errors. The software has been tested extensively, confirming that the mathematical reductions work as predicted and that the operations remain perfectly equivariant to machine precision. The computational cost of these operations scales efficiently, making them practical for large-scale data analysis. The study concludes that this unified description not only resolves the inconsistencies of previous methods but also opens the door to new applications, including the construction of artificial intelligence architectures that can learn from spherical data while respecting its inherent geometry. By lifting the data to the rotation group and bringing it back down with the correct spin, the researcher has provided a robust, error-free way to analyze the curved world we live in and the universe we inhabit.

Technical Summary: Unifying Spherical Convolutions through Lifting

Problem Statement
Data on the sphere ( $S^2$ ) is ubiquitous in fields ranging from cosmology (cosmic microwave background, weak lensing) to geophysics (wind, gravity) and computer vision. A fundamental operation for analyzing such data is convolution. However, the literature contains a fragmented landscape of definitions for spherical convolution. These definitions differ significantly in their assumptions: some restrict kernels to be axisymmetric (invariant under azimuthal rotation), while others allow directional kernels; some produce outputs on the sphere, while others produce outputs on the rotation group $SO(3)$ ; and some accommodate spin fields (tangent vectors, polarization), while others are limited to scalars. Furthermore, existing methods vary in computational cost and, crucially, in whether they guarantee full rotation equivariance (commuting with all $SO(3)$ rotations). There is no unified framework that connects these disparate constructions or clarifies the fundamental trade-offs between directional kernels, spherical output, and equivariance.

Methodology
The author develops a unified description of spherical convolution by "lifting" fields from the sphere to the rotation group $SO(3)$ . The methodology relies on the geometric identification of the sphere as the homogeneous space $S^2 \cong SO(3)/SO(2)$ , where $SO(2)$ is the stabilizer of the North pole.

Lifting and Projection: A spin- $s$ field on the sphere is identified as a section of a line bundle associated with the quotient $SO(3)/SO(2)$ . Mathematically, these are encoded as Mackey functions on $SO(3)$ satisfying a specific transformation law under the stabilizer. The author defines a lift operator $\Lambda_s$ that maps a spin- $s$ field on $S^2$ to a function on $SO(3)$ occupying a single column ( $n = -s$ ) in Wigner space. The projection $\Lambda^\dagger_s$ is the adjoint, mapping functions on $SO(3)$ back to spin- $s$ fields on $S^2$ by integrating over the fiber.
Generalised Lifted Spherical Convolution: The core construction involves three steps:
- Lift a spin- $s$ input field to $SO(3)$ .
- Perform a standard group convolution on $SO(3)$ with an arbitrary kernel $\Psi$ .
- Project the result back to the sphere, potentially to a different spin $s'$ .
Harmonic Analysis: By decomposing the lifted field and the kernel into Wigner expansions, the author derives the harmonic form of this operation. They show that the convolution acts in spin spherical harmonic space as a simple multiplication by a single conjugated Wigner coefficient of the kernel, $\Psi^\ell_{-s', -s}$ , up to a sign.
Classification of Equivariant Maps: Using representation theory (Frobenius reciprocity and Schur's lemma), the author classifies all bounded, rotation-equivariant linear maps between spin fields. They prove that any such map must be a multiplier in harmonic space.

Key Contributions

Unification: The paper demonstrates that every previously defined rotation-equivariant spherical convolution (axisymmetric, directional, group, spin-weighted, and various physical operators like spin raising/lowering) is a special case of the generalised lifted convolution. Differences in existing literature arise solely from constraints on the kernel's Wigner support, the choice of input/output spins, and the inclusion or omission of the lift/projection steps.
Fundamental Limitation: A key theoretical result is the proof that for maps between fields of fixed input and output spin, it is impossible to simultaneously achieve: (1) directional kernels (retaining orientation information), (2) an output defined strictly on the sphere, and (3) full $SO(3)$ equivariance. To retain directional information equivariantly, the output must be a function on $SO(3)$ (or a tuple of spin fields).
Geometric Insight: The work explicitly links the physics of spin-weighted functions (defined by phase changes under local frame rotation) to the mathematics of vector bundles. It identifies spin fields as sections of line bundles over $S^2$ and expresses the rotation phase as the character of the stabilizer evaluated on the "twist" of the section.
Algorithmic Implementation: The author presents exact and efficient harmonic-space algorithms for computing these convolutions for band-limited fields, leveraging sampling theorems on the sphere and rotation group. They provide the open-source s2conv code (built on s2fft and JAX) which implements all operators, including differential operators and spin-changing convolutions.

Results

Numerical Validation: Numerical experiments confirm that the derived reductions (mapping specific literature definitions to the general framework) are correct to machine precision ( $\sim 10^{-16}$ ).
Equivariance: The operators are verified to be rotation-equivariant to machine precision. In contrast, non-equivariant constructions (such as standard correlation or sifting) exhibit errors of order unity when rotated, confirming the theoretical distinction.
Computational Complexity: The complexity of the operators scales as $O(L^3)$ for sphere-to-sphere convolutions and $O(NL^3)$ for directional convolutions with azimuthal band-limit $N$ , where $L$ is the spherical band-limit. This matches the cost of spherical harmonic transforms.
Physical Applications: The framework is applied to real-world data:
- Geophysics: Computing slopes and ridge orientations from topography; deriving gravity anomalies, deflections of the vertical, and gravity gradients from the geoid; and downward continuation of the geomagnetic field.
- Fluid Dynamics: Computing divergence and vorticity from wind and ocean current data, and performing rotation-equivariant smoothing (demonstrating that component-wise smoothing of vector fields introduces coordinate-dependent errors).
- Cosmology: Decomposing cosmic microwave background polarization into E and B modes and reconstructing the polarization map; relating weak lensing convergence and shear via the Kaiser–Squires relation.

Significance
The paper claims to provide the definitive unified framework for rotation-equivariant spherical convolutions. By establishing that all such operations are instances of a single lifted construction, it clarifies the mathematical relationships between disparate methods in the literature. Crucially, it settles the question of what is possible: it proves that one cannot have directional kernels, spherical output, and full equivariance simultaneously for fixed spins, thereby guiding the design of future algorithms and neural networks. The work bridges the gap between the physics literature (spin-weighted functions) and the bundle theory literature, providing a rigorous geometric foundation for handling spin fields on the sphere. The availability of the s2conv code ensures these theoretical advances are immediately applicable to scientific computing and machine learning tasks involving spherical data.

Technical Summary: Unifying Spherical Convolutions through Lifting

More like this