Spherical CNNs: Equivariant Learning on Spheres

• Spherical CNNs are neural architectures that generalize convolution to spherical data by exploiting rotational symmetries and spherical harmonics.
• They compute spherical convolutions in the spectral domain using real spherical harmonics, ensuring mathematical rigor and efficient parameter sharing.
• Their exact rotation equivariance improves data efficiency and robustness, enabling superior performance in applications such as diffusion MRI and 3D shape analysis.

Spherical convolutional neural networks (Spherical CNNs, or S-CNNs) are a class of neural architectures designed to perform convolution-like operations on data defined on the surface of the 2-sphere S² (or, equivalently, equivariant to the rotation group SO(3)). Unlike traditional planar CNNs, which are tailored for Euclidean domains and exploit translation symmetry, Spherical CNNs generalize these notions to accommodate the intrinsic symmetries and geometry of spherical data. This exact rotational equivariance is essential for scientific and engineering domains where the underlying data is naturally defined on spheres, such as in diffusion MRI, 3D shape analysis, cosmology, atmospheric science, and omnidirectional imaging.

1. Mathematical Foundations of Spherical Convolution

The fundamental operation in Spherical CNNs is the spherical convolution, which generalizes planar convolution to the sphere using group theory. Let $f: S^2 \rightarrow \mathbb{R}$ be a spherical signal (e.g., a scalar field such as the apparent diffusion coefficient in diffusion MRI), and $h: S^2 \rightarrow \mathbb{R}$ a trainable spherical filter. The spherical convolution is defined by:

$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$

where $R \in SO(3)$ and $du$ denotes the uniform surface measure on $S^2$. This operation produces a function on $SO(3)$, representing the correlation of $f$ with all rotated versions of $h$.

Filters are commonly parameterized in the real spherical harmonics basis: $$h(u) = \sum_{\ell=0}^L \sum_{m=-\ell}^\ell h_{\ell m}\; Y_{\ell m}(u)$$ allowing efficient computation in the spectral domain. For practical implementation, input and output signals are discretized on suitable spherical grids (such as equiangular, HEALPix, or icosahedral tessellations).

A crucial property is rotation equivariance: For any $h: S^2 \rightarrow \mathbb{R}$0,

$h: S^2 \rightarrow \mathbb{R}$1

with $h: S^2 \rightarrow \mathbb{R}$2. Thus, rotating the input and then applying the convolution yields the same result as convolving and then rotating the output, guaranteeing consistent representation across arbitrary orientations (Cohen et al., 2018, Goodwin-Allcock et al., 2022, Snoussi et al., 2 Apr 2025).

2. Network Architectures and Layer Construction

Spherical CNNs replace or generalize standard CNN layers to operate on spherical data while preserving SO(3) symmetry:

• Input Representation: The initial sphere signal $h: S^2 \rightarrow \mathbb{R}$3 may derive from raw measurements (e.g., ADC profiles in diffusion MRI via $h: S^2 \rightarrow \mathbb{R}$4 (Goodwin-Allcock et al., 2022)) or be synthesized from point clouds, images, or other modalities using appropriate preprocessing.
• Spherical Convolutional Layers: Each layer applies a channel-mixing spherical convolution ($h: S^2 \rightarrow \mathbb{R}$5), typically realized in the spectral domain using spherical harmonics. The SH basis allows band-limiting, parameter sharing across degrees, and efficient computation.
• Nonlinearity: Nonlinearities (e.g., ReLU) are often applied either pointwise in the spatial domain or spectrally in the harmonic domain.
• Pooling and Invariance: Rotation-invariant descriptors are constructed using global pooling over SO(3) (average or max). For regression or classification tasks, this pooled representation feeds into fully connected heads.
• Specialized Architectures: Some networks may include hybrid block structures that interleave spatial-spectral couplings, filter pooling (Goodwin-Allcock et al., 2022), or aggregate radial features (Spezialetti et al., 2020).

S-CNNs differ from planar or standard 3D CNNs in that they explicitly encode the topology and symmetry of the sphere, ensuring architectural robustness to coordinate representation and choice of orientation.

3. Theoretical Properties: Equivariance and Robustness

Spherical CNNs are characterized by their exact rotation equivariance to SO(3) at every layer. This property yields several significant advantages:

• Data Efficiency: The capacity to generalize across orientations reduces the need for training examples spanning all possible rotations. It suffices to train within a single or limited set of reference frames, as the network naturally extrapolates to all orientations (Goodwin-Allcock et al., 2022).
• No Orientation Augmentation Required: Unlike traditional CNNs, Spherical CNNs do not require manual or synthetic data augmentation for rotational coverage (Goodwin-Allcock et al., 2022, Spezialetti et al., 2020).
• Sampling-Scheme Agnosticism: For cases like diffusion MRI, S-CNNs are robust to changes in the underlying measurement/sampling scheme, since continuous signals (e.g., ADC profiles) can be resampled or interpolated to any desired set of spherical directions without necessitating retraining or parameter adjustment (Goodwin-Allcock et al., 2022).
• Stability to Diffeomorphisms: Spherical CNNs exhibit Lipschitz stability to geometric perturbations close to but not exactly rotations (i.e., diffeomorphisms on the sphere), with provable output bounds proportional to the perturbation size. This underpins reliable performance in the presence of mild geometric distortions or non-ideal sampling (Gao et al., 2020).
• Universal Approximation: By parameterizing filters in a band-limited basis and leveraging SO(3) group structure, S-CNNs can approximate any equivariant filter on the sphere to arbitrary accuracy consistent with the symmetry constraints (Goodwin-Allcock et al., 2022).

4. Empirical Performance and Benchmarks

Quantitative evaluations consistently demonstrate the superiority of Spherical CNNs over non-equivariant baselines for tasks with intrinsic spherical or rotational structure:

• Diffusion MRI Parameter Estimation: In scenarios with clinical-level (six-direction) dMRI data, S-CNNs halve the RMSE in high-anisotropy voxels relative to FCNs, maintain accuracy under new gradient schemes (RMSE difference $h: S^2 \rightarrow \mathbb{R}$6 vs. $h: S^2 \rightarrow \mathbb{R}$7 for FCN, $h: S^2 \rightarrow \mathbb{R}$8), and exhibit uniform errors across test orientations. Training on only $h: S^2 \rightarrow \mathbb{R}$9 of the data maintains nearly identical RMSE to full-dataset training due to equivariance (Goodwin-Allcock et al., 2022).
• Canonical Orientation Learning: Compass, leveraging Spherical CNNs, achieves state-of-the-art local reference-frame repeatability and robust shape orientation on point clouds, with orientation accuracy improvements over SHOT, FLARE, TOLDI, and 3DSN (Spezialetti et al., 2020).
• Cross-modal Pose Estimation: Cross-domain S-CNN embeddings enable pose estimation between images and 3D models via equivariant correlation, achieving median errors of $(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$0–$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$1 on ShapeNet (Esteves et al., 2018).
• Generalization Across Sampling Schemes: S-CNNs are agnostic to specific spherical sampling protocols, delivering invariant performance on both standard and novel acquisition schemes (Goodwin-Allcock et al., 2022).

Selected Quantitative Results

Task	S-CNN Metric (RMSE/Acc)	Baseline (FCN/MLP)	Data/Notes
dMRI FA estimation (scheme mismatch)	$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$2	$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$3	$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$4 in RMSE(Goodwin-Allcock et al., 2022)
dMRI orientation generalization (RMSE)	$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$5 (uniform)	up to $(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$6	High-FA voxels, test orientation(Goodwin-Allcock et al., 2022)
Training coverage needed	$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$7 data $(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$8 same RMSE	$(f * h)(R) = \int_{S^2} f(u)\; h(R^{-1}u)\; du$9 needed	Equivariance reduces sample complexity(Goodwin-Allcock et al., 2022)
Surface orientation repeatability (3DMatch)	$R \in SO(3)$0 (Compass)	$R \in SO(3)$1 (SHOT)	Fraction keypoints $R \in SO(3)$2(Spezialetti et al., 2020)
Shape classification, rotated (PointNet+Compass)	$R \in SO(3)$3 (AR)	$R \in SO(3)$4 (PointNet)	ModelNet40, no rot aug(Spezialetti et al., 2020)
Image-3D pose corr. (median error)	$R \in SO(3)$5–$R \in SO(3)$6	—	Cross-domain S-CNN(Esteves et al., 2018)

5. Applications Across Scientific and Engineering Domains

Spherical CNNs demonstrate broad applicability in contexts where data is naturally defined on the sphere or must be analyzed independently of orientation:

• Neuroimaging: Quantitative tissue parameter estimation, disease classification (e.g., Alzheimer's disease) from cortical morphometric measures mapped to the sphere (Feng et al., 2018, Goodwin-Allcock et al., 2022).
• Medical and Scientific Imaging: Rotation-robust analysis for dMRI, fiber orientation distribution estimation, and more general spherical tomographic methods (Snoussi et al., 2 Apr 2025).
• 3D Shape Analysis and Retrieval: Shape classification, orientation learning, and pose-invariant embeddings for objects represented as meshes, point clouds, or rendered views (Spezialetti et al., 2020, Esteves et al., 2018).
• Astrophysics and Cosmology: Detection of non-Gaussianity in cosmic microwave background (CMB) maps via spherical CNN-based regression and classification (Melsen et al., 2024).
• Omnidirectional Vision: Classification and semantic segmentation for 360° images using spherical polyhedron tessellations, mesh-based convolutions, or graph methods (Su et al., 2017, Lee et al., 2018, Jiang et al., 2019).
• Climate and Geoscience: Spherical modeling for weather forecasting, planetary data, and global climate patterns exploiting full-sphere coverage and equivariant architectures (Esteves et al., 2023, Jiang et al., 2019).

6. Extensions, Limitations, and Future Directions

Spherical CNNs have spurred innovations in both theory and practice:

• Filter Parameterization: More expressive architectures support anisotropic (non-zonal) filters using spin-weighted harmonics, partial differential operator (PDO) kernels, or tight framelets for enhanced spatial localization (Esteves et al., 2020, Shen et al., 2021, Li et al., 2022).
• Scalability and Efficiency: Recent work targets scaling S-CNNs to high resolutions and multiple tasks using spectral pooling, hardware-efficient implementations, and hybrid architectures that combine different feature domains (Esteves et al., 2023, Cobb et al., 2020, McEwen et al., 2021).
• Generalization to Other Groups: There is ongoing exploration of convolutional frameworks equivariant to conformal transformations (Möbius group) for broader geometric invariance, as well as methods for data defined on non-spherical manifolds, graphs, and heterogeneous triangulations (Mitchel et al., 2022, Lee et al., 2018, Jiang et al., 2019).
• Limitations: Despite exact theoretical equivariance, practical implementations may face discretization artifacts, computational cost for high bandwidths, and complexity in non-uniform or incomplete sampling schemes. Careful grid and harmonic basis selection is required to balance accuracy and efficiency (Gao et al., 2020, Cobb et al., 2020, Shen et al., 2021).
• Open Problems: Further work is needed on parameter-efficient equivariant nonlinearities, learning on general manifolds, robust representations under non-rigid deformations (diffeomorphisms), and realtime deployment in large-scale or resource-constrained environments (Spezialetti et al., 2020, Gao et al., 2020, Esteves et al., 2023).

Spherical CNNs provide a rigorous, efficient, and geometrically faithful approach for learning on spherical data. By incorporating exact SO(3) symmetry, these models achieve dramatic gains in robustness, generalization, and interpretability for a wide range of modern scientific and engineering applications (Goodwin-Allcock et al., 2022, Snoussi et al., 2 Apr 2025, Spezialetti et al., 2020, Cohen et al., 2018, Gao et al., 2020, Esteves et al., 2023).