Geometric deep learning promises a beautiful mathematical guarantee: if you transform an input (e.g., rotate an image, scale a voxel grid, or spin a sphere), the internal features of a group-equivariant neural network will transform accordingly. In theory, this eliminates the need for brute-force data augmentation and baked-in sample inefficiency.
In practice, however, computers do not operate on continuous manifolds. They operate on discrete grids—pixels, voxels, meshes, and point clouds. This transition from continuous group theory to discrete digital computation introduces a fundamental flaw known as discretization-induced equivariance error. When continuous transformations are mapped onto discrete grids, the theoretical guarantees of equivariance degrade, leading to information leaks, numerical artifacts, and symmetry breaking.
The Root Cause: Grid-Induced Symmetry Breaking
To understand why equivariance fails, consider a standard 2D square pixel grid. This grid naturally possesses a discrete symmetry group: you can rotate an image by $90^\circ$, $180^\circ$, or $270^\circ$, and every pixel will land perfectly on another pixel.
But what happens if you rotate an image by $45^\circ$?
As Sheir A. Zaheer et al. demonstrate in REViT: Roto-reflection Equivariant Convolutional Vision Transformer, transformations that do not align with the grid's cardinal increments force the system to map integer coordinates to fractional coordinates. Because digital hardware cannot store a pixel at coordinate $(1.414, 1.414)$, the framework must approximate the value using interpolation.
Continuous Space Discrete Pixel Grid
(x, y) [ ] [ ] [ ]
| \ Rotate 45° [ ] * [ ] <- Pixel lands between grid points!
v v [ ] [ ] [ ] <- Requires interpolation (blurring)
(x', y')
This interpolation introduces a cascade of errors:
- Kernel Blurring: In Learning Steerable Filters for Rotation Equivariant CNNs, Maurice Weiler shows that rotating small discrete convolutional kernels (such as $3\times3$ or $5\times5$ grids) by arbitrary angles causes severe interpolation artifacts and blurring. The filter loses its high-frequency fidelity, meaning it no longer extracts the same features at different orientations.
- Anisotropy: When continuous space is discretized, interaction potentials and features begin to depend on their orientation relative to the grid axes. As noted in early computational physics frameworks like N. Ishii et al. (Phase shifts in $I=2$ ${\pi}{\pi}$-scattering from two lattice approaches), this loss of rotational invariance introduces systematic directional biases that are entirely artifacts of the coordinate system.
This problem is not unique to flat images. In Spherical CNNs, Taco Cohen et al. point out that while mathematical operations can be defined perfectly for continuous spheres, digital implementations must rely on discrete pixel grids or icosahedral meshes. Zhengyang Shen et al. expand on this in PDO-eS2CNNs: Partial Differential Operator Based Equivariant Spherical CNNs, proving that mapping continuous operators onto discrete mesh structures introduces measurable numerical approximation errors that prevent perfect rotation equivariance in practice.
The Nonlinearity Trap and Aliasing
The discretization gap becomes significantly worse when a network applies pointwise non-linear activation functions (like ReLU).
In a perfectly continuous network, convolutions and activations commute smoothly with geometric transformations. However, in a discrete network, features are represented via spatial or frequency-based samples. As Berfin Inal et al. explain in Adaptive Sampling for Continuous Group Equivariant Neural Networks, applying pointwise nonlinearities requires mapping frequency-based representations back to a spatial grid.
This creates a fundamental mathematical conflict:
- Non-linearities introduce high-frequency harmonics to the signal.
- If the underlying spatial grid is fixed and discrete, these new high frequencies will exceed the Nyquist limit of the grid.
- This causes aliasing—high-frequency signals masquerading as low-frequency noise.
Francesca Bartolucci et al. explore this mechanism deeply in Representation Equivalent Neural Operators: a Framework for Alias-free Operator Learning, demonstrating that these discrete aliasing errors are directly correlated with the violation of fundamental physical properties like translation equivariance. When the grid aliases, the network's response to a transformation leaks information, rendering the theoretical guarantees of the architecture invalid.
The same limitation applies to scale transformations. In Scale-Equivariant Steerable Networks, Ivan Sosnovik et al. show that scale equivariance is inherently bounded by the pixel grid. When an object or filter is downscaled significantly, sub-pixel information is permanently lost, causing a sharp, unavoidable spike in equivariance error.
Mitigating the Gap: How Modern Architectures Fight Back
To bridge the continuous-to-discrete gap, researchers have developed several distinct strategies to bound, bypass, or eliminate discretization errors.
1. Continuous Filter Parameterization via Differential Operators
Instead of defining a filter as a discrete grid of weights and rotating it via interpolation, one can define filters using continuous mathematics and discretize them only at the final step.
In PDO-eConvs: Partial Differential Operator Based Equivariant Convolutions, Zhengyang Shen et al. model convolutional filters as linear combinations of Partial Differential Operators (PDOs). By estimating derivatives via higher-order numerical schemes (like Taylor expansions), they prove that the resulting equivariance error is strictly bounded and scales quadratically with grid spacing ($\mathcal{O}(h^2)$). This provides a predictable mathematical framework where error can be minimized simply by controlling grid density.
2. Lie Group and Spline Interpolation
Another approach involves parameterizing filters over continuous transformation spaces using smooth basis functions. In B-Spline CNNs on Lie groups, Erik J Bekkers leverages the mathematics of Lie groups and B-splines to define continuous convolutions. By using smooth splines, the network can analytically sample transformations onto a grid, minimizing the information leaks and coordinate mismatches that plague standard bilinear interpolation.
3. Exact Alignment with Discrete Subgroups
When absolute mathematical precision is required, researchers often abandon continuous groups entirely in favor of discrete subgroups that match the grid geometry. As Gabriele Cesa argues in A Program to Build E(N)-Equivariant Steerable CNNs, discrete symmetry groups (such as those corresponding to Platonic solids or $90^\circ$ reflections) often outperform continuous groups on grid-based data. Because these discrete symmetries can be aligned exactly with the grid vertices, they completely bypass the interpolation errors inherent to arbitrary continuous rotations.
Summary
The power of equivariant deep learning stems from continuous group theory, but its execution is bottlenecked by discrete hardware. Discretization-induced equivariance error is the tax paid for mapping smooth, continuous symmetries onto rigid pixel, voxel, or mesh grids. Whether through alias-free operator design, continuous PDO parameterization, or restricting models to exact discrete subgroups, modern geometric deep learning is largely a fight to control and bound this hidden approximation gap.