e-equivariant graph neural networks

**E(n)-Equivariant Graph Neural Networks (EGNN)** are **graph neural network architectures that process 3D point clouds (atoms, particles) while guaranteeing that the output transforms correctly under rotations, translations, and reflections** — if the input molecule is rotated by angle $ heta$, all output vectors rotate by exactly $ heta$ (equivariance) and all output scalars remain unchanged (invariance) — achieved through a lightweight coordinate-update mechanism that avoids the expensive spherical harmonics and tensor products used by other equivariant architectures. **What Is EGNN?** - **Definition**: EGNN (Satorras et al., 2021) processes graphs with 3D node positions $mathbf{x}_i in mathbb{R}^3$ and feature vectors $mathbf{h}_i in mathbb{R}^d$. Each layer updates both positions and features: (1) **Message**: $m_{ij} = phi_e(mathbf{h}_i, mathbf{h}_j, |mathbf{x}_i - mathbf{x}_j|^2, a_{ij})$ — messages depend on features and the squared distance (rotation-invariant); (2) **Position Update**: $mathbf{x}_i' = mathbf{x}_i + C sum_{j} (mathbf{x}_i - mathbf{x}_j) phi_x(m_{ij})$ — positions shift along the direction to each neighbor, weighted by a learned scalar; (3) **Feature Update**: $mathbf{h}_i' = phi_h(mathbf{h}_i, sum_j m_{ij})$ — features aggregate messages. - **Equivariance Proof**: The position update uses only the relative direction vector $(mathbf{x}_i - mathbf{x}_j)$ multiplied by a scalar function of invariant quantities (features + distance). When the input is rotated by $R$, the direction vector transforms as $R(mathbf{x}_i - mathbf{x}_j)$, and the scalar coefficient is unchanged (depends only on invariants), so the output position transforms as $Rmathbf{x}_i' + t$ — exactly E(n)-equivariant. Features depend only on distances (invariants) and are therefore rotation-invariant. - **Lightweight Design**: Unlike Tensor Field Networks and SE(3)-Transformers that use spherical harmonics ($Y_l^m$) and Clebsch-Gordan tensor products (expensive $O(l^3)$ operations), EGNN achieves equivariance using only MLPs and Euclidean distance computations — no special mathematical functions, no irreducible representations. This makes EGNN significantly faster and easier to implement. **Why EGNN Matters** - **Molecular Property Prediction**: Molecular properties (energy, forces, dipole moments) depend on the 3D arrangement of atoms, not just the 2D bond graph. EGNN processes 3D coordinates natively and invariantly — predicting the same energy regardless of how the molecule is oriented in space, which is physically required since molecules tumble freely in solution. - **Molecular Dynamics**: Predicting atomic forces for molecular dynamics simulation requires E(3)-equivariant outputs — force on atom $i$ must rotate with the molecule. EGNN's equivariant position updates provide the correct geometric behavior for force prediction, enabling neural network-based molecular dynamics that are orders of magnitude faster than quantum mechanical calculations. - **Foundation for Generative Models**: EGNN serves as the denoising network inside Equivariant Diffusion Models (EDM) — the lightweight equivariant architecture processes noisy 3D atom positions and predicts the denoising direction, generating 3D molecules that respect physical symmetries. Without efficient equivariant architectures like EGNN, 3D molecular generation would be computationally impractical. - **Simplicity vs. Expressiveness Trade-off**: EGNN's simplicity comes at a cost — it uses only scalar messages and pairwise distances, which limits its ability to capture angular information (bond angles, dihedral angles). More expressive models (DimeNet, PaiNN, MACE) incorporate directional information at higher computational cost. EGNN represents the "minimal equivariant" baseline that is fast, simple, and sufficient for many applications. **EGNN vs. Other Equivariant Architectures** | Architecture | Angular Info | Tensor Order | Relative Speed | |-------------|-------------|-------------|----------------| | **EGNN** | Distances only | Scalars + vectors | Fastest | | **PaiNN** | Distance + direction vectors | Up to $l=1$ | Fast | | **DimeNet** | Distances + bond angles | Bessel + spherical harmonics | Moderate | | **MACE** | Multi-body correlations | Up to $l=3+$ | Slower, most accurate | | **SE(3)-Transformer** | Full SO(3) representations | Arbitrary $l$ | Slowest | **EGNN** is **geometry-native neural processing** — understanding the 3D shape of molecules through coordinate updates that mathematically guarantee rotational equivariance, providing the efficient equivariant backbone for molecular property prediction, force field learning, and 3D molecular generation.