Complete Local Frames in GNNs

• Complete local frames GNNs are architectures that incorporate per-node complete, orthonormal coordinate systems to ensure strict equivariance to transformations.
• They leverage geometric primitives and symmetric operations (e.g., SE(3), SO(3)) to facilitate anisotropic message passing and robust feature propagation.
• Applications include molecular property prediction, multi-agent control, and physical simulations, demonstrating improved generalization and efficiency over baseline models.

Complete local frames GNNs are a class of graph neural networks centrally characterized by the incorporation of per-node local coordinate systems (frames) that are constructed so as to be complete, orthonormal, and well-defined regardless of the underlying graph or data geometry. The explicit use of such frames—as opposed to fixed or incomplete bases—enables strict equivariance or invariance to global (and sometimes local) transformations from relevant symmetry groups (e.g., SE(3), SO(3), SE(2)), which is essential in geometric, physical, and dynamical domains. This architectural design improves robustness, generalization, expressiveness, and sample efficiency in tasks ranging from physical systems modeling to distributed control and molecular property prediction.

1. Mathematical Foundation: Local Frames and Group Equivariance

The canonical mathematical object is a local frame at each node of a geometric graph—an orthonormal basis, typically represented by a rotation matrix $F_i \in SO(d)$ for $\mathbb{R}^d$. For any tensorial feature $T$ at node $i$, projections into the local frame are obtained via group representations $D^{(\ell)}(F_i^{-1})T$, where $\ell$ denotes the tensor order and $D^{(\ell)}$ is the Wigner-D or equivalent representation for SO(d) (Du et al., 2021, Filling et al., 10 Nov 2025).

The explicit construction of these local frames carries two principal consequences:

• Completeness and Orthonormality: The frame spans the full physical space, avoiding degeneracy and ensuring well-posed projections for any geometric object.
• Equivariance/Invariance by Design: Under global group action (rotation, translation), both features and local frames are transformed such that the composite output is predictable—typically, equivariant or invariant—allowing the GNN to respect the underlying physical symmetries.

In practice, frame construction uses only geometric primitives—pairwise vectors, cross products, eigendecomposition—eschewing reliance on global axes and thus preventing direction degeneration and numerical instability (Du et al., 2021, Filling et al., 10 Nov 2025).

2. Frame Construction Techniques

The construction of complete local frames varies slightly depending on the domain and desired symmetry:

• SE(3) (3D): At each node $i$, select two non-collinear neighbors $j,k$ and let

$$\begin{aligned} e_1 &= \frac{r_{ij}}{\| r_{ij} \|}, \ e_2 &= \frac{r_{ij} \times r_{ik}}{\| r_{ij} \times r_{ik} \|}, \ e_3 &= e_1 \times e_2, \ F_{i;jk} &= [ e_1 \;|\; e_2 \;|\; e_3 ] \in SO(3). \end{aligned}$$

This basis is robust to geometric degeneracy, as the cross product ensures non-parallelism except in degenerate cases, which are handled by alternative neighbor choices or fallback strategies (Du et al., 2021).

• SO(3) (Molecules): Charge-weighted or spatially informed local PCA frames are constructed by weighted covariance of neighbor positions, then axis disambiguation via eigenvector sign and projection procedures. The requirement $\mathbb{R}^d$0 separates regular from symmetric (ambiguous) cases, leading to deterministic, unique frames suitable for equivariant pooling (Filling et al., 10 Nov 2025).
• SE(2) (Planar Robotics): Identification of planar vectors with complex numbers allows for local rotations to be efficiently represented by complex multiplication $\mathbb{R}^d$1, simplifying both analysis and implementation of rotationally-invariant GNN layers (Honor et al., 3 Apr 2026).
• General Dynamical Systems: If explicit orientation is part of node state, local frames may be defined by Euler angles or constructed from velocity and reference axes, with careful handling of gimbal lock and collinearities using Gram–Schmidt or predefined backbone vectors (Kofinas et al., 2021).

3. Message Passing in Complete Local Frames

In these GNNs, all node and edge features are canonically transformed into their respective local frames. Message-passing proceeds as follows:

• Feature Transport: For each edge $\mathbb{R}^d$2, vector and tensor features from node $\mathbb{R}^d$3 are rotated into $\mathbb{R}^d$4's frame via the relative rotation $\mathbb{R}^d$5, i.e., $\mathbb{R}^d$6, $\mathbb{R}^d$7 (Filling et al., 10 Nov 2025, Du et al., 2021).
• Equivariant Operations: Message functions, aggregation (mean, sum), update (linear or tensorial mixing), and nonlinearities (e.g., phase-magnitude tanh, gated tensor activations) are constructed to be pointwise equivariant with respect to the group action. Composition of equivariant maps preserves equivariance, ensuring correct transformation at any network depth (Honor et al., 3 Apr 2026, Filling et al., 10 Nov 2025, Du et al., 2021).
• Anisotropic Filtering: The dependence of message-passing weights on local-frame features allows for directionally-sensitive, anisotropic filtering, improving the flexibility and inductive bias of the architecture, especially for physical and dynamical systems (Kofinas et al., 2021).

4. Architectural Variants and Domains of Application

Several notable architectures exemplify the complete local frames paradigm:

Model	Symmetry Group	Local Frame Strategy	Key Domain(s)	Reference
CLF-SE(3) GNN	SE(3)	Cross-product basis	Physics, molecules	(Du et al., 2021)
DiMolPol	SO(3)	PCA, weighted axes	Molecular polarizability	(Filling et al., 10 Nov 2025)
Complex GNN (2D)	SO(2)	Complex domain	Planar robot control	(Honor et al., 3 Apr 2026)
LoCS	SE(2)/SE(3)	Gram–Schmidt, kinematic	Interacting dynamics	(Kofinas et al., 2021)
Framelet GNN	Spectral	Tight spectral frame	Multiscale, spectral GNNs	(Liu et al., 2023)

Applications include:

• Learned control of multi-agent systems and decentralized robots in GPS/compass-denied environments (Honor et al., 3 Apr 2026).
• Molecular property prediction: e.g., direct regression of polarizability tensors, without using scalar-only intermediates (Filling et al., 10 Nov 2025).
• Physical simulation and dynamical interaction modeling, where generalization across reference frames and sample efficiency are critical (Du et al., 2021, Kofinas et al., 2021).
• Multiscale graph learning using tight spectral frames (“framelets”) as “local” multiresolution bases, improving oversmoothing and depth scalability due to frame energy preservation (Liu et al., 2023).

5. Empirical Performance and Quantitative Benefits

Empirical studies consistently demonstrate that complete local frame GNNs achieve superior performance by aligning inductive bias with the underlying symmetry:

• Control Tasks (2D): Basis-invariant complex GNNs match the nominal flocking controller with nearly zero error using only 2 layers and 8 hidden units, outperforming real-valued baselines both in trajectory tracking and generalization to novel communication radii or longer rollouts (Honor et al., 3 Apr 2026).
• Molecular Tensor Prediction: Fully tensorial message passing with local frames (S/V/T channels) yields lower mean absolute errors across all polarizability metrics compared to scalar-only GNNs (e.g., Tensor MAE 0.45 vs. 0.53; Anisotropy MAE 0.42 vs. 0.52), and the error under SO(3) perturbation is at machine-precision levels (∼10⁻⁷ relative Frobenius) (Filling et al., 10 Nov 2025).
• Physical Modeling: CLF-based SE(3)-equivariant GNNs achieve 10–15% lower MSE on Newtonian mechanics and ∼5% tighter conformer RMSD on molecular datasets compared to SE(3)-Transformer and EGNN baselines. Ablations confirm full frame completeness is crucial: removing completeness increases MSE by ∼12%, and using naive Gram–Schmidt can degrade performance by >25% (Du et al., 2021).
• Dynamical Systems: The LoCS method yields 15–20% lower test MSE on complex dynamical prediction tasks than variants with incomplete or misaligned frames, and achieves highest sample-efficiency and long-term prediction stability (Kofinas et al., 2021).

6. Limitations, Challenges, and Extensions

Despite their advantages, complete local frame GNNs are subject to domain- and architecture-specific limitations:

• Degenerate Geometries: Local frame construction can be ill-conditioned in collinear or highly symmetric neighborhoods; remedies include explicit neighbor selection, fallback strategies, or learned/robust frame construction (Du et al., 2021, Filling et al., 10 Nov 2025).
• Higher-Order Tensor Complexity: Handling very high-order representations or aggregates (large ℓ) can become computationally prohibitive due to Wigner-D and Clebsch–Gordan complexity (Du et al., 2021).
• Group Coverage: Most methods are equivariant to rotations (SO(d)), but not necessarily to reflections (O(d)\SO(d)), limiting applicability in handedness-sensitive tasks (Filling et al., 10 Nov 2025).
• Spectral Framelets: In spectral GNNs (e.g., Framelet Message Passing), “local frames” are defined in terms of tight frames in the Laplacian eigenbasis rather than geometric coordinate bases. These handle energy preservation and multiscale aggregation but lack explicit spatial directional information (Liu et al., 2023).

Extension directions include robustification of local frame construction, adaptive or learned frames, extension to new groups (e.g., E(3), SE(d)), and broader application domains such as molecular dynamics and large-scale multigraphs (Filling et al., 10 Nov 2025, Du et al., 2021).

7. Theoretical and Practical Significance

The introduction of complete local frames within GNNs represents a systematic and theoretically-grounded approach to embedding geometric and physical symmetries into graph-based models. This architectural paradigm leads to strictly equivariant (or invariant) mappings with respect to physically meaningful transformations, supports multiscale directional reasoning, and enables anisotropic feature propagation.

The practical impact is evidenced by improved generalization, data efficiency, empirical accuracy, and stability across a variety of geometric and dynamical tasks. The approach remains an active area of research, with ongoing developments in scalable construction, domain adaptation, robust handling of degeneracies, and integration with continuous-depth and spectral frameworks (Honor et al., 3 Apr 2026, Filling et al., 10 Nov 2025, Du et al., 2021, Liu et al., 2023, Kofinas et al., 2021).