Polar-Gauge SPD Feature Layer
Implementation & benchmark of arXiv:2607.14204 — A Self-Dual Frame Formalism of the SO(3) Yang-Mills Theory
Source paper: A Self-Dual Frame Formalism of the SO(3) Yang-Mills Theory arXiv:2607.14204 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace a locally oriented three-channel feature frame by its positive-definite polar factor, removing arbitrary SO(3) basis rotations before the feature enters an MLP, attention block, or graph message-passing layer. Process the resulting SPD matrix in log coordinates so the downstream network receives a globally unconstrained symmetric representation rather than a gauge-dependent frame.
Formulas
Mathematical statement
The paper uses pointwise polar decomposition on an orientation-preserving map X from the self-dual frame space to R^3, writing X = Qh, where Q is the orthogonal factor in SO(3) and h = (X^T X)^(1/2) is the positive-definite symmetric factor. Under a local basis change X -> RX with R in SO(3), the factor Q changes while h is unchanged, so h is a gauge-invariant descriptor. The paper also states that a positive frame determines a compatible connection through the invertible map ad_h(a) = [a,h], where a is a connection perturbation and [a,h] is the corresponding bracket-valued form. The neural adaptation uses the algebraic polar slice: for a feature frame X in R^(3x3), compute h = (X^T X + epsilon I)^(1/2), then S = log h and retain the six independent entries of the symmetric matrix S. Epsilon is a positive numerical stabilizer that ensures the Gram matrix remains positive definite.
Implementation notes
Integrate the layer immediately after a network constructs a local 3-by-3 feature frame. In a graph neural network, let each node receive three learned channels associated with three spatial directions or three grouped feature vectors. In an attention model, apply the transformation independently to every token after a linear projection reshaped into a 3-by-3 matrix. For each batch element, take the input frame X, compute G = X^T X + epsilon I, eigendecompose G = U diag(lambda_i) U^T, clamp every lambda_i to at least epsilon, and construct h = U diag(sqrt(lambda_i)) U^T. Then construct S = U diag(0.5 log(lambda_i)) U^T, extract its six upper-triangular entries with vech(S), and concatenate them with optional scalar invariants such as log det(h) before the existing MLP or attention projection. Monitor the smallest eigenvalue of X^T X and optionally add the penalty max(0, tau - lambda_min(X^T X))^2 to discourage degenerate frames. The paper supplies the polar-gauge construction; eigendecomposition, log-SPD coordinates, epsilon, and the conditioning penalty are engineering adaptations. The first cheap experiment should use a small PointNet or graph neural network on ModelNet40, or a synthetic point-cloud classification task, with independently sampled random rotations R in SO(3) applied to every example. Compare a standard baseline with the polar-gauge version at equal parameter count. Evaluate clean accuracy, accuracy under unseen rotations, and the invariance error ||f(RX) - f(X)||_2. The expected signal is substantially lower rotation variance and improved rotated-test accuracy without increased parameter count.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.