# Horizontal Symmetry Quotient Layer

- ID: 802
- Canonical URL: https://synthcore.org/idea/802/horizontal-symmetry-quotient-layer
- API JSON: https://synthcore.org/api/idea/802.json
- API Markdown: https://synthcore.org/api/idea/802.md
- Verification status: unverified
- Source: [arXiv:2607.12350](https://arxiv.org/abs/2607.12350)
- Category: architecture
- Solves: scalability, stability, generalization
- ML areas: graph-nn, cnn, attention, embedding
- Math tags: lie-groups, differential-geometry, geometry, symplectic
- Ratings: usefulness 6/10; difficulty 6/10; novelty 5/10

## Idea description

Replace explicit quotient construction by a differentiable projection that removes learned group-orbit directions from both source and target features. The paper's reduction argument shows that a closed level constraint makes the restricted form horizontal, so the network can operate on invariant coordinates while retaining a measurable residual for symmetry leakage.

## Mathematical statement

For a Lie-group action with infinitesimal generators v_xi^M on M and v_xi^N on N, the paper uses the relative Hamilton equation d_F μ_xi=-iota_{v_xi} varpi and a closed level pair φ with d_F φ_xi=0. On level sets L_M and L_N, the target component satisfies iota_{v_xi^N} i_N*ω=0, and the source component satisfies iota_{v_xi^M} i_M*η=0. Here μ_xi is the moment-map component for Lie-algebra element ξ, φ_xi is its fixed level, i_M and i_N are inclusions of the level sets, ω is the target form, and η is the source form. The engineering consequence is that symmetry-generated tangent directions lie in the kernel of the restricted forms. A practical approximation removes those directions with P=I−V(V^TV+εI)^−1V^T, where V stacks the infinitesimal generators and ε>0 stabilizes the solve.

## Key formulas

- $$d_F\mu_{\xi}=-\iota_{v_{\xi}}\varpi,$$
- $$\iota_{v_{\xi}^{N}}\,i_N^{*}\omega=0,$$
- $$\iota_{v_{\xi}^{M}}\,i_M^{*}\eta=0,$$
- $$P(x)=I-V(x)\bigl(V(x)^TV(x)+\varepsilon I\bigr)^{-1}V(x)^T,\qquad h_{\mathrm{quot}}=P(x)h.$$

## Implementation notes

Insert the layer after an encoder produces feature vectors h_M for source objects and h_N for target objects, before attention, message passing, or classification. Specify a compact Lie-group action or learn its infinitesimal generators from known transformations. For each sample x and Lie-algebra basis element ξ_r, compute v_r(x)=∂_t[g_r(t)·x] at t=0 and stack these columns as V(x) in R^(d×r). Form the differentiable damped projector P=I−V(V^TV+εI)^−1V^T and replace h by Ph. If an explicit level constraint exists, compute a soft level weight w=exp(−||μ(x)−φ||²/τ) and apply the projection mainly to samples with large w; alternatively use P globally and add an invariance loss. Pseudocode: compute h; compute V; solve (V^TV+εI)u=V^Th; set h_q=h−Vu; feed h_q downstream; add lambda_orbit||f(g·x)−f(x)||² and optionally lambda_form||V^T∇_h s||². The theorem supplies the horizontality target, while V, μ, and the level residual are computed from the action or autodiff. Test rotations and translations on synthetic point clouds, then a small image or graph model against augmentation and an equivariant baseline. Measure transformed-input accuracy, orbit sensitivity, FLOPs, and gradient variance. Success is equal or better accuracy with substantially lower orbit sensitivity and at least 20% fewer downstream feature dimensions at matched accuracy.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.