# Minor-consistency regularizer for homogeneous-space coordinates

- ID: 893
- Canonical URL: https://synthcore.org/idea/893/minor-consistency-regularizer-for-homogeneous-space-coordina
- API JSON: https://synthcore.org/api/idea/893.json
- API Markdown: https://synthcore.org/api/idea/893.md
- Verification status: unverified
- Source: [arXiv:2607.14132](https://arxiv.org/abs/2607.14132)
- Category: geometry
- Solves: stability, accuracy, generalization
- ML areas: graph-nn, embedding, training-dynamics
- Math tags: linear-algebra, lie-groups, differential-geometry, geometry
- Ratings: usefulness 6/10; difficulty 6/10; novelty 8/10

## Idea description

When a network learns coordinates q on a homogeneous space from symmetry-generated vector fields, enforce that the predicted Jacobian is compatible with all generator equations using augmented-matrix consistency residuals. This turns the paper's rank and minor criterion into a differentiable regularizer that prevents locally contradictory coordinate derivatives and can produce more stable equivariant representations.

## Mathematical statement

The paper defines r_beta^(alpha)=partial q^alpha/partial x^beta and obtains an overdetermined linear system A(x)r^(alpha)=b^(alpha)(x,q), where A contains coefficients of symmetry vector fields and b contains transformed target components. Consistency is stated as equality of the ranks of the coefficient and augmented matrices, equivalently vanishing of every augmented minor of size m+1: M_i^(alpha)(x,a,q)=0. Here n is the number of equations, m is the number of unknown coordinate derivatives, A is in R^(n by m), b is in R^n, and L=[A b] is in R^(n by (m+1)). Instead of explicitly computing all combinatorial minors, use the equivalent least-squares compatibility residual R=(I-A A^+)b, where A^+ is a damped pseudoinverse; R=0 exactly when b lies in the column space of A, assuming A has the intended rank.

## Key formulas

- $$r_{\beta}^{(\alpha)}=\frac{\partial q^{\alpha}}{\partial x^{\beta}},\qquad A(x)r^{(\alpha)}=b^{(\alpha)}(x,q).$$
- $$M_{i}^{(\alpha)}(x,a,q)=0,\quad i=1,\dots,\binom{n}{m+1},$$
- $$L^{(\alpha)}(x,a,q)=\begin{pmatrix}A(x)&b^{(\alpha)}(x,q)\end{pmatrix},\qquad \operatorname{rank}A=\operatorname{rank}L,$$
- $$\mathcal L_{\mathrm{compat}}=\frac{1}{B}\sum_{j=1}^{B}\left\|\left(I-A_j(A_j^{\mathsf T}A_j+\varepsilon I)^{-1}A_j^{\mathsf T}\right)b_j\right\|_2^2.$$

## Implementation notes

Integrate this into a symmetry-aware coordinate encoder or graph neural network. Assume a known set of n infinitesimal generator equations at each sample x, represented by A(x) with shape n-by-m; the network q_phi(x) supplies coordinates q, and autodiff supplies the Jacobian entries r=J_q(x). The target vector b(x,q) is computed from the generator action; for a prototype, use translations and rotations on R^2, or construct A from fixed graph or Lie-algebra generator matrices. During training compute A_j, b_j, and the damped projector P_perp=I-A_j(A_j^T A_j+epsilon I)^(-1)A_j^T. Add lambda*||P_perp*b_j||^2 to the task loss. If the matrices are tiny, explicitly enumerate every (m+1)-minor of L and compare normalized squared minors against the projector implementation; this validates the adaptation of the paper's rank criterion. Do not differentiate through an unstable exact pseudoinverse: use epsilon between 1e-4 and 1e-2 or solve the normal equations with Cholesky. The paper's contribution is the rank-equivalence condition; the projector residual is the differentiable numerical estimator. First experiment with an MLP encoding points on a rotated synthetic manifold and predicting a symmetry-invariant target, comparing a standard encoder, a conventional equivariance penalty, and this compatibility loss. Measure held-out transformation error, task accuracy, Jacobian singular-value conditioning, and optimization variance. Success is lower held-out transformation error at equal task loss and fewer exploding or ill-conditioned Jacobian batches.

## Disclaimer

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