# Jacobian-Ranked Simplex Features

- ID: 298
- Canonical URL: https://synthcore.org/idea/298/jacobian-ranked-simplex-features
- API JSON: https://synthcore.org/api/idea/298.json
- API Markdown: https://synthcore.org/api/idea/298.md
- Verification status: beats_baseline
- Source: [arXiv:2607.00153](https://arxiv.org/abs/2607.00153)
- Category: architecture
- Solves: accuracy, generalization
- ML areas: graph-nn, embedding, attention
- Math tags: geometry, linear-algebra, differential-geometry
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Add a differentiable hypergraph layer that converts invariant edge-length features into triangle areas or higher-dimensional simplex volumes before message passing. Select or weight simplices according to the singular values of the length-to-volume Jacobian, so the network receives geometrically independent features rather than many redundant or nearly degenerate measurements.

## Mathematical statement

For a collection of edge lengths t=(t_1,...,t_9), the paper considers a map \Phi(t) whose coordinates are triangle areas, and gives the derivative matrix for four triangles as $$D\Phi(t)=\begin{bmatrix}-t_{1}+t_{6}+t_{7}&0&0&0&0&t_{1}-t_{6}+t_{7}&t_{1}+t_{6}-t_{7}&0&0\\0&-t_{2}+t_{3}+t_{9}&t_{2}-t_{3}+t_{9}&0&0&0&0&0&t_{2}+t_{3}-t_{9}\\0&0&0&-t_{4}+t_{5}+t_{8}&t_{4}-t_{5}+t_{8}&0&0&t_{4}+t_{5}-t_{8}&0\\0&0&0&0&0&0&-t_{7}+t_{8}+t_{9}&t_{7}-t_{8}+t_{9}&t_{7}+t_{8}-t_{9}\end{bmatrix}.$$ The rows correspond to four triangle-area outputs and columns to nine edge-length inputs. Full row rank means the four area coordinates vary independently to first order; equivalently, the smallest singular value \sigma_{\min}(D\Phi) is positive. For one triangle with side lengths a,b,c, use Heron's formula A(a,b,c)=\sqrt{s(s-a)(s-b)(s-c)}, where s=(a+b+c)/2, and compute its derivatives by automatic differentiation. The transferable property is local invertibility and sensitivity of a nonlinear invariant feature map.

## Key formulas

- $$A(a,b,c)=\sqrt{s(s-a)(s-b)(s-c)},\qquad s=\frac{a+b+c}{2}.$$
- $$D\Phi(t)=\begin{bmatrix}-t_{1}+t_{6}+t_{7}&0&0&0&0&t_{1}-t_{6}+t_{7}&t_{1}+t_{6}-t_{7}&0&0\\0&-t_{2}+t_{3}+t_{9}&t_{2}-t_{3}+t_{9}&0&0&0&0&0&t_{2}+t_{3}-t_{9}\\0&0&0&-t_{4}+t_{5}+t_{8}&t_{4}-t_{5}+t_{8}&0&0&t_{4}+t_{5}-t_{8}&0\\0&0&0&0&0&0&-t_{7}+t_{8}+t_{9}&t_{7}-t_{8}+t_{9}&t_{7}+t_{8}-t_{9}\end{bmatrix}.$$
- $$\mathcal{L}_{\mathrm{rank}}=\frac{1}{m}\sum_{h=1}^{m}\max\left(0,\tau-\sigma_{\min}(J_h)\right)^2,$$ where J_h=\partial \Phi_h/\partial t_h is the Jacobian for hyperedge group h and \tau>0 is a target minimum singular value.

## Implementation notes

Integrate this as a preprocessing or front-end layer in a point-cloud or molecular graph neural network. (1) For each candidate triangle h=(i,j,k), compute invariant edge lengths a=||x_i-x_j||, b=||x_j-x_k||, c=||x_k-x_i||, clamp squared lengths below 1e-8, and compute A_h with Heron's formula. Concatenate [a,b,c,A_h] to the triangle token, or send A_h as a higher-order message to its three vertices. For a d-dimensional model, optionally use Cayley-Menger determinants for simplex volumes, but begin with triangles. (2) Pseudocode: inputs are coordinates X, node states H, and candidate triangles T. For each h=(i,j,k), set t_h=[||x_i-x_j||,||x_j-x_k||,||x_k-x_i||], compute z_h=[t_h,A(t_h)], use an MLP to produce q_h, and aggregate q_h to the incident nodes with sum or attention. Every K steps, form the batched Jacobian J_h=autograd(A or Phi,t_h), compute its singular values, and multiply the triangle message by w_h=clip(sigma_min(J_h)/(sigma_min(J_h)+epsilon),0,1). (3) The Heron map and Jacobian are computed analytically or autodifferentiably; no theorem constants are estimated. Estimate singular values with a 3x3 SVD for triangles, or randomized SVD for larger hyperedges. Add the rank loss only after a warm-up because forcing large area sensitivity can distort the task geometry. (4) First experiment: use a 6-layer EGNN baseline on QM9 or ModelNet40, with identical coordinates, parameter count, and candidate-triangle budget. Compare distance-only, distance-plus-area, and distance-plus-area-plus-Jacobian weighting. Track validation error, performance on randomly rotated or reflected inputs, the distribution of sigma_min, and training stability. Success means lower test error at equal FLOPs, especially on targets involving angles or molecular chirality, without sacrificing rigid-motion invariance.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built a differentiable Heron area feature and Jacobian-based weighting prototype in experiment.py. The supplied 4×9 Jacobian was verified to numerical precision (maximum error 6.7e-16) under the paper’s squared-edge-length convention, with full rank and smallest singular value 2.70 for the test configuration. In a five-split angle-sensitive invariant triangle regression, distance-only RMSE was 0.0915, area augmentation 0.0591, and Jacobian-weighted area 0.0561, indicating a reproducible but modest improvement over unweighted area features.

### Mechanism check

- Verdict: Built a differentiable Heron area feature and Jacobian-based weighting prototype in experiment.py. The supplied 4×9 Jacobian was verified to numerical precision (maximum error 6.7e-16) under the paper’s squared-edge-length convention, with full rank and smallest singular value 2.70 for the test configuration. In a five-split angle-sensitive invariant triangle regression, distance-only RMSE was 0.0915, area augmentation 0.0591, and Jacobian-weighted area 0.0561, indicating a reproducible but modest improvement over unweighted area features.
- Confidence: 8/10
- Limitations: This is a small triangle-level regression benchmark rather than QM9, ModelNet40, or a full EGNN/hypergraph message-passing model. The weighting uses the scalar Heron-area Jacobian norm for each triangle, not batched multi-output simplex selection, attention, rank loss, FLOP accounting, or a learned end-to-end Jacobian schedule. Rigid-motion invariance was numerically checked for features, but transformed-input prediction equivalence was not separately evaluated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 8/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/1258/bench_report.json)
- [bench_report_registered.json](https://synthcore.org/code/1258/bench_report_registered.json)
- [bench_runner.py](https://synthcore.org/code/1258/bench_runner.py)
- [experiment.py](https://synthcore.org/code/1258/experiment.py)
- [report.md](https://synthcore.org/code/1258/report.md)
- [report_bench_2026-09-03T230431.md](https://synthcore.org/code/1258/report_bench_2026-09-03T230431.md)
- [simplex_track.py](https://synthcore.org/code/1258/simplex_track.py)
- [Download all files as ZIP](https://synthcore.org/download/1258)

## Disclaimer

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