✓✓ Beats tuned baseline 2026

Intrinsic Tangent-Projected Point-Cloud Layer

Implementation & benchmark of arXiv:2609.00216 — Solving the Incompressible Navier-Stokes Equations on Oriented Curved Surfaces Discretized by Point Clouds

Usefulness7/10

Difficulty4/10

Novelty5/10

Architecture Accuracy Stability Generalization Graph nn World model Geometry Differential geometry Geometry Numerical analysis

Source paper: Solving the Incompressible Navier-Stokes Equations on Oriented Curved Surfaces Discretized by Point Clouds arXiv:2609.00216 ⓘ · analyzed Sep 2, 2026

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 Not peer-reviewed.

Idea description

Modify a point-cloud message-passing or neural-operator layer so that scalar gradients, vector features, and vector-to-vector interactions are computed only in the estimated tangent plane of the surface. Projecting both feature values and derivative directions prevents the network from using arbitrary ambient-space normal directions and should improve transfer across differently embedded but intrinsically similar surfaces.

Formulas

$$\mathsf P(x)=I-n(x)n(x)^\top,$$

$$\nabla f=\mathsf P(\overline{\nabla}\,\overline f),$$

$$\nabla\boldsymbol v=\mathsf P(\overline{\nabla}\,\overline{\boldsymbol v})\mathsf P,$$

$$[\nabla\boldsymbol v]_{ij}=\mathsf P_{ik}[\overline{\nabla}\overline{\boldsymbol v}]_{kl}\mathsf P_{lj}.$$

Mathematical statement

For an oriented surface $\mathcal S\subset\mathbb R^3$ with unit normal $n(x)$, let $\mathsf P(x)=I-n(x)n(x)^\top$ be the orthogonal tangent projector. The paper defines the intrinsic gradient of a scalar extension $\bar f$ by $\nabla f=\mathsf P(\bar\nabla\bar f)$, and the intrinsic covariant derivative of an ambient vector extension $\bar{\boldsymbol v}$ by $\nabla\boldsymbol v=\mathsf P(\bar\nabla\bar{\boldsymbol v})\mathsf P$. Here $I$ is the $3\times3$ identity, $n$ is the estimated surface normal, $\bar\nabla$ is the ordinary ambient derivative, and $\nabla$ is the surface derivative. The left projector removes normal components from the derivative output; the right projector restricts the input direction to tangent directions. For a point-cloud network, these identities become a differentiable architectural constraint rather than merely a discretization formula.

Implementation notes

Integrate this into a local point-cloud graph layer whose input is point coordinates $x_i\in\mathbb R^3$, scalar features $s_i$, and tangent vector features $v_i\in\mathbb R^3$. First estimate a unit normal $n_i$ from the centered $k$-nearest-neighbor covariance matrix: take the eigenvector associated with the smallest eigenvalue, orient neighboring normals consistently when possible, and set $P_i=I-n_i n_i^\top$. Before every vector-feature update, enforce $v_i\leftarrow P_i v_i$. For an edge $i\to j$, use the tangent displacement $r_{ij}=P_i(x_j-x_i)$, its norm $d_{ij}=\|r_{ij}\|$, and an MLP message $m_{ij}=\phi([s_i,s_j,v_i,v_j,r_{ij},d_{ij}])$. If the layer produces a vector message, reshape it as an ambient vector and apply $P_i$; if it produces a linear map from vector features, implement the covariant analogue $A_{ij}\leftarrow P_i A_{ij}P_i$. A direct derivative branch can estimate an ambient Jacobian $J_i$ by weighted least squares, solving $\min_J\sum_{j\in N(i)}w_{ij}\|v_j-v_i-J(x_j-x_i)\|^2$, then use $P_iJ_iP_i$ as the intrinsic Jacobian feature. Compute projectors and least-squares coefficients from geometry; learn only the MLP parameters. Start with a six-layer EGNN or PointNet-style baseline on synthetic scalar and tangent-vector fields over spheres, tori, and randomly deformed embeddings, equalizing parameters, FLOPs, and training steps. Pre-register three mechanism tests: the mean normal velocity component $|n_i^\top v_i|$ must remain below $10^{-5}$ after every layer; under a rigid rotation of the entire point cloud, prediction error should change by less than 1%; and training on one embedding and testing on an isometric re-embedding should improve relative error by at least 10% at fixed sample count. Remove only the left and right projectors for the ablation. The transfer is falsified if projection does not reduce normal leakage, gives no rotation or re-embedding benefit, or consistently harms in-distribution accuracy by more than 2%.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 7/10

Built a compact intrinsic point-cloud layer with kNN normal estimation, tangent displacement projection, vector-feature projection, and numerical projector/Jacobian checks. The identities held to float32 precision, and the intrinsic model reduced held-out relative RMSE from 0.0668 to 0.0525, a 21.4% improvement in the fixed noisy rotated-sphere toy setup. This is a promising signal for the claimed effect, but not evidence of broad surface-transfer gains.

Agent confidence: 7/10

Baseline: Relative held-out RMSE 0.0668143
Idea: Relative held-out RMSE 0.0525420; 21.361% improvement; raw mean normal velocity 1.6410 versus projected maximum normal velocity 9.24e-7; projector idempotence max error 1.71e-7

Limitations: Only a small synthetic sphere experiment was tested. Torus/deformed-isometric surfaces, explicit rotation-invariance error comparison, multi-seed statistics, six-layer stacks, learned vector outputs, weighted least-squares Jacobian estimation, FLOP matching, and larger point-cloud scalability were not tested.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2609.00216) Sep 2, 2026

Beats the tuned baseline by 21.7% (p=0.0381, wins 7 / 8 paired seeds; metric: lower is better).

Benchmark: Tangent_surface_vector_regression
Model: shared_mlp
Paired seeds: 8
Baseline mean: 0.0275
Idea mean: 0.0215
Effect (Δ): -0.006 (−21.7%; negative = idea better)
Wins: 7 / 8 paired seeds
p-value: 0.0381 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±19.5%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

This is a small synthetic sphere-based custom track with 400 training and 100 test samples per seed, a shared MLP rather than a full multi-layer message-passing network, 14 epochs, and three learning-rate settings. It tests tangent projection and normal suppression but not torus/deformed-isometric transfer, weighted least-squares Jacobian features, FLOP matching beyond the shared MLP shape, large-scale point clouds, or the proposed <1% rotation-change target; measured rotation prediction change was 25.5%.

How to run: python3 run_stage2.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 2, 2026 latest Worked ✓

Agent confidence: 8/10

Built a structurally matched custom point-cloud benchmark because no built-in track represents tangent-surface geometry. Across 8 paired seeds, the independently trained tangent-projected system reduced MSE from 0.027478 to 0.021507, with delta -0.005971 and permutation p=0.03805, meeting the harness criterion for a significant win. The trained-model mechanism signature confirmed normal leakage suppression (0.6703 raw mean component versus 4.77e-7 projected maximum), but rotation prediction change was 25.5%, so full rotational invariance was not achieved.

Baseline: MSE 0.0274779503; best lr=0.01; 8-seed mean
Idea: MSE 0.0215066459; best lr=0.01; 8-seed mean; absolute improvement 0.0059713044 (21.73%)
Idea wins: 7/8
Mde: 0.0053456830 (19.45% of baseline mean)
Paired delta: -0.0059713044
Permutation p: 0.03805

Benchmark result