✗ Failed on benchmark 2026

Ellipsoidal-Preserving Spherical Feature Stabilizer

Implementation & benchmark of arXiv:2607.09412 — Quantitative stability of the intersection body operator near the ball, and the dynamical origin of the two--dimensional degeneracy

Usefulness6/10

Difficulty6/10

Novelty6/10

Regularization Stability Generalization Sample efficiency Embedding Graph nn Regularization Geometry Harmonic analysis Dynamical systems Linear algebra

Source paper: Quantitative stability of the intersection body operator near the ball, and the dynamical origin of the two--dimensional degeneracy arXiv:2607.09412 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Insert a differentiable intersection-body-inspired map on positive spherical feature fields. The map contracts high-order angular variation while leaving degree-two ellipsoidal structure neutral, providing a principled alternative to generic smoothing that does not erase global anisotropy.

Formulas

$$\rho_{\mathcal{I}K}(u)=\frac{\left\lvert S^{n-2}\right\rvert}{n-1}\,\mathcal{R}\!\left(\rho_K^{\,n-1}\right)(u),\qquad (\mathcal{R}g)(u)=\operatorname{avg}_{v\in S^{n-1}\cap u^\perp}g(v).$$

$$\Phi(f)(u)=\frac{c_n}{\kappa_{n-1}}\,\mathcal{R}\!\left(f^{\,n-1}\right)(u),\qquad c_n=\frac{|S^{n-2}|}{n-1},\qquad \Phi(1)=1.$$

$$\operatorname{gap}(n)=\frac{(n-2)(n+4)}{(n+1)^2},\qquad C(n)=\frac{(n+1)^2}{(n-2)(n+4)}.$$

$$\left\|P_{\ge 4}D(\Phi^2)_1[h]\right\|_{L^2(S^{n-1})}\leq \left(1-\operatorname{gap}(n)\right)\|P_{\ge 4}h\|_{L^2(S^{n-1})},\qquad P_2D(\Phi^2)_1[h]=P_2h.$$

Mathematical statement

For a star body with positive radial function f(u)=rho_K(u), where u belongs to the sphere S^(n-1), the intersection-body map is T(f)(u)=c_n R(f(u)^(n-1))(u). Here c_n=|S^(n-2)|/(n-1), |S^(n-2)| is the surface area of the unit (n-2)-sphere, and R is the spherical Funk transform: Rg(u) is the average of g over the great subsphere S^(n-1) intersected with u-perpendicular. Since T(1)=kappa_(n-1), with kappa_(n-1) the volume of the unit (n-1)-ball, define the normalized map Phi(f)=T(f)/kappa_(n-1), so Phi(1)=1. The paper states that the linearization of Phi squared at the constant function 1 has multiplier exactly 1 on degree-two spherical harmonics, which encode ellipsoidal deformations, while all higher harmonics are contracted. The sharp contraction gap on the higher-mode subspace is gap(n)=(n-2)(n+4)/(n+1)^2. Thus the linear contraction factor is 1-gap(n), and the inverse stability scale is C(n)=(n+1)^2/((n-2)(n+4)). In the neural-network adaptation, P_2 denotes projection onto degree-two harmonics, P_ge4 denotes projection onto harmonics of degree at least four, and P_0 denotes the constant-mode projection.

Implementation notes

(1) Integration point: use this after a spherical or directional encoder produces a positive feature field X with shape [batch, channels, M], sampled at M points u_j on the sphere. Apply it channelwise as a training-time feature-normalization block before the prediction head, or use it only as a regularization branch. Enforce positivity with f=softplus(X)+epsilon. (2) Pseudocode: precompute a matrix A of shape [M,M] approximating the Funk transform, where A[i,j] averages samples near the great circle perpendicular to u_i; row-normalize A. Set c_n=surface_area(S^(n-2))/(n-1) and b=volume_of_unit_ball(n-1). Compute z=(c_n/b)*A*(f**(n-1)), then y=(c_n/b)*A*(z**(n-1)). Return f+alpha*(y-f), with alpha in [0,1], or return y directly. (3) To implement the spectral diagnostic, construct a real spherical-harmonic matrix H on the sampled points. Form P_2=H_2 times the pseudoinverse of H_2 and P_0 from the constant basis vector; approximate P_ge4 by I-P_0-P_2. The paper-derived quantities are the pointwise exponent n-1, Funk averaging, normalization, and predicted contraction 1-gap(n). Estimate A, the projectors, alpha, and the regularization weight empirically. Add lambda times the squared norm of P_ge4 times (y-f), while avoiding penalties on P_2 so ellipsoidal variation is preserved. (4) First experiment: train a small spherical CNN or spherical transformer on ModelNet40 with an icosphere resolution of two. Compare the proposed block against no smoothing and ordinary Gaussian or spectral smoothing at equal parameter count and FLOPs. For n=3, the predicted gap is 7/16, so measure the ratio of high-mode energy after the block to before it; it should be near or below 9/16 in the small-perturbation regime. Success means faster validation-loss reduction, lower accuracy degradation under high-frequency angular perturbations, and preserved accuracy under deliberate ellipsoidal deformations. Failure is clean-accuracy loss or a measured high-mode ratio substantially above the predicted contraction factor.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Significantly worse ✗

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

Stage 1 — Mechanism check agent confidence 8/10

Built a NumPy/SciPy MVP of the normalized n=3 intersection-body map using Fibonacci-sphere quadrature and a local Funk-transform approximation. The numerical linearization reproduced the claimed phenomenon: degree-2 multiplier 0.992 (near neutral), while degree-4 and degree-6 multipliers were 0.547 and 0.371, below the predicted high-mode factor 0.5625; the mixed high-mode ratio was 0.550. The toy classification was saturated at 1.0 accuracy for both baseline and idea, so no downstream accuracy win was demonstrated.

Agent confidence: 8/10

Baseline: Raw spherical feature classifier accuracy: 1.000
Idea: Stabilized feature classifier accuracy: 1.000; degree-2 ratio 0.992, degree-4 ratio 0.547, degree-6 ratio 0.371, mixed high-mode ratio 0.550 versus predicted 0.5625

Limitations: This is a small S^2 NumPy diagnostic, not a spherical CNN or ModelNet40 experiment. The Funk transform is approximated by a finite local-strip neighbor matrix, so the exact degree-2 identity is only approximate. The toy classification task was too easy to measure an accuracy improvement, training-speed gain, FLOP comparison, robustness benefit, or parameter-matched Gaussian-smoothing baseline.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.09412) Sep 4, 2026

Significantly WORSE than the tuned baseline (+7.7%, p=0.0169, wins 1 / 8 paired seeds).

Benchmark: Vision (image classification)
Model: cnn_small
Paired seeds: 8
Baseline mean: 0.7042
Idea mean: 0.7583
Effect (Δ): 0.0542 (+7.7%; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.0169 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±5.9%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 5 configs

Limitations:

The built-in vision benchmark is only a structural proxy for spherical directional fields; no ModelNet40 or true spherical CNN was tested. The spatial Fourier signature is a proxy rather than an exact spherical-harmonic decomposition, and the small 12-epoch CIFAR subset limits statistical power beyond the required 8 paired seeds.

How to run: python3 bench_experiment.py

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

Stage 2 — Benchmark Sep 4, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented the Phi^2-inspired differentiable feature stabilizer in a matched small CNN and ran the full 8-seed vision benchmark with baseline and idea learning-rate sweeps. The baseline error was 0.70417 versus 0.75833 for the best idea setting; paired delta was +0.05417 with permutation p=0.0169, meaning the idea was significantly worse. The trained-model mechanism signature was also not confirmed: observed high-frequency and low-frequency ratios were 7.594 and 16.193 versus predicted 0.5625 and 1.0.

Baseline: vision cnn_small error mean=0.7041667; best lr=0.002; per-seed=[0.7166667,0.7000000,0.6416667,0.6833334,0.7166667,0.7250000,0.7916667,0.6583334]
Comparison: map[delta_mean:0.0541667 idea_wins:1 n_pairs:8 p_value:0.0169 verdict:idea worse (significant)]
Idea: vision stabilized cnn_small error mean=0.7583334; best lr=0.001 alpha=0.75; per-seed=[0.7500001,0.7416667,0.7500001,0.7083334,0.7666667,0.7750000,0.7750000,0.8000001]
Mechanism signature: map[confirmed:false observed_degree2_ratio:16.1934479 observed_high_ratio:7.5937355 predicted_degree2_ratio:1 predicted_high_ratio:0.5625]

Benchmark result