△ Mechanism confirmed, baseline not beaten 2026

Multiplicity-balanced symmetric interaction layer

Implementation & benchmark of arXiv:2608.13753 — Orbit compression and asymptotic contractivity for symmetric Bohnenblust--Hille inequalities

Usefulness7/10

Difficulty4/10

Novelty7/10

Architecture Memory Stability Accuracy Mlp Architecture Regularization Combinatorics Functional analysis Inequalities

Source paper: Orbit compression and asymptotic contractivity for symmetric Bohnenblust--Hille inequalities arXiv:2608.13753 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace a dense degree-m tensor interaction layer by a symmetric orbit-parameterized layer with one parameter per exponent vector and explicit multinomial scaling. This preserves the contribution of all ordered tensor entries represented by one orbit, while reducing parameter count and avoiding the amplitude distortion of unweighted monomial compression.

Formulas

$$N_{\alpha}=\frac{m!}{\prod_{j=1}^{d}\alpha_j!},\qquad P(x)=T(x,\ldots,x)=\sum_{|\alpha|=m}N_{\alpha}T_{\alpha}x^{\alpha},\qquad x^{\alpha}=\prod_{j=1}^{d}x_j^{\alpha_j}.$$

$$\|T\|_{\mathrm{orbit},q}^{q}=\sum_{|\alpha|=m}N_{\alpha}|T_{\alpha}|^{q},\qquad q=\frac{2m}{m+1}.$$

$$\frac{\|T\|_{\mathrm{BH}}}{\|P\|_{\infty}}\leq 1+\frac{C}{m}\quad\text{for sufficiently large }m.$$

$$\widetilde T_{\alpha}=\frac{\theta_{\alpha}}{N_{\alpha}^{1/2}},\qquad P_{\theta}(x)=\sum_{|\alpha|=m}N_{\alpha}^{1/2}\theta_{\alpha}x^{\alpha}.$$

Mathematical statement

For a symmetric m-linear form T on C^d, let alpha=(alpha_1,...,alpha_d) in N^d satisfy |alpha|=sum_j alpha_j=m. All ordered tuples with histogram alpha form a permutation orbit of size N_alpha=m!/prod_j alpha_j!. T_alpha is the common ordered coefficient on that orbit, and x^alpha=prod_j x_j^{alpha_j}. The exact diagonal identity is P(x)=T(x,...,x)=sum_{|alpha|=m}N_alpha T_alpha x^alpha. For the Bohnenblust-Hille exponent q=2m/(m+1), the orbit coefficient quantity is sum_alpha N_alpha|T_alpha|^q. The paper proves an asymptotic complex symmetric norm ratio bounded by 1+C/m for sufficiently large m; the implementation uses the exact multiplicities, not an unknown asymptotic constant. Here N_alpha is a combinatorial multiplicity, T_alpha is the shared ordered-tensor value, and c_alpha=N_alpha T_alpha is the polynomial coefficient.

Implementation notes

Integrate this as a replacement for a dense degree-m symmetric tensor contraction in a polynomial MLP, invariant set encoder, or symmetric feature-interaction block. The input is x in R^d, and the trainable object is one parameter theta_alpha for every exponent vector alpha with |alpha|=m. Precompute the compositions alpha and logN_alpha=lgamma(m+1)-sum_j lgamma(alpha_j+1). Use the variance-balanced feature phi_alpha(x)=exp(0.5*logN_alpha)*x^alpha and output y=sum_alpha theta_alpha phi_alpha(x). Pseudocode: enumerate alpha; compute logN; initialize theta; for each minibatch compute each x^alpha by repeated products or dynamic programming; multiply by exp(0.5 logN); sum theta_alpha*phi_alpha; backpropagate. If coefficient regularization is wanted, add lambda*sum_alpha N_alpha*abs(theta_alpha/sqrt(N_alpha))^q, with q=2m/(m+1). The multiplicities and q are exact; the theorem's constant C is not needed. First compare against an ordered tensor layer, an unweighted orbit layer, and a standard MLP at matched parameter count. Use d=32 and m=4 or 6 on synthetic permutation-invariant regression and a DeepSets classification task. Measure parameters, wall-clock time, activation variance, gradient variance, and validation error. Success is a 3-10x parameter reduction with no accuracy loss and more uniform gradients across monomial supports.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Inconclusive

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

Stage 1 — Mechanism check agent confidence 9/10

Built a compact symmetric orbit interaction experiment with composition enumeration, exact multinomial multiplicities, weighted features, BH orbit norms, and regression comparison. The quantitative predictions were confirmed: parameter reduction matched d^m/C(d+m-1,m), diagonal reconstruction and BH orbit sums had numerical error below 7e-10, and weighted feature second moments scaled by exactly N_alpha. The proposed sqrt(N_alpha) factor preserves ordered-tensor contributions but does not equalize raw activation variance; it amplifies high-multiplicity supports, so the 'balanced' label should be interpreted as coefficient/orbit normalization rather than raw feature-variance balancing.

Agent confidence: 9/10

Baseline: Unweighted orbit regression: validation MSE 9.20039 after 500 updates; 330 parameters; d=8,m=4.
Idea: Multiplicity-weighted orbit regression: validation MSE 8.94852 after 500 updates; 330 parameters; ordered reference would use 4096 parameters. Exact identity errors ranged from 0 to 2.33e-12; BH-sum errors ranged from 8.88e-16 to 6.99e-10; observed variance ratios matched N_alpha for tested multiplicities 1, 2, 6, 12, 24, 30, and 180.

Limitations: Only a small synthetic polynomial regression was tested, not a DeepSets task or dense learned ordered-tensor baseline. Regression optimization was not tuned extensively, timings are implementation-dependent, and the asymptotic Bohnenblust–Hille theorem bound was not independently reproduced.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.13753) Sep 2, 2026

No significant difference vs. the tuned baseline (+8.7%, p=0.1557, wins 1 / 8 paired seeds).

Benchmark: Custom_symmetric_polynomial_regression
Model: interaction_layer
Paired seeds: 8
Baseline mean: 167.829
Idea mean: 182.4809
Effect (Δ): 14.6519 (+8.7%; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.1557 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±13.9%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the custom regression track was tested; the built-in tracks were structurally mismatched. The baseline is a fully independent ordered-tuple parameterization and therefore has substantially more parameters than the orbit model; no matched-parameter MLP, larger datasets, degrees beyond m=4, longer training, or broader architecture study was tested. The mechanism signature measures the trained idea model's feature behavior, while the task comparison remains the primary metric.

How to run: python3 bench_run.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

Built and evaluated a custom permutation-invariant degree-4 polynomial regression track because no built-in track had the required symmetric interaction structure. The exact sqrt(multinomial) feature scaling matched observed feature-variance ratios extremely closely, confirming the mechanism signature, but the idea's best test MSE was worse than the tuned dense ordered baseline (182.48 vs 167.83; paired delta +14.65; permutation p=0.1557). Therefore the mechanism works arithmetically, but the promised neural-network task-level win was not observed.

Baseline: Tuned dense ordered interaction layer, lr=0.01, 20 epochs: test MSE 167.82897 +/- 48.00323 across 8 seeds; 4096 ordered parameters.
Idea: Multiplicity-weighted orbit interaction layer, lr=0.01, 20 epochs: test MSE 182.48091 +/- 44.26628 across 8 seeds; 330 orbit parameters.
Paired delta: idea - baseline = +14.65195 MSE; lower is better.
Parameter reduction: 4096 -> 330, approximately 12.41x fewer parameters.
Permutation p value: 0.1557

Benchmark result