✗ Mechanism failed 2026

Wick-Matching Polynomial Interaction Layer

Implementation & benchmark of arXiv:2608.25651 — Stable Symmetric Series, Differential Operators, and Jack Deformations

Usefulness6/10

Difficulty5/10

Novelty7/10

Architecture Accuracy Sample efficiency Generalization Graph nn Embedding Mlp Regularization Combinatorics Probability Representation theory Algebra

Source paper: Stable Symmetric Series, Differential Operators, and Jack Deformations arXiv:2608.25651 ⓘ · analyzed Aug 29, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained high-order polynomial interaction module with features generated by Gaussian matrix contractions and their exact Wick expansion. The resulting interactions are sums of products of power-sum invariants, with coefficients fixed by perfect-matching counts, providing a low-parameter inductive bias for permutation- or orthogonal-structured data.

Formulas

$$\mathbb{E}\,p_{\lambda}(AZBZ^{t})=\sum_{\delta\in\mathcal{M}_{2n}}p_{\Lambda(\delta,\delta_{\lambda})}(A)p_{\Lambda(\delta,\epsilon)}(B).$$

$$\mathbb{E}\,p_{\lambda}(AZBZ^{t})=\sum_{\mu,\nu\vdash n}a_{\mu\nu}^{\lambda}(2)p_{\mu}(A)p_{\nu}(B),\qquad a_{\mu\nu}^{\lambda}(2)=\#\{\delta:\Lambda(\delta,\delta_{\lambda})=\mu,\ \Lambda(\delta,\epsilon)=\nu\}.$$

$$\widetilde F_{\lambda}(X)=\sum_{\mu,\nu\vdash n}a_{\mu\nu}^{\lambda}(2)\,p_{\mu}(A(X))p_{\nu}(B(X)),\qquad p_{\mu}(A)=\prod_{r=1}^{\ell(\mu)}\operatorname{tr}(A^{\mu_r}).$$

Mathematical statement

Theorem 9.1 states that for a partition lambda of n, p_lambda(M)=product over r of p_{lambda_r}(M), p_k(M)=tr(M^k), and a random matrix Z with independent standard Gaussian entries, E[p_lambda(A Z B Z^t)] = sum over delta in M_{2n} of p_{Lambda(delta,delta_lambda)}(A) p_{Lambda(delta,epsilon)}(B). Here epsilon is a fixed perfect matching on {1,bar{1},...,n,bar{n}}, delta_lambda is the matching encoding the cycle structure of partition lambda, M_{2n} is the set of perfect matchings, and Lambda(delta_1,delta_2) is the partition of connected-component sizes in the union graph delta_1 union delta_2. Grouping matchings by mu=Lambda(delta,delta_lambda) and nu=Lambda(delta,epsilon) gives E[p_lambda(A Z B Z^t)] = sum over mu,nu of a_{mu nu}^lambda(2) p_mu(A) p_nu(B), where a_{mu nu}^lambda(2) is the number of matchings with those two component types. The transferable property is exact closure into a finite basis of trace-power products; the coefficients are combinatorial integers and require no learned fitting.

Implementation notes

Integrate the layer after a token, node, or set encoder and before the task MLP or graph readout. Given X in R^{N times d}, produce two small square matrices A(X), B(X) in R^{q times q}; for stability, use A=U U^T plus epsilon I and B=V V^T plus epsilon I. Choose degree n in {2,3,4}. Enumerate all partitions mu, nu, and lambda of n, then precompute the integer table a[lambda,mu,nu] by enumerating perfect matchings of {1,bar{1},...,n,bar{n}}. In each forward pass compute tracesA[k]=trace(A^k) and tracesB[k]=trace(B^k), construct pA[mu]=product of tracesA[mu_r] and pB[nu]=product of tracesB[nu_r], and form F[lambda]=sum over mu,nu of a[lambda,mu,nu] times pA[mu] times pB[nu]. Normalize F with a running mean and variance, concatenate all F features, and pass them to the prediction head. Compact pseudocode is: A=U@U.T+eps*I; B=V@V.T+eps*I; tracesA[k]=trace(matrix_power(A,k)); tracesB[k]=trace(matrix_power(B,k)); pA[mu]=product(tracesA[part] for part in mu); pB[nu]=product(tracesB[part] for part in nu); F[lambda]=sum(a[lambda,mu,nu]*pA[mu]*pB[nu] for mu,nu); output=head(normalize(F)). The matching coefficients are computed exactly from the paper's combinatorics, while A, B, and their traces are learned quantities. First test this on synthetic invariant polynomial regression and QM9 with a small permutation-invariant set or graph encoder, q=8, and n=3. Compare against DeepSets and the same encoder with an unconstrained degree-3 bilinear head at equal parameter count. Measure test error, performance with 10 percent of the training data, permutation and orthogonal-transformation invariance, and gradient variance. Success means lower test error or better low-data accuracy at comparable FLOPs, with exact features matching Monte Carlo estimates of p_lambda(A Z B Z^t).

Verification

Mechanism failed

Stage 1 · Toy mechanism gate: Failed ✗

Stage 2 · Mechanism transferred to benchmark: Not tested

Stage 2 · Practical benchmark result: Not run

Methodology: Toy-system gate first; the benchmark stage runs only after a pass. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built an exact degree-3 Wick-matching layer with enumerated perfect-match coefficients, PSD matrix construction, and trace-power features. The core identity matched 30,000-sample Gaussian Monte Carlo within 0.29%, confirming the combinatorial expansion. Wick regression had lower RMSE (6.12 vs 8.10), but used substantially more parameters (4,623 vs 1,761), so this is not evidence of a win at comparable parameter count.

Agent confidence: 8/10

Baseline: DeepSets: test RMSE 8.1012, 1,761 parameters, permutation output difference 9.54e-7
Idea: Wick layer: test RMSE 6.1223, 4,623 parameters, permutation output difference 6.59e-5; exact-vs-Monte-Carlo maximum relative error 0.00289

Limitations: Only a small synthetic degree-3 regression task was tested; no QM9 or graph benchmark, FLOP-matched comparison, unconstrained bilinear control, repeated seeds, gradient-variance analysis, orthogonal-transformation test, or running-statistics normalization was evaluated. The regression improvement is confounded by substantially higher parameter count.

How to run: python3 wick_experiment.py

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

Artifacts

Implementation overview ⬇ Download all as ZIP 4 files · code, reports and structured results

📄 report.md 1.3 KB View
📊 results.json 0.7 KB View Raw
📊 results_equal_params.json 0.7 KB View Raw
🐍 wick_experiment.py 6.2 KB View Raw