✗ Failed on benchmark 2026

Entropy-stable split quadratic layer

Implementation & benchmark of arXiv:2608.07355 — On the Role of Split Formulations on Aliasing Errors and Entropy Stability of Discontinuous Galerkin Schemes

Usefulness7/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Speedup Ssm World model Training dynamics Numerical analysis Dynamical systems Functional analysis

Source paper: On the Role of Split Formulations on Aliasing Errors and Entropy Stability of Discontinuous Galerkin Schemes arXiv:2608.07355 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace the naive pseudospectral evaluation of a quadratic neural-operator nonlinearity with a two-point split-form product. Use the entropy-stable (alpha, beta) = (1/3, 2/3) split as the default, or learn alpha under the consistency constraint alpha + beta = 1 while monitoring energy growth. The goal is to suppress weakly underresolved aliasing and prevent long-horizon rollout blow-up without full 2/3-rule zero-padding.

Formulas

$$F(u_i,u_j)=\frac{1}{4}\alpha(u_i^2+u_j^2)+\frac{1}{2}\beta u_i u_j,\qquad \alpha+\beta=1,$$

$$N_i^{\alpha,\beta}(u)=2\sum_j D_{ij}F(u_i,u_j),$$

$$\Delta_{ijk}(\alpha,\beta)=\frac{d_{\Pi}}{2}\frac{\binom{2i}{i}\binom{2j}{j}\binom{2k}{k}}{\binom{2\tau+2}{\tau+1}}\left(\beta(i+j)-\alpha(\tau+2-(i+j))\right),\quad i+j+k=\tau+2,$$

$$\alpha_{\mathrm{ES}}=\frac{1}{3},\qquad \beta_{\mathrm{ES}}=\frac{2}{3},\qquad F_{\mathrm{ES}}(u_i,u_j)=\frac{u_i^2+u_j^2}{12}+\frac{u_i u_j}{3}.$$

Mathematical statement

The paper represents the quadratic two-point flux as F(u_i,u_j) = (alpha/4)(u_i^2 + u_j^2) + (beta/2)u_i u_j, with splitting coefficients alpha and beta satisfying alpha + beta = 1 so that F(u,u) = u^2/2. For a derivative matrix D with the skew/summation-by-parts structure used by the collocation scheme, the split nonlinear derivative is N_i = 2 sum_j D_ij F(u_i,u_j). Theorem 1 states that, for a valid collocation projection Pi of quadrature strength tau and primary aliasing coefficient d_Pi, the leading aliasing coefficient for modal indices 0 <= i,j,k <= p with i+j+k = tau+2 is Delta_ijk(alpha,beta) = (d_Pi/2) [binom(2i,i) binom(2j,j) binom(2k,k) / binom(2tau+2,tau+1)] [beta(i+j) - alpha(tau+2-(i+j))]. Here binom(a,b) is the binomial coefficient, tau describes the quadrature strength, and d_Pi is the projection's primary aliasing coefficient. The result identifies how the split controls the sign and magnitude of dominant unresolved interactions. The entropy-stable (1/3,2/3) split yields a skew-symmetric aliasing pattern needed for long-time stability. In a neural spectral layer, u is the spatial field, D is a fixed spectral derivative operator, and F and N are applied independently over batch and channel or feature groups.

Implementation notes

Integrate this at the quadratic nonlinear block of a Fourier neural operator, neural PDE timestepper, or learned state-space/world model that currently computes a product by inverse FFT, pointwise multiplication, and forward FFT. Let z have shape [batch, channels, grid], and let D be the periodic spectral derivative, implemented for a small prototype as a dense matrix or as FFT multiplication by ik. First create a learned projection z = W x, optionally grouping channels so the quadratic operation acts within each group. Implement the reference operation directly: F_ij = alpha/4*(z_i2 + z_j2) + beta/2*z_i*z_j; N_i = 2*sum_j(D_ij*F_ij). Then insert y = linear(x) + gamma*N, with gamma initialized to 0.1. For the FFT-efficient implementation, use the algebraically equivalent split N = alpha*D(z**2/2) + beta*z*D(z), because 2 sum_j D_ij F(z_i,z_j) reduces to this expression when D has zero row sums and the diagonal/self terms are handled consistently. Use alpha=1/3 and beta=2/3 for the entropy-stable variant. Compare conservative (1,0), skew/entropy-stable (1/3,2/3), symmetric (1/2,1/2), and a zero-padded 2/3-dealiasing baseline. The formulas provide the split flux and predicted leading aliasing coefficient; estimate tau from the grid/quadrature order and estimate d_Pi by measuring the dominant projection error on modal test functions, but do not require these estimates for training. Start with a 1D Burgers or Kuramoto-Sivashinsky dataset on 64 grid points using a small FNO or residual spectral model. Equalize parameters, FFT count, and training steps. Track one-step MSE, rollout MSE, energy E_t = mean(z_t^2)/2, maximum energy growth, and time-to-divergence. The hypothesis is that the entropy-stable split substantially reduces unstable long rollouts relative to the conservative split, while approaching padded-dealiasing accuracy with lower memory and compute.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: No measurable effect

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 and numerically tested the split quadratic spectral derivative with alpha+beta=1, including the entropy-stable (1/3,2/3) choice, padded reference, dense two-point flux, and a toy Burgers rollout. The affine aliasing prediction matched to 8.9e-16, the predicted cancellation roots matched across the index sweep to approximately 1e-16, and the dense flux/FFT identity was checked numerically. Relative to the conservative baseline alpha=1, the entropy-stable split reduced mean absolute energy injection from 0.1317 to 0.1317 (equal here because alpha=1 and 1/3 are symmetric in this metric), reduced padded-reference RMSE from 1.6446 to 1.8598 (worse), but reduced 250-step maximum energy growth from 1.6785x to 1.3644x; thus the mechanism manifested, although it did not dominate all accuracy metrics and the symmetric split was best in this toy.

Agent confidence: 8/10

Baseline: Conservative alpha=1,beta=0: formula slope/intercept agree with theory; mean |<u,N>|=0.1317449; padded-reference RMSE=1.6446275; 250-step max energy ratio=1.6785156; divergence step=250 (no threshold divergence).
Idea: Entropy-stable alpha=1/3,beta=2/3: aliasing cancellation root for (i,j,k)=(1,1,4) observed at 0.3333333 versus predicted 0.3333333; max affine residual=8.88e-16; mean |<u,N>|=0.1317449; padded-reference RMSE=1.8597605; 250-step max energy ratio=1.3643754; divergence step=250. Other cancellation roots also had residuals about 1e-16.

Limitations: This is an analytic/tiny spectral toy, not a trained FNO or PDE dataset experiment. It does not test learned alpha, parameter/FLOP-equalized training, noisy long-horizon supervised rollouts, multiple resolutions, or a rigorous entropy bound; the padded-reference implementation and energy metric are only prototype diagnostics. The rollout compares several fixed split forms and did not produce threshold divergence within 250 steps.

How to run: python3 split_quadratic_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.07355) Sep 2, 2026

No significant difference vs. the tuned baseline (+0.1%, p=0.2846, wins 4 / 8 paired seeds).

Benchmark: Burgers_periodic_split
Model: custom_split_quadratic_net
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0 (+0.1%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.2846 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

Only the custom 32-point periodic Burgers one-step regression was tested. No long-horizon rollout metric, learned alpha, multiple resolutions, or zero-padded trained comparison was run; the generic harness did not expose a trained-model aliasing diagnostic.

How to run: python3 split_bench.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 periodic Burgers PDE track because the built-in dynamics track lacks periodic spatial quadratic-product structure. Across eight paired seeds, alpha=1/3 had mean test MSE 4.3688922346518666e-05 versus 4.365622254454138e-05 for the tuned naive alpha=1 baseline, giving delta_mean +3.269980197728728e-08 and permutation p=0.28455. This is no measurable effect, not a significant win.

Baseline: alpha=1, lr=0.01; mean MSE 4.365622254454138e-05; std 4.107952621194168e-05
Idea: alpha=1/3, lr=0.01; mean MSE 4.3688922346518666e-05; std 4.108227626022683e-05
Paired delta: 3.269980197728728e-08, idea minus baseline
Paired wins: 4/8
Permutation p: 0.28455

Benchmark result