# GQL Safe Residual Layer

- ID: 294
- Canonical URL: https://synthcore.org/idea/294/gql-safe-residual-layer
- API JSON: https://synthcore.org/api/idea/294.json
- API Markdown: https://synthcore.org/api/idea/294.md
- Verification status: failed_benchmark
- Source: [arXiv:2606.31992](https://arxiv.org/abs/2606.31992)
- Category: architecture
- Solves: stability, accuracy, generalization
- ML areas: world-model, loss, training-dynamics
- Math tags: geometry, convex-analysis, linear-algebra, optimization, numerical-analysis
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Insert a scalar flux-correction-style limiter after a neural operator predicts a conservative state or residual. Interpolate between a known-admissible baseline state and the learned high-order candidate, choosing the largest coefficient that satisfies a geometric family of linear inequalities encoding positive density, positive pressure, and subluminal velocity. This retains as much of the neural prediction as possible instead of independently clipping physical variables.

## Mathematical statement

The paper evolves conservative variables using a low-order safe update and a limited antidiffusive update, as in the extracted formulas $\mathbf U^L_{ij}=\mathbf U^n_{ij}-\frac{\Delta t}{\Delta x}(\widehat{\mathcal F}^{L}_{i+1/2,j}-\widehat{\mathcal F}^{L}_{i-1/2,j})-\frac{\Delta t}{\Delta y}(\widehat{\mathcal G}^{L}_{i,j+1/2}-\widehat{\mathcal G}^{L}_{i,j-1/2})$ and $\mathbf U^{n+1}_{ij}=\mathbf U^L_{ij}-\frac{\Delta t}{\Delta x}(\theta_{i+1/2,j}\widehat{\mathcal F}^{A}_{i+1/2,j}-\theta_{i-1/2,j}\widehat{\mathcal F}^{A}_{i-1/2,j})-\frac{\Delta t}{\Delta y}(\theta_{i,j+1/2}\widehat{\mathcal G}^{A}_{i,j+1/2}-\theta_{i,j-1/2}\widehat{\mathcal G}^{A}_{i,j-1/2})$. Here $\mathbf U^L$ is the safe low-order state, $\widehat{\mathcal F}^{A},\widehat{\mathcal G}^{A}$ are antidiffusive fluxes, and $0\leq\theta\leq1$ limits them. The paper also represents coefficients as sign-index dot products, $c_\star=\mathbf s_0\cdot\mathbf V_{c_\star}=\sum_{i=1}^{4}s_i(\mathbf V_{c_\star})_i$, compactly enumerating sign combinations. For the neural adaptation, let $\mathbf U_0$ be a valid baseline and $\mathbf U_1$ the network output, and use $\mathbf U(\theta)=\mathbf U_0+\theta(\mathbf U_1-\mathbf U_0)$. The relativistic admissibility condition is represented by the GQL family $g_{\mathbf s}(\mathbf U)=E-s_DD-\mathbf s_m^\top\mathbf m\geq\varepsilon$ for every auxiliary vector $\mathbf s=(s_D,\mathbf s_m)$ with $\|\mathbf s\|_2\leq1$, together with $D\geq\varepsilon$. Since $E-\sqrt{D^2+\|\mathbf m\|_2^2}=\min_{\|\mathbf s\|\leq1}g_{\mathbf s}(\mathbf U)$, these inequalities enforce the relativistic admissibility condition. For each normal with a negative denominator, the maximal safe interpolation coefficient is $\theta_{\mathbf s}=\frac{g_{\mathbf s}(\mathbf U_0)-\varepsilon}{g_{\mathbf s}(\mathbf U_0)-g_{\mathbf s}(\mathbf U_1)}$; take the minimum over normals. In the basic relativistic case, the worst normal is $\mathbf s^\star=(D_1-D_0,\mathbf m_1-\mathbf m_0)/\|(D_1-D_0,\mathbf m_1-\mathbf m_0)\|$. For more general GQL constraints, the paper’s stereographic parameterization transforms the worst-case minimization into a generalized Rayleigh quotient and obtains the optimum from a small symmetric eigenvalue problem.

## Key formulas

- $$\mathbf U^L_{ij}=\mathbf U^n_{ij}-\frac{\Delta t}{\Delta x}\left(\widehat{\mathcal F}_{i+\frac12,j}^{L}-\widehat{\mathcal F}_{i-\frac12,j}^{L}\right)-\frac{\Delta t}{\Delta y}\left(\widehat{\mathcal G}_{i,j+\frac12}^{L}-\widehat{\mathcal G}_{i,j-\frac12}^{L}\right),$$
- $$\mathbf U^{n+1}_{ij}=\mathbf U^L_{ij}-\frac{\Delta t}{\Delta x}\left(\theta_{i+\frac12,j}\widehat{\mathcal F}_{i+\frac12,j}^{A}-\theta_{i-\frac12,j}\widehat{\mathcal F}_{i-\frac12,j}^{A}\right)-\frac{\Delta t}{\Delta y}\left(\theta_{i,j+\frac12}\widehat{\mathcal G}_{i,j+\frac12}^{A}-\theta_{i,j-rac12}\widehat{\mathcal G}_{i,j-rac12}^{A}\right).$$
- $$g_{\mathbf s}(\mathbf U)=E-s_DD-\mathbf s_m^{\top}\mathbf m,\qquad E-\sqrt{D^2+\|\mathbf m\|_2^2}=\min_{\|\mathbf s\|_2\leq1}g_{\mathbf s}(\mathbf U).$$
- $$\theta=\min\left\{1,\ \min_{\mathbf s:\,g_{\mathbf s}(\mathbf U_1)<\varepsilon}\frac{g_{\mathbf s}(\mathbf U_0)-\varepsilon}{g_{\mathbf s}(\mathbf U_0)-g_{\mathbf s}(\mathbf U_1)}\right\},\qquad \mathbf U_{\mathrm{safe}}=\mathbf U_0+\theta(\mathbf U_1-\mathbf U_0).$$

## Implementation notes

(1) Integration point: apply this layer to the final conservative-state tensor produced by a neural operator, learned finite-volume flux network, or neural ODE time-stepper. Let each predicted state be $\mathbf U_1=(D,\mathbf m,E)$ and obtain $\mathbf U_0$ from the previous valid state, a first-order numerical update, or a positivity-preserving primitive-variable fallback. Do not independently clamp $D$, $E$, and momentum; scale the complete residual toward $\mathbf U_0$. (2) Pseudocode: `delta = U1-U0; theta=ones_like(D); q=norm(concat(D1-D0,m1-m0),dim=channel); s=concat(D1-D0,m1-m0)/(q+1e-12); g0=E0-s_D*D0-dot(s_m,m0); g1=E1-s_D*D1-dot(s_m,m1); bad=(g1 < eps); theta[bad]=clamp((g0[bad]-eps)/(g0[bad]-g1[bad]+1e-12),0,1); density_bad=(D1 < eps); theta[density_bad]=min(theta[density_bad],clamp((D0[density_bad]-eps)/(D0[density_bad]-D1[density_bad]+1e-12),0,1)); Usafe=U0+theta[...,None]*delta`. Recompute primitive variables from `Usafe` and reject the sample if the EOS inversion fails. (3) The GQL identity and interpolation bound are computed analytically. The normal $s^\star$ is the exact minimizer for the Euclidean relativistic constraint. For an extended EOS or additional learned constraints, estimate the worst normal using a small sampled set on the unit sphere, or implement the paper’s stereographic parameterization and solve its $(d+1)\times(d+1)$ symmetric eigenproblem. Use detached $\theta$ initially for robust gradients, then compare straight-through and differentiable versions. (4) First experiment: train a small Fourier neural operator or ConvNet to advance 1D relativistic shock-tube states on a generated dataset, comparing an unconstrained output, primitive-variable parameterization, componentwise clipping, and this residual layer. Track invalid-state rate, rollout horizon before NaNs, shock error, and one-step and long-rollout accuracy. Success is zero invalid states at comparable error, fewer fallback activations than clipping, and materially longer stable rollouts without reducing one-step accuracy.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a NumPy GQL-style residual safety layer for U=(D,m,E), with an exact GQL identity check and robust scalar bisection to find the admissible interpolation boundary. Across two fixed seeds, the limiter produced 100% admissible states versus about 52–53% for the componentwise clipping control, while retaining about 0.67 of the learned residual; the identity error was below 2e-15. The claimed safety phenomenon is real in this toy setting, but the proposed delta-derived worst-normal shortcut was not safe here, so the MVP uses convex-margin bisection instead.

### Mechanism check

- Verdict: Built a NumPy GQL-style residual safety layer for U=(D,m,E), with an exact GQL identity check and robust scalar bisection to find the admissible interpolation boundary. Across two fixed seeds, the limiter produced 100% admissible states versus about 52–53% for the componentwise clipping control, while retaining about 0.67 of the learned residual; the identity error was below 2e-15. The claimed safety phenomenon is real in this toy setting, but the proposed delta-derived worst-normal shortcut was not safe here, so the MVP uses convex-margin bisection instead.
- Confidence: 8/10
- Limitations: No neural operator, training loop, PDE shock-tube rollout, GPU implementation, EOS pressure inversion, runtime comparison, or multidimensional flux update was tested. The implementation uses numerical bisection rather than the paper's generalized eigenvalue/stereographic solver, and the clipping control only clips density and energy independently.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.92165
- Paired wins: 5/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/1256/bench_report.json)
- [gql_experiment.py](https://synthcore.org/code/1256/gql_experiment.py)
- [gql_results.json](https://synthcore.org/code/1256/gql_results.json)
- [gql_track.py](https://synthcore.org/code/1256/gql_track.py)
- [report.md](https://synthcore.org/code/1256/report.md)
- [report_bench_2026-09-03T225149.md](https://synthcore.org/code/1256/report_bench_2026-09-03T225149.md)
- [stage2_bench.py](https://synthcore.org/code/1256/stage2_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/1256)

## Disclaimer

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