# Normal-Cone Certified Priority Weighting

- ID: 2189
- Canonical URL: https://synthcore.org/idea/2189/normal-cone-certified-priority-weighting
- API JSON: https://synthcore.org/api/idea/2189.json
- API Markdown: https://synthcore.org/api/idea/2189.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.12520](https://arxiv.org/abs/2608.12520)
- Category: optimization
- Solves: accuracy, stability, sample-efficiency
- ML areas: loss, optimizer, training
- Math tags: convex-analysis, optimization, geometry
- Ratings: usefulness 7/10; difficulty 6/10; novelty 6/10

## Idea description

Replace hand-tuned exponentially separated coefficients for multiple neural objectives with weights obtained from a local KKT certificate. For L1 hinge penalties, solve a small linear program that maximizes the smallest tier weight while enforcing approximate stationarity of the weighted objective at the current priority solution. This should preserve high-priority behavior more reliably than fixed loss weights while avoiding unnecessarily large coefficients.

## Mathematical statement

Let z be the model output or decision vector, X be a convex feasible set, V_i(z)=sum_j[g_{i,j}(z)]_+ be the L1 hinge violation for priority tier i, and J(z) be the performance loss. The lexicographic solution z^* first minimizes V_1, then V_2, through V_L, and finally J. Define the upper achievement image A={(v_1,...,v_L,j): there exists z in X with v_i>=V_i(z) and j>=J(z)}. The geometric criterion is that a scalar weighted objective exactly supports the lexicographic point when (w_1,...,w_L,1) belongs to the outward normal cone N_A(V_1(z^*),...,V_L(z^*),J(z^*)). For convex differentiable pieces, a KKT certificate requires 0=gradient J(z^*)+sum_i w_i s_i+G_c(z^*)^T lambda, where s_i belongs to the subdifferential of V_i, G_c is the Jacobian of active constraints c(z)<=0, and lambda>=0. For each hinge term, s_{i,j}=0 when g_{i,j}<0, s_{i,j}=gradient g_{i,j} when g_{i,j}>0, and s_{i,j}=alpha gradient g_{i,j}, alpha in [0,1], at g_{i,j}=0. Use a slackened LP to maximize a positive weight margin delta while bounding the stationarity residual by epsilon.

## Key formulas

- $$V_i(z)=\sum_j[g_{i,j}(z)]_+,\qquad [a]_+=\max(a,0),\qquad \min_{z\in X}\;J(z)+\sum_{i=1}^{L}w_iV_i(z).$$
- $$\mathcal{A}=\{(v_1,\ldots,v_L,j):\exists z\in X,\ v_i\ge V_i(z),\ j\ge J(z)\},\qquad (w_1,\ldots,w_L,1)\in N_{\mathcal{A}}(V(z^*),J(z^*)).$$
- $$0=\nabla J(z^*)+\sum_{i=1}^{L}w_i s_i+G_c(z^*)^T\lambda,\qquad s_i\in\partial V_i(z^*),\quad \lambda\ge0.$$$
- $$\max_{w,\lambda,\delta}\ \delta\quad\text{s.t.}\quad w_i\ge\delta,\ 0\le w_i\le w_{\max},\ \left\|\nabla J+\sum_iw_is_i+G_c^T\lambda\right\|_\infty\le\epsilon,\ \lambda\ge0.$$$

## Implementation notes

First integrate this as a loss-weight controller for a neural network with ordered objectives, such as safety, legality, reconstruction, calibration, or auxiliary-task penalties. The network produces predictions y; each tier computes differentiable constraint scores g_{i,j}(y), and the training loss is J(y)+sum_i w_i V_i(y). Begin with a convex last layer or frozen network features so the certificate assumptions are testable, then evaluate the local heuristic in the full network. Every K minibatches, compute the hinge masks, per-tier gradients s_i=gradient(V_i, parameters), and q=gradient(J, parameters). If there are explicit output constraints c(z)<=0, compute their Jacobian G_c and include nonnegative multipliers lambda. Solve the LP maximizing delta subject to w_i>=delta, 0<=w_i<=w_max, lambda>=0, and infinity-norm residual ||q+sum_i w_i s_i+G_c^T lambda||<=epsilon. If delta exceeds delta_min, replace the current weights by the LP solution; otherwise retain the previous weights or invoke sequential tier updates. Compute subgradients and the stationarity residual from the paper's mathematics. Estimate epsilon, w_max, delta_min, and the refresh interval on a validation split. Start with a small MLP on synthetic regression with three ordered constraints, comparing equal weights, fixed 100:10:1 weights, GradNorm, and the certificate controller. Measure tier violations at matched J, frequency of high-tier regressions, weight oscillation, and loss descent at equal FLOPs. Success means lower high-tier violations at equal performance and a positive certificate margin on most batches.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a readable LP-based normal-cone certificate controller and deterministic convex priority-training toy in experiment.py. All three quantitative mechanism predictions were confirmed: inverse-lambda scaling had maximum relative error 1.0e-9, the epsilon=1 feasibility transition was correct at every tested point, and slack-margin scaling had zero absolute error. In the conflicting training toy, the certificate matched equal weighting rather than improving it; fixed 100:10:1 achieved zero violation but at much higher performance loss, so no practical win was demonstrated.

### Mechanism check

- Verdict: Built a readable LP-based normal-cone certificate controller and deterministic convex priority-training toy in experiment.py. All three quantitative mechanism predictions were confirmed: inverse-lambda scaling had maximum relative error 1.0e-9, the epsilon=1 feasibility transition was correct at every tested point, and slack-margin scaling had zero absolute error. In the conflicting training toy, the certificate matched equal weighting rather than improving it; fixed 100:10:1 achieved zero violation but at much higher performance loss, so no practical win was demonstrated.
- Confidence: 7/10
- Limitations: This is a tiny hand-designed convex last-layer problem, not a neural network or minibatch experiment. No GradNorm comparison, explicit constraint multipliers, validation hyperparameter tuning, FLOP/speed measurement, stochastic training, or larger multi-dimensional hinge geometry was tested. The positive result supports the LP mechanism, not a broad empirical superiority claim.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.2116
- Paired wins: 5/0
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/874/bench_report.json)
- [experiment.py](https://synthcore.org/code/874/experiment.py)
- [report.md](https://synthcore.org/code/874/report.md)
- [report_bench_2026-09-01T131527.md](https://synthcore.org/code/874/report_bench_2026-09-01T131527.md)
- [results.json](https://synthcore.org/code/874/results.json)
- [stage2_certified_priority.py](https://synthcore.org/code/874/stage2_certified_priority.py)
- [Download all files as ZIP](https://synthcore.org/download/874)

## Disclaimer

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