# Collision-aware physical-support abstention

- ID: 1030
- Canonical URL: https://synthcore.org/idea/1030/collision-aware-physical-support-abstention
- API JSON: https://synthcore.org/api/idea/1030.json
- API Markdown: https://synthcore.org/api/idea/1030.md
- Verification status: failed_benchmark
- Source: [arXiv:2607.16813](https://arxiv.org/abs/2607.16813)
- Category: architecture
- Solves: accuracy, generalization, stability
- ML areas: embedding, moe-routing
- Math tags: statistics, information-theory, geometry
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Modify a learned sparse encoder so that it distinguishes reliable group activation from unreliable child-ray identity. When test separation information is high but dictionary-orientation information is low, the model should output the active coherent group while abstaining on individual child labels, using a permutation- and sign-invariant support representation.

## Mathematical statement

The paper defines test support information I_S=Tβ_-²s²/σ_+², where T is the number of test observations or replications, β_- is a lower bound on the relevant nonzero coefficient magnitude, s is the within-block separation between physical unit rays, and σ_+ is an upper bound on observation noise scale. It defines dictionary-orientation information I_D=Ns⁶, where N is the number of dictionary-training samples. The important regime is I_S≫1 but I_D≲1: the test can separate coordinates under a known dictionary, but training cannot identify the physical orientation of nearly colliding rays. We transplant the resolution rule: report child-level support only when both I_S≥τ_S and I_D≥τ_D; otherwise report the permutation/sign-invariant parent block. A physical ray is represented as a unit vector modulo sign, and a block prediction is represented as the unordered set of active rays or, when orientation is unresolved, simply the block index.

## Key formulas

- $$I_S=\frac{T\beta_-^{2}s^{2}}{\sigma_+^{2}}$$
- $$I_D=Ns^{6}$$
- $$\text{resolution}=\begin{cases}\text{child support},&I_S\ge\tau_S\ \text{and}\ I_D\ge\tau_D,\\\text{active parent block with child ambiguity},&I_S\ge\tau_S\ \text{and}\ I_D<\tau_D,\\\text{parent inactive/undetermined},&I_S<\tau_S.\end{cases}$$
- $$\operatorname{diam}_{H}(\widehat{\mathcal S},\mathcal S)=O\!\left(s\wedge\frac{1}{\sqrt{N}\,s^{2}}\right)$$

## Implementation notes

Integrate this at the output of a sparse encoder, dictionary-learning layer, or MoE router. Let the encoder produce coefficients z∈R^K and let atoms be grouped into coherent blocks B_g; estimate each block’s collision scale s_g from pairwise atom distances, for example s_g=median_{i,j∈B_g} min(∥d_i-d_j∥,∥d_i+d_j∥), where d_i are unit-normalized atoms. For a minibatch, estimate β_- as a low quantile of selected coefficient magnitudes, estimate σ_+ from held-out residual RMS, use N as the number of samples used to fit the dictionary, and use T as the number of replicated views at inference. Compute I_{S,g}=Tβ_-²s_g²/σ_+² and I_{D,g}=Ns_g⁶. Pseudocode: encode each view; aggregate coefficients over T views; compute block score q_g=max_{i∈B_g}|z_i|; if q_g exceeds the group threshold and I_{S,g}≥τ_S, activate block g; expose individual child indices only if additionally I_{D,g}≥τ_D; otherwise return a child-ambiguous distribution, such as normalized magnitudes within the block, while training the downstream loss on the block-level target. During training, add consistency between sign/permutation-aligned child predictions from two dictionary fits, but do not penalize disagreement within a block when I_D<τ_D. The paper supplies the information scalings and resolution logic; τ_S, τ_D and nuisance estimates are empirical hyperparameters. First test on synthetic sparse mixtures with two nearly colliding atoms, then on MNIST or CIFAR patches with an overcomplete learned dictionary. Compare ordinary top-1 atom prediction against the abstaining encoder at matched reconstruction error. Success is lower physical-ray false-discovery rate and better calibration of child-support confidence, with no loss in parent-block recall; test the predicted transition as N crosses roughly τ_D/s⁶ and as T crosses roughly τ_Sσ_+²/(β_-²s²).

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a reproducible synthetic two-ray collision experiment with sign-invariant ray distance, information-based parent/child resolution, replicated noisy test views, and dictionary-orientation uncertainty. The predicted transitions were observed: the test gate switched at T*=4 for s=0.1, the dictionary gate switched at N*=15,625 for s=0.2, and log-log sweeps measured slopes 2.000 for I_S versus s and 6.000 for I_D versus s. In the low-I_D regime abstention reduced physical child FDR from 0.338 to 0 while preserving parent recall at 1.0; this is a promising mechanism signal, though the orientation uncertainty was simulated from the predicted scaling rather than learned from raw dictionary-training data.

### Mechanism check

- Verdict: Built a reproducible synthetic two-ray collision experiment with sign-invariant ray distance, information-based parent/child resolution, replicated noisy test views, and dictionary-orientation uncertainty. The predicted transitions were observed: the test gate switched at T*=4 for s=0.1, the dictionary gate switched at N*=15,625 for s=0.2, and log-log sweeps measured slopes 2.000 for I_S versus s and 6.000 for I_D versus s. In the low-I_D regime abstention reduced physical child FDR from 0.338 to 0 while preserving parent recall at 1.0; this is a promising mechanism signal, though the orientation uncertainty was simulated from the predicted scaling rather than learned from raw dictionary-training data.
- Confidence: 7/10
- Limitations: This is a toy two-ray experiment, not a trained sparse encoder, dictionary-learning model, MNIST/CIFAR benchmark, or MoE router. Dictionary orientation error was injected with the predicted 1/(sqrt(N)s^2) scale rather than estimated from actual latent training moments, and thresholds were set to 1 rather than calibrated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.57515
- Paired wins: 5/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/422/bench_report.json)
- [collision_abstention.py](https://synthcore.org/code/422/collision_abstention.py)
- [collision_track.py](https://synthcore.org/code/422/collision_track.py)
- [report.md](https://synthcore.org/code/422/report.md)
- [report_bench_2026-08-31T115257.md](https://synthcore.org/code/422/report_bench_2026-08-31T115257.md)
- [results.json](https://synthcore.org/code/422/results.json)
- [run_bench.py](https://synthcore.org/code/422/run_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/422)

## Disclaimer

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