# Profile-Preserving Multislice Noise

- ID: 842
- Canonical URL: https://synthcore.org/idea/842/profile-preserving-multislice-noise
- API JSON: https://synthcore.org/api/idea/842.json
- API Markdown: https://synthcore.org/api/idea/842.md
- Verification status: beats_baseline
- Source: [arXiv:2607.13208](https://arxiv.org/abs/2607.13208)
- Category: sampling
- Solves: accuracy, sample-efficiency, stability
- ML areas: graph-nn, diffusion-sampling, data-augmentation
- Math tags: combinatorics, representation-theory, probability, spectral-theory
- Ratings: usefulness 7/10; difficulty 3/10; novelty 7/10

## Idea description

For an input with exactly $\alpha_a$ occurrences of each state $a\in\{0,\ldots,n-1\}$, corrupt it by repeatedly swapping two positions with different states instead of independently resampling tokens. This defines a Markov process on the connected fixed-profile multislice, preserving global composition exactly and avoiding the distribution shift caused by ordinary categorical masking.

## Mathematical statement

The paper models a colored configuration with $m$ positions as a basis vector in $(\mathbb{C}^{n})^{\otimes m}$, where each position has state $a\in\{0,\ldots,n-1\}$ and profile $\alpha=(\alpha_0,\ldots,\alpha_{n-1})$ satisfies $\sum_a\alpha_a=m$. The multislice adjacency operator is the sum of exchanges of unequal states between two positions. Let $E_{ab}$ be the $n\times n$ matrix unit satisfying $E_{ab}e_b=e_a$ and $E_{ab}e_c=0$ for $c\neq b$, and let $E_{ab}^{(p)}$ act on position $p$. The exchange adjacency is $A_\alpha=\sum_{1\le p<q\le m}\sum_{a\neq b}E_{ab}^{(p)}E_{ba}^{(q)}$. Its degree is $d_\alpha=\frac12(m^2-\sum_a\alpha_a^2)=\sum_{a<b}\alpha_a\alpha_b$, because exactly that many unordered position pairs have unequal states. The random-transposition kernel is $P_\alpha=(1-\rho)I+(\rho/d_\alpha)A_\alpha$, with noise strength $\rho\in[0,1]$; every transition preserves $\alpha$.

## Key formulas

- $$(\mathbb{C}^{n})^{\otimes m},\qquad \alpha=(\alpha_0,\dots,\alpha_{n-1}),\qquad \sum_{a=0}^{n-1}\alpha_a=m.$$
- $$A_\alpha=\sum_{1\le p<q\le m}\sum_{a\neq b}E_{ab}^{(p)}E_{ba}^{(q)}.$$
- $$d_\alpha=\frac{1}{2}\left(m^2-\sum_{a=0}^{n-1}\alpha_a^2\right)=\sum_{a<b}\alpha_a\alpha_b.$$
- $$P_\alpha=(1-\rho)I+\frac{\rho}{d_\alpha}A_\alpha.$$

## Implementation notes

Integrate this at the data-loader or diffusion-noise layer for graph inputs represented by $m$ edge tokens, each taking one of $n$ states. First compute the profile $\alpha$ of each training example and store positions grouped by state. For one corruption step, choose an unordered pair of positions $(p,q)$ uniformly from all pairs with different states, then swap their state labels; repeat $k$ times, where $k$ is sampled from a schedule such as Uniform${\{0,\ldots,K\}}$ or chosen so that the expected number of swaps is $\rho d_\alpha$. This is an exact sampler for the random-transposition kernel $P_\alpha$; no matrix $A_\alpha$ should be materialized. Pseudocode is: `x_noisy=x; for t in range(k): a=random_state_weighted_by(alpha); b=random_state_weighted_by(alpha excluding a); p=random position in bucket[a]; q=random position in bucket[b]; swap(x_noisy[p],x_noisy[q]); update the two buckets`. Train a denoising autoencoder, graph classifier, or discrete diffusion model to predict the original configuration from `x_noisy`, conditioning on $\alpha$ and $k$. The mathematical quantities are the preserved profile, $d_\alpha$, and the exchange transition; the noise schedule and model parameterization are empirical choices. The first cheap experiment should use a small edge-state graph classifier on synthetic fixed-edge-count graph data, comparing independent categorical masking, standard random edge flips, and multislice swaps at equal corruption levels and equal model FLOPs. Measure validation accuracy, denoising cross-entropy, and performance when test profiles differ. Success is lower denoising loss and better fixed-profile generalization without profile violations; a direct invariant check must confirm that every corrupted sample has exactly the original count of every state.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built an exact multislice swap sampler and a CUDA toy denoising experiment. The three mechanism predictions were confirmed: degree matched d=Σ_{a<b}α_aα_b exactly across four profiles, all 7-swap trials had zero profile violations, and empirical one-step self-transition probabilities matched 1−ρ within 0.004 across 12 profile/noise settings. In the toy denoiser at ρ=0.5, multislice noise achieved loss 0.673 and token accuracy 0.906 versus independent masking loss 0.911 and accuracy 0.653, so the promised profile-preservation effect is real in this setup.

### Mechanism check

- Verdict: Built an exact multislice swap sampler and a CUDA toy denoising experiment. The three mechanism predictions were confirmed: degree matched d=Σ_{a<b}α_aα_b exactly across four profiles, all 7-swap trials had zero profile violations, and empirical one-step self-transition probabilities matched 1−ρ within 0.004 across 12 profile/noise settings. In the toy denoiser at ρ=0.5, multislice noise achieved loss 0.673 and token accuracy 0.906 versus independent masking loss 0.911 and accuracy 0.653, so the promised profile-preservation effect is real in this setup.
- Confidence: 9/10
- Limitations: The ML comparison is a small synthetic fixed-profile sequence denoiser rather than a graph model or diffusion system. It tests one corruption strength and one profile, does not measure FLOPs or wall-clock speed, and does not evaluate generalization to shifted profiles or long-run multislice mixing.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 8/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/344/bench_report.json)
- [experiment.py](https://synthcore.org/code/344/experiment.py)
- [report.md](https://synthcore.org/code/344/report.md)
- [report_bench_2026-08-31T014647.md](https://synthcore.org/code/344/report_bench_2026-08-31T014647.md)
- [results.json](https://synthcore.org/code/344/results.json)
- [stage2_multislice_bench.py](https://synthcore.org/code/344/stage2_multislice_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/344)

## Disclaimer

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