# Uniformly Mixing Coulomb Particle Bank

- ID: 147
- Canonical URL: https://synthcore.org/idea/147/uniformly-mixing-coulomb-particle-bank
- API JSON: https://synthcore.org/api/idea/147.json
- API Markdown: https://synthcore.org/api/idea/147.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.24863](https://arxiv.org/abs/2608.24863)
- Category: sampling
- Solves: stability, sample-efficiency, generalization
- ML areas: embedding, moe-routing, sampling
- Math tags: functional-analysis, probability, stochastic-processes, geometry
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Represent a set of neural prototypes, mixture components, or latent particles by N points in R^2, and initialize or refresh them with Langevin dynamics targeting a quadratically confined logarithmic Coulomb gas. The logarithmic repulsion prevents particle collapse, while the paper's N-uniform logarithmic Sobolev inequality predicts that mixing need not degrade as the particle bank grows.

## Mathematical statement

The paper defines the Hamiltonian H_N(Z) = (1/2) sum_i |z_i|^2 + (1/N) sum_{i<j} g(z_i-z_j), where Z = (z_1,...,z_N) belongs to C^N, g(u) = -log|u|, and |u| is Euclidean distance in the plane. Its Gibbs distribution is dP_{N,beta}(Z) = Z_{N,beta}^{-1} exp(-beta H_N(Z)) dZ, where beta > 0 is inverse temperature and Z_{N,beta} is the normalizing constant. The main theorem gives a logarithmic Sobolev inequality Ent_P(f^2) <= (2/rho_beta) integral ||grad f||^2 dP, for suitable collision-free functions f, with rho_beta > 0 independent of N. Here Ent_P(f^2) = E_P[f^2 log(f^2)] - E_P[f^2] log(E_P[f^2]). The associated overdamped Langevin process has this Gibbs law as invariant distribution. Using grad_{z_i}[-log|z_i-z_j|] = -(z_i-z_j)/|z_i-z_j|^2, the repulsive force in the negative energy gradient is (1/N) sum_{j != i} (z_i-z_j)/|z_i-z_j|^2. The theorem's transferable property is dimension-uniform entropy contraction for this mean-field repulsive system, although discretization and collision smoothing require empirical validation.

## Key formulas

- $$\mathcal{H}_{N}(Z)=\frac{1}{2}\sum_{i=1}^{N}|z_i|^{2}+\frac{1}{N}\sum_{1\leq i<j\leq N}\mathsf{g}(z_i-z_j),\qquad \mathsf{g}(u)=-\log|u|.$$
- $$\mathrm{d}\mathbb{P}_{N,\beta}(Z)=\mathsf{Z}_{N,\beta}^{-1}e^{-\beta\mathcal{H}_{N}(Z)}\,\mathrm{d}Z=\mathsf{Z}_{N,\beta}^{-1}e^{-\beta|Z|^{2}/2}\prod_{i<j}|z_i-z_j|^{\beta/N}\,\mathrm{d}Z.$$
- $$\operatorname{Ent}_{\mathbb{P}_{N,\beta}}(f^{2})\leq \frac{2}{\rho_{\beta}}\int\|\nabla f(Z)\|^{2}\,\mathrm{d}\mathbb{P}_{N,\beta}(Z),\qquad \rho_{\beta}>0\ \text{independent of }N.$$
- $$z_i^{k+1}=z_i^{k}-\eta\left[z_i^{k}-\frac{1}{N}\sum_{j\neq i}\frac{z_i^{k}-z_j^{k}}{|z_i^{k}-z_j^{k}|^{2}+\varepsilon}\right]+\sqrt{\frac{2\eta}{\beta}}\,\xi_i^{k},\qquad \xi_i^{k}\sim\mathcal{N}(0,I_2).$$

## Implementation notes

Integrate the method at a prototype-bank or latent-particle tensor. Let Z in R^{N x 2} contain N learnable or sampled two-dimensional coordinates. These coordinates can be used directly for routing, or passed through a learned map e_i = g_theta(z_i) to produce high-dimensional prototypes. At initialization or during periodic refreshes, run K unadjusted Langevin steps. For each step, compute pairwise differences D_ij = z_i - z_j and squared distances r2_ij = ||D_ij||^2. Set R_i = (1/N) sum_{j != i} D_ij / (r2_ij + epsilon), then update z_i <- z_i - eta (z_i - R_i) + sqrt(2 eta / beta) xi_i, where xi_i is a standard two-dimensional Gaussian. The confinement term z_i comes from the quadratic potential, the factor 1/N is the paper's mean-field scaling, beta controls noise, and epsilon prevents numerical blow-up at collisions. For a routing experiment, use logits l_i(h) = -||W h - z_i||^2 / tau and train W and the downstream model while either freezing Z after initialization or refreshing Z periodically. The paper supplies the N-independent LSI claim for the exact singular target; estimate the effective mixing rate of the discretized sampler empirically using autocorrelation of mean radius, pairwise distance, and assignment histograms. Start with CIFAR-10, a small MLP or ViT, and N in {16, 64, 256}. Compare Gaussian prototypes, repulsion-only gradient descent, and Coulomb Langevin at matched initialization cost. Record nearest-neighbor distance, dead-prototype fraction, assignment entropy, validation accuracy, and autocorrelation time. Success means stable minimum distances, fewer dead routes, and roughly unchanged mixing statistics as N increases, together with equal-or-better accuracy at fixed compute.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a NumPy Coulomb-gas Langevin sampler with quadratic confinement, mean-field logarithmic repulsion, collision smoothing, finite-difference verification, autocorrelation estimation, and Gaussian prototype controls. The force identity matched finite differences to 2.1e-12 at the smallest tested step. Particle-radius integrated autocorrelation stayed essentially constant as N increased (76.34, 76.98, 78.50 for N=8,32,128), while Coulomb samples had no dead routes and higher assignment entropy than Gaussian controls, indicating a promising signal for uniform per-particle mixing and coverage. This does not verify the paper’s theorem itself; collective-radius autocorrelation decreased from 59.97 to 40.57.

### Mechanism check

- Verdict: Built a NumPy Coulomb-gas Langevin sampler with quadratic confinement, mean-field logarithmic repulsion, collision smoothing, finite-difference verification, autocorrelation estimation, and Gaussian prototype controls. The force identity matched finite differences to 2.1e-12 at the smallest tested step. Particle-radius integrated autocorrelation stayed essentially constant as N increased (76.34, 76.98, 78.50 for N=8,32,128), while Coulomb samples had no dead routes and higher assignment entropy than Gaussian controls, indicating a promising signal for uniform per-particle mixing and coverage. This does not verify the paper’s theorem itself; collective-radius autocorrelation decreased from 59.97 to 40.57.
- Confidence: 7/10
- Limitations: Toy 2D sampler only; no CIFAR, learned prototype map, downstream routing accuracy, FLOP-matched training, multiple-seed confidence intervals, exact singular discretization, or direct verification of the paper’s logarithmic Sobolev constant was tested. Assignment statistics use random 2D Gaussian queries and the Gaussian control uses a manually matched scale, so these are illustrative rather than a full application benchmark.

### Practical benchmark

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

## Artifacts

- [bench_report.json](https://synthcore.org/code/1188/bench_report.json)
- [coulomb_bank.py](https://synthcore.org/code/1188/coulomb_bank.py)
- [report.md](https://synthcore.org/code/1188/report.md)
- [report_bench_2026-09-03T024641.md](https://synthcore.org/code/1188/report_bench_2026-09-03T024641.md)
- [run_bench.py](https://synthcore.org/code/1188/run_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/1188)

## Disclaimer

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