# Coulomb field corrector for particle-based generator training

- ID: 817
- Canonical URL: https://synthcore.org/idea/817/coulomb-field-corrector-for-particle-based-generator-trainin
- API JSON: https://synthcore.org/api/idea/817.json
- API Markdown: https://synthcore.org/api/idea/817.md
- Verification status: unverified
- Source: [arXiv:2607.12579](https://arxiv.org/abs/2607.12579)
- Category: dynamics
- Solves: stability, accuracy, sample-efficiency
- ML areas: gan, training-dynamics, loss, data-augmentation
- Math tags: optimal-transport, pde, dynamical-systems, numerical-analysis
- Ratings: usefulness 6/10; difficulty 4/10; novelty 7/10

## Idea description

Use one or a few explicit Coulomb transport steps on generated particles as a differentiable or detached corrector, then train the generator to imitate the corrected particles. This separates global distribution matching from the generator parameterization and can reduce adversarial-gradient noise and mode collapse.

## Mathematical statement

Writing the paper's equation as a continuity equation \(\partial_t\rho+\operatorname{div}(\rho v)=0\) identifies the Lagrangian velocity \(v(x)=-\nabla\mathsf g*(\rho-\mu)(x)\). For empirical particles \(\rho_N=N^{-1}\sum_i\delta_{x_i}\), let \(u=\mathsf g*(\rho_N-\mu)\), equivalently \(-\Delta u=\rho_N-\mu\) on the torus. An explicit Euler step is \(x_i^+=x_i-\eta\nabla u(x_i)\). The source contribution should omit the particle's singular diagonal interaction or use a smoothed deposition kernel. The paper establishes PL-type convergence only under positivity and localization assumptions; targets with vacuum and sources far from the target can lack a uniform multiplicative decay rate. Therefore the corrector should use a conservative adaptive step rather than assume a universal constant step size.

## Key formulas

- $$\partial_t\rho-\operatorname{div}\!\left(\rho\nabla\mathsf g*(\rho-\mu)\right)=0,$$
- $$\partial_t\rho+\operatorname{div}(\rho v)=0,\qquad v(x)=-\nabla\mathsf g*(\rho-\mu)(x),$$
- $$x_i^{+}=x_i-\eta\,\nabla u(x_i),\qquad -\Delta u=\rho_N-\mu,\qquad \rho_N=\frac1N\sum_{i=1}^{N}\delta_{x_i},$$
- $$\eta_k=\min\!\left(\eta_{\max},\frac{\gamma\,\mathcal E_k}{\frac1N\sum_i\|\nabla u_k(x_i)\|^2+\varepsilon}\right),\qquad 0<\gamma<1.$$

## Implementation notes

Add a corrector between generator sampling and the generator update. For each minibatch, calculate raw particles \(x_i=G_\theta(z_i)\) and target particles \(y_j\), normalize them into a periodic box, and construct differentiable grid histograms \(r\) and \(m\). Solve the discrete Poisson equation by FFT: for each nonzero frequency index \(k\), set \(\widehat u(k)=\widehat{(r-m)}(k)/\lambda_k\), where \(\lambda_k\) is the positive eigenvalue of the chosen discrete negative Laplacian, and set \(\widehat u(0)=0\). Differentiate the interpolated potential to obtain \(F_i=\nabla u(x_i)\). To avoid self-force artifacts, either compute the source field with particle \(i\) omitted, use Gaussian splats wider than one grid cell, or subtract the analytically known diagonal contribution. Define \(\mathcal E_k=\langle r-m,u\rangle\) and choose \(\eta_k=\min(\eta_{max},\gamma\mathcal E_k/(N^{-1}\sum_i\|F_i\|^2+\varepsilon))\). Apply one to three steps \(\tilde x_i=x_i-\eta_kF_i\). In the detached version, train with \(\|G_\theta(z_i)-\operatorname{stopgrad}(\tilde x_i)\|^2\); in the fully differentiable version, backpropagate through the field computation. The Poisson solve, velocity, and Euler step are mathematical transfers; detachment, smoothing, and step clipping are empirical adaptations. First compare ordinary GAN, MMD, and GAN-plus-corrector on 2-D eight-Gaussians and Swiss roll using the same MLP generator and minibatch size. Record pre/post-correction Coulomb energy, mode coverage, nearest-target distance, gradient variance, radius growth, and wall-clock cost. The desired signal is a substantial immediate energy reduction and improved mode coverage at equal generator updates without particle explosion or target memorization.

## Disclaimer

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