# Minkowski gradient barrier for neural fields

- ID: 990
- Canonical URL: https://synthcore.org/idea/990/minkowski-gradient-barrier-for-neural-fields
- API JSON: https://synthcore.org/api/idea/990.json
- API Markdown: https://synthcore.org/api/idea/990.md
- Verification status: unverified
- Source: [arXiv:2607.15956](https://arxiv.org/abs/2607.15956)
- Category: regularization
- Solves: stability, accuracy, generalization
- ML areas: mlp, loss, regularization, geometry
- Math tags: differential-geometry, pde, geometry, convex-analysis
- Ratings: usefulness 6/10; difficulty 4/10; novelty 5/10

## Idea description

Add a Born–Infeld/Minkowski-gradient barrier to a coordinate MLP so that its spatial gradient remains below a prescribed speed limit, rather than using an ordinary quadratic smoothness penalty. Couple the barrier with a forcing or task loss; under strong forcing, the resulting field should preferentially approach a distance-to-boundary-like profile while avoiding exploding derivatives and oscillatory solutions.

## Mathematical statement

The paper studies the Dirichlet operator -div(
abla u/\sqrt{1-|
abla u|^2}), where u:\Omega    o\mathbb{R} is the scalar field, \Omega\subset\mathbb{R}^N is a bounded domain, and |
abla u|<1 is the spacelike admissibility constraint. The parameters \lambda and \mu weight a linear reaction and nonlinear forcing h(x,u) in -div(
abla u/\sqrt{1-|
abla u|^2})=\lambda u+\mu h(x,u). The associated one-gradient energy density is the adapted Born–Infeld term \Phi(s)=1-\sqrt{1-s} for s=|
abla u|^2<1; its derivative with respect to 
abla u is 
abla u/\sqrt{1-|
abla u|^2}, which diverges as |
abla u|    o1^-. The paper reports that, when the global minimizer is positive and \mu    o+\infty, the limiting profile is \operatorname{dist}(x,\partial\Omega), the largest boundary-zero 1-Lipschitz profile. We exploit both properties: a differentiable barrier discourages gradient violations, while a large forcing term can be used to test whether the network learns a distance-like geometry.

## Key formulas

- $$-\operatorname{div}\left(\frac{\nabla u}{\sqrt{1-|\nabla u|^{2}}}\right)=\lambda u+\mu h(x,u),\qquad u|_{\partial\Omega}=0,$$
- $$\Phi(|\nabla u|^2)=1-\sqrt{1-|\nabla u|^2},\qquad |\nabla u|<1,$$
- $$\frac{\partial}{\partial(\nabla u)}\Phi(|\nabla u|^2)=\frac{\nabla u}{\sqrt{1-|\nabla u|^2}},$$
- $$\mathcal{L}(\theta)=\mathcal{L}_{\mathrm{task}}(u_\theta)+\beta\,\mathbb{E}_{x\sim\Omega}\left[1-\sqrt{\max(\varepsilon,1-\|\nabla_xu_\theta(x)\|_2^2)}\right]+\gamma\,\mathbb{E}_{x\sim\Omega}\left[\max(0,\|\nabla_xu_\theta(x)\|_2-1+\delta)^2\right].$$

## Implementation notes

Integrate this at the loss level of a coordinate MLP u_    heta(x), with x\in\mathbb{R}^N and spatial derivatives obtained by automatic differentiation. Use it first for a PINN or implicit scalar-field task with known boundary samples, such as solving Poisson's equation on a unit square or fitting the distance to the boundary of a disk. The network loss should be: (i) the original PDE residual or supervised field loss, (ii) a boundary loss \mathbb{E}_{x_b\in\partial\Omega}|u_    heta(x_b)|^2, and (iii) the Minkowski barrier above. Pseudocode is: sample interior points x_i and boundary points b_j; compute u_i=u_    heta(x_i) and g_i=
abla_xu_    heta(x_i); set s_i=\|g_i\|_2^2; compute B_i=1-\sqrt{\max(\varepsilon,1-s_i)}; compute the overflow penalty O_i=\max(0,\|g_i\|_2-1+\delta)^2; form L=L_task+\lambda_bL_boundary+\beta\operatorname{mean}(B_i)+\gamma\operatorname{mean}(O_i); update     heta\leftarrow    heta-\eta
abla_    heta L. Use \varepsilon=10^{-6} and \delta=0.02 initially. The paper supplies the geometric flux and its singular behavior; the task loss, \beta, \gamma, and the finite overflow handling are engineering choices. Log the fraction of points with \|g_i\|>1, the 95th/99th percentile gradient norm, boundary error, task residual, and gradient norm of the parameter update. Start with a 4-layer width-128 MLP on a 64    imes64 unit-square grid, compare against no regularizer and the standard penalty \alpha\|
abla u\|^2, and train at equal steps and FLOPs. A successful result is fewer gradient violations, fewer NaNs or training collapses, lower test residual, and—when the target has zero boundary values—a closer match to the true distance profile. For a forcing sweep, increase the analogue of \mu and test whether the normalized field approaches the analytic disk distance 1-\|x\| without developing boundary-layer oscillations.

## Disclaimer

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