Unverified 2026

Distance-Powered Boundary Ansatz for Fractional PINNs

Implementation & benchmark of arXiv:2607.28436 — On boundary regularity for the fractional p-Laplacian with unbounded reactions

Usefulness6/10

Difficulty4/10

Novelty6/10

Architecture Accuracy Stability Sample efficiency Mlp Loss Training dynamics Pde Functional analysis Geometry Convex analysis

Source paper: On boundary regularity for the fractional p-Laplacian with unbounded reactions arXiv:2607.28436 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Represent the predicted solution as $u_{\theta}(x)={\rm d}_{\Omega}(x)^s v_{\theta}(x)$, where $v_{\theta}$ is an unconstrained neural network and ${\rm d}_{\Omega}$ is the distance to the boundary. This builds the fractional Dirichlet boundary layer into the architecture and leaves the network to learn the smoother quotient $u/{\rm d}_{\Omega}^{s}$, which the paper proves extends Hölder-continuously to the boundary when the reaction has sufficient integrability.

Formulas

$$E(u)=\frac{1}{p}\iint_{\mathbb{R}^{N}\times\mathbb{R}^{N}}\frac{|u(x)-u(y)|^{p}}{|x-y|^{N+ps}}\,dx\,dy,$$

$$\min\Big\{1,\,\frac{\rho^{N}}{R_{\Omega}^{N}}\Big\}\frac{C^{-1}}{{\rm d}_{\Omega}^{\alpha}(x_{0})}\leq\int_{\Omega^{c}}\frac{dx}{|x-x_{0}|^{N+\alpha}}\leq\frac{C}{{\rm d}_{\Omega}^{\alpha}(x_{0})},$$

$$u_{\theta}(x)={\rm d}_{\Omega}(x)^{s}v_{\theta}(x),\qquad v_{\theta}(x)=\operatorname{MLP}_{\theta}(x),$$

$$\widehat E(\theta)=\frac{1}{pM}\sum_{m=1}^{M}w_m\frac{|u_{\theta}(x_m)-u_{\theta}(y_m)|^{p}}{|x_m-y_m|^{N+ps}},\qquad x_m,y_m\in\Omega.$$

Mathematical statement

The paper studies the homogeneous nonlocal Dirichlet problem $(-\Delta)^s_pu=f$ in $\Omega$, $u=0$ in $\mathbb{R}^N\setminus\Omega$, where $0<s<1$, $p>1$, $N$ is the spatial dimension, and $(-\Delta)^s_p$ is the Euler derivative of the convex energy $E(u)=\frac1p\iint_{\mathbb{R}^N\times\mathbb{R}^N}\frac{|u(x)-u(y)|^p}{|x-y|^{N+ps}}dxdy$. Here ${\rm d}_{\Omega}(x)=\inf_{z\in\partial\Omega}|x-z|$. The main boundary result states that if $f\in L^q(\Omega)$ with $q>N/s$, then $u$ is $s$-Hölder up to the boundary and $u/{\rm d}_{\Omega}^{s}$ has a Hölder-continuous extension to $\overline\Omega$. The geometric Lemma 2.11 gives, for an exterior-ball condition with radius $\rho$ and inradius $R_{\Omega}$, $\min\{1,\rho^N/R_{\Omega}^N\}C^{-1}{\rm d}_{\Omega}(x_0)^{-\alpha}\leq\int_{\Omega^c}|x-x_0|^{-N-\alpha}dx\leq C{\rm d}_{\Omega}(x_0)^{-\alpha}$ for every $\alpha>0$. The architecture exploits the theorem's smoother quotient $v=u/{\rm d}_{\Omega}^s$; the kernel estimate supplies an optional boundary-dependent sampling weight for the nonlocal energy.

Implementation notes

(1) Integration point: replace the output head of a PINN or neural operator for a fractional Dirichlet problem. Given an input coordinate $x\in\Omega$, compute a differentiable distance approximation $d(x)={\rm d}_{\Omega}(x)$ and return $u_\theta(x)=\max(d(x),\varepsilon)^s v_\theta(x)$. Set the output to exactly zero for sampled exterior points. For simple domains use an analytic signed distance; for arbitrary $C^{1,1}$ domains use a precomputed signed-distance grid with interpolation, or a frozen distance network. (2) Pseudocode: sample interior pairs $(x_m,y_m)$ and reaction points $x_i$; compute $d_x,d_y$; evaluate $u_x=d_x^s v_\theta(x_m)$ and $u_y=d_y^s v_\theta(y_m)$; accumulate the fractional energy term $|u_x-u_y|^p/|x_m-y_m|^{N+ps}$; differentiate it with respect to $\theta$; optionally add a weak-form reaction term $-f(x_i)u_\theta(x_i)$ and source normalization. Use pair weights $w_m\propto1/\pi(x_m,y_m)$ for importance sampling. To focus on the boundary layer, sample one point with distance $r$ from a log-uniform distribution and use the paper's estimate with $\alpha=ps$ to set a diagnostic weight proportional to $d(x)^{-ps}$, clipped to $[w_{\min},w_{\max}]$. (3) Mathematics versus estimation: the exponent $s$ and distance singularity $d^{-\alpha}$ come from the paper; $d(x)$, the Hölder quotient, and Monte Carlo estimates of $E$ are numerical quantities. Do not differentiate through a noisy nearest-boundary search. (4) First experiment: solve the one-dimensional fractional Poisson problem on $\Omega=(-1,1)$ with $p=2$ and manufactured forcing, comparing a standard unconstrained PINN, a linear ansatz $d(x)v_\theta(x)$, and the proposed $d(x)^sv_\theta(x)$. Use equal parameter counts and equal pair samples. Measure relative $L^2$ error, boundary error on points with $d(x)<0.05$, energy residual, and optimizer steps to reach fixed error. Success is lower boundary error and smoother training at equal wall-clock time; a second test should use a 2D disk with random $L^q$ forcing and check whether fewer boundary samples are needed for the same global error.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.