# Fractional Boundary-Factored Neural Solver

- ID: 360
- Canonical URL: https://synthcore.org/idea/360/fractional-boundary-factored-neural-solver
- API JSON: https://synthcore.org/api/idea/360.json
- API Markdown: https://synthcore.org/api/idea/360.md
- Verification status: beats_baseline
- Source: [arXiv:2607.02312](https://arxiv.org/abs/2607.02312)
- Category: architecture
- Solves: accuracy, stability, sample-efficiency
- ML areas: mlp, loss, architecture, training
- Math tags: pde, functional-analysis, approximation-theory
- Ratings: usefulness 6/10; difficulty 4/10; novelty 7/10

## Idea description

For a fractional Dirichlet problem, replace a free coordinate network N_theta(x) with u_theta(x)=d(x)^a N_theta(x), where d(x)=dist(x,boundary) and 0<a<1 is the fractional order. Train the regular quotient v_theta=u_theta/d^a=N_theta and use a weighted gradient loss that reflects the paper's boundary estimate.

## Mathematical statement

The paper studies P u=f in a bounded domain Omega, with u=0 outside Omega, for a fractional-order operator P of order 2a, where 0<a<1. It states that the solution space contains a component of the form d^a times a boundary-value lifting, with d(x)=dist(x,partial Omega). For a>1/2, the gradient estimate is ||d^(1-a+s) nabla(u/d^a)||_{H_q^(s+t)} <= C(||f||_{H_q^t}+||u||_{L_q}), where s and t are small regularity indices, q is an integrability exponent, C is problem-dependent, and the norms are taken on the closure of Omega. The transferable fact is that v=u/d^a is more regular than u, while d^(1-a+s) nabla v is the correctly weighted quantity near the boundary.

## Key formulas

- $$P u=f\text{ on }\Omega,\qquad u=0\text{ on }\mathbb{R}^n\u005csetminus\Omega,$$
- $$u_\theta(x)=d(x)^a v_\theta(x),\qquad d(x)=\operatorname{dist}(x,\partial\Omega),\qquad v_\theta(x)=N_\theta(x),$$
- $$\left\|d^{1-a+s}\nabla\left(\frac{u}{d^a}\right)\right\|_{\dot H_q^{s+t}(\overline\Omega)}\leq C\left(\|f\|_{\overline H_q^t(\Omega)}+\|u\|_{L_q(\Omega)}\right),\qquad a>\frac12,$$
- $$\mathcal L_\theta=\frac{1}{|X|}\sum_{x_i\in X}|P[u_\theta](x_i)-f(x_i)|^2+\lambda_g\frac{1}{|X_b|}\sum_{x_i\in X_b}\left(d(x_i)^{1-a+s}|\nabla v_\theta(x_i)|\right)^2.$$

## Implementation notes

Integrate the construction at the output layer of a coordinate MLP or PINN solving a known fractional PDE. Precompute a nonnegative distance field d(x) for the domain; during optimization use d_delta(x)=sqrt(d(x)^2+delta^2)-delta if exact distance derivatives are unstable, and define u_theta=d_delta^a N_theta. Set the output to zero outside the domain. Each iteration should: (1) sample interior collocation points X, boundary-near points X_b, and forcing values f(x); (2) evaluate v=N_theta(X) and u=d_delta^a v; (3) compute the fractional PDE residual r_i=P[u](x_i)-f(x_i) using the existing fractional-operator discretization or quadrature routine; (4) compute g_i=d_delta(x_i)^(1-a+s) nabla N_theta(x_i) by automatic differentiation; and (5) minimize mean(r_i^2)+lambda_g mean(||g_i||_2^2). The factor d^(1-a+s) comes directly from the paper. The constant C and exact Sobolev norms are not computed; lambda_g, s, and delta are selected by validation. Start with a unit disk or square, a=0.75, a manufactured solution u*=d^a v* with smooth v*, a four-layer width-128 MLP, and equal collocation budgets for the baseline and factored model. Compare against a free-output PINN with a standard boundary penalty. Measure relative L2 error, boundary-strip error, residual variance, and loss decrease per optimizer step. Success means lower boundary error and faster residual reduction at equal compute, especially as a approaches one.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built and tested the fractional boundary-factored MLP in experiment.py. The numerical check reproduced slopes -0.2495 versus expected -0.25 for the raw derivative and 0.3500 versus expected 0.35 for the weighted quotient gradient. In the matched toy run, boundary-strip relative L2 improved from 0.4213 to 0.1598, but global relative L2 worsened from 0.0779 to 0.0977, so the effect is a real boundary-accuracy win rather than an overall accuracy or speed win.

### Mechanism check

- Verdict: Built and tested the fractional boundary-factored MLP in experiment.py. The numerical check reproduced slopes -0.2495 versus expected -0.25 for the raw derivative and 0.3500 versus expected 0.35 for the weighted quotient gradient. In the matched toy run, boundary-strip relative L2 improved from 0.4213 to 0.1598, but global relative L2 worsened from 0.0779 to 0.0977, so the effect is a real boundary-accuracy win rather than an overall accuracy or speed win.
- Confidence: 8/10
- Limitations: The test is a 1D manufactured supervised fitting proxy, not a genuine fractional PDE quadrature/PINN residual experiment. Only one seed, one order, one regularization weight, and one architecture were tested; no higher-dimensional domain or true fractional operator was evaluated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0317
- Paired wins: 6/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/1274/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/1274/bench_report.json)
- [experiment.py](https://synthcore.org/code/1274/experiment.py)
- [fractional_dirichlet_track.py](https://synthcore.org/code/1274/fractional_dirichlet_track.py)
- [report.md](https://synthcore.org/code/1274/report.md)
- [report_bench_2026-09-04T002354.md](https://synthcore.org/code/1274/report_bench_2026-09-04T002354.md)
- [results.json](https://synthcore.org/code/1274/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1274)

## Disclaimer

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