△ Mechanism confirmed, baseline not beaten 2026

Walk-on-Spheres stochastic target layer

Implementation & benchmark of arXiv:2608.09494 — Walk-on-Spheres Monte Carlo and deep neural network approximations of elliptic PDEs with drift and killing

Usefulness7/10

Difficulty5/10

Novelty6/10

Training Sample efficiency Scalability Accuracy Training Mlp Loss Stochastic processes Probability Numerical analysis

Source paper: Walk-on-Spheres Monte Carlo and deep neural network approximations of elliptic PDEs with drift and killing arXiv:2608.09494 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Train a neural network to represent an elliptic solution using Walk-on-Spheres rollouts as stochastic targets instead of evaluating a mesh-based PDE residual. For each input point, recursively jump to a random point on the largest interior sphere, accumulate source contributions, evaluate boundary data at termination, and regress the network output to the resulting Monte Carlo estimate.

Formulas

$$w(x)=\mathbb{E}\left[h_{0}(\tau^{x},x+B_{\tau^{x}})\right]+\mathbb{E}\left[\int_{0}^{\tau^{x}}h_{1}(s,x+B_{s})\,\mathrm{d}s\right].$$

$$r(z)=\operatorname{dist}(z,\partial D),\qquad z_{j+1}=z_{j}+r(z_{j})U_{j},\quad U_{j}\sim\operatorname{Unif}(\mathbb{S}^{d-1}),\qquad \Delta t_{j}=\frac{r(z_{j})^{2}}{2d}.$$

$$\widehat w_{M,K}(x)=\frac{1}{K}\sum_{i=1}^{K}\left[e^{-\lambda t_{M}^{i}}h_{0}(t_{M}^{i},z_{M}^{i})+\sum_{j=0}^{M-1}e^{-\lambda t_{j}^{i}}h_{1}(t_{j}^{i},z_{j}^{i})\Delta t_{j}^{i}\right].$$

$$\operatorname{Err}\lesssim C_{\mathrm{geom}}\delta+C_{\mathrm{trunc}}\left(1+\frac{\beta^{2}\delta^{2}}{2\operatorname{diam}(D)^{2}}\right)^{-M}+\frac{C_{1}}{K}+\frac{C_{2}(M)}{KM^{2}}.$$

Mathematical statement

The paper represents the target function as $w(x)=\mathbb{E}[h_{0}(\tau^{x},x+B_{\tau^{x}})]+\mathbb{E}[\int_{0}^{\tau^{x}}h_{1}(s,x+B_{s})\,\mathrm{d}s]$, where $D\subset\mathbb{R}^{d}$ is a bounded convex domain, $B$ is standard $d$-dimensional Brownian motion, $\tau^{x}=\inf\{t\ge 0:x+B_{t}\in\partial D\}$ is the exit time, $h_{0}$ is terminal boundary data, and $h_{1}$ is a running source term. The Walk-on-Spheres approximation uses the largest ball contained in $D$ around the current point $z$: $r(z)=\operatorname{dist}(z,\partial D)$, a direction $U\sim\operatorname{Unif}(\mathbb{S}^{d-1})$, the next point $z^{+}=z+r(z)U$, and the Brownian exit-time proxy $\Delta t=r(z)^{2}/(2d)$. For killing coefficient $\lambda\ge0$, future contributions are multiplied by $\exp(-\lambda\Delta t)$. With $z_{0}=x$, $z_{j+1}=z_{j}+r(z_{j})U_{j}$, and $t_{j+1}=t_{j}+\Delta t_{j}$, the estimator is $\widehat w_{M,K}(x)=K^{-1}\sum_{i=1}^{K}[\exp(-\lambda t_{M}^{i})h_{0}(t_{M}^{i},z_{M}^{i})+\sum_{j=0}^{M-1}\exp(-\lambda t_{j}^{i})h_{1}(t_{j}^{i},z_{j}^{i})\Delta t_{j}^{i}]$. The extracted bound separates geometric error, a truncation factor $(1+\beta^{2}\delta^{2}/(2\operatorname{diam}(D)^{2}))^{-M}$, and sampling terms proportional to $1/K$ and $1/(KM^{2})$.

Implementation notes

(1) Integration point: use this estimator as a stochastic supervised loss for an MLP $f_{\theta}(x)$ approximating the PDE solution. Sample interior points from $D$ in each minibatch; do not differentiate through random directions or stopping decisions. Boundary data $h_{0}$ and source data $h_{1}$ can be analytic functions or fixed auxiliary networks. (2) Pseudocode: for each point $x_b$ and path $i$, set $z=x_b$, $t=0$, and $y=0$. For $j=0,\ldots,M-1$, compute $r=\operatorname{dist}(z,\partial D)$, sample $U\sim\operatorname{Unif}(\mathbb{S}^{d-1})$, set $\Delta t=r^{2}/(2d)$, add $\exp(-\lambda t)h_{1}(t,z)\Delta t$ to $y$, then update $z\leftarrow z+rU$ and $t\leftarrow t+\Delta t$. After the loop add $\exp(-\lambda t)h_{0}(t,z)$. Average the $K$ returns and minimize $K^{-1}\sum_b(f_{\theta}(x_b)-\widehat w_{M,K}(x_b))^{2}$. Stop early when $r<\delta$ and project $z$ to the boundary. (3) Compute from the mathematics: sphere radius, time increment, killing weight, rollout recursion, and truncation depth. Estimate empirically: target variance, error constants, and whether increasing $M$ or $K$ is more cost-effective. (4) First experiment: solve Poisson and screened-Poisson equations on $[0,1]^d$ and a unit ball for dimensions 2, 4, and 8. Compare against a PINN with the same parameter count and Adam budget. Measure relative test error, wall-clock time, and error versus dimension. Success means lower error at equal training FLOPs or substantially better dimension scaling, with approximately inverse-$K$ reduction in target variance.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significantly worse ✗

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a readable Walk-on-Spheres stochastic target generator and MLP training MVP in wos_experiment.py. The mechanism manifested: target variance followed the predicted 1/K law, with K times variance approximately 0.016 to 0.018, and finite-depth estimates approached the reference from 0.734 at M=2 to 0.9995 at M=64. The learning comparison gave relative RMSE 0.345 for WOS versus 0.350 for exact targets, but WOS took 4.31 seconds versus 1.70 seconds, so no practical win was demonstrated. The checks also identified that the supplied Delta-t formula is half the standard Brownian ball mean exit time under the usual convention.

Agent confidence: 8/10

Baseline: Exact analytic square-Poisson targets: relative RMSE 0.3503, 1.70 seconds
Idea: WOS square-Poisson targets with K=8 and M=25: relative RMSE 0.3450, 4.31 seconds; variance ratio at K=128 was 0.00692 versus predicted 1/128 = 0.0078125

Limitations: Only a tiny 2D square learning experiment was run. Higher dimensions, unit-ball learning, screened Poisson, a true PINN residual baseline, FLOP-normalized comparisons, boundary projection, and corrected-versus-supplied time-step conventions were not fully benchmarked.

How to run: python3 wos_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.09494) Sep 2, 2026

Significantly WORSE than the tuned baseline (+345.4%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Poisson_boundary
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 0.0023
Idea mean: 0.0102
Effect (Δ): 0.0079 (+345.4%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±119.8%
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only registered poisson_boundary and mlp_tiny were tested; no higher-dimensional or screened-Poisson task, FLOP timing, or true residual PINN baseline.

How to run: python3 stage2_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 10/10

The WOS stochastic-target system was significantly worse than the tuned exact-label baseline on the structurally matched PDE benchmark. The sphere identity and inverse-K variance check passed, but the required task metric comparison did not show a win.

Baseline: Mean test MSE 0.0022902884520590305; tuned lr=0.01.
Idea: Mean test MSE 0.01020078727742657; best lr=0.003, K=8, M=24.

Benchmark result