△ Mechanism confirmed, baseline not beaten 2026

PDE Sinkhorn with asymmetric geometric boundaries

Implementation & benchmark of arXiv:2607.17904 — Reflected Schrodinger Bridge Problem over Sub-Riemannian Manifold

Usefulness7/10

Difficulty7/10

Novelty8/10

Sampling Accuracy Stability Sample efficiency Diffusion sampling Diffusion Loss Pde Optimal transport Stochastic processes Numerical analysis

Source paper: Reflected Schrodinger Bridge Problem over Sub-Riemannian Manifold arXiv:2607.17904 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Build a Schrödinger-bridge solver that represents the two Sinkhorn scaling factors as solutions of forward and backward Kolmogorov PDEs, rather than requiring explicit transition-density evaluation. Enforce an oblique Neumann condition on the backward factor and a normal no-flux condition on the forward factor, allowing degenerate diffusion and hard domain boundaries to be handled directly.

Formulas

$$\mathcal L_tu=b(x,t)\cdot\nabla u+\frac12\sum_{i=1}^m\bigl(g_i(x,t)\cdot\nabla\bigr)^2u,$$

$$\partial_t\varphi_t+\mathcal L_t\varphi_t=0,\qquad r(x,t)\cdot\nabla\varphi_t=0\quad\text{on }\partial\mathcal X,$$

$$\partial_t\hat\varphi_t=\mathcal L_t^*\hat\varphi_t,\qquad n(x)\cdot J_t[\hat\varphi_t](x)=0\quad\text{on }\partial\mathcal X,$$

$$\varphi_0\leftarrow\varphi_0\frac{\rho_0}{\varphi_0\hat\varphi_0},\qquad \hat\varphi_1\leftarrow\hat\varphi_1\frac{\rho_1}{\varphi_1\hat\varphi_1},\qquad \rho_t=\varphi_t\hat\varphi_t.$$

Mathematical statement

Let $g_i(x,t)$ be the admissible diffusion fields, $G(x,t)=\sum_i g_i g_i^\top$ their Gram matrix, and $b(x,t)$ the drift. The backward generator is $\mathcal L_tu=b\cdot\nabla u+\tfrac12\sum_i(g_i\cdot\nabla)^2u$. With $n=\nabla\xi/\|\nabla\xi\|$ and $r=Gn/(n^\top Gn)$, the backward Schrödinger factor $\varphi_t$ satisfies $\partial_t\varphi_t+\mathcal L_t\varphi_t=0$ and $r\cdot\nabla\varphi_t=0$ on the boundary. The forward factor $\hat\varphi_t$ evolves under the adjoint Fokker-Planck operator and uses normal no-flux, $n\cdot J_t[\hat\varphi_t]=0$, where $J_t[q]=bq-\tfrac12\sum_i g_i(g_i\cdot\nabla q)$ is the probability current in the directional-diffusion approximation. The bridge density factorizes as $\rho_t=\varphi_t\hat\varphi_t$. Sinkhorn alternates endpoint rescaling of the factors to match prescribed densities $\rho_0$ and $\rho_1$. The transferable result is that boundary conditions are imposed directly on PDE factors and fluxes, avoiding an explicit transition kernel.

Implementation notes

Use this as a low-dimensional constrained generative-model or bridge module where PDE discretization is feasible. Create a mesh or finite-volume grid covering $\mathcal X$, store positive arrays for $\varphi_t$ and $\hat\varphi_t$, and define $\xi$, $n$, $g_i$, and $b$ at cell centers and boundary faces. Discretize $\mathcal L$ with directional derivatives along each $g_i$, using an implicit or positivity-preserving scheme. During every backward solve, impose $r\cdot\nabla\varphi=0$ by eliminating the boundary ghost value along $r=Gn/(n^\top Gn+\varepsilon)$. During every forward solve, impose zero normal numerical flux by setting the boundary-face current to zero rather than reflecting particles individually. Run iterative proportional fitting in log-space: initialize $\varphi_0=1$, propagate it forward, rescale the terminal factor to match $\rho_1$, propagate the adjoint factor backward, rescale the initial factor to match $\rho_0$, and repeat with damping $q\leftarrow(1-\eta)q+\eta q_{\rm new}$. The PDE equations, factorization, and asymmetric boundary conditions are computed from the paper; grid resolution, time-step stability, and damping are empirical. First test a 2D annulus with two separated Gaussian endpoint densities and compare against kernel Sinkhorn, unconstrained diffusion, and particle rejection. Track marginal KL error, boundary mass leakage, bridge action, and iteration count. Success means stable convergence with near-zero boundary flux and lower endpoint error than rejection-based particle SB under anisotropic or rank-deficient diffusion.

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 NumPy/SciPy MVP containing the oblique boundary calculation, conservative finite-volume reflected diffusion, explicit stability sweep, and endpoint Sinkhorn scaling comparison. The mechanism manifested: for G=[[1,c],[c,1]], the oblique residual stayed 0 while the normal-BC residual was exactly c; no-flux propagation conserved mass to 1e-12 or better across diffusion strengths; and positivity held through ratio 1.0 of the predicted dt=7.8125e-5 bound but failed strongly at 1.05. Reflected and unconstrained kernel Sinkhorn both matched endpoint marginals to roughly 1e-15 KL in this easy 1D test, so no accuracy win over the baseline was demonstrated.

Agent confidence: 8/10

Baseline: Unconstrained Gaussian kernel Sinkhorn: initial KL 1.23e-17, terminal KL 5.03e-15; endpoint boundary-cell mass about 9.83e-4.
Idea: Reflected finite-volume/PDE kernel Sinkhorn: initial KL 3.96e-17, terminal KL 4.78e-16; mass conservation error <=8.8e-13, imposed boundary flux exactly 0; oblique BC residual 0 across c=0..0.95 while normal BC residual grew linearly 0..0.95; positivity transition observed at dt/dt_bound between 1.0 and 1.05.

Limitations: This is a low-dimensional numerical MVP, not the requested 2D annulus experiment. The backward oblique condition is verified algebraically and is not yet integrated into a full 2D backward-factor PDE solve; the forward test uses pure diffusion, no drift, and the Sinkhorn comparison uses a precomputed matrix exponential rather than repeated time-stepped factor propagation. No particle-rejection baseline, bridge action, runtime benchmark, or rank-deficient multidimensional directional diffusion was tested.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.17904) Sep 2, 2026

Significantly WORSE than the tuned baseline (n/a, p=0.0152, wins 1 / 8 paired seeds).

Benchmark: Anisotropic_oblique_bvp
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0 (n/a; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.0152 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 0 configs

Limitations:

The custom task is a small manufactured 2D BVP rather than a full Schrödinger-bridge/Sinkhorn PDE solve; the intervention uses boundary derivative penalties rather than explicit forward/backward factor propagation, diffusion discretization, or endpoint rescaling. No annulus geometry, rank-deficient multidimensional diffusion, particle-rejection baseline, bridge action, or large-scale neural architecture was tested.

How to run: python3 run_bench.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Built a local 2D anisotropic oblique-boundary BVP track and an end-to-end boundary-penalty MLP intervention using the same architecture and Adam budget as the baseline. The trained-model mechanism signature was confirmed: mean oblique residual decreased from 0.02720 to 0.01003 and mean normal residual from 0.04277 to 0.00812. However, test MSE was significantly worse for the idea, so it does not win the benchmark metric.

Baseline: Best lr=0.003, weight_decay=0.0; full 8-seed test MSE mean=9.435732454221579e-05, std=2.4013090964758765e-05.
Idea: Best lr=0.003, weight_decay=0.0, bc_weight=0.1; full 8-seed test MSE mean=1.1699842414609309e-04, std=3.095430559427717e-05.
Mechanism signature: Baseline oblique residual=0.02720322646200657, idea oblique residual=0.01003198220860213; baseline normal residual=0.04276570677757263, idea normal residual=0.008119827311020344; confirmed=true
Paired comparison: delta_mean=2.2641099803877296e-05, idea_wins=1/8, permutation p=0.0152, verdict='idea worse (significant)'

Benchmark result