An intrinsic sampling method kept simulated states on mathematically constrained spaces and, in one analytical test, matched the expected regional balance more closely than the comparison’s step-by-step Markov chain Monte Carlo (MCMC) run. The work develops Schrödinger-bridge samplers for problems where an invalid state is outside the permitted space, rather than merely unlikely.
Sampling means generating many states from a target probability distribution. In this work, the tests include spheres, orthogonality-constrained matrices and rotations, all treated as structured state spaces. The sampler’s drift and random motion act through the tangent bundle—the directions that stay within the permitted space—so feasibility is built into the state space.
A geometric route through the constraints
The mathematical core frames sampling as an intrinsic manifold Schrödinger-bridge—an optimal way to steer noisy probability flows—and stochastic-optimal-control problem. Under its stated regularity assumptions, Theorem 1 claims a unique bridge solution and an optimal tangent control given by the manifold gradient of a forward potential, scaled by the noise schedule. Put more simply, the theory describes how to steer noisy motion while keeping it on the allowed space.
The practical R–ASBS version approximates several ideal ingredients. It replaces exact manifold heat-kernel scores and reference bridges with short-time heat-kernel and geodesic approximations, and uses Levi–Civita parallel transport to move directions along the geometry. Extended R–ASBS is meant for more general embedded manifolds, using nearest-point retraction, projection-as-transport and a projected-chord corrector when those geometric primitives are unavailable.
The clearest test came on a sphere
The sharpest numerical contrast came from an analytical distribution on a sphere. In the reported MCMC run, 86% of particles became trapped in one energy basin. R–ASBS placed 43.8% in the northern hemisphere, close to the theoretical ratio of 0.5. The result is specific to that simulation and does not establish that R–ASBS is better than MCMC for all constrained targets.
The spherical work also included models of cosmic rays and earthquakes. The study identifies the cosmic-ray dataset as SD1500 and the earthquake set as worldwide events of magnitude M ≥ 6.5 from 1526 to 2026. Its figures report qualitatively that generated distributions followed the test data, recovered sparse Zenithal clustering and accumulated near real earthquake data.
Further checks moved beyond ordinary surfaces
Another test moved to the Stiefel manifold, where the states are orthogonal matrices. It averaged energy over 5,000 such matrices. The reported figure was said to confirm the stated temperature-limit properties, while expected energy had a non-positive derivative with respect to β.
The extended method was also placed in a closed-loop robotics setup involving a planar manipulator with 10 revolute joints. Obstacles were positioned at (4, 1.5) and (6, −1), and the weak prior had a weight of λ = 0.05. The robotics outcome is described qualitatively rather than with a numerical obstacle-avoidance success rate.
Rotation estimates held up in a stress test
The most detailed results concerned a robust Wahba rotation problem, testing whether the method could recover a rotation while separating useful measurements from outliers. Across outlier ratios of 25%, 50%, 75%, 85%, 90% and 95%, extended R–ASBS reported TLS gaps near zero: 0.143 ± 0.146% at 25% and −0.058 ± 0.029% at 95%. Rotation errors across the tests ranged from 0.733° ± 0.258° to 2.245° ± 1.111°.
Active true inliers ranged from 96.00 ± 2.83% to 98.27 ± 0.76%, while clipped true outliers ranged from 97.73 ± 0.40% to 98.72 ± 0.66%. The values are means with standard deviations over five independent repeats, showing how the method behaved in the tested rotation simulations rather than establishing a universal result across competing samplers.
The practical gap remains open
That distinction matters because the practical samplers are built from approximations to heat-kernel, bridge, transport and denoising quantities. The paper states that these approximations do not have an exact convergence bound or proof. The theorem’s uniqueness result therefore belongs to the idealized construction under its assumptions; the numerical results do not by themselves establish exact convergence of the implemented samplers.
The evidence remains mathematical and numerical, covering spherical, Stiefel, closed-loop kinematic and quaternion-based rotation problems. Several outcomes are qualitative, and the tests leave open how step size, noise schedule, geometric approximation and network training affect target accuracy. Broader comparisons with constrained samplers on other manifolds and larger real datasets are still needed to judge how far the approach generalizes.
Paper data and sources
Original title: Hard-Constrained Sampling on Embedded Riemannian Manifolds via Adjoint Schrödinger Bridges
Authors: Mattia Mosso, Jaemoo Choi, Heng Yang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text