Elliptic Sampling: Optimal Interpolant Schedules for Gaussians and Beyond
Zoloflow achieves a provably optimal score-query bound for Gaussian targets within the algorithm class studied.
The method jointly designs the interpolant schedule and numerical steps used to sample from a target distribution. Its elliptic schedules account explicitly for the target’s condition number and reduce to the classical trigonometric schedule as that number approaches one. For Gaussian targets, the authors prove minimax optimality among oblivious score-span algorithms using exact scores; they also establish a sampling-error bound for log-concave targets. Experiments extend to non-log-concave distributions and image-generation datasets, though the abstract reports no specific performance gains.
ArXiv · AI/CL/LG's note
The method jointly designs the interpolant schedule and numerical steps used to sample from a target distribution. Its elliptic schedules account explicitly for the target’s condition number and reduce to the classical trigonometric schedule as that number approaches one. For Gaussian targets, the authors prove minimax optimality among oblivious score-span algorithms using exact scores; they also establish a sampling-error bound for log-concave targets. Experiments extend to non-log-concave distributions and image-generation datasets, though the abstract reports no specific performance gains.
ArXiv · AI/CL/LG's note
score 5