Abstract
Traditional Markov Chain Monte Carlo sampling methods often struggle withsharp curvatures, intricate geometries, and multimodal distributions. Slicesampling can resolve local exploration inefficiency issues, and Riemanniangeometries help with sharp curvatures. Recent extensions enable slice samplingon Riemannian manifolds, but they are restricted to cases where geodesics areavailable in a closed form. We propose a method that generalizes Hit-and-Runslice sampling to more general geometries tailored to the target distribution,by approximating geodesics as solutions to differential equations. Our approachenables the exploration of the regions with strong curvature and rapidtransitions between modes in multimodal distributions. We demonstrate theadvantages of the approach over challenging sampling problems.