Title:Manifold-based transformation of the probability distribution for convergence in optimization

Abstract:Reconstructing probability distributions from experimental data is a crucial problem across various fields. An effective approach is to optimize a theoretical or computational model of the distribution under an objective functional that evaluates consistency with the experimental data. However, achieving convergence in optimization remains a challenge. Given the manifold structure of the probability distribution space, we demonstrate that transformation of distribution in optimization should be infinitesimal displacements along exponential geodesics. Our theory was validated through the reconstruction of protein conformational ensembles, showing its broad applicability.