摘要：This paper introduces SING, a framework for learning Markov structures in continuous non-Gaussian distributions. The methodology integrates two core components: (1) a conditional independence score matrix derived from the integrated Hessian of the log-density, which theoretically bounds conditional dependencies under logarithmic Sobolev constraints, and (2) deterministic triangular transport maps that couple non-Gaussian distributions to Gaussian references while preserving sparsity patterns corresponding to graph edges. The framework operates through an iterative algorithm that alternates between estimating sparse transport maps and thresholding the score matrix to recover minimal I-maps, supported by theoretical consistency guarantees. By unifying transport-based density estimation with graphical model learning, SING addresses limitations of parametric assumptions in non-Gaussian settings. This approach maintains computational feasibility through sparse parameterizations, enabling scalable structure recovery where conventional Gaussian or linear methods fail.