Density Ratio Estimation with Stein Displacement Fields
The paper ties density-ratio estimation and transport dynamics into one convex optimization problem.
Song Liu models the log density ratio as the negative Stein operator of a base distribution applied to a displacement field, plus normalization. The result is meant to describe distribution shift both statistically and dynamically without estimating the two views separately. Iterating the estimate-and-move step yields push-forward and pull-back inference algorithms, including a way to correct a pretrained sampler without retraining it. The paper tests the approach on simulation-based inference shifts and nonlinear independent component analysis, noting both benefits and limits. ArXiv · AI/CL/LG's note
Song Liu models the log density ratio as the negative Stein operator of a base distribution applied to a displacement field, plus normalization. The result is meant to describe distribution shift both statistically and dynamically without estimating the two views separately. Iterating the estimate-and-move step yields push-forward and pull-back inference algorithms, including a way to correct a pretrained sampler without retraining it. The paper tests the approach on simulation-based inference shifts and nonlinear independent component analysis, noting both benefits and limits. ArXiv · AI/CL/LG's note
score 4