Problem
Research questions and friction points this paper is trying to address.
Density Ratio Estimation
Displacement Fields
Distribution Shift
Stein Operator
Density Ratio Estimation with Stein Displacement Fields
🤖 AI Summary
This study addresses the limitation of existing methods that decouple density ratio estimation from displacement field estimation, thereby relying on post-hoc transformations and hindering end-to-end joint modeling. To overcome this, we introduce the Stein operator to bridge density ratios and displacement fields for the first time. By parameterizing the displacement field to unify density ratio estimation, our approach integrates statistical characterization and dynamic transport into a single convex optimization problem, complemented by an iterative push-forward/pull-back inference mechanism for end-to-end resolution of distribution shifts. We demonstrate that this method effectively calibrates pre-trained samplers in simulation-based inference and accurately fits transformation models in nonlinear independent component analysis tasks, successfully combining theoretical unification with practical efficacy.
📝 Abstract
Density ratios quantify distribution shift from a probability-mass point of view, whereas displacement fields describe, from a dynamical point of view, how one distribution is transported onto another. Although both offer complementary insights, they are usually estimated separately, and converting one into the other requires post-processing. In this paper, we estimate the density ratio between a target and a base distribution by parametrizing it through a displacement field acting on the base: the log-ratio is modeled as minus the Stein operator of the base applied to the field, up to a normalizing constant. This gives both statistical and dynamical descriptions of the distribution shift through a single convex optimization problem. Iterating this estimate-and-move step gives two inference algorithms: push-forward moves the model and corrects a pretrained sampler without retraining it, whereas pull-back moves the data closer to the base and fits a transformation model one layer at a time. Applications to distribution shift in simulation-based inference and to nonlinear independent component analysis illustrate the benefits and limitations of the approach.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Density Ratio Estimation
Stein Displacement Fields
Convex Optimization
Push-forward and Pull-back Algorithms
Distribution Shift
🔎 Similar Papers
No similar papers found.
