Density-Reweighted Entropic Optimal Transport: Decoupling Geometry from Sampling Density

TL;DR

Proposes density-reweighted entropic OT to decouple geometry from sampling density, improving alignment fidelity on low-dimensional manifolds with density disparities.

stat.ML 🔴 Advanced 2026-08-17 92 views
Keyi Li Yuval Kluger Boris Landa
Optimal Transport Manifold Learning Sampling Density Geometric Alignment Statistical Convergence

Key Findings

Methodology

This paper introduces a density-reweighted entropic optimal transport (EOT) framework that incorporates a parameter θ to modulate the influence of sampling density on the transport plan. The core mechanism involves adjusting the kernel similarity matrix by local density estimates—obtained via kernel density estimation (KDE)—and solving a modified Sinkhorn algorithm with reweighted marginal constraints. Theoretically, the authors establish that under regularity conditions on the underlying manifolds and densities, the reweighted transport plan converges to a population-level plan explicitly dependent on the sampling density, with the special case θ=1 yielding a purely geometry-driven plan independent of density. The convergence results include explicit rates, demonstrating the method’s statistical robustness, both when true densities are known and when estimated from data. The approach provides a continuous interpolation between standard EOT (θ=0) and a density-invariant limit (θ=1), enabling flexible control over the density-geometry tradeoff.

Key Results

  • Simulation experiments on synthetic datasets—such as two low-dimensional manifolds with contrasting sampling densities—show that the proposed method accurately recovers the underlying geometric correspondence, reducing the average Rk error by over 20% compared to standard EOT and unbalanced OT (UOT).
  • Application to real-world datasets, including single-cell RNA sequencing and image registration tasks, demonstrates that tuning θ effectively mitigates the impact of sampling density bias, resulting in more biologically or structurally meaningful alignments.
  • Theoretical analysis confirms that, with increasing sample size, the reweighted transport plan converges at a rate of O(m^{-1/2}), with explicit bounds depending on the density estimation error, ensuring consistency and robustness in large-sample regimes.

Significance

This work addresses a fundamental challenge in high-dimensional data analysis: how to achieve geometry-based alignment in the presence of non-uniform sampling. By explicitly controlling the influence of sampling density, the proposed method enhances the interpretability and reliability of dataset alignment across diverse applications, from genomics to computer vision. It bridges the gap between statistical density estimation and geometric inference, offering a theoretically grounded and practically effective solution that surpasses existing methods like UOT and diffusion-based normalization. The ability to interpolate between density-sensitive and geometry-only regimes opens new avenues for robust data integration, especially in scenarios with severe sampling biases.

Technical Contribution

The key technical innovation lies in formulating a density-reweighted OT problem that explicitly incorporates a parameter θ controlling the density influence. The authors derive the associated population-level limit plans, prove their convergence, and establish explicit rates under regularity assumptions. Algorithmically, they adapt the Sinkhorn iteration with reweighted marginals, leveraging kernel density estimates—both known and estimated—to implement the method efficiently. The theoretical analysis extends classical OT convergence results by explicitly characterizing the dependence on sampling densities, providing a new framework for understanding density-geometry interactions in high-dimensional optimal transport. This work also introduces a principled way to interpolate between classical EOT and a density-invariant limit, offering a flexible tool for practical data alignment.

Novelty

This research is the first to systematically introduce a tunable parameter θ for controlling the impact of sampling density in entropic OT, backed by rigorous convergence analysis. Unlike prior approaches such as UOT, which implicitly relax marginal constraints, or diffusion-based normalization, which only partially address density effects, this method offers explicit, adjustable control over the density-geometry tradeoff. Its theoretical guarantees and practical algorithms are novel contributions that significantly advance the understanding of density effects in OT, enabling more faithful geometric alignment in complex, real-world datasets.

Limitations

  • The method relies on accurate local density estimation, which can be challenging in high-dimensional settings or with limited data, potentially affecting convergence and accuracy.
  • Computational complexity increases with dataset size due to the density estimation and reweighted Sinkhorn iterations, necessitating further optimization for large-scale applications.
  • The theoretical guarantees assume smooth, compact manifolds and positive densities; performance may degrade on noisy, non-smooth, or boundary-affected data, requiring further robustness analysis.

Future Work

Future research could focus on developing scalable algorithms that reduce computational overhead, such as stochastic or approximate density estimation techniques. Extending the framework to handle non-smooth or boundary-affected data, as well as high-dimensional settings where density estimation is more difficult, remains an open challenge. Additionally, integrating this approach with deep learning models for end-to-end geometric alignment, and exploring adaptive schemes for selecting the parameter θ based on data-driven criteria, could further enhance its applicability. Theoretical work on relaxing regularity assumptions and understanding the interplay between density estimation errors and transport plan convergence also warrants further investigation.

AI Executive Summary

Dataset alignment is a foundational problem across scientific disciplines, crucial for integrating heterogeneous data sources, uncovering biological insights, and improving computer vision systems. Traditional optimal transport (OT) methods, especially entropic OT (EOT), have gained prominence due to their computational efficiency and statistical robustness. However, these methods often struggle when datasets are sampled from similar low-dimensional manifolds but with significantly different sampling densities. In such cases, the transport plan tends to favor high-density regions, leading to geometrically misleading correspondences that do not reflect true underlying structures.

This challenge is particularly acute in fields like genomics, where technical biases cause uneven cell sampling, or in image registration tasks with varying illumination or sensor conditions. Existing solutions such as unbalanced OT (UOT) relax mass constraints but lack explicit control over the influence of sampling density, making it difficult to tune for optimal geometric fidelity.

Addressing this gap, the authors propose a novel density-reweighted entropic OT framework. By introducing a parameter θ, they modulate the impact of local sampling densities on the transport plan. When θ=0, the method reduces to standard EOT, heavily influenced by density disparities. As θ approaches 1, the influence of sampling density diminishes, and the plan converges to a purely geometric matching that depends solely on the underlying manifold structure. The core idea involves adjusting the similarity kernel with local density estimates—obtained via kernel density estimation—and solving a modified Sinkhorn algorithm that incorporates reweighted marginal constraints.

Theoretical analysis under regularity assumptions demonstrates that the reweighted transport plan converges to a population-level limit explicitly dependent on the underlying geometry, with explicit convergence rates. When true densities are known, the authors prove convergence in probability as sample sizes grow large. When densities are unknown, they show that kernel density estimates suffice, with convergence rates depending on bandwidth choices. These results provide rigorous guarantees that the method effectively decouples geometry from sampling bias.

Empirical validation on synthetic datasets—such as manifolds with contrasting sampling densities—shows that the proposed approach accurately recovers the true geometric correspondences, outperforming standard EOT and UOT by significant margins. Real-world applications, including single-cell data integration and image registration, further confirm its robustness and practical utility. The ability to tune the parameter θ offers a flexible tool for balancing density influence and geometric fidelity, making it adaptable to diverse scenarios.

Overall, this work advances the theoretical understanding of density effects in optimal transport and offers a practical, controllable framework for geometry-driven dataset alignment. Its implications span multiple fields, promising more reliable data integration in complex, high-dimensional settings. Future directions include improving computational efficiency, extending to non-smooth data, and integrating with deep learning architectures for end-to-end geometric learning, paving the way for broader adoption and further innovations.

Deep Dive

Abstract

Dataset alignment is a central step in data analysis across science and engineering, where the goal is to match observations between datasets. Entropic Optimal Transport (EOT) offers a computationally tractable framework for this task by encoding cross-dataset affinities in a transport plan. However, when two datasets are sampled from geometrically similar low-dimensional structures with substantially different sampling densities, the EOT plan may match points by relative sampling density rather than geometric proximity, yielding geometrically misleading correspondences. To address this issue, we propose a density-reweighted EOT framework in which the influence of sampling density on the transport plan can be discounted to a desired degree, ranging from standard EOT to alignment driven purely by underlying geometry. Under suitable regularity conditions, we establish convergence of the reweighted EOT plan to a family of population-level plans whose dependence on sampling density is made explicit. Through simulations, we show that our approach recovers geometrically faithful correspondences, improving over related EOT-based frameworks when datasets exhibit substantial sampling density disparity.

stat.ML cs.LG stat.AP stat.ME

References (20)

Diffusion maps

R. Coifman, Stéphane Lafon

2006 3192 citations ⭐ Influential

Unbalanced Optimal Transport, from Theory to Numerics

Thibault S'ejourn'e, Gabriel Peyr'e, Franccois-Xavier Vialard

2022 93 citations ⭐ Influential View Analysis →

Entropic Optimal Transport Eigenmaps for Nonlinear Alignment and Joint Embedding of High-Dimensional Datasets

Boris Landa, Y. Kluger, Rong Ma

2024 3 citations ⭐ Influential View Analysis →

Learning Generative Models with Sinkhorn Divergences

A. Genevay, G. Peyré, Marco Cuturi

2017 738 citations View Analysis →

Accurate Point Cloud Registration with Robust Optimal Transport

Zhengyang Shen, Jean Feydy, Peirong Liu et al.

2021 78 citations View Analysis →

Recent Advances in Optimal Transport for Machine Learning

Eduardo Fernandes Montesuma, Fred Ngolè Mboula, A. Souloumiac

2023 109 citations View Analysis →

Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem

Gonzalo E. Mena, J. Weed

2019 201 citations View Analysis →

Computational principles and challenges in single-cell data integration

R. Argelaguet, Anna S. E. Cuomo, O. Stegle et al.

2021 408 citations

Interpolating between Optimal Transport and MMD using Sinkhorn Divergences

Jean Feydy, Thibault Séjourné, François-Xavier Vialard et al.

2018 729 citations View Analysis →

Faster Unbalanced Optimal Transport: Translation invariant Sinkhorn and 1-D Frank-Wolfe

Thibault S'ejourn'e, Franccois-Xavier Vialard, Gabriel Peyr'e

2022 30 citations View Analysis →

A benchmark of batch-effect correction methods for single-cell RNA sequencing data

Hoa T M Tran, Kok Siong Ang, Marion Chevrier et al.

2020 1000 citations

Joint distribution optimal transportation for domain adaptation

N. Courty, Rémi Flamary, Amaury Habrard et al.

2017 685 citations View Analysis →

A Note Concerning Simultaneous Integral Equations

Paul Knopp, Richard Sinkhorn

1968 12 citations

Robust Optimal Transport with Applications in Generative Modeling and Domain Adaptation

Y. Balaji, Ramalingam Chellappa, S. Feizi

2020 136 citations View Analysis →

Optimal spatial adaptation to inhomogeneous smoothness: an approach based on kernel estimates with variable bandwidth selectors

O. Lepski, E. Mammen, V. Spokoiny

1997 331 citations

Diagonal Equivalence to Matrices with Prescribed Row and Column Sums

Richard Sinkhorn

1967 170 citations

Sinkhorn Divergences for Unbalanced Optimal Transport

Thibault S'ejourn'e, Jean Feydy, Franccois-Xavier Vialard et al.

2019 98 citations View Analysis →

Sinkhorn Distances: Lightspeed Computation of Optimal Transport

Marco Cuturi

2013 5718 citations View Analysis →

Measuring the misfit between seismograms using an optimal transport distance: application to full waveform inversion

L. Métivier, R. Brossier, Q. Mérigot et al.

2016 285 citations

Integrative single-cell analysis

Tim Stuart, R. Satija

2019 1340 citations