Geometric Distributional Learning via Optimal Transport

This direction is split into other directions: geometry-agnostic distribution comparison. Its last version stays readable and citable here.

Version 1 of 1162 papers165 conceptsfrozen 2026-10-04

In traditional machine learning, we often compare probability distributions using metrics like the Kullback-Leibler (KL) divergence or Total Variation (TV) distance. While mathematically convenient, these measures are "geometrically blind." They treat distributions as abstract sets of points, caring only about whether probability mass overlaps, but ignoring where that mass is located in the underlying space.

A growing body of research suggests that incorporating the geometric structure of the sample space—primarily through Optimal Transport (OT) and the Wasserstein distance—leads to more robust, stable, and expressive models.

The Intuition of Moving Mass

The core limitation of traditional metrics is most apparent when two distributions have non-overlapping supports. If a model predicts a distribution that is slightly shifted from the ground truth, KL divergence will often return an uninformative infinite value, and TV distance will stay at its maximum. Neither metric can tell the model that it is "close" to the target.

As noted by Ievgen Redko, Amaury Habrard, and Marc Sebban in Theoretical Analysis of Domain Adaptation with Optimal Transport, traditional measures fail to account for the manifold structure of data. In contrast, the Wasserstein distance (or Earth Mover's Distance) measures the minimum "work" required to transform one distribution into another by moving mass across the ground metric of the space. This "geometric awareness" allows the metric to remain informative even when distributions are completely disjoint.

Bridging Geometry and Reinforcement Learning

This geometric sensitivity is particularly transformative in Reinforcement Learning (RL). In model-based RL, an agent must learn the dynamics of its environment. Standard probability metrics fail to distinguish between a prediction error that lands an agent in a "similar" state versus a "fundamentally different" one.

Kavosh Asadi and Michael L. Littman demonstrated in Equivalence Between Wasserstein and Value-Aware Model-based Reinforcement Learning that minimizing the Wasserstein distance between true and learned dynamics is mathematically equivalent to optimizing for value-aware planning. By using a metric that respects the distance between states, the agent learns a model that is more useful for achieving its goals.

Similarly, in distributional RL, agents learn the entire distribution of future rewards rather than just the mean. Ke Sun and Bei Jiang introduced SinkhornDRL in Distributional Reinforcement Learning with Regularized Wasserstein Loss, which uses entropic-regularized Wasserstein distances to avoid the "non-crossing" errors common in quantile-based methods, allowing for more stable learning of complex reward structures.

Overcoming the Computational Barrier

Historically, the primary drawback of Optimal Transport has been its cubic computational complexity relative to the number of samples. Researchers have developed several approximations to make these geometric metrics practical:

  1. Entropic Regularization (Sinkhorn Divergence): By adding an entropy term to the OT problem, the computation becomes a series of simple matrix multiplications. Aude Genevay and Marco Cuturi showed in Sample Complexity of Sinkhorn divergences that this approach bridges the gap between OT and kernel methods like Maximum Mean Discrepancy (MMD).
  2. Sliced Wasserstein (SW) Distances: This method projects high-dimensional distributions onto 1D lines, where the Wasserstein distance has a closed-form solution. Soheil Kolouri and Gustavo K. Rohde pioneered Sliced Wasserstein Kernels for Probability Distributions, enabling the use of OT in large-scale computer vision tasks.
  3. Learned Embeddings: In Learning Wasserstein Embeddings, researchers proposed using Siamese networks to map distributions into a Euclidean space where $L_2$ distances approximate the Wasserstein metric, achieving massive speedups for retrieval tasks.

Robustness and Generalization Bounds

The geometric nature of the Wasserstein distance also provides a more natural framework for analyzing model stability and robustness. Traditional information-theoretic bounds on generalization error can be loose because they ignore the physical similarity between different hypotheses.

Borja Rodríguez Gálvez and Mikael Skoglund proved in Tighter Expected Generalization Error Bounds via Wasserstein Distance that OT-based measures provide tighter bounds by accounting for the geometry of the parameter space. Furthermore, in robust optimization, Liyan Xie and Yao Xie demonstrated that using Wasserstein Uncertainty Sets allows for hypothesis testing that is resilient to outliers and "holes" in training data, as the metric naturally understands that a point slightly outside the training set is still "near" the observed distribution.

Papers in this version

Showing 60 of 162 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_geometric_distributional_learning_via_op_5701b2d2,
  title = {Geometric Distributional Learning via Optimal Transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-distributional-learning-via-optimal-transport/rdr_5701b2d21b6048feb48e142af4b2ae70},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_geometric_distributional_learning_via_op_5701b2d2_v1,
  title = {Geometric Distributional Learning via Optimal Transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-distributional-learning-via-optimal-transport/rdr_5701b2d21b6048feb48e142af4b2ae70/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}