Unbalanced Optimal Transport for Incomplete Data

This direction is merged into another direction: unbalanced optimal transport. Its last version stays readable and citable here.

Version 1 of 1157 papers175 conceptsfrozen 2026-10-04

Optimal Transport (OT) is often described through the metaphor of "moving dirt." If you have a pile of earth (a source distribution) and a hole of the same volume (a target distribution), OT finds the most efficient way to move the dirt to fill the hole. This framework has become a cornerstone of machine learning, powering everything from generative models to domain adaptation.

However, the classic formulation has a rigid "weight limit": the total mass of the source and target must be exactly equal. In the real world, this is rarely the case. As noted by Ryoma Sato and colleagues in Fast Unbalanced Optimal Transport on a Tree, this mass conservation constraint is a major bottleneck when dealing with noise, outliers, or datasets of different sizes.

The Problem with Mass Conservation

Standard Optimal Transport requires every single unit of mass in the source to be mapped to the target. If you are comparing a small template of a 3D shape to a large, noisy scene, standard OT will force the template to "stretch" to cover the entire scene, or force the noise in the scene to be mapped back to the template. This leads to distorted alignments that do not reflect the underlying structural similarity.

This issue appears across diverse fields:

Relaxing the Constraints

Unbalanced Optimal Transport (UOT) solves this by relaxing the requirement for strict mass conservation. Instead of forcing the marginals of the transport plan to match the source and target distributions exactly, UOT allows them to deviate.

Mathematically, this is achieved by replacing the hard constraints with "soft" penalty terms, typically using Csiszár divergences like the Kullback-Leibler (KL) divergence. In this regime, the optimization problem doesn't just minimize the cost of moving mass; it also decides whether it is "cheaper" to transport mass or to simply create or destroy it at a certain cost.

Hicham Janati, Boris Muzellec, and Marco Cuturi provided a major theoretical breakthrough in Entropic Optimal Transport between Unbalanced Gaussian Measures has a Closed Form, deriving the first analytical solutions for these problems. This "ground truth" allows researchers to use UOT as a differentiable, robust alternative to the classic Wasserstein distance.

Partial and Gromov-Wasserstein Variants

A specific flavor of this research is Partial Optimal Transport (POT). While UOT uses soft penalties, POT typically seeks to transport a fixed amount of mass (less than the total) while ignoring the rest. Anh Duc Nguyen and Tuan Dung Nguyen have shown in On Partial Optimal Transport that this is essential for tasks like point cloud registration, where only a subset of points should be matched.

When the datasets live in different metric spaces (e.g., comparing a graph to a 3D point cloud), researchers use the Unbalanced Gromov-Wasserstein (UGW) distance. Thibault Séjourné and François‐Xavier Vialard introduced this in The Unbalanced Gromov Wasserstein Distance, enabling isometry-invariant matching that is robust to outliers and missing data.

Scaling to Large Data

The primary challenge for UOT has been computational complexity. Standard solvers often scale quadratically with the number of data points. To address this, several "linearization" and "slicing" techniques have emerged:

  1. Linear Embeddings: Yikun Bai and Soheil Kolouri developed Linear Optimal Partial Transport (LOPT), which embeds unbalanced datasets into a Euclidean space. This reduces the cost of comparing $K$ objects from $O(K^2)$ to $O(K)$.
  2. Sliced UOT: By projecting high-dimensional distributions onto 1D lines, researchers can solve the transport problem much faster. Sliced Optimal Partial Transport (SOPT) leverages this to handle noisy point clouds with significant speedups.
  3. Neural Solvers: For continuous distributions, Alexander Korotin and Arip Asadulaev introduced U-LightOT, a lightweight solver that uses Gaussian Mixture Models to represent transport plans without the need for complex GAN-style optimization.

Why It Works: The Intuition of Robustness

The reason Unbalanced OT works so well for real-world data is its intrinsic denoising property. In a balanced OT problem, an outlier in the source must be transported to something in the target, no matter how far away it is. This "long-distance" transport incurs a massive cost that dominates the optimization.

In the unbalanced setting, the model has a choice: it can pay the transport cost, or it can pay a "destruction" penalty to ignore the outlier. By setting the penalty appropriately, the model naturally focuses on the high-density regions that match well, effectively filtering out noise and focusing on the shared structure between two disparate datasets.

Papers in this version

Showing 60 of 157 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unbalanced_optimal_transport_for_incompl_032ab50d,
  title = {Unbalanced Optimal Transport for Incomplete Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unbalanced-optimal-transport-for-incomplete-data/rdr_032ab50d7b1e466da1108e08c922f8d6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_unbalanced_optimal_transport_for_incompl_032ab50d_v1,
  title = {Unbalanced Optimal Transport for Incomplete Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unbalanced-optimal-transport-for-incomplete-data/rdr_032ab50d7b1e466da1108e08c922f8d6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}