Multivariate Quantiles and Ranks via Optimal Transport

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2104 papers112 conceptsfrozen 2026-10-04

In one dimension, statistics is built on the ability to sort. If you have a list of heights, you can easily find the median, the 95th percentile, or the rank of a specific individual. This "total ordering" of the real line allows us to define quantiles, which characterize the shape of a distribution and identify outliers.

However, as soon as we move to two or more dimensions—modeling joint outcomes like blood pressure and cholesterol, or latitude and longitude—this foundation crumbles. As noted by Haoxu Shu and Giseon Heo in Estimation for bivariate quantile varying coefficient model, there is no natural or unique way to order data points in multi-dimensional space. Is the vector $(1, 5)$ "greater" than $(5, 1)$? Without a clear answer, traditional tools like the Kolmogorov-Smirnov test or quantile regression become mathematically ambiguous.

Recent research has converged on a powerful solution to this problem: using Optimal Transport (OT) to "import" an ordering from a simple reference space into complex, high-dimensional data.

The Geometry of Centrality

To restore the concept of a quantile in higher dimensions, researchers like Alfred Galichon and Marc Hallin propose a "center-outward" perspective. Instead of trying to order points from left to right, we order them by how "deep" they are within a distribution.

In Monge–Kantorovich depth, quantiles, ranks and signs, the authors define multivariate quantiles by mapping a reference distribution (typically a uniform distribution on a unit ball) to the target data distribution. This mapping is achieved through a Monge-Kantorovich transport map.

The intuition is as follows:

  1. Reference Space: Imagine a unit ball where every point has a clear "rank" based on its distance from the center and a "direction" based on its angle.
  2. The Map: We find the most efficient way (the "optimal transport") to push the mass of this ball until it matches the shape of our data cloud.
  3. The Result: Every data point in our complex cloud is now associated with a point in the ball. A data point's "multivariate rank" is simply the position of its corresponding point in the reference ball.

This framework, further explored in Uniform Convergence of Transport Maps, ensures that as we collect more data, these empirical ranks converge to a stable, underlying population truth.

Vector Quantile Regression

This transport-based approach allows for the extension of quantile regression to multiple dependent variables. In Vector quantile regression: An optimal transport approach, Guillaume Carlier and Alfred Galichon demonstrate that multivariate quantiles can be viewed as the gradients of convex functions.

By solving an optimal transport problem, researchers can model how an entire vector of outcomes—such as a set of financial risks or physical measurements—responds to a set of covariates. This moves beyond simply predicting the "average" vector, allowing scientists to see how the "tails" or the "shape" of a joint distribution shifts. This is particularly useful in fields like wind power forecasting, where Jens Schreiber and colleagues use Quantile Surfaces to predict non-Gaussian uncertainty regions for autonomous systems.

Distribution-Free Guarantees

One of the most significant impacts of this research is in Conformal Prediction (CP) and hypothesis testing. Standard CP provides statistical guarantees (e.g., "this prediction region contains the truth 95% of the time") by ranking "non-conformity scores." But as Michal Klein and Louis Béthune point out in Multivariate Conformal Prediction using Optimal Transport, ranking vectors is non-trivial.

By using the Monge-Kantorovich framework, researchers can now:

  • Construct Non-Convex Prediction Regions: Unlike older methods that forced uncertainty into simple ellipses, Optimal transport-based conformal prediction by Gauthier Thurin and others allows for regions that adapt to the actual geometry of the data.
  • Test for Independence: Hongjian Shi and Fang Han developed the first distribution-free and consistent independence tests for vectors of arbitrary dimensions. By using center-outward ranks, these tests remain valid regardless of the underlying distribution, eliminating the need for expensive permutation simulations.

Why It Works

The magic of the optimal transport approach lies in its ability to preserve the topological and measure-theoretic properties of the data. While other methods (like marginal quantiles) treat each dimension separately and ignore correlations, OT-based quantiles treat the vector as a single, cohesive object.

As Xiao-chun Meng and James W. Taylor argue in Scores for Multivariate Distributions and Level Sets, this allows us to define and score "level sets"—the multivariate equivalent of quantile points—providing a principled way to evaluate risks and anomalies in high-dimensional systems. By mapping the "messy" geometry of real-world data back to a "clean" reference ball, we regain the power to sort, rank, and bound the unknown.

Papers in this version

Showing 60 of 104 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 76 papers
  2. Version 1 · 2026-10-04 · 104 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_multivariate_ordering_ambiguity_47c221a0,
  title = {multivariate ordering ambiguity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multivariate-ordering-ambiguity/rdr_47c221a0698a4700af1509f0110a1e92},
  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_multivariate_quantiles_and_ranks_via_opt_47c221a0_v1,
  title = {Multivariate Quantiles and Ranks via Optimal Transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multivariate-quantiles-and-ranks-via-optimal-transport/rdr_47c221a0698a4700af1509f0110a1e92/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}