In univariate statistics, ordering data is trivial. Given a set of numbers, we can easily sort them from smallest to largest. This simple act of sorting underpins many fundamental statistical tools: quantiles tell us where a point sits relative to the rest of the distribution, ranks provide distribution-free non-parametric tests, and signs indicate direction.
However, when moving to multivariate data (vectors in $\mathbb{R}^d$ where $d > 1$), this natural linear ordering vanishes. Is the vector $(1, 5)$ greater or less than $(5, 1)$? Without a total order, classical definitions of quantiles, ranks, and medians break down.
To resolve this ambiguity, a powerful framework has emerged that redefines multivariate sorting through the lens of Optimal Transport (OT). By mapping complex, high-dimensional distributions onto simple, symmetric reference shapes, researchers can establish a rigorous, geometrically sound notion of vector quantiles and "center-outward" ranks.
The Core Intuition: Transporting to a Canonical Reference
Instead of trying to order vectors directly in their native, potentially highly non-convex space, the optimal transport approach maps the data distribution to a well-behaved reference distribution where ordering is mathematically unambiguous.
As introduced by Alfred Galichon et al. in Monge–Kantorovich depth, quantiles, ranks and signs, we can choose a standard reference distribution $\mu_0$, such as a uniform distribution on a unit ball or a unit cube. We then find the unique optimal transport map $T$ that pushes this reference distribution to our target data distribution $\mu$.
Reference Space (Unit Ball) Target Data Space
+---------------+ +---------------+
| . | | * * |
| . o . | ==== Quantile Map (T) => | * X * |
| . | | * * |
+---------------+ +---------------+
Origin (0,0) Median (X)
In this setup:
- The Quantile Map $T$ warps the simple reference space to match the geometry of the empirical data.
- The Rank Map $T^{-1}$ (the inverse map) does the opposite: it takes complex data points and projects them back into the standardized reference space.
Why It Works: Convex Potentials and Monotonicity
Why choose optimal transport maps over any other bijective mapping, such as a standard normalizing flow? The answer lies in a foundational result from optimal transport theory known as Brenier's Theorem.
Brenier's theorem states that under a quadratic cost function, the optimal transport map $T$ pushing a continuous reference distribution to a target distribution is uniquely realized as the gradient of a convex potential function $\psi$:
$$T(u) = \nabla \psi(u)$$
In one dimension, the derivative of a convex function is simply a monotonically increasing function. Thus, expressing multivariate mappings as the gradients of convex potentials directly generalizes the 1D property of monotonicity to higher dimensions.
As demonstrated by Guillaume Carlier et al. in Vector quantile regression: An optimal transport approach, this geometric foundation guarantees that the mapping from a latent rank to a multi-dimensional outcome is unique, stable, and satisfies a multivariate monotonicity property:
$$\langle \nabla \psi(u) - \nabla \psi(v), u - v \rangle \ge 0$$
This property ensures that multivariate quantile contours are non-crossing. In applications like generative modeling and Bayesian inference, this prevents structural collapse. For example, Jesse Sun et al. leverage these convex potentials in Conditional Generative Quantile Networks via Optimal Transport and Convex Potentials to ensure valid conditional distributions, while Jungeum Kim et al. utilize them in Deep Generative Quantile Bayes to construct well-defined, non-overlapping Bayesian credible regions.
Center-Outward Ranks and Signs
When the chosen reference distribution is a uniform unit ball, the inverse transport map $T^{-1}$ provides an elegant geometric decomposition of any data point into a multivariate rank and sign.
When a data point $x$ from a complex distribution is mapped back into the unit ball, its position vector $u = T^{-1}(x)$ yields two intuitive metrics:
- Center-Outward Rank: The magnitude $\|u\| \in [0, 1]$ represents how far the point is from the center of the distribution. Points deep within the data's core map near the origin (the median), while extreme outliers map close to the boundary (magnitude near 1).
- Multivariate Sign: The direction vector $u / \|u\|$ represents the orientation of the data point relative to the distribution's center of mass.
Reference Unit Ball
+-----------+
| ^ | Direction = "Sign"
| | |
| o---> | Distance from center = "Rank"
| |
+-----------+
This elegant formulation enables fully distribution-free non-parametric statistics in arbitrary dimensions. Traditional multivariate tests often require assuming a Gaussian distribution or relying on computationally expensive permutation tests.
By mapping data to a deterministic grid within the unit ball via optimal transport, Hongjian Shi et al. demonstrate in On universally consistent and fully distribution-free rank tests of vector independence and Distribution-free tests of multivariate independence based on center-outward quadrant, Spearman, Kendall, and van der Waerden statistics that researchers can construct universally consistent independence tests. Because the data is transformed into a uniform distribution on a sphere, the resulting tests retain their distribution-free properties regardless of the underlying data's true shape or dimensionality.
Application: Multivariate Conformal Prediction
One of the most immediate practical applications of this framework is in Conformal Prediction (CP), a methodology used to provide rigorous finite-sample coverage guarantees for machine learning models.
Standard conformal prediction relies on ranking scalar calibration scores. When a model outputs multivariate predictions (such as predicting coordinates, bounding boxes, or multi-target regressions), squashing those errors into a single scalar score discards vital correlation structure between the variables.
By applying optimal transport, authors have successfully extended conformal prediction to multi-dimensional outputs. As detailed by Gauthier Thurin et al. in Optimal transport-based conformal prediction and Michal Klein et al. in Multivariate Conformal Prediction using Optimal Transport, mapping multivariate calibration scores to a symmetric reference distribution allows for the construction of flexible, non-convex prediction regions.
Instead of forcing uncertainty sets to be rigid ellipses (as a Mahalanobis distance would), the monotone transport map respects the true geometry of the data. It naturally recovers classical distance-based behavior if the data happens to be normally distributed, but seamlessly adapts to complex, non-Gaussian, or multi-modal distributions while maintaining exact distribution-free coverage guarantees.
Key Foundational Work
Monge–Kantorovich depth, quantiles, ranks and signs (2017) — contributed 4 concept elements to this direction.
Vector quantile regression: An optimal transport approach (2016) — contributed 2 concept elements to this direction.
Vector quantile regression beyond the specified case (2017) — contributed 1 concept elements to this direction.