In one dimension, statistics is built on sorting. If you have a set of scalar numbers, you can easily arrange them from smallest to largest. This simple operation gives rise to foundational statistical tools: the median (the middle number), quantiles (the values dividing the data into percentiles), and rank-based non-parametric tests that remain robust regardless of the underlying distribution.
However, when moving to multivariate data ($d > 1$), this total ordering completely vanishes. Consider two vectors: $x = [1, 5]$ and $y = [5, 1]$. Which one is "larger"? Which one comes first? Without a canonical ordering, extending univariate concepts like quantiles, ranks, and signs to higher dimensions becomes fundamentally ambiguous. As noted by Carlier et al. and Galichon et al. in their work on vector quantile regression (Vector quantile regression: An optimal transport approach; Vector quantile regression beyond the specified case), this limitation has historically prevented researchers from directly modeling how the joint tails of multi-variable systems respond to covariates or external factors.
Geometric Centrality via Statistical Depth
One intuitive way to recover a sense of order in a multi-dimensional cloud of data is to replace the concept of "left-to-right" sorting with an "inside-out" ordering. This is the core philosophy behind statistical depth, which assigns a scalar value to every point indicating how central it is relative to the overall distribution.
A classic example is halfspace depth (or Tukey depth). Dyckerhoff developed exact, efficient algorithms to compute halfspace depth in multi-dimensional spaces (Exact computation of the halfspace depth), allowing researchers to find a robust multivariate median even when the data contains complex geometries or ties.
By ranking points from the most central to the most extreme, statistical depth provides a powerful alternative to traditional sorting. For instance, Ley et al. extended univariate runs tests to bivariate settings to test for central symmetry by ordering data points based on their statistical depth (Depth-based Runs Tests for Bivariate Central Symmetry). In modern machine learning, Seegmiller et al. applied statistical depth to rank and characterize high-dimensional transformer-based text embeddings (Statistical Depth for Ranking and Characterizing Transformer-Based Text Embeddings), identifying typical examples versus outliers within a corpus to improve in-context learning.
The Optimal Transport Solution: Center-Outward Ranks
While statistical depth provides an excellent measure of centrality, it does not naturally preserve all the structural properties of univariate ranks and signs—such as distribution-freeness and the ability to define a unique quantile function. To bridge this gap, a groundbreaking framework emerged by combining multivariate statistics with Optimal Transport (OT) theory.
The key insight, introduced by Galichon et al., is the concept of Monge-Kantorovich depth, quantiles, ranks, and signs (Monge–Kantorovich depth, quantiles, ranks and signs). Instead of trying to sort the multivariate data directly, the authors propose mapping the target distribution to a simple, well-understood reference distribution—typically a uniform distribution on a unit ball.
This mapping (realized as a Brenier map, the gradient of a convex function) acts as a canonical transformation. A data point's center-outward rank is defined by its distance from the center of the reference ball, while its sign is its direction (a unit vector). This elegant formulation ensures that the multivariate ranks behave exactly like their univariate counterparts.
Using these center-outward ranks, Shi et al. developed the first distribution-free and consistent statistical tests for independence between random vectors of arbitrary dimensions (Distribution-free consistent independence tests via center-outward ranks and signs; Distribution-free tests of multivariate independence based on center-outward quadrant, Spearman, Kendall, and van der Waerden statistics). Because the ranks are mapped to a distribution-free reference space, these tests eliminate the need for computationally expensive permutation methods while remaining highly sensitive to non-linear dependencies.
Vector Quantile Regression and Generative Modeling
In univariate settings, quantile regression allows us to model how different parts of a distribution (like the 90th percentile) shift with covariates. In multiple dimensions, however, standard quantile regression suffers from the "quantile crossing" problem, where estimated quantile curves paradoxically intersect.
By leveraging the optimal transport framework, Vector Quantile Regression (VQR) solves this by defining multivariate quantiles as gradients of convex potentials (Vector quantile regression: An optimal transport approach). Because the gradient of a convex function is naturally monotone, the resulting quantile functions are mathematically guaranteed not to cross.
To make this framework practical for complex, high-dimensional real-world data, Rosenberg et al. introduced Fast Nonlinear Vector Quantile Regression (Fast Nonlinear Vector Quantile Regression), utilizing a GPU-accelerated solver to capture highly non-linear relationships. Similarly, Sun et al. extended this to deep generative modeling via Conditional Generative Quantile Networks (Conditional Generative Quantile Networks via Optimal Transport and Convex Potentials), optimizing a dual formulation to learn non-crossing conditional distributions for high-dimensional generation.
High-Dimensional Conformal Prediction
Another critical domain impacted by multivariate ordering ambiguity is Conformal Prediction (CP), the gold standard for distribution-free uncertainty quantification. Standard CP relies on sorting non-conformity scores to determine a threshold for a prediction interval. When the output is a vector of multiple correlated variables, sorting becomes impossible, forcing traditional methods to rely on loose bounding boxes or ellipses.
As highlighted by Klein et al., the lack of a total order makes it difficult to define principled nested sets for multivariate uncertainty quantification (Multivariate Conformal Prediction using Optimal Transport).
To resolve this, Thurin introduced an optimal transport-based conformal prediction framework (Optimal transport-based conformal prediction) that utilizes Monge-Kantorovich ranks to construct flexible, non-convex prediction regions that adapt perfectly to the true geometry of the data. Furthermore, Kondratyev et al. combined Neural Optimal Transport with input-convex neural networks (Neural Optimal Transport Meets Multivariate Conformal Prediction) to learn tight, data-adaptive predictive regions with exact coverage guarantees.
Key Foundational Work
Univariate and multivariate symmetry: statistical inference and distributional aspects (2010) — contributed 1 concept elements to this direction.
Discussion (2010) — contributed 1 concept elements to this direction.
A Color Morphology based on Pareto-dominance Relation and Hypervolume Measure (2008) — contributed 1 concept elements to this direction.