Multivariate Distribution-Free Nonparametric Testing: Generalizing Wilcoxon’s Tests via Optimal Transport

Zhen Huanglabel=e1][email protected], Bodhisattva Senlabel=e2][email protected] · 2025

Read Multivariate Distribution-Free Nonparametric Testing: Generalizing Wilcoxon’s Tests via Optimal Transport by Zhen Huanglabel=e1][email protected], Bodhisattva Senlabel=e2][email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-12236.txt.

Introduction

Zhen Huang and Bodhisattva Sen reviewed in 2025 how optimal transport extends classic rank tests to data with many variables at once. Wilcoxon's tests rank one-dimensional samples, but modern data come as vectors. Transport-based ranks and signs keep the prized traits of the old tests: no distribution assumptions, robustness and strong efficiency. The review compares their power with the standard Hotelling test using relative efficiency bounds.

The bounds reveal two classic phenomena, named for Hodges-Lehmann and Chernoff-Savage, inside multivariate location testing. The transport tests hold up impressively well against the normal-theory champion. The authors also show how the ranks plug into kernel methods, giving consistent tests that need no distribution assumptions. New results cover testing symmetry in many dimensions. The paper serves as a map of a fast-growing corner of statistics.

Catalog record

CREATORZhen Huanglabel=e1][email protected], Bodhisattva Senlabel=e2][email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-12236

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Multivariate Distribution-Free Nonparametric Testing: Generalizing Wilcoxon’s Tests via Optimal Transport Zhen Huanglabel=e1][email protected], Bodhisattva Senlabel=e2][email protected] Abstract This paper reviews recent advancements in the application of optimal transport (OT) to multivariate distribution-free nonparametric testing . Inspired by classical rank-based methods, such as Wilcoxon’s rank-sum and signed-rank tests, we explore how OT-based ranks and signs generalize these concepts to multivariate settings, while preserving key properties, including distribution-freeness, robustness, and efficiency. Using the framework of asymptotic relative efficiency (ARE), we compare the power of the proposed (generalized Wilcoxon) tests against the Hotelling’s T 2 T^{2} test. The ARE lower bounds reveal the Hodges-Lehmann [ 47 ] and Chernoff-Savage [ 13 ] phenomena in the context of multivariate location testing, underscoring the high power and efficiency of the proposed methods. We also demonstrate how OT-based ranks and signs can be seamlessly integrated with more modern techniques, such as kernel methods, to develop universally consistent, distribution-free tests. Additionally, we present novel results on the construction of consistent and distribution-free kernel-based tests for multivariate symmetry, leveraging OT-based ranks and signs. keywords † † email: [ † † email: [ and 1 Introduction Nonparametric testing is a class of statistical methods that do not rely on strong parametric assumptions. Frank Wilcoxon’s 1945 paper [ 94 ] was a significant breakthrough in the history of (classical) nonparametric statistics as it introduced two widely used nonparametric tests: the Wilcoxon’s rank-sum test and the Wilcoxon’s signed-rank test . These two tests utilize the ranks of the (univariate) data to test whether two samples come from the same distribution, and whether one sample has center 0, respectively. Moreover, both these rank-based tests are distribution-free , i.e., the distribution of the test statistics, under the null hypothesis, does not depend on the data generating distribution — a consequence of the fact that for exchangeable data, the vector of ranks follows the uniform distribution over all n! permutations of \{1,\ldots,n\} . Apart from their simplicity, robustness and distribution-freeness, a striking feature of these tests that makes them a strong alternative to the usual two-sample t -test and paired t -test are the efficiency results of Hodges-Lehmann [ 47 ] and Chernoff-Savage [ 13 ] . These results state that Wilcoxon’s tests (with Gaussian score) are never worse than the t -tests in the sense of asymptotic relative efficiency (ARE), against location shift alternatives [ 13 ] . Since Wilcoxon’s seminal work [ 94 ] , the use of signs and ranks to develop distribution-free tests has become widespread in one-dimensional nonparametric statistics [ 58 , 31 , 45 ] . These methods have been extensively applied in various fields, including time series analysis [ 43 , 42 , 46 ] , survival analysis [ 26 , 71 ] , finance [ 3 ] , social science [ 15 ] and genomics [ 72 ] . In multi-dimensional Euclidean space, the absence of a canonical ordering poses a long-standing challenge in defining multivariate generalizations of these tests which preserve the desirable statistical properties from the one-dimensional case. Since these tests rely on ranks, a more fundamental question that arises is: “How can one define multivariate ranks that lead to distribution-free testing procedures?”. Existing extensions of multivariate ranks — such as component-wise ranks [ 7 ] , spatial ranks [ 12 , 56 , 63 ] , depth-based ranks [ 61 , 62 , 96 ] , and Mahalanobis ranks [ 39 , 40 ] (see [ 76 ] and [ 33 ] for recent surveys) — and the corresponding rank-based tests do not achieve exact finite-sample distribution-freeness. A major breakthrough in this regard was recently made in Hallin et al. [ 33 ] , where the authors introduced a notion of multivariate ranks based on the theory of optimal transport (OT) (also see [ 14 ] ). The methodology was later extended by [ 18 , 51 ] . The class of OT-based signs and ranks retains many of the desirable properties of their one-dimensional counterparts. Notably, they are distribution-free in finite samples, and can be used to design distribution-free multivariate analogues of Wilcoxon’s tests. These proposed tests [ 16 , 51 ] also exhibit the Hodges-Lehmann and Chernoff-Savage phenomena, ensuring that their power (in the framework of ARE) is at least as high as the two-sample and one-sample Hotelling’s T^{2} tests (the multivariate analogues of the t -tests) against many subclasses of location shift alternatives. These developments have sparked a surge of research in this area. OT-based signs and ranks have recently been applied to various statistical problems, including the two-sample problem [ 10 , 18 , 16 , 27 ] , independence testing [ 18 , 79 , 27 , 80 , 78 ] , and testing for multivariate symmetry [ 51 ] . They have also been utilized in multiple-output linear regression and MANOVA [ 34 ] , multiple-output quantile regression [ 19 ] , R-estimation for VARMA models [ 35 ] , and robust testing procedures for VAR models [ 36 ] . For a recent review on the applications of OT in statistical problems, see [ 32 ] . In this paper, we provide a brief review of OT-based (multivariate) signs and ranks and associated statistical methodologies. As illustrative applications, we focus on extending Wilcoxon’s distribution-free tests to the multivariate setting. Using the framework of ARE, we compare the power of the proposed (generalized Wilcoxon) tests against the Hotelling’s T^{2} test. The ARE lower bounds reveal the Hodges-Lehmann [ 47 ] and Chernoff-Savage [ 13 ] phenomena in the context of multivariate location testing, underscoring the high power and efficiency of the proposed methods. We further develop universally consistent two-sample and one-sample tests by integrating OT-based ideas with kernel methods [ 29 , 85 ] . We also discuss a general framework for testing multivariate symmetry using OT-based (multivariate) signs and signed-ranks [ 51 ] . Our presentation includes novel results (see Theorems 4.3 , 4.4 and 4.5 ) not previously published — specifically, in Sections 4.2 and 4.3 , we introduce and study a novel kernel method for testing multivariate symmetry, based on maximum mean discrepancy [ 29 ] and the OT-based signed-ranks. To the best of our knowledge, this is the only test for multivariate symmetry that satisfies three key properties: (i) distribution-freeness, (ii) universal consistency, and (iii) applicability to a broad class of multivariate symmetries. Our approach provides a unified framework for constructing distribution-free tests using OT-based ranks that extend beyond one- and two-sample testing. More broadly, it offers a general mechanism for transforming any test statistic into a distribution-free version by substituting the original observations with OT-based ranks (see Remark 3.1 ). The paper is organized as follows: Section 2 introduces the fundamental concepts of OT. In Section 3 , we review OT-based ranks and demonstrate their application in extending Wilcoxon’s rank-sum test, as well as in designing a universally consistent test for the equality of two distributions. Section 4 extends this framework to define OT-based (multivariate) signs and signed-ranks and develops the generalized multivariate Wilcoxon’s signed-rank test [ 51 ] . Our novel methodology and results on testing multivariate symmetry are presented in Sections 4.2 and 4.3 . Further applications of OT-based signs and ranks

Read the complete text