Multivariate Nonparametric Testing in High-Dimensional Data

The rapid proliferation of high-dimensional data across genomics, neuroimaging and environmental studies has posed significant challenges to classical parametric inference. Multivariate nonparametric testing offers a flexible framework for assessing distributional hypotheses when the number of variables approaches or exceeds sample size, alleviating assumptions of Gaussianity and linearity. Modern strategies harness kernel embeddings, distance-based statistics and graph-theoretic constructs to quantify discrepancies between complex multivariate distributions. Central to these approaches are integral probability metrics such as the maximum mean discrepancy and energy distance, which facilitate consistent and computationally tractable tests without explicit density estimation. Rank-based and graph-based methods, including nearest-neighbour graphs and random geometric graphs, provide robust alternatives in the presence of outliers or manifold-supported data. Key considerations in high dimensions include control of type I error under non-asymptotic regimes, power characterisation against “fair” alternatives and efficient approximation of null distributions via resampling or spectral decompositions. Advances in methodology have enabled applications to high-throughput omics, image analysis and network inference, demonstrating resilience to the curse of dimensionality and offering interpretable measures of distributional divergence with global significance for scientific discovery.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent work has illuminated the spectral and statistical factors underlying the power decline of kernel and distance-based nonparametric tests in high dimensions. By distinguishing the challenges of statistic estimation from hypothesis testing, researchers have formalised “fair” alternatives and shown that power may drop polynomially with dimension, while offering theoretical guidance on kernel bandwidth selection. In parallel, novel k-sample procedures based on maximum mean discrepancy have been proposed for testing equality of multiple high-dimensional distributions. These methods derive asymptotic null and alternative distributions, introduce straightforward approaches for accurate null approximation and demonstrate efficacy through simulation studies and real-data analyses. More recently, energy-distance frameworks have been extended to data residing on non-Euclidean manifolds. Permutation-based calibration of two-sample tests on curved supports has achieved strong power against manifold-structured alternatives, broadening applicability to shape analysis, directional data and network topology comparisons.

Multivariate Nonparametric Testing in High-Dimensional Data publication trend

The graph below shows the total number of articles in multivariate nonparametric testing in high-dimensional data across all publications each year (not limited to Nature Index journals).

Technical terms

Kernel-based test: A nonparametric method that measures differences between distributions by embedding data into a feature space via a kernel function.

Maximum mean discrepancy (MMD): A metric that quantifies the distance between probability distributions as the difference of their mean embeddings in a reproducing kernel Hilbert space.

Energy distance: A statistic that captures distributional divergence through expectations of pairwise Euclidean distances between observations.

Reproducing kernel Hilbert space (RKHS): A Hilbert space of functions in which evaluation at any point is expressed as an inner product with a kernel.

Curse of dimensionality: The phenomenon whereby statistical and computational complexity escalates rapidly as the number of variables increases.

References

1. On the empirical estimation of integral probability metrics. Electronic Journal of Statistics (2012).
2. On the Decreasing Power of Kernel and Distance Based Nonparametric Hypothesis Tests in High Dimensions. Proceedings of the AAAI Conference on Artificial Intelligence (2015).
3. Testing Equality of Several Distributions at High Dimensions: A Maximum-Mean-Discrepancy-Based Approach. Mathematics (2023).
4. Manifold energy two-sample test. Electronic Journal of Statistics (2024).
5. Testing multivariate uniformity based on random geometric graphs. Electronic Journal of Statistics (2020).