Depth-based runs tests for bivariate central symmetry

Rainer Dyckerhoff, Christophe Ley, Davy Paindaveine

Abstract

Testing for symmetry is a fundamental task in multivariate statistics, often serving as a prerequisite for various location tests. While univariate symmetry tests are well-established, extending these methods to higher dimensions introduces significant geometric and theoretical challenges. This paper proposes a novel nonparametric procedure for testing bivariate central symmetry by extending the univariate runs test of McWilliams (1990). The proposed method utilizes statistical depth functions to establish a center-outward ordering and introduces the concept of "simplicial runs" to detect asymmetries. The resulting tests are affine-invariant, robust to heavy-tailed distributions, and asymptotically distribution-free.

Introduction

In the univariate setting, McWilliams (1990) introduced a powerful nonparametric procedure based on runs. His approach reorders observations according to their absolute values—effectively their distance from the origin—and then analyzes the sequence of signs. A lack of symmetry is indicated when the number of runs (consecutive sequences of the same sign) is significantly smaller than expected under the null hypothesis. While this test is universally consistent and robust, it is inherently limited to one-dimensional data.

Testing for bivariate central symmetry—where a distribution $P$ is symmetric about the origin if $P(A) = P(-A)$ for all Borel sets $A$—is more complex. Existing bivariate tests often require stringent assumptions, such as the existence of finite fourth-order moments, which makes them unsuitable for heavy-tailed data like the Cauchy distribution. This paper bridges the gap by adapting the logic of McWilliams to the bivariate plane using the tools of statistical depth and simplicial geometry.

Methodology

The proposed bivariate test involves two primary innovations to translate 1D runs into a 2D context:

1. Center-Outward Ordering via Statistical Depth

In 1D, observations are ordered by their distance from the center (absolute value). In 2D, distance alone is insufficient because it ignores directional information. To create a meaningful non-parametric ordering, the authors employ statistical depth functions (such as Halfspace or Simplicial depth). These functions rank points from the "deepest" (most central) to the "shallowest" (outliers). To ensure the center is precisely the origin under the null hypothesis, the points are ranked based on a symmetrized version of the sample.

2. Simplicial Runs

In the univariate case, a run changes when the sign flips, meaning the origin lies in the interval between two consecutive points. In the bivariate setting, the authors define a run based on whether the origin is contained within a simplex (a triangle) formed by three consecutive points in the depth-ordered sequence.

The test statistic, $R_D^{(n)}$, counts these simplicial runs. If the data is asymmetric, the points will tend to cluster on one side of the origin in the depth-ordered sequence, leading to fewer triangles containing the origin and, consequently, a smaller number of runs.

Theoretical Properties

Affine Invariance

A critical property of the proposed test is affine invariance. Because the test relies on the relative geometry of simplices and statistical depth, the results remain unchanged under linear transformations such as rotation, scaling, or shearing. This ensures the test is robust to the coordinate system used to record the data.

Asymptotic Distribution

The authors establish the asymptotic null distribution of the test statistic. For large samples, the normalized statistic follows a normal distribution: $$n^{-1/2}(4R_D^{(n)} - n - 2) \xrightarrow{d} N(0, 11/3)$$ This result allows for the straightforward calculation of p-values without requiring intensive resampling methods, making the test computationally efficient for inference.

Experimental Results

The authors evaluated the performance of the depth-based runs tests through extensive Monte Carlo experiments, using 3,000 independent random samples of size $n=100$.

  • Robustness to Heavy Tails: Unlike "pseudo-Gaussian" tests that fail when variance is infinite, the proposed tests maintained their power when applied to Cauchy distributions.
  • Outlier Resistance: The tests demonstrated high robustness against contamination. When extreme outliers were added to the dataset, the depth-based approach remained stable, whereas competing parametric-adjacent tests saw a total collapse in performance.
  • Consistency Across Kernels: The tests consistently detected asymmetry across various skewing mechanisms (e.g., Azzalini-type skewing and sinh-arcsinh transformations), regardless of the specific depth function utilized (Halfspace, Simplicial, or Simplicial Volume).

Conclusion and Future Work

The depth-based runs test provides a robust, nonparametric alternative for testing bivariate symmetry. By leveraging the geometric properties of statistical depth and simplices, the authors have successfully extended a classic univariate tool into the multivariate domain.

While the bivariate case is well-handled, extending this framework to $k$-dimensions remains a challenge. In 3D and beyond, maintaining the test's properties requires a more complex ranking of directions or angles. Additionally, while the test is empirically consistent, a formal proof of universal consistency in the bivariate setting remains an open mathematical objective for future research.

Summary updated

BibTeX
@inproceedings{dyckerhoff2014depth,
  author = {Rainer Dyckerhoff and Christophe Ley and Davy Paindaveine},
  title = {{Depth-based runs tests for bivariate central symmetry}},
  booktitle = {Kölner},
  year = {2014},
  doi = {10.48550/arxiv.1401.2377},
}
Figure