Technical Summary: Coverage Correlation Coefficient

Problem Statement

The paper addresses the challenge of detecting complex, potentially nonlinear statistical associations between random variables and vectors. While classical measures like Pearson's correlation capture linear relationships and rank-based measures (Spearman, Kendall) capture monotonic associations, they often fail to detect non-monotonic dependencies. Existing nonparametric approaches, such as distance correlation (dCor), Hilbert–Schmidt Independence Criterion (HSIC), and Chatterjee's correlation, have significant limitations. Many are designed primarily for independence testing without providing a direct measure of dependence magnitude. Others, particularly permutation-based tests, suffer from computational intractability in large-scale settings (e.g., genomics) where accurate asymptotic null distributions are required to handle massive multiple testing burdens. Furthermore, Chatterjee's correlation, while efficient and distribution-free, is inherently asymmetric and designed to detect directional functional dependence ($Y = f(X)$), making it less suitable for detecting symmetric or indirect dependencies where variables share a common latent structure.

Methodology

The authors propose the coverage correlation coefficient ($\kappa_{X,Y}$), a novel nonparametric measure of statistical association. The methodology is grounded in geometric intuition and optimal transport theory:

1. Geometric Intuition: The core idea involves covering the unit hypercube $[0, 1]^{d_X + d_Y}$ with small $\ell_\infty$-balls (cubes) centered at the empirical multivariate ranks of the data points. The statistic quantifies the "uncovered volume" (vacancy) of the space outside these balls.

   • Under independence, the points are uniformly distributed, and the uncovered volume converges to $e^{-1}$.
   • Under singular dependence (where the joint distribution is concentrated on a lower-dimensional manifold), the uncovered volume converges to 1.

2. Monge–Kantorovich Ranks: To generalize this to multivariate vectors ($X \in \mathbb{R}^{d_X}, Y \in \mathbb{R}^{d_Y}$), the authors utilize Monge–Kantorovich ranks. These are defined via optimal transport maps that push the empirical marginal distributions to uniform distributions on the unit cube. Specifically, for reference points $U$ and $V$ in the unit cubes, the ranks $R_i$ are obtained by solving optimal assignment problems minimizing the squared Euclidean distance between data points and reference points.

3. Definition: The empirical coverage correlation is defined as:
   $$\kappa_{X,Y}^n := \frac{V_n - e^{-1}}{1 - e^{-1}}$$
   where $V_n$ is the uncovered volume in the unit cube when balls of radius $\gamma = \frac{1}{2n^{1/d}}$ (where $d = d_X + d_Y$) are centered at the joint ranks. The reference points can be random (uniformly distributed) or deterministic (e.g., a regular grid).

4. Theoretical Framework:

   • Consistency: The statistic consistently estimates an $f$-divergence between the joint distribution $P_{(X,Y)}$ and the product of marginals $P_X \otimes P_Y$, with the generator function $f_{cov}(x) = \frac{e^{-x} - e^{-1}}{1 - e^{-1}}$.
   • Null Distribution: Under the null hypothesis of independence, the statistic admits an analytically tractable asymptotic normal distribution, allowing for the construction of $p$-values without resampling.
   • Derandomization: The paper demonstrates that the statistic can be "derandomized" by using deterministic reference points (e.g., regular grids or low-discrepancy sequences) without losing consistency or asymptotic normality, provided certain regularity conditions on the optimal transport maps are met.

Key Contributions

1. Symmetric Measure of Singular Dependence: Unlike Chatterjee's correlation, the coverage correlation is symmetric and specifically designed to detect singular dependencies (where the joint distribution is singular with respect to the product of marginals), capturing relationships of the form $f(X, Y) = 0$ rather than just $Y = f(X)$.
2. Distribution-Free Asymptotics: The method provides a pivotal null distribution that is independent of the marginal distributions, enabling exact asymptotic $p$-values. This is a significant advantage over kernel-based or distance-based methods that often rely on computationally expensive permutation tests.
3. Computational Efficiency: The authors develop an algorithm with $O(n \log n)$ time complexity for univariate marginals and provide efficient implementations in R and Python packages (covercorr).
4. Theoretical Guarantees: The paper establishes consistency for both univariate and multivariate settings (under specific regularity conditions on the optimal transport maps and dimensions) and proves a central limit theorem for the null distribution.
5. Robustness via Bounded Generator: The underlying $f$-divergence uses a bounded generator function, offering robustness against extreme density ratios compared to unbounded divergences like Kullback–Leibler (mutual information).

Results

• Simulation Studies: The paper evaluates the power of the coverage correlation against Chatterjee's correlation, distance correlation, HSIC, KMAc, and USP across various data-generating mechanisms (sinusoidal, zigzag, circle, spiral, Lissajous, and local dependencies).
  • In scenarios involving directional functional dependence (sinusoidal, zigzag), Chatterjee's correlation performs well.
  • In scenarios involving symmetric or indirect dependencies (circle, spiral, Lissajous, where $X$ and $Y$ are functions of a latent variable $U$), the coverage correlation demonstrates superior power compared to Chatterjee's correlation and KMAc.
  • The method maintains well-calibrated Type I error rates, particularly for large sample sizes.
• Real Data Applications:
  • Menstrual Hormone Data: Applied to four reproductive hormones, the coverage correlation was the only method to detect statistically significant dependence for every pair of hormones at the 5% level, aligning with known biological feedback systems.
  • Single-Cell RNA Sequencing: In a dataset of 9,369 CD8+ T cells and 1,000 genes, the coverage correlation identified 54 significant gene pairs that were missed by Pearson, Spearman, and Chatterjee's correlations. Visual inspection of these pairs revealed L-shaped relationships indicative of implicit functional dependence.

Significance and Claims

The paper claims that the coverage correlation coefficient fills a critical gap in statistical association measures by offering a symmetric, distribution-free, and computationally efficient tool for detecting complex, non-monotonic, and singular dependencies. Its ability to provide an analytically tractable asymptotic null distribution makes it particularly well-suited for large-scale pairwise testing scenarios (e.g., genomics) where permutation-based methods are computationally prohibitive.

The authors position the method as a robust alternative to existing divergence-based estimators that require density estimation, noting that the coverage correlation operates directly on ranks and uncovered volumes. While the paper acknowledges a theoretical gap in consistency proofs for dimensions $d_X + d_Y \in \{3, 4\}$, it asserts that the method is theoretically sound for univariate cases and high-dimensional cases ($d_X + d_Y \geq 5$) under regularity conditions, and empirically effective across a broad range of dimensions. The work emphasizes that the statistic measures the extent to which the joint law fails to cover the product space, providing a geometric perspective on dependence that complements existing functional dependence measures.