Coverage Correlation Coefficient

Updated 8 July 2026

Coverage correlation is a nonparametric measure that quantifies the concentration of joint distributions relative to the product of marginals, ranging from 0 for independence to 1 for singular dependence.
It employs Monge–Kantorovich ranks and geometric vacancy (using small hypercubes) to construct a distribution-free statistic that efficiently captures complex, nonlinear associations.
The method features an analytically tractable asymptotic null distribution, making it suitable for large-scale nonparametric dependence screening in high-dimensional settings.

Coverage correlation, or the coverage correlation coefficient, is a nonparametric measure of statistical association designed to quantify the extent to which two random variables have a joint distribution concentrated on a singular subset with respect to the product of the marginals. It consistently estimates an $f$ -divergence between the joint distribution and the product of the marginals, is $0$ if and only if the variables are independent, and is $1$ if and only if the copula is singular. Using Monge–Kantorovich ranks, it naturally extends to random vectors; it is distribution-free, admits an analytically tractable asymptotic null distribution, and can be computed efficiently, which makes it suited to complex, potentially nonlinear association testing in large-scale settings (Yang et al., 8 Aug 2025).

1. Motivation and geometric intuition

Classical measures, such as Pearson, Spearman, or Kendall correlations, mainly capture linear or monotonic associations and may miss complex or non-monotonic dependencies. The coverage correlation was introduced for settings in which dependence may be singular, meaning that the joint distribution of $(X,Y)$ is concentrated on a lower-dimensional, Lebesgue measure-zero subset of the product space. The central object is not a regression slope or a rank concordance count, but the extent to which the empirical joint sample fails to “cover” the ambient space (Yang et al., 8 Aug 2025).

The geometric construction is easiest to state when $X$ and $Y$ have marginals uniform on $[0,1]$ . Given i.i.d. samples $(X_i,Y_i)_{i=1}^n$ , one places $n$ small axis-aligned squares, equivalently $\ell_\infty$ -balls, each of area $0$0, centered at the observed points in $0$1. Under independence, these points cover most of the unit square. If $0$2 is singular with respect to Lebesgue measure, for example when the mass lies on a curve, the union of the balls covers much less area. The uncovered area, termed the vacancy, therefore carries information about dependence: the more concentrated the joint law, the larger the vacancy.

This construction generalizes to higher dimensions. For random vectors $0$3 and $0$4, the same idea uses small cubes of volume $0$5 in $0$6, where $0$7. The paper’s perspective is that singular dependence is detectable through the geometry of coverage rather than through monotonicity or linearity.

2. Empirical definition and rank transformation

To remove marginal effects, the construction first rank-transforms the observations via Monge–Kantorovich ranks. Given reference points $0$8 and $0$9 uniformly spread in $1$0 and $1$1, respectively, the ranks are defined through optimal assignments. For $1$2,

$1$3

and $1$4. The construction is applied likewise to $1$5, after which the joint rank is $1$6 (Yang et al., 8 Aug 2025).

For each $1$7, the method forms an $1$8-ball $1$9 of radius

$(X,Y)$ 0

so that each cube has volume $(X,Y)$ 1. The vacancy is then

$(X,Y)$ 2

where “vol” denotes Lebesgue measure in $(X,Y)$ 3. The coverage correlation coefficient $(X,Y)$ 4 is obtained by normalizing this vacancy using $(X,Y)$ 5, the asymptotic expected vacancy under independence. In the paper’s interpretation, $(X,Y)$ 6 means that the joint law is close to independence, while $(X,Y)$ 7 indicates a highly singular, maximally dependent joint law.

The use of Monge–Kantorovich ranks is essential. Because arbitrary marginals are mapped to a reference uniform grid, the resulting statistic is distribution-free under independence. This rank-based construction also yields the extension from univariate to multivariate $(X,Y)$ 8 and $(X,Y)$ 9 without changing the geometric definition of vacancy.

3. Population target and structural properties

At the population level, the coverage correlation converges to an $X$ 0-divergence between the joint distribution $X$ 1 and the product of marginals $X$ 2. The paper states that this divergence is zero if and only if the variables are independent and one if and only if the joint law is singular with respect to the product measure. A key point is that the statistic consistently estimates this $X$ 3-divergence without explicit density estimation (Yang et al., 8 Aug 2025).

Several structural properties are emphasized. The measure is symmetric, so $X$ 4. It satisfies an information monotonicity statement, $X$ 5 if $X$ 6. The paper also describes lower semicontinuity and information gain in the form that if extra information is added, the measure does not decrease. These properties position coverage correlation as a dependence functional tied to the geometry of support concentration rather than to a specific direction of prediction.

A central comparison is with Chatterjee’s correlation. The paper states that Chatterjee’s measures only functional, directional dependence, with limit $X$ 7 if and only if $X$ 8 is a measurable function of $X$ 9, and is asymmetric. Coverage correlation, by contrast, is symmetric, attains one for any singular relationship, even non-functional, and is zero for independence. This includes implicit relationships such as $Y$ 0, as well as support on a curve or manifold of measure zero.

Property	Statement
Independence characterization	$Y$ 1 iff independence
Maximality	$Y$ 2 iff the joint is singular relative to the product measure
Symmetry	$Y$ 3
Information monotonicity	$Y$ 4 if $Y$ 5

These properties clarify the intended scope of the coefficient. It is not merely another omnibus dependence test; it is specifically tuned to detect concentration of joint mass on sets that are small relative to the product geometry induced by the marginals.

4. Null distribution and hypothesis testing

Under independence, the paper gives an asymptotically normal null distribution: $Y$ 6 where $Y$ 7 has an explicit formula given in the paper (Yang et al., 8 Aug 2025).

The paper further reports an explicit asymptotic variance,

$Y$ 8

The consequence is that the null law is pivotal: it does not depend on the marginal distributions. Combined with the rank transform, this yields a test that is distribution-free under the null for arbitrary marginals.

This analytic null distribution is one of the method’s principal practical advantages. The paper emphasizes that p-values can be computed analytically rather than by permutation, which is critical in large-scale pairwise testing. A plausible implication is that the method is particularly well aligned with settings in which many candidate associations must be screened and permutation calibration would be computationally prohibitive.

5. Computation, scaling, and multivariate use

The computational pipeline has two main components: optimal assignment for the Monge–Kantorovich ranks and geometric computation of the vacancy. The rank computation is an instance of the assignment or optimal transport problem. After ranking, the vacancy is the uncovered volume of a union of $Y$ 9 small hypercubes in $[0,1]$ 0 (Yang et al., 8 Aug 2025).

For the coverage calculation, the paper states complexity bounds of $[0,1]$ 1 in $[0,1]$ 2D via Bentley’s algorithm and $[0,1]$ 3 in higher dimensions, with grid-based partitioning improving real-world performance. Software implementations are reported in both R and Python under the name covercorr. The paper describes the resulting scaling as suitable for large-scale dependence screening.

The same framework extends directly to multivariate pairs. Because the construction is based on ranks and coverage in $[0,1]$ 4, it applies to random vectors $[0,1]$ 5 and $[0,1]$ 6 without requiring density estimates or kernel bandwidth selection. In empirical and simulation studies, the method is described as consistently detecting singular dependencies that Pearson, Spearman, Chatterjee, HSIC, and dCor miss. The reported application areas include exploratory association screening in high-dimensional data such as single-cell RNA-seq and genome-wide marker analyses.

The term “coverage correlation coefficient” is not uniform across the literature. In the statistical sense introduced in “Coverage correlation: detecting singular dependencies between random variables” (Yang et al., 8 Aug 2025), it denotes the vacancy-based, rank-based dependence measure described above. In a different wireless-systems usage, “Spatio-temporal Interference Correlation and Joint Coverage in Cellular Networks” defines a spatio-temporal interference correlation coefficient, also called a coverage correlation coefficient, by

$[0,1]$ 7

and studies its decay with user displacement $[0,1]$ 8, together with joint coverage probability across locations (Krishnan et al., 2016).

A separate group of cellular papers analyzes the impact of correlation between interferers on coverage probability and rate. These papers define interferer correlation coefficients

$[0,1]$ 9

and explicitly state that they do not introduce a new metric called “coverage correlation”; instead, coverage changes are induced by interferer correlation through the correlation matrix and its eigenvalues [(Kumar et al., 2014); (Kumar et al., 2017)]. In another distinct context, coherent noise radar work uses a correlation coefficient $(X_i,Y_i)_{i=1}^n$ 0 between received and reference signals as an adjunct to SNR for performance prediction (Luong et al., 2020).

The phrase also should not be conflated with eigenvalue-based multivariate extensions of Pearson’s coefficient. “The extension of Pearson correlation coefficient, measuring noise, and selecting features” defines a different multivariate coefficient $(X_i,Y_i)_{i=1}^n$ 1 using the maximal eigenvalue or spectral norm of the correlation matrix (Salimi et al., 2024). A plausible implication is that “coverage correlation” in current statistical usage identifies a specific singular-dependence detector, not a generic multivariate correlation construction.

7. Significance for dependence analysis

Coverage correlation is especially sensitive to singular and nonlinear dependencies, including cases in which the support is a curve, a manifold, or another lower-dimensional set, and including implicit relationships that are not functional in either direction. The paper’s central claim is that the coefficient quantifies how much the joint distribution is concentrated relative to the null of independence, via an explicit $(X_i,Y_i)_{i=1}^n$ 2-divergence and a geometric statistic that avoids explicit density estimation (Yang et al., 8 Aug 2025).

Its combination of properties is unusually specific: symmetry, independence characterization, maximal response to singular copulas, multivariate extension through Monge–Kantorovich ranks, analytic null inference, and computational tractability. This suggests a methodological niche distinct from correlation measures optimized for monotonicity, linearity, or directional predictability. In that niche, the coefficient functions simultaneously as a descriptive dependence measure and as a practical test statistic for large-scale exploratory analysis.