Objective Correlation of Variables

• Objective correlation of variables is defined as mathematically derived measures that quantify the degree of dependence between variables, attaining zero under independence and maximal values for deterministic relationships.
• These measures include classical indices like R² and modern nonparametric metrics such as Chatterjee’s ξ, enabling robust analysis of linear, nonlinear, and mixed data types.
• They are essential for feature selection, model evaluation, and exploratory data analysis, providing scalable computation and equitable inference across complex applications.

Objective correlation of variables refers to mathematically well-defined, data-driven measures of association between variables, designed to have intrinsic properties such as being zero under independence and attaining extremal values only under deterministic relationships (often within a specified functional class). These measures aim to quantify dependence in a manner that is robust to data type, invariant (where appropriate) under transformations, and interpretable on universal, often [0,1] or [−1,1] scales. They include both classical coefficients—for linear associations—and a range of modern, nonparametric, or generalized metrics that address complex data structures, higher dimensions, nonlinearities, mixed types, and functional or shape-based restrictions.

1. Classical and Generalized $R^2$ Measures

The archetypal "objective correlation" in statistics is the coefficient of determination, $R^2$, defined in linear regression as the proportion of variance in $Y$ explained by $X$. For $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$, with all classical variance assumptions, $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$, directly interpretable as $\rho^2$, the squared Pearson correlation coefficient in the bivariate case. $R^2$ quantifies signal-to-noise ratio for linear or explicitly modeled relationships.

However, $R^2$ is not meaningful or equitable for general nonlinear or non-parametric dependencies. Approaches such as the generalized $R^2$ via density ratios provide an information-theoretic extension to arbitrary associations. In this framework, association $R^2$0 is defined as

$R^2$1

where $R^2$2 and $R^2$3 are null (independence) and alternative (joint) density estimates via kernel methods. As $R^2$4, $R^2$5 converges to the informational measure of correlation in the sense of Linfoot. $R^2$6 achieves equitability across different functional forms and noise levels, outperforms the Maximal Information Coefficient (MIC) in power and convergence, and generalizes to multivariate and conditional association with appropriate modifications to the null density (e.g., $R^2$7 for controlling on $R^2$8) (Murrell et al., 2013).

2. Nonparametric and Model-Free Dependence Measures

Recently, several measures have been proposed to overcome the limitations of parametric or linear-only coefficients:

• Chatterjee's Coefficient ($R^2$9): For real $Y$0, $Y$1 is defined as the normalized total variation of $Y$2 over $Y$3,

$Y$4

with sample estimator

$Y$5

where $Y$6 is the rank of $Y$7 after sorting $Y$8 by $Y$9. This estimator is distribution-free, consistent, and satisfies the extremal properties: $X$0 if and only if independence, $X$1 if and only if $X$2 almost surely (Chatterjee, 2019).

• Coverage Correlation Coefficient ($X$3): For $X$4, $X$5 is a nonparametric, rank-based, $X$6-divergence estimator quantifying the mass of the joint that is singular relative to the product of the marginals. Using Monge–Kantorovich (transport) ranks, $X$7 is exactly 0 for independence and 1 for singular couplings (copulas), is symmetric, obeys data-processing inequalities, and is efficiently computable in $X$8 in low dimensions (Yang et al., 8 Aug 2025).
• Alternant Recursive Topology (ART) Correlation: ART-based measures (e.g., ARTMIC) use recursive partitioning and exact maximization of restricted mutual information, normalized by provable upper bounds, to provide objective, equitable dependence scores for arbitrary bivariate relationships. Companion statistics assess local randomness, monotonicity asymmetry, and complexity (Liu et al., 2016).
• Label Projection Correlation Statistic (PCor): For mixed continuous and categorical variables, PCor computes expected $X$9-distances between conditional and unconditional projected distributions, is zero if and only if independence, and admits efficient, rank-based $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$0 computation for $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$1 (Liu et al., 27 Apr 2025).
• Copula and Optimal Transport Based Coefficients: Empirical copula transformations combined with Wasserstein distances allow for clustering and targeted dependence coefficients (TFDC) that can emphasize or suppress specific dependence patterns, flexible for non-linear, marginal-invariant, and tail-focused analysis (Marti et al., 2016).

3. Generalized and Distributional Correlations

Objective correlation can be constructed with respect to arbitrary statistical functionals $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$2 (means, quantiles, expectiles, thresholds) by defining $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$3-identification functions $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$4, the associated generalized errors $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$5, and generalized covariances $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$6.

Normalization can be via:

• Cauchy–Schwarz: $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$7
• Fréchet–Hoeffding: Normalization by maximal attainable values under given marginals, leading to sharp bounds and ensuring full range $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$8. This normalization provides that perfect positive dependence achieves $Y_i = \beta_0 + \beta_1 X_i + \varepsilon_i$9, perfect negative dependence $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$0, and independence $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$1 (Fissler et al., 2023).

Distributional (function-valued) objective correlations include:

• Threshold and Quantile Correlations: These localize dependence to regions (thresholds, quantiles) of interest and refine measurement of tail dependence that cannot be isolated by classical tail dependence coefficients.
• Summary Correlations: Integration over the index set of distributional correlations (e.g., quantiles) produces scalar-valued summaries corresponding to classical measures—Spearman's $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$2 as the summary quantile correlation and Pearson's $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$3 as the summary CDF correlation.

4. Objective Measures for Mixed and Higher-Order Data

Several frameworks generalize objective correlation to multivariate, mixed-type, or factorial settings:

• $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$4 Coefficient: Based on a noise-corrected $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$5 statistic from the contingency table, $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$6 is designed for categorical, ordinal, and continuous variables (after binning), reduces to Pearson's $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$7 for bivariate normal input, and is sensitive to nonlinear effects (Baak et al., 2018).
• Unified Correlation via Covariance with $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$8-Transformed Variables: $R^2 = 1 - SS_{\text{res}} / SS_{\text{total}}$9 uses covariances between $\rho^2$0 and an $\rho^2$1-transformed version of $\rho^2$2, with $\rho^2$3 a law-determined functional (e.g., the marginal of $\rho^2$4 or the uniform), thus subsuming variance, Gini mean difference, and entropy-based measures (Asadi et al., 2018).
• Unsigned Correlation Coefficient (UCC): For $\rho^2$5 variables, UCC is built geometrically from determinants of all $\rho^2$6-by-$\rho^2$7 parallelograms formed by standardized data vectors, directly reducing to $\rho^2$8 for $\rho^2$9, and providing a symmetric, scale- and translation-invariant measure of joint linear dependence (Wang et al., 2014).
• Taylor's Multi-Way Correlation: $R^2$0 is derived from the standard deviation of the eigenvalues of the empirical correlation matrix of the variables, generalizing $R^2$1 for $R^2$2, attaining 1 for maximal collinearity, and 0 for mutual independence (Taylor, 2020).

5. Tailored and Functional-Restricted Objective Correlations

Correlations can be tailored to recognize only functional dependencies within a restricted functional class $R^2$3: $R^2$4 Under such constraints, the symmetrized version $R^2$5 is defined, with the property that it is 1 if and only if $R^2$6 for some $R^2$7 and 0 if and only if $R^2$8 and $R^2$9 are independent. For monotone relationships, isotonic regression is applied, and the empirical estimator is based on the fraction of variance in the monotonic projection explained (Cao et al., 2020).

Chatterjee's $R^2$0 is a special case, unrestricted in function class, and affords the same endpoint properties but no additional structure. All these measures have at most logarithmic local power under contiguous alternatives but are robust and interpretable as strength measures across diverse dependence scenarios.

6. Application, Computation, and Practical Implications

Objective correlation measures have broad application:

• In feature screening and model selection, measures such as PCor and Chatterjee's $R^2$1 avoid false positives under independence, are robust to heavy tails/outliers, and detect both linear and complex dependencies (Liu et al., 27 Apr 2025).
• In multiobjective optimization, fixed objective correlation ($R^2$2 parameter in $R^2$3MNK-landscapes) strongly shapes the efficient set structure, connectivity, and tractability, dictating appropriate metaheuristic strategies (Verel et al., 2012).
• For principal component and factor analysis, the absolute correlations between original variables and components are analytically defined as $R^2$4 and directly guide objective variable selection and component truncation (Gniazdowski, 2023).
• Neural network–based measures, such as hyper-occurrence (hoc), contrast activation probabilities on real vs. scrambled data and aggregate these differences to summarize multivariate correlation patterns, fostering robust, regularized feature extraction without explicit model structure (Fontana, 2016).
• In mixed-type and sparse settings, $R^2$5 and related CDF/quantile summary correlations address practical needs for objective and unified scoring, with correction for binning or expected random association (Baak et al., 2018, Fissler et al., 2023).
• Objective-subjective concordance: in psychophysics, high Spearman or other objective correlation between engineered indices and subjective experience—as demonstrated in the case of noise barriers—validates the predictive power and practical adequacy of objective indices under controlled perceptual experiments (Redondo et al., 2024).

The computational complexity varies: many modern estimators (Chatterjee's $R^2$6, PCor, $R^2$7, kernel-density $R^2$8, UCC, mcor) are $R^2$9 to $R^2$0, with higher costs for multivariate integration, dynamic recursive partitioning (ART), or optimal transport in high dimensions.

7. Limitations and Theoretical Guarantees

Most objective correlation measures satisfy symmetry, invariance, sharp bounds (i.e., range of [0,1] or [−1,1]), extremality for deterministic or functional relationships, and vanishing for independence. However, certain measures (e.g., $R^2$1, isotonic correlations) do not have nonzero Pitman local power; i.e., their asymptotic distribution under contiguity is degenerate. Power against local alternatives is nontrivial, but finite-sample performance is often competitive or superior for nonlinear or irregular dependencies. The distribution-free nature and fast convergence of several estimators (notably $R^2$2, $R^2$3, $R^2$4) make them effective for high-throughput discovery (Murrell et al., 2013, Yang et al., 8 Aug 2025, Chatterjee, 2019).

Practical limitations include sensitivity to binning, computational cost in high dimensions, and requirement for careful type handling with mixed or tied data. Objective correlation measures continue to evolve to address complex structures in modern data science, with a focus on robustness, equitability, and statistical interpretability.