Chatterjee’s Rank Correlation

• Chatterjee’s rank correlation is a nonparametric, rank-based statistic that quantifies the extent to which one variable is a function of another, ranging from 0 (independence) to 1 (perfect dependence).
• It is derived from copula theory, features a closed-form for continuous distributions, and supports efficient bootstrap and kernel-based inference with asymptotic normality properties.
• Key limitations include issues with weak continuity and local power against near-independence, which can be mitigated by combining it with classical measures like Spearman’s rho.

Chatterjee’s Rank Correlation Coefficient is a nonparametric, rank-based measure of dependence between random variables, designed to quantify the extent to which one variable is a functional of another, regardless of monotonicity. The statistic is grounded in copula theory, admits a closed-form for continuous distributions, is distribution-free under independence, and is asymptotically normal in a wide range of settings. While it provides powerful tools for functional dependence and independence testing, subtle issues arise regarding weak continuity and local power that distinguish it from classical concordance measures.

1. Definition and Mathematical Formulation

Let $(X, Y)$ be continuous random variables with joint distribution $F_{X,Y}$ and copula $C$. Chatterjee’s rank correlation coefficient, generally denoted $\xi$ or $\xi(X,Y)$, is defined in multiple, equivalent forms:

• Copula Form:

$$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$

where $\partial_1 C(u,v)$ is the partial derivative of the copula with respect to its first argument (Sato, 13 Dec 2025, Ansari et al., 18 Jun 2025, Rockel, 8 Sep 2025).

• Population (Integral) Form:

$$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$

This representation makes it clear that $\xi$ measures the proportion of total variation in $Y$ explained via conditioning on $F_{X,Y}$0 (Sato, 13 Dec 2025, Dalitz et al., 2023).

• Sample Estimator: For an i.i.d. sample $F_{X,Y}$1 of size $F_{X,Y}$2, and no ties in $F_{X,Y}$3,

$F_{X,Y}$4

Here, $F_{X,Y}$5 is the rank of $F_{X,Y}$6 where data are sorted so $F_{X,Y}$7 and $F_{X,Y}$8 is the concomitant value (Sato, 13 Dec 2025, Zhang, 2022, Zhang, 2023).

• Graph-Based Extension: In higher dimensions or for multivariate $F_{X,Y}$9, Azadkia and Chatterjee’s graph-based version is defined using nearest-neighbor pairs in $C$0-space (Ansari et al., 2022, Han et al., 2022).

2. Fundamental Properties

• Range and Characteristic Values: $C$1 is bounded between 0 and 1. $C$2 if and only if $C$3 and $C$4 are independent; $C$5 if and only if $C$6 is a measurable function of $C$7 (Zhang, 2022, Ansari et al., 2022, Ansari et al., 18 Jun 2025).
• Invariance: $C$8 is invariant under strictly increasing transformations of $C$9 or $\xi$0 (Dalitz et al., 2023, Rockel, 8 Sep 2025).
• Directional Nature: Generally, $\xi$1: the statistic is not symmetric. A symmetrized version is given by $\xi$2 (Zhang, 2022).
• Functional Dependence Interpretation: $\xi$3 quantifies the strength of (possibly non-monotone) functional dependence of $\xi$4 on $\xi$5 (Rockel, 8 Sep 2025, Ansari et al., 18 Jun 2025).
• Consistency: Under i.i.d. sampling from a continuous joint distribution, $\xi$6 almost surely (Dalitz et al., 2023, Zhang, 2022).

3. Asymptotic and Finite-Sample Theory

• Asymptotic Normality:
  • Under independence, $\xi$7 (Zhang, 2022, Auddy et al., 2021, Lin et al., 2022, Kroll, 2024).
  • If the variables are not functionally dependent, $\xi$8 is asymptotically normal around its mean, with variance uniformly bounded by 36 (Lin et al., 2022).
• Symmetrized Statistic: The maximum of $\xi$9 and $\xi(X,Y)$0 converges in distribution to a skew-normal limit (Zhang, 2022).
• Local Power and Detection Boundary: For independence testing, the detection boundary for alternatives with $\xi(X,Y)$1 cannot be reached; for many classical alternatives (e.g. Gaussian correlation $\xi(X,Y)$2), only $\xi(X,Y)$3 is detectable (Auddy et al., 2021, Shi et al., 2020). This rate is suboptimal relative to classical measures like Hoeffding's $\xi(X,Y)$4, Blum-Kiefer-Rosenblatt's $\xi(X,Y)$5, and Yanagimoto’s $\xi(X,Y)$6, which detect alternatives at the $\xi(X,Y)$7 regime.
• Minimax-optimality for Strong Dependence: For testing a fixed nonzero level of dependence ($\xi(X,Y)$8), tests based on $\xi(X,Y)$9 achieve the optimal $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$0 rate (Auddy et al., 2021).
• Variance Estimation and Bootstrap: Analytical and $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$1-out-of-$$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$2 bootstrap-based variance estimations are consistent for constructing valid confidence intervals (Dette et al., 2023, Dalitz et al., 2023).

4. Relationship to Other Rank Correlations

• Spearman's $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$3: While both statistics are rank-based, $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$4 focuses on functional dependence, whereas $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$5 captures monotonic association. The possible $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$6 pairs fill a convex region; for stochastically increasing (SI) or decreasing copulas, $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$7, with a maximal difference of $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$8 attainable by an explicit copula (Ansari et al., 18 Jun 2025).
• Spearman’s Footrule $$\xi(C) = 6 \iint_{[0,1]^2} [\partial_1 C(u,v)]^2\,u v\, du\,dv - 2$$9: When $\partial_1 C(u,v)$0 and $\partial_1 C(u,v)$1 are continuous, $\partial_1 C(u,v)$2 equals the footrule of the Markov product of the copula and its transpose. For SI copulas, the region $\partial_1 C(u,v)$3 ($\partial_1 C(u,v)$4) is sharp (Rockel, 8 Sep 2025).
• Extremal Cases: There are explicit rank patterns where $\partial_1 C(u,v)$5 is close to zero but $\partial_1 C(u,v)$6 is near 1 and vice versa, highlighting their complementary sensitivities (Zhang, 2023).
• Combined Tests and Power: Max-type tests combining $\partial_1 C(u,v)$7 with Spearman’s $\partial_1 C(u,v)$8 or Kendall’s $\partial_1 C(u,v)$9 (e.g., $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$0) have favorable properties: they are asymptotically jointly normal under independence, with nontrivial power for both monotone and non-monotone scenarios (Zhang, 2023, Zhang, 2024).

5. Graph-Based and Multivariate Extensions

• Azadkia–Chatterjee Correlation: In the presence of multivariate predictors, the statistic generalizes to the nearest neighbor graph-based estimator,

$$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$1

for $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$2 the nearest neighbor of $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$3 (Ansari et al., 2022, Han et al., 2022, Tran et al., 2024).

• Manifold Adaptivity: When $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$4 lies on an $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$5-dimensional submanifold, the limiting null variance depends only on $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$6, not on the ambient dimension (Han et al., 2022).
• Rank-Based NNG: The rank-vector-based nearest-neighbor graph (Rosenbaum NNG), which uses marginal ranks, achieves full scale invariance and improved finite-sample behavior especially in high-dimensional or rescaled covariate regimes (Tran et al., 2024).
• Multiresponse Generalization: The $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$7 statistic, defined via sequential conditioning and normalization, extends the measure to $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$8, preserving interpretation and properties such as equitability and scale invariance. The estimator $$\xi(X, Y) = \frac{\displaystyle \int \operatorname{Var} \left( \mathbb{E}[ \mathbf{1}\{ Y \geq t \} \mid X] \right) dF_Y(t)}{\displaystyle \int \operatorname{Var}(\mathbf{1}\{ Y \geq t \}) dF_Y(t)}$$9 is strongly consistent and asymptotically normal under mild conditions (Ansari et al., 2022).

6. Practical Computation and Inference

• Computational Complexity: For univariate $\xi$0, computing $\xi$1 requires $\xi$2 time. The nearest-neighbor generalization for multivariate $\xi$3 can also be computed in $\xi$4 with data structures such as KD-trees (Dalitz et al., 2023, Tran et al., 2024).
• Bias and Normalization: The maximum attainable value of $\xi$5 is strictly less than 1 for small $\xi$6, introducing finite-sample bias. Simple normalization (e.g., scaling by its maximal value on the sample) reduces this bias (Dalitz et al., 2023).
• Bootstrap Inference: The $\xi$7-out-of-$\xi$8 bootstrap is consistent for distributional inference on $\xi$9 for both continuous and discrete data, and usually outperforms the classical $Y$0-out-of-$Y$1 bootstrap, especially in terms of coverage accuracy (Dette et al., 2023, Dalitz et al., 2023).
• Kernel Estimation: A kernel estimator for $Y$2 achieves asymptotic normality at a faster rate than the original rank-based estimator, providing improved detection for local alternatives near independence (Azadkia et al., 15 Feb 2026).

7. Limitations, Controversies, and Open Questions

• Lack of Weak Continuity: Chatterjee’s coefficient is not continuous with respect to weak convergence. Sequences of distributions may converge weakly to independence, yet $Y$3 remains nonzero (even 1) along the sequence. This discontinuity is structurally required by the property that $Y$4 iff $Y$5 (Bücher et al., 2024).
• Pathologies in Inference: Due to the above, tests for independence or confidence intervals based on $Y$6 can have trivial power against alternatives arbitrarily close to independence, and uniform confidence intervals may fail to shrink, rendering them uninformative in large samples (Bücher et al., 2024).
• Local Power Deficiency: For certain classical local alternatives (e.g., Gaussian correlation, rotation/mixture alternatives), $Y$7 is rate-suboptimal for independence testing compared to classical U-statistics, missing alternatives at the parametric $Y$8 detection threshold (Shi et al., 2020, Auddy et al., 2021).
• Combined Tests: Combining $Y$9 with monotonicity-sensitive measures like Spearman’s $F_{X,Y}$00 or Kendall’s $F_{X,Y}$01 can mitigate some power deficiencies, creating robust tests across a wide array of dependence structures (Zhang, 2023, Zhang, 2024).

• (Zhang, 2022) On the asymptotic distribution of the symmetrized Chatterjee's correlation coefficient
• (Bücher et al., 2024) On the lack of weak continuity of Chatterjee's correlation coefficient
• (Rockel, 8 Sep 2025) On the exact region between Chatterjee's rank correlation and Spearman's footrule
• (Dalitz et al., 2023) A Simple Bias Reduction for Chatterjee's Correlation
• (Azadkia et al., 15 Feb 2026) Kernel Estimation Of Chatterjee's Dependence Coefficient
• (Zhang, 2023) On relationships between Chatterjee's and Spearman's correlation coefficients
• (Lin et al., 2022) Limit theorems of Chatterjee's rank correlation
• (Sato, 13 Dec 2025) On the epsilon-delta Structure Underlying Chatterjee's Rank Correlation
• (Ansari et al., 2022) A direct extension of Azadkia & Chatterjee's rank correlation to multi-response vectors
• (Auddy et al., 2021) Exact Detection Thresholds and Minimax Optimality of Chatterjee's Correlation Coefficient
• (Dette et al., 2023) A Simple Bootstrap for Chatterjee's Rank Correlation
• (Ansari et al., 18 Jun 2025) The exact region and an inequality between Chatterjee's and Spearman's rank correlations
• (Zhang, 2024) On the extensions of the Chatterjee-Spearman test
• (Han et al., 2022) Azadkia-Chatterjee's correlation coefficient adapts to manifold data
• (Kroll, 2024) Asymptotic Normality of Chatterjee's Rank Correlation
• (Tran et al., 2024) On a rank-based Azadkia-Chatterjee correlation coefficient

Chatterjee’s rank correlation is thus a foundational statistic for nonparametric dependence assessment, functionally complete for general forms of regression-like dependence, computationally tractable, and extensively analyzed both asymptotically and in the finite-sample regime. Its use in practice should be informed by awareness of its unique strengths and fundamental inferential limitations.