---
title: Epsilon-Delta Analysis of Chatterjee's Rank Correlation
url: https://www.emergentmind.com/topics/epsilon-delta-interpretation-of-chatterjee-s-rank-correlation
type: topic
---

# Epsilon-Delta Analysis of Chatterjee's Rank Correlation

Chatterjee’s rank correlation, denoted $\xi(X, Y)$, is a nonparametric functional measuring the degree of association between random variables $X$ and $Y$ that attains 1 for perfect functional dependence, 0 under independence, and interpolates smoothly in between. The $\varepsilon$–$\delta$ interpretation of this coefficient offers a rigorous quantification of its stability, sensitivity, and continuity under perturbations of the joint distribution—whether by gross contamination, weak convergence, or local dependence—providing tight analytical bounds for inference and robustness.

## 1. Fundamental Definitions and Forms

Chatterjee’s sample rank correlation, for i.i.d. data $(X_i, Y_i)$ with ties in $X$ resolved appropriately, is defined as
\[
\xi_n(X, Y) = 1 - \frac{3}{n^2-1}\sum_{i=1}^{n-1}|r_{i+1} - r_i|,
\]
where $r_i$ are the ranks of $Y_{(i)}$ after sorting the data by $X$ [1909.10140]. The population version, critical for asymptotic and contamination analysis, is
\[
\xi(X, Y) = \frac{\int_{\mathbb{R}}\mathrm{Var}(G_X(t))\,d\mu(t)}{\int_{\mathbb{R}}G(t)(1-G(t))\,d\mu(t)},
\]
where $G_X(t) = \mathbb{P}(Y\geq t|X)$ and $G(t) = \mathbb{P}(Y\geq t)$.

Chatterjee’s coefficient admits formulations via local $L^1$ residuals, Markov-product copulas, and conditional variances, which underlie the various $\varepsilon$–$\delta$ analyses [2512.12363, 2503.11390].

## 2. Sensitivity and Robustness via Asymptotic Expected Sensitivity Function

The primary device for $\varepsilon$–$\delta$ robustness is the Asymptotic Expected Sensitivity Function (AESF), defined for a functional $R$ as
\[
\mathrm{AESF}(z; R, F) = \lim_{n\to\infty} (n+1)\,\mathbb{E}_{F}\left[R_{n+1} (X_1,\dots,X_n, z) - R_n(X_1,\dots,X_n)\right],
\]
where $R_n$ is the empirical plug-in estimator. For Chatterjee’s $\xi$, if the supremum $M = \sup_{(x, y)\in\mathbb{R}^2} |\mathrm{AESF}((x,y); \xi, F)| < \infty$, then under $\varepsilon$-contamination
\[
F_\varepsilon = (1-\varepsilon)F + \varepsilon H,
\]
one obtains the first-order contamination bound
\[
|\xi(F_\varepsilon) - \xi(F)| \leq \varepsilon M + o(\varepsilon),
\]
and more conservatively for all $\varepsilon \in [0,1)$,
\[
|\xi(F_\varepsilon) - \xi(F)| \leq \frac{\varepsilon M}{1-\varepsilon}
\]
when the functional is Lipschitz in total variation [2401.05281]. This bound is tight: the worst case occurs when $H$ concentrates mass at the point where $|\mathrm{AESF}|$ is maximized.

For example, in a linear-Gaussian case with $\rho=0.7$, a numerical value $M_{0.7} \approx 1.15$ yields, for $\varepsilon=0.01$, a maximal shift $|\xi(F_\varepsilon)-\xi(F)| \leq 0.012$ under 1% contamination.

## 3. $\varepsilon$–$\delta$ Structure for Functional Dependence and Independence

Chatterjee’s coefficient exhibits explicit $\varepsilon$–$\delta$ equivalences at the endpoints:

- **Functional dependence:** If $Y = f(X)$ a.s., then $\xi=1$. Conversely, if $\xi=1$, $Y$ is almost surely a function of $X$. Finite deviation from noiseless dependence, $\Delta = \mathbb{P}(Y \neq f(X)) < \delta$, yields $1-\xi < \epsilon$ for $\delta=\epsilon D$, where
  \[
  D = \int_\mathbb{R} G(t)(1-G(t))\,d\mu(t)
  \]
  [1909.10140].
- **Independence:** $\xi=0$ if and only if $X$ and $Y$ are independent. For uniform distance from independence $\alpha < \delta$, one gets $\xi < \delta^2/D$. Conversely, if $\xi < \epsilon$, then $\alpha < \sqrt{\epsilon D}$.

These bounds justify the interpretation of $\xi$ as a calibrated, Lipschitz-quantified “distance” from both perfect dependence and independence, with explicit $\varepsilon$–$\delta$ parameters.

## 4. Continuity: Markov Products and Weak Convergence

Continuity properties of $\xi$ in the weak topology deviate from classical rank correlations. Chatterjee’s $\xi$ is not continuous with respect to weak convergence of joint laws, but instead with respect to the law of Markov products $(Y, Y')$ where $Y'$ is conditionally independent given $X$ and $Y'|X \sim Y|X$ [2503.11390].

**Theorem (ε–δ continuity of $\xi$):** For $F_{Y_n}$ continuous and suitable range convergence,
\[
d_P(\mathrm{Law}(Y_n,Y_n'), \mathrm{Law}(Y,Y')) < \delta \implies |\xi(Y_n, X_n) - \xi(Y, X)| < \epsilon,
\]
where $d_P$ is the Prokhorov distance. Copula-based representations yield bounds such as $|\xi(Y_n, X_n) - \xi(Y, X)| \le 6\delta$ when the uniform norm $\|C_{Y_n,Y_n'}-C_{Y,Y'}\|_\infty < \delta$ [2503.11390].

This ensures that small perturbations in the conditional law of $Y|X$, as measured in the appropriate metric (not simply the joint law), produce arbitrarily small effects on $\xi$, with explicit $\varepsilon$–$\delta$ quantification.

## 5. Primitive Local $\varepsilon$–$\delta$ Construction and Empirical Structure

A local $\varepsilon$–$\delta$ perspective frames Chatterjee’s $\xi$ as the limiting residual of a local averaging scheme:
- For $(U, V)$ with $U = F_X(X)$, $V = F_Y(Y)$ (probability-integral transforms), define for $\varepsilon > 0$ the empirical $\varepsilon$-neighborhood of $U_i$ as $\mathcal{N}_\varepsilon(i) = \{j: |U_j-U_i|\leq \varepsilon\}$.
- The local average of $V$ near $U_i$ is $\bar{V}_i(\varepsilon) = \frac{1}{|\mathcal{N}_\varepsilon(i)|}\sum_{j\in\mathcal{N}_\varepsilon(i)} V_j$.
- The mean local $L^1$ residual is $\zeta_n(\varepsilon) = \frac{1}{n}\sum_{i=1}^n |\bar{V}_i(\varepsilon) - V_i|$.
- In the $\varepsilon \to 0$ limit, $\zeta_n(\varepsilon)$ converges to $\mathbb{E}|V - \mathbb{E}[V|U]|$; Chatterjee’s correlation emerges as
  \[
  \xi_n = 1 - \frac{\zeta_n}{\mathbb{E}|V - \mathbb{E} V|} = 1-4\zeta_n,
  \]
  matching the original rank-difference formula [2512.12363].

All $\varepsilon$–$\delta$ operations (local sets, residuals) are invariant under monotone transformations; the probability-integral transform serves only to achieve distribution-freeness.

## 6. Moment-Based Analogues and Unified Framework

Replacement of the local $L^1$ residual with $L^2$ analogues links Chatterjee’s $\xi$ to familiar moment-based indices:
- $\zeta^{(2)} = \mathbb{E}[(V - \mathbb{E}[V|U])^2] = \mathbb{E}[\mathrm{Var}(V|U)]$
- $\eta^{(2)} = 1 - \frac{\mathbb{E}[\mathrm{Var}(V|U)]}{\mathrm{Var}(V)} = \frac{\mathrm{Var}(\mathbb{E}[V|U])}{\mathrm{Var}(V)}$

For jointly Gaussian $(U, V)$, one recovers Pearson’s $R^2$ through this construction, showing the $\varepsilon$–$\delta$ approach unifies rank-based and moment-based dependence measures under a single limiting framework [2512.12363].

## 7. Assumptions, Limitations, and Practical Implications

Rigorous $\varepsilon$–$\delta$ control relies on continuity in $Y$, regularity of $F$, and Hadamard differentiability of $\xi$. The main theoretical limits—tightness of the contamination bound and sharpness of independence/functionality bounds—are achieved under these hypotheses [2401.05281, 1909.10140]. In finite samples, contamination and sampling errors are additive, with the former scaling as $\varepsilon M$ and the latter as $O_p(1/\sqrt{n})$.

A plausible implication is that for statistical inference and robust estimation, Chatterjee’s $\xi$ offers explicit, interpretable robustness margins, with $\varepsilon$–$\delta$ quantification superior to earlier rank-based coefficients where such fine-grained control is unavailable or only asymptotically valid. The local $\varepsilon$–$\delta$ interpretation remains central for applications in dependence quantification, goodness-of-fit, and model diagnostics.

Source: https://www.emergentmind.com/topics/epsilon-delta-interpretation-of-chatterjee-s-rank-correlation