Let $X$ and $Y$ be random variables. Then the maximal correlation $\rho_m(X;Y)$ is defined as $$ \rho_m (X;Y) := \max_{(f(X),g(Y))\in S} \mathbb{E} [f(X)g(Y)] $$ where $S$ is the collection of pairs of real-valued random variables $f(X)$ and $g(Y)$ such that $\mathbb{E}f = \mathbb{E}g = 0$ and $\mathbb{E}f^2 = \mathbb{E}g^2 = 1$.

It is known that $\rho_m \in [0,1]$, and $\rho_m = 0$ if and only if $X$ and $Y$ are independent. I am wondering if anything is known about the "robustness" of this latter fact; are there any known statements of the type, "if $\rho_m$ is small, then $X$ and $Y$ are nearly independent"? For example, suppose $X$ and $Y$ are discrete random variables, each over an alphabet of size $d$. Suppose $\rho_m(X;Y) < 1/d$. Does this imply anything about how close the distribution of $(X,Y)$ is to a product distribution? Or, perhaps there exist random variables that are far from independent yet have arbitrarily small maximal correlation.


For any Borel sets $A$ and $B$ such that $p_A:=P(X\in A)\in(0,1)$ and $r_B:=P(Y\in B)\in(0,1)$, let $q_A:=1-p_A$, $s_B:=1-r_B$, \begin{equation} f(X):=\pm\frac{1_{X\in A}-p_A}{\sqrt{p_A q_A}}, \quad g(Y):=\frac{1_{Y\in B}-r_B}{\sqrt{r_B s_B}}. \end{equation} Then the condition $\rho_m(X;Y) < 1/d$ implies \begin{equation} |P(X\in A,Y\in B)-P(X\in A)P(Y\in B)|\le\tfrac1d\,\sqrt{P(X\in A)P(X\notin A)P(Y\in B)P(Y\notin B)}. \end{equation} The latter inequality holds even when $p_A\in\{0,1\}$ and/or $r_B\in\{0,1\}$.

So, if $d$ is large, then the dependence between $X$ and $Y$ is very weak.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.