0
$\begingroup$

If there are 3 vectors X, Y, Z of the same length, for any $x_i \in X,y_i \in Y,z_i \in Z$, we have $0<x_i<1,0<y_i<1,0<z_i<1$.

The correlation between Z, Y is greater than between X, Y. The inequality could be formulated as $\rho(Z, Y)>\rho(X, Y)$.

I want to ask if the inequality exists: $\rho(XZ, Y)>\rho(X, Y)$. If yes, please demonstrate it. If not, please give an example.

If you couldn't demonstrate it, please give me some clues or references.

$\endgroup$

1 Answer 1

2
$\begingroup$

The inequality does not always hold. Let $D \sim$ Bernoulli$(p)$, independent of $Y$. Let $Y'$ and $Y''$ be two independent copies of $Y$ (also independent of $Y$). Let $X=D Y+(1-D) Y'$ and $Z=DY'' + (1-D)Y$. Define also $m_k=E(Y^k)$ for $k=1,2$. Then it is easy to check that Cov$(X,Y) = p V(Y)$ and Cov$(Z,Y) = (1-p) V(Y)$. Because $X$ and $Z$ have the same distribution as $Y$, $$\rho(X,Y)=p < 1- p = \rho(Z,Y)$$ provided that $p<1/2$. Moreover, $XZ=Y Y'''$ where $Y'''$ is an independent copy of $Y$. Thus, Cov$(XZ,Y) = m_1 V(Y)$ and $V(XZ)=m_2^2 - m_1^4$. Hence, $$\rho(XZ,Y) = m_1 \sqrt{\frac{m_2 - m_1^2}{m_2^2 - m_1^4}} = \frac{m_1}{\sqrt{m_2 + m_1^2}}.$$ Consider a distribution of $Y$ such that $m_2> 3 m_1^2$. Such a distribution exists if $m_1 < 1/\sqrt{3}$ since there exist distributions with support included in $(0,1)$ such that $m_1 > m_2 \geq m_1^2$. For such a distribution, $\rho(XZ,Y)<1/2$ and any $p \in \left(m_1/\sqrt{m_2 + m_1^2}, 1/2\right)$ satisfies $\rho(X,Y) < \rho(Z,Y)$ and $\rho(XZ,Y) < \rho(X,Y)$.

$\endgroup$
4
  • $\begingroup$ Based on your assumption, we have ρ(XZ,Y) >= 0 = ρ(X,Y). (When E(x) =0, ρ(XZ,Y)=ρ(X,Y)). The inequality still holds. $\endgroup$
    – Mac Zhang
    Commented Dec 1, 2021 at 14:45
  • $\begingroup$ Yes, sorry, I first misread your problem. I have completely modified my answer. $\endgroup$
    – bdx77
    Commented Dec 1, 2021 at 14:49
  • $\begingroup$ Hi, many thanks for your answer. I got the same conclusion but the computation process is different from you. As you have already defined X, Z. So we have XZ= [DY+(1-D)Y1][Dy2 + (1-D)Y]. Moreover, we have Cov(XZ, Y) = E(D(1-D))V(Y) = 0, so 0=ρ(XZ,Y)<ρ(X,Y)=p. $\endgroup$
    – Mac Zhang
    Commented Dec 1, 2021 at 15:21
  • $\begingroup$ I don't think we have Cov(XZ, Y) = E(D(1-D))V(Y) with my construction. The covariance is positive (and equal to $m_1$ V(Y)) but small enough to ensure the final inequality. $\endgroup$
    – bdx77
    Commented Dec 1, 2021 at 15:55

Not the answer you're looking for? Browse other questions tagged .