3
$\begingroup$

Let $X,Y$ be two (dependent) random variables with $\mathbb{P}(X\ge 0)=\mathbb{P}(Y\ge 0)=1$.

I want to find a tight lower bound of $\mathbb{E}(XY)$ when $X,Y$ are non-negative, almost surely.

Note that one trivial lower bound is $0$. But I want to know if there is an elegant method in literature for finding a tight lower bound, which has information of only the individual moments of the random variables $X,Y$.

I realize that the following lower bound can be obtained simply by using Cauchy-Scwartz inequality $$\mathbb{E}(XY)\ge \mu_X\mu_Y-\sigma_X\sigma_Y,$$ where $\mu_X=\mathbb{E}X,\ \sigma_X^2=\mathrm{Var}(X)$, and similarly $\mu_Y,\sigma_Y$ defined for $Y$. However, I don't know that if $X,Y$ are non-negative, almost surely, then whether this lower bound is non-negative. A little bit of algebra shows that this claim is equivalent to the following claim $$\frac{\mu_X^2}{\mu_{X^2}}+\frac{\mu_Y^2}{\mu_{Y^2}}\ge 1.$$ However, when $X=Y$, this claim is equivalent to proving that $$\mu_X^2\ge \frac{\mu_{X^2}}{2},$$ which, however, is false in general because Holder's inequality implies that $|\mu_X|\le \frac{\sqrt{\mu_{X^2}}}{\sqrt{2}}$. Any ideas how I can proceed? Thanks in advance.

$\endgroup$
3
  • $\begingroup$ In what terms do you want the tight lower to be? $\endgroup$ Apr 7 '19 at 1:39
  • $\begingroup$ I want the lower bound to be at least non-negative and to be equal to $\mu_{X^2}$ when $X=Y$. $\endgroup$ Apr 7 '19 at 4:51
  • $\begingroup$ "I want the lower bound to be at least non-negative and to be equal to $\mu_{x^2}$ when X=Y" I interpret X=Y as meaning that $X=Y$ in distribution, because, you are assuming no knowledge of the dependency between $X$ and $Y$, so you can not also assume $X=Y$ a.s. Given that interpretation, why would the quoted statement be true? For instance, let $X = 1$ or $3$, w.p. $1/2, 1/2$. Then $EX^2 = 9/2$. Then $Y = 4 - X$ has the same distribution as $X$, but $E(XY) = 3 \ne E(X^2)$. $\endgroup$ Apr 7 '19 at 11:07
3
$\begingroup$

$\newcommand{\de}{\delta} \newcommand{\si}{\sigma} \newcommand{\ep}{\varepsilon}$ Let us present the exact lower bound on $EXY$ in terms of $\mu_1:=\mu_X$, $\mu_2:=\mu_Y$, $\si_1:=\si_X$, $\si_2:=\si_Y$, as follows:

The minimum of $EXY$ over all nonnegative random variables (r.v.'s) $X$ and $Y$ with prescribed positive values of $\mu_1$, $\mu_2$, $\si_1$, $\si_2$ is \begin{equation*} (\mu_1\mu_2-\si_1\si_2)_+, \tag{1} \end{equation*} where $u_+:=\max(0,u)$.

Indeed, by rescaling, without loss of generality \begin{equation*} \mu_1=\mu_2=1. \end{equation*} Let \begin{equation*} v_j:=1+\mu_j^2,\quad r_j:=1/v_j; \end{equation*} everywhere here $j\in\{1,2\}$. Observe that \begin{equation*} \mu_1\mu_2-\si_1\si_2\le0\iff\si_1\si_2\ge1\iff r_1+r_2\le1. \tag{2} \end{equation*} Consider the two possible cases, according to this observation.

Case 1: $r_1+r_2\le1$. Then the bound (1) is $0$. On the other hand, let $S:=\{0,1,2\}$ with the probability measure on $S$ assigning masses $1-r_1-r_2,r_1,r_2$ to the points $0,1,2$, respectively. Let r.v.'s $X$ and $Y$ be defined on $S$ as follows: \begin{equation*} X(0)=X(2)=0,\ X(1)=v_1,\quad Y(0)=Y(1)=0,\ Y(2)=v_2. \end{equation*} Then \begin{equation*} EX=EY=1,\quad \operatorname{Var}X=\si_1^2,\quad \operatorname{Var}Y=\si_2^2,\quad EXY=E0=0. \end{equation*} So, the bound $(1)$ is exact in Case 1.

Case 2: $r_1+r_2\ge1$. Then the bound (1) follows indeed by Cauchy--Schwarz: \begin{equation*} EXY-\mu_1\mu_2=E(X-\mu_1)(Y-\mu_2)\ge-\si_1\si_2. \end{equation*} By (2), in Case 2 we have $\si_1\si_2\le1$. So, we can find $p\in(0,1)$ such that for $q:=1-p$ we have \begin{equation*} \si_2^2\le p/q\le1/\si_1^2. \end{equation*} Let now $S:=\{0,1\}$ with the probability measure on $S$ assigning masses $q,p$ to the points $0,1$, respectively. Let r.v.'s $X$ and $Y$ be defined on $S$ as follows: \begin{equation*} X(0)=1-\si_1\sqrt{p/q},\quad X(1)=1+\si_1\sqrt{q/p}, \end{equation*} \begin{equation*} Y(0)=1+\si_2\sqrt{p/q},\quad Y(1)=1-\si_2\sqrt{q/p}. \end{equation*} Then $X,Y\ge0$, \begin{equation*} EX=EY=1,\quad \operatorname{Var}X=\si_1^2,\quad \operatorname{Var}Y=\si_2^2,\quad EXY=1-\si_1\si_2. \end{equation*} So, the bound $(1)$ is exact in Case 2 as well. $\Box$

You were pretty close to this answer.

$\endgroup$
1
  • $\begingroup$ Great! Thanks for the full answer. $\endgroup$ Apr 8 '19 at 9:12

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.