For $\alpha,\beta\ge 0$, let $X\in\{1,\alpha\}$ and $Y\in\{1,\beta\}$ be two random variables such that $$XY = \begin{cases} \alpha\beta \quad & \text{with probability} \quad p_{11}\\ \alpha \quad & \text{with probability} \quad p_{12}\\ \beta \quad & \text{with probability} \quad p_{21}\\ 1 \quad & \text{with probability} \quad p_{22}\\ \end{cases}$$ (where $p_{11}+p_{12}+p_{21}+p_{22}=1$.) Further define $p_1=p_{11}+p_{12}$ and $p_2=p_{11}+p_{21}$.

Further assume that $X$ and $Y$ are correlated, that is $p_{11} \ge p_1p_2$. (Note that $p_{11}=p_1p_2$ when $X$ and $Y$ are independent, in which case the following trivially holds.)

Define $s=\log\left(\frac{(1-p_1)(1-p_2)}{p_1p_2}\right){\big/}\log\left(\frac{p_{22}}{p_{11}}\right)$, then we would like to show that it holds for any $\alpha,\beta\ge 0$ that

$$E[XY]^s \le E[X^s]E[Y^s].$$

In other words, we want to show the simple real inequality

$$(\alpha \beta p_{11} + \alpha p_{12} + \beta p_{21} + p_{22})^s \le(\alpha^s p_1 + (1 - p_1))(\beta^s p_2 + (1 - p_2)).$$

One can easily check that the two sides are equal when $\alpha=\beta=1$ and when $\alpha=\beta=p_{22}/p_{11}$.

I conjecture that those two points are the only such positions. However, I haven't been able to show even that all solutions must satisfy $\alpha=\beta$.

The problem comes from the study of two-variable hyper-contractive inequalities and would help me greatly in proving certain algorithmic lower bounds. A weaker, but more general inequality was proved by Wolff, but it doesn't seem easily applicable to the above case. If anyone has suggestions for how I may proceed I'd be very grateful.

Update: I should notice that writing $X = f(r) = \hat f_\emptyset + \hat f_{\{1\}} r$ and $Y = g(r') = \hat g_\emptyset + \hat g_{\{1\}} r'$ for some $\{-1,1\}$ valued random variables $r$ and $r'$, one has that $E[XY] = \hat f_\emptyset \hat g_\emptyset + \sigma\rho \hat f_{\{1\}} \hat g_{\{1\}}$ $\le (f_\emptyset^2 + \sigma \hat f_{\{1\}}^2)^{1/2} (g_\emptyset^2 + \rho \hat g_{\{1\}}^2)^{1/2}$ $= \|T_\sigma f\|_2 \|T_\rho g\|_2 $ for any $\sigma\rho = \frac{p_{11} - p_1p_2}{\sqrt{p_1(1-p_1)p_2(1-p_2)}}$. Using a result by Oleszkiewicz one can then bound $\|T_\sigma f\|_2 \|T_\rho g\|_2 \le \|f\|_s \|g\|_t = E[X^s]^{1/s}E[Y^t]^{1/t}$ for values of $s$ and $t$ depending on $\sigma$ and $\rho$. Unfortunately, the application of Cauchy Schwarz appears to be too aggressive, and we don't get the conjectured result above.


The inequality in question is false when e.g. $a=b=0$, $p_{11}+p_{12}=1/2$, $p_{11}=1/10$ and $p_{12}=p_{21}=1/3$.

  • $\begingroup$ Sorry, I should add $b>xy$. For $b=xy$ we get equivalence (easy to check) and for $b<xy$ we should get the inverse inequality. $\endgroup$ – Thomas Dybdahl Ahle Aug 9 '18 at 17:01
  • $\begingroup$ You should update the question to include this condition. $\endgroup$ – Aryeh Kontorovich Nov 14 '18 at 13:07
  • $\begingroup$ I have updated the question with better notation and clearer assumptions. $\endgroup$ – Thomas Dybdahl Ahle Feb 24 '20 at 20:05

Define $q_1 = p_{11} + p_{12}$ and $q_2 = p_{11} + p_{21}$ and set $$ f(1) = \sqrt[s]{\frac{\frac{1}{2} + x}{q_1}} $$ $$ f(-1) = \sqrt[s]{\frac{\frac{1}{2} - x}{1 - q_1}} $$ $$ g(1) = \sqrt[s]{\frac{\frac{1}{2} + y}{q_2}} $$ $$ g(-1) = \sqrt[s]{\frac{\frac{1}{2} - y}{1 - q_2}} $$ where $x, y \in \left[-\frac{1}{2}, \frac{1}{2}\right]$, then your inequality is equivalent to showing that $$ p_{11} f(1)g(1) + p_{12}f(1)g(-1) + p_{21}f(-1)g(1) + p_{22}f(-1)g(-1) \le (q_1 f(1)^s + (1 - q_1)f(-1)^s)^{1/s}(q_2 g(1)^s + (1 - q_2)g(-1)^s)^{1/s} $$ The right hand side is equal to $1$ and the left hand side is equal to the function \begin{align} h(x, y) = A\left(\sqrt[s]{(\frac{1}{2} + x)(\frac{1}{2} + y)} + \sqrt[s]{(\frac{1}{2} - x)(\frac{1}{2} - y)}\right) + B\sqrt[s]{(\frac{1}{2} + x)(\frac{1}{2} - y)} + C\sqrt[s]{(\frac{1}{2} - x)(\frac{1}{2} + y)} \end{align} where $A = \frac{p_{11}}{\sqrt[s]{q_1 q_2}}$, $B = \frac{p_{12}}{\sqrt[s]{q_1(1 - q_2)}}$ and $C = \frac{p_{21}}{\sqrt[s]{q_2(1 - q_2)}}$. So if you can show that $h(x, y)$ is a quasi-concave function then you have the result.


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.