9
$\begingroup$

Given two highly correlated random variables $X$ and $Y$, I'd like to bound the probability that the difference $ |X - Y| $ exceeds some amount: $$ P( |X - Y| > K) < \delta $$

Assume for simplicity that:

  • The correlation coefficient is known to be "high", say : $ \rho_{X,Y}= {covar(X,Y)} / {\sigma_X \sigma_Y} \geq 1 - \epsilon $

  • $X,Y $ are zero mean: $ \mu_x = \mu_y = 0 $

  • $-1 \leq x_i, y_i \leq 1$ (or $ 0 \leq x_i, y_i \leq 1$ if that's any easier)

  • (If it makes things easier, let's say $X,Y $ have identical variance: $\sigma_X^2 = \sigma_Y^2 $)

Not sure how feasible it is to derive a bound on the difference given only the above information (I certainly couldn't get anywhere). A specific solution (if any), mandatory additional restrictions to impose on the distributions, or just advice on an approach would be great.

$\endgroup$
9
$\begingroup$

Even without those simplifying assumptions, a bound can be obtained by combining a couple of usual tools:

In some detail:

$$\sigma^2_{X-Y}=\sigma^2_X+\sigma^2_Y-2·cov(X,Y)$$

$$cov(X,Y)=\sigma_X·\sigma_Y·\rho_{XY}$$

$$\sigma^2_{X-Y}=\sigma^2_X+\sigma^2_Y-2·\sigma_X·\sigma_Y·\rho_{X,Y}$$

According to Chebyshev's inequality, for any random variable $Z$:

$$ \Pr(|Z-\mu|\geq k\sigma) \leq \frac{1}{k^2}$$

Then (and using that $\mu_{X-Y}=\mu_X-\mu_Y)$:

$$ \Pr(|X-Y-\mu_X+\mu_Y|\geq k·\sqrt{\sigma^2_X+\sigma^2_Y-2·\sigma_X·\sigma_Y·\rho_{X,Y}}) \leq \frac{1}{k^2}$$

We can use the proposed simplifying assumptions to get a simpler expression. When:

$$\rho_{X,Y}= {covar(X,Y)} / {\sigma_X \sigma_Y} = 1 - \epsilon $$ $$\mu_x = \mu_y = 0$$ $$\sigma_X^2 = \sigma_Y^2 = \sigma^2$$

Then:

$$\sigma^2_X+\sigma^2_Y-2·\sigma_X·\sigma_Y·\rho_{X,Y} = 2·\sigma^2·(1-(1-\epsilon)) = 2\sigma^2\epsilon$$

And therefore:

$$\Pr(|X-Y|\geq k·\sigma\sqrt{2\epsilon}) \leq \frac{1}{k^2}$$

Interestingly, this result holds even if $\epsilon$ is not small, and if the condition for correlation changes from $=1-\epsilon$ to $\geq 1-\epsilon$, the result doesn't change (because it's already an inequality).

$\endgroup$

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.