0
$\begingroup$

Suppose that we have random variables $X_1$, $X_2$ each drawn independently from Irwin Hall Distribution with same mean of 0 and different variances.

If I have $B_1$ and $B_2$, which are high probability upper bound of $X_1$ and $X_2$ respectively, can I presume the upper bound of $X_1 + X_2$?

In other words, If $Pr[X_1 \gt B_1] \lt \epsilon$, $Pr[X_2 \gt B_2] \lt \epsilon$ holds with negligible probability $\epsilon$, Can I find the tight upper bound $B$ s.t. $Pr[X_1 + X_2 \gt B] \lt \epsilon$?

I thought of two candidates for upper bound B. One is $B_1 + B_2$ and the other is $\sqrt{B_1^2+B_2^2}$.

I thought that first candidate must satisfy the upper bound of $X_1 + X_2$, But it would be a loose bound. Second one would be better and tight upper bound, but I can't prove $Pr[X_1 + X_2 \gt \sqrt{B_1^2+B_2^2}] \lt \epsilon$

I got idea from variance of sum of random variables. If $X_1$ ~ $(0, \sigma_1^2)$ , $X_2$ ~ $(0, \sigma_2^2)$, $X_1 + X_2$ will follow the distribution of $(0, \sigma_1^2 + \sigma_2^2)$ Is this thought reasonable?

More generally, I also want to know the upper bound of $X_1$, $X_2$, ... , $X_n$. If the second candidate is true, I would like to know how to prove it.

Thank you for reading.

$\endgroup$
1
  • $\begingroup$ The distribution of the sum $Y = X_1+X_2$ is given by $P_Y(y) = \int P_{X_1}(x)P_{X_2}(y-x)dx.$ Your high probability upper bound then follows from solving $\int_{B}^\infty P_Y(y)dy < \epsilon. $ Suggest to set this equation up and try to get $B$ from it. Then you can generalize to the sum of $n$ variables $\endgroup$ Commented Aug 3, 2023 at 6:43

1 Answer 1

0
$\begingroup$

The event $X_1+X_2 > B_1+B_2$ is contained in the union of the two events $X_1>B_1$ and $X_2>B_2$. (This assertion is equivalent to the algebraic implication "if $X_1+X_2>B_1+B_2$ then $X_1>B_1$ or $X_2>B_2$", which is equivalent to the easily verified "if $X_1\le B_1$ and $X_2\le B_2$ then $X_1+X_2\le B_1+B_2$".)

In particular, if $Pr[X_1>B_1]<\varepsilon$ and $Pr[X_2>B_2]<\varepsilon$, then $$ Pr[X_1+X_2>B_1+B_2] \le Pr[X_1>B_1] + Pr[X_2>B_2] < 2\varepsilon. $$ It's not hard to show by example that this bound is best possible: let $X_j$ take only values $B_j-1$ and $B_j+2$ with probabilities $1-\delta$ and $\delta$ where $\delta$ is just less than $\varepsilon$, anticorrelated so that $X_1=B_1+2$ and $X_2=B_2+2$ never occur simultaneously.

$\endgroup$

You must log in to answer this question.

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