8
$\begingroup$

This is Lyapunov's inequality for moments of a random variable (the paper can be accessed here):

Let $\{F_{k}\}, k = 1, 2, ..., N$ be an arbitrary sequence of events in $(\Omega, F, P)$.

We have, if $P\left(\bigcup\limits_{k=1}^{n} F_k \right) > 0$,

(1) $$2 \sum_{1\leq j < k \leq N} P(F_{j}F_k) \geq \Bigg[P\bigg(\bigcup_{k=1}^N F_k\bigg)\Bigg]^{-1}\Bigg(\sum_{k=1}^N P(F_k)\Bigg)^{2} - \sum_{k=1}^N P(F_k)$$

Proof: Define r.v. $ X_k(\omega)= \begin{cases} 0, & \text{if $\omega \notin F_k$} \\[2ex] 1, & \text{if $\omega \in F_k$} \end{cases}$

The following identity is evident:

(2) $2 \sum\limits_{1\leq j < k \leq N} P(F_j F_k) = E\left[\left(X_1+...+X_N\right)^{2}\right] - E\left(X_{1}^{2} +...+ X_{N}^{2}\right)$

Now by the Schwarz inequality we have

(3) $[E(X_1+...+X_N)]^2 \leq P(X_1+...+X_{N}>0)E[(X_{1}+...+X_N)]^2$

Since $E(X_k) = E(X_{k}^{2}) = P(F_k), $

$P(X_1+...+X_N>0) = P\left(\bigcup\limits_{k=1}^{N} F_k\right)$ by definition, (1) follows from (2) and (3)

How can I strengthen Lyapunov's inequality for moments of a random variable so that this new inequality:

$$P\bigg(\bigcup_{k=n}^{N} A_k\bigg) \geq \frac{\big(\sum_{k=n}^{N} P(A_k)\big)^{2}}{\sum_{k,j=n}^{N} P(A_k A_j)}$$ holds?

$\endgroup$
5
  • $\begingroup$ I don't quite understand. The inequality you displayed/proved is equivalent to the final inequality you hope to prove? Or are you talking about applying Chung-Erdos for $P(\limsup_n A_n)$? $\endgroup$
    – Tom Chen
    Mar 1 '19 at 15:36
  • $\begingroup$ basically in Lyapunov's inequality, i have an extra term, $\sum_{k=1}^n F_k$ being subtracted and I don't know how i can strengthen that inequality so that term isn't there in the inequality I eventually want to show $\endgroup$
    – user477465
    Mar 1 '19 at 15:54
  • $\begingroup$ But $\sum_{j,k=1}^{N}P(F_j F_k) = 2\sum_{1 \le j < k \le N}^{N}P(F_j F_k) + \sum_{j=1}^{N}P(F_j)$. That extra term should be there, because it would actually be absorbed into the unexpanded sum $\sum_{j,k=1}^{N}P(F_j F_k)$ (with $F$'s replaced with $A$'s) in your final inequality. $\endgroup$
    – Tom Chen
    Mar 1 '19 at 16:19
  • $\begingroup$ im really sorry, i don't know what you mean. can you please explain how it would be absorbed? i'm just confused about that term $\endgroup$
    – user477465
    Mar 1 '19 at 16:33
  • $\begingroup$ No problem, OP. I edited my answer below to reflect what I mean. Let me know if this is clear enough? $\endgroup$
    – Tom Chen
    Mar 1 '19 at 16:55
1
+50
$\begingroup$

The inequality $$2 \sum_{1\leq j < k \leq N} P(F_{j}F_k) \geq \Bigg[P\bigg(\bigcup_{k=1}^N F_k\bigg)\Bigg]^{-1}\Bigg(\sum_{k=1}^N P(F_k)\Bigg)^{2} - \sum_{k=1}^N P(F_k)$$ can be rewritten as \begin{align*} P\bigg(\bigcup_{k=1}^N F_k\bigg) \ge \frac{\Bigg(\sum_{k=1}^N P(F_k)\Bigg)^{2}}{2 \sum_{1\leq j < k \leq N} P(F_{j}F_k) + \sum_{k=1}^N P(F_k)} \end{align*} Note that \begin{align*} \color{red}{2 \sum_{1\leq j < k \leq N} P(F_{j}F_k)} + \color{blue}{\sum_{k=1}^N P(F_k)} &= \color{red}{\sum_{j \neq k} P(F_{j}F_k)} + \color{blue}{\sum_{j = k}P(F_jF_k)} \\ &=\sum_{j,k} P(F_{j}F_k) \end{align*} Note that the red expressions equal because $P(F_j F_k) = P(F_k F_j)$, but since $j < k$ in the first sum, the $2$ doubles for $j > k$ as well. The blue expressions equal because $P(F_k \cap F_k) = P(F_k)$.

And so \begin{align*} P\bigg(\bigcup_{k=1}^N F_k\bigg) \ge \frac{\Bigg(\sum_{k=1}^N P(F_k)\Bigg)^{2}}{\sum_{j,k=1}^{N}P(F_j F_k) } \end{align*} We can reindex the lower index of $1$ to $n$, and change variables $F$ to $A$, and we have the desired final inequality.

Proving $P(\limsup_n A_n) \ge 1/C$ with this bound.

Here is the result I will prove:

If $\sum_{n=1}^{\infty}P(A_n) = \infty$, then
\begin{align*} P\left(\limsup_n A_n\right) \ge \limsup_{n\rightarrow\infty}\frac{(\sum_{k=1}^{n}P(A_k))^2}{\sum_{i,j=1}^{n}P(A_iA_j)} \end{align*}
In particular, it implies a stronger version of the second Borel-Cantelli: If $P(A_i A_j) \le P(A_i)P(A_j)$ for all $i \neq j$, then $P\left(\limsup_n A_n\right) = 1$.

Proof. Let $a_n = (\sum_{k=1}^{n}P(A_k))^2$ and $b_n = \sum_{i,j=1}^n P(A_i A_j)$. By assumption, $a_n \rightarrow \infty$, and by Chung-Erdos, so does $b_n \rightarrow \infty$. Note that $(\sum_{k=m+1}^{n}P(A_k))^2 = (\sqrt{a_n}-\sqrt{a_m})^2$ and \begin{align*} \sum_{i,j = m+1}^{n}P(A_iA_k) = b_n - b_m - \sum_{i=1}^{n}\sum_{j=m+1}^{n}P(A_iA_j) - \sum_{i=1}^{n}\sum_{j=m+1}^{n}P(A_iA_j) \le b_n - b_m \end{align*} And so \begin{align*} P\left(\bigcup_{k=m+1}^{\infty}A_k\right) = \lim_{n\rightarrow \infty}P\left(\bigcup_{k=m+1}^{n}A_k\right) \ge \limsup_{n\rightarrow \infty}\frac{(\sqrt{a}_n-\sqrt{a}_m)^2}{b_n - b_m} = \limsup_{n\rightarrow \infty}\frac{a_n}{b_n} \end{align*} And so $P\left(\limsup_n A_n\right) = \lim_{m\rightarrow \infty}P\left(\bigcup_{k=m+1}^{\infty}A_k\right) \ge \limsup_{n\rightarrow \infty}\frac{a_n}{b_n}.$

$\endgroup$
5
  • $\begingroup$ thank you for being so patient with me. can you just explain how you know $2 \sum_{1\leq j < k \leq N} P(F_{j}F_k) + \sum_{k=1}^N P(F_k) = \sum_{j,k=1}^{N}P(F_k) $ this is true $\endgroup$
    – user477465
    Mar 1 '19 at 17:42
  • $\begingroup$ No worries. I edited my answer with more details on this step! And realized a typo; should be $\sum_{j,k=1}^{N} P(F_j F_k)$, not $\sum_{j,k=1}^{N} P(F_k)$ $\endgroup$
    – Tom Chen
    Mar 1 '19 at 19:14
  • $\begingroup$ i think theres a typo again because you said $\sum P(F_k F_k) = \sum P(F_k)$. also are you assuming these events are independent? bc they aren't supposed to be independent so how do you know that $P(F_k F_j) $ is just $P(F_k)$? $\endgroup$
    – user477465
    Mar 2 '19 at 5:29
  • $\begingroup$ is it true because they are indicator random variables? $\endgroup$
    – user477465
    Mar 2 '19 at 5:32
  • $\begingroup$ The blue portion is $\sum_{j=k} P(F_j F_k)$, not $\sum_{j,k=1}^{N} P(F_j F_k)$; that is, we sum over when $j = k$. So $\sum_{j=k} P(F_j F_k) = \sum_{k=1}^N P(F_k F_k) = \sum_{k=1}^{N} P(F_k)$. $\endgroup$
    – Tom Chen
    Mar 2 '19 at 13:46

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.