This came from an old qualification exam for measure theory:

Suppose $(X,M,\mu)$ is a measure space and $E_1,\ldots, E_N\in M$ with $\mu(E_j\cap E_k)\leq \mu(E_j)/N$ for each $j\neq k$. Show that $$\mu\left(\bigcup_{j=1}^{N}E_j\right) \geq \frac{1}{6}\sum_{j=1}^{N}\mu(E_j)$$

Thoughts: I thought that for $E_1,\ldots,E_n\in M$ that $$\mu\left(\bigcup_{1}^{n}E_j\right)\leq \sum_{1}^{n}\mu(E_j)$$ when $E_j$'s are not disjoint. I know we can use the disjointification trick to get equality. I am not sure how to show the latter though. Any suggestions are greatly appreciated.


Here is my try:

1) Prove by induction that: $$\mu\left(\bigcup_{j=1}^N E_i\right)\ge \sum_{j=1}^N \mu(E_j)-\sum_{i<k}\mu(E_iE_k)$$

2) From $\mu(E_i E_k)\leq \mu(E_i)/N$ we get also $\mu(E_i E_k)\leq \mu(E_k)/N$, so: $$\mu(E_iE_k)\le\frac{\mu(E_i)+\mu(E_k)}{2N}$$

3) Sum the previous inequality for $i<k$ to get: $$\sum_{i< k}\mu(E_iE_k)\le\frac{N-1}{2N}\sum_{j=1}^N \mu(E_j)$$

4) Plug in the Step 1 to get: $$\mu\left(\bigcup_{j=1}^N E_i\right)\ge \frac{N+1}{2N}\sum_{j=1}^N \mu(E_j)$$

5) Trivially prove that $\frac{N+1}{2N}\ge\frac{1}{6}$ (so it looks like you could very well have $1/2$ instead of $1/6$)

| cite | improve this answer | |
  • 1
    $\begingroup$ And the factor $\frac{N+1}N$ is best possible, i.e., it can be attained for each $N.$ $\endgroup$ – bof Jan 5 '17 at 2:35
  • $\begingroup$ Are these steps I need to prove? If so, I am not sure how to prove 1) $\endgroup$ – Wolfy Jan 6 '17 at 22:19
  • $\begingroup$ @Wolfy Plain ol' induction. Induction step: $\mu(\bigcup_{j=1}^{n+1}E_j)=\mu(\bigcup_{j=1}^nE_j)+\mu(E_{n+1})-\mu(\bigcup_{j=1}^n(E_jE_{n+1}))\ge\sum_{i=1}^n\mu(E_j)-\sum_{1\le i<k\le n}\mu(E_iE_k)+\mu(E_{n+1})-\sum_{j=1}^n\mu(E_jE_{n+1})=\sum_{i=1}^{n+1}\mu(E_j)-\sum_{1\le i<k\le n+1}\mu(E_iE_k)$. That's all. $\endgroup$ – Momo Jan 7 '17 at 4:25

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.