0
$\begingroup$

I've seen this used some times and I would like to ask what steps are taken on the way to getting there:

E.g. assuming bounded variance, we can use Chebyshev concentration inequality: for any $t>0$,

$$ P(\lvert X - \mu \rvert > t) \leq \frac{\sigma^2}{t^2} $$

What are the steps I would need to take to find the bound for the length of some interval around $\mu$ that contains probability mass $1-\eta$ of the distribution of $X$, for $0 \leq \eta \leq 1$?

For example

Given bounded fourth moment:

$$ E[(X-\mu)^4] \leq C_4 \bigg(E[(X-\mu)^2]\bigg)^2 $$

the following concentration inequality applies

$$ P(\lvert X - E[X] \rvert \geq t \sigma) \leq \frac{C_4}{t^4} $$

Then let $I_{1-\eta}$ be the interval around $\mu$ containing $1-\eta$ probability mass, then using the concentration inequality we know that:

$$ \text{length}(I_{1-\eta}) \leq \frac{C_4^{\frac{1}{4}}\sigma}{\eta^\frac{1}{4}} $$ What are the steps taken to arrive at this length bound? Taken from this paper, lemma 11 (concentration) and p43 (length)

$\endgroup$
  • $\begingroup$ No general bound exists, apart from the trivial lower bound of $0$ and upper bound of $+\infty:$ you need to specify more about how the interval is constructed or about the distribution of $X.$ Would you like to clarify your question? $\endgroup$ – whuber Mar 30 '19 at 21:16
  • $\begingroup$ @whuber thank you, I have added an example that led to this question $\endgroup$ – eeek Mar 31 '19 at 11:41
0
$\begingroup$

I think this might be how to think of it :

Rewrite the concentration inequality as

$$ \mathbb{P} \bigg( \lvert X - E[X] \rvert \geq t \sigma C_4^\frac{1}{4} \bigg) \leq \frac{1}{t^4} $$

then

$$ \mathbb{P} \bigg( \lvert X - E[X] \rvert \geq \eta^{-\frac{1}{4}} \sigma C_4^\frac{1}{4} \bigg) \leq \eta $$

then if you removed the $\eta$-quantile, the remaining $1-\eta$ mass of points would lie within

$$ \lvert X - E[X] \rvert \leq \frac{C_4^\frac{1}{4}}{\eta^\frac{1}{4}} \sigma $$

| cite | improve this answer | |
$\endgroup$

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.