How many independent measurement of CPU-time are required such that the difference $|\bar{X} - \mu|<0.1$ with probability $0.9$ at least?

93 Views Asked by At

The actual required CPU-time of a workstation session is assumed (due to a long-term study) as a random variable with unknown expected value $\mu$ and known variance $\sigma^2=6.25[s^2]$. How many independent measurement of CPU-time are needed such that with probability of at least $0.9$, you have that the difference $|\bar{X}-\mu|$ is less than $0.1$?

This is task from old exam with different teacher. I ask my teacher and he say he don't show this example in class but give me two method for solve it: Central limit theorem and Chebyshev's inequality. This already help me much and I try solve like this:

I use central limit theorem:

We have that $$\sqrt{n} (\bar{X}-\mu) \text{ for } n \rightarrow \infty \text { is approximately standard normally distributed}$$

Now assume we have very big $n$ such that approximation mistake is very small so we don't need to keep it in mind. Then we can do

$$P(|\bar{X}-\mu|) < \varepsilon) = P(\sqrt{n}|\bar{X}-\mu|<\sqrt{n} \varepsilon)=P(-\sqrt{n}\varepsilon< \sqrt{n}(\bar{X}-\mu)< \sqrt{n} \varepsilon) \\ \approx \Phi(\sqrt{n} \varepsilon)- \Phi(-\sqrt{n}\varepsilon) = 2\Phi(\sqrt{n}\varepsilon)-1$$

Now insert $\varepsilon= 0.1$ in inequality $P(|\bar{X}-\mu|<\varepsilon)\geq 0.9$ then we have

$$2\Phi(0.1 \cdot \sqrt{n})-1 \geq 0.9 \Leftrightarrow \Phi(0.1 \cdot \sqrt{n}) \geq 0.95$$

And from this step I don't know how finish it :/

Can you tell me if this is correct till here and how to finish it? I really think I almost got it but something is missing..

But actually I'm more interested to know how you would do this with Chebyshev? I tried it already but also not sure if correct (see my comment of karakfa's answer).

1

There are 1 best solutions below

5
On BEST ANSWER

Chebyshev inequality won't help you here. With some overall assumptions you can use the normal approximation and work from the z-table.

The sample size you'll come up with should be

$$ n = (\frac{z_{\alpha/2} \sigma}{E})^2 $$

for 90% CI, $z_{\alpha/2} \approx 1.65$, and $E$ is given as $0.1$, $\sigma^2=6.25$

$ n \approx 1700$