3

Let ${X_n}$ be an IID sample such that ${X_i} \sim N(\mu,\sigma^2)$. When both $\mu$ and $\sigma$ are unknown, we construct $t(\hat{\mu},s)=\dfrac{\sqrt{n}(\hat{\mu}-\mu)}{s}$, where $s$ is the sample standard deviation.

The statistic $t(\hat{\mu},s)$ follows t-distribution exactly. I wish to know the asymptotic distribution of $t(\hat{\mu},s)$.

I understand that by Continuous Mapping theorem under certain conditions, if $\hat{\mu} \xrightarrow{d} X \sim N(\mu,\sigma^2/n)$ then $h(\hat{\mu}) \xrightarrow{d} h(X)$. However, in case of $t(\hat{\mu},s)$, it is a function of two random variables. So how to derive its asymptotic distribution?

I also checked Slutsky's theorem. That also requires that at least one of $\hat{\mu}$ and $s$ should converge to a constant. Now by CLT, both converges in distribution to a random variable with normal distribution. However, by combining LLN we may say that $s \xrightarrow{p} \sigma$. Would it be the right way to go?

EDIT: As pointed out in the second comment this is not a duplicate of this. Seemingly $s$ converged in probability to a constant but to a random variable in distribution. This was the main doubt that whether it was okay to ignore the convergence in distribution.

Dayne
  • 2,113
  • 1
  • 7
  • 24
  • 1
    Possible duplicate of [Is there a theorem that says that $\sqrt{n}\frac{\bar{X} - \mu}{S}$ converges in distribution to a normal as $n$ goes to infinity?](https://stats.stackexchange.com/questions/97268/is-there-a-theorem-that-says-that-sqrtn-frac-barx-mus-converges-in) – StubbornAtom Mar 30 '19 at 19:26
  • 1
    @StubbornAtom : I don't agree that this is a duplicate, since this question asks where the gaps are in a proposed argument, rather than just asking what the argument is. – Michael Hardy Mar 30 '19 at 20:09
  • What is the difference between $\hat \sigma$ and $s$? – Alecos Papadopoulos Apr 01 '19 at 14:35
  • No difference. Used $\hat{\sigma}$ by mistake. Corrected now. – Dayne Apr 02 '19 at 00:52

1 Answers1

2

It is not correct that $\widehat\mu$ and $s$ fail to converge in distribution to constants, nor that CLT implies that.

Rather, what CLT says is:

\begin{align} & \frac{\widehat\mu - \mu}{\sigma/\sqrt n} \overset d \longrightarrow \operatorname N(0,1) \\[8pt] \text{and } & \sqrt{n-1} \left. \left(\dfrac{s^2}{\sigma^2}-1\right) \right/\!\!\sqrt 2 \overset d \longrightarrow \operatorname N(0,1). \end{align}

It can be shown to follow that $\widehat\mu$ and $s$ converge in distribution to $\mu$ and $\sigma$ respectively.

And convergence in distribution to a constant entails convergence in probability to that constant.

Michael Hardy
  • 7,094
  • 1
  • 20
  • 38
  • In a more general case, if say neither converged to a constant then is there a way to get the asymptoptic distribution of their function? – Dayne Mar 30 '19 at 23:32