5

This is question 5 is from Staudte and Sheather (1990), Robust estimation and testing.

Let $X_1,\ldots , X_n$ be i.i.d with $$ F_\theta = F(\frac{x}{\theta}),\quad x>0;\theta>0.$$ Assume that $T_n = T_n(X_1,\ldots ,X_n)$ is scale equivariant. Show that $$\mathbb{E}[T_n] = \theta \mathbb{E}_1[T_n]$$.

Using $\int_a^b f(u) du = \sum_{k=a}^b$ and $X\rightarrow \theta X$, my attempt at the question is as follows. \begin{aligned} \mathbb{E}[T_n] &= \int T_n dF(X_n) \\ &= \sum_x T_n P(X) &= \frac{1}{n}[\theta X_1 +,\ldots ,+ \theta X_n] \\ &= \theta\frac{1}{n}[ X_1 +,\ldots ,+ X_n] \\ &= \theta \sum_x T_n P(X) \\ &=\theta \int T_n dF(X_n) \\ &= \theta \mathbb{E}[T_n] \end{aligned}

Is my working out correct? If not, where and why was I wrong?

I should also point out that this is not an assignment question. I need to study this book to gain some background knowledge in robust statistics.

whuber
  • 281,159
  • 54
  • 637
  • 1,101
user7045
  • 133
  • 5
  • 2
    I don't follow some of the equations you've written; some things appear to be missing. The key here is to understand the definitions of *scale family* and *scale equivariant* and then to connect the two. Very little is needed in the way of explicit calculations. – cardinal Aug 20 '12 at 13:41
  • @cardinal, I do not fully understand the topic and that is probably why the question is unclear. But a very useful comment; I will dig deeper into the scale family – user7045 Aug 20 '12 at 16:27
  • No problem. My previous comment was a bit terse. What I meant was more that things like $\int_a^b f(u) \,\mathrm du = \sum_{k=a}^b$ and $[X_1+,\ldots,+X_n]$ are mathematical notation, but with an unclear meaning (to me). (For example, in the first case, there is no actual quantity on the right-hand side.) What you are asking *is* clear, though, and MansT has indicated the right direction to take. If you have follow-up questions as you try to understand the concepts, please consider posting them. :) – cardinal Aug 20 '12 at 16:31

2 Answers2

10

Your random variables belong to what is called a scale family. The first step should be to show that if $X\sim F_1$ then $\theta\cdot X\sim F_\theta$.

A statistic $T_n(X_1,\ldots,X_n)$ is usually said to be scale invariant if $$T_n(\theta X_1,\ldots,\theta X_n)=T_n(X_1,\ldots,X_n),$$ i.e. if rescaling the data leaves the statistic unchanged, but sometimes it is taken to mean that $$T_n(\theta X_1,\ldots,\theta X_n)=\theta T_n(X_1,\ldots,X_n),$$ i.e. that the statistic "scales in the right way", which seems to be the case here. If I understand your notation correctly, you wish to show that this is true for its expected value as well.

Let $X\sim F_\theta$ and $Y=\frac{1}{\theta}X\sim F_1$. Then you can show that

$$\mathbb{E}_\theta(T_n)=\mathbb{E}_\theta T_n(X_1,\ldots,X_n)=\mathbb{E}_\theta\theta T_n\Big(\frac{1}{\theta}X_1,\ldots,\frac{1}{\theta}X_n\Big)\\=\mathbb{E}_1 \theta T_n(Y_1,\ldots,Y_n)=\theta\cdot \mathbb{E}_1(T_n).$$

The crucial step is to show the equality in the line break.

MånsT
  • 10,213
  • 1
  • 46
  • 65
  • 1
    Did you notice that the title and the question use "equivariant," not "invariant"? – whuber Aug 20 '12 at 16:19
  • @manst, this is quit helpful. I will attempt the first step and move on from there. – user7045 Aug 20 '12 at 16:37
  • 1
    @whuber: yes, but it was tagged with invariance and I've seen the two words used as synonyms. Just wanted to avoid possible misunderstandings :) – MånsT Aug 20 '12 at 16:47
  • Good point about the tag. The words aren't synonyms, as you ably discuss. – whuber Aug 20 '12 at 17:40
2

For completeness here is the solution (adding in the missing steps that @manst wanted me to think about).

\begin{align*} \mathbb{E}_\theta[T_n] &= \int \ldots \int t_n (x_1, \ldots , x_n) f_\theta(x_1), \ldots , f_\theta(x_n) \text{d}x_1, \ldots , \text{d}x_n \\ \end{align*} Since $F_\theta(x) = F(\frac{x}{\theta})$ we can write $f_\theta(x) = \frac{1}{\theta}f(\frac{x}{\theta})$, where $f(u) = F'(u)$ \begin{align*} &= \int \ldots \int t_n (x_1, \ldots , x_n) \frac{1}{\theta}f\left(\frac{x_1}{\theta}\right), \ldots , \frac{1}{\theta}f\left(\frac{x_n}{\theta}\right) \text{d}x_1, \ldots , \text{d}x_n \\ \end{align*} let $\textbf{u} = \textbf{x}/\theta$ or $\theta \textbf{u} = \textbf{x}$ \begin{align*} &= \int \ldots \int t_n (\theta u_1, \ldots , \theta u_n) \frac{1}{\theta} f(u_1), \ldots , \frac{1}{\theta}f(u_n) \theta du_1, \ldots , \theta du_n \\ &= \int \ldots \int t_n (\theta u_1, \ldots , \theta u_n) \ f(u_1), \ldots , f(u_n) du_1, \ldots , du_n \\ &= \theta \int \ldots \int t_n ( u_1, \ldots , u_n) \ f(u_1), \ldots , f(u_n) du_1, \ldots , du_n \\ &= \theta \mathbb{E}_1[T_n] \end{align*}

user7045
  • 133
  • 5