Let $X_1,\ldots,X_n$ have the pdf given by $f(x;\theta)=(1/\sigma)e^{-(x-\mu)/\sigma}$. For the statistic $T$ given by $T=\sum_i (X_i - X_{n:1})$, I need to show that $T$ is sufficient for $\sigma$ when $\mu$ is known.

Previously, I have shown that $T'=\sum_i (X_i - \mu)$ is sufficient for $\sigma$ when $\mu$ is known. Is it enough for me to simply say that $T$ is sufficient because it is a one-to-one function of $T'$?


Preliminary note: You have not stated the support of your density, so I am going to assume it is $x \geqslant \mu$, since this gives a valid density function. Under this interpretation, this means that you have a shifted exponential distribution with shift parameter $\mu$ and scale parameter $\sigma$.

Firstly, the function $T$ is not a one-to-one function of $T'$, so no, you can't just assert that. Secondly, the result you are trying to prove is in fact incorrect; the statistic $T$ is not sufficient for $\sigma$ in this case. To prove sufficiency you would generally apply the Fisher-Neyman factorisation theorem, which says that $T$ is sufficient if and only if you can write the likelihood function as:

$$L_\mathbf{x}(\theta) \propto g(\theta, T(\mathbf{x})),$$

where $\theta$ is the unknown parameter. (In other words, the likelihood function depends on the data vector $\mathbf{x}$ only through the statistic $T(\mathbf{x})$.) Denoting $m_n \equiv \min (x_1,...,x_n)$ we have:

$$\begin{equation} \begin{aligned} L_\mathbf{x}(\mu,\sigma) &= \prod_{i=1}^n f(x_i|\mu,\sigma) \\[6pt] &= \prod_{i=1}^n \frac{1}{\sigma} \cdot \exp \Big(- \frac{x_i-\mu}{\sigma} \Big) \cdot \mathbb{I}(x_i \geqslant \mu) \\[6pt] &= \frac{1}{\sigma^n} \cdot \exp \Big(- \frac{n(\bar{x} - \mu)}{\sigma} \Big) \cdot \mathbb{I}(m_n \geqslant \mu) \\[6pt] &= \frac{1}{\sigma^n} \cdot \exp \Big(- \frac{n(\bar{x} - m_n)}{\sigma} - \frac{n(m_n - \mu)}{\sigma} \Big) \cdot \mathbb{I}(m_n - \mu \geqslant 0) \\[6pt] &= \frac{1}{\sigma^n} \cdot \exp \Big(- \frac{n(\bar{x} - m_n)}{\sigma} \Big) \cdot \exp \Big(- \frac{n(m_n - \mu)}{\sigma} \Big) \cdot \mathbb{I}(m_n - \mu \geqslant 0). \\[6pt] \end{aligned} \end{equation}$$

From this decomposition, we can see that the statistic $(\bar{x} - m_n, m_n)$ is minimal sufficient for both parameters, and also minimal sufficient for $\sigma$ when $\mu$ is known. (We can also see that the statistic $m_n$ is minimal sufficient for $\mu$ when $\sigma$ is known.)

| cite | improve this answer | |
  • $\begingroup$ The question doesn't ask whether $T$ is minimal sufficient. If you can express your statistic in terms of $T$--and it looks like you can when $\mu$ is known--then a fortiori $T$ is sufficient. $\endgroup$ – whuber Jun 14 '18 at 21:13
  • $\begingroup$ In this case $T = \bar{x} -m_n$, and from the above decomposition, this doesn't appear to me to be sufficient even when $\mu$ is known. If you disagree, can you show the proposed expression? $\endgroup$ – Ben Jun 14 '18 at 23:46

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.