1: \begin{proof}

2: 	Let

3: 	\[X_{M}:=\frac{\frac{1}{M}\sum\limits_{i=1}^{M}\frac{1}{f_{\G}(Y_{i})}}{|\X_{\G}|},\]

4: 	then

5: 	\[E[X_{M}]=\frac{1}{|\X_{\G}|}E[\frac{1}{f_{\G}(Y_{1})}]=\frac{1}{|\X_{\G}|}\sum_{x}\frac{p_{\G}(x)}{f_{\G}(x)}=\frac{1}{Z_{\G}},\]

6: 	and

7: 	\[\var[X_{M}]=\frac{1}{M|\X_{\G}|^{2}}\var[\frac{1}{f_{\G}(Y_{1})}].\]

8: 	From (\ref{eq:zot}) we can rewrite $\log(\zot_{\G})$ as

9: 	\[

10: 	\log(\zot_{\G}) = g(X_{M}), 

11: 	\]

12: 	where $g(x):=\log(\frac{1}{x})=-\log(x)$, and so $g'(x)=\frac{-1}{x}$.

13: 	Using Taylor expansion of $g$ at $E[X_{M}]$, 

14: 	\[

15: 	\log(\zot_{\G})\simeq g(E(X_{M})) + g'(E(X_{M}))(X_{M}-E(X_{M})),

16: 	\]

17: 	and so

18: 	\begin{eqnarray*}

19: 	\var[\log(\zot_{\G})]&& \simeq \big(g'(E(X_{M}))\big)^{2}\var[X_{M}] \\

20: 	&& = \frac{1}{(E[X_{M}])^{2}} \var[X_{M}] \\

21: 	&& = \frac{Z_{\G}^{2}}{M|\X_{\G}|^{2}} \var[\frac{1}{f_{\G}(Y_{1})}] \\

22: 	\end{eqnarray*}

23: The approximation is only valid in the limit, as $g$ may be approximated as a linear function only when the

24: variance of $X_{M}$ is small.

25: This method of first order approximation is often referred to as the \emph{delta method}. For a more rigorous discussion 

26: on the delta method, see e.g. \cite[Theorem~5.5.24]{Casella:Inference}.

27: %Similarly, one can show

28: %\[

29: %	\lim_{M\rightarrow \infty}M \var[\zot_{\G}(M)] =

30: %	\frac{Z_{\G}^{4}}{|\X_{\G}|^{2}}\var\big[\frac{1}{f_{\G}(Y_1)}\big],

31: %\]

32: %where in this case, one take the function $g$ as $g(x) = \frac{1}{x}$.

33: \end{proof}

34: