= Solution
Given $X=x$, the <posterior expected loss> from reporting $d$ is
$$
\begin{aligned}
\mathbb E\!\left[\theta^{-1}(\theta-d)^2\mid X=x\right]
&=\mathbb E(\theta\mid X=x)-2d
+d^2\mathbb E(\theta^{-1}\mid X=x).
\end{aligned}
$$
This is a strictly convex quadratic in $d$ whenever the conditional inverse moment is finite and positive. Differentiating with respect to $d$ gives
$$
-2+2d\,\mathbb E(\theta^{-1}\mid X=x)=0,
$$
so the <Bayes estimator under reciprocal weighted quadratic loss> is
$$
\boxed{\delta_\pi(x)
=\big(\mathbb E(\theta^{-1}\mid X=x)\big)^{-1}}.
$$
Back to article page
