Minimax posterior regret actions for exponential families of distributions and weighted squared error loss (Q4516143)

Let \({\mathcal X}\) denote the sample space, \(\Theta\) the (one-dimensional) parameter space, \({\mathcal L}\) a class of loss functions, \(\pi\) a fixed prior measure, and let \({\mathcal A}\) denote the class of Bayes actions with respect to \(\pi\) and \({\mathcal L}\). Define for a fixed \(x\in {\mathcal X}\) the posterior regret loss of an action \(a\in{\mathcal A}\) by \(\rho_L(a)=E^{\pi}L(\cdot,a) \), the posterior regret by \(G_L(a)=\rho_L(a)-\inf_{a'}\rho_L(a')\) and the minimax posterior regret (MPR) \(a_0\) by \[ \sup_{L\in{\mathcal L}}G_L(a_0)= \inf_{a'\in{\mathcal A}}\sup_{L\in{\mathcal L}}G_L(a'). \] The author studies the MPR's dependence on the choice of the weight function in the case \[ {\mathcal L}=\{L_r:r\in S\},\quad L_r(\theta,a)=w_r(\theta)(\theta-a)^2. \] In this case \(G_{L_r}(a)=(a-a_r)^2w_r\), where \(w_r=\int_\Theta w_r(\theta)\pi(d\theta|x)\). It is shown that, provided the function \(w_{r'}(\theta)/w_{r''}(\theta)\) is monotone in \(\theta\) for any \(r',r''\), the MPR action \(a_0\) is given by the solution of the equation \(g_1(r_0)=g_2(r_0)\), where \[ g_1(r')=\sup_{r<r'}(a_{r'}-a_r)^2w_r,\quad g_2(r')=\sup_{r>r'}(a_{r'}-a_r)^2w_r. \] This result is applied to a one-parameter family of exponential distributions.