As the title said, for a given state space $\Theta$ and loss function $L(\theta,a)$, if a (randomized) decision rule $\delta$ is a minimax decision rule with respect to $L$, i.e. $$sup_{\theta}E^{X\mid\theta}L(\theta,\delta(X))=inf_{\delta\in\mathcal{D}}sup_{\theta}E^{X\mid\theta}L(\theta,\delta(X))$$ among a class of decision rules $\mathcal{D}=\{\delta:E^{X\mid\theta}L(\theta,\delta(X))<\infty\}$ and it is also admissible, i.e. for all $\delta\neq\delta_1\in\mathcal{D}$ $$E^{X\mid\theta}L(\theta,\delta(X))\leq E^{X\mid\theta}L(\theta,\delta_1(X)),\forall\theta \in \Theta$$ $$E^{X\mid\theta_0}L(\theta_0,\delta(X))\leq E^{X\mid\theta_0}L(\theta_0,\delta_1(X)),\exists\theta_0 \in \Theta$$ Then, must such a decision rule $\delta$ be unique as a function? If so, why; if not, is there any counter-example?

Moreover, what will the answer to the problem change if we assume that $\Theta$ is finite?


It depends on the sampling distribution and on the loss function as well as the dimension of the parameter. For instance, the MLE of a multivariate Gaussian mean is admissible when the dimension is $p=1,2$ under quadratic loss. Since its risk is constant, it is the sole admissible minimax estimator. This does not hold for $p\ge 3$, a phenomenon called the Stein effect. (Brown (1966) shows the wide generality of this phenomenon.)

In Strawderman (1972), the author shows the existence of proper Bayes [hence admissible] minimax estimators of the multivariate Normal mean for dimensions $p\ge 5$. One example of such estimators is associated with the priors$$\theta\sim\text{N}_p(0,\lambda^{-1}\lambda\mathbf{I}_p)\qquad \lambda\sim\lambda^{-a}/(1-a)\quad 0\le a\le 1$$

In Berger and Robert (1990), we extended this result and obtained a family of priors leading to admissible minimax estimators of the multivariate Normal mean. This was further extended by Berger and Strawderman (1996).

In Ghosh and Amin (1981), the authors exhibit a family of proper Bayes [hence admissible] minimax estimators of the p-variate Poisson mean, under the sum of weighted squared error losses, when $p\ge 3$, weights being reciprocals of variances. For instance, for a prior$$\pi(\lambda)\propto\lambda^{m-1}\{1+\lambda\}^{-(n+m)}\qquad 0< m\le p-2,\quad n>0,$$the associated posterior mean $$\mathbb{E}[\lambda|\mathbf{x}]=\dfrac{z+m+p}{z+m+n+p-1}\mathbf{x}\qquad z=\sum_{i=1}^p x_i$$is proper Bayes minimax admissible.

| cite | improve this answer | |
  • 2
    $\begingroup$ So the answer to the question is NO then. $\endgroup$ – kjetil b halvorsen Feb 26 '17 at 15:40
  • 1
    $\begingroup$ @Xi'an Thanks a lot, I accepted it a bit late since I went through your paper. $\endgroup$ – Henry.L Feb 28 '17 at 1:46
  • 1
    $\begingroup$ @Henry.L: I do not have an answer for the finite parameter space case, although I believe there should be unicity and hence admissibility of the minimax rule. $\endgroup$ – Xi'an Feb 28 '17 at 11:12

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.