For $i=1,…,N$, we have the data $X_i \sim \mathcal{N}( \delta \mu_i, \mu_i^2 )$, where $\mu_i>0$ and they are some unknown nuisance parameters. We want to test if there is a shift of the means, i.e. $\delta = 0$ v.s. $\delta \neq 0$. Note that in this model, although data points are not iid, there is an invariance: $Pr(X_i>0)$ are same for all $i$.

We are looking for the minimax test w.r.t. all possible values of $\mu_i$'s, i.e.

\begin{align} &\sup_{\phi} \inf_{\mu_1,…,\mu_N} \mathbb{E}_1[\phi(X_1,...,X_N)]\\ &s.t. \sup_{\mu_1,…,\mu_N} \mathbb{E}_0[\phi(X_1,...,X_N)] \leq \alpha, \end{align} where $\phi$ means a test with probablity $\phi(X_1,...,X_N)$ to reject the null.

We believe the solution is the sign test. The statistic is $Y=\sum_{i=1}^N 1_{\{X_i> 0\}}$, where $1_{\{X_i> 0\}}$ is the indicator function, taking value $1$ if $X_i > 0$ and $0$ otherwise. Assume $N$ is large. The $\alpha$-level test rejects when

\begin{equation} \vert 0.5 \sqrt{N} ( Y/N - 0.5 ) \vert > C_{1-\alpha/2}, \end{equation} where $C_{1-\alpha/2}$ is the $(1-\alpha/2)$-th quantile of standard normal. If we let $\delta=\epsilon/\sqrt{N}$, for some constant $\epsilon$, the above test yields an asymptotic power $1 - \Phi( C_{1-\alpha/2} - \epsilon / \sqrt{\pi \gamma} ) + \Phi( -C_{1-\alpha/2} - \epsilon / \sqrt{\pi \gamma} )$. That is, regardless the value of $\mu_i$'s, the $\alpha$-level sign test has a same asymptotic power.

So far we are trying to prove this test is indeed minimax. However, I find the minimax hypothesis testing is a very narrow topic, and hardly did I find any useful reference. I am pretty confused if I am on the right track or if this problem has been solved before. Any suggestions or references would be helpful.


1 Answer 1


Well, we indeed proved the minimaxity and wrote a paper on it ...


  • $\begingroup$ So you answered your own question and was it publishable? $\endgroup$ Jan 25, 2018 at 2:52
  • $\begingroup$ Yes. It will be submitted shortly. $\endgroup$ Jan 25, 2018 at 6:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.