Test for a specified signal when the noise covariance matrix is unknown (Q1098520)

Let X be distributed as p-variate normal with mean vector \(\mu\) and covariance matrix \(a^{-1}\Sigma\) where a is a known scalar and let S be distributed as Wishart with f degrees of freedom and covariance matrix \(\Sigma\). The problem is to test the hypothesis \(H_ 0: \mu =0\) vs \(H_ 1: \mu =\delta\) (specified) based on observed (X,S) when \(\Sigma\) is unknown and X and S are independent. It is proved that the one-sided t-test on \(f-p+1\) degrees of freedom, based on \[ t=\sqrt{a(f-p+1)}\delta 'S^{-1}X/\sqrt{(1+aX'S^{- 1}X)\delta 'S^{-1}\delta -a(\delta '\quad S^{-1}X)^ 2} \] has larger power at \(\mu =\delta\) than Hotelling's T 2 test. The null robustness of the test t for the spherical distribution \[ const.| S|^{(f-p-1)/2}| \Sigma |^{-(f+1)/2}h(tr \Sigma^{- 1}S+a(X-\mu)\quad '\Sigma^{-1}(X-\mu)) \] is shown. The results are extended to more general situations where X and S have complex p-variate normal and complex \(p\times p\)-variate Wishart distributions, and also where X and S have a spherical distribution.