Universal inadmissibility of least squares estimator (Q1969722)

Let \(X\) be distributed according to a \(p\)-variate normal distribution with mean vector \(\theta\) and covariance matrix \(I\). To estimate \(\theta= (\theta_1,\dots, \theta_p)'\) the losses are taken to be \[ L([\sum_{i=1}^p q_i (\theta_i\cdot d_i)^2 ]^{1/2}),\quad \text{where} q_1\geq q_2\geq\dots\geq q_p> 0 \] are known and \(L(y)\) is a nondecreasing function of \(y\). For the case that \(q_1> q_2\) and at least there are two equal elements among \(q_1,\dots, q_p\), this paper establishes the existence of an estimator universally dominating \(X\) if and only if \(p\geq 4\). For general \(q_i\), this paper gives a lower bound on \(p\) that implies \(X\) is universally inadmissible.