Spike and slab empirical Bayes sparse credible sets (Q2278657)

Consider the sparse normal means model with observed data \(X=(X_1,\ldots,X_n)\) given by \(X_i=\theta_i+\varepsilon_i\), where \(\theta=(\theta_1,\ldots,\theta_n)\in\mathbb{R}^n\) is an unknown parameter vector and \(\varepsilon_1,\ldots,\varepsilon_n\) are IID Gaussian random variables. The true underlying parameter vector is assumed to have at most \(s\) non-zero coordinates, where \(s=s(n)\) is such that \(s/n=o(1)\) and \(s\rightarrow\infty\) as \(n\rightarrow\infty\). The authors consider a Bayesian setting with spike and slab prior distribution \(\Pi\) given by \[ \theta\sim\bigotimes_{i=1}^n[(1-\alpha)\delta_0 + \alpha G(\cdot)]\,, \] where \(\delta_0\) is a Dirac mass at 0, \(G\) is a probability measure with corresponding heavy-tailed density proportional to \((1+|x|)^{-1-\delta}\) for \(x\in\mathbb{R}\) and fixed \(\delta\in(0,2)\), and \(\alpha\) is chosen by maximizing the marginal likelihood. In this setting, adaptive convergence rates are established for distances of the form \(d_q(\theta,\theta^\prime)=\sum_{i=1}^n|\theta_i-\theta^\prime_i|^q\) for \(q\leq2\). Credible sets are constructed as balls centred at the posterior median estimator \(\hat{\theta}\) with radius proportional to \(\int d_q(\theta,\hat{\theta})d\Pi(\theta|X)\). Frequentist coverage results for these credible sets are investigated, which are shown in some cases, and under an excessive-bias condition, to have optimal diameter and coverage properties. Focusing particularly on the case \(q=2\), the authors show that this excessive-bias condition cannot be significantly weakened.