Asymptotic normality of posterior distributions in high-dimensional linear models (Q1290378)

Consider a triangular array setup of the linear regression model \[ y_i= {\mathbf x}_i^T \beta+ \varepsilon_i, \quad i=1,\dots, n, \] where \({\mathbf x}_1,\dots, {\mathbf x}_n\) are \(p\)-dimensional nonstochastic regressors, \(\beta\) is an unknown \(p\)-vector of the regression coefficients, \(y_1,\dots, y_n\) are observations of a real-valued dependent variable, \(\varepsilon_1,\dots, \varepsilon_n\) are independent and identically distributed random errors having a density \(f(\cdot)\), and the \(y_i\), \({\mathbf x}_i\), \(\varepsilon_i\), \(p\) and \(f\) all can depend on \(n\). Let \({\mathbf X}_n^T= ({\mathbf x}_1,\dots, {\mathbf x}_n)\) and put \(A_n= {\mathbf X}_n^T {\mathbf X}_n\) which is assumed to be non-singular. Let \(\Theta\) be the parameter space and \(\beta_0\) be a true (sequence of) parameter point(s). Further, let \(\pi(\cdot)\) be a specified prior. Normalizing \(\beta\) as \({\mathbf u}= A_n^{1/2} (\beta- \beta_0)\), put \[ Z_n({\mathbf u})= \prod_{i=1}^nf(y_i- {\mathbf x}_i^T \beta_0- {\mathbf x}_i^T A^{-1/2}({\mathbf u}))f^{-1}(y_i- {\mathbf x}_i\beta_0), \quad{\mathbf u}\in A^{1/2} (\Theta- \beta_0), \] and set \(Z_n({\mathbf u})=0\) otherwise and define the posterior distribution of \({\mathbf u}\) \[ \pi_n^*({\mathbf u})=\pi(\beta_0+ A_n^{-1/2} ({\mathbf u}) Z_n({\mathbf u}))[\int\pi (\beta_0+ A_n^{-1/2} ({\mathbf w}) Z_n({\mathbf w})) d{\mathbf w}]^{-1}. \] The authors prove that under some conditions \[ \int|\pi_n^*({\mathbf u})= \phi({\mathbf u}, \gamma^{-1} \Delta_n, \gamma^{-1} I)|d{\mathbf u}@>P>> 0, \] where \(\delta_n=- \sum h'(y_i-{\mathbf x}_i^T \beta_0) A_n^{-1/2} {\mathbf x}_i\) with \(h(z)= \log f(z)\), \(\phi(\cdot, \mu, \Sigma)\) is the density of \(N(\mu, \Sigma)\), \(I\) is the identity matrix of order \(p\) and \(\gamma\) is some positive number.