On Optimality of Bayesian Wavelet Estimators

From MaRDI portal





A nonparametric regression model \(y_i=f(i/n)+\varepsilon z_i\), \(i=1,\dots,n,\) is considered, where \(z_i\sim N(0,1)\) i.i.d. A Bayesian wavelet estimation technique is developed for \(f\) in the case when the prior for the wavelet coefficients \(w_{jk}\) is \(w_{jk}\sim \pi_j N(0,\tau_j^2)+(1-\pi_j)\delta(0)\), where \(\tau_j^2=C_1 2^{-\alpha j}\), \(\pi_j=\min(1, C_2 2^{-\beta j})\), and \(\alpha\) and \(\beta\) are hyperparameters. A Bayes factor threshold estimator \(\widetilde w\) and posterior median rule estimator \(\widehat w\) are considered (assymptotically equivalent). Asymptotics of \[ \sup_{w\in b_{p,q}^s(R)} E\| w-w^*\| _{l_2}^2 \] are considered for \(w^*=\widetilde w\) or \(\widehat w\), where \(b^s_{p,q}(R)\) is a Besov ball of radius \(R\). Optimal choice of \(\alpha\) and \(\beta\) is discussed. Simulation results are presented.




Cited in
(31)








This page was built for publication: On Optimality of Bayesian Wavelet Estimators

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4677090)