Regularized-Likelihood Deconvolution with Posterior-Density Reconstruction
Density deconvolution is an ill-posed inverse problem, as recovering the latent distribution amplifies high-frequency noise in the observed data. We propose a two-stage likelihood-based procedure. The first estimator maximizes the convolution likelihood over a regularized Gaussian-mixture sieve, while the second reuses the fitted density as an empirical prior and averages the resulting conditional latent densities over the observations. The two estimators have different population targets under misspecification: posterior reconstruction locally contracts misspecification bias under identifiable convolution, but introduces an additional first-order variance component under correct finite-dimensional specification. We establish observed- and latent-domain convergence rates under Gaussian error. Under ordinary-smooth error, the direct estimator in $L^2$ and the posterior reconstruction in $L^1$ attain the classical deconvolution exponent, up to logarithmic factors. Numerical experiments and an application to Framingham blood-pressure data illustrate their complementary finite-sample behavior.