Deconvolution density estimation with penalized MLE

Deconvolution is the important problem of estimating the distribution of a quantity of interest from a sample with additive measurement error. Nearly all infinite-dimensional deconvolution methods in the literature use Fourier transformations. These methods are mathematically neat, but unstable, and produce bad estimates when signal-noise ratio or sample size are low. A popular alternative is to maximize penalized likelihood for a finite-dimensional basis expansion of the unknown density. We develop a new method to optimize penalized likelihood over the infinite-dimensional space of all functions. This gives the stability of regularized likelihood methods without restricting the space of solutions. Our method compares favorably with state-of-the-art methods on simulated and real data, particularly for small sample size or low signal-noise ratio. We also provide the first results on the consistency and rate of convergence of penalized maximum likelihood estimates for density deconvolution.