Applications of Mellin-Barnes Integrals to Deconvolution Problems

Abstract

In this thesis we study the additive model of errors in variables, which is also known as the deconvolution problem. The objective consists particularly in the reconstruction of the distribution F associated with a random variable X, which is observable only through a sample of a blurred variable Y, due to an additive random error ε with known distribution H. Our initial considerations yield an unbiased estimator for F for various discrete and some continuous distributions. A more general approach then leads us to the symmetrized model of errors in variables. It is obtained by an additional convolution of G with the conjugate error distribution of H, thereby resulting in an error distribution of symmetric type. As a consequence the characteristic function of X can be represented as the limit of a geometric series. By truncation of this series we deduce an approximation of F, which is valid for arbitrary error distributions. This approximation, termed the deconvolution function, converges to F in many cases. To determine the corresponding rates of convergence, techniques from complex calculus and particularly Mellin-Barnes integrals turn out to be appropriate. The latter describe a special class of integrals that can be evaluated by residue analysis. The results are established in a more general setting, which makes them applicable to other Laplace-type integrals. With the aid of the deconvolution function we also construct an estimator for F. The asymptotic properties of its variance, a peculiar integral of dimension two, can be specified by virtue of our findings from the concluding chapter. These results rely on iterated Mellin-Barnes integrals.