PUB - Publikationen an der Universität Bielefeld

The standard implementation of the Maximum Entropy Method (MEM) follows Bryanand deploys a Singular Value Decomposition (SVD) to limit the dimensionality ofthe underlying solution space apriori. Here we present arguments based on theshape of the SVD basis functions and numerical evidence from a mock dataanalysis, which show that the correct Bayesian solution is not in generalrecovered with this approach. As a remedy we propose to extend the search basissystematically, which will eventually recover the full solution space and thecorrect solution. In order to adequately approach problems where anexponentially damped kernel is used, we provide an open-source implementation,using the C/C++ language that utilizes high precision arithmetic adjustable atrun-time. The LBFGS algorithm is included in the code in order to attackproblems without the need to resort to a particular search space restriction.