Pole-fitting for complex functions: enhancing standard techniques by artificial-neural-network classifiers and regressors

Motivated by a use case in theoretical hadron physics, we revisit an application of a pole-sum fit to dressing functions of a confined quark propagator. More precisely, we investigate approaches to determine the number and positions of the singularities closest to the origin for a function that is only known numerically on a specific finite grid of values on the positive real axis. For this problem, we compare the efficiency of standard techniques, like the Levenberg-Marquardt algorithm, to a pure artificial-neural-network approach as well as a combination of these two. This combination is more efficient than any of the two techniques separately. Such an approach is generalizable to similar situations, where the positions of poles of a function in a complex variable must be quickly and reliably estimated from real-axis information alone.