Aug 2 – 6, 2021
Europe/Brussels timezone

Deep Reinforcement Learning based numerical analytic continuation of one-loop diagrams

Aug 5, 2021, 9:40 AM


Parallel contribution H. Statistical Methods for Physics Analysis in the XXI Century Parallels Track H


Andreas Windisch (Department of Physics, Washington University in St. Louis)


Calculating analytic properties of Euclidean propagators is a demanding task, in particular if one considers non-perturbative approaches, such as Dyson-Schwinger equations. At the same time, once calculated in the complex domain, these correlators provide valuable insights into various properties associated with the proagating degree of freedom, and can serve as input to bound state equations. In order to compute those amplitudes in the complex domain, the integration path of the radial component of the loop momentum in the self energy has to be deformed away from the real axis in order to avoid non-analyticities such as poles and/or cuts. While in perturbative settings such deformations can be found manually, this is not feasible when it comes to iterative approaches applied to solve self-consistent integral equations. Utilizing the powerful machinery of Deep Reinforcement Learning, we demonstrate that an autonomous agent can be trained to perform such contour deformations. With our study we provide a proof of principle that such an agent could be deployed in a non-perturbative, iterative setup to compute analytic properties through suitably and automatically deformed integral contours.

Primary author

Andreas Windisch (Department of Physics, Washington University in St. Louis)


Thomas Gallien (Silicon Austria Labs, Graz Austria) Christopher Schwarzlmueller (Silicon Austria Labs, Graz, Austria)

Presentation materials

There are no materials yet.