HISKP Theory Seminar

Physics-informed kernels for sign problems

by Renzo Kapust (University of Heidelberg)

→ Europe/Berlin
Description

Physics-informed kernels (PIKs) [1,2] are a novel generative method that has recently been applied to sign problems in lattice field theory. Generally, any approach to a sign problem must meet two challenges: it  has to provide a formulation in which the sign problem is absent, and  this formulation has to admit efficient sampling. PIKs address both  challenges at once. They transport a real probability distribution to the complex target distribution and flow the samples accordingly.  Crucially, this transport is weight-preserving: each sample retains its statistical weight throughout the flow, and the efficient sampling of the initial real distribution carries over to the complex target. For fermionic sign problems, where the flow may run towards the zeros of the  distribution, the weight-preserving property of the PIK is used  particularly well by considering mixtures of PIK flows.

In this talk, I give a detailed introduction to PIKs, present their extension to fermionic systems, and comment on future work.

[1] F. Ihssen, R. Kapust, J. M. Pawlowski, arXiv:2510.26678
[2] F. Ihssen, R. Kapust, J. M. Pawlowski, arXiv:2603.03159