Speaker
Description
Efficient sampling across topological sectors is one of the central algorithmic challenges in lattice field theory. Approaches based on interpolating distributions, such as parallel tempering with defects and learned-flow methods, improve ergodicity by connecting an easily sampled reference system to the target theory, but their efficiency depends strongly on the choice of interpolation path. We study this optimization problem using two complementary approaches. We first show that Fisher information geometry identifies optimal annealing schedules as geodesics in the space of probability distributions. We then introduce a machine-learning framework that searches directly for efficient interpolation paths while allowing both bulk and defect couplings to vary freely. We present results for the CP$^{N-1}$ model, illustrating how optimized protocols improve generalized tempering strategies and provide a systematic route toward mitigating topological freezing.