We introduce the flow perturbation method, which incorporates optimized stochastic perturbations into the flow. By reweighting trajectories generated by the perturbed flow, our method achieves unbiased sampling of the Boltzmann distribution with orders of magnitude speedup compared to both brute force Jacobian calculations and the Hutchinson estimator.

Please find the arxiv preprint here: https://arxiv.org/abs/2407.10666

Citation update is coming up...

@article{peng2024flow,
  title={Flow Perturbation to Accelerate Unbiased Sampling of Boltzmann distribution},
  author={Peng, Xin and Gao, Ang},
  journal={arXiv preprint arXiv:2407.10666},
  year={2024}
}