Exploring Boltzmann Distributions via Kernel-Based Fokker–Planck Ensemble Transport
Local trajectory-based samplers can relax within a visited basin long before finite trajectories cross barriers to other Boltzmann-relevant regions. We introduce the Fokker–Planck Transformer (FPT), a training-free interacting-ensemble method that makes the instantaneous ensemble distribution a dynamical variable, shifting exploration from trajectory-wise barrier crossing toward distribution-level redistribution. FPT discretizes the current-velocity representation of overdamped Fokker–Planck evolution by combining energetic drift with a kernel estimate of the ensemble score. The continuum flow has the Boltzmann density as its stationary target and monotonically dissipates configurational free energy and relative entropy. A Gaussian row-mass selector couples increasing local occupancy to a vanishing spatial scale without a separate bandwidth limit. For its fresh explicit-Euler realization, pathwise score residuals propagate to distributional and observable errors, while a weak-flux argument gives conditional fixed-horizon convergence to the Fokker–Planck solution. For this realization, a controlled two-dimensional calculation tracks transport from a localized Gaussian toward equilibrium against a numerically propagated Fokker–Planck reference. In a three-dimensional periodic landscape, an exact Fokker–Planck reference shows that a shared-bandwidth approximation populates distant wells far ahead of the continuum solution, which accounts for its faster support discovery; the transported ensemble runs warm, and a Hastings-corrected continuation from its endpoint restores the energy distribution within two steps. Further tests in a 100-dimensional landscape, alanine decapeptide, and 190 docking systems show rapid expansion of Boltzmann-relevant support, with the clearest gains over matched local controls when barriers or configurational complexity limit global exploration. Overall, FPT provides thermodynamically grounded finite-budget support discovery and can seed target-preserving sampling.
Authors
- Yi Qin Gao (ORCID: https://orcid.org/0000-0002-4309-9376)
- Xiaoxuan Yu
- Zichen Sun
- Jun Zhang
Institutions
- Peking University (CN)
- Westlake University (CN)
- Beijing National Laboratory for Molecular Sciences (CN)
- Changping Laboratory (CN)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-19
- DOI
- https://doi.org/10.26434/chemrxiv.15012462/v1
- Primary Topic
- Markov Chains and Monte Carlo Methods
- Type
- preprint