← All papers
First page of Scalable Non-Parametric Bayesian Inference of Free Energy Landscapes and Emission Functions from Observations

Scalable Non-Parametric Bayesian Inference of Free Energy Landscapes and Emission Functions from Observations

Eliane Briand, Helmut Grubmüller

physics.comp-ph Nov 18, 2026 · v1 physics.bio-ph physics.data-an stat.ME
Partially formalizes the convergence proof of the parallel-in-time filtering algorithm in Lean 4/Mathlib, assuming missing pieces such as Birkhoff contraction and SDE theory as hypotheses.
Free-energy landscapes govern processes from protein conformational change to climate oscillations, but are typically observed only indirectly, through observables such as FRET intensity or optical-tweezer transmittance. Neither the landscape shape, nor the emission function mapping latent state to signal, is usually known, and distinct states may produce the same signal. We therefore infer landscape and emission function jointly and non-parametrically, using the exact non-linear Bayesian filtering likelihood, to fully use the information contained in observable trajectories that are often noisy, scarce and costly to acquire. A novel parallel-in-time algorithm makes this tractable: it is up to 27 times faster than sequential filtering on one GPU, scales near-ideally across GPUs, and, given enough resources, has runtime independent of trajectory length. We demonstrate the method on a single-protein optical tweezer recording, where distinct conformations produce the same signal, and beyond biology on the El Nino-Southern Oscillation and financial market volatility.

Free-energy landscapes are usually observed only indirectly, through an unknown and possibly non-injective emission function. Inferring both jointly from noisy, scarce trajectories with the exact non-linear filtering likelihood is computationally expensive.

The latent process is modeled as overdamped Brownian dynamics, and the landscape and emission are represented non-parametrically with wavelet Besov priors. The likelihood comes from exact Bayesian filtering via a diagonalized Fokker–Planck propagator. A parallel-in-time filtering algorithm splits trajectories into blocks that are iterated until the boundary densities converge. Its convergence proof is partially formalized in Lean 4/Mathlib, with Birkhoff contraction and SDE results entering as hypotheses.

The parallel algorithm is up to 27x faster than sequential filtering on one GPU and scales near-ideally across GPUs. The method is applied to single-protein optical tweezer data, the El Niño–Southern Oscillation, and VIX volatility.