Generating Transition Paths by Langevin Bridges

We propose a novel stochastic method to generate paths conditioned to start in an initial state and end in a given final state during a certain time $t_{f}$. These paths are weighted with a probability given by the overdamped Langevin dynamics. We show that these paths can be exactly generated by a non-local stochastic differential equation. In the limit of short times, we show that this complicated non-solvable equation can be simplified into an approximate stochastic differential equation. For longer times, the paths generated by this approximate equation can be reweighted to generate the correct statistics. In all cases, the paths generated by this equation are statistically independent and provide a representative sample of transition paths. In case the reaction takes place in a solvent (e.g. protein folding in water), the explicit solvent can be treated. The method is illustrated on the one-dimensional quartic oscillator.

💡 Research Summary

The paper introduces a novel stochastic framework for generating transition paths that are conditioned to start at a prescribed initial state A (at time t = 0) and end at a prescribed final state B (at a chosen final time t_f). The dynamics of the underlying system are assumed to follow overdamped Langevin equations, dx/dt = −(1/γ)∂U/∂x + η(t), with Gaussian white noise η(t). The authors first rewrite the transition probability P(x_f, t_f | x_0, 0) in a quantum‑mechanical form using the imaginary‑time Schrödinger operator H, which allows them to express the conditional probability Q(x, t) = P(x_f, t_f | x, t) as a matrix element of e^{-(t_f‑t)H}.

By inserting Q into the Fokker‑Planck equation they derive a modified Langevin equation (the “Langevin bridge”)
dx/dt = −D ∂U/∂x + 2D ∂ln Q/∂x + η(t),
which guarantees that any trajectory generated by this stochastic differential equation will end exactly at (x_f, t_f). This bridge equation is exact but non‑local in time because Q requires knowledge of the full propagator e^{-(t_f‑t)H}.

To obtain a practical algorithm the authors propose a short‑time approximation based on the symmetric Trotter splitting: e^{-Ht} ≈ e^{-tV₁/2} e^{-tH₀} e^{-tV₁/2}. This yields an explicit, local stochastic differential equation (Eq. 19)
dx/dt = (x_f − x)/(t_f − t) −