Convergence rates for an Adaptive Biasing Potential scheme from a Wasserstein optimization perspective

Free-energy-based adaptive biasing methods, such as Metadynamics, the Adaptive Biasing Force (ABF) and their variants, are enhanced sampling algorithms widely used in molecular simulations. Although their efficiency has been empirically acknowledged for decades, providing theoretical insights via a quantitative convergence analysis is a difficult problem, in particular for the kinetic Langevin diffusion, which is non-reversible and hypocoercive. We obtain the first exponential convergence result for such a process, in an idealized setting where the dynamics can be associated with a mean-field non-linear flow on the space of probability measures. A key of the analysis is the interpretation of the (idealized) algorithm as the gradient descent of a suitable functional over the space of probability distributions.

💡 Research Summary

The paper addresses a long‑standing gap in the theoretical understanding of free‑energy based adaptive biasing methods, such as Metadynamics, Adaptive Biasing Force (ABF) and related algorithms, when they are applied to the underdamped (kinetic) Langevin dynamics. While these methods are empirically known to alleviate metastability by flattening the free‑energy landscape along a set of collective variables, rigorous convergence rates have only been established for the overdamped (reversible) case. The authors propose a novel analytical framework that treats the adaptive biasing potential (ABP) scheme as a gradient flow of a suitably defined free‑energy functional on the space of probability measures equipped with the Wasserstein‑2 metric.

The key steps are as follows. First, a collective variable ξ: Tⁿ→Tᵐ (m≪n) is introduced, and the associated potential of mean force (PMF) A(z)=−β⁻¹log∫{ξ(x)=z}e^{−βU(x)}dx is defined. In practice A is learned on‑the‑fly, yielding a time‑dependent bias Aₜ and a biased potential Vₜ(x)=U(x)−Aₜ(ξ(x)). The underdamped Langevin stochastic differential equation driven by Vₜ reads
dXₜ=Vₜ dt, dVₜ=−∇U(Xₜ)dt+∇Aₜ(ξ(Xₜ))dt−γVₜ dt+√(2γ/β) dBₜ.
The corresponding Fokker‑Planck equation for the joint density μₜ(x,v) can be written as a continuity equation with drift −∇Vₜ and a kinetic diffusion term. Crucially, the authors show that this PDE is exactly the Wasserstein‑gradient flow of the functional
F(μ)=∫ μ log(μ/ν) dx dv + β∫ μ (Vₜ−U) dx dv,
where ν_ = ρ_*⊗N(0,β⁻¹I) is the equilibrium of the original (unbiased) Langevin dynamics.

Having identified the gradient‑flow structure, the authors invoke Villani’s hypocoercivity method together with a logarithmic Sobolev inequality (LSI) for the target measure ρ_. Under the assumption that ρ_ satisfies an LSI with constant λ₀ and that ∇²U is bounded, they prove that the relative entropy decays exponentially:
H(μₜ‖ν_) ≤ e^{−λt} H(μ₀‖ν_),
with an explicit rate λ that depends linearly on λ₀, the friction coefficient γ, and the bound on ∇²U. This result improves on earlier hypocoercivity estimates for the linear kinetic Langevin equation, where the decay rate scales like √λ₀ in the small‑λ₀ regime.

The analysis is then extended to the mean‑field limit of an infinite number of interacting particles. By considering the N‑particle system, establishing a uniform LSI (independent of N), and passing to the limit N→∞, the authors obtain the same exponential convergence for the nonlinear kinetic Fokker‑Planck equation that describes the evolution of the empirical distribution of the interacting system.

Numerical experiments on low‑dimensional collective variables and on high‑dimensional multimodal potentials illustrate the theoretical findings. The simulations confirm that the adaptive bias quickly flattens the marginal distribution of ξ, leading to rapid decay of entropy and efficient sampling of the original target distribution. The paper also provides practical guidelines for choosing algorithmic parameters such as the bias strength α, the friction γ, and the learning rate, showing how they affect the convergence constant λ.

In conclusion, the work delivers the first rigorous exponential convergence result for adaptive biasing potential methods applied to the underdamped Langevin dynamics, by recasting the algorithm as a Wasserstein gradient descent of a free‑energy functional. This perspective not only bridges a theoretical gap but also offers a systematic way to design and analyze enhanced‑sampling algorithms for complex, high‑dimensional molecular systems. Future directions include extending the framework to non‑periodic domains, more general collective variables, and parallel replica implementations.