Convergence of metadynamics: discussion of the adiabatic hypothesis (Q2240890)

The paper under review analyzes the long term behavior of metadynamic algorithm given in [\textit{A. Laio} and \textit{M. Parrinello}, ``Escaping free-energy minima'', Proc. Natl. Acad. Sci. USA 99, No. 30, 12562--12566 (2002; \url{doi:10.1073/pnas.202427399})] via metadynamics and self interacting stochastic models for polymers. The authors first consider the long-time limit of the time average of the penalizing term as the averaging, and then they reduce the penalization strength to zero in the long-time limit as the vanishing adaption. \textit{J. F. Dama} et al. [``Well-tempered metadynamics converges asymptotically'', Phys. Rev. Lett. 112, Article ID 240602, 6 p. (2014; \url{doi:10.1103/PhysRevLett.112.240602})] and \textit{G. Fort} et al. [Stat. Comput. 27, No. 1, 147--168 (2017; Zbl 1505.62144); J. Stat. Phys. 171, No. 2, 220--268 (2018; Zbl 1392.82003)] analyzed the converging issues of well-tempered metadynamics (a type of metadynamics with vanishing penalization mechanism) using tools from stochastic approximation algorithms. The difficulty with this approach is determining at which rate the penalization strength should decrease, if it vanishes too fast, the system may remain stuck in a metastable state, if it vanishes too slowly, the large fluctuations of the penalization affect the accuracy of the estimator of the free energy discussed in [the second author et al., Free energy computations. A mathematical perspective. London: Imperial College Press (2010; Zbl 1227.82002)]. The averaging approach is biased in nature except in the so-called adiabatic case when the collective walker \((\xi (Q_s))_{s\ge 0}\) follows an autonomous Markov dynamics (The adiabatic case is equivalent to a situation in which the original dynamics has values in the range of the function \(\xi\) and the directly samples \(\exp ( - \beta F(z))dz\) so that one can simplify the identity as a collective variable). The paper under review is interested in the averaging approach and the bias which is introduced as adiabatic assumption does not apply. Section 1 begins with the consistency of metadynamics with adiabatic separation, and proves the convergence of the average in time of the penalty term to minus the free energy. The one-dimensional overdamped Langevin dynamics given by the equation \(dZ_t^{OL}= - F^{'}(Z_t^{OL}) dt + \sqrt{2 \beta^{-1}} dB_t\) has adiabatic separation between the dynamic on \(Z_t^{OL} =\xi (Q_t)\) and the closed and Markovian dynamic on \(Z_t^{OL}\). The metadynamics associated with the overdamped Langevin dynamics evolutes in the following \(\Psi_t^{'} = \frac{d\Psi_t}{dz}\), \[dZ_t = - (F^{'}(Z_t) + \Psi_t^{'}(Z_t))dt + \sqrt{2 \beta^{-1}} dB_t,\ \ \ \ \ \frac{d\Psi_t (z)}{dz}=\gamma \delta^{\varepsilon} (Z_t - z),\tag{1}\] with the deposition rate \(\gamma > 0\) and a smooth even approximation \( \delta^{\varepsilon}\) of the Dirac function. The metadynamics (1) is rewritten in terms of the Fourier series associated with \(\Phi_t = F(z) - \frac{1}{2\pi} \int_T F + \Psi_t\), and project into a finite-dimensional differential equation as Galerkin dimension reduction. Proposition 1.1 shows that the finite dimensional projection gives a positive Harris process and admits a unique invariant probability with ergodicity for all \(f\in L^1 (\mu)\) and the estimation of the measure difference between the metadynamics and the finite dimensional projection one. The model in the discrete space is very close to self-avoiding walks. Proposition 1.3 determines the probability measure for a subclass of piecewise deterministic Markov processes or hybrid processes \((X_t, I_t)\) having an invariant measure, and the link between the metadynamics (1) and the discrete model is built identifying the quantities and the consistency of invariant measures in Proposition 1.4. Theorem 1.5 shows that the discrete model for the hybrid process \((X_t, I_t)\) has exponentially ergodic properties. Theorem 1.6 shows that the ergodic limit exists for \(f\in L^1 (\mu)\), and Theorem 1.7 provides the central limit convergence in law of \(N(0, c_f)\) as \(t\to \infty\) for \(c_f = 2 \int_0^{\infty} E_{\mu}[f(X_0, I_0) f(X_t, I_t)]dt \in [0, \infty)\). The main ingredient for proving these results is the control on return times to a compact set given in Theorem 1.9. The authors also discuss the efficiency of the approach to approximate the free energy at equilibrium and in the transient regime, and the comparison to an estimator based on the nonpenalized dynamics. Section 2 shows that the average of the penalization yields a biased estimator of the free energy in the absence of adiabatic separation for both the continuous and the discrete metadynamics. The dynamics is discretized by an Euler-Maruyama scheme with time scale \(\delta_t = 10^{-4}\) over the time interval \([0, 10^4]\), and in the nonadiabatic metadynamics unfortunately does not hold ``learning the free energy'' by taking time average. The authors modify the discrete model with 4 sites and 2 bins to write an explicit expression for the invariant measure in Lemma 2.1 and Lemma 2.2 with their proofs. Section 3 is devoted to prove the main mathematical results: Theorem 1.5, 1.6 and 1.7 on the adiabatic discrete case. Applying Section 9 of \textit{A. Durmus} et al. [``Piecewise deterministic Markov processes and their invariant measures'', Ann. Inst. H. Poincaré Probab. Statist. 57, No. 3, 1442--1475 (2021; \url{doi:10.1214/20-AIHP1125})] for a large class of hybrid processes, the authors check the subclass is nice enough to ensure the invariant measure for the process. Theorem 3.3 yields a natural Lyapunov function for the process \((X_t, I_t)\) via the Ray-Knight description of the process. Exponential moments for the return time in Theorem 1.9 are proved by a simple comparison argument. The proof of Theorem 1.5 follows from deducing the existence of a Lyapunov function for the continuous time process sampled at integer times, verifying that the set \({\mathcal K}\) is a \textit{petite set} for continuous time to have a Lyapunov function for the chain sampled at integer times, which implies the total variation bound by Theorem 5.2 of \textit{D. Down} et al. [Ann. Probab. 23, No. 4, 1671--1691 (1995; Zbl 0852.60075)] for the \(W_0\)-uniformly ergodic hybrid process \((X_t, I_t)\). The proof of Theorem 1.6 follows from the convergence in total variation which means that ergodic means converge for reasonable test functions to obtain the exponential ergodicity. The proof of Theorem 1.7 is reduced to the result for the process starting with its invariant measure, Slutsky's theorem and the central limit theorem for the equilibrium measure. The key tool on the Ray-Knight representation is provided in subsection 3.5, and the long time convergence analysis on the auxiliary process with generator for common evolutions of independent hybrid processes \(\eta_k^{\pm}\) in Theorem 3.5 is given in Theorem 3.7. The paper ends with the proof of Theorem 3.3 on a natural Lyapunov function for the hybrid process \((X_t, I_t)\). The purpose of metadynamics is to converge faster to an equilibrium state than for the original unbiased dynamics and to obtain in the long time limit an estimate of the free energy as the opposite of the penalty term. The paper studies a parallel between metadynamics and self interacting discrete model for the long time limit in order to the adiabatic separation.