This article is concerned with the mathematical analysis of a family of adaptive importance sampling algorithms applied to diffusion processes. These methods, referred to as Adaptive Biasing Potential methods, are designed to efficiently sample the invariant distribution of the diffusion process, thanks to the approximation of the associated free energy function (relative to a reaction coordinate). The bias which is introduced in the dynamics is computed adaptively; it depends on the past of the trajectory of the process through some time-averages.We give a detailed and general construction of such methods. We prove the consistency of the approach (almost sure convergence of well-chosen weighted empirical probability distribution). We justify the efficiency thanks to several qualitative and quantitative additional arguments. To prove these results , we revisit and extend tools from stochastic approximation applied to self-interacting diffusions, in an original context.One of the main limitations of standard MCMC approaches comes from the fact that the ergodic dynamics are metastable whenever µ is multimodal. In the example above, this happens when V has several local minima. A direct simulation is not able to efficiently and accurately sample the rare transitions between the metastable states, hence the need for advanced Monte-Carlo methods.Many strategies have been proposed, analyzed and applied, to overcome this issue. The associated variance reduction approaches may be divided into two main families. On the one hand, Importance Key words and phrases. adaptive biasing, self-interacting diffusions, free energy computation.
1To simplify notation, we omit the dependence of A ⋆ with respect to the parameter β.The unknows in (2) are the stochastic processes t → X t ∈ T d , t → µ t ∈ P(T d ) (the set of Borel probability distributions on T d , endowed with the usual topology of weak convergence of probability distributions), and t → A t ∈ C ∞ (T m ) (the set of infinitely differentiable functions on T m ). Initial conditions X t=0 = x 0 and µ t=0 = µ 0 are prescribed.An important observation is that the third equation in (2) introduces a coupling between the evolutions of the diffusion process X t and of the probability distribution µ t . Thus the system defines a type of self-interacting diffusion process. However, a comparison with [4] and subsequent articles [5], [6] and [7], reveals a different form of coupling. One of the aims of this article is to study the new arguments which are required for the study of the system (2).The most important quantity in (2) is the random, time-dependent, probability distribution µ t . Observe that its construction requires two successive operations to be performed. First, a weighted occupation measure µ t = µ 0 + t 0 exp −βA r • ξ(X r ) δ Xr dr is computed. Second, this measure is 3 normalized to define a probability distribution, µ t = µt x∈T d µt(dx) . The weights exp −βA r • ξ(X r ) in the definition of µ t are chosen so as to obtain the following consistency result for the es...