1: \begin{abstract}

2: Metadynamics is an established sampling method aimed at reconstructing the free-energy surface

3:  relative to a set of appropriately chosen collective variables.

4: In standard metadynamics the free-energy surface is filled by the addition of Gaussian potentials 

5: of pre-assigned and typically diagonal covariance.  Asymptotically the free-energy surface is  proportional 

6: to the bias deposited. Here we consider the possibility of using Gaussians

7: whose variance is adjusted on the fly to the local properties of the free-energy surface. 

8: We suggest two different prescriptions: one is based on the local diffusivity and the other on the local geometrical properties.

9: We further examine the problem of extracting the free-energy surface when using adaptive Gaussians. 

10: We show that the standard relation between the  bias and the free energy does not hold. 

11: In the limit of narrow Gaussians an explicit correction can be evaluated. 

12: In the general case we propose to use instead a relation between bias and free energy borrowed from umbrella sampling. 

13: This relation holds for all kinds of incrementally deposited bias.  

14:  We illustrate on the case of alanine dipeptide the advantage of using adaptive Gaussians in conjunction with the new free-energy estimator both in terms of accuracy and speed of convergence. 

15: \end{abstract}

16: