1: \begin{abstract}

2: Ensuring a satisfactory statistical convergence of anharmonic 

3: thermodynamic properties requires sampling of many atomic 

4: configurations, however the methods to obtain those necessarily produce 

5: correlated samples, thereby reducing the effective sample size and 

6: increasing the uncertainty compared to purely random sampling. In 

7: previous works procedures have been implemented to accelerate the 

8: computations by first performing simulations using an approximate

9: Hamiltonian which is computationally more efficient than the accurate 

10: one and then using various methods to correct for the resulting error. Those rely 

11: on recalculating the accurate energies of a random subset of 

12: configurations obtained using the approximate Hamiltonian thereby 

13: maximizing the effective sample size. This procedure can be particularly 

14: suitable for calculating thermodynamic properties using 

15: density-functional theory in which case the accurate and approximate 

16: Hamiltonians may be represented by parametrically suitably converged 

17: and non-converged ones. Whereas it is qualitatively known that there 

18: needs to be a sufficient overlap between the phase spaces of the 

19: approximate and the accurate Hamiltonians, the quantitative limits of 

20: applicability and the relative efficiencies of such methods is not well 

21: known. In this paper a statistical analysis is performed first 

22: theoretically and then quantitatively by numerical analysis. The 

23: sampling distributions of different free energy estimators are obtained 

24: and the dependence of their bias and variance with respect to 

25: convergence parameters, simulation times and reference potentials is 

26: estimated.

27: \end{abstract}

28: