“…In our simulations, some easily implementable changes to the method are included, which lead to improved accuracy and to noticeable savings in CPU time. In particular, (i) the density of states is updated only after every Monte Carlo sweep, such that we discard very correlated configurations, (ii) the microcanonical averages are accumulated beginning from the eighth WL level (f 7 ), so that we avoid considering the initial configurations that do not match with those of maximum entropy [21], (iii) a checking parameter ε is used for halting the simulation [22] (the integral of the specific heat over a range of temperatures is calculated using the current density of states during the simulations and the simulations are halted if this quantity varies less then 10 −4 during a whole WL level), and (iv) a single run is performed for all lattice sizes up to the Wang-Landau level f 6 and then, the further simulations begin from these outputs, since up to this point the current density of states is not biased yet and the results we reach are similar to those that would be obtained beginning from the first WL level f 0 [23], a measure that saves about 60% in computational time.…”