As we discussed in our meeting yesterday, I had forgotten to make sure that my sample configurations were uncorrelated. I have now used the subsampleCorrelatedData function in the timeseries module in pymbar. Originally I had 1001 samples, but now I have between 600-1000 in most cases. The number of uncorrelated samples depends on the state point and the property used to determine correlation (as shown below):
The IT_rho0 state point is the lowest density on the supercritical isotherm. Interestingly, the low density isotherm state points have the fewest number of uncorrelated samples when the LJ energy is used but not when pressure is used. Other than those (IT_rho0, IT_rho1, IT_rho2) the number of uncorrelated samples is nearly identical for both LJ energy and pressure and is between 600-1000.
These results suggest that the number of effective samples is not going to change dramatically in most cases. I am somewhat dreading repeating this process for every system although it shouldn't be too difficult. Since the minimum recommended value for Neff is an important result from the manuscript, it is probably important to correct for correlated data.
Just to make sure I am doing this right, to obtain these results all I did was run pymbar.timeseries.subsampleCorrelatedData(x), where x is either the LJ energy or pressure. Is there some other function I should be using from timeseries to test for correlation?
@mrshirts
As we discussed in our meeting yesterday, I had forgotten to make sure that my sample configurations were uncorrelated. I have now used the subsampleCorrelatedData function in the timeseries module in pymbar. Originally I had 1001 samples, but now I have between 600-1000 in most cases. The number of uncorrelated samples depends on the state point and the property used to determine correlation (as shown below):
The IT_rho0 state point is the lowest density on the supercritical isotherm. Interestingly, the low density isotherm state points have the fewest number of uncorrelated samples when the LJ energy is used but not when pressure is used. Other than those (IT_rho0, IT_rho1, IT_rho2) the number of uncorrelated samples is nearly identical for both LJ energy and pressure and is between 600-1000.
These results suggest that the number of effective samples is not going to change dramatically in most cases. I am somewhat dreading repeating this process for every system although it shouldn't be too difficult. Since the minimum recommended value for Neff is an important result from the manuscript, it is probably important to correct for correlated data.
Just to make sure I am doing this right, to obtain these results all I did was run pymbar.timeseries.subsampleCorrelatedData(x), where x is either the LJ energy or pressure. Is there some other function I should be using from timeseries to test for correlation?