search

dmzuckerman / Sampling-Uncertainty

Best Practices article intended for LiveCoMS
36 stars 5 forks source link

Qualitative / semiquantitative Checks #45

Closed dwsideriusNIST closed 6 years ago

dwsideriusNIST commented 6 years ago

Per Reviewer:

"p10: I think that it is dangerous to oppose a "strategy 1" where you do N "independent" repeats and use stddev/sqrt(N) to a "strategy 2" where you do a single simulation and need to worry about correlation times. As a beginner, I would go for strategy 1 because it seems you have much fewer things to worry about. And if I am a "greedy" beginner and I can do 1000 potential evaluations, I would go for 1000x1 timestep, because then my error is reduced. The only problem is that my probabily distribution is entirely determined by the arbitrary way I select my "independent" samples. In other words: you also have very much to worry about correlation times in strategy 1 !!! And it may turn out that if you add sufficient equilibration time for all your replicas, the multi-repeat approach ends up as complicated and more expensive than the single-simulation way."

agrossfield commented 6 years ago

This is an excellent point, and one I should have addressed. I added a new paragraph discussing these issues.