Confidence Limits on Simulation Averages 3.
Expression for confidence limit (error bar) assumes independent samples
- successive configurations in a simulation are (usually) not independent
- block averages are independent for “sufficiently large” blocks
Often 2s is used for error bar (95% confidence interval)
- when reporting error bars it is good practice to state definition
Confidence limits quantify only statistical errors. Sometimes other sources of error are more significant
- systematic errors
poor sampling (non-ergodic)
finite-size effects
insufficient equilibration
- programming errors
- conceptual errors
- limitations of the molecular model