Statistical errors in Monte Carlo estimates of systematic errors

For estimating the effects of a number of systematic errors on a data sample, one can generate Monte Carlo (MC) runs with systematic parameters varied and examine the change in the desired observed result. Two methods are often used. In the unisim method,1 the systematic parameters are varied one at a time by one standard deviation, each parameter corresponding to a MC run. In the multisim method (see footnote 1), each MC run has all of the parameters varied; the amount of variation is chosen from the expected distribution of each systematic parameter, usually assumed to be a normal distribution. The variance of the overall systematic error determination is derived for each of the two methods and comparisons are made between them. If one focuses not on the error in the prediction of an individual systematic error, but on the overall error due to all systematic errors in the error matrix element in data bin m, the number of events needed is strongly reduced because of the averaging effect over all of the errors. For simple models presented here the multisim model was far better if the statistical error in the MC samples was larger than an individual systematic error, while for the reverse case, the unisim model was better. Exact formulas and formulas for the simple toy models are presented so that realistic calculations can be made. The calculations in the present note are valid if the errors are in a linear region. If that region extends sufficiently far, one can have the unisims or multisims correspond to k   standard deviations instead of one. This reduces the number of events required by a factor of k2k2.