Estimate of Errors to be Used in Fits

• One should also mention that there are two versions of depending on whether error taken from ``experiment'' or ``estimated theoretically.''

• In the latter case, it will be a function of the parameters to be varied.

• For example, suppose is number of events in some counting experiment.

  The (experimental)

  but (theory) .

• Now is only Gaussian for large , but then and are essentially same.

• If , then theoretical is to be preferred!---although method itself is in difficulties as not Gaussian.

• It is technically best to go back to maximum likelihood and put in true Poisson distribution for .