2 Errors

  MINUIT gives two different errors from its fitting routines: parabolic
  errors and  MINOS  errors. For functions  which  have  no correlations
  between  the parameters or   the variations are linear, the  parabolic
  error and  the   MINOS errors   should be  the   same.  If  there  are
  correlations or the errors are  asymmetric, you should always use  the
  MINOS errors to be sure.

  Occasionally  MINUIT will claim   to have converged and  the parabolic
  errors are ridiculously small.   This often happens   if you fit  with
  functions having correlated  parameters (like  Gaussians!). The  usual
  fix to this is  to give  the command  `HESSE', which  recalculates the
  covariance  matrix, and then `MINIMIZE' again  or run  `MINOS'. A good
  check is that the parabolic error  should lie between the MINOS errors
  if the problem is linear and be smaller than the MINOS errors if there
  are correlations.

  Note that in  the CERN version of MINUIT  the parabolic error includes
  the correlations between  the parameters. I believe  that was not  the
  case for the C. Rippich version.

  If you want to have 90%  confidence level upper limits for example, it
  is possible to get  these directly by  changing the chi**2 change used
  to  calculate the  error  (command  `ERROR_DEF  1.69').  Note that the
  likelihood is  scaled by a  factor of 2,  so that 1  sigma errors also
  correspond to a likelihood change of 1.

