I would like to ask if it is completely impossible to fix the step size in the minimization of some variables among the variables to be included in the fit using MINUIT (MIGRAD). Sorry I saw a couple of posts about the same issue but didn’t find a solution to work out my problem. I am working with some pre calculated values of two of the variables to fit for and need to avoid the calculation to get more accuracy than the one use for the pre calculation, one reason is to make the fit faster, not having many calls in between two values used in the pre calculation, which lead to the same value of chisq (between 0.9 and 1.0, the fit goes to 0.99, 0.999, 0.99999… which are asigned identical value, the one pre calculated for 1.0). If it is not possible to fix the step size, is there any other way to solve this problem? I thought about changing the tolerance, but not sure that it is what I want.

HI,
You can fix the variables.
The step size you give as input is required to give an order of scale of the variable, which is needed for computing the step sizes of the derivatives.
But, if you know a variable before, the way to to is to fix during the minimisation. If you know it with a very small uncertainty you can try to add a quadratic penalty term to your minimization function

Thanks for your answer. Sorry if I was not clear, but that is not what I want, I don’t want those variables to be fixed, I want them to be marginalized along with others, just necessaryly the step size has to be fixed, so that the fitter tries for the same values used for the pre calculation (I don’t know those variables, but a priori I used certain values of them 0.8,9.9,1.0 to calculate a probability which enters in the calculation of my chisq function). So the fitter has to try 0.8, 0.9, 1.0 … according to the values used in the pre calculation. And this is done along with other variables which are treated as usual, their values can vary freely.

Hi,
What you need to do is basically repeating many minimization for some fixed values of the variables. This is called profiling and it is done for example in Minos for computing the errors on the variables.