Upper limits of 100663294.6 or higher increase start value to a minimal value, starting with approx. 96

[2xB]

Setting an upper limit of more than 100663294.5 (tested with e.g. 100663294.6, 100663295, 1e9) causes iminuit to change small start values. Which values they are exactly changed to depends on the size of the value - 100663294.6 causes a value change to approx. 96, 1e9 causes a value change to approx. 954.

Example code, tested with iminuit==2.31.1:

from iminuit import Minuit

def indiv_lsq(bg):
    return bg**2

m = Minuit(indiv_lsq,bg=1.5)
m.errordef=1
m.print_level=1

m.limits["bg"]=(1e-15, 100663294.5)
m.hesse()
print("a", m.values["bg"]) # prints 'a 1.5000000000000007'

m.limits["bg"]=(1e-15, 100663294.6)
print("b", m.values["bg"]) # prints 'b 1.5000000000000007'
m.hesse()
print("c", m.values["bg"]) # prints 'c 95.99996814728826'

This is definitely not an expected behavior. If this is however intended, Minuit should at least warn or error in this case.

[HDembinski]

Good point. We can raise a warning in that case. This is a side-effect of how Minuit treats bounded parameters. Note that if you have a parameter that can span several orders of magnitude, it is usually better to fit the log of that parameter.

[2xB]

That sounds reasonable! Should I open a bug report in the root project for this, since this should then probably be a message by Minuit2 in general?

[HDembinski]

No, let me investigate first, whether it's something we do or whether it's happening in the C++ library.