Hi Victor,
I am confused.
I think Bruce had the same idea as you: Because each restraint is appended to the vector to be minimized in the least square sense, the important quantity is Sum_restraints ( restraint_i)^2 and so you could just do that ahead of time, and then say that sqrt(Sum_restraints[restraint_i^2]) is a good substitute for the set of restraints. It does have the appealing feature that it separates there interdependence compared to a simple addition (where you could easily get competition between restraints). I think adding in quadrature is not quite right, as not only is the minimum of chi-square important, but also the ability to find the minimum and explore the parameter space. By adding in quadrature, you prevent the restraints from being negative. Well, the sign *is* arbitrary (is it data-model or model-data??), but the ability to switch signs is important. I am not sure whether adding in quadrature is worse than a simple addition, but neither is ideal. For sure, having enough individual restraints is the best approach. --Matt