One can devise non-linear scaling functions of any kind, up to step functions with zero penalty below a certain attribute value and a full value above that. But in common-sense scaling means a simple factor, which is 1/2 in our case.
I firmly believe that these points form a useless discussion, with much ado about (almost) nothing. If really needed one could implement a logistic function with (Bonifarzia's notation) attribute a, penalty p(a) with
p(a) = M+A/(1+exp((a-s)/w)), where A+M is the maximum penalty for high values of a, M the minimum penalty for zero attribute, w sets the width of the increase and s the the midpoint of the change. For small w you have a step at s, for large w you have a linear small slope, and for intermediate values there is an s-shaped increase of the penalty.
Large w and negative M can simulate a linear increase or a proportional one like the now used factor of 0.5.
More complicated models can be found
