It's splitting the probability across all x_i equally, even when they're all massively negatively weighted.

This change would have all the probability going into the null option, in that case, basically.