I think OP's objection is that - even granting that a "badness value" can be assigned to headaches and that 3 people with headaches is worse than 2 - there's no clear reason to suppose that 3 is exactly half again as bad as 2. It may be that the function mapping headaches to badness is logarithmic, or even that it asymptotes towards some limit. In mathematical terms it can be both monotonic and bounded.

Thus, when comparing headaches to a man being tortured, there's no clear reason to suppose that there is a number of headaches that is worse than the torture.