Consider the variant of the Subset Sum problem, where the input is a list of $2 m + 1$ positive integers of sum $2 S$, and the goal is to find a subset with the largest sum that is at most $S$. The problem is NP-hard in general, but it can be solved in polynomial time if the all inputs are 'big', that is, larger than $B\cdot (1-1/3m)$, where $B$ is the largest input and $m$ is the number of inputs.
Proof: under the above assumption, every set of $m+1$ inputs has sum larger than $(m+1)\cdot B\cdot (1-1/3m) = (m+1-1/3-1/3m)B = (m+2/3-1/3m)B > (m+1/2)B$. Since $S \leq (m+1/2)B$, the required subset cannot contain $m+1$ inputs .
On the other hand, every set of $m$ inputs has sum at most $m B$. But $S > (m+1/2)(1-1/3m)B = (m+1/2-1/3-1/6m) = (m+1/6 - 1/6m)B > mB$. Therefore, every set of $m$ inputs is feasible. Therefore, the optimal set is simply the set of $m$ largest items, which can obviously be found in polynomial time.
Note that we can slightly relax the lower bound, from $B\cdot(1-1/3m)$ to $B\cdot (1-1/t m)$ for some $t>2$, and a similar argument works to show that the problem is polynomial.
My goal is to find the exact "transition point" between subset-sum variants that are NP-hard, and those that can be solved in polynomial time. QUESTIONS:
- What is the smallest fraction $r$ such that, when all inputs are larger than $r\cdot B$, the subset-sum problem is in P?
- How to prove that, when the lower bound is smaller than $r \cdot B$, the problem is NP-hard?
