Certificate-Sensitive Subset Sum: Realizing Instance Complexity

Abstract: We present, to our knowledge, the first deterministic, certificate-sensitive algorithm for a canonical NP-complete problem whose runtime provably adapts to the structure of each input. For a Subset-Sum instance $(S, t)$, let $\Sigma(S)$ denote the set of distinct subset sums and define $U = |\Sigma(S)|$. This set serves as an information-theoretically minimal witness, the instance-complexity (IC) certificate. Our solver, IC-SubsetSum, enumerates every element of $\Sigma(S)$ in deterministic time $O(U \cdot n^2)$ and space $O(U \cdot n)$. A randomized variant achieves expected runtime $O(U \cdot n)$. The algorithm's complexity is thus directly governed by the certificate size, and this structure-sensitive performance is paired with a guaranteed worst-case runtime of $O^*(2^{n/2 - \varepsilon})$ for some constant $\varepsilon > 0$, the first such result to strictly outperform classical methods on every instance. We revisit fine-grained reductions that rely on the classical $2^{n/2}$ hardness of SubsetSum and show that these arguments hold only for collision-free instances where $U$ is maximal. IC-SubsetSum reframes this barrier structurally and introduces a new paradigm for certificate-sensitive algorithms across NP-complete problems.