Many discrete optimization problems -- including low-rank quadratic unconstrained binary optimization (QUBO), trimmed covariance, and cardinality-constrained pseudo-Boolean programs -- share a common structure: the gain of any elementary move (bit-flip or swap) can be expressed as an affine form in a low-dimensional feature map $\phi(x) \in \mathbb{R}^R$. We call this the Affine--Predicate Condition, and show that it captures all rank-structured pseudo-Boolean objectives, including higher-order forms via multilinear lifting. We prove an Affine--Predicate Meta--Theorem: under this condition, every locally optimal solution must lie in a chamber of a central hyperplane arrangement in $\mathbb{R}^{R+1}$, defined by the zero-sets of these affine gains. These chambers can be enumerated in $m^{O(R)}$ time, where $m$ is the number of elementary feasible directions. Consequently, exact optimization is in XP time $n^{O(R)}$ in the effective rank $R$, even under uniform matroid (cardinality) constraints. As a corollary, we obtain the first unified $n^{O(r)}$ exact algorithm for rank-$r$ QUBO and its cardinality-constrained extension. This establishes rank as a structural parameter for discrete optimization, analogous to treewidth in algorithmic graph theory and Courcelle-type meta-theorems.
翻译:暂无翻译