On Polytopes with Linear Rank with respect to Generalizations of the Split Closure

TL;DR

This paper investigates the rank of polytopes within the 0-1 cube relative to generalized split cuts, establishing lower bounds that reveal an exponential separation between branch-and-bound proof depth and split rank.

Contribution

It introduces a family of polytopes with high split rank relative to t-dimensional lattice cuts, demonstrating a significant gap between proof depth and split rank bounds.

Findings

Polytopes with empty integer hull can have split rank linear in dimension.

Existence of polytopes with O(n) inequalities and split rank Ω(n).

Exponential separation between branch-and-bound proof depth and split rank.

Abstract

In this paper we study the rank of polytopes contained in the 0-1 cube with respect to $t$-branch split cuts and $t$-dimensional lattice cuts for a fixed positive integer $t$. These inequalities are the same as split cuts when $t=1$ and generalize split cuts when $t > 1$. For polytopes contained in the $n$-dimensional 0-1 cube, the work of Balas implies that the split rank can be at most $n$, and this bound is tight as Cornu\'ejols and Li gave an example with split rank $n$. All known examples with high split rank -- i.e., at least $cn$ for some positive constant $c < 1$ -- are defined by exponentially many (as a function of $n$) linear inequalities. For any fixed integer $t > 0$, we give a family of polytopes contained in $[0,1]^n$ for sufficiently large $n$ such that each polytope has empty integer hull, is defined by $O(n)$ inequalities, and has rank $\Omega(n)$ with respect to $t$-dimensional lattice cuts. Therefore the split rank of these polytopes is $\Omega(n)$. It was shown earlier that there exist generalized branch-and-bound proofs, with logarithmic depth, of the nonexistence of integer points in these polytopes. Therefore, our lower bound results on split rank show an exponential separation between the depth of branch-and-bound proofs and split rank.