Max Gläser,Marc E. Pfetsch
Abstract:This paper investigates linear programming based branch-and-bound using general disjunctions, also known as stabbing planes, for solving integer programs. We derive the first sub-exponential lower bound (in the encoding length $L$ of the integer program) for the size of a general branch-and-bound tree for a particular class of (compact) integer programs, namely $\smash{2^{\Omega(L^{1/12 -\epsilon})}}$ for every $\epsilon >0$. This is achieved by showing that general branch-and-bound admits quasi-feasible monotone real interpolation, which allows us to utilize sub-exponential lower-bounds for monotone real circuits separating the so-called clique-coloring pair. Moreover, this also implies that refuting $\Theta(\log(n))$-CNFs requires size $2^{n^{\Omega(1)}}$ branch-and-bound trees with high probability by considering the closely related notion of infeasibility certificates introduced by Hrubes and Pudlák. One important ingredient of the proof of our interpolation result is that for every general branch-and-bound tree proving integer-freeness of a product $P\times Q$ of two polytopes $P$ and $Q$, there exists a closely related branch-and-bound tree for showing integer-freeness of $P$ or one showing integer-freeness of $Q$. Moreover, we prove that monotone real circuits can perform binary search efficiently.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is about the lower bound of the complexity of the branch - and - bound method for linear programming when using general disjunctions to solve integer linear programs (ILPs). Specifically, the authors aim to prove that there exists a class of compact integer linear programming problems for which the size of any branch - and - bound tree using general disjunctions is at least \(2^{\Omega(L^{1/12-\epsilon})}\), where \(L\) is the encoding length of the integer linear programming problem and \(\epsilon> 0\) is an arbitrarily small positive number.
### Background and Motivation
1. **Branch - and - bound method**: The branch - and - bound method is a commonly used method for solving integer linear programs. It gradually approximates the optimal solution by continuously dividing the problem space. Using general disjunctions (such as "cutting planes" or "stabbing planes") allows for more flexible branching operations, but its complexity analysis has always been a challenge.
2. **Deficiencies in existing research**: Although the branch - and - bound method is very effective in practical applications, theoretically, for the branch - and - bound method using general disjunctions, there are no super - polynomial lower - bound results yet. This leaves a gap in theoretical understanding.
3. **Open problem**: Dadush and Tiwari proposed an open problem, that is, whether there exists a class of integer linear programming problems such that the size of the branch - and - bound tree using general disjunctions must be super - polynomial. This paper conducts research precisely in response to this problem.
### Main Contributions
1. **Lower - bound result**: The authors prove that there exists a class of compact integer linear programming problems whose branch - and - bound tree size is at least \(2^{\Omega(L^{1/12-\epsilon})}\). This is the first time that a sub - exponential lower - bound result has been given for this type of problem.
2. **Interpolation technique**: In order to prove the above lower bound, the authors introduce the concept of "quasi - feasible monotone real interpolation" and use the known lower - bound results of the monotone real circuit separation problem (such as "clique - coloring pairs") to derive the lower bound of the branch - and - bound tree.
3. **Random CNF formulas**: In addition, the authors also prove that for random \(\Theta(\log(n))\)-CNF formulas, the size of the branch - and - bound tree using general disjunctions is at least \(2^{n^{\Omega(1)}}\), and this result holds with high probability.
### Technical Details
1. **Interpolation lemma**: The authors first prove a Structural Interpolation Lemma, showing that for an integer linear programming problem representing the product of two polyhedra, if there exists a branch - and - bound tree proving that there is no integer solution, then there exists a similar branch - and - bound tree proving that one or more of the polyhedra have no integer solution respectively.
2. **Quasi - Farkas certificates**: In order to handle branch - and - bound trees with Farkas certificates, the authors introduce the concept of "quasi - Farkas certificates", allowing in some cases not to verify the validity of Farkas certificates, thus ensuring that the interpolation lemma still holds in branch - and - bound trees with certificates.
3. **Monotone real circuits**: The authors use the complexity results of monotone real circuits, especially Pudlák's lower - bound results on clique - coloring pairs, to derive the lower bound of the branch - and - bound tree.
### Conclusion
This paper successfully proves the lower bound of the complexity of the branch - and - bound tree using general disjunctions for a class of integer linear programming problems by introducing new interpolation techniques and the concept of quasi - Farkas certificates, fills the theoretical gap, and provides new tools and methods for further research.