Combinatorial Benders Cuts

• Combinatorial Benders cuts are polyhedral constraints derived from combinatorial analysis that tailor Benders decomposition to discrete and mixed-integer problems.
• They exploit primal infeasibility and minimal conflicting sets to generate valid inequalities that significantly improve convergence and computational efficiency.
• Applications span robust optimization, SVM training, bin packing, network contagion, and scheduling, demonstrating dramatic speedups and tighter optimality gaps.

Combinatorial Benders cuts are polyhedral constraints derived from explicit combinatorial analysis of discrete problem structure within Benders decomposition frameworks. In contrast to classical (duality-based) Benders cuts, which leverage the dual information of continuous or LP subproblems, combinatorial Benders cuts exploit primal infeasibility, logic, or minimal combinatorial conflict to produce strong, problem-specific valid inequalities. These cuts have demonstrated significant computational impact across integer and mixed-integer programs, logic-based Benders paradigms, stochastic and robust optimization, and high-dimensional combinatorial applications.

1. Foundations and Principle Schemes

Combinatorial Benders cuts emerge from a decomposition in which the "master" subproblem operates over discrete (often binary) variables, while the "subproblem" involves logical or combinatorial feasibility checks or continuous optimization with fixed master decisions. The canonical pattern is as follows:

1. The master problem proposes a value assignment—e.g., binary label selection, item-to-bin assignments, or covering sets.
2. The subproblem evaluates feasibility (for feasibility cuts) or computes an optimal objective (for optimality cuts), given the master solution.
3. If the subproblem is infeasible or suboptimal, a combinatorial analysis (often involving minimal infeasible subsystems, shortest paths, or simulation-driven marginals) identifies a critical subset or structure. This yields a cut that excludes the current—and often many neighboring—infeasible or suboptimal master solutions.
4. The master is re-optimized with the new cut, iterating until convergence.

This approach generalizes logic-based Benders decomposition and is particularly effective when the subproblem either admits efficient (even if problem-specific) combinatorial routines or minimal infeasibility/simulation reasoning can be effectively separated.

2. Minimal Infeasible Subsets and No-Good Cuts

One core class of combinatorial Benders cuts targets infeasibility by explicitly identifying (and cutting off) minimal infeasible sets ("MIS" in (Santana et al., 2022, Côté et al., 2019)). For example, in robust training of hard-margin SVMs:

• The master problem allocates which samples to misclassify via binary variables $z_i$.
• The subproblem checks if the remaining set of data admits a feasible hyperplane separation.
• If infeasible, an MIS $S$ of indices is extracted—no feasible hyperplane separates samples $i\in S$ under margin constraints.
• The cut $\sum_{i\in S} z_i \geq 1$ then blocks all future master solutions that zero out $z_i$ for all $i\in S$.

In 2D bin packing (Côté et al., 2019), after an area-based master relaxes geometric overlap, the subproblem checks for feasible packing. An infeasible item set $S_i$ yields a cut $\sum_{j\in S_i} x_{ij} \leq |S_i|-1$, with possible strengthening by restricting to minimal $C\subseteq S_i$ (MIS) or lifted via auxiliary knapsack relaxations. These no-good cuts prune the master space efficiently and can often be generalized to other minimal conflicting structures (e.g., subtours, unsatisfiable resource configurations).

3. Combinatorial Cut Generation via Explicit Subproblem Algorithms

Certain problem classes admit ultra-efficient combinatorial cut separation strategies:

• For network contagion minimization (Tanınmış et al., 2023), each master solution defines a label-blocking (arc removal) strategy. Rather than solving LP subproblems, the set of infected nodes across all scenarios is computed by running Dijkstra's or shortest-path algorithms. Activation paths for each reachable node yield cut coefficients via path aggregation, producing cuts of the form $\theta^\omega \geq |R| - \sum_k C_k x_k$, matching dual-optimal LP-based coefficients without explicit LP solves.
• In probabilistic covering (Chen et al., 2024), separation of Benders feasibility cuts is strictly combinatorial: by sorting and set operations over covering patterns, optimal cuts are derived in $O(m\log m + \sum_k |S_k|)$, handling up to $10^6$ scenarios per instance.
• Scheduling under logic-based Benders (Avgerinos et al., 2023), local search (e.g., $k$-OPT neighborhoods) is used to exhaustively explore move neighborhoods, and if all are suboptimal, a single Hamming-ball (local branching) cut excludes all such solutions at once.

The key advance across these examples is that full LP dual computation is bypassed in favor of strongly polynomial combinatorial routines that generate integer-valued, facet-defining cuts.

4. Analytical, Simulation-Based, and Logic-Driven Cuts

Beyond infeasibility, combinatorial Benders cuts can encode quantitative optimality information—even when the subproblem cannot be dualized or lacks a tractable LP relaxation:

• In stochastic planning-and-scheduling (Elci et al., 2020), both nogood (assignment-forbidding) cuts and analytic makespan bounds are constructed via explicit reasoning about how the optimal schedule's makespan responds to subset removals, yielding inequalities that closely bound the true recourse in the master variables.
• In stochastic resource allocation with simulation (Forbes et al., 2021), scenario batching and marginal impact estimation produce composite Benders cuts: for monotonic cost functions, the cut $$\theta \geq \phi(x^k) + \sum_j \Delta_j^k (x_j - x_j^k)$$ is valid, with all coefficients $\Delta_j^k$ obtained via simulation—the cuts incorporate real marginal information and are provably strong for submodular or monotonic functions.

The logic-based paradigm (LBBD) thus leverages combinatorial proofs, simulation, or analytic arguments rather than classical convex duality to derive cuts, enabling applicability to domains unavailable to classical (L-shaped) Benders decomposition.

5. Polyhedral and Geometric Facet-Defining Cuts

Combinatorial Benders cuts are not limited to direct combinatorial conflict; they can be tied directly to polyhedral or epigraphical structure:

• In value-function approximation (Ota et al., 26 Sep 2025), corner Benders cuts use conic relaxations constructed via a subproblem basis (“corner”) and project the higher-dimensional value function into the master space. Reverse polar techniques identify a bundle of facet-defining cuts for the epigraph of the corner function, and a row-generation procedure efficiently manages the ray constraint system.
• This multi-cut approach recovers the same LP bound as classical Dantzig-Wolfe (DW) column generation through the aggregation of these facet cuts, and computational experiments on stochastic VRP show orders-of-magnitude speedups and dramatically fewer cuts required to close optimality gaps at the root node compared to both traditional Benders and Lagrangian approaches.

Such geometric combinatorial Benders cuts provide stronger and fewer constraints than classical approaches relying on a single dual-optimality cut, especially when the number of master variables is large or the value function is highly degenerate.

6. Algorithmic Integration, Buffering, and Real-Time Applications

Combinatorial Benders cuts facilitate advanced algorithmic integration strategies:

• In hybrid control policies for robotic systems (Lin, 2024), feasibility and optimality Benders cuts—derived via Farkas certificates and dual solutions from QP subproblems—are stored in online buffers (finite cut pools) and re-used across successive MPC iterations. This enables rapid warm-starting, matching or outperforming off-the-shelf MIQP solvers even under continual environmental changes.
• Efficient greedy subproblem sequencing is possible when cuts are sparse in variables (e.g., time-indexed), dramatically reducing solution times in stagewise problems.
• Selective separation, adaptive cut management, and dynamic aggregation (e.g., block-structured PSCP in (Chen et al., 2024)) further increase tractability for large problem instances.

This approach offers continual improvement and learning over time, opening the door for large-scale, real-time, or repeated-decision applications in stochastic or hybrid systems.

7. Theoretical and Computational Impact

Extensive computational evaluations demonstrate that combinatorial Benders cuts achieve superior performance on a variety of benchmarks:

• On hard-margin SVM training (Santana et al., 2022), combinatorial Benders reduced the average optimality gap from $\sim90\%$ to $\sim36\%$ on the hardest $n=500$ instances, solving 117 new datasets to full optimality.
• In network contagion minimization (Tanınmış et al., 2023), scenario-wise combinatorial cuts enabled root gaps to drop from 80–90% (LP-based BD) to 15–20%, with 5–100x speedup and full solution on previously intractable graphs.
• In large-scale probabilistic covering (Chen et al., 2024), combinatorial separation handled $10^6$-scenario instances an order of magnitude faster than state-of-the-art MIP and automatic Benders methods.
• In scheduling via local-branching cuts (Avgerinos et al., 2023), optimality gaps fell by 50–80% and solution times improved 2–5x, with dramatically fewer nodes explored.

The validity and convergence of combinatorial cuts are established either by polyhedral arguments (deepest supporting hyperplanes), combinatorial minimality, submodularity, or (for block-structured/epigraphical relaxations) by convex duality theorems. Most approaches guarantee finite convergence by the finiteness of possible discrete assignments or minimal infeasible/minimal conflict sets.

A plausible implication is that the combinatorial Benders paradigm enables mathematically strong, domain-tailored, and computationally efficient decomposition methods for integer, hybrid, simulation-based, and logic-rich optimization problems—far beyond the classical settings where Benders decomposition was originally developed.