Logic-Based Benders Decomposition

• Logic-Based Benders Decomposition is a generalization of classical Benders decomposition that integrates combinatorial, constraint-oriented subproblems for enhanced problem solving.
• It employs logic-based inference to generate exclusion and optimality cuts, facilitating rapid convergence even in NP-hard, nonconvex scenarios.
• LBBD strategically combines heterogeneous solvers like MIP, CP, and simulation, proving effective in applications such as scheduling, robust optimization, and hierarchical planning.

Logic-Based Benders Decomposition (LBBD) is a generalization of classical Benders decomposition that enables superior integration of combinatorial, constraint-oriented, and non-convex subproblem structures within the decomposition paradigm. In LBBD, the master problem and subproblems can be solved using heterogeneous optimization technologies (e.g., MIP, CP, MILP, specialized algorithmics), with logic-based inference used for generating valid cuts. The approach extends the reach of decomposition beyond linear programming duality, allowing exact algorithms that leverage problem structure and combinatorial inference, particularly for large-scale, stochastic, and highly structured problems in operations research, scheduling, robust and stochastic optimization, and combinatorial design.

1. Fundamental Principles and Decomposition Structure

LBBD decomposes an optimization problem of the form

$$\min_{x,y}\;\{\,f(x,y)\mid C(x,y),\;C'(x),\;x\in D_x,\;y\in D_y\}$$

by partitioning decision variables into "master variables" $x$ and "subproblem variables" $y$. The master problem determines $x$, subject to master-side constraints, and maintains relaxation variables (such as an explicit bound $z$). For each fixed assignment $x^*$, the subproblem is

$\min_{y}\;\{\,f(x^*,y)\mid C(x^*,y),\;y\in D_y\}$

which may be an arbitrary optimization or feasibility problem—not restricted to LPs or MILPs as in classical Benders.

A key principle is that the subproblem’s solution or infeasibility, together with its proof structure or combinatorial properties, is exploited to generate a “Benders cut” or inference–based relaxation. This “logic-based” cut constrains the master problem and ensures noncycling, eventually enforcing optimality or infeasibility of the overall problem (Hooker, 2019).

LBBD's generality is characterized by:

• The possibility of nonconvex, combinatorial, NP-hard subproblems;
• Cuts derived from combinatorial proofs or inference steps, not LP duality;
• Applicability to domains where natural variable partition permits decoupling (e.g., assignment + scheduling, two-stage stochastic programs, robust combinatorial optimization, hierarchical packing).

2. Construction and Types of Logic-Based Benders Cuts

While the classical Benders cut is obtained via LP dual multipliers, LBBD relies on combinatorial or logical inference extracted from subproblem solutions. Two essential categories of cuts in LBBD are:

• Nogood (no-good) or Exclusion Cuts: Encodes that the current master assignment $x^*$ leads to infeasibility in the subproblem. A minimal infeasible subsystem (IIS) or concrete “nogood” pattern is extracted and used to cut off that assignment via constraints such as

$\sum_{j \in J} x_j \leq |J| - 1$

if the combination $x_j=1$ for all $x$0 was witnessed infeasible in the subproblem (Juvigny et al., 10 Jan 2026, Cappanera et al., 2023). In scheduling, these correspond to specific assignments to machines or slots that break resource constraints or sequencing feasibility.

• Optimality Cuts via Analytical or Inference-Based Arguments: When the subproblem yields a solution with objective value $x$1, logic-based analysis deduces sufficient conditions for achieving $x$2 (e.g., certain assignments/patterns) and produces correspondingly valid lower bounds for all $x$3. A general template is

$x$4

so that if all crucial decisions in $x$5 are maintained, the master variable $x$6 (relaxation of total cost) cannot be less than $x$7.

Cut construction in LBBD is inherently problem-specific. Examples:

• Path-cover and area-bound cuts: For network interdiction or wildfire spread, minimal paths or cutsets are weaponized as cut-supports to force resource placement—i.e., prevention over a minimal vulnerability (Harris et al., 2022).
• Analytical cuts in scheduling: Exploit structural theorems (e.g., additive bounds on makespan changes upon job removal) to propagate lower bounds on performance metrics across the solution space (Elci et al., 2020, Hooker, 2019).
• Feasibility cuts in Answer Set Programming (ASP): No-good sets are codified as ASP integrity constraints to prevent recurrences of infeasible packet–day combinations in health-care scheduling (Cappanera et al., 2023).

The inference dual in LBBD is the process—often not formalized as a mathematical program—of deducing such cuts from the subproblem’s solution structure or logical proof tree, rather than from LP dual multipliers.

3. Algorithmic Variants and Implementation

The basic LBBD algorithm alternates between:

1. Master problem solution: Solve the relaxed master (typically a MILP or other discrete optimization), subject to previously generated logic-based cuts.
2. Subproblem solution: For selected candidate(s) $x$8, solve the subproblem(s) (possibly via CP, problem-specific OR algorithms, or simulation). If feasible, extract structural information to produce strong optimality cuts; if infeasible, generate exclusion/nogood cuts.
3. Cut generation and addition: Formulate the cuts derived from (2) and add to the master.
4. Termination: Iterate until convergence (lower and upper bounds meet), or solution/improvement criteria are satisfied.

Key algorithmic variants include:

• Standard (outer approximation) LBBD: Master and subproblems are solved sequentially in a loop (Hooker, 2019).
• Branch-and-Check (a.k.a. cut-embedding): Subproblems are solved within the branch-and-bound of the master; checks and cuts are generated on-the-fly at integer nodes, reducing re-optimization overhead and enabling lazy constraint management (Elci et al., 2020, Avgerinos et al., 2023).
• Double-oracle LBBD for robust/adaptive problems: For min–max–min combinatorial robust optimization (e.g., $x$9-adaptability), a double-oracle algorithm iterates between scenario and policy oracles as subproblem solvers (Ghahtarani et al., 2022).
• Hybridization with column generation: For routing or vehicle-shift design, master columns are generated via CP or specialized heuristics; LBBD operates in tandem with multiple decomposition layers (Daryalal et al., 2021).
• Simulation-integrated LBBD: Subproblems are simulation-based (not optimization problems); logic-based cuts are derived from monotonicity properties and empirical slopes in the simulation’s response surface (Forbes et al., 2021).

Implementing LBBD effectively requires careful design of cut management (e.g., lazy callbacks, cut pool truncation), warm-starts, and master–subproblem interfacing, often tailored to commercial solvers such as Gurobi, CPLEX, IBM CP Optimizer, or ASP engines such as clingo (Harris et al., 2022, Cappanera et al., 2023).

4. Computational Performance and Scalability Characteristics

LBBD achieves substantial computational efficiencies and scalability gains in several application domains. Notable findings across studies include:

• Substantial cut economy: In large-scale wildfire resource allocation, instances with up to 900 nodes required fewer than 50 logic-based cuts for convergence, yielding order-of-magnitude speedups compared to monolithic big-$y$0 models (Harris et al., 2022).
• Superior scalability for stochastic/robust programs: In two-stage and adaptive problems with combinatorial recourse—e.g., scheduling under uncertainty, robust shortest paths, and $y$1-adaptability—LBBD plus specialized subproblem solvers enable solution of instances in seconds to minutes where extensive form MIPs fail or leave optimality gaps >10% (Elci et al., 2020, Ghahtarani et al., 2022, Guo et al., 2019).
• Strong mastery in structural exploitation: Recursive and hierarchical packing problems, stochastic resource scheduling, and intermodal logistics with penalty objectives all benefit from LBBD’s ability to exploit topology and combinatorial subproblem decomposability (Grus et al., 23 Dec 2025, Avgerinos et al., 2022).
• Efficient convergence for interval min-max regret: For 0–1 problems with interval costs, LBBD precisely captures robustness via cut-based lower bounds, with formal finite convergence even when subproblems are NP-hard (Assunção et al., 2020).

In stochastic and simulation-embedded domains, LBBD’s precise cut-generation using empirical slope information from simulation achieves certified convergence with few MIP solves, dramatically outperforming classical, duality-based decomposition (Forbes et al., 2021).

A synthesized table highlighting LBBD’s comparative performance in recent application studies:

Domain	Monolithic Solve	LBBD Solve Time / Cuts	Notes
Wildfire suppression	Time out on 400+	< 200–600 s / <50 cuts	All solutions provably optimal
Bus fleet electrification	Few solved optimal	10³× speedup	≤1% gaps in heuristic, city-wide cases
Hierarchical packing	Fail except small	Solves deep hierarchies	Outperforms Bottom-Up by 0.2–2.5% in HP
K-adaptability	MILP/SG/BB stall	100% solved up to K=6	Only LBBD+DO scales with K, n
Min-max regret	Classical Benders	Finite convergence	NP-hard separation, still finite

5. Application Domains and Specialized Methodological Adaptations

LBBD’s versatility is reflected in its wide applicability and adaptation to diverse combinatorial and hybrid optimization problems:

• Assignment, scheduling, and cumulative resource planning: Logic-based cuts exploit job assignment topologies for parallel machines, flowshops, and complex cumulative constraints (Hooker, 2019, Elci et al., 2020).
• Intermodal/logistics & network design: Intermodal operations, including first-mile/last-mile logistics with delivery penalties, leverage per-stage decomposition with logic-based optimality cuts, and can be extended to hybrid VRPTW-TSP interlinkages (Avgerinos et al., 2022).
• Stochastic/robust optimization: LBBD is applied to SAA-based two-stage problems with recourse via CP or custom algorithms, K-adaptability under scenario assignment, and regret minimization under interval costs (Elci et al., 2020, Ghahtarani et al., 2022, Assunção et al., 2020).
• Resource–state hybrid scheduling: Project scheduling with time-of-use tariffs, block transitions, and nonlinear energy costs uses ILP master selection of cost-optimal states, with CP subproblem for non-energy-intensive task feasibility/optimization (Juvigny et al., 10 Jan 2026).
• Simulation-embedded inference: Integrates simulation subproblems where only black-box, monotonic performance functions are available, with slope-based logic cuts in the master for rapid convergence (Forbes et al., 2021).
• Answer Set Programming and logic solvers: LBBD enables scalable decomposition in ASP by embedding no-good cut patterns as integrity constraints, dramatically mitigating grounding and solve complexity for real-world scheduling (Cappanera et al., 2023).
• Hierarchical, multi-level recursive structure: Recursive planning (e.g., packing or design) is efficiently handled by recursive master–subproblem stacks, with cut propagation up the hierarchy (Grus et al., 23 Dec 2025).

The approach is not restricted to cost-minimization—it readily supports lexicographic, risk-averse, and robust criteria by propagating cut-based bounds or post-optimality conditions.

6. Convergence, Guarantees, and Theoretical Underpinnings

LBBD algorithms inherit finite convergence guarantees under mild conditions: finite master variable domains and finite valid cuts. Proofs do not require LP duality; instead, monotonicity and cut non-domination ensure progress at each iteration. In cases where subproblems are NP-hard, the validity and completeness of cuts guarantee exactness, given that each new infeasible or suboptimal pattern not previously excluded is covered by a new cut (Assunção et al., 2020, Hooker, 2019).

For robust optimization and $y$2-adaptability, the double-oracle approach embedded in LBBD has finite convergence, as both scenario and recourse spaces are finite (Ghahtarani et al., 2022). In simulation-based applications, monotonicity of the simulated performance measure ensures that slope-based cuts exclude dominated regions in each iteration, converging to global optima (Forbes et al., 2021).

Branch-and-Check variants further guarantee correctness by enforcing cuts at integer feasible nodes, integrating constraint propagation with conventional branch-and-bound searches (Avgerinos et al., 2023, Elci et al., 2020).

7. Comparative Analysis and Methodological Significance

LBBD generalizes and strictly subsumes classical Benders decomposition; the latter emerges as a special case with LP/MIP subproblems and dual-based cuts. LBBD’s main algorithmic and conceptual contributions include:

• Decoupling variable domain and technology: enabling CP, MILP, custom algorithms, or simulation in subproblems, as appropriate to subdomain structure;
• Logical inference as a source of dual bounds/cuts: leveraging nonconvex combinatorial structures (e.g., paths, cutsets, critical assignments);
• Integration and propagation of strong problem-specific cuts into the master, leading to rapid convergence, smaller cut sets, and avoidance of weak LP relaxations.

The literature establishes LBBD as the dominant paradigm for structured, large-scale, or inherently hybrid combinatorial problems, enabled by its capacity to balance solver technology, cut strength, and decomposition granularity. Multiple high-impact real-world case studies—from spectrum repacking to electric fleet investment to healthcare scheduling—demonstrate the framework’s practical tractability and rigorous optimality guarantees (Hooker, 2019, Legault et al., 7 Aug 2025, Cappanera et al., 2023).

LBBD continues to see methodological extension in stochastic, robust, hierarchical, and hybrid-logic optimization, in line with emerging large-scale and AI-integrated OR challenges.