Logic-Based Benders Decomposition

• Logic-Based Benders Decomposition is a method that separates optimization problems into master and subproblems, using logical inference to derive cuts based on structural and combinatorial properties.
• It enables exact decomposition for integer, combinatorial, or nonlinear subproblems by replacing classical LP duality with problem-specific inference techniques.
• LBBD has broad applications in scheduling, robust optimization, and stochastic programming, offering finite convergence under valid cut conditions and finite master domains.

Searching arXiv for recent and foundational papers on Logic-Based Benders Decomposition to ground the article. Found relevant papers including the foundational survey on LBBD and multiple application papers. I’ll use these along with the provided paper data to write the encyclopedia-style article. Retrieving the specific interval min-max regret paper and a few representative application papers for precise citation support. Logic-Based Benders Decomposition (LBBD) is a decomposition paradigm for optimization problems whose variables can be partitioned into a master component and a subproblem component, but whose subproblem is not naturally handled by LP duality. In the foundational formulation, LBBD replaces the dual-multiplier machinery of classical Benders decomposition with inference over an arbitrary optimization or feasibility subproblem, so that cuts are derived from logical, combinatorial, propagation-based, or structural properties of the subproblem rather than from linear dual solutions. This permits exact decomposition when the subproblem is integer, combinatorial, nonlinear, or best solved by specialized methods such as constraint programming, dynamic programming, Answer Set Programming, simulation, or tailored graph algorithms (Hooker, 2019, Elci et al., 2020, Cappanera et al., 2023).

1. Classical Benders as a special case of a broader inference framework

The generic LBBD setting separates a problem into master variables $x\in X$ and a subproblem $S(x)$ with value function

$$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$

The overall problem is then written as

$\min_{x\in X} f(x)+Q(x),$

and the master introduces an auxiliary variable $\theta$ so as to solve

$\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$

subject to valid lower-bounding optimality cuts and feasibility cuts (Hooker, 2019).

What distinguishes LBBD from classical Benders is the source of those cuts. Classical Benders requires an LP, or more generally a convex subproblem, so that dual extreme points and rays generate cuts. LBBD instead relies on an “inference dual”: for a generic minimization subproblem

$\min\{F(y)\mid C(x,y),\ y\in D_y\},$

the corresponding inference dual seeks a proof $P$ that deduces a bound $F(y)\ge v$ from the subproblem structure. The proof may come from nonnegative linear combinations in the LP case, but it may equally come from CP propagation, global constraints, combinatorial arguments, or logical explanations. In that sense, classical Benders is a special case of LBBD, namely the case in which the inference method is LP duality itself (Hooker, 2019, Elci et al., 2020).

This reformulation is not merely terminological. It changes the admissible modeling repertoire. In the surveyed literature, the subproblem is variously a classical 0–1 optimization problem with interval costs, a CP scheduling model with cumulative constraints, a feasibility check in ASP, a simulation-based performance evaluation, a routing/scheduling subproblem with discrete penalties, a double-oracle min–max–min robust problem, a BDD-represented knapsack-like recourse problem, or a Bayesian factor-graph MAP subproblem with logical clustering constraints (Assunção et al., 2020, Elci et al., 2020, Cappanera et al., 2023, Dubey et al., 2024).

2. Formal master–subproblem architecture and cut semantics

In the standard LBBD template, after solving the master at iteration $k$ and obtaining $S(x)$0, one solves the induced subproblem $S(x)$1. If the subproblem is feasible with value $S(x)$2, one adds an optimality cut

$S(x)$3

where $S(x)$4 for all $S(x)$5, and ideally $S(x)$6. If the subproblem is infeasible, one adds a feasibility cut

$S(x)$7

that removes $S(x)$8 and possibly a larger infeasible region of the master search space (Hooker, 2019).

The logic-based character of the method lies in how $S(x)$9 or $$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$0 are derived. In ASP scheduling with feasibility-only subproblems, the subproblem checks whether a day-level assignment can be refined into within-day agendas; if infeasible, it returns a no-good excluding the offending packet-day combination. In that case the master remains a pure optimization over date assignments, and the subproblems contribute only projected feasibility information. Because the objective depends only on master decisions, once all subproblems are feasible the master optimum is globally optimal under valid no-goods and finite domains (Cappanera et al., 2023).

A more explicitly value-oriented instance appears in interval $$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$1-$$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$2 min-max regret optimization. There, for a feasible set $$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$3 and interval costs $$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$4, the robust problem is

$$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$5

For binary $$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$6, the worst-case scenario is attained at the endpoint-induced vector

$$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$7

and the regret becomes

$$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$8

Introducing $$Q(x)= \begin{cases} \text{optimal value of }S(x), & \text{if }S(x)\text{ is feasible},\ +\infty, & \text{if }S(x)\text{ is infeasible and the overall problem is minimization}. \end{cases}$$9 yields a master with cuts indexed by previously discovered classical solutions $\min_{x\in X} f(x)+Q(x),$0: $\min_{x\in X} f(x)+Q(x),$1 The subproblem is not a dual LP but an instance of the classical 0–1 problem with costs $\min_{x\in X} f(x)+Q(x),$2, and the cut is valid because it encodes a lower bound on the regret function induced by the recovered optimal $\min_{x\in X} f(x)+Q(x),$3 (Assunção et al., 2020).

The same architecture extends to stochastic programs with combinatorial recourse. In two-stage planning and scheduling, the first stage is a MILP assignment problem and the second stage is a CP scheduling problem. The master contains variables such as $\min_{x\in X} f(x)+Q(x),$4 or $\min_{x\in X} f(x)+Q(x),$5 for scenario-wise recourse value, while the subproblem solves cumulative-scheduling models parameterized by the current assignment. Valid cuts may be simple nogoods, but the stronger variants are analytical logic-based cuts derived from the structural sensitivity of makespan or tardiness when tasks migrate across facilities (Elci et al., 2020).

3. Convergence theory, finite termination, and branch-and-check

A basic sufficient condition for finite termination is that the master variable domains are finite and that all cuts are valid. Under these assumptions, the master objective values form a monotonically nondecreasing sequence of lower bounds, while the best subproblem-evaluated feasible solution yields a nonincreasing sequence of upper bounds in minimization. LBBD therefore terminates after finitely many iterations with either an optimal solution or a certificate of infeasibility (Hooker, 2019).

Several application papers make this abstract theorem explicit in domain-specific terms. In interval $\min_{x\in X} f(x)+Q(x),$6-$\min_{x\in X} f(x)+Q(x),$7 min-max regret optimization, the separation subproblem

$\min_{x\in X} f(x)+Q(x),$8

is always feasible when $\min_{x\in X} f(x)+Q(x),$9. The crucial strict-progress proposition states that if the stopping condition is not met, the new cut must be indexed by some $\theta$0, where $\theta$1 is the set of previously used indices. Since $\theta$2 is finite, each nonterminating iteration adds a genuinely new index; once $\theta$3, the master coincides with the full robust formulation. This gives finite optimal convergence, including cases where the separation subproblem is NP-hard and not LP-representable unless $\theta$4 (Assunção et al., 2020).

The same logic reappears in the ASP realization. There the master provides an upper bound because it is a relaxation of the full problem, while the best combined master–subproblem solution provides a lower bound. In the feasibility-only variant, the stopping rule simplifies: once every daily subproblem is feasible, the current master solution is globally optimal, again assuming valid cuts and finite domains (Cappanera et al., 2023).

A major algorithmic variant is branch-and-check. Rather than repeatedly re-solving the master from scratch, branch-and-check embeds LBBD inside branch-and-bound: whenever an integer-feasible master solution is found, the subproblem is solved, violated logic-based cuts are added lazily, and the search continues. The foundational survey presents this as solving the master “only once,” while the stochastic planning-and-scheduling study shows that branch-and-check with analytical cuts can outperform both standard LBBD and the integer L-shaped method by several orders of magnitude, largely because it avoids repeated classical-Benders overhead from weak LP relaxations of large scheduling subproblems (Hooker, 2019, Elci et al., 2020).

4. Cut families and strengthening mechanisms

The practical efficacy of LBBD depends heavily on cut design. The surveyed work exhibits several recurring cut families.

One family consists of no-good or feasibility cuts that project a subproblem conflict back to the master. In ASP chronic outpatients scheduling, if a day-level agenda is infeasible, the subproblem returns facts of the form unfeas_subproblem(Patient,Packet,day,gid) and the master injects an integrity constraint forbidding that exact packet-day combination in subsequent iterations. The implementation exploits clingo’s multi-shot solving so that such cuts can be grounded incrementally without re-grounding the whole program (Cappanera et al., 2023).

A second family is analytical logic-based optimality cuts. In stochastic planning and scheduling with cumulative resources, the paper develops makespan cuts of the form

$\theta$5

with additional linear constraints on $\theta$6 involving the release-time span. These cuts are derived from a structural lemma bounding how much the optimal makespan can decrease when a nonempty set of jobs is removed. A weaker but simpler version eliminates $\theta$7 at the cost of domination by the stronger cut. The same study also uses master-side relaxations such as the cumulative “energy bound”

$\theta$8

which tightens the master independently of any incumbent subproblem proof (Elci et al., 2020).

A third family exploits monotonicity of the recourse value. In integrated bus fleet electrification planning, the operational value function $\theta$9 is nonincreasing in the investment vector because the relevant matrices are nonnegative. This yields monotone optimality cuts

$\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$0

along with strengthened feasibility cuts at different aggregation levels, such as single-route fleet infeasibility, multi-route aggregate depot-capacity infeasibility, and fully generic feasibility cuts. That work combines these monotone cuts with LP-based closest/deepest cuts, disaggregation by route, sparse indicator reuse, and partial decomposition of the operational layer (Legault et al., 7 Aug 2025).

A fourth family compresses many local eliminations into a single neighborhood cut. In unrelated-parallel-machine scheduling with sequence-dependent setups and a shared renewable setup resource, one branch-and-check scheme exhaustively explores a formulation-specific $\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$1-OPT neighborhood generated by internal job swaps and starting-job shifts. After proving the neighborhood locally suboptimal, it adds the complementary local-branching inequality

$\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$2

thereby cutting off all schedules in that neighborhood at once. The corresponding bounding-function lemma establishes that the resulting “supercut” still satisfies the standard LBBD validity conditions (Avgerinos et al., 2023).

These examples underscore a central methodological point: LBBD cuts are not a single formula class. They may be conflict clauses, monotone implication cuts, structural sensitivity inequalities, neighborhood elimination cuts, or master-side relaxations. Their unifying feature is that each arises from a sound explanation of the subproblem’s value or infeasibility.

5. Problem classes and application domains

LBBD has been applied to a broad range of exact and heuristic-exact decomposition settings. In robust combinatorial optimization, it provides a formal convergence framework for interval $\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$3-$\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$4 min-max regret problems, including knapsack-like and other NP-hard classical subproblems, by separating the endpoint worst-case scenario and the classical optimization oracle (Assunção et al., 2020). In robust two-stage decision rules, it has been combined with a double-oracle subproblem for the $\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$5-adaptability problem, where the master handles first-stage integer decisions and the subproblem is a min–max–min robust combinatorial optimization model over recourse policies and scenarios; finite termination follows from the finiteness of both the master domain and the generated scenario/policy sets (Harris et al., 2022).

In stochastic programming, LBBD appears in several distinct guises. The planning-and-scheduling study couples a MILP master with CP recourse under many scenarios and directly compares LBBD to the integer L-shaped method, showing the benefit of avoiding LP-relaxation overhead for scheduling recourse (Elci et al., 2020). In stochastic distributed operating-room scheduling, SAA is paired with LBBD optimality cuts and BDD-based Benders cuts; the master assigns surgeries to hospital-day or hospital-day-OR structures, while subproblems decide acceptance and cancellation under realized durations, producing robust schedules with materially lower cancellation rates and higher utilization than deterministic planning (Guo et al., 2019). In bus fleet electrification, the strategic master chooses multi-period fleet and charger investments and the integer operational subproblems schedule hourly service and charging, yielding an accelerated LBBD with preprocessing, partial decomposition, and monotone cuts (Legault et al., 7 Aug 2025).

In scheduling and planning beyond stochastic programming, the ASP chronic-outpatient application decomposes packet-to-day assignment from daily agenda feasibility, adds no-goods through multi-shot clingo, and reports that LBBD solves 30–42% more instances to optimality than monolithic ASP under the same time budgets (Cappanera et al., 2023). Network migration in telecommunications uses an LBBD master over aggregated migration counts per window and CP subproblems for synchronized technician routing and shift construction; this is reinforced by column generation in the master and CP pricing in the subproblem, and is presented as applicable to broader routing problems with node synchronization (Daryalal et al., 2021). Intermodal logistics with delay penalties decomposes intermodal shipment planning from last-mile routing/scheduling and uses inference-based optimality cuts tied to order-to-service assignments, plus feasibility no-goods in the extended connected-hubs model (Avgerinos et al., 2022).

In hybrid optimization–simulation, one paper integrates simulation directly into the LBBD loop for stochastic resource allocation with monotonic performance measures. At each incumbent, simulation produces window-specific shortfalls or augmentation certificates, and the master receives logic-based coverage or augmentation cuts rather than merely post hoc performance estimates (Forbes et al., 2021). In probabilistic inference, MAP estimation for Bayesian factor models is cast in a generalized-Benders/LBBD style: the master carries discrete latent assignments and logical clustering constraints such as must-link, cannot-link, and minimum cluster occupancy, while subproblems optimize continuous latent parameters and generate cuts with convergence certificates via upper and lower bounds (Dubey et al., 2024).

Finally, several papers show LBBD in highly structured spatial and geometric settings. A wildfire suppression model locates suppression resources on a directed graph under minimum-travel-time fire spread and solves the resulting nonlinear integer program by LBBD, benchmarking it against a MIP and an iterated local search metaheuristic (Harris et al., 2022). A multi-level recursive LBBD for hierarchical rectangle packing uses parent-level width–height decisions and child-level packing subproblems, adding plan–act cuts such as

$\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$6

to approximate child feasibility frontiers across hierarchy levels (Grus et al., 23 Dec 2025).

6. Limitations, assumptions, and current research directions

Across the surveyed work, LBBD’s main limitation is that strong cuts are highly problem-specific. The foundational survey states this directly: deriving strong analytical cuts requires structural insight, and while finite convergence follows from valid cuts and finite master domains, practical performance depends on cut strength, cut management, and master-side relaxations (Hooker, 2019). The stochastic planning-and-scheduling study echoes this by observing that introducing extra continuous variables can strengthen cuts but may enlarge the master, and by recommending weaker variants when master size becomes the bottleneck (Elci et al., 2020).

Another recurring assumption is finiteness of the master domain. Several papers rely on it explicitly for convergence: the interval min-max regret proof uses the finiteness of $\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$7; the ASP framework assumes finite domains and valid no-goods; the $\min_{x\in X,\theta\in\mathbb{R}} f(x)+\theta$8-adaptability algorithm assumes bounded discrete recourse and finitely many relevant scenario/policy combinations; the foundational theorem states that if the master variables have finite domains, LBBD terminates finitely (Assunção et al., 2020, Cappanera et al., 2023, Harris et al., 2022, Hooker, 2019). When large continuous master domains are present, additional structure is required.

A third limitation is scalability of subproblem enumeration or cut separation. The neighborhood-supercut approach for scheduling is explicit that exhaustive neighborhood exploration is required to preserve optimality, making acceleration by domination rules or selective invocation essential (Avgerinos et al., 2023). The bus electrification model notes that scalability may be impacted by extremely many scenarios or facilities, motivating scenario aggregation, sampling, multicut strategies, and sparse indicator reuse (Legault et al., 7 Aug 2025). In ASP scheduling, subproblem grounding dominates runtime in the LBBD setting, and stronger no-goods together with parallel day-level solving are identified as natural improvements (Cappanera et al., 2023).

The recent literature also points toward increasingly hybrid decompositions. The intermodal logistics paper proposes extending logic-based cuts with predictive or learning-based mechanisms and parallel subproblem solution (Avgerinos et al., 2022). The hierarchical packing work suggests richer objectives and 3D generalizations (Grus et al., 23 Dec 2025). The optimization–simulation paper suggests a broader role for LBBD in exact sample-average optimization when simulation yields monotonicity-based certificates rather than differentiable recourse models (Forbes et al., 2021). A plausible implication is that LBBD is evolving less as a single algorithm and more as a general exact-inference interface: the master remains a global optimizer over structural decisions, while the subproblem may be CP, ASP, simulation, BDD shortest path, stochastic recourse, or recursive packing, provided that it can return sound logical explanations strong enough to reshape the master search space.