Monotone Benders Cuts

• Monotone Benders cuts are valid inequalities that never relax the master problem, ensuring a nondecreasing lower bound in each iteration.
• They are derived from both classical and logic-based Benders decompositions, using optimality and feasibility cuts that remain valid across master decisions.
• These cuts improve convergence in mixed-integer and stochastic programs by cumulatively refining the feasible region, as evidenced in scheduling and vehicle routing applications.

Monotone Benders cuts are Benders or logic-based Benders cuts whose cumulative addition never relaxes the master problem: each cut is valid for the original problem, is retained, and therefore shrinks the master feasible region while producing a nondecreasing lower bound in minimization. In classical Benders decomposition this behavior follows from successively adding valid optimality and feasibility cuts derived from LP dual solutions or dual rays; in logic-based Benders decomposition (LBBD) the same pattern arises when a proof for a fixed master decision is re-used as a globally valid lower bound or infeasibility certificate over other master decisions (Hooker, 2019). In more recent terminology, the phrase also refers to cuts built directly from monotone recourse functions, such as $$\Theta_p \ge \mathcal{Q}_p(x'_p)\,\mathbf 1[x_p\le x'_p]$$ and $\mathbf 1[x_p\le x'_p]=0$, which are valid when the second-stage value function is nonincreasing in the first-stage variables (Legault et al., 7 Aug 2025).

1. Formal basis in classical and logic-based decomposition

The canonical LBBD setting considers

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

with $x$ as master variables and $y$ as subproblem variables. Fixing $x=\bar x$ defines a subproblem $SP(\bar x)$, and LBBD assumes an inference dual whose optimal value equals the subproblem optimal value. If $v^*$ is the subproblem value and $P^*$ is a proof of $f(\bar x,y)\ge v^*$, the same proof can be re-used at a general $\mathbf 1[x_p\le x'_p]=0$0 to produce a function $\mathbf 1[x_p\le x'_p]=0$1 such that $\mathbf 1[x_p\le x'_p]=0$2 and $\mathbf 1[x_p\le x'_p]=0$3. The master therefore receives a cut

$\mathbf 1[x_p\le x'_p]=0$4

and after $\mathbf 1[x_p\le x'_p]=0$5 iterations the master problem is

$\mathbf 1[x_p\le x'_p]=0$6

Each $\mathbf 1[x_p\le x'_p]=0$7 is a lower bound on the optimal value, $\mathbf 1[x_p\le x'_p]=0$8 is nondecreasing in $\mathbf 1[x_p\le x'_p]=0$9, and if $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$0 is finite, LBBD terminates finitely (Hooker, 2019).

The same structure applies to feasibility subproblems. If $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$1 is infeasible, an inference-dual proof of infeasibility yields a feasibility cut $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$2 that is violated at $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$3 and valid for all feasible solutions of the original problem. The resulting master again shrinks monotonically, and the corresponding sequence of lower bounds is nondecreasing. This is the most direct formal sense in which monotone Benders cuts arise: cuts encode proofs that remain valid after they are added.

Classical Benders decomposition is recovered when the subproblem is an LP and the inference system is nonnegative linear combination and domination. For

$$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$4

fixing $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$5 yields the dual

$$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$6

An optimal dual solution $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$7 produces the valid optimality cut

$$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$8

while a dual ray $$\min \{\, f(x,y)\mid C(x,y),\ C'(x),\ x\in D_x,\ y\in D_y\,\},$$9 from an infeasible subproblem yields a feasibility cut

$x$0

These are globally valid inequalities added cumulatively, so the familiar monotone tightening of classical Benders is a special case of the general LBBD mechanism (Hooker, 2019).

2. Polyhedral interpretation: supporting, facet-defining, and Pareto-optimal cuts

Brandenberg and Stursberg recast Benders cut generation through the alternative polyhedron and the reverse polar of the shifted epigraph of the recourse function. For a two-block problem

$x$1

the recourse function is

$x$2

and Benders decomposition works in the $x$3-space of the epigraph $x$4. If $x$5, the infeasibility of

$x$6

is certified by a point $x$7 in the alternative polyhedron, and each such certificate induces the valid inequality

$x$8

The key structural result is that a relaxed alternative polyhedron is an extended formulation of the reverse polar of $x$9, so cut selection can be posed as optimization over the reverse polar rather than as an arbitrary objective over the alternative polyhedron (Brandenberg et al., 2019).

This viewpoint distinguishes several notions of strength. A cut is supporting if its right-hand side is the support value of $y$0 in its normal direction. A cut is facet-defining if it supports a facet of $y$1, or in subdimensional cases, a maximal supporting face. For $y$2, a cut is Pareto-optimal if no other valid cut implies at least as large a bound on $y$3 for all $y$4 and a strictly larger bound for some $y$5. Brandenberg and Stursberg show that when the cut-generating objective is restricted to the structured form $y$6, the resulting cuts are supporting; for generic admissible objectives they are, except on a lower-dimensional set, facet-defining; and with $y$7 chosen in $y$8, the generated supporting cut is Pareto-optimal (Brandenberg et al., 2019).

In this polyhedral sense, monotonicity is not only the accumulation of valid cuts but also the accumulation of cuts that are locally undominated. Each supporting cut tightens the current outer approximation without slackening it in its own normal direction. Each facet-defining cut adds a maximal face of the target epigraph. Each Pareto-optimal cut avoids master-side dominance. This suggests a stronger interpretation of monotone Benders cuts as a sequence of nonredundant or minimally redundant improvements to the master relaxation.

3. Forms of monotonicity and convergence

The most basic monotonicity statement is

$y$9

because each master problem adds constraints and removes none. Hooker also notes that Benders cuts “partially describe the projection of the feasible set” onto the master variables, so the master progressively approximates that projection more tightly. When a newer cut dominates an older one, keeping both does not harm monotonicity, though one may become redundant. In branch-and-check, the same principle survives inside branch-and-cut: globally valid Benders cuts are injected into the MILP, the LP relaxation becomes increasingly restrictive, and lower bounds behave as in standard Benders, with additional global cuts derived from subproblems (Hooker, 2019).

A common misconception is that monotonicity requires storing every cut forever. In deterministic dynamic programming, Guigues studies Benders-type cuts for value functions $x=\bar x$0 and shows a weaker but sufficient condition: at all previously visited trial points $x=\bar x$1, the approximations satisfy

$x=\bar x$2

Level 1, Territory, and limited memory Level 1 all satisfy this property, even though cuts may be dropped from memory. The new cut is active and exact at the newest trial point,

$x=\bar x$3

and all retained cuts remain valid lower bounds to the true value function. For nonlinear problems this yields asymptotic convergence; in the linear case, under a vertex-finding algorithm such as simplex, there exists $x=\bar x$4 such that the approximations stabilize exactly and the resulting policy is optimal (Guigues, 2017).

This distinction is technically important. The monotonicity of stored cuts, the monotonicity of the induced lower envelope, and the monotonicity of the lower bound at trial or incumbent points are not identical properties. The DDP results show that structured cut deletion can preserve the monotone behavior that matters for convergence, even when the literal set of active cuts is not nested.

4. Cut families and strengthening mechanisms

Within LBBD, monotone behavior emerges through several cut families. Hooker describes dual-information cuts for LP/MIP subproblems, no-good cuts, inference-based cuts, and strengthened cuts that identify smaller subsets of master assignments responsible for infeasibility or high cost. In the scheduling example with assignment variables $x=\bar x$5, facility makespans $x=\bar x$6, and cumulative scheduling subproblems, one strengthened no-good cut is

$x=\bar x$7

which enforces $x=\bar x$8 whenever all jobs in $x=\bar x$9 remain assigned to facility $SP(\bar x)$0. More analytic cuts exploit time windows and energy arguments, for example

$SP(\bar x)$1

and, when release times rather than deadlines matter,

$SP(\bar x)$2

Hooker also emphasizes subproblem relaxations added directly to the master, such as energy inequalities

$SP(\bar x)$3

makespan relaxations $SP(\bar x)$4, and lower-bounding families for tardiness. These are all monotone in the sense that each new inequality only shrinks the feasible assignment space or raises lower bounds (Hooker, 2019).

A later and explicit use of the term appears in integrated bus fleet electrification planning. There the period-$SP(\bar x)$5 operational value function $SP(\bar x)$6 is nonincreasing in the strategic investment vector $SP(\bar x)$7 because $SP(\bar x)$8 and $SP(\bar x)$9 are nonnegative, $v^*$0 appears only on the right-hand side of operational constraints, and increasing $v^*$1 componentwise only expands the operational feasible region. This yields the logic-based cut

$v^*$2

hence the monotone optimality cut

$v^*$3

and, for infeasible subproblems, the monotone feasibility cut

$v^*$4

The same paper derives strengthened monotone feasibility cuts on subsets or aggregates of variables, including a route-level fleet feasibility test and an aggregated depot capacity test. These dominate the generic lower-orthant cut in the sense summarized there as

$v^*$5

and are encoded sparsely through shared indicator variables keyed by thresholds rather than by cut instances (Legault et al., 7 Aug 2025).

These examples clarify that monotone Benders cuts need not be linear. In LBBD and in monotone recourse settings they may be logical, combinatorial, or indicator-based. What is monotone is the validity and cumulative tightening, not the algebraic form.

5. Stronger monotone schemes: corner and disjunctive Benders cuts

Recent work extends the notion of monotone strengthening beyond classical dual cuts. In corner Benders’ cuts, a basis of the higher-dimensional polyhedron $v^*$6 defines a conic relaxation, or corner, $v^*$7 with $v^*$8. The associated relaxed value function $v^*$9 has epigraph

$P^*$0

and reverse-polar separation yields facet-defining inequalities for $P^*$1. Because $P^*$2 implies $P^*$3, any inequality valid for $P^*$4 is also valid for $P^*$5. For a fixed corner, the algorithm adds facet cuts of $P^*$6 and never removes them, so the outer approximation of the value function refines monotonically. Proposition 4.3 further shows that if $P^*$7 is chosen as an optimal corner for the Lagrangian direction, then optimizing over $P^*$8 recovers the Dantzig–Wolfe bound (Ota et al., 26 Sep 2025).

Disjunctive Benders Decomposition strengthens monotone behavior in a different direction: instead of separating only the epigraph of the continuous Benders reformulation, it generates inequalities valid for the convex hull of a split disjunction applied to that reformulation. For a split $P^*$9, the method considers

$f(\bar x,y)\ge v^*$0

and generates the deepest valid inequality for $f(\bar x,y)\ge v^*$1 by solving a normalized cut-generating LP. The dual problem is a projection problem onto the disjunctive hull, so the resulting cut is a supporting hyperplane through the projection point. For mixed-binary models, using the coordinate splits $f(\bar x,y)\ge v^*$2 yields lift-and-project cuts; the paper extends a posteriori strengthening and lifting procedures to the Benders setting and proves that, with these splits, Algorithm 2 converges in finitely many iterations to an optimal solution of the Benders reformulation, eliminating the need to solve the master as a mixed-integer program (Fang et al., 4 Jun 2025).

These two developments suggest complementary generalizations of monotone Benders cuts. Corner cuts exploit basis geometry to generate multiple facet-defining cuts from a single corner. Disjunctive cuts exploit split disjunctions to approximate the integer hull rather than only the continuous epigraph. In both cases, the master is tightened by a nested sequence of valid inequalities with nondecreasing lower bounds.

6. Applications, computational role, and limitations

In practical large-scale models, monotone cuts are rarely the sole mechanism. The bus fleet electrification study formulates a two-stage integer program with integer subproblems and combines preprocessing, partial decomposition, classical LP-based cuts, and monotone Benders cuts derived from integer subproblems and monotonicity. The authors report speedups of up to three orders of magnitude and state that monotone cuts alone are inefficient: the final method generates LP-based cuts in every iteration and invokes monotone cuts only when classical cuts fail to cut off the current master solution. They also report a case in which variants that generated two strengthened route-level feasibility cuts solved an instance to optimality, whereas the version without monotone-cut improvements terminated with five weaker generic infeasibility cuts (Legault et al., 7 Aug 2025).

In value-function approximation for the vehicle routing problem with stochastic demands, corner Benders’ cuts improve the root relaxation and reduce tree size relative to standard Benders and a single Lagrangian cut. On 720 benchmark instances, the reported numbers of solved instances are 600 for the baseline method, 327 for standard Benders with Fischetti normalization, 595 for the single Lagrangian cut, and 636 for corner Benders’ cuts. The paper also reports that corner explores far fewer nodes than lagrange and parada, that about 16% of small-$f(\bar x,y)\ge v^*$3 instances and 14% of large-$f(\bar x,y)\ge v^*$4 instances are solved by corner at the root node, and