Lagrangian-Based Valid Inequalities in Optimization

• Lagrangian-based valid inequalities are constraints redundant for the original model that, when dualized, strengthen relaxation bounds by leveraging problem-specific structure.
• They are applied in QCQP, CDT, and block-structured MIPs to preserve separability and reduce duality gaps, leading to improved computational performance.
• In differential geometry, these inequalities yield sharp curvature bounds by exploiting the total symmetry of the Lagrangian submanifold's cubic form.

Searching arXiv for recent and foundational papers on Lagrangian-based valid inequalities across QCQP, CDT, MIP decomposition, and related inequality frameworks. Lagrangian-based valid inequalities are constraints that are redundant for an original optimization problem but non-redundant for a chosen Lagrangian or lifted relaxation, so that their explicit enforcement or dualization strengthens the resulting bound without destroying the tractable structure that motivated the relaxation. Across continuous and mixed-integer settings, the theme recurs in several distinct forms: supporting hyperplanes added to sharpen an inexact Lagrangian/Shor bound for the Celis–Dennis–Tapia problem (Consolini et al., 2021), triangle inequalities added to McCormick–Shor relaxations of QCQPs and then embedded into an optimal convexification through dual multipliers (Lambert, 2020), monomial and vertex-based redundant constraints added to block-structured MIPs so that the decomposable Lagrangian dual has zero duality gap (Cifuentes et al., 2024), and, in a differential-geometric setting, optimal Chen-type inequalities whose sharp constants depend crucially on the Lagrangian structure of the immersion (Chen et al., 2013). These developments share a common logic: one exploits structure-specific valid inequalities to tighten a relaxation in a way that either preserves separability, improves dual bounds, or yields exactness and rigidity.

1. Conceptual scope and definitions

In optimization, the clearest operational meaning of a Lagrangian-based valid inequality is given by problems where a baseline Lagrangian relaxation is known to be tractable but weak. In the CDT problem, the relaxed ellipsoidal constraint is convex, and the paper strengthens the dual Lagrangian bound by adding one or two linear cuts obtained from supporting hyperplanes of that constraint. These cuts are redundant for the original problem but not for the Lagrangian relaxation, because they exclude Lagrangian minimizers that lie outside the ellipsoid while preserving all primal-feasible points (Consolini et al., 2021). In QCQP, triangle inequalities derived from variable bounds and McCormick underestimators are valid for the original problem, cut points feasible for McCormick alone, and can be dualized inside a partial Lagrangian dual, where their multipliers contribute to the best convex quadratic reformulation (Lambert, 2020). In decomposable MIPs, redundant monomial or vertex-indicator equalities implied by binary coupling are added so that, after dualization, the inner problem remains block-separable while the dual gap is reduced or eliminated (Cifuentes et al., 2024).

A related but structurally different use of “Lagrangian-based inequalities” appears in the geometry of Lagrangian submanifolds. There, the inequalities are not cutting planes for an optimization relaxation, but sharp curvature bounds whose validity and optimal coefficients rely on the algebraic symmetry imposed by the Lagrangian condition. The paper proves pointwise inequalities of the form

$$\delta(n_1,\dots,n_k)\le a(n,k,n_1,\dots,n_k)\|H\|^2+b(n,k,n_1,\dots,n_k)c,$$

with optimal coefficients and explicit equality conditions for Lagrangian submanifolds of complex space forms (Chen et al., 2013). This suggests a broader interpretation: the “Lagrangian-based” qualifier may refer either to inequalities used within Lagrangian duality or to inequalities whose sharpness is made possible by Lagrangian structure itself.

A recurrent distinction is between validity for the primal model and usefulness for the relaxation. A constraint may be implied by the original feasible set and hence redundant in primal space, yet materially tighten a relaxation because that relaxation omits nonlinear, integrality, or coupling structure. The supporting hyperplanes in CDT, the general triangle inequalities in QCQP, and the monomial or vertex equalities in block-structured MIPs all fit this pattern (Consolini et al., 2021, Lambert, 2020, Cifuentes et al., 2024).

2. QCQP: triangle inequalities and optimal convexification

For a general, possibly nonconvex, quadratically constrained quadratic program

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$

the classical lifted relaxation introduces variables $Y_{ij}\approx x_ix_j$ together with McCormick envelopes over the box bounds (Lambert, 2020). The triangle inequalities are obtained by expanding products of three nonnegative factors such as $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$, replacing one bilinear subproduct by a McCormick lower bound, and then linearizing the remaining quadratic terms by substituting $Y_{pq}$ for $x_px_q$. Doing this over all sign patterns and McCormick choices generates 48 candidates, of which exactly 12 per triple are non-redundant with respect to the McCormick envelope $C$ (Lambert, 2020).

These 12 inequalities are the “General Triangle inequalities,” indexed by triples $(i,j,k)$ with $i<j<k$. They reduce to Padberg’s triangle inequalities for the Boolean Quadric Polytope when $\ell_i=0$ and $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$0, so they are continuous-bound generalizations of the BQP triangle facets (Lambert, 2020). The paper shows both validity and non-redundancy: each inequality is valid because it is derived from nonnegative products and McCormick bounds, and specific points can satisfy all McCormick inequalities while violating a triangle inequality, so the latter genuinely cuts the McCormick hull (Lambert, 2020).

The same work places these inequalities inside an SDP/Lagrangian framework. The “Shor + RLT + Triangle” semidefinite relaxation augments the Shor matrix inequality

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$1

with quadratic constraints, diagonal RLT bounds, McCormick bilinear inequalities, and all triangle inequalities written as linear forms in $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$2 (Lambert, 2020). The central theorem states that the optimal value $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$3 of the best convex quadratic relaxation obtained by choosing positive semidefinite matrices $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$4 equals the optimal value $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$5 of this full SDP: $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$6 The corresponding explicit convex quadratic reformulation $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$7 satisfies $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$8, with the matrix $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$9 constructed from optimal dual multipliers associated with quadratic, McCormick, triangle, and Shor constraints (Lambert, 2020).

From the Lagrangian viewpoint, this is the key mechanism. Triangle inequalities appear in two roles. First, they are explicit valid inequalities in the final reformulation $Y_{ij}\approx x_ix_j$0. Second, when dualized, their multipliers are embedded into $Y_{ij}\approx x_ix_j$1, so they shape the optimal convexification of the objective in the $Y_{ij}\approx x_ix_j$2 space (Lambert, 2020). This dual embedding is why the inequalities are “Lagrangian-based” in a strong sense: their contribution is not merely combinatorial tightening of a lifted polyhedron but part of the optimal dual certificate defining the convex relaxation.

Because the full set of inequalities is large, the paper uses a partial Lagrangian dual in which McCormick and triangle inequalities are dualized and handled by a dynamic bundle method. Violated inequalities are separated from primal maximizers of the dualized SDP, active sets are capped by a parameter $Y_{ij}\approx x_ix_j$3, and constraints with small dual multipliers are removed (Lambert, 2020). This produces a Lagrangian-based cutting-plane mechanism that preserves scalability while exploiting the stronger triple-wise structure of triangle inequalities.

The computational effect is explicit on the 135 “unitbox” QCQPs from Bao–Sahinidis–Tawarmalani (2009). With a 2-hour limit, MIQCR-T solves 128 of 135 instances, compared with 119 for MIQCR, 110 for GloMIQO 2, and 109 for BARON 9.0.4. On 109 instances solved by both MIQCR and MIQCR-T, the average root gap decreases from 1.63% to 1.18%. The average number of branch-and-bound nodes is reduced by a factor $Y_{ij}\approx x_ix_j$4, and average B&B CPU time drops from 390 s to 85 s (Lambert, 2020). These figures directly support the claim that the triangle inequalities materially strengthen the node relaxation and accelerate global solution.

3. CDT: supporting hyperplanes as redundant but strengthening cuts

The CDT problem minimizes a quadratic objective over the intersection of the unit ball and a strictly convex quadratic constraint

$Y_{ij}\approx x_ix_j$5

with $Y_{ij}\approx x_ix_j$6 (Consolini et al., 2021). The paper keeps the unit-ball constraint explicit and relaxes only the ellipsoidal constraint through a scalar multiplier $Y_{ij}\approx x_ix_j$7, producing

$Y_{ij}\approx x_ix_j$8

and the dual bound

$Y_{ij}\approx x_ix_j$9

For $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$0, this bound coincides with the Shor semidefinite relaxation bound for CDT (Consolini et al., 2021).

A central structural result is the exactness criterion. The paper introduces the set-valued map $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$1, where $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$2 is the set of minimizers of the Lagrangian subproblem. Under the standing assumptions, exactness is characterized by whether $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$3, equivalently whether some optimal Lagrangian solution lies on the boundary of the ellipsoid (Consolini et al., 2021). The non-exact case occurs when, at the optimal multiplier $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$4, the trust-region subproblem has exactly two minimizers, one in $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$5 and one outside $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$6, so that the Lagrangian/Shor value is strictly below the true optimum (Consolini et al., 2021).

The strengthening mechanism then becomes geometrically transparent. For a boundary point $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$7, convexity of the ellipsoidal constraint yields the supporting hyperplane

$(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$8

This inequality is redundant for the original CDT problem because every feasible point of the ellipsoid satisfies all supporting hyperplanes. It is not redundant for the Lagrangian relaxation because a minimizer outside $(u_i-x_i)(u_j-x_j)(u_k-x_k)\ge 0$9 can violate it (Consolini et al., 2021). The paper chooses $Y_{pq}$0 by projecting a violating Lagrangian minimizer $Y_{pq}$1 onto the boundary via

$Y_{pq}$2

and then defines the halfspace $Y_{pq}$3 by the supporting inequality (Consolini et al., 2021).

If the original dual bound is inexact and the cut removes all Lagrangian minimizers violating the ellipsoid at the previous dual optimum, then the new bound strictly improves. This is formalized by Proposition 4.1: with $Y_{pq}$4, Algorithm DualLagrangian returns $Y_{pq}$5 and a smaller maximizing multiplier $Y_{pq}$6 (Consolini et al., 2021). The paper also constructs a two-cut variant and local adjustment procedures that perturb cut points on $Y_{pq}$7 to remove residual violating minimizers and further improve the bound (Consolini et al., 2021).

The computational results are strong and highly specific. On the 212 “hard” CDT instances from Burer–Anstreicher (2013), the original Lagrangian/Shor bound (LbDual) is strengthened successively by LbOneCut, LbOneAdj, LbTwoCut, and LbTwoAdj. For $Y_{pq}$8, the average relative gaps are 0.20%, 0.08%, 0.05%, 0.03%, and 0.05%, respectively. The bound LbTwoAdj solves 211 of the 212 hard instances, compared with 125 of 212 for LbTwoCut; the abstract summarizes this by stating that one of the proposed bounds solves all but one of the 212 hard instances (Consolini et al., 2021). These facts make the paper one of the clearest demonstrations that a very small number of carefully targeted valid inequalities can nearly close a difficult Lagrangian duality gap.

A plausible implication is that “Lagrangian-based valid inequalities” need not form a large generic family. In CDT, one or two supporting hyperplanes, chosen from the geometry of the violating Lagrangian minimizers, can be enough to transform a weak but tractable dual bound into an almost exact one (Consolini et al., 2021).

4. Block-structured MIPs: redundant monomials, decomposition, and zero duality gap

For block-structured MIPs, the defining tension is between decomposability and dual strength. The prototypical two-block model is

$Y_{pq}$9

Classical Lagrangian relaxation dualizes the coupling $x_px_q$0 with multipliers $x_px_q$1, producing a decomposable dual function $x_px_q$2, but generally with nonzero duality gap because the subproblems are MIPs (Cifuentes et al., 2024).

The 2024 paper proposes adding redundant constraints implied by the binary linking and dualizing them together with the original coupling constraints in a way that preserves block separability (Cifuentes et al., 2024). In general form, if

$x_px_q$3

then one introduces local auxiliary variables $x_px_q$4, adds the redundant constraint $x_px_q$5, and dualizes both this and the original link (Cifuentes et al., 2024). A preliminary structural limit is Proposition 3.1: if either $x_px_q$6 or $x_px_q$7 is affine, then the strengthened dual value is unchanged, so improvement requires genuinely non-affine redundant constraints (Cifuentes et al., 2024).

Two specific constructions are developed.

The first is the monomial-based reformulation, or M-Lagrangian. For a family $x_px_q$8, one adds equalities

$x_px_q$9

introduces local monomial variables $C$0, and dualizes the equalities $C$1 (Cifuentes et al., 2024). If $C$2 is down-closed, Theorem 4.1 establishes $C$3, where $C$4 is the projected convexified feasible set of the extended formulation and $C$5 is a simpler set obtained by enforcing equality over subsets. When $C$6, the M-Lagrangian has zero duality gap: $C$7 The hierarchy $C$8 yields $C$9, with $(i,j,k)$0 recovering the classical Lagrangian dual and $(i,j,k)$1 giving exactness (Cifuentes et al., 2024).

The second is the vertex-based reformulation, or V-Lagrangian. For each binary vertex $(i,j,k)$2, define

$(i,j,k)$3

and add the equalities

$(i,j,k)$4

With variables $(i,j,k)$5, dualizing $(i,j,k)$6 yields a dual whose strong duality is proved in Theorem 5.1: $(i,j,k)$7 The argument uses affine independence of the lifted points $(i,j,k)$8 over binary vertices and a decomposability proposition showing that the convexified linked set coincides with the convex hull of the original feasible set (Cifuentes et al., 2024).

The notable feature is that exactness is achieved without giving up decomposition. The auxiliary monomial or vertex variables are local to each block; only the equalities tying corresponding lifted variables across blocks are dualized. Hence the inner Lagrangian minimization remains fully separable by block and suitable for parallel computing (Cifuentes et al., 2024). This is why the paper frames the construction as zero-duality-gap Lagrangian duals that still admit decomposition.

The paper further extends the framework to general sparse MIPs via the intersection graph and a tree-decomposition. A generic MIP

$(i,j,k)$9

is reformulated over bags of a tree-decomposition, with coupling on shared binary variables. The same monomial lifting is then applied edgewise over bag intersections $i<j<k$0, yielding a decomposable tree-structured M-Lagrangian. Under down-closed families $i<j<k$1, Theorem 6.4 proves the analog $i<j<k$2, and if $i<j<k$3, strong duality again follows (Cifuentes et al., 2024).

The hierarchy also admits multiplicative approximation guarantees for packing and covering MIPs. For the two-block case, with $i<j<k$4, Theorem 6.2 gives for packing under recourse

$i<j<k$5

and Theorem 6.3 gives for covering

$i<j<k$6

(Cifuentes et al., 2024). For general multi-block tree-structured instances, the guarantees are expressed in terms of LP-derived quantities $i<j<k$7 and $i<j<k$8 and the bag-overlap parameter $i<j<k$9, yielding

$\ell_i=0$0

for packing and

$\ell_i=0$1

for covering (Cifuentes et al., 2024).

Preliminary experiments confirm that even low-order lifted valid inequalities strengthen the dual meaningfully. On synthetic block-structured stable-set instances, the quadratic M-Lagrangian (QL, level $\ell_i=0$2) improves the average primal-dual gap relative to the classical Lagrangian dual (L), and the V-Lagrangian (VL) is best on the STAR-STAB class. Specifically, on STAR-STAB the average gaps are 10.0% for Gurobi, 6.0% for L, 4.4% for QL, 3.9% for VL, and 8.6% for SDA; on PATH-STAB they are 10.8% for Gurobi, 3.5% for L, and 1.2% for QL (Cifuentes et al., 2024). The paper explicitly notes that Gurobi is not exploiting decomposability, which is central to the comparative advantage of the Lagrangian-based formulations (Cifuentes et al., 2024).

5. Active inequalities, augmented Lagrangians, and algorithmic identification

A different perspective on Lagrangian-based inequalities appears in the augmented Lagrangian paper on nonlinear programs with inequality constraints (Toussaint, 2014). The work does not add cutting planes in the usual sense, but it develops a mechanism for identifying and stabilizing the inequalities that are active at optimality. The base problem is

$\ell_i=0$3

with multipliers $\ell_i=0$4 for inequalities and $\ell_i=0$5 for equalities (Toussaint, 2014).

The key definition is activity of the $\ell_i=0$6-th inequality: it is active iff $\ell_i=0$7. This leads to the indicator matrix

$\ell_i=0$8

and the proposed augmented Lagrangian

$\ell_i=0$9

Quadratic penalties are therefore applied only to constraints regarded as active, and the gradient contains the effective multiplier-like quantity $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$00 (Toussaint, 2014).

The centered dual update is

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$01

and the paper proves a one-step KKT result in the LP case under fixed activity and linear independence of active constraints: if no previously positive multiplier becomes zero and active rows remain linearly independent, then after one outer iteration the KKT conditions hold at the new primal minimizer (Toussaint, 2014). The anytime update generalizes this to non-stationary $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$02 by solving a small least-squares QP for $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$03, so that penalty-generated gradients are absorbed into updated multipliers (Toussaint, 2014).

The relevance to Lagrangian-based valid inequalities is interpretive rather than combinatorial. The paper can be read as showing how an augmented Lagrangian method dynamically discovers which inequalities are effectively valid and tight at the solution. When the active set is stable, the algorithm behaves like an active-set method that quickly identifies the supporting hyperplanes defining the optimal face; when activity is unstable, performance deteriorates (Toussaint, 2014). This suggests that in many optimization schemes the main issue is not only which valid inequalities exist, but whether the method can robustly learn and maintain the subset that should carry nonzero dual weight.

The empirical contrast in the paper reinforces that point. Any-time Augmented Lagrangian performs well on robot trajectory optimization, where constraint activity is described as moderately stable, but deteriorates on random LPs, where active sets are highly volatile (Toussaint, 2014). This is not a cutting-plane result, yet it bears directly on how Lagrangian methods exploit inequality structure in practice.

6. Lagrangian structure beyond optimization relaxations: sharp curvature inequalities

In differential geometry, the term “Lagrangian-based” refers to inequalities whose strength depends on the Lagrangian condition of a submanifold. The 2013 paper studies Lagrangian submanifolds $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$04 of a complex space form $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$05, including $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$06, $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$07, and $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$08 according to the sign of $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$09 (Chen et al., 2013). A Lagrangian immersion satisfies $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$10, and the cubic form

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$11

is totally symmetric, so the coefficients $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$12 are symmetric in all three indices (Chen et al., 2013). This algebraic symmetry is the decisive structural input.

The paper considers Chen’s $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$13-invariants

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$14

where the infimum is over mutually orthogonal subspaces $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$15 of dimensions $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$16 (Chen et al., 2013). The main result is an optimal intrinsic-extrinsic bound

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$17

with explicit best possible coefficients $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$18 and $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$19 (Chen et al., 2013).

Two cases are distinguished. If $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$20, Theorem 3.1 yields

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$21

where $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$22 (Chen et al., 2013). If $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$23, Theorem 3.2 gives the stronger bound

$\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$24

which strictly improves the coefficient of $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$25 in the full-splitting case (Chen et al., 2013).

The paper’s title, “the final solution,” refers to correctness, optimality, and equality characterization. It corrects an earlier inequality by Chen and Dillen whose proof was incorrect when $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$26, and it shows that the earlier coefficient is not always optimal when $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$27 (Chen et al., 2013). The proof rewrites $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$28 via the Gauss equation, reduces the problem to nonnegativity of quadratic forms in the diagonal components $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$29, analyzes the corresponding block-structured matrix $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$30, and uses an explicit determinant formula from Lemma 2.2 together with Sylvester’s criterion to identify the best constant $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$31 (Chen et al., 2013). Sharpness follows from Lemma 2.1, which states that any fully symmetric triple system can be realized as the cubic form of a Lagrangian immersion in $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$32 at a point, so algebraic extremizers correspond to genuine submanifolds (Chen et al., 2013).

This is not an optimization application, but it is an instructive analogue. The inequality is “Lagrangian-based” because the coefficients and equality conditions are unattainable without the total symmetry imposed by the Lagrangian condition. The left side is intrinsic and the right side mixes extrinsic mean curvature and ambient curvature, so the result plays a role analogous to a sharp valid inequality linking two representations of structure (Chen et al., 2013).

7. Unifying themes, distinctions, and common misconceptions

Across these domains, several common principles emerge.

First, redundancy at the primal level does not imply irrelevance at the relaxation level. Supporting hyperplanes of the CDT ellipsoid, triangle inequalities for QCQP, and monomial or vertex equalities in MIP decomposition are all redundant for the original feasible set, yet they strengthen Lagrangian or lifted relaxations precisely because the relaxation has discarded nonlinear or integrality structure (Consolini et al., 2021, Lambert, 2020, Cifuentes et al., 2024).

Second, the strongest inequalities are usually structure-specific. The general triangle inequalities depend explicitly on the variable bounds $\begin{numcases}{(P)} \min\ f_0(x) \equiv \langle Q_0, xx^T \rangle + c_0^T x \nonumber\ \text{s.t.}\quad f_r(x) \equiv \langle Q_r, xx^T \rangle + c_r^T x \le b_r, & r \in R:=\{1,\dots,m\} \ \ell_i \le x_i \le u_i, & i\in I:=\{1,\dots,n\} \ x_i \in \mathbb{R}, & i\in I, \end{numcases}$33 and on the McCormick system; they are the only 12 of 48 generated triple inequalities that are non-redundant relative to McCormick (Lambert, 2020). The CDT cuts are not generic linearizations but supporting hyperplanes chosen from projections of violating Lagrangian minimizers (Consolini et al., 2021). The M-Lagrangian and V-Lagrangian constraints are carefully selected to preserve block separability after dualization; affine redundant constraints do not help (Cifuentes et al., 2024). In the geometric setting, the curvature inequalities are sharpened only because the Lagrangian cubic form is totally symmetric (Chen et al., 2013).

Third, preservation of tractable structure is central. In QCQP, triangle inequalities are integrated through a partial Lagrangian dual and bundle separation rather than inserted wholesale into a monolithic SDP (Lambert, 2020). In CDT, the strengthened subproblems remain trust-region subproblems with one or two linear constraints, so they retain polynomial-time solvability (Consolini et al., 2021). In block-structured MIPs, the lifted valid inequalities are deliberately chosen so that the inner Lagrangian problem stays decomposable across bags or blocks (Cifuentes et al., 2024).

A common misconception is that stronger valid inequalities must always be generated in large numbers. The CDT results show that one or two carefully chosen cuts can nearly close the full gap of the Lagrangian/Shor bound (Consolini et al., 2021). Another misconception is that exactness and decomposition are inherently opposed in nonconvex discrete problems. The M-Lagrangian and V-Lagrangian constructions show that, at least for binary coupling under appropriate lifted reformulations, one can have both decomposition and zero duality gap (Cifuentes et al., 2024). A third misconception is that Lagrangian-based inequalities are exclusively an optimization-device notion. The curvature results indicate that the phrase can also refer to inequalities whose validity and optimality fundamentally rely on Lagrangian structure, even outside algorithmic duality (Chen et al., 2013).

A plausible implication is that the most effective future developments will continue to be hybrid. The QCQP work combines explicit cuts, SDP structure, dual embedding, and spatial branch-and-bound (Lambert, 2020). The CDT work combines geometric cut generation, exactness analysis, and efficient subproblem solvers (Consolini et al., 2021). The MIP decomposition work combines hierarchical lifting, tree-decomposition, and Lagrangian separability (Cifuentes et al., 2024). The augmented Lagrangian work adds the algorithmic lesson that correctly identifying and stabilizing active inequalities can be as important as inventing new ones (Toussaint, 2014).

Taken together, these papers show that Lagrangian-based valid inequalities are best understood not as a single formal class but as a design principle: exploit problem-specific redundant structure so that, after relaxation or dualization, the bound becomes sharper while the exploitable structure of the subproblem is preserved. Whether the goal is stronger QCQP relaxations, near-exact CDT bounds, zero-gap decomposable MIP duals, rapid active-set identification, or sharp curvature inequalities, the decisive step is the same: use Lagrangian structure to identify the right inequalities, and use the inequalities to reveal the full power of the Lagrangian structure.