KKT-Based Single-Level Reduction

• KKT-based single-level reduction is a methodology that replaces implicit lower-level optimality with explicit KKT conditions, converting bilevel problems into single-level nonlinear programs.
• The approach enables the use of generic optimization solvers and mixed-integer formulations for solving complex leader–follower games, bilevel linear programs, and model predictive control problems.
• Limitations include the potential failure of Slater’s condition and CRCQ, which may lead to discrepancies and feasibility issues between the original bilevel formulation and its MPCC reformulation.

KKT-based single-level reduction is a foundational methodology for translating bilevel (or hierarchical) optimization and equilibrium problems—where decisions at an upper level depend on the solution set of lower-level optimization or equilibrium subproblems—into tractable single-level (often nonconvex) nonlinear programs through the enforcement of Karush–Kuhn–Tucker (KKT) optimality conditions. The resulting Mathematical Program with Complementarity Constraints (MPCC) or other single-level nonlinear programs enable the use of generic optimization solvers, mixed-integer reductions, or further analytical simplifications. The approach is central to modern computational approaches for leader–follower games, bilevel programs with convex lower levels, Stackelberg-Nash equilibrium computation, bilevel linear and (convex) quadratic programming, model predictive control, and multi-agent ordered-preference game equilibria.

1. Formalism and Derivation of KKT-based Single-Level Reformulation

The central principle is to replace implicit lower-level optimality or Nash equilibrium conditions with their explicit KKT system, thereby embedding all optimality, feasibility, and complementarity constraints at once in a single-level program.

For a generic single-leader, multi-follower bilevel game, the problem is posed as

$$\min_{x \in X,\, y} \; F(x, y) \quad\text{s.t.}\quad y \in \mathrm{NEP}(x)$$

where $\mathrm{NEP}(x)$ is the Nash equilibrium set of the parameterized game among the $N$ followers, each solving

$$\begin{aligned} \min_{\hat y_i}\; f_i(x, \hat y_i, y_{-i}) \quad \text{s.t.} \; g_i(x, \hat y_i) \le 0 \end{aligned}$$

for $i = 1, \ldots, N$. Under convexity of each $f_i$ and $g_i$ in $y_i$ and appropriate constraint qualifications, the bilevel equilibrium is characterized by the KKT system for each follower:

• Stationarity: $$\nabla_{y_i}f_i(x,y_i,y_{-i}) + \nabla_{y_i}g_i(x,y_i)^\top \lambda_i = 0$$
• Primal feasibility: $g_i(x,y_i) \le 0$
• Dual feasibility: $\mathrm{NEP}(x)$0
• Complementary slackness: $\mathrm{NEP}(x)$1

Stacking the KKT systems for all followers yields a single-level MPCC for the leader’s problem: $\mathrm{NEP}(x)$2 This reduction transforms the nested equilibrium constraints into explicit algebraic and complementarity constraints amenable to solution by MPCC or mixed-integer programming techniques (Chaipunya et al., 19 Mar 2025).

2. Equivalence Conditions and Constraint Qualifications

Global and local solution equivalence between the original bilevel problem and its KKT-based single-level reformulation is governed by convexity and specific constraint qualifications:

• Global equivalence holds if, for all $\mathrm{NEP}(x)$3, the lower-level follower problems are convex and satisfy Slater's condition (existence of strictly feasible points for inequalities). Theorems in (Chaipunya et al., 19 Mar 2025) establish that every global solution to the original bilevel (SLMFG) corresponds to a global solution of the MPCC, and conversely, whenever these assumptions hold. If Slater fails (e.g., degenerate or nonstrictly feasible constraints), equivalence may be lost; the MPCC may have no feasible point despite the original problem admitting solutions.
• Local equivalence is more nuanced. The sufficiency of local minima transfer depends on whether all Lagrange multipliers at the candidate point yield local MPCC minima (an impractical check in general). With the Constant Rank Constraint Qualification (CRCQ) in place at each follower’s optimal solution, the local equivalence can be checked using only the vertex multipliers (extreme points), reducing the required number of MPCC verifications. CRCQ requires the Jacobian of all subsets of active constraints to maintain constant rank in a neighborhood; its violation can lead to spurious or missing local minima in the MPCC (Chaipunya et al., 19 Mar 2025).

3. Computational Schemes and Linearization

KKT-based reformulations are agnostic regarding the nature of $\mathrm{NEP}(x)$4, $\mathrm{NEP}(x)$5, and $\mathrm{NEP}(x)$6, but for bilevel linear programs, the entire system can be further linearized and encoded as a Mixed-Integer Linear Program (MILP).

Following (Buchheim, 2023), for bilevel linear programs: $\mathrm{NEP}(x)$7 the lower-level LP is replaced by its KKT conditions, and complementarity slackness is linearized via big-M and binary variables: $\mathrm{NEP}(x)$8 for suitably calibrated $\mathrm{NEP}(x)$9, ensuring polynomial encoding length and explicit NP-membership for the bilevel linear optimization class.

For nonlinear but convex lower-level programs (e.g., quadratic), direct enforcement of the complementarity constraints is typically handled via nonlinear solvers for MPCCs, or by employing relaxation or penalty algorithms. Numerical comparisons have shown that duality-based reformulations can, in large or ill-conditioned problems, outperform the KKT-based MPCC in solution quality and robustness, though the MPCC is fastest for small problems (Li et al., 3 Dec 2025).

4. Limitations and Pathological Cases

KKT-based single-level reduction is not universally exact. Notable pathologies and limitations include:

• Failure of Slater’s Condition: Degeneracy or infeasibility of the KKT system at certain $N$0 means the MPCC may not inherit the feasible region of the original bilevel game, as explicitly constructed in counterexamples (Chaipunya et al., 19 Mar 2025).
• Failure of CRCQ: If the constant rank qualification fails, local minima may be admitted by the MPCC that do not correspond to true local optima of the original bilevel or Nash equilibrium problem. Conversely, genuine local solutions might be absent in the reformulation.
• Ill-posed Complementarity Constraints: MPCCs often violate constraint qualifications (e.g., MFCQ or LICQ) at solutions, leading to violations of standard second-order conditions and causing difficulties for typical NLP solvers (Li et al., 3 Dec 2025).

5. Extensions and Analytical Reductions

Recent research has extended KKT-based reductions to large-scale and structured classes of problems. Notable examples include:

• Model Predictive Control (MPC): For bilevel MPC architectures, block-matrix analytical reductions can sidestep MPCC nonconvexity altogether. Under a block-nonsingularity condition, the lower-level MPC QP is condensed analytically into stationarity constraints, yielding a smooth single-level convex problem in reference and control variables, preserving optimality and closed-loop properties (Moriyasu et al., 31 Mar 2026).
• Density-Driven Optimal Control: In multi-agent control, the entire horizon-length KKT system can be analytically eliminated, condensing the regime to a low-dimensional QP or QCQP with linear complexity in the prediction horizon, replacing $N$1 complexity with $N$2 (Martinez et al., 19 Mar 2026).
• Games of Ordered Preference (GOOPs): For multi-player Nash equilibrium problems with hierarchical lexicographic objectives, KKT-based reductions of the full system are exponentially sized. A reduced KKT formulation enforces only essential primal stationarity and aggregates lower-level complementarity, yielding a system whose size grows polynomially with the number of priority levels and players. In the quadratic-linear setting, this reduced system is without loss; in the general nonlinear setting, sufficiency must be checked by second-order conditions (Lee et al., 27 Mar 2026).

6. Numerical Performance and Practical Considerations

Empirical studies reveal that:

• MPCC (KKT-based) reformulations are often the fastest to solve for small- to medium-scale convex bilevel programs but are less reliable in consistently producing globally best bilevel objectives compared to duality-based single-level reformulations. In large or difficult bilevels, duality-based reformulations (e.g., Wolfe, Mond-Weir, etc.) deliver better robustness and accuracy. For the 450 random bilevel test cases in (Li et al., 3 Dec 2025), the MPCC direct approach gave the best or near-best solution in 63 of 450 cases, versus 67–88 for duality-based methods. Relaxation also improves robustness for MPCC, but remains outperformed by strong duality-based approaches by a factor of three to five in dominance (Li et al., 3 Dec 2025).

Effective implementation may require:

• Relaxation or regularization of complementarity in MPCC
• Explicit projection to guarantee true bile