First-Order Optimality Conditions for Mathematical Programming with Equilibrium Constraints

Abstract: We present a systematic introduction to first-order optimality conditions for mathematical programs with equilibrium constraints (MPECs), emphasizing the limitations of classical nonlinear programming techniques. The goal is twofold. First, we explain why a direct application of standard optimality conditions -- based on reformulating MPECs via KKT systems or differentiable exact penalty functions -- is often inadequate, as such approaches typically require strong and restrictive assumptions, including nondegeneracy and smoothness conditions. Second, we develop a first-principles framework for analyzing MPECs by focusing on the geometric structure of the feasible region. In particular, we study stationarity concepts and provide a detailed characterization of the tangent cone at feasible points, which leads to appropriate constraint qualifications tailored to MPECs. These results form the foundation for rigorous first-order analysis and clarify the relationship between the original MPEC formulation and its KKT-based representation, offering practical guidance for handling these inherently challenging optimization problems.

• The paper introduces novel first-order optimality conditions for MPECs that account for the nonconvexity and degeneracy of feasible sets.
• It develops a detailed geometric analysis using tangent cone decompositions and specialized constraint qualifications to correctly characterize stationarity.
• The paper demonstrates that standard NLP reformulations, such as KKT-based methods, fail to capture true local optimality in MPECs due to their inherent nonconvex structure.

First-Order Optimality Conditions for Mathematical Programs with Equilibrium Constraints

Introduction and Scope

Mathematical Programs with Equilibrium Constraints (MPECs) constitute a critical framework for modeling hierarchical optimization problems coupling upper-level decision variables with lower-level equilibria (e.g., variational inequalities or complementarity systems). This paper systematically analyzes first-order optimality conditions for MPECs, emphasizing the inadequacy of direct nonlinear programming (NLP) reformulations (notably KKT-based or penalty methods) due to inherent nonconvexity and degeneracy in MPEC feasible sets. The study develops a geometric and variational analytic approach, scrutinizing tangent cone structures and introducing specialized constraint qualifications (CQs) tailored to the MPEC setting.

Preliminaries: Tangent Cones, Stationarity, and Constraint Qualifications

The feasible set $F$ of an MPEC is expressed as the intersection of the upper-level constraints $Z$ and the graph of the solution mapping of a lower-level variational inequality or complementarity system. At a feasible point $\bar{z} = (\bar{x}, \bar{y})$, the local geometry of $F$ diverges substantially from that of a classical NLP. Specifically, $T(\bar{z}; F)$, the tangent cone to the feasible set, is generally nonconvex and often a union of polyhedral cones, rather than a single polyhedral cone as in standard NLPs under CQs.

Classical first-order necessary conditions rely on multipliers and KKT systems with primal-dual stationarity, but, as shown, these become inadequate in the general MPEC setting where nonconvexity or branching in the feasible set or multiplier set leads to failure of standard CQs. The analysis necessitates new notions:

• Sequence Boundedness CQ (SBCQ): Ensures boundedness of multiplier sequences for convergent sequences in $F$.
• Extreme and Full CQs: Relate the tangent cone to unions of polyhedral cones indexed by extreme points or all points in the multiplier set $M(\bar{z})$.

Geometric Analysis: Tangent Cones and MPEC-Specific Linearized Cones

The paper provides a comprehensive geometric decomposition:

• The tangent cone at $\bar{z}$ is shown to be contained in a nonconvex union of cones constructed from linearizations at each extreme (KKT) multiplier. The MPEC linearized cone $L(\bar{z}; F)$ is generally nonconvex.
• It is rigorously established that, under SBCQ and full CQ, $T(\bar{z}; F) = L(\bar{z}; F)$, and characterizations of stationarity can be equivalently written in terms of this union-of-cones structure.
• There is detailed exposition on the lifted critical cone, directional critical set, and an affine variational inequality structure that captures compatible directions under the lower-level equilibrium constraints.

Failure of Standard NLP Reformulations

A critical result is the demonstration that naïve replacement of the lower-level equilibrium constraint with its KKT system (and subsequent application of standard NLP linearization) is generally unsound, as it linearizes an object with inherently nonconvex tangent structure. The excess directions allowed by the NLP linearization lead to spurious stationarity conditions. Computational examples reinforce that the convex hull of the true tangent cone is generally a proper subset of the NLP linearization, thus confirming the contradictory claim that standard first-order analysis via KKT reformulation misrepresents local optimality for MPECs.

Constraint Qualification Hierarchy and Stationarity Characterizations

The study defines a precise hierarchy:

• Basic CQ: Tangent cone can be represented by a union over some nonempty subset of multipliers.
• Extreme CQ: The union is taken over extreme points of the multiplier set.
• Full CQ: The union is over all multipliers. Under constant rank conditions, inclusion of tangent directions into the union indexed by extreme multipliers is justified.

Equivalence results (stated as rigorous "if and only if" theorems) provide several primal and primal-dual formulations of stationarity, including reducibility to finite verification when the multiplier set is a polytope with finitely many extreme points.

Specialization: NCP-Constrained and Affine MPECs

In the frequently encountered NCP case, the lower-level problem reduces the multiplier set to a singleton, substantially simplifying the tangent cone and stationarity conditions (now described by classical complementarity structures and mixed LCPs). Directions and stationarity decompositions admit closed-form expressions, with degenerate indices controlling the combinatorics of possible systems.

The theoretical results are exemplified with explicit computations for various MPECs, including cases where the multiplier set is a line segment, the tangent cone is a nonconvex union, and the full/SMFCQ fails.

Practical and Theoretical Implications

The development has direct impact on the analysis and algorithmic solution of MPECs, especially for bilevel optimization, equilibrium-constrained control, and engineering design under equilibrium constraints:

• Optimality Verification: Valid stationarity and optimality verification at feasible points in MPECs must utilize the advanced geometric perspective provided, not the faulty NLP-based KKT systems.
• Algorithmic Design: Algorithms premised on NLP homotopy or sequential quadratic programming that ignore the nonconvex union structure of the tangent cone may fail to converge to local minimizers or misidentify stationary points.
• Theory of Variational Analysis: This framework clarifies the necessity and sufficiency of CQs in projecting KKT tangent cones back to the primal feasible set, even when solution mapping fails to be polyhedral or single-valued.

These insights open the possibility of developing more robust MPEC algorithms grounded in variational analysis and set-valued mapping theory, as well as improved sensitivity analysis using this tangent cone framework. Additionally, there are consequences for the computability/verification of stationarity and for the projected, penalized, or relaxation-based solution methods in complex MPECs.

Conclusion

This paper offers a rigorous and comprehensive first-order theory for MPECs, highlighting the geometric and variational obstacles that invalidate standard NLP-based approaches. By characterizing the tangent cone via unions over critical multiplier sets and introducing context-appropriate CQs, the work establishes necessary and sufficient first-order optimality conditions specific to MPEC geometry. The analytic and computational results have substantial ramifications for both theory and algorithm design in mathematical optimization, optimal control, and related domains involving equilibrium constraints.

Reference: "First-Order Optimality Conditions for Mathematical Programming with Equilibrium Constraints" (2605.00388)