Introduction to Exact Penalization for Mathematical Programming with Equilibrium Constraints

Abstract: We present a focused introduction to exact penalty methods for nonlinear programs and mathematical programs with equilibrium constraints (MPECs), emphasizing their connection to modern error bound theory. The goal is twofold. First, we explain how classical optimality conditions can be interpreted through exact penalization, and why such results typically rely on constraint regularity conditions that can be understood as error bounds on perturbations of feasible sets. We then highlight how recent developments based on subanalytic geometry and Lojasiewicz-type inequalities extend this framework beyond classical regularity assumptions, enabling exact penalization under broader analytic conditions. Second, we demonstrate how this theory can be applied in practice to MPECs by reformulating them via KKT systems and constructing exact penalty functions based on residual mappings. Particular attention is given to fractional-order penalties arising from Lojasiewicz error bounds, as well as to improved formulations for special problem classes where sharper exponents can be obtained. These developments provide both theoretical insight and practical guidance for analyzing and solving challenging constrained optimization problems.

• The paper establishes that under subanalytic conditions, an error bound guarantees the existence of an exact penalty function.
• The methodology leverages Łojasiewicz inequalities to relate residuals with feasibility distances, accommodating both regular and degenerate MPEC regimes.
• The analysis offers practical insights for reformulating MPECs, highlighting fractional power penalties and implications for algorithmic convergence.

Exact Penalization in Mathematical Programs with Equilibrium Constraints: Analytic Structure and Algorithmic Implications

Overview and Motivation

This paper provides a systematic exposition of exact penalization techniques for mathematical programming with equilibrium constraints (MPECs), with a particular focus on the analytic underpinnings that justify the use of exact penalty methods in non-classical, nonregular regimes (2605.00387). The main objectives are (i) to relate classical and modern error bound theory directly to the construction of exact penalty functions for MPECs, and (ii) to clarify conditions—especially analytic and geometric—that permit such penalization even in the absence of strong regularity assumptions commonly leveraged in nonlinear programming. The exposition synthesizes traditional optimality conditions, recent subanalytic geometry methods, and practical algorithmic construction for optimization problems with challenging constraint structures.

MPECs: Structural Obstacles and Reformulation

MPECs constitute a challenging class of constrained optimization problems where feasibility is encoded via the solution set of an implicit equilibrium problem, frequently in the form of variational inequalities (VIs), complementarity systems, or generalized equations. Two core structural difficulties arise: (1) the feasible set is defined implicitly by a lower-level equilibrium; (2) standard one-level reformulations introduce complementarity constraints, which render classical constraint qualifications and KKT-based theory inapplicable.

Traditional approaches to MPECs are undermined by the failure of classical constraint qualifications at points of interest, making the development of robust optimality characterizations and solution algorithms particularly challenging. Reformulation via the KKT conditions of the lower-level problem converts the equilibrium constraint into a finite system in upper-level variables, lower-level variables, and multipliers. This conversion is essential for analytic and algorithmic approaches to exact penalization.

Residual Functions and the Principle of Exact Penalization

The core methodology replaces hard (infeasible) equilibrium constraints with a penalty term based on residuals, appended to the objective. The ideal is an "exact" penalty: global minimizers of the penalized problem coincide with those of the original, provided the penalty parameter is sufficiently large. In standard NLPs, such penalties are typically $\ell_1$-type. However, for MPECs—where error bounds are usually weaker—fractional powers of the residual often become necessary.

Penalization is not merely a matter of residual cancellation; an appropriate quantitative relationship (error bound) must link the residual and the distance to feasibility. The paper makes precise that exact penalization is possible if and only if such an analytic error bound exists.

Analytic Framework: Subanalyticity and Łojasiewicz Inequalities

The analytic core rests on characterizing feasible sets and residual maps in the class of subanalytic sets and functions, which is broad enough to accommodate nonsmoothness, piecewise-analyticity, and finite-dimensional generalized equilibrium structures. The critical result leveraged is a Łojasiewicz-type inequality, asserting that for continuous subanalytic functions $\phi,\psi$ and compact subanalytic $S$ with $\phi^{-1}(0) \subseteq \psi^{-1}(0)$, there exist $\rho > 0$, $N^* \in \mathbb{N}$ such that:

$$\rho\, [\psi(x)]^{N^*} \leq \phi(x),\quad \forall x\in S.$$

Specialized to residuals and distances to the feasible set, this yields:

$\operatorname{dist}(x, W) \leq c\, r(x)^{1/N^*},$

with $N^*\ge1$, and typically $N^* > 1$ in degenerate or nonregular cases.

Main Theoretical Result: General and MPEC-Specific Exact Penalty Theorems

The author states and sketches the proof of a general exact penalty theorem: given compactness and subanalyticity of the feasible domain and data, Lipschitz continuity of the objective, and continuity and subanalyticity of the constraining functions, existence of an exact penalty of the form $\phi,\psi$0 is guaranteed for sufficiently large $\phi,\psi$1.

In the MPEC setting, under analogous regularity (subanalyticity, convexity of constraint mappings, KKT equivalence on the feasible set, and bounded multipliers), the KKT-based residual permits an exact penalty with a similar fractional exponent structure, where the bound on multipliers is essential for compactness and uniformity of analytic constants.

The regime splits naturally:

• If classical regularity holds, $\phi,\psi$2 and $\phi,\psi$3 penalties suffice.
• In highly degenerate settings, only H\"older-type error bounds may hold, necessitating penalties with fractional exponents.

Critically, the theory provides an explicit link: Error bound $\phi,\psi$4 existence of exact penalty with exponent $\phi,\psi$5.

Improving Error Bound Exponents: Problem Structure and Explicit Computation

The general subanalytic theory often yields exponents that are inherently nonconstructive (i.e., $\phi,\psi$6 is not computable via the proof). The author discusses special classes—such as quadratic systems with a nonnegativity structure, affine VIs, and NCPs with the uniform P-property—where improved (explicit) error bounds are provable. For example, AVIs admit order-1 bounds, while certain quadratic complementarity systems require order-1/2. In such settings, the penalty formulation becomes more interpretable and enables direct comparison of original and penalized local minima, with stronger stationarity analysis and algorithmic design implications.

Local vs. Global Exactness and Algorithmic Issues

A detailed caveat is given: global exactness does not imply local exactness in both directions. Penalized problems may possess spurious infeasible local minimizers, especially acute in MPECs with complementarity structure. Therefore, careful distinction must be made between convergence to minimizers of the penalized model, convergence to feasibility for the original MPEC, and stationarity notions for both formulations.

Algorithmically, the described workflow involves:

1. Reformulating the MPEC as a KKT system,
2. Building a residual mapping aggregating stationarity, feasibility, and complementarity violations,
3. Penalizing with an analytically justified exponent,
4. Checking feasibility and stationarity a posteriori,
5. Leveraging further problem structure to improve the exponent when possible.

Conceptual Synthesis and Broader Implications

The theory is succinctly encoded as:

$\phi,\psi$7

with several key implications:

• Subanalytic geometry and the Łojasiewicz framework enable exact penalty results well beyond the classical landscape, especially in degenerate cases.
• The algorithmic readiness relies on explicit calculation or estimation of the error bound exponent, which remains an open issue in general.
• For optimization in AI and operations research applications (e.g., hierarchical games, bilevel learning, network equilibrium), the analytic guarantees for exact penalization directly inform surrogate modeling, global solution schemes, and stationarity analysis.
• High degeneracy and the appearance of subanalytic structure are inherent in many modern applications, exacerbating the need for robust penalty approaches that can operate outside the regime of classical regularity.

Conclusion

This paper bridges classical nonlinear programming and modern error bound theory in the context of MPECs by rigorously establishing the existence and structure of exact penalty functions under subanalytic assumptions (2605.00387). The analytic machinery of subanalytic sets and Łojasiewicz inequalities enables exact penalization in regimes where standard regularity is absent, at the conceptual and algorithmic price of fractional power penalties. The results both unify a large body of variational analytic theory and provide a template for future algorithmic and theoretical development, notably in the improvement of computable error bound exponents and in the precise design of penalty-based numerical methods for MPECs and related equilibrium-constraint models.