Equilibrium Constraints (PTEC)

• Equilibrium Constraints (PTEC) are restrictions that require certain variables to satisfy a secondary equilibrium, such as variational inequalities or complementarity conditions, within an optimization model.
• They enable hierarchical and multi-agent frameworks by modeling complex interactions in economics, game theory, and engineering, affecting optimality conditions and algorithmic designs.
• Methodologies include complementarity reformulations, SOS/SDP relaxations, and decomposition techniques that tackle nonconvexity and weakened constraint qualifications.

An equilibrium constraint is a restriction within an optimization or equilibrium problem that imposes the requirement that certain variables themselves must solve (or satisfy) another parameterized equilibrium, variational inequality, inclusion, or saddle-point system. Problems with such constraints are denoted as PTEC (Problems with Equilibrium Constraints) and constitute a vast class of hierarchical and multi-agent models in optimization, economics, engineering, and game theory. The PTEC class subsumes mathematical programs with equilibrium constraints (MPECs), bilevel programs with equilibrium constraints, equilibrium problems with domain restrictions, and constrained Nash or generalized Nash equilibrium problems. The presence of equilibrium constraints fundamentally alters the optimality conditions, constraint qualifications, modeling choices, and algorithmic frameworks applicable to these problems.

1. Mathematical Formulation of Equilibrium-Constrained Problems

Equilibrium constraints can appear as either parameterized variational inequalities, generalized equations, or vector equilibrium problems within a larger optimization or equilibrium model. The general template is as follows (Gfrerer et al., 2016, Uderzo, 2022, Jiao et al., 2019):

• Standard MPEC/MPGE (Mathematical Program with Equilibrium Constraints):

$$\min_{x \in \mathbb{R}^n,\, y \in \mathbb{R}^m} F(x,y) \quad \text{s.t.} \quad 0 \in \phi(x, y) + \widehat{N}_I(y), \quad G(x, y) \leq 0$$

where $\phi$ is an equilibrium mapping, $G$ is an upper-level constraint, and $\widehat{N}_I(y)$ is the regular (Fréchet) normal cone to $I = \{y \mid g(y) \leq 0\}$.

• Vector Equilibrium-Constrained Program (VEP-based MPEC):

$$\min_{\theta \in \Theta,\, x \in \mathbb{R}^n} \varphi(\theta, x) \quad \text{s.t.} \quad x \in \operatorname{SE}(\theta),\, (\theta, x) \in Q$$

where $$\operatorname{SE}(\theta) = \{x \in K(\theta) : f(\theta, x, z) \in C,\, \forall z \in K(\theta)\}$$ and $f$ is a cone-valued bifunction (Uderzo, 2022).

• Complementarity-based (MPCC) Reformulation: Many classical works formulate equilibrium constraints via complementarity constraints using KKT conditions, e.g.,

$$0 = \phi(x, y) + \nabla g(y)^\top \lambda, \quad 0 \leq -g(y) \perp \lambda \geq 0$$

but this introduces extra variables and often fails constraint qualifications (Gfrerer et al., 2016, Gfrerer et al., 2019).

• Polynomial PTECs: For all-polynomial data, an inner variational inequality constraint $p(x, y, v) \geq 0$ for all $\phi$0 can equivalently be replaced via a value function $\phi$1 and the constraint $\phi$2, enabling global semidefinite relaxations (Jiao et al., 2019).

2. Variational Analysis: Stationarity and Constraint Qualifications

The analysis and computation for equilibrium-constrained problems require advanced nonsmooth variational analysis due to the failure of classical regularity (MFCQ/LICQ) at nearly all points. Modern results rely on tools such as tangent/normal cones, metric subregularity, and error bounds (Gfrerer et al., 2016, Gfrerer et al., 2019, Uderzo, 2022):

• Metric Subregularity/Calmness: The mapping $\phi$3 is metrically subregular at $\phi$4 if small residuals in the inclusion imply proximity to feasible points. This property is crucial for deriving necessary optimality conditions and can be checked via first-order (FOSCMS) or second-order (SOSCMS) sufficient conditions on derivatives of data functions (Gfrerer et al., 2016).
• Softer Constraint Qualifications: Unlike MPCC-based CQs, the metric subregularity and first-order sufficient condition for metric subregularity (FOSCMS) are weaker and verifiable without introducing additional multipliers.
• Linearized M-stationarity: Gfrerer & Ye obtain sharper necessary optimality conditions by linearizing the normal cone constraint to the graphical tangent of the regular normal cone and introducing a second-order correction term. This goes beyond classical M-stationarity by requiring no MPCC constraint qualification and captures optimality even when all classical CQs fail (Gfrerer et al., 2019).
• Penalization via Residuals: A penalization scheme replaces the equilibrium constraint with a locally exact term involving the sum of the residual (distance to cone or feasibility) and distance to the geometric set. This approach delivers necessary conditions expressed by Mordukhovich subdifferential calculus for nonsmooth analysis (Uderzo, 2022).

3. Fundamental Classes and Applications

Equilibrium constraints arise in a diverse spectrum of models:

Model Class	Constraint Style	Reference
Bilevel programs with variational	Lower-level VI or GE, possibly vector-valued	(Gfrerer et al., 2016, Uderzo, 2022)
Complementarity-constrained (MPCC)	KKT + complementarity for embedded optimization	(Gfrerer et al., 2016, Gfrerer et al., 2019)
Multi-leader EPECs/EPEC (equilibrium)	Nash or generalized Nash equilibrium, shared KKT conditions	(Kim et al., 2020)
Self-supervised learning pre-train	Local stationarity enforced after $\phi$5 steps (bilevel)	(Cui et al., 27 Aug 2025)
General equilibrium with frictions	Market-clearing embedding portfolio/credit limits	(Abbot, 2017, Pham, 22 Jan 2025)
Ordered/constrained equilibrium	Existence in posets/Banach w/ restricted domains	(Li, 2017)

• Optimization with Lower-Level VIs and GEs: The constraint may state "the lower-level variable $\phi$6 solves a generalized equation parametrized by $\phi$7," formulated as $\phi$8 (Gfrerer et al., 2016, Gfrerer et al., 2019, Uderzo, 2022).
• Multiple Agent Games (EPEC, GNash): Multi-leader, multi-agent Stackelberg games with Nash equilibrium at the lower level and coupling via shared equilibrium are prime examples, with strategic agents each solving a MPEC parameterized by others' actions (Kim et al., 2020).
• Learning-Foundations Models: Implicit equilibrium constraints ensure rapid adaptation for all domains post-pretraining (PTEC for self-supervised learning) (Cui et al., 27 Aug 2025).

4. Numerical Methods and Relaxation Approaches

Algorithmic treatment of PTECs is inherently challenging due to nonconvexity, hierarchical structure, and nonsmoothness:

• Decomposition and PH Methods: In CC-EPEC for electricity markets, leader-based progressive hedging decomposes the game across states, imposing consensus constraints on shared equilibrium variables and solving subproblems in parallel (Kim et al., 2020).
• SOS/SDP Relaxations: For polynomial data, a double Lasserre hierarchy is employed: (1) an inner SOS relaxation approximates the value function representing the equilibrium constraint; (2) an outer moment-SOS relaxation solves the global problem. Under mild conditions, finite convergence and global minimizers are recoverable (Jiao et al., 2019).
• Penalty and First-Order Rewrites: First-order approximation of bilevel equilibrium pretraining schemes (e.g., "first-order MAML") efficiently enforce equilibrium constraints by leveraging chain-rule and ignoring second derivatives when the landscape is flat (Cui et al., 27 Aug 2025).
• Variational Error Bounds and Penalization: Locally exact penalization using error-bound coefficients links distance to the equilibrium set with the residual of the implicit constraint, enabling conversion of the implicit set constraint into a single inequality (Uderzo, 2022).

5. Fundamental Theoretical Results and Examples

• Existence: Under weak conditions (quasi-concavity, compactness, order-completeness), fixed-point and order-theoretic results guarantee the existence of solutions to equilibrium-constrained problems—including in infinite-dimensional partially ordered Banach spaces (Li, 2017).
• Optimality: For both classical and nonconvex problems, necessary conditions include linearized M-stationarity conditions that are strictly stronger than those derived from MPCC reformulations, and hold under weaker, checkable metric subregularity CQs (Gfrerer et al., 2016, Gfrerer et al., 2019).
• Policy Nonmonotonicity: In economic applications, equilibrium constraints due to financial frictions can induce non-monotonic relationships between aggregate output and parameters such as credit limits or productivity due to reallocation and general equilibrium effects (Pham, 22 Jan 2025).
• Game-Theoretic Equilibrium Refinement: Player-compatible equilibrium (PCE) imposes equilibrium constraints on the cross-player relative frequency of trembles, yielding sharper equilibrium selection justified by rational learning and experimentation (Fudenberg et al., 2017).

6. Broader Modeling and Domain Extensions

Equilibrium constraints provide a unifying formalism for:

• Dynamic and Optimal Control with Complementarity: Necessary optimality conditions for optimal control with equilibrium (complementarity) constraints require a two-multiplier representation, reflecting intrinsic nonsmoothness, with strong stationarity available only under strict constraint qualifications (Guo et al., 2016).
• Market Design with Linear Constraints and Pecuniary Externalities: Constrained pseudo-market equilibrium extends the fake-money market mechanism by imposing arbitrary linear (possibly combinatorial) constraints, with pecuniary externalities internalized prices (Echenique et al., 2019).
• Continuous Games and Safety/Chance Constraints: Equilibrium constraints formulated as chance constraints enable safety in multi-agent continuous-action games, solved via smoothed mixed-complementarity formulations (Krusniak et al., 2024).

7. Open Questions and Future Directions

Several open research directions remain:

• Relaxation of Regularity Conditions: Further weakening of metric subregularity, toward purely directionally defined or third-order CQs, and their exploitation in optimality-testing and solver design (Gfrerer et al., 2016, Gfrerer et al., 2019).
• Algorithmic Scalability: Development of decomposition schemes and global relaxations that scale to high-dimensional, mixed-integer, and stochastic equilibria with equilibrium constraints.
• Applications to Learning and AI: Extension of first-order hierarchy-free approximations, such as in bilevel and multi-agent learning systems, to large-scale settings (Li et al., 2023, Cui et al., 27 Aug 2025).
• Refined Equilibrium Concepts: Further embedding of learning dynamics, experimentation, and rational behavior into equilibrium-constraint refinements in Nash and generalized Nash equilibrium modeling (Fudenberg et al., 2017, Li, 2017).

Equilibrium-constrained optimization thus forms a foundational pillar for advanced modeling and computational frameworks across optimization, economics, game theory, and machine learning. The mathematical and algorithmic approaches continue to evolve, leveraging nonsmooth analysis, variational geometry, and hierarchical decomposition to meet the theoretical and practical demands of modern systems.