Optimization Workshop Notes for Mathematical Programming with Equilibrium Constraints Algorithms: Penalty Interior-Point, Implicit-Programming, and Piecewise SQP

Abstract: In this workshop, we discuss several algorithms for mathematical programs with equilibrium constraints (MPECs). The unifying theme is that MPECs are optimization problems whose feasible set contains a lower-level equilibrium system, often written through complementarity or variational-inequality conditions. This destroys the smooth manifold or convex structure that standard nonlinear programming methods rely on. We focus on four algorithmic viewpoints: (i) the classical penalty interior-point algorithm (PIPA); (ii) a monotone-linear complementarity problem (LCP) variant of PIPA that explicitly controls complementarity decay; (iii) an implicit-programming descent method for variational inequality (VI)-constrained MPECs; (iv) piecewise SQP (PSQP), which applies SQP on locally selected smooth pieces. For each method we explain the model, the search direction subproblem, the globalization mechanism, and the meaning of the convergence result. Particular emphasis is placed on what the assumptions are really doing and on the distinction between an attractive algorithmic idea and a fully valid convergence theorem.

• The paper presents a systematic analysis of four canonical approaches for solving MPECs, emphasizing stationarity conditions and convergence properties.
• It introduces penalty interior-point and monotone-LCP variants that balance feasibility with controlled complementarity reduction.
• The work highlights implicit programming and piecewise SQP methods, discussing their computational trade-offs and local convergence advantages.

Algorithms for MPECs: Penalty Interior-Point, Implicit Programming, and Piecewise SQP

Introduction and Problem Structure

Mathematical Programs with Equilibrium Constraints (MPECs) present significant computational and theoretical challenges, primarily due to the intrinsic nonconvex and non-manifold geometry induced by complementarity and variational-inequality lower-level constraints. The standard MPEC can be formulated as an optimization problem in variables $(x, y, w, z)$ with upper-level objective $f(x, y, w, z)$, affine or nonlinear equalities, and complementarity constraints $y \geq 0$, $w \geq 0$, $y^\top w = 0$. The geometry of the feasible region is often stratified, with local degeneracy at intersection points between coordinate axes, thereby precluding application of standard nonlinear programming methods without adaptation.

The notes systematize four canonical algorithmic paradigms for MPECs:

1. Penalty interior-point methods (PIPA) with centrality-driven interior iterates.
2. A monotone-LCP variant of PIPA with explicit control on complementarity decay.
3. Implicit programming descent, leveraging local solution maps for variational inequalities.
4. Piecewise Sequential Quadratic Programming (PSQP), exploiting local smooth piece structure.

Each method is explicated in terms of subproblem formulation, globalization mechanism, and stationarity characterization.

First-Order Stationarity for Complementarity-Constrained MPECs

Strict complementarity is essential for nondegeneracy and guarantees that the index set where both $y_i$ and $w_i$ vanish is empty. First-order stationarity under strict complementarity yields a KKT-like condition with an additional multiplier associated with the complementarity product. For a feasible and strictly complementary point $u^*$ and nonsingular Jacobian block $M^*$, stationarity is equivalent to the existence of multipliers satisfying extended KKT equations, incorporating the effects of complementarity via a multiplier $\xi$. This expresses the coupling between upper and lower levels inherent in MPEC optimality.

Penalty Interior-Point Algorithm (PIPA)

PIPA employs interior-point regularization, maintaining positivity of $f(x, y, w, z)$0, $f(x, y, w, z)$1 and requiring $f(x, y, w, z)$2 with $f(x, y, w, z)$3 being the mean complementarity product. This central path regularization is analogous to those in standard complementarity and nonlinear programs.

The core direction-finding problem consists of a QP:

• Linearizing the upper-level objective,
• Newton step for the lower equilibrium,
• Primal-dual centering for complementarity, all subject to a trust region restricting upper-level motion as a function of infeasibility.

Classical convergence claims for PIPA rested on bounded level sets and nonsingularity, but Leyffer demonstrated that PIPA may converge to nonstationary points due to excess restriction on the upper-level step sizes as infeasibility vanishes. Thus, while PIPA remains a powerful framework, convergence guarantees are conditional rather than universal.

Monotone-LCP Variant of PIPA

For monotone LCP lower levels, the method sharpens the coupling between linear feasibility and complementarity. In the direction-finding QP, both residual and complementarity progress are synchronized explicitly via the "limited complementarity decrease" condition, which prevents complementarity from reducing faster than feasibility. The line search reduces to a quadratic inequality, permitting explicit arc search.

The global theory for this variant realizes the classical penalty method dichotomy: either the penalty sequence diverges yielding stationarity along an update subsequence, or, if the penalty is bounded and step lengths are uniformly positive, limit points are stationary. The critical innovation is prioritizing feasibility over complementarity reduction, a balance that corrects the PIPA pathologies.

Implicit-Programming Descent for VI-Constrained MPECs

In cases where the lower-level solution can be expressed as a (locally) isolated and directionally differentiable function $f(x, y, w, z)$4, implicit programming reduces the MPEC to an upper-level problem in $f(x, y, w, z)$5 only. The direction-finding subproblem is a regularized model on the reduced objective $f(x, y, w, z)$6, using generalized derivatives.

Armijo-type line search is performed along the equilibrium path $f(x, y, w, z)$7, inherently preserving lower-level feasibility. Convergence to stationary points is guaranteed under polyhedrality, closedness, and strong regularity of the implicit map.

Major limitations are computational: evaluation of $f(x, y, w, z)$8 and its directional derivative may require repeated lower-level solves, often including nonsmooth or VI subproblems, which in practice constrains scalability unless the lower-level derivative is tractable.

Piecewise Sequential Quadratic Programming (PSQP)

PSQP addresses the non-manifold feasible set by constructing, at each iteration, a local smooth piece (i.e., a branch corresponding to a selection of the complementarity regime for degenerate indices), and applying standard SQP to that piece. The subproblem is a standard SQP QP, but the admissible piece is determined based on the complementary index structure.

Convergence theory for PSQP is local. Uniform local contraction applies because in a small neighborhood of a solution, only finitely many pieces are relevant; contraction estimates hold uniformly across these cases. This yields Q-superlinear and, under further regularity, Q-quadratic convergence rates, akin to the standard theory for smooth nonlinear programs.

PSQP's central theoretical implication is that one can directly tackle the piecewise smooth geometry of the MPEC, bypassing the need to regularize complementarity constraints, with standard Newton/SQP methods, as long as the correct local piece is identified and strict globalizability is not required.

Comparative Analysis

The synthesis of these canonical methods is as follows:

Method	Local Model Principle	Strengths	Main Weaknesses
PIPA	Linearize lower and perturbed complementarity, keep $f(x, y, w, z)$9	Structured, interpretable	Global theory not robust
LCP-PIPA	Synchronize affinity and complementarity decay	Corrects feasibility bias	Limited to monotone LCP
Implicit Programming	Locally eliminate $y \geq 0$0 via $y \geq 0$1 mapping	Faithful to problem structure	High computational cost
PSQP	Apply SQP on selected local smooth branch	Fast local convergence	Purely local, not global

Implications and Outlook

The results formalize structural limitations of interior-point approaches (most notably the nonnecessity of stationarity due to trust-region artifacts in PIPA) and underscore the importance of balanced progress between upper-level and complementarity violations. PSQP provides a mathematically satisfactory approach for local convergence once strict complementarity can be enforced, highlighting that regularization is not strictly required for superlinear local convergence.

For future algorithm design, these insights suggest that combining centrality, coordinated step-size control, and exploitation of piecewise geometry may yield scalable global-local hybrid methods that balance robustness and fast asymptotic convergence. Advances in efficient implicit differentiation and global piecewise state identification remain critical for increased practical applicability, especially for high-dimensional or black-box lower-level maps.

Conclusion

The discussed algorithmic archetypes each address the principal geometric and analytical challenges of MPECs using distinct strategies: interior-point regularization, synchronized penalty methodologies, reduced-space implicit programming, and active-piecewise smooth optimization. The lecture note elucidates both the analytical framework and the algorithmic subtlety required to make progress in MPECs, delineating both their respective domain of efficacy and their limitations. The continued evolution of globally convergent, computationally efficient, and locally fast MPEC algorithms will necessitate integrating the strengths of these approaches, together with further advances in nonsmooth analysis and structure-exploiting numerical methods.

---

Reference: "Optimization Workshop Notes for Mathematical Programming with Equilibrium Constraints Algorithms: Penalty Interior-Point, Implicit-Programming, and Piecewise SQP" (2604.15690).