Mordukhovich Stationarity in Constrained Optimization

Updated 26 March 2026

Mordukhovich stationarity is the limiting stationarity concept in constrained optimization that leverages the Mordukhovich normal cone and subdifferential to define first-order optimality conditions.
It is instrumental in analyzing nonsmooth, nonconvex, and structured problems, offering robust necessary conditions even when classical regularity and constraint qualifications fail.
Its applications span disjunctive, complementarity, and PDE-constrained programs, bridging theoretical optimality insights with practical algorithmic frameworks.

Mordukhovich Stationarity

Mordukhovich stationarity (commonly referred to as M-stationarity or limiting stationarity) is a central concept in the first-order optimality theory for constrained and variational problems marked by nonsmoothness, nonconvexity, or combinatorial structure. This stationarity notion is formulated in terms of the limiting, or Mordukhovich, normal cone and subdifferential, which provide robust closure properties required for variational analysis but may lack a direct variational or geometric interpretation in pathological or degenerate circumstances. Mordukhovich stationarity is indispensable in the study of disjunctive, complementarity, vanishing, and related structured constraints, playing a pivotal role in both necessary optimality conditions and algorithmic frameworks.

1. Tangent, Fréchet, and Mordukhovich Normal Cones

Let $C \subset \mathbb{R}^n$ be a closed set and $x \in C$ . Denote:

Tangent (Bouligand or contingent) cone:

$T_C(x) = \left\{ w\,:\,\exists\, x_k \in C,\, \tau_k \downarrow 0,\ \frac{x_k - x}{\tau_k} \rightarrow w \right\}.$

Regular (Fréchet) normal cone:

$\hat N_C(x) = \left\{ v\,:\, \langle v, w \rangle \le 0,\ \forall w \in T_C(x) \right\},$

or equivalently,

$\langle v, y-x \rangle \le o(\|y-x\|)\ \forall y \in C.$

Limiting (Mordukhovich) normal cone:

$N_C(x) = \left\{ v \in \mathbb{R}^n\,:\, \exists\, x_k \to x,\ v_k \in \hat N_C(x_k),\ v_k \to v \right\} = \limsup_{x_k \to x} \hat N_C(x_k).$

The inclusion $\hat N_C(x) \subset N_C(x)$ always holds, and $N_C(x)$ is the appropriate construct for variational-analytic closure arguments and generalized differentiation (Pauwels, 2024).

2. Defining Mordukhovich Stationarity

Given a (typically smooth) function $f:\mathbb{R}^n \to \mathbb{R}$ and a closed constraint set $C \subset \mathbb{R}^n$ , introduce the composite objective $F(x) = f(x) + \delta_C(x)$ , where $\delta_C(x)=0$ if $x \in C$ , $+\infty$ otherwise. The Mordukhovich (limiting) subdifferential and stationarity conditions are defined as:

Limiting subdifferential:

$\partial F(x) = \left\{ v \in \mathbb{R}^n\,:\, \exists\, x_k \to x,\, v_k \in \hat \partial F(x_k),\, v_k \to v \right\}.$

When $f$ is $C^1$ , $\partial F(x) = \nabla f(x) + N_C(x)$ (Pauwels, 2024).

Mordukhovich stationarity (criticality):

$0 \in \partial F(x) \quad\Longleftrightarrow\quad -\nabla f(x) \in N_C(x).$

This is the criticality or M-stationarity condition. It includes the Fréchet-stationary points as a subset; i.e., any Fréchet-stationary $x$ is M-stationary, but not necessarily vice versa (Pauwels, 2024).

The variational meaning of $-\nabla f(x)\in N_C(x)$ is that $x$ cannot be immediately ruled out as a local minimizer by considering first-order variations; however, genuine descent directions might still exist if $x$ is not Fréchet-stationary, particularly in nonregular or highly singular circumstances (e.g., sets with combinatorial structure) (Pauwels, 2024, Shikhman, 2024).

3. Role in Optimality Conditions and Constraint Qualifications

M-stationarity is frequently fundamental as a necessary optimality condition in nonsmooth, nonconvex, or disjunctive optimization, especially where classical regular constraint qualifications (LICQ, MFCQ) do not hold. For the constraint system $Q = \{x \mid F(x) \in D\}$ , with $F$ smooth and $D$ a finite union of convex polyhedra, M-stationarity is given by the existence of $\lambda \in N_D(F(x))$ such that

$0 = \nabla f(x) + \nabla F(x)^T \lambda.$

The concept extends naturally to more general settings, such as set-valued constraints ( $0 \in \Phi(x)$ with limiting coderivative $D^*\Phi(x,0)$ ):

$0 \in \partial f(x) + D^*\Phi(x,0)(\lambda).$

M-stationarity is a necessary condition at local minimizers under weak constraint qualifications such as the generalized Guignard CQ (GGCQ), metric (sub)regularity, or outer-semicontinuity of related coderivative mappings (Mehlitz, 2020, Gfrerer, 2016, Shikhman, 2024). These relaxed CQs are often significantly weaker than classical conditions and can be verified directly in many structured nonsmooth settings (complementarity constraints, vanishing/orthogonality/disjunctive systems).

M-stationarity serves as a bridge between the Bouligand (B-) and strong (S-) stationarity notions:

$\text{S-stationarity} \implies \text{M-stationarity} \implies \text{B-stationarity}.$

Under strong regularity, or for convex sets, all these conditions coincide with classical KKT stationarity (Benko et al., 2019).

4. Comparison with Fréchet and Other Stationarity Concepts

A hierarchical chain of stationarity notions emerges in variational analysis:

$\text{Fréchet (}\hat N\text{-stationarity)} \implies \text{Mordukhovich (}N\text{-stationarity)} \implies \text{Clarke (}\overline{N}\text{-stationarity)},$

with

Fréchet stationarity: $-\nabla f(x)\in\hat N_C(x)$
Mordukhovich stationarity: $-\nabla f(x)\in N_C(x)$
Clarke stationarity: $-\nabla f(x)\in\overline{N}_C(x)$

Fréchet stationarity guarantees there are no first-order descent directions and has a strong direct variational interpretation. M-stationarity is more robust under limits and encompasses all local minimizers but can, in degenerate/pathological cases, admit points that are not actual critical points in a geometric sense (so-called "singular saddles") (Pauwels, 2024, Shikhman, 2024).

In many practically relevant problems—such as when $C$ and $f$ are semi-algebraic—generic or regular instances guarantee the equivalence of M-stationarity and Fréchet stationarity at critical points (Pauwels, 2024). Furthermore, projected-gradient algorithms with sufficiently small step-sizes converge only to points that are both M- and Fréchet-stationary, ensuring the exclusion of spurious limiting-only critical points (Pauwels, 2024).

5. M-Stationarity in Structured Nonlinear and Disjunctive Programs

The application of M-stationarity is particularly prominent in structured nonsmooth programs:

Mathematical Programs with Complementarity Constraints (MPCC): In the setting

$\min f(x)\quad\text{s.t.}~g(x)\le0,\ h(x)=0,\ 0\le G(x)\perp H(x)\ge0,$

M-stationarity requires multipliers $\lambda,\eta,\mu,\nu$ satisfying KKT-type conditions, with Mordukhovich-product sign constraints on the biactive index set:

$(\mu_k>0\land\nu_k>0)\ \lor\ \mu_k\nu_k=0~\text{for}~G_k(x)=H_k(x)=0.$

Strong and Clarke stationarity notions correspond to strictly stronger or weaker sign constraints, respectively (Harder, 2020, Harder, 2021).

Disjunctive and Semi-algebraic Constraints: For programs with $F(x)\in D$ where $D$ is a union of convex polyhedral sets, the Mordukhovich normal cone $N_D(F(x))$ is explicitly computable; first-order necessary optimality conditions in terms of M-stationarity require only weak regularity assumptions (Gfrerer, 2016, Benko et al., 2016, Käming et al., 28 Mar 2025).
Sequence/Approximate Stationarity and Constraint Qualifications: In the absence of any regularity condition, every local minimizer is approximately (AM-) stationary; the gap to exact M-stationarity can be bridged by very weak sequential or metric regularity-type conditions (“AM-regularity” or subMFC), which are often easily verifiable (Mehlitz, 2020, Käming et al., 28 Mar 2025).

M-stationarity is a member of a family of stationarity and optimality concepts spanning from strong to weak forms. Several refinements and related notions exist:

Extended M-stationarity (directional): Requires M-stationarity in each critical direction and is equivalent to B-stationarity under generalized Guignard conditions (Gfrerer, 2016).
Strong M-stationarity: Corresponds to the existence of a single “active set” of multipliers satisfying additional nonnegativity constraints, generalizing S-stationarity (Gfrerer, 2016).
Piecewise M-stationarity: Demands KKT-type multipliers exist for each complementarity branch (partition of the biactive index set), coinciding with B-stationarity under tailored ACQ (Wang et al., 24 Mar 2026).
QM/Q-stationarity: Strengthens M-stationarity by imposing it for all convex combinations of linearized tangent cones, readily verifiable by solving quadratic programs (Benko et al., 2016, Benko et al., 2019).

These concepts are instrumental both for theoretical analysis and for the design of robust first-order algorithms: for instance, NCP-based Newton methods solve the M-stationarity system directly, with superlinear local convergence under mild second-order and constraint qualifications (Harder et al., 2020).

7. Applications to Sensitivity, Stability, and Higher-Order Conditions

The limiting normal cone and coderivative structure underlying M-stationarity are pivotal in parametric and stability analysis:

Sensitivity and Robinson Stability: Sharp coderivative calculations for stationary-point mappings are available via explicit second-order formulas, yielding necessary and sufficient conditions for Lipschitzian and Robinson stability under perturbations in both smooth and indefinite quadratic programming (Huyen et al., 2018, Huyen et al., 2018).
Mixed-Order and Directional Asymptotic Stationarity: M-stationarity can be extended to capture higher-order information (order- $\gamma$ stationarity) or to handle irregular/degenerate situations via pseudo-coderivatives. Directional asymptotic regularity conditions serve as weak CQs that guarantee classical M-stationarity of local minimizers, bridging first-order and second-order optimality frameworks (Benko et al., 2022, Benko et al., 2024).
Infinite-Dimensional and PDE-Constrained Optimization: M-stationarity has been established as the correct necessary condition in Lebesgue and Sobolev space MPCCs, ensuring the existence of multipliers satisfying the canonical M-stationarity sign patterns, even in the absence of finite-dimensional convexity (Harder et al., 2021).

References

“A note on stationarity in constrained optimization” (Pauwels, 2024)
“Stationarity in nonsmooth optimization between geometrical motivation and topological relevance” (Shikhman, 2024)
“Optimality conditions for disjunctive programs...” (Gfrerer, 2016)
“New stationarity conditions between strong and M-stationarity for mathematical programs with complementarity constraints” (Harder, 2021)
“M-stationarity for a class of MPCCs in Lebesgue spaces” (Harder et al., 2021)
“Piecewise M-Stationarity and Related Algorithms for Mathematical Programs with Complementarity Constraints” (Wang et al., 24 Mar 2026)
“Approximate stationarity in disjunctive optimization: concepts, qualification conditions, and application to MPCCs” (Käming et al., 28 Mar 2025)
“On the directional asymptotic approach in optimization theory” (Benko et al., 2024)
“Linearized M-stationarity conditions for general optimization problems” (Gfrerer, 2018)
“Reformulation of the M-stationarity conditions as a system of discontinuous equations and its solution by a semismooth Newton method” (Harder et al., 2020)
“Sensitivity Analysis of a Stationary Point Set Map under Total Perturbations. Part 1: Lipschitzian Stability” (Huyen et al., 2018)
“Sensitivity Analysis of a Stationary Point Set Map under Total Perturbations. Part 2: Robinson Stability” (Huyen et al., 2018)
“On estimating the regular normal cone to constraint systems and stationarity conditions” (Benko et al., 2019)
“Asymptotic stationarity and regularity for nonsmooth optimization problems” (Mehlitz, 2020)
“New verifiable stationarity concepts for a class of mathematical programs with disjunctive constraints” (Benko et al., 2016)
“On the directional asymptotic approach in optimization theory Part A: approximate, M-, and mixed-order stationarity” (Benko et al., 2022)