Limiting Normal Cone in Variational Analysis

• Limiting Normal Cone is a generalized normal vector concept for arbitrary closed sets, enabling precise formulation of stationarity in nonconvex settings.
• It is constructed as the Painlevé–Kuratowski outer limit of regular normal cones, capturing intricate boundary and interior behavior.
• Its calculus rules, including chain and intersection rules, are fundamental in deriving optimality conditions for variational inequalities and nonsmooth problems.

A limiting normal cone, also called the Mordukhovich normal cone, is a foundational object in modern variational analysis and nonsmooth optimization. It generalizes the classical concept of a normal vector at smooth boundary points to arbitrary closed sets, supporting precise formulations of stationarity, stability, and sensitivity in settings far beyond convexity. The Mordukhovich limiting normal cone is central to deriving necessary optimality conditions in nonconvex, nonpolyhedral, and variational-inequality-constrained optimization, and it underpins key developments in coderivative calculus, subdifferential theory, and nonsmooth analysis.

1. Preliminaries: Tangent and Normal Cones

Given a closed set $C \subseteq \mathbb{R}^n$ and a point $\bar x \in C$, several notions of tangent and normal cones arise:

• Tangent (contingent) cone:

$$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$

• Fréchet (regular) normal cone:

$$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$

This is always a closed convex cone.

• Limiting (Mordukhovich) normal cone:

$$N_C(\bar x) := \limsup_{x_k \to \bar x,\, x_k \in C} \widehat{N}_C(x_k) = \left\{ v = \lim_{k \to \infty} v_k \;\middle|\; v_k \in \widehat{N}_C(x_k),\, x_k \to \bar x \right\}.$$

$N_C(\bar x)$ is closed but need not be convex. If $C$ is convex, all three cones coincide and equal the classical convex-analytic normal cone $[T_C(\bar x)]^\circ$ (Benko et al., 2017, Benko, 2019).

2. Definition and Construction of the Limiting Normal Cone

The limiting normal cone captures normal vectors not only at $\bar x$ but also those "arising as limits" from sequences of regular normal vectors at nearby points. This nonlocality allows $N_C(\bar x)$ to encode geometry not visible to the Fréchet cone. In practical definitions, it is usually formulated as the Painlevé–Kuratowski outer limit of the regular normal cones: $\bar x \in C$0 Analogous definitions hold in infinite dimensions, with appropriate weak* closure (Hantoute et al., 2017). For epigraphs of convex functions or sublevel sets, normal cone descriptions in terms of (sub)differentials and multipliers are standard (Hantoute et al., 2017).

The limiting normal cone reduces to the convex-analytic normal cone in the convex setting, but, crucially, may be nonconvex or even set-valued in nonconvex or composite structures (Liu et al., 2016).

3. Calculus Rules and Directional Refinements

A fundamental advantage of the limiting normal cone is its amenability to generalized calculus rules—chain rules, sum rules, and intersection rules—that extend classical variational analysis to nonconvex regimes.

• Pre-image (chain) rule: For $\bar x \in C$1 with $\bar x \in C$2 smooth and $\bar x \in C$3 closed,

$\bar x \in C$4

with equality under metric subregularity or related constraint qualifications (Benko, 2019, Benko et al., 2017). Under weaker conditions, only inclusion holds.

• Intersection rule: For $\bar x \in C$5 and suitable regularity,

$\bar x \in C$6

with equality in the convex case. For nonconvex sets, additional qualification conditions such as transversality or metric regularity are required (Benko et al., 2017).

• Directionally limiting normal cone: To further weaken qualification requirements or to analyze "one-sided" sensitivity, the directional limiting normal cone $\bar x \in C$7 is defined for a fixed $\bar x \in C$8 as the set of limits of normals at points approaching $\bar x \in C$9 along $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$0. Calculus rules for directional cones enable finer sensitivity analysis and have weaker (directional) metric subregularity requirements (Benko et al., 2017).

4. Exact Formulas in Key Applications

The power of the limiting normal cone is revealed by explicit formulas for critical sets in optimization and variational inequalities:

• Second-Order Cone Complementarity Set: For

$$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$1

where $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$2 is the Lorentz cone, Ye and Zhou provide an exhaustive casewise formula for $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$3 distinguishing interior, boundary, and zero cases. Each geometric region has an explicit representation, e.g., for $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$4, $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$5,

$$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$6

and for $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$7, $$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$8,

$$T_C(\bar x) := \left\{ u \in \mathbb{R}^n \;\middle|\; \exists t_k \downarrow 0,\, u_k \to u,\, \bar x + t_k u_k \in C \right\}.$$9

(Ye et al., 2016).

• Graph of the Subdifferential of the Nuclear Norm: For the set $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$0, Liu and Pan construct an explicit decomposition involving singular value structure, block partitions, and divided-difference matrices, capturing the difference between regular and limiting cones in the presence of multiplicity at spectral norm thresholds (Liu et al., 2016).
• Intersection of Convex Sublevel Sets: The limiting normal cone to a polyhedral or intersection of convex sublevel sets admits a multiplier–subgradient representation, only involving active indices:

$$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$1

with $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$2 (Hantoute et al., 2017).

• Normal Cone Mapping for Inequality Constraints: For $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$3, coderivative inclusions and explicit formulas involving multipliers, critical cones, and tangent sets detail the construction of $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$4 and related coderivatives (Gfrerer et al., 2016).

5. Restricted, Directional, and Parameterized Variants

• Restricted Limiting Normal Cone: Given sets $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$5 and $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$6, the restricted normal cone $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$7 modifies the construction to only consider sequences from $$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$8. This generalization captures "one-sided" geometry, essential to sharp regularity and convergence results for algorithms such as alternating projections (Bauschke et al., 2012). The following containment always holds:

$$\widehat{N}_C(\bar x) := \left\{ v \in \mathbb{R}^n \;\middle|\; \limsup_{x \to \bar x,\, x \in C} \frac{\langle v, x - \bar x\rangle}{\|x - \bar x\|} \leq 0 \right\}.$$9

and for convex $$N_C(\bar x) := \limsup_{x_k \to \bar x,\, x_k \in C} \widehat{N}_C(x_k) = \left\{ v = \lim_{k \to \infty} v_k \;\middle|\; v_k \in \widehat{N}_C(x_k),\, x_k \to \bar x \right\}.$$0, all constructions coincide with the classical cone.

• Directional Limiting Normal Cone: Specifies limiting normals along a prescribed direction, improving calculus rules for nonconvex feasible sets and enabling directional sensitivity analysis (see properties and sum/intersection/chain rules in (Benko et al., 2017)).
• Parameter-Dependent Systems: Limiting normal cones and their coderivatives provide the fundamental tools for verifying the Aubin property (pseudo-Lipschitz continuity) and tilt stability of solution maps to parameterized variational inequalities, generalized equations, and equilibrium problems (Benko, 2019, Gfrerer et al., 2016).

6. Role in Optimization, Sensitivity, and Variational Analysis

The limiting normal cone is essential in the following domains:

• First-Order and Second-Order Optimality: Necessary conditions (Mordukhovich/M-stationarity, S-stationarity) in mathematical programs with constraints articulated through normal cones—especially for sets with complex structure such as second-order cone complementarity or nuclear norm constraints—are only derivable via $$N_C(\bar x) := \limsup_{x_k \to \bar x,\, x_k \in C} \widehat{N}_C(x_k) = \left\{ v = \lim_{k \to \infty} v_k \;\middle|\; v_k \in \widehat{N}_C(x_k),\, x_k \to \bar x \right\}.$$1 (Ye et al., 2016, Liu et al., 2016).
• Sensitivity and Stability: Properties such as the Aubin property and tilt stability for solution maps are characterized using coderivatives of set-valued maps built from limiting normal cones. Exact formulas greatly aid verifiability (Benko, 2019, Gfrerer et al., 2016).
• Regularity and Constraint Qualification: Calculation of limiting normal cones is closely intertwined with notions of metric subregularity, constraint qualification (such as 2-regularity, 2-LICQ, or superregularity), and directional metric subregularity, which modulate the sharpness of calculus rules and characterization of stability (Gfrerer et al., 2016, Bauschke et al., 2012, Benko et al., 2017).
• Algorithmic Implications: Restricted and directional cones have been employed to formulate new, sharper convergence results for projection methods and to enable semismooth Newton methods for generalized equations (Bauschke et al., 2012, Benko, 2019).

7. Illustrative Examples and Special Cases

Several explicit examples highlight the computational and conceptual role of the limiting normal cone:

• Convex polyhedral sets, where the limiting and regular cones coincide, and explicit formulas are available.
• Epigraphs of nonsmooth functions, where the directional normal cone distinguishes approaching from interior versus tangential directions.
• Second-order cone complementarity sets, where all possible limiting cases, including degenerate and regular situations, are partitioned and classified (Ye et al., 2016).
• Pairs of lines or spheres, where restricted normal cones quantify angles of intersection or break classical symmetries (Bauschke et al., 2012).
• Systems with nonunique multipliers or degenerate tangent directions, where only the limiting construction produces a nontrivial normal (Liu et al., 2016, Gfrerer et al., 2016).

---

The limiting normal cone, in its various incarnations—classical, directional, restricted, parameter-dependent—provides a profound unifying framework for nonsmooth analysis, supporting both deep theoretical understanding and explicit computation in optimization, variational inequalities, and modern approaches to stability and sensitivity in mathematical programming.