Normal Cone¶

At a point of a convex set, collect every vector making a nonpositive inner product with every feasible displacement from that point.

Version: v1 · 2026-10-03 · History
Domain-specific #: 13471
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Convex Analysis → Mathematics
Aliases: Convex normal cone

Core Idea¶

For a convex set \(C\subseteq\mathbb R^n\) and \(x\in C\), the normal cone is \(N_C(x)=\{v:\langle v,y-x\rangle\leq0\text{ for every }y\in C\}\). It collects all vectors pointing outward relative to every feasible displacement from the chosen point. Because the inequality survives addition and nonnegative scaling, the collection is a convex cone. This is a point-indexed geometric object, not merely a single perpendicular arrow.[^{ref-ce6cb9313c63][}ref-ad86641c4e3b]

At an ambient interior point the cone is \(\{0\}\). At a halfspace boundary it is an outward ray, and at a polyhedral corner it is the conic hull of active outward constraint normals. A lower-dimensional affine set differs: its normal cone is the full orthogonal subspace at every point, even at points in its relative interior. The common “interior implies zero” shortcut requires ambient interior.[^{ref-ce6cb9313c63][}ref-140b66087e95]

Scope of Application¶

For a differentiable objective \(f\) minimized over convex \(C\), \(-\nabla f(x)\in N_C(x)\) is a necessary first-order condition: no feasible displacement is an improving descent direction. It is sufficient for a global minimum when \(f\) is also convex. The cone exists without an objective and therefore must not be conflated with KKT conditions or an optimum itself. Generalized nonconvex and nonsmooth normals require their own definitions or qualifications.[^{ref-ce6cb9313c63][}ref-cf7f4cb0c3be]

The same construction appears in unlike convex geometries. For \(C=\{(u,v):u,v\geq0\}\), at the origin the feasible displacements have nonnegative coordinates, so \(N_C(0,0)\) is the negative orthant; only active constraints generate it. For the affine line \(C=\{(u,v):v=0\}\), \(N_C(x)\) is the entire vertical line at every feasible \(x\). Projecting \(z=(2,3)\) to \(x=(2,0)\) gives residual \((0,3)\) in that cone. Both examples map convex carrier, feasible point, all displacements, universal nonpositive pairing and the resulting set of normals.[^{ref-ce6cb9313c63][}ref-140b66087e95]

Clarity¶

The definition distinguishes outward normals from feasible tangent directions: the two are related by negative polarity, not identity. It also separates an active inequality's one-sided ray from an equality's two-sided orthogonal normal space. A topological normal space, algebraic cone or smooth-surface unit normal is not an alias just because the word normal or cone appears.[^{ref-ce6cb9313c63][}ref-140b66087e95]

Manages Complexity¶

The all-\(y\in C\) test can be compressed for a polyhedron \(C=\{x:a_j^\top x\leq b_j\}\): \(N_C(x)\) consists of nonnegative combinations of the \(a_j\) whose constraints are active at \(x\). This makes geometric support checkable from a finite active set. The shortcut is exact for this stated halfspace representation, not a universal replacement for the definition.[^ref-140b66087e95]

Abstract Reasoning¶

State the ambient space, convex set and feasible base point. Test whether a candidate \(v\) pairs nonpositively with every \(y-x\). For polyhedra use active outward normals; for affine equalities allow both signs of orthogonal directions. In optimization, compute the cone first, then test \(-\nabla f(x)\) for membership and separately check whether convexity of \(f\) turns the necessary first-order condition into a sufficient one.[^{ref-ce6cb9313c63][}ref-140b66087e95][^ref-cf7f4cb0c3be]

Knowledge Transfer¶

The literal set–point–displacement–polar pattern transfers between polyhedral constraints and affine projection despite their different cones. Live Convexity is a proposed strict presupposition for this bounded convex-analysis entry; Conical combination is a polyhedral computational neighbor, not its universal genus. The definition's inner-product and pointwise convex geometry remain constitutive, so sharing a portable polarity skeleton does not by itself make the named normal cone a new prime.[^{ref-ce6cb9313c63][}ref-140b66087e95]

[^ref-ce6cb9313c63]: Gabriele Farina, MIT 6.7220 Nonlinear Optimization, Lecture 2 (2025), Definition L2.3, interior/affine examples and projection, pp.3–7. [^ref-140b66087e95]: Gabriele Farina, MIT 6.7220 Nonlinear Optimization, Lecture 3 (2025), halfspace and active-constraint examples, Theorem L3.1, pp.1–3. [^ref-cf7f4cb0c3be]: Gabriele Farina, MIT 6.7220 Nonlinear Optimization, Lecture 4 (2025), convex objective and first-order sufficiency, pp.1–3. [^ref-ad86641c4e3b]: Stephen Boyd and Lieven Vandenberghe, Convex Optimization, Exercise 2.38©, printed p.67.

Relationships to Other Abstractions¶

Current abstraction Normal Cone Domain-specific

Parents (1) — more general patterns this builds on

Normal Cone presupposes Convexity Prime

A convex-analysis normal cone presupposes a convex feasible carrier.

Hierarchy path (1) — routes to 1 parentless root

Normal Cone → Convexity → Optimization

Neighborhood in Abstraction Space¶

Normal Cone sits in a sparse region of the domain-specific corpus (66^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08