🔶

Convex Sets & Functions

The Foundation of Modern Optimization: From Geometry to Metric Spaces

📐

1. What is a Convex Set?

🔷 Geometric Definition

A set is convex if every line segment between two points in the set stays completely inside the set.

Mathematically:

\[ \theta x_1 + (1 - \theta) x_2 \in C, \quad \forall \theta \in [0, 1] \]

🔷 Algebraic Definition

Generally, a set is convex if all convex combinations lie in the set:

\[ \sum \theta_i x_i \in C, \quad \text{where } \sum \theta_i = 1, \theta_i \geq 0 \]

Visual Comparison

Convex Set Line stays inside Non-Convex Set Line goes outside

🤔 Quiz: Equivalence

Prove that the geometric definition (2 points) implies the algebraic definition (n points)!

📜

2. Convexity in General Spaces

Maurice Fréchet & Metric Spaces

Metric spaces were first proposed by French mathematician Maurice Fréchet in his 1906 thesis. A metric space is a pair \((X, d)\) where \(d\) is a distance function satisfying:

  • \( d(x,y) \geq 0 \) (Non-negativity)
  • \( d(x,y) = 0 \iff x=y \) (Identity)
  • \( d(x,y) = d(y,x) \) (Symmetry)
  • \( d(x,y) \leq d(x,z) + d(z,y) \) (Triangle Inequality)
Maurice Fréchet

Maurice Fréchet (1878-1973)

🌐 Menger Convexity

How do we define convexity if we don't have "lines"? Austrian mathematician Karl Menger (1928) proposed using only distance:

A metric space \((X, d)\) is Menger convex if for every distinct \(x, y \in X\), there exists a third point \(z \in X\) such that: \[ d(x,y) = d(x,z) + d(z,y) \]

Karl Menger

Karl Menger

Real Projective Space

Real Projective Space (Boy's Surface)

Morin Surface

Morin Surface (Sphere Eversion)

🧩

3. Examples of Convex Sets

Hyperplanes & Halfspaces

Hyperplane:

\[ \{ x \mid a^T x = b \} \]

Halfspace:

\[ \{ x \mid a^T x \leq b \} \]

Intuition: A hyperplane is like a flat sheet (line in 2D, plane in 3D). A halfspace is everything on one side of it.

Hyperplane & Halfspace
a aᵀx = b aᵀx ≤ b aᵀx > b

Norm Balls

A norm ball is defined as \( \{ x \mid \|x\| \leq 1 \} \). The shape depends on the norm!

L1 Norm (Diamond)

L2 Norm (Circle)

L∞ Norm (Square)

Quiz: Are all norm balls convex?
Answer: Yes, for \( p \geq 1 \). For \( 0 < p < 1 \), the "ball" is star-shaped but NOT convex!

Ellipsoids

\[ \{ x \mid (x-x_c)^T P (x-x_c) \leq 1 \} \]

where \( P \) is symmetric positive definite.

Ellipsoid Visualization

Think of it as a squashed sphere. The eigenvalues of \( P \) determine the axis lengths.

Cones

Norm Cone: \( \{ (x,t) \mid \|x\| \leq t \} \)

Also known as the "Ice Cream Cone" or Second-Order Cone.

Norm Cone Visualization

Polyhedra

A polyhedron is the intersection of a finite number of halfspaces and hyperplanes.

\[ P = \{ x \mid Ax \leq b, Cx = d \} \]

Note: A bounded polyhedron is called a polytope.

Polyhedron Visualization
🛠️

4. Operations Preserving Convexity

1. Intersection

If \( S_1 \) and \( S_2 \) are convex, then \( S_1 \cap S_2 \) is convex.

Generalization: The intersection of any number (even infinite!) of convex sets is convex.

2. Affine Maps

Let \( f(x) = Ax + b \) be an affine function.

  • The image of a convex set under \( f \) is convex.
  • The inverse image of a convex set under \( f \) is convex.

This means if you stretch, rotate, or translate a convex shape, it stays convex!

S f(x) f(S)

Affine map preserves convexity

Examples of Affine Operations

  • Scaling: \( \alpha S = \{ \alpha x \mid x \in S \} \) is convex.
  • Translation: \( S + a = \{ x + a \mid x \in S \} \) is convex.
  • Projection: The shadow of a convex object is convex!

🔍 Spot the Mistake!

Find the error in this statement about unions:

1. The intersection of two convex sets is always convex.

2. A single point is a convex set.

3. The union of two convex sets is always convex.

Previous Chapter

Gradients & Descent

Next Chapter

Analysis