Abhijeet Mulgund's Personal Webpage

Search

Search IconIcon to open search

Equivalent Conditions for Convexity

Last updated Nov 1, 2022

# Statement

Suppose $V$ is a Vector Space of $\mathbb{R}$ and $f: V \to \mathbb{R}$. Then The following are Equivalent:

  1. $f$ is a Convex Function
  2. The Epigraph of $f$, $\text{Epi}(f) \subset V \times \mathbb{R}$ (endowed with the Product Vector Space), is a Convex Set.
  3. (Check this - TODO) First Order Condition for Convexity: Further suppose $V$ is a Normed Vector Space and $f$ has Frechet Derivative $D_{v} f$ at all $v \in V$. For all $x, y \in V$ $$f(x) + D_{x}f(y-x) \leq f(y)$$
  4. (Check this - TODO) Second Order Condition for Convexity: Further suppose $V$ is a Normed Vector Space and $f$ has 2nd Frechet Derivative $D_{v}^{2}f$ at all $v \in V$. For all $x, y \in V$ $$D^{2}_{x}f(y)(y) \geq 0$$
  5. (Check this - TODO) $\forall b \in \mathbb{R}$ , ${x \in V: f(x) \leq b}$ is a Convex Set.
    • I think it is easiest to prove equivalence with (2). Do not (2) $\Rightarrow$ not (5), then (2) $\Rightarrow$ (5).
    • This makes it easy to prove Norms are Convex since it would follow from Balls are Convex

# Proof

(1) $\Rightarrow$ (2): Suppose $(u, a), (v, b) \in \text{Epi}(f)$. Then $f(u) \leq a$ and $f(v) \leq b$. Let $\lambda \in [0,1]$. Then $$f(\lambda u + (1-\lambda)v) \leq \lambda f(u) + (1-\lambda) f(v) \leq \lambda a + (1- \lambda) b$$ so $$\Big( \lambda u + (1- \lambda) v, \lambda a + (1-\lambda)b \Big) = \lambda (u, a) + (1- \lambda) (v, b) \in \text{Epi}(f)$$ and $\text{Epi}(f)$ is a Convex Set. $\checkmark$

(2) $\Rightarrow$ (1): Suppose $u, v \in V$ and $\lambda \in [0,1]$. Then, by definition of Epigraph, $(u, f(u)), (v, f(v)) \in \text{Epi}(f)$. Since $\text{Epi}(f)$ is Convex, we know $$S \ni \lambda (u, f(u)) + (1 - \lambda) (v, f(v)) = \Big(\lambda u + (1- \lambda)v, \lambda f(u) + (1- \lambda)f(v)\Big)$$ Thus, by definition of Epigraph, we have that $$f(\lambda u + (1-\lambda)v) \leq \lambda f(u) + (1-\lambda) f(v)$$ and $f$ is a Convex Function. $\checkmark$