Notes - Continuous Mathematics HT23, Convexity


Flashcards

What does mean for a set $D \subseteq \mathbb{R}^n$ to be convex?


\[\forall x_1, x_2, \forall \alpha \in (0, 1), (1-\alpha)x_1 + \alpha x_2 \in D\]

What does it mean for a function $f : D \to \mathbb{R}$ to be convex where $D$ is also convex?


\[\forall \pmb x_1, \pmb x_2 \in D, \forall \alpha \in (0, 1): f((1 - \alpha)\pmb x_1 + \alpha \pmb x_2) \le (1-\alpha)f(\pmb x_1) + \alpha f(\pmb x_2)\]

What’s the intuitive definition of a function being convex, in terms of secant lines?


Any secant line between two points on the surface is above the surface.

A function $f : D \to \mathbb{R}$ is convex on a convex domain $D$ if

\[\forall \pmb x_1, \pmb x_2 \in D, \forall \alpha \in (0, 1): f((1 - \alpha)\pmb x_1 + \alpha \pmb x_2) \le (1-\alpha)f(\pmb x_1) + \alpha f(\pmb x_2)\]

Can you turn this into the definition of being strictly concave?


\[\forall \pmb x_1, \pmb x_2 \in D, \forall \alpha \in (0, 1): f((1 - \alpha)\pmb x_1 + \alpha \pmb x_2) > (1-\alpha)f(\pmb x_1) + \alpha f(\pmb x_2)\]

A function $f : D \to \mathbb{R}$ is convex if

\[\forall \pmb x_1, \pmb x_2 \in D, \forall \alpha \in (0, 1): f((1 - \alpha)\pmb x_1 + \alpha \pmb x_2) \le (1-\alpha)f(\pmb x_1) + \alpha f(\pmb x_2)\]

Why is it necessary that $D$ is a convex set?


Otherwise $(1 - \alpha)\pmb x _ 1 + \alpha \pmb x _ 2$ might not be defined.

When, in terms of the Hessian, is some $f : D \to \mathbb R$ convex?


\[\mathbf H (f) \succeq 0\]

(i.e. positive semidefinite)

What functions are both convex and concave?


Linear functions

When is $f(\pmb x) = \pmb x^\intercal \mathbf A \pmb x$ convex or concave?


  • Convex: $\mathbf A$ positive semidefinite.
  • Concave: $\mathbf A$ negative semidefinite

If $f$ and $g$ are convex, is $f \times g$?


No.

Suppose $f$ and $g$ are functions. What rules let you remember if $f \circ g$ is convex/concave depending on the convexity of $f$ and $g$?


  • $f$ convex increasing of convex is convex
  • $f$ convex decreasing of concave is convex
  • $f$ concave increasing of concave is concave
  • $f$ concave decreasing of convex is concave.

If $g : \mathbb R \to \mathbb R$ is a strictly increasing function, what trick can you use to find

\[\text{argmin } f(x)\]

?


\[\text{argmin } f(x) = \text{argmin } g(f(x))\]

If you’re optimising against some objective function $f$, which is a tricky expression involving a large product, what might it be easier to optimise instead?


\[\log f\]

If $g : \mathbb{R} \to \mathbb{R}$ is a 1-1 function, what trick can you use to find

\[\text{argmin } f(x)\]

?


\[\text{argmin } f(x) = g(\text{argmin } f(g(x)))\]

How could you solve

\[\min f(x, y) \text{ s.t. } x + y + 1 = 0\]

without resorting to the method of Lagrange multipliers?


Notice that the condition would imply $y = -1 - x$ and then solve

\[\min f(x, -1 - x)\]

How could you solve

\[\min f(x, y) \text{ s.t. } x^2 + y^2 = 1\]

without resorting to the method of Lagrange multipliers?


Reparameterize

\[\min f(\cos \theta, \sin \theta)\]



Related posts