What's the smallest ellipsoid enclosing the half-sphere?

160 Views Asked by At

Let the half-sphere in $\mathbb{R}^n$ be defined by $$ R = \{\mathbf{x} \in \mathbb{R}^n : \Vert\mathbf{x}\Vert_2 \leq 1\;\wedge\;x_1 \geq 0\}\,. $$ Any ellipsoid can be defined using a positive definite matrix $\mathbf{A}$ and a center $\mathbf{a}$. The ellipsoid corresponding to $\mathbf{A}$ and $\mathbf{a}$ is $$ E(\mathbf{A,a}) = \{\mathbf{x} \in \mathbb{R}^n : \mathbf{(x-a)}^T\mathbf{A(x-a)} \leq 1\}\,. $$ The question is: for which pair $\mathbf{A}$, $\mathbf{a}$ where $R\subseteq E(\mathbf{A,a})$ is the volume of $E(\mathbf{A,a})$ the smallest? The volume of $E(\mathbf{A,a})$ being given by $$ \text{Vol}(E(\mathbf{A,a})) = \frac{\text{Vol}(S_n)}{\sqrt{\det \mathbf{A}}} $$ where $S_n$ is the $n$-dimensional unit hypersphere.

Looking at the formula for the volume of an ellipsoid it's clear that in order to minimize the volume we have to maximize $\det \mathbf{A}$ subject to $R\subseteq E(\mathbf{A,a})$ for some $\mathbf{a}$. Also it might be helpful that maximizing $\det \mathbf{A}$ is equivalent to maximizing $\log (\det \mathbf{A})$ which, as it turns out, is concave on the domain of positive definite matrices.

So far I've been able to find this exercise on the topic. However,

  1. I'm not sure if this is the easiest way to answer this question or if it's simply using Slater's condition and the KKT conditions because it wants to practice those.
  2. The overall structure of the argument is still unclear to me. c) shows that the primal optimum is equal to the dual optimum. And d) shows that the proposed $\overline{\mathbf{A}}$ might be an optimum. How does that imply that $\overline{\mathbf{A}}$ indeed is the optimum?