Consider a finite dimensional real Banach space $E$, with norm say $|\cdot|$. Let $N$ denote the set of all norms on $E$. Suppose that $\varphi_1, \varphi_2 \in N$ have unit balls $B_1$ and $B_2$, respectively, and define $$ d(\varphi_1, \varphi_2) = \log \min \{ t \ge 1 : B_{1} \subseteq t B_2 \text{ and } B_{2} \subseteq t B_1\}. $$ Then $(N, d)$ is a metric space.

To every $\varphi \in N$ we may associate a unique ellipsoid $J_\varphi$, called John's ellipsoid, which is characterized by having the maximal Haar measure amongst all ellipsoids contained in the unit ball of $\varphi$. Of course $J_\varphi$ induces a Riemannian norm on $E$, so we may consider $\varphi \mapsto J_\varphi$ as a map from $N$ to itself.

It is a nice exercise to prove that $\varphi \mapsto J_\varphi$ is continuous, but does it have better regularity properties? Is it Holder or Lipschitz? If not, is there a big subset of $N$ for which the map does have better regularity properties (e.g. polytopes with finitely many facets). Perhaps there is another closely-related metric space (with either different objects or a different metric) for which this map has better regularity properties.


In general, functions defined using a max or sup are not smooth and at best Lipschitz. If you want something smoother, there are at least two generalizations of the John ellipsoid available:

The $L^p$ John ellipsoid

A dual (but not in the usual sense) ellipsoid

  • $\begingroup$ for the question I have in mind it is important that we use John's ellipsoid, but we only need Holder or Lipschitz continuity rather than any differentiability. $\endgroup$ May 10 '19 at 3:59
  • $\begingroup$ It should be Lipschitz but I don’t know a proof offhand. $\endgroup$
    – Deane Yang
    May 10 '19 at 12:24
  • $\begingroup$ Do you know a reference where a proof might be? $\endgroup$ May 31 '19 at 3:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.