17

Could you please give me a hint to prove that $\mathbb{R}^n$ with the 1-norm $\lvert x\rvert_1=\lvert x_1\rvert+\cdots+\lvert x_n\rvert$ is not isometric to $\mathbb{R}^n$ with the infinity-norm $\lvert x\rvert_\infty = \max_{i=1,\ldots,n}(\lvert x_i\rvert)$ for $n\gt2$.

I do not see how the critical case $n=2$ enters the picture.

Thank you!

t.b.
  • 78,116
zerhan
  • 171
  • 1
  • 4
  • 6
    Count the number of extreme points for the unit ball. – GEdgar Aug 07 '12 at 11:56
  • @GEdgar: the way I understand it, the question is about purely metric structure, so you can't really define extreme points. – tomasz Aug 07 '12 at 12:11
  • Thank you for your answers. Yes, it would be really great to get a proof that relies only on the metric structure, but I do see now how the critical case $n=2$ enters. – zerhan Aug 07 '12 at 12:16
  • Could you elaborate on what you mean by "how the critical case $n=2$ enters"? Did you find an explicit isometry between the two 2-dimensional spaces? (again look at the unit balls) – t.b. Aug 07 '12 at 12:21
  • 3
    The unit ball for the 1-norm is a cross-polytope with $2*n$ extreme points (in these two cases they are just the vertices) and the unit ball for the infinity norm is just an n-cube with $2^n$ vertices. So for $n=2$ there are in both cases $4$ vertices, but for $n>2$ the infinity unit ball has more. – zerhan Aug 07 '12 at 12:25
  • Moreover it should be intuitively clear that an isometry has to carry extreme points to extreme points, but I am trying to prove that right now. – zerhan Aug 07 '12 at 12:27
  • Yes, I found an explicit isometry for the $2$-dimensional case. – zerhan Aug 07 '12 at 12:28
  • 4
    I guess for knowing that an isometry carries extreme points to extreme points you need to know that surjective isometries between normed spaces are affine (hence linear if they preserve the origin). This isn't trivial: it's the Mazur-Ulam theorem. – t.b. Aug 07 '12 at 12:30
  • Yes, I saw that right now, because I need that the isometry is linear. – zerhan Aug 07 '12 at 12:35
  • @t.b.: That's a nice result! Thanks for the link. :) – tomasz Aug 07 '12 at 13:38
  • 1
    @tomasz: you're very welcome, I like it too :) see this thread for further remarks and pointers. – t.b. Aug 07 '12 at 13:41

2 Answers2

20

But one can do without the Mazur-Ulam theorem just as easily. I find it convenient to put the exponent $p$ in subscript, freeing the superscript for dimension: $\ell_\infty^n$ and $\ell_1^n$

Suppose $F:\ell_\infty^n\to \ell_1^n$ is an isometry (which I do not assume to be surjective or linear). By translation we can achieve $F(0)=0$. Let $V\subset \ell_\infty^n$ be the set of vertices of the unit cube, i.e., the set of all vectors with entries $\pm1$. This set has three interesting properties:

  1. all points are at distance $1$ from the origin
  2. all points are at distance $2$ from one another
  3. the cardinality of the set is $2^n$.

Let $W=F(V)$. The set $W\subset \ell_1^n$ also has the properties 1,2,3 listed above. For each vector $v\in W$ consider its positive and negative supports $\mathrm{supp}^+v=\{1\le j\le n : v_j> 0\}$ and $\mathrm{supp}^-v=\{1\le j\le n : v_j< 0\}$. If $n>2$, then $2^n>2n$. Therefore, we can assume that at least $n+1$ vectors in $W$ have nonempty positive supports. But then at least two of them, say $v'$ and $v''$, have overlapping positive supports. This creates cancellation when we compute distance between them: $\|v'-v''\|_1<\|v'\|_1+\|v''\|_1=2$ , a contradiction. QED

The proof actually shows that $\ell_{\infty}^n$ does not admit an isometric embedding (linear or not) into $\ell_1^N$ for $N<2^{n-1}$.

17

To wrap the comments up (assuming $n \geq 2$ since the cases $n=0,1$ are uninteresting):

From the Mazur–Ulam theorem we know that a surjective distance-preserving map between normed spaces is necessarily affine. After composing with a translation we may assume that our isometry preserves $0$, so we may assume that the isometry is linear right from the start.

A surjective linear isometry must map the extreme points of the unit ball of the domain onto the extreme points of the unit ball of the range. A moment's thought shows that the unit ball in the $\ell^1$-norm has $2n$ extreme points, namely the $n$ coordinate vectors and their $n$ negatives; and that the unit ball in the $\ell^\infty$-norm has $2^n$ extreme points: the vectors all of whose entries are $\pm 1$. Therefore isometry is only possible if $n =2$, where there is an isometry, given by $e_1 \mapsto e_1+e_2$ and $e_2 \mapsto -e_1 + e_2$.

Lee Mosher
  • 120,280
t.b.
  • 78,116