0
$\begingroup$

Let $\mathbb{S}^{n}$ be the canonical simplex of $\mathbb{R}^{n}$ and let $u = (1/n,\dotsc,1/n)$. Is it true that

$$\lVert x - u \rVert_p \leq \lVert y - u \rVert_p$$

implies that

$$\lVert x\rVert_p \leq \lVert y\rVert_p$$

for all $x,y \in \mathbb{S}^{n}$? It seems intuitive to me that this proposition is indeed true because the $p$-norm is minimized by $u$ in the $n$-simplex. However, I can't find a formal proof for that. Any ideas?

$\endgroup$
1
  • 1
    $\begingroup$ Do not use $N$ where $n$ will do. $\endgroup$ Commented Feb 4 at 3:40

1 Answer 1

2
$\begingroup$

The answer is yes if $p=2$ (by the Pythagoras theorem: $\|x\|_2^2=\|x-u\|_2^2+\|u\|_2^2$, because $x-u\perp u$) and, obviously, if $p=1$ or $n\le2$.

Otherwise, the answer is no, because then the value of $\|x-u\|_p$ does not determine the value of $\|x\|_p$.


Details on the last sentence: Suppose that $n\ge3$ and $p\in(0,\infty)\setminus\{1,2\}$. For small $t,s>0$, let $$x(t):=u+t(1,-1,0,\dots,0),\quad y(s):=u+s(1,-1/2,-1/2,0,\dots,0).$$ Then $x(t)$ and $y(s)$ are in the simplex and
$$\|x(t)-u\|_p=\|y(s)-u\|_p$$ if $$t=c_p s,\quad\text{where}\quad c_p:=(2^{-p}+1/2)^{1/p},$$ which latter will be assumed henceforth. Then (for $s\downarrow0$),
$$\|x(t)\|_p^p-n^{1-p}\sim n^{2-p}p(p-1)t^2=n^{2-p}p(p-1)c_p^2 s^2,$$ $$\|y(s)\|_p^p-n^{1-p}\sim \tfrac34\,n^{2-p}p(p-1)s^2.$$ So, if your conjecture were true, we would have $\|x(t)\|_p=\|y(s)\|_p$ and hence $c_p^2 =\tfrac34$, which would imply $p\in\{2,4\}$. But $p=2$ was excluded. It remains to exclude $p=4$.

But, if $p=4$, then $\|x(t)\|_p^p-\|y(s)\|_p^p\sim-3s^3/n$, so that $\|x(t)\|_p\ne\|y(s)\|_p$, the final contradiction.

Further detail, on why $c_p^2 =\tfrac34$ implies $p\in\{2,4\}$. Rewrite $c_p^2 =\tfrac34$ as $g(p):=2^{p-1}-3^{p/2}+1=0$. Note that $2\times3^{-p/2}g'(p)=(4/3)^{p/2}\ln2-\ln3$ is increasing in real $p>0$ from $\ln2-\ln3<0$ to $\infty$. So, for some real $p_*>0$, the function $g$ is decreasing on $(0,p_*]$ and increasing on $[p_*,\infty)$. So, $g$ has at most two positive roots. Also, $2$ and $4$ are roots of $g$, and hence the only positive roots of $g$.

$\endgroup$
5
  • 1
    $\begingroup$ Well, it's also true if $p=1$, for the simple reason that $||x||_1=1$ for every $x$ in the simplex. $\endgroup$ Commented Feb 4 at 2:51
  • $\begingroup$ @TomGoodwillie : Thank you for your comment. This is now fixed. $\endgroup$ Commented Feb 4 at 3:16
  • $\begingroup$ @IosifPinelis : it makes sense to me. Thank you very much for your answer! $\endgroup$ Commented Feb 4 at 12:44
  • $\begingroup$ Here is an alternative argument. Let $v$ be $(0,\frac{1}{n-1},\frac{1}{n-1},\dots)$ and parametrize the line through $u$ and $v$ by writing $x(t)=u+t(v-u)$. Since $t(v-u)$ and $-t(v-u)$ have equal $p$-norm, the conjecture implies that $x(t)$ and $x(-t)$ have equal $p$-norm, so that $||x(t)||_p$ is an even function of $t$ (for small $t$). But $||x(t)||_p^p$ is not even. It is $(\frac{1}{n(n-1)})^p((n-1)^p(1-t)^p+(n-1)(n-1+t)^p)$. The third derivative of $(n-1)^p(1-t)^p+(n-1)(n-1+t)^p$ at $t=0$ is $p(p-1)(p-2)(n-1)^{p-2}(-(n-1)^2+1)$, which is not zero if $n>2$ and $p\neq 1,2$. $\endgroup$ Commented Feb 5 at 3:17
  • $\begingroup$ @TomGoodwillie : Thank you for your further comment. Your argument is indeed a bit simpler. $\endgroup$ Commented Feb 5 at 5:29

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .