1
$\begingroup$

Can we find a universal constant $c>0$ such that for all $p,q\in\Delta:=\lbrace x\in (0,1)^{n}\ \colon\ x_{1}+\dots+x_{n}=1\rbrace$ it is true that \begin{equation} |p_{i}-q_{i}|\le c\left|\ln\frac{p_{i}}{g(p)}-\ln\frac{q_{i}}{g(q)}\right|,\quad i=1,\dots,n \end{equation} where $g(p):=\left(\prod_{i=1}^{n}p_{i}\right)^{\frac{1}{n}}$ is the geometric mean of $p$? The question is of interest since it upper bounds the Euclidean distance on $\Delta$ with the Aitchison distance used in compositional data analysis.
At least in dimension $n=2$ the answer is positive. Indeed, in that case the claim follows simply by the mean-value theorem:

\begin{align*} \frac{|p_{1}-q_{1}|}{|\ln\frac{p_{1}}{g(p)}-\ln\frac{q_{1}}{g(q)}|}=\frac{2|p_{1}-q_{1}|}{|\ln\frac{p_{1}}{1-p_{1}}-\ln\frac{q_{1}}{1-q_{1}}|}=2\xi(1-\xi)\le1, \end{align*} where $\xi\in \overline{p_{1}q_{1}}$.
Unfortunately, so far a failed to prove the general case (if true at all) and I'm grateful for any suggestions. Thanks in advance.

$\endgroup$
2
  • 1
    $\begingroup$ Is it not the case that there are multiple solutions to $p_i/g(p) = const$ when n > 2 ? A dimension counting argument would suggest this is the case. Then $|p_i - q_i| > 0 $ but the rhs = 0 $\endgroup$
    – mike
    Oct 17, 2019 at 12:42
  • $\begingroup$ I'm not sure. Taking logs in $p_{i}=cg(p)$ leads to the linear system $(\operatorname{id}-\frac{1}{n}\mathbf{1}\otimes\mathbf{1})x=\ln c\mathbf{1}$, which as far as I see admits solutions only if $c=1$. In that case solutions are of the form $a\mathbf{1}$ for some $a$. Then transforming back ,we end up with the apparently only solution $p=\frac{1}{n}\mathbf{1}$ since we require $p\in\Delta$. $\endgroup$
    – Tobsn
    Oct 17, 2019 at 13:56

1 Answer 1

1
$\begingroup$

Already in dimension $n=3$ the answer is negative. Let $p=(1/3,1/3,1/3)$ and $q=(1/7,2/7,4/7)$. Then $g(p)=1/3$ and $g(q)=2/7$ so $p_2/g(p)=q_2/g(q)$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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