3
$\begingroup$

Suppose $x_1\ge x_2\ge \cdots \ge x_n\ge 0$ and $y_1\ge y_2\ge\cdots\ge y_n\ge0$ be reals such that for any positive integer $p$, $$ \sum_{i=1}^n x_i^p \geq \sum_{i=1}^n y_i^p. $$

Question: Is there a constant $c\ge 1$ such that for all $x,y$ satisfying the above, we have $cx\succ_w y$.

Here $cx$ denotes the sequence $cx_1, cx_2,\ldots, cx_n$ and $a \succ_w b$ means that $a$ weakly majorizes $b$ from below.

Recall that a sequence $a_1\ge a_2\ge \cdots\ge a_n$ is said to weakly majorize from below another sequence $b_1\ge b_2\ge\cdots\ge b_n$ if for all $1\le k \le n$ $$ \sum_{i=1}^k a_i\ge \sum_{i=1}^k b_i. $$

Update: user35593's example $x=(4,1,1)$ and $y=(3,3,0)$ shows that if there is such a $c$, then $c\ge 1.2$.

Remark 1: If $\sum f(x_i)\ge \sum f(y_i)$ for all convex non-decreasing functions, then we have $x\succ_w y$, a result due to Karamata; Hardy, Littlewood and Pólya. In fact testing against all functions of the form $z\mapsto (z-t)^+$ is enough.

$\endgroup$
  • 1
    $\begingroup$ Choosing $x=(4,1,1)$ and $y=(3,3,0)$ we see that $c$ is at least $1.2$. $\endgroup$ – user35593 Mar 24 '16 at 19:23
5
$\begingroup$

I think there is no such $c$. In fact, for $k$ and $\epsilon \gt 0$ given, there are $n \gg k$ and $x_1\ge x_2\ge \cdots \ge x_n\ge 0$ and $y_1\ge y_2\ge\cdots\ge y_n\ge0$ with $$\sum_{i=1}^n x_i^p \geq \sum_{i=1}^n y_i^p.$$ for all $p\ge 1$ but $$k\cdot \sum_{i=1}^k (x_i - \epsilon) \le\sum_{i=1}^k y_i$$ So $c$ had to be at least $k$.

The sequences are as follows: $$x_1 = 1 + \epsilon, x_2 = \cdots = x_n = \epsilon$$ $$ y_1 = \cdots = y_k = 1, y_{k+1} = \cdots = y_n = 0.$$ Now $$k\cdot \sum_{i=1}^k (x_i - \epsilon) = \sum_{i=1}^k y_i = k$$ So we only have to choose $n \gg k$ s. t. $$\sum_{i=1}^n x_i^p \geq \sum_{i=1}^n y_i^p = k \quad \forall p \ge 1$$ This is clearly possible, since $x_1^p \ge k$ for all but a finite number of $p$, and if $p$ is fixed, $n \ge 1 + \frac{k-1}{\epsilon^p}$ also does the job.

| cite | improve this answer | |
$\endgroup$

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.