4
$\begingroup$

I hope it is okay that I re-post my question from Math.SE. I know it is very specific. I would be grateful for thoughts on how to tackle this problem. Any idea is welcome!

For $n \in \mathbb{N}$, let the positive real numbers $x_1, \dots, x_n, y_1, \dots, y_n$ satisfy the following interlacing inequalities: $$ 0 < x_1 < y_1 < x_2 < y_2 < \dots < x_n < y_n. $$ Show that $$ a_{ij} := \sum_{k=1}^n \frac{y_k^2 - x_k^2}{x_k (y_i^2 - x_k^2) (y_j^2 - x_k^2)}\Big( \prod_{l \neq k} \frac{y_l^2 - x_k^2}{x_l^2 - x_k^2} \Big) > 0 $$ for all $i,j = 1, \dots, n.$

My thoughts so far:

  • The case $i=j$ is simple: $y_l^2 - x_k^2$ and $x_l^2-x_k^2$ have the same sign for all $k,l = 1, \dots, n.$ Since also $y_k^2 > x_k^2$ for all $k=1,\dots,n,$ every summand is positive.
  • The problem is symmetric, so we only need to consider $i > j$. Then $(y_i^2 - x_k^2)(y_j^2-x_k^2) < 0$ if and only if $j < k \le i.$ I tried to apply Lagrange's interpolation formula, defining $$ p(x) := \frac12\prod_{l=1}^n (x_l-x) \prod_{k=1}^n (x_l+x) = \frac12\prod_{k=1}^n (x_l^2 - x^2). $$ This is a polynomial of degree $2n$, and it holds that $$ p'(x_k) = - x_k \prod_{l \neq k} (x_l^2-x_k^2) \qquad \text{ and } \qquad p'(-x_k) = x_k \prod_{l \neq k} (x_l^2 - x_k^2) $$ for $k = 1, \dots, n.$ Since $$ \frac1{(y_i^2-x_k^2)(y_j^2-x_k^2)} = \frac1{y_i^2-y_j^2} \Big( \frac1{y_i^2-x_k^2} - \frac1{y_j^2 - x_k^2}\Big), $$ we may write \begin{align*} a_{ij} &= \frac1{y_i^2 - y_j^2}\Big( \sum_{k=1}^n \frac{f_i(-x_k)}{p'(-x_k)(y_i+x_k)} - \sum_{k=1}^n \frac{f_j(-x_k)}{p'(-x_k)(y_j+x_k)}\Big)\\ &= \frac1{y_i^2 - y_j^2}\Big( \sum_{k=1}^n \frac{f_i(x_k)}{p'(x_k)(y_i-x_k)} - \sum_{k=1}^n \frac{f_j(x_k)}{p'(x_k)(y_j-x_k)}\Big) \end{align*} with $$ f_m(x) := -\frac{\text{sgn}(x)}{y_m+x} \prod_{l=1}^n (y_l^2-x^2), \qquad m \in \{i,j\}. $$ Unfortunately, this is not a polynomial.
  • With a similar approach, one can show that $$ \sum_{k=1}^n \Big( \prod_{l \neq k} \frac{y_l^2 - x_k^2}{x_l^2 - x_k^2} \Big) \frac{y_k^2 - x_k^2}{(y_i^2 - x_k^2) (y_j^2 - x_k^2)} = \delta_{i=j}. $$ Maybe this is helpful?
$\endgroup$

3 Answers 3

4
$\begingroup$

Write $x_i$ for $x_i^2$ and $y_i$ for $y_i^2$, our sum is $$ a_{ij}=\sum_{k=1}^n \frac{f(x_k)}{\prod_{l\ne k} (x_k-x_l)},\,f(t)=-t^{-1/2}\prod_{l\ne i,j} (t-y_j). $$
By Lagrange interpolation, it is a coefficient of $t^{n-1}$ in the polynomial $g(t)$ of degree at most $n-1$, which interpolates the function $f$ in the points $x_1,\dots,x_n$. Thus by Rolle's theorem we have $a_{ij}\cdot (n-1)!=f^{(n-1)}(\theta)$ for a certain positive $\theta$. But $(n-1)$-st derivative of $f$ is clearly positive term-wise: $f=(-1)^{n-1}(c_1t^{-1/2}-c_2t^{1/2}+c_3t^{3/2}-\dots+(-1)^{n-2}c_{n-1}t^{n-5/2})$ for positive coefficients $c_1,c_2,\dots,c_{n-1}$.

Note that we did not use that $x$'s and $y$'s interlace, only that they are positive.

$\endgroup$
8
  • $\begingroup$ Why is it clear that $f^{(n-1)}$ is positive? $\endgroup$ Feb 28, 2017 at 9:53
  • $\begingroup$ look at the formula for $f$ (written now in the answer) and differentiate each term $n-1$ times. The sign is always positive. $\endgroup$ Feb 28, 2017 at 10:46
  • $\begingroup$ I am having trouble with the signs. The Lagrange polynomial is $$\Big(\sum_{k=1}^n \frac{f(x_k)}{\prod_{l \neq k}(x_l-x_k)}\Big)x^{n-1} + ...$$ (not $(x_k - x_l)$). If I follow your proof I get that $(-1)^{n-1}a_{ij} > 0$? $\endgroup$ Feb 28, 2017 at 12:53
  • $\begingroup$ Why? $x_k-x_l$, not $x_l-x_k$ $\endgroup$ Feb 28, 2017 at 13:19
  • $\begingroup$ To interpolate $f$ at the $x_1, \dots, x_n$, I chose $$ g(x) = \sum_{k=1}^n f(x_k) \Big( \prod_{l \neq k} \frac{x-x_l}{x_k-x_l}\Big) = \Big(\sum_{k=1}^n \frac{f(x_k)}{\prod_{l\neq k} x_k-x_l} \Big) x^{n-1} + \tilde{g}(x), $$ for some polynomial $\tilde{g}$ of degree $n-2$. $\endgroup$ Feb 28, 2017 at 13:39
4
$\begingroup$

The expression is equivalent to $$ a_{ij}=\sum_k \frac{\prod_{l\neq i,j} (y_l^2-x_k^2) }{x_k\prod_{l\neq k} (x_l^2-x_k^2)}. $$ This is a linear function in each $y^2_m$ for every $m$. Assume that there exist $y_m\in [x_1,x_n]$. Then the term gets smaller or equal if we replace $y_m$ by $x_1$ or $x_n$. However after the cancellations what is left is an expression of the same form with a smaller $n$. Hence one can prove the inequality by induction.

$\endgroup$
1
  • $\begingroup$ The assumption $y_m \in [x_1,x_n]$ makes this difficult: We are cancelling $y_m$ and $x_1$ (or $x_n$). For the induction to work, we now need another $y_{m'} \in [x_2, x_n]$ etc. $\endgroup$ Feb 28, 2017 at 11:27
2
$\begingroup$

Just assume $x_k, y_k>0$ without interlacing. If necessary by indexing, write $$a=\sum_k \frac{\prod_{l=1}^{n-2}(y_l^2-x_k^2)}{x_k\prod_{l\neq k}^{1,n}(x_l^2-x_k^2)}.$$ If $e_m(y_1^2,\dots,y_{n-2}^2)$ denotes the elementary symmetric polynomials, then $$\prod_{l=1}^{n-2}(y_l^2-x_k^2)=\sum_{r=0}^{n-2}(-1)^rx_k^{2r}\,e_{n-2-r}(y_1^2,\dots,y_{n-2}^2).$$ Since $e_m>0$, to prove $a>0$ it suffices to check that $$(-1)^r\sum_{k=1}^nx_k^{2r-1}\prod_{\ell\neq k}^{1,n}\frac1{x_{\ell}^2-x_k^2}>0 \qquad \text{for each $\,\,0\leq r\leq n-2$}.$$ I claim that there exists a homogeneous symmetric polynomial $P_r(x_1,\dots,x_n)$ with positive coefficients such that $$(-1)^r\sum_{k=1}^nx_k^{2r-1}\prod_{\ell\neq k}^{1,n}\frac1{x_{\ell}^2-x_k^2} =\frac{P_r(x_1,\dots,x_n)}{\prod_{j=1}^nx_j}\prod_{i<j}^{1.n}\frac1{x_i+x_j}.$$ This has now become a new query on MO. Update. There's a cute answer given now. Check it out!

$\endgroup$
1
  • $\begingroup$ Maybe Fedor Petrov's answer helps with your query as well? I can post my full proof if you want. $\endgroup$ Feb 28, 2017 at 18:39

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.