28
$\begingroup$

Let $P$ be a real polynomial of exact degree $2n$ ($n \geq 1$) whose zeros are real numbers and such that \begin{equation*} P(j) \geq 0 \quad \text{for any} \quad j \in \mathbb{Z}. \end{equation*}
Does there exist non-negative real numbers $\alpha_0,\alpha_1,\ldots,\alpha_n,$ with at least one of the $\alpha_i$ non-zero, such that the polynomial \begin{equation*} Q(x) = \sum_{k=0}^{n} \alpha_k P(x+k) \end{equation*} is non-negative on the whole real line, i.e.; $Q(x) \geq 0$ for any $x \in \mathbb{R}$ ?

I would like to add that this question is not merely a mathematical curiosity but pops up naturally while working on spectral transformations of discrete measures.

$\endgroup$
26
  • 2
    $\begingroup$ Yes, this follows from continuity (so $P$ has a minimum) and the fact that $P(x)$ tends to $+\infty$ as $|x|\to\infty$. $\endgroup$
    – user1688
    Jul 20, 2017 at 7:51
  • 4
    $\begingroup$ @WlodAA: I don't see how one can misunderstand this question which is well formulated (except for the missing implicit condition that at least one of the $\alpha_i$ should be nonzero). Also, I do not understand the downvote, the question looks reasonable and difficult to me. $\endgroup$ Jul 20, 2017 at 22:53
  • 8
    $\begingroup$ The good news is that it looks loke there exists a universal (independent of $P$) sequence $\alpha_k={n\choose k}^2$ with this property. The bad news is that I can prove it only for $n=1$ (trivial) and $n=2$ (a bit less trivial). $\endgroup$
    – fedja
    Jul 21, 2017 at 3:49
  • 2
    $\begingroup$ If $\alpha_k$ are fixed, we may without loss of generality suppose that our polynomial has roots at $0,\dots,2k-1$ and $n-k$ pairs of double zeroes. If it helps anyhow. $\endgroup$ Jul 21, 2017 at 11:59
  • 2
    $\begingroup$ by studying the extreme rays of the cone of polynomials which are non-negative at integer points $\endgroup$ Jul 21, 2017 at 12:54

1 Answer 1

30
$\begingroup$

Occasionally I wish somebody could give me a good whack on the head to keep my brains running and the older I get, the more frequently I need it. The problem is actually trivial.

I will prefer to think that $P$ is non-negative on some disjoint with integers arithmetic progression $\Lambda$ with step $1$ . Then we need to show that $$ \sum_{k=0}^n {n\choose k}^2 P(k)\ge 0\,. $$ Let $\lambda$ be a number such that $\Lambda=\{x:\cos(\pi x+\lambda)=0\}$ and put $Q(x)=x(x-1)\dots(x-n)$. Consider the meromorphic function $$ F(z)=\frac{\tan(\pi z+\lambda)-\tan\lambda}{Q(z)^2}P(z)\,. $$ Note that the poles of $F$ are simple and $F(z)$ decays like $|z|^{-2}$ on big circles between the poles of the tangent, so the sum of the residues converges to $0$. Now the residue at the zero $k$ of $Q$ is $\frac{\pi}{(n!)^2\cos^2\lambda}{n\choose k}^2P(k)$ while the residues at the poles $x\in\Lambda$ of the tangent are $-\frac 1{\pi Q(x)^2}P(x)\le 0$. The end.

$\endgroup$
14
  • 11
    $\begingroup$ Trivial. :-) ${} $ $\endgroup$ Jul 22, 2017 at 0:35
  • 5
    $\begingroup$ The interesting part is how did you come up with this identity: $\sum_{k} \binom{n}{k}^{2}P(k) =\frac{(n!)^{2} \cos^{2}\lambda}{\pi^{2}} \sum_{x \in \Lambda}\frac{P(x)}{Q(x)^{2}}$ $\endgroup$ Jul 22, 2017 at 3:53
  • 5
    $\begingroup$ @PaataIvanisvili The interesting part is actually why after teaching graduate complex analysis for the whole semester I didn't see it immediately :lol: It is clear (from the general duality mumbo-jumbo) that if the linear combination in question is, indeed, positive, then some identity like that must exist. Next, all such identities come from the residue theorem, so I need a function with poles at $\Lambda$ and extra poles at $0,\dots,n$ that is decaying at infinity faster than $P$ grows. I certainly know how to place poles with residues of the same sign at $\Lambda$ (continued) $\endgroup$
    – fedja
    Jul 22, 2017 at 4:40
  • 5
    $\begingroup$ That is what tangent is for. Extra poles and the decay dictate $Q^2$ in the denominator. Thus, I already get $\frac{\tan(\pi z+\lambda)}{Q(z)^2}P(z)$ forced upon me. Now, the second order poles are definitely not what I want. So that has to be corrected by offsetting the numerator by $\tan\lambda$ to make it vanish to the first order at integers. Then all that remains is to compute the residues. $\endgroup$
    – fedja
    Jul 22, 2017 at 4:46
  • 5
    $\begingroup$ Long ago I used to see few problems bit more trivial than this. $\endgroup$ Jul 22, 2017 at 8:28

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.