20

Let $a_{1},a_{2},\cdots,a_{n},n\ge 2$ be real numbers,show that $$\dfrac{a_{1}a_{2}+a_{2}a_{3}+\cdots+a_{n-1}a_{n}}{a^2_{1}+a^2_{2}+\cdots+a^2_{n}}\le\cos{\dfrac{\pi}{n+1}}$$

I think this result is interesting.

When $n=2$, clearly

$$\dfrac{a_{1}a_{2}}{a^2_{1}+a^2_{2}}\le\dfrac{1}{2}=\cos{\dfrac{\pi}{3}}=\dfrac{1}{2}$$

When $n=3$, $$\dfrac{a_{1}a_{2}+a_{2}a_{3}}{a^2_{1}+a^2_{2}+a^2_{3}}\le\dfrac{\sqrt{2}}{2}.$$

This is true, because $$a^2_{1}+\dfrac{1}{2}a^2_{2}\ge \sqrt{2}a_{1}a_{2},$$ $$\dfrac{1}{2}a^2_{2}+a^2_{3}\ge\sqrt{2}a_{2}a_{3}.$$

But for general $n$ I cannot prove it.

hardmath
  • 37,015
math110
  • 93,304

4 Answers4

12

Here is a derivation using the calculus of variations.

Given that $$ \sum_{k=1}^na_k^2=1\tag{1} $$ we want to find the maximum of $$ \sum_{k=1}^na_ka_{k-1}\tag{2} $$ where $a_0=a_{n+1}=0$.

We want to find $a_k$ so that $(2)$ is stationary, that is, $$ \begin{align} 0 &=\sum_{k=1}^na_k\,\delta a_{k-1}+a_{k-1}\,\delta a_k\\ &=\sum_{k=1}^n(a_{k-1}+a_{k+1})\,\delta a_k\tag{3} \end{align} $$ for all $\delta a_k$ so that $(1)$ is constant: $$ 0=\sum_{k=1}^na_k\,\delta a_k\tag{4} $$ For both the $a_{k-1}+a_{k+1}$ in $(3)$ and the $a_k$ in $(4)$ to be perpendicular to all the same $\delta a_k$, we must have some $\lambda$ so that $$ a_{k+1}+a_{k-1}=\lambda a_k\tag{5} $$ Solving $(5)$, so that $a_0=a_{n+1}=0$, using the standard methods for linear recurrences yields $$ a_k=r\sin\left(\frac{\pi mk}{n+1}\right)\tag{6} $$ for $m\in\mathbb{Z}$.

We can compute $r$ using $(1)$ and $(6)$: $$ \begin{align} \sum_{k=1}^na_k^2 &=\sum_{k=1}^nr^2\sin^2\left(\frac{\pi mk}{n+1}\right)\\ &=-\frac{r^2}{4}\sum_{k=1}^n\left[\exp\left(\frac{2\pi imk}{n+1}\right)+\exp\left(\frac{-2\pi imk}{n+1}\right)-2\right]\\ &=-\frac{r^2}{4}(-2n-2)\ \Big[m\not\equiv0\bmod{(n+1)}\Big]\\ &=r^2\frac{n+1}{2}\ \Big[m\not\equiv0\bmod{(n+1)}\Big]\tag{7} \end{align} $$ We cannot satisfy $(1)$ when $m\equiv0\bmod{(n+1)}$; therefore, we require $m\not\equiv0\bmod{(n+1)}$, in which case, $$ r=\sqrt{\dfrac2{n+1}}\tag{8} $$

Finally, we need to compute $(2)$ for each $m\not\equiv0\bmod{(n+1)}$: $$ \begin{align} &\frac2{n+1}\sum_{k=1}^n\sin\left(\frac{\pi mk\vphantom{()}}{n+1}\right)\sin\left(\frac{\pi m(k-1)}{n+1}\right)\\ &=\frac1{n+1}\sum_{k=1}^n\left[\cos\left(\frac{\pi\vphantom{()}m}{n+1}\right)-\cos\left(\frac{\pi m(2k-1)}{n+1}\right)\right]\\ &=\frac{n}{n+1}\cos\left(\frac{\pi m}{n+1}\right)+\frac1{n+1}\cos\left(\frac{\pi m}{n+1}\right)\\ &=\cos\left(\frac{\pi m}{n+1}\right)\tag{9} \end{align} $$ The largest that $(9)$ can be if $m\not\equiv0\bmod{(n+1)}$ is $\cos\left(\frac\pi{n+1}\right)$.

Therefore, by homogeneity, $$ \frac{\displaystyle\sum_{k=1}^na_ka_{k-1}}{\displaystyle\sum_{k=1}^na_k^2}\le\cos\left(\frac\pi{n+1}\right)\tag{10} $$

robjohn
  • 345,667
  • Ingenious. However, should you not also prove that it is indeed the maximum and not merely a stationary point and that the maximum has to be a stationary point rather than a corner? In this respect, the quadratic form and eigenvalue approach of @Olivier Bégassat seems superior in dealing with these issues. But of course your $\lambda$ is just the eigenvalue there. However, with the variational (Lagrangian) approach it won't be complete without actually proving the maximality. – Hans Oct 31 '14 at 00:13
  • @Hansen: there are no corner cases in this problem, so the only extremal points are the stationary points. I have included the other solutions of $(5)$ which give the other stationary points, and smaller values for $(2)$. – robjohn Oct 31 '14 at 06:08
  • Yes, the function is smooth on a compact domain so there is no corner case. My point is that it needs to be stated explicitly for completeness. Your edition for the stationary points is nice. Not to be nitpicking, but the rationale for Eq.(5) should be that $\delta a_k$ is DEFINED as the WHOLE set (subspace) of functions that is perpendicular to $a_k$, rather than simply $a_{k_1}+a_k$ and $a_k$ are both perpendicular to $\delta a_k$. The latter reason is insufficient, because merely given two vectors perpendicular to a third does not imply the first two vectors are parallel to each other. – Hans Oct 31 '14 at 16:51
  • @Hansen: for $u$ and $v$ to be perpendicular to all the same vectors, they must be parallel. – robjohn Oct 31 '14 at 17:16
  • It is not clear what you mean by the bold face letter "all". Taken literally, the above statement is not true. Consider $u=(1,0,0),,v=(0,1,0),,w=(0,0,1)$. $u$ and $v$ are both perpendicular to the same vector $w$, yet they are perpendicular not parallel. Only when $w$ is first defined be members of the whole set of vectors perpendicular to $v$, THEN $u$ is perpendicular to $w$ implies $u$ is parallel to $v$. – Hans Oct 31 '14 at 17:44
  • @Hansen: $u$ is not perpendicular to all the same vectors as $v$. For instance, $v$ is perpendicular to $u$ and $u$ is not. They are not both perpendicular to all the same vectors. I have added the word 'both' to make this clearer. – robjohn Oct 31 '14 at 18:40
  • I do not think your statement is clear, even with the word "same". Formally the correct statement is: $(u\wedge v\implies u\perp w,\forall w\perp v) \implies u| v$. Your phrasing does not accurately reflect that. – Hans Oct 31 '14 at 20:16
  • What I am trying to say is $\forall w(u\perp w\iff v\perp w)\implies u|v$ – robjohn Oct 31 '14 at 20:44
  • I know that's what you were saying and the two statements are equivalent --- in the end. However, there is a slight difference. Let $v$ be $a_k$, and etc.. How do you argue naturally, from the outset, $w\perp u\implies w\perp v$ without first invoking $w\perp v\implies w\perp u$? – Hans Oct 31 '14 at 21:03
  • @Hansen: "We want to find $a_k$ so that $\color{#C00000}{\text{$(2)$ is stationary}}$ for all $\delta a_k$ so that $\color{#00A000}{\text{$(1)$ is constant}}$" This says the same thing as "we want to find $a_k$ so that $\color{#C00000}{\text{$a_{k-1}+a_{k+1}$ is perpendicular}}$ to the same $\delta a_k$ to which $\color{#00A000}{\text{$a_k$ is perpendicular}}$". So we are already given that $u$ is perpendicular to all the same vectors to which $v$ is perpendicular. – robjohn Nov 01 '14 at 02:21
9

I don't get the result you are looking for, a mistake may have crept in somewhere, I'll get back tot he question tomorrow morning. What I try to do is to find the spectrum of the matrix $D_n$ defined below. The solution uses the Tchebychev polynomials.

Edit. I made a mistake when identifying (the rescaled version of) the $\chi_n$ with the Tchebychew polynomials: the second polynomial $\tau_2=\frac12\chi_2(-2X)$ is actually equal to $2X^2-\frac12$ and not $T_2=2X^2-1$. I'm not sure I can rescue the proof.


Consider the real symmetric matrix $$D_n=\begin{pmatrix} 0&1\\ 1&0&1\\ &1&0\\ &&&\ddots\\ &&&&0&1\\ &&&&1&0 \end{pmatrix}$$ Then the equation above is equivalent to $$\langle D_nX\,|\,X\rangle\leq2|X|^2\cos\left(\frac{\pi}{n+1}\right)$$ for all nonzero vectors $X\in\mathbb{R}^{n}$, where $|X|^2=\langle X\,|\,X\rangle$. Because $D_n$ is diagonalizable in an orthonormal basis, it is enough to show that all the eigenvalues of $D_n$ are of absolute value less or equal than $2\cos\left(\frac{\pi}{n+1}\right)$.


The characteristic polynomial $\chi_n(X)=\det(D_n-XI)$ of $D_n$ the relations $\chi_1=-X$, $\chi_2=X^2-1$ (and $\chi_3=-X^3+2X$) and an easy manipulation of determinants shows that for all $n\geq 2$, $$\chi_n+X\chi_{n-1}+\chi_{n-2}=0$$

Compare this to the linear recursion satisfied by the Tchebychew polynomials of the first (and second) kind : $$T_n-2XT_{n-1}+T_{n-2}=0$$ The sequence $(\tau_n)=(\frac12\chi_n(-2 X))_n$ satisfies the recursion formula above, and the first two terms agree with the first two Tchebychew polynomials, for $\tau_1=X=T_1$ and $\tau_2=2X^2-1=T_2$ so that for all $n\geq 0$, $\tau_n=T_n$ and $$\chi_n(X)=2T_n\left(-\frac12 X\right)$$ Since the roots of $T_n$ are the $\cos\left(\frac{2k+1}{2n}\pi\right)$ for $k=0,\dots,n-1$, we get the roots of $\chi_n$, i.e. $$\mathrm{Spec}(D_n)=\left\lbrace 2\cos\left(\frac{2k+1}{2n}\pi\right)\left|\right. k=0,\dots,n-1\right\rbrace$$ And so all eigenvalues are (in abolute value) at most $2\cos\left(\frac{\pi}{2n}\right)$.

  • The error is that $\tau_2\ne T_2$. In fact, if $\tau_n=\chi_n(-2X)$, then $\tau_n=U_n(X)$ the Chebyshev Polynomial of the second kind. And we get the same result as in the proposed question. – Omran Kouba Jun 15 '14 at 09:16
  • @OmranKouba You are right! I didn't even bother to check for the Chebychev polynomials of the second kind, what a mistake ^^, thank you! – Olivier Bégassat Jun 15 '14 at 11:07
  • God damn, almost all of these answers to this question are great. Succinct, creative and unexpected. +1 – Fujoyaki Oct 30 '14 at 23:54
7

For $t \in (0,\frac{\pi}{n})$, we have $\sin kt > 0$, for $k \in 1,\cdots,n-1$.

Thus from AM-GM inequality, $\dfrac{\sin (k+1)t}{\sin kt}a_k^2+\dfrac{\sin kt}{\sin (k+1)t}a_{k+1}^2 \ge 2a_ka_{k+1}$,

with equality iff $a_{k+1}\sin kt = a_k\sin (k+1)t$.

Summing over we have, $\displaystyle \sum\limits_{k=1}^{n-1} \left(\dfrac{\sin (k+1)t}{\sin kt}a_k^2+\dfrac{\sin kt}{\sin (k+1)t}a_{k+1}^2\right) \ge 2\sum\limits_{k=1}^{n-1}a_ka_{k+1}$

$\displaystyle \implies 2a_1^2\cos t + \sum\limits_{k=2}^{n} \left(\dfrac{\sin (k-1)t}{\sin kt}+\dfrac{\sin (k+1)t}{\sin kt}\right)a_k^2 \ge \dfrac{\sin (n+1)t}{\sin nt}a_n^2 + 2\sum\limits_{k=1}^{n-1}a_ka_{k+1}$

$\displaystyle \implies 2\cos t\sum\limits_{k=1}^{n}a_k^2 \ge \dfrac{\sin (n+1)t}{\sin nt}a_n^2 + 2\sum\limits_{k=1}^{n-1}a_ka_{k+1}$

Putting, $t = \dfrac{\pi}{n+1}$, $\sin (n+1)t = 0$.

Therefore, $\displaystyle \cos \dfrac{\pi}{n+1}\sum\limits_{k=1}^{n}a_k^2 \ge \sum\limits_{k=1}^{n-1}a_ka_{k+1}$ as desired.

r9m
  • 17,938
4

Let $M_1 = \dfrac{1}{4\cos\dfrac{\pi}{n+1}}$ and $M_{k+1} = \dfrac{1}{4[\cos\dfrac{\pi}{n+1} - M_k]}$. If we can prove $0 <M_k < \cos\dfrac{\pi}{n+1}$ for $k < n -1$, then we have

\begin{align} a_1a_2 \leq \cos\dfrac{\pi}{n+1}a_1^2 + M_1 a_2^2 \\ a_2 a_3 \leq (\cos\dfrac{\pi}{n+1} - M_1)a_2^2 + M_2 a_3^2 \\ \vdots \\ a_{n-1}a_n \leq (\cos\dfrac{\pi}{n+1} - M_{n-2})a_{n-1}^2 + M_{n-1}a_n^2 \end{align}

thus all we need to prove is $0 <M_k < \cos\dfrac{\pi}{n+1}$ for all $ k < n-1$ and $M_{n-1} = \cos\dfrac{\pi}{n+1}$

To do this, we aim at finding constants $x$ and $y$, such that $\exists c$ with \begin{align} \frac{M_{k+1} + x}{M_{k+1} + y} = c \frac{M_{k} + x}{M_{k} + y} \end{align} After simplication, we find we can take $x$ and $y$ as different solutions of $$4\lambda^2 + 4\cos\dfrac{\pi}{n+1}\lambda + 1 = 0$$ and $c = \frac{x}{y}$, i.e.: \begin{align} x = -\frac{1}{2}(\cos\dfrac{\pi}{n+1} + i \sin\dfrac{\pi}{n+1})= -\frac{1}{2}e^{i\theta}\\ y = -\frac{1}{2}(\cos\dfrac{\pi}{n+1} - i \sin\dfrac{\pi}{n+1}) = -\frac{1}{2}e^{-i\theta}\\ c = (\cos\dfrac{\pi}{n+1} + i \sin\dfrac{\pi}{n+1})^2 = e^{i2\theta} \end{align} wiht $\theta = \frac{\pi}{n+1}$

Then we have \begin{align} \frac{M_{n-1} - \frac{1}{2}e^{i\theta}}{M_{n-1} - \frac{1}{2}e^{-i\theta}} = e^{2(n-2)i\theta} \frac{M_{1} - \frac{1}{2}e^{i\theta}}{M_{1} - \frac{1}{2}e^{-i\theta}} \end{align}

Since $M_1 = \dfrac{1}{2(e^{i\theta} + e^{-i\theta})}$, we get

\begin{align} \frac{M_{1} - \frac{1}{2}e^{i\theta}}{M_{1} - \frac{1}{2}e^{-i\theta}} = e^{4i\theta} \end{align}

Thus

\begin{align} \frac{M_{n-1} - \frac{1}{2}e^{i\theta}}{M_{n-1} - \frac{1}{2}e^{-i\theta}} = e^{2ni\theta} = e^{-2i\theta} \end{align}

Then we get easily $M_{n-1} = \dfrac{e^{i\theta} + e^{-i\theta}}{2}=\cos\frac{\pi}{n+1}$

Finally note that if $M_k < \cos\frac{\pi}{n+1} - \frac{1}{4\cos\frac{\pi}{n+1}}$

$$M_{k+1} - \cos\frac{\pi}{n+1} = \dfrac{1 - 4\cos^2\frac{\pi}{n+1} + 4M_k \cos\frac{\pi}{n+1}}{4(\cos\frac{\pi}{n+1} - M_k)} < 0$$

Note also that $M_{k+1} - M_{k} = \dfrac{1 - 4\cos\frac{\pi}{n+1}M_k + 4M_k^2}{4(\cos\frac{\pi}{n+1} - M_k)} > 0$ when $M_k < \cos\frac{\pi}{n+1} $

since we know that $M_{n-2} = \cos\frac{\pi}{n+1} - \frac{1}{4\cos\frac{\pi}{n+1}}$, we can conclude $M_k$ is increasing for $k < n-1$, therefore $0<M_k < \cos\frac{\pi}{n+1}$ for $k < n-1$