3

Problem. Let $n\in \mathbb{N}_{\ge 3}$. Let $0 \le a_1 \le a_2 \le \cdots \le a_n$. Prove or disprove that $$\sqrt{a_1a_2} + \sqrt{a_2a_3} + \ldots + \sqrt{a_na_1} \ge \sqrt[3]{a_1a_2a_3} + \sqrt[3]{a_2a_3a_4} + \ldots + \sqrt[3]{a_na_1a_2}.$$

I saw this problem recently which was originally asked 11 years ago (a Problem Statement Question (PSQ) - which does not meet the current quality standard ).

Some observations:

i) By letting $a_i = x_i^6, \forall i$, the desired inequality becomes $$\sum\limits_{\mathrm{cyc}} x_1^3 x_2^3 \ge \sum\limits_{\mathrm{cyc}} x_1^2 x_2^2 x_3^2.$$

ii) @ivan investigated the case $n = 4, 5, 6$ and found that with the substitution $x_1 = y_1, x_2 = y_1 + y_2, \cdots, x_n = y_1 + y_2 + \cdots + y_n$, the polynomial $g(y_1, y_2, \cdots, y_n) := f(y_1, y_1 + y_2, \cdots, y_1 + y_2 + \cdots + y_n)$ has only one negative coefficient (this term is $- y_1^4 y_2 y_n$). It suffices to prove that this is the case for $n\ge 7$.

iii) Based on @ivan's observation, let $$F(x_1, x_2, \cdots, x_n) := \sum\limits_{\mathrm{cyc}} x_1^3 x_2^3 - \sum\limits_{\mathrm{cyc}} x_1^2 x_2^2 x_3^2 - (x_2 - x_1 - x_n + x_{n-1})^2 x_1^4.$$ Let $$G(y_1, y_2, \cdots, y_n) := F(y_1, y_1 + y_2, \cdots, y_1 + y_2 + \cdots + y_n).$$ It suffices to prove that all coefficients of $G$ are non-negative.

River Li
  • 49,125
  • If this has been posted before then I suggest that you add a link to your question. Probably this one: https://math.stackexchange.com/q/293056/42969 ? – Martin R Sep 02 '24 at 08:49
  • @MartinR I am afraid some users vote to close it when they see that link - a PSQ. For example, this question, in one week, it happened closed -> reopen -> closed -> reopen. Zero tolerance for PSQ. – River Li Sep 02 '24 at 09:09
  • @MartinR I just looked at that link and found one close-vote. It seems no close-vote before I asked my question. – River Li Sep 02 '24 at 09:55
  • As you said, it is a PSQ without context. It should be closed. – Martin R Sep 02 '24 at 09:57
  • 1
    @MartinR That's absolutely right. Perhaps if we do not disclose the link, that question will not be closed, which does not affect the discussion of the problem in my question. In last month, I answered an old question which makes it close. Now I ask a new question which still makes it close. I feel uncomfortable for the closing the other user's question. – River Li Sep 02 '24 at 10:00
  • 1
    I ran some tests picking random $a_i$ which point towards the inequality being true. Tried to prove $\sum_{\text{cyc}}\sqrt{a_ia_{i+1}}\ge\sum_{\text{cyc}}\sqrt[4]{a_ia_{i+1}^2a_{i+2}}\ge\sum_{\text{cyc}}\sqrt[3]{a_ia_{i+1}a_{i+2}}$, but found counter-examples to the right-hand inequality. – Einar Rødland Sep 04 '24 at 06:27
  • 2
    this reminds me of the infamous – dezdichado Sep 17 '24 at 23:25
  • 1
    @dezdichado: Actually, the results provided in your link are exactly what is needed to prove the result using majorization and Karamata's inequality. I'll see if I can add this as a full answer. – Einar Rødland Oct 08 '24 at 09:35
  • @EinarRødland that would be nice - I thought it was only one or two slick transformations away from the inequality in the paper. – dezdichado Oct 08 '24 at 14:04
  • @EinarRødland or is it as simple as doing $a_i\to a_{n+1-i}$ and then letting $a_i = e^{b_i}$, which should preserve the same order for $b_i$ and $e^x$ is convex so we get : $$\sum e^{\frac{b_i+b_{i+1}}{2}}\geq \sum e^{\frac{b_i+b_{i+1} + b_{i+2}}{3}} ?$$ – dezdichado Oct 09 '24 at 16:07
  • 1
    @dezdichado: I've now added a new answer using Karamata's inequality: this and majorization were both new to me. The tricky part is to show that the terms used on the left hand side majorizes the terms used on the right hand side. The page you linked to contains such a proof, but I'll have to read it more carefully to convince myself it is not doing any hand-waving, and instead ended up making my own proof along the same lines. – Einar Rødland Oct 09 '24 at 22:12

3 Answers3

2

I'm leaving my previous approach at the end, but here's my much simpler proof of the inequality.

If $u\le v\le w$, we get $0\le(v-u)(w-v)=uv+vw-uw-v^2$: ie $uv+vw\ge uw+v^2$. We can apply this to $u,v,w=\sqrt{a_i},\sqrt{a_{i+1}},\sqrt{a_{i+2}}$ to get $$ \sqrt{a_ia_{i+1}}+\sqrt{a_{i+1}a_{i+2}}\ge \sqrt{a_ia_{i+2}}+a_{i+1} \quad\text{for}\quad i=1,\ldots,n-2 $$ which makes $$ \begin{align} 2\left(\sqrt{a_ia_{i+1}}+\sqrt{a_{i+1}a_{i+2}}\right)\ge & \sqrt{a_ia_{i+1}}+\sqrt{a_{i+1}a_{i+2}}+\sqrt{a_ia_{i+2}}+a_{i+1}\\ \ge & 3\,\sqrt[3]{a_ia_{i+1}a_{i+2}} + a_{i+1}. \end{align} $$ This makes $$ \sum_{i=1}^{n-2} \sqrt[3]{a_ia_{i+1}a_{i+2}}+\frac{a_{i+1}}3 \le \frac{2}{3}\left(\sqrt{a_1a_2}+\sqrt{a_{n-1}a_n}\right) +\frac43\left(\sqrt{a_2a_3}+\cdots+\sqrt{a_{n-2}a_{n-1}}\right). $$ Applying this, we get $$ \begin{align} \sum_\text{cyc}\sqrt{a_ia_{i+1}} \ge& \sqrt{a_1a_n}+\frac{\sqrt{a_1a_2}+\sqrt{a_{n-1}a_n}}{3} -\frac13\sum_{i=2}^{n-2}\sqrt{a_ia_{i+1}}\\ &+\sum_{i=1}^{n-2}\sqrt[3]{a_ia_{i+1}a_{i+2}} +\frac13\sum_{i=2}^{n-1}a_i. \end{align} $$ Using $D$ to denote the difference between the two sides, $$ \begin{align} D:=&\sum_\text{cyc}\sqrt{a_ia_{i+1}}-\sqrt[3]{a_ia_{i+1}a_{i+2}}\\ \ge& \sqrt{a_1a_n}+\frac{\sqrt{a_1a_2}+\sqrt{a_{n-1}a_n}}{3} -\frac13\sum_{i=2}^{n-2}\sqrt{a_ia_{i+1}}\\ &-\sqrt[3]{a_1a_2a_n}-\sqrt[3]{a_1a_{n-1}a_n} +\frac13\sum_{i=2}^{n-1}a_i \end{align} $$ to which we apply $\sqrt{a_ia_{i+1}}\le(a_i+a_{i+1})/2$ to obtain $$ D\ge \sqrt{a_1a_n}+\frac{\sqrt{a_1a_2}+\sqrt{a_{n-1}a_n}}{3} -\sqrt[3]{a_1a_2a_n}-\sqrt[3]{a_1a_{n-1}a_n} +\frac{a_2+a_{n-1}}6. $$ Next, using $x^sy^{1-s}\le sx+(1-s)y$ for $x,y\ge0$, $0\le s\le1$, entering $x=\sqrt{a_1a_2}, y=a_n, s=2/3$ makes $\sqrt[3]{a_1a_2a_n}\le\frac23\sqrt{a_1a_n}+\frac13\sqrt{a_2}$, etc. This makes $$ \begin{align} D\ge& \frac{\sqrt{a_1a_2}+\sqrt{a_{n-1}a_n}-\sqrt{a_1a_n}}{3} -\frac{a_2+a_{n-1}}{6}\\ =& \frac{\left(\sqrt{a_n}-\sqrt{a_2}\right) \left(\sqrt{a_{n-1}}-\sqrt{a_1}\right)}{3} -\frac{\left(\sqrt{a_{n-1}}-\sqrt{a_2}\right)^2}{6}\\ \ge& \frac{\left(\sqrt{a_{n-1}}-\sqrt{a_2}\right)^2}{6} \ge 0. \end{align} $$ And $D\ge0$ is exactly what we wanted to prove.

We also see from this that equality requires $a_1=a_2=a_{n-1}=a_n$ which means the $a_i$ are all equal.


Here is my previous approach, somewhat too briefly described. I'll try to update it since the middle step was somewhat unusual. I have checked the steps by simulations (random $a_i$) and think the method should be sound, but there may be typos and mistakes even so.

Let $A(a_1,\ldots,a_n)=\sum_\text{cyc}\sqrt{a_ia_{i+1}}$ and $B(a_1,\ldots,a_n)=\sum_\text{cyc}\sqrt[3]{a_ia_{i+1}a_{i+2}}$. We want to show that $(A-B)(a_1,\ldots,a_n)\ge0$ whenever $0\le a_1\le\cdots\le a_n$.

The proof has three steps.

  1. Show that $(A-B)(a_1,\ldots,a_n)\le(A-B)(a_1,\ldots,a_{i-1},a_{i+1},\ldots,a_n)$ for $3\le i\le n-2$.

  2. Starting with $a_1,\ldots,a_n$, for any $\epsilon>0$ we can fill in the gaps between $a_i$ and $a_{i+1}$ forming a new sequence $a'_1\le\cdots\le a'_N$ containing $a_i$ as a subsequence, so that $a'_1=a_1$, $a'_2=a_2$, $a'_{n-1}=a_{N-1}$, $a'_{n}=a_{N}$, and $a'_{i+1}-a'_i<\epsilon$ for $2\le i<N-1$. Then $A(a)-B(a)\ge A(a')-B(a')$ by repeated use of the above inequality starting with $a'_1,\ldots,a'_N$ and removing elements one at a time to obtain $a_1,\ldots,a_n$.

  3. Taking the limit as $\epsilon\rightarrow0$, we get a lower bound on $A(a)-B(a)\ge A(a')-B(a')$ which depends only on $a_1,a_2,a_{n-1},a_{n}$ which can be verified.

Proofs of the inequalities in 1. and 3. use the same approach as explained in the question: express inequality as polynomial in non-negative variables, and check for negative coefficients. I did this using the sympy Python package since I was already doing numerical simulations in Python to test hypotheses, but Sage is a good alternative.

Denoting $a=(a_1,\ldots,a_n)$ and $a^{(i)}=(a_1,\ldots,a_{i-1},a_{i+1},\ldots,a_n)$, $$ \begin{align} A(a)-A(a^{(i)}) =& \sqrt{a_{i-1}a_i}+\sqrt{a_ia_{i+1}}-\sqrt{a_{i-1}a_{i+1}}, \\ B(a)-B(a^{(i)}) =& \sqrt[3]{a_{i-2}a_{i-1}a_{i}}+\sqrt[3]{a_{i-1}a_{i}a_{i+1}} +\sqrt[3]{a_{i}a_{i+1}a_{i+2}} -\sqrt[3]{a_{i-2}a_{i-1}a_{i+1}}-\sqrt[3]{a_{i-1}a_{i+1}a_{i+2}}. \end{align} $$ If $3\le i\le n-2$, we can eliminate $a_{i-2}$ and $a_{i+2}$ by $$ \begin{align} B(a)-B(a^{(i)}) =& \sqrt[3]{a_{i-1}a_{i}a_{i+1}} -\sqrt[3]{a_{i-2}a_{i-1}}\cdot\left(\sqrt[3]{a_{i+1}}-\sqrt[3]{a_{i}}\right) +\sqrt[3]{a_{i+1}a_{i+2}}\cdot\left(\sqrt[3]{a_{i}}-\sqrt[3]{a_{i-1}}\right) \\ \ge& \sqrt[3]{a_{i-1}a_{i}a_{i+1}} -\sqrt[3]{a^2_{i-1}}\cdot\left(\sqrt[3]{a_{i+1}}-\sqrt[3]{a_{i}}\right) +\sqrt[3]{a^2_{i+1}}\cdot\left(\sqrt[3]{a_{i}}-\sqrt[3]{a_{i-1}}\right) \end{align} $$ If we let $a_{i-1}=x^6$, $a_i=(x+y)^6$, $a_{i+1}=(x+y+z)^6$ with $x,y,z\ge0$, we can express $$ \left[A(a^{(i)})-B(a^{(i)})\right] -\left[A(a)-B(a)\right] \ge P(x,y,z) $$ where $P(x,y,z)$ is a degree 6 polynomial with all non-negative coefficients. So, for $x,y,z\ge0$, we get $P(x,y,z)\ge0$, and so $$ A(a^{(i)})-B(a^{(i)}) \ge A(a)-B(a). $$

The original idea was to use induction starting with $n=4$, but that would require an inequality in the opposite direction: I confirmed this using simulations with random sequences.

Instead, I can insert extra values into the sequence, which reduces the difference $A-B$, and prove that $A-B\ge0$ even as $n$ increases. We can do so, except from the first and last two elements in the list which are not changed, and in such a way that we get a new sequence $a'_1\le\cdots\le a'_N$ as described in 2. Since $a'_{i+1}-a'_i<\epsilon=O(1/N)$ for $3\le i\le N-2$, we get $$ \sqrt{a'_ia'_{i+1}}-2\sqrt[3]{a'_ia'_{i+1}a'_{i+2}}+\sqrt{a'_{i+1}a'_{i+2}} = O(\epsilon^2) $$ as long as we take care to spread the $a'_i$ out evenly, eg $a'_i=\sqrt{a'_{i-1}a'_{i+1}}$ for all $a'_i$ not in the original sequence, and so the sum $$ \sum_{i=3}^{n-2} \sqrt{a'_ia'_{i+1}}-2\sqrt[3]{a'_ia'_{i+1}a'_{i+2}}+\sqrt{a'_{i+1}a'_{i+2}} \rightarrow 0\quad\text{as }\epsilon\rightarrow 0,\; N\rightarrow\infty. $$

Using the above limit, what remains of the sum $A(a')-B(a')$ becomes $$ \begin{align} A(a')-B(a')\rightarrow& \sqrt{a_1a_2}+\sqrt{a_1a_n}+\sqrt{a_{n-1}a_n}+\frac{a_2+a_{n-1}}{2} \\ &-\left( \sqrt[3]{a_1a_2^2}+\sqrt[3]{a_1a_2a_n} +\sqrt[3]{a_1a_{n-1}a_n}+\sqrt[3]{a_{n-1}^2a_n} \right) \end{align} $$ where we can enter $a_1=t^6$, $a_2=(t+u)^6$, $a_{n-1}=(t+u+v)^6$, $a_n=(t+u+v+w)^6$ with $t,u,v,w\ge0$ to obtain a polynomial $G(t,u,v,w)$ for the limit. Again, we expand and check the signs of the coefficients, and there is only one negative coefficient: $-t^4uw$. However, this is balanced out by coefficients $3t^4u^2$ and $3t^4w^2$, and the remaining terms are all non-negative. So $G(t,u,v,w)\ge0$ which proves that $$ A(a)-B(a) \ge \lim_{\epsilon\rightarrow0}A(a')-B(a') = G(t,u,v,w) \ge 0. $$

Einar Rødland
  • 11,378
  • 23
  • 39
2

I'm adding a new solution following the approach used in the Mongolian inequality problem referred to by dezdichado. It hasn't been proof-read properly, so read with care.

It uses Karamata's inequality, which is a generalization of Jensen's inequality to sequences $x_1,\ldots,x_n$ and $y_1,\ldots,y_n$ where one majorizes the other.

We say that $(x_1,\ldots,x_n)$ majorizes $(y_1,\ldots,y_n)$ if, for $x^*_i$ and $y^*_i$ denoting the sequences sorted in decreasing order, we have $\sum_{i=1}^k x^*_i\ge \sum_{i=1}^k y^*_i$ for all $k=1,\dots,n$ and with equality for $k=n$: ie $\sum_{i=1}^n x_i=\sum_{i=1}^n y_i$. For any convex function $f$, Karamata's inequality states that $$\sum_{i=1}^n f\left(x^*_i\right) \ge \sum_{i=1}^n f\left(y^*_i\right).$$ We may note that, in terms of the sequences sorted in increasing order $\hat x_1\le\cdots\le\hat x_n$ and $\hat y_1\le\cdots\le\hat y_n$, the sequence $(x_1,\ldots,x_n)$ majorizes $y_1,\ldots,y_n$ if $\sum_{i=1}^k\hat x_i\le\sum_{i=1}^k\hat y_i$ for all $k$ with equality for $k=n$.

Let $a_i=e^{u_i}$ where we for the time being ignore the alternative of $a_i=0$ (eg solve by taking the limit $u_i\rightarrow-\infty$). Define sequences $x_i=\frac{u_i+u_{i+1}}{2}$ and $y_i=\frac{u_i+u_{i+1}+u_{i+2}}{3}$ with $i$ cyclic as before. Using $f(t)=e^t$, we have $$ \sum_i\sqrt{a_ia_{i+1}} = \sum_i f(x_i) \quad\text{and}\quad \sum_i\sqrt[3]{a_ia_{i+1}a_{i+2}} = \sum_i f(y_i). $$ If we can prove that $(x_1,\ldots,x_n)$ majorizes $(y_1,\ldots,y_n)$, then the inequality $\sum_i f(x_i)\ge\sum_i f(y_i)$ follows. The Mongolian inequality problem referred to above contains a proof of this, but here is a slightly different proof.

The sequences $x_i$ and $y_i$ start of being increasing, but with the last few term wrapping around: $$ x_1\le y_1\le x_2\le\cdots\le x_{n-2}\le y_{n-2}\le x_{n-1} \quad\text{and}\quad x_1\le y_n\le x_n\le y_{n-1}\le x_{n-1}. $$ We may express the sums of the sorted values using $(t)_+=\max(0,t)$ to insert the wrapped terms: $$ \sum_{i=1}^k\hat x_i = \sum_{i=1}^k x_i - \left(x_k-x_n\right)_+ \quad\text{and}\quad \sum_{i=1}^k\hat y_i = \sum_{i=1}^k y_i - \left(y_k-y_n\right)_+ - \left(y_{k-1}-y_{n-1}\right)_+. $$ We now want to show that the difference $$ D_k = \sum_{i=1}^k \left(\hat y_i-\hat x_i\right) = \sum_{i=1}^k \left(y_i-x_i\right) - \left(y_k-y_n\right)_+ + \left(x_k-x_n\right)_+ - \left(y_{k-1}-y_{n-1}\right)_+ $$ is non-negative. Note that the terms inside $(\cdot)_+$ are in order $y_k-y_n\ge x_k-x_n\ge y_{k-1}-y_{n-1}$.

If $y_k\le y_n$, the none of the $(\cdot)_+$ terms are positive, and so $6D_k = 2u_{k+2}+u_{k+1}-2u_2-u_1\ge0$.

If $y_{k-1}>y_{n-1}$, all the $(\cdot)_+$ are positive, and $6D_k = u_n+2u_{n-1}-u_k-2u_{k-1}\ge0$.

This leaves the case where $y_k>y_n$ but $y_{k-1}\le y_{n-1}$. In this case, $$ \begin{align} D_k =& \sum_{i=1}^k \left(y_i-x_i\right) - \left(y_k-y_n\right)_+ + \left(x_k-x_n\right)_+\\ \ge& \sum_{i=1}^k \left(y_i-x_i\right) - \left(y_k-y_n\right) + \frac{2}{3}\left(x_k-x_n\right) =\frac{u_{k+1}-u_1}{6} \ge 0. \end{align} $$ We can thus conclude that $(x_1,\ldots,x_n)$ majorizes $(y_1,\ldots,y_n)$, and the inequality therefore follows from Karamata's inequality.

Einar Rødland
  • 11,378
  • 23
  • 39
1

Proof.

By AM-GM, we have $$\sqrt[3]{a_1a_2a_3} \le \frac{\sqrt{a_1a_2} + \sqrt{a_2a_3} + \sqrt{a_3a_1}}{3}, $$ $$\sqrt[3]{a_2a_3a_4} \le \frac{\sqrt{a_2a_3} + \sqrt{a_3a_4} + \sqrt{a_4a_2}}{3}, $$ $$\cdots \cdots \cdots \cdots,$$ $$\sqrt[3]{a_{n-2}a_{n-1}a_n} \le \frac{\sqrt{a_{n-2}a_{n-1}} + \sqrt{a_{n-1}a_n} + \sqrt{a_na_{n-2}}}{3},$$ $$\sqrt[3]{a_{n-1}a_n a_1} \le \frac{\sqrt{a_{n-1}a_n} + \sqrt{a_n a_1} + \sqrt{a_1 a_{n-1}}}{3},$$ and $$\sqrt[3]{a_na_1 a_2} \le \frac{\sqrt{a_n a_1} + \sqrt{a_n a_1} + a_2}{3}.$$ (Note: The last inequality is different from all the previous ones. )

Thus, we have \begin{align*} 3(\mathrm{LHS} - \mathrm{RHS}) &\ge 2\sqrt{a_1a_2} + \sum_{i=2}^{n-1} \sqrt{a_ia_{i+1}} - \sum_{i=1}^{n-2} \sqrt{a_i a_{i+2}} - \sqrt{a_1a_{n-1}} - a_2. \end{align*}

It suffices to prove that $$2\sqrt{a_1a_2} + \sum_{i=2}^{n-1} \sqrt{a_ia_{i+1}} - \sum_{i=1}^{n-2} \sqrt{a_i a_{i+2}} - \sqrt{a_1a_{n-1}} - a_2 \ge 0.$$

Letting $a_i = b_i^2, \forall i$, it suffices to prove that, for all $0\le b_1 \le b_2 \le \cdots \le b_n$, $$2b_1b_2 + \sum_{i=2}^{n-1} b_ib_{i+1} - \sum_{i=1}^{n-2} b_i b_{i+2} - b_1b_{n-1} - b_2^2 \ge 0. \tag{1}$$

We use the Mathematical Induction.

It is clearly true for $n = 3$ since (1) is written as $(b_3 - b_2)(b_2 - b_1) \ge 0$.

Assume that the statement is true for $n$ ($n\ge 3$).

For $n + 1$, we need to prove that, for all $0 \le b_1 \le b_2 \le \cdots \le b_n \le b_{n+1}$, $$2b_1b_2 + \sum_{i=2}^{n} b_ib_{i+1} - \sum_{i=1}^{n-1} b_i b_{i+2} - b_1b_{n} - b_2^2 \ge 0. \tag{2}$$ (2) is written as $$2b_1b_2 + \sum_{i=2}^{n-1} b_ib_{i+1} - \sum_{i=1}^{n-2} b_i b_{i+2} - b_1b_{n-1} - b_2^2$$ $$+ (b_n b_{n+1} - b_{n-1}b_{n+1} + b_1b_{n-1} - b_1 b_n) \ge 0$$ which is written as $$\left(2b_1b_2 + \sum_{i=2}^{n-1} b_ib_{i+1} - \sum_{i=1}^{n-2} b_i b_{i+2} - b_1b_{n-1} - b_2^2\right) + (b_n - b_{n-1})(b_{n+1}-b_1) \ge 0$$ which is true using the inductive hypothesis. Thus, the statement is true for $n+1$.

We are done.

River Li
  • 49,125
  • This looks right to me. Checking the coefficients at the end for general $n$ is a bit of a pain, I guess, but should be doable. I guess it's the last of the top inequalities that gives you the necessary margin for the proof to work. – Einar Rødland Sep 18 '24 at 01:39
  • @EinarRødland Thanks. I will try some approaches. – River Li Sep 18 '24 at 01:53
  • @EinarRødland I found it works by the Mathematica Induction. – River Li Sep 18 '24 at 02:50