27
$\begingroup$

The inequality $\sqrt{\frac{a^b}{b}}+\sqrt{\frac{b^a}{a}}\ge 2$ for all $a,b>0$ was shown here using first-order Padé approximants on each exponent, where the minimum is attained at $a=b=1$.

By empirical evidence, it appears that inequalities of this type hold for an arbitrary number of variables. We can phrase the generalised problem as follows.

Let $(x_i)_{1\le i\le n}$ be a sequence of positive real numbers. Define $\boldsymbol a=\begin{pmatrix}a_1&\cdots&a_n\end{pmatrix}$ such that $a_k=x_k^{x_{k+1}}/x_{k+1}$ for each $1\le k<n$ and $a_n=x_n^{x_1}/x_1$. How do we show that $$\|\boldsymbol a\|_p^p\ge n$$ for any $p\ge1$?

As before, AM-GM is far too weak since the inequality $\displaystyle\|\boldsymbol a\|_p^p\ge 2\left(\prod_{\text{cyc}}\frac{x_1^{x_2}}{x_2}\right)^{1/{2p}}$ does not guarantee the result when at least one $x_i$ is smaller than $1$. We can eliminate the exponent on the denominator by taking $x_i=X_i^{1/p}$ so that $\displaystyle\|\boldsymbol a\|_p^p=\sum_{\text{cyc}}\frac{X_1^{X_2^{1/p}}}{X_2}$ but the approximant approach no longer becomes feasible; even in the case where $p$ is an integer the problem reduces to a posynomial inequality of rational degrees. Perhaps there are some obscure $L^p$-norm/Hölder-type identities of use but I'm at a loss in terms of finding references.

Empirical results: In the interval $p\in[1,\infty)$, Wolfram suggests that the minimum is $n$ (Notebook result) which is obtained when $\boldsymbol a$ is the vector of ones. However, we note that in the interval $p\in(0,1)$, the empirical minimum no longer displays this consistent behaviour as can be seen in this Notebook result. The sequence $\approx(1.00,2.00,2.01,3.36,3.00,4.00)$ appears to increase almost linearly every two values, but I cannot verify it for a larger number of variables due to instability in the working precision.

$\endgroup$
13
  • $\begingroup$ Interesting, I actually was thinking of a different kind of generalization, but I like this one as well. Of course, the usual stuff like when does one term dominate $n$ etc. will give you some easy bounds where the inequality is obvious. $\endgroup$ Oct 16 '21 at 10:38
  • $\begingroup$ Did you do some numerical experiment? Please check $n = 3, p = 1/3$, $x_1 = 1/2, x_2= 3/4, x_3 = 33/8$. $\endgroup$
    – River Li
    Oct 16 '21 at 11:34
  • $\begingroup$ @TheSimpliFire It looks good. Why someone downvoted it? $\endgroup$
    – River Li
    Oct 16 '21 at 14:33
  • $\begingroup$ A counterexample is $(x_1, x_2) = (1/2, 2)$ and $p=1/4$. Unless I made an error, the expression becomes $\approx 1.89 < 2$. Perhaps you need to add the restriction $p \ge 1$. $\endgroup$
    – Martin R
    Oct 18 '21 at 11:56
  • $\begingroup$ @MartinR What if $n\ge 3$? $\endgroup$
    – River Li
    Oct 18 '21 at 12:06
19
+500
$\begingroup$

Proof for $p ≥ 1$

Since $u^p - 1 ≥ p(u - 1)$ for all $u ≥ 0$, it suffices to prove the result for $p = 1$. That follows from

$$\frac{x^y}{y} - 1 ≥ \frac{1 + y \ln x}{y} - 1 = \ln x + \frac1y - 1 ≥ \ln x + \ln \frac1y = \ln x - \ln y$$

by cyclic summation over $(x, y) = (x_i, x_{i + 1})$.

Conjectured proof for $p ≥ \frac12$

Since $u^p - 1 ≥ 2p(u^{\frac12} - 1)$ for all $u ≥ 0$, it suffices to prove the result for $p = \frac12$. Numerical evidence suggests that

$$\left(\frac{x^y}{y}\right)^{\frac12} - 1 ≥ \frac{\ln x}{2\sqrt[4]{1 + \frac13 \ln^2 x}} - \frac{\ln y}{2\sqrt[4]{1 + \frac13 \ln^2 y}}$$

for all $x, y > 0$. If this is true, cyclic summation yields the desired result.

Counterexample for $0 < p < \frac12$

Let $g(x) = \left(\frac{x^{1/x}}{1/x}\right)^p + \left(\frac{(1/x)^x}{x}\right)^p$. Then $g(1) = 2$, $g'(1) = 0$, and $g''(1) = 4p(2p - 1) < 0$, so we have $g(x) < 2$ for $x$ in some neighborhood of $1$. This yields counterexamples for all even $n$:

$$\left(x, \frac1x, x, \frac1x, \dotsc, x, \frac1x\right), \quad x ≈ 1, \quad 0 < p < \frac12.$$

For $n = 3$, the best counterexample seems to be

$$(0.41398215, 0.73186577, 4.77292996), \quad 0 < p < 0.39158477.$$

$\endgroup$
18
  • 1
    $\begingroup$ Your proof is impressive. (+1) It is amazing that there is such a simple proof. $\endgroup$
    – River Li
    Nov 1 '21 at 9:06
  • 1
    $\begingroup$ Let $x=e^X,y=e^Y$ so$$e^{(Xe^Y-Y)/2}\ge1+\frac X{2(1+X^2/3)^{1/4}}-\frac Y{2(1+Y^2/3)^{1/4}}$$For $X\ge0$ it suffices that$$e^{(Xe^Y-Y)/2}\ge1+\frac X2-\frac Y{2(1+Y^2/3)^{1/4}}$$As $\arg\min_Xe^{(Xe^Y-Y)/2}-X/2=-Ye^{-Y}$ it suffices that$$(Y+2)e^{-Y}+\frac Y{(1+Y^2/3)^{1/4}}\ge2$$Equality occurs at $Y=0$ so it suffices to show LHS has derivative $>0$, or$$Y-\log(1+Y)-\frac54\log(3+Y^2)+\log(6+Y^2)>\log\frac2{3^{1/4}}$$Equality occurs at $Y=0$ so it suffices to show LHS has derivative $>0$, or$$1-\frac1{1+Y}-\frac{5Y}{2(3+Y^2)}+\frac{2Y}{6+Y^2}>0$$or $2Y^4-Y^3+17Y^2-18Y+18>0$ which is true. $\endgroup$
    – TheSimpliFire
    Nov 2 '21 at 11:17
  • 1
    $\begingroup$ @Hans I am using $u^p - 1 ≥ p(u - 1)$ to observe that the result for $p = 1$ implies the result for $p ≥ 1$. $\endgroup$ Nov 3 '21 at 0:32
  • 1
    $\begingroup$ I don't know why 100 reputation is given to my answer. It should be given to Anders Kaseorg's excellent answer. $\endgroup$
    – River Li
    Nov 3 '21 at 16:54
  • 1
    $\begingroup$ @RiverLi I agree; I suspect ErikSatie wanted to balance out the reputation. But thanks for starting the bounty, this answer totally deserves it. $\endgroup$
    – TheSimpliFire
    Nov 3 '21 at 17:21
6
$\begingroup$

We prove the result by first deriving a useful inequality for the function $(x,y)\mapsto (x^y/y)^p$.


It is clear that $$(x^y/y)^p \ge \left(\min_{\alpha \in \mathbb{R}_+} (\alpha x)^{\alpha y}/(\alpha y)\right)^p.\tag{1}$$ For fixed $x$ and $y$, we can find the minimum of the function $\alpha\mapsto (\alpha x)^{\alpha y}/(\alpha y)$ by taking derivatives and setting to $0$. Doing this (e.g., in Mathematica) shows that the minimum is achieved by $\alpha^\star = 1/({yW(ex/y)})$, where $W$ is the Lambert W function. Plugged into $(1)$, this gives $$(x^y/y)^p \ge \left[\left(\frac{x/y}{W(ex/y)}\right)^{1/W(e x/y)}W(ex/y)\right]^p$$ Note that the right hand side is a function of the ratio $x/y$ only. It will be convenient to write it as a function of the logarithm of $x/y$, $(x^y/y)^p\ge f(\ln x/y)$ where $$f(a):=\left[\left(\frac{e^a}{W(e^{a+1})}\right)^{1/W(e^{a+1})} W(e^{a+1})\right]^p.$$

We will use that $f(a)$ is a convex function for all $a\in\mathbb{R}$ and $p\ge 1$, since its second derivative is non-negative. This can be verified in Mathematica by running

FullSimplify[D[((Exp[a]/ProductLog[Exp[a + 1]])^(1/ProductLog[Exp[a + 1]])*ProductLog[Exp[a + 1]])^p, {a, 2}], Assumptions -> {Element[a, Reals]}]

which gives $$f''(a) = \frac{p e^{p-\frac{p}{W\left(e^{a+1}\right)}} W\left(e^{a+1}\right)^{p-2} \left((p-1) W\left(e^{a+1}\right)+p\right)}{W\left(e^{a+1}\right)+1}$$ which is non-negative for $p\ge 1$.


We are now ready to prove the result: $$ \sum_{i=1}^{n} \left(\frac{x_{i}^{x_{i+1}}}{x_{i+1}}\right)^p \ge n,\tag{2} $$ where we use the notation $x_{n+1}=x_1$. To begin, write $$\sum_i \left(\frac{x_{i}^{x_{i+1}}}{x_{i+1}}\right)^p \ge \sum_i f\left(\ln \frac{x_{i}}{x_{i+1}}\right) = n \frac{1}{n} \sum_i f\left(\ln \frac{x_{i}}{x_{i+1}}\right) \ge n f\left(\frac{1}{n} \sum_i \ln \frac{x_{i}}{x_{i+1}}\right)\tag{3}$$ where the first inequality uses $(x^y/y)^p\ge f(\ln x/y)$ while the second inequality is Jensen's. Note that $\sum_i \ln \frac{x_{i}}{x_{i+1}}=0$. It is also easy to verify from inspection that $f(0)=1$. Plugging into $(3)$ gives $$n f\left(\frac{1}{n} \sum_i \ln \frac{x_{i}}{x_{i+1}}\right)= n f(0) = n.$$ Combining with $(3)$ gives the desired result, $(2)$.

$\endgroup$
3
  • $\begingroup$ Very nice. (+1) This is the 2nd time I saw this trick (something like $\min_{\alpha \in \mathbb{R}_+} (\alpha x)^{\alpha y}/(\alpha y)$). But I forgot where I saw this trick. $\endgroup$
    – River Li
    Nov 1 '21 at 6:28
  • 1
    $\begingroup$ When proving $f$ convex, you cannot assume $a > 0$. $\endgroup$ Nov 1 '21 at 6:38
  • $\begingroup$ @AndersKaseorg Good point, actually the result does not depend on $a>0$ (I will update the answer) $\endgroup$
    – Artemy
    Nov 1 '21 at 7:19
4
+100
$\begingroup$

@Artemy's proof involves a complicated part of proving $f''(a)\ge 0$. Here is a proof which is a simplification of @Artemy's proof (one can do it by hand; also avoiding the use of the Lambert W function). Perhaps this is also helpful for the case when $p < 1$ (e.g. $\sqrt{\frac{x^y}{y}}+\sqrt{\frac{y^x}{x}}\ge 2$).


Fact 1: For any given $x, y > 0$, there exists a unique $u > 0$ such that $\frac{x}{y} = u\mathrm{e}^{u - 1}$. Furthermore, $$\frac{x^y}{y} \ge u\mathrm{e}^{1 - 1/u}.$$ (The proof is given at the end.)

Using Fact 1, let $$\frac{x_i}{x_{i + 1}} = u_i\mathrm{e}^{u_i - 1}, \,\, i = 1, 2, \cdots, n - 1; \qquad \frac{x_n}{x_1} = u_n\mathrm{e}^{u_n - 1}$$ where $u_1, \cdots, u_n > 0$. We have \begin{align*} 1 &= \frac{x_1}{x_2}\, \frac{x_2}{x_3} \cdots \frac{x_n}{x_1}\\ &= u_1 u_2 \cdots u_n \mathrm{e}^{u_1 + u_2 + \cdots + u_n - n}\\ &\le \left(\frac{u_1 + u_2 + \cdots + u_n}{n}\right)^n \mathrm{e}^{u_1 + u_2 + \cdots + u_n - n} \end{align*} which results in $$\frac{u_1 + u_2 + \cdots + u_n}{n} \ge 1. \tag{1}$$

Let $$f(u) = (u\mathrm{e}^{1 - 1/u})^p.$$ $f(u)$ is convex on $u > 0$ since $f''(u) = (u\mathrm{e}^{1 - 1/u})^p pu^{-4}[(p - 1)u^2 + 2(p - 1)u + p] > 0$ for all $u > 0$. Also, $f(u)$ is strictly increasing on $u > 0$ since $f'(u) = (u\mathrm{e}^{1 - 1/u})^ppu^{-2}(u + 1) > 0$ for all $u > 0$.

Using Fact 1 and (1), using the convexity and monotonicity of $f(u)$, we have \begin{align*} &\left(\frac{x_1^{x_2}}{x_2}\right)^p + \left(\frac{x_2^{x_3}}{x_3}\right)^p + \cdots + \left(\frac{x_n^{x_1}}{x_1}\right)^p\\ \ge\,& f(u_1) + f(u_2) + \cdots + f(u_n)\\ \ge\,& n\, f\left(\frac{u_1 + u_2 + \cdots + u_n}{n}\right) \\ \ge\,& n f(1)\\ \ge\,& n. \end{align*}

We are done.


Proof of Fact 1:

Since $u\mapsto u\mathrm{e}^{u - 1}$ is strictly increasing on $u > 0$, clearly, there exists a unique $u > 0$ such that $\frac{x}{y} = u\mathrm{e}^{u - 1}$.

We have $$\ln x = \ln y + \ln u + u - 1. $$ We need to prove that $$y\ln x - \ln y \ge \ln u + 1 - \frac{1}{u}.$$ It suffices to prove that $$y(\ln y + \ln u + u - 1) - \ln y \ge \ln u + 1 - \frac{1}{u}.$$

Let $$F(y) = y(\ln y + \ln u + u - 1) - \ln y - \left(\ln u + 1 - \frac{1}{u}\right).$$ We have $$F'(y) = \ln y + \ln u + u - \frac{1}{y},$$ and $$F''(y) = \frac{1}{y} + \frac{1}{y^2} > 0.$$ Also, $F(1/u) = 0$ and $F'(1/u) = 0$. Thus, $F(y) \ge F(1/u) = 0$ for all $y > 0$.

We are done.

$\endgroup$
8
  • $\begingroup$ With respect to possible extensions to the case $p < 1$ I'll just mention that the inequality does not hold for $n=2$ and $p=1/4$, see my comment at the question. $\endgroup$
    – Martin R
    Nov 1 '21 at 15:46
  • $\begingroup$ @MartinR I know your comment. But how about $n = 2, 1/2 \le p < 1$, or $n \ge 3, p > 0$? Any counterexample? $\endgroup$
    – River Li
    Nov 1 '21 at 15:49
  • $\begingroup$ I have no idea, I did not search for more (counter)examples. If I had to guess then I would say that $p \ge 1/2$ is sufficient. But I have no facts to back up that conjecture. $\endgroup$
    – Martin R
    Nov 1 '21 at 15:52
  • $\begingroup$ @MartinR Yes, before I think $p \ge 1/2$ might work. Actually, I even guess for $n \ge 3$, $p > 0$ works. A counterexample is expected. $\endgroup$
    – River Li
    Nov 1 '21 at 15:54
  • $\begingroup$ @MartinR A thing I want to tell you: Anders Kaseorg's bound $\left(\frac{x^y}{y}\right)^p - 1 - p \ln x + p \ln y \ge 0$ does not hold for $p=1/2$. $\endgroup$
    – River Li
    Nov 1 '21 at 16:02

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.