Who can help with the following inequality? I can prove it but using some rather ugly approach (e.g. by leveraging the derivative of $\frac{1}{\sqrt{t+1}}+\frac{1}{2}\sqrt{1-\frac{8}{t^2+8}}$ to show this is always less than 1 for $t>0$.

I'm just wondering if we can have some elegant simple prove. I guess we should use Jensen's inequality. Thanks.

Suppose $x,y,z\in R^+$ and $xyz=8$, try to prove that $$\sqrt{\frac{1}{1+x}}+\sqrt{\frac{1}{1+y}}+\sqrt{\frac{1}{1+z}}<2.$$

Please note that the usual AM-GM inequality may not do its trick here as the equality is rather hold on the boundary.


  • 1
    $\begingroup$ I am assuming that $x, y, z$ are all positive? $\endgroup$ – Prism Jul 18 '13 at 20:57
  • $\begingroup$ Yes. Sorry. Let me amend it. $\endgroup$ – user85356 Jul 18 '13 at 21:01
  • $\begingroup$ We might be able to use a slightly different inequality here: since $\|x\|_1\leq \sqrt n \|x\|_2$ (in an $n$ dimensional space), we have $$\sqrt{\frac{1}{1+x}}+\sqrt{\frac{1}{1+y}}+\sqrt{\frac{1}{1+z}}\leq 3 \sqrt{\frac{1}{1+x}+\frac{1}{1+y}+\frac{1}{1+z}}$$ $\endgroup$ – Omnomnomnom Jul 18 '13 at 21:28
  • 1
    $\begingroup$ Oh wow, I'm amazed that $xyz\geq 8$ needs to be used. The 'equality' condition seems to be as $y, z \rightarrow 0$. $\endgroup$ – Calvin Lin Jul 18 '13 at 21:57
  • $\begingroup$ As the LHS is convex, it cannot achieve maximum in the interior, i.e we need one or more of $x=0, y=0$ or $z=0$ satisfied. WLOG $x \le y \le z$, and taking $x \to 0$, we get $z = \frac8{xy} \to \infty$ and LHS $\to 1+\sqrt{\dfrac1{1+y}}+0< 2$, so this is the supremum. The "$8$" as you may see, is purely incidental, it could have been any other number for the same "maximum". $\endgroup$ – Macavity Aug 29 '14 at 6:51

Let $f(x,y,z)=\frac{1}{\sqrt{1+x}}+\frac{1}{\sqrt{1+y}}+\frac{1}{\sqrt{1+z}}$ and WOLOG assume that $x\geq y\geq z$ so we have $x\geq 2$. Consider the following two cases:

  1. $yz\geq 1$ : In this case we have $z\geq1/y$ and thus $$f(x,y,z)\leq f(x,y,1/y)\\ \leq\sqrt{3\left(\frac{1}{1+x}+\frac{1}{1+y}+\frac{y}{1+y}\right)}\\ = \sqrt{3\left(\frac{1}{1+x}+1\right)}\\ \leq \sqrt{3\left(\frac{1}{3}+1\right)}=2.$$ Equation in this case cannot be attained as we cannot have $z=1/y$ and $x=2$ simultaneously.
  2. $yz\leq 1$: Fix the product $yz=p^2$. Since we assumed $y\geq z$ we must have $y\geq p$. Furthermore, we have $p\leq 1$ for which the function $\frac{1}{\sqrt{1+y}}+\frac{1}{\sqrt{1+z}}$ would be decreasing with respect to $y$. Therefore, the mentioned function is minimized at $y=z=p$, i.e., $$\frac{1}{\sqrt{1+y}}+\frac{1}{\sqrt{1+z}}\leq\frac{2}{\sqrt{1+\sqrt{yz}}}\\=\frac{2}{\sqrt{1+\sqrt{8/x}}}.$$ Thus $f(x,y,z)\leq \frac{1}{\sqrt{1+x}}+\frac{2}{\sqrt{1+\sqrt{8/x}}}.$ The RHS is increasing in $x$ so the upper bound as $x\to \infty$ would be $f(x,y,z)<2$.

There might be neater way to show the second case.


Here's a solution using Jensen's inequality.

To get rid of the constraint, let's set $$ x = 2b/a \quad y = 2c/b \quad z = 2a/c $$ the inequality becomes $$ \sqrt{\frac a {a + 2b}} + \sqrt{\frac b {b + 2c}} + \sqrt{\frac c {c + 2a}} < 2 $$ Since $\sqrt t$ is a concave function, applying Jensen's inequality we get $$ \begin{align} \sum_{cyc}\sqrt{\frac a {a + 2b}} &= \sum_{cyc} \frac {a + c} {2(a + b + c)} \sqrt{\frac{4(a + b + c)^2}{(a + c)^2} \frac a {a + 2b}}\\ &\leq \sqrt{\sum_{cyc} \frac{2a(a + b + c)}{(a + c)(a + 2b)}} \end{align} $$ Now, it suffices to expand the expression $E$ inside the last square root.

If we call $N$ its numerator and $D$ its denominator then $4D - N$ results in a sum of monomials in $a, b$ and $c$ with positive coefficients. Therefore $$ E = \frac N D < 4 $$ and the inequality is proved.

  • $\begingroup$ How did you sort out $4D-N$ results in postive weights? I verified it in Matlab and indeed that is true yet just wondering if there is any trick revealing this more easily. $\endgroup$ – user85356 Jul 20 '13 at 14:22
  • $\begingroup$ @user85356 A full expanding here it's $10$ minutes of work without any software. $\endgroup$ – Michael Rozenberg Mar 18 '17 at 8:18

$a=\sqrt{x+1}>1,b=\sqrt{y+1}>1, c=\sqrt{z+1}>1$,for $a,b,c$,at least two of them $>2$ or $\le 2$,WOLG, assume $a,b$ either both $>2$ or both $\le2$, $z=\dfrac{8}{xy}=\dfrac{8}{(a^2-1)(b^2-1)}$,


case I: if $a>2$ and $b>2$

it is trivial $\dfrac{1}{a}+\dfrac{1}{b} <1 ,\dfrac{1}{c}<1$

so it is true in this case.

case II: when $a \le 2,b \le 2$,

now we need to prove:

$\dfrac{1}{a}+\dfrac{1}{b}+\sqrt{\dfrac{(a^2-1)(b^2-1)}{(a^2-1)(b^2-1)+8}}<2 \iff \dfrac{(a^2-1)(b^2-1)}{(a^2-1)(b^2-1)+8}< \left(2-\dfrac{1}{a}-\dfrac{1}{b} \right)^2 \iff \dfrac{(2ab-a-b)^2}{a^2b^2} > \dfrac{(a^2-1)(b^2-1)}{((a^2-1)(b^2-1)+8)}$

note: $(a^2-1)(b^2-1)+8>a^2b^2 \iff 9>a^2+b^2 \iff (a^2\le 4) \cap (b^2 \le 4) $

now we need to prove $(2ab-a-b)^2>(a^2-1)(b^2-1)$

$(2ab-a-b)^2= (a(b-1)+b(a-1))^2 \ge 4ab(a-1)(b-1) \iff 4ab >(a+1)(b+1) \iff (2a>a+1 )\cap (2b>b+1) \iff (a>1) \cap (b>1)$

it is true.

both cases are true. QED

Indeed, when $xyz=2$, this inequality is also true.


The problem can be restated as follows. Find the maximum of the function

$$ f(x,y,z)=\sqrt{\frac{1}{1+x}}+\sqrt{\frac{1}{1+y}}+\sqrt{\frac{1}{1+z}} $$

constrained to $xyz-8=0$. You can use a Lagrange multiplier and proceed.

  • 1
    $\begingroup$ Yes. That can prove. Can we have some more elementary approach? Maybe taking use of the Jensen's inequality? $\endgroup$ – user85356 Jul 19 '13 at 9:22
  • 1
    $\begingroup$ Jensen's inequality doesn't work. $\endgroup$ – chenbai Jul 19 '13 at 11:48

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.