Let $x>0$ and $n$ be a natural number. Prove that: $$\left(\frac{x^n+1}{x^{n-1}+1}\right)^n+\left(\frac{x+1}{2}\right)^n\geq x^n+1.$$

This question is very similar to many contests problems, but I think it's much more harder than contest problem and it's just impossible to solve this problem during a competition.

In my fifth comment I wrote:

I took it here: https://math.stackexchange.com/questions/3304808/refinement-of-a-famous-inequality.

but I think there is no a chance that it will be solved there.

I tried to use $$M_n^2\geq M_{n+1}M_{n-1},$$ where $x>0$ and $M_n=\sqrt[n]{\frac{x^n+1}{2}}$, but without success.

I think a perspective way to solve this problem it's something like the following wrong way:


Thank you!

  • 3
    $\begingroup$ General claim with $a,b,c,d$ is certainly false already for $n=2$: say, check $(5,4,7,1)$. $\endgroup$ – Fedor Petrov Aug 2 '19 at 6:00
  • 2
    $\begingroup$ BTW I am curious where this problem came up. It seems rather finely tuned. $\endgroup$ – GH from MO Aug 2 '19 at 11:45
  • 5
    $\begingroup$ Over at math.se this was tagged "contest math". If this actually is from a contest, then it should not be posed as a problem here until after the contest is over. Does anyone know anything more about this? $\endgroup$ – Todd Trimble Aug 2 '19 at 22:51
  • 2
    $\begingroup$ @Yemon Choi My name is Michael. I don't think that this problem from a contest. Just this problem is very similar to many contest problems. I think this problem is very hard for a contest. $\endgroup$ – Michael Rozenberg Aug 4 '19 at 18:37
  • 3
    $\begingroup$ Is it ok to post on MO a question another user posted on MSE without even mentioning in the question that it was somebody else's idea? $\endgroup$ – Pierre-Yves Gaillard Aug 7 '19 at 12:22

Corrected proof, see GH from MO's comment and answer: A generalization of the inequality gives more flexibility for variations of parameters, which eventually yields a proof. One observation is $\frac{x+1}{2}=\frac{x^b+1}{x^{b-1}+1}$ for $b=1$. If one wants to achieve $(\frac{x^a+1}{x^{a-1}+1})^n+(\frac{x^b+1}{x^{b-1}+1})^n\geq x^n+1$ for all $x\ge0$, looking at derivatives at $x=1$ shows that a necessary condition is $a+b-1\ge n$. In fact, the following generalization holds true for all real $a,b\ge1$ and $x\ge0$: \begin{equation} \left(\frac{x^a+1}{x^{a-1}+1}\right)^{a+b-1}+\left(\frac{x^b+1}{x^{b-1}+1}\right)^{a+b-1}\geq x^{a+b-1}+1. \end{equation} Setting $n=a+b-1$, $a=(n+1)/2+c/2$, $x=y^2$ and $z=y^c$, this inequality is equivalent to \begin{equation} \left(\frac{y^{n+1}+z}{y^{n-1}+z}\right)^n+\left(\frac{y^{n+1}+1/z}{y^{n-1}+1/z}\right)^n\geq y^{2n}+1, \end{equation} where $y>0$ is arbitrary and $z>0$ is in an interval depending on $y$ and $n$. However, the inequality holds true for all $y>0$ and $z>0$. By the $z\leftrightarrow\frac{1}{z}$ and $y\leftrightarrow\frac{1}{y}$ symmetries, we assume in the following $0<z,y\le1$.

Fix $y$ and $n$. We show that the left hand side is monotonically increasing in $z$. Taking that for granted the assertion follows, for we have equality for $z=0$.

Taking the derivative with respect to $z$, the monotonicity is equivalent to \begin{equation} \left(\frac{y^{n+1}+z}{zy^{n+1}+1}\right)^{n-1}\ge \left(\frac{y^{n-1}+z}{zy^{n-1}+1}\right)^{n+1}. \end{equation} (In a previous version, there was a miscalculation, observed by GH from MO, which made the rest of the "proof" easier.)

Raising both sides to the $\frac{1}{(n-1)(n+1)}$-th power, and fixing now $y$ and $z$, the claim follows once we know that \begin{equation} (0,\infty)\to\mathbb R,\;\;t\mapsto\left(\frac{y^t+z}{zy^t+1}\right)^{\frac{1}{t}} \end{equation} is monotonically increasing in $t$. For this GH from MO's answer contains an elegant proof. Here is another one:

Take the derivative (with respect to $t$) of the logarithm of this function, then multiply by $t^2$ and set $w=y^t$. The result is \begin{equation} h(w,z):=\log w\cdot w\cdot(\frac{1}{w+z}-\frac{z}{zw+1})+\log(zw+1)-\log(w+z). \end{equation} We need to show that $h(w,z)\ge0$ for all $w,z\in(0,1)$. The derivative of $h(w,z)$ with respect to $w$ is \begin{equation} \frac{\partial h(w,z)}{\partial w}=\frac{(1-w^2)(1-z^2)z\log w}{((zw+1)(w+z))^2}<0. \end{equation} So $h(w,z)$, for fixed $z$, is decreasing in $w$. Thus $h(w,z)\ge h(1,z)=0$.

  • 1
    $\begingroup$ On the right hand side of your third display, a factor of $\frac{1}{z^2}$ is missing. That is, monotonicity is equivalent to $\left(\frac{y^{n+1}+z}{zy^{n+1}+1}\right)^{n-1}\ge \left(\frac{y^{n-1}+z}{zy^{n-1}+1}\right)^{n+1}$. I added below a detailed proof of this inequality, for any $n\in[1,\infty)$ and $y,z\in[0,1]$. $\endgroup$ – GH from MO Aug 10 '19 at 4:18
  • $\begingroup$ $\frac{x+1}{2}=\frac{x^b+1}{x^{b-1}+1}$ it's $\left(x^{b-1}-1\right)(x-1)=0,$ which for $b\neq1$ gives $x=1$ only. $\endgroup$ – Michael Rozenberg Aug 10 '19 at 10:17
  • 1
    $\begingroup$ @GHfromMO Thank you for spotting the miscalculation and fixing the proof! I added another argument for the final monotonicity assertion. $\endgroup$ – Peter Mueller Aug 10 '19 at 11:15
  • 1
    $\begingroup$ @MichaelRozenberg: Peter Mueller proved the inequality contained in his first display. Setting $a:=n$ and $b:=1$ in this inequality, you obtain the inequality in your post. Alternatively, start from his second display, and set $y:=x^{1/2}$ and $z:=x^{(n-1)/2}$ there. $\endgroup$ – GH from MO Aug 10 '19 at 11:49
  • 2
    $\begingroup$ It's an extremely nice proof! Thank you Peter! @GH from MO Thank you! $\endgroup$ – Michael Rozenberg Aug 10 '19 at 18:28

This is a supplement (correction) to Peter Mueller's nice solution. As he observed, it suffices to show that, for any fixed $n\geq 1$ and $y\in[0,1]$, the function $$z\mapsto\left(\frac{y^{n+1}+z}{y^{n-1}+z}\right)^n+\left(\frac{zy^{n+1}+1}{zy^{n-1}+1}\right)^n,\qquad z\in(0,1),$$ is increasing. (Indeed, $y:=x^{1/2}$ and $z:=x^{(n-1)/2}$ yields the LHS of the OP's inequality, while $y:=x^{1/2}$ and $z:=0$ yields the RHS of the OP's inequality.) Taking the derivative with respect to $z$, the statement becomes $$\left(\frac{y^{n+1}+z}{zy^{n+1}+1}\right)^{n-1}\ge \left(\frac{y^{n-1}+z}{zy^{n-1}+1}\right)^{n+1},\qquad y,z\in(0,1).$$ Let us now fix $y,z\in(0,1)$ and think of $n\geq 1$ as the variable. Taking the logarithm of both sides and dividing by $(n-1)(n+1)$, it suffices to show that the function $$t\mapsto\frac{1}{t}\log\frac{y^t+z}{zy^t+1},\qquad t>0,$$ is increasing. Making the change of variable $w:=y^t$, it suffices to show that the function $$w\mapsto\frac{\log(w+z)-\log(wz+1)}{\log w},\qquad w\in(0,1),$$ is increasing. Writing $w=:\tanh u$ and $z=:\tanh v$, it suffices to show that the function $$u\mapsto\frac{\log\tanh(u+v)}{\log\tanh(u)},\qquad u>0,$$ is increasing. Taking the derivative with respect to $u$, the statement becomes $$\sinh(u)\cdot\cosh(u)\cdot\log\tanh(u)\geq\sinh(u+v)\cdot\cosh(u+v)\cdot\log\tanh(u+v).$$ That is, it suffices to show that the function $$u\mapsto \sinh(u)\cdot\cosh(u)\cdot\log\tanh(u),\qquad u>0,$$ is decreasing. With the notation $s:=-\log\tanh(u)$, we have $$\sinh(u)\cdot\cosh(u)\cdot\log\tanh(u)=\frac{e^{-s}}{\sqrt{1-e^{-2s}}}\cdot\frac{1}{\sqrt{1-e^{-2s}}}\cdot(-s)=\frac{-s}{2\sinh s},$$ hence it suffices to show that the function $$s\mapsto\frac{\sinh s}{s},\qquad s>0$$ is increasing. However, this is clear, because the Taylor series of this function converges everywhere, and it has nonnegative coefficients.


$\newcommand{\s}{\overset{\text{sgn}}=} \newcommand{\Dx}{\text{Dx}} \newcommand{\logDx}{\text{logDx}} \newcommand{\DlogDx}{\text{DlogDx}} \newcommand{\DDDlogDx}{\text{DDDlogDx}} \newcommand{\DDDDDlogDx}{\text{DDDDDlogDx}} \newcommand{\dif}{\text{dif}} \newcommand{\Ddif}{\text{Ddif}} \newcommand{\R}{\mathbb{R}}$ Let us show that the inequality in question holds for all real $n\ge5$; the cases when $n\in\{1,2,3,4\}$ are verified directly. By a comment of Pietro Mayer, without loss of generality $0<x<1$. We shall reduce the problem to the completely algorithmic problem of checking sign patterns of several polynomials in $n,x$, of total degrees $\le11$. This reduction is done in a few steps:

Step 1: Eliminating $(\frac{1+x}2)^n$: The inequality in question can be rewritten as \begin{equation} u(x):=u_n(x):=n \ln \left(\frac{x^n+1}{x^{n-1}+1}\right) -\ln \left(x^n+1-z^n\right)\ge0, \end{equation} where $z:=z_x:=\frac{1+x}2$. Note that \begin{multline*} u'(x)\frac{x (1+x)}n \left(x^{1-n}+x^n+x+1\right) \left(x^n+1-z^n\right) \\ =\Dx:=\left(n \left(1-x^2\right)+\left(x^{2-n}-1\right) \left(1+x^n\right)\right) z^n-(n-1) \left(1-x^2\right) \left(1+x^n\right), \end{multline*} so that \begin{equation} u'(x)\s\Dx\s\logDx(x), \end{equation} where $\s$ denotes the equality in sign and
\begin{equation} \logDx(x):=\logDx_n(x):=n \ln z-\ln \frac{(n-1) \left(1-x^2\right) \left(1+x^n\right)}{n \left(1-x^2\right)+\left(x^{2-n}-1\right) \left(1+x^n\right)}. \end{equation} Here and in the sequel, $\Dx$, $\logDx$, etc. are atomic, "indivisible" symbols; $\Dx$ refers to the derivative (of $u$) in $x$, $\logDx$ refers to a certain kind of logarithmic modification of $\Dx$, etc. Next, let \begin{multline*} \DlogDx(x):=\DlogDx_n(x):= \\ \logDx'(x)(1-x) (1+x) x^{n-1} \left(1+x^n\right) \left(n \left(1-x^2\right)+\left(x^{2-n}-1\right) \left(1+x^n\right)\right) \\ =n^2 (x-1)^2 (x+1) \left(x-x^n\right) x^{n-2}-2 \left(x^n-1\right) \left(x^n+1\right)^2+\frac{n (x-1) \left(x^n+1\right)^2 \left(x^n+x\right)}{x}. \end{multline*} So, we get a polynomial in $x^n$ of degree $3$ over the field $\R(n,x)$ of all real rational functions in $n,x$.

Step 2: Reducing the degree from $3$ to $2$: Let \begin{multline*} \DDDlogDx(x):= \DlogDx''(x) x^{3 - 3 n}\\ =x^{3-3 n} (n (n x-n+2 x+2) (n^2 x^2-n^2+n x^2+2 n-1) x^{n-3} \\ -2 (n-1) n (x-1) (2 n^2 x^2-2 n^2+n x^2-2 n x+3 n+2 x) x^{2 n-4}+n (3 n-1) (3 n x-3 n-6 x+2) x^{3 n-3}) \\ \s\DlogDx''(x). \end{multline*} Taking the second derivative $\DlogDx''(x)$ of the polynomial $\DlogDx(x)$ in $x^n$ of over $\R(n,x)$ kills the free term of that polynomial. Thus, we get the polynomial $\DDDlogDx(x)$ of degree $2$ in $x^{-n}$ of over $\R(n,x)$.

Step 3: Reducing the degree from $2$ to $1$: Let \begin{equation} \DDDDDlogDx(x):= \frac{\DDDlogDx''(x)}{2 (n - 1) n^2 x^{-3 - 2 n}} = A_n(x) - x^n B_n(x), \end{equation} \begin{equation} A_n(x):=\left(2 n^3+3 n^2-5 n-6\right) x^4+\left(-2 n^3+3 n^2+3 n-2\right) x^3+\left(-2 n^3-n^2+5 n-2\right) x^2+\left(2 n^3-5 n^2+n+2\right) x, \end{equation} \begin{equation} B_n(x):=2 n^3+3 n^2-\left(-2 n^3+5 n^2-n-2\right) x^3-\left(2 n^3+n^2-5 n+2\right) x^2-\left(2 n^3-3 n^2-3 n+2\right) x-5 n-6, \end{equation} so that \begin{equation} \DDDlogDx''(x)\s A_n(x) - x^n B_n(x). \end{equation} Thus, we get the polynomial $\DDDDDlogDx(x)$ of degree $1$ in $x^n$ of over $\R(n,x)$.

Step 4: Reducing the degree from $1$ to $0$:
We can see that (under the conditions $n\ge5$ and $0<x<1$, assumed everywhere here) $B_n(x)>0$. So, $\DDDlogDx''(x)<0$ whenever $A_n(x)\le0$.

Further, let \begin{equation} \dif(x) = \dif_n(x) :=\ln\frac{A_n(x)}{B_n(x)} - n \ln x\s A_n(x) - x^n B_n(x)\s \DDDlogDx''(x) \end{equation} wherever $A_n(x)>0$, and then \begin{multline*} \Ddif(x) = \Ddif_n(x) :=\dif'(x)\frac{A_n(x)B_n(x)}{(n+1)(n-2)} \\ =-4 n^5 (x-1)^4 (x+1)^2+4 n^4 (x-1)^4 (x+1)^2+n^3 (x-1)^2 \left(15 x^4+16 x^3-10 x^2+16 x+15\right) \\ -4 n^2 \left(x^2-1\right)^2 \left(5 x^2+x+5\right)+n \left(-x^6+30 x^5+41 x^4-44 x^3+41 x^2+30 x-1\right) \\ +2 \left(3 x^6-6 x^5-11 x^4-36 x^3-11 x^2-6 x+3\right) \\ \s\dif'(x), \end{multline*} finally getting a polynomial in $n,x$.

Now we need to trace the above steps back:

Looking back at the polynomial $A_n(x)$, (for $x\in(0,1)$) we find that $A_n(x)\le0$ iff $x_1\le x\le x_2$, where $x_1=x_1(n)$ and $x_2=x_2(n)$ are the two roots of $A_n(x)$ in $(0,1)$ such that $x_1<x_2$.

Further, $\Ddif<0$ and hence $\dif'<0$ on $(0,x_1]$; and $\Ddif>0$ and hence $\dif'>0$ on $[x_2,1)$. So, $\dif$ decreases on $(0, x_1]$ and increases on $[x_2, 1)$. So, $\dif$ is $+-$ on $(0, x_1]$ (that is, $\dif$ can switch sign at most once on $(0, x_1]$, and only from $+$ to $-$). Similarly, $\dif$ is $-+$ on $[x_2, 1)$.

But also $\dif(1)=0$. So, actually $\dif<0$ on $[x_2, 1)$.

So, $\DDDlogDx''$ is $+-$ on $(0, x_1]$ and $\DDDlogDx'' < 0$ on $[x_2, 1)$.

Also, $A < 0$ and hence $\DDDlogDx'' < 0$ on $[x_1, x_2]$. So, $\DDDlogDx''$ is $+-$ on $(0, 1)$. So, $\DDDlogDx$ is convex-concave on $(0, 1)$. Also, $\DDDlogDx(1)=0$.

So, $\DDDlogDx$ is $+-+$ on $(0, 1)$. So, $\DlogDx$ is convex-concave-convex on $(0, 1)$. Also, $\DlogDx(1)=\DlogDx'(1)=\DlogDx''(1)=0>-8n(n^2-1)=\DlogDx'''(1)$ and $\DlogDx(0+)=2-n<0$. So, $\DlogDx$ is $-+$; so, $\logDx$ is decreasing-increasing.

Also, $\logDx(1-)=0$. So, $\logDx$ is $+-$, and hence so is $\Dx$ (with $z = \frac{1 + x}2$).

Recalling that $u'(x)\s\Dx$, we see that $u_n(x)$ is increasing-decreasing (in $x\in(0,1)$). Also, $u_n(0)=-\ln(1 - 2^{-n})>0$ and $u_n(1)=0$.

Thus, $u>0$, which concludes the proof.

  • $\begingroup$ I have highlighted the steps of the reduction to the polynomial, which should make the idea clearer. $\endgroup$ – Iosif Pinelis Aug 7 '19 at 17:32

Mathematica confirms that for concrete values of $n$, e.g.

n = 30; ForAll[x,x > 0, ((x^n+1)/(x^(n-1)+1))^n + ((x + 1)/2)^n >= x^n + 1];Resolve[%,Reals]


Mathematica fails with it in the general case of an arbitrary positive integer $n$.

  • 1
    $\begingroup$ Mathematica confirms such cases as $n=\pi$ too. $\endgroup$ – user64494 Aug 5 '19 at 5:38
  • 1
    $\begingroup$ I looked for a human's proof. $\endgroup$ – Michael Rozenberg Aug 25 '19 at 4:45

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.