3
$\begingroup$

A previous question was as follows:

Assume that $f\colon[0,1]\to[0,1]$ is a diffeomorphism so that $(f''(x)/f'(x))'<0$ and that $f''(0)=0$. It seems to me that $$\frac{1-f(x)^2}{1-x^2}\le f'(x)$$ on $[0,1]$. But no proof so far.

It was shown that the inequality in question does not hold in general.

The OP of that question then commented that the condition that $f$ be increasing was missing in the OP. So, the question then is whether the inequality in question holds with the just mentioned additional condition, that $f$ be increasing.

This amended problem is much more delicate, and the positive answer to it will be given below.

$\endgroup$

1 Answer 1

5
$\begingroup$

$\newcommand{\pa}{\partial}\newcommand{\diff}{\text{diff}}$Since $f$ is an increasing diffeomorphism, we have $f'>0$, so that \begin{equation*} f'=e^h \end{equation*} for some real-valued function $h$. Then $f''=e^h h'$ and hence $h''=(f''/f')'<0$, by one of the given conditions on $f$. So, the function $h$ is strictly concave. Also, the condition $f''(0)=0$ and the identity $f''=e^h h'$ imply $h'(0)=0$, so that the strictly concave function $h$ is (strictly) decreasing. Moreover, the conditions $f(0)=0$ and $f(1)=1$ imply \begin{equation*} 1=\int_0^1 f'=\int_0^1 e^h<e^{h(0)}, \end{equation*} since $h$ is decreasing. So, \begin{equation*} h(0)>0. \tag{10}\label{10} \end{equation*}

Take now any $x\in(0,1)$. By the strict concavity of $h$, for all $t\in(x,1)$, \begin{equation*} h(t)<h_x(t):=h(0)+\frac tx\,[h(x)-h(0)]. \end{equation*} Therefore and because $f(1)=1$, we have
\begin{equation*} \begin{aligned} f(x)&=1-\int_x^1 dt\, f'(t) \\ &=1-\int_x^1 dt\, e^{h(t)} \\ &>1-\int_x^1 dt\, e^{h_x(t)} \\ & =1-x e^{h(x)}\frac{1-e^{-(1/x-1)[h(0)-h(x)]}}{h(0)-h(x)} \\ & =1-x e^u R , \end{aligned} \end{equation*} where \begin{equation*} R:=R(v):=\frac{1-e^{-z(v-u)}}{v-u},\quad u:=h(x),\quad v:=h(0),\quad z:=1/x-1, \end{equation*} so that \begin{equation*} z>0\quad\text{and}\quad v>\max(0,u), \end{equation*} in view of \eqref{10} and because $h$ is decreasing.

Note that $(v-u)^2 e^{z (v-u)}R'(v)=1+z (v-u)-e^{z (v-u)}<0$, so that $R(v)$ is decreasing in $v$. Since $v>\max(0,u)$, we have $R(v)\le R(u+)=z$ if $u\ge0$ and $R(v)\le R(0)$ if $u<0$. So, \begin{equation*} f(x)>\left\{ \begin{alignedat}{2} f_1(u)&:=\max[0,1-(1-x)e^u] &&\text{ if } u\ge0, \\ f_2(u)&:=1-xe^u\frac{1-e^{z u}}{-u} &&\text{ if } u<0. \end{alignedat} \right. \end{equation*}

Let \begin{equation*} r:=\frac{1-f(x)^2}{f'(x)(1-x^2)}=\frac{1-f(x)^2}{e^u(1-x^2)} \end{equation*} and \begin{equation*} \diff:=1-f(x)^2-f'(x)(1-x^2)=1-f(x)^2-e^u(1-x^2). \end{equation*}

We want to show that $r<1$ or, equivalently, $\diff<0$.

Consider first the case $u\ge0$. If now $(1-x)e^u\le1$, then $f_1(u)=1-(1-x)e^u$ and hence \begin{equation*} r<\frac{1-f_1(u)^2}{e^u(1-x^2)}=\frac{e^u (x-1)+2}{1+x} \le\frac{ (x-1)+2}{1+x}=1, \end{equation*} so that $r<1$ indeed. If $(1-x)e^u>1$, then $f_1(u)=0$ and hence
\begin{equation*} r<\frac{1-f_1(u)^2}{e^u(1-x^2)}=\frac{1}{e^u(1-x^2)} <\frac{1}{1+x}<1, \end{equation*} so that we still have $r<1$.

Consider finally the case $u<0$. Here \begin{equation*} \diff< H(x):=1-f_2(u)^2-u(1-x^2) =1-(1-x^2) e^{w x}-\frac{(e^{w x}+w-e^w)^2}{w^2}, \end{equation*} where $w:=u/x<0$. Let \begin{equation*} H_1(x):=\frac{-w}{e^{w x}}\,H'(x)=w (1-x) (w x+w+2)+2 e^{w x}-2 e^w. \end{equation*} Then $H_1'(x)=-2 w (1+w x-e^{w x})<0$. So, $H_1$ is decreasing in $x\in(0,1)$, to $H_1(1)=0$. So, $H_1>0$. So, $H(x)$ is increasing in $x\in(0,1)$, to $H(1)=0$. So, $H(x)<0$ and hence $\diff<0$. $\quad\Box$

Remark: Inequality $$\frac{1-f(x)^2}{1-x^2}\le f'(x)$$ in the previous post does not make sense for $x=1$, but the slightly more general inequality \begin{equation*} 1-f(x)^2\le f'(x)(1-x^2), \tag{20}\label{20} \end{equation*} whose strict version we have proved for $x\in(0,1)$, obviously holds for all $x\in[0,1]$, by continuity. In fact, \eqref{20} will still hold if the condition $(f''/f')'<0$ is relaxed to the condition that $f'$ is log concave.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .