2
$\begingroup$

I would like to prove the following but I couldn't manage to do it. Let $a>b>0$ be two real numbers. Let $f$ be the function defined as:

$$\forall \sigma>0, ~\forall x\in\mathbb{R},~f_\sigma(x):=\sigma e^{-x^2/\sigma^2}+\sqrt{\pi}x\text{erf}\left(\frac{x}{\sigma}\right).$$

I would like to show that $f_a-f_b$ is log-concave.

One can see $f_\sigma$ as the function that, differentiated two times, gives the gaussian function (with a scaling factor) of mean $0$ and deviation $\sigma$. More formally, one can compute the second derivative and show that: $$f_\sigma''(x) = \frac{e^{-x^2/\sigma^2}}{2\sigma}.$$

Numerically, I am pretty sure that this statement is true. I tried to compute the second derivative of the $\log(f_a-f_b)$ but without sucess (the derivatives are too diffucult to analyse).

Any hints or solutions will be highly appreciated! Thank you.

$\endgroup$
8
  • $\begingroup$ (i) "One can see $f$ as the second anti-derivative of the gaussian distribution (with a scaling factor)." -- What is $f$? What is "the second anti-derivative"? Of what "gaussian distribution"? What is a "second anti-derivative" of a distribution (which latter is a measure)? (ii) Do you have an inequality between $a$ and $b$? $\endgroup$ Commented Aug 20, 2023 at 15:09
  • $\begingroup$ Thank you @IosifPinelis for asking details. I updated the post, I hope it is clearer. $\endgroup$
    – NancyBoy
    Commented Aug 20, 2023 at 15:19
  • $\begingroup$ "One can see $f_\sigma$ as the function that, differentiated two times, gives the gaussian distribution (with a scaling factor) of mean $0$ and deviation $\sigma$." -- As I noted, a distribution cannot be the second derivative of a function. Also, in this case, the second derivative will also depend on $a$. $\endgroup$ Commented Aug 20, 2023 at 15:29
  • $\begingroup$ I updated the post, is it correct now ? $\endgroup$
    – NancyBoy
    Commented Aug 20, 2023 at 15:56
  • $\begingroup$ "where $Z$ does not depend on $\sigma$"... How come? I think this post may require quite a bit of work. $\endgroup$ Commented Aug 20, 2023 at 16:00

1 Answer 1

2
$\begingroup$

This conjecture is true.

Indeed, for $h:=\ln(f_a-f_b)$ and real $x\ne0$ we have \begin{equation} h'(x)=R(x):=\frac{F(x)}{G(x)}, \end{equation} where \begin{equation} F(x):=\sqrt{\pi } \left(\text{erf}\left(\frac{x}{a}\right)-\text{erf}\left(\frac{x}{b}\right)\right), \end{equation} \begin{equation} G(x):= a e^{-x^2/a^2}+\sqrt{\pi } x\, \text{erf}\left(\frac{x}{a}\right)-b e^{-x^2/b^2}-\sqrt{\pi } x \, \text{erf}\left(\frac{x}{b}\right); \end{equation} note that $G(0)=a-b>0$ and $G'(x)=\sqrt{\pi } \left(\text{erf}\left(\frac{x}{a}\right)-\text{erf}\left(\frac{x}{b}\right)\right)>0$ for real $x>0$, so that $G>0$ on $[0,\infty)$.

Since the function $h$ is smooth and even, it is enough to show that $R$ is decreasing on $(0,\infty)$. In what follows, $a>b>0$ and $x>0$, unless otherwise indicated.

Let \begin{equation} R_1(x):=\frac{F'(x)}{G'(x)}, \end{equation} \begin{equation} R_2(x):=\frac{F''(x)}{G''(x)} = -\frac{2 x \left(a^3 e^{x^2/a^2}-b^3 e^{x^2/b^2}\right)}{a^2 b^2 \left(a e^{x^2/a^2}-b e^{x^2/b^2}\right)}; \end{equation} note that $R_2(x)$ is undefined at $x=x_{a,b}$, where \begin{equation} x_{a,b}:=a b \sqrt{\frac{\ln a-\ln b}{a^2-b^2}}. \end{equation}

Then \begin{equation} \begin{aligned} H(t)&:=R'_2(x)\frac{1}{2} a^3 b^3 \left(a e^{x^2/a^2}-b e^{x^2/b^2}\right)^2 \\ & =a^4 \left(b^2-2 t\right)-a b^5 e^{\left(\frac{1}{b^2}-\frac{1}{a^2}\right)t}+a^2 \left(b^4+4 b^2 t\right)-a^5 b e^{\left(\frac{1}{a^2}-\frac{1}{b^2}\right)t}-2 b^4 t, \end{aligned} \end{equation} where $t:=x^2>0$, and \begin{equation} H'(t)\frac{ab}{a^2-b^2}=H_1(u):=\frac{a^4}{u}-2 a^3 b+2 a b^3-b^4 u, \end{equation} where $u:=e^{\left(\frac{1}{a^2}-\frac{1}{b^2}\right)t}>1$. The only root $u$ of the equation $H_1(u)=0$ that may be positive is \begin{equation} u_{a,b}:=\frac{a b^3-a^3 b+\sqrt{a^6 b^2-a^4 b^4+a^2 b^6}}{b^4}. \end{equation} Replacing now $e^{\left(\frac{1}{a^2}-\frac{1}{b^2}\right)t}$ and $e^{\left(\frac{1}{b^2}-\frac{1}{a^2}\right)t}$ in the expression for $H(t)$ by $u_{a,b}$ and $1/u_{a,b}$, respectively, we see that the condition $H'(t)=0$ for some real $t>0$ will imply \begin{equation} \begin{aligned} H(t)&=a^4 \left(-\frac{b^4}{-a^2+\sqrt{a^4-a^2 b^2+b^4}+b^2}+2 b^2-2 t\right) \\ &-a^2 b^2 \left(\sqrt{a^4-a^2 b^2+b^4}-4 t\right)-2 b^4 t, \end{aligned} \end{equation} which is a rather simple algebraic expression, which is actually $<0$ (still assuming $a>b>0$ and $t>0$). So, $H(t)<0$ at any critical point $t>0$ of $H$. Also, $H(0)=-a (a - b)^2 b (a^2 + a b + b^2)<0$ and $H(\infty-):=\lim_{t\to\infty}H(t)=-\infty$. So, $H(t)<0$ for all real $t\ge0$.

So, $R_2$ is decreasing on $(0,x_{a,b})$ and on $(x_{a,b},\infty)$.

Also, $R_1>0$ on $(0,x_{a,b})$, $R_1(x_{a,b})=0$, $R_1<0$ on $(x_{a,b},\infty)$, $R_1(0+)=\infty$, and $R_1(\infty-)=-\infty$.

So, by Table 1.1, $R_1$ is decreasing on $(0,\infty)$. Also, $R'(0)=-\frac2{ab}<0$ and $R(\infty-)=-\infty$. So, again by Table 1.1, $R$ is decreasing on $(0,\infty)$. $\quad\Box$

$\endgroup$
1
  • $\begingroup$ Thank you very much for this very clever answer @Iosif Pinelis :) $\endgroup$
    – NancyBoy
    Commented Aug 21, 2023 at 7:28

You must log in to answer this question.

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