2
$\begingroup$

This is a natural follow-up question related to one of my previous questions at here. Assume that $\rho$ is a log-concave probability density function with support $[0,\infty)$ and fixed mean $\mu >0$. Let $$H[\rho] = \iiint_{y,v,w\geq 0} \rho(v)\rho(w)\rho(y)\left(\frac{|v-y|+|w-y|}{2} - \left|\frac{v+w}{2} -y\right|\right)\,\mathrm{d}v\,\mathrm{d}w\,\mathrm{d}y \,\,(\geq 0) $$ and $$G[\rho] = \iint_{x,y\geq 0} \rho(x)\rho(y)|x-y|\,\mathrm{d}x\,\mathrm{d}y.$$ I am wondering if it is possible to obtain functional inequalities of the form $$H[\rho] \geq f(G[\rho])$$ for some non-negative function $f \colon [0,\infty) \to [0,\infty)$ with $f(0) = 0$. Of course, the best scenario I can hope for is for $f(x) = c\cdot x$ for some $c >0$, but this is might be too good to be true. Thanks for your help!


Remark: taking into account of the comment made by Iosif Pinelis, there is no hope for a non-trivial $f$ if we do not impose any "structural" restrictions on the pdf $\rho$ (such as log-concavity of the density).

$\endgroup$
2
$\begingroup$

$\newcommand{\R}{\mathbb R}$Let $p:=\rho$, $H:=H[p]$, $G:=G[p]$. Let us show that \begin{equation*} f(x)\equiv kx \end{equation*} with \begin{equation*} k=1/14334 \end{equation*} will do. We will not need the restriction that the support of the distribution is $[0,\infty)$.

By approximation, without loss of generality (wlog), $p(x)>0$ for all real $x$ -- for instance, one may approximate $p$ by its convolution $p*g$ with the density $g$ of a centered normal distribution with an arbitrarily small variance. Then $p*g>0$ on $\R$ and $p*g$ is arbitrarily close to $p$ and log concave, since $p$ and $g$ are log concave -- see e.g. this Wikipedia article.

Therefore and because $p$ is a log-concave density, $p$ is continuous and attains its maximum value, say $p_*[>0]$, at some point $c\in\R$, so that $p_*=p(c)\ge p(x)$ for all real $x$. Moreover, again because $p$ is a log-concave density, there exist (unique) real $a$ and $b$ such that \begin{equation*} \text{$a<c<b$ and $p(a)=p(b)=p_*/e$. } \end{equation*} Again by the log-concavity of $p$, \begin{equation*} p(x)\le q(x):= \left\{ \begin{aligned} q_1(x):=p_*\exp\Big\{-\frac{x-c}{a-c}\Big\} &\text{ if }x<a, \\ p_* &\text{ if }a\le x<b, \\ q_2(x):=p_*\exp\Big\{-\frac{x-c}{b-c}\Big\} &\text{ if }x\ge b. \end{aligned} \right. \end{equation*} As an illustration, for $p(x)=xe^{-x}1(x>0)$, here are the graphs $\{(x,p(x))\colon-2\le x\le6\}$ (blue), $\{(x,q(x))\colon-2\le x\le6\}$ (black), $\{(x,q_1(x))\colon a\le x\le c\}$ (dashed), and $\{(x,q_2(x))\colon c\le x\le b\}$ (dashed):

enter image description here

For this particular $p$, we have $c=1$, $a=-W_0\left(-1/e^2\right)=0.15859\dots$, and $b=-W_{-1}\left(-1/e^2\right)=3.1461\dots$, where $W_j$ is the $j$th branch of the Lambert $W$ function.

By shifting, wlog $$a=0.$$ So, \begin{align*} G&\le\iint\limits_{\R^2} q(x)q(y)|x-y|\,dx\,dy \\ &=p_*^2\frac{e^2 b^3+9 e b^3+3 b^3+3 b^2 c-12 e b^2 c-3 b c^2+12 e b c^2}{3 e^2} \\ &\le p_*^2\frac{\left(1+3 e+e^2/3\right) b^3}{e^2}, \end{align*} since $0<c<b$. Moreover, again by the log-concavity of $p$, we have $p\ge p_*/e$ on the interval $[a,b]=[0,b]$, so that $1=\int_\R p\ge\int_0^b p_*/e=bp_*/e$, whence $p_*\le e/b$ and \begin{equation*} G\le (1+3 e+e^2/3) b. \tag{1} \end{equation*}

On the other hand, because $p\ge p_*/e$ on the interval $[a,b]=[0,b]$ and the integrand in the definition of $H$ is $\ge0$, we have \begin{align*} H&\ge\Big(\frac{p_*}e\Big)^3\iiint\limits_{[0,b]^3} \Big(\frac{|x-z|+|y-z|}2 -\Big|\frac{x-z+y-z}2\Big|\Big)\,dx\,dy\,dz \\ &=\Big(\frac{p_*}e\Big)^3\frac{b^4}{24}. \end{align*} Also, $1=\int_\R p\le\int_\R q=p_*b(1+1/e)$, so that $p_*\ge1/(b(1+1/e))$ and hence \begin{equation*} H\ge\Big(\frac1{(e+1)b}\Big)^3\frac{b^4}{24}=\frac b{24(e+1)^3}. \tag{2} \end{equation*}

Comparing (1) and (2), we get \begin{equation*} H\ge\frac G{24(e+1)^3(1+3 e+e^2/3)}\ge\frac G{14334}, \end{equation*} as claimed.

$\endgroup$
13
  • $\begingroup$ Remarkable effort in writing this answer. Although I have one concern, in my problem, I really need to impose that the support of $\rho$ is contained in $[0,\infty)$ (i.e., no mass outside $\mathbb{R}_+$), so in that case, your sentence "By shifting, wlog, $a =0 $" causes me some trouble... $\endgroup$
    – Fei Cao
    Jun 30 '21 at 5:47
  • $\begingroup$ @FeiCao : Your case is just a particular case of the general one dealt with in this answer. So, I do not see any difficulty here. Also, you can do the same calculations with an arbitrary real $a$, without any shifting -- only the calculations will be a bit more complicated. $\endgroup$ Jun 30 '21 at 5:51
  • $\begingroup$ Thanks! Actually I checked your profile and found that you are a very good researcher in the area of probability, I might use (not very sure) this answer for a publishable research project, may I know whether you will be interested in co-author a paper? (I am merely a Ph.D student in applied math at ASU who just finished my fourth year) If you are interested, you can email me at fcao5@asu.edu, and I am very happy to share the potential research project behind it....Best regards. $\endgroup$
    – Fei Cao
    Jun 30 '21 at 6:17
  • 1
    $\begingroup$ @FeiCao : Concerning $G$, take $U=X-Y$ and $V_n=Z_n-W_n$, where $X,Y$ are iid random variables each with density $p$, and $Z_n,W_n$ are iid centered normal random variables each with variance $1/n$, say, such that $Z_n,W_n$ are independent of $X,Y$. Similarly, for $H$. $\endgroup$ Jun 30 '21 at 18:12
  • 1
    $\begingroup$ @FeiCao : I have added the description of these $a,b,c$. $\endgroup$ Jun 30 '21 at 18:52

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.