2
$\begingroup$

A random variable $X$ is "good" for $(a_0, b_0) \in (0,1)^2$ if its characteristic function $\varphi_X(t)$ satisfies the following constraints:

  1. $\forall t : \varphi_X(t) \geq 0$.
  2. $\varphi_X$ is monotonically decreasing on $\mathbb{R}_{+}$.
  3. $\varphi(a_0) = b_0$.

Given $(a_0, b_0)$ and $a_0 < x < 1$, I would like to find a random variable $X^*$ which is "good" for $(a_0, b_0)$ and whose characteristic function $\varphi_X$ has the lowest value at $x$, i.e. for any random variable $X'$ which is good for $(a_0, b_0)$ it holds that $\varphi_X(x) \leq \varphi_{X'}(x)$.

My conjecture is that a Gaussian satisfies this optimality criterion. My idea of proving it was thinking about the connections between $\varphi_X(t)$ and the entropy of $X$ ($h(X):= \mathbb{E}[-\log(X)]$): it is known that a Gaussian maximizes the entropy, so if it can be shown that $\varphi_X(t)$ decrease in a rate depends on the entropy, perhaps the claim would follow. However I couldn't obtain a more rigorous claim.

$\endgroup$
10
  • $\begingroup$ What do you mean by characteristic function? Usually it is $\varphi_X(t) = \mathbb{E}e^{itX}$? But then $\varphi_X(0) = 1 \geq |\varphi_X(t)|$ for any $t \in \mathbb{R}$. $\endgroup$ Aug 18, 2020 at 14:26
  • 1
    $\begingroup$ I mean exactly that, why doesn't it fit with what I wrote? $\endgroup$
    – Ernie
    Aug 18, 2020 at 15:17
  • $\begingroup$ What does it mean for a complex-valued function to be "monotone"? What does $\phi_X(x)\leq \phi_{X'}(x)$ mean? $\endgroup$ Aug 18, 2020 at 15:44
  • $\begingroup$ @JamesMartin: I think (1) is implicitly asking for the characteristic function to take only nonnegative real values, in which case everything else makes sense. $\endgroup$ Aug 18, 2020 at 16:25
  • $\begingroup$ ah ok right. And monotone on $\mathbb{R}_+$. $\endgroup$ Aug 18, 2020 at 16:41

1 Answer 1

2
$\begingroup$

Your minimization problem is clearly a problem of infinite-dimensional linear programming. Usually, such problems do not have a closed-form solution. Your problem, too, seems unlikely to have such a solution in full generality; however, see Remark 1 at the end of this answer.

One thing that can be definitely said is that your conjecture "that a Gaussian satisfies this optimality criterion" is false. Indeed, for any $(a_0,b_0)\in(0,1)^2$, let $c:=c_{a_0,b_0}:=-a_0^{-2}\ln b_0$, so that $c$ is the positive real root of the equation $\phi(a_0)=b_0$, where $\phi(t):=\phi_c(t):=e^{-ct^2}$. Then $\phi$ is a characteristic function and your conditions 1--3 are satisfied with $\phi$ in place of $\varphi=\varphi_X$.

Let now $\psi(t):=\psi_k(t):=\max(0,1-|t|/k)$ for all real $t$, where $k:=k_{a_0,b_0}:=a_0/(1-b_0)$ is the only root of the equation $1-a_0/k=b_0$. Then $\phi$ is a characteristic function and your conditions 1--3 are satisfied as well with $\psi$ in place of $\varphi=\varphi_X$.

However, $k>a_0$ and $\psi(k)=0<\phi(k)$. $\Box$

Remark 1: This also shows that for each $t\ge k=k_{a_0,b_0}$ the function $\psi_k$ is a minimizer of $\varphi(t)$ over all characteristic functions $\varphi$ satisfying your conditions 1--3.

Moreover, note that $t=a_0$ is a positive real root of the equation $\phi_c(t)=\psi_k(t)$. The latter equation may have one or two positive real roots. Let $a_{\max}$ be the largest root of the equation $\phi_c(t)=\psi_k(t)$. It follows that $a_{\max}<k$ and for each $t>a_{\max}$ we have $\psi_k(t)<\phi_c(t)$. In particular, if $a_{\max}=a_0$ (that is, if $a_0$ is the largest root of the equation $\phi_c(t)=\psi_k(t)$), then $a_0<k$ and for each $t\ge a_0$ we have $\psi_k(t)<\phi_c(t)$.

The condition $a_{\max}=a_0$ can be rewritten as $\psi'_k(a_0)\le\phi'_c(a_0)$ (with $c=c_{a_0,b_0}$ and $k=k_{a_0,b_0}$), and then further as $0<b\le b_*$, where $b_*=0.28466\dots$ is the only root $b\in(0,1)$ of the equation $1-b=2b\ln\frac1b$; the necessary and sufficient condition $0<b\le b_*$ for $a_{\max}=a_0$ does not involve the scaling parameter $a_0$ (and of course it should not).

For $a_0=1$, here are the graphs $\{(t,\phi_c(t))\colon t\in[0,k+1)\}$ (blue) and $\{(t,\psi_k(t))\colon t\in[0,k+1)\}$ (gold) (i) for $b=b_*$ and hence $a_{\max}=a_0[=1]$ (left); (ii) for $b=b_*-2/10$ and hence $a_{\max}=a_0[=1]$ (middle); and (ii) for $b=b_*+2/10$ and hence $a_{\max}>a_0[=1]$ (right):

enter image description here

$\endgroup$
7
  • $\begingroup$ Thank you so much. Do you think the solution to my problem can be approximated somehow? Can I even prove there is a solution? $\endgroup$
    – Ernie
    Aug 19, 2020 at 12:19
  • $\begingroup$ @Ernie : I have added that for $t\ge a_0/(1-b_0)$ there is the (rather trivial) minimizer. For $t\in(a_0,a_0/(1-b_0))$, the main problem is that there seems to be no good description of the set of all nonnegative characteristic functions -- see e.g. shorturl.at/twDJK and, in particular, shorturl.at/rDFPZ $\endgroup$ Aug 19, 2020 at 12:50
  • $\begingroup$ @ Iosif Pinelis thanks. Is there a trivial way to see why the Gaussian fails even for $t \in (a_0, a_0/1-b_0)$? Is there a way of saying something meaningful about the minimizer? $\endgroup$
    – Ernie
    Aug 19, 2020 at 12:58
  • $\begingroup$ May you please explain the part after the remark? thanks! $\endgroup$
    – Ernie
    Aug 19, 2020 at 13:45
  • $\begingroup$ @Ernie : As stated now in the remark, $\psi_k$ is better than $\phi_c$ for all $t\ge a_0$ in the case $b\le b_*=0.28\dots$. At this point, I cannot say anything definite for $a\in(a_0,a_{\max})$ and $b>b_*$. $\endgroup$ Aug 19, 2020 at 14:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.