3
$\begingroup$

I am looking for a function $\phi(x)$ such that

$\mathbb{E}_{x\sim\mathcal{N}(0,1)}[\phi(x)^2] = \mathbb{E}_{x\sim\mathcal{N}(0,1)}[\phi'(x)^2]$.

Obvious solutions are $\phi(x) = x$ and $\phi(x) = \exp(x)$. But do you know any other non-trivial solution?

$\endgroup$
5
  • $\begingroup$ Do you get something useful if you expand $\phi$ in terms of Hermite polynomials? $\endgroup$ Commented Jan 30, 2020 at 4:43
  • $\begingroup$ Interesting point! Can you explain why Hermite polynomials are relevant here? $\endgroup$ Commented Jan 30, 2020 at 9:30
  • 2
    $\begingroup$ They are an orthogonal basis for $L^2$ of standard Gaussian measure, or orthonormal after rescaling. So if $\{h_n\}$ are the orthonormal Hermite polynomials, we can write $\phi(x) = \sum_n a_n h_n(x)$, and $\mathbb{E}_{x\sim\mathcal{N}(0,1)}[\phi(x)^2] = \sum_n a_n^2$. Moreover we have $h_n' = \sqrt{n} h_{n-1}$, if I compute correctly, and so at least formally, $\mathbb{E}_{x\sim\mathcal{N}(0,1)}[\phi'(x)^2] = \sum_n n a_n^2$. Thus in order for your equation to hold, we need the coefficients to satisfy $\sum_n a_n^2 = \sum_n n a_n^2$. $\endgroup$ Commented Jan 30, 2020 at 14:08
  • 1
    $\begingroup$ Many polynomials qualify, e.g. any $a + bx + cx^2 + dx^3 + ex^4$ for which $a^2+2 a c+6 a e-c^2-18 c e-12 d^2-135 e^2=0$. $\endgroup$
    – user44143
    Commented Jan 30, 2020 at 14:43
  • 1
    $\begingroup$ Or (after Nate Eldridge's answer): many polynomials qualify, e.g. any $a+bx+cx^2+dx^3+ex^4$ for which $a^2+2ac+6ae-(c+9e)^2-12d^2-54e^2=0$. From this you can see that whenever $c,d,e$ are not all zero, this equation has both a positive root and a negative root for $a$. $\endgroup$
    – user44143
    Commented Jan 30, 2020 at 17:06

3 Answers 3

6
$\begingroup$

In fact, every reasonable function can be made into an example by adding an appropriate constant.

I'll write $Z$ for a standard Gaussian random variable. Recall the Gaussian Poincaré inequality:

Theorem. For every $f \in C^1(\mathbb{R})$ we have $\operatorname{Var}[f(Z)] \le E[f'(Z)^2]$.

Equivalently, this is the fact that the Ornstein-Uhlenbeck "number" operator has spectral gap equal to 1. Perhaps the simplest way to prove the Poincaré inequality is via Hermite polynomials; see Bogachev, Gaussian Measures, Theorem 1.6.4. The statement generalizes directly to absolutely continuous functions (in the appropriate Sobolev space over Gaussian measure)

Corollary. Let $f \in C^1(\mathbb{R})$ with $E[f(Z)^2], E[f'(Z)^2] < \infty$. There exist either one or two real numbers $c$ such that $\phi(x) := f(x) + c$ satisfies $E[\phi(Z)^2] = E[\phi'(Z)^2]$.

Proof. Set $$\begin{align*}\psi(c) &:= E[\phi(Z)^2] - E[\phi'(Z)^2] \\ &= \operatorname{Var}[\phi(Z)] + (E[f(Z)] + c)^2 - E[\phi'(Z)^2] \\ &= \operatorname{Var}[f(Z)] + (E[f(Z)] + c)^2 - E[f'(Z)^2]\end{align*}.$$ Now $\psi(c)$ is a quadratic in $c$ with $\psi(c) \to +\infty$ as $c \to \pm \infty$, and by the Poincaré inequality we have $\psi(-E[f(Z)]) \le 0$. So $\psi$ has either one or two real roots.

Indeed, the only way for the constant $c$ to be unique is if $f$ is a linear function, because that is the only case in which the Poincaré inequality saturates. Again, this can be seen via Hermite polynomials.

$\endgroup$
4
  • 2
    $\begingroup$ A good online reference seems to be ssk.im/blog/poincare-inequalities $\endgroup$
    – user44143
    Commented Jan 30, 2020 at 16:38
  • $\begingroup$ What do you mean by Poincaré inequality saturates? $\phi(x)=\exp(x)$ is also such a function where $c$ is unique, right? $\endgroup$ Commented Feb 1, 2020 at 2:09
  • 1
    $\begingroup$ @user3750444: I mean that we have $\operatorname{Var}[f(Z)] = E[f'(Z)^2]$ if and only if $f$ is linear. $\phi(x) = \exp(x)$ does not have a unique $c$; following the calculation above shows that $\phi(x) = \exp(x) - 2e^{1/2}$ also works. $\endgroup$ Commented Feb 1, 2020 at 2:45
  • $\begingroup$ @NateEldredge Thank you very much! I will cite your answer in my paper. $\endgroup$ Commented Apr 27, 2020 at 4:52
1
$\begingroup$

A family of such functions is given by $$\phi(x)=\phi_c(x):=e^{t_cx+cx^2}$$ for all real $x$, where $c\in[-\frac{1+\sqrt2}2,\frac{-1+\sqrt2}2]$ and $t_c=\pm\sqrt{1 - 8 c + 12 c^2 + 16 c^3}$; then both expected values equal ${e^{{2 t_c^2}/(1-4 c)}}/{\sqrt{1-4 c}}={e^{{2 - 8 c - 8 c^2}}}/{\sqrt{1-4 c}}$.

In particular, choosing here $c=0$, we get two members of this family, given by $\phi_0(x)=e^{\pm x}$ for all real $x$.

Another two members of this family are given by $\phi_c(x)=e^{cx^2}$ for $c\in\{-\frac{1+\sqrt2}2,\frac{-1+\sqrt2}2\}$ and all real $x$.

$\endgroup$
1
$\begingroup$

Another family of examples, now parametrized by an arbitrary function in a certain general class of functions:

Let $g\colon\mathbb R\to\mathbb R$ be any bounded continuously differentiable function with a bounded derivative such that $g(0)\ne0$ and $g'(u)^2\ge1$ if $|u|\le1$. Let us then show that a function $\phi$ of the form $f_c$ for some real $c>0$ will do, where $$f_c(x):=g(cx)$$ for all real $x$; that is, for some real $c>0$ we will have $$Eg(cZ)^2=c^2Eg'(cZ)^2, \tag{1}$$ where $Z\sim N(0,1)$.

Let $$D(c):=Eg(cZ)^2-c^2Eg'(cZ)^2.$$ Then, by dominated convergence, $D(c)$ is continuous in $c\ge0$. Also, $$D(0)=g(0)^2>0.$$

On the other hand, for $c\to\infty$ $$Eg'(cZ)^2\ge Eg'(cZ)^2 1_{|Z|\le1/c}\ge E1_{|Z|\le1/c}=P(|Z|\le1/c)\gtrsim\frac2{c\sqrt{2\pi}},$$ whereas $Eg(cZ)^2$ stays bounded (since $g$ is bounded). So, $D(c)\to-\infty$ as $c\to\infty$.

Therefore and because $D(c)$ is continuous in $c\ge0$, we see that $D(c)=0$ for some real $c>0$, which indeed yields (1).


Clearly, the condition $g'(u)^2\ge1$ if $|u|\le1$ here can be relaxed just to $g'(0)\ne0$.

E.g., we may take $g(x)=1+\sin x$ for all real $x$, and then $D(c)=\frac{3}{2}-\frac{1}{2} e^{-2 c^2}-e^{-c^2} c^2 \cosh \left(c^2\right)$, and the equation $D(c)=0$ has two roots, $c\approx\pm1.73$.

$\endgroup$

You must log in to answer this question.

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