0
$\begingroup$

Given scalars $c_1, c_2 > 0$, how would one go about solving, for non-expansive (i.e 1-Lipschitz) $\phi: \mathbb R \rightarrow (-\infty,+\infty]$, the following equation

$$ \begin{split} \mathbb E_{z \sim \mathcal N(0, 1)}[\phi(z)^2] &= c_1\\ \mathbb E_{z \sim \mathcal N(0, 1)}[\phi'(z)] &= c_2. \end{split} $$

Notation:

  • $\mathcal N(0, 1)$ is the standard Gaussian normal distribution so that $$\mathbb E_{z \sim \mathcal N(0, 1)} [g(z)] = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty \exp(-z^2/2)g(z)dz $$
$\endgroup$
  • $\begingroup$ If the person who downvoted this post could kindly add a comment, it would be very productive. Otherwise, such attitude makes MO look like youtube, where you can downvote a random video and disappear into the wild. Thanks. $\endgroup$ – dohmatob Apr 8 '18 at 11:13
3
$\begingroup$

Following approach is true when $ c \geq 2$.

Let $f(x) := a x $ be a linear function. Then, $\mathrm {prox}_{\lambda f}(z)= z-\lambda a$, and $$ \mathbb c=E_{z \sim \mathcal N(0, 1)}[(\operatorname{prox}_{\lambda f}(z))^2] =\mathbb E_z [(z-a)^2]=2+(\lambda a)^2 $$ So, $(\lambda a) ^2=c-2$, and $f (x) = (c-2)^{.5}x /\lambda^{.5}$ is a solution.

You can choose other model functions, instead of a linear function.

$\endgroup$
  • $\begingroup$ Thanks for the response (upvoted). I've added a second equation that constraints the solution a bit. Any ideas ? Thanks in advance. $\endgroup$ – dohmatob Apr 8 '18 at 11:02

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.