Let X be a standard Gaussian r.v.

Use Stein's characterization $Ef'(X) = E(Xf(X))$

to prove the Poincare inequality $E|f(X)-Ef(X)|^2 \leq E|f'(X)|^2$.

This looks like Markov's inequality could be used to prove the inequality.

However, I'm not quite sure how to piece it together.


You want to prove that $E|f(X)-Ef(X)|^2 \leq E|f'(X)|^2$. This inequality doesn't change if you add a constant to $f(x)$, so without loss of generality you can assume that $Ef(X)= 0$. So, all we need is that $E|f(X)|^2 \leq E|f'(X)|^2$.

To show that we will use Stein's method and define a function $g$ such that $g'(x) - x g(x) = f(x)$. This implies that $f'(x) = g''(x) - x g'(x) - g(x)$.

We will use the following two equalities derived by Stein's Lemma: $$E[ f'(X)g(X) ] + E[ f(X)g'(X) ] = E(f(X)g(X))' = E [ X f(X) g(X)] \,\,\,\,\,\, \textrm{(1)}$$ also $$E[ g'(X) g'(X) ] + E[ g''(X) g(X) ] = E(g'(X)g(X))' = E [ X g'(X) g(X)]\,\,\,\,\,\, \textrm{(2)}$$

So, we have that $E[ f(X)^2 ] = E[ f(X) f(X) ] = E[ (g'(X) - X g(X)) f(X) ] = -E[ f'(X)g(X) ]$, where the last equality follows by $(1)$.

Moreover, we also have that: $$E[ f(X)^2 ] = -E[ f'(X)g(X) ] = -E[ (g ''(X) - X g'(X) - g(X) ) g(X) ] = E[ g'(X)^2 + g(X)^2 ]$$ where we applied $(2)$ for the last equality. This implies that $E[ f(X)^2 ] \le E[ g(X)^2 ]$.

Finally, by the Cauchy-Swartz we have that: $$E[ f(X)^2 ] = -E[ f'(X)g(X) ] \le \sqrt{ E[ f'(X)^2 ] E[ g(X)^2 ] } \le \sqrt{ E[ f'(X)^2 ] E[ f(X)^2 ] }$$

This completes the proof of the Poincare Inequality.

  • $\begingroup$ It might be helpful to say how you know that such a function $g$ exists. Can you write it down explicitly, or are you invoking existence of solutions to ODEs? $\endgroup$ – Nate Eldredge Mar 12 '15 at 15:05
  • 3
    $\begingroup$ Stein has proved the existence of such a solution under mild conditions, e.g. absolute continuity. See lemma 1 in stat.duke.edu/courses/Fall10/sta205/lec/topics/dstn.pdf for a more precise statement. I assumed that such mild conditions hold in this case as It is impossible to prove the statement without any conditions since expectations might not even be defined. $\endgroup$ – thelionkingrafiki Mar 12 '15 at 15:15

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.