6

Prove that there exists a positive constant $C$ such that $$\int \limits_{0}^{\infty} e^{-y}[u(y) - \langle u \rangle]^2 \mathrm{d}y \le C \int \limits_{0}^{\infty} e^{-y}[u'(y)]^2 \mathrm{d}y$$ for all $u \in C^{\infty}_c (\mathbb{R})$, where $$\langle u \rangle = \int \limits_{0}^{\infty} e^{-y}u(y) \mathrm{d}y$$


Attempt: Comparing the two sides, and seeing the limits of integration are the same, my goal is to show $$u(y) - \langle u \rangle \le Cu'(y)$$

I figured I should write $u(y)$ as an integral and follow the equations, but I got stuck with something that doesn't appear to make sense.

$$u(y) = u(0) + \int \limits_{0}^{\infty} u'(y) \mathrm{d}y$$

Then, $$u(y) - \langle u \rangle = u(0) + \int \limits_{0}^{\infty} u'(y) \mathrm{d}y - \int \limits_{0}^{\infty} e^{-y} u(y) \mathrm{d}y = u(0) + \int \limits_{0}^{\infty} u'(y)-e^{-y}u(y) \mathrm{d}y$$

Using the FTC we have that \begin{align*} u(y) - u(0) = \int \limits_{0}^{y} u'(s) \mathrm{d}s \implies & \lim \limits_{y \to \infty} u(y) = u(0) + \lim \limits_{y \to \infty} \int \limits_{0}^{y} u'(s) \mathrm{d}s \\ \implies & 0 = u(0) + \int \limits_{0}^{\infty} u'(s) \mathrm{d}s \\ \implies & u(0) = - \int \limits_{0}^{\infty} u'(s) \mathrm{d}s \end{align*} since $u \in C^{\infty}_{c}(\mathbb{R})$ means $u$ is compactly supported and so vanishes at infinity.

Substituting back in, we would then have that \begin{align*} u(y) - \langle u \rangle &= \int \limits_{0}^{\infty} -u'(y) + u'(y) - e^{-y} u(y) \mathrm{d}y \\ &= - \int \limits_{0}^{\infty} e^{-y} u(y) \mathrm{d}y = - \langle u \rangle \end{align*}

But this doesn't seem right, since it would imply that $u(y) = 0$.

Does anyone have any other approaches to this problem? Did I make a mistake in any of my assumptions?

Thanks in advance.

whpowell96
  • 7,849
JOlv
  • 237
  • 2
    It should be $u(y)=u(0)+\int_0^y u'(s)ds$ not $\infty$ in the upper integral bound. – Idividedbyzero May 15 '25 at 06:34
  • You're right, but taking the limit then as $y \to \infty$ of both sides still leads to the same place. – JOlv May 15 '25 at 07:13
  • @JOIv the $\lim_{y\to\infty}$ only tells you that $\lim_{y\to\infty}u(y) = 0$, which you already knew beforehand – Johannes Moerland May 15 '25 at 07:21
  • @JohannesMoerland Yes, the only reason I'm taking it is to be able to combine the integrals. For that, I need the limits of integration to be the same. – JOlv May 15 '25 at 07:26
  • Can you say anything about where this problem comes from? – Alex Ortiz May 15 '25 at 16:30
  • @JOlv the issue is for general $y \in \operatorname{Supp}(u)$ we get $u(0) + \int_0^\infty u'(s) , \mathrm d s = u(x) \neq u(y)$ where $x$ is the maximal element in the support (assuming the support is a compact interval). – Robertmg May 15 '25 at 17:14
  • @AlexOrtiz I wouldn’t be surprised if this came from a PDE or stochastic process textbook. It looks like a slightly nontrivial extension of a Poincaré result to an unbounded domain with a finite measure (exp weighting). – whpowell96 May 15 '25 at 18:58
  • @AlexOrtiz A professor gave it as part of a problem set, but he has a habit of making up problems without first checking that they work. I was wondering if there was something I wasn’t seeing or if it really does not work. – JOlv May 16 '25 at 00:32
  • @Robertmg By maximal element, do you mean the element in the support that maximizes $u$? – JOlv May 16 '25 at 00:41
  • I have found a few proofs of this claim that use some heavier functional analysis machinery such as the Rellich-Kondrachov theorem or some measure concentration results but that seems like severe overkill – whpowell96 May 16 '25 at 01:24
  • Related: https://math.stackexchange.com/questions/445053/weighted-poincare-inequality https://mathoverflow.net/questions/123924/sub-exponential-tail-implies-poincare-inequality – whpowell96 May 16 '25 at 01:29
  • Also, it’s worth noting that the integrands typically do not satisfy any nice pointwise inequalities. They are only obeyed when integrated over the entire space – whpowell96 May 16 '25 at 03:43

1 Answers1

4

Set $g = (f-f(0))^2$. Then \begin{align*} \int_0^\infty g(x)e^{-x}\,dx &= \int_0^\infty 2(f(x)-f(0))f'(x)e^{-x}\,dx \\&\leq \left(\int_0^\infty 4 g(x)e^{-x}\,dx\right)^{\frac{1}{2}}\left(\int_0^\infty |f'(x)|^2e^{-x}\,dx\right)^{\frac{1}{2}}\,. \end{align*} We deduce that $$ \int_0^\infty g(x)e^{-x}\,dx \leq 4\int_0^\infty |f'(x)|^2e^{-x}\,dx \,. $$ Thus, $$ \mathrm{Var}_\mu[f] = \inf_{c\in \mathbb R}\int_0^\infty (f(x)-f(c))^2e^{-x}\,dx \leq \int_0^\infty g(x)e^{-x}\,dx \leq 4\mathbf E_\mu[|f'(\xi)|^2]\,, $$ where $\mu$ is the one sided exponential measure on $\mathbb R$.

The paper "Poincaré’s inequalities and Talagrand’s concentration phenomenon for the exponential distribution" discusses in detail such inequalities for the exponential distribution.

The above inequality is a special case of the more general family of Poincaré inequalities, which take the form $\mathrm{Var}_\mu f \leq \mathcal E(f,f)$, where $\mathcal E(f,g) = -(f,\mathscr L g)_\mu$ is the Dirichlet form associated to a Markov semigroup $P_t$ with stationary measure $\mu$ and self-adjoint generator $\mathscr L$ such that $P_t f \to \mu(f)$ in $L^2(\mu)$ exponentially fast in $t$.

Andrew
  • 2,257
  • 1
  • 7
  • 13
  • Very nice! I am interested to know the course in which OP encountered this question because this trick to estimate integrals involving $\langle f\rangle$ and $f(0)$ would be hard to come up with without considering the probabilistic interpretation. I am curious if there is an even more elementary way to make that estimate that one might find in a PDE class, for instance. – whpowell96 May 16 '25 at 04:57