Consider the Hilbert Space of weighted-square-integrable functions f(x):

$$ \begin{equation} \int_{-\infty}^{\infty}\frac{f(x)^2}{e^{x}+e^{-x}}dx<\infty. \end{equation} $$

Note this integral is taken over the real line in both directions, and the weighting is roughly exponential. Hilbert Polynomials are complete using a different weighting, while Laguerre polynomials use a similar exponential weighting, but only on the positive real line.

What is a simple proof or reference that polynomials are complete in this space?

  • $\begingroup$ Do you suppose that finding some $g$ such that $g'(u) = \frac{1}{e^{g(u)}+e^{-g(u)}}$ and doing a u-substitution to get $\int_{-\infty}^{\infty}f(g(u))^2 du$ in the integral be a useful approach? $\endgroup$ – Milo Brandt Sep 28 '14 at 17:16
  • $\begingroup$ Do you want to know that/if/why the set $\Bbb{R}[x]$ of polynomials is dense in the given $L^2$ space, or do you want to have an explicit orthonormal basis consisting of polynomials? $\endgroup$ – PhoemueX Sep 28 '14 at 18:03
  • $\begingroup$ @phoemueX: I want to know that polynomials are dense in the space. Then I can easily construct an orthonormal polynomial basis using the Gram-Schmidt procedure. $\endgroup$ – Steven Heston Sep 28 '14 at 20:10

Disclaimer: The following proof is not completely elementary and uses methods from complex analysis as well as the Fourier transform, but at least it does the job:

Assume that the polynomials are not dense in your $L^{2}$-space. This implies the existence of $f\neq0$ in your $L^{2}$space for which $$ \int_{\mathbb{R}}\frac{f\left(x\right)\cdot x^{n}}{e^{x}+e^{-x}}\, dx=0\qquad\forall n\in\mathbb{N}_{0}. $$ Now consider the function $$ \Gamma:\left\{ s\in\mathbb{C}\mid-\frac{1}{2}<\text{Re}\left(s\right)<\frac{1}{2}\right\} \to\mathbb{C},s\mapsto\int_{\mathbb{R}}e^{sx}\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx. $$ Note that $\Gamma$ is well-defined because of \begin{eqnarray*} \int_{\mathbb{R}}\left|e^{sx}\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\right|\, dx & \leq & \sqrt{\int_{\mathbb{R}}\frac{\left|f\left(x\right)\right|^{2}}{e^{x}+e^{-x}}\, dx}\cdot\sqrt{\int_{\mathbb{R}}\frac{\left|e^{sx}\right|^{2}}{e^{x}+e^{-x}}\, dx}\\ & = & \sqrt{\int_{\mathbb{R}}\frac{\left|f\left(x\right)\right|^{2}}{e^{x}+e^{-x}}\, dx}\cdot\sqrt{\int_{\mathbb{R}}\frac{e^{2\text{Re}\left(sx\right)}}{e^{x}+e^{-x}}\, dx}\\ & \leq & \sqrt{\int_{\mathbb{R}}\frac{\left|f\left(x\right)\right|^{2}}{e^{x}+e^{-x}}\, dx}\cdot\sqrt{\int_{\mathbb{R}}e^{\left(2\left|\text{Re}\left(s\right)\right|-1\right)\left|x\right|}\, dx}<\infty. \end{eqnarray*} Also observe that we can bound the integrand on each strip $\left\{ s\in\mathbb{C}\mid-c<{\rm Re}\left(s\right)<c\right\} $, with $c\in\left(0,\frac{1}{2}\right)$, by $\frac{e^{2c\left|x\right|}\cdot\left|f\left(x\right)\right|}{e^{x}+e^{-x}}$ independently(!) of $s$ in that strip, so that standard arguments yield continuity of $\Gamma$.

Now, let $\gamma:\left[a,b\right]\to U$ be a closed (piecewise) $C^{1}$-curve in $U:=\left\{ s\in\mathbb{C}\mid-\frac{1}{2}<{\rm Re}\left(s\right)<\frac{1}{2}\right\} $. Then \begin{eqnarray*} \int_{\gamma}\Gamma\left(z\right)\, dz & = & \int_{a}^{b}\Gamma\left(\gamma\left(t\right)\right)\cdot\gamma'\left(t\right)\, dt\\ & = & \int_{a}^{b}\int_{\mathbb{R}}e^{\gamma\left(t\right)\cdot x}\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\cdot\gamma'\left(t\right)\, dx\, dt\\ & \overset{\left(\ast\right)}{=} & \int_{\mathbb{R}}\int_{a}^{b}e^{\gamma\left(t\right)\cdot x}\cdot\gamma'\left(t\right)\, dt\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx\\ & = & \int_{\mathbb{R}}\int_{\gamma}e^{xz}\, dz\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx\\ & = & 0. \end{eqnarray*} In the last step, we used that $z\mapsto e^{xz}$ is an entire function so that $\int_{\gamma}e^{xz}\, dz$ vanishes by Cauchy's integral theorem. The application of Fubini's theorem in $\left(\ast\right)$ is justified because of \begin{eqnarray*} \int_{\mathbb{R}}\int_{a}^{b}\left|e^{\gamma\left(t\right)\cdot x}\cdot\gamma'\left(t\right)\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\right|\, dt\, dx & = & \int_{a}^{b}\int_{\mathbb{R}}\left|e^{\gamma\left(t\right)\cdot x}\cdot\gamma'\left(t\right)\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\right|\, dx\, dt\\ & \leq & \left\Vert \gamma'\right\Vert _{\sup}\cdot\left(b-a\right)\cdot\int_{\mathbb{R}}\frac{e^{2c\left|x\right|}\left|f\left(x\right)\right|}{e^{x}+e^{-x}}\, dx, \end{eqnarray*} where $c>0$ is chosen with $\left|{\rm Re}\left(\gamma\left(t\right)\right)\right|\leq c<\frac{1}{2}$ for all $t\in\left[a,b\right]$ (such a $c$ exists, because $\gamma\left(\left[a,b\right]\right)\subset U$ is compact). Note that Tonelli's theorem is always applicable for non-negative (measurable) functions.

By Morera's theorem, we conclude that $\Gamma$ is holomorphic on $U$. But for $\left|s\right|<\frac{1}{4}$, we have $$ \sum_{n=0}^{\infty}\frac{\left|\left(sx\right)\right|^{n}}{n!}=e^{\left|sx\right|}\leq e^{\frac{1}{4}\left|x\right|} $$ with $\int_{\mathbb{R}}\frac{e^{\frac{1}{4}\left|x\right|}\left|f\left(x\right)\right|}{e^{x}+e^{-x}}\, dx<\infty$ (as above), so that dominated convergence yields \begin{eqnarray*} \Gamma\left(s\right) & = & \int_{\mathbb{R}}e^{sx}\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx\\ & = & \int_{\mathbb{R}}\lim_{n\to\infty}\sum_{\ell=0}^{n}\frac{\left(sx\right)^{n}}{n!}\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx\\ & = & \lim_{\ell\to\infty}\int_{\mathbb{R}}\left[\sum_{\ell=0}^{n}\frac{\left(sx\right)^{n}}{n!}\right]\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx=0. \end{eqnarray*} By the identity theorem for holomorphic functions, we conclude $\Gamma\equiv0$ and hence $$ 0=\Gamma\left(i\xi\right)=\int_{\mathbb{R}}e^{i\xi x}\cdot\frac{f\left(x\right)}{e^{x}+e^{-x}}\, dx, $$ so that the Fourier transform of the $L^{1}$-function(!) $x\mapsto\frac{f\left(x\right)}{e^{x}+e^{-x}}$ vanishes, which implies $\frac{f\left(x\right)}{e^{x}+e^{-x}}\equiv0$ (a.e.) and hence $f\equiv0$ (a.e.), a contradiction.

| cite | improve this answer | |
  • $\begingroup$ I just noted that one can probably replace the application of Morera's theorem by simply differentiating under the integral sign. $\endgroup$ – PhoemueX Sep 30 '14 at 10:36
  • $\begingroup$ Thanks, @PhoemueX. I had noticed that it is straightforward to show the Fourier transform is holomorphic in a neighborhood of the origin. This approach follows the Wikipedia proof for the Hermite case, except the Hermite Fourier transform is entire. $\endgroup$ – Steven Heston Sep 30 '14 at 18:23

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.