5
$\begingroup$

The common example of a nonseparable Hilbert space comes from the collection of Besicovitch almost periodic function spaces. Starting with $L^p_{\text{loc}}(\mathbb{R})$ we look at those elements finite under the seminorm

$$ \|f\|_{M^p} = \limsup_{R\to\infty} \left(\frac1{2R}\int_{-R}^R \left|f(x)\right|^pdx\right)^{1/p} $$

Modding out (and completing?) by the zero elements nets us a Banach space for $1 \le p <\infty$, and a Hilbert space with the natural inner product for $p=2$. We'll call this space $M^p$.

We can see that $M^2$ is an example of a nonseparable Hilbert space because the collection $e^{i\xi x}$ is orthonormal for all $\xi \in \mathbb{R}$. We can look at the subspace $B^p\subseteq M^p$ of elements spanned by these functions, called the Besicovitch almost periodic functions.

We can see that $B^2\neq M^2$ since there are functions like

$$ f(x) = \left\{\begin{align}1 \ \ \ \ \ x\ge0 \\ -1 \ \ \ \ \ x < 0\end{align}\right. $$

which is orthogonal to all $e^{i\xi x}$, and $\|f\|_2 = 1$.

Question: I can't seem to find any discussion of $M^p$ independent from $B^p$. Is there a standard name for $M^p$? Is there a convenient description of an orthonormal basis for $M^2$?

$\endgroup$
  • 1
    $\begingroup$ I would not call $M^p$ functions (such as your $f$) "almost periodic" at all. $\endgroup$ – Gerald Edgar Mar 2 '14 at 13:13
  • 2
    $\begingroup$ Of course. The $B^p$ functions have all the nice properties, like plenty of $epsilon$-translation numbers, which make them almost periodic. The $M^p$ functions almost just seem like a nice place to house the $B^p$. $\endgroup$ – Greg Zitelli Mar 2 '14 at 15:14
  • $\begingroup$ Simple-minded question: why does the limit in your defining formula exist? or do you have to build that into your definition of $M^p$? $\endgroup$ – Yemon Choi Mar 2 '14 at 19:06
  • $\begingroup$ And if the limit exists for $f$ and $g$, does it follow that it also exists for $f+g$? Greg, you say it is a Banach space, but perhaps you should explain. $\endgroup$ – Gerald Edgar Mar 2 '14 at 19:19
  • 1
    $\begingroup$ The definition is careless. Although relying solely on a supremum might get me into trouble, I think everything will work properly with a limsup (Besicovitch uses these upper means in his own investigation of the almost periodics). Then we look at locally $p$-integrable functions finite under this seminorm (a real seminorm this time), which will be a Banach space after modding out and completing (if necessary). $\endgroup$ – Greg Zitelli Mar 2 '14 at 19:59
2
$\begingroup$

OK. Perhaps a reason $M^p$ is not often studied is: it is not even a vector space. (using the original defintion with lim not limsup.)

Define functions $f$ and $g$ as follows:
$f(x)=0$ if $x<1$,
$f(x)=1$ if $x \ge 1$ and $\{x\}< 1/2$; here, $\{x\} = x-\lfloor x\rfloor$ is the fractional part
$f(x)=-1$ if $x \ge 1$ and $\{x\} \ge 1/2$.

$g(x)=0$ if $x<1$,
$g(x)=f(x)$ if $x \ge 1$ and $\lfloor \log_2(x)\rfloor$ is even,
$g(x)=-f(x)$ if $x \ge 1$ and $\lfloor \log_2(x)\rfloor$ is odd.

Some graphs:

$f(x)$
f

$g(x)$
g

$f(x)+g(x)$
f+g

But note: $$ \lim_{R\to\infty} \left(\frac{1}{2R}\int_{-R}^R |f(x)|^2\right)^{1/2} = \lim_{R\to\infty} \left(\frac{1}{2R}\int_{-R}^R |g(x)|^2\right)^{1/2} = \frac{1}{\sqrt{2}} $$ both exist, while $$ \lim_{R\to\infty} \left(\frac{1}{2R}\int_{-R}^R |f(x)+g(x)|^2\right)^{1/2} $$ does not exist. In fact (do some computations): $$ \limsup_{R\to\infty} \left(\frac{1}{2R}\int_{-R}^R |f(x)+g(x)|^2\right)^{1/2} =\frac{2}{\sqrt{3}}, $$

$$ \liminf_{R\to\infty} \left(\frac{1}{2R}\int_{-R}^R |f(x)+g(x)|^2\right)^{1/2} =\frac{\sqrt{2}}{\sqrt{3}} $$

added
With the "limsup" definition, as suggested by Jean Van Schaftingen, we contradict the parallelogram law, since $$ \|f\|_{M^2} = \|g\|_{M^2} = \frac{1}{\sqrt{2}},\qquad \|f+g\|_{M^2} = \|f-g\|_{M^2} = \frac{2}{\sqrt{3}} $$

$\endgroup$
  • 1
    $\begingroup$ In the definition above, $M^p$ is defined by the limsup being finite, so I think that $M^p$ is indeed a vector space. However your example might show that the semi-norm induced by this limsup on $M^2$ is not Euclidean as it would not satisfy the parallelogram identity. $\endgroup$ – Jean Van Schaftingen Mar 3 '14 at 9:01
  • 5
    $\begingroup$ When questions are changed, answers may not longer apply... editing to say I answered the original question. $\endgroup$ – Gerald Edgar Mar 3 '14 at 14:44

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.