0
$\begingroup$

Assume h is a measure whose Fourier transform vanishes in an interval $[-\Omega,\Omega]$. I'm interested in obtaining inequalities of the form \begin{equation*} \int_{-\delta}^{+\delta}|h|(dt)\le C(\delta,\Omega)\int_{-\delta}^{+\delta}t^2|h|(dt) \end{equation*}

$\endgroup$
  • $\begingroup$ Where does this come from? $\endgroup$ – Igor Rivin Jul 16 '13 at 21:58
  • $\begingroup$ This comes up when trying to prove stability of recovery from low pass measurements using l1 minimization. In that case h would correspond to a signal in the null space of a low pass fourier matrix (operator). Please see the some what related paper below for use of l1 minimization for this purpose stats.stanford.edu/~donoho/Reports/Oldies/SRLS.pdf $\endgroup$ – mohi Jul 16 '13 at 22:40
1
$\begingroup$

If I understand this correctly--that you'd like an inequality like this to hold uniformly over all measures $dh$ whose Fourier transform is $0$ on $[-\Omega, \Omega]$--then you're unfortunately out of luck.

Here's one way of seeing that such an inequality can't be true. Let $g(\xi)$ be a $C^{\infty}$ function that's supported in the interval $[1, 2]$. Define $f(t)$ via $\hat{f} = g$. Note that $f$ is a Schwartz function, so for any $N > 0$ there's a constant $C_{N}$ such that $|f(t)| \leq C_{N} (1 + |t|)^{-N}$. Now introduce a rescaling parameter $r > 0$: let $g_{r}(\xi) = g(\xi / r)$, and define $f_{r}$ by $\hat{f}_{r} = g_{r}$, so that $f_{r}(t) = r f(rt)$. The measures (really just functions) I have in mind to examine are $dh_{r} = f_{r} \, dt$, letting $r \to \infty$.

Since $\widehat{dh}_{r} = g_{r}$ is supported in $[r, 2r]$, for all $r > \Omega$ the measure $dh_{r}$ will satisfy the criterion. The $dh_{r}$ all have the same mass, but as $r \to \infty$ the mass becomes concentrated closer and closer to the origin and so $$ \int_{-\delta}^{\delta} d|h_{r}| \to \| f \|_{1} $$ for any fixed $\delta$. On the other hand, $$ \int_{-\delta}^{\delta} t^{2} \, d|h_{r}| = \int_{-\delta}^{\delta} t^{2} r |f(rt)| \, dt \leq C_{N} \int_{-\delta}^{\delta} \frac{t^{2} r}{(1 + r|t|)^{N}} \, dt = C_{N} r^{-2} \int_{-r \delta}^{r \delta} \frac{s^{2}}{(1 + |s|)^{N}} \, ds, $$ from which we have $$ \int_{-\delta}^{\delta} t^{2} \, d|h_{r}| \to 0 $$ by taking $N > 3$.

Roughly speaking, for an inequality like the one you want to hold, you need some way to make sure the measures can't get concentrated near the origin, since the $t^{2}$ factor then causes problems. I haven't thought this through, but perhaps a condition that $\widehat{dh}$ have compact support, or perhaps at least some rate of decay, might give a result in this direction.

$\endgroup$

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.