3
$\begingroup$

$ \newcommand{\bR}{\mathbb{R}} \newcommand{\diff}{\mathop{}\!\mathrm{d}} $ We fix $\alpha \in (0, 1)$ and $c>0$. Let $\ell : \bR^d \to \bR_+$ be a probability density function such that $$ \|\ell\|_{\infty} + \sup_{\substack{x, y \in {\bR}^d \\ x \neq y}} \frac{| \ell (x)-\ell (y)|}{|x-y|^\alpha} + \int_{\bR^d} |x| \ell (x) \diff x \le c. $$

Let $(p_t)_{t>0}$ be the standard Gaussian heat kernel on $\bR^d$, i.e., $$ p_t (x) := \frac{1}{(4 \pi t)^{\frac{d}{2}}} \exp \left ( -\frac{|x|^2}{4 t} \right ). $$

We denote by $*$ the convolution operation and let $\ell_t := \ell * p_t$.

Is there a constant $c_1>0$ (depending on $d, \alpha, c$) such that we have for $s, t\in [0, 1]$: $$ \int_{\bR^d} (1 + |x|) |\ell_t (x) - \ell_s (x)| \diff x \le c_1 |t-s|^{\frac{\alpha}{2}} \, ? $$

Thank you for your elaboration.

$\endgroup$
6
  • 1
    $\begingroup$ Akira, I hope you don't mind if I changed the term "mollification" with the term convolution: strictly speaking the former one is the convolution with a mollifier, i.e. the operator $p_t\ast(\cdot)$. $\endgroup$ Commented Jul 1 at 13:53
  • 1
    $\begingroup$ @DanieleTampieri Thank you for your edit. Indeed, I meant "convolution" rather than "mollification". $\endgroup$
    – Akira
    Commented Jul 1 at 13:54
  • 1
    $\begingroup$ Are you willing to assume the additional, stronger condition $\int_{\mathbb{R}^d} (1+|x|) \sqrt{\ell (x)} \mathrm{d} x < \infty$ as in your other question mathoverflow.net/questions/474098/… ? Then I think the answer is yes, otherwise no because there is no reason for the integral of $\frac{\vert \ell(x + tZ) - \ell(x) \vert)}{\vert t \vert^{\alpha}}$ under $(1+\vert x\vert) \mathrm{d}x$ to be $o(\vert t\vert^{\alpha/2})$ (compare with the situation of an function which is integrable but it's derivative is not) $\endgroup$
    – unwissen
    Commented Jul 1 at 15:15
  • $\begingroup$ @unwissen I am willing to assume that stronger condition. I hope you can elaborate on such situation in an answer. I am curious how the detail works out $\endgroup$
    – Akira
    Commented Jul 1 at 15:59
  • 1
    $\begingroup$ @unwissen I have asked a new question at mathoverflow.net/questions/474289/…. $\endgroup$
    – Akira
    Commented Jul 1 at 22:33

1 Answer 1

5
$\begingroup$

No, such a bound is not possible.

We will show a counterexample for $d = 1$. For a given $T \in \mathbb{Z}_{+}$, consider the following function $\ell^{(T)} : \mathbb{R} \to \mathbb{R}_{\geq 0}$: on the interval $x\in [-1,1]$ define $\ell(x) := (1 - |x|)$, on the interval $x\in[2\pi T, 2\pi(T+1)$ define $\ell(x) := (1 + \sin(x T))/T$, on $x \in [2\pi T - 1, 2 \pi T]$ take $\ell(x)$ to linearly interpolate between $0$ and $1/T$, similarly on the interval $x\in [2\pi (T+1), 2\pi (T+1) + 1]$ and finally $\ell(x) = 0$ elsewhere.

Here is the plot of the proposed function (not to scale). Plot of the proposed construction l(x)

Clearly $\int \ell(x) \mathrm{d} x \in [1, 10]$, similarly $\int |x| \ell(x) \mathrm{d}(x) \in [0.5, 10]$, and $\ell$ is $1$-Lipschitz.

Now, note that for small enough $t$, as long as $T \gg 1/t$, the difference $\ell_t(X) - \ell(X)$ is very close $\sin(xT)/T$ for $x \in [2 \pi T, 2\pi (T+1)]$, and close to zero everywhere else. This is because $\sin(T x)$ is an eigenfunction of a convolution with a Gaussian kernel $p_t$: if $f_T(x) := \sin(T x)$ then $f_T * p_t = \exp(-t^2 T^2) f_T$. Hence $\int |x| |\ell_t^{(T)}(x) - \ell^{(T)}(x)| \mathrm{d}{x} \approx \int_{2 \pi T}^{2\pi(T+1)} |x| |\sin(Tx)|/|T| \mathrm{d} x \geq \Omega(1)$.

For arbitrarily small $t$, we can chose large enough $T \gg 1/t$, such that the integral $\int |x| |\ell^{(T)}_t - \ell^{(T)}_t| \mathrm{d} x\geq C$, where $C$ is a universal constant.

$\endgroup$
2
  • $\begingroup$ Yes, thanks, it should interpolate between $0$ and $1/T$ there. I think it's supposed to be $sin(xT)/T$ on the interval $[2 \pi T, 2\pi (T+1)]$, though (the function is supposed have very high oscillations, so that convolution with gaussian of width much larger than wavelength averages it out to zero). $\endgroup$ Commented Jul 1 at 16:27
  • $\begingroup$ Yes, sorry, the other correction was nonsense, of course it has to be like this for the derivative to have high mass. Again, a very nice example. $\endgroup$
    – unwissen
    Commented Jul 1 at 16:48

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .