Let $\mathcal{H}$ the Hilbert transform defined by $$\mathcal{H}f(x)= p.v.\int_{-\infty}^{+\infty}\frac{f(x-y)}{y}dy.$$ We know that, for each $1<p<\infty$, it is true that $$||\mathcal{H}f||_{L^p}\leq C_p||f||_{L^p}$$ for some positive constant depending only of $p$.

My question is: Consider $0<\alpha<1$ and $||f||_{C^{\alpha}}$ is the $\alpha^{th}$-Holder norm, e.g, $$||f||_{C^{\alpha}(\Omega)}=\sup_{x\neq y\in\Omega}\frac{|f(x)-f(y)|}{|x-y|^{\alpha}}.$$ Is it true that $$||\mathcal{H}{f}||_{C^{\alpha}}\leq C||f||_{C^{\alpha}}?$$

Edit: I will accept answer of the David as correct. To the other readers, I ask you to read all the answers and comments to understand such a choice.

  • $\begingroup$ Yes, take $C=1$. $\endgroup$ – supinf Feb 11 at 12:03
  • $\begingroup$ adjusted, @supinf. $\endgroup$ – VVCM Feb 11 at 12:08
  • 1
    $\begingroup$ The $L^\infty\to L^\infty$ estimate fails, so there must be some explicit example of a sequence $f_n$ such that $\|Hf_n\|_{L^\infty}/\|f_n\|_{L^\infty}\to \infty$. I guess that the same counterexample would make the $C^\alpha \to C^\alpha$ estimate fail as well. Can you please check? I am interested. $\endgroup$ – Giuseppe Negro Feb 11 at 12:23
  • 2
    $\begingroup$ @GiuseppeNegro Your Idea does not work if the function $f_n$ is not continuous. And i think the heaviside function could be a counterexample for $L^\infty$, but i did not check it. $\endgroup$ – supinf Feb 11 at 14:18
  • 1
    $\begingroup$ @orange, $p.v.$ is the principal value. $\endgroup$ – VVCM Feb 11 at 19:53

Adding another answer because this really has little to do with my previous answer; instead it's a comment on how one might "interpret" things to reconcile the contradiction between my "yes" and supinf's "no".

We need to be a little careful. Let $$H_{\epsilon, A}f(x)=\int_{\epsilon<|y|<A}f(x-y)\frac{dy}y,$$so $$H=\lim_{\epsilon\to0,\\A\to\infty}H_{\epsilon, A}.$$

supinf gave a simple example of $f\in C^\alpha$ such that $H_{\epsilon, A}f(0)\to-\infty$. In fact his example has $Hf(x)=-\infty$ for every $x$, so if we want to talk about the Hilbert transform on $C^\alpha$ we need to modify the definition. Look at it this way:

Of course the $C^\alpha$ norm is just a seminorm. It's clear that $||f||=0$ if and only if $f$ is constant, so we do have a norm on the quotient space $X_\alpha=C^\alpha/\Bbb C$, consisting of $C^\alpha$ modulo constants.

When I said that $H$ was bounded on $C^\alpha$ I should have said it was bounded on $X_\alpha$. In that context we shouldn't expect pointwise convergence; instead we have this:

True Fact. If $f\in C^\alpha$ there exist $g\in C^\alpha$ and constants $c_{\epsilon,A}$ such that $H_{\epsilon, A}f(x)-c_{\epsilon, A}\to g(x)$ for every $x$.

I'm not going to show that $g\in C^\alpha$ here; that's contained in my previous answer. But I will show that the limit $g(x)$ exists (and is finite) for every $x$; this resolves the contradiction given by supinf: If he'd defined $Hf=g$ then he would not have obtained $Hf=-\infty$.

Define $$H=\int f(x-y)\frac{dy}y=\int_{-\infty}^{-1}+\int_{-1}^1+\int_1^\infty=H^{-}+H^0+H^+.$$

First, $H^0$ is no problem. If we say $H_\epsilon^0=\int_{\epsilon<|y|<1}$ then $$H^0f(x)-H_\epsilon^0f(x)=\int_0^\epsilon (f(x-y)-f(x+y)\frac{dy}y;$$since $f(x-y)-f(x+y)=O(y^\alpha)$ this shows that in fact $H_\epsilon^0 f\to H^0f$ uniformly.

Now say $H^+_A=\int_1^A$. We do need to subtract a constant $c_A^+$ to get this to converge. The obvious choice is $c_A^+=H_A^+f(0)$, since that certainly gives convergence for $x=0$. If I did the calculus correctly we have $$\begin{align}H_A^+f(x)-H_A^+(0)&=\int_{1-x}^1f(-y)\frac1{y-x}dy \\&+\int_{1}^{A-x}f(-y)\left(\frac1{y-x}-\frac1y\right)dy \\&-\int_{A-x}^Af(-y)\frac{dy}y.\end{align}$$The first integral on the RHS is independent of $A$ , while the second integral tends to somethhing finite as $A\to\infty$, since $f(y)=O(y^\alpha)$ and $1/(y+x)=1/y=O(1/y^2)$; similarly the tird integral tends to $0$.

Similarly if $H_A^-=\int_{-A}^{-1}$ there exists $c^-_A$ such that $H_A^--c_A^-$ is pointwise convergent; hence $H_{\epsilon,A}-(c^+_A+c^-_A)$ is pointwise convergent.

  • $\begingroup$ This is an issue that also appears for other function spaces; for example, I read in the book of Tao on nonlinear wave equations that negative order Sobolev spaces should be interpreted as "modulo polynomials". I never understood what he meant but now that you write this I think I get it. Thanks! $\endgroup$ – Giuseppe Negro Feb 12 at 18:15
  • 2
    $\begingroup$ Indeed. In any space where $f=\sum f_n$ as "above" the convergence is at best modulo polynomials. Because $\sum\widehat f_n$ can't see what $\hat f$ does at the origin; a distribution with Fourier transform supported at the origin is precisely a polynomial. Here the only polynomials in $C^\alpha$ are the constants. $\endgroup$ – David C. Ullrich Feb 12 at 19:53

I believe the answer is yes. Going to cheat, pulling out a big gun.

If $0<\alpha<1$ then $C^\alpha$ is in fact a Besov space; we have $f\in C^\alpha$ if and only if $$f=\sum_{n\in\Bbb Z} f_n,$$where $\widehat{f_n}$ is supported in the annulus $$A_n=\{\xi:2^{n-1}<|\xi|<2^{n+1}\}$$and $$2^{n\alpha}||f_n||_\infty\le c.$$This sort of decomposition of a Besov space often makes it trivial to determine whether a convolution operator is bounded.

Edit: No, it's not quite right to say $C^\alpha$ is a Besov space. What's actually a Besov space is the quotient $X_\alpha=C^\alpha/\Bbb C$, the space of $C^\alpha$ functions modulo constants. I was trying to avoid technical details, but given the other answer showing that taken literally $H$ does not map $C^\alpha$ to $C^\alpha$ it seems we cannot ignore the issue.

Here for example we're done if we can show that $$||Hf_n||_\infty\le c||f_n||_\infty,$$and that's actually true, even though $H$ is not bounded on $L^\infty$.

Recall that up to an irrelevant constant $$\widehat{Hf}(\xi)=sgn(\xi)\hat f(\xi).$$Choose a Schwarz function $\phi_0$ with $$\widehat\phi_0(\xi)=sgn(\xi)\quad(\xi\in A_0).$$If $\phi_n$ is an appropriate dilate of $\phi_0$ we have $$\widehat{\phi_n}(\xi)=sgn(\xi)\quad(\xi\in A_n)$$and $$||\phi_n||_1=||\phi_0||_1.$$Hence $$Hf_n=\phi_n*f_n,$$and hence $$||Hf_n||_\infty\le||\phi_n||_1||f_n||_\infty=||\phi_0||_1||f_n||_\infty.$$

(The inequality fails for $\alpha=1$; one explanation for why is that $Lip_1$ is not a Besov space...)

Why is that, I've been asked. First, $H$ is certainly not bounded on $L^\infty$; if $f=\chi_{(0,\infty)}$ then $Hf$ is not bounded.

Hence $H$ is not bounded on $Lip_1$. Because, in the sense of distributions, $f\in Lip_1$ if and only if $f'\in L^\infty$. It's clear, say from the Fourier transform, that $H$ commutes with the derivative. So $$||Hf||_{Lip_1}=||(Hf)'||_\infty=||H(f')||_\infty \not\le c||f'||_\infty=c||f||_{Lip_1}.$$

  • $\begingroup$ This is a great answer, very interesting. It would be even greater if you could give some details on why the estimate fails for $\alpha=0$ (the $L^\infty$ case) and $\alpha=1$. $\endgroup$ – Giuseppe Negro Feb 11 at 14:25
  • $\begingroup$ Was Rudin your advisor? $\endgroup$ – orange Feb 11 at 18:57
  • $\begingroup$ @orange Why yes. $\endgroup$ – David C. Ullrich Feb 12 at 14:15
  • $\begingroup$ @DavidC.Ullrich, can I make this Palley-Littlewood characterization for the $L^p$ spaces? $\endgroup$ – VVCM Feb 13 at 13:05
  • 1
    $\begingroup$ Yes, There is a Littlewood-Paley theory for $L^p$, but only for $1 < p < \infty$. Nice answer. The fact that the Hilbert transform is bounded between Hölder classes can also be proved without using that the Hilbert transform is a Fourier multiplier. It holds that Calderón-Zygmund operators are bounded over those classes. $\endgroup$ – Adrián González-Pérez Feb 13 at 21:23

It is not true.

We take the function $$ f(x) = \begin{cases} 1 &:& x\geq 1, \\ x &:& -1<x<1, \\ -1 &:& x\leq -1. \end{cases} $$ First, let us verify that $\|f\|_{C^\alpha} \leq 3$. Let $x,y\in\mathbb [-1,1]$ be given (the other cases for $x,y$ are not that interesting, since $f$ has the same value as $f(1)$ or $f(-1)$). Then $$ \frac{|f(x)-f(y)|}{|x-y|^\alpha} = |x-y|^{1-\alpha} \leq 1+|x-y| \leq 3. $$ Thus $f\in C^\alpha$, i.e. $\|f\|_{C^\alpha}$ is finite.

Calculating $\mathcal Hf$ at $x=0$ yields $$ \mathcal (Hf)(0) = p.v.\int_{-\infty}^\infty \frac{f(-y)}{y} \mathrm dy = \int_{-\infty}^{-1} \frac1y + p.v.\int_{-1}^1 (-1) \mathrm dy + \int_1^\infty \frac{-1}y \mathrm dy = -\infty -2 -\infty = -\infty. $$ Hence $\mathcal Hf$ is not in $C^\alpha$ because it is not continuous. Therefore $\|\mathcal Hf\|=\infty$.

  • $\begingroup$ I have a question, in what sense is $\mathcal H f$ defined, since $f$ is not a usual test function? $\endgroup$ – Calvin Khor Feb 11 at 14:55
  • $\begingroup$ The function $f$ must also decay at infinity for the inequality to hold. This decay is implicit in the Besov decomposition of David. $\endgroup$ – Giuseppe Negro Feb 11 at 14:57
  • $\begingroup$ @CalvinKhor I used the definition given by VVCM in the original question. $\endgroup$ – supinf Feb 11 at 15:02
  • $\begingroup$ Oh, the PV includes a cutoff at infinity. thanks $\endgroup$ – Calvin Khor Feb 11 at 15:04
  • $\begingroup$ @GiuseppeNegro It was not a requirement that the function decays at infinity. Even if the functions vanish at infinity, it might be possible to find a sequence that leads to a contradiction to the inequality conjectured in the original post. $\endgroup$ – supinf Feb 11 at 15:07

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.