The Hilbert transform on the real Hilbert space $L^2(\mathbb R)$ is the singular integral operator $$ \mathcal H(f)(x) := \frac{1}{\pi} \int_{-\infty}^\infty \frac{1}{x-y} f(y) dy. $$

It satisfies $\mathcal H^2=-Id_{L^2(\mathbb R)}$, and in that sense, it is a complex structure on the Hilbert space $L^2(\mathbb R)$ of real-valued, square integrable functions on the real line.

I am wondering if there are other operators $\tilde {\mathcal H}:L^2(\mathbb R)\to L^2(\mathbb R)$ with similar properties.

Question: Does there exists a function $K:\mathbb R^2\to \mathbb R$ with the following properties:

  • The function $K(x,y)$ looks like $\frac{1}{x-y}$ in a neighborhood of the diagonal $x=y$
    (here, by "looks like", I mean for instance as "$K(x,y) = \frac{1}{x-y} +$ smooth function").

  • The singular integral operator $$ \tilde {\mathcal H}(f)(x) := \frac{1}{\pi} \int_{-\infty}^\infty K(x,y) f(y) dy. $$ satisfies $\tilde {\mathcal H}^2=-Id_{L^2(\mathbb R)}$, and thus defines a complex structure on $L^2(\mathbb R)$.

  • The function $K$ goes to zero faster than $\frac{1}{x-y}$ along the antidiagonals.
    Namely, it satisfies $$ \forall x\in \mathbb R,\qquad\qquad \lim_{t\to \infty}\;\;\;\; t\cdot K(t,x-t) = 0. $$

Variant: In case it turns out difficult to produce an example of an operator $\tilde {\mathcal H}:L^2(\mathbb R)\to L^2(\mathbb R)$ as above, I would be happy to replace $L^2(\mathbb R)$ by $L^2(\mathbb R;\mathbb R^n)$, the Hilbert space of $\mathbb R^n$-valued $L^2$ functions on the real line.

In that case, I would be looking for an integral kernel $$ K:\mathbb R^2\to \mathit{Mat}_{n\times n}(\mathbb R) $$ with all the properties listed above.

Right now, I actually believe that such an integral kernel does not exist, but this is purely a gut feeling...
If someone has any ideas about how to prove the non-existence of $\tilde {\mathcal H}:L^2(\mathbb R)\to L^2(\mathbb R)$ with the above properties, then I would very interested to hear them.

  • $\begingroup$ Could you explain the third condition? E.g. when $K=1/(x-y)$ one gets $t \cdot K(t,x-t) = t/(2t-x) $... $\endgroup$ Commented Apr 29, 2011 at 21:53
  • $\begingroup$ @:Piero. That's right. And $lim_{t\to\infty}t/(2t-x)$ is not zero. But if you have something like $K=1/(x-y)^\alpha$ with $\alpha>1$, then you get $lim_{t\to\infty}t/(2t-x)^\alpha=0$. $\endgroup$ Commented Apr 30, 2011 at 2:23
  • $\begingroup$ Well, if K=K(x-y) then the operator is translation invariant, then it is a constant coefficient pseudodifferential operator with a symbol $a(\xi)$, and by your second condition $a(\xi)$ must take only the values $\pm i$. Thus you have jump singularities in the symbol which should always produce a decay of order $\sim t^{-1}$. I guess. $\endgroup$ Commented Apr 30, 2011 at 13:54
  • $\begingroup$ @Piero: I do not require translation invariance. But I agree with you that it looks like it's not possible to achieve my decay condition. $\endgroup$ Commented Apr 30, 2011 at 14:31

2 Answers 2


EDIT: This solution does not satisfy the third condition, which rules out the Hilbert transform itself. So, this is an answer to different question. I do not delete it in hope it may be useful for someone.

Let $\phi(x)$ be a smooth monotone function such that $x-\phi(x)$ has compact support. This is a diffeomorphism of the real line and the pullback $\phi^*$ is a linear operator acting on $L^2(\mathbb R)$. The singular integral operator $$(\phi^*)^{-1}{\mathcal H}\phi^*$$ has all the properties you need.


It is possible also to reason thus. Too fast decay could lead to compactness of the operator if not in $L_2$ then in some suitably defined spaces. And this cannot be since its degree is identity and the last operator is compact only in finite dimension spaces.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.