How can we compute the distributional derivative of the Weierstrass function $$W(x) =\sum_{k=1}^\infty \lambda^{(s-2)k}\sin(\lambda^k x)$$ where $s \in (0,2)$ and $\lambda$ are fixed parameters?

We know that the Weierstrass function is nowhere differentiable. This implies that it does not have a weak derivative. However, since $W \in L^1_{loc}$, we can consider the associated distribution $T_W$ and compute its distributional derivative.

I'm having troubles doing that computation because of the series representation of $W$.

  • $\begingroup$ Your question doesn't make sense as written. What are you actually trying to do, what tools do you have available, and what is the motivation? $\endgroup$ – user296602 Apr 6 '19 at 1:10
  • $\begingroup$ @T.Bongers Sorry, there was a typo (that is, I had mistakenly deleted the main verb of the sentence). $\endgroup$ – Riku Apr 6 '19 at 1:16
  • 1
    $\begingroup$ I'm glad you edited it to fix that, but it doesn't address 2 of the 3 issues in my comment; since this is not a do-my-homework site, please edit the question to actually improve it. $\endgroup$ – user296602 Apr 6 '19 at 1:16
  • $\begingroup$ @T.Bongers Done. $\endgroup$ – Riku Apr 6 '19 at 1:25
  • $\begingroup$ You have two parameters: $s$ and $\lambda$. How is $W$ a function just of $x$? $\endgroup$ – Ted Shifrin Apr 6 '19 at 1:26

For some $r \in (1/2,1)$ let

$$W(x) = \sum_{m=0}^\infty r^m e^{2i\pi 2^m x}$$

The series converges uniformly to a continuous function and for any $\phi \in C^\infty(\Bbb{R/Z})$ smooth $1$-periodic $$\langle W,\phi \rangle = \int_0^1 W(x)\phi(x)dx = \sum_{m=0}^\infty r^m \hat{\phi}(-2^m), \qquad \hat{\phi}(n) = \langle e^{-2i \pi n x},\phi \rangle$$ Which defines a distribution $W \in D'(\Bbb{R/Z})$.


$$W^{(k)}(x) = \lim_{M\to \infty}(2i\pi)^k\sum_{m=0}^M (2^k r)^m e^{2i\pi 2^m x}$$ (with the limit taken in the sense of distribution)

defines a distribution

$$\langle W^{(k)},\phi \rangle = (2i\pi)^k\sum_{m=0}^\infty (2^k r)^m \hat{\phi}(-2^m)$$ the series converges absolutely since $\hat{\phi}(n) = O(n^{-l})$ for every $l$.

  • $\begingroup$ Thank you. 1. Why can we rewrite $W(x)$ in that way for some $r \in (1/2, 1)$? 2. Why can we take $\phi \in C^\infty(\mathbb{R}/\mathbb{Z})$ instead of just $\phi \in C^\infty(\mathbb{R})$? 3. What did you deduce the decay of $\hat \phi$ from? 4. Can we rewrite the series in the last formula in a more explicit way? $\endgroup$ – Riku Apr 6 '19 at 1:31
  • $\begingroup$ If you want to consider $W$ as a distribution on $\Bbb{R}$ then $\langle W,\varphi \rangle_\Bbb{R} = \langle W, P \varphi \rangle_\Bbb{R/Z}$ where $P \varphi(x) = \sum_n \varphi(x+n)$, so it doesn't change anything (as $P$ is a continuous linear map $ C^\infty_c(\Bbb{R}) \to C^\infty(\Bbb{R/Z})$), equivalently replace the Fourier series coefficients of $\phi$ by the Fourier transform of $\varphi$. $\endgroup$ – reuns Apr 6 '19 at 1:39
  • $\begingroup$ The faster-than-polynomial decay of Fourier coefficients of smooth functions is well-known : $|\hat{\phi}(n) (2i\pi n)^l| \le \|\phi^{(l)}\|_{L^1(0,1)}$ $\endgroup$ – reuns Apr 6 '19 at 1:41
  • $\begingroup$ anything unclear ? $\endgroup$ – reuns Apr 7 '19 at 0:07
  • $\begingroup$ Thank you for your comments. Why is it true that there exists $r \in (1/2,1)$ such that $W(x) = \sum_{m=0}^\infty r^m e^{2i\pi 2^m x}$? Could you add some details on that? $\endgroup$ – Riku Apr 7 '19 at 13:07

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.