First, I need to give some definitions and background information:

Define $h(x)=|x|$ for $x\in [-1,1]$. Extend this function to $\mathbb R$ by defining $h(x+2) = h(x)$. Here is a graph of $h$:

enter image description here

Now if we define $g(x) = \sum_{n=0}^\infty {1 \over 2^n}h(2^nx)$ then $g$ is the Takagi function. One can prove that $g$ is continuous and nowhere differentiable. I did both. Continuity follows easily from the Weierstrass M-test and for non-differentiability one can first show it for dyadic points and then for non-dyadic points.

Now I am interested in two different modified versions of the Takagi function:

$$ g_1(x) = \sum_{n=0}^\infty {1\over 2^n}h(3^n x)$$ and $$ g_2(x) = \sum_{n=0}^\infty {1\over 3^n}h(2^n x)$$

These two are easily seen to be continuous.

Could someone please show me how to determine the differentiability of both $g_1$ and $g_2$? I really tried but failed.

  • $\begingroup$ I tried to make the question shorter so that maybe people are more likely to read it. You can see my work in the revision history. $\endgroup$
    – blue
    Jun 23, 2014 at 17:15
  • 2
    $\begingroup$ Although I'm rather convinced that $g_1$ is nowhere differentiable, I don't know how to prove it. For $g_2$, since $\sum \left(\frac{2}{3}\right)^n = 3 < \infty$, we see that that is Lipschitz, and hence by Rademacher's theorem differentiable almost everywhere. I expect $g_2$ is differentiable in every non-dyadic point, but haven't set out to prove it yet. $\endgroup$ Jun 25, 2014 at 9:26

2 Answers 2


I believe Daniel's conjecture, so let me do some work on $g_1(x)$. Let $c=2m/3^k$ for some integer $m$ and some positive integer $k$. I want to show that $g_1(x)$ is not differentiable at $x=c$. Define $\epsilon_a = 2/3^a$ for positive integers $a$. Let's compute $\frac{g_1(c+\epsilon_a)-g_1(c)}{\epsilon_a}$ and show the limit as $a\rightarrow \infty$ does not converge to a real number.

For any non-negative integer $n\geq a$, we have $(2)3^{n-a}$ is an even integer, so: \begin{eqnarray*} h(3^nc + 3^n\epsilon_a) - h(3^nc) &=& h(3^nc + (2)3^{n-a})-h(3^nc)\\ &=& h(3^{n}c) - h(3^nc) \\ &=& 0 \end{eqnarray*}

If $a>n\geq k$ then $2m3^{n-k}$ is an even integer, and the right-derivative of $h(x)$ at $x=2m3^{n-k}$ is $1$. Since $(2)3^{n-a}<1$, we have: \begin{eqnarray*} h(3^nc + 3^n\epsilon_a) - h(3^nc) &=& h(2m3^{n-k} + (2)3^{n-a})-h(2m3^{n-k})\\ &=& (2)3^{n-a} \end{eqnarray*}

In summary, assuming $a>k$:
\begin{eqnarray*} \frac{g_1(c+\epsilon_a)-g_1(c)}{\epsilon_a} &=& \frac{1}{\epsilon_a}\left[\sum_{n=0}^{k}\frac{1}{2^n}[h(3^nc+3^n\epsilon_a)-h(3^nc)] + \sum_{n=k+1}^{a-1}\frac{(2)3^{n-a}}{2^n}\right] \\ &\geq&\frac{1}{\epsilon_a}\left[-\sum_{n=0}^k\frac{3^n\epsilon_a}{2^n} + \sum_{n=k+1}^{a-1}\frac{(2)3^{n-a}}{2^n} \right] \\ &=& -\sum_{n=0}^k(3/2)^n + \sum_{n=k+1}^{a-1}(3/2)^n \end{eqnarray*} and the right-hand-side goes to $\infty$ as $a\rightarrow \infty$. So $g_1(x)$ is not differentiable at $x=c$.


Let $x \in \mathbb{R}$. For $n$, define $a_n,b_n \in 1/3^{n+1}.\mathbb{Z}$ as follow. If $x$ is in some interval $[\frac{2p}{3^n},\frac{2p+1}{3^n})$, put $a_n=\frac{2p}{3^n}$ and $b_n=\frac{2p+2/3}{3^n}$. If $x$ is in some interval $[\frac{2p+1}{3^n},\frac{2p+2}{3^n})$, put $a_n=\frac{2p+4/3}{3^n}$ and $b_n=\frac{2p+2}{3^n}$. In both cases $y \mapsto h(3^n y)$ is of constant slope in some interval of length $1/3^n$ containing $x,a_n,b_n$.

For $k \geq n+1$, $h(3^k a_n) =0$ and $h(3^k b_n)=0$.

For $k \leq n$, $\frac{h(3^k b_n)-h(3^k a_n)}{b_n-a_n} = \varepsilon_k(x).3^k$ where $\varepsilon_k(x)$ is the right-derivative of $h$ at $3^kx$.

Hence $$\frac{g_1(b_n)-g_1(a_n)}{b_n-a_n} = \sum_{k=0}^{n} \varepsilon_k(x) \frac{3^k}{2^k}.$$ So the sequence $(g_1(b_n)-g_1(a_n))/(b_n-a_n)$ does not converge.

Suppose that $g_1$ has derivative $\lambda$ at $x$. Write $\frac{g_1(b_n)-g_1(a_n)}{b_n-a_n}-\lambda = \left( \frac{g_1(b_n)-g_1(x)}{b_n-x} -\lambda \right) \frac{b_n-x}{b_n-a_n} + \left( \frac{g_1(x)-g_1(a_n)}{x-a_n} -\lambda\right) \frac{x-a_n}{b_n-a_n}$. This goes to zero because $\frac{b_n-x}{b_n-a_n}$ and $\frac{x-a_n}{b_n-a_n}$ is bounded.

For $g_2$. Let $x \in \mathbb{R}$. For $h>0$, write $$\frac{g_2(x+h)-g_1(x)}{h} = \sum_{k} g_k(h),$$ with $g_k(h) = \frac{h(2^k(x+h))-h(2^k x)}{3^kh}$. Since $h$ is $1$-Lipchitz, we have $\|g\|_{\infty} \leq (2/3)^k$. Moreover $\lim_{h \to 0} g_k(h) = (2/3)^k h^+(2^kx)$, where $h^+$ is the right-derivative of $h$. Hence $g_2$ has right-derivative $\sum_k (2/3)^k h^+(2^kx)$. Similary $g_2$ has left-derivative $\sum_k (2/3)^k h^-(2^kx)$. Thus $g$ has derivative at all non dyadic rationnals.

  • $\begingroup$ I notice that you posted a few hours after mine, but did not give me any comments. I believe you need to consider points $a_n$, $b_n$ of the form $2m/3^n$, as in my answer, to conclude that your sum can be truncated. Do you agree? $\endgroup$
    – Michael
    Jun 25, 2014 at 19:50
  • $\begingroup$ For example, you seem to claim that $h(3^kb_n)-h(3^ka_n)=0$ if $k>n$. But if $a_n=1/3^n, b_n=2/3^n$, and $k=n+1$, then $h(3^ka_n)=h(3)= 1 \neq h(3^kb_n)=h(6)=0$. $\endgroup$
    – Michael
    Jun 25, 2014 at 19:53
  • $\begingroup$ I also disagree that $h(3^kb_n)-h(3^ka_n) \in \{3^k, -3^k\}$ for all $k \in \{0, \ldots, n\}$. $\endgroup$
    – Michael
    Jun 25, 2014 at 20:15
  • $\begingroup$ You are right, $a_n$ should be of the form $2m/3^n$. For the other I meant $\frac{h(3^n b_n)-h(3^n a_n)}{b_n -a_n} = \pm 3^k$. $\endgroup$
    – user10676
    Jun 25, 2014 at 20:22
  • $\begingroup$ Why do you use the theorem for right-differentiable functions instead of differentiable functions? $\endgroup$
    – blue
    Jun 26, 2014 at 6:45

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.