Is there a continuous function $f:[0,1] \to \mathbb R$ such that $f(x) = 0$ uncountably often and, for every $x$ such that $f(x) = 0$, in any neighbourhood of $x$ there are $a$ and $b$ such that $f(a) > 0$ and $f(b) < 0$?


4 Answers 4


Let $K$ be the Cantor set; recall that $K$ is uncountable. Then the function $x\to d(x,K),$ where $d(x,K)$ denotes the distance from $x$ to $K,$ is continuous on $[0,1].$ Define

$$ f(x) = \begin{cases} 0, x \in K\\ d(x,K)\sin (1/d(x,K)), x \in [0,1]\setminus K. \end{cases}$$

Let $I_1, I_2, \dots $ be the open intervals thrown out in the construction of $K.$ In each $I_n,$ $f$ is continuous. Hence $f$ is continuous on $[0,1]\setminus K.$ If $x_n \to x \in K,$ then $d(x_n,K)\to 0,$ hence so does $f(x_n),$ which implies $f$ is continuous at $x.$ Thus $f$ is continuous on $[0,1].$

Now in any neighborhood of a point in $K,$ one of the $I_n$ will be contained in that neighborhood, hence $f$ will be both positive and negative in that neighborhood because of the $\sin (1/d(x,K))$ term.

Added: I just noticed that I omitted discussing points in $[0,1]\setminus K$ where $f=0.$ That happens at an $x$ such that $d(x,K)=1/(m\pi)$ for some $m\in \mathbb {N}.$ Because each $I_n$ has length $1/3^k$ for some $k,$ such an $x$ could not be the midpoint of any $I_n.$ So then $x$ is to the right or left of the midpoint of whatever $I_n$ it belongs to, and thus $d(\cdot,K)$ will be strictly increasing or decreasing in a neighborhood of $x,$ which implies a change in the sign of $f$ at that $x$ as desired.

  • $\begingroup$ Very clever! How'd you come up with that? $\endgroup$
    – Moya
    Aug 30, 2015 at 23:24
  • 12
    $\begingroup$ $x\sin(1/x)$ and its relatives are the gift that keeps on giving. $\endgroup$
    – zhw.
    Aug 30, 2015 at 23:28

In addition to the carefully constructed examples others have given, I just want to note that with probability $1$ one-dimensional Brownian motion starting at $0$ has these properties.


Absolutely. For example, we can proceed in a similar fashion to the definition of the Cantor function, using the "middle third construction of the Cantor set. In the first stage, we remove the "middle third" $(\frac13,\frac23)$ of $[0,1],$ and on this interval, we define $f(x)=(x-\frac12)^2+c_1$, where $c_1$ is the unique number such that $f(x)\to 0$ as $x\searrow\frac13$ and $x\nearrow\frac23.$ (I leave its determination to you.)

In the second stage, we remove the two "middle thirds" $(\frac19,\frac29)$ and $(\frac79,\frac89)$. On the first of these, we define $f(x)=-(x-\frac16)^2+c_2,$ where $c_2$ is the unique number so that $f(x)\to 0$ as $x\searrow\frac19$ and $x\nearrow\frac29.$ on the second, we define it by $f(x)=-(x-\frac56)^2+c_2,$ and we can show that $f(x)\to 0$ as $x\searrow\frac79$ and $x\nearrow\frac89.$ (I leave it to you to find $c_2$ and verify this.

We proceed in much the same manner. At the $n$th stage, we remove $2^{n-1}$ open intervals of the form $\left(m-\frac1{2\cdot3^n},m+\frac1{2\cdot3^n}\right),$ and there is some $c_n$ such that for any such "middle third" interval, we define the function on that interval by $f(x)=(-1)^{n-1}(x-m)^2+c_n,$ and we have the property that $f(x)\to0$ as $x\searrow m-\frac1{2\cdot3^n}$ and $x\nearrow m+\frac1{2\cdot3^n}.$

Finally, we define $f(x)=0$ on the Cantor set, itself. Then $f$ has exactly the properties you need, and is in fact differentiable everywhere but at the endpoints of the "middle third" intervals, I believe.



Our function will be $0$ on the Cantor set and nowhere else. For convenience we define a spike function $S(a, b)(x) = \mathbb{1}_{[a, b]}(x) \cdot \min\{|a-x|, |b-x|\}$ which is a continuous, positive, compact support function on the interval $[a, b]$. The idea will be to add a countable collection of spike functions to eventually satisfy your conditions.

So, lets approximate our function in a series of countable steps. Firstly, let $f_1 = S(1/3, 2/3)$. The greatest distance from a point with positive function value is now $1/3 = \delta_1$. Let $f_2 = f_1 - (S(1/9, 2/9) + S(7/9, 8/9))$. Now, the greatest distance from a point in $x\in [0, 1]$ with $f_2(x) = 0$ to a point with $y\in [0, 1]$ with $f_2(y)<0$ is now $1/9 = \delta_2$. Proceeding in this way, we would next add $4$ spikes to get $f_3 = f_2 + \sum_{i = 1, 7, 19, 25} S(i/27, (i+1)/27)$, and $\delta_3 = 1/27$, subtract $8$ spikes to get $f_4$

I claim that the limit of the $\{f_n\}$ exists, is well defined, and satisfies your conditions because the $\delta_n\to 0$.

Completed as a pennance for not reading the question properly.

  • 2
    $\begingroup$ The question doesn't ask for the function to be surjective, just that it takes real values. $\endgroup$ Aug 30, 2015 at 22:50
  • $\begingroup$ +1: Nice fix. That's pretty much the direction I was considering, initially, but I decided to try for more differentiability (for no particular reason). $\endgroup$ Aug 30, 2015 at 23:41
  • $\begingroup$ Its the natural solution, though Robert Israel has a nice example/class of examples too. $\endgroup$
    – user24142
    Aug 31, 2015 at 5:49

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.