I learned this question from math.stackexchange, which is equivalent to ask that if $f:[0,1]\to \mathbb{R}$ is a continuous function with bounded variation, does $$g(x):=\lim_{\epsilon\to 0}\frac{f(x+\epsilon)-f(x-\epsilon)}{2\epsilon}$$ exist for every $x\in[0,1]$ imply that $f$ is absolutely continuous?

Moreover, if we do not know whether $f$ is of bounded variation or not, what can we say about the differentiability of $f$? For example, if $g\equiv 0$, will $f$ be a constant?

Any help is appreciated.

Edit: I think the original question cited from math.stackexchange has been solved by the asker himself there, which is based on the Vitali covering theorem for Radon measures on $\mathbb{R}^n$. For my own question, where $f$ is not assumed to be of bounded variation a priori, Jack Huizenga's answer is good enough for me.

  • $\begingroup$ The existence of $g$ for every $x \in [0,1]$ is saying that $f$ is differentiable everywhere, right? I mean, this is not exactly the definition, but close enough, maybe? $\endgroup$ – Daniel Spector Dec 10 '12 at 8:55
  • $\begingroup$ I cannot see why they are close enough, because when $f$ is differentiable everywhere and of bounded variation, I know $f$ is absolutely continuous. $\endgroup$ – qianzhang Dec 10 '12 at 9:36
  • $\begingroup$ @Daniel Spector : take for $f$ any even function, so that $g(0)=0$ is well-defined. Yet it happens that some even functions are not differentiable at 0. You can elaborate easily on this example to show that there exists continuous functions $f$ nowhere differentiable for which $g$ exists everywhere. You can also extend the construction of Alexander Eremenko given below: if you take for $f$ the indicator function of the rational numbers then $g$ is identically zero on the rational numbers, since both $x+\eps$ and $x-\eps$ are irrational or rationa at the smae time. $\endgroup$ – Loïc Teyssier Dec 10 '12 at 9:48
  • $\begingroup$ OK. Thanks for the differentiable correction :). So my thinking is the requirement that $f$ is $BV$ is enough to get rid of these kind of pathologies and regain the fundamental theorem, if $g$ is locally integrable (since the fundamental theorem implies absolute continuity). $\endgroup$ – Daniel Spector Dec 10 '12 at 11:46
  • 1
    $\begingroup$ You can show the following. If $g$ exists everywhere, then $f$ is absolutely continuous outside of a closed nowhere dense set. This is the best you can do in the general case, as given any closed nowehere dense set $S$ there exists a differentiable function which is absolutely continuous outside of $S$ but has infinite variation on any neighbourhood of each point of $S$. However, if $g$ exists everywhere and is locally integrable, then $f$ is absolutely continuous. You can prove this using the Baire category theorem. I don't have time to post a full answer right now, so just leaving a comment. $\endgroup$ – George Lowther Dec 13 '12 at 6:51

The ordinary proof that 0 derivative implies constant rests on the mean value theorem. Your function $g(x)$ is the "symmetric derivative" of $f$, so we should look for a "quasi-mean value theorem" for symmetric derivatives. A quick google search came up with the paper


Here it is shown (Theorem 2) that if $f$ is continuous on $[a,b]$ and symmetric differentiable on $(a,b)$ with symmetric derivative $f^s$, then there are points $\xi,\eta\in (a,b)$ with

$$f^s(\eta) \leq \frac{f(b)-f(a)}{b-a} \leq f^s(\xi).$$

Clearly then if $f^s = 0$ we must have that $f$ is constant.

  • $\begingroup$ Thank you for your answer. Do you have any idea about the absolute continuity of $f$? An immediate corollary of the "quasi-mean value theorem" is that if $f^s$ is bounded, then $f$ is Lipschitz, but I still have no idea about the general situation. $\endgroup$ – qianzhang Dec 10 '12 at 3:10
  1. For continuous $f$, there is a completely elementary proof that $g=0$ implies that $f$ is constant.

By uniform continuity, for every $\delta>0$ and for every $x$ there exists $\epsilon>0$ such that the oscillation of $f$ on $[x-\epsilon,x+\epsilon]$ is at most $2\epsilon\delta$. (Oscillation is $\sup$ minus $\inf$).

Now use the following Lemma: if a closed bounded set $E$ on the real line is covered by intervals such that every point is the center of some interval, then there exists a subcovering such that the intervals intersect at most by two. (No triple of intervals has non-empty intersection). From this we deduce that the total oscillation of $f$ is at most $4\delta$. As $\delta$ is arbitrary, this implies that $f$ is constant.

2 . For discontinuous $f$ the statement is not true. Let $E$ be a finite set on $(0,1)$, $f(x)=1$ for $x\in (0,1)\backslash E$, and arbitrary values on $E$. Then $g=0$.

  • $\begingroup$ Sorry, I cannot see why your statement in the second paragraph is true. $\endgroup$ – qianzhang Dec 10 '12 at 3:03

$f$ is a continuous function thus can be considered as a distribution. For $\phi\in C^1_c(0,1)$, we have $$ \langle f',\phi\rangle=-\int_{\mathbb R} f(x) \phi'(x) dx=\lim_{\epsilon \rightarrow 0} \int_{\mathbb R} f(x) \frac{\phi(x-\epsilon)-\phi(x+\epsilon)}{2\epsilon} dx, $$

$$=\lim_{\epsilon \rightarrow 0} \int_{\mathbb R} \phi(x) \frac{f(x+\epsilon)-f(x-\epsilon)}{2\epsilon} dx=\int \phi(x) g(x) dx, $$ if we know some domination for some $\epsilon_0>0$ $$ \sup_{0<\epsilon<\epsilon_0}\bigl\vert\frac{f(x+\epsilon)-f(x-\epsilon)}{2\epsilon}\bigr\vert \le h(x),\quad h\in L^1. $$ In that case, we get $f'=0$ in the distribution sense, so that $f$ is constant on the connected (0,1).

  • 1
    $\begingroup$ Bazin, thank you for your answer. I think the dominated convergence assumption is too strong for me, and I am only interested in pointwise convergence without further condition. I am sorry for not excluding this situation in my question. $\endgroup$ – qianzhang Dec 9 '12 at 19:24

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.