For $f:[0,1]\to\mathbb{R}$, let $f'_{app}(x)$ denote the approximate derivative (that is, the derivative calculated along some set with density $1$ at $x$, if such a thing exists). Assume that $f'_{app}$ exists and is continuous on $[0,1]$. Does this imply that $f'$ exists on $[0,1]$ (of course immediately $f'$ would also be continuous, since $f'_{app}=f'$ when $f'$ exists).

Put succinctly:

Does $C^1_{app}[0,1]=C^1[0,1]$?

  • $\begingroup$ I am not sure I understand what you call an approximate derivative. Doesn't the indicator function of $\mathbb{Q}$ have approximate derivative $0$? $\endgroup$ – abx Oct 20 '17 at 14:22
  • $\begingroup$ @abx No, at any rational point, the approximate derivative of the indicator function of $\mathbb{Q}$ does not exist. The difference quotient from a rational point $x_1$ to an irrational point $x_2$ approaches $\pm\infty$ as $x_2\to x_1$ from the right or left. $\endgroup$ – Trevor J Richards Oct 20 '17 at 14:47

Yes. Let $g$ be an antiderivative of $f'_{app}$, then $(f-g)'_{app}=0$ , this implies that $h:=f-g=\text{const}$.

Sketch of the proof of this fact. It suffices to prove that $h(x)=h(0)$ a.e. Indeed, if $h(a)\ne 0$, than this would imply $h'_{app}(a)=\infty$. We may fix $c>0$ and $\rho\in (0,1)$ and prove that the measure of $x\in [0,1]: h(x)-h(0)<cx$ is at least $\rho$. This, again, suffices. We say that a point $t\in [0,1]$ is nice if $\mu\{x\in [0,t]:h(x)-h(0)\geqslant cx\}\leqslant (1-\rho)t$, and $t$ is very nice if $t$ is a density point for the set of nice points. Let $a$ be a supremum of very nice points in $[0,1]$. If $a<1$, we get $h(a)-h(0)\leqslant ca$, else $h'_{app}$ would be infinite. Next, there exist a segment $[a,a+\varepsilon]$ such that $h(x)-h(a)<c(x-a)$ for more than $\rho\cdot \varepsilon$ points in this segment. This allows to find a very nice point greater than $a$. A contradiction.

| cite | improve this answer | |
  • 2
    $\begingroup$ You are using a (very plausible) result that if $h'_{app}=0$ on $[0,1]$, then $h$ is constant? I see a (slightly complicated) way to prove this using the Lebesgue density theorem, but is there a more or less trivial way to see this? Thanks $\endgroup$ – Trevor J Richards Oct 20 '17 at 14:51
  • 1
    $\begingroup$ The way I have in mind also uses Lebesgue density theorem. $\endgroup$ – Fedor Petrov Oct 20 '17 at 15:05

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.