So, I have the following question:

Let $C[a,b]$ denote the space of continuous real-valued functions on $[a,b]$ with the sup-metric. Let $C^1[a,b]$ denote the space of continuously differentiable functions on $[a,b]$. Is $C^1[a,b]$ a closed subset of $C[a,b]$?

So, what I've done so far. The sup metric is $d_{\infty}(f,g) = \sup \left\{|f(x) - g(x)| : a \leq x \leq b \right\}$ for $f,g\in C[a,b]$

Clearly $C^1[a,b] \subset C[a,b]$

$C^1[a,b]$ is a closed subset of $C[a,b]$ if $C[a,b]\backslash C^1[a,b]$ is open in $C[a,b]$ - i.e.

$\forall h \in C[a,b]\backslash C^1[a,b]$, $\exists \epsilon > 0$ such that $\left\{ f \in C[a,b] : d_{\infty}(f,h) < \epsilon \right\} \subseteq C[a,b]\backslash C^1[a,b]$

Which seems to be saying that, essentially, for any non-differentiable continuous function $h$ we can find an $\epsilon > 0$ such that when the supremum of the distance between $h$ and all other continuous functions on $[a,b]$ is less than $\epsilon$, then they will also not be differentiable.

This is where I get stuck, not least because I'm not convinced whether that is true or not. Clearly, it's easily possible to create further non-differentiable functions by, for example, moving one point of $h$ less than $\epsilon$. My instinct is that you're not going to be able to turn a non-differentiable function into a differentiable one with an arbitrarily small $\epsilon$, so I think such an $\epsilon$ exists - but this obviously isn't formal (or even likely to be correct)



No -- on the contrary, $C^1[a,b]$ is dense in $C[a,b]$, by the Weierstrass approximation theorem. But that's too big a gun to bring out to show that it is simply not closed.

Instead, to follow your ansatz, Karolis' suggestion is to consider $h(x)=|x|$. That's the archetypical example of a function that's continuous but not differentiable, if $0$ is an interior point of $[a,b]$ (which we can clearly assume without loss of generality).

We're then looking for an $\epsilon$ such that no $f$ that differs from $h$ by at most $\epsilon$ can be differentiable. If we fail in that -- in other words, if for every $\epsilon$ there is a differentiable $f$ close to $h$ -- then we have proved that $C^1[a,b]$ is not closed.

Now consider $f(x)=\sqrt{x^2+\delta}$ where $\delta$ is a small positive constant that possibly depends on $\epsilon$. Is this always differentiable? Can we made it stay within a distance of $\epsilon$ from $|x|$ by choosing $\delta$ small enough?

  • $\begingroup$ On an interval $[a,b]$ you can make it stay within a distance of $\epsilon$ from $|x|$, but not if you considering the cases where $x$ goes to $\pm \infty$. $\endgroup$ – Noble. Oct 18 '13 at 18:20
  • $\begingroup$ @Noble: Fortunately the relevant metric here considers only differences in function values on $[a,b]$. But actually this argument does work for $C^1(\mathbb R)$ and $C(\mathbb R)$ too, because $\sqrt{x^2+\delta} - |x| \approx \frac{\delta}{2 |x|}$ for large $x$. $\endgroup$ – Henning Makholm Oct 18 '13 at 18:35
  • $\begingroup$ Right, and so because this $f(x)$ is differentiable, we've found a counterexample essentially, so it isn't closed? $\endgroup$ – Noble. Oct 18 '13 at 19:04
  • $\begingroup$ @Noble: Correct. $\endgroup$ – Henning Makholm Oct 18 '13 at 19:17

Try building a sequence of differentiable functions, converging to $f(x) = |x|$ (say $[a, b] = [-1, 1]$).

  • $\begingroup$ Sorry, I don't think I'm fully understanding the suggestion. $\endgroup$ – Noble. Oct 18 '13 at 16:54
  • $\begingroup$ A set is closed iff every convergent sequence of points in that set converges to a point in that set. $|x|$ is a function differentiable except for a single point. If you, say, replace that corner with a quarter of a circle, it becomes differentiable. If the radius of that circle is tiny, then by sup metric, the functions are very close. Thus you have a sequence. $\endgroup$ – Karolis Juodelė Oct 19 '13 at 7:10

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.