2

Suppose we have a function $f : \mathbb{R}^n \to \mathbb{R}$, such that for all differentiable paths ${\bf p} : [-1, 1] \to \mathbb{R}^n$ with ${\bf p}(0) = {\bf 0}$, the function $(f\circ{\bf p})$ is differentiable at $0$.

This amazing answer shows that such a function $f$ that is not (Fréchet-)differentiable at ${\bf 0}$ exists. The example given lacks the condition that for a differentiable function at a point, the directional derivatives depend linearly on the direction.

Now I know that even if all directional derivatives exist at a point and they are a linear function of the direction, the function might still not be differentiable there.

My question is, what if we add the condition that for all the smooth* paths ${\bf p}$, $(f\circ{\bf p})'(0) = \nabla f({\bf 0})\cdot{\bf p}'(0)$?

Now, I can write a proof but I'm not sure this is right. This answer claims that for a function $f : \mathbb{R}^n \to \mathbb{R}$, the statement $\lim_{\bf h\to 0} f({\bf h}) = L$ is equivalent to $\lim_{t\to 0} f({\bf g}(t)) = L$ for all smooth paths ${\bf g} : [-1, 1] \to \mathbb{R}^n, \lim_{t\to 0} {\bf g}(t) = {\bf 0}$ (Actually the interval in that answer is $[0, 1]$ but I don't think that matters, and they only consider the case $n=2$, but it seems like it can be easily generalised).

Consider the function $F : \mathbb{R}^n \to \mathbb{R}$, $F({\bf h}) = \frac{|f({\bf h}) - f({\bf 0}) - \nabla f({\bf 0}) \cdot {\bf h}|}{|{\bf h}|}$. According to the argument I made before, if indeed $(f\circ{\bf p})'(0) = \nabla f({\bf 0})\cdot p'(0)$ for all smooth paths ${\bf p}$ with ${\bf p}(0) = {\bf 0}$, then, after a bit of fiddling with limits, $F({\bf p}(t)) \to 0$ as $t\to 0$, so $F({\bf h})\to 0$ as ${\bf h}\to{\bf 0} \iff f$ is differentiable at ${\bf 0}$.

*By smooth path I mean a path of class $C^n$ where $n$ is as high as needed. For this question, it seems like the bare minimum is that the path is differentiable or even differentiable at just at $0$.

  • I think it is correct. My intuition is: maybe you could consider radial coordinate, and use the path argument to show that F(p(t)) -> 0 as t ->0 "uniformly" within the neighborhood of 0. Then, the limit of F(h) under radial coordinate reduced to 1D problem (as long as the convergent speed is uniform). – WHLin Feb 05 '21 at 22:27
  • This question has already been answered here https://math.stackexchange.com/questions/2170970/differentiability-of-composition-with-every-path-implies-differentiability. – No-one Jan 19 '22 at 19:12

0 Answers0