Let $f_n : \mathbb{R}^p \to \mathbb{R}$ such that the $f_n$ are $C^1$ and such that the sequence $(f_n)_{n \in \mathbb{N}}$ converges uniformly to a function $f : \mathbb{R}^p \to \mathbb{R}$ which is $C^1$. Then prove that for all $x \in \mathbb{R}^n$ there is a sequence $(x_n)_{n \in \mathbb{N}}$ which converge to $x$ such that $\mathrm{d}f_n(x_n)$ converges to $\mathrm{d}f(x)$.

I must say that I don't know at all how to do and don't have any intuition of what is really going on here. So we might look at the case qhere $p= 1$.

So we can write :

$$f(a+h) = f(a)+ f'(a)h +o(h)$$ $$\forall n \in \mathbb{N}, f_n(a+h) = f_n(a) + f'_n(a)h +o(h)$$

Hence we have :

$$\mid f'(a)h - f'_n(a)h \mid \leq \mid f(a+h)-f(a) \mid +\mid f(a)-f_n(a) \mid + \mid o(h) \mid$$

Since the function $f_n$ converge uniformly to $f$, we have : $$\mid f'(a)h - f'_{\infty}(a) \mid \leq \mid o(h) \mid$$ And now using we let $h \to 0$ so that :

$$\mid f'(a) -f_\infty'(a) \mid = 0 $$

I don't know if this works, but it feels strange to me since in the case the sequence $x_n$ is just the constant sequence... and moreover if this is correct I don't see at all how to generalise to higher dimensions.

Thank you !

  • $\begingroup$ Your demonstration is not correct, since you don’t know if $lim_{n \to \infty} f’_n(a)$ converges, hence you can note it like $f’_{\infty}(a)$, that’s why you always need to be careful when dealing with double limit $\endgroup$ – Thinking Jan 28 '19 at 23:42
  • 3
    $\begingroup$ For complex or vector valued functions the claim is wrong. Consider the sequence $f_n(x):={1\over n}e^{inx}$, which converges uniformly to $0$, but $|f_n'(x)|=1$ for all $x$ and $n$. $\endgroup$ – Christian Blatter Jan 31 '19 at 10:34

We first prove the result in the case where there is a local maximum. At a maximum, the differential vanishes, so the claim is here:

Lemma. If $(g_n)_n$ is a sequence of $C^1$ functions $\mathbb{R}^p\to \mathbb{R}$ converging uniformly to a $C^1$ function $g$ having a local (strict) maximum at $y$, that is $g(x)<g(y)$ for all $x\neq y$ in a ball neigborhood $B(y,r)$ of $y$, then there exist a sequence $x_n$ such that $\lim x_n=y$ and $dg_n(x_n)=0$ for sufficiently large $n$.

Proof of the Lemma: Pick $N$ sufficiently large so that $\forall n\geq N$, $$sup_{ \|x-y\|=r} g_n(x)<g_n(y).$$ The existence of such an $N$ follows from the fact the corresponding inequality is true for $g$ by hypothesis, and the uniform convergence of $g_n$. For any $n\geq N$, pick $x_n$ to be a maximum of $g_n$ on $B(y,r)$. Because of the previous inequality, $x_n$ is in the interior of the ball $B(y,r)$. So the derivative satisfies $dg_n(x_n)=0$.

Let $x$ be a limit point of a subsequence of $(x_n)$. Since $g_n(x_n)\geq g_n(y)$ by definition of $x_n$, taking the limit we have $g(x)\geq g(y)$, and of course $x_n$ is still in the closed ball $B(y,r)$. So necessarily $x=y$ since $y$ is a local strict maximum of $g$ on $B(y,r)$. This concludes the proof of the Lemma.

Now, to deal with the general case, pick a point $y$ and define $$g_n(x)=f_n(x)-f(x)-\|x-y\|^2.$$ Clearly, this sequence of $C^1$ functions converges to $g(x)=-\|x-y\|^2.$ We now apply the Lemma, so there is a sequence $(x_n)_n$ such that $\lim x_n=y$, and $dg_n(x_n)=0$. But since $$dg_n(x_n).h=df_n(x_n).h-df(x_n).h-2\langle x_n-y, h \rangle,$$ so $$df_n(x_n)=df(x_n)+ 2\langle x_n-y, . \rangle.$$ The linear forms $h\mapsto 2\langle x_n-y, h \rangle$ converges to zero because of Cauchy-Scharwz inequality, and $df(x_n)$ converges to $df(y)$ since $f$ is $C^1$. This concludes the proof.

  • $\begingroup$ This is clever ! I am wondering why it's actually simpler to prove the problem when there is a local minima/maxima. I guess it's because in this case we know what is the value $\mathrm{d}f$ at this point. I'll deliver the bounty tonight. Thank you ! $\endgroup$ – dghkgfzyukz Jan 31 '19 at 18:59

In one dimension we don't need uniform convergence; pointwise converge will do. Also we only need $f$ and $f_n, n=1,2,\dots$ differentiable everywhere, not necessarily $C^1.$

WLOG we can assume $f\equiv 0$ because $f_n(x)-f(x)\to 0$ everywhere and $f_n-f$ is differentiable everywhere.

Fix $x\in \mathbb R.$ Let $\delta > 0.$ Then for $n\in \mathbb N$ the MVT shows there exists $c(n,\delta)\in (x,x+\delta)$ such that

$$f_n(x+\delta)- f_n(x) = f_n'(c(n,\delta))\delta.$$

Since the left side $\to 0$ as $n\to \infty,$ we can make the right side as small as we like by taking $n$ large. We can thus find $N = N_\delta$ such that $n\ge N_\delta$ implies $|f_n'(c(n,\delta))| < \delta.$

Now think of $\delta_k = 1/k, k=1,2,\dots .$ Then from the above there exist integers $0<N_1<N_2 < \cdots$ such that $N_k\le n < N_{k+1}$ implies $|f_n'(c(n,1/k))| < 1/k.$ If we then define

$$x_n = c(n,1/k),\,\, N_k\le n N_{k+1},$$

we have $x_n\to 0$ and $f'(x_n)\to 0.$ (We can let $x_n$ be anything for $1\le n <N_1.$)

  • $\begingroup$ You can't say WLOG $f = 0$ since at the beginning you are assuming $f$ is not necessarily $C^1$. $\endgroup$ – Thinking Feb 5 '19 at 19:02
  • $\begingroup$ @Thinking I don't understand your comment. $\endgroup$ – zhw. Feb 5 '19 at 21:45
  • $\begingroup$ It's not so clear for me that you are not loosing generality by saying $f = 0$, since $f$ is not necessarily $C^1$. $\endgroup$ – Thinking Feb 5 '19 at 22:18
  • $\begingroup$ @Thinking Like I wrote, I am assuming $f$ and $f_n, n=1,2,\dots$ are differentiable everywhere. Where do you think the proof goes wrong? $\endgroup$ – zhw. Feb 5 '19 at 22:22

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.