I have some questions about the following theorem and it's proof.

Theorem. Let $X\subset \mathbb R^n$, open, convex, bounded and $f_n:X \to \mathbb R^m$ differentiable.And let's consider the following:

  1. There exist $H$: $\mathbb R^n\rightarrow L(\mathbb R^n,\mathbb R^m)$ such that $Df_n \rightarrow H$ uniformly in X, where $Df_n$ is the differential of $f_n$.
  2. There exist $x_0 \in $ X such that the sequence $\{f_n(x_0)\}$ converges.

Then there exist $f:X \rightarrow \mathbb R^m$ such that $f_n \rightarrow f$ uniformly over X, $f$ is differentiable and $Df(x)=H(x)$, for all $x\in X$.

Proof. Notice that \begin{align*} & \frac{f(x)-f(x_{0})-H(x_{0})\cdot(x-x_{0})}{\Vert x-x_{0\Vert}}% \color{fuchsia}=\frac{f(x)-f(x_{0})-[f_{n}(x)-f_{n}(x_{0})]}{\Vert x-x_{0\Vert}}\\ & +\frac{f_{n}(x)-f_{n}(x_{0})-\nabla f_{n}(x_{0})\cdot(x-x_{0})}{\Vert x-x_{0\Vert}}+\frac{(\nabla f_{n}(x_{0})-H(x_{0}))\cdot(x-x_{0})}{\Vert x-x_{0\Vert}}\\ & =:I+II+III. \end{align*} Since $X$ is convex, by applying the mean value theorem to the function $$ g_{n,m}(t)=f_{m}(tx+(1-t)x_{0})-f_{n}(tx+(1-t)x_{0}),\quad t\in\lbrack0,1] $$ $\color{fuchsia}{**...(2)**}$

there is $t_{0}$ such that \begin{align*} & f_{m}(x)-f_{m}(x_{0})-[f_{n}(x)-f_{n}(x_{0})]=g_{n,m}(1)-g_{n,m}(0)\\ & =g_{n,m}^{\prime}(t_{0})=(\nabla f_{m}(z_{0})-\nabla f_{n}(z_{0}% ))\cdot(x-x_{0}), \end{align*} where $z_{0}=t_{0}x+(1-t_{0})x_{0}$. By uniform convergence of the gradients, ...(1) $$ \Vert\nabla f_{m}(z)-\nabla f_{n}(z)\Vert\leq\Vert\nabla f_{m}(z)-H(z)\Vert +\Vert\nabla f_{n}(z)-H(z)\Vert\leq2\varepsilon $$ for all $n,m\geq n_{\varepsilon}$ and all $\color{fuchsia}z\in X$. Hence, by Cauchy's inequality \begin{align*} \left\vert \frac{f_{m}(x)-f_{m}(x_{0})-[f_{n}(x)-f_{n}(x_{0})]}{\Vert x-x_{0\Vert}}\right\vert & =\left\vert \frac{(\nabla f_{m}(z_{0})-\nabla f_{n}(z_{0}))\cdot(x-x_{0})}{\Vert x-x_{0\Vert}}\right\vert \\ & \leq\Vert\nabla f_{m}(z_{0})-\nabla f_{n}(z_{0})\Vert\leq2\varepsilon. \end{align*} Since $X$ is bounded, this inequality implies that \begin{align*} \vert f_{m}(x)-f_{n}(x)\vert\le|f_{m}(x_{0})-f_{n}(x_{0})|+\Vert x-x_{0}\Vert 2\varepsilon\le |f_{m}(x_{0})-f_{n}(x_{0})|+2M\varepsilon. \end{align*} and so $\{f_n\}$ is a uniform Cauchy sequence ...(2) and so it converges uniformly to a function $f$. Letting $m\rightarrow\infty$ $\color{fuchsia}{**...(3)**}$ we get $$ \left\vert \frac{f(x)-f(x_{0})-[f_{n}(x)-f_{n}(x_{0})]}{\Vert x-x_{0\Vert}% }\right\vert \leq2\varepsilon $$ for all $n \geq n_{\varepsilon}$. This takes care of $I$. Taking $n =n_{\varepsilon}\color{fuchsia}{**...(4)**} $ and using the fact that $f_{n_{\varepsilon}}$ is differentiable at $x_{0}$ we get that $$ \left\vert \frac{f_{n_{\varepsilon}}(x)-f_{n_{\varepsilon}}(x_{0})-\nabla f_{n_{\varepsilon}}(x_{0})\cdot(x-x_{0})}{\Vert x-x_{0\Vert}}\right\vert \leq\varepsilon $$ for all $x\in X$ with $0<\Vert x-x_{0}\Vert\leq\delta_{\varepsilon}$. This takes care of $II$.

Lastly, by Cauchy's inequality $$ \left\vert \frac{(\nabla f_{n}(x_{0})-H(x_{0}))\cdot(x-x_{0})}{\Vert x-x_{0\Vert}}\right\vert \leq\Vert\nabla f_{n}(x_{0})-H(x_{0})\Vert \leq\varepsilon $$ for all $n\geq n_{\varepsilon}$. In conclusion we have that for all $x\in A$ with $0<\Vert x-x_{0}\Vert\leq\delta_{\varepsilon}$, $$ \left\vert \frac{f(x)-f(x_{0})-H(x_{0})\cdot(x-x_{0})}{\Vert x-x_{0\Vert}% }\right\vert \leq4\varepsilon $$ which implies that $f$ is differentiable at $x_{0}$ with $\nabla f(x_{0})=H(x_{0})$.$\color{fuchsia}{**...(5)**}$ By repeating the proof with $x_0$ replaced by any other point, we get that $f$ is differentiable in $X$. ...(4)

My questions are in $\color{fuchsia}{pink}$ color.

  1. At the beginning why it is that $H(x_0) (x-x_0)= f_{n}(x)-f_{n}(x_{0})$?
  2. How is $g_{m,n}$ defined? I think its domain is $[0,1]$ but I don't know which codomain has.
  3. What is the form of $z$, is it a simple vector in X, or has the form of $z_0$?
  4. Why do we take $m\to\infty$? Could have been $n$? And why taking $m\to\infty$ implies the next inequality?
  5. Why do we take the $n=n_{\epsilon}$? I really don't see this step.
  6. Why do we need to repeat the proof?? Isn't $x_0$ arbitrary?
  7. How can the proof be formal? At the beginning, it is stated that it must be always $\varepsilon>0$, and maybe the $x,x_0\in X$ must be given too? Or they must be given in the middle of the proof? Or where should they be? Or how?

I tried to understand exactly which arguments are unclear to you. Let me know if I didn't cover something. $$ \newcommand{\R}{\Bbb{R}} $$

  1. It's not that $H(x_0)(x-x_0) = f_n(x)-f_n(x_0)$ (it is only true when passing to the limit $n \to \infty$), this step is a common algebraic trick (adding and substracting the same thing): $$ \frac{f(x)-f(x_0)-H(x_0)(x-x_0)}{\| x-x_0 \|} + \frac{f_n(x)-f_n(x_0) + \nabla f_n(x_0)(x-x_0)}{\|x-x_0\|} - \frac{f_n(x)-f_n(x_0) + \nabla f_n(x_0)(x-x_0)}{\|x-x_0\|} $$ Rearrange the terms conveniently and you get the above result.
  2. This is another common trick in multivariate calculus: by hypothesis, $X$ is a convex domain, i.e. contains (by definition) all the points of the segment $xy$ whenever $x, y \in X$. In this specific case, the points are $x$ and $x_0$. We now parametrize the segment between $x$ and $x_0$ in the following way: $$ \phi(t) = tx + (1-t)x_0, \quad t \in [0,1] $$ You can easily check this is actually such a parametrization. Then the proof goes on and defines $$ g_{n,m}(t) = f_m(tx+(1-t)x_0) - f_n(tx+(1-t)x_0), \quad t \in [0,1] $$ Which is a double sequence of functions. Each function has [0,1] as domain, and codomain in $\R^n$: this is because the $f_n$ themselves are defined to take values in a unspecified subset of $\R^n$, which by the way isn't actually relevant to the proof. Now notice that $$ f_m(x) - f_m(x_0) - [f_n(x)-f_n(x_0)] = f_m(x) - f_n(x) - (f_m(x_0) - f_n(x_0)) = g_{n,m}(1) - g_{n,m}(0) $$ We then apply the MVT to each of the $g_{n,m}$: it exists a $t_0 \in [0,1]$ such that $$ \frac{g_{n,m}(1) - g_{n,m}(0)}{1-0} = g_{n,m}(1) - g_{n,m}(0) = g'_{n,m}(t_0) $$ The $t$s where just parameters for the segment $xx_0$, thus to $t_0$ corresponds a point on that segment which we label $z_0$. If we compute $g'_{n,m}$ we get (using the chain rule) $$ g'_{n,m}(t) = (\nabla f_m(tx+(1-t)x_0) - \nabla f_n(tx+(1-t)x_0))(x-x_0) $$ That's why $g'_{n,m}(t_0) = (\nabla f_m(x) - \nabla f_n(x))(x-x_0)$.
  3. $z$ is taken in $X$, thus it is a vector. However it's arbitrary, while $z_0$ is not (it must have the property required by the MVT).
  4. The inequality here is just the same inequality before, but at the limit for $m \to \infty$. It is a property of the limit to preserve (weak) inequalities. We make $m$ go to infinity, and not $n$, because $m$ was actually an auxiliar index: we introduced it to use the uniform convergence of the gradients and prove the first inequality. Instead, $n$ is the index of the sequence $\{f_n\}_{n\in\Bbb{N}}$, thus we cannot dispense of it.
  5. Taking $n = n_\varepsilon$ it's just a way to guarantee that each estimate is cumulative on the previous: it could be, for example, that the second estimate is valid from $n_1 < n_\varepsilon$, thus between $n_1$ and $n_\varepsilon$ we lose the first estimate. We don't want this to happen. It's just a manner to say "for all sufficiently large values of $n$".
  6. I believe this to be a mistake. $x_0$ is fixed by the hypothesis (it's a point where the sequence $\{f_n(x_0)\}_{n\in\Bbb{N}}$ converges. It is $x$ which is arbitrary. By arbitrariety, it follows the proof is valid for any "choice" of $x \in X$, that is you could repeat the proof for any $x \in X$ and it will still be valid.
  7. The proof is sufficiently formal, in my opinion. See the previous point for a clarification on the role of $x$ and $x_0$ in the proof. The condition (which is implicit) $\varepsilon > 0$ descends from the standard $(\varepsilon, \delta)$-definition of limit.
| cite | improve this answer | |

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.