Theorem (Approximation of compact operators) Let $X,Y$ be Banach spaces and $F: X \supset M \to Y$, where $M \ne \emptyset$, be compact. Then for every $n \in \mathbb{N}$ there exists a continuous $F_n: M \to Y$ such that $\max_{v \in M} \| F_n v - F v \| \le \frac{1}{n}$.

Proof. Let $n \in \mathbb{N}$ and $(v_k)_{k = 1}^{m}$ be a finite $\frac{1}{n}$-net for $F(M)$. Then we have $$ \min_{j = 1}^{m} \| F u - v_j \| \le \frac{1}{n} \qquad \forall u \in M. $$ For $u \in M$, $n \in \mathbb{N}$ and $j \in \{1, \ldots, m\}$ let \begin{equation} a_j(u) := \max\left(0, \frac{1}{n} - \| F u - v_j \|\right) \ge 0, \qquad F_n u := \frac{\sum_{j = 1}^{m} a_j(u) v_j}{\sum_{j = 1}^{m} a_j(u)}, \end{equation} which are well-defined and continuous, as is $u \mapsto \| F u - v_j \|$.

For $v \in M$ we have \begin{align} \| F_n v - F v \| & = \left\| \left( \sum_{j = 1}^{m} a_j(v) \right)^{-1} \sum_{j = 1}^{m} a_j(v) v_j - F v \right\| \\ & = \left\| \left( \sum_{j = 1}^{m} a_j(v) \right)^{-1} \sum_{j = 1}^{m} a_j(v) (v_j - F v) \right\| \\ & \le \left( \sum_{j = 1}^{m} a_j(v) \right)^{-1} \left( \sum_{j = 1}^{m} a_j(v) \right) \| v_j - F v \| < \frac{1}{n} \end{align} My question In the last line there is a $v_j$ outside of any summations over $j$, so there must be at least a typo. But I still don't understand, we have that $\min_{j = 1}^{m} \| F v - v_j \| \le \frac{1}{n}$, so how can $\| v_j - F v \|< \frac{1}{n}$ for any $j \in \{1, \ldots, m\}$?

Here's a shorter formulation of the proof, which unfortunately doesn't clear up my misunderstanding.

  • $\begingroup$ If $\|Fv-v_j\|\ge 1/n$ then $\alpha_j(v)=0$. $\endgroup$ – Jochen Sep 26 at 13:37
  • $\begingroup$ @Jochen right. But what about the typo mentioned. What does actually belong there? $\endgroup$ – Viktor Glombik Sep 26 at 14:22
  • $\begingroup$ The closing parenthesis is missplaced. $\endgroup$ – Jochen Sep 26 at 16:17
  • $\begingroup$ Or the term in question has to be replaced by $\max_{j=1,..,m}\min(1/n,\|v_j-Fv\|)$, as individually $a_j(v)\|v_j-Fv\|=a_j(v)\min(1/n,\|v_j-Fv\|)\le a_j(v)/n$. Or just simply put the bound $1/n$ there. $\endgroup$ – Dr. Lutz Lehmann Sep 26 at 17:58
  • $\begingroup$ That depends on how much you want to argue. To get equality you would need that for any index $j$ with $\|Fu-v_j\|<\frac1n$ so that $a_j(u)>0$ you get $\|Fu-v_j\|=\frac1n$, which of course is absurd. However, it seems to be not as easy to find a general bound. While $ a_j(u)\|Fu-v_j\| = a_j(u)(\tfrac1n-a_j(u)) \le\frac1{4n^2}$ I do not see how to extract a bound on the second factor that is smaller than $\frac1n$. $\endgroup$ – Dr. Lutz Lehmann Sep 27 at 10:18

One variant to repair this and to obtain the strict inequality directly is to pick an index $j^*$ with $\|Fu-v_{j^*}\|=\min_j\|F_u-v_j\|<\frac1n$. Then single this index out in the summation \begin{align} \|Fu-F_nu\|\le... &\le\left(\sum_{j:a_j(u)>0}a_j(u)\right)^{-1}\left(\sum_{j:a_j(u)>0}a_j(u)\|Fu-v_j\|\right)\\ &\le\left(\sum_{j:a_j(u)>0}a_j(u)\right)^{-1}\left(a_{j^*}(u)\|Fu-v_{j^*}\|+\frac1n\sum_{j\ne j^*:a_j(u)>0}a_j(u)\right)\\ &=\frac1n-\left(\sum_{j:a_j(u)>0}a_j(u)\right)^{-1}a_{j^*}(u)\left(\frac1n-\|Fu-v_{j^*}\|\right) <\frac1n \end{align}


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.