We know that every linear open map between normed spaces is onto. This fact actually motivates the Open Mapping theorem which gives extra assumptions for converse to hold true. But I am unable to construct counterexample for the first fact, i.e. I am looking for a linear map between normed spaces which is onto but not open. Any hint for such map. Thanks.
3 Answers
Consider the identity $$I : (\ell^1, \|\cdot\|_1) \to (\ell^1, \|\cdot\|_\infty)$$ which is a continuous bijection since $\|\cdot\|_\infty \le \|\cdot\|_1$. However, it is not open.
Indeed, if $I$ were open, it would imply that $I^{-1}$ is bounded i.e. that $\|\cdot\|_1$ is bounded by $\|\cdot\|_\infty$, which is false.
- 46,490
-
I knew this fact that $|\cdot|1$ is not bounded by $|\cdot|\infty$ but I am not coming with proper logic now how to show this. Can you give some hint. – ogirkar Apr 29 '20 at 11:02
-
1@Believer For $n \in \Bbb{N}$ consider $x_n = (\underbrace{1, 1, \ldots, 1}n, 0, 0).$ We have $|x_n|_1 = n$ but $|x_n|\infty = 1$. – mechanodroid Apr 29 '20 at 14:24
Consider $\mathfrak{c}_{00} = \{\text{sequences } f: \mathbb{N} \to \mathbb{R} \text{ with } f(n) \neq 0 \text{ for finitely many } $n$ \}$, normed by supremum. A Hamel basis for $\mathfrak{c}_{00}$ is given by the singleton sequences $f_n : m \to \delta_{mn}$ for $n \in \mathbb{N}$.
Then consider the map $g: \mathfrak{c}_{00} \to \mathfrak{c}_{00}$ given by $g(f)(n) = \frac{1}{n} f(n)$. Then $g$ is linear, $||g|| = 1$ (and hence $g$ is continuous), and $g$ is onto as for any sequence $f: \mathbb{N} \to \mathbb{R}$, the sequence $\tilde{f} : n \to nf(n)$ is still an element of $\mathfrak{c}_{00}$ and satisfies $g(\tilde{f}) = f$. However, $g$ is not open, because for any $f$ with $||f|| < 1$, $$g(f)(n) = \frac{1}{n} f(n) < 1/n.$$ So $g(B(0,1))$ contains no neighborhood of $0$.
It's worth mentioning here that we can see exactly why this specific example fails in the completion of $\mathfrak{c}_{00}$. The completion of $\mathfrak{c}_0$ has an explicit description given by $\mathfrak{c}_0 = \{\text{sequences } f: \mathbb{N} \to \mathbb{R} \text{ so } \lim_{n \to \infty}f(n) = 0\}.$ If we define $g$ as above on $\mathfrak{c}_{0}$, $g$ is not onto because $(1, \frac{1}{2}, \frac{1}{3}, ...) \in \mathfrak{c}_0$, but if it were to happen that $g(f)(n) = \frac{1}{n}$ for all $n$, then $f(n)$ would need to be $1$ for all $n$. This is impossible for $f \in \mathfrak{c}_0$.
- 4,986
- 1
- 12
- 26
-
1
-
@Alex I couldn't figure out why $g(B(0,1))$ contains no nbhd of $0$. – Noob mathematician Apr 28 '20 at 22:11
-
2@Noobmathematician . A nbhd $B$ of $0$ covers$ B(0,r)$ for some $r>0,$ so $(r/2)f_n\in B$ for all $n.$ But if $1/n<r/2$ and $f\in B(0,1)$ then $g(f)\ne (r/2)f_n $ because $g(f)(n)<1/n<r/2=((r/2)f_n)(n).$ – DanielWainfleet Apr 28 '20 at 23:24
Let $V$ be the set of all $f:\Bbb N\to \Bbb R$ such that $\{n\in \Bbb N:f(n)\ne 0\}$ is finite.
For $f,g\in V$ let $\|f-g\|=\max_{n\in \Bbb N}|f(n)-g(n)|. $
And (as usual) for $f,g\in V$ and $n\in \Bbb N$ and $r\in \Bbb R$ let $(f+g)(n)=f(n)+g(n)$ and $(rf)(n)=r\cdot f(n).$
Let $M:V\to V$ where $(M(f))(n)=f(n)/n^2$ for $f\in V$ and $n\in \Bbb N.$
$M$ is linear ( & continuous ). $M$ is also a bijection, so if $M$ was an open map then $M^{-1}$ would be continuous. But $M^{-1}:V\to V$ is linear but unbounded so $M^{-1}$ cannot be continuous.
- 57,985