20

I've been pondering over this problem for a while now, but I can't come up with a proof or even a useful approach...

Let $X$ be am infinite-dimensional normed vector space over $\mathbb{K}$ (that is either $\mathbb{R}$ or $\mathbb{C}$). Then the weak topology $\sigma(X,X^*)$ is not metrizable, i.e. there is no metric $d$ such that the induced topology of $d$ coincides with $\sigma(X,X^*)$.

Can anyone help me with this?

Amarus
  • 1,697

2 Answers2

24

The original answer is correct, and here I only want to elaborate every steps in details (to earn some contribution points if possible :)) Hopefully they are correct.

We claim:

(1) $\text{dim} X^* \geq \text{dim} X$, and equality holds iff $\text{dim} X < \infty$:

This is a well known theorem, e.g. one possible proof is on pages 244-248 of Jacobson's {Lectures in Abstract Algebra: II. Linear Algebra.}


(2) Dual space of $X^*$ of normed vector space $X$ is Banach with norm $||\cdot ||_{X^*}$:

Let $\{T_n\}\subset X^*$ a Cauchy sequence. Then for each fixed $x$, the sequence $\{T_nx\}\subset \Phi$ is a Cauchy sequence, which converges by completeness to some element of $\Phi$ denoted $Tx$. The map $x\mapsto Tx$ is linear; we have to check that it is continuous and that $\lVert T_n-T\lVert_{X^*}\to 0$.

We get $n_0$ such that if $n,m\geq n_0$ then for each $x$ $\lVert T_nx-T_mx\rVert_\Phi\leq\lVert x\rVert $ and letting $m\to+\infty$ we obtain $\lVert T_nx-Tx\rVert_\Phi \leq\lVert x\rVert $ so $\lVert Tx\rVert\leq \lVert x\rVert+ \lVert T_{n_0}\rVert\lVert x\rVert$ and $T$ is continuous.

Fix $\varepsilon>0$. We can find $N$ such that if $n,m\geq N$ and $x\in E$ then $\lVert T_nx-T_mx\rVert_\Phi\leq \varepsilon\lVert x\rVert$. Letting $m\to \infty$, we get for $n\geq N$ and $x\in X$ that $\lVert T_nx-Tx\rVert_\Phi\leq \varepsilon\lVert x\rVert$, and taking the supremum over the $x\neq 0$ we get for $n\geq N$ that $\lVert T-T_n\rVert_{X^*}\leq \varepsilon$.


(3) Every proper subspace of a normed vector space has empty interior.

We only need to show that the only subspace of a normed vector space $X$ that has a non-empty interior, is $X$ itself. Suppose subspace $S$ has a nonempty interior. Then it contains some ball $B(x,r) = \{y : \|y-x\| < r\}$. Now the idea is that every point of $V$ can be translated and rescaled to put it inside the ball $B(x,r)$. Namely, if $z \in V$, then set $y = x + \frac{r}{2 \|z\|} z$, so that $y \in B(x,r) \subset S$. Since $S$ is a subspace, we have $z = \frac{2 \|z\|}{r} (y-x) \in S$. So $S=V$.


(4) Finite-dimensional subspace of normed vector space is closed.

Suppose we have a convergent sequence $\{x_n\}$ such that $||x_n-x||\rightarrow 0$ and thus it is a Cauchy sequence as $||x_n-x_m||\leq ||x_n-x||+||x_m-x||$ . Since $\forall n, x_n=\sum_{i=1}^{K} \alpha_{n,i} x'_i, \alpha_{n,i} \in \Phi$, we know $\forall i, \{\alpha_{n,i} \}_n$ forms a Cauchy sequence in $\Phi$ as well and thus converge to some $\alpha_i \in \Phi$. Clearly, $x= \sum_{i=1}^{K} \alpha_{n,i} x' _i$ is inside the subspace.


By claim 2 we know $(X^*,||\cdot||_{X^*})$ is a Banach space. Suppose the Banach space $(X^*,||\cdot||_{X^*})$ has a countable basis $\{v_n; n\in\mathbb N\}$. Let us denote $X^*_n=[v_1,\dots,v_n]$, the linear span by the first n basis, which is a subspace as $\forall x,y\in X^*_n, \alpha,\beta \in \Phi, $ we know $\alpha x + \beta y = \sum_{i=1}^{n } (\alpha \lambda_{x,i}+ \lambda_{y,i}) v_i \in X^*_n$ where $x=\sum_{i=1}^{n }\lambda_{x,i}v_i, y=\sum_{i=1}^{n }\lambda_{y,i}v_i$. Then we have:

  • $X^*=\bigcup\limits_{n=1}^\infty X^*_n$
  • $X^*_n$ is a finite-dimensional subspace of $X^*$, hence it is closed by claim 4.
  • $X_n^*$ is a proper subspace of $X^*$, so it has empty interior by claim 3.

So we see that $ (\overline{X}^*_n)^\circ = X_n^\circ=\emptyset$, which means that $X^*_n$ is nowhere dense. So $X^*$ is a countable union of nowhere dense subsets and thus is of first category, which contradicts the Baire's category theorem. Therefore, we only need to show that $(X^*,||\cdot||_{X^*})$ has at most countable basis, which would then imply $\text{dim} X^* < \infty$ since otherwise we would have countable basis and it derives contradictions as shown above. Then by claim 1 we know $\text{dim} X < \infty$.


Now we show there exists a countable set $F\subset X^*$ such that every $ f\in X^*$ is a (finite) linear combination of elements in $F$.

Notice a collection of neighborhoods of the form $$B=\left\{x\in X: |f_i(x)|<1,\ f_i\in X^*, 1\leq i\leq K \right\} $$ forms local base of the weak topology, which means that any neighborhood of zero contains some neighborhood of this form.

Then suppose $\{A_\alpha \}_{\alpha\in\mathbb{N}}$ forms a countable local base of weak topology, we know that for each $A_\alpha$ there exists $$B_\alpha = \left\{x\in X: |f^\alpha_i(x)|<1,\ f^\alpha_i\in X^*, \ 1\leq i\leq K^\alpha\right\} \subset A_\alpha.$$ We claim that $$F= \bigcup_{\alpha\in\mathbb{N}} \left\{f^\alpha_i, 1\leq i\leq K^\alpha \right\} ,$$ which is countable, satisfies that every $f\in X^*$ is a (finite) linear combination of elements in $F$. Given $f\in X^*$ continuity gives us that $\{x\in X, |f(x)|<1 \}$ forms a zero neighborhood; so it contains some basic neighborhood $A_\alpha$. Thus $$B_\alpha \subset A_\alpha\subset \{x\in X,\ |f(x)|<1\}.$$ Suppose $f_1(x)=\cdots = f_{K^\alpha}(x)=0$. Then $|f(x)|<1$. The key observation is that we also have $f_1(nx)=\cdots = f_{K^\alpha}(nx)=0$ for all $n>0$. That is $|f(nx)|<1$ for all $n$, which implies $|f(x)|<1/n$ for all $n$ and so $f(x)=0$. This means that $$\bigcup_{j=1}^{K^\alpha}\ker f_j\subset \ker f.$$ This implies (see, for instance Lemma 3.9 in Rudin Functional Analysis) that $f$ is a linear combination of $f_1(x), \cdots, f_{K^\alpha}(x)$.

Martin Argerami
  • 205,756
  • I have edited the last few paragraphs to try and make the argument clearer. – Martin Argerami Dec 29 '21 at 11:50
  • Since any metric space is first countable and if weak topology is metrisizable then its first countable. But this is not true by this post: https://math.stackexchange.com/questions/1417295/the-weak-topology-on-an-infinite-dimensional-linear-space-is-not-first-countable This seems to be an alternative proof ? – Balaji sb Dec 19 '22 at 00:35
  • Is the statement that you cite regarding $dim(X^) \geq dim(X)$ for the algebraic dual $X^$, i.e. all linear functionals, or only those that are in addition continuous (i.e. those that one considers usually in functional Analysis settings) ? – 2000mg Haigo Feb 24 '23 at 15:27
18

Let $X$ be a normed space. You can show that if the weak topology of $X$ admits a countable base of open sets at $0$, then $X$ is finite dimensional:

  • Prove the existence of a countable set $\{\zeta_n\}$ in $X^*$ such that every $\zeta \in X^*$ is a finite linear combination of the $\zeta_n$.
  • Derive from this that $X^*$ is finite dimensional.
  • Deduce that $X$ is finite dimensional.
Seirios
  • 33,157
  • Thanks, I'll try to prove this right away. Still, how do I proceed to show that $\sigma(X,X^*)$ is not metrizable from there? I think I remember this criterion from Wikipedia, but I don't know any proof for it. Oh, and to be sure: by a countable base of open sets at $0$ you mean countable many sets such that every neighbourhood of $0$ contains one of them? Sadly, I don't know the English term for this. – Amarus Jun 20 '13 at 00:27
  • Any metrizable space is first-countable: ${B(x,1/n) \mid n \geq 1}$ is a countable neighbourhood basis. For exact definitions, you can see http://en.wikipedia.org/wiki/First-countable_space and http://en.wikipedia.org/wiki/Neighbourhood_system. – Seirios Jun 20 '13 at 07:10
  • Could you give some hints? I'm stucked in proving the first statement. – Xiang Yu Dec 11 '15 at 09:04
  • I wrote this sketch a long time ago, but if I am right, you may suppose without loss of generality that your countable basis has the form ${ \mathrm{ker}(\zeta_n) \mid n \geq 1 }$, just by definition of the weak topology. Then, use the fact that any $\zeta \in X^*$ is continuous with respect to the weak topology, and look at the link between $\mathrm{ker}(\zeta)$ and the $\mathrm{ker}(\zeta_n)$'s to deduce that $\zeta$ is a finite linear combination of the $\zeta_n$'s. – Seirios Dec 11 '15 at 20:40
  • @Seirios How to prove there exists a countable set ${\zeta_n}$ that is a basis for $X^$? Suppose that $U_n$ is a countable neighborhood base at the origin, then for every ${U_n}$, we can find an open set $V_n={x\in X:|\zeta_{n_1}(x)|<r_{n_1},\cdots,|\zeta_{n_k}(x)|<r_{n_k}}$ with $V_n\subset U_n$, where $\zeta_{n_1},\dots,\zeta_{n_k}\in X^$. I'm stuck here. – Xiang Yu Dec 12 '15 at 10:12
  • Seirios means you can prove it by contradiction. That is you can assume weak topology is metrizable, then weak topology admits a countable base of open set. – Lin Xuelei Feb 13 '19 at 10:26
  • 1
    For anyone reading this years later, it is not true that sets of the form ${\ker\zeta_n}$ form a local base for the weak topology at $0$. See Lingwei's answer for details on how the argument in the answer is implemented. – Martin Argerami Dec 29 '21 at 11:57