Let $\langle X, Y \rangle$ be a separated dual pair of (real) vector spaces. A non-empty family $\mathcal{A}$ of non-empty subsets of $Y$ is called polar if

  1. every $A \in \mathcal{A}$ is bounded (w.r.t. to the duality)
  2. every point of $y$ is contained in some $A \in \mathcal{A}$: $\ \bigcup \{ A \mid A \in \mathcal{A} \} = Y$
  3. $\mathcal{A}$ is directed by inclusion: $\ \forall B, C \in \mathcal{A} \, \exists A \in \mathcal{A}: \, B \cup C \subseteq A$
  4. $\mathcal{A}$ is closed under scalar multiplication: $\ \forall \lambda \in \mathbb{R} \, \forall A \in \mathcal{A}: \, \lambda A \in \mathcal{A}$.

This is the definition as in Wikipedia and many other books in the literature. A polar family $\mathcal{A}$ on $Y$ induces a locally convex topology (the polar topology) on $X$, namely the topology of uniform convergence on $\mathcal{A}$. Due to property 2. this topology is Hausdorff. The weak topology $\sigma(X, Y)$ is the weakest polar topology ($\mathcal{A}$ consists of all finite sets) and the strong topology $\beta(X,Y)$ the strongest polar topology ($\mathcal{A}$ consists of all bounded sets).

Wilansky, "Modern Methods of Topological Vector Spaces" uses a slightly more general definition which uses 1., 3. and

4'. $\mathcal{A}$ is directed under doubling: $\ \forall A \in \mathcal{A} \, \exists B \in \mathcal{A}: 2A \subseteq B$.

So, Wilansky omits property 2. and weakens property 4. He then considers a locally convex topology on $X$ generated by the set of all polars $A^o$ where $A \in \mathcal{A}$. Let us call this a W-polar topology. He shows that this topology is just the topology of uniform convergence on $\mathcal{A}$. Every W-polar topology is weaker than $\beta(X,Y)$. A W-polar topology is then called admissible if it is also finer than $\sigma(X,Y)$. He then shows that a W-polar topology is admissible iff 2'. holds where

2'. $\bigcup \{ A^{oo} \mid A \in \mathcal{A} \} = Y$

is a weakening of property 2.

Question 1: Am I right to say that admissible topologies are just the usual polar topologies as defined above?

Question 2: I was surprised to see Exercise 8-5-105:

"Give an example of a separated nonadmissible W-polar topology."

So the topology in question must be weaker than $\sigma(X,Y)$ or equivalently not satisfying 2'., but still be Hausdorff and W-polar. Do you have some ideas?


This is an old question but perhaps others will benefit from it being answered anyways.

Question 1: Yes, the polar topology is unchanged by taking the saturated hull of $\mathcal A$ (that is, the collection $\{B | B \subseteq A^{\circ \circ} \text{ for some } A \in \mathcal A\}$). C.f. Jarchow Section 8.4, Prop. 1 or Bourbaki Chapter 3, Section 3, No. 1, Prop. 2 for instance. (In these references, for $E, F$ locally convex, the topology on $\mathscr L(E,F)$ of uniform convergence on some collection $\mathfrak S$ of bounded subsets of $E$ is discussed; this contains the situation of $\mathcal A$-convergence described in your question).

Question 2: Here is an example. Let $E= \mathcal D((0,1))$ with its usual inductive limit topology, and consider the duality $\langle E, E' \rangle$. Let $\mathcal B = \{ \{T_{\varphi}\} | \varphi \in E \}$ where $T_{\varphi}$ is the distribution defined by $\langle T_{\varphi}, \psi \rangle = \int \varphi \psi dm$, $\psi \in E$. Let $\mathcal A$ be the saturated hull of $\mathcal B$. Then the $\mathcal A$-topology $\tau_{\mathcal A}$ is a W-polar topology and $\tau_{\mathcal A}$ is just the topology of pointwise convergence in $\mathcal D$ (regarding $\mathcal D$ as a subspace of $\mathcal D'$), and this topology is Hausdorff because $\bigcup \mathcal A$ is $\sigma(\mathcal D', D)$-dense in $\mathcal D'$. But the identity map $(\mathcal D, \tau_{\mathcal A}) \to (\mathcal D, \sigma(\mathcal D, \mathcal D'))$ is not even sequentially continuous: choose test functions $\varphi_n \in E$ with $0 \leq \varphi_n \leq 1$ and which are identically 1 on $[\frac{1}{2n}, \frac{1}{n}]$ and supported in $[\frac{1}{3n}, \frac{2}{n}]$. Then, $\varphi_n \to 0$ in $\tau_{\mathcal A}$ but $\langle 1/x, \varphi_n \rangle = \int_0^1 \frac{1}{x} \varphi_n(x) dm(x) \geq \log 2$ for all $n$.

  • $\begingroup$ Very nice example. Just to recapitulate: $\varphi_n \to 0$ in $\tau_{\mathcal{A}}$ iff $\langle T_\varphi, \varphi_n \rangle = \int_0^1 \varphi(x) \varphi_n(x) \, dx \to 0$ for all $\varphi \in \mathcal{D}$, whereas $\varphi_n \to 0$ in $\sigma(\mathcal{D}, \mathcal{D}')$ iff $\langle T, \varphi_n \rangle \to 0$ for all $T \in \mathcal{D}'$. $T= "\frac{1}{x}"$ (i.e. $T \psi:= \int_0^1 \frac{1}{x} \psi(x) \, dx$) is a distribution that is not of the form $T = T_\varphi$ for some $\varphi \in \mathcal{D}$. So, weak convergence is strictly stronger than $\tau_\mathcal{A}$-convergence. $\endgroup$
    – yada
    Aug 12 '19 at 7:54

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.