4
$\begingroup$

Let $\mathcal{M}_1(\mathbb R)$ denote the space of Borel probability measures on $\mathbb R$. The space is a Polish space (a space which admits a complete, separable, metric) using, say the Levy-Prokhorov metric. For $\mu \in \mathcal{M}_1(\mathbb R)$, let $L^1(\mu)$ denote the Banach space of $\mu$-integrable functions (mod $\mu$-null). Again, for each $\mu$, $L^1(\mu)$ is a Polish space.

Is the space $$\mathbb X = \left\{(\mu,f) \ \middle|\ \mu \in \mathcal{M}_1(\mathbb R),\ f\in L^1(\mu)\right\}$$ a Polish space?

Of course we need to specify a topology. I am assume the most natural topology. (I believe this is the topology generated by sets of the form $$\left\{(\nu,g) \in \mathbb X \ \middle|\ d(\mu,\nu)<\varepsilon \text{ and } \|g-f\|_{L^1(\nu)} < \varepsilon \right\}$$
where $f \in C_0(\mathbb R)$, $\mu \in \mathcal{M}_1(\mathbb R)$, $\varepsilon > 0$, and $d$ is a metric on $\mathcal{M}_1(\mathbb R)$ which makes it a Polish space.)

If $\mathbb X$ is a Polish space, is there a nice metric?

If $\mathbb X$ is not a Polish space, how nice is this space?

The motivation for my question is whether $\mathbb X$ is a computable metric space, so feel free to answer that question as well? :)


Terminology: Feel free to correct me if I am mistaken on the terminology. I originally called $\mathbb X$ a "coproduct" and wrote it as $$\coprod_{\mu \in \mathcal{M}_1(\mathbb R)} L^1(\mu),$$ but I was corrected in the comments. (However, the underlying set is a direct sum of sets, and I think similar notation $\sum_{x:A} B(x)$ is used in homotopy type theory to express the sum of dependent types.)

After doing some research I think I should say the map $\mathbb X \to \mathcal{M}_1(\mathbb R)$ given by $(\mu,f)\mapsto \mu$ is a fiber bundle where $\mathbb X$ is the total space and $\mathcal{M}_1(\mathbb R)$ is the base space. (Although, I could be mistaken.)

Topology: In case I gave the wrong basis for the topology above, I want the following maps to be continuous: $$(\mu,f)\mapsto \mu,\quad (\mu,f)\mapsto \|f\|_{L^1(\mu)},\quad (\mu,f)\mapsto \int fg\,d\mu$$ where $g$ is a bounded continuous function.

$\endgroup$
  • 1
    $\begingroup$ @SebastianGoette I changed my terminology. Hopefully it is (more) correct now. $\endgroup$ – Jason Rute Feb 29 '16 at 20:53
0
$\begingroup$

Yes, $\mathbb X$ is a Polish space, even a computable one, but it doesn't look like it has a nice metric.

(I figured out the answer on my own question, but any other insightful answers or references would still be welcome.)


Suppose $\mathbb A$ is a Polish space and suppose that $\mathbb B(a)$ is a parametrized family of Polish metric spaces, each with a metric $d_a$. Further assume that the $\mathbb B(a)$ are "uniformly presented" in that for each $a \in \mathbb A$ there is dense sequence of points $b_n(a) \in \mathbb B(a)$ such that the following map is continuous: $$a, m, n \mapsto d_a(b_m(a), b_n(a))\quad :\quad \mathbb A \times \mathbb N\times \mathbb N \to \mathbb R.\qquad(*)$$

(As a technical point, assume $d_a(\cdot,\cdot) \leq 1$ for all $a\in \mathbb A$.)

Our goal is to find the appropriate Polish topology on the dependent sum set $$\mathbb{X} = \sum_{a\in \mathbb A} \mathbb B(a) = \{(a,b) \mid a\in \mathbb A, b\in \mathbb B(a)\}.$$


The TL;DR version is that we embed $\mathbb B(a)$ (uniformly) into $[0,1]^\mathbb N$, associate $\mathbb{X}$ with the resulting $G_\delta$ (${\bf \Pi}^0_2$) subspace of $A \times [0,1]^\mathbb N$, and then apply Alexandrov's Theorem which says that all $G_\delta$ subsets of Polish spaces are Polish.


The appropriate topology on $\mathbb{X}$ is the weakest topology where

  1. The first projection is continuous, $$\pi_1(a,b)=a \quad :\quad \mathbb X \to \mathbb A,$$

  2. The distances to the dense set $b_n(a)$ are continuous: $$(a,b) \mapsto (d_a(b,b_n(a)))_{n\in\mathbb{N}} \quad :\quad \mathbb X \to \mathbb[0,1]^\mathbb{N}.$$

The main idea is we will replace $b$ with the more manageable object $(d_a(b,b_n(a)))_{n\in\mathbb{N}} \in \mathbb[0,1]^\mathbb{N}$. Let $f\colon \mathbb X \to \mathbb A \times [0,1]^\mathbb N$ be the injective map $$f(a,b) = (a,(d_a(b,b_n(a)))_{n\in\mathbb{N}}),$$ and give $\mathbb X$ the topology induced by this map. That is, $f$ is a homeomorphism from $\mathbb X$ to the range of $f$, which we will call $\mathbb Y$. This is weakest topology satisfying (1) and (2) above. Moreover, this topology has the following nice properties.

  1. The second projection is a homeomorphism for a fixed $a \in \mathbb A$, $$\pi_2(a,b)=b \quad :\quad \mathbb X_a := \{b \mid (a,b)\in \mathbb X\} \to \mathbb B(a),$$

  2. Our dense sequence $b_n(a)$ embeds continuously $$a,n \mapsto (a,b_n(a)) \quad :\quad \mathbb A \times \mathbb N \to \mathbb X.$$

  3. A subbasis is given by sets of the following form, $$\{(a,b)\in \mathbb X \mid d_\mathbb {A}(a,a_0)<\varepsilon \text{ and }d_a(b,b_n(a))<\varepsilon\}$$where $a_0 \in \mathbb A$ and $\varepsilon >0$.

Now, we show $\mathbb{Y}$, and therefore $\mathbb X$, is Polish. The set $\mathbb{Y}$ is a $G_\delta$ (${\bf \Pi}^0_2$) subspace of $A \times [0,1]^\mathbb N$. This can be seen by looking at the definition of $\mathbb{Y}$, $$\mathbb Y = \{(a,(x_n))\mid \forall \varepsilon\ \exists n\ (x_n < \varepsilon)\} \cup \{(a,(x_n))\mid \forall n,m\ (x_n + x_m \geq d_a(b_n(a),b_m(a)) \leq x_n - x_m)\}$$

Alexandrov's theorem says that every $G_\delta$ subset of a Polish space is Polish. Therefore $\mathbb Y$ (and $\mathbb X$) are Polish.


This proof is effective (including Alexandrov's theorem). So $\mathbb X$ is a computable metric space (effective Polish space) if $\mathbb A$ is a computable metric space and $a,m,n \mapsto d_a(b_n(a),b_m(a))$ is computable.

Unfortunately, however, there doesn't appear to be a nice metric since we applied Alexandrov's theorem. (The exception is when the $\mathbb B(a)$ are compact (in some uniform sense?). Then $\mathbb Y$ is closed in $A \times [0,1]^\mathbb N$ and can use the same metric.)

| cite | improve this answer | |
$\endgroup$

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.