Let $\mathcal{P}_d$ be the space of Borel probability measures on $\mathbb{R}^d$ endowed with the topology of weak convergence of probability measures. Does there exist a map $G: \mathcal{P}_d \to \mathbb{R}^{\infty}$ (the latter with the usual topology of pointwise convergence) $\textbf{which is one-to-one and continuous}$ such that $G(\mathcal{P}_d) \subseteq \mathbb{R}^{\infty}$ is closed?

Some thoughts on this: It is known that for $\{\mu_n\}_n \in \mathcal{P}_d$, the existence of lim$_n$$\int f d\mu_n$ in $\mathbb{R}$ for each $f \in C_b(\mathbb{R}^d)$ yields the existence of a unique $\mu \in \mathcal{P}_d$ such that $\mu_n \to \mu$ weakly. Hence, if one could choose $G$ as $$G(\mu) = \bigg(\int f d\mu\bigg)_{f \in C_b},$$ the answer to my question would be affirmative. Since $C_b$ is neither countable nor separable, such a choice of $G$ is not possible. Replacing $C_b$ by a dense, countable subset of $C_c(\mathbb{R}^d)$ does not work either, since then the limit object $\mu$ will in general only be a sub-probability measure. This allows only to conclude that this map $G$ would turn $G(\mathcal{M}^+_{\leq 1}) \subseteq \mathbb{R}^{\infty}$ into a closed set, where $\mathcal{M}^+_{\leq 1}$ denotes the set of all Borel sub-probability measure endowed with the vague topology.

Is there a way to solve this issue? I was thinking of replacing $C_b$ by the union of a dense, countable subset of $C_c$ and a sequence $(f_l)_{l \geq 1}$ of $C_b$-functions such that $f_l \uparrow 1$ uniformly. Then, the existence of a sub-probability measure $\mu$ as the vague limit of $(\mu_n)_n$ follows and my hope is to obtain $$\mu(\mathbb{R}^d) = lim^l\int f_l d\mu = lim^l \, lim^n \int f_l d\mu_n = lim^n \, lim^l \int f_l d\mu_n = lim^n \mu_n(\mathbb{R}^d) = 1,$$where the interchange of limits is justified by the uniform convergence of $f_l \to1$. However, the second equality uses that the integrals $\int f_l d\mu$ are represented by $lim^n \int f_l d\mu_n$. But the existence of $\mu$ is deduced from the Riesz-Markov-Kakutani representation and from there I do not know how to conclude that the integral representation of the limit object $\mu$ also holds for functions $f$ other than $f \in C_c$.

Is this approach feasible at all? If not, is there any other way to answer my question affirmatively? Thanks a lot in advance!

  • $\begingroup$ I am unable to understand the question. Why can't you take $G(P)=(0,0,..0)$ for all $P$? $\endgroup$ Mar 19 '20 at 10:17
  • $\begingroup$ Sorry for being sloppy. Of course, I forgot to mention one crucial requirement: $G$ should be one-to-one. I will edit the post. Thanks! $\endgroup$
    – Marco
    Mar 19 '20 at 10:41
  • $\begingroup$ You probably want some more properties of $G$. Do you want it to be continuous and linear? $\endgroup$ Mar 19 '20 at 11:52
  • $\begingroup$ Sorry for again being imprecise. I've edited the post again. Linearity is, for my concerns, not necessarily needed I suppose. $\endgroup$
    – Marco
    Mar 19 '20 at 11:57
  • $\begingroup$ I was also sloppy :-). Linearity doesn't even make sense. I should have mentioned convexity instead of linearity. $\endgroup$ Mar 19 '20 at 11:59


In general, every Polish space is homeomorphic to a closed subset of $\mathbb{R}^\infty$; see Kechris, Classical Descriptive Set Theory, Theorem 4.17 (where the space is denoted $\mathbb{R}^{\mathbb{N}}$). And $\mathcal{P}_d$ with its weak topology is Polish; Kechris Theorem 17.23.


I'm a little stumbled, but I cannot find a mistake in the following proof, which would somehow prove the statement of the above answer by Nate Eldredge without the assumption of the space under consideration being Polish (which I did not expect to be true!):

Let $X$ be a separable, metrizable topological space. Fix any metric $d$, which metrics the given topology on $X$. Now define

$$F: X \to \mathbb{R}^{\infty}, F(x) := (d(x,x_1),d(x,x_2),\dots), $$where $(x_n)_n$ is a fixed dense subset of $X$ (such set does not depend on the actual choice of $d$). $\mathbb{R}^{\infty}$ is endowed with the metric $\rho(\alpha,\beta) := \sum_{k\geq1} 2^{-k}|\alpha_k-\beta_k|$, which induces the topology of pointwise convergence. Let's show that $F$ is one-to-one and a homeomorphism between $X$ and $F(X) \subseteq \mathbb{R}^{\infty}$:

Firstly, for $d(x,y) = c > 0$ we find $x_k$ such that $d(x,x_k) < \frac{c}{3}$, thus $d(y,x_k) > \frac{c}{3}$ and hence $F(x) \neq F(y)$, whereby $F$ is one-to-one.

Secondly, for $\epsilon > 0$ and $x \in X$, if $d(x,y)< \epsilon$, then $|d(x,x_i)-d(y,x_i)| < \epsilon$ for each $i$, so that $\rho(F(x),F(y)) < \epsilon$, whereby $F$ is continuous.

Finally, let $\epsilon > 0$ and fix $F(x) \in F(X)$. There exists $x_n$ such that $d(x,x_n) < \frac{\epsilon}{3}$. If $y \in X$ with $d(y,x_n) > \frac{2}{3}\epsilon$, then $|d(x,x_n)-d(y,x_n)| > \frac{\epsilon}{3}$, hence $\rho(F(x),F(y)) > \frac{\epsilon}{3\cdot2^{n}}$. Thereby, $\rho(F(x),F(y)) \leq \frac{\epsilon}{3\cdot2^{n}} \implies d(x,y) < \epsilon, $whereby $F^{-1}$ is continuous from $F(X)$ to $X$.

Since $X$ is closed (as a subset of itself) and $F^{-1}$ is continuous, also $F(X) = (F^{-1})^{-1}(X)$ is closed.

Is there a mistake in my reasoning? (Sorry for posting this question as an answer, it was certainly too long for a comment and I did not want to open a new post for this)

  • $\begingroup$ Your second-to-last paragraph only proves that $F(X)$ is closed in $F(X)$ (i.e. with respect to the subspace topology induced from $\mathbb{R}^\infty$), which is trivial. It does not prove that it is closed in $\mathbb{R}^\infty$. $\endgroup$ Mar 19 '20 at 20:32

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.