I'm generally wary when lifting a result stated unconditionally to a situation where I'm conditioning on a random variable. Consider the following classical result in weak convergence:

Theorem. Let $\phi,\phi_1,\phi_2,\dotsc$ be random measures on a Polish space $S$. Then $\phi_n \to \phi$ in distribution if and only if $\phi_n f \to \phi f$ in distribution for every bounded continuous $f$.

Now consider this application: Let $\eta$ be a random variable, and define $\zeta_n = \Pr[\phi_n \in \cdot \,| \eta]$ and $\zeta = \Pr[\phi \in \cdot \,| \eta]$. I would like to show $\zeta_n \to \zeta$ weakly almost surely.

Naively, one might think to show that, for every bounded continuous $f$, we have $$\Pr [\phi_n f \in \cdot \,|\eta] \to \Pr [\phi f \in \cdot \, | \eta] \text { weakly a.s.}$$ But this fails because of the uncountable number of bounded continuous functions. However, if I knew that there was some countable class $\mathcal C$ of bounded continuous functions such that convergence on this class implied convergence for all bounded continuous functions, then I could establish the a.s. weak convergence on this countable class and then conclude that convergence holds simultaneously, and then I can apply the Theorem above.

Is there anything that I have missed?

Are there example where one might expect to transfer an unconditional result to a conditional setting, but trouble lurks?

  • $\begingroup$ Have I got it right that $\zeta_n$ are random measures on the space $\mathcal{P}(S)$ of measures? That is, they are $\mathcal{P}(\mathcal{P}(S))$-valued random variables? $\endgroup$ – Nate Eldredge May 4 '16 at 0:27
  • $\begingroup$ Yes, that is correct. $\endgroup$ – D.R. May 5 '16 at 3:37
  • $\begingroup$ If $S$ is compact, then $C(S)$ is separable, so you can pick a countable dense subset and use that as $\mathcal{C}$. I think you can do something similar if $S$ is locally compact (let $\mathcal{C}$ be dense in $C_c(S)$). $\endgroup$ – Nate Eldredge May 5 '16 at 3:56
  • $\begingroup$ Interesting idea. And of course Polish does not imply locally compact, e.g., Baire space ${\mathbb N}^{\mathbb N}$. Perhaps there is a counterexample lurking in there. $\endgroup$ – D.R. May 5 '16 at 4:07

There's a fairly well known result due to Varadarajan (I think in this paper, though I can't get through the paywall at the moment), that provides your countable class $\mathcal{C}$ you desire: If $S$ is a Polish space, then there exists a countable family $\mathcal{C}$ of continuous bounded functions of $S$ such that $\mu_n \rightarrow \mu$ in $\mathcal{P}(S)$ if and only if $\mu_n f\rightarrow\mu f$ for every $f \in \mathcal{C}$.

| cite | improve this answer | |
  • 1
    $\begingroup$ Oh right. This is really just the fact that the weak topology on $\mathcal{P}(S)$ is second countable (since it's separable and metrizable). $\endgroup$ – Nate Eldredge May 8 '16 at 18:08

To fill in some details in Dan's answer:

If $S$ is Polish then there does exist a countable $\mathcal{C} \subset C_b(S)$ which determines weak convergence in $\mathcal{P}(S)$, as you desire.

It's a standard fact that if $S$ is Polish then the weak topology on $\mathcal{P}(S)$ is separable and metrizable; see the paper of Varadarajan cited by Dan, or Billingsley's Convergence of Probability Measures which anybody working in this field surely must have on their shelf ;-). As such it is second countable.

Now by definition of the weak topology, open sets of the following form are a basis for $\mathcal{P}(S)$:

$$U_{\mu, f_1, \dots, f_n, \epsilon} := \{\nu : |\mu f_i - \nu f_i| < \epsilon,\; i = 1,\dots,n\}$$

where $\mu \in \mathcal{P}(S)$, $f_1, \dots, f_n \in C_b(S)$ (the bounded continuous functions), and $\epsilon > 0$.

By second countability, there is a basis $\mathcal{U}$ consisting of countably many such sets. Let $\mathcal{C}$ consist of all the functions $f_i$, used in defining all the $U \in \mathcal{U}$; then $\mathcal{C}$ is a countable union of finite sets and hence countable.

Now suppose $f \nu_j \to f\nu$ for all $f \in \mathcal{C}$. Suppose $U \in \mathcal{U}$ is a neighborhood of $\nu$. We can write $U = U_{\mu, f_1, \dots, f_n, \epsilon}$ for some $\mu, f_1, \dots, f_n, \epsilon$. Since $\nu \in U$, we have $|f_i \nu - f_i \mu| < \epsilon$ for all $i$. Let $\delta = \max_i |f_i \nu - f_i \mu| < \epsilon$.

Now $f_1, \dots, f_n \in \mathcal{C}$, so by assumption we have $f_i \nu_j \to f_i \nu$ for each $i$. There is thus $J$ so large that for all $j > J$, we have $|f_i \nu_j - f_i \nu| < \epsilon - \delta$ for each $i$. Then for such $j$ we have $$|f_i \nu_j - f_i \mu| \le |f_i \nu_j - f_i \nu| + |f_i \nu - f_i \mu| < (\epsilon - \delta) + \delta = \epsilon.$$ Hence $\nu_j \in U$. Since $U$ ranges over a basis, it follows that $\nu_j \to \nu$ weakly.

| cite | improve this answer | |

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.