Suppose that $S$ is a separable metric space or Polish. Let $μ_{n},n∈N $ be a random probability measures and let μ be a deterministic probability measure on $S$. That is to say, that the $ μ_{n}$ are measurable maps from a probability space $(Ω,T,P)$ to the space of $M_1(S) $ equipped with the Borel-σ-algebra generated by the topology of weak convergence. Assume that the expected measures $ν_{n}:=Eμ_{n}$, defined via duality as $∫f dν_{n}:=E∫f dμ_{n}$ for all $f∈C_{b}(S)$, converge weakly to μ, i.e. that for all $f∈C_{b}(S)$ we have the convergence of $E∫fdμ_{n}$ to $∫fdμ$.

  1. For all ϵ>0 the sequence $P(d_{BL}(μ_{n},μ)>ϵ)$ converges to zero, where $d_{BL}$ is the bounded Lipschitz metric $$ d_{BL}(μ,ν)=\sup\Big\{\Big\vert∫f dμ−∫f dν\Big|: f:S\to\mathbb{R} \; \mbox{is 1-Lipschitz and $\Vert f\Vert_\infty\leq 1$}\Big\} $$ which completely metrizes the topology of weak convergence)

  2. For all $f∈C_{b}(S)$ it holds that $∫f dμ_{n}$ converges in probability to $∫f dμ$.

please i want to know if 1) implies 2) or if 2) implies 1) or if 1) is equivalent to 2)

and in the case where $μ_{n}$ is not necessary a probability measure but just a random measure, do the same results remain true?


In your setting, the two are equivalent. Let's first show that your assumptions imply that the $\mu_n$, viewed as random variables in $M_1(S)$, are tight. Since $S$ is Polish and the $\nu_n$ converge weakly to a limit, we know from Prokhorov that the sequence $\nu_n$ is tight so, for every $\varepsilon > 0$ there exists a compact set $K_\varepsilon \subset S$ such that $\nu_n(K_\varepsilon) > 1-\varepsilon $, uniformly in $n$. A simple application of Markov's inequality then shows that $P(\mu_n(K_\varepsilon) > 1-\sqrt \varepsilon) > 1-\sqrt \varepsilon$.

Define now $$\mathcal{K}_\delta = \{\mu \in M_1(S)\,:\, \mu(K_{2^{-2m}\delta^2}) > 1- 2^{-m} \delta \quad \forall m \in \mathbb{N}\}.$$ The set $\mathcal{K}_\delta$ is tight and therefore precompact in $M_1(S)$. Furthermore, we know from above that $P(\mu_n(K_{2^{-2m}\delta^2}) > 1-2^{-m} \delta) > 1-2^{-m} \delta$ so that, summing all of these probabilities, one has $P(\mu_n \in \mathcal{K}_\delta) > 1-\delta$. Now that we know that the $\mu_n$ are tight, we know that there are (possibly random) accumulation points $\hat \mu$.

The claim now is that both of your conditions are equivalent to the statement that $\hat \mu = \mu$ almost surely. It is clear that this statement implies your second claim by continuity of the maps $\eta \mapsto \int f\,d\eta$. It also implies the first claim since $d_{BL}$ is a metric for $M_1(S)$ and convergence in law to a deterministic element is the same as convergence in probability. The fact that your claims both imply that $\hat \mu$ cannot be anything other than $\mu$ is pretty much immediate.

I haven't checked the details, but I guess that if the $\mu_n$ are allowed to be random positive measures, then the claim still holds. (Testing with $f=1$ gives an a priori bound on the mass of the measure and identifies the mass of the limit.)


The equivalence of 1. and 2. is essentially a special case of [1] van der Vaart/Wellner (1996), Weak Convergence and Empirical Processes - With Applications to Statistics. (Unfortunately this book mainly covers nonseparable metric spaces, so it may be hard to read.)

That 1. implies 2. even if $\mu$ is arbitrary, is an easy consequence of [1], Lemma 1.9.2 (ii), rewritten to the separable case: $\mu_n \to \mu$ in probability iff every subsequence $(\mu_{n'})$ has a further subsequence $(\mu_{n''})$ with $\mu_{n''} \to \mu$ almost surely.

The metric space $\mathbb{D}$ in [1] here are $(M_1(S),d_{BL})$ resp. $\mathbb{R}$. These spaces are separable.

That 2. implies 1., now for constant $\mu$, is a direct consequnce of [1], Lemma 1.10.2 (iii).


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.