5
$\begingroup$

Let $\mu_n,n\in \mathbb N$ be a random probability measures and let $\mu$ be a deterministic probability measure on $\mathbb R$. That is to say, that the $\mu_n$ are measurable maps from a probability space $(\Omega,\mathcal{T},\mathbf{P})$ to the space of $M_1(\mathbb R)$ equipped with the Borel-$\sigma$-algebra generated by the topology of weak convergence. Assume that the expected measures $\nu_n:=\mathbf E\mu_n$, defined via duality as $\int f~d\nu_n:=\mathbf E\int f~d\mu_n$ for all $f\in C_b(\mathbb R)$, converge weakly to $\mu$, i.e. that for all $f\in C_b(\mathbb R)$ we have the convergence of $\mathbf E\int f d\mu_n$ to $\int f d\mu$. By Levy's continuity Theorem this is equivalent to the statement that the characteristic function $\phi_{\nu_n}(t):=\int e^{itx}~d\nu_n(x)$ converges pointwise to the characteristic function $\phi_\mu$ of $\mu$. Is a similar statement also true for weak convergence in probability the random measures, i.e. is there a connection between the following statements?

  • For all $\epsilon>0$ the sequence $\mathbf{P}(d_{BL}(\mu_n,\mu)>\epsilon)$ converges to zero, where $d_{BL}$ is the bounded Lipschitz metric $$d_{BL}(\mu,\nu)=\sup\left\{\left\lvert\int f~d\mu-\int f~d\nu\right\lvert~:~f\colon\mathbb{R}\to\mathbb{R}\text{ is 1-Lipschitz and }\lVert f\lVert_\infty\leq1\right\}$$ (which completely metrizes the topology of weak convergence)

  • For all $f\in C_b(\mathbb R)$ it holds that $\int f~d\mu_n$ converges in probability to $\int f~d\mu$.

  • For all $t\in \mathbb R$, $\phi_{\mu_n}(t)$ converges to $\phi_\mu(t)$ in probability.

  • On some small interval $[0,\epsilon]$, the sequence $\sup_{0\leq t\leq \epsilon}\lvert \phi_{\mu_n}(t)-\phi_\mu(t)\lvert$ converges in probability to zero.

  • For all $k\in\mathbb N$, the sequence of moments $\int x^k~d\mu_n$ converges in probability to $\int x^k d\mu$.

  • For all $k\in\mathbb N$, the sequence of expected moments $\mathbf E\int x^k~d\mu_n$ converges to $\int x^k d\mu$ and the variances $\mathbf{Var}\int x^k~d\mu_n$ converge to zero.

Clearly, for the last two statements to be equivalent to any of the above, we would need some condition on the determinacy of $\mu$ via its moments. Therefore, for simplicity assume that $\mu$ is subgaussian, and that the $\mu_n$ are uniformly subgaussian in the sense that $\mathbf E\mu_n([-R,R]^c)\leq C e^{-C R^2}$ for some $C$ and all $n,R$.

$\endgroup$
  • $\begingroup$ You defined the expected measure $\nu_{n}:=\mathbf{E}\mu_{n}$ via duality as $\int f~d\nu_n:=\mathbf E\int f~d\mu_n$ for all $f\in C_b(\mathbb R)$. I assume this is by the Riesz-Markov representation theorem. But for this to apply don't you need the space of test functions which are continuous bounded and compactly supported, that is $C_{c}(\mathbb{R})$? Or why is it sufficient to take test functions from the larger space $C_{b}(\mathbb{R})$? $\endgroup$ – Marco Breitig Jul 17 '15 at 12:38
2
$\begingroup$

$\newcommand{\bR}{\mathbb{R}}$ For any Polish space $S$ (separable complete metric space) we denote by $M_1(S)$ the space of Borel probability measures on $S$.

The space $\newcommand{\eP}{\mathscr{P}}$ $\eP:=M_1(\bR)$ is a topological space with respect to the weak convergence. In fact, $\eP$ with this topology is a Polish space.

A random measure $\mu$ on $\bR$ is by definition a measure on $\eP$, i.e.,

$$\mu \in M_1\bigl(\;\eP\;\bigr).$$

The space $C_b(\bR)$ embeds in $C_b\bigl(\eP\bigr)$ $\newcommand{\bsE}{\boldsymbol{E}}$

$$ f\in C_b(\bR)\mapsto \widehat{f}\in C_b\bigl(\eP\bigr), \;\;\widehat{f}(\pi):=\bsE_\pi(f)=\int_{\bR} f(t) \pi(dt),\;\;\forall \pi\in \eP. $$

wher $\bsE_\pi$ denotes the expectation of a random variable with respect to the probability distribution $\pi$. Denote by $\newcommand{\eX}{\mathscr{X}}$ $\eX\subset C_b\bigl(M_1(\bR)\bigr)$ the image of this embedding.

In your question you are given $\newcommand{\bmu}{\boldsymbol{\mu}}$ a sequence $\bmu_n\in M_1(\eP)$ and another measure $\bmu$ with the property that

$$\lim_n\int_{\eP}F(\pi) \bmu_n(d\pi)=\int_{\eP}F(\pi)\bmu(d\pi),\;\;\forall F\in \eX $$

and you ask if the above holds for all $F\in C_b(\eP)$. Clearly this happens iff $\eX$ is dense in $C_b(\eP)$. It is not hard to see that is not the case. Fix a nontrivial, compactly supported continuous function $\rho:\bR\to[0,1]$. Then the bounded, continuous function

$$ Q_\rho:\eP\to\bR,\;\;Q_\rho(\pi)=\left(\int_{\bR}\rho(t) \pi(dt)\,\right)^2, $$

is not in the closure of $\eX$.

$\endgroup$
  • $\begingroup$ I am not entirely sure how your answer relates to my question. Firstly, I'd rather say that a random measure $\mu$ is, by definition, a measurable map $\omega\mapsto\mu(\omega)$ from some probability space $(\Omega,\mathcal{T},\mathbf P)$ to $(M_1(\mathbb R),\mathcal{B})$ where $\mathcal{B}$ is the Borel $\sigma$-algebra generated by the topology of weak convergence. What you defined, is in my mind, the distribution of such a random measure and I don't think that a concept like convergence in probability can be phrased in terms of the distribution. $\endgroup$ – whz Feb 16 '15 at 17:19
  • $\begingroup$ Secondly, in your answer you only assumed the convergence of $\int_{\mathcal{P}}\int_{\mathbb R} f(t)~d\pi(t)~d\mu_n(\pi)$ towards $\int_{\mathcal{P}}\int_{\mathbb R} f(t)~d\pi(t)~d\mu(\pi)$ for all $f\in C_b(\mathbb R)$, which I would call 'convergence in expectation'. It is clear that convergence in expectation is not sufficient on its own. The assumptions in my question are much stronger than convergence in expectation. $\endgroup$ – whz Feb 16 '15 at 17:28
  • $\begingroup$ @whz You are right. $\endgroup$ – Liviu Nicolaescu Feb 17 '15 at 19:59
0
$\begingroup$

I found a way to prove the implication I was mostly interested in (the last statement implies the second): Below my argument, for the sake of completeness.

Fix a function $f$. A sequence of random variables $X_n$ converges to a random variable $X$ in probability if and only if any subsequence $X_{n_m}$ has a further subsequence $X_{n_{m_l}}$ which converges to $X$ almost surely. Apply this to the sequence in question and pick the subsequence $n_{m_l}$ in such a way that $\mathbf{Var}\int x~d\mu_{n_{m_l}}$ is summable. By Borel-Cantelli this implies that $\int x~d\mu_{n_{m_l}}$ converges almost surely to $\int x~d\mu$. Continue inductively and choose a diagonal sequence (for convenience still named the same) such that the sequences $\int x^k~d\mu_{n_{m_l}}$ converge to $\int x^k~d\mu$ almost surely for all $k\in\mathbb N$. If the limiting distribution is uniquely determined by its moment this implies that with full probability $\mu_{n_{m_l}}$ converges weakly to $\mu$, which, in particular, implies that the integral against $f$ converges almost surely.

$\endgroup$

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.