Do the finite product probability measures converge to the infinite product probability measure?

48 Views Asked by At

Is the limit of the finite product probability measures equal to the infinite product probability measure?\

$\newcommand{\N}{\mathbb{N}}$ $\newcommand{\A}{\mathfrak{A}}$ To be precise:
Let $(P_n)_{n\in \N}$ be probability distributions on $(\Omega_n, \A_n)$ respectively.
For $I\subset \N$, set $P_I:=\times_{i\in I}P_i$, the product probability measure on $\bigotimes_{i \in I}\A_i =: \A_I$
We know this exists because of Ionescu-Tulcea's Theorem for the infinite case.
Let $\A'_I:=\times_{i\in I}\A_i$ be the cartesian product.
Let $\Omega_I := \times_{i\in I}\Omega_i$
Let $\pi_I:\A_\N \rightarrow \A_I$ be the projection on the I'th components.
For $A \in \A$, we write $\pi_I(A):=\{\pi_I(a)|a\in A\}$
For $n \in \N$, let $\N_n = \{1,\dots,n\}$
and $\N_{> n} = \N\setminus \N_n = \{n+1,\dots\}$.

Set $Q_n(A) := P_{\N_n}(\pi_{\N_n}(A))$
Note that $Q_n$ is only well-defined if $\pi_{\N_n}(A)$ is a measurable set.
Set $Q(A) := lim_n Q_n(A)$
Since $\pi_{\N_n+1}(A)\times\Omega_{\N_{> n+1}}\subseteq \pi_{\N_n}(A)\times\Omega_{\N_{> n}}$ we have that
$Q_n(A)=P_{\N_n}(\pi_{\N_n}(A))=P_{\N}(\pi_{\N_n}(A)\times\Omega_{\N_{> n}})\geq P_{\N}(\pi_{\N_n+1}(A)\times\Omega_{\N_{> n+1}})=Q_{n+1}(A)$.
So $Q_n(A)$ is non-increasing and bounded by 0. This means the limit exists and Q is well-defined.
Obviously for all rectangles $A=B\times \Omega_{\N{>n}}$ for $n\in \N, B\in \A'_{\N_n}$
we have $Q(A)=Q_n(A)=P_\N(A)$, so if Q is a measure we already know $Q=P_\N$ since they agree on a generator.\

So my first Question is: Is there a counterexample where $\pi_{\N_n}(A)$ is not a measurable set?

My second Question is: Is there a counterexample where $Q_n$ is well-defined, but $Q\neq P_\N$?

1

There are 1 best solutions below

1
On BEST ANSWER

My friend came up with a counterexample to the second question, so I might as well share it.

Let $P_i=U(\{0,1\})$
Let $A_{n,m} = \{0,1\}^n \times \{0\}^m \times \{0,1\}^{\mathbb{N}_{>n+m}} \in \mathfrak{R}$, the measurable rectangles.
We have $P(A_{n,m})={1\over 2^m}$
Let $A_n = \bigcap_{m\in\mathbb{N}}A_{n,m} =\{0,1\}^n\times \{0\}^{\mathbb{N}_{>n}} $
We have $P(A_n) = \lim_m P(A_{n,m}) = 0$
Let $A = \bigcup_{n\in\mathbb{N}} A_n = \{b\in\{0,1\}^\mathbb{N}|\exists n_0\in\mathbb{N}:\forall m\in\mathbb{N}_{>n_0}:b_m=0\}$.
Since $A_n \subseteq A_{n+1}$ we have $A_n \big\uparrow A$ so $P(A) = \lim_n P(A_n) = 0$

While $\forall n\in\mathbb{N}:\pi_{\mathbb{N}_n}(A)=\{0,1\}^n$._n So $\forall n\in\mathbb{N}:Q_n(A)=1$, so $Q(A)=1$
We have $P(A)\neq Q(A)$