18.9 Kolmogorov’s Extension Theorem

Definition 18.9.1 (Consistent).label Let $\bracs{(\Omega_i, \cf_i)}_{i \in I}$ be a family of measurable spaces. For each $J \subset J' \subset I$, denote $\pi_{J}: \prod_{i \in I}X_{i} \to \prod_{j \in J}X_{j}$ and $\pi_{J', J}: \prod_{j \in J'}X_{j} \to \prod_{j \in J}X_{j}$ as the projection maps.

Let $\bracs{\mu_J|J \subset I \text{ finite}}$ such that each $\mu_{J}$ is a probability measure on $\prod_{j \in J}\Omega_{j}$, then $\bracs{\mu_J}$ is consistent if for all $J \subset J' \subset I$, $\mu_{J'}= \mu_{J} \circ \pi_{J', J}^{-1}$.

Lemma 18.9.2 ([Lemma 1.17, Bau14]).label Let $\seq{X_n}$ be topological spaces where for each $n \in \natp$,

  1. (a)

    Every finite measure on $\prod_{j = 1}^{n} X_{j}$ is regular.

  2. (b)

    $X_{n}$ is Hausdorff.

  3. (c)

    $X_{n}$ is separable.

Let $\bracs{\mu_{I}| I \subset \natp \text{ finite}}$ be consistent Borel probability measures, then for any $\seq{B_n}$ where:

  1. (d)

    For each $n \in \nat$, $B_{n} \in \cb_{\prod_{j = 1}^n X_j}$.

  2. (e)

    For each $n \in \nat$, $B_{n+1}\subset B_{n} \times X_{n+1}$.

  3. (f)

    There exists $\eps > 0$ such that $\mu_{[n]}(B_{n}) > \eps$ for all $n \in \natp$.

Then there exists $\seq{K_n}$ such that for every $n \in \natp$,

  1. (1)

    $K_{n} \subset \prod_{j = 1}^{n} X_{j}$ is compact.

  2. (2)

    $K_{n} \subset B_{n}$.

  3. (3)

    $K_{n+1}\subset K_{n} \times X_{n+1}$.

  4. (4)

    $\mu(K_{n}) \ge \eps/2$.

Proof. Let $n \in \natp$. By (c), $\mu_{[n]}$ is a Borel probability measure. Thus $\mu_{[n]}$ is regular by (a). Thus there exists $C_{n} \subset \prod_{j = 1}^{n} X_{j}$ compact such that $C_{n} \subset B_{n}$ and $\mu_{[n]}(B_{n} \setminus C_{n}) < \eps/2^{n+1}$. Let

\[K_{n} = \bigcap_{j = 1}^{n} \pi_{[n], [j]}^{-1}(C_{n})\]

Since each $X_{j}$ is Hausdorff, $K_{n} \subset \prod_{j = 1}^{n} X_{j}$ is compact with $K_{n} \subset B_{n}$ and $K_{n+1}\subset K_{n} \times X_{n+1}$. Moreover,

\[\mu_{[n]}(B_{n} \setminus K_{n}) \le \sum_{j = 1}^{n}\mu_{[n]}\braks{\pi_{[n], [j]}^{-1}(B_j \setminus C_j)}\le \sum_{j = 1}^{n}\mu_{[j]}(B_{j} \setminus C_{j}) \le \eps/2\]

Thus $\mu_{[n]}(K_{n}) \ge \eps/2$.$\square$

Theorem 18.9.3 (Kolmogorov’s Extension Theorem).label Let $\seqi{X}$ be topological spaces where for each $J \subset I$ finite,

  1. (a)

    Every finite measure on $\prod_{j \in J}X_{j}$ is regular.

  2. (b)

    $X_{j}$ is Hausdorff.

  3. (c)

    $X_{j}$ is separable.

Let $\bracs{\mu_{I}| I \subset \natp \text{ finite}}$ be consistent Borel probability measures, then there exists a unique probability measure $\mu: \bigotimes_{i \in I}\cb_{X_i}\to [0, 1]$ such that for any $J \subset I$ finite, $\mu = \mu_{J} \circ \pi_{J}^{-1}$.

Proof [Theorem 1.14, Bau14]. Let

\[\alg = \bracs{\pi_J^{-1}(B)| B \in \bigotimes_{j \in J}\cb_{X_j}, J \subset I \text{ finite}}\]

then $\alg$ is an algebra. For any $B \in \bigotimes_{j \in J}\cb_{X_j}$, define $\mu_{0}(\pi_{J}^{-1}(B)) = \mu_{J}(B)$, then $\mu_{0}: \alg \to [0, 1]$ is well-defined and finitely additive by the consistency of $\bracs{\mu_J}$.

To show that $\mu_{0}$ is a premeasure, it is sufficient to show that for any $\{\pi_{J_n}^{-1}(B_{n})\}_{1}^{\infty}$ with $\pi_{J_n}^{-1}(B_{n}) \downto \emptyset$, $\mu_{0}(\pi_{J_n}^{-1}(B_{n})) \downto 0$. To this end, suppose for contradiction that $\limv{n}\mu_{0}(\pi_{J_n}^{-1}(B_{n})) = \eps > 0$.

By inserting additional elements into the sequence and relabeling the indices, assume without loss of generality that $J_{n} = [n]$ for all $n \in \natp$. By Lemma 18.9.2, there exists $\seq{K_n}$ such that:

  1. (1)

    $K_{n} \subset \prod_{j = 1}^{n} X_{j}$ is compact.

  2. (2)

    $K_{n} \subset B_{n}$.

  3. (3)

    $K_{n+1}\subset K_{n} \times X_{n+1}$.

  4. (4)

    $\mu(K_{n}) \ge \eps/2$.

Let $N \in \natp$ and $x \in \prod_{j = 1}^{N} X_{j}$ such that $x \in \bigcap_{n \ge N}\pi_{[N]}(K_{n})$. By compactness and (4), there exists $x_{N+1}\in X_{N+1}$ such that $(x_{1}, \cdots, x_{N}, x_{N+1}) \in \bigcap_{n > N}\pi_{[N+1]}(K_{n})$. Thus there exists $x \in \prod_{i \in I}X_{i}$ such that $x \in \pi_{[n]}^{-1}(K_{n})$ for all $n \in \natp$, and

\[x \in \bigcap_{n \in \natp}\pi_{[n]}^{-1}(K_{n}) \subset \bigcap_{n \in \natp}\pi_{[n]}^{-1}(B_{n}) \ne \emptyset\]

$\square$