Consider the probability space $(\mathcal{S}, \mathcal{A}, \mu)$. Suppose that the random variables $X_1,..., X_n$ are exchangeable and take values in the Borel space $(\mathcal{X}, \mathcal{B})$. We can prove that the empirical probability measure $P_n$ is a measurable function from $(\mathcal{X}^n, \mathcal{B}^n)$ to $(\mathcal{P}, \mathcal{C}_{\mathcal{P}})$, where $\mathcal{P}$ is the set of all probability measures on $(\mathcal{X}, \mathcal{B})$ (Schervish, Theory of Statistics, Sec. 1.4 Prob. 24.).

I have a problem in understanding where do we use exchangeability.

Indeed, suppose for simplicity $\mathcal{X}=\{0,1\}$.

$\forall x:=(x_1,..., x_n)\in \mathcal{X}^n$, we have

$$ \begin{cases} P_n(x)(\{1\}):=\frac{1}{n}\sum_{i=1}^n 1(x_i=1)\\ P_n(x)(\{0\}):=\frac{1}{n}\sum_{i=1}^n 1(x_i=0)\\ P_n(x)(\{1,0\}):=\frac{1}{n}\sum_{i=1}^n 1(x_i\in \{0,1\})\\ P_n(x)(\emptyset):=\frac{1}{n}\sum_{i=1}^n 1(x_i\in \emptyset)\\ \end{cases} $$ and hence $P_n(x):\mathcal{B}\rightarrow[0,1]$ is a probability measure on $(\mathcal{X}, \mathcal{B})$ assigning $P(x)(B)\in [0,1]$ to each $B\in \mathcal{B}$.

Hence, $P_n: \mathcal{X}^n\rightarrow \mathcal{P}$.

Or (equivalently?), suppose for simplicity $\mathcal{S}:=\{s_1,s_2,s_3\}$, $\mathcal{X}:=\{0,1\}$, $n=2$, $X_1(s_1)=1,X_1(s_2)=1 , X_1(s_1)=0$, and $X_2(s_1)=0,X_2(s_2)=1, X_2(s_1)=0$.

$\forall s:=(s_1,s_2)\in \mathcal{S}^2$, we have

$$ \begin{cases} P_n(s)(\{1\}):=\frac{1}{2}\sum_{i=1}^2 1(X_i(s_i)=1)\\ P_n(s)(\{0\}):=\frac{1}{2}\sum_{i=1}^2 1(X_i(s_i)=0)\\ P_n(s)(\{1,0\}):=\frac{1}{2}\sum_{i=1}^2 1(X_i(s_i)\in \{0,1\})\\ P_n(s)(\emptyset):=\frac{1}{2}\sum_{i=1}^2 1(X_i(s_i)\in \emptyset)\\ \end{cases} $$ and hence $P_n(s):\mathcal{B}\rightarrow[0,1]$ is a probability measure on $(\mathcal{X}, \mathcal{B})$ assigning $P(s)(B)\in [0,1]$ to each $B\in \mathcal{B}$.

Hence, $P_n: \mathcal{S}^n\rightarrow \mathcal{P}$.

Where do we use exchangeability?

Similar question here


We don't. Fix $B\in\mathcal B$ and consider the map $f_B(P):=P(B)$. You have, for a Borel subset $A$ of $\Bbb R$, $$ \eqalign{ \{x\in\mathcal X^n:P_n(x)\in f_B^{-1}(A)\}&=\{x\in\mathcal X^n:P_n(x)(B)\in A\}\cr &=\{x\in\mathcal X^n:{1\over n}\sum_{k=1}^n 1_B(x_k)\in A\},\cr } $$ and each of the maps $x\mapsto 1_B(x_k)$ is $\mathcal B^n/\mathcal B(\Bbb R)$-measurable, hence so is their average. It follows that $\{x\in\mathcal X^n:P_n(x)\in f_B^{-1}(A)\}\in\mathcal B^n$ for each $A\in\mathcal B(\Bbb R)$ and $B\in\mathcal B$, and since $\mathcal C_{\mathcal P}$ is generated by $\{f_B^{-1}(A): A\in\mathcal B(\Bbb R), B\in\mathcal B\}$, the asserted measurability of $P_n$ follows.


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.