2

Let $X,Y$ be Polish spaces and $\mathcal P(X)$ the space of all Borel probability measures on $X$. Fix $\mu\in \mathcal P(X)$ and $\nu \in \mathcal P(Y)$. Let $\gamma \in \Pi(\mu, \nu)$, i.e., $\gamma \in \mathcal P(X \times Y)$ whose marginal on $X$ is $\mu$ and that on $Y$ is $\nu$.

Let $c:X \times Y \to \mathbb R$ and $\varphi:X \to \mathbb R \cup\{-\infty\}$ be measurable. Let $\psi:Y \to \mathbb R \cup\{-\infty\}$ which is not necessarily measurable. We assume $\varphi (x) + \psi (y) = c(x, y)$ for $\gamma$-a.e. $(x, y) \in X \times Y$. This implies there is a $\sigma$-compact subset $S$ of $X \times Y$ such that

  • $\gamma (S) = 1$,
  • $\varphi (x), \psi(y) \neq -\infty$, and
  • $\varphi (x) + \psi (y) = c(x, y)$ for all $(x, y) \in S$.

My goal is to construct a map $\psi': Y \to \mathbb R \cup\{-\infty\}$ such that

  • $\psi' (y) = \psi (y)$ for $\nu$-a.e. $y\in Y$, and
  • $\psi'$ is measurable.

My attempt: The map $$ f: X \times Y \to \mathbb R \cup \{-\infty\} : c(x, y) - \varphi (x) $$ is measurable. Let $S_1, S_2$ be the projections of $S$ into $X,Y$ respectively, then $S_1, S_2$ are measurable and $\mu(S_1) = \nu(S_2) = 1$. For each $y \in S_2$, we pick some $x_y \in S_1$ and define $$ \psi' (y) = \begin{cases} &f(x_y, y) &&\text{if } y \in S_2 \\ &0 &&\text{otherwise}. \end{cases} $$

Then $\psi' = \psi$ $\nu$-a.e. but I'm not sure "my pick" assures that $\psi'$ is measurable.

Could you elaborate on a valid construction of $\psi'$?

Akira
  • 18,439

1 Answers1

1

I will prove the following weaker result.

Theorem: Let $X,Y$ be Polish spaces and $\gamma\in\mathcal P(X\times Y)$. Let $\varphi:X\rightarrow\mathbb R\cup\{-\infty\}$ be measurable, let $c:X\times Y\rightarrow\mathbb R$ be measurable, let $\psi:Y\rightarrow\mathbb R\cup\{-\infty\}$ be a map, and let $S\subseteq X\times Y$ be open and such that $\gamma(S)=1$, $c(x,y)=\varphi(x)+\psi(y)$ for $(x,y)\in S$. Then there exists a measurable set $S^*\subseteq S$ such that $\gamma(S^*)=1$, such that $S^*_2=\{y:(x,y)\in S^*\}\subseteq Y$ is measurable and the map $\psi':Y\rightarrow\mathbb R$, $y\mapsto\unicode{120793}\{y\in S^*_2\}\psi(y)$, is measurable. In particular, for $(x,y)\in S^*$ we have $c(x,y)=\varphi(x)+\psi'(y)$.

We start the proof with the following base case.

Proposition: Let $X,Y$ be measurable spaces and equip $X\times Y$ with the product $\sigma$-algebra and a probability measure $\gamma$. Let $\varphi:X\rightarrow\mathbb R\cup\{-\infty\}$ be measurable, let $\psi:Y\rightarrow\mathbb R\cup\{-\infty\}$, let $c:X\times Y\rightarrow\mathbb R$ be measurable and let $S\subseteq X\times Y$ be such that $c(x,y)=\varphi(x)+\psi(y)$ for $(x,y)\in S$, and there exist measurable $U_i\subseteq X$, $V_i\subseteq Y$, $i\in\mathbb N$, such that $S=\bigcup_i(U_i\times V_i)$. Then the map $\psi':Y\rightarrow\mathbb R$ given by $\psi'(y)=\psi(y)$ for $y\in\bigcup_iV_i$ and $\psi'(y)=0$ otherwise is measurable.

Proof: Assume without loss of generality that for all $i$ either both $U_i,V_i=\emptyset$ are trivial or both $U_i,V_i\neq\emptyset$ are not. Let $S_1=\{x:(x,y)\in S\}$ and notice that $S_1=\bigcup_iU_i$ is measurable. Let $S_2=\{y:(x,y)\in S\}$ and notice that $S_2=\bigcup_iV_i$ is measurable. For $x\in S_1$ there exists $i$ and $y$ with $(x,y)\in U_i\times V_i\subseteq S$, so $\varphi(x)+\psi(y)=c(x,y)>-\infty$ and hence $\varphi(x)>-\infty$. Similarly, we have $\psi(y)>-\infty$ for $y\in S_2$. This shows that $\psi'$ is well-defined since $\psi(y)>-\infty$ for $y\in S_2$. For an event $\mathcal E\subseteq\mathbb R$ let $\psi'^{-1}(\mathcal E)=\mathcal Y_0\cup\mathcal Y_1$ with $\mathcal Y_0=\psi'^{-1}(\mathcal E)\cap(Y\setminus S_2)$ and $\mathcal Y_1=\psi'^{-1}(\mathcal E)\cap S_2$. Notice that $\mathcal Y_0=\emptyset$ if $0\not\in\mathcal E$ and $\mathcal Y_0=Y\setminus S_2$ otherwise, so $\mathcal Y_0$ is measurable. Further, let $\mathcal Y_1=\bigcup_i\mathcal Y_1(i)$ with $\mathcal Y_1(i)=\psi'^{-1}(\mathcal E)\cap V_i$. Let $f:X\times Y\rightarrow\mathbb R$ be given by $f(x,y)=c(x,y)-\varphi(x)$ for $(x,y)\in S$ and $f(x,y)=0$ otherwise. Then $f$ is measurable and we further have $f^{-1}(\mathcal E)\cap(U_i\times V_i)=U_i\times\mathcal Y_1(i)$, but since this box is measurable, so are the sections $U_i$ and $\mathcal Y_1(i)$. This shows that the countable union $\mathcal Y_1$ is measurable and thereby the union $\psi'^{-1}(\mathcal E)$ is measurable. This shows that $\psi'$ is measurable.

Next, I'll briefly summarize a few properties of $\sigma$-compact spaces.

Fact: Let $K$ be $\sigma$-compact. Then there exists an increasing sequence $(K_n)_n$ with $\lim_{n\rightarrow\infty}K_n=K$. Further, every cover of $K$ with open sets admits a countable subcover.

Proof: By definition we have $K=\bigcup_nK'_n$ for countably many compact sets $K'_n$. But then we have $K=\bigcup_nK_n$ for the compact sets $K_n=\bigcup_{i=1}^nK'_i$, which is increasing. Let $K=\bigcup_iU_i$ over $i\in\mathcal I$ with $U_i$ open, and for $n\in\mathbb Z_{>0}$ let $\mathcal I_n\subseteq\mathcal I$ be such that $|\mathcal I_n|<\infty$ and $K_n\subseteq\bigcup_{i}U_i$ over $i\in\mathcal I_n$, using that $K_n$ is compact. This shows that there exists a countable subcover of $K=\bigcup_n\bigcup_{i\in\mathcal I_n}U_i$.

Proof of the theorem: Using that $S$ is open, that $X,Y$ are Polish and that $\gamma(S)=1$ we obtain a $\sigma$-compact subset $K$ of $S$ with $A=\pi^X(K)$, $B=\pi^Y(K)$ being $\sigma$-compact, where $\pi^X(x,y)=x$, $\pi^Y(x,y)=y$, and $\gamma(K)=\mu(A)=\nu(B)=1$, from the theorem here. Using the sup metric on $X\times Y$ for given metrics on $X,Y$, we can write the open set $S=\bigcup_{z\in S}(U_{z}\times V_z)$ as a union of (open) boxes. Using the fact, we obtain a countable subcover $S^*=\bigcup_i(U_i\times V_i)$ of $K\subseteq S^*$. But now, the proposition applies to $S^*$ since $S^*\subseteq S$, so $c(x,y)=\varphi(x)+\psi(y)$ for $(x,y)\in S^*$, and $\gamma(S^*)\ge\gamma(K)=1$. This yields the desired map $\psi'$.

Matija
  • 3,663
  • 1
  • 5
  • 24
  • Please see here for the existence of the $\sigma$-compact set $S$. My concern is that the requirement that $S$ has the form of $S=\bigcup_i(U_i\times V_i)$ is quite restrictive. – Akira Nov 19 '22 at 22:45
  • 1
    I have updated my answer based on my answer to the other question. The initial answer was intended to deal with the base case, which is now the proposition. Regarding further extensions, currently I need that $S$ is open on two separate occasions. First, for the existence of the $\sigma$-compact subset, discussed in the other thread. Second, for the existence of the cover by boxes. – Matija Nov 20 '22 at 11:44