Definition of marginal probability measures using test function spaces

20 Views Asked by At

This is a question from Villani's Topics in Optimal Transportation. Let's define $\mu, \upsilon$ probability measures on some measure spaces $X, Y$ respectively, and $\pi$ a probability measure on the product space $X \times Y$.

In the Introduction chapter, pg. 2, there is an intuitive assertion that for a transference plan $\pi \in P(X \times Y)$ to be admissible, it must verify

$$\int_{Y} d\pi(x,y) = d\mu(x) \, \text{and, similarly} \, \int_{X} d\pi(x,y) = d\upsilon(y)$$

which can be more rigorously expressed as:

$$\pi(A \times Y) = \mu(A) \, \text{and} \, \pi(X \times B) = \upsilon(B)$$

for all $A \subset X, B \subset Y$ measurable subsets, is equivalent to

$$\int_{X \times Y} [\phi(x) + \psi(y)] d\pi(x,y) = \int_{X} \phi(x) d\mu(x) + \int_{Y} \psi (y) d\upsilon(y)$$

for all $\phi, \psi$ in a suitable class of test functions.

My question has two parts:

  1. Where does this result come from? Does it stem from the fact that

$$\pi(A \times Y) = \mu(A) \, \text{and} \, \pi(X \times B) = \upsilon(B)$$ is equivalent to

$$\int_{X \times Y} \mathcal{X}_{A \times Y}(x,y) d\pi(x,y) = \int_{X} \mathcal{X}_{A}(x) d\mu(x) \, \text{and} \, \int_{X \times Y} \mathcal{X}_{X \times B}(x,y) d\pi(x,y) = \int_{Y} \mathcal{X}_{B}(x) d\upsilon(x)$$

  1. How do we know what class of test functions to choose? Villani says that the natural set of admissible test functions is $L^1(d\mu) \times L^1(d\upsilon)$, which he says is equivalent to $L^{\infty}(d\mu) \times L^{\infty}(d\upsilon)$ (I don't see this either), but he also says that the class can be narrowed down in certain situations.