Domination for (finite) product measures equivalent to domination for marginals?

80 Views Asked by At

Let $\Omega=\Omega_1\times\Omega_2$ be a set equipped with a partial ordering that is inherited from partial orderings on $\Omega_1\times\Omega_2$. I.e. $(\omega_1,\omega_2)=\omega\leq\phi=(\phi_1,\phi_2)$ if and only if $\omega_1\leq\phi_1$ and $\omega_2\leq\phi_2$. An increasing set $I$ is a set where $\omega\in I$ and $\omega\leq\phi$ implies that $\phi\in I$.

Let $\nu=\nu_1\times\nu_2$ and $\mu=\mu_1\times\mu_2$ be product probability measures whose marginals are also probability measures. I am interested in when $\nu\preccurlyeq\mu$. This is a stochastic domination or stochastic ordering (more generally domination or ordering of measures). It is equivalent to $\nu(I)\leq\mu(I)$ for any increasing set $I$. Assume that the marginal sigma algebras have whatever properties they need to have, e.g. the sigma-algebra for $\nu_j$ is the same as for $\mu_j$ should be fine. Although, I would like to know the answer in a more general case.

Question: Is $\nu_1\times\nu_2\preccurlyeq\mu_1\times\mu_2$ equivalent to $\nu_1\preccurlyeq\mu_1$ and $\nu_2\preccurlyeq\mu_2$?

Here is my attempt at a proof.

$\nu_1\times\nu_2\preccurlyeq\mu_1\times\mu_2$ if and only if for an arbitrary increasing set $I\subset\Omega$ we have $(\nu_1\times\nu_2)(I)\preccurlyeq(\mu_1\times\mu_2)( I)$. Clearly $I_1\times\Omega_2$ is an increasing set when $I_1\subset\Omega_1$ is an increasing set. This gives us that $$\begin{aligned} \nu_1(I_1)&=\nu_1(I_1)\cdot\nu_2(\Omega_2)\\ &=(\nu_1\times\nu_2)(I_1\times\Omega_2)\\ &\leq(\mu_1\times\mu_2)(I_1\times\Omega_2)\\ &= \mu_1(I_1)\cdot\mu_2(\Omega_2)\\ &= \mu_1(I_1). \end{aligned} $$

So clearly $\nu_1\preccurlyeq\mu_1$ and we can likewise get that $\nu_2\preccurlyeq\mu_2$. Here I used that we are working with probability measures which makes me think the question is answered "no" for at least some product measures.

On the other hand, if we assume $\nu_j\preccurlyeq\mu_j$ for $j=1,2$. Then for an arbitrary increasing set $I\subset\Omega$ we can partition $I$ into "rectangular" sets $(A_1\times I_1)\cup (A_2\times I_2)\cup\cdots\cup (A_N\times I_N)$. We assume that each $A_j$ and $I_j$ are measurable by the appropriate $\nu_i, \mu_i$. The $A_j$ on the left partition the projection of $I$ onto $\Omega_1$ (denote that by $\pi_1(I)$), and the $I_j$ on the right are each increasing subsets of $\Omega_2$. Let's choose the partition fine enough so that

$$\nu(I)\approx \sum_{j=1}^N \nu(A_j\times I_j) ~ \text{ and } ~ \mu(I)\approx \sum_{j=1}^N \mu(A_j\times I_j)$$

Then we get

$$\begin{aligned} \nu(I) &=\int_{\pi_1(I)} \nu_2(I^x)d\nu_1(x)\\ &\approx\sum_{j=1}^N \nu(A_j\times I_j)\\ &=\sum_{j=1}^N \nu_1(A_j) \nu_2(I_j)\\ &\leq \sum_{j=1}^N \nu_1(A_j) \mu_2(I_j)\\ &\approx \int_{\pi_1(I)} \mu_2(I^x) d\nu_1(x)\\ &=\mu(I). \end{aligned}$$

Hence we conclude that $\nu_1\times\nu_2\preccurlyeq\nu_1\times\mu_2$. Similarly we can show that $\nu_1\times\mu_2\preccurlyeq\mu_1\times\mu_2$. This gives us $$\nu=\nu_1\times\nu_2\preccurlyeq\nu_1\times\mu_2\preccurlyeq\mu_1\times\mu_2=\mu.$$

I figure this argument could be more carefully formalized with $\epsilon$'s etc. My question is if the basic ideas seem correct. I know I hand-waved some things in there.

If this works, does it also work for infinite products? I'm interested in nice cases, say $\{0,1\}^{\mathbb N}$ with appropriate metric and/or topology, or more generally a Polish space or complete, separable metric space, something like that. My knowledge at this level is fairly rough...

1

There are 1 best solutions below

3
On

The second part of your proof is not rigorous. Here is an alternative. First observe that for every measure $\nu$ on $(\Omega,\mathcal F)$, Fubini's theorem yields that $$\int_\Omega f \, d\nu = \int_\Omega \int_0^\infty {\bf 1}_{\{t \le f(x)\}} dt \, d\nu(x)= \int_0^\infty \int_\Omega {\bf 1}_{\{t \le f(x)\}} \, d\nu(x) \, dt \,.$$

Therefore, given measures $\nu,\mu$ on $(\Omega,\mathcal F)$, we have $\nu\preccurlyeq\mu$ iff all increasing positive measurable functions $f: \Omega \to \mathbb R$ satisfy $ \int_\Omega f \, d\nu \le \int_\Omega f \, d\mu$.

Now suppose that $\nu_i\preccurlyeq\mu_i$ on $\Omega_i$ for $i=1,2$. Given an increasing positive measurable function $g$ on $\Omega_1 \times \Omega_2$ endowed with the product $\sigma$-algebra, define for $x \in \Omega_1$ $$g_1(x):=\int_{\Omega_2} g_1(x,y) \, d\nu_2$$ and $$G_1(x):=\int_{\Omega_2} g_1(x,y) \, d\mu_2$$ so that $g_1 \le G_1$ are two positive increasing measurable functions on $\Omega_1$. We infer that $$ \int_{\Omega_1 \times \Omega_2} g(x,y) \, d\nu_2 \, d\nu_1 = \int_{\Omega_1} g_1(x) \, d\nu_1 \le \int_{\Omega_1} G_1(x) \, d\nu_1 $$ $$\le \int_{\Omega_1} G_1(x) \, d\mu_1 =\int_{\Omega_1 \times \Omega_2} g(x,y) \, d\mu_2\, d\mu_1 \,. $$