9
$\begingroup$

Let $X,Y$ be locally compact spaces (in my specific case, they are locally compact groups). Suppose that we are given a measure $\mu$ on $X$ and a finite number of quotient maps $p_1,\ldots,p_n:Y\rightarrow X$. Under what conditions on $\mu, p_1,\ldots,p_n$ is there a measure $\eta$ on $Y$ such that the marginals $(p_1)_\ast\eta=\ldots=(p_n)_\ast\eta=\mu$?

EDIT clarification about the type of quotients needed. In the specific problem I am trying to solve, the space $Y=G\times G\times G$ and $X=G\times G$ where the measure $\mu$ is a given Radon probability measure on $X$. The quotients are given by
$p_1(x,y,z)=(x,y)$
$p_2(x,y,z)=(z,y)$
$p_3(x,y,z)=(x,xyz^{-1})$
$p_4(x,y,z)=(z,xyz^{-1})$

So the crucial difficulty is that the quotient maps do not come from a product decomposition of $Y$.

I would like to have a lemma of the form:
lemma If the quotient maps $p_1,\ldots,p_n$ satisfy the following conditions:
$\ldots$
then for every probability measure $\mu$ on $X$, there exists a probability measure $\nu$ on $Y$ such that $(p_i)_\ast\nu=\mu$.

EDIT2 Changed the title to be more in line with common terminology, and clarified the question. Thanks to Benoît.

$\endgroup$
2
  • $\begingroup$ The use of the word "marginal" seems inappropriate, and is at least very misleading as it is usually used when $p_i$ are projection from a Cartesian product. "Push forwards" seems a better name. $\endgroup$ Nov 23, 2013 at 17:54
  • 4
    $\begingroup$ You should state what kind of condition you seek, or what kind of properties you have on $\mu$, to avoid too weak answers (like: when $\mu$ is a Dirac at $x$ and $p_i$ are invertible and the $p_i^{-1}(x)$ do not match, there is no such $\nu$). $\endgroup$ Nov 23, 2013 at 17:59

3 Answers 3

1
$\begingroup$

This kind of problem has been studien in:

J. Hoffmann-Jørgensen The general marginal problem Functional Analysis II Lecture Notes in Mathematics Volume 1242, 1987, pp 77-367

$\endgroup$
0
$\begingroup$

This is not a complete answer, just a simple example showing what kind of things can go wrong.

Consider $Y=A^3$ and $X=A^2$ with $A$ a non-trivial measurable space, and let $p_1(x,y,z)=(x,y)$ and $p_2(x,y,z)=(y,z)$. Then when $\mu$ a probability on $X$ does not have equal marginals, there can be no $\nu$ on $Y$ such that $p_i(\nu)=\mu$.

$\endgroup$
0
$\begingroup$

I hope this helps. For two projections $p_1$ and $p_2$, if there exists such an $\eta$ then $\eta( p_1(S_Y) ) = \eta( p_2( S_Y) )$ for all measurable sets $S_Y \subset Y$. Upon setting $S_1 = p_1(S_Y)$ for an arbitrary $S_Y \subset Y$ we see that $\eta$ must satisfy $\eta( S_1) = \eta( p_2 \circ p_1^{-1}(S_1) )$ for any measure able set $S_1 \subset X$. In otherwords $(p_2 \circ p_1^{-1})^* \eta = \eta$. So we see that the answer to the question is yes if $(p_2 \circ p_1^{-1})^*$ has fixed points. For $n$ projections we need to check $(p_i \circ p_j^{-1})^*$ for $i,j = 1,\dots,n$. This is for a given fixed $\mu$ on $Y$. If we want an $\eta$ for every $\mu$ my guess is that this happens if and only if $(p_i \circ p_j^{-1})^*$ is the identity. However, this would mean that $p_i = p_j$ for all $i,j$. In the case presented here we find $p_3 \circ p_1^{-1}(x,y) = (x,G)$ while $p_1 \circ p_2^{-1}(x,y) = (G,y)$. The first calculation tells us that $\eta$ must be independent of the second component in $G \times G$ while the second calculation tells us that $\eta$ must be independent of the first component of $G \times G$. This $\eta$ must be very restricted, to say the least, and this restricts the class of possible $\mu$'s.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.