5
$\begingroup$

Let $X \subset \mathbb{R}^n$ be compact and consider the measurable space $(X, \mathcal{B}(X))$ where $\mathcal{B}(X)$ is the Borel $\sigma-$algebra over $X$. Denote by $\mu$ the Lebesgue measure on this space. Let $\nu$ be another measure on this space such that, $\nu \ll \mu$, i.e., it's absolutely continuous w.r.t $\mu$.

Can we always find a measurable mapping $T: X \to X$ such that $T_\# \mu = \nu$, i.e., such the push-forward measure induced by $T$ is equivalent to $\nu$? If not, what conditions one needs to impose on $\nu$ for this to hold?

I am not sure how to approach this, but my intuition is that $T$, if it exists, should be non-singular.

$\endgroup$
1
  • 2
    $\begingroup$ Al the least one should assume not only that $\nu\ll \mu$ but also that $\nu(X)=\mu(X)$ otherwise the answer is no in general. Suppose $\nu$ is such that $\nu(X)<\mu(X)$, then no push-forward of $\mu$ is $\nu$: $\mu(T^{-1}(X))=\mu(X)$. $\endgroup$ Commented Feb 17, 2023 at 13:38

2 Answers 2

Reset to default
3
$\begingroup$

Coincidence of the total measures is an obstruction; morally this should be the only obstruction (in a variety of settings (often when there is more regularity to worry about) this is often named a "Moser Lemma" ($\dagger$)).

Indeed, first let's assume that $X\subseteq \mathbb{R}^n$ is bounded, so that $\text{leb}(X)<\infty$; wlog we may also assume $\text{leb}(X)>0$. Then for any measure $\nu$ on $X$ with $\nu(X)=\text{leb}(X)$, there is a bimeasurable ($\triangle$) $T:X\to X$ such that $T_\ast(\nu)=\text{leb}|_X$ as a corollary of the theorem mentioned in my answer at Is there always a mapping between probability measures on the $n$-sphere?. More generally to apply this theorem all that is required is that $\nu$ assigns zero measure to any point in $X$ ($\star$).

In the case of unbounded $X$ the situation is more subtle. For finite $\nu$ the above argument again applies by one-point compactifying $\mathbb{R}^n$.

More generally if $\nu$ is $\sigma$-finite and if there is a partition $\{X_m\}_m$ of $X$ into countably many pieces with $\nu(X_m)=\text{leb}(X_m)$, or even two partitions $\{X_m\}_m$, $\{X_m'\}_m$with $\nu(X_m')=\text{leb}(X_m)$.

(In an earlier version of this answer I had claimed that $X_m = \{x\in X| m\leq |x|<m+1\}$ would work, which is not correct in general.)

$(\dagger)$ Moser had shown in his paper "On the Volume Elements on a Manifold" the (roughly) following: Let $\omega,\eta$ be two volume elements on a smooth manifold $M$, there is a diffeomorphism $f:M\to M$ with $f^\ast(\omega)=\eta$ iff $\int_M \omega=\int_M \eta$.

Added:

($\triangle$) Thus $T:X\to X$ is a measurable bijection such that $T^{-1}:X\to X$ is also measurable.

($\star$) That is, the condition "for any $x\in X$, $\nu(\{x\})=0$" is sufficient for the argument.

$\endgroup$
5
  • $\begingroup$ Thanks for your answer. So if I understood correctly, you are saying (using my notation), that for any measure $\nu$ such that $\nu(X) = \mu(X)$ we can always find a bijective mapping $T:X \to X$ such that $T_{\#} \nu = \mu$ which is equivalent (?) to $T^{-1}_{\#} \mu = \nu$? And you seem to be saying that this relationship is if and only if, i.e., such a $T$ only exists if $\nu(X) = \mu(X)$. $\endgroup$ Commented Feb 17, 2023 at 14:23
  • $\begingroup$ "More generally to apply this theorem all that is required is that ๐œˆ assigns zero measure to any point in ๐‘‹" I don't understand this statement. Do you mean that $\nu$ should have the same null sets of $\mu$? that is, $\nu(A) = 0 $ for all $A$ such that $\nu(A) = 0$? $\endgroup$ Commented Feb 17, 2023 at 14:24
  • $\begingroup$ Finally, I don't see how this follows as a corollary from the theorem in your other answer. Because, there, you talk about bijective mapppings from $X$ to $[0,1]$. $\endgroup$ Commented Feb 17, 2023 at 14:26
  • 2
    $\begingroup$ @Saleh I've added some clarifying statements. For your last question, I take two bimeasurable maps $A,B:X\to [0,1]$ such that the first takes $\nu$ to the Lebesgue measure on the interval and the second takes $\text{leb}$ to Lebesgue measure on the interval, then the bimeasurable map $B^{-1}\circ A$ carries $\nu$ to $\text{leb}$, or alternatively $A^{-1}\circ B$ carries $\text{leb}$ to $\nu$. $\endgroup$ Commented Feb 17, 2023 at 15:05
  • 1
    $\begingroup$ Thanks. The theorem you cite applies for probability distributions, but since we have finite measures and assume $\nu(X)=\mu(X)$ we can get to the same setting probably via some renormalization argument, right? $\endgroup$ Commented Feb 17, 2023 at 15:54
3
$\begingroup$

The statment seems to hold for probability measures on Borel spaces under the assumption that $\mu$ is really the Lebesgue measure on $(X,\mathscr{B})$.

Definition: A measure space $(X,\mathscr{B})$ is Borel if there is a bijective map $\phi:(0,1)\rightarrow X$ such that $\phi$ is $\mathcal{B}((0,1))/\mathscr{B}$ -measurable and $\phi^{-1}$ is $\mathscr{B}/\mathcal{B}(0,1)$ -measurable. Such a map $\phi$ is called measurable isomorphism.

Uncountable Polish spaces for example are Borel spaces (Parthasarathy, K.L., Probability on Metric spaces, AMS Chelsea, 2005 (reprint of 1967 edition)).

Suppose $(X,\mathscr{B})$ is a Borel space and let $\phi$ be as in the definition above. Let $\lambda$ denote the Lebesgue measure restricted to $((0,1),\mathcal{B}(0,1))$. A probability measure $\mu$ on $(X,\mathscr{B})$ is said to be of Lebesgue type if $\mu=\lambda\circ \phi^{-1}$.

For such $\mu$, the problem in the OP then reduces to considering the space $((0,1),\mathcal{B}(0,1),\lambda)$. For any probability measure $\nu$ define $F_\nu(x)=\nu((0,x])$, $0<x<1$, and define the quantile function $Q_\nu:(0,1)\rightarrow\mathbb{R}$ as $$Q_\nu(q)=\inf\{x: F_\nu(x)\geq q\}$$ It can be proven that $$F_\nu(x)\geq q \quad\text{iff}\quad Q_\nu(q)\leq x$$ Hence $$\lambda\big(\{q:Q_\nu(q)\leq x\}\big)=\lambda\big(\{q: F_\nu(x)\geq q\}\big)=F_\nu(x)$$ This means that $\lambda\circ Q^{-1} =\nu$. Notice that in fact $Q(0,1)\subset(0,1)$; thus the push forward of $\lambda$ by $Q$ is $\nu$. Notice that whether $\nu\ll\lambda$ or not does not play a role here.

Edit: If $\mu$ is not equivalent to the Lebesgue measure on $((0,1),\mathcal{B}(0,1))$, then the answer may be no. Consider the measure $\mu=\frac12\lambda+\frac12\delta_{1/2}$ and let $\mu=\lambda$. Clearly $\lambda\ll\mu$. Suppose $T:((0,1),\mathcal{B}(0,1))\rightarrow((0,1),\mathcal{B}(0,1))$ and let $p_0=T(1/2)$. Then $$0=\lambda(\{p_0\})<\frac12\leq \mu\big(T^{-1}(\{p_0\})\big)$$ which shows that no push forward of $\mu$ is $\nu$.

More interestingly perhaps are the cases:

  1. $X$ is a topological locally compact Hausdorff group equipped with the Borel $\sigma$-algebra $\mathscr{B}(X)$, and that admits a finite Haar measure, say $\mu(X)=1$. Is any Borel measure $\nu$ on $(X,\mathscr{B}(X))$ a push forward of $\mu$? The answer in this case escapes me, even for under the assumption $\nu\ll \mu$.
  2. $M$ is a compact manifold of dimension $m\geq 1$ embedded in $\mathbb{R}^n$, for some $n\geq m$, one my also look at the Hausdorff measure $\mu=H^m$ restricted to $(M,\mathscr{B}(M))$. In this case, the answer to the same question above also escapes me.
$\endgroup$
5
  • $\begingroup$ thanks for the answer. Still need some time to grasp the details but the definition you provide looks a bit weird to me. Is it some sort of characterisation of Borel spaces? Wiki page on Borel spaces mentions that a measurable bijection between two Borel spaces is actually bi-measurable. It looks to me that this definition is similar but I'm not sure. For the generic case of of finite measures $\mu(X) = \nu(X) = C \neq 1$ the same result should hold with some renormalization argument, correct? $\endgroup$ Commented Feb 17, 2023 at 16:23
  • 1
    $\begingroup$ @Saleh: what you quote is a (deep) theorem not a definition. That uncountable Polish spaces are Borel spaces (in the sense that I describe in my posting) is also a deep theorem. In any case, the idea of such spaces is to show that it suffices to consider the canonical space $((0,1),\mathcal{B}(0,1),\lambda)$. $\endgroup$ Commented Feb 17, 2023 at 16:26
  • $\begingroup$ The definition you provide also looks more like a property or result to me than a definition. Is it provided in the reference you cite (Parthasarathy)? But anyway, I understood that the goal is considering the canonical space instead. $\endgroup$ Commented Feb 17, 2023 at 16:32
  • 1
    $\begingroup$ @No, it is actually a particular instance of the following definition: Two measurable spaces $(X,\mathscr{B})$ and $(Y,\mathscr{C})$ are (measurable) isomorphic of there is a bijection $\phi:X\rightarrow Y$ such that $\phi$ is $\mathscr{B}/\mathscr{C}$-measurable and $\phi^{-1}$ is $\mathscr{C}/\mathscr{B}$-measurable. $\endgroup$ Commented Feb 17, 2023 at 16:36
  • 1
    $\begingroup$ @Saleh: you may want to repost your question in MO where the professionals can handle this question in a perhaps more general settings: Haar measures in nice topological groups, measures on Remannian manifolds, Hausdorff measure of the right dimension embedded on an Euclidean space, etc. $\endgroup$ Commented Feb 17, 2023 at 16:46

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.