Suppose that $\pi:Y\to X$ is a continuous surjection from one compact metric space to another. Given a regular probability measure $\mu$ on $Y$ with pushforward measure $\nu:=\pi^*\mu$, it is known that there is a family $(\mu_x)_{x\in X}$ of probability measures on $Y$ such that $x\mapsto \mu_x(B)$ is Borel measurable for each Borel subset $B$ of $Y$, such that almost every $\mu_x$ lives on the fiber above $x$, and such that $\int_Y f(y)d\mu(y)=\int_{x\in X}\int_{\pi^{-1}(x)}f(y)d\mu_x(y)d\nu(x)$.

My question is: suppose in advance that $\mathcal{O}$ is some fixed nonempty open subset of $Y$. Is it true that there is a probability measure $\mu$ as above for which $\mu(\mathcal{O})>0$ and such that there is a disintegration as in the previous paragraph for which the map $x\mapsto \mu_x(\mathcal{O})$ is upper semi-continuous? (A previous version of this question asked for continuity.) This is motivated by the fact that any Radon measure on a compact metric space is upper semi-continuous with respect to the Hausdorff metric.

  • $\begingroup$ How about $\mu := \delta_y$ for some fixed $y\in\mathcal{O}$? Then you can take the disintegration to be $\mu_x = \delta_y$ for every $x$, which trivially makes $x\mapsto \mu_x$ continuous and satisfies your requirements. But perhaps you want $\mu$ to have full support? $\endgroup$ May 5, 2018 at 6:44
  • $\begingroup$ Thanks to @TobiasFritz I understand that I do not understand what you're asking. What about the identity map $[0,1]\to [0,1]$? then the disintegration is $\mu_x=\delta_x$ for every $\mu$ and for every $\mathcal{O}\subsetneq [0,1]$, $x\mapsto \mu_x(\mathcal{O})$ is not continuous in any boundary point of $\mathcal{O}$. $\endgroup$
    – Uri Bader
    May 5, 2018 at 7:36

2 Answers 2


Let me start by addressing the question hinted on in the title, which in my view is more interesting than the question eventually asked in the body.

Given $\pi:X\to Y$ as above, we may endow the space $\text{Prob}(Y)$ with the weak* topology and consider its closed subspace $Z=\cup_x \text{Prob}(\pi^{-1}(x))$ consisting of all probability measures which images in $X$ are delta measures. Identifying $\delta_x$ with $x$ we consider the map $\pi_*:Z\to X$. The disintegration of a probability measure $\mu\in\text{prob}(Y)$ is a $\pi_*(\mu)$-a.e defined measurable map $X\to Z$, $x\mapsto \mu_x$, as described in the question. One may ask whether such a map could be represented as a continuous map (and that was my initial interpenetration of the question above).

The answer to this question is that in case $Y=X=[0,1]$ and $\pi:Y\to X$ is the Cantor-Lebesgue function (see https://en.wikipedia.org/wiki/Cantor_function) there are no local continuous sections $X\to Z$. Indeed, any such section is easily seen to be discontinuous at every diadic point of $X=[0,1]$. In particular, the disintegration function of any probability measure on $Y$ could not be represented by a continuous function.

Below is an edit answering the new version in the question's body.

Consider $X=Y=[-1,1]$ and $\mathcal{O}=(0,1]\subset Y$.

Let $\pi_1:Y\to X$ be the identity map. Then for any $\mu\in \text{Prob}(Y)$ we have that $\mu_x=\delta_x$, and the function $x\mapsto \mu_x(\mathcal{O})$ equals $\chi_{\mathcal{O}}$ which is lower semicontinuous, but not upper semi-continuous.

Let $\pi_2:Y\to X$ be given by defined by $\pi_2(x)=\min\{x,0\}$. Then for every $\mu\in \text{Prob}(Y)$ the function $x\mapsto \mu_x(\mathcal{O})$ equals $\mu(\mathcal{O})\cdot\delta_0$ and if $\mu(\mathcal{O})>0$ this function is upper semi-continuous but not lower semi-continuous.

We conclude that no semi-continuity is guaranteed in general.

  • $\begingroup$ I see how this shows that Lebesgue measure doesn’t work. But why couldn’t some other measure still satisfy the conclusion for some given open set? $\endgroup$
    – Isaac
    Apr 30, 2018 at 18:50
  • $\begingroup$ This is not about disintegration of a measure - it is about the fact that every map $X\to Z$, where $Z=\cup \text{Prob}(\pi^{-1}(x))$ endowed with the natural topology, is discontinuous at every diadic point of $X=[0,1]$. $\endgroup$
    – Uri Bader
    Apr 30, 2018 at 19:13
  • $\begingroup$ the above example works for every open set in $X$. if you want an example that continuity fails at just one point, you can use $Y$ being the graph of $\sin 1/x$ on $(0,1]$ union the segment $[-1,1]$ in the y-axis and $X$ being the segment $[0,1]$ in the x-axis. $\endgroup$
    – Uri Bader
    Apr 30, 2018 at 19:37
  • $\begingroup$ Is it possible that if I rephrased my question to merely ask for upper semi-continuity that it would then have a positive answer? Or does your example show that even this is not true? $\endgroup$
    – Isaac
    May 2, 2018 at 3:07
  • $\begingroup$ Not for any reasonable pre-order on the target space that I can think of right now (leaving aside too week ones like the pre-order given by inclusions of supports). And of course, you are most welcome, I am happy to help. $\endgroup$
    – Uri Bader
    May 4, 2018 at 6:08

I know one situation when this happens namely when $X,Y$ are smooth manifolds $\pi$ is a smooth map and $\mu$ is the volume measure defined by a Riemann metric $g_Y$ on $Y$, $\mu=dV_{g_Y}$.

Fix a Riemann metric $g_X$ on $X$ and assume that $\dim Y=N\geq n=\dim X$. Then the coarea formula shows the pushforward $\nu$ has the form $\newcommand{\eH}{\mathscr{H}}$ $$ \nu=\rho dV_{g_X},\;\;\rho(x)=\int_{\pi^{-1}(x)} \frac{1}{J_\pi(y)} d\eH_{N-n}(y), $$ where $J_\pi(y)$ is the Jacobian of the differential $d\pi(y): T_yY\to T_{\pi(y)} X$ and $\eH_k$ denotes the $k$-th dimensional Hausdorff measure on $Y$ defined by the metric $g_Y$

You can then define $\newcommand{\eO}{\mathscr{O}}$ $$ \mu_x (V):= \frac{1}{\rho(x)}\int_{\pi^{-1}(x)\cap V} \frac{1}{J_\pi(y)} d\eH_{N-n}(y),\;\;V\subset Y. $$ Let $U\subset X$ be the set of regular values of $\pi$. This is an open subset since $Y$ is compact. It is also dense. As set $\eO$ you can take $\eO=\pi^{-1}(U)$. If $x\in U$, then $\pi^{-1}(x)$ is a compact smooth submanifold of $X$ and $J_\pi\neq 0$ on $\eO$. The map

$$ U\ni x\mapsto \mu_x(\eO) $$ is continuous. For $x\in X\setminus U$ we have $\mu_x(\eO)=0$. If $\pi$ is a submersion then $U=X$.

  • $\begingroup$ That’s a nice example but unfortunately in my context I cannot assume that my spaces are manifolds. $\endgroup$
    – Isaac
    Apr 30, 2018 at 17:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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