Let $F: \mathbb{R} \to [0,1]$ be cumulative distribution function (cdf). The standard way to build a random variable $\tau$ on $([0,1],\mathcal{B},\text{Leb})$ with $F$ as its cdf is using the Skorokhod representation: For $x \in [0,1]$, let $$ \tau(x) := \inf \{ z : F(z) > x \} $$ be the generalized inverse of $F$.

I'm looking for a more general version of this. Let $X = \mathbb{R}^n$ (to begin with), and let $\mu$ be a probability measure on it. How does one build a random variable $\tau$ on $([0,1],\mathcal{B},\text{Leb})$ that takes values in $X$, so that the pushforward measure $\text{Leb}_* (\tau)$ is $\mu$?

Can this be done if $X$ is a Polish space and $\mu$ is a Borel measure? My feeling that this is generally true, and the result is probably due to von Neumann. I can't remember the right reference though.

Can anyone help? Note that a statement like "an isomorphism between these measure spaces exists" is not what I'm looking for. I want an explicit way to construct $\tau$.

  • $\begingroup$ The isomorphism of measurable spaces is constructive, is it not? So following that proof should let you construct $\tau$. $\endgroup$ Mar 12, 2014 at 1:59
  • $\begingroup$ You're probably right, although I haven't read the proof of the isomorphism theorem (Polish spaces with Borel measures are isomorphic to an interval $[a,b)$). $\endgroup$
    – arjun
    Mar 12, 2014 at 17:13

1 Answer 1


The statement is true for probability measures on a complete separable metric space. I made a slight variation of the Skorokhod representation theorem in my PhD thesis (way long time ago!) so I am more or less cutting and pasting, sorry for the errors. I got the original proof from Ikeda, Watanabe, Stochastic differential equations and diffusion processes, North Holland Mathematical Library 24, North Holland/Kodansha 1989.

Let $(S,d)$ be a complete separable metric space, and $\mu$ a probability measure on $S$ (on the Borel $\sigma$-algebra of $S$). Set $\Omega=[0,1)$, $\mathcal{F}$ the Borel $\sigma$ algebra of $[0,1)$, and $\mathcal{L}$ the Lebesgue measure on $[0,1)$.

Step 1: construction of suitable families of subsets of $S$. For each $k\geq0$ let $(B_m^k)_{m\geq1}$ be balls with radius less than $2^{-(k+1)}$ which cover $S$ and such that $\mu(\partial B_m^k)=0$. This is possible since $\mu$ is a finite measure, hence in each point only a countable set of radii gives balls with ``fat'' boundaries. Set $$ D_1^{k} = B_1^k, \qquad D_2^k = B_2^k\setminus B_1^k, \quad \ldots, \quad D_n^k = B_n^k\setminus\bigcup_{i=1}^{n-1}B_i^k, \quad \ldots $$ and $$ S_{i_1\ldots i_k}=D_{i_1}^1\cap D_{i_2}^2\cap\ldots\cap D_{i_k}^k. $$ It can be easily seen that

  • if $(i_1,\ldots, i_k)$ and $(i_1',\ldots, i_k')$ are different, then $S_{i_1\ldots i_k}$ and $S_{i_1'\ldots i_k'}$ are disjoint,
  • $\bigcup_i S_i=S$ and $\bigcup_i S_{i_1\ldots i_ki}=S_{i_1\ldots i_k}$,
  • the diameter of $S_{i_1\ldots i_k}$ is $\leq 2^{-k}$,
  • $\mu(\partial S_{i_1\ldots i_k})=0$.

Order the $k$-tuples $S_{i_1\ldots i_k}$ by lexicographic order.

Step 2: Define a family of sub--intervals $\Delta_{i_1\ldots i_k}$ (of the type $[a,b)$ in $[0,1)$), such that $$ \mathcal{L}(\Delta_{i_1\ldots i_k}) = \mu(S_{i_1\ldots i_k}), $$ such that $\Delta_{i_1\ldots i_k}$ is on the left of $\Delta_{i_1'\ldots i_k'}$ if $(i_1,\ldots, i_k)<(i_1',\ldots,i_k')$, and such that together they cover $[0,1)$.

Step 3: Define the random variable. For each $(i_1,\ldots, i_k)$ such that $S_{i_1\ldots i_k}$ is non-empty, choose a point $x_{i_1\ldots i_k}$ in the interior of $S_{i_1\ldots i_k}$. For each $\omega\in[0,1)$ set $X_k(\omega)=x_{i_1\ldots i_k}$ if $\omega\in\Delta_{i_1\ldots i_k}$. The random variable is well defined since, if $\Delta_{i_1\ldots i_k}$ is non-empty, then the interior of $S_{i_1\ldots i_k}$ is non-empty. Observe that $d(X_k(\omega),X_{k+p}(\omega))\leq\frac1{2^k}$, hence by the completeness, the limit $X(\omega)=\lim_k X_k(\omega)$ always exists.

Step 4: $X$ has law $\mu$. Indeed $$ \mathbb{P}[X_{k+p}\in\bar S_{i_1\ldots i_k}] = \mathbb{P}[\omega:\omega\in\Delta_{i_1\ldots i_k}] = \mu(S_{i_1\ldots i_k}), $$ and, since each open set $A$ in $S$ is a disjoint countable union of sets $S_{i_1\ldots i_k}$, by Fatou, $\mu(A)\leq\liminf_k\mathbb{P}[X_k\in A]$, that is $X$ has law $\mu$.


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.