Suppose that $X,Y$ are scalar random variables supported on some standard Lebesgue probability space $(\Omega, \mathrm{P})$, such that $X \overset{\mathrm{d}}{=} Y$ in the sense that their pushforward measures are equal, $X_*(\mathrm{P}) = Y_*(\mathrm{P})$. Does there exist a nondegenerate random variable $Z$ on $(\Omega, \mathrm{P})$ satisfying $X + Z \overset{\mathrm{d}}{=} Y + Z$?

In the case that $Y = X\circ T$ for some measure preserving automorphism $T: \Omega \rightarrow \Omega$ (this appears to be often the case by: Random variables with same distribution), and we have in addition the factorization $T = S^2$ for some automorphism $S: \Omega \rightarrow \Omega$, then clearly we can set $Z = X \circ S$, whence

$X + Z = X + X\circ S \overset{\mathrm{d}}{=} X \circ S + X \circ S^2 = Z + Y$.

Are there more general conditions than this, or perhaps conditions under which the answer to the question is negative?

Edit: it also turns out that if $Y = X \circ T$ where $T$ is measure preserving and non-ergodic then $Z$ exists; just take $Z$ such that $Z \circ T = Z$ almost surely and $Z$ non-constant. Then a computation like the one above shows equality in distribution.


There is$^*$ a counterexample in the atomic case, see below, so we will assume that $(\Omega, \mathrm{P})$ is a non-atomic standard Lebesgue probability space (so it is Isomorphic to the unit interval with Lebesgue measure, see https://en.wikipedia.org/wiki/Standard_probability_space). In that setting we claim:

If $X \overset{\mathrm{d}}{=} Y$ are two random variables on $(\Omega, \mathrm{P})$ , then there always exists a nondegenerate random variable $Z$ on $(\Omega, \mathrm{P})$ satisfying $X + Z \overset{\mathrm{d}}{=} Y + Z$.

Proof: We consider three cases.

Case 1 : $X+Y$ is not an almost sure constant. Then defining $Z:=-X-Y$ proves the claim.

Case 2: $X+Y=c$ almost surely but $X$ takes more than 2 values a.s. (That is, for every pair of reals $a,b$ we have $ \mathrm{P}[X \in\{a,b\}]<1.$ ) Then $Z:=X(c-X)$ is nondegenerate. Denoting $f(x)= x+x(c-x)$, we have $f(X)=X+Z$ and $f(Y)=Y+Z$ almost surely, so $X + Z \overset{\mathrm{d}}{=} Y + Z$.

Case 3: $X+Y=c$ almost surely, and there exist $a,b$ such that $ \mathrm{P}[X \in\{a,b\}]=1.$ In this case let $D_a$ be a subset of $\{\omega: X(\omega)=a\}$ with $\mathrm{P}(D_a)=\mathrm{P}(\{\omega: X(\omega)=a\})/2$. (To see that $D_a$ exists we may assume that $(\Omega, \mathrm{P})$ is the unit interval with Lebesgue measure, and observe that the mapping $t \mapsto \mathrm{P}([0,t] \cap \{\omega: X(\omega)=a\})$ is continuous on $[0,1]$ so its image includes $\mathrm{P}(\{\omega: X(\omega)=a\})/2$.) Similarly, let $D_b$ be a subset of $\{\omega: X(\omega)=b\}$ with $\mathrm{P}(D_b)=\mathrm{P}(\{\omega: X(\omega)=b\})/2$. Let $Z(\omega)=1$ if $\omega \in D_a \cup D_b$ and $Z(\omega)=0$ otherwise. Then $Z$ is independent of $X,Y$ so $X + Z \overset{\mathrm{d}}{=} Y + Z$.


(*) Counterexample in the atomic case: If $\Omega=\{0,1\}$ with the uniform measure and maximal $\sigma$-algebra, and $X(\omega)=\omega=1-Y(\omega)$, then for every nonconstant random variable $Z$ on $\Omega$, the laws of $X+Z$ and $Y+Z$ are different; e.g., if $Z(1)>Z(0)$ then the maximum of $X+Z$ exceeds the maximum of $Y+Z$.


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.