Suppose I have two random variables $X,Y$ over some finite set and a function $f$ such that

$\forall z \in Image(f) \, Prob[f(X,Y)=z | X = x] = Prob[f(X,Y) = z]$

Intuitively, I would think that if I now consider the function $h(X,Y) = g(f(X,Y))$ I should have

$\forall w \in Image(h)\, Prob[h(X,Y) = w | X = x] = Prob[h(X,Y) = w].$

However, a simple counterexample shows this is not the case. If we choose $X,Y$ to be uniformly distributed over $\{0,1\}$ and $$f(X,Y) = (-1)^{X+Y}$$ then $f(X,Y)$ and $f(X,Y) | X = x$ both have the same distribution: they are uniformly distributed over $\{-1,1\}.$

However, if I take $g(z) = log_{-1}(z)$, then $h(X,Y) = g(f(X,Y)) = X + Y$ does not have the same distribution as $X + Y | X = x$. For example, we have that $Prob(X+Y=2) = 1/4$, but $Prob(X+Y = 2 | X = 0) = 0.$

What is going on in this counter-example that breaks the intuition? Is it that $\log_{-1}(z)$ is not a well-defined function (it has multiple complex branches)? Is it because $\log_{-1}(z)$ is discontinuous for $z = -1$? Or is the intuition just not true?

  • $\begingroup$ Since you have a counter-example, the intuition is incorrect. Logarithms, complex numbers, and continuity are irrelevant in your example: everything in it could (easily) be described without reference to those things. $\endgroup$ – whuber Aug 15 '16 at 21:01
  • $\begingroup$ Thanks for your comment. I would agree, but the counter-example seems contrived and dependent on a function $\log_{-1}(z)$ which to my knowledge might not even exist. Originally, when I wrote the question, I wrote $f(X,Y) = X + Y \bmod 2$ (with $X,Y$ uniformly distributed over $\mathbb{Z}_2$ and $h(X,Y) = X + Y$ (with $X,Y$ ordinary integers distributed uniformly over $\{0,1\}$.) However, there does not seem to be a function $g(z)$ such that $g(X+Y \bmod 2) = X+Y$ that can be used to define $h(X,Y)$. $\endgroup$ – Asterix Aug 15 '16 at 21:06
  • 1
    $\begingroup$ There's no need to call $g$ a logarithm. If I have understood your exposition, $g$ merely is the function that maps $1$ to $0$ and $-1$ to $1$. Unfortunately your description following "if I take" is murky because it includes a false series of equations: it is not the case that $X+Y$ can be recovered from $f(X,Y)$. $\endgroup$ – whuber Aug 15 '16 at 21:10

The "intuitive" claim in the beginning is correct: \begin{equation} P(h(X,Y) = w \mid X = x) = P(f(X,Y) \in \{z \mid g(z)=w\} \mid X = x) \\ = \sum_{z_i \in \{z \mid g(z)=w\}} P(f(X,Y) = z_i \mid X=x) = \sum_{z_i \in \{z \mid g(z)=w\}} P(f(X,Y) = z_i) = P(f(X,Y) \in \{z \mid g(z)=w\}) = P(h(X,Y) = w). \end{equation}

An alternative way of stating it would be that if $f(X,Y)$ and $X$ are independent, then $g(f(X,Y))$ and $X$ are independent. Applying a deterministic mapping to one of independent random variables cannot lose independence.

The counterexample does not work since there is no function $g$ with the property that $g((-1)^{(X+Y)})=X+Y$ for all values $X+Y$ may attain here (as @whuber pointed out in the comments). The example is equivalent to your first try (mentioned in comments) with $f(X,Y)=X+Y~\mathrm{mod}~2$ upto relabeling the range of $f$. As you found that there was no function in that case the first version, there is no applicable function in the new case, either.

$\mathrm{log}_{-1}$ not being such a function could be interpreted as being due to it having 'multiple branches': if we want $(-1)^w = z$ to imply $\mathrm{log}_{-1}(z)=w$, $\mathrm{log}_{-1}(1)$ should equal both $0$ and $2$, which is a proof-by-contradiction for nonexistence of such a function. We may decide to define if as a multivalued function or perhaps define a function of $(X,Y)$ that picks the correct "branch" based on $X+Y$ -- but then the premises of the initial claim do not hold.


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.