If $Z=f(X)$ for random variables, does $p(y,z)$ have the same maxima as $p(y,x)$ up to the transform $f$?

17 Views Asked by At

Assume real valued random variables $X,Y,Z$ such that $Z = f(X)$ for an injective, smooth $f$ with ${f^{-1}}'\neq 0$.

Suppose we have a joint density $$p(y,x) = p(y|x) p_x(x) = g(y,f(x))p(x)$$ for some $g$ (for example $g = \mathcal{N}(y,g(x),\sigma)$).

We then also have a joint density

$$p(y,z)=p(y|z)p_z(z) = g(y,z)p_x(f^{-1}(z)) |{f^{-1}}'(z)|$$

Now intuitively, the maxima of these joint densities should coincide, i.e. if for fixed $y$, $x^*$ is a maximum of $p(y,x)$ then also $z^*=f(x^*)$ should be a maximum of $p(y,z)$. But it seems not to be correct:

If we apply logarithms to both densities and derivate, we obtain

$$\frac{d}{dx} \ln p(y,x) = (\frac{d}{du}\ln g(y,f(x)))f'(x) + \frac{d}{dx} \ln p(x)$$ $$ \frac{d}{dx} \ln p(y,z) = (\frac{d}{du}\ln g(y,z)) + (\frac{d}{dx}\ln p_x(f^{-1}(z))) ({f^{-1}}'(z)) + \frac{d}{dz} \ln |{f^{-1}}'(z)| ,$$ where by $du$ I denoted derivating w.r.t. the second component for clarity.

Now, I don't really see how to proceed to show equality. For example, the second equation is zero iff $$(\frac{d}{du}\ln g(y,z))\frac{1}{({f^{-1}}'(z))} + (\frac{d}{dx}\ln p_x(f^{-1}(z))) + \frac{\frac{d}{dz} \ln |{f^{-1}}'(z)|}{{f^{-1}}'(z))}=0$$

so now plugging in $z^* = f(x*)$ would yield

$$ 0 + 0 + \frac{\frac{d}{dz} \ln |{f^{-1}}'(z^*)|}{{f^{-1}}'(z^*))},$$

where the last term does not seem to be $0$ necessarily.