I am trying the find $x^*$ defined as: $$ x^* = argmax_x \int f(x,y)p(y)dy $$ where $f(x,y)$ can be an arbitrary positive function and $p(y)$ is a probability density.

Does the location of $x^*$ hold true under a log transformation of $f$? Specifically, does the previous equation imply: $$ x^* = argmax_x \int log \{f(x,y)\} p(y)dy $$

I doubt that this is true, because of the $p(y)$ term in the integral, but I can't quite pin down why not. A formal proof is not necessary - some intuitive explanation (or counterexample) would suffice.

  • $\begingroup$ $\log(f(x, y))$ needs not be integrable to begin with. $\endgroup$ – user251257 Feb 16 '16 at 1:17
  • $\begingroup$ True, but in this case lets assume that the log transformed integrand has an analytical solution (which is one of the reasons we want to do a log transform), while the original formulation may or may not have one. $\endgroup$ – Supratik Paul Feb 16 '16 at 11:51

On $[\frac12, 1]^2$ let $f(x,y) = y$ for $\frac12\le y \le x \le 1$, $f(x,y) = 0$ else. Let $p(y) = 2$ for $\frac12\le y \le 1$. Then, we have $$ g(x) := \int_{1/2}^1 f(x, y) p(y) dy = 2\int_{1/2}^x \underbrace{y}_{\ge 0}\, dy $$ and $\operatorname{argmax} g = 1$. On other hand, we have $$ h(x) := \int_{1/2}^1 \log(f(x,y)) p(y) dy = 2\int_{1/2}^x \underbrace{\log(y)}_{\le 0} dy $$ and thus $\operatorname{argmax} h = \frac{1}{2}$.


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.