So I was trying to prove that the sum of two log convex functions is also log convex. In my attemp I was able to conclude that if

$$f(x)^{1-t} f(y)^t + g(x)^{1-t} g(y)^t \leq (f(x)+g(x))^{1-t} (f(y)+g(y))^t $$

holds for any $t \in [0,1]$ then the proof is completed. I found that another user has asked how to prove the same question as me. This is the question How to prove that the sum of two log-convex functions is log-convex?

The first answer by the user Xiang Yu gets to the same point as me, and then he assume that $f(x)+g(x)=f(y)+g(y)=1$ and then continues the proof. Now I can´t figure out why we can assume this. I was trying to prove that this is indeed the worst case for the inequality to hold, but I am not sure this is true.

Can anybody explain me why we can assume $f(x)+g(x)=f(y)+g(y)=1$ , or maybe tell me what other options are to conclude the proof.

Thank you

  • 3
    $\begingroup$ because the problem is homogeneous, so you can always multiply a positive number to f(x) and f(y) without changing the problem's nature. $\endgroup$
    – Yimin
    Mar 1 '19 at 1:19

I believe the relevant part of the answer in the other question the OP is referring to is

Set $a=f(x),b=f(y),c=g(x),d=g(y)$, then we need to show $$a^\theta b^{1-\theta}+c^\theta d^{1-\theta}\leq (a+c)^\theta(b+d)^{1-\theta}.$$ By dividing $(a+c)^\theta$ and $(b+d)^{1-\theta}$ on both sides, we may assume that $a+c=b+d=1$.

As suggested, dividing both sides by $(a+c)^\theta (b+d)^{1-\theta}$ gives

$$\cfrac{a^\theta b^{1-\theta}+c^\theta d^{1-\theta}}{(a+c)^\theta (b+d)^{1-\theta}} \leq 1 \tag{1}\label{eq1}$$

The question's comment by Yimin refers to "the problem is homogeneous". I believe this means that if, using the question variables, consider $a + c = k$ for some $k$, then let $a = ka_1$ and $c = kc_1$ to get, if $k \neq 0$, that $a_1 + c_1 = 1$. Substituting the redefined $a$ and $c$ values into the LHS of \eqref{eq1} gives

\begin{align} \cfrac{a^\theta b^{1-\theta}+c^\theta d^{1-\theta}}{(a+c)^\theta (b+d)^{1-\theta}} & = \cfrac{\left(ka_1\right)^\theta b^{1-\theta}+\left(kc_1\right)^\theta d^{1-\theta}}{({ka_1}+{kc_1})^\theta (b+d)^{1-\theta}} \\ & = \cfrac{k^\theta a_1^\theta b^{1-\theta}+ k^\theta c_1^\theta d^{1-\theta}}{k^\theta(a_1 + c_1)^\theta (b+d)^{1-\theta}} \\ & = \cfrac{a_1^\theta b^{1-\theta}+ c_1^\theta d^{1-\theta}}{(a_1 + c_1)^\theta (b+d)^{1-\theta}} \tag{2}\label{eq2} \end{align}

As such, $a_1 + c_1 = 1$ also works in the inequality where there is just a relabeling of $a$ to $a_1$ and $c$ to $c_1$. You can do a similar thing for $b + d$.

  • $\begingroup$ Thank you very much, nice explanation!! $\endgroup$
    – JoseSquare
    Mar 1 '19 at 1:43
  • 1
    $\begingroup$ You are welcome. Note this is not my area of expertise, so it wasn't immediately obvious to me either & took me a while to figure it out. $\endgroup$ Mar 1 '19 at 1:44
  • $\begingroup$ I appreciate it. I had no clue that the reason was the homogeneity. Thanks so much for illuminating me :-) $\endgroup$
    – JoseSquare
    Mar 1 '19 at 9:58

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.