I'd like to show that $$ \frac{\log\frac{y_1+y_2}{x_1+x_2}}{\log\frac{z_1+z_2}{x_1+x_2}} \ge \min\left\{ \frac{\log(y_1/x_1)}{\log(z_1/x_1)}, \frac{\log(y_2/x_2)}{\log(z_2/x_2)} \right\}. $$ This would follow if $f(x,y,z) = \frac{\log(y/x)}{\log(z/x)}$ was quasi concave.

One way to show this is if $f(x,y,z)\ge\alpha$ defines a convex set for all $\alpha\ge 0$. Plotting for various values of $\alpha$, this is clearly the case.

From the first order condition, it suffices if $f(\vec y) \le f(\vec x) \implies \nabla f(\vec x)^T(\vec y-\vec x) \ge 0.$ Let $x'\ge y',z'$ be such that $f(x',y',z')\le f(x,y,z)$ then we must show

$$z \log \left(\frac{x}{z}\right) (y x'-x y') \ge y \log \left(\frac{x}{y}\right) (z x'-x z'),$$ which unfortunately doesn't seem any easier.

Alternatively, since $\log x$ is quasi-linear, perhaps it suffices to show that $(x-y)/(x-z)$ is quasi-concave? I don't know a theorem to this effect, however.

Do you see a nice argument I might use to simplify this problem? Perhaps just tackling the original inequality directly?

  • $\begingroup$ Could you clearly define the domain? Can $x/y/z$ all be negative? $\endgroup$ – LinAlg Mar 11 at 19:05
  • $\begingroup$ @LinAlg Right! They are all positive. They are also at most 1, though that doesn't seem to make a difference. $\endgroup$ – Thomas Ahle Mar 11 at 19:50
  • 1
    $\begingroup$ Ok, just to make sure, $\log(z/x)$ can be both negative and positive, right? Or did you mean $x \geq y > 0$ and $x > z > 0$? $\endgroup$ – LinAlg Mar 11 at 19:53
  • $\begingroup$ @LinAlg I didn't even think of that obvious ambiguity! It is always negative, or if we write it as $\log(x/y)/\log(x/z)$. $\endgroup$ – Thomas Ahle Mar 11 at 19:56

Let me give you a possible method. Consider the superlevel set $\{(x,y,z) : \log(y/x) / \log(z/x) \geq \alpha, x \geq y > 0, x > z > 0 \}$. Since $\log(z/x)$ is negative, the sublevel set is equivalent to the set: $$\{(x,y,z) : \log(y/x) \leq \alpha \log(z/x), x \geq y > 0, x > z > 0 \}$$ $$=\{(x,y,z) : y \leq x^{1-\alpha}z^{\alpha}, x \geq y > 0, x > z > 0 \}.$$ Consider the Hessian of $f(x,z) = x^{1-\alpha}z^{\alpha}$. If it is negative semidefinite, the set is convex and you are done. Otherwise, try the same but with $x$ or $z$ on one side.

  • $\begingroup$ The eigenvalues of the Hessian are $0$ and $-(1-\alpha) a x^{-\alpha-1} z^{\alpha-2} \left(x^2+z^2\right)$, so I guess it is negative semidefinite for all $\alpha\le1$. That suffices for my purposes! Is $\{(x,y,z) : y \le f(x,y)\}$ always convex when $f$ is concave? $\endgroup$ – Thomas Ahle Mar 11 at 21:16
  • 1
    $\begingroup$ @ThomasAhle the answer to your last question is yes; the equivalent set in convex optimization is $\{x : f(x) \leq 0\}$ where $f$ is convex. $\endgroup$ – LinAlg Mar 12 at 0:44

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.