$f(x)$ is quasi-convex,

$$x^*\in\arg\min_{x\in C}f(x).$$

How to prove that, for any $a\in C$, $f(x) $ is weakly monotonic in the direction of $(x^*-a)$?

Is this simple result a part of an ancient theorem?

  • 3
    $\begingroup$ You can easily prove it, using the definition of function convexity and the fact that $x^*$ is a minimizer. I'm hesitant to say more, lest I ruin a perfectly good homework problem. $\endgroup$ – prubin Aug 6 '20 at 20:32

By the definition of quasiconvex: $f(x)$ with compact support $C$ is quasiconvex if for two points in the domain $x_1,x_2$ and $w\in[0,1]$ $f(wx_1+(1-w)x_2)\geq \max\{f(x_1),f(x_2)\}$.

Let $x^* = \arg\min_{x\in C}f(x)$ where $C$ is the compact support of $f$. Then consider $x_1,x_2\in [x^*,\infty)$.

Choose $x_2>x_1$. By the definition of quasiconvexity, the secant segment from $(x_1,f(x_1))$ to $(x_2,f(x_2))$ lies below or at the maximum of the segment endpoints $\{f(x_1),f(x_2)\}$. Since $x^*$ is a global minimizer, we can choose $x_1=x^*$ which implies the right limit inequality:

$$\lim_{x_2\downarrow x_1} f(wx_1+(1-w)x_2)-f(x_1)\geq \max\{0,f(x_2)-f(x_1)\}~\forall w\in[0,1].$$ Thus the right derivative is non-negative. This then holds for all $x_1\geq x^*$. Thus $f$ is weakly monotone increasing on $[x^*,\infty)$.

We can do likewise for $x_1,x_2\in(-\infty,x^*]$ using left limits and show that $f$ is weakly monotone decreasing on $(-\infty,x^*]$.

  • $\begingroup$ Thank you kurtosis! I like your method but I am not sure if we are implicitly assuming that $f$ is differentiable (which is not necessarily true). $\endgroup$ – High GPA Aug 8 '20 at 13:59
  • $\begingroup$ I suppose we could forego the limit existing (and hence the derivative) and just have the inequality be true ($\forall\delta>0:0<x_2-x_1\leq\delta\ldots$) and still get weak monotonicity. $\endgroup$ – kurtosis Aug 8 '20 at 19:46

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.