0
$\begingroup$

Prove or disprove that the set $$S:= \left\{p \in \Delta^n \mid \mbox{Var}(X) \leq \alpha \right\}$$ is convex. $p \in \Delta^n$ denotes that $p$ is a probability distribution (i.e., belongs to the probability simplex). $X$ is a random variable such that $p(X = a_i) = p_i$, where $a_1 \leq a_2 \leq \cdots \leq a_n\in \mathbb{R}$.

The question was taken from the exercise 2.15 of Boyd & Vandenberghe's Convex Optimization


My current attempts are to pick two arbitrary points from $S$ and show that the convex combination stays in $S$. I realized that I could use the formula $\mbox{Var}(X) = E[(X - E(X))^2]$, but I don't know how to proceed because the $(\cdot)^2$ in the variance prevents me from distributing the convex combination inside $(\cdot)^2$. Also, $E(X)$ in $X - E(X)$ depends on the probability distribution to provide.

Could anyone provide some pointers? I have been stuck for two days.

$\endgroup$
1
$\begingroup$

Try to check whether the variance is a convex function of $X$.

$\endgroup$
3
  • $\begingroup$ Hi, I am a bit confused about why "variance" could/couldn't be a convex function of $X$. In particular, proving a function is convex often requires us to take two points on the convex function and show that the convex combination of the two points satisfies the Jenson's Inequality, but $X$ is a random variable, how could we take two points? $\endgroup$
    – Fan
    Sep 23 '18 at 21:07
  • $\begingroup$ You should check whether $\mathbf{var}(X)$ is convex in $p$. A convex function has sublevel sets which are convex sets. In this case your set is in terms of $p$ so you need to check whether variance is a convex function of $p$ $\endgroup$
    – Casey
    Sep 23 '18 at 22:18
  • $\begingroup$ @Casey Convexity of sublevel sets is a necessery condition of convexity. The functions for which such sets are convex are called quasi-convex. $\endgroup$
    – szw1710
    Sep 23 '18 at 22:25
0
$\begingroup$

Can you check, for some nontrivial set of $a_i$, and $\alpha=0$, whether the vertices of $\Delta^n$ lie in $S$?

$\endgroup$
0
$\begingroup$

To get an intuitive feel, maybe think about some distributions that you know the variance of (i.e. if $X$ is Bernoulli then $\mathbf{var}(X)= p(1-p)$) then check if it's convex. If you find one counter example then you're done. If the statement is true this won't prove it but the intuition might help

$\endgroup$

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.