What is the largest possible variance of a random variable on $[0; 1]$?

It is evident that it does not exceed $1$, but I doubt, that $1$ is actually possible. The largest variance, for which I found the example is $\frac{1}{4}$. That is the variance of a random variable $X$, such that $P(X = 1) = P(X = 0) = \frac{1}{2}$, but I doubt that it is the largest possible one.

Why is it interesting:

Initially I wanted to find the largest possible second moment of $X - Y$, where $X$ and $Y$ are i.i.d. random variables on $[0; 1]$. Then I found: $$E(X - Y)^2 = E(X^2 - 2XY + Y^2) = EX^2 - 2EXY + EY^2 = 2(EX^2 - EXY + EXY - EXEY + Cov(X, Y)) = 2(EX^2 - {(EX)}^2) = 2VarX$$ And thats where I am now.

This question is partially inspired by: Probability distribution to maximize the expected distance between two points

  • $\begingroup$ BTW, the "motivation" part can be derived more easily: $E[X-Y]^2 = Var(X-Y) + E[X-Y]^2$, but since $X, Y$ are i.i.d., $E[X-Y]=0$ and $Var(X-Y) = 2Var(X)$. $\endgroup$ – antkam Feb 9 at 21:11

The answer is $1/4$. For any probability distribution on $[0,1]$ the point $p=(EX,EX^2)$ will be a point in the convex hull of the set $S = \{(x,x^2):x\in[0,1]\}$. (This is a segment of a parabola.) The variance is the height of $p$ above the set $S$. This is clearly maximized when $p$ lies on the straight line connecting $(0,0)$ to $(1,1)$. By calculus, this is attained at $p=(1/2,1/2)$, which is $1/4$ above the point $(1/2,1/4)\in S$.

That $p$ is in the convex hull of $S$ is a consequence of Caratheodory's theorem: each element $p$ of the convex hull of $S$ is a weighted combination of at most 3 elements of $S$: Let that combination be $p=\sum_{i=1}^3 w_i\cdot(x_i,x_i^2)$, where the $w_i\ge0$ add up to $1$. Now look at the prob distribution for which $PX=x_i)=w_i$. Its first 2 moments are the components of $p$.

In the special case at hand Caratheodory's theorem is trivial. Every point in the convex hull of $S$ is on a chord of $S$. If $p$ is already in $S$, it is of the form $p=(x,x^2)$, and the prob law $P(X=x)=1$ does the trick. Otherwise, the line passing through $(0,0)$ and $p$ cuts $S$ at $q$; the chord in question can be between $(0,0)$ and $q$, and $a$ can be chosen so $P(X=(0,0))=a, P(X=q)=1-a$ does the trick.

More generally, as a comment suggests, this is a fundamental property of a expectation operator. An expectation $ET(X)$ of a vector valued function is a particular weighted average of the possible values of the function $T(x)$. A probability law, one can think, amounts to a choice of weights. In problems like this one, the set of possible values of $ET(X)$ you get as you vary the probability law of $X$ is the convex hull of the set of values of the function $T(x)$.

  • $\begingroup$ perhaps i'm missing something obvious, but why is the point $p$ in the convex hull of $S$? presumably this is a consequence of the Expectation operator, but i cant see an obvious proof. can you help? $\endgroup$ – antkam Feb 9 at 19:58
  • $\begingroup$ @antkam I have edited my answer. $\endgroup$ – kimchi lover Feb 9 at 20:28
  • $\begingroup$ thanks! that theorem is exactly what i was looking for. very neat, especially the geometric interpretation. however, i must say, the theorem is much stronger than the requested result. i have written up a more elementary (but less insightful) proof as alternative answer. $\endgroup$ – antkam Feb 9 at 20:39

Another elementary proof: $X^2\leq X$ since $X\in[0,1]$. Therefore $$ Var(X)=\mathbb E[X^2] - \mathbb E[X]^2\leq \mathbb E[X]-\mathbb E[X]^2 = z-z^2 \leq \frac14. $$ The last indequality valid since function $f(z)=z-z^2$ reaches its maximal value $\frac14$ at point $z=\frac12$.

To reach equality in the first inequality $\mathbb E[X^2]\leq \mathbb E[X]$ we need $X^2=X$ a.s. This is possible only when $X$ takes values $0$ and $1$. The second indequality $z-z^2\leq \frac14$ is equality if $z=\mathbb E[X]=\frac12$.

So the upperbound is $\frac14$ and it is achieved by $\mathbb P(X=0)=\mathbb P(X=1)=\frac12$ only.


A more elementary proof:

Variance is shift-invariant, i.e. let $Y = X - 1/2$ and $Var(Y) = Var(X)$. So we just need to consider $Y$ on $[-1/2, +1/2]$. The requested result follows from these observations:

  • $Y^2 \in [0, 1/4] \implies E[Y^2] \in [0, 1/4] \implies E[Y^2] \le 1/4$

  • $E[Y]^2 \ge 0$ since it is a square.

    • (In fact, $Y \in [-1/2, 1/2] \implies E[Y] \in [-1/2, 1/2] \implies E[Y]^2 \in [0, 1/4]$.)
  • $Var(Y) = E[Y^2] - E[Y]^2 \le 1/4 - 0 = 1/4$

This proves the requested upperbound of $1/4$; which the OP has already shown is achievable. This also shows that achieving this bound requires $E[Y] = 0$ and $E[Y^2]=1/4$ which in turn requires $Y^2 = 1/4$ which in turn requires $Y = \pm 1/2$. The OP's example is obviously the only such distribution.


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.