3
$\begingroup$

This may be really obvious but I don't see it. Let $f:\Omega \to \mathbb R^n$ be integrable with respect to a probability measure $\mu$. Does it follow that $\int_\Omega f \, d\mu$ is in the convex hull (not just the closed convex hull) of the range of $f$?

If the answer is yes, does this remain true if $\mu$ is merely finitely additive?

If $f$ is continuous and $\Omega$ is an interval, the answer is affirmative.

$\endgroup$

1 Answer 1

2
$\begingroup$

Sorry! It was rather easy, so perhaps it should be closed.

The second question clearly gets a negative. Let $\mu$ be a finitely additive measure on $\mathbb Z^+$ that assigns zero to all finite subsets. Let $n=1$. Let $f(k)=1/k$. Then $\int_{\mathbb Z^+} fd\mu =0$, but $0$ is not in the convex hull of the range.

The first question gets a positive. It's clear for $n=1$. Suppose $n>1$ and it's true for smaller dimensions. Let $e=\int_\Omega fd\mu$. Let $C$ be the convex hull of the range of $f$. If $e\notin C$, there is a hyperplane $H$ through $e$ such that $C$ lies to one side of it. Let $p$ be a normal of $H$ such that $p\cdot f(x) \le p\cdot e$ for all $x\in\Omega$. Then $\int_\Omega (p\cdot f) d\mu = p\cdot e$ but $p\cdot f(x) \le p\cdot e$ for all $x$, so $p\cdot f = p\cdot e$ almost everywhere. We can modify $f$ without changing its range so as to ensure $p\cdot f = p\cdot e$ everywhere. But then the range of $f$ lies in the $(n-1)$-dimensional hyperplane $H$, and so by the $(n-1)$-dimensional case we have $e$ in the convex hull of the range of $f$.

$\endgroup$
1
  • $\begingroup$ Note: The link in my questions to the affirmative answer in the continuous case will go to a theorem by Jankovic and Merkle that the integral is a convex combination of $n$ points in the range when $f$ is continuous. By Caratheodory's theorem, without assuming continuity we will have a combination of $n+1$ points in the range. So continuity lets one simplify the convex combination from $n+1$ to $n$ points. That helps me to understand the significance of the Jankovic-Merkle theorem. $\endgroup$ May 1, 2014 at 15:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.