7
$\begingroup$

Let $(X,\mathcal A, \mu)$ be a probability space, $\mathbb E$ a Banach space, and $f:X\to\mathbb E$ a Bochner integrable function.

Does there exist a sequence $(x_k)_{k\ge 1} $ in $X$, and a sequence of non-negative numbers $(\lambda_k)_k$ such that $\sum_{k=1}^\infty\lambda_k =1$, $\sum_{k=1}^\infty\lambda_k\|f(x_k)\|<+\infty$ and $$\int_Xfd\mu=\sum_{k=1}^\infty\lambda_k f(x_k)$$ $$?$$

In other words, are all integral means of $f$ already obtained as means over discrete probability measures? We may assume $f$ is bounded, if it helps.

$\endgroup$
1
  • 1
    $\begingroup$ I think a negative answer should follow from this answer and Choquet's theorem. $\endgroup$ Commented Mar 24 at 17:57

2 Answers 2

5
$\begingroup$

No.
Let $(X, \cal A, \mu)$ be $[0,1]$ with Lebesgue measure.
Let $E = L^2[0,1]$ with inner product $\langle \alpha,\beta\rangle := \int \alpha(t)\overline{\beta(t)}\;dt$.
Define $f : [0,1] \to L^2[0,1]$ by $$ f(t) = \mathbf1_{[0,t]} $$ (I wrote $\mathbf1_S$ for the indicator function of a set $S$.)
Now $f$ is continuous and bounded, so it is Bochner integrable.
The Bochner integral $\psi= \int_0^1 f(t)\;dt$ satisfies: for all $\alpha \in L^2$, $$ \langle \psi, \alpha\rangle =\int_0^1 \psi(s) \overline{\alpha(s)}\;ds =\int_0^1 \int_0^1\mathbf1_{[0,t]}(s)\;dt \;\overline{\alpha(s)}\;ds \\ =\int_0^1 \int_s^1 1\;dt \;\overline{\alpha(s)}\;ds =\int_0^1 (1-s)\;\overline{\alpha(s)}\;ds $$ Therefore $\psi(s) = 1-s$ for almost all $s \in [0,1]$. (Altering $\psi$ on a nullset, we may assume $\psi(s) = 1-s$ for all $s \in [0,1]$.)

Now let us consider $$ \varphi =\sum_{k=1}^\infty\lambda_k f(x_k) , $$ where $\lambda_k$ and $x_k$ are as specified. Can we have $\varphi(s) = 1-s$?
Arguing as before, we conclude $$ \varphi(s) = \sum_{k=1}^\infty\lambda_k \mathbf1_{[0,x_k]}(s) = \sum_{x_k \ge s} \lambda_k $$ for almost all $s$. (Again we may assume this holds for all $s$.) Because the $\lambda_k$ are nonnegative, we see that $\varphi$ is a nonincreasing function. We may assume without loss of generality that the points $x_k$ are all distinct. Because $\varphi(s) = 1-s$, for some $k$ we have $0 < x_k < 1$ and $\lambda_k > 0$. By re-labeling, we may assume $0<x_1<1$ and $\lambda_1 > 0$. Now compute $$ \lim_{s \nearrow x_1} \varphi(s) = \sum_{x_k \ge x_1}\lambda_k, \qquad \lim_{s \searrow x_1} \varphi(s) = \sum_{x_k > x_1}\lambda_k . $$ The function $\varphi$ has a jump of size $\lambda_1 > 0$ at $x_1$. Recalling that $\varphi$ is nonincreasing, we conclude that $\varphi(s) = 1-s$ cannot hold almost everywhere.

Thus $\varphi \ne \psi$.

$\endgroup$
5
  • $\begingroup$ Fantastic example! $\endgroup$ Commented Mar 24 at 18:23
  • $\begingroup$ In order to conclude, can’t we just say that $\sum_{k=1}^\infty\lambda_k \mathbf1_{[0,x_k]}$, seen as a true measurable function $[0,1]\to\mathbb R$, has countable image? On the contrary, any representative of $\psi\in L^2([01,])$ must have image of full measure in $[0,1]$ $\endgroup$ Commented Mar 24 at 18:36
  • 1
    $\begingroup$ Not sure about that "countable image" assertion, in general. A nice example is $\sum_{k=1}^\infty (1/2)^k \epsilon_k$ where $\epsilon_k$ are independent random variables with $\epsilon_k = 1$ and $\epsilon_k = 0$ each with probability $1/2$. Then the sum is uniformly distributed on $[0,1]$. $\endgroup$ Commented Mar 24 at 18:43
  • $\begingroup$ Can we say: for every finite Borel measure $\mu$ on $I:=[0,1]$ the integral $\int_Ifd\mu$ is the distribution function $t\mapsto \mu([t,1])$, so $\mu$ can be recovered from $\int_Ifd\mu$, and different measure give different integrals of $f$. $\endgroup$ Commented Mar 25 at 7:12
  • $\begingroup$ In fact if for a measure $\mu$ on [0,1] we denote $F_\mu$ the distribution function $t\mapsto \mu([t,1])$, then the above $f$ is $x\mapsto F_{\delta_x}=\mathbf1_{[0,x]}$, and the integral writes $\int_IF_{\delta_x}\mu(x)=F_\mu$. $\endgroup$ Commented Mar 25 at 16:02
4
$\begingroup$

I don't think so. Consider the functions $f(x,y)=e^{ixy}, -1<x<1, y\in \mathbb{R}$. Then, $$ \int_{-1}^1 f(x,y) \frac{dx}{2} = \frac{\sin(y)}{y}. $$ The question is if this is representable as $$ \sum_{k=0}^\infty \lambda_k e^{ix_ky} = \frac{\sin(y)}{y} $$ for some $\lambda_k \in (0,1)$ such that $\sum_{k}\lambda_k = 1 $ and $x_k\in (-1,1)$. Then by taking the generalized Fourier transform you can see that the function on the left has a Fourier transform $\sum_k \lambda_k \delta_k$, while on the write it is $\frac{1}{2}\chi_{(-1,1)}$. Probably there is a complex analysis way to see this without taking Fourier transform but I do not see it now

$\endgroup$
3
  • $\begingroup$ Nice idea. The series $\sum \lambda_k f(x_k)$ presumably converges in the norm of $E$. Which (in case $E$ is a space of functions) may or may not imply pointwise convergence. I spent a lot of space in my solution justifying the pointwise convergence. $\endgroup$ Commented Mar 24 at 18:24
  • $\begingroup$ I was probably skipping details, but since the series convergences pointwise the limit must coincide with the norm limit (in any reasonable norm like L2) $\endgroup$ Commented Mar 24 at 18:28
  • $\begingroup$ This is also a nice source of examples. Thank you! $\endgroup$ Commented Mar 24 at 18:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .