6

If $h(x,t)$ is measurable and the measures involved are $\sigma$-finite, does there exist a sequences of functions

$$h_n(x,t) = \sum_{j=1}^{N_n} f_{j,n}(x) 1_{F_{j,n}}(t)$$ where the sets are pairwise disjoint (in $j$) such that $h_n \to h$ pointwise almost everywhere?

I know that if we fix $x$ we can get a sequence like that but we cannot just make the coefficients depending on $x$ since the sets $F_{j,n}$ will be different. Can I somehow combine them? Do I need the $\sigma$-finiteness?

If the answer is positive, we can use this to prove Minkowski's inequality for integrals quite easily.

JT_NL
  • 14,514
  • Sorry if it is obvious to those who know measure theory well (not me, unfortunately!), but in what sense would you like the $h_n$ to converge to $h$? – Matt E Sep 28 '10 at 20:32
  • Pointwise. I'll add it. – JT_NL Sep 28 '10 at 20:46
  • I've heard of a $\sigma$-finite measure, but what is a $\sigma$-finite sigma algebra? –  Sep 29 '10 at 00:42
  • I suspect that the answer is no, but that the required counterexample is pretty complicated. For what it's worth... –  Sep 29 '10 at 00:44
  • The measure is $\sigma$-finite, I'll correct it. – JT_NL Sep 29 '10 at 10:09
  • What measure? The question is entirely about measurability without reference to any measure. Unless by "pointwise" you mean "pointwise almost everywhere". Maybe you could show us your easy proof of Minkowski's inequality, and we could help fill in the gaps. –  Sep 29 '10 at 12:35
  • The measure is just a $\sigma$-finite measure, that are the hypothesis on Minkowski's integral inequality. I suspect it has to do with using Fubini but my "proof" doesn't use that, so I guess that if it is true then you need the $\sigma$-finiteness for this. I'll change it later this day to add my proof. – JT_NL Sep 29 '10 at 15:34

2 Answers2

2

In contrast to William, I think the answer is yes. I had to prove something similar once; here is a suitably adapted argument. It might be more complicated than necessary.

First, reduce to the case that $(\Omega_1, \mu_1), (\Omega_2, \mu_2)$ are finite measure spaces and $h$ is bounded. Let $\mathcal{P}$ be the set of all functions on $\Omega_1 \times \Omega_2$ of the form $F(x,y) = f(x) g(y)$ where $f$ is bounded and measurable and $g$ is simple; such functions can be written in the form you seek. Let $\mathcal{Q}$ be the linear span of $\mathcal{P}$; these functions can also be written in the desired form. Let $\mathcal{L}$ be the closure of $\mathcal{Q}$ in $L^1(\mu_1 \times \mu_2)$ and let $\mathcal{L}_b$ be the bounded functions from $\mathcal{L}$.

Now $\mathcal{L}_b$ is a vector space which is closed under bounded convergence (by the dominated convergence theorem), contains the constants and contains $\mathcal{P}$. $\mathcal{P}$ is closed under multiplication and contains all functions of the form $1_{A \times B}$ with $A \subset \Omega_1$, $B \subset \Omega_2$ measurable; the collection of such sets $A \times B$ generates the product $\sigma$-algebra. By the functional version of the Dynkin $\pi$-$\lambda$ theorem (references below), $\mathcal{L}_b$ contains all bounded measurable functions; in particular it contains $h$.

Since $h$ is in $\mathcal{L}$ which is the $L^1$ closure of $\mathcal{Q}$, there is a sequence $\mathcal{Q} \ni h_n \to h$ in $L^1$. Then some subsequence converges almost everywhere.

For the functional $\pi$-$\lambda$ theorem, see Theorem 8.2 of Bruce Driver's probability notes. A reference is also given to C. Dellacherie, Capacités et processus stochastiques, page 14.

Nate Eldredge
  • 97,710
  • This looks good. I will wait if someone can come up with an easier argument to prove this before I accept it as answer. What was it that you had to prove? – JT_NL Sep 29 '10 at 22:17
  • It was related to http://math.stackexchange.com/questions/1618/is-the-natural-map-lpx-otimes-lpy-to-lpx-times-y-injective. I wanted to prove that $L^p(X) \otimes L^p(Y) \subset L^p(X \times Y)$ was dense. – Nate Eldredge Sep 29 '10 at 22:40
1

The answer here is no. I am assuming you're trying to prove the following:

Given $\sigma$-finite measure spaces $\Omega_1, \Omega_2$, and $\Omega = \Omega_1\times\Omega_2$ carries the product measure, then if $h: \Omega\rightarrow \mathbb{R}$ is a measurable function, then $h$ is point-wise approximated almost everywhere by the sequence $h_n$ given by:

$$h_n(x,t) = \sum_{j=1}^{N_n}f_{j,n}(x)1_{F_{j,n}}(t)$$

Well, in general this is not the case since the sets $F_{j,n}$ will generally depend on $x$. If $\Omega_1$ is countable, the situation is not too bad: the family of sets $\{F_{j,n}(x): x\in \Omega_1\}$ is countable. In general, however, the situation is hopeless. I think using $\Omega_1 = \Omega_2 = \mathbb{R}_{\geq 0}$ and $h(x,t) = x^2 + e^t$, or something to that effect, should do as a counter example. Play with it.

  • I'll try it if that works before I accept your answer. The $\sigma$-algebra is not necessarily countable. – JT_NL Sep 29 '10 at 15:26
  • 2
    $h(x,t) = x^2 + e^t$ is not a counterexample. We can easily approximate $e^t$ by functions of the desired form (take $f_{j,n}$ constant) and $x^2$ even more easily (take $f_{1,1}(x) = x^2$ and $F_{1,1} = \Omega_2$), and add them together. – Nate Eldredge Sep 29 '10 at 18:42
  • @Nate: Indeed! I don't know what I was thinking. –  Sep 30 '10 at 01:41