Vector-valued integration is something I generally try not to think about very much. I have the impression that it can be a sort of "rabbit hole" of a subtlety if one allows it to be. So, I tend to treat it as a black box. Eventually, though, adopting this attitude is bound to make a person feel guilty -- and then it's time to firm things up a bit. I know there are various ways of defining the integral of a Banach space-valued function. What I'm asking about here is an approach that I've always had in the back of my mind as something that "should work", but, when I tried to write down the details, I became less confident. Can anybody help me put this on rigorous footing? Or gently inform me that this approach is not good?

Here is the sort of setting I have in mind. Let $X$ be a Banach space, say over $\mathbb{C}$, and let $f : \mathbb{R} \to X$ be continuous and such that $M : = \int_\mathbb{R} \|f(t)\| \ dt < \infty$. Now, for any $\varphi \in X^*$, we have $$ | \int \varphi \circ f | \leq \int |\varphi(f(t))| \ dt \leq \int \| \varphi \| \| f(t) \| \ dt = M \|\varphi\| $$ so it is not difficult to see that $\varphi \mapsto \int \varphi \circ f : X^* \to \mathbb{C}$ is a bounded linear functional, that is, an element of $X^{**}$.

What I want to do now show is show that this element of the double dual is actually in $X \subset X^{**}$ so that there is an element $I \in X$ satisfying $\varphi(I) = \int \varphi \circ f$ for all $\varphi \in X^*$. One then defines $\int f := I$. The way I always envisioned doing this was by an application of the following fact:

An element $\psi \in X^{**}$ is actually in $X \subset X^{**}$ if (and only if) it is continuous $X^* \to \mathbb{C}$ when $X^*$ has the weak-star topology.

So, let's try to see if $\varphi \mapsto \int \varphi \circ f$ has this kind of continuity. Suppose that $\varphi_i$ is any net in $X^*$, convergent in the weak-start topology to some $\varphi \in X^*$. That is $\varphi_i(x) \to \varphi(x)$ for all $x \in X$. We want to show that $$ \int_\mathbb{R} \varphi_i(f(t)) \ dt \to \int_\mathbb{R} \varphi(f(t)) \ dt.$$ In our favour, we have that $\varphi_i \circ f \to \varphi \circ f$ pointwise over $\mathbb{R}$ (from the weak-star continuity). So, one thinks that Lebesgue's dominated convergence function may apply. But the problems with this are two-fold.

  1. We need a dominating function for the family of functions $t \mapsto \varphi_i(f(t))$. I think maybe the uniform boundedness principle can help here?
  2. The Lebesgue dominated convergence theorem applies to sequences not nets. Unless there is some way to finesse around this issue, I think this may be a more critical error.

To reiterate, the slightly flabby question I'm asking is:

Can the approach to defining $\int f$ I just outlined be "rigourized"?

Thanks for reading.

  • $\begingroup$ I'm not sure how one would best prove that $\Lambda\colon\varphi \mapsto \int \varphi\circ f$ is in (comes from an element of) $X$, but that's basically the Pettis integral. $\endgroup$ – Daniel Fischer Aug 4 '13 at 21:54

Consider first the integral over a compact interval $[a,b]$. Then $\frac{1}{b-a} \int_a^b f(t) dt$ is in the weak$^*$ closure in $X^{**}$ of the convex hull $\mathrm{Conv}(f([a,b]))$ of $f([a,b])$ (morally, this is clear since the integral with respect to a probability measure is a generalized convex combination, the proof uses this property for real-valued functions). By the continuity of $f$ the range $f([a,b])$ is compact in $X$ and hence its convex hull is precompact and thus relatively compact in $X$. Since compact subsets of $X$ are weak$^*$ closed in $X^{**}$ you get that the integral $\int_a^b f(t) dt$ belongs indeed to $X$ (and not only to the bidual).

For the general case you can approximate $f$ by $f \chi_n$ where $\chi_n$ is $1$ on $[-n,n]$, $0$ in $\mathbb R \setminus [-(n+1),n+1]$ and afin-linear on the remaining two intervals. Then you show $\int_{\mathbb R} f_n(t)dt = \int _{-(n+1)}^{n+1} f_n(t) dt \to \int_{\mathbb R} f(t)dt$ in the norm of $X^{**}$ so that the integral is in $X$ (because $X$ is closed in its bidual).

  • $\begingroup$ Do the terms "precompact" and "relatively compact" mean different things to you? $\endgroup$ – Mike F Aug 6 '13 at 6:25
  • $\begingroup$ In Banach (or complete metric) spaces they are equivalent. By definition, a set is precompact, if for every $\varepsilon >0$ the set is covered by finitely many balls of that radius. Relatively compact means that the closure is compact. $\endgroup$ – Jochen Aug 6 '13 at 7:56
  • $\begingroup$ Hmm OK so your "precompact" is my "totally bounded". $\endgroup$ – Mike F Aug 6 '13 at 16:44

From this answer we know that $\int_\mathbb{R}f(t)dt$ exists as some element $x_f\in X$ if

1) $f(\mathbb{R}\setminus A)$ is separable for some null set $A$

2) $\varphi \circ f$ is continuous for every $\varphi\in X^*$

We are going to prove that 1) and 2) holds in our case. For every $n\in\mathbb{N}$ the set $f([-n,n])$ is compact as continuous image of compact set, and hence separable. Thus $f(\mathbb{R})=\bigcup_{n\in\mathbb{N}} f([-n,n])$ is separable as union of separable sets. The second condition holds trivially because $f$ is continuous.

Hence for any $\varphi\in X^*$ we get $$ \int_\mathbb{R}(\varphi\circ f)(t)dt=\varphi\left(\int_\mathbb{R} f(t)dt\right)=\varphi(x_f) $$ So the map $\varphi\mapsto \int_\mathbb{R}(\varphi\circ f)(t)dt$ is just the evaluation of $\varphi$ at $x_f$.

  • $\begingroup$ It occurs to me, since $f$ vanishes at $\infty$ (being continuous and integrable), it extends ctsly to the 1-pt compactification $\mathbb{R} \cup \{\infty\}$ by $\infty \mapsto 0$. So, $f(\mathbb{R}) \cup \{0\}$ is already compact. Just another argument for separability of $f(\mathbb{R})$. $\endgroup$ – Mike F Dec 16 '13 at 4:48

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.