While trying to answer this question, I've come across the notion of the weak*-topology on a set of probability measures. I'd like some clarification about what this means.
More specifically, let $(\Omega, \mathcal{F})$ be a measurable space. We don't assume that $\Omega$ has any metric or topological structure. What does it mean to equip the set $\mathcal{M}$ of probability measures on this space with the weak*-star topology?
I understand that the weak*-topology is the weakest topology on the dual space $V'$ of a normed vector space $V$ that makes the evaluation functionals defined by $\lambda_f(\phi) = \phi(f)$, $\phi \in V'$ and $f \in V$, continuous. What I don't understand is how $\mathcal{M}$ can be equipped with this topology as it's not a vector space.
From what I've read, I think that measures in $\mathcal{M}$ are being identified with linear functionals on a space of measurable functions. For instance, $P \in \mathcal{M}$ gives rise to a linear functional $\phi$ on the normed linear space of bounded $\mathcal{F}$-measurable functions, equipped with the $\sup$-norm, by $\phi(f) := \int f dP$. Is something like this correct? Which underlying vector space of measurable functions should be used?
I would appreciate if someone could please sketch the relevant theory for me and/or refer me to a comprehensive textbook treatment of this topic.
Addendum. My current understanding of this topic is summarized as part of my attempt to answer my own question in the link above.
Best Answer
You can consider the set of probabilities over $\mathcal{F}$ as a subset of the linear space $V(\mathcal{F})$ of all finitely additive (bounded) scalar measures over $\mathcal{F}$ endowed with the variation norm (see Theory of charges (K. Bhaskara Rao, M. Bhaskara Rao), chapter 7). We need to show that this space is the topological dual of another (locally convex) topological vector space.
When $\mathcal{F}$ is a Boolean algebra of subsets of $\Omega$ (which in this case does not have to be a topological space) we define $S(\mathcal{F})$ as the linear space generated by $\{\chi_A:\ A\in\mathcal{F}\}$ (the characteristic functions of the sets in $\mathcal{F}$). $S(\mathcal{F})$ is called the space of simple functions.
Now, for every $f\in S(\mathcal{F})$, $\Vert f\Vert_s:=\sup\vert f\vert<\infty$ so $(S(\mathcal{F}),\Vert\cdot\Vert_s)$ is a normed space. It is not hard to prove that the dual of $(S(\mathcal{F}),\Vert\cdot\Vert_s)$ is the space $V(\mathcal{F})$. In fact, on one hand every $\lambda\in V(\mathcal{F})$ defines the bounded linear functional $f\mapsto \int f\ d\lambda$ (for $f\in S(\mathcal{F})$), on the other, every $x^*\in S(\mathcal{F})^*$ defines on $\mathcal{F}$ the measure $\lambda_{x^*}(A):=x^*(\chi_A)$ (for $A\in\mathcal{F}$). See Topological Riesz spaces and measure theory (Fremlin) for a more complete reference.
Having showed that the set of probabilities over $\mathcal{F}$ is contained in the topological dual of the normed space $S(\mathcal{F})$, it is clear why we can talk about weak$^*$-convergence.
PS: It is worth observing that by the Stone representation Theorem for Boolean rings any Boolean ring $\mathcal{R}$ is isomorphic to the ring of clopen sets in a locally compact Hausdorff space (see Measure theory Vol III (Fremlin), or this survey by Tao). Following this line, the approach showed by Tomasz can be proved to be much closer to the one exposed so far than someone would think. It would be interesting to go through this idea.