# Ergodic Invariant Measures and the Rado graph

Given two or more invariant measures on a structure, there are various ways to combine them to form another invariant measure on the structure. For example, given two invariant measures on a structure, one may construct another invariant measure by taking a mixture, the distribution of the following probabilistic process: first flip a weighted coin to determine which measure to use, and then independently sample from that measure. This provides a general method for finding new invariant measures, but since it is always available to us, we might search for ergodic measures, i.e., ones that are not decomposable as a nontrivial mixture.

Question 1: What Ergodic invariant measures exist on the Rado graph? What about the Henson graph?

Question 2: Can you devise ways of using invariant measures on one structure to produce invariant measures on another? For example, the measures described on the generic bipartite graph can be thought of as elaborations of the measures on the Rado graph. Can such methods for creating invariant measures tell you anything nonobvious about the algebraic or definable closures of the second structure?

-
If by "invariant" you mean invariant under all automorphisms of the graph, then the only invariant measures, up to constant factors, are the counting measures, because the graphs are countable and the automorphism groups are transitive. If you meant some other sort of invariance, please specify it. –  Andreas Blass Aug 23 '11 at 2:50
I am sorry for self-advertising, but may be this paper may clarify something: lanl.arxiv.org/abs/0804.3386 –  Fedor Petrov Aug 23 '11 at 7:29
It didn't occur to me to interpret "measures on the Rado graph" to mean "measures on the collection of all (isomorphic) Rado graphs." That would make the question reasonable, and apparently the cited paper of Fedor Petrov and Anatoly Vershik completely answers it. –  Andreas Blass Aug 24 '11 at 3:21
I should add that the work I mention in the answer below cites the paper of Fedor and Anatoly. –  James Freitag Aug 24 '11 at 14:34

From what I understand they give necessary and sufficient conditions for constructing invariant measures on first order structures. Their condition is roughly that the "definable closure" is trivial, where I put this in quotes because it is not what we normally think of in model theory. In the following, letters represent (possibly) tuples, not just singletons. For a structure $\mathcal M,$ we define $a \in dcl (b)$ if given $\sigma \in Aut (\mathcal M /b)$, $\sigma (a)=a.$ In the case that we demand $\mathcal M$ is saturated, this matches the normal definition. But, in unsaturated structures, this need not be the case. Consider, for instance, a structure with countably many descending nested predicates, $P_n$ with $\mathcal M$ a structure so that there is a single point in the intersection $\cap _n P_n$. Of course, such an element is fixed by any automorphism, but is not definable over any set which does not include it.
Restrict, for the moment to a countable language. Then the dcl given above is equivalent to $a \models \psi (x,b)$ and $x \models \psi (x,b) \rightarrow x=a,$ but where $\psi \in \mathcal L_{\omega _1 ,\omega }.$ Make the appropriate modifications when in a larger language (change the $\omega_1$).
Their conditions says that an invariant measure exists when dcl is trivial (meaning $dcl(a)=a$). So, for instance, these structures must be relational.