1: \begin{abstract}

2: Let $\sigma$ be a first-order signature and let $\mbW_n$ be the set of all $\sigma$-structures with domain $\{1, \ldots, n\}$.

3: By an inference framework we mean a class $\mbF$ of pairs $(\mbbP, L)$, where 

4: $\mbbP = (\mbbP_n : n = 1, 2, 3, \ldots)$ and $\mbbP_n$ is a probability distribution on $\mbW_n$,

5: and $L$ is a logic with truth values in the unit interval $[0, 1]$.

6: An inference framework $\mbF'$ is asymptotically at least as expressive as another

7: inference framework $\mbF$ if for every $(\mbbP, L) \in \mbF$ there is $(\mbbP', L') \in \mbF'$ such that 

8: $\mbbP$ is asymptotically total-variation-equivalent to $\mbbP'$ and for every $\varphi(\bar{x}) \in L$ there is $\varphi'(\bar{x}) \in L'$ such that 

9: $\varphi'(\bar{x})$ is asymptotically equivalent to $\varphi(\bar{x})$ with respect to $\mbbP$.

10: This relation is a preorder and we describe a partial order 

11: (illustrated by Figure~\ref{map of inference frameworks}) on the equivalence classes of some inference frameworks

12: that seem  natural in the context of machine learning and artificial intelligence.

13: Several previous results about asymptotic (or almost sure) equivalence of formulas or convergence in probability

14: can be formulated in terms of relative asymptotic strength of inference frameworks.

15: We incorporate these results in our classification of inference frameworks and prove two new results.

16: Both concern sequences of probability distributions defined by directed graphical models that use ``continuous''

17: aggregation functions.

18: The first considers queries expressed by a logic with truth values in $[0, 1]$ which employs continuous aggregation functions.

19: The second considers queries expressed by a two-valued conditional logic that can

20: express statements about relative frequencies.

21: 

22: \end{abstract}

23: