Bayesian probabilities are usually justified by the Cox theorems, that can be written this way:

*Under some technical assumptions (continuity, etc, etc...), given a set $P$ of objects $A, B, C, \ldots$, with a boolean algebra defined over it with operations $A \wedge B$ (and) and $A | B$ (or) such that*:

1) $A \wedge B = B \wedge A$

2) $A \wedge (B \wedge C) = (A \wedge B) \wedge C$

3) $A | (B \wedge C) = (A|B) \wedge (A|C)$

*and a "valuation":*

$f : P \rightarrow \mathcal{R}$

*there is a strictly monotonic "regraduation function"* $R : \mathcal{R} \rightarrow \mathcal{R}$ *such that, for*:

$R(f(A\wedge B)) = R(f(A)) + R(f(B))$ (sum rule)

and

$R(f(A|B)) = R(f(A) ) R(f(B))$ (product rule)

This theorem allows one to show that any system designed to "evaluate" boolean expressions consistently with a single real number redunds in the laws of classical probability (this can be seen shortly here: arxiv:physics/0403089 and more thoroughly here: arxiv:abs/0808.0012)

Recently this has been extended for valuations of the type $f : P \rightarrow \mathcal{R}^2$ in http://arxiv.org/abs/0907.0909 and they proved that there are just 5 canonical valuations compatible with the underlying Boolean algebra (one of them giving a complex field structure to the "valuation" field).

My question/proposal is: is it possible/interesting/feasible to classify at least a class of valuations of the type:

$f : P \rightarrow W$

where W is a continuous manifold? If we retrict our attention to $\mathcal{R}^n$ for example, is there, for each n, a set of canonical valuations to which all others can be reduced after a regraduation?

If this can be done, are those nice rules for inference in some sense? Are they useful as inference tools in specific situations?