Title:Generalized information structures and their cohomology

Abstract:Baudot and Bennequin introduced a cohomological construction adapted to information theory, called \emph{information cohomology}; it characterizes Shannon entropy as the only nontrivial cohomology class in degree one, for suitable coefficients. This text develops the relation between information cohomology and topos theory. To this end, we redefine information structures, as categories of random variables related by a notion of extension or refinement; classical and quantum probability spaces appear as models (or representations) of them. These generalized information structures form a category with finite products and coproducts. Information cohomology is a derived functor in the abelian category of modules over an algebra of random variables, a subcategory of the presheaf topos associated with the information structure. The cohomology is invariant under isomorphisms of generalized structures. We prove that the relatively free bar construction gives a projective object for the computation of this invariant. We provide detailed computations of $H^0$ and $H^1$ when the coefficients are functionals of classical probabilities, including the degenerate cases. In the process, we also establish the homological nature of Tsallis entropies, as the only cocycles in degree one for certain one-parameter family of modules.