no. 16 · Tuesday 17 November 2026
Operator-valued maximal $f$-divergences for completely positive maps
Abstract
We introduce an operator-valued maximal $f$-divergence for completely positive (CP) maps between von Neumann algebras, associated with an operator convex function $f$ on $(0,+\infty)$. The construction takes values in the extended lower-semibounded self-adjoint part of the codomain von Neumann algebra. We prove independence of the common CP upper bound used in its computation, joint subadditivity, monotonicity under unital precomposition and normal postcomposition, a martingale convergence theorem for normal CP maps, and joint lower semicontinuity in the point-$σ$-weak topology. For normal positive functionals, our construction recovers Hiai's maximal $f$-divergence. As a consequence, we establish its joint weak lower semicontinuity for arbitrary von Neumann algebras, answering a question left open by Hiai. For $η(t)=t log t$, we obtain an operator-valued Belavkin–Staszewski (BS) relative entropy. Moreover, for normal channels with $σ$-finite codomain, we prove that the BS channel divergence of Hollands and Ranallo coincides with the extended norm of our operator-valued divergence. Finally, we give finite-dimensional examples and obtain an explicit formula for a finite-index conditional expectation. In the BS case, this formula reduces to the logarithm of the Jones–Kosaki index.
- fields
- Operator Algebras, Mathematical Physics, Functional Analysis