让下一篇重要论文，不再错过。

原文摘要

A single classical f-divergence has several inequivalent quantum versions, but what distinguishes them operationally is less clear. Matsumoto's reverse tests characterize the maximal quantum f-divergence as the least classical distinguishability needed to prepare two quantum states with one device. Here we introduce retrodictive tests, in which two reverse tests sharing a reference state are coupled via Bayesian retrodiction to define a forward and a reverse stochastic process. In the forward process, the system is prepared by the first device and measured by the Petz transpose of the second; the reverse process interchanges their roles. We seek the least classical distinguishability between these two processes, as measured by the f-divergence between their joint laws, for fixed prescribed states and reference. For full-rank states in finite dimension, we prove that the canonical spectral implementations attain this minimum simultaneously for every operator-convex f. The reference then determines which quantum f-divergence is obtained: a flat reference gives the standard divergences, with the Nussbaum--Szkoła distributions as optimal joint laws; either prescribed state as reference recovers Matsumoto's reverse tests and the maximal divergences; and power references give the geodesic family. In particular, the Umegaki and Belavkin--Staszewski relative entropies arise from the same classical variational principle under different choices of reference. Spectral optimality extends to processes containing a quantum channel; for channels with full-rank Choi operator, the minimum of the forward relative entropy agrees with a known average entropy production.