{"id":"arxiv/2602.09061","url":"https://arxiv.org/abs/2602.09061","tags":["stat.ME","cs.IT","math.ST"],"score":0,"title":"Coherent Information Deletion: Bayesian, Generalized Bayesian and Possibilistic Unlearning","domain":"stat.ME","source":"arxiv","authors":["Hans Montcho","Håvard Rue","Maurizio Filippone"],"reviews":0,"subject":"stat.ME","summary":"Bayes' theorem admits an information processing interpretation due to Zellner (1988): under the Shannon information criterion, the posterior is the unique rule that processes prior and data information without information loss. We revisit these ideas from the perspective of information deletion. Given a posterior based on a complete dataset, what distribution should replace it when a subset of the data is removed? For ordinary Bayesian inference, we show that the corresponding information deletion objective is uniquely minimized by the leave-data-out posterior. We then extend the framework…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-02-08","reactions":{}}