{"id":"arxiv/2608.30246","url":"https://arxiv.org/abs/2608.30246","tags":["math.LO","cs.LG","math.PR"],"score":0,"title":"A Borel Concept Class of VC Dimension One with a Non-PAC Consistent Learner in ZFC","domain":"math.LO","source":"arxiv","authors":["Mateus Jesus de Arruda Campos","Gabriel Fernandes","Vinicius de Oliveira Rodrigues"],"reviews":0,"subject":"math.LO","summary":"The fundamental theorem of statistical learning states that, under suitable measurability assumptions, finite Vapnik--Chervonenkis (VC) dimension guarantees that every proper consistent learning rule is probably approximately correct (PAC). Blumer, Ehrenfeucht, Haussler, and Warmuth showed, assuming the Continuum Hypothesis, that the \"well-behavedness\" condition of the concept class cannot be omitted: they constructed a concept class of Borel sets of VC dimension one admitting a consistent learning rule that is not PAC. We show that the Continuum Hypothesis is unnecessary. Working in…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-08-31","reactions":{}}