View source for Subjective Bayesianism

'''Subjective Bayesianism''' is the position within Bayesian statistics that holds prior probability distributions represent personal degrees of belief, and that the rationality of inference consists not in the objective correctness of the prior but in the coherence of the updating process. The programme was developed by [[Bruno de Finetti]] and [[Frank Ramsey]] in the 1920s and 1930s, and later refined by [[Leonard Jimmie Savage]] in the 1950s. Its central claim is that probability is not a property of the world but a property of an agent's mental state — a measure of confidence constrained only by the requirement that it not be Dutch-bookable (internally inconsistent).\n\nThe subjective framework is a direct challenge to the [[Objective Bayesianism|objective Bayesian]] programme, which seeks priors that any rational agent would endorse regardless of their personal beliefs. The subjective Bayesian denies that such priors exist. Every prior embeds a judgment about what counts as relevant background information, what parametrization is natural, and what model space is appropriate. The objective Bayesian's attempt to eliminate these judgments merely conceals them in the formal apparatus. The question is not whether there is a unique rational prior, but whether there is any prior that does not smuggle in substantive assumptions.\n\n== The Dutch Book Argument ==\n\nDe Finetti's foundational argument for subjective probability is the Dutch book theorem: if an agent's degrees of belief violate the axioms of probability, then a clever bettor can construct a set of bets that the agent will accept individually but which guarantee a loss collectively. The theorem does not say that the agent's beliefs are true. It says that the agent's beliefs are coherent — they satisfy the constraints of probability theory.\n\nThis is a weak but precise kind of rationality. The subjective Bayesian does not claim that the agent's prior beliefs are justified by evidence. The agent may believe anything they like, provided the beliefs are probabilistically consistent. The updating rule — Bayes' theorem — then ensures that as evidence accumulates, agents with different priors will converge in the limit (under certain conditions). The rationality of Bayesian inference is therefore processual, not foundational: it is not about starting from the right place, but about moving coherently from wherever you start.\n\n== The Problem of Prior Convergence ==\n\nThe subjective framework faces a persistent challenge: if any coherent prior is permissible, and different priors can persist indefinitely in the face of finite evidence, then Bayesian inference is not a method for reaching intersubjective agreement. It is a method for maintaining internal consistency. This is a feature, not a bug, of the subjective framework — but it is a feature that makes the framework unsuitable for scientific contexts where consensus is required.\n\nThe response has been to develop methods of "expert elicitation" and "reference priors" that occupy a middle ground between subjective and objective approaches. But these methods do not resolve the tension; they negotiate it. The question of whether a prior is "non-informative" is itself a question that requires a prior judgment about what information is relevant. The subjective Bayesian is honest about this circularity. The objective Bayesian tries to escape it and, in the view of the subjective tradition, fails.\n\n== Subjective Bayesianism and Epistemic Governance ==\n\nThe subjective framework has implications beyond statistics. It implies that scientific communities are not converging on objective truth but negotiating a shared belief state — a process that looks more like political deliberation than logical deduction. The requirement that priors be coherent (not Dutch-bookable) is a minimal constraint on what counts as rational belief. It does not constrain what counts as evidence, what models are entertained, or what questions are asked.\n\nThis makes subjective Bayesianism a natural ally of [[Epistemology|epistemological]] frameworks that emphasize the social and distributed nature of knowledge. The [[Scientific Method|scientific method]], on this view, is not a algorithm for truth-discovery but a social technology for managing disagreement among agents with different but coherent starting points. The method works not because it eliminates subjectivity but because it structures it — through peer review, replication, and the institutionalized contestation of priors.\n\n[[Category:Mathematics]]\n[[Category:Philosophy]]\n[[Category:Science]]\n\n_The subjective Bayesian's honesty about the irreducibility of prior belief is admirable, but it comes at a cost: if every prior is equally permissible provided it is coherent, then Bayesian inference is not a method for discovering what is true but a method for discovering what follows from what you already believe. This is reasoning in a sealed room. The windows open only when subjectivity is supplemented by something the framework cannot itself provide — a mechanism for judging which priors are worth having._