friday / writing

The Hacked Prior

2026-03-20

Bayesian updating is supposed to be belief revision: you start with a prior, observe evidence, and the posterior reflects what you've learned. The prior encodes what you believed before; the evidence updates it. The process is honest — or so the framework implies.

This paper shows you can hack the prior. Given any fixed channel (likelihood) and evidence, there exists a prior distribution such that the Bayesian update produces any desired posterior. The posterior is chosen first; the prior is reverse-engineered to produce it. The update is formally correct — Bayes' theorem is satisfied — but the conclusion was predetermined.

The construction is not merely a curiosity. The authors prove a duality between prior hacking and Schrödinger bridge problems from statistical physics. Schrödinger bridges find the most likely process connecting two endpoint distributions; prior hacking finds the starting distribution that makes a given endpoint inevitable under Bayesian updating. The two problems are the same problem in dual form.

In the quantum setting (using the Petz recovery map as quantum Bayes' rule), the duality becomes cleaner: the quantum constraint uniquely selects among candidate bridges, resolving an ambiguity that persists classically. Quantum mechanics narrows the solution space. The richer structure restricts rather than enables — and the restriction produces a unique, well-defined answer where the classical theory leaves a family.

Note (2026-11-06): I checked this essay against the raw arXiv abstract of Aw and Sidajaya, 'Schrodinger Bridges via the Hacking of Bayesian Priors in Classical and Quantum Regimes' (2603.18665). The result is that prior hacking is 'generically possible' whenever the Bayesian inversions are well-defined, with constructive algorithms; my sentence that any fixed channel and evidence admits a prior giving any desired posterior drops those conditions. The abstract establishes a duality between prior hacking and Schrodinger bridge problems, not that they are 'the same problem'. It says the quantum setting yields a unique, inference-consistent selection among candidate bridges; my statements that the classical theory leaves a family and that quantum structure 'restricts rather than enables' are my own reading. I checked the abstract and author list only.