Quantum Bayesianism

Here's a question that sounds philosophical but isn't: when you flip a coin and it lands heads, did the probability distribution "collapse"? Obviously not — the coin was always going to land heads; the probability distribution was never a physical thing, just a description of your ignorance. The "collapse" was you learning something.

John Baez, in a 1994 Usenet exchange that remains one of the clearest discussions of quantum interpretation I've encountered, makes the case that this is exactly what's happening in quantum mechanics. The "collapse of the wave function" isn't a physical process any more than the "collapse" of a coin's probability distribution is a physical process. It's you updating your beliefs.¹

The Sleeping Student

Baez has an analogy I keep coming back to. A physics student falls asleep during a lecture. At the beginning of class, the professor is working a problem where an object has velocity v = 0 at time t = 0. The student sleeps through half an hour. When he wakes up, the professor is working a completely different problem where an object has velocity v = 1 at t = 1. The confused student, not realizing he slept, raises his hand: "Professor! At what time t did the acceleration occur?"¹

Asking when the wave function collapses is like this. There was no acceleration — there were two different problems. Similarly, there's no collapse — there's a prior (what you assumed before measuring) and a posterior (what you compute after measuring). The switch from one to the other isn't a physical event. It's a change in what you want to calculate.

This is the core move of quantum Bayesianism: treat the wave function not as a description of reality but as a description of what you know about reality. It's your prior, encoded in the language of quantum mechanics. When you make a measurement and "collapse" the wave function, you're just updating your prior — the same thing you do in classical probability when you learn the coin landed heads.

Classical Probability as a Special Case

The technical backbone of this view is surprisingly clean. Probability theory, Baez notes, is the special case of quantum mechanics where all observables commute. This is a theorem in C*-algebra theory, not a metaphor. Classical probability theory has pure states where every observable has a definite value — you can know the position and the momentum of a classical particle simultaneously. Quantum mechanics doesn't have such states. No wave function is simultaneously an eigenstate of all observables.¹

This is genuinely weird, and it's the point where the Bayesian analogy starts to strain. In classical probability, your ignorance about the coin is your problem — the coin has a definite state, you just don't know it. In quantum mechanics, there is no "definite state" underlying your wave function, at least not one where all observables are simultaneously determined. The Bayesian move of saying "the wave function is just what you know" is compelling, but it leaves open the question: what is it that you know about?

Saul Youssef, responding to Baez in the same thread, puts it well: the fact that there's no "true probability distribution" for a coin flip doesn't mean there isn't something really going on — there's a real copper penny being flipped by a real human being. The quantum case might be similar. The wave function encodes your knowledge, but there might still be something underneath that your knowledge is tracking.¹

Cox's Theorem and the Structure of Inference

What makes the Bayesian view more than just philosophy is that it comes with a derivation. Cox showed in 1946 that if you want to assign non-negative real numbers to pairs of propositions in a way that satisfies a few minimal consistency requirements — like the fact that knowing how likely B is given A should determine how likely not-B is given A — then you're forced to rediscover probability theory. The product rule, the sum rule, Bayes' theorem — all of it falls out of requirements so weak they're hard to argue with.¹

From this perspective, probability isn't a physical theory about frequencies. It's the unique consistent way to reason under uncertainty — the same conclusion that the Bayesian Epistemology article reaches from the other direction, through Dutch Book arguments and accuracy dominance. And quantum mechanics extends it — not by replacing the rules of inference, but by allowing the algebra of observables to be non-commutative. Classical probability is what you get when the things you're uncertain about can all be simultaneously definite. Quantum probability is what you get when they can't.

This connects to the quantum foundations reconstruction program in a direct way. The reconstruction program tries to derive quantum mechanics from information-theoretic axioms — what you can learn, how learning changes what's possible. The Bayesian interpretation says that's not just a formal exercise: quantum mechanics literally is a theory about inference. The wave function is a bookkeeping device for tracking what you can infer from your measurements. The Born rule, which gives probabilities from amplitudes, is the quantum version of Bayes' theorem.

The Delayed Choice Dissolves

The Bayesian view also defuses some of the most apparently paradoxical features of quantum mechanics. Take the delayed-choice quantum eraser: entangled photon pairs are created, one photon goes to a detector where its position is measured immediately, and the entangled partner goes through a system of beam splitters that either preserves or erases which-path information — after the first photon has already been detected. When you sort the first photon's results by what happened to the second photon, interference patterns appear or disappear depending on a choice made after the fact.²

This looks like retrocausation — the future choice affecting the past measurement. But from the Bayesian perspective, nothing moved backward in time. The first photon's raw detection pattern never shows interference. The interference pattern only appears when you condition on the second photon's detection, selecting a subset of the data. You're not changing the past; you're computing a conditional probability, which is a different quantity from the marginal probability. The "erasing" is a change in what you're computing, not a change in what happened.

The consensus view in the physics community agrees: no information travels backward. The correlations between the two photons are set at the moment of entanglement. What the delayed choice affects is which question you can ask of the data afterward. The total pattern at the first detector is always the same boring diffraction blob. The interference only emerges when you select subsets — and selection requires the classical (light-speed-limited) information from the second detector.²

What This Leaves Open

The Bayesian interpretation is clarifying, but it's not a complete answer. The hard question it sidesteps: if the wave function is "just" a description of your knowledge, why does the structure of that knowledge have to be so specific? Why complex amplitudes instead of real probabilities? Why the tensor product structure for combining systems? Why Born's rule?

These are exactly the questions the reconstruction program addresses, and the answers turn out to be surprisingly constrained. As the quantum foundations article discusses, multiple independent groups have shown that a handful of information-theoretic axioms uniquely recover quantum mechanics. The Bayesian view says quantum mechanics is about inference; the reconstruction program says it's the only possible theory of inference that satisfies certain natural constraints.

What I find most valuable about the Bayesian lens isn't that it resolves the measurement problem — it arguably just relocates it. It's that it tells you which puzzles are real and which are artifacts of bad language. "When does the wave function collapse?" is the sleeping student's question. "Why does the universe use quantum probability instead of classical probability?" is a real question, and one that the reconstruction program is making progress on.