Why This Matters — The Physics Bridge

This folder has developed probability as mathematics. Physics uses it in three structurally different ways, and they are worth separating, because the same measure-theoretic apparatus plays a different role in each.

Only the middle column requires the generalization of quantum-probability.md; the other two are ordinary Kolmogorov probability applied to different objects.

1. Statistical Mechanics

The microcanonical, canonical, and grand canonical ensembles are probability measures on phase space: normalized Liouville measure on an energy shell, $Z^{-1}{e}^{-\beta H}d\Gamma$, and so on. Liouville's theorem says the Hamiltonian flow preserves $d\Gamma$, so the flow is a measure-preserving transformation and Birkhoff's ergodic theorem (laws-of-large-numbers.md §4) applies directly: time averages equal $E[\cdot \mid \mathcal{I}]$, and equal phase averages exactly when the flow is ergodic.

That is the classical justification of the ensemble formalism, and its limitations are precisely the ones the theorem states:

• Ergodicity is a hypothesis, not a theorem, and it fails for integrable systems (harmonic solids, KAM tori). Where it fails, Birkhoff still applies but the limit is a random variable — the ergodic decomposition, exactly as in exchangeability §3.
• Ergodicity says nothing about time scales. Equality of time and phase averages holds in the infinite-time limit; relaxation times can exceed the age of the universe.
• The modern replacement is typicality: rather than time-averaging, one argues that the overwhelming majority of microstates compatible with a macrostate exhibit the thermodynamic behaviour, "overwhelming majority" being measured by the same Liouville measure. This shifts the burden onto the choice of measure — an instance of the underdetermination that Bertrand's paradox §4 makes vivid, and the reason the past hypothesis must be added by hand to explain the thermodynamic arrow (arrow of time).

Entropy links the two subjects: $S=k_B\log W$ is a statement about the measure of a macrostate, and the Gibbs/Shannon form $-k_B{\sum }_i{p}_i\log p_i$ is a functional of a probability distribution. Fluctuation results (Einstein's relation, the fluctuation–dissipation theorem) are CLT-type statements about $\sqrt{N}$ deviations.

2. Quantum Mechanics

The dictionary, assembled from quantum-probability.md and functional analysis:

Three points where the mathematics of this folder does real work:

• The Born rule's form is a theorem. Gleason's theorem (quantum-probability.md §2) derives $\mathrm{Tr}(\rho \Pi )$ from additivity over orthogonal alternatives, so the quadratic form of $|⟨\psi |\varphi ⟩{|}^2$ is not an extra postulate — a point recorded in QFT remarks.
• Continuous spectra need measures, not eigenvalues. Position and momentum have no eigenvectors in $\mathcal{H}$; "$|x⟩$" is a distributional object, and the probability of finding the particle in $B$ is the spectral measure $⟨\psi ,E(B)\psi ⟩$ — a genuine probability measure on $\mathbb{R}$ in the sense of random-variables.md §2. The rigorous setting is rigged Hilbert space.
• Decoherence produces a classical measure. Interaction with an environment suppresses off-diagonal terms of $\rho$ in a preferred basis; the resulting diagonal $\rho$ is, mathematically, an ordinary probability distribution on a Boolean sublattice. Decoherence explains the emergence of the Boolean structure and thus of Kolmogorov probability, without by itself explaining outcomes.

Path integrals connect the two subjects again: the Euclidean (Wick-rotated) path integral is an integral against a genuine probability measure — Wiener measure or a Gaussian process (brownian-motion.md §4) — whereas the Lorentzian one is an oscillatory integral with no measure behind it (QM path integral, QFT generating functional). That asymmetry is why Euclidean field theory is mathematically better behaved.

3. Randomized Computation

$\mathsf{BPP}$, $\mathsf{RP}$, and $\mathsf{ZPP}$ (complexity-theory.md §7) are defined by acceptance probabilities over the algorithm's internal coin flips — a product measure on $\{0,1{\}}^{\mathrm{poly}(n)}$, entirely classical. Two results from this folder carry the weight:

• Amplification is Hoeffding/Chernoff (expectation.md §6): majority voting over $\mathrm{poly}(n)$ independent runs drives the error to $2^{-\mathrm{poly}(n)}$, which is why the constants in the definition of $\mathsf{BPP}$ are arbitrary.
• What "random bits" are is the subject of algorithmic-randomness.md; the resource-bounded analogue is pseudorandomness, and the conjecture $\mathsf{BPP}=\mathsf{P}$ says that randomness buys no super-polynomial speedup because sufficiently strong circuit lower bounds would let it be simulated.

Quantum computation ($\mathsf{BQP}$) is where columns 2 and 3 meet: amplitudes interfere and only the final measurement produces a Kolmogorov probability, which is exactly why $\mathsf{BQP}$ is not obviously contained in the classical randomized classes.

4. The Common Lesson

Across all three, the mathematics is silent on the same question. The axioms determine the calculus completely and determine the modelling not at all: which measure, on which space, with which conditioning family, is an input in statistical mechanics (the ensemble, the past hypothesis), in quantum mechanics (the state, and the interpretation of the Born probabilities), and in computation (the designer's coin distribution). That is the content of kolmogorov-axioms.md §2, restated three times by physics.

5. Where This Is Used

• Quantum probability — the non-Boolean case in detail.
• Laws of large numbers §4 — the ergodic theorem behind §1.
• Expectation §6 — the concentration bounds behind §3.
• QM postulates, QFT remarks — the physics pages this folder serves.