Disintegration of Measures

Elementary probability defines P(A|B) only when P(B) > 0. The Disintegration Theorem resolves this radically: it decomposes any measure into fibers and gives rigorous meaning to conditional distributions even when P(B) = 0.

• **AlphaGo and conditional probabilities:** DeepMind trained AlphaGo on 50 million Go positions; every move-selection step computes P(move | position). Mathematically this is the conditional measure mu_y delivered by disintegration
• **Spotify Bayesian models:** in 2023 Spotify processed 100 billion streams using Bayesian models built on P(genre | listening history) - formally a disintegration of the joint measure along the observation projection
• **Optimal transport:** a transport plan γ on X×Y disintegrates over X, giving for each x the distribution γ_x of destinations - the standard computational step in Wasserstein distance evaluation
• **Diffusion models (DDPM, score matching):** the score function ∇ log p_t(x) is the gradient of the log marginal density, obtained by integrating conditional measures over the noise schedule

Предварительные знания

• Polish spaces: complete separable metric spaces (R^n, C([0,1]), l^2)
• Borel sigma-algebra and measurable maps
• Pushforward measure: nu = pi_* mu, nu(B) = mu(pi^{-1}(B))
• Conditional expectation E[f | G] as an L^2 projection

• Conditional Expectation
• Measures and Integration

The Disintegration Theorem

In 2014, Google DeepMind trained AlphaGo on 50 million Go positions, with conditional probability at the core of every move-selection step. The rigorous foundation is the Disintegration Theorem: any measure can be decomposed along the fibers of a measurable map, yielding a family of probability measures, one per fiber.

Conditional Measures and Optimal Transport

Spotify processed 100 billion streams in 2023 using Bayesian models built on explicit conditional probability structures. The disintegration theorem gives rigorous meaning to P(A|B) even when P(B) = 0, resolving a fundamental gap left by elementary probability theory.

Disintegration across measure theory and related fields

The Disintegration Theorem bridges abstract measure theory and applied conditional distributions. It unifies the iterated-integral formula, the Fubini-Tonelli theorem, and regular conditional distributions under one roof.

• Optimal transport — Disintegrating the plan γ ∈ Π(μ,ν) over the first coordinate gives the conditional plan γ_x, the key step in numerical OT
• Fubini-Tonelli theorem — Fubini is a special case: for product measures the conditional measures μ_y are equal to the second marginal
• Bayesian inference — The posterior P(θ|x) is the disintegration of the joint P(θ,x) along the observation x
• Diffusion models — Score matching uses conditional densities p(x|t), which are formally the conditional measures from disintegration over the noise schedule

Итоги

• **Disintegration theorem:** for Polish X, measurable pi: X → Y, and measure mu there exists a nu-a.e. unique family {mu_y} concentrated on the fibers pi^{-1}(y) such that mu = ∫ mu_y dν(y)
• **Iterated integration:** ∫_X f dμ = ∫_Y (∫_{pi^{-1}(y)} f dμ_y) dν(y) - generalizes Fubini to non-product measures
• **Regular conditional distribution:** P(A|Y=y) = mu_y(A) resolves conditioning on zero-probability events
• **Bivariate Gaussian:** mu_y = N(ρy, 1-ρ²) is the conditional measure in closed form, a consequence of the closure of the Gaussian family under conditioning
• **OT application:** disintegrating the optimal plan γ over x yields point masses γ_x = δ_{σ(x)} for monotone transport on R