LemmaStatement: Literature-sourcedProof: AI-adaptedPipeline-generatedverified 2026-10-22 (gpt-6.1-sol)

How statement and proof provenance work

The first chip identifies the source of the statement or construction; the second identifies the source of its local proof or verification.

Literature-sourced: the exact statement appears in a cited source; only wording and notation differ.
AI-adapted: a semantically identical restatement of literature-sourced material, modulo indexing, notation, and boundary cases adopted by the library.
AI-generated: a genuinely novel statement formulated by AI, with no source for the claim itself.

These labels describe origin, not correctness: citations and verification chips remain separate evidence.

Population regression and exact-flow meaning of conditional flow matching

Statement

Assume the Axiom of Choice. Let $d, m, n \geq 1$ , let $η$ be a probability measure on $R^{d}$ , and let $π$ be a proper Bayesian joint law on $R^{d} \times R^{m}$ with data marginal $π_{Y}$ and posterior-kernel version $π_{y}$ . Assume $\int ∣ x_{0} ∣^{2} η (d x_{0}) + \int ∣ x_{1} ∣^{2} π (d x_{1}, d y) < \infty$ . Let $t$ , $X_{0}$ and $(X_{1}, Y)$ be independent with law $Unif [0, 1] \otimes η \otimes π$ , where $X_{t} = (1 - t) X_{0} + t X_{1}$ , and $V = X_{1} - X_{0}$ has finite second moment. Write $L (v) : = E ∣ v (t, X_{t}, Y) - V ∣^{2}$ and let ${\hat{L}}_{n} (v)$ be the empirical mean of this squared residual over $n \geq 1$ independent training triples. Then the conditional-expectation class $v^{*} (t, X_{t}, Y) = E [V ∣ σ (t, X_{t}, Y)]$ admits a measurable representative in $(t, x, y)$ and uniquely minimizes $L$ over square-integrable predictors measurable in these coordinates, up to almost-sure equality. For every fixed measurable candidate $v$ with finite population loss, $E [{\hat{L}}_{n} (v)] = L (v) .$ This expectation identity is for a fixed candidate, not a data-dependent fitted field.

If a fitted field satisfies all hypotheses of the global ODE flow lemma, it has an exact triangular terminal map $T (x, y) = (T_{y} (x), y)$ . If also $T_{#} (η \otimes π_{Y}) = π$ , then $T_{y}_{#} η = π_{y}$ as probability measures outside one $π_{Y}$ -null set. These conclusions make no assertion that finite training or numerical solver success establishes the exact joint pushforward identity.

Statement references

Conditional flow matching supplies the training law and empirical objective.

Blechschmidt et al., section 3.2 and Theorem 3.1

Facts & Assumptions

Given: The hypotheses, training law, independent triples and fixed candidate in the Statement; the algorithm supplies the empirical-objective convention.

[A1]

AC is assumed for the conditional-expectation projection and standard-Borel posterior-kernel interfaces. Its implications include Countable Choice and Dependent Choice for the flow lemma. The Axiom of Choice, AC implies DC implies countable choice, Conditional expectation is the l2 orthogonal projection, Misfit evidence and the potential posterior law

[F1]

In a proper Bayesian model, $π$ is the joint unknown/data law and $π_{y}$ is a version of the posterior conditional law; values on a $π_{Y}$ -null set are not fixed by the conditional identity. Misfit evidence and the potential posterior law, Conditional law given a random element

[F2]

For a square-integrable target, the conditional expectation given $σ (t, X_{t}, Y)$ is the unique $L^{2}$ minimizer of the population squared prediction loss, coordinatewise. Conditional expectation given a sigma algebra, Conditional expectation is the l2 orthogonal projection

[F3]

Under AC, given a disintegration kernel for a standard-Borel random element $U$ conditional on a standard-Borel random element $W$ , integrating an integrable real function of $U$ against that kernel, with zero filling where its absolute integral is infinite, gives a measurable function of $W$ representing the conditional expectation almost surely. Finite-dimensional Euclidean Borel spaces are standard Borel. Conditional expectation as a measurable function of the conditioning variable, Finite-dimensional Euclidean space with its Borel sigma-algebra is standard Borel

[F4]

Independent sampling from $Unif [0, 1]$ , $η$ , and $π$ gives the stated product training law; product probabilities have the rectangle formula, and the expectation of a finite sum of integrable losses is the sum of their expectations. For sigma-finite factors, the product measure exists, has the rectangle formula, is sigma-finite, and is unique, The Lebesgue integral is linear on $L^{1} (μ)$

[F5]

Under [A1], a fitted field jointly continuous in $(t, x, y)$ and satisfying the ODE lemma's compact-parameter-uniform integrable global spatial-Lipschitz and origin-growth bounds, holding almost everywhere in time simultaneously for all states and parameters in each compact set, has a unique global flow with bijective fibers and a jointly continuous terminal map; these facts do not certify that its output law is the target. Global Lipschitz ODE flows and the C1 density formula

[F6]

If the exact triangular terminal map has joint pushforward $π$ , its exact conditional fibers push $η$ to $π_{y}$ for $π_{Y}$ -almost every datum; this needs the joint identity as a separate hypothesis. An exact triangular transport disintegrates to the posterior

[F7]

The six-term posterior-computation budget is a telescoping ledger and supplies no learned-flow regression or optimizer bound. Posterior computation error budget

[F8]

Under AC, the joint probability law of two standard-Borel random elements admits a probability disintegration kernel giving the conditional law of the first given the second. Disintegration of a joint law on standard borel spaces

[F0]

The training law, empirical objective and procedure are as specified in Conditional flow matching for inverse posteriors.

Proof

Given: The hypotheses in the Statement and the empirical-objective convention [F0].

1.1A1F0F2F3F4F8givenalgebra

The inequality $∣ V ∣^{2} \leq 2 ∣ X_{1} ∣^{2} + 2 ∣ X_{0} ∣^{2}$ makes $V$ square-integrable. By [A1] and [F2], coordinatewise conditional expectation uniquely minimizes the sum of squared prediction errors among square-integrable predictors measurable with respect to $σ (t, X_{t}, Y)$ , up to almost-sure equality. Set $W = (t, X_{t}, Y)$ , viewed as a random element of $R^{1 + d + m}$ . The maps defining $V$ and $W$ are Borel, and their Euclidean target spaces are standard Borel by [F3]. Thus [A1] and [F8] applied to the joint law of $(V, W)$ supply a kernel $K (w, d u)$ for $V$ given $W = w$ . Apply [F3] to each integrable coordinate function $u \mapsto u_{j}$ : its zero-filled kernel integral defines a real measurable $v_{j}^{*} (w)$ with $v_{j}^{*} (W) = E [V_{j} ∣ σ (W)]$ almost surely. Combining the finitely many coordinates gives the required measurable $v^{*} (t, x, y)$ . For a fixed measurable $v$ with finite loss, every sampled squared residual has expectation $L (v)$ under the specified product law. Linearity in [F4] therefore gives $E [{\hat{L}}_{n} (v)] = n^{- 1} \sum_{i = 1}^{n} L (v) = L (v)$ . This does not apply to substituting a data-dependent fitted field into the fixed-candidate expectation identity.

2.1F1F5F6F7given∎

For a fitted field satisfying all stated ODE hypotheses, [F5] gives its exact flow, joint continuity of the terminal map and bijective fibers. If that map also satisfies the exact joint pushforward hypothesis, [F1] and [F6] identify its image of $η$ with the posterior kernel outside one $π_{Y}$ -null set. A finite optimizer status or a solver-success response supplies neither that joint identity nor an exact numerical endpoint. The bookkeeping budget in [F7] supplies no missing regression or optimizer estimate.

Depends on

Used by

Nothing in the library uses this result yet.

Dependency tree · two levels

120 results within two dependency steps of this one, each drawn at its shortest distance from it. An arrow runs from a result to what uses it, so the chart reads left to right and ends at this result, which carries a heavier outline. Every node is a link to that result. Click elsewhere on the chart to enlarge it.

Sources

Blechschmidt et al., Flow Matching for Fast Posterior Sampling in Bayesian Inverse Problems (standard reference, not scraped)