Exchangeability and de Finetti’s theorem

4 Main Implication: Contractable implies Conditionally i.i.d.

This is the deep direction of de Finetti’s theorem. We formalize three independent proofs.

4.1 Via Martingale (Aldous’ proof)

The martingale approach uses reverse martingale convergence to the tail \(\sigma \)-algebra.

4.1.1 Pair Law Equality

Lemma 12 Pair law equality for contractable sequences

For a contractable sequence and \(k \le m\), the joint distribution of \((X_k, X_{m+1}, X_{m+2}, \ldots )\) equals that of \((X_m, X_{m+1}, X_{m+2}, \ldots )\).

Lemma 13 Contractable distribution equality

For a contractable sequence, the joint distribution of \((X_k, \theta _{m+1} X)\) equals that of \((X_m, \theta _{m+1} X)\) for all \(k \le m\), where \(\theta _n\) is the shift operator.

Lemma 14 Conditional expectation of indicator equals under contractability

For contractable sequences and \(k \le m\): \(\mathbb {E}[\mathbf{1}_{X_k \in B} \mid \mathcal{F}_m] = \mathbb {E}[\mathbf{1}_{X_m \in B} \mid \mathcal{F}_m]\) a.s.

4.1.2 Kallenberg Chain and Convergence

Conditional expectations of indicators given the reverse filtration converge to conditional expectations given the tail \(\sigma \)-algebra.

Lemma 16 Conditional expectation convergence

For \(k \le m\) and measurable \(B\): \(\mathbb {E}[\mathbf{1}_{X_m \in B} \mid \mathcal{F}_m] = \mathbb {E}[\mathbf{1}_{X_k \in B} \mid \mathcal{F}_m]\) a.s.

Lemma 17 Extreme members equal on tail

For any measurable \(B\): \(\mathbb {E}[\mathbf{1}_{X_m \in B} \mid \mathcal{T}] = \mathbb {E}[\mathbf{1}_{X_0 \in B} \mid \mathcal{T}]\) a.s.

4.1.3 Factorization and Directing Measure

Lemma 18 Finite level factorization

For finite products of indicators, the conditional expectation given \(\mathcal{F}_m\) factors as a product of individual conditional expectations.

Lemma 19 Tail factorization from future

The tail \(\sigma \)-algebra factorization follows from the finite-level factorization via reverse martingale convergence.

Lemma 20 Block coordinate conditional independence

Coordinates are conditionally independent given the tail \(\sigma \)-algebra.

Lemma 21 Directing measure is probability measure

The directing measure \(\nu (\omega )(B) = \mathbb {E}[\mathbf{1}_{X_0 \in B} \mid \mathcal{T}](\omega )\) is a probability measure for every \(\omega \) (the bound is pointwise, not merely a.e., via mathlib’s isProbability_condExpKernel).

Lemma 22 Directing measure measurability

The directing measure \(\omega \mapsto \nu (\omega )(B)\) is measurable for each Borel \(B\).

Lemma 23 Conditional law of \(X_n\) equals directing measure

Per-coordinate identification: for every \(n\) and every Borel \(B\), \((\omega \mapsto \nu (\omega )(B).\mathrm{toReal}) \, =^{\mu \text{-a.e.}} \, \mathbb {E}[\mathbf{1}_B \circ X_n \mid \mathcal{T}]\). This is the bridge condition consumed by 24 below.

Lemma 24 Finite product formula

Under contractability, for every strictly monotone selection \(k : \mathrm{Fin}\, m \to \mathbb {N}\) and any kernel \(\nu \) satisfying the per-coordinate bridge 23:

\[ \mathrm{Measure.map}\, (\omega \mapsto (i \mapsto X_{k\, i}(\omega )))\, \mu \, =\, \mu .\mathrm{bind}\, (\omega \mapsto \mathrm{Measure.pi}\, (\lambda \, i,\, \nu (\omega ))). \]

This is the martingale route’s analogue of the L²/Koopman 49: it bundles the rectangle \(\pi \)-system extension and the \(\mathrm{Measure.map} = \mathrm{bind}\) identification into one step driven by per-coordinate conditional expectations.

Theorem 25 Contractable implies Conditionally i.i.d. (via Martingale)

If \((X_n)\) is contractable, then it is conditionally i.i.d. The directing kernel \(\nu (\omega )(B) = \mathbb {E}[\mathbf{1}_{X_0 \in B} \mid \mathcal{T}](\omega )\) is constructed from the tail \(\sigma \)-algebra, identified per-coordinate by 23, and lifted to the joint law by 24 — bypassing the L²/Koopman common-ending step.

4.2 Via L\(^2\) (Elementary proof)

The L\(^2\) approach uses elementary contractability bounds on block averages. This is Kallenberg’s “second proof” and has the lightest dependencies.

Note: This proof applies to real-valued sequences (\(X : \mathbb {N} \to \Omega \to \mathbb {R}\)) with \(L^2\) integrability (i.e., \(\mathbb {E}[X_i^2] {\lt} \infty \) for all \(i\)).

4.2.1 Block Averages and Covariance Structure

Definition 26 Block average (L\(^2\) version)
#

The block average \(A_n = \frac{1}{n}\sum _{i=0}^{n-1} f(X_i)\) for bounded measurable \(f\).

Lemma 27 Contractable covariance structure

For contractable sequences, the covariance \(\mathrm{Cov}(f(X_i), f(X_j))\) is constant for \(i \ne j\).

Lemma 28 L\(^2\) contractability bound

For contractable sequences, certain L\(^2\) norms of block averages are bounded.

4.2.2 Cesaro Convergence

Recorded Kallenberg L\(^2\) bound.

Kallenberg’s Lemma 1.2 gives an L\(^2\) distance bound for weighted averages of an exchangeable sequence. It is recorded in the Lean library as Exchangeability.DeFinetti.ViaL2.kallenberg_L2_bound for completeness, but it is not on the active proof path — the route below uses the contractability-only variant 28 so that the proof of Contractable \(\Rightarrow \) Conditionally i.i.d. does not assume exchangeability.

Lemma 29 Block averages are Cauchy in L\(^2\)

For contractable \(X\) and bounded \(f\), the block Cesaro averages \(\mathrm{blockAvg}\, f\, X\, 0\, n\) form a Cauchy sequence in L\(^2\).

Lemma 30 Cesaro to conditional expectation (L\(^2\))

Block averages converge in L\(^2\) to the conditional expectation given the tail.

Lemma 31 Cesaro to conditional expectation (L\(^1\))

Block averages converge in L\(^1\) to the conditional expectation given the tail.

4.2.3 Directing Measure Construction

The L\(^2\) route now goes through mathlib’s condExpKernel machinery: the Stieltjes construction supplies the candidate directing measure \(\omega \mapsto \nu (\omega )\), and a pair of bridge lemmas identifies its push-forward and integral with the kernel coming from mathlib’s condExpKernel on the tail \(\sigma \)-algebra. The shared common-ending theorem then lifts the indicator-level identity to the full product \(\sigma \)-algebra.

Lemma 32 Directing measure is probability (L\(^2\) version)

The Stieltjes measure constructed from the limiting CDF is a probability measure for every \(\omega \) (pointwise, not merely a.e.). Monotonicity, bounds, and right-continuity of the underlying CDF are inherited directly from mathlib’s stieltjesOfMeasurableRat construction.

Theorem 33 Directing measure equals condExpKernel push-forward

Almost everywhere in \(\omega \), the Stieltjes-built directing measure coincides with the push-forward of mathlib’s condExpKernel \(\mu \) \(\mathcal{T}\) \(\omega \) by \(X_0\). This is the kernel-level bridge from the analytic L\(^2\) construction to mathlib’s conditional-expectation kernel API.

Lemma 34 Directing-measure integral equals conditional expectation

For a bounded measurable function \(f\), the integral \(\int f \, d\nu (\omega )\) against the directing measure equals \(\mathbb {E}[f \circ X_0 \mid \mathcal{T}](\omega )\) almost everywhere. This is the shape consumed by the shared common-ending step.

Theorem 35 Weighted sums converge in L\(^1\)

The weighted sums of indicators converge in L\(^1\).

Theorem 36 Contractable implies Conditionally i.i.d. (via L\(^2\))

If \((X_n)\) is contractable (and real-valued), then it is conditionally i.i.d. Scope note: the L\(^2\) route is currently stated for real-valued sequences via the Stieltjes-based directing-measure construction; extending it to general standard Borel targets would require a measure-disintegration replacement for the Stieltjes step.

4.3 Via Koopman (Mean Ergodic Theorem)

The Koopman approach uses the Mean Ergodic Theorem via the shift operator on L\(^2\). This is Kallenberg’s “first proof” and uses disjoint-block averaging.

Scope note: the Koopman route is currently stated for real-valued sequences (the Mean Ergodic Theorem is invoked through L\(^2(\mu ; \mathbb {R})\) on path space). Generalising to standard Borel targets would require an L\(^2\)-valued ergodic infrastructure beyond what is presently formalised. The default project export (50) handles general standard Borel spaces via the martingale route.

4.3.1 Block Averages and Ergodic Theory

Definition 37 Block average

The block average \(A_{m,n,k}(f)(\omega ) = \frac{1}{n} \sum _{j=0}^{n-1} f(\omega _{k \cdot n + j})\) averages \(f\) over the \(k\)-th block of size \(n\) (indices \([kn, kn+n)\)). For \(n = 0\), the block average is defined as \(0\).

Lemma 38 Koopman-condexp commutation

The conditional expectation operator onto the shift-invariant subspace commutes with the Koopman operator.

Theorem 39 Birkhoff averages converge to condexp

Birkhoff averages converge in L\(^2\) to the conditional expectation given the shift-invariant \(\sigma \)-algebra.

Lemma 40 Block averages converge in \(L^1\)

For a shift-invariant measure, block averages converge in \(L^1\) to the conditional expectation given the shift-invariant \(\sigma \)-algebra: \(\int |A_{m,n,k}(f) - \mathbb {E}[f \circ \pi _0 \mid \mathcal{I}]| \, d\mu \to 0\) as \(n \to \infty \).

4.3.2 Contractability and Factorization

Lemma 41 Conditional expectation lag constancy from exchangeability

For exchangeable sequences, the conditional expectation of a product does not depend on the lag between coordinates.

Lemma 42 Integral product equals block average product

For contractable sequences, integrals of products factor through block averages.

Lemma 43 Product block average L\(^1\) convergence

Products of block averages converge in L\(^1\).

Theorem 44 Conditional expectation product factorization

For contractable sequences, the conditional expectation of a product of indicators factors as a product of conditional expectations.

Lemma 45 Bridge from contractability

For contractable sequences, indicator products satisfy the bridge condition.

Theorem 46 Contractable implies Conditionally i.i.d. (via Koopman)

If \((X_n)\) is contractable, then it is conditionally i.i.d. This proof uses the Mean Ergodic Theorem via the Koopman operator on L\(^2\).