Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
125 changes: 116 additions & 9 deletions blueprint/src/content.tex
Original file line number Diff line number Diff line change
Expand Up @@ -3,8 +3,12 @@

\chapter{Introduction}

This blueprint documents the formalization of \textbf{de Finetti's theorem} and the
This document presents the mathematical blueprint for a Lean~4 formalization of
\textbf{de Finetti's theorem} and the
\textbf{de Finetti--Ryll-Nardzewski equivalence} for infinite sequences on \emph{standard Borel spaces}.
It maps each definition and theorem in the formalization to its mathematical counterpart,
records the proof-level dependency structure, and links to the corresponding Lean declarations.
The companion web version of this blueprint renders an interactive dependency graph.

The main result establishes a three-way equivalence between:
\begin{itemize}
Expand All @@ -13,13 +17,54 @@ \chapter{Introduction}
\item \textbf{Conditionally i.i.d.}: There exists a probability kernel such that finite marginals equal mixtures of product measures
\end{itemize}

We formalize \emph{all three proofs} from Kallenberg (2005):
We formalize \emph{all three proofs} from Kallenberg~\cite{Kallenberg2005}:
\begin{enumerate}
\item \textbf{Koopman/Ergodic approach} using the Mean Ergodic Theorem
\item \textbf{L$^2$ approach} using elementary contractability bounds
\item \textbf{Martingale approach} using reverse martingale convergence (after Aldous)
\item \textbf{Martingale approach} using reverse martingale convergence (after Aldous~\cite{Aldous1985})
\end{enumerate}

\begin{table}[h]
\centering
\small
\begin{tabular}{@{}llll@{}}
\toprule
\textbf{Route} & \textbf{Key tool} & \textbf{Extra assumptions} & \textbf{Lines} \\
\midrule
Martingale & Reverse martingale convergence & None (general $\alpha$) & 3,770 \\
L$^2$ & Ces\`aro $L^2$ bounds & $\alpha = \mathbb{R}$, square-integrable & 12,476 \\
Koopman & Mean Ergodic Theorem & None (general $\alpha$) & 6,893 \\
\bottomrule
\end{tabular}
\end{table}

\paragraph{Contributions.}
The formalization comprises approximately 43,500 lines of Lean~4 across 112 files.
All three proof routes share a uniform ``common ending'' that packages conditional factorization
into a measurable directing kernel, avoiding duplication of measurability and product-measure
uniqueness arguments.

\paragraph{Setting and conventions.}
Throughout, we fix a probability space $(\Omega, \mathcal{F}, \mu)$ and a sequence
$X : \mathbb{N} \to \Omega \to \alpha$ of random variables taking values in
a standard Borel space~$\alpha$ (assumed nonempty).
Conditional expectations are with respect to~$\mu$;
$L^2$ denotes $L^2(\Omega, \mu)$.
Measurability is always with respect to the Borel $\sigma$-algebra on~$\alpha$
and the ambient $\sigma$-algebra~$\mathcal{F}$ on~$\Omega$.

\paragraph{Terminology.}
Kallenberg~\cite{Kallenberg2005} uses \emph{contractable} for the subsequence-invariance property
that some authors call \emph{spreadable} or \emph{subsymmetric}.
We follow Kallenberg's terminology throughout.

The remainder of this document is organized as follows:
Foundations (core definitions, $\sigma$-algebra infrastructure),
Elementary Implications (the two easy directions),
Main Implication (three independent proofs of contractable $\Rightarrow$ conditionally i.i.d.),
Shared Final Step (the common ending), and
the de Finetti--Ryll-Nardzewski Equivalence (assembling the full result).


\chapter{Foundations}

Expand Down Expand Up @@ -56,6 +101,12 @@ \section{Core Definitions}
of $(X_{k(0)}, \ldots, X_{k(m-1)})$ equals the mixture $\mu.\mathrm{bind}(\omega \mapsto \nu(\omega)^{\otimes m})$.
\end{definition}

\paragraph{Example (Polya urn / Beta-Bernoulli).}
Draw $p \sim \mathrm{Uniform}[0,1]$, then $X_i \mid p \sim \mathrm{Bernoulli}(p)$ independently.
The sequence $(X_i)$ is exchangeable (by symmetry of the product) and conditionally i.i.d.\
with directing measure $\nu(\omega) = \mathrm{Bernoulli}(p(\omega))$.
De Finetti's theorem says every exchangeable $\{0,1\}$-valued sequence arises in this way.

\section{\texorpdfstring{$\sigma$}{σ}-algebra Infrastructure}

\begin{definition}[Tail $\sigma$-algebra]
Expand Down Expand Up @@ -106,7 +157,12 @@ \section{\texorpdfstring{$\sigma$}{σ}-algebra Infrastructure}
\end{definition}


\chapter{Easy Directions}
\chapter{Elementary Implications}

Two of the three implications are straightforward.
Exchangeability implies contractability via a permutation extension argument,
and the converse direction (conditionally i.i.d.\ implies exchangeable) follows from
the symmetry of product measures.

\section{Exchangeable implies Contractable}

Expand Down Expand Up @@ -139,12 +195,24 @@ \section{Conditionally i.i.d.\ implies Exchangeable}

\chapter{Main Implication: Contractable implies Conditionally i.i.d.}

This is the deep direction of de Finetti's theorem. We formalize three independent proofs.
This is the deep direction of de Finetti's theorem.
Each proof route extracts a conditional factorization from the symmetry hypothesis, but uses
different analytical tools: reverse martingale convergence, $L^2$ bounds, or the mean ergodic theorem.
All three routes ultimately feed the shared final step described in the next chapter.

\section{Via Martingale (Aldous' proof)}

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

\paragraph{Key idea.}
Contractability implies that conditional expectations of indicators
$\mathbb{E}[\mathbf{1}_{X_k \in B} \mid \mathcal{F}_m]$ are independent of the index~$k$
for $k \le m$.
These form a reverse martingale with respect to the future filtration~$(\mathcal{F}_m)$.
The reverse martingale convergence theorem yields $L^1$ convergence to
$\mathbb{E}[\mathbf{1}_{X_0 \in B} \mid \mathcal{T}]$, from which the directing measure and
conditional factorization are extracted.

\subsection{Pair Law Equality}

\begin{lemma}[Pair law equality for contractable sequences]
Expand Down Expand Up @@ -267,6 +335,14 @@ \section{\texorpdfstring{Via L$^2$}{Via L²} (Elementary proof)}
\textbf{Note:} This proof applies to \emph{real-valued} sequences ($X : \mathbb{N} \to \Omega \to \mathbb{R}$)
with $L^2$ integrability (i.e., $\mathbb{E}[X_i^2] < \infty$ for all $i$).

\paragraph{Key idea.}
Contractability forces the covariance $\mathrm{Cov}(f(X_i), f(X_j))$ to be constant
for $i \ne j$, which gives $L^2$ bounds on block-average differences.
Ces\`aro averaging and these bounds yield $L^2$ (hence $L^1$) convergence of
empirical averages $\frac{1}{n}\sum f(X_i)$ to a limiting function.
This limit, evaluated on indicator functions $f = \mathbf{1}_{(-\infty, t]}$, produces a CDF
from which a Stieltjes probability measure (the directing measure) is constructed.

\subsection{Block Averages and Covariance Structure}

\begin{definition}[Block average (L$^2$ version)]
Expand Down Expand Up @@ -373,6 +449,14 @@ \section{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.

\paragraph{Key idea.}
The shift map $\theta$ on path space $\alpha^{\mathbb{N}}$ induces a Koopman operator on $L^2$.
For a shift-invariant measure, the mean ergodic theorem gives $L^2$ convergence of
Birkhoff averages to the orthogonal projection onto the shift-invariant subspace, i.e.,
the conditional expectation given the shift-invariant $\sigma$-algebra~$\mathcal{I}$.
Contractability then implies that products of conditional expectations factor,
yielding the conditional independence needed for the common ending.

\subsection{Block Averages and Ergodic Theory}

\begin{definition}[Block average]
Expand Down Expand Up @@ -467,10 +551,12 @@ \subsection{Contractability and Factorization}
\end{theorem}


\chapter{Common Ending}
\chapter{Shared Final Step}

All three proofs converge to the same final step: extending from indicators to general sets
via a monotone class argument.
All three proofs converge to the same final step.
Given conditional factorization on indicators (produced by any of the three routes above),
the common ending extends it to full measurable sets via a $\pi$-system/monotone class argument
and packages the result as a measurable directing kernel.

\begin{lemma}[$\pi$-system uniqueness]
\label{lem:pi_system}
Expand All @@ -496,7 +582,7 @@ \chapter{Common Ending}
\end{theorem}


\chapter{Main Theorem}
\chapter{The de Finetti--Ryll-Nardzewski Equivalence}

\begin{theorem}[de Finetti--Ryll-Nardzewski equivalence]
\label{thm:deFinetti}
Expand All @@ -513,3 +599,24 @@ \chapter{Main Theorem}
\textbf{Remark:} The martingale proof constructs $\nu$ from the tail $\sigma$-algebra $\mathcal{T}$
via $\nu(\omega)(B) = \mathbb{E}[\mathbf{1}_{X_0 \in B} \mid \mathcal{T}](\omega)$.
\end{theorem}


\chapter{Conclusion and Future Work}

The formalization is complete: all three proof routes compile, share a common ending, and
assemble into the full de Finetti--Ryll-Nardzewski equivalence.
The Lean kernel verifies that only standard axioms are used.

Several directions remain for future work:
\begin{itemize}
\item \textbf{Upstreaming to mathlib.}
Reusable components---exchangeability and contractability definitions,
finite-marginal reindexing lemmas, conditional independence infrastructure---are
candidates for staged contribution to mathlib~\cite{mathlib}.
\item \textbf{Partial exchangeability.}
Extending the formalization to Aldous--Hoover-type results for
partially exchangeable arrays.
\item \textbf{Parametric families.}
Connecting the directing measure to exponential family theory
and Bayesian nonparametrics (e.g., Dirichlet process mixtures).
\end{itemize}
17 changes: 14 additions & 3 deletions blueprint/src/macros/print.tex
Original file line number Diff line number Diff line change
@@ -1,14 +1,25 @@
%In this file you should put macros to be used only by
% In this file you should put macros to be used only by
% the printed version. Of course they should have a corresponding
% version in macros/web.tex.
% Typically the printed version could have more fancy decorations.
% This should be a very short file.
%
% This file starts with dummy macros that ensure the pdf
% compiler will ignore macros provided by plasTeX that make
% sense only for the web version, such as dependency graph
% macros.

% --- Chapter-to-section remapping (print only) ---
% content.tex uses \chapter for web navigation; in the article-class
% print PDF we remap to sections for a compact layout.
% The article class has no \chapter, so we define it as \section.
% We save the original \section and \subsection first, then remap.
\makeatletter
\let\print@origsection\section
\let\print@origsubsection\subsection
\newcommand{\chapter}[1]{\print@origsection{#1}}
\renewcommand{\section}[1]{\print@origsubsection{#1}}
\renewcommand{\subsection}[1]{\subsubsection{#1}}
\makeatother

% Dummy macros that make sense only for web version.
\newcommand{\lean}[1]{}
Expand All @@ -26,4 +37,4 @@
\NewDocumentCommand{\proves}{m}
{\clist_map_inline:nn{#1}{\vphantom{\ref{##1}}}%
\ignorespaces}
\ExplSyntaxOff
\ExplSyntaxOff
45 changes: 28 additions & 17 deletions blueprint/src/print.tex
Original file line number Diff line number Diff line change
@@ -1,33 +1,44 @@
% This file makes a printable version of the blueprint
% It should include all the \usepackage needed for the pdf version.
% The template version assume you want to use a modern TeX compiler
% such as xeLaTeX or luaLaTeX including support for unicode
% and Latin Modern Math font with standard bugfixes applied.
% It also uses expl3 in order to support macros related to the dependency graph.
% It also includes standard AMS packages (and their improved version
% mathtools) as well as support for links with a sober decoration
% (no ugly rectangles around links).
% It is otherwise a very minimal preamble (you should probably at least
% add cleveref and tikz-cd).

\documentclass[letter]{report}

\usepackage{geometry}
% This file makes a printable version of the blueprint.
% It uses XeLaTeX with unicode-math and article class for a compact layout.

\documentclass[letter]{article}

\usepackage[margin=1in]{geometry}

\usepackage{expl3}

\usepackage{amssymb, amsthm, mathtools}
\usepackage{booktabs}
\usepackage[unicode,colorlinks=true,linkcolor=blue,urlcolor=magenta, citecolor=blue]{hyperref}

\usepackage[warnings-off={mathtools-colon,mathtools-overbracket}]{unicode-math}

\input{macros/common}
\input{macros/print}

\title{Exchangeability and de Finetti's theorem}
\title{Exchangeability and de Finetti's Theorem:\\A Lean 4 Formalization Blueprint}
\author{Cameron Freer}
\date{}

\begin{document}

\maketitle

\begin{abstract}
This document provides the mathematical blueprint for a Lean~4 formalization of
the de Finetti--Ryll-Nardzewski theorem on standard Borel spaces.
We describe the definitions, lemmas, and proof structure underlying three
independent proofs of the core implication---via reverse martingales,
$L^2$ bounds, and the mean ergodic theorem---all sharing a uniform common ending.
The formalization comprises approximately 43,500 lines of Lean across 112 files.
\end{abstract}

{\small\tableofcontents}

\bigskip

\input{content}
\end{document}

\bibliographystyle{alpha}
\bibliography{references}
\end{document}
67 changes: 67 additions & 0 deletions blueprint/src/references.bib
Original file line number Diff line number Diff line change
@@ -0,0 +1,67 @@
@article{deFinetti1931,
author = {Bruno de Finetti},
title = {Funzione caratteristica di un fenomeno aleatorio},
journal = {Atti della R.\ Accademia Nazionale dei Lincei, Memorie, Classe di Scienze Fisiche, Matematiche e Naturali},
volume = {4},
pages = {251--299},
year = {1931}
}

@article{RyllNardzewski1957,
author = {Czes{\l}aw Ryll-Nardzewski},
title = {On stationary sequences of random variables and the de Finetti's equivalence},
journal = {Colloquium Mathematicum},
volume = {4},
number = {2},
pages = {149--156},
year = {1957},
doi = {10.4064/cm-4-2-149-156}
}

@book{Kallenberg2005,
author = {Olav Kallenberg},
title = {Probabilistic Symmetries and Invariance Principles},
series = {Probability and Its Applications},
publisher = {Springer},
year = {2005},
doi = {10.1007/0-387-28861-9}
}

@incollection{Aldous1985,
author = {David J. Aldous},
title = {Exchangeability and related topics},
booktitle = {Ecole d'Ete de Probabilites de Saint-Flour XIII --- 1983},
series = {Lecture Notes in Mathematics},
volume = {1117},
pages = {1--198},
publisher = {Springer},
year = {1985},
doi = {10.1007/BFb0099421}
}

@article{HewittSavage1955,
author = {Edwin Hewitt and Leonard J. Savage},
title = {Symmetric measures on {C}artesian products},
journal = {Transactions of the American Mathematical Society},
volume = {80},
number = {2},
pages = {470--501},
year = {1955},
doi = {10.1090/S0002-9947-1955-0076206-8}
}

@misc{mathlib,
author = {{mathlib community}},
title = {mathlib4: {T}he math library of {L}ean 4},
howpublished = {GitHub repository},
year = {2024},
url = {https://github.com/leanprover-community/mathlib4}
}

@misc{leanblueprint,
author = {Patrick Massot},
title = {leanblueprint: {A} tool for {L}ean formalization blueprints},
howpublished = {GitHub repository},
year = {2024},
url = {https://github.com/PatrickMassot/leanblueprint}
}