|
\documentclass[10pt,twocolumn,conference]{IEEEtran} |
|
|
|
|
|
\usepackage{graphicx} |
|
\usepackage{listings,color} |
|
\usepackage{listings} |
|
\renewcommand{\ttdefault}{lmtt} |
|
\ttfamily |
|
\DeclareFontShape{OT1}{lmtt}{m}{it} |
|
{<->sub*lmtt/m/sl}{} |
|
\lstdefinestyle{default}{language=Anglican, basicstyle=\ttfamily, columns=flexible, showstringspaces=false} |
|
|
|
\usepackage{listings} |
|
\usepackage{color} |
|
\usepackage{xcolor} |
|
\definecolor{blue}{rgb}{0,0.3,0.7} |
|
\definecolor{red}{rgb}{0.60,0.0,0.0} |
|
\definecolor{purple}{rgb}{0.5,0,0.7} |
|
\definecolor{cyan}{rgb}{0.0,0.6,0.5} |
|
\definecolor{gray}{rgb}{0.4,0.4,0.4} |
|
|
|
\lstdefinelanguage{scheme} |
|
{sensitive, alsoletter={:,-,+,*,?,/,!,>,<}, morecomment=[l];, }[comments] |
|
|
|
\lstdefinelanguage{anglican}{morekeywords=[1]{}, |
|
morekeywords=[2]{def, def-, defn, defn-, defmacro, defmulti, defmethod, defstruct, defonce, declare, definline, definterface, defprotocol, defrecord, defstruct, deftype, defproject, ns, }, morekeywords=[3]{->, ->>, .., amap, and, areduce, as->, assert, binding, bound-fn, case, comment, cond, cond->, cond->>, condp, declare, definline, definterface, defmacro, defmethod, defmulti, defn, defn-, defonce, defprotocol, defrecord, defstruct, deftype, delay, doseq, dosync, dotimes, doto, extend-protocol, extend-type, fn, for, future, gen-class, gen-interface, if, if-let, if-not, if-some, import, io!, lazy-cat, lazy-seq, let, letfn, locking, loop, memfn, ns, or, proxy, proxy-super, pvalues, recur, refer-clojure, reify, some->, some->>, sync, time, when, when-first, when-let, when-not, when-some, while, with-bindings, with-in-str, with-loading-context, with-local-vars, with-open, with-out-str, with-precision, with-redefs}, morekeywords=[4]{*, *', +, +', -, -', ->ArrayChunk, ->Vec, ->VecNode, ->VecSeq, -cache-protocol-fn, -reset-methods, /, <, <=, =, ==, >, >=, accessor, aclone, add-classpath, add-watch, agent, agent-error, agent-errors, aget, alength, alias, all-ns, alter, alter-meta!, alter-var-root, ancestors, apply, array-map, aset, aset-boolean, aset-byte, aset-char, aset-double, aset-float, aset-int, aset-long, aset-short, assoc, assoc!, assoc-in, associative?, atom, await, await-for, await1, bases, bean, bigdec, bigint, biginteger, bit-and, bit-and-not, bit-clear, bit-flip, bit-not, bit-or, bit-set, bit-shift-left, bit-shift-right, bit-test, bit-xor, boolean, boolean-array, booleans, bound-fn*, bound?, butlast, byte, byte-array, bytes, cast, char, char-array, char?, chars, chunk, chunk-append, chunk-buffer, chunk-cons, chunk-first, chunk-next, chunk-rest, chunked-seq?, class, class?, clear-agent-errors, clojure-version, coll?, commute, comp, comparator, compare, compare-and-set!, compile, complement, concat, conj, conj!, cons, constantly, construct-proxy, contains?, count, counted?, create-ns, create-struct, cycle, dec, dec', decimal?, delay?, deliver, denominator, deref, derive, descendants, destructure, disj, disj!, dissoc, dissoc!, distinct, distinct?, doall, dorun, double, double-array, doubles, drop, drop-last, drop-while, empty, empty?, ensure, enumeration-seq, error-handler, error-mode, eval, even?, every-pred, every?, ex-data, ex-info, extend, extenders, extends?, false?, ffirst, file-seq, filter, filter-ns-publics, filterv, find, find-keyword, find-ns, find-protocol-impl, find-protocol-method, find-var, first, flatten, float, float-array, float?, floats, flush, fn?, fnext, fnil, force, format, frequencies, future-call, future-cancel, future-cancelled?, future-done?, future?, gensym, get, get-in, get-method, get-proxy-class, get-thread-bindings, get-validator, group-by, hash, hash-combine, hash-map, hash-ordered-coll, hash-set, hash-unordered-coll, identical?, identity, ifn?, in-ns, inc, inc', init-proxy, instance?, int, int-array, integer?, interleave, intern, interpose, into, into-array, ints, isa?, iterate, iterator-seq, juxt, keep, keep-indexed, key, keys, keyword, keyword?, last, line-seq, list, list*, list?, load, load-file, load-reader, load-string, loaded-libs, long, long-array, longs, macroexpand, macroexpand-1, make-array, make-hierarchy, map, map-indexed, map?, mapcat, mapv, max, max-key, memoize, merge, merge-with, meta, method-sig, methods, min, min-key, mix-collection-hash, mod, munge, name, namespace, namespace-munge, neg?, newline, next, nfirst, nil?, nnext, not, not-any?, not-empty, not-every?, not=, ns-aliases, ns-functions, ns-imports, ns-interns, ns-macros, ns-map, ns-name, ns-publics, ns-refers, ns-resolve, ns-unalias, ns-unmap, nth, nthnext, nthrest, num, number?, numerator, object-array, odd?, parents, partial, partition, partition-all, partition-by, pcalls, peek, persistent!, pmap, pop, pop!, pop-thread-bindings, pos?, pr, pr-str, prefer-method, prefers, print, print-ctor, print-simple, print-str, printf, println, println-str, prn, prn-str, promise, proxy-call-with-super, proxy-mappings, proxy-name, push-thread-bindings, quot, rand, rand-int, rand-nth, range, ratio?, rational?, rationalize, re-find, re-groups, re-matcher, re-matches, re-pattern, re-seq, read, read-line, read-string, realized?, record?, reduce, reduce-kv, reduced, reduced?, reductions, ref, ref-history-count, ref-max-history, ref-min-history, ref-set, refer, release-pending-sends, rem, remove, remove-all-methods, remove-method, remove-ns, remove-watch, repeat, repeatedly, replace, replicate, require, reset!, reset-meta!, resolve, rest, restart-agent, resultset-seq, reverse, reversible?, rseq, rsubseq, satisfies?, second, select-keys, send, send-off, send-via, seq, seq?, seque, sequence, sequential?, set, set-agent-send-executor!, set-agent-send-off-executor!, set-error-handler!, set-error-mode!, set-validator!, set?, short, short-array, shorts, shuffle, shutdown-agents, slurp, some, some-fn, some?, sort, sort-by, sorted-map, sorted-map-by, sorted-set, sorted-set-by, sorted?, special-symbol?, spit, split-at, split-with, str, string?, struct, struct-map, subs, subseq, subvec, supers, swap!, symbol, symbol?, take, take-last, take-nth, take-while, test, the-ns, thread-bound?, to-array, to-array-2d, trampoline, transient, tree-seq, true?, type, unchecked-add, unchecked-add-int, unchecked-byte, unchecked-char, unchecked-dec, unchecked-dec-int, unchecked-divide-int, unchecked-double, unchecked-float, unchecked-inc, unchecked-inc-int, unchecked-int, unchecked-long, unchecked-multiply, unchecked-multiply-int, unchecked-negate, unchecked-negate-int, unchecked-remainder-int, unchecked-short, unchecked-subtract, unchecked-subtract-int, underive, unsigned-bit-shift-right, update-in, update-proxy, use, val, vals, var-get, var-set, var?, vary-meta, vec, vector, vector-of, vector?, with-bindings*, with-meta, with-redefs-fn, xml-seq, zero?, zipmap}, morekeywords=[5]{def-cps-fn, defanglican, defm, defquery, defun, defproc, defdist}, morekeywords=[6]{cps-fn, fm, lambda, mem, query, with-primitive-procedures}, morekeywords=[7]{doquery, conditional, collect-by, equalize, exec, infer, log-marginal, print-predicts, rand, rand-int, rand-nth, rand-roulette, stripdown, warmup, ->CRP-process, ->DP-process, ->GP-process, ->bernoulli-distribution, ->beta-distribution, ->binomial-distribution, ->categorical-crp-distribution, ->categorical-distribution, ->categorical-dp-distribution, ->chi-squared-distribution, ->dirichlet-distribution, ->discrete-distribution, ->exponential-distribution, ->flip-distribution, ->gamma-distribution, ->mvn-distribution, ->normal-distribution, ->poisson-distribution, ->sample, ->observe, sample*, observe*, ->uniform-continuous-distribution, ->uniform-discrete-distribution, ->wishart-distribution, CRP, DP, GP, abs, absorb, acos, asin, atan, bernoulli, beta, binomial, categorical, categorical-crp, categorical-dp, cbrt, ceil, chi-squared, cos, cosh, cov, dirichlet, discrete, exp, exponential, flip, floor, gamma, gen-matrix, log, log-gamma-fn, log-mv-gamma-fn, log-sum-exp, map->CRP-process, map->DP-process, map->GP-process, map->bernoulli-distribution, map->beta-distribution, map->binomial-distribution, map->categorical-crp-distribution, map->categorical-distribution, map->categorical-dp-distribution, map->chi-squared-distribution, map->dirichlet-distribution, map->discrete-distribution, map->exponential-distribution, map->flip-distribution, map->gamma-distribution, map->mvn-distribution, map->normal-distribution, map->poisson-distribution, map->uniform-continuous-distribution, map->uniform-discrete-distribution, map->wishart-distribution, mvn, normal, poisson, pow, produce, rint, round, signum, sin, sinh, sqrt, tag, tan, tanh, transform-sample, uniform-continuous, uniform-discrete, wishart, add-log-weight, add-predict, clear-predicts, get-log-weight, get-mem, get-predicts, in-mem?, set-log-weight, set-mem, }, morekeywords=[8]{factor, observe, predict, retrieve, sample, store}, sensitive, alsoletter={:,-,+,*,?,/,!,>,<}, morecomment=[l];, morestring=[b]", keywordsprefix=:, }[keywords,comments,strings] |
|
|
|
\lstset{ |
|
language=anglican, |
|
basicstyle=\ttfamily, |
|
commentstyle=\em\color{gray}, |
|
keywordstyle=[1]\color{red}, |
|
stringstyle=\color{gray}, |
|
keywordstyle=[2]\color{purple}\bfseries, |
|
keywordstyle=[3]\color{red}\bfseries, |
|
keywordstyle=[4]\color{cyan}\bfseries, |
|
keywordstyle=[5]\color{purple}\bfseries, |
|
keywordstyle=[6]\color{red}\bfseries, |
|
keywordstyle=[7]\color{cyan}\bfseries, |
|
keywordstyle=[8]\color{blue}\bfseries |
|
} |
|
|
|
\newcommand{\commentout}[1]{} |
|
\newcommand{\shortversion}[1]{#1} |
|
\newcommand{\longversion}[1]{} |
|
\newcommand{\todo}{\textbf{To do: }} |
|
\newcommand{\hy}[1]{ |
|
\smallskip\noindent\fbox{\begin{minipage}{.97\linewidth}{\bf HY:} |
|
{\rm #1}\end{minipage}}} |
|
\newcommand{\sss}[1]{ |
|
\smallskip\noindent\fbox{\begin{minipage}{.97\linewidth}{\bf SS:} |
|
{\rm #1}\end{minipage}}} |
|
\newcommand{\ok}[1]{ |
|
\smallskip\noindent\fbox{\begin{minipage}{.97\linewidth}{\bf OK:} |
|
{\rm #1}\end{minipage}}} |
|
\newcommand{\ch}[1]{ |
|
\smallskip\noindent\fbox{\begin{minipage}{.97\linewidth}{\bf CH:} |
|
{\rm #1}\end{minipage}}} |
|
|
|
\renewcommand{\hy}[1]{} |
|
\renewcommand{\sss}[1]{} |
|
\renewcommand{\ok}[1]{} |
|
\renewcommand{\ch}[1]{} |
|
|
|
\newcommand{\XI}{X^\omega} |
|
\newcommand{\Xn}{X^n} |
|
\newcommand{\RR}{\mathbb{R}} |
|
\newcommand{\NN}{\mathbb{N}} |
|
\newcommand{\sigalg}[1]{\Sigma_{#1}} |
|
\newcommand{\qbtosig}[1]{\Sigma_{#1}} |
|
\newcommand{\qb}[1]{M_{#1}} |
|
\newcommand{\sigtoqb}[1]{M_{#1}} |
|
\newcommand{\inv}[1]{{#1}^{\textup{-1}}} |
|
\newcommand{\id}[1]{\mathrm{id}_{#1}} |
|
\newcommand{\iid}{\mathrm{iid}} |
|
\newcommand{\iidn}{\mathrm{iid}_n} |
|
\newcommand{\iotan}{\iota_n} |
|
\newcommand{\inviotan}{\inv{\iota}_n} |
|
\newcommand{\op}[1]{{#1}^{\mathrm{op}}} |
|
\newcommand{\defeq}{\stackrel {\textup{def}}=} |
|
\newcommand{\QBS}{\mathbf{QBS}} |
|
\newcommand{\Meas}{\mathbf{Meas}} |
|
\newcommand{\Set}{\mathbf{Set}} |
|
\newcommand{\SMeas}{\mathbf{SMeas}} |
|
\newcommand{\dd}{\mathrm{d}} |
|
\newcommand{\denot}[1]{\llbracket#1\rrbracket} |
|
\newcommand{\Pmonad}{P} |
|
\newcommand{\Giry}{G} |
|
\newcommand{\curry}{\mathsf{curry}} |
|
\newcommand{\uncurry}{\mathsf{uncurry}} |
|
\newcommand{\bindsymbol}{\scalebox{0.5}[1]{$>\!>=$}} |
|
\newcommand{\bind}[2]{#1\mathrel{\bindsymbol} #2} |
|
\newcommand{\bindname}{(\bindsymbol)} |
|
\newcommand{\bindGsymbol}{\scalebox{0.5}[1]{$>\!>=$}_G} |
|
\newcommand{\bindG}[2]{#1\mathrel{\bindGsymbol} #2} |
|
\newcommand{\bindGname}{(\bindGsymbol)} |
|
\newcommand{\munit}{\eta} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
\ifCLASSINFOpdf |
|
\else |
|
\fi |
|
|
|
|
|
|
|
|
|
|
|
|
|
\usepackage{amssymb,amsthm,amsmath,amsfonts,mathtools,stmaryrd} |
|
\usepackage{xy} |
|
\xyoption{all} |
|
\newtheorem{definition}{Definition} |
|
\newtheorem{proposition}[definition]{Proposition} |
|
\newtheorem{theorem}[definition]{Theorem} |
|
\newtheorem{lemma}[definition]{Lemma} |
|
\newtheorem{example}[definition]{Example} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
\hyphenation{op-tical net-works semi-conduc-tor} |
|
|
|
|
|
\begin{document} |
|
\title{A Convenient Category for \\ Higher-Order Probability Theory} |
|
|
|
\author{ |
|
\IEEEauthorblockN{Chris Heunen} |
|
\IEEEauthorblockA{University of Edinburgh, UK} |
|
\and |
|
\IEEEauthorblockN{Ohad Kammar} |
|
\IEEEauthorblockA{University of Oxford, UK} |
|
\and |
|
\IEEEauthorblockN{Sam Staton} |
|
\IEEEauthorblockA{University of Oxford, UK} |
|
\and |
|
\IEEEauthorblockN{Hongseok Yang} |
|
\IEEEauthorblockA{University of Oxford, UK} |
|
} |
|
|
|
|
|
|
|
|
|
|
|
\IEEEoverridecommandlockouts |
|
\IEEEpubid{\makebox[\columnwidth]{978-1-5090-3018-7/17/\$31.00~ |
|
\copyright2017 IEEE \hfill} \hspace{\columnsep}\makebox[\columnwidth]{ }} |
|
|
|
|
|
|
|
|
|
|
|
\maketitle |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
\begin{abstract} |
|
Higher-order probabilistic programming languages allow programmers to write sophisticated models in machine learning and statistics in a succinct and structured way, |
|
but step outside the standard measure-theoretic formalization of probability theory. |
|
Programs may use both higher-order functions and continuous distributions, or even define a probability distribution on functions. |
|
But standard probability theory does not handle higher-order functions well: the category of measurable spaces is not cartesian closed. |
|
|
|
Here we introduce quasi-Borel spaces. |
|
We show that these spaces: |
|
form a new formalization of probability theory replacing measurable spaces; |
|
form a cartesian closed category and so support higher-order functions; |
|
form a well-pointed category and so support good proof principles for equational reasoning; |
|
and support continuous probability distributions. |
|
We demonstrate the use of quasi-Borel spaces for higher-order functions and probability by: showing that a well-known construction of probability theory involving random functions gains a cleaner expression; and generalizing de Finetti's theorem, that is a crucial theorem in probability theory, to quasi-Borel spaces. |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
\end{abstract} |
|
|
|
\IEEEpeerreviewmaketitle |
|
\allowdisplaybreaks |
|
|
|
|
|
\section{Introduction}\label{sec:intro} |
|
To express probabilistic models in machine learning and statistics in a succinct and structured way, it pays to use \emph{higher-order} programming languages, such as Church~\cite{goodman_uai_2008}, Venture~\cite{Mansinghka-venture14}, or Anglican~\cite{wood-aistats-2014}. These languages support advanced features from both |
|
programming language theory and probability theory, while |
|
providing generic inference algorithms for answering probabilistic queries, such as marginalization and posterior computation, for all models written in the language. As a result, the programmer can succinctly express a sophisticated probabilistic model and explore its properties while avoiding the nontrivial busywork of designing a custom inference algorithm. |
|
|
|
|
|
This exciting development comes at a foundational price. |
|
Programs in these languages may combine higher-order functions and continuous |
|
distributions, or even define a probability distribution on functions. |
|
But the standard measure-theoretic formalization of probability theory does not handle higher-order functions well, as the category of measurable spaces is not cartesian closed~\cite{aumann:functionspaces}. |
|
For instance, the Anglican implementation of Bayesian linear regression in Figure~\ref{fig:linearregression} goes beyond the standard measure-theoretic foundation of probability theory, as it defines a probability distribution on functions $\RR \to \RR$. |
|
|
|
\makeatletter |
|
\lst@Key{countblanklines}{true}[t]{\lstKV@SetIf{#1}\lst@ifcountblanklines} |
|
|
|
\lst@AddToHook{OnEmptyLine}{\lst@ifnumberblanklines\else \lst@ifcountblanklines\else \advance\c@lstnumber-\@ne\relax \fi \fi} |
|
\makeatother |
|
\lstset{frame=single, xleftmargin=8.25mm, framexleftmargin=7mm, numbers=left, numberblanklines=false, linewidth=.48\textwidth |
|
} |
|
\begin{figure} |
|
\begin{lstlisting}[style=default,countblanklines=false, basicstyle=\ttfamily\small,escapechar=\|] |
|
(defquery Bayesian-linear-regression |\vskip-.6\baselineskip| |
|
|
|
(let [f (let [s (sample (normal 0.0 3.0)) |
|
b (sample (normal 0.0 3.0))] |
|
(fn [x] (+ (* s x) b)))] |\vskip-.6\baselineskip| |
|
|
|
(observe (normal (f 1.0) 0.5) 2.5) |
|
(observe (normal (f 2.0) 0.5) 3.8) |
|
(observe (normal (f 3.0) 0.5) 4.5) |
|
(observe (normal (f 4.0) 0.5) 6.2) |
|
(observe (normal (f 5.0) 0.5) 8.0) |\vskip-.6\baselineskip| |
|
|
|
(predict :f f))) |
|
\end{lstlisting} |
|
\fbox{\includegraphics[width=.97\linewidth]{posterior.png}} |
|
\caption{Bayesian linear regression in Anglican. The program defines a probability |
|
distribution on functions $\RR \to \RR$. It first samples |
|
a random linear function \texttt{f} by randomly selecting |
|
slope \texttt{s} and intercept \texttt{b}. It then adjusts the probability distribution of the function |
|
to better describe five observations $(1.0,2.5)$, $(2.0,3.8)$, $(3.0,4.5)$, $(4.0,6.2)$ |
|
and $(5.0,8.0)$ by posterior computation. |
|
In the graph, each line has been sampled from the posterior distribution over linear functions.} |
|
\label{fig:linearregression} |
|
\end{figure} |
|
|
|
We introduce a new formalization of probability theory that accommodates higher-order functions. The main notion replacing a measurable space is a \emph{quasi-Borel space}: a set~$X$ equipped with a collection of functions $\qb X \subseteq {[\RR \to X]}$ satisfying |
|
certain conditions (Def.~\ref{def:qbs}). Intuitively, $\qb X$ is the set of random variables of type $X$. |
|
Here $\RR$ means that the randomness of random variables in $\qb X$ |
|
comes from (a probability distribution on) $\RR$, one of the best behaving measurable spaces. |
|
Thus the primitive notion shifts from measurable subset to random variable, which is traditionally a derived notion. |
|
For related ideas see \S\ref{sec:related}. |
|
|
|
Quasi-Borel spaces have good properties and structure. |
|
\begin{itemize} |
|
\item The category of quasi-Borel spaces is \emph{well-pointed}, since a morphism is just a structure-preserving function (\S\ref{sec:quasiborel}). (This is in contrast to~\cite[\S8]{statonyangheunenkammarwood:higherorder}). |
|
|
|
\item The category of quasi-Borel spaces is cartesian closed (\S\ref{sec:structure}), so that it becomes a setting to study probability distributions on \emph{higher-order} functions. |
|
|
|
\item There is a natural notion of probability measure on quasi-Borel spaces (Def.~\ref{def:probabilitymeasure}). |
|
The space of all probability measures is again a |
|
quasi-Borel space, and forms the basis for a commutative \emph{monad} on the category of quasi-Borel spaces (\S\ref{sec:giry}). Thus quasi-Borel spaces form semantics for a probabilistic programming language in the monadic style~\cite{moggi-monads}. |
|
\end{itemize} |
|
We also illustrate the use of quasi-Borel spaces. |
|
\begin{itemize} |
|
\item \emph{Bayesian regression} (\S\ref{sec:example}). |
|
Quasi-Borel spaces are a natural setting for understanding programs such as the one in Figure~\ref{fig:linearregression}: |
|
the prior (Lines 2--4) defines a probability distribution over functions~\lstinline|f|, i.e.\ a measure on $\RR^\RR$, and the posterior (illustrated in the graph), is again a probability measure on $\RR^\RR$, conditioned by the observations (Lines 5--9). |
|
\item \emph{Randomization} (\S\ref{sec:functions}). |
|
A key idea of categorical logic is that $\forall\exists$ statements should become statements |
|
about quotients of objects. |
|
The structure of quasi-Borel spaces allows us to rephrase a crucial randomization lemma in this way. |
|
Classically, it says that every probability kernel arises from a random function. |
|
In the setting of quasi-Borel spaces, it says that |
|
the space of probability kernels $P(\RR)^X$ is a quotient of the space of random functions, $P(\RR^X)$ (Theorem~\ref{theorem:random-quotient}). |
|
Notice that the higher-order structure of quasi-Borel spaces allows us to succinctly state this result. |
|
\item \emph{De Finetti's theorem} (\S\ref{sec:definetti}). |
|
Probability theorists often encounter problems when working with arbitrary probability measures on arbitrary measurable spaces. |
|
Quasi-Borel spaces allow us to better manage the source of randomness. |
|
For example, de Finetti's theorem is a foundational result in Bayesian statistics which says that every exchangeable random sequence can be generated |
|
by randomly mixing multiple independent and identically distributed sequences. |
|
The theorem is known to hold for standard Borel spaces~\cite{deFinetti37} |
|
or measurable spaces that arise from good topologies~\cite{HewittSavage55}, but not for arbitrary measurable spaces~\cite{Dubins1979}. We show that it holds for all quasi-Borel spaces (Theorem~\ref{thm:deFinetti-qbs}). |
|
\end{itemize} |
|
All of this is evidence that quasi-Borel spaces form a convenient category for higher-order probability theory. |
|
|
|
|
|
|
|
|
|
|
|
\section{Preliminaries on probability measures and measurable spaces}\label{sec:prelims} |
|
\begin{definition}\label{def:borel} |
|
The \emph{Borel sets} form the least collection $\Sigma_\RR$ of subsets of $\RR$ that |
|
satisfies the following properties: |
|
\begin{itemize} |
|
\item intervals $(a,b)$ are Borel sets; |
|
\item complements of Borel sets are Borel; |
|
\item countable unions of Borel sets are Borel. |
|
\end{itemize} |
|
\end{definition} |
|
|
|
The Borel sets play a crucial role in probability theory because of the tight connection |
|
between the notion of probability measure and the axiomatization of Borel sets. |
|
|
|
\begin{definition}\label{def:borelprob} |
|
A \emph{probability measure} on $\RR$ is a function $\mu\colon \sigalg\RR\to [0,1]$ satisfying $\mu(\RR)=1$ and $\mu(\biguplus S_i)=\sum \mu(S_i)$ for any countable sequence of disjoint Borel sets $S_i$. |
|
\end{definition} |
|
|
|
The natural generalization gives measurable spaces. |
|
|
|
\begin{definition}\label{def:meas-space} |
|
A \emph{$\sigma$-algebra} on a set $X$ is a nonempty family of subsets of $X$ that is closed under complements and countable unions. A \emph{measurable space} is a pair $(X,\sigalg X)$ of a set $X$ and a $\sigma$-algebra $\sigalg X$ on it. A probability measure on a measurable space $X$ is a function $\mu\colon \sigalg X\to [0,1]$ satisfying $\mu(X)=1$ and $\mu(\biguplus S_i)=\sum \mu(S_i)$ for any countable sequence of disjoint sets $S_i\in\sigalg X$. |
|
\end{definition} |
|
|
|
The Borel sets of the reals form a leading example of a $\sigma$-algebra. |
|
Other important examples are countable sets with their \emph{discrete $\sigma$-algebra}, which contains all subsets. |
|
We can characterize these spaces as standard Borel spaces, but first introduce the appropriate structure-preserving maps. |
|
|
|
\begin{definition}\label{def:measurablefunction} |
|
Let $(X,\sigalg X)$ and $(Y,\sigalg Y)$ be measurable spaces. |
|
A \emph{measurable function} $f\colon X\to Y$ is a function such that $\inv{f}(U)\in\sigalg X$ when $U\in\sigalg Y$. |
|
\end{definition} |
|
|
|
Thus a measurable function $f\colon X\to Y$ lets us \emph{push-forward} a probability measure $\mu$ on $X$ to a probability measure $f_*\mu$ on $Y$ by $(f_*\mu)(U)=\mu(\inv{f}(U))$. Measurable spaces and measurable functions form a category $\Meas$. |
|
|
|
Real-valued measurable functions $f \colon X \to \RR$ can be integrated with respect to a probability measure $\mu$ on $(X,\sigalg X)$. The \emph{integral} of a nonnegative function $f$ is |
|
\[ |
|
\int_X f \,\dd\mu \defeq \sup_{\{U_i\}} \sum_i \left(\mu(U_i) \cdot \inf_{x \in U_i} f(x)\right)\text, |
|
\] |
|
where $\{U_i\}$ ranges over finite partitions of $X$ into measurable subsets. |
|
When $f$ may be negative, its integral is\[ |
|
\int_X f\,\dd\mu \defeq \left(\int_X \max(0,f)\,\dd\mu\right) - \left(\int_X \max(0,-f)\,\dd\mu\right) |
|
\] |
|
when those two integrals exist. |
|
When it is convenient to make the integrated variable explicit, we write $\int_{x \in U} f(x)\,\dd\mu$ for $\int_X (\lambda x.\,f(x) \cdot [x \in U])\,\dd\mu$, where $U \in \sigalg X$ is a measurable subset and $[\varphi]$ has the value $1$ if $\varphi$ holds |
|
and $0$ otherwise. |
|
|
|
\subsection{Standard Borel spaces} |
|
\begin{proposition}[e.g.~\cite{kallenberg}, App.~A1]\label{prop:char-sbs} |
|
For a measurable space $(X,\sigalg X)$ the following are equivalent: |
|
\begin{itemize} |
|
\item $(X,\sigalg X)$ is a retract of $(\RR,\sigalg \RR)$, that is, there exist measurable \smash{$X\xrightarrow f \RR\xrightarrow g X$} such that $g \circ f=\id X$; |
|
\item $(X,\sigalg X)$ is either measurably isomorphic to $(\RR,\sigalg \RR)$ or countable and discrete; |
|
\item $X$ has a complete metric with a countable dense subset and $\sigalg X$ is the least $\sigma$-algebra containing all open sets. |
|
\end{itemize} |
|
\end{proposition} |
|
When $(X,\sigalg X)$ satisfies any of the above conditions, we call it |
|
\emph{standard Borel space}. These spaces play an important role in |
|
probability theory because they enjoy properties that do not hold for |
|
general measurable spaces, such as the existence of conditional |
|
probability kernels~\cite{kallenberg,Preston-Borel08} |
|
and de Finetti's theorem for exchangeable random processes~\cite{Dubins1979}. |
|
|
|
Besides $\RR$, another popular uncountable standard Borel space is $(0,1)$ with |
|
the $\sigma$-algebra ${\{U \cap (0,1) ~|~ U \in \sigalg \RR\}}$. As |
|
the above proposition indicates, these spaces are isomorphic by, for instance, |
|
$\lambda r.\,\frac1{(1+e^{-r})} : \RR \to (0,1)$. |
|
|
|
\subsection{Failure of cartesian closure} |
|
\begin{proposition}[Aumann, \cite{aumann:functionspaces}]\label{prop:not-ccc} |
|
The category $\Meas$ is not cartesian closed: there is no space of functions $\RR\to\RR$. |
|
\end{proposition} |
|
Specifically, the evaluation function |
|
\[\varepsilon\colon \Meas(\RR,\RR)\times \RR\to \RR |
|
\qquad\text{with}\qquad |
|
\varepsilon(f,r)=f(r)\] |
|
is never measurable |
|
$(\Meas(\RR,\RR)\times \RR,\Sigma\otimes \sigalg \RR)\to (\RR,\sigalg \RR)$ |
|
regardless of the choice of $\sigma$-algebra $\Sigma$ |
|
on $\Meas(\RR,\RR)$. |
|
Here, $\Sigma\otimes \sigalg \RR$ is the product $\sigma$-algebra, |
|
generated by rectangles ${(U\times V)}$ for $U\in\Sigma$ and $V\in\sigalg \RR$. |
|
|
|
|
|
|
|
|
|
\section{Quasi-Borel spaces}\label{sec:quasiborel} |
|
The typical situation in probability theory is that there is a fixed measurable space $(\Omega,\sigalg\Omega)$, called the \emph{sample space}, from which all randomness originates, and that observations are made in terms of random variables, which are pairs $(X,f)$ of a measurable space of observations $(X,\sigalg X)$ and a measurable function $f\colon \Omega\to X$. |
|
From this perspective, the notion of measurable function is more important |
|
than the notion of measurable space. In some ways, the |
|
$\sigma$-algebra $\sigalg X$ is only used as an intermediary to restrain the class of measurable functions $\Omega\to X$. |
|
|
|
We now use this idea as a basis for our new notion of space. |
|
In doing so, we assume that our sample space $\Omega$ is the real numbers, |
|
which makes probabilities behave well. |
|
\begin{definition}\label{def:qbs} |
|
A \emph{quasi-Borel space} |
|
is a set $X$ together with a set $\qb X \subseteq [\RR\to X]$ satisfying: |
|
\begin{itemize} |
|
\item $\alpha \circ f \in \qb X$ if $\alpha \in \qb X$ and $f\colon \RR\to\RR$ is measurable; |
|
\item $\alpha \in \qb X$ if $\alpha\colon \RR\to X$ is constant; |
|
\item if ${\RR=\biguplus_{i\in \NN}S_i}$, with each set $S_i$ Borel, |
|
and ${\alpha_1,\alpha_2,\ldots\in \qb X}$, then |
|
${\beta}$ is in ${\qb X}$, where ${\beta(r) = \alpha_i(r)}$ for $r\in S_i$. |
|
\end{itemize} |
|
\end{definition} |
|
The name `quasi-Borel space' is motivated firstly by analogy to quasi-topological spaces (see \S\ref{sec:related}), |
|
and secondly in recognition of the intimate connection to the standard Borel space $\RR$ (see also Prop.~\ref{prop:adjunction}(2)). |
|
|
|
\begin{example}\normalfont |
|
For every measurable space $(X,\sigalg X)$, let $\sigtoqb {\sigalg X}$ be the set of measurable functions $\RR\to X$. |
|
Thus $\sigtoqb {\sigalg X}$ is the set of $X$-valued random variables. |
|
In particular: $\RR$ itself can be considered as a quasi-Borel space, with $\qb \RR$ the set of measurable functions $\RR\to\RR$; the two-element discrete space $2$ can be considered as a quasi-Borel space, with $\qb 2$ the set of measurable functions $\RR\to 2$, which are exactly the characteristic functions of the Borel sets~(Def.~\ref{def:borel}). |
|
\end{example} |
|
|
|
Before we continue, we remark that the notion of quasi-Borel space |
|
is invariant under replacing $\RR$ with a different uncountable standard Borel space. |
|
\begin{proposition} |
|
For any measurable space $(\Omega,\sigalg\Omega)$, any measurable isomorphism $\iota \colon \RR \to \Omega$, |
|
any set $X$, and any set $N$ of functions $\Omega\to X$, the pair |
|
$(X,\{\alpha \circ \iota~|~\alpha\in N\})$ |
|
is a quasi-Borel space if and only if: |
|
\begin{itemize} |
|
\item $\alpha \circ f \in N$ if $\alpha \in N$ and $f\colon \Omega\to\Omega$ is measurable; |
|
\item $\alpha \in N$ if $\alpha\colon \Omega\to X$ is constant; |
|
\item if $\Omega=\biguplus_{i\in \NN}S_i$, with each set $S_i \in \sigalg \Omega$, |
|
and $\alpha_1,\alpha_2,\ldots\in N$, then |
|
$\beta$ is in $N$, where $\beta(r)=\alpha_i(r)$ if $r\in S_i$. |
|
\end{itemize} |
|
\end{proposition} |
|
By Prop.~\ref{prop:char-sbs}, the measurable spaces isomorphic to $\RR$ are the uncountable standard Borel spaces. |
|
Note that the choice of isomorphism~$\iota$ is not important: it does not appear in the three conditions. |
|
|
|
Probability theory typically considers a basic probability measure on the sample |
|
space $\Omega$. Each random variable, that is each measurable function |
|
$\Omega\to X$, then induces a probability measure on $X$ by pushing forward the basic measure. Quasi-Borel spaces take this idea as an axiomatic notion of probability measure. |
|
\begin{definition}\label{def:probabilitymeasure} |
|
A \emph{probability measure} on a quasi-Borel space $(X,\qb X)$ is a pair $(\alpha,\mu)$ of $\alpha \in \qb X$ and a probability measure $\mu$ on $\RR$ (as in Def.~\ref{def:borelprob}). |
|
\end{definition} |
|
|
|
\subsection{Morphisms and integration} |
|
|
|
\begin{definition}\label{def:morphism} |
|
A \emph{morphism} of quasi-Borel spaces $(X,\qb X)\to (Y,\qb Y)$ is a function $f \colon X \to Y$ such that $f \circ \alpha \in \qb Y$ if $\alpha\in \qb X$. Write $\QBS\big((X,\qb X),(Y,\qb Y)\big)$ for the set of morphisms from $(X,\qb X)$ to $(Y,\qb Y)$. |
|
\end{definition} |
|
|
|
In particular, elements of $M_X$ are precisely morphisms $(\RR,\qb\RR)\to (X,\qb X)$, so $\qb X = \QBS\big((\RR,\qb \RR),(X,\qb X)\big)$. |
|
|
|
Morphisms compose as functions, and identity functions are morphisms, |
|
so quasi-Borel spaces form a category $\QBS$. |
|
|
|
\begin{example}\normalfont |
|
There are two canonical ways to equip a set $X$ with a quasi-Borel |
|
space structure. The first structure $\qb X^R$ consists of all |
|
functions $\RR \to X$. The second structure $\qb X^L$ consists of |
|
all functions $\beta \colon \RR \to X$ for which there exist: a countable |
|
subset $I \subseteq \NN$; a measurable $f \colon \RR \to \RR$; a |
|
partition $\RR = \biguplus_{i \in I}S_i$ with every $S_i$ |
|
measurable; and a sequence $(x_i)_{i \in I}$ in $X$, such that |
|
$\beta(r) = x_i$ whenever $f(r) \in |
|
S_i$. These are the right and left adjoints, respectively, to the |
|
forgetful functor from $\QBS$ to $\Set$. |
|
\end{example} |
|
|
|
Def.~\ref{def:morphism} is independent of $\RR$: the sample space may be any uncountable standard Borel space. |
|
|
|
\begin{proposition} |
|
Consider a measurable space $(\Omega,\sigalg\Omega)$ |
|
with a measurable isomorphism $\iota \colon \RR \to \Omega$. |
|
For $i \in {1,2}$, let $X_i$ be a set and $N_i$ a set of functions $\Omega\to X_i$ such that |
|
\[ |
|
M_i = \big(X_i,\{\alpha \circ \iota~|~\alpha\in N_i\} \big) |
|
\] |
|
are quasi-Borel spaces. A function $g \colon X_1 \to X_2$ is a morphism $(X_1,M_1) \to (X_2,M_2)$ if and only if $g \circ \alpha \in N_2$ for $\alpha \in N_1$. |
|
\end{proposition} |
|
|
|
Morphisms between quasi-Borel spaces are analogous to measurable functions |
|
between measurable spaces. The crucial properties of measurable functions |
|
are that they work well with (probability) measures: we can push-forward these measures, and integrate over them. |
|
Morphisms of quasi-Borel spaces also support these constructions. |
|
\begin{itemize} |
|
\item\emph{Pushing forward:} if $f\colon X\to Y$ is a morphism and $(\alpha,\mu)$ is a probability measure on $X$ then $f\circ \alpha$ is by definition in $\qb Y$ and so $(f\circ\alpha,\mu)$ is a probability measure on $Y$. |
|
\item\emph{Integrating:} If $f\colon X\to \RR$ is a morphism of quasi-Borel spaces and $(\alpha,\mu)$ is a probability measure on $X$, the integral of $f$ with respect to $(\alpha,\mu)$ is |
|
\begin{equation}\label{eqn:def-qbs-integration} |
|
\int f\, \dd(\alpha,\mu)\defeq \int_\RR (f\circ \alpha)\,\dd\mu\text. |
|
\end{equation} |
|
So integration formally reduces to integration on $\RR$. |
|
\end{itemize} |
|
|
|
\subsection{Relationship to measurable spaces} |
|
|
|
If we regard a subset $S\subseteq X$ as its characteristic function $\chi_S\colon X\to 2$, then we can regard a $\sigma$-algebra on a set $X$ as a set of characteristic functions $F_X\subseteq [X\to 2]$ satisfying certain conditions. Thus a measurable space (Def.~\ref{def:meas-space}) could equivalently be described |
|
as a pair $(X,F_X)$ of a set $X$ and a collection $F_X\subseteq [X\to 2]$ of characteristic functions. |
|
Moreover, from this perspective, a measurable function $f\colon (X,F_X)\to (Y,F_Y)$ is |
|
simply a function $f\colon X\to Y$ such that $\chi\circ f \in F_X$ if $\chi\in F_Y$. |
|
Thus quasi-Borel spaces shift the emphasis from characteristic functions $X\to 2$ to random variables $\RR\to X$. |
|
|
|
\subsubsection{Quasi-Borel spaces as structured measurable spaces} |
|
A subset $S\subseteq X$ is in the $\sigma$-algebra $\sigalg X$ |
|
of a measurable space $(X,\sigalg X)$ if and only if its |
|
characteristic function $X\to 2$ is measurable. |
|
With this in mind, we \emph{define} a measurable subset of a |
|
quasi-Borel space $(X,\qb X)$ to be a subset $S\subseteq X$ |
|
such that the characteristic function $X\to 2$ is a morphism of quasi-Borel spaces. |
|
\begin{proposition} |
|
The collection of all measurable subsets of a quasi-Borel space $(X,\qb X)$ is characterized as |
|
\begin{equation}\label{eqn:qb-sigma} |
|
\qbtosig {\qb X}\defeq \{U~|~\forall \alpha\in\qb X.\,\inv\alpha(U)\in\sigalg \RR\} |
|
\end{equation} |
|
and forms a $\sigma$-algebra. |
|
\end{proposition} |
|
Thus we can understand a quasi-Borel space as a measurable space $(X,\sigalg X)$ equipped |
|
with a class of measurable functions $\qb X\subseteq [\RR\to X]$ determining the $\sigma$-algebra by $\sigalg X=\qbtosig{\qb X}$ |
|
as in~\eqref{eqn:qb-sigma}. |
|
|
|
Moreover, every morphism $(X,\qb X)\to (Y,\qb Y)$ is also a measurable |
|
function $(X,\qbtosig {\qb X})\to (Y,\qbtosig{\qb Y})$ (but the converse does not hold |
|
in general). |
|
|
|
A probability measure $(\alpha,\mu)$ on a quasi-Borel space $(X,\qb X)$ induces a |
|
probability measure $\alpha_*\mu$ on the underlying measurable space. |
|
Integration as in~\eqref{eqn:def-qbs-integration} matches the standard |
|
definition for measurable spaces. |
|
|
|
\subsubsection{An adjunction embedding standard Borel spaces} |
|
Under some circumstances morphisms of quasi-Borel spaces coincide with measurable functions. |
|
\begin{proposition}\label{prop:adjunction} |
|
Let $(Y,\sigalg Y)$ be a measurable space. |
|
\begin{enumerate} |
|
\item If $(X,\qb X)$ is a quasi-Borel space, a function $X\to Y$ is a measurable function |
|
$(X,\qbtosig{\qb X})\to (Y,\sigalg Y)$ if and only if it is a morphism $(X,\qb X)\to (Y,\sigtoqb {\sigalg Y})$. |
|
\item If $(X,\sigalg X)$ is a standard Borel space, a function $X\to Y$ is a morphism $(X,\sigtoqb{\sigalg X})\to (Y,\sigtoqb {\sigalg Y})$ if and only if it is a measurable function $(X,\sigalg X)\to (Y,\sigalg Y)$. |
|
\end{enumerate} |
|
\end{proposition} |
|
\newcommand{\meastoqbs}{R} |
|
\newcommand{\qbstomeas}{L} |
|
Proposition~\ref{prop:adjunction}(1) means there is an adjunction |
|
\[\xymatrix{\Meas\ar@/_/[rr]_\meastoqbs&&\ar@/_/[ll]_\qbstomeas^\bot \QBS}\] |
|
where |
|
$\qbstomeas(X,\qb X)= (X,\qbtosig{\qb X})$ |
|
and |
|
$\meastoqbs(X,\sigalg X)= (X,\sigtoqb{\sigalg X})$. |
|
Proposition~\ref{prop:adjunction}(2) means that the functor $\meastoqbs$ is full and faithful when restricted to standard Borel spaces. Equivalently, |
|
$L(R(X,\sigalg X))=(X,\sigalg X)$, that is $\sigalg X = \qbtosig {\sigtoqb {\sigalg X}}$ |
|
for standard Borel spaces $(X,\sigalg X)$. |
|
|
|
|
|
|
|
\section{Products, coproducts and function spaces}\label{sec:structure} |
|
Quasi-Borel spaces support products, coproducts, and function spaces. |
|
These basic constructions form the basis for interpreting simple type theory |
|
in quasi-Borel spaces. |
|
\begin{proposition}[Products]\label{prop:qbsproduct} |
|
If $(X_i,\qb {X_i})_{i\in I}$ is a family of quasi-Borel spaces indexed by a set $I$, then $(\prod_iX_i,\qb {\Pi_i X_i})$ is a quasi-Borel space, where $\prod_iX_i$ is the set product, and |
|
\[ |
|
\qb {\Pi_iX_i}\defeq \textstyle{\Big\{f\colon \RR\to \prod_iX_i~|~\forall i.\, (\pi_i\circ f)\in\qb{X_i}\Big\}\text.} |
|
\] |
|
The projections $\prod_i X_i\to X_i$ are morphisms, and provide the structure of a categorical product in $\QBS$. |
|
\end{proposition} |
|
|
|
\begin{proposition}[Coproducts]\label{prop:qbscoproduct} |
|
If $(X_i,\qb{X_i})_{i\in I}$ is a family of quasi-Borel spaces indexed by a countable set $I$, then $(\coprod_iX_i,\qb{\amalg_i X_i})$ is a quasi-Borel space, where $\coprod_iX_i$ is the disjoint union of sets, |
|
\begin{align*} |
|
\qb{\amalg_i X_i}\defeq \{\lambda r.\,(f(r),\alpha_{f(r)}(r)) |
|
\mid\; &f\colon \RR\to I\ \mbox{is measurable},\\ |
|
& (\alpha_{i}\in \qb {X_{i}})_{i\in \mathsf{image}(f)}\}\text, |
|
\end{align*} |
|
and $I$ carries the discrete $\sigma$-algebra. |
|
This space has the universal property of a coproduct in the category $\QBS$. |
|
\end{proposition} |
|
\begin{proof}[Proof notes] |
|
The third condition of quasi-Borel spaces is needed here. |
|
It is a crucial step in showing that for an $I$-indexed family of morphisms $(f_i\colon X_i\to Z)_{i\in I}$, the copairing $[f_i]_{i\in I}\colon\coprod_{i\in I}X_i\to Z$ is again a morphism. |
|
\end{proof} |
|
|
|
\begin{proposition}[Function spaces]\label{prop:functionspace} |
|
If $(X,\qb X)$ and $(Y,\qb Y)$ are quasi-Borel spaces, so is $(Y^X,\qb {Y^X})$, |
|
where ${Y^X\defeq \QBS(X,Y)}$ is the set of morphisms $X\to Y$, and |
|
\[ |
|
\qb{Y^X}\defeq \{\alpha\colon\RR\to Y^X \mid |
|
\mathsf{uncurry}(\alpha)\in\QBS(\RR\times X,Y)\}\text. |
|
\] |
|
The evaluation function $Y^X\times X\to Y$ is a morphism and has the universal property of the function space. |
|
Thus $\QBS$ is a cartesian closed category. |
|
\end{proposition} |
|
\begin{proof}[Proof notes] |
|
The only difficult part is showing that $(Y^X,\qb{Y^X})$ satisfies the third condition of quasi-Borel spaces. |
|
Prop.~\ref{prop:qbscoproduct} is useful here. |
|
\end{proof} |
|
\subsection{Relationship with standard Borel spaces} |
|
Recall that standard Borel spaces can be thought of as a full subcategory |
|
of the quasi-Borel spaces, that is, |
|
the functor $\meastoqbs\colon \Meas\to \QBS$ is full and faithful (Prop.~\ref{prop:adjunction}(2)) |
|
when restricted to the standard Borel spaces. |
|
This full subcategory has the same countable products, coproducts and function spaces |
|
(whenever they exist). |
|
We may thus regard quasi-Borel spaces as a conservative extension of standard Borel spaces that supports simple type theory. |
|
|
|
\begin{proposition} |
|
The functor $\meastoqbs(X,\sigalg X)=(X,\sigtoqb {\sigalg X})$: |
|
\begin{enumerate} |
|
\item preserves products of standard Borel spaces: |
|
$\meastoqbs(\prod_iX_i)=\prod_i\meastoqbs(X_i)$, where $(X_i,\sigalg{X_i})_{i\in I}$ is a countable family of standard Borel spaces; |
|
\item preserves spaces of functions between standard Borel spaces whenever they exist: |
|
if $(Y,\sigalg Y)$ is countable and discrete, |
|
and $(X,\sigalg X)$ is standard Borel, then $\meastoqbs(X^Y)=\meastoqbs(X)^{\meastoqbs(Y)}$; |
|
\item preserves countable coproducts of standard Borel spaces: |
|
$\meastoqbs(\coprod_iX_i)=\coprod_i\meastoqbs(X_i)$, where $(X_i,\sigalg{X_i})_{i\in I}$ is a countable family of standard Borel spaces. |
|
\end{enumerate} |
|
\end{proposition} |
|
Consequently, a standard programming language semantics in standard Borel spaces can be conservatively embedded in quasi-Borel spaces, allowing higher-order functions while preserving all the type theoretic structure. |
|
|
|
We note, however, that in light of Prop.~\ref{prop:not-ccc}, |
|
the quasi-Borel space $\RR^\RR$ does not come from a standard Borel space. |
|
Moreover, the left adjoint $L\colon \QBS\to \Meas$ does not preserve products |
|
in general. |
|
For quasi-Borel spaces $(X,\qb X)$ and $(Y,\qb Y)$, we always have |
|
$\qbtosig{\qb X}\otimes \qbtosig{\qb Y}\subseteq \qbtosig {\qb {X\times Y}}$, |
|
but not always $\supseteq$. Indeed, $\qbtosig{\qb {\RR^\RR}}\otimes \sigalg \RR |
|
\neq \qbtosig {\qb{(\RR^\RR\times \RR)}}$, by Prop.~\ref{prop:not-ccc}. |
|
|
|
\section{A monad of probability measures}\label{sec:giry} |
|
In this section we will show that the probability measures on a quasi-Borel space form a quasi-Borel space again. This gives a commutative monad that generalizes the Giry monad for measurable spaces~\cite{giry:monad}. |
|
\subsection{Monads} |
|
\newcommand{\CCat}{\mathcal C} |
|
We use the Kleisli triple formulation of monads (see e.g.~\cite{moggi-monads}). Recall that a monad on a category $\CCat$ comprises |
|
\begin{itemize} |
|
\item for any object $X$, an object $T(X)$; |
|
\item for any object $X$, a morphism $\munit\colon X\to T(X)$; |
|
\item for any objects $X,Y$, a function |
|
\[\bindname\colon \CCat(X,T(Y))\to \CCat(T(X),T(Y))\text.\] |
|
We write $(t\bindsymbol f)$ for $\bindname(f)(t)$. |
|
\end{itemize} |
|
This is subject to the conditions $( t\bindsymbol \munit)=t$, $( {\munit(x)}\bindsymbol f)=f(x)$, and |
|
$\bind t {(\lambda x.\,(\bind {f(x)} g))}= |
|
\bind{(\bind t f)}g$. |
|
|
|
The intuition is that $T(X)$ is an object of computations returning~$X$, |
|
that $\munit$ is the computation that returns immediately, and that |
|
$t\bindsymbol f$ sequences computations, |
|
first running computation $t$ and then calling $f$ with the result. |
|
|
|
When $\CCat$ is cartesian closed, |
|
a monad is \emph{strong} if $\bindname$ internalizes to an operation |
|
$\bindname\colon (T(Y))^X\to (T(Y))^{T(X)}$, |
|
and then the conditions are understood as expressions in a cartesian closed category. |
|
|
|
\subsection{Kernels and the Giry monad} |
|
We recall the notion of probability kernel, which is a measurable family of probability measures. |
|
\begin{definition} |
|
Let $(X,\sigalg X)$ and $(Y,\sigalg Y)$ be measurable spaces. |
|
A \emph{probability kernel} from $X$ to $Y$ is a function $k:X\times \sigalg Y\to [0,1]$ such that $k(x,-)$ is a probability measure for all $x\in X$ (Def.~\ref{def:meas-space}), and $k(-,U)$ is a measurable function for all $U\in\sigalg Y$ (Def.~\ref{def:measurablefunction}). |
|
\end{definition} |
|
We can classify probability kernels as follows. |
|
Let $\Giry(X)$ be the set of probability measures on $(X,\sigalg X)$. |
|
We can equip this set with the $\sigma$-algebra generated by |
|
${\{\mu\in\Giry(X)~|~\mu(U)<r\}}$, for $U\in\sigalg X$ and $r\in[0,1]$, |
|
to form a measurable space $(\Giry(X),\sigalg{\Giry(X)})$. |
|
A measurable function $X\to \Giry(Y)$ amounts to a probability kernel from $X$ to $Y$. |
|
|
|
The construction $\Giry$ has the structure of a monad, as first discussed by |
|
Giry~\cite{giry:monad}. |
|
A computational intuition is that $\Giry(X)$ is a space of probabilistic computations |
|
over $X$, and this provides a semantic foundation for a first-order probabilistic programming language (see e.g.~\cite{statonyangheunenkammarwood:higherorder}). |
|
The unit $\eta\colon X\to \Giry(X)$ lets $\eta(x)$ be the Dirac measure on $x$, |
|
with $\eta(x)(U)= 1$ if $x\in U$, and $\eta(x)(U)= 0$ if $x\not\in U$. |
|
If $\mu\in \Giry(X)$ and $k$ is a measurable function $X\to \Giry(Y)$, then $(\mu\bindGsymbol k)$ is the measure in $\Giry(Y)$ with |
|
$(\mu\bindGsymbol k)(U)= \int_{x \in X} k(x)(U) \,\dd\mu$. |
|
|
|
\subsection{Equivalent measures on quasi-Borel spaces} |
|
Recall (Def.~\ref{def:probabilitymeasure}) that a probability measure $(\alpha,\mu)$ on a quasi-Borel space $(X,\qb X)$ is a pair $(\alpha,\mu)$ of a function |
|
$\alpha\in\qb X$ and a probability measure $\mu$ on $\RR$. |
|
Random variables are often equated when they describe the same distribution. |
|
Every probability measure $(\alpha,\mu)$ determines |
|
a push-forward measure $\alpha_* \mu$ on the corresponding |
|
measurable space $(X,\qbtosig{\qb X})$, that assigns to $U \subseteq X$ the real number $\mu(\alpha^{-1}(U))$. |
|
We will identify two probability measures when they define the same push-forward measure, and write $\sim$ for this equivalence relation. |
|
|
|
This is a reasonable notion of equality even if we put aside the notion of |
|
measurable space, because two probability measures have the same push-forward measure precisely when they have the same integration operator: $(\alpha,\mu) \sim (\alpha',\mu')$ if and only if $\int f\, \dd(\alpha,\mu) = \int f\,\dd(\alpha',\mu')$ for all morphisms ${f \colon (X,\qb X) \to \RR}$. |
|
Nevertheless, other notions of equivalence could be used. |
|
|
|
\subsection{A probability monad} |
|
We now explain how to build a monad of probability measures on the |
|
category of quasi-Borel spaces, modulo this notion of |
|
equivalence. |
|
This monad $\Pmonad$ will inherit properties |
|
from the Giry monad. |
|
Technically, the functor |
|
$L\colon \QBS\to \Meas$ (Prop.~\ref{prop:adjunction}) is a `monad opfunctor' |
|
taking $\Pmonad$ to the |
|
Giry monad $\Giry$, which means that it extends to a |
|
functor from the Kleisli category of~$\Pmonad$ to the Kleisli category of~$\Giry$~\cite{street-monads}. |
|
\paragraph{On objects} |
|
For a quasi-Borel space $(X,\qb X)$, let |
|
\begin{align*} |
|
P(X) &= \{ (\alpha,\mu) \text{ probability measure on } (X,\qb X)\}\slash\sim \text{,}\\ |
|
\qb{P(X)} & = \{ \beta \colon \RR \to P(X) \mid \exists \alpha \in \qb X.\,\exists g \in \Meas(\RR, G(\RR)).\, |
|
\\ |
|
& \phantom{= \{ \beta \colon \RR \to P(X) \mid\;\;} \forall r \in \RR.\, |
|
\beta(r) = [\alpha,g(r)] \} \text{,} |
|
\end{align*} |
|
where $[\alpha,\mu]$ denotes the equivalence class. |
|
Note that \begin{equation}P(X)\cong \{\alpha_*\mu\in \Giry(X,\qbtosig{\qb X})~|~\alpha\in\qb X,\,\mu\in\Giry(\RR)\}\label{eqn:opfunctor-map}\end{equation} |
|
as sets, and $l_X([\alpha,\mu])= \alpha_*\mu$ defines a measurable injection |
|
$l_X\colon L(P(X))\rightarrowtail \Giry (X,\qbtosig{\qb X})$. |
|
|
|
|
|
\paragraph{Monad unit (return)} |
|
Recall that the constant functions $(\lambda r.x)$ are all in $\qb X$. |
|
For any probability measure $\mu$ on $\RR$, the push-forward measure |
|
$(\lambda r.x)_*\,\mu$ on $(X,\qbtosig{\qb X})$ |
|
is the Dirac measure on $x$, with $((\lambda r.x)_*\,\mu)(U)=1$ if $x\in U$ |
|
and $0$ otherwise. |
|
Thus $(\lambda r.x,\mu) \sim (\lambda r.x,\mu')$ for all measures $\mu,\mu'$ on $\RR$. |
|
The unit of $\Pmonad$ at $(X,\qb X)$ is the morphism |
|
$\munit \colon X\to \Pmonad(X)$ given by |
|
\begin{equation}\label{eq:giry:unit} |
|
\munit_{(X,\qb X)}(x) = [ \lambda r.x,\, \mu ] |
|
\end{equation} |
|
for an arbitrary probability measure $\mu$ on $\RR$. |
|
|
|
\paragraph{Bind} |
|
To define ${\bindname\colon \Pmonad(Y)^X\to (\Pmonad(Y))^{\Pmonad(X)}}$, |
|
suppose $f\colon X\to \Pmonad(Y)$ |
|
is a morphism and $[\alpha,\mu]$ in $\Pmonad(X)$. |
|
Since $f$ is a morphism, there is a measurable $g\colon \RR\to\Giry(\RR)$ |
|
and a function $\beta\in \qb Y$ such that |
|
$(f\circ \alpha)(r) = [\beta,g(r)]$. |
|
Set $(\bind{[\alpha,\mu]} f) = [\beta,\bindG \mu g]$, |
|
where $\bindG\mu g$ is the bind of the Giry monad. |
|
This matches the bind of the Giry monad, since |
|
$(\bindG{(\alpha_*\mu)}{(l_Y \circ f)})=\beta_*(\bindG \mu g)$. |
|
\begin{theorem} |
|
\label{thm:Pmonad} |
|
The data $(P,\eta,\bindname)$ above defines a strong monad on the |
|
category $\QBS$ of quasi-Borel spaces. |
|
\end{theorem} |
|
\begin{proof}[Proof notes] |
|
The monad laws can be reduced to the laws for the monad $\Giry$ on $\Meas$~\cite{giry:monad}. The monad on $\QBS$ is strong because |
|
$\bindname\colon \Pmonad(Y)^X\to (\Pmonad(Y))^{\Pmonad(X)}$ is a |
|
morphism, which is shown by expanding the definitions. |
|
\end{proof} |
|
|
|
\begin{proposition}\label{prop:giryproperties} |
|
The monad~$\Pmonad$ satisfies these properties: |
|
\begin{enumerate} |
|
\item For $f\colon (X,\qb X)\to (Y,\qb Y)$, the functorial action |
|
$\Pmonad(f)\colon P(X)\to P(Y)$ is $[\alpha,\mu] \mapsto [f\circ \alpha,\mu]$. |
|
\item It is a commutative monad, i.e.\ the order of |
|
sequencing doesn't matter: |
|
if $p\in \Pmonad(X)$, $q\in\Pmonad(Y)$, and $f$ is a morphism $X\times Y\to \Pmonad(Z)$, then $\bind p {\lambda x.\,\bind q{\lambda y.\,f(x,y)}}$ equals $\bind q {\lambda y.\,\bind p{\lambda x.\,f(x,y)}}$. |
|
\item The faithful functor ${L\colon \QBS\to \Meas}$ with ${L(X,\qb X)=(X,\qbtosig{\qb X})}$ extends to a faithful functor ${\mathsf{Kleisli}(\Pmonad)\to \mathsf{Kleisli}(\Giry)}$, i.e.\ $(L,l)$ is a monad opfunctor~\cite{street-monads}. |
|
\item When $(X,\sigalg X)$ is a standard Borel space, the map |
|
$l_X$ of Eq.~\eqref{eqn:opfunctor-map} is a measurable isomorphism. |
|
\label{prop:monad-morphism} |
|
\end{enumerate} |
|
\end{proposition} |
|
|
|
|
|
\section{Example: Bayesian regression} |
|
\label{sec:example} |
|
We are now in a position to explain the semantics of the Anglican program in Figure~\ref{fig:linearregression}. |
|
The program can be split into three parts: a prior, a likelihood, and a posterior. |
|
Recall that Bayes' law says that the posterior is proportionate to the |
|
product of the prior and the likelihood. |
|
|
|
\paragraph{Prior} |
|
Lines 2--4 define a prior measure on $\RR^\RR$: |
|
\newcommand*{\mlstinline}[1]{\text{\lstinline|#1|}} |
|
\begin{align*} |
|
\mathit{prior}\defeq |
|
\mlstinline{(let [}&\mlstinline{s (sample (normal 0.0 3.0)) }\\[-4pt]&\mlstinline{b (sample (normal 0.0 3.0))]}\\[-4pt]&\hspace{-0.5cm}\mlstinline{(fn [x] (+ (* s x) b)))} |
|
\end{align*} |
|
|
|
To describe this semantically, observe the following. |
|
\begin{proposition} |
|
\label{prop:change-randomsource} |
|
Let $(\Omega,\sigalg \Omega)$ be a standard Borel space, |
|
and $(X,\qb X)$ a quasi-Borel space. |
|
Let $\alpha\colon R(\Omega,\sigalg \Omega)\to X$ be a morphism and $\mu$ a probability measure on $(\Omega,\sigalg \Omega)$. |
|
Any section-retraction pair $(\Omega\xrightarrow\varsigma \RR \xrightarrow \rho\Omega)=\id\Omega$ has a probability measure $[\alpha\circ \rho,\varsigma_* \mu]\in \Pmonad(X)$, that is independent of the choice of $\varsigma$ and $\rho$. |
|
\label{prop:different-prob-space} |
|
\end{proposition} |
|
Write $[\alpha,\mu]$ for the probability measure in this case. |
|
|
|
Now, the program fragment $\mathit{prior}$ describes the distribution $[\alpha,\nu\otimes \nu]$ |
|
in $\Pmonad(\RR^\RR)$ |
|
where $\nu$ is the normal distribution on $\RR$ with mean $0$ and standard deviation $3$, |
|
and where $\alpha\colon \RR\times \RR\to \RR^\RR$ is given by |
|
$\alpha(s,b)\defeq \lambda r.\,s \cdot r+b$. |
|
Informally, |
|
\begin{equation} |
|
\label{eqn:linear-reg-prior} |
|
\denot{\mathit{prior}}=[\alpha,\nu\otimes\nu]\in \Pmonad(\RR^\RR)\text. |
|
\end{equation} |
|
Figure~\ref{fig:prior} illustrates this measure $[\alpha, \nu \otimes \nu]$. |
|
This denotational semantics can be made compositional, by using |
|
the commutative monad structure of $\Pmonad$ and the cartesian closed structure of the category $\QBS$ (following e.g.~\cite{moggi-monads,statonyangheunenkammarwood:higherorder}), |
|
but in this paper we focus on this example rather than spelling out the general case once again. |
|
|
|
\begin{figure} |
|
\includegraphics[width=\linewidth]{prior.png} |
|
\caption{Illustration of 1000 sampled functions from the prior on $\RR^\RR$ for Bayesian linear regression (\ref{eqn:linear-reg-prior}).} |
|
\label{fig:prior} |
|
\end{figure} |
|
|
|
\newcommand{\nordens}{d} |
|
|
|
\paragraph{Likelihood}Lines 5--9 define the likelihood of the observations: |
|
\begin{align*} |
|
\mathit{obs}\ \ \defeq &\phantom{\mbox{} \dots \mbox{} } \mlstinline{(observe (normal (f 1.0) 0.5) 2.5)}\\&\dots |
|
\mlstinline{ |
|
(observe (normal (f 5.0) 0.5) 8.0)} |
|
\end{align*} |
|
This program fragment has a free variable~$f$ of type~$\RR^\RR$. |
|
Let us focus on line~5 for a moment: |
|
\[\mathit{obs}_1\defeq \mlstinline{(observe (normal (f 1.0) 0.5) 2.5)}\] |
|
Given a function $f\colon \RR\to \RR$, |
|
the likelihood of drawing $2.5$ from a normal distribution |
|
with mean $f(1.0)$ and standard deviation $0.5$ is |
|
\[ |
|
\denot{f\colon\RR^\RR\vdash \mathit{obs}_1}=\nordens(f(1.0),2.5)\text, |
|
\] |
|
where $\nordens\colon \RR^2\to[0,\infty)$ is the density of the normal distribution function with standard deviation $0.5$: |
|
\[ |
|
\nordens(\mu,x) |
|
\ =\ |
|
\sqrt{\tfrac 2\pi}\,e^{-2(x-\mu)^2}\text. |
|
\] |
|
Notice that we use a normal distribution |
|
to allow for some noise in the measurement. Informally, we are not recording an observation that $f(1.0)$ is \emph{exactly} $2.5$, |
|
since this would make regression impossible; rather, $f(1.0)$ is roughly $2.5$. |
|
|
|
Overall, lines 5--9 describe a likelihood weight which is |
|
the product of the likelihoods of the five data points, given |
|
$f\colon \RR^\RR$. |
|
\begin{align*} |
|
\denot{f\colon \RR^\RR\vdash\mathit{obs}} |
|
= \nordens(f(1),2.5) |
|
& |
|
{} |
|
\cdot \nordens(f(2),3.8) |
|
\cdot \nordens(f(3),4.5) |
|
\\ |
|
& |
|
{} |
|
\cdot \nordens(f(4),6.2) |
|
\cdot \nordens(f(5),8.0)\text. |
|
\end{align*} |
|
|
|
\paragraph{Posterior} |
|
We follow the recipe for a semantic posterior given in~\cite{statonyangheunenkammarwood:higherorder}. |
|
Putting the prior and likelihood together gives a probability measure |
|
in~$\Pmonad(\RR^\RR\times [0,\infty))$ |
|
which is found by pushing forward the measure |
|
$\denot{\mathit{prior}}\in\Pmonad(\RR)$ along the function |
|
\shortversion{$(\id\,,\,\denot{\mathit{obs}})\colon \RR^\RR\to \RR^\RR\times [0,\infty)$.} |
|
\longversion{\[(\id\,,\,\denot{\mathit{obs}})\colon \RR^\RR\to \RR^\RR\times [0,\infty)\text.\]} |
|
This push-forward measure |
|
\[\Pmonad(\id\,,\,\denot{\mathit{obs}})\,(\denot{\mathit{prior}})\in \Pmonad(\RR^\RR\times [0,\infty))\] |
|
is a measure over pairs $(f,w)$ of functions together with their likelihood weight. |
|
We now find the posterior by multiplying the prior and the likelihood, and dividing by a normalizing constant. |
|
To do this we define a morphism |
|
\newcommand{\normalize}{\mathit{norm}}\begin{align*} |
|
& |
|
\normalize\colon \Pmonad(X\times [0,\infty))\to \Pmonad(X)\uplus\{\mathsf{error}\} |
|
\\ |
|
&\normalize([(\alpha,\beta),\nu])\defeq |
|
\begin{cases} |
|
[\alpha,\nu_\beta/(\nu_\beta(\RR))]&\text{if }{0\neq \nu_\beta(\RR)\neq \infty}\\ |
|
\mathsf{error}&\text{otherwise}\end{cases} |
|
\end{align*} |
|
where $\nu_\beta\colon \sigalg\RR\to [0,\infty] |
|
\defeq \lambda U.\,\int_{r \in U}(\beta(r))\, \dd \nu$. |
|
The idea is that if $\beta\colon\RR\to [0,\infty)$ and $\nu$ is a probability measure on $\RR$ then |
|
$\nu_\beta$ is always a posterior measure on $\RR$, but it is typically not normalized, |
|
i.e.~$\nu_\beta(\RR)\neq 1$. We normalize it by dividing by the normalizing constant, |
|
as long as this division is well-defined. |
|
|
|
Now, the semantics of the entire program in Figure~\ref{fig:linearregression} |
|
is $\normalize(\Pmonad(\id\,,\,\denot{\mathit{obs}})\,(\denot{\mathit{prior}}))$, which is a measure in $\Pmonad(\RR^\RR)$. Calculating this posterior using Anglican's |
|
inference algorithm \texttt{lmh} gives the plot in the lower half of |
|
Figure~\ref{fig:linearregression}. |
|
|
|
\paragraph{Defunctionalized regression and non-linear regression} |
|
Of course, one can do regression without explicitly considering distributions over the space of all measurable functions, |
|
by instead directly calculating posterior distributions for the slope~$s$ and the intercept~$b$. |
|
For example, one could defunctionalize the program in Fig.~\ref{fig:linearregression} in the style of Reynolds~\cite{defunctionalization}. |
|
But defunctionalization is a whole-program transformation. |
|
By structuring the semantics using quasi-Borel spaces, we are able to work compositionally, without mentioning~$s$ and~$b$ explicitly on lines 5--10. |
|
The internal posterior calculations actually happen at the level of standard Borel spaces, and so a defunctionalized version would be in some sense equivalent, |
|
but from the programming perspective it helps to abstract away from this. |
|
The regression program in Fig.~\ref{fig:linearregression} is quickly adapted to fit other kinds of functions, e.g.\ polynomials, or even programs from a small domain-specific language, simply by changing the prior in Lines~2--4. |
|
|
|
|
|
\section{Random functions}\label{sec:functions} |
|
We discuss random variables and random functions, starting from the traditional setting. |
|
Let $(\Omega,\sigalg\Omega)$ be a measurable space with a probability measure. |
|
A \emph{random variable} is a measurable function |
|
$(\Omega,\sigalg\Omega)\to (X,\sigalg X)$. |
|
A \emph{random function} between measurable spaces |
|
$(X,\sigalg X)$ and $(Y,\sigalg Y)$ is a measurable function |
|
$(\Omega \times X,\sigalg \Omega \otimes \sigalg X)\to (Y,\sigalg Y)$. |
|
|
|
We can push forward a probability measure on $\Omega$ along a random variable |
|
$(\Omega,\sigalg\Omega)\to (X,\sigalg X)$ to get a probability measure on |
|
$X$, but in the traditional setting we cannot push forward a measure along a |
|
random function. Measurable spaces are not cartesian closed (Prop.~\ref{prop:not-ccc}), |
|
and so we cannot form a measurable space $Y^X$ and we cannot curry a random function in general. |
|
|
|
Now, if we revisit these definitions in the setting of quasi-Borel spaces, |
|
we \emph{do} have function spaces, and so we can push forward along random functions. |
|
In fact, this is somewhat tautologous because a probability measure (Def.~\ref{def:probabilitymeasure}) on a function space |
|
is essentially the same thing as a random function: a probability measure on a function |
|
space $(Y,\sigalg Y)^{(X,\sigalg X)}$ is defined to be a pair $(f,\mu)$ of a |
|
probability measure $\mu$ on $\RR$, our sample space, and a morphism |
|
$f \colon \RR\to Y^X$; but to give a morphism $\RR\to Y^X$ is to give a |
|
morphism $\RR \times X \to Y$ (Prop.~\ref{prop:functionspace}) |
|
as in the traditional definition of random function. |
|
|
|
We have already encountered an example of a random function in Section~\ref{sec:example}: the prior for linear regression is a random function from $\RR$ to $\RR$ |
|
over the measurable space $(\RR\times \RR, \sigalg \RR \otimes \sigalg \RR)$ with the measure $\nu\otimes \nu$. |
|
Random functions abound throughout probability theory and stochastic processes. |
|
The following section explores their use in the so-called randomization lemma, which is |
|
used throughout probability theory. By moving to quasi-Borel spaces, |
|
we can state this lemma succinctly (Theorem~\ref{theorem:random-quotient}). |
|
|
|
|
|
\subsection{Randomization} |
|
An elementary but useful trick in probability theory is |
|
that every probability distribution on $\RR$ |
|
arises as a push-forward of the uniform distribution on $[0,1]$. |
|
Even more useful is that this can be done in a parameterized way. |
|
\begin{proposition}[\cite{kallenberg}, Lem.~3.22]\label{prop:randomization} |
|
Let $(X,\sigalg X)$ be a measurable space. |
|
For any kernel ${k\colon X\times \sigalg \RR\to [0,1]}$ |
|
there is a measurable function ${f\colon\RR\times X\to \RR}$ such that |
|
${k(x,U)=\upsilon\{r~|~f(r,x)\in U\}}$, where |
|
$\upsilon$ is the uniform distribution on $[0,1]$. |
|
\end{proposition} |
|
For quasi-Borel spaces we can phrase this more succinctly: it is a result |
|
about a quotient of the space of random functions. |
|
We first define quotient spaces. |
|
\begin{proposition} |
|
Let $(X,\qb X)$ be a quasi-Borel space, let $Y$ be a set, |
|
and let $q\colon X\to Y$ be a surjection. |
|
Then $(Y,\qb Y)$ is a quasi-Borel space with $\qb Y=\{q\circ \alpha~|~\alpha\in \qb X\}$. |
|
\end{proposition} |
|
We call such a space a \emph{quotient} space. |
|
\begin{theorem}\label{theorem:random-quotient} |
|
Let $(X,\qb X)$ be a quasi-Borel space. |
|
The space $(\Pmonad(\RR))^X$ of kernels is a quotient of the |
|
space $\Pmonad(\RR^X)$ of random functions. |
|
\end{theorem} |
|
Before proving this theorem, we use Prop.~\ref{prop:randomization} |
|
to give an alternative characterization of our probability monad. |
|
\begin{lemma}\label{lemma:alt-qb-pmonad} |
|
Let $(X,\qb X)$ be a quasi-Borel space. |
|
The function $q\colon X^\RR\to \Pmonad(X)$ given by $q(\alpha)\defeq [\alpha,\upsilon]$ |
|
is a surjection, with corresponding quotient space |
|
$(\Pmonad(X),\qb{\Pmonad(X)})$: |
|
\begin{align} |
|
\label{eqn:altMPmonad} |
|
M_{\Pmonad(X)}&=\{\lambda r\in\RR.\,[\gamma(r),\upsilon]~|~\gamma\in \qb{X^\RR}\}\text, |
|
\end{align} |
|
where $\upsilon$ is the uniform distribution on $[0,1]$. |
|
\end{lemma} |
|
\begin{proof}[Proof notes] |
|
The direction $(\subseteq)$ follows immediately from Prop.~\ref{prop:randomization}. |
|
For the direction $(\supseteq)$ we must consider $\gamma\in \qb{X^\RR}$ and show that |
|
$(\lambda r\in \RR.\,[\gamma(r),\upsilon])$ is in $\qb{\Pmonad(X)}$. |
|
This follows by considering the kernel $k\colon \RR\to \Giry(\RR\times \RR)$ with |
|
$k(r)=\upsilon\otimes\delta_r$, |
|
so that $[\gamma(r),\upsilon]=[\uncurry(\gamma),k(r)]$. Here we are using Prop.~\ref{prop:change-randomsource}. |
|
\end{proof} |
|
\shortversion{ |
|
\begin{proof}[Proof of Theorem~\ref{theorem:random-quotient}] |
|
Consider the evident morphism |
|
$q\colon \Pmonad(\RR^X)\to (\Pmonad(\RR))^X$ |
|
that comes from the monadic strength. |
|
That is, $(q([\alpha,\mu]))(x)=[\lambda r.\,\alpha(r)(x),\mu]$. |
|
We show that $q$ is a quotient morphism. |
|
|
|
We first show that $q$ is surjective. |
|
To give a morphism $k\colon (X,\qb X)\to \Pmonad(\RR)$ |
|
is to give a measurable function $(X,\qbtosig{\qb X})\to\Giry(\RR)$, |
|
since |
|
$(\Pmonad(\RR),\qb{\Pmonad(\RR)})\cong (\Giry(\RR),\sigtoqb{\sigalg{\Giry(\RR)}})$ (Prop.~\ref{prop:giryproperties}(4))and by using the adjunction between measurable spaces |
|
and quasi-Borel spaces (Prop.~\ref{prop:adjunction}(1)). |
|
Directly, we understand a morphism $k\colon (X,\qb X)\to \Pmonad(\RR)$ |
|
as the kernel $k^\sharp\colon X\times \sigalg \RR\to [0,1]$ with |
|
$k^\sharp(x,U)\defeq \mu_x(\inv{\alpha}_x(U))$ whenever |
|
$k(x)=[\alpha_x,\mu_x]$. The definition of $k^\sharp$ does not depend on the choice |
|
of $\alpha_x,\mu_x$. |
|
|
|
Now we can |
|
use the randomization lemma (Prop.~\ref{prop:randomization}) |
|
to find a measurable function $f_{k^\sharp}\colon \RR\times X\to \RR$ such that |
|
$k^\sharp(x,U)=\upsilon\{r \mid f_{k^\sharp}(r,x)\in U\}$. |
|
In general, if a function ${Y\times X\to Z}$ is jointly measurable |
|
then it is also a morphism from the product quasi-Borel space. |
|
So $f_{k^\sharp}$ is a morphism, |
|
and we can form $(\curry{f_{k^\sharp}}) \colon \RR\to \RR^X$. |
|
So, |
|
\begin{align*} |
|
q([\curry{f_{k^\sharp}},\,\upsilon])(x) |
|
& = [\lambda r.\curry{f_{k^\sharp}}(r)(x),\,\upsilon] |
|
\\ |
|
& = [\lambda r. f_{k^\sharp}(r,x),\,\upsilon] |
|
= k(x)\text{,} |
|
\end{align*} |
|
and $q$ is surjective, as required. |
|
|
|
Finally we show that $\qb{(\Pmonad(\RR))^X}=\{q\circ \alpha~|~ |
|
\alpha\in\qb {\Pmonad(\RR^X)}\}$. |
|
We have $(\supseteq)$ since $q$ is a morphism, so it remains to show $(\subseteq)$. |
|
Consider $\beta\in\qb{(\Pmonad(\RR))^X}$. |
|
We must show that $\beta=q\circ\alpha$ for some $\alpha\in \qb{\Pmonad(\RR^X)}$. |
|
By Prop.~\ref{prop:functionspace}, $\beta\in\qb{(\Pmonad(\RR))^X}$ |
|
means the uncurried function |
|
$(\uncurry\,\beta)\colon\RR\times X\to \Pmonad (\RR)$ is a morphism. |
|
As above, this morphism corresponds to a kernel |
|
$(\uncurry\,\beta)^\sharp\colon (\RR\times X)\times \sigalg \RR\to [0,1]$. |
|
The randomization lemma (Prop.~\ref{prop:randomization}) gives a measurable function $f_{\beta}\colon \RR\times (\RR\times X)\to \RR$ such that |
|
$(\uncurry\,\beta)^\sharp((r,x),U)=\upsilon\{s \mid f_{\beta}(s,(r,x))\in U\}$. |
|
By Prop.~\ref{prop:adjunction}(1) and the fact that |
|
the $\sigma$-algebra of a product quasi-Borel space |
|
$\RR \times (\RR \times X)$ includes the product $\sigma$-algebras |
|
$\sigalg \RR \otimes \sigalg {\qb {\RR \times X}}$, |
|
this function $f_\beta$ is also a morphism. |
|
Define $\gamma\colon \RR\to (\RR^X)^\RR$ by |
|
${\gamma = \lambda r.\,\lambda s.\,\lambda x.\,f_\beta(s,(r,x))}$. |
|
This is a morphism since we can interpret $\lambda$-calculus in a cartesian closed |
|
category. |
|
Define $\alpha\colon \RR\to \Pmonad(\RR^X)$ by |
|
$\alpha(r)= [\gamma(r),\upsilon]$; |
|
this function is in $\qb{\Pmonad(\RR^X)}$ by Lemma~\ref{lemma:alt-qb-pmonad}. |
|
A direct calculation now gives $\beta=q\circ\alpha$, as required. |
|
\end{proof} |
|
} |
|
\longversion{ |
|
\begin{proof}[Proof of Theorem~\ref{theorem:random-quotient}] |
|
We consider the evident morphism |
|
$q\colon \Pmonad(\RR^X)\to (\Pmonad(\RR))^X$ |
|
that comes from the monadic strength. |
|
That is, $(q([\alpha,\mu]))(x)\defeq[\lambda r.\,\alpha(r)(x),\mu]$. |
|
We show that $q$ is a quotient morphism. |
|
|
|
We first show that $q$ is a surjection. |
|
Note that to give a morphism $k\colon (X,\qb X)\to \Pmonad(\RR)$ |
|
is to give a measurable function $(X,\qbtosig{\qb X})\to\Giry(\RR)$, |
|
since |
|
$(\Pmonad(\RR),\qb{\Pmonad(\RR)})\cong (\Giry(\RR),\sigtoqb{\sigalg{\Giry(\RR)}})$ (Prop.~\ref{prop:monad-morphism}(4)) |
|
and by using the adjunction between measurable spaces |
|
and quasi-Borel spaces (Prop.~\ref{prop:adjunction}(1)). |
|
Directly, we understand a morphism $k\colon (X,\qb X)\to \Pmonad(\RR)$ |
|
as the kernel $k^\sharp\colon X\times \sigalg \RR\to [0,1]$ with |
|
$k^\sharp(x,U)\defeq \mu_x(\inv{\alpha_x}(U))$ whenever |
|
$k(x)=[\alpha_x,\mu_x]$. The definition of $k^\sharp$ does not depend on the choice |
|
of $\alpha_x,\mu_x$. |
|
|
|
Now we can |
|
use the randomization lemma (Prop.~\ref{prop:randomization}) |
|
to find a measurable function $f_{k^\sharp}\colon \RR\times X\to \RR$ such that |
|
$k^\sharp(x,U)=\upsilon\{r~|~f_{k^\sharp}(r,x)\in U\}$. |
|
In general, if a function ${Y\times X\to Z}$ is jointly measurable |
|
then it is also a morphism from the product quasi-Borel space |
|
(but the converse is unknown). |
|
So $f_{k^\sharp}$ is a morphism, |
|
and we can form $(\curry{f_{k^\sharp}}) \colon \RR\to \RR^X$. |
|
So, |
|
\begin{align*} |
|
q([\curry{f_{k^\sharp}},\,\upsilon])(x) |
|
& = [\lambda r.\curry{f_{k^\sharp}}(r)(x),\,\upsilon] |
|
\\ |
|
& = [\lambda r. f_{k^\sharp}(r,x),\,\upsilon] |
|
= k(x)\text{,} |
|
\end{align*} |
|
and $q$ is surjective, as required. |
|
|
|
Finally we must show that \[\qb{(\Pmonad(\RR))^X}=\{q\circ \alpha~|~ |
|
\alpha\in\qb {\Pmonad(\RR^X)}\}\text.\] |
|
We have $(\supseteq)$ since $q$ is a morphism, so it remains to show $(\subseteq)$. |
|
Consider $\beta\in\qb{(\Pmonad(\RR))^X}$. |
|
We must show that $\beta=q\circ\alpha$ for some $\alpha\in \qb{\Pmonad(\RR^X)}$. |
|
By Prop.~\ref{prop:functionspace}, $\beta\in\qb{(\Pmonad(\RR))^X}$ |
|
means that the uncurried function |
|
$(\uncurry\,\beta)\colon\RR\times X\to \Pmonad (\RR)$ is a morphism. |
|
As above, this morphism corresponds to a kernel |
|
$(\uncurry\,\beta)^\sharp\colon (\RR\times X)\times \sigalg \RR\to [0,1]$. |
|
We use the randomization lemma (Prop.~\ref{prop:randomization}) |
|
to find a measurable function $f_{\beta}\colon \RR\times (\RR\times X)\to \RR$ such that |
|
$(\uncurry\,\beta)^\sharp((r,x),U)=\upsilon\{s~|~f_{\beta}(s,(r,x))\in U\}$. |
|
By Prop.~\ref{prop:adjunction}(1) and the fact that |
|
the $\sigma$-algebra of a product quasi-Borel space |
|
$\RR \times (\RR \times X)$ includes the product $\sigma$-algebras |
|
$\sigalg \RR \otimes \sigalg {\qb {\RR \times X}}$, |
|
this function $f_\beta$ is also a morphism. |
|
Let $\gamma\colon \RR\to (\RR^X)^\RR$ be given by |
|
$\gamma\defeq \lambda r.\,\lambda s.\,\lambda x.\,f_\beta(s,(r,x))$. |
|
This is a morphism since we can interpret $\lambda$-calculus in a cartesian closed |
|
category. |
|
Let $\alpha\colon \RR\to \Pmonad(\RR^X)$ be given by |
|
$\alpha(r)\defeq [\gamma(r),\upsilon]$; |
|
this function is in $\qb{\Pmonad(\RR^X)}$ by Lemma~\ref{lemma:alt-qb-pmonad}. |
|
Moreover a direct calculation gives $\beta=q\circ\alpha$, as required. |
|
\end{proof} |
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
\section{De Finetti's theorem}\label{sec:definetti} |
|
De Finetti's theorem \cite{deFinetti37} is one of the foundational results in Bayesian statistics. |
|
It says that every exchangeable sequence of random observations on $\RR$ |
|
or another well-behaved measurable space can be modeled accurately by the following |
|
two-step process: first choose a probability |
|
measure on $\RR$ randomly (according to some distribution on probability measures) |
|
and then generate a sequence with independent samples from this measure. Limiting observations to values in a well-behaved space like $\RR$ in the theorem is important: Dubins and Freedman |
|
proved that the theorem fails for a general measurable space~\cite{Dubins1979}. |
|
|
|
In this section, we show that a version of de Finetti's theorem holds for all quasi-Borel spaces, not just $\RR$. Our result does not |
|
contradict Dubins and Freedman's obstruction; probability measures on quasi-Borel spaces may only use $\RR$ as their |
|
source of randomness, whereas those on measurable spaces are allowed to use |
|
any measurable space for the same purpose. As we will show shortly, this careful choice |
|
of random source lets us generalize key arguments in a proof |
|
of de Finetti's theorem~\cite{Austin-IISC13} to quasi-Borel spaces. |
|
|
|
Let $(X,\qb X)$ be a quasi-Borel space and $(\Xn, \qb \Xn)$ |
|
the product quasi-Borel space $\prod_{i = 1}^n X$ for each positive integer $n$. |
|
Recall that $\Pmonad(X)$ consists of equivalence classes $[\beta,\nu]$ of probability measures $(\beta,\nu)$ on $X$. |
|
For $n \geq 1$, define a morphism $\iidn \colon \Pmonad(X) \to \Pmonad(\Xn)$ by |
|
\[ |
|
\textstyle{\iidn([\beta,\nu])} |
|
= \textstyle{\left[\left(\prod_{i = 1}^n \beta \circ \iotan\right),\, |
|
\left(\left(\inviotan\right)_* \bigotimes_{i = 1}^n \nu\right)\right]} |
|
\] |
|
where $\iotan$ is a measurable isomorphism $\RR \to \prod_{i =1}^n \RR$, |
|
and $\bigotimes_{i = 1}^n \nu$ is the product measure formed by $n$ copies of $\nu$. |
|
The name $\iidn$ represents `independent and identically distributed'. |
|
Indeed, $\iidn$ transforms a probability measure $(\beta,\nu)$ on $X$ to the measure of the random sequence in $\Xn$ that independently samples from $(\beta,\nu)$. |
|
The function $\iidn$ is a morphism $\Pmonad(X) \to \Pmonad(\Xn)$ because |
|
it can also be written in terms of the strength of the monad $\Pmonad$. |
|
|
|
Write $(\XI,\qb \XI)$ for the countable product $\prod_{i = 1}^\infty X$. |
|
\begin{definition} |
|
A probability measure $(\alpha,\mu)$ on $\XI$ is \emph{exchangeable} |
|
if for all permutations $\pi$ on positive integers, |
|
\shortversion{ |
|
$[\alpha,\mu] = [\alpha_\pi, \mu]$, |
|
where $\alpha_\pi(r)_i \defeq \alpha(r)_{\pi(i)}$ for all $r$ and $i$. |
|
} |
|
\longversion{ |
|
\[ |
|
[\alpha,\mu] = [\alpha_\pi, \mu]\text, |
|
\] |
|
where $\alpha_\pi(r)_i \defeq \alpha(r)_{\pi(i)}$ for all $r \in \RR$ and $i \geq 1$. |
|
} |
|
\end{definition} |
|
|
|
\newcommand{\projn}[1]{(#1)_{1\dots n}} |
|
\newcommand{\projnname}{\projn{-}} |
|
\begin{theorem}[Weak de Finetti for quasi-Borel spaces] |
|
\label{thm:deFinetti-qbs} |
|
If $(\alpha,\mu)$ is an exchangeable probability measure on $\XI$, |
|
then there exists a probability measure $(\beta,\nu)$ in $\Pmonad(\Pmonad(X))$ such that |
|
for all $n \geq 1$, the measure $(\bind{[\beta,\nu]} \iidn)$ on $\Pmonad(\Xn)$ equals $\Pmonad(\projnname)(\alpha,\mu)$ |
|
when considered as a measure on the |
|
product measurable space $(\Xn,\bigotimes_{i = 1}^n\qbtosig {\qb X})$. |
|
(Here $\projnname\colon \XI\to X^n$ is $\projn x\defeq (x_1,\dots,x_n)$.) |
|
\end{theorem} |
|
In the theorem, $(\beta,\nu)$ represents a random variable that has a probability measure on $X$ |
|
as its value. The theorem says that (every finite prefix of) a sample sequence from $(\alpha,\mu)$ can be generated |
|
by first sampling a probability measure on~$X$ according to $(\beta,\nu)$, then generating |
|
independent $X$-valued samples from the measure, and finally forming a sequence with these samples. |
|
|
|
We call the theorem \emph{weak} for two reasons. First, the $\sigma$-algebra |
|
$\qbtosig {\qb \Xn}$ includes the product $\sigma$-algebra $\bigotimes_{i = 1}^n \qbtosig {\qb X}$, |
|
but we do not know that they are equal; two different probability |
|
measures in $\Pmonad(\Xn)$ may induce the same measure on $(\Xn,\bigotimes_{i = 1}^n \qbtosig {\qb X})$, |
|
although they always induce different measures on $(\Xn,\qbtosig {\qb \Xn})$. In the theorem, |
|
we equate such measures, which lets us use a standard |
|
technique for proving the equality of measures on product $\sigma$-algebras. Second, we are unable to |
|
construct a version of $\iidn$ for infinite sequences, i.e.\ a morphism $\Pmonad(X) \to \Pmonad(\XI)$ |
|
implementing the independent identically-distributed random sequence. The theorem is stated only for finite prefixes. |
|
|
|
The rest of this section provides an overview of our proof of Theorem~\ref{thm:deFinetti-qbs}. |
|
The starting point is to unpack definitions in the theorem, especially those related to quasi-Borel spaces, and to rewrite the statement of the theorem purely in terms of standard measure-theoretic notions. |
|
\begin{lemma} |
|
\label{lemma:deFinetti-qbs:paraphrase} |
|
Let $(\alpha,\mu)$ be an exchangeable probability measure on $\XI$. Then, |
|
the conclusion of Theorem~\ref{thm:deFinetti-qbs} holds if and only if |
|
there exist a probability |
|
measure $\xi \in G(\RR)$, a measurable function |
|
$k \colon \RR \to G(\RR)$, and $\gamma \in \qb X$ |
|
such that for all $n \geq 1$ and all $U_1,\ldots,U_n \in \qbtosig {\qb X}$, |
|
\begin{multline*} |
|
\int_{r \in \RR} \left(\prod_{i=1}^n [\alpha(r)_i \in U_i]\right)\, \dd\mu |
|
\\ |
|
{} |
|
= |
|
\int_{r \in \RR} \prod_{i = 1}^n \left(\int_{s \in \RR} \left[\gamma(s) \in U_i\right] \dd (k(r))\right)\dd\xi\text. |
|
\end{multline*} |
|
Here we express the domain of integration and the integrated variable explicitly to avoid confusion. |
|
\end{lemma} |
|
\shortversion{ |
|
\begin{proof} |
|
Let $(\alpha,\mu)$ be an exchangeable probability measure on~$\XI$. We |
|
unpack definitions in the conclusion of Theorem~\ref{thm:deFinetti-qbs}. |
|
The first definition to unpack is the notion of probability measure in $\Pmonad(\Pmonad(X))$. |
|
Here are the crucial facts that enable this unpacking. First, for every probability measure |
|
$(\beta,\nu)$ on $\Pmonad(X)$, there exist a function $\gamma \colon \RR \to X$ in $\qb X$ |
|
and a measurable $k \colon \RR \to G(\RR)$ such that $\beta(r) = [\gamma,k(r)]$ |
|
for all $r \in \RR$. Second, conversely, for a function $\gamma \in \qb X$, |
|
a measurable $k \colon \RR \to G(\RR)$, and a probability measure $\nu \in G(\RR)$, |
|
the function $(\lambda r.\,[\gamma,k(r)],\nu)$ |
|
is a probability measure in $\Pmonad(\Pmonad(X))$. Thus, we can look for $(\gamma,k,\nu)$ in |
|
the conclusion of the theorem instead of $(\beta,\nu)$. |
|
|
|
The second is the definition of $\bind{[\beta,\nu]}{\iidn}$. Using |
|
$(\gamma,k,\nu)$ instead of $(\beta,\nu)$, we find that |
|
$\bind{[\beta,\nu]}\iidn$ is the measure $[(\prod_{i=1}^n\gamma)\circ \iotan,\;(\inviotan)_*\,(\bind\nu{\lambda r.\,\bigotimes_{i=1}^n k(r)})]$. |
|
|
|
Recall that two measures $p$ and $q$ on the product space |
|
$(\Xn,\bigotimes_{i = 1}^n X)$ are |
|
equivalent when |
|
$p(U_1\times\cdots\times U_{n})$ equals $q(U_1\times\cdots\times U_{n})$ |
|
for all $U_1,\ldots, U_n\in \qbtosig{\qb X}$. Thus we must show that |
|
$(\projnname \circ \alpha)_*\mu)(U_1\times\ldots\times U_{n})$ is equal to $\big((\prod_{i = 1}^n\gamma)_*\,(\bind\nu{\lambda r.\,\bigotimes_{i = 1}^nk(r)})\big)(U_1\times\ldots\times U_{n})$. |
|
This equation is equivalent to the one in the statement of the lemma with $\xi= \nu$. |
|
\end{proof} |
|
} |
|
\longversion{ |
|
\begin{proof} |
|
Let $(\alpha,\mu)$ be an exchangeable probability measure on $\XI$. We |
|
unpack various definitions in the conclusion of Theorem~\ref{thm:deFinetti-qbs}. |
|
This semantics-preserving rewriting and some post-processing will give us |
|
the equivalence claimed in the lemma. |
|
|
|
The first definition to unpack is the notion of probability measure in $\Pmonad(\Pmonad(X))$. |
|
Here are the crucial facts that enable this unpacking. First, for every probability measure |
|
$(\beta,\nu)$ on $\Pmonad(X)$, there exist a function $\gamma \colon \RR \to X$ in $\qb X$ |
|
and a measurable $k \colon \RR \to G(\RR)$ such that for all $r \in \RR$, |
|
\[ |
|
\beta(r) = [\gamma,k(r)]\text. |
|
\] |
|
Second, conversely, for a function $\gamma \colon \RR \to X$ in $\qb X$, |
|
a measurable $k \colon \RR \to G(\RR)$, and a probability measure $\nu$ on~$\RR$, |
|
\[ |
|
(\lambda r.\,[\gamma,k(r)],\nu) |
|
\] |
|
is a probability measure in $\Pmonad(\Pmonad(X))$. Thus, we can look for $(\gamma,k,\nu)$ in |
|
the conclusion of the theorem instead of $(\beta,\nu)$. |
|
|
|
The second is the definition of $\bind{[\beta,\nu]}{\iidn}$. If we do this and use |
|
$(\gamma,k,\nu)$ instead of $(\beta,\nu)$, we have |
|
\begin{multline*} |
|
\textstyle{(\bind{[\beta,\nu]}\iidn)}\\ |
|
{} = \textstyle{[(\prod_{i=1}^n\gamma)\circ \iotan,\;(\inviotan)_*\,(\bind\nu{\lambda r.\,\bigotimes_{i=1}^n k(r)})]} |
|
\end{multline*} |
|
|
|
Now, recall that two measures $p$ and $q$ on the product space |
|
$(\Xn,\bigotimes_{i = 1}^n X)$ are |
|
equivalent if and only if for all $U_1,\ldots, U_n\in \qbtosig{\qb X}$, |
|
\[ |
|
\textstyle {p(U_1\times\ldots\times U_{n})} = \textstyle{q(U_1\times\ldots\times U_{n})\text.} |
|
\] |
|
Thus we must show that |
|
\begin{multline*} |
|
\textstyle{(\projnname \circ \alpha)_*\mu)(U_1\times\ldots\times U_{n})} |
|
\\ |
|
{} = \textstyle{\big((\prod_{i = 1}^n\gamma)_*\,(\bind\nu{\lambda r.\,\bigotimes_{i = 1}^nk(r)})\big)(U_1\times\ldots\times U_{n})} |
|
\end{multline*} |
|
This equation is equivalent to the one in the statement of the lemma, |
|
where $\xi\defeq \nu$. |
|
\end{proof} |
|
} |
|
|
|
Thus we just need to show how to construct $\xi$, $k$ and $\gamma$ in Lemma~\ref{lemma:deFinetti-qbs:paraphrase} |
|
from a given exchangeable probability measure $(\alpha,\mu)$ on $\XI$. |
|
Constructing $\xi$ and $\gamma$ is easy: |
|
\[ |
|
\xi \defeq \mu,\qquad |
|
\gamma \defeq \lambda r.\,\alpha(r)_1\text. |
|
\] |
|
Note that these definitions type-check: $\xi = \mu \in G(\RR)$, |
|
and $\gamma \in \qb X$ because $\alpha \in \qb \XI$ and the first projection |
|
$(-)_1$ is a morphism $\XI \to X$. |
|
|
|
Constructing $k$ is not that easy. We need to use the fact that |
|
$\mu$ is defined over $\RR$, a standard Borel space. This fact itself |
|
holds because all probability measures on quasi-Borel spaces use $\RR$ |
|
as their source of randomness. Define measurable functions |
|
$\alpha_e,\alpha_o \colon (\RR,\sigalg \RR) \to (\XI,\qbtosig {\qb \XI})$ by |
|
\[ |
|
\alpha_e(r)_i \defeq \alpha(r)_{2i}\quad\text{(even),}\qquad |
|
\alpha_o(r)_i \defeq \alpha(r)_{2i-1}\quad\text{(odd)}\text. |
|
\] |
|
Since $\mu$ is a probability measure on $\RR$, there exists a measurable function |
|
$k' \colon (\XI,\qbtosig {\qb \XI}) \to (G(\RR),\sigalg {G(\RR)})$, |
|
called a conditional probability kernel, such that for all measurable $f \colon \RR \to \RR$ |
|
and $U \in \inv{(\alpha_e)}(\qbtosig {\qb \XI})$, |
|
\begin{equation} |
|
\label{eqn:deFinetti-qbs:0} |
|
\int_{r \in U} f(r) \,\dd\mu = \int_{r \in U} \left(\int_\RR f \,\dd((k' \circ \alpha_e)(r))\right) \dd\mu\text. |
|
\end{equation} |
|
Define $k \defeq k' \circ \alpha_e$. |
|
|
|
Our $\xi$, $k$ and $\gamma$ satisfy the requirement in Lemma~\ref{lemma:deFinetti-qbs:paraphrase} |
|
because of the following three properties, which follow from exchangeability of |
|
$(\alpha,\mu)$. |
|
\begin{lemma} |
|
\label{lemma:deFinetti-qbs:odd} |
|
For all $n \geq 1$ and all $U_1,\ldots,U_n \in \qbtosig {\qb X}$, |
|
\[ |
|
\int_{r \in \RR} \left(\prod_{i=1}^n [\alpha(r)_i \in U_i]\right)\, \dd\mu |
|
= |
|
\int_{r \in \RR} \left(\prod_{i=1}^n [\alpha_o(r)_i \in U_i]\right)\, \dd\mu \text. |
|
\] |
|
\end{lemma} |
|
\begin{proof} |
|
Consider $n \geq 1$ and $U_1,\ldots,U_n \in \qbtosig {\qb X}$. |
|
Pick a permutation $\pi$ on positive integers such that |
|
$\pi(i) = 2i-1$ for all integers $1 \leq i \leq n$. |
|
Then, $[\alpha,\mu] = [\alpha_\pi,\mu]$ |
|
by the exchangeability of $(\alpha,\mu)$. Thus |
|
\begin{align*} |
|
\int_{r \in \RR} \left(\prod_{i=1}^n [\alpha(r)_i \in U_i]\right) \dd\mu |
|
& = |
|
\int_{r \in \RR} \left(\prod_{i=1}^n [\alpha_\pi(r)_i \in U_i]\right) \dd\mu |
|
\text, |
|
\end{align*} |
|
from which the statement follows. |
|
\end{proof} |
|
\begin{lemma} |
|
\label{lemma:deFinetti-qbs:marginal} |
|
For all $U \in \qbtosig {\qb X}$ and all $i,j \geq 1$, |
|
\[ |
|
\int_{s \in \RR} \left[\alpha_o(s)_i \in U\right] \dd(k(r)) |
|
= |
|
\int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r)) |
|
\] |
|
holds for $\mu$-almost all $r \in \RR$. |
|
\end{lemma} |
|
\shortversion{ |
|
\begin{proof} |
|
Consider a measurable set ${U \in \qbtosig {\qb X}}$ and ${i,j \geq 1}$. |
|
The function ${\lambda r.\;\int_{s \in \RR} \left[\alpha_o(s)_i \in U\right] \dd(k(r))}:{\RR\to \RR}$ is a conditional expectation of the indicator function ${\lambda s.\,[\alpha_o(s)_i \in U]}$ |
|
with respect to the probability measure $\mu$ and the $\sigma$-algebra |
|
generated by the measurable function $\alpha_e \colon \RR \to (\XI,\qbtosig {\qb \XI})$. |
|
By the almost-sure uniqueness of conditional expectation, it suffices to show that |
|
$\lambda r.\, \int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r))$ |
|
is also a conditional expectation of $\lambda s.\,[\alpha_o(s)_i \in U]$ |
|
with respect to $\mu$ and $\alpha_e$. Pick a measurable subset $V \in \qbtosig {\qb \XI}$. Then: |
|
\begin{align*} |
|
& \int_{r \in \RR} \left[\alpha_e(r) \in V\right] \cdot |
|
\left(\int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r))\right) |
|
\dd\mu |
|
\\ |
|
& \qquad \qquad \qquad {} = |
|
\int_{r \in \RR} \left[\alpha_o(r)_j \in U \wedge \alpha_e(r) \in V\right] \dd\mu |
|
\\ |
|
& \qquad \qquad \qquad {} = |
|
\int_{r \in \RR} \left[\alpha_o(r)_i \in U \wedge \alpha_e(r) \in V\right] \dd\mu\text. |
|
\end{align*} |
|
The first equation holds because the function |
|
$\lambda r.\, \int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r))$ |
|
is a conditional expectation of $\lambda s.\,[\alpha_o(s)_j \in U]$ with respect to $\mu$ and $\alpha_e$. |
|
The second equation follows from the exchangeability of $(\alpha,\mu)$. We have just shown that |
|
$\lambda r.\, \int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r))$ is a conditional |
|
expectation of $\lambda s.\,[\alpha_o(s)_i \in U]$ |
|
with respect to $\mu$ and $\alpha_e$. |
|
\end{proof} |
|
} |
|
\longversion{ |
|
\begin{proof} |
|
Consider a measurable subset $U \in \qbtosig {\qb X}$ and $i,j \geq 1$. |
|
We remind the reader that when viewed as a function on $r \in \RR$, |
|
\[ |
|
\int_{s \in \RR} \left[\alpha_o(s)_i \in U\right] \dd(k(r)) |
|
\] |
|
is a conditional expectation of the indicator function $\lambda s.\,[\alpha_o(s)_i \in U]$ |
|
with respect to the measure $\mu$ and the $\sigma$-algebra |
|
generated by the measurable function $\alpha_e \colon \RR \to (\XI,\qbtosig {\qb \XI})$. |
|
By the almost-sure uniqueness of conditional expectation, it suffices to show that |
|
\[ |
|
\lambda r.\, \int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r)) |
|
\] |
|
is also a conditional expectation of $\lambda s.\,[\alpha_o(s)_i \in U]$ |
|
with respect to $\mu$ and $\alpha_e$. Pick a measurable subset $V \in \qbtosig {\qb \XI}$. Then, |
|
\begin{align*} |
|
& \int_{r \in \RR} \left[\alpha_e(r) \in V\right] \cdot |
|
\left(\int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r))\right) |
|
\dd\mu |
|
\\ |
|
& \qquad \qquad \qquad {} = |
|
\int_{r \in \RR} \left[\alpha_o(r)_j \in U \wedge \alpha_e(r) \in V\right] \dd\mu |
|
\\ |
|
& \qquad \qquad \qquad {} = |
|
\int_{r \in \RR} \left[\alpha_o(r)_i \in U \wedge \alpha_e(r) \in V\right] \dd\mu\text. |
|
\end{align*} |
|
The first equation holds because the function |
|
\[ |
|
\lambda r.\, \int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r)) |
|
\] |
|
is a conditional expectation of $\lambda s.\,[\alpha_o(s)_j \in U]$ with respect to $\mu$ and $\alpha_e$. |
|
The second equation follows from the exchangeability of $(\alpha,\mu)$. We have just shown that |
|
$\lambda r.\, \int_{s \in \RR} \left[\alpha_o(s)_j \in U\right] \dd(k(r))$ is a conditional |
|
expectation of $\lambda s.\,[\alpha_o(s)_i \in U]$ |
|
with respect to $\mu$ and $\alpha_e$. |
|
\end{proof} |
|
} |
|
\begin{lemma} |
|
\label{lemma:deFinetti-qbs:independence} |
|
For all $n \geq 1$ and all $U_1,\ldots,U_n \in \qbtosig {\qb X}$, |
|
\begin{multline*} |
|
\int_{s \in \RR} \left(\prod_{i = 1}^n\left[\alpha_o(s)_i \in U_i\right]\right) \dd(k(r)) |
|
\\ |
|
{} |
|
= |
|
\prod_{i = 1}^n \int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r)) |
|
\end{multline*} |
|
holds for $\mu$-almost all $r \in \RR$. |
|
\end{lemma} |
|
\shortversion{ |
|
\begin{proof}[Proof notes] |
|
Use induction on $n \geq 1$. There is nothing to |
|
prove for the base case $n = 1$. To handle the inductive case, assume that $n > 1$. |
|
Let $U_1,\ldots,U_n$ be subsets in $\qbtosig {\qb X}$. |
|
Define a function $\alpha' \colon \RR \to \XI$ as follows: |
|
\[ |
|
\alpha'(r)_i = |
|
\left\{\begin{array}{ll} |
|
\alpha_o(r)_i & \mbox{if $1 \leq i \leq n-1$} |
|
\\ |
|
\alpha_e(r)_{i-n+1} & \mbox{otherwise.} |
|
\end{array}\right. |
|
\] |
|
Then, $\alpha'$ is in $\qb \XI$, so that $\alpha'$ is a measurable |
|
function $(\RR,\sigalg \RR) \to (\XI,\sigalg {\qb \XI})$. Thus there exists a measurable |
|
$k'_0 \colon (\XI,\qbtosig {\qb \XI}) \to (G(\RR),\sigalg {G(\RR)})$, the |
|
conditional probability kernel, such that |
|
for all measurable functions $f \colon \RR \to \RR$, |
|
$\lambda r.\, \int_\RR f \,\dd((k'_0 \circ \alpha')(r))$ |
|
is a conditional expectation of $f$ with respect to $\mu$ |
|
and the $\sigma$-algebra generated by~$\alpha'$. |
|
Define $k' \colon \RR \to G(\RR) = k'_0 \circ \alpha'$. |
|
Then $k'$ is measurable because so are $k'_0$ and $\alpha'$. |
|
More importantly, for $\mu$-almost all $r \in \RR$, |
|
\begin{equation} |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k(r)) |
|
= |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k'(r))\text. |
|
\label{eqn:deFinetti:3} |
|
\end{equation} |
|
The proof of this equality appears in the full version of this paper. |
|
|
|
Recall that $k = k_0 \circ \alpha_e$ and $k' = k'_0 \circ \alpha'$ are |
|
defined in terms of conditional expectation. Thus, they inherit all the properties |
|
of conditional expectation. In particular, |
|
for $\mu$-almost all $r \in \RR$ |
|
and all measurable $h \colon \RR \to \RR$, |
|
\begin{align} |
|
\begin{split} |
|
& \int_{s \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r)) |
|
\\ |
|
& \quad {}= |
|
\int_{s \in \RR} \left(\int_{t \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(t)_i \in U_i\right] \dd(k'(s))\right) |
|
\dd(k(r))\text, |
|
\end{split} |
|
\label{eqn:deFinetti:4} |
|
\\[1ex] |
|
\begin{split} |
|
& \int_{s \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k'(r)) |
|
\\ |
|
& \quad {}= |
|
\prod_{i = 1}^{n-1} \left[\alpha_o(r)_i \in U_i\right] |
|
\cdot |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k'(r))\text, |
|
\end{split} |
|
\label{eqn:deFinetti:5} |
|
\\[1ex] |
|
\begin{split} |
|
& \int_{s \in \RR} \left(h(s) |
|
\cdot \int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(s))\right) |
|
\dd(k(r)) |
|
\\ |
|
& \quad {}= |
|
\left(\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(r))\right) |
|
\cdot |
|
\left(\int_{s \in \RR} h(s)\, \dd(k(r))\right)\text. |
|
\end{split} |
|
\label{eqn:deFinetti:6} |
|
\end{align} |
|
Using the assumption \eqref{eqn:deFinetti:3} |
|
and the properties \eqref{eqn:deFinetti:4}, \eqref{eqn:deFinetti:5} and \eqref{eqn:deFinetti:6}, |
|
we complete the proof of the inductive case as follows: |
|
for all subsets $V \in \inv{(\alpha_e)}(\qbtosig {\qb \XI})$, |
|
\begin{align*} |
|
& |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\,\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\int_{t \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(t)_i \in U_i\right] |
|
\dd(k'(s))\,\dd(k(r))\,\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\prod_{i = 1}^{n-1} \left[\alpha_o(s)_i \in U_i\right] |
|
\\* |
|
& |
|
\phantom{{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR}} |
|
{} \cdot |
|
\int_{t\in\RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k'(s))\, |
|
\dd(k(r))\, |
|
\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\prod_{i = 1}^{n-1} \left[\alpha_o(s)_i \in U_i\right] |
|
\\* |
|
& |
|
\phantom{ |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR}} |
|
{}\cdot |
|
\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(s))\,\dd(k(r))\,\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\left(\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(r))\right) |
|
\\* |
|
& |
|
\phantom{ |
|
{} = |
|
\int_{r \in V}} |
|
{} |
|
\cdot |
|
\left(\int_{s \in \RR} \prod_{i = 1}^{n-1} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\right)\, |
|
\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\prod_{i = 1}^n \int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] |
|
\dd(k(r))\,\dd\mu\text. |
|
\end{align*} |
|
The first and the second equalities hold because of \eqref{eqn:deFinetti:4} and \eqref{eqn:deFinetti:5}. |
|
The third equality uses \eqref{eqn:deFinetti:3}, and the fourth the equality |
|
in \eqref{eqn:deFinetti:6}. The fifth follows from the induction hypothesis. Our derivation |
|
implies that both |
|
$\lambda r.\,\int_{s \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))$ |
|
and |
|
$\lambda r.\,\prod_{i = 1}^n \int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))$ |
|
are conditional expectations of the same function with respect to $\mu$ and the same $\sigma$-algebra. |
|
So, they are equal for $\mu$-almost all inputs~$r$. |
|
\end{proof} |
|
} |
|
\longversion{ |
|
\begin{proof} |
|
We prove the lemma by induction on $n \geq 1$. There is nothing to |
|
prove for the base case $n = 1$. To handle the inductive case, assume that $n > 1$. |
|
Let $U_1,\ldots,U_n$ be subsets in $\qbtosig {\qb X}$. |
|
Define a function $\alpha' \colon \RR \to \XI$ as follows: |
|
\[ |
|
\alpha'(r)_i = |
|
\left\{\begin{array}{ll} |
|
\alpha_o(r)_i & \mbox{if $1 \leq i \leq n-1$} |
|
\\ |
|
\alpha_e(r)_{i-n+1} & \mbox{otherwise.} |
|
\end{array}\right. |
|
\] |
|
Then, $\alpha'$ is in $\qb \XI$, so that $\alpha'$ is a measurable |
|
function $(\RR,\sigalg \RR) \to (\XI,\sigalg {\qb \XI})$. Thus, there exists a measurable |
|
$k'_0 \colon (\XI,\qbtosig {\qb \XI}) \to (G(\RR),\sigalg {G(\RR)})$, called |
|
conditional probability kernel, such that |
|
for all measurable $f \colon \RR \to \RR$, |
|
\[ |
|
\lambda r.\, \int_\RR f \,\dd((k'_0 \circ \alpha')(r)) |
|
\] |
|
is a conditional expectation of $f$ with respect to $\mu$ |
|
and the $\sigma$-algebra generated by $\alpha'$. Let |
|
\begin{align*} |
|
k' & \colon \RR \to G(\RR) |
|
\\ |
|
k' & = k'_0 \circ \alpha'\text. |
|
\end{align*} |
|
The function $k'$ is measurable because so are $k'_0$ and $\alpha'$. |
|
At the end of this proof, we will show that for $\mu$-almost all $r \in \RR$, |
|
\begin{multline} |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k(r)) |
|
\\ |
|
= |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k'(r))\text. |
|
\label{eqn:deFinetti:3} |
|
\end{multline} |
|
For now, just assume that this equation holds and see how this assumption lets |
|
us complete the proof. |
|
|
|
Recall that $k = k_0 \circ \alpha_e$ and $k' = k'_0 \circ \alpha'$ are |
|
defined in terms of conditional expectation. Thus, they inherit all the properties |
|
of conditional expectation after minor adjustment. In particular, |
|
since the $\sigma$-algebra generated by $\alpha'$ is larger than the one |
|
generated by $\alpha_e$ and it makes $\alpha_i$ measurable for all $1 \leq i \leq (n-1)$, |
|
we have the following equalities: for $\mu$-almost all $r \in \RR$ |
|
and all measurable functions $h \colon \RR \to \RR$, |
|
\begin{multline} |
|
\int_{s \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r)) |
|
\\ |
|
= |
|
\int_{s \in \RR} \left(\int_{t \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(t)_i \in U_i\right] \dd(k'(s))\right) |
|
\dd(k(r))\text, |
|
\label{eqn:deFinetti:4} |
|
\end{multline} |
|
\begin{multline} |
|
\int_{s \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k'(r)) |
|
\\ |
|
= |
|
\prod_{i = 1}^{n-1} \left[\alpha_o(r)_i \in U_i\right] |
|
\cdot |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k'(r))\text, |
|
\label{eqn:deFinetti:5} |
|
\end{multline} |
|
\begin{multline} |
|
\int_{s \in \RR} \left(h(s) |
|
\cdot \int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(s))\right) |
|
\dd(k(r)) |
|
\\ |
|
= |
|
\left(\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(r))\right) |
|
\cdot |
|
\left(\int_{s \in \RR} h(s)\, \dd(k(r))\right)\text. |
|
\label{eqn:deFinetti:6} |
|
\end{multline} |
|
Using the assumption \eqref{eqn:deFinetti:3} |
|
and the properties \eqref{eqn:deFinetti:4}, \eqref{eqn:deFinetti:5} and \eqref{eqn:deFinetti:6}, |
|
we complete the proof of the inductive case as follows: |
|
for all subsets $V \in \inv{(\alpha_e)}(\qbtosig {\qb \XI})$, |
|
\begin{align*} |
|
& |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\,\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\int_{t \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(t)_i \in U_i\right] |
|
\dd(k'(s))\,\dd(k(r))\,\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\prod_{i = 1}^{n-1} \left[\alpha_o(s)_i \in U_i\right] |
|
\\ |
|
& |
|
\phantom{{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR}} |
|
{} \cdot |
|
\int_{t\in\RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k'(s))\, |
|
\dd(k(r))\, |
|
\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR} |
|
\prod_{i = 1}^{n-1} \left[\alpha_o(s)_i \in U_i\right] |
|
\\ |
|
& |
|
\phantom{ |
|
{} = |
|
\int_{r \in V} |
|
\int_{s \in \RR}} |
|
{}\cdot |
|
\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(s))\,\dd(k(r))\,\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\left(\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(r))\right) |
|
\\ |
|
& |
|
\phantom{ |
|
{} = |
|
\int_{r \in V}} |
|
{} |
|
\cdot |
|
\left(\int_{s \in \RR} \prod_{i = 1}^{n-1} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\right)\, |
|
\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\int_{t \in \RR} \left[\alpha_o(t)_n \in U_n\right] \dd(k(r)) |
|
\\ |
|
& |
|
\phantom{{} = \int_{r \in V}} |
|
{}\cdot |
|
\prod_{i = 1}^{n-1} \int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\, |
|
\dd\mu |
|
\\ |
|
& |
|
{} = |
|
\int_{r \in V} |
|
\prod_{i = 1}^n \int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] |
|
\dd(k(r))\,\dd\mu\text. |
|
\end{align*} |
|
The first and the second equalities hold because of \eqref{eqn:deFinetti:4} and \eqref{eqn:deFinetti:5}. |
|
The third equality uses our assumption in \eqref{eqn:deFinetti:3}, and the fourth the equality |
|
in \eqref{eqn:deFinetti:6}. The fifth equality follows from the induction hypothesis. Our derivation |
|
implies that both |
|
\begin{align*} |
|
& \lambda r.\,\int_{s \in \RR} \prod_{i = 1}^{n} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r)) |
|
\\ |
|
& \mbox{and}\quad |
|
\lambda r.\,\prod_{i = 1}^n \int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r)) |
|
\end{align*} |
|
are conditional expectations of the same function with respect to $\mu$ and the same $\sigma$-algebra. |
|
Thus, they are equal for $\mu$-almost all inputs $r \in \RR$. |
|
|
|
It remains to show that the equality in \eqref{eqn:deFinetti:3} holds for $\mu$-almost all $r \in \RR$. |
|
Define two functions $h, h' \colon \RR \to \RR$ as follows: |
|
\begin{align*} |
|
h(r) & = \int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k(r))\text, |
|
\\ |
|
h'(r) & = \int_{s \in \RR} \left[\alpha_o(s)_n \in U_n\right] \dd(k'(r))\text. |
|
\end{align*} |
|
Let $\Sigma$ and $\Sigma'$ be the $\sigma$-algebras generated by $\alpha_e$ and $\alpha'$, respectively. |
|
Then, $\Sigma \subseteq \Sigma'$, the function $h$ is $\Sigma$-measurable and bounded, |
|
and $h'$ is $\Sigma'$-measurable and bounded. Let |
|
$L^2(\RR,\Sigma',\mu)$ be the Hilbert space of the equivalence classes of |
|
square integrable functions that are $\Sigma'$-measurable. |
|
Let $M$ be the subspace of $L^2(\RR,\Sigma',\mu)$ consisting of the equivalence |
|
classes of some square-integrable and $\Sigma$-measurable functions. Then, |
|
\[ |
|
[h] \in M |
|
\quad\mbox{and}\quad |
|
[h'] \in L^2(\RR,\Sigma',\mu)\text, |
|
\] |
|
where $[h]$ and $[h']$ are equivalence classes of $L^2(\RR,\Sigma',\mu)$. Furthermore, $[h]$ is |
|
the projection of $[h']$ to the subspace $M$, because $h$ and |
|
$h'$ are conditional expectations of the same bounded function with respect to $\mu$. |
|
Thus, when $\Vert {-} \Vert_2$ is the $L^2$ norm with respect to the probability measure $\mu$, |
|
\[ |
|
\Vert h \Vert_2 \leq \Vert h' \Vert_2\text. |
|
\] |
|
The equality holds here if and only if $[h] = [h']$, i.e.~$h$ and $h'$ are equal except at some $\mu$-null |
|
set in $\Sigma'$. So, it is sufficient to prove that $\Vert h \Vert_2 = \Vert h' \Vert_2$. We |
|
rewrite $\Vert h \Vert^2_2$ as follows: |
|
\begin{align*} |
|
& \Vert h \Vert^2_2 |
|
\\ |
|
& = \int_\RR h^2 \dd\mu |
|
\\ |
|
& = \int_{r \in \RR} \left(\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n \right] \dd(k(r))\right)^2 \dd\mu |
|
\\ |
|
& = \int_{r \in \RR} \left(\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n \right] \dd((k_0 \circ \alpha_e)(r))\right)^2 \dd\mu |
|
\\ |
|
& = \int_{u \in \XI} \left(\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n \right] \dd(k_0(u))\right)^2 |
|
\dd\left((\alpha_e)_*(\mu)\right) |
|
\\ |
|
& = \int_{u \in \XI} \left(\int_{x \in X} \left[x \in U_n \right] \dd\left((\alpha_o(-)_n)_*(k_0(u))\right)\right)^2 |
|
\\ |
|
& \phantom{= \int_{u \in \XI} \left(\right)} |
|
\dd\left((\alpha_e)_*(\mu)\right) |
|
\\ |
|
& = \int_{(x_0,u) \in X \times \XI} \left(\int_{x \in X} \left[x \in U_n \right] \dd\left((\alpha_o(-)_n)_*(k_0(u))\right)\right)^2 |
|
\\ |
|
& \phantom{= \int_{u \in \XI} \left(\right)} |
|
\qquad\qquad |
|
\dd\left((\alpha_o(-)_n,\alpha_e)_*(\mu)\right)\text. |
|
\end{align*} |
|
At the last line, $X \times \XI$ denotes the product measurable space |
|
$(X \times \XI, \qbtosig {\qb X}\otimes \qbtosig {\qb \XI})$. A similar rewriting gives |
|
\begin{align*} |
|
& \Vert h' \Vert^2_2 = |
|
\\ |
|
& \quad \int_{(x_0,u) \in X \times \XI} \left(\int_{x \in X} \left[x \in U_n \right] \dd\left((\alpha_o(-)_n)_*(k_0'(u))\right)\right)^2 |
|
\\ |
|
& \phantom{= \int_{u \in \XI} \left(\right)} |
|
\qquad\qquad |
|
\dd\left((\alpha_o(-)_n,\alpha'_e)_*(\mu)\right)\text. |
|
\end{align*} |
|
By the exchangeability of $(\alpha,\mu)$, |
|
\begin{equation} |
|
\label{eqn:deFinetti:6.5} |
|
(\alpha_o(-)_n,\alpha_e)_*(\mu) = (\alpha_o(-)_n,\alpha')_*(\mu) |
|
\end{equation} |
|
as probability measures on $(X \times \XI,\qbtosig {\qb {(X \times \XI)}})$. |
|
This equality continues to hold when its LHS and RHS are viewed as |
|
probability measures on $(X \times \XI, \qbtosig {\qb X} \otimes \qbtosig {\qb \XI})$. |
|
Then, the following functions from $X \times \XI$ to $\RR$: |
|
\begin{align} |
|
& \lambda (x_0,u).\, \int_{x \in X} \left[x \in U_n \right] \dd\left((\alpha_o(-)_n)_*(k_0(u))\right) |
|
\label{eqn:deFinetti:7} |
|
\\ |
|
& \mbox{and}\quad |
|
\lambda (x_0,u).\, |
|
\int_{x \in X} \left[x \in U_n \right] \dd\left((\alpha_o(-)_n)_*(k_0'(u))\right) |
|
\label{eqn:deFinetti:8} |
|
\end{align} |
|
are conditional expectations of $\lambda (x,u).\,[x \in U_n]$ with respect to |
|
$(\alpha_o(-)_n,\alpha_e)_*(\mu)$ on $(X \times \XI,\qbtosig {\qb X} \otimes \qbtosig {\qb \XI})$ |
|
and the $\sigma$-algebra generated by the projection $\lambda (x,u).\,u$. This is because |
|
$k_0$ and $k'_0$ are appropriate conditional probability kernels. Concretely, |
|
for all $U \in \qbtosig {\qb \XI}$, we can show that |
|
\begin{align*} |
|
& \int_{(x_0,u) \in X \times U} |
|
\int_{x \in X} \left[x \in U_n \right] \dd((\alpha_o(-)_n)_*(k_0(u))) |
|
\\ |
|
& \phantom{\int_{(x_0,u) \in X \times U} |
|
\int_{x \in X} \left[x \in U_n \right]} |
|
\quad |
|
\dd((\alpha_o(-)_n,\alpha_e)_*(\mu)) |
|
\\ |
|
& = |
|
\int_{u \in U} |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n \right] \dd(k_0(u))\,\dd((\alpha_e)_*(\mu)) |
|
\\ |
|
& = |
|
\int_{r \in \inv{\alpha}_e(U)} |
|
\int_{s \in \RR} \left[\alpha_o(s)_n \in U_n \right] \dd((k_0 \circ \alpha_e)(u))\,\dd\mu |
|
\\ |
|
& = |
|
\int_{r \in \inv{\alpha}_e(U)} |
|
\left[\alpha_o(r)_n \in U_n \right] \dd\mu\text. |
|
\end{align*} |
|
By similar reasoning and the equation in \eqref{eqn:deFinetti:6.5}, |
|
\begin{align*} |
|
& \int_{(x_0,u) \in X \times U} |
|
\int_{x \in X} \left[x \in U_n \right] \dd((\alpha_o(-)_n)_*(k'_0(u))) |
|
\\ |
|
& \phantom{\int_{(x_0,u) \in X \times U} |
|
\int_{x \in X} \left[x \in U_n \right]} |
|
\quad |
|
\dd((\alpha_o(-)_n,\alpha)_*(\mu)) |
|
\\ |
|
& = \int_{(x_0,u) \in X \times U} |
|
\int_{x \in X} \left[x \in U_n \right] \dd((\alpha_o(-)_n)_*(k'_0(u))) |
|
\\ |
|
& \phantom{\int_{(x_0,u) \in X \times U} |
|
\int_{x \in X} \left[x \in U_n \right]} |
|
\quad |
|
\dd((\alpha_o(-)_n,\alpha')_*(\mu)) |
|
\\ |
|
& = |
|
\int_{r \in \inv{\alpha}_e(U)} |
|
\left[\alpha_o(r)_n \in U_n \right] \dd\mu\text. |
|
\end{align*} |
|
The outcomes of these calculations imply that |
|
the functions in \eqref{eqn:deFinetti:7} and \eqref{eqn:deFinetti:8} |
|
are the claimed conditional expectations. Thus, these functions |
|
are equal for $(\alpha_o(-)_n,\alpha_e)_*(\mu)$-almost all $(x_0,u)$. From this it follows that |
|
$\Vert h \Vert_2^2 = \Vert h' \Vert_2^2$, as desired. |
|
\end{proof} |
|
} |
|
The following calculation combines these lemmas and shows that $\xi$, $k$ and $\gamma$ |
|
satisfy the requirement in Lemma~\ref{lemma:deFinetti-qbs:paraphrase}: |
|
\begin{align*} |
|
& \int_{r \in \RR} \prod_{i = 1}^n \left[\alpha(r)_i \in U_i\right] \dd\mu |
|
\\ |
|
& |
|
= \int_{r \in \RR} \prod_{i = 1}^n \left[\alpha_o(r)_i \in U_i\right] \dd\mu |
|
& \mbox{Lem.~\ref{lemma:deFinetti-qbs:odd}} |
|
\\ |
|
& |
|
= \int_{r \in \RR} \left(\int_{s \in \RR} \prod_{i = 1}^n \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\right)\dd\mu |
|
& \mbox{Eq.~\eqref{eqn:deFinetti-qbs:0}} |
|
\\ |
|
& |
|
= \int_{r \in \RR} \prod_{i = 1}^n \left(\int_{s \in \RR} \left[\alpha_o(s)_i \in U_i\right] \dd(k(r))\right) \dd\mu |
|
& \mbox{Lem.~\ref{lemma:deFinetti-qbs:independence}} |
|
\\ |
|
& |
|
= \int_{r \in \RR} \prod_{i = 1}^n \left(\int_{s \in \RR} \left[\alpha_o(s)_1 \in U_i\right] \dd(k(r))\right) \dd\mu |
|
& \mbox{Lem.~\ref{lemma:deFinetti-qbs:marginal}} |
|
\\ |
|
& |
|
= \int_{r \in \RR} \prod_{i = 1}^n \left(\int_{s \in \RR} \left[\gamma(s) \in U_i\right] \dd(k(r))\right) \dd\xi |
|
& \mbox{Def. of $\gamma,\xi$}\text. |
|
\end{align*} |
|
This concludes our proof outline for Theorem~\ref{thm:deFinetti-qbs}. |
|
|
|
\section{Related work}\label{sec:related} |
|
|
|
|
|
\subsection{Quasi-topological spaces and categories of functors} |
|
|
|
|
|
Our development of a cartesian closed category from measurable spaces mirrors the development of cartesian closed categories of topological spaces |
|
over the years. |
|
|
|
For example, quasi-Borel spaces are reminiscent of \emph{subsequential spaces}~\cite{johnstone-topological-topos}: a set $X$ together with a collection of functions $Q\subseteq {[\NN\cup \{\infty\}\to X]}$ satisfying some conditions. |
|
The functions in $Q$ are thought of as convergent sequences. |
|
Another notion of generalized topological space is \emph{C-space}~\cite{xu-escardo}: a set~$X$ together with a collection $Q\subseteq [2^\NN\to X]$ |
|
of `probes' satisfying some conditions; |
|
this is a variation on Spanier's early notion of \emph{quasi-topological space}~\cite{spanier:quasitopologies}. |
|
Another reminiscent notion in the context of differential geometry is a \emph{diffeological space}~\cite{bh-convenient}: |
|
a set $X$ together with a set $Q_U\subseteq [U\to X]$ of `plots' for each open subset $U$ of $\RR^n$ satisfying some conditions. |
|
These examples all form cartesian closed categories. |
|
|
|
A common pattern is that these spaces can be understood as extensional (concrete) sheaves |
|
on an established category of spaces. |
|
Let $\SMeas$ be the category of standard Borel spaces and measurable functions. |
|
There is a functor |
|
\shortversion{ |
|
$J\colon \QBS\to[\op\SMeas,\Set]$ |
|
} |
|
\longversion{ |
|
\[ |
|
J\colon \QBS\to[\op\SMeas,\Set] |
|
\] |
|
} |
|
with $\big(J(X,\qb X))(Y,\sigalg Y\big)\defeq \QBS\big((Y,\sigtoqb{\sigalg Y}),(X,\qb X)\big)$, |
|
which is full and faithful by Prop.~\ref{prop:adjunction}(2). |
|
We can characterize those functors that arise in this way. |
|
\begin{proposition}\label{prop:extensionalpresheaf} |
|
Let $F\colon\op\SMeas\to\Set$ be a functor. |
|
The following are equivalent: |
|
\begin{itemize} |
|
\item $F$ is naturally isomorphic to $J(X,\qb X)$, for some quasi-Borel space $(X,\qb X)$; |
|
\item $F$ preserves countable products and $F$ is extensional: the functions $i_{(X,\sigalg X)}\colon F(X,\sigalg X)\to\Set(X,F(1))$ are injective, where $(i_{(X,\sigalg X)}(\xi))(x)=(F(\ulcorner x\urcorner))(\xi)$, and we consider $x\in X$ as a function $\ulcorner x\urcorner \colon 1\to X$. |
|
\end{itemize} |
|
\end{proposition} |
|
|
|
There are similar characterizations of subsequential spaces~\cite{johnstone-topological-topos}, quasi-topological spaces~\cite{dubuc-concrete-quasitopoi} and diffeological spaces~\cite{bh-convenient}. Prop.~\ref{prop:extensionalpresheaf} |
|
is an instance of a general pattern (e.g.~\cite{bh-convenient,dubuc-concrete-quasitopoi}); |
|
but that is not to say that the definition of quasi-Borel space (Def.~\ref{def:qbs}) |
|
arises automatically. |
|
The method of |
|
extensional presheaves also arises in other models of computation |
|
such as finiteness spaces~\cite{ehrhard-extensional} and realizability models~\cite{rosolini-streicher}. |
|
This work appears to be the first application to probability theory, |
|
although via Prop.~\ref{prop:extensionalpresheaf} there are connections |
|
to Simpson's probability sheaves~\cite{simpson-cippmi}. |
|
|
|
The characterization of Prop.~\ref{prop:extensionalpresheaf} gives a canonical categorical status to quasi-Borel spaces. |
|
It also connects with our earlier work~\cite{statonyangheunenkammarwood:higherorder}, which used the cartesian closed category of countable-product-preserving functors |
|
in $[\op\SMeas, \Set]$. |
|
Quasi-Borel spaces have several advantages over this functor category. |
|
For one thing, they are more concrete, leading |
|
to better intuitions for their constructions. For example, measures in~\cite{statonyangheunenkammarwood:higherorder} are built abstractly from left Kan extensions, whereas for quasi-Borel spaces they have a straightforward concrete definition (Def.~\ref{def:probabilitymeasure}). |
|
For another thing, in contrast to the functor category in~\cite{statonyangheunenkammarwood:higherorder}, quasi-Borel spaces form a well-pointed category: |
|
if two morphisms $(X,\qb X)\to (Y,\qb Y)$ are different |
|
then they disagree on some point in $X$. |
|
From the perspective of semantics of programming languages, |
|
where terms in context $\Gamma\vdash t :A$ are interpreted as morphisms |
|
$\denot t \colon \denot \Gamma\to\denot A$, well-pointedness is a crucial property. |
|
It says that if two open terms |
|
are different, $\denot t\neq \denot u:\denot \Gamma\to\denot A$, |
|
then there is a ground context $\mathcal C \colon 1\to\denot \Gamma$ that |
|
distinguishes them: $\denot{\mathcal C[t]}\neq \denot{\mathcal C[u]}:1\to \denot A$. |
|
|
|
Quasi-Borel spaces add objects to make the category of measurable spaces cartesian closed. Another interesting future direction is to add morphisms to make more objects isomorphic, and so find a cartesian closed subcategory~\cite{steenrod:convenient}. |
|
|
|
|
|
\subsection{Domains and valuations} |
|
In this paper our starting point has been the standard foundation for probability theory, |
|
based on $\sigma$-algebras and probability measures. |
|
An alternative foundation for probability is based on topologies and valuations. |
|
An advantage of our starting point is that we can reference the canon of work on |
|
probability theory. Having said this, an advantage to the approach based on valuations |
|
is that it is related to domain theoretic methods, which have already been used to |
|
give semantics to programming languages. |
|
|
|
Jones and Plotkin~\cite{jones-plotkin} showed that valuations form a monad which is |
|
analogous to our probability monad. However, there is considerable debate |
|
about which cartesian closed category this monad should be based on~(e.g.~\cite{jung-tix,gl-qrb-domains}). |
|
For a discussion of the concerns in the context of programming languages, see e.g.~\cite{escardo-high-type-prob-testing}. |
|
One recent proposal is to use Girard's probabilistic coherence spaces~\cite{etp-pcoh}. |
|
Another is to use a topological domain theory as a cartesian closed category for analysis and probability~(\cite{bss-convenient-domains,pape-streicher,huang-morrisett}). |
|
|
|
Concerns about probabilistic powerdomains have led instead to domains of random variables~(e.g.~\cite{mislove-randvar,barker-monad,scott-stochastic}). |
|
We cannot yet connect formally with this work, but there are many intuitive links. For example, our measures on quasi-Borel spaces (Def.~\ref{def:probabilitymeasure}) are reminiscent of continuous random variables on a dcpo. |
|
|
|
An additional advantage of a domain theoretic approach is that it naturally |
|
supports recursion. We are currently investigating a notion of `ordered quasi-Borel |
|
space', by enriching Prop.~\ref{prop:extensionalpresheaf} over dcpo's. |
|
|
|
\subsection{Other related work} |
|
Our work is related to two recent semantic studies on probabilistic |
|
programming languages. The first is Borgstr\"om et al.'s \emph{operational} (not denotational as |
|
in this paper) semantics for |
|
a higher-order probabilistic programming language with continuous |
|
distributions~\cite{blgs-lambda-prob-untyped}, |
|
which has been used to justify a basic inference algorithm for the language. |
|
Recently, Culpepper and Cobb refined |
|
this operational approach using logical relations~\cite{Culpepper-esop17}. The second study is Freer |
|
and Roy's results on a computable variant of de Finetti's theorem and its implication |
|
on exchangeable random processes implemented in |
|
higher-order probabilistic programming languages~\cite{FreerR12}. One interesting future direction is |
|
to revisit the results about logical relations and computability in these studies with quasi-Borel spaces, |
|
and to see whether they can be extended to spaces other than standard Borel spaces. |
|
|
|
|
|
\section{Conclusion}\label{sec:conclusion} |
|
|
|
|
|
We have shown that quasi-Borel spaces (\S\ref{sec:quasiborel}) |
|
support higher-order functions (\S\ref{sec:structure}) |
|
as well as spaces of probability measures (\S\ref{sec:giry}). |
|
We have illustrated the power of this new formalism by |
|
giving a semantic analysis of Bayesian regression (\S\ref{sec:example}), |
|
by rephrasing the randomization lemma as a quotient-space construction (\S\ref{sec:functions}), |
|
and by showing that it supports de Finetti's theorem (\S\ref{sec:definetti}). |
|
|
|
|
|
|
|
|
|
\shortversion{ |
|
\section*{Acknowledgment} |
|
We thank Radha Jagadeesan and Dexter Kozen for encouraging us to think about a well-pointed cartesian closed category for probability theory, Vincent Danos and Dan Roy for nudging us to work on de Finetti's theorem, Mike Mislove for discussions of quasi-Borel spaces, and Martin Escard\'o for explaining C-spaces, and Alex Simpson for detailed report with many suggestions. This research was supported by a Royal Society Research Fellowship and EPSRC grants EP/L002388/2 and EP/N007387/1, and also by Institute for Information \& communications Technology Promotion (IITP) grant funded by the Korea government (MSIP) (No.R0190-16-2011, Development of Vulnerability Discovery Technologies for IoT Software Security). |
|
} |
|
\longversion{ |
|
\section*{Acknowledgment} |
|
We thank Radha Jagadeesan and Dexter Kozen for encouraging us to think about a well-pointed cartesian closed category for probability theory, Vincent Danos and Dan Roy for nudging us to work on de Finetti's theorem, Mike Mislove for discussions of quasi-Borel spaces, Martin Escard\'o for explaining C-spaces, and Alex Simpson for detailed report with many suggestions. This research was supported by a Royal Society Research Fellowship and EPSRC grants EP/L002388/2 and EP/N007387/1, and also by an Institute for Information \& communications Technology Promotion (IITP) grant funded by the Korea government (MSIP) (No.R0190-16-2011, Development of Vulnerability Discovery Technologies for IoT Software Security). |
|
} |
|
|
|
|
|
|
|
|
|
\bibliographystyle{IEEEtranS} |
|
\bibliography{lics2017} |
|
\end{document} |
|
|