Each result cited is universally quantified over the data in its own statement.
Throughout, Td is the collection of the subsets of Z that are open in (Z,d), a topology on Z by Metric Open Sets Form a Topology, and "open", "closed" and "compact" for subsets of Z refer to (Z,Td). We write Q>0 for the set of positive rational numbers, Bd(q,s) for the open ball (open ball) and Bˉd(q,s)={z∈Z:d(q,z)≤s} for the closed ball. For A∈Y and B∈B(Z) the rectangle A×B belongs to Y⊗B(Z) by Product Sigma-Algebra. Since π(Y×Z)=1, every value of π lies in [0,1] by Basic Properties of a Measure §monotone; the same holds for every probability measure used below. "Measurable" for a real-valued function on Y means measurable with respect to Y and B(R), and "a.e." and "null" refer to μ.
Part 1: existence.
Step 0 (the second marginal). The projection (y,z)↦z is measurable with respect to Y⊗B(Z) and B(Z) by claim 5 of Sections of Product-Measurable Sets and Maps Are Measurable, and Insertion Maps into Products Are Measurable, so by claim 1 of Image Measures, Measures with Densities, and Change of Variables the formula ρ(B)=π(Y×B), for B∈B(Z), defines a measure ρ on (Z,B(Z)) with ρ(Z)=1; it is a Borel measure on (Z,d). Moreover Z=∅, since otherwise Y×Z=∅ and π(Y×Z)=0.
Step 1 (an increasing sequence of compact sets). For each m∈N, Ulam's Theorem: a Finite Borel Measure on a Complete Separable Metric Space is Tight §tight, applied to the finite Borel measure ρ on the complete separable metric space (Z,d) with ε=1/m, shows that the set Am of the compact sets L⊆Z with ρ(Z∖L)≤1/m is nonempty. By Axiom of Countable Choice, applied to the family (Am)m∈N of nonempty subsets of the set of all subsets of Z, there is a sequence (Lm)m∈N with Lm∈Am for every m; that is, each Lm is compact and ρ(Z∖Lm)≤1/m. Put Km=L1∪⋯∪Lm. Each Km is compact: if (Oi)i∈I is a family of open sets with Km⊆⋃i∈IOi, then for each j≤m the criterion Compact Subset Criterion via Open Covers in the Ambient Space gives a finite Jj⊆I with Lj⊆⋃i∈JjOi; the set J=J1∪⋯∪Jm is finite by claim 3 of Peeling an Element off a Finite Set, and Unions of Finite Sets and induction, and Km⊆⋃i∈JOi, so Km is compact by the same criterion. By Compact Subsets of a Metric Space are Closed and Borel §closed and Compact Subsets of a Metric Space are Closed and Borel §borel each Km is closed and belongs to B(Z). Clearly Km⊆Km+1, and ρ(Z∖Km)≤ρ(Z∖Lm)≤1/m by Basic Properties of a Measure §monotone. Let K∞=⋃mKm∈B(Z). For every m we have ρ(Z∖K∞)≤ρ(Z∖Km)≤1/m, so ρ(Z∖K∞)=0, since a positive real number exceeds some 1/m by claim 3 of The Archimedean Property of the Real Numbers. Consequently, by Basic Properties of a Measure §monotone,
π(A×(Z∖K∞))≤π(Y×(Z∖K∞))=ρ(Z∖K∞)=0(A∈Y).(1)
Step 2 (a countable algebra generating B(Z)). Since (Z,d) is separable, fix a countable set D⊆Z that is dense in (Z,Td). It is nonempty: pick z∈Z (Step 0); then z lies in the closure of D, which is Z, and Z is an open set containing z, so Z∩D=∅ by Closure of a Subset of a Topological Space. Hence ((Z,d),D) is a separable metric datum in the sense of Borel Sets and Measurable Maps in a Separable Metric Space; let E be the family of the open balls Bd(q,s) with q∈D and s∈Q>0, which is countable by claim 1 of that lemma. Let Gcl be the family of the closed balls Bˉd(q,s) with q∈D and s∈Q>0. The set Q>0 is countable by claim 2 of The Integers and the Rational Numbers are Countable and claim 3 of Basic Properties of Countable Sets, so D×Q>0 is countable by claim 1 of Products and Powers of Countable Sets, and Gcl, its image under (q,s)↦Bˉd(q,s), is countable by claim 4 of Basic Properties of Countable Sets. The family Gcp={Km:m∈N}, the image of N under m↦Km, is countable by claims 1 and 4 of Basic Properties of Countable Sets. Let G=E∪Gcl∪Gcp. Putting X1=E, X2=Gcl and Xi=Gcp for i≥3, we obtain a family (Xi)i∈N of countable subsets of the set of all subsets of Z whose union is G, so G is countable by A Countable Union of Countable Sets is Countable. It is nonempty, as K1∈G, so by Countable Set there is a sequence (Gi)i∈N whose set of terms is G.
For a set G⊆Z write G1=G and G0=Z∖G. For n∈N and ϵ=(ϵ1,…,ϵn)∈{0,1}n let Pϵ=G1ϵ1∩⋯∩Gnϵn. For z∈Z let ϵn(z)∈{0,1}n have i-th entry 1 exactly when z∈Gi; then z∈Pϵ if and only if ϵ=ϵn(z), so the sets Pϵ, ϵ∈{0,1}n, are pairwise disjoint and cover Z. Let Un be the family of the sets ⋃ϵ∈SPϵ with S⊆{0,1}n (the empty union being ∅). It is finite, being the image of the finite set of subsets of {0,1}n. It contains Z (take S={0,1}n); the complement of ⋃ϵ∈SPϵ is ⋃ϵ∈/SPϵ, and the union of the members given by S and S′ is the member given by S∪S′. Since Pϵ=P(ϵ,1)∪P(ϵ,0) with (ϵ,1),(ϵ,0)∈{0,1}n+1, we have Un⊆Un+1. Let U=⋃nUn; it is countable by claim 2 of Basic Properties of Countable Sets and A Countable Union of Countable Sets is Countable. It contains Z, so it is nonempty, and by Countable Set we fix a sequence (Uj)j∈N whose set of terms is U; for B∈U let jB be the least j∈N with Uj=B. The family U is closed under complements, and if B∈Un and B′∈Un′ then both lie in Umax(n,n′), so B∪B′∈U. Hence U is also closed under B∩B′=Z∖((Z∖B)∪(Z∖B′)) and under B′∖B=B′∩(Z∖B); in particular U is a π-system in the sense of Dynkin's Pi-Lambda Theorem, and ∅∈U. Each Gi is the union of the Pϵ with ϵ∈{0,1}i and ϵi=1, so G⊆U.
We show that the σ-algebra generated by U, denoted σ(U), is B(Z). Open balls are open by Open Ball in a Metric Space is Open, closed balls are closed by claim 3 of Elementary Properties of the Closed Ball in a Metric Space, and open and closed sets belong to B(Z) by claim 1 of Borel Measurability and Bounded Integration on a Metric Space; with Step 1, G⊆B(Z). As B(Z) is a σ-algebra, closed under complements and finite unions and intersections (Sigma-Algebra and Measurable Space), every Pϵ and every member of U belongs to B(Z), so σ(U)⊆B(Z) by claim 2 of Intersections of Sigma-Algebras and Minimality of the Generated Sigma-Algebra. Conversely E⊆U⊆σ(U), so by the same claim and claim 2 of Borel Sets and Measurable Maps in a Separable Metric Space, B(Z)=σ(E)⊆σ(U).
Step 3 (inner approximation by closed and by compact members of U). Call a set B⊆Z closed-approximable if there is a sequence (Ck)k∈N of closed members of U with Ck⊆Ck+1 for every k and ⋃kCk=B. We show that every member of U is closed-approximable.
(a) Let G± consist of the members of G and their complements; all lie in U. Each is closed or open: the sets Bˉd(q,s) and Km are closed, Bd(q,s) is open, Z∖Bd(q,s) is closed and Z∖Bˉd(q,s), Z∖Km are open, by Closed Subset of a Topological Space. A closed member B is closed-approximable by the constant sequence Ck=B. Let V∈U be open and let KV be the family of the members of Gcl contained in V. We claim ⋃KV=V. One inclusion is clear. Let z∈V. By claim 1 of Borel Sets and Measurable Maps in a Separable Metric Space there are p∈D and t∈Q>0 with z∈Bd(p,t)⊆V. As t−d(p,z)>0, claim 2 of The Archimedean Property of the Real Numbers gives n∈N with t<n(t−d(p,z)); then n≥2 (for n=1 this would say d(p,z)<0), and d(p,z)<t(n−1)/n=:t′. Now t′∈Q>0 and t′<t, so Bˉd(p,t′)⊆Bd(p,t)⊆V; thus Bˉd(p,t′)∈KV and z∈Bˉd(p,t′). If V=∅, take Ck=∅. Otherwise KV is nonempty and countable (claim 3 of Basic Properties of Countable Sets), so there is a sequence (Fj)j∈N whose set of terms is KV, and Ck=F1∪⋯∪Fk is closed by claim 2 of Complements, Unions and Intersections of Closed Sets in a Topological Space, belongs to U, increases with k and has union V. So every member of G± is closed-approximable.
(b) If B,B′ are closed-approximable through (Ck) and (Ck′), then B∪B′ and B∩B′ are closed-approximable through (Ck∪Ck′) and (Ck∩Ck′): these are closed members of U by claims 2 and 3 of Complements, Unions and Intersections of Closed Sets in a Topological Space, they increase with k, every point of B∪B′ lies in some Ck or some Ck′, and a point of B∩B′ lying in Ck and in Ck′′ lies in Ck′′∩Ck′′′ for k′′=max(k,k′); the reverse inclusions are clear.
(c) The empty set is closed-approximable by the constant sequence ∅, which is closed by claim 1 of Complements, Unions and Intersections of Closed Sets in a Topological Space. By (a), (b) and induction, every Pϵ, being a finite intersection of members of G±, and every finite union of such sets is closed-approximable; that is, every member of U is.
For j∈N let Cj be the set of the sequences (Ck)k∈N of closed members of U with Ck⊆Ck+1 for every k and ⋃kCk=Uj; it is nonempty, as Uj is closed-approximable. By Axiom of Countable Choice, applied to the family (Cj)j∈N of nonempty subsets of the set of all sequences in U, there is a sequence (cj)j∈N with cj∈Cj for every j. For B∈U let (CB,k)k∈N be the sequence cjB, which belongs to CjB with UjB=B, and put EB,k=CB,k∩Kk. Then EB,k∈U, EB,k is compact by claim 1 of Compactness of Intersections with Closed Sets and of Level Sets of Semicontinuous Functions, EB,k⊆EB,k+1, EB,k⊆B, and ⋃kEB,k=B∩K∞: a point of B∩K∞ lies in some CB,k and some Kk′, hence in EB,k′′ for k′′=max(k,k′).
Step 4 (densities with values in [0,1]). For B∈B(Z) define λB(A)=π(A×B) for A∈Y. Then λB(∅)=0, and for pairwise disjoint Am∈Y the rectangles Am×B are pairwise disjoint with union (⋃mAm)×B, so λB is a measure on (Y,Y). As A×B⊆A×Z, Basic Properties of a Measure §monotone gives λB(A)≤μ(A); so λB is finite and λB(A)=0 whenever μ(A)=0. The measure μ is σ-finite, as μ(Y)=1 (Measure, Measure Space, and Probability Measure).
Let B∈U. By The Radon-Nikodym Theorem for a Finite Measure and a Sigma-Finite Measure, and Uniqueness of Densities §existence there is a measurable h~B:Y→R with h~B≥0 and λB(A)=∫Y1Ah~Bdμ for every A∈Y. For n∈N let tn=1+1/n and An={y∈Y:tn≤h~B(y)}, which belongs to Y by The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §markov. The function 1Anh~B is measurable by claims 1 and 3 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions, it is nonnegative, and the set where it is at least tn is An, since tn>0. By The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §markov and Basic Properties of a Measure §monotone,
tnμ(An)≤∫Y1Anh~Bdμ=λB(An)≤μ(An),
and as μ(An)≤1 is real this gives μ(An)/n≤0, so μ(An)=0. If h~B(y)>1, claim 3 of The Archimedean Property of the Real Numbers gives n with 1/n<h~B(y)−1, so y∈An; hence {h~B>1} is contained in the null set ⋃nAn (The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §null-union). Let hB=min(h~B,1), measurable by claims 1 and 4 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions, with values in [0,1] and equal to h~B off ⋃nAn. For A∈Y the nonnegative measurable functions 1AhB and 1Ah~B agree a.e., so by The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §comparison
π(A×B)=λB(A)=∫Y1AhBdμ(A∈Y, B∈U),(2)
that is, hB is a density of λB with respect to μ. Consequently, for every j∈N the set Hj of the measurable functions h:Y→[0,1] with π(A×Uj)=∫Y1Ahdμ for every A∈Y is nonempty. By Axiom of Countable Choice, applied to the family (Hj)j∈N of nonempty subsets of the set of all functions from Y to R, there is a sequence (fj)j∈N with fj∈Hj for every j. From now on, for B∈U, hB denotes fjB; since UjB=B, it is a measurable function Y→[0,1] satisfying (2).
Step 5 (countably many identities almost everywhere).
(5.1) The constant function 1 is measurable by claim 1 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions, and ∫Y1A⋅1dμ=μ(A)=π(A×Z)=λZ(A) by The Integral of an Indicator Function is the Measure of the Set. So 1 and hZ are densities of λZ, and by The Radon-Nikodym Theorem for a Finite Measure and a Sigma-Finite Measure, and Uniqueness of Densities §uniqueness the set NZ={y:hZ(y)=1} belongs to Y and μ(NZ)=0.
(5.2) Let P be the set of pairs (B,B′)∈U×U with B∩B′=∅; it is countable by claim 1 of Products and Powers of Countable Sets and claim 3 of Basic Properties of Countable Sets. For (B,B′)∈P we have B∪B′∈U, the function hB+hB′ is measurable (claim 2 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions) and nonnegative, and for A∈Y, by Linearity and Monotonicity of the Lebesgue Integral §nonnegative, (2) and Basic Properties of a Measure §additivity applied to the disjoint sets A×B and A×B′,
∫Y1A(hB+hB′)dμ=∫Y1AhBdμ+∫Y1AhB′dμ=π(A×B)+π(A×B′)=π(A×(B∪B′)).
So hB+hB′ and hB∪B′ are densities of λB∪B′, and by The Radon-Nikodym Theorem for a Finite Measure and a Sigma-Finite Measure, and Uniqueness of Densities §uniqueness the set NB,B′={y:hB∪B′(y)=hB(y)+hB′(y)} belongs to Y and has μ(NB,B′)=0. The set P is nonempty, as (∅,∅)∈P, so by Countable Set there is a sequence (pj)j∈N whose set of terms is P. Write Np=NB,B′ for p=(B,B′)∈P, and let Nadd=⋃j∈NNpj, which is the union of the NB,B′ over (B,B′)∈P. Each Npj belongs to Y, so Nadd∈Y because Y is a σ-algebra (Sigma-Algebra and Measurable Space), and μ(Nadd)≤∑j∈Nμ(Npj)=0 by Basic Properties of a Measure §subadditivity; so μ(Nadd)=0, and in particular Nadd is null. If y∈/Nadd and B⊆B′ are members of U, then (B,B′∖B)∈P with union B′, so
hB′(y)=hB(y)+hB′∖B(y)≥hB(y).(3)
(5.3) Let B∈U. Let gB(y)=supkhEB,k(y); it takes values in [0,1] and is measurable by claim 1 of Measurability of Countable Suprema, Bounded Pointwise Limits, Monotone Functions, and Continuous Functions, applied with the bound 1. For k∈N let sB,k=max(hEB,1,…,hEB,k), measurable by claim 4 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions and induction on k. At every y the sequence (sB,k(y))k is nondecreasing with supremum gB(y), and for y∈/Nadd, (3) applied to EB,i⊆EB,k for i≤k gives sB,k(y)=hEB,k(y). Let A∈Y. The measurable functions 1AsB,k (claims 1 and 3 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions) are nondecreasing in k at every point, with supremum 1AgB, so by Monotone Convergence Theorem
∫Y1AgBdμ=ksup∫Y1AsB,kdμ=ksup∫Y1AhEB,kdμ=ksupπ(A×EB,k),
where the second equality is The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §comparison, the two integrands agreeing off the null set Nadd, and the third is (2). The sets A×EB,k increase with k and have union A×(B∩K∞) by Step 3, so by Basic Properties of a Measure §continuity-below the last supremum is π(A×(B∩K∞)). By Basic Properties of a Measure §additivity, Basic Properties of a Measure §monotone and (1), π(A×B)=π(A×(B∩K∞))+π(A×(B∖K∞))=π(A×(B∩K∞)). Hence ∫Y1AgBdμ=λB(A) for every A∈Y, so gB and hB are densities of λB, and by The Radon-Nikodym Theorem for a Finite Measure and a Sigma-Finite Measure, and Uniqueness of Densities §uniqueness the set NB={y:gB(y)=hB(y)} belongs to Y and has μ(NB)=0.
Let N0=NZ∪Nadd∪⋃j∈NNUj; since the terms of (Uj) are exactly the members of U, NB⊆N0 for every B∈U. The set N0 is the union of the sequence NZ,Nadd,NU1,NU2,… of members of Y of μ-measure 0, so N0∈Y because Y is a σ-algebra (Sigma-Algebra and Measurable Space), and μ(N0)=0 by Basic Properties of a Measure §subadditivity.
Step 6 (a premeasure on U for each y∈/N0). Fix y∈Y∖N0 and let τy(B)=hB(y)∈[0,1] for B∈U.
(6.1) τy(Z)=1, as y∈/NZ.
(6.2) τy(B∪B′)=τy(B)+τy(B′) for disjoint B,B′∈U, as y∈/Nadd. Taking B=B′=∅ gives τy(∅)=2τy(∅), so τy(∅)=0.
(6.3) If B⊆B′ in U, then τy(B′∖B)=τy(B′)−τy(B) and τy(B)≤τy(B′), by (6.2).
(6.4) For B1,…,Br∈U, τy(B1∪⋯∪Br)≤τy(B1)+⋯+τy(Br): for r=2, B1∪B2 is the disjoint union of B1 and B2∖B1, so (6.2) and (6.3) give τy(B1∪B2)=τy(B1)+τy(B2∖B1)≤τy(B1)+τy(B2); the general case follows by induction on r.
(6.5) For B∈U, τy(B)=hB(y)=gB(y)=supkτy(EB,k), as y∈/NB. Hence for every real η>0 there is k with τy(B)−τy(EB,k)≤η, since otherwise τy(B)−η would be an upper bound of the τy(EB,k) smaller than their least upper bound.
(6.6) Continuity at ∅. Let (Rn)n∈N be a sequence in U with Rn+1⊆Rn for every n and ⋂nRn=∅. We show that τy(Rn) converges to 0. By (6.3) the sequence (τy(Rn)) is nonincreasing and bounded below by 0; let δ≥0 be its greatest lower bound, and suppose, for a contradiction, that δ>0. Then τy(Rn)≥δ for every n. With δ fixed, for each n let kn be the least k∈N with τy(Rn)−τy(ERn,k)≤δ2−n−1, which exists by (6.5). Put Wn=ER1,k1∩⋯∩ERn,kn∈U. Then Wn+1⊆Wn and Wn⊆ERn,kn⊆Rn. Since Rn⊆Ri for i≤n,
Rn∖Wn=i=1⋃n(Rn∖ERi,ki)⊆i=1⋃n(Ri∖ERi,ki),
so by (6.3) and (6.4)
τy(Rn∖Wn)≤i=1∑n(τy(Ri)−τy(ERi,ki))≤i=1∑nδ2−i−1<2δ,
and therefore, by (6.3), τy(Wn)=τy(Rn)−τy(Rn∖Wn)>δ−δ/2>0. By (6.2), Wn=∅ for every n. Each ERi,ki is compact, hence closed by Compact Subsets of a Metric Space are Closed and Borel §closed, so every Wn is closed by claim 3 of Complements, Unions and Intersections of Closed Sets in a Topological Space, and W1=ER1,k1 is compact. The sets Z∖Wn are open, and since ⋂nWn⊆⋂nRn=∅ they cover W1. By Compact Subset Criterion via Open Covers in the Ambient Space there is a finite J⊆N with W1⊆⋃n∈J(Z∖Wn); J is nonempty because W1=∅. Let n∗ be the largest element of J. For n∈J we have Wn∗⊆Wn, so Z∖Wn⊆Z∖Wn∗, and hence W1⊆Z∖Wn∗. But ∅=Wn∗⊆W1, a contradiction. Thus δ=0. Given a real η>0 there is then n0 with τy(Rn0)<η, and 0≤τy(Rn)≤τy(Rn0)<η for n≥n0; so τy(Rn) converges to 0.
(6.7) Countable additivity. Let (Bi)i∈N be pairwise disjoint members of U whose union B belongs to U. Put Rn=B∖(B1∪⋯∪Bn)∈U. Then Rn+1⊆Rn, and ⋂nRn=∅ because each point of B lies in some Bi and hence not in Rn for n≥i. As B is the disjoint union of B1,…,Bn,Rn, (6.2) and induction give ∑i=1nτy(Bi)=τy(B)−τy(Rn), which converges to τy(B) by (6.6). So the series ∑iτy(Bi) converges and its sum is τy(B), in the sense of Series of Real Numbers §convergent.
By Step 2, (6.2) and (6.7), the family U of subsets of Z and the function τy satisfy the hypotheses of Extension of a Finite Premeasure from an Algebra of Sets to the Generated Sigma-Algebra. By Extension of a Finite Premeasure from an Algebra of Sets to the Generated Sigma-Algebra §extension there is a measure on σ(U)=B(Z) that agrees with τy on U, and by Extension of a Finite Premeasure from an Algebra of Sets to the Generated Sigma-Algebra §uniqueness there is only one; we call it κy. By (6.1), κy(Z)=1, so κy is a probability measure on (Z,B(Z)), and
κy(B)=hB(y)(y∈Y∖N0, B∈U).(4)
For y∈N0 put κy=ρ, a probability measure on (Z,B(Z)) by Step 0. Define κ(y,B)=κy(B) for y∈Y and B∈B(Z); its values lie in [0,1].
Step 7 (measurability and the rectangle identity). Let L be the family of those B∈B(Z) such that y↦κ(y,B) is measurable and
π(A×B)=∫Y1A(y)κ(y,B)μ(dy)for every A∈Y.(5)
U⊆L: for B∈U, (4) gives κ(⋅,B)=1Y∖N0hB+ρ(B)1N0, which is measurable by claims 1, 2 and 3 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions, as N0∈Y. For A∈Y the nonnegative measurable functions 1Aκ(⋅,B) and 1AhB agree off the null set N0, so (5) follows from The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §comparison and (2).
L is a λ-system in the sense of Dynkin's Pi-Lambda Theorem. (1) Z∈U⊆L. (2) Let B⊆B′ be members of L. For each y, κy(B′∖B)=κy(B′)−κy(B) by Basic Properties of a Measure §differences, so κ(⋅,B′∖B) is measurable by claim 2 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions. For A∈Y we have 1Aκ(⋅,B′)=1Aκ(⋅,B)+1Aκ(⋅,B′∖B) pointwise, by Basic Properties of a Measure §additivity for each κy, so by Linearity and Monotonicity of the Lebesgue Integral §nonnegative and (5) for B and B′,
∫Y1Aκ(⋅,B′∖B)dμ=π(A×B′)−π(A×B)=π(A×(B′∖B)),
the subtraction being legitimate since π(A×B)≤1 is real, and the last equality being Basic Properties of a Measure §differences for A×B⊆A×B′, whose difference is A×(B′∖B). So B′∖B∈L. (3) Let (Bn)n∈N be a nondecreasing sequence in L with union B. For each y, κ(y,Bn) is nondecreasing in n by Basic Properties of a Measure §monotone, and, its values lying in [0,1], it converges to κ(y,B), which is also its least upper bound, by Basic Properties of a Measure §continuity-below; hence κ(⋅,B) is measurable by claim 5 of Arithmetic, Absolute Values, and Pointwise Limits of Measurable Real-Valued Functions. For A∈Y the measurable functions 1Aκ(⋅,Bn) are nondecreasing in n with supremum 1Aκ(⋅,B), so by Monotone Convergence Theorem, (5) for each Bn, and Basic Properties of a Measure §continuity-below for the increasing sets A×Bn with union A×B,
∫Y1Aκ(⋅,B)dμ=nsup∫Y1Aκ(⋅,Bn)dμ=nsupπ(A×Bn)=π(A×B).
So B∈L.
By Dynkin's Pi-Lambda Theorem, applied to the π-system U and the λ-system L, we get σ(U)⊆L; with Step 2 and L⊆B(Z) this gives L=B(Z).
Step 8 (conclusion of existence). By Step 6 each κy is a probability measure on (Z,B(Z)), and by Step 7 each κ(⋅,B), B∈B(Z), is measurable with respect to Y; so κ is a probability kernel from (Y,Y) to (Z,B(Z)) by Probability Kernels Between Measurable Spaces §kernel. By Step 7, (5) holds for all A∈Y and B∈B(Z), which is the displayed identity of claim 1, and Integration Against a Probability Kernel: Measurable Sections, the Composite Measure on the Product and the Iterated Integral §rectangles, applied to the probability measure μ, the kernel κ and the measure π, gives π=μ⊗κ.
Part 2: uniqueness. Let κ and κ′ be probability kernels from (Y,Y) to (Z,B(Z)) with π=μ⊗κ=μ⊗κ′, and let U be the family fixed in Step 2, with its enumeration (Uj)j∈N; it is a countable π-system that contains Z and satisfies σ(U)=B(Z). Let B∈U. By Integration Against a Probability Kernel: Measurable Sections, the Composite Measure on the Product and the Iterated Integral §composite, for every A∈Y,
λB(A)=π(A×B)=(μ⊗κ)(A×B)=∫Y1A(y)κ(y,B)μ(dy),
and likewise with κ′. By Probability Kernels Between Measurable Spaces §kernel, κ(⋅,B) and κ′(⋅,B) are measurable with values in [0,1], so both are densities of the finite measure λB (Step 4) with respect to μ, and by The Radon-Nikodym Theorem for a Finite Measure and a Sigma-Finite Measure, and Uniqueness of Densities §uniqueness the set MB={y:κ(y,B)=κ′(y,B)} belongs to Y and has μ(MB)=0. Let M=⋃j∈NMUj, which is the union of the MB over B∈U, since the terms of (Uj) are exactly the members of U. Each MUj belongs to Y, so M∈Y because Y is a σ-algebra (Sigma-Algebra and Measurable Space), and μ(M)≤∑j∈Nμ(MUj)=0 by Basic Properties of a Measure §subadditivity, so μ(M)=0.
We claim that M is exactly the set of y with κy=κy′. If y∈M, then κy(B)=κy′(B) for some B, so κy=κy′. If y∈/M, then κy and κy′ are measures on (Z,B(Z)) with κy(Z)=κy′(Z)=1<∞ that agree on the π-system U generating B(Z), so κy=κy′ by claim 1 of Uniqueness of Finite Measures on a Generating Pi-System and the Density of the Exponential Law. Hence the set of y with κy=κy′ is M, which belongs to Y and has μ-measure 0. ■