Throughout, integrals of [0,∞]-valued measurable functions are those of Lebesgue Integral of a Nonnegative Measurable Function, with the [0,∞] conventions of the statement; π-systems and λ-systems are those of Dynkin's Pi-Lambda Theorem; generated σ-algebras are those of Generated Sigma-Algebra; and σ-algebra manipulations use Sigma-Algebra and Measurable Space. Write Rl for the family of Borel rectangles in Rl. We use two elementary consequences of countable additivity (Measure, Measure Space, and Probability Measure) for a measure m: finite additivity and monotonicity — padding a finite disjoint family with copies of ∅ gives m(U∪V)=m(U)+m(V) for disjoint U,V, whence m(U)≤m(V) for U⊆V, as m(V)=m(U)+m(V∖U) — and continuity from below: if E1⊆E2⊆…, then m(⋃jEj)=limjm(Ej), since the union is the disjoint union of the successive differences Ej∖Ej−1 (with E0=∅) and the partial sums of their masses are the m(Ej).
Step 1 (a uniqueness principle). We first record: let Z be a set, P a π-system of subsets of Z, and m, m′ measures on σ(P) with m(C)=m′(C) for every C∈P; suppose there are pairwise disjoint Z1,Z2,⋯∈P with ⋃nZn=Z and m(Zn)<∞ for every n. Then m=m′.
Indeed, fix n and let Dn be the class of E∈σ(P) with m(E∩Zn)=m′(E∩Zn). It contains Z, both sides being the common finite value m(Zn)=m′(Zn) (as Zn∈P). It is closed under proper differences: for F⊆E both in Dn, finite additivity gives m(E∩Zn)=m((E∖F)∩Zn)+m(F∩Zn), all terms being finite by monotonicity (each is at most m(Zn)<∞), so m((E∖F)∩Zn)=m(E∩Zn)−m(F∩Zn), and likewise for m′. It is closed under increasing countable unions by continuity from below. Hence Dn is a λ-system; it contains P, since for E∈P the set E∩Zn again lies in the π-system P, on which m and m′ agree; so Dn⊇σ(P) by Dynkin's Pi-Lambda Theorem. Finally, for E∈σ(P), countable additivity along the pairwise disjoint sets E∩Zn gives m(E)=∑nm(E∩Zn)=∑nm′(E∩Zn)=m′(E).
Step 2 (claim 1). Induction on l. For l=1 the Borel rectangles are exactly the Borel sets, so generation is trivial; λ is σ-finite by Existence of Lebesgue Measure on the Real Line; and the rectangle values prescribe the measure on every Borel set, so uniqueness is immediate. Let l≥2 and assume claim 1 for l−1.
Generation. Every Borel rectangle A1×⋯×Al=(A1×⋯×Al−1)×Al is a measurable rectangle of Bl−1⊗B(R) (Product Sigma-Algebra), since A1×⋯×Al−1∈Bl−1 by the induction hypothesis; hence σ(Rl)⊆Bl. Conversely, fix A∈B(R) and let CA be the class of all H⊆Rl−1 with H×A∈σ(Rl). Then Rl−1∈CA, as Rl−1×A∈Rl; CA is closed under complements, since (Rl−1∖H)×A=(Rl−1×A)∖(H×A) and σ-algebras are closed under set differences; and it is closed under countable unions, since (⋃mHm)×A=⋃m(Hm×A). So CA is a σ-algebra containing Rl−1, hence CA⊇Bl−1 by the induction hypothesis. Thus every measurable rectangle H×A generating Bl−1⊗B(R) lies in σ(Rl), and Bl⊆σ(Rl).
Rectangle values and σ-finiteness. By Existence and Uniqueness of the Product Measure, λl=λl−1⊗λ is a σ-finite measure on Bl with λl((A1×⋯×Al−1)×Al)=λl−1(A1×⋯×Al−1)λ(Al), which equals λ(A1)⋯λ(Al) by the induction hypothesis.
Uniqueness. Rl is a π-system, intersections being formed coordinatewise: (A1×⋯×Al)∩(A1′×⋯×Al′)=(A1∩A1′)×⋯×(Al∩Al′). For each l-tuple q=(q1,…,ql) of integers put Zq=[q1,q1+1)×⋯×[ql,ql+1); each factor is an interval, hence a Borel set of Lebesgue measure 1 by Existence of Lebesgue Measure on the Real Line, so the Zq are countably many pairwise disjoint members of Rl covering Rl with λl(Zq)=1 by the rectangle values. Any measure on Bl with the stated rectangle values agrees with λl on Rl, so Step 1 gives uniqueness.
Step 3 (claim 2 for l≥2: generation and bimeasurability). Let l≥2, (Y,G,μ) and i be as in claim 2, and write Pl′ for the family of sets R×G with R∈Rl′ and G∈G.
(a) For every l′≥1, Bl′⊗G is generated by Pl′: each R×G is a measurable rectangle of Bl′⊗G, since R∈σ(Rl′)=Bl′ by claim 1; conversely, for fixed G∈G, the class of H⊆Rl′ with H×G∈σ(Pl′) is a σ-algebra (it contains Rl′ as Rl′×G∈Pl′, and is closed under complements and countable unions exactly as in Step 2) containing Rl′, hence containing Bl′ by claim 1; so every generator H×G of Bl′⊗G lies in σ(Pl′). This proves in particular the generation assertion of claim 2.
(b) B(R)⊗(Bl−1⊗G) is generated by the family Q of sets A×(R′×G) with A∈B(R), R′∈Rl−1, G∈G: each such set is a measurable rectangle, since R′×G∈Bl−1⊗G; conversely, for fixed A∈B(R), the class of K⊆Rl−1×Y with A×K∈σ(Q) is a σ-algebra (it contains Rl−1×Y since A×(Rl−1×Y)∈Q, and is closed under complements and countable unions as before) containing the sets R′×G, hence containing Bl−1⊗G by (a) applied with l′=l−1; so every generator A×K lies in σ(Q).
Ψi is a bijection; its inverse Ξ sends (θ,y) to (θi,((θ1,…,θi−1,θi+1,…,θl),y)). For R=A1×⋯×Al∈Rl and G∈G,
Ψi−1(R×G)=Ai×((A1×⋯×Ai−1×Ai+1×⋯×Al)×G)∈Q,
and for A∈B(R), R′=A1′×⋯×Al−1′∈Rl−1, G∈G,
Ξ−1(A×(R′×G))=(A1′×⋯×Ai−1′×A×Ai′×⋯×Al−1′)×G∈Pl.
By the generator criterion of Measurable Function and Real-Valued Measurable Function together with (a) and (b), both Ψi and Ξ are measurable.
Step 4 (claim 2 for l≥2: the image measure). The measures λl−1⊗μ, λl⊗μ and λ⊗(λl−1⊗μ) are defined and σ-finite by Existence and Uniqueness of the Product Measure, their factors being σ-finite by claim 1 and by hypothesis. Let m be the image measure of λ⊗(λl−1⊗μ) under Ψi, a measure on Bl⊗G by Step 3 and claim 1 of Image Measures, Measures with Densities, and Change of Variables. For R=A1×⋯×Al∈Rl and G∈G, the preimage computed in Step 3 and the rectangle property of Existence and Uniqueness of the Product Measure, applied twice, give
m(R×G)=λ(Ai)⋅λl−1(A1×⋯×Ai−1×Ai+1×⋯×Al)⋅μ(G)=λ(A1)⋯λ(Al)μ(G)=(λl⊗μ)(R×G),
using the rectangle values of claim 1 for λl−1 and for λl. The family Pl is a π-system (coordinatewise intersections). By σ-finiteness of μ there are Y1′′,Y2′′,⋯∈G covering Y with μ(Yn′′)<∞; setting Y1=Y1′′ and Yn=Yn′′∖(Y1′′∪⋯∪Yn−1′′) yields pairwise disjoint members of G with ⋃nYn=Y and μ(Yn)≤μ(Yn′′)<∞ by monotonicity. The countably many sets Zq×Yn (with Zq as in Step 2) are pairwise disjoint members of Pl covering Rl×Y, with (λl⊗μ)(Zq×Yn)=μ(Yn)<∞. Step 1 applied to m and λl⊗μ gives m=λl⊗μ.
Step 5 (claims 3 and 4 for l≥2). Write m0=λ⊗(λl−1⊗μ). Let f:Rl×Y→[0,∞] be Bl⊗G-measurable. Then f∘Ψi is measurable, since for real a one has {f∘Ψi>a}=Ψi−1({f>a}), measurable by Step 3; and claim 2 of Image Measures, Measures with Densities, and Change of Variables, applied to T=Ψi whose image measure is λl⊗μ by Step 4, gives
∫Rl×Yfd(λl⊗μ)=∫f∘Ψidm0.
The Sections clause of Tonelli and Fubini Theorems, applied to the σ-finite factors (R,B(R),λ) and (Rl−1×Y,Bl−1⊗G,λl−1⊗μ), shows that for every (θ′,y) the section t↦(f∘Ψi)(t,(θ′,y))=f(Ψi(t,(θ′,y))) is B(R)-measurable, and the Tonelli clause — the second iterated integral of its display, which integrates over the first factor innermost — shows that (θ′,y)↦∫Rf(Ψi(t,(θ′,y)))dλ(t) is Bl−1⊗G-measurable with
∫f∘Ψidm0=∫Rl−1×Y(∫Rf(Ψi(t,(θ′,y)))dλ(t))d(λl−1⊗μ)(θ′,y).
Chaining the two displays proves claim 3. For claim 4, let f:Rl×Y→R be integrable with respect to λl⊗μ; by the integrable clause of claim 2 of Image Measures, Measures with Densities, and Change of Variables, f∘Ψi is integrable with respect to m0 with the same integral, and the Fubini clause of Tonelli and Fubini Theorems — its symmetric statement in the other order, whose exceptional set lies in the second factor — provides exactly the set N∈Bl−1⊗G of (λl−1⊗μ)-measure zero and the integrability statements and identity asserted in claim 4.
Step 6 (claims 2, 3, and 4 for l=1). Under the conventions of the statement: the generation assertion of claim 2 is that B(R)⊗G is generated by the measurable rectangles A×G, which is Product Sigma-Algebra directly; Ψ1 is the identity map of R×Y, trivially a bijection that is measurable with measurable inverse, and the image measure of λ⊗μ under it is λ⊗μ=λ1⊗μ itself (its value on each B∈B(R)⊗G being unchanged); and claims 3 and 4 are precisely the Sections, Tonelli, and symmetric Fubini clauses of Tonelli and Fubini Theorems applied to the σ-finite factors (R,B(R),λ) and (Y,G,μ), in the order that integrates over R innermost. ■