Conditional Measures and Algebras
5.4 Algebras and Maps
Moreover,
[x]A =φ−1(φ(x))
for x /∈ N, and μAx = νφ(x) for some measurable map y → νy defined on a φ∗μ-conull subset of Y. In fact we can take Y = M(X), φ(x) = μAx , andνy =y.
This will be proved later; the conclusion described in Corollary 5.22 is depicted in Fig.5.3.
Fig. 5.3 Eachy∈Y determines an atomφ−1(y) and its conditional measureνy
Lemma 5.23.If X is a compact metric space, and f ∈ L∞(X), then the map
M(X)ν −→
fdν
is Borel measurable. In particular, for a Borel subset X of X, we have that M(X)is a Borel subset of M(X). Moreover, if φ:X →Y is a Borel measurable map between Borel subsets of compact metric spaces, then the induced mapφ∗:M(X)→M(Y)is Borel measurable.
Proof. Starting with continuous functions, we know that
fdν depends continuously onν (by definition of the weak*-topology on M(X)). Arguing just as we did on p.139, this can be extended to show that
fdν depends measurably onν for all indicator functions of open sets, and thence to show that it does so for indicator functions of Borel measurable sets, and finally for anyf ∈L∞. By definition and the argument above, it follows that
M(X) ={μ∈M(X)|μ(XX) = 0}
5.4 Algebras and Maps 147
is Borel measurable. Now letφ:X →Y be measurable and fixr∈R,ε >0 andf ∈C(Y). Then
Of,r,ε =
μ∈M(Y)|fdμ−r< ε is an open set inM(Y), and clearly
φ−∗1Of,r,ε=
ν∈M(X)|f◦φdν−r< ε
is measurable in M(X). Since any open set in M(Y) can be written as a countable union of finite intersections of sets of the formOf,r,εwithf chosen from a dense countable subset ofC(X),r∈Qandε∈Q, the lemma follows.
Proof of Corollary 5.22. Let A = σ({A1, A2, . . .}) be countably- generated. TakingY =M(X) (with the weak*-topology, so thatY is a com- pact metric space) andφ(x) =μAx we can set νy =y and henceνφ(x)=μAx follows at once. LetX be aμ-conull set on which all the statements in The- orem5.14hold. We claim first that (perhaps after enlarging the complement ofX by null sets countably often),A =φ−1BY. By Theorem5.14(2),
χAi(x) =μAx(Ai) (5.12) for almost everyx, and we may assume that this holds for allx∈X. Since {ν | ν(Ai) = 1} ∈BY, (5.12) shows that Ai∩X ∈ φ−1BY and therefore A|X ⊆φ−1BY.
For the reverse directionφ−1BY ⊆A|X, it is sufficient to check this on sets of the formOf,r,εsince these generate the weak*-topology in a countable manner, and by Theorem5.14(1) the set
φ−1 {ν| |
fdν−r|< ε}
={x| |
fdμAx −r|< ε}
isA-measurable for anyf ∈C(X),r∈Randε >0. HenceA|X =φ−1B. Sinceφ:X →Y satisfies A|X =φ−1BY, and the Borelσ-algebraBY
ofY separates points, [x]A =φ−1(φ(x)) follows.
Corollary 5.24.Let φ: (X,BX, μ) → (Y,BY, ν) be a measure-preserving map between Borel probability spaces, and let A ⊆BY be a sub-σ-algebra.
Then
φ∗μφx−1A =νφ(x)A forμ-almost everyx∈X.
Proof.First notice that for anyf ∈L1(Y,BY, ν), Eν(fA)◦φis φ−1A- measurable and
φ−1A
Eν(fA)◦φdμ=
A
Eν(fA) dν
=
A
fdν
=
φ−1A
f◦φdμ.
It follows that
Eν(fA)◦φ=Eμ(f◦φφ−1A).
Thus forf ∈L∞(Y,BY),
fdνφ(x)A =Eν(fA) (φ(x))
=Eμ(f ◦φφ−1A)(x)
=
f◦φdμφx−1A
=
fd
φ∗μφx−1A
(all almost everywhere). Using a dense countable subset ofC(X) completes
the proof.
In the next result we will work with measurability on Borel subsets of compact metric spaces, without reference to a particular measure.
Lemma 5.25.Let X, Y, Z be Borel subsets of compact metric spaces X,Y and Z respectively, and let φZ : X → Z and φY : X → Y be measurable maps. Suppose that φZ is φ−Y1(BY)-measurable. Then there is a measurable mapψ:Y →Z with φZ(x) =ψ◦φY(x)on X, as illustrated in Fig.5.4.
Fig. 5.4 The map constructed in Lemma5.25
This is related to Corollary 5.22, in that it allows us to draw the same conclusion that we can write the conditional measure μAx as a measurable functionνy on the image spaceY wheneverA =φ−1BY.
Proof of Lemma5.25.DefineA =φ−Y1(BY), which is countably-generated sinceBY is. Since the Borelσ-algebraBY ofY separates points,
[x]A =φ−Y1(φY(x))
5.4 Algebras and Maps 149
forx∈X. By assumption, φZ isA-measurable, and henceφZ(x) =φZ(x) whenever [x]A = [x]A, or equivalently whenever φY(x) = φY(x). So we could define
ψ(y) =φZ(x)
whenever y = φY(x) for some x ∈ X, and use some fixed z0 ∈ Z to de- fine ψ(y) = z0 for y ∈ YφY(X). However, it is not clear why this should define a measurable function, so instead we will defineψ:Y →Z by a lim- iting process. In order to do this we will cut the target space Z into small metric balls, and ensure that at each finite stage everything is appropriately measurable. For this we start with the additional assumption thatZ=Z is a compact metric space, and later show that this requirement can be removed.
Since Z is compact, there is a sequence (ξn) of finite Borel measurable partitions ofZ with the property that
σ(ξn)⊆σ(ξn+1)
forn1, and for which every element ofξn has diameter less than n1. We will define related partitionsξXn andξnY ofX andY. The first of these is defined by taking pre-images as follows. SinceφZ isA-measurable we get, for anyP ∈ξn, a set
PX=φ−Z1(P)∈A,
and hence a finite partitionξnX={PX |P ∈ξn} ⊆A ofX. This is illustrated in Fig.5.5.
Fig. 5.5 The partitionsξnonM(X) andξnXonX
We next define the partition ξnY. Note that in the construction of ξnY care needs to be taken, since the set φY(X) is not assumed to be mea-
surable. Since A = φ−Y1BY, we can choose for every PX ∈ ξXn a set PY in BY with PX =φ−Y1PY. Since the various sets PX ∈ξnX are disjoint, we can require the same disjointness for the sets PY. Indeed if sets P and Q in ξn are different, PX = φ−Z1P = φ−Y1P, and QX = φ−Z1Q = φ−Y1QY with QY, PY ∈ BY not disjoint, then replacing QY by QYPY will not change its properties but will ensure disjointness fromPY. Moreover, using a similar argument inductively we can also insist that Pn+1Y ⊆ PnY when- everPn+1X ⊆PnXandPn+1X ∈ξn+1X ,PnX∈ξXn. The setsPY ∈BY forP∈ξnX, together with their common complementQn ∈BY, form a partitionξnY ofY. By construction,φ−1Qn=∅and Qn+1⊇Qn for alln1. Write
Q=
m1
Qm
and define ψ(y) to be some fixed element z0 ∈ Z for any y ∈ Q. We de- fine ψ on YQ as the limit of the sequence of functions ψn : YQ → Z defined as follows. For every PY ∈ ξYn, we choose some zP ∈ P ∈ ξn and define ψn(y) = zP for y ∈ PYQ. Clearly ψn : YQ → Z is measurable, since the partitionξnY is measurable by construction. By construction of the partitions ξn, the diameters of the partition elementsP ∈ξn go to zero, so the functionψdefined by
ψ(y) = lim
n→∞ψn(y)
exists for ally ∈YQ; since it is a pointwise limit of a sequence of measur- able functions, the mapψis also measurable. Ify =φY(x) for somex∈X, then φZ(x) and ψn(y) will belong to the same element of ξn for all n, soφZ(x) =ψ(φY(x)) as required.
Assume now thatZ is only a Borel subset of a compact metric space Z.
The construction above can be used to define a measurable mapψ:Y →Z withφZ =ψ◦φY. Since
∅=φ−Z1ZZ
=φ−Y1
ψ−1ZZ we may replaceψby the measurable map
ψ(y) =
ψ(y) ifψ(y)∈Z, and z0∈Z if not,
for some fixed elementz0∈Z without affecting the fact thatφZ =ψ◦φY.