4 Applications 477
B=[a−m,...,am]∪···∪[a′−m,...,a′m],
we define
μm(B)=pa−m···pam+···+pa′−m···pam′.
Thenμm(X)=1,μm(B)≥0foreveryB∈Cm,and
μm(B∪C)=μm(B)+μm(C)ifB,C∈CmandB∩C=∅.
Every union of cylinder sets of ordermis also a union of cylinder sets of order
m+1, since
[a−m,...,am]=∪a,a′∈A[a,a−m,...,am,a′].
ThusCm⊆Cm+ 1. Moreoverμm+ 1 continuesμm,since
μm+ 1 ([a−m,...,am])=
∑r
j,j′= 1
pjpj′pa−m···pam
=μm([a−m,...,am])
(∑r
j= 1
pj
)(∑r
j′= 1
pj′
)
=μm([a−m,...,am]).
Letμdenote the continuation of allμmtoC=C 0 ∪C 1 ∪....IfB,C∈C,then
B,C∈Cmfor somem. Hence, for givenC∈C, there are only finitely many distinct
B∈Csuch thatB⊆C. Consequently, ifCis the union of a sequence of disjoint sets
Cn ∈ C(n = 1 , 2 ,...),thenCn =∅for all largenandμ(C)=
∑
n≥ 1 μ(Cn).
It follows, by a construction due to Carath ́eodory (1914), thatμcan be uniquely
extended to theσ-algebraBof subsets ofXgenerated byCso that (X,B,μ)isa
probability space. For anyε>0 there exists, for eachB∈B,someC∈Csuch that
μ(B∆C)<ε.
Thetwo-sided Bernoulli shift Bp 1 ,...,pris the mapσ:X→Xdefined byσx=x′,
wherexi′=xi+ 1 for everyi∈Z. It is a measure-preserving transformation of the
probability space (X,B,μ), since
σ−^1 [a−m,...,am]=∪a,a′∈A[a,a′,a−m,...,am]
and hence
μ(σ−^1 [a−m,...,am])=
∑r
j,j′= 1
pjpj′pa−m···pam
=
∑r
j,j′= 1
pjpj′μ([a−m,...,am])=μ([a−m,...,am]).
The Bernoulli shiftB 1 / 2 , 1 / 2 is a model for the random process consisting of bi-infinite
sequences of coin-tossings.