QUESTION:For
any stationary markov process, is the autocorrelation of an interval
the product of the autocorrelations of subintervals?
Summary
A stationary process Zt
has multiplicative autocorrelation when
Cor[Zt,Zr]=Cor[Zt,Zs]Cor[Zs,Zr]
for all t≤s≤r.
Autocorrelation is defined as Cor[Zt,Zs]:=Cov[Zt,Zs]σ2
with σ2=Var(Zt).
A stationary autoregressive process has multiplicative
autocorrelation
1.
However, not all stationary Markov processes have multiplicative
autocorrelation. See the section below about a real-valued 3-state
Markov chain for a counterexample.
Among discrete-time stationary processes, only autoregressive
processes have multiplicative autocorrelation. Some Markov processes
are not obviously autoregressive processes even though technically
they are. For example, all stationary real-valued two-state Markov
chains are autoregressive (and thus also have multiplicative
autocorrelation).
Define what will be shown to be "white noise" of
Zt
as autoregressive process: ϵt:=Zt−ρZt−1
By convenient translation, E[Zt]=0ϵt=0Cov[Zt,Zs]=E[ZtZs]E[Zt2]=σ2E[ZtZt+1]=ρ
Consider any 0]]>n>0.
E[ϵtϵt+n]=E[(Zt−ρZt−1)(Zt+n−ρZt+n−1)]=E[ZtZt+n]+ρ2E[Zt−1Zt+n−1]−ρ(E[ZtZt+n−1]+E[Zt−1Zt+n])=(1+ρ2)ρnσ2−ρ(ρn−1σ2+ρn+1σ2)=0
thus ϵt
satisfies the "white noise" condition for expressing
Zt
as the autoregressive process Zt+1=ρZt+ϵt
QED
Real-valued 2-state Markov chain
For any stationary two-state Markov chain
1Zt,
Cor[Zt,Z0]=Cor[Zt,Zs]Cor[Zs,Z0]
Proof Let q1:=P(Zt=a1)q0:=P(Zt=a0)
Map Zt
to a more convenient Yt:=Zt−a0a1−a0
Since Yt
only equals 0
or 1:
E[Yt]=E[Yt2]=q1
and thus Var(Yt)=q1−q12=q1q0
For convenience let p0:=P(Y1=0∣Y0=1)p1:=P(Y1=1∣Y0=0)s:=p0+p1
Since Yt
is stationary, it follows that qi=pi/s
for i∈{0,1}.
In preparation for induction, assume P(Yt=1∣Y0=1)=q1+q0(1−s)tP(Yt=1∣Y0=0)=q1−q1(1−s)t
It must follow that P(Yt+1=1∣Y0=1)=P(Yt+1=1∣Y1=1)(1−p0)+P(Yt+1=1∣Y1=0)p0=[q1+q0(1−s)t](1−p0)+[q1−q1(1−s)t]p0=q1+[q0(1−p0)−q1p0](1−s)t=q1+[q0(1−p0)−(1−q0)p0](1−s)t=q1+[q0−p0](1−s)t=q1+[q0−q0s](1−s)t=q1+q0(1−s)t+1
and P(Yt+1=1∣Y0=0)=P(Yt+1=1∣Y1=1)p1+P(Yt+1=1∣Y1=0)(1−p1)=[q1+q0(1−s)t]p1+[q1−q1(1−s)t](1−p1)=q1+[q0p1−q1(1−p1)](1−s)t=q1+[(1−q1)p1−q1(1−p1)](1−s)t=q1+[p1−q1](1−s)t=q1+[q1s−q1](1−s)t=q1−q1(1−s)t+1
which completes induction, noting the base case of
t=0
is true.
Due to the convenient mapping to Yt,
E[YtY0]=P(Yt=1∣Y0=1)P(Y0=1)=(q1+q0(1−s)t)q1=q12+q0q1(1−s)t
thus Cov[Yt,Y0]=E[YtY0]−E[Yt]E[Y0]=q12+q0q1(1−s)t−q12=q0q1(1−s)tCor[Yt,Y0]=(1−s)t
QED
Counterexample of Real-Valued 3-State Markov Chain
Let Zt
be a stationary Markov process such that P(Zt=−1)=P(Zt=0)=P(Zt=1)=1/3P(Zt+1=0∣Zt=−1)=1/2P(Zt+1=1∣Zt=0)=1/2P(Zt+1=−1∣Zt=1)=1/2
and for all i∈{−1,0,1},
P(Zt+1=i∣Zt=i)=1/2
Conveniently E[Zt]=0,
thus Cov[Zt,Zs]=E[ZtZs].
For one time step we have P(Z1=1∧Z0=1)=(1/3)(1/2)P(Z1=−1∧Z0=1)=(1/3)(1/2)P(Z1=1∧Z0=−1)=0P(Z1=−1∧Z0=−1)=(1/3)(1/2)
thus autocorrelation of one time step must be positive:
E[Z1Z0]=(1⋅1)16+(−1⋅1)16+(−1⋅−1)16=16
For two time steps P(Z2=1∧Z0=1)=(1/3)(1/2)2P(Z2=−1∧Z0=1)=(1/3)[(1/2)2+(1/2)2]P(Z2=1∧Z0=−1)=(1/3)(1/2)2P(Z2=−1∧Z0=−1)=(1/3)(1/2)2
thus the autocorrelation for two time steps must be negative:
E[Z2Z0]=(1⋅1)112+(−1⋅1)212+(1⋅−1)112+(−1⋅−1)112=−112
thus Cor[Z2,Z0]≠Cor[Z2,Z1]Cor[Z1,Z0]
ReferencesHamiltonJames D.Princeton University PressPrinceton, N.J1994978-0-691-04289-3