The handle https://hdl.handle.net/1887/3135034 holds various files of this Leiden
University dissertation.
Author: Ziemlańska, M.A.
Title: Approach to Markov Operators on spaces of measures by means of equicontinuity Issue Date: 2021-02-10
Chapter 5
Central Limit Theorem for some
non-stationary Markov chains
This chapter is based on:
Jacek Gulgowski, Sander C. Hille, Tomasz Szarek, Maria A. Ziemla´nska. Central Limit Theorem for some non-stationary Markov chains. Studia Mathematica, Number 246 (2019), Pages 109-131.
Abstract:
Using the classical Central Limit Theorem for stationary Markov chains proved by M. I. Gordin and B. A. Lif˘sic in [GL78] we show that it also holds for non–stationary Markov chains provided the transition probabilities satisfy the spectral gap property in the Kantorovich– Rubinstein norm.
5.1
Introduction
In this chapter we aim at showing that the Central Limit Theorem (CLT) obtained for stationary Markov chains by Gordin and Lif˘sic in [GL78] (see also its extension due to M. Maxwell and M. Woodroofe in [MW00]) may be extended to non–stationary ones provided that their transition probabilities satisfy the Spectral Gap Property in the Kantorovich– Rubinstein norm (see condition (A2) below). This condition implies that the iteration of the Markov operator associated to these transition probabilities constitute an equicon-tinuous family of conequicon-tinuous maps on probability measures, equipped with the Dudley metric, see Chapter 1, Theorem 2.8.4. In this way we obtain the result stronger than the almost sure result known in the literature as a quenched Central Limit Theorems (see [Pel15, BI95]).
Recently the CLT was proved for various non–stationary Markov processes (see [KW12, Kuk02]). For more details we refer readers to the book by T. Komorowski et al. [KLO12], where a more detailed description of recent results on central limit theorems is provided. Our result is in the same spirit as the main theorem in [KW12]. However some delicate approximation allow us to obtain the CLT for initial distributions with 2-nd moment finite instead of 2 δ.
In this chapter we introduce notation that deviates from previous chapters, mainly be-cause those used here are more common in the field of probability theory, while the topic considered is specially targeted to an audience from this field.
Suppose that X, ρ is a Polish space. By BX we denote the family of all Borel sets in X. Denote by BbX the set of all bounded Borel measurable functions equipped with
the supremum norm and let CbX be its subset consisting of all bounded continuous
functions.
By M1 and M we denote the spaces of all probability Borel measures and of all Borel
measures on X, respectively. Let π X BX 0, 1 be a transition probability on X and let U BbX BbX be defined by Ufx RXfyπx, dy for every f > BbX.
The operator U is dual to the Markov operator P defined onM and given by the formula P µ RXπx, µdx for µ > M, i.e.
SXfxP µdx S
X
U fxµdx for any f > BbX and µ > M.
In particular, we have P δx πx, for x > X.
5.2. Assumptions
whose transition probability is π. If the distribution of X0 is µ0, then the distribution of
Xn equals Pnµ0, n C 1. By Fn we denote the natural filtration of the chain, i.e. the
increasing family of σ–algebrasFn σXi i B n.
For any x> X and n C 1 we define the measure Pn
x on Xn,BXn, where BXn σA 1 A2 An Ai > BX, i 1, . . . , n by the formula PnxA1 An SXnχA1Anx1, . . . , xnπxn1, dxnπx1, dx2δxdx1 SXnχA1Anx1, . . . , xnP δxn1dxnP δx1dx2P δxdx1.
Here χA1Andenotes the characteristic function of A1An. Obviously, the distribution of the random vector X1, . . . , Xn if X0 x is given by Pnx. On the other hand, if the
distribution of X0 equals µ0, then the distribution of the random vector X1, . . . , Xn is
equal to Pn
µ0 RXPnx µ0dx.
By Px we will denote the probability measure on the Borel σ-algebra of the trajectory
space Xª associated with Xn with X0 x. Further, Ex denotes the expected value with
respect to Px. Analogously, if X0 is distributed with µ0, then Pµ0 RXPx µ0dx.
Similarly, we have then Eµ RXEx µdx.
5.2
Assumptions
Set Mp 1 µ > M1 S X ρx, x0pµdx @ ª for pC 1.We equip the space M1
1 with the Kantorovich-Rubinstein distance
Yµ νY sup VSXf dµ S
X
f dνV for µ, ν> M1 1,
where the supremum is taken over all Lipschitz functions f X R with the Lipschitz constant bounded by 1. If µ and ν are two Borel probability measures on some Polish spaces W and Z respectively, then byCµ, ν we denote the set of all joint Borel probability measures on W Z whose marginals are µ and ν (the so–called coupling). If µ, ν > M1
1,
Theorem 5.2.1. [Rubinstein Theorem [Bog07a], Theorem 8.10.45] The Kantorovich-Rubinstein distanceYµνY between Radon probability measures µ, ν > M1
1 can be represented
in the form
Yµ νY inf
γ>Cµ,νSXX
ρx, y γdx, dy.
Moreover, there exists a measure λ0> Cµ, ν at which the value Yµ νY is attained.
This immediately implies
Yδx δyY ρx, y for x, y> X.
Assumptions:
(A1) We assume that P leaves M1
1 invariant;
(A2) there exist C A 0 and 0 @ q @ 1 such that
YPnµ PnνY B CqnYµ νY (5.1)
for all µ, ν> M1
1 and n> N;
(A3) we assume that there exists µ> M2
1 such that
sup
nC1 SX
ρx, x02Pnµdx @ ª for some (thus all) x0> X. (5.2)
The following theorem is standard but we provide its proof for completeness of our pre-sentation.
Proposition 5.2.1. If (A1) and (A2) hold, then P has a unique invariant measure µ> M1
1 and for any µ> M11 we have YPnµ µY 0 as n ª. Moreover, if (A3) holds,
then µ> M2 1.
Proof. For any µ> M1
1X and any m, n > N, m C n one obtains from (5.1) that
YPnµ PmµY Bmn1Q k 0
YPnkµ Pnk1µY B Cqn
1 qYµ P µY. So Pnµ is a Cauchy sequence in the Y Y-complete space M1
1 ([Vil08], Theorem 6.18).
Hence it converges to some invariant measure µ > M1
5.3. Gordin–Lif˘sic results for stationary case
µ> M1 1X,
YPnµ µ
Y YPnµ PnµY B CqnYµ µY 0 as n ª.
So µ must be unique. Now suppose that (A3) holds and let µ> M2
1 be such that (5.2) is
satisfied. Let x0> X. For any K A 0, one has
sup
nC1 SX
ρx, x0 , K2Pnµdx B sup nC1 SX
ρx, x02Pnµdx M0 @ ª.
Since x( ρx, x0 , K2 is a Lipschitz function on X andYPnµ µY 0, we have
SX ρx, x0 , K2Pnµdx S X
ρx, x0 , K2µdx.
So RX ρx, x0 , K2µdx B M0 for every K. By the Monotone Convergence Theorem
we obtain that RX ρx, x02µdx B M0 too.
5.3
Gordin–Lif˘sic results for stationary case
We start with the following simple consequences of the Gordin and Lif˘sic result on the CLT for stationary Markov chains:
Proposition 5.3.1. Let P be a Markov operator that satisfies (A1) - (A3) and let Xn
be a stationary Markov chain corresponding to P . Then for any bounded Lipschitz function g X R such that RXgdµ 0, where µ is the unique invariant distribution for P , the limit σ2 lim n ªEµ gX0 gXº1 gXn n 2
exists and is finite. Moreover, if σA 0, then
lim n ªP gX0 gXº1 gXn n @ a 1 º 2πσ2 S a ªe y2 2σ2dy for all a> R.
Otherwise, if σ 0, then the sequence
gX0 gX1 gXn
º
n for nC 1
Proof. Let a bounded Lipschitz function g X R such that RXgdµ 0 be given. With no loss of generality we may assume that the Lipschitz constant of g equals 1. From Gordin and Lif˘sic [GL78] it follows that to finish the proof it is enough to show that
ª
Q
n 0
YUngY
L2X,µ@ ª, where U is dual to P . In fact, since RXUigxµ
dx RXgxµdx 0 for all i C 0, we have ª Q n 1 YUngY L2X,µ ª Q n 1SX Ungx2µ dx 1~2 ª Q n 1SX Ungx S XU ngyµ dy2µdx 1~2 B Qª n 1SXSX SUngx UngySµ dy 2 µdx 1~2 B Qª n 1SXSX Cqnρx, yµdy 2 µdx 1~2 B C Qª n 1 qnS XSX ρx, yµdy 2 µdx 1~2 B C Qª n 1 qnS Xρx, x0 SX ρx0, yµdy 2 µdx 1~2 B C1 q12 S X ρx, x02µdx 2 S X ρx0, yµdy 2 1~2 @ ª, by the fact that µ> M2
1. The proof is complete.
5.4
Auxiliary lemmas
We start with a theorem on the existence of a suitable coupling for the trajectories of a given Markov chain.
Theorem 5.4.1. Assume that a Markov operator P satisfies (A1) and (A2). Let l0 be a
positive integer such that Cql0 @ 1, where the constants C, q are given by (5.1). Then there exists a constant κA 0 such that for every integers l C l0 and mC 1 and every two points
x, y> X we have a measure Pml x,y > CPmlx , Pmly satisfying m Q i 1SXlmXlm ρxil, yilPmlx,ydx1, . . . , dxml, dy1, . . . , dyml B κρx, y. (5.3)
5.4. Auxiliary lemmas
Proof. Choose ˜q> Cql0, 1. Fix l C l
0. By induction on m for every two points x, y> X we
construct Pml
x,y> CPmlx , Pmly and prove that
SXlmXlm m Q i 1 ρxil, yil Pmlx,ydx1, . . . , dxml, dy1, . . . , dyml B m Q j 1 ˜ qjρx, y. (5.4) Then the hypothesis will hold with κ q˜1 ˜q1. Fix x, y> X and let m 1. Since
YPlδ
x PlδyY B Cql0ρx, y,
by the Kantorovich–Rubinstein theorem there exists µ1
x,y > CPlδx, Plδy such that
YPlδ
x PlδyY B S XX
ρu, vµ1x,ydu, dv B ˜qρx, y.
From the proof of the Kantorovich–Rubinstein theorem it follows that the function XX ? x, y µ1
x,y > CPlδx, Plδy is measurable if the space CPlδx, Plδy is endowed with
some metric of weak topology (see Theorem 11.8.2 in [Dud02]). We may assume that the measure µ1
x,y is absolutely continuous with respect to the product measure PlδxaPlδy. Let
gx,y X X 0,ª be the Radon–Nikodem density, i.e.
gx,yu, v dµ1 x,y dPlδ xa Plδy u, v u, v> X. Define Plx,yA SXl SXlχAx1, . . . , xl, y1, . . . , ylgx,yxl, ylπxl1, dxlπyl1, dyl πx1, dx2πy1, dy2δxdx1δydy1.
Let A A1 Al Xl for some Borel sets A1, . . . , Al` X. We have
Plx,yA
SXlχA1Alx1, . . . , xl SXlgx,yxl, ylπyl1, dylπy1, dy2δydy1 πxl1, dxlπx1, dx2δxdx1.
Moreover, the measure
is absolutely continuous with respect to Plδ
x. Obviously,
SXlgx,yxl, ylπyl1, dylπy1, dy2δydy1 SXgx,yxl, ylP
lδ
ydyl.
If we show that RXgx,y , ylPlδydyl 1, Plδx-a.s., then
Plx,yA
SXlχA1Alx1, . . . , xlπxl1, dxlπx1, dx2δxdx1 Pl
xA1 Al.
To do this set hxl RXgx,yxl, ylPlδydyl. From the definition of the coupling measure
µ1
x,y for any Borel set B` X we have
Plδ
xB µ1x,yB X S BSX
gx,yxl, ylPlδydyl Plδxdxl
SBhxlPlδxdxl.
Since the above equality holds for an arbitrary Borel set B` X, we obtain that hxl 1,
Plδ
x-a.s., and consequently Plx,y A1AlXl PlxA1Al. In the same way we
show that Pl
x,yXl B1 Bl PylB1 Bl for Borel sets B1, . . . , Bl` X. Hence
Pl
x,y> CPlx, Ply. Finally, we have
SXlXlρxl, ylP
l
x,ydx1, . . . dxl, dy1, . . . , dyl
SXlXlρxl, ylgx,yxl, ylπxl1, dxlπyl1, dylπx1, dx2πy1, dy2δxdx1δydy1 SXXρxl, ylgx,yxl, ylPlδxdxlPlδydyl S
XX
ρxl, ylµ1dxl, dyl B ˜qρx, y
and the first step of the proof is finished.
Assume now that for j 1, . . . , m and arbitrary x, y> X there exists Pjlx,y > CPjlx, Pjly such
that condition (5.4) holds with m replaced by j.
Define the measure Pm1lx,y on Xm1l Xm1l by the formula
Pm1lx,y A S Xm1lXm1l χAx1, . . . , xm1l, y1, . . . , ym1l Pmlx l,yldxl1, . . . , dxm1l, dyl1, . . . , dym1lP l x,ydx1, . . . , dxl, dy1, . . . , dyl
5.4. Auxiliary lemmas
for any Borel set A` Xm1l Xm1l. From the Markov property it easily follows that Pm1lx,y > CPm1lx , Pm1ly . We also have
SXm1lXm1l m1 Q j 1 ρxjl, yjlPm1lx,y dx1, . . . , dxm1l, dy1, . . . , dym1l B SXlXlSXmlXml m1 Q j 2 ρxjl, yjlPmlxl,yldxl1, . . . , dxm1l, dyl1, . . . , dym1l Pl x,ydx1, . . . , dxl, dy1, . . . , dyl S XlXlρxl, ylP l x,ydx1, . . . dxl, dy1, . . . , dyl BQm j 1 ˜ qjS XlXlρxl, ylP l x,ydx1, . . . dxl, dy1, . . . , dyl ˜qρx, y BQm j 1 ˜ qjqρ˜ x, y ˜qρx, y m1 Q j 1 ˜ qjρx, y
by the inductive hypothesis for 1 and m. This completes the proof.
As a consequence of Proposition 5.3.1 and Theorem 5.4.1 we have the following:
Proposition 5.4.1. Let P be a Markov operator that satisfies (A1) - (A3) and let Xn
be a Markov chain corresponding to P . Let l0 be a positive integer such that Cql0 @ 1, where
the constants C, q are given by (5.1) and let l C l0 be given. Set Yn Xnl for n C 0 and
assume that X0 Y0 x for x > X. Then for any bounded Lipschitz function g X R
such that RXgdµ 0, where µ is the unique invariant distribution for P , the limit ˜
σ2 lim n ªEµ
gY0 gY1 . . . gYn
º
n
2
exists and is finite. Moreover, if ˜σ2A 0, then the sequence of random vectors Y
n satisfies
lim
n ªP
gY0 gY1º . . . gYn
n @ a 1 º 2πσ2S a ªe y2 2˜σ2dy for all a> R.
Proof. Without loss of generality we may assume that g X R is bounded by 1 and its Lipschitz constant is also bounded by 1. By ϕx
n, x > X, we denote the characteristic
function of the random variable
gY0 gY1º . . . gYn
where Y0 x. When the distribution of Y0 is equal to µ the characteristic function is denoted by ϕµn. Obviously, ϕµ n t S Xϕ y ntµdy for t> R.
Moreover, from Proposition 5.3.1 it follows that lim n ªϕ µ n t e ˜ σ2t2 2 for t> R.
Theorem 5.4.1 allows us to evaluate the difference of characteristic functions ϕx
nt and
ϕynt for x, y > X and t > R. Fix x, y > X. We have
Sϕx nt ϕyntS S S Xlnexpit gx0 gxl . . . gxnl º n P nl xdx1, dx2, . . . , dxnl SXlnexpit
gy0 gyºl . . . gynl
n P nl y dy1, dy2,, dynlS B ºStS n SXlnXln ρx, y ρxl, xl ρxnl, ynl P nl x,ydx1, . . . , dxnl, dy1, . . . , dynl B ºStS n1 κρx, y for t> R, where Pnl
x,y> CPnlx, Pnly is given by Theorem 5.4.1. Consequently, for t > R we obtain
Sϕx nt ϕµntS Sϕxnt S X ϕyntµdyS B lim n ª StS1 κº n SX ρx, yµdy 0 as n ª. Since limn ªϕµnt et 2˜σ2~2
, we obtain that limn ªϕxnt et
2˜σ2~2
and the proof is complete.
For any Markov chainZn with values in the space X and an arbitrary Lipschitz function
g X R we shall denote SmngZi n Q i m gZi for nC m C 0.
The key point in proving the main theorem is the following lemma:
5.4. Auxiliary lemmas
Markov chain corresponding to P . Assume that g X R is a bounded Lipschitz function. Then there exists a constant M A 0 such that for any n > N the function
x Ex
Sn
0gXº i
n
2
has Lipschitz constant B M.
Proof. Again assume that g X R is a 1–Lipschitz function bounded by 1. Let l0 be a
positive integer such that Cql0 @ 1 with the constants C, q given by (5.1). We have Ex Sn 0gXi º n 2 1 n n Q i 0 ExgXi2 2 n n1 Q i 0 Q j 0@ji@l0,jBn ExgXigXj 2 n n Q i 0 Q j 0Bj@l0, l0jiBn Ex gXi ki,j Q m 1 gXml0ji,
where ki,j are maximal positive integers such that ki,jl0 j i B n. We show that all three
terms are Lipschitzean with a Lipschitz constant independent of n. To do this fix x, y> X. Since the function g 2 is a B 2–Lipschitz function, by (5.1) we have
SExgXi2 EygXi2S B 2YPiδx PiδyY B 2Cρx, y for i 1, . . . , n
and the first term
1 n n Q i 1 ExgXi2 is a B 2C–Lipschitz function.
Further, for jA i we have
EzgXigXj EzgXiESXigXj for all z> X,
where ESXigXj E gXjSFi. Since ESXigXj hXi, where h is a bounded by 1 a B C–Lipschitz function, we obtain that
SExgXigXj EygXigXjS SExgXihXi EygXihXiS
for i 1, . . . , n and consequently the Lipschitz constant of the second term 2 n n1 Q i 1 Q j 0@ji@l0,jBn ExgXigXj
is bounded from above by 4l0C.
Finally, we have Ex gXi ki,j Q m 1 gXml0ji Ex gXiESXi ki,j Q m 1 gXml0j . Using Theorem 5.4.1 we show that the function
X ? z Ez ki,j Q
m 1
gXml0j for all i, j C 0
has a Lipschitz constantB Cκqj, where κ is given by Theorem 5.4.1. Indeed, for any i, j C 0
we have Ez ki,j Q m 1 gXml0j EzESXj ki,j Q m 1 gXml0j. In turn, from Theorem 5.4.1 we obtain that the function
X ? u ru Eu ki,j Q
m 1
gXml0 has a Lipschitz constant B κ. Indeed, fix u, v > X. We have
Sru rvS SEu ki,j Q m 1 gxml0 Ev ki,j Q m 1 gxml0S B S SXki,j l02 ki,j Q m 1 gxml0 gyml0P ki,jl0
u,v dx1, . . . , dxki,jl0; dy1, . . . , dyki,jl0 B SXki,j l02 ki,j Q m 1 ρxml0, yml0P ki,jl0
u,v dx1, . . . , dxki,jl0; dy1, . . . , dyki,jl0 B κρu, v, by Theorem 5.4.1. Since Ez ki,j Q m 1 gXml0j S X ruPjδzdu,
5.4. Auxiliary lemmas the function X? z Ez ki,j Q m 1 gXml0j
has a Lipschitz constant B Cκqj, by (5.1). Moreover its supremum norm is bounded by
n. Since the Lipschitz constant of the function z gzrz is bounded by n Cκqj we
obtain that the function
X? x Ex gXi ki,j Q m 1 gXml0ji has a Lipschitz constant B Cqin Cκqj. Thus the third term
2 n n Q i 0 Q j 0Bj@l0, l0jiBn Ex gXi ki,j Q m 1 gXml0ji has Lipschitz constant bounded by
2 n n Q i 0 l01 Q j 0 Cqin Cκqj B 2C n n Q i 0 l0n Cκqi B 2l0C1 Cκ 1 q . Thus the function
x Ex
Sn
0gXº i
n
2
has Lipschitz constant B M, where
M 2C 4l0C
2l0C1 Cκ
1 q . This completes the proof.
Lemma 5.4.3. Let P be a Markov operator that satisfies (A1) - (A3) and let Xn be a
Markov chain corresponding to P . Assume that g X R is a bounded Lipschitz function. Let x> X and ε A 0. Then there exists K A 0 and N0> N such that
EPnδ x <@ @@ @> Sn 0gXº i m 2 χK,ªW Sn 0gXº i m W =A AA A?B ε (5.5)
for all m> N and n C N0.
such that Cql0 @ 1, where the constants C, q are given by (5.1) and let l C l
0 be given. Set
Yn Xnl for nC 0. We first show that the hypothesis holds if we replace Xn with Yn. Set
Sm
gY0 gY1º . . . gYm
m for mC 0.
The Markov chainYn corresponds to the operator Pland the assumptions of Lemma 5.4.2
are satisfied with P and Xn replaced by Pl and Yn, respectively. Thus the functions
y EySm2 are Lipschitzean with the Lipschitz constant bounded by, say ˜M , independent
of m. The Markov operator P satisfies (A2), so that we may find N0 > N such that
SEPnδ xS
2
m EµSm2S B ˜MYPnδx µY B ε~9l3 for n C N0 and m> N. (5.6)
Moreover, from Theorem 5.4.1 it follows that for any Lipschitz function Θ R R with the Lipschitz constant ˆM A 0 such that we have for all m C 1
SEyΘSm EzΘSmS B
κ ˆM º
mρy, z for y, z> X. Consequently, for some N1 > N we have
SEPnδ
xΘSm EµΘSmS B ε~9l3 for n C N1 and m> N. (5.7)
Let ϕK R R, K C 1 be a Lipschitz function such that
χ0,K1y B ϕKy B χ0,Ky for all x> R.
We are going to show that there exists KC 1 such that
EµSm21 ϕKSSmS @ ε~9l3 for all mC 1. By Proposition 5.3.1 we have EµSm2 σ˜2 as m ª and EµSm2ϕKSSmS 1 º 2πSR x2ϕKSySey 2~2˜σ2 dy as m ª.
5.4. Auxiliary lemmas
Since the integral on the right hand side converges to σ2 as K ª, we obtain that there
exist constants m0 and k0 such that
EµSm21 ϕK0SSmS EµSm2 EµSm2ϕK0SSmS @ ε~9l
3 for all m C m 0.
Enlarging K0 if necessary we obtain
EµSm21 ϕK0SSmS @ ε~9l
3 for all m C 1. (5.8)
Observe that the function y Θy given by the formula Θy y2ϕ
K0SyS is a Lipschitz function. Combining (5.6) - (5.8), we obtain that for nC maxN0, N1 we have
EPnδ x S 2 mχK0,ªSSmS B EPnδxS 2 m1 ϕK0SSmS B SEPnδ xS 2 m EµSm2S SEPnδ xΘSm EµΘSmS EµSm21 ϕK0SSmS B ε~9l3 ε~9l3 ε~9l3 ε~3l3.
Having this we may show that (5.5) holds with K lK0. Set mi maxj i jl B m for
i 0, . . . , l 1. We have EPnδ x gX0 . . . gXº m m 2 χK,ªW gX0 . . . gXº m m W B l2 EPnδ x max 0BiBl1 gXi gXilº . . . gXimil mi 2 χK,ªW gX0 . . . gXº m m W B l2 EPnδ x max 0BiBl1 gXi gXilº . . . gXimil mi 2 χK0,ªW gXi . . . gXº imil mi W B l2 EPnδ x gX0 gXl . . . gXm0l º m1 2 χK0,ªW gX0 gXl . . . gXm0l º m0 W l2 EPnδ x gX1 gX1l . . . gX1m1l º m1 2 χK0,ªW gX1 gX1l . . . gX1m1l º m1 W l2 EPnδ x gXl1 gX2l1 . . . gXl1ml1l º ml1 2 χK0,ªW gXl1 gX2l1 . . . gXl1ml1l º ml1 W B l 2lε~l3 ε.
Since εA 0 is arbitrary, the proof is complete.
Proposition 5.4.2. Let P be a Markov operator that satisfies (A1) - (A3) and let Xn
be a Markov chain corresponding to P . Let l0 be a positive integer such that Cql0 @ 1,
where the constants C, q are given by (5.1) and let l C l0 be given. Assume that Yn Xnl
for nC 0. If g X R is a bounded Lipschitz function, then
lim
m ªEx
gY0 gY1 . . . gYm
º m 2 ˜ σ2, where ˜ σ2 lim m ªEµ gX0 gXº1 . . . gXn n 2 .
Proof. First observe that for any kC 1 we have RRRRR
RRRRR REx
gY0 gY1º . . . gYn
n
2
Ex
gYk gYk1º . . . gYnk
n 2 RRRRR RRRRR R 0 as n ª. Further we have Ex
gYk gYk1º . . . gYnk
n
2
EPkδ x
gY0 gY1º . . . gYn
n
2
,
by the Markov property. On the other hand, the Markov chain Yn corresponds to the
operator Pl and the assumptions of Lemma 5.4.2 are satisfied with P and X
n replaced
with Pl and Y
n, respectively. Thus the functions
X ? x Ex
gY0 gY1 . . . gYn
º
n
2
for nC 0
are Lipschitzean with the Lipschitz constant bounded by, say ˜M , independent of n. Con-sequently, we obtain RRRRR RRRRR REP kδ x
gY0 gY1º . . . gYn
n
2
Eµ
gY0 gY1º . . . gYn
n 2 RRRRR RRRRR R B ˜MYPkδ x µY B ˜M CqkYδx µY.
5.5. The Central Limit Theorem
5.5
The Central Limit Theorem
Now we are in a position to formulate and prove the main theorem of this paper.
Theorem 5.5.1. (Central Limit Theorem) Let P be a Markov operator that satisfies (A1) - (A3) and letXn be a Markov chain corresponding to P . Assume that X0 x for x> X.
Then for any bounded Lipschitz function g X R such that RXgdµ 0, where µ is the unique invariant distribution for P , the sequence of random vectors Xn satisfies
lim n ªP gX0 gX1 . . . gXn º n @ a 1 º 2πσ2 S a ªe y2 2σ2dy for all a> R, if σ2 lim n ªEµ gX0 gX1 . . . gXn º n 2 A 0. Otherwise, if σ 0, the sequence
gX0 gX1 gXn
º
n for nC 1
converges in distribution to 0.
Proof. To prove the theorem we show that for any x> X
lim n ªExexpit gX0 gXº1 . . . gXn n e σ2t2~2 for t> R.
Without loss of generality we may assume that g is bounded by 1 and its Lipschitz constant is also bounded by 1. We will make use of the following formula:
eia 1 ia a2~2 Raa2, where SRaS B 1 and lima 0Ra R0 0.
Fix x> X, t > R 0 and ε A 0. Let k0 > N and η A 0 be such that
S1 σ2
0t2~2kk eσ
2t2~2
S @ ε for Sσ0 σS @ η and k C k0.
Set D supnC1RXRXρu, zPnδ
to Proposition 5.2.1 DB sup nC1 SX ρu, x0Pnδxdu S X ρz, x0µdz @ ª.
For given x, t, ε, by Lemma 5.4.3, we choose K A 0 and N0 > N such that
EPnδ x <@ @@ @> gX0 . . . gXm º m 2 χK,ªW gX0 . . . gXm º m W =A AA A?B ε~2t2 (5.9) for nC N0 and m> N.
Fix lC N0 such that
qlCDStS1 q1 MStS @ ε and qlC@ 1, (5.10) where the constants q, C are given by condition (5.1) and the constant M is given by Lemma 5.4.2. Since sup m,nC1E Pnδ x <@ @@ @> gX0 . . . gXm º m 2=A AA A?@ ª one may choose kC k0 such that
EPnδ x gX0 . . . gXm º m 2 WR ºθt k gX0 . . . gXm º m W χ0,KW gX0 . . . gXº m m W B ε~2t
2 for all m, nC 1 and θ > 0, 1,
(5.11)
by the fact that SRaS 0 as a 0 and Sσm σS @ η for m C k, where
σ2m Eµ gX0 . . . gXº ml m 2 Eµ gXl . . . gXº m m 2 .
This can be done due to the fact that σm tends to σ as m ª. For positive integers u, v
and w, vC l, where l is such that condition (5.10) holds, we set
Iu,v,w Exexpit
gXl . . . gXw . . . gXu1wl . . . gXuw
º
5.5. The Central Limit Theorem
and
Iw Exexpit
gX0 . . . gXº w
w .
Note that limn ªSIn Im,m, n~mS 0 for every m C 1. So let us consider in further detail
the terms Ij,k,m for given k, j B k and m C k. In the following we will use the notation
ESXj1m E SFj1m. We have Ij,k,m Exexpit gXl . . . gXm . . . gXj1ml . . . gXjm º km Ex exp it gXl . . . gXm . . . gXj2ml . . . gXj1m º km ESXj1mexp it gXj1ml . . . gXjm º km Ex exp it gXl . . . gXm . . . gXj2ml . . . gXj1m º km 1 it º kmESXj1mgXj1ml . . . gXjm t2 2kESXj1m gXj1ml . . . gXjm º m 2 t2 kESXj1m <@ @@ @> gXj1ml . . . gXjm º m 2 Rºt kmgXj1ml . . . gXjm =A AA A? . (5.12) Since EµgXi 0, by (A2) we obtain
SEugXl . . . gXmS SEPlδugX0 . . . gXml EµgX0 . . . gXmlS BQm i l SEPiδugX0 EµgX0S B m Q i l YPiδ x µY B qlC1 q1Yδ u µY B qlC1 q1S Xρu, zµdz
and hence we have
ExW it º kmESXj1mgXj1ml . . . gXjmW B q lC1 q1ºStS kmExSX ρXj1m, zµdz B qlC1 q1ºStS kmSXSX ρu, zµdzPj1mδxdu B qlCD1 q1 StS º km.
On the other hand, by Lemma 5.4.2 and (A2) we have RRRRR RRRRR REu gXl . . . gXm º m 2 Eµ gXl . . . gXm º m 2RRRRR RRRRR R RRRRR RRRRR REP lδu gX0 . . . gXml º m 2 Eµ gX0 . . . gXml º m 2 RRRRR RRRRR R B qlCMm l m Yδu µY B q lCM SXρu, zµdz, where M is given in Lemma 5.4.2 and consequently we obtain
ExRRRRRRRRRR R t2 kESXj1m gXj1ml . . . gXjm º m 2 t2 kEµ gXl . . . gXº m m 2 RRRRR RRRRR R B qlMt2 kExSX ρXj1m, zµdz B qlCMt 2 k SXSX ρu, zµdzPj1mδxdu B qlCM Dt2 k.
Finally conditions (5.9) and (5.11) will allow us to evaluate the term
ExRRRRRRRRRR R t2 kESXj1m <@ @@ @> gXj1ml . . . gXjm º m 2 Rºt kmgXj1ml . . . gXjm =A AA A?RRRRRRRRRRR . Indeed, we have ExRRRRRRRRRR R t2 kESXj1m <@ @@ @> gXj1ml . . . gXjm º m 2 Rºt kmgXj1ml . . . gXjm =A AA A?RRRRRRRRRRR B t2 kEx <@ @@ @>ESXj1mRRRRRRRRRR R gXj1ml . . . gXjm º m 2 Rºt kmgXj1ml . . . gXjm RRRRR RRRRR R =A AA A? t2 kEx <@ @@ @>RRRRRRRRRRR gXj1ml . . . gXjm º m 2 Rºt kmgXj1ml . . . gXjm RRRRR RRRRR R =A AA A?
5.6. Example t2 kEPj1mlδx <@ @@ @> gX0 . . . gXml º m 2 W R ºt kmgX0 . . . gXmlW =A AA A? B t2 kEPj1mlδx <@ @@ @> gX0 . . . gXml º m l 2 χK,ªW gX0 . . . gXml º m l W =A AA A? t2 kEPj1mlδx gX0 . . . gXº ml m l 2RRRRR RRRRR RRR ¾ m l m t º k gX0 . . . gXº ml m l RRRRRRRRRRRR χ0,KW gX0 . . . gXml º m l W by (5.9) and (5.11) B t2 kε~2t 2 t2 kε~2t 2 ε~k.
Now from (5.12) it follows that
SIj,k,m Ij1,k,m1 σ2mt2~2kS B qlCD1 q1 StS º km q lCM Dt2 k ε k B 2ε~k for j 1, . . . , k, by condition (5.10) and the fact that mC k. Iterating this formula k-times we obtain
SIk,k,m 1 σm2t
2~2kkS B 2ε
and consequently
SIk,k,m eσ
2t2~2
S B 3ε for all m sufficiently large.
Since ε A 0 was arbitrary, we obtain that limn ªSIk,k, n~k InS 0, which completes the
proof.
5.6
Example
Let X, ρ be a Polish space and let T, A be a measurable space. Let ν A 0,ª be some measure on T . Let p T X 0,ª be a measurable function such that
STpt, xνdt 1 for x> X.
We shall assume that pt, X 0,ª for t > T is a Lipschitzean function. Denote its Lipschitz constant by kt.
Let πt X BX 0, 1, t > T , be a transition probability and let Pt and Ut denote the
corresponding Markov operator and its dual, respectively. We shall consider the Markov chain Xn on some space Ω, F, P corresponding to the action of randomly chosen
tran-sition probabilities with probability depending on potran-sition, i.e. the Markov chain Xn
given by the formula:
P Xn1> SXn x πx, S
T pt, xπtx, νdt for x > X and n C 1. (5.13)
This model generalizes some random dynamical systems studied in [HSl16].
If the distribution of X0 is equal to µ, then the distribution of Xn is given by Pnµ, where
P M M is of the form P µA S
TSXpt, xπtx, Aµdxνdt for A> BX. (5.14)
Theorem 5.6.1. Assume that πtfor t> T are transition probabilities and there exists q @ 1
such that for the Markov operators Pt corresponding to πt we have
YPtµ1 Ptµ2Y B qYµ1 µ2Y for µ1, µ2 > M11.
Moreover, for some x0 > X and the operator Ut dual to Pt, t> T , we have
Ut ρ , x02x B a ρx, x02 b
for some a@ 1 and b A 0. Finally set γt sup
x>XSX
ρz, x0πtx, dz
for some x0> X and assume that
STγtktνdt @ 1 q and
STpt, xγtνdt @ ª for x> X.
Let Xn be a Markov chain given by (5.13). Assume that X0 x for x > X. Then for
any bounded Lipschitz function g X R such that RXgdµ 0, where µ is the unique invariant distribution for P , the sequence of random vectors Xn satisfies
lim n ªP gX0 gXº1 . . . gXn n @ v 1 º 2πσ2S v ªe y2 2σ2dy for v> R,
5.6. Example if σ2 lim n ªEµ gX0 gX1 . . . gXn º n 2 A 0. Otherwise, if σ 0, the sequence
gX0 gX1 gXn
º
n for nC 1
converges in distribution to 0.
Proof. From Theorem 5.5.1 it follows that to finish the proof it is enough to show that the Markov operator P given by (5.14) satisfies conditions (A1) - (A3).
Observe that P is a Feller operator and its dual is of the form U fx S
TSX
pt, xfzπtx, dzνdt for f > CbX. (5.15)
The operator U may be extended to an arbitrary positive unbounded function f and then it is also given by formula (5.15). The operator U can be, in fact, extended to the space of all Lipschitz functions. To show this, observe that for x0> X we have
U ρ , x0x S T pt, x S X ρz, x0πtx, dzνdt B S T pt, xγtνdt B ST Spt, x pt, x0Sγtνdt S T pt, x0γtνdt B ST ktγtνdt S T pt, x0γtνdt @ ª.
Therefore for any Lipschitz function f X R with Lipschitz constant L and x > X we have
SUfxS B USfSx B Sfx0S LUρ , x0x @ ª,
by the fact that the Lipschitz constant of the function SfS is bounded by L.
Set ˆq RT γtktνdt q @ 1 and let f X R be an arbitrary Lipschitz function with Lipschitz constant bounded by 1 and such that fx0 0. We show that then Uf is a
Lipschitz function with Lipschitz constant bounded by ˆq. Indeed, for any x, y> X we have SUfx UfyS VST SXpt, xfzπtx, dzνdt S TSX pt, yfzπty, dzνdtV B ST SXSpt, x pt, ySSfzSπtx, dzνdt ST pt, y VS X fzπtx, dz S X fzπty, dzV νdt B ST kt S X ρz, x0πtx, dz νdt ρx, y q S T pt, yνdt ρx, y B ST ktγtνdt q ρx, y ˆqρx, y.
To verify (A1) we compute SXρx, x0P µdx S
X
U ρ , x0xµdx
SXUρ , x0x Uρ , x0x0µdx Uρ , x0x0
B ˆqSXρx, x0µdx Uρ , x0x0 B S X
ρx, x0µdx Uρ , x0x0 @ ª
for µ> M1 1.
Observe that for any µ1, µ2 > M11 we have
Yµ1 µ2Y sup VS X
fxµ1dx S X
fxµ2dxV ,
where the supremum is taken over all Lipschitz functions f X R with Lipschitz constant bounded by 1 and such that fx0 0. Indeed for any f X R with Lipschitz constant
bounded by 1, we have VSXfxµ1dx S
X
fxµ2dxV VS
Xfx fx0µ1dx SXfx fx0µ2dxV .
To show that (A2) holds fix µ1, µ2> M11. Then we have
YP µ1 P µ2Y sup VS X fxP µ1dx S X fxP µ2dxV ˆ q supVS X ˆ q1U fxµ1dx S X ˆ q1U fxµ2dxV B ˆqsup VS X fxµ1dx S X fxµ2dxV ˆ qYµ1 µ2Y
5.6. Example
Here the supremum is taken over all Lipschitz functions f X R with Lipschitz constant bounded by 1 and such that fx0 0, by the fact that ˆq1U f is a Lipschitz function with
Lipschitz constant bounded by 1. Hence (A2) holds.
Finally, we show that (A3) holds with µ δx0. To do this it is enough to prove that U ρ , x0x2B a ρx, x02 b (5.16)
for some a@ 1 and b A 0. Indeed, iterating the above formula we have
Un ρ , x02x B an ρx, x02 b1 a1 for nC 1 and x > X,
consequently, sup nC1 SX ρx, x02Pnδx0dx B sup nC1a n SX ρx, x02δx0dx b1 a1 b1 a1@ ª. and (A3) will hold. To verify condition (5.16) we compute
U ρ , x0x2 S T pt, x0 S X ρz, x02πtx0, dzνdt S T pt, x0Ut ρ , x02xνdt B STpt, x0a ρx, x02 bνdt a ρx, x02 b.