Central limit theorem for functions of weakly dependent variables

Similar documents
1 Random Variable. Why Random Variable? Discrete Random Variable. Discrete Random Variable. Discrete Distributions - 1 DD1-1

Probability Distribution (Probability Model) Chapter 2 Discrete Distributions. Discrete Random Variable. Random Variable. Why Random Variable?

A GENERALIZATION OF A CONJECTURE OF MELHAM. 1. Introduction The Fibonomial coefficient is, for n m 1, defined by

Some Ideal Convergent Sequence Spaces Defined by a Sequence of Modulus Functions Over n-normed Spaces

Lecture 23: Central Force Motion

Some Remarks on the Boundary Behaviors of the Hardy Spaces

KANTOROVICH TYPE INEQUALITIES FOR THE DIFFERENCE WITH TWO NEGATIVE PARAMETERS. Received April 13, 2010; revised August 18, 2010

Quadratic Harmonic Number Sums

A question of Gol dberg concerning entire functions with prescribed zeros

Optimum Settings of Process Mean, Economic Order Quantity, and Commission Fee

JORDAN CANONICAL FORM AND ITS APPLICATIONS

ATMO 551a Fall 08. Diffusion

Application of Poisson Integral Formula on Solving Some Definite Integrals

LINEAR MOMENTUM Physical quantities that we have been using to characterize the motion of a particle

Existence and multiplicity of solutions to boundary value problems for nonlinear high-order differential equations

[ ] = jω µ [ + jω ε E r

THE LEAST COMMON MULTIPLE OF RANDOM SETS OF POSITIVE INTEGERS. 1. Introduction

THE LEAST COMMON MULTIPLE OF RANDOM SETS OF POSITIVE INTEGERS. 1. Introduction

30 The Electric Field Due to a Continuous Distribution of Charge on a Line

556: MATHEMATICAL STATISTICS I

6 Matrix Concentration Bounds

Queuing Network Approximation Technique for Evaluating Performance of Computer Systems with Hybrid Input Source

Research Article Approximation of Signals (Functions) by Trigonometric Polynomials in L p -Norm

Distributed Adaptive Networks: A Graphical Evolutionary Game-Theoretic View

On generalized Laguerre matrix polynomials

ALOIS PANHOLZER AND HELMUT PRODINGER

Appendix A. Appendices. A.1 ɛ ijk and cross products. Vector Operations: δ ij and ɛ ijk

New problems in universal algebraic geometry illustrated by boolean equations

A Multivariate Normal Law for Turing s Formulae

On the velocity autocorrelation function of a Brownian particle

FINITE DOUBLE SUMS OF KAMPE DE FE RIET S DOUBLE HYPERGEOMETRIC FUNCTION OF HIGHER ORDER

Math 124B February 02, 2012

EN40: Dynamics and Vibrations. Midterm Examination Tuesday March

Game Study of the Closed-loop Supply Chain with Random Yield and Random Demand

3.1 Random variables

On Bounds for Harmonic Topological Index

Introduction to Mathematical Statistics Robert V. Hogg Joeseph McKean Allen T. Craig Seventh Edition

arxiv: v1 [math.nt] 12 May 2017

Soft-Decision Majority Decoding of Reed Muller Codes

Robust Spectrum Decision Protocol against Primary User Emulation Attacks in Dynamic Spectrum Access Networks

FARADAY'S LAW. dates : No. of lectures allocated. Actual No. of lectures 3 9/5/09-14 /5/09

On the Poisson Approximation to the Negative Hypergeometric Distribution

CHAPTER 5: Circular Motion; Gravitation

Chapter 6 Differential Analysis of Fluid Flow

Physics Tutorial V1 2D Vectors

Temporal-Difference Learning

Orbital Angular Momentum Eigenfunctions

1121 T Question 1

On the Quasi-inverse of a Non-square Matrix: An Infinite Solution

Physics 161 Fall 2011 Extra Credit 2 Investigating Black Holes - Solutions The Following is Worth 50 Points!!!

1 Explicit Explore or Exploit (E 3 ) Algorithm

Ch. 4: FOC 9, 13, 16, 18. Problems 20, 24, 38, 48, 77, 83 & 115;

Do not turn over until you are told to do so by the Invigilator.

16 Modeling a Language by a Markov Process

ON THE TWO-BODY PROBLEM IN QUANTUM MECHANICS

In statistical computations it is desirable to have a simplified system of notation to avoid complicated formulas describing mathematical operations.

FARADAY'S LAW dt

Vortex Initialization in HWRF/HMON Models

LECTURE 15. Phase-amplitude variables. Non-linear transverse motion

Method for Approximating Irrational Numbers

Do Managers Do Good With Other People s Money? Online Appendix

Lecture 28: Convergence of Random Variables and Related Theorems

Journal of Computational and Applied Mathematics. Finite element analysis for the axisymmetric Laplace operator on polygonal domains

Adsorption and Desorption Kinetics for Diffusion Controlled Systems with a Strongly Concentration Dependent Diffusivity

Homework 7 Solutions

Results on the Commutative Neutrix Convolution Product Involving the Logarithmic Integral li(

Multiple Criteria Secretary Problem: A New Approach

ON LACUNARY INVARIANT SEQUENCE SPACES DEFINED BY A SEQUENCE OF MODULUS FUNCTIONS

BEST CONSTANTS FOR UNCENTERED MAXIMAL FUNCTIONS. Loukas Grafakos and Stephen Montgomery-Smith University of Missouri, Columbia

ON THE RECURRENCE OF RANDOM WALKS

Study on GPS Common-view Observation Data with Multiscale Kalman Filter. based on correlation Structure of the Discrete Wavelet Coefficients

Section 26 The Laws of Rotational Motion

( ) F α. a. Sketch! r as a function of r for fixed θ. For the sketch, assume that θ is roughly the same ( )

Asymptotically Lacunary Statistical Equivalent Sequence Spaces Defined by Ideal Convergence and an Orlicz Function

Internet Appendix for A Bayesian Approach to Real Options: The Case of Distinguishing Between Temporary and Permanent Shocks

VLSI IMPLEMENTATION OF PARALLEL- SERIAL LMS ADAPTIVE FILTERS

The second law of thermodynamics - II.

t is bounded. Thus, the state derivative x t is bounded. Let y Cx represent the system output. Then y

This paper is dedicated to the memory of Donna L. Wright.

A Bijective Approach to the Permutational Power of a Priority Queue

Pseudonormality and a Lagrange Multiplier Theory for Constrained Optimization 1

8-3 Magnetic Materials

PROBLEM SET #1 SOLUTIONS by Robert A. DiStasio Jr.

arxiv: v1 [physics.pop-ph] 3 Jun 2013

f err (AZ)g(Z)dZ n (A),A 0; da=i<_jh dlj (nljxlj); nlj Sm, ^ E Sm; the space of m m real symmetric matrices parameterlzed by (j), rm<) (2i

SOME SOLVABILITY THEOREMS FOR NONLINEAR EQUATIONS

Analysis of Arithmetic. Analysis of Arithmetic. Analysis of Arithmetic Round-Off Errors. Analysis of Arithmetic. Analysis of Arithmetic

Journal of Inequalities in Pure and Applied Mathematics

Chapter 3: Theory of Modular Arithmetic 38

d 2 x 0a d d =0. Relative to an arbitrary (accelerating frame) specified by x a = x a (x 0b ), the latter becomes: d 2 x a d 2 + a dx b dx c

SUPPLEMENTARY MATERIAL CHAPTER 7 A (2 ) B. a x + bx + c dx

(read nabla or del) is defined by, k. (9.7.1*)

Class 6 - Circular Motion and Gravitation

GENLOG Multinomial Loglinear and Logit Models

PHYSICS OF ASTROPHSYICS - Energy

Assessment of the general non-linear case ..=(LK).", J'

Semicanonical basis generators of the cluster algebra of type A (1)

Chapter 3 Optical Systems with Annular Pupils

An Adaptive Diagonal Loading Covariance Matrix Estimator in Spatially Heterogeneous Sea Clutter Yanling Shi, Xiaoyan Xie

Strong Result for Level Crossings of Random Polynomials. Dipty Rani Dhal, Dr. P. K. Mishra. Department of Mathematics, CET, BPUT, BBSR, ODISHA, INDIA

Transcription:

Int. Statistical Inst.: Poc. 58th Wold Statistical Congess, 2011, Dublin (Session CPS058 p.5362 Cental liit theoe fo functions of weakly dependent vaiables Jensen, Jens Ledet Aahus Univesity, Depatent of Matheatical Sciences Ny Munkegade, 8000 Aahus C, Denak E-ail: jlj@if.au.dk Result We descibe a cental liit theoe fo the su S n = n i=1 X i, X i R p, whee the X i s can be appoxiated by weakly dependent vaiables. The poof is fo i Z, but can be diectly genealized to the case of a ando field, i Z d. The weak dependency is foulated though a set of σ-algebas, D j, j Z. The stong ixing coefficients of these ae α(k, l, d = sup P (A1 A 2 P (A 1 P (A 2, whee the supeu is taken ove sets A i σ ( D j : j I i, i = 1, 2, with I1 k, I 2 l, and the distance d(i 1, I 2 between the two sets I 1 and I 2 is at least d. The cental liit theoe has applications in the study of nonhoogeneous hidden Makov chains (Jensen 2005. Theoe 1. Assue that thee exist δ 0, ɛ 0 > 0, δ 1, δ 2 0, θ > δ 1 + δ 2 + ax { (2 + δ 0 /δ 0, 1 + δ 2, 2 } and constants c 0, c 1, c 2 such that (1 (2 (3 EX i = 0, E X i 2+δ 0 c 0, Va(a S n ɛ 0 n a a R p α(k, l, d c 1 k δ 1 l δ 2 ax{1, d} θ, N X j σ ( D k : d(k, j : E X j X j c 2 θ. Then we have that Va(S n 1/2 S n N p (0, I fo n. (Fo the case of a ando field, X i, i Z ν, the lowe bound on θ is ultiplied by ν. We divide the poof into a nube of subsections. In the fist two subsections we use tuncation to educe the poble to that of bounded vaiables. In the last section the ethod of Bolthausen (1982 is used fo the bounded vaiables. Tuncation We use the tunctation function T M whee T M (x equals x fo x M and equals Mx/ x othewise. Let Q M (x = x T M (x. Using that EX j = 0 we wite the su S n as S n = S n + S N, with (4 S n = [ TM (X j E(T M (X j ] and S n = [ QM (X j E(Q M (X j ], and the idea is to pove that a CLT fo S n fo any fixed M iplies a CLT fo S n. In the lea below we conside fixed values of j, k and fixed unit vectos a j and a k. We define U = a j X j, U M = a j T M (X j and ÛM = a j Q M (X j, and define V, V M and ˆV M siilaly with j eplaced by k. Lea 2. Thee exists a constant c 3, depending on c 1, c 2, δ 0 and θ only, such that [ 1/(2+δ Cov(U, V c 3 c 0 0 + c 2/(2+δ ] { } 0 κ, 0 ax 1, d(j, k whee κ = (θ δ 1 δ 2 δ 0 /(2 + δ 0 > 1. Poof. Following Deo (1973 we expand Cov(U, V = Cov ( U M + ÛM, V M + ˆV M into fou tes and bound each of these. Thus, using the siple bound Q M (x ( x /M 1+δ 0 x we find Cov ( U M, ˆV M 2ME ˆVM 2Mc 0 /M 1+δ 0 = 2c 0 /M δ 0.

Int. Statistical Inst.: Poc. 58th Wold Statistical Congess, 2011, Dublin (Session CPS058 p.5363 Siilaly, Cov (ÛM, V M 2c0 /M δ 0, and Cov (ÛM, ˆV M { Va( Û M Va( ˆV M } 1/2 c0 /M δ 0, since EÛ M 2 E X j 2+δ 0 /M δ 0. To bound Cov ( ( U M, V M we define U M = a j T M X j and V M = a k T M (Xk. Since T M (X j T M (Xj X j Xj we find Cov ( U M UM, V M 2ME U M UM 2c 2 M θ, Cov ( UM, V M VM 2c 2 M θ, and theefoe Cov(U M, V M Cov(UM, V M + 4c 2 M θ. Finally, we use the classical bound (Ibagiov and Linnik, 1971 [17.2.1] Cov(UM, V M 4M 2 α ( 2 + 1, 2 + 1, ax{0, d(j, k 2}. Putting all the tes togethe we obtain Cov(U, V 5c0 /M δ 0 + 4c 2 M θ + 4c 1 M 2 (2 + 1 δ 1+δ 2 ax { 1, d(j, k 2 } θ. Choosing = d(j, k/3 and M = c 1/(2+δ 0 0 d(j, k κ/δ 0 we get the esult of the lea. Lea 3. Let S n be defined in (4. Thee exists a function b(m with b(m 0 fo M such that fo all unit vectos a and fo all n we have Va ( a S n /n b(m. Poof. We use Lea 2 with the ando vaiable X eplaced by Z = Q M (X E(Q M (X. Let 0 < ξ < δ 0 be so lage that κ 1 = (θ δ 1 δ 2 ξ/(2 + ξ > 1. Reebeing the siple inequality Q M (x ( x /M α x we have ET M (X i = EQ M (X i c 0 /M 1+δ 0 and E Q M (X i 2+ξ c 0 /M δ 0 ξ. Thus, eplacing δ 0 by ξ we use the bound (5 E Zi 2+ξ 2 1+ξ { E QM (X i 2+ξ + EQM (X i 2+ξ } 2 1+ξ{( c 0 /M δ 0 ξ + ( c 0 /M 1+δ 0 2+ξ} = c0 (M, ξ. Futheoe, we can appoxiate Z by Q M (X E(Q M (X with a ean eo E Q M (X Q M (X E X X + E T M (X T M (X 2E X X 2c 2 θ. We can now use Lea 2 with δ 0 eplaced by ξ, c 0 eplaced by c 0 (M, ξ and c 2 eplaced by 2c 2. Fo soe constant c 3 and any unit vecto a we then have Cov ( a Q M (X j, a Q M (X k c 3 b(m ax { 1, d(j, k } κ1, whee b(m = c 0 (M, ξ 1/(2+ξ + c 0 (M, ξ 2/(2+ξ. Witing Va(S N as a double su of covaiances we find the esult of the lea with b(m = c 3 b(m[3 + 2/(κ1 1]. Fo (5 we see that c 0 (M, ξ tends to zeo, and theefoe b(m tends to zeo as M tends to infinity. Vaiance Witing Va(a S n as a double su of covaiances we get diectly fo Lea 2 the bound Va(a S n [ 1/(2+δ c 4 n with c 4 = c 3 c 0 0 + c 2/(2+δ ] 0 0 [3 + 2/(κ 1]. Fo Lea 3 we find Va ( a S n / n Va ( a S n/ n 2 Cov ( a S n / n, a S n/ n + Va ( a S n/ n 2 c 4 b(m + b(m 0 fo M.

Int. Statistical Inst.: Poc. 58th Wold Statistical Congess, 2011, Dublin (Session CPS058 p.5364 The assuption of a lowe bound on the vaiance of S n / n in Theoe 1 theefoe gives that (6 Va ( a S n / n /Va ( a S n/ n 1 fo M. As in Ibagiov and Linnik (1971, page 346 we have fo Lea 3 and (6 that it suffices to pove a CLT fo S n fo fixed M to obtain a CLT fo S n. Poof of cental liit theoe fo S n Let a be a fixed unit vecto and define Y i = a (T M (X i ET M (X i and S n = n i=1 Y i = a S n. We saw in the poof of Lea 3 that ET M (X i c 0 /M 1+δ 0 so that fo M sufficiently lage we have Y i M + 1. Futheoe, if we set Yi = T M (Xi ET M (X i we have E Y i Yi E X i Xi c 2 θ and Y i M + 1. We will use the ethod of poof fo Bolthausen (1982. Fo this we need the following estiates. Lea 4. Thee exists a constant c 4 such that Cov ( Y j, Y k c4 (M+1 2 ax{1, d(j, k} γ, Cov ( Y j Y k, Y Y s c4 (M+1 4 ax { 1, d({j, k}, {, s} } γ, and Cov ( Y j, Y k Y Y s c4 (M + 1 4 ax { 1, d(j, {k,, s } } γ, whee γ = θ δ 1 δ 2. Poof. The poof is based on successively eplacing Y i by Yi in the ean of a poduct of Y s. Thus, using the second inequality as an exaple, E(Y j Y k Y Y s E(Yj Yk Y Ys 4(M + 1 3 c 2 θ. Fo inequalities of this fo we obtain Cov(Y j Y k, Y Y s Cov(Yj Yk, Y Y s 2 4(M + 1 3 c 2 θ. Next, the stong ixing iplies (Ibagiov and Linnik, 1971 [17.2.1] Cov(Y j Yk, Y Y s 4(M + 1 4 c 1 [2(2 + 1] δ 1+δ 2 ax { 1, d({j, k}, {, s } 2} θ. Cobining the two inequalities and taking = d({j, k}, {, s}/4 we obtain the second inequality of the lea. Fo a nube we intoduce the notation S i,n =,d(i,j Y j and α n = E(Y i S i,n, i=1 whee eventually will be tending to infinity with n. Fo Lea 4 we find that Va ( Sn / n α n n = 1 n i=1,d(i,j> Cov(Y i, Y j 4c 4(M + 1 2 (γ 1 γ 1, whee γ 1 = θ δ 1 δ 2 1 > 0. Now, conside the case = n ω, with ω > 0. Then the ight hand side above tends to zeo. Since also we have fo (6 that Va( S n is of the sae ode as Va(S n, and we have assued the lowe bound ɛ 0 n fo the latte, we find that α n has a siila lowe bound and Va( S n /α n 1. We ust theefoe show that S n = S n / α n conveges to a standad noal distibution. Poof of CLT fo S n We follow the poof of Bolthausen (1982, whee the definitions of A 1, A 2 and A 3 below can be found. These tes depend on the aguent t of the chaacteistic function fo

Int. Statistical Inst.: Poc. 58th Wold Statistical Congess, 2011, Dublin (Session CPS058 p.5365 S n. One needs to show that E ( A 1 A 2 A 3 0. Using the expession in Bolthausen (1982 and Lea 4 we have with J = {1 i, i, j, j n : d(i, j, d(i, j }, (7 E A 1 2 t 2 = 2 Cov(Y i Y j, Y i Y j J = t2 { 2 Cov(Y i Y j, Y i Y j + J,d(i,i 3 c 4t 2 (M + 1 4 { 2 n(2 + 1 2 2 J,d(i,i <3 } Cov(Y i Y j, Y i Y j (k 2 γ + n(2 + 1 8 6 ( 3 1 + j γ} k=3 = O ( t 2 (M + 1 4 2 /n = O ( t 2 (M + 1 4 /n 1 2ω, whee the last expession follows upon taking = n ω. Fo the A 2 te we have fo Bolthausen (1982 and Lea 4 (8 E A2 nt 2 (M + 1 α 3/2 n ax i j,k=1,d(i,j,d(i,k Cov(Y j, Y k c 4nt 2 (M + 1 3 { 2+1 3/2 (2 + 1 1 + 2 j γ} = O ( t 2 (M + 1 3 /n 1 ω 2. Finally, we need to conside A 3 = 1/2 n i=1 Y i exp { it( S n S i,n }, whee S i,n = S i,n / α n. Consideing the exponential pat, and eplacing all the vaiables by the appoxiating vaiables Yj, we see that E exp { it( S n S i,n } exp { it( S n S i,n } Using this we obtain t E ( S n S i,n ( S n S i,n t nc 2 θ. ( Cov Yi, exp{it( S n S i,n } Cov ( Yi, exp{it( S n S i,n} c2 θ + (M + 1 t nc 2 θ. Fo the ixing we have (Ibagiov and Linnik, 1971 [17.2.1] Cov ( Y i, exp{it( S n S i,n } 4(M + 1c 1 (2 + 1 δ 1 n δ 2 ( 2 θ. Taking = /3 and cobining the two bounds we find fo soe constant c 5 Cov ( Y i, exp{it( S n S i,n } { c 5 (M + 1n δ 2 θ+δ 1 + θ [1 + (M + 1 t n/ α n ] }. Since EY i = 0 the tes in A 3 ae covaiances, and taking = n ω we theefoe have the bound (9 EA 3 = O ( (M + 1[n ωθ+1 + n ω(θ δ 1+δ 2 + 1 2 ]. Thus, if we choose ω such that ω < 1 2, ωθ 1 > 0 and ω(θ δ 1 δ 2 1 2 > 0, which is possible fo the assuption on θ in Theoe 1, we see that all of (7, (8, and (9 tend to zeo. REFERENCES Bolthausen, E. (1982: On the cental liit theoe fo stationay ixing ando vaiables. Ann. Pobab., 10, 1047 1050. Deo, C. M. (1973: A note on epiical pocesses of stong-ixing sequences. Ann. Pobab., 1, 870 875.

Int. Statistical Inst.: Poc. 58th Wold Statistical Congess, 2011, Dublin (Session CPS058 p.5366 Ibagiov, I. A. and Linnik, Yu. V. (1971: Independent and Stationay Sequences of Rando Vaiables. Woltes-Noodhoff Publishing, Goningen. Jensen, J.L. (2005: Context dependent DNA evolutionay odels. Reseach Repot, No. 458, Depatent of Matheatical Sciences, Univesity of Aahus.