ebook img

A central limit theorem for Lebesgue integrals of random fields PDF

0.16 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview A central limit theorem for Lebesgue integrals of random fields

A central limit theorem for Lebesgue integrals of random fields Ju¨rgen Kampf January 5, 2016 6 1 0 Abstract 2 InthispaperweshowacentrallimittheoremforLebesgueintegralsofstationaryBL(θ)-dependent n randomfieldsastheintegrationdomaingrowsinVanHove-sense. Ourmethodistousethe(known) a analogue result for discrete sums. As applications we obtain various multivariate versions of this J central limit theorem. 4 ] 1 Introduction R P Random fields are collections of random variables indexed by the Euclidean space Rd. They have appli- . h cationsinvariousbranchesofscience,e.g.inmedicine[1,13],ingeostatistics[5,15]orinmaterialscience t a [10, 14]. m The aim of the present paper is to establish a central limit theorem for integrals X(t)dt, where Wn [ (Wn)n N isasequenceofcompactsubsetsofRd and(X(t))t Rd isarandomfield. TheRsequence(Wn)n N 1 of inte∈gration domains is assumed to grow in Van Hove-sen∈se (VH-sense), i.e. ∈ v 3 lim λd((bdWn)+Bd)/λd(Wn)=0, n 1 →∞ 5 whereλ denotestheLebesguemeasure,bdW istheboundaryofW Rd,A+B := a+b a A, b B d 0 for two subsets A,B Rd and Bd := x Rd x 1 is the close⊆d Euclidean un{it bal|l. ∈ ∈ } 0 ⊆ { ∈ |k k≤ } Themainresultofthe presentpaperisthe following(the notionofBL(θ)-dependence willbe defined . 1 in Subsection 2.1). 0 6 Theorem 1. Let θ =(θr)r N be a monotonically decreasing zero sequence. Let (X(t))t Rd be a measur- 1 able, stationary, BL(θ)-dep∈endent R-valued random field such that ∈ : v i Cov X(0),X(t) dt< . X ZRd| | ∞ r (cid:0) (cid:1) a Let (Wn)n N be a VH-growing sequence of subsets of Rd. Then ∈ X(t)dt EX(0)λ (W ) Wn − d n (0,σ2), n , R λd(Wn) →N →∞ in distribution, where p σ2 := Cov X(0),X(t) dt. ZRd (cid:0) (cid:1) There is a wide literature on similar results, where mixing conditions are assumed instead of BL(θ)- dependence, see e.g. [4, 7, 8, 9]. For BL(θ)-dependent random fields there are no central limit theorems for Lebesgue integrals up to now. However, there are such results for discrete sums [2] and for Lebesgue measures of excursion sets [3]. In the latter paper the random field is in fact assumed to be quasi- associated, which is a slightly stronger assumption than BL(θ)- dependence. This paper is organized as follows: In Section 2 we collect preliminaries about associated random variables,randomfieldsandfunctionsofboundedvariation. Section3isdevotedtotheproofofthemain theorem. InSection4wepresentseveralexampleshowthemainresultcanbeextendedtoamultivariate centrallimittheorem. Thecasethatthe randomfieldisofthe form(f(X(t)))t Rd forsomedeterministic function f :R Rs and some random R-valued field (X(t))t Rd will be of par∈ticular interest. → ∈ 1 2 PRELIMINARIES 2 2 Preliminaries 2.1 Association concepts In this subsection we introduce different association concepts and discuss their relations. We start with the broadest appearing in this paper, namely BL(θ)-dependence. For finite subsets I,J Rd we put dist(I,J) := min x y : x I, y J , where is the 1 1 ℓ1-norm. For two Lipschitz⊆functions f :Rn1 R and g :{Rkn2− kR we p∈ut ∈ } k·k → → Ψ(n ,n ,f,g)=min n ,n Lip(f)Lip(g), 1 2 1 2 { } where f(x) f(y) Lip(f):=sup | − | x,y Rn, x=y x y | ∈ 6 1 n k − k o denotes the (optimal) Lipschitz constant of a Lipschitz function f :Rn R. Fora randomfield (X(t))t Rd, a finite subsetI = t1,...,tn Rd w→ith n elements anda function f on Rn we abbreviate f(X ):=∈f(X(t ),...,X(t )). I{f such an a}b⊆breviation X appears more than once I 1 n I within one formula, then always the same enumeration of the elements of I has to be used. For a set M let #M denote the number of elements of M. Furthermore, for ∆>0 we put T(∆):= (j /∆,...,j /∆) (j ,...,j ) Zd . 1 d 1 d { | ∈ } Def. 2. Let θ =(θr)r N be a monotonically decreasing sequence with limr θr =0. ∈ →∞ (i) An Rs-valuedrandomfield (X(t))t Rd is called BL(θ)-dependent if for any ∆>1 andany disjoint, finite sets I,J T(∆) with dist(I,∈J) r and all bounded Lipschitz functions f : Rs#I R and · g :Rs#J R ⊆we have ≥ → · → Cov(f(X ),g(X )) Ψ(#I,#J,f,g)∆dθ . I J r ≤ (ii) An Rs-valued random field (X(t))t Zd is called BL(θ)-dependent if for any disjoint, finite sets I,J Zd with dist(I,J) r andall∈boundedLipschitz functions f :Rs#I R andg :Rs#J R · · ⊆ ≥ → → we have Cov(f(X ),g(X )) Ψ(#I,#J,f,g)θ . I J r ≤ Lemma 3. Let θ = (θr)r N be a monotonically decreasing sequence with limr θr = 0. For T = Zd or T =Rd, let (X(n)(t)) ∈,n N, be a sequence of BL(θ)-dependent random fi→e∞lds such that the finite- t T ∈ ∈ dimensionaldistributionsconvergetothoseofafield(X(t)) . Then(X(t)) isalsoBL(θ)-dependent. t T t T ∈ ∈ Proof: Bythedefinitionofconvergenceinprobabilitywegetlim Ef(X(n))g(X(n))=Ef(X )g(X ), n I J I J lim Ef(X(n))=Ef(X ) andlim Eg(X(n))=Eg(X )f→or∞any finite sets I,J T andbounded Lipnsc→h∞itz contiInuous functioIns f :R#In→∞R andJg :R#J RJ, which yields the asserti⊆on. → → Lemma 4. Let θ = (θr)r N be a monotonically decreasing sequence with limr θr = 0. Let (X(t))t T be a BL(θ)-dependent ran∈dom field and let f : Rs Rs′ be a Lipschitz fu→n∞ction. Then there is∈a → monotonically decreasing sequence θ′ = (θr′)r N with limr θr′ = 0 such that (f(X(t)))t T is BL(θ′)- ∈ →∞ ∈ dependent. Proof: For a function f : V W and a finite set I let f denote the function V#I W#I, I → → (x ,...,x ) (f(x ),...,f(x )). We put θ := Lip(f)2 θ . Let I,J be two disjoint finite sets wi1th dist(#I,IJ)7→ r and1 let f˜:Rs#′#II R and g :r′Rs′#J R b·ertwo Lipschitz functions. Then · · ≥ → → Cov(f˜(f (X )),g(f (X ))) min #I,#J Lip(f˜ f ) Lip(g f ) ∆d θ I I J J I J r ≤ { }· ◦ · ◦ · · min #I,#J Lip(f˜) Lip(g) ∆d θ . ≤ { }· · · · r′ An Rs-valued random field (X(t)) is called positively associated (PA) if t T ∈ Cov(f(X ),g(X )) 0 I J ≥ 2 PRELIMINARIES 3 for any finite sets I,J T and functions f : Rs#I R and g : Rs#J R which are bounded and · · ⊆ → → monotonically increasing in every coordinate. For a Lipschitz function f :Rn R we define coordinate-wise Lipschitz constants by → f(x ,...,x ,y ,x ,...,x ) f(x ,...,x ,z ,x ,...,x ) 1 k 1 k k+1 n 1 k 1 k k+1 n Lipk(f)=sup | − y −z − | | k k n | − | x ,...,x ,x ,...,x ,y ,z R, y =z , k 1,...,n . 1 k 1 k+1 n k k k k − ∈ 6 ∈{ } o An Rs-valued random field (X(t)) with E[X (t)2] < ,t T,k = 1,...,s, is called quasi-associated t T k ∈ ∞ ∈ (QA) if s s Cov(f(X ),g(X )) Lip (f) Lip (g) Cov(X (t),X (u)) | I J |≤ t,k · u,l | k l | t I k=1u J l=1 X∈ XX∈ X for any finite sets I,J T and Lipschitz continuous functions f :Rs#I R and g :Rs#J R. · · ⊆ → → ItiswellknownthateveryPArandomfieldisalsoQA,seee.g.Theorem5.3in[2,p.89](thistheorem is only formulated in the special case s=1 and T =Zd, but the proof holds in the present setting). Lemma 5. Let (X(t))t Rd be an Rs-valued QA random field. Assume that there are c > 0 and ǫ > 0 with ∈ Cov(Xi(t1),Xj(t2)) c t1 t2 −d−ǫ ≤ ·k − k∞ fort1,t2 Rd andi,j =1,...,s. Then(X(t))t Rd isBL(θ)-dependentforsomemonotonicallydecreasing zero sequ∈ence θ. ∈ Proof: Let r>0, ∆>1 and let I,J T(∆) be finite with dist(I,J) r, w.l.o.g. #I #J. Moreover, let f :Rs#I R and g :Rs#J R b⊆e bounded and Lipschitz contin≥uous. Then ≤ · · → → s s Cov(f(X ),g(X )) Lip (f) Lip (g)Cov(X (t),X (u)) I J ≤ t,k · u,l k l t I k=1u J l=1 X∈ XX∈ X s2 #I Lip(f) Lip(g) max Cov(X (t),X (u)) k l ≤ · · · · t,k,l u J X∈ s2 #I Lip(f) Lip(g) max c t u d ǫ. − − ≤ · · · · t I ·k − k∞ ∈ u J X∈ We have for fixed t I, if r >1, ∈ t u −d−ǫ ∞ s −d−ǫ # v T(∆) v = s k − k∞ ≤ ∆ · { ∈ |k k∞ ∆} uX∈J s=X⌈r∆⌉(cid:0) (cid:1) = ∞ s −d−ǫ (2s+1)d (2s 1)d ∆ · − − s=X⌈r∆⌉(cid:0) (cid:1) (cid:0) (cid:1) d 1 ∞ − d =∆d+ǫ s d ǫ (1+( 1)d ι 1)(2s)ι − − − − ι − s=X⌈r∆⌉Xι=0 (cid:18) (cid:19) d 1 ∆d+ǫ ∞ − d (1+( 1)d ι 1)2ιs d ǫ+ιds − − − − ≤ ι − Z⌈r∆⌉−1Xι=0(cid:18) (cid:19) ∆d+ǫd−1 d (1+( 1)d−ι−1)2ι(r∆−∆)−d−ǫ+ι+1 ≤ ι − d+ǫ ι 1 ι=0(cid:18) (cid:19) − − X ∆dd−1 d (1+( 1)d−ι−1)2ι(r−1)−d−ǫ+ι+1. ≤ ι − d+ǫ ι 1 ι=0(cid:18) (cid:19) − − X 2 PRELIMINARIES 4 Putting θr :=(c∆3d·ds·2mPadιx=−i=011(cid:0),d.ι.(cid:1).,(s1V+ar((−X1i)(d0−))ι−+1)θ22ι(r−d1+)−ǫ−d−ι−ǫ+1ι+1 ffoorr rr >=11,, we obtain Cov(f(X ),g(X )) min #I,#J Lip(f) Lip(g) ∆dθ . I J r ≤ { }· · · 2.2 Random fields After having introduced the association concepts in subsection 2.1, we will now collect various other preliminaries concerning random fields. The following theorem(see [6, Ch. III, 3]and [12, Prop.3.1])saysthat for stationaryrandomfields § stochastic continuity and measurability are essentially equivalent. Theorem 6. (i) Let(X(t))t Rd beastochasticallycontinuousrandomfield. Thenthereisameasurable modification of (X(t))t Rd∈. ∈ (ii) Let (X(t))t Rd be a stationary and measurable random field. Then (X(t))t Rd is stochastically continuous.∈ ∈ Lemma 7. Let (Xt)t Rd be a stationary, stochastically continuous random field with EX(0)j < for j >0. Then (Xt)t Rd∈is continuous in j-mean. ∞ ∈ Proof: Let (tn)n N be a sequence of points in Rd converging to a point t Rd. Then ∈ ∈ lim E X(t ) X(t)j = lim ∞P(X(t ) X(t)j x)dx= ∞ lim P(X(t ) X(t) √j x)dx=0. n n n n | − | n | − | ≥ n | − |≥ →∞ →∞Z0 Z0 →∞ We have been allowed to interchange limit and integral, since P(X(t ) X(t) √j x) P(X(t ) √jx)+P(X(t) √jx)=2P(X(t) √jx) | n − |≥ ≤ | n |≥ 2 | |≥ 2 | |≥ 2 due to the stationarity and ∞2P(X(t) √jx)dx= ∞2P(2 X(t)j x)dx=2j+1E X(0)j < . | |≥ 2 | · | ≥ | | ∞ Z0 Z0 Lemma 8. Let (X(t))t Rd and (Y(t))t Rd be two stochastically continuous and measurable random fields having the same distrib∈ution. Let A ,∈...,A Rd be bounded Borel sets. Assume that X(t)dt is 1 m ⊆ Ai defined a.s. for i = 1,...,m, i.e. not both the positive part and the negative part of these integrals are R infinite. Then Y(t)dt,..., Y(t)dt are defined a.s. as well and A1 Am R R d X(t)dt,..., X(t)dt = Y(t)dt,..., Y(t)dt . (cid:18)ZA1 ZAm (cid:19) (cid:18)ZA1 ZAm (cid:19) Proof: By the Monotone ConvergenceTheorem, we may assume w.l.o.g. that there is some N N such that X(t) [ N,N] and Y(t) [ N,N] for all t Rd. ∈ ∈ − ∈ − ∈ We define processes (Xn(t))t Rd and (Yn(t))t Rd by putting ∈ ∈ z z z z +1 z z +1 Xn(t ,...,t ):=X 1,..., d , for all t 1, 1 ,...,t d, d , z ,...,z Z. 1 d 1 d 1 d n n ∈ n n ∈ n n ∈ (cid:16) (cid:17) h (cid:17) h (cid:17) We get Xn(t)dt i=1,...,m (1) nZAi (cid:12) o (cid:12) z1 z1+1 zd zd+1 z1 zd (cid:12)= λ A , , X ,..., i=1,...,m d i ∩ n n ×···× n n n n nz1,.X..,zd∈Z (cid:16) (cid:2) (cid:1) (cid:2) (cid:1)(cid:17) (cid:0) (cid:1)(cid:12)(cid:12) o d z1 z1+1 zd zd+1 z1 zd (cid:12) = λ A , , Y ,..., i=1,...,m d i ∩ n n ×···× n n n n nz1,.X..,zd∈Z (cid:16) (cid:2) (cid:1) (cid:2) (cid:1)(cid:17) (cid:0) (cid:1)(cid:12)(cid:12) o = Yn(t)dt i=1,...,m . (cid:12) (2) nZAi (cid:12) o (cid:12) (cid:12) 2 PRELIMINARIES 5 For ǫ,δ >0 we get m m P Xn(t)dt X(t)dt >ǫ P Xn(t) X(t) dt>ǫ − ≤ | − | (cid:16)Xi=1(cid:12)ZAi ZAi (cid:12) (cid:17) (cid:16)Xi=1ZAi (cid:17) (cid:12) (cid:12) E m Xn(t) X(t) dt i=1 Ai| − | ≤ ǫ P R m E Xn(t) X(t) dt = i=1 Ai | − | ǫ P R m δ+P(Xn(t) X(t) >δ) 2N dt i=1 Ai | − | · ≤ ǫ P R (cid:0) (cid:1) m δdt n→∞ i=1 Ai −→ ǫ P R m δ λ (A ) = · i=1 d i . ǫ P The limit relation holds by the MajorizedConvergenceTheorem, since the assumption that (X(t))t R is stochastically continuous implies that Xn(t) converges to X(t). Since δ >0 was arbitrary,we get ∈ m P Xn(t)dt X(t)dt >ǫ n→∞0 − −→ (cid:16)Xi=1(cid:12)ZAi ZAi (cid:12) (cid:17) and the same way (cid:12) (cid:12) m P Yn(t)dt Y(t)dt >ǫ n→∞0. − −→ (cid:16)Xi=1(cid:12)ZAi ZAi (cid:12) (cid:17) Now (2) yields the assertion. (cid:12) (cid:12) 2.3 Functions of bounded variation A function f : R R is said to be of locally bounded variation if there is a monotonically increasing function α:R R→and a monotonically decreasing function β :R R such that f =α+β. We denote → → the set of such functions α and β by A resp. B. We put inf α(x) α A, α(0)=f(0) if x>0 { | ∈ } f+(x):= f(0) if x=0  sup α(x) α A, α(0)=f(0) if x<0. { | ∈ } It is easy to see that f+ A and f− :=f f+ B. We put hf :=f+ f−. ∈ − ∈ − Lemma 9. Let f : R R be a function of locally bounded variation. Then f = g h for a Lipschitz f continuous function g :→R R of Lipschitz constant 1. ◦ → Proof: Foreachx R,forwhichthereist Rwithh (t)=x,defineg(x):=f(t). Nowgiswell-defined, f sincefort ,t Rw∈ithh (t )=h (t ),f is∈constanton[t ,t ]. Clearly,f =g h . Moreover,g -defined 1 2 f 1 f 2 1 2 f onasubsetof∈Rsofar-isLipschitzcontinuouswithLipschitzconstant1. Indeed◦,letx ,x R, x <x , 1 2 1 2 be two points for which there are t ,t R with h (t )=x and h (t )=x . Then ∈ 1 2 f 1 1 f 2 2 ∈ hf(t2) hf(t1)=f+(t2) f+(t1) (f−(t2) f−(t1)) f+(t2) f+(t1)+(f−(t2) f−(t1)) = f(t2) f(t1). − − − − ≥| − − | | − | Hence x x g(x ) g(x ). 2 1 2 1 It rem−ains t≥o |show t−hat g h|as a Lipschitz continuous extension to the whole of R. The domain of g is R minus the union of countable many, disjoint intervals. For a point x lying on the boundary of the domain of g but not in the domain of g, choose a sequence (xn)n N such that g(xn) is defined for all n N. Then (g(xn))n N is a Cauchy sequence, since g is Lipschit∈z continuous, and hence convergent. ∈ ∈ Since (g(xn))n N is convergentfor every such sequence (xn)n N, the limit is independent of the choice of ∈ ∈ thesequence. Sowecanputg(x):=lim g(x ). ItiseasytoseethatthisextensionstillhasLipschitz n n →∞ constant 1. Now all gaps in the domain of g are open intervals. So they can be filled by affine functions. Clearly, the Lipschitz constant is preserved again. 3 THE UNIVARIATE CLT 6 3 The univariate CLT In this section we will prove Theorem 1. Proof: For j = (j ,...,j ) Zd we put Q = d [j ,j +1) and Z(j) := X(t)dt EX(0). We 1 d ∈ j ×i=1 i i Qj − will show that this random field (Z(j))j Zd fulfills the assumptions of TheoreRm 1.12 of [2, p. 178]. The collection ∈ n 1 Zn(j):= nd X(j1+ kn1,...,jd+ knd)−EX(0), j ∈Zd, k1,.X..,kd=1 is BL(θ )-dependent for any n N, where θ := θ . Indeed, let I,J Zd and let f : R#I R and g :R#J′ R beboundedLipsch∈itzfunctions.r′ PutI˜r−=dI+ 1/n,2/n,...,⊆1 d,J˜=J+ 1/n,2/n→,...,1 d, → { } { } nd nd 1 1 f˜:R#Ind R, (x ,...,x ) f x EX(0),..., x EX(0) · → 1,1 #I,nd → nd 1,ℓ− nd #I,ℓ− (cid:16) Xℓ=1 Xℓ=1 (cid:17) and nd nd 1 1 g˜:R#J·nd →R, (x1,1,...,x#J,nd)→g nd x1,ℓ−EX(0),...,nd x#J,ℓ−EX(0) . (cid:16) Xℓ=1 Xℓ=1 (cid:17) Then we have f(Z ) = f˜(X ), g(Z ) = g˜(X ), Lip(f˜) = Lip(f)/nd, Lip(g˜) = Lip(g)/nd and n,I I˜ n,J J˜ dist(I˜,J˜) dist(I,J) d. So ≥ − Cov(f(Z ),g(Z ))=Cov(f˜(X ),g˜(X )) n,I n,J I˜ J˜ min #I nd,#J nd Lip(f˜)Lip(g˜)ndθ r d ≤ { · · } − =min #I,#J Lip(f)Lip(g)θ . { } r′ By Lemma 3, the field (Z(j))j Zd is BL(θ′)-dependent if we can show that the finite-dimensional distributions of (Zn(j))j Zd converg∈e to those of (Z(j))j Zd. First we will show ∈ ∈ lim E Z (j) Z(j) =0, j Zd. (3) n n | − | ∈ →∞ Let ǫ>0. Due to Lemma 7, the field (X(t))t Rd is continuous in 1-mean and hence there is n such that ∈ E X(0) X(t) <ǫ for all t [0, 1]d. | − | ∈ n Since (Xt)t Rd is stationary, this implies ∈ E|X(j1+ kn1,...,jd+ knd)−X(t)|<ǫ for all t∈[j1+ k1n−1,j1+ kn1]×···×[jd+ kdn−1,jd+ knd]. Hence E Z (j) Z(j) <ǫ, which finishes the proof of (3). n Now l|et j(1),−...,j(r|) Zd and let δ >0. From Markov’s inequality we get ∈ r r E Z (j(l)) Z(j(l)) P Z (j(l)) Z(j(l)) >δ l=1 | n − | 0, n . n | − | ≤ δ → →∞ (cid:16)Xl=1 (cid:17) P Sothefinite-dimensionaldistributionsof(Zn(j))j Zd convergetothoseof(Z(j))j Zd andhence(Z(j))j Zd is BL(θ)-dependent. ∈ ∈ ∈ By Lemma 8, the assumption that (X(t))t Rd is stationary implies that (Z(j))j Zd is stationary. Moreover,(Z(j))j Zd is centered, since ∈ ∈ ∈ EZ(0)=E X(t)dt EX(0)= EX(t)dt EX(0)=0. − − Z[0,1)d Z[0,1)d 4 THE MULTIVARIATE CLT 7 Further, Cov Z(0),Z(j) = Cov X(s),X(t) dtds jX∈Zd (cid:0) (cid:1) jX∈ZdZ[0,1)dZj+[0,1)d (cid:0) (cid:1) = Cov X(0),X(t s) dtds Z[0,1)dZRd − (cid:0) (cid:1) = Cov X(0),X(t) dtds Z[0,1)dZRd (cid:0) (cid:1) = Cov X(0),X(t) dt. ZRd (cid:0) (cid:1) We put Q := j Zd j +[0,1)d W and W := j +[0,1)d . As explained in the proof of [3, Thneorem{ 1∈.2], th|e assumption⊆thatn}(Wn)n Nn−is VH-gjr∈oQwning implies that (Qn)n N is regular ∈ S (cid:0) (cid:1) ∈ growing. Now Theorem 1.12 of [2, p. 178] implies that Wn−X(t)dt−λd(Wn−)EX(0) = j∈QnZ(j) (0,σ2), n . √#Q →N →∞ R λd(Wn−) P n q If we can show that Wn\Wn−X(t)dt−λd(Wn\Wn−)EX(0) P 0, n , (4) R λd(Wn) → →∞ p then Slutzki’s theorem will imply the assertion, since, clearly, λd(Wn−)/ λd(Wn) 1. We get → q p Var X(t)dt = Cov(X(s),X(t))dtds (cid:16)ZWn\Wn− (cid:17) ZWn\Wn−ZWn\Wn− Cov(X(s),X(t)) dtds ≤ZWn\Wn−ZRd| | =λ (W W ) Cov(X(0),X(t)) dt. d n\ n− ZRd| | Since (Wn)n N is VH-growing,we get ∈ Var Wn\Wn−X(t)dt = Var Wn\Wn−X(t)dt 0, n . (cid:18)R λd(Wn) (cid:19) (cid:0)R λd(Wn) (cid:1) → →∞ By the Chebyshev inequalitypthis implies (4). 4 The multivariate CLT In this section we extend Theorem 1 in various ways to multivariate central limit theorems. Theorem 10. Let θ = (θr)r N be a monotonically decreasing zero sequence. Let (X(t))t Rd be an Rs- valued random field. Assume∈that (X(t))t Rd is stationary, measurable, BL(θ)-dependent a∈nd fulfills ∈ Cov(X (0),X (t)) dt< , i,j =1,...,s. i j ZRd| | ∞ Let (Wn)n N be a VH-growing sequence of subsets of Rd. Then ∈ X (t)dt EX (0)λ (W ) X (t)dt EX (0)λ (W ) Wn 1 − 1 d n ,..., Wn s − s d n (0,Σ), n , R λd(Wn) R λd(Wn) →N →∞ (cid:16) (cid:17) in distribution, whpere Σ is the matrix with entries p Cov(X (0),X (t))dt, i,j =1,...,s. i j ZRd 4 THE MULTIVARIATE CLT 8 Proof: Let u = (u1,...,us) Rs. Then ( X(t),u )t Rd is BL(θ′)-dependent for a monotonically sdteactrieoansainrygasneqdumenecaesuθr′ab=le(.θWr′)re∈∈hNavweith limr→∞hθr′ = 0idu∈e to Lemma 4. Obviously, (hX(t),ui)t∈Rd is s s Cov( X(0),u , X(t),u )dt= u u Cov(X (0),X (t))dt=uTΣu. i j i j ZRd h i h i i=1j=1 ZRd XX In particular, the integral is defined. So Theorem 1 implies X (t)dt EX (0)λ (W ) X (t)dt EX (0)λ (W ) Wn 1 − 1 d n ,..., Wn s − s d n ,u R λd(Wn) R λd(Wn) D(cid:16) (cid:17) E X(t),u dt E X(0),u λ (W ) = Wnh p i − h i d n (p0,uTΣu), n . R λd(Wn) →N →∞ Since Y,u (0,uTΣu) for a rapndom vector Y (0,Σ), the Theorem of Cram´er and Wold implies h i∼N ∼N the assertion. Corollary 11. Let (X(t))t Rd be a stationary, measurable R-valued random field and let f1,...,fs : R R be functions. Let (W∈n)n N be a VH-growing sequence of subsets of Rd. Assume that one of the → ∈ following conditions holds: (i) The field (X(t))t Rd is BL(θ)-dependent for a monotonically decreasing zero sequence θ =(θr)r N, the maps f ,...,∈f are Lipschitz continuous and ∈ 1 s Cov f (X(0)),f (X(t)) dt< , i,j =1,...,s. i j ZRd ∞ (cid:12) (cid:0) (cid:1)(cid:12) (cid:12) (cid:12) (ii) The field (X(t))t Rd is QA and there are c>0 and ǫ>0 with ∈ Cov(X(0),X(t)) c t d ǫ, t Rd. (5) − − ≤ ·k k∞ ∈ The maps f ,...,f are Lipschitz continuous. 1 s (iii) The field (X(t))t Rd is PA with EX(0)2 < . The maps f1,...,fs are of locally bounded variation with E[h (X(0))∈2]< , i=1,...,s, and t∞here are c>0 and ǫ>0 with fi ∞ Cov h (X(0)),h (X(t)) c t d ǫ, t Rd, i,j =1,...,s. (6) fi fj ≤ ·k k∞− − ∈ (cid:0) (cid:1) Then f (X(t))dt Ef (X(0))λ (W ) f (X(t))dt Ef (X(0))λ (W ) Wn 1 − 1 d n ,..., Wn s − s d n (0,Σ), R λd(Wn) R λd(Wn) →N (cid:16) (cid:17) as n in distribuption, where Σ is the matrix with entries p →∞ Cov f (X(0)),f (X(t)) dt, i,j =1,...,s. i j ZRd (cid:0) (cid:1) Part (i) of this corollary is an immediate consequence of Lemma 4 and Theorem 10. Proof of Corollary 11(ii): The field (X(t))t Rd is BL(θ)-dependent by Lemma 5 and thus Lemma 4 implies that the field (f1(X(t)),...,fs(X(t)))t R∈d is also BL(θ)-dependent. In order to check the integrability assumpt∈ions from part (i), we put N if f (x)< N j − − f(N) :x f (x) if f (x) [ N,N] j 7→ j j ∈ − N if fj(x)>N.  REFERENCES 9 Since (X(t))t Rd is QA, we get ∈ Cov f(N)(X(0)),f(N)(X(t)) Lip(f(N)) Lip(f(N)) Cov(X(0),X(t)) | i j |≤ i · j ·| | (cid:0) (cid:1) Lip(fi) Lip(fj) Cov(X(0),X(t)). ≤ · ·| | By the Monotone Convergence Theorem, applied to both summands of E[f(N)(X(0))f(N)(X(t))] i j − E[f(N)(X(0))] E[f(N)(X(t))], this yields i · j Cov f (X(0)),f (X(t)) Lip(f ) Lip(f ) Cov(X(0),X(t)). i j i j | |≤ · ·| | Moreover,(5) implies (cid:0) (cid:1) Cov X(0),X(t) dt< ZRd ∞ (cid:12) (cid:0) (cid:1)(cid:12) and hence (cid:12) (cid:12) Cov f (X(0)),f (X(t)) dt< , i,j =1,...,s. i j ZRd ∞ (cid:12) (cid:0) (cid:1)(cid:12) So part (i) yields the assertio(cid:12)n. (cid:12) Proof of Corollary 11(iii): Since (X(t))t Rd is PA, the random field (hf1(X(t)),...,hfs(X(t)))t Rd is also PA, see Theorem 1.8(d) of [2, p. 7], a∈nd therefore QA. By Lemma 5 it is BL(θ)-dependent∈for somemonotonicallydecreasingzerosequenceθ. Hence (f1(X(t)),...,fs(X(t)))t Rd isBL(θ′)-dependent for some monotonically decreasing zero sequence θ by Lemma 9 and Lemma 4.∈ ′ Clearly, the field (f1(X(t)),...,fs(X(t)))t Rd is also stationary and measurable. Moreover,(6) implies ∈ Cov h (X(0)),h (X(t)) dt< , i,j =1,...,s. ZRd fi fj ∞ (cid:12) (cid:0) (cid:1)(cid:12) (cid:12) (cid:12) Now Lemma 9 and the QA property of (hf1(X(t)),...,hfs(X(t)))t Rd give ∈ Cov f (X(0)),f (X(t)) dt Cov h (X(0)),h (X(t)) dt< , i,j =1,...,s. ZRd i j ≤ZRd fi fj ∞ (cid:12) (cid:0) (cid:1)(cid:12) (cid:12) (cid:0) (cid:1)(cid:12) So Th(cid:12)eorem 10 yields the assert(cid:12)ion. (cid:12) (cid:12) References [1] R.J. Adler, J.E. Taylor, K.J. Worsley: Applications of Random Fields and Geometry: Founda- tions and case studies. Springer Series in Statistics. New York: Springer. In preparation, 2009. http://webee.technion.ac.il/people/adler/hrf.pdf [2] A.Bulinski,A.Shashkin: LimitTheorems forAssociatedRandomFieldsandRelatedSystems.World Scientific Publishing, New Jersey, 2007. [3] A. Bulinski, E. Spodarev, F. Timmermann: Central limit theorems for the excursion sets volumes of weakly dependent random fields. Bernoulli 18, 2012, 100–118. [4] A. Bulinski, Z. Zhurbenko: A central limit theorem for additive random functions. Theory of Prob- ability and its Applications 21, 1976, 687– 697. [5] J. Chil´es, P. Delfiner: Geostatistics: Modeling Spatial Uncertainty. Wiley, New York, 2007. [6] I. Gikhman, A. Skorokhod: The Theory of Stochastic Processes I, Springer, 2004. [7] V.V.Gorodetskii: Thecentrallimittheoremandaninvarianceprincipleforweaklydependentrandom fields. Soviet Mathematics. Doklady 29(3), 1984,529–532. REFERENCES 10 [8] A.V. Ivanov, N.N. Leonenko: Statistical Analysis of Random Fields. Kluwer, Dordrecht, 1989. [9] N.N. Leonenko: The central limit theorem for homogeneous random fields, and the asymptotic nor- mality of estimators of the regression coefficients. Dopov¯ıd¯ıAkadem¯ı¨ıNauk Ukra¨ınsko¨ıRSR. Seriya A, 1974,699–702. [10] K.Mecke,D.Stoyan: Morphology ofCondensedMatter-Physics andGeometryofSpatiallyComplex Systems. Springer, 2002. [11] C.N.Morris: Naturalexponentialfamilies withquadraticvariance functions.TheAnnalsofStatistics 10, 1982, 65–80. [12] P. Roy: Nonsingular group actions and stationary SαS random fields. Proceedings of the American Mathematical Society 138, 2010, 2195–2202. [13] J.E. Taylor, K.J. Worsley: Detecting sparse signal in random fields, with an application to brain mapping. Journal of the American Statistical Association 102 (479), 2007, 913–928. [14] S. Torquato: Random Heterogeneous Materials - Microstructure and Macroscopic Properties. Springer, 2001. [15] H. Wackernagel: Multivariate Geostatistics: An Introduction with Applications. Springer, 2003.

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.